Nicole Berline Ezra Getzler

Michele Vergne

Heat Kernels and

Dirac Operators

Springer
Nicole Berline
Centre de Mathematiques
Ecole Polytechnique
F-91128 Palaiseau Cedex, France

Ezra Getzler
MIT
Cambridge, MA 02139, USA

Michele Vergne
DMI, Ecole Normale Sup6rieure
45, rue d'Ulm
F-75005 Paris, France

Corrected Second Printing 1996

Mathematics Subject Classification (1991):

58G10,58011,58A10,53C05

Library of Congress Cataloging-in-Publication Data

Berline, Nicole, 1944-
Heat kernels and Dirac operators / Nicole Berline, Ezra Getzler,
Michele Vergne.
p. cm. -- (Grundlehren der mathematischen Wissenschaften
298)
"Second corrected printing 1996"--T.p. verso.
Includes bibliographical references and index.
ISBN 0-387-53340-0 (ai'k. paper)
1. Heat equation. 2. Dirac equation. 3. Index theorems.
4. Differential forms. I. Getzler, Ezra. 1962- II. Vergne,
.

Michele. III. Title. IV. Series.

OA377.B49 1996
515'.353--dc2O 95-51231
CIP

ISBN 3-540-53340-0 Springer-Verlag Berlin Heidelberg New York

ISBN 0-387-53340-0 Springer-Verlag New York Berlin Heidelberg
This work is subject to copyright. All rights are reserved, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilm or in any other way,
and storage in data banks. Duplication of this publication or parts thereof is permitted
only under the provisions of the German Copyright Law of September 9, 1965, in its
current version, and permission for use must always be obtained from Springer-
Verlag. Violations are liable for prosecution under the German Copyright Law.
il Springer-Verlag Berlin Heidelberg 1992
Printed in Germany
Typesetting: By the authors in AMS-LATEX
SPIN 10476724 41/3143-5 4 3 2 10 Printed on acid-free paper
Contents

Introduction 1
Acknowledgements ........................................... 13

Chapter 1. Background on Differential Geometry 14

1.1. Fibre Bundles and Connections ............................ 14
1.2. Riemannian Manifolds ................................... 32
1.3. Superspaces ............................................ 38
1.4. Superconnections ........................................ 43
1.5. Characteristic Classes .................................... 46
1.6. The Euler and Thom Classes ............................. 51
Bibliographic Notes .......................................... 61

Chapter 2. Asymptotic Expansion of the Heat Kernel 63

2.1. Differential Operators ................... 64
2.2. The Heat Kernel on Euclidean Space ....................... 72
2.3. Heat Kernels ............................................ 74
2.4. Construction of the Heat Kernel ........................... 77
2.5. The Formal Solution ..................................... 81
2.6. The Trace of the Heat Kernel ............................. 87
2.7. Heat Kernels Depending on a Parameter .................... 98
Bibliographic Notes .......................................... 101
Chapter 3. Clifford Modules and Dirac Operators 102
3.1. The Clifford Algebra ..................................... 103
3.2. Spinors ................................................ 109
3.3. Dirac Operators ......................................... 113
3.4. Index of Dirac Operators ................................. 122
3.5. The Lichnerowicz Formula ................................ 126
3.6. Some Examples of Clifford Modules ........................ 127
The DeRham Operator .................................. 128
The Signature Operator .................................. 131
The Dirac Operator on a Spin-Manifold .................... 134
The 0-Operator on a Kahler Manifold ...................... 135
Bibliographic Notes .......................................... 141
vi Contents

Chapter 4. Index Density of Dirac Operators 143

4.1. The Local Index Theorem ................................ 143
The Gauss-Bonnet-Chern Theorem ........................ 148
The Hirzebruch Signature Theorem ........................ 150
The Index Theorem for the Dirac Operator ................. 151
The Riemann-Roch-Hirzebruch Theorem ................... 151
4.2. Mehler's Formula ........................................ 153
4.3. Calculation of the Index Density ........................... 157
Bibliographic Notes .......................................... 164
Additional Remarks .......................................... 166
Chapter 5. The Exponential Map and the Index Density 167
5.1. Jacobian of the Exponential Map on Principal Bundles ....... 168
5.2. The Heat Kernel of a Principal Bundle ..................... 172
5.3. Calculus with Grassmann and Clifford Variables ............. 178
5.4. The Index of Dirac Operators ............................. 181
Bibliographic Notes .......................................... 183
Chapter 6. The Equivariant Index Theorem 184
6.1. The Equivariant Index of Dirac Operators .................. 185
6.2. The Atiyah-Bott Fixed Point Formula ...................... 186
6.3. Asymptotic Expansion of the Equivariant Heat Kernel ....... 190
6.4. The Local Equivariant Index Theorem ..................... 193
6.5. Geodesic Distance on a Principal Bundle ................... 197
6.6. The Heat Kernel of an Equivariant Vector Bundle .......... 199
6.7. Proof of Proposition 6.13 ................................. 202
Bibliographic Notes .......................................... 204
Chapter 7. Equivariant Differential Forms 205
7.1. Equivariant Characteristic Classes ......................... 206
7.2. The Localization Formula ................................ 213
7.3. Bott's Formulas for Characteristic Numbers ................. 221
7.4. Exact Stationary Phase Approximation ..................... 224
7.5. The Fourier Transform of Coadjoint Orbits ................. 226
7.6. Equivariant Cohomology and Families ...................... 232
7.7. The Bott Class .......................................... 239
Bibliographic Notes .......................................... 244
Additional Remarks .......................................... 245

Chapter 8. The Kirillov Formula for the Equivariant Index 246

8.1. The Kirillov Formula .................................... 247
8.2. The Weyl and Kirillov Character Formulas ................. 251
8.3. The Heat Kernel Proof of the Kirillov Formula .............. 256
Bibliographic Notes .......................................... 265
Additional Remarks .......................................... 266
Contents vii

Chapter 9. The Index Bundle 267

9.1. The Index Bundle in Finite Dimensions .................... 269
9.2. The Index Bundle of a Family of Dirac Operators ............ 277
9.3. The Chern Character of the Index Bundle .................. 281
9.4. The Equivariant Index and the Index Bundle ................ 292
9.5. The Case of Varying Dimension ........................... 294
9.6. The Zeta-Function of a Laplacian .......................... 299
9.7. The Determinant Line Bundle ............................. 304
Appendix. More on Heat Kernels ............................... 309
Bibliographic Notes .......................................... 315
Chapter 10. The Family Index Theorem 316
10.1. Riemannian Fibre Bundles ............................... 319
10.2. Clifford Modules on Fibre Bundles ........................ 324
10.3. The Bismut Superconnection ............................. 331
10.4. The Family Index Density ................................ 335
10.5. The Transgression Formula ............................... 343
10.6. The Curvature of the Determinant Line Bundle ............. 347
10.7. The Kirillov Formula and Bismut's Index Theorem .......... 350
Bibliographic Notes .......................................... 354
Bibliography 355

List of Notations 360

Index 364
Introduction

Dirac operators on Riemannian manifolds, which were introduced in the arti-

cles of Atiyah and Singer [13] and Lichnerowicz [80], are of fundamental impor-
tance in differential geometry: they occur in situations such as Hodge theory,
gauge theory, and geometric quantization, to name just a few examples. Most
first-order linear differential operators of geometric origin are Dirac operators.
After Atiyah and Singer's fundamental work on the index for general el-
liptic operators, methods based on the heat kernel were applied to prove the
Atiyah-Singer Index Theorem in the special case of Dirac operators, by Pa-
todi [89], Gilkey [64] and Atiyah-Bott-Patodi [8]. In recent years, new insights
into the local index theorem of Patodi and Gilkey have emerged, which have
simplified the proofs of their results, and permitted the extension of the lo-
cal index theorem to other situations. Thus, we felt it worthwhile to write
a book in which the Atiyah-Singer Index Theorem for Dirac operators on
compact Riemannian manifolds and its more recent generalizations would re-
ceive elementary proofs. Many of the theorems which we discuss are due
to J.M. Bismut, although we have replaced his use of probability theory by
classical asymptotic expansion methods.
Our book is based on a simple principle, which we learned from D. Quillen:
Dirac operators are a quantization of the theory of connections, and the super-
trace of the heat kernel of the square of a Dirac operator is the quantization
of the Chern character of the corresponding connection. From this point of
view, the index theorem for Dirac operators is a statement about the relation-
ship between the heat kernel of the square of a Dirac operator and the Chern
character of the associated connection. This relationship holds at the level of
differential forms and not just in cohomology, and leads us to think of index
theory and heat kernels as a quantization of Chern-Weil theory.
Following the approach suggested by Atiyah-Bott and McKean-Singer, and
pursued by Patodi and Gilkey, the main technique used in the e_tD2
book is an
explicit geometric construction of the kernel of the heat operator asso-
ciated to the square of a Dirac operator D. The importance of the heat kernel
is that it interpolates between the identity operator, when t = 0, and the pro-
jection onto the kernel of the Dirac operator D, when t = oo. However, we will
study the heat kernel, and more particularly its restriction to the diagonal, in
its own right, and not only as a tool in understanding the kernel of D.
2 Introduction

Lastly, we attempt to express all of our constructions in such a way that

they generalize easily to the equivariant setting, in which a compact Lie group
G acts on the manifold and leaves the Dirac operator invariant.
We will consider the most general type of Dirac operators, associated to a
Clifford module over a manifold, to avoid restricting ourselves to manifolds
with spin structures. We will also work within Quillen's theory of supercon-
nections, since this is conceptually simple, and is needed for the formulation
of the local family index theorem of Bismut in Chapters 9 and 10.
We will now give a rapid account of some of the main results discussed
in our book. Dirac operators on a compact Riemannian manifold M are
closely related to the Clifford algebra bundle. The Clifford algebra CC(M) at
the point x E M is the associative complex algebra generated by cotangent
vectors a E TxM with relations
al a2 + a2 al = -2(a,, a2),
where (al, a2) is the Riemannian metric on T.,M. If ei is an orthonormal
basis of TM with dual basis ei, then this amounts to saying that CA(M) is
generated by elements ci subject to the relations
(Ci)2=-1, andcc!+cica=0fori#j.
The Clifford algebra CA(M) is a deformation of the exterior algebra AT XM,
and there is a canonical bijection ox : Cx(M) -+ AT*ZM, the symbol map,
defined by the formula
o-:,; ...ci?) = eil A . . . A
The inverse of this map is denoted by c,; : AT xM --+ CA(M).
Let £ be a complex Z2-graded bundle on M, that is, £ = £+ ® £-. We
say that £ is a bundle of Clifford modules, or just a Clifford module, if there
is a bundle map c : T*M -+ End(S) such that
(1) c(al)c(a2) +c(a2)c(al) = -2(al,a2), and
(2) c(a) swaps the bundles £+ and £-.
That is, £., is a Z2-graded module for the algebra CA(M). If M is even-
dimensional, the Clifford algebra CC (M) is simple, and we obtain the decom-
position
End(S) = C(M) ® Endc(M) (£).
From now on, most of our considerations only apply to even-dimensional
oriented manifolds. If M is a spin manifold, that is, a Riemannian manifold
satisfying a certain topological condition, there is a Clifford module S, known
as the spinor bundle, such that End(S) = C(M). On such a manifold, any
Clifford module may be written as a twisted spinor bundle W ® S, with
W = Homc(M)(S,£). Let rM E r(M,C(M)) be the chirality operator in
C(M), given by the formula
rM = idim(M)/2C1 ... cn",
so that I'M=1.
Introduction 3

If £ is a vector bundle on M, let r(M, £) be the space of smooth sections

of £, and let A(M, £) = r(M, AT*M (9 6) be the space of differential forms
on M with values in S. We make the obvious, but crucial, remark that if 6
is a Clifford module, then by the symbol map, the space of sections
r(M, End(£)) = r(M, C(M) ® Endc(M) (£))
is isomorphic to the space of bundle-valued differential forms
A(M, Endc(M) (£)) = r(M, AT*M ® Endc(M) (£)).
Thus, a section k E r(M, End(£)) corresponds to a differential form a(k)
with values in Endc(M)(£). If M is a spin manifold and £ = W ® S is a
twisted spinor bundle, o(k) is a differential form with values in End(W).
A Clifford connection on a Clifford module £ is a connection VE on £
satisfying the formula
[vx, c(a)] = c(VXa)
where a is a one-form on M, X is a vector field, and VXa is the Levi-Civita
derivative of a.
The Dirac operator D associated to the Clifford connection yE is the com-
position of arrows

r(M, £) r(M,T*M ®£) - r(M,£).

With respect to a local frame ei of T*M, D may be written
D=
i
A number of classical first-order elliptic differential operators are Dirac
operators associated to a Clifford connection. Let us list three examples:
(1) The exterior bundle AT*M is a Clifford bundle with Clifford action by
the one-fom a E r(M, T*M) defined by the formula
c(a) = e(a) - E(a)*;

here E(a) : r(M, -i r(M, is the operation of exterior

multiplication by a. The Levi-Civita connection on AT*M is a Clifford
connection. The associated Dirac operator is d+d*, where d is the exterior
differential operator. The kernel of this operator is just the space of
,

harmonic forms on M, which by Hodge's Theorem is isomorphic to the

de Rham cohomology H (M) of M.
(2) If M is a complex manifold and W is a Hermitian bundle over M, the
bundle A(T°,1M)* ® W is a Clifford module, with a a1,o + a0,1 E
A1,0(M) e A,' (M) acting by
c(a) = v(E(ao,1) - E(al,o)*)
The Levi-Civita connection on AT*M preserves A(T°,1M)* and defines
a Clifford connection if M is Kahler, and the Dirac operator associated
4 Introduction

to this connection is vf2_(a + a*). If W is a holomorphic vector bundle

with its canonical connection, the tensor product of this connection with
the Levi-Civita connection on A(T°,1M)* is a Clifford connection with
associated Dirac operator v'2(6 + a*). The kernel of this operator is the
space of harmonic forms on M lying in A°'' (M, W), which by Dolbeault's
theorem is isomorphic to the sheaf cohomology H (M, W).
(3) If M is a spin manifold, its spinor bundle S is a Clifford module, the
Levi-Civita connection is a Clifford connection, and the associated Dirac
operator is known simply as the Dirac operator. Its kernel is the space
of harmonic spinors.
Thus, we see from these examples that the kernel of a Dirac operator often
has a topological, or at least geometric, significance.
The heat kernel (x I e-tD2 I y) E Hom(£y, £,,) of the square of the Dirac
operator D is the kernel of the heat semigroup a-tD2, that is,

(e-tD2 s) (x) _ f (x I e-02 I y)s(y) Idyl for all s E 1 (M, £),

where Idyl is the Riemannian measure on M. The following properties of the

heat kernel are proved in Chapter 2:
(1) it is smooth;
(2) the fact that smooth kernels are trace-class, from which we see that the
kernel of D is finite-dimensional;
(3) the uniform convergence of (x I e-tD2 I y) to the kernel of the projection
onto ker(D) as t --4 oo;
(4) the existence of an asymptotic expansion for (x I e-tD2 I y) at small t,
(where dim(M) = n = 22),
00
(x I e-t°2 y (4-7rt)-te-d(x,y)2/4t T tif2(x, y),
i=O

where fi is a sequence of smooth kernels for the bundle £ given by local

functions of the curvature of V& and the Riemannian curvature of M, and
d(x, y) is the geodesic distance between x and y.
Note that the restriction to the diagonal x (x I e-tD2 I x) is a section of
End(£). The central object of our study will be the behaviour at small time
of the differential form

o,((x I e-tD2 I x)) E A(M,EndC(M)(£)),

obtained by taking the image under the symbol map o of (x I e-tD2 I x).
Let us describe the differential forms which enter in the study of the asymp-
totic expansion of (x I e-tD2 1 x). If g is a unimodular Lie algebra, let j9(X)
Introduction 5

be the analytic function on g defined by the formula

smh(ad X/2)
jg (X) = det
ad X/2 '

it is the Jacobian of the exponential map exp : g --> G. We define in a

neighbourhood of 0 E g to be the square-root of j,, such that 1.
Let R E A2 (M, so (TM)) be the Riemannian curvature of M. Choose a
local orthonormal frame ei of TM, and consider the matrix R with two-form
coefficients,
Rij = (Ref, ej) E A2(M).
Then (S"'R(2/2)) is a matrix with even degree differential form coefficients,
and since the determinant is invariant under conjugation by invertible matri-
ces,
(sinh(R/2))
j(R) = det
is an element of A(M) independent of the frame of TM used in its definition.
Note that the zero-form component of j(R) equals 1. Thus, we can define the
A-genus A(M) of the manifold M by

A(M) = j (R) -1/2 = det1/2 (sink(/R/2)) E A(M);

it is a closed differential form whose cohomology class is independent of the

metric on M. It is a fascinating puzzle that the function jg 1/2 occurs in a
basic formula of representation theory for Lie groups, the Kirillov character
formula, while its cousin, the A-genus, plays a similar role in a basic formula
of differential geometry, the index theorem for Dirac operators. Understanding
the relationship between these two objects is one of the aims of this book.
Note that our normalization of the A-genus, and of other characteristic
classes, is not the same as that preferred by topologists, who multiply the
2k-degree component by (-2lri)k, so that it lies in H2k(M, Q). We prefer to
leave out these powers of -27ri, since it is in this form that they will arise in
the proof of the local index theorem.
Let £ be a Clifford module on M, with Clifford connection £ and curvature
Fs. The twisting curvature Fs/s of £ is defined by the formula
Fs/s = Fe - Rs E A2 (M, Endc(M) (£)),
where
Rs (ei, ej) = 2 (R(ei, ej )ek, et)ckcl.
k<I
If M is a spin manifold with spinor bundle S and S = W ® S, Fe/s is the
curvature of the bundle W.
For a E r(M, C(M)) = r(M, AT*M), we denote the k-form component of
o(a) by vk(a).
6 Introduction

The first four chapters lead up to the proof of the following theorem, which
calculates the leading order term, in a certain sense, of the heat kernel of a
Dirac operator restricted to the diagonal.
e_t02
Theorem 0.1. Consider the asymptotic expansion of (x I I x) at small
times t,
ao
e_tD2
xI I x)
%=o

with coefficients ki E I'(M, C(M) ®EndC(M) (£)). Then

(1) o (ki) = 0 for j > 2i
(2) Let Q(k) = oo2i(ki) E A(M,EndClyti(£)). Then

Q(k) = det1/2(sinh(R/2)
J
In Chapter 4, we give a proof of Theorem 0.1 which relies on an approx-
imation of D2 by a harmonic oscillator, which is easily derived from Lich-
nerowicz's formula for the square of the Dirac operator, and properties of the
normal coordinate system.
Since the zero-form piece of the A-genus equals 1, we recover Weyl's for-
mula
lim (4irt)e(x a-t°2 I x) = rk(£);
t-40
in this sense, Theorem 0.1 is a refinement of Weyl's formula for the square of
a Dirac operator.
Define the index of D to be the integer
ind(D) = dim(ker(D+)) dim(ker(D-)),
where D is the restriction of D to P(M, £t). For example, ind(d+d*) is the
Euler characteristic

Eul(M) = (-1)i dim(Hi(M))

i=0
of the manifold M, while ind(a + a*) is the Euler characteristic of the sheaf
of holomorphic sections
e
Eul(M, W) _ (-1)i dim(Hi(M, W)).
i=0
These indexes are particular cases of the Atiyah-Singer Index Theorem, and
are given by well-known formulas, respectively the Gauss-Bonnet-Chern the-
orem and the Riemann-Roch-Hirzebruch theorem; we will show in Section 4.1
how these formulas follow from Theorem 0.1.
In Chapter 3, we establish some well-known properties of the index, such
as its homotopy invariance, using the formula of McKean and Singer. If E
Introduction 7

is a Z2-graded vector space and A is an endomorphism on E, we define the

supertrace Str(A) of A to be the trace of the operator FA, where r E End(E)
is the chirality operator, which equals ±1 on E±. The McKean-Singer formula
says that for each t > 0, the index of D equals
Str(e-tD2)
ind(D) = = Str(x I e-tD2 I x) Jdxl.
fM
This formula shows that the index of D is given in terms of the restriction to
the diagonal of the heat kernel of D2 at arbitrarily small times, which we know
by the asymptotic expansion of the heat kernel to be given by local formulas
in the curvature of the connection Ve and the Riemannian curvature of M.
Using the McKean-Singer formula, we see that

ind(D) _ (4-7r) -n/2 Str(k,,,/2(x)) dx.

JM
From Theorem 0. 1, we see that
a(kn/2(x))[n] E An(M, Endc(M)(£)) = I(M,EndC(M)(£))
equals the n-form component of
1/2
detl/2 (sinh(R/2) J exp(-FE/S) E A (M, EndC(M) (£)).
The differential form
ch(£/S) = 2-n/2 Str(FM exp(-Fs/S))
is a generalization of the Chern character for Clifford modules, called the
relative Chern character; when M is a spin manifold and S = W ® S, the
form ch(S/S) equals the Chern character ch(W) of the twisting bundle W.
In this way, we obtain the formula
fM
ind(D) = (27ri)-n/2 A(M) ch(£/S).
J
This is the Atiyah-Singer index theorem for the Dirac operator D. Thus,
we see that our main theorem is at the same time a generalization of Weyl's
formula for the leading term of the heat kernel, and of the Atiyah-Singer index
theorem. As conjectured by McKean-Singer, we establish that
e-tD2

lim Str (x I I x)
t-40
exists and is given by the above integrand. This result, due to Patodi [89]
and Gilkey [64], is known as the local index theorem. It is very important,
since it generalizes to certain situations inaccessible to the global index theo-
rem of Atiyah and Singer, such as when the manifold M is no longer closed.
On important example is the signature theorem for manifolds with bound-
ary of Atiyah-Patodi-Singer [Ill. As another example, it is possible in the
case of infinite covering spaces to prove a local index theorem which gives
8 Introduction

the "index per unit volume" of the Dirac operator in question (Connes and
Moscovici [50]).
Note that the proofs of Patodi and Gilkey of the local index theorem re-
quired the calculation of the index of certain Dirac operators on homogeneous
spaces. In the proof which we present, we are able to calculate the index den-
sity by analysis of Lichnerowicz's formula for the square of the Dirac operator,
which was proved in [80].
In Chapter 5, we give another proof of Theorem 0.1. This proof, while
similar in some respects to the first, uses an integral expression for the heat
kernel of D2 in terms of the scalar heat kernel of the orthonormal frame
bundle SO(M); the Jacobian of the exponential map on SO(M) leads to the
appearance of the A-genus. Once more, Lichnerowicz's formula is fundamental
to the proof.
The second topic of this book is an equivariant version of the results of
Chapter 4. Suppose that a compact Lie group G acts on the manifold M
and on the bundle £, and that the operator D is invariant under this action
(for this to be possible, G must act on M by isometries). In this situation,
the action of the group G leaves the kernel of D invariant; in other words,
ker(D) is a representation of G. Thus, the index of D may be generalized to
a character-valued index, defined by the formula
indc(y, D) = Tr(y, ker(D+)) - Tr(-y, ker(D-)).
The analogue of the McKean-Singer formula holds here:

indG(y, D) = fMStr(x I rye e 2 I ) I dxj,

where 'y£ is the action of y E G on the space of sections of the vector bundle
S. In Chapter 6, we prove a formula, due to Gilkey [65], for the distribution
lira Str (x I ye CID' I x).
t->o
It is a product of the delta-function along the fixed point set Mry of the action
of y on M and a function involving the curvature of My, the curvature of
£, and the curvature of the normal bundle N of My. Integrating over M,
we obtain an expression for ind0(y, D) as an integral over My, which is a
theorem due to Atiyah, Segal and Singer [12], [15]. The proof that we give is
modelled on that of Chapter 5, and is hardly more difficult.
The Weyl character formula for representations of compact Lie groups is
a special case of the equivariant index theorem for Dirac operators. By the
Borel-Weil-Bott theorem, a finite-dimensional representation of G may be
realized as the kernel of a Dirac operator on the flag variety G/T of G, where
T is a maximal torus of G. If g is a regular element of G, then the fixed-point
set (G/T)-9 is in one-to-one correspondence with the Weyl group, and the
fixed point formula for the index may be identified with the Weyl character
formula, as is explained by Atiyah-Bott [6].
Introduction 9

Generalizing ideas of Bott [42] and Baum-Cheeger [18], we will show how
the fixed point formula for the equivariant index indG (ex, D) may be rewritten
as an integral over the whole manifold. To do this, we introduce the notion
of equivariant differential forms in Chapter 7; this is of independent interest.
Let M be a manifold with an action of a compact Lie group G; if X is in the
Lie algebra g of G, let XM be the corresponding vector field on M. Let C[g]
denote the algebra of complex valued polynomial functions on g. Let AG(M)
be the algebra of G-invariant elements of C[g] ® A(M), graded by
deg(P (9 a) = 2 deg(P) + deg(a)
for P E C[g] and a E A(M). We define the equivariant exterior differential
dg on C[g] ® A(M) by

(dga)(X) = d(a(X)) - c(X)(a(X)),

where t(X) denotes contraction by the vector field XM. Thus, dg increases
by one the total degree on C[g] ® A(M), and preserves AG(M). Cartan's
homotopy formula implies that dg vanishes on elements of AG(M); thus
(AG (M), dg) is a complex, called the complex of equivariant differential forms;
of course, if G = {1}, this is the deRham complex. Its cohomology HG(M)
is called the equivariant cohomology of the manifold with coefficients in I[8; it
was proved by H. Cartan [45] that this is the real cohomology of the homotopy
quotient M X G EG. The Chern-Weil construction of differential forms associ-
ated to a vector bundle with connection may be generalized to this equivariant
setting.
Integration over M defines a map

:AG(M)-;C[91G
fM

by the formula (fM a) (X) = fm a(X); this map vanishes on equivariantly

exact forms dg a, and hence defines a homomorphism from HH(M) to C[g]G.
An important property of equivariant differential forms is the localization
formula, a generalization of a formula of Bott, which expresses the integral
of an equivariantly closed differential form evaluated at X E g as an integral
over the zero set of the vector field XM. This formula has many applications,
especially when the vector field XM has discrete zeroes.
Using the localization theorem, we can rewrite the formula for the equivari-
ant index as an integral over the entire manifold, at least in a neighbourhood
of the identity in G.

Theorem 0.2. The equivariant index is given by the formula, for X E g

sufficiently small,

indG(e-X, D) = (27ri) -n/2 Ag(X, M) ch(X, E/S),

JM
10 Introduction

where chg (X, £/S) E Ac (M) is a closed equivariant differential form which
specializes to the relative Chern character at X = 0, and A. (X, M) E AG (M)
similarly specializes at X = 0 to the A-genus.
Thus, the integrand is a dg-closed equivariant differential form depending
analytically on X for X near 0 E g, and which coincides with A(M) ch(£/S)
atX=0.
Let G be a compact Lie group, and let T be a finite dimensional irreducible
representation of G. As we will see in Chapter 8, Kirillov's formula for the
characters of the representation T of G is a special case of the above equivari-
ant index theorem: this formula says that there is an orbit M of the coadjoint
representation of G on g* such that
T'r(T(eX)) e'(.f,X) d0(f),
= j8 12(X)
JM
where d13(f) is the Liouville measure on M.
This formula, which unlike the Weyl character formula has a generalization
to any Lie group, inspired our expression for the equivariant index in terms of
equivariant differential forms. Note the similarity with the equivariant index
theorem: the equivariant A-genus corresponds to the factor jg 112(X), while
the equivariant Chern character corresponds to the measure ei(f,x) d$(f).
The final major topic of the book is the local index theorem for families of
Dirac operators. This` theorem, which is due to Bismut, is a relative version
of the local index theorem, in which we consider a fibre bundle 7r : M - B
and a smooth family of Dirac operators Dz, one for each fibre Mz = 7r-1(z),
(z E B). The kernels of the Dirac operators Dz now fit together to form
a vector bundle, if they are of constant dimension. The index of a single
operator D, ind(D) = dim(ker(D+)) - dim(ker(D-)), is now replaced by the
notion of the index bundle, which is the difference bundle
ind(D) = [ker(D+)] - [ker(D-)];
following Atiyah-Singer, we define in Chapter 9 a smooth index bundle even if
the dimension of the kernel ker(Dz) is not constant, but it is no longer canon-
ical. The index theorem of Atiyah-Singer [16] for a family gives a formula for
the Chern character of this index bundle.
It turns out that the formulation of a local version of this family index the-
orem requires the introduction of generalization of connections due to Quillen,
known as superconnections. Let f =H+ ®n- be a Z2-graded vector bundle.
A superconnection on 7-l is an operator acting on A(M, l) of the form
A=A101+A[1] +A12>+...,
where A(,] is a connection preserving the Z2-grading on 7-l, and A[i] for i 0 1
are i-form valued sections of End(7l) which exchange 7-f+ and 7-1- for i even
and preserves them for i odd; usually, A[01 E r(M, End (H)) is supposed to be
self-adjoint as well. The curvature of a superconnection A2 is its square, which
Introduction 11

is an element of A(M, End(7I)) of the form A2 = Quillen showed

that as a de Rham cohomology class, the Chern character of the difference
bundle [7-l+] - [7-1-] is equal to the Chern character of A, defined by the
supertrace
ch(A) = Str(e-A2) = Str(e-Aro]+...),

which is a closed even degree differential form whose cohomology class is

independent of A. Superconnections play a role in two situations: to define the
Chern character on non-compact manifolds, where A[o] grows at infinity, thus
allowing ch(A) to represent an integrable cohomology class, and for infinite
2
dimensional vector bundles, where A[o] is such that e-A101 is trace-class. We
explain the extension of the Chern-Weil construction of characteristic forms
to the case of superconnections in Chapter 1.
Let D be an odd endomorphism of a finite-dimensional superbundle 7-L with
components D+ : 7-lt - 7-lw, such that ker(D) has constant rank, and let A
be a superconnection on 7-l with zero-degree term A[o] equal to D. At the
level of cohomology,
ch(A) = ch(ker(D+)) - ch(ker(D-)).
In Chapter 9, we prove a refinement of this equation which holds at the level
of differential forms.
If t > 0, we may define the rescaled superconnection on 71, by the formula
At = t112D + All] + t-1/2A[2] + ... ,
with Chern character
ch(At) = Str(e-At).
We define a connection on the bundle ker(D) by the formula

where Po is the projection from 7-l onto its sub-bundle ker(D); thus,
ch(Vo) = ch(ker(D+)) - ch(ker(D-)).
Intuitively speaking, as t tends to infinity, the supertrace
Str(e-tn2+o(t"2))

ch(At) =
is pushed onto the sub-bundle ker(D). In fact, we can prove the following
formula,
lim ch(At) = ch(Vo).
t->oo
Let M -+ B be a fibre bundle with compact Riemannian fibres of dimen-
sion n, and let £ be a bundle of modules for the vertical Clifford algebras
CC(M/B) associated to the vertical tangent spaces of M -> B. Bismut
showed that the above picture can be extended to an infinite-dimensional set-
ting when D is a smooth family of Dirac operators on the fibres £Z -> M,
where £x and M,z are the fibres over z E B of £ and M. He introduced the
12 Introduction

infinite-dimensional bundle it*£ whose fibre at z E B is the space of sections

of £ over the fibre Mx,
(i.£)_ = r(Mz7 £Z)
Given a connection on the fibre bundle M -+ B, that is, a choice of horizontal
tangent bundle in TM transverse to the vertical tangent bundle, we may
define a unitary connection on 7r*£, whose formula involves the connection on
E and the mean curvature of the fibres M,z. If we form the superconnection
At = t1/2D+07"E, then the heat kernel of At has a supertrace in A(B), since
+t1/2v7r"ED + (v7r.E)2
At = tD2

is a perturbation of the operator tD2. In this way, we may define ch(At) E

A(B).
Assume that ker(D) has constant rank. The projection Po V""E Po of
the connection V""E defines a connection Vo on the finite dimensional bundle
ker(D). We obtain the following transgression formula:
00
ch(At) - ch(Vo) = d Cf Str(d sS e-Aa) ds I .
t
In general, it is impossible to take the limit t -+ 0 in this formula. However,
Bismut showed that this limit exists if a certain term A[2] is added to A,
proportional to Clifford multiplication by the curvature of the fibre bundle
M -; B; the curvature of this superconnection satisfies a formula very similar
to the Lichnerowicz formula for the square of a Dirac operator. By methods
almost identical to those of Chapter 4, we prove the following theorem in
Chapter 10.

Theorem 0.3. Let R`B'I/B E A2(M,so(T(M/B))) be the curvature of the

vertical tangent bundle of M/B, with A-genus A(M/B). Then, if A is the
Bismut superconnection,

t o ch(At) = (27ri)-n/2 fMOB A(M/B) ch(£/S) E A(B).

This result has turned out to be of interest to mathematicians working in
such varied areas as string theory, Arakelov theory, and the theory of moduli
spaces of Yang-Mills fields. Interestingly, when the fibre bundle M -> B has
a compact structure group G and all of the data are compatible with G, the
above theorem is equivalent to the Kirillov formula for the equivariant index.
The book is not necessarily meant to be read sequentially, and consists of
four groups of chapters:
(1) Chapters 1, which gives various preliminary results in differential geom-
etry, and Chapter 7, on equivariant differential forms, do not depend on
any other chapters.
Acknowledgements 13

(2) Chapters 2, 3 and 4 introduce the main ideas of the book, and take the
reader through the main properties of Dirac operators, culminating in
the local index theorem.
(3) Chapters 5, 6 and 8 are on the equivariant index theorem, and may be
read after the first four chapters, although Chapter 7 is needed in Chap-
ter 8.
(4) Chapters 9 and 10 are on the family index theorem, and can be read
after the first four chapters, except for Sections 9.4 and 10.7, which have
Chapter 8 as a prerequisite.
In the second printing, we have added some additional remarks at the
ends of Chapters 4 and 7, in order to give an idea of how the book would
have changed if we were writing it now. We have also made a number of
small corrections throughout.

Acknowledgements
This book began as a seminar in 1985 at MIT, and we would like to thank
the other participants in the seminar, especially Martin Andler and Varghese
Mathai, for their spirited participation. Discussions with many other people
have been important to us, among whom we would like to single out Jean-
Michel Bismut, Dan Freed and Dan Quillen. Finally, we are pleased to be
able to thank all of those people who read all or part of the book as it
developed and who made many comments which were crucial in improving the
book, both mathematically and stylistically, especially J6an-Frangois Burnol,
Michel Duflo, Sylvie Paycha, Christophe Soule, and Shlomo Sternberg. We
also thank the referee for suggestions which have improved the exposition.
We have not attempted to give a definitive bibliography of this very large
subject, but have only tried to draw attention to the articles that have influ-
enced us.
To all of the following institutes and funds, we would like to express our
gratitude: the Centre for Mathematical Analysis of the ANU, the ENS-Paris,
the Harvard Society of Fellows, the IHES, MIT, and the Universite de Paris-
Sud. We also received some assistance from the CNRS, the NSF, and the
Sloan Foundation.
Chapter 1. Background on Differential Geometry

In this chapter, we will survey some of the notions of differential geometry

that are used in this book: principal bundles and their associated vector
bundles, connections and superconnections, and characteristic classes. For
further details about these topics, we refer the reader to the bibliographic
references at the end of the chapter.

1.1. Fibre Bundles and Connections

Definition 1.1. Let 7r : £ -- M be a smooth map from a manifold £ to
a manifold M. We say that (£, 7r) is a fibre bundle with typical fibre E
over M if there is a covering of M by open sets Ui and diffeomorphisms
Oi : 7r-1(Ui) -> Ui x E, such that -7r : 7r-1(Ui) --> Ui is the composition of Oi
with projection onto the first factor Ui in Ui x E. The space £ is called the
total space of the fibre bundle, and M is called the base.
It follows from this definition that -7r-1(x) is diffeomorphic to E for all
x E M; we will call E "the" fibre of E. In this book, we will often make use
of the convenient convention that the bundles denoted £, V, W, and so on,
have typical fibres E, V, W, etc.
Consider the diffeomorphism Oj o O;, 1 of (Ui fl Uj) x E which we.obtain
from the definition of a fibre bundle. It is a map from Ui fl Uj to the group
of diffeomorphisms Diff(E) of the fibre E.

Definition 1.2. A fibre bundle it : £ --r M is a vector bundle if its typical

fibre is a vector space E, and if the diffeomorphisms Oi may be chosen in such
a way that the diffeomorphisms cj o 0i 1 : {x} x E -*,{x} x E are invertible
linear maps of E for all x E Ui fl Uj.

A section s of a vector bundle S over M is a map s : M -> £ such that

7rs(x) = x for all x E M. In this book, we will use the word smooth to
mean infinitely differentiable. The space of all smooth sections is denoted by
r(M, £), while the space of all compactly supported smooth sections (those
equal to 0 outside a compact subset of M) is denoted by r ,(m, £).
If 0 : M1 -> M2 is a smooth map of manifolds and £ is a vector bundle on
M2, we denote by 0*£ the vector bundle over M1 obtained by pulling back £:
1.1. Fibre Bundles and Connections 15

it is the smooth vector bundle given by restricting the vector bundle M1, x £
over Ml x M2 to the graph P(hi) = {(x1, o(x1)) E M1 x M2} of 0, so that
O*£={(x,v)EM1x£I vE£0l0.
There is induced a pull-back map 0* : I'(M2i £) -* r(M1, r*£).
Definition 1.3. A principal bundle P with structure group G is a fibre
bundle P with a right action of a Lie group G on the fibres, that is, (p g) h
p .(gh) for p E P, g, h E G such that
7r(pg) =,7r(p) forallpEPandgEG,
and such that the action of G is free and transitive on the fibres. The fibres
of the bundle P are diffeomorphic to G itself, and so the base M may be
identified with the quotient manifold PIG.
A vertical vector on a fibre bundle £ with base M is a tangent vector on
£ which is tangential to the fibres: that is, X (lr* f) = 0 for any f E C' (M),
where lr* f = f o 7r. We denote the bundle of vertical tangent vectors by VS;
it is a subbundle of the tangent bundle T£. The space of vertical tangent
vectors to a point p in a principal bundle can be canonically identified with
g, the Lie algebra of G, by using the derivative of the right action of G on P.
If X E g, we denote by Xp the vector field such that if f E COO(P),

(XPf)(p) = f(pexpeX).
de e=0
Thus, we obtain a map from P x g to TP, whose image at p is just the vertical
subspace VpP. The vector field Xp will often be simply denoted by X.
If P is a principal bundle with structure group G and E is a left G-space,
with action of G on it given by the homomorphism
p : G --+ Diff (E),

we can form a fibre bundle over M with E as fibre, called the associated
bundle, by forming the fibred product P xG E, defined by taking the product
P x E and dividing it by the equivalence relation

(p,P(g)f),
for all p E P, g E G and f E E. In particular, if we take for E a vector space
which carries a linear representation of G, we obtain a vector bundle on M.
Every vector bundle with N-dimensional fibres on M is an associated bun-
dle for a principal bundle on M with structure group GL(N), called the bundle
of frames. In this book, we will use vector spaces over both IR and C, and
when it is clear from the context, we will not explicitly mention which field
we take homomorphisms or tensor products over. Thus, when we write here
GL(N), we mean either GL(N, ]R) or GL(N, C), depending on the context.
16 1. Background on Differential Geometry

Proposition 1.4. If £ is a vector bundle over M with fibre RN, let GL(E)
be the bundle whose fibre over the point x E M is the space of all invertible
linear maps from RN to the fibre E. Then GL(E) is a principal bundle for
the group GL(N, R), under the action

(p . g) (v) = p(g . v),

where p : RN --> £x, g E GL (N, R) and v E RN, and £ is naturally isomorphic
to GL(E) XGL(N,R) RN. The same result holds when IR is replaced by C. We
call GL(E) the frame bundle of E.
Proof. Since locally, £ has the form U x RN, it follows that locally, GL(S)
has the form U x GL(N, R), proving that P is a principal bundle.
The isomorphism from GL (6) XGL(N,IR) RN to £ is given by mapping the
equivalence class of (p, v) E GL (S),; X RN to p(v) E £_ ; it is easily seen that
this map is equivariant under the action of GL(N, R), that is, (p g, v) maps
to the same point as (p, g v).
The formation of an associated bundle is functorial with respect to mor-
phisms between G-spaces; in other words, we could consider it as a map from
the category of G-spaces to the category of fibre bundles on the base. Indeed,
if n : El -f E2 is a G-map from one G-space to another, then it induces a
map from P x G El to P X G E2. This map is constructed by the commutative
diagram
Ixrc,
PxEl PxE2

PxGEI ) PXGE2
Definition 1.5. Let ?r: £ --> M be a fibre bundle and let G be a Lie group.
We say that £ is an G-equivariant bundle if G acts smoothly on the left
of both S and M in a compatible fashion, that is,
y 7r = 7r y for all y E G.
If £ is a vector bundle, we require in addition that the action ye : E. -> £y . .
is linear.

The group G acts on the space of sections r(M, £) of an G-equivariant

bundle by the formula
(1.1)

Let g be the Lie algebra of G. We denote by L' (X), X E g, the corresponding

infinitesimal action on r(M, £),

(1.2) G6(X)s = T a=o exp(eX) s,

1.1. Fibre Bundles and Connections 17

which is called the Lie derivative. Clearly, GE (X) is a first-order differential

operator on F(M, 5).
Let TM be the tangent bundle of M. A section X E F(M,TM) is called
a vector field on M. If 0 : M1 -* M2 is a smooth map, it induces a map
TM1 -> TM2, in such a way that
0. (v) ETo(,)M2 ifvETTM1.
In particular, diffeomorphisms 0 of M induce a diffeomorphism 0* of TM.
Informally, TM is a Diff(M)-equivariant vector bundle.
Let GL(M) = GL(TM) be the frame bundle obtained by applying the con-
struction of Proposition 1.4 to the tangent bundle TM; it has structure group
GL(n) with n = dim(M). From this principal bundle, we can construct a vec-
tor bundle on M corresponding to any representation E of GL(n); these vector
bundles are called tensor bundles. We see from this that any tensor bun-
dle carries a natural action of Diff(M), making it into a Diff(M)-equivariant
vector bundle over M.
If V is a tensor bundle, the Lie derivative £'(X) by a vector field X on M
is the first-order differential operator defined on F(M, V) by the formula

.CV(X)s Ot'S,
dt t=o
where Ot is the family of diffeomorphisms generated by X. For example, if
V = TM, then we obtain the Lie bracket .C(X)Y = [X,Y].
One example of a tensor bundle is obtained by taking the exterior algebra
A(II8')*; this leads to the bundle of exterior differentials AT*M. The space of
sections r(M, AT*M) of the bundle of exterior differentials is called the space
of differential forms A(M); it is an algebra graded by
A'' (M) = F(M, AiT*M).
Similarly, we denote by A°(M) the algebra F°(M, AT*M) of compactly sup-
ported differential forms. If a is a differential form on M, we will denote
by a[i] its component in Ai(M). If a E Ai(M), we will write jal for the
degree i of a; such an a is called homogeneous. The space A(M) is a
super-commutative algebra (see Section 3 for the definition of this term), that
is,
aA = (-1)IaH-1010Aa.
We denote by d the exterior differential; it is the unique operator on
A(M) such that
(1) d : A' (M) -> A*+1(M) satisfies d2 = 0;
(2) if f E C°° (M), then df E A' (M) is the one-form such that (df) (X) _
X (f) for a vector field X;
(3) (Leibniz's rule) d is a derivation, that is, if a and Q are homogeneous
differential forms on M, then
d(a A 0) = (da) A 0 + (-1)tala A (d,3).
18 1. Background on Differential Geometry

If a E Ak(M) and X0,.. -, Xk E r(M, TM), then

(1.3) (da)(Xo, ... , Xk) _ >(-1)iXi(a(Xo, ... , Xi, ... , Xk))

i=0
+ E (-1)i+ja(1Xi'Xj]7 Xe,..., Xi, ... , Xj,... , Xk).
O<i<j<k

Definition 1.6. If V is a vector space, the contraction operator b(v)

AV* --+ AV*, (v E V), is the unique operator such that
(1) L(v)a = a(v) if a E V*;
(2) t(v)(aAQ) = (t(v)a) A/3+ (-1)I oI aA(t(v)/3) if a and ,6 are homogeneous
elements of AV*.
The exterior operator e(a) : AV* ---> AV*, is the operation of left exterior
multiplication by a E V*.
We will denote the operation of contraction with a vector field X by
t(X) : X (M) --+ A" (M).
Similarly, on A(M, £), t(X) is defined in such a way that for elements of the
form as, a e A(M) and S E I'(M, £),
t(X)(as) = (t(X)a)s.
On differential forms, the Lie derivative satisfies the two properties,

,C(X)d = dC(X),
,C(X)(t(Y)a) = L([X,Y])a+t(Y)(G(X)a),
from which follows easily E. Cartan's homotopy formula:

(1.4) £(X) = d t(X) + t(X) d.

Note that if a is an n-form, where n is the dimension of M, then this formula
implies that C(fX)a =1(X)(fa) for all f E Coo(M).
If 0 : Ml -> M2 is a smooth map of manifolds, we denote by q5*
A' (M2) -+ A' (Ml) the pull-back of differential forms; it commutes with
exterior product and with d, and hence defines a homomorphism of differen-
tial graded algebras.
Since d2 = 0, we can form the cohomology of the complex (A(M), d), which
is
Hi(A(M), d) = ker(dPAi(M))/ im(dIAi-1iMi).
We denote Hi(A(M),d) by Hi(M), and call H'(M) the deRham coho-
mology groups of M. If da = 0, we say that a is closed, and we write its
class in H'(M) as [a]. If a = dQ for some Q, we say that a is exact; exact
1.1. Fibre Bundles and Connections 19

forms represent zero in H (M). The graded vector space Eao H2(A(M), d)
is a graded algebra, with product defined by
[al] A [a2] = [a1 A a2] for closed forms a1 and a2.
This definition depends on the fact that if a1 and a2 are closed and one of
them is exact, then so is a1 A a2; for example, if a1 = do, then
a1na2=d(/8Aa2).
If M is a non-compact manifold, for example, a vector bundle, the defi-
nition of the de Rham cohomology may be repeated with A(M) replaced by
the differential forms of compact support A°(M). The cohomology of the
complex (A,* (M), d) is denoted HH (M) and called the de Rham cohomology
with compact supports.
Let P be a pricipal bundle with structure group G. There is a representa-
tion of the sections of an associated vector bundle P xG E as functions on the
corresponding principal bundle that is extremely useful in doing calculations
in a coordinate free way.
Proposition 1.7. Let C°° (P, E)G denote the space of equivariant maps from
P to E, that is, those maps s : P -> E that satisfy p(g-1)s(p).
There is a natural isomorphism between F(M, PxGE) and C°°(P, E)G, given
by sending s E C' (P, E)G to sM defined by

sM(x) = {p, s(p)];

here p is any element of 7r-1(x) and [p, s(p)] is the element of P xG E cor-
responding to (p, s(p)) E P X E.
Proof. We first note that the definition of sM is unambiguous; if, for some
point x E M, we had chosen some lifting p g instead of p E 7r-1(x), then
[p, s(p)] would have been replaced by

[p, s(p)1-

To complete the proof, we only have to show that any section of P X G E is

of the form sM for some s E Coo (P, E)G. To do this, we merely define s(p)
to equal the unique v e E such that (p, v) is a representative of sM(x).
Observe the following infinitesimal version of equivariance: a function s in
C°° (P, E) G satisfies the formula
(1.5) forXEg,
where we also denote by p the differential of the representation p.
If E is a vector bundle on M, let A(M, £) denote the space of differential
forms on M with values in £, in other words,
Ak (M, £) = r(M, AkT*M (a £).
20 1. Background on Differential Geometry

The differential forms will be taken over the field of real numbers unless either
they take values in a complex bundle or unless explicitly stated otherwise. If £
is the trivial bundle E = M x E, we may write A' (M, E) instead of Ak (M, £).
When P x G E is an associated bundle on M, we will describe A(M, P X G E)
as a subspace of the space of differential forms on P with values in E.
The group G acts on A(P, E) = A(P) ® E by the formula

g-(O(ge) =g./30g-e.
We denote by A(P, E) G C A(P) ® E the space of invariant forms. Elements
of A(P, E)G satisfy the following infinitesimal form of invariance:

.C(X)a + p(X)a = 0,

for all XEgand aEA(P,E)G.

Definition 1.8. A horizontal differential form on a fibre bundle E is a
differential form a such that t(X)a = 0 for all vertical vector fields X.
A basic differential form on a principal bundle P with structure group
C, taking values in the representation (E, p) of G, is an invariant and hori-
zontal differential form, that is, a form a E A(P, E) which satisfies
(1) g-a=a,gEG;
(2) t(X)a = 0 for any vertical vector field X on P.
The space of all basic forms with values in E is denoted A(P, E)bas

We will not prove the following proposition, since its proof is similar to
that of Proposition 1.7.

Proposition 1.9. If a E A'(P, E)bas, define am E AQ(M, P xG E) by

aM(7r*X1i... ,7r*X,)(x) = Ip,a(X1,... ,Xq)(p)],

where p E P is any point such that 7r(p) = x, and Xi E TTP. Then am is

well-defined, and the map a am from A(P, E)bas to A(M, P xG E) is an
isomorphism.

Let £ be a fibre bundle with base M. The quotient of the tangent bundle
T£ by its subbundle V. is isomorphic to the pull-back 7r*TM of the tangent
bundle of the base to £; that is, we have a short exact sequence of vector
bundles
0->V£-4TS ->7r*TM--+ 0.
There is no canonical choice of splitting of this short exact sequence, but
it is important to have such a splitting in many situations. Such a splitting is
called a connection.
1.1. Fibre Bundles and Connections 21

Definition 1.10. Let 7r : E -> M be a fibre bundle. A connection one-

form is a one-form w E A'(E,VE) such that L(X)w = X for any vertical
vector field on S. The kernel of the one-form w is called the horizontal
bundle, and written HE; it is isomorphic to -7r*TM.
If X is a vector field on the base M, we denote by XE the section of HE
such that 7r.Xe = X; we call XE the horizontal lift of X with respect to
the connection determined by w.

Note that we can equally well specify a connection by the horizontal sub-
bundle HE or by the connection form w. In many situations, however, the
connection form is more convenient.
An important object associated to a connection on a fibre bundle £ is its
curvature S2, which is a section of the bundle A27r*T*M ® V£: if X and
Y are vector fields on M and if Xe and Ye are their horizontal lifts, then
52(X, Y) is the vertical vector field on £ defined by the formula

(1.6) 11(X,Y) _ [X,Y]e - [XE,Ye].

Note that this tensor is local on M, in the sense that for any f E C°°(M)

l(fX,Y) _ (7r*f)52(X, Y);

this follows immediately from the fact that Xe (r* f) = -7r* (X f ).
If P is a principal bundle, we only consider connections on it which are
stable under the action of the structure group G by multiplication on the right.
This condition is most conveniently characterized in terms of the connection
form. Observe that, by the identification of VP with the trivial bundle P x
a connection form on P will be an element of A' (P, g).

Definition 1.11. A connection one-form on a principal bundle P is an

invariant g-valued one-form w E A'(P, 2)G such that t (Xp)w = X if X E g.

A differential form a E A(P, g)G satisfies the formula

(1.7) £(X)a + [X, a] = 0 for all X E g.

Proposition 1.12. The space of connection one forms on a principal bundle
P is an afine space modelled on the space of one forms A'(M,P xG g).

Proof. Once a connection form wo is chosen, all others are described uniquely
in the form wo + a for some a E A' (P, g)ba, . The result now follows from
Proposition 1.9.

The curvature SZ of the connection associated to w is the element of

A2(P, g)b,,s defined by

Q(X, Y) P = PHP[X,YI - [PHPX, PHPY],

22 1. Background on Differential Geometry

where PHP is the projection onto the horizontal sub-bundle of TP, and X
and Y are two vector fields on P. In stating the following proposition, we use
the Lie bracket on A(P, g) defined by
[a®A1,a®A2] = (a A/.3) ® [A,,A2],
for a and 0 in A(P), and Al and A2 in g. This makes A(P, g) into a Lie
superalgebra, as we will see in the Section 1.3. If a is a g-valued one-form and
Xl and X2 are two tangent vectors, then [a, a](Xi, X2) = 2[a(Xl), a(X2)].
Proposition 1.13. The curvature form SZ satisfies the formula
S2 = dw + [w, w].
2
Proof. If X E g, then
t(X)(dw + 2[w, w]) = G(X)w - d(t(X)w) + [X, w] = 0
by (1.7), since d(L(X)w) = dX = 0. Thus, the differential form dw + [w, w]
is seen to be horizontal; as it is clearly invariant, it is seen to be basic. To
complete the proof, we need only evaluate the two-form dw + 2 [w, w] on the
horizontal vectors Y and Z; it is easy to see that we obtain -t([Y, Z])w, which
is exactly equal to 1(Y, Z).
The choice of a connection on a principal bundle P determines connections
on all associated bundles £ = P XG E, by defining HE to be the image of
HP under the projection P x E -> E. We will be especially interested in this
construction for associated vector bundles, since it provides an analogue of
the exterior differential on spaces of twisted differential forms A(M, £), called
a covariant derivative.
Definition 1.14. If £ is a vector bundle over a manifold M, a covariant
derivative on £ is a differential operator
v : r(M, £) + r(M, T*M (9£)
which satisfies Leibniz's rule; that is, if s E r(M, 6) and f E Cc (M), then
V(fs)=df®s+fVs.
Note that a covariant derivative extends in a unique way to a map
V : A' (M, £) , A'+1(M £)
that satisfies Leibniz's rule: if a E Ak(M) and 9 E A(M, £), then
V(aA9) =daA9+ (-1)kaA V9.
This formula is the starting point for the definition of a superconnection,
which generalizes the notion of a covariant derivative and forms the topic of
Section 3.
If X is a vector field on M, we will denote by VX the differential operator
t(X)V, which is called the covariant derivative by the vector field X. It
1.1. Fibre Bundles and Connections 23

depends linearly on X, V fX = f Ox, while the commutator [Ox, f] of Vx

and the operator of multiplication by a smooth function f is multiplication by
X- f.
We see from the definition of a connection that if a E Ak(M,£) and
X0, ... , Xk E F(M, TM), then
k
(1.8)
i=0
+ (_1)i+ja([Xi,Xj]>Xo, ,Xi, ..,Xj,...,Xk)
0<i<j<k
Two covariant derivatives Vo and V1 on a vector bundle £ differ by a
one-form w E A1(M,End(£)):
forsEr(M,£).
If £ = M x E is a trivial vector bundle, then the exterior differential d is
a covariant derivative on £, so any covariant derivative on M x E may be
written
(1.9) Os = ds + w s,
for some w E A1(M,End(£)) = A'(M) ® End(E). We say that w is the
connection one-form of V in the given trivialization of 6.
The curvature of a covariant derivative V is the End(£)-valued two-form
on M defined by
F(X,Y) = [Vx,Vy] -V[X,Y],
for two vector fields X and Y. It is easy to check, using Leibniz's rule, that
F(X, Y) really is a local operator, in other words, that it commutes with
multiplication by a smooth function f :

[F(X,Y),f] = [Vx,Yf] + [Xf,VY] - [X,Y]f = 0.

Proposition 1.15. The operator V2 : A' (M, £) --> A'+2 (M, £) is given by
the action of F E A2(M,End(£)) on A(M,E).
Proof. If £ is a trivial bundle and V = d + w, then F is the End(E)-valued
two-form on M given by the formula
F=dw+wAw.
On the other hand, the square of the operator d + w is also given by the action
ofdw+wAw.
We will study the curvature thoroughly later, in the more general context
of superconnections.
24 1. Background on Differential Geometry

If £ is an associated bundle E = P xG E, then a covariant derivative V

may be formed from a connection one-form w on P, by means of the diagram
d+p(-)
coo (P E) G ) A1(P, E)bas

I.
I'(M, E) -° -} A' (M, E)
To show this, we must check that if s is an equivariant map from P to E,
then ds + p(w)s is basic. But choosing X E g, we see that
t(X) (ds + p(w)s) = X s + p(X)s = 0
since s E C°°(P, E)G. Similarly, the form ds + p(w)s c Al(P, E)G, since
s E CO°(P,E)G. It is easy to verify that V is indeed a covariant derivative.
There is another formula for the covariant derivative when written on the
principal bundle in terms of the projection onto the horizontal subspaces of
TP:
Vs=PHMods forsEC°°(P,E)G.
Indeed, if XH is a horizontal tangent vector, then
t(XH)(d + p(w))s = c(XH)ds
since XH is horizontal, while if Xv is a vertical tangent vector,
t(Xv)(d + p(w))s = 0,
since s E C°° (P, E)G. Thus, if sP E CO° (P, E)G corresponds to the section
s E r(M,6), and X E T(M,TM) is a vector field with horizontal lift XP,
then for any p E P above x,
(Vxs) (x) = In (XP sp) (p)]
Proposition 1.16. There is a one-to-one correspondence between connec-
tions on the frame bundle GL(E) and covariant derivatives on the vector
bundle S.
Proof. If V is a covariant derivative on E, we may construct a one-form
w E Al(P,End(E))G by
w(X)s(p) = (V.,. xs)(p) - (X s)(p) for s E C°°(P,E)G, X E TPP.
This defines w uniquely, since for any (p, v) E P x E, there exists an s E
C' (P, E)G such that s(p) = v. The one-form w satisfies
t(X)w=p(X) for allXEg,
and hence is a connection form.
As an example of this result, let £ = M x E be a trivial vector bundle
on M with connection V = d + w. The frame bundle GL(E) of 6 is equal
1.1. Fibre Bundles and Connections 25

to 7r : M x GL(E) --+ M. Let g-ldg be the Maurer-Cartan one-form in

A1(GL(E), End (E)) defined using the tautological map
g E C' (GL (E), End (E)),
which sends an element of the group GL(E) to the corresponding linear map
in End(E). Then the connection one-form on GL(9) associated to V is
g-1. i*w . + g_1dg.
WGL(e) - 9
In this way, we see that the two different notions of connections, as a
connection on a principal bundle and as a covariant derivative on sections of
a vector bundle, are two faces of a single object.
If the covariant derivative has been defined in terms of a connection w on
a principal bundle P, then we obtain a representation of its curvature as a
basic two-form ci on P with values in g: F = p(Sl), with 92 = dw + [w, w],
since for s E C' (P, E) G,
F(X,Y)s = ([XP,YP] - [X,Y]P)S = -Q(X,Y)s
= p(ci(X,Y))S.
The advantage of constructing a covariant derivative from a connection
on a principal bundle is that this provides a compatible family of covariant
derivatives on all vector bundles associated to P simultaneously, in a way
which is compatible with the natural linear maps between these bundles, and
with the tensor product of bundles. Thus, if we are given a G-equivariant
map is : V --k W and S E r(M, P xG V), then we may form the bundle maps
K:PXGV->PxGWand 1®ic:T*M®PXGV--4T*M®PxGW,and
then
V,c(s) = (1(& r.)(Vs) E A'(M, P xG W).
We now give some ways of contructing new covariant derivatives out of old.
If V is a covariant derivative on £, we obtain a covariant derivative on the
dual bundle £*, by the following formula, for s E r(M, £) and t E r(M, £*):
d(s, t) = (Vs, t) + (s, Vt),
that is, if X E r(M,TM) is a vector field on M,
X (s, t) = (Vxs, t) + (s, Vxt).
Here, we use the notation (s, t) to denote the pairing between £ and S.
If V is a covariant derivative on £i for i = 1, 2, we obtain a covariant
derivative V-01 0-02 on the tensor product by
Ve e2 (S1 (9S2) = Vel sl ®S2 + S1 ®Ve2S2
for si E r(M, Si). We refer to pal®e2 as the tensor product covariant deriva-
tive, and write
Vet19E2 = Vel ®1 + 1®Ve2.
26 1. Background on Differential Geometry

Let M be a manifold on which a compact Lie group G acts smoothly, and

let E be an equivariant bundle over M. We say that a covariant derivative
vE is invariant if it commutes with the action of G,
gT*M®E . v = v. gE

for all g E G. If G is compact, there will exist such connections, since we may
average with respect to the Haar measure on G,
f(gT*M®E)_1 V.gEdg
(1.10)

is an invariant covariant derivative.

Let 0: N -> M be a smooth map, and let £ be a vector bundle on M. If
V is a covariant derivative on £, then the formula
v1*E(fO*s) = df ®Qi*s + f cb* (VEs)
vO*E
for f E Coo (N) and s E F(M, £), defines a covariant derivative on the
bundle O*£, which we call the pull-back of the covariant derivative VE.
In particular, if -y(t) : R ---+ M is a smooth curve in M with tangent vector
(t) E T7(t)M, and if s : Il 8 -+ y*£ is smooth, then V itls(t) is defined to
be the covariant derivative vy*E(s)(t). If £ is the trivial bundle M x E and
vE = d + w, we may write s(t) = (y(t), f (t)) with f : R -* E, and

°y(t)s(t) = ((t) ' dfdt(t)

The parallel transport map along y(t),
Ty(t) E Hom(£,y(o),£y(t))

is defined by solving the ordinary differential equation

(1.11) v,Y(t)T--Y(t) = 0,

with initial condition ry(0) = I.

Suppose U is an open ball in 118' and £ is a vector bundle over U with
connection V. If xi are the coordinates on U and let ai are the corresponding
partial derivatives, let
R Exiai
i
be the radial vector field on U. The following result follows from the theory
of ordinary differential equations.

Proposition 1.17. Parallel transport along rays t - tv, v E U, gives a

smooth trivialization of 6 over U, by identifying the fibres £ with E = £o.
1.1. Fibre Bundles and Connections 27

In terms of this trivialization, the connection one-form on £ is the End(E)-

valued one-form w = >i widxi such that V = d + w, and the curvature of S is
the two-form F with values in End(E) defined by the equation F = dw+wnw.
By the definition of this frame of £, we have t(R)w = 0. Thus the following
formula for the radial derivative of the connection form w holds:
(1.12) £(R)w = [t(R), d]w = t(R)(dw +wnw) = t(R)F.
A useful property of this framing of S is that all of the coefficients of the
Taylor expansion of wi at 0 are determined by the coefficients of the Taylor
expansion of F at 0.
Proposition 1.18. The Taylor expansion of wi at xo = 0 has the following
form:
Xa
wi(x) - - F(ai, 8j)x0x + E 8awi(xo) c!
j Ic I?2

Proof. Expanding the Taylor's series of both sides of (1.12), we obtain

E(Iaj + 1)8awe(xo)xa/a! = E 8aF(ak, ae)x.Xkxa/a!

a a,k

By equating coefficients of xa on both sides, we see from this formula that

ajwi(xo) = -ZF(ai,oj)xo.
Furthermore, it follows that the Taylor coefficients of wi at xo to order m only
depend on those of F(8i, (9j) to order m - 1. 0
A metric on a vector bundle is a smooth family of non-degenerate inner
products on the fibres of E. If £ is a real bundle of rank N, then this amounts
to requiring that £ is associated to a O(N)-bundle P with fibre the standard
representation RN. In fact, the principal bundle may be taken equal to O(S),
the sub-bundle of GL(S) consisting of orthonormal frames. (There are cor-
responding statements with R replaced by C and O(N) replaced by U(N).)
A real vector bundle with metric will be called a Euclidean vector bundle,
while a complex vector bundle with metric will be called a Hermitian vector
bundle.
We say that a covariant derivative V preserves a metric if it satisfies the
formula
d(sl, s2) = (Vs1i 82) + (s1i Vs2) for si E r(M, £).
If V preserves a metric on £, it is easy to see that its curvature Fx(X, Y) lies
in the Lie algebra so(£x) of skew-adjoint endomorphisms of £x.
The covariant derivative V preserves a metric on £ if and only if the parallel
transport map also preserves this metric, since if s E\£,y(o), we have
d
dt
ITy(t)SI2 - 2 (1 T.y (t) S, V . (t)T-Y(t)SJ = 0.
28 1. Background on Differential Geometry

This shows that the trivialization of Proposition 1.17 is compatible with the
metrics on the fibres of E. Thus, if e,, is an orthonormal basis of E = 60, we
obtain smooth sections, which we also denote by ea, which are orthonormal
at each point of U; these sections make up an orthonormal frame.
Let 7r : E - * M be a vector bundle on M. A covariant derivative Ve on £
determines a connection on GL(£), and hence a connection on the total space
£, that is, a splitting
TS =V£ED H£.
The bundle VE is isomorphic to 7r*£. We denote by 0 E A' (£, V£) the
connection one-form of this connection.
Definition 1.19. If E is a vector bundle, its tautological section x E
r(£, n*£) is the smooth section which to a point v E 6 assigns the point
(v,v) E 7r *S.

Proposition 1.20. Let E be a vector bundle with covariant derivative Ve.

V'*6

If denotes the pull-back of the covariant derivative on £ to -7r*£, then

0, thought of as an element of A.1(£,7r*£), equals V"*sx.
Proof. Locally, £ may be trivialized; thus, we assume that 6 = M x E, with
covariant derivative V = d + w. Denote a point of £ by (x, v) E M x E; then
x(x, v) = v, and V"µ E x = dv + w v. Thus
v(XV)x = V +w(X)v.
The frame bundle GL(E) has tangent vectors (X, A) E T(.,,9)GL(£) =
TTM x g((E) at the point (x, g) E GL(E) = M x GL(E). The horizontal
tangent space at (x, g) is
H(.,9) GL(E) _ {(X, -w(X)) I X E TXM}.
Under the map ((x, g), v) H (x, gv) from GL(S) x E to E, the image of
H(;,,e) GL(E) is
= {(X, -w(X)v) I X E TZM}.
Since V *6x vanishes on this space and satisfies
V(o,V)x = V,

we see that it equals 9.

We will now study connections on the tangent bundle TM: these are often
called affine connections.
Definition 1.21. The fundamental one-form B on M is the element of
A'(M,7'M) such that e(X)9 = X for X E TM. If V is a connection on
TM, than the torsion of V is the differential form T = VB E A2(M,TM).
If 7':-- 0, we say that the connection V is torsion-free.
1.1. Fibre Bundles and Connections 29

If X and Y are two vector fields on M, then T (X, Y) E r(M, TM) is given
by the explicit formula
T(X,Y) _ VxY -VyX - [X,Y].
A connection on TM gives rise to a covariant derivative on any vector
bundle which is associated to GL(M), in particular the exterior bundle AT*M;
by naturality, this connection satisfies
Vx(aA0) =Vxan/.3+aAVx,@
for any homogeneous differential forms a and 0, and vector field X. If a is a
k-form and X, Y1i ... , Yk are vector fields, we see that
k
X-a(Yl,... Yk) = (VXa)(Yl,... ,Yk)+a(Y1,... ,VxYi,- ,Yk)-
i=1
In statement of the following proposition, we use the exterior product map
E : T*M ®A*T*M -* A*+1T*M.
Proposition 1.22. If V is a torsion free connection on TM, the exterior
differential is equal to the composition
P(M, AT*M) -°> P(M,T*M (9 AT*M) E, I'(M, A*+1T*M).
Proof. Let us denote the operator defined above by D. Using the fact that V
satisfies Leibniz's rule, we see that D does too:
D(a/) = Da A Q + (-1)I0'Ia n D,(3.
(Note that this part of the proof is independent of whether V is torsion-free
or not.)
Given Leibniz's rule, it clearly suffices to show that D agrees with d on
functions (which is clear), and one-forms. The proof now rests upon the
following formula: for any f E C°O(M),
D2f = D(df) = -(T,df) E A2(M),
where T E A2 (M, TM) is the torsion of the affine connection V. In partic-
ular, if T vanishes, then D agrees with d on one-forms. To prove this, we
choose a local coordinate system on M, and let dxi and Xi = a/8xi be the
corresponding frames of the cotangent and tangent bundles. If we denote the
covariant derivative in the direction Xi by Vi,. D is given by the formula
D = E e(dxi)Di.
i
Thus, D(df) equals
EE(dx')Di(8j f dxj) = >e(dx')8j fVidxj.
ij ij
30 1. Background on Differential Geometry

But by Leibniz's rule,

0 = Di(dxj,Xk) = (Didxi,Xk) + (dx2,V Xk)
so that
D(df) E e(dxidxk)aj f (dx', viXk)
ijk
= ->Is(dxidxk)(df,T(Xi,Xk))
i<k
Let M be a manifold of dimension n. Another example of a vector bundle
associated to the principal bundle GL(M) is the bundle, denoted IAMJS, of
s-densities on M, where s is any real number. (If the manifold M is clear
from the context, we will write simply CAI'.) This is the bundle associated to
the character A H det(A)I-s of GL(n,R). An element of the fibre JAI , is a
function ATTxM - {0} --> lib such that V(AX) = JAI (X) for X# 0. If U
is an open subset of R, we denote by IdxIs the s-density such that
Idxls(alA...an)=1,
where ai is the vector field a/axi corresponding to differentiation in the ith
coordinate direction. Using a partition of unity, we see that we can always
construct a nowhere-vanishing section of IAl8, so that JAI' is a trivializable
bundle.
The importance of the one-density bundle, or density bundle as it is
usually known, comes from the following result, which follows immediately
from the change of variables formula for integrals in several variables.
Proposition 1.23. Let M be a manifold of dimension n. There is a unique
linear form, the integral, denoted by

: r,(M, IAA) ---> R

JM
which is invariant under diffeomorphisms, and agrees in local coordinates
with the Lebesgue integral:

fM f (x) IdxI = fLQn

.f (x) dxl ... dx,,,.

For any vector field X on M and a any compactly supported C1 density,

we see that
(1.13) G(X)a = 0.
IM
Since G(fX)a = f1(X)a+X(f)a if a E r(M, JAI), it is easy to see that
if 3 E r(M, IA11/2) is a half-density on M,
(1.14) C(fX)i3 = fG(X)/3+ 2X(f)3.
1.1. Fibre Bundles and Connections 31

The bundle of densities JAI is very closely related to the bundle of volume
forms AnT*M; the first corresponds to the character I det(A)I-1 of GL(n)
and the second to the character det(A)-1. Unlike IAA, the line-bundle AnT*M
is trivializable if and only if M is orientable. If M is a connected orientable
manifold, its frame bundle GL(M) has two components; an orientation of
M is the choice of one component, which is a principal bundle for the group
GL+(n) of matrices with positive determinant. We say that frames in the
chosen component GL+(M) are oriented. An orientation of a manifold M is
equivalent to the choice of an isomorphism between the bundles A"T*M and
JAM!. If v E An(M) is a volume form on M, the corresponding density vJ is
such that Ivk(X) = v(X) if X E AnTXM is oriented.
If M is an n-dimensional oriented manifold, we define the integral of a
compactly supported differential form a E A,(M) to be the integral of the
density corresponding to a[n]:

f a=
M M
la[njl.

There is a fibred version of this integral. Let -7r : M --+ B be a fibre bundle
with n-dimensional fibre, such that both M and B are oriented. If a E Ak (M)
is a compactly-supported differential form on M, its integral over the fibres
of M --> B is the differential form fM/B a E Ak-n (B) such that

(1.15) fM a A 7r*Q,
f(fM/ B a/ n
for all differential forms 0 on the base B. We sometimes write -7r*a instead
of f M/B a. It follows easily from (1.15) that

(1.16) 7r* (a A 7r*,0) = 7r*a A,3

for all aEA,(M) and A(B).

The integral over the fibres commutes with the exterior differential, in the
sense that

(1.17) dB 1
M/B
a = (-1)n JM/B dMa.
This formula shows that the integral over the fibres induces a map

He (M) _He-n(B)
M/B
on de Rham cohomology.
The following result is an example of a transgression formula: it says that
the cohomology class of the integral over the fibres does not change if the map
it is deformed smoothly.
32 1. Background on Differential Geometry

Proposition 1.24. Let 7r : M x J -> B be a smooth map such that 7rt =

7riMx{t} : M -+ B is a fibre bundle for each t. Then there is a smooth
t-dependent map irt : Ak(M) --> A'-n-1(B) such that
d t
d*ta = (-1)' r.(da) - d(Frtka).
Proof. Think of 7r as a fibre bundle map M x R ---> B x R which preserves
the t-coordinate. Let q : M x JR --4 M be the projection which forgets the
t-coordinate. Given a E Ak(M), the differential form it*q*a c Ak-n(B x IR)
may be decomposed as follows:
7r*q*a = 7rtka + irta A dt,
where this formula serves as the definition of *ta. Applying this formula to
da, we see that on the one hand,
7r*q*da = 7r*t da + dt n Frtkda,

while on the other hand,

7r*q*da = 7r*dq*a
= (-1)nd (7rta + ir;,a n dt)
d7rtdt a
= (_1)ndB7rta + (-1)71 ( + dB'rta n dt.
Equating coefficients of dt, the proposition follows.
Corollary 1.25. The map induced by 7rt on de Rham cohomology with com-
pact support is independent of t.

1.2. Riemannian Manifolds

If M is an n-dimensional manifold, then a Riemannian structure on M is
a metric on the tangent bundle TM. Using the Riemannian structure, we can
construct the orthonormal frame bundle, which is denoted by O(M):
O(M) = { (x, (e1, ... , en)) ei form an orthonormal frame of TXM }
I

An orthonormal frame p E 0(M) over x e M may be considered to be an

isometry u = (u1,... , un) E R H E uiei E TAM from Rn to ,,x. We
will denote by 7r : 0(M) -+ M the projection map. The bundle 0(M) is an
0(n)-principal bundle, and the tangent bundle of M is an associated bundle
of O(M), corresponding to the standard representation Rn of O(n):
TM'- 0(M) xp(n) W
If, in addition, M is oriented, then the bundle of oriented orthonormal frames
is a SO(n)-principal bundle which we denote by SO(M).
1.2. Riemannian Manifolds 33

The tangent bundle TM of a Riemannian manifold M has a canonical

covariant derivative V, called the Levi-Civita connection. This is the
unique covariant derivative which preserves the Riemannian metric,
d(X, Y) = (VX, Y) + (X, VY),
for all vector fields X and Y on M, and which is torsion-free,
vxY - VYX = [X,Y].
It is easily verified from the above two formulas that the Levi-Civita derivative
VxY is given by the formula
(1.18) 2(VxY, Z) = ([X, Y], Z) - ([Y, Z], X) + ([Z, X], Y)
+X(Y,Z) +Y(Z,X) - Z(X, Y).
Denote by so (TM) the bundle 0(M) xo(n) so(n), whose fibre at x E M is
the Lie algebra consisting of antisymmetric endomorphisms of the inner prod-
uct space TIM. The curvature of the Levi-Civita connection is an so(TM)-
valued two-form on M, called the Riemannian curvature, which we will
denote by R E A2(M,so(TM)), that is, if X, Y and Z are vector fields on
M,
R(X, Y)Z = VxVYZ - VyvxZ - V [X,Y]Z.
If X and Y are two vector fields on M, we will write R(X, Y) for the section
of so(TM) given by contracting the two-form R with X and Y, and (RX, Y)
for the two-form obtained by contracting the curvature in its so(TM) part.
For future reference, we list some well-known properties of the Riemannian
curvature.
Proposition 1.26. Let X, Y, Z and W be vector fields on M.
(1) R(X,Y) = -R(Y,X)
(2) (R(W, X)Y, Z) + (Y, R(W, X)Z) = 0
(3) R(X, Y)Z + R(Y, Z)X + R(Z, X)Y = 0
(4) (R(W, X)Y, Z) = (R(Y, Z) W, X)
Proof. (1) says that R is a two-form, and (2) that RZ(W, X) E so(TTM).
Let us prove (3), which holds for the curvature of any torsion-free connec-
tion V on TM. If R is the curvature tensor of V,
vxvyZ-vxvzY=vx[Y,Z],
and similarly if we cycle the three vectors X, Y and Z. In this way, we obtain
three equations which, when added together, give
VxVYZ +VYVZX +VzVxY - DYDxZ - VzvyX - vxVZY
= vx[Y, Z] +VY[Z, X] +Vz[X,Y].
Clearly, this is equivalent to (3).
34 1. Background on Differential Geometry

The last property is a purely algebraic consequence of the first three, which
we leave as an exercise.
When we are given a frame Xi for the tangent bundle with dual frame
Xi, we can write the components of the curvature matrix in the form Rj"kl =
(R(Xk, X1)Xj, Xi); here, the i and j indices refer to the fact that R takes
values in End(TM), and the k and l indices are the two-form indices. Using
the metric, define
(1.19) Rijkl = (R(Xk,XI)Xj,Xi)-
The components Rijkl satisfy Rijkl = Rklij If ei is an orthonormal frame of
TM with dual frame ei, the two-form (Rei, ej) is given by the formula
(1.20) (Rei, ej) _ - Rijklek A el.
k<1

From the curvature tensor, we may form a symmetric tensor

Ricij = E Rikjk E r(M, S2(TM)),
k

called the Ricci curvature, and its trace the scalar rM R1,,,.i,,,, called
the scalar curvature of M.
We denote by Jdxl E I'(M, JAI) the Riernannian density associated to
the Riemannian metric on M, which is such that
JdxJ(ei n ... A e,,,) = 1
when ei is an orthonormal frame of TM. The existence of IdxI corresponds to
the fact that the character I det I of GL(n) used to define the line bundle JAI is
trivial on the structure group O(n), so that the density bundle is canonically
isomorphic to M x R. It follows that IdxI is covariant constant under the
Levi-Civita derivative. We will often write dx instead of IdxI, although this
notation should be restricted to oriented manifolds.
If f is a smooth function on a Riemannian manifold M, the gradient of
M, denoted grad f, is the vector field on M dual to df, that is, for every
vector field X on M,

By a smooth path xt : [0,1] --+ M on a Riemannian manifold, we mean the

restriction to [0, 1] of a smooth map xt : (-e, 1 + e) -+ M for some e > 0. If
xt is such a path, we define its length by the formula

L(x) = f0 (xt, xt) dt,

where by xt we mean the tangent vector in TT, to the path xt at time t. Define
the Riemannian distance between two points d(xo, x1) to be the infimum of
L(xt) over all smooth paths between them.
1.2. Riemannian Manifolds 35

A smooth path xt is called a geodesic if it satisfies the ordinary differential

equation
0.

Like any second-order ordinary differential equation, the equation for a geode-
sic has a unique solution for t small, given the starting point xo and the
derivative x = xo E Two M at the starting point. The end point xi of the
resulting curve is called the exponential of the vector x at xo and is written
expxo x. It is defined if x is small enough. (This terminology arises because
on a compact Lie group with the left and right invariant metric, the geodesics
turn out to be the same as the one-parameter subgroups and we recover the
exponential map of the Lie group.)
The derivative d expxo of the exponential map at xo itself is just the identity
map of Two M, so that by the inverse-function theorem, the exponential map is
a diffeomorphism from a small ball around zero in Ty0 M to a neighbourhood
of xo in M. (The radius of the largest ball in T 0M for which this is true
is called the injectivity radius at xO.) Thus the exponential map defines a
system of coordinates around xo, called the normal coordinate system (this
is also known as the canonical coordinate system). The important property of
this coordinate system is that the rays emanating from the origin in TT0 M map
onto geodesics in M. We will use letters of the form x to denote coordinates
in the normal coordinate system: thus, x E TT0M represents the coordinates
of the point expxo x.
We will now collect some formulas for the pull-back of the metric and
covariant derivatives by the exponential map. In particular we will see that
the ray from 0 to x is the curve of shortest length between xo and exp,,o x
when x is small.
If we apply the construction of Proposition 1.17 to the tangent bundle with
its Levi-Civita connection, we obtain a smooth trivialization of TM over a
normal coordinate chart centred at a point xo E M. Choose an orthonor-
mal frame of the tangent space Two M, with respect to which the coordinate
functions are xi, and the corresponding partial derivatives are 8i; thus, the
vectors 8i are orthonormal at xo, although not in general elsewhere. In this
coordinate system, the point xo has coordinates x = 0. The radial vector field
R is the vector field defined by the formula R = E' l xi8i. Let ei be the
orthonormal frame of TM obtained from the smooth trivialization of TM in
this coordinate patch; thus, ei = 8i + O(lxl) and O7zei = 0.

Proposition 1.27. In the normal coordinate system, the following formulas

hold:
(1) V?ZR=R
(2) R = Ei xiei, and thus (R, R) = 1x12
(3) (R, ai) = xi
(4) d(xo, exp.a x) = 1x1
36 1. Background on Differential. Geometry

Proof. (1) In the normal coordinate system, the curve xt = tx is a geodesic.

Since R(xt) = txt, the geodesic equation gives
VRR = tVXt (t±t) = txt = R.
(2) If we differentiate the function (R, ej) with respect to the vector field
R, we obtain
R(R, ej) _ (VRR, ej) + (R,VRei) = (R,ei)
In other words, (R, ej) is homogeneous of order one. Since

(R, ej) xj (& , ej) = x` + O(1x12),

the result follows from these two equations.

(3) The function (R, ai) is equal to E, x5 (ai, a;), so it must equal xi +
O(Ixi2). On the other hand,

R(R, 8i) = (VRR, 8i) + (R,V100 -

Since the Levi-Civita connection is torsion-free and [7Z, ai] = -ai, it follows
that

(R, V8) _ (R, V 0iR) + (R, [R, Oil)

If we assemble all of the terms that we have obtained, we see that

R(R,ai) = 2aiIRI2 =xi,

from which the result follows.
(4) The equation (3) shows that R is orthogonal for the Riemannian metric
to the vector fields xia1 - xiai. If xt, 0 < t < 1, is any curve from 0 to x,
decompose its tangent vector xt into the sum of a multiple of R and a vector
lying in the span of the vector fields xiaj - x3ai tangent to the sphere. It
follows that
dlxt I
Ixtl? dt
thus the length of the curve xt is greater than or equal to 1x1. Since the length
of the curve xt = tx is equal to Ixi, the formula follows. O
A property of the normal coordinate system is that near the origin, the
coefficients of the metric (ai, o9;). agree with the Kronecker delta Sij up to
second order. Although we will not use the full force of the next proposition
in the rest of this book, we give its proof because of its applications to the
theory of the heat kernel.
1.2. Riemannian Manifolds 37

Proposition 1.28. The Taylor expansion of the functions gij (x), defined by
(ai, aj)X in normal coordinates at xo has the following form:

gij (x) " sij - g > Rikjl(x0)xkxl +

kl IaI>3
(aagij)(x0) a
Proof. To prove this result, we will make use of two distinct frames of the
tangent bundle TM in a normal coordinate chart around a point x0, the
first being given by the vectors ai (which is not an orthonormal frame in
general), and the second by the orthonormal frame ei obtained by radial
parallel transport of the orthonormal frame (ai),0 of The matrix
relating these two frames is the matrix 8 such that

ai0 ej.
With respect to the frame ej, the fundamental one-form

0= dxiai E A'(M,TM)
i
equals E Bjej, where Oi = >i 8 dxi. Thus the R'-valued one-form 0 = (0j)
satisfies the structure equation
dO+w AO=O,

where w is the End(R91)-valued one-form of the LGvi-Civita connection in the

frame ei. Observe that
t?,8 = R = (xl, ... x'),
,

since R = xiei, as follows from Proposition 1.27(2).

Let f = dw + w A w be the curvature of M in this coordinate system; as
in Proposition 1.18, we see that C(R)w = t(R)Q. Substituting the formulas
t-.w = 0 and (,C(R) -1)x = 0, into the identity tR.(dO+w A0) = 0, we obtain

(,C(R) - 1)L(R)9 = (1(R)w)x = (t(7Z)1l)x,

where we think of the curvature St as a matrix of two-forms, so that )x is a
vector of two-forms. Using the formula R = E xiei once more gives

(1.21) t(ak)(L(R) - 1)G(R)Oa = Exixj(cl(ai,ak)aj,ea).

ij

Now ej is expressed as a function of the ai by the inverse of the matrix Bil,

and 8z (xo) = bij. Thus this equation determines the Taylor expansion of
Bb to order m in terms of the Taylor expansion of the curvature coefficients
38 1. Background on Differential Geometry

Rijkl = (c(ak, al)8j, ai)x to order m - 2. In particular, the linear term is

zero, and the quadratic piece may be calculated from (1.21)

Bb = bab - I T. XZx3Rjaib + O(1x13)

Since the metric on M is expressed in terms of the one-form 0 by the formula

k
gij(x) = Bik Bj,
k

the Taylor coefficients of gij may be expressed in terms of those of 0.

The differential d, expxo of the exponential map exp.,, : T,0M --+ M at
the point x e Tx0M is a map from Tx0M to TxM, where x = expxo x. Since
Tx0M and TxM are metric spaces, the Jacobian
j(x) = 1det(d.expxo)
is well defined. By definition, if ei is an orthonormal frame of TM and
ai=Ej8?ej,then j(x)=Idet(0i) 1, so that
(1.22) j(x) = ldet(gij(x))11/2.

The frames 8i and ei of the tangent bundle TM that were introduced a

the start of the above proof will be used constantly throughout this book,
and the reader must be very careful to distinguish them. The first frame is
frequently useful because it satisfies the formula [8i, 8j] = 0, while the frame
ei possesses the advantage of being orthonormal at each point in the normal
coordinate chart. In fact, we will only use the following two consequences of
Propositions 1.18 and 1.28 in the rest of this book:

(1.23) w (x) = -2 F(,9i,8j)xoxi +O(1x12);

j
(1.24) gij(x) = 1 + O(1x12).

1.3. Superspaces
Although this is not a book on supergeometry, we will constantly use the
terminology of super-objects. Recall that a superspace E is a 7L2-graded
vector space
E=E+ED E-,
and that a superalgebra is an algebra A whose underlying vector space is
a superspace, and whose product respects the 7L2-grading; in other words,
Ai Aj C Ai+j. (We will denote the two elements of Z2 by + and -, although
1.3. Superspaces 39

we should really denote them by 0 mod 2 and 1 mod 2.) An exterior algebra
is an example of a superalgebra, with the Z2-grading
AtE AEE.
(-1)==f1
An ungraded vector space E is implicitly Z2-graded with E+ = E and E- _
0. The algebra of endomorphisms End(E) of a superspace is a superalgebra,
when graded in the usual way:
End (E) = Hom(E+, E+) ® Hom(E-, E-),
End- (E) = Hom(E+, E-) ® Hom(E-, E+).

Definition 1.29. A superbundle on a manifold M is a bundle £ _ £+®£-,

where £+ and £- are two vector bundles on M. Thus, the fibres of £ are
superspaces.
If £ is a vector bundle, we will identify it with the superbundle such that
£+=£and£-=0.
There is a general principle in dealing with elements of a superspace, that
whenever a formula involves commuting an element a past another b, one
must insert a sign (__1)1a1 . I bI , where lal is the parity of a, which equals 0 or
1 according to whether the degree of a is even or odd. Thus, the supercom-
mutator of a pair of odd parity elements of a superalgebra is actually their
anticommutator, due to the extra minus sign. We will use the same bracket
notation for this supercommutator as is usually used for the commutator:
[a, b] = ab - (-1)IaI' IbIba
This bracket satisfies the axioms of a Lie superalgebra:
(1) [a, b] +(-1)IaI-IbI[b,a] = 0,
(2) [a, [b, c]] = [[a, b], c] + (-1)laI. III [b, [a, c]]
We say that a superalgebra is super-commutative if its superbracket
vanishes identically: an example is the exterior algebra AE. The following
definition gives the extension of the notion of a trace to the superalgebra
setting.
Definition 1.30. A supertrace on a superalgebra A is a linear form 0 on
A satisfying 0([a, b]) = 0.
On the superalgebra End(E), there is a canonical linear form given by the
formula
Str(a) _ TrE+(a) -'ICE-(a) if a is even,
0 if a is odd.
Proposition 1.31. The linear form Str defined above is a supersrace on
End(E).
40 1. Background on Differential Geometry

Proof. We must verify that Str[a, b] = 0. If a and b have opposite parity, this
is clear, since [a, b] is then odd in parity and hence Str[a, b] = 0.
If a = (o a ) and b = (o b ) are both even, then
([a+, b+]
[a, b] _ 0 [a-0b-])
has vanishing supertrace, since
b_]
TrE+ [a+, b+] = 'IrE- [a , = 0.
If a = (a o ) and b = (b+ o) are both odd, then
a-b++b-a+ 0
ab
[ ]= ( 0 a+b- + b+a
has supertrace

Str[a, b] = TrE+ (a- b+ + b-a+) - TrE- (a+b- + b+a-) = 0.

If E _ E+ ® E- and F = F+ ® F- are two superspaces, then their tensor
product E ® F is the superspace with underlying vector space E 0 F and
grading

(E®F)+=E+®F+®E-®F-,
(E®F)-=E+®F-®E-®F+.
If A and B are superalgebras, then their tensor product A ® B is the
superalgebra whose underlying space is the Z2-graded tensor product of A
and B, and whose product is defined by the rule
(ai ®bi) .(a2 (9 b2) = (-1)Iblf a2Iaia2 ®b1b2.
(This product is imposed by the sign rule.) Note that this Z2-graded product
is not the same as the usual product on A 0 B, so is this algebra is some-
times denoted A®B; however, we will never make use of the ungraded tensor
product, so there is no ambiguity.
Definition 1.32. If A is a supercommutative algebra and E is a superspace,
we extend the supertrace on End(E) to a map
Str : A ®End(E) --* A,
by the formula Str(a (9 M) = a Str(M) for a E A and M E End(E).
This extension of the supertrace still vanishes on supercommutators in
A ® End(E), since
(1.25) [a®M,b®N] = (-1)IMI-(blab®[M,N]
when A is supercommutative.
1.3. Superspaces 41

If El and E2 are two superspaces, we may identify End(Ei (9 E2) with

End(E1) ® End(E2), the action being as follows:
(ai (9 a2)(el (9 e2) (-1)Ia21- IelI
(a,el) ® (a2 e2),
where ai E End(Ei) and ei e E.
If E is a superspace, such that dim(E±) = m±, we define its determinant
line to be the one-dimensional vector space
det(E) = (Am+E+)-1 ® Am- E-.
we make use of the commonly-used notation L-1 for L* when L is a
one-dimensional vector space. The vector space det(E) is one-dimensional,
and if E+ = E-, then det(E) is canonically isomorphic to R.
Definition 1.33. If £ _ £+ ® £- is a superbundle, then its determinant
line bundle det(£) is the bundle
(Am_£-),
det(E) = (A-+£+)-1
®

where m± = rk(£t).
For example, if M is an n-dimensional manifold, det(TM) is the volume-
form bundle AnT*M.
Definition 1.34. A Hermitian superspace is a superspace E such that
both E+ and E- are complex vector spaces with Hermitian structures.
If E is a Hermitian superspace, we say that u E End- (E) is odd self-adjoint
if it has the form
u=
O 0),
where u+ : E+ --> E-, and u- is the adjoint of u+.
Let V be a real vector space with basis ei. We will denote tk = 46k) and
6k = e(ek). We will identify AkV* with (AkV)* by setting (e1, ej) = SJ,
where

I={1 and
J = {1 < j1 <... <jk <dim(V)}
are multi-indices, that is, subsets of {1, ... , dim V}, and er = ei' ... eik, ej
ej, ... ejk. (Here, as occasionally in the rest of this book, we write v w instead
of v A w for the product in an exterior algebra.) Using this identification, it is
easy to see that s(a)* = t(a) E End(AV) and that t(v)* = e(v) E End(AV).
If A E End(V), we denote by )\(A) the unique derivation of the superalgebra
AV which coincides with A on V C AV; it is given by the explicit formula

(1.26) \(A) = 1: (ej, Aek)sjtk.

jk
42 1. Background on Differential Geometry

A non-zero linear map T : AV --* 11 which vanishes on AkV for k < dim(V)
is called a Berezin integral. Let us explain why such a linear map is called
an integral. If V is a real vector space, the superalgebra AV* is the algebra
of polynomial functions on the purely odd superspace with E+ = 0, E- = V.
If ei is a basis for V with dual basis ei, the elements ez E AV* play the
role of coordinate function on V. From this point of view, the contraction
t(v) : AV* -+ AV*, (v E V), is the operation of differentiation in the direction
v. We now see that a Berezin integral is an analogue of the Lebesgue integral:
a Berezin integral T : AV* -* R is a linear form on the function space AV*
which vanishes on "partial derivatives":

(1.27) T t(v)a = 0 for all v E V and a E AV*.

If V is an oriented Euclidean vector space, there is a canonical Berezin

integral, defined by projecting a E AV onto the component of the monomial
e1 A...Ae,,,; here n is the dimension of V, and ei form an oriented orthonormal
basis of V. We will denote this Berezin integral by T:

1, 111 n,
(1.28) T(ei) =
0, otherwise.

If a E AV, we will often denote T (a) by a[,,I, although strictly speaking a[n]
is an element of AThV and not of IR. If A E A 'V, its exponential in the algebra
AV will be denoted by expA A.

Definition 1.35. The Pfaffian of an element A E A2V is the number

PfA(A) = T(expAA)
The Pfaffian of an element A E so(V) is the number

Pf(A) = T (expA E(Aei, ej) ei A ej).

i<j
If V = IL2 with orthonormal basis {el, e2}, and if
A= (0 -0
0 0 ,

then the Pfaffian of A is 0.

The Pfaffian vanishes if the dimension of V is odd, while if it is even,
Pf(A) is a polynomial of homogeneous order n/2 in the components of A. If
the orientation of V is reversed, it changes sign.
Proposition 1.36. The Pfafan of an antisymmetric linear map is a square
root of the determinant:
Pf(A)2 = det(A).
1.4. Superconnections 43

Proof. By the spectral theorem, we can choose an oriented basis ej of V such

that there are real numbers cj, 1 < j < n/2 for which
Ae2,j-1 = ci e2.i
Ae2.j = -cj e23_1.

In this way, we reduce the proof to the case in which V is the vector space R2,
and Ael = 9e2, Ae2 = -Be1. In this case, Pf(A) = 0, while the determinant
ofA=(e o)isG2. r-,

We will frequently use the notation

(1.29) det1/2(A) = Pf(A).

The choice of sign in the square root det1/2(A) depends on the choice of
orientation of E.

1.4. Superconnections
In this section, we describe Quillen's concept of a superconnection. If M is a
manifold and £ = £+ ® £- is a superbundle on M, let A(M, 6) be the space
of £-valued differential forms on M. This space has a Z-grading given by the
degree of a differential form, and we will denote the component of exterior
degree i of a E A(M, £) by a[ ]. In addition, we have the total 7L2-grading,
which we will denote by
A(M, £) = A+ (M, £) ® A- (M, £),
and which is defined by
A'(M,£) _ A2i(M £f) ®EA2a+1(M,£:F);

for example, the space of sections of 6± is contained in A± (M, £).

If g is a bundle of Lie superalgebras on M, then A(M, g) is a Lie superal-
gebra with respect to the Lie superbracket defined by
[a1 ®X1, a2 ®X2] = (-1)IXII I*2l (al n a2) ® [X1, X2]
Likewise, if 6 is a superbundle of modules for g with respect to an action p,
then A(M, 6) is a supermodule for A(M, g), with respect to the action
p(a (9 X) (a ® v) Ipl(a n,Q) ® (p(X) v).

In particular, this construction may be applied to the bundle of Lie superal-

gebras End(S), where 6 is a superbundle on M, since AT*M ® 6 is a bundle
of modules for the superalgebra AT*M ® End(S); we see that A(M, End(£))
44 1. Background on Differential Geometry

is a Lie superalgebra, which has A(M, E) as a supermodule. We will fre-

quently use the fact that any differential operator on A(M, E) which super-
commutes with the action of A(M) is given by the action of an element of
A(M, End(E)); such an operator will be called local. (This is consistent with
the "super" point of view, that A(M) is the algebra of functions on a super-
manifold fibred over M, and that elements of A(M, End(E)) are the zeroth
order differential operators on this supermanifold.)
In defining a superconnection, Quillen abstracted the main properties of
a covariant derivative, namely, that it is an odd operator satisfying Leibniz's
rule. It turns out that many of the results which hold for ordinary connections
continue to hold for superconnections.
Definition 1.37. (1) If £ is a bundle of superspaces over a manifold M,
then a superconnection on E is an odd-parity first-order differential
operator
A: A}(M,E) --' A:F (M,£)
which satisfies Leibniz's rule in the Z2-graded sense: if a E A(M) and
0EA(M,£),then
A(aA0) =daA0+(-1)lalaAA9.
(2) If A is a superconnection on E, then A is extended to act on the space
A(M, End(E)) in a way consistent with Leibniz's rule:
Aa = [A, a] for a E A(M, End(E)).
In order to check that Aa, as defined by this formula, is an element of
A(M, End(E)), we need only observe that the operator [A, a] commutes
with exterior multiplication by any differential form Q E A(M).
Define the curvature of a superconnection A to be the operator A2 on
A(M, £).
Proposition 1.38. The curvature F is a local operator, and hence is given
by the action of a differential form F E A(M, End(S)), which has total degree
even, and satisfies the Bianchi identity AF = 0.
Proof. The operator A2 is local because it supercommutes with exterior mul-
tiplication by any a E A(M):
[A2, e(a)] = [A, [A, s(a)]] = e(d2ca) = 0.

Thus, the curvature A2 is given by the action of a differential form F E

A(M, End (6)) The Bianchi identity is just another way of writing the obvious
-

identity [A, A2] = 0.

It is clear that in the case when the superconnection E is a connection, the
above definition of curvature coincides with that of Section 1.
1.4. Superconnections 45

A superconnection is entirely determined by its restriction to r(M, £),

which may be any operator A : r(M, £t) --> AF (M, £) that satisfies
A(f s) = df s + f As for all f E C°° (M), s E r(M, £).
Indeed, if we define A(a 0 s) = das + (-1)I'IaAs for a E A(M) and s E
r(M, £), this gives the extension of A to all of A(M, £).
In order to better understand what a superconnection consists of, we can
break it up into its homogeneous components A[j], which map r(M, £) to
A1(M, £):
A = A[o] + A[i] + A[2] + .. .

Proposition 1.39. (1) The operator A(,] is a covariant derivative on the

bundle £ which preserves the sub-bundles £+ and £-.
(2) The operators A[ j] for i # 1 are given by the action of differential forms
w[i] E Ai(M,End(£)) on A(M,£), where w[j] lies in A'(M, End- (E)) if
i is even, and in Ai(M, End+(£)) if i is odd.

Proof. Let us decompose Leibniz's rule for A acting on r(M, 6) according to

degree:
n n
A(fs) = EA[ ](fs) = df ®s+ f >A[i]s.
i=o i=o
It follows immediately from this that the operator A(,) is a covariant derivative
on the bundle £; it preserves the Z2-grading of £, since A(,] has odd total
degree, so that it is the direct sum of covariant derivatives on the bundles £t.
On the other hand, for i 54 1, we see that A[j] (f s) = f (A[i] s), in other
words, that A[ j] : r(M, £) -+ A(M, £) is given by the action of a differential
form w[j] in A[i] (M, End(£)); from the fact that A is an odd operator it follows
that w[i] is of odd total degree.
The actions of the operators A and w(°M + A[11 + w(21 + ... on A(M, £)
must agree, since any two operators that agree when restricted to r(M, £) _
A° (M, £) and satisfy Leibniz's rule are equal.
Corollary 1.40. The space of superconnections on F is an affine space mod-
elled on the vector space A- (M, End(£)). Thus, if As is a smooth one-
parameter family of superconnections, then dA,/ds lies in A- (M, End(£))
for each s.

We will call A[i] the covariant derivative component of the supercon-

nection A.
As an example of a superconnection, let us take a trivial bundle on M
with fibre a complex superspace E. If L E COO(M,End-(E)), we may form
a superconnection by adding L to the trivial connection d. The curvature of
46 1. Background on Differential Geometry

the superconnection d + L is then F = dL + L2. Relative to the splitting

E _ E+ ® E-, we may decompose L as
_r0 L"
L - \ L}
0/
where L+ E C°° (M, Hom(E+, E-)) and L- E C-(M, Hom(E-, E+)). In
terms of this decomposition, the curvature of d + L may be written
/L-L} dL-
F = ( dL+ L+L_

If S and Jr are superbundles, then £ ®.F is the superbundle graded by

(£(& F)+=£+®F+®£-®.F-,
(£®F)-=£+®F-®£-®.F+.
If A and 13 are superconnections respectively on £ and F, there is a tensor
product superconnection A ®1 + 1 ®1B on £ ® F defined by the formula
(A®1 +1 0 3)(aA,3) =AaA#+(-1)I0Ia.
for a E A(M, £) and,3 E A(M,.F).
If ¢ : M --+ N is a smooth map and £ is a superbundle on N with
superconnnection A, there is a unique superconnection 4*A on 0*£ such that
0*A(0*a) = 0*(Aa) for a E A(N,6), called the pull-back of A by 0.

1.5. Characteristic Classes

Let S be a superbundle, either real or complex, with superconnection A, on
a manifold M. Following Quillen, we will associate to this superconnection
certain closed differential forms, called the characteristic forms of A. These
give invariants of the superbundle £ in the de Rham cohomology H* (M) of M;
the cohomology classes thus obtained are called the characteristic classes of
S. We will only dealt with the special case of characteristic forms associated
to invariant polynomials of the type Str(A" ). In Chapter 7, we will give a
generalization of the theory of characteristic classes to the equivariant context,
where a Lie group acts on the data.
In order to define the characteristic forms, we need the supertrace map on
the space of differential forms A(M, End(S)). This is defined by means of the
supertrace Str., : End(£)., --p C, defined in each fibre of End(S) by

Str (c d) = Tr(a) - Tr(d).

Applying the construction of Definition 1.32 with auxiliary supercommutative
algebra AT*XM, we obtain a bundle map
Str : AT*M ® End(£) --> AT*M
1.5. Characteristic Classes 47

which, applied to the sections of these bundles, gives the desired supertrace:
Str : A(M, End(£)) - A(M),
characterized by the formula Str(a (9 A) = a Str(A) for a E A(M) and
A E r(M, End(£)). The algebra A(M) is supercommutative, so by (1.25),
this extension behaves like a supertrace from a superalgebra to C, in that it
vanishes on supercommutators. It is clear that the supertrace map preserves
the Z2-gradings of these spaces, that is, it maps A±(M, End (S)) to A± (M).
Suppose f (z) is a polynomial in z. If A2 E A+ (M, End(£)) is the curvature
of the superconnection A on £, then f (A2) is the element of A+ (M, End (9)),
defined by

f (A2) _ t(k!0) (A 2)k_

If we apply the supertrace map
Str : A+ (M, End(S)) -> A+ (M)
to f (A2), we obtain a differential form on M denoted by Str f (A2). We will
call this differential form the characteristic form, or Chern-Weil form, of
A corresponding to the polynomial f (z).
Proposition 1.41. (1) The characteristic form Str(f(A2)) is a closed diff-
erential form of even degree.
(2) (Transgression formula) If At is a differentiable one-parameter family of
superconnections on £, then

Str(f (At )) = d Str( dt '(Ai ))

dt f
where dAt/dt E A(M, End(£)) is the derivative of At with respect to t.
(3) If AO and Al are two superconnections on £, then the differential forms
Str(f (A2)) and Str(f (A2)) lie in the same deRham cohomology class.
Proof. The proof makes use of the following lemma.
Lemma 1.42. For any a E A(M, End(£)), d(Stra) = Str([A, a]).
Proof. In local coordinates on a subset U C M and with respect to a triv-
ialization of S over U, A = d + w, where w E A(U, Eid(£)) is the super-
connection form of A in this coordinate system, and hence Str([A, a]) =
Str([d, a]) + Str([w, a]). The second term Str([w, a]) vanishes, while the first
equals Str(da).
Using this lemma, we see that
d Str(f (A2)) = Str([A, f (A2)]) = 0.
This proves that Str(f (A2)) is closed; the fact that it has even degree is clear,
since f (A2) is even, and Str preserves degree.
48 1. Background on Differential Geometry

Let At be a differentiable one-parameter family of superconnections on £,

with curvatures At. We see that
dAt r - dAt
dt dt ]'
LAt'

(where the bracket is, of course, an anticommutator, since both At and dAt/dt
are odd operators). We now use the following formula: if at : [0, 1] -+
A+ (M, End (E)) is a family of differential forms on M with values in End (E)
of even total degree, then

Str(f (at)) = Str(&tf'(at)),

dt
where at is the derivative of at with respect to t. To prove this, it suffices to
consider the case f(z) = z":
d n-1
dt
Str(at) = Str ( ' atatat -a-1 I

i=O
= n Str(i tat-1)
where the last step uses the fact that is a supertrace in order to replace
Str(at&tat -2-1) by Str(atat-1). Applying this with at = At, we obtain
a

dt
Str(.f (At)) = Strl dtt '(A2))
f
= Str I [At, d t f'(A2)1)
t since [At, At ] = 0

=dStr(dtt f'(A2)).
t

If AO and Al are two superconnections and w = Al - A0, then we may

integrate the transgression formula applied to the family
At = (1 - t)A0 + tA1 = Ao + tw.
We see that

Str(f (A2)) - Str(f (Ao)) = d JO 1 Str(w f'(At )) dt

which shows explicitly that Str(f (Ao)) and Str(f (A2)) differ by an exact
form. p
There is another way to obtain the transgression formula, which is similar
to Proposition 1.24. Introduce the space M x R, with projection q : M x R -+
M. Using the family At of superconnections, we may define a superconnection
Ai on the superbundle q*£, by the formula
09,6(x, s)
(x, s) s)) (x) + ds A
8s
1.5. Characteristic Classes 49

The curvature .F of A is

.F = .F3 - d88 Ads,

where .Fs = As is the curvature of A97 thus,

f(J)=f(Fs-dasAds).

Since ds A ds = 0, we see as in the preceding proof that underneath the

supertrace,

Str(f(.F)) = Str(f(.F,s)) - Str(dsS f'(.F8)) Ads.

Since Str(f (.F)) is closed in A(M x R), we infer that

dStr(f (Fs)) - dStr(da ss f'(.Fs)).

A special case of the above construction of characteristic forms is that
in which the bundle £ is ungraded and A is a covari-
ant derivative V with curvature F E A2(M,End(£)); then Str(V2)k is just
the differential form Tr(Fk) E A2k(M). In this case we can allow f(z)
to be a power series in z, since the curvature F is a nilpotent element in
A(M, End(£)). The differential form Tr(f (F)) is usually called the Chern-
Weil form associated to the invariant power series A -4 Tr(f (A)) on 01(N),
(where N is the rank of £). If A is a superconnection, we can allow f to be
any power series with infinite radius of convergence.
It should be clear that any differential form which is a polynomial

P(Str(fi (A2)), ... , Str(fk(A2)))

is a closed even differential form. Since the ring of invariant polynomials

on gl(N) is generated by the polynomials Tr(ak), the ring of all forms thus
obtained coincides with the classical ring of Chern-Weil forms, in the case of
an ungraded bundle.
The reader is warned that topologists prefer to consider the characteristic
forms defined by the formula Str f (-F/2iri), where F is the curvature of
a covariant derivative V, because the Chern class, the cohomology class of
c(F) = det(1 - F/27ri), defines an element of the integral cohomology of M
only if these negative powers of 27ri are included. However, from the point of
view of the local index theorem, this is not as natural, which is why we have
preferred to give our less conventional definition.
Three cases of the construction given above are of special importance for
us.
50 1. Background on Differential Geometry

(1) If £ _ £+ ®£- is a complex Z2-graded vector bundle and f (z) =

exp(-z), the characteristic form obtained is called the Chern character
form of the superconnection,
Str(e-A2).

(1.30) ch(A) =

The Chern character form is additive with respect to taking the direct sum
of superbundles with superconnections:
(1.31) ch(A ® l3) = ch(A) + ch( )
Furthermore, it is multiplicative with respect to taking the tensor product of
superbundles with superconnections:
(1.32) ch(A (9 1 + 10 3) = ch(A) A ch(l3).
These two formulas explain why the differential form ch(A) is called a char-
acter.
The transgression formula for the Chern character of a family of supercon-
nections gives
ch(At) = -d Str( Ate-At),
dt
which implies that
f1
(1.33) ch(Ao) - ch(Al) = d J Str( dd t e-At) dt.

This proves the following proposition.

Proposition 1.43. The cohomology class of ch(A) is independent of the
superconnection A on 5; we will denote it by ch(S).
In particular, we see that replacing A by its covariant derivative component
V = A[i] does not change the cohomology class of the Chern character:
[ch(A)] = [ch(V)]
However, the connection V preserves the bundles £±, and if we denote by
V+ and V- the restrictions of V to E+ and £-, we have
ch(A) = ch(V) + dct = ch(V+) - ch(V-) + da
for some differential form a. In particular, if we take the corresponding
cohomology classes, we obtain
(1.34) ch(£) = [ch(0+)] - [ch(V-)] = ch(£+) - ch(£-).
(2) If £ is a real vector bundle with covariant derivative V of curvature F,
we associate to it its A-genus form,

(1.35) A(0) = det1/2(si hF/2) E A4-(M,][).

1.6. The Euler and Thom Classes 51

Using the formula det(A) = expTr(logA), we see that A(V) may be under-
stood as the exponential of the characteristic form associated to the power
series log ((z/2)/ sinh(z/2)), that is,
2
1 F/2 l
A(V) = expTr 2 log (
sinh(F/2)))'
The A-genus is multiplicative, in the sense that
A(v1 ® V2) = A(v1) A A(02).
The reason that A(V) only has non-vanishing terms in dimensions divisible
by 4 is that the function f (z) is an even function of z, so f (02) only involves
even powers of the curvature V. In fact, the coefficients of the A-genus of E
are polynomials in what are called the Pontryagin classes of M, but we will
not need to know how these are defined; the first few terms of the expansion
of A(V) in terms of the classes Tr(Fc) are as follows:
(1.36)

A(£,V)=1- 22112Tr(F2)+24360 [ (F4)+24288Tr(F2)2+..

The cohomology class of A(V) will be denoted A(£). In particular, the A-

genus A(M) of a manifold M is defined to be the A-genus of its tangent
bundle TM.
(3) If £ is an ungraded complex vector bundle with a covariant derivative
V with curvature F, we define its Todd genus form by the formula

Td(V) = expTr (log (eFF

1

Thus
Td(V) = det(eFF 1) E A2'(M,C).
Like A(V), the Todd genus is multiplicative. The cohomology class of Td(V)
will be denoted Td(£). In particular, the Todd genus Td(M) of a complex
manifold M is the Todd genus of its tangent bundle TM.
As we will see, the A-genus and its cousin the Todd genus arise quite
naturally in a number of situations in mathematics, for example, in the Atiyah-
Singer index theorem and in the Weyl and Kirillov character formulas.

1.6. The Euler and Thom Classes

In this section, we will define a characteristic class, the Euler class, which
differs from the characteristic classes defined in the last section, in that it is
only defined for oriented bundles. However, we will see that it has a definition
quite analogous to that of the Chern character. This section is a prerequisite
only to reading Section 7.7.
52 1. Background on Differential Geometry

We start by describing the Thom form of an oriented vector bundle. Let

M be a compact oriented manifold and let p : V -+ M be a real oriented
vector bundle over M of rank N. Let j : M -+ V be the inclusion of M into
V as the zero section.
Definition 1.44. A Thom form for the bundle V is a closed differential
form U E AN (V) such that

(27r)-N/2 f IM U = 1 E A°(M).

We will show that a Thom form exists, following a method of Mathai-

Quillen.
The following lemma is a fibred version of the homotopy used to prove the
Poincare lemma.

Lemma 1.45. Let R be the vertical Euler vector field on V, and let ht (v) _
tv. With H : A'(V) --+ A*-'(V) defined by the formula

Hp = f ht
1 (t(R),(j) t-ldt.
o

we have the homotopy formula

- p*(j*,0) = (dH + Hd),Q for all /3 E A(V).
Proof. First, note that the integral converges in the definition of H because
R vanishes at 0. If ,Q E A(V), let Ot = ht,8, and observe that /3o = p*j*Q and
that 01 = Q. Differentiating by t and using integrating the Cartan homotopy
formula
L(R)Q = d(t(R),Q) + t(R)(dld),
we obtain the lemma.

Corollary 1.46. The maps

/3l--r j*Q : H'(V) -+ H'(M) and
al ip*a : H'(M) -+ H'(V)
are inverses to each other.

We need another lemma, which is a consequence of Proposition 1.24.

Lemma 1.47. Let p : V -+ M be a real oriented vector bundle over a mani-

fold M. There is a smooth bilinear map m(cx, 8) from AS(V) x A3,(V) to
Ai+.1-rk(V)-1(V)
such that
(p*p*a) A 0 - a A (p*p*Q) = dm(a, 8) - (m(da,,6) + (-1)b0I m(a, da)).
1.6. The Euler and Thom Classes 53

Proof. Consider the one-parameter family kt of submersions from the Whit-

ney sum V x p V to V given by the formula
(v,w) EVxx V,,1 l tv+(1-t)wEVV.
If a and /3 are differential forms, we may form their external product a ®p
which is a differential form on V xp V. We easily see that
k° (a ®p 0) _ (p*pa) A /3,
k* (a ®p /3) = a A (p*p*l3)
Thus, it suffices to take m(a, 0) equal to

m(a,/3) = f0
1

The following result explains the importance of Thom classes.

Proposition 1.48. Let M be a compact oriented manifold and let p : V
M be a real oriented vector bundle over M of rank N, with Thom form
U E AN(V).
(1) The maps

18 --> (27r) -Nl2p.,8: He* (V) --+ H* -N(M) and

+N(V)
are inverses to each other.

(2) For every closed differential form 0 in A(V),

(27r)-N/2 f U A,6 =
V
f M
j*/3.

Proof. By (1.16), (27r)-N/2p* (U A p*a) = a for all a E A(M).

The other direction of (1) follows from the homotopy formula
(-1)N(27r)-N12m(U, d/3)
0 - (27r)-N/2U A (p*p*f3) = (27r)-N12dm(U,,3) -
which follows from Lemma 1.47 on setting a = (27r) -N/2U.
To prove (2), we use the homotopy H of Lemma 1.45. It follows that if
0EA(V),

(27r)-N/2 f
v
U A,3 = (27r)-N/2 f U
v
A p* (j*/l)

+ (27r)-N/2 f v
U A dH/3 + (21r)-N/2 f v
U A Hd/3.

The first term on the right-hand side equals fm j* f3, the second term vanishes
by Stokes's theorem, since U is closed, and the third term vanishes since 0 is
closed.
54 1. Background on Differential Geometry

In this section, we will construct a Thom form using the Berezin integral.
Let us start with a model problem, the construction of a volume form U on
an oriented Euclidean vector space V with integral fy U = (2ir)N/2; this is a
Thom class for V considered as a vector bundle over a point. If x1, ... , XN
is an orthonormal coordinate system on V, we define U by
U(x) = e-Ixl2/2dx1 A ... A dxN.
We will show later how to remedy the fact that U does not have compact
support.
Now, consider the identity map V --+ V to be an element of A° (V, V), with
exterior differential dx E A' (V, V). The exponential a-idx lies in A(V, AV).
We extend the Berezin integral T : AV --3 R to a map T : A(V, AV) -+ A(V)
by
T(a (9 ) = a E A(V) and l; E AV.

Lemma 1.49. (1) The form U equals

U(x) = e(N)T (e-Ixl2/2-idx

where e(N) = 1 if N is even and i if N is odd.

(2) The integral of U over V equals (27r)N/2.

Proof. Let ek be an orthonormal basis of V dual to the basis xk of V*. We

have
N
T (e-idx) = T ([J(1 - idXk (9 ek))
k=1
= (-i)NT ((dx1 ® el) ... (dxN (9 eN))
//

= (-i)N(-1)N(N-1)12
dX1 A ... A dXN,

and (1) follows, since (-i)N(-1)N(N-1)/2 equals 1 if N is even and -i if N

is odd. The integral of U is calculated using the Gaussian integral
°O

00
e-sl2 dx = 27r.

We will define a Thom form on an oriented Euclidean vector bundle p :

V --> M of rank N by modification of the formula of Lemma 1.49. Choosing
a Euclidean connection Vv on V, we may replace dx by Vvx E .A1(V, p*V),
where x is the tautological section of p*V. However, because the connection
Vv need not be flat, we will see that some modification to the definition of U
is necessary.
Suppose £ is an oriented Euclidean vector bundle of rank N over B. The
Berezin integral T defines a map
T : A(B, AE) -; A(B).
1.6. The Euler and Thom Classes 55

If V is a covariant derivative on £, we obtain a canonical covariant derivative

on A£.
Proposition 1.50. If V is a covariant derivative on £ compatible with the
metric on £, then for any a E A(B, A£), we have
dT(a) =T(Va).
Proof. Let N be the dimension of the fibres of S. The Berezin integral is
given by pairing a section of ANS with the section T E I'(B, AN£*) given
with respect to a local orthonormal oriented frame ei of £* by the formula
el A ... A eN. Since VT = 0, the result follows by Leibniz's rule.
We apply this proposition with B = V and £ = p*V, and with connection
VP*V = p* VV. The space of sections A = A(V, A(p*V)) is a bigraded algebra,
since it may be decomposed as a direct sum of subspaces
Ai,j = A2(V, Ajp*V).

The covariant derivative VP*v defines a map V : Ai'j -> Ai+i,j If s E

F(V,p*V), the contraction a(s) is defined by the formula

a(s)(a ® (Si A ... A sj)) _ sk)a 0 (s1 A ... Ask A... A sj)
k=1
for a E A(V) homogeneous and Sk E r(V,p*V), (1 < k < j). The contraction
a(s) defines a map a(s) : Ai'j -+ Ai'j-'. We will identify so(V) with A2V by
the map
A E so(V) -- E(Aei, ej )ei A ej.
i<j
For any s E A°'1 = F(V,p*V) and a E A, T(a(s)a) = 0, since a(s)a has
no component in A*,N. It follows that the Berezin integral T : A(V, Ap*V)
A(V) also satisfies the formula
dT(a) =T((V +a(s))a)
Let us list some elements of the algebra A:
(1) the tautological section x E A°'1 = T(V,p*V);
(2) the elements 1x12 E A°'0 and Vx E A',', formed from x;
A2,2 =
(3) the curvature F = (Vv)2 E A2(M,so(V)) gives an element of
A2(V, A2p*V) by identifying it with an element of A2(M, A2V), and pulling
it back to V by the projection p : V --+ M; abusing notation, we will write
this pull-back as F also.
Lemma 1.51. (1) Let S2 =x12/2 + iVx + F E A. Then
(V - ia(x))f2 = 0.
56 1. Background on Differential Geometry

(2) If p E C°° (R), define p(1) E A by the formula

N
PA _ k-o p(k) (1x12/2)
kI
(ivx + F)k,

Then (V - ia(x))p(f) = 0.
Proof. By Leibniz's rule, we have V (1x12) = -2a(x)Vx. By the definition of
curvature, we have V(Vx) = a(x)F, while by the Bianchi identity, we have
VF = 0. Combining all of this with the obvious facts that a(x)x = 1x12 and
a(x)1x12 = 0, we see that (V - ia(x))f = 0.
The formula (V - ia(x))p(S2) = 0 now follows from the fact that V - ia(x)
is a derivation of the algebra A.
Note that Sl E EO<k<2 Ak'k, so that p(fl) E Eo<k<N Ak,k We see that
T(p(S2)) E AN(V), and the above lemma shows that
dT(p(S2)) = T ((V - ia(x))p(52)) = 0,
so that T(p(52)) is a closed N-form on V.
We now define the form U on V by choosing p to be the function p(x) _
where e(N) = 1 if N is even and i if N is odd:
(1.37) U= e(N)T(e-sz)
= e(N)T(e-(I"I2/2+ivx+F)).
We will study in greater detail the analogies between this formula and that of
the Chern character in Section 7.7. Observe that the Thom form depends on
the orientation of V (through T), on its metric, and on its connection.
Proposition 1.52. The form (27r)-N/2U has integral 1 over each fibre.
Proof. To calculate f v/M U, it suffices to consider the the case in which M is
a point, so that V is a vector space; this is just Lemma 1.49.
The Thom form U constructed above has rapid decay at infinity instead
of being compactly supported. However, using the diffeomorphism from the
interior of the unit ball bundle of V to V given by the formula
y H y
1y12)1/2'

(1 -

we can pull back the Thom form U to obtain a Thom form with support in
the unit ball bundle of V.
Just as for the Chern character, there are transgression formulas for the
Thom form which show how it changes when the metric is rescaled and when
the connection changes. Let us first consider the effect of rescaling the metric:
this is equivalent to replacing S by
SZt = t2Ix12/2 + itVx + F.
Let Ut denote the Thom form corresponding to nt.
1.6. The Euler and Thom Classes 57

Proposition 1.53. We have the transgression formula

dUt = _ie(N) dT(xe-nt).

dt

Proof. Observe that

dtt = tlxl2 + iVx = i(V - ita(x))x.
Since (V - ita(x))52t = 0, we see that
-dtt a-0t
dt e-sit - = -i(V - ita(x)) (xe "t)
and hence that
dUt
= -ie(N) T ((V - ita(x)) (xe 0t))
dt
=-ie(N)dT(xe-°t).

By a very similar proof, we obtain a transgression formula for the Thom

form under a change of connection. Let Us be the Thom form corresponding
to a smooth family of connections O9 on V compatible with its inner product.
The derivative of this family dVs/ds is an element of 41(M, A2V); we
also denote by dVs/ds the element of = A1(V, A2V) obtained by pulling
A1,2

IX12
back dVs/ds to V, and let 1 + iVsx + Fq be the element of A
corresponding to the connection V8.
Proposition 1.54. We have the transgression formula
dUs
= -e(N) dT dd s e-n' )
Proof. It is easy to see that
dQ_,
= (Vs - ia(x)) d--'
ds ds
Using the formula (Os - ia(x)) , = 0, we see that
d e-spa = ds e-ne = -(Vs - ia(x)) (Y! ens)
and hence that
ddss
= -e(N) T I (V8 - ia(x)) (e_))
_ -e(N) dT I dd s I .

We now restrict attention to even-dimensional V. Let j : M -+ V be

the inclusion of M in V as the zero-section. The pull-back j*U is a closed
58 1. Background on Differential Geometry

differential form on M called the Euler form of V, and denoted X (Vv) .

Thus, if F is the curvature of Vv, we have
(1.38) j *U = X(V') = Pf(-F).

Proposition 1.55. The Euler form X(Vv) is a closed differential form of

degree N on M whose cohomology class depends only on the bundle V and
its orientation and not on the metric and covariant derivative on V.
Proof. Choose a metric on V. If V9 is a one-parameter family of Euclidean
connections on V with corresponding Thom class Us, the transgression for-
mula (1.54) for U3 implies that

X(Vo) - X(Vl = d f o1 T dst

dt
n expn(-Ft)) dt,

which is the analogue of the trangression formula (1.33) for the Chern char-
acter. This shows that the cohomology class of X(V) is independent of the
connection used in its definition.
We can also show that the cohomology class of X(V) is independent of the
metric used in its definition. If and ')1 are two metrics on V, then
we can find an invertible orientation preserving section of End(V) such that
(v, v)1 = (gv, gv)o-
If Vo is a Euclidean covariant derivative for the metric then it is easy
to see that V1 = g'1 Vo g is a Euclidean covariant derivative with respect
to the metric .)1. If we now construct the Euler forms of the two covariant
derivatives Vi with respect to the metrics we see that they are equal.
A case of special importance is that in which the bundle V is the tangent
bundle TM of an oriented even-dimensional Riemannian manifold. In this
case, the differential form X(V), where V is the Levi-Civita connection, is
called the Euler form X(M) of the manifold, and lies in A'a(M), where n
is the dimension of M. For example, if M is a two-dimensional Riemannian

C0
) then M
manifold, and if {el, e2} is an orthonormal basis of T,,M for which R(el, e2) _
.gel A e2.
Let v E I'(M, V) be a section of the vector bundle V. The pull-back v*U
is a closed differential form of degree N on M, given by the formula
v*U = T(e-(Iv12/2+sVvv+F)
The transgression formula Proposition 1.53 implies that v*U is cohomologous
to the Euler class for any v E F(M, V):
/1
x(Vv) - v*U = -i d T (v A e-(O dt.
J0
1.6. The Euler and Thom Classes 59

The above results imply a form of the Gauss-Bonnet-Chern theorem, which

is a formula for the Euler number of a manifold in terms of the integral of its
Euler class. This is the simplest example of an index theorem.
Let v E P(M, TM) be a vector field on M. At a zero p of v, the Lie bracket
X --4 [X, v]
induces an endomorphism Lp(v) of TpM; in a coordinate system in which
p = 0 and v = E 1 v', 9i, Lp(v) is given by the formula
n
LP(v)ai = 1: ('9iO(p))a7
7=1

A zero p of v is called non-degenerate if Lp(v) is invertible. We denote by

v(p, v) E {f1} the sign of the determinant of Lp(v).
If all of the zeros of v are non-degenerate, we say that v is non-degenerate.
If M is compact, such a vector field has only a finite number of zeroes.
A non-degenerate vector field may be constructed with the help of a Morse
function f, that is, a real smooth function on M such that at each critical
point p of f, there is a coordinate system in which
f (p) n ixz2
f= + i , where Ai E {±1}.
i=1
If we choose a Riemannian metric on M which is equal to >i dxi ® dxi
in each of these coordinate neighbourhoods, we see that the vector field
grad f equals E 1 )x9 around the critical point p of f , and hence that
v(p, grad f) = f2 1 A j. The following theorem is Chern's generalization of
the Gauss-Bonnet theorem to arbitrary orientable compact manifolds. The
form in which we derive it here is not the usual statement of the Gauss-
Bonnet-Theorem, because it is phrased in terms of a definition of the Euler
number in terms of the zeroes of a non-degenerate vector field on M; however,
this agrees with the definition of the Euler number as alternating sum of the
Betti numbers of M, as we will prove in Theorem 4.6.
Theorem 1.56 (Poincare-Hopf). If v is a non-degenerate vector field on
an orientable compact manifold M, then

1: v(p, v) = (27r) -'I' IM X(M)

{pI v(p)=0}

In particular, the sum

Eul(M) = E v(p,v)
{pI v(p)=0}

is independent of the non-degenerate vector field; it is called the Euler num-

ber of the manifold.
60 1. Background on Differential Geometry

Proof. Choose a coordinate chart in a neighbourhood UP of each zero p of v.

In this coordinate chart, the tangent bundle is trivialized, T UP ^_' UP x Rn.
The vector field v defines a smooth map from UP to Rn such that v(p) = 0,
with invertible derivative LP(v) at p. It follows that v maps an open subset
VP of UP diffeomorphically to a small ball B C R'1. The orientations on VP
induced by M and by the diffeomorphism v to Rn differ by the sign v(p, v).
We may assume the open sets VP to be disjoint. Choose a Riemannian metric
on M which on VP equals the metric induced by the diffeomorphism into Rn.
Since M is compact, we see that there is a positive constant s > 0 such that
JJvJI > e on the set M\ UP VP.
Let U be the Thom form of TM, with respect to the given Riemannian
metric and its associated Levi-Civita connection. For each t > 0, we have the
formula
fMvt U = fm X(M)
Let 0 : R+ --> [0, 1] be any smooth function such that
0(s)=1 ifs<e2/4
z/i(s)=0 ifs>s2.
Then

(1.39) f Mvt U = f
M
(1 - (IIvII2))vi U + f
M
V)(IIv1I2)vt U.

Since vt U is of the form

n
vt U= e-t' IIv112 /2 tic ,
i=O

where cx2 are differential forms on M, it is rapidly decreasing as a function

of t as t --> oo when JJvI12 > 62/4. We see that the first integral in (1.39) is
a rapidly decreasing function of t. The second integral is a sum of integrals
over the neighbourhoods V.
On VP, the metric and connection are trivial and v = a 1 x281i so that
vt UIvP = tnv(p,v)e t'llXll'/2dxl A... Adxn.
Thus
V)(IIxII2)e-t211X112/Zdx1
= tnv(p, v) f A ... A W.
vT 0(IivII2)vtU
Making the change of variables x -+ t-lx, we see that

lim to f (IIxI12)e-t2 l1X11'/2 dxl

A... A dxn = (21r)n/2O(0) = (27r)n/2.
t->oo n

Taking the sum over the zeroes p of v, the result follows.

Bibliographic Notes 61

There is a generalization of this result, to the case where we only assume

that the zeroes of v are isolated.
Definition 1.57. The index v(p, v) of a vector field v on a manifold M at
an isolated zero v(p) = 0 is defined by choosing a coordinate system around
p, in which case v(p, v) is the degree of the map from the sphere S,, _ {x E
Rn I jlxii = u} to the unit sphere defined by

x E S.,. 1
)
llv(x)ll-'v(x)
The index v(p, v) is an integer which is independent of the coordinate chart
used in its definition, as well as the small parameter u.
We can again choose small balls VP around the zeroes of v such that 11v 11 > e
on the complement of UPVP. The map sending x E VP to v(x) is a covering
map over the set of regular values, and using Sard's theorem, we see that

fvT>
V,(Ilvll )vt U = tnv(p, v) fRn V)(iIvIi2)e
t211X112/2dx1 A ... A dxn.

Theorem 1.58 (Poincare-Hopf). If v is a vector field with isolated zeroes

on a compact oriented even dimensional manifold M, then

E v(p, v) = (27r)-n12 X (M)

{P1 v(P)=O}
IM

The book of Kobayashi-Nomizu [74] is perhaps the best reference for the the-
ory of bundles and connections.

Section 2. For background on Riemannian manifolds, the reader may refer

to the books of Milnor [83], Berger-Gauduchon-Mazet [20] and Kobayashi-
Nomizu.
The proofs of Proposition 1.18 and 1.28 are taken from the second appendix
of Atiyah-Bott-Patodi [8].

Sections 3 and 4. It was Berezin [19], motivated by the problem of understand-

ing geometric quantization for fermions, who emphasized the strong analogy
between calculus on superspaces and ordinary vector spaces.
The notion of a superconnection was developed by Quillen in order to
provide a formalism for a local family index theorem for Dirac operators [93].
This hope was realised by Bismut, as we will see in Chapters 9 and 10. The
notion of superconnection has turned out to be fruitful in other contexts, too.
62 1. Background on Differential Geometry

Section 5. We have presented the theory of characteristic forms in the con-

text of superconnections, following Quillen. The construction of characteristic
forms for connections is due to Chern and Weil; see Kobayashi-Nomizu. Our
interest is more in the characteristic forms themselves than in their cohomol-
ogy classes. The reader interested in the topological point of view may consult
Bott-Tu [44] and Milnor [85]. This section is also an introduction to Chap-
ter 7, where we define equivariant characteristic forms; thus, the bibliographic
notes of that chapter are relevant here.
Section 6. The results on the Thom form and the Euler form presented here
are due to Mathai and Quillen [81], although we give more direct proofs which
use the Berezin integral. The proof of Theorem 1.56 is close to Chern's proof
of the Gauss-Bonnet-Chern theorem [47]; we have presented it in such a way
as to show the analogy with the theorems of later chapters. For more on the
proof of the Poincare-Hopf Theorem 1.58, see Bott and Tu [44].
Chapter 2. Asymptotic Expansion of the Heat Kernel

Given a Riemannian structure on a manifold M, one may construct the scalar

Laplacian 0 = d*d, which is the operator on the space of functions on M
corresponding to the quadratic form

(f, Of) = fM' df) dxl

This operator can be used to study the geometry of the manifold, and is also
important because it arises frequently in physics problems, such as quantum
mechanics and diffusion in curved spaces. All of the eigenfunctions of A are
smooth functions and, if the manifold M is compact, i has a unique self-
adjoint extension. The Laplacian has a heat kernel, which is the function
kt (x, y) on R+ x M x M which solves the equation
at kt (x, y) + A.kt (x, y) = 0,
subject to the initial condition that limt_*0 kt(x, y) is the Dirac distribution
along the diagonal 6(x, y). It turns out that the heat kernel kt(x, y) is a smooth
function whose behaviour when t becomes small reflects the local geometry
of the manifold M. In particular, there is an approximation to kt(x, y) near
the diagonal, of the form
00
(47rt)-,/2e-d(xy)2/4t E
kt(x, y) - tifi(x, y),
i=0

where d(x, y) is the geodesic distance between x and y, the integer n is the
dimension of M, and the functions fi(x, y) are given by explicit formulas when
restricted to the diagonal.
In this chapter, we will construct the heat kernel of a more general class
of operators which generalizes the scalar Laplacian to vector bundles. The
construction is extremely intuitive geometrically, being based on the explicit
formula involving a Gaussian for the heat kernel of the Laplacian on Eu-
clidean space. Since the manifold on which we work is compact, we require
no estimates more sophisticated than the summation of geometric series.
These generalized Laplacians are introduced in Section 1, while the con-
struction of the heat kernel, when M is compact, occupies Sections 2-4. In
Section 5, we give some applications, such as the smoothness of eigenfunctions
of generalized Laplacians and the Weyl formula for the asymptotic number
64 2. Asymptotic Expansion of the Heat Kernel

of eigenfunctions in terms of the volume of the manifold and its dimension.

In Section 6, we discuss the dependence of the heat kernel on the Laplacian if
the coefficients of the Laplacian are varied smoothly.

2.1. Differential Operators

Let M be a manifold, and let E be a vector bundle over M. The alge-
bra of differential operators on £, denoted by D(M, £), is the subalgebra of
End(F(M, £)) generated by elements of r(M, End(£)) acting by multiplica-
tion on r(M, E), and the covariant derivatives Ox, where V is any covariant
derivative on £ and X ranges over all vector fields on M. The algebra D(M, £)
has a natural filtration, defined by letting
Dti(M, £) = r(M, End(-')) span{Vx, ... Ox; I j < i}.
We call an element of Di(M, £) an i-th differential order operator. Note that
if D is an i-th order operator and f is a smooth function, then (ad f )iD is a
zeroth order operator.
If A is any filtered algebra, that is, A = U°° o Ai with Ai C Ai+1 and
Ai A; C Ai+i, we may define the associated graded algebra
00 00
grA=EgriA=EAi/Ai-1
i=0 i=o
where the component in degree i is gri A. The projection from the graded
algebra E°_o Ai to grA is called the symbol map, and written its com-
ponent in degree i is the projection Ai -+ gr A with kernel Ai_1.
We will now apply this formalism to the algebra of differential operators
D(M,£). Let S(TM) be the infinite dimensional graded vector bundle over
M whose fibre at x is the symmetric algebra on TIM.

Proposition 2.1. The associated graded algebra

gr D(M, £) = E Dk (M, £) /Dk-1(M, £)

k_o

to the filtered algebra of differential operators is isomorphic to the space of sec-

tions of the bundle S(TM) ® End(S). The isomorphism ok : grk D(M, £) -->
P(M, Sk(TM) (9 End(£)) is given by the following formula: if D E Dk(M, £),
then for x E M and 6 E TAM,

(2.1) vk (D) (x, t;) = lira t-k (e-itf . D eitf) (x) E End(£x),
t--+00

where f is any smooth function on M such that df (x) =1;.

2.1. Differential Operators 65

Proof. Let D be the differential operator aVx1 -

Vx,E, where X1, ... , Xk are
vector fields on M and a E r(M, End(£)). By Leibniz's rule, the operator
e-itf D - eitf is a differential operator depending polynomially on t, of the
form
(it)ka(x) (X 1(x), e) ... (Xk(x), 6) + 0(tk-1).
It follows that the leading order is a zeroth-order operator, and that the
limit limt_._,oo t-k (e-itf - D - eitf) (x) is independent of which function f with
df (x) = l; we choose, and defines a linear isomorphism from grk D(M, End(£))
to r(M, Sk(TM) (9 End (6)).
By differentiation of (2.1), we see that
ak(D)(x,6) ( i (ad f)kD.
If Di E Dk.; (M, £), i = 1, 2, then
ak1+k2(D1D2)(x,6) =0k1(D1)(x, akz(D2)(x,S)-
f
We will often write a(D) instead of ak(D), if k is the degree of D.
We may identify r(M, S(TM) (9 End(£)) with the subspace of sections
in r(T*M, ir* End(£)) which are polynomial along the fibres of T*M. A
differential operator D of order k is called elliptic if the section ak(D) E
r(T*M, 7r* End(£)) over the cotangent space is invertible over the open set
{(x,l;) 16 O}.
Definition 2.2. Let £ be a vector bundle over a Riemannian manifold M. A
generalized Laplacian on £ is a second-order differential operator H such
that
0'2(H)(x,e) _ ieI2.
Clearly, such an operator is elliptic. An equivalent way of stating this
definition is that in any local coordinate system,
H = - E gij (x)8i8f + first-order part,
i,j
where gi.9 (x) _ (dxi, dxj)X is the metric on the cotangent bundle. By defini-
tion of the symbol, we can characterize a generalized Laplacian as follows:
Proposition 2.3. An operator H is a generalized Laplacian if and only for
any f E C°O (M),
[[H,fl,fl=-2IdfI2.
Let £ be a vector bundle bundle on a Riemannian manifold M, with con-
nection V6 : r(M, £) -+ r(M, T*M(9£). Since M is Riemannian, it possesses
a canonical connection, the Levi-Civita connection V, by means of which we
define the tensor product connection on the bundle T*M ® S. In this way, we
obtain an operator
VT*M®EVS : r(M, s) , r(M,T*M ®T*M (9 £)
as the composition of the two covariant derivatives.
66 2. Asymptotic Expansion of the Heat Kernel

Definition 2.4. The Laplacian AE on- r(M, £) (which we denote by A if

there is no ambiguity) is the second-order differential operator
DES = -Tr(VT*M®EDES).

Here, we denote by Tr(S) E r(M, £) the contraction of an element S E

r(M, T*M ® T*M (9 £) with the metric g c r(M, TM ® TM).
If s E r(M, £) and X and Y are two vector fields, then
(VT*M®eVes)(X,Y) _ (V5 - VY) S.
If ei is a local orthonormal frame of TM, we have the following explicit formula
for AE:
AE = E
; e; -EDVesea J)
On the other hand, with respect to the framing a/axi defined by a coordinate
system around a point in M, the operator AE equals

AS gij(x)f VaVa - Er vat

\
23
ij \ k

where r are the components of the Levi-Civita connection, defined by Viaj =

>k r j ak. Thus, the Laplacian AE is a generalized Laplacian.
In particular the scalar Laplacian is given for f E CI(M) by the formula

If = -Tr(Vdf) gij(x)(aiaj - Er ak) f.

ij k

If £1 and £2 are two bundles on M with connections VEI and DE2, let AEI,
AEz and AEI®E2 be the Laplacians defined with respect to the connections
V&1, VE2 and VSI®E2 = VE® 01+10V62. The following formula is clear:
for si E r(M, £i),
A61062 (S1 (9 s2) = (AEIS1) ® S2 - 2Tr(VElsl (& V-0282) + s1 ® AE2s2.

As we will now show, any generalized Laplacian is of the form i + F,

where AE is the Laplacian associated to some connection DE, and F is a
section of the bundle End(£).

Proposition 2.5. If H is a generalized Laplacian on a vector bundle £, there

exists a connection QE on £ such that for any f E C°° (M), [H, f] is the
differential operator whose action on r(M, £) is given by the formula
(2.2) [H, f] = -2(grad f, V6) + Af,
where A is the scalar Laplacian. Furthermore, F = H - AE is a zeroth-order
operator.
2.1. Differential Operators 67

Proof. Let us define an operator yE from r(M, £) to A' (M, £) by the formula
(fo grad fl, 17-s) = fo (-H(fis) + f, (Hs) + (Ofl)s),
2
for all fo and fl E CO°(M). The relation
0(flf2) = (Afl)f2 - 2(dfl, df2) + fl(of2)
and the formula [[H, fl], f2] = -2(dfl, df2) implies that this definition is con-
sistent, in the sense that
(fograd(flf2),DES) = (foflgrad f2+fof2grad fl,DES),
and that yE is a connection on the bundle £:
[(grad f,VE),h] = [-2[H,f] + 2Of,h] = -2[[H,f],h] = (grad f,dh).
It is now easy to show that F = H - DE is a zeroth-order operator: if f is
any smooth function on M, then
[H - DE, f] = 0,
since both [H, f] and [DE, f] equal -2(grad f, VS) +Af.
To summarize, a generalized Laplacian is constructed from the following
three pieces of geometric data:
(1) a metric g on M, which determines the second-order piece;
(2) a connection Ve on the vector bundle £, which determines the first-order
piece;
(3) a section F of the bundle End(£), which determines the zeroth order
piece.
If S is a vector bundle and £* is its dual vector bundle, then the space
r, (m, £) of compactly supported sections of £ is naturally paired with the
space of sections of £* ® IAA: if we take two sections s1 E r,(M,£) and
s2 E r(M, £* (9 1AI), then their pointwise scalar product (sl, s2). is a section
of the density bundle IAA, so can be integrated, giving a number fM(s1, s2),
the pairing of the two sections. In particular, if £ is a Hermitian bundle, then
the space r, (M, S ®1AI1/2) has a natural inner-product.
Definition 2.6. (1) If Si and £2 are two vector bundles on M and D
r(M, £1) -> r(M, £2) is a differential operator, the formal adjoint D* of D
is the differential operator
D* : r(M, £2 (9 Al)
J--+ r(M, £i (9 JAI)
such that
(Ds, t) = (s, D*t)
IM IM
for s E r°(M, £1) and t E r(M, £2 ®JAI).
68 2. Asymptotic Expansion of the Heat Kernel

(2) If £ is a Hermitian vector bundle and D a differential operator acting

on r(M,£ (9 JAI1/2), we say that D is symmetric if it is equal to its formal
adjoint, D = D*.

It is easy to see that the symbol of the formal adjoint D* is equal to

oi(D*) = oi(D)* ®11Aj.
Although the bundle JAS has no canonical trivialization, in most of this
book, we will deal with Riemmanian manifolds, which have a canonical no-
where vanishing density JdxJ E I'(M, JAI). Using the identification of r(M, £)
with I'(M, £®IAJ) which results from multiplication by Jdxi, the formal adjoint
of D : P(M, £1) --+ r(M, £2) becomes an operator D* : r(M, £2) -> r(M, £1).
However, by taking into account the density bundles, we get the most geo-
metrically invariant formulations of many theorems.
On a Riemannian manifold, the formal adjoint of the exterior differential
d : A' (M) --p A6+1 (M) is an operator from the bundle A"+1TM ® IAA to
the bundle A*TM ®JAI. Using the identification of the bundles A'T*M and
A'TM®JAI which comes from the Riemannian metric on M, we may identify
it with a map d* : A'+1 (M) -* A* (M). Thus, d* is defined by the formula

(dQ, a) I dxI = ()3, d* a) I dxI ,

IM IM
where a and Q are compactly supported differential forms on M, and JdxJ is
the Riemannian density.
If a is a one-form, then d*a is a function called the divergence of a.
Taking )3 = 1 in the preceding equation, we see that

d*a idxJ = 0
fM
if a is a compactly supported one-form.
If a is a one-form, the function Tr(Va) is given by the formula

Tr(Va) _ eia(ei) - a(Ve,ei),

where e,, is a local orthongrmal frame.

Proposition 2.7. The divergence on one forms is given by the formula

d*a = -Tr(Va),
and hence for any compactly supported C1 one-form a, we have

Tr(Va) IdxI = 0.
IM
2.1. Differential Operators 69

Proof. If X is a vector field on M, then

VX E r(M,T*M (9 TM) = r(M,End(TM))
is the endomorphism of TM which acts on Y E r(M,TM) by the formula
(VX) Y = VyX. Since V is torsion free, we see that
£(X)Y = [X, Y] _ VXY - (VX)Y,
hence, using the fact that I dxI is preserved by the Levi-Civita connection, that
=TI (VX)Idxl.
£(X)I dxl
If f E C°°(M), it follows from the formula fm £(X)(f Idxl) = 0 that

f X (f) I dxl =- f
M M
Tr(VX) f ldxl.

If a is the one-form on M corresponding to X under the Riemannian metric

on M, so that X (f) = (a, df ), we see that

f (a,df)Idxl=fM X(f)Idxl=- f
M M
Tr(Da)fIdx!.
The following proposition extends the above formula for the divergence, by
giving a formula for d*a in any degree.
Proposition 2.8. Denote by t : r(M,T*M ®AT*M) --> F(M, AT*M) the
contraction operator on differential forms defined by means of the Rieman-
nian metric on M. Then
d* = -t o VAT*M
Proof. Since the Levi-Civita covariant derivative is torsion free, we see from
Proposition 1.22 that d = e o V. Let /3p E A'(M), ,Qp+l E Ap+1(M), and
let a be the one-form given by the formula a(X) = (/3p, c(X)/3p+1). Using
Leibniz's rule and the fact that at each point x E M, e* = t, we see that
(EVgq,.p-1)a = -(8p, tVQp+1)x + Tr(Da)x.
Integration over M kills Tr(Va), and we obtain the proposition.
We will now calculate the formal adjoint of the Laplacian Y. Recall
that the bundle IAI- of s-densities on a Riemannian manifold has a natural
connection, which preserves the canonical section Idxl8. If DE®lAl is the
Laplacian associated to the connection Ve®lAl on the bundle 6 (9 JAI", we
have
oe®lAia(0IdxI9) =
)Idxls
Because of the following result, it is often more natural to write formulas with
half-densities. Let IM(s1, s2) denote the natural pairing between a compactly
supported section s1 of the bundle £ ® IAI1/2, and a section $2 of the bundle
£* 0 IAI'/2. We consider on the bundle £* the connection V* dual to the
connection Ve on S.
70 2. Asymptotic Expansion of the Heat Kernel

Proposition 2.9. If sl is a compactly supported C2 -section of £®IAI1/2 and

s2 is a C2-section of £* ®IAI1/2, then

(DE®IAI1(2S,
S2) = (sl, 0e*®IAI112s2)
Jm
_ Tr(VE®IAI1(2s,i vE*®IAI"2s2),
M
where the last inner product is defined using the natural pairing between
T*M®£ and T*M®£*.
Proof. If we apply the Levi-Civita covariant derivative V to the density
(VE®IA11/2s1i S2)x, we obtain the continuous section of T*M ® T*M

V(VE®IAI11251,S2)x =
VE*®IAI'1/252)x,
(vT*M®E®IAI1/2VE(&IAI1/2 s1, S2)x + (VE®IAI1/251,

so that
Tr(V(VE®IA1'/2Sl,s2))x =
Tr(DE®IAI1(2s1i pE*(9IAl1/2s2)x.
l _ (DE®IAI112s1i 52)z +
Tr(V(VE®IAI1i2s1i

By Proposition 2.7, the term s2))x vanishes after integra-

tion over M. 17
If 6 has a Hermitian metric, then the bundle £* is naturally identified with
£, and there is a natural inner product on the space P(M, £ (9 IAI1/2). In this
situation, the above proposition has the following consequence.

Corollary 2.10. If the connection V on £ is compatible with the Hermi-

tian metric' on 6, then the Laplacian AE®IAI112 on the bundle ,6 ®IAI1/2 is
symmetric with respect to the inner product on r(M, £ 0 JAI 1/2).

Applying Proposition 2.9 with £ = M x C the trivial line bundle, we see in

particular that the scalar Laplacian satisfies

(Of, f) Idxl = (df,df) Idx1,

IM IM
or, in other words, A f = d*df.
In the following lemma, we collect some formulas for the Laplacian on the
bundle of half-densities in normal coordinates x '--* exp.,, x around a point
xo. These formulas involve the function j (x) on T,,,, M defined in Chapter 1
by the property that the pull-back of the volume form dx on M by the map
expxo equals j(x) dx, or, in other words,

j(x) = I det(d.expxo)I = detl"2(ga,(x))

2.1. Differential Operators 71

The reader should beware the distinction between the symbol x, denoting any
coordinate system on the manifold M, and x, denoting coordinates on the
tangent space Two M, or normal coordinates on M.
Let R be the radial vector field R xiaXi. Denote the Laplacians on
functions and on half-densities by A.

Proposition 2.11. In normal coordinates x around x0, we have

(1) o(fldxl1/2) = (df - f d(logj))Idxll/2, for f E C°°(M)
2
(2) A(f Idxll/2) _ ((jl/2 o A o j-1/2)f )Idxll/2

(3) AIIXII2 = -2(n + R(log j))

(4) If qt(x) is the time-dependent half-density on a normal coordinate chart
in M centred at x0 given by the formula

qt(x) = (47rt)-n/2e 11X112/4tIdx11/2,

then
(at + A - j1/2( A . j-1/2))qt = 0.

Proof. Parts (1) and (2) follow from the formula

V(jl/2ldxl1/2) = Vldxll/2 = 0.
To prove (3), we use integration by parts: if 0 E C,00 (TT0 M), then

f (0, dllxll2) j(x) dx = f (dO, dllxll2) j(x) dx = 2 f (RO) j(x) dx;

here we have used that (dllxll2, dO)/2 = R4, which is a consequence of Prop
sition 1.27 (3). Since the adjoint of the vector field R with respect to the
density Idxl is -n - R, we see that

J(l!xlI2)i(x) dx = -2 f (n + R(log j)) O(x) j(x)dx.

As for (4), Leibniz's rule shows that A(e 11X112/4tldxll/2) equals

A(e-11X112/4t) Idxl1/2 + t-1e-IIX112/4tV ldxll/2 +- e-11X112/4toldxl1/2

and that 0(e-IIXII2/4t) equals

4te-11X112/4t(2(n+R(logj))
- t-111x112)
jl/2(0, j-1/2))qt
From these formulas, it is easy to show that (at + A -
vanishes.
72 2. Asymptotic Expansion of the Heat Kernel

2.2. The Heat Kernel on Euclidean Space

Let V = R be the standard Euclidean vector space of dimension one, and let

0=- d2 2

be its Laplacian. If
qt (x, y) = (47rt)-1/2e (X-Y)2/4t ,
it is easy to check that
a
at
)
+ OX q, (x, y) = 0,

where OX denotes the Laplacian acting in the variable x. This equation is a

special case of Proposition 2.11 (4), since i(x) = 1 on IR.
We will need the following lemma.
Lemma 2.12. The Gaussian integrals
v2/4vk
a(k) _ (47r)-1/2 f e dv

a re given by the formula

k!
k even
a(k) = (k/2)!' '
10, k odd.
Proof. Consider the generating function
k
A(t) = E t a(k) = (47r)-1/2 j e-v2/4+tv dv.
k=o k' m

By the change of variables u = v - 2t, we see that this integral is equal to

A(t) = (47r)-1/2 r e-u2/4+t2 du = et2,

from which the lemma easily follows.

Define the norm on Ce-functions by
k
k<<XERldxkO(x)I.

Let Qt be the operator on functions on ]l8 defined by

(Qt0)(x) = fqt(x, Y)O(Y) dy

= (47rt)-1/2 e-(X-Y)2/4to(Y)
dy.
Jle
2.2. The Heat Kernel on Euclidean Space 73

Proposition 2.13. Let Q be an even number. If 0 is a function on R with

II0IIP+1 finite, then

k
Ako < O(ti/2+1).
Qtcb - E (k!
k=O

Proof. The change of variables y -. x + t1/2v shows that

(QtO)(x) _ (47r) -1/2 J v2/40(x + t1/2v) dv.

fl
If I I 0 I I P+1 < oo, Taylor's formula
e k r
O(x + v) _ E k O(k) (x) + -Q! 1 1 (1 - t)" f (P+1) (x + tv) dt
k=O J0
shows that

O(x + t1/2V) t k/2vk IyI

P+1t(1+1)/2
k! (k)(x)II <

< 0(t(P+1)/2)
II Qtcb -k=0
E a(k) t k!2 c(k) (x) II
The proposition follows immediately from this estimate and the formula of
Lemma 2.12 for a(k).

In this chapter, we will show that there is an analogue of the function

qt (x, y) on any compact Riemannian manifold, and also that the analogue of
the above proposition holds.
In Section 5.2, we will use the fact that on any Euclidean vector space V,
if 0 E C°°(V), there is an asymptotic expansion in powers of t1/2,

f e-IIajj2/4t0(a) da N (47ft)dim(V)/2 E(-t)k(Okq5)(O)

V k=O

which we will denote by

Tpaeymy
e-IIaII2/4t0(a)da.
V

This is proved by the same change of variables as was used in the proof of
the Proposition 2.13.
74 2. Asymptotic Expansion of the Heat Kernel

2.3. Heat Kernels

In this section, we will define the heat kernel of a generalized Laplacian on a
bundle S ®IAI1/2 over a compact Riemannian manifold M. As background to
this definition, let us recall the kernel theorem of Schwartz. If £ is a bundle on
a compact manifold M, the space r-°°(M,£) of generalized sections of a
vector bundle S is defined as the topological dual space of the space of smooth
sections P(M, £® ® JAI). There is a natural embedding P(M, £) c P-O° (M, £).
We will denote by I I I I I a CI-norm on the space Pt (M, £) of CI-sections of a
-

vector bundle £ on M or on M x M.
Let £1 and £2 be vector bundles on M, and let pr1 and pre be the pro-
jections from M x M onto the first and second factor M respectively. We
denote by Si 9£2 the vector bundle
pri£, ®pri£2
over M x M.
We will call a section p E P(M x M, (.'F(g IAI1/2) ® (£* (9 IAI1/2)) a kernel.
Using p, we may define an operator
P : P--(M,£ (& IAI1/2) , P(M,F® IAI1/2)
by

(Ps)(x) =
f EM
p(x, y) s(y)
By this, we mean that (Ps)(x) is the pairing of the distribution in
r--(m, E ®IAI1/2) with the smooth function p(x, ) E P(M, £* (9 IAI1/2).

An operator with a smooth kernel will be called a smoothing operator.

Clearly, the composition of two operators P1i P2 with smooth kernels pl, P2
is the operator with smooth kernel

p(x, y) = f P1 (X, z)p2 (z, y);

zEM
P1 (X, z)p2(z, y) is a density in z, so can be integrated.
The following form of the Schwartz kernel theorem gives a characteri-
zation of operators with smooth kernels.
Proposition 2.14. The map which assigns to the kernel p(x, y) the corre-
sponding operators '-+ Ps is an isomorphism between the space of kernels
P(M x m, (.P (& IAI1/2) ® (£® (9 IAI1/2)) and the space of bounded linear op-
erators from P'O°(M, £ (& IAI"2) to r(M,.F (9 JAI 1/2).
We will sometimes use Dirac's notation for the kernel of an operator P:
if x and y lie in M, then the kernel of an operator P at (x, y) E M X M is
written (x I P I y), so that

(2.4) (PO) (X) = JEM (x I P I y)O(y)-

y
2.3. Heat Kernels 75

There are two quite different ways to investigate the heat kernel of H, that
is, the kernel attached to the operator e-tH (t > 0) by the kernel theorem:
(1) If H is symmetric, if we can show that the operator H has a self-
adjoint extension H that is bounded below, then by the spectral theorem,
the operator e-tH, which we will simply denote by a tH, is well defined
as a bounded operator on the Hilbert space of square integrable sections of
E ®I A 11/2. If, in addition, we can show that a tH extends to an operator
which maps distributional sections to smooth sections, then it follows by the
kernel theorem that the kernel of e-tH is smooth.
(2) The other method, which we will follow, starts by proving the existence
of a smooth kernel pt (x, y) which possesses the properties that the kernel
of the operator e-tH should possess, without first constructing the operator
H. The advantage of this method is that it gives more information about the
operator e-1H than the first method, is more closely related to the geometry of
H, is technically more elementary, and does not require that H be symmetric.
The following definition summarizes the properties that the kernel of an
operator e-tH must have.
Definition 2.15. Let H be a generalized Laplacian on E ® IAI1/2. A heat
kernel for H is a continuous section pt(x, y) of the bundle (Se JAI 1/2) Z (6*
IAI1/2) over 1[8+ x M x M, satisfying the following properties:
(1) pt(x, y) is C' with respect to t, that is, i9pt(x,y)/at is continuous in
(t, x, y),
(2) pt (x, y) is C2 with respect to x, that is, the partial derivatives
2ptX y)
axiaxi
are continuous in (t, x, y) for any coordinate system x;
(3) pt(x, y) satisfies the heat equation
(at+Hx)pt(x,y) = 0;
(4) pt (x, y) satisfies the boundary condition at t = 0: if s is a smooth section
of £ ® IAI1/2, then
lim Pts = s.
t--+o
Here we denote by Pt the operator defined by the kernel pt, and the limit
is meant in the uniform norm IIsIIo = supXEM IIs(x)II, for any metric on
E.

Our first task is to prove that the heat kernel is unique. We consider the
formal adjoint H* of H, which is a generalized Laplacian on E® ®IAI1/2
Lemma 2.16. Assume that the operator H* has a heat kernel, which we
denote by pt . If s(t, x) is a map from 1R+ to the space of sections of E ®IAI1/2
76 2. Asymptotic Expansion of the Heat Kernel

which is C1 in t and C2 in x, such that limt_.,o st = 0 and which satisfies

the heat equation (8t + H)st = 0, then st = 0.

Proof. If sl(t, ) and 82 (t, ) are C2 sections of £ ®JAI 112 for each t > 0, we
have
H*s2(t,
(Hsi(t,'), s2(t, .)) = (sl(t, .), -))-
fM JM

For any u in r(M, E* 0 JAI1/2), consider the function

AO) = f(x,y)EMxM
(s(O, x),pt-e(x, y)u(y))

where 0 < 0 < t. Differentiating with respect to 0 and using the heat equation
Definition 2.15 (3), we see that f (0) is constant. By considering the limits
when 0 tends to t and to 0, we obtain

(s (t, x), u(x)) = 0

LEM

for every uEr(M,£*(& IAI'/2),sothatst=0for allt>0.

Using this lemma, we obtain the following result.

Proposition 2.17. (1) Suppose there exists a heat kernel pt for the operator
H*. Then there is at most one heat kernel pt for H.
(2) Suppose there exist heat kernels pt for H*, and pt for H. Then pt(x, y) _
(pt (y, x))*.
(3) Suppose there exist heat kernels pt for H*, and pt for H. Then the
operators Pts(x) = fm pt (x, y)s(y) dy form a semi-group.

Proof. Consider f (0) = fM((Pes)(x), (Pt Bu)(x)) for 0 < 0 < t, where s and
u are smooth sections of r(M, £ (& IAI1/2); as before, this is independent of 0
and we obtain
(Pt s, u) = (s, Pt u)

by considering limits. This proves (1) and (2).

To show the semigroup property, we must show that ifs E r(M, £(&JAI 1/2)
and 0 > 0, then the unique solution st of the heat equation for H with
boundary condition limt_,o st ='Pes is given by Pt+os; but this follows from
the lemma.
2.4. Construction of the Heat Kernel 77

2.4. Construction of the Heat Kernel

On R', the scalar heat kernel is given by the exact formula
qt(x, y) = (47rt)-n12e-11=-y112/4t

For an arbitrary Riemannian manifold M and bundle £, it is usually impossi-

ble to find an exact expression for the heat kernel pt(x, y). However, for many
problems, the use of an approximate solution is sufficient. The true kernel
pt is then calculated from the approximate solution by a convergent iterative
procedure.
In order to describe this technique, let us consider the following finite
dimensional analogue of the construction of pt. Let V be a finite dimensional
vector space with a linear endomorphism H (we have in mind here though
that V = r(M, £®I11J1/2) and H = Os(&I°I1/2+F); we would like to construct
the family of operators Pt = e-'H
Suppose that we are given a function Kt : l[8+ -> End(V) which is an
approximate solution of the heat equation for small t in the sense that:

Rt = dKtt + HKt = O(ta), for some a > 0,

and such that K0 = 1; such a function is also called a parametrix for the heat
equation. The function Rt is called the remainder.
Definition 2.18. The k-simplex Lk is the following subset of Rk:
1}-

Often,
O f t e n , we will parametrize Ok by the coordinates

Q0=t1, Qi=ti+1-ti, Ok=1-tk,

which satisfy vo + - + ak = 1 and 0 < Qi 1. For t > 0, we will write tIk
for the resealed simplex
. < tk < t}.
t 1 -

Let
L e t vk be the volume of Ok for the Lebesgue measure dti ... dtk. Since
vo = 1 and
1 1
yk-1
Vk = I VOl(tkOk-1) dtk = tk-lvk-1 dtk = ,
0 J0
we see that Vk = 1/k!. The rapid decay of vk will be important in the following
proofs, when we must show that certain series converge.
Theorem 2.19. Let Qt : R+ -* End(V) be defined by the integral

Qt = f tOk
Kt-tkRtk-tk-1... Rte-t1Rt1 dtl ... dtk;
78 2. Asymptotic Expansion of the Heat Kernel

in particular, Q° = K. Then the sum of the convergent series k>0(-i)kQt

is equal to Pt = e-tH, and Pt = Kt + O(t'+a).
Proof. The fundamental theorem of calculus shows that the derivative of the
convolution f o a(t - s)b(s) ds with respect to t equals
t
d (t - s)b(s) ds + a(0)b(t).
J
If we apply this with a(s) = K9 and b(s) R(k)(s) defined by

R(k) (s) R9_tk-l ... "'L2-tl Rtl dt1... dtk-1,

90k-1
then since KO = 1, we obtain
(dt + H)Qt = R(k+l) (t) + R(k) (t)
Thus, the sum with respect to k telescopes:
00

(dt + H) 1:(-i)kQt = 0.
k=0
To estimate Pt - Kt for small t, we use the uniform bounds over t/k,
lKt-tk < Co and JRtj+1 -t: I < Cta,
which lead to the bound
iQt i < CoCktka tk , tk
since vol(tAk) = M.

k!
It is clear from this that the series defining Pt converges, and that Pt =
Kt + O(tl+a). 0
An important special case of this formula is the Volterra series for the
exponential of a perturbed operator: if H = Ho + H1 E End(V) and Kt =
e-tHo, we see that Rt = (d/dt + H)e-'HO = HletHo, and we obtain
00
e-t(Ho+Hl) = e-tHo + E(-i)kjk,
k=1
where
Ik = f e-(t-tk)H0H1e-(tk-tk-1)H0...Hle-t1HO dt1...dtk.
Ok
This formula implies in particular that

e-t(Ho+HI) = (-t)k f e-aotHoHle-vitH0 ...H1e-oktH° do1...dUk

k =O

(2.5) = e-tHo - t f 1 e-atHoHie-(1-a)tH0 do- + .. .

0
2.4. Construction of the Heat Kernel 79

If HZ is a one-parameter smooth family in End(V), by setting Ho = Hz and

H, = edH/dz we obtain the following formula for the derivative of a tH-:
t
(2.6) d (e-tH=) = e-(t-s)Ha dHz asHxds.
dz - fo dz
We now turn to implementing this method in the case which interests
us, namely V = r(M, E (9 IAI1/2) and H = !1£0111 2 + F. In formulas
which follow, we will omit the notation dt, ... dtk for the Lebesgue measure
on tOk, since this is the only measure which we will consider on this space.
Imitating the above method in an infinite dimensional setting, the steps in the
construction of the heat kernel are:
(1) Construction of an approximate solution kt(x, y), and study of the
remainder rt(x, y) = (at + H,,)kt(x, y);
(2) Proof of convergence of the series
00
E(-1)k kt-tk (x, zk)rtk-t,_1(zk, zk-1) ... rt1(zi, y)
k=0 f kfM k

t o the heat kernel pt (x, y).

The construction of kt will be performed in the next section, by solving
the heat equation in a formal power series in t. For the moment, we will
only use the properties of the approximate solution kt(x, y) summarized in
the following theorem, which will be proved in the next section.
Theorem 2.20. For every positive integer N, there exists a smooth one-
parameter family of smooth kernels km (x, y) such that, for every integer 2
(1) for every T > 0, the corresponding operators Kt form a uniformly bounded
family of operators on the space re(M, E ®IAI1/2) for 0 < t < T;
(2) for every s E r (M, c (& IAI1/2), we have limt_,o Kts = s with respect to
the norm 11 IIR

(3) the kernel rt(x, y) = (at + (x, y) satisfies the estimate

Ci(t)t(N-n/2)-e/2
lIrtlle <
We fix N and write kt for kN. For such a kernel kt, we would like to
consider the operator

h%t = I
k Kt-tkRtk-tk-1 ...Rtl
defined by the kernel

qt (x, y) = JA k 1Mk
We will first prove that if N is chosen large enough, this integral is convergent
and that the kernel qt is differentiable to some order depending on N. To do
this, we need some estimates.
80 2. Asymptotic Expansion of the Heat Kernel

Consider the kernel

rt +1 (x, y)
= Okf fMk rt-tk (x, zk)rk-tk-i (zk, zk-1) - rti (zl, y)

Lemma 2.21. If N satisfies N > (n + £)/2, then rt +1 is of class C1 with

respect to x and y, and
IIrt+1
It < Ck+1 t(k+1)(N-n/2)-t/2 Vol(M)ktk/k!
Proof. If N > (n + 1) /2, the section (x, y) ---> rt(x, y) and its derivatives up
to order 2 extend continuously to t = 0, so that the integral defining rt+1 is
convergent. Now, the uniform estimates of the integrand over tLk lead to the
result, the factor tk/k! being equal to the volume of tLk.
We now turn to the estimation of qt (x, y) and its derivatives.
Lemma 2.22. Assume that N > (n + 2)/2, and that I > 1.
(1) The kernel q (x, y) is of class Ce with respect to x and y, and there exists
a constant C such that
Vol(M)k-1tk(N-n/2)-e/2tk/(k
llgt lie < CCk - 1)!
for every k > 1. l
(2) The kernel qt (x, y) is of class C' with respect to t, and
(at + HH)gt (x, y) = rt+1(x y) + rt (x, y)
Proof. We write the kernel qt in the following form:
t
qt (x, y) = f
o
(f EM
kt-s(x, z)rs (z, y)) ds.

Since composition with the kernel ks defines a uniformly bounded family of

operators on re(E (& IAI1/2), the section

b(t, s, x, y) = J EM kt-s (x, z)r(z, y)

depends continuously on s E [0, t], as do its derivatives up to order 2 with
respect to x and y. Furthermore, b(t, t, x, y) = rt (x, y); also

(at + H.) b(t, s, x, y) = f rt-s (x, z)r9 (z, y) = rk+1(x, y),

EM
so that b(t, s, x, y) is differentiable with respect to t, and atb(t, s, x, y) is a
continuous function of s E [0, t]. The proof of (2) now follows by differentiating
qt (x, y) = fo b(t, s, x, y)ds with respect to t.
To prove (1), we use the uniform boundedness of Kt, which shows that for
0<s<t,
Ilb(t,s)Ile <C'Ilr, Ile.
2.5. The Formal Solution 81

Combining this with Lemma 2.21, we obtain

IIb(t, s) lie <
C'Cktk(N-n/2)-e/2
vol( - 1)!
from which (1) follows.
The following theorem shows that we can construct a heat kernel pt starting
from the approximate kernel kt.
Theorem 2.23. Assume that the kernel kt' (x, y) satisfies the conditions of
Theorem 2.20 with N > n/2 + I.
(1) For any 2 such that N > (n+.£+ 1)/2 the series
00

pt(x, y) = E(-1)kgt (x, y)

k=0
converges in the II Ile+1-norm (over M x M) and defines a C'-map from
R+ to re(M x M, (£ ®IAI1/2) ® (E® ®JAl1/2)) which satisfies the heat
equation
(at + HH)pt(x,y) = 0.
(2) The kernel kt approximates pt in the sense that
O(t(N-n/2)-k-a/2+1) when t --> 0.
Ilat (pt - kN) II e =
(3) The kernel pt is a heat kernel for the operator H.
Proof. Our estimate on Ilgt (x, y)Ile+1 proves the absolute convergence of the
series defining pt(x, y) in the Banach space re+l, uniformly with respect to
t. From the equation atgt = rt +1 + rt - H.,qt , we obtain bounds for the
derivative with respect to t in the Ce norm. This shows the convergence of
the series E(-1)katgt in re, uniformly with respect to t, and the equation
(at + HH)pt(x, y) = 0 follows from the same telescoping as in Theorem 2.19.
To show (3), we must show that pt satisfies the initial condition of a heat
kernel. However, we know that kt satisfies this initial condition, so that (3)
follows from (2).

2.5. The Formal Solution

AE191A1112
Let H = + F be a generalized Laplacian on the bundle E ® IAI112
over a compact manifold M. In this section, we will denote the connection VE
on E simply by V. We denote by Idx11/2 the canonical half-density associated
to the Riemannian metric on M. To construct an approximate solution to the
heat equation for H, we start by finding a formal solution to the heat equation
as a series of the form

kt(x,y) = qt(x, y) E00ttDz(x, y, H) Idyll/2,

i=O
82 2. Asymptotic Expansion of the Heat Kernel

where the coefficients (x, y) ' -r Dt (x, y, H) are smooth sections of the bundle
E®E* defined in a neighbourhood of the diagonal of M x M and qt is modelled
on the Euclidean heat kernel: in normal coordinates around y (x = expy x),

(47rt)-n/2e-IIXII2/4tldxll/2
qt(x, y) =
= (47rt)-n/2e-d(z,11)2/4t j (X)-1/2Idx I1/2.

(The precise definition of a formal solution is given below). We fix y E M and

write qt for the section x qt (x, y) of the bundle IAl1/2. Abusing notation
once more, we denote by j1/2 the function on M around y given by j112(x)
j(x)1/2 if x=expyx.
To show the existence of a formal solution of the heat equation, we will
make use of the following formula. Here, we write H for the operator s H
ldxl -1/2 H - s dxl1/2 on F(M, E).

Proposition 2.24. Let B be the operator acting on sections of E in a neigh-

bourhood of y defined by the formula

B= j1/2 o H o j-1/2.
Then, for any time dependent section st of E,
+t-1vjz
(at + H)stqt = ((at + B)st)qt.

Proof. Let t st be a smooth map from R+ to the space of smooth sections

of E around y. By Leibniz's rule, we see that

(at + H) (stgt) = ((at + H)st)gt - 2(Vst, Vqt) + st ((at + A)qt),

where A is the scalar Laplacian. Proposition 2.11 gives explicit formulas for
the last two terms:

-2Vgt = (t-1R + d(logj))gt,

(at + A) qt = j112(0.7-1/2)qt

From this, we see that

t-1VR
(at + H)(stqt) = ((at + H + + Vdlog j +j1/2(Aj-1/2))st) qt-

By Leibniz's rule again, we obtain

j1/2 o H o j-1/2 = H + Vdlogj +j1/2(Aj-1/2)

Proposition 2.24 makes clear what we mean by a formal solution of the heat
equation.
2.5. The Formal Solution 83

Definition 2.25. Let -1)t (x, y) be a formal power series in t whose coefficients
are smooth sections of £ E E* defined in a neighbourhood of the diagonal of
M x M. We will say that qt (x, y)dt(x, y)Idyj1/2 is a formal solution of the
heat equation around y if x H-* (Dt(x, y), considered as a section of the bundle
E ®£y over a neighbourhood of y in M, satisfies the equation
(at + t-1 V7z + B)(Dt(., y) = 0.
We can now prove the existence and uniqueness of the formal solution of
the heat equation. Let x and y be sufficiently close points in M, and write
x = expy x for x E TyM. Denote by -r (x, y) : Ey --+ Ex the parallel transport
along the geodesic curve x3 = expy sx : [0,1] --> M.
Theorem 2.26. There exists a unique formal solution kt(x, y) of the heat
equation
(at + H.)kt(x,y) = 0
of the form
00

kt(x,y) = gt(x,y) tz1bi(x, y, H) Idyl1/2,

i=O
such that -Do(y, y, H) = IE1, E End(Ey). We have the following explicit for-
mula for (Di:

fo
In particular, I)o(x, y) = 7-(x, y).
Proof. In the left side of the equation
00

(0 t +t-1V7z + Bx) Eti1)i(x,y,H) = 0,

i=0
we set the coefficients of each ti equal to zero. This gives the system of
equations:
V7Z.(I)o = 0
(V 7z + i)(Di = -B,4Di_1 if i > 0.
Moreover, we know that Do (y, y, H) = IEar E End(EE), and that 4pi(x, y, H)
should be continuous at x = y.
The parallel transport 7-(x, y) : Ey --> Ex along the geodesic x3 = expy sx :
[0, 1] -+ M satisfies the differential equation V7z7 = 0 with boundary condi-
tion y) = 1, and we see that Do (x, y, H) = -r (x, y).
To obtain a formula for (Di in terms of 1i_1i we consider the function
qi(s) = s''(Di(x,,, y, H)
where xs = expy sx. Then Oi(0) 0 and Vd/dsti = -s'-'(B. 11i-1)(xa,y),
and we obtain the required explicit formula for 15j.
84 2. Asymptotic Expansion of the Heat Kernel

Example 2.27. We can use the explicit iterative formula for (Di(x, y) given
above to calculate the subleading term 4P1(x, x) in the asymptotic expansion
of kt(x, y) on the diagonal. We have
(bi(x,x) =
However, since Do(x, y) = T(x, y), we /see that

-jl(x,x) = +F)i 1/2)(x,x).

Using Proposition 1.28, we see easily that this function equals 3rM(x) -F(x),
where rM(x) is the scalar curvature of M.
Consider the Ey-valued function on TyM defined by
Ui(x, y, H) = r(expy x, y)-l,Di(expy x, y, H) x E TTM.
Let L be the differential operator which acts on the space of S. -valued func-
tions on M around y, defined by the formula
y)-1Bx(r(x,
(LO) (x) = r(x, y)V'(x))
Then we see that
1

Ui(x, y, H) fo S'-1(LUi-1)(x3, y)ds,

where xs = expy sx. It is clear from Propositions 1.18 and 1.28 that the
coefficients Ui,a(y, H) in the Taylor expansion

Ui,a (y, H)x

of the function x ' Ui (x, y, H) are polynomials in the covariant derivatives

at y of the Riemannian curvature, the curvature of the bundle E, and the
potential F, since this is true of the Taylor coefficients of the coefficients of
the operator L at y.
Definition 2.28. If : R+ -> [0,1] is a smooth function such that
hi(s) = 1, if s < s2/4,
Vi(s) = 0, ifs > 62 ,
we will call 0 a cut-off function.
We will now use the formal solution to the heat equation to construct an
approximate solution kN (x, y) satisfying the properties given in Theorem 2.20.
To do this, we truncate the formal series and extend it from the diagonal to
all of M x M by means of a cut-off function &(d(x, y)2) along the diagonal,
where the small constant a is chosen smaller than the injectivity radius of
2.5. The Formal Solution 85

the manifold; thus, the exponential map is a diffeomorphism for lxi < E. We
define the approxiamte solution by the formula
N
(2.7) kt'(x,y) =
i=O
In the next theorem, we will prove that kN possesses the properties that
were used to construct the true heat kernel in Section 3. Indeed, we will see
that it satisfies somewhat stronger estimates than were stated in Theorem 2.20.
Theorem 2.29. Let £ be an even positive integer.
(1) For any T > 0 the kernels kt', 0 < t < T, define a uniformly bounded
family of operators Kt' on rt (M, £ ®IAI1/2), and

tl KNs-s1le=0.

(2) There exist differential operators Dk of order less than or equal to 2k such

j
that Do is the identity and such that for any s E Ft+1(M, £ (D I Al 1/2),
Q/2-9
`K NS _ ` tkDkS I2= O(t(R+I)/2-.7).
II t l
k=0

(3) The kernel rN(x, y) = (8t +H.)kN(x, y) satisfies the estimates

Ct(N-n/2)-k-2/2
Ilat rN IIk <
for some constant C depending on 2 and k.
Proof. We fix N and write Kt for the operator corresponding to the kernel
kN. In a neighbourhood of the diagonal, we write y = expx y, with y E TIM,
and identify £ to S. by parallel transport along the unique geodesic joining
x and y. If s is a section of £ ® IAI1/2, we denote by s(x, y) the function of
y such that s(y) = s(x,y)Idy11/2. We see that
N
(Kts)(x) = (flirt)-"'/2
f AM
e-IIYII2/4t Etityi(x,Y)s(x,Y)Idyl ®Idxl1/2,
i=O

with smooth compactly supported coefficients

(x, Y) Wi (x, Y) E End(£x)
Furthermore, To (x, 0) = Iss.
The statement of the theorem is local in x, so we can assume that the vector
spaces £., and TAM are fixed vector spaces E and V. The change of variables
y = t1/2v on V shows that (Kts)(x) equals
N
(47r)-n/2 e-IIv112/4 tiT,(x,tl/2v)s(x,t1/2v)dv ® Idxl1/2,
V i=O

from which (1) follows.

86 2. Asymptotic Expansion of the Heat Kernel

By Taylor's Theorem, there exist differential operators Dk of order less

than 2k such that
f/2
Kt'S - E tkDkS 1 0
= O(tP/2+1).

k=0

A similar argument may be used to estimate the C2j -norm of

t/2-j
KNS - tkDkS.
k=0

rt T
The bound on (x, y) follows from the formula for by by means of Propo-
sition 2.24, which shows that
N
rN(x, y) _ [(at + t-' Viz + Bx) - (O(d(x, Y)2) E tZ,Dj (x, y, H))] qt (x, y).
i=o

The right-hand side in this formula is made up of two kinds of terms, those
which involve at least one differentiation of the factor '(d(x, y)2), and those
which only involve time derivatives or derivatives of the functions 'Di. The
first type vanishes for d(x, y) < E/2 and so is very easy to estimate, since for
d(x,y) > e/2
y)2)e-d(x,v)2/4t = O(tk)
a° O(d(x,

for any k. From the system of differential equations defining the functions I)i,
we see that the terms which do not involve a derivative of b(d(x, y)2) cancel,
except for one remaining term equal to

tN4t (x, y) N) (x, y),

which may be bounded by tN-n12. The argument used to bound the deriva-
tives of rt(x, y) is similar, once one observes that
ate-x2/t t-1(x2/t)e-x2/t

= = O(t-1),
t-1/2(-2x/t1/2)e-x2/t

axe-x2/t = = O(t-1/2).

As a consequence of Theorems 2.29 and 2.23, we have now proved the

existence of a Ce-heat kernel, for any operator H = De®IAI1/2 + F and any
2. From the unicity of the heat kernel (Proposition 2.17) it follows that the
heat kernel pt (x, y, H) is smooth as a function of (x, y). The heat equation
(at + HH)pt(x, y) = 0 implies of course that it is smooth as a function of
(t, x, y).
Let us summarize the properties of the heat kernel that will be used in the
proof of the local index theorem.
2.6. The Trace of the Heat Kernel 87

Theorem 2.30. Let pt (x, y, H) be the heat kernel of the generalized Lapla-
Ae01°I112 +F.
cian H = There exist smooth sections -Di E r(M x M, E ®E*)
such that, for every N > n = dim(M)/2, the kernel kN(x, y, H) defined by
the formula
N
(47rt)-"/2e-d(x,y)2/4t j(X,y)-1/2JdxI 1/2 (9 Idyl 1,2
h(d(x, y)2) E t2<Di(x, y, H)
i=0
is asymptotic to pt (x, y, H):
II at (pt(x, y, H) - kN (x, y, H)) II$ = O(tN-n/2-a/2-k).
The leading term 4o (x, y, H) is equal to the parallel transport T(x, y) : y
£5 with respect to the connection associated to H along the unique geodesic
joining x and y.
If s E r(M, £ (9 All/2) is a smooth section of E ®IAI1/2, we see from
Theorem 2.29 that Pts, defined by

(Pts)(x) = f pt(x,y,H)s(y)
EM

has an asymptotic expansion in r(M, E (&IAII/2) with respect to t of the form

E tkDks. The heat equation (a/8t+H)Pts = 0 implies that Dk = (-H)k/k!.
Thus, we obtain the following estimate:
(_tH)2
(2.8) IPts - > sI = O(tk+l).
'f
i=0
This asymptotic expansion justifies writing a-tH for the operator Pt.

2.6. The Trace of the Heat Kernel

In this section, we will derive some of the consequences of the existence of a
heat kernel for an operator of the form H = AE19IA!112 + F. In particular, we
show that if H is symmetric, then it has a unique self-adjoint extension H in
the space of square-integrable sections of the bundle £ ® IAI1/2, and the kernel
of the operator a-tH is the heat kernel pt(x, y, H) of H.
We first recall some elementary facts on trace class operators on a Hilbert
space. Let 7-l be a Hilbert space with orthonormal basis ei. An operator A is
a Hilbert-Schmidt operator if
IIAI1as = IIAeiII2 = I(Aei, e,) I2
i ii
is finite. The number IIAIIxs is called the Hilbert-Schmidt norm of A. If A is
a Hilbert-Schmidt operator, so is its adjoint A* and IIAIIHS = IIA*Ilas; also,
88 2. Asymptotic Expansion of the Heat Kernel

if U is a bounded operator on N and A is an Hilbert-Schmidt operator, then

UA and AU are Hilbert-Schmidt operators and IIUAIlHS < IIUII IIAIIHS If A
and B are Hilbert-Schmidt operators, then for any orthonormal basis ei of H,
E 1(ABei, ei)I IIAIIHSIIBIIHS-
i
Let M be a compact manifold and let £ be a Hermitian vector bundle on
M, and denote by rL2 (M,£ (&IAI1/2) the Hilbert space of square-integrable
sections of £ ® IAI 1/2. We will also frequently consider the Hilbert space
rL2 (M, £) where M is a compact manifold with canonical smooth positive
density (for example, a Riemannian manifold with the Riemannian density).
Lemma 2.31. The topological vector space underlying rL2 (M, £ (&IAI1/2) is
independent of the metric on £, although the metric on rL2(M,£ (gIAI1f2)
depends of course on the metric on S. Similarly, the topological vector space
underlying rL2 (M, £) is independent of the metric on £ and the density on
M.
Proof. If hl(., ) and are two metrics on £, then there is an invertible
section A E r(M, End(S)) such that hl (s, s) = h2 (As, As) for all s E r(M, £).
Since A and A'1 induce bounded operators on rL2 (M, £ (9 IAI1/2), the first
part of the lemma follows.
If w1 and w2 are two smooth positive densities on M, then there is an
everywhere positive function 0 E C°° (M) such that w1 = iw2. Since mul-
tiplication by 01/2 and 0-1/2 induce bounded operators on FL2 (M,.6), the
independence of this space on the density on M follows.
An operator K with square-integrable kernel
k(x, y) E rL2 (M x M, (E (& JAI 1/2) Z (6 (& JAI
1/2))

is Hilbert-Schmidt, and

(2.9) IIK11HS = fx,y)EMxM

as follows from the definition of the Hilbert-Schmidt norm,

IIK112. = 1(Kei, e,)12.

ii

In particular, we see that the heat kernel pt of a generalized Laplacian H is

the kernel of a Hilbert-Schmidt operator for all t > 0.
An operator K is said to be trace-class if it has the form AB, where
A and B are Hilbert-Schmidt. For such an operator, the sum Ei(Kei, ei)
is absolutely summable, and the number >i(Kei,ei) is independent of the
Hilbert basis; it is called the trace of K:
TI (K) = E(Kei, ei).
2.6. The Trace of the Heat Kernel 89

Note that if A and B are Hilbert-Schmidt, then Tr(AB*) = (A, B)HS; also
(2.10) Tr(AB) = TY(BA) _ (Ae;, ej) (Bej, e2).
aj

Observe that the restriction of the bundle (S ®IAII/2) ® (£ 0 IAII/2) on

M x M to the diagonal is isomorphic to End(S) o JAI. If
a(x, y) E r(M x M, (S (9 JA 11/2) CK (g (& JAI 1/2))

is a smooth kernel on M x M, denote by Tr(a(x, x)) the section of I A I obtained

by taking the pointwise trace of a(x, x) E r(M, End(S) (9 IAI) acting on the
bundle E.
Proposition 2.32. For all t > 0, the operator Pt with kernel pt the heat
kernel of a generalized Laplacian on a compact manifold M is trace class,
with trace given by the formula

` r(Pt) = J Tr(pt(x, x))-

EM

Proof. This follows from the fact that Pt = (Pt/2)2, and the fact that Pt/2 is
Hilbert-Schmidt. The trace of Pt is given by the formula
Tr(Pt) = (Pt/2, Pt1 2)HS

Ix,v) EM2
pt/2(x,y)pt/2(y,x)

pt(x, x). O
f.EM
Let H = + F be a generalized Laplacian on £, and consider the
Ae®1AI112

operator H as an unbounded operator on rL2(M,£ (&IAI1/2) with domain

equal to the space of smooth sections of £ ® IAI1/2; its formal adjoint H* will
also be considered on the same domain.
Proposition 2.33. The closure of H* is equal to the adjoint of the operator
H with domain r(M, £ (9 In particular, if H is symmetric, then it
IAI1/2).

is essentially self-adjoint; in other words, H is a self-adjoint extension of H,

and is the only one.
Proof. Consider the action of the formal adjoint H* on the space of distribu-
tional sections of £ ® IAI1/2. It is clear that the domain of the adjoint of H
is
D = {s E rL2 (M, £ 0 IAI1/2) I H*s E rL2 (M, E (9 JAI 1/2)1,
where H*s is defined in the distributional sense. Thus, we need to prove that
for s E D, there is a sequence s, of smooth sections of £ ®I Al 1/2 such that s,,
converges to s and H*s, converges to H*s in rL2 (M, £ (9 IAI1/2). We will do
this using the operators Pt with kernel pt(x, y, H); first, we need two lemmas.
90 2. Asymptotic Expansion of the Heat Kernel

Lemma 2.34. If S E rL2 (M, £ ®IAI1/2), then limt-...,oPts = s in the L2-

norm.
Proof. If 0 is a continuous section of £ ®(AI1/2, we know that Ptcb converges
to 0 in the uniform norm, and hence in the L2-norm. Writing s = (s - 0) +
we see that
IIPts-slIL2 <- IIPt(s-0)IIL2+IIPtO-0IIL2+IIs-OIIL2.
Thus, it suffices to establish that the operators Pt are uniformly bounded on
rL2 (M, £ (9IAI1/2) for small t.
Theorem 2.30 shows that it is enough to prove the lemma for the operator
Qt associated to the kernel qt(x, y)Idx11/2 0 Idyll/2 with
(47rt)-n/2e-d(zv)2/4tV) (d(x,y)2)IdxI 1/2 0 Idyl 1/2.
gt(x,y) =
There exists a constant C such that fm Igt(x, y) I dy < C for t > 0. By
Schwarz's inequality, we see that if f E rL2 (M,.6 ®IAI1/2),

f I J qt(x, y) f (y) dy12 dx < J(fqt(x , y) dy) (f qt (x, y)I f(Y) 12 dy) dx
< C f fqt(x,Y)lf(y)l2dxdY < C2IIf1I2

Lemma 2.35. If s E V and Pt is the operator with kernel pt (x, y, H*) _

pt(y, x, H)*, then H*Pt s = Pt H*s.
Proof. If s E I'(M,£ (& IAI1/2), we have HPts = PtHs, since both sides
satisfy the heat equation and Pts - s converges uniformly to 0 as well as all
its derivatives, when t --f 0. Taking adjoints, we see that Pt H* = H*Pt on
r-°°(M, £ (g IAI1/2), and in particular on D.
Given s, we may take the sections s,,, to be given by Pi/ns. As n -> oo,
the first lemma shows that sn --> S in rL2, while the second shows that
H*s,ti = P* H*s, which again converges to H*s by the first lemma.
Applying this result with H replaced by its formal adjoint H*, we see
that the closure of H with domain r(M, £ (& IAI1/2) is the adjoint of H*. In
particular, if H is symmetric, that is H = H*, then H is self-adjoint, and we
obtain Proposition 2.33.
From now on, we consider only the case in which H is symmetric. If we
simultaneously diagonalize the trace-class self-adjoint operators Pt, we obtain
a Hilbert basis of eigensections 0i; the semigroup property for Pt now implies
that for some Ai E R
Ptoi = e-ta;oi
Since Ptq5i is a smooth section of £ 0 AII/2, we see that the eigensections Oi
are smooth. The eigenvalues X,, are bounded below, since the operators Pt are
bounded. The fact that the function Ptoi satisfies the heat equation implies
that H /. = AiOi. Furthermore, since the operators Pt are trace-class, we see
2.6. The Trace of the Heat Kernel 91

that >z e-tAl is finite for each t > 0. This shows that the spectrum of H is
discrete with finite multiplicity. Thus, we have proved the following result:
Proposition 2.36. If His a symmetric generalized Laplacian, then the op-
erator Pt is equal to e-tH. The operator H has discrete spectrum, bounded
below, and each eigenspace is finite dimensional, with eigenvectors given by
smooth sections of Eo JAI 1/2.
We will not be very careful about distinguishing between the operator H
and its closure H, and will write e-1H for the operator pt = e tH. From the
spectral decomposition of H it follows that
e-tAj
Pt (x, y) = E 0j (x) ®0j (y)-

Proposition 2.37. If H is a generalized Laplacian, and if P(0,) is the pro-

jection onto the eigenfunctions of H with positive eigenvalue, then for each
2 E N, there exists a constant C(2) > 0 such that for t sufficiently large,
C(2)e-tAl/2,
II(x I P(o,.)e-t'P(o,oo) 10111 <
where Al is the smallest non-zero eigenvalue of H.
Proof. For t large, we may write
(x I P(o,ac)e-tHP(o,.)
I y)

k1/2(x,z1)(z1 I P(o,oo)e-(t-1)Hp(o,oo) I z2)k1/2(z2,y)-

(z1,z2)EM2
The kernel k1/2 (x, y) is smooth, so we only have to show that

'Al.
I(x 11'(o,oo)e-(t-1)HP(0,.) I y)I2 < Ce
J(.,V)EM2
However, the left-hand side is nothing but the square of the Hilbert-Schmidt
norm of P(o,ac)e-(t-1)HP(0oo). Thus, if 0 < Al < A2 < ... are the positive
eigenvalues of H, repeated according to their multiplicities, we must show
that
E e-2(t-1)Aj
< Ce"1.
j>1
Thus, we see for t large that
e-2(t-1)Aj < e-(t-2),\j ) . sup e-tAj
j>1
j>1 j>1
< e-(t-2)A1) . e-tai < e-Aj) _ e-tA1
j>1 j
- (f.EM x I
P(o,oo)e-HP(o,oo) I x)) e-tAi. 0
92 2. Asymptotic Expansion of the Heat Kernel

In particular, if H is positive, that is, if it is non-negative and has vanishing

kernel, then P(o,00) is the identity, and each e E N, there exists a constant
C(e) > 0 such that for t sufficiently large,

Ilpt(x,y)IIe < C(e)e-tai/2,

where Al is the smallest eigenvalue of H.

The Green operator G of a positive generalized Laplacian H is by def-
inition the inverse of H on L L2 (M, £ ®IAI1/2). Clearly G is bounded on
F L2 (M, £ ®JAI 1/2). In the following theorem, we discuss the properties of G.

Theorem 2.38. Assume that H is a positive generalized Laplacian. Let G

be the inverse of H on FL2 (M, IF (9 IAI1/2). Then we have the integral repre-
sentation
Gk = (k -1 1)! e-tHtk-1 dt

for each natural number k. The operator Gk has a CL kernel for k > 1 +
(dim(M) + 1)/2.

Proof. If k > 1, define the operator Gk by the integral

00
Gk = 1 e-tHtk-1 dt.
(k - 1)! Jo
If 0 is an eigenfunction of H with eigenvalue A, it is easy to see that the
operator Gk applied to 0 equals

0" e-tatk-1 dt) o= A.

- 1).! (f
1
(k o

We will show that for k sufficiently large, the integral

9k (x, y) = (k 1 fpt(x,y)tk_ldt
1)!

converges and defines a kernel which is in fact the kernel of the operator Gk.
We start by splitting the integral into two pieces,

9k(x, y) _
(k
1 1)1 f 1

pt(x, y)tk-1 dt +
(k
1
00
1)! f pt(x, y)tk-1 dt.
We bound the first integral by means of the following lemma.

Lemma 2.39. The family of kernels tk-lpt(x, y), (t < 1), is uniformly
bounded in the Ce-norm when k > 1 + (e + n)/2.
2.6. The Trace of the Heat Kernel 93

Proof. Using the asymptotic expansion for the heat kernel pt (x, y), we see
that it suffices to prove the lemma for the kernel
y)2)t- dim(M)/2e- d(x V)2/4t
qt(x, y) = '5(d(x,
where V) is a smooth function on R+ which equals 1 near zero and 0 in
[e2, oo). Using the fact that 10ae-d(x y)2 /4t[ < Ct-1'!/2, it is easily seen
that for k > 1 + (2 + n)/2, the family of kernels tk-1a"qt(x, y) is uniformly
continuous for0<t<1, a<t.
We will complete the proof by showing that the second integral converges
to a smooth kernel, using Proposition 2.37. Indeed, for any Ce-norm,
C(2) e-tai/2 tk-1 dt
1 IIpt(x, y) Ie tk-1 dt <
(k -1)! floc, (k -1)! 100
which is finite.
In Section 9.6, we will extend this theorem to fractional powers of the Green
operator.
It is straightforward to extend this theorem to an arbitrary self-adjoint
generalized Laplacian H. We define the Green operator of H to be the inverse
of H on ker(H)1, and to vanish on ker(H). The integral representation of
G is similar to the above one, except that we must handle the non-positive
eigenvalues separately: if )L. m < .. < \_1 < 0 are the strictly negative
eigenvalues of H, and if Pa, is the projection onto the eigenspace of H with
eigenvalue \i, then we see that
00
(2.11) Gk = 1 P(o,00)e-txtk-1 dt + J:(),_i)-kPac,
7n
(k -1)! fo i=1
Since the operators P_,\ are smoothing operators, the estimates on the kernel
of Gk may be carried out in exactly the same fashion as when H is positive.
The above theorem has the following important consequence, known as
elliptic regularity.
Corollary 2.40. Let H be a symmetric generalized Laplacian on,6 (9 JAI 1/2.
(1) Ifs E FL2 (M, £ (9 lAI1/2) is such that Hks, in the sense of generalized
sections, is square-integrable for all k, then s is smooth.
(2) If A is a bounded operator on rL2 (M, £ (9 JAI1/2) which commutes with
H, and ifs E I'(M, £ ®IAI1/2) is smooth, then As E P(M, £ (& JAI'/') is
smooth.
Proof. We may assume that H is positive by adding a large positive constant
to H. It follows that 1 = Gk o Hk = Hk o Gk.
Since Gk commutes with H, we see that s = GkHks. By hypothesis,
Hks is square-integrable, hence by the above estimates, GkHks lies in Ce for
2 < 2k - n - 1. Letting k tend to oo, we obtain the first part.
94 2. Asymptotic Expansion of the Heat Kernel

The second part is proved in an analogous way, using the formula

As = (Gk o Hk o A)s = (Gk o A o Hk)s.
If s E r(M, £®IAI1/2) is smooth, then so is Hks; in particular, Hks is square-
integrable. Since A is bounded on the Hilbert space rL2 (M, £ (& JAI 1/2), we
see that AHks E rL2(M,£® 1A11/2). It follows that As = Gk(AHks) is CI
for £ < 2k - dim(M) - 1. Taking k large, we obtain the result.
It follows from Theorem 2.30 and the asymptotic expansion of pt (x, y) that
the trace of the heat operator a-tH possesses an asymptotic expansion in a
Laurent series in t1/2 for small t:
00
Tr(e-tH) _ Tr(pt(x, x)) ^'
(47rt)_n/2 to
x, H)) dx.
M i-o JM
In particular, using the fact that 45o(x,x,H) is equal to IE, we obtain Weyl's
Theorem.

Theorem 2.41 (Weyl). Let \i be the eigenvalues of H, each counted with

multiplicities. Then as t ---; 0, we have the asymptotic formula for the Laplace
transform of the spectral measure of H:

Ei (47rt)-1z/2
rk(£) vol(M) + 0(t-n/2+1)

Note that, at least in principle, it is possible to compute all the coefficients

(Di(x, x, H) in the asymptotic expansion of the heat kernel pt(x, y) explicitly,
so that we can, if we desire, obtain a full asymptotic expansion for the Laplace
transform of the spectral measure of H.
It is possible to derive from this theorem information about the asymp-
totic distribution of the eigenvalues of H at infinity, by means of Karamata's
Theorem.

Theorem 2.42 (Karamata). Let dp(i) be a positive measure on IR+ such

that the integral

f0 00 e"
converges for t > 0, and such that
00
lim t" dp(A) = C
t-40 Jo
for some positive constants a and C. If f (x) is a continuous function on the
unit interval [0, 1], then
00 00
t ota f f(e-t),) e-t' dµ()) =rya) f .f (e-t)ta-ie-t dt.
2.6. The Trace of the Heat Kernel 95

Proof. By Weierstrass's Theorem, we can approximate the function f (x) ar-

bitrarily closely by a polynomial p:
If(x) - p(x) < s for x E [0,1].
It follows that it suffices to prove the limit formula for polynomials on the
unit interval, and, in fact, for the monomial f (x) = xk. But this is a straight-
forward calculation: on the one hand,
00
lim t" e-(k+1)ta dy(A) = C(k + 1)-",
t-->o J0
while on the other,
C e ktt"-le-t dt
r(a) Jo
also equals C(k + 1)-01.
Using Karamata's Theorem we obtain the following restatement of Weyl's
theorem.
Corollary 2.43. If N(A) is the number of eigenvalues of H that are less
than A, then
(A) vol(M) \n/2
N 'k(E)
(47r)n/2r(n/2 + 1)
Proof. Taking in Karamata's Theorem a decreasing sequence of continuous
functions converging to the function equal to x-1 on the interval [1/e, 1] and
zero on the interval [0,1/e), we see that
t1 C
lim t" J
t-,0 o
dµ(a)=
r(a)
f't'-ldt= r(a C+ 1)
The corollary follows by considering a constant a such that H + a is positive,
and applying the above result to the spectral measure Ft = Ez 8a, of H + a,
for which a = n/2 and C = (47r)-n/2 rk(E) vol(M).
It is interesting to see how this works in the simplest possible case, where
M is equal to the circle of circumference 21r, and H is the scalar Laplacian
0. The eigenvalues of the Laplacian are A = 0, with multiplicity 1, and the
squares A = n2 (n > 1), with multiplicity 2. It follows that N(A) = [2v/A- +1].
On the other hand, Weyl's Theorem tells us that

N(A) N
vol(S')
4_rr(3/2)
r
and it is easily checked that the coefficient of f equals 2 (since vol(Sl) = 2ir
and r(3/2) = vr7r/2).
In the next proposition, we will extend the formula for the trace of the heat
kernel to any operator with smooth kernel. The proof is an application of the
Green operator.
96 2. Asymptotic Expansion of the Heat Kernel

Proposition 2.44. If A is an operator with smooth kernel

a(x, y) E r(M x M, (9 (9 JAI 1/2) N (S (& JAI 1/2)),
then A is trace class, with trace given by the formula

Tr(A) = Tr(a(x, x)).

EM
Proof. Let H be a positive generalized Laplacian on the bundle E, with Green
operator G. By Theorem 2.38, the operator Gk has continuous kernel for
k > 1 + (dim(M)/2), and hence is Hilbert-Schmidt.
The operator Ak associated to the smooth kernel H.,,,a(x, y) is Hilbert-
Schmidt, as is any operator with smooth kernel, and A = GkAk is the product
of two Hilbert-Schmidt operators and hence is trace class. The trace of A is
now easy to calculate:
Tr(A) = Tr(GkAk)

I x,y)EM2
gk (x, y)Hka(y, x)

a(x, x).
J.EM
Let D be a differential operator acting on r(M, £ (& IAI1/2). If K is an
operator with smooth kernel, then so are DK and KD, and hence they are
trace class. More precisely, the operator DK has smooth kernel Dxk(x, y)
while integration by parts shows that KD has smooth kernel Dyk(x, y).
Proposition 2.45. Let D be a differential operator acting on the bundle E.
If K is an operator with smooth kernel, then Tr(DK) = Tr(KD).
Proof. Choose a positive generalized Laplacian on the bundle £, and let G
be its Green operator. For sufficiently large N, the operator DGN has a
continuous kernel, and hence is a Hilbert-Schmidt operator. Thus,
Tr(DK) = Tt ((DGN)(HNK)) ='I4'((HNK)(DGN)) by (2.10)
= Tr((HNKD)GN) = Tr(GN(HNKD)) = L(KD).
Let us construct an asymptotic expansion for the kernel of the operator
DPt, where D is a differential operator.
Proposition 2.46. If Pt is the heat kernel of a generalized Laplacian H and
D is a differential operator of degree m on the bundle E, then there exist
smooth kernels Tj(x, y, H, D) such that the kernel (x I DPt I y) satisfies
N
n/2),
(x I DPt I y) - ht (x, y) t''W i(x, y, H, D) II = G(tN
i=-m
2.6. The Trace of the Heat Kernel 97

where ht (x, y) = (47rt)-n/2e-d(1,V)2/4t';(d(x, y)2)Idxll/2 ® Idyll/2, and 0 is a

cut-off function.
Proof. If we apply the differential operator a/axi to the Euclidean heat kernel
(47rt)-n/2e Ix-YI2/4t
gt(x,y) =
we obtain
xi - Yi
-qt (x, Y) 2t
By induction, we see that eX qt has the form

8Xgt(x,Y) = gt(x,Y)
(x- p
013,3 tIQI+j)
I0I+2.7<<I«I

The proof now follows by combining this with Theorem 2.30.

From the proof of Proposition 2.46, we see that the restriction of the kernel
of the operator DPt to the diagonal has the form

(47rt) -n/2 t"Qi(x, x, H, D) Idxl,

since the restriction of the coefficients Wi(x, V, H, D) to the diagonal vanishes

for i < -m/2. It follows that the trace Tr(DPt) has an asymptotic expansion
of the form
Tr(DPt) - (47rt) -n/2 tiai,
i=-[m/2]
where
ai=
fM
It is also possible to give an asymptotic expansion of the trace of the
operator KPt, if K is a smoothing operator on r(M, E (& JAI 1/2).
Proposition 2.47. There is an asymptotic expansion of the form

Tr(KPt) N Tr(K) + tiai.

i=1
Proof. Let k(x, y) Idxl1/2 ® Idyl 1/2 be the kernel of K, and let ky E r(M, £)
be the section x -+ k(x, y). Since Tr(KPt) = Tr(PK), we see that
Tr(KPt) = fM (Ptkx)(x) Idxl.

Applying the asymptotic expansion (2.8), we see that

Ptky -
(-t)iHi ky,
z,
i=0
from which the proposition follows.
98 2. Asymptotic Expansion of the Heat Kernel

2.7. Heat Kernels Depending on a Parameter

Let M be a compact manifold and let £ be a vector bundle over M. Let
(HZ I z E 1k), be a smooth family of generalized Laplacians; in other words,
we are given the following data, from which we construct the generalized
Laplacians HZ in the canonical way:
(1) a family gZ of Riemannian metrics on M depending smoothly on z E 1k;
(2) a family of connections VZ = 176 + wZ on the bundle £, where wZ is a
family of one-forms in A1(M, End (E)) depending smoothly on z c 1k;
(3) a smooth family of potentials FZ E I'(M, End(£)).
Our goal in this section is to prove the following result.

Theorem 2.48. If HZ is a smooth family of generalized Laplacians, then

for each t > 0, the corresponding family of heat kernels pt(x, y, z) defines
a smooth family of smoothing operators on the bundle E. Furthermore, the
derivative of pt(x, y, z) with respect to z is given by Duhamel's formula

8zpt(x, y, z) f (f
0 yiEM
pt-s (x, yi, z) (8zHz)y,ps (yl, y, z)) ds;

alternatively, writing this in operator form, we have

a e tH' _- f t e-(t-s)H (8ZHz)e sr ds.

0

As a first step in the proof, we will prove a result about the dependence of
the formal heat kernel of the family HZ on the parameter z. Since we are only
interested in results local in z, we may assume without loss of generality that
the injectivity radius is everywhere bounded below by e > 0. Let Vi : R+ ->
[0, 1] be a cut-off function.

Lemma 2.49. For each t > 0, the formal solution kN(x, y, z) = kN(x, y, Hz)
depends smoothly on z. Given T > 0 and z lying in a compact subset of R,
the family of kernels azkN(x, y, z), 0 < t < T, form a uniformly bounded
collection of operators on re(M, £ (9 JAI 1/2) for each £ > 0.

Proof. Consider the explicit formula for kN(x, y, z):

N
(47rt)-n/2e-dx(x,y)2/4t0(dz
(x,y)2)T t"Di(x,y,z)jz 1/2 (x, y) ldzxl1/2JdzyI1/2.
ti-o

In this formula, dz (x, y)2 is the distance between x and y E M, jz (x, y) is

the Jacobian of the exponential map, and l dZxl is the Riemannian volume,
all with respect to the Riemannian metric gz. All of these objects depend
2.7. Heat Kernels Depending on a Parameter 99

smoothly on z, as does the parallel transport Tz (x, y) in £ with respect to the

connection W.
We can show, by induction on i, that the terms (Di(x, y, z) in the asymptotic
expansion depend smoothly on z near the diagonal in M x M. Indeed, we
have the following formula for Di(x, y, z) in terms of -Pi-I(x, y, z):
1
y)(x,
T z(x, y, z) Si-l,rx(xs, y)-1 (Bz . 4bi-1)(x8, y, z) ds.
fo

It remains to observe that the operator Bz depends smoothly on z, as is clear

from the formula
Bz=j112oHzojz1/2.

The proof of uniform boundedness of the kernels kN(x, y, z) for z in a

compact subset of R is the same as in the proof of Theorem 2.29.
We now turn to the proof of Theorem 2.48.
Proof. Consider the formula for pt (x, y, z),
00
E(-t)k
k=0 JOk f yi,...,yk)EMk
k ot(x, y1, z)ra t(y1, y2, z) ... r kt(yk, y, z),

where rN (x, y, z) _ (at + HH )k v(x, y, z). By Lemma 2.49, we know that

rN (x, y, z) depends smoothly on z. From now on, restrict z to lie in a bounded
interval. It is easy to see that rN satisfies the uniform estimate
amrN
C(R)tN-n/2_e/2-m
II ax - I e <
where the Ce-norm is that of Pe(M x M, (£ ®IAI1/2) 2 (£ ®IAI1/2)).

Consider the space of all families of CI-sections of £ ®IAI1/2 which are Ck

in the parameter z, with norm
m

m<k
E azr1 IIe

Application of the operator KN (z) with kernel kN (x, y, z), 0 < t < 1, gives a
uniformly bounded family of linear operators for this norm; this follows from
the fact that the operator az Kt (z) equals
m
(i;)(azKN(z))ax
m

combined with the fact that the family 93KN(z) is uniformly bounded on
re(M, £ ® IAI1/2) for t E [0,1]. Thus, we see that the series for az pt(x, y, z)
converges uniformly in the Ce-norm as long as N > m + (Q + n)/2. Since N
is arbitrary, this proves the smoothness of pt(x, y, z) in z.
100 2. Asymptotic Expansion of the Heat Kernel

Let us now prove Duhamel's formula for apt (x, y, z) /8z. Let 0 be a smooth
section of £ 01 AI 1/2. If we apply the operator at + HZ to the section a--Pt-,o E
F(M, £ (9 AI1/2), we obtain

(at + Hz)azPP az(at + HZ)Ptzo - (azHz)Pt q

q5 =
_ -(azHz)PP O-
From this, it is easy to see that

A (x) =
a
(P0) + ff pt-.,(x, yl, z)(azHz) ,1ps(y,, y2, x) qS(y2) ds

satisfies the heat equation (at+HZ)At = 0. We will show that At (x) converges
to zero as t -+ 0; by the uniqueness of solutions to the heat equation, it follows
that it vanishes. Since 0 is an arbitrary element of F(M,£ (9 AI 1/2), this
proves Duhamel's formula.
We have the following asymptotic expansion for Pt b in I'(M, E (9 JAI 1/2):
tHz k
k=0

Furthermore, we may take derivatives of both sides, of this asymptotic expan-

sion for az Pt o. In particular,

to (f yEM azpt
(x, y, z)*(y)) = 0.

On the other hand, the section

f yi,y2)EM2
pt-8(', y,, z)(a=HZ)y,ps(yl, Y2, Z) O(y2)

is a smooth function of (s, t, x, z), and hence

lim f (
t-+0 0
f
(yi,y2)EM2
pt-e(., yl, z)(aZHZ)ylps(yl, y2, z) O(y2)) ds = 0.

The following corollary will be used repeatedly in the rest of the book.
Corollary 2.50. If Hz is a one-parameter smooth family of generalized La-
placians on a vector bundle E over a compact manifold, then

a tH'
(aaz

Str (e-tH9) = -t Str

az
Proof. If pt(x, y, z) is the kernel of the operator etH', then Duhamel's for-
mula shows that
it
a
azpt(x, y, z) ff
0 i EM
p8(x, yl, z)(zHlp8(yl, y, z) ds.
Bibliographic Notes 101

Taking the supertrace of both sides gives

t axZ a-s r) ds
Str(e-tx=) (e(s-t)H
Str
az JO J
J t Str ( x e-tx= ds,
o J
from which the result is clear, since the integrand is independent of s.

Bibliographic Notes
The pioneers of the approach to the analysis of the heat kernel that we present
in this chapter were Minakshisundaram and Pleijel [86]. A more recent work
which presents some applications to Riemannian geometry is the book by
Berger-Gauduchon-Mazet [20]; we have followed this book closely in Sec-
tions 2-4.
Chapter 3. Clifford Modules and Dirac Operators

In this chapter, we begin the study of Dirac operators, which are a gener-
alization of the operator discovered by Dirac in his study of the quantum
mechanics of the electron.
The problem that Dirac asked was the following (he was actually working
in four-dimensional Minkowski space and not in a Riemannian manifold):
what are the first-order differential operators whose square is the Laplacian?
Generalizing the question, one is motivated to ask: if M is a manifold and
E is a vector bundle over M, what are the first-order differential operators D
whose square is a generalized Laplacian on r(M, E)?
At this level of generality, the answer is quite easy to give. The operator
D, if written in local coordinates, takes the form
D = ) ak(x)8k + b(x),

where ak(x) and b(x) are sections of End(s), and it is easily seen that
E. ak(x)8k transforms as a section of the bundle Hom(T*M, End(£)); this
section is -i times the symbol of D introduced in Chapter 2,
a(D)(x,1: G dxk) =i>ak(x)ek
k k

The square of D is
D2 = 2 T(ai(x)aj (x) + ai (x)a'(x))8283 + first order operator,

so that D2 is a generalized Laplacian if and only if for any C and 77 E T XM,

we have
(a(x), C) (a(x), 71) + (a(x), ii)(a(x), rl).,
where is the metric on T* M. This is the defining relation for the
Clifford algebra of TZM, which7we introduce in Section 1, and study the
representations of in Section 2.
In Sections 3-5, we will define and study the notion of a generalized Dirac
operator, and in particular, introduce the index, which is our main object of
interest in this book. Our general application of the term "Dirac operator"
should be carefully distinguished from its more usual use, which reserves
this name for the operators introduced by Lichnerowicz and Atiyah-Singer,
3.1. The Clifford Algebra 103

which act on a twisted spinor bundle. We will discuss this special case in
Section 6. Also in Section 6, we show how some of other the classical operators
of differential geometry are generalized Dirac operators, such as d + d* on a
Riemannian manifold, and a + a* on a Kahler manifold.

3.1. The Clifford Algebra

The first step in defining the Dirac operator is to understand the linear algebra
underlying it: the Clifford algebra.
Definition 3.1. Let V be a real vector space with quadratic form Q (which
need not be non-degenerate). The Clifford algebra of (V, Q), denoted by
C(V, Q), is the algebra over R generated by V with the relations
forvandwinV,
in other words, v2 = -Q(v) for all v E V.
When Q is fixed, we may write C(V) for C(V, Q), and abbreviate Q(a, b)
to (a, b).
The Clifford. algebra of (V, Q) is the solution to the following universal
problem.
Proposition 3.2. If A is an algebra and c : V --> A is a linear map satisfying
c(w)c(v) + c(v)c(w) = -2Q (v, w),
for all v, w E V, then there is a unique algebra homomorphism from C(V, Q)
to A extending the given map from V to A.
The Clifford algebra may be realized as the quotient of the tensor algebra
T(V) by the ideal generated by the set
,7Q =Iv ®w+w®v+2Q(v,w) Iv,w EV}.
Observe that the algebra T(V) is a Z2-graded algebra, or superalgebra,
with Z2-grading obtained from the natural N-grading after reduction mod 2.
Since the generating set of the above ideal is contained in the evenly graded
subalgebra of T(V), it follows that C(V) itself is a superalgebra, C(V) =
C+ (V) ® C- (V), with V C C- (V). We will mainly consider modules over
R or C for the Clifford algebra which are 7L2-graded, by which we mean that.
the module E is a superspace E = E+ ® E-, and the Clifford action is even
with respect to this grading:
C+ (V) E=I C E=,
C- (V) E=' C E:F.

Since the algebra T (V) carries a natural action of the group O(V, Q) of
linear maps on V preserving the quadratic form Q and the above ideal is
104 3. Clifford Modules and Dirac Operators

invariant under this action, it follows that the Clifford algebra carries a natural
action of O(V, Q) as well.
Let a H a* be the anti-automorphism of T(V) such that v E V is sent
to -v. Since the ideal JQ is stable under this map, we obtain an anti-
automorphism a a* of C(V).

Definition 3.3. If Q is a positive-definite quadratic form, we say that a

Clifford module E of C(V) with an inner-product is self-adjoint if c(a*)
c(a)*. This is equivalent to the operators c(v), v E V, being skew-adjoint.

We will denote by c(v) the action of an element of V on a Clifford module

of C(V), and in particular, on the Clifford algebra C(V) itself. If E is a
Z2-graded Clifford module, we denote by Endc(v) (E) the algebra of endo-
morphisms of E supercommuting with the action of C(V).
Our first example of a Clifford module is the exterior algebra of V. To
define the Clifford module action of C(V) on AV, by the above lemma, we
need only specify how V acts on AV. To do this, we introduce the notations
E(v)a for the exterior product of v with a, and t(v) for the contraction with
the covector Q(v, ) E V*. We now define the Clifford action by the formula

(3.1) c(v)a = E(v)a - t(v)a.

To check that this defines a Clifford module action on AV, we use the formula

(3.2) e(v)L(w) + t(w)e(v) = Q(v, W).

If Q is positive-definite, the operator t(v) is the adjoint of e(v), so that the

Clifford module AV is self-adjoint.

Definition 3.4. The symbol map o- : C(V) -+ AV is defined in terms of

the Clifford module structure on AV by

a(a) = c(a)l E AV,

where 1 E A°V is the identity in the exterior algebra AV. Note that a(1) = 1.

Let e1 be an orthogonal basis of V, and denote by c1 the element of C(V)

corresponding to e1.

Proposition 3.5. The symbol map a has an inverse, denoted c : AV

C(V), which is given by the formula

c(e11 A...Aei;) =c1i...c1;.

We call c the quantization map. It follows that C(V) has the same
dimension as AV, namely 2dim(V). In fact, o is an isomorphism of Z2-graded
O(V)-modules.
3.1. The Clifford Algebra 105

The Clifford algebra has a natural increasing filtration C(V) = U j Ci(V),

defined as the smallest filtration such that
Co(V) = R,
C1(V) = V ®R.
It follows that a E C(V) lies in CC(V) if and only if it lies in the span of
elements of the form v1 ... vk, where vj E V and k < i. The following result is
analogous to the identification of gr D(M, E) with sections of S(TM)®End(E)
in the theory of differential operators; in fact, these two results are unified in
the theory of supermanifolds.

Proposition 3.6. The associated graded algebra grC(V) is naturally iso-

morphic to the exterior algebra AV, the isomorphism being. given by sending
v1 A... A vi E ASV to vi (vl .. - vi) E gri C(V ).
The symbol map o- extends the symbol map ai : CC(V) -+ gri C(V) = A'V,
in the sense that if a E Ci(V), then a (a)[ ] = a2(a). The filtration Ci(V) may
be written
i
Ci(V) = >c(AkV).
k=o

Using o, the Clifford algebra C(V) may be identified with the exterior
algebra AV with a twisted, or quantized, multiplication c -Q 0. We will need
the following formula later: if v E V and a E C(V), then
(3.3) v([v, a]) = -2t(v)o(a)
We will now show that if V is a real vector space with positive-definite
scalar product, then there is a natural embedding of the Lie algebra so(V)
into C(V).
Proposition 3.7. The space C2(V) = c(A2V) is a Lie subalgebra of C(V),
with bracket the commutator in C(V). It is isomorphic to the Lie algebra
so(V), under the map r : C2(V) --* so(V) obtained by letting a E C2(V) act
on CI (V) V by the adjoint action:
-r(a) v = [a, v].
Proof. It is easy to check that the map -r(a) really does preserve C1(V), so
defines a Lie algebra homomorphism from C2(V) to gr(V). To see that it
maps into so(V), observe that
Q(r(a) -v,w) +Q(v,-r(a) -w) = -2[[a,v],w] - 2[v, [a,w]].
Since [[v, w], a] = 0, Jacobi's identity shows that this vanishes.
The map r must be an isomorphism, since it is injective and since the
dimensions of C2(V) and so(V) are the same, namely n(n - 1)/2. 0
106 3. Clifford Modules and Dirac Operators

We will make constant use in this book of the fact that a matrix A E so(V)
corresponds to the Clifford element

(3.4) T-1(A) = 2 >(Aei,ej)cicj

i<j

Note the following confusing fact: if we identify A with an element of A2 V

by the standard isomorphism
A E so(V)'--i (Aei,ej)ei Aej,
i<j

then c(A) equals

c(A) = >(Aei, ej )ci,cj,
i<j
which differs from -r-1(A) by a factor of two.
If a E C(V), then we may form its exponential in C(V), which we will
denote by expc(a); on the other hand, if a E AV, we can exponentiate it in
AV, to form expA(a).
Let a E A2V and consider the exponential expA a in the algebra AV. If T
is the Berezin integral on AV introduced in (1.28), then by Proposition 1.36,
we have the following result.
Lemma 3.8. Assume that V is even-dimensional and oriented. If a E C2(V),
so that a(a) E A2V, then
T(exPA a(a)) = 2- dim(V)/2 detl/2(T(a))
We will often use the following formula for the exponential of an element of
C2(V) inside C(V). If v and w are vectors in V such that (v, v) = (w, w)
and (v, w) = 0, then
°O k
(v.w)k
(3.5) k.
k-0
= cost + (sin t)v w.
If we are given a Lie subalgebra of the Lie algebra underlying a finite
dimensional algebra, then we can exponentiate it inside this algebra to obtain
an associated Lie group.
Definition 3.9. The group Spin(V) is the group obtained by exponentiating
the Lie algebra C2(V) inside the Clifford algebra C(V).
The adjoint action T of C2(V) on V exponentiates to an orthogonal action
still denoted by T of Spin(V) on V. For g E Spin(V) and v E V, we have the
fundamental relation
(3.6) 9v9-1 = T(9) v.
3.1. The Clifford Algebra .107

Indeed, writing g = expc(a) with a E A2V, we see from the fact that [a, v]
r(a)v that expc(a)v expc(a)-1 = exp(T(a)) v.
Proposition 3.10. If dim(V) > 1, the homomorphism
r: Spin(V) --+ SO(V)
is a double covering.

Proof. The map 'r is clearly surjective, since the exponential map is surjective
on SO(V). Let g E Spin(V) be such that T(g) = 1. Then [9,v] = 0 for all
v E V. Formula (3.3) implies that c(v)o,(g) = 0 for all v E V, so that g is a
scalar.
If dim(V) > 1, then -1 E Spin(V), as follows from (3.5) with t = 7r.
Consider the canonical anti-automorphism a H a* of C(V). On C2(V),
this anti-automorphism equals multiplication by -1, since if v and w are
orthogonal vectors,
(v w)* = (-w) (-v) = _v. w.
Since

expc(a)* = expc(a*) = expc(-a)

for a E C2 (V), we see that g g* = 1 for all g E Spin(V). For g in the kernel
of r, and hence scalar, the relation g g* = 1 implies that g = ±1. Since
-1 E Spin(V), we see that r is a double cover.
Any Clifford module restricts to a representation of the group Spin(V) C
C+(V); this gives a nice way to construct those representations of Spin(V)
which do not descend to representations of SO(V). These are known as spinor
representations, and will be discussed in the next section.

Proposition 3.11. If the Clifford module E is self-adjoint, then the repre-

sentation of Spin(V) on E is unitary.

Proof. The representation of C2(V) on E is skew-adjoint, since c(a)* _

c(a*) = -c(a) for a E C2(V). It follows that the Clifford action of C2(V)
exponentiates to a unitary representation of Spin(V).
Definition 3.12. Let V be a vector space. If X E End(V), let

Jv (X) _ sinh(X/2)
X/2
- eX/2 - e-X/2
X
and let
Csinh(X/2)1
jv(X) = det(Jv(X)) = det
\ X/2 J
108 3. Clifford Modules and Dirac Operators

Let us compare the two functions a(expc(a)) and expAa(a) from C(V) to
AV. If X E End(V), let
Hv(X) - X/2.
tanh(X/2)
be the analytic map from End(V) to itself which is defined on the set of
X whose spectral radius is less than 27r. Since Hv(0) = 1, there is a well
defined square root BT' 1'(X) E GL(V) if X lies in a neighbourhood of 0. For
a E C2(V), we will denote Hv(-r(a)) by Hv(a), and jv(T(a)) by jv(a). We
denote by a g a the automorphism of the algebra AV corresponding to
an element g E GL(V).
Proposition 3.13. If a E C2(V) = c(A2V) is sufficiently small, we have
o-(expc(a)) = jv 2(a) det(H/ 2(a)) (Hv1'2(a) . exPA(a(a))).
Proof. We can choose an orthonormal basis (ei)$ 1 of V such that
a = 91ele2 + 92e3e4 + ....

In this way we reduce the proof to the case in which V is two-dimensional and
a = 9ele2. In this case, expc a = cos 9 + sin 9e1e2i while jv 2(a) = 9-1 sin 0,
and the matrix Hv(a) equals 9 cot 9.
The preceeding proof shows that on the Cartan subalgebra of so(V) of
matrices of the form
Xe2ti_1 = 29ie2Li 1 < i < dim(V)/2,
Xe2i = -29ie2i-1,
we have the formula
I
jv(X) = 11 I\ 9
r sin
0i )
i=1

It follows that on this subalgebra, jv(X) has an analytic square root

jv2(X)=flsinOi
i=1 Bi

We may now apply the following theorem of Chevalley, which is proved in

Chapter 7 (see Theorem 7.28).
Proposition 3.14. Let G be a -connected compact Lie group with Lie algebra
g,a nd let T be a maximal torus of G, with Lie algebra t and Weyl group
W (G, T). If 0 is an analytic function on t which is invariant under W (G, T),
then it extends to an analytic function on all of g.
Corollary 3.15. The function j/2 (X) has an analytic extension to so(V).
3.2. Spinors 109

Let T be the Berezin integral on AV. Since T(g a) = det(g)T(a), we

obtain the following corollary of Proposition 3.13.
Proposition 3.16. If a E C2(V), then
T(v(expc(a))) = jv 2(a) T(expn o(a))

3.2. Spinors
In this section, we will construct the spinor representation of the Clifford
algebra of a Euclidean vector space V. We will suppose that V has even
dimension n, and, in addition, that an orientation on V has been chosen, to
remove an ambiguity which otherwise exists in defining the Z2-grading of the
spinor space.
In the following lemma, we do not assume that V is even-dimensional.
Lemma 3.17. Let ej, 1 < j < n, be an oriented, orthonormal basis of V,
and define the chirality operator
r = i Pel...en,

where p = n/2 if n is even, and p = (n+1)/2 if n is odd. Then r E C(V)®C

does not depend on the basis of V used in its definition. It satisfies r2 = 1,
and super-anticommutes with v for v E V, in other words, rv = -vr if n is
even, while rv = yr if n is odd.
It follows from this lemma that r belongs to the centre of C(V) ®C if n is
odd, while if n is even and E is a complex Clifford module, we can define a
Z2-grading on E by
E±= {vEEIrv=±v}.
Furthermore, when the dimension of V is divisible by four, r belongs to the
real Clifford algebra C(V), so that in that case, real Clifford modules are also
Z2-graded.
Definition 3.18. A polarization of the complex space V ® C is a subspace
P which is isotropic, that is, Q(w, w) = 0 for all w E P, and such that V®C =
P ® P; here, we extend the bilinear form Q on V ® C by complex linearity.
The polarization is called oriented if there is an oriented orthonormal basis
e;, of V, such that P is spanned by the vectors {e2j_1 - ie2j 1 < j < n/2}.

Proposition 3.19. If V is an even-dimensional oriented Euclidean vector

space, then there is a unique Z2-graded Clifford module S = S+ ® S-, called
the spinor module, such that
C(V) ® C L--- End(S).
In particular, dim(S) = 2f12 and dim(S+) = dim(S-) = 2(n/2)-1.
110 3. Clifford Modules and Dirac Operators

Proof. Given a polarization of V, finding a realization of S is very simple.

The bilinear form Q on V ® C places P and P in duality. The spinor space
S is now defined to be equal to the exterior algebra AP of P. An element v
of V is represented on S by splitting it into its components in P and P; the
first of these acts on AP by exterior product, while the second, which may be
considered to be an element of P* by the above duality, acts by contraction
(there are some constants inserted to make the relations of the Clifford algebra
come out precisely): if s c S = AP, we have

c(w) s = 21/2e(w)s, if w E P,
if wEP=P
To see that C(V) ® C is in fact isomorphic to End(S), we only need to
observe that dim(C(V)) = dim(S)2, and that no element of C(V) acts as
the zero map on S. The uniqueness of the spinor representation is now a
consequence of the fact that the algebra of matrices is simple, that is, it has
a unique irreducible module.
We will now show that if the polarization P of V is oriented, the operator
c(F) on S = AP is equal to (-1)k on AkP, so that

(3.7) S+ = A:: P.

In terms of the vectors we = 2-1/2(e2j_1 - ie2j), we may rewrite r as

r = 2-n/2(wiwl - wlwl) ... (wn/2tD ,2 - 2Un/2wn/2)

Thus, P acts on AP by

(-1)n/2(E(w1)t(w1) - t(wl)e(wi)) ... (E(wn/2)t(wn/2) - b(wn/2)E(wn/2));

this is easily seen to equal (-1)k on AkP.

Since Spin(V) c C+(V), it follows that both S+ and S- are representa-

tions of Spin(V). They are called the half-spinor representations, and have a
great importance in differential geometry.

Definition 3.20. We denote by p the spinor representation of Spin(V) in

S=S+ED S-.
There is a natural supertrace on C(V), defined by

Trs+(a) - Trs_ (a), if a E C+ (V),

(3.8) Str(a) = Trs(ra) =
{0, ifaEC-(V).
We will derive an explicit formula for this supertrace; before doing so, we will
prove a result about abstract supertraces on C(V).
3.2. Spinors 111

Proposition 3.21. If the quadratic form Q is non-degenerate, then there is,

up to a constant factor, a unique supertrace on C(V), equal to T o a. The
supertrace Str(a) defined in (3.8) equals
Str(a) = (-2i)'nl2T o a(a).
Proof. We will show that [C(V), C(V)] = C,,,_1(V). It follows that any su-
pertrace on C(V) must vanish on C,,,_1(V), and hence be proportional to
T o a.
Let (ei)z 1 be an orthonormal basis of V. For any multi-index I C
{1 ... , n}, we form the Clifford product Cl = fl EI Ci; the set
{cjIIC{1,...,n}}
is a basis for C(V). If ill < n, then there is at least one j such that j 0 I,
and we obtain
c(ej) = -2[c?,cjcj].
On the other hand, the space of supertraces on C(V) is at least one-dimension-
al, since we have constructed a non-vanishing supertrace in (3.8). To calculate
the normalization of this particular supertrace, it suffices to calculate its value
on the single element r, which by (3.7) equals dim(AP) = 2f12. In this way,
we obtain the formula Str(a) = (-2i),/2T o a(a).
In particular, for a E C2 (V), we see by Corollary 3.16 that

(3.9) Str(expca) _ (-2i)n/2jl 2(a)T(expAa(a))

Lemma 3.22. If g E Spin(V), then T(a(g))2 = det((1 - 7-(g))/4).
Proof. The calculation is easily reduced to the case in which dim(V) = 2 and
g = cos 9 + sin 9e1e2, so that
det((1 - Tr(g))/4) = (1 - e2ie)(1 - e-tie)/16 = 4 sine 9.
It follows from Lemma 3.22 that if g E Spin(V) is such that det(l - r(g))
is invertible, then T(a(g)) E A'V is non-zero. Define (g) to equal +1 (re-
spectively -1) if T(a(g)) is a positive (negative) element of A'V. With
this notation, we obtain the following formula for the supertrace of the Spin-
representation. This result expresses in a quantitative way the notion that
the spinors are a "square-root" representation.
Proposition 3.23. If g E Spin(V), then
Str(p(g)) = i-n'2e(g)Idet(1 - T(g))I1/2.

When g E SO(V), the eigenvalues of g occur either in conjugate complex

pairs, or are fl, and hence det(1 - g) > 0. The following result does not
mention the spinor representation, but we state it here because the proof
makes use of this representation.
112 3. Clifford Modules and Dirac Operators

Proposition 3.24. If g E SO(V), then the function det(1 - g exp A) of A E

.5o (V) has an analytic square root.
Proof. The function in/2 Str(p(g exp A)), where g" is any element of Spin(V)
which covers g E SO(V), is an analytic square root of det(1 - gexpA).
Definition 3.25. If g E SO(V), then D9(A), A E so(V), is the analytic
square root of det(1 - g exp A) such that D9(0) = Idet(l _g)1/21.
We will use the more suggestive notation det1/2(1 - g exp A) for D9(A).
Next, we show that the spinor representation is self-adjoint.
Proposition 3.26. There is, up to a constant, a unique scalar product on
the spinor space S for which it becomes a self-adjoint Clifford module. With
respect to this metric, the spaces S+ and S- are orthogonal to each other.
Proof. Uniqueness is clear, since it is implied by the irreducibility of S as a
C(V)-module. To show the existence of a scalar product, we construct one
explicitly on the realization AP that we obtained above. Indeed, it is sufficient
to have a scalar product on the space P, which can be extended in the usual
way to all of AP. We choose the scalar product furnished by (w, w), where
w is the complex conjugate of w and is thus an element of P. In particular,
it is immediate that S+ = A+P is orthogonal under this scalar product to
S- = A- P.
To show that this metric makes the Clifford module S self-adjoint, it suffices
to show that the operator c(v) is skew-adjoint for v e V. But an element of
V, being real, has the form w + w; thus it acts on AP by v/'2-(e(w) - a(w)),
which is clearly skew-adjoint.
The next result shows that the spinor space S is the building block for
complex representations of C(V).
Proposition 3.27. If V is an even-dimensional real Euclidean vector space,
then every finite-dimensional Z2-graded complex module E of the Clifford
algebra C(V) is isomorphic to W ®S, for the Z2-graded complex vector space
W = Homc(v) (S, E) which carries a trivial C(V)-action, and End(W) is
isomorphic to Endc(v)(E).
Proof. This follows from the fact that any finite dimensional module for the
matrix algebra End(S) is of the form W ® S. The isomorphism between
HomC(v) (S, E) ®S and E is given by w ®s H w(s).
The space W will be called the twisting space for the Clifford module E.

Since Strs (F) = 2n/2, the supertrace over W of an element F E End(W)

Endc(V) (E) is given by the formula
(3.10) Strw(F) = 2-n/2 StrE(rF).
Motivated by this, we make the following definition.
3.3. Dirac Operators 113

Definition 3.28. The relative supertrace StrE/s : Endcivl (E) -- C

equals
StrE/s(F) = 2-n/2 StrE(rF).
If W is ungraded, the relative supertrace equals the W-trace, and is pro-
portional to the E-trace:
Trw(F) = StrE/s(F) = 2-n/2 TrE(F).
We end this section with some formulas for the special case in which V
is a complex Hermitian vector space; these formulas will be of use in study-
ing the 8-operator on a Kahler manifold. In this case, the group U(V) of
unitary transformations of V is a subgroup of SO(V). Let us consider the
decomposition
V ® C = V1,0 ® V°"1,
where V1"0 and V°"1 are the spaces on which the complex structure J of V
acts by +i and -i respectively. Then V1"0 is a polarization of V ® C and
we can take S = AV""° as the spinor space. On the other hand a complex
transformation of V leaves V1"0 stable, so there is a natural representation
A of U(V) on AV""°. The representation A and the restriction of the spinor
representation p of Spin(V) to T-1 U(V) differ by the character det1 2 of the
double cover -r-'(U(V)) of U(V). More precisely, the infinitesimal represen-
tations of the Lie algebras on S, which we denote also by A and p, are related
by the following result.

Lemma 3.29. Let a E C2(V). If r(a) is complex linear on V, then

p(a) _ A(T(a)) - Trvi,o T(a).
a
Proof. Choose a complex basis wj of V1"0 such that r(a)wj = Ajwj for every
j, and (wj, 0i) = 6ij. Then a = 4 1:1<i<n/2 '\i(wiwi - wi'U10, so that

A(T(a)) = AiE(wi)t('wi),

p(a) = 2 A (E(wi)t('wi) - A(r(a)) -zEA .

From these formulas, the result is evident.

3.3. Dirac Operators

Definition 3.30. If M is a Riemannian manifold, the Clifford bundle
C(M) is the bundle of Clifford algebras over M whose fibre at x E M is
the Clifford algebra C(T*XM) of the Euclidean spaces T**M.
114 3. Clifford Modules and Dirac Operators

The Clifford bundle C(M) is an associated bundle to the orthonormal frame

bundle,
C(M) = O(M) xo(,) C(R').
From this representation of C(M), we see that it inherits a connection, the
Levi-Civita connection, which is compatible with the Clifford product, in the
sense that
V (ab) = (Da)b + a(Vb) for all a and b E r(M, C(M)).
Definition 3.31. The symbol map a
o : C(M) -* AT*M,
which identifies the Clifford bundle C(M) with the exterior bundle AT*M, is
defined by means of the symbol map o : C(T*XM) --3 AT XM.

Definition 3.32. (1) A Clifford module £ on an even-dimensional Rie-

mannian manifold M is a Z2-graded bundle £ on M with a graded action
of the bundle of algebras C(M) on it, which we write as follows:
(a, s) i c(a)s, where a E r(M, C(M)) and s E r(M, £).
(2) If £ is a Clifford module with metric h, for which £+ and £- are orthog-
onal, we say that the Clifford module 6 is self-adjoint if the Clifford
action is self-adjoint at each point of M, in other words, if the operators
c(a) with a E T*M are skew-adjoint.
(3) If W is a vector bundle, then the twisted Clifford module obtained from
6 by twisting with W is the bundle W ® 6, with Clifford action 10 c(a).
The most basic example of a Clifford module is the bundle of spinors. In
order to define this, we need a spin-structure on the manifold.
Definition 3.33. A spin-structure on a manifold M is a Spin(n)-principal
bundle Spin(M) on M such that the cotangent bundle of M is isomorphic to
the associated bundle Spin(M) xspin(,,,)R".
Note that a manifold with a spin-structure is, in particular, an oriented
Riemannian manifold, with frame bundle
SO(M) = Spin(M) xspin(n) SO(n).
Thus the principal bundle Spin(M) is a double cover of SO(M) and inherits
from SO(M) the Levi-Civita connection.
There may be topological obstructions to the existence of a spin-structure
on M, while on the other hand, there may be more than one. This question
is understood using the classification of isomorphism classes of principal 0-
bundles on a manifold G by the cohomology group H' (M, G), combined with
the Bockstein long exact sequence
0 , H' (M, Z2) , H1 (M, Spin (n)) --> H1(M, SO(n)) --> H2(M, Z2)
3.3. Dirac Operators 115

associated to the short exact sequence of groups

1->7L2---Spin(n)-fSO(n)-'1.
The frame bundle SO(M) defines an element a E HI(M, SO(n)) whose image
in H2(M,7L2) may be identified with the second Stieffel-Whitney class of M,
written W2(M). This at least motivates the following result.

Proposition 3.34. An orientable manifold M has a spin-structure if and

only if its second Stieffel- Whitney class w2 (M) E H2 (M, Z2) vanishes. If this
is the case, then the different spin-structures are parametrized by elements of
HI(M,7L2) = Hom(irl(M),Z2).

If M is an even-dimensional spin-manifold, the spinor bundle S is defined

to be the associated bundle
S = Spin(M) xspin(n) S.
It is clear that S is a Clifford module, since the action of C(1R) on S leads
to an action of the associated bundle C(M) = Spin(M) xspin(n,) C(RI) on S.
The following is the analogue for manifolds of Proposition 3.27.

Proposition 3.35. If M is an even-dimensional oriented manifold with spin-

structure, every Clifford module £ is a twisted bundle £ = W ® S.

Proof. We may define a complex vector bundle W such that £ = W ® S

as Clifford modules by the formula W = Endc(M) (S, £). The isomorphism
between Homc(M) (S, £) ®S and £ is given by w ® s - 4 w(s).
Locally, we may trivialize the bundle C(M) by choosing an orthonormal
frame of the cotangent bundle. Over such an open set U, the Clifford bun-
dle C(M) ^_' U x C(RI). Hence, we see that a spinor bundle S always
exists locally, and we can always locally decompose a Clifford module S as
Homc(M) (S, £) ® S.
If a E A(M, C(M)) is a Clifford algebra-valued differential form on M,
we can define an operator c(a) which acts on the space A(M, £) by means of
the following formula: if a and 0 are differential forms on M, a is a Clifford
algebra section, and s is a section of £, all homogeneous with respect to the
Z2-grading, then

(3.11) (c(a (9 a))(0 (9 s) _ (_i)la1 IQI (a A0) ® (c(a)s)

The interest of Clifford modules is that they are naturally associated to a

class of first-order differential operators, known as Dirac operators. We will
use a more general definition of Dirac operators than is usual, since it fits in
better with the notion of a superconnection, and does not make any of the
proofs more difficult.
116 3. Clifford Modules and Dirac Operators

Definition 3.36. A Dirac operator D on a Z2-graded vector bundle £ is

a first-order differential operator of odd parity on £,
D : r(M, £±) --> P(M, £:r),
such that D2 is a generalized Laplacian.
Applying the results of Chapter 2, we obtain the following result. From
now on, when discussing the heat kernel of a generalized Laplacian H on a
Riemannian manifold M and vector bundle £, we will use the Riemannian
density to identify the heat kernel (x I e -1H I y) of H with a section of £®£*
over M x M.
Proposition 3.37. (1) If D is a Dirac operator on a compact manifold M,
then D2, acting on P(M, £), has a smooth heat kernel
e_tO2
(x I I y E P(M x M, £ ®£*).
(2) If the manifold M is compact, the Dirac operator D has finite dimensional
kernel.
In particular, a symmetric Dirac operator D is essentially self-adjoint.
Proof. The operator D2 has a finite dimensional kernel, and hence so does the
Dirac operator D. The proof that D is essentially self-adjoint uses the heat
kernel (x I e-tD I y) in the same way as the proof of the corresponding result
for D2 given in Proposition 2.33 did. 0
The purpose of this section is to work out what are the geometric data
needed to specify a Dirac operator: this is analogous to the one-to-one corre-
spondence between generalized Laplacians and the three pieces of data con-
sisting of a metric on M, a connection on £ and a potential F E r(M, End(S)).
If we are given a Dirac operator on a vector-bundle £, then S inherits a
natural Clifford module structure.
Proposition 3.38. The action of T*M on £ defined by
[D, f] = c(df), where f is a smooth function on M,
is a Clifford action, which is self-adjoint with respect to a metric on £ if the
operator D is symmetric. Conversely, any differential operator D such that
[D, fj = c(df) for all f E CO° (M) is a Dirac operator.
Proof. First of all, this definition is unambiguous, since if f and h are both
smooth functions on M, we have
c(d(f h)) = [D, f h] = f c(dh) + hc(df ).
To check the relations for c(df) to be a Clifford action, we need only observe
that
c(df )2 = 2 L[D, f], [D, f]] = 1 [[D2, f], f] = -I df I2,
3.3. Dirac Operators 117

since D2 is by assumption a generalized Laplacian; we have used the fact that

[[D, f], f] = 0 and that [D, [D, Al] = [D2, A], which is true for any odd operator
D and even operator A. This equation also shows that any operator satisfying
[D, f] = c(df) for all f c C' (M) is in fact a Dirac operator. Finally, if D is
self-adjoint with respect to some metric on £, then for all real functions f on
M,
c(df)* _ [D, f]* _ [.f*, D*] _ -c(df ).
It is clear that there is always at least one Dirac operator on a Clifford
module; we can construct one explicitly by taking the operators >8 c(dxi)8i
on local coordinate patches and combining them by means of a partition of
unity.
If D is a Dirac operator on a bundle £ and A is an odd section of the bundle
End(9), the operator D +A will be another Dirac operator on £ corresponding
to the same Clifford action on S. Conversely, any two Dirac operators on £
corresponding to the same Clifford action will differ by such a section. Indeed,
if Do and D1 are two Dirac operators on £,
[D1 - Do, f] = c(df) - c(df) = 0

for all f E C' (M), and hence D1 - Do may be represented by the action of a
section of End- (E). Thus, the collection of all Dirac operators on a Clifford
module is an affine space modelled on r(M, End- (£)). In order to sharpen
this identification, we will introduce the notion of a Clifford superconnection
on a Clifford module.
Definition 3.39. (1) If DE is a connection on a Clifford module £, we say
that V is a Clifford connection if for any a E r(M, C(M)) and X E
r(M,TM),
[Ox, c(a)] = c(Vxa).
In this formula, Vx is the Levi-Civita covariant derivative extended to
the bundle C(M).
(2) If A is a superconnection on a Clifford module £, we say that A is a
Clifford superconnection if for any a E r(M, c(m)),
(3.12) [A, c(a)] = c(Va).
In this formula, Va is the Levi-Civita covariant derivative of a, which is
an element of A1(M, C(M)).
It follows from this definition that if A is a Clifford superconnection, then
the formula (3.12) holds for all a E A(M, C(M)).
The collection of all Clifford superconnections on .6 is an affine space based
on A- (M, Endc(M) (£)), the space of sections of AT*M ® Endc(M) (£) of odd
total degree; this follows from the fact that if A is a Clifford superconnection
on £ and a E A- (M, EndQ(M) (£)), then A + a is again a Clifford supercon-
nection.
118 3. Clifford Modules and Dirac Operators

The Levi-Civita connection VS on the spinor bundle S of a spin manifold is

a Clifford connection. If wk are the coefficients of the Levi-Civita connection
with respect to an orthonormal frame eti of the tangent bundle, defined by
Va,ej = w 3ek, then (3.4) shows that VS is given in local coordinates by the
formula

(3.13) pat
jk

Proposition 3.40. Let M be a spin manifold and let 8 = W OS be a twisted

spinor bundle.
(1) If Al^' is a superconnection on the bundle W, then the tensor product
superconnection A = AW 0 1 + 1 ® VS on the twisted Clifford module
W ® S is a Clifford superconnection. We call Ai'r' the twisting super-
connection associated to the Clifford superconnection A, and denote it
by Ar/s.
(2) There is a one-to-one correspondence between Clifford superconnections
on the twisted Clifford module W®S and superconnection on the twisting
bundle W, induced by the correspondence Al"V - (Aw ® 1 + 1(9 VS).

Proof. Part (1) follows from the formula for the Clifford action on W ® S,
which sends a E C(M) to 1 ® c(a) E r(M, End(W ® S)). Thus, we see that

[A"V®1+1® V8,1®c(a)]=1®[V8,c(a)]=1®c(Va).
To see that the assignment of a twisted Clifford superconnection on W ® S
to a superconnection on W is a bijection, we simply observe that the space of
Clifford superconnections on W ® S is an affine space modelled on the space
of sections A-(M,EndC(M)()IV ®£)) = A- (M, End(W)), as is the space of
superconnections on W.
If A is a Clifford superconnection on the twisted Clifford module W ® S,
we define an operator Al't' acting on sections of W by the formula

(AWu) ®s = A(u (9 s) - u 0 (VSS),

where s is a non-vanishing section of S. It is easy to see that this formula is

independent of the section s used, since any two sections sl and s2 of S are
related locally by the formula s2 = c(a)se for some section a E r(M, C(M))
of the Clifford bundle C(M). We then see that

A(u (9 S2) - u ® (VS82) = A(u ® (c(a)si) - u ® (Vsc(a)si)

= c(a) (A(u ®sl) - u (&(V8s1))

Corollary 3.41. There exists a Clifford superconnection on any Clifford

module £.
3.3. Dirac Operators 119

Proof. To construct a Clifford connection on the Clifford module E, it is

sufficient to do this locally and then patch the local Clifford connections to-
gether by means of a partition of unity. But locally, we may decompose £ as
W®S. It suffices to take any connection Vw and form the Clifford connection
Vw 1+1®Os.
We wil now study the relationship between Clifford superconnections and
Dirac operators. Let us start with the case of a Clifford connection v-'. We
may define a Dirac operator on ,6 by means of the following composition:

P(M, E) -°- r(M, T*M (9 £) - r(M, £)

In local coordinates, this operator may be written
D= c(dxi)Da,
i
from which it is clear that it is a Dirac operator, that is, that [D, f] = c(df).
Often in the results of this book, we will have to restrict attention to operators
in the subclass of Dirac operators constructed from a Clifford connection. In
terms of an orthonormal frame ei of the tangent bundle TM over an open set
U with dual frame ei of T*M, we may also write
D = E c(e)Ve..
i
If M is a spin manifold, the Dirac operator on S associated to the Levi-
Civita connection Vs is the operator often referred to as the Dirac operator
associated to the spin-structure, with no further qualification. Other authors
reserve the term Dirac operator for twisted versions of this operator; we have
chosen a far more liberal definition of Dirac operators because many results
of index theory deserve to be stated in greater generality.
Let A be a Clifford superconnection on the Clifford module £. We can
define a first-order differential operator on r(M, £) , which we will write as
DA, by composing the superconnection with the Clifford multiplication map:

P(M, £) -& A(M, £) _-. r(M, C(M) (& £) -_c+ r(M, £)

Here, we have made use of the isomorphism
A(M, £) = I'(M, C(M) ® £)
given by the quantization map c : AT*M -; C(M) of Proposition 3.5. With
respect to a local coordinate system and orthonormal frame ei of the tangent
bundle in which

Adx''®8i+ el ®AI,
i=1 Ic{i,...,n}
120 3. Clifford Modules and Dirac Operators

AI being sections of End(£), we have

n
DA = c(dxi)8i + c(el )AI.
i=1 IC{1,...,n}

Proposition 3.42. The map sending a Clifford superconnection A to the

operator DA is a one-to-one correspondence between Clifford supereonnections
and Dirac operators compatible with the given Clifford action on E.
Proof. First, we must check that the operator DA is in fact a Dirac operator
compatible with the given Clifford action. This follows from Leibniz's formula
A(f s) = df ®s + f As E .A(M, £), where f E C°° (M) and s E F(M, £).
If we apply the Clifford map
A(M, £) --> F(M, C(M) ® 6) --+ F(M, £)
to both sides, we obtain the desired equation, namely
DA(f s) = c(df)s + f DAs.
It remains to show that any Dirac operator on £ may be realised in a unique
way in the form DA for some Clifford superconnection A. The uniqueness
is immediately deduced from the fact that any two Clifford superconnections
differ by an element of A- (M, Endc(M) (£)). If Do and D1 are two Dirac
operators on £, D1 - Do may be represented by the action of a section of the
bundle End- (£). However, applying the symbol map
F(M, End- (S)) = F(M, c(m) ®Endc(M) (£))- -p A- (M, End(£))
to this section, we obtain a differential form w E A-(M,Endc(M)(E)). It
follows from this that D1 = Do+c(w). If Do is the Dirac operator associated to
the Clifford superconnection A, we see that D1 is the Dirac operator associated
to the Clifford superconnection A + w.

Proposition 3.43. If A is a Clifford superconnection on 6, then the curva-

ture A2 E A(M, End(£)) of A decomposes under the isomorphism End(S)
C(M) ® Endc(M) (£) as follows:

A2 = RE + Fe/s.
In this formula, RE E A2(M,C(M)) C A2(M,End(£)) is the action of the
Riemannian curvature R of M, on the bundle £, given by the formula

RE(ei, ej) = 4 E(R(ei, ej)ek, et)ci`ci,

kt

and Fe/s E A(M,Endc(M) (£)) is an invariant of A, called the twisting

curvature of the Clifford module E.
3.3. Dirac Operators 121

Proof. Define Fe/S to equal A2 - RE. If we can show that the operation
e(FE/S) of exterior multiplication by FEIS, acting on A(M, 6), commutes
with the operators c(a), where a E P(M,T*M) C r(M,C(M)), then it will
follow immediately that FE/S is actually a differential form with values in
EndC(M)(£). But this commutation follows immediately from the the fact
that A is compatible with the Clifford action, as expressed in (3.12):
[A2, c(a)] _ [A, [A, c(a)]] = [A, c(Da)] = c(V2a).
To finish the proof, we use the fact that V2a = Ra,where R is the Riemannian
curvature of M and that [RE, c(a)] = c(Ra), so that
[A2, c(a)] = c(Ra) = [RE, c(a)] -

Let M be a spin manifold and .6 = W®S be a twisted spinor bundle. If AW

is a superconnection on W, with curvature FW, and if A = AW ®1+1®VS is
the corresponding twisted superconnection on £ = W 0 S, then the curvature
A2 equals
A2 = RE + F' .
Thus, the twisting curvature Fe/s of A equals the curvature FW of the twist-
ing superconnection AW; this explains why we call it the twisting curvature.
In the next proposition we compute the formal adjoint of a Dirac operator.
Recall that the adjoint V* of an ordinary connection V on a Hermitian vector
bundle is defined by the equality of one-forms
d(si, s2) = (Vs1, s2) + (si, V*s2) for all Si E r(M, £).
Likewise, if A = V + w is a superconnection, where w E A- (M, End(S)),
then A* is defined to equal V* + w*, where by w* we mean the differential
form EZ 0(-1)i(i+1)/2w*; the reason for the sign (-1)i(i+i)/2 is that the
differential form dfl . . . dfa = (-1)i(i-1)/2dfi... df1 and taking adjoints will
introduce an additional (-1)i sign.
Proposition 3.44. The adjoint of the Dirac operator DA is the Dirac oper-
ator DA..
Proof. It is easy to see that it suffices to prove this when A is a connection V.
Let Si and s2 be two sections of £ and let X be the vector field on M given by
a(X) = (Si, c(a)s2) for a E A'(M). Thus VX belongs to r(M,TM®T*M).
It is easy to see that
(DAS1, S2)x = (si, DA' S2), rk(VX),.
The integral of the last term vanishes by Proposition 2.7, proving the propo-
sition.
Dirac operators are characterized by the formula [D, f] = c(df) for all
f E C°O(M). In the next proposition, we calculate [D, c(B)] when 0 is a one-
form on M and D is a Dirac operator. Let VE be the connection associated
to the generalized Laplacian D2 on the bundle S.
122 3. Clifford Modules and Dirac Operators

Proposition 3.45. If 0 E A1(M) is dual to the vector field X, then

[D, c(8)] = -2V + c(d9) + d*B,

where we recall that

c(dG) = E(d8)(ei, ej)c(ei)c(ej).

i<j
Proof. It suffices to prove the theorem for differential forms of the type f dg.
Let X = grad g be the vector field dual to dg. Observe that

[D, c(f dg)] = f [D, c(dg)] + c(df) c(dg)

Since c(df) c(dg) = c(df n dg) - (df, dg), we see that

[D, c(f dg)] = f [D, c(dg)] + c(d(fdg)) - (df, dg)-

Now observe that
[D2,
[D, c(dg)] _ [D, [D, g]] = g]
_ -2V + Og,
since D2 is a generalized Laplacian associated to the connection DE. We see
that
[D,c(fdg)] = -2fVX +c(d(fdg))+ f (d*dg) - (df, dg).
Using the fact that d* (f dg) = f Og - (df, dg), the result follows. O

3.4. Index of Dirac Operators

Let £ be a Z2-graded vector bundle on a compact Riemannian manifold M,

and let D : r(M, £) - r(M, £) be a self-adjoint Dirac operator. We denote
by D± the restrictions of D to r(M, £t), so that

0 D-
D= D+ 0 ,

where D- = (D+)*.
An important invariant that can be associated to the operator D is its index.
If E = E+ ® E- is a finite-dimensional superspace, define its dimension to
be
dim(E) = dim(E+) - dim(E-).
The superspace ker(D) is finite dimensional.
3.4. Index of Dirac Operators 123

Definition 3.46. The index space of the self-adjoint Dirac operator D is

its kernel
ker(D) = ker(D+) ® ker(D-).
The index of D is the dimension of the superspace ker(D),
ind(D) = dim(ker(D+)) - dim(ker(D-)).

Lemma 3.47. If u = (° ) is an odd self-adjoint endomorphism of a

o coker(u+) = ker(u-)*, where coker(u+) _
superspace E = E+ ® E-, then
E-/ im(u+).
Proof. Since u- is the adjoint of u+, we have for all e e E+ and f E E- that
(u+e,f) = (e,u f),
so that f E ker(u-) if and only if (u+ e, f) = 0 for all e E E+.
We will now prove an analogous result for self-adjoint Dirac operators.
Thus let £ be a Z2-graded Clifford module on a compact Riemannian manifold
M of even dimension, and let D be a self-adjoint Dirac operator on r(M, £).
As above, we denote by D± the restrictions of D to r(M, £±).
Proposition 3.48. If D is a self-adjoint Dirac operator on a Clifford module
S over a compact manifold M, then
r(M, £:':) = ker(Dt) ® im(D:F),
and hence
ind(D) = dim(ker(D+)) - dim(coker(D+)).
Proof. Since D is self-adjoint, (D2s, s) = (Ds, Ds) and ker(D) = ker(D2).
Let G be the Green operator for the square of D; this is the operator
which inverts D2 on the orthogonal complement of ker(D). Since D2 is a
generalized Laplacian, G preserves r(M, £) by Proposition 2.38. Let P0 be
the projection operator onto ker(D). If s E r(M, £), then s = D(DGs) + Pos
is a decomposition of s into a section D(DGs) in the image of the D and a
section in the kernel of D. This shows that
r(M, £t) = im(D:F) ED ker(D±).
From this proposition, we see that the index ind(D) may be identified with
the dimension of the superspace ker(D+) ®coker(D+) if D is self-adjoint. This
is the usual definition of the index of D+. However, we prefer to phrase the
definition of the index in a more symmetric form, involving both D+ and D-,
so we adopt our first definition of the index. It is also for this reason that we
write ind(D) and not the more usual notation ind(D+).
The index of D is a topological invariant of the manifold M and the Clifford
bundle £: that is, if Dz is a one-parameter family of operators on r(M, £)
which are Dirac operators with respect to a family of metrics gz on M and
Clifford actions cz of C(M,gz) on £, then the index of Dz is independent of
124 3. Clifford Modules and Dirac Operators

z. We will prove this by means of a beautiful formula for ind(D) in terms of

the heat kernel a-tD2 called the McKean-Singer formula.
If A is a trace-class operator on a Z2-graded Hilbert space then its
supertrace is defined as in the finite dimensional case by the formula

Str(A) _ Tr +(A) - Trx- (A) if A is even,

0 if A is odd.

Consider the Hilbert spaces 1-l of L2-sections of E±. The following lemma
is the analogue for the supertrace of Proposition 2.45.
Lemma 3.49. If D is a differential operator on E and if K has a smooth
kernel, then Str[D, K] = 0.
Proof. If K and D are both even, this follows immediately from Lemma 2.45.
On the other hand, if D and K are both odd, that is, D = (D+ o) and
K = (K+ o ), then Str(DK+ KD) = Tr[K-, D+] +Tr[D-, K+] = 0, again
by Proposition 2.45.
The supertrace of the projection onto the kernel of D is clearly equal to
the index of D:
Str(Po) = dim(ker(D+)) - dim(ker(D-)) = ind(D).
The McKean-Singer formula generalizes this formula, and will allow us to use
the kernel of a-tD2 to calculate the index of D.
e-tD2
Theorem 3.50 (McKean-Singer). Let (x I I y) be the heat kernel
of the operator D2. Then for any t > 0

ind(D) = Str(e-tD2 ) = e-tD2 x)) dxProof.

fM
We give two proofs of this formula. The first one uses the spectral
theorem for the essentially self-adjoint operator D2. If A is a real number,
let na be the dimension of the A-eigenspace 1i of the generalized Laplacian
D2, acting on r(M, E+). We have the following formula for the supertrace of
e-tD2:
2
Str(e-tD E(nt - n-)e-'A.

A>0

Since the Dirac operator D commutes with D2, it interchanges 7-la and 7-la ,
and moreover induces an isomorphism between them if A # 0, since in that
case, the composition

is equal to the identity. Thus, nt - na = 0 for A > 0, and we are left with
no - na , which is of course nothing but the index of D.
3.4. Index of Dirac Operators 125

The second proof is more algebraic; we give it because it is a model for

the proof of more general versions of the McKean-Singer formula. Let a(t) be
the function Str(e-tD2). If P1 = 1 - Po is the projection onto the orthogonal
complement of ker(D), then by Lemma 2.37, for t large,

SStr(e-t°2 Ple-tD2p1
_ Po)I = IM Str((x I I x)) dx < Cvol(M)e-tA1/2,

where )1 is the smallest non-zero eigenvalue of D2. This shows that

a(oo) = lim Str(e-tD2) = ind(D).

t--*oo

The proof is completed by showing that the function a(t) is independent of

t, so that a(t) = a(oo) = ind(D). We do this by differentiating with respect
to t, which is legitimate because the heat equation tells us that the operator
d(e_tD2)/dt has a smooth kernel equal to -D2e tD2, and is thus trace-class
fort>0. We see that
-Str(D2e`tDa)

dta(t) _

_ -2 Str([D, De-tD2]),
since D is odd, and this vanishes by Lemma 3.49.

With the McKean-Singer formula in hand, it is easy to see that the index
of a smooth one-parameter family of Dirac operators (Dz I z E R) is a smooth
function of z, since the heat kernel of (Dz)2 is. However, the index is an
integer, so it follows that it is independent of s. Even without using the fact
that the indexe_tD2
is an integer, we can show explicitly that the variation of the
supertrace of vanishes, by means of the following result.

Theorem 3.51. The index is an invariant of the manifold M and Clifford

module S.

Proof. Given a one-parameter family of Dirac operators Dz on £, we see that

(Dz)2 is a one-parameter family of Laplacians. Applying Lemma2.50 to the
McKean-Singer formula, it follows that

d ind(Dz) = d Str(e-t(Dz)2)
dz dz
= -t Str([dd z, We t(° )2]).
Once again, this last quantity vanishes by Lemma 3.49.
126 3. Clifford Modules and Dirac Operators

3.5. The Lichnerowicz Formula

If A is a Clifford connection, there is an explicit formula for D22 which is of

great importance in differential geometry, as we will see when we describe
examples of Clifford modules in the next section.

Theorem 3.52 (Lichnerowicz formula). Let A be a Clifford connection

on the Clifford module S. Denote the Laplacian with respect to the connection
A by AA. If rM is the scalar curvature of M, then

D22 _ AA + c(FE/S) + r4

where FEIs E A2 (M, Endc(M) (E)) is the twisting curvature of the Clifford
connection A defined in Proposition 3.43, and

c(FEIS) = EFEIS(ei,ej)c(ea)c(ei).
i<j

Proof. Choose a covering of the manifold by coordinate patches. In a coordi-

nate system, we may write the Dirac operator as DA = > c(dxi)Ai, where Ai
is the covariant differentiation in the direction 8/axi. Squaring this formula
gives

D2 =2
ij
=2 E[c(dxi),c(dxj)]AiAj + c(dx') [Ai, c(dxj)]Aj

+ Ec(dxi)c(dxj)[Ai,Aj]
ij

gza (AiAj + E r k Ak) + E c(dx2)c(dxj) [Ai, Aj]

ij k i<j

In this calculation, V is the Levi-Civita connection on M, and F are its

coefficients with respect to the frame ai, which satisfy Vidxj = -Fkdxk.
Here, we have used the compatibility of the connection A with the Clifford
action, so that
[Ai, c(dxj)] = c(Vidxj) = -rikc(dxk),

and the symmetry Fik = rki which follows from Viak = V kai, so that

E rzkc(dx')c(dxk) = 2 rZk[c(dxi), c(dxk)].

ik ik
3.6. Some Examples of Clifford Modules 127

Let A2 be the curvature tensor of the connection A, and let ei, (1 < i < n),
be an orthonormal frame of TM. The above calculation gives the formula

DZA = DA + > c(dxi)c(dxj) [Ai, Aj]

ij
(3.14) = AA + E c(e2)c(e2)A2(ei, ej).
i<j

By Proposition 3.43, we see that

c(ex)c(ej)A2(ei, ej) g 1: Rklijc(ei)c(e')c(ek)c(el) + c(Fe1S),

i<j ijkl
where Rijkl are the components of the Riemannian curvature of M.
Observe that c(ei)c(ej)c(ek) equals

1: sgn(o,)c(ea(2))c(ea(i))c(eo-(k)) -Sijc(ek) - Sjkc(ei) +Skic(ej).

0ES3

Using the fact that the antisymmetrization of Rklij over ijk vanishes, we
obtain

E Rklijc(ei)c(eJ)c(ek)c(el) 1: c(ea)c(el)Rjlij + E c(ej)c(el)Rilij.

ijkl iji ijl
We see that

E Rklij c(ei)c(ej)c(ek)c(el) = 2 E c(ei)c(ei) E Rkjk

ijkl ii k

But Eij c(ej)c(e1)Rikjk = - Ei Rikik, since Rikjk = Rikik Hence, the right-
hand side equals -2rMi and the Lichnerowicz formula follows.

3.6. Some Examples of Clifford Modules

In this section, we will show that the classical linear first-order differential
operators of differential geometry are Dirac operators with respect to very
natural Clifford modules and Clifford connections. By classical, we mean an
operator considered by Atiyah and Singer [13]. Our exposition follows closely
that of Atiyah and Bott [6].
The operators that we describe here, rather briefly, are those for which the
index theorem of the next chapter might be considered to be classical; both
the Gauss-Bonnet-Chern theorem and the Riemann-Roch-Hirzebruch theorem
turn out to be special cases, in which we take as our Clifford module respec-
tively the operator d + d* on A(M), and the operator a + a* on AO,* (M, S),
where M is a Kahler manifold and S is a holomorphic Hermitian bundle.
128 3. Clifford Modules and Dirac Operators

The De Rham Operator. Our first example is related to the de Rham op-
erator
'-0 12
0 A° (M) A1(M) + A2 (M) )

On a Riemannian manifold, the bundle AT*M is a Clifford module by the

action of (3.1): if a E r(M, T*M) and 0 E A(M), then c(a)/3 = 3.
Clearly, the Levi-Civita connection on the bundle AT*M is compatible with
this Clifford action.
Proposition 3.53. The Dirac operator associated to the Clifford module
AT*M and its Levi-Civita connection is the operator d + d*, where
d* : A* (M) -f A'-' (M)
is the adjoint of the exterior differential d.
Proof. Since V is torsion-free, Proposition 1.22 shows that d equals e o V,
where V is the Levi-Civita connection. On the other hand, Proposition 2.8
shows that d* = -L o V. It follows that d + d* = (e-0 o V =cop.
The square dd* + d* d of the operator d + d* is called the Laplace-Beltrami
operator. Recall the statement of de Rham's theorem: the cohomology
H'(A(M), d) = ker(di) / im(d;,_1)
of the de Rham complex is isomorphic to the singular cohomology Hi(M, R).
A differential form in the kernel of dd* + d*d is called a harmonic form.
It is useful to consider a more general situation than the above, in which
£' --> M is a Z-graded vector bundle (with £i non-zero for only a finite
number of values of i) with first-order differential operator d : I'(M, S) --+
r(M, £'+1), such that d2 = 0 and A = dd* + d*d is a generalized Laplacian.
We will write Hi (d) for the cohomology of this complex,
Hi(d) = ker(di)/ im(di_1).
In this situation, the operator d + d* is a Dirac operator, since (d + d*)2 =
dd* + d*d is a generalized Laplacian; the Z2-grading on the bundle Ei £i is
obtained by taking the Z-degree modulo two. It is clear from the following
equation that kf er(A) = ker(d + d*)2 =f ker(d) n ker(d*):

M(a, (dd* +d*d)a) = M(da, da) + (d*a, d* a).

Theorem 3.54 (Hodge). If (£, d) is as above, then the kernel ker(Ai) of

the operator Di = di*+ldi + di_idi is finite-dimensional in each dimension,
and naturally isomorphic to Hi(d) by the decomposition ker(di) = im(dz_1)
ker(Aj).
The Dirac operator d+d* on the bundle Ei £i has index equal to the Euler
number of the complex (£, d), defined by Ei(-1)i dim(Hi(d)).
3.6. Some Examples of Clifford Modules 129

Proof. Since the Laplacian A has a smooth heat kernel, we see that it is es-
sentially self-adjoint, has a discrete spectrum with finite multiplicities, and
smooth eigenfunctions. Thus, the space ker(Di) coincides with the zero-
eigenspace of the closure of di*+1di + d1_1di in the Hilbert space rLz (M, £i)
of square-integrable sections of £i.
On the orthogonal complement to ker(A), we can invert the generalized
Laplacian A to obtain its Green operator G, which we extend by zero on
ker(A); moreover, if a E r(M, £i) is smooth, then Ga E r(M, £i) is also
smooth (Corollary 2.38). Observe that A commutes with d and d* as does G.
The operator d*G on r(M, s) satisfies
[d, d*G] = (dd* + d*d)G =1 - Po,
where P0 is the orthogonal projection onto ker(A). Thus, d*G is what is
known as a homotopy between the identity operator on rL2 (M, and Po,
and gives us an explicit decomposition ker(di) = im(dti_1) ® ker(Di).
Thus, we may identify the kernel of the Dirac operator d + d* with He (d).
The dimension of the kernel of d + d* on the space of even-degree sections
Ei r(M, £2i) is equal to >i dim(H2i(d)), while the dimension of its kernel
on Ei r(M, £2i+1) is equal to Ei dim(H2i+1(d)). Thus, the index of d + d*
equals
(dim(H2 (d)) - dim(H2i+1(d))),
which equals the Euler number of (£, d) by definition.
Corollary 3.55. The kernel of the Laplace-Beltrami operator dd* + d*d on
Ai(M) is naturally isomorphic to the deRham cohomology space Hi(M,R).
The index of the Dirac operator d+d* on A(M) is equal to the Euler number

Eul(M) _ (-1)i dim(Hi(M,R))

i=1

of the manifold M.
The Euler number defined using the de Rham cohomology is equal to the
Euler number defined in Theorem 1.56 by means of a vector field with non-
degenerate zeroes; this may be shown using Morse theory, but would take us
too far afield to discuss here.
Corollary 3.56 (Kiinneth). If M and N are two compact manifolds,
He (M x N) ^_' He (M) (9 He (N).
Proof. Choose Riemannian metrics on M and N, with Laplace-Beltrami op-
erators OM and AN. The Laplace-Beltrami AMxN on M x N with the
product metric is AM ®1 + 1® AN. Using the spectral decomposition of the
positive operators OM and AN, we see that the zero eigenspace of AMXN is
the tensor product of the zero eigenspaces of OM and AN.
130 3. Clifford Modules and Dirac Operators

The Lichnerowicz formula for the square of a Dirac operator, when applied
to the operator d+d*, gives a formula for the Laplace-Beltrami operator which
is known as Weitzenbock's formula. Let us denote by LATM the Laplacian
of the exterior bundle AT*M. The bundle AT*M is the bundle associated
to the principal bundle GL(M) by the representation A : GL(n) --> A(Rn)*
of (1.26), so that the curvature of the Levi-Civita connection on the bundle
AT*M is given by the formula

(3.15) ,\(R) = ek A el E RijklE(ei)t(ej)

k<l ii
Hence by (3.14),

(d + d*)2 = QAT*M + 2 > Rijkl (Ek

ijkl
- tk)(El
Using the fact that the antisymmetrization of Rijkl over three indices vanishes,
we see that
EiEjEktlRijkl = 0,

bid ektlRijkl = 0.

Retaining only the non-zero terms, we obtain Weitzenbock's formula: the

decomposition of the Laplace-Beltami operator as a generalized Laplacian is

(3.16) (d + d* )2 _ OAT*M _ E RijklektlE269

ijkl
Observe that on the space of one-forms, the Weitzenbock formula becomes
(d + d*)2 = AAT*M + Ric,
where the Ricci curvature is identified with a section of the bundle End(T*M)
by means of the Riemannian metric on M. This follows from the following
formula,
RijklEat9Ektl
=- RijkleZEkd tl - Rlcij E2d,

where the first term on the right-hand side vanishes on one-forms. In partic-
ular, if the manifold M has Ricci curvature uniformly bounded below by a
positive constant c, then the Laplace-Beltrami operator is bounded below by
c on one-forms, since the operator OAT *M itself positive. By Hodge's The-
orem, this implies that the first de Rham cohomology space of the manifold
vanishes; the Hurewicz theorem then tells us that the manifold has finite fun-
damental group. This result is a simple example of the so-called vanishing
theorems that follow from lower bounds for the square of the Dirac opera-
tor, and which are proved by combining assumptions on curvature and the
Lichnerowicz theorem.
3.6. Some Examples of Clifford Modules 131

There is a generalization of the de Rham complex, in which we twist the

bundle of differential forms by a flat bundle W, with covariant derivative V:

0,A°(M,W) -Vo 4'(M, W) 111+A2(M,W) 122-.

This sequence is a complex precisely when V2 = 0, that is, when the bundle
W is flat. There is a simple way to characterize flat bundles, as an associated
bundle M W, where W is representation of iri(M). Here, M is the
covering space of M, which is a principal bundle with structure group iri (M)
the fundamental group of M, acting on M by means of the so-called deck
transformations.
The Signature Operator. There is a variant of the first example, in which
the Dirac operator is the same but the definition of the Z2-grading on the
Clifford module AT*M is changed. In order to define this new grading, we
use the Hodge star operator, which is defined on an oriented Riemannian
manifold.
Definition 3.57. If V is an oriented Euclidean vector space with complex-
ification VC, the Hodge star operator * on AVC equals the action of the
chirality element I' E C(V) ® C, defined in Proposition 3.17, on the Clifford
module AVC.

This definition differs from the usual one by a power of i in order that
*2=1.
Applying this operator to each fibre of the complexified exterior bundle
ACT*M = AT*M ®R C of an oriented n-dimensional Riemannian manifold
M, we obtain the Hodge star operator
*:A CT*M -> A'-k T*M.
Proposition 3.58. (1) If a E A1(M, C), then * 1e(a)* = (-1)nt(a).
(2) If a and Q are k-forms on M, then

f a A *p =
M
(-1)kn+k(k-1)/2 ip fM(a,,3)dx,

where dx is the Riemannian volume form on M corresponding to the

chosen orientation, and p = n/2 for even and (n + 1) /2 for n odd.
(3) The operator d* equals (-1)n+l *-1 d* on the space of k -forms.
In these formulas, we can replace *-1 by *, since *2 = 1.
Proof. (1) If ei is an orthonormal frame of V, we may assume that a = en,
and then we see that
e(en)c(el) ... c(en) = (-1)n-'c(ei) ... c(en-i)s(en)c(en)
But for any v E V, we have e(v)c(v) = e(v)(e(v) -a(v)) _ -(e(v) - a(v))c(v),
from which the formula follows.
132 3. Clifford Modules and Dirac Operators

(2) Let a,,3 E AkV. We will prove by induction on k that

a A*0 = (-1)kn+k(k-1)/2ip(a,0)el
A -- A en.
For k = 0, )3 E AO V = C, we have by definition
*p = iP)3el A ... A en.
If aEAkV,/3 Ak+1V and e E V, we have
aAeA(*,(3) _ (-1)naA*t(e))3
(-1)n+kn+k(k-1)/2ip(a, /
/ t(e),Q)el A ... A en
_ (-1)n+kn+k(k-1)/2ip(e(e)a,,3)e1
A- A en
(-1)n+kn+k(k-1)/2(-1)kip(a
A e, Q),

which is the desired formula for k + 1.

(3) To prove this, we simply insert (2) in the formula which defines d*, and
apply Stokes' formula: if a E Ak (M) and ,3 E Ak+1(M), then

(a, d*,l3) dx = (da,,3) dx = (-1)(k+1)n+k(k+1)/2i-p da A*Q

M fm fM
= _ ( -1)k+(k+l)n+k(k+1)/2i-p J and * Q
M

= (a,*d*0)dx.

Corollary 3.59 (Poincare duality). Let M be an oriented compact n-dim-

ensional manifold. The bilinear form fm a A )3 induces a non-degenerate
pairing
Hc(M) X Hn-k(M) -> R.
Proof. Let a E .Ak (M) be a harmonic form representing an element of Hk(M);
thus, *a is closed. If fm a A /3 = 0 for all )3 E H'-k (M), then in particular
fm a A (*a) = 0. It follows that fM Ial2 dx = 0, and hence that a = 0.
Suppose that n = dim(M) is even. Then the operator * anticommutes with
the Dirac operator d+d* = d-*d*, hence with the Laplace-Beltrami operator.
Since *2 = 1 and * anticommutes with the action of c(v) = s(v) - t(v), we
can define another Z2-grading on the exterior bundle AT*M ® C using the
operator *; the differential forms satisfying *a +a (respectively *a = -a)
are called self-dual (respectively anti-self-dual). If the dimension of M is
divisible by four, F is actually an element of the real Clifford algebra, so *
operates on the space of real valued forms.
Note as a special case of Proposition 3.58 that if dim(M) = n is even and
a and 0 E An/2(M), then
/M
fM a A,3 = ep (a, *Q) dx,
fm
3.6. Some Examples of Clifford Modules 133

where ep=1 ifn=4l andEp=i ifn=4l+2

Let us recall the definition of the signature of a quadratic form.
Definition 3.60. Let Q be a quadratic form on a real vector space V.
Choose a basis in which
Q(x) = x 2l + ... + x 2p - x 2p+l - ... - xp+q-
2

The signature a(Q) is the number p - q, which depends only on the quadratic
form Q.
If M is a manifold of even dimension n, the bilinear form on Hn/2(M, R)
defined by (a,,3) --> fm a A# satisfies

f a A 0 = (-1)n/2J )ina
M M
Thus, if the dimension of M is divisible by four, this quadratic form is sym-
metric. The signature of M is by definition the signature of the restriction of
this quadratic form to the space Hn'2(M,]R), and we denote it by o(M).
If M is an oriented compact Riemannian manifold of dimension divisible
by four, the signature operator of M is the Dirac operator d + d* for the
Clifford module AT*M with 7L2-grading coming from the Hodge star *.
Proposition 3.61. The index of the signature operator d + d* equals the
signature a(M), so by the McKean-Singer formula,
a(M) = Tr(*e-t°),
where A is the Laplace-Beltrami operator.
Proof. Let ai E A' (M) be a basis for the harmonic forms fk, where k < n/2.
Note that the differential forms at = ai ±*ai form a basis for the direct sum
of the spaces of harmonic forms ?-lk®ln-k since * maps the space of harmonic
forms into itself. Now, at is self-dual and a- is anti-self-dual, thus the index
of the Dirac operator d + d* restricted to Ak(M) ® An-k(M) is zero.
Now consider the harmonic forms of degree n/2. We may find a basis ai
such that *ai = ai for i < m and *ai = -ai for i > m; it is clear that
the index of d + d* with respect to the Z2-grading defined by * is 2m - n.
Since n is divisible by four, the signature form may be rewritten as follows:
if a E H n/2 satisfies *a = ±a, (then

aAa=J (a,*a)dx = f ja12dx;

JM M IM
from this we see that the signature of the manifold is equal to 2m - n as
well.
If W is a Z2-graded vector bundle on M with connection, we may twist
the Dirac operator d + d* by W to obtain a twisted signature operator on the
tensor product AT*M ® W, whose index is denoted a(W). It is clear from
134 3. Clifford Modules and Dirac Operators

the homotopy invariance of the index that the signature Q(W) is an invariant
which depends solely on the manifold M, its orientation, and the vector bundle
W.
The Dirac Operator on a Spin-Manifold. Let S be the spinor bundle
over an even-dimensional spin manifold M. The most basic example of a
Dirac operator is the Dirac operator on S associated to the Levi-Civita con-
nection Vs on S; this operator is often referred to as the Dirac operator
on M. A section of the spinor bundle S lying in the kernel of D is known
as a harmonic spinor. More generally, we may consider the Dirac operator
Dw®s on a twisted spinor bundle W (&S with respect to a Clifford connection
of the form
Vw®s = Ow ®1 +1® Os. If F1^' is the curvature of V'", then the twisting
curvature FE1 s of Ol'w®s equals Fw, and the Lichnerowicz formula becomes
D2 - AW®s + r4
W®s - FW(ei, ej)c(e2)c(ej) +
i<j
In particular, for the spinor bundle S itself, we have
D2=s+r4
Lichnerowicz used this formula to prove the following result.
Proposition 3.62. If M is a compact spin manifold whose scalar curvature
is non-negative, and strictly positive at at least one point, then the kernel of
the Dirac operator on the spinor bundle S vanishes; in particular, its index
is 0.
Proof. Ifs E r(M,S), then

f (D2s,s)dx=fMIIVs1I2dx+4
M
f MrMIISII2dx.
Thus, if Ds = 0, we obtain Vs = 0 and rms = 0. Then (s, s) is constant on
M, hence s = 0.
On a spin-manifold M, the signature operator of the last section may be
rewritten as a twisted Dirac operator in an explicit way. Indeed, from the
isomorphism of Clifford modules
AT*M®C=C(M)®C=End(S)=S*®S,
we see that the Dirac operator d+d* on AT*M is the Dirac operator obtained
by twisting the Dirac operator on the spinor bundle S by the bundle S*. The
two different Z2-gradings on AT*M correspond to the two different ways of
grading the twisting bundle S*: if we treat it as a 7L2-graded bundle, we
recover the grading of AT*M according to the parity of the degree, while if
we treat S* as an ungraded bundle, we recover the grading coming from the
Hodge star *.
3.6. Some Examples of Clifford Modules 135

The a-Operator on a Kahler Manifold. Our last example of a Dirac

operator is important in algebraic geometry.
Definition 3.63. An almost-complex manifold M is a manifold of real
dimension 2n with a reduction of its structure group to the group GL(n, C);
that is, there is given a GL(n, C)-principal bundle P and an isomorphism
between the vector bundles P xGL(n,c) Cn and TM.
An equivalent way of defining an almost-complex structure is to give a
section J of End(TM) satisfying j2 = -1.
On an almost-complex manifold, the complexification of the tangent bundle
splits into two pieces, called the holomorphic and anti-holomorphic tangent
spaces,
TM OR C=T1'0M®T°'1M,
on which J acts by i and -i, respectively. In terms of a local frame Xzf YY,
(1 < i < n), such that JXi = YZ and JYa = -Xi, we see that T1'0M is
spanned by Wj = (Xj - iY ), and T°'1M is spanned by Wj = (Xj + iYj ).
It follows from this splitting that if W is a complex vector bundle on M,
the space of differential forms is bigraded,
A(M W) = E Ap,q(M W) = r(M, Ap(Tl'0M)* ® Aq(T°'1M)* ® W).
O <p, q <n

A Hermitian structure on an almost-complex manifold M is a Hermitian

metric on the tangent space TM. Underlying the Hermitian structure is a
Riemannian structure, and its associated Levi-Civita connection. However,
in general the Levi-Civita connection does not preserve T°"1M and T1'°M.
If M is a Hermitian almost-complex manifold, and W a vector bundle on M,
then the vector bundle of anti-holomorphic differential forms A(TO,'M)* ®W
is a Clifford module, with the following action: if f E A' (M) decomposes as
f = f1,o + f0,1 with f1'0 E A1,O(M) and f°,1 E A°'1(M), then its Clifford
action on v E r(M, A(TO,1M)* ®-W) equals

c(f)v = v2(e(f"0) - b(f°'1))v.

Note that if W is Hermitian, e(f 1,o)_t(fo,i) is skew-adjoint, since f°'1 = T0,
and the Clifford action is self-adjoint.
From now on, we will assume that M is a complex manifold, in other
words, there is a collection of local charts for M parametrized by an open
ball in Cn for which the transition functions are holomorphic. Let zi be a set
of local holomorphic coordinates. Then T1"0M is spanned by the vectors &,'i,
while T°"1M is spanned by the vectors az,. The elements in AA'q(M) have
the form
E A ... A dzip A dzjl n ... A dzja
On a complex manifold, the exterior differential d splits into two pieces,
d = a + a, where a increases the degree of a form (p, q) by (1, 0) and 6
136 3. Clifford Modules and Dirac Operators

increases it by (0, 1), and 82 = as + a8 = 82 = 0. Both of these operators

satisfies Leibniz's rule:
8(an,0) =asAQ+(-1)1a1aA8Q,
a(anQ) asn0+(-1)lalan00.
Now let W be a complex vector bundle over M with covariant derivative
V. With respect to the decomposition T*M OR C = (Tl'OM)* ® (TO'1M)*,
V splits into two pieces V = 01'0 + V°'1, which satisfy the following Leibniz
formulas: if a E A(M) and v E A(M, W), then
V"°(a A v) = as A v+(-1)lala A V',Ov,
V°11(a A v) = as A v+(-1)lala A V°'lv.

Definition 3.64. A holomorphic vector bundle on a complex manifold M

is a complex vector bundle W which has a chart, that is, a covering of M by
open sets Ui and framings Oi : WijUi --+ Ui x CN, in which the transition
functions Oi o 0.- 1 : Ui n Uj - GL(n, C) are complex analytic.
If W is a holomorphic vector bundle over M, there is a unique operator
8 :.A.P'9 (M, W) AP'Q+l (M, W) which in each coordinate patch of a chart
with complex analytic transition functions equals
TL

E,5 (d2z)
i=1
az
The sheaf O(W) of holomorphic sections of a holomorphic bundle W is the
kernel of a acting on the sheaf r(M, w) = .A°'0(M, W) of smooth sections of
W. We say that a covariant derivative V on a holomorphic vector bundle is
holomorphic if V°'1 = a.
Proposition 3.65. If W is a holomorphic vector bundle with a Hermitian
metric h(s, t), there exists a unique holomorphic covariant derivative on W
preserving the metric.
Proof. If V is holomorphic and preserves the Hermitian metric h, then taking
the component of dh(s,t) lying in A°'1(M), we see that
ah(s, t) = h(V°'1s, t) + h(s, O1'0t).
If moreover V°'1 = 8, this equation determines V1'0. It is then easy to check
that V = O''0 + V°" is a covariant derivative.
Let V be a holomorphic covariant derivative on W which preserves a
Hermitian metric h. Since the curvature is skew-adjoint as a section of
A2T *M ® End (W) , we see that (01'0) 2 = 0, so that the curvature of V is
the section of A1'1(M,End(W)) equal to [V',o, V0,1]
3.6. Some Examples of Clifford Modules 137

It is interesting to see how the Hermitian metric and holomorphic covariant

derivative corresponding to it are related in a chart. If the Hermitian metric
on W is given at a point z E M by a Hermitian matrix h(z), then the
corresponding covariant derivative VW equals
Vw =d+h-lah.
Its curvature is the (1, 1)-form
Fh = hylaah - (h-16h) A (h-'A).
If W is a line bundle, h is a positive real function and the curvature is thus
Fh = aalog(h).
If s is a non-vanishing holomorphic section of W over the chart, we see that
Fh = 19a1og(IIsII2)
If M is a complex manifold with a Hermitian metric h, the real part of h
restricted to TM is a Riemannian metric on M, while the imaginary part SZ
of h restricted to TM is a two-form on M: if X and Y are vector fields on
M, then
f (X,Y) = Imh(X,Y) = Imh(Y,X) = -Imh(Y,X) = -1l(Y,X).
If X and Y are vector fields on M, we see that g(X, Y) = 1(JX, Y), where
J is the almost-complex structure on M.
We call a complex manifold M Kahler of SI is a closed two-form. The
following result gives another characterization of Kahler manifolds.
Proposition 3.66. A complex Hermitian manifold M is Kahler if and only
if the bundles T1"°M and T°"M are preserved by the Levi-Civita connection
V of the underlying Riemannian structure, or in other words, if VJ = 0,
where J is the almost-complex structure of M.
Proof. Using the fact that SI(J , ), we see that
((V x J)Y, Z) = (V X (JY), Z) + SZ (V x Y, Z)
=-XSZ(Y,Z)+SZ(V Y,Z)+SZ(Y,V Z)
= -(V SS)(Y,Z).
Since the Levi-Civita connection is torsion free, the right hand side of this
equation when antisymmetrized over {X, Y, Z} is nothing but dSZ(X, Y, Z).
Thus, if V J vanishes, we see that dS2 = 0.
Using the explicit formula for the Levi-Civita connection (1.18), we com-
pute that in a holomorphic coordinate system zi

2(vax; aZk, axL) = 0,

2 (V a=; azk , axe) = axk (az , a ;) - az (az; , azk) = i dSZ(azk, az; >,
2(paj; axk , aSI) = az; (,9.., ,92,) - az (az; ,axk) = i dSt(az; , axk, az ).
138 3. Clifford Modules and Dirac Operators

Thus we see that if dfl = 0, the Levi-Civita covariant derivative preserves sec-
tions of T1 "0M and induces on the complex vector bundle T1>0M the canonical
holomorphic covariant derivative.

This result has the following beautiful consequence.

Proposition 3.67. Let W be a holomorphic vector bundle with Hermitian

metric on a Kahler manifold. The tensor product of the Levi-Civita con-
nection with the canonical connection of W is a Clifford connection on the
Clifford module A(T0,1M)* ®W, with associated Dirac operator v"2_(6+8*).

Proof. It is clear that the Levi-Civita connection is a Clifford connection; a

quick look at the formula for the Clifford action of T*M on A(T°.1M)* shows
that what is needed is that the splitting TM ®R C = T',OM ® T°"M is
preserved by the Levi-Civita connection, in other words, that VJ = 0.
If Zi is a local frame of T1"0M, with dual frame Zi E (Tl.OM)*, then

d= E(Z')Vz, +E(Zi)Vz,
i
from which we see that
a = YE(Zi)vzs.

If we choose a holomorphic trivialisation of W over an open set U C M, so

that W j U = U X CN, the operator &w may be identified over U with the
operator a on A(U, CN). On the other hand, if V is the tensor product of the
Levi-Civita connection of A(Tl,OM)* with the holomorphic connection of W,
then the operator
E(Zt)V 2.,

also coincide with a in the holomorphic trivialisation of W over U, and thus

must be equal to 6W.
It remains to show that

Let a be the one-form on M such that for any /jq E Ar'q (M, W) and /3q+1 E
,,4n,q+l (M W)

a(X) = (Oq, t(X0'1))3q+1)

Since V preserves the splitting TM ®R C = T1"0M ® T°>1M, the contraction
Tr (Va) of the two-form Va with the Hermitian structure on TM ORC equals

Tr(Va) = E Zia(22) - a(Vz.22),

i
3.6. Some Examples of Clifford Modules 139

where Zi is a local orthonormal frame of T 1'0 M with respect to the Hermitian

structure on T1'°M. Thus
\1
= - (I9, / q
(( e(ZZ)OZ )Nq, Qq+1 l
x \ ( L(2 )OZ' )Nq+l x + Tr(Vo
+ .

i i
The last term vanishes after integration over M and we obtain the desired
formula for a*, from which it is clear that ,/(a + 0*) = c o V.
The cohomology of the complex

0 -> A°'0(M,W) A°'1(M,W) A°'2(M,W) -

is called the Dolbeault cohomology of the holomorphic bundle W. Dol-
beault's theorem shows that this is isomorphic to the sheaf cohomology space
H* (M, CA (W)). Hodge's theorem gives an explicit realization of this cohomol-
ogy as the kernel of the Laplace-Beltrami operator on A°'i(M,W).
Theorem 3.68 (Hodge). The kernel of the Dirac operator v f2-(6 + a*) on
the Clifford module A(T°'1M)* ® W is naturally isomorphic to the sheaf co-
homology space H* (M, 0(W)).

Corollary 3.69. The index of a+a* on the Clifford module A(T°'1M)* ®W

is equal to the Euler number of the holomorphic vector bundle W:
ind(a + a*) = Eul(W) = T(-1)i dim(Hi(M, O(W))).
The Lichnerowicz formula for the square of the operator a+a* on a Kahler
manifold is called the Bochner-Kodaira formula. Define the generalized Lapla-
cian 0°' on A°' (M, W) by the formula

f s) dx =f (V°'ls, V°'ls) dx.

Lemma 3.70. If Zi is a local orthonormal frame of T1'0M for the Hermitian

metric on M, we have the formula
O °,'_- (VziOZi - OvZiZi).

Proof. Introduce the (1,0)-form a(X) = (Vxi,os, s). Then

Tr(Da) = E Zia(ZZ) - a(OziZi),
i
since T1'0M is preserved by the Levi-Civita connection. Thus

(-E(VziVZi - VvZizi)s, s) _ (V°'1s, V°'ls)X - Tr(Va)x.

i
The lemma follows by integrating over M, using Proposition 2.7.
140 3. Clifford Modules and Dirac Operators

The canonical line bundle of a Kahler manifold is the holomorphic line

bundle K = Al (Tl,OM)* on M. The curvature of K* = AnT1>0M with
respect to the Levi-Civita connection is the (1,1)-form E(RZi, Zi), where R
is the Riemannian curvature of M.
Proposition 3.71 (Bochner-Kodaira). Let W be an Hermitian holomor-
phic vector bundle over the Kihler manifold M. In a local holomorphic co-
ordinate system,
aa* + a*a = SO,* + E E(dz)t(dzj )Fw®K` (aZj, azi ).
ij
Proof. The proof is similar to the proof of Lichnerowicz's formula. Writing
a = E e(dz)Va/azi and a* = - E L(dzi)Va/azi, we see that
aa* + 6*8 = SO,* + E(dzi)t(dzj)(R+(azj, azi) + Fw(azj, aZi))
ij

where R+ is the curvature of A(TO,1M)*. Thus it only remains to show that

E(dzi)c(dzj)R+(azj,aZi) _ E(dzi)t(dzj)FK`(0Z3,0Zi).

The left side is equal to

E E(Zf)t(Z$)E(Z!)t(Zk*)(R(Zj, Zi)Zk, Zl) _

ijkl

-(Zi)i(Zk*)(R(Zj, Zi)Zk, Zj).

ijk
Using the fact that the Levi-Civita connection is torsion-free, we see that
A(Zj, Zi)Zk + R(Zk, Zj)Zi + R(22, Zk)Zj = 0,
and the fact that R(Zk, Zj) = 0, we see that
R(Zj, Zi)Zk = R(Zk, 22)Zj.
It follows that
E E(2i)t(Zk*)(R(Zj, Zi)Zk, Zj) _ E(2i)t(ZZ)F" (Zk, Zi),
ijk ik

completing the proof of the theorem.

This formula implies some important vanishing results. If L is a Hermitian
holomorphic line bundle with curvature F = E Fijdzi A dzj, we say that L
is positive if the Hermitian form v ,--> F(v, v) on T1"0M is positive. This
positivity condition is related to the existence of global holomorphic sections
of L. For example, if s is a global holomorphic section of W, and x is a
point in M at which Is 12 attains a strict maximum, then it is easy to see that
Fx = 5a log IsI2 is positive on TOM.
Bibliographic Notes 141

Proposition 3.72 (Kodaira). (1) If G is a Hermitian holomorphic line

bundle on a compact Kahler manifold such that G ® K* is positive, then

HZ(M, CM(G)) = 0 fori > 0.

(2) If L is a positive Hermitian holomorphic line bundle and W is a Hermi-
tian holomorphic vector bundle on M, then for m sufficiently large

Ht'(M, o(I (&W)) = 0 for i > 0.

Fn®K.
Proof. (1) Denote by _ FZ ®K*dza A the curvature of L ®K*
d2.7-

®K.
and by A(Fc®K*) the endomorphism >1j e(dzz)t(dzj)F of A$(T°°1M)*.
Then the Bochner-Kodaira formula shows that

IM ((aa* + 6*8)a, a) = f f(V°"la,V°'la) + fm (A(F'®K*)a, a)

> M (A(FG®K)a, a)

The right-hand side is strictly positive if i > 0 and a # 0; thus the conditions
as = 0 and 8*a = 0 imply that a = 0.
(2) For the bundle CI ® W, the Bochner-Kodaira formula gives

aa* + a*a = A" + ma(F'C) + other curvature terms independent of m.

For m sufficiently large the term mA(F4) dominates the other curvature terms,
and we obtain (2).

The reason for the importance of this result is that it shows that for m
large, dim(H°(M, 0(C (9 W))) = Eul(Cm ® W). Riemann-Roch-Hirzebruch
theorem that we prove in the next chapter gives an explicit formula for
Eul(Cm ® W) and hence for dim(H°(M, 0(Gm (9 W))). This combination
of an index theorem with a vanishing theorem is a very powerful technique in
differential geometry.

Bibliographic Notes

The book of Lawson and Michelson [79] is a valuable reference for many of
the topics in this chapter.

Section 1. The theory of Clifford algebras and their representations is clas-

sical; we mention as references Chevalley [49], Atiyah-Bott-Shapiro [9], and
Karoubi [71].
142 3. Clifford Modules and Dirac Operators

Sections 2 and 3. The characterization of spin manifolds in terms of w2(M) of

Proposition 3.34 is proved by Borel and Hirzebruch [38] (see also Milnor [84]
and Lawson and Michelson [79]). The Dirac operator on spin manifolds ap-
pears in Atiyah-Singer [13]; they introduced the idea of twisting a Dirac oper-
ator. Lichnerowicz also studied this operator in [80], proving the fundamental
formula for its square; he also proves the vanishing theorem for harmonic
spinors on manifolds of positive scalar curvature.
Sections 4 and 5. The supertrace formula for the index of a Dirac operator is
formulated by Atiyah-Bott [5] in terms of the zeta-function of D2; the intro-
duction of the heat kernel in this context is due to McKean-Singer [82].
Section 6. We do not attempt here to give references to the original articles
on the geometric Dirac operators discussed in this section. However, there are
some good textbooks on this subject, among which we mention Gilkey [66],
Lawson and Michelson [79] for Dirac operators, and Wells [104] for the 9-
operator. Many examples are contained in the article of Atiyah-Bott [6].
Chapter 4. Index Density of Dirac Operators

Let M be a compact oriented Riemannian manifold of even dimension n.

Let D be a Dirac operator on a Clifford module £ on M associated to a
Clifford connection V. Since D is an elliptic operator, the Atiyah-Singer
Index Theorem gives a formula for the index of D in terms of the Chern
character of £ and the A-genus of M. In this chapter, we will give a proof
of this formula which uses the asymptotic expansion of the heat kernel of D2
for t small. This approach was suggested by Atiyah-Bott [5] and McKean-
Singer [82], and first carried out by Patodi [89] and Gilkey [64]. However, our
method differs from theirs, in that the A-genus and Chern character emerge
in a natural way from purely local calculations, while previous proofs all
required that the index density be determined by calculating the index of a
range of examples.
We will prove a refined form of the local index theorem mentioned above.
We define a filtration on sections of the bundle End(£) over M x R+. We take
the heat kernel of the operator D2 restricted to the diagonal, and calculate
it explicitly modulo lower order terms with respect to this filtration. In this
way, we obtain a generalization of the local index theorem which extracts a
differential form on M from the heat kernel for small times; this differential
form is nothing other than the product of the A-genus A(M) with respect to
e_Fels
the Riemannian curvature and E A(M, Endc(M) £), the exponential
of the twisting curvature of E.

4.1. The Local Index Theorem

Let kt(x, x) = (x I e-tD2 I x) be the restriction of the heat kernel of D2 to the
diagonal. In this chapter, we will calculate the leading term in the asymptotic
expansion of Str(kt(x, x)) as t --+ 0. In particular, when combined with the
McKean-Singer formula, this will imply a formula for the index of a Dirac
operator.
The heat kernel kt(x,x) is a section of the bundle of filtered algebras
End(S) = C(M) ® EndC(M)(£), where the filtration is induced by the fil-
tration of C(T*M) and elements of Endc(M)(5) are given degree zero. We
denote by Ci(M) the sub-bundle of C(M) of Clifford elements of degree less
than or equal to i. The associated bundle of graded algebras is the bundle
144 4. Index Density of Dirac Operators

AT*M ® EndC(M) (£). Let

R/2
A(M) = det1/2
sinh(R/2)
be the A-genus form of the manifold M with respect to the Riemannian cur-
vature R.
Let Fs/S be the twisting curvature of the connection ve, which was defined
in (3.43). The aim of this chapter is to prove the following theorem.
Theorem 4.1. Consider the asymptotic expansion of kt(x,x),
W
kt(x,x) - (47rt) -n/2 E tiki(x)
ii=OO

with coefficients ki E I'(M, C(M) ® EndC(M)(£)).

(1) The coefficient ki is a section of the bundle C2a(M) ®Endc(M)(E).
i=n/2
(2) Let u(k) _ E 0-2i(ki) E A(M,Endc(M)(£)). Then
i=O

v(k) = det1/2 (sinh(R2/2) )

exp(-Fe/S).

This result need not hold for a Dirac operator associated to a general
Clifford superconnection; for a discussion of the modifications which must
be made in the general case, see the Additional Remarks at the end of this
chapter.
The i = 0 piece of Theorem 4.1 is precisely Weyl's formula
lim (47rt)n/2kt(x, x) = Ide .

t-+o
Thus, Theorem 4.1 is an extension of Weyl's theorem in the case of a general-
ized Laplacian which is the square of a Dirac operator associated to a Clifford
connection.
The idea of the proof is to work in normal coordinates x E U around a
point xo E M, where U is an open neighbourhood of zero in V = T"'M.
Using the parallel transport map for the connection De, we may trivialize
the bundle E over the set U, obtaining an identification of I'(U, End(s)) with
C°O(U, End(E)), where E _ £, ,. There is an isomorphism o of the vec-
tor space End(E) = C(V*) ® Endc(v.)(E) with AV* 0 Endc(v.) (E). We
introduce a rescaling on the space of functions on R+ x U with values in
AV* ® EndC(v.) (E) by the formula
n
(&a) (t, x) = E u-i/2 a(ut, u1/2x)
i=0
4.1. The Local Index Theorem 145

Using this rescaling, we may restate Theorem4.1 as follows: if k(t,x) _

o ((exp,o x I e-tee I xo)), then

slim (un/2Suk)I(t,x)=(l,o) = (47r)-n1'2A(M)exp(-FE1s).

The rescaling operator 6, induces a filtration on the algebra of differential

operators acting on C°°(R+ x U, AV* ® Endc(v.) (E)); an operator D has
filtration degree m if
lira umI2S,DSu 1
u-*0
exists. Thus, a polynomial P(x) has degree - deg(P), a polynomial P(t)
has degree -2 deg(P), a derivative 8/3xi has degree one, a derivative 7/8t
has degree two, an exterior multiplication operator ei has degree one, and
an interior multiplication operator ti has degree -1. It will be shown, using
Lichnerowicz's formula, that the operator D2 has degree two; this is why
Theorem 4.1 holds. Indeed, up to operators of lower order, we will see that
the operator D2 may be identified with a harmonic oscillator with differential
form coefficients.
Given r and f E IR, let H be the harmonic oscillator on the real line
d2 r2x2
H=-dx2+ 16 +
In Section 3, we will prove Mehler's formula, which states that the heat kernel
pt (x, y; H) of the harmonic oscillator H is equal to

tr/2
exp(-$t (coth(tr/2)(x2+ y2)-2cosech(tr/2)xy) -tf).
(47rtsinh(tr/2))
The above filtration argument reduces the calculation of (S,4k)(t,x) to the
calculation of the heat kernel of a harmonic oscillator, with differential form
coefficients, in which r is replaced by the Riemannian curvature R of M and
f by the twisting curvature Fe/s of S. The A-genus of the manifold M and
the twisting curvature Fe/s appear as consequences of the above formula for
Pt (0, 0; H).
In the rest of this section, we will discuss some of the consequences of this
theorem. The main consequence of Theorem 4.1 is that it provides the local
index theorem for Dirac operators, which is a strengthening of Atiyah and
Singer's result. By the McKean-Singer formula, for every t > 0,

ind(D) = Str (kt (x, x)) dx.

IM
Theorem 4.1 implies that the integrand itself has a limit when t converges
to zero, if D is associated to a connection. Indeed, as the supertrace of an
146 4. Index Density of Dirac Operators

element a E C(M) vanishes on all elements of Clifford filtration strictly less

than n = dim(M), the first part of Theorem 4.1 implies that

Str(kt(x, x)) - (47rt)

-n/2 t Str(ki(x)),
i>n/2
thus there are no poles in the asymptotic expansion of Str(kt (x, x)). Further-
more, as the left-hand side ind(D) of the McKean-Singer formula is indepen-
dent of t, we necessarily have

ind(D) = (47r)-n/2 JM Str(kn/2(x)) dx,

while the integrals of all other terms fm Str(kj (x)) dx, j n/2, vanish.
To identify the term Str(kn/2(x)) as a characteristic form on M, we must
introduce some notation. If we write a Clifford module £ locally as W ® S,
where S is a spinor bundle and W is an auxiliary bundle, then by (3.28), for
a E r(M, End(W)) = r(M, Endc(M) (£)),
Strw (a) = 2-n/2 Str& (ra),
where r E r(M, C(M)) is the chirality operator. We will avoid making
use of this decomposition, which is often only possible locally. Instead, we
use the relative supertrace of Definition 3.28; the relative supertrace of a E
r(M, Endc(m) (£)) is
Stre/s(a) = 2-n/2 Stre(ra).
We extend the relative supertrace to a linear map
Str8/s : A(M, Endc(M)(S)) -* A(M),
and define, for a Clifford superconnection VE on E, the relative Chern
character form of the bundle £ by the formula
ch(£/S) = Stre/s(exp(-Fe/s)).

The relative Cherri character ch(£/S) is a closed differential form on M.

The cohomology class ofthis differential form is independent of the choice of
Clifford superconnection on £, and will also be denoted by ch(£/S). In the
special case where the Clifford module £ = W®S and Ve is the tensor product
of the Levi-Civita connection on S and a connection on W with curvature Fw,
we see that ch(£/S) = Strw(exp(-Fw)) is the Chern character form of the
twisting bundle W.
If a E r(M, C(M)) and b E r(M, Endc(M)(E)), then the point-wise super-
trace of the section a ® b E r(M, C(M) ® Endc(M) (£)) r(M, End(S)) was
shown in (3.21) to equal the Berezin integral
Stre(a(x) (9 b(x)) = (-2i),/2a"(a(x)) StrE/s(b(x)),
4.1. The Local Index Theorem 147

so that
Str£(kn/2)(x) = (-2i)n/2 Strs/s(0-n(kn/2(x))

Thus Theorem 4.1 implies the following theorem for the index of a Dirac op-
erator associated to a Clifford connection.

Theorem 4.2 (Patodi, Gilkey). Let M be a compact oriented Rieman-

nian manifold of even dimension n, with Clifford module £ and Clifford con-
nection Ve, and let D be the associated Dirac operator. If kt(x, x) IdxI E
r(M, End(E) (9 JAI) is the restriction of the heat kernel of the operator D2 to
the diagonal, then limt_,o Str(kt(x, x)) IdxI exists and is the volume form on
M obtained by taking the n-form piece of

(27ri)-n/2 det1/2
(sinh (R/2)) Strs/s(eXP(-F£/s))
It is important to note that there is no a priori reason to suppose that
the limit limt._,0Strkt(x,x) exists at each point, and indeed, this is not
necessarily true for Dirac operators which are not associated to a Clifford
connection. However, since the index of a Dirac operator is independent of
the Clifford superconnection used to define it, we obtain the Atiyah-Singer
formula for the index of an arbitrary Dirac operator.

Theorem 4.3 (Atiyah-Singer). The index of a Dirac operator on a Clif-

ford module £ over a compact oriented even-dimensional manifold is given
by the cohomological formula

(2iri)-n/2
ind(D) = A(M) ch(£/S).
JM
The factor (27ri)-n/2
is absent from the usual statement of this theorem,
since it is part of the topologist's definition of the characteristic classes, the
purpose of which is to ensure that the characteristic Chern classes are integral.
However, in this book, we prefer the more natural (to the geometer) definitions
of the A-genus form

R/2
A(M) = det1/2
sinh(R/2)

and of the Chern character form ch(£) = Str(exp(-V2)).

In the rest of this section, we will derive the usual formulas for the index
density (27ri)-n/2A(M) ch(S/S) for each of the four classical Dirac operators
that were described in the last chapter. The resulting formulas are special
cases of the Atiyah-Singer Index Theorem for elliptic operators.
148 4. Index Density of Dirac Operators

The Gauss-Bonnet-Chern Theorem. We consider the Clifford bundle

e = AT*M with its Levi-Civita connection V. In order to analyse what
the above index theorem says for the case of the de R.ham operator, we must
calculate Fe/s in this case. We will do this in a local orthonormal frame ei.
We start with some general results about the Clifford module E = AV
of the Clifford algebra C(V), for V an even-dimensional space. From the
isomorphisms
S*®SEnd (S)C(V)OR C-*AV®R C,
we see that in this case, we may identify the auxiliary bundle W with S*.
Thus, there are two natural Z2-gradings on S*: the grading S* = (S*)+
(S*)- induced by the grading on AV, and the grading defined by the chirality
operator, for which S* is purely even. We denote the first corresponding
supertrace on Endc(v)(AV) by Strnv/s; the second one, the relative trace,
we denote by
TrAV/s = 2-n/2 'AV
Let ei be a positive orthonormal basis of V. Consider the following elements
of End(E):
c2 = E(e2) - t(e'), b2 = e(e') + c(e2).
We have the relations [ci, cj] = -20, [b, bi] = 20, and [ci, bj] = 0, so that
the algebra Endc(v) (E) is generated by the elements V. We freely identify
C2 (V) and A2 V by the map a.
Lemma 4.4. If a = aijeiej E A2V, define b(a) = Ei<j aijbibi E
Endc(v)(E). We have the following formulas:

StrAV/s(eb(a)) = (-2i)n'/2jv(a)1/2 PfA(a),

TrAV/s(eb(a)) = 2n/2 det(cosh(-r(a)/2))1/2.

Proof. By definition, Strnv/s(eb(a)) = (-2i)-n/2 StrAV (c1 . . . cn eb(a)) it

suffices to prove the lemma when dim(V) = 2 and a = Oe e2, so that b(a) _
Bb'b2 and exp(9b'b2) = cos 0 + (sin 9)b'b2.
It is easily checked that StrAV(c'c2) = 0 and that StrAV(c1c2b1b2) = -4,
so that
StrAV/s (eb(a)) _ (-2i)-1 StrAV (c'c2 cos O + (sin 9)c'c2bib2)
_ -2i sin 9.
On the other hand, jv(9e'e2)1/2 = 9-1 sin 9, while PfA(9e'e2) = 0. Putting
this all together proves the first formula.
On the other hand, det(coshv(r(9e1e2)/2)1/2 = cos0, while
Tr'AV(esblb2) =TrAV(cos9)+TrAV((sin0)b'b2) =4cos9,

since TrAV(b1b2) = 0. This proves the second formula.

4.1. The Local Index Theorem 149

The curvature tensor (Ve)2 is equal to Eij k<l Ri3kl 62 t3 6k A e, which when
rewritten in terms of the operators ci and bi equals
Rijkl(c2 + bz)(ci - V)ek Ael = -4 E Rijkl(cici - bibj)ek Ael,
ij,k<l ij,k<l
where we have used the antisymmetry of Rijkl in i and j to show that
Rijkl (cibj - b2ci) = E Rijkl (cibj + ci bi) = 0.
ij ij
From this formula and (1.20), it is easy to see that F£/s is given by the
formula
FsIs = -4 1: (Rei,ej)bbj
ij
Let us now calculate the term A(M) Stre/s(exp(-FE/s)) which enters in the
formula for the index of d + d*. Let X(TM) E An(M) be the Euler form of
the manifold M associated to the Levi-Civita connection (1.38).
Proposition 4.5. We have A(M) Stre/s(exp(-FE/s)) = in/2X(TM).
Proof. When a = Ei<j aijei A ej E A2V, we have seen that

jv(a)-1/2 Stress (exp aijb'bj) = (-2i)n/2 PfA(-a).

ij
By analytic continuation of this identity, we see that the same formula remains
true when a = 2 E%-<J (Rei, ej )ei A ej is an element of A2V with two-form
coefficients, and we obtain the proposition.
As a result, we see that the index formula for

Eul(M) = ind(d + d*) = 2(-1)i dim(Hi(M)),

i=o
the Euler number of the manifold, simplifies greatly, and we obtain the fol-
lowing consequence of Theorem 4.1.
Theorem 4.6 (Gauss-Bonnet-Chern). The Euler number of an even-di-
mensional oriented manifold M is given by the formula

Eul(M) = (27r)-n12 fM Pf(-R).

For example, if the manifold is two-dimensional, we recover the classical

Gauss-Bonnet theorem

JIM
Eul(M) = -- r, dx.
Of course, this is possibly the most difficult proof of this basic theorem of
differential geometry, with the exception of proofs using quantum field theory.
150 4. Index Density of Dirac Operators

Indeed, the usual proof of the Gauss-Bonnet-Chern Theorem was given in

Section 1.6, although we did not give there the topological argument, involving
Morse theory, which is needed to show that the Euler number calculated
from a Morse function equals the Euler number calculated from the de Rham
cohomology. However, the above proof does provide a test for the index
theorem.

The Hirzebruch Signature Theorem. To derive Hirzebruch's formula for

the index of the signature operator, we have only to combine the formula that
we have already obtained for FE/S with the general index theorem. Recall that
the grading of AT*M for which the supertrace is computed for the signature
operator is the grading induced by the Hodge star-operator *.
Define the characteristic form L(M) E A4'(M) by

R/2
L(M) = det1/2
(tanh(R/2)

it is known as the L-genus. By Lemma4.4, A(M) TrE/s(exp(-FE/s))

2n/2L(M), and we obtain the following result.

Theorem 4.7 (Hirzebruch). Let M be an oriented Riemannian manifold

of dimension divisible by four; then the signature v(M) is given by the formula

o-(M) = (nri)-n/2 L (M).

IM

It follows by the same calculation that the twisted signature with respect
to an auxiliary vector bundle W is given by the formula

Q(W) = (7ri)-n/2 L(M) ch(W).

IM

This formula is due to Atiyah-Singer [13], and is one of the main steps in
their calculation of the index of an arbitrary elliptic operator on a compact
manifold.
The first non-trivial term in the Taylor expansion of the L-genus may be
easily calculated:

L(M) = 1 + 24 Tr(RZ) 1 ... .

From this, we see that for a four-manifold,

_ Tr(R 2)
o(M)
- fm 24,7r2
.
4.1. The Local Index Theorem 151

The Index Theorem for the Dirac Operator. In the case of the Dirac
operator on a spinor bundle over a spin-manifold, the curvature Fe/S = 0, so
that
ind(D) _ (27ri)-n/2 fM A (M).

This formula has a number of immediate corollaries:

(1) The index of the Dirac operator is independent of the spin-structure.
(2) If the right-hand side (which is always a rational number) is not an integer,
then the manifold M does not have a spin-structure.
(3) If the manifold M has a metric with respect to which the scalar curvature
rM is positive and non-zero, then the integral of the A-genus is zero,
since by Lichnerowicz's vanishing theorem, the index of the Dirac operator
vanishes.
Using the formula for the first few terms in the expansion of A(R) that we
gave in (1.36), we see that if dim(M) = 4, then a(M) = -8ind(D).
Consider the twisted Clifford module 6 = W ® S, where W is an auxiliary
Hermitian vector bundle with a Hermitian connection of curvature Fw. In
the index formula for the twisted Dirac operator D' associated to the Clifford
connection obtained by twisting the Levi-Civita connection on S by the con-
nection of W, the term Fe/s is nothing but the curvature Fw^' of the twisting
bundle, so we obtain the following case of the index theorem, which is in some
ways the most fundamental.

Theorem 4.8 (Atiyah-Singer). Let M be an oriented spin manifold of

even dimension. Let S be the spin bundle and let W be a vector bundle
over M. If Dye, is the twisted Dirac operator on I'(M, w (& S), then

ind(Dw) = (27ri) -n/2 fM A(M) ch(W).

T he Riemann-Roch-Hirzebruch Theorem. Let M be a Kahler manifold,

and consider the Clifford module £ = A(TO,lM)*. Let VE be the Clifford
connection on this Clifford module induced from the Levi-Civita connection.
Let us calculate the curvature FE1S of this Clifford module.
Considering the Riemannian curvature R to be a matrix with two-form
coefficients, the curvature operator (DE)2 is

Y(Rwk, wt)e(wl)c(wk)
kl

where wk is a basis of T1'0(M) and wk is the dual basis, while the End(E)-
valued two-form RE of Proposition 3.43 equals

4 >(Rwk,wt)c(wk)c(2Ul) +.1 (Rzvl,Wk)C(w`)C(wk)

kl kl
152 4. Index Density of Dirac Operators

Thus, if the basis wk is orthonormal, we have

(Ve)2 = Re + 2 E(Rwk,2Uk)
k

If W is a holomorphic Hermitian vector bundle and De is the Clifford

connection on Ao, T*M ® W obtained by twisting the Levi-Civita connection
with the connection of W compatible with the metric, we obtain
Fe/s =.1 TrT1,oM (R+) + FW,
where R+ is the curvature of the bundle T1'0M and Fl't' the curvature of the
bundle W.
Since TM OR C = T1"0M ® T°"1M, we see that
R+
A(M) = det (eR+/2 - e_R+/2 ),

so that
A(M) Tre/s(eXp(-Fe/s)) = Td(M) Tr(exp(-FW)),
where Td(M) is the Todd genus of the complex manifold M:

Td(M) = det (eRR+ i) = det (eR+/2 R e_R+/2) exp (- Tr(R+/2)).

We obtain the following theorem, which is due to Hirzebruch in the case of

projective manifolds, and to Atiyah and Singer in general.
Theorem 4.9 (Riemann-Roch-Hirzebruch). The Euler number of the
holomorphic bundle W is given by the formula

Eul(W) = (27ri)-n12 Td(M) ch(W).

fM
For example, if M is a Riemann surface (has complex dimension one) and
curvature R+ E A2 (M, C), and if L is a line-bundle with curvature F E
A2 (M' C), then Td(M) = 1 - R+/2 and ch(L) = 1- F, and we obtain the
classical Riemann-Roch Theorem,

ind(ac) = dim H°(M, L) - dim H1(M, ,C) =

4a
L(R+ + 2F).
If we take L = C to be the trivial line-bundle in this formula, we see that

dimH°(M) - dim H°'1(M) = 1 - g = 4i f R+,

M
where g = dim H"°(M) = dimH°"1(M) = 2 dimH'(M) is the genus of M.
If we define the degree deg(L) E Z of the line bundle C by the formula
1
deg(L) _ F,
-1 M
4.2. Mehler's Formula 153

we may restate the Riemann-Roch Theorem in its classical form,

Eul(1C) = 1 - g + deg(L).
If M is a compact complex manifold which is not Kahler, then for any holo-
morphic vector bundle W we may still form the Clifford module A(T°,1M)*
W, and the operator V2-(a+6*) is a generalized Dirac operator. The index of
the operator a + a* is equal to the Euler number of the bundle W by Hodge's
Theorem, as for a Kahler manifold,
n/2
ind(a + a*) = Eul(W) = j:(-1)i dimH°''(M,W).
i_°
However, although the operator ,/(a+a*) is a Dirac operator associated to a
superconnection, it is no longer the Dirac operator associated to a connection,
so we cannot expect the local index theorem to hold. Nevertheless, there is a
natural Dirac operator D on this Clifford module which does satisfy the local
index theorem, which is formed simply by taking the Dirac operator associated
to the Levi-Civita connection of the bundle A(T°'1M)* ® W. It is not hard
to extend our above treatment to this operator, and it is easy to see that
its index is given by the same formula as in the Riemann-Roch-Hirzebruch
Theorem:
(2rri)-n12
ind(D) = Td(M) ch(W).
JM
Since the Dirac operator D and ,/(a + a*) are both Dirac operators for the
same Clifford module, they must have the same index. In this way, we see
that the Riemann-Roch-Hirzebruch Theorem is valid on any compact complex
manifold, but not the local formula.

4.2. Mehler's Formula

In this section, we derive Mehler's formula for the heat kernel of the harmonic
oscillator, in the form in which it is needed for the proof of Theorem 4.1. In
order to give an idea of the result in its most basic form, let us consider the
harmonic oscillator on the real line:
z
H=-dz+x. 2

We would like to find a function pt (x, y) satisfying the following requirements:

for each t > 0, the map

0 -- Ot(x) = foo00pt (x, y) 0(y) dy

is bounded on the space of test functions S(R), satisfies the heat equation
(at + HH)gt(x) = 0,
154 4. Index Density of Dirac Operators

and the initial condition

lim A(x, y) 0(y) dy = O(x)

t-o 00

To find a solution to this problem, one is guided by the fact that the operator
H is quadratic in differentiation and multiplication to seek a solution which
is a Gaussian function of x and y; in addition, the solution must clearly be
symmetric in x and y, since the harmonic oscillator is self-adjoint:
pt(x,y) = exp(atx2/2 + btxy + aty2/2 + ct).

Applying the heat operator to this ansatz, and dividing by pt (x, y), we find

pt(x,y)+Hpt(x,y)
(atx2/2 + btxy + aty2/2 + et - (atx + bty)2 - at + x2)pt(x,y)
In particular, we obtain the following ordinary differential equations for the
coefficients:
at/2 = at - 1 = bt and ct = at.
These equations have.the solutions at = - coth(2t + C), bt = cosech(2t + C)
and ct = -1/2log sinh(2t + C) + D, and in order to determine the values of
the integration constants C and D, we simply have to substitute in the initial
conditions for pt(x, y), which show that C = 0 and D = log(2ir)-1/2. Thus,
we have obtained the following solution to the heat equation for H:

pt (x, y) = (27r sinh 2t)-1/2 exp (- ((coth 2t)(x2 + y2) - 2(cosech 2t)xy)) .
2
This formula is known as Mehler's formula.
We will mainly use this formula when y = 0. By change of variables (or
by direct calculation), for any real numbers r, f , it follows that the function
pt (x, r, f) defined by
1/2
(4.1) (47rt)-1/2 (sinh(tr/2)) exp(-(tr/2) coth(tr/2)x2/4t - t f )

is a solution of the equation

_ d2 r2x2
at + +f)pt(x) = 0.
dx2 16

We will now generalize the above argument so that it applies to the formal
heat equation that we use to prove Theorem 4.1, as well as in the proof of the
index theorem for families in Chapter 10.
Let V be an n-dimensional Euclidean vector space with orthonormal basis
ei. Let A be a finite dimensional commutative algebra over C with identity
(in the applications A will be the even part of an exterior algebra).
4.2. Mehler's Formula 155

If R is an n x n antisymmetric matrix with coefficients in A, we associate

to it the A-valued one-form w on V given by

w= 1) . Ri,,x;dxi.

The operator Vi = ai + w(8) = ai + 4 >j Rijxj acts on the space of

A-valued functions on V.
Definition 4.10. Let R be an n x n antisymmetric matrix and let F be an
N x N matrix, both with coefficients in the commutative algebra A. The
generalised harmonic operator H associated to R and F is the differential
operator acting on A(&End(CN)-valued functions on V defined by the formula
HV2)+FE(ai+.114I:Rijxj)z

+F.
i
Consider the following A-valued function defined on the space of n x n
antisymmetric matrices with values in A:

iv (R) = det
/ eR/2 - e- R/2 1
R/2 .

Since jv(0) = 1, jV1/2(tR) is defined for t sufficiently near 0, and is an

analytic function of t with a Taylor series of the form
00
(4.2) v1/2(tR) = 1 + tk.fk(R),
k=1

where fk (R) is a homogeneous polynomial function of degree k > 1 with

respect to the coefficients Rij of the matrix R. Similarly, (tR/2) coth(tR/2)
is an n x n-symmetric matrix with coefficients in A, defined for t small, and
with Taylor expansion E00 1
t2kckR2k. Thus the A-valued quadratic form

(x tR coth(tR) x)

is defined for t small, and has a Taylor series of the form

IIxI12 + E Ckt2k (XIR2k Ix).

k=1

Proposition 4.11. The kernel pt(x, R, F) taking values in A ® End(CN),

defined for t small by the formula

pt(x,R,F) = (47rt)-n/2jV1/2(tR) exp( 4t(xl tR coth(tR) Ix)) exp(-tF)

is a solution of the heat equation
(at + Hx)pt(x) = 0.
156 4. Index Density of Dirac Operators

Proof. We have to prove that dpt/dt = -Hpt. Both sides of this formula
are analytic with respect to the coefficients Rig of the matrix R. Thus it
is sufficient to prove the result when R has real coefficients. Choosing an
appropriate orthonormal basis of V, it suffices to verify the above formula
when V is two dimensional and Rel = re2, Re2 = -rel, so that
H = -(al + 192) - Zr(x2a1 - x1192) - (4r)2(xi + x2) + F

and

Pt (XI, x2) = (4irt) exp(-(tr/2) cot(tr/2)IIXII2/4t) exp(-tF).

sin(tr/2)
Since the function IIXII2 is annihilated by the infinitesimal rotation x2191-x1192,
the fact that pt satisfies the heat equation follows from (4.1) which gives the
heat kernel for the harmonic oscillator on the real line, except that r must be
replaced by ir.
In the applications we will make of the above formula, the coefficients of
the matrices R and F will be nilpotent, so that the kernel will be defined
2 /4t
for all t. Let qt(x) = (4lrt)-n/2e-11_11 be the solution of the heat equation
for the Euclidean Laplacian on V. The kernel pt (x, R, F) has an asymptotic
expansion of the form

pt (x, R, F) = qt (x) k (x),

k=0
for small t, with 4Do = 1.
We will prove again the unicity result for formal solutions of the heat equa-
tion for the harmonic oscillator, even though a theorem of greater generality
has already been proved in Theorem 2.26. Observe that H is the Lapla-
cian on the Euclidean vector space V associated to the non-trivial connection
V = d + w. Let R = Ei xiai be the radial vector field on V. If st is a
A ® End(CN)-valued smooth function on V, then
(at + H)gtst = gt(19t + t-17Z + H)st;
this follows from Proposition 2.24, if we bear in mind that the antisymmetry
of R implies that w(R) = 0, so that VTy = R.
Let Dt (x) be a formal power series in t, whose coefficients are smooth
A® End(CN)-valued functions defined in a neighbourhood of 0 in V. We will
say that qt(x)(Dt(x) is a formal solution of the heat equation (19t + H)pt = 0,
if 4 t (x) satisfies the equation
(0t+t-1R+H)-Pt=0.
Theorem 4.12. For any ao E A ® End(CN), there exists a unique formal
solution pt(x, R, F, ao) of the heat equation
(at + HH)pt(x) = 0
4.3. Calculation of the Index Density 157

of the form
00
>tk.D 1, (X)
pt(x) = qt (x)
k=0
and such that -Do (0) = ao. The function pt (x, R, F, ao) is given by the formula

(47rt)-n/2ju1/2(tR) exp(-4t (x l t2 coth(tR) x>) exp(-tF)ao.

Proof. Although this follows from the results of Chapter 2, it is worth giving
the direct proof in this special case, since it is made so easy by our formula
for the formal solution. On the left side of the equation
00

(at + t-17Z + Hx) > tk'Dk(x) =.0,

k=0
we set the coefficients of each tk to zero. This gives the system of equations
7Z(Do = 0
(7Z+k)4k = -Hx-k_1 if k > 0.
Thus we see that 4)o is the constant function equal to ao; uniqueness is then
clear, since if ao = 0, then'bk = 0 for all k.
The function pt (x, R, F) that we found above is analytic in t for small t and
is a solution of the heat equation, thus the corresponding formal expansion
pt(x, R, F) = qt(x) o tk4k(x), is a formal solution.

4.3. Calculation of the Index Density

Let £ be a graded Clifford module on an even-dimensional Riemannian mani-
fold M, with Clifford connection V and associated Dirac operator D. Recall
Lichnerowicz's formula, Theorem 3.52, for D2

D2 = A + E FE'S (ei, ej )cic3 + rM,

i<j 4

where 0 is the Laplacian of the bundle £ with respect to the connection V6, ei
a local orthonormal frame of TM with dual frame ei of T*M, ci equals c(ei),
and Fs/S(ei,ej) E Endc(M)(£) are the coefficients of the twisting curvature
of the Clifford connection VS.
Fix xo E M and trivialize the vector bundle £ in a neighbourhood of xo by
parallel transport along geodesics. More precisely, let V = Tx0M, E = £xo
and U = {x E V I lIxIl < e}, where e is smaller than the injectivity radius
of the manifold M at xo. We identify U by means of the exponential map
x H expxo x with a neighbourhood of x0 in M. For x = expxo x, the fibre E.
and E are identified by the parallel transport map r(xo, x) : £x -+ E along
the geodesic x9 = expxo sx. Thus the space C°O(U,£) of sections of 9 over
158 4. Index Density of Dirac Operators

U may be identified with the space of E-valued C"-functions on V, defined

in the neighbourhood U. A differential operator D is written in this local
trivialization as
D = E aa(x)ax,
a

where aa(x) E End(E).

Let us compute D2 in the above trivialization. For this, we need to compute
the Clifford action and the connection. Choose an orthonormal basis 8i of
V=Tx°M, with dual basis dxi of Tx°M, and let ci = c(dxi) E End(E). Let
S be the spinor space of V* and let W = Homc(v.) (S, E) be the auxiliary
vector space such that E = S ® W, so that End(E) = End(S) ® End(W)
C(V*) ®End(W). Let ei be the local orthonormal frame obtained by parallel
transport along geodesics from the orthonormal basis ai of TxoM, and let ei
be the dual frame of T * M.

Lemma 4.13. In the trivialization of E by parallel transport along geodesics,

the End(E) -valued function c(ei)x is the constant endomorphism ci.

Proof. Let R be the radial vector field on V. In the trivialization of E given

by parallel transport along geodesics, the covariant derivative VR is, by def-
inition, differentiation by R. By construction, V .ei = 0, and since V. is
a Clifford connection, [De, c(ei)] = c(Vei). Thus R c(ei)x = 0, so that the
Clifford action of the cotangent vector eX is constant and equal to c.
It follows from the above lemma that in this trivialization the bundle
Endc(M) (E), restricted to U, is the trivial bundle with fibre Endc(v) (E)
End(W).

Lemma 4.14. The action of the covariant derivative V. on C' (U, E) is

given by the formula

V. = ai + 4 E Rklijxjckcl + E AM(x)ckcl +gi(x),

j;k<i k<l

where Rklij = (R(8i, 8j)al, ak)ro is the Riemannian curvature at xo, and
fikl(x) = O(Ixl2) E C°°(U),
gi(x) = O(Ixj) E C°O(U, Endc(v.)(E)) r(U,End(W)),
are error terms.

Proof. If Fe is the curvature of the connection Ve, and R is the Riemannian

curvature of M, Proposition 3.43 shows that

Fe = a 57 (R(ai, 8j )ek, el)ckcl dxi A dxj + FE/S (8i, 8j )dxi A dxj,

i<j;k<l
4.3. Calculation of the Index Density 159

where

Fe/s(8i, 8j) E r(U, Endc(v.)(E)) = r(U,End(W)), and

(R(8i, 8j)ek, el) E C°°(U).
Observe that (R(ai, aj)ek, el)x0 = -Rklij, since at the point xo, ei = 8i. If
we apply the formula £(R)w = t(R)FE (see 1.12), we obtain the lemma.
Let pt(x, xo) be the heat kernel of the operator D2. We transfer this kernel
to the neighbourhood U of 0 E V, thinking of it as taking values in End(E),
by writing
k(t, x) = r(xo, x)pt(x, xo),
where x = expxo x. Since we may identify End(E) = C(V*) ® End(W) with
AV* ® End(W) by means of the full symbol map o-, k(t, x) is identified with a
AV* ® End(W)-valued function on U. Consider the space AV* ® End(W) as
a C(V*) ® End(W) module, where the action of C(V*) on AV* is the usual
one c(a) = 6(a) - t(a). The next lemma shows that k(t, x) is a solution to a
heat equation on the open manifold U.
Lemma 4.15. Let L be the differential operator on U C V, with coefficients
in C(V*) ® End(W), defined by the formula

L= - E((D)2 - es V) 41rM +
i<j
FE/s(ei, ej)cc'
a
.

The AV* ® End(W)-valued function k(t, x) satisfies the differential equation

(8t+L)k(t,x)=0.
Proof. Since the kernel pt (x, xo) E £., ® £x*O satisfies the differential equation

(at + Dx)pt(x, xo) = 0,

it follows that T(xo, x)pt (x, xo) E C°O (U, End(E)) satisfies the differential
equation (et + L)T(xo, x)pt(x, xo) = 0, where C(V*) ® End(W) End(E)
acts on the left on E. But under the symbol isomorphism End(E) C(V*) ®
End(W) ^_' AV* ® End(W), the action of C(V*) ® End(W) on E intertwines
with the usual action of C(V*) ® End(W) on AV* ® End(W).
To explain the method which we will use to prove the local index theorem,
let us explain it in a simplified setting. Let H be a generalized Laplacian on
a vector bundle £. If 0 < u < 1 is a small parameter, let us rescale space-time
by sending t to ut and x to u1/2x; t rescales by the same power of u as 11xI12
because the heat equation satisfied by k(t, x) has two space derivatives but
only one time derivative. Consider the rescaled heat kernel k(u, t, x), defined
by

(4.4) k(u, t, x) = un/2k(ut, u1/2x).

160 4. Index Density of Dirac Operators

The factor of un/2 in this definition is included because the heat kernel is a
density in the variable x; to put this in a more elementary way, it is needed
in order that the rescaled kernel continues to satisfy the initial condition
limt_,o k(u, t, x) = 8(x) for all 0 < u < 1. Note that the Euclidean heat
kernel qt(x) _ (4 rt)-n /Ze-jjxjj2/4t is invariant under this rescaling.
If T (t, x) _ E°__0 t'Ti(x) is a formal power series in t with polynomial
coefficients W (x), we define the rescaled series by
00
,p (U, t, x) = E(ut)i p,(u1/2x)
i=0

This is a formal series in u1/2 with coefficients polynomial functions of (t, x).

Proposition 4.16. There exist End(E)-valued polynomials Ti(x) on V with

910(0) = 1, such that the function u -+ k(u, t, x) has an asymptotic expansion
in u1/2 when u -> 0 of the form
o0
k(u,t,x) r., gt(x)E(ut)iT,(ul/2x)
i=O

This expansion is uniform for (t,x) lying in compact subsets of (0, 1) x U

and asymptotic expansions for the derivatives u '--> atOX (k(u, t, x)) may be
obtained by, differentiation.

Proof. By Theorem 2.30, there exist functions Oi E COO(U, End(E)), with

00 (0) = 1, such that for any x E U,
N
k(t,x) -gt(x)E00i(x) < C(N)tN-n/2
i=0
xke-x2/4th
Using the bounds < Cktk/2, we can replace Oi(x) by its Taylor
expansion 4b (x) of order 2(N - i):
N
k(t, x) - qt (x) E tz' i (x) < C'(N)tN-n/2,

i=0
Thus we obtain
N
""
k(u,t,x) - qt(x) 0(ut)i,bi(u1/2x) < C'(N)uN,
i=0

for (t, x) lying in compact subsets of (0, 1) x U and 0 < u < 1, which gives the
desired asymptotic expansion of the function u k(u, t, x). The argument
for the derivatives of k(u, t, x) is similar.
4.3. Calculation of the Index Density 161

The main idea of our proof of the index theorem is to modify the rescaling
of space-time of (4.4) by inclusion of a "rescaling" of AV*. Given 0 < u < 1
and a E C°° (1R x U, AV* (9 End(W)), let Sua equal
n
(&"a) (t, x) = E u-i/2a(ut, u112x) [i] .
i=o
This rescaling has the following effect on operators on C°°(JR+ x U, AV* ®
End(W)):

6.0(x)su 1 = O(u1/2x) for 0 E C°°(U),

6uat6U 1 = u-lat,
buai6uI = u-1/tai,
8ue(a)6u1 = u-1/2e(a)
for a E V*,
,u1/2t(a).
6ut(a)6u l =
Definition 4.17. The rescaled heat kernel r(u, t, x) is defined by
r(u,t,x) = un12(bu,k)(t,X).
The above choice is motivated by the fact that
n
.lim r(u, t = 1, x = 0) = lim u (n-i)/2 ku (xo, xo) [i)
u-+0 u- 0 i=o
The right-hand side of this equation is precisely what we would like to prove
to be equal to
(47r)-nI2A(M) exp(-Fe/s).
It is clear that r(u, t, x) satisfies the differential equation
(at + u8uL6u1)r(u, t, x) = 0.
The most important step in the proof of Theorem 4.1 is the calculation of the
leading term of the asymptotic expansion of L(u) = u&uL5 1 as a function
of u; we show that
L(u) = K + O (ul/2),

where K is a harmonic oscillator of the type that we discussed in the last

section. To show that this implies that r (u, t, x) has a limit as u - 0, we use
the following lemma, which is a consequence of Proposition 4.16.
Lemma 4.18. There exist AV* ®End( W) -valued polynomials' j(t,x) onRx
V, such that for every integer N, the function rN (u, t, x) defined by
2N
rN(u,t,x) = qt (x) uz/2'Yi(t,x)
i=-n
162 4. Index Density of Dirac Operators

approximates r(u, t, x) in the following sense: for N > j + Jal/2, there is a

constant C(N, j, a) such that
IIata0(r(u,t,x) -r"(u,t,x))II <_ C(N,j,a)u
for 0 < u < 1 and (t, x) E (0, 1) x U. Furthermore, yi(0, 0) = 0 if i 0,
while yo(0,0) = 1.
Proof. By Proposition 4.16, there exist AV* 0 End(W)-valued polynomial
functions Ti(x) with X0(0) = 1 such that
N
CtN-n/2.
11k (t, x) - qt (x) E t Wi (x) I I <
i=0
If we project out the p-form component of this equation, we see that
N
k (t, x) [pj - qt (x) E t2'Pi (x) [P] < CtN-n/2.
11
i=0
Rescaling by u E (0,1), we obtain
N
u-P/2gt(x) E (ut)Zi(ul/2x)[P] < C26N-p/2tN-n/2.

r(u,t,x)]p] -
i=0
It is clear that we must define yj(t,x)0,] to be the coefficient of uw/2 in the
sum
(j+p)/2
u-P12 E (ut)iWi(u1/2x)[P],
i=0
which is a polynomial on ][8+ x V with values in APV* ® End(W). It is clear
that the sum yj (t, x) = EP'=0 yj (t, x) (p] satisfies bv,yj = uj/2yj. Furthermore,
y; (t, x) fpl = 0 for j < -p, and hence yj (t, x) = 0 for j < -n.
We see that for (t, x) E (0,1) x U and 0 < u < 1,
2N
ui/2yi CuN.
r(u, t, x) - qt (x) E (t, x) I I <
i=-n/2
The argument for the derivatives is similar. The values of yi (0, 0) are deter-
mined by
00
E ui/2yi(0, 0) = (b. IF) (0, 0) = 1 O
i=-n/2
Using the fact that L(u) = K + 0(u 1/2), which will be proved later, we
can now show that there are no poles in the Laurent series expansion in u1/2
of r(u, t, x),
00
r(u, t, x) N qt(x) E ui/2yi(t, x).
i=-n
4.3. Calculation of the Index Density 163

Expanding the equation (at + L(u))r(u, t, x) = 0 in a Laurent series in u1/2,

we see that the leading term u-a/2gt(x)-y_t(t, x) of the asymptotic expan-
sion of r(u, t, x) satisfies the heat equation (at + K.)(gt(x)y_t(t, x)) = 0.
Since formal solutions of the heat equation for the harmonic oscillator are
uniquely determined by -y-e(0, 0), by the results of the last section, and since
y_Q(0, 0) = 0 for 2 > 0, we see that y_t is identically equal to zero unless
f = 0. In particular, we see that there are no poles in the Laurent expansion
of r(u, t, x) in powers of u1/2.

The other thing that we learn from this argument is that the leading term
of the expansion of r(u, t, x), namely r(0, t, x) = gt(x)yo(t, x), satisfies the
heat equation for the operator L(0) = K, with initial condition

-YO (0, 0) = 1.

Thus, to calculate r(0, t, x), we have only to obtain a formula for K, and to
apply the results of the last section.
In order to state the formula for K, let us denote by R the matrix with
nilpotent entries equal to

(4.5) Rij = (R,:oai, aj) E A2V*,

where R.,o is the Riemannian curvature of M at xO, operating on the algebra

AV* by exterior multiplication. Similarly, let F be the element of A2V*
End(W) obtained by evaluating the twisting curvature FE/s at the point xo;
again, it acts on AV* ® End(W) by exterior multiplication.

Proposition 4.19. The family of differential operators L(u) = u6 .',L6'-, 1 act-

ing on C' (U, AV* ®End(W)) has a limit K when u tends to 0, given by the
formula
1-ERijxj)2

+ F.
i j

Proof. By Lemma4.14, the differential operator V 61u = 1 is

equal to

'9j + 4 RklijXj (8k - uLk) (EL - ut')

j;k<l
+ u-1/2 E fikl(u./2x)(Ek - utk)(el - utl) + u1/2gi(ul/2x)
k<l

Since fikl(x) = O(Ixl2), we see that Va'u has a limit as u --* 0, equal to

ai + 4
1
E
j;k<l
(Rzo)klijxiekel.
164 4. Index Density of Dirac Operators

Using the fundamental symmetry of the curvature Rijkl = Rklii and the defi-
nition (4.5) of Rij = Ek<e(Rxo)jiklekel, the. above limit is seen to equal

Va,0=8i- 4Rijxj.

The operator L(u) equals L1 (u) + L2 (u), where

L1(u) = (De ') 2 + E FE/S (ei, e1) (ul/2x) (ei - utz) (e - utj )
i i<j
L2(u) = 4urM(ul/2X) +ul/2 Ve'
Clearly, lim,,,. L2(u) = 0, since its leading term is the sum of urM(xo)/4 and
0

u1/2 times limo -}O Ei Ooe e,, which we have just shown to be nonsingular.
The first term of L1 (u) has limit

- - E(ai - 4 E
(,76,0) 2

i i j

while its second term converges to

lim E FE(ei, ej)(ul/2x)(e2
- ubi)(Ej - uLj) = E F(8i, aj)EiEj = F,
'u'-0 i<j i<j
since the vector fields ei coincide with 8i when x = 0.
The operator L(0) = K is a generalized harmonic oscillator, in the sense
of Definition4.10, associated to the n x n antisymmetric curvature matrix
-Rij and the N x N-matrix F (where N = dim(W)), with coefficients in the
commutative algebra A+V. Applying Theorem4.12, we obtain the following
result.
Theorem 4.20. The limit limu._,Or(u, t, x) exists, and is given by the for-
mula

(47rt) -n/2 det1/2 (sinhtR/2

tR/2 I exp (- 1 (XI tR2 coth(tR4t2 )
Setting t = 1 and x = 0, we obtain Theorem4.1.

Bibliographic Notes
Theorem 4.1, is proved in Getzler [60], using a symbol calculus for pseudo-
differential operators which combines the usual symbol for pseudodifferential
operators with the Clifford symbol a. It is also in this article that the approx-
imation of D2 by a harmonic oscillator is introduced. Another exposition of
this proof may be found in Roe's book [96].
Bibliographic Notes 165

Earlier proofs which were special cases of the proof of [60] were those
of the Gauss-Bonnet-Chern theorem by Patodi [88], and the Riemann-Roch
theorem for Riemann surfaces by Kotake [77]. The previous proofs of the
local index theorem, by Patodi [89], Gilkey [64], and Atiyah-Bott-Patodi [8],
use power counting and symmetries of the Riemannian curvature to show
that the top-order piece a, (kt(x, x)) must be a characteristic form given by a
universal formula for all n-dimensional manifolds, and then calculate sufficient
special cases of the global index theorem to characterize this characteristic
form completely. Thus, this route to calculating the local index density is
less direct. This approach to the local index theorem is explained in Gilkey's
book [66].
There have been a number of other proofs of the local index theorem sim-
ilar to ours, among which we mention Bismut [28], Friedan-Windey [59] and
Yu [107]. Also, the work of the physicist Alvarez-Gaume [1] was influential in
the development of this point of view.
The Gauss-Bonnet-Chern theorem is due to Chern [47]; we gave an adap-
tation of his original proof in Section 1.6. The local Gauss-Bonnet-Chern
theorem is due to Patodi [88].
The signature theorem is proved by Hirzebruch in [69], using cobordism
theory; in [70], Hirzebruch tells the story of his discovery of this theorem.
The twisted signatures were first calculated by Atiyah-Singer [13], also using
cobordism theory. In the same article, they introduce the twisted Dirac oper-
ator and calculate its index. Their work was motivated by one of the central
problems of analysis: to calculate the index of an arbitrary elliptic differential
operator on a compact manifold. Atiyah and Singer used Bott periodicity to
show that the index of any elliptic pseudodifferential operator may be written
in terms of the numbers av(W). References for their general index theorem are
two seminars directed by Cartan [46] and Palais [87], and of course Atiyah's
Collected Works [2]; an excellent account of Atiyah and Singer's discovery of
their theorem may be found in the introduction to [2].
The local index theorem is proved for the a-operator on Kahler manifolds
by Patodi [88], and for the twisted signature operator by Gilkey [64]. Atiyah-
Bott-Patodi [8] give a unified treatment of the local index theorem for the
classical Dirac operators by Gilkey's method.
The index of the 0-operator was first calculated by Hirzebruch [69], in the
special case of smooth projective varieties; another proof, which proves a much
more general result, and led to the introduction of K-theory into the study
of index problems, is due to Grothendieck [39]. Atiyah and Singer gave the
first proof valid for an arbitrary compact complex manifold, as a corollary of
their general index theorem. The local index theorem for these operators was
proved by Kotake [77] for Riemann surfaces, and by Patodi [88] and Gilkey [64]
in general.
166 4. Index Density of Dirac Operators

Additional Remarks
There is an extension of Theorem 4.1 to Dirac operators associated not to a
Clifford connection but rather to a Clifford superconnection.
Let E be a Clifford module on the manifold M, and let A be a Clifford
superconnection on E. Introduce a family of rescaled superconnections
n
AE _ ail-al/2A[i].
i=0
Let FE/s be the twisting curvature of the superconnection A.
Theorem 4.21. Let Dt be the Dirac operator associated to the Clifford super-
connection At-,, and let k, , (t, x, y) = (x I e-utDt I y) be the heat kernel of
the generalized Laplacian tDt .
(1) There exist sections 1)i E C°°([8+, Ci(M) ® EndC(M) (E)) of the Clifford
bundle of M such that as t --> 0,
(47rut)_n/2
ku(t, x, x) - E ti/2,bi(u, x).
i=0
n
(2) En uj(45i(u, x)) = detl/2 (sinh(uR/2)) exp(-uFE/s)
i=o
The proof of the local index theorem of Getzler [61] is based on the Feynman-
Kac formula in stochastic differential geometry. It is shown in Getzler [62] that
with almost no modification, the same proof works for Theorem 4.21. The
same theorem is implicit in Bismut-Cheeger [33], which also uses stochastic
methods. In both cases, one obtains global estimates) for the error

ku(t, x, x) - (4irut)-n/2 det1/2 (sinuR/2

)l exp(-uFE/s)
as a function of u and x.
There is a generalization of Theorem 4.21 to the infinite dimensional situ-
ation of Chapter 9. This generalization may be used to give an alternative
approach to the family index theorem, and is also an important tool in further
generalizations of this theorem; see Bismut-Cheeger [33] for a typical example
of its use.
Chapter 5. The Exponential Map and the Index
Density

In this chapter, we will give another proof of Theorem 4.1 for a Dirac operator
D, independent of the one in Chapter 4. This proof generalizes easily to obtain
the fixed point formula for the equivariant index of a group of isometries, as
we will see in the next chapter. A feature of the proof is that it gives an
explanation for the striking similarity between the A-genus and the Jacobian
of the exponential map on a Lie group, both of which involve the j-function
sinh(X/2)
(X) X/2
Indeed, the A-genus will appear naturally through the Jacobian of the expo-
nential map on the principal bundle SO(M) over M: this Jacobian can be
computed along the fibres in terms of the curvature of M, and this is where
the j-function makes its appearance, as we will see in Section 1.
Let us explain the idea of this proof. Let P -+ M be a principal bundle
endowed with a connection, with compact structure group G, and let £ be
a vector bundle associated to P, with Laplacian 0£. It is easy to see that
the kernel kt of the heat semi-group a-tAe is obtained from the scalar heat
kernel ht on the manifold P by averaging over the fibres Px - G. This gives
integral formulas for the coefficients of the asymptotic expansion of kt(x,x),
as we will see in Section 2.
Let M be a spin-manifold of dimension n, and let Spin(M) be the corre-
sponding principal bundle with structure group Spin(n); Spin(M) is a double
cover of SO(M). Let D be the Dirac operator on the spin bundle S of M. Since
by Lichnerowicz's formula, the square of the Dirac operator is a generalized
Laplacian, we obtain a formula for the restriction kt(x,x) = (x I e-tD2 I x)
of the heat kernel to the diagonal in terms of the scalar heat kernel ht of the
Riemannian manifold Spin(M): apart from minor factors, kt is given by the
integral

kt(x,x) = ht (p, pg) p(g) -' dg,

J pin(e)
where p : Spin(e) -+ End(S) is the spin representation and p is an element of
Spin(M) lying above x E M. In exponential coordinates, this is a Gaussian
integral over A2Rn, which we identify with the Lie algebra of Spin(e). Using
this formula, it is easy to compute o(k), and the answer only involves the first
168 5. The Exponential Map and the Index Density

term of the asymptotic expansion of ht. The elimination of the singular part
of Str(kt(x, x)) is reduced to the following simple lemma: if is a nilpotent
element of an algebra A (which for us will be an exterior algebra), and if f
is a polynomial in one variable, then

lim (47rt)-1/2 f e-(X-E)214t f(x) dx = f

t-*o I
As we saw in Chapter 2, the first term in the asymptotic expansion of ht can
be computed explicitly, for any Riemannian manifold, in terms of the Jacobian
of the Riemannian exponential map, so we obtain the equality a(k) = A(M).
This argument is extended to the case of a twisted Dirac operator, in which
case ht must be replaced by the heat kernel of a generalized Laplacian associ-
ated to the twisting bundle W, and we obtain a(k) = A(M) exp(-FE/S); the
twisting curvature Fe/S is brought into this formula by the parallel transport
factor in the first term of the asymptotic expansion of ht.

5.1. Jacobian of the Exponential Map on Principal Bundles

Let G be a compact Lie group with Lie algebra g, let M be a compact Rie-
mannian manifold and let it : P --+ M be a principal bundle over M with
structure group G and connection form w. Let us choose on g a G-invariant
positive scalar product; we will denote by fG f (g) dg the integral with res-
pect to the corresponding Riemannian volume form; because the Riemannian
metric is invariant, this is a Haar measure. For p E P above x E M, the
tangent space TpP is the direct sum of the horizontal space HpP, which is
identified with TAM by the projection .7r* : TP --> 7r*TM, and the vertical
space identified with Z. Taking HP and g orthogonal to each other turns P
into a Riemannian manifold, with metric depending on w.
For p E P, we denote by
J(p, a) : TpP -> TqP
the derivative of the map expp at the tangent vector a E TpP. The purpose
of this section is to compute J(p, a) when a E g. This computation is very
similar to the computation of the derivative of the exponential map in a Lie
group.
If a E g, let je (a) equal

je(a) = dete
( 1 - e-ada)
ad a

Note that je (0) = 1.

Proposition 5.1. If dg is a Haar measure on G, then d(exp a) = je (a) da

in exponential coordinates.
5.1. Jacobian of the Exponential Map on Principal Bundles 169

Proof. If G is a compact Lie group with Lie algebra g, then

ada
(5.2) de exp(-a) exp(a + eb) 1E=o = 1 b,
ad a
for a and b E g. In a matrix group, this follows from Duhamel's formula (2.6);
since G is compact, it has a faithful matrix representation.
Let V be an Euclidean vector space with orthonormal basis eti. The Eu-
clidean scalar product on A2V has orthonormal basis {e;,Aej I i < j}. Identify
A2V with so (V) by the map T : A2V -i so (V) of Proposition 3.7, defined by
the formula (T(a)ei,ej) = 2(a,e;, Aej).
We denote by Sl the curvature of the connection w on P. It is a horizontal
g-valued 2-form on P so, for p E P, S2p is an element of A2HPP ® g. If a E g,
the contraction of SZ1, with a is an element of A2Hp P which we will denote
by S2P a.
If X is a vector field on M, we denote by XP its horizontal lift on P. If
a E g, we denote by a the corresponding vertical vector field on P. Since
XP is invariant by G, we see that [a, XP] = 0. From the definition of the
curvature of a connection (1.6), we see that

[XP,YP] = [X, YIP - Q(XP,YP)P,

where c (XP,YP) E g induces a vertical vector field Sl(XP,YP)P on P. We
will usually'write SZ(XP,YP) instead of SZ(X1',YP)P. The formula

Qp9(XP,YP) = 9-1QP(XP,YP)
follows from the G-invariance of w.
The skew-symmetric endomorphism r(Qp a) E End(TM) is given by

T(Qp a)X = 2 E(SZPW, J), a)ez,

and it varies along the fibre containing p by the formula

(5.3) T (SZp, - a) = T (Qp . (9a))

Denote by VM and VP the Levi-Civita covariant derivatives on the tangent

bundles TM and TP.

Lemma 5.2. Let X and Y be vector fields on M, and let a and b be elements
of Z. Then
(1) VaPb = [a, b].
(2) V XP a=
(3) VzPYP = (VXY)P - !Q(XP,YP)
170 5. The Exponential Map and the Index Density

Proof. By the definition of vP, if X, Y and Z are vector fields on P,

2 (V P Y, Z) equals

X (Y' Z) + Y(Z, X) - Z(X, Y) + ([X, Y], Z) - ([Y, Z], X) + ([Z, X], Y).
To prove (1), we observe that (V', b, XP) = 0, while

(Vd b, c) = a ([a, b], c) - .1 ([b, c], a) + ([c, a], b) = z ([a, b], c),
z
since the scalar product on g is G-invariant.
To prove (2), we observe that
2(V XP,YP) =-([XP,YP],a) - 2(T(92 'a)XP,1'P)
by the definition of -r, while (V XP, b) = 0. A similar argument leads to the
proof of (3).
As a consequence of Lemma5.2, we obtain a description of the horizontal
and vertical geodesics in P.

Lemma 5.3. (1) For a c g and p E P, the curve s F+ p exp(sa) is a geodesic,

so that expp a = p exp a.
(2) If x(s). is a geodesic in M, its horizontal lift is a geodesic in P.
Proof. By definition we have 89 (p exp sa) = a, so (1) follows from the relation
Vd a = 0. Similarly (2) follows from Lemma 5.2 (3).
Fix x E M and let V = T.,M. For p E P such that ir(p) = x, identify HpP
with V and TpP with V ® g. Then the map
J(p, a) : TpP --+ TpexpaP
becomes a linear map from V ®g to itself.

Theorem 5.4. The map J(p, a) preserves the subspaces V and V of TpP,
and we have
1 - e-T(np a)/2
J(p,a)Jv =
T (Qp - a) /2
1-e-ada
J(p, a) I e =
ad a
Proof. The formula for J(p, a) 1a follows from (5.2) and Lemma 5.3 (1). To
calculate J(p, a) I v, we need to compute 8t expp(a + tX)l t=o, for X E V.
Introduce p(s, t) = expp s(a + tX) and p(s) = expp(sa). For every t E R,
the curve s H p(s, t) is a geodesic. Put Y(s) = 8tp(s, t) It=o, so that Y(s) is
a vector field along the curve p(s), corresponding to the deformation of the
geodesic p(s) into the geodesic p(s, e); such a vector field is called a Jacobi
vector field. We have J(p,a)X = Y(s)I3=1. Let RP E A2(P,End(TP)) be
5.1. Jacobian of the Exponential Map on Principal Bundles 171

the curvature of the manifold P. The Jacobi vector field Y(s) is determined
by the differential equation
(5.4) VP VP Y(s) = RP(a3p,Y(s)) -asp
with initial conditions Y(O) = 0 and VaPgY(s)I3=o = X. Let us recall how
these equations are obtained: since V is torsion free, taking the covariant
derivative of Vaga8p = 0 gives

0 = Va,Va,asp= VasVa,a3p+R(atp,asp)asp
= VasVa,atp + R(atp, a8p)a5p,
Let X be a vector field on M. Since VPa = 0 and [a, XP] = 0, we see that
RP(a,XP)a = VaVPxPa.

Since exp(sa)a = a, it follows from (5.3) that T(S2pexp(sa) a) = r(Qp . a),

hence is independent of s and
(VP P a) p(3) ` 4 (T(c a)X)P -

It follows that
RP(a,XP) - a = (.Ir(ftp -a))2XP.
If V (s) is an horizontal vector field on P along the curve p(s), the vectors
Va Va V (s) and RP (a, V(s))a are horizontal. Thus the differential equation
(5.4) implies that Y(s) remains horizontal, and identifies with an element y(s)
of V. In this context, the differential equation (5.4) says /that

(as + r(fIp - a)/4)2y(S) = (T(cl - a)/4)2y(s),

so that we obtain the theorem.
For X E End(V), recall Definition 3.12 of j,,,
/sinh X/2
jv( X ) _ det
\ X/2
The Hirzebruch A-genus will appear in the index formula because of the fol-
lowing corollary to the above theorem.
Corollary 5.5. det(J(p, a)) = jg(a)jv(r(fl -a)/2))
Proof. Let A be the matrix r(fZp - a)/2. Since A is antisymmetric, we see that
det(eA/2) = 1, so that
/ \ (sinhA/2)
det I 1 - e_A I = det
A / A/2
There is no conflict of notation between jg (a) and iv (X) since the compact
group G is unimodular, so that jg (a) and jg (ad a) are equal.
172 5. The Exponential Map and the Index Density

5.2. The Heat Kernel of a Principal Bundle

Let P -- M be a principal bundle with compact structure group G and
connection form w. Let (p, E) be a representation of G in a vector space E
and let £ = P x G E be the associated vector bundle on M, with covariant
derivative V associated to the connection w on P. Let Y be the Laplacian
of £; with respect to a local orthonormal frame ei of TM,
AE .-V e,).
i
We can identify the space r(M, £) with the subspace of C°° (P) ® E invariant
under the group G, denoted C°° (P, E)G, as was explained in Proposition 1.7.
In this section, we will relate the Laplacian Y and its heat kernel to the
scalar Laplacian AP on P, which acts on C°°(P).
If 0 E C°° (P) is a function on P, we have the integration formula

f 0(p)dp = (f 0(p9)dg)dx.
P JM G

In this formula, note that fG q(pg)dg depends only on x = 7r(p).

Let Ea, be an orthonormal basis of the Lie algebra g of G and let Cas =
>2a p(Ea,)2 E End(E) be the Casimir operator; it is independent of the
choice of the orthonormal basis Ea, of g and commutes with the action of G
on E. In particular, if the representation p is irreducible, Cas is a scalar.
Proposition 5.6. The Laplacian L coincides with the restriction of the
operator OP ®1 + 1® C to (C°O (P) (9 E)G.
Proof. Let eP be the horizontal lift of the vector field ei on M to the principal
bundle P. The vector fields eP and Ea form a local orthonormal basis of TP.
By (5.2) we see that
AP =- ((eP)2

- (De:ei)P) - EE 2.
i a
In the identification of F(M, £) with (Coo (P) (9 E)G, the operator Ve;, corre-
sponds to eP, while, using the G-invariance of elements of (C°° (P) ® E)G,
the vertical vector field Ea may be replaced by the multiplication operator
-p(Ea); the proposition follows.

Let us compare the semigroups of operators e-t°" on P(M, £) and e-t°v

on COO(P). The Schwartz kernel of e_t°E with respect to the Riemannian
density jdxI of M is denoted by (x I e-t°E y). On pulling back to P, it
becomes an element (pl, p2) H (Pl e_t°E P2) of C°°(P x P) ® End(E)
which satisfies
(pigi I e-tv£ I P292) = P(91)-l(pi I e-toE I p2)P(92)
5.2. The Heat Kernel of a Principal Bundle 173

From Proposition 5.6, we see that

e-toE -tcase tOP
=
Thus we obtain an integral representation of the heat kernel of the Laplacian
of an associated vector bundle 5 in terms of the scalar heat kernel on the
manifold P.
Proposition 5.7. If p1 and P2 are points in the principal bundle P, then

(P1 I
e-to' P2) = e-tcas
I

(p1 I e-t" I P29)P(9)-1d9

JG
Proof. We have
(e-toE
0) (P1) =
e-t"' I e-top I P2) O(P2) dp2
J (P1
= e-tCas f I e-top P29) O(P29) d9) dx
(P1

M \ fG
= e-tcas M f (f P1 I e-toP p29) P(9)-1 dg) 0(P2) dx,

from which the theorem follows.

This theorem leads to an integral representation for the asymptotic expan-
sion of (x I e-t°f I x) in terms of the asymptotic expansion of the heat kernel
of P.
If 0 E C°° (R) is a smooth function with compact support, then by Propo-
sition 2.13, the integral
110 k
f e-a2/4to(a)da N (47rt)1/2 (i (Oko)(0)
k=0
has an asymptotic expansion in powers of t1/2 near t = 0 which depends only
on the Taylor series of 0 at 0.
Let 0 E C°°(g) be a smooth function with compact support on the vector
space g. We will denote by

(5.5) "ymp e-Ilal12/4t¢(a) da

8

the asymptotic expansion in t1/2 of the integral

e-IIaII2/4tO(a) da.
18

This power series has the form tdim(e)/2 EOO0 taxi

If 0 is a smooth function defined only in a neighbourhood U of 0 E g, and
V is a cut-off function on g with support in U and equal to 1 near 0, then
reeymp e_IIaII2/4tO(a)i(a)da
J8 is independent of the choice of 0, so we will write
174 5. The Exponential Map and the Index Density

it simply as ff YmP e-IIx112/4t0(a)da. If -D(t, a) _ E°_o ti4bi(a) is a formal

power series in t with coefficients in C°°(U), we will write
_YMP
J IIaII2/4t4p (t, a)da

for the formal series

00 tjYrnP
e2/4tqi(a)da.

To simplify the statement of the next theorem, we will suppose that the
Casimir Cas of the representation p of G on E is a multiple of the identity.
Theorem 5.8. There exist smooth functions I)i on P x g such that
(p e-taE I p)
a symp °O
(4irt)-(dim(M)+dim(g))/2Ig (p,a)e-IIaII2/4tp(eXpa)-1 jg(a)da.
i =0

Furthermore, denoting T,.(P)M by V, we have

(bo(p, a) = det-1/2(J(p, a))

1/2
= jg (r(fl . a)/2)
1/2(a).'j

V
Proof. Let ht(p1ip2) be the heat kernel of AP on COO(P). By Theorem5.8,
we have
(p l e-toE I p) = e-tCas G
ht(p,pg) p(g)-1 dg
Let fi(t) be a cut-off function which vanishes for t greater than half the
square of the injectivity radius of M.
In Section 2.5, we showed how to obtain a formal series E°_o tii, whose
coefficients (bi are smooth functions in the neighbourhood
{(P1,P2)I d(pl,p2) < r}
of the diagonal of P x P such that if
N
(4irt)-dim(P)/2
hiT (p1 ,P2) = (d(pl,p2)2)e-d(Pi,P2)2/4t
EtiW (pi,p2),
i=0

we have
O(tN-dim(')/2)
ub(Pi,p2) - hN(pi,p2)1I =
By Theorem 2.30 and Corollary 5.5, we have, for a small,
N
ht (p, p exp a) = (4irt)- d,m(P)/2e-IIaII2/4t ti`I`i(P, a) + O(tN-dim(e)/2),

i=0
5.2. The Heat Kernel of a Principal Bundle 175

(a) jv1/2(T(Lp a)/2). The proof is completed by mul-

where (DO (p, a) = j9 1
tiplying through by the formal power series a-tcas = Ek °(-tCas)k/k!, and
inserting the formula dg = jg (a) da.

Let us extend this theorem to the case of a generalized Laplacian on a

twisted bundle W 0 E. Let W be a bundle over M, which we will call the
twisting bundle. We can make the following identifications between space
of sections:

r(M, W ®£) = (r(P,*W) ®E)G,

r(M, End(W (9£)) _ (r(P, zr* End(W)) ® End(E))G.

If F E (r(P, 7r* End(W)) (9 g)G, then applying the map p : U --+ End (E), we
obtain a section

p(F) E (r(P, 7r* End(W)) (9 End(E))G = r(M, End(W (9 £)),

which we will simply denote by F. We will also abbreviate our notation

for the bundles zr*W and zr* End(W) on P to W and End(W). If F E
(r(P, End(W)) ® g)G and a E g, then for all p E P, the contraction Fp a
is a well-defined element of End(W,,), where x = zr(p), and we have the in-
variance for g E G,

(5.6) Fpy a = Fp (ga).

We will consider generalized Laplacians on W ®£ of the form
H = AW®E + F° + F',

where

F° E r(M, End(W)) C r(M, End(W ® E)),

F1 E (r(P, End(W)) (D g)G.

Here, W is a twisting bundle with connection Vw, and we is the Laplacian

on r(M, W (9 E) with respect to the tensor product connection on W ® S.
The main result of this section will be an integral formula for the asymptotic
expansion of the heat kernel of H. As before, we may view the kernel of e -1H
as a section
(pl, p2) (pi I e-tH I P2)
of the bundle W ® W® ® End(E) over P x P. Restricted to the diagonal, the
kernel (p I e-1H I p) gives a section of (r(P, End(W)) (D End(E))G. We will
give an asymptotic expansion of (p I e-tH I p) compatible with this tensor
product decomposition.
176 5. The Exponential Map and the Index Density

Theorem 5.9. Let H = 0W®E + F, where F = F° + F1 is a potential

such that F° E r(M, End(W)) and F1 E r(P, End(W)) (& g)G. Assume that
the Casimir Cas of the representation p is scalar. Then there exist smooth
sections 'j E r(P x g, End(W)) (where we denote by End(W) the pull-back
of End(W) on M to P x g) such that
(p I e-t' I p) ,.,
asymp a0
e-IIa112/4t E t'l)3 (p, a)®p(exp a)-17e(a) da.
(47rt)-(dim(M)+dim(g))/2
f
Furthermore, letting V = T,r(p)M, we have

d)o(p, a) _ ig'12(a)1v1/2(r(1p . a) /2) exp(FF a/2).

Proof. The proof is very similar to the proof of Theorem 5.8. The first step
is to write H as the restriction to r(M, w (& £) _ (r(P, w) (& E) G of a
carefully chosen generalized Laplacian Hp on P. We will thus obtain an
integral formula similar to Theorem 5.7 for the kernel (p I e-l' I p).
The map a H .1Fp1 a E End (W,,) defines a vertical one-form on P. We
associate to the potential F a connection V on lr* W by adding to the pull-back
7r*Ol^' of the connection V"V on W the vertical one-form 2F1:

=VW+ZF'.
Let 0 be the Laplacian of the bundle a*W associated to the Riemannian
structure of P and the connection V.

Lemma 5.10. There exists a potential F E r(P, End(W))G such that H is

the restriction of the operator (0 + F) ® 1 + 1 ® Cas to r(M, w o e) _
(r(P, W) (& E)G.

Proof. Let F = Ea (F' . Ea)2 + F°, where Ea is an orthonormal basis of

We have 4

H= - EVeP - V(Ve=ej)P +>F1 Ea ®p(Ea) +F°,

i a
whereas
2 2

i a
Observe that
VE`+ = Ea)2
(Ea +IF1
a a
_ Ea + 1(F1 Ea)Ea + 2Ea(F1 -Pa) + 1(F1 Ea)2.
a
5.2. The Heat Kernel of a Principal Bundle 177

From the relation (5.6), we see that Fpexp(sEa) - Ea is independent of s, so that

the action of the vector field Ea commutes with the endomorphism F1 Ea,
and we obtain

DEa Eat + (F1- Ea)Ea +

1(F1

Ea)2.
a a

As before, we can replace the second order operator E. Ea on (r(P, W)(&E)G,

by Cas, while we can replace Ea by -p(Ea); the lemma follows easily.
For pi and P2 E P, let us denote by ht(pi, p2) E Hom(WP2, Wp3) the kernel
of As in Theorem 5.7, we have

(P1
I e-t' I P2) = e-teas IG ht(p1,p29)p(9)-' d9-

There exists smooth sections (pl,p2) - I'j (pl,p2) E Hom(Wp2, Wpl), defined
in a neighbourhood of the diagonal of P x P, such that if
N
htr(p1,p2) = (47rt)-dim(P)/2 (d(p17p2)2)e-d(p,,p2)2/4t
EtiIi(Pl,p2),
i=0
we have
Ilht(pl,p2) - hiv(Pl,p2)II = O(tN-dim(e)/2)

Furthermore, the first term (Do(p1i p2) does not depend on F, and is given
by the formula
'Do (p1, p2) = det-1/2 (J(pl, expP1' p2)) T (pl, p2),

where T(p1, p2) E Hom(Wp2, Wpl) is the geodesic parallel transport with
respect to the connection V. We will compute T in the next lemma.
Lemma 5.11. For p E P and a E g, we have
T(p,pexpa) = exp(2Fp -a).
Proof. The function t H T (t) = T (p, p exp ta) is the solution of the differential
equation
a) T (t) = 0, T (0) = 1,
dtT (t) 2 (Fp exp to .
and, by (5.6), we have Fpexpta . a = Fp a.
The rest of the proof of Theorem 5.9 is the same as that of Theorem 5.8,
except that the Do term is multiplied by T (p, p exp a), calculated above.
When we apply the above theorem to the heat kernel of D2 in Section 4,
the factor jv1/2(T(a Stp)/2) in

-1'o(p, a) = je 1/2(a) jv1/2 (T ()p . a)/2) exp(FF . a/2)

178 5. The Exponential Map and the Index Density

will gives rise to the A-genus of M, while the factor exp(Fp' a/2) will lead
to the relative Chern character.

5.3. Calculus with Grassmann and Clifford Variables

Let 0 be a continuous function on R slowly increasing at infinity. For a E Ilk,
we have:
tlimo (47rt) -1/2 e-(x-a)2/4tO(x)dx
f = O(a),

as follows from the change of variable x t1/2(a + x). We will prove a

generalization of this formula, which allows a to lie in a more general algebra.
Let A be a finite dimensional supercommutative algebra with unit. (In
practice, A will be an exterior algebra). Let l; _ (Cl, ... , ln,) be a n-tuple of
even nilpotent elements of A. Let 0 be an A-valued smooth function on R.
We define:

(5.7) O(x + ) = E aJO(x)6J/J!, for x E R'.

JEN'
This sum is finite because the elements 6 are nilpotent. We will write lJx-612
for the function E 1(x j - j) 2.
Lemma 5.12. Let 0 be an A-valued smooth function on Rn slowly increasing
at infinity (as well as all its derivatives). Let 6 Cn) be an n-tuple
of even nilpotent elements of A. Then

tlim e-11x-e112/4tO(x)
o(4irt)-n/2 f n dx = ( ).
Proof. If (D and its derivatives are rapidly decreasing functions on Rn, we
have

f n 6) - 1)(x)) dx = E jI
J#o
J

2n
aJ-D(x) dx = 0.

Thus

f e-IIx-E112/4to(x) dx =
f e-11112/4tq(x + 6) dx
Rn n

= E JI . L.
J
t dx.

In the limit t --+ 0, we obtain the lemma.

Let V be a Euclidean vector space. We denote by expA(a) the exponential
of a E AV, and by g = C2(V) the Lie algebra of Spin(V), which we can
identify as a vector space with A2 V by the symbol map. By the universal
5.3. Calculus with Grassmann and Clifford Variables 179

property of the symmetric algebra (C[g] of polynomial functions on g, the

injection of g into the even part of AV extends to an algebra homomorphism
A : C[g] -- AV.
We can extend the map A to the algebra to C°°(g) by composing with the
Taylor series expansion at 0. Let us describe this homomorphism in the case
in which V = R', with orthonormal basis ea. An element f of COO (g) is a
smooth function f (aid) of the coordinates {ati3 1 < i < j < n} on A21[F1,
and A(f) is the element f (ei n e&) E AV defined in (5.7), which is obtained
by replacing the variables ai.9 by the elements e' A e2 E A2Rn. Denote by
e n e* the antisymmetric matrix whose (i, j)-coefficient equals ei A ei E A2V.
We will use the suggestive notation f (e n e*) for A(f), and will call it the
evaluation off at e n e*.
Since ei n ei = 0, many polynomial expressions in the variables ei n ej will
vanish, and consequently some functions on g will have a vanishing evaluation
at eAe*. In the following lemma, we collect a number of such functions. Recall
that T is the canonical representation g -i so(V)-
Lemma 5.13. (1) Let v, w E V and let fv,.(a) _ (T(a)kv,w) be the cor-
responding coefficient of the matrix -r(a)k. Then f, w(e n e*) = 0 for
k>2.
(2) Let fk(a) = Tr((ada)k), where ad is the adjoint representation of g.
Then fk(e n e*) = 0 fork > 1.
(3) je(e n e*) = 1

Proof. The coefficients of the matrix r(a)2 are linear combinations of mono-
mials aikakj. Thus (1) follows from the relation ek n ek = 0.
The adjoint representation can be written as ad(a) = T-(a) n 1 + 1 A 7(a) for
a E g. Hence it follows from (1) that the evaluation at e A e* of any coefficient
of (ad(a))k vanishes, when k > 2. Since 7-(a) is antisymmetric, we also have
Tr(T(a) A r(a)) = - Tr(r(a)2) and Tr(ad a) = 0. Thus, for any k > 1, the
evaluation at e n e* of Tr((ad a) k) vanishes.
Using the relation det(A) = exp(Tr(logA)), we can write je(a) in the form
j. (a) = exp(Tr(Eck(ada)k)),
k>1

and (3) follows from (2). (Alternatively, (2) and (3) can be seen as a conse-
quence of the fact that the constants are the only elements in AV which are
invariants under the full orthogonal group O(V).)
The following proposition is the crucial technical tool leading to the elimi-
nation of the singular part of the heat kernel index density.
Proposition 5.14. If (D is a smooth slowly increasing function on g, then
tli o (47rt)-dim(e)/2 f e-IlaII2/4t expA(a/2t) -D(a) da = 4)(e n e*).
8
180 5. The Exponential Map and the Index Density

Proof. Since (ez A ej)2 = 0, we obtain by completion of squares the formula

e-11a112/4t expc(a/2t) = expA (-4t E(azj - e2 A ei)2).
_<j

Applying Lemma 5.12 gives the desired formula.

For the proof of Theorem 4.1, we need to replace Grassmann variables by
Clifford variables in Proposition 5.14. The next proposition shows that this is
indeed possible when taking the limit t --> 0. Let St be the automorphism of
AV such that
St(y) = t-112y for 'y E A'V.
As usual, we will identify the spaces C(V) and AV by means of the symbol
map.
Proposition 5.15. If (D is a smooth slowly increasing function on g, and
'y E AkV, then

lim (4irt)-dim(e)/2tk/2 f e-Ila112/4tst(7 expc(a/2))(D(a) da = -y A (e A e*).

t0 a
Proof. Recall the relation (3.13) between the exponentials in the exterior al-
gebra and the Clifford algebra:
j1/2
expc a = (a) det(Hv2(a)) expA a),
where Hv(a) : g -- GL(V) is given by the formula
T(a)/2
Hv(a) =
tanh(T(a)/2)
In particular, the matrix entries of HH1/2(a) are analytic function of the
coefficients of the matrix T(a)2.
Let el be the orthonormal basis of AV obtained from an orthonormal basis
eti of V. We have

St (eXpc (a/2) )

= jv 2(a) det(Hly 2(a)) J(eXpA(a/2t), ei) (Hvl/2(a) Cl, ej) ej.

IJ
By Lemma5.13, the evaluation of jv2(a) det(Hv2(a)) (Hu1/2(a)er,eJ)
at e A e* is equal to SI J. Thus, applying Proposition 5.14, we obtain the result
for -y = 1.
If -y = el A ... A ek, we have
ry expc(a/2) _ (el - t1) ... (ek - tk) expc(a/2)
Applying St, we obtain
t1/26t('y. expc(a/2)) _ (e - ttl) ... (ek - ttk)St(eXpe(a/2))
5.4. The Index of Dirac Operators 181

and taking the limit, we deduce the result from the case y = 1.
We restate the above result in a slightly different way.
Proposition 5.16. Let (D be a smooth function on g defined in a neighbour-
hood of 0. Let y E Cm(V), with symbol am('y) E AmV. Consider the formal
power series

(47rt) - dim(9)/2 e-116112/4t

P(a) ('y expc a) da ^ E
00 tECi,
J9 ymp
i=o

where Ci E C(V). Then the element Ci is of Clifford degree less than 2i + m,

and we have
n/2
L:
i=O
o2i+m(Ci) A 4D (2e A e").

w
Proof. Proposition5.15 asserts that the Laurent series Eti+m/2bt(Ci) has
i=o
no poles, and computes the constant term.

5.4. The Index of Dirac Operators

We now give our second proof of Theorem 4.1. Thus M is an oriented compact
Riemannian manifold of even dimension n. For simplicity, we assume that M
admits a spin structure; this is always true locally, and since the theorem is
local, it is sufficient to prove the theorem in this case.
Let V = Wn, with its standard orthonormal basis ei. Let G = Spin(n),
with Lie algebra g = C2(V), and let T : g -+ so(V) be the map defined
in Proposition 3.7. We denote by 7r : P -* M the principal bundle Spin(M)
which defines the spin structure, and provide it with its Levi-Civita connection
w. Let S2 E A2 (Spin(M), g) be the curvature of the principal bundle Spin(M).
Then the Riemannian curvature R of M is equal to r(S2) E A'(M, so (TM)).
At a point p E P, which corresponds under the two-fold covering Spin(M) -+
SO(M) to an orthonormal frame at the point x = 7r(p) E M, we may identify
both TxM and T xM with the vector space V.
Let p : G --> End(S) be the spinor representation, and let S be the asso-
ciated superbundle of spinors, which carries the Levi-Civita connection VS.
The Casimir Cas of the representation p is given by
Cas = E(eiej)2 = -dim(g) = -n(n - 1)/2,
i<j

so is a scalar.
We suppose given a Hermitian complex vector bundle W --j M and a
Hermitian connection V'^' on W with curvature Fw, and we let D be the
corresponding twisted Dirac operator of the twisted spinor bundle W ® S.
182 5. The Exponential Map and the Index Density

Let kt (x, y) be the heat kernel of the operator D2; along the diagonal, it is a
section of the bundle End(W (& S) = C(M) ® End(W).
Since the operator D2 satisfies the Lichnerowicz formula
D2 = AW®S + V(FW) + 4rM,

we may apply to it the analysis of Section 2. Indeed,

D2 = AW®S + F° + F',

where
F°= 1rM
4

Fp = E Fp (eP, eP) 0 (ei A ej) E (r(P, End(W)) ®g)G.

i<j
By Theorem 5.9, the asymptotic expansion of (x I e -ID' I x) has the form
asymp W
(47rt)-(dim(s)+n)/2
f
9
e-I1a112/4t
E
j=o
tj Dj(a) ® p(exp a) jg(a) da,

where 'j form a series of smooth functions on g with values in End(WW), and
j V1/2(T
-Do(a) = jg 1/2(a) (gp . a)/2) exp(-FP . a/2).
In the identification End(S) ^_' C(V) = AV by the symbol map o,, the matrix
p(exp a) corresponds to the element expC(a) of AV. The first assertion of
Theorem 4.1 follows immediately from Proposition 5.15 with y = 1.
The formula in Proposition 5.15 for the leading order of the above asymp-
totic expansion shows that
v(k) = (Do (2e A e*) jg(2e A e*).
By Lemma 5.13, the function js (2a) evaluated at e A e* equals 1. Thus, it
remains to calculate (Do (2e A e*).
Identifying V and V*, we consider the curvature Fp as an element of
A2 V 0 End (W.) and the Riemannian curvature Rp as an element of A2 V
End(TTM). The function a i-, FP1 a is a End(W,,)-valued linear function on
g, thus its evaluation at e A e* is an element of A2V ® End(WW). Similarly
the function a H T(527, a) is an End(T,,M)-valued linear function on g, thus
its evaluation at e A e* is an element of A2V 0 End(TxM).
Lemma 5.17. We have Fp (e A e*) = Fp and T(f2p .(e A e*)) = R.
Proof. The curvature FP is the element
ei A ej ®Fp (eP, eP)
i<j
of A2V ® End(WW). If Fp E End(WW) 0 g is given by (5.4) and a =
F-i<j aij ei A ej E g, we see that FP a = Ei<j aijFp (eP, el'). Replacing
Bibliographic Notes 183

the scalar variables aij by the Grassmann variables ei A ej, we obtain the first
formula.
The second formula is more subtle. Let us write lp E A2V ®g as
f2p = T ek A el ® (flp(ek,el),ei n ej) ei A ej.
i<j;k<l
Let Eij = T(ei n ej), so that {Eij I i < j} is a basis of so(V). Since the
Riemannian curvature Rp E A2V (g So (V) is the image of the curvature SZp by
the representation T of g in V, we see that
Rp = 57 ek n el ®(SIp(ek, el), ei n ej) Eij.
i<j;k<l
The fundamental symmetry of the Riemannian curvature shows that
RP = ei n ej ® (lp(ek, el), ei A ej) Ekz.
i<j;k<l
For a = >i<j aijei A ej E g, the element f2p a of A2V equals
57 (lp(ek, el), ei A ej) aijek A el,
i<j;k<l
so that
T(Qp . a) = E (Qp(ek, el), ei A ej) aijEk1.
i<j;k<l
Replacing the variable aij by the Grassmann variable einej, we obtain (2).
The function -Do (2a) is the product of the real function jo 112 (2a), the real
function jV112(T(!Qp a)) and the End(W,,)-valued function exp(-Fp a). We
know that jB 112 (2e A e*) = 1. By the preceding lemma, the evaluation of
the function jv1/2(T(Qp a)) at e A e* equals the A-genus of M, while the
evaluation of the End(WW)-valued function exp(-Fp a) at e A e* equals the
element exp(-FW) of A(M, End(W)). This completes our second proof of
Theorem 4.1.

Bibliographic Notes
The method used in this chapter first appeared in Berline-Vergne [25]. This
work was influenced by the articles of Getzler [60] and Bismut [28].
Chapter 6. The Equivariant Index Theorem

Let M be a compact oriented Riemannian manifold of even dimension n,

and let H be a compact group of orientation preserving isometries acting on
M. Let E -3 M be a Clifford module with Clifford connection; if H acts on
E compatibly with the Clifford action and Clifford connection, we call E an
equivariant Clifford module. If D is the Dirac operator on £ associated to the
given data, then D commutes with the action of H; hence, the kernel of D is
a finite-dimensional representation of H. The equivariant index is the virtual
character of H given for y E H by the formula
indH(y, D) = Tr(y, ker D+) - Tr(y, ker D-).
In this chapter, we will prove a generalization of the local index theorem
for D which enables us to calculate indH(y, D) for arbitrary y E H. Let
xI rye-tD2 y) be the equivariant heat kernel of D, that is, the kernel of the
operator rye-tD2. The restriction of (x I ye-tD2 I y) to the diagonal of M is
a section of End (6) which we denote by kt(y, x).
Fix y E H, and denote its fixed point set by M. The main result of
this chapter, Theorem 6.11, is that kt (y, x) has an asymptotic expansion, as
t --> 0, of the form
00
(47rt) - dim(My)/2 E to(Di(-Y, x),
i=0

where the coefficients -Di(y, x) are generalized sections of End(S) which are
supported on M'Y, and such that 1Di (y, x) has Clifford filtration degree 2i +
dim(M) -dim(Mry). We will calculate O2i+dim(M)-dim(M'r)C('Y) x) explicitly,
obtaining a formula of a similar type to that of Chapter 4, which is the case
where H = {1}; this formula involves the curvature of the normal bundle N
of My in M and the action of H on it.
We begin the chapter with some generalities on the equivariant index. In
Section 2, we give a proof of the Atiyah-Bott fixed point formula for the equi-
variant index of the 5-operator when the fixed point set is discrete; this serves
as a simple introduction to the equivariant index theorem. In Section 3 we
state Theorem 6.11, while in Section 4, we show that it implies the equivariant
index formula of Atiyah-Segal-Singer for indH(y, D) as an integral over M'r
and the local formula of Gilkey. The rest of the chapter is devoted to the
proof of Theorem 6.11, by a method generalizing that of Chapter 5.
6.1. The Equivariant Index of Dirac Operators 185

6.1. The Equivariant Index of Dirac Operators

Let M be a compact oriented Riemannian manifold of even dimension n,
on which a compact Lie group H acts by orientation-preserving isometries.
There is induced an action of H on the Clifford bundle C(M), which preserves
the product.
Definition 6.1. An equivariant Clifford module £ over M is an H-
equivariant bundle, with a Clifford module structure and Hermitian inner
product preserved by the group action. We suppose that the action of H
preserves the Z2-grading of S.
Let D be a Dirac operator associated to a Clifford connection Ve on S.
The following lemma is evident.
Lemma 6.2. The action of the group H commutes with the Dirac operator
if and only if the Clifford connection vE is H-invariant.
From now on, we only consider invariant Dirac operators D on S. Since an
element y E H preserves the Dirac operator D, it must map ker(D) to itself,
so that ker(D) becomes a 7L2-graded representation of H. The character of
this representation, that is, the supertrace
Str(y, ker(D)) = Tr(y, ker(D+)) - Tr('y, ker(D-)),
is called the equivariant index of the Dirac operator D, and we will denote it
by indH(y, D); it is an element of the representation ring R(H) of the compact
group H. The purpose of this chapter is to calculate indH(y, D) in terms of
local data at the fixed point set of y acting on M; for y = 1 E H, this is just
the local index theorem of the Chapter 4. The other extreme occurs when the
fixed point set of y consists of isolated points; if D is associated to an elliptic
complex, the formula for indH(y, D) in this case, due to Atiyah and Bott, is
particularly simple to derive, and we will present it in Section 2.
The foundation of the calculation of the equivariant index of a Dirac oper-
ator is the following generalization of the McKean-Singer formula.
Proposition 6.3. Let kt(y, x) IdxI = (x I ye tD2 I x) be the restriction of
the kernel of the operator ye-tD2 to the diagonal. Then for any t > 0,

indH(y, D) = Str(ye-tD2)
= IM dxl.
Proof. We will give two proof of this result, which are transcriptions of the
two proofs of the McKean-Singer formula Theorem 3.50 in its simplest version.
If A is a real number, the A-eigenspace 7-la of the generalized Laplacian D2,
acting on r(M, £}) is a finite dimensional representation of H, with character
ye-tD2:

X". We have the following formula for the supertrace of

Str(ye-tD2) _ (XA (Y) - Xa (Y))e
t
A>o
186 6. The Equivariant Index Theorem

The Dirac operator D induces an isomorphism between 7-1 and 7-la when
A 0. Since D commutes with H, this isomorphism respects the action of
H, so that
Xa (7) - Xa (7) = 0 ford > 0,
and we are left with Xo (y) - Xo (y), which is nothing but the equivariant
index of D.
The second proof of Theorem 3.50 generalizes as follows. If a(y, t) is the
function Str(ye_tD2), we again see by Lemma2.37 that for t large,
Ce-tal/2
Str(ye-toe - 7Pker(D)) I < vol(M),

where Al is the smallest non-zero eigenvalue of D2. This shows that

a(-y, oo) = lim Str(ye_tD2) = indH(y, D).

t-->oo

The proof is completed by showing that the function a(t, y) is independent of

t, so that a(7, t) = a(y, oo) = indH(y, D). Differentiating with respect to t,
we obtain

dta(y,t) _ -Str(yD2e-tD2)
De-'D2]),
_ -2 Str([-yD,
since D is odd and [D, y] = 0. This last line vanishes by Lemma 3.49. E3

Thus, the calculation of the equivariant index is reduced to the calculation

of the limiting generalized section lira kt(y, x) of S.
t-+o

6.2. The Atiyah-Bott Fixed Point Formula

The Atiyah-Bott fixed point formula is a generalization of the Lefschetz fixed
point formula, which gives the supertrace of the action of a diffeomorphism on
the de Rham cohomology of a manifold. In some respects, it is more general
than the equivariant index theorem which we will prove later, since the group
H is allowed to be non-compact. However, it is more restrictive, in that
it only gives a formula for elements y E H whose fixed point set consists
of non-degenerate isolated points, and D must be associated to an elliptic
complex.
Let D : r(M, £') -+ r(M, £i}1) be a differential operator acting on the
space of sections F(M, £) of a Z-graded finite dimensional vector bundle £
over a compact manifold M. We say that (r(M, £), D) is a complex if
D2 = 0. Suppose that the cohomology space ker(D)/im(D) is
finite dimensional.
6.2. The Atiyah-Bott Fixed Point Formula 187

Consider a group H acting on £ - + M and assume that the action of H

on I'(M, £) commutes with D. The group H acts on the graded vector space
H* (D) and we define, for -y E H,
indH(y, D) _ (-1)2 Tr('y, H'(D))
E

If 'y acts on M with isolated nondegenerate fixed points, Atiyah and Bott
gave a fixed point formula for the index of the transformation y, even if H is
a non-compact group.
In this section, we will give a proof of Atiyah-Bott formula for those com-
plexes D for which there exist a Hermitian structure on £ and a Riemannian
metric on M such that the operator D + D* is a Dirac operator. Let £ -+ M
be a H-equivariant vector bundle over a Riemannian manifold M. However,
we do not need to assume that H acts by isometries on M; in particular, H
need not be compact.
If x is an isolated fixed point of the action of an element -y E H on M,
denote by -lx the tangent action at the fixed point x, and by yy the endomor-
phism of the fibre £x induced by y. The point x is called a non-degenerate
fixed point if the endomorphism (1 - yr) of TAM is invertible; for example,
this condition is satisfied if x is an isolated fixed point and H is a compact
group. Let L be a generalized Laplacian on £ and let kt(y, x) I dxI be the
restriction of the kernel of 7e-'L to the diagonal; it is a section of £ ® IAI,
where IAI is the bundle of smooth densities on M.
Recall that a generalized section s E r-O°(M,£) of a bundle £ is a contin-
uous linear form on the space of smooth sections of £* ® Ithus,
Al; the space of
smooth sections F(M, £) of £ may be embedded in F-°°(M, £). We use the
notation fM(s(x), O(x)) I dxI for (s, 0 I dxI ). If X E M, the delta distribution
Sx at x is the generalized section (6, 0) = O(x) of the bundle of densities Al. I

Lemma 6.4. Let X0 E M be a non-degenerate isolated fixed point of y E

H. If X(x) E C°°(M) is a function equal to 1 near xa and vanishing in

IP
a neighbourhood of all other fixed points of the action of y on M, then as
t -+ 0, the section x(x)kt(y, x) of End(S) 0 IAI has a limit in the space of
generalized sections of End(g) ® IAI, given by

tlimo X(x) kt(`y, x) I dxI = det(1 F-°° (M, End(E) ® Al).

a'yxo ) I E

We must compute

lim X(x) kt(y,x) fi(x) ldxl for o E F(M,£).

t--+OJM
Since kt(y,x) = ye(y-ix I e-tL I x), we see that the function x(x)kt(y,x) is
rapidly decreasing as a function of t, except in a small neighbourhood U of
xo. Let V = T,,0M, and let F : U --+ V be a diffeomorphism of U with a
188 6. The Equivariant Index Theorem

small neighbourhood of 0 in V such that dx0F = I. We transport the integral

over M to an integral on a small neighbourhood of 0 in V, and trivialize the
bundle E on this neighbourhood.
The approximation of (x I e-IL I y) shows that

lim X(x) kt('Y, x) O(x) dx!

t-+0 JM
=t mo(4-7rt)-n/2 e-d('Yso x,x)2/4t
O(x) O(x) 1dxl,
IV
where 5(x) is a smooth compactly supported section of End(E) equal to -y-O.
for x = 0. Consider the change of variables x -> t1/2x on V. We see by
Proposition 1.28 that
lim t-'d(-)c1t1/2X, tl/2x)2 = 11yxo x - xjf2,
t-+o
where II - II
is the Euclidean norm on V, and the lemma follows. 0
If D is a complex on a 7L-graded Hermitian vector bundle E --+ M over
a Riemannian manifold M such that D + D* is a Dirac operator, then L =
DD* + D*D = (D + D*)2 is a generalized Laplacian which commutes with
D:
(DD* + D*D)D = DD*D = D(DD* + D*D).
It follows from Hodge's Theorem 3.54 that D + D* is essentially self-adjoint,
that the cohomology H(D) is a finite-dimensional vector space and that
ker(D+D*) = ker(D) flker(D*) is a subspace of ker(D) isomorphic to H(D).
Lemma 6.5. Let L be a generalized Laplacian of the form DD* + D*D with
D2 = 0. Assume that the action of H on r(M, E) commutes with D. Then,
for every t > 0,
indH(y, D) = Str(ye-tL)
Proof. The proof is very similar to that of the equivariant McKean-Singer
formula of Section 1. First, note that the right-hand side of the equation is
independent of t; indeed, its derivative with respect to t is

dt Str(ye-IL) = - Str(y(DD* + D*D)e-tL)

D*e-tL])
= - Str([yD, = 0.
Here, we have used the fact that the supertrace vanishes on supercommutators
and that L and -y commute with D.
When t --> oo, the semigroup a-tL tends to the projection PO onto ker(D+
D*). This subspace need not be fixed by the action of y, since -y does not
commute with D*. However, we have ker(D) = Po(ker(D)) ®Im(D), thus the
supertrace of the operator PoyPo is the index of y, and we obtain
indH(y, D) = Str(PoyPo) = Str(ye-tL).
6.2. The Atiyah-Bott Fixed Point Formula 189

From the above two lemmas, we obtain the following special case of the
Atiyah-Bott fixed point formula.
Theorem 6.6 (Atiyah-Bott). Let £ -- M be a Z-graded Hermitian vector
bundle on a Riemannian manifold M, and let D be a first-order differential
operator on r(M, £) such that D2 = 0, and such that the operator DD*+D*D
is a generalized Laplacian.
Let y be a bundle map of the bundle S covering the action of 'y on M,
and commuting with D. Assume that the action of ry on M has only isolated
non-degenerate fixed points. Then
Str(^GO)
indH('y, D) _
xoEM' I det(1 - rya )I'
where M7 is the fixed point set of the action of y on M.
Let us state the formulas that this theorem gives for the d and 0-complexes.
Corollary 6.7. If M is a compact manifold, and ry is a diffeomorphism of
M with isolated non-degenerate fixed points, then
(-1)2 Tr (- , HZ(M)) _ E e(xo, -y),
i xoEM'r
where e(x°i y) is the sign of the determinant det(1 - ryxa ).
Proof. The operator d acts on sections of the 7G-graded bundle £ = AT*M
and satisfies the condition of the above theorem for any choice of Rieman-
nian metric on M. If y is an endomorphism of a real vector space V, then
Ei(-1)'Tr(y, AiV*) = det(1 - y-1, V).
Note that when y lies in a compact group, the sign of det(1-yxo) is always
positive, since the real eigenvalues of ry are ±1.
Next, we turn to the 6-complex.
Corollary 6.8. If M is a compact complex manifold with holomorphic vector
bundle W - M, and y is a holomorphic transformation of W --> M, then
y acts on the a-cohomology spaces H°'i(M, W). If the action of y on M has
only isolated non-degenerate fixed points, then
.w
(ylo 1

i xoEM'Y
detTxo
io M( 'Yxo )

Proof. The operator a acts on sections of the Z-graded bundle A(T°,1M)*

VV. We have
E(-1)i TrAi(TT61M)'19WIo (,y) = Tr(y o) detToo1M(1
i

Since I det(1 - 7;" 01)1 = detT ooM(1 - yxo) detTo 1M(1 -, -y,-;0 1), the corollary
follows.
190 6. The Equivariant Index Theorem

6.3. Asymptotic Expansion of the Equivariant Heat Kernel

Let M be a compact oriented R.iemannian manifold of even dimension n, and
let H be a compact group of orientation preserving isometries of M. Since an
element y of H is an isometry of M, its fixed point set M7 is a submanifold
of M, which may consist of several components of different dimension. Using
the Riemannian metric on M, we may decompose the tangent bundle along
M7 as
TMIM- =TM7®N,
whereJV is the normal bundle along M7.
Since -y is an isometry of M, the Levi-Civita connection preserves the
decomposition TM I M, = TM7 e N, so induces connections on TM7 and N
which preserve the metrics on these bundles; the connection on TM7 is in fact
the Levi-Civita connection of M. Let R E A2 (M, so (M)) be the Riemannian
curvature of TM, and let R° E A2(M7,so(M7)) and R' E A2(M7,so(N))
be the curvatures of the induced connections on TM7 and N; thus, R° is the
Riemannian curvature of the manifold M. We have the following relation:
(6.1) RIM7=ROE) R'.

Since Ro is the Riemannian curvature of the manifold M7, we see that

R°/2
A(M7) = det1/2
(Slfli/2))
is the A-genus of M7.
The tangent map dy restricted to M7 defines a section of the bundle
SO(M)lm-Y, which preserves the orthogonal decomposition TMI M7 = TM7
N. Furthermore, TM7 is exactly the eigen-bundle with eigenvalue one of dy:
if v E T,,0M is a tangent vector fixed by dx0y for some x° E M7, the curve
expxo tv is fixed by y so that v E TXOM7.
Fix a component M° of M7 and let yl be the transformation induced
on All Mo . Since the only possible real eigenvalue of yl is -1, we see that
det(1 - yl) > 0. Furthermore, since det(yl) = 1, we see that dim(Nxo) =
dim(M) -dim(Mo) must be even. We denote by 2° and 21 the locally constant
functions on M7' such that dim(M°) = 2t° and dim(Nxo) = 2.21 for all x° E
MOB; it is clear that n = Vo + 221. Denote by det(N) the line bundle over M7
which over the component M° of M7 is equal to A2e'N*; this line bundle is
contained in AT*M!M,.
Lemma 6.9. The function det(1-yl) is constant on each component of M.
Proof. If Mo is a component of M7, each fibre Nom, x E M0, of the bundle
N carries a finite-dimensional representation of the closure G(y) of the group
gcuorated by y in the group of isometries of M. Since M is compact, this
is a compact group. Since the dual of the compact group G(y) is discrete,
6.3. Asymptotic Expansion of the Equivariant Heat Kernel 191

the representation must be independent of x E Mo, from which the lemma

follows.

Let -y E SO(V) and A E 50(V). In Definition 3.25, we have defined the

analytic square root det1"2(1 - yexpA). Applying this with y1 E
and A = -R1 E so (A(,) ® A2T*XO M Y, we obtain an element of AT x0 M'Y; as
the point xo varies, we obtain a differential form det1"2(1 - y1 exp(-R1)) E
A(M'Y). It follows from the theory of characteristic classes that this form
is closed. Its zero-degree component is equal to I det(1 - yi)I1/2 so that
det1/2(1 - y1 exp(-R1)) is a well-defined invertible element of A(Mry).
We may trivialize the density bundles IAMI and IAM, I by means of the
canonical sections dxI and Idxol derived from the Riemannian metrics on M
and M"Y respectively. Thus, the delta-function SM, of MY is the generalized
function on M defined by

SMry (x) O(x) I dxI = f O(xo) I dxol ,

IM M7
where I dxI and I dxo I are the Riemannian densities of M and M. If.F is a
bundle over M and v E r(M'Y,.F), then v6M-I is a generalized section of F
over M.
From now on, £ will be a H-equivariant Clifford module on M, with
H-invariant Clifford action and Clifford connection. As usual, we write
End(S) ^_' C(M) ® EndC(M)(£). If M has a spin structure, then £ is iso-
morphic to W ® S for some bundle W, and Endc(M)(£) = End(W).

Lemma 6.10. Along MY, the action of y on £ may be identified with a

section yE of C(N*) ®Endc(M)(£).

Proof. We have yc(a)y-1 = c(ya) for a E r(M, T*M). Thus the action
on £xo of y at a point xo E M7 commutes with the Clifford action of
C(T x0M'Y). It is clear that the commutant of C(T**0M'Y) in End(£.,o) is
C(Nxo) ®EndC(M)(£)X0.

By this lemma, we may form the highest degree symbol vdim(N)('yC),

which is a section of Endc(M) (£) ® det(N) C Endc(M) (£) ® AT*MI M7, since
det(N) = Adim(N)N*.
Let FE/S be the twisting curvature of the Clifford module S. Recall that if
M has a spin-structure with spinor bundle S, we may write £ = W ® S, and
then FE/S = Fw is the curvature of the twisting bundle W. We denote the
restriction of FE/s to M'Y by Fo JS E r(M7, AT*M'Y ® Endc(M) (£)).
Let D be the Dirac operator of the Clifford module £; it is an H-invariant
operator on r(M, £). Let
kt(y, x) I dxI E r(M, c(m) ®Endc(M) (£) 0 IAI)
192 6. The Equivariant Index Theorem

be the restriction of the kernel of the operator rye tD2 to the diagonal. The
main result of this chapter is to show that kt('y,x) has an asymptotic ex-
pansion when t --> 0 in the space of generalized sections r-O° (M, C(M)
Endc(M) (£) (& IAI), and to calculate the leading order of the expansion. This
theorem is a generalization of a result of Gilkey.
Theorem 6.11. The section kt(y) of the bundle C(M) ® Endc(M)(E) ® IAI
has an asymptotic expansion as t -+ 0 in the space of generalized sections of
C(M) ® Endc(M)(£) ® IAI of the form
00
kt(y) '" (47ft)-dim(M')/2 Etti-Pi(Y)
i=0

where 1)i (-y) is a generalized section of C2i+dim(N)(M) ® EndC(M)(E) (9 IAI

supported on M. The symbol of 1)i(y) is given by the formula
dim(M'Y)/2

T,
i=0
a2i+dim(N) (-Di (y)) = 1'(-Y) ' 6Mry ,

where I(y) E I'(M7, AT*M ® Endc(M)(E)) equals

A(MY) exp(-Fp /s) o,di m (N)('YE)

(Y) =
det'/2 1 1-detl/2 1 - i ex Rl
Observe that when y is the identity, this theorem reduces to the local index
theorem of the last chapter, Theorem 4.1.
From the formula
kt('y,x) = -y' (-y-'x I e-tD2 I x),
we see that if yx # x, the function t -> kt(y,x) is rapidly decreasing as
t - 0; thus only generalized sections with support in M7 will appear in the
asymptotics of the generalized section kt(y), and we need only study kt(y) in
a neighbourhood of M1.
The map (xo, v) E N H exp,,0 v defines a diffeomorphism between a neigh-
bourhood of the zero section of N and a neighbourhood U of M. We may
identify £eXp 0 v with £.,0 by parallel transport along the geodesic s E [0, 1]
exp.,, sv; this respects the filtration on End(E) = C(M) ®EndC(M) (6). Thus
kt (y, expx0 v) may be thought of as lying in the fixed filtered algebra End(£x0 ).
If 0 is a smooth function on M with support in U, the integral

I (t)yAxo) = JAr.0 kt ( y, expx0 v) 0(eXpx0 v) I dv I

defines an element of End(E,,0), where I dvI is the Euclidean density of J'J

Using the rapid decrease of kt (y, x) as t -+ 0 for x 0 M'' and a partition of
unity argument, we see that Theorem 6.11 is implied by the following result.
6.4. The Local Equivariant Index Theorem 193

Proposition 6.12. Let MO be the component of the point xo in Mry, and

let 0 be a smooth function on M supported in a small neighbourhood of Mo.
Then
(1) 1(t, -y, 0, xo) has an asymptotic expansion as t -40 of the form
00
(47rt)-to E
I (t, xo) - tzllz(o, xo),
=o

where xo H -bi, (0, xo) E r (Mo, C2(%+tl) (M) ®EndC(M) (s)).

(2) If a(k(ry, 0, xo)) is given by the formula EZ° o a2(i+ti) ((bi(O, xo)), then
a(k('y, 0, xo)) = I(y, xo).O(xo)
We will prove this proposition later in this chapter. However, if xo is an
isolated point of M'Y, the proof is much the same as that of Lemma 6.4, which
we proved in the course of proving the Atiyah-Bott fixed point theorem in the
last section. Indeed we see, using the change of variables v -> t1/2v, that
yxo (xo)
lim I (t, -y, 0, xo) =
t-4o det(1 --yo 0M}

6.4. The Local Equivariant Index Theorem

The Atiyah-Segal-Singer fixed point formula for indH(-y, D) is a formula for
the equivariant index indH(y, D) of an equivariant Dirac operator D as an
integral over the fixed point manifold M. In this section, we will show how
combining Theorem 6.11 with the equivariant McKean-Singer formula leads
to a local version of this theorem.
Let W be an H-equivariant superbundle on M. Fix -y E H. If A is an
H-invariant superconnection on W, denote by FO the restriction to M-1 of the
curvature F = A2 of A. Define the differential form chH(ry, A) E A(M1'), by
the formula
chH(ry, A) = Strw(y exp(-Fo))
This is a closed differential form on M'r whose cohomology class is indepen-
dent of the H-invariant superconnection A, by a proof which is identical to
that for the ordinary Chern character. In this section, we denote chH (y, A)
by chH(ry, W).
If £ is an H-equivariant Clifford module over M with H-invariant super-
connection A, we will define a differential form
chH(y, £/S) E A(MI, det(N)) = r(M7, AT*M" (9 det(N)),
which agrees with chH (y, W) up to a sign when £ = W ® S is an equivariant
twisted spinor bundle. Recall that we have defined a map
Stre/s : r(Mry,Endc(M)(£)) -' Coo (M").
194 6. The Equivariant Index Theorem

If E = W ® S and a c r(M'', End(W)), Stre/s(a) = Strw(a). We extend

Strg/s to a map
Str&/s : r (M", EndC(M) (E) (9 det(N)) -> r(MI, det(N)).
Let Fo lS be the restriction to M-1 of the twisting curvature of A. Then
Stre/s (odim(N) (yE) exp(-Fo /S))
is a differential form on M7 with values in the line bundle det(JV).
Definition 6.13. Define the localized relative Chern character form
2P1
chH (y, E/S) = 1/2(1) StrE/S (adim(N) (yE) exp(-FF /s))
det - yl
E A(M7, det(N)).
Obviously, this rather peculiar definition requires some justification; we will
rewrite it in terms of more familiar objects when M has an H-equivariant
spin-structure; this is by definition a spin-structure determined by a princi-
pal bundle Spin(M) which carries an action of the group H, such that the
projection map Spin(M) --> SO(M) is equivariant.
Proposition 6.14. If M has an H-equivariant spin-structure, then for any
y E H, the manifold M'r is naturally oriented.
Proof. The action of -y on the fibre of the spinor bundle S,,,0 at the point xo E
M7 gives an element' E End(S,,0) C(T*a,oM) such that yc(ce) = c('la)y'
for all cY E NrO. With respect to the decomposition TOM = Vo ® V1, where
Vo = T 0Mry and V1 =.N , we see that y" is an element of C(V, ).
Let ei, 1 < i < 2t0, be an orthonormal basis of Vo . If
eo
ro = expC (2 e2i-1 A e2i) = c(el A ... A e2to) E Spin(Vo ),
i=1
then ro -y E Spin(V*) maps to (-1) x yl E SO(VV) X SO(V1 ).
Let T : AT Xo M -+ R be the Berezin integral. By Lemma 3.28,
IT(ro .5)I = 2-L'l detl/2(1 - yl)-
Thus 0_dim(N)(5') is a non-vanishing section of det(N) such that

I0'dim(N) (y) I = 2-ei det1/2(1 - yl)

Thus 0dim(N) (y) defines a trivialization of det(N), which is the same thing
as an orientation of Al. But given an orientation of M and N*, there is a
unique orientation on MY compatible with these data.
Fix an orientation of M''. Sometimes, there is a natural orientation on MY
(for example, if Mry is discrete), which may not agree with the orientation
which we obtained above. We will write e(y)(xo) = 1 if the two orientations
6.4. The Local Equivariant Index Theorem 195

agree, and e(y)(xo) = -1 if they do not; thus, for fixed y, e(y) is a function
from M'Y to {±1} which is constant on each component.
This discussion implies the following result.

Proposition 6.15. Let M be an H-equivariant even-dimensional spin mani-

fold, and let W be an H-equivariant vector bundle with H-equivariant super-
connection A. Consider the H-equivariant Clifford connection Vw ®1 + 1®
V8 on the Clifford module £ = W ® S. Then
chH('y, -'IS) = e('y) chH('y, W)
Proof. In this case, rye = yi'p' ® y, and

Odim(N) (7e) = e(y)2-el detl/2(1 - yl)_Yl_

We now return to the general case, in which M does not necessarily have
a spin-structure, and state the equivariant local index theorem. Consider the
section of AT*M over My given by
A(Mry) chH(y, £/S)
- 'yi exp(-Ri) )
det1/2 (1
Since M is orientable, we may take the Berezin integral of this section, to
obtain a function on M'r, which we denote by
TM A(Mry) chH('y,S/S) E C°O(M").
detl/2(1 -'y1 eXp(-R''))
We can now state the equivariant index theorem.

Theorem 6.16 (Atiyah-Segal-Singer). The equivariant index indH(y, D)

of an equivariant Dirac operator D associated to an ordinary connection is
given by the formula
indH(y, D) _
dim(M
7 )/2TM (_4(fY)
chH(y, £/S)
i - dim(M)/2
fM_Y detl/2 (1 - y1 eXp(-R'))

Proof. We will show that this theorem is a consequence of Theorem 6.11. Let
U be a tubular neighbourhood of MY and let 0 be a smooth function with
support in U identically equal to 1 in a neighbourhood of M. Using a
partition of unity, we see that

fM Str(kt('(, x)) dx - fu Str(kt('Y, x))'b(x) I dxI

-f-.AV Str(kt(y, expx0 v)) O(exp.0 v)b(xo, v) dxo ldvl.

196 6. The Equivariant Index Theorem

Here, we have rewritten the Riemannian volume jdxl in terms of the Rieman-
nian volume on N and a smooth function b(xo, v) satisfying b(xo, 0) = 1, by
the formula

fM
(x) dxl =
fMl
fNs0
(eXp0 v)b(xo, v) dyl Idxol

Applying Proposition 6.12 with ¢(v) = O(expx0 v) b(xo, v), we see that

I (t, y, 0, xo) = kt(y, expx0 v) O(v) I dvi

fAro
z

has an asymptotic expansion of the form

00
x0) .., (47rt)- dim(M')I2 E ti-bi (x0),
i=O

where the Clifford degree of 4Di(xo) is less than or equal to 2(i + ei) We
now use the fact that the supertrace vanishes on elements of Clifford degree
strictly less than n. It follows from Theorem 6.11 that Str(I(t, y, xo)) has an
asymptotic expansion without singular part, so that

indH(y, D) = lim Str(kt(ry,x)) Idxi

t-O JM
lim Str I (t, y, c, xo) i dxo I
M'r t--40

= fMl Str (dim(M1)/2(XO)) I dxol ,

from which the theorem follows immediately on substituting the formula for
Cn (Ddim(M1)I2).

There are two special cases of this formula in which the geometry is easier
to understand. The first case is that in which M has an H-equivariant spin-
structure, with spinor bundle S. Thus, the Clifford module S is a twisted
spinor bundle W ®S, where W = Homc(M) (S, 6), and D is the twisted Dirac
operator associated to an invariant connection VW on W with curvature Fw.
Then Proposition 6.15 shows that

indH(y, D) = fM (27ri)-eoi-t1s(y) A(M7) chH('Y, W)

det"2(1- y1 exp(-R'))
A further simplification of the formula occurs if in addition the fixed point
set M'' has a spin-structure. In this case, the normal bundle N has associated
to it a spinor bundle S(N), defined by the formula

S(N) = HomC(M7)(SM7, SMI M'r)-

6.5. Geodesic Distance on a Principal Bundle 197

The actions of y on the spinor bundles SM7 and SM induces an automorphism

yl of the vector bundle S(.V). Choosing compatible orientations on Mry and
N, we have
chH('Y,S(N)) = chH('Y,S+(N)) -chH('Y,S (N))
= i-eie(y, MI) det1/2(1 - 7i exp(-R')),
and the equivariant index formula may be rewritten as follows:

indH('Y, D) (27ri)-Q0(-1)tl A(Mry) chH('Y,W)

M7 chH('Y, S(N))

6.5. Geodesic Distance on a Principal Bundle

In this section, we prove a technical result which is basic to the proof of
Proposition 6.12. Let 7r : P -* M be a principal bundle over a Riemannian
manifold, with compact structure group G, connection form w E Al (P) ®
and curvature Q.
We denote by d2 (pl, P2) the square of the geodesic distance between two
points pi and P2 of P. If pl and P2 belong to the same fibre it-1(x), the
horizontal tangent spaces at pl and P2 may be identified with TIM. Given
vl; v2 in TIM, we are going to compute the Taylor expansion up to second
order in t E R of d2(expp, tvl, expp2 tv2). We have p2 = pl exp a for some
a E g. Recall from Theorem 5.4 that
expp, (a + tv) = expp2 (tJ(pl, a)v + o(t)) for v E TIM,
where
e_T(nn
(6 2) (1-11 a) I T,;M - 1 -T(S2p, , a) ./2a)/2 .

Proposition 6.17. For small a, d2(expp, tvl, expP2 tv2) equals

Ila112+t2 ((Vi, J(pl, a)-1vl) - 2(vl, J(pl, a)-'V2) + (v2, J(pl, a)-'V2))+o(t2)
Proof. Let pi(ti) = expp.(tivi), i = 1, 2, and define X(tl,t2) E Tp,(t,)P by
p2(t2) = expp,(t,) X(tl,t2). Thus, we must compute IIX(tl,t2)112 to second
order with respect to (t1, t2).
We identify a neighbourhood of pl in P with a neighbourhood of the origin
in V x g by means of the exponential map; here, V = TIM. Thus, for t small,
Tp,(t)P may be identified with V ® g. Under this identification, the metric
on TT, (t) P coincides with the orthogonal sum metric on V ® g = Tp, P up to
second order.
Let us start by computing X(tl, t2) up to first order. We have
X(t) = X(t,t) =a+tXl +0(t2).
198 6. The Equivariant Index Theorem

Differentiating both sides of the equation exppl(t) X(t) = p2(t) with respect
to t, we obtain
v1 + J(p1, a)Xl = v2,
here, we have used the fact that expel (t) a = expP2 (tv1), which follows because
both sides of this equation are geodesic curves starting at p2 with tangent
vector v1. Thus, we obtain
X (t1, t2) =a+ J(pi, a)-1(t2v2 - tivl) + a(I tl i + It2I)
In particular, since the horizontal and vertical tangent spaces are orthogon
for t1 = t2 = 0, this implies that II X (t1, t2)112 - IIail2 vanishes to at least
second order.
Denote by VP the Riemannian covariant differentiation on P, and by RP
its curvature. Consider the map p : ][83 P given by
p(s, ti, t2) = expP,(t,)(sX(t1, t2))
This map satisfies the equations
oa asp = 0, la9pII2 = IIX(t1,t2)II2, and a8p19-o = X(t1i t2).
Thus, for i = 1, 2,
as (at.;p, asp) = (VP Da at;p) asp) ,

= (v VP 8,,p, a9p),
q by the vanishing of the torsion,
= (RP(a3p, at:p)a8p, asp) ,
= 0, since RP is antisymmetric.
From this, we see that (8t;p, 89p) is an affine function of s. Furthermore, the
vector at,pl s=i is tangent to the curve
t1 H expP, (t1) X (t1, t2) = p2 (t2)
Thus 8t,p19=1 = 0, from which it follows that
(at,P,asp) = (1 - s)(8t,p,asp)I9-o
Writing

at1 I1 aep112 = 2 (Da , asp, a8p)

= 2a8 (at,p, asp)
= -2(at1p, a3p) I9=0I
we obtain the differential equation

at, IIX(t1, t2) 112 = -2 (at1p, X(t1, t2))Pl(tl)

= 2(v1, J(p1, a)-1(tjv1 - t2v2)) + o(It11 + It21)
6.6. The Heat Kernel of an Equivariant Vector Bundle 199

Let IIX(tl,t2)II2 = IIa2+tix11+2t1t2x12--t2x22+o(ti+t2) be the Taylor

expansion of II X (tl, t2)1 12. We see that
a)-lvl),
xll = (vl, J(pl, and
x12 = -(vl, J(pl, a)-lv2)
Now d2(pl(tl),p2(t2)) = d2(p2(t2),expP2exp(_a)tlvl). Thus, exchanging the
roles of (pl, ti) and (p2, t2) we obtain

x22 = (v2, J(p2, -a)-lv2) = (J(p2, a)-1v2, v2)

Since J(pi, a) = J(p2i a), this completes the proof.

6.6. The Heat Kernel of an Equivariant Vector Bundle

As in Chapter 5, let 7r : P -> M be a principal bundle with compact structure
group G, and let £ = P x G E be an associated vector bundle corresponding to
the representation p of G on the vector space E. Let w be a connection one-
form on P, and let 17 be the corresponding covariant derivative on £; we will
always consider P with the Riemannian structure associated to the connection
w. In this section, we will generalize the results of Section 5.2 to the situation
in which a compact group H acts on the left on P by orientation-preserving
isometries, and leaves the connection one-form w invariant.
Also, suppose we are given an auxiliary bundle W with an action of H,
endowed with an H-invariant Hermitian metric and an H-invariant Hermitian
connection. Let F be a potential of the form F = F° + F', where F° E
F(M, End(W)) and Fl E r(P, End(W) ®g)G, as in Section5.2. Assume that
F commutes with the action of -y.
The group H acts on £, hence on the space of sections F(M, £) by the
formula (y s)(x) = ys(y-lx), where y E H and x E M. If Aw®E is the
Laplacian on the bundle W ®£ associated to the connection Vw ®1 + 1® Ve
on W ® £, we will interested in the smooth kernel
(x0 I ),e-t(01'"®-+F) I

We can lift this kernel to the principal bundle P, obtaining there the kernel
I ye-t(Ow(&E+F)
(PO I pi) E F(P x P,W ®W*) ®End(E).

Fix y E H, and let Al be the normal bundle to M7 in M. The exponential

map (x0, v) expxo v, for x° E M'' and v E NNo, gives an isomorphism
between a neighbourhood of the zero section in Al and a neighborhood of M"
in M, and if p E 7r-1(x0), the point expyv projects to expxo v under ir. We
identify We,.pao v and Wx0 by parallel transport along the geodesic. Let 0
200 6. The Equivaxiant Index Theorem

be a smooth function on N,,o with small support. Our aim is to study the
asymptotic behaviour of the time-dependent element of End(W,) (& End(E)

p) = J (expp v I ry exp (-t(OW + F)) I expp v) 0(v) dv.

No
Generalizing Theorem 4.1, we will prove that 1(t, -y, 0, p) has an asymptotic
expansion as a Laurent power series in t1/2 when t - 0, and give an integral
formula for this expansion.
Consider the quadratic form Qa on Ty0M®TZOM defined for v1, v2 E Ty0M
by the formula
/ /
Qa(vl, v2) _ (vi, J(p, a)-lvl) - 2(vl, J(p, a)-lv2) + (v2, J(p, a)-lv2)-
By Proposition 6.17, Qa is the Hessian of the function

(vi, v2) -' d2 (expp vl, exPpexp a v2)

at the critical point vl = v2 = 0, and for a = 0, we have that Qa(vi,V2)I a=o =
Ilvi - v2112. Let Q1(a, y) be the quadratic form on given by

Qi (a, y) (v) = Qa(v, 7v) for v E Nxo;

for a = 0, we have det(Q1(0, y)) = det(1 - yl)2 # 0. Finally, let 7 be the

element of G such that yp = pry".

Theorem 6.18. Assume that the Casimir Cas of the representation p is a

scalar. There exist smooth sections Di E I'(P x g, End(W)) such that as
t - 0, I (t, y, 0, p) is asymptotic to the Laurent series
aeymp 00
(47-t)- dim(P)/2+e1 f 6- 11 aII'/4t a) p((exp
a)-iry)j,, (a) da.
8 i=0

Furthermore, -'o (p, a) is given by the formula

y o exp(Fnl - a/2) 0(0)

,D0(p,a) - det( Q1(a,ry))Il/2 A/2 (a) 7T M (T (np . a) /2)
,
Proof. The proof is a generalization of that of Theorem 5.9. However, we
will only give the details in the scalar case, in which W is the trivial bundle,
since the extension to non-zero potential F is similar to the corresponding
extension in the non-equivariant case, apart from the occurrence of y o in the
formula for Thus, from now on, we study the heat kernel of the operator
L
Let ht(po, pi) be the scalar heat kernel of P and let Cas be the Casimir of
the representation (p, E) of G. The following lemma is the equivariant version
of Proposition 5.7.
6.6. The Heat Kernel of an Equivariant Vector Bundle 201

Lemma 6.19.
po I ye-t0E I pl) = e-teas fG
ht(po, yp19)p(9)-1 dg

Proof. In the identification of r(M, £) with C°°(P, E)G, the action of H on

F(M,6) corresponds to the action on COO(P) given by (ys)(x) = s(y-lx).
Now the operator ye-top on C°O(P) has kernel ht('y-1po,p1) = ht(po,'ypl);
the rest of the proof is as in that of Proposition 5.7.
Identifying v E Txo M with a horizontal tangent vector at p E P, the
geodesic expp tv is a horizontal curve which projects to the curve expxo tv in
M. It follows from the above lemma that
(eXpp v I ye-t0E I
expp v) = e-'Cas fG ht(exp v, -y expp(v) - 9)p(9)-1 dg.

The curve yexpp(ty) is a geodesic in P starting at yp = py with a horizontal

tangent vector which projects onto ylv. Thus yexpp v = expp(yiv) y and,
replacing g by y-'g, we obtain
(eXpp v I ye-toE I expp v) tCasp(9-1'y)
= JG ht(expp v, expp(1'iv) . 9)e d9
When g is far from the identity and v is small enough, the function
t H ht(eXp , v, expp(y1v) . g)
decreases rapidly as t - 0. Thus, since only a small neighbourhood of the
identity contributes asymptotically, we can work in exponential coordinates
on G. Let 0 E C°° (g) be a smooth cut-off function equal to 1 on a small
neighbourhood of 0. Since Id(exp a) I = je (a) Idal, we see that I(t, y, 0, xo) is
asymptotic to
asymp
e-tCas
v, exp (ylv) . ea)p(e-'') O(v) ")(a) j9 (a) J dal Idvi.
Jg xJv o ht (eXpp
Applying Theorem 2.30, we may replace ht (expp v, expp(yiv) exp a) by its
approximation
00
(47rt) - dim(P) /2e-f (a,v)/4t E ti-Di (a, v),
i=0
where f (a, v) = d2 (expp v, expp(yly) exp a) and
Di(a, v) = Ui (expp v, expp('yly) exp a).
Furthermore, Theorem 5.8 shows that
1/2
,Do (a, 0) = jg(a)-1/2.7T (T(cZp . a)/2)).
By Proposition 6.17, the function v H f (a, v) on .N has its only critical
point at v = 0, where its Hessian is the quadratic form v - Q1(a,y)(v).
202 6. The Equivariant Index Theorem

When a = 0, we have Q1(0, -/)(V) = 11(1 - yl)vll2. Thus, for small a, the
quadratic form Q1 (a, y) is not degenerate on AI . By Morse's Lemma, there
exists a local diffeomorphism v H w = Fa(v) of N.,,, such that Fa(0) = 0 and
f(a,v) = Ila1I2+IIWII2 Note that
dv = I det(Q1(a,'Y))I-1/2.
dw w=0
Thus, we see that I(t, y, 0, x0) is asymptotic to
asymp
(47rt)- dim(P)/2e-tCas e-(II.II2+II_112)/4t

00
dv
1(w))O(F,,'(w)) V)(a)p(eXP(a)-1'Y)j2(a) dw Idal Idwl.
i=0

The change of variable w -> t-1/2w gives us the asymptotic expansion of the
right-hand side that we wished to prove, with
0(0)
Do (p, a) _
det(Q1(a, 7))I1/23e/2(a) 3T a) /2)
In the case of a twisting bundle W, we consider the heat kernel ht of the
operator 0 + F of Theorem 5.9, and we use the formula

(p0 I ye-t(AW®e+F) I Pl) = f

G
ht(po,'ypi - g)
`Yr®e-tcap(g-1)
dg.

6.7. Proof of Proposition 6.13

In this section, we will prove Proposition 6.12, and hence Theorem 6.11. The
proof is a generalization of the method of Chapter 5. As in that chapter, we
will take advantage of the fact that the theorem is local to require that the
manifolds M has an equivariant spin structure, with spinor bundle S. Thus,
the Clifford module £ is isomorphic to W ®S, where W is a Hermitian vector
bundle with connection VW.
Let P = Spin(M) be the double-cover of the orthonormal frame bundle
SO (M) corresponding to the spin-structure on M. A point p E Spin(M)
projects to an orthonormal frame at the point xo = 7r(p) E M, which we will
denote by p : V -> T,,0M, where V is the vector space R'. The theorem will
follow from an application of Theorem 6.18, with group G = Spin(n).
Let M0 be a component of the fixed point set y. Decompose V = RI into
the direct sum of V0 = R2e° and V1 = R 2P1. If xo is a point in M0i we will
denote by p a point in the fibre of the bundle Spin(M) at x0 such p restricted
to V0 is a frame of Two M0 C Two M, while p restricted to V1 is a frame of A f.,.
Let %y be the element of Spin(V) such that yu = uy". The element' induces
an endomorphism of S. Since y" lies above a transformation -fl E SO(Nxo)
6.7. Proof of Proposition 6.13 203

such that det(1- yi) # 0, we see that ' belongs to C(Vi), and that its symbol
ate, (7) is non-zero.
We may identify the curvature F W of the connection Vw with an element
of A2 V ®End(WW0 ). Lichnerowicz's formula (Theorem 3.52) shows that D2
OW®s + F° + F1, where Fp° = r /4 and Fp = Fr. Thus, by Theorem 6.18,
we obtain the following result. o
Lemma 6.20. The section I(t, y, 0, x0) has the asymptotic expansion

(4irt)- dim(e)/2-to j ymp e-IIa1I2/4t 00

t(a) p(eXp(a)) jg (a) da,
i=0

where Do (a) equals

/ exp(-F, o .a/2)'Y o O(xo)

det(Q1(-a, _1,)) 11/2 je(a)1/2
j.1/2(T(np
- a/2))
Proposition 5.16 implies the first assertion of Theorem 6.12, as well as giv-
ing us the formula

o'(k('Y, 0, xo)) DO A o2P1(y)

Let us consider the decomposition V = Vo ® Vi = 1[821. ® R2e1. Denote by

eo A eo the antisymmetric n x n-matrix whose (i, j)-coefficient equals ei A e3
if i < 22o and j < 220, and such that all other coefficients vanish. If 4D is
a function on g, we denote by 4D(eo A eo) its evaluation at this element of
® A2Vo . It is clear that D(eo A eo) depends only on the restriction of 1) to
the subspace go = A2Vo of g = A2V.
Since 0'2e1 ('y) E A211 V1, it is clear that

4D(2eAe*) Aa2e1('Y) =' (2eo heo) Ao2e1(5')

We now have the following results:

(1) jg(2eo A eo) = 1
(2) Fl"V - (eo A eo) equals the restriction to M' of the curvature Fw;
(3) if R = R° ®R1 is the decomposition of the curvature of the bundle TMI Mo
of (6.1), corresponding to the decomposition TMI Mo = TMo ® N, it
follows that for a E go = A2Vo,

jv (T (lp . a) /2)) = jvo (T (1Po . a)/2) jv1(T (cP '.a)/2).

It only remains to calculate Q1(2eo A eo, yl).

Lemma 6.21. For v E Vi and a E go, we have

(Q1(a, y)v, v) = 2(v, Jvi (T(SZp a)/2)-1(1 - yl)v).
204 6. The Equivariant Index Theorem

Proof. By Proposition 6.17, we have

a)-iv)-2(v,
(Qi (a, -yi)v, v) = (v, Jv(p, Jv(p, a)-171v)+(71v, Jv(p, a)-171v)
For a E go, the endomorphism Jv (p, a) preserves the decomposition V =
Vo G) V1, and
Jv(p,a)Iv. = Jvi(T(O' -a)/2).
The lemma follows from the fact that Jvl (T(SlP - a)/2) commutes with 7i.
Let A = Jv, (T(SQP1 a)/2)-1. The matrix A -(1 - 71) is not symmetric, so
we must rewrite (Qi (a, -y)v, v) in the form (Bv, v), where B is a symmetric
matrix. Thus, we write
(Qi (a, 7)v, v) = (v, A (1 - 71)v) + (v, (1 - 7i) A*v).
We have A* = eT (OP ' a)/2 A, and hence

- 71) + (1 - 71)* . A* = A -(1 - 71)'(1 _ y

A - (I e'(f2n' a)/2).

Thus for a c- go, we find that

jv1/2(T(Op, a)/2) I det(Q1(a, 71))]-1/2

= jvl/2(T(S2p a/2)) det(1 - 71)-i/2

det(1 - 71 exp(-T(S21 - a)/2)-i/2.

We now replace a by 2eo A eo. Using the fundamental symmetry of the Rie-
mannian curvature as in Lemma 5.17 we see that
jui/2(T(Do
, eo A e0)) = A(M'Y),
which completes the proof of Proposition 6.12.

Bibliographic Notes
The fixed point formula Theorem 6.16 for elliptic operators invariant un-
der a compact Lie group G was proved by Atiyah-Segal [12]; the proof uses
the Atiyah-Singer index theorem [14] and the localization result for the ring
KG(M). The Atiyah-Bott fixed point formula was proved in (5].
The generalization of the local index theorem to a fixed point formula in the
presence of a compact group action, was announced by Patodi [90] in the case
of the 8-operator, and was proved by Donnelly-Patodi [52] for the G-signature.
It was first proved for the twisted G-signature operator by Gilkey [65]. Our
proof of Theorem 6.11 is taken from Berline-Vergne [25], which was inspired
by Bismut [29]. For a proof more in the style of Chapter 4, see the article of
Lafferty-Yu-Zhang [78].
Chapter 7. Equivariant Differential Forms

Although it is rarely possible to calculate the integral of a differential form

exactly, it was a beautiful discovery of Bott that it is sometimes possible
to localize the calculation of such integrals to the zero set of a vector field
on the manifold. In this chapter, we will describe a generalization of this,
a localization formula for equivariant differential forms. Only the results of
Chapter 1 are a prerequisite to reading this chapter.
Let M be an n-dimensional manifold acted on by a Lie group G with
Lie algebra g. A G-equivariant differential form on M is defined to be a
polynomial map a : g -> A(M) such that a(gX) = g a(X) for g E G. The
equivariant exterior differential dg is the operator defined by the formula
(dga)(X) _ (d - c(XM))a(X),
where XM is the vector field on M which generates the action of the one-
parameter group exp tX . The space of equivariant differential forms with
this differential is a complex, with cohomology the equivariant cohomology of
M, denoted HG(M).
The condition that X H a(X) is equivariantly closed means that for each
X E g, the homogeneous components of the differential form
a(X) = a(X) [01 + a(X)[Jl + ...
satisfy the series of relations
t(XM)a(X)[ ] = da(X)[i_2].
These relations imply that a(X) I,,,l is exact outside the set Mo of zeroes of
the vector field XM, if the one-parameter group of diffeomorphism generated
by X is a circle. The localization formula then reduces the integral fm a(X)
to a certain integral over Mo.
Equivariantly closed differential forms arise naturally in a variety of sit-
uations. Bott's formulas for characteristic numbers, the exact stationary
phase formula of Duistermaat-Heckmann for a Hamiltonian action, Harish-
Chandra's formula for the Fourier transform of orbits of the coadjoint repre-
sentation of a compact Lie group, and Chevalley's theorem on the structure
of the algebra of invariant functions on a semi-simple Lie algebra are all con-
sequences of the localization formula. These results will not be needed in the
rest of the book, but are included for their intrinsic interest.
206 7. Equivariant Differential Forms

We will also use equivariant differential forms to give a generalization of the

Chern-Weil map, thereby relating the equivariant cohomology of a G-manifold
M to the cohomology of fibre bundles with M as fibre. As an illustration,
we consider in Section 7 the simple case of a Euclidean vector bundle with an
action of the orthogonal group, which leads to the Mathai-Quillen universal
Thom form of a vector bundle.

7.1. Equivariant Characteristic Classes

Let M be a C°°-manifold with an action of a Lie group G, and let g be the
Lie algebra of G. The group G acts on C°° (M) by the formula (g 0) (x) =
0(g-lx). For X E g, we denote by XM (or sometimes simply X) the vector
field on M given by

(XM . 0) (x) = d - (exp(-eX)x)

E=0

The minus sign is there so that X --> XM is a Lie algebra homomorphism.

On the other hand, if G acts on the right on a manifold P then we must write

(Xp 0) (p) = c(p exp &X) I

E=o

Let C[g] denote the algebra of complex valued polynomial functions on g. We

may view the tensor product C[g] ® A(M) as the algebra of polynomial maps
from g to A(M). The group G acts on an element a E C[g] ® .A(M) by the
formula
forallgEGandXEg.
Let AG(M) = (C[g] ® A(M))G be the subalgebra of G-invariant elements;
an element a of .AG(M) satisfies the relation a(g X) = g a(X), and will be
called an equivariant differential form.
The algebra C[g] ® A(M) has a Z-grading, defined by the formula
deg(P ®a) = 2 deg(P) + deg(a)
for P E C[g] and a E A(M). We define the equivariant exterior differential
d8 on C[g] ® A(M) by the formula
(dea)(X) = d(a(X)) - t(X)(a(X)),
where t(X) denotes contraction by the vector field XM; from now on, we will
frequently write X instead of XM. Thus, d9 increases by one the total degree
on C[g] ® A(M), and preserves AG(M). The homotopy formula L(X)d +
dt(X) = £(X) (see (1.4)) implies that
(d'ca)(X) = -,C(X)a(X),
7.1. Equivariant Characteristic Classes 207

for any a E C[g] 0 A(M), hence (AG(M), d9) is a complex. The elements
of AG(M) such that dga = 0 are called equivariantly closed forms; those of
the form a = dg @ are called equivariantly exact forms. This complex was
introduced by H. Cartan.
Definition 7.1. The equivariant cohomology HG(M) of M is the coho-
mology of the complex (AG(M), do).
We will also consider equivariant differential forms X F-+ a(X) which de-
pend smoothly on X in a neighbourhood of 0 E g, and not polynomially; the
algebra of all of these is written AG (M). Although AG (M) is not a Z-graded
algebra, it has a differential d9, defined by the same formula as before, which
is odd with respect to the Z2-grading.
If H --; G is a homomorphism of Lie groups, there is a pull-back map
AG(M) --> AH(M) on equivariant forms, defined using the restriction map
C[g] --} C[b]. This is a homomorphism of differential graded algebras; that
is, it sends exterior product to exterior product and equivariant exterior diff-
erential dg to exterior differential dry, and hence induces a map HH(M) -->
HH(M).
It is clear that when H is the trivial group {1}, the equivariant cohomology
of M is the ordinary de Rham cohomology. Thus, letting H = {1} and pulling
back by the inclusion of the identity {1} -+ G, we obtain a map AG(M) -
A(M), given explicitly by evaluation a H a(0) at X = 0.
If 0 : N --p M is a map of G-manifolds which intertwines the actions of G,
then pull-back by 0 induces a homomorphism of differential graded algebras
0* : AG(M) -> AG(N). In particular, if N is a G-invariant submanifold of
M, the restriction map maps AG(M) to AG(N).
If M is a compact oriented manifold, we can integrate equivariant differ-
ential forms over M, obtaining a map

: AG(M) -' C[g]G,

JM
by the formula (f M a) (X) = fm a(X); here, if a is a non-homogeneous diff-
erential form, the integral fm a is understood to mean the integral of the
homogeneous component of top degree. If a is equivariantly exact, that is,
a = dev for some v E AG(M), and dim(M) = n, then
a(X)[,,.] = d(v(X)[,,-i]),
so that fm a(X) = 0. Thus, if a is an equivariantly closed form, fm a is an
element of C[g]G which only depends on the equivariant cohomology class of
a.
We will now associate to an equivariant superbundle certain equivariantly
closed forms whose classes in equivariantly cohomology will be called equi-
variant characteristic classes. This construction follows closely the construc-
tion of non-equivariant characteristic forms in Chapter 1.
Recall the definition of a G-equivariant vector bundle Definition 1.5.
208 7. Equivariant Differential Forms

Definition 7.2. If £ is a G-equivariant vector bundle, the space of equivari-

ant differential forms with values in £ is the space

AG(M, £) = (C [o] ®A(M, £))G,

with Z-grading defined by a formula analogous to that on AG(M).

A G-equivariant superbundle £ = £+ ® £- is a superbundle such that £+

and £- are G-equivariant vector bundles. If A is a superconnection on a G-
equivariant superbundle S which commutes with the action of G on A(M, £),
we see that
[A,GE(X)]=0
for all X E g. We say that A is a G-invariant superconnection. If G
is compact, then every G-equivariant bundle admits an invariant connection,
constructed as in (1.10).

Definition 7.3. The equivariant superconnection Ag corresponding to

a G-invariant superconnection A is the operator on C[g] ® A(M, £) defined
by the formula
(Aea)(X) = (A - t(X))(a(X)), X E g,
where t(X) denotes the contraction operator t(XM) on A(M, £).

The justification for this definition is that

Ag(aA0) = dgaA0+(-1)kalaAAgO
for all a E C[g]®A(M) and 0 e C[g]®A(M, E). The operator A. preserves the
subspace AG(M,£) C C[g] ® A(M,£), and we will also denote its restriction
to this space by Ag.
Bearing in mind the formula

dga(X) + £(X)a(X) = 0,
we are motivated to define the equivariant curvature F. of the equivariant
superconnection A. by the formula

e(F9 (X)) = Ag (X) 2 + Ge (X).

If 0 E AG(M, End (6)), let A.0 be the element of AG(M, End(s)) such that
e(Ae0) = [Ag, e(0)].

Proposition 7.4. The equivariant curvature Fg is in AG(M,End(£)),

and satisfies the equivariant Bianchi formula

A9Fg=0.
7.1. Equivariant Characteristic Classes 209

Proof. To prove that F. E AG(M, End(E)), we must show that A9(X)2 +

GE (X) commutes with multiplication by any a E C[g] 0 .A(M):

[As(X)2 +,CE(X),E(a(X))] = [As(X), [As(X),E(a(X))]] + [GE(X),E(a(X))]

= [A9(X), e(d9a(X))] + e(L (X)a(X))
= e(d22a(X) + G(X)a(X)) = 0.
The equivariant Bianchi identity is just another way of writing the obvious
identity [A9(X),A9(X)2 +GE(X)] = 0.
If we expand the definition of F9, we see (identifying F. with the operator
e(F.) on .AG(M,End(£))) that
F9(X) = (A- t(X))2 +,CE(X)

= F - [A,t(X)]+£e(X).
where [t(X), A] = t(X)A+At(X) is the supercommutator, and F = A2 is
the curvature of A. In particular, F. (0) = F. Motivated by this formula for
F9 (X), we make the following definition.
Definition 7.5. The moment of X E g (relative to a superconnection A) is
the element of A+ (M, End(£)) given by the formula
µ(X) = £e(X) - [t(X),A].
Observe that µ = F9 - F is an element of AG(M,End(£)), so that
F9(X) = F+µ(X).
The following formula is a restatement of the equivariant Bianchi identity,
(7.1) Ai(X) = t(X)F.
The use of the word moment is justified by the similarity between this equation
and the definition of the moment of a Hamiltonian vector field in symplectic
geometry, as we will see later.
In the special case in which the invariant superconnection A on the equi-
variant bundle E is a connection V, the above formulas take a simplified form.
Since [V, t(X)] = Vx, we see that the moment µ(X), given by the formula
µ(X) = Le(X) -Vx,
lies in r(M, End(E)), and hence
µ E (g* ®r(M, End(£)))G C AG(M, End(£)).
If the vector field XM vanishes at a point xo E M, the endomorphism
µ(X)(xo) coincides with the infinitesimal action .C(X)(xo) of exp tX in £xo.
We can give a geometric interpretation of the moment µ(X) in this case.
Denote by ir : 9 -+ M the projection onto the base. The action of G on 6
210 7. Equivariant Differential Forms

determines vector fields XE on the total space £ corresponding with X E

The connection V on £ determines a splitting of the tangent space T£,
T£ = 7r*£ ®ir*TM,
into a vertical bundle isomorphic to 7r*£ and a horizontal bundle isomorphic
to 7r*TM.
Proposition 7.6. Let x be the tautological section of the bundle 7r*E over
E. The vertical component of XE may be identified with -p(X)x.
Proof. This follows from Proposition 1.20, once we observe that C'r*e(X)x =
0.

We now construct the equivariant generalization of characteristic forms.

If f (z) is a polynomial in the indeterminate z, then f (FF) is an element of
AG(M, End (9)). When we apply the supertrace map
Str : AG(M, End(£)) -+ AG(M),
we obtain an element of AG(M), which we call an equivariant character-
istic form.
Theorem 7.7. The equivariant differential form Str(f (Fg)) is equivariantly
closed, and its equivariant cohomology class is independent of the choice of
the G-invariant superconnection A.
Proof. If a E AG(M,End(£)), it follows from Lemma 1.42 that
dg Str(a) = Str(Aga).
The equation dg(Str(f(Fg))) = 0 now follows from the equivariant Bianchi
identity AeFg = 0. Similarly, if At is a one-parameter family of G-invariant
superconnections, with equivariant curvature Ft, we have

Str(f(F)) = Str ({A

t
= dg Str( f'(Fg))l
dtg
so that the difference of the two equivariant characteristic forms
Str(f (F')) - Str(f (Fg ))
is the equivariant coboundary of

I 1 Str(
dtg f'(FB)) dt.
As in the non-equivariant case discussed in Section 1.4, we can also allow
f (z) to be a power series with infinite radius of convergence. In particular, the
equivariant Chern character form chg (A) of an equivariant superbundle
7.1. Equivariant Characteristic Classes 211

£ with superconnection A is the element of the analytic completion AG(M)

of AG(M) defined by
(7.2) ch9(A) = Str(exp(-F9)),
in other words, ch9(A)(X) = Str(exp(-F9(X))).
In the case of a G-invariant connection V, the component of exterior degree
zero of the equivariant curvature F. (X) = µ(X) + F is A(X), which satisfies
µ(X) = O(1XI). Thus we can define f(F9(X)) = f(y(X) + F) even if f
is a germ of an analytic function at 0 E g. As an example, we have the
equivariant A-genus A. (V) of a vector bundle £ over a compact manifold
M,
A9(O)(X) = det1/2 Fg(X)/2
sinh(F9 (X) /2) )

Finally, we define the equivariant Euler form. Let £ be a G-equivariant

oriented vector bundle with G-invariant metric and G-invariant connection
V compatible with the metric. The curvature F and the moment p are both
elements of AG (M, so (9)). We define the equivariant Euler form of £ by

Xg(V)(X) = Pf(-F9(X)) = det1/2(-F9(X)),

where the notations are those of Definition 1.35. An argument similar to that
of Theorem 7.7 shows that X9 (V) is an equivariantly closed form, and that its
class in cohomology depends neither on the connection nor on the Euclidean
structure of £, but only on the orientation of £.
We close this section with two important examples of the equivariant mo-
ment map.
Example 7.8. Let us compute the moment µM(X) E r(m, so (TM)) of an
infinitesimal isometry X acting on the tangent bundle of a Riemannian mani-
fold; this is called the Riemannian moment of M. The group G of isome-
tries of M preserves the Levi-Civita connection V on TM. By the definition
of the moment, we see that
µM (X )Y = [X, Y] - OXY,
and from the vanishing of the torsion we obtain
(7.3) µM(X)Y = -VYX,
in other words, µM (X) = -VX. Since L (X) and Vx preserve the metric on
M, µM (X) is skew-symmetric. In this example, (7.1) states that the covariant
derivative of the moment µM is given by the formula
(7.4) [VY, µM (X)] = R(X, Y),
where R is the Riemannian curvature of M.
212 7. Equivariant Differential Forms

The equivariant curvature of a Riemannian manifold M acted on isometri-

cally by a compact group G is defined by the formula R.(X) = R+µM(X) E
AG(M, so(TM)).
Example 7.9. The next example of a moment map is due to Kostant and
Souriau; it is fundamental to the theory of geometric quantization. Let M be
a symplectic manifold with symplectic two-form S2, that is, Il E A2 (M)
is a closed two-form, df l = 0, such that the bilinear form cl (X, Y) on T.M
is non-degenerate for each x E M. If f E C°°(M), then the Hamiltonian
vector field generated by f is defined as the unique vector field H f such that

df = t(Hf)1l.
Assume that a Lie group G acts on M and that the action is Hamiltonian:
this means that, for every X E g, there is given a function µ(X) on M such
that
(1) µ(X) depends linearly on X;
(2) the vector field XM is the Hamiltonian vector field generated by µ(X),
that is
dg(X) _ t,(XM)Q;
(3) p(X) is equivariant, that is, g µ(X) =µ(g X) for g E G.
The symplectic moment map of the action is the C°° map µ : M -> g*
defined by (X, p(m)) = µ(X)(m). One sees easily that
X 1)920(x)=A(X)+Q
is an equivariantly closed form on M such that Q,(0) = Q.
This equivariant differential form may in some cases be identified with i
times the equivariant curvature of an equivariant line bundle. Let .C be a
complex line bundle on M. Suppose that L carries a connection V whose
curvature
(v) 2 = is E A2(M)
equals i times the symplectic form 0 on M. Furthermore, suppose that the
Lie group G acts on the manifold M and on the line bundle L, in such a way
as to preserve the connection V L. It follows immediately that the symplectic
form SI is preserved by the action of G.
We may define the moment of this action, pC(X) = C'c(X) - V j, and the
equivariant curvature
Fg = µn + ifl E A2 (M).
In fact, ulc equals v '-I times the moment u defined above, since by (7.1),
ib(XM)SZ = dµc(X).
It follows that Fg = Mg.
7.2. The Localization Formula 213

7.2. The Localization Formula

If M is a manifold acted on by a compact group G, there is a Riemannian
metric on M which is invariant under the action of G. For example, such
a metric may be formed by averaging any metric on M with respect to the
Haar measure of G. In this section, we will show how such a metric may be
used to study equivariant differential forms.
Proposition 7.10. Let G be a compact Lie group and X --> a(X) be an
equivariantly closed differential form on M. If X E g, let Mo(X) be the set
of zeroes of the vector field XM. Then for each X E g, the differential form
a(X)[n] is exact outside Mo(X).
Proof. Fix X E g and write dx = d - t(X ). Let 0 be a differential form on M
such that L(X)9 = 0 and such that dX9 is invertible outside Mo(X). Such a
differential form may be constructed using a G-invariant Riemannian metric
on M. We define 9(6) = (XM, e) for t; E I'(M,TM), then 0 is a one-form
on M invariant under the action of X such that dx9 = IX12 + d9, which is
clearly invertible outside the set Mo (X) = { X 1 2 = 0}.
On the set M - Mo(X), we have
9 n a(X)
(ce (X)[n] d
= dx9 [n-1]
for every equivariantly closed differential form a on M. Indeed, since dXB =
0, we have
C9na(X))
a(X) = dX
dx9 ),
and the result follows by taking the highest degree piece of each side.
Note that it is essential to assume that G is compact. For example, consider
M = Sl x Sl with coordinates x, y E R/27rZ. Let X be the nowhere-vanishing
vector field (1 + 2 sin x) ay, and let
a(X) 2 (7 cos x + sin 2x) + (1 - 4 sin x)dx n dy.
It is easy to verify that dxa(X) = 0. However, fm a(X) = (2ir)2, so that
a(X) [2) is not exact.
Proposition 7.10 strongly suggests that when G is a compact Lie group
and M is compact, the integral of an equivariantly closed form a(X) depends
only on the restriction of a(X) to Mo (X ). In the rest of this section, we will
prove the localization formula, which expresses the integral fm a(X) of an
equivariantly closed differential form a as an integral over the set of zeroes
of the vector field XM.
We will first state and prove the localization formula in the important
special case where XM has isolated zeroes. Here, at each point p E Mo(X),
the Lie action L(X)l; = [XM,1;] on r(M, TM) gives rise to an invertible
transformation LP of TPM. This can be proved using the exponential map
214 7. Equivariant Differential Forms

with respect to a G-invariant Riemannian metric: if E TpM was annihilated

by LP, all the points of the curve expp(s6) would be fixed by exp(tX). The
transformation LP being the Lie derivative of an action of a compact Lie group,
it has only imaginary eigenvalues. Thus the dimension of M is even and there
exists an oriented basis ei, 1 < i < n, of TTM such that for 1 < i < 2 = n/2,

Lpe2i-1 = Aie2i,
Lpe2i = -Aie2i-1

We have det(Lp) _ A2A2...A2 , and it is natural to take the following square

root (dependent only on the orientation of the manifold):

det1/2(Lp) = A1... ae.

Theorem 7.11. Let G be a compact Lie group with Lie algebra g acting on a
compact oriented manifold M, and let a be an equivariantly closed differential
form on M. Let X E g be such that XM has only isolated zeroes. Then

IM
a(X) = (-27) t \' ,I
a (X)(p )
1/2(Lp )'
pEMo(X)

where e = dim(M)/2, and by a(X)(p), we mean the value of the function

a(X) [0) at the point p E M.

Proof. Let p E Mo(X). Using a G-invariant metric and the exponential

map, the vector field X can be linearized around p. Thus there exist local
coordinates xl, ... , xn around p such that XM is the vector field

XM = Ai(x231 - xia2) + ... +.e(xnan-1 - xn-lan),

and det112(Lp) =Al... At.

Let Bp be the one-form in a neighbourhood Up of p given by

Bp = Aj 1(x2dxl - xldx2) +. .. +Ae 1(xndxn_1 - xn_ldxn).

Then Bp is such that L(X)OP = 0 and B2 (XM) = >i xi = 11x112. Using

a G-invariant partition of unity subordinate to the covering of M by the
G-invariant open sets Up and M - Mo(X) (which may be constructed by
averaging any partition of unity with respect to the action of G), we can
construct a one-form 0 such that £(X)0 = 0, dX9 is invertible outside M0(X),
and such that 0 coincides with Bp in a neighbourhood of p.
7.2. The Localization Formula 215

Consider the neighbourhood BE of p in M given by BE = {x I IIx112 < e}.

If SE = {x I IIx112 =/s}, then
a(X) = lim a(X)
J JM E-40 J
(0Aa(X)1
= 1im
f
E->0 M-UpB'
d

0Aa(X)
dxO ))

hm
P E--+o sE dx0
(The sign change comes from exchanging the interior for the exterior orienta-
tion of S. P.). Let us fix a point p. Near p, 0 = 0P. Resealing the variable x by
a factor of e /2, the sphere SE becomes the unit sphere S1, while 0(dx0)-1

being homogeneous of degree zero is invariant. Then

/ 0Aa(X) _ /' 0AaE(X)
Js dx0 i s1 dxO
where c(X)(x, dx) = a(X)(s1/2x, E1/2dx). When s -> 0, ae(X) tends to the
constant a(X) (p). To prove the proposition, itf remains to compute

- fs, 0(dx0)-1 =

JS1
0(1 - d0)-1 =
l
0(d0)Q-1 =
JB l
(d0)e

But
(d0)' = (-2)e1! (Al ... At)-l dxl A... A dx,.
Since the volume of the 2$-dimensional unit ball equals ire/2!, we obtain the
theorem.
We now turn to the localization formula in the general case. Fix an element
X e g, and let Mo denote the set of zeroes of X.
Proposition 7.12. The set Mo of zeroes of the vector field XM is a sub-
manifold of M, which may have several components of different dimension.
The normal bundle Al of Mo in M is an even-dimensional orientable vector
bundle.
Proof. Fix a G-invariant Riemannian structure on M. If x0 is a zero of XM,
the one parameter group exp tX acts on Two M. If a tangent vector at x0 is
fixed by exp tX, then the geodesic tangent to it is contained in M0, which
proves that Mo is a submanifold of M.
The Lie derivative Y - L(X )Y = [XM, Y] is an endomorphism of Nxo =
Two M/T o Mo which is invertible and antisymmetric. This implies that the
real vector bundle Al can be given a complex structure and is therefore even-
dimensional and orientable. Indeed, observe that the eigenvalues of C(X) in
.IV:,C:o = Nxo ®] C come in pairs ±iAj, with A; > 0. Therefore, we have a
decomposition
N-CO

where N ,, is the sum of the eigenspaces such that Aj > 0.

216 7. Equivariant Differential Forms

Let rk(N) : Mo --> Z be the locally constant function on Mo which gives

the codimension of each component. Let go = {Y E g [Y, X] = 0} be I

the centralizer of X E g, and let Go be the connected Lie subgroup of G

with Lie algebra go. Then Go preserves the submanifold Mo and acts on
the normal bundle N. We choose a Go-invariant Euclidean structure and a
Go-invariant metric connection ON on N. The normal moment of Y E go is
the endomorphism ,N(Y) of N such that

LN(Y) = ON + µN(Y).

In particular, since the vector field XM vanishes on Mo, we have

µN(X ) = [XM, C] for l; E N,
and µN (X) is an invertible transformation of N. Choose an orientation on N
and consider the equivariant Euler form xgo(N) E A+(Mo). Its degree zero
piece equals det(-µN)1/2, and therefore xg0(N)(Y) is invertible in A+(MO)
when Y is sufficiently close to X. We can now state the general localization
formula.

Theorem 7.13. Let G be a compact Lie group acting on a compact manifold

M. Let a be an equivariantly closed form on M. Let X E g, let Mo be the
zero set of the vector field XM and let N be the normal bundle of Mo in
M. Choose an orientation on N and impose the corresponding orientation
on Mo. Then for Y in the centralizer go of X in g and sufficiently close to
X, we have
_rk(N)/2 a(Y)
fM
a(Y) fMO(2 x9o (N ) (Y)
where x20 (N) is the equivariant Euler form of the normal bundle. In partic-
ular,
a(X) = (-27f)rk(N)/2 12
X) + R)
IM 1M0 det / (L (X)
The proof follows from several lemmas. As before, let 0 be the Go-invariant
one-form on M dual to the vector field XM with respect to the G-invariant
Riemannian metric on M,
8(Z) = (XM, Z) for Z E r(M,TM).
Then we have
(dg8) (Y) = dO - (XM, YM), for Y E go.

Lemma 7.14. For all t E R and Y E go,

fm a(Y) = fm et(d0e)(Y)a(Y).
7.2. The Localization Formula 217

Proof. We give two arguments, the second of which extends better to non-
compact manifolds. The first proof follows from the fact that etdgO - 1 is
equivariantly exact:
etde9 - l (0(etd 1_ 1) \
- dg JI
d9

The other argument is analogous to the proof of the McKean-Singer formula

Theorem 3.50: we observe that
d et(d,O)(Y)o,(y)

dt M,f = f (dse)(Y)et(d0B)(Y)a(Y)
M

fM de(OetdHea)(y) = 0,
=
since a is equivariantly closed. Thus, we may set t = 0 without changing the
integral, whence the result.
Thus, we must compute the limit when t -> 0 of
n/2
e(dae)(Y)/ta(y) = e(XY)/ta(Y) tk
IM fM k!
k=0
To do this, we study the form 0 in the neighbourhood of Mo.
By orthogonal projection, the Levi-Civita connection V gives a connection
V'v on the normal bundle N which is compatible with the induced metric.
Identifying Mo with the zero section of N, we obtain a canonical isomorphism
TNIMO = TMIM0.
Consider the moment map pM(Y)Z = -VzY of TM. Since am(Y) com-
mutes with the operator £(X) on Mo, it preserves the decomposition
. TMJM0 =TMo ®N;
thus, the restriction of µM (Y) to N coincides with the moment endomorphism
1'(Y). The endomorphism e(X) of Al coincides with the infinitesimal
action L(X). The action of Go on the manifold N determines vector fields on
N. The vector field Xj.r is vertical and is given at the point (x, y) E Mo x N"

The connection J
by the vector -pm(X)y E N,.
determines a splitting of the tangent space to the total
space N into horizontal and vertical tangent spaces. We denote by (vi, v2) o
the Euclidean structure of the vector bundle N. Consider the one-form 80 on
the total space N given by
(7.5) 00(Z) = (XN, Zv)o
where Zv is the vertical part of a vector field Z on N.
To prove Theorem 7.13, we could proceed as in the proof of Theorem 7.11,
multiplying the one-form 00 by a cut-off function to obtain a one-form 0 which
coincides with 00 in a neighbourhood of Mo. However, we will see in the next
218 7. Equivariant Differential Forms

lemma that this is not necessary, since Bo arises naturally as a scaling limit
of the one-form 0.
Let be a diffeomorphism of a neighbourhood U of the zero section in N
with a neighbourhood of Mo in M, such that
O(x, 0) = x for x E Mo,
do I M, = I on T Al Mo = TM I MO .

An example of such a map is the exponential map 0 (x, y) = expx y, but the
exponential map is difficult to compute explicitly in most cases, so that it is
interesting to permit a different choice of &i. Transport the metric and
the one-form 8 to U, by means of the diffeomorphism ii. Consider the dilation
St, t > 0, of Al given by the formula
St(x, y) = (x, t112y), where x E Mo and y E NM.
We assume that U is preserved by St for t < 1.

Lemma 7.15. (1) dO(Z1, Z2) = -2(µm(X)Z1i Z2) for Z1, Z2 E r(M,TM)
(2) lim t-16t 0 = Bo
t-+o
(3) For Y E go, lim t-1 (X, Y) (.,tii2y) = (µN(X )Y, µN(Y)Y)o
t--+o

Proof. It is easy to calculate dB, using the explicit formula for the exterior
differential of a one-form: if Z1 and Z2 are vector fields on M, then

dO(Z1,Z2) = Z1(X,Z2) -Z2(X,Z1) -(X, [7'1,Z2])

= (Vz1X, Z2) - (Vz2X, Z1)
= -2(pM(X)Z1i Z2).
Since Am(X)Z1 = 0 if Zl is tangent to Mo, we see that dO(Z1iZ2) vanishes
if Z1 or Z2 is tangent to Mo. Consider local coordinates (x, y) on U such
that Mo is the set y = 0. Since 0 vanishes on Mo, as does t(Z1)dO if Z1 is
tangential to Mo, we see that

0= aijk(x,Y)yty dxk + bij(x,Y)Yidy ,

ijk ii
for some smooth coefficients aijk(x, y) and bid (x, y). This shows that

t-1St 8 = E aijk (x, t1/2Y)Yiyjdxk + 57 bij(x, t1/2Y)Yidyj

ijk ij
has a limit when t -+ 0:

lim t-16t 0 = E aijk(x, 0)Yiyjdxk + bij(x, 0)Yidyj.

t--+0 ijk ij
7.2. The Localization Formula 219

In order to obtain (2), it remains to show that

(7.6) ayie(ay,)IY=o = aYieo(ay,)IY=o

and

(7.7) aYiaY,B(axk)1Y=0 = aYiaY,eo(axk)IY=o.

Denote by O the one-form on MO which represents the connection V in the

local basis ayi of the normal bundle N. Thus
va k aYi = e(axk )ayi'

By Proposition 1.20, we have

90 = -(pN(X )y, dy + ©y)0.
Since X vanishes on Mo, we have on MO the formula

ayi(X,0Y2)IY=o = (V8 X,0y3)IMo

_ -(µN(X)aYi,(9Y;)o,
which proves (7.6).
Let us prove 7.7. We have
ayiay,eo(axk)IY=o = -(pm(X)aYi, o(aZk)DY.)o - (N'M(X)aY,,e(axk)aY.i)o.
By (7.4), we have
Vy1Vy2X = -Vyy(/M(X)Y2) _ -^X)vyyY2 - R(X,Y1)Y2
If Z is tangential to Mo, so that pM(X)Z and R(X,Y1) vanish on Mo, we see
that
(Vy1VY2X, Z)IMo = (VYlY2, pm(X)Z)I Mo - (R(X,Y1)Y2, Z)IMo = 0.
It follows that
(7.8) Y1Y2(X,Z)IMo = (VY1X,VY2Z)IMo +(VY2X,VY1Z)IMo
We apply this with Z = axk , Y1 = ayi , and Y2 = ay1. Using the relation
pay. axk = pack ayi and the fact that µM (X) preserves the decomposition
TMIMO =TM0 ®N, we obtain
(VoyiX,paysaxk)IMo =-V(X)aYi,V kaY,)o
-(p ' (X) ay,, e(axk)aY,)o,
and hence

dytay,e(axk)Iy=o = -(e(X)aYi,e(axk)aY,)o - (e(X)ay,,©(axk)ayi)o

= 9yi(9Y,e0 ( ,)IY=O,
proving the second part of the lemma.
220 7. Equivariant Differential Forms

The vector field X, v is vertical and is given at the point (x, y), x E M,
y E Nom, by the vector -µN(X)y E .M . If Y E go, the vector field YM
commutes with XM, and hence is tangent to Mo, so that the function (X, Y)
vanishes to second order on Mo. Equation (7.8) shows that

Y1Y2(X,Y)I Mo = (pN(X)Yi.,1jr(Y)Y2)o + (kCN(X)Y2, K'(Y)Y1)o,

which proves the third part of the lemma. 0
Using the diffeomorphism 0, we can identify the neighbourhood U of the
zero section in N with a neighbourhood U of Mo in M. Outside the neigh-
bourhood U, the function (X, X) is strictly positive, so we can find e > 0
such that if Y E go is sufficiently close to X, (X,Y)y > e for all y U. Since
dO is nilpotent, a partition of unity argument now shows that

lim e-(X,Y)/teda/ta(Y) = lim J e-(X,Y)/tede/ta(Y)0

t->0 JM t-0 u
if 0 is a compactly supported function on U equal to one on a neighbourhood
of Mo.
Let us consider a(Y)o as a differential form on N. If we denote the one-
form a(Y) on Mo and its pull-back to N by the same notation, we see that
limt-o St (a(Y)q5) = a(Y)IMo. Thus, it follows from Lemma 7.15 that
e-(rAN(X)Y,JLM(Y)Y)oedeoa(y)I
lim 8t (e-(X,Y)/tede/ta(Y)0) = Mo
t-+o
Furthermore, for Y near X and y in a neighbourhood of Mo, we have a bound
(X, Y) (.,y) > cI I Y II2 with c > 0, so that by dominated convergence and change
of variables (x, y) i- (x, tl/2y), we see that

lim e-(X,Y)lted6/ta(Y)0 e-(,A1(X)Y,,Ar(Y)Y)oedeo a(Y)I Mo

t oL _ fN
Consider the differential form defined by integration over the fibre,

JN/Mo e - (µN (X )Y,µN (Y)Y)o edeo

We have
00 = -(li'(X)Y,ONY)o
Recall that V is invariant under.C(X), so that [Vu, e (X)] = 0. We see
that dOo equals
-(uN(X)VNY,V y)o - (pN(X)Y,RY)o
It follows that
e-(µN(X)Y,µN(Y)Y)oed0o = e-(I,N(X)Y,(F'N(Y)+RRr)Y)o-(,-A (X)VNY,VAY)o.
7.3. Bott's Formulas for Characteristic Numbers 221

The proof of the theorem is completed by the following lemma. Here, the
element xgo(.N)(Y) = det1/2(-(pN(Y) + RN)) is an invertible element of
A(Mo), for Y sufficiently near X.

Lemma 7.16.

I e-(,,N(X)Y,(µN(y)+RN)Y)o-(µN(X)vNy,o"(y)o - (27r)rk(N)/2
JAr/Ma x90 (.N) (Y)

Proof. Choose a point xo E Mo and let V = .M . Given a local oriented

orthonormal frame of 1V around xo, we may write ON d + 4. We must
show that

e-(,u(X)Y,(µ(Y)+R)Y)-(M(X)dY,dY) _ (-27r)rk(N)/2 det1/2(µ(X))

Iv det(µ(X)(µ(Y) + R))1/2

where µ = Lo E g® ® End(V) and R = R o E A2T*x0Mo ® End(V).

It is a simple matter to evaluate this integral. First we apply the definition
of the Pfaffian, which shows that

-(µ(X)Y,(A(Y)+R)Y)-({i(X)dY,dY)
IV e

_ (-2)n/2 e-(W(X)Y,(W(Y)+R)Y) det1/2(µ(X))dYl A ... A dyn.

v
Recall the formula for a Gaussian integral: if A is a positive-definite endo-
morphism, then
e-(Y,AY) dy = rn/2 det(A) -1/2.
IV

Since µ(X) and µ(Y) + R are antisymmetric and commute, the matrix

-µ(X) (µ(Y) + R)
is symmetric. Its zero-degree component -µ(X)µ(Y) is positive definite when
X = Y, and hence for all Y sufficiently close to X. Hence, we obtain

I V
e-(,u(X)Y,(,u(Y)+R)Y)dyi A... A dyn = 7rn/2 det(-µ(X)(µ(Y) + R))-1/2

= 7rn/2 det-1/2(µ(X)) det-1/2(µ(Y) + R).

7.3. Bott's Formulas for Characteristic Numbers

In the next three sections, we will give some applications of the localization
formula. The first of these predates this formula, and is due to Bott.
222 7. Equivariant Differential Forms

If 41 E C[so(n)]°() is an invariant polynomial function on the Lie algebra

so(n), where n = 22 is even, then 1D is uniquely determined by its restriction
to the Cartan subalgebra t C so(n) of matrices X of the form

Xe2i-1 = xie2i,
Xe2i = -xie2i-1

Let pk be the k-th elementary symmetric polynomials in A, given by the

formula
2 2
Pk = xil ... xik
1Cil <...<ik<e

(In particular, pt = det(X).) If' is O(n)-invariant, the restriction of to t

is a symmetric function of the variables x?, so belongs to the polynomial ring
generated by pi, 1 < i < Q, that is, C[so(n)]0ln) = C[p1, ... ,pe]
The function detl"2(X) = xl ... xe = p1/2 is only invariant under the
group SO(n), and is the additional generator of the ring of SO(n)-invariant
polynomials over the ring of 0(n)-invariant polynomials; it is nothing but the
classical Pfaffian function Pf(X).
Let M be a compact oriented Riemannian manifold of dimension n =
2t, and let R be the curvature of a connection on the tangent bundle TM
compatible with the metric. The Chern-Weil map defines a homomorphism

C[so (n)]SO(n) = C [pi, ... , Pt, pg12] --> A(M),

by the formula and -D(R) is a closed form on M whose coho-

mology class is independent of the choice of connection. Define

f
(-27r)-t
,D(M) =

The number (M) is called a characteristic number of M; it is a classical

theorem of algebraic topology that if -P E Z [pi, -P(M) is an integer.
Suppose that M admits a circular symmetry with isolated fixed points. At
each fixed point p, the action of the infinitesimal generator X of the circle
group G = {e2'ri8} on the tangent space TpM determines an endomorphism
Lp, given in an appropriate oriented basis by

Lpe2i-, = -Aie2i,
Lpe2ti = Aie2i-1,

where Ai are integers called the exponents of the action at p. Recall that by
definition det112(Lp) = al ... Ae, and depends on the orientation of TpM but
not on its metric.
7.3. Bott's Formulas for Characteristic Numbers 223

Theorem 7.17 (Bott). Let M be a manifold on which the circle acts with
isolated fixed points. If 1) is a homogeneous polynomial of degree k < Q,
(D(Lp) (D (M) k = Q,
detl/2(Lp) 0, k < P.

Proof. Identify the Lie algebra of G with {uX I u E R}. Choose a G-invariant
metric, and let uum (X) + R be the equivariant curvature of the associated
Levi-Civita connection. If -D is a polynomial, u --p 4D(upM(X) + R) is an
equivariantly closed form on M, and we see from Theorem 7.11 that
u_$ uLp
(-27r)-'e fM (uµM(X) +R) = detl/2(Ln)
The left hand side of this equation is a polynomial in u with value at zero the
characteristic number (D (M), while the right hand side is a Laurent polynomial
in u. The equality of the two sides implies remarkable cancellations properties
over fixed points of the values of -1) (Lp) if deg(d)) < 2, while the equality of the
constant term of this Laurent polynomial gives the formula when deg(4D) _
2.

When applied to the Pfaffian function 4(X) = Pf(X) on so(n), the char-
acteristic number (-27r)-t fm Pf(R) is (-1)E times the Euler characteristic of
M, by the Gauss-Bonnet-Chern theorem, and we see that the number of fixed
points of the circle action is the Euler characteristic of M. Since the vector
field X generating the circle action is an infinitesimal isometry, it always has
index v(p, X) = 1 at an isolated zero X (p) = 0. Thus, this result is a very
particular case of the Poincare-Hopf theorem, Theorem 1.58.
The general localization theorem allows us to generalize the above formula
to situations in which M° (X) is not zero-dimensional; this extension was
made by Baum and Cheeger. Let us merely mention the case of the Pffafian,
in which we obtain the following result.
Proposition 7.18. Let M be a compact oriented even-dimensional Rieman-
nian manifold, and let X be an isometry of M with fixed point set Mo. Then
the Euler characteristic of M equals the Euler characteristic of M°.
Proof. Observe that
detl/2(-Rg
detl/2(-R9(X))IMo = (X)) det1/2(-R°),
where R° is the Riemannian curvature of M°. By the localization theorem,
we see that
(27r)
-n/2 dim(Mo)/2 det1/2(-R°).
J det"2(-Rg(X)) = fM0
Setting X = 0 and applying the Gauss-Bonnet-Chern theorem to both sides,
the result follows.
224 7. Equivariant Differential Forms

7.4. Exact Stationary Phase Approximation

One important application of the localization formula is the "exact stationary
phase approximation" of Duistermaat-Heckmann.
Let M be a compact manifold of dimension n = 2P, f a smooth function
on M and let dx be a smooth density on M. Let t E IR and consider the
function
F(t) = e'tf dx.
JM
The major contribution to the value of this integral when t tends to infinity
arises from the neighbourhood of stationary points, that is, points where
the differential df of the phase function f vanishes. Assume that the phase
function f is non degenerate, which means the set Mo where the differential
of f vanishes consists in a finite number of points and at these points the
Hessian Hp = Opdf of f is a non-degenerate quadratic form on TpM. Let
a(Hp) be the signature of the quadratic form Hp. If TpM = Tp+ ® Tp is
an orthogonal splitting for Hp such that Hp is positive definite on T+ and
negative definite on T- then a(Hp) = dim(T+) - dim(T-).
If e2 is a basis of TpM such that (Idxl p, el n ... A e,,,) = 1, let

a,, (f, dx) _ det(Hp(ei, ej))1-1/2.

It is not hard to show that when t tends to infinity,

(2) Ieri-(Hp)/4ap(f, dx)eitf (p) +O (t-e`1 ) .

F(t) = pEMo

Duistermaat and Heckman discovered a class of examples of "exact stationary

phase approximation", where the error term in the above formula vanishes.
Let (M, )) be a compact symplectic manifold of dimension n, and let G be
a compact group of Hamiltonian transformations of M. For X E g, let j.c(X)
be the symplectic moment of X, defined in (7.9). Since dµ(X) = t(XM)Q,
the set of points where the one-form dp(X) vanishes coincides with the zero
set Mo(X) of the vector field XM. For p E Mo(X), we denote by Gp(X) the
infinitesimal action of X on TpM. The Liouville form of the symplectic
manifold M is the form

dp = (est/2") =
1n1 = (2ir)e2!'

this is a volume form which defines a canonical orientation on M.

Theorem 7.19 (Duistermaat-Heckman). Let (M, S2) be a compact sym-

plectic manifold, with compact Hamiltonian symmetry group G. If X is an
7.4. Exact Stationary Phase Approximation 225

element of the Lie algebra of G such that Mo(X) consists of a finite number
of points, then
eilL(X)(p)
eiIA(X)di = ie
fm pEM0(X) detl/2(Gp(X))

(The square root of det(Lp(X)) is computed with respect to the canonical

orientation of TpM.)
Proof. The equivariant symplectic form S28 (X) _ {.c(X)+S2 is an equivariantly
closed form on M, as is
ein,(X) = e(X)e.
The theorem follows from the application of the localization theorem, Theo-
rem 7.11 to the integral
ei.Q9(X)
f eiµ(X)d,(3 = (27ri)-e

M JfM
Corollary 7.20. Let X be a Hamiltonian vector field with Hamiltonian f
on a compact symplectic manifold M, such that the flow generated by X is
periodic, and X has discrete zeroes. Then the error term in the stationary-
phase approximation to the integral fm eitf dQ vanishes:
(2'7r)ee7,iv(Hp)/4ap(.f, dx)eZtf(p)
f Meitf di3 = E
pEMo
t
Proof. We need the following lemma.
Lemma 7.21. The Hessian Hp(Y, Z) of the function f at the point p E
M0(X) is given by the formula
Hp(Y, Z) = -ct(Lp(X)Y, Z).
Proof. Since L(Z) f = t(Z)a(X)52, we have
L(Y),C(Z) f = c(Y)(t(Z)t(X)cI)
= t(G(Y)Z)t(X)c + t(Z)t(C(Y)X)n + t(Z)t(X)(L(Y)c ).
Since X vanishes at p, we obtain Hp(Y, Z) = -S1([X,Y], Z), proving the
lemma.
By the lemma, we see that
ap(f, dO) = (2ir)-'l det(Gp(X))I-1/2
Thus, we must prove that for each critical point p of f,
(7.9) sgn(detl/2(Gp(X))) = i-n/2 evric(Hp)/4
Let (V, SZ) be a symplectic vector space, and let A be an invertible semisim-
ple endomorphism of V with purely imaginary eigenvalues. The quadratic
226 7. Equivariant Differential Forms

form H(v, w) = -Sl(Av, w) is non-degenerate, and A is a skew-adjoint endo-

morphism of V with respect to H. Choose a basis for V of vectors pi and qj,
1 < i < 2, such that 9(pi, qj) = bij, f(pi,pj) = 1(gi,gj) = 0, Api = )tigi, and
Aqi = -Aipi We see that
t
sgn(H) = 2 sgn(Xi),
i=1

while

detl/2(A) _ I Ai.
It is now easy to check (7.9).
Of course, using the general localization formula, it is possible to remove
the condition in the above theorem that Mo (X) is zero-dimensional, at the
cost of introducing the curvature of the normal bundle to M0(X).

7.5. The Fourier Transform of Coadjoint Orbits

Let G be a Lie group with Lie algebra g, and let g* be the dual vector space to
the vector space Z. The group G acts on g by the adjoint action, and the dual
action of G on g* is called the coadjoint action. One of the most important
examples of a Hamiltonian action arises in this situation.
Let M be an orbit of the coadjoint representation; thus, for some f E
g*, M = G f . The vector fields XM, X E g, are sections of TM and
span the tangent space at each point, and we have (XM) f = -X f, where
- (X f) (Y) = (f, [X, Y]). Let A (X) be the restriction to M of the linear
form f f (X) on g*.
Lemma 7.22. The form S2(XM,YM) f = -(f, [X, Y]) defines a G-invariant
symplectic form on M. Furthermore the action of G on M is Hamiltonian
and the symplectic moment of X is the function µ(X).
Proof. The form 0 is clearly non-degenerate and G-invariant. The function
µ(X) being the restriction of a linear function obviously satisfies

dµ(X)(YM)f = (YM)f(X) = -(f, [X, Y]),

and hence dlz(X) = t(XM)Sl, which shows that µ(X) is the Hamiltonian for
the vector field XM. Furthermore, it follows that
0 = d2k(X) = dt(XM)S2 = -c(XM)dc,
where in the last equality, we used the invariance of 1, L(XM)1 = 0. Since
the vector fields XM span the tangent space, we see that dS2 = 0.
7.5. The Fourier Transform of Coadjoint Orbits 227

Let d,3 be the Liouville measure on M. By the Fourier transform of an

orbit, we mean the integral

FM(X) = e ,X)df3;
fM
this is a generalized function on g if the orbit M is sufficiently well-behaved.
Such integrals are very important in representation theory. The above lemma
shows that such a Fourier transform is the integral of an equivariantly closed
differential form, since

FM(X) = (27ri)-n12 f easze(X)

M

When G is a compact Lie group, the localization theorem will lead to a

formula for the Fourier transform of the. Liouville measure of M, due to
Harish-Chandra. This formula in turn implies Chevalley's theorem on the
structure of C[B]G. To state Harish-Chandra's formula precisely, we first
recall a few results on the structure of compact Lie groups.
Let Be be a reductive Lie algebra over C; in other words,

Be = 3 E ) Be],

where 3 is the centre of Be and [gc, Be] is semisimple. Consider the collection
of all abelian subalgebras a of Be such that the transformations

{ad X I X E a} c End(gc)
are simultaneously diagonalizable; a maximal subalgebra among this set is
called a Cartan subalgebra, and as is well known, any two Cartan subalgebras
are conjugate under the action of the adjoint group Ad(gc).
Choose a Cartan subalgebra h of Be. If a E F*, define
'(Be). = {X E gc I [H,X] = (a,H)X}, for all H E h.

If a # 0 and (gc)a # 0, a is called a root of h in Be. The set of roots of h in

Be is denoted by A(gc,1)), or A when Cl is fixed. If a E A, then dim(gc)a = 1,
-a E A and dim[(gc)a, (gc)_a] = 1. Furthermore, [(gc)a, (gc)_a] C C), and
there exists a unique element Ha E [(gc)a, (gc)_a] such that (a, Ha) = 2.
For a E A, we may consider the reflection of b about the plane orthogonal
to Ha, sa(H) = H - (a, H)Ha, which satisfies s2 = 1. The subgroup W =
W (gc, h) of transformations of Cl generated by the reflections sa is a finite
group, called the Weyl group. We denote by e(w) = ±1 the determinant of
the transformation w E W.
The Killing form B (X, Y) = Tr (ad X ad Y) is positive definite on the real
span Ea RHa. The vector Ha = 2B(Ha, Ha)-'H. E F) is the unique element
of h such that for all H in C7, B(Ha, H) = (a, H).
228 7. Equivariant Differential Forms

A positive system P of roots is a subset of A such that for a E A, either

a or -a, but not both, belongs to P, and such that if a E P, /3 E P and
a+0EAthen a+/3EP. We define
PP = 1
2 E a.
aEP
A simple root with respect to a fixed positive system P is a root which cannot
be written as the sum of two elements of P. The set {sa I a simple} is a set
of generators for W. If w E W, we denote by Jwl the length of w with respect
to this set of generators.
A compact connected abelian Lie group T is called a torus; it is the quotient
of its Lie algebra t by a lattice. If T acts on a finite dimensional real vector
space V, then V = Vo ®EkEt'\{o} V[k]; where Vo is the subspace of V fixed
by T and V [k] is an even-dimensional real vector space on which the spectrum
of the action of t is ±ik.
If G is a compact connected Lie group, every element of C belongs to a
torus and any two maximal tori of G are conjugate under the adjoint action of
G. The complexification gc of the Lie algebra g of G is a complex reductive Lie
algebra. If T is a maximal torus in G with Lie algebra t, we have g = t ®t,
where t = [t, g], and thus g* = t* ® r*, and tc is a Cartan subalgebra of
gc. Every element of g is conjugate to an element of t by the action of an
element of G, and every element of g* is conjugate to an element A E V. Let
A = A(gc, fe). Roots a E A take imaginary values on t, and iHa E t for
all a E A. The group T acts on t and we write r[za) for the two-dimensional
subspace of r on which X E t acts by the infinitesimal rotation of angle ia(X).
If T is a maximal torus in a compact Lie group G, let
N(T)={gEGIgTg-'=T}
be the normalizer of T in G. The group N(T) acts on t by the restriction of
the adjoint action. Since T acts trivially on t, this defines an action of the
quotient group W(G, T) = N(T)/T, which can be identified with the Weyl
group W(gc, fe). We will denote it by W.
For a compact Lie group G, every coadjoint orbit is of the form MA = G A,
with A E t*. We will give an explicit formula for the Liouville measure of an
orbit MA and for its Fourier transform FA(X). For A E t*, let PA be the set
of roots
PA=fee EAI(A,iHa)>0},
and let to = aEP t[z) The orbit MA may be identified with the homo-
.

geneous space G/Gta), where G(A) is the stabiliser of A. The space to is

isomorphic to the tangent space Te(G/G(A)) at the base point e = G(A). In
particular, if (A, iHa) 0 0 for every a E A, then G(A) = T and to = r.
Such a point A is called regular and its orbit has maximal dimension. The
restriction of minus the Killing form to to is positive definite and determines
a G(A)-invariant inner product on Te(G/G(A)). Thus the homogeneous space
GIG(A) is a Riemannian manifold, with a G-invariant metric which coincides
7.5. The Fourier Transform of Coadjoint Orbits 229

with -B I,, at the point e. Denote by dg the corresponding G-invariant mea-

sure on G/G(A). This measure depends only on the subgroup G(a), which
varies over a finite number of subgroups of G.
Lemma 7.23. The Liouville measure d,Qa of Ma is given by the formula
2a)
a=
d ,3.x
2ir JGIG(a)
0(9a) d9
Ma «EP,
Proof. We can choose X±a such that {ea, fa} is an orthonormal basis of r[8,,,],
where
Xa - X-,,, iXa + ix-":
e., = (Ha,Ha)1/2, and f« =
B B(Ha, H«)1/2
Since 0.\(f,, - A, ea -,\) = we easily obtain the lemma.
The Fourier transform of the measure d,QA is a G-invariant analytic function
F,\ (X) on g, which is clearly determined by its restriction to t.
Theorem 7.24 (Harish-Chandra). Given A E t*, let Wa = {w E W
wA = Al be the stabilizer of A in the Weyl group W. For X E t, X regular,
the Fourier transform Fa (X) = fMA eif (X) d/ja (f) is given by the formula
ea(wa X)

wEWIWa
E IaEPa
Proof. Let X be a regular element of t. In this case the zero set of the vector
field generated by the action of exp tX on g* is the subspace t* of g* which
is fixed by the whole of T. Thus the zero set of the vector field XM, is the
finite set
Mant*={wAIwEW/Wa}.
Applying Theorem7.11, we see that it suffices to compute det1/2(Gwa(X)).
The element 11aE pA fa A ea determines the orientation of to T5MA. Since
,C,\(X)f. = ia(X)ea, we find that det1,2(Ga(X)) = H Epa(ia(X)).

If A is regular, PA is "half" of 0, and up to a sign, the denominator

detw/a (G(X)) _ (WC,, iX)
aEPi
is the same at each fixed point. Thus we obtain the following corollary.
Corollary 7.25 (Harish-Chandra). If Ma is a regular orbit of the coad-
joint representation, then for X E t, X regular,
1 ( ,w)eti(wa,JC)aEPA
FM,,(X) _ fl (a,X)_

wEW l

When \ is regular, we can compute the volume vol(MA) of the symplectic

manifold (M, Q) as the limit when X tends to 0 of FM,, (X).
230 7. Equivariant Differential Forms

Proposition 7.26. Let A be regular and A = iA. Then

f.>0 (a, A)
vol(Ma) =
iia>o(a' P)
Proof. We will also write p for the element of c dual to p c t*, which is half
the sum of the roots in the positive system Pa C A. The volume of MA is the
limit when t tends to 0 of FM,, (tp). Since
t-n/2 [: E(w)es(w\,tP) = t-n/2 E
wjEEW l wEW
= t-n/2 H (et(A,a)/2 - e-t(°,a)/2)
a>0

tends to fa>0 (a, A), when t -* 0, we obtain the proposition.

In particular, for is = p the symplectic volume is 1. Using Lemma 7.23,
which relates the symplectic volume and the Riemannian volume, we obtain
the following result.
Corollary 7.27. The Riemannian volume of the flag manifold G/T equals
[a >o 2ir(a, p)-1.
Using the above results, we will now analyse the structure of the algebra of
G-invariant functions on g. If j : t --> g is the inclusion map, then we obtain
a restriction map
7* : C°°(g)G --, C°O(t)W.
There are similar maps where the space of functions C°° (g) is replaced by
analytic functions C41(g) or polynomials C[g]. We will prove that the restric-
tion map is an isomorphism in all of these cases, by constructing an explicit
inverse for it.
Choose a system P C A of positive roots, and define
(A, ice)
fJ
aEP (p' a)
E C[t*].

Let us denote by 8,, the constant coefficient differential operator on t, given

by

11(p, a)-1 fl aoH.,

aEP aEP
where aft. is differentiation in the direction of the vector iHa.
Let pr be the projection from g = t®t to t. Let 1rs/t be the real polynomial
on t
7re/t(X) = fJ ia(X) E C[t].
aEP
All integrals over G are with respect to the Haar measure of volume one.
7.5. The Fourier Transform of Coadjoint Orbits 231

Theorem 7.28 (Chevalley). If 0 E C°° (t), let c(o) be the function on g

defined by the formula

No)) (X) = 7W f,))(()) dg.

Then c is an isomorphism of C°°(t)W, C 1(t)w and C[t*]w to C°°(g)G,
C"(g)G and C[g]G, and is the inverse of the restriction map j*.
Proof. The theorem follows from the following formula: if X is a regular
element of t, then

(7.10) O(wX)
(8wO)(Pr(gX)) dg =
fG wrw 7,9/t(wX)
To see that j*c(O) = 0 if 0 is invariant under the Weyl group, we apply this
formula to the function 7rg/t4'
It suffices to prove (7.10) for q5(X) = ezl-',X>, A E t* regular. Indeed, this
proves it by continuity for all A, hence for all 0 E C°°(t), and finally, since G
is compact, for all 0 E C°°(t).
For the function q5(X) = ei(a,X), (8.0)(X) = w(iA)e$(a,X) and the left-
hand side of (7.10) becomes
iIPIw(A) f e'(a,9X> dg.
c
If A is regular, this is equal to FM,, (X) up to a constant independent of X. To
evaluate this constant, we compare both sides with X = 0; Proposition 7.26
shows that
f(a)(Pr(gX)) dg = i1P1FMa(X)
Applying Corollary 7.25, we obtain (7.10). 0
When G is a non-compact Lie group and M is a coadjoint orbit of G,
the Fourier transform FM (X) = fm eif (X) d13 may often still be defined as
a generalized function on g. If K is a compact subgroup of G, the action
of K on M is the Hamiltonian action of a compact Lie group, and so, by
Proposition 7.10, for X E t the form (X)d)3 is exact outside the zeroes of
XM. Thus we can hope that when FM admits a restriction-as a generalized
function on the Lie algebra f of K, the localization formula will allow us to
compute this restriction FMI t. We state here a result of this type, where
G is a real semisimple Lie group with maximal compact subgroup K. This
result can be proved using Stokes's theorem and estimates at infinity on the
noncompact manifold M.
Let G be a connected real semisimple Lie group with Lie algebra g, and
let K be a maximal compact subgroup of G with Lie algebra t. Let g =1= ® p
be the Cartan decomposition of g, and let T be a maximal torus of K with
Lie algebra t. If the Cartan subalgebra tc of tc is also a Cartan subalgebra
232 7. Equivariant Differential Forms

of gc, we say that G and K have equal rank. We have then g = t ® r, where
r = [t, g] and we identify t* with a subset of g*.
Let L = A(gc, te), and 0,, = {a I (gc),,,, c pc}; a root belonging to it,,, is
called noncompact. For A E t*, let Pa = {a E A I (A, iHa) > 0}, and let nA
be the number of non-compact roots contained in PA. If
f,.={HEtI (a,H) 0forallaEA}
is the set of regular elements in t, then G t, is an open set of g. The function
FM (X) is a generalized function on g and is analytic on G t,.. The restriction
of FM to G t, is thus determined by its restriction to t,.. When M is of
maximal dimension, the following formula is due to Rossmann.
Theorem 7.29 (Rossman). Let M be a closed orbit of the coadjoint rep-
resentation of a real semisimple Lie group G such that G and K have equal
rank. Let W = W(tc, fc) be the compact Weyl group. Then for X E t,., we
have the following results:
(1) IfMnt*=0, then Fm(X) = 0.
(2) If M = G A with A E t*, and Wa is the subgroup of W stabilizing A,
ei(wa,X )
FM(X) = (_i)n(A) E
[LwEW/Wa iL EPa (wa, X)'

7.6. Equivariant Cohomology and Families

In this section, we will define the equivariant Chern-Weil homomorphism,
which maps the equivariant de Rham complex of a G-manifold M to the or-
dinary de Rham complex of a fibre bundle with fibre M and structure group
G, which we do not need to assume is compact. This is an extension of the
theory of characteristic classes in Section 1.5, although here we only consider
the case of connections, and not of superconnections.
Let P --+ B be a principal bundle with structure group G, with connection
one-form w E A' (P, g) G and curvature 0 = dw + [w, w]. The decomposition
TP = HP$VP of the tangent bundle TP into a horizontal2 subbundle HP =
ker(w) and a vertical subbundle VP determines a projection operator h from
A(P) onto the subalgebra of horizontal forms
A(P)hor={aI t(X)a=0forallXE9}.
With respect to a basis Xi, 1 < i < m, of g, we may write
m m
w% andSZ= 0iXi,
i=1 i=1
where wi are one-forms on P, and S21 are horizontal two-forms.
7.6. Equivariant Cohomology and Families 233

Lemma 7.30. The projection h onto the algebra A(P)hor of horizontal forms
is given by the explicit formula
m
h = J1 (I - wit(Xi))
i=1
(-1)r(r+l)/2Wi1 ...Wi,L(Xil) ... L(Xi,.).
1<il <... <i,<m
Proof. If pi = I - wit(Xi), the relations
[w2, wi] = [L(Xi), L(Xj)] = 0 and [t(Xi), wi] = aij,
imply that L(Xi)pi = [pi, pj] = 0 and p% = pi. It is also clear that pi is
an algebra homomorphism and that (lliR 1 pi)wk = 0. Thus f z"_' 1 pi is the
projection h on the space of horizontal forms.
Let D denote the operator h d h on A(P); under the identification of
A(P)bas with A(B), the restriction of D to A(P)bas corresponds to d, since
h acts as the identity on A(P)hor and d preserves A(P)bas-
Lemma 7.31. (1) D = h - (d - E', 1lit(Xi)), and
(2) D2 + h - (E-1 fit(Xi)) = 0
Proof. Since d is a derivation and hwi = 0, hdwi = SZi, we have for a E A(P),
m
h d h(a) = h d (a - wit(Xi)a)
i=1
M
= h(da - S2iL(Xi)a).
i=1
To calculate D2, we observe that

m
1it(Xi))
i=1

= h (d - Q't.(Xi))
2

ti=1
M m
= h (- do 2L(Xi) - n',C(Xi))
i=1 i=1
since h(dW) = 0.

The following result shows how the covariant derivative Vv on an associ-

ated vector bundle V = P X G V is related to the horizontal projection.
234 7. Equivariant Differential Forms

Proposition 7.32. The covariant derivative Vv on V coincides with the

restriction of the operator D ®I on A(P, V) to A(B, V) = A(P, V)b,,, .

Proof. If a c A(P, V)bas, we have t(X)da = (G(X) ® I)a = -p(X)a. Hence

t(Xi)t(XX)da = 0 and

h(dha) = h(da) = da + Ew'p(XX)a = V''a.

If M is a G-manifold, let M = P xG M be the associated fibre bundle,
with base B and typical fibre M. Since M is the quotient of P x M by a free
action of G, we may identify A(M) with the space A(P x M)bas of forms a
on P x M which are basic with respect to the action of G. The form w is a
connection form for the action of G on P x M. Thus, if we write simply t(X)
instead of t(Xp,M) = t(Xp) + t(XM), the projection h onto the algebra of
G-horizontal forms A(P x M)hor is given by the formula
m
h = fl(I - wit(Xi))
i=1

and the operator D = h . d - h restricts to dM on the space A(M) = A(P x

M)bas
Consider the differential graded algebra (C (g] (9 A(M), d9) of A(M)-valued
polynomial functions on g. For a = f ®/j E C[g] 0.4(M), we define a(1) E
A(P) ® A(M) by a(St) = f (St) ®p.
Definition 7.33. The map 0.: C[g] ®.A(M) -+ A(P x M)hor defined by
0w(a) = h(a(g))
is called the Chern-Weil homomorphism.
Observe that the restriction of cb to AGM = (C[g] 0 A(M))G sends
AG(M) into A(M) A(P X M)bas.

Theorem 7.34. Let G be a Lie group and let P -+ B be a principal bun-

dle with structure group G and connection form w. Let M be a manifold
with smooth action of G. Then the Chern-Weil homomorphism induces a
homomorphism of differential graded algebras

Ou, : (AG(M), de) ((C[g] (9 A(M))G, ds) - (A(M), d)

Proof. Since 0,, clearly preserves products, we have only to show that it
intertwines the differential de on AG(M) and the differential d on A(M). In
fact, we will prove that the map 0, : C[g] ®A(M) --p (A(P x M))hor satisfies
the formula
D.q5 =g4,de.
7.6. Equivariant Cohomology and Families 235

If f ® a E C [g] 0 .A(M), we see by Lemma 7.31 that

DO, ,,(f ® a) = h (d(f (@) (&a) - ®a))

i=1

= h (f (S2) ® (da - > Slit(XE)a))

since h(dQ') = 0 and t(Xi)S23 = 0,

= cb . dg(f (&a).

Note that when M is a point, the above Chern-Weil homomorphism be-

comes a map
0W : C[g]G ---> .A(B),

which is just the ordinary Chern-Weil homomorphism. In the general case, the
map q5, still has a geometric interpretation. The connection on P determines
a horizontal subbundle HM of the tangent bundle TM, where HM is defined
as the image of HP under the projection P x M -> M. We have TM =
HM ® VM, where VM is the vertical tangent bundle. Thus, if (p, m) E
P x M projects to y = [p, m] E M, we have the isomorphism

.7(p,m) : HHM ® T,,6M L--- TV M,

which induces an isomorphism between AHPP 0 AT,- M and AT.-M.

If f ®p E C[g] ® .A(M), then substituting S2 for X E g in f, we obtain
a horizontal form f (Q) p E AHP P, while /3,,,, E AT,nM may be considered
as a vertical form on P x M. If we define (f (9 Q) (it) to be f (12) ® 0, then
by linearity we can extend this to define a(n) for any a E C[g] ® A(M). It
is easy to see that if a is G-invariant, then j(p,m,)(a(S2)(p,,,L)) depends only
on y E M. Thus, we obtain a differential form on M which corresponds to
h(a(S2)), since h(a())) is a basic form on P x M whose restriction to the
subspace AHP ® ATM coincides with that of a(S2).
The Chern-Weil homomorphism has the following functorial properties. A
morphism f : N --> M of G-manifolds gives rise to a G-invariant map of
fibre bundles f p: N= P x G N -+ M = P X G M, and to the commutative
diagram
Ag(M) .A.(M)

fw1
(N)
f;1
Ae A(N)
236 7. Equivariant Differential Forms

Proposition 7.35. Let G be a Lie group, and let M be a compact oriented

manifold with smooth action of G. We obtain the commutative diagram

A,, (M)
'- ) A(M)

m-) A(B)
C[9]G
where the left vertical arrow is the integration over M and the right vertical
arrow is the integration over the fibres of M -* B.
We will now relate the equivariant curvature with the ordinary curvature.
Let E = E+ ® E- be a G-equivariant superbundle on M, with invariant
superconnection A. Let FE = A2 E A(M, End(E)) be the curvature of A and
let µE be its moment. Consider the associated family of vector bundles
£=PX E->M=PXGM.
Then A(M, £) may be identified with the space of basic forms A(P x M, E)bas,
where we consider E as a vector bundle on P x M. The operator
m
hE
i=1
is a projection from A(P x M, E) onto A(P x M, E)hor. Consider the operator
=hE.(dp(9
De
on A(P x M, E), where d p ®1 + 1®A is the pull-back of the superconnection A
on P x M. Clearly De preserves A(M, £) and induces on it a superconnection.
Definition 7.36. The superconnection A£ on A(M, £) is the restriction of
the operator De to A(P x M, £)bas.
Let us compute the curvature F1 = (AE)2 of the superconnection A6. The
operator DE on A(P x M, £) satisfies
Ds (as) = (Da)s + (-1)kkI h(a)DEs
forallaEA(PxM)andsEA(PxM,£),where By the
formula D2 + h(Ez'__ 1 SZV.C(Xi)) = 0 of Lemma 7.31, we see that the operator
m
(DE) 2
+ ho Q,C(Xi)
i=1
commutes with exterior multiplication by any horizontal differential form, and
hence equals hE F. We will now compute Fs as a function of the curvature
FE of A and the moment µE.
Lemma 7.37. (1) DE = hE (dp + A - EZ"' 1 SZia(Xi))
7.6. Equivariant Cohomology and Families 237

(2) The curvature Fe equals FE + SliµE(Xx), and

m
(De) 2 + hE . (> cli'L(Xi)) = hE FE.
i=1
Proof. The proof of this lemma is almost the same as that of Lemma 7.31,
except that the operator d = dp + dM on A(P x M) is replaced by dp + A
on A(P x M, E). Since A is a superconnection and h(wi) = 0,
De = hE (dp +A) hE
= hE (dp + A - >b ` 1 Wi,(Xi)) .
From this, it follows that
(D-0 )2 = hE (dp + A - Em1 n't(Xi.))2.
Using the relation
[dp + A, t(Xi)] = LE(Xi) - I-iE(Xi) + L((Xi)p),
we see that
[dp + A, lit(Xi)] = dfli t(Xi) + l t (Xi) - ci, E(Xi)
It follows that
m m m
(dp+A-Em1 Sl''t(Xi))2 = FE+> S2Z LE(Xi)-> dir t(Xi)-E !QzL(Xi),
i=1 i=1 i=1

and hence that

m m
(DE)2
+ hE ( S2iL(Xi)) = hE -(FE AE(Xi))
i=1 i=1
since h(dS2i) = 0.
Recall that the equivariant superconnection A. is the operator on C[g]
..4(M, E) defined by
(Aea)(X) = A(a(X)) - t(X)(a(X))
and that the operator A9(X)2 + L(X) is given by exterior multiplication by
the equivariant curvature F. (X) = FE + µE(X).
Theorem 7.38. Let G be a Lie group and let P --- B be a principal bundle
with structure group G and connection form w. Let M be a manifold with
smooth action of G, and let E -+ M be a G-equivariant bundle on M with
G-invariant superconnection A.
(1) The map
0,,, : C[g] ® A(M, E) --f A(P x M, E)hor
given by 0,,,(a) = hE(a(f2)) satisfies the formula 0, Ag = DE ,.
238 7. Equivariant Differential Forms

(2) If we also denote by 0w the map

0w : C[g] ®A(M, End(E)) -> A(P x M, End(E))hor,
then 0w(Fg) = FE.
(3) Let the (A) be the equivariant Chern character of the G-equivariant super-
bundle E - M, with G-invariant superconnection A, and let ch(AE) =
Str(e-(A,6)2
) be the Chern character form of the associated family of su-
perbundles £ -> M with superconnection AE. Then the image of chg(A)
by the Chern-Weil homomorphism is the Chern character form of AE
0, (che (A)) = ch(AE).

Proof. The proof of (1) follows the proof in Theorem 7.34 that V 0" = Y'w d,
except that the operator d = dp + dM on A(P x M) is replaced by dp + A.
To prove (2), we must show that
m
(L (X)) _ nZG(Xi) 0.
i=1

Let fi E g* be the dual basis to the basis Xi of g. If f c C[g] and a E A(M),

we have

0. AX) if ®a)) = 0w
i=1
fif ®G(Xi)a)
m
l
=h nif (Q) ®G(Xi)a)
M
= hc C(Xi)(f(0) ®a)),
i=1

since EE"1 n',C(Xi) f (52) = 0.

Part (3) now follows immediately from (2).
Let us see what these formulas become in the case of a trivial principal
bundle P = B x G. A connection one-form on P is just a g-valued one-form
on B,
m
w= ixi.
i=1

Let M be a manifold with G action, and E be a G-equivariant vector bundle

over M with invariant connection V and moment uE E g* ® F(M, End(E)).
The manifold M = P xG M is the direct product B x M, and the bundle
£ = P X G E is the pullback of E by the projection of B x M onto M. Using the
connection form w, we obtain a connection DE on £. The following formula
is easily shown.
7.7. The Bott Class 239

Proposition 7.39. Let w µE _ Em?=l wi p (XX) be the contraction of w with

AE. Then V

7.7. The Bott Class

Let V --3 M be an oriented Euclidean vector bundle of even rank with spin
structure, let S -- M be the corresponding spin superbundle, and Sv -* V be
the pull-back of S to V. In this section, we will describe the "Riemann-Roch"
formula of Mathai and Quillen relating the Chern character of a supercon-
nection on Sv and of the Thom class U(V) of the bundle V. This result
is an application of the functorial properties of the equivariant Chern-Weil
differential forms proved in the last section.
Consider the case where M is a point. Let V be an oriented Euclidean
vector space of dimension n = V. Let G = Spin(V), with Lie algebra g =
A2V C C(V), and let r : g -> so(V) be the action of g on V defined in (3.4).
Let S be the spinor space of V and let p be the representation of G in S.
The trivial bundle Sv = V x S is a G-equivariant vector bundle with action
9(x, S) = (9' x, P(9)S)
Let c : C(V) - End(S) be the spin representation of the Clifford algebra
of V, and consider the odd endomorphism c(x) E r(V, End- (Sr)), where x E
r(V, v x v) is the tautological section of the trivial bundle with fibre V. From
the section c(x) = Ek XkCk, we can construct a G-invariant superconnection
A = d + ic(x) on Sv -+ V with curvature
FV = IIxII2 + idc(x) = IIxII2 + i > dxkck E A(V, End(S)) A(V, C(V)).
k

The vector bundle SV -+ V, with superconnection d + ic(x), is called the

Bott class; by Bott periodicity, it represents an element of the K-theory of
V which generates K(V) as a free module, but we will not discuss this point
of view.
If X = Ei<j Xijei Aej is an element of g, the vector field on V defined by
the action of G on V is
XV = -2EXij(xi(9j - xjai)
i<j
If X E g, the moment µ(X) = G(X) - V x of X is equal to c(X), hence the
equivariant curvature of the superconnection is the End(S)-valued form on V
FV (X) = IIxII2 + idc(x) + c(X) = E xk + i dxkck + E Xklckcl.
k k k<i

In this context, the Bianchi identity [A - t(X), Fg(X)] = 0 becomes

(7.11) (d - c(XV))FV(X) + i[c(x), FV(X)] = 0,
which is also easy to check directly.
240 7. Equivariant Differential Forms

The equivariant Chern character Av of the trivial bundle Sv -- V with

superconnection A = d + ic(x) is the equivariantly closed form on V given by
the formula
Av(X) = Str(exp(-Fv(X)))
= Str(e-(11X112+idc(X)+c(X)))

= Str(e-(11X112-zc(dX)+c(X))).

This differential form decays rapidly at infinity. Such a behaviour would

not have been possible if we had taken instead the Chern character of a
connection. This illustrates one of the applications of superconnections: the
construction of Chern characters which decrease rapidly at infinity on non-
compacts manifolds.
Consider now an oriented Euclidean vector bundle V -+ M of even rank
n = 22 over a compact manifold M, and assume that V has a spin structure.
Thus V is associated to a principal bundle P --* M with structure group G =
Spin(V), in other words, V = P xGV. Let w be a connection form on P with
curvature form 0. The connection w defines a Chern-Weil homomorphism
:AG(V)- A(V).
Let S = P x G S be the corresponding spinor bundle over M, and let Sv =
P xG Sv be a bundle over V defined using the trivial bundle Sv = V x S over
V; Sv is the pull-back of the bundle S - * M to V by the projection V -+ M.
The bundle map c(x) : Sv -+ Sv defines a bundle map c(x) : SV -p Sv.
Consider the G-invariant superconnection A = d + ic(x) on Sv -> V. As in
the preceding section, the connection w allows us to lift the superconnection A
to a superconnection ASV on Sv -> V. (Beware that the base of our principal
bundle P is M, and the fibre of the associated bundle is V, whereas in the
last section, the base was B and the fibre was M.) If we denote by Vu, the
connection induced on the pull-back Sv - i V by the connection on S --> M
with connection form w, then
ASV = V4, + ic(x).
If dx E Al (V, V) is the canonical one-form on V, Lemma 7.30 shows that
0,,,(dx)
Thus the Chern-Weil homomorphism 0,,, : C[g] ® A(V) -> A(P' x V)hor is
obtained by substituting SI for X E g, and the one-form dx + w - x on P x V
for the one-form dx on V. It follows that
(e-(11X112-ic(dX+W - X)+c(fl))
&u, (Av) = Str
The following result follows from the functorial properties of the Chern-Weil
homomorphism.
Proposition 7.40. The closed form 0,, (AV) E A(V) is the Chern character
form ch(ASv) of the bundle Sv -+ V with superconnection ASv.
7.7. The Bott Class 241

The differential form 0,, (AV) has the important property of decaying rapidly
in the fibre directions.
In Section 1.6, we defined the Thom form of an oriented Euclidean vec-
tor bundle V --f M. This may be obtained by a construction analogous to
that used above; we will restrict attention to the case of even-dimensional
V, although a similar construction works in the odd-dimensional case. The
equivariant Thom form Uv of the vector space V may be defined by a formula
analogous to that of AV, except that we work within the Grassman algebra
of V instead of its Clifford algebra. The analogy between the Clifford alge-
bra and the Grassmann algebra leads us replace the C(V)-valued equivariant
curvature Fv of Sv by the AV-valued form fv on V, given by the formula
fv(X) =IIxII2+idx+X =I: x +i1: dxkek+1: Xklek A el.
k k k<1

This is the analogue of the differential form which we denoted S2 in Section 1.6.
If e E V, we denote by LA(e) the derivation of A(V, AV) defined by
1,V (a (-1)Iala forc
and by tA(X) the operator

tA(x) = Exktn(ek)
k=1
It is easy to verify that for every X E g,
(d - t(Xv))fv(X) - 2itA(x)fv(X) = 0.
This formula is the analogue of Part (1) of Proposition 1.51.
Let exp(-fv(X)) be the exponential of fv(X) in the algebra A(V, AV).
Since the operators d, t(Xv) and tA(x) are derivations, we see that
(7.12) (d - tA(XV) - 2itA(x)) exp(-fv(X)) = 0.
The equivariant Thom form UV on V is defined using the Berezin integral
T : A(V, AV) -+ A(V) in place of the supertrace Str : A(V, C(V)) -> A(V):
Uv(X) = T(exp(-fv(X))) = T(e-(IIxII2+idx+x))
More explicitly, Uv (X) is given by a sum over multi-indices I C {1, ... , n}
of the form
Uv(X) = e-IIxII2 1: PI,(X)dxi
IZI even
where PI (X) are homogeneous polynomials of degree (n - III) /2 in X, I' is
the complement of I, and P1, (X) coincides up to a sign with the Pfaffian of
XIl = Xjkej A ek E A2RI .
{(j,k)EI' Xr'Ij<k}
242 7. Equivariant Differential Forms

In particular,

e-1
Uv(X) = Pf(-X)e-11X112 +' ai + e-Ilxll2dxl A ... A dxn,
i=1

where ai E A2i(V) ®(C[g]e-i

Theorem 7.41. (1) The form UV is an equivariantly closed form on V, and
for all X E g, we have

7r_e Uv(X)=1.
v
(2) If i : {0} --> V is the injection of the origin into V,
i.(UV(X)) = Pf(-X).
Proof. It is clear that T(LA(x)a) = 0 for every a E A(V,AV). Applying the
Berezin integral to exp(-fv(X)), we see from (7.12) that
(d - t(Xv))T (exp(-fv(X))) = 0.
The other properties of Uv (X) are obvious.
Definition 7.42. Define U(V) E A(V) by the formulas

U(V) = 2e0w (Uv) = T (e (11x112+i(dx+w x)+SZ)) .

The differential form U(V) is the Thom form of Section 1.6, for the metric
obtain another proof that U(V) is closed, from the properties of
IIXII2/2. We

the Chern-Weil map qw. We also see that

(27r)-e U(V) = 1.
JVIM
Recall the function jv (X ), which we have already met a number of times
in this book:
sinh(r(X)/2)
jv(X) = det
{ (X)/2
By definition,
S1
0w UV 2) = det1/2 I Q/2/2) = A(V)-1.

The following proposition is a generalization of Proposition 3.16.

Proposition 7.43. The equivariant differential forms AV and Uv are re-
lated by the formula
(-2i)-'AV(X) = jV 2(X) UV(X).
7.7. The Bott Class 243

Proof. Choose an oriented orthonormal basis ei of V such that

t
X = E A e2i-1 A e2i
i=1
From this, we see that it suffices to consider the case in which dim(V) = 2
andX=Aeine2. Then
-FF(X) = -IIxII2 - idxici - idx2c2 - Ac1c2
= -IIxII2 - A(ci + i.\-idx2)(c2 - is-idxi) - a-'dxidx2.
If we let Si = c1 + i)-1dx2 and 2 = C2 - iA-idxl, then i = 2 = -1 and
S1 2 + 2 1 = 0, so that e-a£le2 = cos A -1;1e2 sin A. Thus, we see that
e-Fv(X) = e-11X112 (cos. - sinAcic2 - isinX (dxici + dx2c2)
\ +
sin A - A cos A sin A
A2 dxidx2 + dxidx2cic2).

Since Str(cic2) = -2i, we obtain

Str(e-Fv(X))
= 2i sin A e 11X112 (1- A-idxidx2).
On the other hand,
-fv(X) = -IIxII2 - idxlei - idx2e2 - Ae1 A e2
and hence

e-fv(X) = e-IIX112 (1 - idxlei - idx2e2 - Aei A e2 + dxidx2e1 A e2).

It follows that T (e-fv (X)) = -A e-IIxII2 (1 - ),-idxidx2).

The matrix T(X) has eigenvalues ±2i/\, so that jV 2 (X) A-1 sin A. Thus,
it follows that
sin p
Str (e-F' (X)) _ -2i T (e-fv (X))

Thus, we obtain the following result, which is a refinement at the level of

differential forms of the well known "Riemann-Roch" relation between the
Thom classes in K-theory and in cohomology. In cohomology, this formula is
due to Atiyah-Hirzebruch [10].
Theorem 7.44. Let V -> M be an even-dimensional oriented Euclidean
vector bundle with spin structure over a manifold M. The Chern character
ch(A'") of the Thom bundle Sv --+ V with superconnection Asv and the
Thom form U(V) are related by the formula
i' ch(A's") = A(V)-lU(V).
244 7. Equivariant Differential Forms

Bibliographic Notes
Section 1. The complex of equivariant differential forms was introduced in
H. Cartan [45]. He shows that if G is compact, the cohomology HG (M) of
this complex agrees with the topological definition of equivariant cohomology
H' (M x G BG, R). The map between the two theories may be described using
the Chern-Weil map of Section 6, with base the universal classifying space BG
(or a finite-dimensional approximation).
The definitions of the moment of an equivariant connection, the equivari-
ant curvature and equivariant characteristic classes are taken from Berline
and Vergne [23], except that here we allow superconnections instead of just
connections.
Sections 2, 3 and 4. Theorem 7.13 is due to Berline-Vergne [21] when X = Y,
and to Bismut [31] as stated. The original proof used an argument due to
Bott, of approximating MO by a small tubular neighbourhood, similar to
that which we give here for the proof of Theorem 7.11. Here, we use an
argument by Gaussian approximation; it is interesting to observe its close
analogy to the proof of the McKean-Singer formula. Special cases which were
proved earlier include Bott's formulas for characteristic numbers, when the
zero-set of the vector field is discrete [42], corresponding formulas when the
zero-set may have arbitrary codimension proved in the Riemannian case by
Baum and Cheeger [18] and in the holomorphic case by Bott [43], Duistermaat
and Heckman's theorem on the exactness of stationary phase for functions
whose Hamiltonian vector field generates a circle action Corollary [57], and
the localization theorem for equivariant differential forms, proved by Berline
and Vergne [21], and by Atiyah and Bott [7]. A generalization of Theorem 7.11
to algebraic manifolds with singularities has been given by Rossman [98].
A localization theorem for equivariant K-theory was proved by Segal [101],
and by Quillen for equivariant generalized cohomology theories [91]. The
similarity between the localization properties of the complex of equivariant
differential forms and the localization theorem in K-theory were observed by
Berline and Vergne [21] and Witten [105].
Witten has proposed an interesting analogy between the local index the-
orem for the Dirac operator on a manifold M and the localization theorem
in equivariant cohomology for the action of the circle group on on the loop
space LM (see Atiyah [4]). Bismut generalizes their results to twisted Dirac
operators in [31].
Section 5. The homogeneous Hamiltonian actions of Section 5 were introduced
by Kostant [75]. A general relationship between the Fourier transform of a
coadjoint orbit and the character of the corresponding representation was
conjectured by Kirillov [72], [73], based on his work on his formulas for the
characters of compact and nilpotent groups. Further conjectures on this for-
mula are stated in Vergne [102].
Theorem 7.24 is proved in Harish-Chandra [68].
Additional Remarks 245

Further results on the relationship between equivariant cohomology and

Fourier transforms of orbits may be found in Berline and Vergne [22], Duflo-
Heckman-Vergne [54], Duflo-Vergne [55]. The proof we give of Chevalley's
Theorem 7.28 is in the last of these references. Theorem 7.29 is due to Ross-
man [97].
Section 7. The results of this section are due to Mathai and Quillen [81], al-
though we have made greater use of the Berezin integral in the exposition
than they did. As explained in Atiyah-Singer [15], it allows one to pass from
the Atiyah-Singer Theorem for the index of an elliptic operator, expressed in
terms of K-theory, to formulas in terms of characteristic classes.

Additional Remarks
Let M be a compact manifold on which a compact group G acts differentiably.
If g E G, denote by G(g) = {h E G I gh = hg} the centralizer of g, by 9(g)
its Lie algebra, and by M(g) the fixed point set of g acting on M, considered
as a manifold with differentiable G(g)-action.
In Block-Getzler [37] and Duflo-Vergne [56], a de Rham model 1CG(M) for
the G-equivariant K-theory of a G-manifold M is introduced. (In the first
of these articles, the relationship with equivariant cyclic homology is also
described.) An element of 1CG(M) is represented by a collection of equi-
variant differential forms {w9(X)}gEG, called a bouquet, where wg(X) E
,AG(g)(M(g)) is a G(g)-equivariant differential form on M(g). Axioms of
invariance and compatibility for bouquets are given in [37] and [56].
By analogy with the case of equivariant differential forms, one can define an
integral for bouquets. If M has a G-invariant spin-structure, Proposition 6.14
induces a family of orientations on the fixed point sets M(g). Let R°°(G) =
C°°(G)G be the ring of smooth functions on G invariant under the adjoint
action of G. The integral fM 1CG(M) - R°O(G) is defined by the formula
e
wg (X) AB (g) (X, M(g))
f w(9 exP X) = JM(g) detl/2 (1 -
M

Here, A.(g)(X,M(g)) is the equivariant A-genus of M(g), and R9 (g) is the

equivariant curvature of the normal bundle N(g) of the embedding M(g) -+
M. The fact that these germs of smooth functions on G glue together to form
an element of R°° (G) is a corollary of the localization theorem 7.13 ([56]).
There is also a generalization of this formula to spines manifolds.
Chapter 8. The Kirillov Formula for the Equivariant
Index

The character of a finite-dimensional irreducible representation of a compact

Lie group G can be described by the Weyl character formula, which is a spe-
cial case of the fixed point formula for the equivariant index of Chapter 6.
However, there is another formula, the universal character formula of Kiril-
lov[72], which presents the character not as a sum over fixed points but as
an integral over a certain orbit of G in its coadjoint representation on g*;
this second formula is in principle much more general than the first, since it
applies to many cases other than that of compact groups.
Let M be a compact oriented even-dimensional Riemannian G-manifold,
where G is a compact Lie group, and let £ be a G-equivariant Clifford module
over M. In this chapter we will give an analogue of Kirillov's formula which
computes the equivariant index of a Dirac operator on the bundle £ over M
as the integral over M of an equivariantly closed differential form. Let g
be the Lie algebra of G, let AB (X, M), X E g, be the equivariant A-genus
of the tangent bundle of M, and let chs (X, £/S) be the equivariant relative
Chern character of 6 (see (8.2)). If D is a Dirac operator on 6 associated to a
G-invariant connection on £, we call the Kirillov formula the formula for the
equivariant index of D, which holds for X E g sufficiently small,

indG(e-X, D) = (27ri)-dim(M)/2 A. (x, m) che(X, £/S).

J
In Section 1, we will prove this formula by combining the fixed point formula
for the index with the localization formula of the last chapter. We discuss
the special case in which D is an invariant Dirac operator on a homogeneous
space of a compact Lie group in Section 2.
In Section 3, we introduce Bismut's "quantized equivariant differential",
namely, we replace the Dirac operator D by the operator D + 4c(XM), where
XM is the vector field on M corresponding to the Lie algebra element X E g.
Following Bismut, we rewrite the general Kirillov formula as a local index
theorem for this operator; the proof of this theorem follows closely the proof
of the local index theorem of Chapter 4.
8.1. The Kirillov Formula 247

8.1. The Kirillov Formula

In Chapter 6, we proved for the equivariant index of the Dirac operator
indG(y, D) the fixed point formula (where n = dim(M))

(8.1)
(2 r)dim(N)/2 A(M r) chG(y, E/S)
indG(y, D) = fM7 TM I Jdxoj,
(27ri)-/2 detl/2(1 _'YNexp(-RN))
which expresses the equivariant index as an integral over the fixed point set
Mry of -y in M. In this section, we will rewrite this for -y near the identity in
G, in terms of the equivariant cohomology of M; we call the resulting formula
the Kirillov formula, by analogy with Kirillov's formulas for characters of
Lie groups.
Let M be a compact oriented Riemannian manifold of even dimension n,
and let G be a Lie group, with Lie algebra g, acting on M by positively
oriented isometries. Let R be the Riemannian curvature of M, and let

µM E (g* ® r(M,so(M)))G C A2 (M,so(M))

be the Riemannian moment of X defined in (7.3), that is, the skew-endo-
morphism of TM given by
µM(X)Y = -VYX.
Let Re = AM +R E AG (M, so (M)) be the equivariant Riemannian curvature
of M. Let X AB (X, M) be the equivariant A-genus of the tangent bundle
of M,
Rg (X) /2
Ag (X, M) = detl/2
(sinh(R9(X)/2)
Let 6 be a G-equivariant Clifford module over M with G-invariant Her-
mitian metric, and invariant Clifford connection V£. Let £6(X) be the Lie
derivative of the action of X E g on 1'(M, 9), and let

µE (X) = £6 (X) - OX E r(M, End(£))

be the moment of X with respect to the connection Os; thus, µE is an element
of (g* ® r(M, End(E)))G C A2 (M, End(£)).
Recall from (3.7) the canonical map
T:A2T*M-->so (M);

if a E A2(M) and E A'(M), then as operators on E,

[c(a),c()I = c(T(a))
248 8. The Kirillov Formula for the Equivariant Index

If µM is the Riemannian moment of M, then c(7-1(1-iM)) is the element of

A2 (M, End(£)) given by the formula
c(T-1(j M(X)))
= 2 Y(/M(X)ei,ej)c(ez)c(ei)
i<j

where ei, (1 < i < n), is an orthonormal frame of T*M. Thus

[c(T-1('4M(X))),
c(a)] = c(AT*M(X)a)
for all a E A' (M), where µT {M is the moment of the bundle T*M, equal to
minus the adjoint of µM.
Define the twisting moment ,579(X) for the action of X on 6 by the
formula
/-1E/S(X) = /.ie(X) - c(T 1(AM(X)))
µels commutes with c(a) for all a c
Lemma 8.1. The twisting moment
Ai (M), and hence lies in (g* ®r(M, Endc(M) (£)))G C AG(M, EndC(M) (£)).

Proof. Since µs(X) = GS(X) - OX and ITM(X) _ ,CT*M(X) - VT*M, we

see, since V' is a G-invariant Clifford connection, that
[µ£(X),c(a)] = c(,CT+M(X)a) - c(VT'Ma)
= c(,T*M(X)a)
[c(T_1(pM(X)),c(a))
_
We now define the equivariant twisting curvature of £ to be
FB'S(X) = µs/S(X) +Fe/S E A2 (M,Endc(M)(£)).
Note that if M has an equivariant spin-structure with spinor bundle S, so that
£ is isomorphic to the equivariant twisted spinor bundle W ® S, then we may
identify Fg /S E AG(M, Endc(M) (£)) with FF E AG(M, End(W)) under the
isomorphism of bundles Endc(M)(6) = End(W). Define the equivariant
relative Chern character form of 6 to be
(8.2) chs (X, £/S) = Stress (exp(-Fe /S(X))).
In this chapter, we will study the following reformulation of the equivariant
index theorem near the identity in G. We will explain in the next section why
we call this the Kirillov formula.

Theorem 8.2 (Kirillov formula). For X E g sufficiently close to zero,

indG(e-X, D) = (2,7ri) -n/2 f A8 (X,M)the(X,£/S).

M
8.1. The Kirillov Formula 249

Let us first comment on some differences between the fixed point formula
(8.1) and the Kirillov formula for indc(y, D). In the first formula, it is not a
priori clear, and in fact quite remarkable that the right-hand side of the fixed
point formula depends analytically on -y near y = 1, and in particular that it
has a limit when -y --> 1 equal to
(27ri)_n12
ind(D) _ A (M) ch(£/S).
fM
In contrast the analytic behaviour of indG(e-X, D) near X = 0 is exhibited in
the Kirillov formula. However, the fact that this analytic function of X is the
restriction to a neighbourhood of zero of an analytic function of e -X is not
apparent. This fact is an analogue of the integrality property of the A-genus
of a spin manifold.
The Kirillov formula can be deduced from the localization formula, since
the integrals over M of an equivariantly closed form a(X) localize on the set
of zeroes of XM.
We now start the proof of Theorem 8.2. Choose X E g sufficiently close
to zero that the zero set of XM coincides with the fixed point set Mo of
-y = exp(-X). By Proposition 7.12, the bundle N is orientable, and hence so
is Mo. Fix compatible orientations on Mo and its normal bundle N. Let T,v
be the Berezin integral
TAr : r(MO, AN*) - C°°(Mo).
Since ye E P(Mo, C(N*) ® Endc(M) (E)) r(Mo, AN* ® Endc(M) (E)), we
see that
Tg (-ye) E r (Mo, Endc(M) (E)) .
The fixed point formula for the equivariant index states that
N)/2 A(Mo) chG('y, E/S)
indG(y, D) = (2iri) -n/2 f
Mo detl/2(1 -'yNexp(-RN))
where RN is the curvature of the normal bundle N -> Mo, yN is the induced
action of y on N, and chG(y, E/S) is the differential form on Mo given by the
formula
2dim(N)/2
chG('Y,E/S) = Stre/s(TN('ye).exp(-Fols))..
detl/2(1 - yN)
By the localization theorem, we have
(Ag(X, M) chg(X, E/S)) IMo
A(X, M) ch(X, E/S) = (27r)dim(N)/2
fM JMo 1/2
detN (-R.Ar(X))
Thus, we need only prove that, for -y = e-X with X sufficiently small,
A(Mo) chG(y, E/S) (Ag(X, M) chg(X, E/S)) 1 Mo
det1/2 (1 - yAexp(-RN)) det1/2(-Re (X))
250 8. The Kirillov Formula for the Equivariant Index

Observe that the differential forms detN2(-RA( (X)) and chG(y, £/S) both
depend on the orientation of N. We will choose an orientation such that
detN2(.N(X)) > 0. Under the splitting
TMIM0 =TMo®N,
the Riemannian curvature restricted to Mo splits as R = Ro ® RN, where
Ro is the Riemannian curvature of Mo. The Riemannian moment /M(X) of
X restricted to Mo is the infinitesimal action of X on TM, and equals the
moment p/ (X). Thus,

Rg(X) = I 0o Re(X))

It follows that
Rg(X)/2
A,; (X, M) = det1/2
(sinh(Rg(X)/2)
2 RN(X)/2
= detl/2 (sinhh(oRo/2)) detl/2 (sinh(Rg(X)/2) )

The exponential of Rg (X) equals (y 1)-1 exp(RN) where -y = e-X, since

µN (X) may be identified with the action of the vector field X on the bundle N,
and i1 (X) and RN commute. Since Rg E AG (MO, so (N)) is antisymmetric
and N is even-dimensional, we see that
det-1/2(sinh(Rg (X)/2)) = 2d'm(N)/2 det-1/2(1- yNexp(-RN)),
and hence
Ag(X, M) A(Mo)
detAr (Rg (X)) detl/2 (1 - yN exp(-RN))
It remains to show that
chg (X, E/S) I Mo = (-1)dim(N)/2
chG(e-X
, £/S),
Consider the decomposition

A' (X) = c(T-1(AM(X))) +µe/S(X),

where r-1(AM(X)) E A2N*. Since the two terms on the right-hand side
commute, we see on taking the exponential of both sides that
ye = exp(-c(r-1(µ`(X)))) 'exp(-µe/S(X))

Let us write a for exp(c(r-1(.M(X)))) E C(N*). With our choice of orien-

tation of N, TN(a) > 0, and
T,v (a) = 2-dim(N)/2 detl/2(1-'YN)
8.2. The Weyl and Kirillov Character Formulas 251

so that
2dim(N)/2
chg('Y,£/S) = Stre/S(TV(rye)exp(-Fo/S))
detl/2(1 - yN)
= (-l)dim(N)/2 Stre/S(exp(-pe/S(X)) eXp(-FO /S))
= ch.(X, £/S).
This completes the proof of Theorem 8.2.

8.2. The Weyl and Kirillov Character Formulas

Let G be a connected compact Lie group with maximal torus T. The irre-
ducible finite dimensional representations of G were classified by E. Cartan,
and their characters were calculated by H. Weyl. In the notation of Sec-
tion 7.5, let LT be the lattice {X E t I ex = 1}, so that T = tILT, and denote
by LT the dual lattice
LT = {l E t* 1(X) E 27rZ for every X E LT} C t*.
Choose a system P C A of positive roots. The subset XG = ipp + LT of t* is
independent of the choice of P and W-invariant. In the notation of Section 7.5,
we write Weyl's character formula as follows.
Theorem 8.3 (Weyl). If )) is a regular element of XG, there exists a unique
finite-dimensional irreducible representation TA, of 0, such that for X E t,

WEWe(w ) ei(wa,X)
Tr( Tx(e X )) =
11a,=PA (e(a,X)/2 - e-(a,X)/2)
Let us now describe Kirillov's formula for the character of the representa-
tion TA. As usual, for X E g, let
smh(ad X/2)
jg (X) = det
ad X/2
We can define an analytic square root of je (X) on the whole of g. Indeed, if
XEt,

j9 (X)
e(a,X)l2 - e-(a,X)12
\
e(a,X)/2 -
(/a' X/ 1\ (CY, X)
aEA aEP
Thus, by Chevalley's Theorem 7.28, there exists a unique analytic G-invariant
function je/2 on g, which on t is equal to
j1/2 (x) e(a,X)/2 - e-(a,X)l2
aEP (a'X)
252 8. The Kirillov Formula for the Equivariant Index

Theorem 8.4 (Kirillov). Let A be a regular element of XG and let MA, be

the symplectic manifold G A with its canonical symplectic structure Sta. For
X E g, we have
(X)
j1 ,2(X) Tr(TA(eX )) =
19 eif )
dm(ffMi

Proof. Both sides of the equation are G-invariant functions on g. Thus it

is sufficient to check this equality when X E t. The equality then follows
immediately from Corollary 7.25.

In this section, we will explain how these formulas are special cases of two
forms of the equivariant index theorem, respectively the fixed point formula
Theorem 6.16 and the Kirillov formula Theorem 8.2. To do this, we must
find an equivariant Dirac operator whose equivariant index is the character
of TA. As underlying space, we take the flag manifold G/T. Consider GIT
as a Riemannian manifold, with the metric induced by the restriction of the
opposite of the Killing form to r = Te(G/T). Choosing a system P C A of
positive roots, let
_ Xa - X_a _ iXa + ix-,
ea , .f«
IIH0II 11H011

Choose the orientation on GIT such that [aE p ea A fa is the orientation of

r = Te (G/T). The manifold GIT is an even-dimensional oriented Riemannian
manifold with spin-structure.

Lemma 8.5. Let Ag (X, G/T) be the equivariant A-genus of the Riemannian
manifold GIT. Then A.(X,G/T) and jg 12(X) represent the same class in
the equivariant deRham cohomology H,,(G/T).

Proof. Consider the direct sum decomposition g = t ® r of g. From this, we

obtain a direct sum decomposition of the bundle G XT g,
CXT9=(GXTt)®T(G/T).
Since T acts trivially on t, the bundle G xTt is isomorphic to the trivial bundle
GIT x t. The trivial connection on this bundle has vanishing equivariant
curvature, so the equivariant A-genus equals 1. Thus, the bundle G X T g,=with
connection the direct sum of the trivial connection on GIT x t and the Levi-
Civita connection on T (G/T ), has equivariant A-genus equal to A. (X, GIT).
However, since g is a G-module, the bundle G XT g may be trivialized to
G/T x g by the map
[(g, X)] E G XT g -+ E GIT x g.
8.2. The Weyl and Kirillov Character Formulas 253

The trivial connection on G/T x g has equivariant curvature F.(X) = ad X,

so that for the trivial connection,
1/2
A,(X,GXT9)=jg (X)
Since the equivariant cohomology class of the equivariant A-genus is indepen-
dent of the connection used in its definition, we obtain the lemma.
Of course, Ag (X, G/T) and jg 1/2 (X) are not in general equal as differential
forms, since jg 1/2(X) is a function whereas the differential form A(G/T) _
Ag (0, G/T) has non-vanishing higher degree terms in general.
If A E LT, there exists a unique character of T with differential iA on t.
Denote this character by e'' and by La the homogeneous line bundle G xT CA,
where C,\ is the one-dimensional vector space C with the representation e' A
of T. Let us denote by A' (X) the restriction to MA of the linear function
f H (f, X). If A is regular, the map g H g A is an isomorphism of G/T with
Ma = G C. A. Thus, La is a line bundle over Ma.
Proposition 8.6 (Kostant). The homogeneous line bundle La has a G-
invariant canonical Hermitian connection. The equivariant curvature of LA
is the equivariantly closed form i(p'(X)+fla).
Proof. For X E g, define an operator VX on sections of C,\ by the equation
L(X) _VX +iµa(X).
Since the vector fields XM,, span the tangent bundle of Ma, it is only necessary
to verify the following condition in order for OX to be a connection: if XM,
vanishes at a point f E Ma, in other words, if X E g(f), and if ¢ is a section
of La, then Vx (f) = 0. But if X E g(f ),
(L-1(X)0)(f) = i(f,X)O(f) =iµA(X)(f)0(f)
by definition of the homogeneous vector bundle La. Thus V is a connection
with moment map iµ!` (X), from which it follows that its equivaxiant curvature
equals i(µa(X) + SZA).
If G is simply connected, the element ipp E LT, so that if XG = L.
Consider the G-invariant twisted Dirac operator Da associated to the twist-
ing bundle La:
Da : F(G/T, S ®La) -4 T(G/T, S (9L,\).
Its equivariant index
indG(g, DA) = Trker(Dt)(g) - Trker(D,-)(9)
is the character of a virtual representation of G. Denote by s(P, Pa) = ±1
the quotient of the two orientations on r determined by P and Pa. The
following theorem identifies the Weyl and Kirillov character formulas with
the equivariant index theorem for the Dirac operator DA.
254 8. The Kirillov Formula for the Equivariant Index

Theorem 8.7. The fixed point formula for the equivariant index of Da is
given for X E t by the formula
EwEW e(w)ei(wa,X)
indG(e X , D A ) = faEP(e(«,X)12
- e-(a,X)/2)'
Thus, the index indG(g, DA) equals the character of the representation Ta up
to a sign.
The Kirillov formula Theorem 8.2 for the equivariant index of Da at X E g
small is

indG(eX, DA) = E(P,PA) JMA ._1/2(X)if(X) dm(f),

w here dm is the Liouville measure of Ma.

Proof. Let us start with the fixed point formula. The action of T on the flag
manifold leaves stable the points
{wT I w E W(G,T) = N(G,T)/T} C G/T,
and for generic t E T, the action of t on G/T has a finite number of fixed
points indexed by the Weyl group W. Thus it is easy to write the fixed point
formula for the character of the virtual representation ind DA: it equals
eiwa
W (-i)'PI E(W, t, P) /2
(t)
\ wtw-1)I'

W
detT (1 -
where the sign e(w, t, P) is the quotient of the orientations of r determined
respectively by P and by the element p(wtw-1) in the Clifford algebra C(t)
of the Euclidean space t. We have, for t = ex, X E t,

p(ex) = exp(2 E(Xea,fa)eafa)

aEP
= exp (-i (a, X/2)eafa)
aEP
_ (cosh(a, X/2) - i sinh(a, X/2)efa)
aEP
It follows that
E(w,P,e') _ (-1)lPIe(w) 11 sgn(isinh(a,X/2)).
aEP
Since
det,(1 - ex) = fJ (2isinh(a,X/2))2,
aEP
the fixed point formula follows.
8.2. The Weyl and Kirillov Character Formulas 255

Since the Clifford module S®La that we are considering is a twisted spinor
bundle, the Kirillov formula may be rewritten as

ind(DA)(e-X) = (2iri)- dim(G/T)/2

f /T
A9(X, G/T) ch (X, LA)-

By Lemma 8.6, the equivariant Chern character of La equals

.
chg(X, LA) = e i(f,X)e-uhf

The integral on G/T of an equivariantly closed differential form depends only

on its equivariant cohomology class. By Lemma 8.5, we can choose as a repre-
sentative for A. (X,G/T) the function jg 1/2(X). It only remains to compute
the highest degree term of chg(X, LA), which is (-21ri)`lim(G/T)/2e i(f,x)dm.
The replacement of X by -X and a careful comparison of orientations proves
the second formula.

Bott has given another realization of the representation TA, as the a-

cohomology space of a line bundle on the flag manifold G/T. Define r+ _
E«Ep(gr)« Then r+ is a subalgebra of ge and r has a complex structure
J such that r+ _ {X E rc I JX = iX}. There exists a unique G-invariant
complex structure on G/T such that Te'0(G/T) = t+. Furthermore, for any
µ E LT, the line bundle L. = G XT Cp, has a structure of an holomorphic
line bundle over G/T. It is easy to write down the fixed point formula for
the character of the virtual representation E(-1)2H°,i(G/T,.C,) on the a-
cohomology spaces of the line bundle Lµ.

Theorem 8.8. If A E LT, then for X E t,

[1
P Trgo,p(G/T,Gµ)(eX ) - EW
rL,EP(e(a,X)/2 - e-(«,X)/2)

This theorem shows that the virtual representation E(-1)PH°>P(G/T, Lu,)

is zero if )., = µ - ipp is not a regular element of t*. If A is regular, let wa
be the element of W such that w(PA) = P. Then the virtual representation
E(-1)PH°'P(G/T, LN,) is the representation e(wa)TA. In fact a more refined
result, the Borel-Weil-Bott theorem, is true:

H°,P(G/T,,c,.) = f T,\, IWAI =p,

l0, if Iwa 136 P,

where I w I is the length of the Weyl group element w E W (G, T).

256 8. The Kirillov Formula for the Equivariant Index

8.3. The Heat Kernel Proof of the Kirillov Formula

Let M be a compact oriented Riemannian manifold of even dimension n, and
let G be a compact Lie group with Lie algebra g acting on M by orientation-
preserving isometries. Let £ be a G-equivariant Clifford module over M with
G-equivariant Hermitian structure and G-invariant Clifford connection V.
Definition 8.9. The Bismut Laplacian is the second-order differential op-
erator on r(M, E) given by the formula

H(X) = (D + 4c(X))2 +G(X),

where we write c(X) for the Clifford action of the dual of the vector field XM
on S.
The operator H(X) is a generalized Laplacian depending polynomially on
X E g. It may be thought of as a quantum analogue of the equivariant
Riemannian curvature
R0(X) _ (V - ,(X))2 +G(X).
The restriction kt(x, x, X) IdxI of the heat kernel of the semigroup e-tx(x)
to the diagonal is a section of the bundle End(S) = C(M) ®EndC(M) (£) ®JAI.
The asymptotic expansion of kt (x, x, X) as a function of t for t small has the
form
00
kt(x,x,X) " (4-7rt) -n/2 tiki(x,X),
ii=o
where ki (x, X) depends polynomially on X (as we will show later). Denoting
by CQgj the algebra of formal power series on g, we may identify ki with
a section of the equivariant Clifford algebra C[g] ® C(M) ® Endc(M)(£).
The associated graded algebra of this filtered algebra is CQgj ® AT*M ®
Endc(M) (£).
If V is a Euclidean vector space with Clifford algebra C(V*), consider the
algebra CQg ® C(V*) with filtration

Fi((CQg 0 C(V *)) = C[e]k ® Ci-2k(V*),

k<i/2
where C[g]k is the space of homogeneous polynomials on g of degree k. This
filtration is analogous to the grading of the equivariant cohomology complex,
and will be called the equivariant Clifford filtration. The associated graded
algebra is isomorphic to C[g] 0 AV*, with equivariant grading
deg(P (&a) = 2 deg(P) + deg(a) for P E C[g] and a E AV*.
The next theorem is the generalization of Theorem 4.1 to the equivariant con-
text. It generalizes Bismut's infinitesimal Lefschetz formula [31].
8.3. The Heat Kernel Proof of the Kirillov Formula 257

Theorem 8.10. Let kt(x,x,X) = (x I e tx(x) I x) be the restriction of the

heat kernel of the semigroup e-tx(x) to the diagonal. Consider the asymp-
totic expansion
00
kt(x,x,X) - (47rt)Et''ki(x,X).
i=o
(1) The equivariant Clifford degree of ki is less or equal than 2i.
(2) Let a(k) = E°_o Gr[22] (ki) E CQg ®A(M, Endc(M) (E)). Then

a(k)(X) = detl/2 (sin/2)) exp(-FH/S(X))

Before proving this theorem, let us show how it implies the Kirillov formula
for the equivariant index. Recall that the heat equation computation of the
index is based on the McKean-Singer formula Proposition 6.3: for every t > 0,
(rye-tD2
indG (ry, D) Str )

For u E C, we consider the operator

Du = D + uc(X).
Since the operators D and c(X) commute with the action of etX on P(M,E),
so does D.
Proposition 8.11. For all t > 0 and all u E C, we have
indG(expX, D) = Str(exe-tD u
Proof. The result is true for u = 0; we will show that the supertrace on the
right-hand side does not depend on u. By Proposition 2.50, and the fact ex
and the heat operator a-tD' commute, we have
2
Str(exe-tDu) _ -tStr (&e_tt.eX).
Since ex commutes with c(X) and D, we obtain

Str(eXet0) _ -tStr (exe_t) 2

= -t Str((Duc(X) + c(X)Du)exetDu)
_ -tStr([Du,c(X)exetDu])
This vanishes by Lemma 3.49
Let us now set u = 4. It follows from this proposition that for every t > 0,

indG(e-tx, D) = Str(e-t.c(x)et(D-{-c(X)/4)2) = Str(e-tx(x))

258 8. The Kirillov Formula for the Equivariant Index

Thus we obtain the formula

indG(e-tx, D) = Idxj.
fM
Consider the asymptotic expansion in t of both sides of this formula. We
have

Str(kt(x, x, X)) - (41rt)-n/2 E tj Str(kj(x, X)).

j=0
It follows from the first assertion of Theorem 8.10 that if j < n/2, the total
degree of kj is strictly less than n. In particular, its exterior degree is strictly
less than the top degree and the supertrace of kj vanishes. Thus there is no
singular part in the asymptotic development of Str(kt (x, x, X)). Using the
formula Str(a) = (-2i) /2 StrE/S(a)[,), we obtain

indG(e-'X, D) - (27ri) -n/2 tj I dxl

j=0 JM Stre/S (kn/2+j (x, X) [n])
The first assertion of Theorem 8.10 also implies that kn/2+j (X, x)) [n] is a
polynomial in X of degree less or equal to j. The asymptotic expansion of
the left-hand side indG(e-tx, D) of the above equation is a convergent series:

indG(e-'X,
D) 00 (' 'ker(D+)(,C(X)') Trker(D-)(IC(X)j)) .

j=0 3.

Comparing the coefficient of tj in these two power series in t, we see that

only the component of o (kn12+j (X, x)) lying in C[g], ®ATT ,M contributes
to the equivariant index. Using the explicit formula of the theorem for o-(k),
we obtain

indG(e-x, D) = (27ri)-nl2J Ae(X,M) che(X,ElS)

M

which is the Kirillov formula.

Let us now prove Theorem 8.10. Following a plan which should be familiar
by now, we start by proving a generalization of Lichnerowicz's formula. Let
Ox E A'(M) be the one-form associated to X,
Ox (X, ) E A' (M).
Let DE,x be the Clifford connection Ve - 49x on the bundle 6; since OX
is scalar-valued, it is clear that VE,x is again a Clifford connection. Let us
show how Bismut's Laplacian H(X) can be written as a generalized Laplacian
associated to the connection VE,x on E. Let Ox be the Laplacian on E
associated to Vex.
8.3. The Heat Kernel Proof of the Kirillov Formula 259

Proposition 8.12. If FS/S and MS/S(X) are, respectively, the twisting cur-
vature and the twisting moment of X for the connection Vs and rM is the
scalar curvature of M, then
H(X) = Ax + 4rM + c(FE/S) + MS/S(X).
Proof. Recall from Lemma 7.15 that (MM (X )Y, Z) dOx (Y, Z) for Y, Z E
r(M, TM). Since 8x is a scalar-valued one-form, the z twisting curvature
Fe/S,x of the connection Ve,x is equal to F8/S-4d9x. Let Dx = D+.1c(X);
since Dx is the Dirac operator associated to the connection V,6,-X, Lich-
nerowicz's formula shows that

DX = 0-x + 4rM +c(F6IS,-x)

= 0-x + 4rM + c(FS/S) + 4 c(d&x).

By this formula and Proposition 3.45, we see that
A_x-Ox=DX-D?2 x-zc(dOx)
[D, c(Ox)] - a c(dOx)
_-oX+2d*Ox.
But d*Ox = Tr(MM (X)) E COO (M) vanishes, since MM is antisymmetric; this
implies that
Ox - A-x = pX
Since GE (X) = VX + MS (X ), we see that

H(X) = DX +'C-6(X)
= L_x + 4rM + c(FEIS) + VX + MS(X) + 4 c(dOx).
Since Ox = 0-x + VX and pe/S(X) = MS(X) + 4 c(d9x) (by definition of
the twisting moment), the result is proved.
We could model the proof of Theorem 8.10 on the proofs of the simple index
theorem of either Chapter 4 or Chapter 5. We will follow the former. Thus,
as in Section4.3, let V = Tx0M, E = £xu and U = {x E V I lixil < e}, where
e is smaller than the injectivity radius of the manifold M at xo. We identify
U by means of the exponential map x H expxo x with a neighbourhood of xo
in M.
For x = expxo x, the fibre £x and E are identified by the parallel transport
map -r(xo, x) : £x --p E along the geodesic x3 = expx,, sx. Thus the space
I,(U, £) of sections of £ over U may be identified with the space C°°(U, E).
Choose an orthonormal basis 8i of V = T,,M, with dual basis dxi of
T*XOM, and let ci = c(dx) E End(E). Let ei be the local orthonormal frame
obtained by parallel transport along geodesics from the orthonormal basis 8i
of Tx0M, and let ei be the dual frame of T*M.
260 8. The Kirillov Formula for the Equivariant Index

Let S be the spinor space of V* and let W = Homc(v.) (S, E) be the

auxiliary vector space such that E = S ® W, so that End(E) = End(S)
End(W) = C(V*) ®End(W), and Endc(v.) (E) = End(W).
We must study the operator H(X) in the above coordinate system, but with
respect to the frame of the bundle £ obtained by parallel transport along
geodesics emerging from xo with respect to the connection ve°x. Define
a:Uxg -->Cx by the formula

ax(x) = -4 f (t(R)9x)(tx)t-1dt,
0

so that Rax = -4t(R)9x. Let p(X,x) = e'x(x). Then p(X,x) is the

parallel transport map on the trivial line bundle over M with respect to the
connection d - 49x, along the geodesic leading from x and xo.
Lemma 8.13. The conjugate
(Vaix)p(X,x)_1

p(X, x)
of the covariant derivative Ve'X = V. - 49x((9i) is given on elements of
C°°(U,E) by the formula
ai+4 RijklxjckC' - 4 f(X)xi+Efijk(x)c'Ck+gi(x)+(hi(x),X).
fG

j;k<l j j<k
Here, Rijkl = (R(ak, al).oaj, ai).o and uM(X) _ aj)xo are
the Riemannian curvature and moment atxo, and flik(x) E Coo (U), 9i (X) E
CO°(U, Endc(v.) (E)) and hi(x) E C°°(U) ®g* are error terms which satisfy
fijk(x) = O(IxI2), gi(X) = O(JxI), and hi(x) = O(1x12).
Proof. Let us write
p(X, x) (d -
9x)p(X,x)-1

=d+wx,
where
Wx = -(dax + 49x) _ wi(X,x) dxi

satisfies wx (R) = 0 and depends linearly on X E g. As a special case of

(1.12), we have
G(R)w = dt(R)W + t(R)dw = -4t(R)dOx.
The one-form -4t(R)d8x is dual to :AM(X)R.

Taking the Taylor expansion

of both sides of this formula, we see that

wi(X, x) = >2(AM(X )aj, ai).oxi + (X, hi (x)),

4 J

where hi(x) = O(Jx12). Since p(X,x)(V8 X)p(X,x)-1 = V-0 +wi, the lemma
follows from Lemma 4.14.
8.3. The Heat Kernel Proof of the Kirillov Formula 261

Let (x I e-tx(x) I xo) be the heat kernel of the operator H(X); by The-
orem2.48, (x I e-tx(x) I xo) depends smoothly on X E g. We transfer this
kernel to the neighbourhood U of 0 E V, thinking of it as taking values in
End(E), by writing
k (t, x, X) = A(X, x)T (xo, x) (x I e tx(x) I xo ),
where x = expxo x. Since we may identify End(E) = C(V*) 0 End(W) with
AV*®End(W) by means of the full symbol map or, k(t, x,X) is identified with
a AV*®End(W)-valued function on Uxg. Consider the space AV*®End(W)
as a C (V *) ®End (W) module, where the action of C(V*) on AV* is the usual
one c(a) = e(a) - t(a). The following lemma shows that k(t, x, X) maybe
characterized as a solution to a heat equation on the open manifold U; the
proof is analogous to that of Lemma 4.15.
Lemma 8.14. Let L(X) be the differential operator on U, with coefficients
in C(V*) ® End(W), defined by the formula

p(X, x)-1L(X)p(X, x) 4 x(ei))2 - Vveiey + -1Ox(Veiei))

E((Ve: - O 4

+ --1 TM + E FE/s(ei, ej)c o + tie/S(X).

i<j

The function k(t, x, X) E C°°(]R+ x U, AV* ® End(W)) satisfies the differen-

tial equation
(at + L(X))k(t, x, X) = 0.
Using the explicit recurrence formula of Theorem 2.26 and the fact that the
coefficients of L(X) are polynomial in X, we see that the coefficients of the
asymptotic expansion of k(t,x,X) are polynomial in X.
To prove the Kirillov formula, we will rescale the kernel in a way similar to
that used in Section 4.3, except that we will perform an additional resealing
on the Lie algebra variable X E g. If a E C°°(1R+ x U x g, AV* (9 End(W)),
let bv,a, (0 < u < 1), equal
n
(bua) (t, x, X) = 57 u-i"2a(ut, u112x, u 1X)
i=O

Define the resealed heat kernel r(u, t, x, X) by

r(u,t,x,X) = un12(buk)(t,x,X).
We define an operator L(u, X) by the formula ubuL(X)bu 1; it is clear that
r(u, t, x, X) satisfies the heat equation
(8t + L (u, X)) r (u, t, x, X) = 0.
Using Lemma 8.13, we can give an asymptotic expansion for L(u, X). We
will show that
L(u, X) = K(X) +O(u1"2),
262 8. The Kirillov Formula for the Equivariant Index

where K(X) is a harmonic oscillator similar to that which entered in Sec-

tion 4.3.
In order to state the formula for K(X), recall the matrix R with nilpotent
entries equal to Rij = (R,,0 8i, 8j) E A2V* where Rx0 is the Riemannian
curvature of M at xo, and the element F of A2V* ® End(W) obtained by
evaluating the twisting curvature Fe/s at the point xo. Let pM(X) E End(V)
be the evaluation of the moment of the manifold M at the point xo and Lie
algebra element X E g, and let µE/s(X) E End(W) be the evaluation of
the relative moment of the Clifford module E at xo and Lie algebra element
X E g.
Proposition 8.15. The family of differential operators
L(u, X) = uS.uL(X)6,,, 1
acting on COO (U x g, AV* (9 End(W)) has a limit K(X) when u tends to 0,
given by the formula
K(X) (ai - 4 (Rij +A j1(X))xj)2+F+µe/s(X).

i j

Proof. Using the fact that u1/2Suci5u 1 = E(ei) - ut(ei), we see that the diff-
erential operator

Va,X'u = u1/2Su(P(X,x)Va.XP(X,x)-1)6 1 =Ai(u) +A2(u),

where
Ai(u) =8i + 4 E RijklX3(Ek - utk)(! - utt) - I A,M(X)Xj,
j;k<l j
A2(u) =u-1/2 E fijk(ui/2x)(e. - uti)(Ek - ubk)
j<k
+ u1/2gi(u1/2x) + u-1/2(hi(ui/2x), X).
The functions fijk(x), gi(x) and hi(x) vanish to sufficiently high order in x
that limu_,o A2(u) = 0. Thus, we see that DagX'u has a limit as u -> 0, equal
to
8i + .
4 E
j;k<l
RZjkIXlekel
4
j
TAM (X)Xj.
13

Since Rij = >.k<1 Rjikleke.l, the above limit is seen to equal

8i - 1-i E(Rij + pM(X))xj.
j
Let ci(u) = u1/26uciSu i = Ei - uci. The operator L(u, X) equals the sum
of

ej)(ui/2x)c
- i
E FE'S(ei,
i<j
(u)o (u) + (µE/1(X))(u1/2x)
8.3. The Heat Kernel Proof of the Kirillov Formula 263

and
u1/2 S'X' + 4urM(ul/2x).
i
Clearly, these last two terms vanish in the limit u --> 0. The first term has
the limit

i j
while the second and third terms converge to

F(ai,aj)e'6j +µEIS(X)Ix=o = F+FzE/S(X),

i<j

since the vector fields ei coincide with ai when x = 0.

The operator L(0, X) = K(X) is a generalized harmonic oscillator, in
the sense of Definition 4.10, associated to the n x n antisymmetric curvature
matrix Ri. +µM (X) and the N x N-matrix F+µE/s (X) (where N = dim W),
with coefficients in the commutative algebra A+V*. It follows that there is a
unique solution of the heat equation
(at + K(X))r(t, x, X) = 0
such that limt_,o r(t, x, X) = 6(x), given by the formula

(4irt)-n/2 +AM
1/2 exp(-t(F + µe/s(X )))
det (siiri (R(R + µ(M()X))/2 /
x exp(- 1 Cx t(R+AM(X)) coth(t(R+AM(X))Ix))
4t 2 \ 2
If J is an integer, let C[g](J) be the quotient of the algebra of polynomials
on the vector space g by the ideal of polynomials of order J+1. If a E C[g](J),
it has a unique polynomial representative of the form a(X) a0Xa;
we define the norm of a to be sup,, Ia,, 1.
We can think of r (u, t, x, X) as defining an element
rj(u,t,x,X) E C°O(R+ x U, AV* (9 End(W)) ®C[g](J)
by taking the Taylor series expansion of r(u, t, x, X) at X = 0 up to order J,
in other words,

rJ(u,t,x,X) = >
IoI<J
aj.
(a5r(u,t,x,X))x=0X°

Lemma 8.16. There exist polynomials -yi,J(t,x) on R x V, with values in

AV* ® End(W) ® C[g](J), such that for every integer N, the function
qt 2N
rJ (u, t, x, X) = (x) > Ui/2-Yi,J (t, X, X)
i=-n-2J
264 8. The Kirillov Formula for the Equivariant Index

approximates rJ(u, t, x, X) in the following sense:

-rim (u, t,x,X))II AV*®End(W)®C[g](J) <C(J,N, j,a)uN,
(-)It?.

ax'

for 0 < u < 1 and (t, x) lying in a compact subset of (0,1) x U. Furthermore,
yi,J(0, 0) = 0 if i # 0, while yo,J(0, 0) = 1.
Proof. By Theorem 2.48, there exist AV* ® End(W)-valued smooth functions
Oi on U x g such that
N
C(N)tN-n/2
I I k (t, x, X) -qt (x) E ti0i (X, X) I
i=0
for t E (0, 1), x E U, and X in a bounded neighbourhood of 0 E g. Fur-
thermore, we can differentiate this asymptotic expansion with respect to the
parameter X. It follows that there exist AV* ® End(W) ® C[g](J)-valued
smooth functions iJ ,J on U such that
N
J)tN-n/2
k J (t, x, X) - qt (x) E tz?Pi,J (x, X) < C(N,
i=0
for t E (0, 1) and x E U, where kj(t, x, X) is the Taylor expansion of k(t, x, X)
around X = 0 to order J. As in the proof of Proposition 4.16, we can suppose
thati,J is polynomial in x.
Rescaling by U E (0, 1), we obtain
N
rj (u, t, x, X) [p) - u-p/2qt (x) E (ut)'gt i,J (ul/2x u-1X) [p]
i=0

< C(N J)uN-p/2-JtN-n/2

Thus, we see that the function u -+ rJ(u, t) x, X) has an asymptotic expansion

of the desired form, where yj,j(t,x,X)[p] is the coefficient of ui/2 in
00
u-p/2 r(ut)jWj,J(ul/2x, u-1X
)[p]
j=0

Since Oj,J is a polynomial in X of order at most J, we see that (yi,J)[p] = 0

if i < -p/2 - J. The argument for the derivatives is similar. Since
u-p/2,po J(0, u-1X
E ui/2yi,J(0, 0, X )[P) = )[p1

we see that yi, J (0, 0, X) [p] = 0 unless i = p = 0, and that yo, J (0, 0, X) = 1. 0
It is clear that rJ (u, t, x, X) satisfies the differential equation
(8.3) (Bt+LJ(u,X))rJ(u,t,x,X) =0,
where LJ(u,X) is the action induced by the differential operator L(u,X)
on C°° (U) ® AV" ® End(W) ® C [g] (J) . Let us show how this implies that
Bibliographic Notes 265

rj (u, t, x, X) has a limit as u - 0, using the fact that L(u, X) = K(X) +

O(u1/2). Let 2 be the largest integer for which y_,g,j(t,x,X) is non-zero.
Expanding (8.3) in a Laurent series in u1'2, we see that the leading term
u-e/2gt(x)y_t, j(t, x, X) of the asymptotic expansion of rJ(u, t, x, X) satisfies
the heat equation (8t + K(X))(gt(x)y_t,j(t,x,X)) = 0. Since formal solu-
tions of the heat equation (8t + K(X)) (qt (x)y(t, x, X)) 0 for the harmonic
oscillator are uniquely determined by y(0, 0, X), and since yi,1 (0, 0, X) = 0
for i < 0, we see that y_t,a must equal zero unless £ < 0. In particular, we see
that there are no poles in the Laurent expansion of rJ(u, t, x, X) in powers of
u1/2.
We also learn from this argument that the leading term of the expansion of
rj (u, t, x), namely rj (0, t, x, X) = qt (x)yo,j (t, x, X), satisfies the heat equa-
tion for the operator L(0, X) = K(X), with initial condition
'Yo,j(t = 0, x = 0, X) = 1.

Applying Theorem 4.12 to the operator K(X), we obtain the following result.

Proposition 8.17. The limit lim r j (u, t, x, X) exists, and is given by re-
u->o
duction modulo X j+' of the formula

(47rt)det( sinht(R+µ(M ))/2)1/2 x

x exp (-(' + µ (X))/2)
x> -t(F+coth(t(R

Since this is true for arbitrary J, we have only to set t = 1 and x = 0 to

obtain Theorem 8.10.

Bibliographic Notes
The Kirillov formula Theorem 8.2 was proved in Berline-Vergne [24], using
the localization theorem for equivariant differential forms and the equivariant
index theorem of Atiyah-Segal-Singer. The local version of the theorem proved
in Section 3 is a generalization of the results of Bismut [32]; in particular, the
definition of the "quantized equivariant curvature" H(X), and the proofs of
Propositions 8.11 and 8.12 are due to him.
In Sections 9.4 and 10.7, we will explain the relationship between the Kiril-
lov formula and the index theorem for families with compact structure group;
the relationship comes about by considering the family P xG M, where P
is an approximation of the universal bundle EG for G. We will see that the
Kirillov formula is equivalent to Bismut's local family index theorem, in the
case of families with compact structure group.
266 8. The Kirillov Formula for the Equivariant Index

Additional Remarks
There is a formula combining the equivariant index formula of Chapter 6
with the Kirillov formula, proved in Berline-Vergne [24]. This formula may
be formulated using the language of bouquets discussed in the Additional
Remarks at the end of Chapter 7. Introducing a bouquet bch(E/S) which
collects the equivariant relative Chern characters of the bundle E over the
fixed point sets M(g), we have
b
indG(D) = (27ri)-dim(M)/2 J bch(E/S).
M
This type of formula has also been used to express the index of a transversally
elliptic operator (Berline-Vergne [27]; see also Duflo [53]), and to reformulate
results on characters of non-compact groups (Vergne [103]).
Chapter 9. The Index Bundle

Consider a manifold B and a finite dimensional Hermitian superbundle 7.1 =

®7.1- -+ B. Let D be an odd endomorphism of 7-1 with components
D± : 7-l± - 7-lF, such that ker(D) has constant rank, so that the family of
superspaces (ker(Dz) I z E B) forms a superbundle over B, called the index
bundle of D. Let A be a superconnection on 7-l with zero-degree term equal
to the odd endomorphism D of 7-l, and let .7:' = A2 E A(B, End(7-1)) be the
curvature of A; all notations are as in Section 1.4. If we assume that D is
self-adjoint and has kernel of constant rank, then at the level of cohomology,
ch(A) = ch(ker(D)).
In fact, this equation has a refinement at the level of differential forms.
If t > 0, we may define a new superconnection on 7{, by the formula
At = t112 D + A[1] + t-1/2A[21 + ... ,
where A[i] is the connection associated to A. This superconnection has Chern
character form
ch(At) = Str(e-At).

Let Po be the projection from 7-l onto the subbundle ker(D). We may also
define a connection on the superbundle ker(D) by projection,
P0.
In Section 1, we will prove the following result:
lim ch(At) = ch(Vo).
t--+oo
Intuitively speaking, as t tends to infinity, the supertrace
Str(e-t°2+o(t1i2))

ch(At) = Str(e-At) =
is pushed onto the sub-bundle ker(b). The proof, which is given in Section 1,
is substantially more complicated than this simple idea.
Our real interest lies in an infinite-dimensional version of this, in which
the family D of operators on the finite dimensional bundle 7{ is replaced
by a family of Dirac operators. Our technical tools will be the estimates
of Chapter 2 on the heat kernels of generalized Laplacians, and the spectral
theorem.
268 9. The Index Bundle

Consider a fibre bundle 7r : M --> B with compact fibres. We denote by

T(M/B) the vertical tangent bundle, which is the sub-bundle of TM con-
sisting of vertical vectors, and by IA,rj the vertical density bundle, which is
isomorphic to I AM 10 ir* IAB -1. If £ is a superbundle on M, we denote by
it*£ the infinite-dimensional superbundle on B whose fibre at z E B is the
Frechet space
(ir*£) r(Mz7 £z ®JAm.11/2);
here, Mz = 7r-1(z) is the fibre over z E B, and £z is the restriction of the
bundle £ to Mz. The space of sections of it*£ is defined by setting
r(B, it*£) = r(M, £ (& 1A,r11/2),

and the space of differential forms on B with values in 7r*£ by

A(B,ir*E) = r(M,7r*(AT*B) ®£®1A, 11/2).

Let D = (Dz I z E B) be a smooth family of Dirac operators. By a super-

connection on 7r*£ adapted to D, we mean a differential operator A on the
bundle E = ir* (AT*B) ®£ ®1A1,11/2 over M, of odd parity, such that
(1) (Leibniz's rule) if v E A(B) and s E r(M, E), then
A(vs) = (dBV)S + (-1)I"lvAs;
(2) A= D + j:dim(B) A[2j, where A11 :.A*(B,7r*£) --> A*+i(B,7r*£).
In the appendix to this chapter, we show that the curvature of the super-
connection A has a smooth heat kernel (x I e-A2 I y) varying smoothly as a
function of z E B and x, y E M. Because of this we can extend the definition
of the Chern character of a superconnection to the infinite dimensional bundle
7r*£->B.
Suppose that the dimension of ker(Dz) is independent of z E B. In this
case, the vector spaces ker(Dz) combine to form a vector bundle ind(D), which
is the index bundle of the family D; the index of a single Dirac operator is
a special case of this construction. In Section 2, we show that, as in the
finite-dimensional case, we can define a connection on the index bundle by
the formula Vo = Po A[i] Po, and the result of this chapter is a formula
due to Bismut for the Chern form ch(Vo), which is a generalization of the
McKean-Singer formula. Just as in the finite-dimensional case, and by the
same proof, we will show in Section 3 that
lim ch(At) = ch(Vo).
t-->oo
In Section 4, we consider the special case where the family D is associated
to a principal bundle P --+ B with compact structure group G. In this case, it
is possible to calculate directly the Chern character form ch(ker(D)) = ch(Vo)
using the Kirillov formula of the last chapter, and we obtain the formula

ch(ker(D)) = (2iri)- dim(M/B)/2 JM/B A(M/B) ch(S/S),

9.1. The Index Bundle in Finite Dimensions 269

where A(M/B) is the A-genus of the vertical tangent bundle T(M/B), and
dim(M/B) = dim(M) - dim(B) (we assume that M is connected). This
formula is true at the level of cohomology for any family of Dirac operators,
as we will show in the next chapter.
Section5 is devoted to studying the case when the vector spaces ker(Dz)
vary in dimension. The index bundle ind(D) is now only locally a vector
bundle, but it turns out that it has a well-defined Chern character in de Rham
cohomology, and that ch(At) lies in the same cohomology class for all t > 0
if A is a superconnection on 7r*£ such that A[o] = D. In Chapter 10, in the
case where the family D is associated to a Clifford connection, we will present
Bismut's construction of a very particular superconnection for which ch(At)
possesses a limit limt_,o ch(At); combined with the results of this chapter,
this enables us to calculate the Chern character of the index bundle ind(D).
In Section 6, we define the zeta-function and zeta-function determinant of a
generalized Laplacian. Using this, we define in Section 7 the determinant line
bundle det(ir*£, D) associated to a smooth family of Dirac operators, and, in
the case in which the index of DZ is zero, a section det(D+) of det(7r*£, D). The
determinant line bundle has a canonical metric, known as the Quillen metric,
and we define a connection on det(,7r*£, D) preserving this metric given the
data of a connection on the bundle M --3 B and a Hermitian connection on
the bundle £ over M.

9.1. The Index Bundle in Finite Dimensions

In this section, we will discuss in detail the finite-dimensional analogue of the
index bundle, since it is easier to visualize what is happening here than in the
infinite-dimensional case. Notation will be as in Sections 1.4 and 1.5.
Consider a manifold B and a superbundle f = 7-l+ ® 7.1- --> B. Let D
be an odd endomorphism of l with components Dt : 7-l± -; 7{=F, such that
ker(D) has constant rank, so that the family of superspaces (ker(Dz) I z E B)
forms a superbundle over B, called the index bundle of D. (From now
on, points of B will always be denoted by z.) Suppose that x+ and 9.1-
have Hermitian structures for which D- is the adjoint of D+ (so that D is
self-adjoint). If x E f and y E ker(D), then by the assumption that D is
self-adjoint,
(Dx, y) = (x, Dy) = 0.
Denote by 7-lo C l the superbundle ker(D), graded by 710 = ker(D+), and
by 7-li the the image im(D) C H of D. The above equation shows that fi is
the orthogonal complement of 71o,

7-l+ = fl (D'H
7.1- = 7-l0 ®1-li ,
270 9. The Index Bundle

and hence that the bundle ker(D-) is isomorphic to the bundle coker(D+)
7-1-/im(D+), and the bundles ?i and 71j are isomorphic.
Let A = A[o] + A]1] + A[2] +... be a superconnection on 71, with curvature
.7=' = A2 E A(B, End(7-l)). By (1.34), the Chern character form of A,

ch(A) = Str(e .F),

is equal in de Rham cohomology to the difference of the Chern characters of

the bundles 7-l+ and 7-1-.
Let A be a superconnection whose zero-degree term A[o] equals the odd
endomorphism D of R. If we assume that D is self-adjoint and has kernel of
constant rank, then by the above discussion and (1.34), we have the equality
in cohomology

ch(A) = ch(7-l+) - ch(7-1-)

= ch(7-lo) + ch(fi) - ch(fo) - ch(7-li )
= ch(fl) - ch(7-l0-) = ch(ker(D)).
We will refine this equation to the level of differential forms.
Let Po be the orthogonal projection of 71 on Ho, and let P1 = 1 - PO
be the orthogonal projection of 7-l on 7-ll. The endomorphism D+ gives an
isomorphism between 7-1 and 7-1j .
Let A be the superconnection

A= PO
which preserves the spaces A(B,7-lo) and A(B,71k) C A(B,7{).
We will make constant use of the following notation: if K E A(B, End(7-1)),
we write
K= a Q
,
7 b

which means simply that

a PoKP0 P0KP1
ry b P1KPo P1KP1
with a E F(B, End(7-lo)) etc.
Since A commutes with Po, we see that its curvature has the form

.7==A2= R 0
0 S

Denote by Vo the connection on the bundle 7-Lo given by projecting the con-
nection A(1) onto the bundle 7-Lo:

Vo=PoA]1]Po.
9.1. The Index Bundle in Finite Dimensions 271

We filter the algebra M = A(B, End(7{)) by the subspaces

Mi = E Aj (B, End(-H)).
j>i

Lemma 9.1. The differential form R lies in M2, and the curvature of the
connection Do equals R[21 .
Proof. The superconnection Ao = Po - A A. Po on the bundle Ho has curvature
Ao=R. Since we see that

i>2

from which the lemma is clear.

For t > 0, let St be the automorphism of A(B, 7{) which acts on Ai(B, 7-l)
by multiplication by t-i/2. Then At = t1/2St A A. St 1 is again a superconnec-
tion on 7-l, and the decomposition of At into homogeneous components with
respect to the exterior degree is given by the formula
At = t1/2A[o] + A[1[ + t-1/2A[2] + ... .
The curvature -Ft = At of the superconnection At is the operator

and the cohomology class of ch(At) = Str(e-mot) is independent of t, and is

equal to the difference of the Chern characters ch(fo) - ch(lo) for all t > 0.
We will study the limit of ch(At) as t -4 oo; it is remarkable that the following
stronger result holds.
Theorem 9.2. Let 7-l = 7-l+ ® 7-1- be a Hermitian super-vector bundle and
let D be a Hermitian odd endomorphism of It whose kernel has constant rank.
Let A be a superconnection on 7d with zero-degree term equal to D. Fort > 0,
let
At = t1/2bt A. St I = t1/2D + A[1 + t-1/2A[21 + .. .
be the rescaled superconnection, with curvature J. Then fort large,
Ile^Y` - e-8121111 = O(t-1/2)
uniformly on compact subsets of B and for all Ce-norms.
Proof. We start by proving the following lemma.
Lemma 9.3. (1) Under the decomposition 71 = 7to ® R1, the curvature F
can be written as
- XY M2 .M1
Z T E M1 .Mo .

(2) The endomorphism T[o] E I'(B, End(f1)) is equal to Pi D2 P1 and is

positive definite.
272 9. The Index Bundle

(3) Denote the inverse of T[01 on xl by G. The curvature R[2] of the con-
nection Vo on Ho equals
R[2] = X[2] - Y1]GZ[1].

Proof. We write A=A+w,where EM1:

0
0I.

Expanding the right-hand side of the equation

.F = (A + w)2 = .P + [A, w] + w A w,

we obtain
R+µv Po[A,ll]P,
P, [A, V] PO S + vµ

R[2] + u[I]v[11 µ[1]D

mod
M3 M2
Dv[ll D2 M2 M1
If we write
X Y
Z T
we see that
X[2] - Yl]GZ[l] = (R[2] +µ[1]v[1]) - (µ[l]D)G(Dv[I]) = R121. D
The following lemma is the key step in the proof. We will use the fact that
the space of matrices g of the form
g = 1 + K, where K E .Ml,
form a group; to obtain the inverse of such a matrix, we use the formula
(which is a finite sum, since B has finite dimension),
00
(1 + K)-1 = 1 + E(-K)k.
k=1

Lemma 9.4. There exists a matrix g with g - 1 E M1, such that

X Y 91= U 0
gig-1 = g Z T 0 V
Furthermore
U X - YGZ (mod M3),
V T (mod .M1).
9.1. The Index Bundle in Finite Dimensions 273

Proof. To construct a matrix g which puts F in diagonal form, we argue by

induction on dim(B) - i. Thus, assume that we have found gi such that
92-F9tiIZi Xi Yi Mi
T, E Mi MO 7

with Ti - D2 (mod M1); in particular,

0 0
E AA
0 1 - GTT 1-

We write
-1
1 -YzG Xi Yi 1 -YjG X.i Yi
GZi 1 Zi Ti GZi 1 Zi T,
0 -YZG
Since E Mi, we see that
GZi 0

1 -YzG 1 YiG
E M2z-
GZi 1 -GZi 1

We have the following explicit formulas for Xi, Yi, Zi, and Ti mod M2i:
Xi = Xi - 2(YiG)Zi + (YiG)TT(GZZ)
Xi (mod M2i);
Y Yi(1 - GTi) + (Xi - (YG)ZZ)(YjG) E Mj+1;
Zi = (1 - TG)Zi + (GZZ)Xi - (GZi)Yi(GZi) E Mi+1;
Ti = Ti + (GZi)Xi(YZG) + Z2(YG) + (GZZ)Y
Ti (mod M1).
Thus, we can continue the induction.
Now, suppose that we have a matrix g of the required form which diagonal-
izes .7. This implies that
1+K M X Y U 0 1+K M
N 1+L Z T 0 V N 1+L

for some
K M E M1, from which we obtain the equation
N L
X +KX +MZ Y+KY+MT U(1 + K) UM
NX+Z+LZ NY+T+LT VN V(1 +L)
Since X is in M2 and K, L, M, N, Y and Z are all in M1, we see that
(1) V = (T +LT +NY)(1 +L)-1 -T (mod MI),
(2) hence GV - 1 (mod MI),
(3) U = (X + KX + MZ)(1 + K)-1 - X + MZ (mod M3),
(4) Y + MT = UM - KY E M2i
274 9. The Index Bundle

(5) hence, multiplying on the right by G, M - -YG (mod .M2).

Combining the last two formulas, we see that
U - X - YGZ (mod .M3).
By this lemma, we may write
e-t&(U)
e tbt(-W) = 6t(g)-1

0
0
e-tbt(V) St(g)

Lemma 9.5. We have the estimate ie-tbt(v) 1 < Ce et for some positive e,
as well as for all of its derivatives.
Proof. Essentially, this is true because V[o] = T[o] = D2 is positive definite on
'Hi. The proof uses the Volterra series for a tst(v) (see 2.5): if V = D2 + A,
then
e-t6t(V) = 1:(-t)klk,
k>O
where
Ik = f e--otD26t(A)e-Q1tD2St(A)...St(A)e-01k-ltD26t(A)e-°ktD2do-1...da,.
Ok
On the simplex
Ok = &0, ... 0k) E Rk+1 I Ek 0 ai = 1, ai
, 01,

one of the vi must be greater than (k + 1)-1. Since the operator

e-°'titD2

decays exponentially when t - oo, while the operators e-,;tD2, j 54 i, are

bounded, we see that Ik decays exponentially. The sum is finite, since Ik is a
sum of terms of degree at least k with respect to the grading of the differential
forms on B.
Using this lemma and the fact that tbt(U) = R[2) + 0(t-1/2), which is a
consequence of Lemma 9.4, it follows that
St(g)-lle-R[21 +0(t-1/2) O
St(g) + 0(e-Et).

Both St(g) and St(g)-1 = St(g-1) have the form

1 + O(t-1/2) O(t-1/2)
O(t-1/2) 1 + 0(t-1/2)I
It follows that
e -t6t(F) Ie-R[21 +O(t-1/2) O(t-1/2)
(9 . 1) =
O(t-1/2) O(t-1)
The argument is similar for the derivatives with respect to the base, using the
formula of Theorem 2.48.
9.1. The Index Bundle in Finite Dimensions 275

Corollary 9.6. The limit limt_,OO ch(At) exists, and equals the Chern char-
acter of the connection Vo on the superbundle Ho = ker(D).
We will now give an explicit homotopy between ch(At) and limt_,00 ch(At).
Let us write
a(t) = Str(ddtte At) E A(B).
The transgression formula (1.33) states that
d ch(A,) = _da(s).
ds
This may also be seen by the following construction, which is valid for _any
smooth family As of superconnections on a superbundle £ --+ B. Let B =
B xR+, and let E be the superbundle £ x R+ over b, which is the pull-back
to b of S. Define a superconnection A on £ by the formula
(9,3(X, s)
(A/3) (x, s) = (A3/9(., s)) (x) + ds A
8s
The curvature F of A is

=F3-d9 Ads,

where F., = A2 is the curvature of A9. Since (ds)2 = 0, the Volterra series
expansion of
e-7 is

e-7 = e-7e + (f e-u7, d 9 ei1-u)F du I Ads.

Let
a(s) = Str(dAse-1-e).
d
Taking the supertrace of the above formula, we see that
ch(A) = ch(A9) + a(s) A ds,
with respect to the decomposition b = B x R+. Since ch(A) is closed in
A(B) and ch(A9) is closed in A(B), we infer that
dch(A9) = _da(s).
ds
Integrating this formula, we see that
T d
ch(At) - ch(AT) = - f ch(A9) ds
T
d Str (d 88 a-A2,) ds
= itt
T
= d f a(s) ds.
t
276 9. The Index Bundle

Our task is to show that this integral converges as T - oo. We will prove
this by a small modification of the proof of Theorem 9.2.
Theorem 9.7. The differential form
\
a(t) = Str I d tt e-A2 I E A(B)

satisfies the estimates, for t large,

C(e)t-3/2
Ila(t)Ile <
on compact subsets of B, for all Ce-norms.
For all t > 0, the integral

0C a(s) ds E A(B)
it
J
is convergent, and defines a differential form which satisfies the formula

ch(At) - ch(Vo) = d

Proof. We have the following formula for a(t):

IJ a(s) ds.

At e-tst(.r)ll
a( t) = St r (d
dt l

Since
di m(B )
dAt D
dt 2t1 /2
i=2
2t(i+l)/2
we have
dAt 1 0 0
+0 (t -3/2 ).
dt 2t l/2 0 D

Inserting this into (9.1), we see that

dAte_tst(g7) _ 1 0 0 O(t_1)) e-R[21 +O(t-1/2) O(t-1/2)
+
dt 2tl/2 0 D 0(t-1/2) O(t-1)

O(t-3/2) O(t-2)
O(t_1) O(t-3/2)
Since only the diagonal blocks contribute to the supertrace, this proves the
theorem.
In the special case in which A has the form V + D, so that A[j] = 0 for
i > 2, we see that At = V + tl/2 D equals V at t = 0. Furthermore,

a ()
t = 1 Str( D e-A2) = 1O(t-1/2) for t --> 0,
2t1/2 1 O(t-3/2) for t --> oo.
9.2. The Index Bundle of a Family of Dirac Operators 277

Thus, the integral f O° a(t) dt defines a differential form on B, and we have

(9.2) ch(V) - ch(Vo) = d oo a(t) dt.

J0
This is a refinement of the fact that
[ch('H+)] - [ch(x )] = [ch(xo )] - [ch(xo )] E He(B)
to the level of differential forms.

9.2. The Index Bundle of a Family of Dirac Operators

The goal of this chapter is to give an infinite-dimensional version of the last
section, in which the family D of operators on the finite dimensional bundle
7-l is replaced by a family of Dirac operators. The details are slightly more
complicated for two reasons: firstly, because we work in infinite dimensions,
and secondly because we will be interested in the more general case in which
the dimension of ker(DZ) varies as we move about the family.
By a family of manifolds (Mz I z E B), we will mean the family of
fibres M,z = it-1(z) of a smooth fibre bundle -7r : M --+ B. We assume that
the fibres M,,, are compact. However, we will not assume that the base is
compact. Let T (M/B) C TM be the bundle of vertical tangent vectors, with
dual T*(M/B) = T*Ml7r*T*B. To the family M -+ B, we associate the
vertical density bundle
IA,,I = IAT*(M/B)I IAMI ® (Ir*IABI)-1
When restricted to a fibre M,z of M, IA,I may be identified with the bundle
IAM.I of densities along the fibre. We will call a section of IA,,. a vertical
density.
By a family of vector bundles (£z z E B) we mean a (smooth) vector
I

bundle £ -* M, so that £,z is the restriction of the bundle £ to M. For us, £

will always be a complex superbundle with Hermitian metric. To the family
£ -+ M, we associate the infinite-dimensional superbundle 7r*£ = 7r*£+
,7r*£- over B, whose fibre at z E B is the Frechet space r(Mz7 £z (& I JAM. I1/2).
A smooth section of it*£ over B is defined to be a smooth section of £®IAx11/2
over M:
(9.3) r(B,7r*£) = r(M,£ (& IAxI1/2)

The bundle 7r*£ carries a metric, defined by means of the canonical metric on
each fibre r(MM, £z (9 IAm. 11/2). Indeed, if Oz E 7r*£z = r(Mz7 £z o I AMs I1/2),
we see that ICrz(x)12 E r(Mz7IAM I), and hence may be integrated:

I0x12 = f I0z(x)I2.

Mx
278 9. The Index Bundle

The reason for inserting the line bundle IA,I112 is precisely in order to have
such a canonical metric along the fibres 7r..E. The fibre bundle M will have
a metric along the fibres, which trivializes IA,,I. However, unlike in the case
where B is a point, it is important to distinguish between the bundles 7r*E
and z --* r(Mz, £z).
Instead of working with the bundle End(7r*E) of all endomorphisms of 7r,,E,
we will restrict our attention to only those endomorphisms of the fibres which
are the sum of a differential operator and a smoothing operator. Thus, we will
denote by D(E) the bundle of algebras over B whose fibre at z is the algebra
D(MM, Ez (9 J Am. I1/2) of differential operators on Ez ® IAM. I1/2, and whose
smooth sections are families of differential operators Dz whose coefficients in
a local trivialization of M and E depend smoothly on the coordinates in B.
We denote by K(£) the bundle of algebras whose fibre at z is the algebra of
smoothing operators on the bundle Sz ® IAM= I1/2, and whose smooth sections
are smooth families of smoothing operators Kz. (We refer to reader to the
appendix for more details on this bundle.) Since IC(S) is a bundle of modules
for D(E), we may form an algebra from the sums of operators in D(E) and
K(£).
Definition 9.8. The P-endomorphisms of the infinite-dimensional bun-
dle 7r*E are smooth sections of the bundle D(E) + K(£), that is, families of
operators which may be written in the form
Dz + Kz,
where (Dz I Z E B) is a smooth family of differential operators acting on the
family of bundles Ez ® I AM= I1/2, and (Kz I Z E B) is a smooth family of
smoothing operators on the same family of bundles. We write Endp(7r*E) for
the space of smooth sections of this bundle.
We will consider a smooth family D = (Dz I Z E B) E r(B, D(S)) of Dirac
operators on S. Thus, D is an odd operator with respect to the 7L2-grading,
with components Dt : r(B,-7r*EF), and D2 E r(B,D(E)) is
a family of generalized Laplacians. Note that the Dirac operator Dz defines
a Riemannian structure on the fibre Mz (that is, an inner product on the
vector bundle T(M/B) --> M of vertical tangent vectors), and an action of
the Clifford algebra bundle C(T*(M/B)) on E. In particular, we obtain a
canonical trivialization of the bundle IA,r I.
Now assume that for each z E B, the operator Dz is self-adjoint with
respect to the Hermitian structure on the vector bundle £z. The vector space
ker(Dz) has finite dimension. The aim of this section is to define the index
bundle of the family D as a superbundle on B whose fibre at z E B is equal
to ker(Dz) when the dimension of this vector space is independent of z; in
Section 5, we will show that the difference bundle [ker(D+)] - [ker(D-)] makes
sense even without this condition.
If X is not an eigenvalue for any operator of the family Dz, then the su-
perbundle ? [o,a) whose fibre at the point z is the spectral subspace of (D')2
9.2. The Index Bundle of a Family of Dirac Operators 279

associated to the interval [0, A) is a smooth bundle over B, as we will now

prove. This superbundle is also a natural candidate for the index bundle,
since (Dz)+ exchanges the even and odd eigenspaces associated to a given
non-zero eigenvalue.
Lemma 9.9. Let P E I'(B, K(£)) be a smooth family of smoothing opera-
tors on the family of Hermitian vector bundles £ --> M, such that Pz is a
finite-rank projection for all z E B. Then dim(im(P')) is constant in each
component of B, and the spaces (im(Pz) z E B) form a smooth bundle
I

im(Pz) over B.
Proof. The dimension of the vector space im(Pz) is an integer, given by the
formula
dim(im(Pz)) = Tr(Pz) fMTPz
,
where (x I Pz I x) E r(Mz, End (E) (9 1AM= 1) is the restriction to the diagonal
of the kernel of Pz. Since (x I Pz I x) varies smoothly as a function of z
by assumption, the trace varies smoothly as well; thus the dimension of the
image of Pz is constant.
We now prove that the family of vector spaces im(Pz) form a smooth
superbundle. If z° E B, let U be a neighbourhood of z° in B such that
7r : M ---* B may be trivialized over U. In this way, we may replace the base
B by an open ball U C RP, the family M by the product M° x U, and the
bundle £ by the bundle £° x U, where £° -- M° is a superbundle over M°.
The space 7r*£z of sections of £z ® 1AM= X1/2 does not depend on z. This gives
a trivialization of 7r*£ as r(M°i £° ®lAM(, 11/2) x U. Thus Pz is a smooth
family of finite rank projectors on the fixed space r(M°,£° 01AMo1 1/2).
Choose a basis O°, (1 < j < m), of im(P°) at the point 0 E U. We may
extend these sections smoothly to U, by extending the corresponding sections
of £ (9 1A, 1/2 over Mz to -7r-1(U) C M, and we obtain m families of sections
E r(U,7r,£). The sections Pct are smooth as a function of z, lie in
im(P), and are linearly independent for JzJ sufficiently small. Thus, they form
a basis of im(P), for lzJ small. From the smoothness of P with respect to
z, we see that Pz0J1(y) is smooth with respect to (y, z) E M° x U, and from
this we deduce the smoothness of the transition maps between two open sets
U C B with a trivialization of 7r*£ above them.
If A is a real number, let Ua be the subset of B on which A is not an eigen-
value of (Dz)2. Let (P[1 0,A) I z E U),) be the family of orthogonal projectors
onto the spectral subspace [0, A) of (Dz)2, with respect to the L2-norm.
Proposition 9.10. (1) The sets U., form an open covering of B.
(2) The family P(°,,\) lies in r(Ua, K(£)); that is, the operators Pro Al form
a smooth family of smoothing operators on the bundles. (9 JAm. I'/' for
zEU,,.
280 9. The Index Bundle

(3) The vector spaces (9-l[o,Al)z = im(Pro a)) form a smooth finite-dimension-
al superbundle 7-l[o,al C over a.

Proof. As in the proof of Lemma 9.9, we may assume that the base is an
open ball U C R2, that M is a product Mo x U, and that the bundle E is
the pull-back of a superbundle Eo -- Mo. The fibre P(Mo, £o ® IAMO 11/2) of
the bundle 7r*E carries a metric which depends on z E U; however, by
Lemma 2.31, the Hilbert space completion rL2 (Mo, So ® 1AM,, 11/2) does not
depend on the metric. _

The family of heat kernels KZ = e-(° >2 at time t = 1 is a smooth family of

smoothing operators on r(Mo, £o (9 1AMo11/2) by Theorem 2.48, which extends
to a family of operators on the space rL2 (MO, £o (9 IAMo 11/2) of L2-sections,
self-adjoint with respect to the metric -)Z. The projector on the spectral
subspace [0, A) of (Dz)2 is equal to the projector on the spectral subspace
(e-,\,1] of KZ. If e-,\ is not an eigenvalue of KZO, there exists a neighbour-
hood U of zo such that for z E U, e-a is not an eigenvalue of KZ. This proves
that Ua is an open subset of B. The fact that Ua U), = B is clear.
Let C be the circle in the complex plane which crosses the real axis at the
points e-a and 1 + e. The spectral theorem for KZ shows that, for z E Ua,
the projector Pa is given by a contour integral

Pa = L(K- - u)-1 du.

Taking derivatives of this formula with respect to z, we obtain a formula for
8z Pa , as a sum of contour integrals of the form

Ic (KZ - u)-1 (8z Kz) ... (Kz - u)-1(8z kKZ) (KZ - u)-1 du.

The operator (KZ - u)-1 is bounded on the space of L2-sections of £o ®

lAM,11/2 over Mo (since boundedness is independent of the metric and
commutes with the generalized Laplacian (Dz)2. It follows from Corollary 2.40
that if Q is a smoothing operator, then both (KZ - u)-1Q and Q(KZ - u)-1
are smoothing operators. Since the operator 8z KZ has a smooth kernel, we
see that any derivative 8i Pa has smooth kernel, hence that the family Pa is
a smooth family of smoothing operators.
Corollary 9.11. Assume that ker(Dz) has constant dimension for z E B.
Then the vector spaces ker(Dz) form a smooth vector bundle ker(D) over B.
Proof. If ker(Dz) has constant dimension for z E B, then by the local com-
pactness of B, there exists for each relatively compact subset U C B a small
constant s > 0 such that the projector onto ker(Dz) coincides with P, ,z.
If dim(ker(Dz)) is constant, the index bundle ind(D) of the family D is
defined to be the superbundle ker(D). (This is also known as the analytic
index of the family.)
9.3. The Chern Character of the Index Bundle 281

If < Ec, let (P,\ µ) I z E U. fl UN,) be the family of orthogonal projectors

onto the spectral subspace (A, µ) of (Dz)2, so that

1'[0,/') = P[o,a) + P(,\,µ)

Since P(a,1) is a smooth family of projections, the family of vector spaces
?lea µ) = Pea,) forms a smooth superbundle over Ua f1 U. It is clear that
x[o,g) = f[O,A) ®x(A,µ).

9.3. The Chern Character of the Index Bundle

In this section, following Bismut, we will extend Quillen's theory of super-
connections to the infinite dimensional bundle *£ -+ B, thereby obtaining
a formula for the Chern character of ind(D) in terms of heat kernels. Our
treatment is modelled on the finite-dimensional case, which we discussed in
Section 1.
The space of sections of r*£ is defined by r(B, 7r*£) = r(M, £ (& JA, 11/2).

It is natural to define the space of differential forms on B with values in it*£

by
A(B, 7r*£) = r(M, 7r*(AT*B) ®£ (9 IA,11/2).
A differential operator on the space A(B, -7r*£) is by definition a differential
operator on r(M, r*(AT*B) ®£ (9 IA, 11/2). Let
A(B, D(£)) = r(B, AT*B (9 D(£))
be the space of vertical differential operators with differential form coefficients.
If a differential operator D on A(B,,7r*£) supercommutes with the action of
A(B), then this operator is given by the action of an element of A(B, D(£)).
Similarly, we write

A(B,1C(£)) = r(B, AT*B ®K(£))

for the space of smooth families of smoothing operators Kz with differential
form coefficients. Denote by dB the exterior differential on B.

Definition 9.12. Let D be a smooth family of Dirac operators on E. A

superconnection adapted to the family D is a differential operator A on
A(B, -7r*£) of odd parity such that
(1) (Leibniz's rule) for all v E A(B) and 0 E A(B, 7r*£),

A(u) = (dBV)O + (_1)Ivk VA(0);

(2) A = D + Edii(B) A[z), where A(ti) : A*(B,7r*£) -+ 7r.6).
282 9. The Index Bundle

It is easy to see that A[zl supercommutes with A(B) if i # 1, and hence

A[2) E Ai(B, D(£)) for i # 1.
Let us show how to construct a superconnection adapted to a family of
Dirac operators D. It suffices to define a connection V' on the bundle 7r*£,
that is, a differential operator from r(B,ir*£t) to A1(B,7r*£F) such that
V' (f 0) = df A 0 + fvr.E0
for all f E C°°(B) and o E r(B,7r*£). Let us assume that the bundle M/B
possesses the additional structure of a connection, that is, a choice of a split-
ting TM = THM®T(M/B), so that the subbundle THM is isomorphic to the
vector bundle lr*TB. From this, we can define a canonical linear connection
on the vertical tangent space T(M/B) using the projection operator
P:TM-*T(M/B)
with kernel the chosen horizontal tangent space THM. If X is a vector field
on the base B, denote by XM its horizontal lift on M, that is, the vector
field on M which is a section of THM and which projects to X under the
pushforward 7r* : (THM)X --+ T.,(.)B.
Furthermore, let us suppose that the bundle £ over M is provided with a
connection VE compatible with its Hermitian structure.

Proposition 9.13. Let a E r(M, IAMI1/2), 3 E r(B, JAB1-112) and sE

r(M, £). Define the action of V E, where X is a vector field on B, on
s®a®7r*13 E r(M,£ (& IA, 11/2) =r(B,7r*E)
by the formula
V E(s®a(&?r*,0) = (V Ms)®a®ir*,3+s®G(XM)a®ir*,Q+s®a®7r*G(X),8.
(1) This formula is independent of the tensor product decomposition of a sec-
tion s®a®7r*/3 E r(M, 9(9 IA, 11/2) with s E r(M, £), a E r(M, IAMM1/2)
and ,0 E r(B, JABS1/2), and satisfies the formula V fX = fVX for
f E CO°(B). Hence, V" is a connection on the bundle 7r*£ over B.
(2) The connection V'r,E is compatible with the inner product on it*£.

Proof. To show the independence of V '*e on the representation of a section

of 7r*£ as a tensor product s ® a ® it*/3, we must check that if f E COO (M),
then
V'"E((fs) ® a (9 7r*/3) = (s ® (f a) (9 7r* 0),
and that if h E COO(B), then
p'.E(s ® ((7r*h)a) (9 it*Q) = a ® it*(h(3))
These follow easily from Leibniz's rule.
9.3. The Chern Character of the Index Bundle 283

To show that is a connection, we must show that V71 = fVXe for

f E C°° (B). Using Lemma 1.14, we see that
V fx (s ®a ®7r*Q) =7r*.f (oXM s) ®a ®7r*,O
+s ® (7r*fG(XM)a+ 27r*X(f)a) ®7r*Q
+ s (&a(& (7r* f7r*,C(X))fl - 27r*X(f)7r*/Q)
=7r f [ (VXM s) ®a ®7r*N
+s®G(XM)a®7r*,O+s®a®7r*G(X),O].
We must now show that V'r*s preserves the inner-product on 7r*E. This
follows from the assumption that Ve is compatible with the Hermitian metric
on E, so that

,C(X) Is 0 a ®7r*0I2 =,C(X) M/B 2 a2,& 7r*)32

1,81

C(XM) (IsI2 a2 (9 7r*/02)

M/B
=2(s(9 a®7r*,O,OXE(s®a®7r*,O)).
Thus, associated to a connection on the fibre bundle M/B and a connection
on the bundle E, there is a natural superconnection D + V71-0 for the family
of Dirac operators (Dz I Z E B). In Chapter 10, we will also have to consider
superconnections A for which A[2[ does not vanish.
The curvature -'F = A2 E A(B, D(E)) of a superconnection is a vertical
differential operator with differential form coefficients. We have

where D2 is a smooth family of generalized Laplacians and

.F[+] E r(B, AT*B (9 D(E))
is a smooth family of differential operators with differential form coefficients
which raises exterior degree in
AT*ZB ®r(MM, e, (& IAM. I1/2).
The results of Appendix 1 apply in this situation, and we obtain the exis-
tence of a smooth family of heat kernels for F, which we denote by e-ty E
A(B,,AC(E)):
e-t.r = E-tD2 + E(-t)kjk,
k>O
where
-aotD2 [-o1tDa 'F[+] . E-ak-ItD2 .F[t] C-aktD2
Ik= fk e -)2 . .

Since Ik vanishes for k > dim(B), the above sum is finite.

284 9. The Index Bundle

We will show that the Chern character of the index bundle ind(D) can be
computed from a superconnection with A[o] = D, by the same formula as in
Str(e-A2).

the case of a finite dimensional superbundle, namely ch(ind(D)) =

Let K = (Kz I z E B) E A(B,)C(£)) be a smooth family of smoothing
operators with coefficients in A(B), given by a kernel
(x I K I y) E r(M x.r M,7r*AT*B 0 (6 (9 JA, 11/2) Z' (9* (9 JA, 11/2)).
Here M x,- M = { (x, y) E M x M I 7r(x) = 7r(y) } with projections pr1(x, y)
x and pr2(x, y) = y to M, and if E1 and £2 are two vector bundles on M, the
vector bundle £1 ®, £2 over M x, M is given by the formula
£1 ®.r £2 = pri E1 ® pre £2.

There is a supertrace on 1C(£Z) over each fibre MZ of M/B, which gives a

supertrace
Str : r(B, K(£)) --+ C' (B).
Suppose that K E A(B,1C(£)); when restricted to the diagonal, its kernel
(x I KZ I x) is a smooth section of the bundle 7r*AT*B®End(£)®lA,J over M,
and its pointwise supertrace Stre (x I KZ I x) is a section in r(M,7r*AT*B (g
JA, I). This section can be integrated over the fibres to obtain a differential
form on B.
Lemma 9.14. The A(B) -valued supertrace Str : A(B,1C(£)) H A(B) of the
family of operators K is the differential form on B

z --; S tre ((x KZ x))

fM,
If A is a superconnection on 7r*£, then [A, K] is again a family of smoothing
operators with differential form coefficients. The following lemma is one of
the main steps in the extension of Quillen's construction of Chern characters
to the infinite dimensional setting.
Lemma 9.15. dB Str(K) = Str([A, K]) E A(B)
Proof. As in the proof of Proposition 9.9, we may assume that the base B is
an open neighbourhood U C RP of 0 E RP, that the family M is the product
Mo x U, and that the bundle £ is the pull-back to M of the bundle £o --> Mo.
With respect to the identification A(B, 7r*£) = A(U) ®r(Mo, £o (a IAMo 11/2),
we may write
A=dU+EDIdzI,
I
where Di are differential operators on r(Mo, Eo(&IAM011/2) depending smooth-
ly on z E U; we also have

ZKIdz7,
9.3. The Chern Character of the Index Bundle 285

where KI are smoothing operators on r(Mo, £o(&IAM011/2) depending smooth-

ly on z E U. By Lemma3.49, we have Str([DI,Kj]) = 0, so that we may
assume that A = du in this trivialization. We see that
z
[du, K] _ z
dzi h dzz.
azi

On the other hand,

8
du Str(K) _ Str(KI) dzi A dzl
azi
8
_ ( Str (x I K12 x)) dzi A dzl
i
IJM0

= Str (8zi (x Kit x)) dzi A dzz

I,i
= Str([du, K]).

Definition 9.16. The Chern character form of a superconnection A on

the bundle -7r .,S, adapted to a family of Dirac operators D, is the differential
form on B given by the formula

ch(A) = Str(e-A2);

this is well-defined, since e_'2 = e-F e A(B,1C(6)).

Theorem 9.17. Let A be a superconnection on the bundle ir.£ for the family
of Dirac operators D.
(1) The differential form ch(A) is closed.
(2) If AQ is a one-parameter family of superconnection on the bundle 7r*E
for the family of Dirac operators D0., then

-dB Str (dd

do-

Thus, the class of ch(AQ) in de Rham cohomology is a homotopy invariant

of the superconnection A.

Proof. The fact that ch(A) is closed follows from Lemma 9.15, since

dB Str([A, e-]) = 0.
286 9. The Index Bundle

The proof of the homotopy invariance is similar. As in Proposition 1.41, we

argue that

Str Str ([A-, e--F° )

To-

Str([A,, d°e Y-1)

_ -dB Str(d °°e- °),
where all of the steps may be justified by using the fact that dAQ/do- is a
family of vertical differential operators with differential form coefficients.
The fact that the cohomology class of ch(A,) depends only on the homo-
topy class of Da follows as in the finite-dimensional case from the homotopy
invariance of ch(A): if A0 and Al are both superconnections for D, then
a E [0, 1] --> A, = aA1 + (1 - o-)Ao is a one-parameter family of supercon-
nections for D. Similarly, when DQ varies smoothly, we can form the smooth
family of superconnections AQ = DQ + W-6, and the cohomology class of
ch(A0.) will be independent of a.
Our goal in the remainder of this section is to prove the following funda-
mental theorem, which generalizes the results of Section 1 to the case of a
family of Dirac operators. Assume that D is a family of Dirac operators such
that ker(Dz) has constant dimension, so that ker(D) is a superbundle over B.
If Po is the orthogonal projection from 7r*£z to ker(Dz), then Po E r(B, K(£))
is a smooth family of smoothing operators. The following lemma is clear.
Lemma 9.18. The operator Vo defined by the formula Vo = POA[1]Po is a
connection on the superbundle ker(D).
For t > 0, let bt be the automorphism of A(B, 7r*£) which multiplies
A' (B, 7r*£) by t-i/2. Then At = t1/26t -A 8t 1 is a superconnection for the
family of Dirac operators t1/2D.
Theorem 9.19. For t > 0, let
1
At = t1/2bt - A A. 6t = t1/2D + A[1) + t-1/2A[2] + .. .
be the rescaled superconnection, with curvature.Ft = t6t(F). Then fort large,
-e- ° I It <_ C(2)t-1/2
uniformly on compact subsets of M x,r M, for all Ce-norms.
Proof. The proof is formally almost exactly the same as the proof in the finite-
dimensional case given in Section 1, and we give only a few steps. Consider the
algebra M = A(B, Endp(£)) = r(B,7r*AT*B®End.(£)), where Endp(ir*£)
is as in Definition 9.8; we filter M by the subspaces
Mi = 1: AA(B,End, (£)).
j>i
9.3. The Chern Character of the Index Bundle 287

We will also need the algebra N = A(B, )C(£)), filtered in the same way.
Let G = (GZ I z E B) be the family of Green operators GZ of (Dz )2.
Proposition 9.20. The action of left and right multiplication by G preserves
N.
Proof. Since this is a local question, we may suppose that the bundles M and
£ over B are trivial, by replacing B by an open subset U. We may rewrite
the integral representation of G given in (2.11) as follows, where PO is the
projection onto the kernel of D:
G= f e-t(°2+'°) dt -P0.

Thus, we may decompose all ((G + P0)K), where K E N, into terms propor-
tional to 00

f (aa18 t(D2+Po))(az2K)dt,

where al + a2 = a. Repeated application of Theorem2.48 shows that this

integral has the general form
00 e-ti (p2+Po)Die-t2(o2+Po)D2
... e-t', (D2+Po)Dk dt1... dtk,
J()Oc

where k jai I + 1 and Di E r(B, Endp (i*£)). This integral may be bounded
using the exponential decay of (x I e-t(o2+1'°) I y) proved in Proposition 2.37.
Using the formula
,9- (GK) = 8- ((G + Po)K) - 9-(PoK),
we see that GK is in N. The case where G multiplies on the right is simi-
lar.
Let P1 = 1 - PO be the projection onto im(D). If K E M, we will write

K= a PoKPo POKP1
E
N N
-y S P1KPo P1KP1 N M
Let
X
= IZ YI
T
be the curvature of the superconnection A; then X, Y and Z are in N.
Let R[21 be the curvature of the connection Vo on the bundle ker(D). By
the same proof as in Section 1, we have the formula R121 = X[2] - YY11GZ[11.
The space of endomorphisms g E M of the form g = 1 +K, where K E M
form a group, with
00
(1 + K)-1 = 1 + 1:(-K)k.
k=1
288 9. The Index Bundle

Lemma 9.21. There exists g E M with g - 1 E Nl, such that

X Y U 0
9 Z T 9-1 = 0 V
Furthermore
U - X - YGZ (mod N3),
V = T (mod N1).
The proof is the same as in the finite-dimensional case of Section 1, once
we observe that if G is the Green operator of D2, and if Y and Z E N, then
the operators YG and GZ are also in Nz, and Y(1- GT) and (1 - TG)Z are
in N+1. Thus, we may construct g as in Section 1 as a product of matrices
of the form
1 -YG
E1+
GZ 1

Thus, as before, we may put F into block diagonal form; there is a family
of operators K E Nl such that, with g = 1 + K,

.F =g-1 U 0
0 V 9,
where U - R[2] (mod N3) and V e D2 (mod N1). Note that the family of
operators St(V) is for each t > 0 the sum of a family of generalized Laplacians
D 2 and an element of P1 .M 1 P1. It follows from Appendix 2 that the family
of heat kernels e`61(v) is a section in A(B, K(£)) for each t > 0. Furthermore,
the uniqueness of solutions of the heat equation implies that
St(g)-1(e-t0
e-tst(,-) =
(U) e-t0(V))St(9)

If A(t) : (0, oo) -> F(B, )C(£)), we will write A(t) = O(f (t)) if for all e > 0
and 2 E N and each function 0 E CC° (B) of compact support, there is a
constant C(P, e, 0) such that
I l rr* (O) (x) x I A(t) I y) I l e < C(e, e, q) f (t) for all t > e.
Let U be a relatively compact open subset of B. If X is the infimum over
U of the lowest non-zero eigenvalue of the operators D2, then it follows from
Proposition 9.49 that over U,
Pie-t6t(v)Pl = O(e-tA/2)

Thus, we have
1R2
e O(t-1/2) 0
0 [ 6t(9) +St(9) 0 O(e-t,\/2
9.3. The Chern Character of the Index Bundle 289

Since St(g) - 1 = 0(t-1/2), we see that

e-tat(s) __ e- R[21 0 O(t-1/2) O(t-1/2)
0 0 O(t-112) O(t-1)
from which the theorem follows.
Corollary 9.22. The limit
lim ch(At) = ch(PoA[1jPo) E A(B)
t->oo
holds with respect to each Ce-norm on compact subsets of B.
The above corollary is equivalent to the McKean-Singer theorem when B
is a point. In that case, a family of Dirac operators is just a single Dirac
operator D on a Clifford bundle E -+ M and the superconnection A equals
D. The Chern character of the index bundle ind(D) is just a sophisticated
name for its dimension, which is the index of D in the usual sense. On the
other hand, we see that Str(e-t) is nothing but Str(e-tD2), which by the
McKean-Singer theorem equals the index of D.
We will now prove an infinite-dimensional version of the transgression for-
mula, which gives an explicit formula for a differential form whose differential
is the difference ch(At) - limt--+0. ch(At).
Theorem 9.23. Let D be a family of Dirac operators such that ker(Dz) has
constant dimension. The differential form

a(t) = Str (dd t e-At) E A(B)

satisfies the estimates
Cy)t_3/2
Ila(t)JIe <
on compact subsets of B, for all Ce-norms, and
00
ch(At) = lim ch(At) + d a(s) ds
t-+oo Jt
= ch(PoA[1]Po) + d it W a(s) ds.
t
Proof. It is clear that the formula for ch(At) - limt--+0. ch(At), known as the
transgression formula, is an immediate consequence of the estimate on a(t)
ast -+oc.
The proof of the estimate on a(t) is again much the same as that of Theo-
rem 9.7, except that the estimates are all rewritten in terms of the kernels of
the operators in question. Since
dAt 1 'E (1 - i)A[il
CF t-
D - i=2 2t(i+1)/2
290 9. The Index Bundle

and Alzj E M2, we see that if K(t) (0, oo) --+ r(B,1C(£)) is a family of
:

kernels such that K(t) = O(t-s), then

dA
tt K(t) - 2
D 2 K(t) =
O(t-s-3/2).

Furthermore, the proof of Theorem 9.19 showed that

e-tM-'r-') e-R[21 0 O(t-1/2) O(t-1/2)
_ +
0 0 0(t-1/2) O(t-1)
It follows that a(t) is the supertrace of
1 0 0 O(t-1/2) O(t-1/2) O(t-3/2) O(t-3/2)
+ O(t-3/2) _
2t1/2 0 D O(t-1/2) O(t-1) O(t-1) O(t-3/2)
Since only the diagonal blocks contribute to the supertrace, this proves the
theorem.
In the next chapter, following Bismut, we will construct a superconnection
A for the family D of Dirac operators associated to a Clifford connection, such
that ast ---+ 0,
a(t) = O(t-1/2).

From this will follow the existence of a limit limt_,o ch(At), which can in fact
be calculated explicitly, and the following analogue of (9.2):
00
lim ch(At) - ch(Vo) = dJ a(t) dt.
t->o o

Let R E r(B, .C(£)) be a smooth family of self-adjoint smoothing opera-

tors, such that each operator Rz is odd. Consider the perturbed family of
Dirac operators
(9.4) D + R E r(B, End.,(£)).
If we square D + R, we obtain
(D+R)2 = D2 +(DR+RD+R2).
Since D2 is a smooth family of generalized Laplacians and DR + RD + R2 is
a smooth family of smoothing operators, we obtain the following lemma.
Lemma 9.24. The square of the operator D + R has the form H = Ho + K,
where Ho is a smooth family of generalized Laplacians and K is a smooth
family of smoothing operators.
There is no difficulty in extending the results in Sections 2 and 3 to the more
general case where the family D of Dirac operators is everywhere replaced by
a family D + R of the form discussed above. In particular, the vector space
ker((D + R)z) is finite-dimensional for each z E B, the index of (D + R)z
ind((D + R)z) = dim(ker+((D + R)z)) - dim(ker-((D + R)z))
9.3. The Chern Character of the Index Bundle 291

is independent of z E B (and equals ind(Dz), by the homotopy invariance of

the index). As in Proposition 9.20, we can also prove that if the dimension
of ker((D + R)x) is constant for all z E B, then the Green operator GR =
(Gz I z E B) of D + R depends smoothly on z, in the sense that if X is a
family of smoothing operators, then XGR and GRX are families of smoothing
operators.
Note that a superconnection A adapted to D + R will have A[o] = D + R,
and hence cannot be a differential operator when considered as an operator
on F(M, it*AT*B ® £ (9 1A,11/2). Thus, we will allow a superconnection A for
D+R to be the sum of a differential operator on .A(B, it*£) = r(M, lr*AT*B®
£ (& IA, 1) and an element of A(B, Endp (E)). This does not cause any change
in the wording of the analogues of Theorem 9.17 and 9.19 for D + R, nor in
the estimates required in their proof.
First we have the analogue of Theorem 9.17.

Theorem 9.25. Let D+R be the sum of family of Dirac operators D and an
odd self-adjoint family of smoothing operators R. Let A be a superconnection
for the family D + R.
(1) The differential form ch(A) is closed.
(2) If AQ is a one-parameter family of superconnections on the bundle it*£,
then
ch(AQ) = -dB Str (ddo-°° e-Yo )
.-
Thus the class of ch(Ao.) in de Rham cohomology is a homotopy invariant
of the superconnection A.
(3) In particular, the class of ch(A) in deRham cohomology depends only on
the homotopy class of the Dirac operator D.

Next, we have the analogue of Theorem 9.19.

Theorem 9.26. Assume that ker((D + R)z) has constant dimension for z E
B.
(1) The vector spaces ker((D + R)z) form a smooth vector bundle ker(D + R)
over B.
(2) The limit
lim ch(At) = ch(ker(D + R)) E A(B)
t-3oo
holds with respect to each C'-norm on compact subsets of B.

We could also state the analogue of the transgression formula, but since
we will not make use of it, we leave its formulation to the reader.
292 9. The Index Bundle

9.4. The Equivariant Index and the Index Bundle

In this section, we will consider the special case of a family of Dirac operators
on a bundle M --+ B with compact structure group G. We will see that
the Chern character of the index bundle may be calculated using the Kirillov
formula (Theorem 8.2). This section is not needed to understand the rest of
this chapter or the next.
Let N be a compact Riemannian manifold with an action of a compact Lie
group G. Let E be an equivariant Hermitian Clifford bundle over N, with
equivariant Clifford connection DE and associated Dirac operator DN acting
on r(N, E). Then ker(DN) and ker(DN) are finite-dimensional G-modules;
we will denote the actions of G and of its Lie algebra g on ker(DN) by PN.
Let P ---> B be a principal bundle with structure group G, and let M =
P xG N be the associated fibre bundle, with base B and typical fibre N.
We also form the associated bundle £ = P xG E, which we consider to be a
bundle over M; thus, £ defines a family of bundles (£z -> Mz I z E B). A
section of 6 is a G-invariant section of the pull-back of E to P x N, that is a
section 0 E P(P x N, E) such that
0(pg, g-Lx) = g-1O(p, x) for p E P and x E N.
The operator DN acts fibrewise on P(P x N, E) and preserves the subspace
of G-invariant sections. Thus DN induces a family of Dirac operators D =
(Dz I z E B). It is clear that in this special case the vector space ker(DZ) has
constant dimension, and that the superbundle ker(D) over B is induced by
the 7L2-graded G-module ker(DN):
ker(D) = P xG ker(DN).
Using the Riemannian structure on N, we identify IAN 11/2 with the trivial
line bundle. Let 9 = Ei 9 Xi E A' (P) 0 g be a connection one-form on the
principal bundle P --+ B, with curvature two-form SZp E A2(P, g)bas. Since
-7r*£ is an associated vector bundle P xG P(N,E) with infinite dimensional
fibre, we deduce from the connection 9 on P a natural connection on the
bundle B. If £E(Xi) is the action of the Lie algebra element Xi E g
on r(N, E), we see that
va.E = dp + E 0iCE(Xi)
i
on G-invariant sections.
The following lemma follows from the obvious fact that V'-'c preserves
the bundle ker(D), which itself follows from the fact that GE(Xi) preserves
ker(DN).

Lemma 9.27. The projected connection Do on ker(D) may be identified with

the connection on P xG ker(DN) associated to the connection one-form 6.
9.4. The Equivariant Index and the Index Bundle 293

Denote by ch(ker(D)) the Chern character of ker(D) with respect to the

connection Vo. It follows from this lemma that the Chern character form
ch(Vo) is given in terms of the representation PN by the formula
ch(Vo) = Str(e PN(n")).
Denote by Ig the G-invariant analytic function on g given by the formula
I0(X) = indG(e X, DN).
The equivariant Chern-Weil homomorphism 0e maps the equivariant de Rham
complex (Ac (N), de) to the deRham complex (A(M), d). On C[g]G, it coin-
cides with the ordinary Chern-Weil homomorphism, that is, a H a(cp), and
maps into the space of basic forms on P, which may be identified with forms
on B. It follows that
Oo(Ig) = Str(e-PN(n')) = ch(ker(D))

The vertical tangent bundle T(M/B) of M may be identified with the

associated bundle PxGTN. It has a connectionVN,e, obtained by combining
the Levi-Civita connection on TN with the connection on the principal bundle
P as in Section 7.6. Let RM/B be the curvature of the connection VN,e
Denote by A(M/B) the A-genus of the connection VN,e,
RM/B/2
A(M/B) = det1/2
(sjnh(RM/B/2))
To simplify the discussion, we will assume that N has a spin structure
with spinor bundle S(N), and that E = W ® S(N) for some G-equivariant
Hermitian twisting superbundle W with connection VW. Let W = P xG W
be the corresponding Hermitian superbundle over M. Given a connection on
W, we obtain a connection V"V on W; denote by ch(W) the Chern character
form of the connection Vw:
ch(W) = Str(exp(-F ')).
The differential form A(M/B) ch(W) is a closed differential form on M,
whose integral over the fibres of M/B is a closed differential form on B.

Proposition 9.28. The Chern character of the connection Do on the index

bundle ker(D) is given by the formula

ch(Vo) = (27ri)-dim(N)/2 /' A(M/B) ch(W).

J M/B
Proof. The proof is based on the Kirillov formula of Chapter 8:

10(X) = (27ri) - dim (N)/2 L Ag(X,N) ch8(X,W),

294 9. The Index Bundle

where Ag (X, N) and the (X, W) are equivariant characteristic classes given
by the formulas

Rg (X)/2
Ag (X, N) = det1/2
sinh(Rg (X)/2)
chg(X,W) = Str(exp(-Fg (X))).
By Proposition 7.38 (2), we see that the Chern-Weil map 4a : Ac (N) --r
A(M) sends these equivariant characteristic forms to the corresponding char-
acteristic forms of the associated bundles:

ge(Ag(N)) = A(MlB),
Oe(ch9(W))= ch(W)
The result follows from the formula fMJB 000 = 00 o fN

We will see in Chapter 10 that the above formula for ch(ker(D)) holds in
cohomology for any family of Dirac operators. We will also show that the
local version of the family index theorem implies the local Kirillov formula.

9.5. The Case of Varying Dimension

In this section, following Atiyah and Singer, we will show that the index bundle
of a family of Dirac operators over a compact base B may be represented as
the formal difference of two vector bundles, even without the assumption that
ker(b) has constant dimension.
Let us start by recalling the definition of a difference bundle. This is
an equivalence class of superbundles on a manifold B, with respect to the
equivalence relation that £ - F if there are vector bundles 9 and rl such that

E) 9 (D R;
£-®9=F-®7-l.
The class of £ is denoted by [£+] - [£-] and is called the difference bundle
of £. If g is a vector bundle, we will associate to it the difference bundle
191-101.
If B is compact, the difference bundles forms an abelian ring K(B) called
the K-theory of B. The sum in this ring is induced by taking the direct sum,
and the product by taking the tensor product, of superbundle representatives.
It is easily verified that the additive identity is the zero bundle, the multiplica-
tive identity is the class of the trivial line bundle, which we will denote by [C],
and the negative of [£+] - [£-] is [£-] - [£+]. We will not give any further re-
sults about the ring K(B): as explained in the introduction, our emphasis in
this book is on explicit representatives, by superbundles or differential forms,
9.5. The Case of Varying Dimension 295

rather than the classes that they represent in K-theory K(B) or de Rham co-
homology H*(B). We do note, however, that the Chern character defines a
homomorphism of rings from K(B) to the de Rham cohomology H* (B) of B;
this follows from the fact that the Chern character vanishes on superbundles
of the form 9 ®!9 In fact, this map may be shown to induce an isomorphism
of K(B) ® C with the even de Rham cohomology, by the Atiyah-Hirzebruch
spectral sequence.
If ker(D) has constant dimension, we have defined the index bundle to be
the difference bundle represented by ker(D). Since M is compact, there is a
positive e such that UA = M for 0 < A < E. It follows that ind(D) is also
represented by R[o,a) for 0 < A < e. Motivated by this, we would like to define
ind(D) in such a way that when restricted to Ua, it has the superbundle H[o,a)
as representative. The following lemma shows that this approach is consistent
on UaflU,,,,for 0<A<M.
Lemma 9.29. If 0 < A < µ, then the difference bundles represented by f[o,a)
and 'H[o,m) are equal.
Proof. By the isomorphism
x[o,µ) = 'Hjo,a) ®x(a,p),
we see that we must prove that on Ua fl UN,,

x(A,µ) - x(A,1A)
On the bundle 7-la µ), the operator (D+)*D+ = D-D+ has spectrum lying in
the interval (A, µ), and hence is invertible (an explicit inverse is given by the
Green operator G). It follows that the operator

induces the desired isomorphism of bundles.

We will show that if B is compact, there is a difference bundle of the form
[E] - [CN] which when restricted to U,, is equivalent to the difference bundle
7-l[o,a). This will show that ind(D) may be considered as an element of K(B).
We will also generalize Theorem 9.19 to give a formula for the Chern character
of ind(D) in the general case.
We make use of the simple observation that if D is a family of Dirac oper-
ators such that ker(D-) vanishes, then
dim(ker(Dz)) = ind(Dz),
and is thus independent of z. Thus, one way to perturb a family so that
its kernel ker(D) becomes a vector bundle is to modify it in such a way as
to ensure that D+ is surjective for each z, and hence that D- has vanishing
kernel. This is not difficult if we permit ourselves to perturb Dz by a family
of smoothing operators.
296 9. The Index Bundle

We introduce the trick of replacing M by an enlarged version M' = MUB,

obtained by adding one point pz to each fibre Mz of M -> B. Similarly, we
replace £ by the bundle £', which coincides with E over M and whose fibre
over the point pz is a constant vector space V with V+ = GN and V- = 0.
Note that it E' is isomorphic to r .E ® (B X CN), so that
r(B, 7r*£') - r(B, it*£) ® C°°(B, CN).
The family D of Dirac operators corresponds in a natural way to a family of
Dirac operators on the family of vector bundles £' -+ M'.
Let Eli : GN --+ r(B, it*£-) be a linear map from C' to r(B, Let Y)i
be the sections of .7r*£- corresponding to the basis vector ei E CN. We define
a family of smoothing operators Rp E r(B, JC(£')) on the vector bundles
£' --> M' by the formulas
R05 _ E ei ('t/li, O), for O E r(B, rr £) C r(B, 7r*£'),
R,pei = y)i
It is easy to see that R,p is odd and symmetric. We define Dp to equal D+R,p,
given explicitly by the formula
N N
D,+(0®uiei) =D+O +u4i,
i=1 i=1
N
Dip (0) = D (0i, 0) ei,
i=1

for 0 E r(B, 7r£+) and r(B, zr*£-) respectively, and ui E CO°(B). We thus
have a family (D,+ )z of the type described in (9.4).
In order for ker(D;) to vanish, we must demand that D,+ is surjective for
all z E B. The following lemma shows how to choose suitable i,b .
Lemma 9.30. There exists an integer N and a map z/' : CN -+ r(B, it*£-)
such that the map is surjective at each point in z E B.
Proof. By Proposition 3.48, we know that
ker(Dz) ® Dz [r(Mz7£z+ (9 IA,,1112)] = r(MM,£z (9 1A,, 11,2)
at each point z E B. We need to construct a collection ,b of sections of 7r*£-
such that at each point z E B, the projection onto the vector space ker(Dz)
of the sections zk2 span it. We proceed as follows.
Given zo E B, we can find a A > 0 such that zo E U.., a ball U(zo) C B
around zo and sections tfi E r(U(zo), such that for every z E U(zo) the
elements Of form a basis of (710 A))z. If X E C'°(U(zo)) is a cut-off function
on U(zo) equal to 1 around zo, the sections x(z) 0z can be extended to smooth
sections of the bundle it on B. Thus, for each zo E B, there exists a A > 0,
an open set V(zo) C U(zo) and sections 'Vi E r(B,-7r £) such that Of span
9.5. The Case of Varying Dimension 297

(?t[o a))z, and hence their projections onto the vector space ker((D-)z) span
this space, for z E V(zo). Since B is compact, it is covered by a finite number
of sets V(zo), and we may take for (V)i) the union of the sections O for just
these sets.
If A is a positive number, let U,\,,, be the open subset of Ua consisting of
those points z E U), such that the composition of
_ P(O,,))
(CN 0 , (ir*s)z (f[o,a))z
is surjective. If Al is the smallest non-zero eigenvalue of (DZO)2, then zo is in
U,,O if A < A A. Thus, B is covered by the open sets {Ua,,, I A E R}.
Proposition 9.31. (1) If 0 satisfies the conditions of Lemma9.30, we have
the short exact sequence of vector bundles on Ua ,
P10, X)
0 --+ ker(DV,) 'H+
to a) ® (B X CN) P(O, )D ?-l[a ,
a)
--40.

Thus, over Ua, the difference bundles [ker(D.p)] - [CN] and H[o,a) are
equivalent.
(2) The difference bundle [ker(Dp)] - [CN] of K(B) is independent of the
choice of map : CN -+ I'(B,-7r. -) satisfying Lemma9.30.
Proof. Let us show that the bundle map
P[o,a) E) 1 : ker(Dp) C 7r.£+ (D (B x CN) -> 7-l[o a) ® (B X CN)
is injective. Indeed, if (0, u) E ker(Dp) lies in the kernel of this map, then we
see that u = 0, and hence that 0 E ker(D). However, the projection P[o,A) is
the identity on ker(D) if A > 0, and hence 0 = 0.
The bundle map
P[o,a)D.O : x[o a) ® (B X CN) .- R- a)
is surjective by the definition of U.>,.O.
The composition of the two maps
Pl°"\)61 P1°')D
ker(D1) C (B X CN) 7'{[o a) ®(B x CN) [o,A)
is the bundle map from ker(D,,) to which sends (0, u) E ker(D,,) to

P[o,A) (DP[o,A)0) + P[o,A)')(u) = P[o,A) (Do +V)(u)) = P[o,a) Do (0, u),

which vanishes, since (0, u) E ker(D,,). Thus, we see that
P[o e1' CN) P[0'T)D$
0 ---+ ker(D,p) x[o A) ® (B X No "\) 0

is a complex. Let us show that it is a short exact sequence. If ¢ E (7-l[o a)) z

and u E CN are such that P[o A)(Dzcb + i,bz(u)) = 0, then there exists an
298 9. The Index Bundle

a E (7r*£+)z such that Pro A)a = 0 and (D+)zO +'iI?(u) = (D+)za. Thus
- a, u) E ker(D ), and its image by P[o,al is our element (0, u).
It remains to prove the independence of this construction from the map
: CN -- r(B, 7r*£-). Consider another map i : CN -- r(B, -7r*£-). The
bundle ker(D ,) may be identified with a subbundle of ker(D e,) by sending
(0, u) E (7r*£+) z ® CN to u, 0) E (ir*£+) CN ® CN. Consider the
sequence of bundles
0 --> ker(Dp) --3 ker(Dpe.) --> (B x CN) -: 0,
where the last arrow maps (0, u, u) E ker(DV,®,) to u E CN. This sequence is
exact at each point z E B. To see this, choose u E CN. The surjectivity of D
implies that there exists u E CN and 0 E it*£z such that '(u) = Do+,O(u).
Using this proposition, we may define the index bundle of any family of
Dirac operators on a compact base.
Definition 9.32. The index bundle of a smooth family of Dirac operators
over a compact base B is the virtual bundle
ind(D) = [ker(D,,)] - [CN],
where is chosen as in Proposition 9.31.
We can now prove a generalization of the McKean-Singer formula to su-
perconnections associated to families of Dirac operators.
Theorem 9.33. Let A be a superconnection on the bundle 7r*£ for the family
D. Then the cohomology class of ch(A) is equal to the Chern character of the
index bundle ind(D).
Proof. Unlike in the case where ker(D) had constant dimension, the differen-
tial form Str(e-t6t(A2)) may have no limit as t -+ oo when the dimension of
ker(D) is not constant. However, we may replace the family D over M -+ B
by a family Dp over M' - B such that ker(D,,) is a vector bundle.
Let A be a superconnection for the family D, and transfer the operators
A[ j] to operators on A(B, 7r*£'). Let AE denote the superconnection A + eR,y
for the family D + ER,,. Applying Theorem 9.25 to the one-parameter family
AE, we see that the cohomology class of the differential form
Str(e-2)

ch(AE) =
does not depend on e. For e = 1, we know by Theorem 9.26 that ch(AE)
represents the cohomology class of ch(ker(D.,,)), the Chern character of the
vector bundle ker(D,1). Taking e = 0, we see that
ch(Ao) = ch(A) + N E A(B).
It follows that ch(A) and ch(ker(D,,)) - N represent the same class in coho-
mology, from which the theorem follows.
9.6. The Zeta-Function of a Laplacian 299

9.6. The Zeta-Function of a Laplacian

In this section, we give an application of the existence of the asymptotic
expansion for the heat kernel: the construction of the zeta-function of a gen-
eralized Laplacian and the definition of its renormalized determinant. This
construction, applied to a family of Dirac operators, is used in the theory of
determinant line bundles, which we will describe in the next section. The
reader only interested in the family index theorem may omit the rest of this
chapter.
If f E C°°(0, oo) is a function on the positive real line (well-behaved at 0
and oo in a sense which we will describe shortly) the Mellin transform of
f is the function
M[f](s) = r(s) f f(t) is-1 dt-
Lemma 9.35. Let f E C°°(0, oo) be a function with asymptotic expansion
for small t of the form

f (t) ti E fktk/2 '+' g log t,

k>-n
and which decays exponentially at infinity, that is, for some A > 0, and t
sufficiently large,
If (t) I < Ce-t\.
Let -y be the Euler constant. Then
(1) the Mellin transform M[f] is a meromorphic function with poles contained
in the set n/2 - N/2;
(2) the Laurent series of M[f] around s = 0 is -gs-1 + (fo - yg) +O(s).
Let h E C°° (0, oo) be a function with asymptotic expansion for small t of
the form
h(t) r.J hk tk/2'
k>-n
and such that for some A > 0, and t large,
Ce-ta
I h(t) I <

Then f (t) = f °° h(s) ds satisfies the above assumptions.

Proof. Choose Re s large, and split the integration over [0, oo) into the inter-
vals [0,1] and (1, oo):

r(s) M [f] (s) = f

0
1 f (t) to-1 dt + f f (t)
1
is-' dt.
300 9. The Index Bundle

The second term, the contribution of the integral at infinity, is clearly entire,
by the exponential decay of f (t) as t --> oo. The first term may be expanded
in an asymptotic series, and then integrated explicitly, term by term:
1
Jf1 f(t) is-1 dt _ fk tk/2+s-1 dt
° k<K fo

+gJ I (logt)ts-1dt+J O(tK/2+s-1)dt

0 0

+ r(s),
k<K s + k/2 s2

where r(s) is holomorphic in the right-half plane Re s > -K/2. Since the
inverse of the Gamma function F(s)-1 is entire, (1) follows.
Since P(s)-1 = s + ys2 + 0(s3), it follows that M[f] has the Taylor series
around s = 0
-gs-1 + (fo -'Yg) + O(s),
proving part (2) of the lemma.
If h is as in the statement of the lemma, then to show that f °° h(s) ds has
an asymptotic expansion of the required form, we write, for t small,

At) =
J
1 E I hksk/2J ds - J t\(h(s) - E hksk/2 1 ds
k<0 k<0
r
+
J0
1(h(s) - E hksk/2) ds + h(s) ds
k<0 J
from which we obtain

f (t) - -
kO-2
E k/2 + 1
tk/2+1- h_2 log t + c,

where c is the constant

E
c = {k<0JkO-2} k/2 + l
+
J l (h(s)
- hksk/2) ds +
J 00 h(s) ds. El
k<0
If we define f (t) to equal fo-yg, where fo and g are the coefficients
of to and log t in the asymptotic expansion of f, we see that part (2) of the
above lemma implies the formula
d
s_0sM[f] = t I f(t),
ds M
We will call LI Mt.-.o f (t) the renormalized limit of f (t) as t ---r 0.
9.6. The Zeta-Function of a Laplacian 301

If H is a self-adjoint generalized Laplacian on a compact manifold M and

A is a positive real number, the zeta-function of H is the Mellin transform

((s, H, A) = M [Tr(P(a,00)e-tH)] = f Tr(P(a

00)e-tH)ts-1

dt,
r(s)
where P(a,.... ) is the spectral projection associated to H onto the eigenvalues
lying in (A, oo). This function equals Tr(Pla )H-s} for s > n/2, where n is
the dimension of the manifold M.
As an example, consider the scalar Laplacian -d2/dt2 on the circle, with
eigenvalues {0,1,1, 4, 4, ... , n2, n2, ... }. If 0 < A < 1 and s > 1/2, we see
that
00
00
((s, H, ) = I ( J e-tn2ts-1 dt = 2 1: n-23 = 2((2s),
n=1 0 n=1
where ((s) is the Riemann zeta function.
We can understand the behaviour of the zeta-function as a function on the
complex plane by means of the Minakshisundaram-Pleijel asymptotic expan-
sion of the heat kernel.
Proposition 9.35. The function ((s, H, A) possesses a meromorphic exten-
sion to the whole complex plane, and is holomorphic at s = 0.
Proof. As t --+ 0, we have
m
Tr(P(a .)e-tH) = Tr(etH) _ E e-tXk
k=0
00
.., E tkak,
k=-n/2
where A0 < ... < Am < A are the eigenvalues of H lying in (-00, .\], enu-
merated according to multiplicity. The fact that Tr(P(a,00)e-tH) decays ex-
ponentially fast as t -> oo has been proved in Lemma 2.37.
Denote by C'(s, H, A) the derivative of C(s, H, A) with respect to s. Follow-
ing Ray and Singer, we define the zeta-function determinant of a gener-
alized Laplacian H for which 0 is not an eigenvalue to be
det(H) = e s (o,H,o)
This definition is motivated by the fact that if H is a positive endomorphism
of a finite dimensional Hermitian vector space V, and
C(s, H) = M [Tr(e-tH)] = Tr(H-s)

then ('(0, H) log det(H). The zeta-function determinant of a generalized

Laplacian H is a renormalized determinant adapted to the class of generalized
Laplacians.
302 9. The Index Bundle

If H is a self-adjoint Laplacian, let 0 < A < p, and let A < al < .. <
be an enumeration of the eigenvalues of H lying in the interval (A, µ].
The following result is clear.
Proposition 9.36. The zeta-function of H satisfies the formula
m

and hence its derivative at s = 0 satisfies the formula

m
('(0, H, A) = S' (0, H, IL) - .1 log ati.
In the following proposition, we calculate the variation of the zeta-function
S(s, HZ, A) of a family HZ of positive generalized Laplacians on M with res-
pect to the parameter z. We work on the set U. where A is not an eigenvalue
of HZ.

Proposition 9.37. Over the set UA, the derivative with respect to the pa-
rameter z of the zeta-function of a family of generalized Laplacians is given
by the formula
HZ'.\)
a((s' = -sM [Tr(P(a oo)&ZHZ(HZ)-1e-tH1
az
Proof. Using the method of Lemma 9.34, it is not hard to show that the deriva-
tive aC(s, HZ, A)/az of the zeta-function is meromorphic. Thus, we may argue
at large Re s, where both sides of the formula are given by convergent inte-
grals. If we differentiate the expression

S(s, HZ, A) - r(s) f 00 Tr \Pia,0)e-tH'l is-1 dt

for the zeta-function with respect to z and apply Corollary 2.50, we obtain
that
(9.5)

Hz,
A) _ -M [tTr(Pa )aZHZe-thI)]
+ M [Tr r)
Since P = P(\ 00) is a projection, Leibniz's rule shows that
aZP = PaZP + aZPP;
multiplying on the left and right by P, we see that P0ZPP = 0, so that
8ZP = P(aZP)(1- P) + (1 - P)(8ZP)P.
Thus the second term on the right-hand side of (9.5) vanishes. On the other
hand, integration by parts shows that in general,
9.6. The Zeta-Function of a Laplacian 303

M[tf'] _ -sM[f],
so that the first term can be rewritten in the desired form.

The following theorem is the analogue for the zeta-function determinant of

the formula

log det(HZ) = Tr(azHz(Hz)-1)

az
for the variation of the determinant of a family of operators Hz on a finite-
dimensional vector space.

Proposition 9.38. If Hz is a family of generalized Laplacians on a compact

manifold M, then over the set Ua,

a
0, Hz ),) LIM
Tr(P(\,.)azHz(Hz)-le-tH=)

io
Proof. To apply Lemma 9.34 (2), we must show that

(P00)azHz(Hz)-ie-tH=)
Tr

has an asymptotic expansion. Consider the asymptotic expansion

00

lP(a
oo)e-sH=)
1: ak(z) Sk
k=-n/2
Recalling from the proof of the previous proposition that

ax (P(
00)e-eH=)
= -s Tr(P(a 00)azHze sH=),

we see that Tr(P(a )azHze-eH=) has an asymptotic expansion of the form

00

Tr(P(a
00)azHze-sH=) ,., _ 8zak Sk-1.
k=-n/2

Thus, we can apply Lemma 9.34 to conclude that

oo
oo)azHze-aH=) ds =Tr(P(a oo)azHz(Hz)-le-tH=)
ft (P(a

has an asymptotic expansion.

304 9. The Index Bundle

9.7. The Determinant Line Bundle

Using the results of the last section, we will now define the determinant line
bundle associated to a family of Dirac operators, its Quillen metric, and its
natural connection. It is helpful, however, to see first how the theory looks in
finite dimensions.
Let 7-l = 7.1+ ®7{- be a finite-dimensional superbundle over a manifold B,
such that that dim(7-l+) = dim(7-1-) = m. The map
AmD+ : Amf+ , A-R
is a smooth section of the determinant line bundle
det(7i) = (A-H+)-1 ®AmH-,
called the determinant of D+, which we denote by det(D+). The section
det(D+) vanishes at precisely those points where D+ is not invertible. In
particular, if D+ is everywhere invertible, the section det(D+) defines a triv-
ialization of the line bundle det(7-l), since it is nowhere vanishing.
Let V be a connection on the bundle f which preserves the decomposition
7-l = 7-l+ ® 7l-; it induces connections Vdet(71) and VEnd(w) on the bundles
det(H) and End(7-l).
Proposition 9.39. The section det(D+) satisfies the differential equation
Vdet(rc) det(D+) = -det(D+) TrH+((D+)-1VEnd(x)D+)
Proof. We may assume that B is an open set on which 7i has been trivialized,
7{t = U x Ht, so that V = d + w, where w = w+ ® w-. In this frame, the
connection Vdet(rc) is given by the formula d - Str(w). If we replace D+ by a
small perturbation D++ESD+, where 8D E r(B, Hom(7-1+,7l-)), we see that
det(D+ + eBD+) = det(D+(1 + E(D+)-18D+))
= det(D+) det(1 +E(D+)-18D+))
= det(D+) + Edet(D+) Tr((D+)-18D+) + O(E2),
where we use the fact that
d
det(1 + sA) = Tr(A).
de E=0

From this, applying the sign rule for graded tensor products, we see that
Odetr' det(D+) _ -det(D+)Trn+((D+)-1dD+) - Str(w) det(D+).
On the other hand,
Tr,,+ ((D+)-1 VEnd(%) ((D+)-1(dD+ + w-D+ + D+w+))
D+) = Tr.,+
= Tr7j+ ((D+)-1dD+) + Str(w).
9.7. The Determinant Line Bundle 305

We will now show how to define a determinant line bundle using a family
of Dirac operators D parametrized by a base B. Recall that if £ and F are
two vector bundles, then

det(£ ®.T) = det(£) ® det(F).

Let 0 < A < µ be a pair of positive real numbers, and consider the decompo-
sition
x[o,µ) = 7"l[o,a) ® x(a,l+)
over the open set Ua n UN,. From this, we see that

det(x[o,,,)) = det(7-I[o,a)) (9 det(x(a,. )).

Denote by D(a,,,) the Dirac operator restricted to Ho,,,,). The line bundle
det(H(a,N)) over Ua n UN, is trivial, since it has a nowhere-vanishing section
det(D(A )). Thus, the line bundles det(7-l[o,a)) piece together to form a line
bundle over all of B, called the determinant line bundle associated to
the family of Dirac operators D and denoted by det(7r*£, D); to a section s E
F(UanUU, det(7l[o,a))), we associate the section s®det(D µ)) of det(7l[o µ)).
We have assumed that the vector bundle £ has a Hermitian metric, and
hence that 7r*£ has a metric too, given on sections sz and tz E 7r*£z =
r(M., £® (9 1A Iz,2) by

(sz(x),tz(x))e..
LM.
Ma

We would like to define a natural metric on the line bundle det(7r*£, D). From
the metric on 7r*£, we obtain a natural metric on each of the vector bundles
7-l[o,a), which is known as the L2-metric. It follows that each of the line
bundles det(7-l[o,a)) has a natural metric, which it inherits from the metric on
R[o,a) in the natural way. However, the metrics on the bundles det(7-l[o,a))
and det(N(o,, )) over Ua n Uµ are not equal, and they differ by a factor equal
to
det (D+ A))

where the metric is the L2-metric on the trivial line bundle det(7l(a,µ)).
The family of generalized Laplacians D- D+ on 7r*£+ consists of self-adjoint
operators, since (D+)* = D-. The following lemma is clear.

Lemma 9.40. If A < Al < < Am <,u are the eigenvalues of D- D+ lying
between A and p, then
m
11A/2'
Idet(D(A,/L))I
=
i=1
306 9. The Index Bundle

By choosing local trivializations of the bundles M and E over B, we see

that the zeta-function of the family D- D+ is a meromorphic function of s with
smooth dependence on z E B. By Proposition 9.36, its derivative at s = 0
satisfies the formula
m
('(0, D-D+, A) = S'(0, D - - log A .
2=1

From this, we easily see that the metric

IQ = e-s (0,D D+,a)/21

on the line bundle det(f[o,a)) agrees with the corresponding metric on the
line bundle det(7-L[o,,,)), and hence all of these metrics patch together to form
a metric on the line bundle det(ir*E, D), which is the Quillen metric.
If the family D has index zero, then dim(7-1 a)) for all
A > 0, and so that it makes sense to speak of the section a)) of the
determinant line bundle det(1-1[o,,)) over the set Ua. Since

det(D+o, )) = det(D+o a)) 0 det(D+a µ)) E I'(Ua fl U.,

we see that these sections patch together to give a section det(D+) of the line
bundle det(ir*E, D).
Proposition 9.41. Let D be a family of Dirac operators of index zero on a
vector bundle E over the family of manifolds M ---> B. There is a canonical
smooth section det(D+) of the determinant line bundle det(7r*£, D), which
vanishes precisely where D is not invertible. The section det(D+) satisfies
the formula
1 det(D+) IQ = det(D-D+)i/2
where by det(D-D+) we mean the zeta function determinant e-C' (0,o-D+,o)
of the symmetric Laplacian D- D+. (This is the analogue for Dirac operators
of the formula det(A*A) = I det(A) 12 in finite dimensions.)
We will now define a natural connection on the determinant line bundle
det(7r*E, D), compatible with the Quillen metric. To do this, we need a unitary
connection on the bundle 7r*£, for example, the one defined in Section 2
from connections on the bundle M/B and on the Hermitian vector bundle S.
From the connection 0(0,x) = P[o,a)V .eP(o,a) on 1-l[o,a), we obtain a con-
nection on det(7-l[o,a)) compatible with the L2-metric, which we denote by
Qdet('Kja, ;,)) .
The operator D(a,,,.) = P(a,,,)D over the open set U,, has the property
that the bundle ker(D(a,00)) equals 7lto,a). We may define a superconnection
associated to the operator D(a,,,) by the formula
Aa = D(a,.) + V
9.7. The Determinant Line Bundle 307

and its rescaled version

A,\,, = s112D(a 00) + V".e
Define two differential forms a+ (s, A) E A(B), given by traces over 7r*£+ and
it*£- respectively:
3e-Aa,e)
at(s, A) = `h'7l.ef (a8s Tr".e (D(,`,OO)e-A'`,s).

2s1/2

Lemma 9.42. The one-form components of the differential forms a+(s, A)

and a- (s, A) satisfy
a+(s, A)[1] = a- (S' A)[1],
and have asymptotic expansions for small s of the form

a+(s, A)[ll ti sk/2 a+

k=-N
Proof To see that a+ (s, A)[1] = a- (s, A)[1], we use the facts that D* = D and
that V1--c respects the metric on 7r*£.
The one-form component of a+(s, A) is equal to

a±(s, A)[11 = 2s1/2 )[11

)]e-9oz
_ - i r of D(a

_ -2 Tr'.ef ((1 -
2
D±]e-so ),

since P(a,,,) (V"weP(A )) P(A,.) = 0. The result now follows from Propo-
sitions 2.46 and 2.47, which give asymptotic expansions for Tr(DPt) and
Tr(KPt) at small t respectively, where Pt is a heat kernel, D is a differential
operator, and K is a smoothing operator.
By Lemma 9.34, we conclude that the functions f °O aE (s, A)111 ds have
asymptotic expansions for t small. Let
00
Qa = 2 LIM a:' (s, A)111 ds E A'(B).
t-0 Jt
The sum i3a +)3- is real, while the difference as - ,Qa is imaginary.
Lemma 9.43. d('(0, D- D+, A) _ - (i3t + Qa )
Proof. Using the formula ff'O a aH ds = H-le-tH, we see that
00
,3 = - LIM0 TI E+ (P(A ) D-
D+]e-sot)
ds
t /
D+](D-D+)-le-to-o+)
OO)D-[V"'e,
tIM T ,.e+(P(a
308 9. The Index Bundle

and

D-]e se2) ds
i3 = - LIM
t--o jt,:,o
Tr'1".s-

= LIM J'OoTr,.e+(P(a,o) [V"-E, D ]D+e-sD2) ds

t-0
LIMTr,.e+(P(a,oo)[V".e,D-]D+(D-D+)-le-to-o+)
=
t--.o

Adding the one-forms Q+ and j3a together, we obtain

[p".E, D D+] (D D+) -le-to-o+)

Qa + Qa = L I M
o D,.,,+

i
The lemma follows immediately from Proposition 9.38.
We can now define the connection on det(7r,,E, D). On det(H[o,),)), we take
the connection
Odet(?- [o, \)) +)3' .

Lemma 9.44. The section det(D+a,,)) is parallel with respect to the connec-
tion Odet(7{(a,µ)) + (i9 - Qa) on det(f(a,u,)).

Proof. Observe that

,µ))-1
(D( D(a,w) _ DU,µ)
By Proposition 9.39 applied to the bundle 7 l = 7{(a ,), we see that
Vdet(rt(a,,.))
(\,µ) (det(D(a,, )) (D(a,,,))-1 D(a,µ) )
_ -det(D(a,w))(Qµ - +)
The connection Vdet(%[o,j.)) on det(f[o,N,)) is the tensor product of the
connections and Vdet(rc(,,,,,)):
Vdet(?{[0,,)) = vdet(7{(o,,,)) ®1 + 1® Odet(%(,,,M))
Lemma 9.44 now shows that the connections
Vdet(f[o,a)) + ' and Vdet(?{[o,v)) +,Qµ

on the line bundles det(7I[o,a)) and det(f(o,µ)) agree over the set Ua fl Uµ,
when we identify these two bundles by multiplying by the section det(D+a w))
of det(7l(a µ)). Denote the resulting connection on det(a*E, D) by Vdet(r.E,D)

Proposition 9.45. The connection Vdet(,r.e,°) is compatible with the Quillen

metric I IQ on D).
Appendix. More on Heat Kernels 309

Proof. If I I and II . II are two metrics on a line bundle C such that II - II = of I .

and V is a connection compatible with the metric I I, then it is easy to see
that V + df + w will be compatible with the metric II . II for any imaginary
one-form w.
The connection Vdet(x[o,a)) is compatible with the L2-metric I I. Thus, we
see that any connection of the form
vdet(?{[o,a>) - 2d('(O, D- D+, A) +w = Vdet(rc[o,a)) + z(,Oa +Qa) +w
where w is an imaginary one-form, will be compatible with the Quillen metric
(o,D-D ,A)/2I . I. Choosing w = (,O, -,3f )/2, we obtain the desired
I
IQ = e-C'
result.

Appendix. More on Heat Kernels

In this appendix, we will prove a generalization of the construction of heat
kernels of Chapter 2 which we need in the course of this chapter.
Let M be a compact manifold, let £ be a vector bundle on M, and let
A = Ez o Ai be an finite-dimensional algebra with identity graded by the
natural numbers. (In practice, A will be an exterior algebra.) We will denote
by D the algebra of differential operators D(M, £ (& IAI1/2), by K the algebra
of smoothing operators on £®1 AI1/2, and by P the algebra of operators acting
on I'(M, £ (9 IAI 1/2) generated by D and by K. We call an element of P a
P-endomorphism.
Let M be the algebra P ® A. We define a decreasing filtration of the
algebra M by setting Mi = Ej>i P ® A. The space I'(M, £ ® IAI1/2) ® A
is be made into a module for M, by letting P act on r(M, £ (9 IAI1/2), and
letting A act on itself by left multiplication.
We will prove that an operator of the form
.F = Ho + K + .F'(+),
where Ho is a generalized Laplacian, K E K and .F[+) E M1, has a heat kernel
satisfying many of the properties of the heat kernel of a generalized Laplacian.
We define a heat kernel for F to be a continuous map (t, x, y) F-+ pt(x, y) E
(£ (9 IAI1/2)x ® (£* (g IAI1/2)y ® A which is C1 in t, C2 in x, and satisfies the
equation
a
atpt (x, y) + LF pt (x, y) = 0,
with the boundary condition that for every s E r(M, £ (9 I AM I1/2) ® A,

uniformly in x E M.
lim f EM pt(x, y) s(y) = s(x)
t --+o

Let us start with the case where A is the algebra of complex numbers C,
so that there is no component .P[+]. Thus, we wish to construct the heat
310 9. The Index Bundle

kernel of an operator H of the form H = Ho + K, which differs from a

generalized Laplacian Ho by a smoothing operator K. Since we know that
Ho has a smooth heat kernel satisfying certain strong estimates, it is not very
surprising that we can prove the same things for H. To construct the heat
kernel of H, we use a generalization of the Volterra series of Proposition 2.48.

Proposition 9.46. The series

Qt-_ (-t)k
00
f e-aotHOKe aitHO ... Ke-aktHO do,
k=0 k

converges in the sense that the corresponding series of kernels converges for
t > 0, with respect to any Ct-norm, t > 0, to a kernel
qt E r(M x M, (£ (9 JAI1/2) ® (E*,g JAI 1/2)).
The sum is C°O with respect to t and is a solution of the heat equation
(at + (Ho). + KK)gt(x, y) = 0
with the following boundary condition at t = 0: if 0 is a smooth section of
£ ®IAI1/2, then
lim Qts = s in the uniform norm.
t-0
Thus qt(x, y) is a heat kernel for the operator H = Ho+K. Furthermore the
kernel of the difference
00
e-H - e-tHO = E(_t)k eaOtHOKe-altH0 ...Ke-aktH° do,
k=1 JOk
tends to 0 when t --+ 0.
Proof. Since K is a smoothing operator, the operator a-tHOK has a smooth
kernel for all t > 0, and
IIe-tHoKIIP < C(t)IIKIIP,

for some constant C(i) depending on e. It follows that, for k > 1,

f e-aotHOKe_0 tHo ... Ke-aktHo C(e)k+1II KII e

JLk e-C k!

Thus the series >k>1 converges with respect to the Ct-norm, uniformly for
t > 0, with similar estimates for the derivatives with respect to t. It is easy
to verify, as in theorem 2.23, that the complete sum >k>0 is a solution of the
heat equation. The estimate II a tH - e-tHo IIt = 0(t) is clear, and implies the
boundary condition for t = 0.
Appendix. More on Heat Kernels 311

If the operator H = Ho + K is symmetric, then we can extend a number

of properties which hold for a generalized Laplacian Ho to the case where K
is non-zero.
(1) As in Proposition 2.33, H is essentially self-adjoint on the Hilbert space
of L2-sections of £ (9 JAI 1/2.
(2) The unique self-adjoint extension of H has a discrete spectrum bounded
below; each eigenspace is finite-dimensional and is contained in the space
of smooth sections.
(3) The analogue of Proposition 2.37 holds, namely, if H = Ho+K, and if Pl
is the projection onto the eigenfunctions of H with positive eigenvalue,
then for each $ E N, there exists a constant C(P) > 0 such that for t
sufficiently large,

II(x I Pie-"Pi I y)Ile < C(2)e-tal/2,

where \1 is the smallest non-zero eigenvalue of H.
(4) The Green operator G, which is H-1 on ker(H)J- and 0 on ker(H), op-
erates on the space of smooth sections of £ ®JAI1/2, and is given by the
formula
G= PletHPi dt.
f
We now turn to the case of an operator F of the form
F = Ho + K + .F[+],
which differs from a generalized Laplacian by the sum of a smoothing operator
K E 1C and an operator P(+] lying in the ideal Mi. The operators .P and H
differ by an operator P[+l of positive degree in the finite-dimensional graded
algebra A. Since we have already constructed the heat kernel of H = Ho + K,
the Volterra series once more gives a candidate for the heat kernel of .P: for
fixed t > 0, define the operator e-'s to equal
e-t.Y = e-tH + E(-t)1Ik,
k>0

where
Ik = e-QOtH_F[+]e-o,1tH.[+] ... e-Qk_1tHF[+]e-QktH do..
fOk
The sum is finite, since Ik E Mk, and for k large, we know that Ak = 0, and
hence Mk is zero. Thus, we only have to make sense of each term in the sum.
We need the following lemma.
Lemma 9.47. Let D be a differential operator of order k. There exists a
constant C > 0 such that if K is a smoothing operator, one has the following
bounds: for t E [0, T], where T is a positive real number, we have
ll De-tHK111:5 CJJKllk+e
312 9. The Index Bundle

IIKe-tHDIIt S CIIKIIk+e-

Proof. There exists a constant C(t) such that for 0 E Pt(M, £) one has for
t E [0,T], Ile-tH0IIe < C(i)11011e. The bound IIDe-tHKIIt < C(t)IIKIIk+t
follows easily. Then we write
Ke-'H D =
The above bound applies to the kernel of D*e tH*K* , yielding II Ke-tHDII t <
C(e)IIKIIk+t
We can now complete the proof of the following theorem.
Theorem 9.48. There exists a unique heat kernel pt (x, y) for F, which is a
smooth map from t E (0, oo) to r (M x M, (£ (9 IAI1/2) ® (£®®IAI1/2)) .

Proof. We must show that each term Ik has a smooth kernel. On the simplex
Ok, one of the v;, must be greater than (k+1)-1. Since for (k+1)-1 < a- < 1
and fixed t, the operator e-'1H has uniformly smooth kernel, it follows by
iterated application of Lemma 9.47 that the operator
e-too HF[+] e-a1 tH T[+] ... e-ak_ 1 tH F[+] a-ak tH
on

has a smooth kernel which depends continuously (o-o, ... , Qk) E Ok. Thus
the integral makes sense as an operator with smooth kernel. As in Chapter 2,
we see easily that e-tH + >k>o (-t) klk satisfies the heat equation for F. The
boundary condition as t -r 0 holds, since for 0 smooth, each term Ikcb has a
limit as t -> 0; when multiplied by (-t)k, each term contributes zero except
the term Io = e-tH. Uniqueness is proved as in Chapter 2 as a consequence
of the existence of the heat kernel for the adjoint operator F*.
There is an analogue of Proposition 2.37 for operators of the type F.
Proposition 9.49. Suppose F = H+.P[+] where H = Ho + K is the sum of
a symmetric generalized Laplacian Ho and a symmetric smoothing operator
K, and .P[+] E M,. If Pi is the projection onto the positive eigenspace of H
and P, commutes with F[+], then there exists an e > 0 such that the kernel
of the operator e-t' satisfies the estimates
II(x I Pie-tFP, I y) < C(t) e-et
onMxM ast ->oo.
Proof. Using the Volterra series, we may write

k=0
where
Ik = e-QOtH. e-a1tH ....P e-aktH
J
Appendix. More on Heat Kernels 313

It follows that P1IkP1 equals

(P1eaotH pl)(PiT[+1 Pi)(Pie-°1tHP1) ... (P1.F,+,p1)(Ple-aktHP1)

Lk
By the analogue of Proposition 2.37 to the operator H, we see that for t
sufficiently large, and for each Q,

II (x I Pie-tH pi I y) Ilt < CV)e-tat/2

On the simplex Ok, one of the ai must be greater than (k+1)-i. The estimate
now follows easily by induction using Lemma 9.47 combined with the above
decay for a o,tH i]
Let 7r : M -+ B be a family of manifolds over a base B. Denote by M x, M
the fibre-product
Mx,,M={(x,y)EMxMI7r(x)=7r(y)},
which is a fibre bundle over B with fibre at z E B equal to MZ x M. Let
pr1 and pre be the two projections from M x, M to M, and let IA,I be the
vertical density bundle of the fibre bundle it : M --+ B. If £1 and £2 are vector
bundles on M, let Si ®,r £2 be the vector bundle over M x, M given by the
formula
91 9,, E2 = pri £1 ® pre E2.

Let £ -+ M be a family of vector bundles EZ -> M, . We define a smooth

family of smoothing operators acting on the bundles EZ -> MZ along the
fibres of 7r : M --a B to be a family of operators with a kernel
k E r(M x,r M, (6 ®IA,11/2) Z' (6* (9 IA, 11/2)).
When restricted to the fibre M,z x MZ, the kernel k may be viewed as a
kernel kZ in r(MZ x M,z, (6 (9 IAM. I1/2) ®' (£* ® IAM= 11/2)), which defines a
smoothing operator KZ on 7r*SZ = r(MZ, 4) by the formula

(KZO)(x) = fM
kZ(x, y) (y)
for 0 E r(MZ, £Z ® JAM. I1/2). We define the bundle IC(E) over B to be the
bundle whose smooth sections are given by
r(B, K(£)) = r(M x,r M, (£ (9 IArI1/2) ® (E* (9 1A, 11/2));
as explained in Definition 9.8, 1C(£) is a sub-bundle of Endp(E).
Suppose we are given a smooth family of generalized Laplacians HZ along
the fibres of M -1 B; in other words, we are given the following data, from
which we construct the Laplacians HZ in the canonical way:
(1) a smooth family gM/B E r(M, S2(T(M/B))) of metrics along the fibres;
314 9. The Index Bundle

(2) a smooth family of connections acting on £ along the fibres, which we

may suppose to be the restriction of a smooth connection on the bundle
£ - M;
(3) a smooth family of potentials FZ E r(MZ, End(£Z)), in other words, an
element of r(M, End (S)).
We have the following result, which generalizes Theorem 2.48.
Theorem 9.50. If HZ is a smooth family of generalized Laplacians, then
the corresponding heat kernel pt (x, y, z) defines a smooth family of smoothing
operators, that is, a section in r(B, K(£)).
Proof. As usual, around any point zo E B, we can find a neighbourhood on
which the families M and £ are trivialized. Thus, we may replace B by a ball
U C RP centred at zero, M by the trivial bundle Mo x U, and £ by the bundle
£o x U, where £o is a bundle over Mo. Since the changes of coordinates used
to obtain this trivialization are smooth, as are their inverses, we see that the
data used to define the family of generalized Laplacians HZ give a smooth
family of data for defining generalized Laplacians on the bundle £® ® IAM. I1/2,
parametrized by the ball U. By Theorem 2.48, we know that for any t > 0,
the derivative &. pt(x, y, HZ) is a smooth family of smooth kernels on £o, from
which the theorem follows. O
The manifold M embeds inside M x., M as the diagonal M C M x,r M.
If we restrict a kernel k E r(M x,r M, (£ (& JA, 11/2) 2,, (9* (9 IA,, 11/2)) to the
diagonal, it becomes a section of End(E) ®JA, 1, and its pointwise trace Tr(k)
becomes a section of the bundle of vertical densities JA,,I. Such a section
may be integrated over the fibres to give a function on B; we will denote this
integral by fM/B Tr(k(x, x)). In this language, the formula for the trace along
the fibres of the family of operators KZ becomes

Tr(KZ)
= f MIBTr((x I KZ I x)) E C-(B)

We may also state the family version of this theorem. Let A be a bundle
of finite-dimensional graded algebras with identity over B (in practice, A =
AT*B) and let M be the bundle of filtered algebras M = A ® End-p(9). Let
.F E r(B, A ® End-p(£)) be a family of P-endomorphism with coefficients in
A, of the form
T=Ho+K+.F[+],
where Ho E r(B, D(£)) is a smooth family of generalized Laplacians, K E
r(B, )C(£)) is a smooth family of smoothing operators, and T[+] is an element
of r(B,M1) = >ti 1 r(B, Ai 0 Endp(£)).
Theorem 9.51. For each t > 0, the kernel of the operator a-tom is a smooth
family of smoothing operators with coefficients in A, that is, a smooth section
in r(B, A (9 K(£)).
Bibliographic Notes 315

Bibliographic Notes
The main theorem of this chapter, Theorem 9.33, is proved in the first part of
Bismut's article on the family index theorem [30]. This theorem is a synthesis
of the heat kernel approach of McKean and Singer towards the local index
theorem with Quillen's theory of superconnections. Our other main reference
for this chapter is Berline and Vergne [26], where Theorem 9.2 is proved. The
proof of our refinement Theorem 9.19 of Theorem 9.33 is new.
Sections 1, 2 and 3. Atiyah and Singer first defined the family index of a
continuous family of elliptic pseudodifferential operators, as an element of the
K-theory of the base [16]. They then derived a formula for this family index,
generalizing the Riemann-Roch theorem of Grothendieck [39]. For expositions
of topological K-theory, see the books of Atiyah [3] and Karoubi [71].
Bismut, by contrast, obtained a formula for the Chern character of the
family index in the special case of a family of Dirac operators; in this, he
was inspired by Quillen's definition of the Chern character of a superconnec-
tion [93].
The results on the convergence of the transgressed Chern character (Theo-
rems 9.7 and 9.23) are new, but were inspired by an analogue of Theorem 9.23
for complex manifolds due to Gillet and Soule [67].
Section 4. The example presented in this section comes from Section 5 of
Atiyah and Singer [16]. However, since we have the Kirillov formula of Chap-
ter 8 at our disposal, we are able to give a formula for the Chern character
form of the index bundle.
Section 5. The techniques of this section are those of Atiyah and Singer [16].
Section 6. The zeta-function of the Laplace-Beltrami operator was introduced
by Minakshisundaram and Pleijel [86]. This- work was extended to elliptic
pseudodifferential operators by Seeley [99], and applied in the original proof
of the Atiyah-Bott fixed point formula [5] (see Section 6.2). The definition of
the zeta-function determinant of a Laplacian was given by Ray and Singer [94].
Section 7. Our definition of the determinant line bundle is taken from an
article by Bismut and Freed [34]; the construction is based on the ideas of
Quillen [92]. This theory began with the work of physicists, in particular
Alvarez-Gaume and Witten on anomalies of quantum field theories (which is
in fact precisely the theory of the Chern class of the determinant line bundle
of a family of Dirac operators); see for example [106].
Chapter 10. The Family Index Theorem

Let 7r: M ---+ B be a family of oriented Riemannian manifolds (Mz I z E B),

and let £ be a bundle on M such that £x = Elm. is a Clifford module for
each z; suppose in addition that there is a connection Ve given on £ whose
restriction to each bundle £Z is a Clifford connection. Let 7r*£ be the infinite-
dimensional bundle over B whose fibre at z E B is the space r(MZ, £z); let
D = (Dz I z E B) be the family of Dirac operators acting on the fibres of
-7r*£, constructed from the Clifford module structure and Clifford connection
on S. The aim of this chapter is to calculate the Chern character of the
index bundle ind(D) E K(B), by introducing a superconnection for the family
of Dirac operators D whose Chern character is explicitly calculable. (Note
that because the fibres of the bundle M/B have Riemannian metrics, the
line bundle IA1I1/2 has a canonical trivialization; this reconciles the above
definition of it*£ with that in Chapter 9.) This theorem is a generalization of
Theorem 4.1 of Chapter 4; as in that chapter, we must assume that the Dirac
operators Dz are associated to Clifford connections on £z.
We introduce the bundle ]E = lr*AT*B ® £, which has the property that
A(B, it*£) ^_' r(M, ]E).

Recall that a superconnection for the family D is a differential operator A on

the bundle ]E --+ M, which is odd with respect to the total Z2-grading of the
bundle E, such that
(1) (Leibniz's rule) if v E A(B) and s E r(M,]E), then
A(vs) = (dBv)s + (-1)1"1/ As;
(2) A= D +>dim(B)
A[j], where A[j] : AO (B, 7r.6) -+ A*+'(B, 7r. E).
Let At = t112gt , A. bt = Ed_o(B) t(1_i)/2A[i] be the rescaled supercon-
nection corresponding to A. In this chapter, we will describe a particular
superconnection A for the family of Dirac operators D, for which the limit

lim ch(At) = lim Str(ent )

t-+o t-+o
exists in the space A(B). Note that there is no reason a priori why the limit
limt_,o ch(At) should exist; this is analogous to the fact that the local index
theorem is not true for arbitrary Dirac operators, but only those associated to
10. The Family Index Theorem 317

an ordinary Clifford connection. Since the superconnection A was constructed

by Bismut, we will call it the Bismut superconnection.
Recall that in the last chapter, we proved that ch(At) is a differential form
which lies in the de Rham cohomology class of the index bundle ind(D) E
K(B) of the family D. It follows that the limit limt_+o ch(At) lies in the same
cohomology class. The formula for this limit, explained below, gives a way of
calculating the Chern character of the index bundle. Thus, Bismut's theorem
implies a local version of the family index theorem of Atiyah and Singer for
the family D.
To define the Bismut superconnection, we need a connection for the fibra-
tion M/B. In other words, if T(M/B) is the bundle of vertical vectors, we
need a decomposition of TM into a direct sum of T(M/B) and a horizontal
tangent bundle THM, isomorphic to 7r*TB. This allows us to construct a
canonical connection V M/B on T (M/B), with curvature RM/B; we do this in
Section 1. The bundle £ is assumed to be a Clifford module for the Clifford
bundle C(M/B) = C(T*(M/B)). We assume the connection Ve is compati-
ble with the connection VM/B. Denote by Fe/S the twisting curvature of Ve:
if we are given a spin-structure on the bundle T(M/B) with spinors SM/B, so
that £ = SM/B ® W for the bundle W = Homc(M/B) (SM/B, £), the twisting
curvature Fe/S is the curvature Fw of the twisting bundle W. Let
RM/B12
A(M/B) = detl/2 (Sjflh(RM,(B/2))

be the A-genus of the connection VM/B, and let

StrC/S(e_Fe/S)
ch(£/S) =
be the relative Chern character of the Clifford module S. We will prove
Bismut's theorem,

to
lim ch(At) = (27ri)-"/2 f A(M/B) ch(£/S),
B

where n is the dimension of the fibre.

Now make the stronger assumption that the dimension of the kernel ker(D)
is constant, so that ker(D) is a finite-dimensional subbundle of 7r*£. If Po is
the projection from 7r*£ onto the finite dimensional bundle ker(D), we may
form the projected connection Vo = PoA[1]Po on ind(D). The Chern character
ch(At) has a limit as t --+ oo as well, equal to the Chern character of V0:
lim ch(At) = ch(Vo).
t--+oo

We will prove that the integral

f0
00
Str
(e) dt
318 10. The Family Index Theorem

converges in A(B) under the hypothesis that ker(D) is constant in dimension,

and thus derive the transgression formula for the Chern character associated
to the Bismut connection,
00
ch(Vo) _ (2i7r)-n/2 fM1B A(M/B) ch(£/S) +dJ dt.

In Sections 1-3, we study the geometry of a family M/B with a given

horizontal tangent bundle. Thinking of M as a Riemannian manifold with
degenerate metric g = gM/B on T*M, defined to vanish on the horizontal
cotangent vectors, then we can construct a natural Clifford module structure
and Clifford connection on E using one more piece of data, a metric on the
base B.
Having constructed the degenerate Clifford module structure on E, we de-
fine the Bismut superconnection A to be the associated Dirac operator; we
then prove the important result that A does not depend on the horizontal
metric 9B used in the definition of the Clifford module structure and Clifford
connection on E. It is interesting to observe that A has in general terms up
to degree two:
A = A[0] + A[1] + A[2].
Let F = A2 be the curvature of the Bismut superconnection. In Section,3,
we will prove Bismut's explicit formula for F, and explain how it is a gener-
alization of the Lichnerowicz formula.
By the results of Chapter 9, for each t > 0, the heat operator a-t."C' acting
on r(M, ]E) has a kernel
kt E r(M x, M, 7r*AT*B (o (£ ®, £*)).
If we restrict the kernel kt to the diagonal, we obtain a section kt (x, x) of
the bundle of algebras 7r*AT*B ® End(E) over M. The bundle End(E) is
isomorphic to C(M/B) ® Endc(M/B)(£), and applying the symbol map to
C(M/B), we obtain a section of 7r*AT*B ® AT*(M/B) 0 Endc(M/B) (£). Us-
ing the connection on the fibre bundle M -+ B, we obtain an isomorphism
of 7r*AT*B ® AT*(M/B) with AT*M, so that we obtain a differential form
kt(x, x) on M, with values in the bundle Endc(M/B) (£). We will calculate
limt_,0 ch(At) by showing that kt(x, x) has an asymptotic expansion
0
kt(x,x) N (47rt)-n/2T, tik2(x)
i=o

with coefficients ki E Ej:52i A' (M, Endc(M/B)(£)), such that

dim(M)/2
E (ki) [2i] = A(M/B) exp(-Fees).
i=0
From this result, we immediately deduce that limt_,0 ch(At) exists.
10.1. Riemannian Fibre Bundles 319

. In Section 5, we prove the transgression formula, and use this in Section 6

to calculate the "anomaly" formula of Bismut-Freed for the curvature of the
connection on the determinant line bundle introduced in Section 9.7. Finally,
in Section 7, we return to the case studied in Section 9.5 in which the bundle
M/B is associated to a principal bundle P - B with compact structure group
G, and show that in this case the local family index theorem is equivalent to
the local Kirillov formula of Chapter 8.

10.1. Riemannian Fibre Bundles

Let 7r : M --> B be a fibre bundle with Riemannian metrics on the fibres. In
this section, we will describe two different families of connections that may be
constructed on the tangent bundle TM of the total space M. This material
is used in the rest of this chapter in the proof of the family index theorem.
Let it : M ---> B be a fibre bundle. We will denote by T(M/B) the bundle
of vertical tangent vectors. Let us assume that the bundle M/B possesses the
following additional structures:
(1) a connection, that is, a choice of a splitting TM =THM®T(M/B), so
that the subbundle THM is isomorphic to the vector bundle it*TB;
(2) a connection VMIB on T(M/B).
We denote by P the projection operator
P:TM ->T(M/B)
with kernel the chosen horizontal tangent space THM. If X is a vector field on
the base B, denote by XM its horizontal lift on M, that is, the vector field on
M which is a section of THM and which projects to X under the pushforward
7r* : (THM)x -+ T,(x)B. We will often make use of a local frame ei of the
vertical tangent bundle T(M/B), and of a local frame fa of TB, with dual
frames e2 and f a; using the connection, we obtain a local frame of the tangent
bundle TM.
There are three tensors, S, k and Sl, canonically associated to a family of
Riemannian manifolds with connection M/B. These are defined as follows:
(1) The tensor S (the second fundamental form) is the section of the
bundle
End(T(M/B)) ®T*HM?'T*(M/B) ®T(M/B) ®T*HM
defined by
(S(X, 9), Z) = (V BX - P[Z, X], 9)
for Z E r(M,THM), X E r(M,T(M/B)) and 9 E I (M,T*(M/B)). It
is clear that this is a tensor, since P vanishes on THM.
320 10. The Family Index Theorem

(2) The one-form k E A' (M) (the mean curvature) is the trace of S:

k(Z) =Tr'(S(Z)) = E(S(e1, ei), Z).

(3) The tensor S2 is the section of the bundle Hom(A2THM,T(M/B)) over

M defined by the formula
fl(X, Y) = -P[X, Y] for X and Y in r(M, THM).
This is clearly antisymmetric in X and Y, and is a tensor since if we
replace Y by fY, for f E C°°(M), then
1(X, fY) = -f P[X, Y] - X (f) PY = f!(X,Y).
By comparison with (1.6), we see that St may be identified with the cur-
vature of the fibre bundle M/B.
The total exterior differential dM may be expressed in terms of VM/B, $,
S2 and the vertical exterior differential dM/B of the fibre bundle M -a B. We
extend dM/B to an operator on r(M, AT*HM (9 AT*(M/B)) by the formula

dM/B(7C*v ®Q) _ / -1)1"17r*v ®dM/BQ

for v E A(B) and Q E r(M, AT*(M/B)). We cannot define a horizontal
differential on r(M, AT*HM (9 AT*(M/B)), but we may define an operator
6B by means of the connection VM/B on r(M, AT*(M/B)):
MIB
6B(7*v(90) =7r*(dBv)AQ+(-1)I"17r*v®E faAvf. 0.
a

If 0 E r(M,T*(M/B)), we define the contraction (S, 0) E I'(M,T*HM®

T*(M/B)) C A2(M) by the formula
(S, 0) (ei, fa) = (S(ez, 0), fa),
and the contraction (S2, 0) E r(M, A2T*HM) C A2 (M).

Proposition 10.1. dm = dM/B + 6B - E(S, ez)t(et) + E(fl, ea)t(ei)

z z

Proof. It is easy to check that both sides are derivations of A(M), and that
they agree on horizontal forms. Thus, we only need to check that they agree
on the vertical one forms e'. We start by checking this when both sides are
evaluated on ek A fa: on the one hand,

(dMez)(ek, fa) = (e1, [fa, ek]),

while by the definition of S, the right-hand side equals

f
-(V l Be', ek) - (V /Bek, et) + ([fa, ek], ei) = ([fa, ek], e4).
10.1. Riemannian Fibre Bundles 321

We must also evaluate both sides on f, A fa: but the equality comes down to

(dMe2) (fn, fp) _ fR), ey)

Let us now suppose that M --> B is a family of Riemannian manifolds,
thus, on each fibre MZ = 7r-1(z), z E B, there is given a Riemannian metric,
which we will denote by gM/B; in other words, we are given an inner product
on the bundle of vertical tangent vectors T(M/B). There is then a canonical
connection VM/B on T(M/B), constructed as follows.
If we choose a Riemannian metric 9B on the base B and pull it up to
THM by means of the identification THM = 7r*TB given by the connection
on the bundle M/B, we obtain an inner product on the bundle THM, which
we will call a horizontal metric. We may now form the total metric g =
9B ® 9M/B on the tangent bundle TM of M, by means of the identification
TM = THM ® T(M/B). Let V9 be the Levi-Civita connection on TM with
respect to this metric, and define a connection VM/B on the bundle T(M/B)
by projecting this connection

The following proposition shows that the choice of metric 9B on the base is
irrelevant in constructing V M/ B . In what follows, we will denote by g (X, Y)
the inner product of two vector fields X and Y with respect to the metric g,
or simply (X, Y) if the metric g is evident from the context.

Proposition 10.2. The connectionVM/B on T(M/B) is independent of the

metric gB on TB used in its definition.

Proof. This follows easily from the formula (1.18) for the Levi-Civita connec-
tion V9, which we will use constantly in this chapter:
2(VXY, Z) =([X,Y], Z) - ([Y, Z], X) + ([Z, X], Y)
+X(Y, Z) +Y(Z, X) - Z(X,Y).
If the vectors X, Y and Z are all vertical, that is, sections of T (M/B), then
the right-hand side reduces to the Levi-Civita connection on the fibres for the
vertical metric gM/B. On the other hand, if X is horizontal but Y and Z are
vertical, then [Y, Z] is vertical, so that ([Y, Z], X) vanishes, and we see that

2(VX/BY, Z) = (P[X, Y], Z) + (P[Z, X], Y) + X (Y, Z).

From this formula, it is clear that only the vertical metric gM/B and the
vertical projection P are used to define OM/B for X horizontal.
The vertical metric allows us to think of the tensor S as a section of
T*(M/B) ® T*(M/B) ® T*HM. Let us give a more explicit formula for the
tensor S.
322 10. The Family Index Theorem

Lemma 10.3. If VM/B is the connection onT(M/B) associated to a vertical

Riemannian metric gM/B, then S is given by the formula
2(S(X, Y), Z) = Z(X, Y) - (P[Z, X], Y) - (P[Z,Y], X).
Proof. This follows easily from the explicit formula for the Levi-Civita con-
nection.
It is evident from this lemma that (S(X, Y), Z) is symmetric in X and Y.
In fact, it represents the horizontal covariant derivative of the metric.
Let n be the dimension of the fibres of M/B, and let VM/B E An(M) be
the Riemannian volume form along the fibres, where we identify a section of
the bundle AnT*(M/B) with a section of the bundle AnT*M by means of the
connection on the fibre bundle M/B.
n
Lemma 10.4. dvM/B = k AVM/B + ez)t(ei)vM/B
i=1

Proof. We use the formula of Proposition 10.1. Since VM/B preserves the
metric gM/B on T(M/B), and hence VM/BVM/B = 0, we see that

dvM/B = - >(S, e')t(ei)VM/B

i
+ J>, e')t(ei)vM/B
i
Furthermore,

E(S,ei)t(ei)VM/B = E(S(ej,ez),f°)ej Afat(ei)VM/B

i,j,a
= -k(fa)VM/B
If we choose a metric 9B on B, we obtain a Levi-Civita connection on TB,
which may be pulled back to give a connection on the bundle THM = lr*TB.
We can form a new connection on TM taking the direct sum of the connection
VM/B on T(M/B) and 7r*VB on THM (which we will write simply as VB).
We will denote this connection by Ve:
v® = VB ® VM/B.
Note that if we replace g by the rescaled metric ugB + 9M/B, where u > 0,
then VB does not change, and so neither does V.
The connection V e) preserves the metric g, but has non-vanishing torsion
in general. Our next task is to compare the connections V9 and V® on TM.
In order to do this, we introduce a three-tensor w on M which only depends
on the data that we are given for the fibre bundle M --+ B, namely the vertical
metrics gM/B and the choice of connection on M/B.
10.1. Riemannian Fibre Bundles 323

Definition 10.5. Let w E Al(M,A2T*M) be the A2T*M-valued one=form

on M defined by the formula

w(X)(Y, Z) = S(X, Z)(Y) - S(X,Y)(Z)

+ 2(c2(X,z),Y) - a(1(X,Y),z) + 2(c(Y,z),X)

The above formula is antisymmetric in Y and Z, so that w takes its values

in A2T *M, as claimed.
The following result, due to Bismut, is the main result of this section.

Proposition 10.6. The Levi-Civita connection V9 is related to the connec-

tion V by the following formula:
g(V9xY, Z) = g(VXY, Z) +w(X)(Y, Z).

Proof. We will make use of the following lemma

Lemma 10.7. If Y is the horizontal lift of a vector field on B, and if X is

vertical, then [X, Y] is vertical.

Proof. If 0 E C°'(B), then Y(7r*o) = 7r*(7r*Yo), while, since X is vertical,

X(7r*q5) = 0 and (X Y)7r*q5 = 0. This shows that [X,Y](7r*o) = 0, and hence
that [X, Y] is vertical.

Now, observe that (71 Y, Z) - (V Y, Z) is antisymmetric in Y and Z,

because each of the connections V9 and p® preserves the total metric g on
M. The proof of the theorem now consists on a case-by-case examination of
the different situations in which X, Y and Z are horizontal or vertical.
(1) If X, Y and Z are all horizontal lifts of vector fields on the base or all
vertical, it is easy to see that (V Y, Z) equals (V Y, Z) while w(X) (Y, Z)
is zero.
(2) If both Y and Z are vertical, then (V Y, Z) and (V Y, Z) are both equal
to (V BY, Z), and once more w(X)(Y, Z) is zero.
In each of the remaining cases, (V Y, Z) is equal to zero.
(3) If X and Y are vertical and Z is horizontal, then (V Y, Z) equals

-(S(X, Y)Z) = w(X)(Y, Z).

(4) If X and Y are horizontal lifts of vector fields on B and Z is vertical,
then (V Y, Z) equals

2([X,Y],Z)+ 2([Z,Y],X)+.1([Z,X],Y) - 2Z(X,Y) = 2([X,YI,z),

as follows from the above lemma. But this equals -.1 (52(X, Y), Z) by the
definition of Q.
324 10. The Family Index Theorem

(5) If X is vertical and Y and Z are the horizontal lifts of vector fields on B,
then (V Y, Z) is equal to

([X,Y],Z) - It([X,Z],Y) - 2([Y,Z],X)

which equals (Sl(Y, Z), X) by the same argument as for (4). El

a

10.2. Clifford Modules on Fibre Bundles

A fibre bundle M with the data that we considered in the last section may
be thought of as a Riemannian manifold with a degenerate metric go on its
cotangent bundle T*M, which vanishes in the horizontal cotangent directions
T(M/B)-1- C T*M. This metric explodes in the horizontal directions of the
tangent bundle TM of M, so that in contrast to Riemannian geometry, we
distinguish between T*M and TM. Since the Clifford algebra of a vector
space V with vanishing metric is isomorphic to the exterior algebra AV, we
see that the Clifford bundle C(T*M) of M with respect to the degenerate
metric go, which we will denote by Co (M), may be identified with the bundle
of algebras it*AT*B ® C(M/B), where C(M/B) denotes the Clifford bundle
C(T*(M/B)).
In this section, we will generalize the discussion of Clifford modules and
Clifford connections of Chapter 3 to the setting of the Clifford bundle Co (M).
Fix a metric 9B on the base. If U E (0, 1] is a small positive number, let
gu be the metric on T*M equal to the vertical metric 9M/B on the vertical
cotangent vectors, and ugB on the horizontal ones:

9u = gM/B ® ugB.
In particular, go = limu_._,o gu is the degenerate metric that we introduced
above. The family of metrics gu is only a tool in investigating the geometry
of go. Let gu be the dual metric on TM,

9u = 9M/B ®u-1 9B,

which explodes in the horizontal directions as u - 0.
Let Cu (M) = C(T*M, gu) be the Clifford algebra bundle of the bundle of
inner-product spaces (T*M, gu), and denote the canonical quantization map
of Proposition 3.5 from AT*M to Cu(M) by cu. Thus, the Clifford bundle
Co (M) is the limit of the one-parameter family of algebras bundles Cu (M).
Let -ru denote the bundle map from A2T*M to End(T*M) which is defined
implicitly by
[cu(a),cu(6)],
where a E A2T XM and 6 E T*XM; then

Tu (v1 A v2) = 2 (gu (vl, )v2 - 9u (v2, e)vl)

10.2. Clifford Modules on Fibre Bundles 325

In order to understand this definition better, we will write it out explicitly in

terms of an orthogonal frame of T*xM of the form {ei} U { fa}, where ei is
an orthonormal frame for T*(M/B) and f" is an orthonormal frame for T*B,
with respect to the metric ga. In terms of this frame,
2Tu(e'ej)ek = Sikej 6jke' -
fa)ej = Sij fa
l ru(ei
(10.1) Tu(fa f/3)e' = 0

Tu(eZfa)fa = -uSape'
ITu(f0, fo)fry = u (Sa7f'6

- 60ryfa)
Tu(eZej)fa = 0.
It is clear from these formulas that the family of actions Tu has a well-defined
limit as u -> 0, which we denote by T°:
ZTO(eiej)ek = Sikej _ Sjkei
(10.2) 2TO(e'fa)ej =5iifa
TO(fa fO)e' = 0

T°(a)fa=0 forallaEA2T*M.
It is important to note that T° (a) vanishes on T*B.
We will denote the negative of the adjoint of Tu(a) E End(T*M), where
a E A2T*M, by Tu(a) E End(TM), and the negative of the adjoint of T°(a)
by To (a). Note the formula
Zgu(Tu(a)X,Y) = (a,X AY),
for X,Y E r(M,TM) and a E A2T*M.
It is clear that A2T*M has a family of Lie brackets induced on it by the
maps Tu, defined by the formula
Iru[al,a2]u = [Tu(al),TU(a2)].
We denote the limit of [al, a2]u as u --+ 0 by [a/l, a2]o, which satisfies
TO[al,a2]o = [T°(al),T°(a2)]-
Let VM'u, u > 0, be the Levi-Civita connection on TM corresponding to
the metric gu, and let p® be the direct-sum connection introduced in the last
section. In the notation of this section, we may restate Proposition 10.6 as
VM'u = Ve + lru(w).
2
where w is the element of Al (M, A2T*M) of Definition 10.5. It follows that
the family VM'u of connections has a well-defined limit as u --k 0, which we
will denote by VT"M'0, and which is given by the formula
VM'0=0®+2T0(w)-
326 10. The Family Index Theorem

We denote by VT 'M,' and VT *M,0 the dual connections on TM, that is,

(10.3)
pT*M,0 = V® + 2r°(w)

Proposition 10.8. (1) The connection VM,u is torsion free.

(2) The projection of VM,u to the sub-bundle T(M/B), equals VM/B.
(3) Restricted to each fibre, the connection V M,O depends only on the vertical
metric 9M/B and the connection on the fibre bundle M/B.

Proof. For u > 0, the connection VM,u is a Levi-Civita connection, so is

torsion-free. Since limu ,0 VM,u = VM,°, it follows that VM,O is torsion-
free. The connection VM/B on the bundle T(M/B) is by definition equal
to the projection of VM,u onto T(M/B), and taking the limit u -+ 0, we
see that this is true for VM,0. Finally, the fact that the restriction of VM,O
to a fibre of the bundle M/B depends only on the vertical metric gM/B and
the connection on the bundle M/B follows from (10.3), since 7r*VB = 0 on
vertical vectors.

Consider the section of the bundle A2T*M ® A2T*(M/B),

RMIB (W, X) (Y, Z) = (RM/B (W, X )Y, Z) M/B,

where RM/B E A2(M,End(T(M/B))) is the curvature of the connection

VM/B on T(M/B). Using the connection on the bundle M -* B, we may
extend RMI B to a four-tensor on M by setting RMI B (W, X) (Y, Z) = 0 if y
or Z is horizontal. Let R' E A2(M, End(TM)) be the Riemannian curvature
of the metric g,,,, and let RB E A2(B, End(TB)) be the Riemannian curvature
of the metric gB on the base B.

Proposition 10.9. Let R E r(M, A2T*M (9 A2T*M) be the four-tensor de-

fined by the formula

R(W,X,Y,Z) = lim (gu(RL'(W,X)Y,Z)-u-1(RB(W,X)Y,Z)B).

(1) The above limit exists and equals RM/B + V w + .1 [w, w] 0.

(2) When Y and Z are vertical, R(W, X, Y, Z) equals RM/B(W, X, y' Z).
(3) If W, X, Y and Z are vector fields on M, then
R(W,X,Y,Z)+R(X,Y,W,Z)+R(Y,W,X,Z) = 0.
(4) R(W,X,Y,Z) = R(Y,Z,W,X)
 10.2. Clifford Modules on Fibre Bundles 327

Proof. Written as a four-tensor g u (Ru (W, X )Y, Z), the curvature of the metric
gu is a polynomial in u and u-1, and we must show that R is the constant
term in this expansion, by proving that
lim ugu(Ru(W, X) Y, Z) _ (RB(W, X) y' Z)B
u->o
From the formula VM,u = Ve + 27'u(w), we see that
(VM u) 2 _ (Ve)2 + [V ®, Tr(w)] + 7
[Tu(w), T.(w)J
2 8
2Tu(V11w)
ll = (V®)2+ + 2Tu(4[w,w]u])-
Since (Ve)2 = (7r*VB)2 + (VM/B)2, it follows that
gu (Ru (W, X )Y, Z) =u-1(ir*RB (W, X )Y, Z) B + (RM/B (W, X) y' Z)
+ (V®w) (W, X) (Y, Z) + 4 [w, w]u(W, X) (Y, Z),
where we recall that [w,w]u = (Tu)-1([-ru(w),Tu(w)]). The existence of the
limit as u -f 0 now follows from the fact that [w, w]u = [w, w]o + O(u), and
we have

R(W, X, Y, Z) =(RM/B (W, X )y' Z)M/B

+ (VeW)(WX)(Y, Z) + 4[w,w]o(W,X)(Y,Z)
The other properties of R are clear from its definition: for example, (3)
follows from the fact that both Ru and RB, being Riemannian curvatures,
satisfy the desired symmetry.
Let E be a Clifford module along the fibres of the bundle M/B, that is, a
Hermitian vector bundle over M with a skew-adjoint action
c : C(T*(M/B)) -> End(E)
of the vertical Clifford bundle of M/B, and a Hermitian connection VE com-
patible with this action,
[Vi, c(a)] = c(Vx/Ba),
for X E r(M,TM) and a E r(M,T*(M/B)).
Denote by E the vector bundle over M defined by the formula
E _ *AT*B ®E.
This bundle carries a natural action mo of the degenerate Clifford module
Co (M). In order to define this action, it suffices to define the actions of
T*HM C Co (M) and T*(M/B) C Co(M). The Clifford action of a horizontal
cotangent vector a E r(M, T*HM) is given by exterior multiplication
MO (a) = 6(a)
acting on the first factor AT*HM in E, while the Clifford action of a vertical
cotangent vector simply equals its Clifford action on E. This Clifford module
328. 10. The Family Index Theorem

will be the one of the main tools in calculating the family index of the family
D. In order to study E, we will write it as the limit of a family of Clifford
modules for the bundles of Clifford algebras C,,(M), all constructed on the
same underlying vector bundle as E.
In order to define the Clifford action
m,,, : C,, (M) -* End(E) '= End(AT*HM ®E),
it suffices to define the actions of T*HM C C,,,(M) and T*(M/B) C C,,(M).
The Clifford action of a horizontal cotangent vector a E r(M, T*HM) is given
by the formula
mu(a) = e(a) - ut(a),
acting on the first factor AT*HM in E, while the Clifford action of a vertical
cotangent vector simply equals its Clifford action on E. It is a straightforward
task to check that
mu(a)2 = -gu(a, a) for a E r(M,T*M).
In particular, we see that the limiting Clifford module action limu_,o mu is
just the degenerate action mo introduced above.
There are connections VE,0 and VF,,' on the Clifford module lE analogous
to the connections p® and that we have constructed on the bundle
T*M. As before, to construct these connections, we must choose a horizontal
metric 9B on B. The connection y1,6) on lE ^_' Alr*T*B ® 6 is defined by
taking the sum of the Levi-Civita connection 7r*VB on the bundle A-7r*T*B
with the connection De on E,
DE'e=7r*VB®1+1(D De;

the connection VE,u is defined by the formula, inspired by Proposition 10.6,

VF-,u
= VE,® + Zmu(W)
where w is the one-form defined in Section 1.
Proposition 10.10. The connection VE'u is a Clifford connection for the
Clifford action mu of Cu(M) on ]E, in other words,

[oX", mu(a)] = mu(VT'M'ua),

for X E r(M, TM) and a E r(M, T*M). In particular, the connection
DE,1 = lira VE,u = VE>® + 2mo(W)
u->0
is a Clifford connection for the Clifford action mo of the Clifford algebra
bundle CO(M) on lE,

[VX°, mo(a)] =
The restriction of VE'u, and in particular of VE,o, to each fibre of the bundle
M/B is independent of the horizontal metric gB used in its definition.
 Modules on Fibre Bundles 329

Proof. The first step of the proof is to show that VE,® is a Clifford connec-
tion with respect to any of the Clifford actions mu,, if C,,,(M) is given the
connection V :
[V ,mu(a)] = mu(VXa)
This formula follows from considering the two cases, in which X respectively
horizontal and vertical. The proof that VE,u is a Clifford connection for
the action mu and the connection VTM,u on Cu(M) follows by observing
that, almost by definition, [mu (w (X) ), mu (a)] = mu (Tu (w (X) )cx) . Finally,
the independence of the horizontal metric 9B of the restriction of VE,u to a
fibre of the bundle M/B follows as it did for VM,u from the fact that it*VB
vanishes on vertical vectors.
In the rest of this chapter, we will make use of the frame {C2} U { f a} of T*M
introduced in the last section, as well as the dual frame {ei} U {f,,} of TM.
We adopt the convention for indices that i, j, ... , label vertical vectors, a,
Q, ... , label horizontal vectors, while a, b, ... , label all vectors, horizontal
or vertical. With this convention, we denote by ea any one element of the
cotangent frame {ei} U {fa}, and the Clifford action m. (e') by m'U.
In the next section, we will need a formula for the curvature of the Clifford
connection VE°o. First, observe that an easy generalization of Proposition 3.43
shows that the curvature (Ve) 2 of the connection Ve on £ equals

-2 R ab cVe'e' +EFay sea Eb

i<j;a<b a<b

where RM'b = (RM/B (ea, eb)ej, ei) and Fe/s E A.2 (M, Endc(M/B) (£)) is
by definition the twisting curvature of the Clifford connection Ve on the
Clifford module 6 over C(M/B). In the special case in which the fibres
M/B have a spin-structure with spinor bundle S(M/B), we may write £ as
£ = S(M/B) ®W, where the twisting bundle W is a Hermitian vector bundle
with connection Vw, and the twisting curvature Fe/S is then nothing but the
curvature of the connection V1^'.
Let A be the natural action of End(TB) on ATB defined in (1.26). Then
A(RB) is given by the formula (3.15).
Proposition 10.11. The curvature of the connection VE,o on E equals
A(RB) + Zmo(R) + Fels.
Proof. We have
(VE'e + 2mo(w))2 = (VE,®)2 + 2[VE'®,mo(w)] + s[mo(w),m0(W)]
The proposition follows from the formulas
(1) (DE'e)2 = a((VB)2) + (V6)2 = A(RB) + 2mo(RM/B) +Fels,
(2) [VE'®, mo(w)] = mo(V(9w), and
(3) [mo(w), mo(w)] =mo([w,w]o)
330 10. The Family Index Theorem

If the base B has a spin-structure, with associated spinor bundle S(B),

there is another Clifford module on M, this time considered with the non-
degenerate metric g = gB + gMIB, namely 7r*S(B) ® E. Indeed, there is a
natural decomposition
C(M) =' 7r-C(B) ® C(M/B)
of Clifford algebra bundles over M, corresponding to the decomposition T*M
lr*T*B ® T*(M/B). We may introduce two natural connections in this situ-
ation, which are respectively Clifford connections on S(B) ®£ with respect
to the connections y® and O9 on C(M). We will not give the details of the
proof, since it is very similar to that of Proposition 10.10.
Proposition 10.12. (1) The connection
®1 + 1 (g V6
Vs(B)®E,® = O®
on S(B) ® £ is a Clifford connection with respect to the connection y®
on C(M).
(2) Denote by c(w) E A'(M,C(M)) the one form defined by the formula
c(w) = 1 >w(ea)(eb, ec)ea ® c(eb)c(ec).
abc

The connection
2
on the bundle S(B) ® £ is a Clifford connection with respect to the Levi-
Civita connection V on C(M).
(3) If M is a spin manifold with spinor bundle SM, and £ is the relative
spinor bundle S(M/B) =' Homc(B)(lr*S(B),SM), then S(B) ®£ is nat-
urally isomorphic to SM, and V5(B)®e may be identified with the Levi-
Civita connection on S(M).
To finish this section, we give a lemma that we will need later.
Lemma 10.13. If E is any Clifford module for the Clifford algebra bundle
C,,, (M/), with action m, then we have the equality

w(ea)(eb, ec)mumumu = 2,E k(fa)mu - .1 (Q(fa, fp), ei)m,aymamu

abc a a(ji
Proof. Using the fact that w(ea)(eb,ec) vanishes when the three indices a, b
and c are all horizontal and also when both b and c are vertical, the left-hand
side is seen to equal

2w(ei)(ej, fa)mumuma +
ija ia,Q

+ 2w(fa)(ei, fp)mwmuma
a(3i
10.3. The Bismut Superconnection 331

Writing this ought explicitly, we get

-2 S(ei, ej)(fa)mumumu - 2 E(91(fa, fe), ei)mumQrn U.
ij iaji
Since S(ei,ej)(fa) is symmetric in i and j, only the trace of the tensor S
contributes to the first sum, and we obtain our formula.

10.3. The Bismut Superconnection

Let 7r : M -p B be a fibre bundle with data as in the first section, namely
vertical Riemannian metrics and a connection. Recall that given a vector
bundle over M, -7t,,£ is the infinite-dimensional bundle over B whose fibre
at z E B is the space r(MM, £). If £ is a Clifford module for the vertical
Clifford algebra bundle C(M/B), the Dirac operators Dz along the fibres
(Mz z E B), associated to the Clifford module £z over Mz, combine to
I

give a smooth family of Dirac operators D = (Dz z (=- B), which is a

I

smooth section of Endp(7r*£) C End(7r*£) over B. In this section, we will

describe a superconnection, which we call the Bismut superconnection, on
the bundle it*£, with zero-form component D, whose Chern character form
is particularly well-behaved. In view of its fundamental character, we will
present the Bismut superconnection from two different viewpoints which are
not obviously equivalent.
The following definition is as in Section 9.3, except that supress the factor
JA, 11/2, since it is canonically trivialized by the vertical Riemannian metric
on M.
Definition 10.14. The space A(B, 7r*£) of differential forms on B with coef-
ficients in it*£ is the space of sections of the Clifford module IE = 7r* (AT*B) ®£
over M.
We introduce the Bismut superconnection
A : A(B, 7r*£) -+ A(B, 7r*£)
as a Dirac operator for the Clifford module E -> M. This is the motivation
for introducing the whole apparatus of the last section.
Proposition 10.15. The Dirac operator on E, given by the formula

is a superconnection when thought of as an operator on the space A(B, it*£);

we call A the Bismut superconnection. More explicitly, the restriction of
A to a map r(M,£) , r(M,E) is given by the formula
c2vE FEea(V + 2k(fa)) - 4
i a a<O
332
10. The Family Index Theorem

The superconnection A is independent of the horizontal metric gB used in its

definition.
Proof. It is clear that A satisfies the first condition for it to be a supercon-
nection on the bundle 7r*£, namely it is a differential operator on the bundle
E = 7r*AT*B ® £ over M which is odd with respect to the total 7L2-grading.
Next, we check that A satisfies the formula A(vs) = (dBV)S + (-1)1""vAs,
for v e A(B) and s E r(M, B), where dB is the exterior differential on B. It
is easily seen that A(vs) = E0eaV«v+ (-1)I"lvAs. Since the Levi-Civita
connection is torsion free, we see that >a E0 VB v = dBV, as required.
In order to check the explicit formula for A on r(M, £), observe that

A = C` cZV '®+ eaOE'®+ a E w(ea)(eb, ee)momomo.

ti
a abc

The proof of the formula is completed by Lemma 10.13. In particular, we see

that A1ol = D.
The decomposition of the Bismut superconnection with respect to degree
A = A101 + A[i] + A[2] has in general terms up to degree two. Indeed, this
superconnection is a natural example of a superconnection having non-zero
terms of degree higher than one.
Proposition 10.16. The connection part A(,] of the Bismut superconnection
is the natural unitary connection V'*E on the bundle 7r*£ of Proposition 9.13,
bearing in mind the identification of 6 ® IA,I1/2 with £ coming from the
vertical metric 9M/B on M/B.
Proof. Choose a metric gB on B, and let IvBI and IvMI be the Riemannian
densities on B and M with respect to the metrics 9B and gm = gB ® gM/B;
clearly, I vM/ B I = I vM I ®7r* I vB 1-1 It will suffice to check that if X is a vector
field on B, then
.c(XM)IvMI ®7r*IvBI-1+IVMI ®7r*G(X)IvBI-1 =k(X)IvMI ®IvBI-1,

since this will imply that

OXE(S ® IVM11/2 (& 7r*IvB1-1/2) = (V , + 2k(X))s ®IVM11/2 ®7r*IvB1-1/2.
But this is a local result, so we may replace the vertical density I vM I®7r* I vB I-1
by the vertical volume form VM/B in the calculation. By (1.4), we see that
L(XM)vM/B = t(XM)dvM/B + d(G(XM)vM/B)
The second term vanishes, since vM/B is vertical, while the second equals
k(X)vM/B by Lemma 10.4.
In order to compute the Chern character of the superconnection A, we need
the formula for its curvature A2 which we referred to above. Let us introduce
the smooth family AM/B = (Az I z E B) E r(B, End(7r*£)) of generalized
Laplacians along the fibres Mz of the bundle M/B; each Laplacian Oz is just
10.3. The Bismut Superconnection 333

the generalized Laplacian acting on the bundle Ez - MZ corresponding to

the connection VE,OIM. and with zero potential.

Theorem 10.17. Let F£l s E A(M, EndC(M/B)(E)) be the twisting curva-

ture of the Clifford module E, and let rM/B be the scalar curvature of the
fibres of M/B. Then
A2 AM/B
= + !rM/B + E momoFE/s(ea, eb).
a<b

Proof. The proof of this theorem is similar to the proof of Lichnerowicz's

formula, Theorem 3.52. To begin with, denoting the covariant derivative Ve
in the direction ea by Va for any connection V on a bundle over M, we have

A2 = 1 a ]E,0 Mb
mOv6
IE 0
]
ab
_1 E[ma mb v,Ov,0
a b
0 0] a b
ab

+
ab
ma0 [VE,0
a mb]VE°
0 b' + 2i
ab
mamb a,
0 0[0E'0 VF,0
b ]'

Since our frame ea is the union of an orthonormal basis ei of T*(M/B) and a

basis f a of T*B, we see that the first term of this sum is equal to - Ei(vb'0)2.
Using the fact that VE,0 is a Clifford connection, proved in Proposition 10.10,
the second term is equal to
momo(va+M'°eb)V '° _ a c T'M,0e,
6 ec)vb, 0
mOMO(va
ab abc

ac i

Here we have used the facts that VM,O agrees with VM/B when restricted to a
fibre M, that the connection VM,O on TM is torsion-free, and the adjunction
formula
1`(vT.M'Oeb
VT'M,0eb = ac )2c = - (eb vM'0e )Cc
a a e a ec ,
c c

It follows that
A2 = OM/B + momo ([va'°, Ob'°] - 0[e eb]) .

z ab

By Proposition 10.11, Eab m0 6ORE'O(ea, eb) equals

mamoA(RB (e a, eb)) -14E mambm`mdRdab

0 0 0 0 c + E mom4Fel s (ea, eb),
ab abcd ab
334 10. The Family Index Theorem

where Rabcd = R(e°i ed)(eb, ea). The first term of this sum equals
-21 «perybaRB
ya«fi'
«p-Y6

which vanishes since the antisymmetrization of RB over any three indices is

zero. Furthermore, we have
m0 CM d 0Rabcd = mambmam0Rabcd
0 0 0 0 E m0m0m0m0Rabbd
abcd abd abd
a d
= 2 E m'aMdRaidi = -2 jJ Rijii,
adi ii
where we have used for the first equality Proposition 10.9 and antisymmetriza-
tion over the first three indices, and for the second equality the symmetry
Raidi = Rdiaz. But the four-tensor R coincides on vertical tangent vectors
with the Riemannian curvature tensor of the fibre.
We will finish this section by giving another definition of the Bismut super-
connection. Locally, we may assume that the base B is a spin-manifold, with
spinor bundle S(B). Let DM be the Dirac operator on r(M, E 0 7r*S(B))
corresponding to the connection VE®s(B) = Vs(B)®e,® + .1c(w) introduced
in the last section.
If W is a finite-dimensional bundle over B with superconnection A, we
can construct a Clifford superconnection B = A ® 1 + 1 ® Vs(B) on the
twisted Clifford module W ® S(B) over B. Let Da be the Dirac operator on
r(B, W (9 S(B)) associated to this superconnection. We proved in Proposi-
tion 3.42 that the map A -+ Ds defined a one-to-one correspondence between
superconnections on W and Dirac operators on r(B, W (9 S(B)), that is, odd
differential operators D satisfying the identity [D, f] = c(df ), for f E CO° (B).
We will say that Da is the Dirac operator associated to A.
We may identify the spaces of sections r(B, S(B) (9 7r*£) with r(M, £
7r*S(B)). We define a Dirac operator on the Clifford module ir*E ® S(B) to
be an odd differential operator D on the space of sections r(B, 7r*£ ®S(B)) ^_'
r(M, e 0 7r*S(B)) such that [D, 7r* f] = c(7r*(df )) for all f E C°O(B). The
following result is an infinite-dimensional extension of Proposition 3.42.
Proposition 10.18. There is a one-to-one correspondence between super-
connections A on the bundle 7r*£ and Dirac operators on the Clifford module
7r*E ®S(B).
The total Dirac operator DM on r(M,£(&7r*S(B)), when considered as an
operator on the isomorphic space r(B, S(B) 0 7r*£), is a Dirac operator for
the Clifford module 7r*£ ® S(B) over B. Indeed, the fact that [DM, 7r* f ] =
c(7r* (df)) for f E Coo (B) is just a special case of the formula [DM, f] = c(df)
for f E C°°(M). Thus, there is a superconnection A on the bundle 7r*£
whose associated Dirac operator may be identified with DM. In fact, this
superconnection on 7r*£ is just the Bismut superconnection.
10.4. The Family Index Density
335

Theorem 10.19. The Dirac operator on

r(B, 7r*£ ® 8(B)) = r(M, £ ® lr*S(B))
associated to the Bismut superconnection A is the total Dirac operator DM.
Proof. Let 13 = Os(B) ®1 + 1 ®A. If s E r(B, S(B)) and t E r(M, £), we
see that
Dn(s (9 t) = s (9 +Ecas ®Dat
i a
s+ /
2k(fa)cas)®t- 4
/f
E(Q(fa,fp),ei)cac- s®cit.
a a</3 i

Thus, we see that Ds Ea

caVa(B)®e'®+1h,
where his the section of C(M)
given by the formula
h=2 k(fa)Ca - 2 ei)cac'-Rqc$.
a a(3i

By Lemma 10.13, h equals >abc w(ea)(eb, ec)ca'cbec.

Since Vs®s(B) _ Vs(B)®-',n + e(w), the operator DM equals
DM =LaCapa(B)®E,e + 4
w(ea)(eb,ec)cacbcc,
abc
which we see is the same as DA.

10.4. The Family Index Density

In this section, we use the method of Chapter 4 to prove the local family index
theorem of Bismut, that is, to calculate limt_,o ch(At). Let us summarize the
data that we are given:
(1) A fibre bundle 7r : M -> B with a vertical metric gM/B and connection,
that is, a decomposition of the tangent bundle of M into the direct sum
of the vertical tangent bundle T (M/B) and the horizontal tangent bundle
7r*TB; from this, we obtain a connection VM/B on the vertical tangent
bundle T(M/B).
(2) a Clifford module 6 for the vertical Clifford bundle C(M/B), with Clifford
connection DE compatible with VM/B.
Using just this data, we can construct a family of twisted Dirac operators
D = (DZ I z E B), and the Bismut superconnection A, which satisfies A101 _
D. Let n denote the dimension of the fibres M.
Let A. be the finite dimensional algebra AT*ZB. The curvature F = A2 of
A acts on the space A. ® (7r*£)Z of sections of the bundle ]E = Alr*T*B ® £
along the fibre M.
The following lemma is a restatement of Theorem 9.48.
336 10. The Family Index Theorem

Lemma 10.20. For each t > 0, the heat operator a-tF acting on r(M, E)
has a kernel
(x I e-t-, I y) E r(M x,,. M,7r*A® £ ®,r £*),
in the sense that if 0 E r(M, E), we have
(e-t.7=O)(x)
= I x I e-tF I y) 0(y) dy,
M=

where dy is the Riemannian volume form of the fibre MZ and z = 7r(x).

The manifold M embeds inside M x,r M as the diagonal M C M x,, M.
If we restrict the heat kernel of .F to the diagonal, it becomes a section of the
bundle of algebras 7r*AT*B ® End(E). Since £ is a Clifford module over the
Clifford algebra C(M/B), we see that
7r*AT*B ® End(S) = 7r*AT*B ® C(M/B) (9 EndC(M/B)(£)
7r*AT*B ® AT*(M/B) ® Endc(M/B) (£)
= AT*M ® EndC(M/B)(£):
Thus, the restriction to the diagonal kt(x, x) of the heat kernel of F is identi-
fied with a differential form on M with values in the bundle EndC(M/B)(£)
Let A(M/B) be the A-genus of the bundle T(M/B) for the connection
QM/B,

A(M/B) = detl/2 (RM/B/2)

(sinh(RM/B/20'
and let F£/S E r(M, Endc(M/B) (6)) be the twisting curvature of the bun-
dle E. The aim of this section is to prove the following generalization of
Theorem 4.1.

Theorem 10.21. Consider the asymptotic expansion of kt(x,x)

00
kt(x,x) - (47rt) -n/2 `tikz(x)
ii=OO

(1) The coefficient ki lies in Ej<2i .Ai (M, EndC(M/B) (£)) .

Ed`m0(M)/2
(2) The full symbol of kt(x,x), defined by v(k) = o-2i(ki), is given
by the formula

a(k) = A(RMI B) exp(-FE/S) E A(M, EndC(M/B) (£))

Before proving the theorem, let us show how it allows us to compute the
Chern character of the index bundle ind(D) of the family of vertical Dirac
operators D. Recall the operator 6B on A(B) which multiplies differential
forms of degree i by t-i/2. Let
At = t1/26B . A (6B)-1 = tl/2A(O] + A(1] + t-1/2A121
10.4. The Family Index Density 337

be the rescaled Bismut superconnection, with curvature Pt = t8B F (6B)-1.

By Theorem9.19, the Chern character of the bundle ind(D) is represented,
for every t > 0, by the closed differential form ch(At) = Str(e -'t) on B.

Lemma 10.22. Extend the map Str : r(MZ,End(£Z)) -+ COO (M;,) to a map
Str : r(MZ, AZ (gEnd(£Z)) -> AZ ®COO(MZ). Then

ch(At) = fM S (Str(kt(x, x)) dx.

Proof. Using the formula Ft = t6B we see that

ch(At) =
Str(e_,Ft)
= Str(e-t"' Y)

IM
= Str (SB (e-tF))
= SB (Str(e-t-F))

Now, the kernel of the operator e-t-F is by definition equal to kt(x, y), and
the lemma follows now from the fact that

Str(e-tom`) =
.
Let us show that the local family index theorem of Bismut is a corollary of
Strs/S(e-F9/S)

Theorem 10.21. Let ch(£/S) = be the relative Chern charac-

ter of the Clifford module S. Let
TM/B : A(M) -> r(M, lr*AT*B)
be the map given by decomposing the bundle AT*M as a tensor product
AT*(M/B) ® it*AT*B and applying the Berezin integral map to the first
factor, that is, projecting onto AnT*(M/B) ® 7r*AT*B and then dividing by
the vertical Riemannian volume form vM/B in F(M,AnT*(M/B)).

Theorem 10.23 (Bismut). The section

SB (Stre. (kt(x, x))) E r(M, 1r*AT*B)

has a limit when t - 0 equal to

(27ri)-nl2TM/B(A(M/B) ch(£/S)) E r(M,7r*AT*B).
Consequently, the differential form ch(At) E A(B) has a limit when t - 0
given by the formula

tlimoch(At) = (27ri)-n/2fM BA(RM/B) ch(£/S).

338 10. The Family Index Theorem

Proof. Consider the bigrading on AT*XM given by

AP,gZ xM = EAi ®A4T*X(M/B)
P,q

If we decompose the symbol of a E A® ® End(EE) according to this bigrading,

o(a) _ o[P,q] (a)

p,q

where
0'[,,q] (a) E AP, ® AqT*x(M/B) ®Endc(M)(Ex),
we have the formula
Stre. (a) = (-2i) /2 E StrE/S(o[P,n1(a)).
P

Applying this formula to the supertrace of

(47rt)-n/2
kt (x, x) - tiki(x),
i
we see that
8 (Strer(kt(x,x))) ^'
j,P

Since o[P,n] (kj) = 0 if 2j < n + p, we see that there is no singular term in the
asymptotic expansion of 6B (Strex (kt (x, x))) as t --> 0. The explicit formula
of Theorem 10.21 for Ej O2j(kj) implies the formula for the leading order in
the asymptotic expansion.
Bearing in mind the family McKean-Singer theorem of Chapter 9, we obtain
the cohomological form of the Atiyah-Singer family index theorem.
Corollary 10.24 (Atiyah-Singf er).

ch(ind(D)) = (27ri)n/2 M/B A(M/B) ch(E/S) E H2i(B).

We now turn to the proof of Theorem 10.21, which is a fairly straightfor-

ward generalization of Theorem 4.1, the case in which B is a point. We start
by choosing a point z E B, but instead of working with the generalized Lapla-
cian D2 on the manifold M as in Chapter 4, we will work with the curvature
,F' of the Bismut superconnection A restricted to the fibre M, which we may
think of as a generalized Laplacian with coefficients in the exterior algebra
A.
Fix a point xo E M,zi and let V = Txo (M/B) and H = TB be the vertical
and horizontal tangent spaces at xo; thus
T =Tx0M=V ®H.
10.4. The Family Index Density 339

Let U = {x E V I lixii < e}, where e is smaller than the injectivity radius
of the fibre M2. We identify U by the exponential map x '---> exp.0 x with a
neighborhood of xo in M2. Let TM/B (xo, x) be the parallel transport map in
the bundle T (M/B) along the geodesic from x to xo, defined with respect to
the connection VM/B defined in Section 1; since we are working on a single
fibre M, this connection is nothing but the Levi-Civita connection of M2.
Using this map, we can identify the fibre T,,(M/B) with the space V, so that
the space of differential forms A(U) is identified with C' (U, AV*).
Choose an orthonormal basis dxi of V* = T*x0(M/B), and let ei E
r(U,T*(M/B)) = r(U,V*) be the orthonormal frame of T*(M/B) over U
obtained by parallel transport of dxi along geodesics by the Levi-Civita con-
nection on Mx. We denote by ea a local frame of T*M on U consisting of the
union of the cotangent frame ei and of a fixed basis f" of T*ZB.
Let E = £.,O be the fibre of the Clifford module £ at xo, let Sv be the
spinor space of V*, and let W = HomC(v.) (Sv, E), so that E is naturally
isomorphic to Sv ® W. Recall the Clifford action mo of T*M on Alr*T *B ® £,
for which vertical cotangent vectors act by Clifford multiplication on £, and
horizontal tangent vectors act by exterior multiplication on A = A7r*T*B.
Let r1(xo, x) be the parallel transport map in the bundle Az ® £ along the
geodesic from x to xo, defined with respect to the Clifford connection VE,O
of Section 2. Using this map, we can identify the fibre A® ®£. of E at x
with the space AH* ® Sv ® W, and the space of sections P(U, A. (9 £) with
C°°(U, AH* ® Sv ® W).
Although the bundle E over M. is the tensor product of the bundle £ with
the trivial bundle A, the term 2mo(w) in the definition of the connection
VE,0 means that parallel transport with respect to VE,O is not equal simply
to the tensor product of the identity on AZ with parallel transport in £ by the
connection Ve. Using the parallel transport map 1E(xo, x), we can analyse
the Clifford action mo of T*2M on E. by transporting it back to xo, where
it becomes an x-dependent action of T* = H* ® V* on A,, ® E.

Lemma 10.25. If ci is the Clifford action of the cotangent vector dxi on E,

and ea is multiplication by f a in the exterior algebra AH*, we have

mo (ei) = ci + E uiae°
a
mo(fa) = Ear
where ua are smooth functions on U satisfying ua(x) = O(1xD.

Proof. Let 7Z be the radial vector field on U. The frame ei satisfies

0. It follows that
2rO(w(7Z))ez
VM,Oe' = = W(7Z)(ei, f«)e",
a
340 10. The Family Index Theorem

by (10.3), and the fact that w(X, Y) = 0 if both X and Y are vertical. Now,
we use the fact, proved in Proposition 10.10, that VE'O is a Clifford connection:
[VX°, mo(a)] = mo(VM'Oa). This shows that

[V i°,mo(e2)] w(R)(ei,f.)f
a
Integrating this ordinary differential equation with the initial condition that
at x = 0, mo (ei) = ci, we obtain the first formula.
The second formula is clear: since 7-0(a)f' = 0 for all a E A2T xM, the
connection VM,0 coincides with it*VB on the sub-bundle 7r*TB of horizontal
vectors, so that f a = VR f a = 0. C]
We need another simple lemma.
Lemma 10.26. In the trivialization of E over U induced by the parallel
transport map r E (xo, x), the connection VE'0 equals d + O, where

©(ai) = -1 E (R(ai, aj)ea, eb)mambxi + fiab(X)mamb + gi(x);

j;a<b a<b
here fiab(X) = O(1x12) E C°°(U) and gi(x) = O(IxI) E C°°(U) 0 End(W).
Proof. Let R be the radial vector field on U. Since e(R) = 0, the one-form
O is determined (see (1.12)) by the equation
L(R)9 = t(R)
(VE'0)2

= t(R)(A(RB) + 2mo(R) +Fe1S) by Proposition 10.11,

= 2t(R)mo(R) + b(R)Fe/s;
the last line holds because t(R).X (RB) = 0, R being a vertical vector field and
RB being horizontal.
Let Oz be the Laplacian on M,z associated to the connection VE,O, so that
the curvature of the Bismut superconnection equals
.pz
= Az + qrM, + E momoFe/s(ea, eb)
a<b

We can transfer this operator to C°°(U, AT*OEnd(W)) using the quantization

map, that is, by replacing the Clifford action mo(ea) at x = 0 by the action
mi=ei - ti ma=ea.
The resulting operator L is given by the explicit formula
+qrm.+EFbfsmamb.
\ / a<b
Next, we introduce the rescaling operator
6,, : COO(U,AT* (& End(W)) CO°(U,AT* (& End(W)),
10.4. The Family Index Density 341

which is basic to the proof; if a E Cm(U, AiT* (9 End(W)), then

bu(a)(x) = u-i/2 a(u1/2x)
Furthermore, if a is an element of COO (U x R, AiT* (9 End(W)), we define

bu(a)(t, x) = u-i/2a(ut, u1/2x).

By Lemma 10.25,
lim ul/2Su m°(9) . bu 1 = e(9) for 9 E C°°(U,T*).
u-->0
In the next lemma, we calculate the leading term of the asymptotic expansion
of bu0E °Su 1. The following lemma follows immediately from Lemma 10.26.
Lemma 10.27. In the trivialization of Az®£z over U induced by the parallel
transport map rE(x°,x),
u1/26, VE,O 6-1 = ai _.I
. R(ai, aj, ea, eb)xomambxi + O(ul/2)
j;a<b
We can now show that the family of differential operators uSuLS; 1 acting
on C- (U, AT* ® End(W)) has a limit as u -4 0; this is the most important
step in the proof. Since the twisting curvature F = F o's of the Clifford module
£ at x° is an element of A2T* ® End(W), it acts on the space C' (U, AT* (9
End(W)) by multiplication. Similarly, the matrix-coefficients
19j)xo6aEb
aij = (RMI Ba,, aj) x,, = E(RMI B(ea, eb) ai,
a<b

of the curvature of the bundle T(M/B) belong to A2T* and act on the space
C- (U, AT* (9 End (W)) -
Proposition 10.28. The differential operator uSu L Su 1 on C°°(U, AT* ®
End(W)) has a limit K when u tends to 0, equal to
aijxj)2

+ F.
i j

Proof. By Lemma 10.27, the differential operator ul/2Supe=°bu 1 equals

ai -4 (R(ai, aj)x°ea, eb)EaEbxj + O(ul/2)

j;a<b
The fundamental symmetry of the four-tensor R (Proposition 10.9 (3)) shows
that
R(ai, aj, ea, eb) = R(ea, eb, ai, aj) = (RM/B(ea, eb)ai, aj),
from which it follows that
u1/2SuVE,0S;1 = ai - -Iaijxj +O(uX12).
342 10. The Family Index Theorem

The operator u6u L Su 1 is equal to

- Cul/2 oez° bu 1
I2
+ uSu mamb su 4FE/S(ea, eb)(u1/2x)
i / a<b
-u Su Vo°:e, _6.-I + 4rM= (u1/2x).

It is clear that both the third and fourth terms behave as O(ul/2) as u --+ 0,
while
u5umambbu lFE/S(ea, eb)(ul/2x) _ >Ea,bFab + 0(ul/2).
a<b a<b
Putting all of these equations together, and using the fact that ei = ai at
x = 0, we obtain the theorem.
Consider the time-dependent function on U given by transferring the heat
kernel Kx I e-l"' I xo) to the open set U C TT0Mx by means of the parallel
transport map TE (x0, X):
k(t, x) = TE(xo, x) (x I e-t-F I xo), where x = expx0 x.
Then k(t, x) lies in C°°(U, AH* ® End(Sv) (9 End(W)); using the symbol
map End(Sv) -> C(V*) --+ AV*, we may think of k(t,x) as a map from U
to AV* ® AH* 0 End(W) ^_' AT* ® End(W), and is a solution to the heat
equation
(at + L) k(t, x) = 0
with initial condition limt_,0 k(t, x) = S(x). As in Chapter 4, we rescale the
heat kernel as follows.
Definition 10.29. The rescaled heat kernel r(u, t, x) equals

r(u, t, x) = u("-i)/2 k(ut, ul/2x)[i].

E
i=o
The choice of the particular rescaling operator bu is motivated by the fact
that
lim r(u, t, 0) = lim bu(kt(xo)) E AT XOM,
u->o u->°
and the right-hand side of this equation is precisely what we must calculate
to prove Theorem 10.21.
It is clear that r(u, t, x) satisfies the differential equation
(at + uSu L Su l)r(u, t, x) = 0.
The rest of the proof proceeds in exactly the same way as the proof of Theo-
rem 4.1. Lemma 4.18 applies to the operator A2 on Mx just as well as it does
when the base B is a point, and we obtain AT*®End(W)-valued polynomials
10.5. The Transgression Formula 343

'yi (t, x) on ][8 x V such that for every integer N, the function rN (u, t, x) defined
by
2N
qt(x) E ui/2yi(t, x)
i=-n
approximates r(u, t, x) in the following sense: for N > j + IaI/2,
I1Olta.'(r(u,t,x) -rN(u,t,x))II <_ C'(N,j,a)vN,
for 0 < u < 1 and (t, x) E (0,1) x U. Just as in Chapter 4, yi(0, 0) = 0 if
i ,-E 0, while -yo (0, 0) = 1. Using the fact that L(u) = K + O(u1/2), we can
show in the same way as when the base B is a point that there are no poles
in the Laurent series expansion of r(u, t, x), that is, that 'yi(t, x) = 0 for i < 0
so that
r(u, t, x) ,., qt (x) E uZ/2,yi(t, x).
i=0
The limit limu_,o r(u, t, x) = qt (x)'yo(t, x) is a formal solution to the heat
equation
(at + K)gt(x)-yo(t, x) = 0
with initial condition yo (0, 0) = 1. The following theorem follows from Theo-
rem 4.12.

Theorem 10.30. The limit limu.. 0 r(u, t, x) exists and is given by the for-
mula
tRM/B/2(sinh(tRM/B/
(47rt) -n/2 det1/2 exp(-tF)
2))
1 RM/B/2
x exp(-4t
\x coth(tRM/B/2)
I I

Theorem 10.21 follows from this lemma when we set t = 1 and x = 0.

10.5. The Transgression Formula

In this section, we will show that the transgression formula
rT
ch(A3) - ch(AT) = d Str ( att a-At) dt
f.9
has a limit as s -> 0, when A is the Bismut superconnection; of course, for
an arbitrary superconnection for the family D, this is not true.
Let us start with an analogue in finite dimensions of the technique which
we will use. Recall the construction of Section 9.1, in which we associated to
a smooth family of superconnections As on a superbundle E over a manifold
B parametrized by R+ a superconnection A on the superbundle £ = 9 x 1[S+
344 10. The Family Index Theorem

over the manifold B = B x R+. Consider the special case in which A3 is the
rescaled superconnection
(6B)-1,
As = 81126E . A

and let it be the dilation of the corresponding superconnection A on the

superbundle E:
a
At = A3t + ds A s,

with curvature J t = At .

Lemma 10.31. We have

Str (e-39) = Str(e---`) - ta(t) A ds,
s=1

where a(t) equals

a-A2
a (t) = Str (ddtt )

Proof. Since At = At8 + dR+, it follows that at s = 1,

J't =.Ft+tdAt Ads.

The lemma now follows from the Volterra expansion and the definition of
a(t).
The following theorem is proved by applying the family index theorem to
a superconnection over B = B x ]R+ constructed from the Bismut supercon-
nection in the same way as above.

Theorem 10.32. (1) The differential form

a(t) = Str ( dtt a-A2 )

has an asymptotic expansion as t -+ 0 of the form

a l(t)
1 _ j=1
tj/2-laj/2,

fch(
where aj/2 E A(B).
(2) The Chern form ch(At) is given by the transgression formula

= (2i)/2
f MID
A(M/B) ch(E/S) - d J a(s) ds.
0
10.5. The Transgression Formula 345

(3) If dim(ker(Dz)) is independent of z c B, so that ind(D) is a vector bundle

with connection Vo = PoA[l]Po, then

ch(Vo) = lim ch(A3)

8-00
= (2iri)-n/2 LIB A(M/B) ch(S/S) - d J W a(s) ds.

Proof. It is clear that Parts (2) and (3) are an immediate consequence of
Part (1) and the explicit formula for limt_,o ch(At) which comes from the
local family index theorem.
Let I[8+ be the positive real line (0, co), and consider the family of manifolds
M = M x ]f8+ over the base b = B x R+. To prove that the function s F-+ a(s)
is integrable as s -- 0, we will apply the family index theorem to this family,
with respect to the following additional data:
(1) the vertical metric gM,B on M equals s-1gM/B;
(2) the connection on the family M -- B equals the natural extension of
the given connection on the family M/B in the directions tangential to
B, and the trivial one in the direction tangential to R+, for which the
horizontal tangent space is spanned at each point by 8/8s;
(3) the Clifford bundle over M is E x R+, with Clifford connection VE +dx+.
We denote by S, k -and T, the metric and tensors corresponding to the above
data on M, and by A and F the corresponding Bismut superconnection and
its curvature. Let n = dim(M) - dim(B) be the dimension of the fibres of
the family M -> B.

Lemma 10.33. (1) If X and Y are vertical vector fields on M, then the
tensor S(X, Y) equals

S(X, Y) = s S(X, Y) - Zs2 (X, Y) ds.

(2) The one-form k equals k = k - 2sds.

(3) The tensor f1 equals Q.
(4) The Bismut superconnection for the above data on the family M --> B
equals
nds.

Pro Given vector fields on M, we may extend them to vector fields on all
of M by taking the extension which is independent of s E R+. Denote by 83
the vector field on M which generates translation along R+. If X and Y are
346 10. The Family Index Theorem

vertical vector fields on M and Z is horizontal, S(X, Y) may be calculated

as follows:

(S(X,Y),as) = 2as(X,Yif
a
- 2
ass-1(X,I') _ -zs-2(X,Y),

(S(X,Y), Z) = a(Z.(X,Y)1ir - (P[Z,X],Y)M - (P[Z,Y],X)M)

= s-1(S(X Y) Z)

The formula for the one-form k follows by taking the trace of S.

It is easy to see that f2 = 0, since [897 X] = 0 if X is any vector field on
M.
To compare the Bismut superconnections A and A, we use the above for-
mulas, the explicit formula for A given in Proposition 10.15, and the fact that
if a E Al (M), the Clifford action c(a) on £ equals sl/2c(a).

If we take the square of the above formula for the total Bismut supercon-
nection, we see that
A2 As + ass ds,

which has Chern character

ch(A) = ch(A9) - a(s) ds.

The theorem will follow from the local family index theorem applied to the
Bismut superconnection A. By this theorem, we know that b,' Stre (e-tF_) E
A(B) has an asymptotic expansion in t as t --+ 0 of the form

SB Str(e-tF) = Str(e--7et) - Str(aa"St e-tee°, ds

00

tj12(O,j/2 - 01j/2 ds).

.7=0

where Oj/2 E A(B) and aj/2 E A(B) are differential forms on b which do
not involve ds. Using the fact that
aAst aAt
as s=1 at

we may set s = 1, and we obtain the asymptotic expansion

00
a Ate ,fit ti/2-1a.7/2(x,s = 1).
Str
( at
du
J=O
10.6. The Curvature of the Determinant Line Bundle 347

To complete the proof, we must show that ao = 0. We use the explicit

formula for 00 - ao ds furnished by the family index theorem:

co - ao ds = (27ri)-n12 M_/B_A(M/B) ch(E/S).

But the differential form A(M/B) ch(E/S) E A(M) is the pull-back to M of

the differential form A(M/B) ch(E/S) E A(M); hence, it does not involve ds,
and we see that ao vanishes.

10.6. The Curvature of the Determinant Line Bundle

In Section 9.7, we defined the determinant line bundle det(7r*E, D), its Quillen
metric I IQ, whose definition involves the zeta-function determinant of the op-
erators D- D+ restricted to the eigenspace (A, oo), and its natural connection.
In this section, following Bismut and Freed, we will calculate the curvature of
the natural connection on det(7r,,E, D), when D is a family of Dirac operators
associated to a connection on S.
If £ is a finite-dimensional superbundle, there is a relationship between the
Chern character of E with respect to a connection Ve and the curvature of its
determinant line bundle det(E) with respect to the corresponding connection
Vdet(E):
(Vdet(8))2 = -Str((V )2) = ch(Ve)[2].
We will see a similar relationship in infinite dimensions; the curvature of the
connection that we described in Section 9.7 on the line bundle det(ir*£, D) will
be shown to equal

tli mch(At)[2] = (21ri)-n12 ( fMIB A(M/B) ch(E/S)) [2]

Let us first consider a special case, in which ker(D) is a vector bundle, so

det(ir,,E, D) and det(ker(D)) may be identified. In this case, the connec-
tion Vo = POA[1]Po on ker(D) is compatible with the L2-metric, and Theo-
rem 10.32 shows that
00
ch(Vo) = (2i)2 A(M/B) ch(£/S) - d 10 a(s) ds.
f
M/B 0

If we take the two-form component of this formula, we see that

(Vdet(ker(D))) 2 = (21ri) -n/2 (J MI A(M/B) ch(£/S)) - df a(s) [1] ds,
B [2]

from which we see immediately that

oo
Vdet(ker(D)) + f a(s)[i]
l
ds
0
348 10. The Family Index Theorem

is a connection on the bundle det(ker(D)) with curvature

(27ri)-n/2(f A(M/B) ch(E/S)) [2].

This connection is compatible with the Quillen metric. It would be very

interesting to have similar geometric interpretations of the other components
of the transgression formula.
Recall that the Quillen metric is defined on the representative det(f[o,,\))
of the line bundle det(7r*£, D) over Ua as

I IQ =e- C(0,D D+,a)/2I . I,

where I I
is the L2-metric on det (7-l[o,A)) .
The operator over the open set Ua has the property that
the bundle ker(D(A,,,,,)) equals 7-l[o,A). We may define a superconnection Aa
by the formula
Aa = D(a,oo) + V----.
Similarly, we define the rescaled version of this superconnection

A,\s = S1/2D(a oo) + Q7r.e

Define two one-forms c (s,A)[1] E .A'(B), given by traces over lr*£+ and
7r*£- respectively:

± A)[1] =
a(s, (aA 2

ass e-
By Lemma9.42, we see that a+(s, A)[l] = a-(s,,\)[1].
As in Section 9.7, we denote the one-forms 2 LI Mt-.o f too at (s, A) [1] ds by
,Oa ; these satisfy p\+ = j3 . Thus, the sum i3X + Qa is real, while the
difference 0,+\ -#a is imaginary. Also, we proved that

-d('(0, D- D+, \) = /3a +,3T.

Thus, we see that the sum of differential forms Qa + 8a is an exact real

one-form.
We defined the connection Vdet(7,.e) on det(ir*E, D) as follows: on the
bundle det(f[o,a)), we took the connection
Vdet(%[o,a)) + Qa+

Our goal is to calculate the curvature of this connection. In the following

lemma, we calculate the exterior differential of the imaginary one-form /3,+ -
18; -
10.6. The Curvature of the Determinant Line Bundle 349

Lemma 10.34. The difference a+(s,.X) [i] - a- (s,,\) [1] is bounded by a mul-
tiple of s-1/2 in A(UA) for small s, and hence

2 00 (a+ (s, a) - a- (s, A)} [1] ds.

J0
The exterior differential of this differential form is given by the formula

d((3a - 3) + ch (V [o'a)) [2] _ (27ri)-"`/2 M/Ua

(r A(M/B)
eh(-IS)) [2]

Proof. We have seen that

a, (s, A) [1] = - 21 T r".6± (D [V *e, D] e-8D2P(a m))

Since the Bismut superconnection A coincides with D + V"'e up to terms of
degree 2, we see that

a (s)[1] _"'e} (as a ) [1]

2).
= - Z Tr".et (D [V" e, D] a-sD
From this, it follows that
aAs 9 e-AA,s
\
Tr".et I (9s9 I
= - Tr".e} (D {0"'e, D)e sD2P[o,a))
\ [i]

The operator P[o,a) is a smoothing operator, so that by Lemma 9.47, for s

small,
(aAass
a±(s,A)[i] - Tr".ef e -A.)
llt < C(B).
[il
The result now follows from Theorem 10.32.
We now arrive at the main result of this section (this is what is known to
physicists as the "anomaly formula").
Theorem 10.35. The curvature of the connection Vdet(".e,D) equals

(27ri) -n/2 (J A(M/B) ch(E/S))

[2]

Proof. We have already done all of the work. Indeed,

(Vdet(or.e,D))2 = (Vdet(9-l[o,a)) + )3A+) 2
= (Vdet(r[[o,ai))2 + d/3a
(l

= ((Vdet(%[o,a)))2 + 2d(/3a - as ),
since the differential form 0,+ + (3a is closed. The theorem follows from
Lemma 10.34.
350 10. The Family Index Theorem

10.7. The Kirillov Formula and Bismut's Index Theorem

In this section, we will show that the local family index theorem, in the case
in which the family M -+ B is associated to a principal bundle P -+ B with
compact structure group G, is equivalent to the Kirillov formula of Chapter 8.
All of our notations are as in Section 9.4.
We start by calculating the Bismut superconnection A. The vertical tan-
gent bundle T(M/B) = P xG TN has a connection VN"B, obtained by com-
bining the Levi-Civita connection VN on TN with the connection 0 on the
principal bundle P -> B, as explained in Section 7.6. On the other hand, the
connection 9 determines a connection on the fibre bundle M -> B, and hence,
by the construction of Section 1, a connection VMIB on T(M/B).
Lemma 10.36. The two connections VN,e and VM/B onT(M/B) are equal.
Proof. Since this formula is local, we assume that P is the trivial bundle P =
B x G, with connection form 9 = >' 19kXk E A(B, g), so that M = B x N.
If V is a vector field on B, its horizontal lift is V + 9(V)N. If Y and Z are
vector fields on N, then by the definition of VM/B, we see that

2(VV+9(V)NY, Z) = ([9(V)N, Y], Z) + ([Z, 9(V)N],Y) + 9(V)N(Y, Z)

= 2([8(V)N,Y], Z),
where we have used the fact that the vertical metric on M is independent of
the point z in the base B, and that 9(V)N is a Killing vector field. On the
other hand, Lemma 7.37 shows that

V °e(V)NY = V e(V)NY + N'N(0(V)N)Y

= [9(V)N,Y].
In the same way, the bundle £ = P xG E has a connection VE,8, obtaine
by combining the Clifford connection VE on E -+ N with the connection 9.
Using the fact that for any Y E g and a E r(N, T*N),
[V E (Y), c(a)] = c(µ' (Y)a)
we see that VE,e is a Clifford connection on £ compatible with VN,e = VMIB
We denote this connection by VE.
The equivariant Chern-Weil homomorphism 0e maps the equivariant de
Rham complex (AG(N), de) to the de Rham complex (A(M), d). Using the
fact that VN,e and VE,e equal OMB and VE respectively, we see that
oe(Ag) = A(M/B),
Oe (chg (E/S)) = ch(£/S),

where A. and chg(E/S) are the equivariant A-genus of N and the equivariant
relative Chern character of E.
10.7. The Kirillov Formula and Bismut's Index Theorem 351

In the fibre bundle M -> B, the tensor S is equal to zero, and hence
its trace k is also zero. To calculate the Bismut superconnection, we must
describe the curvature tensor Il of M.

Lemma 10.37. If V and W are two vector fields on B, then

fl(V,W) _ -cP(VP,WP)N,
where VP and Wp are the horizontal lifts of V and W to P, flp E A2 (p, g)
is the curvature of P, and for any element X E Z, XN is the corresponding
vector field on N, as well as the induced vertical vector field on M = P xGN.

Since the bundle 7r*E is the associated bundle to the representation of G

on IF (N, E), it has a canonical connection given on the space
A(B, itE) = (A(P) (& P(N, E))bas
by the formula
m
V's _dp+EOkfE(Xk).
k=1
This connection has curvature
m
(v.7r.E)2
fk J E(Xk).
k=1

There is a linear map c : g -+ r(N, End(E)), defind by sending X E g

to C(XN), the operator of Clifford multiplication by the vector field XN E
F(N, TN) associated to X. By means of the formula
m
c(IP) Ok C(Xk),
k=1

we obtain an action of the curvature Q p of P on A(P) ® r(N, E).

Theorem 10.38. (1) The Bismut superconnection of the bundle P XG N is

the restriction to (A(P) ® r(N, E)) bas of the operator
1®DN+Vlr.E+.1c(1p)

(2) The Chern character ch(At) = E A(B) is independent oft.

Proof. From the formula for A given in Proposition 10.15, we see that A(01 _
DN and A[21 = 4c(SZp). To check that

jJ CaVEa = Vir.E
a
352 10. The Family Index Theorem

we may assume that P is trivial, as in the proof of Lemma 10.36. We see that

E«QE
« _ E«vaEs9+9(8 )N
a
_ I ILN(0(8«)N))
«

= E--«L'(9( )N)

_ > 6kGE(Xk)-
k

The connection W.E commutes with DN, since DN is G-invariant. Fur-

thermore, and c(Stp) commute, by the Bianchi identity V 1p = 0. If we
let Qt be the operator

Qt = At - va,E = t1/2D N + 4t 1/2c(St) E r(B , Endv ( ?r* £)) ,

we see that
dAt dAt 1 dAt t, Qt
dt = [ dt
, At] =l d ]
and hence that
-A2 ) -A,
dt
St r(e _ - St r ( Q te 0,
[ , /
since the supertrace vanishes by Lemma 3.49. O
We have seen in Section 9.4 that Kirillov's formula for indG(e-X , DN) im-
plies that

f
(27ri)-n/2
tlimo A(M/B) ch(e/S).
B

We see from the above theorem that the formula

Str(e-A=) (27ri)-n/2
_ A(M/B) ch(£/S),
tl1me
J s
is, in the case of a compact structure group, a consequence of the Kirillov
formula for the equivariant index of DN. We will strengthen this result by
showing that the the local family index theorem of Bismut is implied in this
context by the local Kirillov formula for the operator K(X) = (D+4c(X))2+
£E(X) of Section 8.3.
The operator K(X) is a differential operator depending polynomially on
X, and hence acts on the space C[2] 0 r(N, E), by the formula
(Ks)(X) = K(X)s(X).
10.7. The Kirillov Formula and Bismut's Index Theorem 353

If p is a point in the principal bundle P lying in the fibre of z E B, we may

specialize the operator K(X) to an operator K(S1) acting on A2T .,r(p)B
r(N, E), by replacing X = Fk 1 akXk E g, ak E C, by
M
Qp = E A2T ir(p)B ®g.
i=1

On the other hand, the point p E P gives an identification of the fibres MZ

with N, so that we may also consider the curvature A2 of the Bismut super-
connection to be a differential operator Ap2 acting on A2T*,r(p)B ® P(N, E).

Proposition 10.39. The two operators K(1p) and 2

are equal.

Proof. This follows from Theorem 10.38 and the fact that supercom-
mutes with DN and c(cl):

A2 = (D + 4C(fP))2 + (V7r.E)2
_ (D + IC(gp))2 + L(clp).

From this lemma, we see that the coefficients of the asymptotic expansion
of (p I e-1A2 I p) depend polynomially on 52k, and are obtained from the co-
efficients of the asymptotic expansion of (p I e-tx(x) I p) by replacing X by
c p.
The converse result is also true: the local family index theorem implies the
local Kirillov formula. To show this, we use the following lemma from the
theory of classifying spaces.

Lemma 10.40. There exist fibre bundles PN -> BN for which the Chern-
Weil map C[g]G -+ A(BN) is an injection up to polynomials of degree N, for
anyN>0.
Proof. Let V be a faithful unitary representation of G. From the embedding

U(V (D CN) X U(V (& CN) C U(V (& C2N),

we see that U(V (&CN) acts freely on PN = U(V (gC2N)/ U(V (&C"T ). There is
an embedding of G in U(V ®CN), in which we map g E G to g®I. Let BN =
PN/G. The result follows from the fact that if m > n, iri(U(m)/U(n)) = 0
for i < 2n.

The local index theorem for families for the bundle PN -+ BN implies the
local Kirillov formula for K(X) up to order N. Since N is arbitrary, we
obtain the local Kirillov formula to all orders.
354 10. The Family Index Theorem

Bibliographic Notes
Sections 1-4. The results of the first four sections of this chapter are due to
Bismut [30], although the proof of the main result (Theorem 10.21), the local
family index theorem, follows our proof of the local index theorem in Chap-
ter 4 of this book. In his work, Bismut uses the term Levi-Civita superconnec-
tion where, for obvious reasons, we prefer the term Bismut superconnection.
Theorem 10.19 is new, and gives a natural interpretation of the Bismut super-
connection. For closely related approaches to the family index theorem, see
the articles of Berline-Vergne [26], Donnelly [51] and Zhang [108].
Sections 5 and 6. The "anomaly" formula for the curvature of the natural
connection on the determinant line bundle was proved by Witten in the special
case where there are no zero modes (see for example [106]), based on the work
of many theoretical physicists. Further special cases were proved by Atiyah-
Singer [17] and Quillen [92], and the theorem is stated and proved in full
generality in the article of Bismut and Freed [35]. In this chapter, it appears
in a generalized form inspired by the article of Gillet and Soule [67].
Section 7. The relationship between the family index theorem for a family
with compact structure group and the equivariant index theorem is in Atiyah
and Singer [16]. The remarkable observation that such a relationship holds at
the level of differential operators, as in Proposition 10.39, is due to Bismut [30].
For an extension of the results of this chapter in the direction of algebraic
geometry, see Bismut, Gillet and Soule [36].
Bibliography

[1] L. Alvarez-Gaume. Supersymmetry and the Atiyah-Singer index theorem. Comm.

Math. Phys. 90 (1983), 161-173.
[2] M. F. Atiyah. Collected works, Clarendon Press, Oxford, 1988.
[3] M. F. Atiyah. K-theory, Benjamin, New York, 1967.
[4] M. F. Atiyah. Circular symmetry and stationary phase approximation.
131 (1985), 43-59.
[5] M. F. Atiyah and R. Bott. A Lefschetz fixed point formula for elliptic complexes. I.
Ann. Math. 86 (1967), 374-407.
[6] M. F. Atiyah and R. Bott. A Lefschetz fixed point formula for elliptic complexes. II.
Applications Ann. Math. 88 (1968), 451-491.
(7] M. F. Atiyah and R. Bott. The moment map and equivariant cohomology. Topology
23 (1984), 1-28.
[8] M. F. Atiyah, R. Bott and V. K. Patodi. On the heat equation and the index theorem.
Invent. Math. 19 (1973), 279-330. Errata, Invent. Math. 28 (1975), 277-280.
[9] M. F. Atiyah, R. Bott and A. Shapiro. Clifford modules. Topology 3, Suppl.1 (1964),
3-38.
[10] M. F. Atiyah and F. Hirzebruch. Riemann-Roch theorems for differentiable mani-
folds. Bull. Amer. Math. Soc. 65 (1959), 276-281.
[11] M. F. Atiyah, V. K. Patodi and I. M. Singer. Spectral assymetry and Riemannian ge-
ometry: I. Math. Proc. Camb. Phil. Soc. 77 (1975), 43-69.
[12] M. F. Atiyah and G. B. Segal. The index of elliptic operators II. Ann. Math. 87
(1968), 531-545.
[13] M. F. Atiyah and I. M. Singer. The index of elliptic operators on compact manifolds.
Bull. Amer. Math. Soc. 69 (1963), 422-433.
[14] M. F. Atiyah and I. M. Singer. The index of elliptic operators I. Ann. Math. 87 (1968),
484-530.
[15] M. F. Atiyah and I. M. Singer. The index of elliptic operators III. Ann. Math. 87
(1968), 546-604.
[16] M. F. Atiyah and I. M. Singer. The index of elliptic operators IV. Ann. Math. 93
(1971), 119-138.
[17] M. F. Atiyah and I. M. Singer. Dirac operators coupled to vector potentials. Proc.
Nat. Acad. Sci. U. S. A. 81 (1984), 2597-2600.
[18] P. Baum and J. Cheeger. Infinitesimal isometries and Pontryagin numbers. Topology
8 (1969), 173-193.
[19] F. A. Berezin. The method of second quantization. Pure and applied physics, v. 24.
Academic Press, NewYork, 1966.
[20] M. Berger, P. Gauduchon and E. Mazet. Le spectre dune varietg riemannienne,
Lecture Notes in Mathematics 194, Springer-Verlag, Berlin-Heidelberg-New York,
1971.
356 Bibliography

[21] N. Berline and M. Vergne. Classes caracteristiques equivariantes. Formules de local-

isation en cohomologie equivariante. C. R. Acad. Sci. Paris 295 (1982), 539-541.
[22] N. Berline and M. Vergne. Fourier transforms of orbits of the coadjoint representa-
tion. In Proceedings of the conference on representation theory of reductive groups,
(Park City, Utah, 1982), Progress in Mathematics, vol. 40, Birkauser, Boston, 1983,
pp. 53-75.
[23] N. Berline and M. Vergne. Zeros d'un champ de vecteurs at classes characteristiques
equivariantes. Duke Math. Journal 50 (1983), 539-549.
[24] N. Berline and M. Vergne. The equivariant index and Kirillov character formula.
Amer. J. Math. 107 (1985), 1159-1190.
[25] N. Berline and M. Vergne. A computation of the equivariant index of the Dirac op-
erator. Bull. Soc. Math. France 113 (1985), 305-345.
[26] N. Berline and M. Vergne. A proof of the Bismut local theorem for a family of Dirac
operators. Topology 26 (1987), 435-463.
[27] N. Berline and M. Vergne. L'indice equivariant des operateurs transversalement el-
liptique. To appear, Invent. Math.
[28] J.-M. Bismut. The Atiyah-Singer theorems: a probabilistic approach. J. Panct. Anal.
57 (1984), 56-99.
[29] J: M. Bismut. The Atiyah-Singer theorems: a probabilistic approach. II. J. F1unct.
Anal. 57 (1984), 329-348.
[30] J.-M. Bismut. The index theorem for families of Dirac operators: two heat equation
proofs. Invent. Math. 83 (1986), 91-151.
[31] J: M. Bismut. Localization formulas, superconections, and the index theorem for
families. Comm. Math. Phys. 103 (1986), 127-166.
[32] J.-M. Bismut. The infinitesimal Lefschetz formulas: a heat equation proof. J. Funct.
Anal. 62 (1985), 435-457.
[33] J: M. Bismut and J. Cheeger. q-invariants and their adiabatic limits. J. Amer. Math.
Soc. 2 (1989), 33-70.
[34] J: M. Bismut and D. S. Freed. The analysis of elliptic families: Metrics and connec-
tions on determinant bundles. Comm. Math. Phys. 106 (1986), 159-176.
[35] J: M. Bismut and D. S. Freed. The analysis of elliptic families: Dirac operators, eta
invariants, and the holonomy theorem of Witten. Comm. Math. Phys. 107 (1986),
103-163.
[36] J: M. Bismut, H. Gillet and C. Soule. Analytic torsion and holomorphic determinant
bundles. Comm. Math. Phys. 115 (1988), 49-78, 79-126, 301-351.
[37] J. Block and E. Getzler. Equivariant cyclic homology and equivariant differential
forms. Ann. Sci. Ecole Norm. Sup. (4) 27 (1994), 493-527.
[38] A. Borel and F. Hirzebruch. Characteristic classes and homogeneous spaces.
Amer. J. Math. 80 (1958), 458-538.
[39] A. Borel and J.-P. Serre. Le theoreme de Riemann-Roch (d'apres Grothendieck).
Bull. Soc. Math. France 86 (1958), 97-136.
[40] J. B. Bost. Fibres determinants, determinants regularises et mesures sur les espaces
de modules de courbes complexes, In SQminaire Bourbaki, Asterisque 152-153
(1987), pp. 113-149.
[41] R. Bott. Homogeneous vector bundles. Ann. Math. 66 (1957), 203-248.
[42] R. Bott. Vector fields and characteristic numbers. Michigan Math. J. 14 (1967),
231-244.
Bibliography 357

[43] R. Bott. A residue formula for holomorphic vector fields. J. Diff. Geom. 1 (1967),
311-330.
[44] R. Bott and L. Tu. Differential forms in algebraic topology. Graduate Texts in Math.
82, Springer-Verlag, Berlin-Heidelberg-New York, 1982.
[45] H. Cartan. La transgression dans un groupe de Lie et dans un espace fibre principal.
In Colloque de Topologie, C. B. R. M., Bruxelles, 1950, pp. 57-71.
[46] H. Cartan. Theoreme d'Atiyah-Singer sur 1'indice d'un operateur differentiel.
Seminaire Cartan-Schwartz, 16e annde: 1963/64, Secretariat mathSmatique, Paris,
1965.
[47] S. S. Chern. A simple intrinsic proof of the Gauss-Bonnet formula for closed Rie-
mannian manifolds. Ann. Math. 45 (1944), 741-752.
[48] S. S. Chern and J. Simon. Characteristic forms and geometric invariants. Ann. Math.
99 (1974), 48-69.
[49] C. Chevalley. The algebraic theory of spinors, Columbia University Press, New York,
1954.
[50] A. Connes and H. Moscovici. The L2-index theorem for homogeneous spaces of Lie
groups. Ann. Math. 115 (1982), 291-330.
[51] H. Donnelly. Local index theorem for families. Michigan Math. J. 35 (1988), 11-20.
[52] H. Donnelly and V. K. Patodi. Spectrum and the fixed point set of isometries. Topol-
ogy 16 (1977), 1-11.
[53] M. Duflo. Operateurs transversalement elliptiques et former differentielles
dquivariantes. Sem. Bourbaki (1994-95), no. 791.
[54] M. Duflo, G. Heckman and M. Vergne. Projection d'orbites, formule de Kirillov et
formule de Blattner. Mem. Soc. Math. Prance, 15 (1984), 65-128.
[55] M. Duflo and M. Vergne. Orbites coadjointes et cohomologie equivariante. In The or-
bit method in representation theory, Progress in Mathematics, vol. 82, Birkhauser,
Boston, 1990.
[56] M. Duflo and M. Vergne. Cohomologie 6quivariante et descente. Asterisque 215
(1993), 5-108.
[57] J. J. Duistermaat and G. Heckman. On the variation in the cohomology of the sym-
plectic form of the reduced phase space. Invent. Math. 69 (1982), 259-268. Adden-
dum, 72 (1983), 153-158.
[58] A.Erdelyi. Asymptotic expansions, Dover, New York, 1956.
[59] D. Friedan and P. Windey. Supersymmetric derivation of the Atiyah-Singer index
theorem. Nucl. Phys. B235 (1984), 394-416.
[60] E. Getzler. Pseudodifferential operators on supermanifolds and the index theorem.
Comm. Math. Phys. 92 (1983), 163-178.
[61] E. Getzler. A short proof of the Atiyah-Singer index theorem. Topology 25 (1986),
111-117.
[62] E. Getzler. The Bargmann representation, generalized Dirac operators and the index
of pseudodifferential operators on R1. To appear, Contemporary Math.
[63] P. B. Gilkey. Curvature and the eigenvalues of the Laplacian for elliptic complexes.
Adv. in Math. 10 (1973), 344-382.
[64] P. B. Gilkey. Curvature and the eigenvalues of the Dolbeault complex for Kaehler
manifolds. Adv. in Math. 11 (1973), 311-325.
[65] P. B. Gilkey. Lefschetz fixed point formulas and the heat equation. In Partial diff-
erential equations and geometry. Proceedings of the Park City conference, 1977,
ed. C. Byrnes, Lecture notes in pure and applied math., vol. 48, Marcel Dekker,
NewYork-Basel, 1979, pp. 91-147.
358 Bibliography

[66] P. B. Gilkey. Invariance, the heat equation and the Atiyah-Singer index theorem,
Publish or Perish, Washington, 1984.
[67] H. Gillet and C. Soule. Analytic torsion and the arithmetic Todd genus. Topology 30
(1991), 21-54.
[68] Harish-Chandra. Fourier transforms on a Lie algebra, I. Amer. J. Math. 79 (1957),
193-257.
[69] F. Hirzebruch. Topological methods in algebraic geometry, 3rd ed., Grundlehren der
mathematische Wissenschaften 131, Springer-Verlag, Berlin-Heidelberg-New York,
1966.
[70] F. Hirzebruch. The signature theorem: reminiscences and recreation. In Prospects
in mathematics, Annals of Mathematics Study 70, Princeton Univ. Press, 1971, pp.
3-31.
[71] M. Karoubi. K-theory, an introduction, Grundlehren der mathematische Wis-
senschaften 226, Springer-Verlag, Berlin-Heidelberg-New York, 1978.
[72] A. A. Kirillov. Method of orbits in the theory of unitary representations of Lie groups.
Funktsional Anal. i Prilozhen. 2 (1968), no. 1, 96-98.
[73] A. A. Kirillov. Elements of the theory of representations, Grundlehren der mathe-
matische Wissenschaften 220, Springer-Verlag, Berlin-Heidelberg-New York, 1976.
[74] S. Kobayashi and K. Nomizu. Foundations of differential geometry, J. Wiley and
Sons, New York-London-Sydney, 1963.
[75] B. Kostant. Quantization and unitary representations. In Modern analysis and ap-
plications, Lecture Notes in Mathematics 170, Springer-Verlag, Berlin-Heidelberg-
New York, 1970, pp. 87-207.
[76] T. Kotake. The fixed-point formula of Atiyah-Bott via parabolic operators. Comm.
Pure Appl. Math. 22 (1969), 789-806.
[77] T. Kotake. An analytical proof of the classical Riemann-Roch theorem. In Global
analysis (Proc. Sympos. Pure Math. 16), Amer. Math. Soc., Providence, 1970, pp.
137-146.
[78] J. D. Lafferty, Y. L. Yu and W. P. Zhang. A direct geometric proof of the Lefschetz
fixed point formulas. Trans. Amer. Math. Soc. 329 (1992), 571-583.
[79] H. B. Lawson, Jr. and M-L. Michelson. Spin geometry, Princeton Mathematical Se-
ries 38, Princeton Univ. Press, Princeton, 1989.
[80] A. Lichnerowicz. Spineurs harmoniques. C. R. Acad. Sci. Paris. Sdr. A 257 (1963),
7-9.
[81] V. Mathai and D. Quillen. Superconnections, Thom classes and equivariant differen-
tial forms. Topology 25 (1986), 85-110.
[82] H. McKean and I. M. Singer. Curvature and the eigenvalues of the Laplacian. J. Diff.
Geom. 1 (1967), 43-69.
[83] J. Milnor. Morse theory, Annals of Mathematics Study 51, Princeton Univ. Press,
Princeton, 1963.
[84] J. W. Milnor. Remarks concerning spin manifolds. In Differential and combinatorial
topology, a symposium in honor of Marston Morse, S. S. Cairns (ed.), Princeton
Univ. Press, Princeton, 1965, pp. 55-62.
[85] J. W. Milnor and J. D. Stasheff. Characteristic classes, Annals of Mathematics Study
76, Princeton Univ. Press, Princeton, 1974.
[86] S. Minakshisundaram and A. Pleijel. Some properties of the eigenfunctions of the
Laplace operator on Riemannian manifolds. Canad. J. Math. 1 (1949), 242-256.
Bibliography 359

[87] R. S. Palais, ed. Seminar on the Atiyah-Singer index theorem, Annals of Mathemat-
ics Study 57, Princeton Univ. Press, Princeton, 1965.
[88] V. K. Patodi. Curvature and the eigenforms of the Laplace operator. J. Difj. Geom.
5 (1971), 233-249.
[89] V. K. Patodi. An analytic proof of the Riemann-Roch-Hirzebruch theorem. J. Difj.
Geom. 5 (1971), 251-283.
[90] V. K. Patodi. Holomorphic Lefschetz fixed point formula. Bull. Amer. Math. Soc.
79 (1973), 825-828.
[91] D. Quillen. The spectrum of an equivariant cohomology ring I and II. Ann. Math.
94 (1971), 549-602.
[92] D. Quillen. Determinants of Cauchy-Riemann operators over a Riemann surface.
Funktsional Anal. i Prilozhen. 19 (1985), no. 1, 37-41.
(93] D. Quillen. Superconnections and the Chern character. Topology 24 (1985), 89-95.
[94] D. B. Ray and I. M. Singer. R-torsion and the Laplacian on Riemannian manifolds.
Adv. Math. 7 (1971), 145-210.
[95] D. B. Ray and I. M. Singer. Analytic torsion for complex manifolds. Ann. Math. 98
(1973), 154-177.
[96] J. Roe. Elliptic operators, topology and asymptotic methods, Pitman Res. Notes in
Math. Series 179, Longman Scientific and Technical, Harlow, 1988.
[97] W. Rossman. Kirillov's character formula for reductive groups. Invent. Math. 48
(1978), 207-220.
[98] W. Rossman. Equivariant multiplicities on complex varieties. Astdrisque 113-114
(1989),313-330.
[99] R. Seeley. Complex powers of an elliptic operator. In Proc. Sympos. Pure Math. 10,
Amer. Math. Soc., Providence, 1967, pp. 288-307.
(100] R. Seeley. A proof of the Atiyah-Bott-Lefschetz fixed point formula. An. Acad. Brasil
Cience 41 (1969), 493-501.
[101] G. B. Segal. Equivariant K-theory. Inst. Hautes Etudes Sc. Publ. Math. 34 (1968),
129-151.
[102] M. Vergne. Formule de Kirillov et indice de l'operateur de Dirac. In Proceedings
of the International Congress of Mathematicians, Warszawa, 1983, pp. 921-93.4,
North Holland, Amsterdam-New York Oxford, 1984.
[103] M. Vergne. Geometric quantization and equivariant cohomology. In First European
Congress of Mathematics. I. pp. 249-295. Progress in Mathematics, Birkhaiiser,
Basel-Boston-Berlin.
[104] R. O. Wells, Jr. Differential analysis on complex manifolds. Graduate texts in math-
ematics 65, Springer-Verlag, New York, 1980.
[105] E. Witten. Supersymmetry and Morse theory. J. Dif. Geom. 17 (1982), 661-692.
[106] E. Witten. Global gravitational anomalies. Comm. Math. Phys. 100 (1985), 197-
229.
[107] Y. L. Yu. Local index theorem for Dirac operator. Acta Math. Sinica, New Series
3(2) (1987), 152-169.
[108] W. P. Zhang. The local Atiyah-Singer index theorem for families of Dirac opera,
tors. In Differential geometry and topology, Lecture Notes in Mathematics 1369,
Springer-Verlag, Berlin-Heidelberg-New York, 1989, pp. 351-366.
List of Notations

r(M, s) space of smooth sections of £ 14

r ,(M, E) space of compactly supported smooth sections of £ 14
GL(S) frame bundle of E 16
£(X) Lie derivative acting on vector fields 17
G£ (X) Lie derivative acting on sections of E 16
A(M) space of differential forms on M 17
A, (M) space of compactly supported differential forms on M 17
aN i-form component of a 17
lal degree of a 17
t(v) contraction 18, 69
e(a) exterior multiplication by a 18
H' (M) de Rham cohomology of M 18
H, (M) compactly supported de Rham cohomology of M 19
A(M, £) differential forms on manifold M with values in £ 19, 43
VE1®E2 tensor product of covariant derivatives 25
Try (t) parallel translation along smooth path y 26
7Z Euler (radial) vector field 26, 52
IA11/2 bundle of half-densities 30
i*a integral of a along the fibres of a fibre bundle 31
fM/B a another notation for i*a 31
O(M) orthonormal frame bundle of Riemannian manifold M 32
so(TM) O(M) x0(n) so(n) 33
R(X, Y) Riemannian curvature 33
Ric Ricci curvature 34
rM scalar curvature 34
IdxI Riemannian density 34
grad f gradient of the function f 34
x point in the normal coordinate system around x0 35
j (x) Jacobian of exponential map around xo 38
[a, b] supercommutator 39
Str supertrace 39
det(S) determinant line-bundle of £ 41
T(a) Berezin integral 42
Pf (A) Pfaffian of A 42
List of Notations 361

det1/2 (A) another notation for Pf (A) 43

A superconnection 44
A[i] degree i component of A 45
ch(A) Chern character form of A 49, 285
ch(S) Chern character of £ 50
A(M) A-genus of M 50
Td(M) Todd genus of the almost-complex manifold M 51
X(M) Euler class of the oriented manifold M 58
Eul(M) Euler number of M 59
DE Laplacian associated to £ 66
D* formal adjoint of D 68
d*a divergence of a 68, 127
qt (x, y) heat kernel of Euclidean space 72
asymF
fH
asymptotic expansion of a Gaussian integral 73, 173
P-°° (M, £) space of generalized sections of £ 74
r, (M, £) space of CE-sections of £ 74
Si ® £2 external tensor product of £1 and E2 74
(x I P I y) kernel of the operator Pat (x, y) 74
pt(x, y, H) heat kernel of Laplacian H 75, 87
Ak k-simplex 77
-r (X, y) parallel translation along geodesic joining x and y 83
IIAIIHS Hilbert-Schmidt norm of A 87
FL2 (M, £ (9 JAI1/2) space of square-summable sections of £ (9 JAI 1/2 88
Tr(K) trace of K 88
Gk k-th Green kernel of H 93
C(V) Clifford algebra of V 103
c(v) action of v E V on a Clifford module 104
a (a) symbol map from C(V) to AV 104
Ci (V) degree i subspace of Clifford algebra 104
T canonical homomorphism from C2(V) to So (V) 105
Spin(V) double cover of SO(V) 106
iv 2(X) detl/2 (since 2 2) 108
r chirality element/of C(V) 109
S = S+ ®S- spinor representation of C(V) 109
Endc(v) (E) endorphisms of Clifford module E over C(V) 112
StrE/s relative supertrace for Clifford module E 113
.A(X) action of U(V) on AT1,OV 113
C(M) Clifford bundle of Riemannian manifold M 113
Spin(M) principal Spin(n)-bundle of spin-manifold M 114
Endc(M) (6) endorphisms of Clifford module E over C(M) 114
S spinor bundle 115
D Dirac operator 115
FE/S twisting curvature of Clifford module £ 120
RE Clifford action of Riemannian curvature on 9 120
362 List of Notations

ind(D) index of D 123

Hodge star 131
cr(M) signature of M 133
kt (x, x) heat kernel of D2 restricted to the diagonal 143
6'a rescaling of a 144
Stre/s relative supertrace for Clifford module £ 146
ch(£/S) relative Chern character of Clifford module 6 146
Cas Casimir of a representation 172
C[g] polynomials on the Lie algebra g 179
Mry fixed point set of 'y 184
indH(-y, D) equivariant index of D 185
N normal bundle of MY C M 190
chH (y, A) localized Chern character of A 193
chH(y, £/S) localized twisting Chern character of £ 194
AG(M) space of equivariant differential forms 206
dg equivariant exterior differential 206
Hc(M) equivariant de Rham cohomology 207
AG(M, £) equivariant differential forms with values in 6 208
Ag equivariant superconnection 208
Fg(X) equivariant curvature 208
µ(X) moment map 209
chg (A) equivariant Chern character 210
AM(X) Riemannian moment 211
symplectic form 212
A(X) symplectic moment 212
Ma zero set of vector field X 215
N normal bundle of MO C M 215
xg (N) equivariant Euler class of N 216
d/3 Liouville form 224
Ag (X, M) equivariant A-genus 246
AelS(X) twisting moment 248
,e/S equivariant twisting curvature 248
g
chg (X, £/S) equivariant relative Chern character of £ 248
H(X) Bismut Laplacian 256
V0 projection of a superconnection A onto ker(D) 270, 286
At rescaling of the superconnection A 271, 286
'Ft curvature of At 271, 286
a(t) transgression form of A 276, 289
(Mz I zEB) family of manifolds 277
T(M/B) vertical tangent bundle 277
vertical density bundle 277
family of vector bundles 277
vector bundle of sections along fibres of £ 277
I Z E B) family of Dirac operators 278
List of Notations 363

n1o a) subbundle of zr*£ on which D2 < A 278

Plo,a) spectral projection onto D2 < A 289
Ua subset of B on which A is not an eigenvalue of D2 279
ind(D) family index of D 280, 298
7-l(a,,U) subbundle of on which A < D2 < µ 281
P(a,v) spectral projection onto A < D2 < µ 281
vW.E natural connection on ir,,9 282
A(M/B) A-genus of T (M/B) 293, 336
M If] Mellin transform of f 299
LIM f (t) renormalized limit of f 300
t
(s, H, A) zeta function of H 301
det(H) zeta function determinant of H 301
det(7r*E, D) determinant line bundle associated to the family D 305
det(D+) determinant of D+ (section of det(ir*E, D)) 306
1
- IQ Quillen metric 306
Qdet(7r.E,D) natural connection on det(ir*£, D) 308
Index

adjoint, formal 67, 89 Cartan subalgebra 227

affine connection 28 canonical line bundle 140
A-genus 50, 144 Casimir operator 172
equivariant 211 character formula
almost-complex manifold 135 Kirillov 252
anomaly formula 349 Weyl 251
associated bundle 15 characteristic class, 47
Atiyah, M. F. 147, 150, 189, 195, A-genus 50, 147, 151, 336
243, 295, 338 Chern character 50, 285
Atiyah-Bott fixed point formula 189 equivariant 210
Atiyah-Segal-Singer fixed point Euler class 58, 149
formula 195 L-genus 150
Todd genus 51, 152
characteristic number
Baum, P. 223
Bott's formula 223
Berezin integral 42, 54, 106, 109,
Cheeger, J. 223
146, 195, 241
Chern, S. S. 59
Bianchi identity 44
Chern character 50, 285
Bismut, J.-M. 256, 282, 331, 337, equivariant 210
347
relative 146
Bismut
equivariant 248
Laplacian 256 localized 194
superconnection 331 Chern-Weil homomorphism 47
Bochner-Kodaira formula 139 equivariant 234, 293, 350
Borel-Weil-Bott Theorem 255 Chevalley, C. 108, 231
Bott class 239 chirality operator 109
Bott, R. 189, 223, 255 Clifford algebra 103
bundle quantization map 104
associated 15 supertrace 110
Clifford 113 symbol map 104
equivariant 16 Clifford bundle 113
frame 15, 32 Clifford connection 117, 328
holomorphic 136 Clifford superconnection 117
horizontal 21 Clifford module 103, 114, 127, 324
index 280, 298 equivariant 185
principal 15 twisted 114
Index 365

coadjoint orbit differential form 17

Fourier transform 226 basic 20
cohomology equivariant 206
de Rham 18, 127, 295 harmonic 128
Dolbeault 138 horizontal 20
equivariant 207 integral along the fibre 31
complex manifold 135 self-dual 132
connection 20 volume 31
affine 28 differential operator 64
Clifford 117, 328 elliptic 65
invariant 26, 209 Dirac, P. A. M. 102
Levi-Civita 33 Dirac operator 116
on a principal bundle 21 classical 127, 134
torsion-free 28 index of 123
connection form 21, 28 divergence 68
contraction 18 Dolbeault cohomology 139
covariant derivative 22 Dolbeault's theorem 139
holomorphic 136 Duhamel's formula 98
tensor product 25 Duistermaat-Heckman theorem 224
curvature 21, 23, 44 Duistermaat, J. J. 224
equivariant 208
Ricci 34, 130 equivariant
Riemannian 33 A-genus 211, 247
scalar 34, 126, 134 bundle 16, 208
twisting 120 characteristic class 210
equivariant 248 Chern character 210
cut-off function 84 relative 194, 250
Chern-Weil homomorphism 234,
5-operator 135 293, 350
de Rham cohomology 18, 127, 295 Clifford module 185
compact supports 19 cohomology 207
de Rham's theorem 128 curvature 208
delta distribution 191 differential form 206
density 30 exterior differential 206
half 69 Euler class 211, 216
integral of 30 index 185, 249, 292
Riemannian 34 superconnection 208
vertical 277 twisting curvature 248
determinant, zeta-function 301 Euler class 58, 149
determinant line bundle 41, 304, 347 equivariant 211, 216
connection 308, 348 Euler number 59, 129, 149
curvature 347 holomorphic vector bundle 136,
L2-metric 305 152
Quillen metric 306 exponential map 35
difference bundle 294 Jacobian of 38, 168
366 Index

exterior algebra 104 index density 147

exterior differential 17, 29 equivariant 195
equivariant 206 family 336
index theorem
family 277 local 147, 181
of smoothing operators 313 equivariant 195, 248
family index theorem 336 injectivity radius 35
flag manifold 230, 252
Fourier transform of orbit 227 Jacobian of exponential map 38, 168
frame bundle 16
orthonormal 32 K-theory 294
Freed, D.A. 347 Kahler manifold 137, 151, 255
fundamental one-form 28 Karamata, J. 94
kernel 74
Gauss-Bonnet-Chern theorem 59, 148 Kirillov, A.A. 252
Gaussian integral 54, 72, 173, 178, Kirillov formula 248, 293, 350
221 Kirillov character formula 252
generalized section 74 Kodaira vanishing theorem 141
geodesic 35 Kostant, B. 2531
Gilkey, P.B. 147, 192 Kiinneth theorem 129
gradient 34
Green operator 92, 287 L-genus 150
L2-metric 305
Hamiltonian vector field 212, 225 Laplace-Beltrami operator 128, 133
Harish-Chandra 229 Laplacian, generalized 65
harmonic form 128 Leibniz's rule 17, 22, 44, 281
harmonic oscillator 145, 153 Levi-Civita connection 33
harmonic spinor 134 Lichnerowicz, A. 134
heat equation 75 Lichnerowicz formula 126, 134, 157,
formal solution 83 333
heat kernel 72, 75, 309 Lie bracket 17
asymptotic expansion 87 Lie derivative 16
Heckman, G. 224 E. Cartan's homotopy formula
Hilbert-Schmidt operator 87 18
Hirzebruch, F. 150, 152, 243 Lie superalgebra 39
Hodge star 131 Liouville form 224, 229
Hodge's theorem 128, 139 local index theorem 147, 181
holomorphic vector bundle 136 equivariant 195, 248
horizontal bundle 21 family 336
horizontal lift 21 localization theorem 214, 216

index 123, 145 manifold

equivariant 185, 248, 292 almost-complex 135
family (see index bundle) complex 135
index bundle 269, 280, 298 Hermitian 135
Index 367

Kahler 137, 151, 255 renormalized limit 300

Riemannian 32 Ricci curvature 34, 130
spin 114, 151 Riemann-Roch theorem 152, 243
symplectic 212 Riemann-Roch-Hirzebruch theorem
Mathai, V. 52, 239 152
McKean, H. 124 Riemannian
McKean-Singer formula 124, 145 curvature 33
equivariant 185 manifold 32
family 298 metric 32
mean curvature 320 moment 211, 247
Mehler's formula 154, 263, 343 Rossman, W. 232
Mellin transform 299
Minakshisundaram, S. 101 scalar curvature 34, 126, 134
moment 209 Schwartz kernel theorem 74
Riemannian 211, 247 second fundamental form 319
symplectic 212 section 14
twisting 248 generalized 74.
Segal, G. B. 195
normal coordinates 35, 71 signature 133, 150
signature operator 133
orientation 31 signature theorem 150
simplex 77
P-endomorphism 278, 309 Singer, I. M. 124, 147, 150, 195, 294,
parallel transport 26 338
parametrix 77 smoothing operator 74
Patodi, V. K. 147 family 313
Pfaffian 42, 222 Spin group 106
Pleijel, A. 101 spin structure 114
Poincare duality 132 spinor bundle 115
Poincare lemma 52 spinor module 109
Poincare-Hopf theorem 59, 61, 223 supertrace 110
polarization 109 stationary phase 224
principal bundle 15 superalgebra 38
geodesic distance 197 Lie 39
heat kernel 173, 199 supercommutative 39
superbundle 39
quantization map 104 supercommutator 39
Quillen, D. 43, 46, 52, 239, 306 superconnection 44, 281
Quillen metric 306, 347 Bismut 317, 331
Clifford 117
radial vector field 35 curvature 44
relative Chern character 146 equivariant 208
equivariant 248 tensor product 46
localized 194 superspace 38
relative supertrace 113, 146 Hermitian 41
368 Index

tensor product 40 twisting

supertrace 39, 46 bundle 114, 131, 134, 175
relative 113, 146 curvature 120
spinor 110 equivariant 248
symbol moment 248
Clifford algebra 104
differential operator 64 vanishing theorem 130
symplectic manifold 212 Kodaira 141
Lichnerowicz 134
tangent bundle 16 vertical density 277
vertical 278, 319 Volterra series 78, 275, 310
tautological section 28, 55, 210 volume form 31
tensor product 40 Weitzenbock formula 130
Thom form 52, 241 Weyl, H. 94, 144, 251
Todd genus 51, 152 Weyl character formula 251
torsion 28 Weyl group 227
trace-class 88
transgression formula 31, 47, 57, zeta function 301
276, 289, 344 zeta-function determinant 301

Printing: Mercedesdruck, Berlin

Binding: Buchbinderei Liideritz & Bauer, Berlin