Riemannian Geometry

Lectures given during the fall of 1990

Hörmander, Lars

1990

Document Version:
Other version

Link to publication

Citation for published version (APA):

Hörmander, L. (1990). Riemannian Geometry: Lectures given during the fall of 1990.

Total number of authors:

1

Creative Commons License:

CC BY-NC-ND

General rights
Unless other specific re-use rights are stated the following general rights apply:
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors
and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the
legal requirements associated with these rights.
• Users may download and print one copy of any publication from the public portal for the purpose of private study
or research.
• You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal

Read more about Creative commons licenses: https://creativecommons.org/licenses/

Take down policy
If you believe that this document breaches copyright please contact us providing details, and we will remove
access to the work immediately and investigate your claim.

LUNDUNI
VERSI
TY

PO Box117
22100L und
+4646-2220000
RIEMANNIAN GEOMETRY

Lars Hörmander

Department of Mathematics
Box 118, S-221 00 Lund, Sweden
RIEMANNIAN GEOMETRY
Lectures given during the fall of 1990 by

Lars Hörmander

Department of Mathematics
Box 118, S-221 00 Lund, Sweden
i
Lars Hörmander
Department of Mathematics
University of Lund
Box 118
S-221 00 Lund, Sweden
e-mail lvh@maths.lth.se (internet)

Copyright c 1990 Lars Hörmander

The notes have been typeset by AMS-TEX.

ii
 PREFACE

The inspiration for the theory of partial differential equations has always come from
two main sources, physics and geometry. The interaction between all three areas has
become intensified in recent years. The solution of the index problem by Atiyah and
Singer in the early 60’s forced the people working on differential equations to improve
their knowledge of differential geometry. This was very useful in the subsequent de-
velopment of microlocal analysis which mainly involved symplectic geometry, a topic
which had previously mainly been cultivated in connection with ordinary differential
equations. Riemannian geometry is central in the recent development of gauge theories,
which rely on a mixture of geometry, physics and partial differential equations. There
has also been a great deal of recent activity in the general theory of relativity, that is, in
pseudo-Riemannian geometry — for example, the proof of the positive mass conjecture
and global existence theorems for the vacuum Einstein equations. The solution of some
purely geometric problems, such as the Yamabe problem and the isometric imbedding
problem, have also enriched the theory of non-linear partial differential equations. The
fundamental open problems in the theory of overdetermined systems of linear differ-
ential equations also require a strong background in geometry even to understand the
present state of affairs. All this should be sufficient reason for an analyst to study
geometry seriously.
In a half semester course it is only possible to present a brief outline of the most clas-
sical Riemannian geometry with a few glimpses of more recent developments. However,
if enough interest is manifested, I plan to continue for one or several semesters more
in order to be able to approach the research front, and these lectures should then be a
convenient platform to build on. One could continue in many different directions. One
possibility is to discuss pseudo-Riemannian manifolds with Lorentz signature (general
relativity theory). Another is to discuss relations to the theory of functions of several
complex variables; the presence of a complex structure gives a much richer structure.

Lund in December 1990

Lars Hörmander

iii
iv
 Contents of Lectures on Riemannian Geometry

Chapter I. Curves in a Euclidean space

1.1. Curvature of a curve . . . . . . . . . . . . . . . . . . . . . . . 1

1.2. The two dimensional case . . . . . . . . . . . . . . . . . . . . . 2
1.3. The Frenet formulas . . . . . . . . . . . . . . . . . . . . . . . 4
1.4. Moving frames . . . . . . . . . . . . . . . . . . . . . . . . . . 6

Chapter II. Curvature of submanifolds of a Euclidean space

2.1. Curves on a submanifold of a Euclidean space . . . . . . . . . . . . 11

2.2. The curvature of a hypersurface . . . . . . . . . . . . . . . . . . 17
2.3. Algebraic properties of the curvature tensor . . . . . . . . . . . . . 19

Chapter III. Abstract Riemannian manifolds

3.1. Covariant derivatives and curvature . . . . . . . . . . . . . . . . 27

3.2. Local isometric embedding . . . . . . . . . . . . . . . . . . . . 47
3.3. Spaces of constant curvature . . . . . . . . . . . . . . . . . . . 52
3.4. Conformal geometry . . . . . . . . . . . . . . . . . . . . . . . 60
3.5. The curvature of a submanifold . . . . . . . . . . . . . . . . . . 67

Chapter IV. Exterior differential calculus in Riemannian geometry

4.1. Submanifolds of a Euclidean space . . . . . . . . . . . . . . . . . 69

4.2. Abstract Riemannian manifolds . . . . . . . . . . . . . . . . . . 74
4.3. The Gauss-Bonnet theorem in higher dimensions . . . . . . . . . . . 77
4.4. The Pontrjagin forms and classes . . . . . . . . . . . . . . . . . 80

Chapter V. Connections, curvature and Chern classes

5.1. Connections in a vector bundle . . . . . . . . . . . . . . . . . . . 83

5.2. Chern classes . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.3. Lie groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.4. Principal and induced bundles . . . . . . . . . . . . . . . . . . . 97
v
Chapter VI. Linear differential operators in Riemannian geometry

6.1. Metric elliptic operators . . . . . . . . . . . . . . . . . . . . . 107

6.2. The exterior algebra of a Euclidean vector space . . . . . . . . . . 116
6.3. The Hodge decomposition . . . . . . . . . . . . . . . . . . . . 120
6.4. Heat equations . . . . . . . . . . . . . . . . . . . . . . . . 127
6.5. Hirzebruch’s index formula and Gilkey’s theorem . . . . . . . . . . 130
6.6. Operators of Dirac type . . . . . . . . . . . . . . . . . . . . . 139
6.7. Clifford and spinor algebra . . . . . . . . . . . . . . . . . . . 140
6.8. Clifford and spinor analysis . . . . . . . . . . . . . . . . . . . 153
6.9. Hermite polynomials and Mehler’s formula . . . . . . . . . . . . 161
6.10. The local index formula for twisted Dirac operators . . . . . . . . 166

Appendix A: Prerequisites from multilinear algebra . . . . . . . . . . . . 171

Appendix B: The calculus of differential forms . . . . . . . . . . . . . . 179

Appendix C: The Frobenius theorem . . . . . . . . . . . . . . . . . . 185

Appendix D: Invariants of O(n) . . . . . . . . . . . . . . . . . . . . 189

Notes and references . . . . . . . . . . . . . . . . . . . . . . . . . 193

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197

vi
 CHAPTER I

CURVES IN A EUCLIDEAN SPACE

Summary. In Section 1.1 we shall just recall the definition of the curvature of a curve in
a Euclidean vector space. The special two dimensional case where the curvature can be
given a sign is studied in somewhat greater depth in Section 1.2. We define the torsion
of a curve and prove the Frenet formulas in Section 1.3. This leads to a discussion of
moving frames in Section 1.4 which is a preparation for the Riemannian geometry in
Chapter IV.

1.1. Curvature of a curve. Let V be a finite dimensional vector space over R

1
with a symmetric scalar product (x, y) defined for x, y ∈ V and the norm kxk = (x, x) 2 .
By a C k curve in V we mean a C k map

(1.1.1) I ∋ t 7→ x(t) ∈ V

where I is an interval ⊂ R, k ≥ 1, and x′ (t) 6= 0 for every t. (Self-intersections are

allowed.) If J is another interval on R and ϕ : J → I is a C k bijection with ϕ′ > 0,
then we regard the curve defined by x ◦ ϕ as the same curve as (1.1.1) with just a
change of parametrization. The arc length s(t) on (1.1.1) is defined up to an additive
constant by
ds(t)/dt = kx′ (t)k.
If we define X(s(t)) = x(t) then X is also in C k and kX ′ (s)k = 1. If s̃ is another
parameter with this property, then ds̃/ds = 1 so s̃ − s is a constant.
Differentiation of the equation 1 = kX ′ (s)k2 = (X ′ (s), X ′ (s)) with respect to s gives

2(X ′′ (s), X ′ (s)) = 0.

Thus the second derivative X ′′ (s) is orthogonal to the tangent X ′ (s) of the curve.
To interpret it geometrically we consider a circle with two perpendicular radii y, z of
length R and center x0 ; it is parametrized by

x(s) = x0 + y cos(s/R) + z sin(s/R), thus x′ (s) = (−y/R) sin(s/R) + (z/R) cos(s/R).

Since (y, y) = (z, z) = R2 and (y, z) = 0, we confirm using the second equation that
kx′ (s)k2 = 1. The second derivative

x′′ (s) = −(y/R2 ) cos(s/R) − (z/R2 ) sin(s/R) = (x0 − x(s))/R2

1
2 I. CURVES IN A EUCLIDEAN SPACE

is directed toward the center of the circle and kx′′ (s)k = 1/R.
Definition 1.1.1. If s 7→ x(s) ∈ V is a C 2 curve parametrized by the arc length,
then x′ (s) is the unit tangent vector at x(s). If x′′ (s) 6= 0, then n = x′′ (s)/kx′′ (s)k
is called the principal normal of the curve at x(s), 1/kx′′ (s)k is called the radius
of curvature, and the circle with center at x(s) + x′′ (s)/kx′′ (s)k2 lying in the plane
spanned by the vectors x′ (s) and x′′ (s) at x(s) is called the osculating circle. One calls
κ = kx′′ (s)k the curvature at x(s) even if κ = 0; otherwise x′′ (s) = κn.
From the discussion above it is clear that when the curvature is not 0 then the
osculating circle is the only one which has a tangency of higher order with the curve
at x(s).
Now suppose that we have a curve t 7→ x(t) with x′ (t) 6= 0 which may not have the
arc length as parameter. If we write x(t) = X(s(t)) as above, we obtain
x′ (t) = X ′ (s)ds/dt, x′′ (t) = X ′′ (s)(ds/dt)2 + X ′ (s)d2 s/dt2 .
Since X ′′ (s) is orthogonal to X ′ (s) we have
X ′′ (s)(ds/dt)2 = x′′ (t) − (x′′ (t), X ′ (s))X ′ (s)
or equivalently,
(1.1.2) X ′′ (s) = x′′ (t)/kx′ (t)k2 − x′ (t)(x′′ (t), x′ (t))/kx′ (t)k4 .
Exercise 1.1.1. Calculate the curvature of a C 2 curve t 7→ x(t) with arbitrary
parametrization.
Exercise 1.1.2. Show that the tangential component of x′′ (t) is equal to d2 s/dt2
times the unit tangent vector, and that the normal component is (ds/dt)2 κ times the
principal normal.
1.2. The two dimensional case. For curves in the plane R2 we can attach a
sign to the curvature (which depends on the orientation of the plane). If s 7→ x(s)
is a curve with the arc length s as parameter and if x′′ (s) 6= 0, then we change the
direction of the normal n(s) so that x′ (s) and n(s) are positively oriented, that is,
det(x′ (s), n(s)) > 0 and define the curvature κ(s) so that x′′ (s) = κ(s)n(s) as before.
Thus κ > 0 means that moving along the curve one sees the curve to the left of the
tangent.
Consider now a closed simple C 2 curve
R/LZ ∋ s 7→ x(s) ∈ R2
where s is still the arc length and x(s) = x(s′ ) if and only if s − s′ ∈ LZ. We can write
x′ (s) = (cos θ(s), sin θ(s)), 0 ≤ s ≤ L,
where θ is a C 1 function uniquely determined up to an integer multiple of 2π. Since
x′′ (s) = (− sin θ, cos θ)dθ/ds = κn = κ(− sin θ, cos θ),
we have κ = dθ/ds. The integral
Z L Z L
(1.2.1) κ ds = dθ = θ(L) − θ(0)
0 0
is obviously a multiple of 2π.
 THE TWO DIMENSIONAL CASE 3

Theorem 1.2.1. The integral (1.2.1) of the curvature of a simple closed curve is
±2π with the positive sign if the curve lies entirely to the left of some tangent when
one moves around it in the positive direction.
Proof. The normalized chord direction

x(t, s) = (x(t) − x(s))/kx(t) − x(s)k, 0≤s<t≤L

becomes a continuous function in the closed triangle T = {(t, s); 0 ≤ s ≤ t ≤ L} if we

define x(s, s) = x′ (s). Since T is simply connected we can find a continuous function
ϕ(t, s) in T such that ϕ(s, s) = θ(s), 0 ≤ s ≤ L, and

x(t, s) = (cos ϕ(t, s), sin ϕ(t, s)), (t, s) ∈ T.

We want to find

θ(L) − θ(0) = ϕ(L, L) − ϕ(L, 0) + ϕ(L, 0) − ϕ(0, 0) = 2(ϕ(L, 0) − ϕ(0, 0)),

where the last equality follows from the fact that x(L, s) = −x(s, 0). Choose the origin
of the parametrization so that x2 (0) = mins x2 (s). Then ϕ(s, 0) only varies between 0
and π. If x′1 (0) > 0 then ϕ(0, 0) = 0 and ϕ(L, 0) = π so the integral of the curvature
becomes 2π. If x′1 (0) < 0 the sign changes, which proves the theorem.
Exercise 1.2.1. Show that when x′1 (0) = 1 in the preceding argument, then the
variation d(y) of the argument of (x(s) − y)/kx(s) − yk for 0 ≤ s ≤ L is either 0 or 2π
for all points y in the complement Ω of the curve, and that Ωe = {y ∈ Ω; d(y) = 0} is
an open connected unbounded set while Ωi = {y ∈ Ω; d(y) = 2π} is an open connected
bounded set, the interior of the curve. (Jordan’s curve theorem.) Hint: Examine first
a neighborhood of x(0), then a neighborhood of the curve.
Exercise 1.2.2. Show that if Γ : [0, T ] ∋ s 7→ x(s) ∈ R3 is a simple closed C 2
curve parametrized by the arc length, and N ∈ S 2 , then s 7→ x(s) − N (x(s), N ) is a
C 2 curve in the orthogonal plane of N for almost all N . Write an expression for its
RT
curvature and conclude that if κ(s) is the curvature of Γ, then 0 κ(s) ds ≥ 2π. (See
also Milnor [1].)
Exercise 1.2.3. Show that if

R/T Z ∋ t 7→ x(t) ∈ R2

is a simple closed C 2 curve, then

Z T
L= | det(x′ (t), x′′ (t))|1/3 dt
0

is independent of the parametrization. It is called the affine length. Show that L3 /A

is invariant under affine transformations if A is the area of the interior of the curve,
calculate the quotient for an ellipse. (By a theorem of Blaschke this is the least upper
bound for all convex curves; see Burago and Zalgaller [1], p. 7.)
4 I. CURVES IN A EUCLIDEAN SPACE

1.3. The Frenet formulas. We shall now return to the study of a C k curve
(1.1.1) taking the arc length s as parameter. If the curvature is not equal to 0 at
x(s0 ), then the principal normal n(s) is defined for s near s0 . If k ≥ 3 we can take
its derivative n′ (s). Since (n(s), n(s)) = 1 we obtain 2(n′ (s), n(s)) = 0, and since
(n(s), x′ (s)) = 0 we have

(n′ (s), x′ (s)) + (n(s), x′′ (s)) = 0, that is, (n′ (s), x′ (s)) = −κ(s).

Hence n′ (s) + κ(s)x′ (s) is orthogonal to the plane spanned by x′ (s) and n(s). The
length τ (s) of this vector is called the torsion of the curve at x(s). If τ (s) 6= 0, then
normalization gives a unit vector b(s), called the binormal of the curve at x(s), and we
have
n′ (s) = −κ(s)x′ (s) + τ (s)b(s).
The procedure can be continued by differentiation of b with respect to s. Differentiation
of the equations

(b(s), x′ (s)) = 0, (b(s), n(s)) = 0, (b(s), b(s)) = 1

gives

(b′ (s), x′ (s)) = −(b(s), x′′ (s)) = 0,

(b′ (s), n(s)) = −(b(s), n′ (s)) = −τ (s), (b′ (s), b(s)) = 0.

If dim V = 3 the vectors x′ (s), n(s) and b(s) form a complete orthonormal system, so
we get the third of the Frenet formulas

x′′ (s) = κ(s)n(s)

n′ (s) = −κ(s)x′ (s) + τ (s)b(s)
b′ (s) = − τ (s)n(s).

However, the procedure becomes more illuminating in the higher dimensional case if
we use somewhat different notation.
Consider a C k+1 curve (1.1.1) with arbitrary parametrization, such that for all t ∈ I
the derivatives x′ (t), . . . , x(k) (t) are linearly independent but x′ (t), . . . , x(k+1) (t) are
linearly dependent. Clearly k ≤ dim V . For any given sufficiently differentiable curve
this is true if I is replaced by any interval in the dense open subset where the rank
of x′ (t), x′′ (t), . . . is locally maximal. The hypothesis is independent of the choice of
parametrization, and so is the linear span Ej (t) of x′ (t), . . . , x(j) (t) when j ≤ k. In
fact, if t̃ is another parameter then
j
dj x dj x

dt
− ∈ Ej−1
dt̃j dt̃ dtj
 THE FRENET FORMULAS 5

so this follows inductively. Application of the Gram-Schmidt orthogonalization proce-

dure to the sequence x′ (t), . . . , x(k) (t) gives orthonormal vectors e1 (t), . . . , ek (t):
X
e1 (t) = x′ (t)/c1 (t), . . . , ej (t) = x(j) (t) − (x(j) (t), ei (t))ei (t) /cj (t),


i<j

where cj (t) > 0 is chosen so that kej (t)k = 1. These vectors at x(t) do not depend
on the choice of parametrization; e1 (t) is the unit tangent vector, e2 (t) is the principal
normal, e3 (t) is the binormal, and so on.
Assume now for the sake of simplicity that the parameter t is the arc length s. By
the construction and the hypothesis that Ek+1 (s) = Ek (s) we have
k
X
e′i (s) = ωij (s)ej (s), i ≤ k,
j=1

where ωij (s) = (e′i (s), ej (s)) = 0 if j > i + 1. If we differentiate the equations

(ei (s), ej (s)) = δij

expressing the orthonormality of e1 (s), . . . , ek (s), we obtain

0 = (e′i (s), ej (s)) + (ei (s), e′j (s)) = ωij (s) + ωji (s),

so the matrix (ωij (s)) is skew symmetric. Thus ωii (s) = 0, and since ωij (s) = 0 when
j > i + 1, this is also true when i > j + 1. If we set κj (s) = ωj,j+1 (s), 1 ≤ j < k, it
follows that the matrix (ωij ) has the special form for k = 4, say:

0 κ1 0 0
 
 −κ1 0 κ2 0 
.
0 −κ2 0 κ3

0 0 −κ3 0
Here κ1 (s) is of course the curvature, and κ2 is the torsion. The Frenet formulas can
now be written

(1.3.1) e′1 (s) = κ1 (s)e2 (s), e′j (s) = −κj−1 (s)ej−1 (s) + κj (s)ej+1 (s), 1 < j < k,
e′k (s) = −κk−1 (s)ek−1 (s).

This system of differential equations has a unique solution with given initial data. If
these are orthonormal then the solution remains orthonormal, and the solution must
belong to the space Ek (s0 ), for there is a solution with e1 (s), . . . , ek (s) contained in
this space which is equal to e1 (s0 ), . . . , ek (s0 ) at s0 , so Ek (s) is independent of s and
the curve lies in an affine subspace of dimension k.
Exercise 1.3.1. Let k ≥ 3 and determine the Taylor expansion of x(s) with error
o(s3 ) at 0 ∈ I in terms of x(0), e1 (0), e2 (0), e3 (0) and the Taylor expansions of κ1 and
κ2 at 0.
6 I. CURVES IN A EUCLIDEAN SPACE

1.4. Moving frames. In the preceding section we studied orthonormal vectors

depending on a parameter which were adapted to a given curve. However, some of the
arguments are of a much more general nature so we shall take them up once more.
Let V again be a Euclidean space of dimension n < ∞, and denote by F (V ) the
set of all orthonormal frames e1 , . . . , en ∈ V . Since F (V ) ⊂ V n it makes sense to say
that a curve I ∋ t 7→ f (t) ∈ F (V ) is differentiable. Writing f (t) = (e1 (t), . . . , en (t))
we have by definition of F (V )

(ei (t), ej (t)) = δij

and conclude as in Section 1.3 by differentiating that

n
X
(1.4.1) dei /dt = ωij (t)ej (t), where ωij (t) = −ωji (t).
j=1

This can be interpreted as follows. If e = (e1 , . . . , en ) ∈ F (V ) and f = (f1 , . . . , fn ) ∈

F (V ), then
n
X n
X
(1.4.2) ei = Oij fj , where Oik Ojk = δij .
j=1 k=1

Thus O t O = I, where I is the identity matrix, or equivalently t O O = I. Such matrices

are called orthogonal, and one denotes by O(n) the set of orthogonal n × n matrices,
which is a group under matrix multiplication. When (1.4.2) is valid we write e = Of .
With any fixed e0 we can now write e(t) = O(t)e0 where O(t) is differentiable with
values in O(n), and we have de(t)/dt = (dO(t)/dt)e0 . In particular, if t = 0 and we
take e0 = e(0), then O(0) = I, and (1.4.1) means that

(1.4.3) O(t) differentiable at 0 with values in O(n), O(0) = I

=⇒ O ′ (0) is skew symmetric.

The proof is obvious: the equation O(t) t O(t) = I implies

O ′ (0) t O(0) + O(0) tO ′ (0) = 0, that is, O ′ (0) + t O ′ (0) = 0.

This means that infinitesimal rotations are defined by skew symmetric matrices. Every
such infinitesimal rotation ω gives rise to a one parameter group of rotations O(t), that
is a function R ∋ t 7→ O(t) ∈ O(n) such that

(1.4.4) O(t)O(s) = O(t + s), O ′ (0) = ω.

The first condition implies O(0) = I, and differentiation with respect to s gives O ′ (t) =
O(t)ω when s = 0, hence
X∞
tω
O(t) = e = (tω)j /j!.
0
 MOVING FRAMES 7

The properties (1.4.4) follow at once, and since t O(t) = O(−t) it follows that O(t) ∈
O(n).
Let us consider the examples of lowest dimension. If n = 2 then
   
0 θ tω cos(tθ) sin(tθ)
ω= , e = ,
−θ 0 − sin(tθ) cos(tθ)

so etω means rotation by the angle tθ in the negative direction. If n = 3 then

 
0 θ3 −θ2
ω = −θ3
 0 θ1  .
θ2 −θ1 0

Noting that ωθ = 0 it is easy to see that etω means rotation by the angle t|θ| around
θ.
Exercise 1.4.1. Show that if
 
0 θ3 −θ2
ω = −θ3
 0 θ1  ,
θ2 −θ1 0

then etω = A + B cos(t|θ|) + C sin(t|θ|) and determine the 3 × 3 matrices A, B, C. Use

this to find the curve in R3 with constant curvature and torsion which passes through
the origin with tangent, principal normal and binormal along the positive x1 , x2 and
x3 axes.
Exercise 1.4.2. Show that if n is odd and S is a skew symmetric n × n matrix,
then the equation Sθ = 0 has at least one solution θ 6= 0, and show that etS θ = θ.
Show for any n that if S is a skew symmetric n × n matrix 6= 0, then there is a two
dimensional plane W ⊂ Rn such that W and its orthogonal complement W ⊥ are left
invariant by S and therefore by etS . Describe the structure of S and etS geometrically
in general. Hint: Use that S/i is hermitian symmetric in Cn .
Exercise 1.4.3. Show that if O ∈ O(n), then Rn is an orthogonal direct sum
W+ ⊕ W− ⊕ W such that ±O is the identity on W± and W is the orthogonal direct
sum of twodimensional subspaces where O is a rotation. Conclude that the component
of the identity in O(n) is the subgroup SO(n) of elements with determinant 1, and
that every element in SO(n) is of the form eS for some skew symmetric S. Hint: Use
that O is unitary in Cn .
We end this section with a few exercises giving some information on Lie groups and
Lie algebras which will be useful later on. Let M(V ) be the set of linear transformations
in the finite dimensional vector space V , and let GL(V ) be the subset of invertible
ones. Then GL(V ) is a group under multiplication; the identity is denoted by I. With
some Euclidean norm in V we use the corresponding operator norm in M(V ). When
V = Rn we write M(n) and GL(n) instead of M(Rn ) and GL(Rn ).
8 I. CURVES IN A EUCLIDEAN SPACE

Exercise 1.4.4. Show that

P∞ n
(1) eX = 0 X /n! converges in M(n) if X ∈ M(n), and that e
X
∈ GL(n),
−X X
e e = I. P
∞
(2) log(I + X) = 1 (−1)n−1 X n /n converges in M(n) if X ∈ M(n) and kXk < 1,
and elog(I+X) = I + X then.
(3) log eX = X if kXk < log 2.
(4) if R ∋ t 7→ A(t) ∈ GL(n) is a continuous one parameter subgroup, then
Z ε Z t+ε
A(t) A(s) ds = A(s) ds;
0 t

deduce that A : R → M(n) is differentiable and that A(t) = etX where X =

A′ (0) ∈ M(n).
(5) if X, Y ∈ M(n), then

eX+Y = lim (eX/ν eY /ν )ν .

ν→∞

Exercise 1.4.5. Let G be a closed subgroup of GL(n), and

g = {X ∈ M(n); etX ∈ G, t ∈ R}.

Show that
(1) g is a linear subspace of M(n).
(2) if g ∈ G and X ∈ g, then

getX g −1 = etY , t ∈ R,

where Y = gXg −1 ∈ g. One writes Y = Ad(g)X also (the adjoint representa-

tion).
(3) if Y, X ∈ g, then t 7→ Ad(etY ) is a one parameter subgroup of GL(g), hence
Ad(etY ) = et ad(Y ) where ad(Y ) ∈ M(g) (the infinitesimal adjoint representa-
tion); show that ad(Y )X = [Y, X] = Y X − XY and deduce that g is a Lie
algebra, that is, closed also under commutators.
(4) show that {eX ; X ∈ g, kXk < 1} is a neighborhood of the identity in G. Hint:
Assume that this is false and take a sequence Xk → 0 in M(n) such that
eXk ∈ G. Write Xk = Yk + Zk where Zk ∈ g and 0 6= Yk ∈ g⊥ (the orthogonal
complement with respect to some Euclidean norm in M(n)). Note that

G ∋ eXk e−Zk = eYk +Zk e−Zk = eWk , where Wk = Yk + O(|Yk ||Zk |).

Pass to a subsequence such that Wk /kWk k → W and get a contradiction by

proving that W ∈ g ∩ g⊥ .
 MOVING FRAMES 9

Exercise 1.4.6. Show that if U ∈ SU(n), the group of unitary n × n matrices

with determinant 1, and h ∈ H0 (n), the space of n × n hermitian symmetric matrices
with trace 0, then U hU ∗ ∈ H0 (n), which gives a homomorphism SU(n) → O(n2 − 1).
Describe the one parameter subgroups of SU(n), and prove that SU(2) → SO(3) is
surjective, and in fact a double cover. Show that SU(2) is simply connected by proving
that
SU(2) ∋ U 7→ (Re U11 , Im U11 , Re U21 , Im U21 ) ∈ S 3
is a bijection on the unit sphere S 3 ⊂ R4 , and write down the inverse.
10 I. CURVES IN A EUCLIDEAN SPACE
 CHAPTER II

CURVATURE OF SUBMANIFOLDS OF A EUCLIDEAN SPACE

Summary. In Section 2.1 we introduce the notions of first and second fundamental
forms for a submanifold of a Euclidean space. After introducing the Christoffel symbols
and geodesic curvature, we define the Riemann curvature tensor and connect it with
the second fundamental form through the Gauss equations. Section 2.2 is devoted to
the special case of a hypersurface, and in particular the Gauss map. Basic algebraic
(symmetry) properties of the curvature tensor are given in Section 2.3. The discussion
includes the first Bianchi identity and the decomposition of the curvature tensor with
special emphasis on the four dimensional case.

2.1. Curves on a submanifold of a Euclidean space. Let M be a C µ

submanifold of dimension n in a finite dimensional vector space V of dimension N .
This means that to every X0 ∈ M there is a neighborhood U in V and
(1) a map F ∈ C µ (U, RN−n ) such that F ′ is surjective and M ∩ U = {X ∈
U ; F (X) = 0};
(2) there is a neighborhood ω of 0 in Rn and a map f ∈ C µ (ω, V ) with f (0) = X0
and injective differential, such that f (ω) = M ∩ U .
These conditions are equivalent. In fact, given F as in (1) we can choose G ∈
C (U, Rn ) such that (F ′ , G′ ) is of rank N at X0 and G(X0 ) = 0. By the implicit
µ

function theorem there is an inverse Φ ∈ C µ defined in a neighborhood of 0 in RN .

Since X = Φ(F (X), G(X)) when X is in a neighborhood of X0 , it follows that a
neighborhood of X0 in M is parametrized by X = Φ(0, y) where y = G(X) is in a
neighborhood of 0 in Rn . Conversely, (F (Φ(0, y)), G(Φ(0, y))) = (0, y), so Φ(0, y) ∈ M .
Choosing G as the projection on a suitable n dimensional coordinate plane, we obtain
a representation (2) where N − n coordinates are C µ functions of the other n.
On the other hand, given f as in (2) we can choose g ∈ C µ (RN−n , V ) with g(0) = 0
so that (x, y) 7→ f (x) + g(y) has bijective differential at (0, 0). Let ϕ = (ϕ1 , ϕ2 ) be
an inverse from a neighborhood of X0 to a neighborhood of 0 in Rn × RN−n . Then
ϕ(f (x)) = (x, 0) for x in a neighborhood of 0 in Rn , and X = f (ϕ1 (X)) + g(ϕ2 (X)) =
f (ϕ1 (X)) if ϕ2 (X) = 0, so (1) is valid if F = ϕ2 .
If I ∋ t 7→ X(t) is a C ν curve for some ν ≤ µ which is contained in M , and if
f is a parametrization of M near X(t0 ) = X0 as in (2) above, then we can write
X(t) = f (x(t)) where x(t) ∈ ω, if t is close to t0 . The local coordinates x(t) are also
C ν functions of t, for f has a C µ left inverse ϕ by the inverse function theorem, and
x(t) = ϕ(X(t)). Thus it is equivalent to say that X(t) is a C ν function with values in
V or that the local coordinates x(t) are C ν functions of t. (In the same way we see that
a change of local coordinates is made by a C µ map.) This will be the only available
11
12 II. CURVATURE OF SUBMANIFOLDS OF A EUCLIDEAN SPACE

definition for abstract manifolds in Chapter III. We shall consistently use upper case
letters for points in V and lower case letters for points in the parameter space. It is
the latter which will survive in Chapter III.
If ν ≥ 1 then the tangent of our curve has the direction
n
dX df (x(t)) X dxj
(2.1.1) = = fj (x(t)); fj (x) = ∂f (x)/∂xj = ∂j f (x).
dt dt 1
dt

Thus the tangent vectors of all C 1 curves passing through f (x) span the plane

Tf (x) M = f ′ (x)Rn ,

called the tangent plane of M at f (x). It is of course independent of the choice of f ,

but f ′ (x) identifies it with Rn in a manner which is not invariant. From now on we
assume that V is a Euclidean vector space. The arc length s on the curve is then given
by
n
X
(2.1.2) (ds/dt)2 = gjk (x(t))τj τk ; τj = dxj /dt, gjk (x) = (fj (x), fk (x)).
j,k=1

One calls the quadratic form (2.1.2) the first fundamental form of M . It is a quadratic
form depending on x ∈Pω, which can invariantly be regarded as a quadratic form in
n
Tf (x) M , for Rn ∋ τ 7→ 1 τj fj (x) = X ∈ Tf (x) M is a bijection, and the form is equal
to the square of the norm of the tangent vector X in V .
Assuming from now on that µ ≥ 2 and that we have a C 2 curve, we compute the
second derivative:
n n
d2 f (x(t)) X d2 xj X dxj dxk
(2.1.3) 2
= 2
fj (x(t)) + fjk (x(t)),
dt j=1
dt dt dt
j,k=1

where fjk (x) = ∂j ∂k f (x). Second derivatives of x just occur in the first sum, and the
coefficients there are tangent vectors. If we project on the normal plane Nf (x) M of
Tf (x) M , we can eliminate this sum and obtain:
Theorem 2.1.1 (Meusnier). Let hjk (x), j, k = 1, . . . , n, be the orthogonal projec-
tion of fjk (x) = ∂j ∂k f (x) in the normal plane Nf (x) M of M at f (x). For every C 2
curve on M with the unit tangent vector f ′ (x)τ at f (x), the sum
n
X
(2.1.4) hjk (x)τj τk
j,k=1

is then equal to the curvature times the orthogonal projection of the principal normal
in Nf (x) M .
We can regard (2.1.4) as a quadratic form H in Tf (x) M with values in Nf (x) M ; by
Theorem 2.1.1 it is then independent of the choice of parametrization. One calls (2.1.4)
 CURVES ON A SUBMANIFOLD OF A EUCLIDEAN SPACE 13

the second fundamental form of M . Classically it was defined for hypersurfaces (in R3 ).
In that case, an orientation of M identifies Nf (x) M with R, so one can then regard H
as a real valued quadratic form. We shall sometimes introduce an orthonormal basis in
Nf (x) M and will then be able to write H as an N − n tuple of scalar quadratic forms.
We shall now consider the tangential components of (2.1.3). Since f1 (x), . . . , fn (x)
form a basis in Tf (x) M , we can write
n
X
(2.1.5) fik (x) = Γik l (x)fl (x) + hik (x),
l=1
for the normal component is hik (x) by definition. To calculate the coefficients Γik l we
take the scalar product with fj which gives
Xn
(2.1.6) Γik l glj = (fik , fj ).
l=1
By a miracle one can compute the right-hand side by means of the coefficients of the
first fundamental form and their derivatives, for we have
∂k gij = (fik , fj ) + (fi , fjk )
∂i gjk = (fji , fk ) + (fj , fki ),
∂j gki = (fkj , fi ) + (fk , fij ).
The desired quantity occurs in the first two equations so we add them and subtract
the third, which gives
(2.1.7) Γikj = (fik , fj ) = 21 (∂k gij + ∂i gjk − ∂j gki ),
where the first equality is a new definition. Note the symmetry in the indices i, k, and
that
(2.1.7)′ ∂k gij = Γikj + Γjki .
If (g ij ) denotes the inverse of the matrix (gij ) we obtain from (2.1.6) and (2.1.7)
Xn n
X
l lj
(2.1.8) Γik = g (fik , fj ) = g lj Γikj .
j=1 j=1

Definition 2.1.2. The functions Γikj and Γik j are called Christoffel symbols of
the first and second kind; they are determined by the first fundamental form.
 
l
Remark. The classical notation was [ij, k] and instead of Γijk and Γij l .
ij
√
Exercise 2.1.1. Prove that if g = det(gij ), then ∂l g = 2g i Γil i , that is, ∂l g =
P
√
g i Γil i .
P

Summing up the preceding discussion, we can now write (2.1.3) in the form
n n n
d2 f (x(t)) X d2 xj X
j dxi dxk

X dxi dxk
(2.1.9) 2
= 2
+ Γik f j (x) + hik (x).
dt j=1
dt dt dt dt dt
i,k=1 i,k=1
We discussed the normal component in Theorem 2.1.1, and we can now examine the
tangential component:
14 II. CURVATURE OF SUBMANIFOLDS OF A EUCLIDEAN SPACE

Theorem 2.1.3.P Let s → f (x(s)) be a C 2 curve in M , parametrized by the arc

length, that is, gjk (x) dxj /ds dxk /ds = 1. Then the curvature times the orthogonal
projection of the principal normal in Tf (x) M is equal to
n  2
X d xj X dxi dxk 
(2.1.10) + Γik j fj (x).
j=1
ds2 ds ds
i,k

Thus the vector (2.1.10) does not depend on the choice of parametrization f . The
length is called the geodesic curvature of the curve at f (x), and the direction is called
the geodesic principal normal direction.
In (2.1.10) we have just computed the derivative of the tangent vector along the
curve. However, the argument is much more general. Suppose that we have a C 1
vector field v defined in a neighborhood of the curve and that v is everywhere tangent
to M . Then we can write the vector at f (x) as
n
X
v(x) = vj (x)fj (x),
1

and we obtain
n
dv(x(t)) X
= ((∂k vj (x))fj (x) + vj (x)fjk (x))x′k .
dt
j,k=1

By (2.1.5) it follows that

n n
dv(x(t)) X X
(2.1.11) = (∂k vj + Γik j vi )x′k fj (x) mod Nf (x) .
dt i=1
j,k=1

This is called the covariant derivative of the vector field along the curve. Note that
one does not need to know the embedding function in order to compute (2.1.11); it
suffices to know the first fundamental form. (The right-hand side of (2.1.11) is also well
defined if v is just given on the curve.) This aspect will be discussed systematically in
Chapter III.
Now we pass to determining the tangential components of the derivatives of a C 1
normal vector field n(x) at f (x). We can write
n
X
∂i n(x) = nik (x)fk (x) mod Nf (x) .
k=1

Since differentiation of the equation (n, fj ) = 0 gives

n
X
(∂i n, fj ) + (n, fij ) = 0, that is, nik gkj = −(hij , n)
k=1
 CURVES ON A SUBMANIFOLD OF A EUCLIDEAN SPACE 15

it follows that
n
X
(2.1.12) nik = − (hij , n)g kj .
j=1
P
This means that y 7→ yi ∂i n(x) mod Nf (x) regarded as a linear transformation in
Tf (x) is precisely the linear transformation defined by the scalar product (H, −n) of
the second fundamental form and −n, using the identification of quadratic forms and
linear transformations given by the first fundamental form.
It is now easy to find the tangential component of fijk = ∂i ∂j ∂k f . (The following
formulas should be understood in the sense of distribution theory if M is only in C 2 ,
but one can also assume that M ∈ C 3 and use approximation to extend P the final
2 l
formulas where no third derivatives occur to the C case.) Since fij = Γij fl + hij
we obtain
n
X n
X n
X
fkij = ∂k fij = (∂k Γij m + Γij l Γlk m − (hij , hkl )g lm )fm mod Nf (x) .
m=1 l=1 l=1

Since fkij − fjik = 0, we have proved

Theorem 2.1.4 (The Gauss equations). The first and the second fundamental
form are related by the equations
n
X
m m m
R ijk = ∂j Γik − ∂k Γij + (Γik l Γlj m − Γij l Γlk m )
l=1
(2.1.13) n
X
(hik , hjl ) − (hij , hkl ) g lm ;


=
l=1

here the first equality is a definition.

The second equality in (2.1.13) suggests that we should introduce
n
X
(2.1.14) Rlijk = glm Rm ijk
m=1

for then it takes the simple form

(2.1.15) Rlijk = (hik , hjl ) − (hij , hkl ).

To rewrite the first equality in (2.1.13) we note that in view of (2.1.7)′

n
X n
X n
X n
X
m s m
glm ∂j Γik + Γik Γsjl = ∂j glm Γik − Γik m (Γljm + Γmjl )
m=1 s=1 m=1 m=1
n
X
+ Γik m Γmjl .
m=1
16 II. CURVATURE OF SUBMANIFOLDS OF A EUCLIDEAN SPACE

Using the cancellation between the last two sums we obtain

n
X
(2.1.13)′ Rlijk = ∂j Γikl − ∂k Γijl + (Γij m Γlkm − Γik m Γljm ).
m=1

Again we note that the expression (2.1.13)′ shows that Rijkl can be computed from
the first fundamental form alone while (2.1.15) only involves the second fundamental
form. From (2.1.15) we also see that the corresponding 4-linear form
n
X
1 2 3 4
R(t , t , t , t ) = Rijkl t1i t2j t3k t4l ; t1 , . . . , t4 ∈ R n ,
i,j,k,l=1

is antisymmetric in the pair t1 , t2 and in the pair t3 , t4 but symmetric for exchange
of the pairs. Considered as a 4-linear form in f ′ (x)t1 , . . . , f ′ (x)t4 ∈ Tf (x) M the form
R is independent of the choice of parametrization since thisP is true for the symmetric
bilinear map H : Tf (x) M × Tf (x) M → Nf (x) M defined by hij t1i t2j . In fact, for the
corresponding form R e we have by (2.1.15)

(2.1.15)′ e 1 , t2 , t3 , t4 ) = H(t1 , t3 ), H(t2 , t4 ) − H(t1 , t4 ), H(t2 , t3 ) .

R(t

Definition 2.1.5. The 4-linear form R

e on T M is called the Riemann curvature
tensor of M .
The curvature tensor has symmetries in addition to those already mentioned. We
shall discuss them in Section 2.3. However, already here we introduce the Ricci tensor
which is obtained by contraction of the Riemann curvature tensor,
n
X n
X n
X
(2.1.16) Rjl = g ik Rijkl = Rk jkl = Rk lkj .
i,k=1 k=1 k=1

The corresponding bilinear form X

Rjl t1j t2l
is symmetric and invariantly defined on the tangent space of M , for it is the trace
of the linear transformation corresponding to the bilinear form (t, s) 7→ R(t, t1 , s, t2 )
and the first fundamental form. Taking the trace once more one obtains the scalar
curvature
X
(2.1.17) S= g jl Rjl .

Exercise 2.1.2. Show that with the notation in Exercise 2.1.1

n
X n
X n
X
j l
Rik = ∂j Γik − 1
2 ∂i ∂k log g + 1
2 Γik ∂l log g − Γij l Γlk j .
j=1 l=1 l,j=1
 THE CURVATURE OF A HYPERSURFACE 17

2.2. The curvature of a hypersurface. To clarify the geometric meaning of

the curvature tensor we shall devote this section to a discussion of hypersurfaces in
RN . (The case N = 3 is the classical one of course.) Thus we assume now that
dim V = N = n + 1. Let n(x) denote one of the unit normals of M at f (x). Then we
can write
hij (x) = h̃ij (x)n(x)
where h̃ij is now a scalar. It is clear that this scalar quadratic form corresponds
to an invariantly
P P defined quadratic form on the tangent space of M . The quotient
h̃ij ti tj / gij ti tj is by Meusnier’s theorem the curvature at f (x) of a curve in M
with tangent vector f ′ (x)t and the principal normal in the direction of the normal n(x)
of the surface, that is, with vanishing geodesic curvature. To understand the quotient
it is natural to diagonalize the forms simultaneously, thus introduce new coordinates
s1 , . . . , sn such that
n
X n
X n
X n
X
h̃ij ti tj = Ki s2i , gij ti tj = s2i .
i,j=1 i=1 i,j=1 i=1
P
This is always possible since gij ti tj is positive definite. The eigenvalues Ki are the
solutions of the equation

(2.2.1) det(h̃ij − λgij ) = 0.

In particular, the symmetric functions are given by the coefficients in this equation,
n
X n
X n
Y
ij
(2.2.2) Ki = h̃ij g , Ki = det(h̃ij )/ det(gij ).
i=1 i,j=1 i=1

Definition 2.2.1. For an oriented hypersurface the eigenvectors of the second fun-
damental form with respect to the first fundamental form, both regarded as quadratic
forms on the tangent space Tx M of M at x, are called principal curvature directions at
x. The corresponding eigenvalues are called principal curvatures. Their product and
sum, given by (2.2.2), are called the total (or Gauss) curvature and the mean curvature
respectively.
Exercise 2.2.1. Write down explicitly the formulas for the mean curvature and
the total curvature of a surface of the form xn+1 = ϕ(x1 , . . . , xn ) in Rn+1 , where
ϕ ∈ C2.
Since (n, n) = 1 we have (n, ∂i n) = 0, so (2.1.12) gives
n
X
∂i n = − h̃ij g kj fk ,
j,k=1

or equivalently
n
X n
X n
X
( ti ∂i n, s k fk ) = − h̃ij ti sj .
i=1 k=1 i,j=1
18 II. CURVATURE OF SUBMANIFOLDS OF A EUCLIDEAN SPACE

Thus the differential of the Gauss map

(2.2.3) γ : M ∋ f (x) → n(x) ∈ S n

is the linear transformation in Tf (x) M corresponding to minus the second fundamental

form and the first fundamental form:
Theorem 2.2.2. If γ is the Gauss map M → S n where M is an oriented hypersur-
face and S n is the unit sphere, then the differential γ ′ : Tx M → Tγ(x) S n , x ∈ M , can
be regarded as a map Tx M → Tx M since Tx M and Tγ(x) S n are parallel. The princi-
pal curvature directions are then eigenvectors of −γ ′ with the principal curvatures as
eigenvalues.
The last statement is a classical theorem of Olinde Rodrigues. In particular we see
that γ is a local diffeomorphism precisely when the total curvature is not 0.
Exercise 2.2.2. Let M0 be an open subset of M which is mapped diffeomorphically
into S n by γ, and let f be a continuous function with support in the range γ(M0 ) ⊂ S n .
Show that Z Z
f dS = (f ◦ γ)|K| dM

where dS and dM are the Euclidean volume elements of S n and of M .

We shall now compute the Riemann curvature tensor of a hypersurface. If e1 , . . . , en
is an orthonormal system of principal curvature directions at a point x ∈ M , then
n
X
h̃(t) = Kj (t, ej )2 , t ∈ Tx (M ),
j=1

where Kj are the corresponding principal curvatures. For t1 , . . . , t4 ∈ Tx (M ) we have

(t1 , ei )(t3 , ei )(t2 , ej )(t4 , ej ) − (t1 , ei )(t4 , ei )(t2 , ej )(t3 , ej )

+ (t1 , ej )(t3 , ej )(t2 , ei )(t4 , ei ) − (t1 , ej )(t4 , ej )(t2 , ei )(t3 , ei )
(t1 , ei ) (t1 , ej ) (t3 , ei ) (t3 , ej )
= .
(t2 , ei ) (t2 , ej ) (t4 , ei ) (t4 , ej )

If we multiply by 21 Ki Kj and sum, it follows that

n
e 1 , t2 , t3 , t4 ) = 1
X (t1 , ei ) (t1 , ej ) (t3 , ei ) (t3 , ej )
(2.2.4) R(t Ki Kj .
2 (t2 , ei ) (t2 , ej ) (t4 , ei ) (t4 , ej )
i,j=1

We can of course equally well sum for i < j, omitting the factor 1/2. In the particular
case where n = 2 we only have one term then and conclude that R(t e 1 , t2 , t3 , t4 ) is the
product of the total curvature K = K1 K2 and the areas of the parallelograms spanned
by t1 , t2 and by t3 , t4 . Since R is determined by the first fundamental form, we have
in particular proved:
 ALGEBRAIC PROPERTIES OF THE CURVATURE TENSOR 19

Theorem 2.2.3 (Teorema egregium of Gauss). For a surface in R3 the total

curvature is determined by the first fundamental form; it equals R(t1 , t2 , t1 , t2 ) if t1 , t2
are two orthogonal unit tangent vectors.
Note that the mean curvature is not determined by the first fundamental form, for
a plane and a cylinder have the same fundamental form with suitable coordinates but
the mean curvatures are different. In Chapter III we shall return to the historical
background of Gauss’ discovery.
To compute the Ricci tensor, we take i 6= j in (2.2.4), let t1 = t3 = ek and sum
over k. The only contributions 6= 0 occur when k = i or k = j, so the quadratic form
corresponding to the symmetric Ricci tensor is
X
(2.2.5) t 7→ Ki Kj (t, ej )2 .
i6=j

When n = 2 it is the Gauss curvature times the firstP fundamental form. For any n
it has diagonal form, and the diagonal elements are j;j6=i Ki Kj , for i = 1, . . . , n.
This is the sum of the Gauss curvatures of the sections with three dimensional planes
containing n, ei and ej for some j 6= i. The scalar curvature is
X n
X X
(2.2.6) S= Ki Kj = ( Ki )2 − Ki2 ,
i6=j 1

which is twice the Gauss curvature when n = 2.

2.3. Algebraic properties of the curvature tensor. Recall that the curvature
tensor Rijkl defined by (2.1.15) for a submanifold of RN is antisymmetric in the pair
ij and in the pair kl but symmetric for exchange of these pairs. This implies that

R(t1 , t2 , t1 , t2 )/(g(t1 , t1 )g(t2 , t2 ) − g(t1 , t2 )2 )

only depends on the two plane spanned by the tangent vectors t1 and t2 ; the de-
nominator is the square of the area of the parallelogram spanned by them. Because
of Theorem 2.2.3 one calls this quotient the sectional curvature for the two plane in
Tx M .
We shall now determine if there are additionals restrictions on thePtensors which
may occur at a point. This is a simple problem in linear algebra. Let hjk xj xk be a
n r
quadratic form in R with coefficients hjk = hkj in R where r = N − n, and define
as in (2.1.15)

(2.3.1) Rijkl = (hik , hjl ) − (hil , hjk ).

Since we put no condition on r = N − n and the different coordinates in Rr give

additive contributions, the set T of tensors Rijkl which can occur is the linear space
generated by (2.2.4) where we take only K1 and K2 different from 0. Let E be the
vector space of all Rijkl , i, j, k, l = 1, . . . , n with the obvious symmetries

(2.3.2) Rijkl = −Rjikl = −Rijlk = Rklij .

20 II. CURVATURE OF SUBMANIFOLDS OF A EUCLIDEAN SPACE

P 2
E is a Euclidean space with the norm square Rijkl , so every linear form on E can
be written X
R 7→ Rijkl Sijkl

where S ∈ E. If the linear form vanishes on the term in (2.2.4) with i = 1, j = 2, then
we have for the 4-linear form defined by S

(2.3.3) S(e1 , e2 , e1 , e2 ) = 0,

for (2.3.2) implies that the four terms in R̃ give the same contribution. Since

S(e1 , e2 + λe1 , e1 , e2 + λe1 ) = S(e1 , e2 , e1 , e2 ), λ ∈ R,

we must have (2.3.3) for arbitrary e1 , e2 ∈ Rn , not necessarily orthogonal. Polarization

gives

(2.3.3)′ S(e1 , e2 , e1 , e4 ) = 0, e1 , e2 , e4 ∈ Rn ,

for this is a symmetric bilinear form in e2 and e4 . Trying to polarize again we just
obtain
S(e1 , e2 , e3 , e4 ) + S(e3 , e2 , e1 , e4 ) = 0,
or if we use the symmetries (2.3.2)

S(e1 , e2 , e3 , e4 ) = S(e1 , e4 , e2 , e3 ).

Thus S(e1 , e2 , e3 , e4 ) is invariant under circular permutations of e2 , e3 , e4 . By (2.3.2)

an exchange of e1 and e2 or e3 and e4 changes the sign, so

S(e1 , e2 , e3 , e4 ) = sgn πS(eπ(1) , eπ(2), eπ(3) , eπ(4) )

if π is any permutation of 1, 2, 3, 4. This means that S is an alternating 4-linear form

(hence equal to 0 if n ≤ 3). Conversely, every alternating form satisfies (2.3.3), so T is
the orthogonal space in E of the alternating forms. These are spanned by the exterior
products
X 4
Y
i 4
(e1 , . . . , e4 ) 7→ det(t , ej )i,j=1 = sgn π (tπ(i) , ei )
π 1

where the sum is taken over all permutations of 1, . . . , 4 and t1 , . . . , t4 ∈ Rn . Thus

R ∈ T if and only if R ∈ E and
X
sgn πR(tπ(1), . . . , tπ(4)) = 0.
π

Because of the symmetries (2.3.2) we can rearrange the arguments so that t1 stands
first and the permutation is positive, so the condition is equivalent to

(2.3.4) R(t1 , t2 , t3 , t4 ) + R(t1 , t4 , t2 , t3 ) + R(t1 , t3 , t4 , t2 ) = 0.

 ALGEBRAIC PROPERTIES OF THE CURVATURE TENSOR 21

This is called the first Bianchi identity. It is often written Ri[jkl] = 0 where the bracket
is read as summation over all circular permutations. The symmetry conditions are not
independent of each other. The condition on Rijlk in (2.3.2) is obviously a consequence
of the others, and the condition on Rklij follows from the first two equalities in (2.3.2)
and the Bianchi identity. In fact, they give

R(t1 , t2 , t3 , t4 )
= −R(t1 , t4 , t2 , t3 ) − R(t1 , t3 , t4 , t2 ) = R(t4 , t1 , t2 , t3 ) + R(t3 , t1 , t4 , t2 )
= −R(t4 , t3 , t1 , t2 ) − R(t4 , t2 , t3 , t1 ) − R(t3 , t2 , t1 , t4 ) − R(t3 , t4 , t2 , t1 )
= 2R(t3 , t4 , t1 , t2 ) + R(t2 , t4 , t3 , t1 ) + R(t2 , t3 , t1 , t4 )
= 2R(t3 , t4 , t1 , t2 ) − R(t2 , t1 , t4 , t3 ) = 2R(t3 , t4 , t1 , t2 ) − R(t1 , t2 , t3 , t4 )

which proves the remaining symmetry in (2.3.3).

For a submanifold M of Rn+r with dimension n the Gauss curvature at X ∈ M
of the intersection with a plane of dimension 2 + r containing the normal NX M and
two tangent vectors t1 , t2 is by Theorem 2.2.3 equal to R(t1 , t2 , t1 , t2 ) divided by the
square of the area of the parallelogram spanned by t1 and t2 . Now a tensor R ∈ T is
uniquely determined by such curvatures, for

R∈T, R(t1 , t2 , t1 , t2 ) = 0, if t1 , t2 ∈ Rn =⇒ R = 0.

In fact, the condition on R here is the condition (2.3.3), which means that R is orthog-
onal to T , hence equal to 0.
A symmetric bilinear form in ν variables has
ν(ν + 1)/2 independent coefficients
ν
whereas an antisymmetric bilinear form has 2 = ν(ν − 1)/2 independent coefficients.
Since we can interpret E as a space of symmetric bilinear forms on the space Rn ∧ Rn
which has dimension ν = n(n−1)/2,
it is clear that dim E = ν(ν+1)/2. The orthogonal
space of T in E has dimension n4 , so it follows that

dim T = ν(ν + 1)/2 − ν(n − 2)(n − 3)/12

= ν(3n(n − 1) + 6 − n2 + 5n − 6)/12 = n2 (n2 − 1)/12.

We sum up the results proved so far in a theorem:

Theorem 2.3.1. A 4-linear form R in Rn can occur as the Riemann curvature
tensor for some n dimensional submanifold of a Euclidean space if and only if it has
the symmetry properties (2.3.2) and satisfies the Bianchi identity (2.3.4). Such tensors
form a linear space T of dimension n2 (n2 − 1)/12. When R ∈ T then R is uniquely
determined by R(t1 , t2 , t1 , t2 ) for t1 , t2 ∈ Rn .
It is clear that one can find r depending only on n such that for every R ∈ T there is
a solution h of (2.3.1) with values in Rr . The dimension of the space of such quadratic
forms h is rn(n + 1)/2, so it follows from the Morse-Sard theorem that we must have
rn(n + 1)/2 ≥ n2 (n2 − 1)/12, that is,

(2.3.5) r ≥ n(n − 1)/6.

22 II. CURVATURE OF SUBMANIFOLDS OF A EUCLIDEAN SPACE

Exercise 2.3.1. Use the fact that (2.3.1) is invariant if h is replaced by Oh where
O ∈ O(r) to show that (2.3.5) can be strengthened to

n(n + 1)s − s(s − 1) ≥ n2 (n2 − 1)/6 for some s ∈ {1, 2, . . . , r},

√ √
which means that r ≥ (1 − 1/ 3)(n2 + 1) + (1 − 3)n + O(1/n) /2.

Classical results to be discussed in Chapter III will prove that one can always take
r = n(n − 1)/2. Using representation theory Berger, Bryant and Griffiths [1] have
n−1


proved that one can always take r = 2 + 2, but the best value of r does not seem
to be known. Such a value is an obvious lower bound for the codimension with which a
general Riemannian manifold can be locally C 2 embedded. In Chapter III we shall see
that a local embedding with high regularity is usually not possible with codimension
n


lower than 2 .
The full linear group GL(n) acts on T by

(gR)(t1, . . . , t4 ) = R(g −1 t1 , . . . , g −1 t4 ), g ∈ GL(n), R ∈ T , tj ∈ Rn .

There is no invariant subspace for this operation (see Berger, Bryant and Griffiths [1]).
However, in the context where we encountered the Riemann curvature tensor only the
operation of O(n) is natural, since it corresponds to changing orthonormal basis in
the tangent space. At the end of Section 2.1 we also defined the Ricci tensor, which
belongs to the space S 2 (Rn ) of symmetric bilinear forms in Rn . The passage from
the Riemann tensor to the Ricci tensor commutes with the operation of O(n), so the
kernel W consisting of all R ∈ T such that
n
X
(2.3.6) Rijil = 0, j, l = 1, . . . , n,
i=1

is invariant under O(n). (In view of the symmetries (2.3.2) the contraction of R with
respect to any pair of indices is equal to 0 if R ∈ W.) Since the Euclidean norm we
introduced in E above is invariant under O(n), the orthogonal complement of W in T is
also invariant under O(n). We shall determine the decomposition of a general R ∈ T
in components belonging to these spaces, but first we shall make some elementary
remarks on how O(n) acts on S 2 (Rn ).
The metric scalar product (·, ·) in Rn is of course invariant under O(n), and so is
its orthogonal space consisting of symmetric forms which are traceless (with respect to
the Euclidean form). If a O(n) invariant subspace of S 2 (Rn ) contains
P a form which is
not proportional to the metric form, then it contains a form λj xj yj with λ1 6= λ2 .
Hence it also contains the form with λ1 and λ2 interchanged and therefore the form
x1 y1 − x2 y2 and so the form xj yj − xk yk for arbitrary j and k. The linear hull contains
all diagonal forms with zero trace, so the only O(n) invariant subspaces of S 2 (Rn )
are the multiples of the metric form and the forms with zero trace. If n ≥ 3, as we
assume now, it follows at once from (2.2.5) that both the metric form and other forms
can occur as Ricci tensors, so the map T (Rn ) → S 2 (Rn ) corresponding to passage
 ALGEBRAIC PROPERTIES OF THE CURVATURE TENSOR 23

from the Riemann tensor to the Ricci tensor is surjective. (It is easy to show that any
symmetric tensor can in fact occur as Ricci tensor for a manifold with codimension
r = 2.)
From the surjectivity just proved it follows that
P the orthogonal space of W in T
is of dimension n(n + 1)/2. If R ∈ T and S = Rijij is the scalar curvature while
Bij = Rij − Sδij /n is the traceless Ricci tensor, then we claim that
(2.3.7) Rijkl = Wijkl
+ (Bik δjl − Bil δjk + Bjl δik − Bjk δil )/(n − 2) + S(δik δjl − δil δjk )/n(n − 1)
decomposes R in an element in W and two orthogonal ones, one which has traceless
Ricci tensor and one which has Ricci tensor proportional to the metric tensor. To
verify this it suffices to show that the last two terms are in T , for they are obviously
orthogonal to W and the contractions with respect to i, k are
(0 − Bjl + nBjl − Bjl )/(n − 2) = Bjl resp. S(nδjl − δjl )/n(n − 1) = Sδjl /n.
The symmetries (2.3.2) are obviously valid, and the Bianchi condition is then equivalent
to orthogonality to all antisymmetric 4-linear forms, which is equally obvious since
every term is symmetric in some pair of indices.
Definition 2.3.2. The component W ∈ W of R ∈ T defined by (2.3.7) is called
the Weyl tensor or the conformal curvature tensor. We have
(2.3.7)′ Rijkl = Wijkl
+ (Rik δjl − Ril δjk + Rjl δik − Rjk δil )/(n − 2) − S(δik δjl − δil δjk )/(n − 1)(n − 2)

The reason for the terminology will be clear in Chapter III, but some motivation
will be provided already by the following
Exercise 2.3.2. Calculate the Riemann curvature tensor of a sphere of radius R
in Rn+1 and decompose it according to (2.3.7).
We have proved that O(n) acts on T ⊖W with just two invariant subspaces, displayed
in the decomposition (2.3.7), but the proof that W has no O(n) invariant subspace
requires a bit of invariant theory. We must therefore refer the reader to the proof given
in Berger, Gauduchon and Mazet [1, pp. 76–78].
Exercise 2.3.3. Prove that
X X
Rijkl Riklj = − 12 Rijkl 2 .
i,j,k,l i,j,k,l

When n = 3 then n(n + 1)/2 = 6 = n2 (n2 − 1)/12 so there is no Weyl tensor. When
n > 4 it is known that W has no SO(n) invariant subspace (see references in Besse
[1, p. 49]). However, this is not true when n = 4, and since further decomposition of
W is then relevant in Yang-Mills theory, we shall discuss it briefly. (For details and
references we refer to Atiyah, Hitchin and Singer [1], Besse [1].) Since the 4-linear
forms R with the symmetries (2.3.2) can be identified with symmetric bilinear forms
on ∧2 (Rn ), we shall first discuss how SO(n) acts on ∧2 (Rn ).
24 II. CURVATURE OF SUBMANIFOLDS OF A EUCLIDEAN SPACE

Lemma 2.3.3. ∧2 (Rn ) has no proper invariant subspace under the action of SO(n)
if n 6= 4, but ∧2 (R4 ) = ∧+ ⊕ ∧− where ∧± are irreducible SO(4) invariant subspaces
of dimension 3.
Proof. If S is a skew symmetric n × n matrix, then Rn = V0 ⊕ V1 ⊕ · · · ⊕ Vj where
SV0 = 0 and Vi is of dimension 2 and SVi = Vi when i = 1, . . . , j. (See Exercise 1.4.2.)
This means that every element in ∧2 (Rn ) can be transformed by the action of SO(n)
to the form

(2.3.8) c1 e1 ∧ e2 + · · · + cj e2j−1 ∧ e2j ,

where e1 , . . . , en is the standard basis in Rn and 2j ≤ n. Suppose now that ∧ is

some SO(n) invariant subspace 6= {0} of ∧2 (Rn ), and let (2.3.8) be an element in ∧
with c1 . . . cj 6= 0. If j = 1, then ∧ will contain every element of the form (2.3.8) so
∧ = ∧2 (Rn ) then. If 2j < n, then e1 → −e1 , en → −en defines an element in SO(n)
which just replaces c1 by −c1 , so e1 ∧ e2 ∈ ∧ and ∧ = ∧2 (Rn ) then. Assume now that
2j = n. Then e1 ↔ e3 , e2 ↔ e4 defines an element in SO(n) which interchanges c1
and c2 ; by subtraction we find that e1 ∧ e2 − e3 ∧ e4 ∈ ∧ unless c1 = c2 . The lemma is
now proved if n > 4 unless n = 2j and c1 = c2 = · · · = cj . Since e1 → e3 → e5 → e1
defines an element in SO(n) we can then conclude by subtraction that

(e1 − e3 ) ∧ e2 + (e3 − e5 ) ∧ e4 + (e5 − e1 ) ∧ e6 ∈ ∧.

Since e1 − e3 + e3 − e5 + e5 − e1 = 0, the rank of this two vector is just 4 and the lemma
follows also in this case.
We are now just left with the case n = 4, and then we know that ∧ must contain
either e1 ∧ e2 + e3 ∧ e4 or e1 ∧ e2 − e3 ∧ e4 = e1 ∧ e2 + e4 ∧ e3 . In the first case
∧ must contain the space ∧+ spanned by ei ∧ ej + ek ∧ el where i, j, k, l is an even
permutation of 1, 2, 3, 4, and in the other case ∧ contains the space ∧− defined using
the odd permutations. If ϕ± ∈ ∧± then

ϕ+ ∧ ϕ+ = (ϕ+ , ϕ+ )ω, ϕ− ∧ ϕ− = −(ϕ− , ϕ− )ω, ϕ+ ∧ ϕ− = 0, (ϕ+ , ϕ− ) = 0,

where ω = e1 ∧ e2 ∧ e3 ∧ e4 and (·, ·) is the Euclidean scalar product of two vectors.

Thus W± are the eigenspaces with eigenvalues ±1 of the quadratic form ϕ ∧ ϕ/ω with
respect to (ϕ, ϕ), which proves the invariance under SO(n), since ω is invariant.
Exercise 2.3.4. Let n = 4 and take
√ √ √
(e1 ∧ e2 ± e3 ∧ e4 )/ 2, (e1 ∧ e3 ± e4 ∧ e2 )/ 2, (e1 ∧ e4 ± e2 ∧ e3 )/ 2

as basis in ∧± . With this basis the SO(4) action on ∧2 (R4 ) gives a homomorphism
SO(4) → SO(3)×SO(3). Show that the corresponding homomorphism of Lie algebras
so(4) → so(3) ⊕ so(3) is an isomorphism and determine it explicitly in terms of the
corresponding 4 × 4 and 3 × 3 skew symmetric matrices.
Remark. The decomposition when n = 4 will perhaps be more evident when we
have introduced the ∗ operator on forms later on.
 ALGEBRAIC PROPERTIES OF THE CURVATURE TENSOR 25

Still with n = 4 we now interpret the space E of 4-linear forms on R4 satisfying

(2.3.2) as the space of symmetric bilinear forms on ∧2 (R4 ) = ∧+ ⊕ ∧− . This space has
a SO(4) invariant block matrix decomposition

S 2 (∧+ ) ⊕ S 2 (∧− ) ⊕ Hom(∧+ , ∧− ),

with spaces of dimensions 6, 6 and 9. However, since we have an SO(3) action on the
first two spaces they decompose further into the multiples I(∧± ) of the metric form
and the tracless forms S02 (∧± ), so we have in fact an SO(4) invariant decomposition

S02 (∧+ ) ⊕ I(∧+ ) ⊕ S02 (∧− ) ⊕ I(∧− ) ⊕ Hom(∧+ , ∧− ),

where the spaces now have dimensions 5, 1, 5, 1, 9. Since the space of Weyl tensors
W has dimension 10, and the space of traceless Ricci tensors has dimension 9, it is
now clear that Hom(∧+ , ∧− ) is precisely the space of traceless Ricci tensors while
W = W+ ⊕ W− with W± = S02 (∧± ). The purely antisymmetric tensor and the
pure scalar curvature part S(δik δjl − δil δjk )/12 are found in the two invariant one
dimensional spaces; the antisymmetric tensors are generated by the differences of the
two metric tensors whereas the scalar curvature term corresponds to their sum. Thus
a general curvature tensor in T is represented in block matrix form by
 
W+ + SI/12 B
t
B W− + SI/12

where B corresponds to the traceless Ricci tensor and the traceless symmetric W±
together define the Weyl tensor. We leave the details of the verification for the energetic
reader.
The SO(n) invariant classification of the curvature tensors gives rise to important
classes of Riemannian manifolds:
Definition 2.3.4. A Riemannian manifold is called an Einstein manifold if the
traceless Ricci tensor vanishes, and it is called conformally flat if the Weyl tensor
vanishes. An oriented manifold of dimension 4 is called self-dual if the Weyl tensor is
in W+ and anti-self-dual if it is in W− .
26 II. CURVATURE OF SUBMANIFOLDS OF A EUCLIDEAN SPACE
 CHAPTER III

ABSTRACT RIEMANNIAN MANIFOLDS

Summary. In Section 3.1 we extend the intrinsic results proved in Chapter II to abstract
Riemannian manifolds, taking the notion of covariant differentiation as the main tool.
After introducing geodesic coordinates we also discuss various geometrical interpretations
of curvature and prove the classical Gauss-Bonnet theorem. Section 3.2 is devoted to
embedding theorems which show that abstract Riemannian manifolds are not really more
general than the submanifolds of a Euclidean space studied in Chapter II. In Section
3.3 we show that a Riemannian manifold with vanishing curvature tensor is flat, that
is, locally isomorphic to Rn ; more generally we also discuss manifolds with constant
curvature. Section 3.4 is devoted to the transformation rules for the curvature under
conformal changes of metric. For dimensions ≥ 4 it is proved that vanishing of the Weyl
tensor is necessary and sufficient for the existence of a flat conformal metric. The Yamabe
problem to find conformal metrics of constant scalar curvature is also mentioned briefly,
but we leave the study of it for another chapter.

3.1. Covariant derivatives and curvature. In Chapter II we studied n dimen-

sional C k submanifolds M of some RN . Locally they were parametrized by maps
ω ∋ x 7→ κ(x) ∈ M where ω is an open set in Rn and κ ∈ C k is injective with injective
differential. If ω̃ ∋ x̃ 7→ κ̃(x̃) ∈ M is another local parametrization, it follows that the
map
x 7→ x̃ = κ̃−1 ◦ κ(x)

is defined in the open set κ−1 (κ(ω) ∩ κ̃(ω̃)) ⊂ ω which is in C k since κ̃ has a C k left
inverse.
An abstract C k manifold M of dimension n is by definition a Hausdorff topological
space provided with a family of homeomorphisms, κ : ω → f (ω) ⊂ M with open range,
such that κ̃−1 ◦ κ ∈ C k (κ−1 (κ(ω) ∩ κ̃(ω̃)) if κ̃ is another member of the family. We shall
always assume that M has a countable dense subset. Then it is well known that M
can be embedded as a submanifold of R2n+1 , so the difference between submanifolds
of RN and abstract manifolds is more one of principle and attitude than of substance.
The tangent bundle T (M ) was defined for a submanifold of RN as the set of all
(x, t) with x ∈ M and t in the tangent space Tx (M ) of M in RN . It has a natural
generalization for an abstract manifold. One just takes Tx (M ) to be the n dimensional
vector space of first order differential operators at x which annihilate constants. In a
n
local coordinate
Pn patchj ω ⊂ R with local coordinates x, the first order operators can
n
be written 1 tj ∂/∂x , which identifies T (M ) with ω × R over ω. (From now on we
shall use superscripts for the coordinates to conform with the conventions of tensor
calculus.) If one switches to coordinates x̃ as above, then (x, t) is identified with (x̃, t̃)
if κ(x) = κ̃(x̃) and (κ̃−1 κ)′ t = t̃, which means that t̃ = (∂ x̃/∂x)t.
27
28 III. ABSTRACT RIEMANNIAN MANIFOLDS

When studying submanifolds of RN we took from the Euclidean metric in RN the

first fundamental form which we wrote
n
X
2
(3.1.1) ds = gij (x)dxi dxj
i,j=1

in terms of local coordinates. A manifold is said to be Riemannian if it is provided

with such a form:
Definition 3.1.1. A C k manifold, where k ≥ 1, is said to be Riemannian if there is
given in each fiber Tx M of T M a positive definite quadratic form such that in local
coordinates it has the form (3.1.1) with gij ∈ C k−1 .
Shifting to other local coordinates x̃ gives
X X
g̃ij (x̃)dx̃i dx̃j = g̃ij (x̃)(∂ x̃i /∂xν )dxν (∂ x̃j /∂xµ )dxµ ,

or in matrix notation (gij (x)) = t (∂ x̃/∂x)(g̃ij (x̃))(∂ x̃/∂x), so it is sufficient to assume

that the coefficients in (3.1.1) are in C k−1 for a set of coordinate patches covering M .
We note in passing that

det(gij (x)) = (det ∂ x̃/∂x)2 det(g̃ij (x̃)),

which means that det(gij (x)) dx, where dx is Lebesgue measure in Rn , is an invariant
p

definition of a positive measure (positive C k−1 density) on M , often denoted by dvol(x).

We shall now show that the results in Chapter II which are intrinsic in the sense
that they can be expressed in terms of the first fundamental form alone remain valid
for abstract Riemannian manifolds. This could be done by proving that every abstract
Riemannian manifold can at least locally or to a high degree of approximation be
embedded isometrically in RN for some N . We shall see in Section 3.2 that this is
possible. However, we shall now give a direct approach which only relies on Chapter
II for motivation and not for any proofs. This means that we shall establish invari-
ance under changes of coordinates even when we know from Chapter II that this is
true for submanifolds of RN . It would be tedious to mention precise differentiability
assumptions, so for the time being we assume that M is a C ∞ Riemannian manifold.
Let X and Y be two smooth vector fields, in local coordinates represented by
n
X n
X
j j
X= X ∂/∂x , Y = Y j ∂/∂xj .
1 1

(2.1.11) suggests that another vectorfield, the covariant derivative of Y in the direction
X, should be invariantly defined by
n
X n
X
(3.1.2) ∇X Y = (XY j + Γik j X k Y i )∂/∂xj ,
j=1 i,k=1
 COVARIANT DERIVATIVES AND CURVATURE 29

where Γik j is defined in terms of the metric by the last expression in (2.1.8), (2.1.7).
We could verify by direct computation that this definition is indeed independent of the
choice of local coordinates, but instead we shall prove that in a local coordinate patch
the linear differential operator defined by (3.1.2) has the properties

(3.1.3) ∇X Y − ∇Y X = [X, Y ],
(3.1.4) (∇Z X, Y ) + (X, ∇Z Y ) = Z(X, Y ),

where Z is a third vector field. From (3.1.3) it follows that ∇X Y is linear in X as well
as in Y . We shall also prove that ∇X is uniquely determined by (3.1.3), (3.1.4), which
implies the invariance. The proof of (3.1.3), where [X, Y ] = XY − Y X denotes the
j
commutator of the vector fields, follows immediately from the fact that P Γjik , defined
by (2.1.7), (2.1.8), is symmetric in i, k. To prove (3.1.4) we write Z = Z ∂/∂xj and
obtain
n
X n
X
j
(∇Z X, Y ) = (ZX + Γik j Z k X i )gjl Y l
j,l=1 i,k=1
n
X n
X
j l
= gjl (ZX )Y + Z k Γikl X i Y l .
j,l=1 i,k,l=1

If we exchange X and Y and add using (2.1.7)′ , it follows that (3.1.4) holds. To prove
the uniqueness means proving that if the right hand sides of (3.1.3), (3.1.4) are replaced
by 0, then ∇X Y must be 0 for arbitrary X and Y . This follows since

(∇X Y, Z) = −(Y, ∇X Z) = −(Y, ∇Z X)

= (∇Z Y, X) = (∇Y Z, X) = −(Z, ∇Y X) = −(Z, ∇X Y )

where we have alternated using the homogeneous forms of (3.1.3), (3.1.4). This implies
∇X Y = 0 as claimed. Thus we have proved:
Theorem 3.1.2. For smooth vector fields X, Y on M the vector field defined in
local coordinates by (3.1.2) is invariantly defined and the covariant differentiation ∇ is
characterized by (3.1.3), (3.1.4).
Note that ∇X Y (x) is defined even if Y is only defined on a curve with tangent X at
x. Taking for X and Y the unit tangent of a curve we can define the geodesic curvature
and geodesic principal normal as the norm and the direction of ∇X Y . This generalizes
the definition given after Theorem 2.1.3 for submanifolds of RN .
In the Euclidean case the differential operators ∇X and ∇Y act componentwise as
the scalar operators X and Y , so their commutator is ∇[X,Y ] . We shall now examine
to what extent this is true in the Riemannian case, so we form with three vector fields
X, Y, Z
S = (∇X ∇Y − ∇Y ∇X − ∇[X,Y ] )Z.
30 III. ABSTRACT RIEMANNIAN MANIFOLDS

If X is replaced by the product ϕX with some smooth function ϕ, then S is just

multiplied by ϕ, for since [ϕX, Y ] = ϕ[X, Y ] − (Y ϕ)X and

(3.1.5) ∇ϕX = ϕ∇X , ∇Y (ϕW ) = ϕ∇Y W + (Y ϕ)W,

the other terms −(Y ϕ)∇X Z + (Y ϕ)∇X Z will cancel. Thus S contains no derivatives
of X, hence no derivatives of Y by the skew symmetry. This is also true for Z since

∇X (ϕ∇Y Z + (Y ϕ)Z) − ∇Y (ϕ∇X Z + (Xϕ)Z)

= ϕ(∇X ∇Y − ∇Y ∇X )Z + (Xϕ)(∇Y Z − ∇Y Z) + (Y ϕ)(∇X Z − ∇X Z) + ([X, Y ]ϕ)Z.

We can therefore calculate S using local coordinates and assuming that the components
of X, Y, Z are constants, hence [X, Y ] = 0. Then
X X X 
∇ X ∇Y Z = ∇X Γjl i Y l Z j ∂/∂xi = ∂k Γjl i + Γkν i Γjl ν X k Y l Z j ∂/∂xi .
i,j,l ijkl ν

If we subtract the analogous formula for ∇Y ∇X Z and recall the first definition (2.1.13)
of the Riemann curvature tensor, it follows that
X
(3.1.6) (∇X ∇Y − ∇Y ∇X − ∇[X,Y ] )Z = Ri jkl X k Y l Z j ∂/∂xi .
ijkl

In particular, this proves that the definition of the Riemann curvature tensor given
by (2.1.13) is coordinate independent and does define a tensor of type 1, 3. (2.1.13)′
follows as before.
If [0, a] ∋ t 7→ γ(t) ∈ M is a (piecewise) C 1 curve and X0 ∈ Tγ(0) M , then there is
a unique vector field X(t) ∈ Tγ(t) M , t ∈ [0, a], along the curve with X(0) = X0 and
∇γ ′ (t) X(t) = 0. In fact, in local coordinates this is the Cauchy problem for a linear
system of differential equations

dX j (t) X dγ k (t) i
+ Γik j (γ(t)) X (t) = 0, j = 1, . . . , n.
dt dt
i,k

One calls X(t) the parallel translation of X0 along the curve.

Exercise 3.1.1. Show that in the local coordinates the parallel translation of X0 from
x to x + εek , where ek is the kth basis vector in Rn , is

X0 + (A(ε) + 12 A(ε)2 + O(ε3 ))X0 , as ε → 0; A(ε)ji = −εΓik j (x + 21 εek ).

e0 (ε) − X0 )/ε2 if X
Use this to find the limit as ε → 0 of (X e0 is the parallel translation
of X0 around the square from x to x + εek to x + ε(ek + el ) to x + εel to x.
 COVARIANT DERIVATIVES AND CURVATURE 31

Covariant differentiation can be extended to arbitrary tensors. Let first ω be a one

form, that is, a section of the cotangent bundle T ∗ (M ). To ω corresponds a vector field
Y , for the metric tensor identifies Tx∗ (M ) and Tx (M ) for every x; in local coordinates
n
X n
X X X
k kj
Y = g ωj , ωj = gjk Y k , if ω = ωj dxj , Y = Y k ∂/∂xk .
1 1

Sometimes the notation Y = ω ♯ and ω = Y ♭ is used for the “musical” isomorphisms

Tx∗ → Tx → Tx∗ raising and lowering indices. With that notation we now have an
obvious invariant definition of covariant differentiation of one forms ω,

∇X ω = (∇X ω ♯ )♭ .

By (3.1.4) we obtain for any vector field Y

X(ω ♯ , Y ) = (∇X ω ♯ , Y ) + (ω ♯ , ∇X Y ),

or if we use the notation h·, ·i for the duality between T ∗ and T ,

(3.1.7) Xhω, Y i = h∇X ω, Y i + hω, ∇X Y i.

This could also have been taken as a definition of ∇X ω. In local coordinates (3.1.7)
means that
n
X n
X n
X
h∇X ω, Y i = X ωj Y j − ωj (XY j + Γkl j X k Y l )
j=1 j=1 k,l=1
n
X n
X
j
= (Xωj )Y − ωl Γlkj X k Y j
j=1 j,k,l=1

so we obtain
n
X
(3.1.2) ′
(∇X ω)j = Xωj − Γkj l X k ωl .
k,l=1

ωj dxj be a one form. Show that for vector fields X, Y we

P
Exercise 3.1.2. Let ω =
have X
(∇X ∇Y − ∇Y ∇X − ∇[X,Y ] )ω = − Rj ikl X k Y l ωj dxi .

Exercise 3.1.3. Show that for tensor fields f of type k, l, that is, sections of

T (M ) ⊗ · · · ⊗ T (M ) ⊗ T ∗ (M ) ⊗ · · · ⊗ T ∗ (M )
| {z } | {z }
k times l times
32 III. ABSTRACT RIEMANNIAN MANIFOLDS

and a vector field X the formula in local coordinates

k X
n
i ...i µiν+1 ...ik
X
(3.1.2) ′′
(∇X f )ij11...ik
...jl = Xfji11...j
...ik
l
+ Γµj iν fj11...jlν−1 Xj
ν=1 µ,j=1
l
X n
X
− Γjν j µ fji11...j
...ik
ν−1 µjν+1 ...jl
Xj
ν=1 µ,j=1

defines invariantly a tensor field ∇X f of type k, l, and that this extension of ∇X is

uniquely determined by the product rule

∇X (Y1 ⊗ · · · ⊗ Yk ⊗ ω1 ⊗ · · · ⊗ ωl )
= (∇X Y1 ) ⊗ Y2 ⊗ · · · ⊗ ωl + · · · + Y1 ⊗ · · · ⊗ ωl−1 ⊗ ∇X ωl ,

where Y1 , . . . , Yk are vector fields and ω1 , . . . , ωl are one forms. Prove that the product
rule also holds for arbitrary tensor products.
To round off the definition of ∇X , we define ∇X f = Xf if f is just a function on M .
Recall that contraction of a tensor means in local coordinates that one of the upper
indices is put equal to one of the lower indices followed by summation over that index.
In particular, the contraction of Y ⊗ ω where Y is a vector and ω a one form is the
scalar product hω, Y i discussed above. For decomposable tensors as in the preceding
exercise, the contraction also means precisely taking scalar product of one of the vector
factors and one of the one form factors. Since (3.1.7) shows that for Y ⊗ ω it does
not matter if one first applies ∇X and then contracts or if one first contracts and then
applies ∇X , it follows that covariant differentiation commutes with contraction.
Let f again by a tensor field of type k, l. Since ∇X f at every point is just a linear
function of X, not depending on its derivatives, we can regard ∇X f as the contraction
of X and a tensor field ∇f of type k, l + 1,

k X
n
i ...i µiν+1 ...ik
X
(3.1.2)′′′ (∇f )ij11...i i1 ...ik
...jl ,j = ∂j fj1 ...jl +
k
Γµj iν fj11...jlν−1
ν=1 µ=1
l X
X n
− Γjν j µ fji11...j
...ik
ν−1 µjν+1 ...jl
ν=1 µ=1

Note the notation , j for the added index. Also note that ∇X f is defined at x even if f
is only defined on a curve passing through x with tangent X. A tensor f defined along
a curve t 7→ x(t) ∈ M can therefore be differentiated with respect to t as ∇x′ (t) f .
If X andP Y are two vector fields and g is the metric tensor, of type 0, 2, then
i j
(X, Y ) = gij X Y is the contraction of g ⊗ X ⊗ Y with respect to both pairs of
indices. Thus ∇(X, Y ) is the contraction of

(∇g) ⊗ X ⊗ Y + g ⊗ (∇X) ⊗ Y + g ⊗ X ⊗ ∇Y.

 COVARIANT DERIVATIVES AND CURVATURE 33

On the other hand, by (3.1.4) it is the contraction of the last two terms only. Hence it
follows that

(3.1.8) ∇g = 0.

Exercise 3.1.4. Prove (3.1.8) directly from the definitions. Show also that ∇g ♯♯ = 0.
Prove that if X, Y, Z are vector fields then ∇X ∇Y Z is the sum of the contraction of
(∇2 Z) ⊗ Y ⊗ X and that of (∇Z) ⊗ ∇X Y . Deduce that in local coordinates
X
Z i ,kj − Z i ,jk = Ri ljk Z l .
l

Exercise 3.1.5. Let f be a tensor field of type k, l, and compute the antisymmetric
part of ∇∇f , that is, in local coordinates,

fji11...j
...ik
l ,rs
− fji11...j
...ik
l ,sr
.

Exercise 3.1.6. Let V be a linear subspace of the tensor product Rn ⊗ · · · ⊗ Rn with

k + l factors, such that V is invariant under the action of the orthogonal group O(n)
on the tensor product. Show that if M is a Riemannian manifold of dimension n and
f is a tensor of type k, l at x ∈ M , then the components fiik+11 ...ik
...ik+l with respect to
any orthonormal basis for Tx M are in V if this is true for one orthonormal basis. Such
tensors therefore form a vector subbundle V of the bundle of tensors of type k, l. Prove
that if f is a section of V and X is a vector field, then ∇X f is a section of V.
We shall now discuss the notion of curvature from a somewhat different view point
which is close to its origin. Surveyors and cartographers perceived quite early the
need for solving spherical triangles (on the earth) which are too large for application
of Euclidean trigonometry but too small to make it numerically convenient to use
spherical trigonometry. (A spherical triangle is bounded by arcs of great circles.) It
was known at least since the 16th centrury that the sum of the angles in a spherical
triangle is equal to π + S/R2 where S is the area and R is the radius. Legendre
observed that when the corner angles are not close to 0 or π, then the “spherical
excess” S/R2 is approximately equally divided as an angular excess of S/3R2 at each
corner, compared to a Euclidean triangle with the same sides. The practical rule
is then to apply Euclidean trigonometry with this correction. In his fundamental
paper on surface theory Gauss [1] devoted much attention to Legendre’s theorem and
determined the next order of approximation (which gives a larger excess opposite a
shorter side), and for general surfaces he derived an analogue of Legendre’s theorem
with S/3R2 replaced by SK/3. Here K is what we know as Gauss’ curvature, which
occurs in this context naturally as an intrinsic property of the surface. Riemann [1]
followed the path initiated by Gauss in his extension to dimensions > 2. We shall now
follow these papers in principle but using modern notation of course.
The first point in Gauss [1] was to define an analogue of the great circle arcs on
the sphere. These are the shortest paths between any two of its points which have
34 III. ABSTRACT RIEMANNIAN MANIFOLDS

no antipodal points between them. To define an analogue for Riemannian manifolds,

geodesic arcs, we assume given two points x0 , x1 in the same coordinate patch ω and
look for a smooth curve [0, 1] ∋ t 7→ x(t) in ω with x(0) = x0 , x(1) = x1 , such that
Z 1 Z 1 n
1
X
s= ds = ( gjk (x)dxj /dtdxk /dt) 2 dt
0 0 j,k=1

is minimized. By Schwarz’ inequality we have

Z 1 n
X
2
s ≤ gjk (x)dxj /dtdxk /dt dt
0 j,k=1

with equality if and only if ds/dt is constant, that is, the parameter t is proportional
to the arc length. This can always be achieved, so the minimum problem is equivalent
to minimizing
Z 1 n
X
(3.1.9) I= gjk (x)dxj /dtdxk /dt dt
0 j,k=1

for all smooth curves from x0 to x1 . This has the advantage that one can expect a
unique solution if the points are sufficiently close. If the minimum is attained, then
the Euler equations
n n
d X dxk X dxi dxk
(3.1.10) 2 gjk (x) = ∂j gik (x) , j = 1, . . . , n,
dt dt dt dt
k=1 i,k=1

must be valid. (Replace x(t) by x(t) + εy(t) where y(0) = y(1) = 0, put the derivative
with respect to ε equal to 0 when ε = 0, and integrate by parts.) Carrying out the
differentiation in the left-hand side we obtain the equations
n n
X d 2 xk X dxi dxk
2 gjk (x) 2 + (2∂i gjk (x) − ∂j gik (x)) = 0, j = 1, . . . , n.
dt dt dt
k=1 i,k=1

The second sum does not change if we replace the parenthesis by

∂i gjk (x) + ∂k gji (x) − ∂j gik (x) = 2Γikj

so the differential equations (3.1.10) can be written

n
′ d 2 xk X i
k dx dx
j
(3.1.10) + Γij = 0.
dt2 i,j=1
dt dt

These equations mean that the covariant derivative of the tangent vector dx/dt along
the curve is equal to 0, that is, that the geodesic curvature vanishes (cf. Theorem
 COVARIANT DERIVATIVES AND CURVATURE 35

2.1.3). The derivation as Euler equations of an invariant variational problem shows

that the equations are invariant under a change of variables, and we know that also
from the invariance of the covariant derivative.
Without deciding yet if a solution of the minimum problem for I really exists, we
can define:
Definition 3.1.3. An integral curve of the differential equations (3.1.10) (or equivalently
(3.1.10)′ ) is called a geodesic curve.
By the basic existence theorems for ordinary differential equations the equations
(3.1.10) have a unique solution for small t with prescribed initial data x = x0 , dx/dt =
v ∈ Tx0 when t = 0. The solution x(x0 , t, v) is a C ∞ function and depends on tv rather
than on t and v, because (3.1.10) is independent of t and homogeneous in dt. This
means that x(x0 , t, v) = X(x0 , tv) where X is a C ∞ function from a neighborhood of
the zero section in T (M ) to M with X(x0 , 0) = x0 and ∂v X(x0 , 0) equal to the identity.
By the inverse function theorem we can therefore introduce X as local coordinates in
a neighborhoodP of x0 . If we do that for fixed x0 , and denote the metric in the new
coordinates by Gjk (X)dX j dX k , the fact that t 7→ tX is a solution of (3.1.10) for
every X means that

∂ X X
(3.1.10)′′ 2 Gjk (tX)X k = (∂j Gik )(tX)X iX k .
∂t

by X j and sum, it follows that Gjk (tX)X j X k is independent of

P
If we multiply
j k
P
t, thus (Gjk (tX) − Gjk (0))X X = 0. (This also follows at once by covariant
differentiation.) The derivative with respect to X j is
X X
2 (Gjk (tX) − Gjk (0))X k + t (∂j Gik )(tX)X i X k

which must therefore vanish. Combined with (3.1.10)′′ this gives

X
2Sj + 2t∂Sj /∂t = 0, Sj = (Gjk (tX) − Gjk (0))X k .

Thus tSj is independent of t, hence equal to 0, so we obtain

n
X n
X
k
(3.1.11) Gjk (X)X = Gjk (0)X k , j = 1, . . . , n.
1 1

Conversely, if we have a coordinate system such that (3.1.11) is valid, then reversing the
preceding argument shows that the straight lines through the origin (with respect to
the parameters) are geodesics. We could make a linear change of coordinates to make
Gjk (0) equal to the identity matrix. However, with pseudo-Riemannian geometry in
view, where gjk will be non-singular but not positive definite, we prefer to leave Gjk (0)
arbitrary. The coordinate systems now obtained are called geodesic coordinates; they
are uniquely determined up to a linear transformation (up to an orthogonal one if one
has insisted on the Euclidean normal form).
36 III. ABSTRACT RIEMANNIAN MANIFOLDS

From Cauchy-Schwarz’ inequality and (3.1.11) it follows that

n
 X  21
j k
Gjk (X)dX dX
j,k=1
n
X n
 X  12 n
 X  12
j k j k
≥ Gjk (0)X dX / Gjk (0)X X =d Gjk (0)X j X k ,
j,k=1 j,k=1 j,k=1

and this implies that when X is in the domain of the geodesic coordinates, then the
shortest path to the origin is indeed the geodesic ray. Thus we have proved that
geodesics do give the shortest path between any two of its points which are sufficiently
close. (The importance of the latter restriction was already clear for the sphere.)
So far in this chapter we have only discussed local properties of Riemannian man-
ifolds. A Riemannian manifold is a metric space with the distance s(x, y) between
x, y ∈ M defined as the infimum of the lengths of differentiable curves from x to y.
We add a theorem of a global nature:
Theorem 3.1.4 (Hopf-Rinow). Let M be a connected Riemannian manifold. Then
the following properties are equivalent:
(1) M is a complete metric space.
(2) Every geodesic in M can be extended indefinitely in both directions.
(3) There exists a point x ∈ M such that all geodesics starting at x can be extended
indefinitely in both directions.
(4) Every closed bounded subset of M is compact.
They imply that any two points x, y ∈ M can be joined by a geodesic of length s(x, y).
Proof. (1) =⇒ (2): Let R ⊃ (a, b) ∋ s 7→ x(s) be a geodesic with maximal interval
of definition and the arc length as parameter. Since s(x(s1 ), x(s2 )) ≤ |s1 − s2 |, the
sequence x(b − 1/k) is a Cauchy sequence if b < ∞. By the local existence theorem
for ordinary differential equations applied in a neighborhood of the limit, the geodesic
with initial data x(b − 1/k), x′ (b − 1/k) can be extended for an interval δ independent
of k for large k, which shows that (a, b) is not a maximal interval of definition and
proves (2).
The implications (4) =⇒ (1) and (2) =⇒ (3) are trivial. Assume now that (3) is
valid. This means that we have a globally defined geodesic exponential map γ : Tx M ∋
v 7→ X(1) ∈ M , giving the value at time 1 of the geodesic with initial data x, v. Since
γ is continuous and B(r) = {v ∈ Tx M ; |v|x ≤ r} is compact, it follows that γB(r) is
compact. We know already that γB(r) is equal to B(r) b = {y ∈ M ; s(x, y) ≤ r} for
small r, and it is obvious that γB(r) ⊂ B(r) for every r > 0. If we prove that equality
b
holds for every r, then (4) will be proved.
Let r1 be the supremum of all r such that γB(̺) = B(̺)b when ̺ < r. If r1 = ∞ our
claim is true so assume that r1 < ∞. Then γB(r1 ) = B(r1 ), for B(r
b b 1 ) is the closure of

∪r<r1 B(r
b 1 ) = γ ∪r<r B(r) ⊂ γB(r1 ),
1
 COVARIANT DERIVATIVES AND CURVATURE 37

and γB(r1 ) is compact and therefore closed. Now we claim that γB(r) = B(r) b for
r < r1 + δ if δ is small enough. If y ∈ B(r) \ B(r1 ), we can for large ν find a path
b b
of length < s(x, y) + 1/ν from x to y. If zν is the last point in B(r b 1 ) = γB(r1 ), then
s(x, zν ) = r1 and s(zν , y) + r1 ≤ s(x, y) + 1/ν. If z is a limit point of the sequence zν
in the compact set γB(r1 ), it follows that s(z, y) ≤ s(x, y) − r1 < δ. In view of the
compactness of γB(r) we can choose δ independently of z so small that this implies
that there is a geodesic from z to y of length s(x, y) − r1 . We also have a geodesic of
length r1 from x to y. If these did not fit together to one geodesic, we could get a path
shorter than s(x, y) from x to y by smoothing out the broken geodesic near z, which
is a contradiction completing the proof of Theorem 3.1.4.
Variations of geodesics are controlled by the Jacobi differential equations:
Theorem 3.1.5. Let γ ∈ C 3 (ω, M ) where ω is an open set in R2 , and assume that
t 7→ γ(t, s) is for fixed s a geodesic with parameter proportional to the arc length, when
(t, s) ∈ ω. Let T = γ∗ ∂/∂t be the tangent of the geodesic and let X = γ∗ ∂/∂s be a vector
field along the geodesic describing the direction in which it moves. Then X satisfies
the Jacobi differential equation ∇T ∇T X = R(T, X)T , where in local coordinates
X
R(T, X)Z = Ri jkl Z j T k X l ∂/∂xi .

Proof. If γ ′ has rank 2 then the range is locally a two dimensional surface where the
vector fields T and X are defined and commute, since ∂/∂t and ∂/∂s commute. In
this surface we have ∇T T = 0 for t 7→ γ(t, s) is a geodesic. Since [T, X] = 0 we have
∇T X − ∇X T = 0, so (3.1.6) gives

∇T ∇T X = ∇T ∇X T = [∇T , ∇X ]T + ∇X ∇T T = ∇[T,X] T + R(T, X)T = R(T, X)T,

which proves the Jacobi differential equation in the closure of the open subset of ω
where γ ′ has rank 2. In the complement in ω we have locally γ(t, s) = γ0 (ψ(t, s))
where γ0 (t) = γ(t, s0) for a suitable s0 . The geodesic equation means that ψ(t, s) is
linear in in t for fixed s. Hence X = aT where a = (∂ψ/∂s)/(∂ψ/∂t) is linear in t, and

∇T X = (∇T a)∇T, ∇T ∇T X = (∇T ∇T a)T = 0,

which completes the proof.

Exercise 3.1.7. Show that if M0 is a submanifold of the Riemannian manifold M of
codimension ν, then one can at every point in M0 choose local coordinates x1 , . . . , xn
in M such that M0 is defined by x1 = · · · = xν = 0 and
ν  j
X
k x , if 1 ≤ j ≤ ν
gjk (x)x =
k=1
0, if j > ν.

Hint: This means precisely that the rays t 7→ (tx1 , . . . , txν , xν+1 , . . . , xn ) are
geodesics orthogonal to M0 . (Note that when ν = 1 the condition means that g1k (x) =
δ1k .)
38 III. ABSTRACT RIEMANNIAN MANIFOLDS

Exercise 3.1.8. Show that if M0 is a geodesic curve in a Riemannian manifold M ,

then one can in a neighborhood of any compact interval on M0 find coordinates x =
(x1 , . . . , xn ) such that M0 is defined by x′ = (x1 , . . . , xn−1 ) = 0 and

gjk (x) = δjk + Gjk (x′ , xn ) + O(|x′ |3 ),

where Gjk (x′ , xn ) is a quadratic form in x′ depending on xn and

n−1
X
Gjk (x′ , xn )xk = 0, j = 1, . . . , n.
1

(Hint: Use Exercise 3.1.6 and make the vector fields ∂/∂xj parallel along M0 .) Con-
clude that if the differential equations

2d2 X j /dxn2 = ∂Gnn (X ′ , xn )/∂xj , j = 1, . . . , n − 1

have a solution X ′ (xn ) 6≡ 0 vanishing when xn = a and xn = b, then an interval on M0

containing [a, b] strictly cannot minimize the distance between its endpoints. Show that
the Jacobi differential equation along M0 for a vector field (X ′ (xn ), X n (xn )) means
that X n is a linear function of xn and that X ′ (xn ) satisfies the preceding differential
equations.
It follows from (3.1.11) that if we expand G in a Taylor series,

G = G0 + G1 + G2 + . . .

where Gj is homogeneous of degree j inPX, then G1 = 0, for the equations

G1jk (X)X k = 0 imply that if G1jk (X) = Gjkl X l , then Gjkl is symmetric in the
P
first two indices and antisymmetric in the last two. The permutations

jkl → jlk → ljk → lkj → klj → kjl → jkl

must therefore change the sign although we get back the same elements, which proves
the claim.
The first interesting term is therefore G2 . We shall write

G2 (X; Y ) = hG2 (X)Y, Y i,

which is a symmetric quadratic form in X as well as in Y and has the fundamental

property

(3.1.12) ∂G2 (X; Y )/∂Y = 0, if Y = X.

In particular, G2 (X; X) = 0. Since the dimension of the space of quadratic forms in

n variables is n(n + 1)/2 and that of cubic forms is n(n + 1)(n + 2)/6, it is easily seen
that the space of forms satisfying (3.1.12) is of dimension

(n(n + 1)/2)2 − n2 (n + 1)(n + 2)/6 = n2 (n2 − 1)/12.

 COVARIANT DERIVATIVES AND CURVATURE 39

We have seen in Theorem 2.3.1 that this is precisely the dimension of the space T of
curvature tensors. We can polarize G2 to a 4-linear form G2 (X1 , X2 ; Y1 , Y2 ) which is
symmetric in X1 , X2 as well as in Y1 , Y2 and has the property

G2 (X; Y ) = G2 (X, X; Y, Y ).

Then (3.1.12) yields

(3.1.12)′ G2 (X1 , X2 ; X3 , X4 ) + G2 (X3 , X1 ; X2 , X4 ) + G2 (X2 , X3 ; X1 , X4 ) = 0.

From (3.1.12)′ we easily obtain G2 (X, X; Y, Y ) = −2G2 (X, Y ; X, Y ), hence the sym-
metry

(3.1.13) G2 (X; Y ) = G2 (Y ; X).

Since G′jk (0) = 0, the corresponding Christoffel symbols vanish at the origin. (Note
that the Christoffel symbols are not tensors, for otherwise we could not make them
equal to 0 at a point by changing coordinates. The fact that we have raised and
lowered an index in the same way as for tensors might suggest otherwise, but it is just
for linear transformations that they behave as tensors.) Thus the formulas (2.1.13)′
for the curvature tensor simplify at 0 to

Rlijk = 21 (∂j ∂i Gkl + ∂k ∂l Gij − ∂j ∂l Gik − ∂i ∂k Gjl )

after cancellation of two terms ∂j ∂k Gil . This means that for the corresponding 4-linear
forms we have

(3.1.14) R(X1 , X2 ; X3 , X4 ) = G2 (X1 , X4 ; X2 , X3 ) + G2 (X2 , X3 ; X1 , X4 )

− G2 (X2 , X4 ; X1 , X3 ) − G2 (X1 , X3 ; X2 , X4 ).

Here we have only used so far that G1 = 0. For geodesic coordinates we have the
symmetry (3.1.13), which simplifies (3.1.14) to

(3.1.14)′ R(X1 , X2 ; X3 , X4 ) = 2G2 (X1 , X4 ; X2 , X3 ) − 2G2 (X1 , X3 ; X2 , X4 ).

Using (3.1.12)′ we obtain on the other hand

−6G2 (X1 , X2 ; X3 , X4 ) = R(X1 , X4 ; X2 , X3 ) + R(X1 , X3 ; X2 , X4 ),

(3.1.15)
−3G2 (X, X; Y, Y ) = R(X, Y ; X, Y ).

We have a one to one correspondence between the two 4-linear forms G2 (X1 , . . . , X4 )
and R(X1 , . . . , X4 ) with the properties:
symmetry for G2 when X1 ↔ X2 or X3 ↔ X4 or (X1 , X2 ) ↔ (X3 , X4 ); for R when
(X1 , X2 ) ↔ (X3 , X4 );
antisymmetry for R when X1 ↔ X2 or X3 ↔ X4 ;
40 III. ABSTRACT RIEMANNIAN MANIFOLDS

circular antisymmetry for both G2 and R as described in (3.1.12)′ . For R this is

again the first Bianchi identity.
Exercise 3.1.9. Show that at the center of a geodesic coordinate system we have
X
Γijk (x) = 1
3 (Rikjl (0) + Riljk (0))xl + O(|x|2 ).
l

What is the analogue for Γij k ?

Exercise 3.1.10. Show that at the center of a geodesic coordinate system we have
X X
Rijkl 2 = 12 Gijkl 2 .
i,j,k,l i,j,k,l

The preceding observations were expressed in words by Riemann [1, p. 279]: “Führt
man diese GrössenPein, so wird für unendlich kleine Werthe von x das Quadrat des
Linienelements = dx2 , das Glied der Nächsten Ordnung in demselben aber gleich
einem homogenen Ausdruck zweiten Grades der n(n − 1)/2 Grössen (x1 dx2 − x2 dx1 ),
(x1 dx3 − x3 dx1 ), . . . , also eine unendlich kleine Grösse der vierten Dimension, so
dass man eine endliche Grösse erhält wenn man sie durch das Quadrat des unendlich
kleinen Dreiecks dividirt, in dessen Eckpunkten die Werthe der Veränderlichen sind
(0, 0, 0, . . . ), (x1 , x2 , x3 , . . . ), (dx1 , dx2 , dx3 , . . . ). Diese Grösse behält denselben Werth,
so lange die Grössen x und dx in denselben binären Linearformen enthalten sind, oder
so lange die beiden kürzesten Linien von den Werthen 0 bis zu den Werthen x und
von den Werthen 0 bis zu den Werthen dx in demselben Flächenelement bleiben, und
hängt also nur von Ort und Richtung desselben ab. Sie wird offenbar = 0, wenn
die
P dargestellte Manningfaltigkeit eben, d.h. das Quadrat des Linienelements auf
2
dx reducirbar ist, und kann daher als das Mass der in diesem Punkte in dieser
Flächenrichtung stattfindenden Abweichung der Mannigfaltigkeit von der Ebenheit
angesehen werden. Multiplicirt mit −3/4 wird sie der Grösse gleich, welche Herr
Geheimer Hofrath Gauss das Krümmungsmass einer Fläche genannt hat.”
The factor 4 here comes from the fact that Riemann divided by the square of the
area of a triangle and not the corresponding parallelogram. The factor −3 is the same
as in (3.1.15) above, and we shall now show that it is also closely connected to the
denominator 3 in Legendre’s theorem.
Still with geodesic coordinates we shall consider the geodesic triangle with corners
at 0, εY ,P εZ when ε is small. For the sides from 0 the lengths are ε|Y | and ε|Z| where
|X|2 = Gjk (0)X j X k ; we denote the corresponding scalar product by (·, ·). The
third side is not as easy to determine since we do not know the geodesic. However, it
is clear that the square of its length is equal to ε2 times the square of the length of the
geodesic from Y to Z for the metric
X
Gjk (εX)dX j dX k = |dX|2 + ε2 G2 (X; dX) + O(ε3 ).
 COVARIANT DERIVATIVES AND CURVATURE 41

This geodesic must differ from the straight line segment [0, 1] ∋ t 7→ Y + t(Z − Y ) by
O(ε2 ), and since the straight line is a geodesic for the metric |dX|2 , we obtain for the
geodesic distance εa between εY and εZ
Z 1
2 2 2
a = |Y − Z| + ε G2 (Y + t(Z − Y ); Z − Y ) dt + O(ε3 )
0
= |Y − Z|2 + ε2 G2 (Y ; Z) + O(ε3 ),

for G2 (Y + t(Z − Y ); Z − Y ) = G2 (Y ; Z − Y ) = G2 (Y ; Z). The Riemannian angle αr

at 0 is defined by (Y, Z) = |Y ||Z| cos αr , and the angle αe opposite εa in the Euclidean
triangle with sides ε|Y |, ε|Z|, εa is given by the cosine theorem

a2 = |Y |2 + |Z|2 − 2|Y ||Z| cos αe , hence

2|Y ||Z|(cos αr − cos αe ) = ε2 G2 (Y ; Z) + O(ε3 ).

Thus δ = αr − αe = O(ε2 ) and

G2 (Y ; Z) |εY ||εZ| sin αr

δ=− + O(ε3 ).
(|Y ||Z| sin αr )2 2

Since the first factor is −1/3 times the sectional curvature in the Y Z plane (cf. (3.1.15)
and Theorem 2.2.3), and the second factor is the area of the geodesic triangle +O(ε3 ),
we have proved Legendre’s theorem. Note that the total angle excess is equal to the
sectional curvature times the area +O(ε3 ).
Exercise 3.1.11. Assuming that the earth is a sphere with circumference 40000 km,
and that T is an equilateral geodesic triangle on the earth with the base equal to 100
km and angles 60◦ at the base, estimate how many seconds of arc the third angle differs
from 60◦ .
In the two dimensional case considered by Gauss we can make a subdivision of the
geodesic triangle T by geodesics joining the midpoints of each side. Repeating this
subdivision and noting that angle excess is additive when we subdivide, we conclude
in the limit that
ZZ
(3.1.16) α+β+γ = π+ K dS
T

where α, β, γ are the angles at the corners, K is the Gauss curvature, and dS is the
Riemannian area measure. This is Gauss’ part of the Gauss-Bonnet theorem; it was
extended by Bonnet to more general regions. This we shall now do with a different
more analytical proof which will introduce some ideas which will be important later
on.
Let us assume that we have a Riemannian metric in a simply connected open set
ω ⊂ R2 . We shall generalize Theorem 1.2.1 by calculating the integral of the signed
42 III. ABSTRACT RIEMANNIAN MANIFOLDS

geodesic curvature κg over a simple closed curve s 7→ x(s) ∈ ω, 0 ≤ s ≤ L. We assume

that s is the arc length, that is,

2
X dxj (s) dxk (s)
gjk (x(s)) = 1,
ds ds
j,k=1

and we denote by e(s) the unit tangent vector e(s) = x′ (s). The covariant derivative
of e along the curve has the components

2
j dej X dxk
(∇e e) = + Γik j (x(s))ei (s) ,
ds ds
i,k=1

but we shall avoid explicit calculations using this expression. Let n be the unit vector
orthogonal to e such that e, n is positively oriented. Then the signed geodesic curvature
κg is given by
κg = (∇e e, n),
and our task is to calculate the integral
Z L
(3.1.17) κg ds.
0

This is the integral of the differential form

2
X 2
X
(3.1.18) κ= gjl (dej + Γik j (x)ei dxk )nl
j,l=1 k,i=1

in the unit circle bundle

2
X
S(T ω) = {(x, w) ∈ T (ω); gjk (x)wj wk = 1},
j,k=1

the integral being taken along the curve γ : s 7→ (x(s), e(s)). (Recall that n is uniquely
determined by e.) Here e(s) = x′ (s), but we shall simplify the problem by generalizing
it, so we allow any closed C 1 curve
R γ in S(T ω) now.
We shall first examine how γ κ depends on the choice of the unit vector e. Any
other choice ẽ can be written

ẽ = e cos θ + n sin θ, thus ñ = −e sin θ + n cos θ,

where θ is uniquely determined modulo 2π, so dθ is uniquely determined. We obtain

∇ẽ = (∇e) cos θ + (∇n) sin θ + (−e sin θ + n cos θ)dθ,

 COVARIANT DERIVATIVES AND CURVATURE 43

and since (e, e) = (n, n) = 1, (e, n) = 0, it follows from (3.1.4) that

(∇e, e) = 0, (∇n, n) = 0, (∇e, n) + (e, ∇n) = 0.

Hence
(∇ẽ, ñ) = (∇e, n) + dθ,
R R RL
so κ = γ κ + 0 dθ if γ̃ is the curve s 7→ (x(s), ẽ(s)).
γ̃
If v and ṽ are two arbitrary vector fields with no zeros defined along the curve
s 7→ x(s), then we can normalize them with respect to the Riemannian metric to unit
RL
vector fields e and ẽ and introduce the variation 0 dθ of the angle from e to ẽ along the
curve. This is an integer multiple of 2π which must be independent of the Riemannian
metric. In fact, if we replace gjk by λgjk + (1 − λ)δjk , 0 ≤ λ ≤ 1, then the angle
variation depends continuously on λ so it must have the same value when λ = 0 as
when λ = 1. If for example one of the vector fields is tangent to the curve and the
other is ∂/∂x1 , then it follows from Theorem 1.2.1 that the angle variation between
them is ±2π. R
We shall now determine the integral γ κ when e is a unit vector field defined in the
whole of ω, for example the normalization of the vector field ∂/∂x1 . The advantage
of this is that we can then pull the form κ back by the R map x 7→ (x, e(x)) to a form
in ω with integral over the curve s 7→ x(s) equal to γ κ. (It also follows that γ is a
boundary in S(T ω).) The interior ω ′ of the curve is ⋐ ω since ω was assumed simply
connected, and we shall calculate the integral of the differential form over ∂ω ′ using
Stokes’ formula. If εi = ∂/∂xi and we write ∇i = ∇εi for the sake of brevity, the
differential form to integrate is

2
X
(∇i e, n) dxi .
i=1

The integral over ∂ω ′ is equal to the integral over ω ′ of

(3.1.19) ∇1 (∇2 e, n) −∇2 (∇1 e, n) = ((∇1 ∇2 −∇2 ∇1 )e, n) +(∇2 e, ∇1 n) −(∇1 e, ∇2 n).

Here we have used (3.1.4). The last two terms are equal to 0 since ∇j e has the direction
of n and ∇k n has the direction of e. By (3.1.6) the first term, on the right is equal
to R(n, e, ε1 , ε2 ) = −R(e, n, ε1 , ε2 ). The vectors e, n are positively oriented and span
√
a parallelogram with area 1 while ε1 , ε2 span one of area g where g = det(gjk ). In
√
view of Theorem 2.2.3 we therefore conclude that (3.1.19) is equal to −K g, where
K is the total (Gaussian) curvature, and we have proved:
Theorem 3.1.6 (Gauss-Bonnet). Let ω ⊂ R2 be a simply connected open set with
Riemannian metric, and let ω ′ ⋐ ω be simply connected with C 2 boundary. Then
Z Z
(3.1.20) κg ds + K dS = 2π,
∂ω ′ ω′
44 III. ABSTRACT RIEMANNIAN MANIFOLDS

where K is the total curvature, κg is the geodesic curvature of ∂ω ′ with the orientation
making ω ′ lie to the left of ∂ω ′ , ds is the arc length of ∂ω ′ , and dS is the Riemannian
√
area element g dx in ω ′ .
In the proof we have actually proved
√
(3.1.21) dκ = π ∗ (−K gdx1 ∧ dx2 )

where π is the projection S(T ω) → ω and κ is defined by (3.1.18). In fact, since

exterior differentiation commutes with pullbacks, we have just proved that the two
sides of (3.1.21) are equal when pulled back by any local section x 7→ (x, e(x)) of
S(T ω). This implies (3.1.21), for if u is a differential form in a fiber space over a
manifold M and s∗ u = 0 for every local section of M , then u = 0 if the degree of u
does not exceed dim M . (Verify this as an exercise.) Note that κ cannot be obtained
by lifting a form from ω; in fact, we have shown that the restriction of κ to any fiber
of S(T ω) is equal to the natural one form on the oriented circle.
In Theorem 3.1.6 we assumed that ∂ω ′ was in C 2 , so (3.1.20) does not quite cover
(3.1.16). However, it is easy to extend the theorem to the case where ∂ω ′ is only
piecewise C 2 . To get a closed curve in S(T ω) we must then add at each corner x
a circular arc in the fiber connecting the incoming tangent to the outgoing tangent;
alternatively we can approximate ω ′ by domains with the corners rounded off. This
gives
Z X Z
′
(3.1.20) κg ds + αj + K dS = 2π,
∂ω ′ ω′

where αj denote the exterior angles at the corners and κg is integrated only over the
smooth part of ∂ω ′ . For a geodesic triangle, the three angles are π − αj , so the sum is
X Z
3π − αj = π + K dS.
ω′

Thus (3.1.20)′ contains (3.1.16). It is also easy to obtain (3.1.20) from (3.1.16) applied
to a triangulation of a polygonal approximation of ω ′ . We leave this also as an exercise.
We shall now discuss the case of a compact oriented Riemannian manifold M of
dimension 2 (without boundary). Suppose that M is decomposed by geodesic arcs
into a finite number ν2 of geodesic polygons ωj . Denote the interior angles of ωj by
βjk . Then (3.1.20)′ yields
X Z
(π − βjk ) + K dS = 2π.
k ωj

The number of terms in the sum is equal to the number of sides of ωj . If ν0 and ν1
denote the total number of corners and sides occurring in some ωj , we get by adding
Z
K dS + 2πν1 − 2πν0 = 2πν2 ,
M
 COVARIANT DERIVATIVES AND CURVATURE 45

for there will be altogether 2ν1 terms π in the left-hand side. Thus

1
Z
(3.1.22) K dS = ν2 − ν1 + ν0 ,
2π M

so the right-hand side is independent of how the decomposition of M is made; it is the

Euler characteristic of M .
We can approach the calculation of the integral in (3.1.22) in another way starting
from a function f on M which has only non-degenerate critical points. Then the vector
field F = (df )♯ , usually denoted grad f , has only finitely many zeros. Let

Γ = {(x, F (x)/kF (x)k); F (x) 6= 0} ⊂ S(T M ).

This is a manifold whose boundary consists of the circles over the zeros of F , that
is, the critical points of f . For at such a point we can introduce geodesic coordinates
diagonalizing the quadratic terms in f , that is,

gjk (x) = δjk + O(|x|2 ), f (x) = f (0) + (f1 (x1 )2 + f2 (x2 )2 )/2 + O(|x|3 ).

Then we have F (x) = (f1 x1 , f2 x2 ) + O(|x|2 ) and

p
F (x)/kF (x)k = (f1 x1 , f2 x2 )/ (f1 x1 )2 + (f2 x2 )2 + O(|x|).

When x winds around a small circle |x| = ε then F (x)/kF (x)k winds around the unit
circle in the same or opposite direction depending on the sign of f1 f2 . The boundary
of Γ is the limit of the image of the negatively oriented circle of radius ε as ε → 0, so
it consists of the circles in S(T M ) with orientation opposite to the sign of the Hessian
of f at the critical point. If we integrate (3.1.21) over Γ it follows now from Stokes’
formula that Z Z Z X
− K dS = dκ = κ = −2π εj
Γ ∂Γ

where εj is the sign of the Hessian of f at the critical point. Thus we have
Theorem 3.1.7. If M is a compact two dimensional
R oriented Riemannian manifold
with total curvature K and no boundary, then M K dS/2π is for every real valued
function on M with only non-degenerate critical points equal to the sum of the signs
of the Hessian of f at the critical points. The integral is also equal to the Euler
characteristic.
Instead of the vector field F = (df )♯ we could have used here any vector field with
only non-degenerate fixed points (that is, zeros). We shall later on give an extension
due to Chern of the preceding arguments which is applicable to any oriented manifold of
even dimension. The problem is to find the appropriate differential forms in S(T M ).
To do so we need a systematic approach to Riemannian geometry using differential
forms; the proof of Theorem 3.1.6 was meant to motivate the need for that.
R
We shall now supplement Theorem 3.1.7 by studying the integral M K dS for a
compact oriented hypersurface M ⊂ Rn+1 of any dimension n. We denote by γ the
46 III. ABSTRACT RIEMANNIAN MANIFOLDS

Gauss map (2.2.3); the direction of the normal n is chosen so that a positive system of
tangent vectors followed by n is a positive system in Rn+1 . We choose the orientation
of S n so that with this definition the normal at x ∈ S n is −x. Set γ̌ = −γ, which is
then the identity map if M = S n . The degree D of the map γ̌ is then defined by
Z Z
∗
γ̌ u = D u
M Sn

where u is an arbitrary n-form on S n . We choose for u the volume form on S n , which

means that u(t1 , . . . , tn ) is the n-dimensional volume with sign of the parallelepiped
spanned by the tangent vectors t1 , . . . , tn at a point on S n . By Theorem 2.2.2
Y

(γ̌ ∗ u)(t1 , . . . , tn ) = u(γ̌ ′ t1 , . . . , γ̌ ′ tn ) = Kj u(t1 , . . . , tn ).

Now t1 , . . . , tn have as tangent vectors of M the same orientation as they have as

tangent vectors of S n at −n, since the normal is n there. Hence γ̌ ∗ u equals K times
the volume form of M , and we obtain
Z Z
(3.1.23) K dS = D dS,
M Sn

where dS denotes the area element in M and that in S n in the two integrals.
When n = 2 it follows in view of Theorem 3.1.7 that the degree of the mapping −γ
is equal to half the Euler characteristic. Later on we shall extend this to arbitrary n
such that (n − 2)/4 is an integer.
We shall finally prove an important result on the covariant derivative of the curvature
tensor. For a geodesic system of coordinates the Christoffel symbols are O(|x|), so the
non-linear terms in (2.1.13)′ are O(|x|2 ). Hence we obtain

Rijkl = 21 (∂j ∂k gil + ∂i ∂l gjk − ∂i ∂k gjl − ∂j ∂l gik ) + O(|x|2 ),

and it follows that

Rijkl,m = 21 (∂j ∂m ∂k gil + ∂i ∂l ∂m gjk − ∂i ∂m ∂k gjl − ∂j ∂l ∂m gik )

at the center of the geodesic coordinates. Note that the first and the last term are
equal apart from the sign and a circular permutation of the indices klm, and that this
is also true for the middle terms. Hence we obtain the second Bianchi identity

(3.1.24) Rijkl,m + Rijlm,k + Rijmk,l = 0,

or more briefly Rij[kl,m] = 0. Since Rijkl,m is a tensor, this is of course true for any
system of coordinates. Contraction of (3.1.24) with respect to the indices ik and the
indices jl, that is, multiplication by g ik g jl and summation gives, since contraction and
the musical isomorphisms commute with covariant differentiation
X X
S,m − g ik Rim,k − g jl Rjm,l = 0,
 LOCAL ISOMETRIC EMBEDDING 47

that is,
X
(3.1.25) ∂k S = 2 g ij Rik,j .

Here S is the scalar curvature and Rij is the Ricci tensor.

By Definition 2.3.4 a Riemannian manifold M is said to be an Einstein manifold
if Rij = f gij for some function f on M . This implies that S = nf , where n is
the dimension of M . Since Rik,j = gik ∂j f in view of (3.1.8), using (3.1.25) gives
n∂k f = 2∂k f , k = 1, . . . , n. If n > 2 and M is connected, it follows that f is a
constant, so we have proved:
Theorem 3.1.8. If M is a connected Einstein manifold of dimension > 2, then the
Ricci tensor is a constant multiple of the metric tensor.
3.2. Local isometric embedding. Let us assume given a smooth Riemannian
metric in a neighborhood of 0 in Rn . We want to find a map x 7→ f (x) ∈ RN defined
in a neighborhood of 0 so that the given metric is the same as the metric
P introduced on
the embedded manifold by the Euclidean metric, that is, |df (x)|2 = gjk (x)dxj dxk ,
or explicitly

(3.2.1) (∂j f (x), ∂k f (x)) = gjk (x), 1 ≤ j ≤ k ≤ n.

Altogether there are n(n+1)/2 equations, so it follows from general theorems discussed
at the end of this section that smooth solutions do not exist in general unless N ≥ n(n+
1)/2. We shall now discuss a classical theorem of Janet and Cartan (see Jacobowitz
[1] and references there) which shows that a local real analytic solution always exists
when N = n(n + 1)/2 if gjk are real analytic. The idea of the proof is to argue by
induction with respect to n, and extend a local isometric embedding of Rn−1 × {0} by
solving a Cauchy problem. To do so one must cut down the number of equations to
solve:
Lemma 3.2.1. The equations (3.2.1) are valid in a ball with center at 0 if and only
if they hold for 1 ≤ j ≤ k < n when xn = 0 and in addition we have the equations
(∂n2 f , ∂j f ) = Γnnj , 1 ≤ j ≤ n,
(3.2.2)
(∂n2 f , ∂i ∂j f ) = ∂n Γijn + (∂i ∂n f, ∂j ∂n f ) − 21 ∂i ∂j gnn , 1 ≤ i ≤ j < n,

and the boundary conditions

(∂n f, ∂j f ) = gnj , 1 ≤ j ≤ n, xn = 0,
(3.2.3)
(∂n f, ∂i ∂j f ) = Γijn , 1 ≤ i ≤ j < n, xn = 0.

Proof. The first n equations (3.2.2) and the second set of initial conditions (3.2.3)
follow from the definition (2.1.7) of the Christoffel symbols. To prove the second set
of equations (3.2.2) we note that

∂n Γijn = ∂n (∂n f, ∂i ∂j f ) = (∂n2 f, ∂i ∂j f ) + (∂n f, ∂n ∂i ∂j f ),

(3.2.4) 1
∂∂ g
2 i j nn
= 21 ∂i ∂j (∂n f, ∂n f ) = (∂n f, ∂n ∂i ∂j f ) + (∂i ∂n f, ∂j ∂n f ).
48 III. ABSTRACT RIEMANNIAN MANIFOLDS

Subtraction yields the remaining equations (3.2.2).

Now assume that we have a solution of (3.2.2), (3.2.3) such that (3.2.1) is valid when
xn = 0 for 1 ≤ j ≤ k < n. The first equation (3.2.2) with j = n shows that

∂n (∂n f, ∂n f ) = 2Γnnn = ∂n gnn ,

so using the first boundary condition (3.2.3) with j = n we obtain (∂n f, ∂n f ) = gnn
also when xn 6= 0. From the first equation (3.2.2) with j < n we now obtain

∂n (∂n f, ∂j f ) = Γnnj + (∂n f, ∂j ∂n f )

= Γnnj + 21 ∂j (∂n f, ∂n f ) = Γnnj + 12 ∂j gnn = ∂n gnj ,

and using the boundary condition we conclude that (∂n f, ∂j f ) = gnj . Now we have
for 1 ≤ i ≤ j < n

∂n (∂i f, ∂j f ) = (∂i ∂n f, ∂j f ) + (∂i f, ∂j ∂n f )

= ∂i (∂n f, ∂j f ) + ∂j (∂i f, ∂n f ) − 2(∂n f, ∂i ∂j f ) = ∂i gjn + ∂j gin − 2(∂n f, ∂i ∂j f ).

When xn = 0 this is equal to ∂i gjn + ∂j gin − 2Γijn = ∂n gij by the second set of initial
conditions (3.2.3). Differentiating again we obtain

∂n2 (∂i f, ∂j f ) = ∂n ∂i gjn + ∂n ∂j gin − 2∂n Γijn = ∂n2 gij ,

where we have used (3.2.4). In view of the initial conditions it follows that
(∂i f, ∂j f ) = gij for 1 ≤ i ≤ j < n, which proves the lemma.
As already mentioned we shall start from a local isometric embedding of Rn−1 , and
we must choose it so that the equations (3.2.3) with j < n can be solved. This requires
a stronger condition on the embedding, which we formulate with Rn−1 replaced by
Rn in the following definition:
Definition 3.2.2. A C 2 map x 7→ f (x) from a neighborhood of 0 in Rn to RN is said
to be free at 0 if the derivatives

∂i f (0), ∂i∂j f (0), 1≤i≤j≤n

are linearly independent. Their linear hull is then called the osculating space at 0.
It is clear that f is free at every point in a neighborhood of 0 if f is free at 0. If a
free map exists then N ≥ n + n(n + 1)/2, the dimension of the osculating space. Note
that when n is replaced by n − 1, this condition becomes N ≥ (n − 1) + (n − 1)n/2 =
n(n + 1)/2 − 1.
gjk (x)dxj dxk is a real analytic Riemann-
P
Theorem 3.2.3 (Janet-Cartan). If
ian metric in a neighborhood of 0 in Rn then there is a local real analytic isometric
embedding in RN with N = n(n + 1)/2. It can be chosen free if N = n(n + 3)/2.
 LOCAL ISOMETRIC EMBEDDING 49

Proof. When n = 1 the statement is obvious. We may therefore assume that n > 1
and that the theorem has already been proved for dimensions smaller than n. Without
restriction we may assume that the coordinates are geodesic or at least that

gjk (x) − δjk = O(|x|2 ), as x → 0.

Using the inductive hypothesis we choose a free real analytic map f0 from a neighbor-
hood of 0 in Rn−1 to RN−1 , where N = n(n + 1)/2, such that for x′ in a neighborhood
of 0
(∂j f0 (x′ ), ∂k f0 (x′ )) = gjk (x′ , 0), 1 ≤ j ≤ k < n.
We want to find f satisfying (3.2.2) so that f (x′ , 0) = (f0 (x′ ), 0) and (3.2.3) is valid.
The equations

(∂n f (x′ , 0), ∂j f0 (x′ )) = gnj (x′ , 0), 1 ≤ j < n,

′ ′ ′
(∂n f (x , 0), ∂i ∂j f0 (x )) = Γijn (x , 0), 1 ≤ i ≤ j < n,

determine uniquely the component v(x′ ) of ∂n f (x′ , 0) in RN−1 , because f0 is free, and
v(0) = 0 since the right-hand sides of these equations vanish at 0. The only remaining
equation (3.2.3) can now be written

(∂n fN (x′ , 0))2 = gnn (x′ , 0) − kv(x′ )k2

where k · k denotes the Euclidean norm. Since the right-hand side is positive when
x′ = 0 we have a unique analytic positive solution. Summing up, we have well defined
analytic boundary conditions
p
(3.2.5) f (x′ , 0) = (f0 (x′ ), 0), ∂n f (x′ , 0) = (v(x′ ), gnn (x′ , 0) − kv(x′ )k2 ).

The last coordinate of ∂n f (x′ , 0) is not 0, so ∂j f (0), 1 ≤ j ≤ n, and ∂i ∂j f (0), 1 ≤ i ≤

j < n, form a basis for RN . The equations (3.2.2) can therefore be solved for ∂n2 f ,

(3.2.6) ∂n2 f = Φ(x, {∂j f }j≤n , {∂i ∂j f }i<n,j≤n ),

where Φ is analytic in a neighborhood of the initial data at 0. This is a Kovalevsky

system so it has a unique analytic solution with the data (3.2.5) in a neighborhood of
the origin.
It remains to show that we can get a free embedding in a space of dimension N + n.
To do so, we first construct an embedding in RN+n−1 by changing (3.2.5) to

f (x′ , 0) = (f0 (x′ ), 0, . . . , 0),

| {z }
(3.2.5)′ n times
p
′ ′
∂n f (x , 0) = (v(x ), gnn (x′ , 0) − kv(x′ )k2 − |x′ |2 , x′ ).

The equations (3.2.2) still have a unique analytic solution (3.2.6) if we require that
∂n2 f shall lie in the linear hull of ∂j f , 1 ≤ j ≤ n and ∂i ∂j f , 1 ≤ i ≤ j < n. Then it is
50 III. ABSTRACT RIEMANNIAN MANIFOLDS

clear that all derivatives ∂j f and ∂i ∂j f with i ≤ j are linearly independent at 0 with
the exception of ∂n2 f , and since we have an isometric embedding
P in R
N+n−1
we could
not hope for more. If f is such an embedding for the metric gjk dx dx − 4(xn dxn )2 ,
j k

then the embedding (f, (xn )2 ) in RN+n will be free.

If the metric is not real analytic, the proof breaks down for there is no reason
then why the Cauchy problem should be solvable. Of course one can still use the
completely elementary formal part of the Cauchy-Kovalevsky theorem and conclude
that if gjk ∈ C ∞ then f can be chosen so that (3.2.1) holds with an error vanishing of
infinite order at 0. Even for n = 2 it is not known whether the first part of Theorem
3.2.3 is valid in the C ∞ case when the Gauss curvature has a non-simple zero at the
origin. (See e.g. Jacobowitz [2].) However, the second part of Theorem 3.2.3 remains
true as will now be shown following Günther [1].
Assume now just that gij ∈ C 2+̺ for some ̺ ∈ (0, 1), that is, that ∂ α gij is
Hölder continuous of order ̺ when |α| ≤ 2. Replacing gij by the second order Tay-
lor polynomial we obtain from Theorem 3.2.3 a free embedding f with values in RN ,
N = n(n + 3)/2, defined in a neighborhood of 0 such that

∂ α (gij (x) − (∂i f (x), ∂j f (x)) = O(|x|̺+2−|α| ) as x → 0.

Choose ϕ ∈ C0∞ (Rn ) such that ϕ(x) = 1 when |x| ≤ 1 and ϕ(x) = 0 when |x| ≥ 2,
and set for small ε


hij (x) = ϕ(x/ε) gij (x) − (∂i f (x), ∂j f (x)) .

Then |x| ≤ 2ε if x ∈ supp hij , gij (x) = hij (x) + (∂i f (x), ∂j f (x)) if |x| < ε, and

(3.2.7) sup |∂ α hij (x)| = O(ε̺+2−|α| ), if |α| ≤ 2.

Let Ω ⊂ Rn be a ball with center at 0 so small that f is free in Ω. It suffices to

show that for small ε there is a map u ∈ C 2+̺ from Ω to RN such that |d(f + u)|2 =
|df |2 + hij dxi dxj , that is,
P

(3.2.8) (∂i u, ∂j f ) + (∂j u, ∂i f ) + (∂i u, ∂j u) = hij .

We choose ε so small that hij vanishes on ∂Ω. If u P vanishes on ∂Ω then (3.2.8) is

equivalent to the equation obtained by letting ∆ = ∂k2 act on both sides, which
gives


∆hij = ∆ (∂i u, ∂j f ) + (∂j u, ∂i f ) + ∂i (∆u, ∂j u) + ∂j (∆u, ∂i u) + Tij (u)



= ∂i (∆u, ∂j u) + ∆(∂j f, u) + ∂j (∆u, ∂i u) + ∆(∂i f, u) − 2∆(∂i ∂j f, u) + Tij (u),

where Tij (u) is quadratic in the second derivatives of u. We can simplify the equations
by adding just n equations which annihilate the first two terms in all these equations
and conclude that (3.2.8) is valid if

∆(∂i f,u) = −(∂i u, ∆u), 1 ≤ i ≤ n,

1
∆(∂i ∂j f,u) = 2
(Tij (u) − ∆hij ), 1 ≤ i ≤ j ≤ n,
 LOCAL ISOMETRIC EMBEDDING 51

in Ω and u = 0 on ∂Ω. Let G be the operator solving the Dirichlet problem for the
Laplacian in Ω, thus ∆Gψ = ψ in Ω and Gψ = 0 on ∂Ω if ψ ∈ C(Ω). Then these
conditions are fulfilled if

(∂i f,u) = −G(∂i u, ∆u), 1 ≤ i ≤ n,

(3.2.9)
(∂i ∂j f,u) = 21 G(Tij (u) − ∆hij ), 1 ≤ i ≤ j ≤ n.

The vectors ∂i f (x), ∂i ∂j f (x) with 1 ≤ i ≤ j ≤ n form a basis for RN for every x ∈ Ω
since f is free, so we have for every U ∈ RN
X X
U= ϕi (∂i f, U ) + ϕij (∂i ∂j f, U ),
1≤i≤n 1≤i≤j≤n

where ϕi , ϕij are analytic in Ω. Thus the equations (3.2.9) are equivalent to u =
H + T (u) where
X X X
H = − 12 ϕij hij , T (u) = − ϕi G(∂i u, ∆u) + 1
2 ϕij GTij (u).
1≤i≤j≤n 1≤i≤n 1≤i≤j≤n

Standard Hölder estimates in the theory of elliptic equations show that T (u) is
continuous in C 2+̺ and a contraction operator in the convex subset where
α
P
|α|≤2 sup |∂ u| is small enough. In view of (3.2.7) it follows that the equation
u = H + T (u) has a unique solution there if ε is small enough. Arguing by stan-
dard elliptic theory we conclude that u is as smooth as h if h has additional regularity.
Admitting these basic facts without proof here, we obtain:
gjk (x)dxj dxk is a Riemannian metric in a neighborhood of 0
P
Theorem 3.2.4. If
in R with coefficients in C 2+̺ for some non-integer ̺ > 0, then there is a local C 2+̺
n

free isometric embedding in RN with N = n(n + 3)/2.

Note that the solution we have found to the first order equations (3.2.8) is not
smoother than the right hand side. This is unavoidable because of the Gauss equations
and gives rise to the analytical difficulties of the problem which were first overcome
by Nash [1]. A great deal is known also about global isometric embeddings. We
must confine ourselves here to referring to Gromov and Rohlin [1], Berger, Bryant and
Griffiths [1], and the references in these papers.
We shall finally justify the statement made above that a smooth isometric embedding
in RN does not exist in general unless N ≥ n(n + 1)/2. It is notationally more
convenient to prove a general form of this statement. Let u denote a C ∞ function
defined in a neighborhood of 0 in Rn with values in RN . Denote by Jm u the m-jet of
u,
Jm u = {∂ α u}|α|≤m ,
and let Φ(x, Jm u) be a C ∞ function of x and Jm u with values in Rν . Thus the
equation Φ(x, Jm u) = v is a general system of ν differential equations of order m for
N unknowns.
52 III. ABSTRACT RIEMANNIAN MANIFOLDS

Theorem 3.2.5. If for all v in an open subset of C ∞ (Rn , Rν ) one can find a function
u ∈ C ∞ (Rn , RN ) such that Φ(x, Jm u) = v in a neighborhood of 0, then N ≥ ν.
Proof. Let µ be some large integer to be chosen later. We shall not really use the full
differential equation but only the much weaker condition

(3.2.10) Jµ Φ(x, Jm u)(0) = Jµ v(0).

The space Eµ of µ-jets at 0 of functions in C
∞ (Rn , Rν ), to which the right-hand

side belongs, has dimension ν µ+n , for µ+n


n n
is the number of multi-indices α =
(α1 , . . . , αn ) with α1 + · · · + αn ≤ µ. The left-hand
side is a C ∞ function of Jµ+m u(0),
which belongs to a space of dimension N µ+m+n n . By the Morse-Sard theorem the
range is of measure 0 in Eµ if
    n
µ+m+n µ+n Y µ+j
N <ν , that is, N < ν .
n n j=1
µ+m+j

If N < ν then this condition is fulfilled for large µ, which proves the theorem and even
more: If N < ν then the equation Φ(x, Jm u) = v cannot even be satisfied to order µ
at a fixed point when µ is large, unless Jµ v is exceptional there in the sense that it
belongs to a set of measure 0.
For the isometric embedding problem one can improve the estimate of µ obtained
from the preceding proof by using the special properties of the equations. (See Exercise
2.3.1 and Gromov-Rokhlin [1].) Whether local isometric embedding of low regularity
is possible in dimensions below the Janet-Cartan dimension n(n + 1)/2 does not seem
to be known except for the result of Nash [2], Kuipers [1]; they showed that even a
global C 1 isometric embedding is possible in essentially the same dimension where a
C 1 embedding exists, hence always in 2n + 1 dimensions. However, for C k embeddings
with k ≥ 2 the situation is quite different as shown by (2.3.5), because the Gauss
equations are then available, and the problem does not seem to have been studied
then.
3.3. Spaces of constant curvature. Let M be a connected Riemannian manifold
of dimension n, and recall from Section 2.3 that the sectional curvature of M at a
point x ∈ M for the two plane spanned by t1 , t2 ∈ Tx M is defined by

(3.3.1) R(t1 , t2 , t1 , t2 )/(g(t1 , t1 )g(t2 , t2 ) − g(t1 , t2 )2 ).

If this is a function f (x) independent of the direction of the two plane then

R(t1 , t2 , t3 , t4 ) = f (x)(g(t1, t3 )g(t2 , t4 ) − g(t1 , t4 )g(t2 , t3 ))

by the uniqueness statement in Theorem 2.3.1, for the right-hand side has all the
symmetry properties defining T . Thus the Weyl tensor and the traceless Ricci tensor
are equal to 0. If this is true for every point in M , then M is an Einstein manifold, and
 SPACES OF CONSTANT CURVATURE 53

if n > 2 it follows from Theorem 3.1.8 that f is a constant, so the sectional curvature is
independent of x also. (This is a classical theorem of F. Schur, far older than Theorem
3.1.8.)
Definition 3.3.1. A connected Riemannian manifold M is said to have constant cur-
vature K if the curvature tensor is given by

(3.3.2) Rijkl = K(gik gjl − gil gjk ),

with a constant K.
From (3.3.2) it follows that the Ricci tensor is (n − 1)Kgij and that the scalar
curvature is n(n − 1)K. Thus a manifold of constant curvature is an Einstein manifold
but the converse is not true when n > 3 since the Weyl tensor may not be equal to 0
then. Since the covariant derivative of the metric tensor is equal to 0, this is also true
for the curvature tensor (3.3.2).
An obvious example of a manifold of constant zero curvature is Rn with the standard
Euclidean metric. We shall now prove that locally there are no others.
Theorem 3.3.2. A Riemannian manifold M with curvature tensor identically equal
to 0 is flat in P
the sense that at every point one can choose local coordinates such that
the metric is (dxj )2 .
Proof. Choose first some arbitrary local coordinates x = (x1 , . . . , xn ) varying over a
ball Ω with center at the origin. Set ∇i = ∇∂i where ∂i = ∂/∂xi . Since [∂i , ∂j ] = 0 it
follows from (3.1.6) and the hypothesis that [∇i , ∇j ] = 0. For every v0 ∈ Rn we can
therefore find a unique vector field v in Ω with v(0) = v0 and ∇i v = 0, i = 1, . . . , n. In
fact, we shall prove inductively for ν = 1, . . . , n that there is such a vector field defined
in Ων = {x ∈ Ω; xj = 0, j > ν} with ∇i v = 0, i ≤ ν. This is obvious when ν = 1,
for the equation ∇1 v = 0 is just a linear system of differential equations with leading
term ∂v/∂x1 which we can solve with initial value v0 . If ν > 1 and ṽ has the required
property with ν replaced by ν − 1, then we can find v defined in Ων with ∇ν v = 0 and
v = ṽ when xν = 0. If µ < ν it follows that ∇ν ∇µ v = ∇µ ∇ν v = 0 in Ων , and since
∇µ v = ∇µ ṽ = 0 when xν = 0, it follows that ∇µ v = 0 in Ων . When ν = n, the claim
is proved.
Now choose vector fields v1 , . . . , vn in Ω such that ∇vj = 0 for j = 1, . . . , n and
v1 , . . . , vn form an orthonormal basis at 0. Since d(vj , vk ) = (∇vj , vk ) + (vj , ∇vk ) = 0,
this follows at every point in Ω. We have

[vj , vk ] = ∇vj vk − ∇vk vj = 0,

so there are new coordinates y in a neighborhood of 0 such that vj = ∂/∂y j , j =

1, . . . , n. But this means that the metric is the Euclidean metric in these coordinates.
An equally obvious example of a manifold of constant positive curvature K is a
1
sphere with radius R = K − 2 in Rn+1 . In fact, O(n + 1) is a transitive group of
isometries, and the isotropy group O(n) leaving a given point x ∈ S n fixed is transitive
on the two planes in the tangent plane at x. Analytically the sphere is defined by
54 III. ABSTRACT RIEMANNIAN MANIFOLDS

|x|2 + (xn+1 )2 = R2 where x = (x1 , . . . , xn ). In the hemisphere where xn+1 > 0 we

can use x as parameter, and noting that hx, dxi + xn+1 dxn+1 = 0 we obtain

ds2 = |dx|2 + hx, dxi2 /(xn+1 )2 = |dx|2 + Khx, dxi2 + O(|x|4 )

where K = 1/R2 . Thus gij = δij + Kxi xj + O(|x|4 ), and we obtain when x = 0 by
(3.1.14)

i l
Rijkl = K
2 (∂j ∂k x x + ∂i ∂l xj xk − ∂i ∂k xj xl − ∂j ∂l xi xk ) = K(δjl δik − δil δkj ).

This we knew already of course, for when n = 2 the total curvature is 1/R2 = K.
For the restriction to the hyperboloid H = {x ∈ Rn+1 ; (xn+1 )2 = |x|2 + R2 , xn+1 >
0} of the hyperbolic (Lorentz) metric |dx|2 − |dxn+1 |2 we have

ds2 = |dx|2 − hx, dxi2 /(xn+1 )2 = |dx|2 + Khx, dxi2 + O(|x|4 )

where K = −1/R2 < 0 now. The metric is positive definite since |x|2 < (xn+1 )2 on H.
The preceding calculation for the sphere gives that

Rijkl = K(δjl δik − δil δkj )

at the origin. To prove that the curvature is constant it suffices to note that the group
of
PnLorentz transformations, that is, linear transformations preserving the Lorentz form
j 2
1 (x ) − (xn+1 )2 acts isometrically and transitively on H.
The parametrization of S n above covers only a half sphere. We can cover the whole
sphere minus one point by means of the stereographic projection. If X ∈ S n then the
stereographic projection x ∈ Rn from the point (0, . . . , 0, −R) to the tangent plane at
the antipodal point is obtained from the equations

X = (0, . . . , 0, −R) + t(x, 2R), |X| = R, that is, t = 4R2 /(4R2 + |x|2 ).

Thus

|dX|2 = |tdx + xdt|2 + 4R2 dt2 = t2 |dx|2 + (4R2 + |x|2 )dt2 + 2tdthx, dxi = t2 |dx|2 ,

for dt(4R2 + |x|2 ) + 2thx, dxi = 0. Hence the metric is

ds2 = (1 + |x|2 /4R2 )−2 |dx|2 = |dx|2 (1 − K 2

2 |x| ) + O(|x|4 )

where K = 1/R2 again. We leave as an exercise to calculate the curvature at 0 using

this expression for the metric.
For the hyperboloid H we can similarly use stereographic projection from
(0, . . . , 0, −R) on the tangent plane at (0, . . . , 0, R). This gives the equations

X = (0, . . . , 0, −R) + t(x, 2R), X ∈ H, that is, t = 4R2 /(4R2 − |x|2 ).

 SPACES OF CONSTANT CURVATURE 55

The stereographic projection just fills the open ball with radius 2R now. We obtain
n
X
|dX j |2 − |dX n+1 |2 = |tdx + xdt|2 − 4R2 dt2
1
= t2 |dx|2 + (|x|2 − 4R2 )dt2 + 2tdthx, dxi = t2 |dx|2 .
With K = −1/R2 we therefore obtain the metric
(3.3.3) ds2 = (1 + K|x|2 /4)−2 |dx|2 .
For K = −1 this is the Poincaré model of non-Euclidean geometry. By an inversion
we pass to the half space model. First we move the boundary point (0, . . . , 0, −2R) to
0 by introducing
y = x + (0, . . . , 0, 2R); 4R2 − |x|2 = −|y|2 + 4Ry n .
With z = y/|y|2 we make an inversion at the origin and obtain

|dx|2 = |dy|2 = |dz/|z|2 − 2zhz, dzi/|z|4 |2 = |dz|2 /|z|4 ,

4Ry n − |y|2 = (4Rz n − 1)/|z|2
so the metric is 16R4 (4Rz n − 1)−2 |dz|2 or after a translation, x = z − (0, . . . , 0, 1/4R)
(3.3.4) ds2 = R2 |dx|2 /(xn )2 .
Exercise 3.3.1. Verify directly that the metric (3.3.4) in the half space where xn > 0
has constant curvature −1/R2 .
Exercise 3.3.2. Find geodesic coordinates at 0 for the metric (3.3.3).
We obtained the Poincaré model starting from an immersion in Minkowski space.
Schur [1] gave an isometric immersion of a part of the n dimensional hyperbolic space
with curvature −1 in R2n−1 as follows. Set
1 n−1
X 1 + iX 2 = eix /xn , . . . , X 2n−3 + iX 2n−2 = eix /xn .
Then
2n−1
X n−1
X
j 2 n −2
(dX ) = (x ) (dxj )2 + (n − 1)(xn )−4 (dxn )2 + (dX 2n−1 )2
1 1

which is equal to the Poincaré metric if

(dX 2n−1 )2 = (dxn )2 ((xn )2 − (n − 1))(xn )−4 .
Choosing an integral of this√differential equation we obtain an isometric immersion of
the half space where xn > n − 1. According to a reference in Gromov and Rokhlin
[1] it was proved in Liber [1] that a local isometric embedding is not possible in R2n−2 .
A classical result of Hilbert states that for n = 2 it is not possible to embed the whole
hyperbolic plane in R3 ; this was extended by Efimov [1] to arbitrary complete two
dimensional surfaces with a negative upper bound for the curvature.
We shall now prove an extension of Theorem 3.3.2 to any constant curvature:
56 III. ABSTRACT RIEMANNIAN MANIFOLDS

Theorem 3.3.3. A Riemannian manifold with constant curvature K is isometric near

any point to a sphere, if K > 0, to Rn if K = 0 and to hyperbolic space if K < 0.
Proof. Assume that we have geodesic coordinates in a ball Ω with center at 0, gij (0) =
δij . Then the theorem states that the metric tensor gij is uniquely determined by K
in a neighborhood of the origin. This will be proved by deriving differential equations
along the rays, that is, the integral curves of the radial vector field
n
X
̺= xj ∂ j .
1

We shall prove in Lemma 3.3.4 below that for j = 1, . . . , n there is a unique vector
field ej such that

(3.3.5) ∇̺ ej = 0, ej (0) = ∂j .

(This is not quite obvious since the radial vector field vanishes at 0.) Since

∇̺ (ej , ek ) = (∇̺ ej , ek ) + (ej , ∇̺ ek ) = 0,

where (·, ·) denotes scalar product in the Riemannian metric, it follows that e1 , . . . , en
are an orthonormal system at every point in Ω. Let θ 1 , . . . , θ n be the one forms
biorthogonal to e1 , . . . , en , that is,

hθ j , ek i = δkj , j, k = 1, . . . , n.
i
P
Then t = i hθ , tiei for every tangent vector, so
X X
gjk tj tk = hθ i , ti2 .
i

Writing X
θi = Aij dxj , that is, Aij = hθ i , ∂j i,

we obtain gjk = i Aij Aik , so it is enough to find the coefficients Aij .

P

Since h∇̺ θ j , ek i = 0 for j, k = 1, . . . , n, we have ∇̺ θ j = 0 also. Thus, by (3.1.3),

̺Aij = hθ i , ∇̺ ∂j i = hθ i , ∇j ̺ − ∂j i,

for [̺, ∂j ] = −∂j . Now hθ i , ̺i = xi , for ∇̺ ̺ = ̺ since the radial direction is parallel, so
̺hθ i , ̺i = hθ i , ̺i, which means that hθ i , ̺i is homogeneous of P degree 1. At the origin
i i
we have θ = dx , so the assertion follows. It means that ̺ = xi ei . Thus
X X
(̺ + 1)Aij = hθ i , ∇j ̺i = ∂j xi − h∇j θ i , ̺i = δji − xk h∇j θ i , ek i = δji + xk Bkj
i
.

Here we have introduced

i
Bjk = hθ i , ∇k ej i
 SPACES OF CONSTANT CURVATURE 57

and must now obtain a differential equation for this new quantity. By (3.1.6)
i
̺Bjk = hθ i , ∇̺ ∇k ej i = −hθ i , ∇k ej i + R(ei , ej , ̺, ∂k ),

for [̺, ∂k ] = −∂k and ∇̺ ej = 0. Thus

X
i
(̺ + 1)Bjk = R(ei , ej , ̺, el )hθ l , ∂k i.
l

Since the covariant derivative of R is equal to 0, we have

̺R(ei , ej , ̺, el ) = R(ei , ej , ∇̺ ̺, el ) = R(ei , ej , ̺, el ),

so this is a homogeneous function of degree 1. At the origin we have

R(ei , ej , ∂m , el ) = Rijml = K(δik δjl − δil δjk ),

so it follows that X
R(ei , ej , ̺, el ) = xm Rijml.
m

Summing up and introducing polar coordinates x = rω, we have the differential equa-
tions
d X
(rAij ) = δji + rBkji
ωk ,
dr
d i
X
(rBjk )= Rijml ω m rAlk
dr
l,m

for the functions rAij and rBjk

i
which vanish at the origin. This determines them
uniquely and proves the theorem when we have established the following
Lemma 3.3.4. Let ̺ be the radial vector field in the ball Ω in a geodesic system of
coordinates. For every v0 ∈ Rn one can then find a unique C ∞ vector field v with
v(0) = v0 and ∇̺ v = 0. It follows that ∇v = 0 at the origin.
Proof. We have to solve a singular Cauchy problem of the form
X
(3.3.6) xj ∂j v + Av = 0, v(0) = v0 ,

where (Av)j = Γik j xk v i has coefficients vanishing of second order at the origin. If
P
we introduce polar coordinates, x = rω, the problem takes the form
X
∂v j /∂r + Γik j (rω)ω k v i = 0, v = v0 when r = 0.

It is immediately clear that we have a unique solution in [0, R) × S n−1 , if Ω is the

ball with radius R, and it is a C ∞ function there. We have to verify that it is a C ∞
58 III. ABSTRACT RIEMANNIAN MANIFOLDS

function in the original variables x1 , . . . , xn . To do so we observe that (3.3.6) can be

solved in terms of formal power series

v = v0 + v1 + v2 + . . .

where vj is homogeneous of degree j. In fact, (3.3.6) can be written

v1 + 2v2 + · · · + A(v0 + v1 + . . . ) = 0

and gives v1 = 0, 2v2j +

P j k i
γik x v0 and so on, where γik j is the first order Taylor
expansion of Γik j . Now write v = v0 + · · · + vN−1 + w for some large N . Then w = 0
when r = 0, and X
∂wj /∂r + Γik j (rω)ω k wi = f j (r, ω)

where f = O(r N−1 ). This implies that w = O(r N ), and since ∂/∂xν = ων ∂/∂r +
h∂ω/∂xν , ∂/∂ωi, where ∂ω/∂xν is homogeneous of degree −1, it follows that w ∈ C N−1
as a function of the original variables, the derivatives vanishing at 0. This completes
the proof of the lemma and of the theorem.
The tools introduced to prove Theorem 3.3.3 also give another useful result on the
connection between the curvature tensor and the metric tensor.
(x)dxj dxk be the metric form for a geodesic coordinate
P
Theorem 3.3.5. Let gjkP
system centered at 0, thus gjk (x)xk = xj . Any derivative ∂ α gjk (0) can then be
expressed as a polynomial in the components of the curvature tensor and its covariant
derivatives of order ≤ |α| − 2.
Proof. It suffices to prove the theorem with covariant derivatives replaced by deriva-
tives ∂ α with respect to the coordinates. In fact, the difference

(3.3.7) Rijkl,α1 ...αm − ∂αm . . . ∂α1 Rijkl

is at the origin a polynomial in derivatives of R of order < m and derivatives of the

metric tensor g of order ≤ m since the Christoffel symbol just contains the first order
derivatives of g. If we already know that these derivatives of g can be expressed in
terms of derivatives of R of order ≤ m − 2, it follows that (3.3.7) can be expressed in
terms of derivatives of R of order < m. By induction with respect to m we therefore
conclude that every polynomial in derivatives of R of order ≤ m can be written as a
polynomial in covariant derivatives of R of order ≤ m.
Thus it suffices to examine how the Taylor expansion of R determines that of g.
Since (gij (0)) is the identity it suffices in fact to consider the inverse matrix g ij , which
is somewhat more convenient since for every t ∈ Rn , regarded as a covector,
X X X
g ij ti tj = hei , ti2 , hence g ij = em i em j .
m

Here em are the vector fields in the proof of Theorem 3.3.3, said to form a synchronous
frame for T M . If we prove that derivatives of em of order ≤ ν at 0 are polynomials in
those of R of order ≤ ν − 2, the theorem will be proved.
 SPACES OF CONSTANT CURVATURE 59

In the proof of Theorem 3.3.3 we saw that

X
hθ i , (∇̺ + 1)∇l em i = R(ei , em , ̺, ∂l), that is, ((∇̺ + 1)∇l em )i = Ri jkl ejm xk .
j,k

The operator (̺ + 1) multiplies a homogeneous function by the degree of homogeneity

plus 1. If we take the homogeneous terms of order µ in the Taylor expansion we find
for µ ≥ 0 that the derivatives of order µ of ∇l em at 0 are polynomials in those of R
and of em , of total order at most µ − 1. Since the first order derivatives of em vanish
at 0 it follows inductively that all derivatives of em at 0 of order µ + 1 are polynomials
in those of R of order ≤ µ − 1, which completes the proof. Taking the first order term
in ∇l em above we also obtain
X
(∇l em )i = 12 Ri mkl (0)xk + O(|x|2 ),
k

which gives the following result which will be needed in Section 6.10:
Theorem 3.3.6. If e1 , . . . , en is the synchronous frame, equal to the basis vectors at
the origin of a geodesic coordinate system, then
X
∇l ej = 21 Ri jkl (0)xk ei + O(|x|2 ),
i,k

and the same result holds for the dual frame in T ∗ M .

Exercise 3.3.3. Show that withPthe notation in Theorem 3.3.6 ∂k (∇l ej − Γjl i ei )(0) is
symmetric in k and l. Express gjk (x)dxj dxk in terms of R(0) with an error O(|x|3 ).
Remark. For a manifold of constant curvature, the curvature tensor is a polynomial
in the metric tensor. Hence it follows inductively from Theorem 3.3.5 that we can
calculate all derivatives of the metric tensor at the origin of a geodesic system of
coordinates. If we already knew that the metric is analytic, this would give another
proof of Theorem 3.3.3, so the connection between Theorems 3.3.3 and 3.3.5 is quite
close.
Exercise 3.3.4. Let M and N be Riemannian manifolds of the same constant curvature
K such that N is complete. Let γ : [0, 1] ∋ t 7→ x(t) ∈ M be a smooth arc, and let f0
be an isometry of a neighborhood of x(0) on a neighborhood of some point y(0) ∈ N .
(1) Prove that, restricting f0 if necessary, one can find a smooth arc γ̃ : [0, 1] ∋ t 7→
y(t) ∈ N and for every t ∈ [0, 1] an isometry ft of a neighborhood U (t) of x(t)
on a neighborhood of y(t) such that ft = fs in U (t) ∩ U (s) for all s, t ∈ [0, 1].
(2) Show that γ̃ and the germ of f1 at x(1) are uniquely determined by γ and the
germ of f0 at x(0).
(3) Prove that if M is simply connected, then y(1) and the germ of f1 are uniquely
determined by x(0), x(1) and f0 .
(4) Deduce that if M is complete and simply connected, then f0 can be extended
to a locally isometric map f : M → N , and that M is then for K > 0, K = 0
or K < 0 isomorphic to a sphere, to Rn , or to hyperbolic space.
60 III. ABSTRACT RIEMANNIAN MANIFOLDS

3.4. Conformal geometry. A Riemannian metric allows one to measure both

angles and distances. In some cases only the angles are interesting, so metrics differing
by a scalar factor give the same results. Also in analytical contexts where the metric
originates from the principal part of a second order differential operator, one is often
able to simplify by multiplication of the equation and therefore the metric by a non-
vanishing factor.
Definition 3.4.1. If M and M
f are Riemannian manifolds with metrics g and g̃, then a
diffeomorphism f : M → M f is said to be conformal if f ∗ g̃ = e2ϕ g for some smooth ϕ.
When M = M f we shall assume that f is the identity unless otherwise stated.
The two dimensional case is very special so we shall discuss it first.
Theorem 3.4.2. Every Riemannian manifold M of dimension 2 is locally conformal
to the flat space R2 .
P2
Proof. A metric j,k=1 gjk dxj dxk in a neighborhood of 0 in R2 is conformal to the
Euclidean metric if and only if g12 = 0 and g11 = g22 . This means precisely that the
one forms dx1 and dx2 are orthogonal and of equal pointwise norm.
Assuming that gjk = δjk + O(|x|2 ), as we may, we set ω1 = dx1 and
p
ω2 = (g 11 dx2 − g 12 dx1 )/ g 11 g 22 − (g 12 )2 .
Then ω1 and ω2 are orthogonal and have the same norm; ω2 − dx2 is O(|x|2 ). We can
choose a complex valued function u with u(0) 6= 0 such that d(u(ω1 + iω2 )) = 0, for
this means that
du ∧ (ω1 + iω2 ) + ud(ω1 + iω2 ) = 0,
which reduces to ∂u/∂x1 + i∂u/∂x2 = 0 at the origin, so it is an elliptic differential
equation. In a connected neighborhood of the origin we can now write
u(ω1 + iω2 ) = dy 1 + idy 2
where y 1 and y 2 are real valued and dy 1 , dy 2 are orthogonal and of equal norm, for dy 1
and dy 2 are obtained from ω1 and ω2 by an orthogonal transformation and a rotation.
If we take y 1 and y 2 as new coordinates, the metric is a multiple of the Euclidean
metric (dy 1 )2 + (dy 2 )2 , which proves the theorem.
Remark. What we have used here is that an oriented two dimensional Riemannian
manifold has an analytic structure; we have chosen y 1 + iy 2 analytic with respect to
it.
From now on we assume that the dimension n is at least 3.
Theorem 3.4.3. If g̃ = e2ϕ g, then we have for the corresponding curvature tensors
eijkl = Rijkl + gil ϕjk + gjk ϕil − gik ϕjl − gjl ϕik + (gil gjk − gik gjl )|∇ϕ|2
(3.4.1) e−2ϕ R
where ϕij = ϕ,ij − ∂i ϕ∂j ϕ, calculated in terms of the metric g.
A direct verification is possible but laborious and uninteresting. We shall therefore
prepare the proof with an elementary lemma which allows us to avoid messy calcula-
tions.
 CONFORMAL GEOMETRY 61

Lemma 3.4.4. If a ∈ Rn then the metric (1 + 2hx, ai + |x|2 |a|2 )−2 |dx|2 in Rn is flat
except at the singularity where |a|2 x = −a 6= 0.
Proof. If we take y = |a|2 x + a, the metric becomes |y|−4 |dy|2 , and the inversion
z = y/|y|2 reduces it to |dz|2 , so it is flat. (See also the argument leading to (3.3.4).)
Proof of Theorem 3.4.3. If ϕ is a constant, the statement is obvious, so we may assume
that ϕ(0) = 0. Next assume that g is the Euclidean metric in Rn and that

ϕ(x) = ϕa (x) = − log(1 + 2hx, ai + |x|2 |a|2 ) = −2hx, ai − |x|2 |a|2 + 2hx, ai2 + O(|x|3 )

as for the conformal factor in Lemma 3.4.4. Then we have

ϕij = 2(2ai aj − δij |a|2 − 2ai aj ) = −2δij |a|2

and it follows at once that the right-hand side of (3.4.1) vanishes, so Theorem 3.4.3 is
valid in this case. In the general case we write ϕ(x) = ϕa (x) + q(x) + O(|x|3 ) where
q is quadratic, and we assume that gjk (x) = δjk + Gjk (x) + O(|x|3 ) where Gjk is a
quadratic form. Then
X
g̃ = e2ϕa g0 + G + 2qg0 + O(|x|3 ); g0 = (dxj )2 .

Only the first term contributes to the Christoffel symbols Γ eijk for the metric g̃ at 0.
′
From the linearity of the formula (2.1.13) in the second order derivatives we conclude
e is the sum of the curvature tensor for the metric e2ϕa g0 , which is 0, the curvature
that R
tensor R and the curvature tensor for the metric g0 + 2qg0 . The latter has the ijkl
component
gil ∂j ∂k q + gjk ∂i ∂l q − gjl ∂i ∂k q − gik ∂j ∂l q
which agrees with the terms in (3.4.1) of second order in ϕ. The proof is complete.
If we contract (3.4.1) in the indices jl we obtain the transformation law for the Ricci
curvature:
X
(3.4.2) Reik = Rik + (2 − n)ϕik − ( g jl ϕjl + (n − 1)|∇ϕ|2 )gik .

Another contraction gives the transformation rule for the scalar curvature:

(3.4.3) e2ϕ Se = S + 2(1 − n)∆ϕ − (n − 1)(n − 2)|∇ϕ|2 ,

where we have introduced the Laplace-Beltrami operator

X
(3.4.4) ∆ϕ = g ij ϕ,ij .

With this notation we can rewrite (3.4.2) in the form

(3.4.2)′ eij = Rij + (2 − n)ϕij − (∆ϕ + (n − 2)|∇ϕ|2 )gij .

R
62 III. ABSTRACT RIEMANNIAN MANIFOLDS

If we multiply (3.4.3) by gij /(2 − 2n) and add to (3.4.2)′ , we eliminate ∆ϕ and obtain

R
eij + Sg̃
e ij /(2 − 2n) = Rij + Sgij /(2 − 2n) + (2 − n)Φij , Φij = ϕij + 12 |∇ϕ|2 gij .

If we note that (3.4.1) can be written

(3.4.1)′ e−2ϕ R
eijkl = Rijkl + gil Φjk + gjk Φil − gik Φjl − gjl Φik

it follows that

−2ϕ
 g̃il R
ejk + g̃jk R
eil − g̃ik R
ejl − g̃il R
eik S(g̃
e il g̃jk − g̃ik g̃jl ) 
e Rijkl +
e −
n−2 (n − 1)(n − 2)
gil Rjk + gjk Ril − gik Rjl − gil Rik S(gil gjk − gik gjl )
= Rijkl + − .
n−2 (n − 1)(n − 2)

Comparison with (2.3.7)′ shows that this means precisely that

(3.4.5) e−2ϕ W
fijkl = Wijkl

for the corresponding Weyl tensors. Hence we have the following corollary to Theorem
3.4.3:
Corollary 3.4.4. Under the hypotheses of Theorem 3.4.3 we have the transformation
law (3.4.5) for the corresponding Weyl tensors.
Exercise 3.4.1. Show that if g̃ = e2ϕ g then the corresponding covariant derivatives are
related by
e X Y = ∇X Y + (Xϕ)Y + (Y ϕ)X − (X, Y )(dϕ)♯ ,
∇

where the scalar product and ♯ are defined by the metric g.

A metric is called conformally flat if it is conformal to a flat metric. From Corollary
3.4.4 it follows at once that a necessary condition for this is that the Weyl tensor
vanishes. (When H. Weyl introduced this tensor he called it the conformal curvature
tensor. This term is still used occasionally, and the term Weyl tensor may then have a
different meaning. This is the case in Gerretsen [1], where somewhat different proofs of
the following results can be found.) We shall now examine if it is a sufficient condition
too.
If the Weyl tensor of the Riemannian manifold M with metric tensor gij is equal to
0, then the Weyl tensor of the manifold M f = M with metric tensor g̃ij = e2ϕ gij is 0
for arbitrary ϕ, so the curvature tensor of M f vanishes if and only if the Ricci tensor is
′
equal to 0, that is by (3.4.2)

Rij + (2 − n)ϕij − (∆ϕ + (n − 2)|∇ϕ|2 )gij = 0, i, j = 1, . . . , n.

 CONFORMAL GEOMETRY 63

To clarify the role of the Weyl tensor we shall discuss these equations for the vanishing
of the Ricci P
tensor Reij without assuming at first that W = 0. Let Φ be the one form
∇ϕ = dϕ = ∂i ϕdxi . Then the preceding equations can be written
X
(3.4.6) Rij + (2 − n)(Φi,j − Φi Φj ) − ( g kl Φk,l + (n − 2)|Φ|2 )gij = 0.
k,l

Conversely, if we can find a one form satisfying these equations in an open set Ω ⊂ M ,
we obtain using (3.1.2)′

0 = Φi,j − Φj,i = ∂j Φi − ∂i Φj

since Γij l is symmetric in i, j. Thus dΦ = 0 in Ω, so Φ = dϕ for some ϕ if Ω is simply

connected, and ϕ will then have the desired properties.
Contraction of (3.4.6) gives (see (3.4.3))
X
S + 2(1 − n) g kl Φk,l − (n − 2)(n − 1)|Φ|2 = 0,
k,l

or equivalently
X
2(n − 1)( g kl Φk,l + (n − 2)|Φ|2 ) = S + (n − 2)(n − 1)|Φ|2 .
k,l

We can therefore rewrite the equations (3.4.6) in the form

(3.4.7) Φi,j = Φi Φj + ωij − 12 |Φ|2 gij ,

(3.4.8) ωij = (Rij − Sgij /(2n − 2))/(n − 2).

Note that (2.3.7)′ can be written

(3.4.9) Rijkl = Wijkl + ωik gjl − ωil gjk + ωjl gik − ωjk gil .

If we introduce the definition Φi,j = ∂j Φi − l Γij l Φl in (3.4.7), we obtain a system

P
of the form (C.6) discussed in Theorem C.3 in the appendix, with Φ in the role of the
y variables there. To find the Frobenius integrability condition (C.7) we should apply
∂k and subtract the equation with j and k interchanged, and then use the equations
(3.4.6) to express the derivatives of Φ. Clearly this gives the same result as if we take
the covariant derivative of (3.4.6) and use that
X
(3.4.10) Φi,jk − Φi,kj = Rν ijk Φν .
ν

Here the right-hand side can be expressed using (3.4.9), and we have

Φi,jk = Φi,k Φj + Φi Φj,k + ωij,k − 21 gij ∂k |Φ|2 .

64 III. ABSTRACT RIEMANNIAN MANIFOLDS

(3.4.6) implies Φj,k = Φk,j , as already observed, and

(3.4.11) Φi,k Φj − Φi,j Φk = ωik Φj − ωij Φk − 21 |Φ|2 (gik Φj − gij Φk ).

Furthermore,
X
(3.4.12) 1
∇ |Φ|2
2 k
= (∇k Φ, Φ) = g lm Φl,k Φm
l,m
X X
= g lm (Φl Φk + ωlk − 21 glk |Φ|2 )Φm = g lm ωlk Φm + 12 Φk |Φ|2 .
l,m l,m

With Φν denoting the coordinates of Φ♯ , it follows from (3.4.9) that the right-hand
side of (3.4.10) can be written
X X X
Wνijk Φν + gik ωνj Φν − gij ωνk Φν + ωik Φj − ωij Φk .
ν ν ν

The left-hand side is

ωik Φj − ωij Φk − 12 |Φ|2 (gik Φj − gij Φk ) + ωij,k − ωik,j

X X
− gij ( ωνk Φν + 12 |Φ|2 Φk ) + gik ( ωνj Φν + 21 |Φ|2 Φj ).
ν ν

After cancellation the compatibility conditions (3.4.10) therefore simplify to

X
(3.4.10)′ Wνijk Φν = ωij,k − ωik,j , i, j, k = 1, . . . , n.
ν

If W = 0, then these conditions no longer involve Φ at all, so they are necessary for
the existence of any solution at all to (3.4.6), and they imply the local existence of a
solution with given value at a point. On the other hand, if there is a solution with
∇ϕ given at a point, then it follows that both sides of (3.4.10)′ must vanish, since the
right-hand side is independent of Φ, hence W = 0 and

(3.4.10)′′ ωij,k − ωik,j = 0, i, j, k = 1, . . . , n.

We shall now show that when n > 3 there is a second miracle: the conditions
(3.4.10)′′ are always fulfilled if the Weyl tensor is equal to 0. (This is not true when
n = 3 which is not surprising since the Weyl tensor is always 0 then.) In fact,
X
(3.4.13) (n − 3)(ωjk,i − ωik,j ) = W l kij,l .
l

(I owe the following calculations to Anders Melin.) First note that the definition of
the Ricci tensor and the second Bianchi identity give
X X X X
Rjk,i − Rik,j = Rl klj,i − Rl kli,j = (Rl klj,i + Rl kil,j ) = Rl kij,l .
l l l l
 CONFORMAL GEOMETRY 65

By (3.4.9) it follows that

(3.4.14) X X
Rjk,i − Rik,j = W l kij,l + g lm (ωmi,l gkj − ωmj,l gki + ωkj,l gmi − ωki,l gmj )
l lm
X X X
l
= W kij,l + ωkj,i − ωki,j + ω l i,l gkj − ω l j,l gki ,
l l l

for ∇g = 0. Since

(3.4.8)′ Rij = (n − 2)ωij + Sgij /(2n − 2),

we have
Rij,l = (n − 2)ωij,l + S,l gij /(2n − 2).
If we multiply by g il and contract in i, l, recalling (3.1.25), we obtain
X X
1
S
2 ,j
= (n − 2) ω l j,l + S,j /(2n − 2), hence ω l j,l = S,j /(2n − 2).

(3.4.13) follows if we use this result and (3.4.8)′ in (3.4.14), for S drops out. Hence we
have proved (3.4.13) and the following
Theorem 3.4.5. A Riemannian manifold of dimension > 3 is conformally flat if
and only if the Weyl tensor vanishes. A Riemannian manifold of dimension 3 is
conformally flat if and only if the integrability conditions (3.4.10)′ are valid.
We shall finally discuss how the Laplace-Beltrami operator defined by (3.4.4) is
changed when one passes to a conformal metric. First we give an equivalent and often
more useful definition:
Proposition 3.4.6. For every u ∈ C 2 we have
1 1
X X
(3.4.15) g ij u,ij = g − 2 ∂j (g 2 g ij ∂i u),
i,j i,j

where g = det(gij ).
Proof. If u, v are in C 2 with support in the same coordinate patch, then
Z X
1
g ij ∂i u∂j vg 2 dx
i,j

is invariantly defined, dx denoting the Lebesgue measure in the local coordinates, for
1
g 2 dx is the invariant volume element dvol(x). Since integration by parts shows that
the integral is equal to
Z X
1 1
− v g − 2 ∂j (g 2 g ij ∂i u) dvol(x)
66 III. ABSTRACT RIEMANNIAN MANIFOLDS

we conclude that the right-hand side of (3.4.15) is invariantly defined, and so is the
left-hand side. Both are equal at the center of a geodesic coordinate system, which
proves the proposition.
Exercise 3.4.2. Prove (3.4.15) by explicit computation using Exercise 2.1.1.
Denote by ∆e the Laplace-Beltrami operator defined using the conformal metric
2ϕ
g̃ = e g. When n = 2 it follows from the second expression for ∆ in (3.4.15) that ∆ e =
e−2ϕ ∆; hence the harmonic functions satisfying the homogeneous Laplace-Beltrami
equation are the same for two conformal metrics. This is not true when n > 2, but we
shall now prove a substitute result in that case. With a constant a to be chosen later
we have
e aϕ u) = e−nϕ g − 21 1
X
∆(e ∂j (e(n−2)ϕ g 2 g jk ∂k (eaϕ u))
X
= e(a−2)ϕ (∆u + (n − 2 + 2a) g jk ∂j ϕ∂k u) + F u,

where F does not depend on u. Now we choose a = 1 − n2 so that the first order terms
disappear. If the formula is applied with u = e−aϕ , it follows that

−F = e2(a−1)ϕ ∆(e−aϕ ) = e(a−2)ϕ (a2 |∇ϕ|2 − a∆ϕ);

the last equality is justified using geodesic coordinates. By (3.4.3) we have

a|∇ϕ|2 − ∆ϕ = 12 ((2 − n)|∇ϕ|2 − 2∆ϕ) = (e2ϕ Se − S)/(2n − 2),

so we have since −a/(2n − 2) = (n − 2)/(4n − 4)

e aϕ u) = e(a−2)ϕ (∆u + (n − 2)/(4n − 4)(e2ϕ Se − S)u), or

∆(e
S(n
e − 2) S(n − 2)
(3.4.16) (∆
e− )(eϕ(2−n)/2 u) = e−ϕ(n+2)/2 (∆ − )u.
4n − 4 4n − 4
The operator ∆ − S(n − 2)/(4n − 4) is called the conformal Laplacian. The transfor-
mation law (3.4.16) makes it easy to pass from solutions for one metric to solutions for
another conformal one. Note that in a flat space, where S = 0, the conformal Laplacian
is the standard Laplacian, so we shall be able to study its solutions by studying the
conformal Laplacian in a conformally equivalent situation. In the Euclidean case one
may for example use the stereographic projection to obtain a compact situation. The
applications are perhaps even more striking in pseudo-Riemannian geometry where a
related conformal map can be used to map Minkowski space into a bounded part of
the Einstein universe, but that will not be discussed in the present notes.
If Se is prescribed, then (3.4.3) is a non-linear elliptic equation for ϕ which can be
solved locally without difficulty. Thus there always exists locally a conformal metric
with zero scalar curvature. The corresponding global questions are very hard, however.
The Yamabe problem to find for a given compact Riemannian manifold a conformal
metric with constant scalar curvature was not completely solved until 1984, when R.
Schoen cleared up the hardest exceptional cases. We hope to return to this in another
chapter.
 THE CURVATURE OF A SUBMANIFOLD 67

3.5. The curvature of a submanifold. Chapter II was restricted to the study of

submanifolds M of RN , but in this chapter we have extended all results given there on
the interior geometry of M to abstract Riemannian manifolds. We shall now extend
the results related to the embedding in RN also by discussing submanifolds of a general
Riemannian manifold.
Let Mf be a Riemannian manifold and let M be a smooth submanifold. We shall
denote the covariant differentiations in M and in Mf by ∇ and ∇ e respectively.

Theorem 3.5.1. If X and Y are vector fields in M then

(3.5.1) e X Y − ∇X Y = h(X, Y ),
∇

where h is a symmetric bilinear form in Tx M with values in the normal plane Nx M =

Tx M
f ⊖ Tx M for every x ∈ M .

Proof. The theorem states in particular that τ ∇

e X Y = ∇X Y if τ is the map T M f|M →
T M defined by orthogonal projection in each fiber. We begin by proving this weaker
result. To do so we note that the identities (3.1.3), (3.1.4) for ∇
e show that if X, Y, Z
are vector fields in M , then we have there

τ∇
eXY − τ∇
e Y X = τ [X, Y ] = [X, Y ],

(τ ∇
e Z X, Y ) + (X, τ ∇
e Z Y ) = (∇
e Z X, Y ) + (X, ∇
e Z Y ) = Z(X, Y ),

hence τ ∇
e X Y = ∇X Y by the uniqueness statement in Theorem 3.1.2.
From (3.1.3) applied to ∇ and ∇
e we obtain

e X Y − ∇X Y = ∇
∇ e Y X + [X, Y ] − ∇Y X − [X, Y ] = ∇
e Y X − ∇Y X,

which proves that the left-hand side of (3.5.1) is symmetric in X and Y . If ϕ ∈ C0∞
then

e X (ϕY ) − ∇X (ϕY ) = ∇
∇ e ϕY X − ∇ϕY X = ϕ(∇
e Y X − ∇Y X) = ϕ(∇
e X Y − ∇X Y ).

Hence the left-hand side of (3.5.1) at x ∈ M is a symmetric bilinear form in X(x) and
Y (x), with values in Nx (M ), which completes the proof.
Definition 3.5.2. The symmetric bilinear map h in Tx M with values in Nx M defined
by (3.5.1) is called the second fundamental form of M with respect to M
f.

In the particular case where M f = RN Theorem 3.1.5 gives back the expression
(2.1.11) for the covariant derivative. We shall now also give an extension of the Gauss
equations (2.1.13), (2.1.15). Let X, Y, Z, W be vector fields in M . By (3.1.6) we have


(∇X ∇Y − ∇Y ∇X − ∇[X,Y ] )Z, W = R(W, Z, X, Y ),
68 III. ABSTRACT RIEMANNIAN MANIFOLDS

where R is the Riemann curvature tensor of M , and we have a similar formula with ∇
and R replaced by ∇e and R.
e (We can extend X, Y, Z, W to vector fields on Mf, but on
M the result is independent of the extension.) By Theorem 3.5.1 we have

(∇X ∇Y Z, W ) = (∇
e X ∇Y Z, W ) = (∇ e Y Z, W ) − (∇
eX∇ e X h(Y, Z), W ).

Since (h(Y, Z), W ) = 0 on M , it follows from (3.1.4) and Theorem 3.5.1 that


e X h(Y, Z), W ) = (h(Y, Z), ∇
−(∇ e X W ) = h(Y, Z), h(X, W ) .

Hence

(∇X ∇Y Z, W ) = (∇ e Y Z, W ) + h(Y, Z), h(X, W ) ,
eX∇

and since (∇[X,Y ] − ∇

e [X,Y ] )Z is a normal vector, we obtain

R(W, Z, X, Y ) = R(W,
e Z, X, Y ) + h(Y, Z), h(X, W ) − h(X, Z), h(Y, W ) ,

which proves the following extension of the Gauss equations (2.1.13), (2.1.15):
Theorem 3.5.3. If M is a smooth submanifold of a Riemannian manifold M f, with
Riemann curvature tensors R and R, respectively, and t1 , t2 , t3 , t4 ∈ Tx M , then
e



(3.5.2) R(t1 , t2 , t3 , t4 ) = R(t
e 1 , t2 , t3 , t4 ) + h(t1 , t3 ), h(t2 , t4 ) − h(t1 , t4 ), h(t2 , t3 ) ,

where h is the second fundamental form of M with respect to M

f.
 CHAPTER IV

EXTERIOR DIFFERENTIAL CALCULUS

IN RIEMANNIAN GEOMETRY

Summary. Chapters II and III have been based on the tensor calculus of Ricci. We
shall now discuss an alternative approach due to E. Cartan using the calculus of exterior
differential forms and moving orthogonal frames systematically. In Section 4.1 we recon-
sider the study in Chapter II of submanifolds of Euclidean space, and in Section 4.2 we
discuss abstract Riemannian manifolds, as in Chapter III. Using the tools developed we
then return to the Gauss-Bonnet theorem in Section 4.3 and prove its analogue in higher
dimensions following S. S. Chern. Pontrjagin classes are then defined in Section 4.4.

4.1. Submanifolds of a Euclidean space. In the proof of the Gauss-Bonnet the-

orem (Theorem 3.1.6) we have seen a convincing example of the usefulness of working
with arbitrary orthonormal frames in the tangent bundle of a Riemannian manifold.
We shall now use such methods also to study submanifolds of a Euclidean vector space
V , of dimension N . Let F0 (V ) be the set of all orthonormal frames e1 , . . . , eN ∈ V . If
we choose a fixed frame (e01 , . . . , e0N ) ∈ F0 (V ), then any other orthonormal frame can
be written uniquely in the form Oe01 , . . . , Oe0N where O ∈ O(N ), the orthogonal group.
(See Section 1.4.) Another choice of e01 , . . . , e0N gives an identification with O(N ) which
only differs by a right translation in O(N ); in particular we see that F0 (V ) is a C ∞
(in fact real analytic) manifold.
In F0 (V ) we have N functions

ej : F0 (V ) ∋ (e1 , . . . , eN ) 7→ ej ∈ V.

The differential dej is a linear form on the tangent space of F0 (V ) with values in V ,
so it can be written
N
X
(4.1.1) dej = ωkj ek ,
k=1

where ωjk = (ek , dej ) are scalar one forms on F0 (V ). Since

0 = d(ej , ek ) = (dej , ek ) + (ej , dek ),

we have

(4.1.2) ωjk + ωkj = 0, j, k = 1, . . . , N.

Since
N
X N
X N
X N
X
0 = d2 ej = (dωkj )ek − ωkj ∧ dek = (dωkj )ek − ωlj ∧ ωkl ek ,
k=1 k=1 k=1 k,l=1
69
70 IV. EXTERIOR DIFFERENTIAL CALCULUS IN RIEMANNIAN GEOMETRY

we obtain
N
X
(4.1.3) dωkj + ωkl ∧ ωlj = 0, j, k = 1, . . . , N.
l=1

The equations (4.1.2), (4.1.3) are of course just another way of writing the results of
Section 1.4 such as (1.4.1).
If W ⊂ V is a subspace of dimension n, the set of orthonormal frames in F0 (V )
such that e1 , . . . , en ∈ W , hence en+1 , . . . , eN ∈ W ⊥ , is a submanifold F0 (V, W ) which
we can identify with the group O(n) × O(N − n). It is clear that the restriction of the
form ωjk above to F0 (V, W ) vanishes if j ≤ n and k > n. When j ≤ n and k ≤ n,
it is equal to the analogous form on F0 (W ) pulled back to F0 (V, W ) by the obvious
surjective map F0 (V, W ) → F0 (W ).
Now let M be a C ∞ submanifold of V , of dimension n, and set

(4.1.4) Fx = F0 (V, Tx M ), x ∈ M,

which is the set of orthonormal frames at x such that e1 , . . . , en span the tangent space
Tx M while en+1 , . . . , eN span the normal space. We shall use the notation α, β, . . . for
indices running from 1 to n, the notation r, s, . . . for indices running from n +1, . . . , N ,
and A, B, . . . for indices running from 1 to N . It is clear that F (M ) = ∪x∈M {x} × Fx ,
as a subset of M × F0 (V ) is a C ∞ fiber space over M . Composing the projection
p : F (M ) → M with the embedding M → V we get a map F (M ) → V , which we also
denote by p. In addition we have N maps eA : F (M ) → V . Since the range of the
differential dp at a point in Fx is equal to Tx M , we can write
n
X
(4.1.5) dp = ωα eα ,
α=1

and we have as in (4.1.1), (4.1.2)

N
X
(4.1.6) deA = ωBA eB , A = 1, . . . , N,
B=1
(4.1.7) ωAB + ωBA = 0, A, B = 1, . . . , N.

Here ωα and ωAB are scalar differential forms on F (M ). From (4.1.5) we obtain
X X X X
0 = d2 p = (dωα )eα − ωα ∧ deα = (dωα )eα − ωα ∧ ωBα eB

or if we separate components with indices ≤ n and > n:

n
X
(4.1.8) dωα + ωβ ∧ ωβα = 0, α = 1, . . . , n,
β=1
Xn
(4.1.9) ωα ∧ ωαr = 0, r = n + 1, . . . , N.
α=1
 SUBMANIFOLDS OF A EUCLIDEAN SPACE 71

Differentiation of (4.1.6) gives

X X X X
0= (dωBA )eB − ωCA ∧ deC = (dωBA )eB + ωBC ∧ ωCA eB ,

hence
N
X
(4.1.10) dωBA + ωBC ∧ ωCA = 0, A, B = 1, . . . , N.
C=1

In particular,
(4.1.11)
n
X
dωαβ + ωαγ ∧ ωγβ = Ωαβ , α, β = 1, . . . , n, where
γ=1
N
X
(4.1.12) Ωαβ = − ωαr ∧ ωrβ , α, β = 1, . . . , n.
r=n+1

The restriction of ωα to a fiber Fx is equal to 0 for α = 1, . . . , n, since p is constant

in Fx , and since p is surjective these forms are at every point in Fx a basis for covectors
orthogonal to Fx . Now the restriction of ωαr to a fiber Fx must vanish, for when x is
fixed we have the situation discussed for F0 (V, W ) above. Hence we can write
n
X
ωαr = λrαβ ωβ
β=1

with uniquely determined coefficients λrαβ ∈ C ∞ (F (M )). Here λrαβ = λrβα according
to (4.1.9). Hence
N
X N
X n
X
Ωαβ = ωαr ∧ ωβr = λrα̺ λrβσ ω̺ ∧ ωσ
r=n+1 r=n+1 ̺,σ=1

is also a form with vanishing restriction to the fibers. The forms ωα , ωαβ and Ωαβ ,
α, β = 1, . . . , n, can be obtained by pulling back forms to F (M ) from the fiber space
P (M ) over M with fiber at x consisting of the orthonormal frames in Tx M , for the
definitions above are already applicable in P (M ) without reference to the normal
vectors.
To clarify the meaning of the preceding equations we shall express them in terms
of our earlier notation. Thus consider a subset of M with a local parametrization
x 7→ f (x), where x varies in an open subset of Rn . We can consider the coordinates
x1 , . . . , xn as functions on M and lift them to F (M ). With the notation fα = ∂f /∂xα ,
which is a vector field on M , we then obtain
n
X n
X n
X
dp = fβ (x)dxβ = ωα eα , thus ωα = (fβ , eα )dxβ .
β=1 α=1 β=1
72 IV. EXTERIOR DIFFERENTIAL CALCULUS IN RIEMANNIAN GEOMETRY

Here dxβ is a differential form on M pulled back to F (M ), and the coefficients are
functions on F (M ). Let s be a local section of F (M ) and set s∗ eα = Eα . Then we
have
n
X n
X
s∗ ωβα = (dEα , Eβ ) = (∂Eα /∂xi , Eβ )dxi = (∇fi Eα , Eβ )dxi .
i=1 i=1

By the product rule (3.1.4) for ∇ and (3.1.6) we obtain

X
s∗ dωβα = ds∗ ωβα = (∇fj ∇fi Eα , Eβ ) + (∇fi Eα , ∇fj Eβ ) dxj ∧ dxi


X X
= 21 ((∇fj ∇fi − ∇fi ∇fj )Eα , Eβ )dxj ∧ dxi + (∇fi Eα , Eγ )(∇fj Eβ , Eγ )dxj ∧ dxi
X X
= 21 R(Eα , Eβ , fi , fj )dxj ∧ dxi + s∗ ωγα ∧ ωβγ .

If we compare this result with (4.1.11), it follows that

n
X
(4.1.13) Ωαβ = 1
2 R(eα , eβ , fi , fj )dxi ∧ dxj .
i,j=1

fi dxi =
P P
If we use that dp = ωγ eγ , and evaluate (4.1.13) on a pair of tangent
vectors, we conclude that

n
X
′ 1
(4.1.13) Ωαβ = 2
R(eα , eβ , eγ , eδ )ωγ ∧ ωδ .
γ,δ=1

Thus the forms Ωαβ contain exactly the same information as the Riemann curvature
tensor. From (4.1.13)′ we also see again that Ωαβ is a linear combination with co-
efficients in C ∞ (P (M )) of forms pulled back from M to P (M ). By the expression
(2.1.13)′ for the curvature tensor it also follows that Ωαβ is independent of the imbed-
ding of M . In Section 4.2 we shall give an independent proof of that by reconsidering
the preceding formulas for an abstract Riemannian manifold.
We shall now also calculate the forms ωαr . With the same notation as above and
(2.1.5), we have
X X
s∗ ωrα = (dEα , Er ) = (∂Eα /∂xj , Er )dxj = (h(fj , Eα ), Er )dxj
X
= (h(Eβ , Eα ), Er )s∗ ωβ ,
β

fj dxj =
P P
since dp = eβ ωβ . Hence we obtain

n
X
(4.1.14) ωrα = (h(eβ , eα ), er )ωβ , α = 1, . . . , n, r = n + 1, . . . , N.
β=1
 SUBMANIFOLDS OF A EUCLIDEAN SPACE 73

Using (4.1.12) we therefore obtain

n
X X
(4.1.15) Ωαβ = ωαr ∧ ωβr = (h(eγ , eα ), h(eδ , eβ ))ωγ ∧ ωδ ,
r=1 γ,δ

which in view of (4.1.13) shows that (4.1.12) is equivalent to the Gauss equations.
We shall finally express the degree of the Gauss mapping for a hypersurface in terms
of differential forms (see (3.1.23)). Thus consider a compact oriented hypersurface
M ⊂ Rn+1 . We can then restrict ourselves to positively oriented frames in F (M ) such
that en+1 has the positive normal direction. The product

(4.1.16) ω1,n+1 ∧ · · · ∧ ωn,n+1

is the pullback to F (M ) of a differential form of degree n on M . It is clear that it

is locally equal to such a form multiplied by a function Pn on F (M ), so we just have to
show that it does not change if we replace eα by β=1 Oβα eβ , α = 1, . . . , n, where
O ∈ SO(n). An easy calculation which we leave as an exercise shows that (4.1.16) is
just multiplied by the determinant of O, which is equal to 1. To compute (4.1.16) we
may therefore choose e1 , . . . , en as directions of principal curvature, which means that

ωα,n+1 = (den+1 , eα ) = −Kα ωα

whereQKα are the principal curvatures. Hence (4.1.16) is equal to (−1)n Kω where
n
K = 1 Kα is the total curvature and ω = ω1 ∧ · · · ∧ ωn is the volume form on M
(pulled back to F (M )). If D is the degree of the reflected Gauss map γ̌, we obtain in
view of (3.1.23)
Z
n n
Dvol(S ) = (−1) ω1,n+1 ∧ · · · ∧ ωn,n+1
M
Z X
n
= (−1) εi1 ...in ωi1 ,n+1 ∧ · · · ∧ ωin ,n+1 /n!
M

where εi1 ...in is +1 for even and −1 for odd permutations of 1, . . . , n and 0 otherwise.
(The integrands here and below should be read as the forms on M which pull back to
them.) If n is even we can pair the factors using the Gauss equations (4.1.12), which
gives
Z
n
X
n
(4.1.17) Dvol(S ) = (−1) 2 εi1 ...in Ωi1 i2 ∧ Ωi3 i4 ∧ · · · ∧ Ωin−1 in /n!.
M

The forms in the right-hand side are well defined for any abstract Riemannian manifold,
and we shall take (4.1.17) as our starting point when generalizing the Gauss-Bonnet
theorem to higher dimensions.
74 IV. EXTERIOR DIFFERENTIAL CALCULUS IN RIEMANNIAN GEOMETRY

4.2. Abstract Riemannian manifolds. In this section M will denote a general

abstract C ∞ Riemannian manifold of dimension n. The fiber bundle P (M ) over M
with fiber Px at x ∈ M consisting of orthonormal frames for Tx M is still well defined.
We denote the projection P (M ) → M by p. In a neighborhood U of any point in
M we can choose a C ∞ orthonormal basis for T U and use it to identify P (U ) with
U × O(n). For i = 1, . . . , n we have C ∞ maps ei : P (M ) → T (M ) mapping an element
in P (M ) to the ith element in the frame, in the tangent space at the base point. We
denote by ωj the one forms on P (M ) which vanish on the fibers such that
n
X
(4.2.1) dp = ωj ej .
1

(Since all indices run from 1 to n now, we shall often omit the range.) We shall now
extend (4.1.8).
Theorem 4.2.1. There are uniquely determined one forms ωij on P (M ), i, j =
1, . . . , n, such that
n
X
(4.2.2) ωij + ωji = 0, dωj + ωk ∧ ωkj = 0, i, j = 1, . . . , n.
k=1

Proof. Uniqueness. Let ω̃ij be forms with

X
ω̃ij + ω̃ji = 0, ωk ∧ ω̃kj = 0.

Since ω1 , . . . , ωn are linearly independent at every point, it follows if we extend to a

basis for T ∗ P (M ) at a point that ω̃kj must be a linear combination of ω1 , . . . , ωn , that
is, X
ω̃kj = λkji ωi , where λkji − λijk = 0.
Since also λkji = −λjki by assumption, we obtain

λijk = −λjik = −λkij,

so a circular permutation of the indices changes the sign. After three circular permu-
tations we conclude that λijk = 0.
Existence. In view of the uniqueness it suffices to prove existence locally, so we may
assume that M is an open subset of Rn , with coordinates denoted by (x1 , . . . , xn ), and
we identify the tangent space of Rn with Rn . Since
X
dp = (dx1 , . . . , dxn ) = ωj ej ,

where p is a function with values in Rn , we obtain

X X X
(dωj )ej − ωj ∧ dej = 0, that is, dωj = ωk ∧ (dek , ej ).
 ABSTRACT RIEMANNIAN MANIFOLDS 75

(The scalar product is of course taken in the Riemannian metric.) The restriction of
(dek , ej ) to a fiber of P (M ) is antisymmetric in j and k, so we can find one forms Bjk
which are also antisymmetric in j, k such that (dek , ej ) + Bjk vanishes on the fibers.
Then we obtain X X
dωi = Bik ∧ ωk + Aijk ωj ∧ ωk

where Aijk is antisymmetric in j and k. Now set

X
ωki = Bik + (Akij + Aijk + Ajik )ωj .
j

Since
Akij + Aijk + Aikj + Akji = 0
it follows that ωik is antisymmetric in its indices, and
X X X
ωi ∧ ωki = ωi ∧ (Bik + Akij ωj ) = dωk ,
j

because Aijk + Ajik is symmetric in i and j. This proves the existence.

Exercise 4.2.1. Show that for every section s of P (M ) we have

s∗ ωkj = −(∇s∗ ek , s∗ ej ) = (∇s∗ ej , s∗ ek ),

and that this determines ωkj uniquely.

Next we shall extend (4.1.11) and (4.1.13)′ :
Theorem 4.2.2. If ωj and ωjk are the forms on P (M ) defined in Theorem 4.2.1,
then the two forms
n
X
(4.2.3) Ωik = dωik + ωij ∧ ωjk , i, k = 1, . . . , n,
j=1

are antisymmetric in the indices i and k, we have

n
X
(4.2.4) Ωik ∧ ωk = 0, i = 1, . . . , n,
k=1

and there are C ∞ functions Rijkl on P (M ) such that

n
X
1
(4.2.5) Ωik = 2 Rikjl ωj ∧ ωl , i, k = 1, . . . , n.
l,j=1

Rijkl is antisymmetric in i, j and in k, l.

76 IV. EXTERIOR DIFFERENTIAL CALCULUS IN RIEMANNIAN GEOMETRY

Proof. Exterior differentiation of the second part of (4.2.2) gives

X X X X X
0= dωjk ∧ ωk − ωjk ∧ dωk = (dωjk + ωji ∧ ωik ) ∧ ωk = Ωjk ∧ ωk .
i

This proves (4.2.4). From the discussion of vector spaces at the beginning of Section
4.1 we know that the restriction of Ωik to any fiber must vanish. (See (4.1.3).) We can
therefore write X X
Ωik = θikj ∧ ωj + 21 Rikjl ωj ∧ ωl .
Here Rikjl are C ∞ functions on P (M ) and θikj are linear combinations with such
coefficients of forms extending ω1 , . . . , ωn to a basis for one forms at a point in P (M ).
The forms θikj are skew symmetric in i and k since Ωik are and θikj is unique. By
(4.2.4) we must have X
θikj ∧ ωk ∧ ωj = 0
k,j

and we can conclude that θikj are skew symmetric in k, j. Recalling the first part of
the proof of Theorem 4.2.1 we can therefore conclude that θikj = 0. Hence we obtain
(4.2.5), where the coefficients are uniquely determined if we require skew symmetry in
j, l. The proof is complete.
If we enter the expression (4.2.5) in (4.2.4) we find that

Ri[kjl] = 0,

where the notation means summation over the circular permutations of k, j, l. In

view of (4.1.13)′ we can identify Rijkl with the Riemann curvature tensor, expressed
in the basis e1 , . . . , en corresponding to a point in P (M ), so this is again the first
Bianchi identity (2.3.4). To derive the second Bianchi identity we take the exterior
differential
P of (4.2.3) and eliminate dωij and dωjk using the same equation. Two sums
ωil ∧ ωlj ∧ ωlk cancel and we obtain
n
X n
X
(4.2.6) dΩik − Ωij ∧ ωjk + ωij ∧ Ωjk = 0.
j=1 j=1

To show that this is equivalent to the second Bianchi identity (3.1.24) we choose a
system of geodesic coordinates x1 , . . . , xn with center at a given point. By Lemma 3.3.4
there are unique orthonormal vector fields E1 , . . . , En near 0 such that Ej = ∂/∂xj at
the origin and the covariant derivative along the radial vector field vanishes. Thus we
obtain a section s of P (M ) which we can use to pull back (4.2.6) from P (M ) to M .
By Lemma 3.3.4 the first order derivatives of Ej vanish at 0, and since the forms s∗ ωi
are biorthogonal,
P ∗ their first derivatives also vanish at 0. Hence it follows from (4.2.2)
that s ωij ∧ s∗ ωj = 0 at 0. Since s∗ ωij is skew symmetric in i, j, it follows (see the
proof of uniqueness in Theorem 4.2.1) that s∗ ωij = 0 at 0, so (4.2.6) gives ds∗ Ωik = 0.
By (4.2.5) we have
X
s∗ Ωik = 21 R(Ei , Ek , Ej , El )s∗ ωj ∧ s∗ ωl ,
 THE GAUSS-BONNET THEOREM IN HIGHER DIMENSIONS 77

where R is the Riemann curvature tensor, so we obtain

X
∂Rikjl /∂xr dxr ∧ dxj ∧ dxl = 0

at the origin, if we recall that the derivatives of s∗ ωj and Ej vanish there. At the center
of a geodesic coordinate system the Christoffel symbols vanish so covariant derivatives
are equal to the corresponding partial derivatives, and we obtain

Rik[jl,r] = 0,

which is the second Bianchi identity (3.1.24).

Summing up, the equations (4.2.1), (4.2.2), (4.2.3) define the one forms ωi , ωij and
the two forms Ωij on P (M ) uniquely. The equations (4.2.4) and (4.2.6) expressing the
first and the second Bianchi identity are valid, and the curvature forms Ωik have the
form (4.2.5) where R denotes the components of the Riemann curvature tensor in the
moving frame in P (M ).
4.3. The Gauss-Bonnet theorem in higher dimensions. In Section 3.1 we
proved the Gauss-Bonnet theorem by showing that the curvature form in the right-
hand side of (3.1.21) lifted to the sphere (circle) bundle is exact. In this section
we shall prove a similar result for the form suggested by (4.1.17). Thus let M be
a compact oriented Riemannian manifold of even dimension n, let P+ (M ) be the
positively oriented part of the frame bundle, and let S(M ) be the unit sphere bundle
{(x, e); x ∈ M, e ∈ Tx M, (e, e) = 1}. We can regard P+ (M ) as a fiber space over
S(M ) with the projection (e1 , . . . , en ) 7→ en . Indices running from 1 to n − 1 will
be denoted by α, β. The restrictions of the differential forms ωαn to the fibers of
P+ (M ) → S(M ) are equal to 0 by the discussion of F0 (V, W ) in Section 4.1, with
V = Tx M and W = Ren . We can therefore regard them as linear combinations with
coefficients depending on e1 , . . . , en−1 of forms lifted from S(M ). The same is true for
Ωjk in view of (4.2.5). If a linear combination of products of these forms is invariant
under SO(n − 1) operating on e1 , . . . , en−1 , it follows that it is the pullback of a form
on S(M ).
Lemma 4.3.1. If the pullback F of a form f on S(M ) to P+ (M ) can be written
Pn−1
F = α,β=1 ωαβ ∧ Fαβ , then it is equal to 0.
Proof. We may assume that α < β in the sum. The restriction of these forms ωαβ
to a fiber of P (M ) → S(M ) are linearly independent (cf. Section 1.4). If π is the
projection P (M ) → S(M ) we can therefore for every tangent vector t of S(M ) find
a unique tangent vector τ of P (M ) such that π ′ τ = t and ωαβ (τ ) = 0, α < β. Since
the multilinear form F vanishes on the linear space of such tangent vectors at any
p ∈ P (M ), it follows that f vanishes on the tangent space of S(M ) at π(q), which
proves the lemma.
The form we wish to integrate is the one in (4.1.17):
X
(4.3.1) ∆0 = εi1 ...in Ωi1 i2 ∧ Ωi3 i4 ∧ · · · ∧ Ωin−1 in .
78 IV. EXTERIOR DIFFERENTIAL CALCULUS IN RIEMANNIAN GEOMETRY

We shall first verify that it is the pullback of a form on M . Since (4.2.5) shows that ∆0
is a linear combination with coefficients in C ∞ (P (M )) of pullbacks of forms on M , it
suffices to show thatP ∆0 is invariant under the map of P (M ) into itselfPreplacing P
a frame
′
e1 , . . . , en by eP
k = O e
i ik i where O ∈ SO(n) is constant. Since ω e
k k = ωk′ e′k
we have ωi = Oik ωk′ then, hence
X X X
dωi = Oik dωk′ = − ′
Oik ωkl ∧ ωl′ = − ′
Oik ωkl ∧ ωj Ojl ,
k k,l j,k,l
X X
′
so ωij = Oik Ojl ωkl , Ωij = Oik Ojl Ω′kl ,
k,l k,l

and we obtain
X
∆0 = εi1 ...in Oi1 j1 . . . Oin jn Ω′j1 j2 ∧ · · · ∧ Ω′jn−1 jn .

The sum over i1 , . . . , in of the coefficients is the determinant of n columns in the

orthogonal matrix O; and since O ∈ SO(n) it follows that it is equal to εj1 ...jn . This
proves the invariance of ∆0 , so it is the pullback to P (M ) of a form of degree n on M ,
which we also denote by ∆0 . Hence it is a closed form.
In the same way we verify for k ≤ n2 − 1 that the n − 1 form
X
(4.3.2) Φk = εα1 ...αn−1 Ωα1 α2 ∧ · · · ∧ Ωα2k−1 α2k ∧ ωα2k+1 n ∧ · · · ∧ ωαn−1 n

is the pullback to P (M ) of a form on S(M P ). It suffices to note that a rotation of the

′ ′
first n − 1 elements in a frame to eα = Oβα
Peβ keeping en = en transforms Ωαβ as
′
in the discussion of ∆0 above while ωαn = Oαβ ωβn . We leave the details of the
repetition of the argument as an exercise and form the differential of Φk . We claim
that
X
(4.3.3) dΦk = k εα1 ...αn−1 Ωα1 α2 ∧ · · · ∧ dΩα2k−1 α2k ∧ ωα2k+1 n ∧ · · · ∧ ωαn−1 n
X
+(n −2k −1) εα1 ...αn−1 Ωα1 α2 ∧· · ·∧Ωα2k−1 α2k ∧dωα2k+1 n ∧ωα2k+2 n ∧· · ·∧ωαn−1 n .

Here we have used that there will be no change of sign in terms where d acts on a factor
Ω. Differential forms of even degree commute, so the differential of such a factor can be
moved to the right of the others without any sign change, and this corresponds to an
even number of inversions of the indices αj . Thus we obtain k identical contributions
when differentiating on the Ω factors. When the factor ωα2k+j n is differentiated we
obtain a sign factor (−1)j−1 which is compensated by j − 1 inversions if it is moved
to the left of the other factors ω, so we get n − 2k − 1 identical contributions of this
kind, which proves (4.3.3).
We shall now calculate the right-hand side of (4.3.3) using that
X X
dΩαβ = Ωαj ∧ ωjβ − ωαj ∧ Ωjβ ,
j j
X
dωαn = Ωαn − ωαj ∧ ωjn .
j
 THE GAUSS-BONNET THEOREM IN HIGHER DIMENSIONS 79

In view of Lemma 4.3.1 we may omit terms containing a factor ωαβ when simplifying
(4.3.3), so we may replace dΩαβ by Ωαn ∧ ωnβ − ωαn ∧ Ωnβ = −Ωαn ∧ ωβn + Ωβn ∧ ωαn
and dωαn by Ωαn . To handle the first sum in (4.3.3) we now introduce for 0 ≤ k < n2
the n forms
(4.3.4) X
Ψk = 2(k + 1) εα1 ...αn−1 Ωα1 α2 ∧ · · · ∧ Ωα2k−1 α2k ∧ Ωα2k+1 n ∧ ωα2k+2 n ∧ · · · ∧ ωαn−1 n .
As for the forms Φk one verifies at once that they are forms on S(M ) pulled back to
P (M ). In view of Lemma 4.3.1 we obtain
(4.3.5) dΦk = −Ψk−1 − ((n − 2k − 1)/(2k + 2))Ψk ,
for the difference between the two sides is the pullback to P (M ) of a form on S(M ),
and every term contains a factor ωαβ . For k = 0 we must interpret Ψ−1 as 0, for then
there are no factors Ω to differentiate.
When k = p − 1 where p = n2 (recall that n is even) we obtain
X
Ψp−1 = n εα1 ...αn−1 Ωα1 α2 ∧ · · · ∧ Ωαn−1 n = ∆0 ,
so we can write ∆0 as a differential by successive elimination of Ψk , k = p−1, p−2, . . . , 0
using (4.3.5). This gives
p−1
X p−1
Y
p−k


(4.3.6) ∆0 = dΦ, Φ= (−1) Φk (2j + 2)/(n − 2j − 1) .
k=0 j=k

Using (4.3.6) instead of (3.1.21) we can now proceed as in the proof of Theorem
3.1.7. Choose f ∈ C ∞ (M ) with only non-degenerate critical points, and let F =
(df )♯ = grad f be the corresponding vector field. We recall that
Γ = {(x, F (x)/kF (x)k); F (x) 6= 0} ⊂ S(M ),
is a manifold with boundary ∂Γ consisting of the fibers Sx over the finitely many
critical points of f , with positive or negative orientation when the sign of the Hessian
of f at x is negative or positive, respectively. The integral of ∆0 over M is equal to the
integral over Γ of the pullback to S(M ), which by Stokes’ formula and (4.3.6) is equal
to the sum of the integrals of Φ over Sx (M ) when f ′ (x) = 0. Since the restriction of
Ωαβ to Sx (M ) is equal to 0, we only get contributions from Φ0 ,
Z Z
p p
(−1) Φ = 2 p!/(2p − 1)!! Φ0 = −(2p p!/(2p − 1)!!)(n − 1)!vol(S n−1 )
Sx (M ) Sx (M )

by the calculation which led to (4.1.17). As in (4.1.17) we shall divide by n!vol(S n ).

Note that if B n is the unit ball in Rn then
Z 1
n
n n−1 n+1 n
vol(S )/vol(S ) = ((n + 1)/n)vol(B )/vol(B ) = (n + 1)/n (1 − t2 ) 2 dt
−1
Z 1
n
= (n + 1)/n (1 − s) 2 s−1/2 ds = (2p + 1)Γ(p + 1)Γ( 21 )/(2pΓ(p + 32 ))
0
= (2p + 1)p!/(2p 21 . . . (p + 21 )) = 2p (p − 1)!/(2p − 1)!!.
80 IV. EXTERIOR DIFFERENTIAL CALCULUS IN RIEMANNIAN GEOMETRY

This gives Z
p n
(−1) /(n!vol(S )) Φ = − 12 .
Sx (M )

Here Sx (M ) has been oriented so that e1 , . . . , en−1 is a positively oriented frame in the
tangent space at en ∈ Sx (M ) if e1 , . . . , en−1 , en , is positively, that is, en , e1 , . . . , en−1
is negatively oriented, so the sphere has been oriented as the boundary of its exterior.
Hence we obtain
Z X
p n
(4.3.7) (−1) /(n!vol(S )) ∆0 = 12 εx ,
M f ′ (x)=0

where εx is the sign of the Hessian of f at x. (We can check the sign by noting that
for M = S n , embedded in Rn+1 as the unit sphere, the right-hand side is equal to 1 if
f is one of the coordinates in Rn+1 , so the sign agrees with (4.1.17).) Using a method
of Chern we have now proved the following extension due to Allendoerfer, Fenchel and
Weil of the Gauss-Bonnet theorem:
Theorem 4.3.2. If M is a compact oriented Riemannian manifold of even dimension
n = 2p, then (4.3.7) is valid with ∆0 denoting the form on M with pullback to P (M )
defined by (4.3.1) and any f ∈ C ∞ (M ) with only non-degenerate critical points, εx
denoting the sign of the Hessian at a critical point x.
As observed after Theorem 3.1.7 we could also replace (df )♯ by any C ∞ vector field F
with only non-degenerate zeros, with the sum taken over the zeros of F and εx defined
as the index of the zero. The right-hand side of (4.3.7) is the Euler characteristic of
M , which we cannot define until Chapter VI though. Without having this concept
available we can still remark that the right-hand side is completely independent of the
Riemannian metric, so the integral in (4.3.7) can only depend on the differentiable
structure of M .
4.4. Pontrjagin forms and classes. The proof that the form ∆0 defined by (4.3.1)
is the pullback of a form on M suggests the definition of other forms on P (M ) with
the invariance properties which guarantees that they are such pullbacks. The most
important ones come from the determinant

(4.4.1) det(δij + λΩij )ni,j=1 ,

where λ is an indeterminate. The determinant is well defined since differential forms

of even order commute. If as in the discussion
P of ∆0 we make a rotation of P (M )
with O ∈ O(n), then Ωij is replaced by k,l Oki Olj Ωkl , which means that (4.4.1) is
multiplied by the square of det O, which is equal to 1. This proves that the coefficient
of any power of λ in (4.4.1) is the pullback to P (M ) of a form in M . (Note that we
have not assumed here that M is oriented.) We can write the coefficient of λk in the
form
 
X I
(4.4.2) sgn Ωi1 j1 ∧ · · · ∧ Ωik jk /k!
J
I,J
 PONTRJAGIN FORMS AND CLASSES 81

where I = (i1 , . . . , ik
) and J = (j1 , . . . , jk ) are sequences of k different indices between
1 and n, and sgn JI is defined as 0 if they are not permutations of each other and
as the sign of the permutation otherwise. The sum (4.4.2) is equal to 0 if k is odd,
I


for if we exchange I and J then sgn J does not change but the following differential
forms are multiplied by −1. It follows from the Bianchi identities (4.2.6) that the forms
(4.4.2) are closed. In fact, since the restrictions of the forms ωjk , j < k, to the fibers
of P (M ) are linearly independent, an obvious analogue of Lemma 4.3.1 shows that we
may calculate the differential as if dΩjk were equal to 0. With a normalization which
will be explained later on, we introduce:
Definition 4.4.1. The Pontrjagin form Pk of a Riemannian manifold M is the form on
M of degree 4k which pulled back to the frame bundle P (M ) is the term of degree 4k
in det(δij + Ωij /2π) where Ωij are the curvature forms (4.2.3). Thus Pk is given by
(4.4.2) with k replaced by 2k, divided by (2π)2k . The de Rham cohomology class of
the closed form Pk is called the Pontrjagin class of M .
The definition of the Pontrjagin form depends heavily on the Riemannian metric,
but we shall now prove that the Pontrjagin classPis independent of it. On any C ∞
manifold one can define a Riemannian metric by ϕj gj where {ϕj } is a partition of
unity subordinate to a covering with coordinate patches and gj is the Euclidean metric
in the local coordinates. Thus the Pontrjagin classes are well defined on arbitrary
compact C ∞ manifolds.
Theorem 4.4.2. If M1 and M2 are Riemannian manifolds with metrics P g1Mand g2 ,
respectively, and if M1 × M2 is given the metric g1 + g2 , then the sum Pk of the
∗
P M1 ∗
P M2
Pontrjagin forms of M is equal to p1 Pk ∧ p2 Pk , where pj is the projection
M1 × M2 → Mj . The Pontrjagin class of a differentiable manifold is independent of
the metric.
Proof. On P (M1 ) ×  P (M2) ⊂ P (M1 × M2 ) it is clear that the matrices ωjk and Ωjk
∗ 0
are block matrices where the diagonal entries are the corresponding matrices
0 ∗
for M1 and M2 , lifted to P (M1 × M2 ). This proves the first statement. Now denote by
M a C ∞ manifold with two different metrics g0 and g1 . Choose an increasing function
χ ∈ C ∞ (R) with χ = 0 in (−∞, 1/3) and χ = 1 in (2/3, ∞). On M × R we introduce
the metric
g = χ(t)g1 + (1 − χ(t))g0 + (dt)2 ,
where t is the coordinate in R. From the first part of the proof it follows that on
M × (−∞, 1/3) (resp. M × (2/3, ∞)) the Pontrjagin forms Pk on M × R are equal to
the Pontrjagin forms Pk0 (resp. Pk1 ) of (M, g0 ) (resp. (M, g1 )) pulled back to M × R.
If
it (x) = (x, t) ∈ M × R, x ∈ M, t ∈ R,
then Pkt = i∗t Pk , t = 0, 1. Since i0 and i1 are homotopic and Pk is closed, it follows
that Pk0 and Pk1 are in the same cohomology class. The proof is complete.
82 IV. EXTERIOR DIFFERENTIAL CALCULUS IN RIEMANNIAN GEOMETRY
 CHAPTER V

CONNECTIONS, CURVATURES AND CHERN CLASSES

Summary. This chapter is not restricted to Riemannian geometry but rather a discus-
sion of analogues for a general vector bundle of the study in Chapters III and IV of the
tangent bundle of a Riemannian manifold. In Section 5.1 we introduce connections as
differential operators on sections of a vector bundle and define the curvature as in Section
3.1. We also give an analogue of the arguments in Chapter IV using exterior differential
forms, for this is essential in Section 5.2 when we define Chern classes, generalizing the
Pontrjagin classes in Section 4.4. A brief introduction to Lie groups is given in Section 5.3
in preparation for a study of principal bundles and associated vector bundles in Section
5.4.

5.1. Connections in vector bundles. The essential point in the expression for
the Riemann curvature tensor in (3.1.6) was the covariant differentiation ∇. Recall
that for a vector field u on M , that is, a section of the tangent bundle T M , ∇u is a
section of T ∗ M ⊗ T M such that for ϕ ∈ C ∞ (M )
(5.1.1) ∇(ϕu) = ϕ∇u + (dϕ) ⊗ u,
and u 7→ ∇u is a linear operator. Here T M can be replaced by any other vector bundle:
Definition 5.1.1. Let M be a C ∞ manifold and E a C ∞ vector bundle over M . Then
a linear map ∇ : C ∞ (M, E) → C ∞ (M, T ∗ M ⊗ E) is called a connection if (5.1.1) is
valid for u ∈ C ∞ (M, E) and ϕ ∈ C ∞ (M ).
In this definition E may be a real or a complex vector bundle provided that ϕ is
in the same category. From (5.1.1) it follows at once that ∇u = 0 in a neighborhood
of x if u = 0 in a neighborhood of x, for we can then choose ϕ ∈ C ∞ (M ) equal to 0
in a neighborhood of x so that u = ϕu. Thus ∇ is a local operator. If e1 , . . . , eN ∈
C ∞ (M, E) form a basis in Ex for all x in an open set ω ⊂ M , then
X X X
∇( ϕj ej ) = ϕj ∇ej + dϕj ⊗ ej .

which proves that ∇ is a first order differential operator.

To put Definition 5.1.1 in the right context we digress to discuss linear differential
operators between sections of vector bundles. If M is a C ∞ manifold and E, F two
vector bundles over M , then a linear map P : C ∞ (M, E) → C ∞ (M, F ) from sections
of E to sections of F is a differential operator if supp P u ⊂ supp u for all u. By a
theorem of Peetre this means that if x1 , . . . , xn are local coordinates in an open subset
′
ω of M where E is identified with ω × RN and F is identified with ω × RN , then P
is given by a matrix of partial differential operators in the usual sense:
N
X
(5.1.2) (P u)j (x) = Pjk uk (x), j = 1, . . . , N ′ .
1
83
84 V. CONNECTIONS, CURVATURES AND CHERN CLASSES

(We could also have taken (5.1.2) as our definition for it is obvious that this property
is invariant under change of local coordinates and bases for E and F .) If all Pjk are
of order m, then P is said to be of order m, and the principal part is defined as the
matrix of principal parts of Pjk , obtained when (∂1 , . . . , ∂n ) is replaced by (ξ1 , . . . , ξn )
in the terms of order m. An invariant formulation is that if 0 6= ξ ∈ Tx∗ and we choose
ϕ ∈ C ∞ with dϕ = ξ at x, then

(5.1.3) p(ξ)u(x) = lim t−m e−tϕ P (etϕ u)(x) ∈ Fx .

t→∞

In fact, using the local representation (5.1.2) we see immediately that this limit exists
and is linear in u(x) and a homogeneous polynomial of degree m in ξ. The invariance
of (5.1.3) shows that p(ξ) ∈ L(Ex , Fx ) is invariantly defined. In the particular case of
a first order operator, p is also linear in ξ, so p can be considered as a vector bundle
map from Tx∗ M ⊗ E to F , and then we have

(5.1.4) P (ϕu) = ϕP u + p(dϕ ⊗ u).

A connection in E is therefore a first order differential operator from sections of E to

sections of T ∗ M ⊗ E whose principal symbol is the identity map in T ∗ M ⊗ E.
Proposition 5.1.2. There exist connections in any vector bundle E. If ∇ is a con-
nection in E, then every connection in E can be written in the form ∇ + A where A
is a section of Hom(E, T ∗ M ⊗ E) ∼
= T ∗M ⊗ E ⊗ E ∗.
Proof. If E is a trivial vector bundle M × CN , say, then we can define

∇(u1 , . . . , uN ) = (du1 , . . . , duN ).

If we take a covering of M with open subsets Uν where E is a trivial bundle, the

trivialization gives a connection ∇ν for PE restricted to Uν . If χν is a subordinate
partition of unity, it is clear that ∇ = χν ∇ν is a connection. If A is the difference
between two connections, it follows from (5.1.1) that

A(ϕu) = ϕAu,

so A is a bundle map, as claimed.

If X is a vector field and u ∈ C ∞ (M, E) we shall just as in the Riemannian case
write ∇X u = hX, ∇ui ∈ C ∞ (M, E). Motivated by (3.1.6) we take another vector field
Y and form

(5.1.5) R∇ (X, Y )u = (∇X ∇Y − ∇Y ∇X − ∇[X,Y ] )u.

R∇ (X, Y )u(x) depends only on the values of X, Y and u at x, for the proof of this
fact for covariant differentiation only used that it is a connection. Thus (5.1.5) is at x
a skew symmetric bilinear form in the tangent vectors X(x), Y (x) and a linear form
in u(x) ∈ Ex , with values in Ex , so R∇ is a two form with values in Ex ⊗ Ex∗ .
 CONNECTIONS IN VECTOR BUNDLES 85

Definition 5.1.3. The curvature R∇ of the connection ∇ in E is the two form with
values in E ⊗ E ∗ defined by (5.1.5).
In the spirit of Chapter IV we can also argue more elegantly by extending the
connection to forms, that is, sections of ∧p T ∗ M ⊗ E. Note that if u is such a section
and ψ is a scalar q form, then ψ ∧ u ∈ ∧q+p T ∗ M ⊗ E can be uniquely defined so that
the multiplication is bilinear and ψ ∧ (ϕ ⊗ v) = (ψ ∧ ϕ) ⊗ v, if v is a section of E and
ϕ is a scalar p form.
Proposition 5.1.4. Given a connection ∇ in the vector bundle E, there is a unique
differential operator d∇ : C ∞ (M, ∧p T ∗ M ⊗ E) → C ∞ (M, Λp+1 T ∗ M ⊗ E), for any p,
such that

(5.1.6) d∇ (ϕ ∧ u) = (−1)p ϕ ∧ ∇u + (dϕ) ∧ u,

if ϕ is a scalar p form and u is a section of E. We have d∇ = ∇ if p = 0, and (5.1.6)

remains valid if u is a section of ∧q T ∗ M ⊗ E and ∇u is replaced by d∇ u.
Proof. Uniqueness is obvious so it suffices to prove the other statements locally, as-
suming that E is a trivial bundle. If also ∇ is trivial, the statement is just standard
calculus of exterior differential forms. If we add to the trivial connection an operator
A of order 0, as in Proposition 5.1.2, then d∇ u just gets an additional term (−1)p Au
if u is a section of ∧p T ∗ M ⊗ E, and these contributions on the two sides of (5.1.6)
cancel.
If u is a section of E, it follows from (5.1.6) for scalar ϕ that

d∇ d∇ (ϕu) = d∇ (ϕ∇u + dϕ ∧ u) = ϕd∇ d∇ u + dϕ ∧ ∇u − dϕ ∧ ∇u = ϕd∇ d∇ u,

which shows that d∇ d∇ is a two form with values in Hom(E, E) ∼

= E ⊗ E ∗ . We claim
that

(5.1.7) d∇ d∇ = R ∇ .

For the proof we note that in terms of local coordinates x1 , . . . , xn in M , with εj =

∂/∂xj , we have
X X
d∇ u = dxi ∇εi u, d∇ d∇ u = − dxi ∧ dxj ∇εj ∇εi ;

X i εi and Y = Y j εj this means that

P P
if X =
X
(d∇ d∇ u)(X, Y ) = − (X i Y j − X j Y i )∇εj ∇εi u
X
= X j Y i (∇εj ∇εi − ∇εi ∇εj )u = R∇ (X, Y )u,

which proves (5.1.7).

86 V. CONNECTIONS, CURVATURES AND CHERN CLASSES

One can also argue as in Sections 4.1 and 4.2. Since we are only interested in forms
on the base manifold we shall study pullbacks of forms analogous to those studied
there. Note that if s is a local section of P (M ) and we write s∗ eα = Eα , which is a
local basis
P for T M , we saw in the proof of (4.1.13) that s∗ ωαβ = (∇Eα , Eβ ), that is,
∇Eα = (s∗ ωαβ )Eβ . We take this as our starting point now.
Let e1 , . . . , eN be a local basis for the vector bundle E, that is, sections of E over
an open set U which form a basis in the fiber Ex for every x ∈ U . Omitting the tensor
product sign ⊗, we can then write
N
X
(5.1.8) ∇ei = ωji ej , i = 1, . . . , N,
j=1

where ωij are one forms. Thus

N
X N
X N
X
∇ ∇
(5.1.9) d d ei = (dωji )ej − ωki ∇ek = Ωji ej , i = 1, . . . , N,
j=1 k=1 j=1
N
X
(5.1.10) Ωji = dωji + ωjk ∧ ωki , i, j = 1, . . . , N.
k=1
P
If u = ui ei is a general local section of E, it follows that
N
X N
X N
X
d∇ d∇ u = ui Ωji ej = Ωij uj )ei ,
i,j=1 i=1 j=1

so (Ωij ) is the matrix of R∇ .

Exercise 5.1.1. Verify by direct computation that if e′i =
P
bji ej is another local basis,
′ −1
then Ω = B ΩB for the corresponding matrices of curvature forms and B = (bij ).
From (5.1.10) we obtain the differential Bianchi identity
N
X N
X
(5.1.11) dΩji = Ωjk ∧ ωki − ωjk ∧ Ωki
k=1 k=1

by repeating the proof of (4.2.6). With matrix notation this can be written more
succinctly
(5.1.11)′ dΩ = [Ω, ω],
where the matrix multiplication is made using the wedge product of course. (See also
the discussion following (5.3.8).)
The characterization of the covariant differentiation in Theorem 3.1.2 involved the
conditions (3.1.3) and (3.1.4). The first is only meaningful for the tangent bundle, but
the second has a general analogue:
Definition 5.1.5. A real (complex) vector bundle is called Euclidean (Hermitian) if in
each fiber there is given a Euclidean (Hermitian) scalar product (·, ·) such that (e, e)
is a C ∞ function on E.
 CONNECTIONS IN VECTOR BUNDLES 87

Proposition 5.1.6. In every real (complex) vector bundle it is possible to introduce a

Euclidean (Hermitian) structure.
Proof. In a coordinate patch U where the bundle is ∼= U ×RN (∼ = U ×CN ), we can just
N N
take the standard Euclidean (Hermitian) form in R (in C ). By piecing together
with a partition of unity as in the proof of Proposition 5.1.2, we obtain a global form.
Proposition 5.1.7. In a real (complex) vector bundle with Euclidean (Hermitian)
structure, it is always possible to find a connection ∇ such that

(5.1.12) d (u, v) = (∇u, v) + (u, ∇v),

when u and v are sections of E. For the corresponding curvature matrices in terms of
orthonormal bases this implies that Ωij = −Ωji , thus Ωij = −Ωji in the real case.
Proof. By the Gram-Schmidt orthogonalization procedure one can in some neighbor-
hood U of any point find an orthonormal basis for E, which means that the identifi-
cation of E over U with U × RN (resp. U × CN ) carries the Euclidean (Hermitian)
form in the bundle to the standard form in RN (in CN ). The trivial connection in
U × RN (in U × CN ) obviously has the property (5.1.12), so we have such a connection
in the restriction of the bundle E to U . Piecing together such local connections by a
partition of unity as in the proof of Proposition 5.1.2, we obtain a global connection
such that (5.1.12) holds.
If e1 , . . . , eN is a local orthonormal basis, it follows from (5.1.8) that

ωji = (∇ei , ej ).

Since (ei , ej ) = δij , we obtain using (5.1.12) that

N
X
ωij + ωji = 0, hence Ωji = −dωij − ωkj ∧ ωik = −Ωij ,
k=1

which completes the proof.

So far we have only discussed connections in a fixed bundle. However, applying a
connection to a section of a bundle one gets a section of another bundle, the tensor
product with the cotangent bundle, so we must also examine how to obtain a connection
there. Thus consider now two C ∞ vector bundles E and F over M , with connections
denoted ∇E and ∇F . In E ⊕ F we get an obvious connection: A section consists of
one section of E and one section of F , and we can apply ∇E and ∇F to them. The
tensor product is not quite so obvious:
Proposition 5.1.8. Given connections ∇E and ∇F in E and in F , there is one and
only one connection ∇ in E ⊗ F such that

(5.1.13) ∇(u ⊗ v) = (∇E u) ⊗ v + u ⊗ (∇F v),

if u and v are sections of E and of F .

88 V. CONNECTIONS, CURVATURES AND CHERN CLASSES

Proof. The uniqueness is obvious so it suffices to prove existence locally. Thus we may
assume that there is a global basis v1 , . . . , vN for sections of F . Every section s of
PN
E ⊗ F can then uniquely be written in the form s = 1 uj ⊗ vj , where uj is a section
of E. If (5.1.13) holds we must have
N
X N
X
E
∇s = (∇ uj ) ⊗ vj + uj ⊗ (∇F vj ),
1 1

PN
and we define ∇ now by this equation. If v = 1 ϕj vj with scalar ϕj ∈ C ∞ , and if u
is a section of E, we have by definition
X X
∇(u ⊗ v) = ∇E (ϕj u) ⊗ vj + ϕj u ⊗ (∇F vj )
X X X
= ∇E u ⊗ ϕj vj + dϕj (u ⊗ vj ) + u ⊗ ϕj (∇F vj ) = (∇E u) ⊗ v + u ⊗ (∇F v),

where we have used first that ∇E is a connection, then that ∇F is a connection. Thus
(5.1.13) is valid which shows that the construction is independent of the choices made.
That ∇ is a connection follows at once from (5.1.13) if we replace u by ϕu and use
that ∇E is a connection.
If E is a vector bundle with connection ∇ and E ∗ is the dual bundle, with fiber Ex∗
equal to the dual space of Ex , then there is a unique connection ∇∗ in E ∗ such that

(5.1.14) dhu, vi = h∇u, vi + hu, ∇∗ vi

for arbitrary sections u and v of E and E ∗ . We leave for the reader the simple verifi-
cation that (5.1.14) defines a connection ∇∗ .
5.2. Chern classes. If T is a linear transformation in a finite dimensional vector
space V , then the determinant of T can be defined as the determinant of the matrix
A representing T with respect to a coordinate system in V . This is independent of
the choice of coordinates, for if x′ = Bx, where B is an invertible matrix, is another
system of coordinates, then the matrix BAB −1 for T in the new coordinates has the
same determinant.
Now let ϕ be any polynomial in the vector space M(N ) of N ×N (complex) matrices
which is invariant in the sense that

(5.2.1) ϕ(A) = ϕ(B −1 AB)

for every B ∈ GL(N ), the group of invertible matrices in M(N ). This is of course
equivalent to

(5.2.1)′ ϕ(AB) = ϕ(BA)

for all A, B ∈ M(N ). An example is ϕ(A) = det(I + λA) for an arbitrary λ, hence also
the coefficient of λk for any k. (Every invariant polynomial is in fact a polynomial in
these special ones, but we shall not give the proof here.)
 CHERN CLASSES 89

Lemma 5.2.1. If ϕ is an invariant polynomial, then the derivative of ϕ(A) in the

direction [A, C] vanishes for any C ∈ M(N ).
Proof. If B = I + εC, then B −1 AB = A + ε[A, C] + O(ε2 ). Since (ϕ(BAB −1 ) −
ϕ(A))/ε = 0, the statement follows when ε → 0.
When we use the lemma below it is convenient to assume that ϕ is homogeneous,
of degree m, and use the symmetric multilinear form Φ(A1 , . . . , Am ) such that ϕ(A) =
Φ(A, . . . , A). The conclusion in the lemma is then that

(5.2.2) mΦ([A, C], A, . . . , A) = 0, A, C ∈ M(N ).

Now let Ω be the matrix of curvature forms for a vector bundle E with fiber dimension
N , defined by means of a local basis. Since all two forms commute, we can form ϕ(Ω)
unambiguously, and since another choice of local basis just replaces Ω by B −1 ΩB for
some invertible matrix B (cf. Exercise 5.1.1), it follows that ϕ(Ω) is independent of
the choice of local basis, so we obtain a globally defined differential form for which we
shall also use the notation ϕ(Ω).
Theorem 5.2.2. If ϕ is an invariant polynomial in M(N ) then ϕ(Ω) is a closed
differential form, and its cohomology class does not depend on the choice of connection.
Proof. Locally we write
ϕ(Ω) = Φ(Ω, . . . , Ω)
where the curvature matrix Ω satisfies the Bianchi identity (5.1.11)′ . Since the forms
Ωij have even degree, we can apply d separately to each argument in the multilinear
form Φ. Two forms commute with arbitrary differential forms so we obtain using
(5.2.2)
dϕ(Ω) = mΦ(dΩ, Ω . . . , Ω) = mΦ([Ω, ω], Ω, . . . , Ω) = 0.
The proof that the cohomology class is independent of the choice of connection is
essentially the same as the proof of Theorem 4.4.2. We keep the notation used there.
If ∇0 and ∇1 are two connections for E over M , and Ee is the vector bundle E pulled
back to M × R by the projection M × R → M , then

e = (1 − χ)∇0 + χ∇1 + ∂/∂t dt

∇

is a connection for E
e on M × R. Note that if x ∈ M then the fiber of E e over (x, t) is
always equal to Ex , so the derivative of a section with respect to t is well defined. It
is now obvious that the matrices of curvature forms for the connections ∇0 and ∇1 on
M are pullbacks from M to M × R by the homotopic maps i0 and i1 of the matrices of
curvature forms for ∇,
e which proves that the cohomology class of ϕ(Ω) is independent
of the choice of connection.
Definition 5.2.3. If E is a complex vector bundle over M and Ω is the (local) matrix
of curvature forms with respect to a connection, then the cohomology class of

(5.2.3) det(I + iΩ/2π),

90 V. CONNECTIONS, CURVATURES AND CHERN CLASSES

where I is the unit matrix, is called the total Chern class of E. It can be written
1 + c1 + c2 + . . . , where cj is of degree 2j and is called the jth Chern class.
If E is the complexification of the tangent bundle of a Riemannian manifold with
the covariant differentiation as connection, it is clear that (−1)j c2j is the Pontrjagin
class of degree 4j and that c2j+1 = 0. More generally, if E is the complexification of
a real vector bundle, it follows from Propositions 5.1.6 and 5.1.7 that we can choose
a connection such that Ωij = −Ωji is a real form. That was all we used to prove
that (4.4.2) vanishes for odd k. For a general complex vector bundle it follows from
Propositions 5.1.6 and 5.1.7 that we can choose a connection such that Ωij = −Ωji ,
which implies that complex conjugation preserves (5.2.3). Hence we obtain:
Proposition 5.2.4. The Chern classes of any complex vector bundle are real. For the
complexification of a real vector bundle the Chern classes cj vanish when j is odd.
This proposition motivates the factor i in (5.2.3). To explain the factor 2π we shall
now discuss the case of a complex line bundle L in a manifold M . We introduce a
Hermitian structure in L and choose an open covering {Uj } such that L is trivial, that
is, has a section sj with norm 1 at every point in Uj . An arbitrary section can then in
Uj be written in the form uj sj . In Uj ∩ Uk we have uj sj = uk sk then, hence writing
gjk sj = sk we obtain |gjk (x)| = 1, x ∈ Uj ∩ Uk , and
uj = gjk uk .
The functions gjk are the transition functions of L. With a connection ∇ in L satisfying
(5.1.12) we now set
∇sj = Γj sj
where iΓj is a real one form in Uj by (5.1.12). The curvature form is then
Ω = dΓj + Γj ∧ Γj = dΓj in Uj ,
so c1 is the cohomology class of the two form equal to idΓj /(2π) in Uj for every j.
Note that
Γk sk = ∇sk = ∇(gjk sj ) = gjk Γj sj + sj dgjk in Uj ∩ Uk ,
which means that
Γk = Γj + dgjk /gjk , dΓj = dΓk in Uj ∩ Uk .
If the cohomology class is equal to 0, then there is a global one form Γ in M such
that d(Γ − Γj ) = 0 in Uj for every j. If Uj has been chosen simply connected, it follows
that we can find a function ϕj in Uj such that Γ − Γj = dϕj in Uj ; we can choose Γ
and ϕj purely imaginary. Since Γk − Γj = dϕj − dϕk in Uj ∩ Uk , this means that
dgjk /gjk = dϕj − dϕk ,
so cjk = gjk eϕk −ϕj is locally constant, that is, constant in each component of Uj ∩ Uk .
If we replace the sections sj by sj eϕj , we get new transition functions which are locally
constant. But in a line bundle with locally constant transition functions we can define
a connection ∇ by ∇sj = 0 in Uj , and then all Γj become 0. Thus we have proved:
 LIE GROUPS 91

Proposition 5.2.5. The Chern class of a line bundle vanishes if and only if it can be
defined by locally constant transition functions.
We shall now consider the example where M is the sphere S 2 of dimension 2. If L
is a complex line bundle on S 2 , we can find a non-zero section s1 in a neighborhood
U1 of the upper hemisphere and another s2 in a neighborhood U2 of the lower one. We
write as before ∇sj = Γj sj and recall that the Chern form is idΓj /2π in Uj ,

Γ1 = Γ2 + dg21 /g21 in U1 ∩ U2 .

Choose a one form Γ in S 2 equal to Γ2 in a neighborhood of the lower hemisphere.

Then the Chern class contains the form c = i(dΓ1 − dΓ)/2π in U1 which vanishes in
the lower hemisphere, and by Stokes’ theorem
Z Z Z Z
c = i/2π (Γ1 − Γ) = i/2π (Γ1 − Γ2 ) = −i/2π dg12 /g12 ,
S2

the last three integrals taken over the equator. This is the winding number of g12 on
the
R equator, thus an integer. In this example the effect of the factor 2π is therefore that
c can precisely take integer values, so that the Chern class is an integer cohomology
class. It would take us too far into a discussion of cohomology theory to give a general
form of this result here; the example should suffice as a motivation.
5.3. Lie groups. In Section 4.1 we worked systematically with forms on the or-
thonormal frame bundle of a Riemannian manifold. Each fiber becomes isomorphic to
the orthogonal group O(n) when a point in it is distinguished. In Section 5.4 we shall
clarify the results by putting them in a more general context. In order to cover the
results in Sections 5.1 and 5.2 we shall also need other groups than O(n). Subgroups of
the full linear groups suffice for our purposes, but the arguments become conceptually
more transparent in an abstract setting where O(n) is replaced by an arbitrary Lie
group, so we shall give a brief introduction to this concept here.
Definition 5.3.1. A group G which has also the structure of a C ∞ manifold is called
a Lie group if
G × G ∋ (a, b) 7→ ab−1 ∈ G
is a C ∞ map.
In particular, the hypothesis implies that b 7→ b−1 and (a, b) 7→ ab are C ∞ maps. It
would suffice to make the latter assumption:
Exercise 5.3.1. Show that if G is a group which is also a C ∞ manifold, then G is a
Lie group if G × G ∋ (a, b) 7→ ab ∈ G is a C ∞ map.
Let La be the left translation La x = ax, a, x ∈ G, which maps the identity e ∈ G
to a, thus L′a : Te → Ta . If X0 ∈ Te , we define a C ∞ vector field X on G by

X(a) = L′a X0 ∈ Ta .
92 V. CONNECTIONS, CURVATURES AND CHERN CLASSES

For arbitrary a, b ∈ G we have L′a X(b) = L′a L′b X0 = L′ab X0 = X(ab) because La Lb =
Lab by the associative law. Thus X is a left invariant vector field. If ω0 ∈ Te∗ we define
similarly a C ∞ one form on G so that ω is equal to ω0 at the identity and ω(X) is a
constant if X is a left invariant vector field. Then we have L∗a ω = ω for every a ∈ G,
and this left invariance together with the value ω0 at e determine ω.
Let X1 , . . . , Xn be left invariant vector fields such that X1 (e), . . . , Xn (e) form a
basis for Te , and let ω 1 , . . . , ω n be the left invariant one forms which are biorthogonal
at e and hence at any a ∈ G. Then we have
n
X
i
(5.3.1) dω = 1
2 cijk ω j ∧ ω k ,
j,k=1

for some constants cijk with cijk = −cikj . In fact, this is obviously true at the identity,
and both sides are left invariant, hence equal everywhere. The constants cijk are called
the structural constants of the Lie group. (They depend in an obvious way on the
choice of basis in Te∗ .) We can also express (5.3.1) in terms of the left invariant vector
fields if we use that by (C.4)

hXj ∧ Xk , dω i i = Xj hXk , ω i i − Xk hXj , ω i i + h[Xk , Xj ], ω ii = h[Xk , Xj ], ω i i.

The left-hand side is cijk , so we obtain

X
(5.3.1)′ [Xk , Xj ] = cijk Xi .

Proposition 5.3.2. In the tangent space g of a Lie group G at the identity, there is
a natural bilinear antisymmetric product g × g ∋ X, Y 7→ [X, Y ] defined by taking the
commutator of the left invariant vector fields extending X and Y . We have the Jacobi
identity

(5.3.2) [X, [Y, Z]] + [Y, [Z, X]] + [Z, [X, Y ]] = 0, X, Y, Z ∈ g.

This means that the structural constants satisfy the conditions

n
(cµνi cνjk + cµνj cνki + cµνk cνij ) = 0,
X
(5.3.2)′ µ = 1, . . . , n.
ν=1

One calls g a Lie algebra.

Proof. Only the Jacobi identity remains to prove, and it follows since it is valid for
commutators of arbitrary vector fields just by the associativity of the products.
In Section 1.4 we discussed one parameter subgroups of the orthogonal group and
of the general linear group. We shall now do the same for an arbitrary Lie group, so
let us assume that ϕ : R → G is a C 1 function such that

(5.3.3) ϕ(s + t) = ϕ(s)ϕ(t), s, t ∈ R.

 LIE GROUPS 93

This implies that ϕ(0) = e, and differentiation with respect to t gives when t = 0, if
we write ϕ(s)ϕ(t) = Lϕ(s) ϕ(t),

ϕ′ (s) = L′ϕ(s) X0 = X(ϕ(s)),

where X0 = ϕ′ (0) and X is the left invariant vector field with X(e) = X0 . This
system of differential equations in the local coordinates of G at e has a unique solution
with ϕ(0) = e, when |s| < 2δ, say. We claim that (5.3.3) is then valid if |s| < δ and
|t| < δ. Fix s and denote the left (right) side of (5.3.3) by ϕ1 (t) (resp. ϕ2 (t)). Then
ϕ1 (0) = ϕ2 (0) = ϕ(s), and
ϕ′1 (t) = X(ϕ1 (t)), ϕ′2 (t) = L′ϕ(s) X(ϕ(t)) = X(ϕ(s)ϕ(t)) = X(ϕ2 (t)),

so the uniqueness theorems for ordinary differential equations show that ϕ1 (t) = ϕ2 (t)
when |t| < δ. Now we can define ϕ(s) uniquely for arbitrary s by setting

ϕ(s) = ϕ(s/N )N
where N is a positive integer so large that |s/N | < δ. This does not depend on N ,
for ϕ(s/N ) = ϕ(s/N M )M for every positive integer M , hence ϕ(s/N )N = ϕ(s/M )M
if |s/M | < δ too. It is clear that (5.3.3) is valid for all s, t.
Proposition 5.3.3. There is a unique C ∞ map exp from the Lie algebra g of G to
G such that the differential at the origin is the identity g → Te (G) = g and R ∋ t 7→
exp(tX) is a one parameter subgroup for every X ∈ g. For X, Y in a neighborhood
of 0 in g we have exp X exp Y = exp ϕ(X, Y ) where ϕ(X, Y ) is a C ∞ function with
values in g defined in a neighborhood of 0 in g × g such that
(5.3.4) ϕ(X, Y ) = X + Y + 12 [X, Y ] + O(|X||Y |(|X| + |Y |)).

Proof. Let X1 , . . . , Xn be P
a basis for the left invariant vector fields. The one parameter
subgroup with derivative aj Xj (e) at 0 is the solution of the Cauchy problem
n
X
dϕ(t)/dt = aj Xj (ϕ(t)), ϕ(0) = e.
1

There is a unique solution for |t| < δ if |a| ≤ 1. It is a C ∞ function of a and t depending
only on ta which we define to be exp(ta). The extension of the map is done as before
by writing exp X = (exp(X/N ))N for large N .
By the implicit function theorem the exponential map is a diffeomorphism of a
neighborhood of the origin in g on a neighborhood of the identity in G. The function
ϕ is therefore well defined as the composition of exp X exp Y with the inverse of the
exponential map, ϕ ∈ C ∞ , and ϕ(sX, tX) = (s + t)X for small sX, tX. Hence it
follows from Taylor’s formula applied first in X, then in Y , that
ZZ
ϕ(X, Y ) − X − Y = ϕ′′XY (sX, tY )(X, Y ) ds dt
0≤s,t≤1
= B(X, Y ) + O(|X||Y |(|X| + |Y |))
94 V. CONNECTIONS, CURVATURES AND CHERN CLASSES

where B(X, Y ) is a bilinear g valued form with B(X, X) = 0, hence skew symmetric.
If X
e is the left invariant vector field equal to X at the identity, pushed back to g from
G by the inverse of the exponential map, then

X(Z)
e = L′Z (0)X, LZ = ϕ(Z, ·), hence
X(Z)
e = X + B(Z, X) + O(|Z|2 ), as Z → 0.

If Ye is a left invariant vector field equal to Y at 0, then we have a similar formula for
Ye , hence
[X,
e Ye ](0) = B(X, Y ) − B(Y, X) = 2B(X, Y ).

The left-hand side is [X, Y ], which completes the proof.

Note that for the full linear group the definition of the exponential map here agrees
with the exponential of matrices used in Section 1.4.
Proposition 5.3.4. If X ∈ g and a ∈ G, then

(5.3.5) a(exp X)a−1 = exp(Ad(a)X),

where Ad(a) ∈ GL(g) is the differential at e of the map G ∋ x 7→ axa−1 ∈ G. In a

neighborhood of the identity Ad(a) is given by

Ad(exp Y ) = exp(ad Y ), ad Y X = [Y, X].

Proof. The map R ∋ t 7→ a(exp(tX))a−1 ∈ G is a one parameter group, hence equal

to t 7→ exp(tZ) where Z is the derivative at 0 of t 7→ a(exp(tX))a−1 ∈ G, hence
equal to Ad(a)X, which is a linear function of X. It is also clear that R ∋ t 7→
Ad(exp(tY )) ∈ GL(g) is a one parameter subgroup, hence equal to exp(t ad(Y )) for
some linear transformation ad(Y ) in g.To determine ad(Y ) we use that by Proposition
5.3.3
exp(tY ) exp(sX) exp(−tY ) = exp(sX + st[Y, X] + O(st(|s| + |t|))),
and this implies that Ad(exp(tY ))X = X + t[Y, X] + O(t2 ), hence that ad(Y )X =
[Y, X]. The proof is complete.
Formula (5.3.4) can be refined to the Campbell-Hausdorff formula

∞
X X
ϕ(X, Y ) = (−1)ν+1 ν −1 (ad X)α1 (ad Y )β1 . . . (ad X)αν (ad Y )βν −1 Y /cαβ ,
ν=1 αi +βi 6=0
Xν ν
Y
cαβ = (αi + βi ) αi !βi !,
i=1 i=1

for X and Y in a neighborhood of 0. (When βν = 0 the last factor should be replaced

by (ad X)αν −1 X.) In particular this shows that the Lie algebra determines the group,
 LIE GROUPS 95

and using (5.3.4) one can also show that there is a local Lie group with any given Lie
algebra. We shall not prove the Campbell-Hausdorff formula here but shall give another
proof that the group operation is uniquely determined near e by the Lie algebra.
Introducing in a neighborhood of the identity in G the canonical coordinates in g
provided by the exponential map has the advantage that the one parameter groups
become rays through the origin. Identifying g with Rn , we denote by X1 , . . . , Xn the
left invariant vector fields which are equal to the unit vectors in Rn at the origin. Then
we know that for a ∈ Rn , s ∈ R and |a||s| small enough the map s 7→ sa is a one
parameter subgroup. P The derivative is a, and the left
P invariant vector field with this
j j
value at the origin is a Xj , so it follows that a = a Xj (sa), that is,
X
(5.3.6) x= xj Xj (x), |x| < δ.
P j
In the left-hand side x denotes of course the radial vector field ̺ = x ∂/∂xj . Taking
the scalar product with the left invariant forms ω k gives h̺, ω k i = xk . We shall derive
equations for the determination of ω i by writing down (5.3.1) explicitly.
differential P
With ω i = ωνi dxν the equation means that
n
X
(5.3.1)′′ ∂ωµi /∂xν − ∂ωνi /∂xµ = cijk ωνj ωµk , i, ν, µ = 1, . . . , n,
j,k=1

or if we multiply by xν and add, recalling that xν ωνi = xi ,

P

X
(̺ + 1)ωµi = δµi + cijk xj ωµk .

This means that

∂ X
tωµi (tθ)) = δµi + cijk θ j tωµk (tθ)
∂t
which is a system of ordinary differential equations with constant coefficients for tωµi (tθ)
from which ωµi can be uniquely obtained. It follows that ωµi is uniquely determined and
analytic in a neighborhood of 0 in the canonical coordinates. We shall now prove that
also the multiplication law is uniquely determined and analytic in a neighborhood of the
identity. (Additional arguments using the conditions (5.3.2)′ show that all equations
(5.3.1)′ are fulfilled and that there is a local Lie group with arbitrarily prescribed Lie
algebra. However, we shall not give the details here but content ourselves with the
local uniqueness of the group.) To do so we note that if ϕ(t) = a exp(tX), where a ∈ G
and X ∈ g, then
ϕ′ (t) = L′ϕ(t) X = X(ϕ(t)),
e ϕ(0) = a,

if Xe is the left invariant vector field equal to X at the identity. This has a unique
solution which is analytic in a and tX, when a is in a neighborhood of e in G and tX
is in a neighborhood of 0 in g. The proof that the structure constants determine the
multiplication uniquely in a neighborhood of the identity is now complete. Note that it
would have been enough to assume that G is a C 3 manifold with C 3 group operations
96 V. CONNECTIONS, CURVATURES AND CHERN CLASSES

to conclude that it is actually an analytic manifold. Hilbert’s fifth problem was to

show that it suffices to assume that G is a topological manifold with continuous group
operations. The proof was completed in the early 1950’s by Gleason, Montgomery and
Zippin. The proof also showed that any topological group having a neighborhood of
the identity containing no subgroup 6= {e} is a Lie group. In particular, every closed
subgroup of a Lie group is a Lie group, which is also easily proved directly (cf. Exercise
1.4.5).
The special case of formula (5.3.1) for the orthogonal group is (4.1.3). Before return-
ing to a discussion of connections we shall put (5.3.1) in another form which avoids the
explicit use of a basis in g. Recall that having chosen a basis X1 , . . . , Xn for g = Te G,
we defined the left invariant form ω j at a ∈ G by
n
X
j j
ω (L′a X) =x , if X = xj Xj ∈ Te G.
1

Xj ω j (L′a X) = X, X ∈ Te G, so the differential form ω = Xj ω j with values

P P
Thus
in g is simply defined by

(5.3.7) ω(L′a X) = X, X ∈ g = Te G.

In Ta G it is therefore the pullback by La−1 of the identity in g = Te G.

We can now write (5.3.1) in the compact form

(5.3.1)′′′ dω = − 12 [ω ∧ ω],

where by (5.3.1) and (5.3.1)′

X X
(5.3.8) [ω ∧ ω] = − cijk Xi ω j ∧ ω k = [Xj , Xk ]ω j ∧ ω k .

To justify the notation we note that in general, if ω1 , ω2 are differential forms with
values in finite dimensional vector spaces W1 and W2 , then we can define ω1 ∧ ω2
uniquely as a form with values in the tensor product W1 ⊗ W2 so that it is equal to
(w1 ⊗ w2 )σ1 ∧ σ2 if ωj = wj σj , j = 1, 2, with wj ∈ Wj and scalar differential forms σj .
In fact, this expression is bilinear in w1 , σ1 and in w2 , σ2 , and the space of forms with
values in Wj is the tensor product of Wj and the space of scalar forms. If we have
a bilinear map β : W1 × W2 → W3 , a third vector space, then composition with the
corresponding map β̃ : W1 ⊗ W2 → W3 gives a form β̃ω1 ∧ ω2 with values in W3 . In
particular, if W1 = W2 = g and β is the Lie bracket we get precisely the definition of
[ω ∧ ω] in (5.3.8). Of course the equation (5.3.1)′′′ has the same contents as (5.3.1) or
(5.3.1)′′ but it is much more compact thanks to the coordinate free notation.
In g there is a natural symmetric bilinear form

(5.3.9) B(X, Y ) = Tr(ad X ad Y ), X, Y ∈ g,

called the Killing form. It is invariant under the adjoint action,

(5.3.10) B(Ad(a)X, Ad(a)Y ) = B(X, Y ), a ∈ G, X, Y ∈ g.

 PRINCIPAL AND ASSOCIATED BUNDLES 97

In fact, it follows from (5.3.4) that

(5.3.11) [Ad(a)X, Ad(a)Z] = Ad(a)[X, Z],
if we expand a exp X exp(−Z) exp X exp Za−1 at X = Z = 0 inserting factors a−1 a
between the exponentials. With A = Ad(a) (5.3.11) means that (ad(AX))A = A ad X,
hence
B(AX, AY ) = Tr (ad(AX) ad(AY )) = Tr (A(ad X ad Y )A−1 ) = B(X, Y ),
which proves (5.3.10). In many cases of interest to us the Killing form gives a natural
Euclidean metric in g, invariant under Ad a, a ∈ G:
Proposition 5.3.5. If G is a compact Lie group, then the Killing form is negative
semi-definite; we have B(X, X) = 0 if and only if ad X = 0.
Proof. The compactness of G guarantees that there is a Euclidean scalar product
(X, Y ), X, Y ∈ g which is invariant under Ad a for every a ∈ G. (If (X, Y ) is
an arbitrary Euclidean scalar product we just have to replace it by the integral of
(Ad aX, Ad aY ) over G with respect to the invariant measure on the group.) With
corresponding orthonormal coordinates in g, the matrix of Ad a is then orthogonal for
every a ∈ G, hence the matrix (Xjk ) of ad X is skew
P symmetric forPevery X ∈ g, as
2
proved in Section 1.4. But then Tr(ad X ad X) = Xjk Xkj = − Xjk ≤ 0, with
equality only if ad X = 0.
Definition 5.3.6. A Lie group is called semi-simple if the Killing form is non-degenerate.
For a semi-simple compact Lie group changing the sign of the Killing form therefore
gives a natural Euclidean metric in the Lie algebra. Let us look at the examples which
will be needed later on; to deal with them we do not really need Proposition 5.3.5 since
everything is done quite explicitly. The first example is the orthogonal group O(N ).
As we saw in Section 1.4, the Lie algebra consists of the skew symmetric matrices so
the Killing form is negative definite (see the proof of Proposition 5.3.5). Let us now
consider instead the unitary group U(N ) of N × N unitary matrices. The Lie algebra
consists of the matrices iH where H is hermitian symmetric. If ad Hh = [H, h] = 0
for every Hermitian symmetric matrix h, then H is a multiple of the identity. This
is clear if H has diagonal form, and we can always reduce to that case by a unitary
transformation. Hence the Killing form vanishes in a one dimensional space, so U(N )
is not semi-simple. However, the subgroup SU(N ) is semi-simple, for its Lie algebra
consists of the matrices iH where H is Hermitian with zero trace. If [H, h] = 0 for all
h satisfying the same condition, we conclude again that H is a multiple of the identity,
and since the trace vanishes it follows that H = 0. At least the groups SU(N ) with
N = 2, 3, 4, 5, 8, 16 have been proposed by physicists in connection with unified field
theories.
5.4. Principal and associated bundles. Let us begin by recalling the definition
of a C ∞ complex vector bundle E with fiber dimension N over a C ∞ manifold M .
First of all, E is supposed to be a C ∞ manifold with
(i) a C ∞ map π : E → M , called the projection;
(ii) a vector space structure in each fiber Ex = π −1 (x).
98 V. CONNECTIONS, CURVATURES AND CHERN CLASSES

These data are required to be locally trivial which means that

(iii) for every point in M there is a neighborhood U and a C ∞ diffeomorphism
ϕ : π −1 (U ) → U × CN such that ϕ restricts to a linear isomorphism Ex →
{x} × CN ∼ = CN for every x ∈ M .
Let {Ui }i∈I be an open covering of M such that for every i ∈ I there is a diffeomorphism
ϕi : π −1 (Ui ) → Ui × CN with the properties listed above. Then we can regard

gij = ϕi ϕ−1
j

as a C ∞ map Ui ∩ Uj → GL(N, C), called a transition function, and we have the

cocycle conditions

gij gji = identity in Ui ∩ Uj

(5.4.1)
gij gjk gki = identity in Ui ∩ Uj ∩ Uk .

The bundle E can be reconstructed using the system of transition functions gij . In
e be the set of all (i, x, w) ∈ I × M × CN with x ∈ Ui and the equivalence
fact, let E
relation

(5.4.2) (i, x, w) ∼ (j, x′ , w′ ) if x = x′ and w′ = gji w.

The reflexivity and symmetry of this relation follow from the first part of (5.4.1), and
the second part of (5.4.1) gives the transitivity. For any given family of transition func-
tions with values in GL(N, C) satisfying (5.4.1), it is clear that the space of equivalence
classes of E,
e with the projection induced by the map E e ∋ (i, x, w) → x ∈ M is a C ∞
complex vector bundle of fiber dimension N . (The linear structure is inherited from
CN of course.) If the transition matrices are obtained from a given vector bundle E
as above, it is also clear that we get back an isomorphic bundle.
For a real vector bundle there is no real change in the argument except that the
transition functions gij will take their values in the smaller group GL(N, R). The
group can be reduced further if we recall that by Proposition 5.1.6 it is always possible
to introduce a Hermitian (Euclidean) structure in a complex (real) vector bundle. The
local trivializations can then be chosen so that they respect the structure, for the local
frames in an arbitrary local trivialization can be orthonormalized using the Gram-
Schmidt procedure. If only such trivializations are used, then the transition functions
take their values in U(N ) (resp. O(N )). If we have a real vector bundle which is
oriented, that is, each fiber is oriented and there exist local trivializations respecting
the orientation, then gij takes its values in SO(N ).
To avoid looking at a large number of cases we now consider an arbitrary Lie group
G. Assume that we have an open covering {Ui }i∈I of the C ∞ manifold M and C ∞
transition functions gij : Ui ∩ Uj → G satisfying (5.4.1). Let ̺ be any representation
of G in a finite dimensional vector space F , that is, ̺(g) is for every g ∈ G a linear
transformation in F such that

(5.4.3) ̺(g1 )(̺(g2 )w) = ̺(g1 g2 )w, w ∈ F, g1 , g2 ∈ G.

 PRINCIPAL AND ASSOCIATED BUNDLES 99

Following the reconstruction of E above we now form the set of all (i, x, w) ∈ I ×M ×F
with x ∈ Ui . Then

(5.4.2)′ (i, x, w) ∼ (j, x′ , w′ ) if x = x′ and w′ = ̺(gji )w,

is an equivalence relation, for if (i, x, w) ∼ (j, x, w′ ) ∼ (k, x, w′′ ), then

w′′ = ̺(gkj )w′ = ̺(gkj )(̺(gji )w) = ̺(gkj gji )w = ̺(gki )w,

so (i, x, w) ∼ (k, x, w′′ ). The set of equivalence classes is a vector bundle F̺ with fiber
dimension dim F and the projection induced by the map (i, x, w) 7→ x ∈ M . The
restriction π −1 (Ui ) is identified with Ui × F by the map

Ui × F ∋ (x, w) 7→ (i, x, w)∼ ∈ F̺ ,

and the corresponding transition functions are ̺(gij ). If the representation ̺ of G on F

is faithful, so that ̺(g) =identity implies g = e, this means that the transition functions
gij can be recovered from F̺ . It is clear that changing coverings and trivializations
gives an isomorphic bundle F̺ , for if we have two such sets then each defines a bundle
which is isomorphic to that defined by the union.
In the preceding discussion we assumed that F is a vector space, but the discussion
of the equivalence relation used only (5.4.3), so we could use any space where G
acts. In particular, taking F = G and ̺ equal to left multiplication, we obtain a
principal bundle P . Note that if (i, x, g) ∼ (j, x, g ′), then (i, x, ga) ∼ (j, x, g ′ a), so right
translation by a ∈ G will be defined on the set P of equivalence classes. Thus the
conditions in the following definition are all fulfilled:
Definition 5.4.1. Let P and M be C ∞ manifolds, let π : P → M be a C ∞ map, and
let G be a Lie group. Then P is called a principal G bundle over M with projection
π if
(i) G acts freely to the right on P , that is, we have a map P ×G ∋ (p, g) 7→ pg ∈ P
such that (pg1 )g2 = p(g1 g2 ) if g1 , g2 ∈ G, pe = p, and pg 6= p for every p ∈ P
if G ∋ g 6= e;
(ii) For every p ∈ P the set pG = {pa; a ∈ G}, which is in one to one correspon-
dence with G, is equal to the fiber π −1 (π(p)) containing p;
(iii) P is locally trivial, that is, for every point in M there is a neighborhood U and
a C ∞ diffeomorphism ϕ : π −1 (U ) → U × G such that

ϕ(pg) = ϕ(p)g, p ∈ π −1 (U ), g ∈ G,

if we define (x, a)g = (x, ag) when x ∈ M and a, g ∈ G.

For the principal bundle constructed above using the transition functions gij we
have obvious trivializations of π −1 (Ui ) which give back the transition functions. Now
we can pass directly from the principal bundle to any one of the bundles F̺ ; they are
100 V. CONNECTIONS, CURVATURES AND CHERN CLASSES

said to be associated to the principal bundle. To do so we note that G acts to the right
(see Definition 5.4.1, (i)) on P × F by

(P × F ) × G ∋ ((p, w), a) 7→ (pa, ̺(a−1)w) = ψa (p, w).

In fact,

ψab (p, w) = (p(ab), ̺((ab)−1w) = ((pa)b, ̺(b−1)(̺(a−1 w)) = ψb (ψa (p, w)),

which is just the definition of action to the right. We define two elements in P × F to
be equivalent if they are carried into each other by some element in G, thus (pa, w) ∼
(p, ̺(a)w), and form the quotient (P × F )/G by this equivalence relation. The map

Ui × G × F ∋ (x, g, w) 7→ ((x, i, g)∼, w) ∈ (P × F )Ui

is a bijection, and each equivalence class under the right action of G contains precisely
one element with g = e; (x, i, g, w) is equivalent to (x, i, e, ̺(g)w) under the G action.
If x ∈ Uj also, then (x, i, e, w) and (x, j, gji, w) define the same element in P × F , and
under the G action the latter is equivalent to (x, j, e, ̺(gji)w). Thus our trivializations
of (P ×F )/G over Uj and Ui are related by the transition function ̺(gji ), which means
that (P × F )/G is isomorphic to F̺ . We have now proved:
Proposition 5.4.2. If P is a principal G bundle over M , and ̺ is a representation
of G on a vector space F , then the quotient of P × F by the right G action defined by

(P × F ) × G ∋ ((p, w), g) 7→ (pg, ̺(g −1)w)

is a vector bundle with fiber dimension dim F , said to be associated with P and ̺,
denoted by P ×̺ F .
Let us now as an example consider the case of a real vector bundle E of fiber
dimension N over M . Let P be the frame bundle over M with fiber Px consisting
of all bases (t1 , . . . , tN ) for Ex . The full linear group GL(N, R) acts to the right
on PP , mapping (t1 , . . . , tN ) ∈ Px and A = (ajk ) ∈ GL(N, R) to (t′1 , . . . , t′N ), where
′
tj = akj tk , and it is clear that P is a principal GL(N, R) bundle. The vector bundle
associated to P and the natural action of GL(N, R) in RN is isomorphic to E. In
fact, the map

N
X
N 1 N
Px × R ∋ (t1 , . . . , tN , x , . . . , x ) 7→ x j tj ∈ E x
1

is constant on the G orbits, for

N
X N
X N
X
xj t′j = xj akj tk = (Ax)k tk .
j=1 j,k=1 k=1
 PRINCIPAL AND ASSOCIATED BUNDLES 101

Hence it defines an isomorphism (P × RN )/G → E. If we instead take ̺(A) = t A−1 ,

then the associated bundle is the dual bundle of E.
After introducing a Euclidean structure in E, we can instead consider the orthonor-
mal frame bundle P which is a principal O(N ) bundle. The bundle associated to the
natural action of O(N ) on RN is again E.
We shall now carry the notion of connection over from a real vector bundle E to the
corresponding frame bundle P . This will lead to a natural definition of a connection
in a principal bundle which induces a connection in the sense already defined on every
associated vector bundle. Let ∇ be a connection in E, choose a basis e1 , . . . , eN for
the sections of E on a coordinate patch U ⊂ M , and write
N
X
∇ei = ωji ej , i = 1, . . . , N,
j=1

PN
as in (5.1.8). For a general section u = 1 ui ei we have

N
X N
X N
X N
X
∇u = dui ei + ui ωji ej = (dui + uj ωij )ei .
i=1 i,j=1 i=1 j=1

The equation ∇u = 0 can therefore be written

N
X
(5.4.4) dui + ωij uj = 0.
j=1

We say that u is parallel along a curve t 7→ γ(t) if (∇γ(t) u)(γ(t)) ≡ 0. This is a

linear system of N differential equations for N unknowns, so the unique solvability of
the Cauchy problem means that there is a unique parallel section of E along γ with
prescribed value at a point on γ.
PN
We can apply this to make a parallel transport of a frame Ej = i=1 aij ei , j =
1, . . . , N , along γ. By (5.4.4) the equations ∇Ei = 0 can be written
N
X
(5.4.5) daij + ωik akj = 0, i, j = 1, . . . , N.
k=1
P
If λj Ej = 0 at some point on γ then this is true everywhere, so if we prescribe
linearly independent initial data we get a frame along γ.
Now we can regard the local coordinates x1 , . . . , xn in M together with the coordi-
nates aij with det(aij ) 6= 0 as coordinates in the frame bundle P over U . The equations
(5.4.5) have an invariant meaning independent of the choice of these coordinates. If
they are satisfied for the restriction to a regular curve Γ ⊂ P , then Γ cannot be a
tangent to a fiber, for if dx = 0 it follows from (5.4.5) that daij = 0, contradicting that
the curve is regular. Thus the equations (5.4.5) define at every point p ∈ P a plane in
Tp P of dimension n which is mapped bijectively on Tπ(p) M by the differential of the
102 V. CONNECTIONS, CURVATURES AND CHERN CLASSES

base projection π : P → M . The P right action

P P of an element g = (gjk ) ∈ GL(N, R)
maps the frame E1 , . . . , EN to gkj Ek = i ( k aik gkj )ei , j = 1, . . . , N , which just
means that the coordinates A = (aik ) are replaced by Ag. The equations (5.4.5) are
obviously invariant under this right multiplication, so the “horizontal” planes defined
by (5.4.5) have the properties in the following definition:
Definition 5.4.3. If P is a principal G bundle over the n dimensional manifold M ,
with projection π : P → M , then a connection on P is a differentiable assignment
P ∋ p 7→ Hp ⊂ Tp P , where Hp is an n dimensional subspace such that
(i) π ′ : Hp → Tπ(p) M is bijective;
(ii) Hpa = Ra′ Hp , a ∈ G, where Ra denotes the right action on P of a ∈ G.
One calls Hp the horizontal space at p.
In the example of the frame bundle of a real vector bundle E with fiber dimension
N , we saw above that the horizontal space is defined by the vanishing of N 2 differential
forms (5.4.5), and N 2 is the dimension of the group GL(N, R). We shall now show
that in the general case the horizontal space is defined by means of a natural differential
form on P with values in the Lie algebra g of G. This is a consequence of the following
two facts.
(a) If p ∈ P then the tangent space Tp P is by (i) in Definition 5.4.3 the direct sum
of Hp and the kernel of π ′ : Hp → Tπ(p) M , that is, the tangent space Tp0 P at
p of the fiber π −1 (π(p)) through p. Thus we have a well defined projection
(5.4.6) vp : Tp → Tp0 P, vp Hp = 0, (Id − vp )Tp0 = 0.
(b) The right action G ∋ a 7→ pa is by Definition 5.4.1 for every p ∈ P a diffeomor-
phism of G on the fiber of P through p. Hence the differential at the identity
is a linear bijection
(5.4.7) γ(p) : g → Tp0 P.

The composition ωp = γ(p)−1 vp of the projection (5.4.6) with the inverse γ(p)−1 is
thus a linear map from Tp P to g, that is, a one form on P with values in g, which
vanishes precisely in the horizontal spaces Hp . To define it we have used part (i) of
Definition 5.4.3, but we must also express condition (ii) there in terms of ω. To do so
we note that the condition means that
vpa = Ra′ vp (Ra′ )−1 , hence ωpa = γ(pa)−1 Ra′ vp (Ra′ )−1 = γ(pa)−1 Ra′ γ(p)ωp (Ra′ )−1 .
If X, Y ∈ g and
Ra′ γ(p)X = γ(pa)Y
then t 7→ p exp(tX)a and t 7→ pa exp(tY ) have the same derivative for t = 0, hence
t 7→ p exp(tX) and t 7→ pa exp(tY )a−1 = p exp(Ad a (tY )) have the same derivatives for
t = 0. By the injectivity of γ(p) we conclude that X = Ad a Y , that is, Y = Ad(a−1 )X,
so
ωpa Ra′ = Ad(a−1 )ωp .
The left-hand side is the pullback Ra∗ ω in Tp . We have now proved the first half of the
following:
 PRINCIPAL AND ASSOCIATED BUNDLES 103

Theorem 5.4.4. For a connection in a principal G bundle P , the projection on the

tangent space of the fiber along the horizontal space followed by the inverse of (5.4.7)
is a one form on P with values in the Lie algebra g of G such that

(5.4.8) Ra∗ ω = Ad(a−1 )ω, a ∈ G,

(5.4.9) ωp (γ(p)X) = X, X ∈ g,

where Ra is the right action of a on P . Conversely, if ω is a one form on P with

values in g satisfying (5.4.8) and (5.4.9), then Hp = {t ∈ Tp P ; ht, ωi = 0} defines a
connection in P in the sense of Definition 5.4.3. One calls ω the connection form.
Proof. If ω satisfies (5.4.9) then Hp is transversal to the fiber and of codimension
dim g in Tp P , which proves (i) in Definition 5.4.3. Condition (ii) follows from (5.4.8)
by reversing the proof of (5.4.8).
The next goal is to show that a connection in a principal G bundle P gives rise to
a unique connection in every associated vector bundle. This must be done so that in
the motivating example of a real vector bundle above we get back the connection we
started from. Let s be a section over an open set U ⊂ M of the vector bundle F̺
associated to P , F and the representation ̺ of G on F , and let t ∈ Tx M , x ∈ U . We
may assume that there is a section p of P in U , and this defines a map w : U → F such
that s is the image (p, w)∼ of the section (p, w) of P × F . (Recall that P̺ is the set of
equivalence classes of P × F under the right G action (p, w) 7→ (pa, ̺(a−1 )w), a ∈ G.)
If p′ (x)t ∈ Tp(x) were horizontal, we would in the special case of the frame bundle of
a vector bundle E have Pp = (e1 , . . . , eN ) with ∇t ej (x) = 0, and w = (w1 , . . . , wN ), so
we would have ∇t s = ht, dwj iej . This leads us to require that

(5.4.10) ∇t s(x) = (p, ht, dwi)∼, if p′ (x)t ∈ Hp(x) .

(ht, dwi is well defined since w is a function with values in a vector space.) If p′ (x)t
is not horizontal, we replace p and w by p̃ = p exp(−ϕ) and w̃ = ̺(exp ϕ)w where
ϕ : U → g, ϕ(x) = 0. If ht, ϕ′ (x)i = X ∈ g, then

p̃′ (x)t = p′ (x)t − γ(p)X, ht, dw̃i = ht, dwi + ̺′ (X)w,

where ̺′ = ̺′e is a linear map from g to linear maps in F . Thus p̃′ (x)t ∈ Hp(x) if and
only if vp p′ (x)t = γ(p)X, that is, X = hp′ (x)t, ωp(x) i = ht, p∗ ωi, so (5.4.10) leads to

∼
(5.4.11) ∇t s(x) = p(x), ht, dw(x)i + ̺′ (ht, p∗ ωi)w(x) .

The preceding arguments show that (5.4.11) is the only definition satisfying (5.4.10),
but it remains to show that it does not depend on p(x). To do so we must prove that
(5.4.11) does not change if p is replaced by pa = Ra p and w is replaced by ̺(a−1 )w
with a constant a ∈ G. If we do that then ht, p∗ ωi is replaced by

ht, p∗ Ra∗ ωi = ht, p∗ Ad(a−1 )ωi,

104 V. CONNECTIONS, CURVATURES AND CHERN CLASSES

by (5.4.8), and since ̺(a−1 ba) = ̺(a−1 )̺(b)̺(a) we have

̺′ (Ad(a−1 )X) = ̺(a−1 )̺′ (X)̺(a), X ∈ g.

Thus the second component in (5.4.11) is replaced by

ht, d̺(a−1)wi + ̺(a−1 )̺′ (ht, p∗ ωi)̺(a)̺(a−1)w

which is precisely ̺(a−1 ) times the value in (5.4.11). The first component is replaced
by p(x)a, so the equivalence class under the G action is unchanged. Hence (5.4.11)
gives a unique definition of ∇t s(x), t ∈ Tx M , when s is a section of F̺ . It is obvious
that ∇t s(x) is linear in t and linear in s, and we have for ϕ ∈ C ∞

∇t (ψs) = ψ∇t s + ht, dψis,

so ∇ is a connection in F̺ .
To motivate the definition of the curvature form of a principal bundle with connec-
tion we shall write the preceding formulas explicitly in the case of the frame bundle P
of a vector bundle, which is a principal G = GL(N, R) bundle. As above we choose a
basis e1 , . . . , eN for the sections of E in a coordinate patch U ⊂ M and define forms
ωij by (5.1.8). As coordinates in P |U we use the local coordinates x1 , . . . , xn in U and
the components A = (aij ) ∈ G of a general frame
X
Ej = aij ei , j = 1, . . . , N.

The Lie algebra g is identified with the space M(N ) of N × N matrices, which is
mapped bijectively to the tangent to the fiber of P at E1 , . . . , EN by

d
M(N ) ∋ X 7→ (AetX )|t=0 = AX.
dt
The horizontal space is defined by the equations (5.4.5), hence

N
X
(Aω)ij = daij + ωik akj
k=1

if ω denotes the connection form in Definition 5.1.3, with values in M(N ), for both
sides vanish in the horizontal space and are equal in the fiber direction. If ω e denotes
the matrix (ωij ), this means that the connection form ω satisfies

Aω = dA + ω e A,

which should be compared to the solution of Exercise 5.1.1. Hence

dA ∧ ω + Adω = (dω e )A − ω e ∧ dA,

 PRINCIPAL AND ASSOCIATED BUNDLES 105

or after insertion of dA = Aω − ω e A

(5.4.12) A(dω + ω ∧ ω) = (dω e + ω e ∧ ω e )A.

The matrix of two forms dω e + ω e ∧ ω e in the right-hand side of (5.4.12) is the matrix
Ω in (5.1.10).
It is now clear that we should define the curvature form Ω for a principal G bundle
with g valued connection form ω by

(5.4.13) Ω = dω + 12 [ω ∧ ω].

Here the right-hand side is defined as explained in Section 5.3, using the bracket in
g. When G = GL(N, R), g = M(N ), this is equal to ω ∧ ω. We shall now prove a
general analogue of (5.4.12) where transformation by the matrix A is replaced by the
adjoint action.
Theorem 5.4.5. Let P be a principal G bundle with connection form ω. The curva-
ture form Ω defined by (5.4.13) is then for every p ∈ P an antisymmetric bilinear form
in Tp P/Tp0 P with values in g, and

(5.4.14) Ra∗ Ω = Ad(a−1 )Ω, a ∈ G,

(5.4.15) Ω(t1 , t2 ) = dω(t1 , t2 ), if t1 , t2 ∈ Hp .

Proof. (5.4.15) is obvious since

[ω ∧ ω](t1 , t2 ) = [ω(t1 ), ω(t2 )] = 0 when t1 ∈ Hp .

For a ∈ G we have in view of (5.4.8) and (5.3.11)

Ra∗ Ω = dRa∗ ω + 21 [Ra∗ ω, Ra∗ ω] = Ad(a−1 )Ω

by (5.4.8). What remains is to prove that Ω(t1 , t2 ) = 0 if t1 ∈ Tp P and t2 ∈ Tp0 P . If

also t1 ∈ Tp0 P then this follows from (5.3.1)′′′ , for if we identify a fiber of P with G,
then it follows from (5.4.9) that the restriction of the connection form ω to the fiber
is the form ω in Section 5.3. Hence it suffices to prove that

Ω(t1 , t2 ) = 0, if t ∈ Hp P, t2 ∈ Tp0 P.

Since ω(t1 ) = 0 we have [ω ∧ ω](t1 , t2 ) = 0, so the claim is that dω(t1 , t2 ) = 0. To prove

this we choose a horizontal vector field h with h(p) = t1 , write t2 = γ(p)X, and denote
by Xe the vector field equal to γ(q)X at q ∈ P , corresponding to the infinitesimal right
action of G in the direction X. By (C.5) we must show that

−h[h, X],
e ωi + hhX,
e ωi − Xhh,
e ωi = 0.
106 V. CONNECTIONS, CURVATURES AND CHERN CLASSES

Now hh, ωi = 0 since h is horizontal, and hX,

e ωi = X is constant by (5.4.9), so the
last two terms vanish. The proof will be complete if we prove that [h, X]
e is horizontal.
Now

e = lim (Rexp(εX) )∗ h − h /ε
[h, X]
ε→0

as is immediately seen in a coordinate system where the vector field X̃ is constant.

Here
(Ra )∗ h(p) = Ra′ h(pa−1 )
is horizontal for every a ∈ G, by condition (ii) in Definition 5.4.3, which proves that
[h, X]
e is also horizontal.

Theorem 5.4.5 means that at every p ∈ P we can regard the curvature form Ω as the
pullback of a two form with values in g at the base point π(p), and this form transforms
by Ad(a−1 ) under right translation by a. Thus we have obtained an extension of
(5.4.12) with considerable precision added when we can reduce the group G to a
subgroup.
Exercise 5.4.1. Show using (5.4.14) that if ̺ : G → F is a representation and ̺′ = ̺′e ,
then ̺′ (p∗ Ω) is a two form which at x ∈ M only depends on p(x), and that

̺′ ((pa)∗ Ω) = ̺(a−1 )̺′ (p∗ Ω)̺(a),

which means that ̺′ (p∗ Ω) defines a two form on M with values in End(P ×̺ F ).
For a Riemannian manifold of dimension n the tangent bundle is associated to the
principal O(n) bundle of orthonormal frames. So are all tensor bundles and more
generally the subbundles associated to O(n) invariant subspaces of tensor products
of a number of factors Rn . All the covariant differentiations discussed in Section
3.1.5, including that in Exercise 3.1.5, are therefore induced by the same Riemannian
connection in the orthonormal frame bundle, and its curvature form is essentially
equivalent to the Riemann curvature tensor by (4.2.5).
 CHAPTER VI

LINEAR DIFFERENTIAL OPERATORS

IN RIEMANNIAN GEOMETRY

Summary. In Section 6.1 we discuss second order elliptic operators on sections of a

vector bundle such that the principal part is a multiple of the identity. Such operators
occur frequently in geometry, and their analysis is particularly elementary thanks to a
classical construction of a parametrix due to Hadamard. In the later sections we shall
apply the conclusions to the de Rham complex and to Dirac operators. Section 6.2 is a
digression where we discuss from a purely algebraic point of view topics such as the ∗
operator on differential forms. Section 6.3 is then devoted to Hodge theory, culminating
in a discussion of the Hirzebruch signature operator, Weitzenböck decomposition and
Bochner’s vanishing theorem. The corresponding heat equations are studied in Section
6.4. Gilkey’s theorem on invariant forms with non-negative weights is proved in Section
6.5; the required background in invariant theory is developed in Appendix D. This the-
orem implies that there is a local index formula for the Hirzebruch signature operator.
Rather than determining coefficients we then pass to a discussion of Dirac operators,
started in Section 6.6 and leading in Section 6.10 to Getzler’s completely constructive
proof of the local index theorem for such operators. This requires a great deal of back-
ground material concerning Clifford algebras and spinors given in Sections 6.7 and 6.8.
We also need the classical Mehler formula for Hermite polynomials and some extensions
presented in Section 6.9.

6.1. Metric elliptic operators. Already in (3.4.4) and Proposition 3.4.6 we de-
fined the Laplace operator acting on a scalar function P
on a Riemannian manifold. It
is a second order operator with principal symbol ξ 7→ g jk ξj ξk . (See (5.1.3) for the
definition of the principal symbol.) However, in geometrical questions the most com-
mon objects are sections of vector bundles, such as tensor bundles, rather than scalar
functions. We shall therefore consider a more general class of operators here.
Definition 6.1.1. If E is a C ∞ vector bundle on a C ∞ manifold M , then a second order
differential operator P : C ∞ (M, E) → C ∞ (M, E) is said to be metric if the principal
symbol p(x, ξ) : Ex → Ex is a positive multiple p0 (x, ξ)IEx of the identity in Ex for
every ξ ∈ Tx M .
The positive definite quadratic form p0 defines a symmetric tensor of type 2, 0,
and its dual tensor, of type 0, 2, is a Riemannian metric. With local coordinates
x1 , . . . , xn P
in M and corresponding coordinates x1 , . . . , xn , ξ1 , .P
. . , ξn in T ∗ M , we have
p0 (x, ξ) = g jk (x)ξj ξk , and the Riemannian metric is ds2 = gjk (x)dxj dxk , where
jk −1
(gjk ) = (g ) . Thus a metric operator is associated with a unique Riemannian
metric, hence the term “metric operator” (which is used by some authors but not
universally).

Typeset by AMS-TEX
107
108 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

An example of a metric operator is obtained if we choose a connection ∇E in E and

define
n
X
(6.1.1) Pu = g ij (∇
e E ∇E u)ij ,
i,j=1

where ∇e E is the connection in E ⊗T ∗ M defined by ∇E and the Levi-Civita connection

∇ (covariant differentiation) in T ∗ M according to Proposition 5.1.8. If ϕ ∈ C ∞ (M ) is
a scalar function, then with (·)s denoting symmetrization in T ∗ M × T ∗ M

e E ∇E (ϕu) = ∇
∇ e E (ϕ∇E u + (dϕ)u) = ϕ∇
e E ∇E u + 2((dϕ)∇E u)s + (∇dϕ)u,

hence

(6.1.2) P (ϕu) = ϕP u + 2h(dϕ)♯ , ∇E ui + (∆ϕ)u.

Now suppose that P is any metric operator in E, introduce the associated Riemannian
metric, and define for t ∈ Tx M

(6.1.3) ht, ∇E ui = 12 (P (ϕu) − ϕP u − (∆ϕ)u),

if ϕ ∈ C ∞ (M ) and ϕ′ (x) = t♭ . We claim that the right-hand side only depends on t

and that (6.1.3) defines a connection. With local coordinates x1 , . . . , xn in M and a
local basis for E, we can write

n
X n
X
Pu = g ij ∂i ∂j u + cj ∂j u + cu,
i,j=1 j=1

where u = (u1 , . . . , uN ) and c1 , . . . , cn , c are N × N matrices. Then the right-hand side

of (6.1.3) can be written

n n n
− 21 1
X X X
ij i
g ∂i ϕ∂j u + 1
2 (c − g ∂j (g 2 g ij ))∂i ϕ u
i,j=1 i=1 j=1

which is a linear function of ϕ′ , hence of t = ϕ′♯ , with principal part equal to the
identity. If we write
X
P = g ij (∇
e E ∇E u)ij + Q,

it follows from (6.1.2) and the way we have defined ∇E that Q(ϕu) = ϕQu, ϕ ∈ C ∞ ,
so Q is a differential operator of order 0. We have proved:
 METRIC ELLIPTIC OPERATORS 109

Proposition 6.1.2. If P is a metric differential operator in the vector bundle E on

M , then there is a Riemannian metric in M , a connection ∇E in E, and a section c
of Hom E = E ⊗ E ∗ , such that
X
(6.1.4) Pu = g ij (∇
e E ∇E u)ij + cu, u ∈ C ∞ (M, E),

e E is the connection in E ⊗T ∗ M defined by ∇E and the Levi-Civita connection.

where ∇
The Riemannian metric, ∇E and c are uniquely determined by P .
Note that no first order terms are visible in (6.1.4); they have been absorbed in
the connection ∇E . We shall see many explicit formulas of the type (6.1.4) where c is
related to the Riemannian curvature tensor.
Using a classical approach due to Hadamard we shall now construct a parametrix
for any metric operator, that is, a distribution section Fa of E such that P F − aδx
for given x ∈ M and a ∈ Ex is in C ∞ or at least highly differentiable. Here the
Dirac function is defined in terms of coordinates such that g = det(gjk ) = 1 at x. For
√
general local coordinate x1 , . . . , xn it must be defined as δ(x1 , . . . , xn )/ g; division by
the Riemannian density changes the distribution density δ(x1 , . . . , xn ) to a distribution
independent of the coordinates. In our construction we shall use geodesic coordinates
centered at x, so gjk (0) = δjk which makes g(0) = 1. To simplify the construction one
should also use a frame in E which is adapated to the geodesic coordinates.
With local coordinates x1 , . . . , xn in M and a local frame for E we write u =
(u1 , . . . , uN ) and

n n
− 12 1
X X
jk
(6.1.5) Pu = g ∂j (g g ∂k u) +
2 cj ∂j u + cu,
j,k=1 j=1

where u is differentiated componentwise as a function in Rn with values in RN . Then

(6.1.3) means that

n
X
E
ht, ∇ ui = 1
2
(P (ϕu) − ϕP u − (∆ϕ)u) = (dϕ, du) + 1
2
cj (∂j ϕ)u,
j=1

hence
n
X
E
∇ u = du + 1
2 cj gjk dxk u.
j,k=1
Pn
If ̺ = 1 xj ∂j is the radial vector field, we obtain

n
X n
X
(6.1.6) ∇E
̺ u = ̺u +
1
2 ck xk u, ck = cj gjk .
k=1 j=1
110 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

Lemma 6.1.3. In a neighborhood of the origin in the local coordinates one can choose
the local frame e1 , . . . , eN for E so that ∇E
̺ eν = 0, ν = 1, . . . , N , where ̺ is the
radial vector field. Such a frame is called synchronous; it is uniquely determined by
e1 (0), . . . , eN (0).
Proof. By (6.1.6) we want all eν to be solutions of the system of differential equations

n
X n
X
xj ∂ j u + 1
2 ck xk u = 0.
j=1 k=1

We shall prove that there is a unique solution with given value at the origin. At the
origin the equations imply that ∂j u(0) = 21 cj (0)u(0), so we set

n
X
v= exp( 21 cj (0)xj )u.
j=1

Then

X n
X n
X n
X
j j j
exp( 21 1
cj (0)xj u + O(|x|2 )u ,


x ∂j v = cj (0)x ) x ∂j u + 2
j=1 j=1 j=1

so the equation becomes

n
X
xj ∂j v + Rv = 0,
j=1

where R vanishes of second order at the origin. But this equation is of the form
discussed in the proof of Lemma 3.3.4, which completes the proof.
Remark. We have assumed aboved that the vector bundle is real, but there is no
difference in statements or proofs if it is complex.
In the following construction we shall use geodesic coordinates and a frame for E
satisfying the conclusions in Lemma 6.1.3, that is,

n
X n
X n
X
j k j k
(6.1.7) gjk (x)c (x)x = gjk (0)c (x)x = ck (x)xk = 0.
j,k=1 j,k=1 k=1

The operator is then so well approximated by the constant coefficient Laplacian ∆

that a fundamental solution of ∆ is a good first approximation for a parametrix of P .
To get higher order approximations we shall also have to use fundamental solutions of
powers of ∆. To avoid some minor technical difficulties and at the same time prepare
for a study of heat equations later on, we shall consider ∆ + z and P + z instead of ∆
and P for suitable z ∈ C.
 METRIC ELLIPTIC OPERATORS 111

Lemma 6.1.4. If z ∈ C\R+ then there is for every integer ν ≥ 0 a unique distribution
Fν ∈ S ′ (Rn ) such that

(6.1.8) (∆ + z)ν+1 Fν = δ0 .

xβ ∂ α Fν is a bounded continuous function if |α| < |β| + 2(ν + 1) − n, and

(6.1.9) (∆ + z)Fν = Fν−1 , ν > 0,

(6.1.10) 2ν∂Fν /∂x = xFν−1 , ν > 0.

Proof. The equation (6.1.8) means that (z − |ξ|2 )ν+1 Fbν = 1 if Fbν is the Fourier trans-
form of Fν . Hence
Z
Fν (x) = (2π) −n
eihx,ξi (z − |ξ|2 )−ν−1 dξ

with the integral taken in the sense of distribution theory. Thus xβ ∂ α Fν is the inverse
Fourier transform of (i∂ξ )β (iξ)α (z − |ξ|2 )−ν−1 which is integrable if |α| − |β| − 2(ν +
1) < −n, so xβ ∂ α Fν is then a bounded continuous function. (6.1.9) is an obvious
consequence of the fact that Fν is the unique solution of (6.1.8). The Fourier transform
of ∂Fν /∂xj is
iξj (z − |ξ|2 )−ν−1 = i∂ξj (z − |ξ|2 )−ν /2ν,
which proves (6.1.10) and ends the proof of the lemma.
1
The uniqueness implies that Fν is a function of |x| = (x21 + · · · + x2n ) 2 , and it would
be easy to express Fν in terms of Bessel functions. We shall not do so but note that
Fν (x) = O(|x|2(ν+1)−n log |x|) as x → 0, where the logarithm can be dropped if n is
odd or 2(ν + 1) < n. This follows from the corresponding well known facts for the
Laplacian. In particular, Fν ∈ L1 .
If f ∈ D ′ (R), we have for x 6= 0
n
X n
X
jk 2 2
′
g (x)∂k f (|x| ) = 2f (|x| ) g jk (x)xk
k=1 k=1
= 2f ′ (|x|2 )xj = ∂j f (|x|2 ), j = 1, . . . , n,
Pn
for k=1 gjk (x)xk = xj since the coordinates are geodesic. When f (|x|2 ) is replaced
by Fν the first and last expressions are equal as distributions in Rn since both are
locally integrable functions. Hence it follows from (6.1.9), (6.1.10) that

n
X
g jk (x)∂k Fν = (2ν)−1 xj Fν−1 , if ν > 0;
k=1
n
X
∂j (g jk (x)∂k Fν ) + zFν = Fν−1 , if ν ≥ 0,
j,k=1
112 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

where F−1 = δ0 . If uν ∈ C ∞ with values in CN , it follows now from (6.1.5) that

X
(P + z)(uν Fν ) = (P uν )Fν + (uν + ν −1 (huν + xj ∂j uν ))Fν−1 , ν > 0;
(6.1.11) X
(P + z)(u0 F0 ) = (P u0 )F0 + u0 (0)δ0 + 2(hu0 + xj ∂j u0 )f ′ (|x|2 ),

where F0 = f (|x|2 ). Here we have used (6.1.7) and introduced the notation
1 1
X
(6.1.12) h = 12 xj g − 2 ∂ j g 2 .

If we sum the equations (6.1.11) for ν = 0, . . . , µ, we obtain

µ
X
(6.1.13) (P + z) uν Fν = u0 (0)δ0 + (P uµ )Fµ ,
ν=0

if the coefficients uν satisfy the differential equations

n
X
(6.1.14) νuν + huν + xj ∂j uν + νP uν−1 = 0, ν = 0, . . . , µ;
j=1

here the undefined term P u−1 should be omitted when ν = 0. These equations are
chosen so that the terms involving F0 , . . . , Fµ−1 drop out. Note that the error term
(P uµ )Fµ in the right-hand side of (6.1.13) is as smooth as desired if µ is chosen large.
The first equation (6.1.14) can be integrated explicitly, for
n n
1 1 1 1
X X
(6.1.12)′ h= 1
2
xj g − 2 ∂ j g 2 = g − 4 xj ∂j (g 4 ).
j=1 j=1

This means that

1
(6.1.15) u0 (x) = u0 (0)g(x)− 4 ,

and that the other equations (6.1.13) can be written in the form
n
1 1
X
(6.1.14) ′
(ν + xj ∂j )(g 4 uν ) + νg 4 P uν−1 = 0, ν = 1, . . . , µ.
j=1

A differential equation of the form

n
X
(6.1.16) (ν + xj ∂j )v = f
1

in a starshaped neighborhood of the origin in Rn has a unique smooth solution if ν is

a positive integer and f is smooth. In fact, with polar coordinates it can be written

(ν + r∂/∂r)v = f, thus ∂(r ν v)/∂r = r ν−1 f,

 METRIC ELLIPTIC OPERATORS 113

which gives the unique smooth solution

Z 1
(6.1.17) v(x) = tν−1 f (tx) dt.
0

Thus the equations (6.1.14) have a unique solution when u0 (0) is prescribed.
The functions uν obtained depend linearly on u0 (0) = e(0), so we can write them
in the form ũν (x)e(0), where ũ(x) is a N × N matrix. The corresponding linear trans-
formation from E0 to Ex , both of which have been identified with CN , is independent
of the choice of basis made. In fact, by the uniqueness in Lemma 6.1.3 the basis is
uniquely determined up to a linear transformation independent of x. From now on we
drop tilde from the notation, so uν (x) ∈ L(E0 , Ex ) ∼
= Ex ⊗ E0∗ and u0 (0) is the identity
in E0 . Note that uν is independent of z but the distributions Fν depend on z.
For every y ∈ M the exponential map Ty M ∋ X 7→ expy (X) at y gives geodesic
coordinates centered at y when we introduce an orthonormal basis in Ty M . Using
Lemma 6.1.3 we then construct a synchronous local frame for E, so that P can be
written in the form (6.1.5) with (6.1.7) fulfilled, in a neighborhood of the origin. The
functions uν obtained depend of course in a C ∞ fashion on the parameters y. Now
the distributions Fν (X) pull back to distributions Fν (x, y) defined in a neighborhood
of the diagonal in M × M , which only depend on the geodesic distance between x and
y, and are locally integrable in y for fixed x, hence also in x for fixed y. We multiply
the functions uν by a C ∞ cutoff function which is 1 near the center of the geodesic
coordinates and has support in a small neighborhood and pull them back to M × M .
Since Fν (X) ∈ C ∞ when X ∈ Rn \ {0}, this does not introduce any new singularities
in (6.1.13). We have then obtained C ∞ functions uν ∈ C ∞ (M × M, E ⊠ E ∗ ) where
E ⊠ E ∗ is the vector bundle with fiber Ex ⊗ Ey∗ at (x, y) ∈ M × M , such that Fν (x, y)
is defined in a neighborhood of supp uν and for every µ
µ
X
(Px + z) uν Fν − δdiag ∈ C 2µ+1−n (M × M, E ⊠ E ∗ ).
ν=0

Here δdiag is the kernel of the identity map in C ∞ (M, E), that is, the distribution in
D ′ (M × M, E ⊠ E ∗ ) such that
Z
δdiag (x, y)ϕ(y) dvol(y) = ϕ(x), ϕ ∈ C0∞ (M, E),

p
which means that in terms of local coordinates δdiag (x, y) is equal to δ(x − y)/ g(y)
times the identity matrix. We can choose F ∈ D ′ (M × M, E ⊠ E ∗ ) so that
µ−1
X
F− uν Fν ∈ C 2µ+1−n (M × M, E ⊠ E ∗ ), ∀µ.
ν=0

In fact, since uν Fν ∈ C 2ν+1−n we can choose ψν ∈ C ∞ (M × M, E ⊠ E ∗ ) so that all

seminorms in C 2ν+1−n of uν Fν − ψν over a compact set Kν in M × M are < 2−ν . If
114 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

Kν contains an arbitrary
P∞ compact set in M × M when ν is large enough, this implies
that the series F = 0 (uν Fν − ψν ) converges in D ′ (M × M, E ⊠ E ∗ ), and F has the
required properties. Thus

(6.1.18) (Px + z)F (x, y) = δ + R(x, y), R ∈ C ∞ (M × M, E ⊠ E ∗ ).

If M is not compact it is useful to take the cutoffs in the definition so close to the
diagonal that the maps

supp F ∋ (x, y) 7→ x ∈ M and supp F ∋ (x, y) 7→ y ∈ M

are proper. One calls F a proper right parametrix then.

If h ∈ C ∞ (M, E) then
Z
x 7→ F (x, y)h(y) dvol(y) = (F h)(x) is in C ∞ (M, E).

Since F is in C ∞ outside the diagonal, it is sufficient to verify this in local coordinates.

Introducing geodesic coordinates Y centered at x as integration variables we see that
the terms uν (x, y)Fν (x, y) give a C ∞ contribution, for with these coordinates Fν will
only depend on Y and not on x, and uν and h are in C ∞ as functions of Y and x.
This proves the assertion. In the same way we see that the operator F ∗ defined by
the kernel F (y, x) maps C ∞ (M, E ∗ ) to C ∞ (M, E ∗ ). We have hF ϕ, ψi = hϕ, F ∗ ψi,
ϕ ∈ C0∞ (M, E), ψ ∈ C0∞ (M, E ∗ ), where h·, ·i is the integral with respect to dvol(x) of
the scalar product between the fibers Ex and Ex∗ .Hence both F and F ∗ have continuous
extensions to D ′ .
The adjoint operator P ∗ : C ∞ (M, E ∗ ) → C ∞ (M, E ∗ ) is defined similarly so that
hP ϕ, ψi = hϕ, P ∗ ψi. From the equation (6.1.18), which for the corresponding operators
means that

(6.1.18)′ (P + z)F ϕ = ϕ + Rϕ, ϕ ∈ C0∞ (M, E),

it follows by taking adjoints that F ∗ is a left parametrix of P ∗ + z,

(6.1.18)′′ F ∗ (P ∗ + z)ψ = ψ + R∗ ψ, ψ ∈ C0∞ (M, E ∗ ).

The principal symbol of P ∗ is p0 (x, ξ) times the identity in E ∗ , so if we apply (6.1.18)′′

with P replaced by P ∗ we obtain a distribution G ∈ C ∞ (M × M, E ⊠ E ∗ ) with the
same regularity properties as F which is a proper left parametrix of P . Then

F ϕ − Gϕ = (G(P + z) − S)F ϕ − G((P + z)F − R)ϕ = (GR − SF )ϕ, ϕ ∈ C0∞ (M, E),

where R and S have C ∞ kernels. Hence F − G is the kernel of GR − SF , which is in

C ∞ since one of the factors in each term is. (The compositions are well defined since
we take properly supported parametrices.) We may therefore conclude that F is also
a left parametrix, so (6.1.18) can be strengthened to

(6.1.18)′′′ (P + z)F ϕ = ϕ + R1 ϕ, F (P + z)ϕ = ϕ + R2 ϕ, ϕ ∈ C0∞ (M, E),

 METRIC ELLIPTIC OPERATORS 115

where both R1 and R2 have kernels in C ∞ (M × M, E ⊠ E ∗ ). We have now forged

all the tools required to discuss existence and uniqueness theorems for the operator P
when M is a compact manifold. However, we first observe that the condition z ∈ / R+
is easily removed here. In fact, for any z we can modify the definition of Fν in Lemma
6.1.4 to the inverse Fourier transform of (1 − χ(ξ))(z − |ξ|2 )−ν−1 where χ ∈ C0∞ is
equal to 1 in a neighborhood of the real zeros of z − |ξ|2 . Then δ0 is replaced by δ0 − R
in (6.1.8), where Rb = χ, hence R ∈ S; (6.1.9) remains valid, and in (6.1.10) we just get
an additional term in S. The new smooth terms do not affect the arguments above,
so (6.1.18)′′′ holds for arbitrary z. In what follows we take z = 0, but we shall use
general z later on to handle heat equations.

Theorem 6.1.5. If M is a compact C ∞ manifold, E a C ∞ vector bundle over M ,

and P a smooth metric differential operator in C ∞ (M, E), then
(i) Ker P = {u ∈ C ∞ (M, E); P u = 0} is finite dimensional;
(ii) u ∈ D ′ (M, E), P u ∈ C ∞ (M, E) implies u ∈ C ∞ (M, E);
(iii) the equation P u = f ∈ D ′ (M, E) has a solution u ∈ D ′ (M, E) if and only
if hf, vi = 0 for all v ∈ Ker P ∗ , which is a finite dimensional subspace of
C ∞ (M, E ∗ ).

Proof. To prove (ii) we note that F f = F P u = u + R2 u by (6.1.18)′′′ , and F f ∈ C ∞ ,

R2 u ∈ C ∞ since f ∈ C ∞ and R2 ∈ C ∞ . Hence u ∈ C ∞ . It follows also that if
u ∈ Ker P , then u + R2 u = 0, so Fredholm theory for the operator R2 with C ∞ kernel
proves that Ker P is finite dimensional. If P u = f then

hf, vi = hP u, vi = hu, P ∗ vi = 0, if v ∈ Ker P ∗ ,

and Ker P ∗ is finite dimensional by (i) applied to P ∗ . To solve the equation P u = f

with f ∈ D ′ (M, E) we set u = v +F f and get the equivalent equation P v = f −P F f =
−R1 f . Since R1 f ∈ C ∞ it follows that it suffices to discuss solvability of P u = f when
f ∈ C ∞ . With u = F w the equation becomes w +R1 w = f , which by Fredholm theory
can be solved for all f such that hf, hj i = 0, j = 1, . . . , N , where h1 , . . . , hN is a basis
in C ∞ (M, E ∗ ) for solutions of h + R1∗ h = 0. Thus the equation P u = f can be solved
if hf, hj i = 0, j = 1, . . . , N . Hence it can be solved if (hf, h1 i, . . . , hf, hN i) ∈ V ,

V = {hP u, hj i; u ∈ C ∞ (M, E)} ⊂ CN .

PN ∗
If (λ 1 , . . . , λN ) is a normal of the linear space V , then 1 λj hj ∈ Ker P , so
λj hf, hj i = 0 if hf, Ker P ∗ i = 0. Hence (hf, h1 i, . . . , hf, hN i) ∈ V , which completes
P
the proof.

It is easy to give other spaces than C ∞ and D ′ , such as Sobolev spaces and Hölder
spaces, for which (ii) and (iii) in Theorem 6.1.5 are true, but we shall postpone intro-
ducing these spaces until we need them.
Theorem 6.1.5 has an important corollary for first order operators:
116 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

Corollary 6.1.6. Let M be a compact C ∞ manifold, E and F two C ∞ vector bundles

over M , and

P : C ∞ (M, E) → C ∞ (M, F ), Q : C ∞ (M, F ) → C ∞ (M, E)

first order differential operators such that QP is a metric operator. Then

(i) Ker P = {u ∈ C ∞ (M, E); P u = 0} is finite dimensional;
(ii) u ∈ D ′ (M, E), P u ∈ C ∞ (M, F ) implies u ∈ C ∞ (M, E);
(iii) the equation Qv = g ∈ D ′ (M, E) has a solution v ∈ D ′ (M, F ) if and only
if hg, ui = 0 for all u ∈ Ker Q∗ , which is a finite dimensional subspace of
C ∞ (M, E ∗ ).

Proof. P u = 0 implies QP u = 0, and P u ∈ C ∞ (M, F ) implies QP u ∈ C ∞ (M, E), so

(i), (ii) follow from (i), (ii) in Theorem 6.1.5. Since P ∗ Q∗ = (QP )∗ is a metric elliptic
operator, the finite dimensionality of Ker Q∗ follows from (i). The equation Qv = g has
a solution if the equation QP u = g can be solved, which by Theorem 6.1.5 is possible
if g is orthogonal to the finite dimensional space Ker(QP ). As at the end of the proof
of Theorem 6.1.5 it follows that the equation Qv = g can be solved precisely when
hv, Ker Q∗ i = 0.
If p and q are the principal symbols of P and Q, then the hypothesis implies that

(6.1.19) q(ξ)p(ξ) = r(ξ)IEx , where r(ξ) > 0, 0 6= ξ ∈ Tx∗ M.

Conversely, if P is given and such a linear symbol q exists, then we can choose Q so
that the hypotheses in the theorem are fulfilled. When E and F have the same fiber
dimension, then (6.1.19) implies that p(ξ)q(ξ) = r(ξ)IFx , for p(ξ)q(ξ)p(ξ) = r(ξ)p(ξ)
and p(ξ) is invertible. Hence the roles of P and Q may be interchanged, so the equation
P u = f ∈ D ′ (M, F ) has a solution u ∈ D ′ (M, E) if and only if u is orthogonal to
Ker P ∗ , which is a finite dimensional subspace of C ∞ (M, F ∗ ); we have statements
analogous to (i) and (ii) for Q.
6.2. The exterior algebra of a Euclidean vector space. As a preparation for
the Hodge theory in Section 6.3 we shall discuss here some elementary algebraic aspects
of exterior differential forms. Let V be a vector space of dimension n < ∞ over R,
and denote its dual space by V ∗ . Recall that the space ∧p V ∗ of alternating p linear
forms on V is spanned by the forms θ1 ∧ · · · ∧ θp , θj ∈ V ∗ , defined by

(6.2.1) V × · · · × V ∋ (v1 , . . . , vp ) 7→ dethvk , θj ipj,k=1 .

| {z }
p times

In fact, if L is such a form and e1 , . . . , en is a basis for V , θ1 , . . . , θn the dual basis for
V ∗ , then
X X
L(v1 , . . . , vp ) = L( hv1 , θj1 iej1 , . . . , hvp , θjp iejp )
dethvk , θjl ipk,l=1 L(ej1 , . . . , ejp )/p!.
X X
= hv1 , θj1 i . . . hvp , θjp iL(ej1 , . . . , ejp ) =
 THE EXTERIOR ALGEBRA OF A EUCLIDEAN VECTOR SPACE 117

Thus the forms (6.2.1) give a basis for ∧p V ∗ even if θ1 , . . . , θp are restricted to elements
in a fixed basis for V ∗ .
The forms

(6.2.1)′ V ∗ × · · · × V ∗ ∋ (θ1 , . . . , θp ) 7→ dethvk , θj ipj,k=1 ,

| {z }
p times

span ∧p V for reasons of symmetry. We have used the duality between ∧p V and ∧p V ∗
such that

(6.2.2) hv1 ∧ · · · ∧ vp , θ1 ∧ · · · ∧ θp i 7→ dethvk , θj ipj,k=1 .

(Many authors use another definition where the right-hand side is divided by p!, for
that is the duality inherited from the natural duality of the tensor products. See
Sternberg [1, p. 19] for a discussion of this point. Kobayashi and Nomizu [1] use the
division by p!, which should be kept in mind when comparing identities.) The existence
and uniqueness of a bilinear form on ∧p V × ∧p V ∗ such that (6.2.2) holds is obvious,
for there is a unique bilinear form such that (6.2.2) holds when vk are chosen in a fixed
basis for V and θj in a fixed basis for V ∗ , and in view of the multilinearity and skew
symmetry of the two sides they must then be equal for arbitrary v1 , . . . , vp , θ1 , . . . , θp .
Assume now that V is a Euclidean vector space with scalar product denoted by (·, ·).
The linear form v ♭ : V ∋ v ′ 7→ (v, v ′ ) is an element 6= 0 in V ∗ , so we have a bijection
V ∋ v 7→ v ♭ ∈ V ∗ with inverse denoted by ♯; these are the musical isomorphisms we
used already in Chapter III. In particular, this gives a norm in V ∗ , which is of course
the dual norm
kθk = sup |hv, θi|/kvk,
06=v∈V

and the scalar product of θ1 , θ2 ∈ V ∗ is

(θ1 , θ2 ) = hθ1♯ , θ2 i = (θ1♯ , θ2♯ )

where we have used the convention that (·, ·) denotes scalar product in a Euclidean
space while h·, ·i denotes the bilinear form in a space and its dual. The map θ 7→ θ ♯
extends to a map ∧p V ∗ ∋ ϕ 7→ ϕ♯ ∈ ∧p V , and we define

(6.2.3) (ϕ, ψ) = hϕ♯ , ψi, ϕ, ψ ∈ ∧p V ∗ .

If ε1 , . . . , εn is an orthonormal basis in V ∗ then ej = ε♯j is an orthonormal basis for V

which is biorthogonal to the basis in V ∗ , so we obtain
X X X
(6.2.4) (ϕ, ψ) = ϕI ψI /p!, if ϕ = ϕI εI /p!, ψ = ψI εI /p!,
|I|=p |I|=p |I|=p

where ϕI and ψI are antisymmetric in the p indices of I = (i1 , . . . , ip ), all ranging

from 1 to n, and εI = εi1 ∧ · · · ∧ εip . In fact, hε♯I , εJ i equals 0 unless the indices in
118 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

I are different and J is a permutation of them; in that case, we have the sign of the
permutation. This means that for every I we get p! equal non-zero terms ϕI ψJ hε♯I , εJ i
which proves (6.2.4). By (6.2.4) it is clear that (6.2.3) gives a Euclidean structure in
∧p V ∗ . We can suppress the factorials in (6.2.4) if we sum only for strictly increasing
P′
multiindices I, that is, i1 < · · · < ip ; when we do that we shall use the notation .
∗
Note that our definition (6.2.3) did not use a basis for V or V . If O is an orthogonal
transformation in V and we define

(O ∗ ϕ)(v1 , . . . , vp ) = ϕ(Ov1 , . . . , Ovp ), v1 , . . . , vp ∈ V,

it follows that we have orthogonal invariance:

(6.2.5) (O ∗ ϕ, O ∗ ψ) = (ϕ, ψ), ϕ, ψ ∈ ∧p V ∗ .

∧n V ∗ is a one dimensional Euclidean vector space so it contains precisely two ele-

ments ε with kεk = 1. A choice of one of them means choosing an orientation of V ∗ ;
a basis ε1 , . . . , εn for V ∗ is defined to be positively oriented if ε1 ∧ · · · ∧ εn is a positive
multiple of ε. If we fix an orientation ε then

ϕ ∧ ψ = B(ϕ, ψ)ε, ϕ ∈ ∧p V ∗ , ψ ∈ ∧n−p V ∗

defines a non-degenerate bilinear form on ∧p V ∗ × ∧n−p V ∗ , hence a duality between

these spaces. Thus we obtain a linear isomorphism of ∧p V ∗ on the dual of ∧n−p V ∗ ,
which we have identified with ∧n−p V ∗ using (6.2.3), so we have an isomorphism ∗ :
∧p V ∗ → ∧n−p V ∗ such that

(6.2.6) ϕ ∧ ψ = ε(∗ϕ, ψ), ϕ ∈ ∧p V ∗ , ψ ∈ ∧n−p V ∗ .

If ε1 , . . . , εn is an orthonormal basis in V ∗ with ε1 ∧ · · · ∧ εn = ε, then

(6.2.7) ∗(εi1 ∧ · · · ∧ εip ) = σεip+1 ∧ · · · ∧ εin ,

if (i1 , . . . , in ) is a permutation with sign σ of 1, . . . , n.

Exercise 6.2.1. Show that ∗(v ∧ w) is the usual vector product if v, w are vectors in
R3 .
It is obvious that the square of the ∗ operator must be an automorphism of ∧p V ∗
for every p. Under the hypotheses of (6.2.7) we obtain

∗(εp+1 ∧ · · · ∧ εn ) = (−1)p(n−p) ε1 ∧ · · · ∧ εp ,

for the permutation (p + 1, . . . , n, 1, . . . , p) of (1, . . . , n) has p(n − p) inversions. This

means that

(6.2.8) ∗ ∗ ϕ = (−1)p(n−p) , ϕ ∈ ∧p .
 THE EXTERIOR ALGEBRA OF A EUCLIDEAN VECTOR SPACE 119

This is independent of the choice of orientation, for changing the orientation means
replacing ∗ by −∗, which preserves any formula where there is an even number of
factors ∗.
It is often convenient to write (6.2.6) in a different way. If we replace ϕ by ∗ϕ, we
obtain
ε(ϕ, ψ) = (−1)p(n−p) (∗ϕ) ∧ ψ = ψ ∧ ∗ϕ,
hence
(6.2.6)′ ψ ∧ ∗ϕ = ε(ϕ, ψ), ϕ, ψ ∈ ∧n−p V ∗ .
Using this formula it is easy to write down the ∗ operator explicitly for an arbitrary
basis ε1 , . . . , εn in V ∗ . Set g jk = (εj , εk ), and let I, J be multiindices of length p, n − p,
such that I, J is a positive permutation of 1, . . . , n. If the basis ε1 , . . . , εn is positively
oriented, that is, ε = ε1 ∧ · · · ∧ εn /kε1 ∧ · · · ∧ εn k, then taking ψ = εI in (6.2.6)′ gives
(6.2.9)
′
X
p
1
(∗ϕ)J = (ϕ, εI )/kε1 ∧ · · · ∧ εn k = |K| = pϕK det g kν Iµ ν,µ=1 / det(g νµ )nν,µ=1 2 .
2
If n is even, n = 2l, then ∗ maps ∧l V ∗ into itself, and the square is (−1)l . For a
positively oriented orthonormal basis θ1 , . . . , θ2l in V ∗ we have
2
∗(θ1 ∧ · · · ∧ θl + cθl+1 ∧ · · · ∧ θ2l ) = θl+1 ∧ · · · ∧ θ2l + (−1)l cθ1 ∧ · · · ∧ θl .
If l is even, then taking c = ±1 we see using such elements in ∧l V
∗ that ∧l V ∗ = ∧+ ⊕∧−
where ∧+ and ∧− are subspaces of the same dimension 21 2ll and ∗λ = ±λ when
λ ∈ ∧± . This is the general form of the decomposition we noticed in Lemma 2.3.3
when n = 4. If l is odd, we must take c = ±i, so it is the complexification ∧l VC∗ which
is the direct sum of the eigenspaces of ∗ with eigenvectors ±i.
To generalize this to ∧∗ V ∗ = ⊕np=0 ∧p V ∗ , still with n = 2l, we define

(6.2.10) τ (ϕ) = ip(p−1)+l ∗ ϕ, ϕ ∈ ∧p VC∗ .

Since p(p − 1) is even this is real if l is even and purely imaginary if l is odd, so it is
necessary to go to the complexification then. When ϕ ∈ ∧pC we have

τ 2 ϕ = ip(p−1)+l+(2l−p)(2l−p−1)+l (−1)p(2l−p) ϕ.
The exponent of −1 here is congruent to p2 mod 2 and the exponent of i is 4l2 −
4lp + 2p2 ≡ 2p2 mod 4, so τ 2 ϕ = ϕ. If p 6= l it follows that ∧p VC∗ ⊕ ∧2l−p VC∗ is the
direct sum of the eigenspaces
of τ corresponding to the eigenvalues ±1, and these must
have the same dimension 2l p ∗
p since none intersects ∧ VC ⊕ {0}. Summing up, we have
proved:
Proposition 6.2.1. If the oriented Euclidean vector space V is of even dimension 2l,
then the complexification ∧∗ VC∗ of ∧∗ V ∗ is the direct sum of the eigenspaces ∧± of the
operator τ defined by (6.2.10) corresponding to the eigenvalues ±1. ∧+ and ∧− have
the same dimension, and ∧± is the direct sum of its intersections with ∧p VC∗ +∧2l−p VC∗
when 0 ≤ p ≤ l.
120 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

6.3. The Hodge decomposition. Let M be a C ∞ Riemannian manifold. Then

the exterior powers ∧p Tx∗ M are the fibers of a vector bundle ∧p T ∗ M with the obvious
trivializations provided by the basis dx1 , . . . , dxn in T ∗ M over a local coordinate patch.
We shall denote by λp the space of sections C ∞ (M, ∧p T ∗ M ), that is, the C ∞ p forms
on M . If ϕ, ψ ∈ λp , then (ϕ, ψ)(x) is a C ∞ function on M as defined in Section 6.2
using the Euclidean metric in Tx M . We set
Z
(6.3.1) (ϕ, ψ) = (ϕ, ψ)(x) dvol(x), ϕ, ψ ∈ λp , supp ϕ ∩ supp ψ ⋐ M.
M
If M is oriented we can use (6.2.6)′ to write the integral in a more convenient way in
terms of the ∗ operator. With local coordinates x1 , . . . , xn which are positively oriented
we have
1
ε = dx1 ∧ · · · ∧ dxn /kdx1 ∧ · · · ∧ dxn k = (det gjk ) 2 dx1 ∧ · · · ∧ dxn ,
which means that
Z
(6.3.1) ′
(ϕ, ψ) = ϕ ∧ ∗ψ, ϕ, ψ ∈ λp , supp ϕ ∩ supp ψ ⋐ M.
M
This formula makes it easy to compute the formal adjoint of the exterior differential
operator d : λp → λp+1 . It is defined by
(dϕ, ψ) = (ϕ, d∗ ψ), ϕ ∈ λp , ψ ∈ λp+1 , supp ϕ ∩ supp ψ ⋐ M.
By Stokes’ formula we obtain
Z Z Z
p
(dϕ, ψ) = (dϕ) ∧ ∗ψ = d(ϕ ∧ ∗ψ) − (−1) ϕ∧d∗ψ
Z
p+1+p(n−p)
= (−1) ϕ ∧ ∗ ∗ d ∗ ψ.

Since p2 − p = p(p − 1) is an even number, it follows that

(6.3.2) d∗ = (−1)np+1 ∗ d ∗ on λp+1 .
Since the right-hand side contains two factors ∗ it does not change if the orientation
is changed, so the formula can also be used for non-oriented manifolds if one keeps in
mind that intermediate products such as d∗ make no sense.
Assume now that M is a compact Riemannian manifold. Recall that the de Rham
cohomology groups are defined by
H p (M ) = {ψ ∈ λp ; dψ = 0}/dλp−1 .
Since λp is a pre-Hilbert space with the scalar product (6.3.1) it is natural to try to
find in each class an element with minimal norm, for if it exists it must be unique. If
dψ = 0 and ψ minimizes the norm in its cohomology class, then
kψk2 ≤ kψ + dϕk2 , ϕ ∈ λp−1 ,
or equivalently,
(ψ, dϕ) = 0, ϕ ∈ λp−1 .
Thus the minimum property is equivalent to d∗ ψ = 0.
 THE HODGE DECOMPOSITION 121

Theorem 6.3.1. On a compact Riemannian manifold the equations dψ = 0, d∗ ψ = 0

for ψ ∈ λp (M ) are equivalent to ∆ψ = 0, where ∆ is the Hodge Laplacian

∆ = dd∗ + d∗ d : λp → λp .

Proof. It is obvious that ∆ψ = 0 if dψ = 0 and d∗ ψ = 0. On the other hand, if

∆ψ = 0, then

0 = (∆ψ, ψ) = (dd∗ ψ, ψ) + (d∗ dψ, ψ) = (d∗ ψ, d∗ ψ) + (dψ, dψ),

so d∗ ψ = dψ = 0.
Definition 6.3.2. The differential forms satisfying the equation ∆ψ = 0 are called
harmonic forms.
To justify the notation ∆ and the term “harmonic” we shall calculate the Hodge
Laplacian, starting with the case where ϕ ∈ λp (Rn ) and the metric in Rn is the
standard Euclidean. Write
X′
ϕ= |I| = pϕI dxI
P′
where denotes summation over increasing sequences I = (i1 , . . . , ip ) and dxI =
dx ∧ · · · ∧ dxip . Then
i1

′
X ′
X
i I
dϕ = I, i∂i ϕI dx ∧ dx , ∗ϕ = I, JϕI dxJ sgn(I, J)

where J is an increasing sequence of n − p indices and I, J is a permutation of 1, . . . , n.

For the Dirichlet integral we now obtain if ϕ ∈ λp has compact support
Z ′  
2 ∗ 2
X i I
(∆ϕ, ϕ) = kdϕk + kd ϕk = ∂i ϕI ∂i′ ϕI ′ sgn ′
i I′
   
j J I J

+ ∂j ϕI ∂j ′ ϕI ′ sgn ′ ′ sgn dx.
j J I′ J′

If I 6= I ′ then the contribution is 0 unless the difference is in just one index. Assume
for example that I = (2, . . . , p + 1), I ′ = (1, . . . , p). Then we must have i = j ′ = 1,
i′ = j = p + 1, and we obtain
   
i I 1 2 ... p+1
sgn = sgn = (−1)p ,
i′ I′ p+1 1 ... p
   
j J p+1 1 p+2 ...
sgn = sgn = −1
j′ J′ 1 p+1 p+2 ...
   
I J 2 ... p +1 1 p+2 ...
sgn = sgn = (−1)p .
I′ J′ 1 ... p p+1 p+2 ...
122 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

If we integrate by parts moving derivatives to the first factor the two terms will therefore
give contributions which cancel each other, and the only remaining terms are those with
I = I ′ and i = i′ ∈ J or j = j ′ ∈ I, hence
Z X ′ Xn
2 2
∗
−∂i2 ϕI ϕI dx.


(∆ϕ, ϕ) = kdϕk + kd ϕk = I
i=1

Since ∆ is formally self-adjoint, polarization gives that

′
X n
X
(6.3.3) ∆ϕ = I −∂i2 ϕI dxI .
i=1

Thus the Hodge Laplacian operates componentwise as minus the classical Laplacian;
the minus sign is natural since the Hodge Laplacian is a positive operator by its defi-
nition. We shall write ∆H whenever confusion seems possible.
To compute ∆ generally we can use the formula
∆ = (−1)n(p−1)+1 d ∗ d ∗ +(−1)np+1 ∗ d ∗ d on λp ,
which follows from (6.3.2) and can be used locally even if M is not orientable. Choose
geodesic coordinates x1 , . . . , xn centered at the point where we want to calculate ∆ϕ.
Then gjk (x) = δjk + O(|x|2 ), by (3.1.15) the second derivatives of gjk are linear func-
tions of the Riemann curvature tensor, and gjk + g jk − 2δjk = O(|x|3 ). We can use
(6.2.9) to express the ∗ operator. When we compute ∆H ϕ(x) the result will be the
same as for the Euclidean metric unless two derivatives fall on a coefficient involving
g jk . Hence (6.3.3) only needs modification by a term of order 0.
Since ϕ can also be regarded as a P skew symmetric tensor field, we could also form
another Laplacian by the contraction g jk ϕ,jk . At the center of a geodesic coordinate
system this will also apart from lower order terms be the classical Laplacian acting on
each component of ϕ, so
X
∆H ϕ + g jk ϕ,jk is of order 0.

When p = 0 there P isjkno term d ∗ d∗ so no second order derivatives of gjk can occur,
hence −∆H ϕ = g ϕ,jk is the standard Laplace-Beltrami operator. We shall return
to the exact calculation for p 6= 0 later on in this section. For the moment it suffices
for us just to observe that the calculation of the leading term just made shows that
the principal symbol of the Hodge Laplacian is −|ξ|2 . Hence ∆H (or rather −∆H )
satisfies the hypotheses of Theorem 6.1.5. Thus the space Hp of harmonic p forms is
finite dimensional, and every ψ ∈ λp can be written in the form
ψ = h + ∆ϕ
where h is the orthogonal projection of ψ on Hp and ϕ ∈ λp is uniquely determined
mod Hp . Here we have of course used that ∆H is formally self-adjoint. With ϕ+ = dϕ
and ϕ− = d∗ ϕ, it follows that
(6.3.4) ψ = h + dϕ− + d∗ ϕ+ , h ∈ Hp , ϕ− ∈ λp−1 , ϕ+ ∈ λp+1 .
 THE HODGE DECOMPOSITION 123

The three terms in this decomposition are mutually orthogonal, for dh = d∗ h = 0 by

Theorem 6.3.1 and

(h, dϕ− ) = (d∗ h, ϕ− ) = 0, (h, d∗ ϕ+ ) = (dh, ϕ+ ) = 0,

(dϕ− , d∗ ϕ+ ) = (d2 ϕ− , ϕ+ ) = 0.

Hence the decomposition (6.3.4) is unique (but ϕ− and ϕ+ are not). If dψ = 0 then
dd∗ ϕ+ = 0, which implies

(dd∗ ϕ+ , ϕ+ ) = (d∗ ϕ+ , d∗ ϕ+ ) = 0,

so dψ = 0 is equivalent to d∗ ϕ+ = 0. Hence we have proved:

Theorem 6.3.3 (Hodge). Every C ∞ p form on a compact C ∞ Riemannian manifold
has a unique decomposition (6.3.4) as the sum of one harmonic p form, one form in
the range of d, and one form in the range of d∗ . Every de Rham cohomology class
contains exactly one harmonic form, so H p (M ) can be identified with the space Hp of
harmonic p forms.
An immediate consequence is of course that H p (M ) is finite dimensional. If M is
oriented we also conclude that H p (M ) is isomorphic to H n−p (M ) (Poincaré duality).
In fact, we have ∆∗ = ∗∆ since

(∆ϕ, ψ) = (dϕ, dψ) + (∗d ∗ ϕ, ∗d ∗ ψ) = (∗d ∗ ∗ϕ, ∗d ∗ ∗ψ) + (d ∗ ϕ, d ∗ ψ) = (∆ ∗ ϕ, ∗ψ),

which implies

(∆ ∗ ϕ, ψ) = (∆ ∗ ∗ϕ, ∗ψ) = (∗ ∗ ∆ϕ, ∗ψ) = (∗∆ϕ, ψ).

Hence ∗ gives an isomorphism Hp → Hn−p .

Assume now that M is a compact oriented C ∞ Riemannian manifold. Then the
operator
D = d + d∗ : λ∗ → λ∗
is defined, and it satisfies the hypotheses of Corollary 6.1.6 since

(6.3.5) D2 = dd + dd∗ + d∗ d + d∗ d∗ = dd∗ + d∗ d = ∆.

From (6.3.5) it follows that the kernel consists of the harmonic forms, and D is formally
selfadjoint by its definition. Hence the range is the orthogonal space of the harmonic
forms, and the index is equal to 0. However, one can obtain operators with non-trivial
index by restricting D to suitable subbundles of ∧∗ T ∗ M .
First note that D maps forms of even (odd) degree to forms of odd (even) degree.
If we write
λeven = ⊕λ2p , λodd = ⊕λ2p+1 ,
it follows that the restriction Deo to λeven is a differential operator from λeven to λodd
with adjoint equal to the restriction of D to λodd , with values in λeven . Hence the
124 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

kernel of Deo is ⊕H2p , and the range is the orthogonal space of ⊕H2p+1 in λodd . This
means that
n
X n
X
eo p p
(6.3.6) ind D = (−1) dim H = (−1)p dim H p (M ),
p=0 p=0

which is the Euler characteristic of M . If we accept the fact that the Euler characteristic
is equal to the number of critical points of functions on M , counted with signs, then
Theorem 4.3.2 means that the index of Deo can be expressed in terms of the curvature
forms. This is the model for the index theorems which will be a major subject of this
chapter.
Assume now that M is a compact oriented C ∞ Riemannian manifold of even dimen-
sion n = 2l. By Proposition 6.2.1 the map τ defined in ∧∗ Tx∗ for every x ∈ M gives a
decomposition
(∧∗ T ∗ M )C = ∧+ ⊕ ∧− ,
where ∧± are complex vector bundles of fiber dimension 2n−1 . We claim that

(6.3.7) Dτ = −τ D,

which will prove that D maps sections of ∧+ (∧− ) to sections of ∧− (∧+ ). To prove
(6.3.7) we let ϕ be a p form. Since the dimension n is even, we have d∗ = − ∗ d∗, hence

2
Dτ ϕ = ip(p−1)+l (d − ∗d∗) ∗ ϕ = ip(p−1)+l (d ∗ ϕ − (−1)p ∗ dϕ)
τ Dϕ = τ (dϕ − ∗d ∗ ϕ) = i(p+1)p+l ∗ dϕ − i(p−1)(p−2)+l ∗ ∗d ∗ ϕ
= ip(p−1)+l (−1)p ∗ dϕ − (−1)p−1 (−1)2l−p+1 d ∗ ϕ

= ip(p−1)+l (−1)p ∗ dϕ − d ∗ ϕ .

This proves (6.3.7).

We shall now determine the index of the restriction D+ : C ∞ (M, ∧+ ) → C ∞ (M, ∧− )
of D. The adjoint of D+ is the restriction D− : C ∞ (M, ∧− ) → C ∞ (M, ∧+ ), so the
kernels H± of D± are the harmonic forms h with τ h = ±h, and by definition

ind D+ = dim H+ − dim H− .

k 2l−k
HC ⊕HC is mapped to itself by τ , and if k < l then the eigenspaces with eigenvalues
±1 are
k
{h ± τ h; h ∈ HC }
l
so they have the same dimension. What remains is to examine how τ splits HC . There
2
l
we have τ = i ∗ which is equal to ∗ if l is even and equal to i∗ if l is odd. For odd l this
l
means that HC is split according to the eigenspaces of the operator ∗ corresponding
to eigenvalues ±i, and since ∗ is a real operator in Hl , they have the same dimension
which gives no contribution to the index. On the other hand, if l = 2k is even, that is,
 THE HODGE DECOMPOSITION 125

n ≡ 0 mod 4, then τ is a real operator and it suffices to consider the eigenspaces of ∗

acting in the real vector space Hl corresponding to eigenvalues ±1. If ϕ ∈ Hl then
Z Z
(6.3.8) (ϕ, τ ϕ) = ϕ∧∗∗ϕ = ϕ∧ϕ
M

is positive (negative) when ϕ is in the eigenspace of ∗ with eigenvalue +1 (−1), so the

index of D+ is equal to the index of the quadratic form
Z
l
(6.3.9) H ∋ ϕ 7→ ϕ ∧ ϕ.
M

2
(If l is odd then ϕ ∧ ϕ = 0 since ϕ ∧ ϕ = (−1)l ϕ ∧ ϕ, so (6.3.9) vanishes then.)
We can give (6.3.9) an interpretation which does not involve harmonic forms. To
do so we note that the bilinear form
Z
(6.3.10) (ϕ, ψ) 7→ ϕ∧ψ
M

on {θ ∈ ∧l ; dθ = 0} induces a bilinear form on H l (M ). In fact, if ϕ = dθ, then

ϕ∧ψ = d(θ∧ψ) so (6.3.10) vanishes then, and the same is true if ψ = dθ. Interchanging
2
ϕ and ψ multiplies the integrand in (6.3.10) by (−1)l , so the form is symmetric if l is
even and it is skew symmetric if l is odd.
Definition 6.3.4. The signature of a compact oriented Riemannian manifold M of even
dimension 2l is the signature of the quadratic form induced in H l (M ) by (6.3.10), if l
is even, and it is 0 if l is odd.
With this definition we have proved

(6.3.11) ind D+ = the signature of M.

The expression of ind D+ in terms of Pontrjagin classes will be discussed later on.
We shall close this section by making a few
P additional calculations involving ∗, d∗
and ∆. Let ϕ be a one form and write ϕ = ϕj dxj in local coordinates. By (6.2.9)
n
n√
X
∗ϕ= ϕk g ki (−1)i−1 dx1 ∧ . . . |∧dx i
{z } ∧ · · · ∧ dx g
k,i=1 omit
n
n√
X
= Φi (−1)i−1 dx1 ∧ . . . ∧dx i
| {z } ∧ · · · ∧ dx g,
i=1 omit

i√
where Φ = ϕ♯ . Hence d ∗ ϕ = g)dx1 ∧ · · · ∧ dxn , so
P
i (∂i (Φ

n
− 12 1
X
(6.3.12) ∗
d ϕ = − ∗ d ∗ ϕ = −g ∂i (Φi g 2 ) = − div Φ,
i=1
126 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

where the last equality is a definition. The invariance of div Φ when Φ is a vector field
is also clear since for χ ∈ C0∞ (M ) we have
Z Z
(6.3.13) − χ div Φ dvol = hΦ, dχi dvol.
M M

This is also true if χ = 1 and Φ has compact support, that is, the integral of the
divergence of a vector field with compact support is equal to 0. We can also write
X
(6.3.14) div Φ = Φi ,i ,

for (6.3.14) is true in a geodesic coordinate system. (See also the proof of (3.4.15) and
Exercise 3.4.2.) Thus
Z X n
(6.3.14) ′
Φi ,i dvol = 0,
M i=1

which is a convenient formula to use for partial integration of expressions involving

covariant differentiation.
For the one form ϕ we have
X X
dϕ = ∂i ϕj dxi ∧ dxj = ϕj,i dxi ∧ dxj ,

since ϕj,i = ∂i ϕj + k Γij k ϕk and Γij k = Γji k . However, one must remember when
P
using this expression that ϕj,i is not antisymmetric in general. We shall now compute
d∗ dϕ by taking another one form ψ with compact support and calculating (dϕ, dψ).
At a point where gij = δij the pointwise scalar product is
n
X n
X n
X
1
2 (ϕj,i − ϕi,j )(ψj,i − ψi,j ) = (ϕj,i ψj,i − ϕi,j ψj,i ) = (∇ϕ, ∇ψ) − Φi ,j Ψj ,i ,
i,j=1 i,j=1 i,j=1

where Φ = ϕ♯ and Ψ = ψ ♯ . We have written

n
X ′ ′
(∇ϕ, ∇ψ)(x) = g ii g jj ϕj,i ϕj ′ ,i′
i,j,i′ ,j ′ =1

for the natural scalar product in Tx∗ M ⊗ Tx∗ M . The last expression is invariant so it is
always equal to (dϕ, dψ)(x), which proves that
Z Z X n
(dϕ, dψ) = (∇ϕ, ∇ψ) dvol − Φi ,j Ψj ,i dvol.
M M i,j=1

′
In
P thei lastj integral we can integrate by parts using (6.3.14) applied to the vector field
j Φ ,j Ψ , which gives
Xn
∗ ∗
d dϕ = ∇ ∇ϕ + Φi ,ji dxj .
i,j
 HEAT EQUATIONS 127

Here ∇∗ is the adjoint of the operator ∇ from one forms, that is, sections of T ∗ M , to
sections of T ∗ M ⊗ T ∗ M , with the scalar product above. Now we add
n
X n
X
dd∗ ϕ = −d Φi ,i = − Φi ,ij dxj
i=1 i,j=1

It follows from (3.1.6) (see Exercise 3.1.4) that

n
X n
X n
X
i i i l
(Φ ,ji −Φ ,ij ) = R lij Φ = Rlj Φl ,
i=1 i,l=1 l=1

where the last R is the Ricci tensor. Thus we have proved

n
X
(6.3.15) ∆H ϕ = ∇∗ ∇ϕ + Rij Φi dxj .
i,j=1

This makes the remarks following (6.3.3) explicit. The identity (6.3.15) is of the form
(6.1.4), corresponding to the Levi-Civita connection in T ∗ M , and c has been identified
as the Ricci curvature. There are similar formulas for forms of degree p > 1, due to
Weitzenböck [1]; see also de Rham [1, p. 131]. Formulas of the same structure as
(6.3.15) are called Weitzenböck decompositions in general.
To show the importance of (6.3.15) we now assume that M is compact and con-
nected, and take the scalar product of (6.3.15) with ϕ. This gives
Z X
2
(6.3.16) (∆H ϕ, ϕ) = k∇ϕk + Rij Φi Φj dvol, ϕ ∈ λ1 , Φ = ϕ♯ .
M

If ϕ is a harmonic one form, that is, ∆ϕ = 0, and the Ricci tensor is non-negative, it
follows from (6.3.16) that both terms in the right-hand side must vanish. Hence the
covariant differential of ϕ is equal to 0, and ϕ♯ is at every point contained in the kernel
of the Ricci tensor. If the Ricci tensor is strictly positive definite at some point, then ϕ
must vanish in a neighborhood. But kϕ(x)k is a constant since ∇ϕ = 0, so we obtain
the following theorem of Bochner:
Theorem 6.3.5. If h is a harmonic one form on a compact, connected Riemannian
manifold M with non-negative Ricci tensor, then ∇h = 0 and h♯ is in the kernel of R
at every point. Hence dim H 1 (M ) ≤ n. If in addition R is strictly positive definite at
some point, then h = 0, hence H 1 (M ) = 0 then.
Theorem 6.3.5 has been the starting point of much work on the connection between
the topology of a manifold and its curvature.
6.4. Heat equations. The first step to a calculation of the index of the operators
Deo and D+ introduced in Section 6.3 is to study the heat equations associated with the
corresponding metric elliptic operators. Let P be a metric differential operator acting
on sections of the C ∞ vector bundle E over the compact C ∞ Riemannian manifold
128 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

M . By E we shall also denote the lifting of E to R × M . We want to solve the Cauchy

problem to find u ∈ C ∞ (R+ × M, E) such that

(6.4.1) ∂t u = P u in [0, ∞) × M, u(0) = v,

where v is given in C ∞ (M, E). Formally this means that u = etP v, t ≥ 0, where

e−tz
Z
tP −1
e = (2πi) dz
Γ z+P

with an integration contour Γ such as a sector < π of the unit circle with the two
tangents in the direction of the bisector of the first and fourth quadrants. (We have
defined the principal symbol without the customary factors i, so our conventions make
P bounded above.) The distributions Fν (x) in (6.1.8), which depend on z ∈ C \ R+ ,
are inverse Fourier transforms of ξ 7→ (z − |ξ|2 )−ν−1 , and
Z
2
(2πi) −1
(z − |ξ|2 )−ν−1 e−tz dz = e−t|ξ| (−t)ν /ν!.
Γ

If we multiply Fν by e−tz /(2πi) and integrate over Γ, we thus obtain at least formally
2
the inverse Fourier transform of (−t)ν e−t|ξ| /ν!, that is,
2
(6.4.2) Hν (t, x) = (−t)ν H0 (t, x)/ν!, where H0 (t, x) = (4πt)−n/2 e−|x| /4t
, t > 0,

is the fundamental solution of the scalar heat equation. The identities (6.1.9), (6.1.10)
are transformed to the obvious identities

(∆ − ∂t )((−t)ν H0 (t, x)/ν!) = (−t)ν−1 H0 (t, x)/(ν − 1)!,

−2t∂H0 (t, x)/∂x = xH0 (t, x).

Recall that in (6.1.13) the coefficients uν were independent of z. Repeating the proof
of (6.1.13) we now obtain with the same coefficients uν
µ
X
(6.4.3) (P − ∂t ) uν Hν = (P uµ )Hµ .
0

Since ze−tz = −∂t e−tz , this can easily be justified also by an integration
Pµ of (6.1.13)
−tz
multiplied by e . Since H0 (t, ·) → δ0 as t → +0, it is clear that 0 uν Hν → δ0 as
t → +0. If we define Hν (t, x) = 0 when t ≤ 0, then Hν ∈ C k (R×U ), if νP > k +n/2 and
∞ µ
U is a geodesic coordinate patch in M . Every Hν is C outside (0, 0), so 0 uν Hν (t, x)
is a good parametrix for (∂t − P ) when µ is large.
Using a parametrix it is easy to solve the Cauchy problem (6.4.1). At first we put the
inhomogeneity in the equation instead of the data and consider the Cauchy problem

(6.4.1)′ (∂t − P )u = f, u(t, ·) = 0 when t < 0,

 HEAT EQUATIONS 129

where f (t, ·)P= 0 when t < 0 and f is smooth. To do so we define H

e locally by cutoffs
∞ ∞
of the sum 0 uν Hν with the terms approximated by C functions vanishing in the
lower half space, just as in the parametrix F in Section 6.1. Thus H e is a section of
∗
E ⊠ E on R × M × M with t ≥ 0 in the support and

(∂t − Px )H(t,
e x, y) = R(t, x, y) when t > 0; H(t,
e ·, ·) → δdiag as t → +0,

where R ∈ C ∞ (R × M × M, E ⊠ E ∗ ) and R(t, x, y) = 0 when t < 0. Now set

ZZ
u(t, x) = H(t
e − s, x, y)w(s, y) ds dvol(y).
s<t

The integral with respect to y converges to w(t, x) as s → t, so we obtain

ZZ
(∂t − P )u(t, x) = w(t, x) + R(t − s, x, y)w(s, y) ds dvol(y).
s<t

The equation (6.4.1)′ now becomes a Volterra integral equation

ZZ
w(t, x) + R(t − s, x, y)w(s, y) ds dvol(y) = f (t, x)
s<t

with C ∞ kernel, translation invariant in t, so by simple iteration one obtains a unique

solution ZZ
w(t, x) = f (t, x) + K(t − s, x, y)f (s, y) ds dvol(y).
s<t

Here K ∈ C ∞ (R × M × M, E ⊠ E ∗ ) also vanishes when t < 0. Hence

ZZ
H(t, x, y) = H(t,
e x, y) + H(t
e − s, x, z)K(s, z, y) ds dvol(z)
0<s<t

is an exact fundamental solution for the Cauchy problem. (The product here is the
duality between Ez∗ and Ez .) The correction term is infinitely differentiable and van-
ishes for t < 0, so the asymptotic behavior at the diagonal is still given by the sum in
(6.4.3).
Now the Cauchy problem (6.4.1) has the exact solution
Z
u(t, x) = H(t, x, y)v(y) dvol(y).

There is no other solution. In fact, for the scalar product with a solution U of the
adjoint equation (∂t + P ∗ )U = 0 we have

∂t hu, U i = h∂t u, U i + hu, ∂t U i = hP u, U i + hu, −P ∗ U i = 0.

130 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

We can choose U as a solution for t < T which is equal to a given section of E ∗ when
t = T by the preceding existence proof, for the change of sign in front of ∂t changes
the direction of “time”. Since

hu(T, ·), U (T, ·i = hu(0, ·), U (0, ·)i

it follows that u(T, ·) is uniquely determined by the initial data v. Hence H is also
uniquelyP determined. (It would therefore have been enough to start from an approx-
µ
imation 0 uν Hν with a fixed large µ; the resulting H is then independent of µ.) In
particular it follows that
∞
X
(6.4.4) H(t, x, x) ∼ tj−n/2 hj (x), t → +0,
0

where hj ∈ C ∞ (M, E ⊗ E ∗ ). If one introduces geodesic coordinates at y ∈ M and

a corresponding synchronous frame for E, with the connection defined in Proposition
6.1.2, then the coefficients in the expansion at y are given by (6.1.15) and the recursion
formulas (6.1.14). For operators such as D+ which are entirely determined by a Rie-
mannian geometry in M , it follows from Theorem 3.3.5 that hj (y) can be expressed
as a polynomial in the Riemann curvature tensor and its covariant derivatives at y.
(The trivialization of the bundles ∧± is then given by the Levi Civita radial parallel
translation.) This will be the starting point for the discussion of the corresponding
index in the following section.
6.5. Hirzebruch’s index formula and Gilkey’s theorem. Let M be an oriented
Riemannian manifold of even dimension, and consider the operator D+ from sections
of ∧+ to ∧− defined in Section 6.3, using the decomposition of the exterior algebra by
the eigenvalues of the map τ defined in Section 6.2. Let H ± (t, x, y) be the fundamental
solutions of ∂t + D±∗ D± constructed in Section 6.4; recall that D∓ is the adjoint of
D± . In (6.3.11) we expressed the index of D+ in terms of the intersection form of the
cohomology in the middle dimension. We shall now give an analytical expression in
terms of the heat kernels:
Proposition 6.5.1. For every t > 0 we have
Z
(6.5.1) ind D = (Tr H + (t, x, x) − Tr H − (t, x, x)) dvol(x).
+

Here H ± (t, x, x) is a linear transformation in ∧±x .

Proof. Let us consider the decomposition of C ∞ (M, ∧+ ) given by the eigenspaces of
the self-adjoint operator D+∗ D+ . The spectrum is discrete, for if

kD+∗ D+ ukL2 + kukL2 ≤ 1,

then u belongs to a compact subset of L2 . In fact, there is a parametrix F such that

u = F D+∗ D+ u + Ru
 HIRZEBRUCH’S INDEX FORMULA AND GILKEY’S THEOREM 131

where R is an integral operator with C ∞ kernel, and F is continuous from L2 to

the space of sections with first derivatives in L2 , because first derivatives of F are
integrable. The set Γ+ ∞
λ of eigenfunctions with eigenvalue λ is a subset of C (M, ∧+ ),
and
L2 (M, ∧+ ) = ⊕λ Γ+λ (∧+ ).

If ϕ ∈ Γ+
λ then

ϕ = F λϕ + Rϕ = · · · = (F λ)N ϕ + (I + F λ + · · · + (F λ)N−1 )Rϕ.

When N > n/4 the kernel of F N is square integrable in each variable, so we obtain
N
(6.5.2) sup |ϕ(x)| ≤ C(1 + λ)1+ 4 kϕkL2 , ϕ ∈ Γλ .

The eigenspaces Γ− λ of D D = D+ D+∗ have similar properties and their orthogonal

−∗ −

direct sum is L2 (M, ∧− ).

Since D+∗ D+ u = λu implies D+ D+∗ (D+ u) = λD+ u, and D+ u = 0 implies u = 0
if λ 6= 0, the restriction of D+ to Γ+ −
λ is then an injective map into Γλ . In the same
+
way we see that D+∗ defines an injective map Γ− λ → Γλ , so these spaces have the same
dimension if λ 6= 0. Since Γ± ±
0 = Ker D , it follows that
X
(6.5.3) ind D+ = dim Γ+ −
0 − dim Γ0 = χ(λ)(dim Γ+ −
λ − dim Γλ ),

χ(λ) dim Γ± −tλ

P
if χ(0) = 1. We shall prove that λ converges when χ(λ) = e , t > 0,
±
R
and that the sum is equal to Tr H (t, x, x) dvol(x). This implies (6.5.1).
If ϕ ∈ Γ+
λ then (∂t + D
+∗ +
D )(e−tλ ϕ) = 0, hence
Z
e ϕ(x) = H + (t, x, y)ϕ(y) dvol(y).
−tλ

If α ∈ ∧+x it follows that

Z
e−tλ
(ϕ(x), α) = (ϕ(y), H +(t, x, y)∗α) dvol(y).

If we let ϕ run through a complete orthonormal system of eigenfunctions ϕj , with

corresponding eigenvalues λj , then Parseval’s formula gives
X
e−2tλj |(ϕj , α)|2 ≤ Ct−n kαk2 , 0 < t < 1,

and if we sum over all α in an orthonormal basis in ∧+x we get for 0 < t < 1
X X
(6.5.4) e−2tλj |ϕj (x)|2 ≤ C ′ t−n , hence e−2tλ dim Γ+λ ≤C t
′′ −n
.

The last bound follows by integrating the first over M , and it proves the convergence
of the sum in (6.5.3) when χ(λ) = e−tλ , t > 0.
132 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

If f is a finite linear combination of the eigenfunctions ϕj , then

Z X
H + (t, x, y)f (y) dvol(y) = e−tλj ϕj (x)(f, ϕj ).

Since such sections of ∧+ are dense in the continuous sections it follows in view of
(6.5.2) and (6.5.4) that for every f ∈ ∧+x
X
H + (t, x, x)f = e−tλj ϕj (x)(f, ϕj (x)).

Let fν , ν = 1, . . . , 2n−1 , be an orthonormal basis in ∧+x . If we choose f = fν , take

the scalar product with fν in ∧+x and sum, it follows that
X X
Tr H + (t, x, x) = e−tλj |(ϕj (x), fν )|2 = e−tλj |ϕj (x)|2 .
j,ν j

Integration over M gives

Z X
Tr H + (t, x, x) dvol(x) = e−tλ dim Γ+
λ.

We have an analogous formula for H − , so (6.5.1) follows from (6.5.3).

By (6.4.4) there is an asymptotic expansion

∞
n
X
+ −
Tr H (t, x, x) − Tr H (t, x, x) ∼ tj− 2 hj (x),
0

where hj is a polynomial in the components of the Riemann curvature tensor and its
covariant derivatives. The product hj (x) dvol(x) is a density on the oriented manifold
M . If we reverse the orientation, then τ is replaced by −τ , which means that the
spaces ∧+ and ∧− and therefore the kernels H + and H − are interchanged. Thus hj is
replaced by −hj , which means that we can regard hj (x) dvol(x) as a n form ωj . From
(6.5.1) it follows then that
Z
(6.5.5) ωj = 0, j < n/2,
M
Z
(6.5.6) ωn/2 = ind D+ .
M

Our aim now is to prove that ωj = 0 when j < n/2, which is a very much stronger
statement than the integral condition (6.5.5), and that ωn/2 is a polynomial in the
Pontrjagin forms. The following simple observation is crucial:
 HIRZEBRUCH’S INDEX FORMULA AND GILKEY’S THEOREM 133

Lemma 6.5.2. If the metric ds2 is replaced by λ2 ds2 where λ is a positive constant,
then ∆ is multiplied by λ−2 and ωj is multiplied by λn−2j .
Proof. In a local coordinate system gjk is replaced by λ2 gjk and g jk by λ−2 g jk , which
means that ∆ is replaced by λ−2 ∆. Now

λ−2 D+∗ D+ + ∂t = λ−2 (D+∗ D+ + ∂τ ),

if τ = λ−2 t. Hence the initial value problem for λ−2 D+∗ D+ + ∂t has the solution
Z
H + (t/λ2 , x, y)f (y) dvol(y)

where H + and dvol(y) belong to the metric ds2 . For the metric λ2 ds2 the vol-
ume element is λn dvol(y), so the new fundamental solution is H + (t/λ2 , x, y)λ−n .
The product of H ± (t/λ2 , x, x)λ−n by the new volume element λn dvol(x) is equal to
H + (t/λ2 , x, x) dvol(x), so
n n
X X
tj− 2 ωj is replaced by λn−2j tj− 2 ωj ,

which proves the lemma.

When j < n/2 the weight of ωj is positive in the sense that the power of λ in Lemma
6.5.2 is positive, and the weight is 0 when j = n/2. We shall prove a theorem of Gilkey
which states that this suffices to make the conclusions about ωj mentioned above. Of
course we also have to use that if x1 , . . . , xn is a geodesic coordinate system centered
at p, then
ωj (p) = Φj (R, R′ , . . . , R(k) ) dx1 ∧ · · · ∧ dxn ,
where Φj is a polynomial in the curvature tensor R and its covariant derivatives of
order ≤ k when x = 0. Note that the local definition of ωj makes these forms defined
for every Riemannian manifold of even dimension.
Definition 6.5.3. A q form invariant of Riemannian manifolds M of dimension n is a
function which to every such manifold assigns a q form ω on M such that for every
p ∈ M and geodesic coordinate system x1 , . . . , xn centered at p we have for some κ
′
X
(6.5.7) ω(p) = |I| = qΦI (R, . . . , R(κ) ) dxI ,

where ΦI is a polynomial in the curvature tensor and its covariant derivatives of order
≤ k at p, which is independent of M and p. The invariant is said to have weight k if
ω is multiplied by λk when the first fundamental form is multiplied by λ2 .
Let V be the Euclidean vector space Rn . We can regard the curvature tensor R as
N4 N4+j
an element of V and R(j) as an element of V , so (6.5.7) is a polynomial map
k
M 4+j
O
q
Φ : W → ∧ V, W = Wj , Wj = V.
j=0
134 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

The orthogonal group operates both in W and in ∧q V , and we denote the operation of
O by O ∗ in both cases. Since an orthogonal transformation of x1 , . . . , xn gives a new
geodesic system of coordinates centered at p, we must have

Φ(O ∗ w) = O ∗ Φ(w)

if w = (R, . . . , R(k) ) for some choice of a Riemannian metric. (Note that already the
L4
Bianchi identities show that R is not an arbitrary element in V .) Hence we have

Φ(w) = (O −1 )∗ Φ(O ∗ w)

for all such w. Taking the average of the right-hand side over the orthogonal group,
we obtain a polynomial Φ e which can still be used in (6.5.7) and has the advantage of
being equivariant, that is,

e ∗ w) = O ∗ Φ(w),
Φ(O e ∀w ∈ W.

To simplify notation we assume in what follows that Φ already has this property. It
is clear that if we split Φ into a sum of polynomials which are homogeneous in each of
the variables wj ∈ Wj , then all the terms will be equivariant.
A polynomial map Ψ(w) of degree ν in w ∈ Wj can be written in one and only one
way in the form Ψp (w, . . . , w) where Ψp (w1 , . . . , wν ) is a symmetric multilinear map
in w1 , . . . , wν ∈ Wj ; the polarization Ψp is given by
ν
Y
Ψp (w1 , . . . , wν ) = hwj , ∂/∂wiΨ(w)/ν!.
j=1
Np
Ψp defines a linear map on Wj which is equivariant if Ψ is. If we use polarization
for each Wj we may conclude that Φ is a sum of polynomials each of which is induced
by a linear equivariant map
ON
ϕ: V → ∧q V.

The fundamental theorem on O(n) invariants, Theorem D.1, gives a complete descrip-
NN
tion of such maps when q = 0. One calls a linear form ϕ on V elementary if
N = 2k and
ϕ(v1 ⊗ · · · ⊗ v2k ) = (v1 , v2 ) . . . (v2k−1 , v2k ),
that is,
X N
O
ϕ(ξ) = ξα1 α1 α2 α2 ...αk αk , ξ∈ V,
or if ϕ differs from this linear form just by a permutation of the indices.
NN
Theorem 6.5.4. Every invariant linear form V → R is a linear combination of
elementary forms; in particular there are no such forms 6= 0 unless N is even.
This is just a reformulation of Theorem D.1, so we pass to:
 HIRZEBRUCH’S INDEX FORMULA AND GILKEY’S THEOREM 135

NN
Corollary 6.5.5. If ϕ : V → ∧q V is equivariant and not identically 0, then N −q
is an even integer 2r ≥ 0, and ϕ is a linear combination of elementary maps, that is,
maps which apart from a permutation of the indices are of the form
X
ϕ(ξ) = ξα1 α1 α2 α2 ...αr αr [β1 ...βq ] dxβ1 ∧ · · · ∧ dxβq ,
α,β

where [. . . ] denotes alternation over the enclosed indices, that is, summation over all
permutations after multiplication by the sign of the permutation.
NN+q
Proof. We define an invariant form ϕ̃ : V → R by

ϕ̃(v1 ⊗ · · · ⊗ vN+q ) = (ϕ(v1 , . . . , vN ), vN+1 ∧ · · · ∧ vN+q ).

It follows from Theorem 6.5.4 that ϕ̃ = 0 unless N + q is even. If N + q is even then

ϕ̃(v1 , . . . , vN+q ) is a linear combination of elementary forms and is preserved by alter-
nation of the last q vectors followed by division by q!. This eliminates elementary forms
containing a scalar product of two vectors vN+1 , . . . , vN+q , for if they are interchanged
the sign of the permutation is changed but the term is not affected otherwise. Hence
ϕ̃ = 0 if q > N . If N = q + 2r, where r ≥ 0 is even, then ϕ̃ is a linear combination of
forms of the type

(v1 , . . . , vN+q ) 7→ (v1 , v2 ) . . . (v2r−1 , v2r )(vN+1−q , vN+1 ) . . . (vN , vN+q ).

Alternation gives the forms in the corollary.

We shall now apply the corollary to the q form invariant ω in Definition 6.5.3. When
α = (α1 , . . . , αν ) is a sequence of |α| = ν indices between 1 and n, we shall denote by
Rα the corresponding components of the ν − 4th covariant derivative of the Riemann
curvature tensor Rijkl . By an elementary monomial of degree r in R we shall mean an
expression of the form
X∗
(6.5.8) m(R) = R α1 R α2 . . . R αr ,
q

where the summation indicates alternation of precisely q indices and pairwise contrac-
tion of the others in the total index sequence α1 α2 . . . αr , where the number of indices
shall exceed q by an even number (or 0). What we have proved so far is that with
coefficients am ∈ R we have
X
(6.5.7)′ ω(p) = am m(R).

The number of possible terms is reduced if one examines the weights:

136 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

Lemma 6.5.6. The weight of m(R) is 2r + q − |α| where |α| = |α1 | + . . . |αr |.
Proof. With a fixed system of local coordinates the Christoffel symbols Γikj are mul-
tiplied by λ2 if the metric is multiplied by λ2 , but the Christoffel symbols Γik j remain
unchanged. By (2.1.13) it follows that Rijkl is multiplied by λ2 , and so are the covari-
ant derivatives. If t1 , . . . , tq are tangent vectors, then m(R) evaluated with respect to
these in a fixed coordinate system is
X∗
(6.5.8)′ Rα1 . . . Rαr g i1 i2 . . . tj11 . . . tjqq ,
q

where for each pair i1 i2 of indices to be contracted there is a factor g i1 i2 while j1 , . . . , jq

correspond to indices to be alternated. In fact, (6.5.8)′ is invariant under coordinate
changes and agrees with (6.5.8) at the center of a geodesic coordinate system. The
number of indices in the contraction is |α| − q, so replacing g by λ2 g multiplies (6.5.8)′
by λ2r λ−(|α|−q) , which proves the lemma.
Write |αi | = 4 + εP
i where εi is the number of covariant differentiations which occur
in Rαi , and let ε = εi be the total number of them. Then it follows from Lemma
6.5.6 that for the terms in (6.5.7)′ the weight of m(R) is equal to q − 2r − ε. Thus
one must alternate with respect to a larger number of indices to get a higher weight,
which explains why invariants of higher weight have a simpler structure.
Theorem 6.5.7 (Gilkey). Every q form invariant of positive weight for Riemannian
manifolds is equal to 0, and every q form invariant of weight 0 is contained in the ring
generated by the Pontrjagin forms, which all have weight 0.
Proof. The curvature tensor and its covariant derivative have the symmetry properties

(6.5.9) Rijkl = −Rjikl , Rijkl = −Rijlk , Rijkl = Rklij ,

(6.5.10) Ri[jkl] = 0, Rij[kl,m] = 0 (the Bianchi identities).

Such identities are preserved by covariant differentiation since it commutes with the
permutation group acting on the tensor product. Thus differentiation of the first
Bianchi identity gives
Rijkl,m + Riljk,m + Riklj,m = 0
if we write out all terms explicitly. The alternation of Rijkl over three arbitrary indices
is zero by the first Bianchi identity and the symmetries (6.5.9), and this is also true
for Rijkl,m . When one alternates over mij or mkl this is the second Bianchi identity.
If one alternates over jkm and notes that

Rijkl,m − Rikjl,m + Riljk,m = 0,

it follows that two times the alternation is 0, for exchanging k and j changes the sign
of the permutation. This proves the claim. By covariant differentiation we conclude
that the alternation of Rα over three of the first four or five indices is always equal to
0.
 HIRZEBRUCH’S INDEX FORMULA AND GILKEY’S THEOREM 137

If the term m(R) is not zero it follows that the alternation in (6.5.8) involves at
most two of the first four indices in each factor. Altogether we can then alternate in
at most 2r + ε indices, hence q ≤ 2r + ε, which means that the weight of m(R) is
≤ 0. This proves the first part of Gilkey’s theorem. If the weight is 0 we have also
found that q = 2r + ε and that we must alternate over all indices corresponding to
covariant differentiation and in addition two of the first four indices in each factor. If
a covariant differentiation occurs in some factor, it follows that we alternate over three
of the first five indices in it, which implies that m(R) = 0. Hence ε = 0, that is, no
covariant derivatives occur. Furthermore we have alternation with respect to precisely
two indices in each factor Rijkl . If they are the first two, we can use the symmetry
Rijkl = Rklij to replace them by the last two. If they are the middle two, we can use
that
Rijkl − Rikjl = Rilkj = (Rilkj − Riljk )/2
to replace them by the last two indices, at the expense of a factor 12 . Hence it follows
that in a geodesic coordinate system ω is a sum of products of expressions of the form
X
Ri1 i2 j1 j2 Ri2 i3 j3 j4 . . . Rik i1 j2k−1 j2k dxj1 ∧ · · · ∧ dxj2k .
i,j

With the notation in (4.1.13) we recognize this as 2k times the form

(6.5.11) Ωi1 i2 ∧ Ωi2 i3 ∧ · · · ∧ Ωik i1 .

Hence the Gilkey theorem is now a consequence of the following:

Lemma 6.5.8. The Pontrjagin forms are universal polynomials in the forms
(6.5.11).
Proof. We proved in Section 4.4 that the forms (6.5.11) are closed forms on M , lifted
to P (M ). The Pontrjagin forms are the forms of degree 4, 8, . . . in

det(δij + (2π)−1 Ωij ).

Let us now note that for n × n matrices Aij with complex coefficients we have
n
Y n
X
det(λδij − Aij ) = (λ − λj ) = λn−j (−1)j cj ,
1 0

where λj are the eigenvalues and cj the elementary symmetric functions of them. We
have for every positive integer k
n
X X
sk = λkj = Ai1 i2 Ai2 i3 . . . Aik i1 ,
j=1

for the sum is invariant under conjugation of the matrix A, and the formula is valid
for diagonal matrices. Since c1 = s1 , and by Newton’s formulas

sk = c1 sk−1 − c2 sk−2 + · · · + (−1)k kck , 2 ≤ k ≤ n,

138 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

we can for every k ≤ n express sk as a polynomial in c1 , . . . , ck and express ck as

a polynomial in s1 , . . . , sk . But these polynomial identities in the coefficients of Aij
remain valid if we replace Aij by the commuting forms Ωij , which completes the proof
of the lemma and of the Gilkey theorem.
For an oriented compact manifold M of dimension 4k we can apply the Gilkey
theorem to the forms ωj in (6.5.5), (6.5.6) which occur in the expansion
∞
X
+ −
tj−2k ωj .


Tr H (t, x, x) − H (t, x, x) dvol(x) ∼
j=0

The conclusion is that

(6.5.5)′ ωj = 0, j < 2k;

(6.5.6)′ ω2k = Lk (p1 , . . . , pk ).

Here Lk is a polynomial of weight k if pj are indeterminates of weight j, so replacing

pj by the Pontrjagin form of degree 4j gives a a form of highest degree 4k which is
equal to ω2k . Integration of (6.5.5)′ gives back (6.5.5), and integration of (6.5.6)′ gives
by (6.5.6) and (6.3.11)
Z
′′ +
(6.5.6) ind D = sign M = Lk (p1 , . . . , pk )
M

for every oriented manifold M of dimension 4k. One can now determine the coefficients
of L by specializing M to manifolds for which the signature and the Pontrjagin classes
are easy to describe, such as products of complex projective spaces. We refer to Atiyah,
Bott and Patodi [1] for the calculation of Lk and only give the result. Set
X Y Y X
(6.5.12) Lk = xj / tanh xj , where (1 + x2j ) = pk .

This should be understood as follows. To calculate Lk we take m ≥ k variables xj

and note that xj / tanh xj is a power series in x2j . Collecting the terms of degree 2k
in the product for j = 1, . . . , m gives a polynomial Lk in the elementary symmetric
functions p1 , p2 , . . . of the x21 , . . . , x2m which has weight k when pj is given the weight
j. The polynomial is independent of the choice of m ≥ k, and one verifies that it is the
only polynomial which makes (6.5.6)′′ valid for products of complex projective spaces.
Apart from the verification of this we have now proved
Theorem 6.5.9 (Hirzebruch). The signature of an oriented Riemannian manifold
M of dimension 4k is given by (6.5.6)′′ where the polynomial Lk is defined by (6.5.12)
and p1 , . . . are the Pontrjagin forms of M .
We also refer to Atiyah, Bott and Patodi [1] for an extension of Theorem 6.5.9 to the
signature operator with coefficients in a vector bundle, and for the arguments required
to go from there to the general index theorem for elliptic pseudo-differential operators.
 OPERATORS OF DIRAC TYPE 139

We have followed that paper here apart from substituting the Hadamard construction
of a parametrix for pseudo-differential operator theory. The Hadamard construction
is not only more elementary, it fits precisely with the differential geometric context.
For the operator Deo having the Euler characteristic as index, it was first proved by
Patodi [1] that a corresponding phenomenon occurs, leading to the Gauss-Bonnet-
Chern formula. A variant of his proof using ideas of supersymmetry from physics is
given in Cycon, Froese, Kirsch and Simon [1, Chapter 12]. (See page 258 for a criticism
of the methods used here which do not go all the way to a direct computation of the
coefficients in the index formula.) More recently, a direct computational proof has
been obtained for twisted Dirac operators by Bismut [1] and Getzler [1,2]. Our next
aim is to give an exposition of the methods of Getzler.
6.6. Operators of Dirac type. We started this chapter with a study of second
order metric elliptic operators, but all the geometric applications in Sections 6.3 and
6.5 concerned first order operators D such that −D∗ D is metric. (The minus sign
comes from the convention for defining the principal symbol introduced in Section 5.1,
which does not contain the factor i which is customary in pseudo-differential operator
theory.) The rest of this chapter will be devoted to a more systematic search for such
operators.
Definition 6.6.1. If M is a C ∞ manifold and E0 , E1 two C ∞ Hermitian vector bundles
on M with the same fiber dimension, then a first order differential operator

D : C ∞ (M, E0 ) → C ∞ (M, E1 )

is said to be of Dirac type if −D∗ D : C ∞ (M, E0 ) → C ∞ (M, E0 ) is a metric differential

operator.
Here the formal adjoint D∗ is defined using the hermitian metrics in E0 , E1 , and
some positive density in M ; the definition is independent of the choice of density since
it only influences terms of lower order. If σ(x, ξ) : E0x → E1x is the symbol of D at
(x, ξ) ∈ T ∗ M , then σ(x, ξ) is a linear transformation depending linearly on ξ, and the
definition means that

(6.6.1) σ(x, ξ)∗σ(x, ξ) = p(x, ξ)IE0x ,

where σ(x, ξ)∗ : E1x → E0x is the adjoint with respect to the Hermitian metrics and
p(x, ξ) is a positive definite quadratic form in Tx∗ M . Since E0x and E1x have the same
dimension, it follows from (6.6.1) that

(6.6.1)′ σ(x, ξ)σ(x, ξ)∗ = p(x, ξ)IE1x ,

so D∗ is also of Dirac type. When D is an operator of Dirac type on M , we give M

the Riemannian structure defined by the dual of the quadratic form p(x, ξ).
A Dirac type operator D : C ∞ (M, E0 ) → C ∞ (M, E1 ) is called symmetric if E0 = E1
and D∗ = D. Thus −D2 is then a metric operator. This is the context in which Dirac
140 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

originally introduced what is now known as a Dirac operator. He wanted to write the
Klein-Gordon equation

∂ 2 /∂t2 − ∂ 2 /∂x2 − ∂ 2 /∂y 2 − ∂ 2 /∂z 2 + m2

as the square of a first order operator in order to obtain a relativistically invariant

operator similar to the Schrödinger equation, and found that this could be done using a
4×4 system of first order operators. Our Dirac operators are analogous but correspond
to a positive definite metric rather than one of Lorentz signature. In part of Section
6.7 we shall avoid making assumptions on the signature in order to cover the original
case also.
For a general operator D : C ∞ (M, E0 ) → C ∞ (M, E1 ) of Dirac type, a symmetric
operator of Dirac type C ∞ (E0 ⊕ E1 ) → C ∞ (E0 ⊕ E1 ) is defined by
   
0 D∗ 0 σ(x, ξ)∗
(6.6.2) ; the symbol is .
D 0 σ(x, ξ) 0
It maps sections of E0 ⊂ E0 ⊕E1 to sections of E1 ⊂ E0 ⊕E1 and vice versa. Conversely,
a symmetric Dirac type operator with this property in C ∞ (M, E), E = E0 ⊕ E1 , is
always obtained from a Dirac type operator C ∞ (M, E0 ) → C ∞ (M, E1 ) as in (6.6.2).
If σ(x, ξ) is the principal symbol of a symmetric Dirac type operator in C ∞ (M, E),
then

(6.6.3) σ(x, ξ) = σ(x, ξ)∗, σ(x, ξ)2 = p(x, ξ)IEx .

The first step in the study of (symmetric) Dirac operators is to determine matrices
σ(x, ξ) depending linearly in ξ, which satisfy (6.6.3) for a fixed x. This leads to the
definition of Clifford algebras and spinors, which will be studied in Section 6.7. We can
then define (twisted) Dirac operators in Section 6.8, where we also prove an analogue
of the Weitzenböck formula due to Lichnerowicz. Section 6.9 is an analytical interlude
devoted to the classical Mehler formula giving the heat kernel for the harmonic oscil-
lator explicitly, and to some extensions needed here. We are then prepared to prove
the local index theorem for Dirac operators in Section 6.10.
6.7. Clifford and spinor algebra. Let V be a real vector space, q a quadratic
form in V , E a complex vector space, and σ : V → End(E) = L(E, E) a linear map
such that

(6.7.1) σ(v)2 = q(v)IE ,

as in (6.6.3). Every linear map L

σ : V → End(E) can be uniquely extended to a linear
∞
map σ̃ from the tensor algebra 0 ⊗k V to End(E) with

σ̃(v1 ⊗ · · · ⊗ vk ) = σ(v1 ) . . . σ(vk ), v1 , . . . , vk ∈ V,

for the right-hand side is a multilinear function of v1 , . . . , vk . The condition (6.7.1)

means that for v ∈ V and arbitrary tensors t1 , t2

σ̃(t1 ⊗ v ⊗ v ⊗ t2 ) − σ̃(t1 ⊗ q(v) ⊗ t2 ) = 0,

 CLIFFORD AND SPINOR ALGEBRA 141

that is, σ̃ vanishes on the two sided ideal I generated in the tensor algebra by all
elements of the form v ⊗v −q(v) · 1 with v ∈ V , so σ̃ induces an algebra homomorphism
from the quotient algebra to End(E). Note that with the notation q also for the
polarized form

(v1 ± v2 ) ⊗ (v1 ± v2 ) − q(v1 ± v2 ) · 1 ∈ I =⇒ v1 ⊗ v2 + v2 ⊗ v1 − 2q(v1 , v2 ) · 1 ∈ I.

If we choose a basis ε1 , . . . , εn in V which diagonalizes q,

n
X n
X
(6.7.2) q( ξj εj ) = qj ξj2 , ξ ∈ Rn ,
1 1

it follows that I is generated by all elements of the form

(6.7.3) εj ⊗ εk + εk ⊗ εj , j 6= k; εj ⊗ εj − qj · 1; j, k = 1, . . . , n.

Every element in the tensor algebra is a linear combination of tensor products of

ε1 , . . . , εn , so using (6.7.3) we see that modulo I it is congruent to an element of the
form
X
(6.7.4) ai1 ,...,ij εi1 ⊗ · · · ⊗ εij , i1 < i2 < · · · < ij .
j≤n

Here ai1 ,...,ij is a linear form on the tensor algebra vanishing in I, defined by

n
qiµi ,
Y
ι
ai1 ,...,ij (εk1 ⊗ · · · ⊗ εkν ) = (−1)
i=1

if the indices i1 , . . . , ij occur 1 + 2µi1 , . . . , 1 + 2µij times among k1 , . . . , kν , the other

indices i ≤ n occur 2µi times, and ι is the number of index pairs in k1 , . . . , kν which
occur in the wrong order. Otherwise ai1 ,...,ij (εk1 ⊗ · · · ⊗ εkν ) = 0. This follows since
the definition makes the form ai1 ,...,ij vanish on I while it gives the desired coefficient
for tensors of the form (6.7.4).
Definition 6.7.1. If V is a real vector space of dimension n and q a quadratic form
in V , then
L∞ the Clifford algebra Cl(V, q) is defined to be the quotient of the full tensor
algebra 0 ⊗k V by the twosided ideal I generated by the elements v ⊗ v − q(v) · 1,
v ∈ V . Elements in V are identified with their images in Cl(V, q), and the product of
two elements x and y in Cl(V, q) is denoted x · y.
We have already proved most of the following result:
Theorem 6.7.2. If ε1 , . . . , εn is a basis for V giving q the form (6.7.2), then every
element in the associative Clifford algebra Cl(V, q) can be written in one and only one
way in the form (6.7.4), with coefficients in R, so the dimension of Cl(V, q) over R is
2n . The images Cl0 (V, q) and Cl1 (V, q) in Cl(V, q) of the tensors of even (odd) rank
142 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

are represented by sums (6.7.4) with j even (odd), so the dimensions are 2n−1 . We
have
Cli (V, q) · Clj (V, q) ⊂ Cli+j mod 2 (V, q),
which gives a Z2 grading of the algebra Cl(V, q). The set Cl[j] (V, q) of elements of the
form (6.7.4) with j fixed ∈ {0, . . . , n} is a linear subspace of Cl(V, q), equal to the linear
span of all images of elements of the form
X
(6.7.5) sgn πvπ(1) ⊗ · · · ⊗ vπ(j) , v1 , . . . , vj ∈ V,
π

where π is any permutation of 1, . . . , j; thus it is independent of the choice of diago-

nalizing basis ε1 , . . . , εn . We have
M M
Cl0 (V, q) = Cl[j] (V, q), Cl1 (V, q) = Cl[j] (V, q).
j even j odd

There is a unique linear map Cl(V, q) ∋ x 7→ t x ∈ Cl(V, q) (transposition) such that

t
(x · y) = t y · t x and t x = x if x ∈ V . The constant term Q(x) in t x · x is equal to
the constant term in x · t x and defines invariantly a quadratic form Q on Cl(V, q) such
that Q(1) = 1 and

Q(v) = q(v), Q(x) = Q(t x), Q(v · x) = Q(x · v) = q(v)Q(x),

if v ∈ V , x ∈ Cl(V, q). If x is the class of (6.7.4) and (6.7.2) holds, then

X
Q(x) = qi1 . . . qij a2i1 ...ij .
j≤n

Proof. The transposition is inherited from the maps v1 ⊗ · · · ⊗ vk → vk ⊗ · · · ⊗ v1 in

the tensor algebra, for they vanish on I. Only the statement about Cl[j] remains to
be verified. If j is fixed in (6.7.4) we can extend the sum to all indices i1 , . . . , ij ∈
{1, . . . , n} provided that we divide by j! and extend the definition of the coefficients in
a skew symmetric way. Hence every such element is a linearP combination of elements
n
of the form (6.7.5). On the other hand, suppose that vi = k=1 vi,k εk , i = 1, . . . , j.
Then the tensor (6.7.5) can be written
X X

sgn πvπ(1),k1 . . . vπ(j),kj εk1 ⊗ · · · ⊗ εkj ,
k1 ,...,kj π

and since the sum over π is skew symmetric in k1 , . . . , kj , we have an element of the
form (6.7.4) with j fixed.
If x is the class of the form (6.7.4) with P a basis ε1 , . . . , εn satisfying (6.7.2), then
t t
x · x and x · x both have the constant term a2i1 ...ij qi1 . . . qij , for εi1 · · · εij · εk1 · · · εkl
can only be a constant if there is some ν ∈ {i1 , . . . , ij } ∩ {k1 , . . . , kl } with qν = 0 or
{i1 , . . . , ij } = {k1 , . . . , kl }. The definition of Q is obviously invariant since it makes
 CLIFFORD AND SPINOR ALGEBRA 143

no reference to the choice of a basis, and the other statements are obvious since for
example t (v · x) · (v · x) = t x · v · v · x.
Note that Cl[j] · Cl[k] is not contained in Cl[j+k] but may also have components in
Cl[i] with i < j + k.
Exercise 6.7.1. Show that when v1 , . . . , v4 ∈ V we have with all products taken in
Cl(V, q)
[v1 · v2 − v2 · v1 , v3 · v4 − v4 · v3 ] = 4 q(v1 , v4 )(v2 · v3 − v3 · v2 ) − q(v1 , v3 )(v2 · v4 − v4 · v2 )


+ q(v2 , v3 )(v1 · v4 − v4 · v1 ) − q(v2 , v4 )(v1 · v3 − v3 · v1 ) .
Thus Cl[2] (V, q) is a Lie algebra. Deduce that if q is positive definite and ε1 , . . . , εn is
an orthonormal basis, then
X X X
[ aij εi · εj , bk,l εk · εl ] = 4 [a, b]ij εi · εj
i,j k,l i,j

if a and b are skew symmetric matrices. (Compare with the Lie algebra of SO(n).)
When q = 0, the Clifford algebra becomes the exterior algebra ∧∗ V on V . However,
the case of interest to us here is primarily the case where q is positive definite. The
negative definite case is often preferred but the difference is not important in the
arguments below where we introduce the complexification of Cl, for all non-degenerate
quadratic forms in a complex vector space are equivalent. However, before doing so
we introduce an element in the Clifford algebra which will be very important later on.
Theorem 6.7.3. If V is oriented and q is non-degenerate, then
n q
Y
(6.7.6) ν = ε1 · · · εn / |qj |
1
is independent of the choice of positively oriented diagonalizing basis ε1 , . . . , εn for V .
We have
n
1
Y
(6.7.7) ν 2 = (−1) 2 n(n−1) sgn qj , νu = (−1)j(n−1) uν, if u ∈ Cl[j] (V, q).
1

Proof. As in the proof of Theorem 6.7.2 we have

n q
1 X Y
ν= sgn πεπ(1) · · · επ(n) / |qj |.
n! π 1
P
If ε̃i = ci,k εk is another diagonalizing positively oriented basis, then det(cik ) > 0
and
n q
Y X
n! |q̃j |ν̃ = sgn π ε̃π(1) · · · ε̃π(n)
1 π
X X
= sgn πcπ(1),k1 . . . cπ(n),kn εk1 · · · εkn
k1 ,...,kn π
X
= det(ci,k ) sgn(k1 , . . . , kn )εk1 · · · εkn
144 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

where all k1 , . . . , kn are different in the last sum. Now we have

ξ˜j ε̃j ) = q̃j ξ˜j2 , ξk = ξ˜j cj,k .

X X X X X
qj ξj2 = q( ξj εj ) = q(
Qn Qn
Hence 1 q̃j = 1 qj | det(ci,k )|2 which proves that
n q
Y n q
Y
det(ci,k ) |qj | = |q̃j |,
1 1

since the determinant is positive. Thus ν̃ = ν. The proof of (6.7.7) is straightforward:

n
Y n
Y
2
ν = ε1 · · · εn ε1 · · · εn |qj | = (−1)n−1+···+1 ε21 · · · ε2n / |qj |
1 1
n
1
Y qj
= (−1) 2 n(n−1) ,
1
|qj |

n q
Y
|qj |νεj = (−1)n−j ε1 · · · εj−1 qj εj+1 · · · εn ,
1
n q
Y
|qj |εj ν = (−1)j−1 ε1 · · · εj−1 qj εj+1 · · · εn .
1

Hence νεj = (−1)n−1 εj ν, which immediately gives the second part of (6.7.7).
A complete description of the real Clifford algebras is fairly complicated (see Atiyah,
Bott and Shapiro [1]) so we content ourselves with discussing the complexification
ClC (V, q) obtained by allowing the coefficients in (6.7.4) to be complex, or equivalently,
replacing V by the complexification VC in the definition by the tensor algebra. All non-
degenerate quadratic forms in a complex vector space are equivalent under complex
linear coordinate transformations, and we shall only discuss ClC (n) = ClC (Rn , e)
where e is the Euclidean quadratic form in Rn .
Theorem 6.7.4. There are complex algebra isomorphisms
k k k
(6.7.8) ClC (2k) ∼
= End(C2 ), ClC (2k + 1) ∼
= End(C2 ) ⊕ End(C2 ).

Proof. We start with dimensions 1 and 2. The algebra ClC (1) consists of all (α, β) ∈ C2
with componentwise addition and

(α, β)(α′ , β ′ ) = (αα′ + ββ ′ , αβ ′ + βα′ ).

Since
αα′ + ββ ′ ± (αβ ′ + βα′ ) = (α ± β)(α′ ± β ′ ),
 CLIFFORD AND SPINOR ALGEBRA 145

the map (α, β) → (α+β, α−β) changes the operations in ClC (1) to the coordinatewise
operations in C ⊕ C.
To determine ClC (2) we define a linear map σ : R2 → End(C2 ) by
   
1 0 0 i
σ(ε1 ) = , σ(ε2 ) = ,
0 −1 −i 0

where ε1 , ε2 are the basis vectors in R2 . Since

 
2 2 0 i
σ(ε1 ) = σ(ε2 ) = I, σ(ε1 )σ(ε2 ) = −σ(ε2 )σ(ε1 ) = ,
i 0

it follows that σ extends to a bijective homomorphism ClC (2) → End(C2 ).

For larger n we shall obtain (6.7.8) inductively if we prove that

(6.7.9) ClC (n + 2) ∼
= ClC (n) ⊗ ClC (2), n ≥ 1.

To prove (6.7.9) we define a linear map σ : Rn+2 → ClC (n) ⊗ ClC (2) by

iεj ⊗ εn+1 · εn+2 , if j ≤ n,


σ(εj ) =
1 ⊗ εj , if j = n + 1, n + 2.

Since σ(εj )2 = 1 ⊗ 1 for j ≤ n + 2 and σ(εj )σ(εk ) + σ(εk )σ(εj ) = 0 if j 6= k, the map
σ induces a homomorphism σ̃ : ClC (n + 2) → ClC (n) ⊗ ClC (2). The range of σ̃ is a
subalgebra containing 1 ⊗ ClC (2) and ClC (n) ⊗ 1, so σ̃ is surjective, henced bijective
because the dimensions agree. Now the isomorphisms (6.7.8) follow inductively since
for arbitrary (complex) vector spaces E and F we have

(End E) ⊗ (End F ) ∼
= End(E ⊗ F ).

In fact, End E ∼ = E ⊗ E ∗ , End F ∼= F ⊗ F ∗ , End(E ⊗ F ) ∼

= E ⊗ F ⊗ E ∗ ⊗ F ∗ , and the
resulting isomorphism ι respects the product structure for if S1 , S2 ∈ End(E), T1 , T2 ∈
End F , then ι(S1 ⊗ T1 ) times ι(S2 ⊗ T2 ) and ι(S1 S2 ⊗ T1 T2 ) both map e ⊗ f ∈ E ⊗ F
to (S1 S2 e) ⊗ (T1 T2 f ). The proof is complete.
If we identify R2k with Ck , we get a natural isomorphism

(6.7.10) µ : ClC (2k) ∼

= End(∧∗ Ck ),

where ∧∗ Ck denotes the full complex exterior algebra over Ck . (Since ∧∗ Ck has
complex dimension 2k , the two sides of (6.7.10) are isomorphic by Theorem 6.7.4.)
To define the isomorphism we must recall a construction made in Section 6.2. For
w ∈ W = Ck we denote by Tw the exterior multiplication Tw u = w ∧ u, u ∈ ∧∗ W , and
by Tw∗ we denote the adjoint with respect to the natural hermitian scalar product in
∧∗ W . If w is the basis vector e1 ∈ W , then

Tw u = e1 ∧ u0 , if u = u0 + e1 ∧ u1 ,
146 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

where u0 , u1 are in the subalgebra generated by e2 , . . . , ek . If u is of degree ν and

v = v0 + e1 ∧ v1 is of degree ν + 1, then

(Tw u, v) = (e1 ∧ u0 , v0 + e1 ∧ v1 ) = (e1 ∧ u0 , e1 ∧ v1 ) = (u0 , v1 ),

which means that Tw∗ v = v1 . Hence

kTw vk2 + kTw∗ vk2 = ke1 ∧ v0 k2 + kv1 k2 = kv0 k2 + kv1 k2 = kvk2 .

By unitary invariance and homogeneity we conclude that

Tw∗ Tw + Tw Tw∗ = |w|2 I, w ∈ W,

where I is the identity in ∧∗ W , and |w|2 is the square of the norm of w as an element
in Ck , or equivalently as an element in R2k . Since Tw2 = 0 and Tw∗ 2 = 0, it follows that

(6.7.11) µ(w)2 = |w|2 I, if µ(w) = Tw + Tw∗ .

It is clear that µ is real linear, so we can extend µ first to an algebra homomorphism

Cl(2k) → End(∧∗ Ck ), and then to a complex algebra homomorphism (6.7.10). It is
k
actually an isomorphism, for both sides are isomorphic to End(C2 ), by Theorem 6.7.4,
and the kernel of µ must vanish by the following well known lemma:
Lemma 6.7.5. If E is a vector space and J is a two sided ideal in End(E), then
J = {0} or J = End(E).
Proof. Let E = CN and T ∈ J, rank T = r > 0. The product AT B ∈ J consists of
an invertible matrix T1 in the upper left r × r corner with zeros elsewhere, if A, B are
non-singular, B maps the last N − r basis vectors to the kernel of T and A maps the
range to the plane spanned by the first r basis vectors. Multiplying once more to the
left we make T1 equal to the identity. Adding such matrices for different choices of
basis vectors we conclude that the identity is in J, so J = End(E).
Summing up, we have proved:
Theorem 6.7.6. The complex extension of µ, defined by (6.7.11) gives a natural iso-
morphism (6.7.10).
Recall that our purpose is to find linear maps V → End(E) satisfying (6.7.1). We
have found that the existence of such a map means precisely that E is a Cl(V, q) module.
Theorem 6.7.6 gives an example, for the map µ makes ∧∗ (Ck ) a Cl(R2k , e) module. To
define an associated vector bundle on a Riemannian manifold we would need to have
a representation of O(2k) on ∧∗ (Ck ). To define the appropriate representations we
digress to discuss the relation between the orthogonal group and the Clifford algebra.
Proposition 6.7.7. If q is a positive definite quadratic form in the real vector space
R, then

Pin(V, q) = {v1 · · · vj ; vi ∈ V, q(vi ) = 1, for i = 1, . . . , j} ⊂ Cl(V, q),

Spin(V, q) = {v1 · · · v2j ; vi ∈ V, q(vi ) = 1, for i = 1, . . . , 2j} = Pin(V, q) ∩ Cl0 (V, q),
 CLIFFORD AND SPINOR ALGEBRA 147

are multiplicative groups.

Proof. The transpose vj · · · v1 of v1 · · · vj is an inverse.
If v ∈ V , q(v) = 1, then

v · x · v = v · (−v · x + 2q(v, x)) = −x + 2q(v, x)v, x ∈ V.

Now x − 2q(v, x)v is the orthogonal reflection of x in the plane orthogonal to v, so

x 7→ v · x · v is this reflection followed by reflection in the origin. If we define

u∗ = (−1)j t u if u ∈ Clj (V, q), j = 0, 1,

thus (v1 · · · vl )∗ = (−1)l vl · · · v1 , v1 , . . . , vl ∈ V , it follows if u ∈ Pin(V, q) that

(6.7.12) V ∋ x 7→ u · x · u∗ , x ∈ V,

is a product of orthogonal reflections in V , and it preserves orientation if and only if

u ∈ Spin(V, q). Hence (6.7.12) defines a homomorphism

τ : Pin(V, q) → O(V, q),

such that Spin(V, q) is the inverse image of SO(V, q). The maps Pin(V, q) → O(V, q)
and Spin(V, q) → SO(V, q) are surjective by the following lemma, which is very close
to Exercise 1.4.3:
Lemma 6.7.8. Every element in O(n) is a product of at most n orthogonal reflections
in Rn .
Proof. We can extend an orthogonal transformation O in Rn to a unitary transforma-
tion in Cn . The projection in Rn of an eigenvector in Cn is an invariant subspace V1
for O of dimension 1 or 2, and the orthogonal space V1⊥ is also invariant since

(OV1⊥ , V1 ) = (V1⊥ , O −1 V1 ) = 0.

Hence it suffices to prove the lemma in dimension 1 and dimension 2. In the first case
an orthogonal transformation is a reflection or the identity, in the second case it is a
reflection or a rotation by an angle θ and therefore the product of reflections in two
lines with an angle θ/2 between them.
Theorem 6.7.9. If q is a positive definite quadratic form in V , then the homomor-
phisms τ : Pin(V, q) → O(V, q) and τ : Spin(V, q) → SO(V, q) are surjective with
kernel {±1}. Both Pin(V, q) and Spin(V, q) are compact subsets of Cl(V, q), Spin(V, q)
is arcwise connected and so is Pin(V, q) \ Spin(V, q). If dim V ≥ 3 then Spin(V, q) is
simply connected.
Proof. We have already proved surjectivity, and it is clear that ±1 is in the kernel.
Suppose that u ∈ Pin(V, q) and that u · x · u∗ = x, x ∈ V . Then u ∈ Spin(V, q) since
τ (u) preserves the orientation, so u∗ · u = 1, hence u · x = x · u and

x · u · x = q(x)u.
148 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

With an orthonormal basis ε1 , . . . , εn in V , we can write (see (6.7.4)

X
u= ai1 ,...,ij εi1 · · · εij , i1 < · · · < ij , j even.

If we take x = εl and note that εl · εi1 · · · εij · εl = ±εi1 · · · εij with the minus sign when
l ∈ {i1 , . . . , ij }, we conclude that only a term with j = 0 can occur. This means that
u is a real number, and since uxu = x, x ∈ V , we have u = ±1.
The compactness follows from the bound for the number of reflections in Lemma
6.7.8. To prove that Spin(V, g) is connected it suffices to note that v1 · · · v2k , where
vj ∈ V and q(vj ) = 1, is connected to v1 · v1 · · · v1 = 1 if each vj is connected to v1 in
the connected unit sphere in V .
What remains is to prove that Spin(V, q) is simply connected if dim V ≥ 3. In
the proof we shall need that τ is a local homeomorphism. It suffices to prove that at
±1. Take disjoint compact neighborhoods U± of ±1 in Spin(V, q) with U− = −1 · U+
and, using the compactness, a compact neighborhood U of the identity in SO(V, q)
such that τ −1 U ⊂ U+ ∪ U− . By the results already proved it follows that τ maps the
′
neighborhoods U± = U± ∩ τ −1 U of ±1 bijectively, hence homeomorphically, on U . For
the proof we shall also need the following elementary lemma:
Lemma 6.7.10. Every closed loop in SO(n), n ≥ 2, is homotopic to a loop in SO(2),
embedded in SO(n) as SO(2) × In−2 .
Proof. Let R/Z ∋ t 7→ O(t) ∈ SO(n) be the given loop, n ≥ 3. If εn = (0, . . . , 0, 1) is
left fixed by O(t) for every t, then O(t) ∈ SO(n − 1) × 1, and the lemma follows by
induction from lower dimensions. In the general case we first regularize O(t) to a C 1
map. Then the curve {O(t)εn ; t ∈ R/Z} is of measure 0 in S n−1 , so we can choose
ξ ∈ S n−1 such that ±ξ ∈ / {O(t); t ∈ R/Z}. Let O1 ∈ SO(n) map ξ to εn . Our loop is
homotopic to the loop t 7→ O1 O(t) = O(t),
e and O(t)ε
e n 6= ±εn for every t. Hence

O(t)ε
e n = εn cos θ(t) + ξ(t) sin θ(t)

with uniquely determined continuous θ(t) ∈ (0, π) and ξ(t) ∈ S n−1 orthogonal to
εn .Denote by O2 (t, s) the rotation by the angle −sθ(t) in the εn ξ(t) plane. Then

(R/Z) × [0, 1] ∋ (t, s) 7→ O2 (t, s)O(t)

e

is a homotopy connecting the loop t 7→ O(t)

e to the loop t 7→ O2 (t, 1)O(t)
e which leaves
εn fixed. As observed at the beginning of the proof, the lemma is then proved by
induction.
End of proof of Theorem 6.7.9. Let R/Z ∋ t 7→ x(t) ∈ Spin(Rn , e) be a closed loop.
Then t 7→ τ (x(t)) is a closed loop in SO(n), so by Lemma 6.7.10 there is a homotopy
to a loop in t 7→ O(t) × In−2 where O(t) ∈ SO(2) consists of rotation in R2 by the
angle 2πkt. We can lift the homotopy to one in Spin(Rn , e) connecting the given loop
to the loop
t 7→ (1, 0, . . . , 0) · (cos(πkt), − sin(πkt), 0, . . . , 0).
 CLIFFORD AND SPINOR ALGEBRA 149

Since it is closed it follows that k is even. When n ≥ 3 a homotopy to a constant loop

is given by

(t, s) 7→ (1, 0, . . . , 0) · (cos( 21 πs) cos(πkt), − cos( 12 πs) sin(πkt), sin( 21 πs), 0, . . . , 0)

where t ∈ R/Z, s ∈ [0, 1]. Hence Spin(Rn , e) is simply connected.

Remarks. 1. Comparison of the result with Exercise 1.4.6 shows that Spin(R3 , e) ∼ =
SU(2). We could also have proved that Spin(Rn , e) is simply connected starting from
this fact.
2. Any closed loop γ : R/Z ∋ t 7→ O(t) ∈ SO(n), n ≥ 3, can be lifted to an arc
[0, 1] ∋ t 7→ x(t) ∈ Spin(Rn , e), and x(1) = ±x(0) by Theorem 6.7.9. If x(1) = x(0)
we have a closed loop in Spin(Rn , e), so it is homotopic to a point, hence γ is trivial.
However, if x(1) = −x(0) we do not have a closed loop in Spin(Rn , e), and this remains
true for the lifting of any loop homotopic to γ. This proves that the fundamental
group of SO(n) is Z2 , with the element 6= 0 represented by the projection of any arc in
Spin(Rn , e) connecting 1 to −1. The proof of Theorem 6.7.9 would have been shorter
if we had assumed the fundamental group of SO(n) known, but we have chosen to
determine it at the same time.
3. Since SO(2) ∼= S(1) ∼
= Spin(R2 , e), the fundamental group is equal to Z and the
correspondence is given by the winding number of a loop, when n = 2.
Using Exercise 1.4.5 it is easy to see that Spin(V, q) and Pin(V, q) are analytic
manifolds in Cl(V, q). However, we prefer to give a direct proof which also gives an
explicit correspondence between the Lie algebras of Spin(V, q) and SO(V, q), which of
course are isomorphic since the groups are locally isomorphic.
Theorem 6.7.11. If A = (ajk ) is a real skew symmetric n × n matrix, and ε1 , . . . , εn
is an orthonormal basis in V , then
n
X
(6.7.13) exp( ajk εj · εk ),
j,k=1

defined by the Taylor expansion which converges in Cl(V, q), is an element in Spin(V, q).
The image in SO(V, q) has the matrix exp(4A) in the basis ε1 , . . . , εn . Thus (6.7.13)
gives a bijection of a neighborhood of 0 in the space of skew symmetric matrices on a
neighborhood of the identity in Spin(V, q), which is therefore an analytic manifold with
tangent space Cl[2] (V, q) at the identity, and
n
X
so(n) ∋ 4(ajk )nj,k=1 7→ ajk ej ek ∈ Cl[2] (V, q)
j,k=1

is the isomorphism of Lie algebras corresponding to the local isomorphism of

SO(V, q), identified with SO(n), and Spin(V, q).
Note that this explains Exercise 6.7.1, including the factor 4.
150 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

Proof. Choose
P∞ a norm k·k in Cl(V, q) such that kx·yk ≤ kxkkyk. The formal expansion
exp x = 0 xν /ν! converges for every x ∈ Cl(V, q), for the norm of the term is bounded
by kxkν /ν!. The sum is an analytic function of x, and we obtain exp x · exp(−x) =
exp 0 = 1 by rearranging the terms in the product. For small t ∈ R let
n
Y q
P (t) = ·Pij (t), Pij (t) = εi · ( 1 − a2ij t2 εi + aij tεj ),
i,j=1

with the product taken say in the lexicographical order. We have Pjj = 1, Pij (t) ∈
Spin(V, q) for any i, j, and
n
X
Pij (t) = 1 + aij tεi · εj + O(t2 ), hence P (t) = 1 + tx + O(t2 ), x = aij εi · εj ,
i,j=1

which implies that P (1/ν)ν → exp x as ν → ∞, for

kP (1/ν)ν − (1 + x/ν)ν k ≤ (1 + kxk/ν)ν ((1 + O(1/ν 2 ))ν − 1) → 0,

ν
X
ν
(1 + x/ν) = xµ (1 − 1/ν) . . . (1 − (µ − 1)/ν)/µ! → exp x.
0

Hence exp x ∈ Spin(V, q). To find the corresponding orthogonal transformation we

note that by (6.7.12)

 0,
 if k 6= i, k 6= j
∗ 2
τ (Pij (t))εk = Pij (t)εk Pij (t) = εk + −2akj tεj + O(t ), if k = i 6= j , hence

2aik tεi + O(t2 ), if k = j 6= i

X X X
τ (P (t))εk = εk − 2 akj tεj + 2 aik tεi + O(t2 ) = εk + 4t aik εi + O(t2 ).
j i i

ν
This implies that τ (P (1/ν)P ) = (τ (P (1/ν)))ν → e4A where A is the skew symmetric
operator in V with Aεk = i aik εi , that is, with matrix (ajk ) in the basis ε1 , . . . , εn .
The proof is complete.
From now on we shall use the abbreviated notation Pin(n) and Spin(n) for
Pin(Rn , e) and Spin(Rn , e). The composition
µ
(6.7.14) ̺ : Pin(2k) → Cl(2k) → ClC (2k) → End(∧∗ Ck ))

with µ as in (6.7.10) is a representation since it is a homomorphism mapping 1 to

the identity. It is not quite the desired representation of O(2k), but just a projective
representation, determined up to a factor ±1; we shall come back to this point in
Section 6.8.
 CLIFFORD AND SPINOR ALGEBRA 151

Proposition 6.7.12. The representation ̺ of Pin(2k) defined by (6.7.14) is unitary

and irreducible.
Proof. That ̺ is unitary follows from (6.7.11), for µ(v) is self-adjoint with µ(v)2 = 1
if v ∈ V and |v| = 1. Irreducibility is a consequence of the fact that Pin(2k) generates
the algebra Cl(2k), for µ is an isomorphism.
The map µ(v) in (6.7.11) takes forms of even (odd) degree to forms of odd (even)
degree. Hence the degree of µ(z)w has the same parity as w ∈ ∧∗ (Ck ) if z ∈ Cl0 (2k)
while it is opposite if z ∈ Cl1C (2k). This means that the restriction of ̺ to Spin(2k)
splits into the direct sum of two representations

(6.7.15) D±
1 : Spin(2k) → Aut(S± (2k)), S+ (2k) = ∧even (Ck ), S− (2k) = ∧odd (Ck ),
2

called the half-spin representations. One calls (6.7.14) the spin representation.
Proposition 6.7.13. The half-spin representations are unitary and irreducible.
Proof. If W ⊂ S+ (2k) is an invariant subspace for D+ ∗ k
1 , then W ⊕ (̺(ε1 )W ) ⊂ ∧ (C )
2
is an invariant subspace for ̺, so W = {0} or W = S+ (2k) by Proposition 6.7.11,
which also shows that the representations are unitary.
In the odd dimensional case we can use the isomorphism ClC (2k − 1) → Cl0C (2k)
mapping v ∈ R2k−1 to iv · ε2k , where ε2k is the last basis vector in R2k . This means
[j] 2
that the elements in CC (2k − 1) are multiplied by ij , followed by right multiplication
by ε2k if j is odd. Hence we get an inclusion Spin(2k − 1) → Spin(2k), so (6.7.15) gives
a representation
+
D
1
2
Spin(2k − 1) → Spin(2k) −→ Aut(S+ (2k)),
which is also denoted by D+ −
1 . Similarly we get a representation D 1 in S− (2k), but
2 2
since
+
D−1 µ(ε2k ) = µ(ε2k )D 1 on S+ (2k),
2 2

it is an equivalent representation.
For reasons analogous to Proposition 6.5.1 we shall have to calculate the difference
between the traces of an endomorphism K+ of S+ (2k) and one K− of S− (2k). We
combine them to
 
K+ 0
K= ∈ End(S(2k)) = End(∧∗ (Ck )) ∼= ClC (2k),
0 K−

and we write

(6.7.16) Str K = Tr K+ − Tr K− .

One calls Str K the supertrace of K; it is defined for any endomorphism in a space S
with a given decomposition S = S+ ⊕ S− .
152 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

Theorem 6.7.14 (The Berezin-Patodi formula). If z ∈ ClC (2k) and ν is the

[2k]
“volume element” ε1 · · · ε2k (as in (6.7.6)), then the component of z in ClC is
((2i)−k Str µ(z))ν.
Before the proof we shall discuss some properties of ν in addition to those given in
Theorem 6.7.3. In doing so we must make our identification of R2k with Ck explicit.
We shall use the identification

R2k ∋ (ξ1 , . . . , ξ2k ) 7→ (ξ1 + iξ2 , . . . , ξ2k−1 + iξ2k ) ∈ Ck

which is customary in complex analysis.

Lemma 6.7.15. The center of Spin(2k) is equal to {±1, ±ν}. The restriction of ̺(ν)
to S± (2k) is equal to ±ik .
Proof. The image τ (x) in SO(2k) of an element x in the center of Spin(2k) commutes
with every O ∈ SO(2k). For every ω ∈ S 2k−1 it follows that O(τ (x)ω) = τ (x)Oω =
τ (x)ω, if O ∈ SO(2k) leaves ω fixed. Hence τ (x)ω = ±ω, and it follows that τ (x) is
± the identity. By Theorem 6.7.9 we conclude that there are at most four elements in
the center. From the second part of (6.7.7) it follows that νu = uν if u ∈ Spin(2k), so
±1, ±ν are in the center and it can contain no other elements.
Since ν ∈ Spin(2k) we know that S± (2k) is invariant under ̺(ν), and since the
restriction of ̺ to Spin(2k) is an irreducible representation on S± (2k) commuting with
̺(ν), it follows that ̺(ν) is a constant in each of these spaces. From (6.7.7) we know
that ̺(ν)2 = ̺(ν 2 ) = (−1)k , so the constant values must be ±ik . By (6.7.7) we have
ε1 · ν = −ν · ε1 , hence ̺(ε1 )̺(ν) = −̺(ν)̺(ε1 ) which proves that we have opposite
signs in S± (2k). To find the sign in S+ (2k) it suffices to calculate ̺(ν)1. Recall that
ε1 , . . . , ε2k are the basis vectors in R2k , and that we have identified ε2j−1 with ej and
ε2j with iej , if e1 , . . . , ek are the basis vectors in Ck . By induction for decreasing j it
follows from (6.7.11) that

̺(ε2j · · · ε2k )1 = ik−j+1 ej ∧ · · · ∧ ek ,

for Te∗j−1 annihilates the right-hand side. Starting from the case j = 1 we obtain by
induction for increasing j

̺(ε2j−1 · · · ε1 · ε2 · · · ε2k )1 = ik ej+1 ∧ · · · ∧ ek ,

for Tej+1 annihilates the right hand side while Te∗j+1 removes the factor ej+1 . When
j = k we obtain ̺(ν)1 = ik , for

ε2k−1 · · · ε1 · ε2 · · · ε2k = ε1 · ε2 · ε3 · ε4 · · · ε2k−1 · ε2k

since ε1 · ε2 , ε3 · ε4 , . . . commute with the other factors which allows us to to move them
out to the left starting from the middle. The lemma is proved.
Proof of Theorem 6.7.14. From Lemma 6.7.15 it follows that

(6.7.17) ik Str µ(z) = Tr(µ(ν)µ(z)) = Tr µ(νz).

 CLIFFORD AND SPINOR ANALYSIS 153

[2k]
If we set w = νz, then z = (−1)k νw by (6.7.7), so the component of z in ClC is
(−1)k w0 ν where w0 is the constant term in w. We shall prove that
(6.7.18) Tr µ(w) = 2k w0 .
In view of (6.7.17) this will show that (−1)k w0 = (−2)−k Tr µ(w) = (2i)−k Str µ(z),
which is the Berezin-Patodi formula. To prove (6.7.18) we note first that µ(1) is the
identity in a space of dimension 2k , so the trace is 2k . We also have to show that for
j = 1, . . . , 2k we have Tr µ(w) = 0 if
w = εi1 · · · εij , 1 ≤ i1 < · · · < ij ≤ 2k.
If j is even we write w = εi1 · w1 = −w1 · εi1 and obtain
Tr µ(w) = Tr(µ(εi1 )µ(w1 )) = Tr(µ(w1 )µ(εi1 )) = − Tr µ, hence Tr µ(w) = 0,
for Tr(AB) = Tr(BA). If j is odd we choose i ∈ {1, . . . , 2k} \ {i1 , . . . , ij }. Then we
have w = εi · εi · w = −εi · w · εi , and it follows that
Tr µ(w) = Tr(µ(εi )2 µ(w)) = Tr(µ(εi )µ(w)µ(εi )) = − Tr µ(w),
hence Tr µ(w) = 0. The proof is complete.
The important feature of the Berezin-Patodi formula is that the information about
the supertrace is contained in the highest part of the Clifford algebra. As we shall
see in Section 6.10, the parametrix construction will show that the jth term has no
[2j]
component above the level ClC , and that the term there can be calculated for j ≤ k;
we shall precisely need the kth term to compute an index.
6.8. Clifford and spinor analysis. Let M be a C ∞ Riemannian manifold of
dimension n. For every x ∈ M we have defined in Section 6.7 a Clifford algebra
Clx (M ) = Cl(Tx∗ , gx∗ ), where gx∗ is the quadratic form in Tx∗ M dual to the metric form
gx in Tx M . It is clear that Clx (M ) is the fiber at x of a vector bundle Cl(M ), which
is a quotient of ⊕j≤n ⊗j T ∗ M . We can also view Cl(M ) as a bundle associated to the
orthonormal frame bundle and the representation of the orthogonal group on Cl(Rn , e)
induced by the natural action of O(n) on Rn . In either way we see that the Levi-Civita
connection is defined in Cl(M ).
As motivated in Section 6.6, Dirac operators involve bundles with a special structure:
Definition 6.8.1. A (complex) vector bundle E on M is called a Clifford bundle if for
every x ∈ M we have a map
Clx (M ) × Ex ∋ (v, ϕ) 7→ v · ϕ ∈ Ex
which makes Ex a Clx (M ) module, and the map Cl(M ) ⊕ E → E is in C ∞ . We shall
say that E is a Hermitian Clifford bundle if E is provided with a Hermitian metric
such that the map σ(v) : ϕ 7→ v · ϕ in End Ex is self-adjoint for every v ∈ Tx∗ M .
Note that the last assumption implies that (σ(v)ϕ, σ(v)ϕ) = (σ(v)2 ϕ, ϕ) =
|v|2x (ϕ, ϕ), where |v|x is the norm in Tx∗ M , so σ(v) is an isometry if |v|x = 1.
This implies that the whole group Pin(Tx∗ , gx∗ ) acts as a group of isometries on Ex .
154 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

Proposition 6.8.2. In a Hermitian Clifford bundle one can always define a connec-
tion ∇E compatible with the metric in the sense of (5.1.12) such that for every vector
field X, one form v and section ϕ of E we have

(6.8.1) ∇E E
X (v · ϕ) = (∇X v) · ϕ + v · ∇X ϕ.

Here ∇ is the Levi-Civita connection.

A connection compatible with the metric and satisfying (6.8.1) is called a Clifford
connection.
Proof. If (6.8.1) is valid and v ∈ Tx∗ , |v|x = 1, then it follows that

∇E E
X ϕ(x) = v · ∇X (V · ϕ)(x),

if the one form V is chosen with ∇V = 0 and V = v at x. This determines V up to

second order terms which do not affect the right-hand side. Repeating this argument
we obtain

(6.8.2) ∇E E ∗
X ϕ(x) = u · ∇X (u · ϕ)(x), u ∈ Pin(Tx , gx∗ ),

where u∗ should be extended so that ∇u∗ = 0 at x. By Proposition 5.1.7 we can

always choose a connection ∇E which is compatible with the Hermitian metric. There
is no reason why it should satisfy (6.8.2), but we can force it to do so by passing to
the connection Z
∇X ϕ(x) = u · ∇E
e E ∗
X (u · ϕ)(x) du

where du is the invariant measure on the compact group Pin. Since the metric is
invariant under Pin by assumption, this is still a connection compatible with the metric,
e E , which means that (6.8.1) holds at x when ∇v = 0 at
and (6.8.2) is now valid for ∇ Pn
x. In general we can for any given point x write v = 1 ψj vj in a neighborhood of x,
where ∇vj = 0 at x and ψj ∈ C ∞ . Then we obtain
X X
e E (v · ϕ) =
∇ (Xψj )vj · ϕ + e E ϕ = (∇X v) · ϕ + v · ∇
ψj vj · ∇ e E ϕ,
X X X

e E is a connection, then that ∇ is a connection. This

where we have used first that ∇ X
completes the proof.
For any connection ∇E in a Clifford bundle E we can define a first order differential
operator D in C ∞ (M, E) by composing

∇E : C ∞ (M, E) → C ∞ (M, T ∗ ⊗ E)

with the Clifford multiplication map

m : C ∞ (M, T ∗ ⊗ E) → C ∞ (M, E)
 CLIFFORD AND SPINOR ANALYSIS 155

extending the bilinear map Tx∗ × Ex ∋ (v, ϕ) 7→ v · ϕ ∈ Ex . For ψ ∈ C ∞ (M ) and

ϕ ∈ C ∞ (M, E) we have

e−tψ D(etψ ϕ) = m((dψ) ⊗ ϕ) + Dϕ

so the principal symbol as defined in Section 5.1 is at (x, ξ) ∈ Tx∗ the map

Ex ∋ w 7→ ξ · w.

In other words,

(6.8.3) D(ψϕ) = ψDϕ + (dψ) · ϕ, ψ ∈ C ∞ (M ), ϕ ∈ C ∞ (M, E).

The square of the symbol is equal to |ξ|2 times the identity in Ex , and if E is a
Hermitian Clifford bundle it follows that D is of Dirac type. We can say more if the
connection is well chosen:
Proposition 6.8.3. If ∇E is a Clifford connection in the Hermitian Clifford bundle
E, then D is skew symmetric.
Proof. We must show that if ϕ and ψ are in C0∞ (M, E), then
Z


(6.8.4) (Dϕ, ψ) + (ϕ, Dψ) dvol = 0.
M

At any point we can find an orthonormal basis e1 , . . . , en for the vector fields, hence
a dual orthonormal frame ε1 , . . . , εn for the one forms, with ∇εj = 0 for all j at the
chosen point. Since
Xn
E
∇ ϕ= εj ⊗ ∇Eej ϕ,
1

we have
n
X
Dϕ = εj · ∇E
ej ϕ.
1

Define a (complex) vector field X by

X(v) = (v · ϕ, ψ) = (ϕ, v · ψ),

where v is a one form. Taking v = εj we obtain at a point where ∇εj = 0 for every j,
using (6.8.1) and (5.1.12),
n
X n
X
(∇ej X)(εj ) = (εj · ∇E E
ej ϕ, ψ) + (εj · ϕ, ∇ej ψ) = (Dϕ, ψ) + (ϕ, Dψ).
1 1

Since X has compact support and the divergence in the left-hand side is independent
of the choice of frames (cf. (6.3.14)), we conclude using (6.3.14)′ that (6.8.4) holds.
156 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

Remark. Usually one includes a factor i in the definition to make D symmetric. How-
ever, we have stayed here with the definition of principal symbol given in Section 5.1
and hope to be consistent.
We shall say that a Clifford module E is graded if a direct sum decomposition
E = E 0 ⊕E 1 is given which is compatible with the grading in Cl(M ), that is, Clix ·Exj ⊂
Exi+j mod 2 . The Clifford connection can clearly be chosen so that it respects this
decomposition, and then the Dirac operator takes the form (6.6.2).
Exercise 6.8.1. Prove that if ∇E is a Clifford connection and v is a one form, then

(6.8.5) D(v · ϕ) = −v · Dϕ + 2∇E

v ♯ ϕ + (Dv) · ϕ

where D is the Dirac type operator corresponding to the Clifford module Cl(M ), that
is, with the notation in the proof of Proposition 6.8.3,
X
Dv = εj · ∇ej v.

Prove that D(dψ) = ∆ψ if ψ ∈ C ∞ (M ), and conclude that

(6.8.6) D2 (ψϕ) = ψD2 ϕ + 2∇(dψ)♯ ϕ + (∆ψ)ϕ, ψ ∈ C ∞ (M ), ϕ ∈ C ∞ (M, E).

We shall now prove Weitzenböck type formulas, similar to (6.3.15), for a Dirac type
operator D corresponding to a Hermitian Clifford bundle E with a Clifford connection.
Then D is skew adjoint and the principal symbol of D2 is |ξ|2 . For the connection
∇ : C ∞ (M, E) → C ∞ (M, T ∗ ⊗ E) we can form the adjoint with respect to the metric
in E and the metric in T ∗ ⊗ E such that
n
X n
X
2
k εj ⊗ wj k = kwj k2 , wj ∈ Ex ,
1 1

if ε1 , . . . , εn is an orthonormal basis in Tx∗ . (This is the Hilbert-Schmidt metric; prove as

an exercise that it is independent of the choice of basis ε1 , . . . , εn .) Then the principal
symbol of ∇∗ ∇ is −|ξ|2 , which proves that D2 + ∇∗ ∇ is of order ≤ 1. We shall now
show that it is of order 0, which is no surprise in view of Proposition 6.1.2. For the
proof we take χ ∈ C0∞ (M, R), ϕ ∈ C ∞ (M, E) and recall from Exercise 6.8.1 that

(6.8.7) D2 (χϕ) = χD2 ϕ + 2∇X ϕ + (∆χ)ϕ, X = (dχ)♯ .

Since ∇ is a connection we have

∇(χϕ) = χ∇ϕ + (dχ) ⊗ ϕ.

If ψ ∈ C ∞ (M, R) it follows that

(∇∗ ∇(χϕ), ψ) = (∇(χϕ), ∇ψ) = (χ∇ϕ, ∇ψ) + ((dχ) ⊗ ϕ, ∇ψ) = (∇ϕ, ∇(χψ))
− (∇ϕ, (dχ) ⊗ ψ) + ((dχ) ⊗ ϕ, ∇ψ) = (χ∇∗ ∇ϕ, ψ) − (∇X ϕ, ψ) + (ϕ, ∇X ψ).
 CLIFFORD AND SPINOR ANALYSIS 157

Since the connection is compatible with the Hermitian metric we have with scalar
products in the fiber at x

(ϕ, ∇X ψ)(x) + (∇X ϕ, ψ)(x) = ∇X (ϕ, ψ)(x),

so using (6.3.13) to integrate by parts we obtain

(∇∗ ∇(χϕ), ψ) = (χ∇∗ ∇ϕ, ψ) − 2(∇X ϕ, ψ) − ((div X)ϕ, ψ),

that is,
∇∗ ∇(χϕ) = χ∇∗ ∇ϕ − 2∇X ϕ − (div X)ϕ.
Combining this result with (6.8.7) we conclude that

(D2 + ∇∗ ∇)(χϕ) = χ(D2 + ∇∗ ∇)ϕ,

which means that D2 + ∇∗ ∇ is of order 0. More precisely, we have

Proposition 6.8.4. If D is the Dirac type operator in a Hermitian Clifford bundle E
over M corresponding to a Clifford connection ∇, then
n
X
2 ∗ 1
(6.8.8) D ϕ + ∇ ∇ϕ = 2 εk · εj · R∇ (ek , ej )ϕ, ϕ ∈ C ∞ (M, E),
j,k=1

where ε1 , . . . , εn and e1 , . . . , en are local dual orthonormal frames for the one forms
and the vector fields, and R∇ is the curvature of E with the connection ∇, defined by
(5.1.5).
Proof. Writing ∇E for the connection in E to distinguish it from the Levi-Civita
connection in T ∗ M , we have
n
X
Dϕ = εj · ∇E
ej ϕ, hence
1
n
X Xn n
X
2
D ϕ= εk · ∇E
ek (εj · ∇E
ej ϕ) = εk · (∇ek εj ) · ∇E
ej ϕ + εk · εj · ∇E E
ek ∇ej ϕ,
j,k=1 j,k=1 j,k=1

since ∇E is a Clifford connection. The second sum is equal to

n
X X
∇E E
e j ∇e j ϕ + εk · εj · (∇E E E E
ek ∇ej − ∇ej ∇ek )ϕ, and
j=1 j>k
E
(∇E E E E
ek ∇ej − ∇ej ∇ek )ϕ = R
∇
(ek , ej )ϕ + ∇E
[ek ,ej ] ϕ

εj ⊗ ∇E ∞
P
by (5.1.5). Since ∇ϕ = ej ϕ we obtain if ψ ∈ C0 (M, E) has support in the
coordinate patch U where the local frames are defined
X
(∇∗ ∇ϕ, ψ) = (∇ϕ, ∇ψ) = (∇E E
ej ϕ, ∇ej ψ).
158 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

When x ∈ U we have with scalar product taken only in the fiber at x

(∇E E E E E E
ej ϕ, ∇ej ψ)(x) = ∇ej (∇ej ϕ, ψ)(x) − (∇ej ∇ej ϕ, ψ)(x).

If χ ∈ C0∞ (U ) then
Z Z Z
∇ej χ dvol = hej , dχi dvol = − χ div ej dvol,
U U

which proves that

n
X n
X
∇∗ ∇ϕ = − ∇E E
e j ∇e j ϕ − (div ej )∇E
ej ϕ.
j=1 j=1

Summing up, we have found that

n
X E
2 ∗ 1
(D + ∇ ∇)ϕ − 2 εk · εj · R∇ (ek , ej )ϕ
j,k=1
n
X n
X n
X
= εk · (∇ek εj ) · ∇E
ej ϕ − div ej ∇E
ej ϕ + 1
2 εk · εj · ∇E
[ek ,ej ] ϕ.
j,k=1 j=1 j,k=1

For any point x ∈ U we can choose ϕ with given value and ∇ϕ = 0 at x, which makes
the right-hand side equal to 0 at x. But we have already proved that the left-hand side
only depends on ϕ(x), which completes the proof of (6.8.8).
As an example we shall derive (6.3.16) again by applying Proposition 6.8.4 to the
Dirac type operator d + d∗ in C ∞ (M, ∧∗C M ) (see (6.3.5)). Recall that by Exercise
3.1.2 and (5.1.5) the matrix of R∇ (X, Y ) for the cotangent bundle with the Levi-
Civita connection is − k,l R ikl X k Y l , i, j = 1, . . . , n, in the coordinate frame. With
j
P

Ri jkl denoting the components

P of the Riemann curvature tensor in the frame ε1 , . . . , εn
instead, we obtain if ϕ = ϕl εl is a one form
X
R∇ (ek , ej )ϕ = − Ri lkj ϕi εl .
l,i

Hence the right-hand side of (6.8.8) becomes

n
X
(6.8.9) − 12 εk · εj · εl Ri lkj ϕi .
i,j,k,l=1

If j, k, l are different then εk · εj · εl is invariant under circular permutations, so the

first Bianchi identity shows that the sum of such terms is equal to 0. If k = j then
 CLIFFORD AND SPINOR ANALYSIS 159

Ri lkj = 0 so the only contributions come when k = l or j = l; the terms common to

these cases vanish. Hence (6.8.9) is equal to
n
X n
X X
i i
(6.8.10) − 12 (−εj R kkj + εk R jkj )ϕi =− εj Ri kjk ϕi = − εj Ri j ϕi ,
i,j,k=1 i,j,k=1

where R denotes the Ricci tensor in the last formula. This gives (6.3.16). The
Weitzenböck formulas for forms of higher degree could be obtained in the same way,
but then we would have to work out the curvature form in the higher exterior powers
using Exercise 5.4.1.

We shall now pass to the main examples of Dirac operators, based on the half spin
representations D±1 in (6.7.15). These are not representations of SO(2k) but of the
2
covering group Spin(2k), so we cannot define associated vector bundles for an arbitrary
Riemannian manifold.
Definition 6.8.5. An oriented Riemannian manifold M of dimension n has a spin struc-
ture if there exists a principal Spin(n) bundle Pe on M , which is a double cover of the
oriented orthonormal frame bundle P , such that with the map Spin(n) → SO(n) de-
fined by (6.7.12) we have a commutative diagram, with horizontal arrows defined by
the right group action,
Pe × Spin(n) −−−−→ Pe
 
 
y y
P × SO(n) −−−−→ P.

In a neighborhood U of any point in M we can choose a spin structure just by

identifying P with U × SO(n) and Pe with U × Spin(n). However, the existence of
a spin structure in the large requires the vanishing of an element in H 2 (M, Z2 ) (the
Stiefel-Whitney class), and there may exist inequivalent spin structures when this
condition is fulfilled.
Assume now that M is an oriented manifold of dimension n = 2k with a spin
structure. Then the half spin bundles

S± (Pe) = Pe ×D± S± (2k)

1
2

and the full spin bundle S(Pe) = S+ (Pe) ⊕ S− (Pe) are defined. The latter is associated
with the direct sum ̺ of the representations D± 1 , which is the restriction of (6.7.14)
2
to Spin(2k). The Clifford bundle can be regarded as the bundle associated with the
representation of Spin(2k) on Cl(2k) by the representation

Spin(2k) × Cl(2k) ∋ (a, w) 7→ awa∗ ∈ Cl(2k)

for this is a representation since a∗ is the inverse of a ∈ Spin(2k), and it gives the
orthogonal transformation τ (a) when applied to an element in R2k .
160 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

Proposition 6.8.6. The spin bundle S(Pe) is a Hermitian Clifford bundle, and the
Levi-Civita connection is a Clifford connection in S(P̃ ). For the corresponding Dirac
operator we have the Lichnerowicz formula

(6.8.11) D2 + ∇∗ ∇ = −S/4,

where S is the scalar curvature.

Proof. Since the spin representation is unitary, the Hermitian metric in S(2k) induces
one in S(Pe). The map
Cl(2k) × S(2k) → S(2k)
defined by the identification of ClC (2k) with End(S(2k)), induces for every x ∈ M a
Clifford module structure Clx (M ) × Sx (Pe) → Sx (Pe). In fact, when

(p, x, w) ∈ Pex × Cl(2k) × S(2k)

and (pa−1 , x′ , w′ ) with a ∈ Spin(2k) define the same element in Clx (M ) × Sx (Pe), then
x′ = a · x · a∗ and w′ = a · w, which implies that x′ · w′ = a · (x · w), hence that (p, x · w)
and (pa−1 , x′ · w′ ) = (pa−1 , a · (x · w)) define the same element in S(P̃ ). It is clear
that the multiplication so defined depends smoothly on x. Multiplication by a vector
v ∈ Tx∗ M is self-adjoint, since by our original definition it is the sum of an operator
and its adjoint, so S(Pe ) is a Clifford module. At the center of a geodesic coordinate
system the Levi-Civita connection cannot be distinguished from the flat connection in
any of the bundles involved, so it is obvious that we have a Clifford connection.
With Ri jkl denoting the components of the Riemannian curvature tensor in the
oriented orthonormal frame ε1 , . . . , εn of Tx∗ , with dual frame e1 , . . . , en in Tx , we
have seen above that the curvature R∇ (ek , ej ) of the cotangent bundle has the skew
symmetric matrix −(Rl ikj )l=1,...,n l=1,...,n
i=1,...,n = −(Rlikj )i=1,...,n . By Theorem 6.7.11 this skew
symmetric matrix as an element in so(n) corresponds in the Lie algebra of Spin(n) to
1
P
− 4 i,l Rlikj εi · εl . Hence the right-hand side of (6.8.8) becomes
X X
− 81 Rlikj εk · εj · εi · εl = − 14 εj Rlj · εl = − 14 S,
i,j,k,l

where we have used (6.8.10). This proves (6.8.11).

The Lichnerowicz formula has a corollary completely parallel to Bochner’s Theorem
6.3.5:
Corollary 6.8.7. If M is a compact Riemannian manifold of dimension 2k with spin
structure and non-negative not identically vanishing scalar curvature, then there are
no harmonic spinors in M , that is, if ϕ ∈ D ′ (M, S) and Dϕ = 0, then ϕ = 0. If
the scalar curvature is identically 0, then the space of harmonic spinors has dimension
≤ 2k .
 HERMITE POLYNOMIALS AND MEHLER’S FORMULA 161

Proof. If Dϕ = 0 it follows that ϕ ∈ C ∞ since D is elliptic, and taking the scalar

product of (6.8.11) with ϕ we obtain

k∇ϕk2 + 41 (Sϕ, ϕ) = 0.

Hence ϕ is parallel, and if S is not identically 0 we obtain ϕ = 0 since ϕ must vanish in

some open set. If S is identically 0 we can just conclude that ϕ is uniquely determined
by its values at a point.
If F is a Hermitian vector bundle over the Riemannian manifold M of dimension
2k with spin structure defined by the principal Spin(2k) bundle Pe , then the tensor
product E = S(P̃ ) ⊗ F is a Clifford bundle with Cl(M ) acting on the first factor. If F
has a connection ∇F compatible with the metric, then the connection ∇E in E, given
according to Proposition 5.1.8 by the Levi-Civita connection in S(Pe) and ∇F in F ,
is a Clifford connection. The easy verification is left as an exercise. Hence we get a
twisted Dirac operator DF in C ∞ (M, E) = C ∞ (M, E+ ) ⊕ C ∞ (M, E− ), interchanging
sections of E± = S± (Pe) ⊗ F . Since the curvature of E is the sum of the curvature of
S(Pe) and that of F , tensored respectively with the identity in F and that in S(Pe), the
Weitzenböck formula (6.8.8) now takes the form
X F
(6.8.12) DF2 ϕ + ∇∗ ∇ϕ = − 14 Sϕ − 12 εi · εj I ⊗ R∇ (ei , ej ),

where I is the identity in the spin bundle. We shall need (6.8.12) in Section 6.10.
6.9. Hermite polynomials and Mehler’s formula. The proof of the local index
formula for twisted Dirac operators will culminate in an identification of the difference
of the traces of the heat kernels involved with the heat kernel belonging to an operator
closely related to the harmonic oscillator. As a preparation we shall now give the
classical background.
The Hermite polynomial Hn (x), x ∈ R, of order n is defined by
2 2
(6.9.1) Hn (x) = ex (−d/dx)n (e−x ) = (2x)n + . . . ,
2
which by Taylor’s formula for e−(x−z) implies that we have the generating function
∞
X Hn (x) n 2
(6.9.2) z = e2xz−z .
n=0
n!
2
They are orthogonal with respect to the weight function e−x , for if n < m we have
Z Z
−x2 2
Hn (x)Hm (x)e dx = Hn (x)(−d/dx)m e−x dx = 0
R R

since the integrand vanishes after n + 1 ≤ m integrations by parts. When n = m we

obtain instead
√
Z Z
−x2 2
Hn (x)Hn (x)e dx = Hn(n) e−x dx = 2n n! π.
R R
162 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

Summing up,

√
Z
2
(6.9.3) Hn (x)Hm (x)e−x dx = δnm 2n n! π
R

1 2
p √
Hence the functions e− 2 x Hn (x)/ 2n n! π form an orthonormal system in L2 (R). It
1 2
is complete, for if u ∈ L2 (R) is orthogonal to all of them, then U (x) = u(x)e− 2 x is
orthogonal to all polynomials; the Fourier transform is then an entire analytic function
with all derivatives equal to 0 at 0, hence identically 0, which proves that u = 0.
If in the identity
∞
1 2 1 2 2
X
Hn (x)e− 2 x z n /n! = e− 2 x +2xz−z
0

we apply the differential operator

(6.9.4) L = −d2 /dx2 + x2 − 1,

noting that −(2z − x)2 + x2 = 2z(2x − 2z), it follows that

∞ ∞
X
− 21 x2 d 1 2 2 X 1 2
L(Hn (x)e )z /n! = 2z e 2 x +2xz−z =
n
Hn (x)e− 2 x 2nz n /n!,
0
dz 0

which means that

1 2
(6.9.5) (L − 2n)(Hn (x)e− 2 x ) = 0,

so we have a complete set of eigenfunctions of L, with eigenvalues 2n.

To calculate the polynomials we introduce a convenient formalism. We can write
∞
−z 2
X
(6.9.6) e = hn z n /n!, h2n+1 = 0, h2n = (−1)n (2n)!/n!.
0

P n n
The right-hand side is obtained from the power series h z /n! = ehz when hn is
replaced by hn . Such a substitution is legitimate in any equality between two power
series in h provided that no convergence difficulties occur, for the coefficients of corre-
sponding powers of h are identical. If we apply this to the identity
∞
X
2xz hz 2xz+hz
e e =e = (2x + h)n z n /n!
0

we obtain with E denoting the substitution of hn for hn

(6.9.7)  
n
X
n−2k n X
Hn (x) = E(2x + h) = (2x) h2k = n! (2x)n−2k (−1)k /((n − 2k)!k!).
2k
2k≤n 2k≤n
 HERMITE POLYNOMIALS AND MEHLER’S FORMULA 163

Now we wish to determine the kernel of e−tL when t > 0, which is equal to
X 1 2
+y 2 )
√
e−2tn Hn (x)Hn (y)e− 2 (x /(2n n! π).

This means that we must calculate

∞
X
(6.9.8) Hn (x)Hn (y)(s/2)n/n!, when 0 < s < 1.
0

By (6.9.7) and (6.9.2) the left-hand side is E applied to

∞
X 2 2
(6.9.9) (h + 2x)n Hn (y)(s/2)n/n! = ey(h+2x)s−(h+2x) s /4
,
0

where the exponent simplifies to

−h2 s2 /4 + h(ys − xs2 ) + 2xys − x2 s2 .

The expansion of the right-hand side is not obvious unless the coefficient of h vanishes,
that is, y = xs, but then we can use that
∞ ∞
−h2 s2 /4
X
n 2 n
X (2n)!
Ee = (−1) h2n (s /4) /n! = (s2 /4)n .
0 0
n!2

13 − 12
Since (2n)!/(4n n!2 ) = . . . ( 21 + n − 1)/n! = (−1)n


22 n
, we conclude that
2 2
p
s /4
(6.9.10) Ee−h = 1/ 1 − s2 .

We have now proved that

2 2
X p
Hn (x)Hn (y)(s/2)n/n! = e2xys−x s
/ 1 − s2 , if y = xs.

In view of the symmetry we may exchange x and y in our earlier calculations which
gives
2 2 2 2 2 2 2
p
Ee−h s /4+h(xs−ys )+2xys−y s = e2xys−x s / 1 − s2 , if y = xs.
With the notation v = xs − ys2 = xs(1 − s2 ) we have (y 2 − x2 )s2 = x2 (s2 − 1)s2 =
−v 2 /(1 − s2 ) and conclude that for any v
2 2 2
/(1−s2 )
p
s /4+hv
(6.9.11) Ee−h = e−v / 1 − s2 .

This is the expansion we needed, so we now obtain for arbitrary x, y

2 2 2 2 2
X p
Hn (x)Hn (y)(s/2)n/n! = e−(ys−xs ) /(1−s )+2xys−x s / 1 − s2 .
164 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

This simplifies to
∞
2 2
s +y 2 s2 −2xys)/(1−s2 )
X p
(6.9.12) Hn (x)Hn (y)(s/2)n/n! = e−(x / 1 − s2 ,
0

which is known as Mehler’s formula. I owe the trick in the proof to Marcel Riesz who
showed it to me in a private conversation with him in 1952. My notes state that the
proof goes back to the 1920’s. I have inserted the operator E to justify the formal
argument.
From (6.9.12) it follows that
∞
2
− 12 1
+y 2 )
X
π Hn (x)Hn (y)(s/2)ne− 2 (x /n!
0
2
1
+y 2 )(1+s2 )−4xys)/(1−s2 )
p
= e− 2 ((x / π(1 − s2 ),
and with s = e−2t we obtain the kernel of e−tL as

√
exp − ( 21 (x2 + y 2 ) cosh 2t − xy)/ sinh 2t et / 2π sinh 2t.
The factor et is caused by the term −1 in L = −d2 /dx2 + x2 − 1. If we define
L1 = −d2 /dx2 + x2 , it follows that the kernel of e−tL1 is

√
K(t, x, y) = exp − 12 ((x2 + y 2 ) cosh 2t − 2xy)/ sinh 2t) / 2π sinh 2t.
√
To find the kernel of the operator La = −d2 /dx2 + a2 x2 we take x̃ = x a as new
variable so that La = aL e 1 . Then
√
Z
−tLa e1 ˜
−taL
(e f )(x) = (e f )(x̃) = K(ta, x̃, ỹ)f (ỹ/ a)dỹ
√ √ √
Z
= K(ta, x a, y a)f (y) a dy,
√ √ √
so the kernel of e−tLa is aK(ta, x a, y a), that is,
r
1 2at  1 2at 2 2


(6.9.13) √ exp − (x + y ) cosh 2at − 2xy .
4πt sinh 2at 4t sinh 2at
This formula gives immediately the kernel of a harmonic oscillator
−∆ + hAx, xi
when A is a positive diagonal matrix, the only difference is that we get a product of
such factors. By the orthogonal invariance the kernel is therefore in general, when the
dimension is n,
s √
1 2 At
(6.9.14) n det √
(4πt) 2 sinh 2 At
√ √ √
 1 2 At 2 At 2 At 
× exp − √ x, x + √ y, y − 2 √ x, y .
4t tanh 2 At tanh 2 At sinh 2 At
 HERMITE POLYNOMIALS AND MEHLER’S FORMULA 165

This is how the result is stated by Getzler.

We shall encounter a somewhat different operator, namely
n
X n
X
(6.9.15) L=− (−∂j − i Ωjk xk )2 ,
j=1 k=1

where Ω is a real skew symmetric matrix and i is the imaginary unit. At least formally
we can write L = L0 + L1 , where
n
X X
L0 = − ∂j2 + hΩx, Ωxi, L1 = −2i Ωjk xk ∂j .
1

Formally the operators commute, for

X
[∆, L1 ] = −4i Ωjk ∂j ∂k = 0,
X
[L1 , hΩ2 x, xi] = −4i Ωjk xk (Ω2 x)j = −4ihΩx, Ω(Ωx)i = 0.

Formally this means that e−tL = e−tL0 e−tL1 , but this is rather delicate since L1 is a
formally self-adjoint operator not bounded from below. To proceed, at first formally,
we shall now assume that n = 2, which is no essential restriction since every skew
symmetric Ω is the direct sum of such operators and trivial ones.
With n = 2 we now assume that
   2 
0 a 2 a 0
Ω= , hence − Ω = ,
−a 0 0 a2
L0 = −∆ + a2 |x|2 , L1 = −2ihΩx, ∂i.

The kernel of e−tL0 is

1 2at  1 2at

exp − (|x|2 + |y|2 ) cosh(2at) − 2hx, yi .
4πt sinh(2at) 4t sinh(2at)

The vector field hΩx, ∂i = ax2 ∂/∂x1 − ax1 ∂/∂x2 is infinitesimal generator of rotation
with the speed a, so for real s we conclude that eshΩx,∂i is rotation by the angle sa,
that is, composition with the map x 7→ (x1 cos(sa) −x2 sin(sa), x1 sin(sa) +x2 cos(sa)),
so we can calculate eshOx,∂i e−tL0 by just composing in the x variables with this map.
By analytic continuation to s = 2it this gives for the inner parenthesis in the exponent

cosh(2at)(|x|2 + |y|2 − 2x1 y1 − 2x2 y2 ) + 2i sinh(2at)(x2 y1 − x1 y2 )

= cosh(2at)|x − y|2 + 2i sinh(2at)hΩx, yi/a,

and leads to the kernel

1 2at  1  2at 
exp − |x − y|2 + 4ithΩx, yi .
4πt sinh(2at) 4t tanh 2at
166 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

By analytic continuation
√ it follows at once that ∂/∂t + Lx annihilates the kernel, and
taking x = y + z t we find at once that it converges to δ as t → 0, which easily
justifies that we√have indeed found the fundamental solution. With the suggestive
notation |Ω| = −Ω2 it follows that the heat kernel for the operator (6.9.15) with a
general real skew symmetric Ω is
s
1 2|Ω|t  1 2|Ω|t 
(6.9.16) n det exp − (x − y), x − y + 4ithΩx, yi .
(4πt) 2 sinh 2|Ω|t 4t tanh 2|Ω|t

Note that z/ sinh z and z/ tanh z are analytic functions of z 2 in a neighborhood of the
real axis. For small t the right-hand side of (6.9.16) is therefore well defined even if Ωjk
are not real valued but take their values in some commutative finite dimensional alge-
bra, such as ∧∗ Ck , and it will still be a fundamental solution. This will be important
in Section 6.10.
6.10. The local index formula for twisted Dirac operators. Let M be a
Riemannian manifold of even dimension n with a spin structure, and let F be a Her-
mitian vector bundle on M with a connection compatible with the metric. At the end
of Section 6.8 we defined the skew adjoint twisted Dirac operator DF in C ∞ (M, E)
where E = E+ ⊕ E− and E± = S± (Pe) ⊗ F . It maps sections of E± to sections of E∓ .
By Proposition 6.5.1 we know that the index of DF+ : C ∞ (M, E+ ) → C ∞ (M, E− ) for
2
t > 0 is equal to the difference between the trace of etDF on C ∞ (M, E+) and the trace
2
on C ∞ (M, E− ). Thus it is the supertrace of etDF on C ∞ (M, E) with the grading by
2
E+ and E− . Let K(t, x, y) ∈ Hom(Ey , Ex ) be the kernel of etDF . We want to show
in analogy to (6.5.5)′ and (6.5.6)′ that the supertrace of K(t, x, x) for fixed x, without
integration over M , has an asymptotic expansion in non-negative integer powers of
t and calculate the constant term. The crucial point is the Berezin-Patodi formula
(Theorem 6.7.14), which has an obvious extension to Clx (M )C ⊗ Fx . Let

Lx(j) = ⊕ν≤min(j,n/2) Clx[2ν] (M )C ⊗ End Fx ,

be the natural filtration of the even part of End E; recall that End S(P̃ ) ∼
= Cl(M )C .
(j)
When j ≥ n/2 then Lx is the whole even part.
We have seen in Section 6.4 that using geodesic coordinates centered at y and
synchronous frames for E and T ∗ M , hence for S(Pe), we have for fixed y an asymptotic
expansion
∞
X
K(t, x, y) ∼ H0 (t, x) (−t)ν uν (x)/ν!
0
1
where H0 is the fundamental solution of the scalar heat equation, u0 = g(x)− 4 and
u1 , u2 , . . . are determined successively by integrating the equations (6.1.14) (or equiv-
alently (6.1.14)′ ). What makes the supertrace on the diagonal accessible to explicit
calculation is the following fact:
 THE LOCAL INDEX FORMULA FOR TWISTED DIRAC OPERATORS 167

(ν)
Lemma 6.10.1. The coefficients uν (0) are in Ly for ν = 0, 1, . . .
This is clear when ν = 0 and will be proved by induction for increasing ν. The
statement is void when ν ≥ n/2, but in the course of the proof we shall isolate the
terms which are not better than the lemma states, and when ν = n/2 this will give uν
mod Lk−1 y , which is all that we need. Before the proof we need some preliminaries. The
first point is to express DF2 in terms of geodesic coordinates and synchronous frames
ε1 , . . . , εn for T ∗ M , f1 , . . . , fN for F . With the dual frame e1 , . . . , en in T M we write
for l = 1, . . . , n

n
X N
X
(6.10.1) ∇el εj = e lj k εk , j = 1, . . . , n,
Γ ∇e l f j = ΓF k
lj fk , j = 1, . . . , N.
k=1 k=1

e lj k = −Γ
Here Γ e lk j since the frame is orthonormal, and by Theorem 3.3.6 we have

n
X
(6.10.2) eklj (x) =
Γ 1
Rkjil (0)xi + O(|x|2 ).
2
i=1

Note that since ∇k εj = ∂k εj at 0, it follows that εj (x) − εj (0) = O(|x|2 ). (Γ

e lj k are not
∗
quite the Christoffel symbols since they refer P to kthe synchronous frame for T M and
1 n
not the frame dx , . . . , dx .) Since ∇̺ fj = k x ∇∂k fj (x) = 0, taking the first order
terms in the Taylor expansion shows that ∇∂k fj (0) = 0 for all j and k, hence

(6.10.3) ΓF k
lj (0) = 0, for l = 1, . . . , n; j, k = 1, . . . , N.

In the cotangent bundle we have

X n
X
∇e l ( ϕj εj ) = hel , ∂iϕj + elj k ϕj εk .
Γ
j,k=1

The matrix (Γe lj k )j=1,...,n is in so(n) for fixed l, and as in the proof of Proposition 6.8.6
k=1,...,n
it corresponds to multiplication by − 1
P e k
4
Γlj εj · εk in the Clifford algebra, so
j,k

n
X n
X N
X
1 e lj k εj · εk + ΓF k


(6.10.4) ∇ϕ = εl ⊗ hel , ∂i − 4 Γ lj · ϕ.
l=1 j,k=1 j,k=1

P
To find the adjoint acting on ψ = εl ⊗ ψl we form

n
X n
X N
X
1 elj k εj · εk + ΓF k


(ψ, ∇ϕ) = (ψl , hel , ∂i − 4 Γ lj · ϕ).
l=1 j,k=1 j,k=1
168 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

Here the scalar product is of course taken using the Riemannian volume form, which
is 1 + O(|x|2 ) times the Lebesgue measure. Hence we obtain by partial integration

X n
X N
X
∗ 1 k eF k


∇ ψ=− hel , ∂i − Γlj εj · εk + Γ ψl ,
e
e
4 lj
j,k=1 j,k=1

where Γe lj k and Γ
e eF k also satisfy (6.10.2), (6.10.3) but include the terms arising when
lj
derivatives fall on el or on the Riemannian volume density. Summing up, we obtain in
view of (6.8.12)
n
X F
(6.10.5) DF2 = − 14 S − 1
2 εi · εj R∇ (ei , ej )
i,j=1
n
X n
X N
X n
X N
X
k Fk k
1 1
ΓF k



+ hel , ∂i − Γlj εj · εk + Γlj hel , ∂i − 4 Γlj εj · εk + .
e
e e e
4 lj
l=1 j,k=1 j,k=1 j,k=1 j,k=1

[2]
The coefficients can be viewed as elements in ClC ⊗ End Fy or just End Fy .
Proof of Lemma 6.10.1. We are now ready to prove that for the coefficients uν con-
structed for DF2 as in Section 6.1 we have
X
(6.10.6) uν = uνµ , uνµ ∈ L(µ)
y , uνµ = O(|x|2µ−2ν ), if µ > ν.
µ≤2ν

This is obvious when ν = 0. Assuming that (6.10.6) holds for a certain ν we shall
prove that
X
(6.10.6)′ DF2 uν = vνµ , vνµ ∈ L(µ)
y , vνµ = O(|x|2µ−2ν−2 ), if µ > ν + 1.
µ≤2(ν+1)

This is obvious for the terms coming from the terms in (6.10.5) after the first sum. To
handle the others we observe that multiplication by ΓF k
lj does not raise the Clifford
degree but increases the order of the zeros. Multiplication by Γ̃lj k εj · εk may raise the
Clifford degree by at most 2, but at the same time one gets a factor vanishing at 0.
Differentiation reduces the order of the zero but does not affect the Clifford degree.
Altogether, in the terms coming from the first sum in (6.10.5) where the Clifford degree
is raised 4 units we also get a compensating factor O(|x|2 ), and O(|x|2µ−2ν+2 ) =
O(|x|2(µ+2)−2(ν+1) ). In those where the Clifford degree is raised 2 units we do not
lose any zero at 0, and O(|x|2µ−2ν ) = O(|x|2(µ+1)−2(ν+1) ). In the terms where the
Clifford degree is not raised we cannot lose more than two zeros, and O(|x|2µ−2ν−2 ) =
O(|x|2µ−2(ν+1) ), which completes the proof of (6.10.6)′ .
The integration of the transport equations (6.1.14) for uν+1 , using (6.1.14)′ and
(6.1.17), does not affect the order of the zeros at 0, so (6.10.6) follows by induction,
which proves the lemma.
 THE LOCAL INDEX FORMULA FOR TWISTED DIRAC OPERATORS 169

In the proof the estimate for the order of the zeros was always too low except for
terms coming from the simplified operator
n n n
X X X F
1 i 2 1
(6.10.7) (∂l − 8 Rkjil (y)x εj · εk ) − 2 εi · εj R∇ (ei , ej )(y),
l=1 i,j,k=1 i,j=1

where we have put in the argument y instead of 0 since 0 was the geodesic coordinate
of y. Hence limt→0 Str K(t, y, y) exists and depends only on the Riemann curvature
tensor and the curvature forms of F at y.
An explicit computation can be obtained from the results related to Mehler’s for-
mula given in Section 6.9. As a first step we observe that in the computation of the
component of un/2 in Cl[n]y C ⊗ End Fy there will be no contributions where two equal
factors εi in the Clifford algebra have been multiplied, and different εi anticommute.
The component is therefore equal to the term of degree n obtained if one replaces
Clifford multiplication by exterior multiplication throughout. With the notation
n
X F
(6.10.8) Ωli = − 18 Rkjil (y)dxj ∧ dxk , ΩF = − 12 R∇ (ei , ej )(y)dxi ∧ dxj ,
j,k=1

we must therefore determine the form of degree n in the fundamental solution of

∂/∂t − L where
n
X n
X
(6.10.9) L= (∂j + Ωjk xk )2 + ΩF
l=1 k=1

acts on functions in Rn with values in (∧∗ Cn ) ⊗ Fy . The second term acts only in the
F
factor Fy and commutes with the first, so it only contributes to etL a factor etΩ , and
for the other part we get the fundamental solution from (6.9.16) with Ω replaced by
Ω/i. This means that |Ω|2 is replaced by Ω2 , so the kernel with the pole at 0 is
r
1 2Ωt  1 D 2Ωt E F
(6.10.10) n det exp − x, x etΩ ,
(4πt) 2 sinh 2Ωt 4t tanh 2Ωt

which simplifies to
r
−n 2Ωt F
(6.10.10) ′
(4πt) 2 det etΩ
sinh 2Ωt

when x = 0. By the Berezin-Patodi formula the limit as t → 0 of Str K(t, y, y) is

(2i)n/2 times the coefficient of dx1 ∧ · · · ∧ dxn here, which does not depend on t. Recall
now that the total Chernclass of F is det(I + iΩF /2π); one calls

Ch(F ) = Tr exp(iΩF /2π)

170 LINEAR DIFFERENTIAL OPERATORS IN RIEMANNIAN GEOMETRY

the Chern character of F . The term independent of t in (6.10.10)′ can be calculated

using any convenient choice of t, and we choose t = i/2π. After multiplication by
(2i)n/2 this gives the form
s
(Ωi/π)
det Ch(F ).
sinh(Ωi/π)

Introducing the Riemann curvature matrix

X X
(6.10.11) ΩR
li =
1
2
Rkjil dxj ∧ dxk , (ΩR )li = g lk ΩR
ki

instead of Ω we obtain finally the form

s
(ΩR /4πi)
(6.10.12) det Ch(F ).
sinh(ΩR /4πi)

We have proved the local index theorem:

Theorem 6.10.2. The index of the twisted Dirac operator DF+ in the even dimensional
spin manifold M is the integral of the form (6.10.12) over M . More precisely, the term
in (6.10.12) of maximal degree divided by the normalized positively oriented n form is
equal to the limit as t → 0 of the supertrace of the kernel of etDF on the diagonal.
We refer to Atiyah-Bott-Patodi [1] for the topological arguments required to derive
the general index theorem from Theorem 6.10.2; see also Roe [1] for a discussion of the
passage to the Hirzebruch signature theorem.
APPENDIX A. PREREQUISITES FROM MULTILINEAR ALGEBRA

Let V be a vector space over K = R or K = C of finite dimension n. Then we can

choose a basis e1 , . . . , en ∈ V so that every x ∈ V is a linear combination
n
X
x= xj ej
1

with uniquely determined coordinates xj ∈ K. If e′j is another basis then

n
X
e′k = cjk ej
1

for some cjk ∈ K with non-zero determinant; if we write

X
x= x′k e′k

it follows that
n
X
xj = cjk x′k
1

so bases and coordinates are transformed by inverse transposed matrices.

The dual vector space V ′ (sometimes denoted V ∗ ) is the space of linear forms on V ,
with values in K. With a basis in V as above a linear form L can be expanded as
n
X n
X X
L(x) = L( xk ek ) = L(ek )xk = L(ek )εk (x)
1 1

where X
εk ( xj ej ) = xk , that is, εk (ej ) = δjk .
Thus ε1 , . . . , εn form a dual basis in V ′ which is also a vector space of dimension n
with the natural definition of addition of linear forms and multiplication of them by
scalars. When ξ ∈ V ′ and x ∈ V we usually write hx, ξi for the linear form ξ(x) to
emphasize that it is bilinear, that is, linear in x for fixed ξ (because ξ is a linear form)
and linear in ξ for fixed x (by definition of the vector operations in V ′ ). The bilinear
form is non-degenerate, that is

hx, ξi = 0 ∀x ∈ V =⇒ ξ = 0; hx, ξi = 0 ∀ξ ∈ V ′ =⇒ x = 0.
171
172 APPENDIX A. PREREQUISITES FROM MULTILINEAR ALGEBRA

By the symmetry here we can identify (V ′ )′ with V .

If T : V1 → V2 is a linear map, then a unique linear map T ′ : V2′ → V1′ (also denoted
T ∗ or t T ) called the adjoint or transposed map is defined by

hT x, ηi = hx, T ′ ηi, x ∈ V1 , η ∈ V2′ ,

for the left-hand side is for fixed η a linear form on V1 , hence an element T ′ η ∈ V1′ ,
and it depends linearly on η. T1 and T1′ have the same rank.
If V and W are two vector spaces and B : V × W ∋ x, y 7→ B(x, y) ∈ K is a bilinear
form, then we get in the same way a map W → V ′ and an adjoint map V → W ′ .
The form is called non-degenerate if one (and therefore both) are bijective. Thus the
bilinear forms on V × W can be identified either with the linear maps L(W, V ′ ) or the
linear maps L(V, W ′ ).
Every bilinear form B on V ×V can in one and only one way be written B = B0 +B1
where
B0 (x, y) = B0 (y, x), x, y ∈ V ; B1 (x, y) = −B1 (y, x), x, y ∈ V ;
one calls B0 symmetric and B1 skew symmetric, and we have

B0 (x, y) = (B(x, y) + B(y, x))/2, B1 (x, y) = (B(x, y) − B(y, x))/2; x, y ∈ V.

With a basis e1 , . . . , en for V and corresponding coordinates

n
X n
X n
X
B( xj ej , yk ek ) = bjk xj yk , bjk = B(ej , ek ).
1 1 j,k=1

From the matrix (bjk ) for B one obtains the matrices forPB0 and B1 as (bjk ± bkj )/2.
Only B0 is determined by the quadratic form B(x, x) = bjk xj xk ; we have

B0 (x, y) = (B(x + y, x + y) − B(x − y, x − y))/4 (polarization).

Thus we have a one to one correspondence between quadratic forms and symmetric
bilinear forms. A quadratic form Q is called non-degenerate if the corresponding
symmetric bilinear form B is non-degenerate. If the bilinear form is degenerate we
can find z 6= 0 so that B(V, z) = 0, B(z, V ) = 0, which implies that Q(x + tz) =
B(x + tz, x + tz) = B(x, x) = Q(x) for any t ∈ K, so Q is defined in the quotient space
V /Kz. In what follows we shall use the same notation for a quadratic form and the
corresponding symmetric bilinear form.
If G is a non-degenerate quadratic form in V , then the polarized form defines a
bijection g : V → V ′ and we can define a dual quadratic form G′ in V ′ by

G′ (gx) = G(x), x ∈ V.

If we introduce dual bases in V and V ′ and (gjk ) is the symmetric matrix for G, then
the matrix for G′ is (g jk ) if (g jk ) is the inverse of the matrix gjk .
 APPENDIX A. PREREQUISITES FROM MULTILINEAR ALGEBRA 173

If H is another quadratic form in V , then a linear transformation T : V → V is

defined by the identity
H(x, y) = G(T x, y), x, y ∈ V,
for the corresponding symmetric bilinear forms. With the notation g, h for the maps
V → V ′ defined by G and H, we have T = g −1 h. The map T is self-adjoint with
respect to the bilinear form G since H is symmetric.
There is a unique linear form on L(V, V ), called the trace and denoted Tr, such that
for all x, ξ ∈ V × V ′

Tr Tx,ξ = hx, ξi, if Tx,ξ is defined by Tx,ξ y = xhy, ξi, y ∈ V.

In fact, if ej and εj are dual bases and T ∈ L(V, V ), then

n
X n
X X
Ty = T( ej hy, εj i) = T ej hy, εj i = TT ej ,εj y,
1 1

so the condition requires that

n
X
Tr T = hT ej , εj i.
1
Pn
If we define
PnTr T by this sum for a fixed basis, we obtain for arbitrary x = 1 xj ej ∈ V
′
and ξ = 1 ξj εj ∈ V
n
X n
X n
X
Tr Tx,ξ = hxhej , ξi, εj i = hx, εj ihej , ξi = h ej hx, εj i, ξi = hx, ξi
1 1 1

which proves the existence of the form Tr with the required properites, independent
of the choice
Pn of basis. Note that if Tjk is the matrix of T in terms of a basis then
Tr T = 1 Tjj .
Let V and W be two finite dimensional vector spaces over K. The bilinear forms on
V × W ′ form another vector space called the tensor product of V and W and denoted
′

V ⊗ W , and we have a bilinear map V × W → V ⊗ W mapping x, y ∈ V × W to the

bilinear form
V ′ × W ′ ∋ ξ, η 7→ hx, ξihy, ηi
which is denoted x ⊗ y. (Recall that V is identified with the dual of V ′ and W with the
dual of W ′ , which is behind the definition.) If e1 , . . . , en is a basis of V with dual basis
ε1 , . . . , εn , and f1 , . . . , fN , ϕ1 , . . . , ϕN are dual bases for W and W ′ , then a bilinear
form on V ′ × W ′ can be written
n
X N
X
B(ξ, η) = B( hej , ξiεj , hfk , ηiϕk )
1 1
n X
X N
= B(εj , ϕk )hej , ξihfk , ηi, ξ, η ∈ V ′ × W ′ ,
j=1 k=1
174 APPENDIX A. PREREQUISITES FROM MULTILINEAR ALGEBRA

which means that

n X
X N
B= B(εj , ϕk )ej ⊗ fk .
j=1 k=1

This proves that V ⊗ W is a vector space of dimension dim V dim W with basis ej ⊗ fk ,
j = 1, . . . , n, k = 1, . . . , N . If x1 , . . . , xn are the coordinates of an element in V in
the basis e1 , . . . , en and y1 , . . . , yN are the coordinates of an element in W in the basis
f1 , . . . , fN , then xj yk are the coordinates of x ⊗ y.
If T is a linear map from V ⊗ W to a third vector space Z, then

S : V × W ∋ x, y 7→ T (x ⊗ y) ∈ Z

is a bilinear map. Every bilinear map S : V × W → Z has a unique representation of

this form: If e1 , . . . , en and f1 , . . . , fN are bases in V and W we define T as the linear
mapP with T (ej ⊗ P fk ) = S(ej , fk ) on the basis elements in V ⊗ W , and conclude that if
x= xj ej , y = yk fk then
X X
S(x, y) = xj yk T (ej ⊗ fk ) = T ( xj yk ej ⊗ fk )
X X
= T (( xj ej ) ⊗ ( yk fk )) = T (x ⊗ y).

Thus the bilinear map V × W ∋ x, y 7→ x ⊗ y ∈ V ⊗ W is universal in the sense that all

other bilinear maps from V × W can be factored through it. The dual space of V ⊗ W
is identified with V ′ ⊗ W ′ in a natural way so that hx ⊗ y, ξ ⊗ ηi = hx, ξihy, ηi if x ∈ V ,
ξ ∈ V ′ , y ∈ W and η ∈ W ′ .
The linear transformations L(V, W ) from V to W were identified above with bilinear
forms on W ′ × V , that is, elements of W ⊗ V ′ ,

L(V, W ) ∼
= W ⊗ V ′; in particular, L(V, V ) ∼
= V ⊗ V ′.

We can now look at the trace in a new way. The bilinear form

V × V ′ ∋ x, ξ 7→ hx, ξi

defines a linear map L(V, V ) ∼= V ⊗ V ′ → K, which is precisely the trace.

We can continue to define tensor products of more than two vector spaces. For
example V1 ⊗ V2 ⊗ V3 is the space of trilinear forms on V1′ × V2′ × V3′ ; it is isomorphic to
(V1 ⊗ V2 ) ⊗ V3 and to V1 ⊗ (V2 ⊗ V3 ) and is universal for trilinear maps in V1 × V2 × V3 .
The verification is left as an exercise. Also the trace map can be generalized: If W
and V are finite dimensional vector spaces then the trilinear map

W × V × V ′ ∋ w, x, ξ 7→ whx, ξi ∈ W

defines a linear map W ⊗ V ⊗ V ′ → W called contraction. If Tijk , i = 1, . . . , N ,

j, k = 1, . . . , n, are the coordinates of an element in W ⊗ V ⊗ V ′ with respect to a
basis f1 , . . . , fN in W and dual bases e1 , . . . , en and ε1 , . . . , εn in V and V ′ , then the
 APPENDIX A. PREREQUISITES FROM MULTILINEAR ALGEBRA 175

Pn
coordinates of the contraction are j=1 Tijj . This contraction operation can of course
be applied to any tensor product

W1 ⊗ · · · ⊗ V ⊗ · · · ⊗ V ′ ⊗ Wk

containing two dual vector spaces; contraction gives an element in the tensor product
where they are both removed.
In Riemannian geometry where V is the tangent space and V ′ the cotangent space, it
is customary to put coordinate indices corresponding to factors V in a tensor product as
superscripts and to put indices corresponding to factors V ′ as subscripts. Contraction
then means putting a superscript equal to a subscript and summing over it.
If B is a bilinear form in V ⊗V and a non-degenerate quadratic form G is given in V ,
then we can define the trace of B with respect to G as follows: We have B ∈ V ′ ⊗ V ′ ,
and G gives an identification V ∼ = V ′ , so B can be identified with an element in V ⊗ V ′
for which the trace is defined. In terms
P of kj the matrices (bjk ) and (gjk ) for B and G
with respect to a basis, the trace is bjk g where (g jk ) is the inverse of the matrix
(gjk ).
We have defined the tensor product V ⊗V as the space of bilinear forms on V ′ . Now
every bilinear form on V ′ can be written as the sum of one symmetric and one skew
symmetric bilinear form in one and only one way. We shall denote by S 2 (V ) ⊂ V ⊗ V
the space of symmetric bilinear forms on V ′ and by ∧2 V ⊂ V ⊗ V the space of skew
symmetric bilinear forms on V ′ . Thus the tensor product V ⊗ V is the direct sum of
the symmetric tensor product S 2 (V ) and the exterior product ∧2 V ,

V ⊗ V = S 2 (V ) ⊕ ∧2 V.

For an arbitrary positive integer k we define the symmetric tensor product S k (V ) as

the space of symmetric k linear forms L on V ′ × · · · × V ′ , that is, forms such that for
every permutation π of 1, . . . , k

L(ξ1 , . . . , ξk ) = L(ξπ(1), . . . , ξπ(k)), ξ1 , . . . , ξk ∈ V ′ .

If L ∈ S k (V ), then
e : V ′ ∋ ξ 7→ L(ξ, . . . , ξ)
L
is a homogeneous polynomial of degree k in the sense that

(t1 , . . . , tj ) 7→ L(t
e 1 ξ 1 + · · · + tj ξ j )

for arbitrary ξ1 , . . . , ξj ∈ V ′ is a polynomial in (t1 , . . . , tj ) ∈ Rj ; it suffices of course to

have this for a basis in V ′ . We can recover L(ξ1 , . . . , ξk ) as the coefficient of k!t1 . . . tk
in L(t
e 1 ξ1 + · · · + tk ξk ). Conversely, given a homogeneous polynomial L e in V ′ of degree
k, we can define L(ξ1 , . . . , ξk ) as the coefficient of k!t1 . . . tk in L(t
e 1 ξ1 + · · · + tk ξk ). The
definition is symmetric in ξ1 , . . . , ξk . Since

e 1 ξ + · · · + tk ξ) = (t1 + · · · + tk )k L(ξ),
L(t e
176 APPENDIX A. PREREQUISITES FROM MULTILINEAR ALGEBRA

this implies that L(ξ, . . . , ξ) = L(ξ),

e and since L(te 1 ξ1 + · · · + tk+1 ξk+1 ) has no terms
except
k!t1 . . . tk L(ξ1 , . . . , ξk ) + · · · + k!t2 . . . tk+1 L(ξ2 , . . . , ξk+1 )
not containing the square of any tj , it follows by taking tk+1 = t1 that L(ξ1 , . . . , ξk )
is linear in ξ1 . Thus L ∈ S k (V ), so we have identified S k (V ) with the space of homo-
geneous polynomials of degree k in V ′ ; this is the general meaning of the polarization
discussed above for quadratic forms.
Exercise A.1. Determine the dimension of S k (V ) in terms of dim V and k.
Similarly we define ∧k V as the space of alternating forms L on V ′ , that is, forms
with
L(ξ1 , . . . , ξk ) = sgn πL(ξπ(1), . . . , ξπ(k) ), ξ1 , . . . , ξk ∈ V ′ .
If as before e1 , . . . , en is a basis of V and ε1 , . . . , εn the dual basis of V ′ , then we can
write
X X
L(ξ1 , . . . , ξk ) = L( hξ1 , ej1 iεj1 , . . . , hξk , ejk iεjk )
X X
= hξ1 , ej1 i . . . hξk , ejk iL(εj1 , . . . , εjk ) = dethξi , ejl iki,l=1 L(εj1 , . . . , εjk )/k!.

Thus the forms

(A.1) V k ∋ (ξ1 , . . . , ξk ) 7→ dethxi , ξj iki,j=1 .

give a basis for ∧k V if we restrict x1 , . . . , xk to ej1 , . . . , ejk where e1 , . . . , en is a fixed

basis for V and 1 ≤ j1 < · · · < jk ≤ n. (The linear independence of the forms follows
since
(A.1) is the only one which is not 0 on εj1 , . . . , εjk .) Thus the dimension of ∧k V
is nk , defined as 0 if k > n.
The form (A.1) is denoted by x1 ∧ · · · ∧ xk . The map

V k ∋ (x1 , . . . , xk ) 7→ x1 ∧ · · · ∧ xk

is obviously multilinear and alternating. If W is another vector space and T : ∧k V →

W is a linear map, then

(A.2) S : V k ∋ (x1 , . . . , xk ) 7→ T (x1 ∧ · · · ∧ xk ) ∈ W

is also multilinear and alternating. Conversely, every alternating multilinear map S :

V k → W has as unique representation of the form (A.2). For if e1 , . . . , en is a basis
for V , with dual basis ε1 , . . . , εn for V ′ , then
X X
S(x1 , . . . , xk ) = S( hx1 , εj1 iej1 , . . . , hxk , εjk iejk )
X
= S(ej1 , . . . , ejk ) dethxi , εjl iki,l=1 ,
1≤j1 <···<jk ≤n
X
x1 ∧ · · · ∧ xk = ej1 ∧ · · · ∧ ejk dethxi , εjl iki,l=1
1≤j1 <···<jk ≤n
 APPENDIX A. PREREQUISITES FROM MULTILINEAR ALGEBRA 177

so S is of the form (A.2) if T is the linear map ∧k V → W defined by

T (ej1 ∧ · · · ∧ ejk ) = S(ej1 , . . . , ejk ), 1 ≤ j1 < · · · < jk ≤ n,

for the basis elements in ∧k V . (However, S is of course independent of the choice of

basis.)
The alternating multilinear map

V k+κ ∋ (x1 , . . . , xk+κ ) 7→ x1 ∧ · · · ∧ xk ∧ xk+1 ∧ · · · ∧ xk+κ ∈ ∧k+κ V

is for fixed xk+1 , . . . , xk+κ an alternating multilinear map from V k , so it defines a linear
map
T (xk+1 , . . . , xk+κ ) : ∧k V → ∧k+κ V.
For every w ∈ ∧k V the map

V κ ∋ (xk+1 , . . . , xk+κ ) 7→ T (xk+1 , . . . , xk+κ )w ∈ ∧k+κ V

is alternating and multilinear so it defines a linear map T ′ (w) : ∧κ V → ∧k+κ V . It

depends linearly on w, so

∧k V × ∧κ V ∋ (w, w′ ) 7→ T ′ (w)w′ ∈ ∧k+κ V

is a bilinear map. It is characterized by the fact that it maps x1 ∧ · · · ∧ xk ∈ ∧k V and

xk+1 ∧ · · · ∧ xk+κ ∈ ∧κ V to x1 ∧ · · · ∧ xk+κ ∈ ∧k+κ V . With the notation ∧ also for
this map, we have thus found that there is a uniquely defined bilinear map

(A.3) ∧k V × ∧κ V ∋ (w, w′ ) 7→ w ∧ w′ ∈ ∧k+κ V

such that

(A.4) (x1 ∧ · · · ∧ xk ) ∧ (xk+1 ∧ · · · ∧ xk+κ ) = x1 ∧ · · · ∧ xk+κ , x1 , . . . , xk+κ ∈ V.

It is called the exterior product. It makes

n
M
∗
∧ V = ∧k V,
k=0

where ∧0 V = K, an associative algebra. Since

x1 ∧ · · · ∧ xk ∧ xk+1 ∧ · · · ∧ xk+κ = (−1)kκ xk+1 ∧ · · · ∧ xk+κ ∧ x1 ∧ · · · ∧ xk ,

the algebra is not commutative but we have

(A.5) w ∧ w′ = (−1)kκ w′ ∧ w, if w ∈ ∧k V, w′ ∈ ∧κ V.
178 APPENDIX A. PREREQUISITES FROM MULTILINEAR ALGEBRA

Note that w and w′ commute unless the degrees k and κ of w and w′ are both odd;
then they anticommute.
If V1 and V2 are finite dimensional vector spaces, then every linear map A : V1 → V2
induces a linear map ∧k A : ∧k V1 → ∧k V2 . In fact, the map

V1k ∋ (x1 , . . . , xk ) 7→ (Ax1 ) ∧ · · · ∧ (Axk ) ∈ ∧k V2

is alternating and multilinear so there is a unique linear map ∧k A : ∧k V1 → ∧k V2 such

that

(A.6) (∧k A)(x1 ∧ · · · ∧ xk ) = (Ax1 ) ∧ · · · ∧ (Axk ), x1 , . . . , xk ∈ V.

We can define the exterior powers ∧k V ′ of the dual V ′ of V in the same way. The
multilinear map

V k × V ′k ∋ (x1 , . . . , xk , ξ1 , . . . , ξk ) 7→ dethxi , ξj iki,j=1

is alternating in the first k and in the last k variables. Hence it induces a bilinear form
on ∧k V × ∧k V ′ . It defines a duality for if e1 , . . . , en and ε1 , . . . , εn are dual bases for V
and for V ′ , then the bases ej1 ∧ · · · ∧ ejk and εi1 ∧ · · · ∧ εik with 1 ≤ j1 < · · · < jk ≤ n
and 1 ≤ i1 < · · · < ik ≤ n are dual with respect to this form. Thus we have a natural
duality between ∧k V and ∧k V ′ such that

(A.7) hx1 ∧ · · · ∧ xk , ξ1 ∧ · · · ∧ ξk i = dethxi , ξj iki,j=1 , x1 , . . . , xk ∈ V, ξ1 , . . . , ξk ∈ V ′ .

(Many authors use another definition where the right-hand side is divided by k!, for
that is the duality inherited from the natural duality of the tensor products. See
Sternberg [1, p. 19] for a discussion of this point. Kobayashi and Nomizu [1] use the
division by k!, which should be kept in mind when comparing identities.)
 APPENDIX B. THE CALCULUS OF DIFFERENTIAL FORMS

Let M be a C ∞ manifold of dimension n. Then the exterior power ∧k Tx∗ M of the

n
cotangent space at x ∈ M is a real vector space of dimension k . If M ⊂ Rn , then

Tx∗ M ∼= Rn , and ∧k Tx∗ M has the basis dxj1 ∧ · · · ∧ dxjk , where 1 ≤ j1 < · · · < jk ≤ n;
this is the alternating multilinear form

(Rn )k ∋ (t1 , . . . , tk ) 7→ det(ti jl )ki,l=1 .

For any C 1 map f : N → M , where N is another smooth manifold, the adjoint of the
differential
f ′ (y) : Ty N → Tf (y) M
is a map Tf∗(y) M → Ty∗ N inducing a map

(B.1) ∧k Tf∗(y) M → ∧k Ty∗ N,

which is bijective if f ′ (y) is bijective. In particular, if f is a coordinate system, Rn ⊃

ω → f (ω) ⊂ M , we obtain an identification of x∈f (ω) ∧k Tx∗ M and ω×∧k Rn . Thus the
S

vector spaces ∧k Tx∗ M define a vector bundle ∧k T ∗ M over M . In a coordinate patch

with local coordinates x1 , . . . , xn , a local frame for ∧k T ∗ M is given by the exterior
products
dxj1 ∧ · · · ∧ dxjk , 1 ≤ j1 < · · · < jk ≤ n.
We shall denote the space of sections of ∧k T ∗ M by λk (M ); in terms of local coordinates
a section u can thus be written
X
(B.2) u= aj1 ...jk (x)dxj1 ∧ · · · ∧ dxjk .
1≤j1 <···<jk ≤n

Exterior multiplication of forms is done just by formal multiplication observing the

anticommutation rule dxi ∧ dxj = −dxj ∧ dxi .
If f : N → M is a smooth map, then the map (B.1) defines for every k a linear map

f ∗ : λk (M ) → λk (N ).

One calls f ∗ u ∈ λk (N ) the pullback of u ∈ λk (M ), from M to N . If x and y are local

coordinates in M and N and u has the form (B.2), x = f (y), then
X
(B.3) f ∗u = aj1 ...jk (f (y))df j1 (y) ∧ · · · ∧ df jk (y),
1≤j1 <···<jk ≤n
179
180 APPENDIX B. THE CALCULUS OF DIFFERENTIAL FORMS

where df j (y) =
P j
∂f (y)/∂y idy i . In fact, if ϕ is a function on M (such as a local
coordinate), then

hf ∗ dϕ, ti = hdϕ, f ′ ti = hd(ϕ ◦ f ), ti = hd(f ∗ ϕ), ti, t ∈ T N,

by the chain rule, which means that f ∗ dϕ = d(f ∗ ϕ). We have natural rules of com-
putation such as (f g)∗ = g ∗ f ∗ ; the notation with the upper star is meant to be a
reminder of this reversal of factors.
We shall now define the exterior differential of a k form. We do this first in a
fixed local coordinate system and verify the independence of the chosen coordinates
afterwards. For the form u in (B.2) we thus define
X
(B.4) du = daj1 ...jk (x) ∧ dxj1 ∧ · · · ∧ dxjk .
1≤j1 <···<jk ≤n

It follows at once from this definition that

(B.5) d(u ∧ v) = (du) ∧ v + (−1)k u ∧ dv, u ∈ λk , v ∈ λκ ;

in view of the linearity it suffices to verify (B.5) when u = adxj1 ∧ · · · ∧ dxjk and
v = bdxjk+1 ∧ · · · ∧ dxjk+κ , and then it follows from the fact that d(ab) = adb + bda;
commutation of db through u gives the sign (−1)k . If u is a smooth function we have

(B.6) d2 u = 0.

∂j udxj , so
P
In fact, du =
X X
d2 u = d(∂j u) ∧ dxj = ∂k ∂j udxk ∧ dxj = 0

since dxk ∧ dxj is antisymmetric in k and j while ∂k ∂j u is symmetric. In view of (B.2)

and (B.5) it follows at once that (B.6) is valid for u ∈ λk for any k.
Repeated use of (B.5) shows that for arbitrary smooth functions f0 , . . . , fk we have

d(f0 df1 ∧ · · · ∧ dfk ) = df0 ∧ df1 ∧ · · · ∧ dfk .

If u ∈ λk (ω), where ω is an open subset of Rn , and f : ω ′ → ω is a smooth map from

′
an open subset ω ′ of Rn , it follows that

(B.7) f ∗ (du) = df ∗ u, u ∈ λk (ω).

In fact, if u is given by (B.2), then du is given by (B.4) and f ∗ u is given by (B.3),

hence X
df ∗ u(x) = daj1 ...jk (f (y)) ∧ df j1 (y) ∧ · · · ∧ df jk (y),
1≤j1 <···<jk ≤n

which proves (B.7). In particular, it follows from (B.7) that our definition (B.4) of the
exterior differential is independent of the choice of local coordinates.
Locally there is a converse of (B.6):
 APPENDIX B. THE CALCULUS OF DIFFERENTIAL FORMS 181

Theorem B.1 (Poincaré’s lemma). Let v ∈ λk+1 (X) where X is a convex open
set in Rn , assume that the coefficients of v are in C µ and that dv = 0. Then there is
a form u ∈ λk (X) with C µ coefficients such that du = v.
Proof. We may assume that 0 ∈ X. Then X b = {(x, t) ∈ Rn × R; tx ∈ X} is an
open set containing X × [0, 1], and f (x, t) = tx is a C ∞ map from X b to X, so we can
form
f ∗ v = ft∗ v + dt ∧ wt
where ft∗ is the pullback of v when t is regarded as a parameter, so it is a differential
form which only involves the differentials of the coordinates in X. Since df ∗ v = f ∗ dv =
0 it follows that
0 = dt ∧ (∂(ft∗ v)/∂t − dx wt ) + . . .
where the dots indicate a form which has no factor dt and dx wt is the differential of
wt for fixed t. Hence
∂(ft∗ v)/∂t = dx wt ,
and integration from 0 to 1 gives
Z 1
f1∗ v − f0∗ v = du, u= wt dt.
0

But f1∗ v = v and f0∗ v = 0 since f1 is the identity and f0 maps X to {0}, which proves
the theorem.
Poincaré’s lemma reflects the fact that the topology of a convex set in Rn is very
simple; in general there is a topological obstruction:
Definition B.2. If X is a smooth manifold then the quotient H k (X) of the closed
forms {u ∈ λk (X); du = 0} by the linear subspace {dv; v ∈ λk−1 (X)} of exact forms,
k > 0, is called the de Rham cohomology of degree k. The residue class in H k (X) of
a form u ∈ λk (X) with du = 0 is called the cohomology class of u.
Exercise B.1. Prove that if M and N are smooth manifolds, f : M × [0, 1] → N
is a smooth map, and u ∈ λk (N ) is a closed form, then the forms f (·, t)∗u ∈ λk (M )
are in the same cohomology class for all t ∈ [0, 1].
Poincaré’s lemma is closely related to Stokes’ formula, which we shall now discuss.
First we define the integral of a form u ∈ λn (M ) with compact support over M when
M is an oriented manifold of dimension n. To do so we first assume that the support
is contained in a local coordinate patch with positively oriented local coordinates x ∈
ω ⊂ Rn . In terms of these coordinates we can then write

u = a(x)dx1 ∧ · · · ∧ dxn

and we define Z Z
u= a(x)dx,
182 APPENDIX B. THE CALCULUS OF DIFFERENTIAL FORMS

where dx is the Lebesgue measure. If the support of u is also contained in a coordinate

patch with positively oriented local coordinates y, then

u = a(x(y))dx1(y) ∧ · · · ∧ dxn (y) = a(x(y)) det(∂xi (y)/∂y j )ni,j=1

R
so in terms of these coordinates u should be
Z
a(x(y)) det(∂xi (y)/∂y j )ni,j=1 dy.

The positive orientation means precisely that the Jacobian here is positive, so Rthe
definitions in terms of the two coordinate systems
P agree. Now we can define u
in general by using a partition of unity 1 =
P ϕj to split u into a finite sum u =
ϕj u where each term has support in a local coordinate patch and using these local
coordinates set
XZ
u= ϕj u.

The definition is independent of the choice of partition of unityP and corresponding

local coordinates, for if we have another partition
R of unity 1 = ψk with each term
supported by a coordinate patch, then ψk ϕj u has the same value if we use the
coordinates associated with ϕj or those associated with ψk , and summing over k or
over j we conclude that the definition using the partition of unity {ϕj } is equivalent
to that using {ψk }.

Remark. We have not given a precise definition of orientation above. One way
to do so is to say that an orientation on a manifold of dimension n is a n form o,
the orientation form, which is everywhere different from 0; orientation forms differing
by multiplication with a positive function are considered equivalent. A system of
local coordinates x1 , . . . , xn is then positively oriented if with these coordinates o =
g(x)dx1 ∧ · · · ∧ dxn where g(x) > 0. The manifold is said to be oriented by o > 0.

The integral over M of a k form is defined as 0 if k < n, so we have a definition

of the integral of any form with compact support. The definition of the integral over
a submanifold N of M is an immediate consequence: For a RsubmanifoldR ∗ we have an
∗
embedding i : N → M , and for a form u on M we define N u = N i u if i u has
compact support.
Now assume that X is an open subset of the oriented manifold M , with a C ∞
boundary ∂X for the sake of simplicity. At any boundary point we can then choose
local coordinates in M varying over the unit ball B in Rn , say, such that the points
in X correspond to B− = {x ∈ B; x1 < 0}. If u ∈ λn−1 (B) has compact support in B,
it follows that
Z Z
(B.8) u= du,
B0 B−
 APPENDIX B. THE CALCULUS OF DIFFERENTIAL FORMS 183

where B0 = {x ∈ B; x1 = 0}. In fact,

n
X
u= (−1)j−1 uj (x)dx1 ∧ · · · ∧ dxj−1 ∧ dxj+1 ∧ · · · ∧ dxn ,
1
n
X
du = ( ∂j uj )dx1 ∧ · · · ∧ dxn ,
1

so the statement is that with x′ = (x2 , . . . , xn )

Z Z n
X
′ ′
u1 (0, x ) dx = ∂j uj dx,
x1 <0 1

which is obvious.
Now we can orient ∂X uniquely by using at points in ∂X positively oriented local
coordinates in M such that X is located as just described where x1 < 0; the coordinates
x2 , . . . , xn in ∂X are then by definition positively oriented. From the local formula
(B.8) it follows by using a partition of unity that
Z Z
(B.9) u= du, Stokes’ formula,
∂X X

if u is a form with compact support in M . The same formula holds if X is an open

subset with C 1 boundary in an oriented submanifold of M ; this is an immediate
consequence in view of the definition of integration over a submanifold.
Given a k + 1 form v with
R dv = 0 it follows from (B.9) that there does not exist a k
form u with du = v unless X v = 0 for any compact oriented submanifold X without
boundary. This is the starting point of homology theory, but it would take us too far
to pursue the matter here.
184 APPENDIX B. THE CALCULUS OF DIFFERENTIAL FORMS
 APPENDIX C. THE FROBENIUS THEOREM

To study flat structures we shall need the Frobenius existence theorem for first order
systems of differential equations. One of the standard forms of this result is as follows:
Theorem C.1. Let v1 , . . . , vr be C ∞ vector fields in a neighborhood of 0 in Rn
such that

(C.1) v1 (0), . . . , vr (0) are linearly independent,

Xr
(C.2) [vi , vj ] = cijk vk , i, j = 1, . . . , r,
k=1

where [v, w] denotes the commutator of v and w, and cijk ∈ C ∞ . Then there exist new
local coordinates y1 , . . . , yn in a neighborhood of 0 such that
r
X
∂/∂yi = bij vj , i = 1, . . . , r.
j=1

Thus the solutions of the equations vj u = 0, j = 1, . . . , r are in a neighborhood of the

origin precisely the functions of yr+1 , . . . , yn .
Proof. The proof, which can be found in Hörmander [1, Appendix C1], is by
induction with respect to n. Invariance under change of variables and non-singular
recombinations of the vj shows that one may assume that v1 = ∂/∂x1 while

n
X
vj = vjl ∂/∂xl , j = 2, . . . , r.
l=2

By the inductive hypothesis we may also assume that vjl = 0 for j = 2, . . . , r if l > r
and x1 = 0. By (C.1)
r
X r
X
∂vjl /∂x1 = [v1 , vj ]xl = c1jk vk xl = c1jk vkl , j = 2, . . . , r.
k=2 k=2

Hence vjl = 0 in a neighborhood of 0 if l > r since this is true when x1 = 0, which

proves the theorem.
185
186 APPENDIX C. THE FROBENIUS THEOREM

The geometric interpretation is as follows: The vector fields v1 , . . . , vr span at every

point x a linear space Fx of dimension r in the tangent space. The condition (C.2)
is the necessary and sufficient condition in order that through every point there is a
manifold M of dimension r such that the tangent plane of M is Fx at every x ∈ M .
Now we can also define the linear space Fx by means of its orthogonal spaces Fx⊥ in
the cotangent space, that it, by n − r conditions

(C.3) ωj = 0, j = r + 1, . . . , n,

where ωr+1 , . . . , ωn are linearly independent differential one forms at the origin. We
want to rewrite the Frobenius condition (C.2) in terms of these forms. To do so we
need a lemma:
Lemma C.2. If X and Y are C 1 vector fields and ω is a C 1 one form, then

(C.4) hX ∧ Y, dωi = h[Y, X], ωi + XhY, ωi − Y hX, ωi.

Proof. If ω = du for some u then (C.4) is the definition of the commutator vector
field. If ω is replaced by ϕω, then both sides are multiplied by ϕ and in addition we
get on the left-hand side a term hX ∧ Y, dϕ ∧ ωi which is equal to the additional term

hX, dϕihY, ωi − hY, dϕihX, ωi

in the right-hand side. Hence (C.4) follows in general.

Let us now return to the Frobenius theorem. In a neighborhood of the origin we
can extend the system of vector fields v1 , . . . , vr to a basis v1 , . . . , vn . Let ωj be the
biorthogonal basis of one forms, that is,

hvj , ωk i = δjk , j, k = 1, . . . , n.

Then
n
X n
X
(C.5) [vi , vj ] = cijk vk , i, j = 1, . . . , n =⇒ dωk = − 12 cijk ωi ∧ ωj ,
k=1 i,j=1

for it follows from (C.4) that

X
hvj ∧ vi , dωk i = cijl hvl , ωk i + vj hvi , ωk i − vi hvj , ωk i
X
= cijk = 12 hvi ∧ vj , cµνk ωµ ∧ ων i.

We can reverse the implication in (C.5) if we demand that cijk is antisymmetric in i, j

(which is automatic in the left-hand side). The Frobenius condition (C.2) stating that
cijk = 0 if k > r and i, j ≤ r is therefore equivalent to

(C.2)′ dωj is in the ideal of forms generated by ωr+1 , . . . , ωn , if j > r.

 APPENDIX C. THE FROBENIUS THEOREM 187

Another equivalent formulation of the Frobenius theorem concerns complete inte-

grability of a system of differential equations

(C.6) ∂yµ (x)/∂xν = Fµν (x, y), ν = 1, . . . , n, µ = 1, . . . , m,

where Fµν are smooth functions defined in a neighborhood of (x0 , y0 ) ∈ Rn+m . The
system is called completely integrable if there is a smooth solution in a neighborhood
of x0 with y(x0 ) given in a neighborhood of y0 . Since the equations give
m
X
∂ 2 yµ /∂xν ∂xκ = ∂Fµν /∂xκ + ∂Fµν /∂yσ Fσκ
σ=1

a necessary condition is that in a neighborhood of (x0 , y0 ) we have for µ = 1, . . . , m

and ν, κ = 1, . . . , n
m
X m
X
(C.7) ∂Fµν /∂xκ + ∂Fµν /∂yσ Fσκ = ∂Fµκ /∂xν + ∂Fµκ /∂yσ Fσν .
σ=1 σ=1

This condition is also sufficient. In fact, the equations (C.6) mean precisely that the
graph of y(x) shall be defined by the equations
n
X
ωµ = dyµ − Fµν dxν = 0.
ν=1

Since
X X X
dωµ = − ∂Fµν /∂xκ dxκ ∧ dxν − ∂Fµν /∂yσ (ωσ + Fσκ dxκ ) ∧ dxν
κ,ν
X
=− ∂Fµν /∂yσ ωσ

when (C.7) holds, the condition (C.2)′ is fulfilled. Hence we have:

Theorem C.3. For the system (C.6) to have a solution in a neighborhood of x0 with
y(x0 ) arbitrarily prescribed near y0 it is necessary and sufficient that the integrability
condition (C.7) is valid near (x0 , y0 ). The solution is then unique.
188 APPENDIX C. THE FROBENIUS THEOREM
 APPENDIX D. INVARIANTS OF O(n)

A polynomial f (x1 , . . . , xN ) in xj ∈ V = Rn , j = 1, . . . , N , is called a O(n) invariant

if

(D.1) f (Ox1 , . . . , OxN ) = f (x1 , . . . , xN ), O ∈ O(n).

Sometimes f is then called an even invariant, and one calls f an odd invariant if

(D.2) f (Ox1 , . . . , OxN ) = det Of (x1 , . . . , xN ), O ∈ O(n).

In both cases we have

(D.3) f (Ox1 , . . . , OxN ) = f (x1 , . . . , xN ), O ∈ SO(n).

where SO(n) = {O ∈ O(n); det O = 1}. On the other hand, it follows from (D.3) that

f (Ox1 , . . . , OxN ) = g(x1 , . . . , xN ), O ∈ O(n), det O = −1,

where g is independent of the choice of O and satisfies (D.3). Hence f = 12 (f + g) +

1
2 (f − g) is the sum of one even and one odd invariant.

Theorem D.1. Every even invariant f (x1 , . . . , xN ) is a polynomial in the scalar

products (xj , xk ), j, k = 1, . . . , n. Every odd invariant is a sum of such polynomials
multiplied by the determinant det(xi1 , . . . , xin ) of n vectors xi . In particular, there are
no odd invariants 6= 0 if N < n.
Proof. The proof is by induction. The statement is obvious when n = 1 so we
assume that n > 1 and that the theorem has already been proved when n is replaced
by n − 1. Let V1 = Rn−1 , identified with the plane {x ∈ Rn ; xn = 0}. If we choose
for O the reflection in V1 which just changes the sign of xn and leaves V1 fixed, it
follows that f (x1 , . . . , xN ) = 0 if all xj ∈ V1 and f is an odd invariant; if f is an
even invariant then f (x1 , . . . , xN ) = p((xj , xk )j,k=1,...,N ) by the inductive hypothesis.
Since every hyperplane can be carried to the special position V1 by a transformation
in SO(n), these statements remain true for arbitrary x1 , . . . , xN which do not span
V . In particular, the theorem is true if N < n. When N ≥ n the proof requires an
identity due to Capelli which will now be discussed.
The Capelli identity concerns the differential operators

Dj,k = hxj , ∂/∂xk i

189
190 APPENDIX D. INVARIANTS OF O(n)

which operate on polynomials f (x1 , . . . , xN ) in N vectors x1 , . . . , xN ∈ V = Rn . If

f is homogeneous in each vector xj , then Dj,k increases the degree in xj by 1 and
decreases that in xk by 1, if j 6= k, but changes no degree if j = k. We shall write
Dj,k Dj ′ ,k′ for the standard composition of the operators, but write Dj,k #Dj ′ ,k′ for
the formal (commutative) multiplication, defined as if the coefficients were constant.
Capelli’s identity states that

DN,N + N − 1 DN,N−1 . . . DN,1

DN−1,N DN−1,N−1 + N − 2 . . . DN,1
(D.4) .. .. .. .. f
. . . .
D1,N D1,N−1 . . . D1,1
0, if N > n

=
det(x1 , . . . , xn ) det(∂/∂x1 , . . . , ∂/∂xn )f, if N = n.

The diagonal elements are here Dk,k + (k − 1), the off diagonal elements are Dj,k , and
the expansion of the determinant shall be made so that the elements are multiplied in
the order of their columns. To prove (D.4) we first consider the “minors” formed from
the last two columns. We have

Dj,2 Dk,1 − Dk,2 Dj,1 = Dj,2 #Dk,1 − Dk,2 #Dj,1 + δk,2 Dj,1 − δj,2 Dk,1 ,

that is,

(Dj,2 + δj,2 )Dk,1 − (Dk,2 + δk,2 )Dj,1 = Dj,2 #Dk,1 − Dk,2 #Dj,1 .

Thus the minors of (D.4) in the left-hand side are the same as in a determinant with
diagonal elements Dk,k expanded using the formal product. We shall prove by induc-
tion for increasing k that this is true for the minors taken from the last k ≥ 2 columns.
When k = N this will prove (D.4), for

0, if N > n

dethxj , ξk iN
j,k=1 =
det(x1 , . . . , xn ) det(ξ1 , . . . , ξn ), if N = n.

By the inductive hypothesis we have to calculate

(Dik ,k + (k − 1)δik ,k )Dik−1 ,k−1 # . . . #Di1 ,1

= Dik ,k #Dik−1 ,k−1 # . . . #Di1 ,1 + (k − 1)δik ,k Dik−1 ,k−1 # . . . #Di1 ,1
k−1
X
+ δiν ,k Dik−1 ,k−1 # . . . #Diν+1 ,ν+1 #Dik ,ν #Diν−1 ,ν−1 # . . . #Di1 ,1 .
ν=1

We shall let i1 , . . . , ik run through all permutations of k fixed indices, multiply by the
sign of the permutation and sum. Now the permutations

iν , ik−1 , . . . , iν+1 , ik , iν−1 , . . . , i1 and ik , ik−1 , . . . , i1

 APPENDIX D. INVARIANTS OF O(n) 191

have opposite signs since they differ by one inversion. Hence the terms in the last sum
will cancel the preceding one, which completes the proof of the inductive statement
and hence of the Capelli identity.
End of Proof of Theorem D.1. With the Capelli identity available we can now
finish the proof. By the first part of the proof we may assume that N ≥ n and that
the theorem has already been proved when N is replaced by N − 1. Furthermore we
may assume that f is homogeneous in each of the vectors x1 , . . . , xN , for if we split f
into a sum of polynomials of such separate homogeneities, then each term must be an
PN
invariant. Let rj be the degree of homogeneity with respect to xj , and let |r| = 1 rj
be the total degree of homogeneity. We order all multiindices r = (r1 , . . . , rN ) first
for increasing |r|, and then lexocographically for fixed |r|, that is, first according to
increasing r1 , then for fixed r1 according to increasing r2 , and so on. If r1 = 0 then
f depends on N − 1 vectors so the theorem is true then by inductive hypothesis. By
still another inductive argument we may therefore assume that r1 > 0 and that the
theorem has already been proved for lower values of r.
Now consider (D.4). If N = n, then g = det(∂/∂x1 , . . . , ∂/∂xn)f is an odd (even)
invariant if f is an even (odd) invariant, and the degree of g is lower than the degree
of f , so we know that g has the form stated in the theorem. Since

det(x1 , . . . , xn )2 = det(xj , xk )nj,k=1 ,

it follows that the right-hand side has the stated form for any N ≥ n. In the left-hand
side of (D.4) the diagonal term is cf where c = r1 (r2 + 1) . . . (rN + N − 1) 6= 0 since
r1 6= 0. All other terms contain some factor Dj,k with j 6= k. For a given term we
choose the factor Dj,k with j 6= k which is furthest to the right. To the right of it we
then have only the factors Dk−1,k−1 . . . D1,1 which multiply f by r1 . . . (rk−1 + k − 2).
We must have j > k since two factors must not be chosen in the same row. The full
expression of the term is of the form ajk Dj,k f where ajk is a product of operators
Dp,q . But Dj,k f is also an invariant and its total degree is equal to that of f while
its lexicographic order is lower since Dj,k lowers the degree in xk at the expense of
a raise of degree in xj for some j > k. Thus Dj,k has the desired form by inductive
hypothesis. So has ajk Dj,k f , for

Dp,q (xr , xs ) = (xp , xs )δq,r + (xp , xr )δq,s ,

X
Dp,q det(xi1 , . . . , xin ) = δq,ij det(xi1 , . . . , xij−1 , xp , xij+1 , . . . , xin ).

Hence this is also true of f , which completes the proof.

192 APPENDIX D. INVARIANTS OF O(n)
 NOTES AND REFERENCES

Chapter 1. Starting with the notion of curvature for a curve, which is as old as
calculus, we study the tangent and normals associated to a curve which leads up to a
first encounter with the idea of moving orthogonal frames. It goes back to Euler and
is basic for the methods of E. Cartan presented in Chapter IV. We introduce it in the
elementary context of curves in a Euclidean space and take the opportunity to give at
the same time a first introduction to Lie groups, restricted to subgroups of the linear
group. A more extensive discussion is given in Section 5.3, and the reader who wants
more information is referred to Chevalley [1], Helgason [1], Sternberg [1], Warner [1]
and Weyl [1].
Chapters 2, 3. Principal curvatures of surfaces in R3 were also known to Euler,
but it was Gauss [1] who discovered that the total curvature is an inner property of the
surface. Reading Gauss [1] is still very pleasant, and one will recognize many of the
ideas presented in Chapters 2 and 3 there. The extension to higher dimensions outlined
in Riemann [1], with emphasis on ideas rather than formulas, is equally enjoyable to
read. The later formal development of his ideas by Bianchi, Christoffel, Ricci and
others is the main theme of Chapters 2 and 3. Our presentation is in the same spirit as
Klingenberg [1,2] and to some extent the classic Eisenhart [1,2]. One can find a good
presentation in Berger, Gauduchon and Mazet [1], and excellent summaries are given in
Aubin [1], Besse [1]. Chapter 2 ends with some detailed analysis of the curvature tensor
at a point taken from Atiyah, Hitchin and Singer [1], where much more information is
available.
To give a bridge between the study of submanifolds of Rn in Chapter 2 and abstract
manifolds in Chapter 3 we have devoted much space in Chapter 3 to the problem of
embedding an abstract manifold in a Euclidean space. After the classical results of
Cartan and Janet for the analytic case, the big advance on the problem was made by
Nash [1,2]; recently a technically simpler variation has been found by Günther [1,2].
For additional information on global embedding theorems we refer to Gromov and
Rohlin [1], Griffiths [1], Griffiths and Jensen [1], Jacobowitz [1,2] and references in
these papers.
The discussion of spaces of constant curvature in Section 3.3 has been taken to
a large extent from Kobayashi and Nomizu [1]. This is a very careful and useful
presentation of basic differential geometry, though provided with very little motivation
for the reader.
Chapter 4. In this chapter we leave the Ricci tensor calculus which dominated
Chapter 3 and introduce differential forms in the spirit of E. Cartan. The reader can
193
194 NOTES AND REFERENCES

consult Chern [1,2], Kobayashi and Nomizu [1] and Sternberg [1] for a more thorough
discussion of these topics. Section 4.3 is almost entirely taken from Chern [1].
Chapter 5. We have started the chapter with a discussion of connections in a vector
bundle, as they are encountered naturally by an analyst. However, at the end we also
give the more geometrical approach using connection forms on a principal bundle. The
purpose of this material is to give the language required to state index theorems for
elliptic differential operators and to give the basic concepts required in gauge theory.
Also in this chapter one can consult Kobayashi and Nomizu [1] or Sternberg [1] for
most of the topics covered.
Chapter 6. The term “metric operator” is taken from Günther [3], and Proposition
6.1.2 can also be found there. The Hadamard construction of parametrices (Hadamard
[1], see also e.g. Hörmander [1, Chapter 17]) is made more transparent by using the
connection provided by the operator itself, and this is useful in the proof of the local
index theorem in Section 6.10. Section 6.2 is devoted to the algebra of differential
forms, and Hodge theory is presented in Section 6.3. For a more detailed discussion
of differential forms on a Riemannian manifold including Hodge theory, the reader can
also consult de Rham [1] or Warner [1]. In Section 6.4 the Hadamard construction of
parametrices is adapted to heat equations associated to metric operators. This gives
the analytical tools required for the discussion of Hirzebruch’s signature theorem in
Section 6.5. (For additional background to this result one should consult Hirzebruch
[1].) The presentation here follows Atiyah, Bott and Patodi [1] in principle, but we
have substituted the parametrix construction of Hadamard for the application of pseu-
dodifferential operators. Sections 6.6–6.10 are devoted to a direct proof of the local
index formula for Dirac operators, following Bismut [1] and particularly Getzler [1, 2].
The presentation owes much to Roe [1], and we have also benefited from some lecture
notes of M. Taylor.
Unwritten chapters. One of the possible directions for a continuation is to cover
the solution of the Yamabe problem, following the excellent exposition in Lee and
Parker [1]. After this introduction to non-linear problems in geometry one might
study some gauge theory. Another natural direction is to study pseudo-Riemannian
manifolds of Lorentz signature. There are a number of papers exploiting conformal
invariance to prove global existence theorems for non-linear hyperbolic systems, most
recently Christodoulou and Klainerman [1] where the stability of Minkowski space un-
der small perturbations is proved. This requires of course some preparations concerning
general relativity theory. The classical survey by Pauli [1] is still very readable. More
recent information can be found in Bergmann [1], but this reference is directed towards
mathematical physicists rather than mathematicians interested in physics. With some
background from these sources, including the Schwarzschild solution, the way is also
open to discuss the interesting proof of the positive mass conjecture in Schoen and Yau
[1,2], Witten [1] and Parker and Taubes [1].
 Bibliography
M. F. Atiyah [1], Classical groups and classical differential operators on manifolds, Differential op-
erators on manifolds, Edition Cremonese, Roma, 1975, CIME conference in Varenna September
1975..
M. F. Atiyah, R. Bott and V. K. Patodi [1], On the heat equation and the index theorem, Inv. math.
19 (1973), 279–330, Correction 28 (1975), 277–280..
M. F. Atiyah, R. Bott and A. Shapiro, Clifford modules, Topology 3 (1964), 3–38, Suppl. 1..
M. F. Atiyah, N. J. Hitchin and I. M. Singer [1], Self-duality in four-dimensional Riemannian geom-
etry, Proc. R. Soc. Lond. A. 362 (1978), 425–461.
T. Aubin [1], Nonlinear analysis on manifolds. Monge-Ampère equations, Springer-Verlag, New York,
Heidelberg, Berlin, 1982.
E. Berger, R. Bryant and Ph. Griffiths [1], The Gauss equations and rigidity of isometric embeddings,
Duke Math. J. 50 (1983), 803–892.
M. Berger, P. Gauduchon and E. Mazet [1], Le spectre d’une variété Riemannienne, Springer-Verlag,
1971, SLN 194..
P. G. Bergmann [1], The general theory of relativity, Handbuch der Physik IV, Springer-Verlag, 1962.
A. L. Besse [1], Einstein Manifolds, Springer-Verlag, 1987, Erg. d. Math. Band 10..
J.M. Bismut [1], The Atiyah-Singer theorem for classical elliptic operators: a probabilistic approach.
I. The index theorem., J. Func. Anal. 57 (1984), 56–99.
Yu. D. Burago and V. A. Zalgaller [1], Geometric inequalities, Springer-Verlag, 1988, Grundl.d.Math.
Wiss. 285..
S. S. Chern [1], On the curvature integral in a Riemannian manifold, Ann. of Math. 46 (1945),
674–684.
S.S. Chern [2], Topics in differential geometry, Inst. for Adv. Study, Princeton, N.J., 1951.
C. Chevalley [1], Theory of Lie groups I, Princeton University Press, Princeton, N.J., 1946.
D. Christodoulou and S. Klainerman [1], The global nonlinear stability of the Minkowski space, Ann.
of Math., Manuscript of 565 pages, to appear..
H. L. Cycon, R. G. Froese, W. Kirsch and B. Simon [1], Schrödinger operators with applications
to quantum mechanics and global geometry, Springer-Verlag, 1987, Texts and Monographs in
Physics..
N. V. Efimov [1], Generation of singularities on surfaces of negative curvature, Mat. Sb. 64 (1964),
286–320, (Russian; translation in AMS translations.).
L. Eisenhart [1], Introduction to differential geometry, Princeton University Press, Princeton, N. J.,
1940.
L. Eisenhart [2], Riemannian geometry, Princeton University Press, Princeton, N. J., 1926.
C. F. Gauss [1], Disquisitiones generales circa superficies curvas, Comm. soc. reg. scient. Gött. 6
(1823–27), German translation in Ostwald’s Klassiker der exakten Wissenschaften 5(1889), Eng-
lish translation in “General investigations of curved surfaces”, Raven Press, Hewlett, N.Y. (1965),
which also contains an earlier version..
J. C. H. Gerretsen [1], Lectures on tensor calculus and differential geometry, P. Noordhoff N. V.,
Groningen, 1962.
E. Getzler [1], Pseudodifferential operators on supermanifolds and the Atiyah-Singer index theorem,
Comm. Math. Phys. 92 (1983), 163–178.
E. Getzler [2], A short proof of the local Atiyah-Singer index theorem, Topology 25 (1986), 111–117.
P. Gilkey [1], Curvature and the eigenvalues of the Laplacian for elliptic complexes, Adv. in Math.
10 (1973), 344–382.
Ph. A. Griffiths [1], Exterior differential systems and the calculus of variations, Birkhäuser, Boston,
Basel, Stuttgart, 1983, Progress in Math. 25..
Ph. A. Griffiths and G. R. Jensen [1], Differential systems and isometric embeddings, Princeton
University Press, Princeton, N. J., 1987, Annals of Math. Studies 114..
M. L. Gromov and V. A. Rokhlin [1], Embeddings and immersions in Riemannian geometry., Russian
Math. Surveys 25:5 (1970), 1–57, Russian original in Usp. Mat. Nauk 25:5(155), 3–62 (1970)..
195
196 BIBLIOGRAPHY

M. Günther [1], On the perturbation problem associated to isometric embeddings of Riemannian

manifolds, Ann. Global Anal. Geom. 7 (1989), 69–77.
M. Günther [2], Zum einbettungssatz von J. Nash, Math. Nachr. 144 (1989), 165–187.
P. Günther [3], Huyghens’ principle and hyperbolic equations, Academic Press, Boston and London,
1988.
J. Hadamard, Le problème de Cauchy et les équations aux dérivées partielles hyperboliques, Hermann,
Paris, 1932.
S. Helgason [1], Differential geometry and symmetric spaces, Academic Press, New York, London,
1962.
F. Hirzebruch [1], Topological methods in algebraic geometry, Springer-Verlag, Berlin, Göttingen,
Heidelberg, 1966, Grundl. d. Math. Wiss. 131..
L. Hörmander [1], The analysis of linear partial differential operators III, Springer-Verlag, Berlin,
Heidelberg, New York, Tokyo, 1985.
N. Kuipers [1], On C 1 isometric imbeddings, Indag. Math. 18 (1955), 545–556, 683–689.
H. Jacobowitz [1], Local isometric embeddings, Seminar on differential geometry, Princeton University
Press, 1982, Editor S.-T. Yau, Ann. of Math. Studies 102..
H. Jacobowitz [2], Local isometric embeddings of surfaces into Euclidean four space, Ind. Univ. Math.
J. 21 (1971), 249–254.
W. Klingenberg [1], Eine Vorlesung über Differentialgeometrie, Springer-Verlag, Berlin, Göttingen,
Heidelberg, 1973, Heidelberger Taschenbücher..
W. Klingenberg [2], Riemannian geometry, Walter de Gruyter, Berlin, 1982.
S. Kobayashi and K. Nomizu [1], Foundations of differential geometry I-II, Interscience Publ, New
York, N. Y., London, 1963, 1969.
J.M. Lee and T. H. Parker [1], The Yamabe problem, Bull. Amer. Math. Soc. 17 (1987), 37–91.
A. E. Liber [1], On a class of Riemannian spaces having constant negative curvature, Uch. Zap.
Saratov. Gos. Univ. Ser. Fiz.-Mat. 1:2 (1938), 105–122.
J. Milnor [1], On total curvature of closed space curves, Math. Scand. 1 (1953), 289–296.
J. Nash [1], The imbedding problem for Riemannian manifolds, Ann. of Math. 63 (1956), 20–63.
J. Nash [2], C 1 isometric imbeddings, Ann. of Math. 60 (1954), 383–396.
T. Parker and C. H. Taubes [1], On Witten’s proof of the positive energy theorem, Comm. Math. Phys.
84 (1982), 223–238.
V. K. Patodi [1], Curvature and the eigenforms of the Laplace operator, J. Diff. Geometry 5 (1971),
233–249.
W. Pauli [1], Relativitätstheorie, Teubner Verlag, Leipzig, Berlin, 1921, Reprint from Enc. d. math.
Wiss. V 19..
G. de Rham, Variétés différentiables, Hermann, Paris, 1960, Actualités scient. et ind. 1222.
B. Riemann [1], Über die Hypothesen welche der Geometrie zu Grunde liegen, Werke, ed. H. Weber,
Teubner Verlag, Leipzig, 1892, pp. 372–387.
J. Roe, Elliptic operators, topology and asymptotic methods, Longman Scientific & Technical, Harlow,
Essex and New York, 1988, Pitman Research Notes in Mathematics 179..
R. Schoen and S.-T. Yau [1], On the proof of the positive dmass conjecture in general relativity theory,
Comm. math. phys. 65 (1979), 45–76.
R. Schoen and S.-T. Yau [2], Proof of the positive mass theorem, Comm. math. phys. 79 (1981),
231–260.
F. Schur [1], Ueber die Deformation der Räume constanten Riemann’schen Krümmingsmasses, Math.
Ann. 27 (1886), 163–176.
S. Sternberg [1], Lectures on Differential Geometry, Prentice Hall, Englewood Cliffs, N.J., 1964.
F. W. Warner [1], Foundations of differentiable manifolds and Lie groups, Scott, Foresman and Co.,
Glenview, Ill., London, 1971.
R. Weitzenböck, Invariantentheorie, Nordhoff, Groningen, 1923.
H. Weyl [1], Classical groups, Princeton University Press, Princeton, N. J., 1946.
E. Witten [1], A new proof of the positive energy theorem, Comm. math. phys. 80 (1981), 381–402.
 197

Index

Affine length . . . . . . . . 3 Frenet formulas . . . . . . . 4

Anti-self-dual manifold . . . 25 Frobenius theorem . . . . . 185
Associated bundle . . . . . 100 Fundamental solution . . . 110
Berezin-Patodi formula . . . 153 Gauss equations . . . . . . 15
Binormal . . . . . . . . . . 4 Gauss map . . . . . . . . 17
Bochner’s theorem . . . . . 127 Gauss-Bonnet theorem . 43, 80
Capelli identity . . . . . . 189 Geodesic arc . . . . . . . 35
Chern character . . . . . . 170 Geodesic curvature . . . 14, 29
Chern class . . . . . . . . 90 Geodesic principal normal 14, 29
Complete manifold . . . . . 36 Gilkey theorem . . . . . . 136
Christoffel symbols . . . . . 13 Gradient . . . . . . . . . 45
Clifford algebra . . . . . . 140 Half-spin representations . . 151
Clifford bundle . . . . . . 153 Harmonic forms . . . . . . 123
Clifford connection . . . . . 154 Harmonic spinors . . . . . 160
Completely integrable system 187 Heat equation . . . . . . . 127
Conformal curvature tensor 23, 62 Hermite polynomials . . . . 161
Conformal Laplacian . . . . 66 Hermitian Clifford bundle . . 154
Conformal manifolds . . . . 60 Hermitian vector bundle . . 86
Conformally flat . . . . . . 62 Hirzebruch’s signature theorem 138
Connection . . . . . . . . 83 Hodge Laplacian . . . . . . 121
Connection form . . . . . . 103 Hodge theorem . . . . . . 123
Constant curvature . . . . . 52 Hopf-Rinow theorem . . . . 36
Contraction . . . . . . . . 32 Horizontal space . . . . . . 101
Covariant derivative . . 14, 28 Hyperbolic space . . . . . . 55
Curvature . . . . . . . . . . 2 Invariant polynomial . . . . 89
Curvature of connection . . 85 Isometric embedding . . . . 47
Curvature of principal bundle 105 Jacobi differential equation . 37
de Rham cohomology . 120, 181 Jacobi identity . . . . . . 93
Dirac operator . . . . . . . 153 Janet-Cartan theorem . . . 48
Dirac type operator . . . . 139 Killing form . . . . . . . . 96
Divergence of vector field . . 126 Left invariant vector field . . 92
Einstein manifold . . . . . 25 Left translation . . . . . . 92
Euclidean vector bundle . . 86 Legendre’s theorem . . 33, 41
First fundamental form . . . 12 Levi-Civita connection . . . 108
Euler characteristic . . 45, 124 Lichnerowicz formula . . . . 160
First Bianchi identity . . 21, 76 Lie algebra . . . . . . . . . 8
Free group action . . . . . 99 Lie group . . . . . . . . 8, 92
Free embedding . . . . . . 48 Local index theorem . . . . 170
198

Metric operator . . . . . . 107 Second fundamental form 12, 67

Mehler’s formula . . . . . . 164 Sectional curvature . . . . . 19
Meusnier’s theorem . . . . 12 Selfdual manifold . . . . . 25
Normal plane . . . . . . . 12 Semi-simple Lie group . . . 97
Orthonormal frame bundle 70, 74 Signature of manifold . . . . 125
Orthogonal matrix . . . . . . 6 Spherical excess . . . . . . 33
Parallel translation . . . . . 30 Spin group . . . . . . . . 146
Parametrix . . . . . . . . 109 Spin representations . . . . 151
Pin group . . . . . . . . . 146 Spin structure . . . . . . . 159
Star operator . . . . . . . 118
Polarization . . . . . 134, 172
Structural constants . . . . 93
Pontrjagin class . . . . . . 82
Submanifold . . . . . . . 11
Pontrjagin form . . . . . . 82
Supertrace . . . . . . . . 151
Principal bundle . . . . . . 100 Synchronous frame . . . 58, 108
Principal normal . . . . . . . 2 Tangent vector . . . . . . . 2
Principal symbol . . . . . . 85 Tau operator . . . . . . . 119
Proper parametrix . . . . . 114 Teorema egregium . . . . . 19
Radius of curvature . . . . . . 2 Torsion . . . . . . . . . . . 4
Ricci tensor . . . . . . . . 16 Traceless Ricci tensor . . . . 23
Riemann curvature tensor . . 16 Transition functions . . . . 90
Riemannian manifold . . . . 28 Twisted Dirac operator . . . 161
Right translation . . . . . 93 Weyl tensor . . . . . . . . 23
Scalar curvature . . . . . . 16 Weitzenböck formula . . . . 127
Second Bianchi identity . 46, 76 Yamabe problem . . . . . . 66