INTRODUCTION TO

DIFFERENTIAL GEOMETRY

Joel W. Robbin Dietmar A. Salamon

UW Madison ETH Zürich

28 March 2021
ii
Preface

These are notes for the lecture course “Differential Geometry I” given by the
second author at ETH Zürich in the fall semester 2017. They are based on
a lecture course1 given by the first author at the University of Wisconsin–
Madison in the fall semester 1983.
One can distinguish extrinsic differential geometry and intrinsic differ-
ential geometry. The former restricts attention to submanifolds of Euclidean
space while the latter studies manifolds equipped with a Riemannian metric.
The extrinsic theory is more accessible because we can visualize curves and
surfaces in R3 , but some topics can best be handled with the intrinsic theory.
The definitions in Chapter 2 have been worded in such a way that it is easy
to read them either extrinsically or intrinsically and the subsequent chapters
are mostly (but not entirely) extrinsic. One can teach a self contained one
semester course in extrinsic differential geometry by starting with Chapter 2
and skipping the sections marked with an asterisk like §2.8.
This document is designed to be read either as a .pdf file or as a printed
book.
We thank everyone who pointed out errors or typos in earlier versions
of this book. In particular, we thank Charel Antony and Samuel Trautwein
for many helpful comments.

21 November 2019 Joel W. Robbin and Dietmar A. Salamon

1
Extrinsic Differential Geometry

iii
iv
Contents

1 What is Differential Geometry? 1

1.1 Cartography and Differential Geometry . . . . . . . . . . . . 1
1.2 Coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Topological Manifolds* . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Smooth Manifolds Defined* . . . . . . . . . . . . . . . . . . . 10
1.5 The Master Plan . . . . . . . . . . . . . . . . . . . . . . . . . 13

2 Foundations 15
2.1 Submanifolds of Euclidean Space . . . . . . . . . . . . . . . . 15
2.2 Tangent Spaces and Derivatives . . . . . . . . . . . . . . . . . 24
2.2.1 Tangent Space . . . . . . . . . . . . . . . . . . . . . . 24
2.2.2 Derivative . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.2.3 The Inverse Function Theorem . . . . . . . . . . . . . 31
2.3 Submanifolds and Embeddings . . . . . . . . . . . . . . . . . 34
2.4 Vector Fields and Flows . . . . . . . . . . . . . . . . . . . . . 38
2.4.1 Vector Fields . . . . . . . . . . . . . . . . . . . . . . . 38
2.4.2 The Flow of a Vector Field . . . . . . . . . . . . . . . 41
2.4.3 The Lie Bracket . . . . . . . . . . . . . . . . . . . . . 46
2.5 Lie Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.5.1 Definition and Examples . . . . . . . . . . . . . . . . . 52
2.5.2 The Lie Algebra of a Lie Group . . . . . . . . . . . . . 55
2.5.3 Lie Group Homomorphisms . . . . . . . . . . . . . . . 58
2.5.4 Closed Subgroups . . . . . . . . . . . . . . . . . . . . 62
2.5.5 Lie Groups and Diffeomorphisms . . . . . . . . . . . . 67
2.5.6 Smooth Maps and Algebra Homomorphisms . . . . . . 69
2.5.7 Vector Fields and Derivations . . . . . . . . . . . . . . 71
2.6 Vector Bundles and Submersions . . . . . . . . . . . . . . . . 72
2.6.1 Submersions . . . . . . . . . . . . . . . . . . . . . . . 72
2.6.2 Vector Bundles . . . . . . . . . . . . . . . . . . . . . . 74

v
vi CONTENTS

2.7 The Theorem of Frobenius . . . . . . . . . . . . . . . . . . . . 80

2.8 The Intrinsic Definition of a Manifold* . . . . . . . . . . . . . 87
2.8.1 Definition and Examples . . . . . . . . . . . . . . . . . 87
2.8.2 Smooth Maps and Diffeomorphisms . . . . . . . . . . 92
2.8.3 Tangent Spaces and Derivatives . . . . . . . . . . . . . 93
2.8.4 Submanifolds and Embeddings . . . . . . . . . . . . . 95
2.8.5 Tangent Bundle and Vector Fields . . . . . . . . . . . 97
2.8.6 Coordinate Notation . . . . . . . . . . . . . . . . . . . 99
2.9 Consequences of Paracompactness* . . . . . . . . . . . . . . . 101
2.9.1 Paracompactness . . . . . . . . . . . . . . . . . . . . . 101
2.9.2 Partitions of Unity . . . . . . . . . . . . . . . . . . . . 103
2.9.3 Embedding in Euclidean Space . . . . . . . . . . . . . 107
2.9.4 Leaves of a Foliation . . . . . . . . . . . . . . . . . . . 113
2.9.5 Principal Bundles . . . . . . . . . . . . . . . . . . . . 114

3 The Levi-Civita Connection 121

3.1 Second Fundamental Form . . . . . . . . . . . . . . . . . . . . 121
3.2 Covariant Derivative . . . . . . . . . . . . . . . . . . . . . . . 127
3.3 Parallel Transport . . . . . . . . . . . . . . . . . . . . . . . . 129
3.4 The Frame Bundle . . . . . . . . . . . . . . . . . . . . . . . . 137
3.4.1 Frames of a Vector Space . . . . . . . . . . . . . . . . 137
3.4.2 The Frame Bundle . . . . . . . . . . . . . . . . . . . . 138
3.4.3 Horizontal Lifts . . . . . . . . . . . . . . . . . . . . . . 141
3.5 Motions and Developments . . . . . . . . . . . . . . . . . . . 146
3.5.1 Motion . . . . . . . . . . . . . . . . . . . . . . . . . . 146
3.5.2 Sliding . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
3.5.3 Twisting and Wobbling . . . . . . . . . . . . . . . . . 151
3.5.4 Development . . . . . . . . . . . . . . . . . . . . . . . 154
3.6 Christoffel Symbols . . . . . . . . . . . . . . . . . . . . . . . . 160
3.7 Riemannian Metrics* . . . . . . . . . . . . . . . . . . . . . . . 166
3.7.1 Existence of Riemannian Metrics . . . . . . . . . . . . 166
3.7.2 Two Examples . . . . . . . . . . . . . . . . . . . . . . 169
3.7.3 The Levi-Civita Connection . . . . . . . . . . . . . . . 170
3.7.4 Basic Vector Fields in the Intrinsic Setting . . . . . . 173

4 Geodesics 175
4.1 Length and Energy . . . . . . . . . . . . . . . . . . . . . . . . 175
4.1.1 The Length and Energy Functionals . . . . . . . . . . 175
4.1.2 The Space of Paths . . . . . . . . . . . . . . . . . . . . 178
4.1.3 Characterization of Geodesics . . . . . . . . . . . . . . 180
CONTENTS vii

4.2 Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

4.3 The Exponential Map . . . . . . . . . . . . . . . . . . . . . . 191
4.3.1 Geodesic Spray . . . . . . . . . . . . . . . . . . . . . . 191
4.3.2 The Exponential Map . . . . . . . . . . . . . . . . . . 192
4.3.3 Examples and Exercises . . . . . . . . . . . . . . . . . 195
4.3.4 Geodesics in Local Coordinates . . . . . . . . . . . . . 196
4.4 Minimal Geodesics . . . . . . . . . . . . . . . . . . . . . . . . 197
4.4.1 Characterization of Minimal Geodesics . . . . . . . . . 197
4.4.2 Local Existence of Minimal Geodesics . . . . . . . . . 199
4.4.3 Examples and Exercises . . . . . . . . . . . . . . . . . 203
4.5 Convex Neighborhoods . . . . . . . . . . . . . . . . . . . . . . 205
4.6 Completeness and Hopf–Rinow . . . . . . . . . . . . . . . . . 209
4.6.1 Geodesic Completeness . . . . . . . . . . . . . . . . . 209
4.6.2 Global Existence of Minimal Geodesics . . . . . . . . . 210
4.6.3 Proof of the Hopf–Rinow Theorem . . . . . . . . . . . 212
4.7 Geodesics in the Intrinsic Setting* . . . . . . . . . . . . . . . 217
4.7.1 Intrinsic Distance . . . . . . . . . . . . . . . . . . . . . 217
4.7.2 Geodesics and the Levi-Civita Connection . . . . . . . 219
4.7.3 Examples and Exercises . . . . . . . . . . . . . . . . . 220

5 Curvature 223
5.1 Isometries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
5.2 The Riemann Curvature Tensor . . . . . . . . . . . . . . . . . 232
5.2.1 Definition and Gauß–Codazzi . . . . . . . . . . . . . . 232
5.2.2 Covariant Derivative of a Global Vector Field . . . . . 234
5.2.3 A Global Formula . . . . . . . . . . . . . . . . . . . . 237
5.2.4 Symmetries . . . . . . . . . . . . . . . . . . . . . . . . 239
5.2.5 Riemannian Metrics on Lie Groups . . . . . . . . . . . 241
5.3 Generalized Theorema Egregium . . . . . . . . . . . . . . . . 244
5.3.1 Pushforward . . . . . . . . . . . . . . . . . . . . . . . 244
5.3.2 Theorema Egregium . . . . . . . . . . . . . . . . . . . 245
5.3.3 Gaußian Curvature . . . . . . . . . . . . . . . . . . . . 250
5.4 The Group of Isometries* . . . . . . . . . . . . . . . . . . . . 255
5.4.1 The Myers–Steenrod Theorem . . . . . . . . . . . . . 255
5.4.2 The Topology on the Space of Isometries . . . . . . . 257
5.4.3 Killing Vector Fields . . . . . . . . . . . . . . . . . . . 260
5.4.4 Proof of the Myers–Steenrod Theorem . . . . . . . . . 264
5.4.5 Isometries of Compact Lie Groups . . . . . . . . . . . 273
5.4.6 Examples and Exercises . . . . . . . . . . . . . . . . . 278
5.5 Curvature in Local Coordinates* . . . . . . . . . . . . . . . . 281
viii CONTENTS

6 Geometry and Topology 285

6.1 The Cartan–Ambrose–Hicks Theorem . . . . . . . . . . . . . 285
6.1.1 Homotopy . . . . . . . . . . . . . . . . . . . . . . . . . 285
6.1.2 The Global C-A-H Theorem . . . . . . . . . . . . . . . 287
6.1.3 The Local C-A-H Theorem . . . . . . . . . . . . . . . 293
6.2 Flat Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
6.3 Symmetric Spaces . . . . . . . . . . . . . . . . . . . . . . . . 300
6.3.1 Symmetric Spaces . . . . . . . . . . . . . . . . . . . . 301
6.3.2 Covariant Derivative of the Curvature . . . . . . . . . 303
6.3.3 Examples and Exercises . . . . . . . . . . . . . . . . . 306
6.4 Constant Curvature . . . . . . . . . . . . . . . . . . . . . . . 308
6.4.1 Sectional Curvature . . . . . . . . . . . . . . . . . . . 308
6.4.2 Constant Sectional Curvature . . . . . . . . . . . . . . 309
6.4.3 Examples and Exercises . . . . . . . . . . . . . . . . . 313
6.4.4 Hyperbolic Space . . . . . . . . . . . . . . . . . . . . . 314
6.5 Nonpositive Sectional Curvature . . . . . . . . . . . . . . . . 320
6.5.1 The Cartan–Hadamard Theorem . . . . . . . . . . . . 320
6.5.2 Cartan’s Fixed Point Theorem . . . . . . . . . . . . . 326
6.5.3 Positive Definite Symmetric Matrices . . . . . . . . . . 328
6.6 Positive Ricci Curvature* . . . . . . . . . . . . . . . . . . . . 333
6.6.1 The Ricci Tensor . . . . . . . . . . . . . . . . . . . . . 333
6.6.2 The Bonnet–Myers Theorem . . . . . . . . . . . . . . 334
6.6.3 Positive Sectional Curvature . . . . . . . . . . . . . . 336
6.7 Scalar Curvature* . . . . . . . . . . . . . . . . . . . . . . . . 337
6.7.1 Definition and Basic Properties . . . . . . . . . . . . . 337
6.7.2 Constant Scalar Curvature . . . . . . . . . . . . . . . 340
6.8 The Weyl Tensor* . . . . . . . . . . . . . . . . . . . . . . . . 343
6.8.1 Definition and Basic Properties . . . . . . . . . . . . . 343
6.8.2 Conformal Invariance . . . . . . . . . . . . . . . . . . 345
6.8.3 Self-Dual Four-Manifolds . . . . . . . . . . . . . . . . 346

A Notes 351
A.1 Maps and Functions . . . . . . . . . . . . . . . . . . . . . . . 351
A.2 Normal Forms . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
A.3 Euclidean Spaces . . . . . . . . . . . . . . . . . . . . . . . . . 354

References 357

Index 362
Chapter 1

What is Differential
Geometry?

1.1 Cartography and Differential Geometry

Carl Friedrich Gauß (1777-1855) is the father of differential geometry. He
was (among many other things) a cartographer and many terms in modern
differential geometry (chart, atlas, map, coordinate system, geodesic, etc.)
reflect these origins. He was led to his Theorema Egregium (see 5.3.1) by
the question of whether it is possible to draw an accurate map of a portion
of our planet. Let us begin by discussing a mathematical formulation of this
problem.
Consider the two-dimensional sphere S 2 sitting in the three-dimensional
Euclidean space R3 . It is cut out by the equation
x2 + y 2 + z 2 = 1.
A map of a small region U ⊂ S 2 is represented mathematically by a one-to-
one correspondence with a small region in the plane z = 0. In this book we
will represent this with the notation φ : U → φ(U ) ⊂ R2 and call such an
object a chart or a system of local coordinates (see Fugure 1.1).
What does it mean that φ is an “accurate” map? Ideally the user would
want to use the map to compute the length of a curve in S 2 . The length of
a curve γ connecting two points p, q ∈ S 2 is given by the formula
Z 1
L(γ) = |γ̇(t)| dt, γ(0) = p, γ(1) = q,
0
so the user will want the chart φ to satisfy L(γ) = L(φ ◦ γ) for all curves γ.
It is a consequence of the Theorema Egregium that there is no such chart.

1
2 CHAPTER 1. WHAT IS DIFFERENTIAL GEOMETRY?

U φ

Figure 1.1: A chart

Perhaps the user of such a map will be content to use the map to plot
the shortest path between two points p and q in U . This path is called a
geodesic and is denoted by γpq . It satisfies L(γpq ) = dU (p, q), where

dU (p, q) = inf{L(γ) | γ(t) ∈ U, γ(0) = p, γ(1) = q}

so our less demanding user will be content if the chart φ satisfies

dU (p, q) = dE (φ(p), φ(q))

where dE (φ(p), φ(q)) is the length of the shortest path in the plane. It is
also a consequence of the Theorema Egregium that there is no such chart.
Now suppose our user is content to have a map which makes it easy to
navigate close to the shortest path connecting two points. Ideally the user
would use a straight edge, magnetic compass, and protractor to do this.
S/he would draw a straight line on the map connecting p and q and steer a
course which maintains a constant angle (on the map) between the course
and meridians. This can be done by the method of stereographic projection.
This chart is conformal (which means that it preserves angles). According
to Wikipedia stereographic projection was known to the ancient Greeks
and a map using stereographic projection was constructed in the early 16th
century. Exercises 3.7.5, 3.7.12, and 6.4.22 use stereographic projection; the
latter exercise deals with the Poincaré model of the hyperbolic plane. The
hyperbolic plane provides a counterexample to Euclid’s Parallel Postulate.
Exercise 1.1.1. It is more or less obvious that for any surface M ⊂ R3
there is a unique shortest path in M connecting two points if they are
sufficiently close. (This will be proved in Theorem 4.5.3.) This shortest
path is called the minimal geodesic connecting p and q. Use this fact to
prove that the minimal geodesic joining two points p and q in S 2 is an arc
of the great circle through p and q. (This is the intersection of the sphere
with the plane through p, q, and the center of the sphere.) Also prove that
the minimal geodesic connecting two points in a plane is the straight line
1.1. CARTOGRAPHY AND DIFFERENTIAL GEOMETRY 3

n
p

φ(p)

Figure 1.2: Stereographic Projection

segment connecting them. Hint: Both a great circle in a sphere and a line
in a plane are preserved by a reflection. (See also Exercise 4.2.5 below.)

Exercise 1.1.2. Stereographic projection is defined by the condition that

for p ∈ S 2 \ n the point φ(p) lies in the xy-plane z = 0 and the three
points n = (0, 0, 1), p, and φ(p) are collinear (see Figure 1.2). Using the
formula that the cosine of the angle between two unit vectors is their inner
product prove that φ is conformal. Hint: The plane through p, q, and n
intersects the xy-plane in a straight line and the sphere in a circle through
n. The plane through n, p, φ(p), and the center of the sphere intersects the
sphere in a meridian. A proof that stereographic projection is conformal
can be found in [20, page 248]. The proof is elementary in the sense that
it doesn’t use calculus. An elementary proof can also be found online at
http://people.reed.edu/jerry/311/stereo.pdf.

Exercise 1.1.3. It may seem fairly obvious that you can’t draw an accurate
map of a portion of the earth because the sphere is curved. However the
cylinder C = {(x, y, z) ∈ R3 | x2 + y 2 = 1} is also curved, but the map
ψ : R2 → C defined by ψ(s, t) = (cos(t), sin(t), s) preserves lengths of curves,
i.e. L(ψ ◦ γ) = L(γ) for any curve γ : [a, b] → R2 . Prove this.

1.1.4. Standard Notations. The standard notations N, N0 , Z, Q, R,

C denote respectively the natural numbers (= positive integers), the non-
negative integers, the integers, the rational numbers, the real numbers, and
the complex numbers. We denote the identity map of a set X by idX and
the n × n identity matrix by 1ln or simply 1l. The notation V ∗ is used for the
dual of a vector space V , but when K is a field like R or C the notation K∗
is sometimes used for the multiplicative group K \ {0}. The terms smooth,
infinitely differentiable, and C ∞ are all synonymous.
4 CHAPTER 1. WHAT IS DIFFERENTIAL GEOMETRY?

1.2 Coordinates
The rest of this chapter defines the category of smooth manifolds and smooth
maps between them. Before giving the precise definitions we will introduce
some terminology and give some examples.

Definition 1.2.1. A chart on a set M is a pair (φ, U ) where U is a subset

of M and φ : U → φ(U ) is a bijection from U to an open set φ(U ) in Rm .
An atlas on M is a collection

A = {(φα , Uα )}α∈A

of charts such that the domains Uα cover M , i.e.

[
M= Uα .
α∈A

The idea is that if φ(p) = (x1 (p), . . . , xm (p)) for p ∈ U , then the func-
tions xi form a system of local coordinates defined on the subset U of M .
The dimension of M should be m since it takes m numbers to uniquely spec-
ify a point of U . We will soon impose conditions on charts (φ, U ), however
for the moment we are assuming nothing about the maps φ (other than that
they are bijective).

Example 1.2.2. Every open subset U ⊂ Rm has an atlas consisting of a

single chart, namely (φ, U ) = (idU , U ) where idU denotes the identity map
of U .

Example 1.2.3. Assume that Ω ⊂ Rm is an open set, that M is a subset

of the product Rm × Rn = Rm+n , and that

h : Ω → Rn

is a continuous map whose graph is M , i.e.

graph(h) := {(x, y) ∈ Ω × Rn | y = h(x)} = M.

Let U = graph(h) = M and let φ(x, y) = x be the projection of U onto Ω.

Then the pair (φ, U ) is a chart on M . The inverse map is given by

φ−1 (x) = (x, h(x))

for x ∈ Ω = φ(U ). Thus M has again an atlas consisting of a single chart.

1.2. COORDINATES 5

Example 1.2.4. The m-sphere

S m = p = (x0 , . . . , xm ) ∈ Rm+1 | x20 + · · · + x2m = 1



has an atlas consisting of the 2m + 2 charts φi± : Ui± → Dm where Dm is the

open unit disk in Rm , Ui± = {p ∈ S m | ± xi > 0}, and φi± is the projection
which discards the ith coordinate. (See Example 2.1.13 below.)
Example 1.2.5. Let

A = AT ∈ R(m+1)×(m+1)

be a symmetric matrix and define a quadratic form F : Rm+1 → R by

F (p) := pT Ap, p = (x0 , . . . , xm ).

After a linear change of coordinates the function F has the form

F (p) = x20 + · · · + x2k − x2k+1 − · · · − x2r .

(Here r + 1 is the rank of the matrix A.) The set M = F −1 (1) has an atlas
of 2m + 2 charts by the same construction as in Example 1.2.4, in fact S m
is the special case where A = 1lm+1 , the (m + 1) × (m + 1) identity matrix.
(See Example 2.1.12 below for another way to construct charts.)
Figure 1.3 enumerates the familiar quadric surfaces in R3 . The para-
boloids are examples of graphs as in Example 1.2.3 with Ω = R2 and n = 1,
and the ellipsoid and the two hyperboloids are instances of the quadric hyper-
surfaces defined in Example 1.2.5. The sphere is an instance of the ellipsoid
(with a = b = c = 1) and the cylinder is a limit of the ellipsoid as well as of
the elliptic hyperboloid of one sheet (as a = b = 1 and c → ∞). The pictures
were generated by computer using the parameterizations

x = a cos(t) cos(s), y = b sin(t) cos(s), z = c sin(s)

for the ellipsoid,

x = a cos(t) cosh(s), y = b sin(t) cosh(s), z = c sinh(s),

for the elliptic hyperboloid of one sheet, and

x = a cos(t) sinh(s), y = b sin(t) sinh(s), z = ±c cosh(s)

for the elliptic hyperboloid of two sheets. These quadric surfaces will be
often used in the sequel to illustrate important concepts.
6 CHAPTER 1. WHAT IS DIFFERENTIAL GEOMETRY?

Unit Sphere x2 + y 2 + z 2 = 1

x2 y 2 z 2
Ellipsoid + 2 + 2 =1
a2 b c

Cylinder x2 + y 2 = 1

Elliptic
x2 y 2 z 2
Hyperboloid + 2 − 2 =1
a2 b c
(of one sheet)

Elliptic
x2 y 2 z 2
Hyperboloid + 2 − 2 = −1
a2 b c
(of two sheets)

Hyperbolic
x2 y 2
Paraboloid z= − 2
a2 b

Elliptic
x2 y 2
Paraboloid z= + 2
a2 b

Figure 1.3: Quadric Surfaces

1.2. COORDINATES 7

In the following two examples K denotes either the field R of real numbers
or the field C of complex numbers, K∗ := {λ ∈ K | λ 6= 0} denotes the
corresponding multiplicative group, and V denotes a vector space over K.
Example 1.2.6. The projective space of V is the set of lines (through
the origin) in V . In other words,
P (V ) = {` ⊂ V | ` is a 1-dimensional K-linear subspace}
When K = R and V = Rm+1 this is denoted by RPm and when K = C
and V = Cm+1 this is denoted by CPm . For our purposes we can identify the
spaces Cm+1 and R2m+2 but the projective spaces CPm and RP2m are very
different. The various lines ` ∈ P (V ) intersect in the origin, however, after
the identification P (V ) = {[v] | v ∈ V \ {0}} with [v] := K∗ v = Kv \ {0} the
elements of P (V ) become disjoint, i.e. P (V ) is the set of equivalence classes
of an equivalence relation on the open set V \ {0}. Assume that V = Km+1
and define an atlas on P (V ) as follows. For each integer i = 0, 1, . . . , m
define Ui := {[v] | v = (x0 , . . . , xm ), xi 6= 0} and define φi : Ui → Km by
 
x0 xi−1 xi+1 xm
φi ([v]) = ,..., , ..., .
xi xi xi xi
This atlas consists of m + 1 charts.
Example 1.2.7. For each positive integer k the set
Gk (V ) := {` ⊂ V | ` is a k-dimensional K-linear subspace}
is called the Grassmann manifold of k-planes in V . Thus G1 (V ) = P (V ).
Assume that V = Kn and define an atlas on Gk (V ) as follows. Let e1 , . . . , en
be the standard basis for Kn , i.e. ei is the ith column of the n × n identity
matrix 1ln . Each partition {1, 2, . . . , n} = I ∪ J, I = {i1 < · · · < ik },
J = {j1 < · · · < jn−k } of the first n natural numbers determines a direct
sum decomposition Kn = V = VI ⊕ VJ via the formulas
VI = Kei1 + · · · + Keik , VJ = Kej1 + · · · + Kejn−k .
Let UI denote the set of all k-planes ` ∈ Gk (V ) which are transverse to VJ ,
i.e. such that ` ∩ VJ = {0}. The elements of UI are precisely those k-planes
of the form ` = graph(A), where A : VI → VJ is a linear map. Define the
map φI : UI → Kk×(n−k) by the formula
n−k
X
φI (`) = (ars ), Aeir = ars ejs , r = 1, . . . , k.
s=1
Exercise: Prove that the set of all pairs (φI , UI ) as I ranges over the subsets
of {1, . . . , n} of cardinality k form an atlas.
8 CHAPTER 1. WHAT IS DIFFERENTIAL GEOMETRY?

1.3 Topological Manifolds*

Definition 1.3.1. A topological manifold is a topological space M such
that each point p ∈ M has an open neighborhood U which is homeomorphic
to an open subset of a Euclidean space.
Brouwer’s Invariance of Domain Theorem asserts that, when U ⊂ Rm
and V ⊂ Rn are nonempty open sets and φ : U → V is a homeomorphism,
then m = n. This means that if M is a connected topological manifold
and some point of M has a neighborhood homeomorphic to an open subset
of Rm , then every point of M has a neighborhood homeomorphic to an open
subset of that same Rm . In this case we say that M has dimension m or
is m-dimensional or is an m-manifold. Brouwer’s theorem is fairly difficult
(see [17, p. 126] for example) but if φ is a diffeomorphism, then the result is
an easy consequence of the invariance of the rank in linear algebra and the
chain rule. (See equation (1.4.1) below.)
By definition, a topological m-manifold M admits an atlas where every
chart (φ, U ) of the atlas is a homeomorphism φ : U → φ(U ) from an open
set U ⊂ M to an open set φ(U ) ⊂ Rm . The following definition and lemma
explain when a given atlas determines a topology on M .
Definition 1.3.2. Let M be a set. Two charts (φ1 , U1 ) and (φ2 , U2 ) on M
are said to be topologically compatible iff φ1 (U1 ∩ U2 ) and φ2 (U1 ∩ U2 )
are open subsets of Rm and the transition map
φ21 := φ2 ◦ φ−1
1 : φ1 (U1 ∩ U2 ) → φ2 (U1 ∩ U2 )

is a homeomorphism. An atlas is said to be a topological atlas iff any two

charts in this atlas are topologically compatible.
Lemma 1.3.3. Let A = {(φα , Uα )}α∈A be an atlas on a set M .
(i) If A is a topological atlas, then the collection
 φα (U ∩ Uα ) is an open subset of Rm
  
U := U ⊂ M   (1.3.1)
for every α ∈ A
is a topology on M , and with this topology each Uα is an open subset of M
and each φα is a homeomorphism. Thus M is a topological manifold with
the topology (1.3.1).
(ii) If M is a topological manifold and each Uα is an open set and each φα
is a homeomorphism, then A is a topological atlas and the topology (1.3.1)
coincides with the topology of M .
Proof. Exercise.
1.3. TOPOLOGICAL MANIFOLDS* 9

If M is a topological manifold, then the collection of all charts (U, φ)

on M such that U is open and φ is a homeomorphism is a topological atlas.
It is the unique maximal topological atlas in the sense that it contains
every other topological atlas as in part (ii) of Lemma 1.3.3. However, we will
often consider smaller, even finite, topological atlases, and by Lemma 1.3.3
each of these determines the topology of M .
Exercise 1.3.4. Show that the atlas in each example in §1.2 is a topological
atlas. Conclude that each of these examples is a topological manifold.
Any subset S ⊂ X of a topological space X inherits a topology from X,
called the relative topology of S. A subset U0 ⊂ S is called relatively
open in S (or S-open) iff there is an open set U ⊂ X such that U0 = U ∩ S.
A subset A0 ⊂ S is called relatively closed (or S-closed) iff there is a
closed set A ⊂ X such that A0 = A ∩ S. The relative topology on S is the
finest topology such that the inclusion map S → X is continuous.
Exercise 1.3.5. Show that the relative topology satisfies the axioms of a
topology (i.e. arbitrary unions and finite intersections of S-open sets are S-
open, and the empty set and S itself are S-open). Show that the complement
of an S-open set in S is S-closed and vice versa.
Exercise 1.3.6. Each of the sets defined in Examples 1.2.2, 1.2.3, 1.2.4,
and 1.2.5 is a subset of some Euclidean space Rk . Show that the topology
in Exercise 1.3.4 is the relative topology inherited from the topology of Rk .
The topology on Rk is of course the metric topology defined by the distance
function d(x, y) = |x − y|.
If ∼ is an equivalence relation on a topological space X, then the quo-
tient space Y := X/∼ := {[x] | x ∈ X} is the set of all equivalence classes
[x] := {x0 ∈ X | x0 ∼ x}. The map π : X → Y defined by π(x) = [x]
will be called the obvious projection. The quotient space inherits the
quotient topology from Y . Namely, a set V ⊂ Y is open in this topol-
ogy iff the preimage π −1 (V ) is open in X. This topology is the coarsest
topology on Y such that projection π : X → Y is continuous. Since the op-
eration V 7→ π −1 (V ) commutes with arbitrary unions and intersections the
quotient topology obviously satisfies the axioms of a topology.
Exercise 1.3.7. Show that the topology on the projective space P (V ) in
Example 1.2.6 determined by the atlas is the quotient topology inherited
from the open set V \ {0}. Express the Grassmann manifold Gk (V ) in
Example 1.2.7 as a quotient space and show that the topology determined
by the atlas is the quotient topology. (Recall that in both examples V = Kn
with K = R or K = C.)
10 CHAPTER 1. WHAT IS DIFFERENTIAL GEOMETRY?

1.4 Smooth Manifolds Defined*

Let U ⊂ Rn and V ⊂ Rm be open sets. A map f : U → V is called smooth
iff it is infinitely differentiable, i.e. iff all its partial derivatives
∂ α1 +···+αn f
∂αf = , α = (α1 , . . . , αn ) ∈ Nn0 ,
∂xα1 1 · · · ∂xαnn
exist and are continuous. In later chapters we will sometimes write C ∞ (U, V )
for the set of smooth maps from U to V .
Definition 1.4.1. Let U ⊂ Rn and V ⊂ Rm be open sets. For a smooth
map f = (f1 , . . . , fm ) : U → V and a point x ∈ U the derivative of f at x
is the linear map df (x) : Rn → Rm defined by

d  f (x + tξ) − f (x)
df (x)ξ :=  f (x + tξ) = lim , ξ ∈ Rn .
dt t=0 t→0 t
This linear map is represented by the Jacobian matrix of f at x which
will also be denoted by
∂f1 ∂f1
 
(x) · · · (x)
 ∂x1 ∂xn 

df (x) :=  .
.. .
..  ∈ Rm×n .

 
 ∂fm 
(x) · · · ∂fm
∂xn (x)
∂x1
Note that we use the same notation for the Jacobian matrix and the corre-
sponding linear map from Rn to Rm .
The derivative satisfies the chain rule. Namely, if U ⊂ Rn , V ⊂ Rm ,
and W ⊂ R` are open sets and f : U → V and g : V → W are smooth map,
then g ◦ f : U → W is smooth and
d(g ◦ f )(x) = dg(f (x)) ◦ df (x) : Rn → R` (1.4.1)
for every x ∈ U . Moreover the identity map idU : U → U is always smooth
and its derivative at every point is the identity map of Rn . This implies that,
if f : U → V is a diffeomorphism (i.e. f is bijective and f and f −1 are both
smooth), then its derivative at every point is an invertible linear map. This
is why the Invariance of Domain Theorem (discussed after Definition 1.3.1)
is easy for diffeomorphisms: if f : U → V is a diffeomorphism, then the
Jacobian matrix df (x) ∈ Rm×n is invertible for every x ∈ U and so m = n.
The Inverse Function Theorem (see Theorem A.2.2 in Appendix A.2) is a
kind of converse.
1.4. SMOOTH MANIFOLDS DEFINED* 11

Definition 1.4.2 (Smooth Manifold). Let M be a set. A chart on M

is a pair (φ, U ) where U ⊂ M and φ is a bijection from U to an open sub-
set φ(U ) ⊂ Rm of some Euclidean space. Two charts (φ1 , U1 ) and (φ2 , U2 )
are said to be smoothly compatible iff φ1 (U1 ∩ U2 ) and φ2 (U1 ∩ U2 ) are
both open in Rm and the transition map

φ21 = φ2 ◦ φ−1
1 : φ1 (U1 ∩ U2 ) → φ2 (U1 ∩ U2 ) (1.4.2)

is a diffeomorphism. A smooth atlas on M is a collection A of charts

on M any two of which are smoothly compatible and such that the sets U ,
as (φ, U ) ranges over the elements of A , cover M (i.e. for every p ∈ M
there is a chart (φ, U ) ∈ A with p ∈ U ). A maximal smooth atlas is an
atlas which contains every chart which is smoothly compatible with each of
its members. A smooth manifold is a pair consisting of a set M and a
maximal smooth atlas A on M .

Lemma 1.4.3. If A is a smooth atlas, then so is the collection A of all

charts smoothly compatible with each member of A . The smooth atlas A is
obviously maximal. In other words, every smooth atlas extends uniquely to
a maximal smooth atlas.

Proof. Let (φ1 , U1 ) and (φ2 , U2 ) be charts in A and let x ∈ φ1 (U1 ∩ U2 ).

Choose a chart (φ, U ) ∈ A such that φ−1 1 (x) ∈ U . Then φ1 (U ∩ U1 ∩ U2 ) is
an open neighborhood of x in Rm and the transition maps

φ ◦ φ−1
1 : φ1 (U ∩ U1 ∩ U2 ) → φ(U ∩ U1 ∩ U2 ),
φ2 ◦ φ−1 : φ(U ∩ U1 ∩ U2 ) → φ2 (U ∩ U1 ∩ U2 )

are smooth by definition of A . Hence so is their composition. This shows

that the map φ2 ◦ φ−1
1 : φ1 (U1 ∩ U2 ) → φ2 (U1 ∩ U2 ) is smooth near x. Since x
was chosen arbitrary, this map is smooth. Apply the same argument to its
inverse to deduce that it is a diffeomorphism. Thus A is a smooth atlas.

Definitions 1.4.2 and 1.3.2 are mutatis mutandis the same, so every
smooth atlas on a set M is a fortiori a topological atlas, i.e. every smooth
manifold is a topological manifold. (See Lemma 1.3.3.) Moreover the defi-
nitions are worded in such a way that it is obvious that every smooth map
is continuous.

Exercise 1.4.4. Show that each of the atlases from the examples in §1.2 is a
smooth atlas. (You must show that the transition maps from Exercise 1.3.4
are smooth.)
12 CHAPTER 1. WHAT IS DIFFERENTIAL GEOMETRY?

When A is a smooth atlas on a topological manifold M one says that A

is a smooth structure on the (topological) manifold M iff A ⊂ B,
where B is the maximal topological atlas on M . When no confusion can
result we generally drop the notation for the maximal smooth atlas as in the
following exercise.

Exercise 1.4.5. Define the notion of a continuous map between topological

manifolds and of a smooth map between smooth manifolds via conttinuity,
respectively smoothness, in local coordinates. Let M , N , and P be smooth
manifolds and f : M → N and g : N → P be smooth maps. Prove that
the identity map idM is smooth and that the composition g ◦ f : M → P
is a smooth map. (This is of course an easy consequence of the chain
rule (1.4.1).)

Remark 1.4.6. It is easy to see that a topological manifold can have

many distinct smooth structures. For example, {(idR , R)} and {(φ, R)}
where φ(x) = x3 are atlases on the real numbers which extend to distinct
smooth structures but determine the same topology. However these two
manifolds are diffeomorphic via the map x 7→ x1/3 . In the 1950’s it was
proved that there are smooth manifolds which are homeomorphic but not dif-
feomorphic and that there are topological manifolds which admit no smooth
structure. In the 1980’s it was proved in dimension m = 4 that there are
uncountably many smooth manifolds that are all homeomorphic to R4 but
no two of them are diffeomorphic to each other. These theorems are very
surprising and very deep.

A collection of sets and maps between them is called a category iff the
collection of maps contains the identity map of every set in the collection and
the composition of any two maps in the collection is also in the collection.
The sets are called the objects of the category and the maps are called the
morphisms of the category. An invertible morphism whose inverse is also in
the category is called an isomorphism. Some examples are the category of
all sets and maps, the category of topological spaces and continuous maps
(the isomorphisms are the homeomorphisms), the category of topological
manifolds and continuous maps between them, and the category of smooth
manifolds and smooth maps (the isomorphisms are the diffeomorphisms).
Each of the last three categories is a subcategory of the preceding one.
Often categories are enlarged by a kind of “gluing process”. For example,
the “global” category of smooth manifolds and smooth maps was constructed
from the “local” category of open sets in Euclidean space and smooth maps
between them via the device of charts and atlases. (The chain rule shows
1.5. THE MASTER PLAN 13

that this local category is in fact a category.) The point of Definition 1.3.2
is to show (via Lemma 1.3.3) that topological manifolds can be defined
in a manner analogous to the definition we gave for smooth manifolds in
Definition 1.4.2.
Other kinds of manifolds (and hence other kinds of geometry) are de-
fined by choosing other local categories, i.e. by imposing conditions on the
transition maps in Equation (1.4.2). For example, a real analytic manifold
is one where the transition maps are real analytic, a complex manifold is one
whose coordinate charts take values in Cn and whose transition maps are
holomorphic diffeomorphisms, and a symplectic manifold is one whose coor-
dinate charts take values in R2n and whose transition maps are canonical
transformations in the sense of classical mechanics. Thus CPn is a complex
manifold and RPn is a real analytic manifold.

1.5 The Master Plan

In studying differential geometry it is best to begin with extrinsic differential
geometry which is the study of the geometry of submanifolds of Euclidean
space as in Examples 1.2.3, and 1.2.5. This is because we can visualize
curves and surfaces in R3 . However, there are a few topics in the later
chapters which require the more abstract definition 1.4.2 even to say inter-
esting things about extrinsic geometry. There is a generalization to these
manifolds involving a structure called a Riemannian metric. We will call
this generalization intrinsic differential geometry. Examples 1.2.6 and 1.2.7
fit into this more general definition so intrinsic differential geometry can be
used to study them.
Since an open set in Euclidean space is a smooth manifold the definition
of a submanifold of Euclidean space (see §2.1 below) is mutatis mutandis
the same as the definition of a submanifold of a manifold. The definitions
in Chapter 2 are worded in such a way that it is easy to read them either
extrinsically or intrinsically and the subsequent chapters are mostly (but not
entirely) extrinsic. Those sections which require intrinsic differential geom-
etry (or which translate extrinsic concepts into intrinsic ones) are marked
with a *.
14 CHAPTER 1. WHAT IS DIFFERENTIAL GEOMETRY?
Chapter 2

Foundations

This chapter introduces various fundamental concepts that are central to

the fields of differential geometry and differential topology. Both fields con-
cern the study of smooth manifolds and their diffeomorphisms. The chapter
begins with an introduction to submanifolds of Euclidean space and smooth
maps (§2.1), to tangent spaces and derivatives (§2.2), and to submanifolds
and embeddings (§2.3). In §2.4 we move on to vector fields and flows and
introduce the Lie bracket of two vector fields. Lie groups and their Lie alge-
bras, in the extrinsic setting, are the subject of §2.5, which includes a proof
of the Closed Subgroup Theorem. In §2.6 we introduce vector bundles over
a manifold as subbundles of a trivial bundle and in §2.7 we prove the theo-
rem of Frobenius. The last two sections of this chapter are concerned with
carrying over all these concepts from the extrinsic to the intrinsic setting
and can be skipped at first reading (§2.8 and §2.9).

2.1 Submanifolds of Euclidean Space

To carry out the Master Plan §1.5 we must (as was done in [37]) extend
the definition of smooth map to maps f : X → Y between subsets X ⊂ Rk
and Y ⊂ R` which are not necessarily open. In this case a map f : X → Y is
called smooth iff for each x0 ∈ X there exists an open neighborhood U ⊂ Rk
of x0 and a smooth map F : U → R` that agrees with f on U ∩ X. A
map f : X → Y is called a diffeomorphism iff f is bijective and f and f −1
are smooth. When there exists a diffeomorphism f : X → Y then X and Y
are called diffeomorphic. When X and Y are open these definitions coin-
cide with the usage in §1.4.

15
16 CHAPTER 2. FOUNDATIONS

Exercise 2.1.1 (Chain Rule). Let X ⊂ Rk , Y ⊂ R` , Z ⊂ Rm be arbitrary

subsets. If f : X → Y and g : Y → Z are smooth maps, then so is the
composition g ◦ f : X → Z. The identity map id : X → X is smooth.
Exercise 2.1.2. Let E ⊂ Rk be an m-dimensional linear subspace and
. , vm be a basis of E. Then the map f : Rm → E defined by
let v1 , . .P
f (x) := m i=1 xi vi is a diffeomorphism.
Definition 2.1.3. Let k, m ∈ N0 . A subset M ⊂ Rk is called a smooth
m-dimensional submanifold of Rk iff every point p ∈ M has an open
neighborhood U ⊂ Rk such that U ∩ M is diffeomorphic to an open sub-
set Ω ⊂ Rm . A diffeomorphism
φ:U ∩M →Ω
is called a coordinate chart of M and its inverse
ψ := φ−1 : Ω → U ∩ M
is called a (smooth) parametrization of U ∩ M .

U M M φ Ω

Figure 2.1: A coordinate chart φ : U ∩ M → Ω.

In Definition 2.1.3 we have used the fact that the domain of a smooth
map can be an arbitrary subset of Euclidean space and need not be open
(see page 15). The term m-manifold in Rk is short for m-dimensional sub-
manifold of Rk . In keeping with the Master Plan §1.5 we will sometimes say
manifold rather than submanifold of Rk to indicate that the context holds
in both the intrinsic and extrinsic settings.
Lemma 2.1.4. If M ⊂ Rk is a nonempty smooth m-manifold, then m ≤ k.
Proof. Fix an element p0 ∈ M , choose a coordinate chart φ : U ∩ M → Ω
with p0 ∈ U and values in an open subset Ω ⊂ Rm , and denote its in-
verse by ψ := φ−1 : Ω → U ∩ M . Shrinking U , if necessary, we may as-
sume that φ extends to a smooth map Φ : U → Rm . This extension satis-
fies Φ(ψ(x)) = φ(ψ(x)) = x and hence dΦ(ψ(x))dψ(x) = id : Rm → Rm for
all x ∈ Ω, by the chain rule. Hence the derivative dψ(x) : Rm → Rk is in-
jective for all x ∈ Ω, and hence m ≤ k because Ω is nonempty. This proves
Lemma 2.1.4.
2.1. SUBMANIFOLDS OF EUCLIDEAN SPACE 17

Example 2.1.5. Consider the 2-sphere

M := S 2 = (x, y, z) ∈ R3 | x2 + y 2 + z 2 = 1


depicted in Figure 2.2 and let U ⊂ R3 and Ω ⊂ R2 be the open sets

U := (x, y, z) ∈ R3 | z > 0 , Ω := (x, y) ∈ R2 | x2 + y 2 < 1 .

 

The map φ : U ∩ M → Ω given by

φ(x, y, z) := (x, y)

is bijective and its inverse ψ := φ−1 : Ω → U ∩ M is given by

p
ψ(x, y) = (x, y, 1 − x2 − y 2 ).

Since both φ and ψ are smooth, the map φ is a coordinate chart on S 2 .

Similarly, we can use the open sets z < 0, y > 0, y < 0, x > 0, x < 0 to cover
S 2 by six coordinate charts. Hence S 2 is a manifold. A similar argument
shows that the unit sphere S m ⊂ Rm+1 (see Example 2.1.13 below) is a
manifold for every integer m ≥ 0.

Figure 2.2: The 2-sphere and the 2-torus.

Example 2.1.6. Let Ω ⊂ Rm be an open set and h : Ω → Rk−m be a smooth

map. Then the graph of h is a smooth submanifold of Rm × Rk−m = Rk :

M = graph(h) := {(x, y) | x ∈ Ω, y = h(x)} .

It can be covered by a single coordinate chart φ : U ∩ M → V where

U := Ω × Rk−m , φ is the projection onto Ω, and ψ := φ−1 : Ω → U is given
by ψ(x) = (x, h(x)) for x ∈ Ω.

Exercise 2.1.7 (The case m = 0). Show that a subset M ⊂ Rk is a 0-

dimensional submanifold if and only if M is discrete, i.e. for every p ∈ M
there exists an open set U ⊂ Rk such that U ∩ M = {p}.
18 CHAPTER 2. FOUNDATIONS

Exercise 2.1.8 (The case m = k). Show that a subset M ⊂ Rm is an

m-dimensional submanifold if and only if M is open.
Exercise 2.1.9 (Products). If Mi ⊂ Rki is an mi -manifold for i = 1, 2,
show that M1 × M2 is an (m1 + m2 )-dimensional submanifold of Rk1 +k2 .
Prove that the m-torus Tm := (S 1 )m is a smooth submanifold of Cm .
The next theorem characterizes smooth submanifolds of Euclidean space.
In particular condition (iii) will be useful in many cases for verifying the
manifold condition. We emphasize that the sets U0 := U ∩ M that appear
in Definition 2.1.3 are open subsets of M with respect to the relative topology
that M inherits from the ambient space Rk and that such relatively open
sets are also called M -open (see page 9).
Theorem 2.1.10 (Manifolds). Let m and k be integers with 0 ≤ m ≤ k.
Let M ⊂ Rk be a set and p0 ∈ M . Then the following are equivalent.
(i) There exists an M -open neighborhood U0 ⊂ M of p0 and a diffeomor-
phism
φ 0 : U0 → Ω 0
onto an open set Ω0 ⊂ Rm .
(ii) There exist open sets U, Ω ⊂ Rk and a diffeomorphism φ : U → Ω such
that p0 ∈ U and
φ(U ∩ M ) = Ω ∩ (Rm × {0}) .

(iii) There exists an open set U ⊂ Rk and a smooth map f : U → Rk−m

such that p0 ∈ U , the derivative df (p) : Rk → Rk−m is surjective for every
point p ∈ U ∩ M , and

U ∩ M = f −1 (0) = {p ∈ U | f (p) = 0} .

Moreover, if (i) holds, then the diffeomorphism φ : U → Ω in (ii) can be

chosen such that U ∩ M ⊂ U0 and φ(p) = (φ0 (p), 0) for every p ∈ U ∩ M .
Proof. First assume (ii) and denote the diffeomorphism in (ii) by

φ = (φ1 , φ2 , . . . , φk ) : U → Ω ⊂ Rk .

Then part (i) holds with U0 := U ∩ M , Ω0 := {x ∈ Rm | (x, 0) ∈ Ω}, and

φ0 := (φ1 , . . . , φm )|U0 : U0 → Ω0 ,

and part (iii) holds with f := (φm+1 , . . . , φk ) : U → Rk−m . This shows that
part (ii) implies both (i) and (iii).
2.1. SUBMANIFOLDS OF EUCLIDEAN SPACE 19

We prove that (i) implies (ii). Let φ0 : U0 → Ω0 be the coordinate chart

in part (i), let ψ0 := φ−1
0 : Ω0 → U0 be its inverse, and denote

x0 := φ0 (p0 ) ∈ Ω0 .

Then, by Lemma 2.1.4, the derivative dψ0 (x0 ) : Rm → Rk is an injective

linear map. Hence there exists a matrix B ∈ Rk×(k−m) such that

det([dψ0 (x0 ) B]) 6= 0.

Define the map ψ : Ω0 × Rk−m → Rk by

ψ(x, y) := ψ0 (x) + By.

Then the k × k-matrix dψ(x0 , 0) = [dψ0 (x0 ) B] ∈ Rk×k is nonsingular, by

choice of B. Hence, by the Inverse Function Theorem A.2.2, there exists
an open neighborhood Ω e ⊂ Ω0 × Rk−m of (x0 , 0) such that U e ⊂ Rk
e := ψ(Ω)
is open and ψ|Ωe : Ω e →U e is a diffeomorphism. In particular, the restriction
of ψ to Ωe is injective. Now the set
n  o n  o
Ue0 := ψ0 (x)  x ∈ Ω0 , (x, 0) ∈ Ωe = p ∈ U0  (φ0 (p), 0) ∈ Ω
e ⊂M

is M -open and contains p0 . Hence, by the definition of the relative topology,

there exists an open set W ⊂ Rk such that U e0 = W ∩ M . Define

e ∩ W,
U := U e ∩ ψ −1 (W ).
Ω := Ω

Then U ∩ M = U e0 and ψ restricts to a diffeomorphism from Ω to U .

Now let (x, y) ∈ Ω. We claim that

ψ(x, y) ∈ M ⇐⇒ y = 0. (2.1.1)

If y = 0, then obviously ψ(x, y) = ψ0 (x) ∈ M . Conversely, let (x, y) ∈ Ω and

suppose that p := ψ(x, y) ∈ M . Then p ∈ U ∩ M = U e ∩W ∩M =U e0 ⊂ U0
and hence (φ0 (p), 0) ∈ Ω, by definition of U0 . This implies
e e

ψ(φ0 (p), 0) = ψ0 (φ0 (p)) = p = ψ(x, y).

Since the pairs (x, y) and (φ0 (p), 0) both belong to the set Ω
e and the re-
striction of ψ to Ω is injective we obtain x = φ0 (p) and y = 0. This
e
proves (2.1.1). It follows from (2.1.1) that the map φ := (ψ|Ω )−1 : U → Ω
satisfies φ(U ∩ M ) = {(x, y) ∈ Ω | ψ(x, y) ∈ M } = Ω ∩ (Rm × {0}). Thus we
have proved that (i) implies (ii).
20 CHAPTER 2. FOUNDATIONS

We prove that (iii) implies (ii). Let f : U → Rk−m be as in part (iii).

Then p0 ∈ U and the derivative df (p0 ) : Rk → Rk−m is a surjective linear
map. Hence there exists a matrix A ∈ Rm×k such that
 
A
det 6= 0.
df (p0 )

Define the map φ : U → Rk by

 
Ap
φ(p) := for p ∈ U.
f (p)

Then det(dφ(p0 )) 6= 0. Hence, by the Inverse Function Theorem A.2.2, there

exists an open neighborhood U 0 ⊂ U of p0 such that Ω0 := φ(U 0 ) is an open
subset of Rk and the restriction

φ0 := φ|U 0 : U 0 → Ω0

is a diffeomorphism. In particular, the restriction φ|U 0 is injective. Moreover,

it follows from the assumptions on f and the definition of φ that

U 0 ∩ M = p ∈ U 0  f (p) = 0 = p ∈ U 0  φ(p) ∈ Rm × {0}

   

and hence
φ0 (U 0 ∩ M ) = Ω0 ∩ Rm × {0} .

Hence the diffeomorphism φ0 : U 0 → Ω0 satisfies the requirements of part (ii).

This proves Theorem 2.1.10.

Definition 2.1.11. Let U ⊂ Rk be an open set and f : U → R` be a smooth

map. An element c ∈ R` is called a regular value of f iff, for all p ∈ U ,
we have

f (p) = c =⇒ df (p) : Rk → R` is surjective.

Otherwise c is called a singular value of f . Theorem 2.1.10 asserts that,

if c is a regular value of f , then the preimage

M := f −1 (c) = p ∈ U  f (p) = c
 

is a smooth (k − `)-dimensional submanifold of Rk .

2.1. SUBMANIFOLDS OF EUCLIDEAN SPACE 21

Examples and Exercises

Example 2.1.12. Let A = AT ∈ Rk×k be a symmetric matrix and define
the function f : Rk → R by

f (x) := xT Ax.

Then df (x)ξ = 2xT Aξ for x, ξ ∈ Rk and hence the linear map df (x) : Rk → R
is surjective if and only if Ax 6= 0. Thus c = 0 is the only singular value of f
and, for c ∈ R \ {0}, the set
n o
M := f −1 (c) = x ∈ Rk  xT Ax = c


is a smooth manifold of dimension m = k − 1.

Example 2.1.13 (The sphere). As a special case of Example 2.1.12 con-

sider the case k = m + 1, A = 1l, and c = 1. Then f (x) = |x|2 and so we
have another proof that the unit sphere
n o
S m = x ∈ Rm+1  |x|2 = 1


in Rm+1 is a smooth m-manifold. (See Examples 1.2.4 and 2.1.5.)

Example 2.1.14. Define the map f : R3 × R3 → R by f (x, y) := |x − y|2 .

This is another special case of Example 2.1.12 and so, for every r > 0, the
set
M := (x, y) ∈ R3 × R3  |x − y| = r
 

is a smooth 5-manifold.

Example 2.1.15 (The 2-torus). Let 0 < r < 1 and define f : R3 → R by

f (x, y, z) := (x2 + y 2 + r2 − z 2 − 1)2 − 4(x2 + y 2 )(r2 − z 2 ).

This map has zero as a regular value and M := f −1 (0) is diffeomorphic to

the 2-torus T2 = S 1 × S 1 . An explicit diffeomorphism is given by

(eis , eit ) 7→ (1 + r cos(s)) cos(t), (1 + r cos(s)) sin(t), r sin(s) .

This example corresponds to the second surface in Figure 2.2.

p
Exercise: Show that f (x, y, z) = 0 if and only if ( x2 + y 2 − 1)2 + z 2 = r2 .
Verify that zero is a regular value of f .
22 CHAPTER 2. FOUNDATIONS

Example 2.1.16. The set

M := (x2 , y 2 , z 2 , yz, zx, xy) | x, y, z ∈ R, x2 + y 2 + z 2 = 1



is a smooth 2-manifold in R6 . To see this, define an equivalence relation on

the unit sphere S 2 ⊂ R3 by p ∼ q iff q = ±p. The quotient space (the set of
equivalence classes) is called the real projective plane and is denoted by

RP2 := S 2 /{±1}.

(See Example 1.2.6.) It is equipped with the quotient topology, i.e. a sub-
set U ⊂ RP2 is open, by definition, iff its preimage under the obvious projec-
tion S 2 → RP2 is an open subset of S 2 . Now the map f : S 2 → R6 defined
by
f (x, y, z) := (x2 , y 2 , z 2 , yz, zx, xy)
descends to a homeomorphism from RP2 onto M . The submanifold M is
covered by the local smooth parameterizations
p
Ω → M : (x, y) 7→ f (x, y, 1 − x2 − y 2 ),
p
Ω → M : (x, z) 7→ f (x, 1 − x2 − z 2 , z),
p
Ω → M : (y, z) 7→ f ( 1 − y 2 − z 2 , y, z),

defined on the open unit disc Ω ⊂ R2 . We remark the following.

(a) M is not the preimage of a regular value under a smooth map R6 → R4 .
(b) M is not diffeomorphic to a submanifold of R3 .
(c) The projection Σ := (yz, zx, xy) | x, y, z ∈ R, x2 + y 2 + z 2 = 1 of M


onto the last three coordinates is called the Roman surface and was dis-
covered by Jakob Steiner. The Roman surface can also be represented as
the set of solutions (ξ, η, ζ) ∈ R3 of the equation η 2 ζ 2 + ζ 2 ξ 2 + ξ 2 η 2 = ξηζ.
It is not a submanifold of R3 .
Exercise: Prove this. Show that M is diffeomorphic to a submanifold of R4 .
Show that M is diffeomorphic to RP2 as defined in Example 1.2.6.

Exercise 2.1.17. Let V : Rn → R be a smooth function and define the

Hamiltonian function H : Rn × Rn → R (kinetic plus potential energy) by
1 2
H(x, y) := |y| + V (x).
2
Prove that c is a regular value of H if and only if it is a regular value of V .
2.1. SUBMANIFOLDS OF EUCLIDEAN SPACE 23

Exercise 2.1.18. Consider the general linear group

GL(n, R) = g ∈ Rn×n | det(g) 6= 0


Prove that the derivative of the function f = det : Rn×n → R is given by

df (g)v = det(g) trace(g −1 v)
for every g ∈ GL(n, R) and every v ∈ Rn×n . Deduce that the special linear
group
SL(n, R) := {g ∈ GL(n, R) | det(g) = 1}
is a smooth submanifold of Rn×n .
Example 2.1.19. The orthogonal group
n o
O(n) := g ∈ Rn×n | g T g = 1l

is a smooth submanifold of Rn×n . To see this, denote by

n o
Sn := S ∈ Rn×n | S T = S

the vector space of symmetric matrices and define f : Rn×n → Sn by

f (g) := g T g.
Its derivative df (g) : Rn×n → Sn is given by
df (g)v = g T v + v T g.
This map is surjective for every g ∈ O(n): if g T g = 1l and S = S T ∈ Sn ,
then the matrix v := 21 gS satisfies
1 1 1 1
df (g)v = g T gS + (gS)T g = S + S T = S.
2 2 2 2
Hence 1l is a regular value of f and so O(n) is a smooth manifold. It has
the dimension
n(n + 1) n(n − 1)
dim O(n) = n2 − dim Sn = n2 − = .
2 2
Exercise 2.1.20. Prove that the set
M := (x, y) ∈ R2 | xy = 0


is not a submanifold of R2 . Hint: If U ⊂ R2 is a neighborhood of the

origin and f : U → R is a smooth map such that U ∩ M = f −1 (0), then
df (0, 0) = 0.
24 CHAPTER 2. FOUNDATIONS

p + Tp M
Tp M v p
0

Figure 2.3: The tangent space Tp M and the translated tangent space p+Tp M .

2.2 Tangent Spaces and Derivatives

The main reason for first discussing the extrinsic notion of embedded mani-
folds in Euclidean space as explained in the Master Plan §1.5 is that the
concept of a tangent vector is much easier to digest in the embedded case:
it is simply the derivative of a curve in M , understood as a vector in the
ambient Euclidean space in which M is embedded.

2.2.1 Tangent Space

Definition 2.2.1. Let M ⊂ Rk be a smooth m-dimensional manifold and
fix a point p ∈ M . A vector v ∈ Rk is called a tangent vector of M at p
iff there exists a smooth curve γ : R → M such that
γ(0) = p, γ̇(0) = v.
The set
Tp M := {γ̇(0) | γ : R → M is smooth, γ(0) = p}
of tangent vectors of M at p is called the tangent space of M at p.
Theorem 2.2.3 below shows that Tp M is a linear subspace of Rk . As does
any linear subspace it contains the origin; it need not actually intersect M .
Its translate p + Tp M touches M at p; this is what you should visualize
for Tp M (see Figure 2.3).
Remark 2.2.2. Let p ∈ M be as in Definition 2.2.1 and let v ∈ Rk . Then

∃ε > 0 ∃γ : (−ε, ε) → M such that
v ∈ Tp M ⇐⇒
γ is smooth, γ(0) = p, γ̇(0) = v.
To see this suppose that γ : (−ε, ε) → M is a smooth curve with γ(0) = p
e : R → M by
and γ̇(0) = v. Define γ
 
εt
γ
e(t) := γ √ , t ∈ R.
ε2 + t 2
Then γe is smooth and satisfies γ
e(0) = p and γ e˙ (0) = v. Hence v ∈ Tp M .
2.2. TANGENT SPACES AND DERIVATIVES 25

Theorem 2.2.3 (Tangent spaces). Let M ⊂ Rk be a smooth m-dimen-

sional manifold and fix a point p ∈ M . Then the following holds.
(i) Let U0 ⊂ M be an M -open set with p ∈ U0 and let φ0 : U0 → Ω0
be a diffeomorphism onto an open subset Ω0 ⊂ Rm . Let x0 := φ0 (p) and
let ψ0 := φ−1
0 : Ω0 → U0 be the inverse map. Then
 
Tp M = im dψ0 (x0 ) : Rm → Rk .

(ii) Let U, Ω ⊂ Rk be open sets and φ : U → Ω be a diffeomorphism such

that p ∈ U and φ(U ∩ M ) = Ω ∩ (Rm × {0}). Then

Tp M = dφ(p)−1 (Rm × {0}) .

(iii) Let U ⊂ Rk be an open neighborhood of p and f : U → Rk−m be a

smooth map such that 0 is a regular value of f and U ∩ M = f −1 (0). Then

Tp M = ker df (p).

(iv) Tp M is an m-dimensional linear subspace of Rk .

Proof. Let ψ0 : Ω0 → U0 and x0 ∈ Ω0 be as in (i) and let φ : U → Ω be as

in (ii). We prove that

im dψ0 (x0 ) ⊂ Tp M ⊂ dφ(p)−1 (Rm × {0}) . (2.2.1)

To prove the first inclusion in (2.2.1), choose a constant r > 0 such that

Br (x0 ) := {x ∈ Rm | |x − x0 | < r} ⊂ Ω0 .

Now let ξ ∈ Rm and choose ε > 0 so small that

ε |ξ| ≤ r.

Then x0 + tξ ∈ Ω0 for all t ∈ R with |t| < ε. Define γ : (−ε, ε) → M by

γ(t) := ψ0 (x0 + tξ) for − ε < t < ε.

Then γ is a smooth curve in M satisfying


d 
γ(0) = ψ0 (x0 ) = p, γ̇(0) =  ψ0 (x0 + tξ) = dψ0 (x0 )ξ.
dt t=0

Hence it follows from Remark 2.2.2 that dψ0 (x0 )ξ ∈ Tp M , as claimed.

26 CHAPTER 2. FOUNDATIONS

To prove the second inclusion in (2.2.1) we fix a vector v ∈ Tp M . Then,

by definition of the tangent space, there exists a smooth curve γ : R → M
such that γ(0) = p and γ̇(0) = v. Let U ⊂ Rk be as in (ii) and choose ε > 0
so small that γ(t) ∈ U for |t| < ε. Then
φ(γ(t)) ∈ φ(U ∩ M ) ⊂ Rm × {0}
for |t| < ε and hence

d 
dφ(p)v = dφ(γ(0))γ̇(0) =  φ(γ(t)) ∈ Rm × {0}.
dt t=0
This shows that v ∈ dφ(p)−1 (Rm × {0}) and thus we have proved (2.2.1).
Now the sets im dψ0 (x0 ) and dφ(p)−1 (Rm × {0}) are both m-dimensional
linear subspaces of Rk . Hence it follows from (2.2.1) that these subspaces
agree and that they both agree with Tp M . Thus we have proved asser-
tions (i), (ii), and (iv).
We prove (iii). Let v ∈ Tp M . Then there is a smooth curve γ : R → M
such that γ(0) = p and γ̇(0) = v. For t sufficiently small we have γ(t) ∈ U ,
where U ⊂ Rk is the open set in (iii), and f (γ(t)) = 0. Hence

d 
df (p)v = df (γ(0))γ̇(0) =  f (γ(t)) = 0.
dt t=0
This implies Tp M ⊂ ker df (p). Since Tp M and the kernel of df (p) are both
m-dimensional linear subspaces of Rk , we deduce that Tp M = ker df (p).
This proves part (iii) and Theorem 2.2.3.

Exercise 2.2.4. Let M ⊂ Rk be a smooth m-dimensional manifold and

let pi ∈ M be a sequence that converges to a point p ∈ M . Let τi be a
sequence of nonzero real numbers and let v ∈ Rk such that
pi − p
lim τi = 0, lim = v.
i→∞ i→0 τi
Prove that v ∈ Tp M . Hint: Use part (iii) of Theorem 2.2.3.
Example 2.2.5. Let A = AT ∈ Rk×k be a nonzero matrix as in Exam-
ple 2.1.12 and let c 6= 0. Then part (iii) of Theorem 2.2.3 asserts that the
tangent space of the manifold
n  o
M = x ∈ Rk  xT Ax = c


at a point x ∈ M is the (k − 1)-dimensional linear subspace

n  o
Tx M = ξ ∈ Rk  xT Aξ = 0 .

2.2. TANGENT SPACES AND DERIVATIVES 27

Example 2.2.6. As a special case of Example 2.2.5 with A = 1l and c = 1

we find that the tangent space of the unit sphere S m ⊂ Rm+1 at a point
x ∈ S m is the orthogonal complement of x:
Tx S m = x⊥ = ξ ∈ Rm+1 | hx, ξi = 0 .


Here hx, ξi = m m+1 .

P
i=0 xi ξi denotes the standard inner product on R
Exercise 2.2.7. What is the tangent space of the 5-manifold
M := (x, y) ∈ R3 × R3 | |x − y| = r


at a point (x, y) ∈ M ? (See Exercise 2.1.14.)

Example 2.2.8. Let H(x, y) := 21 |y|2 + V (x) be as in Exercise 2.1.17 and
let c be a regular value of H. If (x, y) ∈ M := H −1 (c), then
T(x,y) M = {(ξ, η) ∈ Rn × Rn | hy, ηi + h∇V (x), ξi = 0} .
Here ∇V := (∂V /∂x1 , . . . , ∂V /∂xn ) : Rn → Rn denotes the gradient of V .
Exercise 2.2.9. The tangent space of SL(n, R) at the identity matrix is the
space
sl(n, R) := T1l SL(n, R) = ξ ∈ Rn×n | trace(ξ) = 0


of traceless matrices. (Prove this, using Exercise 2.1.18.)

Example 2.2.10. The tangent space of O(n) at g is
n o
Tg O(n) = v ∈ Rn×n | g T v + v T g = 0 .

In particular, the tangent space of O(n) at the identity matrix is the space
of skew-symmetric matrices
n o
o(n) := T1l O(n) = ξ ∈ Rn×n | ξ T + ξ = 0

To see this, choose a smooth curve R → O(n) : t 7→ g(t). Then g(t)T g(t) = 1l
for all t ∈ R and, differentiating this identity with respect to t, we ob-
tain g(t)T ġ(t) + ġ(t)T g(t) = 0 for every t. Hence every matrix v ∈ Tg O(n)
satisfies the equation g T v + v T g = 0. With this understood, the claim fol-
lows from the fact that g T v + v T g = 0 if and only if the matrix ξ := g −1 v
is skew-symmetric and that the space of skew-symmetric matrices in Rn×n
has dimension n(n − 1)/2.
Exercise 2.2.11. Let Ω ⊂ Rm be an open set and h : Ω → Rk−m be a
smooth map. Prove that the tangent space of the graph of h at a point
(x, h(x)) is the graph of the derivative dh(x) : Rm → Rk−m :
M = {(x, h(x)) | x ∈ Ω} , T(x,h(x)) M = {(ξ, dh(x)ξ) | ξ ∈ Rm } .
28 CHAPTER 2. FOUNDATIONS

Exercise 2.2.12 (Monge coordinates). Let M be a smooth m-manifold

in Rk and suppose that p ∈ M is such that the projection Tp M → Rm × {0}
is invertible. Prove that there exists an open set Ω ⊂ Rm and a smooth
map h : Ω → Rk−m such that the graph of h is an M -open neighborhood
of p (see Example 2.1.6). Of course, the projection Tp M → Rm × {0} need
k


not be invertible, but it must be invertible for at least one of the m choices
of the m-dimensional coordinate plane. Hence every point of M has an M -
open neighborhood which may be expressed as a graph of a function of some
of the coordinates in terms of the others as in e.g. Example 2.1.5.

2.2.2 Derivative
A key purpose behind the concept of a smooth manifold is to carry over
the notion of a smooth map and its derivatives from the realm of first year
analysis to the present geometric setting. Here is the basic definition. It ap-
peals to the notion of a smooth map between arbitrary subsets of Euclidean
spaces as introduced on page 15.

Definition 2.2.13. Let M ⊂ Rk be an m-dimensional smooth manifold and

f : M → R`

be a smooth map. The derivative of f at a point p ∈ M is the map

df (p) : Tp M → R`

defined as follows. Given a tangent vector v ∈ Tp M choose a smooth curve

γ:R→M

satisfying
γ(0) = p, γ̇(0) = v.
Now define the vector df (p)v ∈ R` by

d  f (γ(h)) − f (p)
df (p)v :=  f (γ(t)) = lim . (2.2.2)
dt t=0 h→0 h

That the limit on the right in equation (2.2.2) exists follows from our
assumptions. We must prove, however, that the derivative is well defined,
i.e. that the right hand side of (2.2.2) depends only on the tangent vector v
and not on the choice of the curve γ used in the definition. This is the
content of the first assertion in the next theorem.
2.2. TANGENT SPACES AND DERIVATIVES 29

Theorem 2.2.14 (Derivatives). Let M ⊂ Rk be an m-dimensional smooth

manifold and f : M → R` be a smooth map. Fix a point p ∈ M . Then the
following holds.
(i) The right hand side of (2.2.2) is independent of γ.
(ii) The map df (p) : Tp M → R` is linear.
(iii) If N ⊂ R` is a smooth n-manifold and f (M ) ⊂ N , then

df (p)Tp M ⊂ Tf (p) N.

(iv) (Chain Rule) Let N be as in (iii), suppose that f (M ) ⊂ N , and

let g : N → Rd be a smooth map. Then

d(g ◦ f )(p) = dg(f (p)) ◦ df (p) : Tp M → Rd .

(v) If f = id : M → M , then df (p) = id : Tp M → Tp M .

Proof. We prove (i). Let v ∈ Tp M and γ : R → M be as in Definition 2.2.13.

By definition there is an open neighborhood U ⊂ Rk of p and a smooth
map F : U → R` such that

F (p0 ) = f (p0 ) for all p0 ∈ U ∩ M.

Let dF (p) ∈ R`×k denote the Jacobian matrix (i.e. the matrix of all first
partial derivatives) of F at p. Then, since γ(t) ∈ U ∩ M for t sufficiently
small, we have

dF (p)v = dF (γ(0))γ̇(0)

d 
=  F (γ(t))
dt t=0

d 
=  f (γ(t)).
dt t=0

The right hand side of this identity is independent of the choice of F while
the left hand side is independent of the choice of γ. Hence the right hand
side is also independent of the choice of γ and this proves (i). Assertion (ii)
follows immediately from the identity

df (p)v = dF (p)v

just established.
30 CHAPTER 2. FOUNDATIONS

Assertion (iii) follows directly from the definitions. Namely, if γ is as in

Definition 2.2.13, then β := f ◦ γ : R → N is a smooth curve in N satisfying
β(0) = f (γ(0)) = f (p) =: q, β̇(0) = df (p)v =: w.
Hence w ∈ Tq N . Assertion (iv) also follows directly from the definitions.
If g : N → Rd is a smooth map and β, q, w are as above, then

d 
d(g ◦ f )(p)v =  g(f (γ(t)))
dt t=0

d 
=  g(β(t))
dt t=0
= dg(q)w
= dg(f (p))df (p)v.
This proves (iv). Assertion (v) follows again directly from the definitions
and this proves Theorem 2.2.14.
Corollary 2.2.15 (Diffeomorphisms). Let M ⊂ Rk be a smooth m-mani-
fold and N ⊂ R` be a smooth n-manifold and let f : M → N be a diffeomor-
phism. Then m = n and the derivative df (p) : Tp M → Tf (p) N is a vector
space isomorphism with inverse
df (p)−1 = df −1 (f (p)) : Tf (p) N → Tp M
for all p ∈ M .
Proof. Define g := f −1 : N → M so that g ◦ f = idM and f ◦ g = idN .
Then it follows from Theorem 2.2.14 that, for p ∈ M and q := f (p) ∈ N , we
have dg(q) ◦ df (p) = id : Tp M → Tp M and df (p) ◦ dg(q) = id : Tq N → Tq N.
Hence df (p) : Tp M → Tq N is a vector space isomorphism and its inverss is
given by dg(q) = df (p)−1 : Tq N → Tp M. Hence m = n and this proves
Corollary 2.2.15.
Exercise 2.2.16. Let M ⊂ Rk be a smooth manifold and let f : M → R` be
a smooth map. Let pi ∈ M be a sequence that converges to a point p ∈ M ,
let τi be a sequence of nonzero real numbers, and let v ∈ Tp M such that
pi − p
lim τi = 0, lim = v.
i→∞ i→∞ τi
(See Exercise 2.2.4.) Prove that
f (pi ) − f (p)
lim = df (p)v.
i→∞ τi
Hint: Use the local extension F of f in the proof of Theorem 2.2.14.
2.2. TANGENT SPACES AND DERIVATIVES 31

2.2.3 The Inverse Function Theorem

Corollary 2.2.15 is analogous to the corresponding assertion for smooth maps
between open subsets of Euclidean space. Likewise, the inverse function
theorem for manifolds is a partial converse of Corollary 2.2.15.

M N
p0 f q0
U0 V0

φ0 ψ0
~
n ~ f ~ n
R U0 x0 V0 R

Figure 2.4: The Inverse Function Theorem.

Theorem 2.2.17 (Inverse Function Theorem). Assume that M ⊂ Rk

and N ⊂ R` are smooth m-manifolds and f : M → N is a smooth map.
Let p0 ∈ M and suppose that the derivative

df (p0 ) : Tp0 M → Tf (p0 ) N

is a vector space isomorphism. Then there exists an M -open neighbor-

hood U ⊂ M of p0 such that V := f (U ) ⊂ N is an N -open subset of N and
the restriction f |U : U → V is a diffeomorphism.

Proof. Choose coordinate charts φ0 : U0 → Ue0 , defined on an M -open neigh-

borhood U0 ⊂ M of p0 onto an open set U e0 ⊂ Rm , and ψ0 : V0 → Ve0 , defined
on an N -open neighborhood V0 ⊂ N of f (p0 ) onto an open set Ve0 ⊂ Rm .
Shrinking U0 , if necessary, we may assume that f (U0 ) ⊂ V0 . Then the map

fe := ψ0 ◦ f ◦ φ−1
0 : U0 → V0
e e

(see Figure 2.4) is smooth and its derivative dfe(x0 ) : Rm → Rm is bijective

at x0 := φ0 (p0 ). Hence the Inverse Function Theorem A.2.2 asserts that
there exists an open neighborhood U e ⊂U e0 of x0 such that Ve := fe(U
e ) is an
open subset of Ve0 and the restriction of fe to U
e is a diffeomorphism from U e
−1 e −1 e
to V . Hence the assertion holds with U := φ0 (U ) and V := ψ0 (V ). This
e
proves Theorem 2.2.17.
32 CHAPTER 2. FOUNDATIONS

Regular Values
Definition 2.2.18 (Regular value). Let M ⊂ Rk be a smooth m-manifold,
let N ⊂ R` be a smooth n-manifold, and let f : M → N be a smooth map.
An element q ∈ N is called a regular value of f iff, for every p ∈ M
with f (p) = q, the derivative df (p) : Tp M → Tf (p) N is surjective.
Theorem 2.2.19 (Regular Values). Let f : M → N be as in Defini-
tion 2.2.18 and let q ∈ N be a regular value of f . Then the set

P := f −1 (q) = {p ∈ M | f (p) = q}

is a smooth submanifold of Rk of dimension m − n and, for each point p ∈ P ,

its tangent space at p is given by

Tp P = ker df (p) = {v ∈ Tp M | df (p)v = 0} .

Proof 1. Fix a point p0 ∈ P and choose a linear map A : Rk → Rm−n such

that the restriction of A to the (m − n)-dimensional linear subspace

ker df (p0 ) = {v ∈ Tp0 M | df (p0 ) = 0} ⊂ Tp0 M ⊂ Rk

is a vector space isomorphism. Define the map F : M → N × Rm−n by

F (p) := (f (p), Ap)

for p ∈ M . The derivative of F at p0 is given by dF (p0 )v = (df (p0 )v, Av)

for v ∈ Tp0 M and is a vector space isomorphism. Hence, by Theorem 2.2.17
there exists an M -open neighborhood U ⊂ M of p0 such that V := F (U )
is an open subset of N × Rm−n in the relative topology and the restric-
tion F |U : U → V is a diffeomorphism. Hence the P -open set U ∩ P is
diffeomorphic to the open set Ω := {y ∈ Rm−n | (q, y) ∈ V } ⊂ Rm−n by the
diffeomorphism φ : U ∩ P → Ω, defined by φ(p) := Ap for p ∈ U ∩ P , whose
inverse is the smooth map ψ : Ω → U ∩ P given by ψ(y) = (F |U )−1 (q, y)
for y ∈ Ω. This shows that P is a smooth (m − n)-manifold in Rk .
Now let p ∈ P and v ∈ Tp P . Then there exists a smooth curve γ : R → P
such that γ(0) = p and γ̇(0) = v. Since f (γ(t)) = 0 for all t, we have

d 
df (p)v =  f (γ(t)) = 0
dt t=0

and so v ∈ ker df (p). Hence Tp P ⊂ ker df (p) and equality holds because
both Tp P and ker df (p) are (m − n)-dimensional linear subspaces of Rk .
This proves Theorem 2.2.19.
2.2. TANGENT SPACES AND DERIVATIVES 33

Proof 2. Here is another proof of Theorem 2.2.19 in local coordinates. Fix

a point p0 ∈ P and choose a coordinate chart φ0 : U0 → φ0 (U0 ) ⊂ Rm on
an M -open neighborhood U0 ⊂ M of p0 . Likewise, choose a coordinate
chart ψ0 : V0 → ψ0 (V0 ) ⊂ Rn on an N -open neighborhood V0 ⊂ N of q.
Shrinking U0 , if necessary, we may assume that f (U0 ) ⊂ V0 . Then the
point c0 := ψ0 (q) is a regular value of the map
f0 := ψ0 ◦ f ◦ φ−1 n
0 : φ0 (U0 ) → R .

Namely, if x ∈ φ0 (U0 ) satisfies f0 (x) = c0 , then p := φ−1

0 (x) ∈ U0 ∩ P , so the
−1
maps dφ0 (x) : R → Tp M , df (p) : Tp M → Tq N , and dψ0 (q) : Tq N → Rn
m

are all surjective, hence so is their composition, and by the chain rule this
composition is the derivative df0 (x) : Rm → Rn . With this understood, it
follows from Theorem 2.1.10 that the set
f0−1 (c0 ) = x ∈ φ0 (U0 ) | f (φ−1

0 (x)) = q = φ0 (U0 ∩ P )

is a manifold of of dimension m − n contained in the open set φ0 (U0 ) ⊂ Rm .

Using Definition 2.1.3 and shrinking U0 further, if necessary, we may assume
that the set φ0 (U0 ∩ P ) is diffeomorphic to an open subset of Rm−n . Com-
posing this diffeomorphism with φ0 we find that U0 ∩ P is diffeomorphic to
the same open subset of Rm−n . Since the set U0 ⊂ M is M -open, there
exists an open set U ⊂ Rk such that U ∩ M = U0 , hence U ∩ P = U0 ∩ P ,
and so U0 ∩ P is a P -open neighborhood of p0 . Thus we have proved
that every element p0 ∈ P has a P -open neigborhood that is diffeomor-
phic to an open subset of Rm−n . Thus P ⊂ Rk is a manifold of dimen-
sion m − n (Definition 2.1.3). The proof that the tangent spaces of P are
given by Tp P = ker df (p) remains unchanged and this completes the second
proof of Theorem 2.2.19.

Definition 2.2.20. Let M ⊂ Rk and N ⊂ R` be smooth m-manifolds. A

smooth map f : M → N is called a local diffeomorphism iff its deriva-
tive df (p) : Tp M → Tf (p) N is a vector space isomorphism for every p ∈ M .
Example 2.2.21. The inclusion of an M -open subset U ⊂ M into M and
the map R → S 1 : t 7→ eit are examples of local diffeomorphisms.
Exercise 2.2.22. Prove that the image of a local diffeomorphism is an open
subset of the target manifold. Hint: Use the Inverse Function Theorem.
In the terminology introduced in §2.3 and §2.6.1 below, local diffeomor-
phisms are both immersions and submersions. In particular, if f : M → N
is a local diffeomorphism, then every element q ∈ N is a regular value of f
and its preimage f −1 (q) is a discrete subset of M .
34 CHAPTER 2. FOUNDATIONS

2.3 Submanifolds and Embeddings

This section deals with subsets of a manifold M that are themselves mani-
folds as in Definition 2.1.3. Such subsets are called submanifolds of M .
Definition 2.3.1 (Submanifold). Let M ⊂ Rk be an m-dimensional man-
ifold. A subset L ⊂ M is called a submanifold of M of dimension `, iff L
itself is an `-manifold.
Definition 2.3.2 (Embedding). Let M ⊂ Rk be an m-dimensional mani-
fold and N ⊂ R` be an n-dimensional manifold. A smooth map f : N → M
is called an immersion iff its derivatve df (q) : Tq N → Tf (q) M is injective
for every q ∈ N . It is called proper if, for every compact subset K ⊂ f (N ),
the preimage f −1 (K) = {q ∈ N | f (q) ∈ K} is compact. The map f is called
an embedding iff it is a proper injective immersion.
Remark 2.3.3. In our definition of proper maps it is important that the
compact set K is required to be contained in the image of f . The literature
also contains a stronger definition of proper which requires that f −1 (K) is
a compact subset of N for every compact subset K ⊂ M , whether or not K
is contained in the image of f . This holds if and only if the map f is proper
in the sense of Definition 2.3.2 and has an M -closed image. (Exercise!)

U0 φ 0(U0)
M φ0
p
0
P

Figure 2.5: A coordinate chart adapted to a submanifold.

Theorem 2.3.4 (Submanifolds). Let M ⊂ Rk be an m-dimensional man-

ifold and N ⊂ R` be an n-dimensional manifold.
(i) If f : N → M is an embedding, then f (N ) is a submanifold of M .
(ii) If P ⊂ M is a submanifold, then the inclusion P → M is an embedding.
(iii) A subset P ⊂ M is a submanifold of dimension n if and only if, for
every p0 ∈ P , there exists a coordinate chart φ : U → Rm on an M -open
neighborhood U of p0 such that φ(U ∩ P ) = φ(U ) ∩ (Rn × {0}) (Figure 2.5).
(iv) A subset P ⊂ M is a submanifold of dimension n if and only if, for ev-
ery p0 ∈ P , there exists an M -open neighborhood U ⊂ M of p0 and a smooth
map g : U → Rm−n such that 0 is a regular value of g and U ∩ P = g −1 (0).
Proof. See page 36.
2.3. SUBMANIFOLDS AND EMBEDDINGS 35

Lemma 2.3.5 (Embeddings). Let M and N be as in Theorem 2.3.4,

let f : N → M be an embedding, let q0 ∈ N , and define

P := f (N ), p0 := f (q0 ) ∈ P.

Then there exists an M -open neighborhood U ⊂ M of p0 , an N -open neigh-

borhood V ⊂ N of q0 , an open neighborhood W ⊂ Rm−n of the origin, and
a diffeomorphism F : V × W → U such that, for all q ∈ V and all z ∈ W ,

F (q, 0) = f (q), (2.3.1)

F (q, z) ∈ P ⇐⇒ z = 0. (2.3.2)
Proof. Choose any coordinate chart φ0 : U0 → Rm on an M -open neighbor-
hood U0 ⊂ M of p0 . Then d(φ0 ◦ f )(q0 ) = dφ0 (f (q0 )) ◦ df (q0 ) : Tq0 N → Rm
is injective. Hence there is a linear map B : Rm−n → Rm such that the map

Tq0 N × Rm−n → Rm : (w, ζ) 7→ d(φ0 ◦ f )(q0 )w + Bζ (2.3.3)

is a vector space isomorphism. Define the set

Ω := (q, z) ∈ N × Rm−n | f (q) ∈ U0 , φ0 (f (q)) + Bz ∈ φ0 (U0 ) .



This is an open subset of N × Rm−n and we define F : Ω → M by

F (q, z) := φ−1
0 (φ0 (f (q)) + Bz) .

This map is smooth, it satisfies F (q, 0) = f (q) for all q ∈ f −1 (U0 ), and
the derivative dF (q0 , 0) : Tq0 N × Rm−n → Tp0 M is the composition of the
map (2.3.3) with dφ0 (p0 )−1 : Rm → Tp0 M and so is a vector space isomor-
phism. Thus the Inverse Function Theorem 2.2.17 asserts that there is an
N -open neighborhood V0 ⊂ N of q0 and an open neighborhood W0 ⊂ Rm−n
of the origin such that V0 × W0 ⊂ Ω, the set U0 := F (V0 × W0 ) is M -open,
and the restriction of F to V0 × W0 is a diffeomorphism onto U0 . Thus we
have constructed a diffeomorphism F : V0 × W0 → U0 that satisfies (2.3.1).
We claim that the restriction of F to the product V × W of sufficiently
small open neighborhoods V ⊂ N of q0 and W ⊂ Rm−n of the origin also
satisfies (2.3.2). Otherwise, there exist sequences qi ∈ V0 converging to q0
and zi ∈ W0 \ {0} converging to zero such that F (qi , zi ) ∈ P . Hence there
exists a sequence qi0 ∈ N such that F (qi , zi ) = f (qi0 ). This sequence converges
to f (q0 ). Since f is proper we may assume, passing to a suitable subsequence
if necessary, that qi0 converges to a point q00 ∈ N . Then

f (q00 ) = lim f (qi0 ) = lim F (qi , zi ) = f (q0 ).

i→∞ i→∞
36 CHAPTER 2. FOUNDATIONS

Since f is injective, this implies q00 = q0 . Hence (qi0 , 0) ∈ V0 × W0 for i

sufficiently large and F (qi0 , 0) = f (qi0 ) = F (qi , zi ). This contradicts the fact
that the map F : V0 × W0 → M is injective, and proves Lemma 2.3.5.

Proof of Theorem 2.3.4. We prove (i). Let q0 ∈ N , denote p0 := f (q0 ) ∈ P ,

and choose a diffeomorphism F : V × W → U as in Lemma 2.3.5. Then the
set V ⊂ N is diffeomorphic to an open subset of Rn (after schrinking V
if necessary), the set U ∩ P is P -open because U ⊂ M is M -open, and we
have U ∩ P = {F (q, 0) | q ∈ V } = f (V ) by (2.3.1) and (2.3.2). Hence the
map f : V → U ∩ P is a diffeomorphism whose inverse is the composition
of the smooth maps F −1 : U ∩ P → V × W and V × W → V : (q, z) 7→ q.
Hence a P -open neighborhood of p0 is diffeomorphic to an open subset of Rn .
Since p0 ∈ P was chosen arbitrary, this shows that P is an n-dimensional
submanifold of M .
We prove (ii). The inclusion ι : P → M is obviously smooth and in-
jective (it extends to the identity map on Rk ). Moreover, Tp P ⊂ Tp M for
every p ∈ P and the derivative dι(p) : Tp P → Tp M is the obvious inclusion
for every p ∈ P . That ι is proper follows immediately from the definition.
Hence ι is an embedding.
We prove (iii). If a coordinate chart φ0 as in (iii) exists, then the
set U0 ∩ P is P -open and is diffeomorphic to an open subset of Rn . Since the
point p0 ∈ P was chosen arbitrary this proves that P is an n-dimensional
submanifold of M . Conversely, suppose that P is an n-dimensional sub-
manifold of M and let p0 ∈ P . Choose any coordinate chart φ0 : U0 → Rm
of M defined on an M -open neighborhood U0 ⊂ M of p0 . Then φ0 (U0 ∩ P )
is an n-dimensional submanifold of Rm . Hence Theorem 2.1.10 asserts
that there are open sets V, W ⊂ Rm with p0 ∈ V ⊂ φ0 (U0 ) and a diffeo-
morphism ψ : V → W such that

φ0 (p0 ) ∈ V, ψ(V ∩ φ0 (U0 ∩ P )) = W ∩ (Rn × {0}).

Now define U := φ−1 0 (V ) ⊂ U0 . Then p0 ∈ U , the chart φ0 restricts to a

diffeomorphism from U to V , the composition φ := ψ ◦ φ0 |U : U → W is a
diffeomorphism, and φ(U ∩ P ) = ψ(V ∩ φ0 (U0 ∩ P )) = W ∩ (Rn × {0}).
We prove (iv). That the condition is sufficient follows directly from
Theorem 2.2.19. To prove that it is necessary, assume that P ⊂ M , is a
submanifold of dimension n, fix an element p0 ∈ P , and choose a coordi-
nate chart φ : U → Rm on an M -open neighborhood U ⊂ M of p0 as in
part (iii). Define the map g : U → Rm−n by g(p) := (φn+1 (p), . . . , φm (p))
for p ∈ U . Then 0 is a regular value of g and g −1 (0) = U ∩ P . This proves
Theorem 2.3.4.
2.3. SUBMANIFOLDS AND EMBEDDINGS 37

Example 2.3.6. Let S 1 ⊂ R2 ∼ = C be the unit circle and consider the map
1 2
f : S → R given by f (x, y) := (x, xy). This map is a proper immersion but
is not injective (the points (0, 1) and (0, −1) have the same image under f ).
The image f (S 1 ) is a figure 8 in R2 and is not a submanifold (Figure 2.6).

Figure 2.6: A proper immersion.

Example 2.3.7. Consider the restriction of the map f in Example 2.3.6 to

the submanifold N := S 1 \ {(0, −1)}. The resulting map f : N → R2 is an
injective immersion but it is not proper. It has the same image as before
and hence f (N ) is not a manifold.
Example 2.3.8. The map f : R → R2 given by f (t) := (t2 , t3 ) is proper
and injective, but is not an embedding (its derivatuve at t = 0 is not
in-
2 3 2

jective). The image of f is the set f (R) = C := (x, y) ∈ R | x = y (see
Figure 2.7) and is not a submanifold. (Prove this!)

Figure 2.7: A proper injection.

Example 2.3.9. Define the map f : R → R2 by f (t) := (cos(t), sin(t)). This

map is an immersion, but it is neither injective nor proper. However, its
image is the unit circle in R2 and hence is a submanifold of R2 . The
map R → R2 : t 7→ f (t3 ) is not an immersion and is neither injective nor
proper, but its image is still the unit circle.
38 CHAPTER 2. FOUNDATIONS

2.4 Vector Fields and Flows

2.4.1 Vector Fields
Definition 2.4.1 (Vector Field). Let M ⊂ Rk be a smooth m-manifold.
A (smooth) vector field on M is a smooth map X : M → Rk such that

X(p) ∈ Tp M

for every p ∈ M . The set of smooth vector fields on M will be denoted by

n o
Vect(M ) := X : M → Rk | X is smooth, X(p) ∈ Tp M for all p ∈ M .

Exercise 2.4.2. Prove that the set of smooth vector fields on M is a real
vector space.
Example 2.4.3. Denote the standard cross product on R3 by
 
x 2 y3 − x 3 y2
x × y :=  x3 y1 − x1 y3 
x 1 y2 − x 2 y1

for x, y ∈ R3 . Fix a vector ξ ∈ S 2 and define the maps X, Y : S 2 → R3 by

X(p) := ξ × p, Y (p) := (ξ × p) × p.

These are vector fields with zeros ±ξ. Their integral curves (see Defini-
tion 2.4.6 below) are illustrated in Figure 2.8.

Figure 2.8: Two vector fields on the 2-sphere.

Example 2.4.4. Let M := R2 . A vector field on M is then any smooth

map X : R2 → R2 . As an example consider the vector field

X(x, y) := (x, −y).

This vector field has a single zero at the origin and its integral curves are
illustrated in Figure 2.9.
2.4. VECTOR FIELDS AND FLOWS 39

Figure 2.9: A hyperbolic fixed point.

Example 2.4.5. Every smooth function f : Rm → R determines a gradient

vector field  
∂f
 ∂x 
 1 
 ∂f 
 
∂x  : Rm → Rm .
 
X = ∇f :=  2
 . . 
 . 
 
 ∂f 
∂xm

Definition 2.4.6 (Integral curves). Let M ⊂ Rk be a smooth m-manifold,

let X ∈ Vect(M ) be a smooth vector field on M , and let I ⊂ R be an open
interval. A smooth map γ : I → M is called an integral curve of X iff it
satisfies the equation
γ̇(t) = X(γ(t))

for every t ∈ I.

Theorem 2.4.7. Let M ⊂ Rk be a smooth m-manifold and X ∈ Vect(M )

be a smooth vector field on M . Fix a point p0 ∈ M . Then the following
holds.
(i) There is an open interval I ⊂ R containing 0 and a smooth curve
γ : I → M satisfying the equation

γ̇(t) = X(γ(t)), γ(0) = p0 (2.4.1)

for every t ∈ I.
(ii) If γ1 : I1 → M and γ2 : I2 → M are two solutions of (2.4.1) on open
intervals I1 and I2 containing 0, then γ1 (t) = γ2 (t) for every t ∈ I1 ∩ I2 .
40 CHAPTER 2. FOUNDATIONS

Proof. We prove (i). Let φ0 : U0 → Rm be a coordinate chart on M , defined

on an M -open neighborhood U0 ⊂ M of p0 . The image of φ0 is an open set

Ω := φ0 (U0 ) ⊂ Rm

and we denote the inverse map by ψ0 := φ−1 0 : Ω → M . Then, by Theo-

rem 2.2.3, the derivative dψ0 (x) : Rm → Rk is injective and its image is the
tangent space Tψ0 (x) M for every x ∈ Ω. Define f : Ω → Rm by

f (x) := dψ0 (x)−1 X(ψ0 (x)), x ∈ Ω.

This map is smooth and hence, by the basic existence and uniqueness the-
orem for ordinary differential equations in Rm (see [46]), the equation

ẋ(t) = f (x(t)), x(0) = x0 := φ0 (p0 ), (2.4.2)

has a solution x : I → Ω on some open interval I ⊂ R containing 0. Hence

the function
γ := ψ0 ◦ x : I → U0 ⊂ M

is a smooth solution of (2.4.1). This proves (i).

The local uniqueness theorem asserts that two solutions γi : Ii → M
of (2.4.1) for i = 1, 2 agree on the interval (−ε, ε) ⊂ I1 ∩ I2 for ε > 0
sufficiently small. This follows immediately from the standard uniqueness
theorem for the solutions of (2.4.2) in [46] and the fact that x : I → Ω is a
solution of (2.4.2) if and only if γ := ψ0 ◦ x : I → U0 is a solution of (2.4.1).
To prove (ii) we observe that the set

I := I1 ∩ I2

is an open interval containing zero and hence is connected. Now consider

the set
A := {t ∈ I | γ1 (t) = γ2 (t)} .

This set is nonempty, because 0 ∈ A. It is closed, relative to I, because the

maps γ1 : I → M and γ2 : I → M are continuous. Namely, if ti ∈ I is a
sequence converging to t ∈ I, then γ1 (ti ) = γ2 (ti ) for every i and, taking
the limit i → ∞, we obtain γ1 (t) = γ2 (t) and hence t ∈ A. The set A is also
open by the local uniqueness theorem. Since I is connected it follows that
A = I. This proves (ii) and Theorem 2.4.7.
2.4. VECTOR FIELDS AND FLOWS 41

2.4.2 The Flow of a Vector Field

Definition 2.4.8 (The flow of a vector field). Let M ⊂ Rk be a smooth
m-manifold and X ∈ Vect(M ) be a smooth vector field on M . For p0 ∈ M
the maximal existence interval of p0 is the open interval
[   I ⊂ R is an open interval containing 0 
I(p0 ) := I .
and there is a solution γ : I → M of (2.4.1)

By Theorem 2.4.7 equation (2.4.1) has a solution γ : I(p0 ) → M . The flow

of X is the map φ : D → M defined by

D := {(t, p0 ) | p0 ∈ M, t ∈ I(p0 )}

and φ(t, p0 ) := γ(t), where γ : I(p0 ) → M is the unique solution of (2.4.1).

Theorem 2.4.9. Let M ⊂ Rk be a smooth m-manifold and X ∈ Vect(M )
be a smooth vector field on M . Let φ : D → M be the flow of X. Then the
following holds.
(i) D is an open subset of R × M .
(ii) The map φ : D → M is smooth.
(iii) Let p0 ∈ M and s ∈ I(p0 ). Then

I(φ(s, p0 )) = I(p0 ) − s (2.4.3)

and, for every t ∈ R with s + t ∈ I(p0 ), we have

φ(s + t, p0 ) = φ(t, φ(s, p0 )). (2.4.4)

Proof. See page 42.

Lemma 2.4.10. Let M , X, D, φ be as in Theorem 2.4.9 and let K ⊂ M be

a compact set. Then there exists an M -open set U ⊂ M and an ε > 0 such
that K ⊂ U , (−ε, ε) × U ⊂ D, and φ is smooth on (−ε, ε) × U .
Proof. In the case where M = Ω is an open subset of Rm this was proved
in [47, Thm 4.1.4]. Using local coordinates we deduce (as in the proof
of Theorem 2.4.7) that, for every p ∈ M , there exists an M -open neighbor-
hood Up ⊂ M of p and an εp > 0 such that (−εp , εp )×Up ⊂ D and the restric-
tion of φ to (−εp , εp ) × Up is smooth. Using this observation Sfor every p ∈ K
(and the axiom of choice) we obtain an M -open cover K ⊂ p∈K Up . Since
the set K is compact there exists a finite subcover K ⊂ Up1 ∪ · · · ∪ UpN =: U .
Now define ε := min{εp1 , . . . , εpN } to deduce that (−ε, ε) × U ⊂ D and φ is
smooth on (−ε, ε) × U . This proves Lemma 2.4.10.
42 CHAPTER 2. FOUNDATIONS

Proof of Theorem 2.4.9. We prove (iii). The map γ : I(p0 ) − s → M defined

by γ(t) := φ(s+t, p0 ) is a solution of the initial value problem γ̇(t) = X(γ(t))
with γ(0) = φ(s, p0 ). Hence I(p0 ) − s ⊂ I(φ(s, p0 )) and equation (2.4.4)
holds for every t ∈ R with s + t ∈ I(p0 ). In particular, with t = −s, we
have p0 = φ(−s, φ(s, p0 )). Thus we obtain equality in equation (2.4.3) by
the same argument with the pair (s.p0 ) replaced by (−s, φ(s, p0 )).
We prove (i) and (ii). Let (t0 , p0 ) ∈ D so that p0 ∈ M and t0 ∈ I(p0 ).
Suppose t0 ≥ 0. Then K := {φ(t, p0 ) | 0 ≤ t ≤ t0 } is a compact subset
of M . (It is the image of the compact interval [0, t0 ] under the unique
solution γ : I(p0 ) → M of (2.4.1).) Hence, by Lemma 2.4.10, there is an
M -open set U ⊂ M and an ε > 0 such that

K ⊂ U, (−ε, ε) × U ⊂ D,

and φ is smooth on (−ε, ε) × U . Choose N so large that t0 /N < ε. Define

U0 := U and, for k = 1, . . . , N , define the sets Uk ⊂ M inductively by

Uk := {p ∈ U | φ(t0 /N, p) ∈ Uk−1 } .

These sets are open in the relative topology of M .

We prove by induction on k that (−ε, kt0 /N + ε) × Uk ⊂ D and φ is
smooth on (−ε, kt0 /N + ε) × Uk . For k = 0 this holds by definition of ε
and U . If k ∈ {1, . . . , N } and the assertion holds for k − 1, then we have

p ∈ Uk =⇒ p ∈ U, φ(t0 /N, p) ∈ Uk−1

=⇒ (−ε, ε) ⊂ I(p), (−ε, (k − 1)t0 /N + ε) ⊂ I(φ(t0 /N, p))
=⇒ (−ε, kt0 /N + ε) ⊂ I(p).

Here the last implication follows from (2.4.3). Moreover, for p ∈ Uk and
t0 /N − ε < t < kt0 /N + ε, we have, by (2.4.4), that

φ(t, p) = φ(t − t0 /N, φ(t0 /N, p))

Since φ(t0 /N, p) ∈ Uk−1 for p ∈ Uk the right hand side is a smooth map
on the open set (t0 /N − ε, kt0 /N + ε) × Uk . Since Uk ⊂ U , φ is also a
smooth map on (−ε, ε) × Uk and hence on (−ε, kt0 /N + ε) × Uk . This
completes the induction. With k = N we have found an open neighborhood
of (t0 , p0 ) contained in D, namely the set (−ε, t0 + ε) × UN , on which φ is
smooth. The case t0 ≤ 0 is treated similarly. This proves (i) and (ii) and
Theorem 2.4.9.

Definition 2.4.11. A vector field X ∈ Vect(M ) is called complete if, for

each p0 ∈ M , there is an integral curve γ : R → M of X with γ(0) = p0 .
2.4. VECTOR FIELDS AND FLOWS 43

Lemma 2.4.12. Let M ⊂ Rk be a compact manifold. Then every vector

field on M is complete.
Proof. Let X ∈ Vect(M ). It follows from Lemma 2.4.10 with K = M
that there exists an ε > 0 such that (−ε, ε) ⊂ I(p) for all p ∈ M . By Theo-
rem 2.4.9 this implies I(p) = R for all p ∈ M . Hence X is complete.
Let M ⊂ Rk be a smooth manifold and X ∈ Vect(M ). Then
X is complete ⇐⇒ I(p) = R ∀ p ∈ M ⇐⇒ D = R × M.
Assume X is complete, let φ : R × M → M be the flow of X, and define
the map φt : M → M by φt (p) := φ(t, p) for t ∈ R and p ∈ M . Then
Theorem 2.4.9 asserts that φt is smooth for every t ∈ R and that
φs+t = φs ◦ φt , φ0 = id (2.4.5)
for all s, t ∈ R. In particular this implies that φt ◦ φ−t = φ−t ◦ φt = id.
Hence φt is bijective and (φt )−1 = φ−t , so each φt is a diffeomorphism.
Exercise 2.4.13. Let M ⊂ Rk be a smooth manifold. A vector field X on M
is said to have compact support iff there exists a compact subset K ⊂ M
such that X(p) = 0 for every p ∈ M \ K. Prove that every vector field with
compact support is complete.
We close this subsection with an important observation about incomplete
vector fields. The lemma asserts that an integral curve on a finite existence
interval must leave every compact subset of M .
Lemma 2.4.14. Let M ⊂ Rk be a smooth m-manifold, let X ∈ Vect(M ),
let φ : D → M be the flow of X, let K ⊂ M be a compact set, and let p0 ∈ M
be an element such that
I(p0 ) ∩ [0, ∞) = [0, b), 0 < b < ∞.
Then there exists a number 0 < tK < b such that
tK < t < b =⇒ φ(t, p0 ) ∈ M \ K
Proof. By Lemma 2.4.10 there exists an ε > 0 such that (−ε, ε) ⊂ I(p) for
every p ∈ K. Choose ε so small that ε < b and define
tK := b − ε > 0.
Choose a real number tK < t < b. Then I(φ(t, p0 )) ∩ [0, ∞) = [0, b − t) by
equation (2.4.3) in part (ii) of Theorem 2.4.9. Since 0 < b − t < b − tK = ε,
this shows that (−ε, ε) 6⊂ I(φ(t, p0 )) and hence φ(t, p0 ) ∈
/ K. This proves
Lemma 2.4.14.
44 CHAPTER 2. FOUNDATIONS

The next corollary is an immediate consequence of Lemma 2.4.14. In

this formulation the result will be used in §4.6 and in §5.4.
Corollary 2.4.15. Let M ⊂ Rk be a smooth m-manifold, let X ∈ Vect(M ),
and let γ : (0, T ) → M be an integral curve of X. If there exists a compact
set K ⊂ M that contains the image of γ, then γ extends to an integral curve
of X on the interval (−ρ, T + ρ) for some ρ > 0.
Proof. Here is another more direct proof that does not rely on Lemma 2.4.10.
Since K is compact, there exists a constant c > 0 such that |X(p)| ≤ c for
all p ∈ K. Since γ(t) ∈ K for 0 < t < T , this implies
Z t  Z t Z t
 
|γ(t) − γ(s)| =  γ̇(r) dr ≤
  |γ̇(r)| dr = |X(γ(r))| dr ≤ c(t − s)
s s s

for 0 < s < t < T . Thus the limit p0 := limt&0 γ(t) exists in Rk and, since K
is a closed subset of Rk , we have p0 ∈ K ⊂ M . Define γ0 : [0, T ) → M by

p0 , for t = 0,
γ0 (t) :=
γ(t), for 0 < t < T.
We prove that γ0 is differentiable at t = 0 and γ̇0 (0) = X(p0 ). To see this, fix
a constant ε > 0. Since the curve [0, T ) → Rk : t 7→ X(γ(t)) is continuous,
there exists a constant δ > 0 such that
0<t≤δ =⇒ |X(γ(t)) − X(p0 )| ≤ ε.
Hence, for 0 < s < t ≤ δ, we have
Z t 
 
|γ(t) − γ(s) − (t − s)X(p0 )| =  (γ̇(r) − X(p0 )) dr

Zs t 
 
=  (X(γ(r)) − X(p0 )) dr

s
Z t
≤ |X(γ(r)) − X(p0 )| dr
s
≤ (t − s)ε.
Take the limit s → 0 to obtain
 
 γ(t) − p0  |γ(t) − γ(s) − (t − s)X(p0 )|
 − X(p 0  = lim
)  ≤ε
 t s→0 t−s
for 0 < t ≤ δ. Thus γ0 is differentiable at t = 0 with γ̇0 (0) = X(p0 ), as
claimed. Hence γ extends to an integral curve γ e : (−ρ, T ) → M of X for
some ρ > 0 via γe(t) := φ(t, p0 ) for −ρ < t ≤ 0 and γ
e(t) := γ(t) for 0 < t < T .
Here φ is the flow of X. That γ also extends beyond t = T , follows by re-
placing γ(t) with γ(T −t) and X with −X. This proves Corollary 2.4.15.
2.4. VECTOR FIELDS AND FLOWS 45

The Group of Diffeomorphisms

Let us denote the space of diffeomorphisms of M by

Diff(M ) := {φ : M → M | φ is a diffeomorphism} .

This is a group. The group operation is composition and the neutral element
is the identity. Now equation (2.4.5) asserts that the flow of a complete
vector field X ∈ Vect(M ) is a group homomorphism

R → Diff(M ) : t 7→ φt .

This homomorphism is smooth and is characterized by the equation

d t
φ (p) = X(φt (p)), φ0 (p) = p
dt

for all p ∈ M and t ∈ R. We will often abbreviate this equation in the form

d t
φ = X ◦ φt , φ0 = id. (2.4.6)
dt

Exercise 2.4.16 (Isotopy). Let M ⊂ Rk be a compact manifold and I ⊂ R

be an open interval containing 0. Let

I × M → Rk : (t, p) 7→ Xt (p)

be a smooth map such that Xt ∈ Vect(M ) for every t. Prove that there is
a smooth family of diffeomorphisms I × M → M : (t, p) 7→ φt (p) satisfying

d
φt = Xt ◦ φt , φ0 = id (2.4.7)
dt

for every t ∈ I. Such a family of diffeomorphisms

I → Diff(M ) : t 7→ φt

is called an isotopy of M . Conversely prove that every smooth isotopy

I → Diff(M ) : t 7→ φt is generated (uniquely) by a smooth family of vector
fields I → Vect(M ) : t 7→ Xt .
46 CHAPTER 2. FOUNDATIONS

2.4.3 The Lie Bracket

Let M ⊂ Rk and N ⊂ R` be smooth m-manifolds and X ∈ Vect(M ) be
smooth vector field on M . If ψ : N → M is a diffeomorphism, then the
pullback of X under ψ is the vector field on N defined by

(ψ ∗ X)(q) := dψ(q)−1 X(ψ(q)) (2.4.8)

for q ∈ N . If φ : M → N is a diffeomorphism, then the pushforward of X

under ψ is the vector field on N defined by

(φ∗ X)(q) := dφ(φ−1 (q))X(φ−1 (q)) (2.4.9)

for q ∈ N .

Lemma 2.4.17. Let M ⊂ Rk , N ⊂ R` , and P ⊂ Rn be smooth m-dimen-

sional submanifolds and let X ∈ Vect(M ) and Z ∈ Vect(P ). Then

φ∗ X = (φ−1 )∗ X (2.4.10)

and
(ψ ◦ φ)∗ X = ψ∗ φ∗ X, (ψ ◦ φ)∗ Z = φ∗ ψ ∗ Z. (2.4.11)

Proof. Equation (2.4.10) follows from the fact that

dφ−1 (q) = dφ(φ−1 (q))−1 : Tq N → Tφ−1 (q) M

for all q ∈ N (Corollary 2.2.15) and the equations in (2.4.11) follow directly
from the chain rule (Theorem 2.2.14). This proves Lemma 2.4.17.

We think of a vector field on M as a smooth map

X : M → Rk

that satisfies the condition X(p) ∈ Tp M for every p ∈ M . Ignoring this

condition temporarily, we can differentiate X as a map from M to Rk and
its derivative at p is then a linear map

dX(p) : Tp M → Rk .

In general, this derivative will no longer take values in the tangent space
Tp M . However, if we have two vector fields X and Y on M , then the next
lemma shows that the difference of the derivative of X in the direction Y
and of Y in the direction X does take values in the tangent spaces of M .
2.4. VECTOR FIELDS AND FLOWS 47

Lemma 2.4.18. Let M ⊂ Rk be a smooth m-manifold and X, Y ∈ Vect(M )

be complete vector fields. Denote by

R → Diff(M ) : t 7→ φt , R → Diff(M ) : t 7→ ψ t

the flows of X and Y , respectively. Fix a point p ∈ M and consider the

smooth map γ : R → M defined by

γ(t) := φt ◦ ψ t ◦ φ−t ◦ ψ −t (p). (2.4.12)

Then γ̇(0) = 0 and

√
1

d 
γ t = γ̈(0)
dt t=0
 2

d 
= ((φs )∗ Y ) (p)
ds s=0 (2.4.13)

d 
∗

=  ψ t X (p)
dt t=0
= dX(p)Y (p) − dY (p)X(p) ∈ Tp M.

Exercise 2.4.19. Let

√ γ : R → Rk be a C 2 -curve

and assume γ̇(0) = 0.
d −2 1

Prove that dt t=0 γ( t) = limt→0 t γ(t) − γ(0) = 2 γ̈(0).
Proof of Lemma 2.4.18. Define the map β : R2 → M by

β(s, t) := φs ◦ ψ t ◦ φ−s ◦ ψ −t (p)

for s, t ∈ R. Then γ(t) = β(t, t) and

∂β
(0, t) = X(p) − dψ t (ψ −t (p))X(ψ −t (p)), (2.4.14)
∂s
∂β
(s, 0) = dφs (φ−s (p))Y (φ−s (p)) − Y (p). (2.4.15)
∂t
for all s, t ∈ R. Hence
∂β ∂β
γ̇(0) = (0, 0) + (0, 0) = 0.
∂s ∂t
This implies the first equality in (2.4.13) by Exercise (2.4.19). To prove
the remaining assertions, note that β(s, 0) = β(0, t) = p, hence the second
derivatives ∂ 2 β/∂s2 and ∂ 2 β/∂t2 vanish at s = t = 0, and therefore

∂2β
γ̈(0) = 2 (0, 0). (2.4.16)
∂s∂t
48 CHAPTER 2. FOUNDATIONS

Combining equations (2.4.15) and (2.4.16) we find

 
1 ∂  ∂β d 
γ̈(0) = (s, 0) = dφs (φ−s (p))Y (φ−s (p))
2 ∂s s=0 ∂t ds s=0

d 
= ((φs )∗ Y ) (p)).
ds s=0

Likewise, combining equations (2.4.14) and (2.4.16) we find

 
1 ∂  ∂β d 
γ̈(0) = (0, t) = −  dψ t (ψ −t (p))X(ψ −t (p))
2 ∂t t=0 ∂s dt t=0

d
=  dψ −t (ψ t (p))X(ψ t (p))

dt t=0

d 
=  dψ t (p)−1 X(ψ t (p))
dt t=0

d 
∗

=  ψ t X (p)).
dt t=0

In both cases the right hand side is the derivative of a smooth curve in the
tangent space Tp M and so is itself an element of Tp M . Moreover, we have

1 ∂ 
γ̈(0) = dφs (φ−s (p))Y (φ−s (p))
2 ∂s s=0
 
∂  ∂ 
= φs ◦ ψ t ◦ φ−s (p)
∂s s=0 ∂t t=0
 
∂  ∂ 
= φs ◦ ψ t ◦ φ−s (p)
∂t t=0 ∂s s=0

∂ 
X(ψ t (p)) − dψ t (p)X(p)


=
∂t t=0

 
∂  ∂ 
= dX(p)Y (p) −  ψ t ◦ φs (p)
∂t t=0 ∂s s=0
 
∂  ∂ 
= dX(p)Y (p) − ψ t ◦ φs (p)
∂s s=0 ∂t t=0

∂ 
= dX(p)Y (p) − Y (φs (p))
∂s s=0
= dX(p)Y (p) − dY (p)X(p).

This proves Lemma 2.4.18.

2.4. VECTOR FIELDS AND FLOWS 49

Definition 2.4.20 (Lie Bracket). Let M ⊂ Rk be a smooth manifold and

let X, Y ∈ Vect(M ) be smooth vector fields on M . The Lie bracket of X
and Y is the vector field [X, Y ] ∈ Vect(M ) defined by
[X, Y ](p) := dX(p)Y (p) − dY (p)X(p). (2.4.17)
Warning: In the literature on differential geometry the Lie bracket of two
vector fields is often (but not always) defined with the opposite sign. The
rationale behind the present choice of the sign will be explained in §2.5.7.
Lemma 2.4.21. Let M ⊂ Rk and N ⊂ R` be smooth manifolds, let X, Y, Z
be smooth vector fields on M , and let
φ:N →M
be a diffeomorphism. Then
φ∗ [X, Y ] = [φ∗ X, φ∗ Y ], (2.4.18)
[X, Y ] + [Y, X] = 0, (2.4.19)
[X, [Y, Z]] + [Y, [Z, X]] + [Z, [X, Y ]] = 0. (2.4.20)
The last equation is called the Jacobi identity
Proof. Let R → Diff(M ) : t 7→ ψ t be the flow of Y . Then the map
R → Diff(N ) : t 7→ φ−1 ◦ ψ t ◦ φ
is the flow of the vector field φ∗ Y on N . Hence, by Lemma 2.4.17 and
Lemma 2.4.18, we have

∗ ∗ d 
∗
[φ X, φ Y ] =  φ−1 ◦ ψ t ◦ φ φ∗ X
dt t=0

d 
∗
=  φ∗ ψ t X
dt t=0
= φ∗ [X, Y ].
This proves (2.4.18). Equation (2.4.19) is obvious. To prove (2.4.20), let φt
be the flow of X. Then by (2.4.18) and (2.4.19) and Lemma 2.4.18 we have

d 
[[Y, Z], X] =  (φt )∗ [Y, Z]
dt t=0

d 
=  [(φt )∗ Y, (φt )∗ Z]
dt t=0
= [[Y, X], Z] + [Y, [Z, X]]
= [Z, [X, Y ]] + [Y, [Z, X]].
This proves Lemma 2.4.21.
50 CHAPTER 2. FOUNDATIONS

Definition 2.4.22. A Lie algebra is a real vector space g equipped with

a skew-symmetric bilinear map g × g → g : (ξ, η) 7→ [ξ, η] that satisfies the
Jacobi identity [ξ, [η, ζ]] + [η, [ζ, ξ]] + [ζ, [ξ, η]] = 0 for all ξ, η, ζ ∈ g.

Example 2.4.23. The Vector fields on a smooth manifold M ⊂ Rk form a

Lie algebra with the Lie bracket (2.4.17). The space gl(n, R) = Rn×n of real
n × n-matrices is a Lie algebra with the Lie bracket

[ξ, η] := ξη − ηξ.

It is also interesting to consider subspaces of gl(n, R) that are invariant under

this Lie bracket. An example is the space
n o
o(n) := ξ ∈ gl(n, R)  ξ T + ξ = 0


of skew-symmetric n × n-matrices. It is a nontrivial fact that every finite-

dimensional Lie algebra is isomorphic to a Lie subalgebra of gl(n, R) for
some n. For example, the cross product defines a Lie algebra structure
on R3 and the resulting Lie algebra is isomorphic to o(3).

Remark 2.4.24. There is a linear map Rm×m → Vect(Rm ) : ξ 7→ Xξ which

assigns to a matrix ξ ∈ gl(m, R) the linear vector field Xξ : Rm → Rm
given by Xξ (x) := ξx for x ∈ Rm . This map preserves the Lie bracket,
i.e. [Xξ , Xη ] = X[ξ,η] , and hence is a Lie algebra homomorphism.

To understand the Lie bracket geometrically, consider again the curve

γ(t) := φt ◦ ψ t ◦ φ−t ◦ ψ −t (p)

in Lemma 2.4.18, where φt and ψ t are the flows of the vector fields X and Y ,
respectively. Since γ̇(0) = 0, Exercise 2.4.19 asserts that
√ √ √ √

1 d 
[X, Y ](p) = γ̈(0) =  φ t ◦ ψ t ◦ φ− t ◦ ψ − t (p). (2.4.21)
2 dt t=0

Geometrically this means that by following first the backward flow of Y

for time ε, then the backward flow of X for time ε, then the forward flow
of Y for time ε, and finally the forward flow of X for time ε, we will not, in
general, get back to the original point p where we started but approximately
obtain an “error” ε2 [X, Y ](p). An example of this (which we learned from
Donaldson) is the mathematical formulation of parking a car.
2.4. VECTOR FIELDS AND FLOWS 51

Example 2.4.25 (Parking a Car). The configuration space for driving a

car in the plane is the manifold M := C × S 1 , where S 1 ⊂ C denotes the
unit circle. Thus a point in M is a pair p = (z, λ) ∈ C × C with |λ| = 1. The
point z ∈ C represents the position of the car and the unit vector λ ∈ S 1
represents the direction in which it is pointing. The left turn is represented
by a vector field X and the right turn by a vector field Y on M . These
vector field are given by
X(z, λ) := (λ, iλ) , Y (z, λ) := (λ, −iλ) .
Their Lie bracket is the vector field
[X, Y ](z, λ) = (−2iλ, 0).
This vector field represents a sideways move of the car to the right. And a
sideways move by 2ε2 can be achieved by following a backward right turn
for time ε, then a backward left turn for time ε, then a forward right turn
for time ε, and finally a forward left turn for time ε.
This example can be reformulated by identifying C with R2 via z = x+iy
and representing a point in the unit circle by the angle θ ∈ R/2πZ via
λ = eiθ . In this formulation the manifold is M = R2 × R/2πZ, a point in M
is represented by a triple (x, y, θ) ∈ R3 , the vector fields X and Y are
X(x, y, θ) := (cos(θ), sin(θ), 1) , Y (x, y, θ) := (cos(θ), sin(θ), −1) ,
and their Lie bracket is [X, Y ](x, y, θ) = 2(sin(θ), − cos(θ), 0).
Lemma 2.4.26. Let X, Y ∈ Vect(M ) be complete vector fields on a man-
ifold M and φt , ψ t ∈ Diff(M ) be the flows of X and Y , respectively. Then
the Lie bracket [X, Y ] vanishes if and only if the flows of X and Y commute,
i.e. φs ◦ ψ t = ψ t ◦ φs for all s, t ∈ R.
Proof. If the flows of X and Y commute, then the Lie bracket [X, Y ] vanishes
by Lemma 2.4.18. Conversely, suppose that [X, Y ] = 0. Then we have
d s d
(φ )∗ Y = (φs )∗ (φr )∗ Y = (φs )∗ [X, Y ] = 0
ds dr
for every s ∈ R and hence
(φs )∗ Y = Y. (2.4.22)
Fix a real number s and define the curve γ : R → M by γ(t) := φs (ψ t (p))
for t ∈ R. Then γ(0) = φs (p) and
γ̇(t) = dφs (ψ t (p))Y (ψ t (p)) = ((φs )∗ Y ) (γ(t)) = Y (γ(t))
for all t. Here the last equation follows from (2.4.22). Since ψ t is the flow of
Y we obtain γ(t) = ψ t (φs (p)) for all t ∈ R and this proves Lemma 2.4.26.
52 CHAPTER 2. FOUNDATIONS

2.5 Lie Groups

Combining the concept of a group and a manifold, it is interesting to consider
groups which are also manifolds and have the property that the group op-
eration and the inverse define smooth maps. We shall only consider groups
of matrices.

2.5.1 Definition and Examples

Definition 2.5.1 (Lie Group). A nonempty subset G ⊂ Rn×n is called a
Lie group iff it is a submanifold of Rn×n and a subgroup of GL(n, R), i.e.
g, h ∈ G =⇒ gh ∈ G
(where gh denotes the product of the matrices g and h) and
g∈G =⇒ det(g) 6= 0 and g −1 ∈ G.
(Since G 6= ∅ it follows from these conditions that the identity matrix 1l is
an element of G.)
Example 2.5.2. The general linear group G = GL(n, R) is an open subset
of Rn×n and hence is a Lie group. By Exercise 2.1.18 the special linear group
SL(n, R) = {g ∈ GL(n, R) | det(g) = 1}
is a Lie group and, by Example 2.1.19, the special orthogonal group
n o
SO(n) := g ∈ GL(n, R) | g T g = 1l, det(g) = 1

is a Lie group. In fact every orthogonal matrix has determinant ±1 and

so SO(n) is an open subset of O(n) (in the relative topology).
In a similar vein the group GL(n, C) := {g ∈ Cn×n | det(g) 6= 0} of com-
plex matrices with nonzero (complex) determinant is an open subset of Cn×n
and hence is a Lie group. As in the real case, the subgroups
SL(n, C) := {g ∈ GL(n, C) | det(g) = 1} ,
U(n) := {g ∈ GL(n, C) | g ∗ g = 1l} ,
SU(n) := {g ∈ GL(n, C) | g ∗ g = 1l, det(g) = 1}
are submanifolds of GL(n, C) and hence are Lie groups. Here g ∗ := ḡ T
denotes the conjugate transpose of a complex matrix.
Exercise 2.5.3. Prove that SL(n, C), U(n), and SU(n) are Lie groups.
Prove that SO(n) is connected and that O(n) has two connected components.
2.5. LIE GROUPS 53

Exercise 2.5.4. Prove that GL(n, C) can be identified with the group
 
0 −1l
G := {Φ ∈ GL(2n, R) | ΦJ0 = J0 Φ} , J0 := .
1l 0
Hint: Use the isomorphism Rn × Rn → Cn : (x, y) 7→ x + iy. Show that a
matrix Φ ∈ R2n×2n commutes with J0 if and only if it has the form
 
X −Y
Φ= , X, Y ∈ Rn×n .
Y X
What is the relation between the real determinant of Φ and the complex
determinant of X + iY ?
Exercise 2.5.5. Let J0 be as in Exercise 2.5.4 and define
n o
Sp(2n) := Ψ ∈ GL(2n, R) | ΨT J0 Ψ = J0 .

This is the symplectic linear group. Prove that Sp(2n) is a Lie group.
Hint: See [36, Lemma 1.1.12].
Example 2.5.6 (Unit Quaternions). The Quaternions form a four-
dimensional associative unital algebra H, equipped with a basis 1, i, j, k.
The elements of H are vectors of the form
x = x0 + ix1 + jx2 + kx3 x0 , x1 , x2 , x3 ∈ R. (2.5.1)
The product structure is the bilinear map H × H → H : (x, y) 7→ xy, deter-
mined by the relations
i2 = j2 = k2 = −1, ij = −ji = k, jk = −kj = i, ki = −ik = j.
This product structure is associative but not commutative. The quaternions
are equipped with an involution H → H : x 7→ x̄, which assigns to a quater-
nion x of the form (2.5.1) its conjugate x̄ := x0 − ix1 − jx2 − kx3 . This
involution satisfies the conditions
x + y = x̄ + ȳ, xy = ȳx̄, xx̄ = |x|2 , |xy| = |x| |y|
p
for x, y ∈ H, where |x| := x20 + x22 + x22 + x23 denotes the Euclidean norm
of the quaternion (2.5.1). Thus the unit quaternions form a group
Sp(1) := {x ∈ H | |x| = 1}
with the inverse map x 7→ x̄. Note that the group Sp(1) is diffeomorphic
to the 3-sphere S 3 ⊂ R4 under the isomorphism H ∼ = R4 . Warning: The
unit quaternions (a compact Lie group) are not to be confused with the
symplectic linear group in Exercise 2.5.5 (a noncompact Lie group) despite
the similarity in notation.
54 CHAPTER 2. FOUNDATIONS

Let G ⊂ GL(n, R) be a Lie group. Then the maps

G × G → G : (g, h) 7→ gh, G → G : g 7→ g −1

are smooth (see [47]). Fixing an element h ∈ G we find that the derivative
of the map G → G : g 7→ gh at g ∈ G is given by the linear map

Tg G → Tgh G : gb 7→ gbh. (2.5.2)

Here gb and h are both matrices in Rn×n and gbh denotes the matrix prod-
uct. In fact, if gb ∈ Tg G, then, since G is a manifold, there exists a smooth
curve γ : R → G with γ(0) = g and γ̇(0) = gb. Since G is a group we obtain
a smooth curve β : R → G given by β(t) := γ(t)h. It satisfies β(0) = gh and
so gbh = β̇(0) ∈ Tgh G.
The linear map (2.5.2) is obviously a vector space isomorphism whose
inverse is given by right multiplication with h−1 . It is sometimes convenient
to define the map Rh : G → G by

Rh (g) := gh

for g ∈ G (right multiplication by h). This is a diffeomorphism and the linear

map (2.5.2) is the derivative of Rh at g, so

dRh (g)b
g = gbh for gb ∈ Tg G.

Similarly, each element g ∈ G determines a diffeomorphism Lg : G → G,

given by
Lg (h) := gh
for h ∈ G (left multiplication by g). Its derivative at h ∈ G is again given by
matrix multiplication, i.e. the linear map dLg (h) : Th G → Tgh G is given by

dLg (h)b
h = gb
h h ∈ Th G.
for b (2.5.3)

Since Lg is a diffeomorphism its derivative dLg (h) : Th G → Tgh G is again a

vector space isomorphism for every h ∈ G.

Exercise 2.5.7. Prove that the map G → G : g 7→ g −1 is a diffeomorphism

and that its derivative at g ∈ G is the vector space isomorphism

Tg G → Tg−1 G : v 7→ −g −1 vg −1 .

Hint: Use [47] or any textbook on first year analysis.

2.5. LIE GROUPS 55

2.5.2 The Lie Algebra of a Lie Group

Let
G ⊂ GL(n, R)
be a Lie group. Its tangent space at the identity matrix 1l ∈ G is called the
Lie algebra of G and will be denoted by

g = Lie(G) := T1l G.

This terminology is justified by the fact that g is in fact a Lie algebra, i.e.
it is invariant under the standard Lie bracket operation

[ξ, η] := ξη − ηξ

on the space Rn×n of square matrices (see Lemma 2.5.9 below). The proof
requires the notion of the exponential matrix. For ξ ∈ Rn×n and t ∈ R
we define
∞ k k
X t ξ
exp(tξ) := . (2.5.4)
k!
k=0

A standard result in first year analysis asserts that this series converges
absolutely (and uniformly on compact t-intervals), that the map

R → Rn×n : t 7→ exp(tξ)

is smooth and satisfies the differential equation

d
exp(tξ) = ξ exp(tξ) = exp(tξ)ξ, (2.5.5)
dt
and that
exp((s + t)ξ) = exp(sξ) exp(tξ), exp(0ξ) = 1l (2.5.6)
for all s, t ∈ R. This shows that the matrix exp(tξ) is invertible for each t
and that the map R → GL(n, R) : t 7→ exp(tξ) is a group homomorphism.

Exercise 2.5.8. Prove the following analogue of (2.4.12). For ξ, η ∈ g

√ √ √ √

d 
exp( tξ) exp( tη) exp(− tξ) exp(− tη) = [ξ, η] (2.5.7)
dt t=0


In other words, the infinitesimal Lie group commutator is the matrix com-
mutator. (Compare Equations (2.5.7) and (2.4.21).)
56 CHAPTER 2. FOUNDATIONS

Lemma 2.5.9. Let G ⊂ GL(n, R) be a Lie group and denote by g := Lie(G)

its Lie algebra. Then the following holds.
(i) If ξ ∈ g, then exp(tξ) ∈ G for every t ∈ R.
(ii) If g ∈ G and η ∈ g, then gηg −1 ∈ g.
(iii) If ξ, η ∈ g, then [ξ, η] = ξη − ηξ ∈ g.
Proof. We prove (i). For every g ∈ G we have a vector space isomor-
phism g = T1l G → Tg G : ξ 7→ ξg as in (2.5.2). Hence each element ξ ∈ g
determines a vector field Xξ ∈ Vect(G), defined by

Xξ (g) := ξg ∈ Tg G, g ∈ G. (2.5.8)

By Theorem 2.4.7 there is an integral curve γ : (−ε, ε) → G satisfying

γ̇(t) = Xξ (γ(t)) = ξγ(t), γ(0) = 1l.

By (2.5.5), the curve (−ε, ε) → Rn×n : t 7→ exp(tξ) satisfies the same initial
value problem and hence, by uniqueness, we have exp(tξ) = γ(t) ∈ G for
all t ∈ R with |t| < ε. Now let t ∈ R and choose N ∈ N such that  Nt  < ε.
Then exp( Nt ξ) ∈ G and hence it follows from (2.5.6) that
 N
t
exp(tξ) = exp ξ ∈ G.
N
This proves (i).
We prove (ii). Consider the smooth curve γ : R → Rn×n defined by

γ(t) := g exp(tη)g −1 .

By (i) we have γ(t) ∈ G for every t ∈ R. Since γ(0) = 1l we have

gηg −1 = γ̇(0) ∈ g.

This proves (ii).

We prove (iii). Define the smooth map η : R → Rn×n by

η(t) := exp(tξ)η exp(−tξ).

By (i) we have exp(tξ) ∈ G and, by (ii), we have η(t) ∈ g for every t ∈ R.

Hence [ξ, η] = η̇(0) ∈ g. This proves (iii) and Lemma 2.5.9.

By Lemma 2.5.9 the curve γ : R → G defined by γ(t) := exp(tξ)g is the

integral curve of the vector field Xξ in (2.5.8) with initial condition γ(0) = g.
Thus Xξ is complete for every ξ ∈ g.
2.5. LIE GROUPS 57

Lemma 2.5.10. If ξ ∈ g and γ : R → G is a smooth curve satisfying

γ(s + t) = γ(s)γ(t), γ(0) = 1l, γ̇(0) = ξ, (2.5.9)

then γ(t) = exp(tξ) for every t ∈ R.

Proof. For every t ∈ R we have
 
d  d 
γ̇(t) = γ(s + t) = γ(s)γ(t) = γ̇(0)γ(t) = ξγ(t).
ds s=0 ds s=0

Hence γ is the integral curve of the vector field Xξ in (2.5.8) with γ(0) = 1l.
This implies γ(t) = exp(tξ) for every t ∈ R, as claimed.

Example 2.5.11. Since the general linear group GL(n, R) is an open subset
of Rn×n its Lie algebra is the space of all real n × n-matrices

gl(n, R) := Lie(GL(n, R)) = Rn×n .

The Lie algebra of the special linear group is

 
sl(n, R) := Lie(SL(n, R)) = ξ ∈ gl(n, R)  trace(ξ) = 0

(see Exercise 2.2.9) and the Lie algebra of the special orthogonal group is
n o
so(n) := Lie(SO(n)) = ξ ∈ gl(n, R)  ξ T + ξ = 0 = o(n)


(see Example 2.2.10).

Exercise 2.5.12. Prove that the Lie algebras of the general linear group
over C, the special linear group over C, the unitary group, and the special
unitary group are given by

gl(n, C) := Lie(GL(n, C)) = Cn×n ,

 
sl(n, C) := Lie(SL(n, C)) = ξ ∈ gl(n, C)  trace(ξ) = 0 ,
u(n) := Lie(U(n)) = ξ ∈ gl(n, R)  ξ ∗ + ξ = 0 ,
 

su(n) := Lie(SU(n)) = ξ ∈ gl(n, C)  ξ ∗ + ξ = 0, trace(ξ) = 0 .

 

These are vector spaces over the reals. Determine their real dimensions.
Which of these are also complex vector spaces?
Exercise 2.5.13. Let G ⊂ GL(n, R) be a subgroup. Prove that G is a Lie
group if and only if it is a closed subset of GL(n, R) in the relative topology.
58 CHAPTER 2. FOUNDATIONS

Exercise 2.5.14. Let V be a finite-dimensional vector space. Prove that

the vector space g := V × End(V ) is a Lie algebra with the Lie bracket

[(u, A), (v, B)] := (Av − Bu, AB − BA) (2.5.10)

for u, v ∈ V and A, B ∈ End(V ). Find the corresponding Lie group. Find

an embedding of g into End(R × V ) as a Lie subalgebra.

Exercise 2.5.15. Let (V, ω) be a 2n-dimensional symplectic vector space,

so ω : V × V → R is a nondegenerate skew-symmetric bilinear form. The
Heisenberg algebra of (V, ω) is the Lie algebra h := V × R with the Lie
bracket of two elements (v, t), (v 0 , t0 ) ∈ V × R defined by

(v, t), (v 0 , t0 ) := 0, ω(v, v 0 ) .

 

(2.5.11)

Find a corresponding Lie group structure on H = V × R. Embed H as a

Lie subgroup into GL(n + 2, R) and find a formula

for the exponential map.
Hint: Take V = R × R and ω (x, y), (x , y ) = hx, y i − hy, x0 i.
n n 0 0 0

2.5.3 Lie Group Homomorphisms

Let G, H be Lie groups and g, h be Lie algebras. A Lie group homomor-
phism from G to H is a smooth map ρ : G → H that is a group homomor-
phism. A Lie group isomorphism is a bijective Lie group homomorphism
whose inverse is also a Lie group homomorphism. A Lie group automor-
phism is a Lie group isomorphism from a Lie group to itself. A Lie algebra
homomorphism from g to h is a linear map Φ : g → h that preserves the
Lie bracket. A Lie algebra isomorphism is a bijective Lie algebra homo-
morphism. A Lie algebra automorphism is a Lie algebra isomorphism
from a Lie algebra to itself.

Lemma 2.5.16. Let G and H be Lie groups and denote their Lie algebras
by g := Lie(G) and h := Lie(H). Let ρ : G → H be a Lie group homomor-
phism and denote its derivative at 1l ∈ G by

ρ̇ := dρ(1l) : g → h.

Then ρ̇ is a Lie algebra homomorphism. Moreover,

ρ(exp(ξ)) = exp(ρ̇(ξ)), ρ(gξg −1 ) = ρ(g)ρ̇(ξ)ρ(g)−1

for all ξ ∈ g and all g ∈ G.

2.5. LIE GROUPS 59

Proof. The proof has three steps.

Step 1. For all ξ ∈ g and t ∈ R we have ρ(exp(tξ)) = exp(tρ̇(ξ)).
Fix an element ξ ∈ g. Then exp(tξ) ∈ G for every t ∈ R by Lemma 2.5.9.
Thus we can define a curve γ : R → H by γ(t) := ρ(exp(tξ)). Since ρ is
smooth, this is a smooth curve in H and, since ρ is a group homomorphism
and the exponential map satisfies (2.5.6), our curve γ satisfies the conditions
γ(s + t) = γ(s)γ(t), γ(0) = 1l, γ̇(0) = dρ(1l)ξ = ρ̇(ξ).
Hence γ(t) = exp(tρ̇(ξ)) by Lemma 2.5.10. This proves Step 1.
Step 2. For all g ∈ G and η ∈ g we have ρ̇(gηg −1 ) = ρ(g)ρ̇(η)ρ(g)−1 .
Define the smooth curve γ : R → G by γ(t) := g exp(tη)g −1 . It takes values
in G by Lemma 2.5.9. By Step 1 we have
ρ(γ(t)) = ρ(g)ρ(exp(tη))ρ(g)−1 = ρ(g) exp(tρ̇(η))ρ(g)−1
for every t. Since γ(0) = 1l and γ̇(0) = gηg −1 we obtain
ρ̇(gηg −1 ) = dρ(γ(0))γ̇(0)

d 
=  ρ(γ(t))
dt t=0

d
=  ρ(g) exp(tρ̇(η))ρ(g)−1
dt t=0
= ρ(g)ρ̇(η)ρ(g)−1 .
This proves Step 2.
Step 3. For all ξ, η ∈ g we have ρ̇([ξ, η]) = [ρ̇(ξ), ρ̇(η)].
Define the curve η : R → g by η(t) := exp(tξ)η exp(−tξ) for t ∈ R. It takes
values in the Lie algebra of G by Lemma 2.5.9 and η̇(0) = [ξ, η]. Hence

d 
ρ̇([ξ, η]) =  ρ̇ (exp(tξ)η exp(−tξ))
dt t=0

d 
=  ρ (exp(tξ)) ρ̇(η)ρ (exp(−tξ))
dt t=0

d 
=  exp (tρ̇(ξ)) ρ̇(η) exp (−tρ̇(ξ))
dt t=0
= [ρ̇(ξ), ρ̇(η)] .
Here the first equality follows from the fact that ρ̇ is linear, the second
equality follows from Step 2 with g = exp(tξ), and the third equality follows
from Step 1. This proves Step 3 and Lemma 2.5.16.
60 CHAPTER 2. FOUNDATIONS

Exercise 2.5.17. A Lie group homomorphism ρ : G → H is uniquely deter-

mined by the Lie algebra homomorphism ρ̇ whenever G is connected. Hint:
If ρ1 , ρ2 : G → H are Lie group homomorphisms such that ρ̇1 = ρ̇2 , prove
that the set A := {g ∈ G | ρ1 (g) = ρ2 (g)} is both open and closed.
Exercise 2.5.18. If ρ : G → H is a bijective Lie group homomorphism,
then ρ−1 : H → G is smooth and hence ρ is a Lie group isomorphism. Hint:
Use Lemma 2.5.16 to prove that ρ̇ : g → h is injective. If ρ̇ is not surjective,
show that ρ has no regular value in contradiction to Sard’s theorem.
Example 2.5.19. The complex determinant defines a Lie group homomor-
phism det : U(n) → S 1 . The associated Lie algebra homomorphism is
˙ : u(n) → iR = Lie(S 1 ).
trace = det
Example 2.5.20 (Unit Quaternions and SU(2)). The Lie group SU(2)
is diffeomorphic to the 3-sphere. Every matrix in SU(2) can be written as
 
x0 + ix1 x2 + ix3
g= , x20 + x21 + x22 + x23 = 1. (2.5.12)
−x2 + ix3 x0 − ix1
Here the xi are real numbers. They can be interpreted as the coordinates
of a unit quaternion x = x0 + ix1 + jx2 + kx3 ∈ Sp(1) (see Example 2.5.6).
The reader may verify that the map Sp(1) → SU(2) : x 7→ g in (2.5.12) is a
Lie group isomorphism.
Exercise 2.5.21 (The double cover of SO(3)). Identify the imaginary
part of H with R3 and write a vector ξ ∈ R3 = Im(H) as a purely imaginary
quaternion ξ = iξ1 + jξ1 + kξ3 . Prove that if ξ ∈ Im(H) and x ∈ Sp(1),
then xξ x̄ ∈ Im(H). Define the map ρ : Sp(1) → SO(3) by
ρ(x)ξ := xξ x̄
for x ∈ Sp(1) and ξ ∈ Im(H). Prove that the linear map ρ(x) : R3 → R3 is
represented by the 3 × 3-matrix
 2
x0 + x21 − x22 − x23

2(x1 x2 − x0 x3 ) 2(x1 x3 + x0 x2 )
ρ(x) =  2(x1 x2 + x0 x3 ) x20 + x22 − x23 − x21 2(x2 x3 − x0 x1 )  .
2(x1 x3 − x0 x2 ) 2(x2 x3 + x0 x1 ) x0 + x23 − x21 − x22
2

Show that ρ is a Lie group homomorphism. Find a formula for the map
ρ̇ := dρ(1l) : sp(1) → so(3)
and show that it is a Lie algebra isomorphism. For x, y ∈ Sp(1) prove
that ρ(x) = ρ(y) if and only if y = ±x.
2.5. LIE GROUPS 61

Example 2.5.22. Let g be a finite-dimensional Lie algebra. Then the set

 
 Φ is a bijective linear map and 


Aut(g) := Φ : g → g  Φ[ξ, η] = [Φξ, Φη] (2.5.13)

for all ξ, η ∈ g
  

of Lie algebra automorphisms of g is a Lie group. Its Lie algebra is the

space of derivations on g denoted by
 
 δ is a linear map and

 
Der(g) := δ : g → g  δ [ξ, η] = [δ ξ, η] + [ξ, δ η] . (2.5.14)

for all ξ, η ∈ g
  

Now suppose that g = Lie(G) is the Lie algebra of a Lie group G. Then
there is a map Ad : G → Aut(g) defined by

Ad(g)η := gηg −1 (2.5.15)

for g ∈ G and η ∈ g. Part (ii) of Lemma 2.5.9 asserts that Ad(g) maps g
to itself for every g ∈ G. It follows directly from the definitions that the
map Ad(g) : g → g is a Lie algebra automorphism for every g ∈ G and that
the map Ad : G → Aut(g) is a Lie group homomorphism. The associated
Lie algebra homomorphism is the linear map ad : g → Der(g) defined by

ad(ξ)η := [ξ, η] (2.5.16)

˙ we compute
for ξ, η ∈ g. To verify the equation ad = Ad
 
˙ d  d 
Ad(ξ)η =  Ad(exp(tξ))η =  exp(tξ)η exp(−tξ) = [ξ, η].
dt t=0 dt t=0

Exercise 2.5.23. Let g be any Lie algebra. Define the map ad : g → End(g)
by (2.5.16) and prove that the endomorphism ad(ξ) : g → g is a derivation
for every ξ ∈ g. Prove that ad : g → Der(g) is a Lie algebra homomorphism.

Exercise 2.5.24. Let g be any finite-dimensional Lie algebra. Prove that

the group Aut(g) in (2.5.13) is a Lie subgroup of GL(g) with the Lie al-
gebra Lie(Aut(g)) = Der(g). Hint: Show that a linear map δ : g → g is
a derivation if and only if the linear map exp(tδ) : g → g is a Lie algebra
automorphism for every t ∈ R. Use the Closed Subgroup Theorem 2.5.26.
62 CHAPTER 2. FOUNDATIONS

2.5.4 Closed Subgroups

This section deals with subgroups of a Lie group G that are also submanifolds
of G. Such subgroups are called Lie subgroups. We assume throughout
that G ⊂ GL(n, R) is a Lie group with the Lie algebra g := Lie(G) = T1l G.

Definition 2.5.25 (Lie Subgroup). A subset H ⊂ G is called a Lie sub-

group of G iff it is both a subgroup and a smooth submanifold of G.

A useful general criterion is the Closed Subgroup Theorem which asserts

that a subgroup H ⊂ G is a Lie subgroup if and only if it is a closed subset
of G. This was first proved in 1929 by John von Neumann [40] for the special
case G = GL(n, R) and then in 1930 by Élie Cartan [9] in full generality.

Theorem 2.5.26 (Closed Subgroup Theorem). Let H be a subgroup

of G. Then the following are equivalent.
(i) H is a smooth submanifold (and hence a Lie subgroup) of G.
(ii) H is a closed subset of G.
It (i) holds, then the Lie algebra of H is the space
 
h := η ∈ g  exp(tη) ∈ H for all t ∈ R . (2.5.17)

Proof of Theorem 2.5.26 (i) =⇒ (ii) and (2.5.17). Assume that H is a Lie
subgroup of G and let h ⊂ g be defined by (2.5.17). We prove that h is
the Lie algebra of H. Assume first that η ∈ h. Then the curve γ : R → G
defined by γ(t) := exp(tη) for t ∈ R takes values in H and satisfies γ(0) = 1l
and γ̇(0) = η, and this implies η ∈ T1l H = Lie(H). Conversely, if η ∈ Lie(H),
then Lemma 2.5.9 asserts that exp(tη) ∈ H for all t ∈ R and hence η ∈ h.
This shows that h = Lie(H).
Next we prove in three steps that H is a closed subset of G. Choose any
inner product on g, denote by |·| the associated norm, and denote by h⊥ ⊂ g
the orthogonal complement of h with respect to this inner product.
Step 1. There exist open neighborhoods V ⊂ H of 1l and W ⊂ h⊥ of the
origin such that the map φ : V × W → G, defined by

φ(h, ξ) := h exp(ξ)

for h ∈ V and ξ ∈ W , is a diffeomorphism from V × W onto an open neigh-

borhood U = φ(V × W ) ⊂ G of 1l.
The derivative of the map H × h⊥ → G : (h, ξ) 7→ h exp(ξ) at the point (1l, 0)
is bijective. Hence Step 1 follows from the Inverse Function Theorem.
2.5. LIE GROUPS 63

Step 2. There exists a δ > 0 such that, if ξ, ξ 0 ∈ h⊥ satisfy |ξ|, |ξ 0 | < δ

and exp(ξ 0 ) exp(−ξ) ∈ H, then ξ = ξ 0 .
Let V, W, φ be as in Step 1, choose an open neigborhood V 0 ⊂ G of 1l such
that V 0 ∩ H = V , and choose a constant δ > 0 such that the following holds.
(a) If ξ ∈ h⊥ satisfies |ξ| < δ, then ξ ∈ W .
(b) If ξ, ξ 0 ∈ g satisfy |ξ|, |ξ 0 | < δ, then exp(ξ 0 ) exp(−ξ) ∈ V 0 .
Let ξ, ξ 0 ∈ h⊥ such that |ξ|, |ξ 0 | < δ and h := exp(ξ 0 ) exp(−ξ) ∈ H. Then we
have ξ, ξ 0 ∈ W by (a) and h ∈ V 0 ∩ H = V by (b). Also φ(h, ξ) = φ(1l, ξ 0 )
and so ξ = ξ 0 , because φ is injective on V × W . This proves Step 2.
Step 3. Let hi be a sequence in H that converges to an element g ∈ G.
Then g ∈ H.
Let φ : V × W → U be as in Step 1 and let δ > 0 be as in Step 2. Since the
sequence h−1 −1
i g converges to 1l, there exists an i0 ∈ N such that hi g ∈ U for
all i ≥ i0 . Hence, for each i ≥ i0 , there exists a unique pair (h0i , ξi ) ∈ V × W
such that h−1 0
i g = hi exp(ξi ). This sequence satisfies limi→∞ ξi = 0. Hence
there exists an integer i1 ≥ i0 such that |ξi | < δ for all i ≥ i1 . Since
hi h0i exp(ξi ) = g = hj h0j exp(ξj ),
we also have exp(ξi ) exp(−ξj ) = (hi h0i )−1 hj h0j ∈ H for all i, j ≥ i1 . By Step 3,
this implies ξi = ξj for all i, j ≥ i1 . Hence ξi = limj→∞ ξj = 0 for all i ≥ i1
and so g = hi h0i ∈ H. This proves Step 3.
By Step 3 the Lie subgroup H is a closed subset of G. Thus we have
proved that (i) implies (ii) and (2.5.17) in Theorem 2.5.26.
The proof of the converse implication requires three preparatory lemmas.
Lemma 2.5.27. Let ξ ∈ g and let γ : R → G be a curve that is differentiable
at t = 0 and satisfies γ(0) = 1l and γ̇(0) = ξ. Then
exp(tξ) = lim γ(t/k)k (2.5.18)
k→∞
for every t ∈ R.
Proof. Fix a nonzero real number t and define ξk := k γ(t/k) − 1l ∈ Rn×n

for k ∈ N. Then
γ(t/k) − γ(0)
lim ξk = t lim = tγ̇(0) = tξ
k→∞ k→∞ t/k
and hence
ξk k
 
exp(tξ) = lim 1l + = lim γ(t/k)k .
k→∞ k k→∞
(See [47, Satz 1.5.2].) This proves Lemma 2.5.27.
64 CHAPTER 2. FOUNDATIONS

Lemma 2.5.28. Let H ⊂ G be a closed subgroup. Then the set h in (2.5.17)

is a Lie subalgebra of g
Proof. Let ξ, η ∈ h and define the curve γ : R → H by

γ(t) := exp(tξ) exp(tη)

for t ∈ R. This curve is smooth and satisfies γ(0) = 1l and γ̇(0) = ξ + η.

Since H is closed, it follows from Lemma 2.5.27 that

exp(t(ξ + η)) = lim γ(t/k)k ∈ H

k→∞

for all t ∈ R and so ξ + η ∈ h by definition. Thus h is a linear subspace of g.

Now fix an element ξ ∈ h. If h ∈ H, then

exp(sh−1 ξh) = h−1 exp(sξ)h ∈ H

for all s ∈ R and hence h−1 ξh ∈ h by definition. Take h = exp(tη) with η ∈ h

to obtain exp(−tη)ξ exp(tη) ∈ h for all t ∈ R. Differentiating this curve
at t = 0 gives [ξ, η] ∈ h and this proves Lemma 2.5.28.

Lemma 2.5.29. Let H ⊂ G be a closed subgroup and let h ⊂ g be the Lie

subalgebra in (2.5.17). Let ξ ∈ g, let (ξi )i∈N be a sequence in g, and let (τi )i∈N
be a sequence of positive real numbers such that

exp(ξi ) ∈ H, ξi 6= 0

for all i ∈ N and

ξi
lim τi = 0, lim ξi = 0, lim = ξ.
i→∞ i→∞ i→∞ τi

Then ξ ∈ h.
Proof. Fix a real number t. Then, for each i ∈ N, there exists a unique
integer mi ∈ Z such that mi τi ≤ t < (mi + 1)τi . The sequence mi satisfies
ξi
lim mi τi = t, lim mi ξi = lim mi τi = tξ
i→∞ i→∞ i→∞ τi
and hence

exp(tξ) = lim exp(mi ξi ) = lim exp(ξi )mi ∈ H.

i→∞ i→∞

Thus exp(tξ) ∈ H for every t ∈ R and so ξ ∈ h by (2.5.17). This proves

Lemma 2.5.29.
2.5. LIE GROUPS 65

Proof of Theorem 2.5.26 (ii) =⇒ (i). Choose any inner product on g and
define a Riemannian metric on G by |v| := |g −1 v| for g ∈ G and v ∈ Tg G.
Let H ⊂ G be a closed subgroup of G and define the set h ⊂ g by (2.5.17).
Then h is a Lie subalgebra of g by Lemma 2.5.28. Define
k := dim(h), ` := dim(g) ≥ k,
and choose a basis η1 , . . . , η` of g such that the vectors η1 . . . . , ηk form a basis
of h and ην ∈ h⊥ for ν > k. Let h0 ∈ H and define the map Θ : R` → G by
Θ(t1 , . . . , t` ) := h0 exp(t1 η1 + · · · + tk ηk ) exp(tk+1 ηk+1 + · · · + t` η` ).
Then Θ(0) = h0 , Θ(Rk × {0}) ⊂ H, and the derivative dΘ(0) : R` → Th0 G
is bijective. Hence the inverse function theorem asserts that Θ restricts to a
diffeomorphism from an open neighborhood Ω ⊂ R` of the origin to an open
neighborhood U := Θ(Ω) ⊂ G of h0 that satisfies
Θ Ω ∩ (Rk × {0}) ⊂ U ∩ H.


Θ(0) = h0 ,
We prove the following.
Claim. There exists an open set Ω0 ⊂ R` such that
Θ Ω0 ∩ (Rk × {0}) = U0 ∩ H,


0 ∈ Ω0 ⊂ Ω, U0 := Θ(Ω0 ). (2.5.19)
Assume, by contradiction, that such an open set Ω0 does not exist. Then
there exists a sequence ti = (t1i , . . . , t`i ) ∈ R` such that
lim ti = 0, ti ∈ Ω \ (Rk × {0}), Θ(ti ) ∈ H.
i→∞

Define
k `
!
X X
hi := h0 exp tνi ην ∈ H, ξi := tνi ην ∈ h⊥ \ {0}.
ν=1 ν=k+1

Then hi exp(ξi ) = Θ(ti ) ∈ H and hence

lim ξi = 0, ξi 6= 0, exp(ξi ) = h−1
i Θ(ti ) ∈ H.
i→∞

Passing to a subsequence, if necessary, we may assume that the sequence

ξi /|ξi | converges. Denote its limit by ξ := limi→∞ ξi /|ξi |. Then ξ ∈ h by
Lemma 2.5.29 and ξ ∈ h⊥ by definition. Since |ξ| = 1, this is a contradiction.
This contradiction proves the Claim. Thus there does, after all, exist an
open set Ω0 ⊂ R` that satisfies (2.5.19), and the map Θ−1 : U0 → Ω0 is then
a coordinate chart on G which satisfies Θ−1 (U0 ∩ H) = Ω0 ∩ (Rk × {0}).
Hence H is a submanifold of G and this proves Theorem 2.5.26.
66 CHAPTER 2. FOUNDATIONS

Remark 2.5.30. Theorem 2.5.26 was considerably strengthened by Ana-

tolij Ivanovich Malcev [35] in 1945. He proved that a subgroup H ⊂ G is
closed if and only if every element ξ ∈ g = Lie(G) satisfies
   
exp(tξ)  t ∈ R ⊂ H =⇒ exp(tξ)  t ∈ R ⊂ H. (2.5.20)
For an exposition see Hilgert–Neeb [21, Corollary 13.4.6].
Example 2.5.31. Choose a nonzero vector (ω1 , . . . , ωn ) ∈ Rn such that at
least one of the ratios ωi /ωj is irrational. Then the subgroup
Sω := {(e2πitω1 , e2πitω2 , . . . , e2πitωn ) | t ∈ R} ⊂ (S 1 )n ∼
= Tn
of the torus is not closed. Similar examples exist in any Lie group that
contains a torus of dimension at least two.
Example 2.5.32 (Center). The center of a group G is the subgroup
 
Z(G) := g ∈ G  gh = hg for all h ∈ G . (2.5.21)
If G ⊂ GL(n, R) is a Lie group, then it follows directly from the definition
that the center is a closed subgroup of G. Hence by Theorem 2.5.26 it is a
Lie subgroup of G. If G is connected, then the Lie algebra of the center Z(G)
is the center of the Lie algebra g = Lie(G), defined by
 
Z(g) := ξ ∈ g  [ξ, η] = 0 for all η ∈ g . (2.5.22)
Exercise: Prove this. Hint: Let G ⊂ GL(n, R) be a connected Lie group
with Lie algebra g. Prove that for every ξ ∈ g the following are equivalent.
(a) exp(tξ) ∈ Z(G) for all t ∈ R.
(b) gξg −1 = ξ for all g ∈ G.
(c) [ξ, η] = 0 for all η ∈ g.
The assumption of the group G being connected is needed in the proof
that (c) implies (b). Examine the formula (2.5.17) in the case H = Z(G).
Example 2.5.33. Let G ⊂ GL(n, R) be a compact Lie group with the Lie
 
algebra g := Lie(G) and let ξ ∈ g. Then the set Tξ := exp(tξ)  t ∈ R is
a closed subgroup of G. Hence, by Theorem 2.5.26, it is a Lie subgroup
of G (called the torus generated by ξ). Exercise: Prove that Tξ is a
connected abelian subgroup of G.
Exercise 2.5.34. Let G ⊂ GL(n, R) be a Lie group with the Lie algebra g
and let ξ : R → g be a smooth function. Prove that the differential
equation γ̇(t) = ξ(t)γ(t), γ(0) = 1l, has a unique solution γ : R → G.
(Hint: Prove the existence of a solution γ : R → GL(n, R) and show that
the set {t ∈ R | γ(t) ∈ G} is open and closed.)
2.5. LIE GROUPS 67

2.5.5 Lie Groups and Diffeomorphisms

There is a natural correspondence between Lie groups and Lie algebras on
the one hand and diffeomorphisms and vector fields on the other hand. We
summarize this correspondence in the following table
Lie groups Diffeomorphisms
G ⊂ GL(n, R) Diff(M )
g = Lie(G) = T1l G Vect(M ) = Tid Diff(M )
exponential map flow of a vector field
t 7→ exp(tξ) t 7→ φt = “ exp(tX)00
adjoint representation pushforward
ξ 7→ gξg −1 X 7→ φ∗ X
Lie bracket on g Lie bracket of vector fields
[ξ, η] = ξη − ηξ [X, Y ] = dX · Y − dY · X.
To understand the correspondence between the exponential map and the
flow of a vector field compare equation (2.4.6) with equation (2.5.5). To un-
derstand the correspondence between the adjoint representation and push-
forward observe that
 
d  −1 d 
φ∗ Y =  φ ◦ ψ ◦ φ ,t
gηg =  g exp(tη)g −1 ,
−1
dt t=0 dt t=0
where ψtdenotes the flow of Y . To understand the correspondence between
the Lie brackets recall that
 
d  t d 
[X, Y ] =  (φ )∗ Y, [ξ, η] =  exp(tξ)η exp(−tξ),
dt t=0 dt t=0
where φt denotes the flow of X. We emphasize that the analogy between
Lie groups and Diffeomorphisms only works well when the manifold M is
compact so that every vector field on M is complete. The next exercise gives
another parallel between the Lie bracket on the Lie algebra of a Lie group
and the Lie bracket of two vector fields.
Exercise 2.5.35. Let G ⊂ GL(n, R) be a Lie group with Lie algebra g and
let ξ, η ∈ g. Define the smooth curve γ : R → G by
γ(t) := exp(tξ) exp(tη) exp(−tξ) exp(−tη).
Show that γ̇(0) = 0 and 21 γ̈(0) = [ξ, η] (cf. Exercise 2.5.8 and Lemma 2.4.18).
Exercise 2.5.36. Let G ⊂ GL(n, R) be a Lie group with Lie algebra g and
let ξ, η ∈ g. Show that [ξ, η] = 0 if and only if the exponential maps com-
mute, i.e. exp(sξ) exp(tη) = exp(tη) exp(sξ) = exp(sξ + tη) for all s, t ∈ R.
How can this observation be deduced from Lemma 2.4.26?
68 CHAPTER 2. FOUNDATIONS

Definition 2.5.37. Let M ⊂ Rk be a smooth manifold and let G ⊂ GL(n, R)

be a Lie group. A (smooth) group action of G on M is a smooth map

G × M → M : (g, p) 7→ φg (p) (2.5.23)

that for each pair g, h ∈ G satisfies the condition

φg ◦ φh = φgh , φ1l = id. (2.5.24)

If (2.5.23) is a smooth group action, then the infinitesimal action of the

Lie algebra g := Lie(G) on M is the map g → Vect(M ) : ξ 7→ Xξ defined by

d 
Xξ (p) :=  φexp(yξ) (p) (2.5.25)
dt t=0
for ξ ∈ g and p ∈ M .
Exercise 2.5.38. Let (2.5.23) be a smooth group action of a Lie group G
on a manifold M . Prove that

Xg−1 ξg = φ∗g Xξ , X[ξ,η] = [Xξ , Xη ] (2.5.26)

for all g ∈ G and all ξ, η ∈ g = Lie(G).

Exercise 2.5.39. Show that the maps GL(m, R) × Rm → Rm : (gx) 7→ gx,
SO(m + 1) × S m → S m : (g, x) 7→ gx, and R × S 1 → S 1 : (θ, z) 7→ eiθ z are
smooth group actions. Verify the formulas in (2.5.26) in these examples.
A smooth group action of a Lie group G on a manifold M can the thought
of as a “Lie group homomorphism”

G → Diff(M ) : g 7→ φg . (2.5.27)

While the group Diff(M ) is infinite-dimensional, and so cannot cannot be a

Lie group in the formal sense, it has many properties in common with Lie
groups as explained above. For example, one can define what is meant by a
smooth path in Diff(M ) and extend formally the notion of a tangent vector
(as the derivative of a path through a given element of Diff(M )) to this
setting. In particular, the tangent space of Diff(M ) at the identity can then
be identified with the space of vector fields Tid Diff(M ) = Vect(M ), and the
infinitesimal action in (2.5.25) is the Lie algebra homomorphism associated
to the “Lie group homomorphism” (2.5.27). In fact, we have chosen the sign
in the definition of the Lie bracket of vector fields so that the map ξ 7→ Xξ
is a Lie algebra homomorphism and not a Lie algebra anti-homomorphism.
This will be discussed further in §2.5.7.
2.5. LIE GROUPS 69

2.5.6 Smooth Maps and Algebra Homomorphisms

Let M be a smooth submanifold of Rk . Denote by F (M ) := C ∞ (M, R)
the space of smooth real valued functions f : M → R. Then F (M ) is a
commutative unital algebra. Each p ∈ M determines a unital algebra ho-
momorphism εp : F (M ) → R defined by εp (f ) = f (p) for p ∈ M .
Theorem 2.5.40. Every unital algebra homomorphism ε : F (M ) → R has
the form ε = εp for some p ∈ M .
Proof. Assume that ε : F (M ) → R is an algebra homomorphism.
Claim. For all f, g ∈ F (M ) we have ε(g) = 0 =⇒ ε(f ) ∈ f (g −1 (0)).
Indeed, the function f − ε(f ) · 1 lies in the kernel of ε and so the func-
tion h := (f − ε(f ) · 1)2 + g 2 also lies in the kernel of ε. There must be at
least one point p ∈ M where h(p) = 0 for otherwise 1 = ε(h)ε(1/h) = 0.
For this point p we have f (p) = ε(f ) and g(p) = 0, hence p ∈ g −1 (0), and
therefore ε(f ) = f (p) ∈ f (g −1 (0)). This proves the claim.
The theorem asserts that there exists a p ∈ M such that every f ∈ F (M )
satisfies ε(f ) = f (p). Assume, by contradiction, that this is false. Then for
every p ∈ M there exists a function f ∈ F (M ) such that f (p) 6= ε(f ). Re-
place f by f − ε(f ) to obtain f (p) 6= 0 = ε(f ). Now use the axiom of choice
to obtain a family of functions fp ∈ F (M ), one for every p ∈ M , such
that fp (p) 6= 0 = ε(fp ) for all p ∈ M . Then the set Up := fp−1 (R \ {0}) is
an M -open neighborhood of p for every p ∈ M . Choose a sequence of S com-
pact sets Kn ⊂ M such that Kn ⊂ intM (Kn+1 ) for all n and M = n Kn .
Then, for each n, there is a gn ∈ F (M ) (a finite sum of the form i fp2i ) such
P
that ε(gn ) = 0 and gn (q) > 0 for all q ∈ Kn . If M is compact, this is already
a contradiction because a positive function cannot belong to the kernel of ε.
Otherwise, choose f ∈ F (M ) such that f (q) ≥ n for all q ∈ M \ Kn and
all n ∈ N. Then ε(f ) ∈ f (gn−1 (0)) ⊂ f (M \ Kn ) ⊂ [n, ∞) by the claim and
so ε(f ) ≥ n for all n. This is a contradiction and proves Theorem 2.5.40.

Now let N be another smooth submanifold (say of R` ) and let C ∞ (M, N )

denote the space of smooth maps from M to N . A homomorphism from
F (N ) to F (M ) is a (real) linear map Φ : F (N ) → F (M ) that satisfies

Φ(f g) = Φ(f )Φ(g), Φ(1) = 1.

Let Hom(F (N ), F (M )) denote the space of homomorphisms from F (N )

to F (M ). An automorphism of the algebra F (M ) is a bijective homo-
morphism Φ : F (M ) → F (M ). The automorphisms of F (M ) form a group
denoted by Aut(F (M )).
70 CHAPTER 2. FOUNDATIONS

Corollary 2.5.41. The pullback operation

C ∞ (M, N ) → Hom(F (N ), F (M )) : φ 7→ φ∗
is bijective. In particular, the map Diff(M ) → Aut(F (M )) : φ 7→ φ∗ is an
anti-isomorphism of groups.
Proof. This is an exercise with hint. Let Φ : F (N ) → F (M ) be a unital
algebra homomorphism. By Theorem 2.5.40 there exists a map φ : M → N
such that εp ◦ Φ = εφ(p) for all p ∈ M . Prove that f ◦ φ : M → R is smooth
for every smooth map f : N → R and deduce that φ is smooth.
Remark 2.5.42. The pullback operation is functorial, i.e.
(ψ ◦ φ)∗ = ψ ∗ ◦ φ∗ , id∗M = idF (M ) .
for φ ∈ C ∞ (M, N ) and ψ ∈ C ∞ (N, P ). Here id denotes the identity map
of the space indicated in the subscript. Hence Corollary 2.5.41 may be
summarized by saying that the category of smooth manifolds and smooth
maps is anti-isomorphic to a subcategory of the category of commutative
unital algebras and unital algebra homomorphisms.
Exercise 2.5.43. If M is compact, then there is a slightly different way to
prove Theorem 2.5.40. An ideal in F (M ) is a linear subspace J ⊂ F (M )
satisfying the condition f ∈ F (M ), g ∈ J =⇒ f g ∈ J . A maximal
ideal in F (M ) is an ideal J ( F (M ) such that every ideal J 0 ( F (M )
containing J is equal to J . Prove that, if M is compact and J ⊂ F (M )
is an ideal with the property that for every p ∈ M there is an f ∈ J
with f (p) 6= 0, then J = F (M ). Deduce that each maximal ideal in F (M )
has the form Jp := {f ∈ F (M ) | f (p) = 0} for some p ∈ M .
Exercise 2.5.44. If M is compact, give another proof of Corollary 2.5.41
as follows. The set Φ−1 (Jp ) is a maximal ideal in F (N ) for each p ∈ M .
Use Exercise 2.5.43 to deduce that there is a unique map φ : M → N such
that Φ−1 (Jp ) = Jφ(p) for all p ∈ M . Show that φ is smooth and φ∗ = Φ.
Exercise 2.5.45. It is a theorem of ring theory that, when I ⊂ R is an ideal
in a ring R, the quotient ring R/I is a field if and only if the ideal I is max-
imal. Show that the kernel of the ring homomorphism εp : F (M ) → R of
Theorem 2.5.40 is the ideal Jp of Exercise 2.5.43. Conclude that M is com-
pact if and only if every maximal ideal J in F (M ) is of the form J = Jp
for some p ∈ M . Hint: The functions of compact support form an ideal. It
can be shown that if M is not compact and J is a maximal ideal contain-
ing all functions of compact support, then the quotient field F (M )/J is a
non-Archimedean ordered field which properly contains R.
2.5. LIE GROUPS 71

2.5.7 Vector Fields and Derivations

A derivation of F (M ) is a linear map δ : F (M ) → F (M ) that satisfies
δ(f g) = δ(f )g + f δ(g).
and the derivations form a Lie algebra denoted by Der(F (M )). We may
think of Der(F (M )) as the Lie algebra of Aut(F (M )) with the Lie bracket
given by the commutator. By Theorem 2.5.40 the pullback operation
Diff(M ) → Aut(F (M )) : φ 7→ φ∗ (2.5.28)
can be thought of as a Lie group anti-isomorphism. Differentiating it at the
identity φ = id gives a linear map
Vect(M ) → Der(F (M )) : X 7→ LX . (2.5.29)
Here the operator LX : F (M ) → F (M ) is given by the derivative of a
function f in the direction of the vector field X, i.e.

d 
LX f := df · X =  f ◦ φt ,
dt t=0

where φtdenotes the flow of X. Since the map (2.5.29) is the derivative
of the “Lie group” anti-homomorphism (2.5.28) we expect it to be a Lie
algebra anti-homomorphism. Indeed, one can show that
L[X,Y ] = LY LX − LX LY = −[LX , LY ] (2.5.30)
for X, Y ∈ Vect(M ). This confirms that our sign in the definition of the
Lie bracket is consistent with the standard conventions in the theory of
Lie groups. In the literature the difference between a vector field and the
associated derivation LX is sometimes neglected in the notation and many
authors write Xf := df · X = LX f , thus thinking of a vector field on a
manifold M as an operator on the space of functions. With this notation one
obtains the equation [X, Y ]f = Y (Xf ) − X(Y f ) and here lies the origin for
the use of the opposite sign for the Lie bracket in many books on differential
geometry.
Exercise 2.5.46. Prove that the map (2.5.29) is bijective. Hint: Fix a
derivation δ ∈ Der(F (M )) and prove the following. Fact 1: If U ⊂ M is
an open set and f ∈ F (M ) vanishes on U , then δ(f ) vanishes on U . Fact 2:
If p ∈ M and the derivative df (p) : Tp M → R is zero, then (δ(f ))(p) = 0.
(By Fact 1, the proof of Fact 2 can be reduced to an argument in local
coordinates.)
Exercise 2.5.47. Verify the formula (2.5.30).
72 CHAPTER 2. FOUNDATIONS

2.6 Vector Bundles and Submersions

2.6.1 Submersions
Let M ⊂ Rk be a smooth m-manifold and N ⊂ R` be a smooth n-manifold.
A smooth map f : N → M is called a submersion iff its derivative

df (q) : Tq N → Tf (q) M

is surjective for every q ∈ N .

N q0

g f

M U p0

Figure 2.10: A local right inverse of a submersion.

Lemma 2.6.1. Let M ⊂ Rk be a smooth m-manifold, N ⊂ R` be a smooth

n-manifold, and f : N → M be a smooth map. The following are equivalent.
(i) f is a submersion.
(ii) For every q0 ∈ N there is an M -open neighborhood U of p0 := f (q0 ) and
a smooth map g : U → N such that g(f (q0 )) = q0 and f ◦ g = id : U → U .
Thus f has a local right inverse near every point in N (see Figure 2.10).

Proof. We prove that (i) implies (ii). Since the derivative

df (q0 ) : Tq0 N → Tp0 M

is surjective we have n ≥ m and

dim ker df (q0 ) = n − m.

Hence there is a linear map A : R` → Rn−m whose restriction to the kernel

of df (q0 ) is bijective. Now define the map ψ : N → M × Rn−m by

ψ(q) := (f (q), A(q − q0 ))

for q ∈ N . Then ψ(q0 ) = (p0 , 0) and its derivative

dψ(q0 ) : Tq0 N → Tp0 M × Rn−m

2.6. VECTOR BUNDLES AND SUBMERSIONS 73

sends w ∈ Tq0 N to (df (q0 )w, Aw) and is therefore bijective. Hence it follows
from the inverse function theorem for manifolds (Theorem 2.2.17) that there
exists an N -open neighborhood V ⊂ N of q0 such that the set

W := ψ(N ) ⊂ M × Rn−m

is an open neighborhood of (p0 , 0) and ψ|V : V → W is a diffeomorphism.

Let
U := {p ∈ M | (p, 0) ∈ W }
and define the map g : U → N by

g(p) := ψ −1 (p, 0).

Then p0 ∈ U , g is smooth and

(p, 0) = ψ(g(p)) = (f (g(p)), A(g(p) − q0 )).

Hence f (g(p)) = p for all p ∈ U and

g(p0 ) = ψ −1 (p0 , 0) = q0 .

This shows that (i) implies (ii). The converse is an easy consequence of the
chain rule and is left to the reader. This proves Lemma 2.6.1

Corollary 2.6.2. The image of a submersion f : N → M is open.

Proof. If p0 = f (q0 ) ∈ f (N ), then the neighborhood U ⊂ M of p0 in

Lemma 2.6.1 (ii) is contained in the image of f .

Corollary 2.6.3. If N is a nonempty compact manifold, M is a connected

manifold, and f : N → M is a submersion, then f is surjective and M is
compact.

Proof. The image f (N ) is an open subset of M by Corollary 2.6.2, it is a

relatively closed subset of M because N is compact, and it is nonempty
because N is nonempty. Since M is connected this implies that f (N ) = M .
In particular, M is compact.

Exercise 2.6.4. Let f : N → M be a smooth map. Prove that the

sets {q ∈ N | df (q) is injective} and {q ∈ N | df (q) is surjective} are open (in
the relative topology of N ).
74 CHAPTER 2. FOUNDATIONS

2.6.2 Vector Bundles

Let M ⊂ Rk be an m-dimensional smooth manifold.
Definition 2.6.5. A (smooth) vector bundle (over M of rank n) is a
smooth submanifold E ⊂ M × R` of dimension m + n such that, for every
point p ∈ M , the set
n o
Ep := v ∈ R` | (p, v) ∈ E

is an n-dimensional linear subspace of R` (called the fiber of E over p).

A vector bundle E over M is equipped with a smooth map
π:E→M
defined by π(p, v) := p. This map is called the canonical projection of E.
If E ⊂ M × R` is a vector bundle, then a (smooth) section of E is a
smooth map s : M → R` such that s(p) ∈ Ep for every p ∈ M .
A section s : M → R` of a vector bundle E over M determines a smooth
map σ : M → E which sends the point p ∈ M to the pair (p, s(p)) ∈ E.
This map satisfies π ◦ σ = id. It is sometimes convenient to abuse notation
and eliminate the distinction between s and σ. Thus we will sometimes use
the same letter s for the map from M to R` and the map from M to E.
Definition 2.6.6. Let M ⊂ Rk be a smooth m-manifold. The set
T M := {(p, v) | p ∈ M, v ∈ Tp M }
is called the tangent bundle of M .
The tangent bundle is a subset of M × Rk and, for every p ∈ M , its
fiber Tp M is an m-dimensional linear subspace of Rk by Theorem 2.2.3.
However, it is not immediately obvious from the definition that T M is a
submanifold of M × Rk . This will be proved below. The sections of T M are
the vector fields on M .
Exercise 2.6.7. Let f : M → N be a smooth map between manifolds.
Prove that the tangent map T M → T N : (p, v) 7→ (f (p), df (p)v) is smooth.
Exercise 2.6.8. Let M ⊂ Rk be a smooth m-manifold and let φ : U → Ω
be a smooth coordinate chart on an M -open set U ⊂ M with values in an
open set Ω = φ(U ) ⊂ Rm . Prove that the map φe : T U → Ω × Rm defined
e v) := (φ(p), dφ(p)) for p ∈ U and v ∈ Tp M is a diffeomorphism. It
by φ(p,
is called a standard coordinate chart on T M . Deduce that T M is
a smooth 2m-dimensional submanifold of M × Rk and hence is a smooth
vector bundle over M . (See also Corollary 2.6.12 below.)
2.6. VECTOR BUNDLES AND SUBMERSIONS 75

Exercise 2.6.9. Let V ⊂ R` be an n-dimensional linear subspace. The

orthogonal projection of R` onto V is the matrix Π ∈ R`×` that satisfies

Π = Π2 = ΠT , im Π = V. (2.6.1)

Prove that there is a unique matrix Π ∈ R`×` satisfying (2.6.1). Prove that,
for every symmetric matrix S = S T ∈ R`×` , the kernel of S is the orthogonal
complement of the image of S. If D ∈ R`×n is any injective matrix whose
image is V , prove that det(DT D) 6= 0 and

Π = D(DT D)−1 DT . (2.6.2)

Theorem 2.6.10 (Vector Bundles). Let M ⊂ Rk be a smooth m-manifold

and let E ⊂ M × R` be a subset. Assume that, for every p ∈ M , the set
n o
Ep := v ∈ R` | (p, v) ∈ E (2.6.3)

is an n-dimensional linear subspace of R` . Let Π : M → R`×` be the map

that assigns to each p ∈ M the orthogonal projection of R` onto Ep , i.e.

Π(p) = Π(p)2 = Π(p)T , im Π(p) = Ep . (2.6.4)

Then the following are equivalent.

(i) E is a vector bundle.
(ii) For every p0 ∈ M and every v0 ∈ Ep0 there is a smooth map s : M → R`
such that s(p0 ) = v0 and s(p) ∈ Ep for all p ∈ M .
(iii) The map Π : M → R`×` is smooth.
(iv) For every p0 ∈ M there is an open neighborhood U ⊂ M of p0 and a
diffeomorphism π −1 (U ) → U × Rn : (p, v) 7→ Φ(p, v) = (p, Φp (v)) such that
the map Φp : Ep → Rn is an isometric isomorphism for all p ∈ U .
(v) For every p0 ∈ M there is an open neighborhood U ⊂ M of p0 and a
diffeomorphism π −1 (U ) → U × Rn : (p, v) 7→ Φ(p, v) = (p, Φp (v)) such that
the map Φp : Ep → Rn is a vector space isomorphism for all p ∈ U .
Condition (i) implies that the projection π : E → M is a submersion. In (ii)
the section s can be chosen to have compact support, i.e. there exists a
compact subset K ⊂ M such that s(p) = 0 for all p ∈ M \ K.

Proof. See page 77.

76 CHAPTER 2. FOUNDATIONS

Definition 2.6.11. The maps Φ : π −1 (U ) → U × Rn in Theorem 2.6.10 are

called local trivializations of E. They fit into commutative diagrams

π −1 (U )
Φ / U × Rn .
NNN pp
NNN
N ppppp
π NNNN p
ppp pr1
NN' pw pp
U
Corollary 2.6.12. Let M ⊂ Rk be a smooth m-manifold. Then T M is a
vector bundle over M and hence is a smooth 2m-manifold in Rk × Rk .
Proof. Let φ : U → Ω be a coordinate chart on an M -open set U ⊂ M with
values in an open subset Ω ⊂ Rm . Denote its inverse by ψ := φ−1 : Ω → M .
By Theorem 2.2.3 the linear map dψ(x) : Rm → Rk is injective and its image
is Tψ(x) M for every x ∈ Ω. Hence the map D : U → Rk×m defined by
D(p) := dψ(φ(p)) ∈ Rk×m
is smooth and, for every p ∈ U , the linear map D(p) : Rm → Rk is injec-
tive and its image is Tp M . Thus the function ΠT M : M → Rk×k defined
by (2.6.4) with Ep = Tp M is given by
 −1
ΠT M (p) = D(p) D(p)T D(p) D(p)T for p ∈ U.

Hence ΠT M is smooth and so T M is a vector bundle by Theorem 2.6.10.

Let M ⊂ Rk be an m-manifold, N ⊂ R` be an n-manifold, f : N → M
be a smooth map, and E ⊂ M × Rd be a vector bundle. The pullback
bundle is the vector bundle f ∗ E → N defined by
n o
f ∗ E := (q, v) ∈ N × Rd | v ∈ Ef (q)

and the normal bundle of E is the vector bundle E ⊥ → M defined by

n o
E ⊥ := (p, w) ∈ M × Rd | hv, wi = 0 ∀ v ∈ Ep .

Corollary 2.6.13. The pullback and normal bundles are vector bundles.
Proof. Let Π = ΠE : M → Rd×d be the map defined by (2.6.4). This map
is smooth by Theorem 2.6.10. Moreover, the corresponding maps for f ∗ E
and E ⊥ are given by
∗E ⊥
Πf = ΠE ◦ f : N → Rd×d , ΠE = 1l − ΠE : M → Rd×d .
These maps are smooth and hence it follows from Theorem 2.6.10 that f ∗ E
and E ⊥ are vector bundles.
2.6. VECTOR BUNDLES AND SUBMERSIONS 77

Proof of Theorem 2.6.10. We first assume that E is a vector bundle and

prove that π : E → M is a submersion. Let σ : M → E denote the zero
section given by σ(p) := (p, 0). Then π ◦σ = id and hence it follows from the
chain rule that the derivative dπ(p, 0) : T(p,0) E → Tp M is surjective. Now
it follows from Exercise 2.6.4 that for every p ∈ M there is an ε > 0 such
that the derivative dπ(p, v) : T(p,v) E → Tp M is surjective for every v ∈ Ep
with |v| < ε. Consider the map fλ : E → E defined by

fλ (p, v) := (p, λv).

This map is a diffeomorphism for every λ > 0. It satisfies

π = π ◦ fλ

and hence

dπ(p, v) = dπ(p, λv) ◦ dfλ (p, v) : T(p,v) E → Tp M.

Since dfλ (p, v) is bijective and dπ(p, λv) is surjective for λ < ε/ |v| it follows
that dπ(p, v) is surjective for every p ∈ M and every v ∈ Ep . Thus the
projection π : E → M is a submersion for every vector bundle E over M .
We prove that (i) implies (ii). Let p0 ∈ M and v0 ∈ Ep0 . We have
already proved that π is a submersion. Hence it follows from Lemma 2.6.1
that there exists an M -open neighborhood U ⊂ M of p0 and a smooth map

σ0 : U → E

such that
π ◦ σ0 = id : U → U, σ0 (p0 ) = (p0 , v0 ).
Define the map s0 : U → R` by

(p, s0 (p)) := σ0 (p) for p ∈ U.

Then s0 (p0 ) = v0 and s0 (p) ∈ Ep for all p ∈ U . Now choose ε > 0 such that

{p ∈ M | |p − p0 | < ε} ⊂ U

and choose a smooth cutoff function β : Rk → [0, 1] such that β(p0 ) = 1

and β(p) = 0 for |p − p0 | ≥ ε. Define s : M → R` by

β(p)s0 (p), if p ∈ U,
s(p) :=
0, if p ∈
/ U.

This map satisfies the requirements of (ii).

78 CHAPTER 2. FOUNDATIONS

We prove that (ii) implies (iii). Thus we assume that E satisfies (ii).
Choose p0 ∈ M and a basis v1 , . . . , vn of Ep0 . By (ii) there exists smooth
sections s1 , . . . , sn : M → R` of E such that si (p0 ) = vi for i = 1, . . . , n. Now
there exists an M -open neighborhood U ⊂ M of p0 such that the vec-
tors s1 (p), . . . , sn (p) are linearly independent, and hence form a basis of Ep
for every p ∈ U . Hence, for every p ∈ U , we have
Ep = imD(p), D(p) := [s1 (p) · · · sn (p)] ∈ R`×n .
By Exercise 2.6.9, this implies Π(p) = D(p)(D(p)T D(p))−1 D(p)T for ev-
ery p ∈ U . Thus every p0 ∈ M has a neighborhood U such that the re-
striction of Π to U is smooth. This shows that (ii) implies (iii).
We prove that (iii) implies (iv). Fix a point p0 ∈ M and choose a ba-
sis v1 , . . . , vn of Ep0 . For p ∈ M define
D(p) := [Π(p)v1 · · · Π(p)vn ] ∈ R`×n
Then D : M → R`×n is a smooth map and D(p0 ) has rank n. Hence the set
U := {p ∈ M | rankD(p) = n} ⊂ M
is an open neighborhood of p0 and Ep = imD(p) for all p ∈ U . Thus
π −1 (U ) = {(p, v) ∈ E | p ∈ U } ⊂ E
is an open set containing π −1 (p0 ). Define the map Φ : π −1 (U ) → U × Rn by
 −1/2
Φp (v) := D(p)T D(p) D(p)T v


Φ(p, v) := p, Φp (v) ,
for p ∈ U and v ∈ Ep . This map is bijective and its inverse is given by
 −1/2
Φ−1 (p, ξ) = p, Φ−1 −1 T


p (ξ) , Φ p (ξ) = D(p) D(p) D(p) ξ

for p ∈ U and ξ ∈ Rn . Thus Φ is a diffeomorphism and |Φp (v)| = |v| for

all p ∈ U and all v ∈ Ep . This shows that (iii) implies (iv).
That (iv) implies (v) is obvious.
We prove that (v) implies (i). Shrinking U if necessary, we may as-
sume that there exists a coordinate chart φ : U → Ω with values in an open
set Ω ⊂ Rm . Then the composition (φ × id) ◦ Φ : π −1 (U ) → Ω × Rn is a dif-
feomorphism. Thus E ⊂ Rk × R` is a manifold of dimension m + n and this
proves Theorem 2.6.10.
Exercise 2.6.14. Define the notion of an isomorphism between two vector
bundles E and F over M . Construct a vector bundle E ⊂ S 1 × R2 of rank 1
that does not admit a global trivialization, i.e. that is not isomorphic to the
trivial bundle S 1 × R. Such a vector bundle is called a Möbius strip.
2.6. VECTOR BUNDLES AND SUBMERSIONS 79

The Implicit Function Theorem

Next we carry over the Implicit Function Theorem in Corollary A.2.6 to
smooth maps on vector bundles.
Theorem 2.6.15 (Implicit Function Theorem).
Let M ⊂ Rk be a smooth m-manifold, let N ⊂ Rk be a smooth n-manifold,
let E ⊂ M × R` be a smooth vector bundle of rank n, let W ⊂ E be open,
and let f : W → N be a smooth map. For p ∈ M define fp : Wp → N by
Wp := {v ∈ Ep | (p, v) ∈ W } , fp (v) := f (p, v).
Let p0 ∈ M such that 0 ∈ Wp0 and dfp0 (0) : Tp0 M → Tq0 N is bijective,
where q0 := f (p0 , 0) ∈ N . Then there exists a constant ε > 0, open neighbor-
hoods U0 ⊂ M of p0 and V0 ⊂ N of q0 , and a smooth map h : U0 × V0 → R`
such that {(p, v) ∈ E | p ∈ U0 , |v| < ε} ⊂ W and
h(p, q) ∈ Ep , |h(p, q)| < ε (2.6.5)
for all (p, q) ∈ U0 × V0 and
fp (v) = 0 ⇐⇒ v = h(p, q) (2.6.6)
for all (p, q) ∈ U0 × V0 , and all v ∈ Ep with |v| < ε.
Proof. Choose a coordinate chart ψ : V → Rn on an open set V ⊂ N con-
taining q0 . Choose an open neighborhood U ⊂ M of p0 such that (p, 0) ∈ W
and f (p, 0) ∈ V for all p ∈ U , there is a coordinate chart φ : U → Ω ⊂ Rm ,
and there is a local trivialization Φ : π −1 (U ) → U × Rn as in Theorem 2.6.10
with |Φp (v)| = |v| for p ∈ U and v ∈ Ep . Define Br := {ξ ∈ Rn | |ξ| < r} and
choose r > 0 so small that Φ−1 (U × Br ) ⊂ W and f ◦ Φ−1 (U × Br ) ⊂ V .
Define the map F : Ω × Rn × Br → Rn by
F (x, y, ξ) := ψ ◦ f ◦ Φ−1 φ−1 (x), ξ − y

for (x, y) ∈ Ω × Rn and ξ ∈ Br . Let x0 := φ(p0 ) and y0 := ψ(q0 ). Then

we have F (x0 , y0 , 0) = 0 and the derivative d3 F (x0 , y0 , 0) : Rn → Rn of F
with respect to ξ at (x0 , y0 , 0) is bijective. Hence Corollary A.2.6 asserts
that there exist open neighborhoods U0 ⊂ U of p0 and V0 ⊂ V of q0 , a con-
stant 0 < ε < r, and a smooth map g : φ(U0 ) × ψ(V0 ) → Bε such that
F (x, y, ξ) = 0 ⇐⇒ g(x, y) = ξ
for all (x, y) ∈ φ(U0 ) × ψ(V0 ) and all ξ ∈ Bε . Thus the map
h : U0 × V 0 → R ` , h(p, q) := Φ−1


p g(φ(p), ψ(q)) ,
satisfies the requirements of Theorem 2.6.15.
80 CHAPTER 2. FOUNDATIONS

2.7 The Theorem of Frobenius

Let M ⊂ Rk be an m-dimensional manifold and n be a nonnegative integer.
A subbundle of rank n of the tangent bundle T M is a subset E ⊂ T M
that is itself a vector bundle of rank n over M , i.e. it is a submanifold
of T M and the fiber Ep = {v ∈ Tp M | (p, v) ∈ E} is an n-dimensional linear
subspace of Tp M for every p ∈ M . Note that the rank n of a subbundle
is necessarily less than or equal to m. In the literature a subbundle of the
tangent bundle is sometimes called a distribution on M . We shall, however,
not use this terminology in order to avoid confusion with the concept of a
distribution in the functional analytic setting.
Definition 2.7.1. Let M ⊂ Rk be an m-dimensional manifold and E ⊂ T M
be a subbundle of rank n. The subbundle E is called involutive if, for any
two vector fields X, Y ∈ Vect(M ), we have

X(p), Y (p) ∈ Ep ∀ p ∈ M =⇒ [X, Y ](p) ∈ Ep ∀ p ∈ M. (2.7.1)

The subundle E is called integrable if, for every p0 ∈ M , there exists a

submanifold N ⊂ M such that p0 ∈ N and Tp N = Ep for every p ∈ N .
A foliation box for E (see Figure 2.11) is a coordinate chart φ : U → Ω
on an M -open subset U ⊂ M with values in an open set Ω ⊂ Rn × Rm−n
such that the set Ω ∩ (Rn × {y}) is connected for every y ∈ Rm−n and, for
every p ∈ U and every v ∈ Tp M , we have

v ∈ Ep ⇐⇒ dφ(p)v ∈ Rn × {0}.

U φ Ω

Figure 2.11: A foliation box.

Theorem 2.7.2 (Frobenius). Let M ⊂ Rk be an m-dimensional manifold

and E ⊂ T M be a subbundle of rank n. Then the following are equivalent.
(i) E is involutive.
(ii) E is integrable.
(iii) For every p0 ∈ M there exists a foliation box φ : U → Ω with p0 ∈ U .
Proof. See page 81.
2.7. THE THEOREM OF FROBENIUS 81

It is easy to show that (iii) =⇒ (ii) =⇒ (i) (see below). The hard part of
the theorem is to prove that (i) =⇒ (iii). This requires the following lemma.
Lemma 2.7.3. Let E ⊂ T M be an involutive subbundle and X ∈ Vect(M )
be a complete vector field such that X(p) ∈ Ep for every p ∈ M . Denote by

R → Diff(M ) : t 7→ φt

the flow of X. Then, for all t ∈ R and all p ∈ M , we have

dφt (p)Ep = Eφt (p) . (2.7.2)

Proof. See page 83.

Lemma 2.7.3 implies Theorem 2.7.2. We prove first that (iii) implies (ii).
Let p0 ∈ M , choose a foliation box φ : U → Ω for E with p0 ∈ U , and define

N := (p ∈ U | φ(p) ∈ Rn × {y0 }}

where (x0 , y0 ) := φ(p0 ) ∈ Ω. Then N satisfies the requirements of (ii).

We prove that (ii) implies (i). Choose two vector fields X, Y ∈ Vect(M )
that satisfy X(p), Y (p) ∈ Ep for all p ∈ M and fix a point p0 ∈ M . Then,
by (ii), there exists a submanifold N ⊂ M containing p0 such that Tp N = Ep
for every p ∈ N . Hence the restrictions X|N and Y |N are vector fields
on N and so is the restriction of the Lie bracket [X, Y ] to N . Thus we
have [X, Y ](p0 ) ∈ Tp0 N = Ep0 as claimed.
We prove that (i) implies (iii). Thus we assume that E is an involutive
subbundle of T M and fix a point p0 ∈ M . By Theorem 2.6.10 there exist
vector fields X1 , . . . , Xn ∈ Vect(M ) such that Xi (p) ∈ Ep for all i and p and
the vectors X1 (p0 ), . . . , Xn (p0 ) form a basis of Ep0 . Using Theorem 2.6.10
again we find vector fields Y1 , . . . , Ym−n ∈ Vect(M ) such that the vectors

X1 (p0 ), . . . , Xn (p0 ), Y1 (p0 ), . . . , Ym−n (p0 )

form a basis of Tp0 M . Using cutoff functions as in the proof of Theo-

rem 2.6.10 we may assume without loss of generality that the vector fields Xi
and Yj have compact support and hence are complete (see Exercise 2.4.13).
Denote by φt1 , . . . , φtn the flows of the vector fields X1 , . . . , Xn , respectively,
and by ψ1t , . . . , ψm−n
t the flows of the vector fields Y1 , . . . , Ym−n . Define the
map
ψ : Rn × Rm−n → M
by
y
ψ(x, y) := φx1 1 ◦ · · · ◦ φxnn ◦ ψ1y1 ◦ · · · ◦ ψm−n
m−n
(p0 ).
82 CHAPTER 2. FOUNDATIONS

By Lemma 2.7.3, this map satisfies

∂ψ
(x, y) ∈ Eψ(x,y) (2.7.3)
∂xi

for all x ∈ Rn and y ∈ Rm−n . Moreover,

∂ψ ∂ψ
(0, 0) = Xi (p0 ), (0, 0) = Yj (p0 ),
∂xi ∂yj

and so the derivative

dψ(0, 0) : Rn × Rm−n → Tp0 M

is bijective. Hence, by the Inverse Function Theorem 2.2.17, there exists an

open neighborhood Ω ⊂ Rn × Rm−n of the origin such that the set

U := ψ(Ω) ⊂ M

is an M -open neighborhood of p0 and ψ|Ω : Ω → U is a diffeomorphism.

Thus the vectors ∂ψ/∂xi (x, y) are linearly independent for every (x, y) ∈ Ω
and, by (2.7.3), form a basis of Eψ(x,y) . Hence

φ := (ψ|Ω )−1 : U → Ω

is a foliation box and this proves Theorem 2.7.2, assuming Lemma 2.7.3.

To complete the proof of the Frobenius theorem it remains to prove

Lemma 2.7.3. This requires the following result.

Lemma 2.7.4. Let E ⊂ T M be an involutive subbundle. If β : R2 → M is

a smooth map such that

∂β ∂β
(s, 0) ∈ Eβ(s,0) , (s, t) ∈ Eβ(s,t) , (2.7.4)
∂s ∂t
for all s, t ∈ R, then
∂β
(s, t) ∈ Eβ(s,t) , (2.7.5)
∂s
for all s, t ∈ R.

Proof. See page 83.

2.7. THE THEOREM OF FROBENIUS 83

Lemma 2.7.4 implies Lemma 2.7.3. Let X ∈ Vect(M ) be a complete vector

field satisfying X(p) ∈ Ep for every p ∈ M and let φt be the flow of X.
Choose a point p0 ∈ M and a vector v0 ∈ Ep0 . By Theorem 2.6.10 there is a
vector field Y ∈ Vect(M ) with values in E such that Y (p0 ) = v0 . Moreover
this vector field may be chosen to have compact support and hence it is
complete (see Exercise 2.4.13). Thus there is a solution γ : R → M of the
initial value problem
γ̇(s) = Y (γ(s)), γ(0) = p0 .
Define β : R2 → M by
β(s, t) := φt (γ(s))
for s, t ∈ R. Then
∂β
(s, 0) = γ̇(s) = Y (γ(s)) ∈ Eβ(s,0) ,
∂s
∂β
(s, t) = X(β(s, t)) ∈ Eβ(s,t)
∂t
for all s, t ∈ R. Hence it follows from Lemma 2.7.4 that
∂β
dφt (p0 )v0 = dφt (γ(0))γ̇(0) =(0, t) ∈ Eφt (p0 )
∂s
for every t ∈ R. This proves Lemma 2.7.3, assuming Lemma 2.7.4.
Proof of Lemma 2.7.4. Given any point p0 ∈ M we choose a coordinate
chart φ : U → Ω, defined on an M -open set U ⊂ M with values in an
open set Ω ⊂ Rn × Rm−n , such that p0 ∈ U and dφ(p0 )Ep0 = Rn × {0}.
Shrinking U , if necessary, we find that for every p ∈ U the linear sub-
space dφ(p)Ep ⊂ Rn × Rm−n is the graph of a matrix A ∈ R(m−n)×n . Thus
there exists a smooth map A : Ω → R(m−n)×n such that, for every p ∈ U ,
dφ(p)Ep = {(ξ, A(x, y)ξ) | ξ ∈ Rn } , (x, y) := φ(p) ∈ Ω. (2.7.6)
For (x, y) ∈ Ω define the linear maps
∂A ∂A
(x, y) : Rn → R(m−n)×n , (x, y) : Rm−n → R(m−n)×n
∂x ∂y
by
n m−n
∂A X ∂A ∂A X ∂A
(x, y) · ξ := ξi (x, y), (x, y) · η := ηj (x, y),
∂x ∂xi ∂y ∂yj
i=1 j=1

for ξ = (ξ1 , . . . , ξn ) ∈ Rn and η = (η1 , . . . , ηm−n ) ∈ Rm−n . We prove the

following.
84 CHAPTER 2. FOUNDATIONS

Claim 1. Let (x, y) ∈ Ω, ξ, ξ 0 ∈ Rn and define η, η 0 ∈ Rm−n by η := A(x, y)ξ

and η 0 := A(x, y)ξ 0 . Then
   
∂A ∂A 0 ∂A 0 ∂A 0
(x, y) · ξ + (x, y) · η ξ = (x, y) · ξ + (x, y) · η ξ.
∂x ∂y ∂x ∂y

The graphs of the matrices A(z) determine a subbundle E e ⊂ Ω × Rm with

the fibers
ez := (ξ, η) ∈ Rn × Rm−n | η = A(x, y)ξ

E

for z = (x, y) ∈ Ω. This subbundle is the image of the restriction

E|U := {(p, v) | p ∈ U, v ∈ Ep }

under the diffeomorphism T M |U → Ω × Rm : (p, v) 7→ (φ(p), dφ(p)v) and

hence it is involutive. Now fix two elements ξ, ξ 0 ∈ Rn and define the vector
fields ζ, ζ 0 : Ω → Rm by

ζ(z) := (ξ, A(z)ξ), ζ 0 (z) := (ξ 0 , A(z)ξ 0 ), z ∈ Ω.

Then ζ and ζ 0 are sections of E

e and their Lie bracket [ζ, ζ 0 ] is given by

[ζ, ζ 0 ](z) = 0, dA(z)ζ 0 (z) ξ − (dA(z)ζ(z)) ξ 0 .

e is involutive the Lie bracket [ζ, ζ 0 ] must take values in the graph
Since E
of A. Hence the right hand side vanishes and this proves Claim 1.

Claim 2. Let I, J ⊂ R be open intervals and let z = (x, y) : I × J → Ω be

a smooth map. Fix two points s0 ∈ I and t0 ∈ J and assume that

∂y
∂x
(s0 , t0 ) = A x(s0 , t0 ), y(s0 , t0 ) (s0 , t0 ), (2.7.7)
∂s ∂s
∂y
∂x
(s, t) = A x(s, t), y(s, t) (s, t) (2.7.8)
∂t ∂t

for all s ∈ I and t ∈ J. Then

∂y
∂x
(s0 , t) = A x(s0 , t), y(s0 , t) (s0 , t) (2.7.9)
∂s ∂s

for all t ∈ J.
2.7. THE THEOREM OF FROBENIUS 85

Equation (2.7.9) holds by assumption for t = t0 . Moreover, dropping the

argument z(s0 , t) = z = (x, y) for notational convenience we obtain
∂2y ∂2x
   
∂ ∂y ∂x ∂A ∂x ∂A ∂y ∂x
−A· = −A − · + ·
∂t ∂s ∂s ∂s∂t ∂s∂t ∂x ∂t ∂y ∂t ∂s
∂2y ∂2x
  
∂A ∂x ∂A ∂x ∂x
= −A − · + · A
∂s∂t ∂s∂t ∂x ∂t ∂y ∂t ∂s
2 2
  
∂ y ∂ x ∂A ∂x ∂A ∂x ∂x
= −A − · + · A
∂s∂t ∂s∂t ∂x ∂s ∂y ∂s ∂t
2 2
 
∂ y ∂ x ∂A ∂x ∂A ∂y ∂x
= −A − · + ·
∂s∂t ∂s∂t ∂x ∂s ∂y ∂s ∂t
  
∂A ∂y ∂x ∂x
+ · −A
∂y ∂s ∂s ∂t
  
∂A ∂y ∂x ∂x
= · −A
∂y ∂s ∂s ∂t
Here the second step follows from (2.7.8), the third step follows from Claim 1,
and the last step follows by differentiating equation (2.7.8) with respect to s.
Define the curve η : J → Rm−n by
∂y
∂x
η(t) := (s0 , t) − A x(s0 , t), y(s0 , t) (s0 , t).
∂s ∂s
By (2.7.7) and what we have just proved, the curve η satisfies the linear
differential equation
 
∂A
∂x
η̇(t) = x(s0 , t), y(s0 , t) · η(t) (s0 , t), η(t0 ) = 0.
∂y ∂t
Hence η(t) = 0 for all t ∈ J. This proves (2.7.9) and Claim 2.
Now let β : R2 → M be a smooth map satisfying (2.7.4) and fix a real
number s0 . Consider the set W := {t ∈ R  ∂s β(s0 , t) ∈ Eβ(s0 ,t) }. By going

to local coordinates, we obtain from Claim 2 that W is open. Moreover, W
is obviously closed, and W 6= ∅ because 0 ∈ W by (2.7.4). Hence W = R.
Since s0 ∈ R was chosen arbitrarily, this proves (2.7.5) and Lemma 2.7.4.
Any subbundle E ⊂ T M determines an equivalence relation on M via
there is a smooth curve γ : [0, 1] → M
p0 ∼ p1 ⇐⇒ (2.7.10)
such that γ(0) = p0 , γ(1) = p1 , γ̇(t) ∈ Eγ(t) ∀ t
If E is integrable, this equivalence relation is called a foliation and the
equivalence class of p0 ∈ M is called the leaf of the foliation through p0 .
The next example shows that the leaves do not need to be submanifolds
86 CHAPTER 2. FOUNDATIONS

Example 2.7.5. Consider the torus M := S 1 × S 1 ⊂ C2 with the tangent

bundle
T M = (z1 , z2 , iλ1 z1 , iλ2 z2 ) ∈ C4 | |z1 | = |z2 | = 1, λ1 , λ2 ∈ R .


Let ω1 , ω2 be real numbers and consider the subbundle

E := (z1 , z2 , itω1 z1 , itω2 z2 ) ∈ C4 | |z1 | = |z2 | = 1, t ∈ R .


The leaf of this subbundle through z = (z1 , z2 ) ∈ T2 is given by

n  o
L= eitω1 z1 , eitω2 z2  t ∈ R .


It is a submanifold if and only if the quotient ω1 /ω2 is a rational number

(or ω2 = 0). Otherwise each leaf is a dense subset of T2 .
Exercise 2.7.6. Prove that (2.7.10) defines an equivalence relation for every
subbundle E ⊂ T M .
Exercise 2.7.7. Each subbundle E ⊂ T M of rank 1 is integrable.
Exercise 2.7.8. Consider the manifold M = R3 . Prove that the sub-
bundle E ⊂ T M = R3 × R3 with fiber Ep = (ξ, η, ζ) ∈ R3 | ζ − yξ = 0

over p = (x, y, z) ∈ R3 is not integrable and that any two points in R3 can
be joined by a path tangent to E.
Exercise 2.7.9. Consider the manifold M = S 3 ⊂ R4 = C2 and define
E := (z, ζ) ∈ C2 × C2 | |z| = 1, ζ ⊥ z, iζ ⊥ z .


Thus the fiber

Ez ⊂ Tz S 3 = z ⊥
is the maximal complex linear subspace of Tz S 3 . Prove that E has real
rank 2 and is not integrable.
Exercise 2.7.10. Let E ⊂ T M be an involutive subbundle of rank n and
let L ⊂ M be a leaf of the foliation determined by E. A subset V ⊂ L is
called L-open iff it can be written as a union of submanifolds N of M
with tangent spaces Tp N = Ep for p ∈ N . Prove that the L-open sets form
a topology on L (called the intrinsic topology). Prove that the obvious
inclusion ι : L → M is continuous with respect to the intrinsic topology on L.
Prove that the inclusion ι : L → M is proper if and only if the intrinsic
topology on L agrees with the relative topology inherited from M (called
the extrinsic topology).
Remark 2.7.11. It is surprisingly difficult to prove that each closed leaf L
of a foliation is a submanifold of M . A proof due to David Epstein [13] is
sketched in §2.9.4 below.
2.8. THE INTRINSIC DEFINITION OF A MANIFOLD* 87

2.8 The Intrinsic Definition of a Manifold*

It is somewhat restrictive to only consider manifolds that are embedded in
some Euclidean space. Although we shall see that (at least) every compact
manifold admits an embedding into a Euclidean space, such an embedding is
in many cases not a natural part of the structure of a manifold. In particular,
we encounter manifolds that are described as quotient spaces and there are
manifolds that are embedded in certain infinite-dimensional Hilbert spaces.
For this reason it is convenient, at this point, to introduce a more general
intrinisc definition of a manifold. (See Chapter 1 for an overview.) This
requires some background from point set topology that is not covered in
the first year analysis courses. We shall then see that all the definitions and
results of this chapter carry over in a natural manner to the intrinsic setting.
We begin by recalling the intrinsing definition of a smooth manifold in §1.4.

2.8.1 Definition and Examples

Uα Uβ

φα φβ
φβα

Figure 2.12: Coordinate charts and transition maps.

Definition 2.8.1 (Smooth m-Manifold). Let m ∈ N0 and M be a set.

A chart on M is a pair (φ, U ) where U ⊂ M and φ is a bijection from U
to an open set φ(U ) ⊂ Rm . Two charts (φ1 , U1 ), (φ2 , U2 ) are called com-
patible iff φ1 (U1 ∩ U2 ) and φ2 (U1 ∩ U2 ) are open and the transition map

φ21 = φ2 ◦ φ−1
1 : φ1 (U1 ∩ U2 ) → φ2 (U1 ∩ U2 ) (2.8.1)

is a diffeomorphism. A smooth atlas on M is a collection A of charts

on M any two of which are compatible and such that the sets U , as (φ, U )
ranges over A , cover M (i.e. for every p ∈ M there is a chart (φ, U ) ∈ A
with p ∈ U ). A maximal smooth atlas is an atlas which contains every
chart which is compatible with each of its members. A smooth m-manifold
is a pair consisting of a set M and a maximal atlas A on M .
88 CHAPTER 2. FOUNDATIONS

In Lemma 1.4.3 it was shown that, if A is an atlas, then so is the

collection A of all charts compatible with each member of A . Moreover,
the atlas A is maximal, so every atlas extends uniquely to a maximal atlas.
For this reason, a manifold is usually specified by giving its underlying set M
and some atlas on M . Generally, the notation for the atlas is suppressed and
the manifold is denoted simply by M . The members of the atlas are called
coordinate charts or simply charts on M . By Lemma 1.3.3 a smooth
m-manifold admits a unique topology such that, for each chart (φ, U ) of the
smooth atlas, the set
U ⊂M
is open and the bijection
φ : U → φ(U )
is a homeomorphism onto the open set φ(U ) ⊂ Rm . This topology is called
the intrinsic topology of M and is described in the following definition.
Definition 2.8.2. Let M be a smooth m-manifold. The intrinsic topology
on the set M is the topology induced by the charts, i.e. a subset

W ⊂M

is open in the intrinsic topology iff φ(U ∩ W ) is an open subset of Rm for

every chart (φ, U ) on M .
Remark 2.8.3. Let M ⊂ Rk be smooth m-dimensional submanifold of Rk
as in Definition 2.1.3. Then the set of all diffeomorphisms (φ, U ∩ M ) as
in Definition 2.1.3 form a smooth atlas as in Definition 2.8.1. The intrin-
sic topology on the resulting smooth manifold is the same as the relative
topology defined in §1.3.
Remark 2.8.4. A topological manifold is a topological space such that
each point has a neighborhood U homeomorphic to an open subset of Rm .
Thus a smooth manifold (with the intrinsic topology) is a topological man-
ifold and its maximal smooth atlas A is a subset of the set A0 of all
pairs (φ, U ) where U ⊂ M is an open set and φ is a homeomorphism from U
to an open subset of Rm . One says that the maximal smooth atlas A is a
smooth structure on the topological manifold M iff the topology of M is
the intrinsic topology of the smooth structure and every chart of the smooth
structure is a homeomorphism. As explained in §1.4 a topological manifold
can have many distinct smooth structures (see Remark 1.4.6). However, it
is a deep theorem beyond the scope of this book that there are topological
manifolds which do not admit any smooth structure.
2.8. THE INTRINSIC DEFINITION OF A MANIFOLD* 89

Example 2.8.5. The complex projective space CPn is the set

CPn = ` ⊂ Cn+1 | ` is a 1-dimensional complex subspace


of complex lines in Cn+1 . It can be identified with the quotient space

CPn = Cn+1 \ {0} /C∗

of nonzero vectors in Cn+1 modulo the action of the multiplicative group

C∗ = C \ {0} of nonzero complex numbers. The equivalence class of a
nonzero vector z = (z0 , . . . , zn ) ∈ Cn+1 will be denoted by
[z] = [z0 : z1 : · · · : zn ] := {λz | λ ∈ C∗ }
and the associated line is ` = Cz. An atlas on CPn is given by the open
cover Ui := {[z0 : · · · : zn ] | zi 6= 0} for i = 0, 1, . . . , n and the coordinate
charts φi : Ui → Cn are
 
z0 zi−1 zi+1 zn
φi ([z0 : · · · : zn ]) := ,..., , ,..., . (2.8.2)
zi zi zi zi
Exercise: Prove that each φi is a homeomorphism and the transition maps
are holomorphic. Prove that the manifold topology is the quotient topology,
i.e. if π : Cn+1 \ {0} → CPn denotes the obvious projection, then a sub-
set U ⊂ CPn is open if and only if π −1 (U ) is an open subset of Cn+1 \ {0}.
Example 2.8.6. The real projective space RPn is the set
RPn = ` ⊂ Rn+1 | ` is a 1-dimensional linear subspace


of real lines in Rn+1 . It can again be identified with the quotient space
RPn = Rn+1 \ {0} /R∗

of nonzero vectors in Rn+1 modulo the action of the multiplicative group

R∗ = R \ {0} of nonzero real numbers, and the equivalence class of a nonzero
vector x = (x0 , . . . , xn ) ∈ Rn+1 will be denoted by
[x] = [x0 : x1 : · · · : xn ] := {λx | λ ∈ R∗ } .
An atlas on RPn is given by the open cover
Ui := {[x0 : · · · : xn ] | xi 6= 0}
and the coordinate charts φi : Ui → Rn are again given by (2.8.2), with zj
replaced by xj . The arguments in Example 2.8.5 show that these coordinate
charts form an atlas and the manifold topology is the quotient topology. The
transition maps are real analytic diffeomorphisms.
90 CHAPTER 2. FOUNDATIONS

Example 2.8.7. The real n-torus is the topological space

Tn := Rn /Zn

equipped with the quotient topology. Thus two vectors x, y ∈ Rn are equiv-
alent iff their difference x − y ∈ Zn is an integer vector and we denote by
π : Rn → Tn the obvious projection which assigns to each vector x ∈ Rn its
equivalence class
π(x) := [x] := x + Zn .
Then a set U ⊂ Tn is open if and only if the set π −1 (U ) is an open subset
of Rn . An atlas on Tn is given by the open cover

Uα := {[x] | x ∈ Rn , |x − α| < 1/2} ,

parametrized by vectors α ∈ Rn , and the coordinate charts φα : Uα → Rn

defined by φα ([x]) := x for x ∈ Rn with |x − α| < 1/2. Exercise: Show
that each transition map for this atlas is a translation by an integer vector.

Example 2.8.8. Consider the complex Grassmannian

Gk (Cn ) := {V ⊂ Cn | v is a k-dimensional complex linear subspace} .

This set can again be described as a quotient space Gk (Cn ) ∼

= Fk (Cn )/U(k).
Here n o
Fk (Cn ) := D ∈ Cn×k | D∗ D = 1l

denotes the set of unitary k-frames in Cn and the group U(k) acts on Fk (Cn )
contravariantly by D 7→ Dg for g ∈ U(k). The projection

π : Fk (Cn ) → Gk (Cn )

sends a matrix D ∈ Fk (Cn ) to its image V := π(D) := im D. A subset

U ⊂ Gk (Cn ) is open if and only if π −1 (U ) is an open subset of Fk (Cn ). Given
a k-dimensional subspace V ⊂ Cn we can define an open set UV ⊂ Gk (Cn ) as
the set of all k-dimensional subspaces of Cn that can be represented as graphs
of linear maps from V to V ⊥ . This set of graphs can be identified with the
complex vector space HomC (V, V ⊥ ) of complex linear maps from V to V ⊥
and hence with C(n−k)×k . This leads to an atlas on Gk (Cn ) with holomorphic
transition maps and shows that Gk (Cn ) is a manifold of complex dimension
kn − k 2 . Exercise: Verify the details of this construction. Find explicit
formulas for the coordinate charts and their transition maps. Carry this
over to the real setting. Show that CPn and RPn are special cases.
2.8. THE INTRINSIC DEFINITION OF A MANIFOLD* 91

Example 2.8.9 (The real line with two zeros). A topological space M
is called Hausdorff iff any two points in M can be separated by disjoint
open neighborhoods. This example shows that a manifold need not be a
Hausdorff space. Consider the quotient space
M := R × {0, 1}/ ≡
where [x, 0] ≡ [x, 1] for x 6= 0. An atlas on M consists of two coordinate
charts φ0 : U0 → R and φ1 : U1 → R where
Ui := {[x, i] | x ∈ R} , φi ([x, i]) := x
for i = 0, 1. Thus M is a 1-manifold. But the topology on M is not
Hausdorff, because the points [0, 0] and [0, 1] cannot be separated by disjoint
open neighborhoods.
Example 2.8.10 (A 2-manifold without a countable atlas). Consider
the vector space X = R × R2 with the equivalence relation
either y1 = y2 6= 0, t1 + x1 y1 = t2 + x2 y2
[t1 , x1 , y1 ] ≡ [t2 , x2 , y2 ] ⇐⇒
or y1 = y2 = 0, t1 = t2 , x1 = x2 .
For y 6= 0 we have [0, x, y] ≡ [t, x − t/y, y], however, each point (x, 0) on
the x-axis gets replaced by the uncountable set R × {(x, 0)}. Our manifold
is the quotient space M := X/ ≡. This time we do not use the quotient
topology but the topology induced by our atlas (see Definition 2.8.2). The
coordinate charts are parametrized by the reals: for t ∈ R the set Ut ⊂ M
and the coordinate chart φt : Ut → R2 are given by
Ut := {[t, x, y] | x, y ∈ R} , φt ([t, x, y]) := (x, y).
A subset U ⊂ M is open, by definition, iff φt (U ∩ Ut ) is an open subset of R2
for every t ∈ R. With this topology each φt is a homeomorphism from Ut
onto R2 and M admits a countable dense subset S := {[0, x, y] | x, y ∈ Q}.
However, there is no atlas on M consisting of countably many charts. (Each
coordinate chart can contain at most countably many of the points [t, 0, 0].)
The function f : M → R given by f ([t, x, y]) := t + xy is smooth and each
point [t, 0, 0] is a critical point of f with value t. Thus f has no regular
value. Exercise: Show that M is a path-connected Hausdorff space.
In Theorem 2.9.12 we will show that smooth manifolds whose topology is
Hausdorff and second countable are precisely those that can be embedded in
Euclidean space. Most authors tacitly assume that manifolds are Hausdorff
and second countable and so will we after the end of the present chapter.
However before §2.9.1 there is no need to impose these hypotheses.
92 CHAPTER 2. FOUNDATIONS

2.8.2 Smooth Maps and Diffeomorphisms

Our next goal is to carry over all the definitions from embedded manifolds
in Euclidean space to the intrinsic setting.

Definition 2.8.11. Let

(M, {(φα , Uα )}α∈A ), (N, {(ψβ , Vβ )}β∈B )

be smooth manifolds. A map f : M → N is called smooth iff it is continu-

ous and the map

fβα := ψβ ◦ f ◦ φ−1
α : φα (Uα ∩ f
−1
(Vβ )) → ψβ (Vβ ) (2.8.3)

is smooth for every α ∈ A and every β ∈ B. It is called a diffeomorphism

iff it is bijective and f and f −1 are smooth. The manifolds M and N are
called diffeomorphic iff there exists a diffeomorphism f : M → N .

The reader may check that the notion of a smooth map is independent
of the atlas used in the definition, that compositions of smooth maps are
smooth, and that sums and products of smooth maps from M to R are
smooth.

Exercise 2.8.12. Let M be a smooth m-dimensional manifold with an atlas

A = {(φα , Uα )}α∈A .

Consider the quotient space

[
M
f := {α} × φα (Uα )/ ∼
α∈A

where
def −1
(α, x) ∼ (β, y) ⇐⇒ φ−1
α (x) = φβ (y).

for α, β ∈ A, x ∈ φα (Uα ), and y ∈ φβ (Uβ ). Define an atlas on M

f by

eα := {[α, x] | x ∈ φα (Uα )} ,
U φeα ([α, x]) := x.

Prove that M
f is a smooth m-manifold and that it is diffeomorphic to M .

Exercise 2.8.13. Prove that CP1 is diffeomorphic to S 2 . Hint: Stereo-

graphic projection.
2.8. THE INTRINSIC DEFINITION OF A MANIFOLD* 93

2.8.3 Tangent Spaces and Derivatives

In the situation where M is a submanifold of Euclidean space and p ∈ M we
have defined the tangent space of M at p as the set of all derivatives γ̇(0) of
smooth curves γ : R → M that pass through p = γ(0). We cannot do this
for manifolds in the intrinsic sense, as the derivative of a curve has yet to be
defined. In fact, the purpose of introducing a tangent space of M is precisely
to allow us to define what we mean by the derivative of a smooth map. There
are two approaches. One is to introduce an appropriate equivalence relation
on the set of curves through p and the other is to use local coordinates.
Definition 2.8.14. Let M be a manifold with an atlas A = {(φα , Uα )}α∈A
and let p ∈ M . Two smooth curves γ0 , γ1 : R → M with γ0 (0) = γ1 (0) = p
are called p-equivalent iff for some (and hence every) α ∈ A with p ∈ Uα
we have  
d  d 
φα (γ0 (t)) =  φα (γ1 (t)).
dt t=0 dt t=0
p
We write γ0 ∼ γ1 iff γ0 is p-equivalent to γ1 and denote the equivalence class
of a smooth curve γ : R → M with γ(0) = p by [γ]p . Every such equivalence
class is called a tangent vector of M at p. The tangent space of M at p
is the set of equivalence classes
Tp M := {[γ]p | γ : R → M is smooth and γ(0) = p} . (2.8.4)
Definition 2.8.15. Let M be a manifold with an atlas A = {(φα , Uα )}α∈A
and let p ∈ M . The A -tangent space of M at p is the quotient space
p
TpA M :=
[
{α} × Rm / ∼, (2.8.5)
p∈Uα

where the union runs over all α ∈ A with p ∈ Uα and

p
d φβ ◦ φ−1


(α, ξ) ∼ (β, η) ⇐⇒ α (x)ξ = η, x := φα (p).
The equivalence class is denoted by [α, ξ]p and is called a tangent vector.
In Definition 2.8.14 it is not immediately obvious that the set Tp M
in (2.8.4) is a vector space. However, the quotient space TpA M in (2.8.5) is
obviously a vector space of dimension m and there is a natural bijection
  
A d 
Tp M → Tp M : [γ]p 7→ α,  φα (γ(t)) . (2.8.6)
dt t=0 p

This bijection induces a vector space structure on the set Tp M . In other

words, the set Tp M in (2.8.4) admits a unique vector space structure such
that the map Tp M → TpA M in (2.8.6) is a vector space isomorphism.
94 CHAPTER 2. FOUNDATIONS

Exercise 2.8.16. Verify the phrase “and hence every” in Definition 2.8.14
and deduce that the map Tp M → TpA M in (2.8.6) is well defined. Show
that it is bijective.
From now on we will use either Definition 2.8.14 or Definition 2.8.15 or
both, whichever way is most convenient, and drop the superscript A .
Definition 2.8.17. For each smooth curve γ : R → M with γ(0) = p we
define the derivative γ̇(0) ∈ Tp M as the equivalence class
  
∼ d 
γ̇(0) := [γ]p = α,  φα (γ(t)) ∈ Tp M.
dt t=0 p

Definition 2.8.18. If f : M → N is a smooth map between two manifolds

(M, {(φα , Uα )}α∈A ) and (N, {(ψβ , Vβ )}β∈B ), we define the derivative
df (p) : Tp M → Tf (p) N
by the formula
df (p)[γ]p := [f ◦ γ]f (p) (2.8.7)
for each smooth curve γ : R → M with γ(0) = p. Here we use (2.8.4).
Under the isomorphism (2.8.6) this corresponds to the linear map
df (p)[α, ξ]p := [β, dfβα (x)ξ]f (p) , x := φα (p), (2.8.8)
for α ∈ A with p ∈ Uα and β ∈ B with f (p) ∈ Vβ , where fβα is given
by (2.8.3).
Remark 2.8.19. Think of N = Rn as a manifold with a single coordinate
chart, namely the identity map ψβ = id : Rn → Rn . For every q ∈ N = Rn
the tangent space Tq N is then canonically isomorphic to Rn via (2.8.5).
Thus for every smooth map f : M → Rn the derivative of f at p ∈ M is a
linear map df (p) : Tp M → Rn , and the formula (2.8.8) reads
df (p)[α, ξ]p = d(f ◦ φ−1
α )(x)ξ, x := φα (p).
This formula also applies to maps defined on some open subset of M . In
particular, with f = φα : Uα → Rm we have
dφα (p)[α, ξ]p = ξ.
Thus the map dφα (p) : Tp M → Rm is the canonical vector space isomor-
phism determined by α.
With these definitions the derivative of f at p is a linear map and we have
the chain rule for the composition of two smooth maps as in Theorem 2.2.14.
2.8. THE INTRINSIC DEFINITION OF A MANIFOLD* 95

2.8.4 Submanifolds and Embeddings

Definition 2.8.20. Let M be a smooth m-manifold and n ∈ {0, 1, . . . , m}.
A subset
N ⊂M

is called an n-dimensional submanifold of M if, for every element p ∈ N ,

there exists a local coordinate chart

φ:U →Ω

for M , defined on an an open neighborhood U ⊂ M of p and with values in

an open set Ω ⊂ Rn × Rm−n , such that

φ(U ∩ N ) = Ω ∩ (Rn × {0}) .

By Theorem 2.1.10 an m-manifold

M ⊂ Rk

in the sense of Definition 2.1.3 is a submanifold of Rk in the sense of Defi-

nition 2.8.20. By Theorem 2.3.4 the notion of a submanifold N ⊂ M of a
manifold M ⊂ Rk in Definition 2.3.1 agrees with the notion of a submanifold
in Definition 2.8.20.

Exercise 2.8.21. Let N be a submanifold of M . Show that if M is Haus-

dorff, so is N , and if M is paracompact, so is N .

Exercise 2.8.22. Let N be a submanifold of M and let P be a submanifold

of N . Prove that P is a submanifold of M . Hint: Use Theorem 2.1.10.

Exercise 2.8.23. Let N be a submanifold of M . Prove that there exists an

open set U ⊂ M such that N ⊂ U and N is closed in the relative topology
of U .

All the theorems we have proved for embedded manifolds and their proofs
carry over almost word for word to the present setting. For example we have
the inverse function theorem, the notion of a regular value, the implicit
function theorem, the notion of an immersion, the notion of an embedding
as a proper injective immersion, and the fact from Theorem 2.3.4 that a
subset P ⊂ M is a submanifold if and only if it is the image of an embedding.
96 CHAPTER 2. FOUNDATIONS

Example 2.8.24 (Veronese embedding). The map

CP2 → CP5 : [z0 : z1 : z2 ] 7→ [z02 : z12 : z22 : z1 z2 : z2 z0 : z0 z1 ]

is an embedding. (Exercise: Prove this.) It restricts to an embedding of

the real projective plane into RP5 and also gives rise to embeddings of RP2
into R4 as well as to the Roman surface: an immersion of RP2 into R3 . (See
Example 2.1.16.) There are similar embeddings
 
n+d
CPn → CPN −1 , N := ,
d

for all n and d, defined in terms of monomials of degree d in n + 1 variables.

These are the Veronese embeddings.

Example 2.8.25 (Plücker embedding). The Grassmannian G2 (R4 ) of

2-planes in R4 is a smooth 4-manifold and can be expressed as the quotient
of the space F2 (R4 ) of orthonormal 2-frames in R4 by the orthogonal group
O(2). (See Example 2.8.8.) Write an orthonormal 2-frame in R4 as a matrix
 
x0 y0
 x1 y1 
D=
 x2
, DT D = 1l.
y2 
x3 y3

Then the map f : G2 (R4 ) → RP5 , defined by

f ([D]) := [p01 : p02 : p03 : p23 : p31 : p12 ], pij := xi yj − xj yi ,

is an embedding and its image is the quadric

X := f (G2 (R4 )) = p ∈ RP5 | p01 p23 + p02 p31 + p03 p12 = 0 .



(Exercise: Prove this.) There are analogous embeddings

 
n N −1 n
f : Gk (R ) → RP , N := ,
k

for all k and n, defined in terms of the k × k-minors of the (orthonormal)

frames. These are the Plücker embeddings.
2.8. THE INTRINSIC DEFINITION OF A MANIFOLD* 97

2.8.5 Tangent Bundle and Vector Fields

Let M be a m-manifold with an atlas A = {(φα , Uα )}α∈A . The tangent
bundle of M is defined as the disjoint union of the tangent spaces, i.e.
[
T M := {p} × Tp M = {(p, v) | p ∈ M, v ∈ Tp M } .
p∈M

Denote by π : T M → M the projection given by π(p, v) := p. Recall the

notion of a submersion as a smooth map between smooth manifolds, whose
derivative is surjective at each point.

Lemma 2.8.26. The tangent bundle of M is a smooth 2m-manifold with

coordinate charts
eα := π −1 (Uα ) → φα (Uα ) × Rm ,
φeα : U φeα (p, v) := (φα (p), dφα (p)v) .

The projection π : T M → M is a surjective submersion If M is second

countable and Hausdorff, so is T M .

Proof. For each pair α, β ∈ A the set

eβ ) = φα (Uα ∩ Uβ ) × Rm
eα ∩ U
φeα (U

is open in Rm × Rm and the transition map

φeβα := φeβ ◦ φe−1

α : φα (Uα ∩ Uβ ) → φβ (Uα ∩ Uβ )
e e e e e e

is given by
φeβα (x, ξ) = (φβα (x), dφβα (x)ξ)
for x ∈ φα (Uα ∩ Uβ ) and ξ ∈ Rm where

φβα := φβ ◦ φ−1
α .

Thus the transition maps are all diffeomorphisms and so the coordinate
charts φeα define an atlas on T M . The topology on T M is determined by
this atlas via Definition 2.8.2. If M has a countable atlas, so does T M . The
remaining assertions are easy exercises.

Definition 2.8.27. Let M be a smooth m-manifold. A (smooth) vector

field on M is a collection of tangent vectors X(p) ∈ Tp M , one for each
point p ∈ M , such that the map M → T M : p 7→ (p, X(p)) is smooth. The
set of smooth vector fields on M will be denoted by Vect(M ).
98 CHAPTER 2. FOUNDATIONS

Associated to a vector field is a smooth map M → T M whose composi-

tion with the projection π : T M → M is the identity map on M . Strictly
speaking this map should be denoted by a symbol other than X, for exam-
ple by X.
e However, it is convenient at this point, and common practice,
to slightly abuse notation and denote the map from M to T M also by X.
Thus a vector field can be defined as a smooth map

X : M → TM

such that
π ◦ X = id : M → M.
Such a map is also called a section of the tangent bundle.
Now suppose A = {(φα , Uα )}α∈A is an atlas on M and X : M → T M
is a vector field on M . Then X determines a collection of smooth maps

Xα : φα (Uα ) → Rm

given by
Xα (x) := dφα (p)X(p), p := φ−1
α (x), (2.8.9)
for x ∈ φα (Uα ). We can think of each Xα as a vector field on the open set
φα (Uα ) ⊂ Rm , representing the vector field X on the coordinate patch Uα .
These local vector fields Xα satisfy the condition

Xβ (φβα (x)) = dφβα (x)Xα (x) (2.8.10)

for x ∈ φα (Uα ∩ Uβ ). This equation can also be expressed in the form

Xα |φα (Uα ∩Uβ ) = φ∗βα Xβ |φβ (Uα ∩Uβ ) . (2.8.11)

Conversely, any collection of smooth maps Xα : φα (Uα ) → Rm satisfy-

ing (2.8.10) determines a unique vectorfield X on M via (2.8.9). Thus we
can define the Lie bracket of two vector fields X, Y ∈ Vect(M ) by

[X, Y ]α (x) := [Xα , Yα ](x) = dXα (x)Yα (x) − dYα (x)Xα (x) (2.8.12)

for α ∈ A and x ∈ φα (Uα ). It follows from equation (2.4.18) in Lemma 2.4.21

that the local vector fields

[X, Y ]α : φα (Uα ) → Rm

satisfy (2.8.11) and hence determine a unique vector field [X, Y ] on M via

[X, Y ](p) := dφα (p)−1 [Xα , Yα ](φα (p)), p ∈ Uα . (2.8.13)

2.8. THE INTRINSIC DEFINITION OF A MANIFOLD* 99

Thus the Lie bracket of X and Y is defined on Uα as the pullback of the Lie
bracket of the vector fields Xα and Yα under the coordinate chart φα . With
this understood all the results in §2.4 about vector fields and flows along with
their proofs carry over word for word to the intrinsic setting whenever M is a
Hausdorff space. This includes the existence and uniquess result for integral
curves in Theorem 2.4.7, the concept of the flow of a vector field in Defini-
tion 2.4.8 and its properties in Theorem 2.4.9, the notion of completeness
of a vector field (that the integral curves exist for all time), and the various
properties of the Lie bracket such as the Jacobi identity (2.4.20), the formu-
las in Lemma 2.4.18, and the fact that the Lie bracket of two vector fields
vanishes if and only if the corresponding flows commute (see Lemma 2.4.26).
One can also carry over the notion of a subbundle E ⊂ T M of rank n to
the intrinsic setting by the condition that E is a smooth submanifold of T M
and intersects each fiber Tp M in an n-dimensional linear subspace

Ep := {v ∈ Tp M | (p, v) ∈ E} .

Then the characterization of subbundles in Theorem 2.6.10 and the theorem

of Frobenius 2.7.2 including their proofs also carry over to the intrinsic
setting whenever M is a Hausdorff space.

2.8.6 Coordinate Notation

Fix a coordinate chart φα : Uα → Rm on an m-manifold M . The components
of φα are smooth real valued functions on the open subset Uα of M and it
is customary to denote them by

x1 , . . . , xm : Uα → R.

The derivatives of these functions at p ∈ Uα are linear functionals

dxi (p) : Tp M → R, i = 1, . . . , m. (2.8.14)

They form a basis of the dual space

Tp∗ M := Hom(Tp M, R).

(A coordinate chart on M can in fact be characterized as an m-tuple of real

valued functions on an open subset of M whose derivatives are everywhere
linearly independent and which, taken together, form an injective map.)
The dual basis of Tp M will be denoted by
∂ ∂
1
(p), . . . , m (p) ∈ Tp M. (2.8.15)
∂x ∂x
100 CHAPTER 2. FOUNDATIONS

Thus 
∂ 1, if i = j,
dxi (p) (p) = δji :=
∂xj 0, if i 6= j,
for i, j = 1, . . . , m and ∂/∂xi is a vector field on the coordinate patch Uα .
For each p ∈ Uα it is the canonical basis of Tp M determined by φα . In the
notation of (2.8.5) and Remark 2.8.19 we have
∂
(p) = [α, ei ]p = dφα (p)−1 ei ,
∂xi
where ei = (0, . . . , 0, 1, 0, . . . , 0) (with 1 in the ith place) denotes the stan-
dard basis vector of Rm . In other words, for all ξ = (ξ 1 , . . . , ξ m ) ∈ Rm and
all p ∈ Uα , the tangent vector

v := dφα (p)−1 ξ ∈ Tp M

is given by
m
X ∂
v = [α, ξ]p = ξi (p). (2.8.16)
∂xi
i=1
Thus the restriction of a vector field X ∈ Vect(M ) to Uα has the form
m
X ∂
X|Uα = ξi ,
∂xi
i=1

where ξ 1 , . . . , ξ m : Uα → R are smooth real valued functions. If the map

Xα : φα (Uα ) → Rm

is defined by (2.8.9), then

Xα ◦ φ−1 1 m
α = (ξ , . . . , ξ ).

The above notation is motivated by the observation that the derivative of

a smooth function f : M → R in the direction of a vector field X on a
coordinate patch Uα is given by
m
X ∂f
LX f |Uα = ξi .
∂xi
i=1

Here the term ∂f /∂xi is understood as first writing f as a function of

x1 , . . . , xm , then taking the partial derivative, and afterwards expressing this
partial derivative again as a function of i
∂ −1

p. Thus ∂f /∂x is the shorthand
notation for the function ∂xi (f ◦ φα ) ◦ φα : Uα → R.
2.9. CONSEQUENCES OF PARACOMPACTNESS* 101

2.9 Consequences of Paracompactness*

In geometry it is often necessary to turn a construction in local coordinates
into a global geometric object. A key technical tool for such “local to global”
constructions is an existence theorem for partitions of unity.

2.9.1 Paracompactness
The existence of a countable atlas is of fundamental importance for almost
everything we will prove about manifolds. The next two remarks describe
several equivalent conditions.

Remark 2.9.1. Let M be a smooth manifold and denote by

U ⊂ 2M

the topology induced by the atlas as in Definition 2.8.2. Then the following
are equivalent.
(a) M admits a countable atlas.
(b) M is σ-compact, i.e. there is a sequence of compact
S∞ subsets Ki ⊂ M
such that Ki ⊂ int(Ki+1 ) for every i ∈ N and M = i=1 Ki .
(c) Every open cover of M has a countable subcover.
(d) M is second countable, i.e. there is a countable collection of open sets
B ⊂ U such that every open set U ∈ U is a union of open sets from the
collection B. (B is then called a countable base for the topology of M .)
That (a) =⇒ (b) =⇒ (c) =⇒ (a) and (a) =⇒ (d) follows directly from the
definitions. The proof that (d) implies (a) requires the construction of a
countable refinement and the axiom of choice. (A refinement of an open
cover {Ui }i∈I is an open cover {Vj }j∈J such that each set Vj is contained in
one of the sets Ui .)

Remark 2.9.2. Let M and U be as in Remark 2.9.1 and suppose in ad-

dition that M is a connected Hausdorff space. Then the existence of a
countable atlas is also equivalent to each of the following conditions.
(e) M is metrizable, i.e. there is a distance function d : M × M → [0, ∞)
such that U is the topology induced by d.
(f ) M is paracompact, i.e. every open cover of M has a locally finite
refinement. (An open cover {Vj }j∈J is called locally finite iff every p ∈ M
has a neighborhood that intersects only finitely many Vj .)
102 CHAPTER 2. FOUNDATIONS

That (a) implies (e) follows from the Urysohn Metrization Theorem
which asserts (in its original form) that every normal second countable topo-
logical space is metrizable [38, Theorem 34.1]. A topological space M is
called normal iff points are closed and, for any two disjoint closed sets
A, B ⊂ M , there are disjoint open sets U, V ⊂ M such that A ⊂ U and
B ⊂ V . It is called regular iff points are closed and, for every closed set
A ⊂ M and every b ∈ M \ A, there are disjoint open sets U, V ⊂ M such
that A ⊂ U and b ∈ V . It is called locally compact if, for every open
set U ⊂ M and every p ∈ U , there is a compact neighborhood of p con-
tained in U . It is easy to show that every manifold is locally compact and
every locally compact Hausdorff space is regular. Tychonoff ’s Lemma
asserts that a regular topological space with a countable base is normal [38,
Theorem 32.1]. Hence it follows from the Urysohn Metrization Theorem
that every Hausdorff manifold with a countable base is metrizable. That (e)
implies (f) follows from a more general theorem which asserts that every
metric space is paracompact (see [38, Theorem 41.4] and [45]). Conversely,
the Smirnov Metrization Theorem asserts that a paracompact Haus-
dorff space is metrizable if and only if it is locally metrizable, i.e. every
point has a metrizable neighborhood (see [38, Theorem 42.1]). Since ev-
ery manifold is locally metrizable this shows that (f) implies (e). Thus we
have (a) =⇒ (e) ⇐⇒ (f) for every Hausdorff manifold.
The proof that (f) implies (a) does not require the Hausdorff property
but we do need the assumption that M is connected. (A manifold with
uncountably many connected components, each of which is paracompact, is
itself paracompact but does not admit a countable atlas.) Here is a sketch.
If M is a paracompact manifold, then there is a locally finite open cover
{Uα }α∈A by coordinate charts. Since each set Uα has a countable dense
subset, the set {α ∈ A | Uα ∩ Uα0 6= ∅} is at most countable for each α0 ∈ A.
Since M is connected we can reach each point from Uα0 through a finite
sequence of sets Uα1 , . . . , Uα` with Uαi−1 ∩ Uαi 6= ∅. This implies that the
index set A is countable and hence M admits a countable atlas.

Remark 2.9.3. A Riemann surface is a 1-dimensional complex manifold

(i.e. the coordinate charts take values in C and the transition maps are
holomorphic) with a Hausdorff topology. It is a deep theorem in the theory
of Riemann surfaces that every connected Riemann surface is necessarily
second countable (see [2]). Thus pathological examples of the type discussed
in Example 2.8.10 cannot be constructed with holomorphic transition maps.
2.9. CONSEQUENCES OF PARACOMPACTNESS* 103

Exercise 2.9.4. Prove that every manifold is locally compact. Find an ex-
ample of a manifold M and a point p0 ∈ M such that every closed neighbor-
hood of p0 is non-compact. Hint: The example is necessarly non-Hausdorff.

Exercise 2.9.5. Prove that a manifold M admits a countable atlas if and

only if it is σ-compact if and only if every open cover of M has a countable
subcover if and only if it is second countable. Hint: The topology of Rm is
second countable and every open subset of Rm is σ-compact.

Exercise 2.9.6. Prove that every submanifold M ⊂ Rk (Definition 2.1.3)

is second countable.

Exercise 2.9.7. Prove that every connected component of a manifold M is

an open subset of M and is path-connected.

2.9.2 Partitions of Unity

Definition 2.9.8. Let M be a smooth manifold. A partition of unity on
M is a collection of smooth functions

θα : M → [0, 1], α ∈ A,

such that each point p ∈ M has an open neighborhood V ⊂ M on which only

finitely many θα do not vanish, i.e.

# {α ∈ A | θα |V 6≡ 0} < ∞, (2.9.1)

and, for every p ∈ M , we have

X
θα (p) = 1. (2.9.2)
α∈A

If {Uα }α∈A is an open cover of M , then a partition of unity {θα }α∈A (indexed
by the same set A) is called subordinate to the cover iff each θα is
supported in Uα , i.e.

supp(θα ) := {p ∈ M | θα (p) 6= 0} ⊂ Uα .

Theorem 2.9.9 (Partitions of unity). Let M be a smooth manifold whose

topology is paracompact and Hausdorff. Then, for every open cover of M ,
there exists a partition of unity subordinate to that cover.

Proof. See page 105.

104 CHAPTER 2. FOUNDATIONS

Lemma 2.9.10. Let M be a smooth manifold with a Hausdorff topology.

Then, for every open set V ⊂ M and every compact set K ⊂ V , there exists
a smooth function κ : M → [0, ∞) with compact support such that κ(p) > 0
for every p ∈ K and supp(κ) ⊂ V .

Proof. Assume first that K = {p0 } is a single point. Since M is a mani-

fold it is locally compact. Hence there is a compact neighborhood C ⊂ V
of p0 . Since M is Hausdorff C is closed and hence the set U := int(C)
is a neighborhood of p0 whose closure U ⊂ C is compact and contained
in V . Shrinking U , if necessary, we may assume that there is a coordinate
chart φ : U → Ω with values in some open neighborhood Ω ⊂ Rm of the
origin such that φ(p0 ) = 0. (Here m is the dimension of M .) Now choose a
smooth function κ0 : Ω → [0, ∞) with compact support such that κ0 (0) > 0.
Then the function κ : M → [0, 1], defined by κ|U := κ0 ◦ φ and κ(p) := 0
for p ∈ M \ U is supported in V and satisfies κ(p0 ) > 0. This proves the
lemma in the case where K is a point.
Now let K be any compact subset of V . Then, by the first part of
the proof, there is a collection of smooth functions κp : M → [0, ∞),
one for every p ∈ K, such that κp (p) > 0 and supp(κp ) ⊂ V . Since K
is compact
 there are finitely
many points p1 , . . . , pk ∈ K P
such that the
sets p ∈ M | κpj (p) > 0 cover K. Hence the function κ := j κpj is sup-
ported in V and is everywhere positive on K. This proves Lemma 2.9.10.

Lemma 2.9.11. Let M be a topological space. If {Vi }i∈I is a locally finite

collection of open sets in M , then
[ [
Vi = Vi
i∈I0 i∈I0

for every subset I0 ⊂ I.

S S
Proof. The set i∈I0 V i is obviously contained in the closure of i∈I0 Vi .
S
To prove the converse choose a point p0 ∈ M \ i∈I0 V i . Since the collec-
tion {Vi }i∈I is locally finite, there exists an open neighborhood U of p0 such
that the set I1 := {i ∈ I | Vi ∩ U 6= ∅} is finite. Hence the set
[
U0 := U \ Vi
i∈I0 ∩I1

is an open neighborhood
S of p0 and we have U0 ∩ Vi = ∅ for every i ∈ I0 .
Hence p0 ∈
/ i∈I0 Vi . This proves Lemma 2.9.11.
2.9. CONSEQUENCES OF PARACOMPACTNESS* 105

Proof of Theorem 2.9.9. Let {Uα }α∈A be an open cover of M . We prove in

four steps that there is a partition of unity subordinate to this cover. The
proofs of steps one and two are taken from [38, Lemma 41.6].
Step 1. There is a locally finite open cover {Vi }i∈I of M such that, for
every i ∈ I, the closure V i is compact and contained in one of the sets Uα .
Denote by V ⊂ 2M the set of all open sets V ⊂ M such that V is compact
and V ⊂ Uα for some α ∈ A. Since M is a locally compact Hausdorff
space the collection V is an open cover of M . (If p ∈ M , then there is an
α ∈ A such that p ∈ Uα ; since M is locally compact there is a compact
neighborhood K ⊂ Uα of p; since M is Hausdorff K is closed and thus
V := int(K) is an open neighborhood of p with V ⊂ K ⊂ Uα .) Since M is
paracompact the open cover V has a locally finite refinement {Vi }i∈I . This
cover satisfies the requirements of Step 1.
is a collection of compact sets Ki ⊂ Vi , one for each i ∈ I,
Step 2. There S
such that M = i∈I Ki .
Denote by W ⊂ 2M the set of all open sets W ⊂ M such that W ⊂ Vi for
some i. Since M is a locally compact Hausdorff space, the collection W is an
open cover of M . Since M is paracompact W has a locally finite refinement
{Wj }j∈J . By the axiom of choice there is a map

J → I : j 7→ ij

such that
W j ⊂ V ij ∀ j ∈ J.

Since the collection {Wj }j∈J is locally finite, we have

[ [
Ki := Wj = W j ⊂ Vi
ij =i ij =i

by Lemma 2.9.11. Since V i is compact so is Ki .

Step 3. There is a partition of unity subordinate to the cover {Vi }i∈I .
Choose a collection of compact sets Ki ⊂ Vi for i ∈ I as in Step 2. Then,
by Lemma 2.9.10 and the axiom of choice, there is a collection of smooth
functions κi : M → [0, ∞) with compact support such that

supp(κi ) ⊂ Vi , κi |Ki > 0 ∀ i ∈ I.

106 CHAPTER 2. FOUNDATIONS

Since the cover {Vi }i∈I is locally finite the sum

X
κ := κi : M → R
i∈I

is locally finite (i.e. each point in M has a neighborhood in which only

finitely many terms do not vanish) and thus defines a smooth function on M .
This function is everywhere positive, because each summand is nonnegative
and, for each p ∈ M , there is an i ∈ I with p ∈ Ki so that κi (p) > 0. Thus
the funtions χi := κi /κ define a partition of unity satisfying supp(χi ) ⊂ Vi
for every i ∈ I as required.

Step 4. There is a partition of unity subordinate to the cover {Uα }α∈A .

Let {χi }i∈I be the partition of unity constructed in Step 3. By the axiom
of choice there is a map I → A : i 7→ αi such that Vi ⊂ Uαi for every i ∈ I.
For α ∈ A define θα : M → [0, 1] by
X
θα := χi .
αi =α

Here the sum runs over all indices i ∈ I with αi = α. This sum is locally
finite and hence is a smooth function on M . Moreover, each point in M has
an open neighborhood in which only finitely many of the θα do not vanish.
Hence the sum of the θα is a well defined function on M and
X X X X
θα = χi = χi ≡ 1.
α∈A α∈A αi =α i∈I

This shows that the functions θα form a partition of unity. To prove the
inclusion supp(θα ) ⊂ Uα we consider the open sets

Wi := {p ∈ M | χi (p) > 0}

for i ∈ I. Since Wi ⊂ Vi this collection is locally finite. Hence, by

Lemma 2.9.11, we have
[ [ [ [
supp(θα ) = Wi = Wi = supp(χi ) ⊂ V i ⊂ Uα .
αi =α αi =α αi =α αi =α

This proves Theorem 2.9.9.

2.9. CONSEQUENCES OF PARACOMPACTNESS* 107

2.9.3 Embedding in Euclidean Space

Theorem 2.9.12. Let M be a second countable smooth m-manifold with a
Hausdorff topology. Then there exists an embedding f : M → R2m+1 with a
closed image.

Proof. The proof has five steps.

Step 1. Let U ⊂ M be an open set and let K ⊂ U be a compact set.
Then there exists an integer k ∈ N, a smooth map f : M → Rk , and an
open set V ⊂ M , such that K ⊂ V ⊂ U , the restriction f |V : V → Rk is an
injective immersion, and f (p) = 0 for all p ∈ M \ U .
Choose a smooth atlas A = {(φα , Uα )}α∈A on M such that, for each α ∈ A,
either Uα ⊂ U or Uα ∩ K = ∅. Since M is a paracompact Hausdorff man-
ifold, Theorem 2.9.9 asserts that there exists a partition of unity {θα }α∈A
subordinate to the open cover {Uα }α∈A of M . Since the sets Uα with Uα ⊂ U
form an open cover of K and K is a compact subset of M , there exist finitely
many indices α1 , . . . , α` ∈ A such that

K ⊂ Uα1 ∪ · · · ∪ Uα` =: V ⊂ U.

Let k := `(m + 1) and, for i = 1, . . . , `, abbreviate

φi := φαi , θi := θαi .

Define the smooth map f : M → Rk by

 
θ1 (p)
 θ1 (p)φ1 (p) 
 
f (p) := 
 .. 
for p ∈ M.
 . 

 θ` (p) 
θ` (p)φ` (p)

Then the restriction f |V : V → Rk is injective. Namely, if p0 , p1 ∈ V satisfy

f (p0 ) = f (p1 ),

then    
I := i  θi (p0 ) > 0 = i  θi (p1 ) > 0 6= ∅
and, for i ∈ I, we have θi (p0 ) = θi (p1 ), hence φi (p0 ) = φi (p1 ), and so p0 = p1 .
Moreover, for every p ∈ K the derivative df (p) : Tp M → Rk is injective, and
this proves Step 1.
108 CHAPTER 2. FOUNDATIONS

Step 2. Let f : M → Rk be an injective immersion and let A ⊂ R(2m+1)×k

be a nonempty open set. Then there exists a matrix A ∈ A such that the
map Af : M → R2m+1 is an injective immersion.
The proof of Step 2 uses the Theorem of Sard (see [1, 37]). The sets
 
W0 := (p, q) ∈ M × M  p 6= q ,
 
W1 := (p, v) ∈ T M  v 6= 0

are open subsets of smooth second countable Hausdorff 2m-manifolds and

the maps

F0 : A × W0 → R2m+1 , F1 : A × W1 → R2m+1 ,

defined by

F0 (A, p, q) := A(f (p) − f (q)), F1 (A, p, v) := Adf (p)v

for A ∈ A, (p, q) ∈ W0 , and (p, v) ∈ W1 , are smooth. Moreover, the zero

vector in R2m+1 is a regular value of F0 because f is injective and of F1
because f is an immersion. Hence it follows from the intrinsic analogue of
Theorem 2.2.19 that the sets
M0 := F0−1 (0) = (A, p, q) ∈ A × W0  Af (p) = Af (q) ,
 

M1 := F −1 (0) = (A, p, v) ∈ A × W1  Adf (p)v = 0

 
1

are smooth manifolds of dimension

dim M0 = dim M1 = (2m + 1)k − 1.

Since M is a second countable Hausdorff manifold, so are M0 and M1 .

Hence the Theorem of Sard asserts that the canonical projections

M0 → A : (A, p, q) 7→ A =: π0 (A, p, q),

M1 → A : (A, p, v) 7→ A =: π1 (A, p, v),

have a common regular value A ∈ A. Since

dim M0 = dim M1 < dim A,

this implies
A ∈ A \ (π0 (M0 ) ∪ π1 (M1 )) .
Hence Af : M → R2m+1 is an injective immersion and this proves Step 2.
2.9. CONSEQUENCES OF PARACOMPACTNESS* 109

If M is compact, the result follows from Steps 1 and 2 with K = U = M .

In the noncompact case the proof requires two more steps to construct an
embedding into R4m+4 and a further step to reduce the dimension to 2m + 1.
Step 3. Assume M is not compact. Then there exists a sequence of open
sets Ui ⊂ M , a sequence of smooth functions ρi : M → [0, 1], and a sequence
of compact sets Ki ⊂ Ui such that

supp(ρi ) ⊂ Ui , Ki = ρ−1
i (1) ⊂ Ui , Ui ∩ Uj = ∅
S∞
for all i, j ∈ N with |i − j| ≥ 2 and M = i=1 Ki .

Since M is second countable, there exists a sequenceSof compact sets Ci ⊂ M

such that Ci ⊂ int(Ci+1 ) for all i ∈ N and M = i∈N Ci (Remark 2.9.1).
Define the compact sets Bi ⊂ M by C0 := ∅ and

Bi := Ci \ Ci−1 for i ∈ N.
S
Then M = i∈N Bi and, for all i, j ∈ N with j ≥ i + 2, we have

Bi ⊂ Ci ⊂ int(Cj−1 ), Bj ⊂ Cj \ int(Cj−1 )

and so Bi ∩ Bj = ∅. Since M is metrizable by Remark 2.9.2, there exists

a distance function d : M × M → [0, ∞) that induces the intrinsic topology
on M . Define
[
Ai := Bj , εi := d(Ai , Bi ) = inf d(p, q).
p∈Ai ,q∈Bi
j∈N\{i−1,i,i+1}

Then Ai is a closed subset of M , because any convergent sequence in M must

belong to a finite union of the Bj . Since Ai ∩ Bi = ∅, this implies εi > 0.
For i ∈ N define the set Ui ⊂ M by
 
Ui := p ∈ M  there exists a q ∈ Bi with d(p, q) < εi /3 .

Then {Ui }i∈N is a sequence of open subsets of M such that Bi ⊂ Ui ⊂ Ci+1

for all i ∈ N and Ui ∩ Uj = ∅ for |i − j| ≥ 2. In particular, each set Ui has a
compact closure.
For each i there exists of a partition of unity subordinate to the open
cover M = Ui ∪ (M \ Bi ) and hence a smooth function ρi : M → [0, 1] such
that supp(ρi ) ⊂ Ui and ρi |Bi ≡ 1. Define Ki := ρ−1
i (1) = {p ∈ Ui | ρi (p) = 1}
for i ∈ N. Then
S Ki is a compact set and Bi ⊂ Ki ⊂ Ui for each i ∈ N.
Hence M = i∈N Ki and this proves Step 3.
110 CHAPTER 2. FOUNDATIONS

Step 4. Assume M is not compact. Then there exists an embedding

f : M → R4m+4

with a closed image and a pair of orthonormal vectors x, y ∈ R4m+4 such

that, for every ε > 0, there exists a compact set K ⊂ M with
 
 f (p) 
sup inf   − sx − ty  < ε. (2.9.3)
p∈M \K s,t∈R |f (p)|

Assume M is not compact and let Ki , Ui , ρi be as in Step 3. Then, by

Steps 1 and 2, there exists a sequence of smooth maps gi : M → R2m+1 such
that gi |M \Ui ≡ 0, the restriction gi |Ki : Ki → R2m+1 is injective, and the
derivative dgi (p) : Tp M → R2m+1 is injective for all p ∈ Ki and all i ∈ N.
Let ξ ∈ R2m+1 be a unit vector and define the maps fi : M → R2m+1 by
 
gi (p)
fi (p) := ρi (p) iξ + q  (2.9.4)
1 + |gi (p)|2

for p ∈ M and i ∈ N. Then the restriction fi |Ki : Ki → R2m+1 is injective,

the derivative dfi (p) : Tp M → R2m+1 is injective for all p ∈ Ki , and

supp(fi ) ⊂ Ui , fi (Ki ) ⊂ B1 (iξ), fi (M ) ⊂ Bi+1 (0).

Define the maps f odd , f ev : M → R2m+1 and ρodd , ρev : M → R by


odd ρ2i−1 (p), if i ∈ N andSp ∈ U2i−1 ,
ρ (p) :=
0, if p ∈ M \ i∈N U2i−1 ,

f2i−1 (p), if i ∈ N andSp ∈ U2i−1 ,
f odd (p) :=
0, if p ∈ M \ i∈N U2i−1 ,

ρ2i (p), if i ∈ N andSp ∈ U2i ,
ρev (p) :=
0, if p ∈ M \ i∈N U2i ,

f2i (p), if i ∈ N andSp ∈ U2i ,
f ev (p) :=
0, if p ∈ M \ i∈N U2i ,

and define the map f : M → R4m+4 by

 
f (p) := ρodd (p), f odd (p), ρev (p), f ev (p)

for p ∈ M .
2.9. CONSEQUENCES OF PARACOMPACTNESS* 111

We prove that f is injective. To see this, note that

 
2i − 2 < f odd (p) < 2i,

p ∈ K2i−1 =⇒
|f ev (p)| < 2i + 1,
(2.9.5)
− 1 < |f ev (p)| < 2i + 1,

p ∈ K2i =⇒ 2i odd
f (p) < 2i + 2,

Now let p0 , p1 ∈ M such that f (p0 ) = f (p1 ). Assume first that p0 ∈ K2i−1 .
odd odd
S
Then ρ (p1 ) = ρ (p0 ) = 1 and hence p1 ∈ j∈N K2j−1 . By (2.9.5), we
also have 2i − 2 < |f odd (p1 )| = |f odd (p0 )| < 2i and hence p1 ∈ K2i−1 . This
implies f2i−1 (p1 ) = f odd (p1 ) = f odd (p0 ) = f2i−1 (p0 ) and so p0 =S p1 . Now
assume p0 ∈ K2i . Then ρev (p1 ) = ρev (p0 ) = 1 and hence p1 ∈ j∈N K2j .
By (2.9.5), we also have 2i − 1 < |f ev (p1 )| = |f ev (p0 )| < 2i + 1, so p1 ∈ K2i ,
which implies f2i (p1 ) = f ev (p1 ) = f ev (p0 ) = f2i (p0 ), and so again p0 = p1 .
This shows that f is injective. That f is an immersion follows from the fact
that the derivative dfi (p) is injective for all p ∈ Ki and all i ∈ N.
We prove that f is proper and has a closed image. Let (pν )ν∈N be a
sequence in M such that the sequence (f (pν ))ν∈N in R4m+4 is bounded.
Choose i ∈S N such that |f odd (pν )| < 2i and |f ev (pν )| < 2i + 1 for all ν ∈ N.
Then pν ∈ 2i j=1 Kj for all ν ∈ N by (2.9.5). Hence (pν )ν∈N has a convergent
subsequence. Thus f : M → R4m+4 is an embedding with a closed image.
Next consider the pair of orthonormal vectors
x := (0, ξ, 0, 0), y := (0, 0, 0, ξ)
in R4m+4 = R × R2m+1 × R × R2m+1 . Let (pν )ν∈N be a sequence in M
that does not have a convergent subsequence and choose a sequence iν ∈ N
such that pν ∈ K2iν −1 ∪ K2iν for all ν ∈ N. Then iν tends to infinity.
If pν ∈ K2iν −1 for all ν, then we have lim supν→∞ |f odd (pν )|−1 |f ev (pν )| ≤ 1
by (2.9.5). Passing to a subsequence, still denoted by (pν )ν∈N , we may
assume that the limit λ := limν→∞ |f odd (fν )|−1 |f ev (pν )| exists. Then
 odd 
f (pν ) 1 |f ev (pν )| λ
0 ≤ λ ≤ 1, lim =√ , lim =√ ,
ν→∞ |f (pν )| 1 + λ2 ν→∞ |f (pν )| 1 + λ2
and it follows from (2.9.4) that
f odd (pν ) f ev (pν )
lim = ξ, lim = λξ.
ν→∞ |f odd (pν )| ν→∞ |f odd (pν )|

This implies
 
f (pν ) ξ λξ 1 λ
lim = 0, √ , 0, √ =√ x+ √ y.
ν→∞ |f (pν )| 1+λ 2 1 + λ2 1+λ2 1 + λ2
112 CHAPTER 2. FOUNDATIONS

Similarly, if pν ∈ K2iν for all ν, there exists a subsequence such that the
limit λ := limν→∞ |f ev (pν )|−1 |f odd (pν )| exists and, by (2.9.4), this implies
 
f (pν ) λξ ξ λ 1
lim = 0, √ , 0, √ =√ x+ √ y.
ν→∞ |f (pν )| 1 + λ2 1 + λ2 1 + λ2 1 + λ2
This shows that the vectors x and y satisfy the requirements of Step 4.
Step 5. There exists an embedding f : M → R2m+1 with a closed image.
For compact manifolds the result was proved in Steps 1 and 2 and for m = 0
the assertion is obvious, because then M is a finite or countable set with
the discrete topology. Thus assume that M is not compact and m ≥ 1.
Choose f : M → R4m+4 and x, y ∈ R4m+4 as in Step 4 and define
(  )
the vectors Ax and Ay

A := A ∈ R(2m+1)×(4m+4)  .

 are linearly independent

Since m ≥ 1, this is a nonempty open subset of R(2m+1)×(4m+4) . We prove

that the map Af : M → R2m+1 is proper and has a closed image for ev-
ery A ∈ A. To see this, fix a matrix A ∈ A. Let (pν )ν∈N be a sequence in M
that does not have a convergent subsequence. Then by Step 4 there exists a
subsequence, still denoted by (pν )ν∈N , and real numbers s, t ∈ R such that
f (pν )
s2 + t2 = 1, lim = sx + ty, lim |f (pν )| = ∞.
ν→∞ |f (pν )| ν→∞

This implies
Af (pν )
lim = sAx + tAy 6= 0
ν→∞ |f (pν )|

and hence limν→∞ |Af (pν )| = ∞. Thus the preimage of every compact sub-
set of R2m+1 under the map Af : M → R2m+1 is a compact subset of M ,
and hence Af is proper and has a closed image (Remark 2.3.3).
Now it follows from Step 2 that there exists a matrix A ∈ A such that the
map Af : M → R2m+1 is an injective immersion. Hence it is an embedding
with a closed image. This proves Step 5 and Theorem 2.9.12.

The Whitney Embedding Theorem asserts that every second count-

able Hausdorff m-manifold M admits an embedding f : M → R2m . The
proof is based on the Whitney Trick and goes beyond the scope of this
book. The next exercise shows that Whitney’s theorem is sharp.
Remark 2.9.13. The manifold RP2 cannot be embedded into R3 . The same
is true for the Klein bottle K := R2 / ≡ where the equivalence relation is
given by [x, y] ≡ [x + k, ` − y] for x, y ∈ R and k, ` ∈ Z.
2.9. CONSEQUENCES OF PARACOMPACTNESS* 113

2.9.4 Leaves of a Foliation

Let M be an m-dimensional paracompact Hausdorff manifold and E ⊂ T M
be an integrable subbundle of rank n. Let L ⊂ M be a closed leaf of the
foliation determined by E. Then L is a smooth n-dimensional submanifold
of M . Here is a sketch of David Epstein’s proof of this fact in [13].
(a) The space L with the intrinsic topology admits the structure of a mani-
fold such that the obvious inclusion ι : L → M is an injective immersion.
This is an easy exercise. For the definition of the intrinsic topology see
Exercise 2.7.10. The dimension of L is n.
(b) If f : X → Y is a continuous map between topological spaces such that Y
is paracompact and there is an open cover {Vj }j∈J of Y such that f −1 (Vj )
is paracompact for each j, then X is paracompact. To see this, we may
assume that the cover {Vj }j∈J is locally finite. Now let {Uα }α∈A be an open
cover of X. Then the sets Uα ∩ f −1 (Vj ) define an open cover of f −1 (Vj ).
Choose a locally finite refinement {Wij }i∈Ij of this cover. Then the open
cover {Wij }j∈J, i∈Ij of M is a locally finite refinement of {Uα }α∈A .
(c) The intrinsic topology of L is paracompact. This follows from (b) and
the fact that the intersection of L with every foliation box is paracompact
in the intrinsic topology.
(d) The intrinsic topology of L is second countable. This follows from (a)
and (c) and the fact that every connected paracompact manifold is second
countable (see Remark 2.9.2).
(e) The intersection of L with a foliation box consists of at most countably
many connected components. This follows immediately from (d).
(f ) If L is a closed subset of M , then the intersection of L with a foliation
box has only finitely many connected components. To see this, we choose
a transverse slice of the foliation at p0 ∈ L, i.e. a connected submanifold
T ⊂ M through p0 , diffeomorphic to an open ball in Rm−n , whose tangent
space at each point p ∈ T is a complement of Ep . By (d) we have that T ∩ L
is at most countable. If this set is not finite, even after shrinking T , there
must be a sequence pi ∈ (T ∩L)\{p0 } converging to p0 . Using the holonomy
of the leaf (obtained by transporting transverse slices along a curve via a
lifting argument) we find that every point p ∈ T ∩ L is the limit point of a
sequence in (T ∩ L) \ {p}. Hence the one-point set {p} has empty interior in
the relative topology of T ∩ L for each p ∈ T ∩ L. Thus T ∩ L is a countable
union of closed subsets with empty interior. Since T ∩L admits the structure
of a complete metric space, this contradicts the Baire category theorem.
(g) It follows immediately from (f) that L is a submanifold of M .
114 CHAPTER 2. FOUNDATIONS

2.9.5 Principal Bundles

An interesting class of foliations arises from smooth Lie group actions.
Let G ⊂ GL(N, R) be a compact Lie group and let P be a smooth m-
manifold whose topology is Hausdorff and second countable. A smooth
(contravariant) G-action on P is a smooth map
P × G → P : (p, g) 7→ pg (2.9.6)
that satisfies the conditions
(pg)h = p(gh), p1l = p (2.9.7)
for all p ∈ P and all g, h ∈ G. Fix any such group action. Then every
group element g ∈ G determines a diffeomorphism P → P : p 7→ pg, whose
derivative at p ∈ P is denoted by Tp P → Tpg P : v 7→ vg. Every Lie alge-
bra element ξ ∈ g := Lie(G) = T1l G determines a vector field Xξ ∈ Vect(P )
which assigns to each p ∈ P the tangent vector

d 
Xξ (p) := pξ :=  p exp(tξ) ∈ Tp P. (2.9.8)
dt t=0
The linear map g → Vect(P ) : ξ 7→ Xξ is called the infinitesimal action.
It is a Lie algebra anti-homomorphism because the group action is contra-
variant. (Exercise: Prove that [Xξ , Xη ] = −X[ξ,η] for ξ, η ∈ g.) The group
action (2.9.6) is said to be with finite isotropy iff the isotropy subgroup
Gp := {g ∈ G | pg = p}
is finite for all p ∈ P . The isotropy subgroup Gp is a Lie subgroup of G
with Lie algebra gp := Lie(Gp ) = {ξ ∈ g | Xξ (p) = 0}. Since G is compact,
this shows that Gp is a finite subgroup of G if and only if gp = {0}
or, equivalently, the map g → Tp P : ξ 7→ Xξ (p) = pξ is injective. Thus, in
the finite isotropy case, the group action determines an involutive subbun-
dle E ⊂ T P with the fibers Ep := pg = {Xξ (p) | ξ ∈ g} for p ∈ P . When G
is connected, the leaves of the corresponding foliation are the group or-
bits pG := {pg | g ∈ G}. These are the elements of the orbit space
P/G := {pG | p ∈ P } .
There is a natural projection π : P → P/G defined by π(p) := pG for p ∈ P
and the orbit space P/G is equipped with the quotient topology (a sub-
set U ⊂ P/G is open if and only if π −1 (U ) is an open subset of P ). The
group action is called free iff Gp = {1l} for all p ∈ P . The next theorem
shows that, in the case of a free action, the quotient space admits a unique
smooth structure such that the projection π : P → P/G is a submersion.
2.9. CONSEQUENCES OF PARACOMPACTNESS* 115

Theorem 2.9.14 (Principal Bundle). Let P be a smooth m-manifold

whose topology is Hausdorff and second countable. Suppose P is equipped
with a smooth contravariant action of a compact Lie group G and assume
the group action is free. Then dim(G) ≤ m and B := P/G admits a unique
smooth structure such that the projection π : P → B is a submersion. The
intrinsic topology of B, induced by the smooth structure, agrees with the
quotient topology, and it is Hausdorff and second countable.

Proof. For each p ∈ P the map G → P : g 7→ pg is an embedding and this

implies k := dim(G) ≤ dim(P ) = m. Define n := m − k. A local slice of
the group action is a smooth map ι : Ω → P , defined on an open set Ω ⊂ Rn ,
such that the map Ω × G → P : (x, g) 7→ ι(x)g is an embedding. With this
understood, we prove the assertions in five steps.
Step 1. For every p0 ∈ P there exists a local slice ι0 : Ω0 → P , defined on
an open neighborhood Ω0 ⊂ Rn of the origin, such that ι0 (0) = p0 .
Choose a coordinate chart φ : V → Rm on an open neighborhood V ⊂ P
of p0 such that φ(p0 ) = 0 and φ(V ) = Rm . Define v1 , . . . , vm ∈ Tp0 P by

dφ(p0 )vi := ei for i = 1, . . . , m,

where e1 , . . . , em is the standard basis of Rm . Reorder the coordinates

on Rm , if necessary, such that the vectors v1 , . . . , vn project to a basis of the
quotient space Tp0 P/p0 g. Define ι : Rn → P by

ι(x1 , . . . , xn ) := φ−1 (x1 , . . . , xn , 0, . . . , 0)

and define the map ψ : Rn × G → P by

ψ(x, g) := ι(x)g for x ∈ Rn and g ∈ G.

Then ψ is smooth and its derivative dψ(0, 1l) : Rn × g → Tp0 P is given by

n
X
dψ(0, 1l)(b
x, ξ) = x
bi vi + p0 ξ
i=1

for xb = (b bn ) ∈ Rn and ξ ∈ g. Hence dψ(0, 1l) is bijective and so it

x1 , . . . , x
follows from the Inverse Function Theorem 2.2.17 that there exist open
neighborhoods Ω ⊂ Rn of 0, Ω1 ⊂ G of 1l, and W ⊂ P of p0 such that the
restricted map
ψ1 := ψ|Ω×Ω1 : Ω × Ω1 → W
is a diffeomorphism.
116 CHAPTER 2. FOUNDATIONS

Next we prove that there exists an open neigborhood Ω0 ⊂ Ω of the

origin such that the restricted map

ψ0 := ψ|Ω0 ×G : Ω0 × G → P

is injective. Suppose otherwise that no such neighborhood Ω0 exists. Then

there exist sequences (xi , gi ), (x0i , gi0 ) ∈ Ω × G such that (xi , gi ) 6= (x0i , gi0 )
and ψ(xi , gi ) = ψ(x0i , gi0 ) for all i and the sequences (xi )i∈N and (x0i )i∈N in Ω
converge to the origin. Since G is compact we may assume, by passing to
a subsequence if necessary, that the sequences (gi )i∈N and (gi0 )i∈N converge.
Denote the limits by

g := lim gi ∈ G, g 0 := lim gi0 ∈ G.

i→∞ i→∞

Then
p0 g = lim ι(xi )gi = lim ι(x0i )gi0 = p0 g 0
i→∞ i→∞

and so g = because the group action is free. Thus the sequence (gi0 gi−1 )i∈N
g0
in G converges to 1l and hence belongs to the set Ω1 for i sufficiently large.
Since
ψ1 (xi , 1l) = ι(xi ) = ι(x0i )gi0 gi−1 = ψ1 (x0i , gi0 gi−1 )
for all i, this contradicts the injectivity of ψ1 . Thus we have proved that
the map ψ0 : Ω0 × G → P is injective for a suitable neighborhood Ω0 ⊂ Ω
of the origin. That it is an immersion is a direct consequence of the formula

g g −1 ) g = dψ0 (x, 1l)(b

x, gbg −1 ) g



dψ0 (x, g)(b
x, gb) = dι(x)b
x + ι(x)(b

for all x ∈ Ω0 , x b ∈ Rn , g ∈ G, and gb ∈ Tg G, and the fact that the deriva-

tive dψ0 (x, 1l) is bijective for all x ∈ Ω0 (even for all x ∈ Ω).
Thus we have proved that ψ0 : Ω0 × G → P is an injective immersion.
Shrinking Ω0 further, if necessary, we may assume that Ω0 has a compact
closure and that ψ is injective on Ω0 × G. This implies that ψ0 is proper.
Namely, if (xi , gi )i∈N is a sequence in Ω0 × G and (x, g) ∈ Ω0 × G such
that ψ0 (x, g) = limi→∞ ψ0 (xi , gi ), then there is a subsequence (xiν , giν )ν∈N
that converges to a pair (x0 , g 0 ) ∈ Ω0 × G. This subsequence satisfies

ψ(x0 , g 0 ) = lim ψ0 (xiν , giν ) = ψ(x, g).

ν→∞

Since ψ is injective on Ω0 × G, this implies x = x0 and g = g 0 . Thus

every subsequence of (xi , gi )i∈N has a further subsequence that converges
to (x, g) and so the sequence (xi , gi )i∈N itself converges to (x, g). Thus the
map ψ0 : Ω0 × G → P is a proper injective immersion and this proves Step 1.
2.9. CONSEQUENCES OF PARACOMPACTNESS* 117

Step 2. Let ι : Ω → P be a local slice. Then the set U := π(ι(Ω)) ⊂ B is

open in the quotient topology and the map π ◦ ι : Ω → U is a homeomorphism
with respect to the quotient topology on U .
The map ψ : Ω × G → P , defined by ψ(x, g) := ι(x)g for x ∈ Ω and g ∈ G,
is an embedding. Hence W := ψ(Ω × G) is an open G-invariant subset
of P and ψ : Ω × G → W is a G-equivariant homeomorphism. Moreover,
for every element p ∈ P , we have π(p) ∈ U if and only if there exists
an element x ∈ Ω and an element g ∈ G such that p = ι(x)g = ψ(x, g).
Thus π −1 (U ) = ψ(Ω × G) = W is an open subset of P , and so U is an open
subset of B = P/G with respect to the quotient topology. The continuity
of π ◦ ι : Ω → U follows directly from the definition. Moreover, if Ω0 ⊂ Ω is
an open set and U 0 := π(ι(Ω0 )), then π −1 (U 0 ) = ψ(Ω0 × G) is open by the
same argument, and so U 0 ⊂ B is open with respect to the quotient topology.
Thus π ◦ ι : Ω → U is a homeomorphism and this proves Step 2.
Step 3. By Step 1 there exists a collection ια : Ωα → P , α ∈ A, of local
slices such that the sets Uα := π(ια (Ωα )) cover the orbit space B = P/G.
For α ∈ A define
φα := (π ◦ ια )−1 : Uα → Ωα .
Then A = {(φα , Uα )}α∈A is a smooth structure on B which renders the
canonical projection π : P → B into a submersion. Moreover, this smooth
structure is compatible with the quotient topology on B.
For α, β ∈ A define Ωαβ := φα (Uα ∩ Uβ ) and φβα := φβ ◦ φ−1 α : Ωαβ → Ωβα .
We must prove that φβα is smooth. To see this, define ψα : Ωα × G → P
by ψα (x, g) := ια (x)g for α ∈ A, x ∈ Ωα , and g ∈ G. Then ψα is a diffeo-
morphism onto its image and ψα (Ωαβ × G) = ψβ (Ωβα × G) = π −1 (Uα ∩ Uβ ).
For x ∈ Ωαβ the element φβα (x) ∈ Ωβα is the projection of ψβ−1 ◦ ψα (x, 1l)
onto the first factor. Thus φβα is smooth and so is its inverse φαβ . This
shows that {(Uα , φα )}α∈A is a smooth structure on B. Second, π is a sub-
mersion with respect to this smooth structure, because φα ◦ π ◦ ψα (x, g) = x
for all α ∈ A, all x ∈ Ωα , and all g ∈ G. Third, this smooth structure is com-
patible with the quotient topology by Step 2. This proves Step 3.
Step 4. There is only one smooth structure on B with respect to which the
projection π : P → B is a submersion.
Fix any smooth structure on B for which the projection π : P → B is a
submersion. Then the dimension of B is n = dim(P ) − dim(G), and so the
smooth structure consists of bijections φα : Uα → Ωα from subsets Uα ⊂ B
onto open sets Ωα ⊂ Rn such that the sets Uα cover B and the transition
maps are diffeomorphisms between open subsets of Rn .
118 CHAPTER 2. FOUNDATIONS

We prove that the intrinsic topology on B agrees with the quotient topol-
ogy. To see this, fix a subset U ⊂ B. Then the following are equivalent.
(a) U is open with respect to the intrinsic topology on B.
(b) φα (U ∩ Uα ) is open in Rn for all α ∈ A.
(c) π −1 (U ∩ Uα ) is open in P for all α ∈ A.
(d) π −1 (U ) is open in P .
(e) U is open with respect to the quotient topology on B.
The equivalence of (a) and (b) follows from the definition of the intrinsic
topology. That (b) implies (c) follows from the three observations that the
set π −1 (Uα ) is open in P , the map φα ◦ π : π −1 (Uα ) → Ωα is continuous,
and (φα ◦ π)−1 (φα (U ∩ Uα )) = π −1 (U ∩ Uα ). That (c) implies (b) follows
from the fact that the map φα ◦ π : π −1 (Uα ) → Ωα is a submersion and hence
maps the open set π −1 (U ∩ Uα ) onto an open subset of Ωα (Corollary 2.6.2).
The equivalence of (c) and (d) follows from the fact that the map π : P → B
is continuous and Uα ⊂ B is open (both with respect to the intrinsic topology
on B) and so π −1 (Uα ) is open in P for all α ∈ A. The equivalence of (d)
and (e) follows from the definition of the quotient topology on B.
Now let ι : Ω → P be a local slice and define the set U := π(ι(Ω)) ⊂ B
and the map φ := (π ◦ ι)−1 : U → Ω. Then the composition

φα ◦ φ−1 = φα ◦ π ◦ ι : φ(U ∩ Uα ) → φα (U ∩ Uα )

is a homeomorphism between open subsets of Rn . Moreover, φα ◦ φ−1 is the

composition of the smooth maps ι : {x ∈ Ω | π(ι(x)) ∈ Uα } → π −1 (U ∩ Uα ),
π : π −1 (U ∩ Uα ) → U ∩ Uα , and φα : U ∩ Uα → φα (U ∩ Uα ). So φα ◦ φ−1 is
smooth and its derivative is everywhere bijective because π is a submersion
and the kernel of dπ(ι(x)) is transverse to the image of dι(x). Thus φα ◦ φ−1
is a diffeomorphism by the Inverse Function Theorem and this proves Step 4.
Step 5. The quotient topology on B is a Hausdorff and second countable.
Let ια : Ωα → P for α ∈ A be a collection of local slices such that the
sets Uα := π(ια (Ωα )) cover B. Then the open sets π −1 (Uα ) form an open
cover of P and so there is a countable subcover. Thus B is second count-
able. To prove that B is Hausdorff, fix two distinct elements b0 , b1 ∈ B and
choose p0 , p1 ∈ P such that π(p0 ) = b0 and π(p1 ) = b1 . Then p0 G and p1 G
are disjoint compact subsets of P and hence can be separated by disjoint
open subsets U0 , U1 ⊂ P , because P is a Hausdorff space. Now for i = 0, 1
the set Vi := {p ∈ P | pG ⊂ Ui } is open (exercise) and contains the orbit pi G.
Hence W0 := π(V0 ) and W1 := π(V1 ) are disjoint open subsets of B such
that b0 ∈ W0 and b1 ∈ W1 . This proves Step 5 and Theorem 2.9.14.
2.9. CONSEQUENCES OF PARACOMPACTNESS* 119

Example 2.9.15. There are many important examples of free group actions
and principal bundles. A class of examples arises from orthonormal frame
bundles (§3.4). The complex projective space B = CPn arises from the
action of the circle G = S 1 on the unit sphere P = S 2n+1 ⊂ Cn+1 (Exam-
ple 2.8.5). The real projective space B = RPn arises from the action of the
finite group G = Z/2Z on the unit sphere P = S n ⊂ Rn+1 (Example 2.8.6).
The complex Grassmannian B = Gk (Cn ) arises from the action of G = U(k)
on the space P = Fk (Cn ) of unitary k-frames in Cn (Example 3.7.6). If G is
a Lie group and K ⊂ G is a compact subgroup, then by Theorem 2.9.14 the
homogeneous space G/K admits a unique smooth structure such that the
projection π : G → G/K is a submersion. The example SL(2, C)/SU(2) can
be identified with hyperbolic 3-space (§6.4.4), the example U(n)/O(n) can
be identified with the space of Lagrangian subspaces of a symplectic vector
space [36, Lemma 2.3.2], the example Sp(2n)/U(n) can be identified with
Siegel upper half space or the space of compatible linear complex structures
on a symplectic vector space [36, Lemma 2.5.12], and the example G2 /SO(4)
can be identified with the associative Grassmannian [51, Remark 8.4]. The
last three examples go beyond the scope of the present book.

Standing Assumption
We have seen that all the results in the present chapter carry over to the
intrinsic setting, assuming that the topology of M is Hausdorff and paracom-
pact. In fact, in many cases it is enough to assume the Hausdorff property.
However, these results mainly deal with introducing the basic concepts like
smooth maps, embeddings, submersions, vector fields, flows, and verifying
their elementary properties, i.e. with setting up the language for differen-
tial geometry and topology. When it comes to the substance of the subject
we shall deal with Riemannian metrics and they only exist on paracompact
Hausdorff manifolds. Another central ingredient in differential topology is
the theorem of Sard and that requires second countability. To quote Moe
Hirsch [22]: “Manifolds that are not paracompact are amusing, but they
never occur naturally and it is difficult to prove anything about them.”
Thus we will set the following convention for the remaining chapters.
We assume from now on that each intrinsic manifold M
is Hausdorff and second countable and hence is also paracompact.
For most of this text we will in fact continue to develop the theory for
submanifolds of Euclidean space and indicate, wherever necessary, how to
extend the definitions, theorems, and proofs to the intrinsic setting.
120 CHAPTER 2. FOUNDATIONS
Chapter 3

The Levi-Civita Connection

For a submanifold of Euclidean space the inner product on the ambient space
determines an inner product on each tangent space, the first fundamental
form. The second fundamental form is obtained by differentiating the map
which assigns to each point in M ⊂ Rn the orthogonal projection onto the
tangent space (§3.1). The covariant derivative of a vector field along a curve
is the orthogonal projection of the derivative in the ambient space onto the
tangent space (§3.2). We will show how the covariant derivative gives rise
to parallel transport (§3.3), examine the frame bundle (§3.4), discuss mo-
tions without “sliding, twisting, and wobbling”, and prove the development
theorem (§3.5).
In §3.6 we will see that the covariant derivative is determined by the
Christoffel symbols in local coordinates and thus carries over to the in-
trinsic setting. The intrinsic setting of Riemannian manifolds is explained
in §3.7. The covariant derivative takes the form of a family of linear op-
erators ∇ : Vect(γ) → Vect(γ), one for every smooth curve γ : I → M , and
these operators are uniquely characterized by the axioms of Theorem 3.7.8.
This family of linear operators is the Levi-Civita connection.

3.1 Second Fundamental Form

Let M ⊂ Rn be a smooth m-manifold. Then each tangent space of M is an
m-dimensional real vector space and hence is isomorphic to Rm . Thus any
two tangent spaces Tp M and Tq M are of course isomorphic to each other.
While there is no canonical isomorphism from Tp M to Tq M we shall see
that every smooth curve γ in M connecting p to q induces an isomorphism
between the tangent spaces via parallel transport of tangent vectors along γ.

121
122 CHAPTER 3. THE LEVI-CIVITA CONNECTION

Throughout we use the standard inner product on Rn given by

hv, wi = v1 w1 + v2 w2 + · · · + vn wn

for v = (v1 , . . . , vn ) ∈ Rn and w = (w1 , . . . , wn ) ∈ Rn . The associated Eu-

clidean norm will be denoted by
p q
|v| = hv, vi = v12 + v22 + · · · + vn2

for v = (v1 , . . . , vn ) ∈ Rn . When M ⊂ Rn is a smooth m-dimensional sub-

manifold, a first observation is that each tangent space of M inherits an inner
product from the ambient space Rn . The resulting field of inner products is
called the first fundamental form.
Definition 3.1.1. Let M ⊂ Rn be a smooth m-dimensional submanifold.
The first fundamental form on M is the field which assigns to each p ∈ M
the bilinear map
gp : Tp M × Tp M → R
defined by
gp (v, w) = hv, wi (3.1.1)
for v, w ∈ Tp M .
A second observation is that the inner product on the ambient space also
determines an orthogonal projection of Rn onto the tangent space Tp M for
each p ∈ M . This projection can be represented by the matrix Π(p) ∈ Rn×n
which is uniquely determined by the conditions

Π(p) = Π(p)2 = Π(p)T , (3.1.2)

and
Π(p)v = v ⇐⇒ v ∈ Tp M (3.1.3)
for p ∈ M and v ∈ Rn (see Exercise 2.6.9).
Lemma 3.1.2. The map Π : M → Rn×n defined by (3.1.2) and (3.1.3) is
smooth.
Proof. This follows directly from Theorem 2.6.10 and Corollary 2.6.12. More
explicitly, if U ⊂ M is an open set and φ : U → Ω is a coordinate chart onto
an open subset Ω ⊂ Rm with the inverse ψ := φ−1 : Ω → U , then
 −1
Π(p) = dψ(φ(p)) dψ(φ(p))T dψ(φ(p)) dψ(φ(p))T

for p ∈ U and this proves Lemma 3.1.2.

3.1. SECOND FUNDAMENTAL FORM 123

ν( p)
TpM

Figure 3.1: A unit normal vector field.

Example 3.1.3 (Gauß Map). Let M ⊂ Rm+1 be a submanifold of codi-

mension one. Then T M ⊥ is a vector bundle of rank one (Corollary 2.6.13),
and so each fiber Tp M ⊥ is spanned by a unit vector ν(p) ∈ Rm , determined
by Tp M up to a sign. By Theorem 2.6.10 each p0 ∈ M has an open neigh-
borhood U ⊂ M on which there exists a smooth map

ν : U → Rm+1

satisfying
ν(p) ⊥ Tp M, |ν(p)| = 1 (3.1.4)
for all p ∈ U (see Figure 3.1). Such a map ν is called a Gauß map. The
function Π : M → Rn×n is in this case given by

Π(p) = 1l − ν(p)ν(p)T (3.1.5)

for p ∈ U .

Example 3.1.4. Let M = S 2 ⊂ R3 . Then ν(p) = p and so

1 − x2
 
−xy −xz
Π(p) = 1l − ppT =  −yx 1 − y 2 −yz 
−zx −zy 1 − z 2

for p = (x, y, z) ∈ S 2 .

Example 3.1.5 (Möbius Strip). Consider the submanifold

 
 x = (1 + r cos(θ/2)) cos(θ),

 
3
M := (x, y, z) ∈ R  y = (1 + r cos(θ/2)) sin(θ),
z = r sin(θ/2), r, θ ∈ R, |r| < ε
  

for ε > 0 sufficiently small. Show that there does not exist a global smooth
function ν : M → R3 satisfying (3.1.4).
124 CHAPTER 3. THE LEVI-CIVITA CONNECTION

Example 3.1.6. Let U ⊂ Rn be an open set and f : U → Rn−m be a smooth

function such that 0 ∈ Rn−m is a regular value of f and U ∩ M = f −1 (0).
Then Tp M = ker df (p) and
 −1
Π(p) = 1l − df (p)T df (p)df (p)T df (p)
for every p ∈ U ∩ M .
Example 3.1.7. Let Ω ⊂ Rm be an open set and ψ : Ω → M be a smooth
embedding. Then Tψ(x) M = im dψ(x) and
 −1
Π(ψ(x)) = dψ(x) dψ(x)T dψ(x) dψ(x)T
for every x ∈ Ω.
Next we differentiate the map Π : M → Rn×n in Lemma 3.1.2. The
derivative at p ∈ M takes the form of a linear map
dΠ(p) : Tp M → Rn×n
which, as usual, is defined by

d 
dΠ(p)v :=  Π(γ(t)) ∈ Rn×n
dt t=0
for v ∈ Tp M , where γ : R → M is chosen such that γ(0) = p and γ̇(0) = v
(see Definition 2.2.13). We emphasize that the expression dΠ(p)v is a matrix
and can therefore be multiplied by a vector in Rn .
Lemma 3.1.8. For all p ∈ M and v, w ∈ Tp M we have
dΠ(p)v w = dΠ(p)w v ∈ Tp M ⊥ .

Proof. Choose a smooth path γ : R → M and a vector field X : R → Rn

along γ such that
γ(0) = p, γ̇(0) = v, X(0) = w.
For example, we can choose X(t) := Π(γ(t))w. Then
X(t) = Π(γ(t))X(t)
for every t ∈ R. Differentiate this equation to obtain


Ẋ(t) = Π(γ(t))Ẋ(t) + dΠ(γ(t))γ̇(t) X(t). (3.1.6)
Hence
dΠ(γ(t))γ̇(t) X(t) = 1l − Π(γ(t)) Ẋ(t) ∈ Tγ(t) M ⊥



(3.1.7)
for every t ∈ R and, with t = 0, we obtain (dΠ(p)v) w ∈ Tp M ⊥ .
3.1. SECOND FUNDAMENTAL FORM 125

Now choose a smooth map

R2 → M : (s, t) 7→ γ(s, t)

satisfying
∂γ ∂γ
γ(0, 0) = p, (0, 0) = v, (0, 0) = w,
∂s ∂t
(for example by doing this in local coordinates) and denote

∂γ ∂γ
X(s, t) := (s, t) ∈ Tγ(s,t) M, Y (s, t) := (s, t) ∈ Tγ(s,t) M.
∂s ∂t
Then
∂Y ∂2γ ∂X
= =
∂s ∂s∂t ∂t
and hence, using (3.1.7), we obtain
   
∂γ ∂γ ∂γ
dΠ(γ) = dΠ(γ) X
∂t ∂s ∂t

∂X
= 1l − Π(γ)
∂t

∂Y
= 1l − Π(γ)
 ∂s
∂γ
= dΠ(γ) Y
∂s
 
∂γ ∂γ
= dΠ(γ) .
∂s ∂t

With s = t = 0 we obtain

dΠ(p)w v = dΠ(p)v w ∈ Tp M ⊥

and this proves Lemma 3.1.8.

Definition 3.1.9. The collection of symmetric bilinear maps

hp : Tp M × Tp M → Tp M ⊥ ,

defined by
hp (v, w) := (dΠ(p)v)w = (dΠ(p)w)v (3.1.8)
for p ∈ M and v, w ∈ Tp M is called the second fundamental form on M .
126 CHAPTER 3. THE LEVI-CIVITA CONNECTION

Example 3.1.10. Let M ⊂ Rm+1 be an m-manifold and ν : M → S m be

a Gauß map so that Tp M = ν(p)⊥ for every p ∈ M (see Example 3.1.3).
Then Π(p) = 1l − ν(p)ν(p)T and hence

hp (v, w) = −ν(p)hdν(p)v, wi

for p ∈ M and v, w ∈ Tp M .

Exercise 3.1.11. Choose a splitting Rn = Rm × Rn−m and write the ele-

ments of Rn as tuples (x, y) = (x1 , . . . , xm , y1 , . . . , yn−m ) Let M ⊂ Rn be a
smooth m-dimensional submanifold such that p = 0 ∈ M and

T0 M = Rm × {0}, T0 M ⊥ = {0} × Rn−m .

By the implicit function theorem, there are open neighborhoods Ω ⊂ Rm

and V ⊂ Rn−m of zero and a smooth map f : Ω → V such that

M ∩ (Ω × V ) = graph(f ) = {(x, f (x)) | x ∈ Ω} .

Thus f (0) = 0 and df (0) = 0. Prove that the second fundamental form
hp : Tp M × Tp M → Tp M ⊥ is given by the second derivatives of f , i.e.
 
m
X ∂2f
hp (v, w) = 0, (0)vi wj 
∂xi ∂xj
i,j=1

for v, w ∈ Tp M = Rm × {0}.

Exercise 3.1.12. Let M ⊂ Rn be an m-manifold. Fix a point p ∈ M and

a unit tangent vector v ∈ Tp M so that |v| = 1 and define

L := {p + tv + w | t ∈ R, w ⊥ Tp M } .

Let γ : (−ε, ε) → M ∩ L be a smooth curve such that γ(0) = p, γ̇(0) = v,

and |γ̇(t)| = 1 for all t. Prove that

γ̈(0) = hp (v, v).

Draw a picture of M and L in the case n = 3 and m = 2.

3.2. COVARIANT DERIVATIVE 127

X(t)
γ(t)
M

Figure 3.2: A vector field along a curve.

3.2 Covariant Derivative

Definition 3.2.1. Let I ⊂ R be an open interval and let γ : I → M be a
smooth curve. A vector field along γ is a smooth map X : I → Rn such
that X(t) ∈ Tγ(t) M for every t ∈ I (see Figure 3.2). The set of smooth vector
fields along γ is a real vector space and will be denoted by
Vect(γ) := X : I → Rn | X is smooth and X(t) ∈ Tγ(t) M ∀ t ∈ I .


The first derivative Ẋ(t) of a vector field along γ at t ∈ I will, in general,

not be tangent to M . We may decompose it as a sum of a tangent vector
and a normal vector in the form


Ẋ(t) = Π(γ(t))Ẋ(t) + 1l − Π(γ(t)) Ẋ(t),
where Π : M → Rn×n is defined by (3.1.2) and (3.1.3). The tangential
component of this decomposition plays an important geometric role. It is
called the covariant derivative of X at t.
Definition 3.2.2 (Covariant Derivative). Let I ⊂ R be an open inter-
val, let γ : I → M be a smooth curve, and let X ∈ Vect(γ). The covariant
derivative of X is the vector field ∇X ∈ Vect(γ), defined by
∇X(t) := Π(γ(t))Ẋ(t) ∈ Tγ(t) M (3.2.1)
for t ∈ I.
Lemma 3.2.3 (Gauß–Weingarten Formula). The derivative of a vector
field X along a curve γ is given by
Ẋ(t) = ∇X(t) + hγ(t) (γ̇(t), X(t)). (3.2.2)
Here the first summand is tangent to M and the second summand is orthog-
onal to the tangent space of M at γ(t).
Proof. This is equation (3.1.6) in the proof of Lemma 3.1.8.
128 CHAPTER 3. THE LEVI-CIVITA CONNECTION

It follows directly from the definition that the covariant derivative along
a curve γ : I → M is a linear operator ∇ : Vect(γ) → Vect(γ). The following
lemma summarizes the basic properties of this operator.
Lemma 3.2.4 (Covariant Derivative). The covariant derivative satisfies
the following axioms for any two open intervals I, J ⊂ R.
(i) Let γ : I → M be a smooth curve, let λ : I → R be a smooth function,
and let X ∈ Vect(γ). Then
∇(λX) = λ̇X + λ∇X. (3.2.3)

(ii) Let γ : I → M be a smooth curve, let σ : J → I be a smooth function

and let X ∈ Vect(γ). Then
∇(X ◦ σ) = σ̇(∇X ◦ σ). (3.2.4)

(iii) Let γ : I → M be a smooth curve and let X, Y ∈ Vect(γ). Then

d
hX, Y i = h∇X, Y i + hX, ∇Y i. (3.2.5)
dt
(iv) Let γ : I × J → M be a smooth map, denote by ∇s the covariant deriva-
tive along the curve s 7→ γ(s, t) (with t fixed), and denote by ∇t the covariant
derivative along the curve t 7→ γ(s, t) (with s fixed). Then
∇s ∂t γ = ∇t ∂s γ. (3.2.6)
d
Proof. Part (i) follows from the Leibniz rule dt (λX) = λ̇X + λẊ and (ii)
d
follows from the chain rule dt (X ◦ σ) = σ̇(Ẋ ◦ σ). To prove part (iii), use
the orthogonal projections Π(γ(t)) : Rn → Tγ(t) M to obtain
d
hX, Y i = hẊ, Y i + hX, Ẏ i
dt
= hẊ, Π(γ)Y i + hΠ(γ)X, Ẏ i
= hΠ(γ)Ẋ, Y i + hX, Π(γ)Ẏ i
= h∇X, Y i + hX, ∇Y i
Part (iv) holds because the second derivatives commute and this proves
Lemma 3.2.4.
Part (i) in Lemma 3.2.4 asserts that the operator ∇ is what is called a
connection, part (iii) asserts that it is compatible with the first fundamental
form, and part (iv) asserts that it is torsion-free. Theorem 3.7.8 below as-
serts that these conditions (together with an extended chain rule) determine
the covariant derivative uniquely.
3.3. PARALLEL TRANSPORT 129

3.3 Parallel Transport

Definition 3.3.1 (Parallel Vector Field). Let I ⊂ R be an interval and
let γ : I → M be a smooth curve. A vector field X along γ is called parallel
if
∇X(t) = 0
for all t ∈ I.
Example 3.3.2. Assume m = k so that M ⊂ Rm is an open set. Then a
vector field along a smooth curve γ : I → M is a smooth map X : I → Rm .
Its covariant derivative is equal to the ordinary derivative ∇X(t) = Ẋ(t)
and hence X is is parallel if and only if it is constant.
Remark 3.3.3. For every X ∈ Vect(γ) and every t ∈ I we have

∇X(t) = 0 ⇐⇒ Ẋ(t) ⊥ Tγ(t) M.

In particular, γ̇ is a vector field along γ and ∇γ̇(t) = Π(γ(t))γ̈(t). Hence γ̇

is a parallel vector field along γ if and only if γ̈(t) ⊥ Tγ(t) M for all t ∈ I.
We will return to this observation in Chapter 4.
In general, a vector field X along a smooth curve γ : I → M is parallel
if and only if Ẋ(t) is orthogonal to Tγ(t) M for every t and, by the Gauß–
Weingarten formula (3.2.2), we have

∇X = 0 ⇐⇒ Ẋ = hγ (γ̇, X).

The next theorem shows that any given tangent vector v0 ∈ Tγ(t0 ) M extends
uniquely to a parallel vector field along γ.
Theorem 3.3.4 (Existence and Uniqueness). Let I ⊂ R be an interval
and γ : I → M be a smooth curve. Let t0 ∈ I and v0 ∈ Tγ(t0 ) M be given.
Then there is a unique parallel vector field X ∈ Vect(γ) such that X(t0 ) = v0 .
Proof. Choose a basis e1 , . . . , em of the tangent space Tγ(t0 ) M and let

X1 , . . . , Xm ∈ Vect(γ)

be vector fields along γ such that

Xi (t0 ) = ei , i = 1, . . . , m.

(For example choose Xi (t) := Π(γ(t))ei .) Then the vectors Xi (t0 ) are lin-
early independent. Since linear independence is an open condition there is a
130 CHAPTER 3. THE LEVI-CIVITA CONNECTION

constant ε > 0 such that the vectors X1 (t), . . . , Xm (t) ∈ Tγ(t) M are linearly
independent for every t ∈ I0 := (t0 − ε, t0 + ε) ∩ I. Since Tγ(t) M is an
m-dimensional real vector space this implies that the vectors Xi (t) form a
basis of Tγ(t) M for every t ∈ I0 . We express the vector ∇Xi (t) ∈ Tγ(t) M in
this basis and denote the coefficients by aki (t) so that
m
X
∇Xi (t) = aki (t)Xk (t).
k=1

The resulting functions aki : I0 → R are smooth. Likewise, if X : I → Rn is

any vector field along γ, then there are smooth functions ξ i : I0 → R such
that
Xm
X(t) = ξ i (t)Xi (t) for all t ∈ I0 .
i=1

The derivative of X is given by

m 
X 
Ẋ(t) = ξ˙i (t)Xi (t) + ξ i (t)Ẋi (t)
i=1

and the covariant derivative by

m 
X 
∇X(t) = ξ˙i (t)Xi (t) + ξ i (t)∇Xi (t)
i=1
m
X m
X m
X
= ξ˙i (t)Xi (t) + i
ξ (t) aki (t)Xk (t)
i=1 i=1 k=1
m m
!
X X
= ξ˙k (t) + k i
ai (t)ξ (t) Xk (t)
k=1 i=1

for t ∈ I0 . Hence ∇X(t) = 0 if and only if

a11 (t) · · · a1m (t)

 

˙ + A(t)ξ(t) = 0,
ξ(t) A(t) :=  .. ..
.
 
. .
am
1 (t) · · ·
m
am (t)

Thus we have translated the equation ∇X = 0 over the interval I0 into a time
dependent linear ordinary differential equation. By a theorem in Analysis II
(see [47, Lemma 4.4.3]), this equation has a unique solution for any initial
condition at any point in I0 . Thus we have proved that every t0 ∈ I is
3.3. PARALLEL TRANSPORT 131

contained in an interval I0 ⊂ I, open in the relative topology of I, such

that, for every t1 ∈ I0 and every v1 ∈ Tγ(t1 ) M , there exists a unique parallel
vector field X : I0 → Rn along γ|I0 satisfying X(t1 ) = v1 . We formulate this
condition on the interval I0 as a logical formula:
∀ t1 ∈ I0 ∀ v1 ∈ Tγ(t1 ) M ∃! X ∈ Vect(γ|I0 )
(3.3.1)
such that ∇X = 0 and X(t1 ) = v1 .
If two I-open intervals I0 , I1 ⊂ I satisfy this condition and have nonempty
intersection, then their union I0 ∪ I1 also satisfies (3.3.1). (Prove this!) Now
define
[
J := {I0 ⊂ I | I0 is an I-open interval, I0 satisfies (3.3.1), t0 ∈ I0 } .

This interval J satisfies (3.3.1). Moreover, it is nonempty and, by defi-

nition, it is open in the relative topology of I. We prove that it is also
closed in the relative topology of I. Thus let (ti )i∈N be a sequence in J
converging to a point t∗ ∈ I. By what we have proved above, there ex-
ists a constant ε > 0 such that the interval I ∗ := (t∗ − ε, t∗ + ε) ∩ I satis-
fies (3.3.1). Since the sequence (ti )i∈N converges to t∗ , there exists an i ∈ N
such that ti ∈ I ∗ . Since ti ∈ J there exists an interval I0 ⊂ I, open in the
relative topology of I, that contains t0 and ti and satisfies (3.3.1). Hence
the interval I0 ∪ I ∗ is open in the relative topology of I, contains t0 and t∗ ,
and satisfies (3.3.1). This shows that t∗ ∈J. Thus we have proved that the
interval J is nonempty, and open and closed in the relative topology of I.
Hence J = I and this proves Theorem 3.3.4.
Definition 3.3.5 (Parallel Transport). Let I ⊂ R be an interval and
let γ : I → M be a smooth curve. For t0 , t ∈ I we define the map
Φγ (t, t0 ) : Tγ(t0 ) M → Tγ(t) M
by Φγ (t, t0 )v0 := X(t) where X ∈ Vect(γ) is the unique parallel vector field
along γ satisfying X(t0 ) = v0 . The collection of maps Φγ (t, t0 ) for t, t0 ∈ I
is called parallel transport along γ.
Recall the notation
γ ∗ T M = (s, v) | s ∈ I, v ∈ Tγ(s) M


for the pullback tangent bundle. This set is a smooth submanifold of I ×Rn .
(See Theorem 2.6.10 and Corollary 2.6.13.) The next theorem summarizes
the properties of parallel transport. In particular, the last assertion shows
that the covariant derivative can be recovered from the parallel transport
maps.
132 CHAPTER 3. THE LEVI-CIVITA CONNECTION

Theorem 3.3.6 (Parallel Transport). Let γ : I → M be a smooth curve

on an interval I ⊂ R.
(i) The map Φγ (t, s) : Tγ(s) M → Tγ(t) M is linear for all s, t ∈ I.
(ii) For all r, s, t ∈ I we have

Φγ (t, s) ◦ Φγ (s, r) = Φγ (t, r), Φγ (t, t) = id.

(iii) For all s, t ∈ I and all v, w ∈ Tγ(s) M we have

hΦγ (t, s)v, Φγ (t, s)wi = hv, wi.

Thus Φγ (t, s) : Tγ(s) M → Tγ(t) M is an orthogonal transformation.

(iv) If J ⊂ R is an interval and σ : J → I is a smooth map, then

Φγ◦σ (t, s) = Φγ (σ(t), σ(s)).

for all s, t ∈ J.
(v) The map

I × γ ∗ T M → γ ∗ T M : (t, (s, v)) 7→ (t, Φγ (t, s)v)

is smooth.
(vi) For all X ∈ Vect(γ) and t, t0 ∈ I we have
d
Φγ (t0 , t)X(t) = Φγ (t0 , t)∇X(t).
dt
Proof. Assertion (i) holds because the sum of two parallel vector fields
along γ is again parallel and the product of a parallel vector field with a
constant real number is again parallel. Assertion (ii) follows directly from
the uniqueness statement in Theorem 3.3.4.
We prove (iii). Fix a number s ∈ I and two tangent vectors

v, w ∈ Tγ(s) M.

Define the vector fields X, Y ∈ Vect(γ) along γ by

X(t) := Φγ (t, s)v, Y (t) := Φγ (t, s)w.

These vector fields are parallel. Thus, by equation (3.2.5) in Lemma 3.2.4,
we have
d
hX, Y i = h∇X, Y i + hX, ∇Y i = 0.
dt
Hence the function I → R : t 7→ hX(t), Y (t)i is constant and this proves (iii).
3.3. PARALLEL TRANSPORT 133

We prove (iv). Fix an element s ∈ J and a tangent vector v ∈ Tγ(σ(s)) M .

Define the vector field X along γ by

X(t) := Φγ (t, σ(s))v

for t ∈ I. Thus X is the unique parallel vector field along γ that satisfies

X(α(s)) = v.

Denote
e := γ ◦ σ : J → M,
γ e := X ◦ σ : I → Rn
X
Then X
e is a vector field along γ
e and, by the chain rule, we have
d e d
X(t) = X(σ(t)) = σ̇(t)Ẋ(σ(t)).
dt dt
Projecting orthogonally onto the tangent space Tγ(σ(t)) M we obtain

∇X(t)
e = σ̇(t)∇X(σ(t)) = 0

for every t ∈ J. Hence X e is the unique parallel vector field along γ

e that
satisfies X(s)
e = v. Thus

Φγe (t, s)v = X(t)

e = X(σ(t)) = Φγ (σ(t), σ(s))v.

This proves (iv).

We prove (v). Fix a point t0 ∈ I, choose an orthonormal basis e1 , . . . , em
of Tγ(t0 ) M , and define Xi (t) := Φγ (t, t0 )ei for t ∈ I and i = 1, . . . , m. Thus
Xi ∈ Vect(γ) is the unique parallel vector field along γ such that Xi (t0 ) = ei .
Then by (iii) we have
hXi (t), Xj (t)i = δij
for all i, j ∈ {1, . . . , m} and all t ∈ I. Hence the vectors X1 (t), . . . , Xm (t)
form an orthonormal basis of Tγ(t) M for every t ∈ I. This implies that, for
each s ∈ I and each tangent vector v ∈ Tγ(s) M , we have
m
X
v= hXi (s), viXi (s).
i=1

Since each vector field Xi is parallel it satisfies Xi (t) = Φγ (t, s)Xi (s). Hence
m
X
Φγ (t, s)v = hXi (s), viXi (t) (3.3.2)
i=1

for all s, t ∈ I and v ∈ Tγ(s) M . This proves (v)

134 CHAPTER 3. THE LEVI-CIVITA CONNECTION

We prove (vi). Let X1 , . . . , Xm ∈ Vect(γ) be as in the proof of (v). Thus

every vector field X along γ can be written in the form
m
X
X(t) = ξ i (t)Xi (t), ξ i (t) := hXi (t), X(t)i.
i=1

Since the vector fields Xi are parallel we have

m
X
∇X(t) = ξ˙i (t)Xi (t)
i=1

for all t ∈ I. Hence

m
X m
X
Φγ (t0 , t)X(t) = ξ i (t)Xi (t0 ), Φγ (t0 , t)∇X(t) = ξ˙i (t)Xi (t0 ).
i=1 i=1

Evidently, the derivative of the first sum with respect to t is equal to the
second sum. This proves (vi) and the theorem.

Remark 3.3.7. For s, t ∈ I we can think of the linear map

Φγ (t, s)Π(γ(s)) : Rn → Tγ(t) M ⊂ Rn

as a real n × n matrix. The formula (3.3.2) in the proof of (v) shows that
this matrix can be expressed in the form
m
X
Φγ (t, s)Π(γ(s)) = Xi (t)Xi (s)T ∈ Rn×n .
i=1

The right hand side defines a smooth matrix valued function on I × I and
this is equivalent to the assertion in (v).

Remark 3.3.8. It follows from assertions (ii) and (iii) in Theorem 3.3.6
that
Φγ (t, s)−1 = Φγ (s, t) = Φγ (t, s)∗

for all s, t ∈ I. Here the linear map Φγ (t, s)∗ : Tγ(t) M → Tγ(s) M is under-
stood as the adjoint operator of Φγ (t, s) : Tγ(s) M → Tγ(t) M with respect to
the inner products on the two subspaces of Rn inherited from the Euclidean
inner product on the ambient space.
3.3. PARALLEL TRANSPORT 135

The two theorems in this section carry over verbatim to any smooth
vector bundle E ⊂ M × Rn over a manifold. As in the case of the tangent
bundle one can define the covariant derivative of a section of E along γ as
the orthogonal projection of the ordinary derivative in the ambient space Rn
onto the fiber Eγ(t) . Instead of parallel vector fields one then speaks about
horizontal sections and one proves as in Theorem 3.3.4 that there is a unique
horizontal section along γ through any point in any of the fibers Eγ(t0 ) . This
gives parallel transport maps from Eγ(s) to Eγ(t) for any pair s, t ∈ I and
Theorem 3.3.6 carries over verbatim to all vector bundles E ⊂ M × Rn . We
spell this out in more detail in the case where E = T M ⊥ ⊂ M × Rn is the
normal bundle of M .
Let γ : I → M be a smooth curve. A normal vector field along γ is
a smooth map Y : I → Rn such that Y (t) ⊥ Tγ(t) M for every t ∈ I. The set
of normal vector fields along γ will be denoted by

Vect⊥ (γ) := Y : I → Rn | Y is smooth and Y (t) ⊥ Tγ(t) M for all t ∈ I .



This is again a real vector space. The covariant derivative of a normal

vector field Y ∈ Vect⊥ (γ) at t ∈ I is defined as the orthogonal projection of
the ordinary derivative onto the orthogonal complement of Tγ(t) M and will
be denoted by
∇⊥ Y (t) := 1l − Π(γ(t)) Ẏ (t).


(3.3.3)
Thus the covariant derivative defines a linear operator

∇⊥ : Vect⊥ (γ) → Vect⊥ (γ).

There is a version of the Gauß–Weingarten formula for the covariant deriva-

tive of a normal vector field. This is the content of the next lemma.
Lemma 3.3.9. Let M ⊂ Rn be a smooth m-manifold. For p ∈ M and
u ∈ Tp M define the linear map hp (u) : Tp M → Tp M ⊥ by


hp (u)v := hp (u, v) = dΠ(p)u v (3.3.4)

for v ∈ Tp M . Then the following holds.

(i) The adjoint operator hp (u)∗ : Tp M ⊥ → Tp M is given by

hp (u)∗ w = dΠ(p)u w, w ∈ Tp M ⊥ .


(3.3.5)

(ii) If I ⊂ R is an interval, γ : I → M is a smooth curve, and Y ∈ Vect⊥ (γ),

then the derivative of Y satisfies the Gauß–Weingarten formula

Ẏ (t) = ∇⊥ Y (t) − hγ(t) (γ̇(t))∗ Y (t). (3.3.6)

136 CHAPTER 3. THE LEVI-CIVITA CONNECTION

Proof. Since Π(p) ∈ Rn×n is a symmetric matrix for every p ∈ M so is the

matrix dΠ(p)u for every p ∈ M and every u ∈ Tp M . Hence

hv, hp (u)∗ wi = hhp (u)v, wi


= dΠ(p)u v, w



= v, dΠ(p)u w

for every v ∈ Tp M and every w ∈ Tp M ⊥ . This proves (i).

To prove (ii) we observe that, for Y ∈ Vect⊥ (γ) and t ∈ I, we have

Π(γ(t))Y (t) = 0.

Differentiating this identity we obtain

Π(γ(t))Ẏ (t) + dΠ(γ(t))γ̇(t) Y (t) = 0

and hence


Ẏ (t) = Ẏ (t) − Π(γ(t))Ẏ (t) − dΠ(γ(t))γ̇(t) Y (t)
= ∇⊥ Y (t) − hγ(t) (γ̇(t))∗ Y (t)

for t ∈ I. Here the last equation follows from (i) and the definition of ∇⊥ .
This proves Lemma 3.3.9.

Theorem 3.3.4 and its proof carry over to the normal bundle T M ⊥ .
Thus, if γ : I → M is a smooth curve, then for all s ∈ I and w ∈ Tγ(s) M ⊥
there is a unique normal vector field Y ∈ Vect⊥ (γ) such that

∇⊥ Y ≡ 0, Y (s) = w.

This gives rise to parallel transport maps

Φ⊥ ⊥
γ (t, s) : Tγ(s) M → Tγ(t) M
⊥

defined by
Φ⊥
γ (t, s)w := Y (t)

for s, t ∈ I and w ∈ Tγ(s) M ⊥ , where Y is the unique normal vector field along
γ satisfying ∇⊥ Y ≡ 0 and Y (s) = w. These parallel transport maps satisfy
exactly the same conditions that have been spelled out in Theorem 3.3.6
for the tangent bundle and the proof carries over verbatim to the present
setting.
3.4. THE FRAME BUNDLE 137

3.4 The Frame Bundle

3.4.1 Frames of a Vector Space
Let V be an m-dimensional real vector space. A frame of V is a basis
e1 , . . . , em of V . It determines a vector space isomorphism e : Rm → V via
m
X
eξ := ξ i ei , ξ = (ξ 1 , . . . , ξ m ) ∈ Rm .
i=1

Conversely, each isomorphism e : Rm → V determines a basis e1 , . . . , em

of V via ei = e(0, . . . , 0, 1, 0 . . . , 0), where the coordinate 1 appears in the
ith place. The set of vector space isomorphisms from Rm to V will be
denoted by

Liso (Rm , V ) := {e : Rm → V | e is a vector space isomorphism} .

The general linear group GL(m) = GL(m, R) (of nonsingular real m × m-

matrices) acts on this space by composition on the right via

GL(m) × Liso (Rm , V ) → Liso (Rm , V ) : (a, e) 7→ a∗ e := e ◦ a.

This is a contravariant group action in that

a∗ b∗ e = (ba)∗ e, 1l∗ e = e

for a, b ∈ GL(m) and e ∈ Liso (Rm , V ). Moreover, the action is free, i.e. for
all a ∈ GL(m) and e ∈ Liso (Rm , V ), we have

a∗ e = e ⇐⇒ a = 1l.

It is transitive in that for all e, e0 ∈ Liso (Rm , V ) there is a group element

a ∈ GL(m) such that e0 = a∗ e. Thus we can identify the space Liso (Rm , V )
with the group GL(m) via the bijection

GL(m) → Liso (Rm , V ) : a 7→ a∗ e0

induced by a fixed element e0 ∈ Liso (Rm , V ). This identification is not

canonical; it depends on the choice of e0 . The space Liso (Rm , V ) admits a
bijection to a group but is not itself a group.
138 CHAPTER 3. THE LEVI-CIVITA CONNECTION

3.4.2 The Frame Bundle

Definition 3.4.1. Let M ⊂ Rn be a smooth m-dimensional submanifold.
The frame bundle of M is the set

F(M ) := {(p, e) | p ∈ M, e ∈ F(M )p } , (3.4.1)

where F(M )p is the space of frames of the tangent space at p, i.e.

F(M )p := Liso (Rm , Tp M ).

Define a right action of GL(m) on F(M ) by

a∗ (p, e) := (p, a∗ e) = (p, e ◦ a) (3.4.2)

for a ∈ GL(m) and (p, e) ∈ F(M ).

One can think of a frame e ∈ Liso (Rm , Tp M ) as a linear map from Rm

to Rn whose image is Tp M and hence as an n × m-matrix of rank m.
The basis of Tp M associated to this frame is given by the columns of the
matrix e ∈ Rn×m . Thus the frame bundle F(M ) of an embedded mani-
fold M ⊂ Rn is a subset of the Euclidean space Rn × Rn×m .

Lemma 3.4.2. The frame bundle

F(M ) ⊂ Rn × Rn×m

is a smooth manifold of dimension m + m2 , the group action

GL(m) × F(M ) → F(M ) : (a, (p, e)) 7→ a∗ (p, e)

is smooth, and the projection

π : F(M ) → M

defined by π(p, e) := p for (p, e) ∈ F(M ) is a surjective submersion. The

orbits of the GL(m)-action on F(M ) are the fibers of this projection, i.e.

GL(m)∗ (p, e) = π −1 (p) ∼

= F(M )p

for (p, e) ∈ F(M ), and the group GL(m) acts freely and transitively on each
of these fibers.
3.4. THE FRAME BUNDLE 139

Proof. Let U ⊂ M be an M -open set. A moving frame over U is a se-

quence of m smooth vector fields E1 , . . . , Em ∈ Vect(U ) on U such that the
vectors E1 (p), . . . , Em (p) form a basis of Tp M for each p ∈ U . Any such
moving frame gives a bijection
U × GL(m) → F(U ) : (p, a) 7→ a∗ (p, E(p)) = (p, E(p) ◦ a),
where
E(p) := (E1 (p), . . . , Em (p)) ∈ F(M )p
for p ∈ U . This bijection (when composed with a parametrization of U )
gives a parametrization of the open set F(U ) in F(M ). The assertions of
the lemma then follow from the fact that the diagram
U × GL(m) / F(U )
EE
EE ~~
EE
EE ~~~
EE ~
~~
pr1 EEE ~
EE ~~ π
EE ~~~
EE ~ ~
"
U
commutes. More precisely, suppose that there exists a coordinate chart
φ:U →Ω
with values in an open set Ω ⊂ Rm , and denote its inverse by
ψ := φ−1 : Ω → U.
Then the open set
F(U ) = π −1 (U ) = {(p, e) ∈ F(M ) | p ∈ U } = (U × Rn×m ) ∩ F(M )
is parametrized by the map


Ω × GL(m) → F(U ) : (x, a) 7→ ψ(x), dψ(x) ◦ a .
This map is amooth and so is its inverse


F(U ) → Ω × GL(m) : (p, e) 7→ φ(p), dφ(p) ◦ e .
These are the desired coordinate chart on F(M ). Thus F(M ) is a smooth
manifold of dimension m + m2 . Moreover, in these coordinates the projec-
tion π : F(U ) → U is the map Ω × GL(m) → Ω : (x, a) 7→ x and so π is a
submersion. The remaining assertions follow directly from the definitions
and this proves Lemma 3.4.2.
140 CHAPTER 3. THE LEVI-CIVITA CONNECTION

The frame bundle F(M ) is a principal bundle over M with struc-

ture group GL(m). More generally, a principal bundle over a manifold B
with structure group G is a smooth manifold P equipped with a surjective
submersion π : P → B and a smooth contravariant action

G × P → P : (g, p) 7→ pg

by a Lie group G such that π(pg) = π(p) for all p ∈ P and g ∈ G and such
that the group G acts freely and transitively on the fiber Pb := π −1 (b) for
each b ∈ B. In this book we shall mostly be concerned with the frame bundle
of a manifold M and the orthonormal frame bundle.

Definition 3.4.3. The orthonormal frame bundle of M is the set

n o
O(M ) := (p, e) ∈ Rn × Rn×m | p ∈ M, im e = Tp M, eT e = 1lm×m .

If we denote by ei := e(0, . . . , 0, 1, 0, . . . , 0) (with 1 as the ith argument) the

basis of Tp M induced by the isomorphism e : Rm → Tp M , then we have

e1 , . . . , em is an
eT e = 1l ⇐⇒ hei , ej i = δij ⇐⇒
orthonormal basis.

Thus O(M ) is the bundle of orthonormal frames of the tangent spaces Tp M

or the bundle of orthogonal isomorphisms e : Rm → Tp M . It is a principal
bundle over M with structure group O(m).

Exercise 3.4.4. Prove that O(M ) is a submanifold of F(M ) and that the
obvious projection π : O(M ) → M is a submersion. Prove that the action
of GL(m) on F(M ) restricts to an action of the orthogonal group O(m)
on O(M ) whose orbits are the fibers

O(M )p := e ∈ Rn×m  (p, e) ∈ O(M )

 

= e ∈ Liso (Rm , Tp M )  eT e = 1l .
 

Hint: If φ : U → Ω is a coordinate chart on M with inverse ψ : Ω → U ,

then
ex := dψ(x)(dψ(x)T dψ(x))−1/2 : Rm → Tψ(x) M

is an orthonormal frame of the tangent space Tψ(x) M for every x ∈ Ω.

3.4. THE FRAME BUNDLE 141

3.4.3 Horizontal Lifts

We have seen in Lemma 3.4.2 that the frame bundle F(M ) is a smooth
submanifold of Rn × Rn×m . Next we examine the tangent space of F(M ) at
a point (p, e) ∈ F(M ). By Definition 2.2.1, this tangent space is given by
 
 R → F(M ) : t 7→ (γ(t), e(t)) 


T(p,e) F(M ) = (γ̇(0), ė(0))  is a smooth curve satisfying .

γ(0) = p and e(0) = e
  

The next lemma gives an explicit formula for this tangent space in terms of
the second fundamental form hp : Tp M × Tp M → Tp M ⊥ in Definition 3.1.9.
Compare this formula with Lemma 4.3.1 in the next chapter.
Lemma 3.4.5. Let M ⊂ Rn be a smooth m-dimensional submanifold. Then
the tangent space of F(M ) at (p, e) is given by
(  )
 pb ∈ Tp M,
eb ∈ Rn×m , and

T(p,e) F(M ) = (b
p, eb)  . (3.4.3)
 1l − Π(p) eb = hp (bp)e

Proof. We prove the inclusion “⊂” in (3.4.3). Let (b p, eb) ∈ T(p,e) F(M ) and
choose a smooth curve R → F(M ) : t 7→ (γ(t), e(t)) such that

γ(0) = p, e(0) = e, γ̇(0) = pb, ė(0) = eb.

Fix a vector ξ ∈ M and define the vector field X ∈ Vect(γ) by X(t) := e(t)ξ
for t ∈ R. Then the Gauß–Weingarten formula (3.2.2) asserts that

ė(t)ξ = Ẋ(t)
= ∇X(t) + hγ(t) (γ̇(t), X(t))
= Π(γ(t))ė(t)ξ + hγ(t) (γ̇(t), e(t)ξ)

for all t ∈ R. Take t = 0 to obtain

(1l − Π(p))b
eξ = hp (b
p, eξ) = hp (b
p)eξ

for all ξ ∈ Rm . This proves the inclusion “⊂” in (3.4.3). Equality holds
because both sides of the equation are (m+m2 )-dimensional linear subspaces
of Rn × Rn×m . This proves Lemma 3.4.5.

It is convenient to consider two kinds of curves in F(M ), namely vertical

curves with constant projections to M and horizontal lifts of curves in M .
We denote by L(Rm , Tp M ) the space of linear maps from Rm to Tp M .
142 CHAPTER 3. THE LEVI-CIVITA CONNECTION

Definition 3.4.6 (Horizontal Lift). Let γ : R → M be a smooth curve.

A smooth curve β : R → F(M ) is called a lift of γ iff

π ◦ β = γ.

Any such lift has the form β(t) = (γ(t), e(t)) with e(t) ∈ Liso (Rm , Tγ(t) M ).
The associated curve of frames e(t) of the tangent spaces Tγ(t) M is called a
moving frame along γ. A curve

β(t) = (γ(t), e(t)) ∈ F(M )

is called horizontal or a horizontal lift of γ iff the vector field

X(t) := e(t)ξ

along γ is parallel for every ξ ∈ Rm . Thus a horizontal lift of γ has the form

β(t) = (γ(t), Φγ (t, 0)e) (3.4.4)

for some e ∈ Liso (Rm , Tγ(0) M ).

Lemma 3.4.7. (i) The tangent space of F(M ) at (p, e) ∈ F(M ) is the direct
sum
T(p,e) F(M ) = H(p,e) ⊕ V(p,e)

of the horizontal space

 
H(p,e) := (v, hp (v)e)  v ∈ Tp M (3.4.5)

and the vertical space

V(p,e) := {0} × L(Rm , Tp M ). (3.4.6)

(ii) The vertical space V(p,e) at (p, e) ∈ F(M ) is the kernel of the linear map

dπ(p, e) : T(p,e) F(M ) → Tp M.

(iii) A curve β : R → F(M ) is horizontal if and only if it is tangent to the

horizontal spaces, i.e. β̇(t) ∈ Hβ(t) for every t ∈ R.
(iv) If β : R → F(M ) is a horizontal curve, so is a∗ β for every a ∈ GL(m).
3.4. THE FRAME BUNDLE 143

Proof. The proof has four steps.

Step 1. Let (p, e) ∈ F(M ). Then V(p,e) = ker dπ(p, e) ⊂ T(p,e) F(M ).
Since π is a submersion, the fiber π −1 (p) is a submanifold of F(M ) by
Theorem 2.2.19 and T(p,e) π −1 (p) = ker dπ(p, e). Now let (b
p, eb) ∈ ker dπ(p, e).
Then there exists a vertical curve β : R → F(M ) with π ◦ β ≡ p such that
β(0) = (p, e), β̇(0) = (b
p, eb).
Any such curve has the form β(t) := (p, e(t)) where e(t) ∈ Liso (Rm , Tp M ).
Hence pb = 0 and eb = ė(0) ∈ L(Rm , Tp M ). This shows that
ker dπ(p, e) ⊂ V(p,e) . (3.4.7)
Conversely, for every eb ∈ L(Rm , Tp M ), the curve
R → L(Rm , Tp M ) : t 7→ e(t) := e + tb
e
takes values in the open set Liso (Rm , Tp M ) for t sufficiently small and
hence β(t) := (p, e(t)) is a vertical curve with β̇(0) = (0, eb). Thus
V(p,e) ⊂ ker dπ(p, e) ⊂ T(p,e) F(M ). (3.4.8)
Combining (3.4.7) and (3.4.8) we obtain Step 1 and part (ii).
Step 2. Let (p, e) ∈ F(M ). Then H(p,e) ⊂ T(p,e) F(M ). Moreover, every
horizontal curve β : R → F(M ) satisfies β̇(t) ∈ Hβ(t) for all t ∈ R.
Fix a tangent vector v ∈ Tp M , let γ : R → M be a smooth curve satisfy-
ing γ(0) = p and γ̇(0) = v, and let β : R → F(M ) be the horizontal lift of γ
with β(0) = (p, e). Then
β(t) = (γ(t), e(t)), e(t) := Φγ (t, 0)e.
Fix a vector ξ ∈ Rm and consider the vector field
X(t) := e(t)ξ = Φγ (t, 0)eξ
along γ. This vector field is parallel and hence, by the Gauß–Weingarten
formula, it satisfies
ė(0)ξ = Ẋ(0) = hγ(0) (γ̇(0), X(0)) = hp (v)eξ.
Here we have used (3.3.4). Thus
(v, hp (v)e) = (γ̇(0), ė(0)) = β̇(0) ∈ Tβ(0) F(M ) = T(p,e) F(M ).

and so H(p,e) ⊂ T(p,e) F(M ). Moreover, β̇(0) = (v, hp (v)e) ∈ H(p,e) = Hβ(0)
and this proves Step 2.
144 CHAPTER 3. THE LEVI-CIVITA CONNECTION

Step 3. We prove part (i).

We have V(p,e) ⊂ T(p,e) F(M ) by Step 1 and H(p,e) ⊂ T(p,e) F(M ) by Step 2.
Moreover H(p,e) ∩ V(p,e) = {0} and so T(p,e) F(M ) = H(p,e) ⊕ V(p,e) for dimen-
sional reasons. This proves Step 3.
Step 4. We prove parts (iii) and (iv).
By Step 2 every horizontal curve β : R → F(M ) satisfies β̇(t) ∈ Hβ(t) . Con-
versely, let R → F(M ) : t 7→ β(t) = (γ(t), e(t)) be a smooth curve satisfy-
ing β̇(t) ∈ Hβ(t) for all t. Then ė(t) = hγ(t) (γ̇(t))e(t) for all t. By the Gauß–
Weingarten formula (3.2.2) this implies that the vector field X(t) = e(t)ξ
along γ is parallel for every ξ ∈ Rm , so β is horizontal. This proves part (iii).
Part (iv) follows from (iii) and the fact that the horizontal tangent bun-
dle H ⊂ T F(M ) is invariant under the induced action of the group GL(m)
on T F(M ). This proves Lemma 3.4.7.

−1
F(M)p = π (p)

F(M)
(p,e)

M p

Figure 3.3: The frame bundle.

The reason for the terminology introduced in Definition 3.4.6 is that one
draws the extremely crude picture of the frame bundle displayed in Fig-
ure 3.3. One thinks of F(M ) as “lying over” M . One would then represent
the equation γ = π ◦ β by the following commutative diagram:
F(M ) ;
8
β
ppppp
p
ppp
π
ppp γ /M

R
hence the word “lift”. The vertical space is tangent to the vertical line in
Figure 3.3 while the horizontal space is transverse to the vertical space. This
crude imagery can be extremely helpful.
3.4. THE FRAME BUNDLE 145

Exercise 3.4.8. The group GL(m) acts on F(M ) by diffeomorphisms. Thus

for each a ∈ GL(m) the map
F(M ) → F(M ) : (p, e) 7→ a∗ (p, e) = (p, e ◦ a)
is a diffeomorphism of F(M ). The derivative of this diffeomorphism is a
diffeomorphism of the tangent bundle T F(M ) and this is called the induced
action of GL(m) on T F(M ). Prove that the horizontal and vertical sub-
bundles are invariant under the induced action of GL(m) on T F(M ).
Exercise 3.4.9. Prove that H(p,e) ⊂ T(p,e) O(M ) and that
0 0
T(p,e) O(M ) = H(p,e) ⊕ V(p,e) , V(p,e) := V(p,e) ∩ T(p,e) O(M ),
for every (p, e) ∈ O(M ).
The following definition introduces an important class of vector fields on
the frame bundle that will play a central role in Section 3.5. They will be
used to prove the Development Theorem 3.5.21 in §3.5.4 below.
Definition 3.4.10 (Basic Vector Field). Every vector ξ ∈ Rm determines
a vector field Bξ ∈ Vect(F(M )) defined by


Bξ (p, e) := eξ, hp (eξ)e (3.4.9)
for (p, e) ∈ F(M ). This vector field is horizontal, i.e.
Bξ (p, e) ∈ H(p,e) ,
and projects to eξ, i.e.
dπ(p, e)Bξ (p, e) = eξ
for all (p, e) ∈ F(M ). These two conditions determine the vector field Bξ
uniquely. It is called the basic vector field corresponding to ξ.
Exercise 3.4.11. (i) Prove that every basic vector field Bξ ∈ Vect(F(M ))
is tangent to the orthonormal frame bundle O(M ).
(ii) Let R → F(M ) : t 7→ (γ(t), e(t)) be an integral curve of the vector
field Bξ and a ∈ GL(m). Prove that R → F(M ) : t 7→ a∗ β(t) = (γ(t), a∗ e(t))
is an integral curve of Ba−1 ξ .
(iii) Prove that the vector field Bξ ∈ Vect(F(M )) is complete for all ξ ∈ Rm
if and only if the restricted vector field Bξ |O(M ) ∈ Vect(O(M )) on the or-
thonormal frame bundle is complete for all ξ ∈ Rm .
Definition 3.4.12. The manifold M is called complete if, for every smooth
curve ξ : R → Rm and every element (p0 , e0 ) ∈ F(M ), there exists a smooth
curve β : R → F(M ) such that β(0) = (p0 , e0 ) and β̇(t) = Bξ(t) (β(t)) for
all t ∈ R.
146 CHAPTER 3. THE LEVI-CIVITA CONNECTION

3.5 Motions and Developments

Our aim in this sections is to define motion without sliding, twisting, or
wobbling. This is the motion that results when a heavy object is rolled,
with a minimum of friction, along the floor. It is also the motion of the large
snowball a child creates as it rolls it into the bottom part of a snowman.
We shall eventually justify mathematically the physical intuition that
either of the curves of contact in such ideal rolling may be specified arbi-
trarily; the other is then determined uniquely. Thus for example the heavy
object may be rolled along an arbitrary curve on the floor; if that curve is
marked in wet ink, another curve will be traced in the object. Conversely,
if a curve is marked in wet ink on the object, the object may be rolled so as
to trace a curve on the floor. However, if both curves are prescribed, it will
be necessary to slide the object as it is being rolled, if one wants to keep the
curves in contact.
We assume throughout this section that M and M 0 are two m-dimensio-
nal submanifolds of Rn . Objects on M 0 will be denoted by the same letter
as the corresponding objects on M with primes affixed. Thus for example
Π0 (p0 ) ∈ Rn×n denotes the orthogonal projection of Rn onto the tangent
space Tp0 M 0 , ∇0 denotes the covariant derivative of a vector field along a
curve in M 0 , and Φ0γ 0 denotes parallel transport along a curve in M 0 .

3.5.1 Motion
Definition 3.5.1. A motion of M along M 0 (on an interval I ⊂ R) is
a triple (Ψ, γ, γ 0 ) of smooth maps
Ψ : I → O(n), γ : I → M, γ0 : I → M 0
such that
Ψ(t)Tγ(t) M = Tγ 0 (t) M 0 ∀ t ∈ I.
Note that a motion also matches normal vectors, i.e.
⊥
Ψ(t)Tγ(t) M ⊥ = Tγ 0 (t) M 0 ∀ t ∈ I.
Remark 3.5.2. Associated to a motion (Ψ, γ, γ 0 ) of M along M 0 is a family
of (affine) isometries ψt : Rn → Rn defined by
ψt (p) := γ 0 (t) + Ψ(t) p − γ(t)


(3.5.1)
for t ∈ I and p ∈ Rn . These isometries satisfy
ψt (γ(t)) = γ 0 (t), dψt (γ(t))Tγ(t) M = Tγ 0 (t) M 0 ∀ t ∈ I.
3.5. MOTIONS AND DEVELOPMENTS 147

Remark 3.5.3. There are three operations on motions.

Reparametrization. If (Ψ, γ, γ 0 ) is a motion of M along M 0 on an in-
terval I ⊂ R and σ : J → I is a smooth map between intervals, then the
triple
(Ψ ◦ σ, γ ◦ σ, γ 0 ◦ σ)
is a motion of M along M 0 on the interval J.
Inversion. If (Ψ, γ, γ 0 ) is a motion of M along M 0 , then

(Ψ−1 , γ 0 , γ)

is a motion of M 0 along M .
Composition. If (Ψ, γ, γ 0 ) is a motion of M along M 0 on an interval I
and (Ψ0 , γ 0 , γ 00 ) is a motion of M 0 along M 00 on the same interval, then

(Ψ0 Ψ, γ, γ 00 )

is a motion of M along M 00 .

We now give the three simplest examples of “bad” motions; i.e. motions
which do not satisfy the concepts we are about to define. In all three of
these examples, p is a point of M and M 0 is the affine tangent space to M
at p:
M 0 := p + Tp M = {p + v | v ∈ Tp M } .

Example 3.5.4 (Pure Sliding). Take a nonzero tangent vector v ∈ Tp M

and let
γ(t) := p, γ 0 (t) = p + tv, Ψ(t) := 1l.
Then γ̇(t) = 0, γ̇ 0 (t) = v 6= 0, and so

Ψ(t)γ̇(t) 6= γ̇ 0 (t).

(See Figure 3.4.)

M’
p

Figure 3.4: Pure sliding.

148 CHAPTER 3. THE LEVI-CIVITA CONNECTION

Example 3.5.5 (Pure Twisting). Let γ and γ 0 be the constant curves

γ(t) = γ 0 (t) = p

and take Ψ(t) to be the identity on Tp M ⊥ and any curve of rotations on the
tangent space Tp M . As a concrete example with m = 2 and n = 3 one can
take M to be the sphere of radius one centered at the point (0, 1, 0) and p
to be the origin:

M := (x, y, z) ∈ R3 | x2 + (y − 1)2 + z 2 = 1 ,

p := (0, 0, 0).

Then M 0 is the (x, z)-plane and A(t) is any curve of rotations in the (x, z)-
plane, i.e. about the y-axis Tp M ⊥ . (See Figure 3.5.)

M’
p

Figure 3.5: Pure twisting.

Example 3.5.6 (Pure Wobbling). This is the same as pure twisting

except that Ψ(t) is the identity on Tp M and any curve of rotations on
Tp M ⊥ . As a concrete example with m = 1 and n = 3 one can take M to be
the circle of radius one in the (x, y)-plane centered at the point (0, 1, 0) and
p to be the origin:

M := (x, y, 0) ∈ R3 | x2 + (y − 1)2 = 1 ,

p := (0, 0, 0).

Then M 0 is the x-axis and Ψ(t) is any curve of rotations in the (y, z)-plane,
i.e. about the axis M 0 . (See Figure 3.6.)

M’ p

Figure 3.6: Pure wobbling.

3.5. MOTIONS AND DEVELOPMENTS 149

3.5.2 Sliding
When a train slides on the track (e.g. in the process of stopping suddenly),
there is a terrific screech. Since we usually do not hear a screech, this means
that the wheel moves along without sliding. In other words the velocity of
the point of contact in the train wheel M equals the velocity of the point
of contact in the track M 0 . But the track is not moving; hence the point of
contact in the wheel is not moving. One may explain the paradox this way:
the train is moving forward and the wheel is rotating around the axle. The
velocity of a point on the wheel is the sum of these two velocities. When
the point is on the bottom of the wheel, the two velocities cancel.
Definition 3.5.7. A motion (Ψ, γ, γ 0 ) is said to be without sliding iff it
satisfies Ψ(t)γ̇(t) = γ̇ 0 (t) for every t.
Here is the geometric picture of the no sliding condition. As explained
in Remark 3.5.2 we can view a motion as a smooth family of isometries
ψt (p) := γ 0 (t) + Ψ(t) p − γ(t)

acting on the manifold M with γ(t) ∈ M being the point of contact with M 0 .
Differentiating the curve t 7→ ψt (p) which describes the motion of the point
p ∈ M in the space Rn we obtain
d
ψt (p) = γ̇ 0 (t) − Ψ(t)γ̇(t) + Ψ̇(t) p − γ(t) .


dt
Taking p = γ(t0 ) we find

d 
ψt (γ(t0 )) = γ̇ 0 (t0 ) − Ψ(t0 )γ̇(t0 ).
dt 
t=t0
This expression vanishes under the no sliding condition. In general the
curve t 7→ ψt (γ(t0 )) will be non-constant, but (when the motion is without
sliding) its velocity will vanish at the instant t = t0 ; i.e. at the instant when
it becomes the point of contact. In other words the motion is without sliding
if and only if the point of contact is motionless.
We remark that, if the motion is without sliding, we have:
 0 
γ̇ (t) = |Ψ(t)γ̇(t)| = |γ̇(t)|

so that the curves γ and γ 0 have the same arclength:

Z t1 Z t1
 0 
γ̇ (t) dt = |γ̇(t)| dt
t0 t0

on any interval [t0 , t1 ] ⊂ I. Hence any motion with γ̇ = 0 and γ̇ 0 6= 0 is not

without sliding (such as the example of pure sliding above).
150 CHAPTER 3. THE LEVI-CIVITA CONNECTION

Exercise 3.5.8. Give an example of a motion where |γ̇(t)| = |γ̇ 0 (t)| for
every t but which is not without sliding.

Example 3.5.9. We describe mathematically the motion of the train wheel.

Let the center of the wheel move right parallel to the x-axis at height one
and the wheel have radius one and make one revolution in 2π units of time.
Then the track M 0 is the x-axis and we take

M := (x, y) ∈ R2 | x2 + (y − 1)2 = 1 .


Choose
γ(t) := (cos(t − π/2), 1 + sin(t − π/2))
= (sin(t), 1 − cos(t)),
0
γ (t) := (t, 0),

and define Ψ(t) ∈ GL(2) by

 
cos(t) sin(t)
Ψ(t) := .
− sin(t) cos(t)

The reader can easily verify that this is a motion without sliding. A fixed
point p0 on M , say p0 = (0, 0), sweeps out a cycloid with parametric equa-
tions
x = t − sin(t), y = 1 − cos(t).
(Check that (ẋ, ẏ) = (0, 0) when y = 0; i.e. for t = 2nπ.)

Remark 3.5.10. These same formulas give a motion of a sphere M rolling

without sliding along a straight line in a plane M 0 . Namely in coordinates
(x, y, z) the sphere has equation

x2 + (y − 1)2 + z 2 = 1,

the plane is y = 0 and the line is the x-axis. The z-coordinate of a point is
unaffected by the motion. Note that the curve γ 0 traces out a straight line
in the plane M 0 and the curve γ traces out a great circle on the sphere M .

Exercise 3.5.11. The operations of reparametrization, inversion, and com-

position respect motion without sliding; i.e. if (Ψ, γ, γ 0 ) and (Ψ0 , γ 0 , γ 00 ) are
motions without sliding on an interval I and σ : J → I is a smooth map
between intervals, then the motions (Ψ ◦ σ, γ ◦ σ, γ 0 ◦ σ), (Ψ−1 , γ 0 , γ), and
(Ψ0 Ψ, γ, γ 00 ) are also without sliding.
3.5. MOTIONS AND DEVELOPMENTS 151

3.5.3 Twisting and Wobbling

A motion (Ψ, γ, γ 0 ) on an intervall I ⊂ R transforms vector fields along γ
into vector fields along γ 0 by the formula

X 0 (t) = (ΨX)(t) := Ψ(t)X(t) ∈ Tγ 0 (t) M 0

for t ∈ I and X ∈ Vect(γ); so X 0 ∈ Vect(γ 0 ).

Lemma 3.5.12. Let (Ψ, γ, γ 0 ) be a motion of M along M 0 on an interval
I ⊂ R. Then the following are equivalent.
(i) The instantaneous velocity of each tangent vector is normal, i.e. for t ∈ I
⊥
Ψ̇(t)Tγ(t) M ⊂ Tγ 0 (t) M 0 .

(ii) Ψ intertwines covariant differentiation, i.e. for X ∈ Vect(γ)

∇0 (ΨX) = Ψ∇X.

(iii) Ψ transforms parallel vector fields along γ into parallel vector fields
along γ 0 , i.e. for X ∈ Vect(γ)

∇X = 0 =⇒ ∇0 (ΨX) = 0.

(iv) Ψ intertwines parallel transport, i.e. for s, t ∈ I and v ∈ Tγ(s) M

Ψ(t)Φγ (t, s)v = Φ0γ 0 (t, s)Ψ(s)v.

A motion that satisfies these conditions is called without twisting.

Proof. We prove that (i) is equivalent to (ii). A motion satisfies the equation

Ψ(t)Π(γ(t)) = Π0 (γ 0 (t))Ψ(t)

for every t ∈ I. This restates the condition that Ψ(t) maps tangent vectors
of M to tangent vectors of M 0 and normal vectors of M to normal vectors
of M 0 . Differentiating the equation X 0 (t) = Ψ(t)X(t) we obtain

Ẋ 0 (t) = Ψ(t)Ẋ(t) + Ψ̇(t)X(t).

Applying Π0 (γ 0 (t)) this gives

∇0 X 0 = Ψ∇X + Π0 (γ 0 )Ψ̇X.

Hence (ii) holds if and only if Π0 (γ 0 (t))Ψ̇(t) = 0 for every t ∈ I. Thus we

have proved that (i) is equivalent to (ii). That (ii) implies (iii) is obvious.
152 CHAPTER 3. THE LEVI-CIVITA CONNECTION

We prove that (iii) implies (iv). Let t0 ∈ I and v0 ∈ Tγ(t0 ) M . Define the
vector field X ∈ Vect(γ) by X(t) := Φγ (t, t0 )v0 for t ∈ I and let X 0 := ΨX.
Then ∇X = 0, hence ∇0 X 0 = 0 by (iii), and hence

X 0 (t) = Φ0γ 0 (t, t0 )X 0 (t0 ) = Φ0γ 0 (t, t0 )Ψ(t0 )v0

for all t ∈ I. Since X 0 (t) = Ψ(t)X(t) = Ψ(t)Φγ (t, t0 )v0 , this implies (iv).
We prove that (iv) implies (ii). Let X ∈ Vect(γ) and X 0 := ΨX. By (iv)
we have
Φ0γ 0 (t0 , t)X 0 (t) = Ψ(t0 )Φγ (t0 , t)X(t).
Differentiating this equation with respect to t at t = t0 and using Theo-
rem 3.3.6, we obtain ∇0 X 0 (t0 ) = Ψ(t0 )∇X(t0 ). This proves the lemma.

Lemma 3.5.13. Let (Ψ, γ, γ 0 ) be a motion of M along M 0 on an interval

I ⊂ R. Then the following are equivalent.
(i) The instantaneous velocity of each normal vector is tangent, i.e. for t ∈ I

Ψ̇(t)Tγ(t) M ⊥ ⊂ Tγ 0 (t) M 0 .

(ii) Ψ intertwines normal covariant differentiation, i.e. for Y ∈ Vect⊥ (γ)

⊥
∇0 (ΨY ) = Ψ∇⊥ Y.

(iii) Ψ transforms parallel normal vector fields along γ into parallel normal
vector fields along γ 0 , i.e. for Y ∈ Vect⊥ (γ)
⊥
∇⊥ Y = 0 =⇒ ∇0 (ΨY ) = 0.

(iv) Ψ intertwines parallel transport of normal vector fields, i.e. for s, t ∈ I

and w ∈ Tγ(s) M ⊥
⊥
Ψ(t)Φ⊥ 0
γ (t, s)w = Φ γ 0 (t, s)Ψ(s)w.

A motion that satisfies these conditions is called without wobbling.

The proof that the four conditions in Lemma 3.5.13 are equivalent is
word for word analogous to the proof of Lemma 3.5.12 and will be omitted.
In summary a motion is without twisting iff tangent vectors at the point
of contact are rotating towards the normal space and it is without wobbling
iff normal vectors at the point of contact are rotating towards the tangent
space. In case m = 2 and n = 3 motion without twisting means that the
instantaneous axis of rotation is parallel to the tangent plane.
3.5. MOTIONS AND DEVELOPMENTS 153

Remark 3.5.14. The operations of reparametrization, inversion, and com-

position respect motion without twisting, respectively without wobbling; i.e.
if (Ψ, γ, γ 0 ) and (Ψ0 , γ 0 , γ 00 ) are motions without twisting, respectively with-
out wobbling, on an interval I and σ : J → I is a smooth map between
intervals, then the motions (Ψ ◦ σ, γ ◦ σ, γ 0 ◦ σ), (Ψ−1 , γ 0 , γ), and (Ψ0 Ψ, γ, γ 00 )
are also without twisting, respectively without wobbling.
Remark 3.5.15. Let I ⊂ R be an interval and t0 ∈ I. Given curves
γ : I → M and γ 0 : I → M 0 and an orthogonal matrix Ψ0 ∈ O(n) such that
Ψ0 Tγ(t0 ) M = Tγ 0 (t0 ) M 0
there is a unique motion (Ψ, γ, γ 0 ) of M along M 0 (with the given γ and γ 0 )
without twisting or wobbling satisfying the initial condition:
Ψ(t0 ) = Ψ0 .
Indeed, the path of matrices Ψ : I → O(n) is uniquely determined by the
conditions (iv) in Lemma 3.5.12 and Lemma 3.5.13. It is given by the explicit
formula
Ψ(t)v = Φ0γ 0 (t, t0 )Ψ0 Φγ (t0 , t)Π(γ(t))v
⊥ (3.5.2)
+ Φ0 γ 0 (t, t0 )Ψ0 Φ⊥


γ (t0 , t) v − Π(γ(t))v
for t ∈ I and v ∈ Rn . We prove below a somewhat harder result where the
motion is without twisting, wobbling, or sliding. It is in this situation that γ
and γ 0 determine one another (up to an initial condition).
Remark 3.5.16. We can now give another interpretation of parallel trans-
port. Given γ : R → M and v0 ∈ Tγ(t0 ) M take M 0 to be an affine subspace of
the same dimension as M . Let (Ψ, γ, γ 0 ) be a motion of M along M 0 without
twisting (and, if you like, without sliding or wobbling). Let X 0 ∈ Vect(γ 0 )
be the constant vector field along γ 0 (so that ∇0 X 0 = 0) with value
X 0 (t) = Ψ0 v0 , Ψ0 := Ψ(t0 ).
Let X ∈ Vect(γ) be the corresponding vector field along γ so that
Ψ(t)X(t) = Ψ0 v0
Then X(t) = Φγ (t, t0 )v0 . To put it another way, imagine that M is a ball. To
define parallel transport along a given curve γ roll the ball (without sliding)
along a plane M 0 keeping the curve γ in contact with the plane M 0 . Let γ 0
be the curve traced out in M 0 . If a constant vector field in the plane M 0 is
drawn in wet ink along the curve γ 0 , it will mark off a (covariant) parallel
vector field along γ in M .
Exercise 3.5.17. Describe parallel transport along a great circle in a sphere.
154 CHAPTER 3. THE LEVI-CIVITA CONNECTION

3.5.4 Development
A development is an intrinsic version of motion without sliding or twisting.

Definition 3.5.18. A development of M along M 0 (on an interval I)

is a triple (Φ, γ, γ 0 ) where γ : I → M and γ 0 : I → M 0 are smooth paths
and Φ is a family of orthogonal isomorphisms

Φ(t) : Tγ(t) M → Tγ 0 (t) M 0

parametrized by t ∈ I, such that

Φ(t)γ̇(t) = γ̇ 0 (t) (3.5.3)

for all t ∈ I and Φ intertwines parallel transport, i.e.

Φ(t)Φγ (t, s) = Φ0γ 0 (t, s)Φ(s) (3.5.4)

for all s, t ∈ I. In particular, the family Φ of isomorphisms is smooth, i.e.

if X is a smooth vector field along γ, then the formula X 0 (t) := Φ(t)X(t)
defines a smooth vector field along γ 0 .

Lemma 3.5.19. Let I ⊂ R be an interval, γ : I → M and γ 0 : I → M 0

be smooth curves, and Φ(t) : Tγ(t) M → Tγ 0 (t) M 0 be a family of orthogonal
isomorphisms parametrized by t ∈ I. Then the following are equivalent.
(i) (Φ, γ, γ 0 ) is a development.
(ii) Φ satisfies (3.5.3) and

∇0 (ΦX) = Φ∇X (3.5.5)

for all X ∈ Vect(γ).

(iii) There exists a motion (Ψ, γ, γ 0 ) without sliding and twisting such that

Φ(t) = Ψ(t)|Tγ(t) M for all t ∈ I. (3.5.6)

(iv) There exists a motion (Ψ, γ, γ 0 ) of M along M 0 without sliding, twisting,

and wobbling that satisfies (3.5.6).

Proof. That (3.5.4) is equivalent to (3.5.5) was proved in Lemma 3.5.12.

This (i) is equivalent to (ii). That (iv) implies (iii) and (iii) implies (i) is
obvious. To prove that (i) implies (iv) choose any t0 ∈ I and any orthogonal
matrix Ψ0 ∈ O(n) such that Ψ0 |Tγ(t0 ) M = Φ(t0 ) and define Ψ(t) : Rn → Rn
by (3.5.2). This proves Lemma 3.5.19.
3.5. MOTIONS AND DEVELOPMENTS 155

Remark 3.5.20. The operations of reparametrization, inversion, and com-

position yield developments when applied to developments; i.e. if (Φ, γ, γ 0 ) is
a development of M along M 0 on an interval I, (Φ0 , γ 0 , γ 00 ) is a development
of M 0 along M 00 on the same interval I, and σ : J → I is a smooth map of
intervals, then the triples

(Φ ◦ σ, γ ◦ σ, γ 0 ◦ σ), (Φ−1 , γ 0 , γ)), (Φ0 Φ, γ, γ 00 )

are all developments.

Theorem 3.5.21 (Development Theorem). Let p0 ∈ M and t0 ∈ R,

let γ 0 : R → M 0 be a smooth curve, and let

Φ0 : Tp0 M → Tγ 0 (t0 ) M 0

be an orthogonal isomorphism. Then the following holds.

(i) There exists a development (Φ, γ, γ 0 |I ) on some open interval I ⊂ R
containing t0 that satisfies the initial condition

γ(t0 ) = p0 , Φ(t0 ) = Φ0 . (3.5.7)

(ii) Any two developments (Φ1 , γ1 , γ 0 |I1 ) and (Φ2 , γ2 , γ 0 |I2 ) as in (i) on two
intervals I1 and I2 agree on the intersection I1 ∩ I2 , i.e.

γ1 (t) = γ2 (t), Φ1 (t) = Φ2 (t)

for every t ∈ I1 ∩ I2 .
(iii) If M is complete, then (i) holds with I = R.

Proof. Let γ : R → M be any smooth curve such that

γ(t0 ) = p0

and, for t ∈ R, define the linear map

Φ(t) : Tγ(t) M → Tγ 0 (t) M 0

by
Φ(t) := Φ0γ 0 (t, t0 )Φ0 Φγ (t0 , t). (3.5.8)
This is an orthogonal transformation for every t and it intertwines parallel
transport. However, in general Φ(t)γ̇(t) will not be equal to γ̇ 0 (t).
156 CHAPTER 3. THE LEVI-CIVITA CONNECTION

To construct a development that satisfies (3.5.3), we choose an orthonor-

mal frame e0 : Rm → Tp0 M and, for t ∈ R, define e(t) : Rm → Tγ(t) M by

e(t) := Φγ (t, t0 )e0 . (3.5.9)

We can think of e(t) as a real n × m-matrix and the map

R → Rn×m : t 7→ e(t)

is smooth. In fact, the map t 7→ (γ(t), e(t)) is a smooth path in the frame
bundle F(M ). Define the smooth map ξ : R → Rm by

γ̇ 0 (t) = Φ0γ 0 (t, t0 )Φ0 e0 ξ(t). (3.5.10)

We prove the following.

Claim: The triple (Φ, γ, γ 0 ) is a development on an interval I ⊂ R if and
only if the path t 7→ (γ(t), e(t)) satisfies the differential equation

(γ̇(t), ė(t)) = Bξ(t) (γ(t), e(t)) (3.5.11)

for every t ∈ I, where Bξ(t) ∈ Vect(F(M )) denotes the basic vector field
associated to ξ(t) ∈ Rm (see equation (3.4.9)).
The triple (Φ, γ, γ 0 ) is a development on I if and only if

Φ(t)γ̇(t) = γ̇ 0 (t)

for every t ∈ I. By (3.5.8) and (3.5.10) this is equivalent to the condition

Φ0γ 0 (t, t0 )Φ0 Φγ (t0 , t)γ̇(t) = γ̇ 0 (t) = Φ0γ 0 (t, t0 )Φ0 e0 ξ(t),

hence to
Φγ (t0 , t)γ̇(t) = e0 ξ(t),
and hence to
γ̇(t) = Φγ (t, t0 )e0 ξ(t) = e(t)ξ(t) (3.5.12)
for every t ∈ I. By (3.5.9) and the Gauß–Weingarten formula, we have

ė(t) = hγ(t) (γ̇(t))e(t)

for every t ∈ R. Hence it follows from (3.4.9) that (3.5.12) is equivalent

to (3.5.11). This proves the claim.
Parts (i) and (ii) follow directly from the claim. Part (iii) follows from
the claim and Definition 3.4.12. This proves Theorem 3.5.21.
3.5. MOTIONS AND DEVELOPMENTS 157

Remark 3.5.22. As any two developments (Φ1 , γ1 , γ 0 |I1 ) and (Φ2 , γ2 , γ 0 |I2 )
on two intervals I1 and I2 that satisfy the initial condition (3.5.7) agree
on I1 ∩ I2 there is a development defined on I1 ∪ I2 . Hence there is a
unique maximally defined development (Φ, γ, γ 0 |I ), defined on a maximal
interval I = I(t0 , p0 , Φ0 ), associated to the initial data t0 , p0 , Φ0 .
Denote the space of initial data by
 t ∈ R, p ∈ M, Φ : Tp M → Tγ 0 (t) M 0
  
P := (t, p, Φ)   , (3.5.13)
is an orthogonal transformation

define the set D ⊂ R × P by

D := {(t, t0 , p0 , Φ0 ) | (t0 , p0 , Φ0 ) ∈ P, t ∈ I(t0 , p0 , Φ0 )} (3.5.14)

and let
D → P : (t, t0 , p0 , Φ0 ) 7→ (t, γ(t), Φ(t)), (3.5.15)
be the map which assigns to each (t0 , p0 , Φ0 ) ∈ P and each t ∈ I(t0 , p0 , Φ0 )
the value at time t of the unique development (Φ, γ, γ 0 |I ) associated to the
inital condition (t0 , p0 , Φ0 ) on the maximal time interval I = I(t0 , p0 , Φ0 ).
Then the space P has a natural structure of a smooth manifold (in the
intrinsic setting), and it follows from Theorem 2.4.9 and the proof of Theo-
rem 3.5.21 that D is an open subset of R × P and the map (3.5.5) is smooth.
The smooth structure on P can be understood as follows. The space

O(γ 0 ) = (t, e0 ) | (γ 0 (t), e0 ) ∈ O(M 0 )



is the pullback of the orthonormal frame bundle O(M 0 ) → M 0 under the

curve γ 0 : R → M 0 or, equivalently, is the orthonormal frame bundle of the
pullback tangent bundle (γ 0 )∗ T M . Thus O(γ 0 ) is a smooth submanifold
of R × Rn×m . The group O(m) acts diagonally on O(γ 0 ) × O(M ) and the
action is free. Hence the quotient (O(γ 0 ) × O(M ))/O(m) is a smooth man-
ifold by Theorem 2.9.14, and it can be naturally identified with P via the
bijection [(t, e0 ), (p, e)] 7→ (t, p, e0 ◦ e−1 ).

Remark 3.5.23. The statement of Theorem 3.5.21 is essentially symmetric

in M and M 0 as the operation of inversion carries developments to develop-
ments. Hence given

γ : R → M, p00 ∈ M 0 , t0 ∈ R, Φ0 : Tγ(t0 ) M → Tp00 M 0 ,

we may speak of the development (Φ, γ, γ 0 ) corresponding to γ with initial

conditions γ 0 (t0 ) = p00 and Φ(t0 ) = Φ0 .
158 CHAPTER 3. THE LEVI-CIVITA CONNECTION

Corollary 3.5.24 (Motions). Let p0 ∈ M and t0 ∈ R, let γ 0 : R → M 0 be

a smooth curve, and let Ψ0 ∈ O(n) be a matrix such that

Ψ0 Tp0 M = Tγ 0 (t0 ) M 0 .

Then the following holds.

(i) There exists a motion (Ψ, γ, γ 0 |I ) without sliding, twisting and wobbling
on some open interval I ⊂ R containing t0 that satisfies the initial condi-
tion γ(t0 ) = p0 and Ψ(t0 ) = Ψ0 .
(ii) Any two motions as in (i) on two intervals I1 and I2 agree on the
intersection I1 ∩ I2 .
(iii) If M is complete, then (i) holds with I = R.

Proof. Theorem 3.5.21 and Remark 3.5.15.

Corollary 3.5.25 (Completeness). The following are equivalent.

(i) M is complete, i.e. for every smooth curve ξ : R → Rm and every
element (p0 , e0 ) ∈ F(M ), there exists a smooth curve β : R → F(M ) such
that β̇(t) = Bξ(t) (β(t)) for all t ∈ R and β(0) = (p0 , e0 ) (Definition 3.4.12).
(ii) For every smooth curve ξ : R → Rm and every element (p0 , e0 ) ∈ O(M ),
there is a smooth curve α : R → O(M ) such that α̇(t) = Bξ(t) (α(t)) for ev-
ery t ∈ R and α(0) = (p0 , e0 ).
(iii) For every smooth curve γ 0 : R → Rm , every p0 ∈ M , and every orthogo-
nal isomorphism Φ0 : Tp0 M → Rm there exists a development (Φ, γ, γ 0 ) of M
along M 0 = Rm on all of R that satisfies γ(0) = p0 and Φ(0) = Φ0 .

Proof. We have already noted that the basic vector fields are all tangent to
the orthonormal frame bundle O(M ) ⊂ F(M ). Now note that if a smooth
curve I → F(M ) : t 7→ β(t) = (γ(t), e(t)) on an interval I ⊂ R satisfies the
differential equation
β̇(t) = Bξ(t) (β(t))
for all t, then so does the curve

I → F(M ) 7→ a∗ β(t) = (γ(t), e(t) ◦ a)

for every a ∈ GL(m, R). Since any frame e0 : Rm → Tp0 M can be car-
ried to any other (in particular an orthonormal one) by a suitable ma-
trix a ∈ GL(m, R), this shows that (i) is equivalent to (ii).
That (i) implies (iii) was proved in Theorem 3.5.21.
3.5. MOTIONS AND DEVELOPMENTS 159

We prove that (iii) implies (ii). Fix a smooth map ξ : R → Rm and an

element (p0 , e0 ) ∈ O(M ). Define

Φ0 := e−1
0 : Tp0 M → R
m

and Z t
0
γ (t) := ξ(s) ds ∈ Rm for t ∈ R.
0
By (ii) there exists a development (Φ, γ, γ 0 ) of M along Rm on all of R that
satisfies the initial conditions

γ(0) = p0 , Φ(0) = Φ0 .

Then

Φ(t) = Φ0 Φγ (0, t) : Tγ(t) M → Rm , Φ(t)γ̇(t) = γ̇ 0 (t) = ξ(t)

for all t ∈ R by Definition 3.5.18. Define

e(t) := Φγ (t, 0)e0 = Φ(t)−1 : Rm → Tγ(t) M

for t ∈ R. Then (γ, e) : R → F(M ) is a smooth curve that satisfies the initial
condition (γ(0), e(0)) = (p0 , e0 ) and the differential equation

γ̇(t) = Φ(t)−1 ξ(t) = e(t)ξ(t),

ė(t) = hγ(t) (γ̇(t))e(t) = hγ(t) (e(t)ξ(t))e(t)
by the Gauß–Weingarten formula. Thus

(γ̇(t), ė(t)) = Bξ(t) (γ(t), e(t))

for all t ∈ R. This proves Corollary 3.5.25.

It is of course easy to to give an example of a manifold which is not com-

plete; e.g. if (Φ, γ, γ 0 ) is any development of M along M 0 , then M \ {γ(t1 )}
is not complete as the given development is only defined for t 6= t1 . In §4.6
we give equivalent characterizations of completeness. In particular, we will
see that any compact submanifold of Rn is complete.
Exercise 3.5.26. An affine subspace of Rn is a subset of the form
 
E = p + E = p + vv ∈ E

where E ⊂ Rn is a linear subspace and p ∈ Rn . Prove that every affine

subspace of Rn is a complete submanifold.
160 CHAPTER 3. THE LEVI-CIVITA CONNECTION

3.6 Christoffel Symbols

The goal of this subsection is to examine the covariant derivative in local
coordinates on an embedded manifold M ⊂ Rn of dimension m. Let

φ:U →Ω

be a coordinate chart, defined on an M -open subset U ⊂ M with values in

an open set Ω ⊂ Rm , and denote its inverse by

ψ := φ−1 : Ω → U ⊂ M.

At this point it is convenient to use superscripts for the coordinates of a

vector x ∈ Ω. Thus we write

x = (x1 , . . . , xm ) ∈ Ω.

If p = ψ(x) ∈ U is the corresponding element of M , then the tangent space

of M at p is the image of the linear map dψ(x) : Rm → Rn (Theorem 2.2.3)
and thus two tangent vectors v, w ∈ Tp M can be written in the form
m
X ∂ψ
v = dψ(x)ξ = ξi (x),
∂xi
i=1
m (3.6.1)
X
j∂ψ
w = dψ(x)η = η (x)
∂xj
j=1

for ξ = (ξ 1 , . . . , ξ m ) ∈ Rm and η = (η 1 , . . . , η m ) ∈ Rm . Recall that the re-

striction of the inner product in the ambient space Rn to the tangent space
is the first fundamental form gp : Tp M × Tp M → R (Definition 3.1.1). Thus
m
X
gp (v, w) = hv, wi = ξ i gij (x)η j , (3.6.2)
i,j=1

where the functions gij : Ω → R are defined by

 
∂ψ ∂ψ
gij (x) := i
(x), j (x) for x ∈ Ω. (3.6.3)
∂x ∂x

In other words, the first fundamental form is in local coordinates represented

by the matrix valued function g = (gij )mi,j=1 : Ω → R
m×m .
3.6. CHRISTOFFEL SYMBOLS 161

X(t)
γ(t)
U
M ψ
φ
Ω
ξ(t)
c(t)

Figure 3.7: A vector field along a curve in local coordinates.

Now let c = (c1 , . . . , cm ) : I → Ω be a smooth curve in Ω, defined on an

interval I ⊂ R, and consider the curve

γ =ψ◦c:I →M

(see Figure 3.7). Our goal is to describe the operator X 7→ ∇X on the space
of vector fields along γ in local coordinates. Let X : I → Rn be a vector
field along γ. Then

X(t) ∈ Tγ(t) M = Tψ(c(t)) M = im dψ(c(t)) : Rm → Rn

for every t ∈ I and hence there exists a unique smooth function

ξ = (ξ 1 , . . . , ξ m ) : I → Rm

such that
m
X ∂ψ
X(t) = dψ(c(t))ξ(t) = ξ i (t) (c(t)). (3.6.4)
∂xi
i=1

Differentiate this identity to obtain

m m
X ∂ψ X ∂2ψ
Ẋ(t) = ξ˙i (t) i (c(t)) + ξ i (t)ċj (t) (c(t)). (3.6.5)
∂x ∂xi ∂xj
i=1 i,j=1

We examine the projection ∇X(t) = Π(γ(t))Ẋ(t) of this vector onto the tan-
gent space of M at γ(t). The first summand on the right in (3.6.5) is already
162 CHAPTER 3. THE LEVI-CIVITA CONNECTION

tangent to M . For the second summand we simply observe that the vec-
tor Π(ψ(x))∂ 2 ψ/∂xi ∂xj lies in tangent space Tψ(x) M and can therefore be
expressed as a linear combination of the basis vectors ∂ψ/∂x1 , . . . , ∂ψ/∂xm .
The coefficients will be denoted by Γkij (x). Thus there exist smooth func-
tions Γkij : Ω → R for i, j, k = 1, . . . , m such that

m
∂2ψ X ∂ψ
Π(ψ(x)) i j
(x) = Γkij (x) k (x) (3.6.6)
∂x ∂x ∂x
k=1

for all x ∈ Ω and all i, j ∈ {1, . . . , m}. The coefficients Γkij : Ω → R are called
the Christoffel symbols associated to the coordinate chart φ : U → Ω. To
sum up we have proved the following.

Lemma 3.6.1. Let c : I → Ω be a smooth curve and define

γ := ψ ◦ c : I → M.

If ξ : I → Rm is a smooth map and X ∈ Vect(γ) is given by (3.6.4), then

its covariant derivative at time t ∈ I is given by
 
m m
X
ξ˙k (t) +
X ∂ψ
∇X(t) = Γkij (c(t))ξ i (t)ċj (t) (c(t)), (3.6.7)
∂xk
k=1 i,j=1

where the Γkij are the Christoffel symbols defined by (3.6.6).

Our next goal is to understand how the Christoffel symbols are deter-
mined by the metric in local coordinates. Recall from equation (3.6.2) that
the inner products on the tangent spaces inherited from the standard Eu-
clidean inner product on the ambient space Rn are in local coordinates
represented by the matrix valued function

g = (gij )m
i,j=1 : Ω → R
m×m

given by
 
∂ψ ∂ψ
gij := , . (3.6.8)
∂xi ∂xj Rn

We shall see that the Christoffel symbols are completely determined by the
functions gij : Ω → R. Here are first some elementary observations.
3.6. CHRISTOFFEL SYMBOLS 163

Remark 3.6.2. The matrix g(x) ∈ Rm×m is symmetric and positive definite
for every x ∈ Ω. This follows from the fact that the matrix dψ(x) ∈ Rn×m
has rank m and the matrix g(x) is given by
g(x) = dψ(x)T dψ(x)
Thus ξ T g(x)ξ = |dψ(x)ξ|2 > 0 for all ξ ∈ Rm \ {0}.
Remark 3.6.3. For x ∈ Ω we have det(g(x)) > 0 by Remark 3.6.2 and
so the matrix g(x) is invertible. Denote the entries of the inverse ma-
trix g(x)−1 ∈ Rm×m by g k` (x). They are determined by the condition
m 
X
jk k 1, if i = k,
gij (x)g (x) = δi =
0, if i 6= k.
j=1

Since g(x) is symmetric and positive definite, so is its inverse matrix g(x)−1 .
In particular, we have g k` (x) = g `k (x) for all x ∈ Ω and all k, ` ∈ {1, . . . , m}.
Remark 3.6.4. Suppose that X, Y ∈ Vect(γ) are vector fields along our
curve γ = ψ ◦ c : I → M and ξ, η : I → Rm are defined by
m m
X
i ∂ψ X ∂ψ
X(t) = ξ (t) i (c(t)), Y (t) = η j (t) (c(t)).
∂x ∂xj
i=1 j=1

Then the inner product of X(t) and Y (t) is given by

m
X
hX(t), Y (t)i = ξ i (t)gij (c(t))η j (t).
i,j=1

Lemma 3.6.5 (Christoffel Symbols). Let Ω ⊂ Rm be an open set and

let gij : Ω → R for i, j = 1, . . . , m be smooth functions such that each ma-
trix (gij (x))m k
i,j=1 is symmetric and positive definite. Let Γij : Ω → R be
smooth functions for i, j, k = 1, . . . , m. Then the Γkij satisfy the conditions
m
∂gij X 
Γkij = Γkji , = gik Γk
j` + gjk Γk
i` (3.6.9)
∂x`
k=1

for i, j, k, ` = 1, . . . , m if and only if they are given by

m  
k
X
k` 1 ∂g`i ∂g`j ∂gij
Γij = g + − . (3.6.10)
2 ∂xj ∂xi ∂x`
`=1

If the Γkij are defined by (3.6.6) and the gij by (3.6.8), then the Γkij sat-
isfy (3.6.9) and hence are given by (3.6.10).
164 CHAPTER 3. THE LEVI-CIVITA CONNECTION

Proof. Suppose that the Γkij are given by (3.6.6) and the gij by (3.6.8). Let

c : I → Ω, ξ, η : I → Rm

be smooth functions and suppose that the vector fields X, Y along the curve

γ := ψ ◦ c : I → M

are given by
m m
X ∂ψ X ∂ψ
X(t) := ξ i (t) (c(t)), Y (t) := η j (t) (c(t)).
∂xi ∂xj
i=1 j=1

Dropping the argument t in each term, we obtain from Remark 3.6.4 and
Lemma 3.6.1 that
X
hX, Y i = gij (c)ξ i η j ,
i,j
 
X X
hX, ∇Y i = gik (c)ξ i η̇ k + Γkj` (c)η j ċ`  ,
i,k j,`
 
X X
h∇X, Y i = gkj (c) ξ˙k + Γki` (c)ξ i ċ`  η j .
j,k i,`

Hence it follows from equation (3.2.5) in Lemma 3.2.4 that

d
0 = hX, Y i − hX, ∇Y i − h∇X, Y i
dt !
X X ∂gij
= gij (c)ξ˙ η + gij (c)ξ η̇ +
i j i j i j
(c)ξ η ċ `
∂x`
i,j `
X X
i k
− gik (c)ξ η̇ − gik (c)Γj` (c)ξ i η j ċ`
k

i,k i,j,k,`
X X
− gkj (c)ξ˙k η j − gkj (c)Γki` (c)ξ i η j ċ`
j,k i,j,k,`
!
X ∂gij X X
= (c) − gik (c)Γkj` (c) − gjk (c)Γki` (c) ξ i η j ċ` .
∂x`
i,j,` k k

This holds for all smooth maps c : I → Ω and ξ, η : I → Rm , so the Γkij satisfy
the second equation in (3.6.9). That they are symmetric in i and j is obvious.
3.6. CHRISTOFFEL SYMBOLS 165

To prove that (3.6.9) is equivalent to (3.6.10), define

m
X
Γ`ij := g`k Γkij . (3.6.11)
k=1

Then (3.6.9) is equivalent to

∂gij
Γ`ij = Γ`ji , = Γij` + Γji` . (3.6.12)
∂x`

and (3.6.10) is equivalent to

 
1 ∂g`i ∂g`j ∂gij
Γ`ij = j
+ i
− . (3.6.13)
2 ∂x ∂x ∂x`

If the Γ`ij are given by (3.6.13), then they satisfy

Γ`ij = Γ`ji

and

∂gij ∂gi` ∂gj` ∂gji ∂gj` ∂gi`

2Γij` + 2Γji` = `
+ − + + −
∂x ∂xj ∂xi ∂x` ∂xi ∂xj
∂gij
=2 `
∂x

for all i, j, `. Conversely, if the Γ`ij satisfy (3.6.12), then

∂gij
= Γij` + Γji` ,
∂x`
∂g`i
= Γ`ij + Γi`j = Γ`ij + Γij` ,
∂xj
∂g`j
= Γ`ji + Γj`i = Γ`ij + Γji` .
∂xi
Take the sum of the last two minus the first of these equations to obtain

∂g`i ∂g`j ∂gij

+ − = 2Γ`ij .
∂xj ∂xi ∂x`

Thus (3.6.12) is equivalent to (3.6.13) and so (3.6.9) is equivalent to (3.6.10).

This proves Lemma 3.6.5.
166 CHAPTER 3. THE LEVI-CIVITA CONNECTION

3.7 Riemannian Metrics*

We wish to carry over the fundamental notions of differential geometry to
the intrinsic setting. First we need an inner product on the tangent spaces
to replace the first fundamental form in Definition 3.1.1. This is the content
of Definition 3.7.1 and Lemma 3.7.4 below. Second we must introduce the
covariant derivative of a vector field along a curve. With this understood all
the definitions, theorems, and proofs in this chapter carry over in an almost
word by word fashion to the intrinsic setting.

3.7.1 Existence of Riemannian Metrics

We will always consider norms that are induced by inner products. But in
general there is no ambient space that can induce an inner product on each
tangent space. This leads to the following definition.
Definition 3.7.1. Let M be a smooth m-manifold. A Riemannian metric
on M is a collection of inner products
Tp M × Tp M → R : (v, w) 7→ gp (v, w),
one for every p ∈ M , such that the map
M → R : p 7→ gp (X(p), Y (p))
is smooth for every pair of vector fields X, Y ∈ Vect(M ). We will also
denote the inner product by hv, wip and drop the subscript p if the base
point is understood from the context. A smooth manifold equipped with a
Riemannian metric is called a Riemannian manifold.
Example 3.7.2. If M ⊂ Rn is a smooth submanifold, then a Riemannian
metric on M is given by restricting the standard inner product on Rn to
the tangent spaces Tp M ⊂ Rn . This is the first fundamental form of an
embedded manifold (see Definition 3.1.1).
More generally, assume that M is a Riemannian m-manifold in the in-
trinsic sense of Definition 3.7.1 with an atlas A = {(φα , Uα )}α∈A . Then the
Riemannian metric g determines a collection of smooth functions
gα = (gα,ij )m
i,j=1 : φα (Uα ) → R
m×m
, (3.7.1)
one for each α ∈ A, defined by
ξ T gα (x)η := gp (v, w), φα (p) = x, dφα (p)v = ξ, dφα (p)w = η, (3.7.2)
for x ∈ φα (Uα ) and ξ, η ∈ Rm .
3.7. RIEMANNIAN METRICS* 167

Each matrix gα (x) is symmetrix and positive definite. Note that the
tangent vectors v and w in (3.7.2) can also be written in the form

v = [α, ξ]p , w = [α, η]p .

Choosing standard basis vectors

ξ = ei , η = ej

in Rm we obtain
∂
[α, ei ]p = dφα (p)−1 ei =: (p)
∂xi
and hence  
∂ −1 ∂ −1
gα,ij (x) = (φ (x)), j (φα (x)) . (3.7.3)
∂xi α ∂x
For different coordinate charts the maps gα and gβ are related through the
transition map

φβα := φβ ◦ φ−1
α : φα (Uα ∩ Uβ ) → φβ (Uα ∩ Uβ )

via
gα (x) = dφβα (x)T gβ (φβα (x))dφβα (x) (3.7.4)

for x ∈ φα (Uα ∩ Uβ ). Equation (3.7.4) can also be written in the shorthand

notation
gα = φ∗βα gβ

for α, β ∈ A.

Exercise 3.7.3. Every collection of smooth maps

gα : φα (Uα ) → Rm×m

with values in the set of positive definite symmetric matrices that satis-
fies (3.7.4) for all α, β ∈ A determines a global Riemannian metric via (3.7.2).

In this intrinsic setting there is no canonical metric on M (such as the

metric induced by Rn on an embedded manifold). In fact, it is not completely
obvious that a manifold admits a Riemannian metric and this is the content
of the next lemma.
168 CHAPTER 3. THE LEVI-CIVITA CONNECTION

Lemma 3.7.4. Every paracompact Hausdorff manifold admits a Rieman-

nian metric.

Proof. Let m be the dimension of M and let A = {(φα , Uα )}α∈A be an atlas

on M . By Theorem 2.9.9 there is a partition of unity {θα }α∈A subordinate
to the open cover {Uα }α∈A . Now there are two equivalent ways to construct
a Riemannian metric on M .
The first method is to carry over the standard inner product on Rm to
the tangent spaces Tp M for p ∈ Uα via the coordinate chart φα , multiply
the resulting Riemannian metric on Uα by the compactly supported function
θα , extend it by zero to all of M , and then take the sum over all α. This
leads to the following formula. The inner product of two tangent vectors
v, w ∈ Tp M is defined by
X
hv, wip := θα (p)hdφα (p)v, dφα (p)wi, (3.7.5)
p∈Uα

where the sum runs over all α ∈ A with p ∈ Uα and the inner product is
the standard inner product on Rm . Since supp(θα ) ⊂ Uα for each α and the
sum is locally finite we find that the function

M → R : p 7→ hX(p), Y (p)ip

is smooth for every pair of vector fields X, Y ∈ Vect(M ). Moreover, the right
hand side of (3.7.5) is symmetric in v and w and is positive for v = w 6= 0
because each summand is nonnegative and each summand with θα (p) > 0 is
positive. Thus equation (3.7.5) defines a Riemannian metric on M .
The second method is to define the functions

gα : φα (Uα ) → Rm×m

by
X
gα (x) := θγ (φ−1 T
α (x))dφγα (x) dφγα (x) (3.7.6)
γ∈A

for x ∈ φα (Uα ) where each summand is defined on φα (Uα ∩ Uγ ) and is

understood to be zero for x ∈ / φα (Uα ∩ Uγ ). We leave it to the reader to
verify that these functions are smooth and satisfy the condition (3.7.4) for
all α, β ∈ A. Moreover, the formulas (3.7.5) and (3.7.6) determine the same
Riemannian metric on M . (Prove this!) This proves Lemma 3.7.4.
3.7. RIEMANNIAN METRICS* 169

3.7.2 Two Examples

Example 3.7.5 (Fubini–Study Metric). The complex projective space
carries a natural Riemannian metric, defined as follows. Identify CPn with
the quotient of the unit sphere S 2n+1 ⊂ Cn+1 by the diagonal action of
the circle S 1 , i.e. CPn = S 2n+1 /S 1 . Then the tangent space of CPn at the
equivalence class
[z] = [z0 : · · · : zn ] ∈ CPn
of a point z = (z0 , . . . , zn ) ∈ S 2n+1 can be identified with the orthogonal
complement of Cz in Cn+1 . Now choose the inner product on T[z] CPn to
be the one inherited from the standard inner product on Cn+1 via this
identification. The resulting metric on CPn is called the Fubini–Study
metric. Exercise: Prove that the action of U(n + 1) on Cn+1 induces a
transitive action of the quotient group
PSU(n + 1) := U(n + 1)/S 1
by isometries. If z ∈ S 2n+1 , prove that the unitary matrix
g := 2zz ∗ − 1l
descends to an isometry φ on CPn with fixed point p := [z] and dφ(p) = −id.
Show that, in the case n = 1, the pullback of the Fubini–Study metric on CP1
under the stereographic projection
 
x1 + ix2
S 2 \ {(0, 0, 1)} → CP1 \ {[0 : 1]} : (x1 , x2 , x3 ) 7→ 1 :
1 − x3
is one quarter of the standard metric on S 2 .
Example 3.7.6. Think of the complex Grassmannian Gk (Cn ) of k-planes
in Cn as a quotient of the space
n o
Fk (Cn ) := D ∈ Cn×k | D∗ D = 1l

of unitary k-frames in Cn by the right action of the unitary group U(k).

The space Fk (Cn ) inherits a Riemannian metric from the ambient Euclidean
space Cn×k . Show that the tangent space of Gk (Cn ) at a point Λ = imD,
with D ∈ Fk (Cn ) can be identified with the space
n o
TΛ Gk (Cn ) = Db ∈ Cn×k | D∗ D
b =0 .

Define the inner product on this tangent space to be the restriction of the
standard inner product on Cn×k to this subspace. Exercise: Prove that
the unitary group U(n) acts on Gk (Cn ) by isometries.
170 CHAPTER 3. THE LEVI-CIVITA CONNECTION

3.7.3 The Levi-Civita Connection

A subtle point in this discussion is how to extend the notion of covariant
derivative to general Riemannian manifolds. In this case the idea of project-
ing the derivative in the ambient space orthogonally onto the tangent space
has no obvious analogue. Instead we shall see how the covariant derivatives
of vector fields along curves can be characterized by several axioms and
these can be used to define the covariant derivative in the intrinsic setting.
An alternative, but somewhat less satisfactory, approach is to carry over
the formula for the covariant derivative in local coordinates to the intrinsic
setting and show that the result is independent of the choice of the coor-
dinate chart. Of course, these approaches are equivalent and lead to the
same result. We formulate them as a series of exercises. The details are
straightforward.
Assume throughout that M is a Riemannian m-manifold with an atlas

A = {(φα , Uα )}α∈A

and suppose that the Riemannian metric is in local coordinates given by

gα = (gα,ij )m
i,j=1 : φα (Uα ) → R
m×m

for α ∈ A. These functions satisfy (3.7.4) for all α, β ∈ A.

Definition 3.7.7. Let f : N → M be a smooth map between manifolds. A

vector field along f is a collection of tangent vectors

X(q) ∈ Tf (q) M,

one for each q ∈ N , such that the map

N → T M : q 7→ (f (q), X(q))

is smooth. The space of vector fields along f will be denoted by Vect(f ).

As before we will not distinguish in notation between the collection of

tangent vectors X(q) ∈ Tf (q) M and the associated map N → T M and
denote them both by X. The following theorem introduces the Levi-Civita
connection as a collection of linear operators ∇ : Vect(γ) → Vect(γ), one for
each smooth curve γ : I → M .
3.7. RIEMANNIAN METRICS* 171

Theorem 3.7.8 (Levi-Civita Connection). There exists a unique col-

lection of linear operators

∇ : Vect(γ) → Vect(γ)

(called the covariant derivative), one for every smooth curve γ : I → M

on an open interval I ⊂ R, satisfying the following axioms.
(Leibniz Rule) For every smooth curve γ : I → M , every smooth function
λ : I → R, and every vector field X ∈ Vect(γ), we have

∇(λX) = λ̇X + λ∇X. (3.7.7)

(Chain Rule) Let Ω ⊂ Rn be an open set, let c : I → Ω be a smooth curve,

let γ : Ω → M be a smooth map, and let X be a smooth vector field along γ.
Denote by ∇i X the covariant derivative of X along the curve xi 7→ γ(x) (with
the other coordinates fixed). Then ∇i X is a smooth vector field along γ and
the covariant derivative of the vector field X ◦ c ∈ Vect(γ ◦ c) is
n
X
∇(X ◦ c) = ċj (t)∇j X(c(t)). (3.7.8)
j=1

(Riemannian) For any two vector fields X, Y ∈ Vect(γ) we have

d
hX, Y i = h∇X, Y i + hX, ∇Y i. (3.7.9)
dt

(Torsion-free) Let I, J ⊂ R be open intervals and γ : I × J → M be

a smooth map. Denote by ∇s the covariant derivative along the curve
s 7→ γ(s, t) (with t fixed) and by ∇t the covariant derivative along the curve
t 7→ γ(s, t) (with s fixed). Then

∇s ∂t γ = ∇t ∂s γ. (3.7.10)

Proof. The proof is based on a reformulation of the axioms in local co-

ordinates. The (Leibnitz Rule) and (Chain Rule) axioms assert that the
covariant derivative is in local coordinates given by Christoffel symbols Γkij
as in equation (3.6.7) in Lemma 3.6.1. The (Riemannian) and (Torsion-free)
axioms assert that the Christoffel symbols satisfy the equations in (3.6.9)
and hence, by Lemma 3.6.5, are given by (3.6.10). (See also Exercise 3.7.10.)
This proves Theorem 3.7.8.
172 CHAPTER 3. THE LEVI-CIVITA CONNECTION

Exercise 3.7.9. The Christoffel symbols of the Riemannian metric are

the functions Γkα,ij : φα (Uα ) → R. defined by
m  
X 1 ∂gα,`i ∂gα,`j ∂gα,ij
Γkα,ij := gαk` j
+ i
− (3.7.11)
2 ∂x ∂x ∂x`
`=1

(see Lemma 3.6.5). Prove that they are related by the equation

X ∂φkβα
0
∂ 2 φkβα
0
X 0
 ∂φi0 ∂φj 0
βα βα
Γk = + Γkβ,i0 j 0 ◦ φβα .
∂xk α,ij ∂xi ∂xj ∂xi ∂xj
k i0 ,j 0

for all α, β ∈ A.
Exercise 3.7.10. Denote ψα := φ−1α : φα (Uα ) → M . Prove that the covari-
ant derivative of a vector field
m
X ∂ψα
X(t) = ξαi (t) (cα (t))
∂xi
i=1

along γ = ψα ◦ cα : I → M is given by
 
m m
X
ξ˙αk (t) +
X ∂ψα
∇X(t) = Γkα,ij (c(t))ξαi (t)ċjα (t) (cα (t)). (3.7.12)
∂xk
k=1 i,j=1

Prove that ∇X is independent of the choice of the coordinate chart.

Exercise 3.7.11. Let Ω ⊂ R2 be open and λ : Ω → (0, ∞) be a smooth
function. Let g : Ω → R2×2 be given by
 
λ(x) 0
g(x) = .
0 λ(x)

Compute the Christoffel symbols Γkij via (3.6.10).

Exercise 3.7.12. Let φ : S 2 \ {(0, 0, 1)} → C be the stereographic projec-

tion, given by  
p1 p2
φ(p) := ,
1 − p3 1 − p3
Prove that the metric g : R2 → R2×2 has the form g(x) = λ(x)1l where
4
λ(x) := for x = (x1 , x2 ) ∈ R2 .
(1 + |x|2 )2
3.7. RIEMANNIAN METRICS* 173

3.7.4 Basic Vector Fields in the Intrinsic Setting

Let M be a Riemannian m-manifold with an atlas A = {(φα , Uα )}α∈A .
Then the frame bundle (3.4.1) admits the structure of a smooth manifold
eα := π −1 (Uα ) and coordinate charts
with the open cover U
eα → φα (Uα ) × GL(m)
φeα : U
given by
φeα (p, e) := (φα (p), dφα (p)e) .
The derivatives of the horizontal curves in Definition 3.4.6 form a horizon-
tal subbundle H ⊂ T F(M ) of the tangent bundle of the Frame bundle
whose fibers H(p,e) over an element (p, e) ∈ F(M ) can in local coordinates
be described as follows. Let
x := φα (p), a := dφα (p)e ∈ GL(m). (3.7.13)
and let (b a) ∈ Rm × Rm×m . This pair has the form
x, b
(b
x, b
a) = dφeα (p, e)(b
p, eb), p, eb) ∈ H(p,e) ,
(b (3.7.14)
if and only if
m
xi aj`
X
ak`
b =− Γkα,ij (x)b (3.7.15)
i,j=1

for k, ` = 1, . . . , m, where the functions Γkα,ij : φα (Uα ) → R are the Christof-

fel symbols defined by (3.7.11). Thus a tangent vector (b p, eb) ∈ T(p,e) F(M )
is horizontal if and only if its coordinates (b x, b
a) in (3.7.14) satisfy (3.7.15).
Hence, for every vector ξ ∈ Rm , there exists a unique horizontal vector
field Bξ ∈ Vect(F(M )) (the basic vector field associated to ξ) such that
dπ(p, e)Bξ (p, e) = eξ
for all (p, e) ∈ F(M ). This vector field assigns to a pair (p, e) ∈ F(M ) with
the coordinates (x, a) ∈ Rm × GL(m) as in (3.7.13) the horizontal tangent
vector (bp, eb) ∈ H(p,e) ⊂ T(p,e) F(M ) whose coordinates (b a) ∈ Rm × Rm×m
x, b
satisfy (3.7.15) and x b = aξ.
Exercise 3.7.13. Verify the equivalence of (3.7.14) and (3.7.15). Prove
that the notion of a horizontal tangent vector of F(M ) is independent of
the choice of the coordinate chart. Hint: Use Exercise 3.7.9.
Exercise 3.7.14. Examine the orthonormal frame bundle O(M ) in the
intrinsic setting.
Exercise 3.7.15. Carry over the proofs of Theorem 3.3.4, Theorem 3.3.6,
and Theorem 3.5.21 to the intrinsic setting.
174 CHAPTER 3. THE LEVI-CIVITA CONNECTION
Chapter 4

Geodesics

This chapter introduces geodesics in Riemannian manifolds. It begins in §4.1

by introducing geodesics as extremals of the energy and length functionals
and characterizing them as solutions of a second order differential equation.
In §4.2 we show that minimizing the length with fixed endpoints gives rise
to an intrinsic distance function d : M × M → R which induces the topol-
ogy M inherits from the ambient space Rn . §4.3 introduces the exponential
map, §4.4 shows that geodesics minimize the length on short time intervals,
§4.5 establishes the existence of geodesically convex neighborhoods, and §4.6
shows that the geodesic flow is complete if and only if (M, d) is a complete
metric space, and that in the complete case any two points are joined by a
minimal geodesic. §4.7 discusses geodesics in the intrinsic setting.

4.1 Length and Energy

4.1.1 The Length and Energy Functionals
The concept of a geodesic in a manifold generalizes that of a straight line in
Euclidean space. A straight line has parametrizations of form t 7→ p + σ(t)v
where σ : R → R is a diffeomorphism and p, v ∈ Rn . Different choices of σ
yield different parametrizations of the same line. Certain parametrizations
are preferred, for example those parametrizations which are “proportional
to the arclength”, i.e. where σ(t) = at + b for constants a, b ∈ R, so that
the tangent vector σ̇(t)v has constant length. The same distinctions can
be made for geodesics. Some authors use the term geodesic to include all
parametrizations of a geodesic while others restrict the term to cover only
geodesics parametrized proportional to the arclength. We follow the latter

175
176 CHAPTER 4. GEODESICS

course, referring to the more general concept as a “reparametrized geodesic”.

Thus a reparametrized geodesic need not be a geodesic.
We assume throughout that M ⊂ Rn is a smooth m-manifold.
Definition 4.1.1 (Length and Energy). Let I = [a, b] ⊂ R be a com-
pact interval with a < b and let γ : I → M be a smooth curve in M . The
length L(γ) and the energy E(γ) are defined by
Z b
L(γ) := |γ̇(t)| dt, (4.1.1)
a

1 b
Z
E(γ) := |γ̇(t)|2 dt. (4.1.2)
2 a
A variation of γ is a family of smooth curves γs : I → M , where s
ranges over the reals, such that the map R × I → M : (s, t) 7→ γs (t) is
smooth and γ0 = γ. The variation {γs }s∈R is said to have fixed endpoints
iff γs (a) = γ(a) and γs (b) = γ(b) for all s ∈ R.
Remark 4.1.2. The length of a continuousPfunction γ : [a, b] → Rn can
be defined as the supremum of the numbers N i=1 |γ(ti ) − γ(ti−1 )| over all
partitions a = t0 < t1 < · · · < tN = b of the interval [a, b]. By a theorem in
first year analysis [47] this supremum is finite whenever γ is continuously
differentiable and is given by (4.1.1).
We shall sometimes suppress the notation for the endpoints of a, b ∈ I.
When γ(a) = p and γ(b) = q we say that γ is a curve from p to q. One
can always compose γ with an affine reparametrization t0 = a + (b − a)t to
obtain a new curve γ 0 (t) := γ(t0 ) on the unit interval 0 ≤ t ≤ 1. This new
curve satisfies L(γ 0 ) = L(γ) and E(γ 0 ) = (b − a)E(γ). More generally, the
length L(γ), but not the energy E(γ), is invariant under reparametrization.
Remark 4.1.3 (Reparametrization). Let I = [a, b] and I 0 = [a0 , b0 ] be
compact intevrals. If γ : I → Rn is a smooth curve and σ : I 0 → I is a
smooth function such that σ(a0 ) = a, σ(b0 ) = b, and σ̇(t) ≥ 0 for all t ∈ I 0 ,
then
L(γ ◦ σ) = L(γ). (4.1.3)
To see this, we compute
Z b0   Z b0 
 d 0 
 0 0 

L(γ ◦ σ) = γ(σ(t )) dt = γ̇(σ(t ))σ̇(t0 ) dt0 = L(γ).
 
 0
a0 dt

a0


Here second equation follows from the chain rule and the fact that σ̇(t0 ) ≥ 0
for all t0 ∈ [a0 , b0 ], and the third equation follows from the change of variables
formula for the Riemann integral. This proves equation (4.1.3).
4.1. LENGTH AND ENERGY 177

Theorem 4.1.4 (Characterization of Geodesics). Let I = [a, b] ⊂ R be

a compact interval and let γ : I → M be a smooth curve. Then the following
are equivalent.
(i) γ is an extremal of the energy functional, i.e.

d 
E(γs ) = 0
ds s=0
for every variation {γs }s∈R of γ with fixed endpoints.
(ii) γ is parametrized proportional to the arclength, i.e. the veloc-
ity |γ̇(t)| ≡ c ≥ 0 is constant, and either γ is constant, i.e. γ(t) = p = q for
all t ∈ I, or c > 0 and γ is an extremal of the length functional, i.e.

d 
L(γs ) = 0
ds s=0
for every variation {γs }s∈R of γ with fixed endpoints.
(iii) The velocity vector of γ is parallel, i.e. ∇γ̇(t) = 0 for all t ∈ I.
(iv) The acceleration of γ is normal to M , i.e. γ̈(t) ⊥ Tγ(t) M for all t ∈ I.
(v) If (Φ, γ, γ 0 ) is a development of M along M 0 = Rm , then γ 0 : I → Rm is
a straight line parametrized proportional to the arclength, i.e. γ̈ 0 ≡ 0.
Proof. See page 180.

Definition 4.1.5 (Geodesic). A smooth curve γ : I → M on an interval I

is called a geodesic iff its restriction to each compact subinterval satisfies
the equivalent conditions of Theorem 4.1.4. So γ is a geodesic if and only if

∇γ̇(t) = 0 for all t ∈ I. (4.1.4)

By the Gauß–Weingarten formula (3.2.2) with X = γ̇ this is equivalent to

γ̈(t) = hγ(t) (γ̇(t), γ̇(t)) for all t ∈ I. (4.1.5)

Remark 4.1.6. (i) The conditions (i) and (ii) in Theorem 4.1.4 are mean-
ingless when I is not compact because then the curve has at most one
endpoint and the length and energy integrals may be infinite. However, the
conditions (iii), (iv), and (v) in Theorem 4.1.4 are equivalent for smooth
curves on any interval, compact or not.
(ii) The function s 7→ E(γs ) associated to a smooth variation is always
smooth and so condition (i) in Theorem 4.1.4 is meaningful. However, more
care has to be taken in part (ii) because the function s 7→ L(γs ) need not be
differentiable. It is differentiable at s = 0 whenever γ̇(t) 6= 0 for all t ∈ I.
178 CHAPTER 4. GEODESICS

4.1.2 The Space of Paths

Fix two points p, q ∈ M and a compact interval I = [a, b] and denote by
 
Ωp,q := Ωp,q (I) := γ : I → M  γ is smooth and γ(a) = p, γ(b) = q

the space of smooth curves in M from p to q, defined on the interval I. Then

the length and energy are functionals L, E : Ωp,q → R and their extremal
points can be understood as critical points as we now explain.
We may think of the space Ωp,q as a kind of “infinite-dimensional man-
ifold”. This is to be understood in a heuristic sense and we use these terms
here to emphasize an analogy. Of course, the space Ωp,q is not a manifold
in the strict sense of the word. To begin with it is not embedded in any
finite-dimensional Euclidean space. However, it has many features in com-
mon with manifolds. The first is that we can speak of smooth curves in Ωp,q .
Of course Ωp,q is itself a space of curves in M . Thus a smooth curve in Ωp,q
would then be a curve of curves, namly a map R → Ωp,q : s 7→ γs that as-
signs to each real number s a smooth curve γs : I → M satisfying γs (a) = p
and γs (b) = q. We shall call such a curve of curves smooth iff the associated
map R × I → M : (s, t) 7→ γs (t) is smooth. Thus smooth curves in Ωp,q are
the variations of γ with fixed endpoints introduced in Definition 4.1.1.
Having defined what we mean by a smooth curve in Ωp,q we can also
differentiate such a curve with respect to s. Here we can simply recall that,
since M ⊂ Rn , we have a smooth map R × I → Rn and the derivative of
the curve s 7→ γs in Ωp,q can simply be understood as the partial derivative
of the map (s, t) 7→ γs (t) with respect to s. Thus, in analogy with embed-
ded manifolds, we define the tangent space of the space of curves Ωp,q
at γ as the set of all derivatives of smooth curves R → Ωp,q : s 7→ γs passing
through γ, i.e.
   
∂  
Tγ Ωp,q := γs R → Ωp,q : s 7→ γs is smooth and γ0 = γ .

∂s s=0 
∂ 

Let us denote such a partial derivative by X(t) := ∂s γ (t) ∈ Tγ(t) M .
s=0 s
Thus we obtain a smooth vector field along γ. Since γs (a) = p and γs (b) = q
for all s, this vector field must vanish at t = a, b. This suggests the formula

Tγ Ωp,q = {X ∈ Vect(γ) | X(a) = 0, X(b) = 0} . (4.1.6)

That every tangent vector of the path space Ωp,q at γ is a vector field along γ
vanishing at the endpoints follows from the above discussion. The converse
inclusion is the content of the next lemma.
4.1. LENGTH AND ENERGY 179

Lemma 4.1.7. Let p, q ∈ M , γ ∈ Ωp,q , and X ∈ Vect(γ) with X(a) = 0

and X(b) = 0. Then there exists a smooth map R → Ωp,q : s 7→ γs such that

∂ 
γ0 (t) = γ(t), γs (t) = X(t) for all t ∈ I. (4.1.7)
∂s  s=0
Proof. The proof has two steps.
Step 1. There exists smooth map M × I → Rn : (r, t) 7→ Yt (r) with compact
support such that Yt (r) ∈ Tr M for all t ∈ I and r ∈ M , Yt (r) = 0 for all t ∈ I
and all r ∈ M \ K, and Ya (r) = Yb (r) = 0 for all r ∈ M .
Define Zt (r) := Π(r)X(t) for t ∈ I and r ∈ M . Choose an open n
S set U ⊂ R
such that γ(I) ⊂ U and U ∩ M is compact (e.g. take U := a≤t≤b Bε (γ(t))
for ε > 0 sufficiently small). Now let β : Rn → [0, 1] be a smooth cutoff
function with support in the unit ball such that β(0) = 1 and define the
vector fields Yt by Yt (r) := β(ε−1 (r − γ(t)))Zt (r) for t ∈ I and r ∈ M .
Step 2. We prove the lemma.
The vector field Yt : M → T M in Step 1 is complete for each t. Thus
there exists a unique smooth map R × I → M : (s, t) 7→ γs (t) such that,
for each t ∈ I, the curve R → M : s 7→ γs (t) is the unique solution of the
∂
differential equation ∂s γs (t) = Yt (γs (t)) with γ0 (t) = γ(t). These maps γs
satisfies (4.1.7) by Step 1.
We can now define the derivative of the energy functional E at γ
in the direction of a tangent vector X ∈ Tγ Ωp,q by

d 
dE(γ)X := E(γs ), (4.1.8)
ds  s=0
where s 7→ γs is as in Lemma 4.1.7. Similarly, the derivative of the length
functional L at γ in the direction of X ∈ Tγ Ωp,q is defined by

d 
dL(γ)X := L(γs ). (4.1.9)
ds  s=0
To define (4.1.8) and (4.1.9) the functions s 7→ E(γs ) and s 7→ L(γs ) must
be differentiable at s = 0. This is true for E but it only holds for L
when γ̇(t) 6= 0 for all t ∈ I. Second, we must show that the right hand sides
of (4.1.8) and (4.1.9) depend only on X and not on the choice of {γs }s∈R .
Third, we must verify that dE(γ) : Tγ Ωp,q → R and dL(γ) : Ωp,q → R are lin-
ear maps. This is an exercise in first year analysis (see also the proof of Theo-
rem 4.1.4). A curve γ ∈ Ωp,q is is then an extremal point of E (respectively L
when γ̇(t) 6= 0 for all t) if and only if dE(γ) = 0 (respectively dL(γ) = 0).
Such a curve is also called a critical point of E (respectively L).
180 CHAPTER 4. GEODESICS

4.1.3 Characterization of Geodesics

Proof of Theorem 4.1.4. The equivalence of (iii) and (iv) follows directly
from the equations ∇γ̇(t) = Π(γ(t))γ̈(t) and ker(Π(γ(t))) = Tγ(t) M ⊥ .
We prove that (i) is equivalent to (iii) and (iv). Let X ∈ Tγ Ωp,q and
choose a smooth curve of curves R → Ωp,q : s 7→ γs satisfying (4.1.7). Then
the function (s, t) 7→ |γ̇s (t)|2 is smooth and hence

d 
dE(γ)X = E(γs )
ds s=0
1 b
 Z
d 
= |γ̇s (t)|2 dt
ds s=0 2 a
1 b ∂ 
Z 
= |γ̇s (t)|2 dt
2 a ∂s s=0
Z b   (4.1.10)
∂ 
= γ̇(t), γ̇s (t) dt
a ∂s s=0
Z bD E
= γ̇(t), Ẋ(t) dt
a
Z b
=− hγ̈(t), X(t)i dt.
a

That (iii) implies (i) follows directly from this identity. To prove that (i) im-
plies (iv) we argue indirectly and assume that there exists a point t0 ∈ [0, 1]
such that γ̈(t0 ) is not orthogonal to Tγ(t0 ) M . Then there exists a vec-
tor v0 ∈ Tγ(t0 ) M such that hγ̈(t0 ), v0 i > 0. We may assume without loss
of generality that a < t0 < b. Then there exists a constant ε > 0 such
that a < t0 − ε < t0 + ε < b and

t0 − ε < t < t 0 + ε =⇒ hγ̈(t), Π(γ(t))v0 i > 0.

Now choose a smooth cutoff function β : I → [0, 1] such that β(t) = 0 for
all t ∈ I with |t − t0 | ≥ ε and β(t0 ) = 1. Define X ∈ Tγ Ωp,q by

X(t) := β(t)Π(γ(t))v0 for t ∈ I.

Then hγ̈(t), X(t)i ≥ 0 for all t and hγ̈(t0 ), X(t0 )i > 0. Hence
Z b
dE(γ)X = − hγ̈(t), X(t)i dt < 0
a

and so γ does not satisfy (i). Thus (i) is equivalent to (iii) and (iv).
4.1. LENGTH AND ENERGY 181

We prove that (i) is equivalent to (ii). Assume first that γ satisfies (i).
Then γ also satisfies (iv) and hence γ̈(t) ⊥ Tγ(t) M for all t ∈ I. This implies
1d
0 = hγ̈(t), γ̇(t)i = |γ̇(t)|2 .
2 dt
Hence the function I → R : t 7→ |γ̇(t)|2 is constant. Choose c ≥ 0 such
that |γ̇(t)| ≡ c. If c = 0, then γ(t) is constant and so γ(t) ≡ p = q. If c > 0,
then
 Z b
d 
dL(γ)X = |γ̇s (t)| dt
ds s=0 a
Z b 
∂ 
= |γ̇s (t)| dt
a ∂s s=0

Z b   
−1 ∂ 
= |γ̇(t)| γ̇(t), γ̇ s (t) dt
a ∂s s=0
1 bD
Z E
= γ̇(t), Ẋ(t) dt
c a
1
= dE(γ)X.
c
Thus, in the case c > 0, γ is an extremal point of E if and only if it is an
extremal point of L. Hence (i) is equivalent to (ii).
We prove that (iii) is equivalent to (v). Let (Φ, γ, γ 0 ) be a development
of M along M 0 = Rm . Then γ̇ 0 (t) = Φ(t)γ̇(t) and dt d
Φ(t)X(t) = Φ(t)∇X(t)
for all X ∈ Vect(γ) and all t ∈ I. Take X = γ̇ to obtain γ̈ 0 (t) = Φ(t)∇γ̇(t) for
all t ∈ I. Thus ∇γ̇ ≡ 0 if and only if γ̈ 0 ≡ 0. This proves Theorem 4.1.4.
Remark 4.1.3 shows that reparametrization by a nundecreasing surjective
map σ : I 0 → I gives rise to map
Ωp,q (I) → Ωp,q (I 0 ) : γ 7→ γ ◦ σ
which preserves the length functional, i.e.
L(γ ◦ σ) = L(γ)
for all γ ∈ Ωp,q (I). Thus the chain rule in infinite dimensions should assert
that if γ◦σ is an extremal (i.e. critical) point of L, then γ is an extremal point
of L. Moreover, if σ is a diffeomorphism, the map γ 7→ γ ◦ σ is bijective and
should give rise to a bijective correspondence between the extremal points
of L on Ωp,q (I) and those on Ωp,q (I 0 ). Finally, if the tangent vector field γ̇
vanishes nowhere, then γ can be parametrized by the arclength. This is
spelled out in more detail in the next exercise.
182 CHAPTER 4. GEODESICS

Exercise 4.1.8. Let γ : I = [a, b] → M be a smooth curve such that

γ̇(t) 6= 0

for all t ∈ I and define

Z b
T := L(γ) = |γ̇(t)| dt.
a

(i) Prove that there exists a unique diffeomorphism σ : [0, T ] → I such that
Z t
0 0
σ(t ) = t ⇐⇒ t = |γ̇(s)| ds
a

for all t0 ∈ [0, T ] and all t ∈ [a, b]. Prove that γ 0 := γ ◦ σ : [0, T ] → M is
parametrized by the arclength, i.e. |γ̇ 0 (t0 )| = 1 for all t0 ∈ [0, T ].
(ii) Prove that
Z b
dL(γ)X = − hV̇ (t), X(t)i dt, V (t) := |γ̇(t)|−1 γ̇(t). (4.1.11)
a

Hint: See the relevant formula in the proof of Theorem 4.1.4.

(iii) Prove that γ is an extremal point of L if and only if the curve γ 0 in
part (i) is a geodesic.
(iv) Prove that γ is an extremal point of L if and only if there exists a
geodesic γ 0 : I 0 → M and a diffeomorphism σ : I 0 → I such that γ 0 = γ ◦ σ.

Next we generalize this exercise to cover the case where γ̇ is allowed to

vanish. Recall from Remark 4.1.6 that the function s 7→ L(γs ) need not be
differentiable. As an example consider the case where γ = γ0 is constant
(see also Exercise 4.4.12 below).

Exercise 4.1.9. Let γ : I → M be a smooth curve and let X ∈ Tγ Ωp,q (I).

Choose a smooth curve of curves R → Ωp,q (I) : s 7→ γs that satisfies (4.1.7).
Prove that the one-sided derivatives of the function s 7→ L(γs ) exist at s = 0
and satisfy the inequalities
Z   Z 
 d  
− Ẋ(t) dt ≤ L(γs ) ≤ Ẋ(t) dt.
    
I ds s=0 I

Exercise 4.1.10. Let (Φ, γ, γ 0 ) be a development of M along M 0 . Show

that γ is a geodesic in M if and only if γ 0 is a geodesic in M 0 .
4.2. DISTANCE 183

4.2 Distance
Assume that M ⊂ Rn is a connected smooth m-dimensional submanifold.
Two point p, q ∈ M are of distance |p − q| apart in the ambient Euclidean
space Rn . In this section we define a distance function which is more in-
timately tied to M by minimizing the length functional over the space of
curves in M with fixed endpoints. Thus it may happen that two points
in M have a very short distance in Rn but can only be joined by very long
curves in M (see Figure 4.1). This leads to the intrinsic distance in M .
Throughout we denote by I = [0, 1] the unit interval and, for p, q ∈ M , by

Ωp,q := {γ : [0, 1] → M | γ is smooth and γ(0) = p, γ(1) = q} (4.2.1)

the space of smooth paths on the unit interval joining p to q. Since M is

connected the set Ωp,q is nonempty for all p, q ∈ M . (Prove this!)

p
q
M p q

Figure 4.1: Curves in M .

Definition 4.2.1. The intrinsic distance between two points p, q ∈ M is

the real number d(p, q) ≥ 0 defined by

d(p, q) := inf L(γ). (4.2.2)

γ∈Ωp,q

The inequality d(p, q) ≥ 0 holds because each curve has nonnegative length
and the inequality d(p, q) < ∞ holds because Ωp,q 6= ∅.
Remark 4.2.2. Every smooth curve γ : [0, 1] → Rn with endpoints γ(0) = p
and γ(1) = q satisfies the inequality
Z 1 Z 1 
 
L(γ) = |γ̇(t)| dt ≥ 
 γ̇(t) dt = |p − q| .
0 0

Thus d(p, q) ≥ |p − q|. For γ(t) := p + t(q − p) we have equality and hence
the straight lines minimize the length among all curves from p to q.
184 CHAPTER 4. GEODESICS

Lemma 4.2.3. The function d : M × M → [0, ∞) defines a metric on M :

(i) If p, q ∈ M satisfy d(p, q) = 0, then p = q.
(ii) For all p, q ∈ M we have d(p, q) = d(q, p).
(iii) For all p, q, r ∈ M we have d(p, r) ≤ d(p, q) + d(q, r).

Proof. By Remark 4.2.2 we have d(p, q) ≥ |p − q| for all p, q ∈ M and this

e(t) := γ(1 − t)
proves part (i). Part (ii) follows from the fact that the curve γ
has the same length as γ and belongs to Ωq,p whenever γ ∈ Ωp,q . To prove
part (iii) fix a constant ε > 0 and choose curves γ0 ∈ Ωp,q and γ1 ∈ Ωq,r
such that L(γ0 ) < d(p, q) + ε and L(γ1 ) < d(q, r) + ε. By Remark 4.1.3 we
may assume without loss of generality that γ0 (1 − t) = γ1 (t) = q for t > 0
sufficiently small. Under this assumption the curve

γ0 (2t), for 0 ≤ t ≤ 1/2,
γ(t) :=
γ1 (2t − 1), for 1/2 < t ≤ 1

is smooth. Moreover, γ(0) = p and γ(1) = r and so γ ∈ Ωp,r . Thus

d(p, r) ≤ L(γ) = L(γ0 ) + L(γ1 ) < d(p, q) + d(q, r) + 2ε.

Hence d(p, r) < d(p, q) + d(q, r) + 2ε for every ε > 0. This proves part (iii)
and Lemma 4.2.3.

Remark 4.2.4. It is natural to ask if the infimum in (4.2.2) is always

attained. This is easily seen not to be the case in general. For example,
let M result from the Euclidean space Rm by removing a point p0 . Then
the distance d(p, q) = |p − q| is equal to the length of the line segment from p
to q and any other curve from p to q is longer. Hence if p0 is in the interior
of this line segment, the infimum is not attained. We shall prove below that
the infimum is attained whenever M is complete.

q
p

Figure 4.2: A geodesic on the 2-sphere.

4.2. DISTANCE 185

Example 4.2.5. Let

M := S 2 = p ∈ R3 | |p| = 1


be the unit sphere in R3 and fix two points p, q ∈ S 2 . Then d(p, q) is the
length of the shortest curve on the 2-sphere connecting p and q. Such a
curve is a segment on a great circle through p and q (see Figure 4.2) and its
length is
d(p, q) = cos−1 (hp, qi), (4.2.3)
where hp, qi denotes the standard inner product, and we have
0 ≤ d(p, q) ≤ π.
(See Example 4.3.11 below for details.) By Lemma 4.2.3 this defines a metric
on S 2 . Exercise: Prove directly that (4.2.3) is a distance function on S 2 .
We now have two topologies on our manifold M ⊂ Rn , namely the topol-
ogy determined by the metric d in Lemma 4.2.3 and the relative topology
inherited from Rn . The latter is also determined by a distance function,
namely the extrinsic distance function defined as the restriction of the Eu-
clidean distance function on Rn to the subset M . We denote it by
d0 : M × M → [0, ∞), d0 (p, q) := |p − q| . (4.2.4)
A natural question is if these two metrics d and d0 induce the same topology
on M . In other words is a subset U ⊂ M open with respect to d0 if and only
if it is open with respect to d? Or, equivalently, does a sequence pν ∈ M
converge to p0 ∈ M with respect to d if and only if it converges to p0 with
respect to d0 ? Lemma 4.2.7 answers this question in the affirmative.
Exercise 4.2.6. Prove that every translation of Rn and every orthogonal
transformation preserves the lengths of curves.
Lemma 4.2.7. For every p0 ∈ M we have
d(p, q)
lim = 1.
p,q→p0 |p − q|
Proof. See page 187.
Lemma 4.2.8. Let p0 ∈ M and let φ0 : U0 → Ω0 be a coordinate chart onto
an open subset of Rm such that its derivative dφ0 (p0 ) : Tp0 M → Rm is an
orthogonal transformation. Then
d(p, q)
lim = 1.
p,q→p0 |φ0 (p) − φ0 (q)|
Proof. See page 188.
186 CHAPTER 4. GEODESICS

The lemmas imply that the topology M inherits as a subset of Rm , the

topology on M determined by the metric d, and the topology on M induced
by the local coordinate systems on M are all the same.
Corollary 4.2.9. For every subset U ⊂ M the following are equivalent.
(i) U is open with respect to the metric d in (4.2.2).
(ii) U is open with respect to the metric d0 in (4.2.4).
(iii) For every coordinate chart φ0 : U0 → Ω0 of M onto an open sub-
set Ω0 ⊂ Rm the set φ0 (U0 ∩ U ) is an open subset of Rm .
Proof. By Remark 4.2.2 we have

|p − q| ≤ d(p, q) (4.2.5)

for all p, q ∈ M . Thus the identity idM : (M, d) → (M, d0 ) is Lipschitz

continuous with Lipschitz constant one and so every d0 -open subset of M
is d-open. Conversely, let U ⊂ M be a d-open subset of M and let p0 ∈ U
and ε > 0. Then, by Lemma 4.2.7, there exists a constant δ > 0 such that
all p, q ∈ M with |p − p0 | < δ and |q − p0 | < δ satisfy

d(p, q) ≤ (1 + ε)|p − q|.

Since U is d-open, there exists a constant ρ > 0 such that

Bρ (p0 , d) ⊂ U.

With  
ρ
ρ0 := min δ,
1+ε
this implies Bρ0 (p0 , d0 ) ⊂ U . Namely, if p ∈ M satisfies

|p − p0 | < ρ0 ≤ δ,

then
d(p, p0 ) ≤ (1 + ε)|p − p0 | < (1 + ε)ρ0 ≤ ρ
and so p ∈ U . Thus U is d0 -open and this proves that (i) is equivalent to (ii).
That (ii) implies (iii) follows from the fact that each coordinate chart φ0
is a homeomorphism. To prove that (iii) implies (i), we argue indirectly
and assume that U is not d-open. Then there exists a sequence pν ∈ M \ U
that converges to an element p0 ∈ U . Let φ0 : U0 → Ω0 be a coordinate
chart with p0 ∈ U0 . Then limν→∞ |φ0 (pν ) − φ0 (p0 )| = 0 by Lemma 4.2.8.
Thus φ0 (U0 ∩ U ) is not open and so U does not satisfy (iii).
4.2. DISTANCE 187

Tp 0M ⊥
M

p0
Tp 0M

Figure 4.3: Locally, M is the graph of f .

Proof of Lemma 4.2.7. By Remark 4.2.2 the estimate |p − q| ≤ d(p, q) holds

for all p, q ∈ M . The lemma asserts that, for all p0 ∈ M and all ε > 0, there
exists a d0 -open neighborhood U0 ⊂ M of p0 such that all p, q ∈ U0 satisfy

|p − q| ≤ d(p, q) ≤ (1 + ε)|p − q|. (4.2.6)

Let p0 ∈ M and ε > 0, and define x : Rn → Tp0 M and y : Rn → Tp0 M ⊥ by

x(p) := Π(p0 )(p − p0 ), y(p) := (1l − Π(p0 )) (p − p0 ),

where Π(p0 ) : Rn → Tp0 M denotes the orthogonal projection as usual.

Then the derivative of the map x|M : M → Tp0 M at p = p0 is the iden-
tity on Tp0 M . Hence the Inverse Function Theorem 2.2.17 asserts that
the map x|M : M → Tp0 M is locally invertible near p0 . Extending this
inverse to a smooth map from Tp0 M to Rn and composing it with the
map y : M → Tp0 M ⊥ , we obtain a smooth map

f : Tp0 M → Tp0 M ⊥

and an open neighborhood W ⊂ Rn of p0 such that

p∈M ⇐⇒ y(p) = f (x(p))

for all p ∈ W (see Figure 4.3). Moreover, by definition the map f satisfies

f (0) = 0 ∈ Tp0 M ⊥ , df (0) = 0 : Tp0 M → Tp0 M ⊥ .

Hence there exists a constant δ > 0 such that, for every x ∈ Tp0 M , we have

|df (x)bx|
|x| < δ =⇒ x + f (x) ∈ W and kdf (x)k = sup < ε.
b∈Tp0 M
06=x |bx|
188 CHAPTER 4. GEODESICS

Define
U0 := {p ∈ M ∩ W | |x(p)| < δ} .
Given p, q ∈ U0 let γ : [0, 1] → M be the curve whose projection to the x-axis
is the straight line joining x(p) to x(q), i.e.
x(γ(t)) = x(p) + t(x(q) − x(p)) =: x(t),
y(γ(t)) = f (x(γ(t))) = f (x(t)) =: y(t).
Then γ(t) ∈ U0 for all t ∈ [0, 1] and
Z t
L(γ) = |ẋ(t) + ẏ(t)| dt
0
Z t
= |ẋ(t) + df (x(t))ẋ(t)| dt
0
Z t


≤ 1 + kdf (x(t))k |ẋ(t)| dt
0
Z t
≤ (1 + ε) |ẋ(t)| dt
0
= (1 + ε) |x(p) − x(q)|
= (1 + ε) |Π(p0 )(p − q)|
≤ (1 + ε) |p − q| .
Hence d(p, q) ≤ L(γ) ≤ (1 + ε) |p − q| and this proves Lemma 4.2.7.
Proof of Lemma 4.2.8. By assumption we have
|dφ0 (p0 )v| = |v|
for all v ∈ Tp0 M . Fix a constant ε > 0. Then, by continuity of the derivative,
there exists a d0 -open neighborhood M0 ⊂ M of p0 such that for all p ∈ M0
and all v ∈ Tp M we have
(1 − ε) |dφ0 (p)v| ≤ |v| ≤ (1 + ε) |dφ0 (p)v| .
Thus for every curve γ : [0, 1] → M0 we have
(1 − ε)L(φ0 ◦ γ)) ≤ L(γ) ≤ (1 + ε)L(φ0 ◦ γ).
One is tempted to take the infimum over all curves γ : [0, 1] → M0 joining
two pints p, q ∈ M0 to obtain the inequality
(1 − ε) |φ0 (p) − φ0 (q)| ≤ d(p, q) ≤ (1 + ε) |φ0 (p) − φ0 (q)| . (4.2.7)
However, we must justify these inequalities by showing that the infimum
over all curves in M0 agrees with the infimum over all curves in M joining
the points p and q.
4.2. DISTANCE 189

It suffices to show that the inequalities hold on a smaller heighbor-

hood M1 ⊂ M0 of p0 . Choose such a smaller neighborhood M1 such that the
open set φ0 (M1 ) is a convex subset of Ω0 . Then the right inequality in (4.2.7)
follows by taking the curve γ : [0, 1] → M1 from γ(0) = p to γ(1) = q such
that φ0 ◦ γ : [0, 1] → φ0 (M1 ) is a straight line. To prove the left inequality
in (4.2.7) we use the fact that M0 is d-open by Lemma 4.2.7. Hence, after
shrinking M1 if necessary, there exists a constant r > 0 such that

p0 ∈ M1 ⊂ Br (p0 , d) ⊂ B3r (p0 , d) ⊂ M0 .

Then, for p, q ∈ M1 we have d(p, q) ≤ 2r while L(γ) ≥ 4r for any curve γ

from p to q which leaves M0 . Hence the distance d(p, q) of p, q ∈ M1 is the
infimum of the lengths L(γ) over all curves γ : [0, 1] → M0 that join γ(0) = p
to γ(1) = q. This proves the left inequality in (4.2.7) and Lemma 4.2.8.

A next question one might ask is the following. Can we choose a coor-
dinate chart φ : U → Ω on M with values in an open set Ω ⊂ Rm so that
the length of each smooth curve γ : [0, 1] → U is equal to the length of the
curve c := φ ◦ γ : [0, 1] → Ω? We examine this question by considering the
inverse map ψ := φ−1 : Ω → U . Denote the components of x and ψ(x) by

x = (x1 , . . . , xm ) ∈ Ω, ψ(x) = (ψ 1 (x), . . . , ψ n (x)) ∈ U.

Given a smooth curve [0, 1] → Ω : t 7→ c(t) = (c1 (t), . . . , cm (t)) we can write
the length of the composition γ = ψ ◦ c : [0, 1] → M in the form
1
Z  
d 
L(ψ ◦ c) =  ψ(c(t)) dt
 dt 
0
v
Z 1u n  2
uX d ν
= t ψ (c(t)) dt
0 dt
ν=1
v !2
Z 1u n m ν
uX X ∂ψ
= t (c(t))ċi (t) dt
0 ∂xi
ν=1 i=1
v
u n X m
Z 1 uX
∂ψ ν ∂ψ ν
= t
i
(c(t)) j (c(t))ċi (t)ċj (t) dt
0 ∂x ∂x
ν=1 i,j=1
v
u m
Z 1uX
= t ċi (t)gij (c(t))ċj (t) dt.
0 i,j=1
190 CHAPTER 4. GEODESICS

Here the functions gij : Ω → R are defined by

n
∂ψ ν ∂ψ ν
 
X ∂ψ ∂ψ
gij (x) := (x) (x) = (x), j (x) . (4.2.8)
∂xi ∂xj ∂xi ∂x
ν=1

Thus we have a smooth function g = (gij ) : Ω → Rm×m with values in the

positive definite matrices given by g(x) = dψ(x)T dψ(x) such that
Z 1q
L(ψ ◦ c) = ċ(t)T g(c(t))ċ(t) dt (4.2.9)
0

for every smooth curve c : [0, 1] → Ω. Thus the condition L(ψ ◦ c) = L(c)
for every such curve is equivalent to

gij (x) = δij

for all x ∈ Ω or, equivalently,

dψ(x)T dψ(x) = 1l. (4.2.10)

This means that ψ preserves angles and areas. The next example shows that
for M = S 2 it is impossible to find such coordinates.

α
β Aγ

Figure 4.4: A spherical triangle.

Example 4.2.10. Consider the manifold M = S 2 . If there is a diffeomor-

phism ψ : Ω → U from an open set Ω ⊂ R2 onto an open set U ⊂ S 2 that
satisfies (4.2.10), it has to map straight lines onto arcs of great circles and
it preserves the area. However, the area A of a spherical triangle bounded
by three arcs on great circles satisfies the angle sum formula

α + β + γ = π + A.

(See Figure 4.4.) Hence there can be no such map ψ.

4.3. THE EXPONENTIAL MAP 191

4.3 The Exponential Map

4.3.1 Geodesic Spray
The tangent bundle T M is a smooth 2m-dimensional manifold in Rn × Rn
by Corollary 2.6.12. The next lemma characterizes the tangent bundle of
the tangent bundle. Compare this with Lemma 3.4.5.
Lemma 4.3.1. The tangent space of T M at (p, v) ∈ T M is given by
  
n
 b ∈ Tp M and
n p
p, vb) ∈ R × R 
T(p,v) T M = (b
. (4.3.1)
1l − Π(p) vb = hp (b
p, v)

Proof. We prove the inclusion “⊂” in (4.3.1). Let (b p, vb) ∈ T(p,v) T M and
choose a smooth curve R → T M : t 7→ (γ(t), X(t)) such that

γ(0) = p, X(0) = v, γ̇(0) = pb, Ẋ(0) = vb.

Then the Gauß–Weingarten formula (3.2.2) asserts that

Ẋ(t) = ∇X(t) + hγ(t) (γ̇(t), X(t))

and hence (1l − Π(γ(t)))Ẋ(t) = hγ(t) (γ̇(t), X(t)) for all t ∈ R. Take t = 0
to obtain (1l − Π(p))b
v = hp (b
p, v). This proves the inclusion “⊂” in (4.3.1).
Equality holds because both sides of the equation are 2m-dimensional linear
subspaces of Rn × Rn .

By Lemma 4.3.1 a smooth map S = (S1 , S2 ) : T M → Rn × Rn is a vector

field on T M if and only if

S1 (p, v) ∈ Tp M, (1l − Π(p))S2 (p, v) = hp (S1 (p, v), v)

for all (p, v) ∈ T M . A special case is where S1 (p, v) = v. Such vector fields
correspond to second order differential equations on M .
Definition 4.3.2 (Spray). A vector field S ∈ Vect(T M ) is called a spray
iff it has the form S(p, v) = (v, S2 (p, v)) where S2 : T M → Rn is a smooth
map satisfying

(1l − Π(p))S2 (p, v) = hp (v, v), S2 (p, λv) = λ2 S2 (p, v) (4.3.2)

for all (p, v) ∈ T M and λ ∈ R. The vector field S ∈ Vect(T M ) defined by

S(p, v) := (v, hp (v, v)) ∈ T(p,v) T M (4.3.3)

for p ∈ M and v ∈ Tp M is called the geodesic spray.

192 CHAPTER 4. GEODESICS

4.3.2 The Exponential Map

Lemma 4.3.3. Let γ : I → M be a smooth curve on an open interval I ⊂ R.
Then γ is a geodesic if and only if the curve I → T M : t 7→ (γ(t), γ̇(t)) is
an integral curve of the geodesic spray S in (4.3.3).
Proof. A smooth curve I → T M : t 7→ (γ(t), X(t)) is an integral curve of S
if and only if
γ̇(t) = X(t), Ẋ(t) = hγ(t) (X(t), X(t))
for all t ∈ I. By equation (4.1.5), this holds if and only if γ is a geodesic
and γ̇ = X.

Combining Lemma 4.3.3 with Theorem 2.4.7 we obtain the following

existence and uniqueness result for geodesics.
Lemma 4.3.4. Let M ⊂ Rn be an m-dimensional submanifold.
(i) For every p ∈ M and every v ∈ Tp M there is an ε > 0 and a smooth
curve γ : (−ε, ε) → M such that
∇γ̇ ≡ 0, γ(0) = p, γ̇(0) = v. (4.3.4)

(iI) If γ1 : I1 → M and γ2 : I2 → M are geodesics and t0 ∈ I1 ∩ I2 with

γ1 (t0 ) = γ2 (t0 ), γ̇1 (t0 ) = γ̇2 (t0 ),
then γ1 (t) = γ2 (t) for all t ∈ I1 ∩ I2 .
Proof. Lemma 4.3.3 and Theorem 2.4.7.

Definition 4.3.5 (Exponential Map). For p ∈ M and v ∈ Tp M the in-

terval
[  
 I is an open interval containing 0 and there is a
Ip,v := I ⊂ R .
geodesic γ : I → M satisfying γ(0) = p, γ̇(0) = v
is called the maximal existence interval for the geodesic through p in the
direction v. For p ∈ M define the set Vp ⊂ Tp M by
Vp := {v ∈ Tp M | 1 ∈ Ip,v } . (4.3.5)
The exponential map at p is the map
expp : Vp → M
that assigns to every tangent vector v ∈ Vp the point expp (v) := γ(1),
where γ : Ip,v → M is the unique geodesic satisfying γ(0) = p and γ̇(0) = v.
4.3. THE EXPONENTIAL MAP 193

p
M

Figure 4.5: The exponential map.

Lemma 4.3.6. (i) The set

V := {(p, v) | p ∈ M, v ∈ Vp } ⊂ T M

is open and the map V → M : (p, v) 7→ expp (v) is smooth.

(ii) If p ∈ M and v ∈ Vp , then

Ip,v = {t ∈ R | tv ∈ Vp }

and the geodesic γ : Ip,v → M with γ(0) = p and γ̇(0) = v is given by

γ(t) = expp (tv), t ∈ Ip,v .

Proof. Part (i) follows directly from Lemma 4.3.3 and Theorem 2.4.9. To
prove part (ii), fix an element p ∈ M and a tangent vector v ∈ Vp , and
let γ : Ip,v → M be the unique geodesic with γ(0) = p and γ̇(0) = v. Fix a
nonzero real number λ and define the map γλ : λ−1 Ip,v → M by

γλ (t) := γ(λt) for t ∈ λ−1 Ip,v .

Then γ̇λ (t) = λγ̇(λt) ans γ̈λ (t) = λ2 γ̈(λt) and hence

∇γ̇λ (t) = Π(γλ (t))γ̈λ (t) = λ2 Π(γ(λt))γ̈(λt) = λ2 ∇γ̇(λt) = 0

for every t ∈ λ−1 Ip,v . This shows that γλ is a geodesic with

γλ (0) = p, γ̇λ (0) = λv.

In particular, we have λ−1 Ip,v ⊂ Ip,λv . Interchanging the roles of v and λv

we obtain λ−1 Ip,v = Ip,λv . Thus

λ ∈ Ip,v ⇐⇒ 1 ∈ Ip,λv ⇐⇒ λv ∈ Vp

and
γ(λ) = γλ (1) = expp (λv)
for λ ∈ Ip,v . This proves Lemma 4.3.6.
194 CHAPTER 4. GEODESICS

Since expp (0) = p by definition, the derivative of the exponential map

at v = 0 is a linear map from Tp M to itself. This derivative is the identity
map as illustrated in Figure 4.5 and proved in the following corollary.
Corollary 4.3.7. The map expp : Vp → M is smooth and its derivative at
the origin is d expp (0) = id : Tp M → Tp M .
Proof. The set Vp is an open subset of the linear subspace Tp M ⊂ Rn ,
with respect to the relative topology, and hence is a manifold. The tan-
gent space of Vp at each point is Tp M . By Lemma 4.3.6 the exponential
map expp : Vp → M is smooth and its derivative at the origin is given by

d 
d expp (0)v =  expp (tv) = γ̇(0) = v,
dt t=0
where γ : Ip,v → M is once again the unique geodesic through p in the
direction v. This proves Corollary 4.3.7.
Corollary 4.3.8. Let p ∈ M and, for r > 0, denote
Br (p) := {v ∈ Tp M | |v| < r} .
If r > 0 is sufficiently small, then Br (p) ⊂ Vp , the set
Ur (p) := expp (Br (p))
is an open subset of M , and the restriction of the exponential map to Br (p)
is a diffeomorphism from Br (p) to Ur (p).
Proof. This follows directly from Corollary 4.3.7 and Theorem 2.2.17.
Definition 4.3.9 (Injectivity Radius). Let M ⊂ Rn be a smooth m-
manifold. The injectivity radius of M at p is the supremum of all real
numbers r > 0 such that Br (p) ⊂ Vp and the restriction of the exponential
map expp to Br (p) is a diffeomorphism onto its image
Ur (p) := expp (Br (p)).
It will be denoted by
 
 Br (p) ⊂ Vp and

 
inj(p) := inj(p; M ) := sup r > 0  expp : Br (p) → Ur (p) .

is a diffeomorphism
  

The injectivity radius of M is the infimum of the injectivity radii of M

at p over all p ∈ M . It will be denoted by
inj(M ) := inf inj(p; M ).
p∈M
4.3. THE EXPONENTIAL MAP 195

4.3.3 Examples and Exercises

Example 4.3.10. The exponential map on Rm is given by

expp (v) = p + v for p, v ∈ Rm .

For every p ∈ Rm this map is a diffeomorphism from Tp Rm = Rm to Rm and

hence the injectivity radius of Rm is infinity.

Example 4.3.11. The exponential map on S m is given by

sin(|v|)
expp (v) = cos(|v|)p + v
|v|

for every p ∈ S m and every nonzero tangent vector v ∈ Tp S m = p⊥ . The re-

striction of this map to the open ball of radius r in Tp M is a diffeomorphism
onto its image if and only if r ≤ π. Hence the injectivity radius of S m
at every point is π. Exercise: Given p ∈ S m and 0 6= v ∈ Tp S m = p⊥ ,
prove that the geodesic γ : R → S m with γ(0) = p and γ̇(0) = v is given
by γ(t) = cos(t |v|)p + sin(t|v|)
|v| v for t ∈ R. Show that, in the case 0 ≤ |v| ≤ π
there is no shorter curve in S m connecting p and q and deduce that the
intrinsic distance on S m is given by d(p, q) = cos−1 (hp, qi) for p, q ∈ S m (see
Example 4.2.5 for m = 2).

Example 4.3.12. Consider the orthogonal group O(n) ⊂ Rn×n with the
standard inner product hv, wi := trace(v T w) on Rn×n . The orthogonal
projection Π(g) : Rn×n → Tg O(n) is given by

1 
Π(g)v := v − gv T g
2
and the second fundamental form by

hg (v, v) = −gv T v.

Hence a curve γ : R → O(n) is a geodesic if and only if γ T γ̈ + γ̇ T γ̇ = 0 or,

equivalently, γ T γ̇ is constant. This shows that geodesics in O(n) have the
form γ(t) = g exp(tξ) for g ∈ O(n) and ξ ∈ o(n). It follows that

expg (v) = g exp(g −1 v) = exp(vg −1 )g

for g ∈ O(n) and v ∈ Tg O(n). In particular, for g = 1l the exponential

map exp1l : o(n) → O(n) agrees with the exponential matrix.
196 CHAPTER 4. GEODESICS

Exercise 4.3.13. What is the injectivity radius of the 2-torus T2 = S 1 × S 1 ,

the punctured 2-plane R2 \ {(0, 0)}, and the orthogonal group O(n)?
Exercise 4.3.14. If p0 ∈ M and 0 < r < inj(p0 ; M ), then there exists an
open neighborhood U0 ⊂ M of p0 such that r < inj(p; M ) for all p ∈ U0 .
Hint: First show that U0 can be chosen such that Br (p) ⊂ Vp for all p ∈ U0 .
Then show that expp : Br (p) → M is an injective immersion for p near p0 .

4.3.4 Geodesics in Local Coordinates

Lemma 4.3.15. Let M ⊂ Rn be an m-dimensional manifold and choose a
coordinate chart φ : U → Ω with inverse ψ := φ−1 : Ω → U. Let Γkij : Ω → R
be the Christoffel symbols defined by (3.6.6) and let c : I → Ω be a smooth
curve. Then the curve γ := ψ ◦ c : I → M is a geodesic if and only if c
satisfies the 2nd order differential equation
m
X
k
c̈ + Γkij (c)ċi ċj = 0 (4.3.6)
i,j=1

for k = 1, . . . , m.
Proof. This follows immediately from the definition of geodesics and equa-
tion (3.6.7) in Lemma 3.6.1 with X = γ̇ and ξ = ċ.
We remark that Lemma 4.3.15 gives rise to another proof of Lemma 4.3.4
that is based on the existence and uniqueness of solutions of second order
differential equations in local coordinates.
Exercise 4.3.16. Let Ω ⊂ Rm be an open set and g = (gij ) : Ω → Rm×m
be a smooth map with values in the space of positive definite symmetric
matrices. Consider the energy functional
Z 1
E(c) := L(c(t), ċ(t)) dt
0
on the space of paths c : [0, 1] → Ω, where L : Ω × Rm → R is defined by
m
1 X i
L(x, ξ) := ξ gij (x)ξ j . (4.3.7)
2
i,j=1

The Euler–Lagrange equations of this variational problem have the form

d ∂L ∂L
k
(c(t), ċ(t)) = (c(t), ċ(t)), k = 1, . . . , m. (4.3.8)
dt ∂ξ ∂xk
Prove that the Euler–Lagrange equations (4.3.8) are equivalent to the geo-
desic equations (4.3.6), where the Γkij : Ω → R are given by (3.6.10).
4.4. MINIMAL GEODESICS 197

4.4 Minimal Geodesics

4.4.1 Characterization of Minimal Geodesics
Any straight line segment in Euclidean space is the shortest curve joining its
endpoints. The analogous assertion for geodesics in a manifold M is false;
consider for example an arc which is more than half of a great circle on a
sphere. In this section we consider curves which realize the shortest distance
between their endpoints.

Lemma 4.4.1. Let I = [a, b] be a compact interval, let γ : I → M be a

smooth curve, and define p := γ(a) and q := γ(b). Then the following are
equivalent.
(i) γ is parametrized proportional to the arclength, i.e. |γ̇(t)| = c is constant,
and γ minimizes the length, i.e. L(γ) ≤ L(γ 0 ) for every smooth curve γ 0 in M
joining p and q.
(ii) γ minimizes the energy, i.e. E(γ) ≤ E(γ 0 ) for every smooth curve γ 0
in M joining p and q.

Proof. See page 198.

Definition 4.4.2 (Minimal Geodesic). A smooth curve γ : I → M on

a compact interval I ⊂ R is called a minimal geodesic iff it satisfies the
equivalent conditions of Lemma 4.4.1.

Remark 4.4.3. (i) Condition (i) says that (the velocity |γ̇| is constant
and) L(γ) = d(p, q), i.e. that γ is a shortest curve from p to q. It is not
precluded that there be more than one such γ; consider for example the
case where M is a sphere and p and q are antipodal.
(ii) Condition (ii) implies that

d 
E(γs ) = 0
ds s=0

for every smooth variation R × I → M : s 7→ γs (t) of γ with fixed endpoints.

Hence a minimal geodesic is a geodesic.
(iii) Finally, we remark that L(γ) (but not E(γ)) is independent of the
parametrization of γ. Hence, if γ is a minimal geodesic, then L(γ) ≤ L(γ 0 )
for every γ 0 (from p to q) whereas E(γ) ≤ E(γ 0 ) for those γ 0 defined on (an
interval the same length as) I.
198 CHAPTER 4. GEODESICS

Proof of Lemma 4.4.1. We prove that (i) implies (ii). Let (c) be the (con-
stant) value of |γ̇(t)|. Then
(b − a)c2
L(γ) = (b − a)c, E(γ) = .
2
Then, for every smooth curve γ 0 : I → M with γ 0 (a) = p and γ 0 (b) = q, we
have
4E(γ)2 = c2 L(γ)2
≤ c2 L(γ 0 )2
Z b 2
2
 0 
=c  γ̇ (t) dt

a
Z b
2 γ̇ 0 (t)2 dt

≤ c (b − a)
a
= 2(b − a)c2 E(γ 0 )
= 4E(γ)E(γ 0 ).
Here the fourth step follows from the Cauchy–Schwarz inequality. Now
divide by 4E(γ) to obtain E(γ) ≤ E(γ 0 ).
We prove that (ii) implies (i). We have already shown in Remark 4.4.3
that (ii) implies that γ is a geodesic. It is easy to dispose of the case where
M is one-dimensional. In that case any γ minimizing E(γ) or L(γ) must be
monotonic onto a subarc; otherwise it could be altered so as to make the
integral smaller. Hence suppose M is of dimension at least two. Suppose,
by contradiction, that L(γ 0 ) < L(γ) for some curve γ 0 from p to q. Since
the dimension of M is bigger than one, we may approximate γ 0 by a curve
whose tangent vector nowhere vanishes, i.e. we may assume without loss of
generality that γ̇ 0 (t) 6= 0 for all t. Then we can reparametrize γ 0 proportional
to arclength without changing its length, and by a further transformation
we can make its domain equal to I. Thus we may assume without loss of
generality that γ 0 : I → M is a smooth curve with γ 0 (a) = p and γ 0 (b) = q
such that |γ 0 (t)| = c0 and

(b − a)c0 = L(γ 0 ) < L(γ) = (b − a)c.

This implies c0 < c and hence

(b − a)c0 2 (b − a)c2
E(γ 0 ) = < = E(γ).
2 2
This contradicts (ii) and proves Lemma 4.4.1.
4.4. MINIMAL GEODESICS 199

4.4.2 Local Existence of Minimal Geodesics

The next theorem asserts the existence of minimal geodesics. It shows
that the set Ur (p) = expp (Br (p)) in Definition 4.3.9 is actually the open
ball Ur (p) = {q ∈ M | d(p, q) < r} whenever r ≤ inj(p; M ).
Theorem 4.4.4 (Existence of Minimal Geodesics). Let M ⊂ Rn be a
smooth m-manifold, fix a point p ∈ M , and let r > 0 be smaller than the
injectivity radius of M at p. Let v ∈ Tp M such that |v| < r. Then

d(p, q) = |v| , q := expp (v),

and a curve γ ∈ Ωp,q has minimal length L(γ) = |v| if and only if there is a
smooth map β : [0, 1] → [0, 1] satisfying

β(0) = 0, β(1) = 1, β̇ ≥ 0

such that γ(t) = expp (β(t)v) for 0 ≤ t ≤ 1.

Proof. See page 201.

Ur

Figure 4.6: The Gauß Lemma.

Lemma 4.4.5 (Gauß Lemma). Let M , p, r be as in Theorem 4.4.4,

let I ⊂ R be an open interval, and let w : I → Vp be a smooth curve whose
norm
|w(t)| =: r
is constant. Define
α(s, t) := expp (sw(t))
for (s, t) ∈ R × I with sw(t) ∈ Vp . Then
 
∂α ∂α
, ≡ 0.
∂s ∂t
Thus the geodesics through p are orthogonal to the boundaries of the embed-
ded balls Ur (p) in Corollary 4.3.8 (see Figure 4.6).
200 CHAPTER 4. GEODESICS

Proof of Lemma 4.4.5. For every t ∈ I we have

α(0, t) = expp (0) = p

and so the assertion holds for s = 0, i.e.

 
∂α ∂α
(0, t), (0, t) = 0.
∂s ∂t
Moreover, each curve s 7→ α(s, t) is a geodesic, i.e.

∂α ∂2α
∇s = Π(α) 2 ≡ 0.
∂s ∂s
By Theorem 4.1.4, the function
 
 ∂α 
s 7→  (s, t)

∂s
is constant for every t, so that
   
 ∂α
 (s, t) =  ∂α (0, t) = |w(t)| = r
  
 ∂s   ∂s  for (s, t) ∈ R × I.

This implies
     
∂ ∂α ∂α ∂α ∂α ∂α ∂α
, = ∇s , + , ∇s
∂s ∂s ∂t ∂s ∂t ∂s ∂t
2
 
∂α ∂ α
= , Π(α)
∂s ∂s∂t
∂α ∂ 2 α
 
= Π(α) ,
∂s ∂s∂t
∂α ∂ 2 α
 
= ,
∂s ∂s∂t
1 ∂  ∂α 2
 
=
2 ∂t  ∂s 
= 0.

Since the function h ∂α ∂α

∂s , ∂t i vanishes for s = 0 we obtain
 
∂α ∂α
(s, t), (s, t) = 0
∂s ∂t
for all s and t. This proves Lemma 4.4.5.
4.4. MINIMAL GEODESICS 201

Proof of Theorem 4.4.4. Let r > 0 be as in Corollary 4.3.8 and let v ∈ Tp M

such that 0 < |v| =: ε < r. Denote q := expp (v) and let γ ∈ Ωp,q . Assume
first that


γ(t) ∈ expp B ε (p) = U ε ∀ t ∈ [0, 1].
Then there is a unique smooth function [0, 1] → Tp M : t 7→ v(t) such that
|v(t)| ≤ ε and γ(t) = expp (v(t)) for every t. The set

I := {t ∈ [0, 1] | γ(t) 6= p} = {t ∈ [0, 1] | v(t) 6= 0} ⊂ (0, 1]

is open in the relative topology of (0, 1]. Thus I is a union of open intervals
in (0, 1) and one half open interval containing 1. Define β : [0, 1] → [0, 1]
and w : I → Tp M by

|v(t)| v(t)
β(t) := , w(t) := ε .
ε |v(t)|

Then β is continuous, both β and w are smooth on I,

β(0) = 0, β(1) = 1, w(1) = v,

and
|w(t)| = ε, γ(t) = expp (β(t)w(t))
for all t ∈ I. We prove that L(γ) ≥ ε. To see this let α : [0, 1] × I → M be
the map of Lemma 4.4.5, i.e.

α(s, t) := expp (sw(t)).

Then γ(t) = α(β(t), t) and hence

∂α ∂α
γ̇(t) = β̇(t) (β(t), t) + (β(t), t)
∂s ∂t
for every t > 0. Hence it follows from Lemma 4.4.5 that
 2  2
2  ∂α
2
 ∂α
|γ̇(t)| = β̇(t)  (β(t), t) +  (β(t), t) ≥ β̇(t)2 ε2
  
∂s ∂t

for every t ∈ I. Hence

Z 1 Z Z   Z
L(γ) = |γ̇(t)| dt = |γ̇(t)| dt ≥ ε β̇(t) dt ≥ ε β̇(t) dt = ε.
 
0 I I I
202 CHAPTER 4. GEODESICS

Here the last equation follows by applying the fundamental theorem of cal-
culus to each interval in I and using the fact that β(0) = 0 and β(1) = 1.
If L(γ) = ε, we must have
∂α
(β(t), t) = 0, β̇(t) ≥ 0 for all t ∈ I.
∂t
Thus I is a single half open interval containing 1 and on this interval the
condition ∂α∂t (β(t), t) = 0 implies ẇ(t) = 0. Since w(1) = v we have w(t) = v
for every t ∈ I. Hence γ(t) = expp (β(t)v) for every t ∈ [0, 1]. It follows
that β is smooth on the closed interval [0, 1] (and not just on I). Thus we
have proved that every γ ∈ Ωp,q with values in U ε has length L(γ) ≥ ε with
equality if and only if γ is a reparametrized geodesic. But if γ ∈ Ωp,q does not
take values only in U ε , there must be a T ∈ (0, 1) such that γ([0, T ]) ⊂ U ε
and γ(T ) ∈ ∂Uε . Then L(γ|[0,T ] ) ≥ ε, by what we have just proved,
and L(γ|[T,1] ) > 0 because the restriction of γ to [T, 1] cannot be constant;
so in this case we have L(γ) > ε. This proves Theorem 4.4.4.

The next corollary gives a partial answer to our problem of finding length
minimizing curves. It asserts that geodesics minimize the length locally.
Corollary 4.4.6. Let M ⊂ Rn be a smooth m-manifold, let I ⊂ R be an
open interval, and let γ : I → M be a geodesic. Fix a point t0 ∈ I. Then
there exists a constant ε > 0 such that
t0 − ε < s < t < t0 + ε =⇒ L(γ|[s,t] ) = d(γ(s), γ(t)).
Proof. Since γ is a geodesic its derivative has constant norm |γ̇(t)| ≡ c (see
Theorem 4.1.4). Choose δ > 0 so small that the interval [t0 − δ, t0 + δ] is
contained in I. Then there is a constant r > 0 such that r ≤ inj(γ(t))
whenever |t − t0 | ≤ δ. Choose ε > 0 such that
ε < δ, 2εc < r.
If t0 − ε < s < t < t0 + ε, then
γ(t) = expγ(s) ((t − s)γ̇(s))
and
|(t − s)γ̇(s)| = |t − s| c < 2εc < r ≤ inj(γ(s)).
Hence it follows from Theorem 4.4.4 that
L(γ|[s,t] ) = |t − s| c = d(γ(s), γ(t)).
This proves Corollary 4.4.6.
4.4. MINIMAL GEODESICS 203

4.4.3 Examples and Exercises

Exercise 4.4.7. How large can the constant ε in Corollary 4.4.6 be chosen
in the case M = S 2 ? Compare this with the injectivity radius.
Remark 4.4.8. We conclude from Theorem 4.4.4 that
  

Sr (p) := q ∈ M  d(p, q) = r = expp v ∈ Tp M | |v| = r (4.4.1)
for 0 < r < inj(p; M ). The Gauß Lemma 4.4.5 shows that the geodesic
rays [0, 1] → M : s 7→ expp (sv) emanating from p are the orthogonal tra-
jectories to the concentric spheres Sr (p).
Exercise 4.4.9. Let
M ⊂ R3
be of dimension two and suppose that M is invariant under the (orthogonal)
reflection about some plane E ⊂ R3 . Show that E intersects M in a geodesic.
(Hint: Otherwise there would be points p, q ∈ M very close to one an-
other joined by two distinct minimal geodesics.) Conclude for example that
the coordinate planes intersect the ellipsoid (x/a)2 + (y/b)2 + (z/c)2 = 1 in
geodesics.
Exercise 4.4.10. Choose geodesic normal coordinates near p ∈ M via
m
!
X
q = expp xi (q)ei ,
i=1

where e1 , . . . , em is an orthonormal basis of Tp M (see Corollary 4.5.4 below).

Then we have xi (p) = 0 and
m
( )
X  i 2
Br (p) = {q ∈ M | d(p, q) < r} = q ∈ M 
 x (q) < r2 (4.4.2)
i=1

for 0 < r < inj(p; M ). Hence Theorem 4.5.3 below asserts that Br (p) is
convex for r > 0 sufficiently small.
(i) Show that it can happen that a geodesic in Br (p) is not minimal. Hint:
Take M to be the hemisphere {(x, y, z) ∈ R3 | x2 + y 2 + z 2 = 1, z > 0} to-
gether with the disc {(x, y, z) ∈ R3 | x2 + y 2 ≤ 1, z = 0}, but smooth the cor-
ners along the circle x2 + y 2 = 1, z = 0. Take p = (0, 0, 1) and r = π/2.
(ii) Show that, if r > 0 is sufficiently small, then the unique geodesic γ
in Br (p) joining two points q, q 0 ∈ Br (p) is minimal and that in fact any
curve γ 0 from q to q 0 which is not a reparametrization of γ is strictly longer,
i.e. L(γ 0 ) > L(γ) = d(q, q 0 ).
204 CHAPTER 4. GEODESICS

Exercise 4.4.11. Let γ : I = [a, b] → M be a smooth curve with end-

points γ(a) = p and γ(b) = q and nowhere vanishing derivative, i.e. γ(t) 6= 0
for all t ∈ I. Prove that the following are equivalent.
(i) The curve γ is an extremal of the length functional, i.e. every
smooth map R × I → M : (s, t) 7→ γs (t) with γs (a) = p and γs (b) = q for
all s satisfies 
d 
L(γs ) = 0.
ds s=0

(ii) The curve γ is a reparametrized geodesic, i.e. there exists a smooth

map σ : [a, b] → [0, 1] with σ(a) = 0, σ(b) = 1, σ̇(t) ≥ 0 for all t ∈ I, and a
vector v ∈ Tp M such that

q = expp (v), γ(t) = expp (σ(t)v)

for all t ∈ I. (We remark that the hypothesis γ̇(t) 6= 0 implies that σ is
actually a diffeomorphism, i.e. σ̇(t) > 0 for all t ∈ I.)
(iii) The curve γ minimizes the length functional locally, i.e. there ex-
ists an ε > 0 such that L(γ|[s,t] ) = d(γ(s), γ(t)) for every closed subinter-
val [s, t] ⊂ I of length t − s < ε.
It is often convenient to consider curves γ where γ̇(t) is allowed to vanish
for some values of t; then γ cannot (in general) be parametrized by arclength.
Such a curve γ : I → M can be smooth (as a map) and yet its image may
have corners (where γ̇ necessarily vanishes). Note that a curve with corners
can never minimize the distance, even locally.
Exercise 4.4.12. Show that conditions (ii) and (iii) in Exercise 4.4.11 are
equivalent, even without the assumption that γ̇ is nowhere vanishing. De-
duce that, if γ : I → M is a shortest curve joining p to q, i.e. L(γ) = d(p, q),
then γ is a reparametrized geodesic.
Show by example that one can have a variation {γs }s∈R of a reparame-
trized geodesic γ0 = γ for which the map s 7→ L(γs ) is not even differentiable
at s = 0. (Hint: Take γ to be constant. See also Exercise 4.1.9.)
Show, however, that conditions (i), (ii) and (iii) in Exercise 4.4.11 remain
equivalent if the hypothesis that γ̇ is nowhere vanishing is weakened to the
hypothesis that γ̇(t) 6= 0 for all but finitely many t ∈ I. Conclude that a bro-
ken geodesic is a reparametrized geodesic if and only if it minimizes arclength
locally. (A broken geodesic is a continuous map γ : I = [a, b] → M for
which there exist a = t0 < t1 < · · · < tn = b such that γ|[ti−1 ,ti ] is a geodesic
for i = 1, . . . , n. It is thus a geodesic if and only if γ̇ is continuous at the
break points, i.e. γ̇(t− +
i ) = γ̇(ti ) for i = 1, . . . , n − 1.)
4.5. CONVEX NEIGHBORHOODS 205

4.5 Convex Neighborhoods

A subset of an affine space is called convex iff it contains the line segment
joining any two of its points. The definition carries over to a submanifold M
of Euclidean space (or indeed more generally to any manifold M equipped
with a spray) once we reword the definition so as to confront the difficulty
that a geodesic joining two points might not exist nor, if it does, need it be
unique.

Definition 4.5.1 (Geodesically Convex Sets). Let M ⊂ Rn be a smooth

m-dimensional manifold. A subset U ⊂ M is called geodesically con-
vex iff, for all p0 , p1 ∈ U , there exists a unique geodesic γ : [0, 1] → U such
that γ(0) = p0 and γ(1) = p1 .

It is not precluded in Definition 4.5.1 that there be other geodesics from p

to q which leave and then re-enter U , and these may even be shorter than
the geodesic in U .

Exercise 4.5.2. (a) Find a geodesically convex set U in a manifold M and

points p0 , p1 ∈ U such that the unique geodesic γ : [0, 1] → U with γ(0) = p0
and γ(1) = p1 has length L(γ) > d(p0 , p1 ). Hint: An interval of length
bigger than π in S 1 .
(b) Find a set U in a manifold M such that any two points in U can be
joined by a minimal geodesic in U , but U is not geodesically convex. Hint:
A closed hemisphere in S 2 .

Theorem 4.5.3 (Convex Neighborhoods). Let M ⊂ Rn be a smooth

m-dimensional submanifold and fix a point p0 ∈ M . Let φ : U → Ω be any
coordinate chart on an open neighborhood U ⊂ M of p0 with values in an
open set Ω ⊂ Rm . Then the set

Ur := {p ∈ U | |φ(p) − φ(p0 )| < r} (4.5.1)

is geodesically convex for r > 0 sufficiently small.

Proof. See page 206.

Corollary 4.5.4. Let M ⊂ Rn be a smooth m-manifold and let p0 ∈ M .

Then, for r > 0 sufficiently small, the open ball

Ur (p0 ) := {p ∈ M | d(p0 , p) < r} (4.5.2)

is geodesically convex.
206 CHAPTER 4. GEODESICS

Proof. Choose an orthonormal basis e1 , . . . , em of Tp0 M and define

Ω := {x ∈ Rm | |x| < inj(p0 ; M )} ,

(4.5.3)
U := {p ∈ M | d(p0 , p) < inj(p0 ; M )} .
Define the map ψ : Ω → U by
m
!
X
ψ(x) := expp0 xi e i (4.5.4)
i=1

for x = (x1 , . . . , xm ) ∈ Ω. Then ψ is a diffeomorphism and d(p0 , ψ(x)) = |x|

for all x ∈ Ω by Theorem 4.4.4. Hence its inverse

φ := ψ −1 : U → Ω (4.5.5)

satisfies φ(p0 ) = 0 and |φ(p)| = d(p0 , p) for all p ∈ U . Thus

Ur (p0 ) = {p ∈ U | |φ(p) − φ(p0 )| < r} for 0 < r < inj(p0 ; M )

and so Corollary 4.5.4 follows from Theorem 4.5.3.

Definition 4.5.5 (Geodesically Normal Coordinates). The coordi-

nate chart φ : U → Ω in (4.5.4) and (4.5.5) sends geodesics through p0 to
straight lines through the origin. Its components x1 , . . . , xm : U → R are
called geodesically normal coordinates at p0 .
Proof of Theorem 4.5.3. Assume without loss of generality that φ(p0 ) = 0.
Let Γkij : Ω → R be the Christoffel symbols of the coordinate chart and,
for x ∈ Ω, define the quadratic function Qx : Rm → R by
m   m
X 2 X
Qx (ξ) := ξk − xk Γkij (x)ξ i ξ j .
k=1 i,j,k=1

Shrinking U , if necessary, we may assume that

 
m
X  1
max  xk Γkij (x) ≤ for all x ∈ Ω.
 
i,j=1,...,m   2m
k=1

Then, for all x ∈ Ω and all ξ ∈ Rm we have

m
!2
2 1 X  i 1 2
Qx (ξ) ≥ |ξ| − ξ  ≥ |ξ| ≥ 0.
2m 2
i=1

Hence Qx is positive definite for every x ∈ Ω.

4.5. CONVEX NEIGHBORHOODS 207

Now let γ : [0, 1] → U be a geodesic and define

c(t) := φ(γ(t))

for 0 ≤ t ≤ 1. Then, by Lemma 4.3.15, c satisfies the differential equation

X
c̈k + Γkij (c)ċi ċj = 0.
i,j

Hence
d2 |c|2 d 2 |ċ|2
= hċ, ci = | ċ| + hc̈, ci = Qc ( ċ) ≥ ≥0
dt2 2 dt 2
and so the function t 7→ |φ(γ(t))|2 is convex. Thus, if γ(0), γ(1) ∈ Ur for
some r > 0, it follows that γ(t) ∈ Ur for all t ∈ [0, 1].
Consider the exponential map

V = {(p, v) ∈ T M | v ∈ Vp } → M : (p, v) 7→ expp (v)

in Lemma 4.3.6. Its domain V is open and the exponential map is smooth.
Since it sends the pair (p0 , 0) ∈ V to expp0 (0) = p0 ∈ U , it follows from con-
tinuity that there exist constants ε > 0 and r > 0 such that

p ∈ Ur , v ∈ Tp M, |v| < ε =⇒ v ∈ Vp , expp (v) ∈ U. (4.5.6)

Moreover, we have

d expp0 (0) = id : Tp0 M → Tp0 M

by Corollary 4.3.7. Hence the Implicit Function Theorem 2.6.15 asserts that
the constants ε > 0 and r > 0 can be chosen such that (4.5.6) holds and there
exists a smooth map h : Ur × Ur → Rn that satisfies the conditions

h(p, q) ∈ Tp M, |h(p, q)| < ε (4.5.7)

for all p, q ∈ Ur and

expp (v) = q ⇐⇒ v = h(p, q) (4.5.8)

for all p, q ∈ Ur and all v ∈ Tp M with |v| < ε. In particular, we have

h(p0 , p0 ) = 0

and expp (h(p, q)) = q for all p, q ∈ Ur .

208 CHAPTER 4. GEODESICS

Fix two constants ε > 0 and r > 0 and a smooth map h : Ur × Ur → Rn

such that (4.5.6), (4.5.7), (4.5.8) are satisfied. We show that any two
points p, q ∈ Ur are joined by a geodesic in Ur . Let p, q ∈ Ur and define
γ(t) := expp (th(p, q)) for 0 ≤ t ≤ 1.
This curve γ : [0, 1] → M is well defined by (4.5.6) and (4.5.7), it is a
geodesic satisfying γ(0) = p ∈ Ur by Lemma 4.3.6, it satisfies γ(1) = q ∈ Ur
by (4.5.8), it takes values in U by (4.5.6) and (4.5.7), and so γ([0, 1]) ⊂ Ur
because the function [0, 1] → R : t 7→ |φ(γ(t))|2 is convex.
We show that there exists at most one geodesic in Ur joining p and q.
Let p, q ∈ Ur and let γ : [0, 1] → Ur be any geodesic such that γ(0) = p
and γ(1) = q. Define v := γ̇(0) ∈ Tp M . Then γ(t) = expp (tv) for 0 ≤ t ≤ 1
by Lemma 4.3.6. We claim that |v| < ε. Suppose, by contradiction, that
|v| ≥ ε.
Then
ε
T := ≤1
|v|
and, for 0 < t < T , we have |tv| < ε and expp (tv) = γ(t) ∈ Ur and so
h(p, γ(t)) = tv.
by (4.5.8). Thus
|h(p, γ(t))| = t|v| for 0 < t < T.
Take the limit t % T to obtain
|h(p, γ(T ))| = T |v| = ε
in contradiction to (4.5.7). This contradiction shows that |v| < ε. Since
expp (v) = γ(1) = q ∈ Ur
it follows from (4.5.8) that v = h(p, q). This proves Theorem 4.5.3.
Remark 4.5.6. Theorem 4.5.3 and its proof carry over to general sprays
(see Definition 4.3.2).
Exercise 4.5.7. Consider the set Ur (p) = {q ∈ M | d(p, q) < r} for p ∈ M
and r > 0. Corollary 4.5.4 asserts that this set is geodesically convex for r
sufficiently small. How large can you choose r in the cases
M = S2, M = T2 = S 1 × S 1 , M = R2 , M = R2 \ {0}.
Compare this with the injectivity radius. If the set Ur (p) in these exam-
ples is geodesically convex, does it follow that every geodesic in Ur (p) is
minimizing?
4.6. COMPLETENESS AND HOPF–RINOW 209

4.6 Completeness and Hopf–Rinow

For a Riemannian manifold there are different notions of completeness.
First, in §3.4 completeness was defined in terms of the completeness of time
dependent basic vector fields on the frame bundle (Definition 3.4.10). Sec-
ond, there is a distance function

d : M × M → [0, ∞)

defined by equation (4.2.2) so that we can speak of completeness of the

metric space (M, d) in the sense that every Cauchy sequence converges.
Third, there is the question of whether geodesics through any point in any
direction exist for all time; if so we call a Riemannian manifold geodesically
complete. The remarkable fact is that these three rather different notions of
completeness are actually equivalent and that, in the complete case, any two
points in M can be joined by a shortest geodesic. This is the content of the
Hopf–Rinow theorem. We will spell out the details of the proof for embedded
manifolds and leave it to the reader (as a straight forward exercise) to extend
the proof to the intrinsic setting.

4.6.1 Geodesic Completeness

Definition 4.6.1. Let M ⊂ Rn be an m-dimensional manifold. Given a
point p ∈ M we say that M is geodesically complete at p if, for ev-
ery tangent vector v ∈ Tp M , there exists a geodesic γ : R → M (on the en-
tire real axis) satisfying γ(0) = p and γ̇(0) = v (or equivalently Vp = Tp M
where Vp ⊂ Tp M is defined by (4.3.5)). The manifold M is called geodesi-
cally complete iff it is geodesically complete at every point p ∈ M .

Definition 4.6.2. Let (M, d) be a metric space. A subset A ⊂ M is called

bounded iff
sup d(p, p0 ) < ∞
p∈A

for some (and hence every) point p0 ∈ M .

Example 4.6.3. A manifold M ⊂ Rn can be contained in a bounded subset

of Rn and still not be bounded with respect to the metric (4.2.2). An
example is the 1-manifold M = (x, y) ∈ R2 | 0 < x < 1, y = sin(1/x) .

Exercise 4.6.4. Let (M, d) be a metric space. Prove that every compact
subset K ⊂ M is closed and bounded. Find an example of a metric space
that contains a closed and bounded subset that is not compact.
210 CHAPTER 4. GEODESICS

Theorem 4.6.5 (Completeness). Let M ⊂ Rn be a connected m-dimen-

sional manifold and let d : M × M → [0, ∞) be the distance function defined
by (4.1.1), (4.2.1), and (4.2.2). Then the following are equivalent.
(i) M is geodesically complete.
(ii) There exists a point p ∈ M such that M is geodesically complete at p.
(iii) Every closed and bounded subset of M is compact.
(iv) (M, d) is a complete metric space.
(v) M is complete, i.e. for every smooth curve ξ : R → Rm and every ele-
ment (p0 , e0 ) ∈ F(M ) there exists a smooth curve β : R → F(M ) satisfying
β̇(t) = Bξ(t) (β(t)), β(0) = (p0 , e0 ). (4.6.1)

(vi) The basic vector field Bξ ∈ Vect(F(M )) is complete for every ξ ∈ Rm .

(vii) For every smooth curve γ 0 : R → Rm , every p0 ∈ M , and every or-
thogonal isomorphism Φ0 : Tp0 M → Rm there exists a development (Φ, γ, γ 0 )
of M along Rm on all of R that satisfies γ(0) = p0 and Φ(0) = Φ0 .
Proof. The proof relies on Theorem 4.6.6 below.

4.6.2 Global Existence of Minimal Geodesics

Theorem 4.6.6 (Hopf–Rinow). Let M ⊂ Rn be a connected m-manifold
and let p ∈ M . Assume M is geodesically complete at p. Then, for ev-
ery q ∈ M , there exists a geodesic γ : [0, 1] → M such that
γ(0) = p, γ(1) = q, L(γ) = d(p, q).
Proof. See page 215.

Theorem 4.6.6 implies Theorem 4.6.5. That (i) implies (ii) follows directly
from the definitions.
We prove that (ii) implies (iii). Thus assume that M is geodesically
complete at the point p0 ∈ M and let K ⊂ M be a closed and bounded
subset. Then r := supq∈K d(p0 , q) < ∞. Hence Theorem 4.6.6 asserts that,
for every q ∈ K, there exists a vector v ∈ Tp0 M such that |v| = d(p0 , q) ≤ r
and expp0 (v) = q. Thus
K ⊂ expp0 (B r (p0 )), B r (p0 ) = {v ∈ Tp0 M | |v| ≤ r} .
Then B := {v ∈ Tp0 M | |v| ≤ r, expp0 (v) ∈ K} is a closed and bounded sub-
set of the Euclidean space Tp0 M . Hence B is compact and K = expp0 (B).
Since the exponential map expp0 : Tp0 M → M is continuous it follows that K
is compact. This shows that (ii) implies (iii).
4.6. COMPLETENESS AND HOPF–RINOW 211

We prove that (iii) implies (iv). Thus assume that every closed and
bounded subset of M is compact and choose a Cauchy sequence pi ∈ M .
Choose i0 ∈ N such that d(pi , pj ) ≤ 1 for all i, j ∈ N with i, j ≥ i0 . Define
c := max d(p1 , pi ) + 1.
1≤i≤i0

Then d(p1 , pi ) ≤ d(p1 , pi0 ) + d(pi0 , pi ) ≤ d(p1 , pi0 ) + 1 ≤ c for all i ≥ i0 and
so d(p1 , pi ) ≤ c for all i ∈ N. Hence the set {pi | i ∈ N} is bounded and so
is its closure. By (iii) this implies that the sequence pi has a convergent
subsequence. Since pi is a Cauchy sequence, this implies that pi converges.
Thus we have proved that (iii) implies (iv).
We prove that (iv) implies (v). Fix a smooth curve ξ : R → Rm and
an element (p0 , e0 ) ∈ F(M ). Assume, by contradiction, that there exists
a real number T > 0 such that there exists a solution β : [0, T ) → F(M )
of equation (4.6.1) that cannot be extended to the interval [0, T + ε) for
any ε > 0. Write β(t) =: (γ(t), e(t)) so that γ and e satisfy the equations
γ̇(t) = e(t)ξ(t), ė(t) = hγ(t) (γ̇(t))e(t), γ(0) = p0 , e(0) = e0 .
⊥ M for all η ∈ Rm and therefore
This implies e(t)η ∈ Tγ(t) M and ė(t)η ∈ Tγ(t)
d d
hη, e(t)T e(t)ζi = he(t)η, e(t)ζi = hė(t)η, e(t)ζi + he(t)η, ė(t)ζi = 0
dt dt
for all η, ζ ∈ Rm and all t ∈ [0, T ). Thus the function t 7→ e(t)T e(t) is con-
stant, hence
|e(t)η|
e(t)T e(t) = eT
0 e0 , ke(t)k = sup = ke0 k (4.6.2)
06=η∈Rm |η|
for 0 ≤ t < T , hence
|γ̇(t)| = |e(t)ξ(t)| ≤ ke0 k |ξ(t)| ≤ ke0 k sup |ξ(s)| =: cT
0≤s≤T

and so d(γ(s), γ(t)) ≤ L(γ|[s,t] ) ≤ (t − s)cT for 0 ≤ s < t < T . Since (M, d)
is a complete metric space, this shows that the limit p1 := limt%T γ(t) ∈ M
exists. Thus the set K := γ([0, T )) ∪ {p1 } ⊂ M is compact and so is the set
n o
Ke := (p, e) ∈ F(M ) | p ∈ K, eT e = eT e0 ⊂ F(M ).
0

By equation (4.6.2) the curve [0, T ) → R × F(M ) : t 7→ (t, γ(t), e(t)) takes
values in the compact set [0, T ] × K
e and is the integral curve of a vector field
on the manifold R × F(M ). Hence Corollary 2.4.15 asserts that [0, T ) cannot
be the maximal existence interval of this integral curve, a contradiction. This
shows that (iv) implies (v).
212 CHAPTER 4. GEODESICS

That (v) implies (vi) follows by taking ξ(t) ≡ ξ in (v).

We prove that (vi) implies (i). Fix an element p0 ∈ M and a tan-
gent vector v0 ∈ Tp0 M . Let e0 ∈ Liso (Rm , Tp0 M ) be any isomorphism and
choose ξ ∈ Rm such that e0 ξ = v0 . By (vi) the vector field Bξ has a unique
integral curve R → F(M ) : t 7→ β(t) = (γ(t), e(t)) with
β(0) = (p0 , e0 ).
Thus
γ̇(t) = e(t)ξ, ė(t) = hγ(t) (e(t)ξ)e(t),
and hence
γ̈(t) = ė(t)ξ = hγ(t) (e(t)ξ)e(t)ξ = hγ(t) (γ̇(t), γ̇(t)).
By the Gauß–Weingarten formula, this implies ∇γ̇(t) = 0 for every t and
hence γ : R → M is a geodesic with γ(0) = p0 and γ̇(0) = e0 ξ = v0 . Thus M
is geodesically complete and this shows that (vi) implies (i).
The equivalence of (v) and (vii) was established in Corollary 3.5.25 and
this shows that Theorem 4.6.6 implies Theorem 4.6.5.

4.6.3 Proof of the Hopf–Rinow Theorem

The proof of Theorem 4.6.6 relies on the next two lemmas.
Lemma 4.6.7. Let M ⊂ Rn be a connected m-manifold and p ∈ M . Sup-
pose ε > 0 is smaller than the injectivity radius of M at p and denote
Sε (p) := p0 ∈ M | d(p, p0 ) = ε .

Σ1 (p) := {v ∈ Tp M | |v| = 1} ,
Then the map Σ1 (p) → Sε (p) : v 7→ expp (εv) is a diffeomorphism and, for
all q ∈ M , we have
d(p, q) > ε =⇒ d(Sε (p), q) = d(p, q) − ε.
Proof. By Theorem 4.4.4, we have
d(p, expp (v)) = |v| for all v ∈ Tp M with |v| ≤ ε
and
d(p, p0 ) > ε for all p0 ∈ M \ expp (v) | v ∈ Tp M, |v| ≤ ε .


This shows that Sε (p) = expp (εΣ1 (p)) and, since ε is smaller than the injec-
tivity radius, the map
Σ1 (p) → Sε (p) : v 7→ expp (εv)
is a diffeomorphism.
4.6. COMPLETENESS AND HOPF–RINOW 213

To prove the second assertion, let q ∈ M such that

r := d(p, q) > ε.

Fix a constant δ > 0 and choose a smooth curve γ : [0, 1] → M such that

γ(0) = p, γ(1) = q, L(γ) ≤ r + δ.

Choose t0 > 0 such that γ(t0 ) is the last point of the curve on Sε (p), i.e.

γ(t0 ) ∈ Sε (p), γ(t) ∈

/ Sε (p) for t0 < t ≤ 1.

Then

d(γ(t0 ), q) ≤ L(γ|[t0 ,1] )

= L(γ) − L(γ|[0,t0 ] )
≤ L(γ) − ε
≤ r + δ − ε.

This shows that d(Sε (p), q) ≤ r + δ − ε for every δ > 0 and therefore

d(Sε (p), q) ≤ r − ε.

Moreover,
d(p0 , q) ≥ d(p, q) − d(p, p0 ) = r − ε

for all p0 ∈ Sε (p). Thus

d(Sε (p), q) = r − ε

and this proves Lemma 4.6.7.

Lemma 4.6.8 (Curve Shortening Lemma). Let M ⊂ Rn be an m-mani-

fold, let p ∈ M , and let ε be a real number such that

0 < ε < inj(p; M ).

Then, for all v, w ∈ Tp M , we have

|v| = |w| = ε, d(expp (v), expp (w)) = 2ε =⇒ v + w = 0.

214 CHAPTER 4. GEODESICS

Figure 4.7: Two unit tangent vectors.

Proof. We will prove that, for all v, w ∈ Tp M , we have

d(expp (δv), expp (δw))
lim = |v − w| . (4.6.3)
δ→0 δ
Assume this holds and suppose, by contradiction, that there exist two tan-
gent vectors v, w ∈ Tp M such that
|v| = |w| = 1, d(expp (εv), expp (εw)) = 2ε, v + w 6= 0.
Then
|v − w| < 2
(see Figure 4.7). Thus by (4.6.3) there exists a constant 0 < δ < ε such that
d(expp (δv), expp (δw)) < 2δ.
Then
d(expp (εv), expp (εw))
≤ d(expp (εv), expp (δv)) + d(expp (δv), expp (δw)) + d(expp (δw), expp (εw))
< ε − δ + 2δ + ε − δ = 2ε
and this contradicts our assumption.
It remains to prove (4.6.3). For this we observe that
d(expp (δv), expp (δw))
lim
δ→0 δ  
d(expp (δv), expp (δw)) expp (δv) − expp (δw)
= lim  
δ→0 expp (δv) − expp (δw) δ
 
expp (δv) − expp (δw)
= lim
δ→0 δ
 expp (δv) − p expp (δw) − p 
 
= lim 
 − 
δ→0 δ δ 
= |v − w| .
Here the second equality follows from Lemma 4.2.7.
4.6. COMPLETENESS AND HOPF–RINOW 215

Proof of Theorem 4.6.6. By assumption M ⊂ Rn is a connected submani-

fold, and p ∈ M is given such that the exponential map expp : Tp M → M is
defined on the entire tangent space at p. Fix a point q ∈ M \ {p} so that

0 < r := d(p, q) < ∞.

Choose a constant ε > 0 smaller than the injectivity radius of M at p and

smaller than r. Then, by Lemma 4.6.7, we have

d(Sε (p), q) = r − ε.

Hence there exists a tangent vector v ∈ Tp M such that

d(expp (εv), q) = r − ε, |v| = 1.

Define the curve γ : [0, r] → M by

γ(t) := expp (tv) for 0 ≤ t ≤ r.

By Lemma 4.3.6, this is a geodesic and it satisfies γ(0) = p. We must

prove that γ(r) = q and L(γ) = d(p, q). Instead we will prove the follow-
ing stronger statement.
Claim. For every t ∈ [0, r] we have

d(γ(t), q) = r − t.

In particular, γ(r) = q and L(γ) = r = d(p, q).

Consider the subset

I := {t ∈ [0, r] | d(γ(t), q) = r − t} ⊂ [0, r].

This set is nonempty, because ε ∈ I, it is obviously closed, and

t∈I =⇒ [0, t] ⊂ I. (4.6.4)

Namely, if t ∈ I and 0 ≤ s ≤ t, then

d(γ(s), q) ≤ d(γ(s), γ(t)) + d(γ(t), q) ≤ t − s + r − t = r − s

and
d(γ(s), q) ≥ d(p, q) − d(p, γ(s)) ≥ r − s.
Hence d(γ(s), q) = r − s and hence s ∈ I. This proves (4.6.4).
216 CHAPTER 4. GEODESICS

We prove that I is open (in the relative topology of [0, r]). Let t ∈ I
be given with t < r. Choose a constant ε > 0 smaller than the injectivity
radius of M at γ(t) and smaller than r − t. Then, by Lemma 4.6.7 with p
replaced by γ(t), we have
d(Sε (γ(t)), q) = r − t − ε.
Next we choose w ∈ Tγ(t) M such that
|w| = 1, d(expγ(t) (εw), q) = r − t − ε.
Then
d(γ(t − ε), expγ(t) (εw)) ≥ d(γ(t − ε), q) − d(expγ(t) (εw), q)
= (r − t + ε) − (r − t − ε)
= 2ε.
The converse inequality is obvious, because both points have distance ε
to γ(t) (see Figure 4.8).

Sε (γ (t)) exp (ε w)
γ(t)
γ(t)
ε ε q
p r−t−ε

Figure 4.8: The proof of the Hopf–Rinow theorem.

Thus we have proved that

d(γ(t − ε), expγ(t) (εw)) = 2ε.
Since
γ(t − ε) = expγ(t) (−εγ̇(t)),
it follows from Lemma 4.6.8 that
w = γ̇(t).
Hence expγ(t) (sw) = γ(t + s) and this implies that
d(γ(t + ε), q) = r − t − ε.
Thus t + ε ∈ I and, by (4.6.4), we have [0, t + ε] ∈ I. Thus we have proved
that I is open. In other words, I is a nonempty subset of [0, r] which is
both open and closed, and hence I = [0, r]. This proves the claim and
Theorem 4.6.6.
4.7. GEODESICS IN THE INTRINSIC SETTING* 217

4.7 Geodesics in the Intrinsic Setting*

4.7.1 Intrinsic Distance
Let M be a connected smooth manifold (§2.8) equipped with a Riemannian
metric (§3.7). Then we can define the length of a curve γ : [0, 1] → M
by the formula (4.1.1) and it is invariant under reparametrization as in
Remark 4.1.3. The distance function d : M × M → R is then given by
the same formula (4.2.2). We prove that it still defines a metric on M and
that this metric induces the same topology as the smooth structure.
Lemma 4.7.1. Let M be a connected smooth Riemannian manifold and
define the function d : M × M → [0, ∞) by (4.1.1), (4.2.1), and (4.2.2).
Then d is a metric and induces the same topology as the smooth structure.
Proof. The proof has three steps.
Step 1. Fix a point p0 ∈ M and let φ : U → Ω be a coordinate chart of M
onto an open subset Ω ⊂ Rm such that p0 ∈ U . Then there exists an open
neighborhood V ⊂ U of p0 and constants δ, r > 0 such that

δ |φ(p) − φ(p0 )| ≤ d(p, p0 ) ≤ δ −1 |φ(p) − φ(p0 )| (4.7.1)

for every p ∈ V and d(p, p0 ) ≥ δr for every p ∈ M \ V .

Denote the inverse of the coordinate chart φ by ψ := φ−1 : Ω → M and
define the map g = (gij )m m×m by g (x) := h ∂ψ (x), ∂ψ (x)i
i,j=1 : Ω → R ij ∂xi ∂xj ψ(x)
for x ∈ Ω. Then a smooth curve γ : [0, 1] → U has the length
Z 1q
L(γ) = ċ(t)T g(c(t))ċ(t) dt, c(t) := φ(γ(t)). (4.7.2)
0

Let x0 := φ(p0 ) ∈ Ω and choose r > 0 such that B r (x0 ) ⊂ Ω. Then there is
a constant δ ∈ (0, 1] such that
q
δ |ξ| ≤ ξ T g(x)ξ ≤ δ −1 |ξ| (4.7.3)

for all x ∈ Br (x0 ) and ξ, η ∈ Rm . Define V := φ−1 (Br (x0 )) ⊂ U .

Now let p ∈ V and denote x := φ(p) ∈ Br (x0 ). Then, for every smooth
curve γ : [0, 1] → V with γ(0) = p0 and γ(1) = p, the curve c := φ ◦ γ takes
values in Br (x0 ) and satisfies c(0) = x0 and c(1) = x. Hence, by (4.7.2)
and (4.7.3), we have
Z 1 Z 1 
 
L(γ) ≥ δ |ċ(t)| dt ≥ δ 
 ċ(t) dt = δ |x − x0 | .
0 0
218 CHAPTER 4. GEODESICS

If γ : [0, 1] → M is a smooth curve with endpoints γ(0) = p0 and γ(1) = p

whose image is not entirely contained in V , then there exists a T ∈ (0, 1]
such that γ(t) ∈ V for 0 ≤ t < T and γ(T ) ∈ ∂V , so c(t) = φ(γ(t)) ∈ Br (x0 )
for 0 ≤ t < T and |c(T ) − x0 | = r. Hence, by the above argument, we have

L(γ) ≥ δr.

This shows that d(p0 , p) ≥ δr for p ∈ M \ V and d(p0 , p) ≥ δ |φ(p) − φ(p0 )|

for p ∈ V . If p ∈ V , x := φ(p), and c(t) := x0 + t(x − x0 ), then γ := ψ ◦ c is a
smooth curve in V with γ(0) = p0 and γ(1) = p and, by (4.7.2) and (4.7.3),
Z 1
L(γ) ≤ δ −1 |ċ(t)| dt = δ −1 |x − x0 | .
0

This proves Step 1.

Step 2. d is a distance function.
Step 1 shows that d(p, p0 ) > 0 for every p ∈ M \ {p0 } and hence d satisfies
condition (i) in Lemma 4.2.3. The proofs of (ii) and (iii) remain unchanged
in the intrinsic setting and this proves Step 2.
Step 3. The topology on M induced by d agrees with the topology induced
by the smooth structure.
Assume first that W ⊂ M is open with respect to the manifold topology
and let p0 ∈ W . Let φ : U → Ω be a coordinate chart of M onto an open
subset Ω ⊂ Rm such that p0 ∈ U , and choose V ⊂ U and δ, r as in Step 1.
Then φ(V ∩ W ) is an open subset of Ω containing the point φ(p0 ). Hence
there exists a constant 0 < ε ≤ δr such that Bδ−1 ε (φ(p0 )) ⊂ φ(V ∩ W ).
Thus by Step 1 we have d(p, p0 ) ≥ δr ≥ ε for all p ∈ M \ V . Hence, if p ∈ M
satisfies d(p, p0 ) < ε, then p ∈ V , so |φ(p) − φ(p0 )| < δ −1 d(p, p0 ) < δ −1 ε
by (4.7.1), and therefore φ(p) ∈ φ(V ∩ W ). Thus Bε (p0 ; d) ⊂ W and this
shows that W is open with respect to d.
Conversely, assume that W ⊂ M be open with respect to d and choose
a coordinate chart φ : U → Ω onto an open set Ω ⊂ Rm . We must prove
that φ(U ∩ W ) is an open subset of Ω. To see this, choose x0 ∈ φ(U ∩ W )
and let p0 := φ−1 (x0 ) ∈ U ∩ W . Now choose V ⊂ U and δ, r as in Step 1.
Choose ε > 0 such that Bδ−1 ε (p0 ; d) ⊂ W and Bε (x0 ) ⊂ φ(V ). Let x ∈ Rn
such that |x − x0 | < ε. Then x ∈ φ(V ) and therefore p := φ−1 (x) ∈ V . This
implies d(p, p0 ) < δ −1 |φ(p) − φ(p0 )| = δ −1 |x − x0 | < δ −1 ε, thus p ∈ W ∩ U ,
and so x = φ(p) ∈ φ(W ∩ U ). Thus φ(W ∩ U ) is an open, and so W is open
in the manifold topology of M . This proves Step 3 and Lemma 4.7.1.
4.7. GEODESICS IN THE INTRINSIC SETTING* 219

4.7.2 Geodesics and the Levi-Civita Connection

With the covariant derivative understood (Theorem 3.7.8), we can define
geodesics on M as smooth curves γ : I → M that satisfy the equation

∇γ̇ = 0,

as in Definition 4.1.5. Then all the above results about geodesics, as well
as their proofs, carry over almost verbatim to the intrinsic setting. In
particular, geodesics are in local coordinates described by equation (4.3.6)
(Lemma 4.3.15) and they are the critical points of the energy functional

1 1
Z
E(γ) := |γ̇(t)|2 dt
2 0
on the space Ωp,q of all paths γ : [0, 1] → M with fixed endpoints γ(0) = p
and γ(1) = q. Here we use the fact that Lemma 4.1.7 extends to the in-
trinsic setting via the Embedding Theorem 2.9.12. So for every vector
field X ∈ Vect(γ) along γ with X(0) = 0 and X(1) = 0 there exists a curve
of curves R → Ωp,q : s 7→ γs with γ0 = γ and ∂s γs |s=0 = X. Then, by the
properties of the Levi-Civita connection, we have

1 1
Z
dE(γ)X = ∂s |∂t γs (t)|2 dt
2 0
Z 1
= hγ̇(t), ∇t X(t)i dt
0
Z 1
=− h∇t γ̇(t), X(t)i dt.
0

The right hand side vanishes for all X if and only if ∇γ̇ ≡ 0 (Theorem 4.1.4).
With this understood, we find that, for all p ∈ M and v ∈ Tp M , there exists
a unique geodesic γ : Ip,v → M on a maximal open interval Ip,v ⊂ R con-
taining zero that satisfies γ(0) = p and γ̇(0) = v (Lemma 4.3.4). This gives
rise to a smooth exponential map expp : Vp = {v ∈ Tp M | 1 ∈ Ip,v } → M as
in §4.3 which satisfies d expp (0) = id : Tp M → Tp M as in Corollary 4.3.7.
This leads directly to the injectivity radius, the Gauß Lemma 4.4.5, the local
length minimizing property of geodesics in Theorem 4.4.4, and the Convex
Neighborhood Theorem 4.5.3. Also the proof of the equivalence of metric
and geodesic completeness in Theorem 4.6.5 and of the Hopf–Rinow Theo-
rem 4.6.6 carry over verbatim to the intrinsic setting of general Riemannian
manifolds. The only place where some care must be taken is in the proof of
the Curve Shortening Lemma 4.6.8 as is spelled out in Exercise 4.7.2 below.
220 CHAPTER 4. GEODESICS

4.7.3 Examples and Exercises

Exercise 4.7.2. Choose a coordinate chart φ : U → Ω with φ(p0 ) = 0 such
that the metric in local coordinates satisfies

gij (0) = δij .

Refine the estimate (4.7.1) in the proof of Lemma 4.7.1 and show that
d(p, q)
lim = 1.
p,q→p0 |φ(p) − φ(q)|
This is the intrinsic analogue of Lemma 4.2.8. Use this to prove that equa-
tion (4.6.3) continues to hold for all Riemannian manifolds, i.e.
d(expp (δv), expp (δw))
lim = |v − w|
δ→0 δ
for p ∈ M and v, w ∈ Tp M . With this understood, the proof of the Curve
Shortening Lemma 4.6.8 carries over verbatim to the intrinsic setting.
Exercise 4.7.3. The real projective space RPn inherits a Riemannian met-
ric from S n as it is a quotient of S n by an isometric involution. Prove that
each geodesic in S n with its standard metric descends to a geodesic in RPn .
Exercise 4.7.4. Let f : S 3 → S 2 be the Hopf fibration defined by
 
f (z, w) = |z|2 − |w|2 , 2Re z̄w, 2Im z̄w

Prove that the image of a great circle in S 3 is a nonconstant geodesic in S 2

if and only if it is orthogonal to the fibers of f , which are also great circles.
Here we identify S 3 with the unit sphere in C2 . (See also Exercise 2.5.21.)
Exercise 4.7.5. Prove that a nonconstant geodesic γ : R → S 2n+1 de-
scends to a nonconstant geodesic in CPn with the Fubini–Study metric (see
Example 3.7.5) if and only if γ̇(t) ⊥ Cγ(t) for every t ∈ R.
Exercise 4.7.6. Consider the manifold
n o
Fk (Rn ) := D ∈ Rn×k  DT D = 1l


of orthonormal k-frames in Rn , equipped with the Riemannian metric in-

herited from the standard inner product

hX, Y i := trace(X T Y )

on the space of real n × k-matrices.

4.7. GEODESICS IN THE INTRINSIC SETTING* 221

(a) Prove that

n o
TD Fk (Rn ) = X ∈ Rn×k  DT X + X T D = 0 ,

n o
TD Fk (Rn )⊥ = DA  A = AT ∈ Rk×k .


and that the orthogonal projection Π(D) : Rn×k → TD Fk (Rn ) is given by

1
Π(D)X = X − D DT X + X T D .


2
(b) Prove that the second fundamental form of Fk (Rn ) is given by
1
hD (X)Y = − D X T Y + Y T X


2
for D ∈ Fk (Rn ) and X, Y ∈ TD Fk (Rn ).
(c) Prove that a smooth map R → Fk (Rn ) : t 7→ D(t) is a geodesic if and
only if it satisfies the differential equation
D̈ = −DḊT Ḋ. (4.7.4)
Prove that the function DT Ḋ is constant for every geodesic in Fk (Rn ). Com-
pare this with Example 4.3.12.
Exercise 4.7.7. Let Gk (Rn ) = Fk (Rn )/O(k) be the real Grassmannian of
k-dimensional subspaces in Rn , equipped with a Riemannian metric as in
Example 3.7.6. Prove that a geodesics R → Fk (Rn ) : t 7→ D(t) descends
to a nonconstant geodesic in Gk (Rn ) if and only if DT Ḋ ≡ 0 and Ḋ 6≡ 0.
Deduce that the exponential map on Gk (Rn ) is given by
  1/2   −1/2  1/2 
Tb Tb Tb
expΛ (Λ) = im D cos
b D D
b +D D D
b b sin D D
b

for Λ ∈ Fk (Rn ) and Λ b ∈ TΛ Fk (Rn ) \ {0}. Here we identify the tangent

space TΛ Fk (Rn ) with the space of linear maps from Λ to Λ⊥ , and choose
the matrices D ∈ Fk (Rn ) and Db ∈ Rn×k such that

Λ = im D, DT D
b = 0, Λ b : Rk → Λ⊥ = ker DT .
b ◦D =D

Prove that the group O(n) acts on Gk (Rn ) by isometries. Which subgroup
acts trivially?
Exercise 4.7.8. Carry over Exercises 4.7.6 and 4.7.7 to the complex Grass-
mannian Gk (Cn ). Prove that the group U(n) acts on Gk (Cn ) by isometries.
Which subgroup acts trivially?
222 CHAPTER 4. GEODESICS
Chapter 5

Curvature

This chapter begins by introducing the notion of an isometry (§5.1). It

shows that isometries of embedded manifolds preserve the lengths of curves
and can be characterized as diffeomorphisms whose derivatives preserve the
inner products. The chapter then moves on to the Riemann curvature ten-
sor (§5.2). The next section is devoted to the generalized Gauß Theorema
Egregium which asserts that isometries preserve geodesics, the covariant
derivative, and the Riemann curvature tensor (§5.3). The chapter includes
a study of the isometries of a compact Lie group with a bi-invariant metric
and a proof of the Myers–Steenrod Theorem, which asserts that the group
of isometries of any Riemannian manifold is a finite-dimensional Lie group
(§5.4). The final section examines the Riemann curvature tensor in local co-
ordinates and shows how the definitions and results of the present chapter
carry over to the intrinsic setting of Riemannian manifolds (§5.5).

5.1 Isometries
Characterization of Isometries
Let M and M 0 be connected submanifolds of Rn . An isometry is an isomor-
phism of the intrinsic geometries of M and M 0 . Recall the definition of the
intrinsic distance function
d : M × M → [0, ∞)
in §4.2 by Z 1
d(p, q) := inf L(γ), L(γ) = |γ̇(t)| dt
γ∈Ωp,q 0
for p, q ∈ M . Let d0 denote the intrinisic distance function on M 0 .

223
224 CHAPTER 5. CURVATURE

Theorem 5.1.1 (Isometries). Let φ : M → M 0 be a bijective map. Then

the following are equivalent.
(i) φ intertwines the distance functions on M and M 0 , i.e.

d0 (φ(p), φ(q)) = d(p, q)

for all p, q ∈ M .
(ii) φ is a diffeomorphism and

dφ(p) : Tp M → Tφ(p) M 0

is an orthogonal isomorphism for every p ∈ M .

(iii) φ is a diffeomorphism and

L(φ ◦ γ) = L(γ)

for every smooth curve γ : [a, b] → M .

The bijection φ is called an isometry iff it satisfies these equivalent condi-
tions. In the case M = M 0 the isometries φ : M → M form a group denoted
by I(M ) and called the isometry group of M .

Proof. See page 226.

Lemma 5.1.2. For every p ∈ M there exists a constant ε > 0 such that,
for all v, w ∈ Tp M with 0 < |w| < |v| < ε, we have

|w|
d(expp (w), expp (v)) = |v| − |w| =⇒ w= v. (5.1.1)
|v|

Proof. See page 225.

Remark 5.1.3. It follows from the triangle inequality and Theorem 4.4.4
that

d(expp (v), expp (w)) ≥ d(expp (v), p) − d(expp (w), p) = |v| − |w|

whenever 0 < |w| < |v| < inj(p). Lemma 5.1.2 asserts that equality can
only hold when w is a positive multiple of v or, to put it differently, that the
distance between expp (v) and expp (w) must be strictly bigger that |v| − |w|
whenever w is not a positive multiple of v.
5.1. ISOMETRIES 225

Proof of Lemma 5.1.2. As in Corollary 4.3.8 we denote

Bε (p) := {v ∈ Tp M | |v| < ε} ,

Uε (p) := {q ∈ M | d(p, q) < ε} .

By Theorem 4.4.4 and the definition of the injectivity radius, the exponential
map at p is a diffeomorphism expp : Bε (p) → Uε (p) for ε < inj(p). Choose
0 < r < inj(p). Then the closure of Ur (p) is a compact subset of M . Hence
there is a constant ε > 0 such that ε < r and ε < inj(p0 ) for every p0 ∈ Ur (p).
Since ε < r we have

ε < inj(p0 ) ∀ p0 ∈ Uε (p). (5.1.2)

Thus expp0 : Bε (p0 ) → Uε (p0 ) is a diffeomorphism for every p0 ∈ Uε (p).

Define p1 := expp (w) and p2 := expp (v). Then, by assumption, we have
d(p1 , p2 ) = |v| − |w| < ε. Since p1 ∈ Uε (p) it follows from our choice of ε that
ε < inj(p1 ). Hence there is a unique tangent vector v1 ∈ Tp1 M such that

|v1 | = d(p1 , p2 ) = |v| − |w| , expp1 (v1 ) = p2 .

Following first the shortest geodesic from p to p1 and then the shortest
geodesic from p1 to p2 we obtain (after suitable reparametrization) a smooth
γ : [0, 2] → M such that

γ(0) = p, γ(1) = p1 , γ(2) = p2 ,

and

L(γ|[0,1] ) = d(p, p1 ) = |w| , L(γ|[1,2] ) = d(p1 , p2 ) = |v| − |w| .

Thus L(γ) = |v| = d(p, p2 ). Hence, by Theorem 4.4.4, there is a smooth

function β : [0, 2] → [0, 1] satisfying

β(0) = 0, β(2) = 1, β̇(t) ≥ 0, γ(t) = expp (β(t)v)

for every t ∈ [0, 2]. This implies

expp (w) = p1 = γ(1) = expp (β(1)v), 0 ≤ β(1) ≤ 1.

Since w and β(1)v are both elements of Bε (p) and expp is injective on Bε (p),
this implies w = β(1)v. Since β(1) ≥ 0 we have β(1) = |w| / |v|. This
proves (5.1.1) and Lemma 5.1.2.
226 CHAPTER 5. CURVATURE

Proof of Theorem 5.1.1. That (ii) implies (iii) follows from the definition of
the length of a curve. Namely
Z b 
d 
L(φ ◦ γ) =  dt φ(γ(t)) dt
 
a
Z b
= |dφ(γ(t))γ̇(t)| dt
a
Z b
= |γ̇(t)| dt
a
= L(γ).
In the third equation we have used (ii). That (iii) implies (i) follows imme-
diately from the definition of the intrinsic distance functions d and d0 .
We prove that (i) implies (ii). Fix a point p ∈ M and choose ε > 0 so
small that ε < inj(p) and that the assertion of Lemma 5.1.2 holds for the
point p0 := φ(p) ∈ M 0 . Then there is a unique homeomorphism
Φp : Bε (p) → Bε (φ(p))
such that the following diagram commutes.
Φp
Tp M ⊃ Bε (p) / Bε (φ(p)) ⊂ Tφ(p) M 0 .
expp exp0φ(p)
 φ 
M ⊃ Uε (p) / Uε (φ(p)) ⊂ M0
Here the vertical maps are diffeomorphisms and φ : Uε (p) → Uε (φ(p)) is a
homeomorphism by (i). Hence Φp : Bε (p) → Bε (φ(p)) is a homeomorphism.
Claim 1. The map Φp satisfies the following equations for every v ∈ Bε (p)
and every t ∈ [0, 1]:
exp0φ(p) (Φp (v)) = φ(expp (v)), (5.1.3)
|Φp (v)| = |v| , (5.1.4)
Φp (tv) = tΦp (v). (5.1.5)
Equation (5.1.3) holds by definition. To prove (5.1.4) we observe that, by
Theorem 4.4.4, we have
|Φp (v)| = d0 (φ(p), exp0φ(p) (Φp (v)))
= d0 (φ(p), φ(expp (v)))
= d(p, expp (v))
= |v| .
5.1. ISOMETRIES 227

Here the second equation follows from (5.1.3) and the third equation from (i).
Equation (5.1.5) holds for t = 0 because Φp (0) = 0 and for t = 1 it is a
tautology. Hence assume 0 < t < 1. Then

d0 (exp0φ(p) (Φp (tv)), exp0φ(p) (Φp (v))) = d0 (φ(expp (tv)), φ(expp (v)))
= d(expp (tv), expp (v))
= |v| − |tv|
= |Φp (v)| − |Φp (tv)| .

Here the first equation follows from (5.1.3), the second equation from (i),
the third equation from Theorem 4.4.4 and the fact that |v| < inj(p), and
the last equation follows from (5.1.4). Since 0 < |Φp (tv)| < |Φp (v)| < ε we
can apply Lemma 5.1.2 and obtain
|Φp (tv)|
Φp (tv) = Φp (v) = tΦp (v).
|Φp (v)|
This proves Claim 1.
By Claim 1, Φp extends to a bijective map Φp : Tp M → Tφ(p) M 0 via
1
Φp (v) := Φp (δv),
δ
where δ > 0 is chosen so small that δ |v| < ε. The right hand side of
this equation is independent of the choice of δ. Hence the extension is well
defined. It is bijective because the original map Φp is a bijection from Bε (p)
to Bε (φ(p)). The reader may verify that the extended map satisfies the
conditions (5.1.4) and (5.1.5) for all v ∈ Tp M and all t ≥ 0.
Claim 2. The extended map Φp : Tp M → Tφ(p) M 0 is linear and preserves
the inner product.
It follows from the equation (4.6.3) in the proof of Lemma 4.6.8 that
d(expp (tv), expp (tw))
|v − w| = lim
t→0 t
d0 (φ(expp (tv)), φ(expp (tw)))
= lim
t→0 t
d0 (exp0φ(p) (Φp (tv))), exp0φ(p) (Φp (tw))))
= lim
t→0 t
d (expφ(p) (tΦp (v))), exp0φ(p) (tΦp (w))))
0 0
= lim
t→0 t
= |Φp (v) − Φp (w)| .
228 CHAPTER 5. CURVATURE

Here the second equation follows from (i), the third from (5.1.3), the fourth
from (5.1.4), and the last equation follows again from (4.6.3). By polariza-
tion we obtain

2hv, wi = |v|2 + |w|2 − |v − w|2

= |Φp (v)|2 + |Φp (w)|2 − |Φp (v) − Φp (w)|2
= 2hΦp (v), Φp (w)i.

Thus Φp preserves the inner product. Hence, for all v1 , v2 , w ∈ Tp M , we

have

hΦp (v1 + v2 ), Φp (w)i = hv1 + v2 , wi

= hv1 , wi + hv2 , wi
= hΦp (v1 ), Φp (w)i + hΦp (v2 ), Φp (w)i
= hΦp (v1 ) + Φp (v2 ), Φp (w)i.

Since Φp is surjective, this implies

Φp (v1 + v2 ) = Φp (v1 ) + Φp (v2 )

for all v1 , v2 ∈ Tp M . With v1 = v and v2 = −v we obtain

Φp (−v) = −Φp (v)

for every v ∈ Tp M and by (5.1.5) this gives

Φp (tv) = tΦp (v)

for all v ∈ Tp M and t ∈ R. This proves Claim 2.

Claim 3. φ is smooth and dφ(p) = Φp .

By (5.1.3) we have

φ = exp0φ(p) ◦Φp ◦ exp−1

p : Uε (p) → Uε (φ(p)).

Since Φp is linear, this shows that the restriction of φ to the open set Uε (p)
is smooth. Moreover, for every v ∈ Tp M we have
 
d d
dφ(p)v =  φ(expp (tv)) =  exp0φ(p) (tΦp (v)) = Φp (v).
dt t=0 dt t=0

Here we have used equations (5.1.3) and (5.1.5) as well as Lemma 4.3.6.
This proves Claim 3 and Theorem 5.1.1.
5.1. ISOMETRIES 229

Exercise 5.1.4. Prove that every isometry ψ : Rn → Rn is an affine map

ψ(p) = Ap + b

where A ∈ O(n) and b ∈ Rn . Thus ψ is a composition of translation and

rotation. Hint: Let e1 , . . . , en be the standard basis of Rm . Prove that any
two vectors v, w ∈ Rn that satisfy

|v| = |w|

and
|v − ei | = |w − ei | for i = 1, . . . , n
must be equal.

Remark 5.1.5. If ψ : Rn → Rn is an isometry of the ambient Euclidean

space with ψ(M ) = M 0 , then certainly φ := ψ|M is an isometry from M
onto M 0 . On the other hand, if M is a plane manifold

M = {(0, y, z) ∈ R3 | 0 < y < π/2}

and M 0 is the cylindrical manifold

M 0 = {(x, y, z) ∈ R3 | x2 + y 2 = 1, x > 0, y > 0},

Then the map φ : M → M 0 defined by

φ(0, y, z) := (cos(y), sin(y), z)

is an isometry which is not of the form φ = ψ|M . Indeed, an isometry of the

form φ = ψ|M necessarily preserves the second fundamental form (as well
as the first) in the sense that

dψ(p)hp (v, w) = h0ψ(p) (dψ(p)v, dψ(p)w)

for v, w ∈ Tp M but in the example h vanishes identically while h0 does not.

We may thus distinguish two fundamental question:

I. Given M and M 0 when are they extrinsically isomorphic, i.e. when is

there an ambient isometry ψ : Rn → Rn with ψ(M ) = M 0 ?

II. Given M and M 0 when are they intrinsically isomorphic, i.e. when is
there an isometry φ : M → M 0 from M onto M 0 ?
230 CHAPTER 5. CURVATURE

As we have noted, both the first and second fundamental forms are
preserved by extrinsic isomorphisms while only the first fundamental form
need be preserved by an intrinsic isomorphism (i.e. an isometry).
A question which occurred to Gauß (who worked for a while as a cartog-
rapher) is this: Can one draw a perfectly accurate map of a portion of the
earth? (i.e. a map for which the distance between points on the map is pro-
portional to the distance between the corresponding points on the surface
of the earth). We can now pose this question as follows: Is there an isom-
etry from an open subset of a sphere to an open subset of a plane? Gauß
answered this question negatively by associating an invariant, the Gaußian
curvature
K : M → R,
to a surface M ⊂ R3 . According to his Theorema Egregium
K0 ◦ φ = K
for an isometry φ : M → M 0 . The sphere has positive curvature; the plane
has zero curvature; hence the perfectly accurate map does not exist. Our
aim is to explain these ideas.

Local Isometries
We shall need a concept slightly more general than that of “isometry”.
Definition 5.1.6. A smooth map φ : M → M 0 is called a local isometry
iff its derivative
dφ(p) : Tp M → Tφ(p) M 0
is an orthogonal linear isomorphism for every p ∈ M .
0
Remark 5.1.7. Let M ⊂ Rn and M 0 ⊂ Rm be manifolds and φ : M → M 0
be a map. The following are equivalent.
(i) φ is a local isometry.
(ii) For every p ∈ M there are open neighborhoods U ⊂ M and U 0 ⊂ M 0
such that the restriction of φ to U is an isometry from U onto U 0 .
That (ii) implies (i) follows immediately from Theorem 5.1.1. On the other
hand (i) implies that dφ(p) is invertible so that (ii) follows from the inverse
function theorem.
Example 5.1.8. The map
R → S 1 : θ 7→ eiθ
is a local isometry but not an isometry.
5.1. ISOMETRIES 231

Exercise 5.1.9. Let M ⊂ Rn be a compact connected 1-manifold. Prove

that M is diffeomorphic to the circle S 1 . Define the length of a compact
connected Riemannian 1-manifold. Prove that two compact connected 1-
manifolds M, M 0 ⊂ Rn are isometric if and only if they have the same
length. Hint: Let γ : R → M be a geodesic with |γ̇(t)| ≡ 1. Show that
γ is not injective; otherwise construct an open cover of M without finite
subcover. If t0 < t1 with γ(t0 ) = γ(t1 ), show that γ̇(t0 ) = γ̇(t1 ); otherwise
show that γ(t0 + t) = γ(t1 − t) for all t and find a contradiction.
The next result asserts that two local isometries that have the same value
and the same derivative at a single point must agree everywhere, provided
that the domain is connected.
0
Lemma 5.1.10. Let M ⊂ Rn and M 0 ⊂ Rn be smooth m-manifolds and
assume that M is connected. Let φ : M → M 0 and ψ : M → M 0 be local
isometries and let p0 ∈ M such that
φ(p0 ) = ψ(p0 ) =: p00 , dφ(p0 ) = dψ(p0 ) : Tp0 M → Tp00 M 0 .
Then φ(p) = ψ(p) for every p ∈ M .
Proof. Define the set
M0 := {p ∈ M | φ(p) = ψ(p), dφ(p) = dψ(p)} .
This set is obviously closed. We prove that M0 is open. Let p ∈ M0 and
choose U ⊂ M and U 0 ⊂ M 0 as in Remark 5.1.7 (ii). Denote
Φp := dφ(p) = dψ(p) : Tp M → Tp0 M 0 , p0 := φ(p) = ψ(p)
Then it follows from equation (5.1.3) in the proof of Theorem 5.1.1 that
there exists a constant ε > 0 such that Uε (p) ⊂ U and Uε (p0 ) ⊂ U 0 and
q ∈ Uε (p) =⇒ φ(q) = exp0p0 ◦Φp ◦ exp−1
p (q) = ψ(q).

Hence Uε (p) ⊂ M0 . Thus M0 is open, closed, and nonempty. Since M is

connected it follows that M0 = M and this proves Lemma 5.1.10.
Exercise 5.1.11. (i) If a sequence of local isometries φi : M → M 0 con-
verges uniformly to a local isometry φ : M → M 0 , then it converges in the
C ∞ topology. Hint: Let p ∈ M . Then every sufficiently small tangent vec-
tor v ∈ Tp M satisfies the equation dφ(p)v = (exp0φ(p) )−1 (φ(expp (v))). Use
this to prove that dφi (p) converges to dφ(p). Deduce that φi converges to φ
uniformly with all derviatives in a neighborhood of p.
(ii) The C ∞ topology on the space of local isometries from M to M 0 agrees
with the C 0 topology.
232 CHAPTER 5. CURVATURE

5.2 The Riemann Curvature Tensor

5.2.1 Definition and Gauß–Codazzi
Let M ⊂ Rn be a smooth manifold and γ : R2 → M be a smooth map.
Denote by (s, t) the coordinates on R2 . Let Z ∈ Vect(γ) be a smooth vector
field along γ, i.e. Z : R2 → Rn is a smooth map such that Z(s, t) ∈ Tγ(s,t) M
for all s and t. The covariant partial derivatives of Z with respect to
the variables s and t are defined by
∂Z ∂Z
∇s Z := Π(γ) , ∇t Z := Π(γ) .
∂s ∂t
In particular ∂s γ = ∂γ/∂s and ∂t γ = ∂γ/∂t are vector fields along γ and we
have
∇s ∂t γ − ∇t ∂s γ = 0
as both terms on the left are equal to Π(γ)∂s ∂t γ. Thus ordinary partial
differentiation and covariant partial differentiation commute. The analogous
formula (which results on replacing ∂ by ∇ and γ by Z) is in general false.
Instead we have the following.
Definition 5.2.1. The Riemann curvature tensor assigns to each p ∈ M
the bilinear map

Rp : Tp M × Tp M → L(Tp M, Tp M )

characterized by the equation

Rp (u, v)w = ∇s ∇t Z − ∇t ∇s Z (0, 0) (5.2.1)

for u, v, w ∈ Tp M where γ : R2 → M is a smooth map and Z ∈ Vect(γ) is a

smooth vector field along γ such that

γ(0, 0) = p, ∂s γ(0, 0) = u, ∂t γ(0, 0) = v, Z(0, 0) = w. (5.2.2)

We must prove that R is well defined, i.e. that the right hand side of
equation (5.2.1) is independent of the choice of γ and Z. This follows from
the Gauß–Codazzi formula which we prove next. Recall that the second fun-
damental form can be viewed as a linear map hp : Tp M → L(Tp M, Tp M ⊥ )
and that, for u ∈ Tp M , the linear map hp (u) ∈ L(Tp M, Tp M ⊥ ) and its
dual hp (u)∗ ∈ L(Tp M ⊥ , Tp M ) are given by

hp (u)∗ w = dΠ(p)u w



hp (u)v = dΠ(p)u v,

for v ∈ Tp M and w ∈ Tp M ⊥ .
5.2. THE RIEMANN CURVATURE TENSOR 233

Theorem 5.2.2. The Riemann curvature tensor is well defined and given
by the Gauß–Codazzi formula

Rp (u, v) = hp (u)∗ hp (v) − hp (v)∗ hp (u) (5.2.3)

for u, v ∈ Tp M .

Proof. Let u, v, w ∈ Tp M and choose a smooth map γ : R2 → M and a

smooth vector field Z along γ such that (5.2.2) holds. Then, by the Gauß–
Weingarten formula (3.2.2), we have

∇t Z = ∂t Z − hγ (∂t γ)Z


= ∂t Z − dΠ(γ)∂t γ Z



= ∂t Z − ∂t Π ◦ γ Z.

Hence



∂s ∇t Z = ∂s ∂t Z − ∂s ∂t Π ◦ γ Z





= ∂s ∂t Z − ∂s ∂t Π ◦ γ Z − ∂t Π ◦ γ ∂s Z





= ∂s ∂t Z − ∂s ∂t Π ◦ γ Z − dΠ(γ)∂t γ ∇s Z + hγ (∂s γ)Z
= ∂s ∂t Z − ∂s ∂t Π ◦ γ Z − hγ (∂t γ)∇s Z − hγ (∂t γ)∗ hγ (∂s γ)Z.

Interchanging s and t and taking the difference we obtain

∂s ∇t Z − ∂t ∇s Z = hγ (∂s γ)∗ hγ (∂t γ)Z − hγ (∂t γ)∗ hγ (∂s γ)Z

+ hγ (∂s γ)∇t Z − hγ (∂t γ)∇s Z.

Here the first two terms on the right are tangent to M and the last two
terms on the right are orthogonal to Tγ M . Hence


∇s ∇t Z − ∇t ∇s Z = Π(γ) ∂s ∇t Z − ∂t ∇s Z
= hγ (∂s γ)∗ hγ (∂t γ)Z − hγ (∂t γ)∗ hγ (∂s γ)Z.

Evaluating the right hand side at s = t = 0 we find that

∇s ∇t Z − ∇t ∇s Z (0, 0) = hp (u)∗ hp (v)w − hp (v)∗ hp (u)w.

This proves the Gauß–Codazzi equation and shows that the left hand side
is independent of the choice of γ and Z. This proves Theorem 5.2.2.
234 CHAPTER 5. CURVATURE

5.2.2 Covariant Derivative of a Global Vector Field

So far we have only defined the covariant derivatives of vector fields along
curves. The same method can be applied to global vector fields. This leads
to the following definition.

Definition 5.2.3 (Covariant derivative). Let M ⊂ Rn be an m-dimen-

sional submanifold and X be a vector field on M . Fix a point p ∈ M and
a tangent vector v ∈ Tp M . The covariant derivative of X at p in the
direction v is the tangent vector

∇v X(p) := Π(p)dX(p)v ∈ Tp M,

where Π(p) ∈ Rn×n denotes the orthogonal projection onto Tp M .

Remark 5.2.4. If γ : I → M is a smooth curve on an interval I ⊂ R and

X ∈ Vect(M ) is a smooth vector field on M , then X ◦ γ is a smooth vector
field along γ. The covariant derivative of X ◦ γ is related to the covariant
derivative of X by the formula

∇(X ◦ γ)(t) = ∇γ̇(t) X(γ(t)). (5.2.4)

Remark 5.2.5 (Gauß–Weingarten formula). Differentiating the equa-

tion X = ΠX (understood as a function from M to Rn ) and using the
notation ∂v X(p) := dX(p)v for the derivative of X at p in the direction v
we obtain the Gauß–Weingarten formula for global vector fields:

∂v X(p) = ∇v X(p) + hp (v)X(p). (5.2.5)

Remark 5.2.6 (Levi-Civita connection). Differentiating a vector field

Y on M in the direction of another vector field X we obtain a vector field
∇X Y ∈ Vect(M ) defined by

(∇X Y )(p) := ∇X(p) Y (p)

for p ∈ M . This gives rise to a family of linear operators

∇X : Vect(M ) → Vect(M ),

one for every vector field X ∈ Vect(M ), and the assignment

Vect(M ) → L(Vect(M ), Vect(M )) : X 7→ ∇X

5.2. THE RIEMANN CURVATURE TENSOR 235

is itself a linear operator. This operator is called the Levi-Civita connec-

tion on the tangent bundle T M . It satisfies the conditions

∇f X (Y ) = f ∇X Y, (5.2.6)

∇X (f Y ) = f ∇X Y + (LX f ) Y (5.2.7)
LX hY, Zi = h∇X Y, Zi + hY, ∇X Zi, (5.2.8)
∇Y X − ∇X Y = [X, Y ], (5.2.9)
for all X, Y, Z ∈ Vect(M ) and f ∈ F (M ), where LX f = df ◦ X and
[X, Y ] ∈ Vect(M ) denotes the Lie bracket of the vector fields X and Y .
The next lemma asserts that the Levi-Civita connection is uniquely deter-
mined by (5.2.8) and (5.2.9).

Lemma 5.2.7 (Uniqueness Lemma). There is a unique linear operator

Vect(M ) → L(Vect(M ), Vect(M )) : X 7→ ∇X

satisfying equations (5.2.8) and (5.2.9) for all X, Y, Z ∈ Vect(M ).

Proof. Existence follows from the properties of the Levi-Civita connection.

We prove uniqueness. Let X 7→ DX be any linear operator from Vect(M )
to L(Vect(M ), Vect(M )) that satisfies (5.2.8) and (5.2.9). Then we have

LX hY, Zi = hDX Y, Zi + hY, DX Zi,

LY hX, Zi = hDY X, Zi + hX, DY Zi,
−LZ hX, Y i = −hDZ X, Y i − hX, DZ Y i.

Adding these three equations we find

LX hY, Zi + LY hZ, Xi − LZ hX, Y i

= 2hDX Y, Zi + hDY X − DX Y, Zi
+hX, DY Z − DZ Y i + hY, DX Z − DZ Xi
= 2hDX Y, Zi + h[X, Y ], Zi + hX, [Z, Y ]i + hY, [Z, X]i.

The same equation holds for the Levi-Civita connection and hence

hDX Y, Zi = h∇X Y, Zi.

This implies DX Y = ∇X Y for all X, Y ∈ Vect(M ).

236 CHAPTER 5. CURVATURE

Remark 5.2.8 (The Levi-Civita connection in local coordinates).

Let φ : U → Ω be a coordinate chart on an open set U ⊂ M with values in
an open set Ω ⊂ Rm . In such a coordinate chart a vector field X ∈ Vect(M )
is represented by a smooth map

ξ = (ξ 1 , . . . , ξ m ) : Ω → Rm

defined by
ξ(φ(p)) = dφ(p)X(p)

for p ∈ U . If Y ∈ Vect(M ) is represented by η, then ∇X Y is represented by

the map
m m
X ∂η k i X k i j
(∇ξ η)k := ξ + Γij ξ η . (5.2.10)
∂xi
i=1 i,j=1

Here the Γkij : Ω → R are the Christoffel symbols defined by

m  
X
k` 1 ∂g`i ∂g`j ∂gij
Γkij := g j
+ i
− , (5.2.11)
2 ∂x ∂x ∂x`
`=1

where gij is the metric tensor and g ij is the inverse matrix so that
X
gij g jk = δik
j

(see Lemma 3.6.5). This formula can be used to prove the existence state-
ment in Lemma 5.2.7 and hence define the Levi-Civita connection in the
intrinsic setting.

Exercise 5.2.9. In the proof of Lemma 5.2.7 we did not actually use
that the operator DX : Vect(M ) → Vect(M ) is linear nor that the oper-
ator X 7→ DX is linear. Prove directly that if a map

DX : L(M ) → L(M )

satisfies (5.2.8) for all Y, Z ∈ Vect(M ), then DX is linear. Prove that every
map
Vect(M ) → L(Vect(M ), Vect(M )) : X 7→ DX

that satisfies (5.2.9) is linear.

5.2. THE RIEMANN CURVATURE TENSOR 237

5.2.3 A Global Formula

Lemma 5.2.10. For X, Y, Z ∈ Vect(M ) we have

R(X, Y )Z = ∇X ∇Y Z − ∇Y ∇X Z + ∇[X,Y ] Z. (5.2.12)

Proof. Fix a point p ∈ M . Then the right hand side of equation (5.2.12) at
p remains unchanged if we multiply each of the vector fields X, Y, Z by a
smooth function f : M → [0, 1] that is equal to one near p. Choosing f with
compact support we may therefore assume that the vector fields X and Y
are complete. Let φs denote the flow of X and ψ t the flow of Y . Define the
map γ : R2 → M by

γ(s, t) := φs ◦ ψ t (p), s, t ∈ R.

Then
∂s γ = X(γ), ∂t γ = (φs∗ Y )(γ).

Hence, by Remark 5.2.5, we have

∇s (Z ◦ γ) = (∇X Z) (γ), ∇t (Z ◦ γ) = ∇φs∗ Y Z (γ).

This implies



∇s ∇t (Z ◦ γ) = ∇∂s γ ∇φs∗ Y Z (γ) + ∇∂s φs∗ Y Z (γ).

Since 
∂ 
φs Y = [X, Y ]
∂s s=0 ∗
and ∂s γ = X(γ) we obtain

∇s ∇t (Z ◦ γ)(0, 0) = ∇X ∇Y Z(p) + ∇[X,Y ] Z(p),

∇t ∇s (Z ◦ γ)(0, 0) = ∇Y ∇X Z(p).

Hence


Rp (X(p), Y (p))Z(p) = ∇s ∇t (Z ◦ γ) − ∇t ∇s (Z ◦ γ) (0, 0)
= ∇X ∇Y Z(p) − ∇Y ∇X Z(p) + ∇[X,Y ] Z(p).

This proves Lemma 5.2.10.

238 CHAPTER 5. CURVATURE

Remark 5.2.11. Equation (5.2.12) can be written succinctly as

[∇X , ∇Y ] + ∇[X,Y ] = R(X, Y ). (5.2.13)

This can be contrasted with the equation

[LX , LY ] + L[X,Y ] = 0 (5.2.14)

for the operator LX on the space of real valued functions on M .

Remark 5.2.12. Equation (5.2.12) can be used to define the Riemann

curvature tensor. To do this one must again prove that the right hand side
of equation (5.2.12) at p depends only on the values X(p), Y (p), Z(p) of the
vector fields X, Y, Z at the point p. For this it suffices to prove that the map

Vect(M ) × Vect(M ) × Vect(M ) → Vect(M ) : (X, Y, Z) 7→ R(X, Y )Z

is linear over the Ring F (M ) of smooth real valued functions on M , i.e.

R(f X, Y )Z = R(X, f Y )Z = R(X, Y )f Z = f R(X, Y )Z (5.2.15)

for X, Y, Z ∈ Vect(M ) and f ∈ F (M ). The formula (5.2.15) follows from

the equations (5.2.6), (5.2.7), (5.2.14), and [X, f Y ] = f [X, Y ] − (LX f )Y. It
follows from (5.2.15) that the right hand side of (5.2.12) at p depends only
on the vectors X(p), Y (p), Z(p). The proof requires two steps. One first
shows that if X vanishes near p, then the right hand side of (5.2.12) vanishes
at p (and similarly for Y and Z). Just multiply X by a smooth function
equal to zero at p and equal to one on the support of X; then f X = X and
hence the vector field R(X, Y )Z = R(f X, Y )Z = f R(X, Y )Z vanishes at p.
Second, we choose a local frame E1 , . . . , Em ∈ Vect(M ), i.e. vector fields
that form a basis of Tp M for each p in some open set U ⊂ M . Then we may
write
Xm Xm X m
X= ξ i Ei , Y = η j Ej , Z= ζ k Ek
i=1 j=1 k=1

in U . Using the first step and the F (M )-multilinearity we obtain

m
X
R(X, Y )Z = ξ i η j ζ k R(Ei , Ej )Ek
i,j,k=1

in U . If X 0 (p) = X(p), then ξ i (p) = ξ 0 i (p) so if X(p) = X 0 (p), Y (p) = Y 0 (p),

Z(p) = Z 0 (p), then (R(X, Y )Z)(p) = (R(X 0 , Y 0 )Z 0 )(p) as required.
5.2. THE RIEMANN CURVATURE TENSOR 239

5.2.4 Symmetries
Theorem 5.2.13. The Riemann curvature tensor satisfies
R(Y, X) = −R(X, Y ) = R(X, Y )∗ , (5.2.16)
R(X, Y )Z + R(Y, Z)X + R(Z, X)Y = 0, (5.2.17)
hR(X, Y )Z, W i = hR(Z, W )X, Y i, (5.2.18)
for X, Y, Z, W ∈ Vect(M ). Equation (5.2.17) is the first Bianchi identity.
Proof. The first equation in (5.2.16) is obvious from the definition and the
second follows from the Gauß–Codazzi formula (5.2.3). Alternatively, choose
a smooth map γ : R2 → M and two vector fields Z, W along γ. Then
0 = ∂s ∂t hZ, W i − ∂t ∂s hZ, W i
= ∂s h∇t Z, W i + ∂s hZ, ∇t W i − ∂t h∇s Z, W i − ∂t hZ, ∇s W i
= h∇s ∇t Z, W i + hZ, ∇s ∇t W i − h∇t ∇s Z, W i − hZ, ∇t ∇s W i
= hR(∂s γ, ∂t γ)Z, W i − hZ, R(∂s γ, ∂t γ)W i.
This proof has the advantage that it carries over to the intrinsic setting. We
prove the first Bianchi identity using (5.2.9) and (5.2.12):
R(X, Y )Z + R(Y, Z)X + R(Z, X)Y
= ∇X ∇Y Z − ∇Y ∇X Z + ∇[X,Y ] Z + ∇Y ∇Z X − ∇Z ∇Y X + ∇[Y,Z] X
+∇Z ∇X Y − ∇X ∇Z Y + ∇[Z,X] Y
= ∇[Y,Z] X − ∇X [Y, Z] + ∇[Z,X] Y − ∇Y [Z, X] + ∇[X,Y ] Z − ∇Z [X, Y ]
= [X, [Y, Z]] + [Y, [Z, X]] + [Z, [X, Y ]].
The last term vanishes by the Jacobi identity. We prove (5.2.18) by com-
bining the first Bianchi identity with (5.2.16):
hR(X, Y )Z, W i − hR(Z, W )X, Y i
= −hR(Y, Z)X, W i − hR(Z, X)Y, W i − hR(Z, W )X, Y i
= hR(Y, Z)W, Xi + hR(Z, X)W, Y i + hR(W, Z)X, Y i
= hR(Y, Z)W, Xi − hR(X, W )Z, Y i
= hR(Y, Z)W, Xi − hR(W, X)Y, Zi.
Note that the first line is related to the last by a cyclic permutation. Re-
peating this argument we find
hR(Y, Z)W, Xi − hR(W, X)Y, Zi = hR(Z, W )X, Y i − hR(X, Y )Z, W i.
Combining the last two identities we obtain (5.2.18). This proves Theo-
rem 5.2.13.
240 CHAPTER 5. CURVATURE

Remark 5.2.14. We may think of a vector field X on M as a section of

the tangent bundle. This is reflected in the alternative notation
Ω0 (M, T M ) := Vect(M ).
A 1-form on M with values in the tangent bundle is a collection of
linear maps A(p) : Tp M → Tp M , one for every p ∈ M , which is smooth
in the sense that for every smooth vector field X on M the assignment
p 7→ A(p)X(p) defines again a smooth vector field on M . We denote by
Ω1 (M, T M )
the space of smooth 1-forms on M with values in T M . The covariant deriva-
tive of a vector field Y is such a 1-form with values in the tangent bundle
which assigns to every p ∈ M the linear map Tp M → Tp M : v 7→ ∇v Y (p).
Thus we can think of the covariant derivative as a linear operator
∇ : Ω0 (M, T M ) → Ω1 (M, T M ).
The equation (5.2.6) asserts that the operators X 7→ ∇X indeed determine
a linear operator from Ω0 (M, T M ) to Ω1 (M, T M ). Equation (5.2.7) as-
serts that this linear operator ∇ is a connection on the tangent bundle
of M . Equation (5.2.8) asserts that ∇ is a Riemannian connection and
equation (5.2.9) asserts that ∇ is torsion-free. Thus Lemma 5.2.7 can
be restated as asserting that the Levi-Civita connection is the unique
torsion-free Riemannian connection on the tangent bundle.
Exercise 5.2.15. Extend the notion of a connection to a general vector bun-
dle E, both as a collection of linear operators ∇X : Ω0 (M, E) → Ω0 (M, E),
one for every vector field X ∈ Vect(M ), and as a linear operator
∇ : Ω0 (M, E) → Ω1 (M, E)
satisfying the analogue of equation (5.2.7). Interpret this equation as a Leib-
niz rule for the product of a function on M with a section of E. Show that
∇⊥ is a connection on T M ⊥ . Extend the notion of curvature to connections
on general vector bundles.
Exercise 5.2.16. Show that the field which assigns to each p ∈ M the
multi-linear map Rp⊥ : Tp M × Tp M → L(Tp M ⊥ , Tp M ⊥ ) characterized by
R⊥ (∂s γ, ∂t γ)Y = ∇⊥ ⊥ ⊥ ⊥
s ∇t Y − ∇ t ∇s Y

for γ : R2 → M and Y ∈ Vect⊥ (γ) satisfies the equation

Rp⊥ (u, v) = hp (u)hp (v)∗ − hp (v)hp (u)∗
for p ∈ M and u, v ∈ Tp M .
5.2. THE RIEMANN CURVATURE TENSOR 241

5.2.5 Riemannian Metrics on Lie Groups

We begin with a calculation of the Riemann curvature tensor on a Lie sub-
group of the orthogonal group O(n) with the Riemannian metric inherited
from the standard inner product

hv, wi := trace(v T w) (5.2.19)

on the ambient space gl(n, R) = Rn×n . This fits into the extrinsic setting
used throughout most of this book. By Theorem 2.5.26 every Lie subgroup
of O(n) is a closed subset of O(n) and hence is compact.
Example 5.2.17. Let G ⊂ O(n) be a Lie subgroup and let

g := Lie(G) = T1l G

be the Lie algebra of G. Consider the Riemannian metric on G induced by

the inner product (5.2.19) on Rn×n . Then the Riemann curvature tensor
on G can be expressed in terms of the Lie bracket (see item (d) below).
(a) The maps g 7→ ag, g 7→ ga, g 7→ g −1 are isometries of G for every a ∈ G.
(b) Let γ : R → G be a smooth curve and X ∈ Vect(γ) be a smooth vector
field along γ. Then the covariant derivative of X is given by
d 1
γ(t)−1 ∇X(t) = γ(t)−1 X(t) + γ(t)−1 γ̇(t), γ(t)−1 X(t) .

(5.2.20)
dt 2
(Exercise: Prove equation (5.2.20). Hint: Since g ⊂ o(n) we have the
identity trace((ξη + ηξ)ζ) = 0 for all ξ, η, ζ ∈ g.)
(c) A smooth map γ : R → G is a geodesic if and only if there exist matrices
g ∈ G and ξ ∈ g such that

γ(t) = g exp(tξ). (5.2.21)

For G = O(n) we have seen this in Example 4.3.12 and in the general
case this follows from equation (5.2.20) with X = γ̇. Hence the exponential
map exp : g → G defined by the exponential matrix (as in §2.5) agrees with
the time-1-map of the geodesic flow (as in §4.3).
(d) The Riemann curvature tensor on G is given by
1
g −1 Rg (u, v)w = − [[g −1 u, g −1 v], g −1 w] (5.2.22)
4
for g ∈ G and u, v, w ∈ Tg G. Note that the first Bianchi identity is equivalent
to the Jacobi identity. (Exercise: Prove equation (5.2.22).)
242 CHAPTER 5. CURVATURE

Definition 5.2.18 (Bi-invariant metric). Let G ⊂ GL(n, R) be a Lie

group and let g = Lie(G) = T1l G be its Lie algebra. A Riemannian metric
on G is called bi-invariant iff it has the form
hv, wig := hvg −1 , wg −1 i (5.2.23)
for g ∈ G and v, w ∈ Tg G, where h·, ·i is an inner product on the Lie algebra g
that is invariant under conjugation, i.e. it satisfies the equation
hξ, ηi = hgξg −1 , gηg −1 i. (5.2.24)
for all ξ, η ∈ g and all g ∈ G.
Exercise 5.2.19. Prove that the Riemannian metric induced by (5.2.19)
on any Lie subgroup G ⊂ O(n) is bi-invariant.
Exercise 5.2.20. Use the Haar measure ([49, Chapter 8]) to prove that
every compact Lie group admits a bi-invariant Riemannian metric.
Exercise 5.2.21. Prove that all the assertions in Example 5.2.17 carry over
verbatim to any Lie group equipped with a bi-invariant Riemann metric.
Exercise 5.2.22. Prove that, if an inner product on the Lie algebra g of a
Lie group G is invariant under conjugation, then it satisfies the equation
h[ξ, η], ζi = hξ, [η, ζ]i (5.2.25)
for all ξ, η, ζ ∈ g. If G is connected, prove that, conversely, equation (5.2.25)
implies (5.2.24). An inner product on an arbitrary Lie algebra g is called
invariant iff it satisfies equation (5.2.25).
Exercise 5.2.23. Let g be a finite-dimensional Lie algebra and denote the
linear subspace spanned by all vectors of the form [ξ, η] by
 
[g, g] := span [ξ, η]  ξ, η ∈ g . (5.2.26)
If g is equipped with an invariant inner product, prove that [g, g]⊥ = Z(g)
is the center of g (Example 2.5.32) and hence g = [g, g] ⊕ Z(g). Prove that
the Heisenberg algebra h in Exercise 2.5.15 satisfies [h, h] = Z(h) and hence
does not admit an invariant inner product.
Example 5.2.24. Every finite-dimensional Lie algebra g admits a natural
symmetric bilinear form κ : g × g → R that satisfies equation (5.2.25). It is
called the Killing form and is defined by


κ(ξ, η) := trace ad(ξ)ad(η) , ξ, η ∈ g, (5.2.27)
where ad : g → Der(g) is the adjoint representation in Example 2.5.22. The
Killing form may have a kernel (which always contains the center of g), and
even if it is nondegenerate, it may be indefinite.
5.2. THE RIEMANN CURVATURE TENSOR 243

Exercise 5.2.25. Prove that κ([ξ, η], ζ) = κ(ξ, [η, ζ]) for all ξ, η, ζ ∈ g.
Exercise 5.2.26. Assume that g admits an invariant inner product. For
each ξ ∈ g prove that the derivation ad(ξ) is skew-adjoint with respect to this
inner product and deduce that κ(ξ, ξ) = −trace(ad(ξ)∗ ad(ξ)) = −|ad(ξ)|2 .
Deduce that the Killing form is nondegenerate whenever g has a trivial
center and admits an invariant inner product.
Example 5.2.27. Let G ⊂ GL(n, R) be any Lie subgroup (not necessarily
contained in O(n)), and let g := Lie(G) = T1l G be its Lie algebra. Fix
any inner product on the Lie algebra g (not necessarily invariant under
conjugation) and consider the Riemannian metric on G defined by
hv, wig := hvg −1 , wg −1 i (5.2.28)
for v, w ∈ Tg G. This metric is called right invariant.
(a) The map g 7→ ga is an isometry of G for every a ∈ G.
(b) Define the linear map A : g → End(g) by
1 
hA(ξ)η, ζi = hξ, [η, ζ]i − hη, [ζ, ξ]i − hζ, [ξ, η]i (5.2.29)
2
for ξ, η, ζ ∈ g. Then A is the unique linear map that satisfies
A(ξ) + A(ξ)∗ = 0, A(η)ξ − A(ξ)η = [ξ, η]
for all ξ, η ∈ g, where A(ξ)∗ is the adjoint operator with respect to the inner
product on g. Let γ : R → G be a smooth curve and X ∈ Vect(γ) be a
smooth vector field along γ. Then the covariant derivative of X is given by
 
d −1 −1 −1
∇X = (Xγ ) + A(γ̇γ )Xγ γ. (5.2.30)
dt
(Exercise: Prove this. Moreover, if the inner produt on g is invariant, prove
that A(ξ)η = − 21 [ξ, η] for all ξ, η ∈ g.)
(c) A smooth curve γ : R → G is a geodesic if and only if it satisfies
d
(γ̇γ −1 ) + A(γ̇γ −1 )γ̇γ −1 = 0. (5.2.31)
dt
(Exercise: G is complete.)
(d) The Riemann curvature tensor on G is given by
 
Rg (ξg, ηg)ζg g −1 = A ([ξ, η]) + [A (ξ) , A (η)] ζ


(5.2.32)

for g ∈ G and ξ, η, ζ ∈ g. (Exercise: Prove this.)

244 CHAPTER 5. CURVATURE

5.3 Generalized Theorema Egregium

We will now show that geodesics, covariant differentiation, parallel trans-
port, and the Riemann curvature tensor are all intrinsic, i.e. they are in-
tertwined by isometries. In the extrinsic setting these results are somewhat
surprising since these objects are all defined using the second fundamental
form, whereas isometries need not preserve the second fundamental form in
any sense but only the first fundamental form.
Below we shall give a formula expressing the Gaußian curvature of a
surface M 2 in R3 in terms of the Riemann curvature tensor and the first
fundamental form. It follows that the Gaußian curvature is also intrinsic.
This fact was called by Gauß the “Theorema Egregium” which explains the
title of this section.

5.3.1 Pushforward
0
We assume throughout this section that M ⊂ Rn and M 0 ⊂ Rn are smooth
submanifolds of the same dimension m. As in §5.1 we denote objects on M 0
by the same letters as objects in M with primes affixed. In particular, g 0
denotes the first fundamental form on M 0 and R0 denotes the Riemann
curvature tensor on M 0 .
Let φ : M → M 0 be a diffeomorphism. Using φ we can move objects
on M to M 0 . For example the pushforward of a smooth curve γ : I → M is
the curve
φ∗ γ := φ ◦ γ : I → M 0 ,
the pushforward of a smooth function f : M → R is the function

φ∗ f := f ◦ φ−1 : M 0 → R,

the pushforward of a vector field X ∈ Vect(γ) along a curve γ : I → M is

the vector field φ∗ X ∈ Vect(φ∗ γ) defined by

(φ∗ X)(t) := dφ(γ(t))X(t)

for t ∈ I, and the pushforward of a global vector field X ∈ V ect(M ) is the

vector field φ∗ X ∈ Vect(M 0 ) defined by

(φ∗ X)(φ(p)) := dφ(p)X(p)

for p ∈ M . Recall that the first fundamental form on M is the Riemannian

metric g defined as the restriction of the Euclidean inner product on the
5.3. GENERALIZED THEOREMA EGREGIUM 245

ambient space to each tangent space of M . It assigns to each p ∈ M the

bilinear map gp ∈ Tp M × Tp M → R given by

gp (u, v) = hu, vi, u, v ∈ Tp M.

Its pushforward is the Riemannian metric which assigns to each p0 ∈ M 0 the

inner product (φ∗ g)p0 : Tp0 M 0 × Tp0 M 0 → R defined by

(φ∗ g)φ(p) (dφ(p)u, dφ(p)v) := gp (u, v)

for p := φ−1 (p0 ) ∈ M and u, v ∈ Tp M . The pushforward of the Riemann

curvature tensor is the tensor which assigns to
each p0 ∈ M 0 the bilinear
map (φ∗ R)p0 : Tp0 M 0 × Tp0 M 0 → L Tp0 M 0 , Tp0 M 0 , defined by

(φ∗ R)φ(p) (dφ(p)u, dφ(p)v) := dφ(p)Rp (u, v) dφ(p)−1

for p := φ−1 (p0 ) ∈ M and u, v ∈ Tp M .

5.3.2 Theorema Egregium

Theorem 5.3.1 (Theorema Egregium). The first fundamental form, co-
variant differentiation, geodesics, parallel transport, and the Riemann cur-
vature tensor are intrinsic. This means that for every isometry φ : M → M 0
the following holds.
(i) φ∗ g = g 0 .
(ii) If X ∈ Vect(γ) is a vector field along a smooth curve γ : I → M , then

∇0 (φ∗ X) = φ∗ ∇X, (5.3.1)

and if X, Y ∈ Vect(M ) are global vector fields, then

∇0φ∗ X φ∗ Y = φ∗ (∇X Y ). (5.3.2)

(iii) If γ : I → M is a geodesic, then φ ◦ γ : I → M 0 is a geodesic.

(iv) If γ : I → M is a smooth curve, then for all s, t ∈ I, we have

Φ0φ◦γ (t, s)dφ(γ(s)) = dφ(γ(t))Φγ (t, s). (5.3.3)

(v) φ∗ R = R0 .
246 CHAPTER 5. CURVATURE

Proof. Assertion (i) is simply a restatement of Theorem 5.1.1. To prove (ii)

we choose a local smooth parametrization ψ : Ω → U of an open set U ⊂ M ,
defined on an open set Ω ⊂ Rm , so that ψ −1 : U → Ω is a coordinate chart.
Suppose without loss of generality that γ(t) ∈ U for all t ∈ I and define
c : I → Ω and ξ : I → Rm by
m
X ∂ψ
γ(t) = ψ(c(t)), X(t) = ξ i (t) (c(t)).
∂xi
i=1

Recall from equations (3.6.6) and (3.6.7) that

 
m m
X
ξ˙k (t) +
X ∂ψ
∇X(t) = Γkij (c(t))ċi (t)ξ j (t) k (c(t)),
∂x
k=1 i,j=1

where the Christoffel symbols Γkij : Ω → R are defined by

m
∂2ψ X ∂ψ
Π(ψ) i j
= Γkij k .
∂x ∂x ∂x
k=1

Now consider the same formula for φ∗ X using the parametrization

ψ 0 := φ ◦ ψ : Ω → U 0 := φ(U ) ⊂ M 0 .
The Christoffel symbols Γ0 kij : Ω → R associated to this parametrization of
U 0 are defined by the same formula as the Γkij with ψ replaced by ψ 0 . But
the metric tensor for ψ agrees with the metric tensor for ψ 0 :
  0
∂ψ ∂ψ 0
 
∂ψ ∂ψ
gij = , = , .
∂xi ∂xj ∂xi ∂xj

Hence it follows from Lemma 3.6.5 that Γ0 kij = Γkij for all i, j, k. This implies
that the covariant derivative of φ∗ X is given by
 
m m
X X ∂ψ 0
∇0 (φ∗ X) = ξ˙k + Γkij (c)ċi ξ j  k (c)
∂x
k=1 i,j=1
 
m m
X
ξ˙k +
X ∂ψ
= dφ(ψ(c)) Γkij (c)ċi ξ j  k (c)
∂x
k=1 i,j=1
= φ∗ ∇X.
This proves (5.3.1). Equation (5.3.2) follows immediately from (5.3.1) and
Remark 5.2.4.
5.3. GENERALIZED THEOREMA EGREGIUM 247

Here is a second proof of (ii). For every vector field X ∈ Vect(M ) we

define the operator DX : Vect(M ) → Vect(M ) by

DX Y := φ∗ (∇φ∗ X φ∗ Y ) .

Then, for all X, Y ∈ Vect(M ), we have

DY X − DX Y = φ∗ (∇φ∗ Y φ∗ X − ∇φ∗ X φ∗ Y ) = φ∗ [φ∗ X, φ∗ Y ] = [X, Y ].

Moreover, it follows from (i) that

φ∗ LX hY, Zi = Lφ∗ X hφ∗ Y, φ∗ Zi

= h∇φ∗ X φ∗ Y, φ∗ Zi + hφ∗ Y, ∇φ∗ X φ∗ Zi
= hφ∗ DX Y, φ∗ Zi + hφ∗ Y, φ∗ DX Zi


= φ∗ hDX Y, Zi + hY, DX Zi .

and hence LX hY, Zi = hDX Y, Zi + hY, DX Zi for all X, Y, Z ∈ Vect(M ).

Thus the operator X 7→ DX satisfies equations (5.2.8) and (5.2.9) and, by
Lemma 5.2.7, it follows that DX Y = ∇X Y for all X, Y ∈ Vect(M ). This
completes the second proof of (ii).
We prove (iii). Since φ preserves the first fundamental form it also
preserves the energy of curves, namely

E(φ ◦ γ) = E(γ)

for every smooth map γ : [0, 1] → M . Hence γ is a critical point of the energy
functional if and only if φ ◦ γ is a critical point of the energy functional.
Alternatively it follows from (ii) that
 
0 d
∇ φ ◦ γ = ∇0 φ∗ γ̇ = φ∗ ∇γ̇
dt

for every smooth curve γ : I → M . If γ is a geodesic, the last term vanishes

and hence φ◦γ is a geodesic as well. As a third proof we can deduce (iii) from
the formula φ(expp (v)) = expφ(p) (dφ(p)v) in the proof of Theorem 5.1.1.
We prove (iv). For t0 ∈ I and v0 ∈ Tγ(t0 ) M define

X(t) := Φγ (t, t0 )v0 , X 0 (t) := Φ0φ◦γ (t, t0 )dφ(γ(t0 ))v0 .

By (ii) the vector fields X 0 and φ∗ X along φ ◦ γ are both parallel and they
agree at t = t0 . Hence X 0 (t) = φ∗ X(t) for all t ∈ I and this proves (5.3.3).
248 CHAPTER 5. CURVATURE

We prove (v). Fix a smooth map γ : R2 → M and a smooth vector

field Z along γ, and define

γ 0 = φ ◦ γ : R2 → M 0 , Z 0 := φ∗ Z ∈ Vect(γ 0 ).

Then it follows from (ii) that

R0 (∂s γ 0 , ∂t γ 0 )Z 0 = ∇0s ∇0t Z 0 − ∇0t ∇0s Z 0

= φ∗ (∇s ∇t Z − ∇t ∇s Z)
= dφ(γ)R(∂s γ, ∂t γ)Z
= (φ∗ R)(∂s γ 0 , ∂t γ 0 )Z 0 .

This proves (v) and Theorem 5.3.1.

The assertions of Theorem 5.3.1 carry over in modified form to local

isometries φ : M → M 0 . In particular, the pushforward of a vector field
on M under φ is only defined when φ is a diffeomorphism while the pushfor-
ward of a vector field along a curve is defined for any smooth map φ. Also,
the pushforward of the Riemann curvature tensor under a local isometry is
only defined locally, and local isometries satisfy the local analogue of the
first assertion in Theorema Egregium by definition.

Corollary 5.3.2 (Theorema Egregium for Local Isometries). Every

local isometry φ : M → M 0 has the following properties.
(i) Every vector field X along a smooth curve γ : I → M satisfies (5.3.1).
(ii) If γ : I → M is a geodesic, then so is φ ◦ γ : I → M 0 .
(iii) Parallel transport along a smooth curve γ : I → M satisfies (5.3.3).
(iv) The curvature tensors R of M and R0 of M 0 are related by the formula
0
Rφ(p) (dφ(p)u, dφ(p)v) = dφ(p)Rp (u, v)dφ(p)−1 (5.3.4)

for all p ∈ M and all u, v ∈ Tp M .

Proof. Let p0 ∈ M . Then, by the Inverse Function Theorem 2.2.17, there

exists an open neighborhood U ⊂ M of p0 such that the restriction of φ
to U is a diffeomorphism onto the open set U 0 := φ(U ) ⊂ M 0 and so is an
isometry. Hence by Theorem 5.3.1 the assertions (i) and (ii) hold for the the
restriction of γ to I0 := γ −1 (U ) (a union of subintervals of I) and (iv) holds
for all p ∈ U . Since these are local statements and p0 was chosen arbitrary,
this proves (i), (ii), and (iv). Part (iii) follows directly from (i) as in the
proof of of Theorem 5.3.1 and this proves Corollary 5.3.2.
5.3. GENERALIZED THEOREMA EGREGIUM 249

The following corollaries derive two important consequences of Corol-

lary 5.3.2. For sufficiently small tangent vectors equation (5.3.5) below al-
ready appeared in the proof of Theorem 5.1.1 and was used in Lemma 5.1.10
and Exercise 5.1.11. When M is not complete, recall the notation Vp ⊂ Tp M
for the domain of the exponential map of M at a point p (Definition 4.3.5).
For p0 ∈ M 0 denote the domain of the exponential map by Vp00 ⊂ Tp0 M 0 .
Corollary 5.3.3. Let φ : M → M 0 be a local isometry and let p ∈ M .
0
Then dφ(p)Vp ⊂ Vφ(p) and, for every v ∈ Vp ,

φ(expp (v)) = exp0φ(p) (dφ(p)v). (5.3.5)

Proof. Let v ∈ Vp ⊂ Tp M and define
γ(t) := expp (tv)
for 0 ≤ t ≤ 1. Then γ : [0, 1] → M is a geodesic by Lemma 4.3.6, and hence
so is the curve γ 0 := φ ◦ γ : [0, 1] → M 0 by Corollary 5.3.2. Moreover,
γ 0 (0) = φ(γ(0)) = φ(p), γ̇ 0 (0) = dφ(γ(0))γ̇(0) = dφ(p)v
by the chain rule. Hence it follows from the definition of the exponential
0
map (Definition 4.3.5) that dφ(p)v ∈ Vφ(p) and

exp0φ(p) (dφ(p)v) = γ 0 (1) = φ(γ(1)) = φ(expp (v)).

This proves Corollary 5.3.3.

The next corollary establishes the basic relation between local isome-
tries and developments and lies at the heart of the existence criterion for
isometries in the Cartan–Ambrose–Hicks Theorem 6.1.7 in §6.1 below.
Corollary 5.3.4. Let φ : M → M 0 be a local isometry and let γ : I → M be
a smooth curve on an interval I. Fix an element t0 ∈ I and define
p0 := γ(t0 ), q0 := φ(p0 ), Φ0 := dφ(p0 ). (5.3.6)
Then there exists a unique development (Φ, γ, γ 0 ) of M along M 0 on the
entire interval I satisfying the initial conditions
γ 0 (t0 ) = q0 , Φ(t0 ) = Φ0 . (5.3.7)
This development is given by
γ 0 (t) = φ(γ(t)), Φ(t) = dφ(γ(t)) (5.3.8)
for t ∈ I.
250 CHAPTER 5. CURVATURE

Proof. Define
γ 0 (t) := φ(γ(t)), Φ(t) := dφ(γ(t))
for t ∈ I. Then
γ̇ 0 (t) = Φ(t)γ̇(t)
for all t ∈ I by the chain rule, and every vector field X along γ satisfies

Φ∇X = ∇0 (ΦX)

by Corollary 5.3.2. Hence (Φ, γ, γ 0 ) is a development by Lemma 3.5.19.

By (5.3.6) this development satisfies the initial condition (5.3.7). Hence
the assertion follows from the uniqueness result for developments in Theo-
rem 3.5.21. This proves Corollary 5.3.4.

5.3.3 Gaußian Curvature

As a special case we shall now consider a hypersurface

M ⊂ Rm+1 ,

i.e. a smooth submanifold of codimension one. We assume that there exists

a smooth map
ν : M → Rm+1
such that, for every p ∈ M , we have

ν(p) ⊥ Tp M, |ν(p)| = 1.

Such a map always exists locally (see Example 3.1.3). Note that ν(p) is
an element of the unit sphere in Rm+1 for every p ∈ M and hence we can
regard ν as a map from M to S m :

ν : M → Sm.

Such a map is called a Gauß map for M . Note that if ν : M → S 2 is a

Gauß map, so is −ν, but this is the only ambiguity when M is connected.
Differentiating ν at p ∈ M we obtain a linear map

dν(p) : Tp M → Tν(p) S m = Tp M

Here we use the fact that Tν(p) S m = ν(p)⊥ and, by definition of the Gauß
map ν, the tangent space of M at p is also equal to ν(p)⊥ . Thus dν(p) is a
linear map from the tangent space of M at p to itself.
5.3. GENERALIZED THEOREMA EGREGIUM 251

Definition 5.3.5. The Gaußian curvature of the hypersurface M is the

real valued function
K:M →R
defined by

K(p) := det dν(p) : Tp M → Tp M
for p ∈ M . (Replacing ν by −ν has the effect of replacing K by (−1)m K;
so K is independent of the choice of the Gauß map when m is even.)
Remark 5.3.6. Given a subset B ⊂ M the set ν(B) ⊂ S m is often called
the spherical image of B. If ν is a diffeomorphism on a neighborhood of
B, the change of variables formula for an integral gives
Z Z
µS = |K|µM .
ν(B) B

Here µM and µS denote the volume elements

R on M and S m , respectively.
Introducing the notation AreaM (B) := B µM we obtain the formula

AreaS (ν(B))
|K(p)| = lim .
B→p AreaM (B)

This says that the curvature at p is roughly the ratio of the (m-dimensional)
area of the spherical image ν(B) to the area of B where B is a very small open
neighborhood of p in M . The sign of K(p) is positive when the linear map
dν(p) : Tp M → Tp M preserves orientation and negative when it reverses
orientation.
Remark 5.3.7. We see that the Gaußian curvature is a natural general-
ization of Euler’s curvature for a plane curve. Indeed if M ⊂ R2 is a
1-manifold and p ∈ M , we can choose a curve γ = (x, y) : (−ε, ε) → M such
that γ(0) = p and |γ̇(s)| = 1 for every s. This curve parametrizes M by the
arclength and the unit normal vector pointing to the right with respect to
the orientation of γ is ν(x, y) = (ẏ, −ẋ). This is a local Gauß map and its
derivative (ÿ, −ẍ) is tangent to the curve. The inner product of the latter
with the unit tangent vector γ̇ = (ẋ, ẏ) is the Gaußian curvature. Thus

dx d2 y dy d2 x dθ
K := 2
− 2
=
ds ds ds ds ds
where s is the arclength parameter and θ is the angle made by the normal
(or the tangent) with some constant line. With this convention K is positive
at a left turn and negative at a right turn.
252 CHAPTER 5. CURVATURE

Exercise 5.3.8. The Gaußian curvature of a sphere of radius r is constant

and has the value r−m .
Exercise 5.3.9. Show that the Gaußian curvature of the surface z = x2 −y 2
is −4 at the origin.

K>0 K=0 K<0

Figure 5.1: Positive and negative Gaußian curvature.

We now restrict to the case of surfaces, i.e. of 2-dimensional submani-

folds of R3 . Figure 5.1 illustrates the difference between positive and nega-
tive Gaußian curvature in dimension two.
Theorem 5.3.10 (Gaußian curvature). Let M ⊂ R3 be a surface and fix
a point p ∈ M . If u, v ∈ Tp M is a basis, then
hR(u, v)v, ui
K(p) = . (5.3.9)
|u|2 |v|2 − hu, vi2
Moreover,
R(u, v)w = −K(p)hν(p), u × viν(p) × w (5.3.10)
for all u, v, w ∈ Tp M .
Proof. The orthogonal projection of R3 onto the tangent space
Tp M = ν(p)⊥
is given by the 3 × 3-matrix
Π(p) = 1l − ν(p)ν(p)T .
Hence
dΠ(p)u = −ν(p)(dν(p)u)T − dν(p)u ν(p)T .

Here the first summand is the second fundamental form, which maps Tp M
to Tp M ⊥ , and the second summand is its dual, which maps Tp M ⊥ to Tp M .
Thus

T
hp (v) = ν(p) dν(p)v : Tp M → Tp M ⊥ ,
hp (u)∗ = dν(p)u ν(p)T : Tp M ⊥ → Tp M.


5.3. GENERALIZED THEOREMA EGREGIUM 253

By the Gauß–Codazzi formula this implies

Rp (u, v)w = hp (u)∗ hp (v)w − hp (v)∗ hp (u)w


T

T
= dν(p)u dν(p)v w − dν(p)v dν(p)u w
= hdν(p)v, widν(p)u − hdν(p)u, widν(p)v
and hence
hRp (u, v)w, zi = hdν(p)u, zihdν(p)v, wi − hdν(p)u, wihdν(p)v, zi. (5.3.11)
Now fix four tangent vectors u, v, w, z ∈ Tp M and consider the composition
A B C
R3 −→ R3 −→ R3 −→ R3
of the linear maps
Aξ := ξ 1 ν(p) + ξ 2 u + ξ 3 v,

dν(p)η, if η ⊥ ν(p),
Bη :=
η, if η ∈ Rν(p),
 
hζ, ν(p)i
Cζ :=  hζ, zi  .
hζ, wi
This composition is represented by the matrix
 
1 0 0
CBA =  0 hdν(p)u, zi hdν(p)v, zi  .
0 hdν(p)u, wi hdν(p)v, wi
Hence, by (5.3.11), we have
hRp (u, v)w, zi = det(CBA)
= det(A) det(B) det(C)
= hν(p), u × viK(p)hν(p), z × wi
= −K(p)hν(p), u × vihν(p) × w, zi.
This implies (5.3.10) and
hRp (u, v)v, ui = K(p)hν(p), u × vi2
= K(p) |u × v|2
 
= K(p) |u|2 |v|2 − hu, vi2 .

This proves Theorem 5.3.10.

254 CHAPTER 5. CURVATURE

Remark 5.3.11. Equation (5.3.9) implies

 
hRp (u, v)w, zi = K(p) hu, zihv, wi − hu, wihv, zi (5.3.12)

for all p ∈ M and all u, v, w, z ∈ Tp M . This is proved in Theorem 6.4.8

below. Exercise: Deduce this formula from (5.3.10).

Corollary 5.3.12 (Theorema Egregium of Gauß). The Gaußian cur-

vature is intrinsic, i.e. if
φ : M → M0

is an isometry of surfaces in R3 , then

K = K 0 ◦ φ : M → R.

Proof. Theorem 5.3.1 and Theorem 5.3.10.

Exercise 5.3.13. For m = 1 the Gaußian curvature is clearly not intrinsic

as any two curves are locally isometric (parameterized by arclength). Show
that the curvature K(p) is intrinsic for even m while its absolute value |K(p)|
is intrinsic for odd m ≥ 3. Hint: We still have the equation (5.3.11) which,
for z = u and v = w, can be written in the form
 
hdν(p)u, ui hdν(p)u, vi
hRp (u, v)v, ui = det .
hdν(p)v, ui hdν(p)v, vi

Thus, for every orthonormal basis v1 , . . . , vm of Tp M , the 2 × 2 minors of

the matrix
(hdν(p)vi , vj i)i,j=1,...,m

are intrinsic. Hence everything reduces to the following assertion.

Lemma. The determinant of an m × m matrix is an expression in its 2 × 2
minors if m is even; the absolute value of the determinant is an expression
in the 2 × 2 minors if m is odd and greater than or equal to 3.
The lemma is proved by induction on m. For the absolute value, note the
formula

det(A)m = det(det(A)1lm ) = det(AB) = det(A) det(B)

for an m × m-matrix A where B is the transposed matrix of cofactors.

5.4. THE GROUP OF ISOMETRIES* 255

5.4 The Group of Isometries*

The goal of this section is to prove that the group I(M ) of isometries of
a connected smooth manifold M admits a natural structure of a finite-
dimensional Lie group. This theorem was proved by Myers and Steenrod [39]
in 1939. The proof given below has certain analogies with the proof of the
Closed Subgroup Theorem 2.5.26 which will be indicated in the text. The
results of this section are not used elsewhere in the book and it can be
omitted at first reading.

5.4.1 The Myers–Steenrod Theorem

Assume throughout that M ⊂ Rn is a nonempty connected smooth m-
manifold. To state the main result, it is convenient to introduce the space
  
 p, q ∈ M and Φ : Tp M → Tq M
G := (q, Φ, p) 
 . (5.4.1)
is an orthogonal isomorphism
This space is a groupoid, i.e. a category in which every morphism is an
isomorphism. The space of objects is the manifold M , the morphisms
from p ∈ M to q ∈ M are the triples of the form (q, Φ, p) ∈ G, the identity
morphism from p to itself is the triple (p, 1l, p), the inverse of (q, Φ, p) ∈ G is
the triple (p, Φ−1 , q), and the composition of a morphism (q, Φ, p) ∈ G from p
to q with a morphism (r, Ψ, q) ∈ G from q to r is the triple (r, ΨΦ, p).
The space G is a smooth manifold (in the intrinsic sense). To see this,
consider the diagonal action of the orthogonal group O(m) on the prod-
uct of the orthonormal frame bundle O(M ) with itself (Definition 3.4.3).
This action is free and the map π : O(M ) × O(M ) → G which sends the
pair ((q, e0 ), (p, e)) to the triple (q, e0 ◦ e−1 , p) ∈ G descends to a bijection
from the quotient (O(M ) × O(M ))/O(m) to G. By Theorem 2.9.14 there is
a unique smooth structure on G such that the map π : O(M ) × O(M ) → G
is a submersion. With this structure the maps s, t : G → M defined by
s(q, Φ, p) := p, t(q, Φ, p) := q
are smooth, the map e : M → G defined by e(p) := (p, 1l, p) is an embedding,
the inverse map i : G → G defined by i(q, Φ, p) := (p, Φ−1 , q) is a smooth
involution, the map s × t : G × G → M × M is a submersion, and the
composition map m : (s × t)−1 (∆) → G defined by
m((r, Ψ, q), (q, Φ, p)) := (r, ΨΦ, p)
is smooth. (Here ∆ := {(q, q) | q ∈ M } is the diagonal in M × M .) These
properties assert that the tuple (G, s, t, e, i, m) is a smooth groupoid.
256 CHAPTER 5. CURVATURE

For each p ∈ M define

  
 q ∈ M and Φ : Tp M → Tq M
Gp := (q, Φ) 
 . (5.4.2)
is an orthogonal isomorphism
This space is a submanifold of G because s : G → M is a submersion. The
space Gp is also a smooth submanifold of Rn × L(Tp M, Rn ) and is dif-
feomorphic to the orthonormal frame bundle O(M ) ⊂ Rn × Rn×m via the
map Gp → O(M ) : (q, Φ) 7→ (q, Φ ◦ e) for any orthonormal frame e of Tp M .
Since M is connected, Lemma 5.1.10 asserts that the map
ιp : I(M ) → Gp , ιp (φ) := (φ(p), dφ(p)), (5.4.3)
is injective for every p ∈ M . Denote the image of this map by
 
Ip := ιp (I(M )) = (φ(p), dφ(p))  φ ∈ I(M ) ⊂ Gp . (5.4.4)
In the following theorem we do not assume that M is complete. For a space
of smooth functions or maps on M we use the term C ∞ topology to mean
the topology of uniform convergence with all derivatives on each compact
subset of M . Likewise we use the term C 0 topology to mean the topology of
uniform convergence on each compact subset of M . The latter is also called
compact-open topology (because a basis of the topology consists of sets
of maps, one for each compact subset of the source and each open subset of
the target, that send the given compact subset of the source into the given
open subset of the target).
Theorem 5.4.1 (Myers–Steenrod). Let M ⊂ Rn be a nonempty con-
nected smooth m-manifold. Then the following holds.
(i) There exists a unique smooth structure on the isometry group I(M ) such
that the map ιp : I(M ) → Gp in (5.4.3) is an embedding for every p ∈ M .
The topology induced by this smooth structure agrees with the C 0 topology
and with the C ∞ topology on I(M ) and dim(I(M )) ≤ m(m + 1)/2.
(ii) With the smooth structure in part (i) the maps
I(M ) × I(M ) → I(M ) : (ψ, φ) 7→ ψ ◦ φ
and
I(M ) → I(M ) : φ 7→ φ−1
are smooth. Thus I(M ) is a finite-dimensional Lie group.
(iii) The Lie algebra of I(M ) is the space VectK,c (M ) of complete Killing
vector fields (defined below).
Proof. See page 271.
5.4. THE GROUP OF ISOMETRIES* 257

5.4.2 The Topology on the Space of Isometries

The next lemma shows that for each p ∈ M the set Ip in (5.4.4) is a closed
subset of Gp and that the map ιp : I(M ) → Ip in (5.4.3) is a homeomorphism
with respect to the C ∞ topology on I(M ).

Lemma 5.4.2. Fix two elements p0 , q0 ∈ M , let Φ0 : Tp0 M → Tq0 M be an

orthogonal isomorphism, and let φi : M → M be a sequence of isometries
Then the following are equivalent.
(i) The sequence ιp0 (φi ) ∈ Ip0 converges to the pair (q0 , Φ0 ) ∈ Gp0 , i.e.

lim φi (p0 ) = q0 , lim dφi (p0 ) = Φ0 . (5.4.5)

i→∞ i→∞

(ii) There exists an isometry φ : M → M such that φ(p0 ) = q0 , dφ(p0 ) = Φ0 ,

and φi converges to φ in the C ∞ topology.

Proof. That (ii) implies (i) follows directly from the definitions. We prove in
three steps that (i) implies (ii). For p ∈ M and r > 0 denote by Vp ⊂ Tp M
the domain of the exponential map of M at p (Definition 4.3.5), and de-
fine Ur (p) := {q ∈ M | d(p, q) < r} and Br (p) := {v ∈ Tp M | |v| < r}.
Step 1. Assume (5.4.5) holds and choose a real number 0 < r ≤ inj(p0 ; M ).
Then r ≤ inj(q0 ; M ) and φi converges on the open set Ur (p0 ) in the C ∞
topology to the isometry φ0 := expq0 ◦ Φ0 ◦ exp−1
p0 : Ur (p0 ) → Ur (q0 ).

Since r ≤ inj(p0 ; M ), Corollary 5.3.3 asserts that

Br (φi (p0 )) = dφi (p0 )Br (p0 ) ⊂ dφi (p0 )Vp0 ⊂ Vφi (p0 )

and φi ◦ expp0 = expφi (p0 ) ◦ Φi : Br (p0 ) → Ur (φi (p0 )). Thus the map

expφi (p0 ) = φi ◦ expp0 ◦ Φ−1

i : Br (φi (p0 )) → Ur (φi (p0 ))

is a diffeomorphism and so r ≤ inj(φi (p0 ); M ) for each i ∈ N. Since φi (p0 )

converges to q0 , this implies Br (q0 ) ⊂ Vq0 (Exercise 4.3.14). Hence the se-
quence of maps expφi (p0 ) ◦ Φi converges to the map expq0 ◦ Φ0 in the C ∞
topology on Br (p0 ). Hence the sequence of isometries

φi = expφi (p0 ) ◦ Φi ◦ exp−1

p0 : Ur (p0 ) → M

converges in the C ∞ topology to the smooth map

φ0 := expq0 ◦ Φ0 ◦ exp−1
p0 : Ur (p0 ) → expq0 (Br (q0 )).
258 CHAPTER 5. CURVATURE

The map φ0 satisfies d(φ0 (p), φ0 (q)) = limi→∞ d(φi (p), φi (q)) = d(p, q) for
all p, q ∈ Ur (p0 ) and so by Theorem 5.1.1 is a difffeomorphism from Ur (p0 )
onto the open set expq0 (Br (q0 )). Hence the map
expq0 = φ0 ◦ expp0 ◦ Φ−1
0 : Br (q0 ) → expq0 (Br (q0 ))

is a diffeomorphism, hence r ≤ inj(q0 ; M ) by Definition 4.3.9, and hence

Theorem 4.4.4 asserts that expq0 (Br (q0 )) = Ur (q0 ). This proves Step 1.
Step 2. The sequence φi converges in the C ∞ topology on all of M to a
smooth map φ : M → M .
Define the set M0 := {p ∈ M | the sequence (φi (p), dφi (p)) converges} . This
set is nonempty because p0 ∈ M0 . We prove that M0 is open. Fix any
element p ∈ M0 and define q := limi→∞ φi (p) and Φ := limi→∞ dφi (p).
Choose a real number r such that 0 < r < inj(p; M ). Then Step 1 as-
serts that r ≤ inj(q; M ) and the sequence φi |Ur (p) converges to the diffeo-
morphism expq ◦ Φ ◦ exp−1 p : Ur (p) → Ur (q) in the C
∞ topology on U (p),
r
and hence Ur (p) ⊂ M0 . This shows that M0 is open, that φi |M0 converges
in the C ∞ -topology to a smooth map φ : M0 → M , and that φ(M0 ) is an
open subset of M .
We prove that M0 is closed. Let pν ∈ M0 be a sequence that con-
verges to an element p ∈ M . Then there exists a real number r > 0 such
that r < inj(pν ; M ) for all ν. Hence Br (pν ) ⊂ M0 for all ν by Step 1.
Choose ν so large that d(pν , p) < r to obtain p ∈ M0 . This shows that M0 is
closed. Since M is connected, we deduce that M0 = M . This proves Step 2.
Step 3. The map φ : M → M in Step 2 is an isometry.
We claim that the triple p0 , q0 , Φ0 in (5.4.5) satisfies
lim φ−1
i (q0 ) = p0 , lim dφ−1 −1
i (q0 ) = Φ0 . (5.4.6)
i→∞ i→∞

Namely, the sequence d(φ−1 i (q0 ), p0 ) = d(q0 , φi (p0 )) converges to zero by as-
sumption, and by Step 1 the derivatives dφi converge to dφ uniformly in
some neighborhood U0 ⊂ M of p0 . Since φ−1 i (q0 ) ∈ U0 for i sufficiently
large, this implies that the sequence dφi (φ−1 i (q 0 )) converges to dφ(p0 ) = Φ0
and hence the sequence dφi (φ−1 i (q 0 ))−1 = dφ−1 (q ) converges to Φ−1 . This
i 0 0
proves (5.4.6). By (5.4.6) and Step 2 the sequence φ−1 i converges in the C ∞
topology to a smooth map ψ : M → M . By uniform convergence on com-
pact subsets of M we have ψ ◦ φ = id and φ ◦ ψ = id, so φ : M → M is a
diffeomorphism. Moreover, since φi is an isometry for each i and converges
pointwise to φ, we have d(φ(p), φ(q)) = limi→∞ d(φi (p), φi (q)) = d(p, q) for
all p, q ∈ M , so φ is an isometry. This proves Step 3 and Lemma 5.4.2.
5.4. THE GROUP OF ISOMETRIES* 259

The next goal is to verify that the spaces Ip in (5.4.4) are diffeomorphic
to each other. For p0 , p1 ∈ M define the map Fp1 ,p0 : Ip0 → Ip1 by
Fp1 ,p0 (φ(p0 ), dφ(p0 )) := (φ(p1 ), dφ(p1 )), φ ∈ I(M ). (5.4.7)
This map is well-defined by Lemma 5.1.10, because M is connected. Col-
lectively, these maps give rise to a map F : M × I → I defined by
 
I := (φ(p), dφ(p), p)  p ∈ M, φ ∈ I(M ) ⊂ G,
(5.4.8)
F(p1 , (φ(p0 ), dφ(p0 ), p0 )) := (φ(p1 ), dφ(p1 ), p1 )
for p0 , p1 ∈ M and φ ∈ I(M ). The next lemma uses the concept of a smooth
map on an arbitrary subset of Euclidean space as in the beginning of §2.1.
Lemma 5.4.3. The map F : M × I → I is smooth. In particular, for each
pair of points p0 , p1 ∈ M , the map Fp1 ,p0 : Ip0 → Ip1 is a diffeomorphism.
Proof. Let p0 , p1 ∈ M and choose a smooth path γ : I = [0, 1] → M with
the endpoints γ(0) = p0 and γ(1) = p1 . Let Uγ ⊂ Gp0 be the set of all
pairs (q0 , Φ0 ) ∈ Gp0 such that there exists a development (Φ, γ, γ 0 ) on the
interval I that satisfies
γ 0 (0) = q0 , Φ(t) = Φ0 . (5.4.9)
By Remark 3.5.22, the set Uγ is open in Gp0 and the map Uγ → G that sends
the pair (q0 , Φ0 ) ∈ Uγ to the pair (γ 0 (1), Φ(1)) ∈ Gp1 is smooth. This shows
that there is a unique diffeomorphism
Fγ : Uγ → Uγ −1 , γ −1 (t) := γ(1 − t), Uγ ⊂ Gp0 , Uγ −1 ⊂ Gp1
that satisfies the condition
Fγ (γ 0 (0), Φ(0)) = (γ 0 (1), Φ(1)) (5.4.10)
for every development (Φ, γ, γ 0 )
of M along M on the interval I. The inverse
of Fγ is the smooth map Fγ −1 : Uγ −1 → Uγ . If φ ∈ I(M ), then by Corol-
lary 5.3.4 there exists a development (Φ, γ, γ 0 ) on I satisfying the initial
conditions γ 0 (0) = φ(p0 ) and Φ(0) = dφ(p0 ), and it is given by
γ 0 (t) = φ(γ(t)), Φ(t) = dφ(γ(t)) (5.4.11)
for t ∈ I. Hence (φ(p0 ), dφ(p0 )) ∈ Uγ and by (5.4.10) and (5.4.11) we have
Fγ (φ(p0 ), dφ(p0 )) = (φ(p1 ), dφ(p1 )) = Fp1 ,p0 (φ(p0 ), dφ(p0 ))
for every φ ∈ I(M ). Thus Ip0 ⊂ Uγ and Ip1 ⊂ Uγ −1 and Fγ |Ip0 = Fp1 ,p0 .
Hence Fp1 ,p0 is smooth. The smoothness of F follows from the smooth
dependence of the map Fγ on the curve γ, the verification of which we leave
to the reader. This proves Lemma 5.4.3.
260 CHAPTER 5. CURVATURE

5.4.3 Killing Vector Fields

Killing vector fields are defined as those vector fields on M whose flows
are one-parameter families of isometries. Assume throughout that M is a
nonempty connected smooth m-dimensional submanifold of Rn . Let X be
a vector field on M and denote by

R × M ⊃ D → M : (t, p) 7→ φ(t, p) = φt (p)

the flow of X (Definition 2.4.8). Then Theorem 2.4.9 asserts that D is

an open subset of R × M and φ is smooth. Thus for every t ∈ R the
set Dt := {p ∈ M | (t, p) ∈ D} is open in M and the map φt : Dt → D−t is a
diffeomorphism with the inverse φ−t : D−t → Dt .

Lemma 5.4.4. In this situation the following are equivalent.

(i) For every t ∈ R the diffeomorphism φt : Dt → D−t is an isometry.
(ii) For every p ∈ M and every pair of tangent vectors u, v ∈ Tp M , we have

h∇u X(p), vi + hu, ∇v X(p)i = 0. (5.4.12)

Proof. Let p ∈ M and v ∈ Tp M . Choose a smooth curve α : R → M such

that α(0) = p and α̇(0) = v, let Ω := {(s, t) ∈ R2 | (s, α(t)) ∈ D}, and define
the map γ : Ω → M by γ(s, t) := φt (α(s)) for (s, t) ∈ Ω. Then

∂s γ(0, t) = dφt (p)v, ∂t γ = X ◦ γ, ∇s ∂t γ = ∇∂s γ X(γ).

Thus the formula ∇t ∂s γ = ∇s ∂t γ in Lemma 3.2.4 shows that

∇t dφt (p)v = ∇dφt (p)v X(φt (p)) (5.4.13)

for all t ∈ R and this implies

d  t 2
dφ (p)v  = 2 ∇dφt (p)v X(φt (p)), dφt (p)v .


(5.4.14)
dt
The right hand side vanishes for all p, v, t if and only if X satisfies (i), and the
left hand side vanishes for all p, v, t if and only if the flow of X satisfies (ii).
This proves Lemma 5.4.4.

Definition 5.4.5. A vector field X ∈ Vect(M ) is called a Killing vector

field (named after Wilhelm Karl Joseph Killing [29]) iff it satisfies equa-
tion (5.4.12) for all p ∈ M and all u, v ∈ Tp M . The space of Killing vector
fields will be denoted by VectK (M ).
5.4. THE GROUP OF ISOMETRIES* 261

The next lemma shows that the space of Killing vector fields is a finite-
dimensional Lie subalgebra of the Lie algebra Vect(M ) of vector fields on M .
Part (ii) is the linear counterpart of Lemma 5.1.10.
Lemma 5.4.6. Let M ⊂ Rn be a connected smooth m-manifold. Then the
following holds.
(i) If X is a Killing vector field on M and ψ : M → M is an isometry, then
the pullback ψ ∗ X is a Killing vector field. If X and Y are Killing vector
fields on M , then so is their Lie bracket [X, Y ].
(ii) If X is a Killing vector field on M and there exists a p0 ∈ M such that

X(p0 ) = 0, ∇X(p0 ) = 0, (5.4.15)

then X(p) = 0 for all p ∈ M .

(iii) If M is complete and X is a Killing vector field, then X is complete.
Proof. We prove part (i). If φt : Dt → D−t is the flow of X, then

ψ −1 ◦ φt ◦ ψ : ψ −1 (Dt ) → ψ −1 (D−t )

is the flow of ψ ∗ X. Hence the flow of ψ ∗ X is a one-parameter family of

isometries by Lemma 5.4.4 and so, again by Lemma 5.4.4, ψ ∗ X is a Killing
vector field. Now assume that both X and Y are Killing vector fields and
denote by ψ t the flow of Y . Then each ψ t is an isometry by Lemma 5.4.4
and [X, Y ] = dtd
t=0
(ψ t )∗ X by Lemma 2.4.18. Hence the pullback (ψ t )∗ X
is a Killing vector field for each t by what we have already proved, and so
is their limit as t tends to zero. This proves (i).
We prove part (ii). Let φt : Dt → D−t be the flow of X. Then it follows
from equations (5.4.13) and (5.4.15) that φt (p0 ) = p0 and dφt (p0 ) = id for
all t. By Lemma 5.1.10 this implies that the isometry φt : Dt → D−t is the
identity for every t ∈ R. Thus Dt = D−t = M for all t and X(p) = 0 for
all p ∈ M . This proves (ii).
We prove part (iii). Thus assume that M is complete and X is a
Killing vector field. Let γ : I → M be an integral curve of X on its
maximal existence interval I ⊂ R and denote p := γ(0). Differentiate the
d
equation γ̇(t) = X(γ(t)) to obtain dt |γ̇(t)|2 = 2h∇γ̇(t) X(γ(t)), γ̇(t)i = 0 for
all t ∈ I. Hence the functionR I → R : t 7→ |γ̇(t)| is constant and this implies
t
the inequality d(p, γ(t)) ≤ 0 |γ̇(s)| ds = t|X(p)| for all t ∈ I. Since M is
complete, the closed ball of radius R about p is compact for every R > 0
(Theorem 4.6.5). Hence the restriction of γ to any bounded subinterval
of I is contained in a compact subset of M amd by Corollary 2.4.15 this
implies I = R. This proves (iii) and Lemma 5.4.6.
262 CHAPTER 5. CURVATURE

The proof of Theorem 5.4.1 is somewhat analogous to the proof of the

Closed Subgroup Theorem 2.5.26. The first parallel is in part (i) of the next
lemma, which asserts that the space of complete Killing vector fields is a Lie
subalgebra of Vect(M ) and can be viewed as an anlogue of Lemma 2.5.28.
Part (ii) is the analogue of Lemma 2.5.27.
Lemma 5.4.7. (i) The set VectK,c (M ) of complete Killing vector fields
on M is a Lie subalgebra of VectK (M ).
(ii) Let R × M → M : (t, p) 7→ ψt (p) be a smooth map such that ψt is an
d

isometry for every t ∈ R and define X(p) := dt t=0 ψt (p) for p ∈ M . Then X
is a complete Killing vector field.
Proof. The proof has three steps.
Step 1. Let U, V ⊂ M be nonempty open sets, let φ : U → V be an isometry,
and let φk : M → M be a sequence of isometries that converges uniformly
on every compact subset of U to φ. Then φ extends uniquely to an isometry
from M to M and φk converges in the C ∞ topology to this extension.
Fix an element p0 ∈ U . Then by part (i) of Exercise 5.1.11 we have
lim φk (p0 ) = φ(p0 ), lim dφk (p0 ) = dφ(p0 ).
k→∞ k→∞

Hence the assertion of Step 1 follows from Lemma 5.4.2.

Step 2. We prove part (ii).
That X is a Killing vector field follows from the same argument as in
Lemma 5.4.4 with time dependent vector fields. Denote by φt : Dt → D−t
the flow of X. Then the sequence of isometries ψt/k k : M → M converges

to φt uniformly on every compact subset of Dt (see Exercise 5.4.10 below).

Hence Step 1 asserts that φt extends to an isometry on all of M when-
ever Dt is nonempty, in particular for small t. The extended isometries still
satisfy φs+t = φs ◦ φt and ∂t φt = X ◦ φt . Thus Dt = M for small t and so for
all t, because φ−s (D−s ∩ Dt ) ⊂ Ds+t (Theorem 2.4.9). This proves Step 2.
Step 3. We prove part (i).
Let X, Y ∈ VectK,c (M ), let φt be the flow of X, and let ψ t be the flow of Y .
d
Then dt t=0
φt ◦ ψ t (p) = X(p) + Y (p) for all p ∈ M , and so X + Y is a com-
plete Killing vector field by Step 2. Hence VectK,c (M ) is a vector space. It is
finite-dimensional by Lemma 5.4.6. Moreover, (ψ −t ◦φs ◦ψ t )s∈R is the flow of
the pullback vector field (ψ t )∗ X, hence (ψ t )∗ X is a complete Killing vector
field for each t ∈ R, and hence so is the Lie bracket [X, Y ] = dt d
|t=0 (ψ t )∗ X
by finite-dimensionality. This proves Step 3 and Lemma 5.4.7.
5.4. THE GROUP OF ISOMETRIES* 263

Exercise 5.4.8. Let M ⊂ Rn be a smooth manifold and let X be a vector

field on M . Consider the following condition.
(K) If γ : I → M is a geodesic, then X ◦γ ∈ Vect(γ) is a Jacobi field along γ,
i.e. it satisfies the differential equation ∇∇(X ◦ γ) + R(X ◦ γ, γ̇)γ̇ = 0.
Prove that every Killing vector field satisfies (K). If M is compact, prove
that a vector field satisfies (K) if and only if it is a Killing vector field. Find a
vector field on a noncompact manifold that satisfies (K) but is not a Killing
vector field. Hint: Differentiate the function t 7→ hX(γ(t)), γ̇(t)i twice.
Exercise 5.4.9 (Grönwall’s inequality). If a, c ≥ 0 and σ : [0, T ] → R is
a continuous function satisfying
Z t
0 ≤ σ(t) ≤ a + c α(s) ds for 0 ≤ t ≤ T, (5.4.16)
0
then σ(t) ≤ aect for every real number 0 ≤ t ≤ T . Hint: The function
Z t
τ (t) := (aecs − σ(s)) ds
0
satisfies τ̇ (t) ≥ cτ (t) for all t. Differentiate the function t 7→ e−ct τ (t) to show
that τ is nonnegative and nondecreasing.
Exercise 5.4.10. Let T, c, ε be positive real numbers, let M ⊂ Rn be a
smooth m-dimensional submanifold, and let [0, T ]×M → Rn : (t, p) 7→ Xt (p)
and [0, T ] × M → Rn : (t, p) 7→ Yt (p) be continuous maps that satisfy the
conditions Xt (p), Yt (p) ∈ Tp M and
|Xt (p) − Yt (p)|Rn ≤ ε, |Xt (p) − Xt (q)|Rn ≤ c|p − q|Rn (5.4.17)
for all p, q ∈ M and all t ∈ [0, T ]. If the curves β, γ : [0, T ] → M are solutions
of the differential equations
β̇(t) = Xt (β(t)), γ̇(t) = Yt (γ(t)),
prove that  
|β(t) − γ(t)|Rn ≤ |β(0) − γ(0)|Rn + T ε ect (5.4.18)
for 0 ≤ t ≤ T . Relax the continuity requirement in t on the vector fields to
piecewise continuity. Hint: Define the function σ : [0, T ] → R by
σ(t) := |β(t) − γ(t)|Rn .
Show that Z t
σ(t) ≤ σ(0) + tε + c σ(s) ds
0
and use Grönwall’s inequality.
264 CHAPTER 5. CURVATURE

5.4.4 Proof of the Myers–Steenrod Theorem

With these preparations we are ready to prove Theorem 5.4.1. The following
lemma is the heart of the proof. It is the analogue of Lemma 2.5.29 in the
proof of the Closed Subgroup Theorem 2.5.26.
Lemma 5.4.11. Fix an element p0 ∈ M and let φi ∈ I(M ) be a sequence
of isometries such that φi 6= id for all i ∈ N and
lim φi (p0 ) = p0 , lim dφi (p0 ) = 1l. (5.4.19)
i→∞ i→∞
Let v0 ∈ Rn , let A0 : Tp0 M → Rn be a linear map, and let τi be a sequence
of positive real numbers such that limi→∞ τi = 0 and, for all v ∈ Tp0 M ,
φi (p0 ) − p0 dφi (p0 )v − v
lim = v0 , lim = A0 v. (5.4.20)
i→∞ τi i→∞ τi
Then there exists a complete Killing vector field X ∈ VectK,c (M ) such that
X(p0 ) = v0 , dX(p0 ) = A0 . (5.4.21)
For every p ∈ M and every v ∈ Tp M this vector field satisfies
φi (p) − p dφi (p)v − v
lim = X(p), lim = dX(p)v. (5.4.22)
i→∞ τi i→∞ τi
and the convergence is is uniform on compact subsets of T M .
Proof. By (5.4.19), (5.4.20), and Exercise 2.2.4 we have
(p0 , A0 ) ∈ T(p0 ,1l) Gp0 .
Let p ∈ M and recall the definition of the map Fp,p0 : Ip0 → Ip in (5.4.7).
By Lemma 5.4.3 this map extends to a diffeomorphism Fγ from an open sub-
set Uγ ⊂ Gp0 containing the set Ip0 to an open subset Uγ −1 ⊂ Gp containing
the set Ip and it satisfies Fγ (p0 , 1l) = (p, 1l). Define
(X(p), A(p)) := dFγ (p0 , 1l)(v0 , A0 ) ∈ T(p,1l) Gp .
Since Fγ (φi (p0 ), dφi (p0 )) = (φi (p), dφi (p)) for all i, it follows from (5.4.20)
and Exercise 2.2.16 that, for all v ∈ Tp M , we have
φi (p) − p dφi (p)v − v
lim = X(p), lim = A(p)v. (5.4.23)
i→∞ τi i→∞ τi
This formula shows that the pair (X(p), A(p)) ∈ T(p,1l) Gp is independent of
the choice of the extension Fγ used to define it. Moreover, it follows from the
smoothness of the map F in Lemma 5.4.3 that the map p 7→ (X(p), A(p))
is smooth and that the convergence in (5.4.23) is uniform as the pair (p, v)
varies over any compact subset of T M .
5.4. THE GROUP OF ISOMETRIES* 265

Now let v ∈ Tp M and define γ(t) := expp (tv) for t ∈ Ip,v . Then
φi (γ(t)) − γ(t) φi (p) − p
X(γ(t)) − X(p) = lim − lim
i→∞ τi i→∞ τi
φi (γ(t)) − φi (γ(0)) − γ(t) + γ(0)
= lim
i→∞ τi
Z t
dφi (γ(s))γ̇(s) − γ̇(s)
= lim ds
i→∞ 0 τi
Z t
= A(γ(s))γ̇(s) ds.
0

Here the last step uses uniform convergence on compact sets in (5.4.23).
Divide by t and use the continuity of the curve t 7→ A(γ(t))γ̇(t) to obtain
1 t X(expp (tv)) − X(p)
Z
A(p)v = lim A(γ(s))γ̇(s) ds = lim .
t→0 t 0 t→0 t
Hence, by Exercise 2.2.16, we deduce that, for all p ∈ M and all v ∈ Tp M ,

A(p)v = dX(p)v.

By (5.4.20) and (5.4.23) this shows that X satisfies (5.4.21) and (5.4.22).
By (5.4.22) and Exercise 2.2.4 we have (X(p), dX(p)) ∈ T(p,1l) Gp and this
implies hv, dX(p)vi = 0 for all v ∈ Tp M . Here is a more direct proof of this
crucial fact. Namely, by (5.4.22) we have
|dφi (p)v − v|2
lim = |dX(p)v|2 .
i→∞ τi2
Hence
hv, dφi (p)v − vi
hv, dX(p)vi = lim
i→∞ τi
hv, dφi (p)vi − |v|2
= lim
i→∞ τi
2hv, dφi (p)vi − |v|2 − |dφi (p)v|2
= lim
i→∞ 2τi
|dφi (p)v − v|2 τi
= − lim
i→∞ τi2 2
=0
for all p ∈ M and all v ∈ Tp M . This shows that X is a Killing vector field
and so it remains to prove that X is complete.
266 CHAPTER 5. CURVATURE

Let φt : Dt → D−t be the flow of X and, for p ∈ M , denote by

 
Ip := t ∈ R  p ∈ Dt

the maximal existence interval for the solution γ of the initial value prob-
lem γ̇(t) = X(γ(t)), γ(0) = p. We prove in three steps that X is complete.
Step 1. Let p ∈ M and let T > 0 such that T |X(p)| < inj(p; M ). Then

[−T, T ] ⊂ Ip .

Define γ(t) := φt (p) for t ∈ Ip . Since X is a Killing vector field, the diffeo-
morphism φt : Dt → D−t is an isometry by Lemma 5.4.4, and hence

|γ̇(t)| = |X(φt (p))| = |dφt (p)X(p)| = |X(p)|,

for all t ∈ Ip . This implies d(p, γ(t)) ≤ t|X(p)| for all t ∈ Ip , and hence γ
cannot leave the compact set UT |X(p)| (p) = {q ∈ M | d(p, q) ≤ T |X(p)|} on
any subinterval I ⊂ [−T, T ]. Hence [−T, T ] ⊂ Ip by Corollary 2.4.15 and
this proves Step 1.
Step 2. Let p ∈ M . If t ∈ R satisfies |tX(p)| < inj(p; M ) and the sequence
of integers mi ∈ Z is chosen such that

mi τi ≤ t < (mi + 1)τi , (5.4.24)

then φt (p) = limi→∞ φm i

i (p).

Let T > 0 such that T |X(p)| < inj(p; M ). Then [0, T ] ⊂ Ip by Step 1 and
the set
K := φt (p)  0 ≤ t ≤ T
 

is compact and |X(q)| = |X(p)| for all q ∈ K. By Lemma 4.2.7 there exists
a constant δ > 0 such that, for all q, q0 , q1 ∈ M ,

d(q0 , q1 )
q ∈ K, d(q, q0 ) ≤ δ, d(q, q1 ) ≤ δ =⇒ ≤ 2. (5.4.25)
|q0 − q1 |

Fix a real number 0 < ε ≤ 1 and choose i0 ∈ N such that, for all i ≥ i0 ,

τi |X(p)| < δ, (T + τi )|X(p)| < inj(p; M ), sup d(q, φi (q)) < δ, (5.4.26)
q∈K
   τ 
 φi (q) − q   φ i (q) − q 
sup 
 − X(q) < ε, sup 
 − X(q) < ε. (5.4.27)
q∈K τi q∈K τi
5.4. THE GROUP OF ISOMETRIES* 267

Then, for all i ≥ i0 and all q ∈ K, we have d(q, φτi (q)) ≤ τi |X(p)| < δ
and d(q, φi (q)) < δ by (5.4.26), and hence by (5.4.25) and (5.4.27),
sup d(φi (q), φτi (q)) ≤ 2 sup |φi (q) − φτi (q)| ≤ 4τi ε. (5.4.28)
q∈K q∈K

Now choose a real number 0 ≤ t ≤ T and choose mi ∈ N0 as in (5.4.24).

Then, for k = 1, . . . , mi − 1 we have φkτi (p) ∈ K and hence
 
d φk+1
i (p), φ (k+1)τi
(p)

 

≤ d φi φki (p) , φi φkτi (p) + d φi φkτi (p) , φτi φkτi (p)



  

= d φki (p), φkτi (p) + d φi φkτi (p) , φτi φkτi (p)


 
≤ d φki (p), φkτi (p) + 4τi ε.

Here the last inequality holds for i ≥ i0 by (5.4.28). By induction this implies
d(φm i
i (p), φ
mi τi
(p)) ≤ 4mi τi ε ≤ 4T ε.
for all i ≥ i0 . Hence
lim φm i
i (p) = lim φ
m i τi
(p) = φt (p)
i→∞ i→∞
and this proves Step 2 for t ≥ 0. For t ≤ 0 the argument is similar.
Step 3. Ip = R for every p ∈ M .
Fix an element p ∈ M and real number T > 0 such that T |X(p)| < inj(p; M ).
Choose the sequence mi ∈ N0 such that τi mi ≤ T < (mi + 1)τi . Then Step 2
asserts that φT (p) = limi→∞ φm mi
i (p). Since φi : M → M is an isometry, it
i

mi
follows that T |X(p)| < inj(φi (p); M ) for all i ∈ N and hence
T |X(p)| < inj φT (p); M .

By Step 1 this implies [0, T ] ⊂ IφT (p) . Hence [0, 2T ] ⊂ Ip and, by Step 2,
φ2T (p) = lim φm φT (p) .
i


i→∞ i

Continue by induction to obtain for all k ∈ N that

T |X(p)| < inj φkT (p); M

and hence [0, (k + 1)T ] ⊂ Ip and

φ(k+1)T (p) = lim φm φkT (p) .
i


i→∞ i

Thus [0, ∞) ⊂ Ip and the same argument shows that (−∞, 0] ⊂ Ip . Hence X
is complete. This proves Step 3 and Lemma 5.4.11.
268 CHAPTER 5. CURVATURE

The next lemma establishes the smooth structure on I(M ), in analogy

to the proof of the Closed Subgroup Theorem 2.5.26.
Lemma 5.4.12. For every p ∈ M the set Ip is a smooth submanifold of Gp
and its tangent space at (q, Φ) ∈ Ip is given by
 
T(q,Φ) Ip = (X(q), dX(q)Φ)  X ∈ VectK,c (M ) . (5.4.29)

Proof. By Lemma 3.4.5 with Rm replaced by Tp M the tangent space of Gp

at an element (q, Φ) ∈ Gp is given by
  

  q
b ∈ L(Tp M, Rn ),
 qb ∈ T M, Φ 



T(q,Φ) Gp = (b
q , Φ)  1l − Π(q) Φ = hq (b
q )Φ, and . (5.4.30)
b  b
  b Φvi = 0 ∀ u, v ∈ Tp M. 
  hΦu, Φvi
b + hΦu, 

If (q, Φ) ∈ Gp and X ∈ Vect(M ), then (1l − Π(q))dX(q)Φ = hq (X(q))Φ by

the Gauß–Weingarten formula in Remark 5.2.5. Moreover, if X is a Killing
vector field, then hΦu, dX(q)Φvi + hdX(q)Φu, Φvi = 0 for all u, v ∈ Tp M , so
it follows from (5.4.30) that the pair (b
q , Φ)
b with qb = X(q) and Φ
b = dX(q)Φ
is a tangent vector of Gp at (q, Φ).
Now abbreviate

m(m + 1)
k := dim VectK,c (M ) ≤ = dim(Gp ) =: `.
2
Fix an isometry φ0 : M → M and define

(q0 , Φ0 ) := (φ0 (p), dφ0 (p)) = ιp (φ0 ) ∈ Ip .

We must construct a coordinate chart on Gp in a neighborhood of the

point (q0 , Φ0 ) ∈ Ip with values in an open set Ω ⊂ R` that sends Ip to the
intersection of Ω with Rk × {0}. For this we first choose a basis Y1 , . . . , Yk
of VectK,c (M ) and then a basis η1 = (b q1 , Φ
b 1 ), . . . , η` = (b
q` , Φ
b ` ) of the tangent
space T(q0 ,Φ0 ) Gp such that



ηj = (bqj , Φ
b j ) = Yj φ0 (p) , dYj φ0 (p) dφ0 (p) , j = 1, . . . , k.

Next we choose any smooth map

R`−k → Gp : (tk+1 , . . . , t` ) 7→ ι(tk+1 , . . . , t` )

such that ι(0) = (q0 , Φ0 ) and

∂ι
(0, . . . , 0) = ηj = (b
qj , Φ
b j ), j = k + 1, . . . , `.
∂tj
5.4. THE GROUP OF ISOMETRIES* 269

For a complete Killing vector field X ∈ VectK,c (M ) let ψX ∈ I(M ) be the

time-1 map of the flow of X and define the map Θ : R` → Gp by
 
Θ t1 , . . . , t` := ψX (q), dψX (q)Φ ,

X := t1 Y1 + · · · + tk Yk , (5.4.31)
(q, Φ) := ι tk+1 , . . . , t`

for t1 , . . . , t` ∈ R` . Then

Θ(0) = ι(0) = (q0 , Φ0 ) = ιp (φ0 ) ∈ Ip

and
dΘ(0) t̂1 , . . . t̂` = t̂1 η1 + · · · + t̂` η`

for t̂1 , . . . , t̂` ∈ R` . Thus the derivative dΘ(0) : R` → T(q0 ,Φ0 ) Gp is a bi-

jective linear map and so the Inverse Function Theorem asserts that the
map Θ restricts to a diffeomorphism from a sufficiently small open neigh-
borhood Ω ⊂ R` of the origin onto its image ι(Ω) ⊂ Gp which is an open
neighborhood in Gp of the point Θ(0) = ιp (φ0 ) ∈ Ip . With this understood,
the assertion that Ip is a smooth submanifold of Gp is a direct consequence
of the following Claim.
Claim. There exists an open set Ω0 ⊂ R` such that

Θ Ω0 ∩ (Rk × {0}) = U0 ∩ Ip ,


0 ∈ Ω0 ⊂ Ω, U0 := Θ(Ω0 ). (5.4.32)

Suppose, by contradiction, that such
an open set Ω0 does not exist. Then
there exist sequences ti = t1i , . . . , t`i ∈ R` and φi ∈ I(M ) such that

ti ∈ Ω \ Rk × {0} ,


lim ti = 0, Θ(ti ) = ιp (φi ) ∈ Ip .
i→∞

Define

Xi := t1i Y1 + · · · + tki Yk , (qi , Φi ) := ι tk+1 , . . . , t`i .

i



Then ψXi (qi ), dψXi (qi )Φi = Θ(ti ) = ιp (φi ) = φi (p), dφi (p) and hence
−1 −1
qi = (ψXi
◦ φi )(p), Φi = dψXi (qi )−1 dφi (p) = d(ψX i
◦ φi )(p).
−1
◦ φi ) = (qi , Φi ) = Θ 0, . . . , 0, tk+1 , . . . , t`i is still a sequence in


Thus ιp (ψXi i
the set Θ(Ω \ (Rk × {0})) ∩ Ip that converges to ιp (φ0 ). Thus we may as-
sume without loss of generality that t1i = t2i = · · · = tki = 0 and so

ι(ti ) = ιp (φi ), lim ιp (φi ) = ιp (φ0 ). (5.4.33)

i→∞
270 CHAPTER 5. CURVATURE

Then φi ◦ φ−1
0 converges to the identity in the C
∞ topology by Lemma 5.4.2.

Since φi 6= φ0 , we have

|d(φi ◦ φ−1
0 )(p)v − v|
τi := (φi ◦ φ−1
 
0 )(p) − p + sup >0 (5.4.34)

06=v∈Tp M |v|

for all i by Lemma 5.1.10, and limi→∞ τi = 0. Hence, by Exercise 2.2.4 there
exists a tangent vector (v0 , A0 ) ∈ T(p,1l) Gp and a subsequence, still denoted
by φi , such that, for all v ∈ Tp M ,

(φi ◦ φ−1
0 )(p) − p d(φi ◦ φ−1
0 )(p)v − v
lim = v0 , lim = A0 v. (5.4.35)
i→∞ τi i→∞ τi
It follows from (5.4.35) and Lemma 5.4.11 that there exists a complete
Killing vector field X ∈ VectK,c (M ) such that

X(p) = v0 , dX(p) = A0 .

This vector field is nonzero because (v0 , A0 ) 6= 0 by (5.4.34). Moreover, by

equation (5.4.22) in Lemma 5.4.11, with φi replaced by φi ◦ φ−1 0 and the
pair (p, v) replaced by the pair (φ0 (p), dφ0 (p)v), we obtain

φi (p) − φ0 (p)
X(φ0 (p)) = lim ,
i→∞ τi
dφi (p)v − dφ0 (p)v
dX(φ0 (p))dφ0 (p)v = lim
i→∞ τi
for v ∈ Tp M . Since by (5.4.33) the sequence

(φi (p), dφi (p)) = ιp (φi ) = ι(ti )

converges to (φ0 (p), dφ0 (p)) = ι(0), the pair (X(φ0 (p)), dX(φ0 (p))dφ0 (p))
belongs to the image of the derivative dι(0) by Exercise 2.2.4, and hence
is a nonzero linear combination of the vectors ηk+1 , . . . , η` . It is also a
linear combination of the vectors η1 , . . . , ηk , because X is a complete Killing
vector field and so is a linear combination of the vector fields Y1 , . . . , Yk .
This is a contradiction, and this contradiction proves the Claim. Thus there
does, after all, exist an open set Ω0 ⊂ R` that satisfies (5.4.32), and the
map Θ−1 : U0 → Ω0 is then a coordinate chart on Gp which satisfies

Θ−1 (U0 ∩ Ip ) = Ω0 ∩ (Rk × {0}).

Hence Ip is a submanifold of Gp and this proves Lemma 5.4.12.

5.4. THE GROUP OF ISOMETRIES* 271

Proof of Theorem 5.4.1. By Lemma 5.4.3 and Lemma 5.4.12 there exists
a unique smooth structure on I(M ) such that the map ιp : I(M ) → Gp
in (5.4.3) is an embedding for every p ∈ M . That the topology induced
by this smooth structure agrees with the C ∞ topology was established in
Lemma 5.4.2. That it also agrees with the compact open topology (i.e. with
the C 0 topology of uniform convergence on conpact sets) is the content of
Exercise 5.1.11. This proves part (i).
We prove part (ii). Recall the definition of the map

F :M ×I →I

in (5.4.8) and the target map t : G → M , the inverse map i : G → G, and the
composition map m : (s × t)−1 (∆) → G in the beginning of §5.4.1. These
maps are all smooth and turn I into a smooth subgroupoid of G. For
each p ∈ M they endow the submanifold Ip ⊂ Gp with the structure of a Lie
group as follows. The unit is the element

ep := (p, 1l) ∈ Ip (5.4.36)

The product is the map mp : Ip × Ip → Ip defined by




(mp (η, ξ), p) := m F t(ξ, p), (η, p) , ξ, p (5.4.37)

for ξ, η ∈ Ip , and the inverse map is the involution ip : Ip → Ip defined by

(ip (ξ), p) := F p, i(ξ, p) (5.4.38)

for ξ ∈ Ip . The maps mp in (5.4.37) and ip in (5.4.38) are smooth by

definition. To show that they define a group structure and that the map

ιp : I(M ) → Ip

in (5.4.3) is a group isomorphism, we must verify the identities

ιp (id) = ep ,
mp (ιp (ψ), ιp (φ)) = ιp (ψ ◦ φ), (5.4.39)
−1
ip (ιp (φ)) = ιp (φ )

for all φ, ψ ∈ I(p). The first equation in (5.4.39) follows directly from the
definitions. To prove the remaining equations, fix two isometries φ and ψ
and define ξ, η, ζ ∈ Ip by

ξ := ιp (φ), η := ιp (ψ), ζ := ιp (ψ ◦ φ).

272 CHAPTER 5. CURVATURE

Then, by the chain rule,

 
(ζ, p) = ψ(φ(p)), dψ(φ(p))dφ(p), p



= m ψ(φ(p)), dψ(φ(p)), φ(p) , φ(p), dφ(p), p



= m F φ(p), (ψ(p), dψ(p), p) , φ(p), dφ(p), p



= m F t(ξ, p), (η, p) , ξ, p
= (mp (η, ξ), p).

Here the last equality follows from the definition of the map mp in (5.4.37).
This proves the second equation in (5.4.39).
To verify the third equation in (5.4.39), we compute

ιp (φ−1 ), p = φ−1 (p), dφ−1 (p), p

= F p, (φ−1 (φ(p)), dφ−1 (φ(p)), φ(p))

= F p, (p, dφ(p)−1 , φ(p))

= F p, i(φ(p), dφ(p), p)


= F p, i(ξ, p)
= (ip (ξ), p).

Here the last equality follows from the definition of the map ip in (5.4.38).
This proves the third equation in (5.4.39) and part (ii).
That the Lie algebra of Ip is isomorphic to the space VectK,c (M ) of
complete Killing vector fields under the Lie algebra isomorphism

VectK,c (M ) → Tep Ip : X 7→ (X(p), dX(p))

follows from Lemma 5.4.12. This proves part (iii) and Theorem 5.4.1.

Corollary 5.4.13. Let (p, e) ∈ O(M ). Then there exists a constant ε > 0
with the following significance. If φ : M → M is an isometry that satisfies

d(p, φ(p)) < ε, |e − dφ(p)e|Rn×m < ε,

then there exists a complete Killing vector field X ∈ VectK,c (M ) whose flow
has the time-1-map ψX = φ.

Proof. Use the construction of the map Θ : Ω → Gp and the Claim in the
proof of Lemma 5.4.12 with φ0 = id to obtain φ ∈ Θ(Ω ∩ (Rk × {0})).
5.4. THE GROUP OF ISOMETRIES* 273

5.4.5 Isometries of Compact Lie Groups

In the following theorem we denote by I0 (M ) the connected component of
the identity in the group of all isometries of a manifold M .

Theorem 5.4.14. Let G ⊂ GL(n, R) be a compact connected Lie group,

equipped with a bi-invariant Riemannian metric. If φ ∈ I0 (G), then there
exist elements a, b ∈ G such that

φ(g) = φa,b (g) := agb−1 for all g ∈ G. (5.4.40)

Proof. See page 277.

The proof of Theorem 5.4.14 is based on four lemmas. The heart of the
proof is Lemma 5.4.16. It equires the following preparatory result, which
makes use of the Killing form in Example 5.2.24. The case where the Lie
algebra has a nontrivial center is then dealt with in Lemma 5.4.18.

Lemma 5.4.15. Let g be a finite-dimensional Lie algebra.

(i) If g admits an invariant inner product and has a trivial center, then the
Killing form on g is nondegenerate.
(ii) If the Killing form on g is nondegenerate, then the Lie algebra homo-
morphism ad : g → Der(g) is bijective.

Proof. Part (i) was explained in Exercise 5.2.26. To prove part (ii), assume
that g has a nondegenerate Killing form. Then the map ad : g → Der(g)
is injective because any ξ ∈ ker(ad) satisfies κ(ξ, η) = 0 for all η ∈ g. Now
choose a basis ξ1 , . . . , ξk of g. Since the Killing form is nondegenerate, there
exists a dual basis η1 , . . . , ηk ofP g such that κ(ξi , ηj ) = δij for all i and j.
With these basesPwe have ζ = i
κ(ηi , ζ)ξi for all ζ ∈ g. Now let δ ∈ Der(g)
and define ξ := i trace δ ad(ξi ) ηi ∈ g. Then, for all ζ ∈ g,

X


trace δ ad(ζ) = κ(ηi , ζ)trace δ ad(ξi ) = κ(ξ, ζ) = trace ad(ξ)ad(ζ) .
i

Thus the derivation ε := δ − ad(ξ) satisfies trace(ε ad(ζ)) = 0 for all ζ ∈ g.

Since [ε, ad(η)] = ad(εη), this implies




κ εη, ζ = trace ad(εη)ad(ζ) = trace [ε, ad(η)]ad(ξ)



= trace ε[ad(η), ad(ζ)] = trace ε ad([η, ζ]) = 0

for all η, ζ ∈ g. Hence ε = 0 by nondegeneracy of the Killing form and

hence δ = ad(ξ). This proves part (ii) and Lemma 5.4.15.
274 CHAPTER 5. CURVATURE

Lemma 5.4.16. Let g be a finite-dimensional Lie algebra with trivial center

that admits an invariant inner product and let δ : g → g be a linear map such
that, for all ξ, η, ζ ∈ g,
δ[[ξ, η], ζ] = [[δξ, η], ζ] + [[ξ, δη], ζ] + [[ξ, η], δζ]. (5.4.41)
Then there exists an element ξδ ∈ g such that δξ = [ξδ , ξ] for all ξ ∈ g.
Proof. Recall the definition of the space [g, g] ⊂ g in Exercise 5.2.23. Since g
admits an invariant inner product, we have [g, g]⊥ = Z(g), and since the
center of g is trivial, this implies
[g, g] = g. (5.4.42)
Now define the skew-symmetric bilinear map Bδ : g × g → g by
Bδ (ξ, η) := δ[ξ, η] − [δξ, η] − [ξ, δη] (5.4.43)
for ξ, η ∈ g. Then equation (5.4.41) can be expressed in the form
Bδ ([ξ, η], ζ) + [Bδ (ξ, η), ζ] = 0 (5.4.44)
for ξ, η, ζ ∈ g. By (5.4.42) there exists a basis of g consisting of vectors of
the form ei = [ξi , ηi ]. Define the linear map Aδ : g → g by
Aδ ei := −Bδ (ξi , ηi ).
Then Bδ (ei , ζ) = [Aδ ei , ζ] for all i and ζ by (5.4.44). Hence
Bδ (ξ, η) = [Aδ ξ, η] = [ξ, Aδ η] (5.4.45)
for all ξ, η ∈ g. Here the second equality holds by the skew-symmetry of Bδ .
By the Jacobi identity and equations (5.4.44) and (5.4.45) we have
2[Bδ (ξ, η), ζ] = [[ξ, Aδ η], ζ] + [[Aδ ξ, η], ζ]
= −[[ζ, ξ], Aδ η] − [[Aδ η, ζ], ξ] + [[Aδ ξ, η], ζ]
= −Bδ ([ζ, ξ], η) − [Bδ (η, ζ), ξ] + [[Aδ ξ, η], ζ]
= [Bδ (ζ, ξ), η] + Bδ ([η, ζ], ξ) + [[Aδ ξ, η], ζ]
= [[ζ, Aδ ξ], η] + [[η, ζ], Aδ ξ] + [[Aδ ξ, η], ζ]
=0
for all ξ, η, ζ ∈ g. Since the center of g is trivial, we find that Bδ (ξ, η) = 0
for all ξ, η ∈ g and so δ is a derivation. Hence δ is in the image of the
adjoint representation ad : g → Der(g) by Lemma 5.4.15 and this proves
Lemma 5.4.16.
5.4. THE GROUP OF ISOMETRIES* 275

Lemma 5.4.17. The assertion of Theorem 5.4.14 holds under the additional
assumption that the center of the Lie algebra g = Lie(G) is trivial.
Proof. Let φ ∈ I0 (G) and choose a smooth isotopy [0, 1] → I0 (G) : t 7→ φt
joining the identity φ0 = id to φ1 = φ. For 0 ≤ t ≤ 1 define the diffeomor-
phism ψt : G → G by ψt (g) := φt (1l)−1 φt (g). Then each ψt is an isometry
and the path [0, 1] → I0 (G) : t 7→ ψt satisfies

ψ0 = id, ψt (1l) = 1l

for all t. By Theorem 5.3.1 the derivatives Ψt := dψt (1l) : g → g preserve

the Riemann curvature tensor of G at 1l and by Example 5.2.17 (for the
standard metric on Lie subgroups of O(n)) and Exercise 5.2.21 (for general
bi-invariant Riemannian metrics) this translates into the condition

Ψt [[ξ, η], ζ] = [[Ψt ξ, Ψt η], Ψt ζ] (5.4.46)

for all t and all ξ, η, ζ ∈ g. Hence the endomorphism δt := Ψ−1 d

t dt Ψt : g → g
satisfies (5.4.41) for each t. Since g has a trivial center and is equipped with
an invariant inner product, Lemma 5.4.16 asserts that there exists a smooth
path [0, 1] → g : t 7→ ξt such that Ψ−1 d
t dt Ψt = ad(ξt ) for all t. Thus

d
Ψt = Ψt ad(ξt ), Ψ0 = 1l. (5.4.47)
dt
Now let [0, 1] → G : t 7→ bt be the solution of the differential equation
d
bt = bt ξt , b0 = 1l, (5.4.48)
dt
d
and define Φt := Ad(bt ) (Example 2.5.22). Then dt Φt = Φt ad(ξt ) for all t
and Φ0 = 1l. Thus Φt = Ψt and so

Ψt η = bt ηb−1
t (5.4.49)

for all t and η. Take t = 1, define b := b1 , and use Lemma 5.1.10 (uniqueness
of local isometries) to obtain ψ1 (g) = bgb−1 for all g ∈ G. Hence

φ(g) = φ(1l)ψ1 (g) = φ(1l)bgb−1 = agb−1

for all g ∈ G with a := φ(1l)b. This proves Lemma 5.4.17.

We will denote by Z0 (G) ⊂ Z(G) the connected component of 1l in the

center of G. Then Z0 (G) ⊂ G is a compact connected abelian Lie subgroup
of G (Example 2.5.32). Since G is connected, its Lie algebra is the center
of g, i.e. Lie(Z0 (G)) = Z(g).
276 CHAPTER 5. CURVATURE

Lemma 5.4.18. Let G ⊂ GL(n, R) be a compact connected Lie group,

equipped with a bi-invariant Riemannian metric, and let φ ∈ I0 (G). Then
φ(hg) = φ(gh) = φ(g)h = hφ(g) (5.4.50)
for all g ∈ G and all h ∈ Z0 (G).
Proof. The proof relies on the following basic observations.
(a) Z0 (G) is a compact connected Lie subgroup of G and Lie(Z0 (G)) = Z(g).
(b) exp(ξ + η) = exp(ξ) exp(η) for all ξ, η ∈ Z(g).
(c) The exponential map exp : Z(g) → Z0 (G) is surjective.
(d) Λ := {ξ ∈ Z(g) | exp(ξ) = 1l} is a discrete additive subgroup of Z(g)
which spans Z(g).
Part (a) was noted above, part (b) follows from Exercise 2.5.36 because the
Lie algebra Z(g) is commutative, and part (c) follows from the Hopf–Rinow
Theorem 4.6.6. That the set Λ is an additive subgroup of Z(g) follows
directly from (b). Moreover, by (a) and (b) the exponential map restricts to
a local diffeomorphism exp : Z(g) → Z0 (G). Hence Λ is discrete and by (c)
the exponential map descends to a proper map from Z(g)/Λ onto Z0 (G).
Since Z0 (G) is compact, the lattice Λ spans the vector space Z(g).
Now assume φ(1l) = 1l and define Φ := dφ(1l) : g → g. Then
φ(exp(ξ)) = exp(Φξ) (5.4.51)
for all ξ ∈ g by Corollary 5.3.3 and Example 5.2.17. Moreover, Φ is an
orthogonal transformation of g that preserves the Riemann curvature tensor
(Theorem 5.3.1). Thus |[Φξ, Φη]| = |[ξ, η]| for all ξ, η ∈ g by Example 5.2.17.
So, if ξ ∈ Z(g), then [ξ, η] = 0 for all η, hence [Φξ, Φη] = 0 for all η, and
hence Φξ ∈ Z(g). This shows that
ΦZ(g) = Z(g). (5.4.52)
By (5.4.51) and (5.4.52) we have ξ ∈ Λ if and only if φ(exp(ξ)) = 1l if and
only if exp(Φξ) = 1l if and only if Φξ ∈ Λ, so that ΦΛ = Λ. Since φ is isotopic
to the identity through isometries, by assumption, there exists a smooth
path of orthogonal transformations Φt of g from Φ0 = 1l to Φ1 = Φ that
satisfy Φt Λ = Λ for all t. Thus Φξ = ξ for all ξ ∈ Λ and so for all ξ ∈ Z(g)
because the lattice spans the subspace Z(g) by (d). Hence, by (5.4.51)
and (c), we obtain φ(h) = h for all h ∈ Z0 (G) whenever φ(1l) = 1l.
To prove equation (5.4.50) in general, fix an element g ∈ G and define the
diffeomorphism φg : G → G by φg (h) := φ(g)−1 φ(gh) for h ∈ G. Then φg is
an isometry and φg (1l) = 1l. Moreover, φg ∈ I0 (G), because G is connected.
Hence φg (h) = h for all h ∈ Z0 (G) and this proves Lemma 5.4.18.
5.4. THE GROUP OF ISOMETRIES* 277

Proof of Theorem 5.4.14. Let φ ∈ I0 (G). By Lemma 5.4.18, the isome-

try φ : G → G descends to a diffeomorphism

φ̄ : Ḡ → Ḡ, Ḡ := G/Z0 (G).

The Lie algebra of the quotient group Ḡ is the quotient space

ḡ := g/Z(g) ∼
= Z(g)⊥

and the invariant inner product on g restricts to an invariant inner product

on the orthogonal complement of Z(g). This defines a bi-invariant Rieman-
nian metric on Ḡ. The diffeomorphism φ̄ is an isometry with respect to this
metric, because
dφ(g) : Tg G → Tφ(g) G
is an orthogonal transformation which sends a tangent vector gη ∈ Tg G
with η ∈ Z(g) to the tangent vector

dφ(g)gη = φ(g)η ∈ Tφ(g) G

by (5.4.50), and hence it sends the subspace gZ(g)⊥ ⊂ Tg G to the sub-

space φ(g)Z(g)⊥ ⊂ Tφ(g) G. Apply Lemma 5.4.17 to the isometry φ̄ to obtain
elements a, b ∈ G whose equivalence classes ā, b̄ ∈ Ḡ satisfy

φ̄(ḡ) = āḡ b̄−1 for all ḡ ∈ Ḡ.

Then it follows from equation (5.4.50) in Lemma 5.4.18 that φ(g) = agb−1
for all g ∈ G. This proves Theorem 5.4.14.

Corollary 5.4.19. Let G be a compact connected Lie group equipped with

a bi-invariant Riemannian metric. Then the map

G × G → I0 (G) : (a, b) 7→ φa,b , φa,b (g) := agb−1 ,

descends to a Lie Group isomorphism

ρG : (G × G)/Z(G) → I0 (G). (5.4.53)

Proof. The map G × G → I0 (G) : (a, b) 7→ φa,b is a group homomorphism by

definition, is smooth by Theorem 5.4.1, and is surjective by Theorem 5.4.14.
Moreover, φa,b is the identity if and only if a = b ∈ Z(G). Hence the map ρG
in (5.4.53) is a Lie group isomorphism, with the smooth structure on the
quotient group (G × G)/Z(G) determined by Theorem 2.9.14. This proves
Corollary 5.4.19.
278 CHAPTER 5. CURVATURE

5.4.6 Examples and Exercises

Throughout we denote by I0 (M ) ⊂ I(M ) the connected component of the
identity in the group of isometries of a Riemannian manifold M .
Example 5.4.20. Theorem 5.4.14 establishes a one-to-one correspondence
between isometries φ ∈ I0 (G) that satisfy φ(1l) = 1l (the case a = b) and Lie
group automorphisms of G in the identity component. This correspondence
does not extend to other connected components. For example, in the case

G = Tn = Rn /Zn

the group of automorphisms of Tn is the infinite group

Aut(Tn ) = GL(n, Z)

of integer matrices with determinant ±1, while the group of isometries that
fix the origin is the finite subgroup

O(n, Z) ⊂ GL(n, Z)

of orthogonal integer matrices. Also, if G is not abelian, then the isome-

try g 7→ g −1 is not a group automorphism.
Exercise 5.4.21. Examine the case G = SU(2) (Example 2.5.20) and de-
duce that there exists a Lie group isomorphism

SO(4) ∼


= SU(2) × SU(2) /{±1l}.

Exercise 5.4.22. Let G be a compact Lie group equipped with a bi-

invariant Riemannian metric and let g := Lie(G). Show that the formula
 

(ζ, δ), (ζ 0 , δ 0 ) := δζ 0 − δ 0 ζ, δ, δ 0 + 41 ad ζ, ζ 0
    
(5.4.54)

for ζ, ζ 0 ∈ g and δ, δ 0 ∈ Der(g) defines a Lie bracket on g × Der(g). Show

that the map VectK (G) → g × Der(g) : X 7→ (X(1l), ∇X(1l)) identifies the
Lie algebra VectK (G) = Lie(I(G)) of Killing vector fields on G with the
Lie algebra g × Der(g). Show that the homomorphism (5.4.53) induces the
surjective Lie algebra homomorphism
 
g × g → g × Der(g) : (ξ, η) 7→ ξ − η, 21 ad(ξ + η) , (5.4.55)

whose kernel consist of all pairs (ξ, η) ∈ g × g that satisfy ξ = η ∈ Z(g).

Show that dim(Der(g)) = dim(g) − dim(Z(g)).
5.4. THE GROUP OF ISOMETRIES* 279

Some of the examples of isometry groups discussed below will be ex-

plained in more detail in Chapter 6.

Example 5.4.23. The isometry group of Rm is the group of all affine trans-
formations of Rm with orthogonal linear part (Exercise 5.1.4).

Example 5.4.24. We will see in §6.4.3 that the isometry group of the m-
sphere S m ⊂ Rm+1 is the group

I(S m ) = O(m + 1)

of orthogonal transformations of the ambient space (Example 6.4.16). In

the case m = 2 a theorem of Smale [54] asserts that the inclusion

O(3) = I(S 2 ) ,→ Diff(S 2 )

of the isometry group of the 2-sphere into the infinite-dimensional group of

all diffeomorphisms of the 2-sphere is a homotopy equivalence.

Example 5.4.25. In §6.4.4 we will introduce the hyperbolic space Hm . Its

isometry group is the group

I(Hm ) = O(m, 1)

of all linear transformations of Rm+1 that preserve the quadratic form Q

in (6.4.9) (Exercise 6.4.24). In the case m = 2 the identity component of
this group is isomorphic to the Lie group

PSL(2, C) = SL(2, C)/{±1l}.

Geometrically this can be understood by examining the upper half space

model of H2 (Exercise 6.4.26).

The preceding three examples are the constant sectional curvature man-
ifolds discussed in §6.2 and §6.4 below. In all three cases the isometry group
has the maximal dimension
m(m + 1)
dim(I(M )) =
2
and is diffeomorphic to the orthonormal frame bundle O(M ), so these exam-
ples satisfy the condition Ip = Gp in the notation of §5.4.1. We will prove in
Corollary 6.4.13 that a complete, connected, simply connected manifold M
satisfies Ip = Gp if and only if it has constant sectional curvature.
280 CHAPTER 5. CURVATURE

Exercise 5.4.26. Consider the incomplete 2-manifolds

M0 := R2 \ {(0, 0)}, M1 := R2 \ Z2 .

Prove that for i = 0, 1 every isometry of Mi extends to an isometry of R2 and

so I(Mi ) is a subgroup of the Lie group I(R2 ) of affine maps with orthogonal
linear part (Example 5.4.23). The isometry group of M0 is isomorphic to
the Lie group O(2). The isometry group of M1 is discrete and is an example
of a so-called wallpaper group (of which there are 17). The Lie algebra
of Killing vector fields in both cases has dimension 3. Which Killing vector
fields are not complete? Hint: A Killing vector field on a connected open
set M ⊂ R2 is a smooth map M → R2 : (x, y) 7→ (u(x, y), v(x, y)) that
satisfies the equations ∂x u = ∂y v = 0 and ∂y u + ∂x v = 0. Deduce that the
map (u, v) is affine and has a skew-symmetric linear part.

Example 5.4.27. The identity component of the isometry group of the

complex projective space CPn (Example 2.8.5) with the Fubini–Study metric
(Example 3.7.5) is the group

I0 (CPn ) = PSU(n + 1) = SU(n + 1)/Z(SU(n + 1)) ∼ = U(n + 1)/U(1).

Here Z(SU(n + 1)) = λ1l | λ ∈ S 1 , λn+1 = 1 ∼


= Z/(n + 1)Z is the center
of the group SU(n + 1). We emphasize that the dimension n(n + 2) of
the isometry group I(CPn ) is smaller than the dimension n(2n + 1) of the
orthonormal frame bundle O(CPn ) unless n = 0 or n = 1.
In the case n = 1 the projective line CP1 is isometric to the 2-sphere
by stereographic projection (Exercise 2.8.13 and Example 3.7.5). Hence the
identity component PSU(2) of the isometry group of CP1 is isomorphic to
the identity component SO(3) of the isometry group of S 2 . An explicit
isomorphism is discussed in Exercise 2.5.21.
The full isometry group I(CPn ) has two connected components. In the
case n = 2 it is an open question whether the inclusion I(CP2 ) ,→ Diff(CP2 )
of the isometry group of CP2 into the group of all diffeomorphisms of CP2 is
a homotopy equivalence. By deep results of Gromov [14] and Taubes [56] a
positive answer to this question is equivalent to the assertion that the space
of symplectic forms on CP2 in a fixed cohomology class is contractible (the
symplectic uniqueness conjecture for CP2 ). It is not even known whether this
space is connected or, equivalently, whether every diffeomorphism of CP2
that induces the identity on cohomology is isotopic to the identity. For a
more detailed discussion see [50, Example 3.4] and [36, Example 13.4.1].
5.5. CURVATURE IN LOCAL COORDINATES* 281

Example 5.4.28. The identity component I0 (S 2 ×S 2 ) of the isometry group

of the product manifold M = S 2 × S 2 is the product group SO(3) × SO(3).
In contrast to Smales’ Theorem the inclusion

I0 (S 2 × S 2 ) ,→ Diff 0 (S 2 × S 2 )

into the group of diffeomorphisms of S 2 ×S 2 that are isotopic to the identity

is not a homotopy equivalence. For example, if

φx,θ : S 2 → S 2

denotes the rotation about the axis through x ∈ S 2 by the angle θ ∈ R/2πZ
and the diffeomorphism ψθ : S 2 × S 2 → S 2 × S 2 is defined by

ψθ (x, y) := (x, φx,θ (y)), x, y ∈ S 2 ,

then the loop R/2πZ → Diff 0 (S 2 × S 2 ) : θ 7→ ψθ is not contractible and

neither is any of its iterates. Thus Diff 0 (S 2 ×S 2 ) has an infinite fundamental
group while the fundamental group of SO(3) × SO(3) is finite. For S 2 ×S 2 it
is an open question whether every diffeomorphism that induces the identity
on cohomology is isotopic to the identity. For a more detailed discussion
see [50, Example 3.5] and [36, Example 13.4.2].

5.5 Curvature in Local Coordinates*

Riemann
Let M ⊂ Rk be an m-dimensional manifold and let

φ = ψ −1 : U → Ω

be a local coordinate chart on an open set U ⊂ M with values in an open

set Ω ⊂ Rm . Define the vector fields E1 , . . . , Em along ψ by

∂ψ
Ei (x) := (x) ∈ Tψ(x) M.
∂xi
These vector fields form a basis of Tψ(x) M for every x ∈ Ω and the coeffi-
cients gij : Ω → R of the first fundamental form are

gij = hEi , Ej i .
282 CHAPTER 5. CURVATURE

Recall from Lemma 3.6.5 that the Christoffel Γkij : Ω → R are the coefficients
of the Levi-Civita connection, defined by
m
X
∇i Ej = Γkij Ek
k=1

and that they are given by the formula

m
X 1
Γkij := g k`


∂i gj` + ∂j gi` − ∂` gij .
2
`=1

` : Ω → R and R
Define the coefficients Rijk ijk` : Ω → R of the Riemann cur-
vature tensor by
m
X
`
R(Ei , Ej )Ek = Rijk E` (5.5.1)
`=1

and
m
X
ν
Rijk` := hR(Ei , Ej )Ek , E` i = Rijk gν` . (5.5.2)
ν=1

These coefficients are given by

m 
X 
`
Rijk = ∂i Γ`jk − ∂j Γ`ik + Γ`iν Γνjk − Γ`jν Γνik . (5.5.3)
ν=1

The coefficients of the Riemann curvature tensor have the symmetries

Rijk` = −Rjik` = −Rij`k = Rk`ij (5.5.4)

and the first Bianchi identity has the form

` ` `
Rijk + Rjki + Rkij = 0, Rijk` + Rjki` + Rkij` = 0. (5.5.5)

Warning: Care must be taken with the ordering of the indices. Some
`
authors use the notation Rkij `
for what we call Rijk and R`kij for what we
call Rijk` .

Exercise 5.5.1. Prove equations (5.5.3), (5.5.4), and (5.5.5). Use (5.5.3)
to give an alternative proof of Theorem 5.3.1.
5.5. CURVATURE IN LOCAL COORDINATES* 283

Gauß
If M ⊂ Rn is a 2-manifold (not necessarily embedded in R3 ), we can use
equation (5.3.9) as the definition of the Gaußian curvature K : M → R.
Let ψ : Ω → U be a local parametrization of an open set U ⊂ M defined on
an open set Ω ⊂ R2 . Denote the coordinates in R2 by (x, y) and define the
functions E, F, G : Ω → R by

E := |∂x ψ|2 , F := h∂x ψ, ∂y ψi, G := |∂y ψ|2 .

We abbreviate D := EG − F 2 . Then the composition of the Gaußian curva-

ture K : M → R with the parametrization ψ is given by

∂y F − 12 ∂x G
 
E F
1 1
K ◦ψ = det  F G 2 ∂y G

D2 1 1 1 2 1 2
2 ∂x E ∂x F − 2 ∂y E − 2 ∂y E + ∂x ∂y F − 2 ∂x G
1
 
E F 2 ∂ y E
1 1
− 2 det  F G 2 ∂x G 
D 1 1
∂y E 2 ∂x G 0
2 
1 ∂ E∂x G − F ∂y E
= − √ √
2 D ∂x E D
 
1 ∂ 2E∂x F − F ∂x E − E∂y E
+ √ √ .
2 D ∂y E D
This expression simplifies dramatically when F = 0 and we get
 
1 ∂ ∂x G ∂ ∂y E
K ◦ψ =− √ √ + √ (5.5.6)
2 EG ∂x EG ∂y EG
Exercise 5.5.2. Prove that the Riemannian metric
4
E=G= , F = 0,
(1 + x2 + y 2 )2

on R2 has constant constant curvature K = 1 and the Riemannian metric

4
E=G= , F = 0,
(1 − x2 − y 2 )2

on the open unit disc has constant curvature K = −1.

284 CHAPTER 5. CURVATURE
Chapter 6

Geometry and Topology

In this chapter we address what might be called the “fundamental problem

of intrinsic differential geometry”: when are two manifolds isometric? The
central tool for addressing this question is the Cartan–Ambrose–Hicks The-
orem (§6.1). In the subsequent sections we will use this result to examine
flat spaces (§6.2), symmetric spaces (6.3), and constant sectional curvature
manifolds (§6.4). The chapter then examines manifolds of nonpositive sec-
tional curvature and includes a proof of the Cartan Fixed Point Theorem
(§6.5). The last three sections introduce the Ricci tensor and show that com-
plete manifolds with positive Ricci tensor are compact (§6.6) and discuss the
scalar curvature (§6.7) and the Weyl tensor (§6.8).

6.1 The Cartan–Ambrose–Hicks Theorem

The Cartan–Ambrose–Hicks Theorem answers the question (at least locally)
when two manifolds are isometric. In general the equivalent conditions given
there are probably more difficult to verify in most examples than the con-
dition that there exist an isometry. However, under additional assumptions
it has many important consequences. The section starts with some basic
observations about homotopy and simple connectivity.

6.1.1 Homotopy
Definition 6.1.1. Let M be a manifold and let I = [a, b] be a compact
interval. A (smooth) homotopy of maps from I to M is a smooth map
γ : [0, 1] × I → M . We often write γλ (t) = γ(λ, t) for λ ∈ [0, 1] and t ∈ I
and call γ a (smooth) homotopy between γ0 and γ1 . We say the

285
286 CHAPTER 6. GEOMETRY AND TOPOLOGY

homotopy has fixed endpoints if γλ (a) = γ0 (a) and γλ (b) = γ0 (b) for all
λ ∈ [0, 1]. (See Figure 6.1.)
We remark that a homotopy and a variation are essentially the same
thing, namely a curve of maps (curves). The difference is pedagogical. We
used the word “variation” to describe a curve of maps through a given
map; when we use this word we are going to differentiate the curve to find
a tangent vector (field) to the given map. The word “homotopy” is used
to describe a curve joining two maps; it is a global rather than a local
(infinitesimal) concept.

γ1
M

γ0

Figure 6.1: A homotopy with fixed endpoints.

Definition 6.1.2. The manifold M is called simply connected iff for

any two curves γ0 , γ1 : [a, b] → M with γ0 (a) = γ1 (a) and γ0 (b) = γ1 (b) there
exists a homotopy from γ0 to γ1 with endpoints fixed. (The idea is that the
space Ωp,q of curves from p to q is connected.)
Remark 6.1.3. Two smooth maps γ0 , γ1 : [a, b] → M with the same end-
points can be joined by a continuous homotopy if and only if they can be
joined by a smooth homotopy. This follows from the Weierstrass approxi-
mation theorem.
Remark 6.1.4. The topological space Ωp,q of all smooth curves in M with
the endpoints p and q is connected for some pair of points p, q ∈ M if and
only if it is connected for every pair of points p, q ∈ M . (Prove this!)
Example 6.1.5. The Euclidean space Rm is simply connected, for any two
curves γ0 , γ1 : [a, b] → Rm with the same endpoints can be joined by the
homotopy γλ (t) := γ0 (t) + λ(γ1 (t) − γ0 (t)).
The punctured plane C \ {0} is not simply connected, for the curves

γn (t) := e2πint , 0 ≤ t ≤ 1,

are not homotopic with fixed endpoints for distinct n.

Exercise 6.1.6. Prove that the m-sphere S m is simply connected for m 6= 1.
6.1. THE CARTAN–AMBROSE–HICKS THEOREM 287

6.1.2 The Global C-A-H Theorem

0
Theorem 6.1.7 (Global C-A-H theorem). Let M ⊂ Rn and M 0 ⊂ Rn
be nonempty, connected, simply connected, complete m-manifolds. Fix two
elements p0 ∈ M and p00 ∈ M 0 and let Φ0 : Tp0 M → Tp00 M 0 be an orthogonal
linear isomorphism. Then the following are equivalent.
(i) There exists an isometry φ : M → M 0 satisfying
φ(p0 ) = p00 , dφ(p0 ) = Φ0 . (6.1.1)

(ii) If (Φ, γ, γ 0 ) is a development satisfying the initial condition

γ(0) = p0 , γ 0 (0) = p00 , Φ(0) = Φ0 , (6.1.2)
then
γ(1) = p0 =⇒ γ 0 (1) = p00 , Φ(1) = Φ0
(iii) If (Φ0 , γ0 , γ00 ) and (Φ1 , γ1 , γ10 ) are developments satisfying (6.1.2), then
γ0 (1) = γ1 (1) =⇒ γ00 (1) = γ10 (1).

(iv) If (Φ, γ, γ 0 ) is a development satisfying (6.1.2), then Φ∗ Rγ = Rγ0 0 .

Proof. See page 288.

p0
M’

Figure 6.2: Diagram for Example 6.1.8.

Example 6.1.8. Before giving the proof let us interpret the conditions in
case M and M 0 are two-dimensional spheres of radius r and r0 respectively in
three-dimensional Euclidean space R3 . Imagine that the spheres are tangent
at p0 = p00 . Clearly the spheres will be isometric exactly when r = r0 .
Condition (ii) says that if the spheres are rolled along one another without
sliding or twisting, then the endpoint γ 0 (1) of one curve of contact depends
only on the endpoint γ(1) of the other and not on the intervening curve γ(t).
By Theorem 5.3.10 the Riemann curvature of a 2-manifold at p is determined
by the Gaußian curvature K(p); and for spheres we have K(p) = 1/r2 .
288 CHAPTER 6. GEOMETRY AND TOPOLOGY

Exercise 6.1.9. Let γ be the closed curve which bounds an octant as shown
in the diagram for Example 6.1.8. Find γ 0 .

Exercise 6.1.10. Show that in case M is two-dimensional, the condition

Φ(1) = Φ0 in Theorem 6.1.7 may be dropped from (ii).

Proof of Theorem 6.1.7. We first prove a slightly different theorem. Namely,

we weaken condition (i) to assert that φ is a local isometry (i.e. not neces-
sarily bijective), and prove that this weaker condition is equivalent to (ii),
(iii), and (iv) whenever M is connected and simply connected and M 0 is
complete. Thus we drop the hypotheses that M be complete and M 0 be
connected and simply connected.
We prove that (i) implies (ii). Given a development as in (ii) we have,
by Corollary 5.3.4,

γ 0 (1) = φ(γ(1)) = φ(p0 ) = p00 , Φ(1) = dφ(γ(1)) = dφ(p0 ) = Φ0 ,

as required.
We prove that (ii) implies (iii) when M 0 is complete. Choose develop-
ments (Φi , γi , γi0 ) for i = 0, 1 as in (iii). Define a curve γ : [0, 1] → M by
“composition” 
γ0 (2t), 0 ≤ t ≤ 1/2,
γ(t) :=
γ1 (2 − 2t), 1/2 ≤ t ≤ 1,
so that γ is continuous and piecewise smooth and γ(1) = p0 . By Theo-
rem 3.5.21 there is a development (Φ, γ, γ 0 ) on the interval [0, 1] satisfy-
ing (6.1.2) (because M 0 is complete). Since γ(1) = p0 it follows from (ii)
that
γ 0 (1) = p00 , Φ(1) = Φ0 .
By the uniqueness of developments and the invariance under reparametriza-
tion, we have

(Φ0 (2t), γ0 (2t), γ00 (2t)),


0 0 ≤ t ≤ 1/2,
(Φ(t), γ(t), γ (t)) =
(Φ1 (2 − 2t), γ1 (2 − 2t), γ10 (2 − 2t)), 1/2 ≤ t ≤ 1.

Hence γ00 (1) = γ 0 (1/2) = γ10 (1) as required.

We prove that (iii) implies (i) when M 0 is complete and M is connected.
Define the map φ : M → M 0 as follows. Fix an element p ∈ M . Since M is
connected, there exists a smooth curve γ : [0, 1] → M such that γ(0) = p0
and γ(1) = p. Since M 0 is complete, there exists a development (Φ, γ, γ 0 )
with γ 0 (0) = p00 and Φ(0) = Φ0 (Theorem 3.5.21). Now define φ(p) := γ 0 (1).
6.1. THE CARTAN–AMBROSE–HICKS THEOREM 289

By (iii) the endpoint p0 := γ 0 (1) is independent of the choice of the curve γ,

and so φ is well-defined. We prove that this map φ satisfies the following
(a) If (Φ, γ, γ 0 ) is a development satisfying γ(0) = p0 , γ 0 (0) = p00 , Φ(0) = Φ0 ,
then φ(γ(t)) = γ 0 (t) for 0 ≤ t ≤ 1.
(b) If p, q ∈ M satisfy 0 < d(p, q) < inj(p; M ) and d(p, q) < inj(φ(p), M 0 ),
then d0 (φ(p), φ(q)) = d(p, q).
That φ satisfies (a) follows directly from the definition and the fact that the
triple (Φt , γt , γt0 ) defined by

Φt (s) := Φ(st), γt (s) := γ(st), γt0 (s) := γ 0 (st)

for 0 ≤ s ≤ 1 is a development.
To prove (b), choose a vector v ∈ Tp M such that

|v| = d(p, q) expp (v) = q

(Theorem 4.4.4) and let γ : [0, 1] → M be a smooth curve with

γ(0) = p0 , γ(t) = expp ((2t − 1)v)

for 21 ≤ t ≤ 1. Let (Φ, γ, γ 0 ) be the unique development of M along M 0

satisfying γ 0 (0) = p00 and Φ(0) = Φ0 (Theorem 3.5.21). Then

γ 0 ( 12 ) = φ(p), γ 0 (1) = φ(q)

by (a) and the restriction of γ 0 to the interval [ 12 , 1] is a geodesic by part (ii) of

Lemma 3.5.19 with X = γ̇. Hence γ 0 (t) = expφ(p) ((2t − 1)v 0 ) for 12 ≤ t ≤ 1,
where the tangent vector v 0 ∈ Tφ(p) M 0 is given by v 0 := γ̇ 0 ( 21 ) = Φ( 12 )v and
hence satisfies |v 0 | = |v| = d(p, q) < inj(φ(p), M 0 ). Thus it follows from The-
orem 4.4.4 that d0 (φ(p), φ(q)) = d0 (φ(p), exp0φ(p) (v 0 )) = |v 0 | = d(p, q) and this
proves (b). It follows from (b) and Theorem 5.1.1 that φ is a local isometry.
We prove that (i) implies (iv). Given a development as in (ii) we have

γ 0 (t) = φ(γ(t)), Φ(t) = dφ(γ(t))

for every t, by Corollary 5.3.4. Hence it follows from part (iv) of Corol-
lary 5.3.2 (Theorema Egregium for local isometries) that

Φ(t)∗ Rγ(t) = (φ∗ R)γ 0 (t) = Rγ0 0 (t)

for all t as required.

290 CHAPTER 6. GEOMETRY AND TOPOLOGY

We prove that (iv) implies (iii) when M 0 is complete and M is simply

connected. Choose developments (Φi , γi , γi0 ) for i = 0, 1 as in (iii). Since M
is simply connected there is a homotopy

[0, 1] × [0, 1] → M : (λ, t) 7→ γ(λ, t) = γλ (t)

from γ0 to γ1 with endpoints fixed. By Theorem 3.5.21 there is, for each λ,
a development (Φλ , γλ , γλ0 ) on the interval [0, 1] with initial conditions

γλ0 (0) = p00 , Φλ (0) = Φ0

(because M 0 is complete). The proof of Theorem 3.5.21 also shows that

γλ (t) and Φλ (t) depend smoothly on both t and λ. We must prove that

γ10 (1) = γ00 (1).

To see this we will show that, for each fixed t, the curve

λ 7→ (Φλ (t), γλ (t), γλ0 (t))

is a development; then by the definition of development we have that the

curve λ 7→ γλ0 (1) is smooth and

∂λ γλ0 (1) = Φλ (1)∂λ γλ (1) = 0

as required.
First choose a basis e1 , . . . , em of Tp0 M and extend it to obtain vector
fields Ei ∈ Vect(γ) along the homotopy γ by imposing the conditions that
the vector fields t 7→ Ei (λ, t) be parallel, i.e.

∇t Ei (λ, t) = 0, Ei (λ, 0) = ei . (6.1.3)

Then the vectors E1 (λ, t), . . . Em (λ, t) form a basis of Tγλ (t) M for all λ and t.
Second, define the vector fields Ei0 along γ 0 by

Ei0 (λ, t) := Φλ (t)Ei (λ, t) (6.1.4)

so that ∇0t Ei0 = 0. Third, define the functions ξ 1 , . . . , ξ m : [0, 1]2 → R by

m
X m
X
0
∂t γ =: i
ξ Ei , ∂t γ = ξ i Ei0 . (6.1.5)
i=1 i=1

Here the second equation follows from (6.1.4) and the fact that Φλ ∂t γ = ∂t γ 0 .
6.1. THE CARTAN–AMBROSE–HICKS THEOREM 291

Now consider the vector fields

X 0 := ∂λ γ 0 , Yi0 := ∇0λ Ei0 (6.1.6)
along γ 0 . They satisfy the equations
m
X m

X
∇0t X 0 = ∇0t ∂λ γ 0 = ∇0λ ∂t γ 0 = ∇0λ ξ i Ei0 = ∂λ ξ i Ei0 + ξ i Yi0

i=1 i=1

and
∇0t Yi0 = ∇0t ∇0λ Ei0 − ∇0λ ∇0t Ei0 = R0 (∂t γ 0 , ∂λ γ 0 )Ei0 .
To sum up we have X 0 (λ, 0) = Yi0 (λ, 0) = 0 and
m
X
∇0t X 0 = ∂λ ξ i Ei0 + ξ i Yi0 , ∇0t Yi0 = R0 (∂t γ 0 , ∂λ γ 0 )Ei0 .


(6.1.7)
i=1

On the other hand, the vector fields

X 0 := Φλ ∂λ γ, Yi0 := Φλ ∇λ Ei (6.1.8)
along γ 0 satisfy the same equations, namely
m
!
X
∇0t X 0 = Φλ ∇t ∂λ γ = Φλ ∇λ ∂t γ = Φλ ∇λ i
ξ Ei
i=1
m
X m
X
∂λ ξ i Ei + ξ i ∇λ Ei = ∂λ ξ i Ei0 + ξ i Yi0 ,



= Φλ
i=1 i=1
∇0t Yi0 = Φλ ∇t ∇λ Ei − ∇λ ∇t Ei = Φλ R(∂t γ, ∂λ γ)Ei

= R0 (Φλ ∂t γ, Φλ ∂λ γ)Φλ Ei = R0 (∂t γ 0 , X 0 )Ei0 .

Here the last but one equation follows from (iv).
Since the tuples (6.1.6) and (6.1.8) satisfy the same differential equa-
tion (6.1.7) and vanish at t = 0 they must agree. Hence
∂λ γ 0 = Φλ ∂λ γ, ∇0λ Ei0 = Φλ ∇λ Ei
for i = 1, . . . , m. This says that λ 7→ (Φλ (t), γλ (t), γλ0 (t)) is a development.
For t = 1 we obtain ∂λ γ 0 (λ, 1) = 0 as required.
Now the modified theorem (where φ is a local isometry) is proved. The
original theorem follows immediately. Condition (iv) is symmetric in M
and M 0 . Thus, if we assume (iv), there are local isometries φ : M → M 0
and ψ : M 0 → M satisfying φ(p0 ) = p00 , dφ(p0 ) = Φ0 and ψ(p00 ) = p0 ,
dψ(p0 ) = Φ−1 0 . But then ψ ◦ φ is a local isometry with ψ ◦ φ(p0 ) = p0 and
d(ψ ◦ φ)(p0 ) = id. Hence ψ ◦ φ is the identity. Similarly φ ◦ ψ is the identity
so φ is bijective (and ψ = φ−1 ) as required. This proves Theorem 6.1.7.
292 CHAPTER 6. GEOMETRY AND TOPOLOGY

Remark 6.1.11. The proof of Theorem 6.1.7 shows that the various im-
plications in the weak version of the theorem (where φ is only a local
isometry) require the following conditions on M and M 0 :
(i) always implies (ii), (iii), and (iv);
(ii) implies (iii) whenever M 0 is complete;
(iii) implies (i) whenever M 0 is complete and M is connected;
(iv) implies (iii) whenever M 0 is complete and M is simply connected.

Remark 6.1.12. The proof that (iii) implies (i) in Theorem 6.1.7 can be
slightly shortened by using the following observation. Let φ : M → M 0 be
a map between smooth manifolds. Assume that φ ◦ γ is smooth for every
smooth curve γ : [0, 1] → M . Then φ is smooth.

Corollary 6.1.13. Let M and M 0 be nonempty, connected, simply con-

nected, complete Riemannian manifolds and let φ : M → M 0 be a local
isometry. Then φ is bijective and hence is an isometry.

Proof. This follows by combining the weak and strong versions of the global
C-A-H Theorem 6.1.7. Fix a point p0 ∈ M and define

p00 := φ(p0 ), Φ0 := dφ(p0 ).

Then the tuple M, M 0 , p0 , p00 , Φ0 satisfies condition (i) of the weak version
of Theorem 6.1.7. Hence this tuple also satisfies condition (iv) of Theo-
rem 6.1.7. Since M and M 0 are connected, simply connected, and complete
we may apply the strong version of Theorem 6.1.7 to obtain an isometry

ψ : M → M0

satisfying
ψ(p0 ) = p00 , dψ(p0 ) = Φ0 .
Since every isometry is also a local isometry and M is connected it follows
from Lemma 5.1.10 that φ(p) = ψ(p) for all p ∈ M . Hence φ is an isometry,
as required.

Remark 6.1.14. Refining the argument in the proof of Corollary 6.1.13 one
can show that a local isometry φ : M → M 0 must be surjective whenever M
is complete and M 0 is connected. None of these assumptions can be removed.
(Take an isometric embedding of a disc in the plane or an embedding of
a complete space M into a space with two components, one of which is
isometric to M .)
6.1. THE CARTAN–AMBROSE–HICKS THEOREM 293

Likewise, one can show that a local isometry φ : M → M 0 must be

injective whenever M is complete and connected and M 0 is simply connected.
Again none of these asumptions can be removed. (Take a covering R → S 1 ,
or a covering of a disjoint union of two isometric complete simply connected
spaces onto one copy of this space, or some noninjective immersion of a disc
into the plane and choose the pullback metric on the disc.)

6.1.3 The Local C-A-H Theorem

Theorem 6.1.15 (Local C-A-H Theorem). Let M and M 0 be smooth
m-manifolds, let p0 ∈ M and p00 ∈ M 0 , and let Φ0 : Tp0 M → Tp00 M 0 be an
orthogonal linear isomorphism. Let r > 0 be smaller than the injectvity radii
of M at p0 and of M 0 at p00 and define

Ur0 := p0 ∈ M 0 | d0 (p00 , p0 ) < r .


Ur := {p ∈ M | d(p0 , p) < r} ,

Then the following are equivalent.

(i) There exists an isometry φ : Ur → Ur0 satisfying (6.1.1).
(ii) If (Φ, γ, γ 0 ) is a development on an interval I ⊂ R with 0 ∈ I, satisfying
the initial condition (6.1.2) as well as γ(I) ⊂ Ur and γ 0 (I) ⊂ Ur0 , then

γ(1) = p0 =⇒ γ 0 (1) = p00 , Φ(1) = Φ0 .

(iii) If (Φ0 , γ0 , γ00 ) and (Φ1 , γ1 , γ10 ) are developments as in (ii), then

γ0 (1) = γ1 (1) =⇒ γ00 (1) = γ10 (1).

(iv) If v ∈ Tp0 M with |v| < r and

γ(t) := expp0 (tv), γ 0 (t) := exp0p0 (tΦ0 v), Φ(t) := Φ0γ 0 (t, 0)Φ0 Φγ (0, t),
0

then Φ(t)∗ Rγ(t) = Rγ0 0 (t) for 0 ≤ t ≤ 1.

If these equivalent conditions are satisfied, then

φ(expp0 (v)) = exp0p0 (Φ0 v)

0

for all v ∈ Tp0 M with |v| < r.

Proof. See page 294.

294 CHAPTER 6. GEOMETRY AND TOPOLOGY

Lemma 6.1.16. Let p ∈ M and v, w ∈ Tp M such that |v| < inj(p). For
0 ≤ t ≤ 1 define

∂ 

γ(t) := exp(tv), X(t) := exp p t(v + λw) ∈ Tγ(t) M.
∂λ λ=0
Then

∇t ∇t X = R(γ̇, X)γ̇, X(0) = 0, ∇t X(0) = w. (6.1.9)

A vector field along γ satisfying the first equation in (6.1.9) is called a Ja-
cobi field along γ.
Proof. Define

γ(λ, t) := expp (t(v + λw)), X(λ, t) := ∂λ γ(λ, t)

for all λ and t. Since γ(λ, 0) = p for all λ we have X(λ, 0) = 0 and
d

∇t X(λ, 0) = ∇t ∂λ γ(λ, 0) = ∇λ ∂t γ(λ, 0) = v + λw = w.
dλ
Moreover, ∇t ∂t γ = 0 and hence

∇t ∇t X = ∇t ∇t ∂λ γ
= ∇t ∇λ ∂t γ − ∇λ ∇t ∂t γ
= R(∂t γ, ∂λ γ)∂t γ
= R(∂t γ, X)∂t γ.

This proves Lemma 6.1.16.

Proof of Theorem 6.1.15. The proofs (i) =⇒ (ii) =⇒ (iii) =⇒ (i) =⇒ (iv)
are as before; the reader might note that when L(γ) ≤ r we also have
L(γ 0 ) ≤ r for any development so that there are plenty of developments
with γ : [0, 1] → Ur and γ 0 : [0, 1] → Ur0 . The proof that (iv) implies (i) is
a little different since (iv) here is somewhat weaker than (iv) of the global
theorem: the equation Φ∗ R = R0 is only assumed for certain developments.
Hence assume (iv) and define φ : Ur → Ur0 by

φ := exp0p0 ◦Φ0 ◦ exp−1 0

p0 : Ur → Ur .
0

We must prove that φ is an isometry. Thus we fix a point q ∈ Ur and a

tangent vector u ∈ Tq M and choose v, w ∈ Tp M with |v| < r such that

expp0 (v) = q, d expp0 (v)w = u. (6.1.10)

6.1. THE CARTAN–AMBROSE–HICKS THEOREM 295

Define γ : [0, 1] → Ur , γ 0 : [0, 1] → Ur0 , X ∈ Vect(γ), and X 0 ∈ Vect(γ 0 ) by


∂ 
γ(t) = expp0 (tv), X(t) := expp0 (t(v + λw))
∂λ λ=0

∂ 
γ 0 (t) = exp0p0 (tΦ0 v), X 0 (t) := exp0p0 (t(Φ0 v + λΦ0 w)).
0 ∂λ λ=0 0

Then, by definition of φ, we have

γ 0 := φ ◦ γ, dφ(γ)X = X 0 . (6.1.11)

By Lemma 6.1.16, X is a solution of (6.1.9) and X 0 is a solution of

∇t ∇t X 0 = R0 (∂t γ 0 , X 0 )∂t γ 0 , X(λ, 0) = 0, ∇t X(λ, 0) = Φ0 w. (6.1.12)

Now define Φ(t) : Tγ(t) M → Tγ 0 (t) M 0 by

Φ(t) := Φ0γ 0 (t, 0)Φ0 Φγ (0, t).

Then Φ intertwines covariant differentiation. Since γ̇ and γ̇ 0 are parallel

vector fields with γ̇ 0 (0) = Φ0 v = Φ(0)γ̇(0), we have

Φ(t)γ̇(t) = γ̇ 0 (t)

for every t. Moreover, it follows from (iv) that Φ∗ Rγ = Rγ0 0 . Combining this
with (6.1.9) we obtain

∇0t ∇0t (ΦX) = Φ∇t ∇t X = R0 (Φγ̇, ΦX)Φγ̇ = R0 (γ̇ 0 , ΦX)γ̇ 0 .

Hence the vector field ΦX along γ 0 also satisfies the initial value prob-
lem (6.1.12) and thus
ΦX = X 0 = dφ(γ)X.
Here we have also used (6.1.11). Using (6.1.10) we find

γ(1) = expp0 (v) = q, X(1) = d expp0 (v)w = u,

and so
dφ(q)u = dφ(γ(1))X(1) = X 0 (1) = Φ(1)u.
Since Φ(1) : Tγ(1) M → Tγ 0 (1) M 0 is an orthogonal transformation this gives

|dφ(q)u| = |Φ(1)u| = |u| .

Hence φ is an isometry as claimed. This proves Theorem 6.1.15.

296 CHAPTER 6. GEOMETRY AND TOPOLOGY

6.2 Flat Spaces

Our aim in the next few sections is to give applications of the Cartan-
Ambrose-Hicks Theorem. It is clear that the hypothesis Φ∗ R = R0 for all
developments will be difficult to verify without drastic hypotheses on the
curvature. The most drastic such hypothesis is that the curvature vanishes
identically.

Definition 6.2.1. A Riemannian manifold M is called flat iff the Riemann

curvature tensor R vanishes identically.

Theorem 6.2.2. Let M ⊂ Rn be a smooth m-manifold.

(i) M is flat if and only if every point has a neighborhood which is isometric
to an open subset of Rm , i.e. at each point p ∈ M there exist local coordinates
x1 , . . . , xm such that the coordinate vectorfields Ei = ∂/∂xi are orthonormal.
(ii) Assume M is connected, simply connected, and complete. Then M is
flat if and only if there is an isometry φ : M → Rm onto Euclidean space.

Proof. Assertion (i) follows immediately from Theorem 6.1.15 and (ii) fol-
lows immediately from Theorem 6.1.7.

Exercise 6.2.3. Carry over the Cartan–Ambrose–Hicks theorem and The-

orem 6.2.2 to the intrinsic setting.

Exercise 6.2.4. A one-dimensional manifold is always flat.

Exercise 6.2.5. If M1 and M2 are flat, so is M = M1 × M2 .

Example 6.2.6. By Exercises 6.2.4 and 6.2.5 the standard torus

Tm = z = (z1 , . . . , zm ) ∈ Cm  |z1 | = · · · = |zm | = 1

 

is flat.

Exercise 6.2.7. For a, b > 0 and c ≥ 0 define M ⊂ C3 by

n  u vo
M := M (a, b, c) := (u, v, w) ∈ C3  |u| = a, |v| = b, w = c .

ab
Then M is diffeomorphic to a torus (a product of two circles) and M is flat. If
a0 , b0 > 0 and c0 ≥ 0, prove that there is an isometry φ from M 0 = M (a0 , b0 , c0 )
to M = M (a, b, c) if and only if the triples (a0 , b0 , c0 ) and (a, b, c) are related
by a permutation.
6.2. FLAT SPACES 297

Hint: Show first that an isometry φ : M 0 → M that satisfies the condi-

tion φ(a0 , b0 , c0 ) = (a, b, c) must have the form
 0 α  0 β  0 γ  0 δ  0 α+γ  0 β+δ !
0 0 0 u v u v u v
φ(u , v , w ) = a 0 0
,b 0 0
,c 0
a b a b a b0

for integers α, β, γ, δ that satisfy αδ − βγ = ±1. Show that this map φ is an

isometry if and only if
2 2
a0 + c0 = α2 a2 + γ 2 b2 + (α + γ)2 c2 ,
2
c0 = αβa2 + γδb2 + (α + γ)(β + δ)c2 ,
2 2
b0 + c0 = β 2 a2 + δ 2 b2 + (β + δ)2 c2 .
Exercise 6.2.8 (Developable manifolds). Let n = m + 1 and let E(t) be
a one-parameter family of hyperplanes in Rn . Then there exists a smooth
map u : R → Rn such that
E(t) = u(t)⊥ , |u(t)| = 1, (6.2.1)
for every t. We assume that u̇(t) 6= 0 for every t so that u(t) and u̇(t) are
linearly independent. Show that
L(t) := u(t)⊥ ∩ u̇(t)⊥ = lim E(t) ∩ E(s). (6.2.2)
s→t

Thus L(t) is a linear subspace of dimension m − 1. Now let γ : R → Rn be

a smooth map such that
hγ̇(t), u(t)i = 0, hγ̇(t), u̇(t)i =
6 0 (6.2.3)
for all t. This means that γ̇(t) ∈ E(t) and γ̇(t) ∈
/ L(t); thus E(t) is spanned
by L(t) and γ̇(t). For t ∈ R and ε > 0 define
L(t)ε := {v ∈ L(t) | |v| < ε} .
Let I ⊂ R be a bounded open interval such that the restriction of γ to the
closure of I is injective. Prove that, for ε > 0 sufficiently small, the set
[ 
M0 := γ(t) + L(t)ε
t∈I

is a smooth manifold of dimension m = n − 1. A manifold which arises this

way is called developable. Show that the tangent spaces of M0 are the
original subspaces E(t), i.e.
Tp M0 = E(t) for p ∈ γ(t) + L(t)ε .
298 CHAPTER 6. GEOMETRY AND TOPOLOGY

(One therefore calls M0 the “envelope” of the hyperplanes γ(t) + E(t).)

Show that M0 is flat (Hint: use Gauß–Codazzi). If (Φ, γ, γ 0 ) is a devel-
opment of M0 along Rm , show that the map φ : M0 → Rm , defined by

φ(γ(t) + v) := γ 0 (t) + Φ(t)v

for v ∈ L(t)ε , is an isometry onto an open set M00 ⊂ Rm . Thus a development

“unrolls” M0 onto the Euclidean space Rm . When n = 3 and m = 2 one
can visualize M0 as a twisted sheet of paper (see Figure 6.3).

Figure 6.3: Developable surfaces.

Remark 6.2.9. Given a codimension-1 submanifold

M ⊂ Rm+1

and a curve γ : R → M we may form the osculating developable M0 to M

along γ by taking
E(t) := Tγ(t) M.
This developable has common affine tangent spaces with M along γ as

Tγ(t) M0 = E(t) = Tγ(t) M

for every t. This gives a nice interpretation of parallel transport: M0 may be

unrolled onto a hyperplane where parallel transport has an obvious meaning
and the identification of the tangent spaces thereby defines parallel transport
in M . (See Remark 3.5.16.)

Exercise 6.2.10. Each of the following is a developable surface in R3 .

(i) A cone on a plane curve Γ ⊂ H, i.e.

M = {tp + (1 − t)q | t > 0, q ∈ Γ}

where H ⊂ R3 is an affine hyperplane, p ∈ R3 \H, and Γ ⊂ H is a 1-manifold.

6.2. FLAT SPACES 299

(ii) A cylinder on a plane curve Γ, i.e.

M = {q + tv | q ∈ Γ, t ∈ R}

where H and Γ are as in (i) and v is a fixed vector not parallel to H. (This
is a cone with the cone point p at infinity.)
(iii) The tangent developable to a space curve γ : R → R3 , i.e.

M = {γ(t) + sγ̇(t) | |t − t0 | < ε, 0 < s < ε} ,

where γ̇(t0 ) and γ̈(t0 ) are linearly independent and ε > 0 is sufficiently small.
(iv) The paper model of a Möbius strip (see Figure 6.3).

Figure 6.4: A circular one-sheeted hyperboloid.

Remark 6.2.11. A 2-dimensional submanifold M ⊂ R3 is called a ruled

surface iff there is a straight line in M through every point. Every devel-
opable surface is ruled, however, there are ruled surfaces that are not devel-
opable. An example is the manifold M = {γ(t) + sγ̈(t) | |t − t0 | < ε, |s| < ε}
where γ : R → R3 is a smooth curve with |γ̇| ≡ 1 and γ̈(t0 ) 6= 0, and ε > 0 is
sufficiently small; this surface is not developable in general. Other examples
are the elliptic hyperboloid of one sheet
 2
y2 z2
 
3 x

M := (x, y, z) ∈ R  2 + 2 − 2 = 1 (6.2.4)
a b c

depicted in Figure 6.4, the hyperbolic paraboloid

x2 y 2
  
3

M := (x, y, z) ∈ R  z = 2 − 2 . (6.2.5)
a b

(both with two straight lines through every point in M ), Plücker’s conoid
  
3 2
 2 2xy
M := (x, y, z) ∈ R  x + y 6= 0, z = 2 , (6.2.6)
x + y2
300 CHAPTER 6. GEOMETRY AND TOPOLOGY

the helicoid
(  )
 x + iy
3 iαz
M := (x, y, z) ∈ R  p =e , (6.2.7)
x2 + y 2
and the Möbius strip
    
 cos(s) cos(s/2) cos(s)
t  s ∈ R and
 
M :=  sin(s)  +  cos(s/2) sin(s)   . (6.2.8)
2  −1 < t < 1 
0 sin(s/2)


These five surfaces have negative Gaußian curvature. The Möbius strip
in (6.2.8) is not developable, while the paper model of the Möbius strip is.
The helicoid in (6.2.7) is a minimal surface, i.e. its mean curvature (the
trace of the second fundamental form) vanishes. A minimal surface which
is not ruled is the catenoid
M := {(x, y, z) ∈ R3 | x2 + y 2 = c2 cosh (z/c)}.
(Exercise: Prove all this.)

6.3 Symmetric Spaces

In the last section we applied the Cartan-Ambrose-Hicks Theorem in the flat
case; the hypothesis Φ∗ R = R0 was easy to verify since both sides vanish. To
find more general situations where we can verify this hypothesis note that
for any development (Φ, γ, γ 0 ) satisfying the initial conditions
γ(0) = p0 , γ 0 (0) = p00 , Φ(0) = Φ0 ,
we have
Φ(t) = Φ0γ 0 (t, 0)Φ0 Φγ (0, t)
so that the hypothesis Φ∗ R = R0 is certainly implied by the three hypotheses
Φγ (t, 0)∗ Rp0 = Rγ(t)
Φ0γ 0 (t, 0)∗ Rp0 0 = Rγ0 0 (t)
0

(Φ0 )∗ Rp0 = Rp0 0 .

0

The last hypothesis is a condition on the initial linear isomorphism

Φ0 : Tp0 M → Tp00 M 0
while the former hypotheses are conditions on M and M 0 respectively,
namely, that the Riemann curvature tensor is invariant by parallel trans-
port. It is rather amazing that this condition is equivalent to a simple
geometric condition as we now show.
6.3. SYMMETRIC SPACES 301

6.3.1 Symmetric Spaces

Definition 6.3.1. A Riemannian manifold M is called symmetric about
the point p ∈ M iff there exists a (necessarily unique) isometry

φ:M →M

satisfying
φ(p) = p, dφ(p) = −id; (6.3.1)
M is called a symmetric space iff it is symmetric about each of its points.
A Riemannian manifold M is called locally symmetric about the point
p ∈ M if, for r > 0 sufficiently small, there exists an isometry

φ : Ur (p, M ) → Ur (p, M ), Ur (p, M ) := {q ∈ M | d(p, q) < r} ,

satisfying (6.3.1); M is called a locally symmetric space iff it is locally

symmetric about each of its points.

Remark 6.3.2. The proof of Theorem 6.3.4 below will show that, if M is
locally symmetric, the isometry φ : Ur (p, M ) → Ur (p, M ) with φ(p) = p
and dφ(p) = −id exists whenever 0 < r ≤ inj(p).

Exercise 6.3.3. Every symmetric space is complete. Hint: If γ : I → M is

a geodesic and φ : M → M is a symmetry about the point γ(t0 ) for t0 ∈ I,
then φ(γ(t0 + t)) = γ(t0 − t) for all t ∈ R with t0 + t, t0 − t ∈ I.

Theorem 6.3.4. Let M ⊂ Rn be an m-dimensional submanifold. Then the

following are equivalent.
(i) M is locally symmetric.
(ii) The covariant derivative ∇R (defined below) vanishes identically, i.e.

(∇v R)p (v1 , v2 )w = 0

for all p ∈ M and v, v1 , v2 , w ∈ Tp M .

(iii) The curvature tensor R is invariant under parallel transport, i.e.

Φγ (t, s)∗ Rγ(s) = Rγ(t) (6.3.2)

for every smooth curve γ : R → M and all s, t ∈ R.

Proof. See page 305.

302 CHAPTER 6. GEOMETRY AND TOPOLOGY

Corollary 6.3.5. Let M and M 0 be locally symmetric spaces and fix two
points p0 ∈ M and p00 ∈ M 0 , and let Φ0 : Tp0 M → Tp00 M 0 be an orthogonal
linear isomorphism. Let r > 0 be less than the injectivity radius of M at p0
and the injectivity radius of M 0 at p00 . Then the following holds.
(i) There exists an isometry φ : Ur (p0 , M ) → Ur (p00 , M 0 ) with φ(p0 ) = p00
and dφ(p0 ) = Φ0 if and only if Φ0 intertwines R and R0 , i.e.
(Φ0 )∗ Rp0 = Rp0 0 . (6.3.3)
0

(ii) Assume M and M 0 are connected, simply connected, and complete.

Then there exists an isometry φ : M → M 0 with φ(p0 ) = p00 and dφ(p0 ) = Φ0
if and only if Φ0 satisfies (6.3.3).
Proof. In (i) and (ii) the “only if” statement follows from Theorem 5.3.1
(Theorema Egregium) with Φ0 := dφ(p0 ). To prove the “if” statement, let
(Φ, γ, γ 0 ) be a development satisfying γ(0) = p0 , γ 0 (0) = p00 , and Φ(0) = Φ0 .
Since R and R0 are invariant under parallel transport, by Theorem 6.3.4, it
follows from the discussion in the beginning of this section that Φ∗ R = R0 .
Hence assertion (i) follows from the local C-A-H Theorem 6.1.15 and (ii)
follows from the global C-A-H Theorem 6.1.7

Corollary 6.3.6. A connected, simply connected, complete, locally symmet-

ric space is symmetric.
Proof. Corollary 6.3.5 (ii) with M 0 = M , p00 = p0 , and Φ0 = −id.

Corollary 6.3.7. A connected symmetric space M is homogeneous; i.e.

given p, q ∈ M there exists an isometry φ : M → M with φ(p) = q.
Proof. If M is simply connected, the assertion follows from part (ii) of Corol-
lary 6.3.5 with M = M 0 , p0 = p, p00 = q, and Φ0 = Φγ (1, 0) : Tp M → Tq M ,
where γ : [0, 1] → M is a curve from p to q. If M is not simply connected,
we can argue as follows. There is an equivalence relation on M defined by
p∼q : ⇐⇒ ∃ isometry φ : M → M 3 φ(p) = q.
Let p, q ∈ M and suppose that d(p, q) < inj(p). By Theorem 4.4.4 there
is a unique shortest geodesic γ : [0, 1] → M connecting p to q. Since M is
symmetric there is an isometry φ : M → M such that φ(γ(1/2)) = γ(1/2)
and dφ(γ(1/2)) = −id. This isometry satisfies φ(γ(t)) = γ(1 − t) and hence
φ(p) = q. Thus p ∼ q whenever d(p, q) < inj(p). This shows that each
equivalence class is open, hence each equivalence class is also closed, and
hence there is only one equivalence class because M is connected. This
proves Corollary 6.3.7.
6.3. SYMMETRIC SPACES 303

6.3.2 Covariant Derivative of the Curvature

For two vector spaces V, W and an integer k ≥ 1 we denote by Lk (V, W )
the vector space of all multi-linear maps from V k = V × · · · × V to W .
Thus L1 (V, W ) = L(V, W ) is the space of all linear maps from V to W .
Definition 6.3.8. The covariant derivative of the Riemann curvature
tensor assigns to every p ∈ M a linear map

(∇R)p : Tp M → L2 (Tp M, L(Tp M, Tp M ))

such that


(∇R)(X)(X1 , X2 )Y = ∇X R(X1 , X2 )Y − R(∇X X1 , X2 )Y
(6.3.4)
− R(X1 , ∇X X2 )Y − R(X1 , X2 )∇X Y

for all X, X1 , X2 , Y ∈ Vect(M ). We also use the notation

(∇v R)p := (∇R)p (v)

for p ∈ M and v ∈ Tp M so that

(∇X R)(X1 , X2 )Y := (∇R)(X)(X1 , X2 )Y

for all X, X1 , X2 , Y ∈ Vect(M ).

Remark 6.3.9. One verifies easily that the map

Vect(M )4 → Vect(M ) : (X, X1 , X2 , Y ) 7→ (∇X R)(X1 , X2 )Y,

defined by the right hand side of equation (6.3.4), is multi-linear over the
ring of functions F (M ). Hence it follows as in Remark 5.2.12 that ∇R is
well defined, i.e. that the right hand side of (6.3.4) at p ∈ M depends only
on the tangent vectors X(p), X1 (p), X2 (p), Y (p).
Remark 6.3.10. Let γ : I → M be a smooth curve on an interval I ⊂ R
and
X1 , X2 , Y ∈ Vect(γ)
be smooth vector fields along γ. Then equation (6.3.4) continues to hold
with X replaced by γ̇ and each ∇X on the right hand side replaced by the
covariant derivative of the respective vector field along γ:

(∇γ̇ R)(X1 , X2 )Y = ∇(R(X1 , X2 )Y ) − R(∇X1 , X2 )Y

(6.3.5)
− R(X1 , ∇X2 )Y − R(X1 , X2 )∇Y.
304 CHAPTER 6. GEOMETRY AND TOPOLOGY

Theorem 6.3.11. (i) If γ : R → M is a smooth curve such that γ(0) = p

and γ̇(0) = v, then

d 
(∇v R)p =  Φγ (0, t)∗ Rγ(t) (6.3.6)
dt t=0

(ii) The covariant derivative of the Riemann curvature tensor satisfies the
second Bianchi identity
(∇X R)(Y, Z) + (∇Y R)(Z, X) + (∇Z R)(X, Y ) = 0. (6.3.7)
Proof. We prove (i). Let v1 , v2 , w ∈ Tp M and choose parallel vector fields
X1 , X2 , Y ∈ Vect(γ) along γ satisfying the initial conditions X1 (0) = v1 ,
X2 (0) = v2 , Y (0) = w. Thus
X1 (t) = Φγ (t, 0)v1 , X2 (t) = Φγ (t, 0)v2 , Y (t) = Φγ (t, 0)w.
Then the last three terms on the right vanish in equation (6.3.5) and hence
(∇v R)(v1 , v2 )w = ∇(R(X1 , X2 )Y )(0)

d 
= Φ R (X1 (t), X2 (t))Y (t)
dt t=0 γ(0,t) γ(t)

d 
= Φ R (Φγ (t, 0)v1 , Φγ (t, 0)v2 )Φγ (t, 0)w
dt t=0 γ(0,t) γ(t)

d 

= Φγ (0, t)∗ Rγ(t) (v1 , v2 )w.
dt t=0

Here the second equation follows from Theorem 3.3.6. This proves (i).
We prove (ii). Choose a smooth function γ : R3 → M and denote by
(r, s, t) the coordinates on R3 . If Y is a vector field along γ, we have


(∇∂r γ R)(∂s γ, ∂t γ)Y = ∇r R(∂s γ, ∂t γ)Y − R(∂s γ, ∂t γ)∇r Y
−R(∇r ∂s γ, ∂t γ)Y − R(∂s γ, ∇r ∂t γ)Y
= ∇r (∇s ∇t Y − ∇t ∇s Y ) − (∇s ∇t − ∇t ∇s ) ∇r Y
+R(∂t γ, ∇r ∂s γ)Y − R(∂s γ, ∇t ∂r γ)Y.
Permuting the variables r, s, t cyclically and taking the sum of the resulting
three equations we obtain
(∇∂r γ R)(∂s γ, ∂t γ)Y + (∇∂s γ R)(∂t γ, ∂r γ)Y + (∇∂t γ R)(∂r γ, ∂s γ)Y
= ∇r (∇s ∇t Y − ∇t ∇s Y ) − (∇s ∇t − ∇t ∇s ) ∇r Y
+∇s (∇t ∇r Y − ∇r ∇t Y ) − (∇t ∇r − ∇r ∇t ) ∇s Y
+∇t (∇r ∇s Y − ∇s ∇r Y ) − (∇r ∇s − ∇s ∇r ) ∇t Y.
The terms on the right cancel out. This proves Theorem 6.3.11.
6.3. SYMMETRIC SPACES 305

Proof of Theorem 6.3.4. We prove that (iii) implies (i). This follows from
the local Cartan–Ambrose–Hicks Theorem 6.1.15 with
p00 = p0 = p, Φ0 = −id : Tp M → Tp M.
This isomorphism satisfies
(Φ0 )∗ Rp = Rp .
Hence it follows from the discussion in the beginning of this section that
Φ∗ R = R 0
for every development (Φ, γ, γ 0 ) of M along itself satisfying
γ(0) = γ 0 (0) = p, Φ(0) = −id.
Hence, by the local C-A-H Theorem 6.1.15, there is an isometry
φ : Ur (p, M ) → Ur (p, M )
satisfying
φ(p) = p, dφ(p) = −id
whenever 0 < r < inj(p; M ).
We prove that (i) implies (ii). By Theorem 5.3.1 (Theorema Egregium),
every isometry φ : M → M 0 preserves the Riemann curvature tensor and
covariant differentiation, and hence also the covariant derivative of the Rie-
mann curvature tensor, i.e.
φ∗ (∇R) = ∇0 R0 .
Applying this to the local isometry φ : Ur (p, M ) → Ur (p, M ) we obtain
∇dφ(p)v R φ(p) (dφ(p)v1 , dφ(p)v2 ) = dφ(p) (∇v R) (v1 , v2 )dφ(p)−1

for all v, v1 , v2 ∈ Tp M . Since

dφ(p) = −id
this shows that ∇R vanishes at p.
We prove that (ii) imlies (iii). If ∇R vanishes, then equation (6.3.6) in
Theorem 6.3.11 shows that the function
s 7→ Φγ (t, s)∗ Rγ(s) = Φγ (t, 0)∗ Φγ (0, s)∗ Rγ(s)
is constant and hence is everywhere equal to Rγ(t) . This implies (6.3.2) and
completes the proof of Theorem 6.3.4.
306 CHAPTER 6. GEOMETRY AND TOPOLOGY

Covariant Derivative of the Curvature in Local Coordinates

Let φ : U → Ω be a local coordinate chart on M with values in an open set
Ω ⊂ Rm , denote its inverse by ψ := φ−1 : Ω → U, and let
∂ψ
Ei (x) := (x) ∈ Tψ(x) M, x ∈ Ω, i = 1, . . . , m,
∂xi
be the local frame of the tangent bundle determined by this coordinate
chart. Let Γkij : Ω → R denote the Christoffel symbols and Rijk
` : Ω → R the

coefficients of the Riemann curvature tensor so that

X X
∇i Ej = Γkij Ek , R(Ei , Ej )Ek = `
Rijk E` .
k `

Given i, j, k, ` ∈ {1, . . . , m} we can express the vector field (∇Ei R)(Ej , Ek )E`
along ψ for each x ∈ Ω as a linear combination of the basis vectors Ei (x).
This gives rise to functions
ν
∇i Rjk` :Ω→R

defined by X
ν
(∇Ei R)(Ej , Ek )E` =: ∇i Rjk` Eν . (6.3.8)
ν
These functions are given by
µ
X
ν ν
∇i Rjk` = ∂i Rjk` + Γνiµ Rjk`
µ
(6.3.9)
Γµij Rµk` Γµik Rjµ` Γµi` Rjkµ
X X X
ν ν ν
− − − .
µ µ µ

The second Bianchi identity has the form

ν ν ν
∇i Rjk` + ∇j Rki` + ∇k Rij` = 0. (6.3.10)

Exercise: Prove equations (6.3.9) and (6.3.10). Warning: As in §5.5,

care must be taken with the ordering of the indices. Some authors use the
ν for what we call ∇ Rν .
notation ∇i R`jk i jk`

6.3.3 Examples and Exercises

Example 6.3.12. Every flat manifold is locally symmetric.

Example 6.3.13. If M1 and M2 are (locally) symmetric, so is M1 × M2 .

6.3. SYMMETRIC SPACES 307

Example 6.3.14. M = Rm with the standard metric is a symmetric space.

Recall that the isometry group I(Rm ) consists of all affine transformations
of the form
φ(x) = Ax + b, A ∈ O(m), b ∈ Rm .
(See Exercise 5.1.4.) The isometry with fixed point p ∈ Rm and dφ(p) = −id
is given by φ(x) = 2p − x for x ∈ Rm .

Example 6.3.15. The flat tori of Exercise 6.2.7 in the previous section are
symmetric (but not simply connected). This shows that the hypothesis of
simple connectivity cannot be dropped in part (ii) of Corollary 6.3.5.

Example 6.3.16. Below we define manifolds of constant curvature and

show that they are locally symmetric. The simplest
example, after a flat
space, is the unit sphere S m = x ∈ Rm+1 | |x| = 1 . The symmetry φ of


the sphere about a point p ∈ M is given by

φ(x) := −x + 2hp, xip

for x ∈ S m . This extends to an orthogonal linear transformation of the

ambient space. In fact the group of isometries of S m is the group O(m + 1)
of orthogonal linear transformations of Rm+1 (see Example 6.4.16 below).
In accordance with Corollary 6.3.7 this group acts transitively on S m .

Example 6.3.17. A compact two-dimensional manifold of constant neg-

ative curvature is locally symmetric (as its universal cover is symmetric)
but not homogeneous (as closed geodesics of a given period are isolated).
Hence it is not symmetric. This shows that the hypothesis that M be simply
connected cannot be dropped in Corollary 6.3.6.

Example 6.3.18. The real projective space RPn with the metric inherited
from S n is a symmetric space and the orthogonal group O(n+1) acts on it by
isometries. The complex projective space CPn with the Fubini–Study metric
in Example 3.7.5 is a symmetric space and the unitary group U(n + 1) acts
on it by isometries. The complex Grassmannian Gk (Cn ) in Example 3.7.6
is a symmetric space and the unitary group U(n) acts on it by isometries.
(Exercise: Prove this.)

Example 6.3.19. The simplest example of a symmetric  space which is not

of constant curvature is the orthogonal group O(n) = g ∈ Rn×n | g T g = 1l

with the Riemannian metric (5.2.19) of Example 5.2.17. The symmetry φ

about the point a ∈ O(n) is given by φ(g) = ag −1 a. This discussion extends
to every Lie subgroup G ⊂ O(n). (Exercise: Prove this.)
308 CHAPTER 6. GEOMETRY AND TOPOLOGY

6.4 Constant Curvature

In the §5.3 we saw that the Gaußian curvature of a two-dimensional surface
is intrinsic: we gave a formula for it in terms of the Riemann curvature
tensor and the first fundamental form. We may use this formula to define the
Gaußian curvature for any two-dimensional manifold (even if its codimension
is greater than one). We make a slightly more general definition.

6.4.1 Sectional Curvature

Definition 6.4.1. Let M ⊂ Rn be a smooth m-dimensional submanifold.
Let p ∈ M and let E ⊂ Tp M be a 2-dimensional linear subspace of the tan-
gent space. The sectional curvature of M at (p, E) is the number

hRp (u, v)v, ui

K(p, E) = , (6.4.1)
|u|2 |v|2 − hu, vi2

where u, v ∈ E are linearly independent (and hence form a basis of E).

The right hand side of (6.4.1) remains unchanged if we multiply u or v

by a nonzero real number or add to one of the vectors a real multiple of the
other; hence it depends only on the linear subspace spanned by u ad v.

Example 6.4.2. If M ⊂ R3 is a 2-manifold, then by Theorem 5.3.10 the

sectional curvature K(p, Tp M ) = K(p) is the Gaußian curvature of M at p.
More generally, for any 2-manifold M ⊂ Rn (whether or not it has codimen-
sion one) we define the Gaußian curvature of M at p by

K(p) := K(p, Tp M ). (6.4.2)

Example 6.4.3. If M ⊂ Rm+1 is a submanifold of codimension one and

ν : M → S m is a Gauß map, then the sectional curvature of a 2-dimensional
subspace E ⊂ Tp M spanned by two linearly independent tangent vectors
u, v ∈ Tp M is given by

hu, dν(p)uihv, dν(p)vi − hu, dν(p)vi2

K(p, E) = . (6.4.3)
|u|2 |v|2 − hu, vi2

This follows from equation (5.3.11) in the proof of Theorem 5.3.10 which
holds in all dimensions. In particular, when M = S m , we have ν(p) = p
and hence K(p, E) = 1 for all p and E. For a sphere of radius r we have
ν(p) = p/r and hence K(p, E) = 1/r2 .
6.4. CONSTANT CURVATURE 309

Example 6.4.4. Let G ⊂ O(n) be a Lie subgroup equipped with the Rie-
mannian metric
hv, wi := trace(v T w)
for v, w ∈ Tg G ⊂ Rn×n . Then, by Example 5.2.17, the sectional curvature
of G at the identity matrix 1l is given by
1
|[ξ, η]|2
K(1l, E) =
4
for every 2-dimensional linear subspace E ⊂ g = Lie(G) = T1l G with an
orthonormal basis ξ, η.
Exercise 6.4.5. Let E ⊂ Tp M be a 2-dimensional linear subspace, let r > 0
be smaller than the injectivity radius of M at p, and let N ⊂ M be the 2-
dimensional submanifold given by
N := expp ({v ∈ E | |v| < r}) .
Show that the sectional curvature K(p, E) of M at (p, E) agrees with the
Gaußian curvature of N at p.
Exercise 6.4.6. Let p ∈ M ⊂ Rn and let E ⊂ Tp M be a 2-dimensional lin-
ear subspace. For r > 0 let L denote the ball of radius r in the (n − m + 2)-
dimensional affine subspace of Rn through p and parallel to the vector sub-
space E + Tp M ⊥ :
n o
L = p + v + w | v ∈ E, w ∈ Tp M ⊥ , |v|2 + |w|2 < r2 .

Show that, for r sufficiently small, L ∩ M is a 2-dimensional manifold with

Gaußian curvature KL∩M (p) at p given by KL∩M (p) = K(p, E).

6.4.2 Constant Sectional Curvature

Definition 6.4.7. Let k ∈ R and m ≥ 2 be an integer. An m-manifold
M ⊂ Rn is said to have constant sectional curvature k iff K(p, E) = k
for every p ∈ M and every 2-dimensional linear subspace E ⊂ Tp M .
Theorem 6.4.8. Let M ⊂ Rn be an m-manifold and fix an element p ∈ M
and a real number k. Then the following are equivalent.
(i) K(p, E) = k for every 2-dimensional linear subspace E ⊂ Tp M .
(ii) The Riemann curvature tensor of M at p is given by
 
hRp (v1 , v2 )v3 , v4 i = k hv1 , v4 ihv2 , v3 i − hv1 , v3 ihv2 , v4 i (6.4.4)

for all v1 , v2 , v3 , v4 ∈ Tp M .
310 CHAPTER 6. GEOMETRY AND TOPOLOGY

Proof. That (ii) implies (i) follows directly from the definition of the sec-
tional curvature in (6.4.1) by taking v1 = v4 = u and v2 = v3 = v in (6.4.4).
Conversely, assume (i) and define the multi-linear map Q : Tp M 4 → R by
 
Q(v1 , v2 , v3 , v4 ) := hRp (v1 , v2 )v3 , v4 i − k hv1 , v4 ihv2 , v3 i − hv1 , v3 ihv2 , v4 i .

Then, for all u, v, v1 , v2 , v3 , v4 ∈ Tp M , the map Q satisfies the equations

Q(v1 , v2 , v3 , v4 ) + Q(v2 , v1 , v3 , v4 ) = 0, (6.4.5)
Q(v1 , v2 , v3 , v4 ) + Q(v2 , v3 , v1 , v4 ) + Q(v3 , v1 , v2 , v4 ) = 0, (6.4.6)
Q(v1 , v2 , v3 , v4 ) − Q(v3 , v4 , v1 , v2 ) = 0, (6.4.7)
Q(u, v, u, v) = 0. (6.4.8)
Here the first three equations follow from Theorem 5.2.13 and the last follows
from the definition of Q and the hypothesis that the sectional curvature
is K(p, E) = k for every 2-dimensional linear subspace E ⊂ Tp M .
We must prove that Q vanishes. Using (6.4.7) and (6.4.8) we find
0 = Q(u, v1 + v2 , u, v1 + v2 )
= Q(u, v1 , u, v2 ) + Q(u, v2 , u, v1 )
= 2Q(u, v1 , u, v2 )
for all u, v1 , v2 ∈ Tp M . This implies

0 = Q(u1 + u2 , v1 , u1 + u2 , v2 )
= Q(u1 , v1 , u2 , v2 ) + Q(u2 , v1 , u1 , v2 )
for all u1 , u2 , v1 , v2 ∈ T pM . Hence
Q(v1 , v2 , v3 , v4 ) = −Q(v3 , v2 , v1 , v4 )
= Q(v2 , v3 , v1 , v4 )
= −Q(v3 , v1 , v2 , v4 ) − Q(v1 , v2 , v3 , v4 ).
Here the second equation follows from (6.4.5) and the last from (6.4.6). Thus
1 1
Q(v1 , v2 , v3 , v4 ) = − Q(v3 , v1 , v2 , v4 ) = Q(v1 , v3 , v2 , v4 )
2 2
for all v1 , v2 , v3 , v4 ∈ Tp M and, repeating this argument,
1
Q(v1 , v2 , v3 , v4 ) = Q(v1 , v2 , v3 , v4 ).
4
Hence Q ≡ 0 as claimed. This proves Theorem 6.4.8.
6.4. CONSTANT CURVATURE 311

Remark 6.4.9. The symmetric group S4 on four symbols acts naturally

on the space L4 (Tp M, R) of multi-linear maps from Tp M 4 to R. The condi-
tions (6.4.5), (6.4.6), (6.4.7), and (6.4.8) say that the four elements

a = id + (12)
c = id + (123) + (132)
b = id − (34)
d = id + (13) + (24) + (13)(24)

of the group ring of S4 annihilate Q. This suggests an alternate proof of

Theorem 6.4.8. A representation of a finite group is completely reducible
so one can prove that Q = 0 by showing that any vector in any irreducible
representation of S4 which is annihilated by the four elements a, b, c and
d must necessarily be zero. This can be checked case by case for each
irreducible representation. (The group S4 has 5 irreducible representations:
two of dimension 1, two of dimension 3, and one of dimension 2.)

If M and M 0 are two m-dimensional manifolds with constant curvature k,

then every orthogonal isomorphism Φ : Tp M → Tp0 M 0 intertwines the Rie-
mann curvature tensors by Theorem 6.4.8. Hence by the appropriate version
(local or global) of the C-A-H Theorem we have the following corollaries.

Corollary 6.4.10. Every Riemannian manifold with constant sectional cur-

vature is locally symmetric.

Proof. Theorem 6.3.4 and Theorem 6.4.8.

Corollary 6.4.11. Let M and M 0 be m-dimensional Riemannian manifolds

with constant curvature k and let p ∈ M and p0 ∈ M 0 . If r > 0 is smaller
than the injectivity radii of M at p and of M 0 at p0 , then for every orthogonal
isomorphism
Φ : Tp M → Tp0 M 0
there exists an isometry

φ : Ur (p, M ) → Ur (p0 , M 0 )

such that
φ(p) = p0 , dφ(p) = Φ.

Proof. This follows from Corollary 6.3.5 and Corollary 6.4.10. Alternatively
one can use Theorem 6.4.8 and the local C-A-H Theorem 6.1.15.
312 CHAPTER 6. GEOMETRY AND TOPOLOGY

Corollary 6.4.12. Any two connected, simply connected, complete Rieman-

nian manifolds with the same constant sectional curvature and the same
dimension are isometric.

Proof. Theorem 6.4.8 and the global C-A-H Theorem 6.1.7.

Corollary 6.4.13. Let M ⊂ Rn be a connected, simply connected, complete

manifold. Then the following are equivalent.
(i) M has constant sectional curvature.
(ii) For every pair of points p, q ∈ M and every orthogonal linear isomor-
phism Φ : Tp M → Tq M there exists an isometry φ : M → M such that

φ(p) = q, dφ(p) = Φ.

Proof. That (i) implies (ii) follows immediately from Theorem 6.4.8 and the
global C-A-H Theorem 6.1.7. Conversely assume (ii). Then, for every pair
of points p, q ∈ M and every orthogonal linear isomorphism

Φ : Tp M → Tq M,

it follows from Theorem 5.3.1 (Theorema Egregium) that

Φ∗ Rp = Rq

and so
K(p, E) = K(q, ΦE)
for every 2-dimensional linear subspace E ⊂ Tp M . Since, for every pair of
points p, q ∈ M and of 2-dimensional linear subspaces

E ⊂ Tp M, F ⊂ Tq M,

we can find an orthogonal linear isomorphism Φ : Tp M → Tq M such that

ΦE = F,

this implies (i).

Corollary 6.4.13 asserts that a connected, simply connected, complete

Riemannian m-manifold M has constant sectional curvature if and only if
the isometry group I(M ) acts transitively on its orthonormal frame bun-
dle O(M ). Note that, by Lemma 5.1.10, this group action is also free.
6.4. CONSTANT CURVATURE 313

6.4.3 Examples and Exercises

Example 6.4.14. Any flat Riemannian manifold has constant sectional
curvature k = 0.

Example 6.4.15. The manifold

M = Rm

with its standard metric is, up to isometry, the unique connected, simply
connected, complete Riemannian m-manifold with constant sectional curva-
ture
k = 0.

Example 6.4.16. For m ≥ 2 the unit sphere

M = Sm

with its standard metric is, up to isometry, the unique connected, simply
connected, complete Riemannian m-manifold with constant sectional curva-
ture
k = 1.
Hence, by Corollary 6.4.12, every connected simply connected, complete
Riemannian manifold with positive sectional curvature k = 1 is compact.
Moreover, by Corollary 6.4.13, the isometry group I(S m ) is isomorphic to
the group O(m + 1) of orthogonal linear transformations of Rm+1 . Thus,
by Corollary 6.4.13, the orthonormal frame bundle O(S m ) is diffeomorphic
to O(m + 1). This follows also from the fact that, if

v1 , . . . , v m

is an orthonormal basis of Tp S m = p⊥ then

p, v1 , . . . , vm

is an orthonormal basis of Rm+1 .

Example 6.4.17. A product of spheres is not a space of constant sectional

curvature, but it is a symmetric space. Exercise: Prove this.

Example 6.4.18. For n ≥ 4 the orthogonal group O(n) is not a space of

constant sectional curvature, but it is a symmetric space and has nonnegative
sectional curvature (see Example 6.4.4).
314 CHAPTER 6. GEOMETRY AND TOPOLOGY

6.4.4 Hyperbolic Space

Fix an integer m ≥ 2. The hyperbolic space Hm is, up to isometry, the
unique connected, simply connected, complete Riemannian m-manifold with
constant sectional curvature k = −1. A model for Hm can be constructed
as follows. A point in Rm+1 will be denoted by

p = (x0 , x), x0 ∈ R, x = (x1 , . . . , xm ) ∈ Rm .

Let Q : Rm+1 × Rm+1 → R denote the symmetric bilinear form given by

Q(p, q) := −x0 y0 + x1 y1 + · · · + xm ym (6.4.9)

for p = (x0 , x), q = (y0 , y) ∈ Rm+1 . Since Q is nondegenerate the space

Hm := p = (x0 , x) ∈ Rm+1 | Q(p, p) = −1, x0 > 0



is a smooth m-dimensional submanifold of Rm+1 and the tangent space

of Hm at p is given by

Tp Hm = v ∈ Rm+1 | Q(p, v) = 0 .


For p = (x0 , x) ∈ Rm+1 and v = (ξ0 , ξ) ∈ Rm+1 we have

p
p ∈ Hm ⇐⇒ x0 = 1 + |x|2 ,
hξ, xi
v ∈ Tp Hm ⇐⇒ ξ0 = p .
1 + |x|2

Now let us define a Riemannian metric on Hm by

gp (v, w) := Q(v, w)
= hξ, ηi − ξ0 η0
(6.4.10)
hξ, xihη, xi
= hξ, ηi −
1 + |x|2

for v = (ξ0 , ξ) ∈ Tp Hm and w = (η0 , η) ∈ Tp Hm .

Theorem 6.4.19. Hm is a connected, simply connected, complete Rieman-

nian m-manifold with constant sectional curvature k = −1.

Proof. See page 315.

6.4. CONSTANT CURVATURE 315

We remark that the manifold Hm does not quite fit into the extrinsic
framework of most of this book as it is not exhibited as a submanifold
of Euclidean space but rather of “pseudo-Euclidean space”: the positive
definite inner product hv, wi of the ambient space Rm+1 is replaced by a
nondegenerate symmetric bilinear form Q(v, w). However, all the theory
developed thus far goes through (reading Q(v, w) for hv, wi) provided we
make the additional hypothesis (true in the example M = Hm ) that the
first fundamental form gp = Q|Tp M is positive definite. For then Q|Tp M is
nondegenerate and we may define the orthogonal projection Π(p) onto Tp M
as before. The next lemma summarizes the basic observations; the proof is
an exercise in linear algebra.
Lemma 6.4.20. Let Q be a symmetric bilinear form on a vector space V
and for each subspace E of V define its orthogonal complement by

E ⊥Q := {u ∈ V | Q(u, v) = 0 ∀v ∈ E} .

Assume Q is nondegenerate, i.e. V ⊥Q = {0}. Then, for every linear sub-

space E ⊂ V , we have

V = E ⊕ E ⊥Q ⇐⇒ E ∩ E ⊥Q = {0},

i.e. E ⊥Q is a vector space complement of E if and only if the restriction

of Q to E is nondegenerate.
Proof of Theorem 6.4.19. The proofs of the various properties of Hm are
entirely analogous to the corresponding proofs for S m . Thus the unit normal
field to Hm is given by
ν(p) = p
for p ∈ Hm although the “square of its length” is

Q(p, p) = −1.

For p ∈ Hm we introduce the Q-orthogonal projection Π(p) of Rm+1 onto

the subspace Tp Hm . It is characterized by the conditions

Π(p)2 = Π(p), ker Π(p) ⊥Q imΠ(p), imΠ(p) = Tp Hm ,

and is given by the explicit formula

Π(p)v = v + Q(v, p)p

for v ∈ Rm+1 .
316 CHAPTER 6. GEOMETRY AND TOPOLOGY

The covariant derivative of a vector field X ∈ Vect(γ) along a smooth

curve γ : R → Hm is given by

∇X(t) = Π(γ(t))Ẋ(t)
= Ẋ(t) + Q(Ẋ(t), γ(t))γ(t)
= Ẋ(t) − Q(X(t), γ̇(t))γ(t).

The last identity follows by differentiating the equation Q(X, γ) ≡ 0. This

can be interpreted as the hyperbolic Gauß–Weingarten formula as follows.
For p ∈ Hm and u ∈ Tp Hm we introduce, as before, the second fundamental
form
hp (u) : Tp Hm → (Tp Hm )⊥Q
via

hp (u)v := dΠ(p)u v
and denote its Q-adjoint by

hp (u)∗ : (Tp Hm )⊥Q → Tp Hm .

For every p ∈ Rm+1 we have


  d

dΠ(p)u v =  v + Q(v, p + tu)(p + tu)
dt t=0
= Q(v, p)u + Q(v, u)p,

where the first summand on the right is tangent to Hm and the second
summand is Q-orthogonal to Tp Hm . Hence

hp (u)v = Q(v, u)p, hp (u)∗ w = Q(w, p)u (6.4.11)

for v ∈ Tp Hm and w ∈ (Tp Hm )⊥Q .

With this understood, the Gauß-Weingarten formula

Ẋ = ∇X + hγ (γ̇)X

extends to the present setting. The reader may verify that the operators

∇ : Vect(γ) → Vect(γ)

thus defined satisfy the axioms of Theorem 3.7.8 and hence define the Levi-
Civita connection on Hm .
6.4. CONSTANT CURVATURE 317

Now a smooth curve γ : I → Hm is a geodesic if and only if it satisfies

the equivalent conditions

∇γ̇ ≡ 0 ⇐⇒ γ̈(t) ⊥Q Tγ(t) Hm ∀ t ∈ I ⇐⇒ γ̈ = Q(γ̇, γ̇)γ.

A geodesic must satisfy the equation

d
Q(γ̇, γ̇) = 2Q(γ̈, γ̇) = 0
dt
because γ̈ is a scalar multiple of γ, and so Q(γ̇, γ̇) is constant. Let p ∈ Hm
and v ∈ Tp Hm be given with Q(v, v) = 1. Then the geodesic γ : R → Hm
with γ(0) = p and γ̇(0) = v is given by

γ(t) = cosh(t)p + sinh(t)v, (6.4.12)

where
et + e−t et − e−t
cosh(t) := , sinh(t) := .
2 2
In fact we have γ̈(t) = γ(t) ⊥Q Tγ(t) Hm . It follows that the geodesics
exist for all time and hence Hm is geodesically complete. Moreover, being
diffeomorphic to Euclidean space, Hm is connected and simply connected.
It remains to prove that Hm has constant sectional curvature k = −1.
To see this we use the Gauß–Codazzi formula in the hyperbolic setting, i.e.

Rp (u, v) = hp (u)∗ hp (v) − hp (v)∗ hp (u). (6.4.13)

By equation (6.4.11), this gives

hRp (u, v)v, ui = Q(hp (u)u, hp (v)v) − Q(hp (v)u, hp (u)v)

= Q(Q(u, u)p, Q(v, v)p) − Q(Q(u, v)p, Q(u, v)p)
= −Q(u, u)Q(v, v) + Q(u, v)2
= −gp (u, u)gp (v, v) + gp (u, v)2

for all u, v ∈ Tp Hm . Hence, for every p ∈ M and every 2-dimensional linear

subspace E ⊂ Tp M with a basis u, v ∈ E we have

hRp (u, v)v, ui

K(p, E) = = −1.
gp (u, u)gp (v, v) − gp (u, v)2

This proves Theorem 6.4.19.

318 CHAPTER 6. GEOMETRY AND TOPOLOGY

Exercise 6.4.21. Prove that the pullback of the metric on Hm under the
diffeomorphism p 
Rm → Hm : x 7→ 1 + |x|2 , x
is given by s
hx, ξi2
|ξ|x = |ξ|2 −
1 + |x|2
or, equivalently, by the metric tensor,
xi xj
gij (x) = δij − (6.4.14)
1 + |x|2
for x = (x1 , . . . , xm ) ∈ Rm .
Exercise 6.4.22. The Poincaré model of hyperbolic space is the open
unit disc Dm ⊂ Rm equipped with the Poincaré metric
2 |η|
|η|y =
1 − |y|2
for y ∈ Dm and η ∈ Rm = Ty Dm . Thus the metric tensor is given by
4δij
gij (y) =  2 , y ∈ Dm . (6.4.15)
2
1 − |y|

Prove that the diffeomorphism

!
1 + |y|2 2y
Dm → Hm : y 7→ 2,
1 − |y| 1 − |y|2
is an isometry with the inverse
x
Hm → Dm : (x0 , x) 7→ .
1 + x0
Interpret this map as a stereographic projection from the south pole (−1, 0).
Exercise 6.4.23. The composition of the isometries in Exercise 6.4.21 and
Exercise 6.4.22 is the diffeomorphism Rm → Dm : x 7→ y given by
x 2y p 1 + |y|2
y=p , x= , 1 + |x|2 = .
1 + |x|2 + 1 1 − |y|2 1 − |y|2
Prove that this is an isometry intertwining the Riemannian metrics (6.4.14)
and (6.4.15). Find a formula for the geodesics in the Poincaré disc Dm .
Hint: Use Exercise 6.4.25 below.
6.4. CONSTANT CURVATURE 319

Exercise 6.4.24. Prove that the isometry group of Hm is the pseudo-ortho-

gonal group
  
m
 Q(gv, gw) = Q(v, w)
I(H ) = O(m, 1) := g ∈ GL(m + 1)   .
for all v, w ∈ Rm+1
Thus, by Corollary 6.4.13, the orthonormal frame bundle O(Hm ) is diffeo-
morphic to O(m, 1).
Exercise 6.4.25. Prove that the exponential map
expp : Tp Hm → Hm
is given by
p 
p  sinh Q(v, v)
expp (v) = cosh Q(v, v) p + p v (6.4.16)
Q(v, v)
for v ∈ Tp Hm = p⊥Q . Prove that this map is a diffeomorphism for every
p ∈ Hm . Thus any two points in Hm are connected by a unique geodesic.
Prove that the intrinsic distance function on hyperbolic space is given by
d(p, q) = cosh−1 (Q(p, q)) (6.4.17)
for p, q ∈ Hm . Compare this with Example 4.3.11.
Exercise 6.4.26. In the case m = 2 the Poincaré model of H2 in Exer-
cise 6.4.22 is the open unit disc D ⊂ C in the complex plane. It can be
identified with the upper half plane
H = {z ∈ C | Im(z) > 0}
via the diffeomorphism
i−z
H → D : z 7→ .
i+z
Show that the pullback of the Poincaré metric on D under this diffeomor-
phism is the Riemannian metric on H given by
|b
z|
|b
z |z =
y
for z = x + iy ∈ H and zb ∈ Tz H = C. Show that the isometries of H (in the
identity component) have the form
 
az + b a b
φ(z) = , ∈ SL(2, R),
cz + d c d
and deduce that the Lie group PSL(2, R) := SL(2, R)/{±1l} is isomorphic to
the identity component of O(2, 1).
320 CHAPTER 6. GEOMETRY AND TOPOLOGY

6.5 Nonpositive Sectional Curvature

In the previous section we have seen that any two points in a connected,
simply connected, complete manifold M of constant negative curvature are
joined by a unique geodesic (Exercise 6.4.25). Thus the entire manifold M
is geodesically convex and its injectivity radius is infinity. This continues
to hold in much greater generality for manifolds with nonpositive sectional
curvature. It is convenient, at this point, to extend the discussion to Rie-
mannian manifolds in the intrinsic setting. In particular, at some point in
the proof of the main theorem of this section and in our main example, we
shall work with a Riemannian metric that does not arise (in any obvious
way) from an embedding.
Definition 6.5.1. A Riemannian manifold M is said to have nonpos-
itive sectional curvature iff K(p, E) ≤ 0 for every p ∈ M and every 2-
dimensional linear subspace E ⊂ Tp M or, equivalently, hRp (u, v)v, ui ≤ 0 for
all p ∈ M and all u, v ∈ Tp M . A nonempty, connected, simply connected,
complete Riemannan manifold with nonpositive sectional curvature is called
a Hadamard manifold.

6.5.1 The Cartan–Hadamard Theorem

The next theorem shows that every Hadamard manifold is diffeomorphic to
Euclidean space and has infinite injectivity radius. This is in sharp contrast
to positive curvature manifolds as the example M = S m shows.
Theorem 6.5.2 (Cartan–Hadamard). Let M be a connected, simply con-
nected, complete Riemannan manifold. Then the following are equivalent.
(i) M has nonpositive sectional curvature.
(ii) The derivative of each exponential map is length increasing, i.e.
 
v  ≥ |b
d expp (v)b v|
for all p ∈ M and all v, vb ∈ Tp M .
(iii) Each exponential map is distance increasing, i.e.
d(expp (v0 ), expp (v1 )) ≥ |v0 − v1 |
for all p ∈ M and all v0 , v1 ∈ Tp M .
Moreover, if these equivalent conditions are satisfied, then the exponential
map expp : Tp M → M is a diffeomorphism for every p ∈ M . Thus any two
points in M are joined by a unique geodesic.
Proof. See page 321.
6.5. NONPOSITIVE SECTIONAL CURVATURE 321

Exercise 6.5.3. Let ξ : [0, ∞) → Rn be a smooth function such that

ξ(0) = 0, ˙
ξ(0) 6= 0, ξ(t) 6= 0 ∀ t > 0.

Prove that the function f : [0, ∞) → R given by f (t) := |ξ(t)| is smooth.

Hint: The function η : [0, ∞) → Rn defined by
 −1
t ξ(t), for t > 0,
η(t) := ˙
ξ(0), for t = 0,

is smooth. Show that f is differentiable and f˙ = |η|−1 hη, ξi.

˙

Exercise 6.5.4. Let ξ : R → Rn be a smooth function such that

ξ(0) = 0, ¨ = 0.
ξ(0)

Prove that there exist constants ε > 0 and c > 0 such that, for all t ∈ R,

|t| < ε =⇒ |ξ(t)|2 |ξ(t)|

˙ 2 − hξ(t), ξ(t)i
˙ 2
≤ c |t|6 .

Hint: Write
ξ(t) = tv + η(t), ˙ = v + η̇(t)
ξ(t)
with η(t) = O(t3 ) and η̇(t) = O(t2 ). Show that the terms of order 2 and 4
cancel in the Taylor expansion at t = 0.
Proof of Theorem 6.5.2. We prove that (i) implies (ii). Fix a point p ∈ M
and two tangent vectors v, vb ∈ Tp M . Assume without loss of generality
that vb 6= 0 and define the curve γ : R → M and the vector field X ∈ Vect(γ)
along γ by

∂ 
γ(t) := expp (tv), X(t) := v )) ∈ Tγ(t) M
expp (t(v + sb (6.5.1)
∂s s=0
for t ∈ R. Then

X(0) = 0, X(t) = d expp (tv)tb

v, ∇X(0) = vb 6= 0. (6.5.2)

To see this, define the map β : R2 → M by β(s, t) := expp (t(v + sb v )). It

satisfies β(0, t) = γ(t), ∂s β(0, t) = X(t), β(s, 0) = p, and ∂t β(s, 0) = v + sbv
for all s, t ∈ R. Hence ∇X(0) = ∇t ∂s β(0, 0) = ∇s ∂t β(0, 0) = vb. Moreover,
the curve β(s, ·) is a geodesic for every s, and hence Lemma 6.1.16 asserts
that X = ∂s β(0, ·) is a Jacobi field along γ, i.e.

∇∇X = R(γ̇, X)γ̇. (6.5.3)

322 CHAPTER 6. GEOMETRY AND TOPOLOGY

It follows from Exercise 6.5.3 with ξ(t) := Φγ (0, t)X(t) that the function

[0, ∞) → R : t 7→ |X(t)|

is smooth and 
d 
|X(t)| = |∇X(0)| = |b
v| .
dt t=0
Moreover, for t > 0, we have
d2 d hX, ∇Xi
2
|X| =
dt dt |X|
|∇X|2 + hX, ∇∇Xi hX, ∇Xi2
= −
|X| |X|3 (6.5.4)
|X|2 |∇X|2 − hX, ∇Xi2 hX, R(γ̇, X)γ̇i
= +
|X|3 |X|
≥ 0.
Here the third equality follows from the fact that X is a Jacobi field along γ,
and the inequality follows from the nonpositive sectional curvature condition
in (i) and from the Cauchy–Schwarz inequality. Thus the second derivative
of the function [0, ∞) → R : t 7→ |X(t)| − t |b v | is nonnegative; so its first
derivative is nondecreasing and it vanishes at t = 0; thus

|X(t)| − t |b
v| ≥ 0

for every t ≥ 0. In particular, for t = 1 we obtain

 
v  = |X(1)| ≥ |b
d expp (v)b v| .

as claimed. Thus we have proved that (i) implies (ii).

We prove that (ii) implies (i). Assume, by contradiction, that (ii) holds
but there exists a point p ∈ M and a pair of vectors v, vb ∈ Tp M such that

hRp (v, vb)v, vbi < 0. (6.5.5)

Define γ : R → M and X ∈ Vect(γ) by (6.5.1) so that (6.5.2) and (6.5.3)

are satisfied. Thus X is a Jacobi field with

X(0) = 0, ∇X(0) = vb 6= 0.

Hence it follows from Exercise 6.5.4 with ξ(t) := Φγ (0, t)X(t) that there is a
constant c > 0 such that, for t > 0 sufficiently small, we have the inequality

|X(t)|2 |∇X(t)|2 − hX(t), ∇X(t)i2 ≤ ct6 .

6.5. NONPOSITIVE SECTIONAL CURVATURE 323

Moreover, it follows from (6.5.1) and (6.5.2) that

X(t)
lim γ̇(t) = v, lim = vb.
t&0 t&0 t

Hence, by (6.5.5) there exist constants δ > 0 and ε > 0 such that

|X(t)| ≥ δt, hX(t), R(γ̇(t), X(t))γ̇(t)i ≤ −εt2 ,

for t > 0 sufficiently small. By (6.5.4) this implies

d2 |X|2 |∇X|2 − hX, ∇Xi2 hX, R(γ̇, X)γ̇i ct3 εt

|X| = + ≤ − .
dt2 |X|3 |X| δ3 δ

Integrate this inequality over an interval [0, t] with ct2 < εδ 2 to obtain

d d 
|X(t)| <  |X(t)| = |∇X(0)|
dt dt t=0

Integrating this inequality again gives |X(t)| < t |∇X(0)| for small positive t.
Hence it follows from (6.5.2) that
 
v  = |X(t)| < t |∇X(0)| = t |b
d expp (tv)tb v|

for t > 0 sufficiently small. This contradicts (ii).

We prove that (ii) implies that the exponential map expp : Tp M → M is
a diffeomorphism for every p ∈ M . By (ii) expp is a local diffeomorphism, i.e.
its derivative d expp (v) : Tp M → Texpp (v) M is bijective for every v ∈ Tp M .
Hence we can define a Riemannian metric on M 0 := Tp M by pulling back
the metric on M under the exponential map. To make this more explicit we
choose a basis e1 , . . . , em of Tp M and define the map

ψ : Rm → M

by
m
!
X
i
ψ(x) := expp x ei
i=1

for x = (x1 , . . . , xm ) ∈ Rm .
Define the metric tensor by
 
∂ψ ∂ψ
gij (x) := (x), j (x) , i, j = 1, . . . , m.
∂xi ∂x
324 CHAPTER 6. GEOMETRY AND TOPOLOGY

Then (Rm , g) is a Riemannian manifold (covered by a single coordinate

chart) and ψ : (Rm , g) → M is a local isometry, by definition of g. The
manifold (Rm , g) is clearly connected and simply connected. Moreover, for
every tangent vector ξ = (ξ 1 , . . . , ξ n ) ∈ Rm = T0 Rm , the curve

R → Rm : t 7→ tξ

is a geodesic with respect to g (because ψ is a local isometry and the image of

the curve under ψ is a geodesic in M ). Hence it follows from Theorem 4.6.5
that (Rm , g) is complete. Since both (Rm , g) and M are connected, simply
connected, and complete, it follows from Corollary 6.1.13 that the local
isometry ψ is bijective. Thus the exponential map

expp : Tp M → M

is a diffeomorphism as claimed. It follows that any two points in M are

joined by a unique geodesic.
We prove that (ii) implies (iii). Fix a point p ∈ M and two tangent
vectors v0 , v1 ∈ Tp M . Let γ : [0, 1] → M be the geodesic with endpoints

γ(0) = expp (v0 ), γ(1) = expp (v1 )

and let v : [0, 1] → Tp M be the unique curve satisfying

expp (v(t)) = γ(t)

for all t. Then

v(0) = v0 , v(1) = v1 ,
and

d(expp (v0 ), expp (v1 )) = L(γ)

Z 1
 
= d expp (v(t))v̇(t) dt
0
Z 1
≥ |v̇(t)| dt
Z0 1 
 
≥  v̇(t) dt
0
= |v1 − v0 | .

Here the third inequality follows from (ii). This shows that (ii) implies (iii).
6.5. NONPOSITIVE SECTIONAL CURVATURE 325

We prove that (iii) implies (ii). Fix a point p ∈ M and a tangent

vector v ∈ Tp M and denote

q := expp (v).

By (iii) the exponential map expq : Tq M → M is injective and, since M

is complete, it is bijective (see Theorem 4.6.6). Hence there is a unique
geodesic from q to any other point in M and therefore, by Theorem 4.4.4,
we have
|w| = d(q, expq (w)) (6.5.6)

for every w ∈ Tq M . Now define

φ := exp−1
q ◦ expp : Tp M → Tq M.

This map satisfies

φ(v) = 0.

Moreover, it is differentiable in a neighborhood of v and, by the chain rule,

dφ(v) = d expp (v) : Tp M → Tq M.

Now choose w := φ(v + vb) with vb ∈ Tp M . Then

expq (w) = expq (φ(v + vb)) = expp (v + vb)

and hence, by (6.5.6),

|φ(v + vb))| = |w| = d(q, expq (w)) = d(expp (v), expp (v + vb)) ≥ |b
v| .

Here the last inequality follows from (iii). This gives

 
v  = |dφ(v)b
d expp (v)b v|
|φ(v + tb
v )|
= lim
t→0 t
|tb
v|
≥ lim
t→0 t
= |b
v| .

Thus we have proved that (iii) implies (ii) and this completes the proof of
Theorem 6.5.2.
326 CHAPTER 6. GEOMETRY AND TOPOLOGY

6.5.2 Cartan’s Fixed Point Theorem

Recall from Definition 6.5.1 that a Hadamard manifold is a nonempty, con-
nected, simply connected, complete Riemannian manifold of nonpositive
sectional curvature.
Theorem 6.5.5 (Cartan). Let M be a Hadamard manifold and let G be a
compact topological group that acts on M by isometries. Then there exists
a point p ∈ M such that gp = p for every g ∈ G.
Proof. The proof has three steps and follows the argument given by Bill
Casselmann in [10]. The second step is Serre’s uniqueness result for the
circumcentre of a bounded set in a semi-hyperbolic space.
Step 1. Let m ∈ M and v ∈ Tm M and define
p0 := expm (−v), p1 := expm (v).
Then
d(p0 , p1 )2
2d(m, q)2 + ≤ d(p0 , q)2 + d(p1 , q)2
2
for every q ∈ M .
By Theorem 6.5.2 the exponential map expm : Tm M → M is a diffeomor-
phism. Hence d(p0 , p1 ) = 2|v|. Now let q ∈ M . Then there is a unique
tangent vector w ∈ Tm M such that
q = expm (w), d(m, q) = |w|.
Since the exponential map is expanding, by Theorem 6.5.2, we have
d(p0 , q) ≥ |w + v|, d(p1 , q) ≥ |w − v|.
Hence
d(m, q)2 = |w|2
|w + v|2 + |w − v|2
= − |v|2
2
d(p0 , q)2 + d(p1 , q)2 d(p0 , p1 )2
≤ − .
2 4
This proves Step 1.
Step 2. For p ∈ M and r ≥ 0 denote by B(p, r) ⊂ M the closed ball of
radius r centered at p. Let Ω ⊂ M be a nonempty bounded set and define
rΩ := inf {r > 0 | there exists a p ∈ M such that Ω ⊂ B(p, r)} .
Then there exists a unique point pΩ ∈ M such that Ω ⊂ B(pΩ , rΩ ).
6.5. NONPOSITIVE SECTIONAL CURVATURE 327

We prove existence. Choose a sequence ri > rΩ and a sequence pi ∈ M such

that
Ω ⊂ B(pi , ri ), lim ri = rΩ .
i→∞

Choose q ∈ Ω. Then d(q, pi ) ≤ ri for every i. Since the sequence ri is

bounded and M is complete, it follows that pi has a convergent subsequence,
still denoted by pi . Its limit pΩ := limi→∞ pi satisfies Ω ⊂ B(pΩ , rΩ ).
We prove uniqueness. Let p0 , p1 ∈ M such that

Ω ⊂ B(p0 , rΩ ) ∩ B(p1 , rΩ ).

Since the exponential map expp : Tp M → M is a diffeomorphism, by Theo-

rem 6.5.2, there exists a unique vector v0 ∈ Tp0 M such that p1 = expp0 (v0 ).
Denote the midpoint between p0 and p1 by

1


m := expp0 2 v0 .

Then it follows from Step 1 that

d(p0 , q)2 + d(p1 , q)2 d(p0 , p1 )2

d(m, q)2 ≤ −
2 4
d(p , p ) 2
2 0 1
≤ rΩ −
4

for every q ∈ Ω. Since supq∈Ω d(m, q) ≥ rΩ , by definition of rΩ , it follows

that d(p0 , p1 ) = 0 and hence p0 = p1 . This proves Step 2.

Step 3. We prove Theorem 6.5.5.

Let q ∈ M and consider the group orbit Ω := {gq | g ∈ G}. Since G is

compact, this set is bounded. Let rΩ ≥ 0 and pΩ ∈ M be as in Step 2. Then

Ω ⊂ B(pΩ , rΩ ).

Since G acts on M by isometries, this implies

Ω = gΩ ⊂ B(gpΩ , rΩ )

for every g ∈ G. Hence it follows from the uniqueness statement in Step 2

that gpΩ = pΩ for every g ∈ G. This proves Step 3 and Theorem 6.5.5.
328 CHAPTER 6. GEOMETRY AND TOPOLOGY

6.5.3 Positive Definite Symmetric Matrices

We close this section with an example of a nonpositive sectional curvature
manifold which plays a key role in Donaldson’s beautiful paper on Lie algebra
theory [11]. Let n be a positive integer and consider the space
n o
P := P ∈ Rn×n | P T = P > 0 (6.5.7)

of positive definite symmetric n × n-matrices. (The notation “P > 0” means

hx, P xi > 0 for every nonzero vector x ∈ Rn .) Thus P is an open subset of
the vector space n o
S := S ∈ R n×n T
|S = S

of symmetric matrices and hence the tangent space of P is TP P = S

for every P ∈ P. However, we do not use the metric inherited from the
inclusion into S but define a Riemannian metric by

hS1 , S2 iP := trace S1 P −1 S2 P −1


(6.5.8)

for P ∈ P and S1 , S2 ∈ S = TP P.

Theorem 6.5.6. The space P with the Riemannian metric (6.5.8) is a

connected, simply connected, complete Riemannian manifold with nonposi-
tive sectional curvature. Moreover, P is a symmetric space and the group
GL(n, R) of nonsingular n × n-matrices acts on P by isometries via

g∗ P := gP g T (6.5.9)

for g ∈ GL(n, R) and P ∈ P.

Proof. See page 329.

Remark 6.5.7. The paper [11] by Donaldson contains an elementary direct

proof that the manifold P with the metric (6.5.8) satisfies the assertions of
Theorem 6.5.2.

Remark 6.5.8. The submanifold

P0 := {P ∈ P | det(P ) = 1} (6.5.10)

of positive definite symmetric matrices with determinant one is totally geo-

desic (see Remark 6.5.9 below). Hence all the assertions of Theorem 6.5.6
(with GL(n, R) replaced by SL(n, R)) remain valid for P0 .
6.5. NONPOSITIVE SECTIONAL CURVATURE 329

Remark 6.5.9. Let M be a Riemannian manifold and L ⊂ M be a sub-

manifold. Then the following are equivalent.
(i) If γ : I → M is a geodesic on an open interval I such that 0 ∈ I and

γ(0) ∈ L, γ̇(0) ∈ Tγ(0) L,

then there is a constant ε > 0 such that γ(t) ∈ L for |t| < ε.
(ii) If γ : I → L is a smooth curve on an open interval I and Φγ denotes
parallel transport along γ in M , then

Φγ (t, s)Tγ(s) L = Tγ(t) L ∀ s, t ∈ I.

(iii) If γ : I → L is a smooth curve on an open interval I and X ∈ Vect(γ)

is a vector field along γ (with values in T M ), then

X(t) ∈ Tγ(t) L ∀ t ∈ I =⇒ ∇X(t) ∈ Tγ(t) L ∀ t ∈ I.

A submanifold that satisfies these equivalent conditions is called totally

geodesic
Exercise 6.5.10. Prove the equivalence of (i), (ii), (iii) in Remark 6.5.9.
Hint: Choose suitable coordinates and translate each of the three assertions
into conditions on the Christoffel symbols.
Exercise 6.5.11. Prove that P0 is a totally geodesic submanifold of P.
Exercise 6.5.12. Prove that the space P is diffeomorphic to the quotient
space GL(n, R)/O(n) via polar decomposition. Prove that the space P0
is diffeomorphic to the quotient space SL(n, R)/SO(n). Prove that, in the
case n = 2, the space P0 is isometric to the hyperbolic space H2 .
Proof of Theorem 6.5.6. The manifold P is obviously connected and simply
connected as it is a convex open subset of a finite-dimensional vector space.
The remaining assertions will be proved in five steps.
Step 1. Let I → P : t 7→ P (t) be a smoth path in P and I → S : t 7→ S(t)
be a vector field along P . Then the covariant derivative of S is given by
1 1
∇S = Ṡ − SP −1 Ṗ − Ṗ P −1 S. (6.5.11)
2 2

The formula (6.5.11) determines a family of linear operators on the spaces

of vector fields along paths that satisfy the torsion-free condition

∇s ∂t P = ∇t ∂s P
330 CHAPTER 6. GEOMETRY AND TOPOLOGY

for every smooth map R2 → P : (s, t) 7→ P (s, t) and the Leibniz rule

∇ hS1 , S2 iP = h∇S1 , S2 iP + hS1 , ∇S2 iP

for any two vector fields S1 and S2 along P . These two conditions determine
the covariant derivative uniquely (see Lemma 3.6.5 and Theorem 3.7.8).
Step 2. The geodesics in P are given by

γ(t) = P exp(tP −1 S)
= exp(tSP −1 )P (6.5.12)
1/2 −1/2 −1/2 1/2
=P exp(tP SP )P

for P ∈ P, S ∈ S = TP P, and t ∈ R. In particular, P is complete.

The curve γ : R → P defined by (6.5.12) satisfies

γ̇(t) = S exp(tP −1 S) = SP −1 γ(t).

Hence it follows from Step 1 that

∇γ̇(t) = γ̈(t) − γ̇(t)γ(t)−1 γ̇(t) = γ̈(t) − SP −1 γ̇(t) = 0

for every t ∈ R. Hence γ is a geodesic. Note also that the curve γ : R → P

in (6.5.12) satisfies γ(0) = P and γ̇(0) = S.
Step 3. The curvature tensor on P is given by
1 1
RP (S, T )A = − SP −1 T P −1 A − AP −1 T P −1 S
4 4 (6.5.13)
1 1
+ T P SP A + AP −1 SP −1 T
−1 −1
4 4
for P ∈ P and S, T, A ∈ S .
Choose smooth maps P : R2 → P and A : R2 → S (understood as a vector
field along P ) and denote S := ∂s P and T := ∂t P . Then

R(S, T )A = ∇s ∇t A − ∇t ∇s A

and ∂s T = ∂t S. By Step 1 we have

1 1
∇s A = ∂s A − AP −1 S − SP −1 A,
2 2
1 1
∇t A = ∂t A − AP −1 T − T P −1 A.
2 2
6.5. NONPOSITIVE SECTIONAL CURVATURE 331

Hence
1 1
R(S, T )A = ∂s ∇t A − (∇t A)P −1 S − SP −1 (∇t A)
2 2
1 1
− ∂t ∇s A + (∇s A)P T + T P −1 (∇s A).
−1
2 2
Now Step 3 follows by a direct calculation which we leave to the reader.
Step 4. The manifold P has nonpositive sectional curvature.
By Step 3 with A = T and equation (6.5.8) we have
= trace SP −1 (RP (S, T )T )P −1


hS, RP (S, T )T iP
1
= − trace SP −1 SP −1 T P −1 T P −1


4
1
− trace SP −1 T P −1 T P −1 SP −1


4
1
+ trace SP −1 T P −1 SP −1 T P −1


4
1
+ trace SP −1 T P −1 SP −1 T P −1


4
1
= − trace SP −1 T P −1 T P −1 SP −1


2
1
+ trace SP −1 T P −1 SP −1 T P −1


2
1   1
= − trace X T X + trace X 2 ,


2 2
where X := P −1/2 SP −1 T P −1/2 . Write X =: (xij )i,j=1,...,n . Then, by the
Cauchy–Schwarz inequality, we have
X X
trace(X 2 ) = xij xji ≤ x2ij = trace(X T X)
i,j i,j

for every matrix X ∈ Rn×n . Hence hS, RP (S, T )T iP ≤ 0 for all P ∈ P and
all S, T ∈ S . This proves Step 4.
Step 5. P is a symmetric space.
Fix an element A ∈ P and define the map φ : P → P by
φ(P ) := AP −1 A
for P ∈ P. This map is a diffeomorphism, fixes the matrix A = φ(A), and
its derivative at P ∈ P is given by
dφ(P )S = −AP −1 SP −1 A for S ∈ S .
332 CHAPTER 6. GEOMETRY AND TOPOLOGY

Hence dφ(A) = −id. Moreover, for all P ∈ P and all S ∈ S , we have

(dφ(P )S)φ(P )−1 = −AP −1 SA−1
and therefore

2  
2 
|dφ(P )S|2φ(P ) = trace AP −1 SA−1 = trace P −1 S = |S|2P .
Hence φ is an isometry and this proves Theorem 6.5.6.
Remark 6.5.13. Theorem 6.5.6 carries over to the complex setting as fol-
lows. Replace P by the space
Q := Q ∈ Cn×n  Q∗ = Q > 0
 
(6.5.14)
of positive definite Hermitian matrices. Here Q∗ denotes the conjugate trans-
posed matrix of Q ∈ Cn×n and the notation “Q > 0” means z ∗ Qz > 0 for
every nonzero vector z ∈ Cn . Thus Q is an open subset of the vector space
H := H ∈ Cn×n  H ∗ = H
 

of Hermitian matrices. Define the Riemannian metric on Q by

hH1 , H2 iQ := Re trace(H1 Q−1 H2 Q−1 )

for Q ∈ Q and two Hermitian matrices H1 , H2 ∈ H = TQ Q. Then all the

assertions of Theorem 6.5.6 (with GL(n, R) replaced by GL(n, C)) carry over
to Q. The proof is verbatim the same, with the transposed matric replaced
by the conjugate tramsposed matrix and the trace replaced by the real part
of the trace.
Remark 6.5.14. The set
Q0 := Q ∈ Q  det(Q) = 1
 
(6.5.15)
of positive definite Hermitian matrices with determinant one is a totally
geodesic submanifold of Q. Hence all the assertions of Theorem 6.5.6
(with GL(n, R) replaced by SL(n, C)) remain valid for Q0 .
Exercise 6.5.15. Verify that Theorem 6.5.6 remains valid for Q. Prove
that Q0 is a totally geodesic submanifold of Q.
Exercise 6.5.16. Prove that the space Q is diffeomorphic to the quotient
space GL(n, C)/U(n) via polar decomposition. Prove that the space Q0
is diffeomorphic to the quotient space SL(n, C)/SU(n). Prove that, in the
case n = 2, the space Q0 is isometric to the hyperbolic space H3 .
The space Q0 ∼= SL(n, C)/SU(n) (with nonpositive sectional curvature)
can be viewed as a kind of dual space of the Lie group SU(n) (with nonneg-
ative sectional curvature).
6.6. POSITIVE RICCI CURVATURE* 333

6.6 Positive Ricci Curvature*

In this section we prove that every complete connected manifold M ⊂ Rn
whose Ricci curvature satisfies a uniform positive lower bound is necessarily
compact. If the sectional curvature is constant and positive, this follows
from Corollary 6.4.12 as was noted in Example 6.4.16.

6.6.1 The Ricci Tensor

Definition 6.6.1 (Ricci tensor). Let M ⊂ Rn be an m-dimensional sub-
manifold and fix an element p ∈ M . The Ricci tensor of M at p is the
symmetric bilinear form

Ricp : Tp M × Tp M → R

defined by
m
X
Ricp (u, v) := hRp (ei , u)v, ei i, (6.6.1)
i=1
where e1 , . . . , em is an orthonormal basis of Tp M . The Ricci tensor is inde-
pendent of the choice of this orthonormal frame and is symmetric by equa-
tions (5.2.16) and (5.2.18) in Theorem 5.2.13.

The Ricci Tensor in Local Coordinates

Let φ : U → Ω be a local coordinate chart on an open set U ⊂ M with values
in an open set Ω ⊂ Rm , denote its inverse by ψ := φ−1 : Ω → U, and let
∂ψ
Ei (x) := (x) ∈ Tψ(x) M, x ∈ Ω, i = 1, . . . , m,
∂xi
be the local frame of the tangent bundle determined by this coordinate chart.
Denote the coefficients of the first fundamental form by gij := hEi , Ej i and
` : Ω → R so that
the coefficients of the Riemann curvature tensor by Rijk
X
`
R(Ei , Ej )Ek = Rijk E` .
`

(see Section 5.5). Then

m
X m
X
ν
Ricij := Ric(Ei , Ej ) = Rνij = Rνijµ g µν . (6.6.2)
ν=1 µ,ν=1

(Exercise: Prove this.)

334 CHAPTER 6. GEOMETRY AND TOPOLOGY

6.6.2 The Bonnet–Myers Theorem

Theorem 6.6.2 (Bonnet–Myers). Let M ⊂ Rn be a complete, connected
manifold of dimension m ≥ 2 and suppose that there exists a δ > 0 such that
Ricp (v, v) ≥ (m − 1)δ |v|2 (6.6.3)
√
for every p ∈ M and every v ∈ Tp M . Then d(p, q) ≤ π/ δ for all p, q ∈ M
and hence M is compact.
Proof. See page 335.

Lemma 6.6.3. Let R × [0, 1] → M : (s, t) 7→ γs (t) be a smooth map such

that γ := γ0 : [0, 1] → M is a geodesic and γs (0) = γ0 (0) and γs (1) = γ0 (1)
∂ 
for all s ∈ R. Define X ∈ Vect(γ) by X(t) := ∂s γ (t) for 0 ≤ t ≤ 1.
s=0 s
Then
Z 1
d2 
 
2
E(γ s ) = |∇t X| − hR(X, γ̇)γ̇, Xi dt. (6.6.4)
ds2  s=0 0

Proof. In the proof of Theorem 4.1.4 we have seen that

Z 1
d
E(γs ) = − h∇t ∂t γs (t), ∂s γs (t)i dt
ds 0

for all s ∈ R (see equation (4.1.10)). Differentiate this equation again with
respect to s and use the identity ∇s ∂t = ∇t ∂s to obtain
Z 1
d2

d 
E(γ s ) = − h∇t ∂t γs , ∂s γs i dt
ds2 ds s=0 0
Z 1 Z 1
=− h∇s ∇t ∂t γs , ∂s γs i dt − h∇t ∂t γs , ∇s ∂s γs i dt
0 0
Z 1
=− hR(∂s γs , ∂t γs )∂t γs , ∂s γs i dt
0
Z 1 Z 1
− h∇t ∇t ∂s γs , ∂s γs i dt − h∇t ∂t γs , ∇s ∂s γs i dt.
0 0

Now take s = 0 to obtain

Z 1 Z 1
d2 

E(γs ) = − hR(X, γ̇)γ̇, Xi dt − h∇t ∇t X, Xi dt
ds2 s=0 0 0
Z 1 
= |∇t X|2 − hR(X, γ̇)γ̇, Xi dt.
0

This proves Lemma 6.6.3.

6.6. POSITIVE RICCI CURVATURE* 335

Proof of Theorem 6.6.2. Let p, q ∈ M . By the Hopf–Rinow Theorem 4.6.6

there exists a geodesic γ : [0, 1] → M such that
γ(0) = p, γ(1) = q, L(γ) = d(p, q).
Let X ∈ Vect(γ) be a vector field along γ such that X(0) = 0 and X(1) = 0
and define γs (t) := expγ(t) (sX(t)) for s ∈ R and 0 ≤ t ≤ 1. Then s = 0 is
the absolute minimum of the function R → R : s 7→ E(γs ) by Lemma 4.4.1.
d2
Hence ds 2 |s=0 E(γs ) ≥ 0 and by Lemma 6.6.3 this implies
Z 1 Z 1
hR(X, γ̇)γ̇, Xi dt ≤ |∇t X(t)|2 dt. (6.6.5)
0 0
Now assume p 6= q and choose an orthonormal frame E1 , . . . , Em along γ
such that E1 = γ̇/|γ̇| and ∇t Ei ≡ 0 for i = 1, . . . m. Define
Xi (t) := sin(πt)Ei (t)
for i = 1, . . . m and 0 ≤ t ≤ 1. Then |∇t Xi (t)| = π cos(πt) for all i and t and
m
X
δ(m − 1) |γ̇(t)|2 ≤ Ricγ(t) (γ̇(t), γ̇(t)) = hR(Ei (t), γ̇(t))γ̇(t), Ei (t)i
i=2

for 0 ≤ t ≤ 1. Multiply this inequality by sin2 (πt), integrate

R1 over the unit
interval, and use the identities |γ̇(t)| = d(p, q) and 0 sin2 (πt) dt = 1/2 to
obtain the estimate
Z 1
δ(m − 1) 2
d(p, q) = δ(m − 1) sin2 (πt) |γ̇(t)|2 dt
2 0
Xm Z 1
≤ hR(Xi (t), γ̇(t))γ̇(t), Xi (t)i dt
i=2 0
Xm Z 1
≤ |∇t Xi (t)|2 dt
i=2 0
Z 1
= (m − 1) π 2 cos2 (πt) dt
0
π 2 (m − 1)
= .
2
Here the third step uses (6.6.5). Since m ≥ 2 it follows from this estimate
that d(p, q)2 ≤ π 2 /δ and this proves Theorem 6.6.2.
A direct consequence of Theorem 6.6.2 is that every complete manifold
with positive Ricci curvature has a compact universal cover and hence has
a finite fundamental group.
336 CHAPTER 6. GEOMETRY AND TOPOLOGY

6.6.3 Positive Sectional Curvature

Corollary 6.6.4. Let M ⊂ Rn be a complete, connected manifold of dimen-
sion m ≥ 2 and suppose that there exists a δ > 0 such that
K(p, E) ≥ δ
for every p ∈ M and every 2-dimensional linear subspace E ⊂ Tp M . Then
π
d(p, q) ≤ √
δ
for all p, q ∈ M and hence M is compact.
Proof. The condition K ≥ δ implies (6.6.3) with m := dim(M ) and hence
the assertion follows from Theorem 6.6.2. This proves Corollary 6.6.4.

The example of the m-sphere shows that the estimate in Corollary 6.6.4
is sharp. Namely, M := S m has sectional curvature K = 1 and diameter π.
The paraboloid M := (x, y, z) ∈ R3 | z = x2 + y 2 } has positive Gaußian
curvature and so positive Ricci curvature (Lemma 6.7.2) but is noncompact.
Remark 6.6.5 (Sphere Theorem). The Topological Sphere Theorem
asserts that every complete, connected, simply connected Riemannian m-
manifold M whose sectional curvature satisfies the estimate
1/4 < K(p, E) ≤ 1
for every p ∈ M and every 2-dimensional linear subspace E ⊂ Tp M must be
homeomorphic to the m-sphere. The Differentiable Sphere Theorem
asserts under the same assumptions that M is diffeomorphic to S m .
The problem goes back to a question posed by Heinz Hopf [24, 25] in the
1920s. After intermediate results by Rauch [43] (with 1/4 replaced by 3/4)
and others, the Toplogical Sphere Theorem was proved in 1961 by Berger [5]
and Klingenberg [30]. The Differentiable Sphere Theorem was proved in
2007 by Brendle and Schoen [6, 7, 8]. They even weakened the assumption
to 0 < maxE K(p, E) < 4 minE K(p, E) for all p ∈ M , where the maximum
and minimum are taken over all 2-dimensional linear subspaces E ⊂ Tp M .
The Topological Sphere Theorem is sharp, as the suitably scaled Fubini–
Study metric on complex projective space satisfies 1/4 ≤ K(p, E) ≤ 1 for
all p and E. The Differentiable Sphere Theorem is a significant improve-
ment, because in many dimensions there exist smooth m-manifolds that are
homeomorphic, but not diffeomorphic, to S m . These are the so-called ex-
otic spheres and many of those do not even admit metrics of positive scalar
curvature [23]. (For the definition of scalar curvature see §6.7.)
6.7. SCALAR CURVATURE* 337

6.7 Scalar Curvature*

This section introduces the scalar curvature and explains how it is related to
the Ricci tensor and the Riemann curvature tensor in dimensions 2 and 3.
The section also includes a brief discussion of several problems in differential
geometry in which the scalar curvature plays a central role.

6.7.1 Definition and Basic Properties

Let m be a positive integer and let M ⊂ Rn be an m-manifold. For each
element p ∈ M denote by Rp : Tp M × Tp M → End(Tp M ) the Riemann cur-
vature tensor and by Ricp : Tp M × Tp M → R the Ricci tensor of M at p.

Definition 6.7.1 (Scalar curvature). Fix an element p ∈ M . The scalar

curvature S(p) of M at p is the trace of the Ricci tensor and is given by
m
X m
X
S(p) := Ricp (ei , ei ) = hRp (ei , ej )ej , ei i, (6.7.1)
i=1 i,j=1

where e1 , . . . , em is an orthonormal basis of Tp M . The scalar curvature is

independent of the choice of this orthonormal frame.

Lemma 6.7.2. Assume m = 2 and let K : M → R be the Gaußian curvature

of M in (6.4.2). Then

S(p) = 2K(p), Ricp (u, v) = K(p)hu, vi, (6.7.2)

 
hRp (u, v)w, zi = K(p) hu, zihv, wi − hu, wihv, zi (6.7.3)
for all p ∈ M and u, v, w, z ∈ Tp M .

Proof. By definition, the Gaußian curvature is given by

hRp (u, v)v, ui
K(p) = (6.7.4)
|u|2 |v|2 − hu, vi2
for every pair of linearly independent tangent vectors u, v ∈ Tp M . Take u
to be a unit vector orthogonal to v to obtain Ricp (v, v) = K(p)|v|2 for
every v ∈ Tp M . Since Ricp : Tp M × Tp M → R is a symmetric bilinear form,
this implies the second equality in (6.7.2). With this understood the first
equality in (6.7.2) follows directly from the definition of the scalar curvature
in (6.7.1). Equation (6.7.3) follows from (6.7.4) by Theorem 6.4.8. This
proves Lemma 6.7.2.
338 CHAPTER 6. GEOMETRY AND TOPOLOGY

Lemma 6.7.3. Assume m = 3. Then

hRp (u, v)w, zi = Ricp (v, w)hu, zi − Ricp (u, w)hv, zi
+ Ricp (u, z)hv, wi − Ricp (v, z)hu, wi (6.7.5)
S(p)  
− hu, zihv, wi − hu, wihv, zi
2
for all p ∈ M and all u, v, w, z ∈ Tp M .

Proof. Choose an orthonormal basis e1 , e2 , e3 of the tangent space Tp M and

define the endomorphism

Qp : Tp M → Tp M

by X
Qp u := Rp (u, ei )ei
i
for u ∈ Tp M . The right hand side of this equation is independent of the
choice of the orthonormal basis and the endomorphism Qp satusfies

trace(Qp ) = S(p)

and
hQp u, vi = Ricp (u, v)
for all u, v ∈ Tp M . With this notation equation (6.7.5) takes the form
 
S(p)
Rp (u, v)w = Ricp (v, w)u + hv, wi Qp u − u
2
  (6.7.6)
S(p)
− Ricp (u, w)v − hu, wi Qp v − v .
2

It suffices to verify equation (6.7.6) in the following three cases.

(a) u, v are linarly dependent.
(b) u, v, w are orthonormal.
(c) u, v are orthonormal and w = v.
In the case (a) both sides of equation (6.7.6) vanish. In the case (b) we have

Rp (u, v)w = hRp (u, v)w, uiu + hRp (u, v)w, viv
= Ricp (v, w)u − Ricp (u, w)v,

and this is equivalent to (6.7.6).

6.7. SCALAR CURVATURE* 339

In the case (c) equaton (6.7.6) takes the form

S(p)
Rp (u, v)v = Ricp (v, v)u − Ricp (u, v)v + Qp u − u. (6.7.7)
2
To verify this formula, choose a unit vector w that is orthogonal to u and v.
Then it follows from the definition of S(p) and Qp that
S(p)
= hRp (u, v)v, ui + hRp (w, v)v, wi + hRp (u, w)w, ui
2
= Ricp (v, v) + hRp (u, w)w, ui,
and
Qp u = Rp (u, v)v + Rp (u, w)w
= Rp (u, v)v + hRp (u, w)w, viv + hRp (u, w)w, uiu
S(p)
= Rp (u, v)v + Ricp (u, v)v + u − Ricp (v, v)u.
2
This proves (6.7.7) and Lemma 6.7.3.

Scalar Curvature in Local Coordinates

Let φ : U → Ω be a local coordinate chart on an open set U ⊂ M with values
in an open set Ω ⊂ Rm , denote its inverse by
ψ := φ−1 : Ω → U,
and denote by
Ei (x) := ∂i ψ(x) ∈ Tψ(x) M
for x ∈ Ω and i = 1, . . . , m the local frame of the tangent bundle determined
by this coordinate chart. Denote the coefficients of the first fundamental
form by gij := hEi , Ej i, of the Ricci tensor by
Ricij := Ric(Ei , Ej ),
`
and of the Riemann curvature tensor by Rijk` and Rijk so that
m
X
`
Rijk` = hR(Ei , Ej )Ek , E` i, R(Ei , Ej )Ek = Rijk E` .
`=1

Then the scalar curvature is the function S : Ω → R given by

m
X m
X m
X
S= Ricij g ij = ν
Rνij g ij = Rijk` g jk g i` . (6.7.8)
i,j=1 i,j,ν=1 i,j,k,`=1

(Exercise: Prove this.)

340 CHAPTER 6. GEOMETRY AND TOPOLOGY

Positive Scalar Curvature

An important question for a compact smooth manifold M is whether or not
it admits a Riemannian metric of positive scalar curvature. A theorem of
Lichnerowicz [34] asserts that, if M is a compact spin manifold of dimen-
sion m = 4n that admits a Riemannian metric of positive scalar curvature,
then a certain characteristic class of this manifold (the A-genus)
b must van-
ish. The definitions of the terms that appear in this sentence (spin structure
and A-genus)
b as well as in the proof, which involves the Dirac operator, the
Atiyah–Singer index theorem, and the Weitzenböck formula, go beyond the
scope of the present book. For an exposition see [48, Theorem 6.30].
A nonlinear variant of Lichnerowicz’ theorem asserts that a compact
oriented smooth 4-manifold with b+ +
2 − b1 odd and b2 > 1 that admits a
Riemannian metric of positive scalar curvature has vanishing Seiberg-Witten
invariants (see [48, Proposition 7.32]).
In another direction Gromov and Lawson [15] proved that if M1 and M2
are two compact manifolds of dimension m ≥ 3 that admit Riemannian met-
rics of positive scalar curvature, then so does their connected sum M1 #M2
(see also [48, Theorem 2.18]).
In the late 1970’s Schoen and Yau [53] proved, using minimal surfaces,
that the torus Tm = Rm /Zm does not admit a metric of positive scalar
curvature for m ≤ 7. We remark that the A-genus
b of the torus vanishes and
so Lichnerowicz’ theorem does not apply. In [15] Gromov and Lawson refined
the techniques of Lichnerowicz to prove that, for any m, the m-torus does not
admit a metric of positive scalar curvature. In fact, they proved that for any
compact spin manifold M of dimension m the connected sum N := M #Tm
does not admit a metric of positive scalar curvature. Moreover, they proved
that if N admits a metric of nonnegative scalar curvature, then this metric
must be flat and N must be the standard m-torus.

6.7.2 Constant Scalar Curvature

An interesting class of Riemannian metrics consists of those that have con-
stant scalar curvature. In dimension m = 2 the existence of such a metric
is the content of the uniformisation theorem. The proof involves the
solution of the Kazdan-Warner equation and goes beyond the scope of this
book. For an exposition see [48, Theorem 2.20 and Theorem D.1].
In dimension 2 Lemma 6.7.2 shows that the constant scalar curvature
condition is equivalent to constant sectional curvature. However, in higher
dimensions the constant scalar curvature condition is far more general. By
6.7. SCALAR CURVATURE* 341

Corollary 6.4.12 every compact simply connected m-manifold with con-

stant sectional curvature is diffeomorphic to the m-sphere, while constant
scalar curvature metrics exist on every compact manifold. Examples are
the Fubini-Study metric on complex projective space (Example 2.8.5 and
Example 3.7.5), locally symmetric spaces (Theorem 6.3.4), and products of
Riemannian manifolds with constant scalar curvature.
Definition 6.7.4. Let M be a Riemannian manifold with the metric g. A
Riemannian metric g 0 on M is called conformally equivalent to g iff there
exists a smooth function λ : M → (0, ∞) such that g 0 = λg. The set of all
such Riemannian metrics is called the conformal class of g.
Remark 6.7.5 (Yamabe problem). The Yamabe problem asserts that
the conformal class of every Riemannian metric on a compact m-manifold M
of dimension m ≥ 2 contains a metric of positive scalar curvature.
This problem was formulated in 1960 by Hidehiko Yamabe and was even-
tually settled in the affirmative in 1984 by the combined work of Hide-
hiko Yamabe [58], Thierry Aubin [4], Neil Trudinger [57], and Richard
Schoen [52]. The proof for a compact manifold M of dimension m > 2
relies on finding a positive function f : M → R and a real number c that
satisfy the Yamabe equation
4(m − 1)
∆g f + Sg f = cf (m+2)/(m−2) . (6.7.9)
m−2
Here Sg : M → R denotes the scalar curvature of the Riemannian metric g
and ∆g denotes its Laplace–Beltrami operator. In local coordinates this
operator is given by the formula
1 X ∂
ij
p ∂
∆g = p i
g det(g) j . (6.7.10)
det(g) i,j ∂x ∂x

If f : M → (0, ∞) is a positive solution of (6.7.9), then the Riemannian

metric f 4/(m−2) g has the scalar curvature c. Exercise: Prove this. Hint:
Show first that
Su2 g = u−2 Sg + 2(m − 1)u−3 ∆g u − (m − 1)(m − 4)u−4 |du|2g . (6.7.11)

Then take f := um/2−1 and use the identities

|dun |2g = n2 u2n−2 |du|2g , (6.7.12)

∆g un = nun−1 ∆g u + n(n − 1)un−2 |du|2g (6.7.13)

for a smooth function u : M → (0, ∞) and a real number n > 0.
342 CHAPTER 6. GEOMETRY AND TOPOLOGY

Examples of constant scalar curvature metrics arise from the Einstein

condition (see Lemma 6.7.7 below).
Definition 6.7.6 (Einstein metric). A Riemannian manifold M is called
an Einstein manifold iff its Ricci tensor is a scalar multiple of the first
fundamental form, i.e. there exists a smooth function λ : M → R
Ricp (u, v) = λ(p)hu, vi (6.7.14)
for all p ∈ M and all u, v ∈ Tp M . It follows from the definitions that the
factor λ in (6.7.14) is related to the scalar curvature S by
S
. λ= (6.7.15)
m
Lemma 6.7.2 shows that every Riemannian metric on a 2-manifold is an
Einstein metric and the factor λ = K is the Gaußian curvature.
Lemma 6.7.7. Let M be an Einstein manifold of dimension m ≥ 3. Then
the scalar curvature of M is locally constant.
Proof. Let p ∈ M . Then there exists a local orthonormal frame of the
tangent bundle E1 , . . . , Em ∈ Vect(U ) in a neighborhood U of p such that the
covariant derivatives ∇Ei all vanish at p. (Exercise: Prove this.) Denote
the deriviative of a function f at p in the direction Ei (p) by ∂i f . Then
X
0= h(∇Ei R)(Ej , Ek )Ek , Ej i
j,k
X X
+ h(∇Ej R)(Ek , Ei )Ek , Ej i + h(∇Ek R)(Ei , Ei )Ek , Ej i
j,k j,k
X
= ∂i hR(Ej , Ek )Ek , Ej i
j,k
X X X X
+ ∂j hR(Ek , Ei )Ek , Ej i + ∂k hR(Ei , Ei )Ek , Ej i
j k k j
X X
= ∂i S − ∂j Ric(Ei , Ej ) − ∂k Ric(Ei , Ek )
j k
m−2
= ∂i S.
m
Here the the first equality follows from the second Bianchi identity (6.3.7),
the second holds at p because ∇Ei (p) = 0, the third follows from the def-
initions of Ric and S, and the last uses the identity Ric(Ei , Ej ) = δij S/m,
which holds by (6.7.14) and (6.7.15) Since m ≥ 3 it follows that the deriva-
tive of S vanishes everywhere, and this proves Lemma 6.7.7.
6.8. THE WEYL TENSOR* 343

Examples of Einstein metrics include all constant sectional curvature

metrics by Theorem 6.4.8, the Fubini-Study metric on complex projective
space, and all metrics with vanishing Ricci tensor. Examples of the lat-
ter are Calabi–Yau metrics on complex manifolds and G2 -structures on 7-
manifolds. These are again subjects that go far beyond the scope of this
book. In general, the construction of Einstein metrics and the question of
their existence is a highly nontrivial problem in differential geonetry. The
study of this problem has a long history and there are many deep theorems
and interesting open questions about this subject.

6.8 The Weyl Tensor*

This section introduces the Weyl tensor and explains some of its basic prop-
erties. The section closes with brief discussions of locally conformally flat
metrics and self-dual four-manifolds.

6.8.1 Definition and Basic Properties

Let m be a positive integer and let M ⊂ Rn be an m-manifold. For each
element p ∈ M denote by Rp : Tp M × Tp M → End(Tp M ) the Riemann cur-
vature tensor and by Ricp : Tp M × Tp M → R the Ricci tensor of M at p.
Also let S : M → R be the scalar curvature in Definition 6.7.1.
Definition 6.8.1 (Weyl tensor). Assume m ≥ 3. The Weyl tensor of M
at an element p ∈ M is the bilinear map
Wp : Tp M × Tp M → End(Tp M )
defined by
hWp (u, v)w, zi := hRp (u, v)w, zi
1  
− Ricp (v, w)hu, zi − Ricp (u, w)hv, zi
m−2
1   (6.8.1)
− Ricp (u, z)hv, wi − Ricp (v, z)hu, wi
m−2
S(p)  
+ hu, zihv, wi − hv, zihu, wi
(m − 1)(m − 2)
for u, v, w, z ∈ Tp M .
Lemma 6.7.3 shows that the Weyl tensor vanishes in dimension three.
In higher dimensions the Weyl tensor may be nonzero. The next lemma
summarizes the basic algebraic properties of the Weyl tensor.
344 CHAPTER 6. GEOMETRY AND TOPOLOGY

Lemma 6.8.2. The Weyl tensor Wp : Tp M × Tp M → End(Tp M ) at an el-

ement p ∈ M is a skew-symmetric bilinear map with values in the space of
skew-adjoint endomorphisms of Tp M and, for all u, v, w, z ∈ Tp M and every
orthonormal basis e1 , . . . , em of Tp M , it satisfies

hWp (u, v)w, zi = hWp (w, z)u, vi, (6.8.2)

Wp (u, v)w + Wp (v, w)u + Wp (w, u)v = 0, (6.8.3)

m
X
hWp (ei , u)v, ei i = 0. (6.8.4)
i=1

Proof. The skew-symmetry of the Weyl tensor and equation (6.8.2) follow
directly from the definition and Theorem 5.2.13. It then follows from (6.8.2)
that Wp (u, v) is a skew-adjoint endomorphism of Tp M for all u, v ∈ Tp M .
The verification of the Bianchi identity (6.8.3) is a straight forward compu-
tation which we leave as an exercise. To prove (6.8.4), let u, v ∈ Tp M and
choose an orthonormal basis e1 , . . . , em of Tp M . Then
m
X m
X
hWp (ei , u)v, ei i = hRp (ei , u)v, ei i
i=1 i=1
m
1 X 
− Ricp (u, v)hei , ei i − Ricp (ei , v)hu, ei i
m−2
i=1
m 
1 X 
− Ricp (ei , ei )hu, vi − Ricp (u, ei )hei , vi
m−2
i=1
m 
S(p) X 
+ hei , ei ihu, vi − hu, ei ihei , vi
(m − 1)(m − 2)
i=1
= Ricp (u, v)
m−1
− Ricp (u, v)
m−2
S(p) 1
− hu, vi + Ricp (u, v)
m−2 m−2
S(p)
+ hu, vi
m−2
= 0.

This proves (6.8.4) and Lemma 6.8.2.

6.8. THE WEYL TENSOR* 345

The Weyl Tensor in Local Coordinates

Let φ : U → Ω be a local coordinate chart on an open set U ⊂ M with val-
ues in an open set Ω ⊂ Rm , denote its inverse by ψ := φ−1 : Ω → U and
let Ei (x) := ∂i ψ(x) ∈ Tψ(x) M for x ∈ Ω and i = 1, . . . , m be the be the local
frame of the tangent bundle determined by this coordinate chart. Denote
the coefficients of the first fundamental form by gij := hEi , Ej i, of the Ricci
tensor by Ricij := Ric(Ei , Ej ), and of the Riemann curvature tensor by Rijk ` .

Let S = i,j Ricij g ij = m ν ij

P P
i,j,ν Rνij g be the scalar curvature in local coor-
dinates. Then the cofficients Wijk ` : Ω → R of the Weyl tensor are defined
` E . and they can be expresses in the form
P
by W (Ei , Ej )Ek = ` Wijk `
X
ν
Wijk` := hW (Ei , Ej )Ek , E` i = Wijk gν`
ν
X
ν S  
= Rijk gν` + gi` gjk − gik gj` (6.8.5)
ν
(m − 1)(m − 2)
1  
− Ricjk gi` − Ricik gj` + Rici` gjk − Ricj` gik .
m−2

6.8.2 Conformal Invariance

Definition 6.8.3 (Locally conformally flat metric). Let M be a Rie-
mannian manifold. The metric g on M is called locally conformally flat,
iff for each p ∈ M there exists a Riemannian metric on M that is confor-
mally equivalent to g (see Definition 6.7.4) and flat in a neighborhood of p.
By the local C-A-H Theorem 6.1.15 a Riemannian m-manifold M is lo-
cally conformally flat if and only if each p ∈ M has an open neighborhood U
that is conformally diffeomorphic to an open subset Ω ⊂ Rm , i.e. there
exists a coordinate chart φ : U → Ω and a smooth function λ : U → (0, ∞)
such that |v| = λ(p)|dφ(p)v|Rm for all p ∈ U and all v ∈ Tp M .
A remarkable property of the Weyl tensor is that it remains unchanged
under multiplication of the Riemannian metric by a positive function and
so is an invariant of the conformal class of the metric. This is easy to see
when the function is constant. In that case the Riemann curvature tensor
and the Ricci tensor remain unchanged, the scalar curvature gets multiplied
by the inverse, and so the Weyl tensor remains unchanged. As Remark 6.7.5
shows, the situation is more complicated when instead of multiplying the
metric by a constant, we multiply it by a nonconstant positive function. The
conformal invariance of the Weyl tensor can then be proved by a somewhat
cumbersome calculation in local coordinates.
346 CHAPTER 6. GEOMETRY AND TOPOLOGY

Remark 6.8.4. This discussion shows that the Weyl tensor vanishes for
every Riemannian metric that is locally conformally flat. In fact, it turns
out that in dimension m ≥ 4 the Weyl tensor of M vanishes if and only if
the Riemannian metric on M is locally conformally flat (see [32]). More-
over, a theorem of Kuiper [26, 31] asserts that every compact, connected,
simply connected Riemannian m-manifold that is locally conformally flat is
conformally diffeomorphic to the m-sphere with its constant sectional curva-
ture metric. Thus every compact, connected, simply connected Riemannian
manifold of dimension m ≥ 4 that is not diffeomorphic to S m must have a
nonvanishing Weyl tensor.

6.8.3 Self-Dual Four-Manifolds

The lowest dimension in which the study of the Weyl tensor is interesting
is m = 4. To explain this, it is useful to consider the notion of an oriented
manifold.
Definition 6.8.5 (Orientation). An orientation of an m-manifold M is
a collection of orientations of the tangent spaces Tp M , one for each p ∈ M ,
that depend continuously on p, i.e. if E1 , . . . , Em are pointwise linearly inde-
pendent vector fields in a connected open neighborhood U ⊂ M of p and the
vectors E1 (p), . . . , Em (p) form a positive basis of Tp M , then for every q ∈ U
the vectors E1 (q), . . . , Em (q) form a positive basis of Tq M . In the intrinsic
language an oriented manifold is one equipped with an atlas such that all the
transition maps are orientation preserving diffeomorphisms.
Definition 6.8.6 (2-Form). Let M be a smooth manifold. A 2-form on M
is a collection of skew-symmetric bilinear maps ωp : Tp M × Tp M → R, one
for each p ∈ M , which is smooth in the sense that for every pair of smooth
vector fields X, Y on M the assignment p 7→ ωp (X(p), Y (p)) defines a smooth
function on M . The space of all 2-forms on M is denoted by Ω2 (M ).
In the similar vein the Weyl tensor of a Riemannian m-manifold M can
be thought of as 2-form with values in the endomorphism bundle End(T M ).
By the symmetry properties in Lemma 6.8.2 the Weyl tensor induces a linear
map W : Ω2 (M ) → Ω2 (M ) via the formula
X
(Wω)p (u, v) := hWp (u, v)ei , ej iωp (ei , ej ) (6.8.6)
1≤i<j≤m

for ω ∈ Ω2 (M ), p ∈ M , and u, v ∈ Tp M , where e1 , . . . , em is an orthonormal

basis of Tp M . The right hand side of equation (6.8.6) is independent of the
choice of this orthonormal basis and is a 2-form by Lemma 6.8.2.
6.8. THE WEYL TENSOR* 347

Now let M be an oriented Riemannian 4-manifold. Then a 2-form ω

on M is called self-dual iff it satisfies the condition

ωp (e0 , e1 ) = ωp (e2 , e3 ) (6.8.7)

for every p ∈ M and every positive orthonormal basis e0 , e1 , e2 , e3 of Tp M . It

is called anti-self-dual iff it satisfies (6.8.7) for every p ∈ M and every neg-
ative orthonormal basis e0 , e1 , e2 , e3 of Tp M . Thus ω is anti-self-dual if and
only if it is self-dual for the opposite orientation. Denote the space of self-
dual 2-forms by Ω2,+ (M ) and the space of anti-self-dual 2-forms by Ω2,− (M ).
Then there is a direct sum decomposition

Ω2 (M ) = Ω2,+ (M ) ⊕ Ω2,− (M ).

Lemma 6.8.7. Let M be an oriented Riemannian 4-manifold. Then the

linear operator W : Ω2 (M ) → Ω2 (M ) in (6.8.6) preserves the subspace of
self-dual 2-forms and the subspace of anti-self-dual 2-forms.
Proof. Fix any orthonormal basis e0 , e1 , e2 , e3 of Tp M and abbreviate

wijk` := hWp (ei , ej )ek , e` i

for i, j, k, ` ∈ {0, 1, 2, 3}. Then by Lemma 6.8.2 we have

wijk` = −wjik` = −wij`k = wk`ij , wijk` + wjki` + wkij` = 0, (6.8.8)

w0ij0 + w1ij1 + w2ij2 + w3ij3 = 0 (6.8.9)

for all i, j, k, `. It follows from (6.8.8) and (6.8.9) that

w0102 = w2331 , w0103 = w2312 , w0203 = w3112 ,

(6.8.10)
w2302 = w0131 , w2303 = w0112 , w3103 = w0212 .

Namely, by P (6.8.8) each of these six identities is equivalent to an equation

of the form 3i=0 wijki = 0 with j 6= k and this holds by (6.8.9). Use (6.8.8)
and (6.8.9) again to obtain

w0101 − w2323 − w0202 + w3131

= w0220 + w1221 + w3223 − w0110 − w2112 − w3113
=0

and hence, by cyclic permutation of 1, 2, 3,

ε := w0101 − w2323 = w0202 − w3131 = w0303 − w1212 .

348 CHAPTER 6. GEOMETRY AND TOPOLOGY

Since w0101 + w0202 + w0303 = 0 by (6.8.9), this implies

3ε = w1221 + w2332 + w1331 = w1221 − w0330 = w0303 − w1212 = ε

Thus ε = 0 and so

w0101 = w2323 , w0202 = w3131 , w0303 = w1212 . (6.8.11)

Now assume that τ is a nonzero self-dual 2-form and fix an element p ∈ M .

Then there exists a positive orthonormal basis e0 , e1 , e2 , e3 of Tp M and a
real number λ 6= 0 such that τ (e0 , e1 ) = τ (e2 , e3 ) = λ and τ (ei , ej ) = 0
for all other pairs i, j. Hence (Wτ )(ei , ej ) = λw01ij + λw23ij for all i and j.
Take λ = 1 and use the equations (6.8.8), (6.8.10), and (6.8.11) to obtain

(Wτ )(e0 , e1 ) = w0101 + w2301 = w0123 + w2323 = (Wτ )(e2 , e3 ),

(Wτ )(e0 , e2 ) = w0102 + w2302 = w0131 + w2331 = (Wτ )(e3 , e1 ),
(Wτ )(e0 , e3 ) = w0103 + w2303 = w0112 + w2312 = (Wτ )(e1 , e2 ).

Thus Wτ is self-dual. The anti-self-dual case follows by reversing the orien-

tation. This proves Lemma 6.8.7.

Lemma 6.8.8. Let M be an oriented Riemannian 4-manifold and denote

by W : Ω2 (M ) → Ω2 (M ) the linear operator determined by the Weyl tensor
via (6.8.6). Then the following are equivalent.
(i) If τ ∈ Ω2 (M ) is anti-self-dual, then Wτ = 0.
(ii) The Weyl tensor W satisfies satisfies the equation

Wp (e0 , e1 ) = Wp (e2 , e3 ) (6.8.12)

for every p ∈ M and every positive orthonormal basis e0 , e1 , e2 , e3 of Tp M .

Proof. We prove that (i) implies (ii). Thus assume (i) and choose a positive
orthonormal basis e0 , e1 , e2 , e3 of Tp M . Let τ be the 2-form defined by

τ (e0 , e1 ) := 1, τ (e2 , e3 ) := −1

and τ (ei , ej ) := 0 for all other pairs i, j. Then τ is anti-self-dual and

hence Wτ = 0 by (i). Thus it follows from (6.8.2) and (6.8.6) that

hW (e0 , e1 )u − W (e2 , e3 )u, vi = (Wτ )(u, v) = 0

for all u, v ∈ Tp M and this proves (ii).

6.8. THE WEYL TENSOR* 349

Conversely, suppose that (ii) holds, choose a positive orthonormal ba-

sis e0 , e1 , e2 , e3 of Tp M , and let τ be an anti-self-dual 2-form. Then
λ1 := τ (e0 , e1 ) = −τ (e2 , e3 ),
λ2 := τ (e0 , e2 ) = −τ (e3 , e1 ),
λ3 := τ (e0 , e3 ) = −τ (e1 , e2 )
and hence
Wτ = λ1 h(W (e0 , e1 ) − W (e2 , e3 ))·, ·i
+ λ2 h(W (e0 , e2 ) − W (e3 , e1 ))·, ·i
+ λ3 h(W (e0 , e3 ) − W (e1 , e2 ))·, ·i = 0
by (ii). This proves Lemma 6.8.8.
Definition 6.8.9. An oriented Riemannian 4-manifold is called self-dual
if its Weyl tensor satisfies the equivalent conditions of Lemma 6.8.8. It is
called anti-self-dual iff it is self-dual for the opposite orientation.
Examples of self-dual 4-manifolds are all 4-manifolds with constant sec-
tional curvature by Theorem 6.4.8, or more generally all locally conformally
flat 4-manifolds such as S 1 × S 3 . Other examples are the complex projec-
tive plane with its Fubini–Study metric (Examples 2.8.5 and 3.7.5), and
Ricci flat Kähler surfaces with the opposite of the complex orientation (the
K3-surface and the Enriques surface). Compact simply connected smooth
4-manifolds with signature zero that are not diffeomorphic to the 4-sphere,
such as S 2 × S 2 or the one-point blowup of the projective plane, do not
admit any self-dual metrics by Kuiper’s theorem (Remark 6.8.4), because
every self-dual metric on such a manifold is locally conformally flat. For a
survey of these basic examples see Kalafat [27].
The study of self-dual 4-manifolds was initiated by Penrose [42] and
Atiyah–Hitchin–Singer [3] and was later taken up by many authors including
Taubes [55], LeBrun [33], and Donaldson [12].
In [12] Donaldson proposes to study the moduli space of self-dual metrics
on a compact oriented smooth 4-manifold (without boundary) modulo the
action of the group of diffeomorphisms. This is a very difficult problem. The
self-duality equation is a system of nonlinear partial differential equations
and they do give rise to a “finite-dimensional moduli space”. However, this
space may be highly singular and it is noncompact unless it is empty. It
would need to be “compactified” in a suitable way, one would need to find a
way to understand the singularities, and one would have assign to it some
kind of “virtual fundamental class” to obtain numerical invariants by pairing
this class with suitable cohomology classes in the ambient space.
350 CHAPTER 6. GEOMETRY AND TOPOLOGY
Appendix A

Notes

A.1 Maps and Functions

The notation
f :X→Y
means that f is a function which assigns to every point x in the set X a
point f (x) in the set Y . When Y = R we express this by saying that f is
a real valued function defined on the set X and, if Y is a vector space, we
may say that f is a vector valued function. However in general it is better
to say that f is a map from X to Y and call the set X the source of the
map and the set Y its target. The graph of f is the set
graph(f ) := {(x, y) ∈ X × Y | y = f (x)}.
We always distinguish two maps with the same graph when their targets are
different.
A map f : X → Y is said to be
   
 injective   f (x1 ) = f (x2 ) =⇒ x1 = x2 
surjective iff ∀y ∈ Y ∃x ∈ X s.t. y = f (x)
bijective it is both injective and surjective.
   

Then
(a) f is injective ⇐⇒ it has a left inverse g : Y → X (i.e. g ◦ f = idX );
(b) f is surjective ⇐⇒ it has a right inverse g : Y → X (i.e. f ◦ g = idY );
(c) f is bijective ⇐⇒ it has a two sided inverse f −1 : Y → X.
(Item (b) is the Axiom of Choice.)

351
352 APPENDIX A. NOTES

The analogous principle holds for linear maps: if A ∈ Rm×n , then the
linear map Rn → Rm : x 7→ Ax is

(a) injective ⇐⇒ BA = 1ln for some B ∈ Rn×m ;

(b) surjective ⇐⇒ AB = 1lm for some B ∈ Rn×m ;

(c) bijective ⇐⇒ A is invertible (i.e. m = n and det(A) 6= 0).

(Here 1lk is the k × k identity matrix.) However, this principle fails com-
pletely for continuous maps: the map f : [0, 2π) → S 1 defined by f (θ) =
(cos θ, sin θ) is continuous and bijective but its inverse is not continuous.
(Here S 1 ⊂ R2 is the unit circle x2 + y 2 = 1.)

A.2 Normal Forms

The Fundamental Idea of Differential Calculus is that near a point x0 ∈ U
a smooth map f : U → V behaves like its linear approximation, i.e.

f (x) ≈ f (x0 ) + df (x0 )(x − x0 ).

The Normal Form Theorem from Linear Algebra says that if A ∈ Rm×n has
rank r, then there are invertible matrices P ∈ Rm×m and Q ∈ Rn×n such
that  
−1 1lr 0r×(n−r)
P AQ = .
0(m−r)×r 0(m−r)×(n−r)
By the Fundamental Idea we can expect an analogous theorem for smooth
maps.

Theorem A.2.1 (Local Normal Form for Smooth Maps). Let U ⊂ Rn

and V ⊂ Rm be open, x0 ∈ U , and f : U → V be smooth. Assume that the
derivative df (x0 ) ∈ Rm×n has rank r. Then there is an open neighborhood
U0 of x0 in U , an open neighborhood V0 of f (x0 ) in V , a diffeomorphism
φ : U1 × U2 ⊂ Rr × Rn−r , a diffeomorphism ψ : V0 → U1 × V2 ⊂ Rr × Rm−r ,
such that φ(x0 ) = (0, 0), ψ(f (x0 )) = (0, 0), and

ψ −1 ◦ f ◦ φ(x, y) = (x, g(x, y)) and dg(0, 0) = 0

for (x, y) ∈ U1 × U2 .

The Local Normal Form Theorem is an easy consequence of the Inverse

Function Theorem.
A.2. NORMAL FORMS 353

Theorem A.2.2 (Inverse Function Theorem). Let U ⊂ Rn , V ⊂ Rm ,

x0 ∈ U and f : U → V be a smooth map. If df (x0 ) is invertible, then (m = n
and) there are neighborhoods U0 of x0 in U and V0 of f (x0 ) in V so that the
restriction f|U0 : U0 → V0 is a diffeomorphism.

Here follow some other consequences of the Inverse Function Theorem.

The terms submersion and immersion are defined in §2.6.1 and Defini-
tion 2.3.2 of §2.3.

Corollary A.2.3 (Submersion Theorem). When r = m the diffeomor-

phisms φ and ψ in Theorem A.2.1 may be chosen so that the local normal
form is
ψ −1 ◦ f ◦ φ(x, y) = x.

Corollary A.2.4 (Immersion Theorem). When r = n the diffeomor-

phisms φ and ψ in Theorem A.2.1 may be chosen so that the local normal
form is
ψ −1 ◦ f ◦ φ(x) = (x, 0).

Corollary A.2.5 (Rank Theorem). If the rank of df (x) = r for all x ∈ U ,

then for every x0 ∈ U the diffeomorphisms φ and ψ in Theorem A.2.1 may
be chosen so that the local normal form is

ψ −1 ◦ f ◦ φ(x) = (x, 0).

Corollary A.2.6 (Implicit Function Theorem). Let U ⊂ Rm × Rn be

an open set, let F : U → Rn be smooth, and let (x0 , y0 ) ∈ U with x0 ∈ Rm
and y0 ∈ Rn . Define the partial derivative d2 F (x0 , y0 ) ∈ Rn×n by

d 
d2 F (x0 , y0 )v :=  F (x0 , y0 + tv)
dt t=0

for v ∈ Rn . Assume that F (x0 , y0 ) = 0 and that d2 F (x0 , y0 ) is invertible.

Then there exist neighborhoods U0 of x0 in Rm and V0 of y0 in Rn and a
smooth map g : U0 → V0 such that

U0 × V0 ⊂ U, g(x0 ) = y0

and
F (x, y) = 0 ⇐⇒ y = g(x)
for x ∈ U0 and y ∈ V0 .
354 APPENDIX A. NOTES

A.3 Euclidean Spaces

This is the arena of Euclidean geometry; i.e. every figure which is studied in
Euclidean geometry is a subset of Euclidean space. To define it one could
proceed axiomatically as Euclid did; one would then verify that the ax-
ioms characterized Euclidean space by constructing “Cartesian Coordinate
Systems” which identify the n-dimensional Euclidean space E n with the n-
dimensional numerical space Rn . This program was carried out rigorously
by Hilbert. We shall adopt the mathematically simpler but philosophically
less satisfying course of taking the characterization as the definition.
We shall use three closely related spaces: n-dimensional Euclidean affine
space E n , n-dimensional Euclidean vector space En , and the space Rn of
all n-tuples of real numbers. The distinction among them is a bit pedantic,
especially if one views as the purpose of geometry the interpretation of
calculations on Rn . The purpose for distinguishing these three spaces is the
same as in elementary vector calculus; it aids geometric intuition. Here is
the precise definition.

Definition A.3.1. An n-dimensional Euclidean vector space is a real n-

dimensional vector space En equipped with a (real valued symmetric positive
definite) inner product En × En → R : (v, w) 7→ hv, wi. An n-dimensional
Euclidean affine space consists of a set E n and an n-dimensional Euclidean
vector space En and maps

E n × E n → En : (p, q) 7→ p − q,
E n × En → E n : (p, v) 7→ p + v

satisfying the axioms

p + 0 = p, p + (v + w) = (p + v) + w, q + (p − q) = p

for all p, q ∈ E n and all v, w ∈ En . The vector p − q ∈ En is called the

vector from q to p and the point p + v is called the translate of p by v.
It follows easily that each choice of a point o ∈ E n determines a bijection
v 7→ o + v from En onto E n . The inner product on En equips the space E n
with a metric via the formula
p
|p − q| = hp − q, p − qi, p, q ∈ En .

The standard Euclidean space of dimension n is E n = EnP = Rn with the

usual matrix algebra operations (x ± y) = x ± y , hx, yi = i xi y i .
i i i
A.3. EUCLIDEAN SPACES 355

Lemma A.3.2. Any choice of an origin o ∈ E n and an orthonormal ba-

sis e1 , . . . , en for En determines an isometric bijection:
n
X
n n 1 n
R → E : (x , . . . , x ) 7→ o + xi ei
i=1

(the inverse of which is called a Cartesian coordinate system on E n ).

Lemma A.3.3. If E n → Rn : p 7→ (x1 , . . . , xn ), (y 1 , . . . , y n ) are two Carte-

sian coordinate systems, the change of coordinates map has the form
n
aji xi (p) + v i ,
X
y j (p) =
j=1

where the matrix a = (aji ) ∈ Rn×n is an orthogonal matrix and v ∈ Rn .

Example A.3.4. Any n-dimensional affine subspace of some numerical

space Rk (with k > n) is an example of a Euclidean space. The corre-
sponding vector space En is the unique vector subspace of Rk for which:

E n = o + En

for o ∈ E n . This subspace is independent of the choice of o ∈ E n . Note

that En contains the “preferred” point 0 while E n has no preferred point.
Such spaces E n and En would arise in linear algebra by taking E n to be
the space of solutions of k − n independent inhomogeneous linear equations
in k unknowns while En is the space of solutions of the corresponding ho-
mogeneous equations. The correspondence between E n and En illustrates
the mantra

The general solution of an inhomogeneous system of linear equa-

tions is a particular solution plus the general solution of the cor-
responding homogeneous linear system.

This discussion shows that a Euclidean space E n is an n-dimensional

manifold with its Cartesian coordinate systems whose tangent space at each
point is naturally isomorphic to En . Thus it is natural to introduce sub-
manifolds of Euclidean space as submanifolds of E n whose tangent spaces
are then linear subspaces of the associated vector space En . Instead we have
chosen in this book for simplicity of the exposition to describe manifolds as
subsets of the vector space Rn equipped with its standard inner product.
356 APPENDIX A. NOTES
Bibliography

[1] Ralph Herman Abraham & Joel William Robbin, Transversal Mappings and
Flows. Benjamin Press, 1967.

[2] Lars Valerian Ahlfors & Leo Sario, Riemann Surfaces. Princeton University
Press, 1960.

[3] Michael Francis Atiyah & Nigel James Hitchin & Isodore Manuel Singer, Self-
duality in four-dimensional Riemannian geometry. Prococeedings of the Royal
Society of London, Series A 362 (1978), 425–461.

[4] Thierry Émilien Flavien Aubin, Équations différentielles non linéaires et

problème de Yamabe concernant la courbure scalaire. Journal de Mathéma-
tiques Pures et Appliquées 55 (1976), 269–296.

[5] Marcel Berger, Les variétés Riemanniennes 1/4-pincées, Annali della Scuola
Normale Superiore di Pisa 14 (1960) , 161–170.

[6] Simon Brendle, Ricci Flow and the Sphere Theorem. Graduate Studies in
Mathematics 111. AMS, Providence, RI, 2010.

[7] Simon Brendle & Richard Melvin Schoen, Manifolds with 1/4-pinched curva-
ture are space forms. Journal of the American Mathematical Society 22 (2009),
287–307. https://arxiv.org/pdf/0705.0766.pdf

[8] Simon Brendle & Richard Melvin Schoen, Curvature, Sphere Theorems, and
the Ricci Flow. Bulletin of the American Mathematical Society 48 (2011),
1–32. https://arxiv.org/pdf/1001.2278.pdf

[9] Élie Joseph Cartan, La théorie des groupes finis et continus et l’Analysis Situs.
Mémorial des Sciences Mathématiques 42 (1930), 1–61.

[10] Bill Casselmann, Symmetric spaces of semi-simple groups. Essays on repre-

sentations of real groups, August, 2012. http://www.math.ubc.ca/~cass/
research/pdf/Cartan.pdf

357
358 BIBLIOGRAPHY

[11] Simon Kirwan Donaldson, Lie algebra theory without algebra. In “Alge-
bra, Arithmetic, and Geometry, In honour of Yu. I. Manin”, edited by
Yuri Tschinkel and Yuri Zarhin, Progress in Mathematics, Birkhäuser 269,
2009, pp. 249–266. https://arxiv.org/pdf/math/0702016.pdf

[12] Simon Kirwan Donaldson, Towords enumerative theories for structures on 4-

manifolds. Seminar, 15 September 2020. https://www.youtube.com/watch?
v=44hwQKQggbA

[13] David Bernard Alper Epstein, Foliations with all leaves compact. Annales de
l’institute Fourier 26 (1976), 265–282.

[14] Mikhael Leonidovich Gromov, Pseudoholomorphic curves in symplectic man-

ifolds. Inventiones Mathematicae 82 (1985), 307–347.

[15] Michael Leonidovich Gromov & Herbert Blaine Lawson, The classification of
simply connected manifolds of positive scalar curvature, Annals of Mathemat-
ics 111 (1980), 423–434.

[16] Victor Guillemin & Alan Pollack, Differential Topology. Prentice-Hall, 1974.

[17] Allen Hatcher, Algebraic Topology, Cambridge, 2002.

[18] Sigurour Helgason, Differential Geometry, Lie Groups, and Symmetric Spaces.
Graduate Studies in Mathematics 35, AMS, 2001.

[19] Noel Justin Hicks, Notes on Differential Geometry, Van Nostrand Rheinhold
Mathematical Studies, 1965.

[20] David Hilbert & Stephan Cohn-Vossen, Geometry and the Imagination.
Chelsea Publishing Company, 1952.

[21] Joachim Hilgert & Karl-Hermann Neeb, Structure and Geometry of Lie groups.
Springer Monographs in Mathematics, Springer-Verlag, 2012.

[22] Morris William Hirsch, Differential Topology. Springer, 1994.

[23] Nigel James Hitchin, Harmonic spinors. Advances in Mathematics 14 (1974),

1–55.

[24] Heinz Hopf, Zum Clifford–Kleinschen Raumproblem. Mathematische Annalen

95 (1926), 313–339.

[25] Heinz Hopf, Differentialgeometrie und topologische Gestalt. Jahresbericht der

Deutschen Mathematiker-Vereinigung 41 (1932), 209–229.

[26] Ralph Howard, Kuiper’s theorem on conformally flat manifolds. Lecture Notes,
University of South Caroline, Columbia, July 1996. http://ralphhoward.
github.io/SemNotes/Notes/conformal.pdf
BIBLIOGRAPHY 359

[27] Mustafa Kalafat, Locally conformally flat and self-dual structures on simple
4-manifolds. Proceedings of the Gökova Geometry-Topology Conference 2012,
International Press, Somerville, MA, 2013, pp 111–122. http://gokovagt.
org/proceedings/2012/ggt12-kalafat.pdf

[28] John Leroy Kelley, General Topology. Graduate Texts in Mathematics, 27.
Springer 1975.

[29] Wilhelm Karl Joseph Killing, Über die Grundlagen der Geometrie. Journal
für die reine und angewandte Mathematik 109 (1892), 121–186.

[30] Wilhelm Klingenberg, Über Riemannsche Mannigfaltigkeiten mit positiver

Krümmung. Commentarii Mathematici Helvetici 35 (1961), 47–54.

[31] Nicolaas Hendrik Kuiper, On conformally flat spaces in the large. Annals of
Mathematics 50 (1949), 916–924.

[32] Wolfgang Kühnel, Differential geometry: Curves-surfaces-manifolds, Second

Edition. Student Mathematical Library, 16, AMS, Providence, RI, 2006.

[33] Claude LeBrun, On the topology of self-dual 4-manifolds. Proceedings of the

American Mathematical Society 98 (1986), 637–640.

[34] André Lichnerowicz, Spineurs harmoniques, Comptes rendus de l’Académie

des Sciences, Série A-B 257 (1963), 7–9.

[35] Anatolij Ivanovich Malcev, On the theory of the Lie groups in the large.
Matematichevskii Sbornik N. S. 16 (1945), 163–189.
Corrections to my paper ”On the theory of Lie groups in the large”. Matem-
atichevskii Sbornik N. S. 19 (1946), 523–524.

[36] Dusa McDuff & Dietmar Arno Salamon, Introduction to Symplectic Topology,
Third Edition. Oxford University Press, 2017.

[37] John Willard Milnor, Topology from the Differentiable Viewpoint. The Uni-
versity Press of Virginia 1969.

[38] James Raymond Munkres, Topology, Second Edition. Prentice Hall, 2000.

[39] Sumner Byron Myers & Norman Earl Steenrod, The group of isometries of a
Riemannian manifold. Annals of Mathematics 40 (1939), 400–416.

[40] John von Neumann, Über die analytischen Eigenschaften von Gruppen lin-
earer Transformationen und ihrer Darstellungen. Mathematische Zeitschrift
30 (1929), 3–42.

[41] Takushiro Ochiai & Tsunero Takahashi, The group of isometries of a left
invariant Riemannian metric on a Lie group. Mathematische Annalen 223
(1976), 91–96.
360 BIBLIOGRAPHY

[42] Roger Penrose, Nonlinear gravitons and curved twistor theory. The riddle of
gravitation – on the occasion of the 60th birthday of Peter G. Bergmann
(Proc. Conf., Syracuse Univ., Syracuse, NY., 1975), General Relativity and
Gravitation 7 (1976), 31–52.

[43] Harry Earnest Rauch, A contribution to differential geometry in the large.

Annals of Mathematics 54 (1951), 38–55.

[44] Joel William Robbin, Extrinsic Differential Geometry. 1982.

http://www.math.wisc.edu/~robbin/0geom.pdf

[45] Mary Ellen Rudin, A new proof that metric spaces are paracompact. Proceed-
ings of the American Mathematical Society 20 (1969), 603.

[46] Dietmar Arno Salamon, Analysis I. Lecture Course, ETHZ, HS 2016.

[47] Dietmar Arno Salamon, Analysis II. Lecture Course, ETHZ, FS 2017.
https://people.math.ethz.ch/~salamon/PREPRINTS/analysis2.pdf

[48] Dietmar Arno Salamon, Spin Geometry and Seiberg–Witten invariants. Un-
published Manuscript, 1999. https://people.math.ethz.ch/~salamond/
PREPRINTS/witsei.pdf

[49] Dietmar Arno Salamon, Measure and Integration. European Mathematical

Society, Textbooks in Mathematics, 2016. https://people.math.ethz.ch/
~salamond/PREPRINTS/measure.pdf
[50] Dietmar Arno Salamon, Uniqueness of symplectic structures. Acta Mathemat-
ica Vietnamica 38 (2013), 123–144.

[51] Dietmar Arno Salamon & Thomas Walpuski, Notes on the Octonions. Pro-
ceedings of the 23rd Gökova Geometry-Topology Conference 2016, edited
by S. Akbulut, D. Auroux, and T. Önder, International Press, Somerville,
Massachusetts, 2017, pp 1-85. https://people.math.ethz.ch/~salamon/
PREPRINTS/Octonions.pdf

[52] Richard Melvin Schoen, Conformal deformation of a Riemannian metric to

constant scalar curvature. Journal of Differerential Geometry 20 (1084), 479–
495.

[53] Richard Melvin Schoen & Shing-Tung Yau, On the structure of manifolds with
positive scalar curvature. Manuscripta Mathematica 28 (1979), 159–183.

[54] Stephen Smale, Diffeomorphisms of the 2-sphere. Proceedings of the American

Mathematical Society 10 (1959), 621–626.

[55] Clifford Henry Taubes, Self-dual Yang-Mills connections on non-self-dual 4-

manifolds. Journal of Differential Geometry 17 (1982), 139–170.
BIBLIOGRAPHY 361

[56] Clifford Henry Taubes, SW =⇒ Gr: From the Seiberg–Witten equations to

pseudoholomorphic curves. Journal of the American Mathematical Society 9
(1996), 845–918.

[57] Neil Sidney Trudinger, Remarks concerning the conformal deformation of Rie-
mannian structures on compact manifolds. Annali della Scuola Normale Su-
periore di Pisa 22 (1968), 265–274.

[58] Hidehiko Yamabe, On a deformation of Riemannian structures on compact

manifolds. Osaka Journal of Mathematics 12 (1960), 21–37.
Index

affine subspace, 159 Cartan–Hadamard Theorem, 320

anti-self-dual Cartesian coordinate system, 355
four-manifold, 349 category, 12
atlas, 4, 87 catenoid, 300
maximal, 9, 11 chain rule, 8, 10, 15, 29
smooth, 11 Christoffel symbols, 162, 172
topological, 8 Closed Subgroup Theorem, 62
automorphism compact-open topology, 256
of a Lie algebra, 58, 61 compatible, 87
of a Lie group, 58 smoothly, 11
of F (M ), 69 topologically, 8
Axiom of Choice, 351 complete, 145
geodesically, 209
basic vector field, 145, 173 vector field, 42
Bianchi identity complex projective space, 89
first, 239 conformal
in local coordinates, 282 chart, 2
second, 304 class of metrics, 341
in local coordinates, 306 diffeomorphism, 345
Bonnet–Myers Theorem, 334 invariance of Weyl tensor, 345
Brouwer, 8 conformally flat, 345
bundle connection, 240
frame, 138 Levi-Civita, 240
principal, 115, 140 Riemannian, 240
tangent, 74, 97 torsion-free, 240
vector, 74 constant
scalar curvature, 340
Cartan Fixed Point Theprem, 326 sectional curvature, 309
Cartan–Ambrose–Hicks Theorem Convex Neighborhood Theorem, 205
global, 287 coordinate chart, 4, 11, 16, 87
weak version, 292 frame bundle, 173
local, 293 geodesically normal, 206

362
INDEX 363

Monge, 28 of the length functional, 179

smoothly compatible, 11, 87 developable manifold, 297
tangent bundle, 74, 97 osculating, 298
topologically compatibe, 8 tangent, 299
countable base, 101 development, 154
covariant derivative diffeomorphic, 15, 92
along a curve, 127, 171 diffeomorphism, 10, 15, 30, 92
of a global vector field, 234 conformal, 345
of a normal vector field, 135 local, 33
of the curvature, 303 dimension, 4
partial, 232 distance function, 183
critical point intrinsic setting, 217
of the energy functional, 179
of the length functional, 179 Einstein metric, 342
curvature ellipsoid, 5
Gaußian, 251, 252 elliptic hyperboloid, 299
in local coordinates, 283 embedding, 34
Ricci, 333 M into Euclidean space, 107
in local coordinates, 333 Plücker, 96
positive, 334 RP2 into R6 , 22
Riemann, 232 Veronese, 96
in local coordinates, 282 energy of a curve, 176
scalar, 337 Euclidean space, 354
constant, 340 Euler’s curvature, 251
in local coordinates, 339 Euler–Lagrange equations, 196
positive, 340 exponential map, 192
sectional, 308 distance increasing, 320
constant, 309 exponential matrix, 55
nonpositive, 320 extremal
positive, 336 of the energy functional, 177
Weyl, 343 of the length functional, 177
in local coordinates, 345
Curve Shortening Lemma, 213 fiber of a vector bundle, 74
intrinsic setting, 220 first fundamental form, 122
flat, 296
derivation flow, 41
on a Lie algebra, 61 foliation, 85
on F (M ), 71 closed leaf, 113
derivative, 10, 28 foliation box, 80
of the energy functional, 179 four-manifold
364 INDEX

self-dual, 349 fixed point, 326

frame bundle, 138, 173 free, 114, 137
coordinate chart, 173 infinitesimal, 68, 114
horizontal tangent vector, 173 orbit space, 114
orthonormal, 140 smooth, 68, 114
Frobenius’ theorem, 80 transitive, 137
Fubini–Study metric, 169 groupoid, 255
functorial, 70 smooth, 255
fundamental form
first, 122 Hadamard manifold, 320
second, 125 Hausdorff space, 91
Heisenberg algebra, 58
Gauß Lemma, 199 helicoid, 300
Gauß map, 123, 250 homogeneous space, 119, 302
Gauß–Codazzi formula, 233 homomorphism
Gauß–Weingarten formula of algebras, 69
along a curve, 127 of Lie algebras, 58
for global vector fields, 234 of Lie groups, 58
for normal vector fields, 135 homotopy, 285
Gaußian curvature, 251, 252, 308 with fixed endpoints, 286
in local coordinates, 283 Hopf fibration, 220
general linear group, 23, 57 Hopf–Rinow Theorem, 210
geodesic, 177 horizontal
broken, 204 curve, 142
in local coordinates, 196 lift, 142
intrinsic setting, 219 subbundle, 173
minimal, 2, 197 tangent space, 142, 173
global existence, 210 hyperbolic space, 314
local existence, 199 Poincaré model, 2, 318
spray, 191 upper half plane, 319
geodesically hyperboloid, 299
complete, 209 hypersurface, 250
convex, 205
normal coordinates, 206 ideal
Grönwall’s inequality, 263 in F (M ), 70
Grassmann manifold, 7, 90 immersion, 34
great circle, 2 Implicit Function Theorem, 79, 353
group action, 68 injectivity radius, 194
contravariant, 137 integrable subbundle of T M , 80
finite isotropy, 114 integral curve, 39
INDEX 365

intrinsic infinitesimal action, 68, 114

distance, 183, 217 invariant inner product, 242
topology, 86, 88 isomorphism, 58
Invariance of Domain, 8 of a Lie group, 55
Inverse Function Theorem, 31 of vector fields, 50
involutive subbundle of T M , 80 Lie bracket of vector fields, 49, 99
isometry, 224 Lie group, 52
group, 224, 256 automorphism, 58
local, 230 bi-invariant metric, 242
of a Lie group, 273 center, 66
isomorphism, 12 group action, 68
of Lie algebras, 58 homomorphism, 58
of Lie groups, 58 isomorphism, 58
isotopy, 45 right invariant metric, 243
isotropy subgroup, 114 symplectic, 53
Lie subgroup, 62
Jacobi field, 294
lift of a curve, 142
Jacobi identity, 49
local coordinates, 4
Jacobian matrix, 10
local diffeomorphism, 33
Killing form, 242 local isometry, 230
nondegenerate, 243, 273 local trivialization, 76
Killing vector field, 260 locally finite
complete, 261, 262 open cover, 101
Klein bottle, 112 sum, 106
Kuiper’s Theorem, 346 locally symmetric space, 301

Laplace–Beltrami operator, 341 manifold, 8, 11, 16, 87

leaf of a foliation, 85, 113 complex, 13
closed, 113 developable, 297
extrinsic topology, 86 dimension, 8
intrinsic topology, 86 flat, 296
length of a curve, 176 oriented, 346
Levi-Civita connection, 171, 235 osculating developable, 298
axioms, 171 real analytic, 13
Lie algebra, 50 smooth, 11, 16, 87
automorphism, 58, 61 symplectic, 13
center, 66 tangent developable, 299
derivation, 61 topological, 8, 88
Heisenberg, 58 maximal existence interval, 41
homomorphism, 50, 58 for geodesics, 192
366 INDEX

maximal ideal in F (M ), 70 projection

maximal smooth atlas, 11 obvious, 9
maximal topological atlas, 9 orthogonal, 75
mean curvature, 300 projective space, 7
metrizable, 101 complex, 89
minimal geodesic, 197 real, 89
existence, 199, 210 proper, 34
minimal surface, 300 pullback
Möbius strip, 78, 123, 300 of a vector bundle, 76
paper model, 299 of a vector field, 46
Monge coordinates, 28 pushforward
motion, 146 of a vector field, 46
without sliding, 149
without twisting, 151 quadric surface, 5
without wobbling, 152 Quaternions, 53
moving frame, 139, 142 quotient space, 9
Myers–Steenrod Theorem, 256 quotient topology, 9

normal bundle, 76 real projective space, 22, 89

refinement, 101
1-form with values in T M , 240
regular value, 20, 32
oriented manifold, 346
relative topology, 9
orthogonal group, 23
relatively closed, 9
paraboloid, 299 relatively open, 9
elliptic, 6 Ricci tensor, 333
hyperbolic, 6 in local coordinates, 333
paracompact, 101 positive, 334
parallel Riemann curvature tensor, 232
transport, 131 covariant derivative, 303
vector field, 129 in local coordinates, 306
parametrized first Bianchi identity, 239
by the arclength, 182 in local coordinates, 282
proportional to arclength, 177 second Bianchi identity, 304
partition of unity, 103 Riemann surface, 102
existence, 103 Riemannian manifold, 166
subordinate to a cover, 103 Einstein, 342
Plücker embedding, 96 flat, 296
Plücker’s conoid, 299 homogeneous, 302
Poincaré metric, 318 locally symmetric, 301
principal bundle, 115, 140 symmetric, 301
INDEX 367

Riemannian metric, 13, 166 spray, 191

bi-invariant, 242 stereographic projection, 3
conformal class, 341 structure group, 140
distance function, 183, 217 subbundle, 80, 99
existence, 168 integrable, 80
Roman surface, 22 involutive, 80
ruled surface, 299 submanifold, 34, 95
totally geodesic, 329
scalar curvature, 337 submersion, 72
constant, 340 surface, 252
Yamabe problem, 341 quadric, 5
in local coordinates, 339 symmetric space, 301
positive, 340 is homogeneous, 302
second Bianchi identity, 304 locally, 301
second countable, 101
second fundamental form, 125 tangent bundle, 74, 97
section of a vector bundle, 74 coordinate chart, 74, 97
sectional curvature, 308 integrable subbundle, 80
constant, 309 involutive subbundle, 80
nonpositive, 320 of the frame bundle, 141
positive, 336 of the tangent bundle, 191
self-dual tangent space, 24, 93
four-manifold, 349 horizontal, 142, 173
simply connected, 286 of the frame bundle, 141
singular value, 20 of the tangent bundle, 191
Smirnov Metrization Theorem, 102 vertical, 142
smooth, 10 tangent vector, 24, 93
atlas, 11 Theorema Egregium, 245, 254
group action, 68 topological atlas, 8
groupoid, 255 topological manifold, 8, 88
manifold, 11, 16, 87 topological space
map, 10, 15, 92 countable base, 101
structure, 12, 88 Hausdorff, 91
vector bundle, 74 locally compact, 102
smoothly compatible, 11 metrizable, 101
special normal, 102
linear group, 23, 57 paracompact, 101
unitary group, 57 regular, 102
sphere, 5 second countable, 101
Sphere Theorem, 336 σ-compact, 101
368 INDEX

topologically compatibe, 8 Whitney Embedding Theorem, 112

topology
compact-open, 256 Yamabe problem, 341
intrinsic, 88
quotient, 9
relative, 9
torus, 90
totally geodesic submanifold, 329
transition map, 8, 11, 87
2-form, 346
anti-self-dual, 347
self-dual, 347
Tychonoff’s Lemma, 102

uniformization theorem, 340

unitary group, 57
Urysohn Metrization Theorem, 102

variation of a curve, 176

vector bundle, 74
canonical projection, 74
local trivialization, 76
section, 74
vector field, 38, 97
along a curve, 127
along a map, 170
basic, 145, 173
flow, 41
integral curve, 39
Killing, 260
Lie bracket, 49
normal, 135
parallel, 129
with compact support, 43
Veronese embedding, 96
vertical tangent space, 142

wallpaper group, 280

Weyl tensor, 343
in local coordinates, 345
self-dual, 349