An Introduction to

Riemannian Geometry
with Applications to Mechanics and Relativity
Leonor Godinho and Jose Natario
Lisbon, 2012
Contents
Preface 3
Chapter 1. Dierentiable Manifolds 5
1. Topological Manifolds 5
2. Dierentiable Manifolds 11
3. Dierentiable Maps 15
4. Tangent Space 17
5. Immersions and Embeddings 24
6. Vector Fields 29
7. Lie Groups 36
8. Orientability 48
9. Manifolds with Boundary 51
10. Notes on Chapter 1 54
Chapter 2. Dierential Forms 61
1. Tensors 61
2. Tensor Fields 68
3. Dierential Forms 70
4. Integration on Manifolds 76
5. Stokes Theorem 79
6. Orientation and Volume Forms 82
7. Notes on Chapter 2 84
Chapter 3. Riemannian Manifolds 91
1. Riemannian Manifolds 91
2. Ane Connections 98
3. Levi-Civita Connection 101
4. Minimizing Properties of Geodesics 107
5. Hopf-Rinow Theorem 114
6. Notes on Chapter 3 117
Chapter 4. Curvature 119
1. Curvature 119
2. Cartan Structure Equations 126
3. Gauss-Bonnet Theorem 135
4. Manifolds of Constant Curvature 141
5. Isometric Immersions 148
1
2 CONTENTS
6. Notes on Chapter 4 155
Chapter 5. Geometric Mechanics 157
1. Mechanical Systems 157
2. Holonomic Constraints 166
3. Rigid Body 170
4. Non-Holonomic Constraints 184
5. Lagrangian Mechanics 193
6. Hamiltonian Mechanics 201
7. Completely Integrable Systems 209
8. Symmetry and Reduction 216
9. Notes on Chapter 5 230
Chapter 6. Relativity 233
1. Galileo Spacetime 233
2. Special Relativity 235
3. The Cartan Connection 245
4. General Relativity 247
5. The Schwarzschild Solution 252
6. Cosmology 263
7. Causality 268
8. Hawking Singularity Theorem 276
9. Penrose Singularity Theorem 288
10. Notes on Chapter 6 295
Bibliography 297
Index 299
Preface
This book is based on a one-semester course taught since 2002 at In-
stituto Superior Tecnico (Lisbon) to mathematics, physics and engineering
students. Its aim is to provide a quick introduction to dierential geometry,
including dierential forms, followed by the main ideas of Riemannian geom-
etry (minimizing properties of geodesics, completeness and curvature). Pos-
sible applications are given in the nal two chapters, which have themselves
been independently used for one-semester courses on geometric mechanics
and general relativity. We hope that these will give mathematics students
a chance to appreciate the usefulness of Riemannian geometry, and physics
and engineering students an extra motivation to learn the mathematical
background.
It is assumed that the reader has a basic knowledge of linear algebra,
multivariable calculus and dierential equations, as well as elementary no-
tions of topology and algebra. For the convenience of the reader (especially
physics and engineering students), we have summarized the main deni-
tions and results from this background material at the end of each chapter
as needed.
To help the reader test and consolidate his/her understanding, and also
to introduce important ideas and examples not treated in the main text, we
have included more than 300 exercises. The detailed solutions are available
in a companion volume.
We now give a short description of the contents of each chapter.
Chapter 1 discusses the basic concepts of dierential geometry: dier-
entiable manifolds and maps, vector elds and the Lie bracket. In addition,
we give a brief overview of Lie groups and Lie group actions.
Chapter 2 is devoted to dierential forms, covering the standard topics:
wedge product, pull-back, exterior derivative, integration and the Stokes
Theorem.
Riemannian manifolds are introduced in Chapter 3, where we treat the
Levi-Civita connection, minimizing properties of geodesics and the Hopf-
Rinow theorem.
Chapter 4 addresses the notion of curvature. In particular, we use the
powerful computational method given by the Cartan structure equations
to prove the Gauss-Bonnet Theorem. Constant curvature and isometric
embeddings are also discussed.
3
4 PREFACE
Chapter 5 gives an overview of geometric mechanics, including holo-
nomic and non-holonomic systems, Lagrangian and Hamiltonian mechanics,
completely integrable systems and reduction.
Chapter 6 treats general relativity, starting with a geometric introduc-
tion to special relativity. The Einstein equation is motivated via the Cartan
connection formulation of Newtonian gravity, and the basic examples of
the Schwarzschild solution (including black holes) and cosmology are stud-
ied. We conclude with a discussion of causality and the celebrated Hawking
and Penrose Singularity Theorems, which, although unusual in introductory
texts, are very interesting applications of Riemannian geometry.
Finally, we want to thank the many colleagues and students who read
this text, or parts of it, for their many comments and corrections. Special
thanks are due to our colleague and friend Pedro Gir ao.
CHAPTER 1
Dierentiable Manifolds
This chapter introduces the basic notions of dierential geometry.
The rst section studies topological manifolds of dimension n, which
is the rigorous mathematical concept corresponding to the intuitive notion
of continuous n-dimensional spaces. Several examples are discussed, par-
ticularly in dimension 2 (surfaces).
Section 2 specializes to dierentiable manifolds, on which one can
dene dierentiable functions (Section 3) and tangent vectors (Section
4). Important examples of dierentiable maps, namely immersions and
embeddings, are examined in Section 5.
Vector elds and their ows are the main topic of Section 6. It is
shown that there is a natural dierential operation between vector elds,
called the Lie bracket, which plays an important role in dierential geom-
etry.
Section 7 is devoted to the important class of dierentiable manifolds
which are also groups, the so-called Lie groups. It is shown that to each
Lie group one can associate a Lie algebra, i.e. a vector space equipped
with a Lie bracket, which essentially determines the group. Quotients of
manifolds by Lie groups are also treated.
Orientability of a manifold (which generalizes the intuitive notion of
having two sides) and manifolds with boundary are studied Sections
8 and 9. Both these notions are necessary to formulate the Stokes Theorem,
which will be studied in Chapter 2.
1. Topological Manifolds
We will begin this section by studying spaces that are locally like R
n
,
meaning that there exists a neighborhood around each point which is home-
omorphic to an open subset of R
n
.
Definition 1.1. A topological manifold M of dimension n is a topo-
logical space with the following properties:
(i) M is Hausdor, that is, for each pair p
1
, p
2
of distinct points of M
there exist neighborhoods V
1
, V
2
of p
1
and p
2
such that V
1
V
2
= .
(ii) Each point p M possesses a neighborhood V homeomorphic to an
open subset U of R
n
.
(iii) M satises the second countability axiom, that is, M has a
countable basis for its topology.
5
6 1. DIFFERENTIABLE MANIFOLDS
Conditions (i) and (iii) are included in the denition to prevent the
topology of these spaces from being too strange. In particular, the Hausdor
axiom ensures that the limit of a convergent sequence is unique. This, along
with the second countability axiom, guarantees the existence of partitions of
unity (cf. Section 7.2 of Chapter 2), which, as we will see, are a fundamental
tool in dierential geometry.
Remark 1.2. If the dimension of M is zero then M is a countable set
equipped with the discrete topology (every subset of M is an open set).
If dimM = 1, then M is locally homeomorphic to an open interval; if
dimM = 2, then it is locally homeomorphic to an open disk, etc.
(a)
(b)
(c)
Figure 1. (a) S
1
; (b) S
2
; (c) Torus of revolution.
Example 1.3.
(1) Every open subset M of R
n
with the subspace topology (that is,
U M is an open set if and only if U = M V with V an open
set of R
n
) is a topological manifold.
(2) (Circle) The circle
S
1
= (x, y) R
2
: x
2
+y
2
= 1
with the subspace topology is a topological manifold of dimension
1. Conditions (i) and (iii) are inherited from the ambient space.
Moreover, for each point p S
1
there is at least one coordinate axis
which is not parallel to the vector n
p
normal to S
1
at p. The projec-
tion on this axis is then a homeomorphism between a (suciently
small) neighborhood V of p and an interval in R.
(3) (2-sphere) The previous example can be easily generalized to show
that the 2-sphere
S
2
= (x, y, z) R
3
: x
2
+y
2
+z
2
= 1
1. TOPOLOGICAL MANIFOLDS 7
with the subspace topology is a topological manifold of dimension
2.
(4) (Torus of revolution) Again as in the previous examples, we can
show that the surface of revolution obtained by revolving a circle
around an axis that does not intersect it is a topological manifold
of dimension 2.
(5) The surface of a cube is a topological manifold (homeomorphic to
S
2
).
Example 1.4. We can also obtain topological manifolds by identifying
edges of certain polygons by means of homeomorphisms. The edges of a
square, for instance, can be identied in several ways (see Figures 2 and 3):
(1) (Torus) The torus T
2
is the quotient of the unit square Q =
[0, 1]
2
R
2
by the equivalence relation
(x, y) (x + 1, y) (x, y + 1),
equipped with the quotient topology (cf. Section 10.1).
(2) (Klein bottle) The Klein bottle K
2
is the quotient of Q by the
equivalence relation
(x, y) (x + 1, y) (1 x, y + 1).
(3) (Projective plane) The projective plane RP
2
is the quotient of Q
by the equivalence relation
(x, y) (x + 1, 1 y) (1 x, y + 1).
(a)
(b)

=
Figure 2. (a) Torus (T
2
); (b) Klein bottle (K
2
).
Remark 1.5.
8 1. DIFFERENTIABLE MANIFOLDS

=

=
Figure 3. Projective plane (RP
2
).
(1) The only compact connected 1-dimensional topological manifold is
the circle S
1
(see [Mil97]).
(2) The connected sum of two topological manifolds M and N is
the topological manifold M#N obtained by deleting an open set
homeomorphic to a ball on each manifold and gluing the bound-
aries, which must be homeomorphic to spheres, by a homeomor-
phism (cf. Figure 4). It can be shown that any compact con-
nected 2-dimensional topological manifold is homeomorphic either
to S
2
or to connected sums of manifolds from Example 1.4 (see
[Blo96, Mun00]).
#

=
Figure 4. Connected sum of two tori.
If we do not identify all the edges of the square, we obtain a cylinder or
a Mobius band (cf. Figure 5). These topological spaces are examples of
manifolds with boundary.
Definition 1.6. Consider the closed half space
H
n
= (x
1
, . . . , x
n
) R
n
: x
n
0.
A topological manifold with boundary is a Hausdor space M, with a
countable basis of open sets, such that each point p M possesses a neigh-
borhood V which is homeomorphic either to an open subset U of H
n
H
n
,
or to an open subset U of H
n
, with the point p identied to a point in H
n
.
1. TOPOLOGICAL MANIFOLDS 9
(a)
(b)

=
Figure 5. (a) Cylinder; (b) Mobius band.
The points of the rst type are called interior points, and the remaining
are called boundary points.
The set of boundary points M is called the boundary of M, and is a
manifold of dimension (n 1).
Remark 1.7.
1. Making a paper model of the Mobius band, we can easily verify
that its boundary is homeomorphic to a circle (not to two disjoint
circles), and that it has only one side (cf. Figure 5).
2. Both the Klein bottle and the real projective plane contain Mobius
bands (cf. Figure 6). Deleting this band on the projective plane, we
obtain a disk (cf. Figure 7). In other words, we can glue a Mobius
band to a disk along their boundaries and obtain RP
2
.
(a) (b)
Figure 6. Mobius band inside (a) Klein bottle; (b) Real
projective plane.
10 1. DIFFERENTIABLE MANIFOLDS

=
Figure 7. Disk inside the real projective plane.
Two topological manifolds are considered the same if they are homeo-
morphic. For example, spheres of dierent radii in R
3
are homeomorphic,
and so are the two surfaces in Figure 8. Indeed, the knotted torus can be
obtained by cutting the torus along a circle, knotting it and gluing it back
again. An obvious homeomorphism is then the one which takes each point
on the initial torus to its nal position after cutting and gluing.

=
Figure 8. Two homeomorphic topological manifolds.
Exercises 1.8.
(1) Which of the following sets (with the subspace topology) are topo-
logical manifolds?
(a) D
2
= (x, y) R
2
[ x
2
+y
2
< 1;
(b) S
2
p (p S
2
);
(c) S
2
p, q (p, q S
2
, p ,= q);
(d) (x, y, z) R
3
[ x
2
+y
2
= 1;
(e) (x, y, z) R
3
[ x
2
+y
2
= z
2
;
(2) Which of the manifolds above are homeomorphic?
(3) Show that the Klein bottle K
2
can be obtained by gluing two
Mobius bands together through a homeomorphism of the boundary.
(4) Show that:
(a) M#S
2
= M for any 2-dimensional topological manifold M;
(b) RP
2
#RP
2
= K
2
;
(c) RP
2
#T
2
= RP
2
#K
2
.
(5) A triangulation of a 2-dimensional topological manifold M is a
decomposition of M in a nite number of triangles (i.e. subsets
homeomorphic to triangles in R
2
) such that the intersection of any
two triangles is either a common edge, a common vertex or empty
2. DIFFERENTIABLE MANIFOLDS 11
(it is possible to prove that such a triangulation always exists). The
Euler characteristic of M is
(M) := V E +F,
where V , E and F are the number of vertices, edges and faces of
a given triangulation (it can be shown that this is well dened,
i.e. does not depend on the choice of triangulation). Show that:
(a) dding a vertex to a triangulation does not change (M);
(b) (S
2
) = 2;
(c) (T
2
) = 0;
(d) (K
2
) = 0;
(e) (RP
2
) = 1;
(f) (M#N) = (M) +(N) 2.
2. Dierentiable Manifolds
Recall that an n-dimensional topological manifold is a Hausdor space
with a countable basis of open sets such that each point possesses a neigh-
borhood homeomorphic to an open subset of R
n
. Each pair (U, ), where
U is an open subset of R
n
and : U (U) M is a homeomorphism of
U to an open subset of M, is called a parameterization. The inverse
1
is called a coordinate system or chart, and the set (U) M is called a
coordinate neighborhood. When two coordinate neighborhoods overlap,
we have formulas for the associated coordinate change (cf. Figure 9). The
idea to obtain dierentiable manifolds will be to choose a sub-collection of
parameterizations so that the coordinate changes are dierentiable maps.
M
W
U

R
n
R
n

Figure 9. Parameterizations and overlap maps.

Definition 2.1. An n-dimensional dierentiable or smooth mani-
fold is a topological manifold of dimension n and a family of parameteri-
zations

: U

M dened on open sets U

R
n
, such that:
12 1. DIFFERENTIABLE MANIFOLDS
(i) the coordinate neighborhoods cover M, that is,

(U

) = M;
(ii) for each pair of indices , such that
W :=

(U

(U

) ,= ,
the overlap maps

:
1

(W)
1

(W)

:
1

(W)
1

(W)
are C

;
(iii) the family / = (U

) is maximal with respect to (i) and (ii),

meaning that if
0
: U
0
M is a parameterization such that
1
0

and
1

0
are C

for all in /, then (U

0
,
0
) is in /.
Remark 2.2.
(1) Any family / = (U

) that satises (i) and (ii) is called a

C

-atlas for M. If / also satises (iii) it is called a maximal

atlas or a dierentiable structure.
(2) Condition (iii) is purely technical. Given any atlas / = (U

)
on M, there is a unique maximal atlas

/ containing it. In fact, we
can take the set

/ of all parameterizations that satisfy (ii) with
every parameterization on /. Clearly /

/, and one can easily
check that

/ satises (i) and (ii). Also, by construction,

/ is
maximal with respect to (i) and (ii). Two atlases are said to be
equivalent if they dene the same dierentiable structure.
(3) We could also have dened C
k
-manifolds by requiring the coordi-
nate changes to be C
k
-maps (a C
0
-manifold would then denote a
topological manifold).
Example 2.3.
(1) The space R
n
with the usual topology dened by the Euclidean met-
ric is a Hausdor space and has a countable basis of open sets. If,
for instance, we consider a single parameterization (R
n
, id), condi-
tions (i) and (ii) of Denition 2.1 are trivially satised and we have
an atlas for R
n
. The maximal atlas that contains this parameter-
ization is usually called the standard dierentiable structure
on R
n
. We can of course consider other atlases. Take, for instance,
the atlas dened by the parameterization (R
n
, ) with (x) = Ax
for a non-singular (n n)-matrix A. It is an easy exercise to show
that these two atlases are equivalent.
(2) It is possible for a manifold to possess non-equivalent atlases: con-
sider the two atlases (R,
1
) and (R,
2
) on R, where
1
(x) = x
and
2
(x) = x
3
. As the map
1
2

1
is not dierentiable at the
origin, these two atlases dene dierent (though, as we shall see, dif-
feomorphic) dierentiable structures (cf. Exercises 2.5.4 and 3.2.6).
2. DIFFERENTIABLE MANIFOLDS 13
(3) Every open subset V of a smooth manifold is a manifold of the same
dimension. Indeed, as V is a subset of M, its subspace topology
is Hausdor and admits a countable basis of open sets. Moreover,
if / = (U

) is an atlas for M and we take the U

s for
which

(U

) V ,= , it is easy to check that the family of

parameterizations

/ = (

U
), where

U

=
1

(V ), is an
atlas for V .
(4) Let M
nn
be the set of n n matrices with real coecients. Re-
arranging the entries along one line, we see that this space is
just R
n
2
, and so it is a manifold. By Example 3, we have that
GL(n) = A M
nn
[ det A ,= 0 is also a manifold of dimension
n
2
. In fact, the determinant is a continuous map from M
nn
to R,
and GL(n) is the preimage of the open set R0.
(5) Let us consider the n-sphere
S
n
= (x
1
, . . . , x
n+1
) R
n+1
[ (x
1
)
2
+ + (x
n+1
)
2
= 1
and the maps

+
i
: U R
n
S
n
(x
1
, . . . , x
n
) (x
1
, . . . , x
i1
, g(x
1
, . . . , x
n
), x
i
, . . . , x
n
),

i
: U R
n
S
n
(x
1
, . . . , x
n
) (x
1
, . . . , x
i1
, g(x
1
, . . . , x
n
), x
i
, . . . , x
n
),
where
U = (x
1
, . . . , x
n
) R
n
[ (x
1
)
2
+ + (x
n
)
2
< 1
and
g(x
1
, . . . , x
n
) = (1 (x
1
)
2
(x
n
)
2
)
1
2
.
Being a subset of R
n+1
, the sphere (equipped with the subspace
topology) is a Hausdor space and admits a countable basis of open
sets. It is also easy to check that the family (U,
+
i
), (U,

i
)
n+1
i=1
is
an atlas for S
n
, and so this space is a manifold of dimension n (the
corresponding charts are just the projections on the hyperplanes
x
i
= 0).
(6) We can dene an atlas for the surface of a cube Q R
3
making
it a smooth manifold: Suppose the cube is centered at the origin
and consider the map f : Q S
2
dened by f(x) = x/|x|. Then,
considering an atlas (U

) for S
2
, the family (U

, f
1

)
denes an atlas for Q.
Remark 2.4. There exist topological manifolds which admit no dier-
entiable structures at all. Indeed, Kervaire presented the rst example (a
10-dimensional manifold) in 1960 [Ker60], and Smale constructed another
one (of dimension 12) soon after [Sma60]. In 1956 Milnor [Mil07] had
14 1. DIFFERENTIABLE MANIFOLDS
already given an example of a 8-manifold which he believed not to admit a
dierentiable structure, but that was not proved until 1965 (see [Nov65]).
Exercises 2.5.
(1) Show that two atlases /
1
and /
2
for a smooth manifold are equiv-
alent if and only if /
1
/
2
is an atlas.
(2) Let M be a dierentiable manifold. Show that a set V M is open
if and only if
1

(V ) is an open subset of R
n
for every parameter-
ization (U

) of a C

atlas.
(3) Show that the two atlases on R
n
from Example 2.3.1 are equivalent.
(4) Consider the two atlases on R from Example 2.3.2, (R,
1
) and
(R,
2
), where
1
(x) = x and
2
(x) = x
3
. Show that
1
2

1
is
not dierentiable at the origin. Conclude that the two atlases are
not equivalent.
(5) Recall from elementary vector calculus that a surface S R
3
is
a set such that, for each p S, there is a neighborhood V
p
of p in
R
3
and a C

map f
p
: U
p
R (where U
p
is an open subset of R
2
)
such that S V
p
is the graph of z = f
p
(x, y), or x = f
p
(y, z), or
y = f
p
(x, z). Show that S is a smooth manifold of dimension 2.
(6) (Product manifold) Let (U

), (V

) be two atlases for

two smooth manifolds M and N. Show that the family (U

) is an atlas for the product M N. With the dif-

ferentiable structure generated by this atlas, M N is called the
product manifold of M and N.
(7) (Stereographic projection) Consider the n-sphere S
n
with the sub-
space topology and let N = (0, . . . , 0, 1) and S = (0, . . . , 0, 1) be
the north and south poles. The stereographic projection from
N is the map
N
: S
n
N R
n
which takes a point p S
n
N
to the intersection point of the line through N and p with the hy-
perplane x
n+1
= 0 (cf. Figure 10). Similarly, the stereographic
projection from S is the map
S
: S
n
S R
n
which takes a
point p on S
n
S to the intersection point of the line through S
and p with the same hyperplane. Check that (R
n
,
1
N
), (R
n
,
1
S
)
is an atlas for S
n
. Show that this atlas is equivalent to the atlas
on Example 2.3.5. The maximal atlas obtained from these is called
the standard dierentiable structure on S
n
.
(8) (Real projective space) The real projective space RP
n
is the set
of lines through the origin in R
n+1
. This space can be dened as
the quotient space of S
n
by the equivalence relation x x that
identies a point to its antipodal point.
(a) Show that the quotient space RP
n
= S
n
/ with the quotient
topology is a Hausdor space and admits a countable basis of
open sets. (Hint: Use Proposition 10.2).
3. DIFFERENTIABLE MAPS 15
N
p
S
n

N
(p)
Figure 10. Stereographic projection.
(b) Considering the atlas on S
n
dened in Example 2.3.5 and the
canonical projection : S
n
RP
n
given by (x) = [x], dene
an atlas for RP
n
.
(9) We can dene an atlas on RP
n
in a dierent way by identify-
ing it with the quotient space of R
n+1
0 by the equivalence
relation x x, with R0. For that, consider the sets
V
i
= [x
1
, . . . , x
n+1
][ x
i
,= 0 (corresponding to the set of lines
through the origin in R
n+1
that are not contained on the hyper-
plane x
i
= 0) and the maps
i
: R
n
V
i
dened by

i
(x
1
, . . . , x
n
) = [x
1
, . . . , x
i1
, 1, x
i
, . . . , x
n
].
Show that:
(a) the family (R
n
,
i
) is an atlas for RP
n
;
(b) this atlas denes the same dierentiable structure as the atlas
on Exercise 2.5.8.
(10) (A non-Hausdor manifold) Let M be the disjoint union of R with
a point p and consider the maps f
i
: R M (i = 1, 2) dened by
f
i
(x) = x if x R0, f
1
(0) = 0 and f
2
(0) = p. Show that:
(a) the maps f
1
i
f
j
are dierentiable on their domains;
(b) if we consider an atlas formed by (R, f
1
), (R, f
2
), the corre-
sponding topology will not satisfy the Hausdor axiom.
3. Dierentiable Maps
In this book the words dierentiable and smooth will be used to mean
innitely dierentiable (C

).
Definition 3.1. Let M and N be two dierentiable manifolds of dimen-
sion m and n, respectively. A map f : M N is said to be dierentiable
(or smooth, or C

) at a point p M if there exist parameterizations (U, )

of M at p (i.e. p (U)) and (V, ) of N at f(p), with f((U)) (V ),
16 1. DIFFERENTIABLE MANIFOLDS
such that the map

f :=
1
f : U R
m
R
n
is smooth.
The map f is said to be dierentiable on a subset of M if it is dieren-
tiable at every point of this set.
As coordinate changes are smooth, this denition is independent of the
parameterizations chosen at f(p) and p. The map

f :=
1
f : U
R
m
R
n
is called a local representation of f and is the expression of f
on the local coordinates dened by and . The set of all smooth functions
f : M N is denoted C

(M, N), and we will simply write C

(M) for
C

(M, R).
M N
U V
f

f
R
m
R
n

Figure 11. Local representation of a map between manifolds.

A dierentiable map f : M N between two manifolds is continuous
(cf. Exercise 3.2.2). Moreover, it is called a dieomorphism if it is bijective
and its inverse f
1
: N M is also dierentiable. The dierentiable
manifolds M and N will be considered the same if they are dieomorphic,
i.e. if there exists a dieomorphism f : M N. A map f is called a local
dieomorphism at a point p M if there are neighborhoods V of p and
W of f(p) such that f[
V
: V W is a dieomorphism.
For a long time it was thought that, up to a dieomorphism, there was
only one dierentiable structure for each topological manifold (the two dier-
ent dierentiable structures in Exercises 2.5.4 and 3.2.6 are dieomorphic
cf. Exercise 3.2.6). However, in 1956, Milnor [Mil56] presented examples of
manifolds that were homeomorphic but not dieomorphic to S
7
. Later, Mil-
nor and Kervaire [Mil59, KM63] showed that more spheres of dimension
greater than 7 admitted several dierentiable structures. For instance, S
19
has 73 distinct smooth structures and S
31
has 16, 931, 177. More recently,
in 1982 and 1983, Freedman [Fre82] and Gompf [Gom83] constructed ex-
amples of non-standard dierentiable structures on R
4
.
4. TANGENT SPACE 17
Exercises 3.2.
(1) Prove that Denition 3.1 does not depend on the choice of param-
eterizations.
(2) Show that a dierentiable map f : M N between two smooth
manifolds is continuous.
(3) Show that if f : M
1
M
2
and g : M
2
M
3
are dierentiable maps
between smooth manifolds M
1
, M
2
and M
3
, then g f : M
1
M
3
is also dierentiable.
(4) Show that the antipodal map f : S
n
S
n
, dened by f(x) = x,
is dierentiable.
(5) Using the stereographic projection from the north pole
N
: S
2

N R
2
and identifying R
2
with the complex plane C, we can
identify S
2
with C, where is the so-called point at inn-
ity. A Mobius transformation is a map f : C C
of the form
f(z) =
az +b
cz +d
,
where a, b, c, d C satisfy ad bc ,= 0 and satises

= 0,

0
=
for any C 0. Show that any Mobius transformation f, seen
as a map f : S
2
S
2
, is a dieomorphism. (Hint: Start by showing that
any Mobius transformation is a composition of transformations of the form g(z) =
1
z
and h(z) = az + b).
(6) Consider again the two atlases on R from Example 2.3.2 and Exer-
cise 2.5.4, (R,
1
) and (R,
2
), where
1
(x) = x and
2
(x) =
x
3
. Show that:
(a) the identity map i : (R,
1
) (R,
2
) is not a dieomorphism;
(b) the map f : (R,
1
) (R,
2
) dened by f(x) = x
3
is a dif-
feomorphism (implying that although these two atlases dene
dierent dierentiable structures, they are dieomorphic).
4. Tangent Space
Recall from elementary vector calculus that a vector v R
3
is said
to be tangent to a surface S R
3
at a point p S if there exists a
dierentiable curve c : (, ) S R
3
such that c(0) = p and c(0) = v
(cf. Exercise 2.5.5). The set T
p
S of all these vectors is a 2-dimensional vector
space, called the tangent space to S at p, and can be identied with the
plane in R
3
which is tangent to S at p.
To generalize this to an abstract n-dimensional manifold we need to nd
a description of v which does not involve the ambient Euclidean space R
3
.
To do so, we notice that the components of v are
v
i
=
d(x
i
c)
dt
(0),
18 1. DIFFERENTIABLE MANIFOLDS
S
v
p
T
p
S
Figure 12. Tangent vector to a surface.
where x
i
: R
3
R is the i-th coordinate function. If we ignore the ambient
space, x
i
: S R is just a dierentiable function, and
v
i
= v(x
i
),
where, for any dierentiable function f : S R, we dene
v(f) :=
d(f c)
dt
(0).
This allows us to see v as an operator v : C

(S) R, and it is clear that this

operator completely determines the vector v. It is this new interpretation
of tangent vector that will be used to dene tangent spaces for manifolds.
Definition 4.1. Let c : (, ) M be a dierentiable curve on a
smooth manifold M. Consider the set C

(p) of all functions f : M R

that are dierentiable at c(0) = p. The tangent vector to the curve c at
p is the operator c(0) : C

(p) R given by
c(0)(f) =
d(f c)
dt
(0).
A tangent vector to M at p is a tangent vector to some dierentiable curve
c : (, ) M with c(0) = p. The tangent space at p is the space T
p
M
of all tangent vectors at p.
Choosing a parameterization : U R
n
M around p, the curve c is
given in local coordinates by the curve in U
c(t) :=
_

1
c
_
(t) = (x
1
(t), . . . , x
n
(t)),
4. TANGENT SPACE 19
and
c(0)(f) =
d(f c)
dt
(0) =
d
dt
_
_
_(

f
..
f ) (
c
..

1
c)
_
_
_
[
t=0
=
=
d
dt
_

f(x
1
(t), . . . , x
n
(t))
_
[
t=0
=
n

i=1

f
x
i
( c(0))
dx
i
dt
(0) =
=
_
n

i=1
x
i
(0)
_

x
i
_

1
(p)
_
(

f).
Hence we can write
c(0) =
n

i=1
x
i
(0)
_

x
i
_
p
,
where
_

x
i
_
p
denotes the operator associated to the vector tangent to the
curve c
i
at p given in local coordinates by
c
i
(t) = (x
1
, . . . , x
i1
, x
i
+t, x
i+1
, . . . , x
n
),
with (x
1
, . . . , x
n
) =
1
(p).
Example 4.2. The map : (0, ) (, ) S
2
given by
(, ) = (sin cos , sin sin , cos )
parameterizes a neighborhood of the point (1, 0, 0) =
_

2
, 0
_
. Conse-
quently,
_

_
(1,0,0)
= c

(0) and
_

_
(1,0,0)
= c

(0), where
c

(t) =
_

2
+t, 0
_
= (cos t, 0, sin t);
c

(t) =
_

2
, t
_
= (cos t, sin t, 0).
Note that, in the notation above,
c

(t) =
_

2
+t, 0
_
and c

(t) =
_

2
, t
_
.
Moreover, since c

and c

are curves in R
3
,
_

_
(1,0,0)
and
_

_
(1,0,0)
can
be identied with the vectors (0, 0, 1) and (0, 1, 0).
Proposition 4.3. The tangent space to M at p is an n-dimensional
vector space.
Proof. Consider a parameterization : U R
n
M around p and
take the vector space generated by the operators
_

x
i
_
p
,
T
p
:= span
_
_

x
1
_
p
, . . . ,
_

x
n
_
p
_
.
20 1. DIFFERENTIABLE MANIFOLDS
It is easy to show (cf. Exercise 4.9.1) that these operators are linearly inde-
pendent. Moreover, each tangent vector to M at p can be represented by a
linear combination of these operators, so the tangent space T
p
M is a subset
of T
p
. We will now see that T
p
T
p
M. Let v T
p
; then v can be written
as
v =
n

i=1
v
i
_

x
i
_
p
.
If we consider the curve c : (, ) M, dened by
c(t) = (x
1
+v
1
t, . . . , x
n
+v
n
t)
(where (x
1
, . . . , x
n
) =
1
(p)), then
c(t) = (x
1
+v
1
t, . . . , x
n
+v
n
t)
and so x
i
(0) = v
i
, implying that c(0) = v. Therefore v T
p
M.
Remark 4.4.
(1) The basis
_
_

x
i
_
p
_
n
i=1
determined by the chosen parameterization
around p is called the associated basis to that parameterization.
(2) Note that the denition of tangent space at p only uses functions
that are dierentiable on a neighborhood of p. Hence, if U is an
open set of M containing p, the tangent space T
p
U is naturally
identied with T
p
M.
If we consider the disjoint union of all tangent spaces T
p
M at all points
of M, we obtain the space
TM =
_
pM
T
p
M = v T
p
M [ p M,
which admits a dierentiable structure naturally determined by the one on
M (cf. Exercise 4.9.8). With this dierentiable structure, this space is called
the tangent bundle. Note that there is a natural projection : TM M
which takes v T
p
M to p (cf. Section 10.3).
Now that we have dened tangent space, we can dene the derivative
at a point p of a dierentiable map f : M N between smooth manifolds.
We want this derivative to be a linear transformation
(df)
p
: T
p
M T
f(p)
N
of the corresponding tangent spaces, to be the usual derivative (Jacobian)
of f when M and N are Euclidean spaces, and to satisfy the chain rule.
Definition 4.5. Let f : M N be a dierentiable map between smooth
manifolds. For p M, the derivative of f at p is the map
(df)
p
: T
p
M T
f(p)
N
v
d (f c)
dt
(0),
where c : (, ) M is a curve satisfying c(0) = p and c(0) = v.
4. TANGENT SPACE 21
Proposition 4.6. The map (df)
p
: T
p
M T
f(p)
N dened above is a
linear transformation that does not depend on the choice of the curve c.
Proof. Let (U, ) and (V, ) be two parameterizations around p and
f(p) such that f((U)) (V ) (cf. Figure 13). Consider a vector v T
p
M
M N
U V
f

f
R
m
R
n

c
c

p
v
(df)
p
(v)
Figure 13. Derivative of a dierentiable map.
and a curve c : (, ) M such that c(0) = p and c(0) = v. If, in local
coordinates, the curve c is given by
c(t) := (
1
c)(t) = (x
1
(t), . . . , x
m
(t)),
and the curve := f c : (, ) N is given by
(t) :=
_

1

_
(t) =
_

1
f
_
(x
1
(t), . . . , x
m
(t))
= (y
1
(x(t)), . . . , y
n
(x(t))),
then (0) is the tangent vector in T
f(p)
N given by
(0) =
n

i=1
d
dt
_
y
i
(x
1
(t), . . . , x
m
(t))
_
[
t=0
_

y
i
_
f(p)
=
n

i=1
_
m

k=1
x
k
(0)
_
y
i
x
k
_
(x(0))
_
_

y
i
_
f(p)
=
n

i=1
_
m

k=1
v
k
_
y
i
x
k
_
(x(0))
_
_

y
i
_
f(p)
,
where the v
k
are the components of v in the basis associated to (U, ). Hence
(0) does not depend on the choice of c, as long as c(0) = v. Moreover, the
components of w = (df)
p
(v) in the basis associated to (V, ) are
w
i
=
m

j=1
y
i
x
j
v
j
,
22 1. DIFFERENTIABLE MANIFOLDS
where
_
y
i
x
j
_
is an n m matrix (the Jacobian matrix of the local repre-
sentation of f at
1
(p)). Therefore, (df)
p
: T
p
M T
f(p)
N is the linear
transformation which, on the basis associated to the parameterizations
and , is represented by this matrix.
Remark 4.7. The derivative (df)
p
is sometimes called dierential of
f at p. Several other notations are often used for df, as for example f

, Df
and f

.
Example 4.8. Let : U R
n
M be a parameterization around a
point p M. We can view as a dierentiable map between two smooth
manifolds and we can compute its derivative at x =
1
(p)
(d)
x
: T
x
U T
p
M.
For v T
x
U

= R
n
, the i-th component of (d)
x
(v) is
n

j=1
x
i
x
j
v
j
= v
i
(where
_
x
i
x
j
_
is the identity matrix). Hence, (d)
x
(v) is the vector in T
p
M
which, in the basis
_
_

x
i
_
p
_
associated to the parameterization , is repre-
sented by v.
Given a dierentiable map f : M N we can also dene a global
derivative df (also called push-forward and denoted f

) between the cor-

responding tangent bundles:
df : TM TN
T
p
M v (df)
p
(v) T
f(p)
N.
Exercises 4.9.
(1) Show that the operators
_

x
i
_
p
are linearly independent.
(2) Let M be a smooth manifold, p a point in M and v a vector tangent
to M at p. Show that if v can be written as v =

n
i=1
a
i
(

x
i
)
p
and
v =

n
i=1
b
i
(

y
i
)
p
for two basis associated to dierent parameteri-
zations around p, then
b
j
=
n

i=1
y
j
x
i
a
i
.
(3) Let M be an n-dimensional dierentiable manifold and p M.
Show that the following sets can be canonically identied with
T
p
M (and therefore constitute alternative denitions of the tan-
gent space):
4. TANGENT SPACE 23
(a) (
p
/ , where (
p
is the set of dierentiable curves c : I R
M such that c(0) = p and is the equivalence relation dened
by
c
1
c
2

d
dt
(
1
c
1
)(0) =
d
dt
(
1
c
2
)(0)
for some parameterization : U R
n
M of a neighborhood
of p.
(b) (, v

) : p

(U

) and v

R
n
/ , where / = (U

)
is the dierentiable structure and is the equivalence relation
dened by
(, v

) (, v

) v

= d(
1

(p)
(v

).
(4) (Chain Rule) Let f : M N and g : N P be two dierentiable
maps. Then gf : M P is also dierentiable (cf. Exercise 3.2.3).
Show that for p M,
(d(g f))
p
= (dg)
f(p)
(df)
p
.
(5) Let : (0, +) (0, ) (0, 2) R
3
be the parameterization of
U = R
3
(x, 0, z) [ x 0 and z R by spherical coordinates,
(r, , ) = (r sin cos , r sin sin , r cos ).
Determine the Cartesian components of

r
,

and

at each point
of U.
(6) Compute the derivative (df)
N
of the antipodal map f : S
n
S
n
at the north pole N.
(7) Let W be a coordinate neighborhood on M, let x : W R
n
be a
coordinate chart and consider a smooth function f : M R. Show
that for p W, the derivative (df)
p
is given by
(df)
p
=


f
x
1
(x(p))(dx
1
)
p
+ +


f
x
n
(x(p))(dx
n
)
p
,
where

f := f x
1
.
(8) (Tangent bundle) Let (U

) be a dierentiable structure on
M and consider the maps

: U

R
n
TM
(x, v) (d

)
x
(v) T
(x)
M.
Show that the family (U

R
n
,

) denes a dierentiable struc-

ture for TM. Conclude that, with this dierentiable structure, TM
is a smooth manifold of dimension 2 dimM.
(9) Let f : M N be a dierentiable map between smooth manifolds.
Show that:
(a) df : TM TN is also dierentiable;
(b) if f : M M is the identity map then df : TM TM is also
the identity;
24 1. DIFFERENTIABLE MANIFOLDS
(c) if f is a dieomorphism then df : TM TN is also a dieo-
morphism and (df)
1
= df
1
.
(10) Let M
1
, M
2
be two dierentiable manifolds and

1
: M
1
M
2
M
1

2
: M
1
M
2
M
2
the corresponding canonical projections.
(a) Show that d
1
d
2
is a dieomorphism between the tangent
bundle T(M
1
M
2
) and the product manifold TM
1
TM
2
.
(b) Show that if N is a smooth manifold and f
i
: N M
i
(i = 1, 2)
are dierentiable maps, then d(f
1
f
2
) = df
1
df
2
.
5. Immersions and Embeddings
In this section we will study the local behavior of dierentiable maps
f : M N between smooth manifolds. We have already seen that f is
said to be a local dieomorphism at a point p M if dimM = dimN and
f transforms a neighborhood of p dieomorphically onto a neighborhood of
f(p). In this case, its derivative (df)
p
: T
p
M T
f(p)
N must necessarily be
an isomorphism (cf. Exercise 4.9.9). Conversely, if (df)
p
is an isomorphism
then the Inverse Function Theorem implies that f is a local dieomorphism
(cf. Section 10.4). Therefore, to check whether f maps a neighborhood of
p dieomorphically onto a neighborhood of f(p), one just has to check that
the determinant of the local representation of (df)
p
is nonzero.
When dimM < dimN, the best we can hope for is that (df)
p
: T
p
M
T
f(p)
N is injective. The map f is then called an immersion at p. If f is an
immersion at every point in M, it is called an immersion. Locally, every
immersion is (up to a dieomorphism) the canonical immersion of R
m
into
R
n
(m < n) where a point (x
1
, . . . , x
m
) is mapped to (x
1
, . . . , x
m
, 0, . . . , 0).
This result is known as the Local Immersion Theorem.
Theorem 5.1. Let f : M N be an immersion at p M. Then
there exist local coordinates around p and f(p) on which f is the canonical
immersion.
Proof. Let (U, ) and (V, ) be parameterizations around p and q =
f(p). Let us assume for simplicity that (0) = p and (0) = q. Since f
is an immersion, (d

f)
0
: R
m
R
n
is injective (where

f :=
1
f is
the expression of f in local coordinates). Hence we can assume (changing
basis on R
n
if necessary) that this linear transformation is represented by
the n m matrix
_
_
I
mm

0
_
_
,
5. IMMERSIONS AND EMBEDDINGS 25
where I
mm
is the mm identity matrix. Therefore, the map
F : U R
nm
R
n
(x
1
, . . . , x
n
)

f(x
1
, . . . , x
m
) + (0, . . . , 0, x
m+1
, . . . , x
n
),
has derivative (dF)
0
: R
n
R
n
given by the matrix
_
_
I
mm
[ 0
+
0 [ I
(nm)(nm)
_
_
= I
nn
.
Applying the Inverse Function Theorem, we conclude that F is a local dif-
feomorphism at 0. This implies that F is also a local dieomorphism at 0,
and so F is another parameterization of N around q. Denoting the canon-
ical immersion of R
m
into R
n
by j, we have

f = F j f = F j
1
,
implying that the following diagram commutes:
M (

U)
f
( F)(

V ) N
F
R
m

U
j

V R
n
(for possibly smaller open sets

U U and

V V ). Hence, on these new
coordinates, f is the canonical immersion.
Remark 5.2. As a consequence of the Local Immersion Theorem, any
immersion at a point p M is an immersion on a neighborhood of p.
When an immersion f : M N is also a homeomorphism onto its
image f(M) N with its subspace topology, it is called an embedding.
We leave as an exercise to show that the Local Immersion Theorem implies
that, locally, any immersion is an embedding.
Example 5.3.
(1) The map f : R R
2
given by f(t) = (t
2
, t
3
) is not an immersion
at t = 0.
(2) The map f : R R
2
dened by f(t) = (cos t, sin 2t) is an immer-
sion but it is not an embedding (it is not injective).
(3) Let g : R R be the function g(t) = 2 arctan(t) + /2. If f is the
map dened in (2) then h := f g is an injective immersion which
is not an embedding. Indeed, the set S = h(R) in Figure 14 is not
the image of an embedding of R into R
2
. The arrows in the gure
mean that the line approaches itself arbitrarily close at the origin
but never self-intersects. If we consider the usual topologies on R
and on R
2
, the image of a bounded open set in R containing 0 is
not an open set in h(R) for the subspace topology, and so h
1
is
not continuous.
26 1. DIFFERENTIABLE MANIFOLDS
S
Figure 14
(4) The map f : R R
2
given by f(t) = (e
t
cos t, e
t
sin t) is an embed-
ding of R into R
2
.
If M N and the inclusion map i : M N is an embedding, M is said
to be a submanifold of N. Therefore, an embedding f : M N maps
M dieomorphically onto a submanifold of N. Charts on f(M) are just
restrictions of appropriately chosen charts on N to f(M) (cf. Exercise 5.9.3).
A dierentiable map f : M N for which (df)
p
is surjective is called a
submersion at p. Note that, in this case, we necessarily have m n. If
f is a submersion at every point in M it is called a submersion. Locally,
every submersion is the standard projection of R
m
onto the rst n factors.
Theorem 5.4. Let f : M N be a submersion at p M. Then
there exist local coordinates around p and f(p) for which f is the standard
projection.
Proof. Let us consider parameterizations (U, ) and (V, ) around p
and f(p), such that f((U)) (V ), (0) = p and (0) = f(p). In
local coordinates f is given by

f :=
1
f and, as (df)
p
is surjective,
(d

f)
0
: R
m
R
n
is a surjective linear transformation. By a linear change
of coordinates on R
n
we may assume that (d

f)
0
=
_
I
nn
[
_
. As in
the proof of the Local Immersion Theorem, we will use an auxiliary map F
that will allow us to use the Inverse Function Theorem,
F : U R
m
R
m
(x
1
, . . . , x
m
)
_

f(x
1
, . . . , x
m
), x
n+1
, . . . , x
m
_
.
Its derivative at 0 is the linear map given by
(dF)
0
=
_
_
I
nn
[
+
0 [ I
(mn)(mn)
_
_
.
The Inverse Function Theorem then implies that F is a local dieomorphism
at 0, meaning that it maps some open neighborhood of this point

U U,
5. IMMERSIONS AND EMBEDDINGS 27
dieomorphically onto an open set W of R
m
containing 0. If
1
: R
m
R
n
is the standard projection onto the rst n factors, we have
1
F =

f, and
hence

f F
1
=
1
: W R
n
.
Therefore, replacing by := F
1
, we obtain coordinates for which f
is the standard projection
1
onto the rst n factors:

1
f =
1
f F
1
=

f F
1
=
1
.

Remark 5.5. This result is often stated together with the Local Immer-
sion Theorem in what is known as the Rank Theorem.
Let f : M N be a dierentiable map between smooth manifolds of
dimensions m and n, respectively. A point p M is called a regular point
of f if (df)
p
is surjective. A point q N is called a regular value of f if
every point in f
1
(q) is a regular point. A point p M which is not regular
is called a critical point of f. The corresponding value f(p) is called a
critical value. Note that if there exists a regular value of f then m n.
We can obtain dierentiable manifolds by taking inverse images of regular
values.
Theorem 5.6. Let q N be a regular value of f : M N and assume
that the level set L := f
1
(q) = p M : f(p) = q is nonempty. Then L
is a submanifold of M and T
p
L = ker(df)
p
T
p
M for all p L.
Proof. For each point p f
1
(q), we choose parameterizations (U, )
and (V, ) around p and q for which f is the standard projection
1
onto the
rst n factors, (0) = p and (0) = q (cf. Theorem 5.4). We then construct
a dierentiable structure for L := f
1
(q) in the following way: take the sets
U from each of these parameterizations of M; since f =
1
, we have

1
(f
1
(q)) =
1
1
(
1
(q)) =
1
1
(0)
= (0, . . . , 0, x
n+1
, . . . , x
m
) [ x
n+1
, . . . , x
m
R,
and so

U :=
1
(L) = (x
1
, . . . , x
m
) U : x
1
= = x
n
= 0;
hence, taking
2
: R
m
R
mn
, the standard projection onto the last mn
factors, and j : R
mn
R
m
, the immersion given by
j(x
1
, . . . , x
mn
) = (0, . . . , 0, x
1
, . . . , x
mn
),
the family (
2
(

U), j) is an atlas for L.

Moreover, the inclusion map i : L M is an embedding. In fact, if A
is an open set in L contained in a coordinate neighborhood then
A =
__
R
n
( j)
1
(A)
_
U
_
L
is an open set for the subspace topology on L.
28 1. DIFFERENTIABLE MANIFOLDS
We will now show that T
p
L = ker (df)
p
. For that, for each v T
p
L, we
consider a curve c on L such that c(0) = p and c(0) = v. Then (f c)(t) = q
for every t and so
d
dt
(f c) (0) = 0 (df)
p
c(0) = (df)
p
v = 0,
implying that v ker (df)
p
. As dimT
p
L = dim(ker (df)
p
) = m n, the
result follows.
Given a dierentiable manifold, we can ask ourselves if it can be embed-
ded into R
K
for some K N. The following theorem, which was proved by
Whitney in [Whi44a, Whi44b] answers this question and is known as the
Whitney Embedding Theorem.
Theorem 5.7. (Whitney) Any dierentiable manifold M of dimension
n can be embedded in R
2n
(and, provided that n > 1, immersed in R
2n1
).
Remark 5.8. By the Whitney Embedding Theorem, any smooth man-
ifold M of dimension n is dieomorphic to a submanifold of R
2n
.
Exercises 5.9.
(1) Show that any parameterization : U R
m
M is an embedding
of U into M.
(2) Show that, locally, any immersion is an embedding, i.e., if f : M
N is an immersion and p M, then there is an open set W M
containing p such that f[
W
is an embedding.
(3) Let N be a manifold. Show that M N is a submanifold of N of
dimension m if and only if, for each p M, there is a coordinate
system x : W R
n
around p on N, for which M W is dened
by the equations x
m+1
= = x
n
= 0.
(4) Consider the sphere
S
n
=
_
x R
n+1
: (x
1
)
2
+ (x
n+1
)
2
= 1
_
.
Show that S
n
is an n-dimensional submanifold of R
n+1
and that
T
x
S
n
=
_
v R
n+1
: x, v = 0
_
,
where , is the usual inner product on R
n
.
(5) Let f : M N be a dierentiable map between smooth manifolds
and consider submanifolds V M and W N. Show that if
f(V ) W then f : V W is also a dierentiable map.
(6) Let f : M N be an injective immersion. Show that if M is
compact then f(M) is a submanifold of N.
6. VECTOR FIELDS 29
6. Vector Fields
A vector eld on a smooth manifold M is a map that to each point
p M assigns a vector tangent to M at p:
X : M TM
p X(p) := X
p
T
p
M.
The vector eld is said to be dierentiable if this map is dierentiable.
The set of all dierentiable vector elds on M is denoted by X(M). Locally
we have:
Proposition 6.1. Let W be a coordinate neighborhood on M (that is,
W = (U) for some parameterization : U M), and let x :=
1
: W
R
n
be the corresponding coordinate chart. Then, a map X : W TW is a
dierentiable vector eld on W if and only if,
X
p
= X
1
(p)
_

x
1
_
p
+ +X
n
(p)
_

x
n
_
p
for some dierentiable functions X
i
: W R (i = 1, . . . , n).
Proof. Let us consider the coordinate chart x = (x
1
, . . . , x
n
). As X
p

T
p
M, we have
X
p
= X
1
(p)
_

x
1
_
p
+ +X
n
(p)
_

x
n
_
p
for some functions X
i
: W R. In the local chart associated with the
parameterization (U R
n
, d) of TM, the local representation of the map
X is

X(x
1
, . . . , x
n
) = (x
1
, . . . , x
n
,

X
1
(x
1
, . . . , x
n
), . . . ,

X
n
(x
1
, . . . , x
n
)).
Therefore X is dierentiable if and only if the functions

X
i
: U R are
dierentiable, i.e., if and only if the functions X
i
: W R are dierentiable.

A vector eld X is dierentiable if and only if, given any dierentiable

function f : M R, the function
X f : M R
p X
p
f := X
p
(f)
is also dierentiable (cf. Exercise 6.11.1). This function X f is called the
directional derivative of f along X. Thus one can view X X(M) as a
linear operator X : C

(M) C

(M).
Let us now take two vector elds X, Y X(M). In general, the operators
XY and Y X will involve derivatives of order two, and will not correspond
to vector elds. However, the commutator XY Y X does dene a vector
eld.
30 1. DIFFERENTIABLE MANIFOLDS
Proposition 6.2. Given two dierentiable vector elds X, Y X(M)
on a smooth manifold M, there exists a unique dierentiable vector eld
Z X(M) such that
Z f = (X Y Y X) f
for every dierentiable function f C

(M).
Proof. Considering a coordinate chart x : W M R
n
, we have
X =
n

i=1
X
i

x
i
and Y =
n

i=1
Y
i

x
i
.
Then,
(X Y Y X) f
= X
_
n

i=1
Y
i


f
x
i
_
Y
_
n

i=1
X
i


f
x
i
_
=
n

i=1
_
(X Y
i
)


f
x
i
(Y X
i
)


f
x
i
_
+
n

i,j=1
_
X
j
Y
i

2

f
x
j
x
i
Y
j
X
i

2

f
x
j
x
i
_
=
_
n

i=1
_
X Y
i
Y X
i
_

x
i
_
f,
and so, at each point p W, one has ((X Y Y X) f) (p) = Z
p
f,
where
Z
p
=
n

i=1
_
X Y
i
Y X
i
_
(p)
_

x
i
_
p
.
Hence, the operator X Y Y X is a derivation at each point, and con-
sequently denes a vector eld. Note that this vector eld is dierentiable,
as (X Y Y X) f is smooth for every smooth function f : M R.
The vector eld Z is called the Lie bracket of X and Y , and is denoted
by [X, Y ]. In local coordinates it is given by
(1) [X, Y ] =
n

i=1
_
X Y
i
Y X
i
_

x
i
.
We say that two vector elds X, Y X(M) commute if [X, Y ] = 0. We
leave the proof of the following properties of the Lie bracket as an exercise.
Proposition 6.3. Given X, Y, Z X(M), we have:
(i) Bilinearity: for any , R,
[X +Y, Z] = [X, Z] +[Y, Z]
[X, Y +Z] = [X, Y ] +[X, Z];
(ii) Antisymmetry:
[X, Y ] = [Y, X];
6. VECTOR FIELDS 31
(iii) Jacobi identity:
[[X, Y ], Z] + [[Y, Z], X] + [[Z, X], Y ] = 0;
(iv) Leibniz Rule: For any f, g C

(M),
[f X, g Y ] = fg [X, Y ] +f(X g)Y g(Y f)X.
The space X(M) of vector elds on M is a particular case of a Lie
algebra:
Definition 6.4. A vector space V equipped with an antisymmetric bi-
linear map [, ] : V V V (called a Lie bracket) satisfying the Jacobi
identity is called a Lie algebra. A linear map F : V W between Lie alge-
bras is called a Lie algebra homomorphism if F([v
1
, v
2
]) = [F(v
1
), F(v
2
)]
for all v
1
, v
2
V . If F is bijective then it is called a Lie algebra isomor-
phism.
Given a vector eld X X(M) and a dieomorphism f : M N
between smooth manifolds, we can naturally dene a vector eld on N using
the derivative of f. This vector eld, the push-forward of X, is denoted
by f

X and is dened in the following way: given p M,

(f

X)
f(p)
:= (df)
p
X
p
.
This makes the following diagram commute:
TM
df
TN
X f

X
M
f
N
Let us now turn to the denition of integral curve. If X X(M) is a
smooth vector eld, an integral curve of X is a smooth curve c : (, )
M such that c(t) = X
c(t)
. If this curve has initial value c(0) = p, we denote
it by c
p
and we say that c
p
is an integral curve of X at p.
Considering a parameterization : U R
n
M on M, the integral
curve c is locally given by c :=
1
c. Applying (d
1
)
c(t)
to both sides of
the equation dening c, we obtain

c(t) =

X( c(t)),
where

X = d
1
X is the local representation of X with respect to the
parameterizations (U, ) and (TU, d) on M and on TM (cf. Figure 15).
This equation is just a system of n ordinary dierential equations:
(2)
d c
i
dt
(t) =

X
i
( c(t)), for i = 1, . . . , n.
The (local) existence and uniqueness of integral curves is then determined
by the Picard-Lindelof Theorem of ordinary dierential equations (see for
example [Arn92]), and we have
32 1. DIFFERENTIABLE MANIFOLDS
M
U
R
n

c
c
X

X
Figure 15. Integral curves of a vector eld.
Theorem 6.5. Let M be a smooth manifold and let X X(M) be a
smooth vector eld on M. Given p M, there exists an integral curve
c
p
: I M of X at p (that is, c
p
(t) = X
cp(t)
for t I = (, ) and
c
p
(0) = p). Moreover, this curve is unique, meaning that any two such
curves agree on the intersection of their domains.
This integral curve, obtained by solving (2), depends smoothly on the
initial point p (see [Arn92]).
Theorem 6.6. Let X X(M). For each p M there exists a neighbor-
hood W of p, an interval I = (, ) and a mapping F : W I M such
that:
(i) for a xed q W the curve F(q, t), t I, is an integral curve of
X at q, that is, F(q, 0) = q and
F
t
(q, t) = X
F(q,t)
;
(ii) the map F is dierentiable.
The map F : W I M dened above is called the local ow of X
at p. Let us now x t I and consider the map

t
: W M
q F(q, t) = c
q
(t).
dened by the local ow. The following proposition then holds:
6. VECTOR FIELDS 33
Proposition 6.7. The maps
t
: W M above are local dieomor-
phisms and satisfy
(3) (
t

s
)(q) =
t+s
(q),
whenever t, s, t +s I and
s
(q) W.
Proof. First we note that
dc
q
dt
(t) = X
cq(t)
and so
d
dt
(c
q
(t +s)) = X
cq(t+s)
.
Hence, as c
q
(t + s)[
t=0
= c
q
(s), the curve c
cq(s)
(t) is just c
q
(t + s), that is,

t+s
(q) =
t
(
s
(q)). We can use this formula to extend
t
to
s
(W) for
all s I such that t + s I. In particular,
t
is well dened on
t
(W),
and (
t

t
)(q) =
0
(q) = c
q
(0) = q for all q W. Thus the map
t
is
the inverse of
t
, which consequently is a local dieomorphism (it maps W
dieomorphically onto its image).
A collection of dieomorphisms
t
: M M
tI
, where I = (, ),
satisfying (3) is called a local 1-parameter group of dieomorphisms.
When the interval of denition I of c
q
is R, this local 1-parameter group
of dieomorphisms becomes a group of dieomorphisms. A vector eld
X whose local ow denes a 1-parameter group of dieomorphisms is said
to be complete. This happens for instance when the vector eld X has
compact support.
Theorem 6.8. If X X(M) is a smooth vector eld with compact sup-
port then it is complete.
Proof. For each p M we can take a neighborhood W and an interval
I = (, ) such that the local ow of X at p, F(q, t) = c
q
(t), is dened on
WI. We can therefore cover the support of X (which is compact) by a nite
number of such neighborhoods W
k
and consider an interval I
0
= (
0
,
0
)
contained in the intersection of the corresponding intervals I
k
. If q is not
in supp(X), then X
q
= 0 and so c
q
(t) is trivially dened on I
0
. Hence we
can extend the map F to M I
0
. Moreover, condition (3) is true for each

0
/2 < s, t <
0
/2, and we can again extend the map F, this time to
M R. In fact, for any t R, we can write t = k
0
/2 +s, where k Z and
0 s <
0
/2, and dene F(q, t) := F
k
(F(q, s),
0
/2).
Corollary 6.9. If M is compact then all smooth vector elds on M
are complete.
We nish this section with an important result, whose proof is left as an
exercise (cf. Exercise 6.11.12).
34 1. DIFFERENTIABLE MANIFOLDS
Theorem 6.10. Let X
1
, X
2
X(M) be two complete vector elds. Then
their ows
1
,
2
commute (i.e.,
1,t

2,s
=
2,s

1,t
for all s, t R) if
and only if [X
1
, X
2
] = 0.
Exercises 6.11.
(1) Let X : M TM be a dierentiable vector eld on M and, for
a smooth function f : M R, consider its directional derivative
along X dened by
X f : M R
p X
p
f.
Show that:
(a) (X f)(p) = (df)
p
X
p
;
(b) the vector eld X is smooth if and only if Xf is a dierentiable
function for any smooth function f : M R;
(c) the directional derivative satises the following properties: for
f, g C

(M) and R,
(i) X (f +g) = X f +X g;
(ii) X (f) = (X f);
(iii) X (fg) = fX g +gX f.
(2) Prove Proposition 6.3.
(3) Show that (R
3
, ) is a Lie algebra, where is the cross product
on R
3
.
(4) Let X
1
, X
2
, X
3
X(R
3
) be the vector elds dened by
X
1
= y

z
z

y
, X
2
= z

x
x

z
, X
3
= x

y
y

x
,
where (x, y, z) are the usual Cartesian coordinates.
(a) Compute the Lie brackets [X
i
, X
j
] for i, j = 1, 2, 3.
(b) Show that spanX
1
, X
2
, X
3
is a Lie subalgebra of X(R
3
), iso-
morphic to (R
3
, ).
(c) Compute the ows
1,t
,
2,t
,
3,t
of X
1
, X
2
, X
3
.
(d) Show that
i,

2

j,

2
,=
j,

2

i,

2
for i ,= j.
(5) Give an example of a non-complete vector eld.
(6) Let N be a dierentiable manifold, M N a submanifold and
X, Y X(N) vector elds tangent to M, i.e., such that X
p
, Y
p

T
p
M for all p M. Show that [X, Y ] is also tangent to M, and that
its restriction to M coincides with the Lie bracket of the restrictions
of X and Y to M.
(7) Let f : M N be a smooth map between manifolds. Two vector
elds X X(M) and Y X(N) are said to be f-related (and we
write Y = f

X) if, for each q N and p f

1
(q) M, we have
(df)
p
X
p
= Y
q
. Show that:
(a) The vector eld X is f-related to Y if and only if, for any
dierentiable function g dened on some open subset W of N,
6. VECTOR FIELDS 35
(Y g) f = X (g f) on the inverse image f
1
(W) of the
domain of g;
(b) For dierentiable maps f : M N and g : N P between
smooth manifolds and vector elds X X(M), Y X(N) and
Z X(P), if X is f-related to Y and Y is g-related to Z, then
X is (g f)-related to Z.
(8) Let f : M N be a dieomorphism between smooth manifolds.
Show that f

[X, Y ] = [f

X, f

Y ] for every X, Y X(M). There-

fore, f

induces a Lie algebra isomorphism between X(M) and

X(N).
(9) Let f : M N be a dierentiable map between smooth manifolds
and consider two vector elds X X(M) and Y X(N). Show
that:
(a) if the vector eld Y is f-related to X then any integral curve
of X is mapped by f into an integral curve of Y ;
(b) the vector eld Y is f-related to X if and only if the local ows
F
X
and F
Y
satisfy f (F
X
(p, t)) = F
Y
(f(p), t) for all (t, p) for
which both sides are dened.
(10) (Lie derivative of a function) Given a vector eld X X(M), we
dene the Lie derivative of a smooth function f : M R in the
direction of X as
L
X
f(p) :=
d
dt
((f
t
)(p))
[
t=0
,
where
t
= F(, t), for F the local ow of X at p. Show that
L
X
f = X f, meaning that the Lie derivative of f in the direction
of X is just the directional derivative of f along X.
(11) (Lie derivative of a vector eld) For two vector elds X, Y X(M)
we dene the Lie derivative of Y in the direction of X as
L
X
Y :=
d
dt
((
t
)

Y )
[
t=0
,
where
t

tI
is the local ow of X. Show that:
(a) L
X
Y = [X, Y ];
(b) L
X
[Y, Z] = [L
X
Y, Z] + [Y, L
X
Z], for X, Y, Z X(M);
(c) L
X
L
Y
L
Y
L
X
= L
[X,Y ]
.
(12) Let X, Y X(M) be two complete vector elds with ows , .
Show that:
(a) given a dieomorphism f : M M, we have f

X = X if and
only if f
t
=
t
f for all t R;
(b)
t

s
=
s

t
for all s, t R if and only if [X, Y ] = 0.
36 1. DIFFERENTIABLE MANIFOLDS
7. Lie Groups
A Lie group G is a smooth manifold which is at the same time a group,
in such a way that the group operations
GG G
(g, h) gh
and
G G
g g
1
are dierentiable maps (where we consider the standard dierentiable struc-
ture of the product on GG cf. Exercise 2.5.6).
Example 7.1.
(1) (R
n
, +) is trivially an abelian Lie group.
(2) The general linear group
GL(n) = n n invertible real matrices
is the most basic example of a nontrivial Lie group. We have seen
in Example 2.3.4 that it is a smooth manifold of dimension n
2
.
Moreover, the group multiplication is just the restriction to
GL(n) GL(n)
of the usual multiplication of n n matrices, whose coordinate
functions are quadratic polynomials; the inversion is just the re-
striction to GL(n) of the usual inversion of nonsingular matrices
which, by Cramers rule, is a map with rational coordinate func-
tions and nonzero denominators (only the determinant appears on
the denominator).
(3) The orthogonal group
O(n) = A /
nn
[ A
t
A = I
of orthogonal transformations of R
n
is also a Lie group. We can
show this by considering the map f : A A
t
A from /
nn

= R
n
2
to the space o
nn

= R
1
2
n(n+1)
of symmetric n n matrices. Its
derivative at a point A O(n), (df)
A
, is a surjective map from
T
A
/
nn

= /
nn
onto T
f(A)
o
nn

= o
nn
. Indeed,
(df)
A
(B) = lim
h0
f(A+hB) f(A)
h
= lim
h0
(A +hB)
t
(A+hB) A
t
A
h
= B
t
A+A
t
B,
and any symmetric matrix S can be written as B
t
A + A
t
B with
B =
1
2
(A
1
)
t
S =
1
2
AS. In particular, the identity I is a regular
value of f and so, by Theorem 5.6, we have that O(n) = f
1
(I) is a
submanifold of /
nn
of dimension
1
2
n(n1). Moreover, it is also a
Lie group as the group multiplication and inversion are restrictions
7. LIE GROUPS 37
of the same operations on GL(n) to O(n) (a submanifold) and have
values on O(n) (cf. Exercise 5.9.5).
(4) The map f : GL(n) R given by f(A) = det A is dierentiable,
and the level set f
1
(1) is
SL(n) = A /
nn
[ det A = 1,
the special linear group. Again, the derivative of f is surjective
at a point A GL(n), making SL(n) into a Lie group. Indeed, it
is easy to see that
(df)
I
(B) = lim
h0
det (I +hB) det I
h
= tr B
implying that
(df)
A
(B) = lim
h0
det (A +hB) det A
h
= lim
h0
(det A) det (I +hA
1
B) det A
h
= (det A) lim
h0
det (I +hA
1
B) 1
h
= (det A) (df)
I
(A
1
B) = (det A) tr(A
1
B).
Since det (A) = 1, for any k R, we can take the matrix B =
k
n
A
to obtain (df)
A
(B) = tr
_
k
n
I
_
= k. Therefore, (df)
A
is surjective
for every A SL(n), and so 1 is a regular value of f. Consequently,
SL(n) is a submanifold of GL(n). As in the preceding example, the
group multiplication and inversion are dierentiable, and so SL(n)
is a Lie group.
(5) The map A det A is a dierentiable map from O(n) to 1, 1,
and the level set f
1
(1) is
SO(n) = A O(n) [ det A = 1,
the special orthogonal group or the rotation group in R
n
,
which is then an open subset of O(n), and therefore a Lie group of
the same dimension.
(6) We can also consider the space /
nn
(C) of complex n n matri-
ces, and the space GL(n, C) of complex n n invertible matrices.
This is a Lie group of real dimension 2n
2
. Moreover, similarly to
what was done above for O(n), we can take the group of unitary
transformations on C
n
,
U(n) = A /
nn
(C) [ A

A = I,
where A

is the adjoint of A. This group is a submanifold of

/
nn
(C)

= C
n
2

= R
2n
2
, and a Lie group, called the unitary
group. This can be seen from the fact that I is a regular value of
the map f : A A

A from /
nn
(C) to the space of self-adjoint
38 1. DIFFERENTIABLE MANIFOLDS
matrices. As any element of /
nn
(C) can be uniquely written as a
sum of a self-adjoint with an anti-self-adjoint matrix, and the map
A iA is an isomorphism from the space of self-adjoint matrices
to the space of anti-self-adjoint matrices, we conclude that these
two spaces have real dimension
1
2
dim
R
/
nn
(C) = n
2
. Hence,
dimU(n) = n
2
.
(7) The special unitary group
SU(n) = A U(n) [ det A = 1,
is also a Lie group now of dimension n
2
1 (note that A det (A)
is now a dierentiable map from U(n) to S
1
).
As a Lie group G is, by denition, a manifold, we can consider the
tangent space at one of its points. In particular, the tangent space at the
identity e is usually denoted by
g := T
e
G.
For g G, we have the maps
L
g
: G G
h g h
and
R
g
: G G
h h g
which correspond to left multiplication and right multiplication by g.
A vector eld on G is called left-invariant if (L
g
)

X = X for every
g G, that is,
((L
g
)

X)
gh
= X
gh
or (dL
g
)
h
X
h
= X
gh
,
for every g, h G. There is, of course, a vector space isomorphism between
g and the space of left-invariant vector elds on G that, to each V g,
assigns the vector eld X
V
dened by
X
V
g
:= (dL
g
)
e
V,
for any g G. This vector eld is left-invariant as
(dL
g
)
h
X
V
h
= (dL
g
)
h
(dL
h
)
e
V = (d(L
g
L
h
))
e
V = (dL
gh
)
e
V = X
V
gh
.
Note that, given a left-invariant vector eld X, the corresponding element
of g is X
e
. As the space X
L
(G) of left-invariant vector elds is closed under
the Lie bracket of vector elds (because, from Exercise 6.11.8, (L
g
)

[X, Y ] =
[(L
g
)

X, (L
g
)

Y ]), it is a Lie subalgebra of the Lie algebra of vector elds

(see Denition 6.4). The isomorphism X
L
(G)

= g then determines a Lie
algebra structure on g. We call g the Lie algebra of the Lie group G.
Example 7.2.
(1) If G = GL(n), then gl(n) = T
I
GL(n) = /
nn
is the space of nn
matrices with real coecients, and the Lie bracket on gl(n) is the
commutator of matrices
[A, B] = AB BA.
7. LIE GROUPS 39
In fact, if A, B gl(n) are two n n matrices, the corresponding
left-invariant vector elds are given by
X
A
g
= (dL
g
)
I
(A) =

i,k,j
x
ik
a
kj

x
ij
X
B
g
= (dL
g
)
I
(B) =

i,k,j
x
ik
b
kj

x
ij
,
where g GL(n) is a matrix with components x
ij
. The ij-component
of [X
A
, X
B
]
g
is given by X
A
g
(X
B
)
ij
X
B
g
(X
A
)
ij
, i.e.
[X
A
, X
B
]
ij
(g) =
_
_

l,m,p
x
lp
a
pm

x
lm
_
_
_

k
x
ik
b
kj
_

_
_

l,m,p
x
lp
b
pm

x
lm
_
_
_

k
x
ik
a
kj
_
=

k,l,m,p
x
lp
a
pm

il

km
b
kj

k,l,m,p
x
lp
b
pm

il

km
a
kj
=

m,p
x
ip
(a
pm
b
mj
b
pm
a
mj
)
=

p
x
ip
(AB BA)
pj
(where
ij
= 1 if i = j and
ij
= 0 if i ,= j is the Kronecker
symbol). Making g = I, we obtain
[A, B] = [X
A
, X
B
]
I
= AB BA.
From Exercise 6.11.6 we see that this will always be the case when
G is a matrix group, that is, when G is a subgroup of GL(n) for
some n.
(2) If G = O(n) then its Lie algebra is
o(n) = A /
nn
[ A
t
+A = 0.
In fact, we have seen in Example 7.1.3 that O(n) = f
1
(I), where
the identity I is a regular value of the map
f : /
nn
o
nn
A A
t
A.
Hence, o(n) = T
I
G = ker(df)
I
= A /
nn
[ A
t
+ A = 0 is the
space of skew-symmetric matrices.
(3) If G = SL(n) then its Lie algebra is
sl(n) = A /
nn
[ tr A = 0.
40 1. DIFFERENTIABLE MANIFOLDS
In fact, we have seen in Example 7.1.4 that SL(n) = f
1
(1), where
1 is a regular value of the map
f : /
nn
R
A det A.
Hence, sl(n) = T
I
G = ker(df)
I
= A /
nn
[ tr A = 0 is the
space of traceless matrices.
(4) If G = SO(n) = A O(n) [ det A = 1, then its Lie algebra is
so(n) = T
I
SO(n) = T
I
O(n) = o(n).
(5) Similarly to Example 7.2.2, the Lie algebra of U(n) is
u(n) = A /
nn
(C) [ A

+A = 0,
the space of skew-hermitian matrices.
(6) To determine the Lie algebra of SU(n), we see that SU(n) is the
level set f
1
(1), where f(A) = det A, and so
su(n) = ker(df)
I
= A u(n) [ tr(A) = 0.
We now study the ow of a left-invariant vector eld.
Proposition 7.3. Let F be the local ow of a left-invariant vector eld
X at a point h G. Then the map
t
dened by F (that is,
t
(g) = F(g, t))
satises
t
= R
t(e)
. Moreover, the ow of X is globally dened for all t R.
Proof. For g G, R
t(e)
(g) = g
t
(e) = L
g
(
t
(e)). Hence,
R

0
(e)
(g) = g e = g
and
d
dt
_
R
t(e)
(g)
_
=
d
dt
(L
g
(
t
(e))) = (dL
g
)
t(e)
_
d
dt
(
t
(e))
_
= (dL
g
)
t(e)
_
X
t(e)
_
= X
gt(e)
= X
R

t
(e)
(g)
,
implying that R
t(e)
(g) = c
g
(t) =
t
(g) is the integral curve of X at g.
Consequently, if
t
(e) is dened for t (, ), then
t
(g) is dened for
t (, ) and g G. Moreover, condition (3) in Section 6 is true for each
/2 < s, t < /2 and we can extend the map F to G R as before: for
any t R, we write t = k/2 + s where k Z and 0 s < /2, and dene
F(g, t) := F
k
(F(g, s), /2) = gF(e, s)F
k
(e, /2).
Remark 7.4. A homomorphism F : G
1
G
2
between Lie groups is
called a Lie group homomorphism if, besides being a group homomor-
phism, it is also a dierentiable map. Since

t+s
(e) =
s
(
t
(e)) = R
s(e)

t
(e) =
t
(e)
s
(e),
the integral curve t
t
(e) denes a group homomorphism between (R, +)
and (G, ).
7. LIE GROUPS 41
Definition 7.5. The exponential map exp : g G is the map that, to
each V g, assigns the value
1
(e), where
t
is the ow of the left-invariant
vector eld X
V
.
Remark 7.6. If c
g
(t) is the integral curve of X at g and s R, it is easy
to check that c
g
(st) is the integral curve of sX at g. On the other hand, for
V g one has X
sV
= sX
V
. Consequently,

t
(e) = c
e
(t) = c
e
(t 1) = F(e, 1) = exp (tV ),
where F is the ow of tX
V
= X
tV
.
Example 7.7. If G is a group of matrices, then for A g,
exp A = e
A
=

k=0
A
k
k!
.
In fact, this series converges for any matrix A and the map h(t) = e
At
satises
h(0) = e
0
= I
dh
dt
(t) = e
At
A = h(t)A.
Hence, h is the ow of X
A
at the identity (that is, h(t) =
t
(e)), and so
exp A =
1
(e) = e
A
.
Let now G be any group and M be any set. We say that G acts on
M if there is a homomorphism from G to the group of bijective mappings
from M to M, or, equivalently, writing
(g)(p) = A(g, p),
if there is a mapping A : GM M satisfying the following conditions:
(i) if e is the identity in G, then A(e, p) = p, p M;
(ii) if g, h G, then A(g, A(h, p)) = A(gh, p), p M.
Usually we denote A(g, p) by g p.
Example 7.8.
(1) Let G be a group and H G a subgroup. Then H acts on G by
left multiplication: A(h, g) = h g for h H, g G.
(2) GL(n) acts on R
n
through A x = Ax for A GL(n) and x R
n
.
The same is true for any subgroup G GL(n).
For each p M we can dene the orbit of p as the set G p := g p [
g G. If G p = p then p is called a xed point of G. If there is a
point p M whose orbit is all of M (i.e. G p = M), then the action is
said to be transitive. Note that when this happens, there is only one orbit
and, for every p, q M with p ,= q, there is always an element of the group
g G such that q = g p. The manifold M is then called a homogeneous
42 1. DIFFERENTIABLE MANIFOLDS
space of G. The stabilizer (or isotropy subgroup) of a point p M is
the group
G
p
= g G [ g p = p.
The action is called free if all the stabilizers are trivial.
If G is a Lie group and M is a smooth manifold, we say that the action
is smooth if the map A : G M M is dierentiable. In this case, the
map p g p is a dieomorphism. We will always assume the action of
a Lie group on a dierentiable manifold to be smooth. A smooth action is
said to be proper if the map
GM M M
(g, p) (g p, p)
is proper (recall that a map is called proper if the preimage of any compact
set is compact cf. Section 10.5).
Remark 7.9. Note that a smooth action is proper if and only if, given
two convergent sequences p
n
and g
n
p
n
in M, there exists a convergent
subsequence g
n
k
in G. If G is compact this condition is always satised.
The orbits of the action of G on M are equivalence classes of the equiv-
alence relation given by p q q G p (cf. Section 10.1). For that
reason, the quotient (topological) space M/ is usually called the orbit
space of the action, and denoted by M/G.
Proposition 7.10. If the action of a Lie group G on a dierentiable
manifold M is proper, then the orbit space M/G is a Hausdor space.
Proof. The relation p q q G p is an open equivalence relation
(cf. Section 10.1). Indeed, since p g p is a homeomorphism, the set

1
((U)) = g p [ p U and g G =

gG
g U is an open subset of M
for any open set U in M, meaning that (U) is open (here : M M/G
is the quotient map). Therefore we just have to show that the set
R = (p, q) M M [ p q
is closed (cf. Proposition 10.2). This follows from the fact that R is the
image of the map
GM M M
(g, p) (g p, p)
which is continuous and proper, hence closed (cf. Section 10.5).
Under certain conditions the orbit space M/G is naturally a dieren-
tiable manifold.
Theorem 7.11. Let M be a dierentiable manifold equipped with a free
proper action of a Lie group G. Then the orbit space M/G is naturally a
dierentiable manifold of dimension dimM dimG, and the quotient map
: M M/G is a submersion.
7. LIE GROUPS 43
Proof. By the previous proposition, the quotient M/G is Hausdor.
Moreover, this quotient satises the second countability axiom because M
does so and the equivalence relation dened by G is open. It remains to
be shown that M/G has a natural dierentiable structure for which the
quotient map is a submersion. We do this only in the case of a discrete
(i.e. zero-dimensional) Lie group (c.f. Remark 1.2).
In this case, we just have to prove that for each point p M there exists
a neighborhood U p such that g U h U = for g ,= h. This guarantees
that each point [p] M/G has a neighborhood [U] homeomorphic to U,
which we can assume to be a coordinate neighborhood. Since G acts by
dieomorphisms, the dierentiable structure dened in this way does not
depend on the choice of p [p]. Since the charts of M/G are obtained from
charts of M, the overlap maps are smooth. Therefore M/G has a natural
dierentiable structure for which : M M/G is a local dieomorphism
(as the coordinate expression of [
U
: U [U] is the identity map).
Showing that g U h U = for g ,= h is equivalent to showing
that g U U = for g ,= e. Assume that this did not happen for any
neighborhood U p. Then there would exist a sequence of open sets U
n
p
with U
n+1
U
n
,

+
n=1
U
n
= p and a sequence g
n
G e such that
g
n
U
n
U
n
,= . Choose p
n
g
n
U
n
U
n
. Then p
n
= g
n
q
n
for some
q
n
U
n
. We have p
n
p and q
n
p. Since the action is proper, g
n
admits a convergent subsequence g
n
k
. Let g be its limit. Making k +
in q
n
k
= g
n
k
p
n
k
yields g p = p, implying that g = e (the action is free).
Because G is discrete, we would then have g
n
k
= e for suciently large k,
which is a contradiction.
Example 7.12.
(1) Let S
n
= x R
n+1
[

n
i=1
(x
i
)
2
= 1 be equipped with the action
of G = Z
2
= I, I given by I x = x (antipodal map). This
action is proper and free, and so the orbit space S
n
/G is an n-
dimensional manifold. This space is the real projective space RP
n
(cf. Exercise 2.5.8).
(2) The group G = R 0 acts on M = R
n+1
0 by multiplica-
tion: t x = tx. This action is proper and free, and so M/G is a
dierentiable manifold of dimension n (which is again RP
n
).
(3) Consider M = R
n
equipped with an action of G = Z
n
dened by:
(k
1
, . . . , k
n
) (x
1
, . . . , x
n
) = (x
1
+k
1
, . . . , x
n
+k
n
).
This action is proper and free, and so the quotient M/G is a
manifold of dimension n. This space with the quotient dieren-
tiable structure dened in Theorem 7.11 is called the n-torus and
is denoted by T
n
. It is dieomorphic to the product manifold
S
1
S
1
and, when n = 2, is dieomorphic to the torus of
revolution in R
3
.
Quotients by discrete group actions determine coverings of manifolds.
44 1. DIFFERENTIABLE MANIFOLDS
Definition 7.13. A smooth covering of a dierentiable manifold B is
a pair (M, ), where M is a connected dierentiable manifold, : M B
is a surjective local dieomorphism, and, for each p B, there exists a
connected neighborhood U of p in B such that
1
(U) is the union of disjoint
open sets U

M (called slices), and the restrictions

of to U

are
dieomorphisms onto U. The map is called a covering map and M is
called a covering manifold.
Remark 7.14.
(1) It is clear that we must have dimM = dimB.
(2) Note that the collection of mutually disjoint open sets U

must
be countable (M has a countable basis).
(3) The bers
1
(p) M have the discrete topology. Indeed, as
each slice U

is open and intersects

1
(p) in exactly one point,
this point is open in the subspace topology.
Example 7.15.
(1) The map : R S
1
given by
(t) = (cos(2t), sin(2t))
is a smooth covering of S
1
. However, the restriction of this map to
(0, +) is a surjective local dieomorphism which is not a covering
map.
(2) The product of covering maps is clearly a covering map. Thus we
can generalize the above example and obtain a covering of T
n
=
S
1
S
1
by R
n
.
(3) In Example 7.12.1 we have a covering of RP
n
by S
n
.
A dieomorphism h : M M, where M is a covering manifold, is called
a deck transformation (or covering transformation) if h = , or,
equivalently, if each set
1
(p) is carried to itself by h. It can be shown that
the group G of all covering transformations is a discrete Lie group whose
action on M is free and proper.
If the covering manifold M is simply connected (cf. Section 10.5), the
covering is said to be a universal covering. In this case, B is dieomorphic
to M/G. Moreover, G is isomorphic to the fundamental group
1
(B) of
B (cf. Section 10.5).
The Lie Theorem states that for a given Lie algebra g there exists
a unique simply connected Lie group

G whose Lie algebra is g. If a Lie
group G also has g as its Lie algebra, then there exists a unique Lie group
homomorphism :

G G which is a covering map. The group of deck
transformations is, in this case, simply ker(), and hence G is dieomorphic
to

G/ ker(). In fact, G is also isomorphic to

G/ ker(), which has a natural
group structure (ker() is a normal subgroup).
Example 7.16.
7. LIE GROUPS 45
(1) In the universal covering of S
1
of Example 7.15.1 the deck trans-
formations are translations h
k
: t t + k by an integer k, and so
the fundamental group of S
1
is Z.
(2) Similarly, the deck transformations of the universal covering of T
n
are translations by integer vectors (cf. Example 7.15.2), and so the
fundamental group of T
n
is Z
n
.
(3) In the universal covering of RP
n
from Example 7.15.3, the only
deck transformations are the identity and the antipodal map, and
so the fundamental group of RP
n
is Z
2
.
Exercises 7.17.
(1) (a) Given two Lie groups G
1
, G
2
, show that G
1
G
2
(the direct
product of the two groups) is a Lie group with the standard
dierentiable structure on the product.
(b) The circle S
1
can be identied with the set of complex numbers
of absolute value 1. Show that S
1
is a Lie group and conclude
that the n-torus T
n
= S
1
. . . S
1
is also a Lie group.
(2) (a) Show that (R
n
, +) is a Lie group, determine its Lie algebra
and write an expression for the exponential map.
(b) Prove that, if G is an abelian Lie group, then [V, W] = 0 for
all V, W g.
(3) We can identify each point in
H = (x, y) R
2
[ y > 0
with the invertible ane map h : R R given by h(t) = yt+x. The
set of all such maps is a group under composition; consequently,
our identication induces a group structure on H.
(a) Show that the induced group operation is given by
(x, y) (z, w) = (yz +x, yw),
and that H, with this group operation, is a Lie group.
(b) Show that the derivative of the left translation map L
(x,y)
:
H H at a point (z, w) H is represented in the above
coordinates by the matrix
_
dL
(x,y)
_
(z,w)
=
_
y 0
0 y
_
.
Conclude that the left-invariant vector eld X
V
X(H) de-
termined by the vector
V =

x
+

y
h T
(0,1)
H (, R)
is given by
X
V
(x,y)
= y

x
+y

y
.
(c) Given V, W h, compute [V, W].
46 1. DIFFERENTIABLE MANIFOLDS
(d) Determine the ow of the vector eld X
V
, and give an expres-
sion for the exponential map exp : h H.
(e) Conrm your results by rst showing that H is the subgroup
of GL(2) formed by the matrices
_
y x
0 1
_
with y > 0.
(4) Consider the group
SL(2) =
__
a b
c d
_
: ad bc = 1
_
,
which we already know to be a 3-manifold. Making
a = p +q, d = p q, b = r +s, c = r s,
show that SL(2) is dieomorphic to S
1
R
2
.
(5) For A gl(n), consider the dierentiable map
h : R R0
t det e
At
and show that:
(a) this map is a group homomorphism between (R, +) and (R0, );
(b) h

(0) = tr A;
(c) det(e
A
) = e
tr A
.
(6) (a) If A sl(2), show that there is a R iR such that
e
A
= cosh I +
sinh

A.
(b) Show that exp : sl(2) SL(2) is not surjective.
(7) Consider the vector eld X X(R
2
) dened by
X =
_
x
2
+y
2

x
.
(a) Show that the ow of X denes a free action of R on M =
R
2
0.
(b) Describe the topological quotient space M/R. Is the action
above proper?
(8) Let M = S
2
S
2
and consider the diagonal S
1
-action on M given
by
e
i
(u, v) = (e
i
u, e
2i
v),
where, for u S
2
R
3
and e
i
S
1
, e
i
u denotes the rotation
of u by an angle around the z-axis.
(a) Determine the xed points for this action.
(b) What are the possible nontrivial stabilizers?
7. LIE GROUPS 47
(9) Let G be a Lie group and H a closed Lie subgroup, i.e. a subgroup
of G which is also a closed submanifold of G. Show that the action
of H in G dened by A(h, g) = h g is free and proper.
(10) (Grassmannian) Consider the set H GL(n) of invertible matrices
of the form
_
A 0
C B
_
,
where A GL(k), B GL(n k) and C /
(nk)k
.
(a) Show that H is a closed Lie subgroup of GL(n). Therefore H
acts freely and properly on GL(n) (cf. Exercise 7.17.9).
(b) Show that the quotient manifold
Gr(n, k) := GL(n)/H
can be identied with the set of k-dimensional subspaces of R
n
(in particular Gr(n, 1) is just the projective space RP
n1
).
(c) The manifold Gr(n, k) is called the Grassmannian of k-planes
in R
n
. What is its dimension?
(11) Let G and H be Lie groups and F : G H a Lie group homomor-
phism. Show that:
(a) (dF)
e
: g h is a Lie algebra homomorphism;
(b) if (dF)
e
is an isomorphism then F is a local dieomorphism;
(c) if F is a surjective local dieomorphism then F is a covering
map.
(12) (a) Show that R SU(2) is a four dimensional real linear subspace
of /
22
(C), closed under matrix multiplication, with basis
1 =
_
1 0
0 1
_
, i =
_
i 0
0 i
_
,
j =
_
0 1
1 0
_
, k =
_
0 i
i 0
_
,
satisfying i
2
= j
2
= k
2
= ijk = 1. Therefore this space can
be identied with the quaternions (cf. Section 10.1). Show
that SU(2) can be identied with the quaternions of Euclidean
norm equal to 1, and is therefore dieomorphic to S
3
.
(b) Show that if n R
3
is a unit vector, which we identify with a
quaternion with zero real part, then
exp
_
n
2
_
= 1 cos
_

2
_
+nsin
_

2
_
is also a unit quaternion.
48 1. DIFFERENTIABLE MANIFOLDS
(c) Again identifying R
3
with quaternions with zero real part,
show that the map
R
3
R
3
v exp
_
n
2
_
v exp
_

n
2
_
is a rotation by an angle about the axis dened by n.
(d) Show that there exists a surjective homomorphismF : SU(2)
SO(3), and use this to conclude that SU(2) is the universal
covering of SO(3).
(e) What is the fundamental group of SO(3)?
8. Orientability
Let V be a nite dimensional vector space and consider two ordered
bases = b
1
, . . . , b
n
and

= b

1
, . . . , b

n
. There is a unique linear
transformation S : V V such that b

i
= S b
i
for every i = 1, . . . , n. We say
that the two bases are equivalent if det S > 0. This denes an equivalence
relation that divides the set of all ordered basis of V into two equivalence
classes. An orientation for V is an assignment of a positive sign to the
elements of one equivalence class and a negative sign to the elements of the
other. The sign assigned to a basis is called its orientation and the basis
is said to be positively oriented or negatively oriented according to its
sign. It is clear that there are exactly two possible orientations for V .
Remark 8.1.
(1) The ordering of the basis is very important. If we interchange the
positions of two basis vectors we obtain a dierent ordered basis
with the opposite orientation.
(2) An orientation for a zero-dimensional vector space is just an as-
signment of a sign +1 or 1.
(3) We call the standard orientation of R
n
to the orientation that
assigns a positive sign to the standard ordered basis.
An isomorphism A : V W between two oriented vector spaces carries
equivalent ordered bases of V to equivalent ordered bases of W. Hence, for
any ordered basis , the sign of the image A is either always the same as the
sign of or always the opposite. In the rst case, the isomorphism A is said
to be orientation preserving, and in the latter it is called orientation
reversing.
An orientation of a smooth manifold consists on a choice of orientations
for all tangent spaces T
p
M. If dimM = n 1, these orientations have to
t together smoothly, meaning that for each point p M there exists a
parameterization (U, ) around p such that
(d)
x
: R
n
T
(x)
M
preserves the standard orientation of R
n
at each point x U.
8. ORIENTABILITY 49
Remark 8.2. If the dimension of M is zero, an orientation is just an
assignment of a sign (+1 or 1), called orientation number, to each point
p M.
Definition 8.3. A smooth manifold M is said to be orientable if it
admits an orientation.
Proposition 8.4. If a smooth manifold M is connected and orientable
then it admits precisely two orientations.
Proof. We will show that the set of points where two orientations agree
and the set of points where they disagree are both open. Hence, one of them
has to be M and the other the empty set. Let p be a point in M and let
(U

), (U

) be two parameterizations centered at p such that d

is
orientation preserving for the rst orientation and d

is orientation preserv-
ing for the second. The map
_
d(
1

)
_
0
: R
n
R
n
is either orientation
preserving (if the two orientations agree at p) or reversing. In the rst case,
it has positive determinant at 0, and so, by continuity,
_
d(
1

)
_
x
has
positive determinant for x in a neighborhood of 0, implying that the two
orientations agree in a neighborhood of p. Similarly, if
_
d(
1

)
_
0
is
orientation reversing, the determinant of
_
d(
1

)
_
x
is negative in a
neighborhood of 0, and so the two orientations disagree in a neighborhood
of p.
Let O be an orientation for M (i.e. a smooth choice of an orientation O
p
of T
p
M for each p M), and O the opposite orientation (corresponding
to taking the opposite orientation O
p
at each tangent space T
p
M). If O

is another orientation for M, then, for a given point p M, we know that

O

p
agrees either with O
p
or with O
p
(because a vector space has just two
possible orientations). Consequently, O

agrees with either O or O on

M.
An alternative characterization of orientability is given by the following
proposition, whose proof is left as an exercise.
Proposition 8.5. A smooth manifold M is orientable if and only if
there exists an atlas / = (U

) for which all the overlap maps

1

are orientation-preserving.
An oriented manifold is an orientable manifold together with a choice
of an orientation. A map f : M N between two oriented manifolds
with the same dimension is said to be orientation preserving if (df)
p
is
orientation preserving at all points p M, and orientation reversing if
(df)
p
is orientation reversing at all points p M.
Exercises 8.6.
50 1. DIFFERENTIABLE MANIFOLDS
(1) Prove that the relation of being equivalent between ordered basis
of a nite dimensional vector space described above is an equiva-
lence relation.
(2) Show that a dierentiable manifold M is orientable i there exists
an atlas / = (U

) for which all the overlap maps

1

are orientation-preserving.
(3) (a) Show that if a manifold M is covered by two coordinate neigh-
borhoods V
1
and V
2
such that V
1
V
2
is connected, then M is
orientable.
(b) Show that S
n
is orientable.
(4) Let M be an oriented n-dimensional manifold and c : I M a
dierentiable curve. A smooth vector eld along c is a dier-
entiable map V : I TM such that V (t) T
c(t)
M for all t I
(cf. Section 2 in Chapter 3). Show that if V
1
, . . . , V
n
: I M are
smooth vector elds along c such that V
1
(t), . . . , V
n
(t) is a basis of
T
c(t)
M for all t I then all these bases have the same orientation.
(5) We can see the Mobius band as the 2-dimensional submanifold of
R
3
given by the image of the immersion g : (1, 1) R R
3
dened by
g(t, ) =
__
1 +t cos
_

2
__
cos ,
_
1 +t cos
_

2
__
sin , t sin
_

2
__
.
Show that the Mobius band is not orientable.
(6) Let f : M N be a dieomorphism between two smooth man-
ifolds. Show that M is orientable if and only if N is orientable.
If, in addition, both manifolds are connected and oriented, and
(df)
p
: T
p
M T
f(p)
N preserves orientation at one point p M,
show that f is orientation preserving.
(7) Let M and N be two oriented manifolds. We dene an orientation
on the product manifold M N (called product orientation)
in the following way: If = a
1
, . . . , a
m
and = b
1
, . . . , b
n

are ordered bases of T

p
M and T
q
N, we consider the ordered basis
(a
1
, 0), . . . , (a
m
, 0), (0, b
1
), . . . , (0, b
n
) of T
(p,q)
(M N)

= T
p
M
T
q
N. We then dene an orientation on this space by setting the
sign of this basis equal to the product of the signs of and . Show
that this orientation does not depend on the choice of and .
(8) Show that the tangent bundle TM is always orientable, even if M
is not.
(9) (Orientable double covering) Let M be a non-orientable n-dimensional
manifold. For each point p M we consider the set O
p
of the (two)
equivalence classes of bases of T
p
M. Let M be the set
M = (p, O
p
) [ p M, O
p
O
p
.
9. MANIFOLDS WITH BOUNDARY 51
Given a parameterization (U, ) of M consider the maps : U
M dened by
(x
1
, . . . , x
n
) =
_
(x
1
, . . . , x
n
),
_
_

x
1
_
(x)
, . . . ,
_

x
n
_
(x)
__
,
where x = (x
1
, . . . , x
n
) U and
_
_

x
1
_
(x)
, . . . ,
_

x
n
_
(x)
_
repre-
sents the equivalence class of the basis
_
_

x
1
_
(x)
, . . . ,
_

x
n
_
(x)
_
of T
(x)
M.
(a) Show that these maps determine the structure of an orientable
dierentiable manifold of dimension n on M .
(b) Consider the map : M M dened by (p, O
p
) = p. Show
that is dierentiable and surjective. Moreover, show that,
for each p M, there exists a neighborhood V of p with

1
(V ) = W
1
W
2
, where W
1
e W
2
are two disjoint open
subsets of M, such that restricted to W
i
(i = 1, 2) is a
dieomorphism onto V .
(c) Show that M is connected (M is therefore called the ori-
entable double covering of M).
(d) Let : M M be the map dened by (p, O
p
) = (p, O
p
),
where O
p
represents the orientation of T
p
M opposite to O
p
.
Show that is a dieomorphism which reverses orientations
satisfying = and = id.
(e) Show that any simply connected manifold is orientable.
9. Manifolds with Boundary
Let us consider again the closed half space
H
n
= (x
1
, . . . , x
n
) R
n
[ x
n
0
with the topology induced by the usual topology of R
n
. Recall that a map
f : U R
m
dened on an open set U H
n
is said to be dierentiable
if it is the restriction to U of a dierentiable map

f dened on an open
subset of R
n
containing U (cf. Section 10.2). In this case, the derivative
(df)
p
is dened to be (d

f)
p
. Note that this derivative is independent of the
extension used since any two extensions have to agree on U.
Definition 9.1. A smooth n-manifold with boundary is a topological
manifold with boundary of dimension n and a family of parameterizations

: U

H
n
M (that is, homeomorphisms of open sets U

of H
n
onto
open sets of M), such that:
(i) the coordinate neighborhoods cover M, meaning that

(U

) =
M;
(ii) for each pair of indices , such that
W :=

(U

(U

) ,= ,
52 1. DIFFERENTIABLE MANIFOLDS
the overlap maps

:
1

(W)
1

(W)

:
1

(W)
1

(W)
are smooth;
(iii) the family / = (U

) is maximal with respect to (i) and (ii),

meaning that, if
0
: U
0
M is a parameterization such that

0

1
and
1

0
are C

for all in /, then

0
is in /.
Recall that a point in M is said to be a boundary point if it is
on the image of H
n
under some parameterization (that is, if there is a
parameterization : U H
n
M such that (x
1
, . . . , x
n1
, 0) = p for
some (x
1
, . . . , x
n1
) R
n1
), and that the set M of all such points is
called the boundary of M. Notice that dierentiable manifolds are partic-
ular cases of dierentiable manifolds with boundary, for which M = .
Proposition 9.2. The boundary of a smooth n-manifold with boundary
is a dierentiable manifold of dimension n 1.
Proof. Suppose that p is a boundary point of M (an n-manifold with
boundary) and choose a parameterization

: U

H
n
M around
p. Letting V

:=

(U

), we claim that

(U

) = V

, where U

=
U

H
n
and V

= V

M. By denition of boundary, we already know

that

(U

) V

, so we just have to show that V

(U

). Let
q V

and consider a parameterization

: U

around q, mapping
an open subset of H
n
to an open subset of M and such that q

(U

).
If we show that

(U

(U

) we are done. For that, we prove that

_

_
(U

) U

. Indeed, suppose that this map

1

takes
a point x U

to an interior point (in R

n
) of U

. As this map is a
dieomorphism, x would be an interior point (in R
n
) of U

. This, of course,
contradicts the assumption that x U

. Hence,
_

_
(U

) U

and so

(U

(U

).
The map

then restricts to a dieomorphism from U

onto V

,
where we identify U

with an open subset of R

n1
. We obtain in this
way a parameterization around p in M, and it is easily seen that these
parameterizations dene a dierentiable structure on M.
Remark 9.3. In the above proof we saw that the denition of a bound-
ary point does not depend on the parameterization chosen, meaning that, if
there exists a parameterization around p such that p is an image of a point
in H
n
, then any parameterization around p maps a boundary point of H
n
to p.
The denition of orientability can easily be extended to manifolds with
boundary. We then have the following result.
Proposition 9.4. Let M be an orientable manifold with boundary. Then
M is also orientable.
9. MANIFOLDS WITH BOUNDARY 53
Proof. If M is orientable we can choose an atlas (U

) on M for
which the determinants of the derivatives of all overlap maps are positive.
With this atlas we can obtain an atlas (U

) for M in the way

described in the proof of Proposition 9.2. For any overlap map
(
1

)(x
1
, . . . , x
n
) = (y
1
(x
1
, . . . , x
n
), . . . , y
n
(x
1
, . . . , x
n
))
we have
(
1

)(x
1
, . . . , x
n1
, 0) = (y
1
(x
1
, . . . , x
n1
, 0), . . . , y
n1
(x
1
, . . . , x
n1
, 0), 0)
and
(
1

)(x
1
, . . . , x
n1
) = (y
1
(x
1
, . . . , x
n1
, 0), . . . , y
n1
(x
1
, . . . , x
n1
, 0)).
Consequently, denoting (x
1
, . . . , x
n1
, 0) by ( x, 0),
(d(
1

))
( x,0)
=
_
_
(d(
1

))
x
[
+
0 [
y
n
x
n
( x, 0)
_
_
and so
det (d(
1

))
( x,0)
=
y
n
x
n
( x, 0) det (d(
1

))
x
.
However, xing x
1
, , x
n1
, we have that y
n
is positive for positive values
of x
n
and is zero for x
n
= 0. Consequently,
y
n
x
n
( x, 0) > 0, and so
det (d(
1

))
x
> 0.

Hence, choosing an orientation on a manifold with boundary M induces

an orientation on the boundary M. The convenient choice, called the
induced orientation, can be obtained in the following way. For p M
the tangent space T
p
(M) is a subspace of T
p
M of codimension 1. As we
have seen above, considering a coordinate system x : W R
n
around p,
we have x
n
(p) = 0 and (x
1
, . . . , x
n1
) is a coordinate system around p in
M. Setting n
p
:=
_

x
n
_
p
(called an outward pointing vector at p),
the induced orientation on M is dened by assigning a positive sign to an
ordered basis of T
p
(M) whenever the ordered basis n
p
, of T
p
M is
positive, and negative otherwise. Note that, since
y
n
x
n
(
1
(p)) > 0 (in the
above notation), the sign of the last component of n
p
does not depend on the
choice of coordinate system. In general, the induced orientation is not the
one obtained from the charts of M by simply dropping the last coordinate
(in fact, it is (1)
n
times this orientation).
Exercises 9.5.
(1) Show with an example that the product of two manifolds with
boundary is not always a manifold with boundary.
54 1. DIFFERENTIABLE MANIFOLDS
(2) Let M be a manifold without boundary and N a manifold with
boundary. Show that the product MN is a manifold with bound-
ary. What is (M N)?
(3) Show that a dieomorphism between two manifolds with boundary
M and N maps the boundary M dieomorphically onto N.
10. Notes on Chapter 1
10.1. Section 1. We begin by briey reviewing the main concepts and
results from general topology that we will need (see [Mun00] for a detailed
exposition).
(1) A topology on a set M is a collection T of subsets of M having
the following properties:
(i) the sets and M are in T ;
(ii) the union of the elements of any sub-collection of T is in T ;
(iii) the intersection of the elements of any nite sub-collection of
T is in T .
A set M equipped with a topology T is called a topological space.
We say that a subset U M is an open set of M if it belongs to
the topology T . A neighborhood of a point p M is simply an
open set U T containing p. A closed set F M is a set whose
complement M F is open. The interior int A of a subset A M
is the largest open set contained in A, and its closure A is the
smallest closed set containing A. Finally, the subspace topology
on A M is T
A
:= U A
UT
.
(2) A topological space (M, T ) is said to be Hausdor if for each pair
of distinct points p
1
, p
2
M there exist neighborhoods U
1
, U
2
of p
1
and p
2
such that U
1
U
2
= .
(3) A basis for a topology T on M is a collection B T such that, for
each point p M and each open set U containing p, there exists
a basis element B B for which p B U. If B is a basis for a
topology T then any element of T is a union of elements of B. A
topological space (M, T ) is said to satisfy the second countability
axiom if T has a countable basis.
(4) A map f : M N between two topological spaces is said to be
continuous if for each open set U N the preimage f
1
(U) is an
open subset of M. A bijection f is called a homeomorphism if
both f and its inverse f
1
are continuous.
(5) An open cover for a topological space (M, T ) is a collection U

T such that

= M. A subcover is a sub-collection V

which is still an open cover. A topological space is said to

be compact if every open cover admits a nite subcover. A sub-
set A M is said to be a compact subset if it is a compact
topological space for the subspace topology. It is easily seen that
continuous maps carry compact sets to compact sets.
10. NOTES ON CHAPTER 1 55
(6) A topological space is said to be connected if the only subsets
of M which are simultaneously open and closed are and M. A
subset A M is said to be a connected subset if it is a connected
topological space for the subspace topology. It is easily seen that
continuous maps carry connected sets to connected sets.
(7) Let (M, T ) be a topological space. A sequence p
n
in M is said to
converge to p M if for each neighborhood V of p there exists an
N N for which p
n
V for n > N. If (M, T ) is Hausdor, then
a convergent sequence has a unique limit. If in addition (M, T )
is second countable, then F M is closed if and only if every
convergent sequence in F has limit in F, and K M is compact
if and only if every sequence in K has a sublimit in K.
(8) If M and N are topological spaces, the set of all Cartesian products
of open subsets of M by open subsets of N is a basis for a topology
on M N, called the product topology. Note that with this
topology the canonical projections are continuous maps.
(9) An equivalence relation on a set M is a relation with the
following properties:
(i) reexivity: p p for every p M;
(ii) symmetry: if p q then q p;
(iii) transitivity: if p q and q r then p r.
Given a point p M, we dene the equivalence class of p as the
set
[p] = q M [ q p.
Note that p [p] by reexivity. Whenever we have an equivalence
relation on a set M, the corresponding set of equivalence classes
is called the quotient space, and is denoted by M/. There is a
canonical projection : M M/, which maps each element of M
to its equivalence class. If M is a topological space, we can dene
a topology on the quotient space (called the quotient topology)
by letting a subset V M/ be open if and only if the set
1
(V )
is open in M. The map is then continuous for this topology.
We will be interested in knowing whether some quotient spaces are
Hausdor. For that, the following denition will be helpful.
Definition 10.1. An equivalence relation on a topological
space M is called open if the map : M M/ is open, i.e., if
for every open set U M, the set [U] := (U) is open.
We then have
Proposition 10.2. Let be an open equivalence relation on
M and let R = (p, q) M M [ p q. Then the quotient space
is Hausdor if and only if R is closed in M M.
Proof. Assume that R is closed. Let [p], [q] M/ with
[p] ,= [q]. Then p q, and (p, q) / R. As R is closed, there are open
56 1. DIFFERENTIABLE MANIFOLDS
sets U, V containing p, q, respectively, such that (U V ) R = .
This implies that [U] [V ] = . In fact, if there were a point
[r] [U] [V ], then r would be equivalent to points p

U and
q

V (that is p

r and r q

). Therefore we would have

p

(implying that (p

, q

) R), and so (U V ) R would not

be empty. Since [U] and [V ] are open (as is an open equivalence
relation), we conclude that M/ is Hausdor.
Conversely, let us assume that M/ is Hausdor. If (p, q) / R,
then p q and [p] ,= [q], implying the existence of open sets

U,

V
M/ containing [p] and [q], such that

U

V = . The sets U :=

1
(

U) and V :=
1
(

V ) are open in M and (U V ) R = . In

fact, if that was not so, there would exist points p

U and q

V
such that p

. Then we would have [p

] = [q

], contradicting the
fact that

U

V = (as [p

] (U) =

U and [q

] (V ) =

V ).
Since (p, q) UV (MM)R and U V is open, we conclude
that (M M) R is open, and hence R is closed.
10.2. Section 2.
(1) Let us begin by reviewing some facts about dierentiability of maps
on R
n
. A function f : U R dened on an open subset U of
R
n
is said to be continuously dierentiable on U if all partial
derivatives
f
x
1
, . . . ,
f
x
n
exist and are continuous on U. In this
book, the words dierentiable and smooth will be used to mean
innitely dierentiable, that is, all partial derivatives

k
f
x
i
1x
i
k
exist and are continuous on U. Similarly, a map f : U R
m
,
dened on an open subset of R
n
, is said to be dierentiable or
smooth if all coordinate functions f
i
have the same property, that
is, if they all possess continuous partial derivatives of all orders. If
the map f is dierentiable on U, its derivative at each point of
U is the linear map Df : R
n
R
m
represented in the canonical
bases of R
n
and R
m
by the Jacobian matrix
Df =
_

_
f
1
x
1

f
1
x
n
.
.
.
.
.
.
f
m
x
1

f
m
x
n
_

_.
A map f : A R
m
dened on an arbitrary set A R
n
(not
necessarily open) is said to be dierentiable on A is there exists
an open set U A and a dierentiable map

f : U R
m
such that
f =

f[
A
.
10.3. Section 4.
(1) Let E, B and F be smooth manifolds and : E B a dieren-
tiable map. Then : E B is called a ber bundle with basis
B, total space E and ber F if
10. NOTES ON CHAPTER 1 57
(i) the map is surjective;
(ii) there is a covering of B by open sets U

and dieomorphisms

:
1
(U

) U

F such that for every b U

we have

(
1
(b)) = b F.
10.4. Section 5.
(1) (The Inverse Function Theorem) Let f : U R
n
R
n
be a
smooth function and p U such that (df)
p
is a linear isomorphism.
Then there exists an open subset V U containing p such that
f[
V
: V f(V ) is a dieomorphism. Moreover,
(d(f[
V
)
1
)
f(q)
= ((d(f[
V
))
q
)
1
for all q V .
10.5. Section 7.
(1) A group is a set G equipped with a binary operation : GG G
satisfying:
(i) Associativity: g
1
(g
2
g
3
) = (g
1
g
2
) g
3
for all g
1
, g
2
, g
3
G;
(ii) Existence of identity: There exists an element e G such
that e g = g e = g for all g G;
(iii) Existence of inverses: For all g G there exists g
1
G
such that g g
1
= g
1
g = e.
If the group operation is commutative, meaning that g
1
g
2
= g
2
g
1
for all g
1
, g
2
G, the group is said to be abelian. A subset H G
is said to be a subgroup of G if the restriction of to H H is
a binary operation on H, and H with this operation is a group.
A subgroup H G is said to be normal if ghg
1
H for all
g G, h H. A map f : G H between two groups G and H is
said to be a group homomorphism if f(g
1
g
2
) = f(g
1
) f(g
2
)
for all g
1
, g
2
G. An isomorphism is a bijective homomorphism.
The kernel of a group homomorphism f : G H is the subset
ker(f) = g G [ f(g) = e, and is easily seen to be a normal
subgroup of G.
(2) Let M and N be topological manifolds. A map f : M N is
called proper if the preimage f
1
(K) of any compact set K N
is compact. If f is also continuous then f is closed, i.e. f maps
closed sets to closed sets. To see this, let F M be a closed set,
and consider a convergent sequence q
n
in f(F) with q
n
q. It
is easily seen that the closure K of the set q
n
[ n N is compact,
and since f is proper, then so is f
1
(K). For each n N choose
p
n
F such that f(p
n
) = q
n
. Then p
n
f
1
(K), and so p
n
must
have a sublimit p F (since F is closed). If p
n
k
is a subsequence
which converges to p we have q
n
k
= f(p
n
k
) f(p) (because f is
continuous). Therefore q = f(p) f(F), and f(F) is closed.
(3) Let f, g : X Y be two continuous maps between topological
spaces and let I = [0, 1]. We say that f is homotopic to g if
58 1. DIFFERENTIABLE MANIFOLDS
there exists a continuous map H : I X Y such that H(0, x) =
f(x) and H(1, x) = g(x) for every x X. This map is called a
homotopy.
Homotopy of maps forms an equivalence relation in the set of
continuous maps between X and Y . As an application, let us x a
base point p on a manifold M and consider the homotopy classes
of continuous maps f : I M such that f(0) = f(1) = p (these
maps are called loops based at p), with the additional restriction
that H(t, 0) = H(t, 1) = p for all t I. This set of homotopy
classes is called the fundamental group of M relative to the base
point p, and is denoted by
1
(M, p). Among its elements there is
the class of the constant loop based at p, given by f(t) = p
for every t I. Note that the set
1
(M, p) is indeed a group with
operation (composition of loops) dened by [f][g] := [h], where
h : I M is given by
h(t) =
_
f(2t) if t [0,
1
2
]
g(2t 1) if t [
1
2
, 1]
.
The identity element of this group is the equivalence class of the
constant loop based at p.
If M is connected and this is the only class in
1
(M, p), M is
said to be simply connected. This means that every loop through
p can be continuously deformed to the constant loop. This property
does not depend on the choice of point p, and is equivalent to the
condition that any closed path may be continuously deformed to a
constant loop in M.
(4) Quaternions are a generalization of the complex numbers intro-
duced by Hamilton in 1843, when he considered numbers of the
form a +bi +cj +dk with a, b, c, d R and
i
2
= j
2
= k
2
= ijk = 1.
Formally, the set H of quaternions is simply R
4
with
1 = (1, 0, 0, 0)
i = (0, 1, 0, 0)
j = (0, 0, 1, 0)
k = (0, 0, 0, 1)
and the bilinear associative product dened by the Hamilton formu-
las (and the assumption that 1 is the identity). With these deni-
tions, H is a division ring, that is, (H0, ) is a (non-commutative)
group and multiplication is distributive with respect to addition.
The real part of a quaternion a +bi +cj +ik is a, whereas its
vector part is bi +cj +dk. Quaternions with zero vector part are
identied with real numbers, while quaternions with zero real part
10. NOTES ON CHAPTER 1 59
are identied with vectors in R
3
. The norm of a quaternion is the
usual Euclidean norm.
10.6. Bibliographical notes. The material in this chapter is com-
pletely standard, and can be found in almost any book on dierential ge-
ometry (e.g. [Boo03, dC93, GHL04]). Immersions and embeddings are
the starting point of dierential topology, which is studied in [GP73,
Mil97]. Lie groups and Lie algebras are a huge eld of Mathematics, to
which we could not do justice. See for instance [BtD03, DK99, War83].
More details on the fundamental group and covering spaces can be found
for instance in [Mun00].
CHAPTER 2
Dierential Forms
This chapter is devoted to dierential forms, a fundamental tool in
dierential geometry.
Section 1 reviews the notions of tensors and tensor product, and
introduces alternating tensors and their exterior product.
Tensor elds, which are natural generalizations of vector elds, are
discussed in Section 2, where a new operation, the pull-back of a covari-
ant tensor eld by a smooth map, is dened. Section 3 studies elds of
alternating tensors, or dierential forms, and their exterior derivative.
Important ideas such as the Poincare Lemma, de Rham cohomology
or the Lie derivative, which will not be central to the remainder of this
book, are discussed in the exercises.
The integral of a dierential form on a smooth manifold in dened
in Section 4. This makes use of another fundamental tool in dierential
geometry, namely the existence of partitions of unity.
The far-reaching Stokes Theorem is proved in Section 5, and some of
its consequences are explored in the exercises.
Finally, Section 6 studies the relation between orientability and the ex-
istence of special dierential forms, called volume forms.
1. Tensors
Let V be an n-dimensional vector space. A k-tensor on V is a real
multilinear function (meaning linear in each variable) dened on the product
V V of k copies of V . The set of all k-tensors is itself a vector space
and is usually denoted by T
k
(V

).
Example 1.1.
(1) The space of 1-tensors T
1
(V

) is equal to V

, the dual space of
V , that is, the space of real-valued linear functions on V .
(2) The usual inner product on R
n
is an example of a 2-tensor.
(3) The determinant is an n-tensor on R
n
.
Given a k-tensor T and an m-tensor S, we dene their tensor product
as the (k +m)-tensor T S given by
T S(v
1
, . . . , v
k
, v
k+1
, . . . , v
k+m
) := T(v
1
, . . . , v
k
) S(v
k+1
, . . . , v
k+m
).
This operation is bilinear and associative, but not commutative (cf. Exer-
cise 1.15.1).
61
62 2. DIFFERENTIAL FORMS
Proposition 1.2. If T
1
, . . . , T
n
is a basis for T
1
(V

) = V

(the dual
space of V ), then the set T
i
1
T
i
k
[ 1 i
1
, . . . , i
k
n is a basis of
T
k
(V

), and therefore dimT

k
(V

) = n
k
.
Proof. We will rst show that the elements of this set are linearly
independent. If
T :=

i
1
, ,i
k
a
i
1
i
k
T
i
1
T
i
k
= 0,
then, taking the basis v
1
, . . . , v
n
of V dual to T
1
, . . . , T
n
, meaning that
T
i
(v
j
) =
ij
(cf. Section 7.1), we have T(v
j
1
, . . . , v
j
k
) = a
j
1
j
k
= 0 for every
1 j
1
, . . . , j
k
n.
To show that T
i
1
T
i
k
[ 1 i
1
, . . . , i
k
n spans T
k
(V

), we
take any element T T
k
(V

) and consider the k-tensor S dened by
S :=

i
1
, ,i
k
T(v
i
1
, . . . , v
i
k
)T
i
1
T
i
k
.
Clearly, S(v
i
1
, . . . , v
i
k
) = T(v
i
1
, . . . , v
i
k
) for every 1 i
1
, . . . , i
k
n, and so,
by linearity, S = T.
If we consider k-tensors on V

, instead of V , we obtain the space T

k
(V )
(note that (V

)

= V , as it is shown in Section 7.1). These tensors are

called contravariant tensors on V , while the elements of T
k
(V

) are called
covariant tensors on V . Note that the contravariant tensors on V are the
covariant tensors on V

. The words covariant and contravariant are related
to the transformation behavior of the tensor components under a change of
basis in V , as explained in Section 7.1.
We can also consider mixed (k, m)-tensors on V , that is, multilinear
functions dened on the product V V V

V

of k copies
of V and m copies of V

. A (k, m)-tensor is then k times covariant and m
times contravariant on V . The space of all (k, m)-tensors on V is denoted
by T
k,m
(V

, V ).
Remark 1.3.
(1) We can identify the space T
1,1
(V

, V ) with the space of linear maps
from V to V . Indeed, for each element T T
1,1
(V

, V ), we dene
the linear map from V to V , given by v T(v, ). Note that
T(v, ) : V

R is a linear function on V

, that is, an element of

(V

)

= V .
(2) Generalizing the above denition of tensor product to tensors de-
ned on dierent vector spaces, we can dene the spaces T
k
(V

)
T
m
(W

) generated by the tensor products of elements of T

k
(V

) by
elements of T
m
(W

). Note that T
k,m
(V

, V ) = T
k
(V

) T
m
(V ).
We leave it as an exercise to nd a basis for this space.
1. TENSORS 63
A tensor is called alternating if, like the determinant, it changes sign
every time two of its variables are interchanged, that is, if
T(v
1
, . . . , v
i
, . . . , v
j
, . . . , v
k
) = T(v
1
, . . . , v
j
, . . . , v
i
, . . . , v
k
).
The space of all alternating k-tensors is a vector subspace
k
(V

) of T
k
(V

).
Note that, for any alternating k-tensor T, we have T(v
1
, . . . , v
k
) = 0 if
v
i
= v
j
for some i ,= j.
Example 1.4.
(1) All 1-tensors are trivially alternating, that is,
1
(V

) = T
1
(V

) =
V

.
(2) The determinant is an alternating n-tensor on R
n
.
Consider now S
k
, the group of all possible permutations of 1, . . . , k.
If S
k
, we set (v
1
, . . . , v
k
) = (v
(1)
, . . . , v
(k)
). Given a k-tensor T
T
k
(V

) we can dene a new alternating k-tensor, called Alt(T), in the fol-

lowing way:
Alt(T) :=
1
k!

S
k
(sgn ) (T ),
where sgn is +1 or 1 according to whether is an even or an odd permu-
tation. We leave it as an exercise to show that Alt(T) is in fact alternating.
Example 1.5. If T T
3
(V

),
Alt(T)(v
1
, v
2
, v
3
) =
1
6
(T(v
1
, v
2
, v
3
) +T(v
3
, v
1
, v
2
) +T(v
2
, v
3
, v
1
)
T(v
1
, v
3
, v
2
) T(v
2
, v
1
, v
3
) T(v
3
, v
2
, v
1
)) .
We will now dene the wedge product between alternating tensors: if
T
k
(V

) and S
m
(V

), then T S
k+m
(V

) is given by
T S :=
(k +m)!
k! m!
Alt(T S).
Example 1.6. If T, S
1
(V

) = V

, then
T S = 2 Alt(T S) = T S S T,
implying that T S = S T and T T = 0.
It is easy to verify that this product is bilinear. To prove associativity
we need the following proposition.
Proposition 1.7.
(i) Let T T
k
(V

) and S T
m
(V

). If Alt(T) = 0 then
Alt(T S) = Alt(S T) = 0;
(ii) Alt(Alt(T S) R) = Alt(T S R) = Alt(T Alt(S R)).
Proof.
64 2. DIFFERENTIAL FORMS
(i) Let us consider
(k +m)! Alt(T S)(v
1
, . . . , v
k+m
) =

S
k+m
(sgn ) T(v
(1)
, . . . , v
(k)
)S(v
(k+1)
, . . . , v
(k+m)
).
Taking the subgroup Gof S
k+m
formed by the permutations of 1, . . . , k +m
that leave k + 1, . . . , k +m xed, we have

G
(sgn )T(v
(1)
, . . . , v
(k)
)S(v
(k+1)
, . . . , v
(k+m)
) =
=
_

G
(sgn )T(v
(1)
, . . . , v
(k)
)
_
S(v
k+1
, . . . , v
k+m
)
= k! (Alt(T) S) (v
1
, . . . , v
k+m
) = 0.
Then, since G decomposes S
k+m
into disjoint right cosets G := [
G, and for each coset

G
(sgn )(T S)(v
(1)
, . . . , v
(k+m)
) =
= (sgn )

G
(sgn ) (T S)(v
( (1))
, . . . , v
( (k+m))
)
= (sgn )k! (Alt(T) S)(v
(1)
, . . . , v
(k+m)
) = 0,
we have that Alt(T S) = 0. Similarly, we prove that Alt(S T) = 0.
(ii) By linearity of the operator Alt and the fact that Alt Alt = Alt
(cf. Exercise 1.15.3), we have
Alt(Alt(S R) S R) = 0.
Hence, by (i),
0 = Alt(T (Alt(S R) S R))
= Alt(T Alt(S R)) Alt(T S R),
and the result follows.

Using these properties we can show the following.

Proposition 1.8. (T S) R = T (S R).
Proof. By Proposition 1.7, for T
k
(V

), S
m
(V

) and R

l
(V

), we have
(T S) R =
(k +m+l)!
(k +m)! l!
Alt((T S) R)
=
(k +m+l)!
k! m! l!
Alt(T S R)
1. TENSORS 65
and
T (S R) =
(k +m+l)!
k! (m +l)!
Alt(T (S R))
=
(k +m+l)!
k! m! l!
Alt(T S R).

We can now prove the following theorem.

Theorem 1.9. If T
1
, . . . , T
n
is a basis for V

, then the set

T
i
1
T
i
k
[ 1 i
1
< . . . < i
k
n
is a basis for
k
(V

), and
dim
k
(V

) =
_
n
k
_
=
n!
k!(n k)!
.
Proof. Let T
k
(V

) T
k
(V

). By Proposition 1.2,
T =

i
1
,...,i
k
a
i
1
i
k
T
i
1
T
i
k
and, since T is alternating,
T = Alt(T) =

i
1
,...,i
k
a
i
1
i
k
Alt(T
i
1
T
i
k
).
We can show by induction that Alt(T
i
1
T
i
k
) =
1
k!
T
i
1
T
i
2
T
i
k
.
Indeed, for k = 1, the result is trivially true, and, assuming it is true for k
basis tensors, we have, by Proposition 1.7, that
Alt(T
i
1
T
i
k+1
) = Alt(Alt(T
i
1
T
i
k
) T
i
k+1
)
=
k!
(k + 1)!
Alt(T
i
1
T
i
k
) T
i
k+1
=
1
(k + 1)!
T
i
1
T
i
2
T
i
k+1
.
Hence,
T =
1
k!

i
1
,...,i
k
a
i
1
i
k
T
i
1
T
i
2
T
i
k
.
However, the tensors T
i
1
T
i
k
are not linearly independent. Indeed,
due to anticommutativity, if two sequences (i
1
, . . . i
k
) and (j
1
, . . . j
k
) dier
only in their orderings, then T
i
1
T
i
k
= T
j
1
T
j
k
. In addition,
if any two of the indices are equal, then T
i
1
T
i
k
= 0. Hence, we can
avoid repeating terms by considering only increasing index sequences:
T =

i
1
<<i
k
b
i
1
i
k
T
i
1
T
i
k
66 2. DIFFERENTIAL FORMS
and so the set T
i
1
T
i
k
[ 1 i
1
< . . . < i
k
n spans
k
(V

).
Moreover, the elements of this set are linearly independent. Indeed, if
0 = T =

i
1
<<i
k
b
i
1
i
k
T
i
1
T
i
k
,
then, taking a basis v
1
, . . . , v
n
of V dual to T
1
, . . . , T
n
and an increasing
index sequence (j
1
, . . . , j
k
), we have
0 = T(v
j
1
, . . . , v
j
k
)= k!

i
1
<<i
k
b
i
1
i
k
Alt(T
i
1
T
i
k
)(v
j
1
, . . . , v
j
k
)
=

i
1
<<i
k
b
i
1
i
k

S
k
(sgn ) T
i
1
(v
j
(1)
) T
i
k
(v
j
(k)
).
Since (i
1
, . . . , i
k
) and (j
1
, . . . , j
k
) are both increasing, the only term of the
second sum that may be dierent from zero is the one for which = id.
Consequently,
0 = T(v
j
1
, . . . , v
j
k
) = b
j
1
j
k
.

The following result is clear from the anticommutativity shown in Ex-

ample 1.6.
Proposition 1.10. If T
k
(V

) and S
m
(V

), then
T S = (1)
km
S T.
Remark 1.11.
(1) Another consequence of Theorem 1.9 is that dim(
n
(V

)) = 1.
Hence, if V = R
n
, any alternating n-tensor in R
n
is a multiple of
the determinant.
(2) It is also clear that
k
(V

) = 0 if k > n. Moreover, the set
0
(V

)
is dened to be equal to R (identied with the set of constant
functions on V ).
A linear transformation F : V W induces a linear transformation
F

: T
k
(W

) T
k
(V

) dened by
(F

T)(v
1
, . . . , v
k
) = T(F(v
1
), . . . , F(v
k
)).
This map has the following properties.
Proposition 1.12. Let V, W, Z be vector spaces, let F : V W and
H : W Z be linear maps, and let T T
k
(W

) and S T
m
(W

). We
have:
(1) F

(T S) = (F

T) (F

S);
(2) If T is alternating then so is F

T;
(3) F

(T S) = (F

T) (F

S);
(4) (F H)

= H

.
Another important fact about alternating tensors is the following.
1. TENSORS 67
Theorem 1.13. Let F : V V be a linear map and let T
n
(V

).
Then F

T = (det A)T, where A is any matrix representing F.

Proof. As
n
(V

) is 1-dimensional and F

is a linear map, F

is just
multiplication by some constant C. Let us consider an isomorphism H
between V and R
n
. Then, H

det is an alternating n-tensor in V , and so

F

det = CH

det. Hence
(H
1
)

det = C det (H F H
1
)

det = C det A

det = C det,
where A is the matrix representation of F induced by H. Taking the stan-
dard basis in R
n
, e
1
, . . . , e
n
, we have
A

det (e
1
, . . . , e
n
) = C det(e
1
, . . . , e
n
) = C,
and so
det (Ae
1
, . . . , Ae
n
) = C,
implying that C = det A.
Remark 1.14. By the above theorem, if T
n
(V

) and T ,= 0, then
two ordered basis v
1
, . . . , v
n
and w
1
, . . . , w
n
are equivalently oriented if
and only if T(v
1
, . . . , v
n
) and T(w
1
, . . . , w
n
) have the same sign.
Exercises 1.15.
(1) Show that the tensor product is bilinear and associative but not
commutative.
(2) Find a basis for the space T
k,m
(V

, V ) of mixed (k, m)-tensors.
(3) If T T
k
(V

), show that
(a) Alt(T) is an alternating tensor;
(b) if T is alternating then Alt(T) = T;
(c) Alt(Alt(T)) = Alt(T).
(4) Prove Proposition 1.10.
(5) Prove Proposition 1.12.
(6) Let T
1
, . . . , T
k
V

. Show that
(T
1
T
k
)(v
1
, . . . , v
k
) = det [T
i
(v
j
)].
(7) Show that Let T
1
, . . . , T
k

1
(V

) = V

are linearly independent

if and only if T
1
T
k
,= 0.
(8) Let T
k
(V

) and let v V . We dene contraction of T by v,

(v)T, as the (k 1)-tensor given by
((v)T)(v
1
, . . . , v
k1
) = T(v, v
1
, . . . , v
k1
).
Show that:
(a) (v
1
)((v
2
)T) = (v
2
)((v
1
)T);
(b) if T
k
(V

) and S
m
(V

) then
(v)(T S) = ((v)T) S + (1)
k
T ((v)S).
68 2. DIFFERENTIAL FORMS
2. Tensor Fields
The denition of vector eld can be generalized to tensor elds of general
type. For that, we denote by T

p
M the dual of the tangent space T
p
M at a
point p in M (usually called the cotangent space to M at p).
Definition 2.1. A (k, m)-tensor eld is a map that to each point
p M assigns a tensor T T
k,m
(T

p
M, T
p
M).
Example 2.2. A vector eld is a (0, 1)-tensor eld (or a 1-contravariant
tensor eld), that is, a map that to each point p M assigns the 1-
contravariant tensor X
p
T
p
M.
Example 2.3. Let f : M R be a dierentiable function. We can
dene a (1, 0)-tensor eld df which carries each point p M to (df)
p
, where
(df)
p
: T
p
M R
is the derivative of f at p. This tensor eld is called the dierential of f.
For any v T
p
M we have (df)
p
(v) = v f (the directional derivative of f
at p along the vector v). Considering a coordinate system x : W R
n
, we
can write v =

n
i=1
v
i
_

x
i
_
p
, and so
(df)
p
(v) =

i
v
i


f
x
i
(x(p)),
where

f = f x
1
. Taking the coordinate functions x
i
: W R, we can
obtain 1-forms dx
i
dened on W. These satisfy
(dx
i
)
p
_
_

x
j
_
p
_
=
ij
and so they form a basis of each cotangent space T

p
M, dual to the coordinate
basis
_
_

x
1
_
p
, ,
_

x
n
_
p
_
of T
p
M. Hence, any (1, 0)-tensor eld on W
can be written as =

i
dx
i
, where
i
: W R is such that
i
(p) =

p
(
_

x
i
_
p
). In particular, df can be written in the usual way
(df)
p
=
n

i=1

f
x
i
(x(p))(dx
i
)
p
.
Remark 2.4. Similarly to what was done for the tangent bundle, we can
consider the disjoint union of all cotangent spaces and obtain the manifold
T

M =
_
pM
T

p
M
called the cotangent bundle of M. Note that a (1, 0)-tensor eld is just a
map from M to T

M dened by
p
p
T

p
M.
This construction can be easily generalized for arbitrary tensor elds.
2. TENSOR FIELDS 69
The space of (k, m)-tensor elds is clearly a vector space since linear
combinations of (k, m)-tensors are still (k, m)-tensors. If W is a coordinate
neighborhood of M, we know that
_
(dx
i
)
p
_
is a basis for T

p
M and that
_
_

x
i
_
p
_
is a basis for T
p
M. Hence, the value of a (k, m)-tensor eld T at
a point p W can be written as the tensor
T
p
=

a
j
1
jm
i
1
i
k
(p)(dx
i
1
)
p
(dx
i
k
)
p

_

x
j
1
_
p

_

x
jm
_
p
where the a
j
1
jm
i
1
i
k
: W R are functions which at each p W give us the
components of T
p
relative to these bases of T

p
M and T
p
M. Just as we did
with vector elds, we say that a tensor eld is dierentiable if all these
functions are dierentiable for all coordinate systems of the maximal atlas.
Again, we only need to consider the coordinate systems of an atlas, since all
overlap maps are dierentiable (cf. Exercise 2.8.1).
Example 2.5. The dierential of a smooth function f : M R is
clearly a dierentiable (1, 0)-tensor eld, since its components


f
x
i
x on a
given coordinate system x : W R
n
are smooth.
An important operation on covariant tensors is the pull-back by a
smooth map.
Definition 2.6. Let f : M N be a dierentiable map between smooth
manifolds. Then, each dierentiable k-covariant tensor eld T on N denes
a k-covariant tensor eld f

T on M in the following way:

(f

T)
p
(v
1
, . . . , v
k
) = T
f(p)
((df)
p
v
1
, . . . , (df)
p
v
k
),
for v
1
, . . . , v
k
T
p
M.
Remark 2.7. Notice that (f

T)
p
is just the image of T
f(p)
by the linear
map (df)

p
: T
k
(T

f(p)
N) T
k
(T

p
M) induced by (df)
p
: T
p
M T
f(p)
N
(cf. Section 1). Therefore the properties f

(T + S) = (f

T) + (f

S)
and f

(T S) = (f

T) (f

S) hold for all , R and all appropriate

covariant tensor elds T, S. We will see in Exercise 2.8.2 that the pull-back of
a dierentiable covariant tensor eld is still a dierentiable covariant tensor
eld.
Exercises 2.8.
(1) Find the relation between coordinate functions of a tensor eld in
two overlapping coordinate systems.
(2) Show that the pull-back of a dierentiable covariant tensor eld is
still a dierentiable covariant tensor eld.
(3) (Lie derivative of a tensor eld) Given a vector eld X X(M),
we dene the Lie derivative of a k-covariant tensor eld T
along X as
L
X
T :=
d
dt
(
t

T)
[
t=0
,
70 2. DIFFERENTIAL FORMS
where
t
= F(, t) with F the local ow of X at p.
(a) Show that
L
X
(T(Y
1
, . . . , Y
k
)) = (L
X
T)(Y
1
, . . . , Y
k
)
+T(L
X
Y
1
, . . . , Y
k
) +. . . +T(Y
1
, . . . , L
X
Y
k
),
i.e., show that
X (T(Y
1
, . . . , Y
k
)) = (L
X
T)(Y
1
, . . . , Y
k
)
+T([X, Y
1
], . . . , Y
k
) +. . . +T(Y
1
, . . . , [X, Y
k
]),
for all vector elds Y
1
, . . . , Y
k
(cf. Exercises 6.11.10 and 6.11.11
in Chapter 1).
(b) How would you dene the Lie derivative of a (k, m)-tensor
eld?
3. Dierential Forms
Fields of alternating tensors are very important objects called forms.
Definition 3.1. Let M be a smooth manifold. A form of degree k
(or k-form) on M is a eld of alternating k-tensors dened on M, that is,
a map that, to each point p M, assigns an element
p

k
(T

p
M).
The space of k-forms on M is clearly a vector space. By Theorem 1.9,
given a coordinate system x : W R
n
, any k-form on W can be written as
=

I
dx
I
where I = (i
1
, . . . , i
k
) denotes any increasing index sequence of integers
in 1, . . . , n, dx
I
is the form dx
i
1
dx
i
k
, and the
I
s are functions
dened on W. It is easy to check that the components of in the basis
dx
i
1
dx
i
k
are
I
. Therefore is a dierentiable (k, 0)-tensor (in
which case it is called a dierential form) if the functions
I
are smooth
for all coordinate systems of the maximal atlas. The set of dierential k-
forms on M is represented by
k
(M). From now on we will use the word
form to mean a dierential form.
Given a smooth map f : M N between dierentiable manifolds, we
can induce forms on M from forms on N using the pull-back operation
(cf. Denition 2.6), since the pull-back of a eld of alternating tensors is
still a eld of alternating tensors.
Remark 3.2. If g : N R is a 0-form, that is, a function, the pull-back
is dened as f

g = g f.
It is easy to verify that the pull-back of forms satises the following
properties, the proof of which we leave as an exercise:
Proposition 3.3. Let f : M N be a dierentiable map and ,
forms on N. Then,
3. DIFFERENTIAL FORMS 71
(i) f

( +) = f

+f

;
(ii) f

(g) = (g f)f

= (f

g)(f

) for any function g C

(N);
(iii) f

( ) = (f

) (f

);
(iv) g

(f

) = (f g)

for any map g C

(L, M), where L is a

dierentiable manifold.
Example 3.4. If f : M N is dierentiable and we consider coordinate
systems x : V R
m
, y : W R
n
respectively on M and N, we have y
i
=

f
i
(x
1
, . . . , x
m
) for i = 1, . . . , n and

f = y f x
1
the local representation
of f. If =

I

I
dy
I
is a k-form on W, then by Proposition 3.3,
f

= f

I
dy
I
_
=

I
(f

I
)(f

dy
I
) =

I
(
I
f)(f

dy
i
1
) (f

dy
i
k
).
Moreover, for v T
p
M,
(f

(dy
i
))
p
(v) = (dy
i
)
f(p)
((df)
p
v) =
_
d(y
i
f)
_
p
(v),
that is, f

(dy
i
) = d(y
i
f). Hence,
f

I
(
I
f) d(y
i
1
f) d(y
i
k
f)
=

I
(
I
f) d(

f
i
1
x) d(

f
i
k
x).
If k = dimM = dimN = n, then the pull-back f

can easily be computed

from Theorem 1.13, according to which
(4) (f

(dy
1
dy
n
))
p
= det (d

f)
x(p)
(dx
1
dx
n
)
p
.
Given any form on M and a parameterization : U M, we can
consider the pull-back of by and obtain a form dened on the open set
U, called the local representation of on that parameterization.
Example 3.5. Let x : W R
n
be a coordinate system on a smooth
manifold M and consider the 1-form dx
i
dened on W. The pull-back

dx
i
by the corresponding parameterization := x
1
is a 1-form on an
open subset U of R
n
satisfying
(

dx
i
)
x
(v) = (

dx
i
)
x
_
_
n

j=1
v
j
_

x
j
_
x
_
_
= (dx
i
)
p
_
_
n

j=1
v
j
(d)
x
_

x
j
_
x
_
_
= (dx
i
)
p
_
_
n

j=1
v
j
_

x
j
_
p
_
_
= v
i
= (dx
i
)
x
(v),
for x U, p = (x) and v =

n
j=1
v
j
_

x
j
_
x
T
x
U. Hence, just as we had
_

x
i
_
p
= (d)
x
_

x
i
_
x
, we now have (dx
i
)
x
=

(dx
i
)
p
, and so (dx
i
)
p
is the
1-form in W whose local representation on U is (dx
i
)
x
.
72 2. DIFFERENTIAL FORMS
If =

I

I
dx
I
is a k-form dened on an open subset of R
n
, we dene
a (k + 1)-form called exterior derivative of as
d :=

I
d
I
dx
I
.
Example 3.6. Consider the form =
y
x
2
+y
2
dx +
x
x
2
+y
2
dy dened on
R
2
0. Then,
d = d
_

y
x
2
+y
2
_
dx +d
_
x
x
2
+y
2
_
dy
=
y
2
x
2
(x
2
+y
2
)
2
dy dx +
y
2
x
2
(x
2
+y
2
)
2
dx dy = 0.
The exterior derivative satises the following properties:
Proposition 3.7. If , ,
1
,
2
are forms on R
n
, then
(i) d(
1
+
2
) = d
1
+d
2
;
(ii) if is k-form, d( ) = d + (1)
k
d;
(iii) d(d) = 0;
(iv) if f : R
m
R
n
is smooth, d(f

) = f

(d).
Proof. Property (i) is obvious. Using (i), it is enough to prove (ii) for
= a
I
dx
I
and = b
J
dx
J
:
d( ) = d(a
I
b
J
dx
I
dx
J
) = d(a
I
b
J
) dx
I
dx
J
= (b
J
da
I
+a
I
db
J
) dx
I
dx
J
= b
J
da
I
dx
I
dx
J
+a
I
db
J
dx
I
dx
J
= d + (1)
k
a
I
dx
I
db
J
dx
J
= d + (1)
k
d.
Again, to prove (iii), it is enough to consider forms = a
I
dx
I
. Since
d = da
I
dx
I
=
n

i=1
a
I
x
i
dx
i
dx
I
,
we have
d(d) =
n

j=1
n

i=1

2
a
I
x
j
x
i
dx
j
dx
i
dx
I
=
n

i=1

j<i
_

2
a
I
x
j
x
i


2
a
I
x
i
x
j
_
dx
j
dx
i
dx
I
= 0.
To prove (iv), we rst consider a 0-form g:
f

(dg) = f

_
n

i=1
g
x
i
dx
i
_
=
n

i=1
_
g
x
i
f
_
df
i
=
n

i,j=1
__
g
x
i
f
_
f
i
x
j
_
dx
j
=
n

j=1
(g f)
x
j
dx
j
= d(g f) = d(f

g).
3. DIFFERENTIAL FORMS 73
Then, if = a
I
dx
I
, we have
d(f

) = d((f

a
I
)df
I
) = d(f

a
I
) df
I
+ (f

a
I
)d(df
I
) = d(f

a
I
) df
I
= (f

da
I
) (f

dx
I
) = f

(da
I
dx
I
) = f

(d)
(where df
I
denotes the form df
i
1
df
i
k
), and the result follows.
Suppose now that is a dierential k-form on a smooth manifold M. We
dene the (k +1)-form d as the smooth form that is locally represented by
d

for each parameterization

: U

M, where

:=

is the local
representation of , that is, d = (
1

(d

) on

(U). Given another

parameterization

: U

M such that W :=

(U

(U

) ,= , it is
easy to verify that
(
1

.
Setting f equal to
1

, we have
f

(d

) = d(f

) = d

.
Consequently,
(
1

= (
1

(d

)
= (f
1

(d

)
= (
1

(d

),
and so the two denitions agree on the overlapping set W. Therefore d
is well dened. We leave it as an exercise to show that the exterior deriv-
ative dened for forms on smooth manifolds also satises the properties of
Proposition 3.7.
Exercises 3.8.
(1) Prove Proposition 3.3.
(2) (Exterior derivative) Let M be a smooth manifold. Given a k-form
in M we can dene its exterior derivative d without using local
coordinates: given k + 1 vector elds X
1
, . . . , X
k+1
(M),
d(X
1
, . . . , X
k+1
) :=
k+1

i=1
(1)
i1
X
i
(X
1
, . . . ,

X
i
, . . . , X
k+1
)+

i<j
(1)
i+j
([X
i
, X
j
], X
1
, . . . ,

X
i
, . . . ,

X
j
, . . . , X
k+1
),
where the hat indicates an omitted variable.
(a) Show that d dened above is in fact a (k + 1)-form in M,
that is,
(i) d(X
1
, . . . , X
i
+Y
i
, . . . , X
k+1
) =
d(X
1
, . . . , X
i
, . . . , X
k+1
) +d(X
1
, . . . , Y
i
, . . . , X
k+1
);
(ii) d(X
1
, . . . , fX
j
, . . . , X
k+1
)=fd(X
1
, . . . , X
k+1
) for any
dierentiable function f;
(iii) d is alternating.
74 2. DIFFERENTIAL FORMS
(b) Let x : W R
n
be a coordinate system of M and let =

I
a
I
dx
i
1
dx
i
k
be the expression of in these coordi-
nates (where the a
I
s are smooth functions). Show that the
local expression of d is the same as the one used in the local
denition of exterior derivative, that is,
d =

I
da
I
dx
i
1
dx
i
k
.
(3) Show that the exterior derivative dened for forms on smooth man-
ifolds satises the properties of Proposition 3.7.
(4) Show that:
(a) if = f
1
dx +f
2
dy +f
3
dz is a 1-form on R
3
then
d = g
1
dy dz +g
2
dz dx +g
3
dx dy,
where (g
1
, g
2
, g
3
) = curl(f
1
, f
2
, f
3
);
(b) if = f
1
dy dz + f
2
dz dx + f
3
dx dy is a 2-form on R
3
,
then
d = div(f
1
, f
2
, f
3
) dx dy dz.
(5) (De Rham cohomology) A k-form is called closed if d = 0.
If it exists a (k 1)-form such that = d then is called
exact. Note that every exact form is closed. Let Z
k
be the set of
all closed k-forms on M and dene a relation between forms on Z
k
as follows: if and only if they dier by an exact form, that
is, if = d for some (k 1)-form .
(a) Show that this relation is an equivalence relation.
(b) Let H
k
(M) be the corresponding set of equivalence classes
(called the k-dimensional de Rham cohomology space of
M). Show that addition and scalar multiplication of forms
dene indeed a vector space structure on H
k
(M).
(c) Let f : M N be a smooth map. Show that:
(i) the pull-back f

carries closed forms to closed forms and

exact forms to exact forms;
(ii) if on N then f

on M;
(iii) f

induces a linear map on cohomology f

: H
k
(N)
H
k
(M) naturally dened by f

[] = [f

];
(iv) if g : L M is another smooth map, then (f g)

=
g

.
(d) Show that the dimension of H
0
(M) is equal to the number of
connected components of M.
(e) Show that H
k
(M) = 0 for every k > dimM.
(6) Let M be a manifold of dimension n, let U be an open subset of
R
n
and let be a k-form on R U. Writing as
= dt

I
a
I
dx
I
+

J
b
J
dx
J
,
3. DIFFERENTIAL FORMS 75
where I = (i
1
, . . . , i
k1
) and J = (j
1
, . . . , j
k
) are increasing index
sequences, (x
1
, . . . , x
n
) are coordinates in U and t is the coordinate
in R, consider the operator Q dened by
Q()
(t,x)
=

I
__
t
t
0
a
I
ds
_
dx
I
,
which transforms k-forms in R U into (k 1)-forms.
(a) Let f : V U be a dieomorphism between open subsets
of R
n
. Show that the induced dieomorphism

f := id f :
R V RU satises

Q = Q

f

.
(b) Using (a), construct an operator Q which carries k-forms on
RM into (k1)-forms and, for any dieomorphism f : M
N, the induced dieomorphism

f := id f : R M R N
satises

f

Q = Q

f

. Show that this operator is linear.

(c) Considering the operator Qdened in (b) and the inclusion i
t
0
:
M RM of M at the level t
0
, dened by i
t
0
(p) = (t
0
, p),
show that

t
0
= dQ + Qd, where : R M M
is the projection on M.
(d) Show that the maps

: H
k
(M) H
k
(R M) and i

t
0
:
H
k
(RM) H(M) are inverses of each other (and so H
k
(M)
is isomorphic to H
k
(R M)).
(e) Use (d) to show that, for k > 0 and n > 0, every closed k-form
in R
n
is exact, that is, H
k
(R
n
) = 0 if k > 0.
(f) Use (d) to show that, if f, g : M N are two smoothly
homotopic maps between smooth manifolds (meaning that
there exists a smooth map H : RM N such that H(t
0
, p) =
f(p) and H(t
1
, p) = g(p) for some xed t
0
, t
1
R), then
f

= g

.
(g) We say that M is contractible if the identity map id : M
M is smoothly homotopic to a constant map. Show that R
n
is contractible.
(h) (Poincare Lemma) Let M be a contractible smooth manifold.
Show that every closed form on M is exact, that is, H
k
(M) = 0
for all k > 0.
(7) (Lie derivative of a dierential form) Given a vector eld X
X(M), we dene the Lie derivative of a form along X as
L
X
:=
d
dt
(
t

)
[
t=0
,
where
t
= F(, t) with F the local ow of X at p (cf. Exer-
cise 2.8.3). Show that the Lie derivative satises the following
properties:
(a) L
X
(
1

2
) = (L
X

1
)
2
+
1
(L
X

2
);
76 2. DIFFERENTIAL FORMS
(b) d(L
X
) = L
X
(d);
(c) Cartan formula: L
X
= (X)d +d((X));
(d) L
X
((Y )) = (L
X
Y ) +(Y )L
X

(cf. Exercise 6.11.11 on Chapter 1 and Exercise 1.15.8).

4. Integration on Manifolds
Before we see how to integrate dierential forms on manifolds, we will
start by studying the R
n
case. For that let us consider an n-form dened
on an open subset U of R
n
. We already know that can be written as

x
= a(x) dx
1
dx
n
,
where a : U R is a smooth function. The support of is, by denition,
the closure of the set where ,= 0 that is,
supp = x R
n
:
x
,= 0.
We will assume that this set is compact (in which case is said to be
compactly supported). We dene
_
U
=
_
U
a(x) dx
1
dx
n
:=
_
U
a(x) dx
1
dx
n
,
where the integral on the right is a multiple integral on a subset of R
n
. This
denition is almost well-behaved with respect to changes of variables in R
n
.
Indeed, if f : V U is a dieomorphism of open sets of R
n
, we have from
(4) that
f

= (a f)(det df)dy
1
dy
n
,
and so
_
V
f

=
_
V
(a f)(det df)dy
1
dy
n
.
If f is orientation preserving, then det (df) > 0, and the integral on the
right is, by the Change of Variables Theorem for multiple integrals in R
n
(cf. Section 7.2), equal to
_
U
. For this reason, we will only consider ori-
entable manifolds when integrating forms on manifolds. Moreover, we will
also assume that supp is always compact to avoid convergence problems.
Let M be an oriented manifold, and let / = (U

) be an atlas
whose parameterizations are orientation-preserving. Suppose that supp is
contained in some coordinate neighborhood W

(U

). Then we dene
_
M
:=
_
U

=
_
U

.
Note that this does not depend on the choice of coordinate neighborhood: if
supp is contained in some other coordinate neighborhood W

(U

),
then

= f

, where f :=
1

is orientation preserving, and hence

_
U

=
_
U

=
_
U

.
4. INTEGRATION ON MANIFOLDS 77
To dene the integral in the general case we use a partition of unity
(cf. Section 7.2) subordinate to the cover W

of M, i.e., a family of dif-

ferentiable functions on M,
i

iI
, such that:
(i) for every point p M, there exists a neighborhood V of p such
that V supp
i
= except for a nite number of
i
s;
(ii) for every point p M,

iI

i
(p) = 1;
(iii) 0
i
1 and supp
i
W

i
for some element W

i
of the cover.
Because of property (i), supp (being compact) intersects the supports of
only nitely many
i
s. Hence we can assume that I is nite, and then
=
_

iI

i
_
=

iI

i
=

iI

i
with
i
:=
i
and supp
i
W

i
. Consequently we dene:
_
M
:=

iI
_
M

i
=

iI
_
U
i

i
.
Remark 4.1.
(1) When supp is contained in one coordinate neighborhood W, the
two denitions above agree. Indeed,
_
M
=
_
W
=
_
W

iI

i
=
_
U

iI

i
_
=
_
U

iI

i
=

iI
_
U

i
=

iI
_
M

i
,
where we used the linearity of the pull-back and of integration on
R
n
.
(2) The denition of integral is independent of the choice of partition
of unity and the choice of cover. Indeed, if
j

jJ
is another
partition of unity subordinate to another cover

compatible
with the same orientation, we have by (1)

iI
_
M

i
=

iI

jJ
_
M

j

and

jJ
_
M

j
=

jJ

iI
_
M

i

j
.
(3) It is also easy to verify the linearity of the integral, that is,
_
M
a
1
+b
2
= a
_
M

1
+b
_
M

2
.
for a, b R and
1
,
2
two n-forms on M.
(4) The denition of integral can easily be extended to oriented mani-
folds with boundary.
78 2. DIFFERENTIAL FORMS
Exercises 4.2.
(1) Let M be an n-dimensional dierentiable manifold. A subset N
M is said to have zero measure if the sets
1

(N) U

have zero
measure for every parameterization

: U

M in the maximal
atlas.
(a) Prove that in order to show that N M has zero measure it
suces to check that the sets
1

(N) U

have zero measure

for the parameterizations in an arbitrary atlas.
(b) Suppose that M is oriented. Let
n
(M) be compactly
supported and let W = (U) be a coordinate neighborhood
such that M W has zero measure. Show that
_
M
=
_
U

,
where the integral on the right-hand side is dened as above
and always exists.
(2) Let x, y, z be the restrictions of the Cartesian coordinate functions
in R
3
to S
2
, oriented so that (1, 0, 0); (0, 1, 0) is a positively ori-
ented basis of T
(0,0,1)
S
2
, and consider the 2-form
= xdy dz +ydz dx +zdx dy
2
(S
2
).
Compute the integral
_
S
2

using the parameterizations corresponding to

(a) spherical coordinates;
(b) stereographic projection.
(3) Consider the manifolds
S
3
=
_
(x, y, z, w) R
4
: x
2
+y
2
+z
2
+w
2
= 2
_
;
T
2
=
_
(x, y, z, w) R
4
: x
2
+y
2
= z
2
+w
2
= 1
_
.
The submanifold T
2
S
3
splits S
3
into two connected components.
Let M be one of these components and let be the 3-form
= zdx dy dw xdy dz dw.
Compute the two possible values of
_
M
.
(4) Let M and N be n-dimensional manifolds, f : M N an ori-
entation preserving dieomorphism and
n
(N) a compactly
supported form. Prove that
_
N
=
_
M
f

.
5. STOKES THEOREM 79
5. Stokes Theorem
In this section we will prove a very important theorem.
Theorem 5.1. (Stokes) Let M be an n-dimensional oriented smooth
manifold with boundary, let be a (n 1)-dierential form on M with
compact support, and let i : M M be the inclusion of the boundary M
in M. Then
_
M
i

=
_
M
d,
where we consider M with the induced orientation (cf. Section 9 in Chap-
ter 1).
Proof. Let us take a partition of unity
i

iI
subordinate to an open
cover of M by coordinate neighborhoods compatible with the orientation.
Then =

iI

i
, where we can assume I to be nite ( is compactly
supported), and hence
d = d

iI

i
=

iI
d(
i
).
By linearity of the integral we then have,
_
M
d =

iI
_
M
d(
i
) and
_
M
i

iI
_
M
i

(
i
).
Hence, to prove this theorem, it is enough to consider the case where supp
is contained inside one coordinate neighborhood of the cover. Let us then
consider a (n 1)-form with compact support contained in a coordinate
neighborhood W. Let : U W be the corresponding parameterization,
where we can assume U to be bounded (supp(

) is compact). Then, the

representation of on U can be written as

=
n

j=1
a
j
dx
1
dx
j1
dx
j+1
dx
n
,
(where each a
j
: U R is a C

-function), and

d = d

=
n

j=1
(1)
j1
a
j
x
j
dx
1
dx
n
.
The functions a
j
can be extended to C

-functions on H
n
by letting
a
j
(x
1
, , x
n
) =
_
a
j
(x
1
, , x
n
) if (x
1
, . . . , x
n
) U
0 if (x
1
, . . . , x
n
) H
n
U.
80 2. DIFFERENTIAL FORMS
If W M = , then i

= 0. Moreover, if we consider a rectangle I in H

containing U dened by equations b
j
x
j
c
j
(j = 1, . . . , n), we have
_
M
d =
_
U
_
_
n

j=1
(1)
j1
a
j
x
j
_
_
dx
1
dx
n
=
n

j=1
(1)
j1
_
I
a
j
x
j
dx
1
dx
n
=
n

j=1
(1)
j1
_
R
n1
_
_
c
j
b
j
a
j
x
j
dx
j
_
dx
1
dx
j1
dx
j+1
dx
n
=
n

j=1
(1)
j1
_
R
n1
_
a
j
(x
1
, . . . , x
j1
, c
j
, x
j+1
, . . . , x
n
)
a
j
(x
1
, . . . , x
j1
, b
j
, x
j+1
, . . . , x
n
)
_
dx
1
dx
j1
dx
j+1
dx
n
= 0,
where we used the Fubini Theorem (cf. Section 7.3), the Fundamental The-
orem of Calculus and the fact that the a
j
s are zero outside U. We conclude
that, in this case,
_
M
i

=
_
M
d = 0.
If, on the other hand, W M ,= we take a rectangle I containing
U now dened by the equations b
j
x
j
c
j
for j = 1, . . . , n 1, and
0 x
n
c
n
. Then, as in the preceding case, we have
_
M
d=
_
U
_
_
n

j=1
(1)
j1
a
j
x
j
_
_
dx
1
dx
n
=
n

j=1
(1)
j1
_
I
a
j
x
j
dx
1
dx
n
= 0 + (1)
n1
_
R
n1
__
cn
0
a
n
x
n
dx
n
_
dx
1
dx
n1
= (1)
n1
_
R
n1
_
a
n
(x
1
, . . . , x
n1
, c
n
) a
n
(x
1
, . . . , x
n1
, 0)
_
dx
1
dx
n1
= (1)
n
_
R
n1
a
n
(x
1
, . . . , x
n1
, 0) dx
1
dx
n1
.
To compute
_
M
i

we need to consider a parameterization of M dened

on an open subset of R
n1
which preserves the standard orientation on R
n1
when we consider the induced orientation on M. For that, we can for
instance consider the set

U = (x
1
, . . . , x
n1
) R
n1
[ ((1)
n
x
1
, x
2
, . . . , x
n1
, 0) U
and the parameterization :

U :M given by
(x
1
, . . . , x
n1
) :=
_
(1)
n
x
1
, x
2
, . . . , x
n1
, 0
_
.
Recall that the orientation on M obtained from by just dropping the last
coordinate is (1)
n
times the induced orientation on M (cf. Section 9 in
Chapter 1). Therefore gives the correct orientation. The local expression
of i : M M on these coordinates (

i :

U U such that

i =
1
i )
is given by

i(x
1
, . . . , x
n1
) =
_
(1)
n
x
1
, x
2
, . . . , x
n1
, 0
_
.
5. STOKES THEOREM 81
Hence,
_
M
i

=
_

=
_

U
(i )

=
_

U
(

i)

=
_

.
Moreover,

=

i

j=1
a
j
dx
1
dx
j1
dx
j+1
dx
n
=
n

j=1
(a
j

i) d

i
1
d

i
j1
d

i
j+1
d

i
n
= (1)
n
(a
n

i) dx
1
dx
n1
,
since d

i
1
= (1)
n
dx
1
, d

i
n
= 0 and d

i
j
= dx
j
, for j ,= 1 and j ,= n.
Consequently,
_
M
i

= (1)
n
_

U
(a
n

i) dx
1
dx
n1
= (1)
n
_

U
a
n
_
(1)
n
x
1
, x
2
, . . . , x
n1
, 0
_
dx
1
dx
n1
= (1)
n
_
R
n1
a
n
(x
1
, x
2
, . . . , x
n1
, 0) dx
1
dx
n1
=
_
M
d
(where we have used the Change of Variables Theorem).
Remark 5.2. If M is an oriented n-dimensional dierentiable manifold
(that is, a manifold with boundary M = ), it is clear from the proof of
the Stokes Theorem that
_
M
d = 0
for any (n1)-dierential form on M with compact support. This can be
viewed as a particular case of the Stokes Theorem if we dene the integral
over the empty set to be zero.
Exercises 5.3.
(1) Use the Stokes Theorem to conrm the result of Exercise 4.2.3.
(2) (Homotopy invariance of the integral ) Recall that two maps f
0
, f
1
:
M N are said to be smoothly homotopic if there exists a dif-
ferentiable map H : R M N such that H(0, p) = f
0
(p) and
H(1, p) = f
1
(p) (cf. Exercise 3.8.6). If M is a compact oriented
manifold of dimension n and is a closed n-form on N, show that
_
M
f

0
=
_
M
f

1
.
(3) (a) Let X X(S
n
) be a vector eld with no zeros. Show that
H(t, p) = cos(t)p + sin(t)
X
p
|X
p
|
82 2. DIFFERENTIAL FORMS
is a smooth homotopy between the identity map and the an-
tipodal map, where we make use of the identication
X
p
T
p
S
n
T
p
R
n+1

= R
n+1
.
(b) Using the Stokes Theorem, show that
_
S
n
> 0,
where
=
n+1

i=1
(1)
i+1
x
i
dx
1
dx
i1
dx
i+1
dx
n+1
and S
n
= x R
n+1
: |x| 1 has the orientation induced
by the standard orientation of R
n+1
.
(c) Show that if n is even then X cannot exist. What about when
n is odd?
6. Orientation and Volume Forms
In this section we will study the relation between orientation and dier-
ential forms.
Definition 6.1. A volume form (or volume element) on a manifold
M of dimension n is an n-form such that
p
,= 0 for all p M.
The existence of a volume form is equivalent to M being orientable.
Proposition 6.2. A manifold M of dimension n is orientable if and
only if there exists a volume form on M.
Proof. Let be a volume form on M, and consider an atlas (U

).
We can assume without loss of generality that the open sets U

are con-
nected. We will construct a new atlas from this one whose overlap maps
have derivatives with positive determinant. Indeed, considering the repre-
sentation of on one of these open sets U

R
n
, we have

= a

dx
1

dx
n

,
where the function a

cannot vanish, and hence must have a xed sign.

If a

is positive, we keep the corresponding parameterization. If not, we

construct a new parameterization by composing

with the map

(x
1
, . . . , x
n
) (x
1
, x
2
, . . . , x
n
).
Clearly, in these new coordinates, the new function a

is positive. Repeating
this for all coordinate neighborhoods we obtain a new atlas for which all the
functions a

are positive, which we will also denote by (U

). Moreover,
whenever W :=

(U

(U

) ,= , we have

= (
1

. Hence,
a

dx
1

dx
n

= (
1

(a

dx
1

dx
n

)
= (a

)(det(d(
1

))) dx
1

dx
n

6. ORIENTATION AND VOLUME FORMS 83

and so det(d(
1

)) > 0. We conclude that M is orientable.

Conversely, if M is orientable, we consider an atlas (U

) for which
the overlap maps
1

are such that det d(

1

) > 0. Taking a
partition of unity
i

iI
subordinate to the cover of M by the corresponding
coordinate neighborhoods, we may dene the forms

i
:=
i
dx
1
i
dx
n
i
with supp
i
= supp
i

i
(U

i
). Extending these forms to M by making
them zero outside supp
i
, we may dene the form :=

iI

i
. Clearly
is a well dened n-form on M so we just need to show that
p
,= 0
for all p M. Let p be a point in M. There is an i I such that

i
(p) > 0, and so there exist linearly independent vectors v
1
, . . . , v
n
T
p
M
such that (
i
)
p
(v
1
, . . . , v
n
) > 0. Moreover, for all other j Ii we have
(
j
)
p
(v
1
, . . . , v
n
) 0. Indeed, if p /

j
(U

j
), then (
j
)
p
(v
1
, . . . , v
n
) = 0.
On the other hand, if p

j
(U

j
), then by (4)
dx
1
j
dx
n
j
= det(d(
1

i
))dx
1
i
dx
n
i
and hence
(
j
)
p
(v
1
, . . . , v
n
) =

j
(p)

i
(p)
(det(d(
1

i
)))(
i
)
p
(v
1
, . . . , v
n
) 0.
Consequently,
p
(v
1
, . . . , v
n
) > 0, and so is a volume form.
Remark 6.3. Sometimes we call a volume form an orientation. In this
case the orientation on M is the one for which a basis v
1
, . . . , v
n
of T
p
M
is positive if and only if
p
(v
1
, . . . , v
n
) > 0.
If we x a volume form
n
(M) on an orientable manifold M, we can
dene the integral of any compactly supported function f C

(M, R) as
_
M
f :=
_
M
f
(where the orientation of M is determined by ). If M is compact, we dene
its volume to be
vol(M) :=
_
M
1 =
_
M
.
Exercises 6.4.
(1) Show that M N is orientable if and only if both M and N are
orientable.
(2) Let M be a compact oriented manifold with volume element

n
(M). Prove that if f > 0 then
_
M
f > 0. (Remark: In particular,
the volume of a compact manifold is always positive).
(3) Let M be a compact orientable manifold of dimension n, and let
be an (n 1)-form in M.
(a) Show that there exists a point p M for which (d)
p
= 0.
(b) Prove that there exists no immersion f : S
1
R of the unit
circle into R.
84 2. DIFFERENTIAL FORMS
(4) Let f : S
n
S
n
be the antipodal map. Recall that the n-
dimensional projective space is the dierential manifold RP
n
=
S
n
/Z
2
, where the group Z
2
= 1, 1 acts on S
n
through 1 x = x
and (1) x = f(x). Let : S
n
RP
n
be the natural projection.
(a) Prove that
k
(S
n
) is of the form =

for some

k
(RP
n
) if and only if f

= .
(b) Show that RP
n
is orientable if and only if n is odd, and that
in this case,
_
S
n

= 2
_
RP
n
.
(c) Show that for n even the sphere S
n
is the orientable double
covering of RP
n
(cf. Exercise 8.6.9 in Chapter 1).
(5) Let M be a compact oriented manifold with boundary and

n
(M) a volume element. The divergence of a vector eld X
X(M) is the function div(X) such that
L
X
= (div(X))
(cf. Exercise 3.8.7). Show that
_
M
div(X) =
_
M
(X).
(6) (Brouwer Fixed Point Theorem)
(a) Let M be an n-dimensional compact orientable manifold with
boundary M ,= . Show that there exists no smooth map
f : M M satisfying f[
M
= id.
(b) Prove the Brouwer Fixed Point Theorem: Any smooth
map g : B B of the closed ball B := x R
n
: [x[ 1
to itself has a xed point, that is, a point p B such that
g(p) = p. (Hint: For each point x B, consider the ray rx starting at
g(x) and passing through x. There is only one point f(x) dierent from g(x) on
rx B. Consider the map f : B B).
7. Notes on Chapter 2
7.1. Section 1.
(1) Given a nite dimensional vector space V we dene its dual space
as the space of linear functionals on V .
Proposition 7.1. If v
1
, . . . , v
n
is a basis for V then there is
a unique basis T
1
, . . . , T
n
of V

dual to v
1
, . . . , v
n
, that is, such
that T
i
(v
j
) =
ij
.
Proof. By linearity, the equations T
i
(v
j
) =
ij
dene a unique
set of functionals T
i
V

. Indeed, for any v V , we have v =

7. NOTES ON CHAPTER 2 85

n
j=1
a
j
v
j
and so
T
i
(v) =
n

j=1
a
j
T
i
(v
j
) =
n

j=1
a
j

ij
= a
i
.
Moreover, these uniquely dened functionals are linearly indepen-
dent. In fact, if
T :=
n

i=1
b
i
T
i
= 0,
then, for each j = 1, . . . , n, we have
0 = T(v
j
) =
n

i=1
b
i
T
i
(v
j
) = b
j
.
To show that T
1
, . . . , T
n
generates V

, we take any S V

and
set b
i
:= S(v
i
). Then, dening T :=

n
i=1
b
i
T
i
, we see that S(v
j
) =
T(v
j
) for all j = 1, . . . , n. Since v
1
, . . . , v
n
is a basis for V , we
have S = T.
Moreover, if v
1
, . . . , v
n
is a basis for V and T
1
, . . . , T
n
is its
dual basis, then, for any v =

a
j
v
j
V and T =

b
i
T
i
V

, we
have
T(v) =
n

j=i
b
i
T
i
(v) =
n

i,j=1
a
j
b
i
T
i
(v
j
) =
n

i,j=1
a
j
b
i

ij
=
n

i=1
a
i
b
i
.
If we now consider a linear functional F on V

, that is, an element

of (V

)

, we have F(T) = T(v

0
) for some xed vector v
0
V .
Indeed, let v
1
, . . . , v
n
be a basis for V and let T
1
, . . . , T
n
be its
dual basis. Then if T =

n
i=1
b
i
T
i
, we have F(T) =

n
i=1
b
i
F(T
i
).
Denoting the values F(T
i
) by a
i
, we get F(T) =

n
i=1
a
i
b
i
= T(v
0
)
for v
0
=

n
i=1
a
i
v
i
. This establishes a one-to-one correspondence
between (V

)

and V , and allows us to view V as the space of linear

functionals on V

. For v V and T V

, we write v(T) = T(v).

(2) Changing from a basis v
1
, . . . , v
n
to a new basis v

1
, . . . , v

n
in
V , we obtain a change of basis matrix S, whose jth column is
the vector of coordinates of the new basis vector v

j
in the old basis.
We can then write the symbolic matrix equation
(v

1
, . . . , v

n
) = (v
1
, . . . , v
n
)S.
The coordinate (column) vectors a and b of a vector v V (a
contravariant 1-tensor on V ) with respect to the old basis and to
the new basis are related by
b =
_
_
_
b
1
.
.
.
b
n
_
_
_ = S
1
_
_
_
a
1
.
.
.
a
n
_
_
_ = S
1
a,
86 2. DIFFERENTIAL FORMS
since we must have (v

1
, . . . , v

n
)b = (v
1
, . . . , v
n
)a = (v

1
, . . . , v

n
)S
1
a.
On the other hand, if T
1
, . . . , T
n
and T

1
, . . . , T

n
are the dual
bases of v
1
, . . . , v
n
and v

1
, . . . , v

n
, we have
_
_
_
T
1
.
.
.
T
n
_
_
_(v
1
, . . . , v
n
) =
_
_
_
T

1
.
.
.
T

n
_
_
_
_
v

1
, . . . , v

n
_
= I
(where, in the symbolic matrix multiplication above, each coordi-
nate is obtained by applying the covectors to the vectors). Hence,
_
_
_
T
1
.
.
.
T
n
_
_
_
_
v

1
, . . . , v

n
_
S
1
= I S
1
_
_
_
T
1
.
.
.
T
n
_
_
_
_
v

1
, . . . , v

n
_
= I,
implying that
_
_
_
T

1
.
.
.
T

n
_
_
_ = S
1
_
_
_
T
1
.
.
.
T
n
_
_
_.
The coordinate (row) vectors a = (a
1
, . . . , a
n
) and b = (b
1
, . . . , b
n
)
of a 1-tensor T V

(a covariant 1-tensor on V ) with respect to

the old basis T
1
, . . . , T
n
and to the new basis T

1
, . . . , T

n
are
related by
a
_
_
_
T
1
.
.
.
T
n
_
_
_ = b
_
_
_
T

1
.
.
.
T

n
_
_
_ aS
_
_
_
T

1
.
.
.
T

n
_
_
_ = b
_
_
_
T

1
.
.
.
T

n
_
_
_
and so b = aS. Note that the coordinate vectors of the covariant 1-
tensors on V transform like the basis vectors of V (that is, by means
of the matrix S) whereas the coordinate vectors of the contravariant
1-tensors on V transform by means of the inverse of this matrix.
This is the origin of the terms covariant and contravariant.
7.2. Section 4.
(1) (Change of Variables Theorem) Let U, V R
n
be open sets, let
g : U V be a dieomorphism and let f : V R be an integrable
function. Then
_
V
f =
_
U
(f g)[ det dg[.
(2) To dene smooth objects on manifolds it is often useful to dene
them rst on coordinate neighborhoods and then glue the pieces
together by means of a partition of unity.
7. NOTES ON CHAPTER 2 87
Theorem 7.2. Let M be a smooth manifold and 1 an open
cover of M. Then there is a family of dierentiable functions on
M,
i

iI
, such that:
(i) for every point p M, there exists a neighborhood U of p such
that U supp
i
= except for a nite number of
i
s;
(ii) for every point p M,

iI

i
(p) = 1;
(iii) 0
i
1 and supp
i
V for some element V 1.
Remark 7.3. This collection
i
of smooth functions is called
partition of unity subordinate to the cover 1.
Proof. Let us rst assume that M is compact. For every
point p M we consider a coordinate neighborhood W
p
=
p
(U
p
)
around p contained in an element V
p
of 1, such that
p
(0) = p and
B
3
(0) U
p
(where B
3
(0) denotes the ball of radius 3 around 0).
Then we consider the C

-functions (cf. Figure 1)

: R R
x
_
e
1
(x1)(x2)
if 1 < x < 2
0 otherwise
,
h : R R
x
_
2
x
(t) dt
_
2
1
(t) dt
,
: R
n
R
x h([x[) .
Notice that h is a decreasing function with values 0 h(x) 1,
equal to zero for x 2 and equal to 1 for x 1. Hence, we can
consider bump functions
p
: M [0, 1] dened by

p
(q) =
_
_
_
(
1
p
(q)) if q
p
(U
p
)
0 otherwise.
Then supp
p
=
p
(B
2
(0))
p
(B
3
(0)) W
p
is contained inside
an element V
p
of the cover. Moreover,
p
(B
1
(0))
pM
is an open
cover of M and so we can consider a nite subcover
p
i
(B
1
(0))
k
i=1
such that M =
k
i=1

p
i
(B
1
(0)). Finally we take the functions

i
=

p
i

k
j=1

p
j
.
88 2. DIFFERENTIAL FORMS
x

h
1 2
Figure 1
Note that

k
j=1

p
j
(q) ,= 0 since q is necessarily contained inside
some
p
i
(B
1
(0)) and so
i
(q) ,= 0. Moreover, 0
i
1,

i
= 1
and supp
i
= supp
p
i
V
p
i
.
If M is not compact we can use a compact exhaustion, that
is, a sequence K
i

iN
of compact subsets of M such that K
i

int K
i+1
and M =

i=1
K
i
. The partition of unity is then obtained
as follows. The family
p
(B
1
(0))
pM
is a cover of K
1
, so we can
consider a nite subcover of K
1
,
_

p
1
(B
1
(0)), . . . ,
p
k
1
(B
1
(0))
_
.
By induction, we obtain a nite number of points such that
_

p
i
1
(B
1
(0)), . . . ,
p
i
k
i
(B
1
(0))
_
covers K
i
int K
i1
(a compact set). Then, for each i, we consider
the corresponding bump functions

p
i
1
, . . . ,
p
i
k
i
: M [0, 1].
Note that
p
1
i + +
p
i
k
i
> 0 for every q K
i
int K
i1
(as there is
always one of these functions which is dierent from zero). As in the
compact case, we can choose these bump functions so that supp
p
i
j
is contained in some element of 1. We will also choose them so that
supp
p
i
j

_
K
i+1
K
i2
(an open set). Hence,
p
i
j

iN,1jk
i
is
locally nite, meaning that, given a point p M, there exists
an open neighborhood V of p such that only a nite number of
these functions is dierent from zero in V . Consequently, the sum

i=1

k
i
j=1

p
i
j
is a positive, dierentiable function on M. Finally,
7. NOTES ON CHAPTER 2 89
making

i
j
=

p
i
j

i=1

k
i
j=1

p
i
j
,
we obtain the desired partition of unity (subordinate to 1).
Remark 7.4. Compact exhaustions always exist on manifolds.
In fact, if U is a bounded open set of R
n
, one can easily construct
a compact exhaustion K
i

iN
for U by setting
K
i
=
_
x U : dist(x, U)
1
n
_
.
If M is a dierentiable manifold, one can always take a countable
atlas / = (U
j
,
j
)
jN
such that each U
j
is a bounded open set,
thus admitting a compact exhaustion K
j
i

iN
. Therefore
_
_
_
_
i+j=l

j
_
K
j
i
_
_
_
_
lN
is a compact exhaustion of M.
7.3. Section 5. (Fubini Theorem) Let A R
n
and B R
m
be com-
pact intervals and let f : A B R be a continuous function. Then
_
AB
f =
_
A
__
B
f(x, y)dy
1
dy
m
_
dx
1
dx
n
=
_
B
__
A
f(x, y)dx
1
dx
n
_
dy
1
dy
m
.
7.4. Bibliographical notes. The material in this chapter can be found
in most books on dierential geometry (e.g. [Boo03, GHL04]). A text en-
tirely dedicated to dierential forms and their applications is [dC94]. The
study of de Rham cohomology leads to a beautiful and powerful theory,
whose details can be found for instance in [BT82].
CHAPTER 3
Riemannian Manifolds
This chapter initiates the study of Riemannian geometry.
Section 1 introduces the concept of a Riemannian metric on a smooth
manifold, which is simply a tensor eld determining an inner product at each
tangent space. This allows the denitions of length, angle and volume.
Section 2 discusses ane connections, which provide a notion of par-
allelism of vectors along curves, and consequently of geodesics (curves
whose tangent vector is parallel). Riemannian manifolds carry a special con-
nection, called the Levi-Civita connection (Section 3), whose geodesics
have special distance-minimizing properties (Section 4).
The Hopf-Rinow Theorem, relating the properties of a Riemannian
manifold as a metric space to the properties of its geodesics, is proved in
Section 5.
1. Riemannian Manifolds
The metric properties of R
n
(distances, angles, volumes) are determined
by the canonical Cartesian coordinates. In a general dierentiable manifold,
however, there are no such preferred coordinates; to dene distances, angles
and volumes we must add more structure, namely a special tensor eld called
a Riemannian metric.
Definition 1.1. A tensor g T
2
(T

p
M) is said to be
(i) symmetric if g(v, w) = g(w, v) for all v, w T
p
M;
(ii) nondegenerate if g(v, w) = 0 for all w T
p
M implies v = 0;
(iii) positive denite if g(v, v) > 0 for all v T
p
M 0.
A covariant 2-tensor eld g is said to be symmetric, nondegenerate or
positive denite if g
p
is symmetric, nondegenerate or positive denite for all
p M. If x : V R
n
is a local chart, we have
g =
n

i,j=1
g
ij
dx
i
dx
j
in V , where
g
ij
= g
_

x
i
,

x
j
_
.
It is easy to see that g is symmetric, nondegenerate or positive denite if
and only if the matrix (g
ij
) has these properties (see Exercise 1.10.1).
91
92 3. RIEMANNIAN MANIFOLDS
Definition 1.2. A Riemannian metric on a smooth manifold M is
a symmetric positive denite smooth covariant 2-tensor eld g. A smooth
manifold M equipped with a Riemannian metric g is called a Riemannian
manifold, and is denoted by (M, g).
A Riemannian metric is therefore a smooth assignment of an inner prod-
uct to each tangent space. It is usual to write
g
p
(v, w) = v, w
p
.
Example 1.3. (Euclidean n-space) It should be clear that M = R
n
and
g =
n

i=1
dx
i
dx
i
dene a Riemannian manifold.
Proposition 1.4. Let (N, g) be a Riemannian manifold and f : M N
an immersion. Then f

g is a Riemannian metric in M (called the induced

metric).
Proof. We just have to prove that f

g is symmetric and positive de-

nite. Let p M and v, w T
p
M. Since g is symmetric,
(f

g)
p
(v, w) = g
f(p)
((df)
p
v, (df)
p
w) = g
f(p)
((df)
p
w, (df)
p
v) = (f

g)
p
(w, v).
On the other hand, it is clear that (f

g)
p
(v, v) 0, and
(f

g)
p
(v, v) = 0 g
f(p)
((df)
p
v, (df)
p
v) = 0 (df)
p
v = 0 v = 0
(as (df)
p
is injective).
In particular, any submanifold M of a Riemannian manifold (N, g) is
itself a Riemannian manifold. Notice that, in this case, the induced metric
at each point p M is just the restriction of g
p
to T
p
M T
p
N. Since R
n
is a Riemannian manifold (cf. Example 1.3), we see that any submanifold
of R
n
is a Riemannian manifold. The Whitney Theorem then implies that
any manifold admits a Riemannian metric.
It was proved in 1954 by John Nash [Nas56] that any compact n-
dimensional Riemannian manifold can be isometrically embedded in R
N
for N =
n(3n+11)
2
(that is, embedded in such a way that its metric is induced
by the Euclidean metric of R
N
). Gromov [Gro70] later proved that one
can take N =
(n+2)(n+3)
2
. Notice that, for n = 2, Nashs result gives an
isometric embedding of any compact surface in R
17
, and Gromovs in R
10
.
In fact, Gromov has further showed that any surface isometrically embeds
in R
5
. This result cannot be improved, as the real projective plane with the
standard metric (see Exercise 1.10.3) cannot be isometrically embedded into
R
4
.
Example 1.5. The standard metric on
S
n
= x R
n+1
: |x| = 1
1. RIEMANNIAN MANIFOLDS 93
is the metric induced on S
n
by the Euclidean metric on R
n+1
. A parame-
terization of the open set
U = x S
n
: x
n+1
> 0
is for instance
(x
1
, . . . , x
n
) =
_
x
1
, . . . , x
n
,
_
1 (x
1
)
2
. . . (x
n
)
2
_
,
and the corresponding coecients of the metric tensor are
g
ij
=
_

x
i
,

x
j
_
=
ij
+
x
i
x
j
1 (x
1
)
2
. . . (x
n
)
2
.
Two Riemannian manifolds will be regarded as the same if they are
isometric.
Definition 1.6. Let (M, g) and (N, h) be Riemannian manifolds. A
dieomorphism f : M N is said to be an isometry if f

h = g. Similarly,
a local dieomorphism f : M N is said to be a local isometry if f

h = g.
A Riemannian metric allows us to compute the length |v| = v, v
1
2
of any vector v TM (as well as the angle between two vectors with the
same base point). Therefore we can measure the length of curves.
Definition 1.7. If (M, , ) is a Riemannian manifold and c : [a, b]
M is a dierentiable curve, the length of c is
l(c) =
_
b
a
| c(t)|dt.
The length of a curve segment does not depend on the parameterization
(see Exercise 1.10.5).
Recall that if M is an orientable n-dimensional manifold then it possesses
volume elements, that is, dierential forms
n
(M) such that
p
,= 0
for all p M. Clearly, there are as many volume elements as dierentiable
functions f C

(M) without zeros.

Definition 1.8. If (M, g) is an orientable Riemannian manifold,

n
(M) is said to be a Riemannian volume element if

p
(v
1
, . . . , v
n
) = 1
for any orthonormal basis v
1
, . . . , v
n
of T
p
M and all p M.
Notice that if M is connected there exist exactly two Riemannian volume
elements (one for each choice of orientation). Moreover, if is a Riemannian
volume element and x : V R is a chart compatible with the orientation
induced by , one has
= fdx
1
. . . dx
n
for some positive function
f =
_

x
1
, . . . ,

x
n
_
.
94 3. RIEMANNIAN MANIFOLDS
If S is the matrix whose columns are the components of

x
1
, . . . ,

x
n
on
some orthonormal basis with the same orientation, we have
f = det S =
_
det
_
S
2
__1
2
=
_
det
_
S
t
S
__1
2
= (det(g
ij
))
1
2
since clearly S
t
S is the matrix whose (i, j)-th entry is the inner product
g
_

x
i
,

x
j
_
= g
ij
.
A Riemannian metric , on M determines a linear isomorphism
g
:
T
p
M T

p
M for all p M dened by
g
(v)(w) = v, w for all v, w T
p
M.
This extends to an isomorphism between X(M) and
1
(M). In particular,
we have
Definition 1.9. Let (M, g) be a Riemannian manifold and f : M R
a smooth function. The gradient of f is the vector eld grad f associated
to the 1-form df through the isomorphism determined by g.
Exercises 1.10.
(1) Let g =

n
i,j=1
g
ij
dx
i
dx
j
T
2
(T

p
M). Show that:
(a) g is symmetric if and only if g
ij
= g
ji
(i, j = 1, . . . , n);
(b) g is nondegenerate if and only if det(g
ij
) ,= 0;
(c) g is positive denite if and only if (g
ij
) is a positive denite
matrix;
(d) if g is nondegenerate, the map
g
: T
p
M T

p
M given by

g
(v)(w) = g(v, w) for all v, w T
p
M is a linear isomorphism;
(e) if g is positive denite then g is nondegenerate.
(2) Prove that any dierentiable manifold admits a Riemannian struc-
ture without invoking the Whitney Theorem. (Hint: Use partitions of
unity).
(3) (a) Let (M, g) be a Riemannian manifold and let G be a dis-
crete Lie group acting freely and properly on M by isometries.
Show that M/G has a natural Riemannian structure (called
the quotient structure).
(b) How would you dene the standard metric on the standard
n-torus T
n
= R
n
/Z
n
?
(c) How would you dene the standard metric on the real pro-
jective n-space RP
n
= S
n
/Z
2
?
(4) Recall that given a Lie group G and x G, the left translation by
x is the dieomorphism L
x
: G G given by L
x
(y) = xy for all
y G. A Riemannian metric g on G is said to be left-invariant
if L
x
is an isometry for all x G. Show that:
(a) g(, ) , is left-invariant if and only if
v, w
x
= (dL
x
1)
x
v, (dL
x
1)
x
w
e
for all x G and v, w T
x
G, where e G is the identity and
,
e
is an inner product on the Lie algebra g = T
e
G;
(b) the standard metric on S
3
= SU(2) is left-invariant;
1. RIEMANNIAN MANIFOLDS 95
(c) the metric induced on O(n) by the Euclidean metric of /
nn

=
R
n
2
is left-invariant.
(5) We say that a dierentiable curve : [, ] M is obtained from
the curve c : [a, b] M by reparameterization if there exists a
smooth bijection f : [, ] [a, b] (the reparameterization) such
that = c f. Show that if is obtained from c by reparameteri-
zation then l() = l(c).
(6) Let (M, g) be a Riemannian manifold and f C

(M). Show that

if a R is a regular value of f then grad(f) is orthogonal to the
submanifold f
1
(a).
(7) Let (M, g) be an oriented Riemannian manifold with boundary.
(a) Given two 1-forms ,
1
(M), we can dene their inner
product , as the inner product of the associated vector
elds. If k 1, we dene the inner product of :=
1

k
and :=
1

k
(with
1
, . . . ,
k
,
1
, . . . ,
k

1
(M))
to be , = det (
i
,
j
). By linearity, we can dene the
inner product of any two k-forms ,
k
(M). Show that
this inner product is well dened, i.e., does not depend on the
representations of and . Compute , for the following
2-forms in R
3
:
:= a dx dy +b dy dz +c dz dx;
:= e dx dy +f dy dz +g dz dx.
(Remark: For k = 0 we dene the inner product of functions f, g to be the
usual product fg).
(b) (Hodge -operator) Consider the linear isomorphism :
k
T

p
M

nk
T

p
M dened as follows: if
1
, . . . ,
k
,
k+1
, . . . ,
n
is any
positively oriented orthonormal basis of T

p
M then (
1

k
) =
k+1

n
. Show that is well dened.
(c) We can dene :
k
(M)
nk
(M) by setting ()
p
:=
(
p
) for all p M and
k
(M). Write out an expression
for in local coordinates, and show that it is a dierential
form.
(d) Prove that for all f, g C

(M, R) and ,
k
(M)
(i) (f +g ) = f +g ;
(ii) = (1)
k(nk)
;
(iii) = = ,
M
;
(iv) ( ) = ( ) = , ;
(v) , = , ,
where
M
= 1 is the Riemannian volume element determined
by the metric g and the orientation of M.
(e) (Divergence Theorem) Let X X(M) be a vector eld on M
and
X

1
(M) the 1-form determined by X. Dening the
divergence of X to be div X := d
X
, show that if M is
96 3. RIEMANNIAN MANIFOLDS
compact then
_
M
div X
M
=
_
M
X, n
M
where n is the outward-pointing unit vector eld on M.
(f) Assume that M = . Show that
(, ) :=
_
M
,
M
is an inner product on
k
(M). Moreover, show that (, ) =
_
M
=
_
M
and (, ) = (, ).
(g) Dene the linear operator :
k
(M)
k1
(M) as :=
(1)
k
(
1
)d. Show that:
(i) = (1)
n(k+1)+1
d;
(ii) = (1)
k
d;
(iii) = (1)
k+1
d;
(iv) = 0;
(v) (d, ) = (, ).
(h) (Laplacian) Consider the Laplacian operator := d +d :

k
(M)
k
(M). Show that if , are dierential forms and
f is a dierentiable function,
(i) = ;
(ii) (, ) = (, );
(iii) = 0 (d = 0 and = 0);
(iv) f = div (grad(f)) ;
(v) (fg) = fg +gf 2grad(f), grad(g).
(i) A harmonic form is a dierential form such that = 0.
Show that if M is connected then any harmonic function on
M must be constant, and any harmonic n-form (n = dimM)
must be a constant multiple of the volume element
M
.
(j) Assume the following result (Hodge decomposition): Any
k-form on a compact oriented Riemannian manifold M can
be uniquely decomposed in a sum =
H
+ d + , where

H
is a harmonic form,
k1
(M) and
k+1
(M).
Show that any cohomology class on M (cf. Exercise 3.8.5 in
Chapter 2) can be uniquely represented by a harmonic form.
(k) (Green formula) Let M be a compact Riemannian manifold
with boundary. The normal derivative of a smooth map
f : M R is the dierentiable map dened on M by
f
n
:= grad(f), n, where n is the outward-pointing unit nor-
mal vector eld on M. Show that
_
M
(f
1
f
2
f
2
f
1
)
M
=
_
M
_
f
1
f
2
n
f
2
f
1
n
_

M
.
1. RIEMANNIAN MANIFOLDS 97
(l) Let M be a compact Riemannian manifold with boundary, and
suppose that f 0 in MM and that one of the following
boundary conditions holds:
(i) f[
M
0 (Dirichlet condition);
(ii)
f
n
0 (Neumann condition).
Show that f 0 in the rst case, and that f is constant in
the second case.
(8) (Degree of a map) Let M, N be compact, connected oriented man-
ifolds of dimension n, and let f : M N be a smooth map.
(a) Show that there exists a real number k (called the degree of f,
and denoted by deg(f)) such that, for any n-form
n
(N),
_
M
f

= k
_
N
.
(Hint: Use the Hodge decomposition mentioned in Exercise 7).
(b) If f is not surjective then there exists an open set W N
such that f
1
(W) = . Deduce that if f is not surjective
then k = 0.
(c) Show that if f is an orientation-preserving dieomorphism
then deg(f) = 1, and that if f is an orientation-reversing dif-
feomorphism then deg(f) = 1.
(d) Let f : M N be surjective and let q N be a regular value
of f. Show that f
1
(q) is a nite set and that there exists a
neighborhood W of q in N such that f
1
(W) is a disjoint union
of opens sets V
i
of M with f[
V
i
: V
i
W a dieomorphism.
(e) Admitting the existence of a regular value of f, show that
deg(f) is an integer. (Remark: The Sard Theorem guarantees that
the set of critical values of a dierentiable map f between manifolds with the
same dimension has zero measure, which in turn guarantees the existence of a
regular value of f).
(f) What is the degree of the natural projection : S
n
RP
n
for n odd?
(g) Given n N, indicate a smooth map f : S
1
S
1
of degree n.
(h) Let S
n
R
n+1
be the unit sphere with the metric induced by
the Euclidean metric of R
n+1
. Let X be a vector eld tangent
to S
n
such that |X| = 1. Consider the map F
t
: S
n
R
n+1
given by F
t
(x) = cos(t)x + sin(t)X
x
. Show that F
t
is a
smooth map of S
n
on S
n
, and dene k(t) = deg(F
t
). Show
that the map t k(t) is continuous.
(i) What are the values of k(0) and k(1)? Show that if n is even
then there exists no vector eld X on S
n
such that X
p
,= 0 for
all p S
n
.
98 3. RIEMANNIAN MANIFOLDS
2. Ane Connections
If X and Y are vector elds in Euclidean space, we can dene the di-
rectional derivative
X
Y of Y along X. This denition, however, uses
the existence of Cartesian coordinates, which no longer holds in a general
manifold. To overcome this diculty we must introduce more structure:
Definition 2.1. Let M be a dierentiable manifold. An ane con-
nection on M is a map : X(M) X(M) X(M) such that
(i)
fX+gY
Z = f
X
Z +g
Y
Z;
(ii)
X
(Y +Z) =
X
Y +
X
Z;
(iii)
X
(fY ) = (X f)Y +f
X
Y
for all X, Y, Z X(M) and f, g C

(M, R) (we write

X
Y := (X, Y )).
The vector eld
X
Y is sometimes known as the covariant derivative
of Y along X.
Proposition 2.2. Let be an ane connection on M, let X, Y X(M)
and p M. Then (
X
Y )
p
T
p
M depends only on X
p
and on the values
of Y along a curve tangent to X at p. Moreover, if x : W R
n
are local
coordinates on some open set W M and
X =
n

i=1
X
i

x
i
, Y =
n

i=1
Y
i

x
i
on this set, we have
(5)
X
Y =
n

i=1
_
_
X Y
i
+
n

j,k=1

i
jk
X
j
Y
k
_
_

x
i
where the n
3
dierentiable functions
i
jk
: W R, called the Christoel
symbols, are dened by
(6)
x
j

x
k
=
n

i=1

i
jk

x
i
.
Proof. It is easy to show that an ane connection is local, that is, if
X, Y X(M) coincide with

X,

Y X(M) in some open set W M then

X
Y =

Y on W (see Exercise 2.6.1). Consequently, we can compute

X
Y for vector elds X, Y dened on W only. Let W be a coordinate neigh-
borhood for the local coordinates x : W R
n
, and dene the Christoel
symbols associated with these local coordinates through (6). Writing out

X
Y =

n
i=1
X
i
x
i

_
_
n

j=1
Y
j

x
j
_
_
and using the properties listed in denition (2.1) yields (5). This formula
shows that (
X
Y )
p
depends only on X
i
(p), Y
i
(p) and (XY
i
)(p). Moreover,
2. AFFINE CONNECTIONS 99
X
i
(p) and Y
i
(p) depend only on X
p
and Y
p
, and (X Y
i
)(p) =
d
dt
Y
i
(c(t))
[
t=0
depends only on the values of Y
i
(or Y ) along a curve c whose tangent vector
at p = c(0) is X
p
.
Remark 2.3. Locally, an ane connection is uniquely determined by
specifying its Christoel symbols on a coordinate neighborhood. However,
the choices of Christoel symbols on dierent charts are not independent,
as the covariant derivative must agree on the overlap.
A vector eld dened along a dierentiable curve c : I M is a
dierentiable map V : I TM such that V (t) T
c(t)
M for all t I. An
obvious example is the tangent vector c(t). If V is a vector eld dened along
the dierentiable curve c : I M with c ,= 0, its covariant derivative
along c is the vector eld dened along c given by
DV
dt
(t) :=
c(t)
V = (
X
Y )
c(t)
for any vector elds X, Y X(M) such that X
c(t)
= c(t) and Y
c(s)
= V (s)
with s (t , t +) for some > 0. Note that if c(t) ,= 0 such extensions
always exist. Proposition 2.2 guarantees that (
X
Y )
c(t)
does not depend
on the choice of X, Y . In fact, if in local coordinates x : W R
n
we have
x
i
(t) := x
i
(c(t)) and
V (t) =
n

i=1
V
i
(t)
_

x
i
_
c(t)
,
then
DV
dt
(t) =
n

i=1
_
_
V
i
(t) +
n

j,k=1

i
jk
(c(t)) x
j
(t)V
k
(t)
_
_
_

x
i
_
c(t)
.
Definition 2.4. A vector eld V dened along a curve c : I M is
said to be parallel along c if
DV
dt
(t) = 0
for all t I. The curve c is called a geodesic of the connection if c is
parallel along c, i.e., if
D c
dt
(t) = 0
for all t I.
In local coordinates x : W R
n
, the condition for V to be parallel
along c is written as
(7)

V
i
+
n

j,k=1

i
jk
x
j
V
k
= 0 (i = 1, . . . , n).
This is a system of rst order linear ODEs for the components of V . By
the Picard-Lindelof Theorem, given a curve c : I M, a point p c(I) and
100 3. RIEMANNIAN MANIFOLDS
a vector v T
p
M, there exists a unique vector eld V : I TM parallel
along c such that V (0) = v, which is called the parallel transport of v
along c.
Moreover, the geodesic equations are
(8) x
i
+
n

j,k=1

i
jk
x
j
x
k
= 0 (i = 1, . . . , n).
This is a system of second order (nonlinear) ODEs for the coordinates of
c(t). Therefore the Picard-Lindelof Theorem implies that, given a point
p M and a vector v T
p
M, there exists a unique geodesic c : I M,
dened on a maximal open interval I such that 0 I, satisfying c(0) = p
and c(0) = v.
We will now dene the torsion of an ane connection . For that, we
note that, in local coordinates x : W R
n
, we have

X
Y
Y
X =
n

i=1
_
_
X Y
i
Y X
i
+
n

j,k=1

i
jk
_
X
j
Y
k
Y
j
X
k
_
_
_

x
i
= [X, Y ] +
n

i,j,k=1
_

i
jk

i
kj
_
X
j
Y
k

x
i
.
Definition 2.5. The torsion operator of a connection on M is the
operator T : X(M) X(M) X(M) given by
T(X, Y ) =
X
Y
Y
X [X, Y ],
for all X, Y X(M). The connection is said to be symmetric if T = 0.
The local expression of T(X, Y ) makes it clear that T(X, Y )
p
depends
linearly on X
p
and Y
p
. In other words, T is the (2, 1)-tensor eld on M
given in local coordinates by
T =
n

i,j,k=1
_

i
jk

i
kj
_
dx
j
dx
k

x
i
(recall from Remark 1.3 in Chapter 2 that any (2, 1)-tensor T T
2,1
(V

, V )
is naturally identied with a bilinear map
T
: V

V

V

= V

through

T
(v, w)() := T(v, w, ) for all v, w V, V

).
Notice that the connection is symmetric if and only if
X
Y
Y
X =
[X, Y ] for all X, Y X(M). In local coordinates, the condition for the
connection to be symmetric is

i
jk
=
i
kj
(i, j, k = 1, . . . , n)
(hence the name).
Exercises 2.6.
3. LEVI-CIVITA CONNECTION 101
(1) (a) Show that if X, Y X(M) coincide with

X,

Y X(M) in some
open set W M then
X
Y =

Y on W. (Hint: Use bump

functions with support contained on W and the properties listed in denition
(2.1)).
(b) Obtain the local coordinate expression (5) for
X
Y .
(c) Obtain the local coordinate equations (7) for the parallel trans-
port law.
(d) Obtain the local coordinate equations (8) for the geodesics of
the connection .
(2) Determine all ane connections on R
n
. Of these, determine the
connections whose geodesics are straight lines c(t) = at + b (with
a, b R
n
).
(3) Let be an ane connection on M. If
1
(M) and X X(M),
we dene the covariant derivative of along X,
X

1
(M),
by

X
(Y ) = X ((Y )) (
X
Y )
for all Y X(M).
(a) Show that this formula denes indeed a 1-form, i.e., show that
(
X
(Y )) (p) is a linear function of Y
p
.
(b) Show that
(i)
fX+gY
= f
X
+g
Y
;
(ii)
X
( +) =
X
+
X
;
(iii)
X
(f) = (X f) +f
X

for all X, Y X(M), f, g C

(M) and ,
1
(M).
(c) Let x : W R
n
be local coordinates on an open set W M,
and take
=
n

i=1

i
dx
i
.
Show that

X
=
n

i=1
_
_
X
i

j,k=1

k
ji
X
j

k
_
_
dx
i
.
(d) Dene the covariant derivative
X
T for an arbitrary tensor
eld T in M, and write its expression in local coordinates.
3. Levi-Civita Connection
In the case of a Riemannian manifold, there is a particular choice of con-
nection, called the Levi-Civita connection, with special geometric prop-
erties.
Definition 3.1. A connection in a Riemannian manifold (M, , )
is said to be compatible with the metric if
X Y, Z =
X
Y, Z +Y,
X
Z
102 3. RIEMANNIAN MANIFOLDS
for all X, Y, Z X(M).
If is compatible with the metric, then the inner product of two vector
elds V
1
and V
2
, parallel along a curve, is constant along the curve:
d
dt
V
1
(t), V
2
(t) =

c(t)
V
1
(t), V
2
(t)
_
+

V
1
(t),
c(t)
V
2
(t)
_
= 0.
In particular, parallel transport preserves lengths of vectors and angles be-
tween vectors. Therefore, if c : I M is a geodesic, then | c(t)| = k is
constant. If a I, the length s of the geodesic between a and t is
s =
_
t
a
| c(u)| du =
_
t
a
k du = k(t a).
In other words, t is an ane function of the arclength s (and is therefore
called an ane parameter). In particular, this shows that the parameters
of two geodesics with the same image are ane functions of each other).
Theorem 3.2. (Levi-Civita) If (M, , ) is a Riemannian manifold then
there exists a unique connection on M which is symmetric and compatible
with , . In local coordinates (x
1
, . . . , x
n
), the Christoel symbols for this
connection are

i
jk
=
1
2
n

l=1
g
il
_
g
kl
x
j
+
g
jl
x
k

g
jk
x
l
_
(9)
where
_
g
ij
_
= (g
ij
)
1
.
Proof. Let X, Y, Z X(M). If the Levi-Civita connection exists then
we must have
X Y, Z =
X
Y, Z +Y,
X
Z;
Y X, Z =
Y
X, Z +X,
Y
Z;
Z X, Y =
Z
X, Y X,
Z
Y ,
as is compatible with the metric. Moreover, since is symmetric, we
must also have
[X, Z], Y =
X
Z, Y +
Z
X, Y ,
[Y, Z], X =
Y
Z, X +
Z
Y, X,
[X, Y ], Z =
X
Y, Z
Y
X, Z.
Adding these six equalities, we obtain the Koszul formula
2
X
Y, Z = X Y, Z +Y X, Z Z X, Y
[X, Z], Y [Y, Z], X +[X, Y ], Z.
Since , is nondegenerate and Z is arbitrary, this formula determines

X
Y . Thus, if the Levi-Civita connection exists, it must be unique.
3. LEVI-CIVITA CONNECTION 103
To prove existence, we dene
X
Y through the Koszul formula. It is
not dicult to show that this indeed denes a connection (cf. Exercise 3.3.1).
Also, using this formula, we obtain
2
X
Y
Y
X, Z = 2
X
Y, Z 2
Y
X, Z = 2[X, Y ], Z
for all X, Y, Z X(M), and hence is symmetric. Finally, again using the
Koszul formula, we have
2
X
Y, Z + 2Y,
X
Z = 2X Y, Z
and therefore the connection dened by this formula is compatible with the
metric.
Choosing local coordinates (x
1
, . . . , x
n
), we have
_

x
i
,

x
j
_
= 0 and
_

x
i
,

x
j
_
= g
ij
.
Therefore the Koszul formula yields
2
_

x
j

x
k
,

x
l
_
=

x
j
g
kl
+

x
k
g
jl

x
l
g
jk

_
n

i=1

i
jk

x
i
,

x
l
_
=
1
2
_
g
kl
x
j
+
g
jl
x
k

g
jk
x
l
_

i=1
g
il

i
jk
=
1
2
_
g
kl
x
j
+
g
jl
x
k

g
jk
x
l
_
.
This linear system is easily solved to give (9).
Exercises 3.3.
(1) Show that the Koszul formula denes a connection.
(2) We introduce in R
3
, with the usual Euclidean metric , , the con-
nection dened in Cartesian coordinates (x
1
, x
2
, x
3
) by

i
jk
=
ijk
,
where : R
3
R is a smooth function and

ijk
=
_
_
_
+1 if (i, j, k) is an even permutation of (1, 2, 3)
1 if (i, j, k) is an odd permutation of (1, 2, 3)
0 otherwise.
Show that:
(a) is compatible with , ;
(b) the geodesics of are straight lines;
(c) the torsion of is not zero in all points where ,= 0 (therefore
is not the Levi-Civita connection unless 0);
(d) the parallel transport equation is

V
i
+
3

j,k=1

ijk
x
j
V
k
= 0

V +( x V ) = 0
104 3. RIEMANNIAN MANIFOLDS
(where is the cross product in R
3
); therefore, a vector paral-
lel along a straight line rotates about it with angular velocity
x.
(3) Let (M, g) and (N, g) be isometric Riemannian manifolds with Levi-
Civita connections and

, and let f : M N be an isometry.
Show that:
(a) f

X
Y =

fX
f

Y for all X, Y X(M);

(b) if c : I M is a geodesic then f c : I N is also a geodesic.
(4) Consider the usual local coordinates (, ) in S
2
R
3
dened by
the parameterization : (0, ) (0, 2) R
3
given by
(, ) = (sin cos , sin sin , cos ).
(a) Using these coordinates, determine the expression of the Rie-
mannian metric induced on S
2
by the Euclidean metric of R
3
.
(b) Compute the Christoel symbols for the Levi-Civita connec-
tion in these coordinates.
(c) Show that the equator is the image of a geodesic.
(d) Show that any rotation about an axis through the origin in R
3
induces an isometry of S
2
.
(e) Show that the geodesics of S
2
traverse great circles.
(f) Find a geodesic triangle (i.e. a triangle whose sides are im-
ages of geodesics) whose internal angles add up to
3
2
.
(g) Let c : R S
2
be given by c(t) = (sin
0
cos t, sin
0
sin t, cos
0
),
where
0

_
0,

2
_
(therefore c is not a geodesic). Let V be a
vector eld parallel along c such that V (0) =

is well
dened at (sin
0
, 0, cos
0
) by continuity). Compute the an-
gle by which V is rotated when it returns to the initial point.
(Remark: The angle you have computed is exactly the angle by which the
oscillation plane of the Foucault pendulum rotates during a day in a place at
latitude

2

0
, as it tries to remain xed with respect to the stars in a rotating
Earth).
(h) Use this result to prove that no open set U S
2
is isometric
to an open set W R
2
with the Euclidean metric.
(i) Given a geodesic c : R R
2
of R
2
with the Euclidean metric
and a point p / c(R), there exists a unique geodesic c : R R
2
(up to reparameterization) such that p c(R) and c(R)
c(R) = (parallel postulate). Is this true in S
2
?
(5) Recall that identifying each point in
H = (x, y) R
2
[ y > 0
with the invertible ane map h : R R given by h(t) = yt + x
induces a Lie group structure on H (cf. Exercise 7.17.3 in Chap-
ter 1).
3. LEVI-CIVITA CONNECTION 105
(a) Determine the left-invariant metric induced by the Euclidean
inner product
g = dx dx +dy dy
in h = T
(0,1)
H (cf. Exercise 1.10.4). (Remark: H endowed with this
metric is called the hyperbolic plane).
(b) Compute the Christoel symbols of the Levi-Civita connection
in the coordinates (x, y).
(c) Show that the curves , : R H given in these coordinates
by
(t) =
_
0, e
t
_
(t) =
_
tanh t,
1
cosh t
_
are geodesics. What are the sets (R) and (R)?
(d) Determine all images of geodesics.
(e) Show that, given two points p, q H, there exists a unique
geodesic through them (up to reparameterization).
(f) Give examples of connected Riemannian manifolds contain-
ing two points through which there are (i) innitely many
geodesics (up to reparameterization); (ii) no geodesics.
(g) Show that no open set U H is isometric to an open set V
R
2
with the Euclidean metric. (Hint: Show that in any neighborhood
of any point p H there is always a geodesic quadrilateral whose internal angles
add up to less than 2).
(h) Does the parallel postulate hold in the hyperbolic plane?
(6) Let (M, , ) be a Riemannian manifold with Levi-Civita connec-
tion

, and let (N, , ) be a submanifold with the induced met-
ric and Levi-Civita connection .
(a) Show that

X
Y =
_

Y
_

for all X, Y X(N), where

X,

Y are any extensions of X, Y

to X(M) and

: TM[
N
TN is the orthogonal projection.
(b) Use this result to indicate curves that are, and curves that are
not, geodesics of the following surfaces in R
3
:
(i) the sphere S
2
;
(ii) the torus of revolution;
(iii) the surface of a cone;
(iv) a general surface of revolution.
(c) Show that if two surfaces in R
3
are tangent along a curve,
then the parallel transport of vectors along this curve in both
surfaces coincides.
(d) Use this result to compute the angle by which a vector
V is rotated when it is parallel transported along a circle on
106 3. RIEMANNIAN MANIFOLDS
the sphere. (Hint: Consider the cone which is tangent to the sphere along
the circle (cf. Figure 1); notice that the cone minus a ray through the vertex is
isometric to an open set of the Euclidean plane).
V
0
V
0
V
0
V
V

Figure 1. Parallel transport along a circle on the sphere.

(7) Let (M, g) be a Riemannian manifold with Levi-Civita connection
. Show that g is parallel along any curve, i.e., show that

X
g = 0
for all X X(M) (cf. Exercise 2.6.3).
(8) Let (M, g) be a Riemannian manifold with Levi-Civita connection
, and let
t
: M M be a 1-parameter group of isometries. The
vector eld X X(M) dened by
X
p
:=
d
dt [
t=0

t
(p)
is called the Killing vector eld associated to
t
. Show that:
(a) L
X
g = 0 (cf. Exercise 2.8.3);
(b) X satises
Y
X, Z +
Z
X, Y = 0 for all vector elds
Y, Z X(M);
(c) if c : I M is a geodesic then

c(t), X
c(t)
_
is constant.
(9) Recall that if M is an oriented dierential manifold with volume
element
n
(M), the divergence of X is the function div(X)
such that
L
X
= (div(X))
(cf. Exercise 6.4.5 in Chapter 2). Suppose that M has a Riemannian
metric and that is a Riemannian volume element.
4. MINIMIZING PROPERTIES OF GEODESICS 107
(a) Show that this denition of divergence coincides with the def-
inition in Exercise 1.10.7.
(b) Show that at each point p M,
div(X) =
n

i=1

Y
i
X, Y
i
,
where Y
1
, . . . , Y
n
is an orthonormal basis of T
p
M and is
the Levi-Civita connection.
(10) Let M be an oriented Riemannian manifold of dimension 3. The
curl of a vector eld X X(M) is the vector eld curl(X) as-
sociated to the 1-form d
X
, where
X

1
(M) is the 1-form
associated to X (cf. Exercise 1.10.7). Show that:
(a) curl(grad(f)) = 0 for f C

(M, R);
(b) div(curl(X)) = 0 for X X(M);
(c) curl(X) =

3
i,j,k=1

ijk

Y
j
X, Y
k
_
Y
i
, where Y
1
, Y
2
, Y
3
is a
positive basis of orthonormal vector elds and
ijk
was dened
on Exercise 3.3.2.
4. Minimizing Properties of Geodesics
Let M be a dierentiable manifold with an ane connection . As we
saw in Section 2, given a point p M and a tangent vector v T
p
M, there
exists a unique geodesic c
v
: I M, dened on a maximal open interval
I R, such that 0 I, c
v
(0) = p and c
v
(0) = v. Consider now the curve
: J M dened by (t) = c
v
(at), where a R and J is the inverse image
of I by the map t at. We have
(t) = a c
v
(at),
and consequently

=
a cv
(a c
v
) = a
2

cv
c
v
= 0.
Thus is also a geodesic. Since (0) = c
v
(0) = p and (0) = a c
v
(0) = av,
we see that is the unique geodesic with initial velocity av T
p
M (that
is, = c
av
). Therefore, we have c
av
(t) = c
v
(at) for all t I. This property
is sometimes referred to as the homogeneity of geodesics. Notice that we
can make the interval J arbitrarily large by making a suciently small. If
1 I, we dene exp
p
(v) = c
v
(1). By homogeneity of geodesics, we can
dene exp
p
(v) for v in some open neighborhood U of the origin in T
p
M.
The map exp
p
: U T
p
M M thus obtained is called the exponential
map at p.
Proposition 4.1. There exists an open set U T
p
M containing the
origin such that exp
p
: U M is a dieomorphism onto some open set
V M containing p (called a normal neighborhood).
108 3. RIEMANNIAN MANIFOLDS
M
T
p
M
v
p
exp
p
(v)
Figure 2. The exponential map.
Proof. The exponential map is clearly dierentiable as a consequence
of the smooth dependence of the solution of an ODE on its initial data
(cf. [Arn92]). If v T
p
M is such that exp
p
(v) is dened, we have, by
homogeneity, that exp
p
(tv) = c
tv
(1) = c
v
(t). Consequently,
_
d exp
p
_
0
v =
d
dt
exp
p
(tv)
[
t=0
=
d
dt
c
v
(t)
[
t=0
= v.
We conclude that
_
d exp
p
_
0
: T
0
(T
p
M)

= T
p
M T
p
M is the identity map.
By the Inverse Function Theorem, exp
p
is then a dieomorphism of some
open neighbourhood U of 0 T
p
M onto some open set V M containing
p = exp
p
(0).
Example 4.2. Consider the Levi-Civita connection in S
2
with the stan-
dard metric, and let p S
2
. Then exp
p
(v) is well dened for all v T
p
S
2
,
but it is not a dieomorphism, as it is clearly not injective. However, its
restriction to the open ball B

(0) T
p
S
2
is a dieomorphism onto S
2
p.
Now let (M, , ) be a Riemannian manifold and its Levi-Civita con-
nection. Since , denes an inner product in T
p
M, we can think of T
p
M
as the Euclidean n-space R
n
. Let E be the vector eld dened on T
p
M0
by
E
v
=
v
|v|
,
4. MINIMIZING PROPERTIES OF GEODESICS 109
and dene X := (exp
p
)

E on V p, where V M is a normal neighbor-

hood. We have
X
exp
p
(v)
=
_
d exp
p
_
v
E
v
=
d
dt
exp
p
_
v +t
v
|v|
_
[
t=0
=
d
dt
c
v
_
1 +
t
|v|
_
[
t=0
=
1
|v|
c
v
(1).
Since | c
v
(1)| = | c
v
(0)| = |v|, we see that X
exp
p
(v)
is the unit tangent
vector to the geodesic c
v
. In particular, X must satisfy

X
X = 0.
For > 0 such that B

(0) U := exp
1
p
(V ), we dene the normal ball
with center p and radius as the open set B

(p) := exp
p
(B

(0)), and
the normal sphere of radius centered at p as the compact submani-
fold S

(p) := exp
p
(B

(0)). We will now prove that X is (and hence the

geodesics through p are) orthogonal to normal spheres. For that, we choose
a local parameterization : W R
n1
S
n1
T
p
M, and use it to dene
a parameterization : (0, +) W T
p
M through
(r,
1
, . . . ,
n1
) = r(
1
, . . . ,
n1
)
(hence (r,
1
, . . . ,
n1
) are spherical coordinates on T
p
M). Notice that

r
= E,
since
E
(r,)
= E
r()
= () =

r
(r, ),
and so
(10) X = (exp
p
)

r
.
Since

i
is tangent to r = , the vector elds
(11) Y
i
:= (exp
p
)

i
are tangent to S

(p). Notice also that

_
_

i
_
_
=
_
_
_

i
_
_
_ = r
_
_
_

i
_
_
_ is propor-
tional to r, and consequently

i
0 as r 0, implying that (Y
i
)
q
0
p
as q p. Since exp
p
is a local dieomorphism, the vector elds X and Y
i
are linearly independent at each point. Also,
[X, Y
i
] =
_
(exp
p
)

r
, (exp
p
)

i
_
= (exp
p
)

_

r
,

i
_
= 0
(cf. Exercise 6.11.8 in Chapter 1), or, since the Levi-Civita connection is
symmetric,

X
Y
i
=
Y
i
X.
110 3. RIEMANNIAN MANIFOLDS
To prove that X is orthogonal to the normal spheres S

(p), we show that

X is orthogonal to each of the vector elds Y
i
. In fact, since
X
X = 0 and
|X| = 1, we have
X X, Y
i
=
X
X, Y
i
+X,
X
Y
i
= X,
Y
i
X =
1
2
Y
i
X, X = 0,
and hence X, Y
i
is constant along each geodesic through p. Consequently,
X, Y
i
(exp
p
v) =
_
X
exp
p
(v)
, (Y
i
)
exp
p
(v)
_
= lim
t0
_
X
exp
p
(tv)
, (Y
i
)
exp
p
(tv)
_
= 0
(as |X| = 1 and (Y
i
)
q
0
p
as q p), and so every geodesic through p is
orthogononal to all normal spheres centered at p. Using this we obtain the
following result.
Proposition 4.3. Let : [0, 1] M be a dierentiable curve such that
(0) = p and (1) S

(p), where S

(p) is a normal sphere. Then l() ,

and l() = if and only if is a reparameterized geodesic.
Proof. We can assume that (t) ,= p for all t (0, 1), since otherwise
we could easily construct a curve : [0, 1] M with (0) = p, (1) =
(1) S

(p) and l( ) < l(). For the same reason, we can assume that
([0, 1)) B

(p). We can then write

(t) := exp
p
(r(t)n(t)),
where r(t) (0, ] and n(t) S
n1
are well dened for t (0, 1]. Note that
r(t) can be extended to [0, 1] as a smooth function. Then
(t) = (exp
p
)

( r(t)n(t) +r(t) n(t)) .

Since n(t), n(t) = 1, we have n(t), n(t) = 0, and consequently n(t) is
tangent to B
r(t)
(0). Noticing that n(t) =
_

r
_
r(t)n(t)
, we conclude that
(t) = r(t)X
(t)
+Y (t),
where X = (exp
p
)

r
and Y (t) = r(t)(exp
p
)

n(t) is tangent to S
r(t)
(p), and
hence orthogonal to X
(t)
. Consequently,
l() =
_
1
0

r(t)X
(t)
+Y (t), r(t)X
(t)
+Y (t)
_1
2
dt
=
_
1
0
_
r(t)
2
+|Y (t)|
2
_1
2
dt

_
1
0
r(t)dt = r(1) r(0) = .
It should be clear that l() = if and only if |Y (t)| 0 and r(t) 0
for all t [0, 1]. In this case, n(t) = 0 (implying that n is constant), and
(t) = exp
p
(r(t)n) = c
r(t)n
(1) = c
n
(r(t)) is, up to reparameterization, the
geodesic through p with initial condition n T
p
M.
4. MINIMIZING PROPERTIES OF GEODESICS 111
Definition 4.4. A piecewise dierentiable curve is a continuous
map c : [a, b] M such that the restriction of c to [t
i1
, t
i
] is dierentiable
for i = 1, . . . , n, where a = t
0
< t
1
< < t
n1
< t
n
= b. We say that c
connects p M to q M if c(a) = p and c(b) = q.
The denition of length of a piecewise dierentiable curve oers no
diculties. It should also be clear that Proposition 4.3 easily extends to
piecewise dierentiable curves, if we now allow for piecewise dierentiable
reparameterizations. Using this extended version of Proposition 4.3 as well
as the properties of the exponential map and the invariance of length under
reparameterization, one easily shows the following result.
Theorem 4.5. Let (M, , ) be a Riemannian manifold, p M and
B

(p) a normal ball centered at p. Then, for each point q B

(p), there
exists a geodesic c : I M connecting p to q. Moreover, if : J M is
any other piecewise dierentiable curve connecting p to q, then l() l(c),
and l() = l(c) if and only if is a reparameterization of c.
Conversely, we have
Theorem 4.6. Let (M, , ) be a Riemannian manifold and p, q M.
If c : I M is a piecewise dierentiable curve connecting p to q and
l(c) l() for any piecewise dierentiable curve : J M connecting p to
q then c is a reparameterized geodesic.
To prove this theorem, we need the following denition.
Definition 4.7. A normal neighborhood V M is called a totally
normal neighborhood if there exists > 0 such that V B

(p) for all

p V .
We will now prove that totally normal neighborhoods always exist. To do
so, we recall that local coordinates (x
1
, . . . , x
n
) on M yield local coordinates
(x
1
, . . . , x
n
, v
1
, . . . , v
n
) on TM labeling the vector
v
1

x
1
+. . . +v
n

x
n
.
The geodesic equations,
x
i
+
n

j,k=1

i
jk
x
j
x
k
= 0 (i = 1, . . . , n),
correspond to the system of rst order ODEs
_
x
i
= v
i
v
i
=

n
j,k=1

i
jk
v
j
v
k
(i = 1, . . . , n).
These equations dene the local ow of the vector eld X X(TM) given
in local coordinates by
X =
n

i=1
v
i

x
i

n

i,j,k=1

i
jk
v
j
v
k

v
i
,
112 3. RIEMANNIAN MANIFOLDS
called the geodesic ow. As it was seen in Chapter 1, for each point
v TM there exists an open neighborhood W TM and an open interval
I R containing 0 such that the local ow F : W I TM of X is
well dened. In particular, for each point p M we can choose an open
neighborhood U containing p and > 0 such that the geodesic ow is well
dened in W I with
W = v T
q
M [ q U, |v| < .
Using homogeneity of geodesics, we can make the interval I as large as we
want by making suciently small. Therefore, for small enough we can
dene a map G : W M M by G(v) := (q, exp
q
(v)). Since exp
q
(0) = q,
the matrix representation of (dG)
0
in the above local coordinates is
_
I 0
I I
_
,
and hence G is a local dieomorphism. Reducing U and if necessary, we
can therefore assume that G is a dieomorphism onto its image G(W), which
contains the point (p, p) = G(0
p
). Choosing an open neighborhood V of p
such that V V G(W), it is clear that V is a totally normal neighborhood:
for each point q V we have q exp
q
(B

(0)) = G(W) (q M)
q V , that is, exp
q
(B

(0)) V .
Notice that given any two points p, q in a totally normal neighborhood
V , there exists a geodesic c : I M connecting p to q such that any
other piecewise dierentiable curve : J M connecting p to q satises
l() l(c) (and l() = l(c) if and only if is a reparameterization of c).
The proof of Theorem 4.6 is now an immediate consequence of the following
observation: if c : I M is a piecewise dierentiable curve connecting p to
q such that l(c) l() for any curve : J M connecting p to q, then c
must be a reparameterized geodesic in each totally normal neighborhood it
intersects.
Exercises 4.8.
(1) Let (M, g) be a Riemannian manifold and f : M R a smooth
function. Show that if | grad(f)| 1 then the integral curves of
grad(f) are geodesics, using:
(a) the denition of geodesic;
(b) the minimizing properties of geodesics.
(2) Let M be a Riemannian manifold and the Levi-Civita connection
on M. Given p M and a basis v
1
, . . . , v
n
for T
p
M, we consider
the parameterization : U R
n
M of a normal neighborhood
given by
(x
1
, . . . , x
n
) = exp
p
(x
1
v
1
+. . . +x
n
v
n
)
(the local coordinates (x
1
, . . . , x
n
) are called normal coordinates).
Show that:
(a) in these coordinates,
i
jk
(p) = 0 (Hint: Consider the geodesic equa-
tion);
(b) if v
1
, . . . , v
n
is an orthonormal basis then g
ij
(p) =
ij
.
4. MINIMIZING PROPERTIES OF GEODESICS 113
(3) Let G be a Lie group endowed with a bi-invariant Riemannian
metric (i.e., such that L
g
and R
g
are isometries for all g G), and
let i : G G be the dieomorphism dened by i(g) = g
1
.
(a) Compute (di)
e
and show that
(di)
g
=
_
dR
g
1
_
e
(di)
e
_
dL
g
1
_
g
for all g G. Conclude that i is an isometry.
(b) Let v g = T
e
G and c
v
be the geodesic satisfying c
v
(0) =
e and c
v
(0) = v. Show that if t is suciently small then
c
v
(t) = (c
v
(t))
1
. Conclude that c
v
is dened in R and
satises c
v
(t + s) = c
v
(t)c
v
(s) for all t, s R. (Hint: Recall
that any two points in a totally normal neighborhood are connected by a unique
geodesic in that neighbourhood).
(c) Show that the geodesics of G are the integral curves of left-
invariant vector elds, and that the maps exp (the Lie group
exponential) and exp
e
(the geodesic exponential at the iden-
tity) coincide.
(d) Let be the Levi-Civita connection of the bi-invariant metric
and X, Y two left-invariant vector elds. Show that

X
Y =
1
2
[X, Y ].
(4) Use Theorem 4.6 to prove that if f : M N is an isometry and
c : I M is a geodesic then f c : I N is also a geodesic.
(5) Let f : M M be an isometry whose set of xed points is a
connected 1-dimensional submanifold N M. Show that N is the
image of a geodesic.
(6) Let (M, , ) be a Riemannian manifold whose geodesics can be
extended for all values of their parameters, and let p M.
(a) Let X and Y
i
be the vector elds dened on a normal ball
centered at p as in (10) and (11). Show that Y
i
satises the
Jacobi equation

X
Y
i
= R(X, Y
i
)X,
where R : X(M) X(M) X(M) X(M), dened by
R(X, Y )Z =
X

Y
Z
Y

X
Z
[X,Y ]
Z,
is called the curvature operator (cf. Chapter 4). (Remark:
It can be shown that (R(X, Y )Z)p depends only on Xp, Yp, Zp).
(b) Consider a geodesic c : R M parameterized by the arclength
such that c(0) = p. A vector eld Y along c is called a Jacobi
eld if it satises the Jacobi equation along c,
D
2
Y
dt
2
= R( c, Y ) c.
114 3. RIEMANNIAN MANIFOLDS
Show that Y is a Jacobi eld with Y (0) = 0 if and only if
Y (t) =

s
exp
p
(tv(s))
[
s=0
with v : (, ) T
p
M satisfying v(0) = c(0).
(c) A point q M is said to be conjugate to p if it is a critical
value of exp
p
. Show that q is conjugate to p if and only if there
exists a nonzero Jacobi eld Y along a geodesic c connecting
p = c(0) to q = c(b) such that Y (0) = Y (b) = 0. Conclude
that if q is conjugate to p then p is conjugate to q.
(d) The manifold M is said to have nonpositive curvature if
R(X, Y )X, Y 0 for all X, Y X(M). Show that for such
a manifold no two points are conjugate.
(e) Given a geodesic c : I M parameterized by the arclength
such that c(0) = p, let t
c
be the supremum of the set of values
of t such that c is the minimizing curve connecting p to c(t)
(hence t
c
> 0). The cut locus of p is dened to be the set of
all points of the form c(t
c
) for t
c
< +. Determine the cut
locus of a given point p M when M is:
(i) the torus T
n
with the standard metric;
(ii) the sphere S
n
with the standard metric;
(iii) the projective space RP
n
with the standard metric.
Check in these examples that any point in the cut locus is
either conjugate to p or joined to p by two geodesics with the
same length but dierent images.
5. Hopf-Rinow Theorem
Let (M, g) be a Riemannian manifold. The existence of totally normal
neighborhoods implies that it is always possible to connect two suciently
close points p, q M by a minimizing geodesic. We now address the same
question globally.
Example 5.1.
(1) Given two distinct points p, q R
n
there exists a unique (up to
reparameterization) geodesic for the Euclidean metric connecting
them.
(2) Given two distinct points p, q S
n
there exist at least two geodesics
for the standard metric connecting them which are not reparame-
terizations of each other.
(3) If p ,= 0 then there exists no geodesic for the Euclidean metric in
R
n
0 connecting p to p.
In many cases (for example in R
n
0) there exist geodesics which
cannot be extended for all values of its parameter. In other words, exp
p
(v)
is not dened for all v T
p
M.
5. HOPF-RINOW THEOREM 115
Definition 5.2. A Riemannian manifold (M, , ) is said to be geodesi-
cally complete if, for every point p M, the map exp
p
is dened in T
p
M.
There exists another notion of completeness of a connected Riemannian
manifold, coming from the fact that any such manifold is naturally a metric
space.
Definition 5.3. Let (M, , ) be a connected Riemannian manifold and
p, q M. The distance between p and q is dened as
d(p, q) = infl() [ is a piecewise dierentiable curve connecting p to q.
Notice that if there exists a minimizing geodesic c connecting p to q then
d(p, q) = l(c). The function d : M M [0, +) is indeed a distance, as
stated in the following proposition (whose proof is left as an exercise).
Proposition 5.4. (M, d) is a metric space, that is, d satises:
(i) Positivity: d(p, q) 0 and d(p, q) = 0 if and only if p = q;
(ii) Symmetry: d(p, q) = d(q, p);
(iii) Triangle inequality: d(p, r) d(p, q) +d(q, r),
for all p, q, r M. The metric space topology induced on M coincides with
the topology of M as a dierentiable manifold.
Therefore, we can discuss the completeness of M as a metric space (that
is, whether Cauchy sequences converge). The fact that completeness and
geodesic completeness are equivalent is the content of the following theorem.
Theorem 5.5. (Hopf-Rinow) Let (M, , ) be a connected Riemannian
manifold and p M. The following assertions are equivalent:
(i) M is geodesically complete;
(ii) (M, d) is a complete metric space;
(iii) exp
p
is dened in T
p
M.
Moreover, if (M, , ) is geodesically complete then for all q M there
exists a geodesic c connecting p to q with l(c) = d(p, q).
Proof. It is clear that (i) (iii).
We begin by showing that if (iii) holds then for all q M there exists
a geodesic c connecting p to q with l(c) = d(p, q). Let d(p, q) = . If = 0
then q = p and there is nothing to prove. If > 0, let (0, ) be such
that S

(p) is a normal sphere (which is a compact submanifold of M). The

continuous function x d(x, q) will then have a minimum point x
0
S

(p).
Moreover, x
0
= exp
p
(v), where |v| = 1. Let us consider the geodesic
c
v
(t) = exp
p
(tv). We will show that q = c
v
(). For that, we consider the set
A = t [0, ] [ d(c
v
(t), q) = t.
Since the map t d(c
v
(t), q) is continuous, A is a closed set. Moreover,
A ,= , as clearly 0 A. We will now show that no point t
0
[0, ) can be
the maximum of A, which implies that the maximum of A must be (hence
d(c
v
(), q) = 0, that is, c
v
() = q, and so c
v
connects p to q and l(c
v
) = ).
116 3. RIEMANNIAN MANIFOLDS
p
r
q
x
0
y
0
Figure 3. Proof of the Hopf-Rinow Theorem.
Let t
0
A[0, ), r = c
v
(t
0
) and (0, t
0
) such that S

(r) is a normal
sphere. Let y
0
be a minimum point of the continuous function y d(y, q)
on the compact set S

(r). We will show that y

0
= c
v
(t
0
+ ). In fact, we
have
t
0
= d(r, q) = + min
yS

(r)
d(y, q) = +d(y
0
, q),
and so
(12) d(y
0
, q) = t
0
.
The triangle inequality then implies that
d(p, y
0
) d(p, q) d(y
0
, q) = ( t
0
) = t
0
+,
and, since the piecewise dierentiable curve which connects p to r through
c
v
and r to y
0
through a geodesic has length t
0
+, we conclude that this is
a minimizing curve, hence a (reparameterized) geodesic. Thus, as promised,
y
0
= c
v
(t
0
+). Consequently, equation (12) can be written as
d(c
v
(t
0
+), q) = (t
0
+),
implying that t
0
+ A, and so t
0
cannot be the maximum of A.
We can now prove that (iii) (ii). To do so, we begin by showing that
any bounded closed subset K M is compact. Indeed, if K is bounded
then K B
R
(p) for some R > 0, where
B
R
(p) = q M [ d(p, q) < R.
As we have seen, p can be connected to any point in B
R
(p) by a geodesic
of length smaller than R, and so B
R
(p) exp
p
_
B
R
(0)
_
. Since exp
p
:
T
p
M M is continuous and B
R
(0) is compact, the set exp
p
_
B
R
(0)
_
is also
compact. Therefore K is a closed subset of a compact set, hence compact.
6. NOTES ON CHAPTER 3 117
Now, if p
n
is a Cauchy sequence in M, then its closure is bounded, hence
compact. Thus p
n
must have a convergent subsequence, and therefore
must itself converge.
Finally, we show that (ii) (i). Let c be a geodesic dened for t < t
0
,
which we can assume without loss of generality to be normalized, that is,
| c(t)| = 1. Let t
n
be an increasing sequence of real numbers converging to
t
0
. Since d(c(t
m
), c(t
n
)) [t
m
t
n
[, we see that c(t
n
) is a Cauchy sequence.
As we are assuming M to be complete, we conclude that c(t
n
) p M,
and it is easily seen that c(t) p as t t
0
. Let B

(p) be a normal ball

centered at p. Then c can be extended past t
0
in this normal ball.
Corollary 5.6. If M is compact then M is geodesically complete.
Proof. Any compact metric space is complete.
Corollary 5.7. If M is a closed connected submanifold of a complete
connected Riemannian manifold with the induced metric then M is complete.
Proof. Let M be a closed connected submanifold of a complete con-
nected Riemannian manifold N. Let d be the distance determined by the
metric on N, and let d

be the distance determined by the induced metric

on M. Then d d

, as any curve on M is also a curve on N. Let p

n
be
a Cauchy sequence on (M, d

). Then p
n
is a Cauchy sequence on (N, d),
and consequently converges in N to a point p M (as N is complete and
M is closed). Since the topology of M is induced by the topology of N, we
conclude that p
n
p on M.
Exercises 5.8.
(1) Prove Proposition 5.4.
(2) Consider R
2
(x, 0) [ 3 x 3 with the Euclidean metric.
Determine B
7
(0, 4).
(3) (a) Prove that a connected Riemannian manifold is complete if
and only if the compact sets are the closed bounded sets.
(b) Give an example of a connected Riemannian manifold contain-
ing a noncompact closed bounded set.
(4) A Riemannian manifold (M, , ) is said to be homogeneous if,
given any two points p, q M, there exists an isometry f : M
M such that f(p) = q. Show that any homogeneous Riemannian
manifold is complete.
6. Notes on Chapter 3
6.1. Section 5. In this Section we use several denitions and results
about metric spaces, which we now discuss. A metric space is a pair
(M, d), where M is a set and d : M M [0, +) is a map satisfying the
properties enumerated in Proposition 5.4. The set
B

(p) = q M [ d(p, q) <

118 3. RIEMANNIAN MANIFOLDS
is called the open ball with center p and radius . The family of all such
balls is a basis for a Hausdor topology on M, called the metric topology.
Notice that in this topology p
n
p if and only if d(p
n
, p) 0. Although a
metric space (M, d) is not necessarily second countable, it is still true that
F M is closed if and only if every convergent sequence in F has limit in
F, and K M is compact if and only if every sequence in K has a sublimit
in K.
A sequence p
n
in M is said to be a Cauchy sequence if for all > 0
there exists N N such that d(p
n
, p
m
) < for all m, n > N. It is easily
seen that all convergent sequences are Cauchy sequences. The converse,
however, is not necessarily true (but if a Cauchy sequence has a convergent
subsequence then it must converge). A metric space is said to be complete
if all its Cauchy sequences converge. A closed subset of a complete metric
space is itself complete.
A set is said to be bounded if it is a subset of some ball. For instance,
the set of all terms of a Cauchy sequence is bounded. It is easily shown that
if K M is compact then K must be bounded and closed (but the converse
is not necessarily true). A compact metric space is necessarily complete.
6.2. Bibliographical notes. The material in this chapter can be found
in most books on Riemannian geometry (e.g. [Boo03, dC93, GHL04]).
For more details on general ane connections, see [KN96]. Bi-invariant
metrics on a Lie group are examples of symmetric spaces, whose beautiful
theory is studied for instance in [Hel01].
CHAPTER 4
Curvature
This chapter addresses the fundamental notion of curvature of a Rie-
mannian manifold.
Section 1 introduces the curvature operator of a general ane con-
nection, and, for Riemannian manifolds, the equivalent (more geometric)
notion of sectional curvature. The Ricci curvature tensor and the
scalar curvature, which will be important in Chapter 6, are also dened.
Section 2 establishes the Cartan structure equations, a powerful
computational method which employs dierential forms to calculate the cur-
vature. These equations are used in Section 3 to prove the Gauss-Bonnet
Theorem, relating the curvature of a compact surface to its topology.
Complete Riemannian manifolds with constant curvature are dis-
cussed in Section 4. These provide important examples of curved geometries.
Finally, the relation between the curvature of a Riemannian manifold
and the curvature of a submanifold (with the induced metric) is studied in
Section 5.
1. Curvature
As we saw in Exercise 3.3.4 of Chapter 3, no open set of the 2-sphere
S
2
with the standard metric is isometric to an open set of the Euclidean
plane. The geometric object that locally distinguishes these two Riemannian
manifolds is the so-called curvature operator, which appears in many
other situations (cf. Exercise 4.8.6 in Chapter 3).
Definition 1.1. The curvature R of a connection is a correspon-
dence that to each pair of vector elds X, Y X(M) associates the map
R(X, Y ) : X(M) X(M) dened by
R(X, Y )Z =
X

Y
Z
Y

X
Z
[X,Y ]
Z.
Hence, R is a way of measuring the non-commutativity of the connection.
We leave it as an exercise to show that this denes a (3, 1)-tensor (called
the Riemann tensor), meaning that
(i) R(fX
1
+gX
2
, Y )Z = fR(X
1
, Y )Z +gR(X
2
, Y )Z,
(ii) R(X, fY
1
+gY
2
)Z = fR(X, Y
1
)Z +gR(X, Y
2
)Z,
(iii) R(X, Y )(fZ
1
+gZ
2
) = fR(X, Y )Z
1
+gR(X, Y )Z
2
,
for all vector elds X, X
1
, X
2
, Y, Y
1
, Y
2
, Z, Z
1
, Z
2
X(M) and all smooth
functions f, g C

(M). Choosing a coordinate system x : V R

n
on M,
119
120 4. CURVATURE
this tensor can be locally written as
R =
n

i,j,k,l=1
R
l
ijk
dx
i
dx
j
dx
k

x
l
,
where each coecient R
l
ijk
is the l-coordinate of the vector eld R(

x
i
,

x
j
)

x
k
,
that is,
R
_

x
i
,

x
j
_

x
k
=
n

l=1
R
l
ijk

x
l
.
Using the fact that [

x
i
,

x
j
] = 0, we have
R
_

x
i
,

x
j
_

x
k
=
x
i

x
j

x
k

x
j

x
i

x
k
=
x
i
_
n

m=1

m
jk

x
m
_

x
j
_
n

m=1

m
ik

x
m
_
=
n

m=1
_

x
i

m
jk

x
j

m
ik
_

x
m
+
n

l,m=1
(
m
jk

l
im

m
ik

l
jm
)

x
l
=
n

l=1
_

l
jk
x
i


l
ik
x
j
+
n

m=1

m
jk

l
im

m=1

m
ik

l
jm
_

x
l
,
and so
R
l
ijk
=

l
jk
x
i


l
ik
x
j
+
n

m=1

m
jk

l
im

m=1

m
ik

l
jm
.
Example 1.2. Consider M = R
n
with the Euclidean metric and the cor-
responding Levi-Civita connection (that is, with Christoel symbols
k
ij

0). Then R
l
ijk
= 0, and the curvature R is zero. Thus, we can also interpret
the curvature as a measure of how much a connection on a given manifold
diers from the Levi-Civita connection of the Euclidean space.
When the connection is symmetric (as in the case of the Levi-Civita
connection), the tensor R satises the so-called Bianchi Identity.
Proposition 1.3. (Bianchi Identity) If M is a manifold with a sym-
metric connection then the associated curvature satises
R(X, Y )Z +R(Y, Z)X +R(Z, X)Y = 0.
Proof. This property is a direct consequence of the Jacobi identity of
vector elds. Indeed,
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
=
X
(
Y
Z
Z
Y ) +
Y
(
Z
X
X
Z) +
Z
(
X
Y
Y
X)

[X,Y ]
Z
[Y,Z]
X
[Z,X]
Y,
1. CURVATURE 121
and so, since the connection is symmetric, we have
R(X, Y )Z +R(Y, Z)X +R(Z, X)Y
=
X
[Y, Z] +
Y
[Z, X] +
Z
[X, Y ]
[Y,Z]
X
[Z,X]
Y
[X,Y ]
Z
= [X, [Y, Z]] + [Y, [Z, X]] + [Z, [X, Y ]] = 0.

We will assume from this point on that (M, g) is a Riemannian manifold

and its Levi-Civita connection. We can dene a new covariant 4-tensor,
known as the curvature tensor:
R(X, Y, Z, W) := g(R(X, Y )Z, W).
Notice that because the metric is nondegenerate the curvature tensor con-
tains the same information as the Riemann tensor. Again, choosing a coor-
dinate system x : V R
n
on M, we can write this tensor as
R(X, Y, Z, W) =
_
_
n

i,j,k,l=1
R
ijkl
dx
i
dx
j
dx
k
dx
l
_
_
(X, Y, Z, W)
where
R
ijkl
= g
_
R
_

x
i
,

x
j
_

x
k
,

x
l
_
= g
_
n

m=1
R
m
ijk

x
m
,

x
l
_
=
n

m=1
R
m
ijk
g
ml
.
This tensor satises the following symmetry properties.
Proposition 1.4. If X, Y, Z, W are vector elds in M and is the
Levi-Civita connection, then
(i) R(X, Y, Z, W) +R(Y, Z, X, W) +R(Z, X, Y, W) = 0;
(ii) R(X, Y, Z, W) = R(Y, X, Z, W);
(iii) R(X, Y, Z, W) = R(X, Y, W, Z);
(iv) R(X, Y, Z, W) = R(Z, W, X, Y ).
Proof. Property (i) is an immediate consequence of the Bianchi iden-
tity, and property (ii) holds trivially.
Property (iii) is equivalent to showing that R(X, Y, Z, Z) = 0. Indeed,
if (iii) holds then clearly R(X, Y, Z, Z) = 0. Conversely, if this is true, we
have
R(X, Y, Z +W, Z +W) = 0 R(X, Y, Z, W) +R(X, Y, W, Z) = 0.
Now, using the fact that the Levi-Civita connection is compatible with the
metric, we have
X
Y
Z, Z =
X

Y
Z, Z +
Y
Z,
X
Z
and
[X, Y ] Z, Z = 2
[X,Y ]
Z, Z.
122 4. CURVATURE
Hence,
R(X, Y, Z, Z) =
X

Y
Z, Z
Y

X
Z, Z
[X,Y ]
Z, Z
= X
Y
Z, Z
Y
Z,
X
Z Y
X
Z, Z
+
X
Z,
Y
Z
1
2
[X, Y ] Z, Z
=
1
2
X (Y Z, Z)
1
2
Y (X Z, Z)
1
2
[X, Y ] Z, Z
=
1
2
[X, Y ] Z, Z
1
2
[X, Y ] Z, Z = 0.
To show (iv), we use (i) to get
R(X, Y, Z, W) + R(Y, Z, X, W) + R(Z, X, Y, W) = 0
R(Y, Z, W, X) + R(Z, W, Y, X) + R(W, Y, Z, X) = 0
R(Z, W, X, Y ) + R(W, X, Z, Y ) + R(X, Z, W, Y ) = 0
R(W, X, Y, Z) + R(X, Y, W, Z) + R(Y, W, X, Z) = 0
and so, adding these and using (iii), we have
R(Z, X, Y, W) +R(W, Y, Z, X) +R(X, Z, W, Y ) +R(Y, W, X, Z) = 0.
Using (ii) and (iii), we obtain
2R(Z, X, Y, W) 2R(Y, W, Z, X) = 0.

An equivalent way of encoding the information about the curvature of

a Riemannian manifold is by considering the following denition.
Definition 1.5. Let be a 2-dimensional subspace of T
p
M and let
X
p
, Y
p
be two linearly independent elements of . Then, the sectional
curvature of is dened as
K() :=
R(X
p
, Y
p
, X
p
, Y
p
)
|X
p
|
2
|Y
p
|
2
X
p
, Y
p

2
.
Note that |X
p
|
2
|Y
p
|
2
X
p
, Y
p

2
is the square of the area of the paral-
lelogram in T
p
M spanned by X
p
, Y
p
, and so the above denition of sectional
curvature does not depend on the choice of the linearly independent vec-
tors X
p
, Y
p
. Indeed, when we change of basis on , both R(X
p
, Y
p
, X
p
, Y
p
)
and |X
p
|
2
|Y
p
|
2
X
p
, Y
p

2
change by the square of the determinant of the
change of basis matrix (cf. Exercise 1.12.3.). We will now see that knowing
the sectional curvature of every section of T
p
M completely determines the
curvature tensor on this space.
Proposition 1.6. The Riemannian curvature tensor at p is uniquely
determined by the values of the sectional curvatures of sections (that is, 2-
dimensional subspaces) of T
p
M.
1. CURVATURE 123
Proof. Let us consider two covariant 4-tensors R
1
, R
2
on T
p
M sat-
isfying the symmetry properties of Proposition 1.4. Then the tensor T :=
R
1
R
2
also satises these symmetry properties. We will see that, if the val-
ues R
1
(X
p
, Y
p
, X
p
, Y
p
) and R
2
(X
p
, Y
p
, X
p
, Y
p
) agree for every X
p
, Y
p
T
p
M
(that is, if T(X
p
, Y
p
, X
p
, Y
p
) = 0 for every X
p
, Y
p
T
p
M), then R
1
= R
2
(that is, T 0). Indeed, for all vectors X
p
, Y
p
, Z
p
T
p
M, we have
0 = T(X
p
+Z
p
, Y
p
, X
p
+Z
p
, Y
p
) = T(X
p
, Y
p
, Z
p
, Y
p
) +T(Z
p
, Y
p
, X
p
, Y
p
)
= 2T(X
p
, Y
p
, Z
p
, Y
p
),
and so
0 = T(X
p
, Y
p
+W
p
, Z
p
, Y
p
+W
p
) = T(X
p
, Y
p
, Z
p
, W
p
) +T(X
p
, W
p
, Z
p
, Y
p
)
= T(Z
p
, W
p
, X
p
, Y
p
) T(W
p
, X
p
, Z
p
, Y
p
),
that is, T(Z
p
, W
p
, X
p
, Y
p
) = T(W
p
, X
p
, Z
p
, Y
p
). Hence T is invariant by
cyclic permutations of the rst three elements and so, by the Bianchi Iden-
tity, we have 3T(X
p
, Y
p
, Z
p
, W
p
) = 0.
A Riemannian manifold is called isotropic at a point p M if its
sectional curvature is a constant K
p
for every section T
p
M. Moreover,
it is called isotropic if it is isotropic at all points. Note that every 2-
dimensional manifold is trivially isotropic. Its sectional curvature K(p) :=
K
p
is called the Gauss curvature.
Remark 1.7. As we will see later, the Gauss curvature measures how
much the local geometry of the surface diers from the geometry of the
Euclidean plane. For instance, its integral over a disk D on the surface gives
the angle by which a vector is rotated when parallel-transported around
the boundary of D (cf. Exercise 2.8.7). Alternatively, its integral over the
interior of a geodesic triangle is equal to the dierence between the sum
of the inner angles of and (cf. Exercise 3.6.6). We will also see that
the sectional curvature of an n-dimensional Riemannian manifold is actually
the Gauss curvature of special 2-dimensional submanifolds, formed by the
geodesics tangent to the sections (cf. Exercise 5.7.5).
Proposition 1.8. If M is isotropic at p and x : V R
n
is a coordinate
system around p, then the coecients of the Riemannian curvature tensor
at p are given by
R
ijkl
(p) = K
p
(g
ik
g
jl
g
il
g
jk
).
Proof. We rst dene a covariant 4-tensor A on T
p
M as
A :=
n

i,j,k,l=1
K
p
( g
ik
g
jl
g
il
g
jk
) dx
i
dx
j
dx
k
dx
l
.
124 4. CURVATURE
We leave it as an exercise to check that A satises the symmetry properties
of Proposition 1.4. Moreover,
A(X
p
, Y
p
, X
p
, Y
p
) =
n

i,j,k,l=1
K
p
( g
ik
g
jl
g
il
g
jk
) X
i
p
Y
j
p
X
k
p
Y
l
p
= K
p
_
X
p
, X
p
Y
p
, Y
p
X
p
, Y
p

2
_
= R(X
p
, Y
p
, X
p
, Y
p
),
and so we conclude from Proposition 1.6 that A = R.
Definition 1.9. A Riemannian manifold is called a manifold of con-
stant curvature if it is isotropic and K
p
is the same at all points of M.
Example 1.10. The Euclidean space is a manifold of constant curvature
K
p
0. We will see the complete classication of (complete, connected)
manifolds of constant curvature in Section 4.
Another geometric object, very important in General Relativity, is the
so-called Ricci tensor.
Definition 1.11. The Ricci curvature tensor is the covariant 2-
tensor locally dened as
Ric(X, Y ) :=
n

k=1
dx
k
_
R
_

x
k
, X
_
Y
_
.
The above denition is independent of the choice of coordinates. In-
deed, we can see Ric
p
(X
p
, Y
p
) as the trace of the linear map from T
p
M to
T
p
M given by Z
p
R(Z
p
, X
p
)Y
p
, hence independent of the choice of basis.
Moreover, this tensor is symmetric. In fact, choosing an orthonormal basis
E
1
. . . , E
n
of T
p
M we have
Ric
p
(X
p
, Y
p
) =
n

k=1
R(E
k
, X
p
)Y
p
, E
k
=
n

k=1
R(E
k
, X
p
, Y
p
, E
k
)
=
n

k=1
R(Y
p
, E
k
, E
k
, X
p
) =
n

k=1
R(E
k
, Y
p
, X
p
, E
k
) = Ric
p
(Y
p
, X
p
).
Locally, we can write
Ric =
n

i,j=1
R
ij
dx
i
dx
j
where the coecients R
ij
are given by
R
ij
:= Ric
_

x
i
,

x
j
_
=
n

k=1
dx
k
_
R
_

x
k
,

x
i
_

x
j
_
=
n

k=1
R
k
kij
,
that is, R
ij
=

n
k=1
R
k
kij
.
1. CURVATURE 125
Incidentally, note that we obtained a (2, 0)-tensor from a (3, 1)-tensor.
This is an example of a general procedure called contraction, where we
obtain a (k 1, m 1)-tensor from a (k, m)-tensor. To do so, we choose
two indices on the components of the (k, m)-tensor, one covariant and other
contravariant, set them equal and then sum over them, thus obtaining the
components of a (k 1, m1)-tensor. On the example of the Ricci tensor,
we took the (3, 1)-tensor

R dened by the curvature,

R(X, Y, Z, ) = (R(X, Y )Z),

chose the rst covariant index and the rst contravariant index, set them
equal and summed over them:
Ric(X, Y ) =
n

k=1

R
_

x
k
, X, Y, dx
k
_
.
Similarly, we can use contraction to obtain a function (0-tensor) from
the Ricci tensor (a covariant 2-tensor). For that, we rst need to dene a
new (1, 1)-tensor eld T using the metric,
T(X, ) := Ric(X, Y ),
where Y is such that (Z) = Y, Z for every vector eld Z. Then, we
set the covariant index equal to the contravariant one and add, obtaining
a function S : M R called the scalar curvature. Locally, choosing a
coordinate system x : V R
n
, we have
S(p) :=
n

k=1
T
_

x
k
, dx
k
_
=
n

k=1
Ric
_

x
k
, Y
k
_
,
where, for every vector eld Z on V ,
Z
k
= dx
k
(Z) = Z, Y
k
=
n

i,j=1
g
ij
Z
i
Y
j
k
.
Therefore, we must have Y
j
k
= g
jk
(where (g
ij
) = (g
ij
)
1
), and hence Y
k
=

n
i=1
g
ik
x
i
. We conclude that the scalar curvature is locally given by
S(p) =
n

k=1
Ric
_

x
k
,
n

i=1
g
ik

x
i
_
=
n

i,k=1
R
ki
g
ik
=
n

i,k=1
g
ik
R
ik
.
(since Ric is symmetric).
Exercises 1.12.
(1) (a) Show that the curvature operator satises
(i) R(fX
1
+gX
2
, Y )Z = fR(X
1
, Y )Z +gR(X
2
, Y )Z;
(ii) R(X, fY
1
+gY
2
)Z = fR(X, Y
1
)Z +gR(X, Y
2
)Z;
(iii) R(X, Y )(fZ
1
+gZ
2
) = fR(X, Y )Z
1
+gR(X, Y )Z
2
,
for all vector elds X, X
1
, X
2
, Y, Y
1
, Y
2
, Z, Z
1
, Z
2
X(M) and
smooth functions f, g C

(M).
126 4. CURVATURE
(b) Show that (R(X, Y )Z)
p
T
p
M depends only on X
p
, Y
p
, Z
p
.
Conclude that R denes a (3, 1)-tensor. (Hint: Choose local coor-
dinates around p M).
(2) Recall that if G is a Lie group endowed with a bi-invariant Rie-
mannian metric, is the Levi-Civita connection and X, Y are two
left-invariant vector elds then

X
Y =
1
2
[X, Y ]
(cf. Exercise 4.8.3 in Chapter 3). Show that if Z is also left-
invariant, then
R(X, Y )Z =
1
4
[Z, [X, Y ]].
(3) Show that |X
p
|
2
|Y
p
|
2
X
p
, Y
p

2
gives us the square of the area
of the parallelogram in T
p
M spanned by X
p
, Y
p
. Conclude that the
sectional curvature does not depend on the choice of the linearly
independent vectors X
p
, Y
p
, that is, when we change of basis on ,
both R(X
p
, Y
p
, X
p
, Y
p
) and |X
p
|
2
|Y
p
|
2
X
p
, Y
p

2
change by the
square of the determinant of the change of basis matrix.
(4) Show that Ric is the only independent contraction of the curvature
tensor: choosing any other two indices and contracting, one either
gets Ric or 0.
(5) Let M be a 3-dimensional manifold. Show that the curvature tensor
is entirely determined by the Ricci tensor.
(6) Let (M, g) be an n-dimensional isotropic Riemannian manifold with
sectional curvature K. Show that Ric = (n 1)Kg and S =
n(n 1)K.
(7) Let g
1
, g
2
be two Riemannian metrics on a manifold M such that
g
1
= g
2
, for some constant > 0. Show that:
(a) the corresponding sectional curvatures K
1
and K
2
satisfy K
1
() =

1
K
2
() for any 2-dimensional section of a tangent space of
M;
(b) the corresponding Ricci curvature tensors satisfy Ric
1
= Ric
2
;
(c) the corresponding scalar curvatures satisfy S
1
=
1
S
2
.
(8) If is not the Levi-Civita connection can we still dene the Ricci
curvature tensor Ric? Is it necessarily symmetric?
2. Cartan Structure Equations
In this section we will reformulate the properties of the Levi-Civita con-
nection and of the Riemannian curvature tensor in terms of dierential
forms. For that we will take an open subset V of M where we have de-
ned a eld of frames X
1
, . . . , X
n
, that is, a set of n vector elds that,
at each point p of V , form a basis for T
p
M (for example, we can take a
coordinate neighborhood V and the vector elds X
i
=

x
i
; however, in gen-
eral, the X
i
are not associated to a coordinate system). Then we consider
2. CARTAN STRUCTURE EQUATIONS 127
a eld of dual coframes, that is, 1-forms
1
, . . . ,
n
on V such that

i
(X
j
) =
ij
. Note that, at each point p V ,
1
p
, . . . ,
n
p
is a basis for
T

p
M. From the properties of a connection, in order to dene
X
Y we just
have to establish the values of

X
i
X
j
=
n

k=1

k
ij
X
k
,
where
k
ij
is dened as the k
th
component of the vector eld
X
i
X
j
on the
basis X
i

n
i=1
. Note that if the X
i
are not associated to a coordinate system
then the
k
ij
cannot be computed using formula (9), and, in general, they are
not even symmetric in the indices i, j (cf. Exercise 2.8.1). Given the values
of the
k
ij
on V , we can dene 1-forms
k
j
(j, k = 1, . . . , n) in the following
way:
(13)
k
j
:=
n

i=1

k
ij

i
.
Conversely, given these forms, we can obtain the values of
k
ij
through

k
ij
=
k
j
(X
i
).
The connection is then completely determined from these forms: given two
vector elds X =

n
i=1
a
i
X
i
and Y =

n
i=1
b
i
X
i
, we have

X
X
j
=

n
i=1
a
i
X
i
X
j
=
n

i=1
a
i

X
i
X
j
=
n

i,k=1
a
i

k
ij
X
k
(14)
=
n

i,k=1
a
i

k
j
(X
i
) X
k
=
n

k=1

k
j
(X) X
k
and hence

X
Y =
X
_
n

i=1
b
i
X
i
_
=
n

i=1
_
(X b
i
)X
i
+b
i

X
X
i
_
(15)
=
n

j=1
_
X b
j
+
n

i=1
b
i

j
i
(X)
_
X
j
.
Note that the values of the forms
k
j
at X are the components of
X
X
j
relative to the eld of frames, that is,
(16)
i
j
(X) =
i
(
X
X
j
) .
The
k
j
are called the connection forms. For the Levi-Civita connection,
these forms cannot be arbitrary. Indeed, they have to satisfy certain equa-
tions corresponding to the properties of symmetry and compatibility with
the metric.
128 4. CURVATURE
Theorem 2.1. (Cartan) Let V be an open subset of a Riemannian man-
ifold M on which we have dened a eld of frames X
1
, . . . , X
n
. Let

1
, . . . ,
n
be the corresponding eld of coframes. Then the connection
forms of the Levi-Civita connection are the unique solution of the equations
(i) d
i
=

n
j=1

j

i
j
,
(ii) dg
ij
=

n
k=1
(g
kj

k
i
+g
ki

k
j
),
where g
ij
= X
i
, X
j
.
Proof. We begin by showing that the Levi-Civita connection forms,
dened by (13), satisfy (i) and (ii). For this, we will use the following
property of 1-forms (cf. Exercise 3.8.2 of Chapter 2):
d(X, Y ) = X ((Y )) Y ((X)) ([X, Y ]).
We have

Y
X =
Y
_
_
n

j=1

j
(X)X
j
_
_
=
n

j=1
_
Y
j
(X) X
j
+
j
(X)
Y
X
j
_
,
which implies
(17)
i
(
Y
X) = Y
i
(X) +
n

j=1

j
(X)
i
(
Y
X
j
).
Using (16) and (17), we have
_
_
n

j=1

j

i
j
_
_
(X, Y ) =
n

j=1
_

j
(X)
i
j
(Y )
j
(Y )
i
j
(X)
_
=
n

j=1
_

j
(X)
i
(
Y
X
j
)
j
(Y )
i
(
X
X
j
)
_
=
i
(
Y
X) Y (
i
(X))
i
(
X
Y ) +X (
i
(Y )),
and so
_
_
d
i

j=1

j

i
j
_
_
(X, Y ) =
= X (
i
(Y )) Y (
i
(X))
i
([X, Y ])
_
_
n

j=1

j

i
j
_
_
(X, Y )
=
i
(
X
Y
Y
X [X, Y ]) = 0.
Note that equation (i) is equivalent to symmetry of the connection. To show
that (ii) holds, we notice that
dg
ij
(Y ) = Y X
i
, X
j
,
2. CARTAN STRUCTURE EQUATIONS 129
and that, on the other hand,
_
n

k=1
g
kj

k
i
+g
ki

k
j
_
(Y ) =
n

k=1
g
kj

k
i
(Y ) +g
ki

k
j
(Y )
=
_
n

k=1

k
i
(Y ) X
k
, X
j
_
+
_
n

k=1

k
j
(Y ) X
k
, X
i
_
=
Y
X
i
, X
j
+
Y
X
j
, X
i
.
Hence, equation (ii) is equivalent to
Y X
i
, X
j
=
Y
X
i
, X
j
+X
i
,
Y
X
j
,
for every i, j, that is, it is equivalent to compatibility with the metric (cf. Ex-
ercise 2.8.2). We conclude that the Levi-Civita connection forms satisfy (i)
and (ii).
To prove unicity, we take 1-forms
j
i
(i, j = 1, . . . , n) satisfying (i) and
(ii). Using (14) and (15), we can dene a connection, which is necessarily
symmetric and compatible with the metric. By uniqueness of the Levi-Civita
connection, we have uniqueness of the set of forms
j
i
satisfying (i) and (ii)
(note that each connection determines a unique set of n
2
connection forms
and vice-versa).
Remark 2.2. Given a eld of frames on some open set, we can perform
Gram-Schmidt orthogonalization to obtain a smooth eld of orthonormal
frames E
1
, . . . , E
n
. Then, as g
ij
= E
i
, E
j
=
ij
, equations (i) and (ii)
above become
(i) d
i
=

n
j=1

j

i
j
,
(ii)
j
i
+
i
j
= 0.
In addition to connection forms, we can also dene curvature forms.
Again we consider an open subset V of M where we have a eld of frames
X
1
, . . . , X
n
(hence a corresponding eld of dual coframes
1
, . . . ,
n
).
We then dene 2-forms
l
k
(k, l = 1, . . . , n) by

l
k
(X, Y ) :=
l
(R(X, Y )X
k
),
for all vector elds X, Y in V (i.e., R(X, Y )X
k
=

n
l=1

l
k
(X, Y )X
l
). Using
the basis
i

j

i<j
for 2-forms, we have

l
k
=

i<j

l
k
(X
i
, X
j
)
i

j
=

i<j

l
(R(X
i
, X
j
)X
k
)
i

j
=

i<j
R
l
ijk

i

j
=
1
2
n

i,j=1
R
l
ijk

i

j
,
where the R
l
ijk
are the coecients of the curvature relative to these frames:
R(X
i
, X
j
)X
k
=
n

l=1
R
l
ijk
X
l
.
130 4. CURVATURE
The curvature forms satisfy the following equation.
Proposition 2.3. In the above notation,
(iii)
j
i
= d
j
i

n
k=1

k
i

j
k
, for every i, j = 1, . . . , n.
Proof. We will show that
R(X, Y )X
i
=
n

j=1

j
i
(X, Y )X
j
=
n

j=1
__
d
j
i

n

k=1

k
i

j
k
_
(X, Y )
_
X
j
.
Indeed,
R(X, Y )X
i
=
X

Y
X
i

Y

X
X
i

[X,Y ]
X
i
=
=
X
_
n

k=1

k
i
(Y )X
k
_

Y
_
n

k=1

k
i
(X)X
k
_

k=1

k
i
([X, Y ])X
k
=
n

k=1
_
X (
k
i
(Y )) Y (
k
i
(X))
k
i
([X, Y ])
_
X
k
+
+
n

k=1

k
i
(Y )
X
X
k

n

k=1

k
i
(X)
Y
X
k
=
n

k=1
d
k
i
(X, Y )X
k
+
n

k,j=1
_

k
i
(Y )
j
k
(X) X
j

k
i
(X)
j
k
(Y ) X
j
_
=
n

j=1
_
d
j
i
(X, Y )
n

k=1
(
k
i

j
k
)(X, Y )
_
X
j
.

Equations (i), (ii) and (iii) are known as the Cartan structure equa-
tions. We list these equations below, as well as the main denitions.
(i) d
i
=

n
j=1

j

i
j
,
(ii) dg
ij
=

n
k=1
(g
kj

k
i
+g
ki

k
j
),
(iii) d
j
i
=
j
i
+

n
k=1

k
i

j
k
,
where
i
(X
j
) =
ij
,
k
j
=

n
i=1

k
ij

i
and
j
i
=

k<l
R
j
kli

k

l
.
Remark 2.4. If we consider a eld of orthonormal frames E
1
, . . . , E
n
,
the above equations become:
(i) d
i
=

n
j=1

j

i
j
,
(ii)
j
i
+
i
j
= 0,
(iii) d
j
i
=
j
i
+

n
k=1

k
i

j
k
(and so
j
i
+
i
j
= 0).
Example 2.5. For a eld of orthonormal frames in R
n
with the Eu-
clidean metric, the curvature forms must vanish (as R = 0), and we obtain
the following structure equations:
(i) d
i
=

n
j=1

j

i
j
,
2. CARTAN STRUCTURE EQUATIONS 131
(ii)
j
i
+
i
j
= 0,
(iii) d
j
i
=

n
k=1

k
i

j
k
.
To nish this section, we will consider in detail the special case of a 2-
dimensional Riemannian manifold. In this case, the structure equations for
a eld of orthonormal frames are particularly simple: equation (ii) implies
that there is only one independent connection form (
1
1
=
2
2
= 0 and

1
2
=
2
1
), which can be computed from equation (i):
d
1
=
2

2
1
;
d
2
=
1

2
1
.
Equation (iii) then yields that there is only one independent curvature form

2
1
= d
2
1
. This form is closely related to the Gauss curvature of the mani-
fold.
Proposition 2.6. If M is a 2-dimensional manifold, then for an or-
thonormal frame we have
2
1
= K
1

2
, where the function K is the
Gauss curvature of M (that is, its sectional curvature).
Proof. Let p be a point in M and let us choose an open set containing
p where we have dened a eld of orthonormal frames E
1
, E
2
. Then
K = R(E
1
, E
2
, E
1
, E
2
) = R
1212
,
and consequently

2
1
=
2
1
(E
1
, E
2
)
1

2
=
2
(R(E
1
, E
2
)E
1
)
1

2
= R(E
1
, E
2
)E
1
, E
2

1

2
= R
1212

1

2
= K
1

2
.

Note that K does not depend on the choice of the eld of frames, since it
is a sectional curvature (cf. Denition 1.5), and, since
1

2
is a Riemannian
volume form, neither does the curvature form (up to a sign). However, the
connection forms do. Let E
1
, E
2
, F
1
, F
2
be two elds of orthonormal
frames on an open subset V of M. Then
_
F
1
F
2
_
=
_
E
1
E
2
_
S
where S : V O(2) has values in the orthogonal group of 2 2 matrices.
Note that S has one of the following two forms
S =
_
a b
b a
_
or S =
_
a b
b a
_
,
where a, b : V R are such that a
2
+b
2
= 1. The determinant of S is then
1 depending on whether the two frames have the same orientation. We
have the following proposition.
Proposition 2.7. If E
1
, E
2
and F
1
, F
2
have the same orientation
then, denoting by
2
1
and
2
1
the corresponding connection forms, we have

2
1

2
1
= , where := a db b da.
132 4. CURVATURE
Proof. Denoting by
1
,
2
and
1
,
2
the elds of dual coframes
corresponding to E
1
, E
2
and F
1
, F
2
, we dene the column vectors of
1-forms
:=
_

1

2
_
and :=
_

1

2
_
and the matrices of 1-forms
A :=
_
0
2
1

2
1
0
_
and

A :=
_
0
2
1

2
1
0
_
.
The relation between the frames can be written as
= S
1
= S
(cf. Section 7.1 in Chapter 2), and the Cartan structure equations as
d = A and d =

A .
Therefore
d = S d +dS = S

A +dS S
1

= S

A S
1
+dS S
1
=
_
S

AS
1
dS S
1
_
,
and unicity of solutions of the Cartan structure equations implies
A = S

AS
1
dSS
1
.
Writing this out in full one obtains
_
0
2
1

2
1
0
_
=
_
0
2
1

2
1
0
_

_
a da +b db b da a db
a db b da a da +b db
_
,
and the result follows (we also obtain a da + b db = 0, which is clear from
det S = a
2
+b
2
= 1).
Let us now give a geometric interpretation of := a db b da. Locally,
we can dene at each point p M the angle (p) between (E
1
)
p
and (F
1
)
p
.
Then the change of basis matrix S has the form
_
a b
b a
_
=
_
cos sin
sin cos
_
.
Hence,
= a db b da = cos d (sin ) sin d (cos )
= cos
2
d + sin
2
d = d.
Therefore, integrating along a curve yields the angle by which F
1
rotates
with respect to E
1
along the curve.
In particular, notice that is closed. This is also clear from
d = d
2
1
d
2
1
= K
1

2
+K
1

2
= 0
(
1

2
=
1

2
since the two elds of frames have the same orientation).
We can use the connection form
2
1
to dene the geodesic curvature
of a curve on an oriented Riemannian 2-manifold M. Let c : I M be
a smooth curve in M parameterized by its arclength s (hence | c(s)| = 1).
2. CARTAN STRUCTURE EQUATIONS 133
Let V be a neighborhood of a point c(s) in this curve where we have a eld
of orthonormal frames E
1
, E
2
satisfying (E
1
)
c(s)
= c(s). Note that it is
always possible to consider such a eld of frames: we start by extending the
vector eld c(s) to a unit vector eld E
1
dened on a neighborhood of c(s),
and then consider a unit vector eld E
2
orthogonal to the rst, such that
E
1
, E
2
is positively oriented. Since

E
1
E
1
=
1
1
(E
1
)E
1
+
2
1
(E
1
)E
2
=
2
1
(E
1
)E
2
,
the covariant acceleration of c is

c(s)
c(s) =
E
1
(s)
E
1
(s) =
2
1
(E
1
(s))E
2
(s).
We dene the geodesic curvature of the curve c to be
k
g
(s) :=
2
1
(E
1
(s))
(thus [k
g
(s)[ = |
c(s)
c(s)|). It is a measure of how much the curve fails to
be a geodesic at c(s). In particular, c is a geodesic if and only if its geodesic
curvature vanishes.
Exercises 2.8.
(1) Let X
1
, . . . , X
n
be a eld of frames on an open set V of a Rie-
mannian manifold (M, , ) with Levi-Civita connection . The
associated structure functions C
k
ij
are dened by
[X
i
, X
j
] =
n

k=1
C
k
ij
X
k
.
Show that:
(a) C
i
jk
=
i
jk

i
kj
;
(b)
i
jk
=
1
2

n
l=1
g
il
(X
j
g
kl
+X
k
g
jl
X
l
g
jk
)
+
1
2
C
i
jk

1
2

n
l,m=1
g
il
_
g
jm
C
m
kl
+g
km
C
m
jl
_
;
(c) d
i
+
1
2

n
j,k=1
C
i
jk

k
= 0, where
1
, . . . ,
n
is the eld
of dual coframes.
(2) Let X
1
, . . . , X
n
be a eld of frames on an open set V of a Rie-
mannian manifold (M, , ). Show that a connection on M is
compatible with the metric on V if and only if
X
k
X
i
, X
j
=
X
k
X
i
, X
j
+X
i
,
X
k
X
j

for all i, j, k.
(3) Compute the Gauss curvature of:
(a) the sphere S
2
with the standard metric;
(b) the hyperbolic plane, i.e., the upper half-plane
H = (x, y) R
2
[ y > 0
with the metric
g =
1
y
2
(dx dx +dy dy)
134 4. CURVATURE
(cf. Exercise 3.3.5 of Chapter 3).
(4) Determine all surfaces of revolution with constant Gauss curvature.
(5) Let M be the image of the parameterization : (0, +) R R
3
given by
(u, v) = (ucos v, usin v, v),
and let N be the image of the parameterization : (0, +) R
R
3
given by
(u, v) = (ucos v, usin v, log u).
Consider in both M and N the Riemannian metric induced by the
Euclidean metric of R
3
. Show that the map f : M N dened by
f((u, v)) = (u, v)
preserves the Gauss curvature but is not a local isometry.
(6) Consider the metric
g = A
2
(r)dr dr +r
2
d d +r
2
sin
2
d d
on M = I S
2
, where r is a local coordinate on I R and (, )
are spherical local coordinates on S
2
.
(a) Compute the Ricci tensor and the scalar curvature of this met-
ric.
(b) What happens when A(r) = (1 r
2
)

1
2
(that is, when M is
locally isometric to S
3
)?
(c) And when A(r) = (1 + r
2
)

1
2
(that is, when M is locally
isometric to the hyperbolic 3-space)?
(d) For which functions A(r) is the scalar curvature constant?
(7) Let M be an oriented Riemannian 2-manifold and let p be a point
in M. Let D be a neighborhood of p in M homeomorphic to a disc,
with a smooth boundary D. Consider a point q D and a unit
vector X
q
T
q
M. Let X be the parallel transport of X
q
along D
in the positive direction. When X returns to q it makes an angle
with the initial vector X
q
. Using elds of positively oriented
orthonormal frames E
1
, E
2
and F
1
, F
2
such that F
1
= X, show
that
=
_
D
K.
Conclude that the Gauss curvature of M at p satises
K(p) = lim
Dp

vol(D)
.
(8) Compute the geodesic curvature of a positively oriented circle on:
(a) R
2
with the Euclidean metric and the usual orientation;
(b) S
2
with the usual metric and orientation.
3. GAUSS-BONNET THEOREM 135
(9) Let c be a smooth curve on an oriented 2-manifold M as in the
denition of geodesic curvature. Let X be a vector eld parallel
along c and let be the angle between X and c(s) along c in the
given orientation. Show that the geodesic curvature of c, k
g
, is
equal to
d
ds
. (Hint: Consider two elds of orthonormal frames E
1
, E
2
and
F
1
, F
2
positively oriented such that E
1
=
X
X
and F
1
= c).
3. Gauss-Bonnet Theorem
We will now use the Cartan structure equations to prove the Gauss-
Bonnet Theorem, relating the curvature of a compact surface to its topol-
ogy. Let M be a compact, oriented, 2-dimensional manifold and X a vector
eld on M.
Definition 3.1. A point p M is said to be a singular point of X
if X
p
= 0. A singular point is said to be an isolated singularity if there
exists a neighborhood V M of p such that p is the only singular point of
X in V .
Since M is compact, if all the singularities of X are isolated then they
are in nite number (as otherwise they would accumulate on a non-isolated
singularity).
To each isolated singularity p V of X X(M) one can associate an
integer number, called the index of X at p, as follows:
(i) x a Riemannian metric in M;
(ii) choose a positively oriented orthonormal frame F
1
, F
2
, dened on
V p, such that
F
1
=
X
|X|
,
let
1
,
2
be the dual coframe and let
2
1
be the corresponding con-
nection form;
(iii) possibly shrinking V , choose a positively oriented orthonormal frame
E
1
, E
2
, dened on V , with dual coframe
1
,
2
and connection
form
2
1
;
(iv) take a neighborhood D of p in V , homeomorphic to a disc, with a
smooth boundary D, endowed with the induced orientation, and de-
ne the index I
p
of X at p as
2I
p
=
_
D
,
where :=
2
1

2
1
is the form in Proposition 2.7.
Recall that satises = d, where is the angle between E
1
and F
1
.
Therefore I
p
must be an integer. Intuitively, the index of a vector eld
X measures the number of times that X rotates as one goes around the
singularity anticlockwise, counted positively if X itself rotates anticlockwise,
and negatively otherwise.
136 4. CURVATURE
Example 3.2. In M = R
2
the following vector elds have isolated sin-
gularities at the origin with the indicated indices (cf. Figure 1):
(1) X
(x,y)
= (x, y) has index 1;
(2) Y
(x,y)
= (y, x) has index 1;
(3) Z
(x,y)
= (y, x) has index 1;
(4) W
(x,y)
= (x, y) has index 1.
(1) (2)
(3) (4)

E
1
=

x
F
1
=
X
|X| F
1
=
Y
|Y |
F
1
=
Z
|Z|
F
1
=
W
|W|
Figure 1. Computing the indices of the vector elds X, Y ,
Z and W.
We will now check that the index is well dened. We begin by observing
that, since is closed, I
p
does not depend on the choice of D. Indeed, the
boundaries of any two such discs are necessarily homotopic (cf. Exercise 5.3.2
in Chapter 2). Next we prove that I
p
does not depend on the choice of the
frame E
1
, E
2
. More precisely, we will show that
I
p
= lim
r0
1
2
_
Sr(p)

2
1
,
where S
r
(p) is the normal sphere of radius r centered at p. Indeed, if r
1
>
r
2
> 0 are radii of normal spheres, one has
(18)
_
Sr
1
(p)

2
1

_
Sr
2
(p)

2
1
=
_

12
d
2
1
=
_

12
K
1

2
=
_

12
K,
3. GAUSS-BONNET THEOREM 137
where
12
= B
r
1
(p) B
r
2
(p). Since K is continuous, we see that
_
_
Sr
1
(p)

2
1

_
Sr
2
(p)

2
1
_
0
as r
1
0. Therefore, if r
n
is a decreasing sequence of positive numbers
converging to zero, the sequence
_
_
Srn
(p)

2
1
_
is a Cauchy sequence, and therefore converges. Let
I
p
:= lim
r0
1
2
_
Sr(p)

2
1
.
Taking the limit as r
2
0 in (18) one obtains
_
Sr
1
(p)

2
1
2I
p
=
_
Br
1
(p)
K =
_
Br
1
(p)
K
1

2
=
_
Br
1
(p)
d
2
1
=
_
Sr
1
(p)

2
1
,
and hence
2I
p
=
_
Sr
1
(p)
=
_
Sr
1
(p)

2
1

2
1
= 2I
p
.
Finally, we show that I
p
does not depend on the choice of Riemannian
metric. Indeed, if ,
0
, ,
1
are two Riemannian metrics on M, it is easy
to check that
,
t
:= (1 t),
0
+t,
1
is also a Riemannian metric on M, and that the index I
p
(t) computed using
the metric ,
t
is a continuous function of t (cf. Exercise 3.6.1). Since I
p
(t)
is an integer for all t [0, 1], we conclude that I
p
(0) = I
p
(1).
Therefore I
p
depends only on the vector eld X X(M). We are now
ready to state the Gauss-Bonnet Theorem:
Theorem 3.3. (Gauss-Bonnet) Let M be a compact, oriented, 2-dimensional
manifold and let X be a vector eld in M with isolated singularities p
1
, . . . , p
k
.
Then
(19)
_
M
K = 2
k

i=1
I
p
i
for any Riemannian metric on M, where K is the Gauss curvature.
Proof. We consider the positively oriented orthonormal frame F
1
, F
2
,
with
F
1
=
X
|X|
,
138 4. CURVATURE
dened on M
k
i=1
p
i
, with dual coframe
1
,
2
and connection form

2
1
. For r > 0 suciently small, we take B
i
:= B
r
(p
i
) such that B
i
B
j
=
for i ,= j and note that
_
M\
k
i=1
B
i
K =
_
M\
k
i=1
B
i
K
1

2
=
_
M\
k
i=1
B
i
d
2
1
=
_

k
i=1
B
i

2
1
=
k

i=1
_
B
i

2
1
,
where the B
i
have the orientation induced by the orientation of B
i
. Taking
the limit as r 0 one obtains
_
M
K = 2
k

i=1
I
p
i
.

Remark 3.4.
(1) Since the right-hand side of (19) does not depend on the metric,
we conclude that
_
M
K is the same for all Riemannian metrics on
M.
(2) Since the left-hand side of (19) does not depend on the vector eld
X, we conclude that (M) :=

k
i=1
I
p
i
is the same for all vector
elds on M with isolated singularities. This is the so-called Euler
characteristic of M.
(3) Recall that a triangulation of M is a decomposition of M in a
nite number of triangles (i.e., images of Euclidean triangles by
parameterizations) such that the intersection of any two triangles
is either a common edge, a common vertex or empty (it is pos-
sible to prove that such a triangulation always exists). Given a
triangulation, one can construct a vector eld X with the following
properties (cf. Figure 2):
(a) each vertex is a singularity which is a sink, that is,
X = x

x
y

y
for certain local coordinates (x, y) centered at the singularity;
(b) the interior of each 2-dimensional face contains exactly one
singularity which is a source, that is
X = x

x
+y

y
for certain local coordinates (x, y) centered at the singularity;
(c) each edge is formed by integral curves of the vector eld and
contains exactly one singularity which is not a vertex.
3. GAUSS-BONNET THEOREM 139
It is easy to see that all singularities are isolated, that the singular-
ities at the vertices and 2-dimensional faces have index 1 and that
the singularities at the edges have index 1. Therefore,
(M) = V E +F,
where V is the number of vertices, E is the number of edges and
F is the number of 2-dimensional faces on any triangulation. This
is the denition we used in Exercise 1.8.5 of Chapter 1.
Figure 2. Vector eld associated to a triangulation.
Example 3.5.
(1) Choosing the standard metric in S
2
, we have
(S
2
) =
1
2
_
S
2
1 =
1
2
vol(S
2
) = 2.
From this we can derive a number of conclusions:
(a) there is no zero curvature metric on S
2
, for this would imply
(S
2
) = 0;
(b) there is no vector eld on S
2
without singularities, as this
would also imply (S
2
) = 0;
(c) for any triangulation of S
2
, one has V E + F = 2. In par-
ticular, this proves Eulers formula for convex polyhedra with
triangular 2-dimensional faces, as these clearly yield triangu-
lations of S
2
.
(2) As we will see in Section 4, the torus T
2
has a zero curvature metric,
and hence (T
2
) = 0. This can also be seen from the fact that there
exist vector elds on T
2
without singularities.
140 4. CURVATURE
Exercises 3.6.
(1) Show that if ,
0
, ,
1
are two Riemannian metrics on M then
,
t
:= (1 t),
0
+t,
1
is also a Riemannian metric on M, and that the index I
p
(t) com-
puted using the metric ,
t
is a continuous function of t.
(2) (Gauss-Bonnet Theorem for non-orientable manifolds) Let (M, g)
be a compact, non-orientable, 2-dimensional Riemannian manifold
and let : M M be its orientable double covering (cf. Exer-
cise 8.6.9 in Chapter 1). Show that:
(a) (M) = 2(M);
(b) K =

K, where K is the Gauss curvature of the Riemannian

metric g :=

g on M;
(c) 2(M) =
1
2
_
M
K.
(Remark: Even though M is not orientable, we can still dene the integral of a
function f on M through

M
f =
1
2

f; with this denition, the Gauss-Bonnet

Theorem holds for non-orientable Riemannian 2-manifolds).
(3) (Gauss-Bonnet Theorem for manifolds with boundary) Let M be
a compact, oriented, 2-dimensional manifold with boundary and
let X be a vector eld in M transverse to M (i.e., such that
X
p
, T
p
M for all p M), with isolated singularities p
1
, . . . , p
k

M M. Prove that
_
M
K +
_
M
k
g
= 2
k

i=1
I
p
i
for any Riemannian metric on M, where K is the Gauss curvature
of M and k
g
is the geodesic curvature of M.
(4) Let (M, g) be a compact orientable 2-dimensional Riemannian man-
ifold, with positive Gauss curvature. Show that any two non-self-
intersecting closed geodesics must intersect each other.
(5) Let M be a dierentiable manifold and f : M R a smooth
function.
(a) (Hessian) Let p M be a critical point of f (i.e. (df)
p
= 0).
The Hessian of f at p is the map (Hf)
p
: T
p
M T
p
M R
given by
(Hf)
p
(v, w) =

2
ts[
s=t=0
(f )(s, t),
where : U R
2
M is such that (0, 0) = p,

s
(0, 0) = v
and

t
(0, 0) = w. Show that (Hf)
p
is a well-dened symmetric
2-tensor.
(b) (Morse Theorem) If (Hf)
p
is nondegenerate then p is called a
nondegenerate critical point. Assume that M is compact
4. MANIFOLDS OF CONSTANT CURVATURE 141
and f is a Morse function, i.e. all its critical points are non-
degenerate. Prove that there is only a nite number of critical
points. Moreover, show that if M is 2-dimensional then
(M) = ms +n,
where m, n and s are the numbers of maxima, minima and
saddle points respectively. (Hint: Choose a Riemannian metric on M
and consider the vector eld X := grad f).
(6) Let (M, g) be a 2-dimensional Riemannian manifold and M a
geodesic triangle, i.e., an open set homeomorphic to an Euclidean
triangle whose sides are images of geodesic arcs. Let , , be the
inner angles of , i.e., the angles between the geodesics at the
intersection points contained in . Prove that for small enough
one has
+ + = +
_

K,
where K is the Gauss curvature of M, using:
(a) the fact that
_

K is the angle by which a vector parallel-

transported once around rotates;
(b) the Gauss-Bonnet Theorem for manifolds with boundary.
(Remark: We can use this result to give another geometric interpretation of the
Gauss curvature: K(p) = lim
p
++
vol()
).
(7) Let (M, g) be a simply connected 2-dimensional Riemannian mani-
fold with nonpositive Gauss curvature. Show that any two geodesics
intersect at most in one point. (Hint: Note that if two geodesics intersected
in more than one point then there would exist a geodesic biangle, i.e., an open set
homeomorphic to a disc whose boundary is formed by the images of two geodesic arcs).
4. Manifolds of Constant Curvature
Recall that a manifold is said to have constant curvature if all sectional
curvatures at all points have the same constant value K. There is an easy
way to identify these manifolds using their curvature forms.
Lemma 4.1. If M is a manifold of constant curvature K, then, around
each point p M, all curvature forms
j
i
satisfy
(20)
j
i
= K
i

j
,
where
1
, . . . ,
n
is any eld of orthonormal coframes dened on a neigh-
borhood of p. Conversely, if on a neighborhood of each point of M there is
a eld of orthonormal frames E
1
, . . . , E
n
such that the corresponding eld
of coframes
1
, . . . ,
n
satises (20) for some constant K, then M has
constant curvature K.
142 4. CURVATURE
Proof. If M has constant curvature K then

j
i
=

k<l

j
i
(E
k
, E
l
)
k

l
=

k<l

j
(R(E
k
, E
l
)E
i
)
k

l
=

k<l
R(E
k
, E
l
)E
i
, E
j

k

l
=

k<l
R
klij

k

l
=

k<l
K(
ki

lj

kj

li
)
k

l
= K
i

j
.
Conversely, let us assume that there is a constant K such that on a neigh-
borhood of each point p M we have
j
i
= K
i

j
. Then, for every
section of the tangent space T
p
M, the corresponding sectional curvature
is given by
K() = R(X, Y, X, Y )
where X, Y are two linearly independent vectors spanning (which we
assume to span a parallelogram of unit area). Using the eld of orthonormal
frames around p, we have X =

n
i=1
X
i
E
i
and Y =

n
i=1
Y
i
E
i
and so,
K() =
n

i,j,k,l=1
X
i
Y
j
X
k
Y
l
R(E
i
, E
j
, E
k
, E
l
)
=
n

i,j,k,l=1
X
i
Y
j
X
k
Y
l

l
k
(E
i
, E
j
)
= K
n

i,j,k,l=1
X
i
Y
j
X
k
Y
l

k

l
(E
i
, E
j
)
= K
n

i,j,k,l=1
X
i
Y
j
X
k
Y
l
_

k
(E
i
)
l
(E
j
)
k
(E
j
)
l
(E
i
)
_
= K
n

i,j,k,l=1
X
i
Y
j
X
k
Y
l
(
ik

jl

jk

il
)
= K
_
|X|
2
|Y |
2
X, Y
2
_
= K.

Let us now see an example of how we can use this lemma.

Example 4.2. The n-dimensional hyperbolic space of radius a > 0,
H
n
(a), is the open half-space
(x
1
, . . . x
n
) R
n
[ x
n
> 0
equipped with the Riemannian metric
g
ij
(x) =
a
2
(x
n
)
2

ij
.
4. MANIFOLDS OF CONSTANT CURVATURE 143
This Riemannian manifold has constant sectional curvature K =
1
a
2
. In-
deed, using the above lemma, we will show that on H
n
(a) there is a eld of
orthonormal frames E
1
, . . . , E
n
whose dual eld of coframes
1
, . . . ,
n

satises
(21)
j
i
= K
i

j
for K =
1
a
2
. For that, let us consider the natural coordinate system x :
H
n
(a) R
n
and the corresponding eld of coordinate frames X
1
, . . . , X
n

with X
i
=

x
i
. Since
X
i
, X
j
=
a
2
(x
n
)
2

ij
,
we obtain a eld of orthonormal frames E
1
, . . . , E
n
with E
i
=
x
n
a
X
i
, and
the corresponding dual eld of coframes
1
, . . . ,
n
where
i
=
a
x
n
dx
i
.
Then
d
i
=
a
(x
n
)
2
dx
i
dx
n
=
1
a

i

n
=
n

j=1

1
a

jn

i
_
,
and so, using the structure equations
d
i
=
n

j=1

j

i
j

j
i
+
i
j
= 0,
we can guess that the connection forms are given by
i
j
=
1
a
(
in

jn

i
).
Indeed, we can easily verify that these forms satisfy the above structure
equations, and hence must be the connection forms by unicity of solution of
these equations. With these forms it is now easy to compute the curvature
forms
j
i
using the third structure equation
d
j
i
=
n

k=1

k
i

j
k
+
j
i
.
We have
d
j
i
= d
_
1
a
(
jn

in

j
)
_
=
1
a
2
(
jn

i

n

in

j

n
)
and
n

k=1

k
i

j
k
=
1
a
2
n

k=1
(
kn

in

k
) (
jn

kn

j
)
=
1
a
2
n

k=1
(
kn

jn

i

k

kn

i

j
+
in

kn

k

j
)
=
1
a
2
(
jn

i

n

i

j
+
in

n

j
),
144 4. CURVATURE
and so,

j
i
= d
j
i

n

k=1

k
i

j
k
=
1
a
2

i

j
.
We conclude that K =
1
a
2
.
The Euclidean spaces R
n
have constant curvature equal to zero. More-
over, we can easily see that the spheres S
n
(r) R
n+1
of radius r have
constant curvature equal to
1
r
2
(cf. Exercise 5.7.2). Therefore we have ex-
amples of manifolds with arbitrary constant negative (H
n
(a)), zero (R
n
) or
positive (S
n
(r)) curvature in any dimension. Note that all these examples
are simply connected and are geodesically complete. Indeed, the geodesics
of the Euclidean space R
n
traverse straight lines, S
n
(r) is compact and the
geodesics of H
n
(a) traverse either half circles perpendicular to the plane
x
n
= 0 and centered on this plane, or vertical half lines starting at the plane
x
n
= 0 (cf. Exercise 4.7.4).
Every simply connected geodesically complete manifold of constant cur-
vature is isometric to one of these examples, as it is stated in the following
theorem (whose proof can be found in [dC93]). In general, if the manifold
is not simply connected (but still geodesically complete), it is isometric to
the quotient of one of the above examples by a free and proper action of
a discrete subgroup of the group of isometries (it can be proved that the
group of isometries of a Riemannian manifold is always a Lie group).
Theorem 4.3. (Killing-Hopf) Let M be a connected, geodesically com-
plete Riemannian manifold with constant curvature K.
(1) If M is simply connected then it is isometric to one of the following:
S
n
_
1

K
_
if K > 0, R
n
if K = 0, or H
n
_
1

K
_
if K < 0.
(2) If M is not simply connected then M is isometric to a quotient

M/, where

M is one of the above simply connected manifolds and
is a nontrivial discrete subgroup of the group of isometries of

M
acting properly and freely on

M.
Example 4.4. Let

M = R
2
. Then the subgroup of isometries cannot
contain isometries with xed points (since it acts freely). Hence can only
contain translations and gliding reections (that is, reections followed by a
translation in the direction of the reection axis). Moreover, is generated
by at most two elements, one of which may be assumed to be a translation
(cf. Exercise 4.7.6). Therefore we have:
(1) if is generated by one translation, then the resulting surface will
be a cylinder;
(2) if is generated by two translations we obtain a torus;
(3) if is generated by a gliding reection we obtain a Mobius band;
(4) if is generated by a translation and a gliding reection we obtain
a Klein bottle.
4. MANIFOLDS OF CONSTANT CURVATURE 145
These are all the possible examples of geodesically complete Euclidean sur-
faces (2-dimensional manifolds of constant zero curvature).
Example 4.5. The group of orientation-preserving isometries of the hy-
perbolic plane H
2
is PSL(2, R) = SL(2, R)/id, acting on H
2
through
_
a b
c d
_
z :=
az +b
cz +d
,
where we make the identication R
2
= C (cf. Exercise 4.7.8 and Section 6.1).
To nd orientable hyperbolic surfaces, that is, surfaces with constant cur-
vature K = 1, we have to nd discrete subgroups of PSL(2, R) acting
properly and freely on H
2
. Here there are many more possibilities. As an
example, we can consider the group = f generated by the translation
f(z) = z + 2. The resulting surface is known as pseudosphere and is
homeomorphic to a cylinder (cf. Figure 3). However, the width of the end
where y + converges to zero, while the width of the end where y 0
converges to +. Its height towards both ends is innite. Note that this
surface has geodesics which transversely auto-intersect a nite number of
times (cf. Figure 4).
Other examples can be obtained by considering hyperbolic polygons
(bounded by geodesics) and identifying their sides through isometries. For
instance, the surface in Figure 5-(b) is obtained by identifying the sides of the
polygon in Figure 5-(a) through the isometries g(z) = z+2 and h(z) =
z
2z+1
.
Choosing other polygons it is possible to obtain compact hyperbolic sur-
faces. In fact, there exist compact hyperbolic surfaces homeomorphic to any
topological 2-manifold with negative Euler characteristic (the Gauss-Bonnet
Theorem does not allow non-negative Euler characteristics).
2 0 2 4

=
Figure 3. Pseudosphere.
Example 4.6. To nd Riemannian manifolds of constant positive cur-
vature we have to nd discrete subgroups of isometries of the sphere that
act properly and freely. Let us consider the case where K = 1. Then
146 4. CURVATURE
2 0 2 4
Figure 4. Trajectories of geodesics on the pseudosphere.
1 0 1

=
(a) (b)
Figure 5. (a) Hyperbolic polygon, (b) Pair of pants.
O(n + 1) (cf. Exercise 4.7.11). Since it must act freely on S
n
, no el-
ement of id can have 1 as an eigenvalue. We will see that, when n is
even, S
n
and RP
n
are the only geodesically complete manifolds of constant
curvature 1. Indeed, if A , then A is an orthogonal (n + 1) (n + 1)
matrix and so all its eigenvalues have absolute value equal to 1. Moreover,
its characteristic polynomial has odd degree (n+1), and so it has a real root,
equal to 1. Consequently, A
2
has 1 as an eigenvalue, and so it has to be
the identity. Hence, A = A
1
= A
t
, and so A is symmetric, implying that
all its eigenvalues are real. The eigenvalues of A are then either all equal
to 1 (if A = id) or all equal to 1, in which case A = id. We conclude
that = id implying that our manifold is either S
n
or RP
n
. If n is odd
there are other possibilities which are classied in [Wol78].
Exercises 4.7.
4. MANIFOLDS OF CONSTANT CURVATURE 147
(1) Show that the metric of H
n
(a) is a left-invariant metric for the Lie
group structure induced by identifying (x
1
, . . . , x
n
) H
n
(a) with
the ane map g : R
n1
R
n1
given by
g(t
1
, . . . , t
n1
) = x
n
(t
1
, . . . , t
n1
) + (x
1
, . . . , x
n1
).
(2) Prove that if the forms
i
in a eld of orthonormal coframes satisfy
d
i
=
i
(with a 1-form), then the connection forms
j
i
are
given by
j
i
= (E
i
)
j
(E
j
)
i
=
i
j
. Use this to conrm the
results in Example 4.2.
(3) Let K be a real number and let = 1+(
K
4
)

n
i=1
(x
i
)
2
. Show that,
for the Riemannian metric dened on R
n
by
g
ij
(p) =
1

2

ij
,
the sectional curvature is constant equal to K.
(4) Show that any isometry of the Euclidean space R
n
which preserves
the coordinate function x
n
is an isometry of H
n
(a). Use this fact
to determine all the geodesics of H
n
(a).
(5) (Schur Theorem) Let M be a connected isotropic Riemannian man-
ifold of dimension n 3. Show that M has constant curvature.
(Hint: Use the structure equations to show that dK = 0).
(6) To complete the details in Example 4.4, show that:
(a) the isometries of R
2
with no xed points are either translations
or gliding reections;
(b) any discrete group of isometries of R
2
acting properly and
freely is generated by at most two elements, one of which may
be assumed to be a translation.
(7) Let f, g : R
2
R
2
be the isometries
f(x, y) = (x, y + 1) and g(x, y) = (x + 1, y)
(thus f is a gliding reection and g is a translation). Check that
R
2
/f is homeomorphic to a Mobius band (without boundary),
and that R
2
/f, g is homeomorphic to a Klein bottle.
(8) Let H
2
be the hyperbolic plane. Show that:
(a) The formula
_
a b
c d
_
z :=
az +b
cz +d
(ad bc = 1)
denes an action of PSL(2, R) := SL(2, R)/id on H
2
by
orientation-preserving isometries;
(b) for any two geodesics c
1
, c
2
: R H
2
, parameterized by the
arclength, there exists g PSL(2, R) such that c
1
(s) = gc
2
(s)
for all s R;
148 4. CURVATURE
(c) if f : H
2
H
2
is an orientation-preserving isometry then
it must be a holomorphic function (cf. Section 6.1). Con-
clude that all orientation-preserving isometries are of the form
f(z) = g z for some g PSL(2, R).
(9) Check that the isometries g(z) = z + 2 and h(z) =
z
2z+1
of the
hyperbolic plane in Example 4.5 identify the sides of the hyperbolic
polygon in Figure 5.
(10) A tractrix is the curve described parametrically by
_
x = u tanh u
y = sech u
(u > 0)
(its name derives from the property that the distance between any
point in the curve and the x-axis along the tangent is constant equal
to 1). Show that the surface of revolution generated by rotating a
tractrix about the x-axis (tractroid) has constant Gauss curvature
K = 1. Determine an open subset of the pseudosphere isometric
to the tractroid. (Remark: The tractroid is not geodesically complete; in fact,
it was proved by Hilbert in 1901 that any surface of constant negative curvature
embedded in Euclidean 3-space must be incomplete).
(11) Show that the group of isometries of S
n
is O(n + 1).
(12) Let G be a compact Lie group of dimension 2. Show that:
(a) G is orientable;
(b) (G) = 0;
(c) any left-invariant metric on G has constant curvature;
(d) G is the 2-torus T
2
.
5. Isometric Immersions
Many Riemannian manifolds arise as submanifolds of other Riemannian
manifolds, by taking the induced metric (e.g. S
n
R
n+1
). In this section,
we will analyze how the curvatures of the two manifolds are related.
Let f : N M be an immersion of an n-manifold N on an m-manifold
M. We know from Section 5 of Chapter 1 that for each point p N there
is a neighborhood V N of p where f is an embedding onto its image.
Hence f(V ) is a submanifold of M. To simplify notation, we will identify
V with f(V ), and proceed as if f were the inclusion map. Let , be a
Riemannian metric on M and let , be the metric induced on N by f
(which is therefore called an isometric immersion). For every p V , the
tangent space T
p
M can be decomposed as
T
p
M = T
p
N (T
p
N)

.
Therefore, every element v of T
p
M can be written uniquely as v = v

+v

,
where v

T
p
N is the tangential part of v and v

(T
p
N)

is the normal
part of v. Let

and be the Levi-Civita connections of (M, , ) and
(N, , ), respectively. Let X, Y be two vector elds in V N and let
5. ISOMETRIC IMMERSIONS 149

X,

Y be two extensions of X, Y to a neighborhood W M of V . Using the
Koszul formula, we can easily check that

X
Y =
_

Y
_

(cf. Exercise 3.3.6 in Chapter 3). We dene the second fundamental form
of N as
B(X, Y ) :=

Y
X
Y.
Note that this map is well dened, that is, it does not depend on the exten-
sions

X,

Y of X, Y (cf. Exercise 5.7.1). Moreover, it is bilinear, symmetric,
and, for each p V , B(X, Y )
p
(T
p
N)

depends only on the values of X

p
and Y
p
.
Using the second fundamental form, we can dene, for each vector n
p

(T
p
N)

, a symmetric bilinear map H

np
: T
p
N T
p
N R through
H
np
(X
p
, Y
p
) = B(X
p
, Y
p
), n
p
.
The corresponding quadratic form is often called the second fundamental
form of f at p along the vector n
p
.
Finally, since H
np
is bilinear, there exists a linear map S
np
: T
p
N T
p
N
satisfying
S
np
(X
p
), Y
p
= H
np
(X
p
, Y
p
) = B(X
p
, Y
p
), n
p

for all X
p
, Y
p
T
p
N. It is easy to check that this linear map is given by
S
np
(X
p
) = (

X
n)

p
,
where n is a local extension of n
p
normal to N. Indeed, since

Y , n = 0 on
N and

X is tangent to N, we have on N
S
n
(X), Y = B(X, Y ), n =

Y
X
Y, n
=

Y , n =

X

Y , n

Y ,

X
n
=

X
n,

Y = (

X
n)

, Y .
Therefore
S
np
(X
p
), Y
p
= (

X
n)

p
, Y
p

for all Y
p
T
p
N.
Example 5.1. Let N be a hypersurface in M, i.e. let dimN = n and
dimM = n + 1. Consider a point p V (a neighborhood of N where f
is an embedding), and a unit vector n
p
normal to N at p. As the linear
map S
np
: T
p
N T
p
N is symmetric, there exists an orthonormal basis
of T
p
N formed by eigenvectors (E
1
)
p
, . . . , (E
n
)
p
(called principal di-
rections at p) corresponding to the set of real eigenvalues
1
, . . . ,
n
of
S
np
(called principal curvatures at p). The determinant of the map S
np
(equal to the product
1

n
) is called the Gauss curvature of f and
H :=
1
n
tr S
np
=
1
n
(
1
+ +
n
) is called the mean curvature of f. When
150 4. CURVATURE
n = 2 and M = R
3
with the Euclidean metric, the Gauss curvature of f is
in fact the Gauss curvature of N as dened in Section 1 (cf. Example 5.5).
Example 5.2. If, in the above example, M = R
n+1
with the Euclidean
metric, we can dene the Gauss map g : V N S
n
, with values on
the unit sphere, which, to each point p V , assigns the normal unit vector
n
p
. Since n
p
is normal to T
p
N, we can identify the tangent spaces T
p
N and
T
g(p)
S
n
and obtain a well-dened map (dg)
p
: T
p
N T
p
N. Choosing a
curve c : I N with c(0) = p and c(0) = X
p
T
p
N, we have
(dg)
p
(X
p
) =
d
dt
(g c)
[
t=0
=
d
dt
n
c(t)
[
t=0
= (

c
n)
p
,
where we used the fact

is the Levi-Civita connection for the Euclidean
metric. However, since |n| = 1, we have
0 = c(t) n, n = 2

c
n, n,
implying that
(dg)
p
(X
p
) = (

c
n)
p
= (

c
n)

p
= S
np
(X
p
).
We conclude that the derivative of the Gauss map at p is (dg)
p
= S
np
.
Let us now relate the curvatures of N and M.
Proposition 5.3. Let p be a point in N, let X
p
and Y
p
be two linearly
independent vectors in T
p
N T
p
M and let T
p
N T
p
M be the 2-
dimensional subspace generated by these vectors. Let K
N
() and K
M
()
denote the corresponding sectional curvatures in N and M, respectively.
Then
K
N
() K
M
() =
B(X
p
, X
p
), B(Y
p
, Y
p
) |B(X
p
, Y
p
)|
2
|X
p
|
2
|Y
p
|
2
X
p
, Y
p

2
.
Proof. Observing that the right-hand side depends only on , we can
assume, without loss of generality, that X
p
, Y
p
is orthonormal. Let X, Y be
local extensions of X
p
, Y
p
, dened on a neighborhood of p in N and tangent
to N, also orthonormal. Let

X,

Y be extensions of X, Y to a neighborhood
of p in M. Moreover, consider a eld of frames E
1
, . . . , E
n+k
, also dened
on a neighborhood of p in M, such that E
1
, . . . , E
n
are tangent to N with
E
1
= X and E
2
= Y on N, and E
n+1
, . . . , E
n+k
are normal to N (n+k = m).
Then, since B(X, Y ) is normal to N,
B(X, Y ) =
k

i=1
B(X, Y ), E
n+i
E
n+i
=
k

i=1
H
E
n+i
(X, Y ) E
n+i
.
5. ISOMETRIC IMMERSIONS 151
On the other hand,
K
N
() K
M
() = R
N
(X
p
, Y
p
, X
p
, Y
p
) +R
M
(

X
p
,

Y
p
,

X
p
,

Y
p
)
= (
X

Y
X +
Y

X
X +
[X,Y ]
X
+

[

X,

Y ]

X)
p
, Y
p

= (
X

Y
X +
Y

X
X +

X)
p
, Y
p
,
where we have used the fact that

[

X,

Y ]

X
[X,Y ]
X is normal to N (cf. Ex-
ercise 5.7.1). However, since on N

X =

Y
(B(X, X) +
X
X) =
=

Y
_
k

i=1
H
E
n+i
(X, X)E
n+i
+
X
X
_
=
k

i=1
_
H
E
n+i
(X, X)

Y
E
n+i
+

Y (H
E
n+i
(X, X))E
n+i
_
+

Y

X
X,
we have

X, Y =
k

i=1
H
E
n+i
(X, X)

Y
E
n+i
, Y +

Y

X
X, Y .
Moreover,
0 =

Y E
n+i
, Y =

Y
E
n+i
, Y +E
n+i
,

Y
Y
=

Y
E
n+i
, Y +E
n+i
, B(Y, Y ) +
Y
Y
=

Y
E
n+i
, Y +E
n+i
, B(Y, Y )
=

Y
E
n+i
, Y +H
E
n+i
(Y, Y ),
and so

X, Y =
k

i=1
H
E
n+i
(X, X)H
E
n+i
(Y, Y ) +

Y

X
X, Y
=
k

i=1
H
E
n+i
(X, X)H
E
n+i
(Y, Y ) +
Y

X
X, Y .
Similarly, we can conclude that

X, Y =
k

i=1
H
E
n+i
(X, Y )H
E
n+i
(X, Y ) +
X

Y
X, Y ,
152 4. CURVATURE
and then
K
N
() K
M
() =
=
k

i=1
_
(H
E
n+i
(X
p
, Y
p
))
2
+H
E
n+i
(X
p
, X
p
)H
E
n+i
(Y
p
, Y
p
)
_
= |B(X
p
, Y
p
)|
2
+B(X
p
, X
p
), B(Y
p
, Y
p
).

Example 5.4. Again in the case of a hypersurface N, we choose an

orthonormal basis (E
1
)
p
, . . . , (E
n
)
p
of T
p
N formed by eigenvectors of S
np
,
where n
p
(T
p
N)

. Hence, considering a section of T

p
N generated by
two of these vectors (E
i
)
p
, (E
j
)
p
, and using B(X
p
, Y
p
) = S
np
(X
p
), Y
p
n
p
,
we have
K
N
()K
M
() =
= |B((E
i
)
p
, (E
j
)
p
)|
2
+B((E
i
)
p
, (E
i
)
p
), B((E
j
)
p
, (E
j
)
p
)
= S
np
((E
i
)
p
), (E
j
)
p

2
+S
np
((E
i
)
p
), (E
i
)
p
S
np
((E
j
)
p
), (E
j
)
p

=
i

j
.
Example 5.5. In the special case where N is a 2-manifold, and M = R
3
with the Euclidean metric, we have K
M
0 and hence K
N
(p) =
1

2
, as
promised in Example 5.1. Therefore, although
1
and
2
depend on the
immersion, their product depends only on the intrinsic geometry of N.
Gauss was so pleased by this discovery that he called it his Theorema
Egregium (Remarkable Theorem).
Let us now study in detail the particular case where N is a hypersurface
in M = R
n+1
with the Euclidean metric. Let c : I N be a curve in N
parameterized by arc length s and such that c(0) = p and c(0) = X
p
T
p
N.
We will identify this curve c with the curve f c in R
n+1
. Considering the
Gauss map g : V S
n
dened on a neighborhood V of p in N, we take
the curve n(s) := (g c)(s) in S
n
. Since

is the Levi-Civita connection
corresponding to the Euclidean metric in R
3
, we have

c
c, n = c, n. On
the other hand,

c
c, n = B( c, c) +
c
c, n = B( c, c), n = H
n
( c, c).
Hence, at s = 0, H
g(p)
(X
p
, X
p
) = c(0), n
p
. This value k
np
:= c(0), n
p
is
called the normal curvature of c at p. Since k
np
is equal to H
g(p)
(X
p
, X
p
),
it does not depend on the curve, but only on its initial velocity. Because
H
g(p)
(X
p
, X
p
) = S
g(p)
(X
p
), X
p
, the critical values of these curvatures
subject to |X
p
| = 1 are equal to
1
, . . . ,
n
, and are called the principal
curvatures. This is why in Example 5.1 we also called the eigenvalues of
S
np
principal curvatures. The Gauss curvature of f is then equal to the
product of the principal curvatures, K =
1
. . .
n
. As the normal curvature
does not depend on the choice of curve tangent to X
p
at p, we can choose c
5. ISOMETRIC IMMERSIONS 153
to take values on the 2-plane generated by X
p
and n
p
. Then c(0) is parallel
to the normal vector n
p
, and
[k
n
[ = [ c(0), n[ = | c(0)| = k
c
,
where k
c
:= | c(0)| is the so-called curvature of the curve c at c(0). The
same formula holds if c is a geodesic of N (cf. Exercise 5.7.6).
Example 5.6. Let us consider the following three surfaces: the 2-sphere,
the cylinder and the saddle surface z = xy.
(1) Let p be any point on the sphere. Intuitively, all points of this
surface are on the same side of the tangent plane at p, implying
that both principal curvatures have the same sign (depending on
the chosen orientation), and consequently that the Gauss curvature
is positive at all points.
(2) If p is any point on the cylinder, one of the principal curvatures
is zero (the maximum or the minimum, depending on the chosen
orientation), and so the Gauss curvature is zero at all points.
(3) Finally, if p is a point on the saddle surface z = xy then the princi-
pal curvatures at p have opposite signs, and so the Gauss curvature
is negative.
Exercises 5.7.
(1) Let M be a Riemannian manifold with Levi-Civita connection

,
and let N be a submanifold endowed with the induced metric and
Levi-Civita connection . Let

X,

Y X(M) be local extensions
of X, Y X(N). Recall that the second fundamental form of the
inclusion of N in M is the map B : T
p
N T
p
N (T
p
N)

dened
at each point p N by
B(X, Y ) :=

Y
X
Y.
Show that:
(a) B(X, Y ) does not depend on the choice of the extensions

X,

Y ;
(b) B(X, Y ) is orthogonal to N;
(c) B is symmetric, i.e. B(X, Y ) = B(Y, X);
(d) B is bilinear;
(e) B(X, Y )
p
depends only on the values of X
p
and Y
p
;
(f)

[

X,

Y ]

X
[X,Y ]
X is orthogonal to N.
(2) Let S
n
(r) R
n+1
be the n dimensional sphere of radius r.
a) Choosing at each point the outward pointing normal unit vec-
tor, what is the Gauss map of this inclusion?
b) What are the eigenvalues of its derivative?
c) Show that all sectional curvatures are equal to
1
r
2
(so S
n
(r)
has constant curvature
1
r
2
).
(3) Let (M, , ) be a Riemannian manifold. A submanifold N M is
said to be totally geodesic if the the geodesics of N are geodesics
of M. Show that:
154 4. CURVATURE
(a) N is totally geodesic if and only if B 0, where B is the
second fundamental form of N;
(b) if N is the set of xed points of an isometry then N is totally
geodesic. Use this result to give examples of totally geodesic
submanifolds of R
n
, S
n
and H
n
.
(4) Let N be a hypersurface in R
n+1
and let p be a point in N. Show
that if K(p) ,= 0 then
[K(p)[ = lim
Dp
vol(g(D))
vol(D)
,
where g : V N S
n
is the Gauss map and D is a neighborhood
of p whose diameter tends to zero.
(5) Let (M, , ) be a Riemannian manifold, p a point in M and a
section of T
p
M. For B

(p) := exp
p
(B

(0)) a normal ball around p

consider the set N
p
:= exp
p
(B

(0) ). Show that:

a) the set N
p
is a 2-dimensional submanifold of M formed by the
segments of geodesics in B

(p) which are tangent to at p;

b) if in N
p
we use the metric induced by the metric in M, the
sectional curvature K
M
() is equal to the Gauss curvature of
the 2-manifold N
p
.
(6) Let (M, , ) be a Riemannian manifold with Levi-Civita connec-
tion

and let N be a hypersurface in M. The geodesic curva-
ture of a curve c : I R M, parameterized by arclength, is
k
g
(s) = |

c(s)
c(s)|. Show that the absolute values of the principal
curvatures are the geodesic curvatures (in M) of the geodesics of N
tangent to the principal directions. (Remark: In the case of an oriented
2-dimensional Riemannian manifold, kg is taken to be positive or negative according
to the orientation of c(s),

c(s)
c(s) cf. Section 2).
(7) Use the Gauss map to compute the Gauss curvature of the following
surfaces in R
3
:
(a) the paraboloid z =
1
2
_
x
2
+y
2
_
;
(b) the saddle surface z = xy.
(8) (Surfaces of revolution) Consider the map f : R (0, 2) R
3
given by
f(s, ) = (h(s) cos , h(s) sin , g(s))
with h > 0 and g smooth maps such that
(h

(s))
2
+ (g

(s))
2
= 1.
The image of f is the surface of revolution S with axis Oz, obtained
by rotating the curve (s) = (h(s), g(s)), parameterized by the
arclength s, around that axis.
(a) Show that f is an immersion.
(b) Show that f
s
:= (df)
_

s
_
and f

:= (df)(

) are orthogonal.
6. NOTES ON CHAPTER 4 155
(c) Determine the Gauss map and compute the matrix of the sec-
ond fundamental form of S associated to the frame E
s
, E

,
where E
s
:= f
s
and E

:=
1
|f

|
f

.
(d) Compute the mean curvature H and the Gauss curvature K
of S.
(e) Using these results, give examples of surfaces of revolution
with:
(i) K 0;
(ii) K 1;
(iii) K 1;
(iv) H 0 (not a plane).
(Remark: Surfaces with constant zero mean curvature are called minimal surfaces;
it can be proved that if a compact surface with boundary has minimum area among
all surfaces with the same boundary then it must be a minimal surface).
6. Notes on Chapter 4
6.1. Section 5. In the exercises of Section 5 we need several deni-
tions and facts about complex functions of a complex variable which we will
discuss briey (refer for instance to [Ahl79] for a detailed exposition)
1. A complex function f of a complex variable is said to be holomor-
phic on some open set U C if it has derivative
f

(z) = lim
h0
f(z +h) f(z)
h
for all z U. If we write z = x +iy and
f(z) = f(x +iy) = u(x, y) +iv(x, y)
with u, v real valued functions dened on U R
2
= C, it is easy
to see that when the derivative exits one has
f

(z) =
f
x
=
u
x
+i
v
x
= i
f
y
=
v
y
i
u
y
.
If follows that a holomorphic function must satisfy the so-called
Cauchy-Riemann equations
u
x
=
v
y
;
u
y
=
v
x
.
Conversely, it is easy to check that if u and v have continuous rst
order partial derivatives satisfying the Cauchy-Riemann equations
then f = u +iv is holomorphic.
156 4. CURVATURE
2. It can be shown that every holomorphic function is locally the sum
of a convergent power series. Consequently, the zeros of a holo-
morphic function that does not vanish identically are isolated.
3. An important class of holomorphic functions are the linear frac-
tional transformations or Mobius transformations
f(z) =
az +b
cz +d
with adbc ,= 0. It is easy to see that each of these transformations
is a composition of the following types of transformations:
(a) translations: z z +b;
(b) rotations: z az, [a[ = 1;
(c) homotheties: z rz, r > 0;
(d) inversions: z 1/z,
and so it is clear that they carry straight lines and circles to either
straight lines or circles.
The special values f() =
a
c
and f(
d
c
) = can be introduced
as limits for z and z d/c, and so, using the stereographic
projection, we can see f as a map from the sphere to itself. Noting
that both straight lines and circles in the plane correspond to circles
in the sphere, we can say that a Mobius transformation, seen as a
map on the sphere, carries circles into circles.
6.2. Bibliographical notes. The material in this chapter can be found
in most books on Riemannian geometry (e.g. [Boo03, dC93, GHL04]).
The proof of the Gauss-Bonnet Theorem (due to S. Chern) follows [dC93,
CCL00] closely. See [KN96, Jos02] to see how this theorem ts within the
general theory of characteristic classes of ber bundles. A more elementary
discussion of isometric immersions of surfaces in R
3
(including a proof of the
Gauss-Bonnet Theorem) can be found in [dC76, Mor98].
CHAPTER 5
Geometric Mechanics
This chapter uses Riemannian Geometry to give a geometric formulation
of Newtonian Mechanics.
Section 1 presents the notion of an abstract mechanical system. Sec-
tion 2 explains how holonomic constraints yield nontrivial examples of
such systems, as for instance the rigid body, which is studied in detail in
Section 3. Non-holonomic constraints are considered in Section 4.
Section 5 presents the Lagrangian formulation of mechanics, includ-
ing the Noether Theorem, which relates symmetries to conservation laws.
The dual Hamiltonian formulation is described in Section 6, and used
in Section 7 to formulate the theory of completely integrable systems.
Finally, Section 8 discusses symplectic and Poisson manifolds and re-
duction.
1. Mechanical Systems
In Mechanics one studies the motions of particles or systems of particles
subject to known forces.
Example 1.1. The motion of a single particle in n-dimensional space is
described by a curve x : I R R
n
. It is generally assumed that the force
acting on the particle depends only on its position and velocity. Newtons
Second Law requires that the particles motion satises the second order
ordinary dierential equation
m x = F(x, x),
where F : R
n
R
n
R
n
is the force acting on the particle and m > 0
is the particles mass. Therefore the solutions of this equation describe the
possible motions of the particle.
It will prove advantageous to make the following generalization:
Definition 1.2. A mechanical system is a triple (M, , , T), where:
(i) M is a dierentiable manifold, called the conguration space;
(ii) , is a Riemannian metric on M yielding the mass operator :
TM T

M, dened by
(v)(w) = v, w
for all v, w T
p
M and p M;
157
158 5. GEOMETRIC MECHANICS
(iii) T : TM T

M is a dierentiable map satisfying T(T

p
M) T

p
M
for all p M, called the external force.
A motion of the mechanical system is a solution c : I R M of the
Newton equation

_
D c
dt
_
= T( c).
Remark 1.3. In particular, the geodesics of a Riemannian manifold
(M, , ) are the motions of the mechanical system (M, , , 0) (describing
a free particle on M).
Example 1.4. For the mechanical system comprising a single particle
moving in n-dimensional space, the conguration space is clearly R
n
. If we
set
v, w := mv, w
for all v, w R
n
, where , is the Euclidean inner product in R
n
, then the
Levi-Civita connection of , will still be the trivial connection, and
D x
dt
= x.
Setting
(22) T(x, v)(w) := F(x, v), w
for all v, w R
n
, we see that

_
D x
dt
_
= T(x, x)
_
D x
dt
_
(v) = T(x, x)(v) for all v R
n
m x, v = F(x, x), v for all v R
n
m x = F(x, x).
Hence the motions of the particle are the motions of the mechanical system
(R
n
, , , T) with T dened by (22).
Definition 1.5. Let (M, , , T) be a mechanical system. The external
force T is said to be:
(i) positional if T(v) depends only on (v), where : TM M is the
natural projection;
(ii) conservative if there exists U : M R such that T(v) = (dU)
(v)
for all v TM (the function U is called the potential energy).
Remark 1.6. In particular any conservative force is positional. A me-
chanical system whose exterior force is conservative is called a conservative
mechanical system.
Definition 1.7. Let (M, , , T) be a mechanical system. The kinetic
energy is the dierentiable map K : TM R given by
K(v) =
1
2
v, v
for all v TM.
1. MECHANICAL SYSTEMS 159
Example 1.8. For the mechanical system comprising a single particle
moving in n-dimensional space, one has
K(v) :=
1
2
mv, v.
Theorem 1.9. (Conservation of Energy) In a conservative mechanical
system (M, , , dU), the mechanical energy E(t) = K( c(t)) + U(c(t))
is constant along any motion c : I R M.
Proof.
dE
dt
(t) =
d
dt
_
1
2
c(t), c(t) +U(c(t))
_
=
_
D c
dt
(t), c(t)
_
+ (dU)
c(t)
c(t)
=
_
D c
dt
_
( c) T( c)( c) = 0.

A particularly simple example of a conservative mechanical system is

(M, , , 0), whose motions are the geodesics of (M, , ). In fact, the
motions of any conservative system can be suitably reinterpreted as the
geodesics of a certain metric.
Definition 1.10. Let (M, , , dU) be a conservative mechanical sys-
tem and h R such that
M
h
:= p M [ U(p) < h , = .
The Jacobi metric on the manifold M
h
is given by
v, w := 2 [h U(p)] v, w
for all v, w T
p
M
h
and p M
h
.
Theorem 1.11. (Jacobi) The motions of a conservative mechanical sys-
tem (M, , , dU) with mechanical energy h are, up to reparameterization,
geodesics of the Jacobi metric on M
h
.
Proof. We shall need the two following lemmas, whose proofs are left
as exercises.
Lemma 1.12. Let (M, , ) be a Riemannian manifold with Levi-Civita
connection and , = e
2
, a metric conformally related to ,
(where C

(M)). Then the Levi-Civita connection

of , is given
by

X
Y =
X
Y +d(X)Y +d(Y )X X, Y grad
for all X, Y X(M) (where the gradient is taken with respect to , ).
Lemma 1.13. A curve c : I R M is a reparameterized geodesic of
a Riemannian manifold (M, , ) if and only if it satises
D c
dt
= f(t) c
for some dierentiable function f : I R.
160 5. GEOMETRIC MECHANICS
We now prove the Jacobi Theorem. Let c : I R M be a motion of
(M, , , dU) with mechanical energy h. Then Lemma 1.12 yields

D c
dt
=
D c
dt
+ 2d( c) c c, c grad ,
where

D
dt
is the covariant derivative along c with respect to the Jacobi metric
and e
2
= 2(h U). The Newton equation yields

_
D c
dt
_
= dU
D c
dt
= grad U = e
2
grad ,
and by conservation of energy
c, c = 2K = 2(h U) = e
2
.
Consequently we have

D c
dt
= 2d( c) c,
which by Lemma 1.13 means that c is a reparameterized geodesic of the
Jacobi metric.
A very useful expression for writing the Newton equation in local coor-
dinates is the following.
Proposition 1.14. Let (M, , , T) be a mechanical system. If (x
1
, . . . , x
n
)
are local coordinates on M and (x
1
, . . . , x
n
, v
1
, . . . , v
n
) are the local coordi-
nates induced on TM then

_
D c
dt
(t)
_
=
n

i=1
_
d
dt
_
K
v
i
(x(t), x(t))
_

K
x
i
(x(t), x(t))
_
dx
i
.
In particular, if T = dU is conservative then the equations of motion are
d
dt
_
K
v
i
(x(t), x(t))
_

K
x
i
(x(t), x(t)) =
U
x
i
(x(t))
(i = 1, . . . , n).
Proof. Recall that the local coordinates (x
1
, . . . , x
n
, v
1
, . . . , v
n
) on TM
parameterize the vector
n

i=1
v
i

x
i
which is tangent to M at the point with coordinates (x
1
, . . . , x
n
). Therefore,
we have
K(x
1
, . . . , x
n
, v
1
, . . . , v
n
) =
1
2
n

i,j=1
g
ij
(x
1
, . . . , x
n
)v
i
v
j
,
where
g
ij
=
_

x
i
,

x
j
_
1. MECHANICAL SYSTEMS 161
are the components of the metric in this coordinate system. Consequently,
K
v
i
=
n

j=1
g
ij
v
j
and hence
K
v
i
(x(t), x(t)) =
n

j=1
g
ij
(x(t)) x
j
(t),
leading to
d
dt
_
K
v
i
(x(t), x(t))
_
=
n

j=1
g
ij
(x(t)) x
j
(t) +
n

j,k=1
g
ij
x
k
(x(t)) x
k
(t) x
j
(t).
Moreover,
K
x
i
=
1
2
n

j,k=1
g
jk
x
i
v
j
v
k
,
and hence
K
x
i
(x(t), x(t)) =
1
2
n

j,k=1
g
jk
x
i
(x(t)) x
j
(t) x
k
(t).
We conclude that
d
dt
_
K
v
i
(x(t), x(t))
_

K
x
i
(x(t), x(t)) =
n

j=1
g
ij
(x(t)) x
j
(t) +
n

j,k=1
_
g
ij
x
k
(x(t))
1
2
g
jk
x
i
(x(t))
_
x
j
(t) x
k
(t).
On the other hand, if v, w T
p
M are written as
v =
n

i=1
v
i

x
i
, w =
n

i=1
w
i

x
i
then we have
(v)(w) =
n

i,j=1
g
ij
v
i
w
j
=
n

i,j=1
g
ij
v
i
dx
j
(w),
and hence
(v) =
n

i,j=1
g
ij
v
i
dx
j
=
n

i,j=1
g
ij
v
j
dx
i
.
Therefore

_
D c
dt
(t)
_
=
n

i,j=1
g
ij
(x(t))
_
_
x
j
(t) +
n

k,l=1

j
kl
(x(t)) x
k
(t) x
l
(t)
_
_
dx
i
.
162 5. GEOMETRIC MECHANICS
Since
n

j=1
g
ij

j
kl
=
1
2
n

j,m=1
g
ij
g
jm
_
g
ml
x
k
+
g
mk
x
l

g
kl
x
m
_
=
1
2
_
g
il
x
k
+
g
ik
x
l

g
kl
x
i
_
,
we have
n

j,k,l=1
g
ij
(x(t))
j
kl
(x(t)) x
k
(t) x
l
(t)
=
1
2
n

k,l=1
_
g
il
x
k
(x(t)) +
g
ik
x
l
(x(t))
g
kl
x
i
(x(t))
_
x
k
(t) x
l
(t)
=
n

j,k=1
_
g
ij
x
k
(x(t))
1
2
g
jk
x
i
(x(t))
_
x
j
(t) x
k
(t),
which completes the proof.
Example 1.15.
(1) (Particle in a central eld) Consider a particle of mass m > 0
moving in R
2
under the inuence of a conservative force whose
potential energy U depends only on the distance r =
_
x
2
+y
2
to the origin, U = u(r). The equations of motion are most easily
solved when written in polar coordinates (r, ), dened by
_
x = r cos
y = r sin
.
Since
dx = cos dr r sin d,
dy = sin dr +r cos d,
it is easily seen that the Euclidean metric is written in these coor-
dinates as
, = dx dx +dy dy = dr dr +r
2
d d,
and hence
K
_
r, , v
r
, v

_
=
1
2
m
_
(v
r
)
2
+r
2
_
v

_
2
_
.
Therefore we have
K
v
r
= mv
r
,
K
v

= mr
2
v

,
K
r
= mr
_
v

_
2
,
K

= 0,
1. MECHANICAL SYSTEMS 163
and consequently the Newton equations are written
d
dt
(m r) mr

2
= u

(r),
d
dt
_
mr
2

_
= 0.
Notice that the angular momentum
p

:= mr
2

is constant along the motion. This conservation law can be traced

back to the fact that neither K nor U depend on .
(2) (Christoel symbols for the 2-sphere) The metric for the 2-sphere
S
2
R
3
is written as
, = d d + sin
2
d d
in the usual local coordinates (, ) dened by the parameterization
(, ) = (sin cos , sin sin, cos )
(cf. Exercise 3.3.4 in Chapter 3). A quick way to obtain the Christof-
fel symbols in this coordinate system is to write out the Newton
equations for a free particle (of mass m = 1, say) on S
2
. We have
K
_
, , v

, v

_
=
1
2
_
_
v

_
2
+ sin
2
(v

)
2
_
and hence
K
v

= v

,
K
v

= sin
2
v

,
K

= sin cos (v

)
2
,
K

= 0.
Consequently the Newton equations are written
d
dt
_

_
sin cos
2
= 0

sin cos
2
= 0,
d
dt
_
sin
2

_
= 0 + 2 cot

= 0.
Since these must be the equations for a geodesic on S
2
, by compar-
ing with the geodesic equations
x
i
+
2

j,k=1

i
jk
x
j
x
k
= 0 (i = 1, 2),
one immediately reads o the nonvanishing Christoel symbols:

= sin cos ,

= cot .
Exercises 1.16.
(1) Generalize Examples 1.1, 1.4 and 1.8 to a system of k particles
moving in R
n
.
164 5. GEOMETRIC MECHANICS
(2) Let (M, , , T) be a mechanical system. Show that the Newton
equation denes a ow on TM, generated by the vector eld X
X(TM) whose local expression is
X = v
i

x
i
+
_
_
n

j=1
g
ij
(x)F
j
(x, v)
n

j,k=1

i
jk
(x)v
j
v
k
_
_

v
i
,
where (x
1
, . . . , x
n
) are local coordinates on M, (x
1
, . . . , x
n
, v
1
, . . . , v
n
)
are the local coordinates induced on TM, and
T =
n

i=1
F
i
(x, v)dx
i
on these coordinates. What are the xed points of this ow?
(3) (Harmonic oscillator) The harmonic oscillator (in appropriate
units) is the conservative mechanical system (R, dx dx, dU),
where U : R R is given by
U(x) :=
1
2

2
x
2
.
(a) Write the equation of motion and its general solution.
(b) Friction can be included in this model by considering the ex-
ternal force
T
_
u
d
dx
_
= dU 2kudx
(where k > 0 is a constant). Write the equation of motion of
this new mechanical system and its general solution.
(c) Generalize (a) to the n-dimensional harmonic oscillator, whose
potential energy U : R
n
R is given by
U(x
1
, . . . , x
n
) :=
1
2

2
_
_
x
1
_
2
+. . . + (x
n
)
2
_
.
(4) Consider the conservative mechanical system (R, dx dx, dU).
Show that:
(a) the ow determined by the Newton equation on TR

= R
2
is
generated by the vector eld
X = v

x
U

(x)

v
X(R
2
);
(b) the xed points of the ow are the points of the form (x
0
, 0),
where x
0
is a critical point of U;
(c) if x
0
is a maximum of U with U

(x
0
) < 0 then (x
0
, 0) is an
unstable xed point;
(d) if x
0
is a minimum of U with U

(x
0
) > 0 then (x
0
, 0) is a stable
xed point, with arbitrarily small neighborhoods formed by
periodic orbits.
1. MECHANICAL SYSTEMS 165
(e) the periods of these orbits converge to 2U

(x
0
)

1
2
as they
approach (x
0
, 0);
(f) locally, any conservative mechanical system (M, , , dU)
with dimM = 1 is of the form above.
(5) Prove Lemma 1.12. (Hint: Use the Koszul formula).
(6) Prove Lemma 1.13.
(7) If (M, , ) is a compact Riemannian manifold, it is known that
there exists a nontrivial periodic geodesic. Use this fact to show
that if M is compact then any conservative mechanical system
(M, , , dU) admits a nontrivial periodic motion.
(8) Recall that the hyperbolic plane is the upper half plane
H =
_
(x, y) R
2
[ y > 0
_
with the Riemannian metric
, =
1
y
2
(dx dx +dy dy)
(cf. Exercise 3.3.5 in Chapter 3). Use Proposition 1.14 to compute
the Christoel symbols for the Levi-Civita connection of (H, , )
in the coordinates (x, y).
(9) (Kepler problem) The Kepler problem (in appropriate units)
consists in determining the motion of a particle of mass m = 1 in
the central potential
U =
1
r
.
(a) Show that the equations of motion can be integrated to
r
2

= p

,
r
2
2
+
p

2
2r
2

1
r
= E,
where E and p

are integration constants.

(b) Use these equations to show that u =
1
r
satises the linear
ODE
d
2
u
d
2
+u =
1
p

2
.
(c) Assuming that the pericenter (i.e. the point in the particles
orbit closer to the center of attraction r = 0) occurs at = 0,
show that the equation of the particles trajectory is
r =
p

2
1 + cos
,
where
=
_
1 + 2p

2
E.
(Remark: This is the equation of a conic section with eccentricity in polar
coordinates).
166 5. GEOMETRIC MECHANICS
(d) Characterize all geodesics of R
2
(0, 0) with the Riemannian
metric
, =
1
_
x
2
+y
2
(dx dx +dy dy) .
Show that this manifold is isometric to the surface of a cone
with aperture

3
.
2. Holonomic Constraints
Many mechanical systems involve particles or systems of particles whose
positions are constrained (for example, a simple pendulum, a particle moving
on a given surface, or a rigid system of particles connected by massless rods).
To account for these we introduce the following denition:
Definition 2.1. A holonomic constraint on a mechanical system
(M, , , T) is a submanifold N M with dimN < dimM. A curve c :
I R M is said to be compatible with N if c(t) N for all t I.
Example 2.2.
(1) A particle of mass m > 0 moving in R
2
subject to a constant
gravitational acceleration g is modelled by the mechanical system
(R
2
, , , mg dy), where
v, w := mv, w
(, being the Euclidean inner product on R
2
). A simple pendu-
lum is obtained by connecting the particle to a xed pivoting point
by an ideal massless rod of length l > 0. Assuming the pivoting
point to be the origin, this corresponds to the holonomic constraint
N = (x, y) R
2
[ x
2
+y
2
= l
2

(dieomorphic to S
1
).
(2) Similarly, a particle of mass m > 0 moving in R
3
subject to a
constant gravitational acceleration g is modelled by the mechanical
system (R
3
, , , mg dz), where
v, w := mv, w
(, being the Euclidean inner product on R
3
). Requiring the
particle to move on a surface of equation z = f(x, y) yields the
holonomic constraint
N = (x, y, z) R
3
[ z = f(x, y).
(3) A system of k particles of masses m
1
, . . . , m
k
moving freely in R
3
is modelled by the mechanical system (R
3k
, , , 0), where
(v
1
, . . . , v
k
), (w
1
, . . . , w
k
) :=
k

i=1
m
i
v
i
, w
i

2. HOLONOMIC CONSTRAINTS 167

(, being the Euclidean inner product on R
3
). A rigid body
is obtained by connecting all particles by ideal massless rods, and
corresponds to the holonomic constraint
N =
_
(x
1
, . . . , x
k
) R
3k
[ |x
i
x
j
| = d
ij
for 1 i < j k
_
.
If at least three particles are not collinear, N is easily seen to be
dieomorphic to R
3
O(3).
Keeping the particles on the holonomic constraint requires an additional
external force (provided by the rods or by the surface in the examples above).
Definition 2.3. A reaction force on a mechanical system with holo-
nomic constraint (M, , , T, N) is a map ! : TN T

M satisfying
!(T
p
N) T

p
M for all p N such that, for each v TN, there is a
solution c : I R N of the generalized Newton equation

_
D c
dt
_
= (T +!)( c)
with initial condition c(0) = v.
For any holonomic constraint there exist in general innite possible
choices of reaction forces. The following denition yields a particularly use-
ful criterion for selecting reaction forces.
Definition 2.4. A reaction force in a mechanical system with holo-
nomic constraint (M, , , T, N) is said to be perfect, or to satisfy the
DAlembert principle, if

1
(!(v)) (T
p
N)

for all v T
p
N and p N.
Remark 2.5. The variation of the kinetic energy of a solution of the
generalized Newton equation is
dK
dt
=
_
D c
dt
, c
_
= T( c)( c) +!( c)( c) = T( c)( c) +

1
(!( c)), c
_
.
Therefore, a reaction force is perfect if and only if it neither creates nor
dissipates energy along any motion compatible with the constraint.
Example 2.6. In each of the examples above, requiring the reaction
force to be perfect amounts to the following assumptions.
(1) Simple pendulum: The force transmitted by the rod is purely
radial (i.e. there is no damping);
(2) Particle on a surface: The force exerted by the surface is or-
thogonal to it (i.e. the surface is frictionless);
(3) Rigid body: The cohesive forces do not dissipate energy.
The next result establishes the existence and uniqueness of perfect reac-
tion forces.
168 5. GEOMETRIC MECHANICS
Theorem 2.7. Given any mechanical system with holonomic constraint
(M, , , T, N), there exists a unique reaction force ! : TN T

M sat-
isfying the DAlembert principle. The solutions of the generalized Newton
equation

_
D c
dt
_
= (T +!)( c)
are exactly the motions of the mechanical system (N, , , T
N
), where
, is the metric induced on N by , and T
N
is the restriction of T to
N. In particular, if T = dU is conservative then T
N
= d (U[
N
).
Proof. Recall from Section 5 of Chapter 4 that if

is the Levi-Civita
connection of (M, , ) and is the Levi-Civita connection of (N, , )
then

X
Y =
_

Y
_

for all X, Y X(N), where

X,

Y are any extensions of X, Y to X(M) (as

usual, v = v

+ v

designates the unique decomposition arising from the

splitting T
p
M = T
p
N (T
p
N)

for each p N). Moreover, the second

fundamental form of N,
B(X, Y ) =

Y
X
Y =
_

Y
_

,
is well dened, and B(X, Y )
p
(T
p
N)

is a symmetric bilinear function of

X
p
, Y
p
for all p N.
Assume that a perfect reaction force ! exists; then the solutions of the
generalized Newton equation satisfy

c
c =
1
(T( c)) +
1
(!( c)).
Since by hypothesis
1
!is orthogonal to N, the component of this equation
tangent to N yields

c
c =
1
N
(T
N
( c))
(where
N
: TN T

N is the mass operator on N) as for any v TN one

has

1
(T( c))
_

, v =
1
(T( c)), v = T( c)(v) = T
N
( c)(v) =
1
N
(T
N
( c)), v.
Hence c is a motion of (N, , , T
N
).
On the other hand, the component of the generalized Newton equation
orthogonal to N yields
B( c, c) =
_

1
(T( c))
_

+
1
(!( c)).
Therefore, if ! exists then it must satisfy
(23) !(v) = (B(v, v))
_
_

1
(T(v))
_

_
for all v TN. This proves uniqueness.
To prove existence, dene ! through (23), which certainly guarantees
that
1
(!(v)) (T
p
N)

for all v T
p
N and p N. Given v TN, let
2. HOLONOMIC CONSTRAINTS 169
c : I R N be the motion of the mechanical system (N, , , T
N
) with
initial condition v. Then

c
c =
c
c +B( c, c) =
1
N
(T
N
( c)) +
_

1
(T( c))
_

+
1
(!( c))
=
_

1
(T( c))
_

+
_

1
(T( c))
_

+
1
(!( c)) =
1
(T( c)) +
1
(!( c)).

Example 2.8. To write the equation of motion of a simple pendulum

with a perfect reaction force, we parameterize the holonomic constraint N
using the map : (, ) R
2
dened by
() = (l sin, l cos )
(so that = 0 labels the stable equilibrium position). We have
d
d
=
dx
d

x
+
dy
d

y
= l cos

x
+l sin

y
,
and hence the kinetic energy of the pendulum is
K
_
v
d
d
_
=
1
2
m
_
vl cos

x
+vl sin

y
, vl cos

x
+vl sin

y
_
=
1
2
ml
2
v
2
.
On the other hand, the potential energy is given by
U(x, y) = mgy,
and hence its restriction to N has the local expression
U() = mgl cos .
Consequently the equation of motion is
d
dt
_
K
v
_
,

_
_

_
,

_
=
U

()

d
dt
_
ml
2

_
= mgl sin

=
g
l
sin .
Notice that we did not have to compute the reaction force.
Exercises 2.9.
(1) Use spherical coordinates to write the equations of motion for the
spherical pendulum of length l, i.e. a particle of mass m > 0
moving in R
3
subject to a constant gravitational acceleration g
and the holonomic constraint
N = (x, y, z) R
3
[ x
2
+y
2
+z
2
= l
2
.
Which parallels of N are possible trajectories of the particle?
170 5. GEOMETRIC MECHANICS
(2) Write the equations of motion for a particle moving on a frictionless
surface of revolution with equation z = f(r) (where r =
_
x
2
+y
2
)
under a constant gravitational acceleration g.
(3) Write and solve the equations of motion for a free dumbbell, i.e. a
system of two particles of masses m
1
and m
2
connected by a mass-
less rod of length l, moving in:
(a) R
2
;
(b) R
3
.
(Hint: Use the coordinates of the center of mass, i.e. the point along the rod at a
distance
m
2
m
1
+m
2
l from m
1
).
(4) The double pendulum of lengths l
1
, l
2
is the mechanical system
dened by two particles moving in R
2
subject to a constant gravi-
tational acceleration g and the holonomic constraint
N = (x
1
, x
2
) R
4
[ |x
1
| = l
1
and |x
1
x
2
| = l
2
.
(dieomorphic to the 2-torus T
2
).
(a) Write the equations of motion for the double pendulum using
the parameterization : (, ) (, ) N given by
(, ) = (l
1
sin, l
1
cos , l
1
sin +l
2
sin , l
1
cos l
2
cos ).
(b) Linearize the equations of motion around = = 0. Look for
solutions of the linearized equations satisfying = k, with
k R constant (normal modes). What are the periods of
the ensuing oscillations?
3. Rigid Body
Recall that a rigid body is a system of k particles of masses m
1
, . . . , m
k
connected by massless rods in such a way that their mutual distances remain
constant. If in addition we assume that a given particle is xed (at the origin,
say) then we obtain the holonomic constraint
N =
_
(x
1
, . . . , x
k
) R
3k
[ x
1
= 0 and |x
i
x
j
| = d
ij
for 1 i < j k
_
.
If at least three particles are not collinear, this manifold is dieomorphic to
O(3). In fact, if we x a point (
1
, . . . ,
k
) in N then any other point in N
is of the form (S
1
, . . . , S
k
) for a unique S O(3). A motion in N can
therefore be specied by a curve S : I R O(3). The trajectory in R
3
of the particle with mass m
i
will be given by the curve S
i
: I R R
3
,
whose velocity is

S
i
(where we use O(3) /
33
(R)

= R
9
to identify
T
S
O(3) with an appropriate subspace of /
33
(R)). Therefore the kinetic
energy of the system along the motion will be
K =
1
2
n

i=1
m
i

S
i
,

S
i
,
where , is the Euclidean inner product on R
3
.
3. RIGID BODY 171
Now O(3), and hence N, has two dieomorphic connected components,
correspondig to matrices of positive or negative determinant. Since any
motion necessarily occurs in one connected component, we can take our
conguration space to be simply SO(3). To account for continuum rigid
bodies, we make the following generalization:
Definition 3.1. A rigid body with a xed point is any mechanical
system of the form (SO(3), , , T), with
V, W :=
_
R
3
V , W dm
for all V, W T
S
SO(3) and all S SO(3), where , is the usual Eu-
clidean inner product on R
3
and m (called the mass distribution of the
reference conguration) is a positive nite measure on R
3
, not supported
on any straight line through the origin, and satisfying
_
R
3
||
2
dm < +.
Example 3.2.
(1) The rigid body composed by k particles of masses m
1
, . . . , m
k
cor-
responds to the measure
m =
k

i=1
m
i

i
,
where

i
is the Dirac delta centered at the point
i
R
3
.
(2) A continuum rigid body with (say) compactly supported integrable
density function : R
3
[0, +) is described by the measure m
dened on the Lebesgue -algebra by
m(A) :=
_
A
()d
3
.
Remark 3.3. The rotational motion of a general rigid body can in many
cases be reduced to the motion of a rigid body with a xed point (cf. Exer-
cise 3.20.2). Unless otherwise stated, from this point onwards we will take
rigid body to mean rigid body with a xed point.
Proposition 3.4. The metric , dened on SO(3) by a rigid body
is left-invariant, that is, any left translation is an isometry.
Proof. Since left multiplication by a xed matrix R SO(3) is a linear
map L
R
: /
33
(R) /
33
(R), we have (dL
R
)
S
V = RV T
RS
SO(3) for
any V T
S
SO(3). Consequently,
(dL
R
)
S
V, (dL
R
)
S
W = RV, RW =
_
R
3
RV , RW dm
=
_
R
3
V , W dm = V, W
(as R SO(3) preserves the Euclidean inner product).
172 5. GEOMETRIC MECHANICS
Therefore there exist at most as many rigid bodies as inner products on
so(3)

= R
3
, i.e., as real symmetric positive denite 3 3 matrices (cf. Ex-
ercise 1.10.4 in Chapter 3). In fact, we shall see that any rigid body can be
specied by 3 positive numbers.
Proposition 3.5. The metric , dened on SO(3) by a rigid body
is given by
V, W = tr
_
V JW
t
_
,
where
J
ij
=
_
R
3

j
dm.
Proof. We just have to notice that
V, W =
_
R
3
3

i=1
_
_
3

j=1
V
ij

j
_
_
_
3

k=1
W
ik

k
_
dm
=
3

i,j,k=1
V
ij
W
ik
_
R
3

k
dm =
3

i,j,k=1
V
ij
J
jk
W
ik
.

Proposition 3.6. If S : I R SO(3) is a curve and is the

Levi-Civita connection on (SO(3), , ) then

S, V =
_
R
3

S, V dm
for any V T
S
SO(3).
Proof. We consider rst the case in which the rigid body is non-
planar, i.e. m is not supported in any plane through the origin. In this
case, the metric , can be extended to a at metric on /
33
(R)

= R
9
by the same formula
V, W =
_
R
3
V , W dm
for all V, W T
S
/
33
(R) and all S /
33
(R). Indeed, this formula
clearly denes a symmetric 2-tensor on /
33
(R). To check positive def-
initeness, we notice that if V T
S
/
33
(R) is nonzero then its kernel is
contained on a plane through the origin. Therefore, the continuous function
V , V is positive on a set of positive measure, and hence
V, V =
_
R
3
V , V dm > 0.
This metric is easily seen to be at, as the components of the metric on the
natural coordinates of /
33
(R) are the constant coecients J
ij
. Therefore
all Christoel symbols vanish on these coordinates, and the corresponding
3. RIGID BODY 173
Levi-Civita connection

is the trivial connection. If S : I R /
33
(R)
is a curve then

S =

S.
Since , is the metric induced on SO(3) by , , we see that for any
curve S : I R SO(3) one has

S =
_

S
_

=

S

,
and hence

S, V =

, V =

S, V =
_
R
3

S, V dm
for any V T
S
SO(3).
For planar rigid bodies the formula can by obtained by a limiting pro-
cedure (cf. Exercise 3.20.3).
We can use this result to determine the geodesics of (SO(3), , ). A
remarkable shortcut (whose precise nature will be discussed in Section 5)
can be obtained by introducing the following quantity.
Definition 3.7. The angular momentum of a rigid body whose mo-
tion is described by S : I R SO(3) is the vector
p(t) :=
_
R
3
_
(S(t)) (

S(t))
_
dm
(where is the usual cross product on R
3
).
Theorem 3.8. If S : I R SO(3) is a geodesic of (SO(3), , )
then p(t) is constant.
Proof. We have
p =
_
R
3
_
(

S) (

S) + (S) (

S)
_
dm =
_
R
3
_
(S) (

S)
_
dm.
Take any v R
3
. Then
Sv, p =
_
Sv,
_
R
3
_
(S) (

S)
_
dm
_
=
_
R
3
_
Sv, (S) (

S)
_
dm
=
_
R
3
_

S, (Sv) (S)
_
dm =
_
R
3
_

S, S(v )
_
dm,
where we have used the invariance of , det(, , ) under even permu-
tations of its arguments and the fact that the cross product is equivariant
under multiplication by S SO(3).
To complete the proof we will need the following lemma, whose proof is
left as an exercise.
174 5. GEOMETRIC MECHANICS
Lemma 3.9. There exists a linear isomorphism : so(3) R
3
such that
A = (A)
for all R
3
and A so(3). Moreover, ([A, B]) = (A) (B) for all
A, B so(3) (that is, is a Lie algebra isomorphism between so(3) and
(R
3
, )).
Returning to the proof, let V so(3) be such that (V ) = v. Then
SV T
S
SO(3) and
Sv, p =
_
R
3
_

S, SV
_
dm =

S, SV = 0
(as S : I R SO(3) is a geodesic). Since v R
3
is arbitrary, we see that
p = 0 along the motion.
If S : I R SO(3) is a curve then

S = SA for some A so(3). Let
us dene := (A). Then
p =
_
R
3
[(S) (SA)] dm =
_
R
3
S [ (A)] dm
= S
_
R
3
[ ( )] dm.
This suggests the following denition.
Definition 3.10. The linear operator I : R
3
R
3
dened as
I(v) :=
_
R
3
[ (v )]dm
is called the rigid bodys moment of inertia tensor.
Proposition 3.11. The moment of inertia tensor of any given rigid
body is a symmetric positive denite linear operator, and the corresponding
kinetic energy map K : TSO(3) R is given by
K(V ) =
1
2
V, V =
1
2
SA, SA =
1
2
I, ,
for all V T
S
SO(3) and all S SO(3), where V = SA and = (A).
Proof. We start by checking that I is symmetric:
Iv, w =
_
R
3
[ (v )] dm, w =
_
R
3
(v ), w dm
=
_
R
3
v , w dm = v, Iw.
In particular we have
I, =
_
R
3
, dm =
_
R
3
A, A dm
=
_
R
3
SA, SA dm = 2K(V ).
3. RIGID BODY 175
The positive deniteness of I is an immediate consequence of this formula.

Corollary 3.12. Given any rigid body there exist three positive num-
bers I
1
, I
2
, I
3
(principal moments of inertia) and an orthonormal basis
of R
3
, e
1
, e
2
, e
3
(principal axes), such that Ie
i
= I
i
e
i
(i = 1, 2, 3).
The principal moments of inertia are the three positive numbers which
completely specify the rigid body (as they determine the inertia tensor,
which in turn yields the kinetic energy). To compute these numbers we must
compute the eigenvalues of a matrix representation of the inertia tensor.
Proposition 3.13. The matrix representation of the inertia tensor in
the canonical basis of R
3
is
_
_
_
_
_
_
_
R
3
(y
2
+z
2
) dm
_
R
3
xy dm
_
R
3
xz dm

_
R
3
xy dm
_
R
3
(x
2
+z
2
) dm
_
R
3
yz dm

_
R
3
xz dm
_
R
3
yz dm
_
R
3
(x
2
+y
2
) dm
_
_
_
_
_
_
.
Proof. Let u
1
, u
2
, u
3
be the canonical basis of R
3
. Then
I
ij
= Iu
i
, u
j
=
_
R
3
(u
i
), u
j
dm.
Using the vector identity
u (v w) = u, wv u, vw
for all u, v, w R
3
, we have
I
ij
=
_
R
3

||
2
u
i
, u
i
, u
j
_
dm =
_
R
3
_
||
2

ij

i

j
_
dm.

We can now write the equations for the geodesics of (SO(3), , ), that
is, the equations of motion of a rigid body in the absence of external forces.
This mechanical system is commonly known as the Euler top.
Proposition 3.14. The equations of motion of the Euler top are given
by the Euler equations
I

= (I) .
Proof. We just have to notice that
p = SI.
Therefore
0 = p =

SI +SI

= SAI +SI

= S
_
(I) +I

_
.

176 5. GEOMETRIC MECHANICS

Remark 3.15. Any point R
3
in the rigid body traverses a curve
x(t) = S(t) with velocity
x =

S = SA = S( ) = (S) (S) = (S) x.
Therefore := S is the rigid bodys instantaneous angular velocity:
at each instant, the rigid body rotates about the axis determined by with
angular speed ||. Consequently, is the angular velocity as seen in the
(accelerated) rigid bodys rest frame.
In the basis e
1
, e
2
, e
3
of the principal axes, the Euler equations are
written
_

_
I
1

1
= (I
2
I
3
)
2

3
I
2

2
= (I
3
I
1
)
3

1
I
3

3
= (I
1
I
2
)
1

2
.
Since I is positive denite (hence invertible), we can change variables to
P := I. Notice that p = SP, i.e. P is the (constant) angular momentum
vector as seen in rigid bodys rest frame. In these new variables, the Euler
equations are written

P = P
_
I
1
P
_
.
In the basis e
1
, e
2
, e
3
of the principal axes, these are
_

P
1
=
_
1
I
3

1
I
2
_
P
2
P
3

P
2
=
_
1
I
1

1
I
3
_
P
3
P
1

P
3
=
_
1
I
2

1
I
1
_
P
1
P
2
.
Proposition 3.16. If I
1
> I
2
> I
3
, the stationary points of the Euler
equations are given by
P = e
i
( R, i = 1, 2, 3),
and are stable for i = 1, 3 and unstable for i = 2.
Proof. Since there are no external forces, the kinetic energy K, given
by
2K = I, =

P, I
1
P
_
=
_
P
1
_
2
I
1
+
_
P
2
_
2
I
2
+
_
P
3
_
2
I
3
,
is conserved. This means that the ow dened by the Euler equations is
along ellipsoids with semiaxes of lengths

2KI
1
>

2KI
2
>

2KI
3
. On
the other hand, since p is constant along the motion, we have a second
conserved quantity,
|p|
2
= |P|
2
=
_
P
1
_
2
+
_
P
2
_
2
+
_
P
3
_
2
.
Therefore the ow is along spheres. The integral curves on a particular
sphere can be found by intersecting it with the ellipsoids corresponding to
dierent values of K, as shown in Figure 2.
3. RIGID BODY 177
e
1
e
2
e
3
Figure 1. Integral curves of the Euler equations.
Remark 3.17. Since = I
1
P, Proposition 3.16 is still true if we
replace P with . The equilibrium points represent rotations about the
principal axes with constant angular speed, as they satisfy = I
i
P, and
hence = I
i
p is constant. If the rigid body is placed in a rotation state close
to a rotation about the axes e
1
or e
3
, P will remain close to these axes, and
hence Se
1
or Se
3
will remain close to the xed vector p. On the other hand,
if the rigid body is placed in a rotation state close to a rotation about the
axis e
2
, then P will drift away from e
2
(approaching e
2
before returning
to e
2
), and hence Se
2
will drift away from the xed vector p (approaching
p before returning to p). This can be illustrated by throwing a rigid body
(say a brick) in the air, as its rotational motion about the center of mass is
that of a rigid body with a xed point (cf. Exercise 3.20.2). When rotating
about the smaller or the larger axis (i.e. the principal axes corresponding to
the larger or the smaller moments of inertia cf. Exercise 3.20.6) it performs
a stable rotation, but when rotating about the middle axis it ips in midair.
If the rigid body is not free, one must use parameterizations of SO(3).
Definition 3.18. The Euler angles correspond to the local coordinates
(, , ) : SO(3) (0, ) (0, 2) (0, 2) dened by
S(, , ) =
_
_
cos sin 0
sin cos 0
0 0 1
_
_
_
_
1 0 0
0 cos sin
0 sin cos
_
_
_
_
cos sin 0
sin cos 0
0 0 1
_
_
.
178 5. GEOMETRIC MECHANICS
The geometric interpretation of the Euler angles is sketched in Figure 2:
if the rotation carries the canonical basis e
x
, e
y
, e
z
to a new orthonormal
basis e
1
, e
2
, e
3
, then is the angle between e
3
and e
z
, is the angle
between the line of intersection of the planes spanned by e
1
, e
2
and e
x
, e
y

(called the nodal line) and the x-axis, and is the angle between e
1
and
the nodal line.
e
1
e
2
e
3
e
x
e
y
e
z

nodal line
horizontal plane
Figure 2. Euler angles.
The general expression of the kinetic energy in the local coordinates of
TSO(3) associated to the Euler angles is quite complicated; here we present
it only in the simpler case I
1
= I
2
.
Proposition 3.19. If I
1
= I
2
then the kinetic energy of a rigid body in
the local coordinates (, , , v

, v

, v

) of TSO(3) is given by
K =
I
1
2
_
_
v

_
2
+ (v

)
2
sin
2

_
+
I
3
2
_
v

+v

cos
_
2
.
A famous model which can be studied using this expression is the so-
called Lagrange top, corresponding to an axisymmetric rigid body in a
constant gravity eld g. The potential energy for the corresponding me-
chanical system is
U := g
_
R
3
S, e
z
dm = MgS, e
z
,
3. RIGID BODY 179
where M = m(R
3
) is the total mass and
:=
1
M
_
R
3
dm
is the position of the center of mass in the rigid bodys frame. By axisym-
metry, the center of mass satises = le
3
for some l R, and so
U = Mgl cos .
Exercises 3.20.
(1) Show that the bilinear form, dened on SO(3) by a rigid body
is indeed a Riemannian metric.
(2) A general rigid body (i.e. with no xed points) is any mechanical
system of the form (R
3
SO(3), , , T), with
(v, V ), (w, W) :=
_
R
3
v +V , w +W dm
for all (v, V ), (w, W) T
(x,S)
R
3
SO(3) and (x, S) R
3
SO(3),
where , is the usual Euclidean inner product on R
3
and m is
a positive nite measure on R
3
not supported on any straight line
and satisfying
_
R
3
||
2
dm < +.
(a) Show that one can always translate m in such a way that
_
R
3
dm = 0
(i.e. the center of mass of the reference conguration is placed
at the origin).
(b) Show that for this choice the kinetic energy of the rigid body
is
K(v, V ) =
1
2
Mv, v +
1
2
V, V ,
where M = m(R
3
) is the total mass of the rigid body and
, is the metric for the rigid body (with a xed point)
determined by m.
(c) Assume that there exists a dierentiable function F : R
3
R
3
such that
T(x, S, v, V )(w, W) =
_
R
3
F(x +S), w +W dm.
Show that, if
_
R
3
(S) F(x +S) dm = 0
for all (x, S) R
3
SO(3), then the projection of any motion
on SO(3) is a geodesic of (SO(3), , ).
(d) Describe the motion of a rigid body falling in a constant grav-
itational eld, for which F = ge
z
is constant.
180 5. GEOMETRIC MECHANICS
(3) Prove Proposition 3.6 for a planar rigid body. (Hint: Include the planar
rigid body in a smooth one-parameter family of non-planar rigid bodies).
(4) Prove Lemma 3.9.
(5) Show that I
1
I
2
+ I
3
(and cyclic permutations). When is I
1
=
I
2
+I
3
?
(6) Determine the principal axes and the corresponding principal mo-
ments of inertia of:
(a) a homogeneous rectangular parallelepiped with mass M, sides
2a, 2b, 2c R
+
and centered at the origin;
(b) a homogeneous (solid) ellipsoid with mass M, semiaxes a, b, c
R
+
and centered at the origin. (Hint: Use the coordinate change
(x, y, z) = (au, bv, cw)).
(7) A symmetry of a rigid body is an isometry S O(3) which pre-
serves the mass distribution (i.e. m(SA) = m(A) for any measur-
able set A R
3
). Show that:
(a) SIS
t
= I, where I is the matrix representation of the inertia
tensor;
(b) if S is a reection in a plane then there exists a principal axis
orthogonal to the reection plane;
(c) if S is a nontrivial rotation about an axis then that axis is
principal;
(d) if moreover the rotation is not by then all axes orthogonal
to the rotation axis are principal.
(8) Consider a rigid body satisfying I
1
= I
2
. Use the Euler equations
to show that:
(a) the angular velocity satises
=
1
I
1
p ;
(b) if I
1
= I
2
= I
3
then the rigid body rotates about a xed axis
with constant angular speed (i.e. is constant);
(c) if I
1
= I
2
,= I
3
then precesses (i.e. rotates) about p with
angular velocity

pr
:=
p
I
1
.
(9) Many asteroids have irregular shapes, and hence satisfy I
1
< I
2
<
I
3
. To a very good approximation, their rotational motion about
the center of mass is described by the Euler equations. Over very
long periods of time, however, their small interactions with the Sun
and other planetary bodies tend to decrease their kinetic energy
while conserving their angular momentum. Which rotation state
do asteroids approach?
3. RIGID BODY 181
(10) Due to its rotation, the Earth is not a perfect sphere, but an oblate
spheroid; therefore its moments of inertia are not quite equal, sat-
isfying approximately
I
1
= I
2
,= I
3
;
I
3
I
1
I
1

1
306
.
The Earths rotation axis is very close to e
3
, but precesses around
it (Chandler precession). Find the period of this precession (in
the Earths frame).
(11) Consider a rigid body whose motion is described by the curve
S : R SO(3), and let be the corresponding angular veloc-
ity. Consider a particle with mass m whose motion in the rigid
bodys frame is given by the curve : R R
3
. Let f be the
external force on the particle, so that its equation of motion is
m
d
2
dt
2
(S) = f.
(a) Show that the equation of motion can be written as
m

= F m ( ) 2m

m

where f = SF. (The terms following F are the so-called iner-
tial forces, and are known, respectively, as the centrifugal
force, the Coriolis force and the Euler force).
(b) Show that if the rigid body is a homogeneous sphere rotating
freely (like the Earth, for instance) then the Euler force van-
ishes. Why must a long range gun in the Northern hemisphere
be aimed at the left of the target?
(12) (Poinsot Theorem) The inertia ellipsoid of a rigid body with
moment of inertia tensor I is the set
E = R
3
[ I, = 1.
Show that the inertia ellipsoid of a freely moving rigid body rolls
without slipping on a xed plane orthogonal to p (that is, the
contact point has zero velocity at each instant). (Hint: Show that
any point S(t)(t) where the ellipsoid is tangent to a plane orthogonal to p satises
S(t)(t) =
1

2K
(t)).
(13) Prove Proposition 3.19. (Hint: Notice that symmetry demands that the ex-
pression for K must not depend neither on nor on ).
(14) Consider the Lagrange top.
(a) Write the equations of motion and determine the equilibrium
points.
182 5. GEOMETRIC MECHANICS
(b) Show that there exist solutions such that , and

are con-
stant, which in the limit [ [ [

[ (fast top) satisfy

Mgl
I
3

.
(15) (Precession of the equinoxes) Due to its rotation, the Earth is not
a perfect sphere, but an oblate ellipsoid; therefore its moments of
inertia are not quite equal, satisfying approximately
I
1
= I
2
,= I
3
;
I
3
I
1
I
1

1
306
(cf. Exercise 10). As a consequence, the combined gravitational
attraction of the Moon and the Sun disturbs the Earths rotation
motion. This perturbation can be approximately modelled by the
potential energy U : SO(3) R given in the Euler angles (, , )
by
U =

2
2
(I
3
I
1
) cos
2
,
where
2

168 days.
(a) Write the equations of motion and determine the equilibrium
points.
(b) Show that there exist solutions such that , and

are con-
stant, which in the limit [ [ [

[ (as is the case with the
Earth) satisfy

2
(I
3
I
1
) cos
I
3

.
Given that for the Earth 23

, determine the approximate

value of the period of (t).
(16) (Pseudo-rigid body) Recall that the (non planar) rigid body metric
is the restriction to SO(3) of the at metric on GL(3) given by
V, W = tr(V JW
t
),
where
J
ij
=
_
R
3

j
dm.
(a) What are the geodesics of the Levi-Civita connection for this
metric? Is (GL(3), , ) geodesically complete?
(b) The Euler equation and the continuity equation for an
incompressible uid with velocity eld u : R R
3
R
3
and
3. RIGID BODY 183
pressure p : R R
3
R are
u
t
+ (u )u = p,
u = 0,
where
=
_

x
1
,

x
2
,

x
3
_
is the usual operator of vector calculus.
Given a geodesic S : R GL(3), we dene
x(t, ) = S(t)
u(t, x) =

S(t) =

S(t)S
1
(t)x.
Show that the velocity eld u satises the Euler equation (with
p = 0), but not the continuity equation.
(c) Let f : GL(3) R be given by f(S) = det S. Show that
f
S
ij
= cof(S)
ij
(where cof(S) is the matrix of the cofactors of S), and conse-
quently
df
dt
= (det S) tr(

SS
1
).
So the continuity equation is satised if we impose the con-
straint det S(t) = 1.
(d) Show that the holonomic constraint SL(3) GL(3) satises
the DAlembert Principle if and only if
_

S
_
= (t)df
det S = 1
Assuming that J is invertible, show that the equation of mo-
tion can be rewritten as

S =
_
S
1
_
t
J
1
.
(e) Show that the geodesics of (SL(3), , ) yield solutions of
the Euler equation with
p =

2
x
t
_
S
1
_
t
J
1
S
1
x
which also satisfy the continuity equation.
(Remark: More generally, it is possible to interpret the Euler equation on an
open set U R
n
as a mechanical system on the group of dieomorphisms of
U (which is an innite dimensional Lie group); the continuity equation imposes
the holonomic constraint corresponding to the subgroup of volume-preserving
dieomorphisms, and the pressure is the perfect reaction force associated to this
constraint).
184 5. GEOMETRIC MECHANICS
4. Non-Holonomic Constraints
Some mechanical systems are subject to constraints which force the mo-
tions to proceed in certain admissible directions. To handle such constraints
we must rst introduce the corresponding geometric concept.
Definition 4.1. A distribution of dimension m on a dierentiable
manifold M is a choice of an m-dimensional subspace
p
T
p
M for each
p M. The distribution is said to be dierentiable if for all p M there
exists a neighborhood U p and vector elds X
1
, . . . , X
m
X(U) such that

q
= span
_
(X
1
)
q
, . . . , (X
m
)
q
_
for all q U.
Equivalently, is dierentiable if for all p M there exists a neighbor-
hood U p and 1-forms
1
, . . . ,
nm

1
(U) such that

q
= ker
_

1
_
q
ker
_

nm
_
q
for all p U (cf. Exercise 4.15.1). We will assume from this point on that
all distributions are dierentiable.
Definition 4.2. A non-holonomic constraint on a mechanical sys-
tem (M, , , T) is a distribution on M. A curve c : I R M is said
to be compatible with if c(t)
c(t)
for all t I.
Example 4.3.
(1) (Wheel rolling without slipping) Consider a vertical wheel of radius
R rolling without slipping on a plane. Assuming that the motion
takes place along a straight line, we can parameterize any position
of the wheel by the position x of the contact point and the angle
between a xed radius of the wheel and the radius containing
the contact point (cf. Figure 3); hence the conguration space is
R S
1
.

x
O
Figure 3. Wheel rolling without slipping.
If the wheel is to rotate without slipping, we must require that
x = R

along any motion; this is equivalent to requiring that the

4. NON-HOLONOMIC CONSTRAINTS 185
motion be compatible with the distribution dened on R S
1
by
the vector eld
X = R

x
+

,
or, equivalently, by the kernel of the 1-form
= dx Rd.
(2) (Ice skate) A simple model for an ice skate is provided by a line seg-
ment which can either move along itself or rotate about its middle
point. The position of the skate can be specied by the Cartesian
coordinates (x, y) of the middle point and the angle between the
skate and the x-axis (cf. Figure 4); hence the conguration space
is R
2
S
1
.
x
y
Figure 4. Ice skate.
If the skate can only move along itself, we must require that
( x, y) be proportional to (cos , sin); this is equivalent to requiring
that the motion be compatible with the distribution dened on
R
2
S
1
by the vector elds
X = cos

x
+ sin

y
, Y =

,
or, equivalently, by the kernel of the 1-form
= sin dx + cos dy.
One may wonder whether there exists any connection between holonomic
and non-holonomic constraints. To answer this question, we must make a
small digression.
Definition 4.4. A foliation of dimension m on an n-dimensional dif-
ferentiable manifold M is a family T = L

A
of subsets of M (called
leaves) satisfying:
(1) M =
A
L

;
(2) L

= if ,= ;
186 5. GEOMETRIC MECHANICS
(3) each leaf L

is pathwise connected, that is, if p, q L

then
there exists a continuous curve c : [0, 1] L

such that c(0) = p

and c(1) = q;
(4) for each point p M there exists an open set U p and local
coordinates (x
1
, . . . , x
n
) : U R
n
such that the connected compo-
nents of the intersections of the leaves with U are the level sets of
(x
m+1
, . . . , x
n
) : U R
nm
.
Remark 4.5. The coordinates (x
1
, . . . , x
m
) provide local coordinates on
the leaves, which are therefore images of injective immersions. In particular,
the leaves have well dened m-dimensional tangent spaces at each point,
and consequently any foliation of dimension m denes an m-dimensional
distribution. Notice however that in general the leaves are not (embedded)
submanifolds of M (cf. Exercise 4.15.2).
Definition 4.6. An m-dimensional distribution on a dierential man-
ifold M is said to be integrable if there exists an m-dimensional foliation
T = L

A
on M such that

p
= T
p
L
p
for all p M, where L
p
is the leaf containing p. The leaves of T are called
the integral submanifolds of the distribution.
Integrable distributions are particularly simple. For instance, the set
of points q M which are accessible from a given point p M by a curve
compatible with the distribution is simply the leaf L
p
through p. If the leaves
are embedded submanifolds, then an integrable non-holonomic restriction
reduces to a family of holonomic restrictions. For this reason, an integrable
distribution is sometimes called a semi-holonomic constraint, whereas a
non-integrable distribution is called a true non-holonomic constraint.
It is therefore important to have a criterion for identifying integrable
distributions.
Definition 4.7. Let be a distribution on a dierentiable manifold M.
A vector eld X X(M) is said to be compatible with if X
p

p
for
all p M. We denote by X() the linear subspace of X(M) formed by all
vector elds which are compatible with .
Theorem 4.8. (Frobenius) A distribution is integrable if and only if
X, Y X() [X, Y ] X().
The proof of this theorem can be found in [War83] (see also Exer-
cise 4.15.3). If is locally given by m vector elds X
1
, . . . , X
m
, then to
check integrability it suces to check whether [X
i
, X
j
] =

m
k=1
C
k
ij
X
k
for
locally dened functions C
k
ij
(cf. Exercise 4.15.4). The next proposition
(whose proof is left as an exercise) provides an alternative criterion.
4. NON-HOLONOMIC CONSTRAINTS 187
Proposition 4.9. An m-dimensional distribution on an n-manifold
M is integrable if and only if
d
i

1

nm
= 0 (i = 1, . . . , n m)
for all locally dened sets of dierential forms
1
, . . . ,
nm
whose kernels
determine .
Since the condition of the Frobenius Theorem is local, this condition
needs to be checked only for sets of dierential forms whose domains form
an open cover of M.
Example 4.10.
(1) (Wheel rolling without slipping) Recall that in this case the con-
straint is given by the kernel of the 1-form
= dx Rd.
Since d = 0, we see that this is a semi-holonomic constraint,
corresponding to an integrable distribution. The leaves of the dis-
tribution are the submanifolds with equation x = x
0
+R.
(2) (Ice skate) Recall that in this case the constraint is given by the
kernel of the 1-form
= sin dx + cos dy.
Since
d = (cos d dx sin d dy) (sin dx + cos dy)
= d dx dy ,= 0,
we see that this is a true non-holonomic constraint.
In a Riemannian manifold (M, , ), any distribution determines an
orthogonal distribution

, given by

p
= (
p
)

T
p
M.
Hence we have two orthogonal projections

: TM and

: TM

.
The set of all external forces T : TM T

M satisfying
T (v) = T
_
v

_
for all v TM is denoted by F

.
Definition 4.11. A reaction force on a mechanical system with non-
holonomic constraints (M, , , T, ) is a force ! F

such that the solu-

tions of the generalized Newton equation

_
D c
dt
_
= (T +!)( c)
with initial condition in are compatible with . The reaction force is said
to be perfect, or to satisfy the DAlembert principle, if

1
(!(v))

p
188 5. GEOMETRIC MECHANICS
for all v T
p
M, p M.
Just like in the holonomic case, a reaction force is perfect if and only if
it neither creates nor dissipates energy along any motion compatible with
the constraint.
Theorem 4.12. Given a mechanical system with non-holonomic con-
straints (M, , , T, ), there exists a unique reaction force ! F

satisfy-
ing the DAlembert principle.
Proof. We dene the second fundamental form of the distribution
at a point p M as the map B : T
p
M
p

p
given by
B(v, w) = (
X
Y )

,
where X X(M) and Y X() satisfy X
p
= v and Y
p
= w. To check that
B is well dened, let Z
1
, . . . , Z
n
be a local orthonormal frame such that
Z
1
, . . . , Z
m
is a basis for and Z
m+1
, . . . , Z
n
is a basis for

. Then

X
Y =
X
_
m

i=1
Y
i
Z
i
_
=
m

i=1
_
_
(X Y
i
)Z
i
+
n

j,k=1

k
ji
X
j
Y
i
Z
k
_
_
,
where the functions
k
ij
are dened by

Z
i
Z
j
=
n

k=1

k
ij
Z
k
.
Consequently,
B(v, w) = (
X
Y )

=
n

i=1
m

j=1
n

k=m+1

k
ij
X
i
Y
j
Z
k
depends only on v = X
p
and w = Y
p
, and is a bilinear map. Incidentally,
the restriction of B to
p

p
is symmetric for all p M if and only if

k
ij
=
k
ji

Z
i
Z
j
, Z
k
=
Z
j
Z
i
, Z
k
[Z
i
, Z
j
], Z
k
= 0
for all i, j = 1, . . . , m and all k = m + 1, . . . , n, i.e. if and only if is
integrable. In this case, B is, of course, the second fundamental form of the
leaves.
Let us assume that ! exists. Then any motion c : I R M with
initial condition on is compatible with and satises
D c
dt
=
1
(T( c)) +
1
(!( c)).
The projection of this equation on

yields
B( c, c) =
_

1
(T( c))
_

+
1
(!( c))
(recall that
D c
dt
=
c
c). Therefore, if ! exists then it must be given by
!(v) = (B(v, v))
_
_

1
(T(v))
_

_
4. NON-HOLONOMIC CONSTRAINTS 189
for any v , and by !(v) = !
_
v

_
for any v TM (as ! F

). This
proves uniqueness of !.
To prove existence, we just have to show that for this choice of ! the
solutions of the generalized Newton equation with initial condition on are
compatible with . Consider the system
_
c =

m
i=1
v
i
Z
i
D c
dt
=
1
(T( c))
_

1
(T( c))
_

+B( c, c)
. (24)
When written in local coordinates, this is a system of rst order ODEs with
n+m unknowns x
1
(t), . . . , x
n
(t), v
1
(t), . . . , v
m
(t). Since the second equation
is just
D c
dt
=
_

1
(T( c))
_

+
_
D c
dt
_

_
D c
dt
_

=
_

1
(T( c))
_

,
we see that this equation has only m nonvanishing components in the local
frame Z
1
, . . . , Z
n
. Therefore, (24) is a system of (n+m) rst order ODEs
on n+m unknowns, and has a unique local solution for any initial condition.
If c(0)
c(0)
, we can always choose v
1
(0), . . . , v
m
(0) such that
c(0) =
m

i=1
v
i
(0) (Z
i
)
c(0)
.
The solution of (24) with initial condition (x
1
(0), . . . , x
n
(0), v
1
(0), . . . , v
m
(0))
must then, by uniqueness, be the solution of
D c
dt
=
1
(T( c)) +
1
(!( c))
with initial condition c(0). On the other hand, it is, by construction, com-
patible with .
Example 4.13. (Wheel rolling without slipping) Recall that in this case
the constraint is given by the kernel of the 1-form
= dx Rd.
Since
1
!is orthogonal to the constraint for a perfect reaction force !, the
constraint must be in the kernel of !, and hence ! = for some smooth
function : TM R.
If the kinetic energy of the wheel is
K =
M
2
(v
x
)
2
+
I
2
_
v

_
2
then

_
D c
dt
_
= M xdx +I

d.
Just to make things more interesting, consider a constant gravitational
acceleration g and suppose that the plane on which the wheel rolls makes
190 5. GEOMETRIC MECHANICS
an angle with respect to the horizontal, so that there exists a conservative
force with potential energy
U = Mgxsin .
The equation of motion is therefore

_
D c
dt
_
= dU +!( c) M xdx +I

d = Mg sin dx +dx Rd.

The motion of the wheel will be given by a solution of this equation which
also satises the constraint equation, i.e. a solution of the system of ODEs
_

_
M x = Mg sin +
I

= R
x = R

.
This system is easily solved to yield
_

_
x(t) = x
0
+v
0
t

2
t
2
(t) =
0
+
v
0
R
t

2R
t
2
=
I
R
2
where
=
g sin
1 +
I
MR
2
and x
0
, v
0
,
0
are integration constants.
Physically, the reaction force can be interpreted as a friction force ex-
erted by the plane on the wheel. This force opposes the translational motion
of the wheel but accelerates its spinning motion. Therefore, contrary to in-
tuition, there is no dissipation of energy: all the translational kinetic energy
lost by the wheel is restored as rotational kinetic energy.
A perfect reaction force guarantees, as one would expect, conservation
of energy.
Theorem 4.14. Let (M, , , dU, ) be a conservative mechanical sys-
tem with non-holonomic constraints. If the reaction force ! satises the
DAlembert principle then the mechanical energy E := K + U is constant
along any motion with initial condition in .
Exercises 4.15.
(1) Show that an m-dimensional distribution on an n-manifold M is
dierentiable if and only if for all p M there exists a neighborhood
U p and 1-forms
1
, . . . ,
nm

1
(U) such that

q
= ker
_

1
_
q
ker
_

nm
_
q
for all q U.
4. NON-HOLONOMIC CONSTRAINTS 191
(2) Show that the foliation
T =
_
(x, y) R
2
[ y =

2x +
_
R
of R
2
induces a foliation T

on T
2
= R
2
/Z
2
whose leaves are not
(embedded) submanifolds.
(3) Let be an integrable distribution. Show that X, Y X()
[X, Y ] X().
(4) Using the Frobenius Theorem show that an m-dimensional distri-
bution is integrable if and only if each local basis of vector elds
X
1
, . . . , X
m
satises [X
i
, X
j
] =

m
k=1
C
k
ij
X
k
for locally dened
functions C
k
ij
. (Remark: Since the condition of the Frobenius Theorem is local,
this condition needs to be checked only for local bases whose domains form an open
cover of M).
(5) Prove Proposition 4.9. (Hint: Recall from Exercise 3.8.2 in Chapter 2 that
d(X, Y ) = X (Y ) Y (X) ([X, Y ]) for any
1
(M) and X, Y X(M)).
(6) Let M be an n-dimensional dierentiable manifold with an ane
connection . Show that the parallel transport of vectors is deter-
mined by a distribution on TM, which is integrable if and only
if the curvature of vanishes.
(7) Prove Theorem 4.14.
(8) (Ice skate) Recall that our model for an ice skate is given by the
non-holonomic constraint dened on R
2
S
1
by the kernel of the
1-form = sin dx + cos dy.
(a) Show that the ice skate can access all points in the congu-
ration space: given two points p, q R
2
S
1
there exists a
piecewise smooth curve c : [0, 1] R
2
S
1
compatible with
such that c(0) = p and c(1) = q. Why does this show that
is non-integrable?
(b) Assuming that the kinetic energy of the skate is
K =
M
2
_
(v
x
)
2
+ (v
y
)
2
_
+
I
2
_
v

_
2
and that the reaction force is perfect, show that the skate
moves with constant speed along straight lines or circles. What
is the physical interpretation of the reaction force?
(c) Determine the motion of the skate moving on an inclined plane,
i.e. subject to a potential energy U = Mg sin x.
(9) Consider a vertical wheel of radius R moving on a plane.
(a) Show that the non-holonomic constraint corresponding to the
condition of rolling without slipping or sliding is the distribu-
tion determined on the conguration space R
2
S
1
S
1
by
the 1-forms

1
= dx Rcos d,
2
= dy Rsin d,
192 5. GEOMETRIC MECHANICS
where (x, y, , ) are the local coordinates indicated in Fig-
ure 5.
(b) Assuming that the kinetic energy of the wheel is
K =
M
2
_
(v
x
)
2
+ (v
y
)
2
_
+
I
2
(v

)
2
+
J
2
(v

)
2
and that the reaction force is perfect, show that the wheel
moves with constant speed along straight lines or circles. What
is the physical interpretation of the reaction force?
(c) Determine the motion of the vertical wheel moving on an in-
clined plane, i.e. subject to a potential energy U = Mg sin x.

x
y
z
O
R
Figure 5. Vertical wheel on a plane.
(10) Consider a sphere of radius R and mass M rolling without slipping
on a plane.
(a) Show that the condition of rolling without slipping is
x = R
y
, y = R
x
,
where (x, y) are the Cartesian coordinates of the contact point
on the plane and is the angular velocity of the sphere.
5. LAGRANGIAN MECHANICS 193
(b) Show that if the spheres mass is symmetrically distributed
then its kinetic energy is
K =
M
2
_
x
2
+ y
2
_
+
I
2
, ,
where I is the spheres moment of inertia and , is the Eu-
clidean inner product.
(c) Using as coordinates on the bers of TSO(3), show that
D c
dt
= x

x
+ y

y
+ .
(Hint: Recall from Exercise 4.8.3 in Chapter 3 that the integral curves of left-
invariant vector elds on a Lie group with a bi-invariant metric are geodesics).
(d) Since we are identifying the bers of TSO(3) with R
3
, we can
use the Euclidean inner product to also identify the bers of
T

SO(3) with R
3
. Show that under this identication the non-
holonomic constraint yielding the condition of rolling without
slipping is the distribution determined by the kernels of the
1-forms

x
:= dx Re
y
,
y
:= dy +Re
x
(where e
x
, e
y
, e
z
is the canonical basis of R
3
). Is this distri-
bution integrable? (Hint: Show that any two points of R
2
SO(3) can
be connected by a piecewise smooth curve compatible with the distribution).
(e) Show that the sphere moves along straight lines with constant
speed and constant angular velocity orthogonal to its motion.
(f) Determine the motion of the sphere moving on an inclined
plane, i.e. subject to a potential energy U = Mg sin x.
(11) (The golfer dilemma) Show that the center of a symmetric sphere
of radius R, mass M and moment of inertia I rolling without slip-
ping inside a vertical cylinder of radius R+a moves with constant
angular velocity with respect to the axis of the cylinder while oscil-
lating up and down with a frequency
_
I
I+MR
2
times the frequency
of the angular motion.
5. Lagrangian Mechanics
Let M be a dierentiable manifold, p, q M and a, b R such that
a < b. Let us denote by ( the set of dierentiable curves c : [a, b] M such
that c(a) = p and c(b) = q.
Definition 5.1. A Lagrangian function on M is a dierentiable map
L : TM R. The action determined by L on ( is the map S : ( R
given by
S(c) :=
_
b
a
L( c(t))dt.
194 5. GEOMETRIC MECHANICS
We can look for the global minima (or maxima) of the action by consid-
ering curves on (.
Definition 5.2. A variation of c ( is a map : (, ) ( (for
some > 0) such that (0) = c and the map : (, ) [a, b] M given
by (s, t) := (s)(t) is dierentiable. The curve c is said to be a critical
point of the action if
d
ds[
s=0
S((s)) = 0
for any variation of c.
Notice that the global minima (or maxima) of S must certainly be at-
tained at critical points. However, as it is usually the case, a critical point
is not necessarily a point of minimum (or maximum). It turns out that the
critical points of the action are solutions of second order ODEs.
Theorem 5.3. The curve c ( is a critical point of the action de-
termined by the Lagrangian L : TM R if and only if it satises the
Euler-Lagrange equations
d
dt
_
L
v
i
(x(t), x(t))
_

L
x
i
(x(t), x(t)) = 0 (i = 1, . . . , n)
for any local chart (x
1
, . . . , x
n
) on M, where (x
1
, . . . , x
n
, v
1
, . . . , v
n
) is the
corresponding local chart induced on TM.
Proof. Assume rst that the image of c is contained on the domain of
a local chart (x
1
, . . . , x
n
). Let : (, ) ( be a variation of c. Setting
x(s, t) := (x )(s, t), we have
S((s)) =
_
b
a
L
_
x(s, t),
x
t
(s, t)
_
dt,
and hence
d
ds[
s=0
S((s)) =
_
b
a
n

i=1
L
x
i
_
x(0, t),
x
t
(0, t)
_
x
i
s
(0, t) dt
+
_
b
a
n

i=1
L
v
i
_
x(0, t),
x
t
(0, t)
_

2
x
i
st
(0, t) dt.
Dierentiating the relations x(s, a) = x(p), x(s, b) = x(q) with respect to s
one obtains
x
s
(0, a) =
x
s
(0, b) = 0.
Consequently, the second integral above can be integrated by parts to yield

_
b
a
n

i=1
d
dt
_
L
v
i
_
x(0, t),
x
t
(0, t)
__
x
i
s
(0, t) dt,
5. LAGRANGIAN MECHANICS 195
and hence
d
ds[
s=0
S((s)) =
_
b
a
n

i=1
_
L
x
i
(x(t), x(t))
d
dt
_
L
v
i
(x(t), x(t))
__
w
i
(t) dt,
where we have set x(t) := (x c)(t) and w(t) :=
x
s
(0, t). This shows that
if c satises the Euler-Lagrange equations then c is a critical point of the
action.
To show the converse, we notice that any smooth function w : [a, b] R
n
satisfying w(a) = w(b) = 0 determines a variation : (, ) ( given in
local coordinates by x(s, t) = x(t) + sw(t), satisfying
x
s
(0, t) = w(t). In
particular, if : [a, b] R is a smooth positive function with (a) = (b) =
0, we can take
w
i
(t) := (t)
_
L
x
i
(x(t), x(t))
d
dt
_
L
v
i
(x(t), x(t))
__
.
Therefore if c is a critical point of the action we must have
_
b
a
n

i=1
_
L
x
i
(x(t), x(t))
d
dt
_
L
v
i
(x(t), x(t))
__
2
(t) dt = 0,
and hence c must satisfy the Euler-Lagrange equations.
The general case (in which the image of c is not contained in the domain
of the local chart) is left as an exercise.
Corollary 5.4. The motions of any conservative mechanical system
(M, , , dU) are the critical points of the action determined by the La-
grangian L := K U.
Therefore we can nd motions of conservative systems by looking for
minima, say, of the action. This variational approach is often very useful.
The energy conservation in a conservative system is, in fact, a particular
case of a more general conservation law, which holds for any Lagrangian.
Before we state it we need the following denitions.
Definition 5.5. The ber derivative of a Lagrangian function L :
TM R at v T
p
M is the linear map (FL)
v
: T
p
M R given by
(FL)
v
(w) :=
d
dt [
t=0
L(v +tw)
for all w T
p
M.
Definition 5.6. If L : TM R is a Lagrangian function then its
associated Hamiltonian function H : TM R is dened as
H(v) := (FL)
v
(v) L(v).
Theorem 5.7. The Hamiltonian function is constant along the solutions
of the Euler-Lagrange equations.
196 5. GEOMETRIC MECHANICS
Proof. In local coordinates we have
H(x, v) =
n

i=1
v
i
L
v
i
(x, v) L(x, v).
Consequently, if c : I R M is a solution of the Euler-Lagrange equa-
tions, given in local coordinates by x = x(t), then
d
dt
(H( c(t))) =
d
dt
_
n

i=1
x
i
(t)
L
v
i
(x(t), x(t)) L(x(t), x(t))
_
=
n

i,j=1
x
i
(t)
L
v
i
(x(t), x(t)) +
n

i=1
x
i
(t)
d
dt
_
L
v
i
(x(t), x(t))
_

i=1
x
i
(t)
L
x
i
(x(t), x(t))
n

i=1
x
i
(t)
L
v
i
(x(t), x(t)) = 0.

Example 5.8. If (M, , , dU) is a conservative mechanical system

then its motions are the solutions of the Euler-Lagrange equations for the
Lagrangian L : TM R given by
L(v) =
1
2
v, v U((v))
(where : TM M is the canonical projection). Clearly,
(FL)
v
(w) =
1
2
d
dt [
t=0
v +tw, v +tw = v, w,
and hence
H(v) = v, v
1
2
v, v +U((v)) =
1
2
v, v +U((v))
is the mechanical energy.
The Lagrangian formulation is particularly useful for exploring the rela-
tion between symmetry and conservation laws.
Definition 5.9. Let G be a Lie group acting on a manifold M. The
Lagrangian L : TM R is said to be G-invariant if
L((dg)
p
v) = L(v)
for all v T
p
M, p M and g G (where g : M M is the map p g p).
We will now show that if a Lagrangian is G-invariant then to each ele-
ment V g there corresponds a conserved quantity. To do so, we need the
following denitions.
5. LAGRANGIAN MECHANICS 197
Definition 5.10. Let G be a Lie group acting on a manifold M. The
innitesimal action of V g on M is the vector eld X
V
X(M) dened
as
X
V
p
:=
d
dt [
t=0
(exp(tV ) p) = (dA
p
)
e
V,
where A
p
: G M is the map A
p
(g) = g p.
Theorem 5.11. (Noether) Let G be a Lie group acting on a manifold
M. If L : TM R is G-invariant then J
V
: TM R dened as J
V
(v) :=
(FL)
v
_
X
V
_
is constant along the solutions of the Euler-Lagrange equations
for all V g.
Proof. Choose local coordinates (x
1
, . . . , x
n
) on M and let (y
1
, . . . , y
m
)
be local coordinates centered at e G. Let A : G M M be the action
of G on M, written in these local coordinates as
(A
1
(x
1
, . . . , x
n
, y
1
, . . . , y
m
), . . . , A
n
(x
1
, . . . , x
n
, y
1
, . . . , y
m
)).
Then the innitesimal action of V =

m
a=1
V
a
y
a
has components
X
i
(x) =
m

a=1
A
i
y
a
(x, 0)V
a
.
Since L is G-invariant, we have
L
_
A
1
(x, y), . . . , A
n
(x, y),
n

i=1
A
1
x
i
(x, y)v
i
, . . . ,
n

i=1
A
n
x
i
(x, y)v
i
_
= L(x
1
, . . . , x
n
, v
1
, . . . , v
n
).
Setting y = y(t) in the above identity, where (y
1
(t), . . . , y
m
(t)) is the ex-
pression of the curve exp(tV ) in local coordinates, and dierentiating with
respect to t at t = 0, we obtain
n

i=1
m

a=1
L
x
i
(x, v)
A
i
y
a
(x, 0)V
a
+
n

i,j=1
m

a=1
L
v
i
(x, v)

2
A
i
y
a
x
j
(x, 0)v
j
V
a
= 0

i=1
L
x
i
(x, v)X
i
(x) +
n

i,j=1
L
v
i
(x, v)
X
i
x
j
(x)v
j
= 0.
In these coordinates,
J
V
(x, v) =
n

i=1
L
v
i
(x, v)X
i
(x).
198 5. GEOMETRIC MECHANICS
Therefore, if c : I R M is a solution of the Euler-Lagrange equations,
given in local coordinates by x = x(t), we have
d
dt
_
J
V
( c(t))
_
=
d
dt
_
n

i=1
L
v
i
(x(t), x(t))X
i
(x(t))
_
=
n

i=1
d
dt
_
L
v
i
(x(t), x(t))
_
X
i
(x(t)) +
n

i,j=1
L
v
i
(x(t), x(t))
X
i
x
j
(x(t)) x
j
(t)
=
n

i=1
d
dt
_
L
v
i
(x(t), x(t))
_
X
i
(x(t))
n

i=1
L
x
i
(x(t), x(t))X
i
(x(t)) = 0.

Remark 5.12. Notice that the map g V X

V
X(M) is linear.
Since (FL)
v
is also linear, we can see J
V
as a linear map g V J
V

(TM). Therefore the Noether theorem yields m = dimg independent

conserved quantities.
Example 5.13. Consider a conservative mechanical system consisting of
k particles with masses m
1
, . . . , m
k
moving in R
3
under a potential energy
U : R
3k
R which depends only on the distances between them. The
motions of the system are the solutions of the Euler-Lagrange equations
obtained from the Lagrangian L : TR
3k
R given by
L(x
1
, . . . , x
k
, v
1
, . . . , v
k
) =
1
2
k

i=1
m
i
v
i
, v
i
U(x
1
, . . . , x
k
).
This Lagrangian is clearly SO(3)-invariant, where the action of SO(3) on
R
3k
is dened through
S (x
1
, . . . , x
k
) = (Sx
1
, . . . , Sx
k
).
The innitesimal action of V so(3) is the vector eld
X
V
(x
1
,...,x
k
)
= (V x
1
, . . . , V x
k
) = ((V ) x
1
, . . . , (V ) x
k
),
where : so(3) R
3
is the isomorphism in Lemma 3.9. On the other hand,
(FL)
(v
1
,...,v
k
)
(w
1
, . . . , w
k
) =
k

i=1
m
i
v
i
, w
i
.
Therefore, the Noether Theorem guarantees that the quantity
J
V
=
k

i=1
m
i
x
i
, (V )x
i
=
k

i=1
m
i
(V ), x
i
x
i
=
_
(V ),
k

i=1
m
i
x
i
x
i
_
is conserved along the motion of the system for any V so(3). In other
words, the systems total angular momentum
Q :=
k

i=1
m
i
x
i
x
i
5. LAGRANGIAN MECHANICS 199
is conserved.
Exercises 5.14.
(1) Complete the proof of Theorem 5.3.
(2) Let (M, , ) be a Riemannian manifold. Show that the critical
points of the arclength, i.e., of the action determined by the La-
grangian L : TM R given by
L(v) = v, v
1
2
(where we must restrict the action to curves with nonvanishing
velocity) are reparameterized geodesics.
(3) (Brachistochrone curve) A particle with mass m moves on a curve
y = y(x) under the action of a constant gravitational eld, cor-
responding to the potential energy U = mgy. The curve satises
y(0) = y(d) = 0 and y(x) < 0 for 0 < x < d.
(a) Assuming that the particle is set free at the origin with zero
velocity, show that its speed at each point is
v =
_
2gy,
and that therefore the travel time between the origin and point
(d, 0) is
S = (2g)

1
2
_
d
0
(1 +y
2
)
1
2
(y)

1
2
dx,
where y

=
dy
dx
.
(b) Show that the curve y = y(x) which corresponds to the mini-
mum travel time satises the dierential equation
d
dx
__
1 +y
2
_
y

= 0.
(c) Check that the solution of this equation satisfying y(0) =
y(d) = 0 is given parametrically by
_
x = R Rsin
y = R +Rcos
where d = 2R. (Remark: This curve is called a cycloid, because it is
the curved traced out by a point on a circle which rolls without slipping on the
xx-axis).
(4) (Charged particle in a stationary electromagnetic eld) The motion
of a particle with mass m > 0 and charge e R in a stationary
electromagnetic eld is determined by the Lagrangian L : TR
3
R
given by
L =
1
2
mv, v +eA, v e ,
where , is the Euclidean inner product, C

(R
3
) is the elec-
tric potential and A X(R
3
) is the magnetic vector potential.
200 5. GEOMETRIC MECHANICS
(a) Show that the equations of motion are
m x = eE +e x B,
where E = grad is the electric eld and B = curl A is
the magnetic eld.
(b) Write an expression for the Hamiltonian function and use the
equations of motion to check that it is constant along any
motion.
(5) (Restricted 3-body problem) Consider two particles with masses
(0, 1) and 1, set in circular orbit around each other. We identify
the plane of the orbit with R
2
and place the center of mass at the
origin. In the rotating frame where the particles are at rest they
are placed at, say, p
1
= (1 , 0) and p
2
= (, 0). The motion of
a third particle with negligible mass in this frame is determined by
the Lagrangian L : T
_
R
2
p
1
, p
2

_
R given by
L(x, y, v
x
, v
y
) =
1
2
_
(v
x
)
2
+ (v
y
)
2
_
+xv
y
yv
x
+
1
2
_
x
2
+y
2
_
+

r
1
+
1
r
2
,
where r
1
, r
2
: R
2
R are the Euclidean distances to p
1
, p
2
.
(a) Find the equations of motion and the Hamiltonian function.
(Remark: Notice the centrifugal and Coriolis force terms cf. Exercise 3.20.11).
(b) Compute the equilibrium points (i.e. the points corresponding
to stationary solutions) which are not on the x-axis.
(6) Consider the mechanical system in Example 5.13.
(a) Use the Noether Theorem to prove that the total linear mo-
mentum
P :=
k

i=1
m
i
x
i
is conserved along the motion.
(b) Show that the systems center of mass, dened as the point
X =

k
i=1
m
i
x
i

k
i=1
m
i
,
moves with constant velocity.
(7) Generalize Example 5.13 to the case in which the particles move in
an arbitrary Riemannian manifold (M, , ), by showing that given
any Killing vector eld X X(M) (cf. Exercise 3.3.8 in Chapter 3)
the quantity
J
X
=
k

i=1
m
i
c
i
, X
is conserved, where c
i
: I R M is the motion of the particle
with mass m
i
.
(8) Consider the action of SO(3) on itself by left multiplication.
6. HAMILTONIAN MECHANICS 201
(a) Show that the innitesimal action of B so(3) is the right-
invariant vector eld determined by B.
(b) Use the Noether Theorem to show that the angular momentum
of the free rigid body is constant.
(9) Consider a satellite equipped with a small rotor, i.e. a cylinder
which can spin freely about its axis. When the rotor is locked the
satellite can be modelled by a free rigid body with inertia tensor
I. The rotors axis passes through the satellites center of mass,
and its direction is given by the unit vector e. The rotors mass is
symmetrically distributed around the axis, producing a moment of
inertia J.
(a) Show that the conguration space for the satellite with un-
locked rotor is the Lie group SO(3) S
1
, and that its motion
is a geodesic of the left-invariant metric corresponding to the
kinetic energy
K =
1
2
I, +
1
2
J
2
+J, e,
where the R
3
is the satellites angular velocity as seen on
the satellites frame and R is the rotors angular speed
around its axis.
(b) Use the Noether Theorem to show that l = J(+, e) R
and p = S(I + Je) R
3
are conserved along the motion
of the satellite with unlocked rotor, where S : R SO(3)
describes the satellites orientation.
6. Hamiltonian Mechanics
We will now see that under certain conditions it is possible to study the
Euler-Lagrange equations as a ow on the cotangent bundle with special
geometric properties.
Let M be an n-dimensional manifold. The set
TM T

M :=
_
qM
T
q
M T

q
M
has an obvious dierentiable structure: if (x
1
, . . . , x
n
) are local coordinates
on M then (x
1
, . . . , x
n
, v
1
, . . . , v
n
, p
1
, . . . , p
n
) are the local coordinates on
TM T

M which label the pair (v, ) T

q
M T

q
M, where
v =
n

i=1
v
i

x
i
, =
n

i=1
p
i
dx
i
,
and q M is the point with coordinates (x
1
, . . . , x
n
). For this dierentiable
structure, the maps
1
: TM T

M TM and
2
: TM T

M T

M
given by
1
(v, ) = v and
2
(v, ) = are submersions.
202 5. GEOMETRIC MECHANICS
Definition 6.1. The extended Hamiltonian function corresponding
to a Lagrangian L : TM R is the map

H : TM T

M R given by

H(v, ) := (v) L(v).

In local coordinates, we have

H(x
1
, . . . , x
n
, v
1
, . . . , v
n
, p
1
, . . . , p
n
) =
n

i=1
p
i
v
i
L(x
1
, . . . , x
n
, v
1
, . . . , v
n
),
and hence
d

H =
n

i=1
_
p
i

L
v
i
_
dv
i
+
n

i=1
v
i
dp
i

i=1
L
x
i
dx
i
.
Thus any critical point of any restriction of

H to a submanifold of the form
T
q
M (for xed q M and T

q
M) must satisfy
p
i

L
v
i
(x
1
, . . . , x
n
, v
1
, . . . , v
n
) = 0 (i = 1, . . . , n).
It follows that the set of all such critical points is naturally a 2n-dimensional
submanifold S TMT

M such that
1[
S
: S TM is a dieomorphism.
If
2[
S
: S T

M is also a dieomorphism then the Lagrangian is said

to be hyper-regular. In this case,
2[
S

1[
S
1
: TM T

M is a ber-
preserving dieomorphism, called the Legendre transformation.
Given a hyper-regular Lagrangian, we can use the maps
1[
S
and
2[
S
to make the identications TM

= S

= T

M. Since the Hamiltonian func-

tion H : TM R is clearly related to the extended Hamiltonian function
through H =

H
1[
S
1
, we can under these identications simply write
H =

H[
S
. Therefore
dH =
n

i=1
v
i
dp
i

i=1
L
x
i
dx
i
(here we must think of (x
1
, . . . , x
n
, v
1
, . . . , v
n
, p
1
, . . . , p
n
) as local functions
on S such that both (x
1
, . . . , x
n
, v
1
, . . . , v
n
) and (x
1
, . . . , x
n
, p
1
, . . . , p
n
) are
local coordinates). On the other hand, thinking of H as a function on the
cotangent bundle, we obtain
dH =
n

i=1
H
x
i
dx
i
+
n

i=1
H
p
i
dp
i
.
Therefore we must have
_

_
H
x
i
=
L
x
i
H
p
i
= v
i
(i = 1, . . . , n),
6. HAMILTONIAN MECHANICS 203
where the partial derivatives of the Hamiltonian must be computed with re-
spect to the local coordinates (x
1
, . . . , x
n
, p
1
, . . . , p
n
) and the partial deriva-
tives of the Lagrangian must be computed with respect to the local coordi-
nates (x
1
, . . . , x
n
, v
1
, . . . , v
n
).
Proposition 6.2. The Euler-Lagrange equations for a hyper-regular La-
grangian L : TM R dene a ow on TM. This ow is carried by the Le-
gendre transformation to the ow dened on T

M by the Hamilton equa-

tions
_

_
x
i
=
H
p
i
p
i
=
H
x
i
(i = 1, . . . , n).
Proof. The Euler-Lagrange equations can be cast as a system of rst
order ordinary dierential equations on TM as follows.
_

_
x
i
= v
i
d
dt
_
L
v
i
_
=
L
x
i
(i = 1, . . . , n).
Since on S one has
p
i
=
L
v
i
, v
i
=
H
p
i
,
L
x
i
=
H
x
i
,
we see that this system reduces to the Hamilton equations in the local coor-
dinates (x
1
, . . . , x
n
, p
1
, . . . , p
n
). Since the Hamilton equations clearly dene
a ow on T

M, the Euler-Lagrange equations must dene a ow on TM.

Example 6.3. The Lagrangian for a conservative mechanical system
(M, , , dU) is written in local coordinates as
L(x
1
, . . . , x
n
, v
1
, . . . , v
n
) =
1
2
n

i,j=1
g
ij
(x
1
, . . . , x
n
)v
i
v
j
U(x
1
, . . . , x
n
).
The Legendre transformation is given in these coordinates by
p
i
=
L
v
i
=
n

j=1
g
ij
v
j
(i = 1, . . . , n),
and is indeed a ber-preserving dieomorphism, whose inverse is given by
v
i
=
n

j=1
g
ij
p
j
(i = 1, . . . , n).
As a function on the tangent bundle, the Hamiltonian is (cf. Example 5.8)
H =
1
2
n

i,j=1
g
ij
v
i
v
j
+U.
204 5. GEOMETRIC MECHANICS
Using the Legendre transformation, we can see the Hamiltonian as the fol-
lowing function on the cotangent bundle.
H =
1
2
n

i,j,k,l=1
g
ij
g
ik
p
k
g
jl
p
l
+U =
1
2
n

k,l=1
g
kl
p
k
p
l
+U.
Therefore the Hamilton equations for a conservative mechanical system are
_

_
x
i
=
n

j=1
g
ij
p
j
p
i
=
1
2
n

k,l=1
g
kl
x
i
p
k
p
l

U
x
i
(i = 1, . . . , n).
The ow dened by the Hamilton equations has remarkable geometric
properties, which are better understood by introducing the following deni-
tion.
Definition 6.4. The canonical symplectic potential is the 1-form

1
(T

M) given by

(v) := ((d)

(v))
for all v T

(T

M) and all T

M, where : T

M M is the
natural projection. The canonical symplectic form on T

M is the 2-
form
2
(T

M) given by = d.
In local coordinates, we have
(x
1
, . . . , x
n
, p
1
, . . . , p
n
) = (x
1
, . . . , x
n
)
and
v =
n

i=1
dx
i
(v)

x
i
+
n

i=1
dp
i
(v)

p
i
.
Consequently,
(d)

(v) =
n

i=1
dx
i
(v)

x
i
,
and hence

(v) = ((d)

(v)) =
n

i=1
p
i
dx
i
_
_
n

j=1
dx
j
(v)

x
j
_
_
=
n

i=1
p
i
dx
i
(v).
We conclude that
=
n

i=1
p
i
dx
i
,
and consequently
=
n

i=1
dp
i
dx
i
.
6. HAMILTONIAN MECHANICS 205
Proposition 6.5. The canonical symplectic form is closed (d = 0)
and non-degenerate. Moreover,
n
= is a volume form (in
particular T

M is always orientable, even if M itself is not).

We leave the proof of this proposition as an exercise (cf. Exercise 6.15.1).
Recall from Exercise 1.15.8 in Chapter 2 that if v T
p
M then (v) T

p
M
is the covector given by
((v)) (w) = (v, w)
for all w T
p
M. Therefore the rst statement in Proposition 6.5 is equiv-
alent to saying that the map T
p
M v (v) T

p
M is a linear isomor-
phism for all p M.
The key to the geometric meaning of the Hamilton equations is contained
in the following result.
Proposition 6.6. The Hamilton equations are the equations for the ow
of the vector eld X
H
satisfying
(X
H
) = dH.
Proof. The Hamilton equations yield the ow of the vector eld
X
H
=
n

i=1
_
H
p
i

x
i

H
x
i

p
i
_
.
Therefore
(X
H
) = (X
H
)
n

i=1
(dp
i
dx
i
dx
i
dp
i
)
=
n

i=1
_

H
x
i
dx
i

H
p
i
dp
i
_
= dH.

Remark 6.7. Notice that H completely determines X

H
, as is nonde-
generate. By analogy with the Riemannian case, X
H
is sometimes called
the symplectic gradient of H. The vector eld X
H
is usually referred to
as the Hamiltonian vector eld determined by H.
Definition 6.8. The Hamiltonian ow generated by F C

(T

M)
is the ow of the unique vector eld X
F
X(T

M) such that
(X
F
) = dF.
The ow determined on T

M by a hyper-regular Lagrangian is therefore

a particular case of a Hamiltonian ow (in which the generating function is
the Hamiltonian function). We will now discuss the geometric properties of
general Hamiltonian ows, starting with the Hamiltonian version of energy
conservation.
Proposition 6.9. Hamiltonian ows preserve their generating func-
tions.
206 5. GEOMETRIC MECHANICS
Proof. We have
X
F
F = dF(X
F
) = ((X
F
)) (X
F
) = (X
F
, X
F
) = 0,
as is alternating.
Proposition 6.10. Hamiltonian ows preserve the canonical symplectic
form: if
t
: T

M T

M is a Hamiltonian ow then
t

= .
Proof. Let F C

(T

M) be the function whose Hamiltonian ow is

t
. Recall from Exercise 3.8.7 in Chapter 2 that the Lie derivative of
along X
F
X(T

M),
L
X
F
=
d
dt [
t=0

,
can be computed by the Cartan formula:
L
X
F
= (X
F
)d +d((X
F
)) = d(dF) = 0.
Therefore
d
dt

=
d
ds[
s=0
(
t+s
)

=
d
ds[
s=0
(
s

t
)

=
d
ds[
s=0

(
s

)
=
t

d
ds[
s=0

=
t

L
X
F
= 0.
We conclude that

= (
0
)

= .

Theorem 6.11. (Liouville) Hamiltonian ows preserve the integral with

respect to the symplectic volume form: if
t
: T

M T

M is a Hamiltonian
ow and F C

(T

M) is a compactly supported function then

_
T

M
F
t
=
_
T

M
F.
Proof. This is a simple consequence of the fact that
t
preserves the
symplectic volume form, since

(
n
) = (
t

)
n
=
n
.
Therefore
_
T

M
F
t
=
_
T

M
(F
t
)
n
=
_
T

M
(F
t
)
t

(
n
)
=
_
T

(F
n
) =
_
T

M
F
n
=
_
T

M
F
(cf. Exercise 4.2.4 in Chapter 2).
Theorem 6.12. (Poincare Recurrence) Let
t
: T

M T

M be a
Hamiltonian ow and K T

M a compact set invariant under

t
. Then
for each open set U K and each T > 0 there exist U and t T such
that
t
() U.
6. HAMILTONIAN MECHANICS 207
Proof. Let F C

(T

M) be a compactly supported smooth function

with values in [0, 1] such that F() = 1 for all K (this is easily con-
structed using, for instance, a partition of unity). Let G C

(T

M) be a
smooth function with values in [0, 1] and compact support contained in U
such that
_
T

M
G > 0.
Consider the open sets U
n
:=
nT
(U). If these sets were all disjoint then
one could dene functions

G
N
C

(M) for each N N as

G
N
() =
_
(G
nT
)() if U
n
and n N
0 otherwise
.
These functions would have compact support contained in K (K is invariant
under
t
) and values in [0, 1], and hence would satisfy

G
N
F. Therefore
we would have
_
T

M
F
_
T

G
N
=
N

n=1
_
T

M
G
nT
= N
_
T

M
G
for all N N, which is absurd. We conclude that there must exist m, n N
(with, say, n > m) such that
U
m
U
n
,=
mT
(U)
nT
(U) ,= U
(nm)T
(U) ,= .
Choosing t = (n m)T and
t
(U
t
(U)) =
t
(U) U yields the
result.
We can use the symplectic structure of the contangent bundle to dene
a new binary operation on the set of dierentiable functions on T

M.
Definition 6.13. The Poisson bracket of two dierentiable functions
F, G C

(T

M) is F, G := X
F
G.
Proposition 6.14. (C

(T

M), , ) is a Lie algebra, and the map

that associates to a function F C

(T

M) its Hamiltonian vector eld

X
F
X(T

M) is a Lie algebra homomorphism, i.e.

(i) F, G = G, F;
(ii) F +G, H = F, H +G, H;
(iii) F, G, H +G, H, F +H, F, G = 0;
(iv) X
F,G
= [X
F
, X
G
]
for any F, G, H C

(T

M) and any , R.
Proof. We have
F, G = X
F
G = dG(X
F
) = ((X
G
))(X
F
)
= (X
G
, X
F
) = (X
F
, X
G
),
208 5. GEOMETRIC MECHANICS
which proves the anti-symmetry and bilinearity of the Poisson bracket. On
the other hand,
(X
F,G
) = dF, G = d(X
F
G) = d(dG(X
F
)) = d((X
F
)dG)
= L
X
F
dG = L
X
F
((X
G
)) = (L
X
F
X
G
) +(X
G
)L
X
F

= ([X
F
, X
G
])
(cf. Exercise 3.8.7 in Chapter 2). Since is non-degenerate, we have
X
F,G
= [X
F
, X
G
].
Finally,
F, G, H +G, H, F +H, F, G
= F, X
G
H G, X
F
H X
F,G
H
= X
F
(X
G
H) X
G
(X
F
H) [X
F
, X
G
] H = 0.

Exercises 6.15.
(1) Prove Proposition 6.5.
(2) Let (M, , ) be a Riemannian manifold,
1
(M) a 1-form and
U C

(M) a dierentiable function.

(a) Show that the Euler-Lagrange equations for the Lagrangian
L : TM R given by
L(v) =
1
2
v, v +(v)
p
U(p)
for v T
p
M yield the motions of the mechanical system
(M, , , T), where
T(v) = (dU)
p
(v)(d)
p
for v T
p
M.
(b) Show that the mechanical energy E = K + U is conserved
along the motions of (M, , , T) (which is therefore called a
conservative mechanical system with magnetic term).
(c) Show that L is hyper-regular and compute the Legendre trans-
formation.
(d) Find the Hamiltonian H : T

M R and write the Hamilton

equations.
(3) Let c > 0 be a positive number, representing the speed of light, and
consider the open set U := v TR
n
[ |v| < c, where | | is the
Euclidean norm. The motion of a free relativistic particle of mass
m > 0 is determined by the Lagrangian L : U R given by
L(v) := mc
2
_
1
|v|
2
c
2
.
(a) Show that L is hyper-regular and compute the Legendre trans-
formation.
7. COMPLETELY INTEGRABLE SYSTEMS 209
(b) Find the Hamiltonian H : T

R
n
R and write the Hamilton
equations.
(4) Show that in the Poincare Recurrence Theorem the set of points
U such that
t
() U for some t T is dense in U. (Remark:
It can be shown that this set has full measure).
(5) Let (M, , ) be a compact Riemannian manifold. Show that for
each normal ball B M and each T > 0 there exist geodesics
c : R M with | c(t)| = 1 such that c(0) B and c(t) B for
some t T.
(6) Let (x
1
, . . . , x
n
, p
1
, . . . , p
n
) be the usual local coordinates on T

M.
Compute X
x
i , X
p
i
, x
i
, x
j
, p
i
, p
j
and p
i
, x
j
.
(7) Show that the Poisson bracket satises the Leibniz rule
F, GH = F, GH +F, HG
for all F, G, H C

(T

M).
7. Completely Integrable Systems
We now concentrate on studying the Hamiltonian ow of a Hamiltonian
function H C

(T

M). We already know that H is constant along its

Hamiltonian ow, so that it suces to study this ow along the level sets
of H. This can be further simplied if there exist additional nontrivial
functions F C

(T

M) such that
X
H
F = 0 H, F = 0.
Definition 7.1. A function F C

(T

M) is said to be a rst inte-

gral of H if H, F = 0.
In general, there is no reason to expect that there should exist nontrivial
rst integrals other than H itself. In the special cases when these exist, they
often satisfy additional conditions.
Definition 7.2. The functions F
1
, . . . , F
m
C

(T

M) are said to be
(i) in involution if F
i
, F
j
= 0 (i, j = 1, . . . , m);
(ii) independent at T

M if (dF
1
)

, . . . , (dF
m
)

(T

M) are
linearly independent covectors.
Proposition 7.3. If F
1
, . . . , F
m
C

(T

M) are in involution and are

independent at some point T

M then m n.
We leave the proof of this proposition as an exercise. The maximal case
m = n is especially interesting.
Definition 7.4. The Hamiltonian H is said to be completely inte-
grable if there exist n rst integrals F
1
, . . . , F
n
in involution which are in-
dependent on a dense open set U T

M.
Example 7.5.
210 5. GEOMETRIC MECHANICS
(1) If M is 1-dimensional and dH ,= 0 on a dense open set of T

M
then H is completely integrable.
(2) (Particle in a central eld) Recall Example 1.15 where a particle of
mass m > 0 moves in a central eld. The corresponding Lagrangian
function is
L
_
r, , v
r
, v

_
=
1
2
m
_
(v
r
)
2
+r
2
_
v

_
2
_
u(r),
and so the Legendre transformation is given by
p
r
=
L
v
r
= mv
r
and p
r
=
L
v

= mr
2
v

.
The Hamiltonian function is then
H (r, , p
r
, p

) =
p
r
2
2m
+
p

2
2mr
2
+u(r).
By the Hamilton equations,
p

=
H

= 0,
and hence p

is a rst integral. Since

dH =
_

2
mr
3
+u

(r)
_
dr +
p
r
m
dp
r
+
p

mr
2
dp

,
we see that dH and dp

are independent on the dense open set of

T

R
2
formed by the points whose polar coordinates (r, , p
r
, p

) are
well dened and do not satisfy
u

(r)
p

2
mr
3
= p
r
= 0
(i.e. are not on a circular orbit cf. Exercise 7.17.4). Therefore this
Hamiltonian is completely integrable.
Proposition 7.6. Let H be a completely integrable Hamiltonian with
rst integrals F
1
, . . . , F
n
in involution, independent in the dense open set
U T

M, and such that X

F
1
, . . . , X
Fn
are complete on U. Then each
nonempty level set
L
f
:= U : F
1
() = f
1
, . . . , F
n
() = f
n

is a submanifold of dimension n, invariant for the Hamiltonian ow of H,

admitting a locally free action of R
n
which is transitive on each connected
component.
Proof. All points in U are regular points of the map F : U R
n
given by F() = (F
1
(), . . . , F
n
()); therefore all nonempty level sets L
f
:=
F
1
(f) are submanifolds of dimension n.
Since X
H
F
i
= 0 for i = 1, . . . , n, the level sets L
f
are invariant for the
ow of X
H
. In addition, we have X
F
i
F
j
= F
i
, F
j
= 0, and hence these
7. COMPLETELY INTEGRABLE SYSTEMS 211
level sets are invariant for the ow of X
F
i
. Moreover, these ows commute,
as [X
F
i
, X
F
j
] = X
F
i
,F
j

= 0 (cf. Theorem 6.10 in Chapter 1).

Consider the map A : R
n
L
f
L
f
given by
A(t
1
, . . . , t
n
, ) = (
1,t
1

n,tn
)(),
where
i,t
: L
f
L
f
is the ow of X
F
i
. Since these ows commute, this
map denes an action of R
n
on L
f
. On the other hand, for each L
f
,
the map A

: R
n
L
f
given by A

(t
1
, . . . , t
n
) = A(t
1
, . . . , t
n
, ) is a local
dieomorphism at the origin, as
(dA

)
0
(e
i
) =
d
dt [
t=0

i,t
() = (X
F
i
)

and the vector elds X

F
i
are linearly independent. Therefore the action is
locally free (that is, for each point L
f
there exists an open neighborhood
U R
n
of 0 such that A(t, ) ,= for all t U 0), meaning that the
isotropy groups are discrete. Also, the action is locally transitive (i.e. each
point L
f
admits an open neighborhood V L
f
such that every V
is of the form = A(t, ) for some t R
n
), and hence transitive on each
connected component (for given L
f
both the set of points L
f
which
are of the form = A(t, ) for some t R
n
and the set of points which are
not are open).
The isotropy subgroups of the action above are discrete subgroups of
R
n
. We next describe the structure of such subgroups.
Proposition 7.7. Let be a discrete subgroup of R
n
. Then there ex-
ist k 0, 1, . . . , n linearly independent vectors e
1
, . . . , e
k
such that =
span
Z
e
1
, . . . , e
k
.
Proof. If = 0 then we are done. If not, let e 0. Since is
discrete, the set
e [ 0 < 1
is nite (and nonempty). Let e
1
be the element in this set which is closest
to 0. Then
span
R
e = span
Z
e
1

(for otherwise e
1
would not be the element in this set closest to 0). If
= span
Z
e
1
then we are done. If not, let e span
Z
e
1
. Then the set
e +
1
e
1
[ 0 < ,
1
1
is nite (and nonempty). Let e
2
be the element in this set which is closest
to span
R
e
1
. Then
span
R
e, e
1
= span
Z
e
1
, e
2
.
Iterating this procedure yields the result.
212 5. GEOMETRIC MECHANICS
Proposition 7.8. Let L

f
be the connected component of L
f
. Then
L

f
is dieomorphic to T
k
R
nk
, where k is the number of generators of the
isotropy subgroup

. In particular, if L

f
is compact then it is dieomorphic
to the n-dimensional torus T
n
.
Proof. Since the action A : R
n
L

f
L

f
is transitive, the local
dieomorphism A

: R
n
L

f
is surjective. On the other hand, because

is discrete, the action of

on R
n
by translation is free and proper, and we
can form the quotient R
n
/

, which is clearly dieomorphic to T

k
R
nk
.
Finally, it is easily seen that A

induces a dieomorphism R
n
/

= L

f
.
We are now in position to understand the Hamiltonian ow on a com-
pletely integrable system. For that we need the following denition.
Definition 7.9. A linear ow on the torus T
n
= R
n
/Z
n
is the projec-
tion of the ow
t
: R
n
R
n
given by

t
(x) = x +t.
The frequencies of the linear ow are the components
1
, . . . ,
n
of .
Theorem 7.10. (Arnold-Liouville) Let H be a completely integrable
Hamiltonian with n rst integrals F
1
, . . . , F
n
C

(T

M) in involution,
independent on the dense open set U T

M. If the connected components

of the level sets of the map (F
1
, . . . , F
n
) : U R
n
are compact then they are
n-dimensional tori, invariant for the ow of X
H
. The ow of X
H
on these
tori is a linear ow (for an appropriate choice of coordinates).
Proof. Since the connected components of the level sets of (F
1
, . . . , F
n
)
are compact, the vector elds X
F
i
are complete. All that remains to be seen
is that the ow of X
H
on the invariant tori is a linear ow. It is clear that
the ow of each X
F
i
is linear in the coordinates given by Proposition 7.8.
Since X
H
is tangent to the invariant tori, we have X
H
=

n
i=1
f
i
X
F
i
for
certain functions f
i
. Now
0 = X
F
i
,H
= [X
F
i
, X
H
] =
n

j=1
(X
F
i
f
j
)X
F
j
,
and hence each function f
i
is constant on the invariant torus. We conclude
that the ow of X
H
is linear.
We next explore in detail the properties of linear ows on the torus.
Definition 7.11. Let
t
: T
n
T
n
be a linear ow. The time average
of a function f C

(T
n
) along
t
is the map
f(x) := lim
T+
1
T
_
T
0
f(
t
(x))dt
(dened on the set of points x T
n
where the limit exists).
7. COMPLETELY INTEGRABLE SYSTEMS 213
Definition 7.12. The frequencies R
n
of a linear ow
t
: T
n
T
n
are said to be independent if they are linearly independent over Q, i.e. if
k, , = 0 for all k Z
n
0.
Theorem 7.13. (Birkho Ergodicity) If the frequencies R
n
of a
linear ow
t
: T
n
T
n
are independent then the time average of any
function f C

(T
n
) exists for all x T
n
and
f(x) =
_
T
n
f.
Proof. Since T
n
= R
n
/Z
n
, the dierentiable functions on the torus
arise from periodic dierentiable functions on R
n
, which can be expanded
as uniformly convergent Fourier series. Therefore it suces to show that
the theorem holds for f(x) = e
2ik,x)
with k Z
n
.
If k = 0 then both sides of the equality are 1, and the theorem holds.
If k ,= 0 that the right-hand side of the equality is zero, whereas the
left-hand side is
f(x) = lim
T+
1
T
_
T
0
e
2ik,x+t)
dt
= lim
T+
1
T
e
2ik,x)
e
2ik,)T
1
2ik,
= 0
(where we used the fact that k, ,= 0).
Corollary 7.14. If the frequencies of a linear ow
t
: T
n
T
n
are
independent then
t
(x) [ t 0 is dense on the torus for all x T
n
.
Proof. If
t
(x) [ t 0 were not dense then it would not intersect an
open set U T
n
. Therefore any nonnegative function f C

(T
n
) with
nonempty support contained in U would satisfy f(x) = 0 and
_
T
n
f > 0,
contradicting the Birkho Ergodicity Theorem.
Corollary 7.15. If the frequencies of a linear ow
t
: T
n
T
n
are
independent and n 2 then
t
(x) is not periodic.
Remark 7.16. The qualitative behavior of the Hamiltonian ow gen-
erated by completely integrable Hamiltonians is completely understood.
Complete integrability is however a very strong condition, not satised by
generic Hamiltonians. The Komolgorov-Arnold-Moser (KAM) Theo-
rem guarantees a small measure of genericity by establishing that a large
fraction of the invariant tori of a completely integrable Hamiltonians sur-
vives under small perturbations of the Hamiltonian, the ow on these tori
remaining linear with the same frequencies. On the other hand, many invari-
ant tori, including those whose frequencies are not independent (resonant
tori), are typically destroyed.
Exercises 7.17.
214 5. GEOMETRIC MECHANICS
(1) Show that if F, G C

(T

M) are rst integrals, then F, G is

also a rst integral.
(2) Prove Proposition 7.3.
(3) Consider a surface of revolution M R
3
given in cylindrical coor-
dinates (r, , z) by
r = f(z),
where f : (a, b) (0, +) is dierentiable.
(a) Show that the geodesics of M are the critical points of the
action determined by the Lagrangian L : TM R given in
local coordinates by
L(, z, v

, v
z
) =
1
2
_
(f(z))
2
(v

)
2
+
_
(f

(z))
2
+ 1
_
(v
z
)
2
_
.
(b) Show that the curves given in local coordinates by = constant
or f

(z) = 0 are images of geodesics.

(c) Compute the Legendre transformation, show that L is hyper-
regular and write an expression in local coordinates for the
Hamiltonian H : T

M R.
(d) Show that H is completely integrable.
(e) Show that the projection on M of the invariant set
L
(E,l)
:= H
1
(E) p

1
(l)
(E, l > 0) is given in local coordinates by
f(z)
l

2E
.
Use this fact to conclude that if f has a strict local maximum
at z = z
0
then the geodesic whose image is z = z
0
is stable,
i.e. geodesics with initial condition close to the point in TM
with coordinates (
0
, z
0
, 1, 0) stay close to the curve z = z
0
.
(4) Recall from Example 7.5 that a particle of mass m > 0 moving in a
central eld is described by the completely integrable Hamiltonian
function
H (r, , p
r
, p

) =
p
r
2
2m
+
p

2
2mr
2
+u(r).
(a) Show that there exist circular orbits of radius r
0
whenever
u

(r
0
) 0.
(b) Verify that the set of points where dH and dp

are not inde-

pendent is the union of these circular orbits.
(c) Show that the projection of the invariant set
L
(E,l)
:= H
1
(E) p

1
(l)
on R
2
is given in local coordinates by
u(r) +
l
2
2mr
2
E.
7. COMPLETELY INTEGRABLE SYSTEMS 215
(d) Conclude that if u

(r
0
) 0 and
u

(r
0
) +
3u

(r
0
)
r
0
> 0
then the circular orbit of radius r
0
is stable.
(5) In General Relativity, the motion of a particle in the gravitational
eld of a point mass M > 0 is given by the Lagrangian L : TU R
written in cylindrical coordinates (u, r, ) as
L(u, r, , u, r,

) =
1
2
_
1
2M
r
_
u
2
+
1
2
_
1
2M
r
_
1
r
2
+
1
2
r
2

2
,
where U R
3
is the open set given by r > 2M (the coordinate u
is called the time coordinate, and in general is dierent from the
proper time of the particle, i.e. the parameter t of the curve).
(a) Show that L is hyper-regular and compute the corresponding
Hamiltonian H : T

U R.
(b) Show that H is completely integrable.
(c) Show that there exist circular orbits of any radius r
0
> 2M,
with H < 0 for r
0
> 3M, H = 0 for r
0
= 3M and H > 0 for
r
0
< 3M. (Remark: The orbits with H > 0 are not physical, since they
correspond to speeds greater than the speed of light; the orbits with H = 0 can
only be achieved by massless particles, which move at the speed of light).
(d) Show that the set of points where dH, dp
u
and dp

are not
independent (and p
u
,= 0) is the union of these circular orbits.
(e) Show that the projection of the invariant cylinder
L
(E,k,l)
:= H
1
(E) p
u
1
(k) p

1
(l)
on U is given in local coordinates by
l
2
r
2

_
1
2M
r
_
1
k
2
2E.
(f) Conclude that if r
0
> 6M then the circular orbit of radius r
0
is stable.
(6) Recall that the Lagrange top is the mechanical system determined
by the Lagrangian L : TSO(3) R given in local coordinates by
L =
I
1
2
_

2
+
2
sin
2

_
+
I
3
2
_

+ cos
_
2
Mgl cos ,
where (, , ) are the Euler angles, M is the tops mass and l is
the distance from the xed point to the center of mass.
(a) Compute the Legendre transformation, show that L is hyper-
regular and write an expression in local coordinates for the
Hamiltonian H : T

SO(3) R.
(b) Prove that H is completely integrable.
(c) Show that the solutions found in Exercise 3.20.14 are stable
for [ [ [

[ if [

[ is large enough.
216 5. GEOMETRIC MECHANICS
(7) Show that the the Euler top with I
1
< I
2
< I
3
denes a completely
integrable Hamiltonian on T

SO(3).
(8) Consider the sequence formed by the rst digit of the decimal ex-
pansion of each of the integers 2
n
for n N
0
:
1, 2, 4, 8, 1, 3, 6, 1, 2, 5, 1, 2, 4, 8, 1, 3, 6, 1, 2, 5, . . .
The purpose of this exercise is to answer the following question: is
there a 7 in this sequence?
(a) Show that if R Q then
lim
n+
1
n + 1
n

k=0
e
2ik
= 0.
(b) Prove the following discrete version of the Birkho Ergodicity
Theorem: if a dierentiable function f : R R is periodic
with period 1 and R Q then for all x R
lim
n+
1
n + 1
n

k=0
f(x +k) =
_
1
0
f(x)dx.
(c) Show that log 2 is an irrational multiple of log 10.
(d) Is there a 7 in the sequence above?
8. Symmetry and Reduction
The symplectic structure on the cotangent bundle can be generalized to
arbitrary manifolds.
Definition 8.1. A symplectic manifold is a pair (M, ), where M
is a dierentiable manifold and
2
(M) is nondegenerate and closed.
Example 8.2. If M is an orientable surface and
2
(M) is a volume
form on M then (M, ) is a symplectic manifold. In fact, is necessarily
non-degenerate (if (v) = 0 for some nonvanishing v T
p
M then
p
= 0),
and d = 0 trivially.
All denitions and results of Sections 6 and 7 (Hamiltonian ow and its
properties, Liouvillie and Poincare Recurrence Theorems, Poisson bracket,
completely integrable systems and the Arnold-Liouville Theorem) are read-
ily extended to arbitrary symplectic manifolds. In fact, all symplectic man-
ifolds are locally the same (i.e. there is no symplectic analogue of the cur-
vature), as we now show.
Theorem 8.3. (Darboux) Let (M, ) be a symplectic manifold and p
M. Then there exist local coordinates (x
1
, . . . , x
n
, p
1
, . . . , p
n
) around p such
that
=
n

i=1
dp
i
dx
i
(in particular the dimension of M is necessarily even).
8. SYMMETRY AND REDUCTION 217
Proof. We begin by observing that is of the form above if and only if
x
i
, x
j
= p
i
, p
j
= 0 and p
i
, x
j
=
ij
for i, j = 1, . . . , n (cf. Exercise 4).
Clearly we must have m := dimM 2 (otherwise = 0 would be
degenerate). Let P C

(M) be a function with (dP)

p
,= 0, let X
P
be the
corresponding Hamiltonian vector eld and let T M be a hypersurface
not tangent to (X
P
)
p
(cf. Figure 6). Then X
P
is not tangent to T on
some neighborhood V of p. Possibly reducing V , we can dene a smooth
function Q on V by the condition that
Q(q)
(q) T for each q V , where

t
is the ow of X
P
. Notice that T V = Q
1
(0), implying that X
Q
is
tangent to T, and so (X
P
)
p
, (X
Q
)
p
is a linearly independent set. This
means that (dP)
p
, (dQ)
p
is linearly independent, and so, reducing V if
necessary, (P, Q) can be extended to a system of local coordinates around
p. If m = 2 then we are done, because (Q, P) are local coordinates and
P, Q = X
P
Q = 1.
If m > 2 then, since X
P
is not tangent to T, the level set P
1
(P(p))
intersects T on a (m 2)-dimensional manifold S T. Since P, Q = 1,
we have X
Q
P = Q, P = 1, and so X
Q
is not tangent to S. If q S and
v
1
, . . . , v
m2
is a basis for T
q
S then (X
P
)
q
, (X
Q
)
q
, v
1
, . . . , v
m2
is a basis
for T
q
M. Moreover, we have (X
P
, v
i
) = dP(v
i
) = 0 (as P is constant in
S), and similarly (X
Q
, v
i
) = dQ(v
i
) = 0. We conclude that the matrix
((v
i
, v
j
)) must be nonsingular, that is, i

must be nondegenerate, where

i : S V is the inclusion map. Since di

= i

d = 0, we see that (S, i

)
is a symplectic manifold. Given any function F C

(S), we can extend it

to T by making it constant along the ow of X
Q
, and then to V by making
it constant along the ow of X
P
. Since [X
P
, X
Q
] = X
P,Q
= X
1
= 0, the
ows of X
P
and X
Q
commute, and so the extended function (which we still
call F) satises P, F = X
P
F = 0 and Q, F = X
Q
F = 0, that is,
X
F
P = F, P = 0 and X
F
Q = F, Q = 0. This implies that X
F
is
tangent to S, and so X
F
coincides on S with the Hamiltonian vector eld
determined by F on (S, i

). In the same way, the Poisson bracket F, G

of the extensions to V of two functions F, G C

(S) satises
X
P
F, G = P, F, G = P, F, G+F, P, G = 0, G+F, 0 = 0,
and similarly X
Q
F, G = 0, implying that F, G is the extension of
the Poison bracket on (S, i

). Thefore, if the Darboux Theorem holds

for (S, i

), meaning that we have m 2 = 2n 2 and local coordinates

(x
1
, . . . , x
n1
, p
1
, . . . , p
n1
) with x
i
, x
j
= p
i
, p
j
= 0 and p
i
, x
j
=
ij
for i, j = 1, . . . , n 1, then, making x
n
= Q and p
n
= P, we have the result
for (M, ).
In fact, to have Hamiltonian ows all that is required is the existence of
a Poisson bracket. This suggests a further generalization.
Definition 8.4. A Poisson manifold is a pair (M, , ), where M
is a dierentiable manifold and , , called the Poisson bracket, is a Lie
bracket on C

(M) satifying the Leibniz rule, that is,

218 5. GEOMETRIC MECHANICS
T
S
X
P
p
q

Q(q)
(q)
Figure 6. Proof of the Darboux Theorem.
(i) F, G = G, F;
(ii) F +G, H = F, H +F, H;
(iii) F, G, H +G, H, F +H, F, G = 0;
(iv) F, GH = F, GH +F, HG
for any , R and F, G, H C

(M).
Example 8.5.
(1) Any symplectic manifold (M, ) is naturally a Poisson manifold
(M, , ) (cf. Exercise 6.15.7).
(2) Any smooth manifold M can be given a Poisson structure, namely
the trivial Poisson bracket , := 0. This is not true for symplectic
structures, even if M is even-dimensional (cf. Exercise 8.23.4).
(3) If , is the Euclidean inner product on R
3
then the formula
F, G(x) := x, grad F grad G
denes a nontrivial Poisson bracket on R
3
(cf. Example 8.22).
The bilinearity and Leibniz rule properties of the Poisson bracket imply
that F, is a derivation (hence a vector eld) for any F C

(M). This
allows us to dene Hamiltonian ows.
Definition 8.6. If (M, , ) is a Poisson manifold and F C

(M)
then the Hamiltonian vector eld generated by F is the vector eld X
F

X(M) such that
X
F
G = F, G
for any function G C

(M).
Proposition 8.7. The map C

(M) F X
F
X(M) is a Lie
algebra homomorphism between (C

(M), , ) and (X(M), [, ]), that is,

(i) X
F+G
= X
F
+X
G
;
(ii) X
F,G
= [X
F
, X
G
]
8. SYMMETRY AND REDUCTION 219
for all , R and F, G, H C

(M).
Proof. Property (i) is immediate from the bilinearity of the Poisson
bracket. Property (ii) arises from the Jacobi identity, as
X
F,G
H = F, G, H = F, G, H G, F, H
= X
F
(X
G
H) X
G
(X
F
H)
= [X
F
, X
G
] H
for any F, G, H C

(M).
The functions in the kernel of the homomorphism F X
F
are called
the Casimir functions, and are simply the functions F C

(M) that
Poisson commute with all other functions, that is, such that F, G = 0
for all G C

(M). Notice that Casimir functions are constant along any

Hamiltonian ow. The image of the homomorphism F X
F
is the set
of Hamiltonian vector elds, which in particular forms a Lie subalgebra of
(X(M), [, ]).
Example 8.8.
(1) If (M, ) is a symplectic manifold then the Casimir functions are
just the (locally) constant functions.
(2) If , := 0 is the trivial Poisson bracket on a smooth manifold M
then any function is a Casimir function, and the only Hamiltonian
vector eld is the zero eld.
(3) If , is the Poisson bracket dened on R
3
by the formula
F, G(x) := x, grad F grad G
then C(x) := |x|
2
is a Casimir function, as grad C = 2x and so
C, F(x) = 2x, x grad F = 2grad F, x x = 0
for any smooth function F C

(R
3
). It follows that the Hamilton-
ian vector elds are necessarily tangent to the spheres of constant
C (and in particular must vanish at the origin).
Since the Poisson bracket can be written as
F, G = X
F
G = dG(X
F
) = dF(X
G
),
we see that F, G(p) is a linear function of both (dF)
p
and (dG)
p
. Therefore
the Poisson bracket determines a bilinear map B
p
: T

p
M T

p
M R for
all p M.
Definition 8.9. The antisymmetric (0, 2)-tensor eld B satisfying
F, G = B(dF, dG)
is called the Poisson bivector.
220 5. GEOMETRIC MECHANICS
Using the identication T
p
M

= (T

p
M)

, we have
X
F
(dG) = dG(X
F
) = X
F
G = F, G = B(dF, dG) = ((dF)B)(dG),
where the contraction of a covector with the Poisson bivector is dened
in the same way as the contraction of a vector with an alternating tensor
(cf. Exercise 1.15.8 in Chapter 2). Therefore we have
X
F
= (dF)B,
and so the set of all possible values of Hamiltonian vector elds at a given
point p M is exactly the range of the map T

p
M ()B T
p
M.
Theorem 8.10. (Kirilov) Let (M, , ) be a Poisson manifold such that
the rank of the map T

p
M ()B T
p
M is constant (as a function
of p M). Then M is foliated by symplectic submanifolds (S,
S
) such that
F, G(p) = F[
S
, G[
S
(p)
for all p M, where S is the symplectic leaf containing p.
Proof. Since the rank r of the map T

p
M ()B T
p
M is
constant, the range
p
of this map has dimension r for all p M, and
so determines a distribution of dimension r in M. By construction, all
Hamiltonian vector elds are compatible with this distribution, and it is clear
that for each p M there exist F
1
, . . . , F
r
C

(M) such that is spanned

by X
F
1
, . . . , X
Fr
on a neighbourhood of p. Since [X
F
i
, X
F
j
] = X
F
i
,F
j

for
i, j = 1, . . . , r, the distribution is integrable, and so M is foliated by
r-dimensional leafs S with T
p
S =
p
for all p S. If , T

p
M then
B(, ) = (()B) = (()B) depends only on the restrictions of
and to
p
, that is, B restricts to

p
. Moreover, this restriction
is nondegenerate, since for each nonvanishing

p
there exists a vector
v = ()B
p
such that (v) ,= 0. It is then easy to check that the Poisson
parethesis determined in each leaf S by the restriction of B to T

S T

S
arises from a symplectic form on S (cf. Exercise 8.23.4).
Remark 8.11. Kirilovs Theorem still holds in the general case, where
the rank of the map T

p
M ()B T
p
M is not necessarily constant.
In this case the symplectic leaves do not necessarily have the same dimension,
and form what is called a singular foliation.
Example 8.12.
(1) If (M, ) is a symplectic manifold then there is only one symplectic
leaf (M itself).
(2) If , := 0 is the trivial Poisson bracket on a smooth manifold M
then the Poisson bivector vanishes identically and the symplectic
leaves are the zero-dimensional points.
(3) If , is the Poisson bracket dened on R
3
by the formula
F, G(x) := x, grad F grad G
8. SYMMETRY AND REDUCTION 221
then the Poisson bivector at x R
3
is given by
B(v, w) = x, v w = w, x v
for any v, w R
3
, where we use the Euclidean inner product ,
to make the identication (R
3
)

= R
3
. Therefore at x R
3
we
have
(v)B = x v,
and so the range of B at x is the tangent space to the sphere S
x
of radius |x| centered at the origin. The symplectic leaves are
therefore the spheres S
x
(including the origin, which is a singular
leaf), and if x ,= 0 the symplectic form on S
x
is given by
(v, w) =
1
|x|
2
x, v w
for v, w T
x
S
x
(that is, is
1
|x|
times the standard volume form).
Indeed, if F C

(R
3
) and v T
x
S
x
then we have
(X
F
, v) =
1
|x|
2
x, X
F
v =
1
|x|
2
v, x ((grad F)B)
=
1
|x|
2
v, x (x grad F)
=
1
|x|
2

v, x, grad Fx |x|
2
grad F
_
= v, grad F = dF(v).
Next we consider the geometric properties of Hamiltonian ows, that
is, ows generated by Hamiltonian vector elds. Just like in the symplectic
case, we have a Hamiltonian version of energy conservation.
Proposition 8.13. Hamiltonian ows preserve their generating func-
tions.
Proof. If F C

(M) then
X
F
F = F, F = F, F = 0.

Recall that in the symplectic case Hamiltonian ows preserve the sym-
plectic form. To obtain the analogue of this property in Poisson geometry
we make the following denition.
Definition 8.14. A Poisson map f : M N between two Poisson
manifolds (M, , ) and (N, , ) is a dierentiable map such that
F, G f = F f, G f
for all F, G C

(N).
As one would expect, Poisson maps preserve Hamiltonian ows.
222 5. GEOMETRIC MECHANICS
Proposition 8.15. If (M, , ) and (N, , ) are Poisson manifolds,
f : M N is a Poisson map and F C

(N) then
f

X
Ff
= X
F
.
Proof. We just have to notice that given G C

(N) we have
(f

X
Ff
) G = X
Ff
(G f) = F f, G f = F, G = X
F
G.

Finally, we show that Hamiltonian ows preserve the Poisson bracket.

Proposition 8.16. Hamiltonian ows are Poisson maps.
Proof. Let
t
: M M be the Hamiltonian ow generated by the
function F C

(M). If G C

(M) is another function we have

d
dt
(G
t
) =
d
dt
(
t

G) =
d
ds[
s=0
((
t+s
)

G) =
d
ds[
s=0
((
t

s
)

G)
=
d
ds[
s=0
(
s

(
t

G)) = X
F
(
t

G) = F,
t

G.
Given G, H C

(M), let K
t
C

(M) be the function

K
t
:= G, H
t
G
t
, H
t
=
t

G, H
t

G,
t

H.
Clearly K
0
= 0. Since the Poisson bracket is bilinear, we have
d
dt
K
t
=
d
dt
(
t

G, H)
_
d
dt
(
t

G),
t

H
_

G,
d
dt
(
t

H)
_
= X
F
(
t

G, H) F,
t

G,
t

H
t

G, F,
t

H
= X
F
(
t

G, H) F,
t

G,
t

H = X
F
K
t
.
Regarding K
t
as a function K dened on I M, where I R is the interval
of denition of
t
, we see that it satises
_
_

t
X
F
_
K = 0
K(0, p) = 0 for all p M
.
Integrating from 0M along the integral curves of

t
X
F
we then obtain
K
t
= 0 for all t I.
We are now ready to discuss symmetry and reduction.
Definition 8.17. Let G be a Lie group acting on a Poisson manifold
(M, , ). The action is said to be:
(1) Poisson if for each g G the map A
g
: M M given by A
g
(p) :=
g p is a Poisson map;
(2) Hamiltonian if for each V g there exists a function J(V )
C

(M) such that the innitesimal action X

V
is the Hamiltonian
vector eld generated by J(V ), that is, X
V
= X
J(V )
.
8. SYMMETRY AND REDUCTION 223
If G is connected then Proposition 8.16 guarantees that a Hamiltonian
action is Poisson (cf. Exercise 8.23.5). Notice that because X
V
is a linear
function of V we can take J(V ) to be a linear function of V , and thus think
of J as a map J : M g

. This map is called the momentum map for

the action.
Theorem 8.18. (Noether, Hamiltonian version) If the action of the Lie
group G on the Poisson manifold (M, , ) is Hamiltonian with momentum
map J : M g

and H C

(M) is G-invariant then J is constant along

the Hamiltonian ow of H.
Proof. Since H is G-invariant we have for any V g
X
V
H = 0 X
J(V )
H = 0 J(V ), H = 0 X
H
J(V ) = 0.

Example 8.19. The relation between the Hamiltonian and the La-
grangian versions of the Noether Theorem is made clear by the following
important example. Let M be a dierentiable manifold, and let G be a Lie
group acting on M. We can lift this action to the symplectic (hence Poisson)
manifold T

M by the formula
g = A

g
1

for all T

M, where A
g
: M M is given by A
g
(p) = g p for all p M.
It is easy to check that this formula indeed denes an action of G on T

M,
mapping each cotangent space T

p
M to T

gp
M.
Let (x
1
, . . . , x
n
) be local coordinates on M and let (x
1
, . . . , x
n
, p
1
, . . . , p
n
)
be the corresponding local coordinates on T

M. Let (y
1
, . . . , y
m
) be local
coordinates on G centered at the identity e G such that (y
1
, . . . , y
m
)
parameterizes the inverse of the element paramaterized by (y
1
, . . . , y
m
) (this
can be easily accomplished by using the exponential map). If in these coor-
dinates the action A : GM M of G on M is given by
(A
1
(x
1
, . . . , x
n
, y
1
, . . . , y
m
), . . . , A
n
(x
1
, . . . , x
n
, y
1
, . . . , y
m
))
then we have
A

g
1
_
n

i=1
p
i
dx
i
_
=
n

i,j=1
p
i
A
i
x
j
(x, y)dx
j
,
and so the lift of the action of G to T

M is written
_
A
1
(x, y), . . . , A
n
(x, y),
n

i=1
A
i
x
1
(x, y)p
i
, . . . ,
n

i=1
A
i
x
n
(x, y)p
i
_
.
Therefore the innitesimal action of V :=

m
a=1
V
a
y
a
on T

M is
n

i=1
X
i
(x)

x
i

n

i,j=1
X
j
x
i
(x)p
j

p
i
=
n

i=1
J
p
i

x
i

n

i=1
J
x
i

p
i
,
224 5. GEOMETRIC MECHANICS
where
X
i
(x) =
m

a=1
A
i
y
a
(x, 0)V
a
are the components of the innitesimal action of V on M and
J =
n

i=1
X
i
(x)p
i
.
We conclude that the lift of the action of G to T

M is Hamiltonian with
momentum map J : T

M g

given by
J()(V ) = (X
V
),
where X
V
X(M) is the innitesimal action of V on M. Notice that J(V ) is
exactly the image by the Legendre Transformation of the conserved quantity
J
V
in the Lagrangian version of the Noether theorem.
Theorem 8.20. (Poisson reduction) If the action of G on (M, , ) is
free, proper and Poisson then M/G is naturally a Poisson manifold (identi-
fying C

(M/G) with the G-invariant functions in C

(M)), and the natural

projection : M M/G is a Poisson map. In particular, carries the
Hamiltonian ow of G-invariant functions on M to the Hamiltonian ow of
the corresponding functions in M/G.
Proof. We just have to observe that the if the action is Poisson then
the Poisson bracket of G-invariant functions is G-invariant.
If G is a Lie group then G acts on G by left multiplication, and the
lift of this action to T

G is free, proper and Hamiltonian. If moreover G is

connected then the action is Poisson, and we have the following result.
Theorem 8.21. (Lie-Poisson reduction) If G is a connected Lie group
then the quotient Poisson bracket on T

G/G g

is given by
F, H() := ([dF, dH])
for all F, H C

(g

), where dF, dH g

g. If (p
1
, . . . , p
m
) are linear
coodinates on g

corresponding to the basis

1
, . . . ,
m
then
F, H =
m

a,b,c=1
p
a
C
a
bc
F
p
b
H
p
c
,
where C
a
bc
are the structure constants associated to the dual basis X
1
, . . . , X
m

of g.
Proof. If we think of
1
, . . . ,
m
as left-invariant 1-forms on G then
the canonical symplectic potential on T

G is
=
m

a=1
p
a

a
,
8. SYMMETRY AND REDUCTION 225
(for simplicity we identify
a
and

a
, where : T

G G is the natural
projection). Now from Exercise 2.8.1 in Chapter 4 we know that
d
a
=
1
2
m

b,c=1
C
a
bc

b

c
,
and so the canonical symplectic form is
= d =
m

a
dp
a

a

1
2
m

a,b,c=1
p
a
C
a
bc

b

c
=
m

a
dp
a

a
dp
a

a,b,c=1
p
a
C
a
bc

c
.
If F C

(T

G) is G-invariant then it only depends on the coordinates

(p
1
, . . . , p
m
) along the bers, and so
dF =
m

a=1
F
p
a
dp
a
.
Setting
X
F
:=
m

a=1

a
X
a
+
m

a=1

p
a
,
where X
1
, . . . , X
m
is the dual basis of g, we then have
(X
F
) =
m

a=1

a
dp
a

a,b,c=1
p
a
C
a
bc

c
+
m

a=1

a
.
From (X
F
) = dF we then obtain
X
F
=
m

a=1
F
p
a
X
a
+
m

a,b,c=1
p
a
C
a
bc
F
p
b

p
c
,
implying that if H C

(T

G) is also G-invariant then

F, H = X
F
H =
m

a,b,c=1
p
a
C
a
bc
F
p
b
H
p
c
.
Notice that as covectors on g

we have dp
a
= X
a
, and so, by denition of
the structure functions C
a
bc
,
F, H =
m

a,b,c=1
p
a

a
([X
b
, X
c
])
F
p
b
H
p
c
=
m

a,b,c=1
p
a

a
__
F
p
b
X
b
,
H
p
c
X
c
__
=
m

a
p
a

a
([dF, dH]).

226 5. GEOMETRIC MECHANICS

Example 8.22. Lie-Poisson reduction on T

SO(3) yields the Poisson

bracket
F, G(x) := x, F G
on so(3)

= (R
3
)

= R
3
, where we used Lemma 3.9 to identify so(3) with
(R
3
, ) and the Euclidean inner product , to make (R
3
)

= R
3
.
Exercises 8.23.
(1) Consider the symplectic structure on
S
2
= (x, y, z) R
3
: x
2
+y
2
+z
2
= 1
determined by the usual volume form. Compute the Hamiltonian
ow generated by the function H(x, y, z) = z.
(2) Let (M, ) be a symplectic manifold. Show that:
(a) =

n
i=1
dp
i
dx
i
if and only if x
i
, x
j
= p
i
, p
j
= 0 and
p
i
, x
j
=
ij
for i, j = 1, . . . , n.
(b) M is orientable;
(c) if M is compact then cannot be exact. (Remark: In particular if
M is compact and all closed 2-forms on M are exact then M does not admit a
symplectic structure; this is the case for all even-dimensional spheres S
2n
with
n > 1).
(3) Let (M, , ) be a Riemannian manifold,
1
(M) a 1-form and
U C

(M) a dierentiable function.

(a) Show that := +

d is a symplectic form on T

M, where
is the canonical symplectic form and : T

M M is the
natural projection ( is called a canonical symplectic form
with magnetic term).
(b) Show that the Hamiltonian ow generated by a function

H
C

(T

M) with respect to the symplectic form is given by

the equations
_

_
x
i
=


H
p
i
p
i
=

H
x
i
+
n

j=1
_

j
x
i


i
x
j
_
x
j
.
(c) The map F : T

M T

M given by
F() :=
p
for T

p
M is a ber-preserving dieomorsm. Show that F
carries the Hamiltonian ow dened in Exercise 6.15.2 to the
Hamiltonian ow of

H with respect to the symplectic form ,
where

H() :=
1
2
, +U(p)
8. SYMMETRY AND REDUCTION 227
for T

p
M. (Remark: Since the projections of the two ows on M
coincide, we see that the magnetic term can be introduced by changing either
the Lagrangian or the symplectic form).
(4) Let (M, , ) be a Poisson manifold, B the Poisson bivector and
(x
1
, . . . , x
n
) local coordinates on M. Show that:
(a) B can be written in these local coordinates as
B =
n

i,j=1
B
ij

x
i


x
j
,
where B
ij
= x
i
, x
j
for i, j = 1, . . . , n;
(b) the Hamiltonian vector eld generated by F C

(M) can be
written as
X
F
=
p

i,j=1
B
ij
F
x
i

x
j
;
(c) the components of B must satisfy
n

l=1
_
B
il
B
jk
x
l
+B
jl
B
ki
x
l
+B
kl
B
ij
x
l
_
= 0
for all i, j, k = 1, . . . , n;
(d) if , arises from a symplectic form then (B
ij
) = (
ij
)
1
;
(e) if B is nondegenerate then it arises from a symplectic form.
(5) Let G be a connected Lie group and U G a neighborhood of the
identity. Show that:
(a) G =
+
n=1
U
n
, where U
n
= g
1
g
n
[ g
1
, . . . , g
n
U;
(b) if G acts on a Poisson manifold (M, , ) and the action is
Hamiltonian then it is Poisson.
(6) Let G be a connected Lie group with a free, proper, Hamiltonian
action on a Poisson manifold (M, , ), and let H C

(M) be
G-invariant. Show that if p M/G is a xed point of

X
H

X(M/G) (where : M M/G is the quotient map) then the ow
of X
H
on
1
(p) is given by orbits of 1-parameter subgroups of G.
(7) The Lie group SO(2) S
1
acts on M = R
2
(0, 0) through
e
i
(r, ) = (r, +),
where (r, ) are polar coordinates on M and , should be under-
stood mod 2.
(a) Write an expression for the innitesimal action X
V
X(M)
of V so(2)

= R.
(b) Determine the momentum map for the lift of this action to the
cotangent bundle.
(c) Write an expression for the Poisson bivector of T

M with
the canonical symplectic structure in the usual coordinates
(r, , p
r
, p

).
228 5. GEOMETRIC MECHANICS
(d) Calculate the Poisson bivector of the Poisson manifold Q :=
T

M/SO(2) R
3
. What are the symplectic leaves of this
manifold? Give an example of a nonconstant Casimir function.
(e) Consider the Hamiltonian H : T

Q R given by
H(r, , p
r
, p

) =
p
r
2
2
+
p

2
2r
2
+u(r).
Show that H is SO(2)-invariant, and determine its Hamilton-
ian ow on the reduced Poisson manifold Q.
(f) Use Noethers Theorem to obtain a quantity conserved by the
Hamiltonian ow of H on T

M.
(8) Recall that the upper half plane H = (x, y) R
2
: y > 0 has a
Lie group structure, given by the operation
(x, y) (z, w) := (yz +x, yw),
and that the hyperbolic plane corresponds to the left-invariant met-
ric
g :=
1
y
2
(dx dx +dy dy)
on H (cf. Exercise 7.17.3 in Chapter 1 and Exercise 3.3.5 in Chap-
ter 3). The geodesics are therefore determined by the Hamiltonian
function K : T

H R given by
K(x, y, p
x
, p
y
) =
y
2
2
_
p
x
2
+p
2
y
_
.
(a) Determine the lift to T

H of the action of H on itself by left

translation, and check that it preserves the Hamiltonian K.
(b) Show that the functions
F(x, y, p
x
, p
y
) = yp
x
and G(x, y, p
x
, p
y
) = yp
y
are also H-invariant, and use this to obtain the quotient Pois-
son structure on T

H/H. Is this a symplectic manifold?

(c) Write an expression for the momentum map for the action of
H on T

H, and use it to obtain a nontrivial rst integral I

of the geodesic equations. Show that the projection on H of
a geodesic for which K = E, p
x
= l and I = m satises the
equation
l
2
x
2
+l
2
y
2
2lmx +m
2
= 2E.
Assuming l ,= 0, what are these curves?
(9) Recall that the Euler top is the mechanical system determined by
the Lagrangian L : TSO(3) R given by
L =
1
2
I, ,
8. SYMMETRY AND REDUCTION 229
where are the left-invariant coordinates on the bers resulting
from the identications
T
S
SO(3) = L
S
(so(3))

= so(3)

= R
3
.
(a) Show that if we use the Euclidean inner product , to iden-
tify (R
3
)

with R
3
then the Legendre transformation is written
P = I,
where P are the corresponding left-invariant coordinates on
T

SO(3).
(b) Write the Hamilton equations on the reduced Poisson manifold
T

SO(3)/SO(3)

= R
3
. What are the symplectic leaves? Give
an example of a nonconstant Casimir function.
(c) Compute the momentum map for the lift to T

SO(3) of the
action of SO(3) on itself by left translation.
(10) Let (P
1
, P
2
, P
3
) be the usual left-invariant coordinates on the bers
of T

SO(3), and consider the functions (

1
,
2
,
3
) dened through
= S
for each S SO(3), where R
3
is a xed vector. Show that for
i, j = 1, 2, 3:
(a) P
i
, P
j
=

3
k=1

ijk
P
k
;
(b)
i
,
j
= 0;
(c) P
i
,
j
=

3
k=1

ijk

k
,
where

ijk
=
_
_
_
+1 if (i, j, k) is an even permutation of (1, 2, 3)
1 if (i, j, k) is an odd permutation of (1, 2, 3)
0 otherwise.
(Hint: Show that

= along any motion of the Euler top, where is the
angular velocity in the Euler tops frame, and regard
(P
i
)
2
2
as the limit of the Euler
top Hamiltonian when I
i
= 1 and I
j
+ for j ,= i).
(11) If on Exercise 10 we set = ge
z
, where g is the (constant) gravi-
tational acceleration, then the motion of a rigid body (with a xed
point) of mass M and moment of inertia I, whose center of mass
has position vector L R
3
in its frame, is given by the Hamiltonian
ow of
H =
1
2
P, I
1
P +M, SL.
(a) Show that H is S
1
-invariant for the lift to T

SO(3) of the
action of S
1
on SO(3) determined by e
i
S = R

S, where
R

:=
_
_
cos sin 0
sin cos 0
0 0 1
_
_
(corresponding to rotations about the z-axis).
230 5. GEOMETRIC MECHANICS
(b) Determine the momentum map for this action.
(c) Show that the functions P and are S
1
-invariant, and that
the Poisson bracket on the quotient manifold T

SO(3)/S
1
=
(SO(3)/S
1
) R
3
= S
2
R
3
is determined by the Poisson
brackets of these functions.
(d) Use the functions P and to write the equations of motion on
the quotient, and give an example of a nonconstant Casimir
function.
9. Notes on Chapter 5
9.1. Section 1. Throughout this chapter, starting at the exercises of
Section 1, we need several denitions and facts related to stability of xed
points of vector elds in R
n
(refer for instance to [Arn92, GH02] for addi-
tional details). In order to study nonlinear systems
(25)
_
x = f(x)
x(0) = x
0
(x R
n
)
one usually starts by nding the zeros of f, called xed points, equilib-
ria or stationary solutions. A xed point x is called stable if for each
neighborhood U of x there exists another (possibly smaller) neighborhood
V of x such that if x
0
V then x(t) U for each t > 0 where the solution
is dened. The behavior of solutions near x can, in many situations, be
studied by linearizing (25) at x and analyzing the resulting (linear) system
(26)
_

= A
(0) =
0
( R
n
)
where A := (df)
x
. This linear system has a global solution
(
0
, t) = e
tA

0
,
where e
tA
can be seen as a map from R
n
to R
n
dening the ow of the vector
eld A. If we put A in the Jordan canonical form then it is clear that this
ow has the following invariant subspaces:
E
s
:= spanv
1
, . . . , v
ns
(stable subspace);
E
u
:= spanu
1
, . . . , u
nu
(unstable subspace);
E
c
:= spanw
1
, . . . , w
nc
(center subspace),
where v
1
, . . . , v
ns
are the n
s
generalized eigenvectors corresponding to eigen-
values with negative real part, u
1
, . . . , u
nu
are the n
u
generalized eigenvectors
corresponding to eigenvalues with positive real part, and w
1
, . . . , w
ns
are the
n
c
generalized eigenvectors corresponding to eigenvalues with zero real part.
If E
c
= then x is called a hyperbolic or nondegenerate xed point
of f. In this case the Hartman-Grobman Theorem tells us that there
exists a homeomorphism from a neighborhood of x in R
n
to a neighborhood
of 0 in R
n
which takes the orbits of the non-linear ow of (25) to those of
9. NOTES ON CHAPTER 5 231
the linear ow e
tA
of (26). The asymptotic behavior of solutions near x,
and consequently its stability type, is then determined by the eigenvalues
of A.
9.2. Section 3. In this Section we use several denitions and results
about measure theory, which we will now briey discuss (see [Rud86] for
further details).
(1) A collection / of subsets of a set X is said to be a -algebra on
X if it satises the following properties:
(a) X /;
(b) if A / then X A /;
(c) if A =
+
i=1
A
i
, with A
i
/ for every i, then A /.
If / is a -algebra on X then the elements of / are called the
measurable sets.
(2) A positive measure is a (nonconstant) function : /[0, +]
which is countably additive, i.e. such that if A
i
is a countable
family of disjoint measurable sets, then

_
+
_
i=1
A
i
_
=
+

i=1
(A
i
).
If (A) < + for every set A / then is called a nite
measure. A measure space (X, /, ) is then a set X equipped
with a -algebra / and a positive measure .
(3) (Dirac measure) A very simple example of a positive measure is
the Dirac measure
x
0
which is dened as follows: given x
0
X,

x
0
(A) :=
_
1 if x
0
A
0 if x
0
/ A
for any subset A X.
(4) (Lebesgue measure) Let X = R
n
. For any A R
n
one denes the
outer measure

(A) of A as

(A) := inf

I(
vol(I)
where the inmum is taken over all countable collections of intervals
whose union contains A. Then one denes A to be Lebesgue
measurable if, for every S R
n
,

(S) =

(S A) +

(S A).
The Lebesgue measurable sets form a -algebra, and the Lebesgue
measure of a Lebesgue measurable set A is dened as (A) :=

(A).
(5) If X is a measure space and Y is a topological space then a function
f : X Y is called measurable if f
1
(E) is a measurable set in
X for every open set E Y .
232 5. GEOMETRIC MECHANICS
(6) (Integration of positive functions) A nonnegative measurable func-
tion s on a measure space X whose image consists of only nitely
many points is called simple. If a
1
, . . . , a
n
are the distinct nonzero
values of s and A / we dene the integral of s over A as
_
A
s d :=
n

i=1
a
i
(A s
1
(a
i
)).
If f : X [0, +] is measurable and A / one denes the
Lebesgue integral of f over A with respect to the positive mea-
sure as
_
A
fd := sup
_
A
s d,
where the supremum is taken over all simple functions such that
0 s f.
(7) (Integration of real functions) Given a positive measure one de-
nes the set /
1
() of Lebesgue integrable functions as the set
of all real measurable functions f on X for which
_
X
[f[ d < +.
If A / then we dene
_
A
f :=
_
A
f
+

_
A
f

,
where
f
+
:= maxf, 0 and f

:= minf, 0,
are the positive and negative parts of f.
(8) A measurable set A X is said to be a zero measure set if
(A) = 0. A full-measure set is a set whose complement is a
zero-measure set.
9.3. Bibliographical notes. The material in this chapter follows [Oli02]
and [Arn97] closely. There are of course many other excellent books on Me-
chanics, both traditional [GPS02] and geometric [AM78, MR99]. Non-
holonomic systems (including control theory) are treated in greater detail
in [Blo03, BL05]. For more information on completely integrable systems
see [CB97, Aud96].
CHAPTER 6
Relativity
This chapter studies one of the most important applications of Riemann-
ian geometry, namely general relativity.
Section 1 discusses Galileo spacetime, the geometric structure under-
lying Newtonian mechanics, which hinges on the existence of arbitrarily fast
motions; if, however, a maximum speed is assumed to exist, then it must be
replaced by Minkowski spacetime, whose geometry is studied in special
relativity (Section 2).
Section 3 shows how to include Newtonian gravity in Galileo spacetime
by introducing the symmetric Cartan connection. Trying to general-
ize this procedure leads to general Lorentzian manifolds satisfying the
Einstein eld equation, of which Minkowski spacetime is the simplest
example (Section 4).
Other simple solutions are analyzed in the subsequent sections: the
Schwarzschild solution, modeling the gravitational eld outside spher-
ically symmetric bodies or black holes (Section 5), and the Friedmann-
Lematre-Robertson-Walker models of cosmology, describing the be-
havior of the Universe as a whole (Section 6).
This chapter concludes with a discussion of the causal structure of a
Lorentz manifold (Section 7), in preparation for the proof of the Singularity
Theorems of Hawking (Section 8) and Penrose (Section 9).
1. Galileo Spacetime
The set of all physical occurrences can be modeled as a connected 4-
dimensional manifold M, which we call spacetime, and whose points we
refer to as events. We assume that M is dieomorphic to R
4
, and that
there exists a special class of dieomorphisms x : M R
4
, called iner-
tial frames. An inertial frame yields global coordinates (x
0
, x
1
, x
2
, x
3
) =
(t, x, y, z). We call the coordinate t : M R the time function associated
to a given inertial frame. Two events p, q M are said to be simultaneous
on that frame if t(p) = t(q). The level functions of the time function are
therefore called simultaneity hypersurfaces. The distance between two
simultaneous events p, q M is given by
d(p, q) =

_
3

i=1
(x
i
(p) x
i
(q))
2
.
233
234 6. RELATIVITY
The motion of a particle is modeled by a smooth curve c : I M such
that dt( c) ,= 0. A special class of motions is formed by the motions of free
particles, i.e. particles which are not acted upon by any external force.
The special property that inertial frames have to satisfy is that the motions
of free particles are always represented by straight lines. In other words, free
particles move with constant velocity relative to inertial frames (Newtons
law of inertia). In particular, motions of particles at rest in an inertial
frame are motions of free particles.
Inertial frames are not unique: if x : M R
4
is an inertial frame and
T : R
4
R
4
is an invertible ane transformation then T x is another
inertial frame. In fact, any two inertial frames must be related by such an
ane transformation (cf. Exercise 1.1.3).
The Galileo spacetime, which underlies Newtonian mechanics, is ob-
tained by further requiring that inertial frames should:
(1) agree on the time interval between any two events (and hence on
whether two given events are simultaneous);
(2) agree on the distance between simultaneous events.
Therefore, up to translations and reections, all coordinate transfor-
mations between inertial frames belong to the Galileo group Gal(4), the
group of linear orientation-preserving maps which preserve time functions
and the Euclidean structures of the simultaneity hypersurfaces.
When analyzing problems in which only one space dimension is impor-
tant, we can use a simpler 2-dimensional Galileo spacetime. If (t, x) are
the spacetime coordinates associated to an inertial frame and T Gal(2)
is a Galileo change of basis to a new inertial frame with global coordinates
(t

, x

), then

:= T
_

t
_
=

t
+v

x

:= T
_

x
_
=

x
with v R, since we must have t = t

, and so
dt
_

t

_
= dt

_

t

_
= 1,
and we want the orientation-preserving map T to be an isometry of the
simultaneity hypersurface t = 0 t

= 0. The change of basis matrix

is then
S =
_
1 0
v 1
_
,
with inverse
S
1
=
_
1 0
v 1
_
.
2. SPECIAL RELATIVITY 235
Therefore the corresponding coordinate transformation is
_
t

= t
x

= x vt
(v R)
(Galileo transformation), and hence the new frame is moving with veloc-
ity v with respect to the old one (as the curve x

= 0 is the curve x = vt).

Notice that S
1
is obtained from S simply by reversing the sign of v, as one
would expect, as the old frame must be moving relative to the new one with
velocity v. We shall call this observation the Relativity Principle.
Exercises 1.1.
(1) (Lucas problem) By the late 19
th
century there existed a regular
transatlantic service between Le Havre and New York. Every day
at noon (GMT) a transatlantic ship would depart Le Havre and
another one would depart New York. The journey took exactly
seven days, so that arrival would also take place at noon (GMT).
Therefore, a transatlantic ship traveling from Le Havre to New
York would meet a transatlantic ship just arriving from New York
at departure, and another one just leaving New York on arrival.
Besides these, how many other ships would it meet? At what times?
What was the total number of ships needed for this service? (Hint:
Represent the ships motions as curves in a 2-dimensional Galileo spacetime).
(2) Check that free particles move with constant velocity relative to
inertial frames.
(3) Let f : R
n
R
n
(n 2) be a bijection that takes straight lines to
straight lines. Show that f must be an ane function, i.e., that
f(x) = Ax +b
for all x R
n
, where A GL(n, R) and b R
n
.
(4) Prove that the Galileo group Gal(4) is the subset of GL(4, R)
formed by matrices of the form
_
1 0
v R
_
where v R
3
and R SO(3). Conclude that Gal(4) is isomorphic
to the group of orientation-preserving isometries of the Euclidean
3-space R
3
.
(5) Show that Gal(2) is a subgroup of Gal(4).
2. Special Relativity
The Galileo spacetime requirement that all inertial observers should
agree on the time interval between two events is intimately connected with
the possibility of synchronizing clocks in dierent frames using signals of
arbitrarily high speeds. Experience reveals that this is actually impossible.
Instead, there appears to exist a maximum propagation speed, the speed of
236 6. RELATIVITY
light (approximately 300,000 kilometers per second), which is the same at
all events and in all directions. A more accurate requirement is then that
any two inertial frames should
(1) agree on whether a given particle is moving at the speed of light.
Notice that we no longer require that dierent inertial frames should
agree on the time interval between two events, or even on whether two given
events are simultaneous. However we still require that any two inertial
frames should
(2) agree on the distance between events which are simultaneous in both
frames.
It is convenient to choose units such that the speed of light is 1 (for in-
stance measuring time in years and distance in light-years). Fix a particular
inertial frame with coordinates (x
0
, x
1
, x
2
, x
3
). A free particle moving at the
speed of light on an inertial frame x : R
4
R will be a straight line whose
tangent vector
v = v
0

x
0
+v
1

x
1
+v
2

x
2
+v
3

x
3
must satisfy
(v
0
)
2
= (v
1
)
2
+ (v
2
)
2
+ (v
3
)
2
, (27)
so that the distance travelled equals the elapsed time. In other words, v
must satisfy v, v = 0, where
v, w := v
0
w
0
+v
1
w
1
+v
2
w
2
+v
3
w
3
=
3

,=0

,
with (

) = diag(1, 1, 1, 1). Notice that , is a symmetric non-degenerate

tensor which is not positive denite; we call it the Minkowski (pseudo)
inner product. The coordinate basis
_

x
0
,

x
1
,

x
2
,

x
3
_
is an orthonormal basis for this inner product (cf. Exercise 2.2.1), as
_

x

,

x

_
=

(, = 0, 1, 2, 3).
Since we used a particular inertial frame to dene the Minkowski inner
product, we must now check that it is well dened (i.e., it is independent
of the inertial frame we chose to dene it). Let (x
0
, x
1
, x
2
, x
3
) be the
coordinates associated to another inertial frame. The analogue of (27) on
the new inertial frame implies that the vectors

x
0


x
i
2. SPECIAL RELATIVITY 237
(i = 1, 2, 3) must be tangent to a motion at the speed of light. By assump-
tion (1), given a motion of a free particle at the speed of light, all inertial
observers must agree that the particle is moving at this (maximum) speed.
Therefore we must have
_

x
0


x
i
,

x
0


x
i
_
= 0.
This implies that
_

x
0
,

x
0
_
=
_

x
i
,

x
i
_
;
_

x
0
,

x
i
_
= 0.
Similarly, we must have
_

2

x
0
+

x
i
+

x
j
,

2

x
0
+

x
i
+

x
j
_
= 0
(i ,= j), and hence
_

x
i
,

x
j
_
= 0.
Since , is non-degenerate, we conclude that there must exist k ,= 0 such
that
_

x

,

x

_
= k

(, = 0, 1, 2, 3).
The simultaneity hypersurfaces x
0
= const. and x
0
= const. are 3-
planes in R
4
. If they are parallel, they will coincide for appropriate values of
the constants; otherwise, they must intersect along 2-planes of events which
are simultaneous in both frames. In either case there exist events which are
simultaneous in both frames. Let v ,= 0 be a vector connecting two such
events. Then dx
0
(v) = dx
0
(v) = 0, and hence
v =
3

i=1
v
i

x
i
=
3

i=1
v
i

x
i
.
By assumption (2), we must have
3

i=1
_
v
i
_
2
=
3

i=1
_
v
i
_
2
.
Consequently, from
3

i=1
_
v
i
_
2
= v, v =
_
3

i=1
v
i

x
i
,
3

i=1
v
i

x
i
_
= k
3

i=1
_
v
i
_
2
,
238 6. RELATIVITY
we conclude that we must have k = 1. Therefore the coordinate basis
_

x
0
,

x
1
,

x
2
,

x
3
_
must also be an orthonormal basis. In particular, this means that the
Minkowski inner product , is well dened (i.e., it is independent of the in-
ertial frame we choose to dene it), and that we can identify inertial frames
with orthonormal bases of (R
4
, , ).
Definition 2.1. (R
4
, , ) is said to be the Minkowski spacetime.
The length of a vector v R
4
is [v[ = [v, v[
1
2
.
The study of the geometry of Minkowski spacetime is usually called
special relativity. A vector v R
4
is said to be:
(1) timelike if v, v < 0; in this case, there exists an inertial frame
(x
0
, x
1
, x
2
, x
3
) such that
v = [v[

x
0
(cf. Exercise 2.2.1), and consequently any two events p and p + v
occur on the same spatial location in this frame, separated by a
time interval [v[;
(2) spacelike if v, v > 0; in this case, there exists an inertial frame
(x
0
, x
1
, x
2
, x
3
) such that
v = [v[

x
1
(cf. Exercise 2.2.1), and consequently any two events p and p + v
occur simultaneously in this frame, a distance [v[ apart;
(3) lightlike, or null, if v, v = 0; in this case any two events p and
p+v are connected by a motion at the speed of light in any inertial
frame.
The set of all null vectors is called the light cone, and it is in a way the
structure that replaces the absolute simultaneity hypersurfaces of Galileo
spacetime. It is the boundary of the set of all timelike vectors, which has
two connected components; we represent by C(v) the connected component
that contains a given timelike vector v. A time orientation for Minkowski
spacetime is a choice of one of these components, whose elements are said
to be future-pointing; this is easily extended to nonzero null vectors.
An inertial frame (x
0
, x
1
, x
2
, x
3
) determines a time orientation, namely
that for which the future-pointing timelike vectors are the elements of C
_

x
0
_
.
Up to translations and reections, all coordinate transformations between in-
ertial frames belong to the (proper) Lorentz group SO
0
(3, 1), the group of
linear maps which preserve orientation, time orientation and the Minkowski
inner product (hence the light cone).
A curve c : I R R
4
is said to be timelike if c, c < 0. Timelike
curves represent motions of particles with nonzero mass, since only for these
2. SPECIAL RELATIVITY 239
curves it is possible to nd an inertial frame in which the particle is instanta-
neously at rest. In other words, massive particles must always move at less
than the speed of light (cf. Exercise 2.2.13). The proper time measured
by the particle between events c(a) and c(b) is
(c) :=
_
b
a
[ c(s)[ds.
p
null vector
timelike future-pointing vector
spacelike vector

y
Figure 1. Minkowski geometry (traditionally represented
with the t-axis pointing upwards).
When analyzing problems in which only one space dimension is impor-
tant, we can use a simpler 2-dimensional Minkowski spacetime. If (t, x) are
the spacetime coordinates associated to an inertial frame and T SO
0
(1, 1)
is a Lorentzian change of basis to a new inertial frame with global coordi-
nates (t

, x

), we must have

:= T
_

t
_
= cosh u

t
+ sinhu

x

:= T
_

x
_
= sinh u

t
+ cosh u

x
with u R (cf. Exercise 2.2.3). The change of basis matrix is
S =
_
cosh u sinh u
sinhu cosh u
_
,
with inverse
S
1
=
_
cosh u sinhu
sinh u cosh u
_
.
240 6. RELATIVITY
Therefore the corresponding coordinate transformation is
_
t

= t cosh u xsinh u
x

= xcosh u t sinhu
(Lorentz transformation), and hence the new frame is moving with ve-
locity v = tanh u with respect to the old one (as the curve x

= 0 is the curve
x = vt; notice that [v[ < 1). The matrix S
1
is obtained from S simply
by reversing the sign of u, or, equivalently, of v; therefore, the Relativity
Principle still holds for Lorentz transformations.
Moreover, since
_
cosh u =
_
1 v
2
_

1
2
sinh u = v
_
1 v
2
_

1
2
,
one can also write the Lorentz transformation as
_
t

=
_
1 v
2
_

1
2
t v
_
1 v
2
_

1
2
x
x

=
_
1 v
2
_

1
2
x v
_
1 v
2
_

1
2
t
.
In everyday life situations, we deal with frames whose relative speed is much
smaller that the speed of light, [v[ 1, and with events for which [x[ [t[
(distances traveled by particles in one second are much smaller that 300,000
kilometers). An approximate expression for the Lorentz transformations in
these situations is then
_
t

= t
x

= x vt
which is just a Galileo transformation. In other words, the Galileo group is
a convenient low-speed approximation of the Lorentz group.
Suppose that two distinct events p and q occur in the same spatial loca-
tion in the inertial frame (t

, x

),
q p = t

= t

cosh u

t
+ t

sinhu

x
= t

t
+ x

x
.
We see that the time separation between the two events in a dierent inertial
frame (t, x) is bigger,
t = t

cosh u > t

.
Loosely speaking, moving clocks run slower when compared to stationary
ones (time dilation).
If, on the other hand, two distinct events p and q occur simultaneously
in the inertial frame (t

, x

),
q p = x

= x

sinhu

t
+ x

cosh u

x
= t

t
+ x

x
,
then they will not be simultaneous in the inertial frame (t, x), where the
time dierence between them is
t = x

sinhu ,= 0
2. SPECIAL RELATIVITY 241
(relativity of simultaneity).
Finally, consider two particles at rest in the inertial frame (t

, x

). Their
motions are the lines x

= x

0
and x

= x

0
+ l

. In the inertial frame (t, x),

these lines have equations
x =
x

0
cosh u
+vt and x =
x

0
+l

cosh u
+vt,
which describe motions of particles moving with velocity v and separated
by a distance
l =
l

cosh u
< l

.
Loosely speaking, moving objects shrink in the direction of their motion
(length contraction).
Exercises 2.2.
(1) Let , be a nondegenerate symmetric 2-tensor on an n-dimensional
vector space V . Show that there always exists an orthonor-
mal basis v
1
, . . . , v
n
, i.e. a basis such that v
i
, v
j
=
ij
, where

ii
= 1 and
ij
= 0 for i ,= j. Moreover, show that s =

n
i=1

ii
(known as the signature of , ) does not depend on the choice
of orthonormal basis.
(2) Consider the Minkowski inner product , on R
4
with the standard
time orientation.
(a) Let v R
4
be timelike and future-pointing. Show that:
(i) if w R
4
is timelike or null and future-pointing then
v, w < 0;
(ii) if w R
4
is timelike or null and future-pointing then
v +w is timelike and future-pointing;
(iii) v

:= w R
4
[ v, w = 0 is a hyperplane contain-
ing only spacelike vectors (and the zero vector).
(b) Let v R
4
be null and future-pointing. Show that:
(i) if w R
4
is timelike or null and future-pointing then
v, w 0, with equality if and only if w = v for some
> 0;
(ii) if w R
4
is timelike or null and future-pointing then
v +w is timelike or null and future-pointing, being null
if and only if w = v for some > 0;
(iii) v

is a hyperplane containing only spacelike and null

vectors, all of which are multiples of v.
(c) Let v R
4
be spacelike. Show that v

is a hyperplane
containing timelike, null and spacelike vectors.
(3) Show that if (t, x) are the spacetime coordinates associated to an
inertial frame and T SO
0
(1, 1) is a Lorentzian change of basis to
242 6. RELATIVITY
a new inertial frame with global coordinates (t

, x

), we must have

= T
_

t
_
= cosh u

t
+ sinhu

x

= T
_

x
_
= sinhu

t
+ cosh u

x
for some u R.
(4) (Twin paradox) Twins Alice and Bob part on their 20
th
anniversary:
while Alice stays on the Earth (which is approximately an inertial
frame), Bob leaves at 80% of the speed of light towards Planet
X, 8 light-years away from the Earth, which he therefore reaches
10 years later (as measured in the Earths frame). After a short
stay, Bob returns to the Earth, again at 80% of the speed of light.
Consequently, Alice is 40 years old when they meet again.
(a) How old is Bob at this meeting?
(b) How do you explain the asymmetry in the twins ages? Notice
that, from Bobs point of view, he is the one who is stationary,
while the the Earth moves away and back again.
(c) Imagine that each twin has a very powerful telescope. What
does each of them see? In particular, how much time elapses
for each of them as they see their twin experiencing one year?
(Hint: Notice that light rays are represented by null lines, i.e. lines whose
tangent vector is null; therefore, if event p in Alices history is seen by Bob at
event q then there must exist a future-directed null line connecting p to q).
(5) (Car and garage paradox) A 5-meter long car moves at 80% of light
speed towards a 4-meter long garage with doors at both ends.
(a) Compute the length of the car in the garages frame, and show
that if the garage doors are closed at the right time the car
will be completely inside the garage for a few moments.
(b) Compute the garages length in the cars frame, and show that
in this frame the car is never completely inside the garage.
How do you explain this apparent contradiction?
(6) Let (t

, x

) be an inertial frame moving with velocity v with respect

to the inertial frame (t, x). Prove the velocity addition formula:
if a particle moves with velocity w

in the frame (t

, x

), the particles
velocity in the frame (t, x) is
w =
w

+v
1 +w

v
.
What happens when w

= 1?
(7) (Hyperbolic angle)
(a) Show that
(i) so(1, 1) =
__
0 u
u 0
_
[ u R
_
;
2. SPECIAL RELATIVITY 243
(ii) exp
_
0 u
u 0
_
=
_
cosh u sinhu
sinh u cosh u
_
:= S(u);
(iii) S(u)S(u

) = S(u +u

).
(b) Consider the Minkowski inner product , on R
2
with a given
time orientation. If v, w R
2
are unit timelike future-pointing
vectors then there exists a unique u R such that w = S(u)v
(which we call the hyperbolic angle between v and w). Show
that:
(i) [u[ is the length of the curve formed by all unit timelike
vectors between v and w;
(ii)
1
2
[u[ is the area of the region swept by the position vector
of the curve above;
(iii) hyperbolic angles are additive;
(iv) the velocity addition formula of Exercise 6 is simply the
formula for the hyperbolic tangent of a sum.
(8) (Generalized twin paradox) Let p, q R
4
be two events connected
by a timelike straight line l. Show that the proper time between
p and q measured along l is bigger than the proper time between
p and q measured along any other timelike curve connecting these
two events. In other words, if an inertial observer and a (necessar-
ily) accelerated observer separate at a given event and are rejoined
at a later event, then the inertial observer always measures a big-
ger (proper) time interval between the two events. In particular,
prove the reversed triangle inequality: if v, w R
4
are timelike
vectors with w C(v) then [v +w[ [v[ +[w[.
(9) (Doppler eect) Use the spacetime diagram in Figure 2 to show that
an observer moving with velocity v away from a source of light of
period T measures the period to be
T

= T
_
1 +v
1 v
(Remark: This eect allows astronomers to measure the radial velocity of stars and
galaxies relative to the Earth).
(10) (Aberration) Suppose that the position in the sky of the star Sir-
ius makes an angle with the x-axis of a given inertial observer.
Show that the angle

measured by a second inertial observer mov-

ing with velocity v = tanh u along the x-axis of the rst observer
satises
tan

=
sin
cosh ucos + sinhu
.
(11) Minkowski geometry can be used in many contexts. For instance,
let l = R

t
represent the motion of an observer at rest in the
atmosphere and choose units such that the speed of sound is 1.
(a) Let : R
4
R the map such that (p) is the t coordinate of
the event in which the observer hears the sound generated at
244 6. RELATIVITY
t
x
x = vt
T
T

Figure 2. Doppler eect.

p. Show that the level surfaces of are the conical surfaces

1
(t
0
) =
_
p R
4
[ t
0

t
p is null and future-pointing
_
.
(b) Show that c : I R
4
represents the motion of a supersonic
particle i
_
c,

t
_
,= 0 and c, c > 0.
(c) Argue that the observer hears a sonic boom whenever c is tan-
gent to a surface = constant. Assuming that c is a straight
line, what does the observer hear before and after the boom?
(12) Let c : R R
4
be the motion of a particle in Minkowski spacetime
parameterized by the proper time .
(a) Show that
c, c = 1
and
c, c = 0.
Conclude that c is the particles acceleration as measured in
the particles instantaneous rest frame, i.e., in the inertial
frame (t, x, y, z) for which c =

t
. For this reason, c is called
the particles proper acceleration, and [ c[ is interpreted as
the acceleration measured by the particle.
(b) Compute the particles motion assuming that it is moving
along the x-axis and measures a constant acceleration [ c[ = a.
(c) Consider a spaceship launched from the Earth towards the
center of the Galaxy (at a distance of 30,000 light-years) with
a = g, where g represents the gravitational acceleration at the
surface of the Earth. Using the fact that g 1 year
1
in units
such that c = 1, compute the proper time measured aboard
3. THE CARTAN CONNECTION 245
the spaceship for this journey. How long would the journey
take as measured from the Earth?
(13) (The faster-than-light missile) While conducting a surveillance mis-
sion on the home planet of the wicked Klingons, the Enterprise un-
covers their evil plan to build a faster-than-light missile and attack
the Earth, 12 light-years away. Captain Kirk immediately orders
the Enterprise back to the Earth at its top speed (
12
13
of the speed
of light), and at the same time sends out a radio warning. Unfor-
tunately, it is too late: eleven years later (as measured by them),
the Klingons launch their missile, moving at 12 times the speed of
light. Therefore the radio warning, traveling at the speed of light,
reaches the Earth at the same time as the missile, twelve years after
its emission, and the Enterprise arrives at the ruins of the Earth
one year later.
(a) How long does the Enterprises trip take according to its crew?
(b) On the Earths frame, let (0, 0) be the (t, x) coordinates of the
event in which the Enterprise sends the radio warning, (11, 0)
the coordinates of the missiles launch, (12, 12) the coordi-
nates of the Earths destruction and (13, 12) the coordinates
of the Enterprises arrival at the Earths ruins (cf. Figure 3).
Compute the (t

, x

) coordinates of the same events on the

Enterprises frame.
(c) Plot the motions of the Enterprise, the Klingon planet, the
Earth, the radio signal and the missile on the Enterprises
frame. Does the missile motion according to the Enterprise
crew make sense?
(Remark: This exercise is based on an exercise in [TW92]).
3. The Cartan Connection
Let (x
0
, x
1
, x
2
, x
3
) = (t, x, y, z) be an inertial frame on Galileo spacetime,
which we can therefore identify with R
4
. Recall that Newtonian gravity
is described by a gravitational potential : R
4
R. This potential
determines the motions of free-falling particles through
d
2
x
i
dt
2
=

x
i
(i = 1, 2, 3), and is, in turn, determined by the matter density function
: R
4
R through the Poisson equation

x
2
+

2

y
2
+

2

z
2
= 4
(we are using units in which Newtons universal gravitation constant G
is set equal to 1). The vacuum Poisson equation (corresponding to the
case in which all matter is concentrated on singularities of the gravitational
246 6. RELATIVITY
t
x (0, 0)
(11, 0)
(12, 12)
(13, 12)
Planet Klingon Earth
Enterprise
missile
radio signal
Figure 3. Faster-than-light missile.
potential) is the well known Laplace equation

x
2
+

2

y
2
+

2

z
2
= 0.
Notice that the equation of motion is the same for all particles, regard-
less of their mass. This observation, dating back to Galileo, was made into
the so-called Equivalence Principle by Einstein. It implies that a grav-
itational eld determines special curves on Galileo spacetime, namely the
motions of free-falling particles. These curves are the geodesics of a sym-
metric connection, known as the Cartan connection, dened through the
nonvanishing Christoel symbols

i
00
=

x
i
(i = 1, 2, 3)
(cf. Exercise 3.1.1), corresponding to the nonvanishing connection forms

i
0
=

x
i
dt.
It is easy to check that the Cartan structure equations

= d

+
3

=0

4. GENERAL RELATIVITY 247

still hold for arbitrary symmetric connections, and hence we have the non-
vanishing curvature forms

i
0
=
3

j=1

x
j
x
i
dx
j
dt.
The Ricci curvature tensor of this connection is
Ric =
_

x
2
+

2

y
2
+

2

z
2
_
dt dt
(cf. Exercise 3.1.2), and hence the Poisson equation can be written as
Ric = 4 dt dt.
In particular, the Laplace equation can be written as
Ric = 0.
Exercises 3.1.
(1) Check that the motions of free-falling particles are indeed geodesics
of the Cartan connection. What other geodesics are there? How
would you interpret them?
(2) Check the formula for the Ricci curvature tensor of the Cartan
connection.
(3) Show that the Cartan connection is compatible with Galileo
structure, i.e., show that:
(a)
X
dt = 0 for all X X(R
4
) (cf. Exercise 2.6.3 in Chapter 3);
(b) if E, F X(R
4
) are tangent to the simultaneity hypersurfaces
and parallel along some curve c : R R
4
, then E, F is
constant.
(4) Show that if the Cartan connection has nonzero curvature then it
is not the Levi-Civita connection of any pseudo-Riemannian metric
on R
4
(cf. Section 4).
4. General Relativity
Gravity can be introduced in Newtonian mechanics through the sym-
metric Cartan connection, which preserves Galileo spacetime structure. A
natural idea for introducing gravity in special relativity is then to search for
symmetric connections preserving the Minkowski inner product. To formal-
ize this, we introduce the following denition.
Definition 4.1. A pseudo-Riemannian manifold is a pair (M, g),
where M is a connected n-dimensional dierentiable manifold and g is a
symmetric nondegenerate dierentiable 2-tensor eld (g is said to be a pseudo-
Riemannian metric in M). The signature of a pseudo-Riemannian
manifold is just the signature of g at any tangent space. A Lorentzian
manifold is a pseudo-Riemannian manifold with signature n 2.
248 6. RELATIVITY
The Minkowski spacetime (R
4
, , ) is obviously a Lorentzian mani-
fold. It is easily seen that the Levi-Civita Theorem still holds for pseudo-
Riemannian manifolds. In other words, given a pseudo-Riemannian mani-
fold (M, g) there exists a unique symmetric connection which is compat-
ible with g (given by the Koszul formula). Therefore there exists just one
symmetric connection preserving the Minkowski metric, which is the trivial
connection (obtained in Cartesian coordinates by taking all Christoel sym-
bols equal to zero). Notice that the geodesics of this connection are straight
lines, corresponding to motions of free particles, which in particular do not
feel any gravitational eld.
To introduce gravity through a symmetric connection we must therefore
consider more general 4-dimensional Lorentzian manifolds, which we will still
call spacetimes. These are no longer required to be dieomorphic to R
4
,
nor to have inertial charts. The study of the geometry of these spacetimes
is usually called general relativity.
Each spacetime comes equipped with its unique Levi-Civita connection,
and hence with its geodesics. If c : I R M is a geodesic, then c, c is
constant, as
d
ds
c(s), c(s) = 2
_
D c
ds
(s), c(s)
_
= 0.
A geodesic is called timelike, null, or spacelike according to whether
c, c < 0, c, c = 0 or c, c > 0 (i.e. according to whether its tangent vector
is timelike, spacelike or null). By analogy with the Cartan connection, we
will take timelike geodesics to represent the free-falling motions of massive
particles. This ensures that the Equivalence Principle holds. Null geodesics
will be taken to represent the motions of light rays.
In general, any curve c : I R M is said to be timelike if c, c < 0.
In this case, c represents the motion of a particle with nonzero mass (which
is accelerating unless c is a geodesic). The proper time measured by the
particle between events c(a) and c(b) is
(c) =
_
b
a
[ c(s)[ds,
where [v[ = [v, v[
1
2
for any v TM.
To select physically relevant spacetimes we must impose some sort of
constraint. By analogy with the formulation of the Laplace equation in
terms of the Cartan connection, we make the following denition.
Definition 4.2. The Lorentzian manifold (M, g) is said to be a vac-
uum solution of the Einstein eld equation if its Levi-Civita connection
satises Ric = 0.
The general Einstein eld equation is
Ric = 8T,
4. GENERAL RELATIVITY 249
where T is the so-called reduced energy-momentum tensor of the mat-
ter content of the spacetime. The simplest model of such a matter content
is that of a pressureless perfect uid, which is described by a rest den-
sity function C

(M) and a unit velocity vector eld U X(M)

(whose integral lines are the motions of the uid particles). The reduced
energy-momentum tensor for this matter model is
T =
_
+
1
2
g
_
,
where
1
(M) is the 1-form associated to U by the metric g. Conse-
quently, the Einstein eld equation for this matter model is
Ric = 4(2 +g)
(compare this with the Poisson equation in terms of the Cartan connection).
It turns out that spacetimes satisfying the Einstein eld equation for appro-
priate choices of T model astronomical phenomena with great accuracy.
Exercises 4.3.
(1) Show that the signature of a pseudo-Riemannian manifold (M, g)
is well dened, i.e., show that the signature of g
p
T
2
(T
p
M) does
not depend on p M.
(2) Let (M, g) be a pseudo-Riemannian manifold and f : N M an
immersion. Show that f

g is not necessarily a pseudo-Riemannian

metric on N.
(3) Let (M, g) be the (n + 1)-dimensional Minkowski spacetime, i.e.,
M = R
n+1
and
g = dx
0
dx
0
+dx
1
dx
1
+ +dx
n
dx
n
.
Let i : N M be the inclusion map, where
N := v M : v, v = 1 and v
0
> 0.
Show that (N, i

g) is the n-dimensional hyperbolic space H

n
.
(4) (Fermi-Walker transport) Let c : I R R
4
be a timelike curve
in Minkowski space parameterized by the proper time, U := c the
tangent unit vector and A := c the proper acceleration. A vector
eld V : I R
4
is said to be Fermi-Walker transported along
c if
DV
d
= V, AU V, UA.
(a) Show that U is Fermi-Walker transported along c.
(b) Show that if V and W are Fermi-Walker transported along c
then V, W is constant.
(c) If V, U = 0 then V is tangent at U to the submanifold
N := v R
4
: v, v = 1 and v
0
> 0,
250 6. RELATIVITY
which is isometric to the hyperbolic 3-space (cf. Exercise 3).
Show that, in this case, V is Fermi-Walker transported if and
only if it is parallel transported along U : I N.
(d) Assume that c describes a circular motion with constant speed
v. Let V be a Fermi-Walker transported vector eld, tangent
to the plane of the motion, such that V, U = 0. Compute
the angle by which V rotates (or precesses) after one revolu-
tion. (Remark: It is possible to prove that the angular momentum vector of
a spinning particle is Fermi-Walker transported along its motion and orthogonal
to it; the above precession, which has been observed for spinning particles such
as electrons, is called the Thomas precession).
(5) (Twin paradox on a cylinder) The quotient of Minkowski space-
time by the discrete isometry group generated by the translation
(t, x, y, z) = (t, x + 8, y, z) is a (at) vacuum solution of the Ein-
stein eld equation. Assume that the Earths motion is represented
by the line x = y = z = 0, and that once again Bob departs at 80%
of the speed of light along the x-axis, leaving his twin sister Alice on
the Earth, on their 20
th
anniversary (cf. Exercise 2.2.5). Because
of the topology of space, the two twins meet again after 10 years
(as measured on the Earth), without Bob ever having accelerated.
(a) Compute the age of each twin in their meeting.
(b) From Bobs viewpoint, it is the Earth which moves away from
him. How do you explain the asymmetry in the twins ages?
(6) (Rotating frame)
(a) Show that the metric of Minkowski spacetime can be written
as
g = dt dt +dr dr +r
2
d d +dz dz
by using cylindrical coordinates (r, , z) in R
3
.
(b) Let > 0 and consider the coordinate change given by =

+ t. Show that in these coordinates the metric is written

as
g =(1
2
r
2
)dt dt +r
2
dt d

+r
2
d

dt
+dr dr +r
2
d

+dz dz.
(c) Show that in the region U = r <
1

the coordinate curves of

constant (r,

, z) are timelike curves corresponding to (accel-

erated) observers rotating rigidly with respect to the inertial
observers of constant (r, , z).
(d) The set of the rotating observers is a 3-dimensional smooth
manifold with local coordinates (r,

, z), and there exists a

natural projection : U . We introduce a Riemannian
metric h on as follows: if v, w T
(p)
then
h(v, w) = g
_
v

, w

_
,
4. GENERAL RELATIVITY 251
where, for each u T
(p)
, the vector u

T
p
U satises
(d)
p
u

= u and g
_
u

,
_

t
_
p
_
= 0.
Show that h is well dened and
h = dr dr +
r
2
1
2
r
2
d

+dz dz.
(Remark: This is the metric resulting from local distance measurements be-
tween the rotating observers; Einstein used the fact that this metric has cur-
vature to argue for the need to use non-Euclidean geometry in the relativistic
description of gravity).
(e) The image of a curve c : I R U consists of simultaneous
events from the point of view of the rotating observers if c is
orthogonal to

t
at each point. Show that this is equivalent to
requiring that ( c) = 0, where
= dt
r
2
1
2
r
2
d

.
In particular, show that, in general, synchronization of the
rotating observers clocks around closed paths leads to incon-
sistencies. (Remark: This is the so-called Sagnac eect; it must be taken
into account when synchronizing the very precise atomic clocks on the GPS
system ground stations, because of the Earths rotation).
(7) (Static spacetime) Let (, h) be a 3-dimensional Riemannian man-
ifold and consider the 4-dimensional Lorentzian manifold (M, g)
determined by M := R and
g := e
2 ( )
dt dt +

h,
where t is the usual coordinate in R, : M is the natural
projection and : R is a smooth function.
(a) Let c : I R M be a timelike geodesic parameterized by
the proper time, and := c. Show that
D
d
= (1 +h( , )) G,
where G = grad() is the vector eld associated to d by
h and can be thought of as the gravitational eld. Show that
this equation implies that the quantity
E
2
:= (1 +h( , ))e
2
is a constant of motion.
(b) Let c : I R M be a null geodesic, c its reparameterization
by the coordinate time t, and := c. Show that is a
geodesic of the Fermat metric
l := e
2 ( )
h.
252 6. RELATIVITY
(Hint: Use Lemma 1.12 in Chapter 5).
(c) Show that the vacuum Einstein eld equation for g is equiva-
lent to
div G = h(G, G);
Ric = d +d d,
where div G is the divergence of G, Ric and are the Ricci
curvature and the Levi-Civita connection of h, and d is the
tensor dened by (d)(X, Y ) := (
X
d) (Y ) for all X, Y
X() (cf. Exercises 2.6.3 and 3.3.9 in Chapter 3).
5. The Schwarzschild Solution
The vacuum Einstein eld equation is nonlinear, and hence much harder
to solve than the Laplace equation. One of the rst solutions to be discovered
was the so-called Schwarzschild solution, which can be obtained from the
simplifying hypotheses of time independence and spherical symmetry, i.e. by
looking for solutions of the form
g = A
2
(r)dt dt +B
2
(r)dr dr +r
2
d d +r
2
sin
2
d d
for unknown positive smooth functions A, B : R R. Notice that this
expression reduces to the Minkowski metric in spherical coordinates for A
B 1).
It is easily seen that the Cartan structure equations still hold for pseudo-
Riemannian manifolds. We have
g =
0

0
+
r

r
+

with

0
= A(r)dt;

r
= B(r)dr;

= rd;

= r sin d,
and hence
0
,
r
,

is an orthonormal coframe. The rst structure

equations,
d

=
3

=0

;
dg

=
3

=0
g

+g

,
5. THE SCHWARZSCHILD SOLUTION 253
which on an orthonormal frame are written as
d

=
3

=0

0
0
=
i
i
= 0;

0
i
=
i
0
;

i
j
=
j
i
(i, j = 1, 2, 3), together with
d
0
=
A

B

r
dt;
d
r
= 0;
d

=
1
B

r
d;
d

=
sin
B

r
d + cos

d,
yield the nonvanishing connection forms

0
r
=
r
0
=
A

B
dt;

r
=
r

=
1
B
d;

r
=
r

=
sin
B
d;

= cos d.
The curvature forms can be computed from the second structure equations

= d

+
3

=0

,
and are found to be

0
r
=
r
0
=
A

B A

AB
3

r

0
;

0
=
A

rAB
2

0
;

0
=
A

rAB
2

0
;

r
=
r

=
B

rB
3

r
;

r
=
r

=
B

rB
3

r
;

=
B
2
1
r
2
B
2

.
254 6. RELATIVITY
Thus the components of the curvature tensor on the orthonormal frame
can be read o from the curvature forms using

<
R

.
and can in turn be used to compute the components of the Ricci curvature
tensor Ric on the same frame. The nonvanishing components of Ric on this
frame turn out to be
R
00
=
A

B A

AB
3
+
2A

rAB
2
;
R
rr
=
A

B A

AB
3
+
2B

rB
3
;
R

= R

=
A

rAB
2
+
B

rB
3
+
B
2
1
r
2
B
2
.
Thus the vacuum Einstein eld equation Ric = 0 is equivalent to the
ODE system
_

_
A

A

A

AB
+
2A

rA
= 0
A

A

A

AB

2B

rB
= 0
A

A

B

B

B
2
1
r
= 0

_
A

A
+
B

B
= 0
_
A

A
_

+ 2
_
A

A
_
2
+
2A

rA
= 0
2B

B
+
B
2
1
r
= 0
.
The last equation can be immediately solved to yield
B =
_
1
2m
r
_

1
2
,
where m R is an integration constant. The rst equation implies that
A =

B
for some constant > 0. By rescaling the time coordinate t we
can assume that = 1. Finally, it is easily checked that the second ODE
is identically satised. Therefore there exists a one-parameter family of
solutions of the vacuum Einstein eld equation of the form we sought, given
by
g =
_
1
2m
r
_
dt dt +
_
1
2m
r
_
1
dr dr (28)
+r
2
d d +r
2
sin
2
d d.
To interpret this family of solutions, we compute the proper acceleration
(cf. Exercise 2.2.12) of the stationary observers, whose motions are the
integral curves of

t
. If E
0
, E
r
, E

, E

is the orthonormal frame obtained

5. THE SCHWARZSCHILD SOLUTION 255
by normalizing
_

t
,

r
,

_
(hence dual to
0
,
r
,

), we have

E
0
E
0
=
3

=0

0
(E
0
)E

=
r
0
(E
0
)E
r
=
A

AB

0
(E
0
)E
r
=
m
r
2
_
1
2m
r
_

1
2
E
r
.
Therefore, each stationary observer is accelerating with a proper acceleration
G(r) =
m
r
2
_
1
2m
r
_

1
2
away from the origin, to prevent falling towards it. In other words, they
are experiencing a gravitational eld of intensity G(r), directed towards the
origin. Since G(r) approaches, for large values of r, the familiar acceleration
m/r
2
of the Newtonian gravitational eld generated by a point particle of
mass m, we interpret the Schwarzschild solution as the general relativistic
eld of a point particle of mass m. Accordingly, we will assume that m > 0
(notice that m = 0 corresponds to Minkowski spacetime).
When obtaining the Schwarzschild solution we assumed A(r) > 0, and
hence r > 2m. However, it is easy to check that (28) is also a solution
of the Einstein vacuum eld equation for r < 2m. Notice that the coordi-
nate system (t, r, , ) is singular at r = 2m, and hence covers only the two
disconnected open sets r > 2m and r < 2m. Both these sets are geodesi-
cally incomplete, as for instance radial timelike or null geodesics cannot be
extended as they approach r = 0 or r = 2m (cf. Exercise 5.1.7). While
this is to be expected for r = 0, as the curvature blows up along geodesics
approaching this limit, this is not the case for r = 2m. It turns out that
it is possible to t these two open sets together to obtain a solution of the
Einstein vacuum eld equation regular at r = 2m. To do so, we introduce
the so-called Painleve time coordinate
t

= t +
_
_
2m
r
_
1
2m
r
_
1
dr.
In the coordinate system (t

, r, , ), the Schwarzschild metric is written

g = dt

dt

+
_
dr +
_
2m
r
dt

_
dr +
_
2m
r
dt

_
+r
2
dd+r
2
sin
2
dd.
This expression is nonsingular at r = 2m, and is a solution of the Einstein
vacuum eld equation for r > 2m and r < 2m. By continuity, it must
be a solution also at r = 2m.
The submanifold r = 2m is called the event horizon, and is ruled by
null geodesics. This is easily seen from the fact that

t

=

t
becomes null
at r = 2m, and hence its integral curves are (reparameterizations of) null
geodesics.
The causal properties of the Schwarzschild spacetime are best under-
stood by studying the light cones, i.e. the set of tangent null vectors at
256 6. RELATIVITY
each point. For instance, radial null vectors v = v
0
t

+v
r
r
satisfy

_
v
0
_
2
+
_
v
r
+
_
2m
r
v
0
_
2
= 0 v
r
=
_
1
_
2m
r
_
v
0
.
For r 2m we obtain approximately the usual light cones of Minkowski
spacetime. As r approaches 2m, however, the light cones tip over towards
the origin, becoming tangent to the event horizon at r = 2m (cf. Figure 4).
Since the tangent vector to a timelike curve must be inside the light cone,
we see that no particle which crosses the event horizon can ever leave the
region r = 2m (which for this reason is called a black hole). Once inside
the black hole, the light cones tip over even more, forcing the particle into
the singularity r = 0.
t

r
r = 2m
Figure 4. Light cones in Painleve coordinates.
Notice that the Schwarzschild solution in Painleve coordinates is still
not geodesically complete at the event horizon, as radial timelike and null
geodesics are incomplete to the past as they approach r = 2m (cf. Exer-
cise 5.1.7). Physically, this is not important: black holes are thought to
form through the collapse of (approximately) spherical stars, whose surface
follows a radial timelike curve in the spacetime diagram of Figure 4. Since
only outside the star is there vacuum, the Schwarzschild solution is expected
to hold only above this curve, thereby removing the region of r = 2m lead-
ing to incompleteness. Nevertheless, it is possible to glue two copies of the
Schwarzschild spacetime in Painleve coordinates to obtain a solution of the
vacuum Einstein eld equation which is geodesically incomplete only at the
two copies of r = 0. This solution, known as the Kruskal extension,
contains a black hole and its time-reversed version, known as a white hole.
5. THE SCHWARZSCHILD SOLUTION 257
For some time it was thought that the curvature singularity at r = 0
was an artifact of the high symmetry of Schwarzschild spacetime, and that
more realistic models of collapsing stars would be singularity-free. Hawking
and Penrose proved that this is not the case: once the collapse has begun,
no matter how asymmetric, nothing can prevent a singularity from forming
(cf. Sections 8 and 9).
Exercises 5.1.
(1) Let (M, g) be a 2-dimensional Lorentzian manifold.
(a) Consider an orthonormal frame E
0
, E
1
on an open set U
M, with associated coframe
0
,
1
. Check that the Cartan
structure equations are

0
1
=
1
0
;
d
0
=
1

0
1
;
d
1
=
0

0
1
;

0
1
= d
0
1
.
(b) Let F
0
, F
1
be another orthonormal frame such that F
0

C(E
0
), with associated coframe
0
,
1
and connection form

0
1
. Show that =
0
1

0
1
is given locally by = du, where u
is the hyperbolic angle between F
0
and E
0
(cf. Exercise 2.2.7).
(c) Consider a triangle U whose sides are timelike geodesics,
and let , and be the hyperbolic angles between them
(cf. Figure 5). Show that
= +
_

0
1
,
where, following the usual convention for spacetime diagrams,
we orient U so that E
0
, E
1
is negative.
(d) Provide a physical interpretation for the formula above in the
case in which (M, g) is a totally geodesic submanifold of the
Schwarzschild spacetime obtained by xing (, ) (cf. Exer-
cise 5.7.3 in Chapter 4).
(2) Consider the Schwarzschild spacetime with local coordinates (t, r, , ).
An equatorial circular curve is a curve given in these coordi-
nates by (t(), r(), (), ()) with r() 0 and ()

2
.
(a) Show that the conditions for such a curve to be a timelike
geodesic parameterized by its proper time are
_

t = 0
= 0
r
2
=
m
r
2

t
2
_
1
3m
r
_

t
2
= 1
.
258 6. RELATIVITY

Figure 5. Timelike geodesic triangle.

Conclude that massive particles can orbit the central mass in
circular orbits for all r > 3m.
(b) Show that there exists an equatorial circular null geodesic for
r = 3m. What does a stationary observer placed at r = 3m,
=

2
see as he looks along the direction of this null geodesic?
(c) The angular momentum vector of a free-falling spinning par-
ticle is parallel-transported along its motion, and orthogonal
to it (cf. Exercise 4.3.4). Consider a spinning particle on a cir-
cular orbit around a pointlike mass m. Show that the angular
momentum vector precesses by an angle
= 2
_
1
_
1
3m
r
_1
2
_
,
after one revolution, if initially aligned with the radial direc-
tion. (Remark: The above precession, which has been observed for spinning
quartz spheres in orbit around the Earth during the Gravity Probe B experiment,
is called the geodesic precession).
(3) (Gravitational redshift) We consider again the Schwarzschild space-
time with local coordinates (t, r, , ).
(a) Show that the proper time interval measured by a station-
ary observer between two events on his history is
=
_
1
2m
r
_1
2
t,
5. THE SCHWARZSCHILD SOLUTION 259
t
r
r
0
r
1
T
T

Figure 6. Gravitational redshift.

where t is the dierence between the time coordinates of
the two events. (Remark: This eect has been measured experimentally;
loosely speaking, gravity delays time).
(b) Show that if (t(s), r(s), (s), (s)) is a geodesic then so is
(t(s) + t, r(s), (s), (s)) for any t R.
(c) Use the spacetime diagram in Figure 6 to show that if a sta-
tionary observer at r = r
0
measures a light signal to have
period T, a stationary observer at r = r
1
measures a period
T

= T

_
1
2m
r
1
1
2m
r
0
for the same signal. (Remark: This gravitational redshift has been mea-
sured experimentally, conrming that spacetime must be curved in Minkowski
spacetime one would necessarily have T = T

).
(4) Let (M, g) be the region r > 2m of the Schwarzschild solution with
the Schwarzschild metric. The set of all stationary observers in
M is a 3-dimensional smooth manifold with local coordinates
(r, , ), and there exists a natural projection : M . We
introduce a Riemannian metric h on as follows: if v, w T
(p)

then
h(v, w) = g
_
v

, w

_
,
260 6. RELATIVITY
where, for each u T
(p)
, the vector u

T
p
U satises
(d)
p
u

= u and g
_
u

,
_

t
_
p
_
= 0
(cf. Exercise 4.3.6).
(a) Show that h is well dened and
h =
_
1
2m
r
_
1
dr dr +r
2
d d +r
2
sin
2
d d.
(b) Show that h is not at, but has zero scalar curvature.
(c) Show that the equatorial plane =

2
is isometric to the rev-
olution surface generated by the curve z(r) =
_
8m(r 2m)
when rotated around the z-axis (cf. Figure 7).
(Remark: This is the metric resulting from local distance measurements between the
stationary observers; loosely speaking, gravity deforms space).
Figure 7. Surface of revolution isometric to the equatorial plane.
(5) In this exercise we study in detail the timelike and null geodesics of
the Schwarzschild spacetime. We start by observing that the sub-
manifold =

2
is totally geodesic (cf. Exercise 5.7.3 in Chapter 4).
By adequately choosing the angular coordinates (, ), one can al-
ways assume that the initial condition of the geodesic is tangent to
this submanifold; hence it suces to study the timelike and null
geodesics of the 3-dimensional Lorentzian manifold (M, g), where
g =
_
1
2m
r
_
dt dt +
_
1
2m
r
_
1
dr dr +r
2
d d.
(a) Show that

t
and

are Killing elds (cf. Exercise 3.3.8 in

Chapter 3).
(b) Conclude that the equations for a curve c : R M to be
a future-directed geodesic (parameterized by proper time if
5. THE SCHWARZSCHILD SOLUTION 261
timelike) can be written as
_

_
g( c, c) =
g
_

t
, c
_
= E
g
_

, c
_
= L

_
r
2
= E
2

_
+
L
2
r
2
_
_
1
2m
r
_
_
1
2m
r
_

t = E
r
2
= L
where E > 0 and L are integration constants, = 1 for time-
like geodesics and = 0 for null geodesics.
(c) Show that if L ,= 0 then u =
1
r
satises
d
2
u
d
2
+u =
m
L
2
+ 3mu
2
.
(d) For situations where relativistic corrections are small one has
mu 1, and hence the approximate equation
d
2
u
d
2
+u =
m
L
2
holds for timelike geodesics. Show that the solution to this
equation is the conic section given in polar coordinates by
u =
m
L
2
(1 + cos(
0
)),
where the integration constants 0 and
0
are the eccen-
tricity and the argument of the pericenter.
(e) Show that for 1 this approximate solution satises
u
2

2m
L
2
u
m
2
L
4
.
Argue that timelike geodesics close to circular orbits where
relativistic corrections are small yield approximate solutions
of the equation
d
2
u
d
2
+
_
1
6m
2
L
2
_
u =
m
L
2
_
1
3m
2
L
2
_
,
and hence the pericenter advances by approximately
6m
r
radians per revolution. (Remark: The rst success of general relativity
was due to this eect, which explained the anomalous precession of Mercurys
perihelion 43 arcseconds per century).
(f) Show that if one neglects relativistic corrections then null
geodesics satisfy
d
2
u
d
2
+u = 0.
262 6. RELATIVITY
Show that the solution to this equation is the equation for a
straight line in polar coordinates,
u =
1
b
sin(
0
),
where the integration constants b > 0 and
0
are the impact
parameter (distance of closest approach to the center) and
the angle between the line and the x-axis.
(g) Assume that mu 1. Let us include relativistic corrections
by looking for approximate solutions of the form
u =
1
b
_
sin +
m
b
v
_
(where we take
0
= 0 for simplicity). Show that v is an
approximate solution of the equation
d
2
v
d
2
+v = 3 sin
2
,
and hence u is approximately given by
u =
1
b
_
sin +
m
b
_
3
2
+
1
2
cos(2) +cos + sin
__
,
where and are integration constants.
(h) Show that for the incoming part of the null geodesic ( 0)
one approximately has
u = 0 =
m
b
(2 +) .
Similarly, show that for the outgoing part of the null geodesic
( ) one approximately has
u = 0 = +
m
b
(2 ) .
Conclude that varies by approximately
= +
4m
b
radians along its path, and hence the null geodesic is deected
towards the center by approximately
4m
b
radians. (Remark: The measurement of this deection of light by the Sun
1.75 arcseconds was the rst experimental conrmation of general relativity,
and made Einstein a world celebrity overnight).
(6) (Birkho Theorem) Prove that the only Ricci-at Lorentzian metric
given in local coordinates (t, r, , ) by
g = A
2
(t, r)dtdt+B
2
(t, r)drdr+r
2
dd+r
2
sin
2
dd
6. COSMOLOGY 263
is the Schwarzschild metric. Loosely speaking, spherically symmet-
ric mass congurations do not radiate.
(7) (a) Show that the radial timelike or null geodesics in the regions
r > 2m and r < 2m of the Schwarzschild spacetime can-
not be extended as they approach r = 0 or r = 2m.
(b) Show that the radial timelike or null geodesics in the Painleve
extension of the Schwarzschild spacetime can be extended to
the future, but not to the past, as they approach r = 2m.
(c) Show that radial observers satisfying
dr
dt

=
_
2m
r
in the Painleve coordinates are free-falling, and that t

is their
proper time.
(d) What does a stationary observer see as a particle falls into a
black hole?
(8) Show that an observer who crosses the horizon will hit the singu-
larity in a proper time interval m.
6. Cosmology
Cosmology studies the behavior of the Universe as a whole. Experimen-
tal observations (chiey that of the cosmic background radiation) suggest
that space is isotropic at the Earths location. Assuming the Copernican
Principle that the Earths location in the Universe is not in any way special,
we take an isotropic (hence constant curvature) 3-dimensional Riemannian
manifold (, h) as our model of space. We can always nd local coordinates
(r, , ) on such that
h = a
2
_
1
1 kr
2
dr dr +r
2
d d +r
2
sin
2
d d
_
,
where a > 0 is the radius of space and k = 1, 0, 1 according to whether
the curvature is negative, zero or positive (cf. Exercise 6.1.1). Allowing for
the possibility that the radius of space may be varying in time, we take
our model of the Universe to be (M, g), where M = R and
g = dt dt +a
2
(t)
_
1
1 kr
2
dr dr +r
2
d d +r
2
sin
2
d d
_
.
These are the so-called Friedmann-Lematre-Robertson-Walker (FLRW)
models of cosmology.
One can easily compute the Ricci curvature for the metric g. We have
g =
0

0
+
r

r
+

264 6. RELATIVITY
with

0
= dt;

r
= a(t)
_
1 kr
2
_

1
2
dr;

= a(t)rd;

= a(t)r sin d,
and hence
0
,
r
,

is an orthonormal coframe. The rst structure

equations yield

0
r
=
r
0
= a
_
1 kr
2
_

1
2
dr;

0
= ard;

0
= ar sind;

r
=
r

=
_
1 kr
2
_1
2
d;

r
=
r

=
_
1 kr
2
_1
2
sin d;

= cos d.
The curvature forms can be computed from the second structure equa-
tions, and are found to be

0
r
=
r
0
=
a
a

0

r
;

0
=
a
a

0
=
a
a

r
=
r

=
_
k
a
2
+
a
2
a
2
_

r
;

r
=
r

=
_
k
a
2
+
a
2
a
2
_

r
;

=
_
k
a
2
+
a
2
a
2
_

.
The components of the curvature tensor on the orthonormal frame can
be read o from the curvature forms, and can in turn be used to compute
the components of the Ricci curvature tensor Ric on the same frame. The
nonvanishing components of Ric on this frame turn out to be
R
00
=
3 a
a
;
R
rr
= R

= R

=
a
a
+
2 a
2
a
2
+
2k
a
2
.
At very large scales, galaxies and clusters of galaxies are expected to
behave as particles of a pressureless uid, which we take to be our matter
6. COSMOLOGY 265
model. By isotropy, the average spatial motion of the galaxies must vanish,
and hence their unit velocity vector eld must be

t
(corresponding to the
1-form dt). Therefore the Einstein eld equation is
Ric = 4(2dt dt +g),
which is equivalent to the ODE system
_

3 a
a
= 4
a
a
+
2 a
2
a
2
+
2k
a
2
= 4

_
a +
a
2
2a
+
k
2a
= 0
=
3 a
4a
.
The rst equation allows us to determine the function a(t), and the
second yields (which in particular must be a function of the t coordinate
only; this is to be taken to mean that the average density of matter at
cosmological scales is spatially constant). On the other hand, the quantity
4a
3
3
= aa
2
is constant, since
d
dt
_
aa
2
_
=
d
dt
_
a a
2
2
+
ka
2
_
= a a a +
a
3
2
+
k a
2
= 0.
Hence we have
a =

a
2
for some integration constant (we take > 0 so that > 0). Substituting
in the equation for a(t) we get the rst order ODE
a
2
2


a
=
k
2
.
This can be used to show that a(t) is bounded if and only if k = 1 (cf. Exer-
cise 6.1.3). Moreover, in all cases a(t) vanishes (and hence a(t), a(t) and (t)
blow up) for some value of t, usually taken to be t = 0. This singularity is
called the Big Bang of the solution dened for t > 0. It was once thought
to be a consequence of the high degree of symmetry of the FLRW mod-
els. Hawking and Penrose, however, showed that the Big Bang is actually a
generic feature of cosmological models (cf. Sections 8 and 9).
The function
H(t) =
a
a
is (somewhat confusingly) called the Hubble constant. It is easy to see
from the above equations that
H
2
+
k
a
2
=
8
3
.
266 6. RELATIVITY
Therefore, in these models one has k = 1, k = 0 or k = 1 according to
whether the average density of the Universe is smaller than, equal to or
bigger than the so-called critical density

c
=
3H
2
8
.
These models were the standard models for cosmology for a long time.
Currently, however, things are thought to be slightly more complicated
(cf. Exercise 6.1.6).
Exercises 6.1.
(1) Show that the Riemannian metric h given in local coordinates
(r, , ) by
h = a
2
_
1
1 kr
2
dr dr +r
2
d d +r
2
sin
2
d d
_
has constant curvature K =
k
a
2
.
(2) The motions of galaxies and groups of galaxies in the FLRW models
are the integral curves of

t
. Show that these are timelike geodesics,
and that the time coordinate t is the proper time of such observers.
(3) Recall that in a FLRW model the radius of space, a(t), evolves
according to the ODE
a
2
2


a
=
k
2
a =

a
2
.
Show that:
(a) a(t) vanishes in nite time (assume that this happens at t = 0);
(b) if k = 1 or k = 0 then the solution can be extended to all
values of t > 0;
(c) if k = 1 then the solution cannot be extended past t = 2
(Big Crunch);
(d) if k = 1 then no observer can circumnavigate the Universe, no
matter how fast he moves;
(e) the solution can be given parametrically by:
(i) k = 1:
_
a = (1 cos u)
t = (u sin u)
;
(ii) k = 0:
_
a =

2
u
2
t =

6
u
3
;
(iii) k = 1:
_
a = (cosh u 1)
t = (sinh u u)
.
6. COSMOLOGY 267
(4) Show that the FLRW model with k = 1 is isometric to the hyper-
surface with equation
_
x
2
+y
2
+z
2
+w
2
= 2
t
2
8
in the 5-dimensional Minkowski spacetime (R
5
, g) with metric
g = dt dt +dx dx +dy dy +dz dz +dw dw.
(5) (A model of collapse) Show that the radius of a free-falling spherical
shell r = r
0
in a FLRW model changes with proper time in exactly
the same fashion as the radius of a free-falling spherical shell in a
Schwarzschild spacetime of mass parameter m moving with energy
parameter E (cf. Exercise 5.1.5), provided that
_
m = r
0
3
E
2
1 = kr
0
2
.
Therefore these two spacetimes can be matched along the 3-dimen-
sional hypersurface determined by the spherical shells motion to
yield a model of collapsing matter. Can you physically interpret
this model?
(6) Show that if we allow for a cosmological constant R, i.e. for
an Einstein equation of the form
Ric = 4(2 +g) + g
then the equations for the FLRW models become
_

_
a
2
2


a


6
a
2
=
k
2
4
3
a
3
=
Analyze the possible behaviors of the function a(t). (Remark: It is
currently thought that there exists indeed a positive cosmological constant, also known
as dark energy. The model favored by experimental observations seems to be k = 0,
> 0).
(7) Consider the 5-dimensional Minkowski spacetime (R
5
, g) with met-
ric
g = dt dt +dx dx +dy dy +dz dz +dw dw.
Show that the induced metric on each of the following hypersurfaces
determines FLRW models for which the matter is a pressureless
uid with cosmological constant with the indicated parameters.
(a) (Einstein universe) The cylinder of equation
x
2
+y
2
+z
2
+w
2
=
1

,
satises k = 1, > 0 and =

4
.
268 6. RELATIVITY
(b) (de Sitter universe) The sphere of equation
t
2
+x
2
+y
2
+z
2
+w
2
=
3

satises k = 1, > 0 and = 0.

7. Causality
In this section we will study the causal features of spacetimes. This is
a subject which has no parallel in Riemannian geometry, where the metric
is positive denite. Although we will focus on 4-dimensional Lorentzian
manifolds, the discussion can be easily generalized to any dimension n 2.
A spacetime (M, g) is said to be time-orientable if there exists a vector
eld X X(M) such that X, X < 0. In this case, we can dene a time
orientation on each tangent space T
p
M (which is, of course, isometric to
Minkowski spacetime) by choosing the timelike vectors in the connected
component C(X
p
) to be future-pointing.
Assume that (M, g) is time-oriented (i.e. time-orientable with a def-
inite choice of time orientation). A timelike curve c : I R M is said
to be future-directed if c is future-pointing. The chronological future
of p M is the set I
+
(p) of all points to which p can be connected by
a future-directed timelike curve. A future-directed causal curve is a
curve c : I R M such that c is timelike or null and future-pointing
(if nonzero). The causal future of p M is the set J
+
(p) of all points
to which p can be connected by a future-directed causal curve. Notice that
I
+
(p) is simply the set of all events which are accessible to a particle with
nonzero mass at p, whereas J
+
(p) is the set of events which can be causally
inuenced by p (as this causal inuence cannot propagate faster than the
speed of light). Analogously, the chronological past of p M is the set
I

(p) of all points which can be connected to p by a future-directed timelike

curve, and the causal past of p M is the set J

(p) of all points which

can be connected to p by a future-directed causal curve.
In general, the chronological and causal pasts and futures can be quite
complicated sets, because of global features of the spacetime. Locally, how-
ever, causal properties are similar to those of Minkowski spacetime. More
precisely, we have the following statement:
Proposition 7.1. Let (M, g) be a time-oriented spacetime. Then each
point p
0
M has an open neighborhood V M such that the spacetime
(V, g) obtained by restricting g to V satises:
(1) V is a normal neighborhood of each of its points, and given p, q V
there exists a unique geodesic (up to reparameterization) joining p
to q (i.e. V is geodesically convex);
(2) q I
+
(p) if and only if there exists a future-directed timelike geo-
desic connecting p to q;
(3) J
+
(p) = I
+
(p);
7. CAUSALITY 269
(4) q J
+
(p) I
+
(p) if and only if there exists a future-directed null
geodesic connecting p to q.
Proof. Let U be a normal neighborhood of p
0
and choose normal
coordinates (x
0
, x
1
, x
2
, x
3
) on U, given by the parameterization
(x
0
, x
1
, x
2
, x
3
) = exp
p
0
(x
0
v
0
+x
1
v
1
+x
2
v
2
+x
3
v
3
),
where v
0
, v
1
, v
2
, v
3
is a basis of T
p
0
(M) (cf. Exercise 4.8.2 in Chapter 3).
Let D : U R be the dierentiable function
D(p) :=
3

=0
(x

(p))
2
,
and let us dene for each > 0 the set
B

:= p U [ D(p) < ,
which for suciently small is dieomorphic to an open ball in T
p
0
M.
Assume, for simplicity, that U is of this form.
Let us show that there exists > 0 such that if c : I R B

is a
geodesic then all critical points of D(t) := D(c(t)) are strict local minima.
In fact, setting x

(t) := x

(c(t)), we have

D(t) = 2
3

=0
x

(t) x

(t);

D(t) = 2
3

=0
( x

(t))
2
+ 2
3

=0
x

(t) x

(t)
= 2
3

,=0
_

=0

(c(t))x

(t)
_
x

(t) x

(t),
and for suciently small the matrix

=0

is positive denite on B

.
Consider the map F : W TM M M, dened on some open
neighborhood W of 0 T
p
0
M by
F(v) = ((v), exp(v)).
As we saw in the Riemannian case (cf. Chapter 3, Section 4), this map is
a local dieomorphism at 0 T
p
0
M. Choosing > 0 suciently small and
reducing W, we can assume that F maps W dieomorphically to B

,
and that exp(tv) B

for all t [0, 1] and v W (as otherwise it would be

possible to construct a sequence v
n
0 T
p
0
M such that exp(v
n
) ,p
0
).
Finally, set V = B

. If p, q V and v = F
1
(p, q), then c(t) = exp
p
(tv)
is a geodesic connecting p to q whose image is contained in B

. If its im-
age were not contained in V , there would necessarily exist a point of local
270 6. RELATIVITY
maximum of D(t), which cannot occur. Therefore, there is a geodesic in V
connecting p to q. Since exp
p
is a dieomorphism onto V , this geodesic is
unique (up to reparameterization). This proves (1).
To prove assertion (2), we start by noticing that if there exists a future-
directed timelike geodesic connecting p to q then it is obvious that q I
+
(p).
Suppose now that q I
+
(p); then there exists a future-directed timelike
curve c : [0, 1] V such that c(0) = p and c(1) = q. Choose normal
coordinates (x
0
, x
1
, x
2
, x
3
) given by the parameterization
(x
0
, x
1
, x
2
, x
3
) = exp
p
(x
0
E
0
+x
1
E
1
+x
2
E
2
+x
3
E
3
),
where E
0
, E
1
, E
2
, E
3
is an orthonormal basis of T
p
M with E
0
timelike and
future-pointing. These are global coordinates in V , since F : W V V
is a dieomorphism. Dening
W
p
(q) :=
_
x
0
(q)
_
2
+
_
x
1
(q)
_
2
+
_
x
2
(q)
_
2
+
_
x
3
(q)
_
2
=
3

,=0

(q)x

(q),
with (

) = diag(1, 1, 1, 1), we have to show that W

p
(q) < 0. Let W
p
(t) :=
W
p
(c(t)). Since x

(p) = 0 ( = 0, 1, 2, 3), we have W

p
(0) = 0. Setting
x

(t) := x

(c(t)), we obtain

W
p
(t) = 2
3

,=0

(t) x

(t);

W
p
(t) = 2
3

,=0

(t) x

(t) + 2
3

,=0

(t) x

(t),
and consequently (recalling that
_
d exp
p
_
0
= id)

W
p
(0) = 0;

W
p
(0) = 2 c(0), c(0) < 0.
Therefore there exists > 0 such that W
p
(t) < 0 for t (0, ).
Using the same ideas as in the Riemannian case (cf. Chapter 3, Sec-
tion 4), it is easy to prove that the level surfaces of W
p
are orthogonal to
the geodesics through p. Therefore, if c
v
(t) = exp
p
(tv) is the geodesic with
initial condition v T
p
M, we have
(grad W
p
)
cv(1)
= a(v) c
v
(1),
where the gradient of a function is dened as in the Riemannian case (notice
however that in the Lorentzian case a smooth function f decreases along
7. CAUSALITY 271
the direction of grad f if grad f is timelike). Now

(grad W
p
)
cv(t)
, c
v
(t)
_
=
d
dt
W
p
(c
v
(t)) =
d
dt
W
p
(c
tv
(1))
=
d
dt
_
t
2
W
p
(c
v
(1))
_
= 2tW
p
(c
v
(1)),
and hence

(grad W
p
)
cv(1)
, c
v
(1)
_
= 2W
p
(c
v
(1)).
On the other hand,

(grad W
p
)
cv(1)
, c
v
(1)
_
= a(v) c
v
(1), c
v
(1)
= a(v)v, v = a(v)W
p
(c
v
(1)).
We conclude that a(v) = 2, and therefore
(grad W
p
)
cv(1)
= 2 c
v
(1).
Consequently, grad W
p
is tangent to geodesics through p, being future-
pointing on future-directed geodesics.
Suppose that W
p
(t) < 0. Then (grad W
p
)
c(t)
is timelike future-pointing,
and so

W(t) =
_
(grad W
p
)
c(t)
, c(t)
_
< 0,
as c(t) is also timelike future-pointing (cf. Exercise 2.2.2). We conclude that
we must have W
p
(t) < 0 for all t [0, 1]. In particular, W
p
(q) = W
p
(1) < 0,
and hence there exists a future-directed timelike geodesic connecting p to q.
To prove assertion (3), let us see rst that I
+
(p) J
+
(p). If q I
+
(p),
then q is the limit of a sequence of points q
n
I
+
(p). By (2), q
n
= exp
p
(v
n
)
with v
n
T
p
M timelike future-pointing. Since exp
p
is a dieomorphism, v
n
converges to a causal future-pointing vector v T
p
M, and so q = exp
p
(v)
can be reached from p by a future-directed causal geodesic. The converse
inclusion J
+
(p) I
+
(p) holds in general (cf. Proposition 7.2).
Finally, (4) is obvious from (3) and the fact that exp
p
is a dieomorphism
onto V .
This local behavior can be used to prove the following global result,
whose proof is left as an exercise (cf. Exercise 7.10.2).
Proposition 7.2. Let (M, g) be a time oriented spacetime and p M.
Then:
(1) I
+
(p) is open;
(2) J
+
(p) I
+
(p);
(3) I
+
(p) = int J
+
(p)
(4) if r J
+
(p) and q I
+
(r) then q I
+
(p);
(5) if r I
+
(p) and q J
+
(r) then q I
+
(p).
The generalized twin paradox (cf. Exercise 2.2.8) also holds locally for
general spacetimes. More precisely, we have the following statement:
272 6. RELATIVITY
Proposition 7.3. Let (M, g) be a time-oriented spacetime, p
0
M and
V M a geodesically convex open neighborhood of p
0
. The spacetime (V, g)
obtained by restricting g to V satises the following property: if p, q V
with q I
+
(p), c is the timelike geodesic connecting p to q and is any
timelike curve connecting p to q, then () (c), with equality if and only
if is a reparameterization of c.
Proof. Any timelike curve : [0, 1] V satisfying (0) = p, (1) = q
can be written as
(t) = exp
p
(r(t)n(t)),
for t [0, 1], where r(t) 0 and n(t), n(t) = 1. We have
(t) = (exp
p
)

( r(t)n(t) +r(t) n(t)) .

Since n(t), n(t) = 1, we have n(t), n(t) = 0, and consequently n(t) is
tangent to the level surfaces of the function v v, v. We conclude that
(t) = r(t)X
(t)
+Y (t),
where X is the unit tangent vector eld to timelike geodesics through p
and Y (t) = r(t)(exp
p
)

n(t) is tangent to the level surfaces of W

p
(hence
orthogonal to X
(t)
). Consequently,
() =
_
1
0

r(t)X
(t)
+Y (t), r(t)X
(t)
+Y (t)
_

1
2
dt
=
_
1
0
_
r(t)
2
[Y (t)[
2
_1
2
dt

_
1
0
r(t)dt = r(1) = (c),
where we have used the facts that is timelike, r(t) > 0 for all t [0, 1]
(as is future-pointing) and (c) = r(1) (as q = exp
p
(r(1)n(1))). It should
be clear that () = (c) if and only if [Y (t)[ 0 Y (t) 0 (Y (t) is
spacelike or zero) for all t [0, 1], implying that n is constant. In this case,
(t) = exp
p
(r(t)n) is, up to reparameterization, the geodesic through p with
initial condition n T
p
M.
There is also a local property characterizing null geodesics.
Proposition 7.4. Let (M, g) be a time-oriented spacetime, p
0
M and
V M a geodesically convex open neighborhood of p
0
. The spacetime (V, g)
obtained by restricting g to V satises the following property: if for p, q V
there exists a future-directed null geodesic c connecting p to q and is a
causal curve connecting p to q then is a reparameterization of c.
Proof. Since p and q are connected by a null geodesic, we conclude
from Proposition 7.1 that q J
+
(p) I
+
(p). Let : [0, 1] V be a
causal curve connecting p to q. Then we must have (t) J
+
(p) I
+
(p)
7. CAUSALITY 273
for all t [0, 1], since (t
0
) I
+
(p) implies (t) I
+
(p) for all t > t
0
(see
Proposition 7.2). Consequently, we have
_
(grad W
p
)
(t)
, (t)
_
= 0,
where W
p
was dened in the proof of Proposition 7.1. The formula
(grad W
p
)
cv(1)
= 2 c
v
(1),
which was proved for timelike geodesics c
v
with initial condition v T
p
M,
must also hold for null geodesics (by continuity). Hence grad W
p
is tangent
to the null geodesics ruling J
+
(p) I
+
(p) and future-pointing. Since (t) is
also future-pointing, we conclude that is proportional to grad W
p
(cf. Ex-
ercise 2.2.8), and therefore must be a reparameterization of a null geodesic
(which must be c).
Corollary 7.5. Let (M, g) be a time-oriented spacetime and p M.
If q J
+
(p) I
+
(p) then any future-directed causal curve connecting p to q
must be a reparameterized null geodesic.
For physical applications, it is important to require that the spacetime
satises reasonable causality conditions. The simplest of these conditions
excludes time travel, i.e. the possibility of a particle returning to an event
in its past history.
Definition 7.6. A spacetime (M, g) is said to satisfy the chronology
condition if it does not contain closed timelike curves.
This condition is violated by compact spacetimes:
Proposition 7.7. Any compact spacetime (M, g) contains closed time-
like curves.
Proof. Taking if necessary the time-orientable double covering (cf. Ex-
ercise 7.10.1), we can assume that (M, g) is time-oriented. Since I
+
(p) is an
open set for any p M, it is clear that I
+
(p)
pM
is an open cover of M.
If M is compact, we can obtain a nite subcover I
+
(p
1
), . . . , I
+
(p
N
). Now
if p
1
I
+
(p
i
) for i ,= 1 then I
+
(p
1
) I
+
(p
i
), and we can exclude I
+
(p
1
)
from the subcover. Therefore, we can assume without loss of generality that
p
1
I
+
(p
1
), and hence there exists a closed timelike curve starting and
ending at p
1
.
A stronger restriction on the causal behavior of the spacetime is the
following:
Definition 7.8. A spacetime (M, g) is said to be stably causal if there
exists a global time function, i.e. a smooth function t : M R such that
grad(t) is timelike.
In particular, a stably causal spacetime is time-orientable. We choose
the time orientation dened by grad(t), so that t increases along future-
directed timelike curves. Notice that this implies that no closed timelike
274 6. RELATIVITY
curves can exist, i.e. any stably causal spacetime satises the chronology
condition. In fact, any small perturbation of a stably causal spacetime still
satises the chronology condition (cf. Exercise 7.10.4).
Let (M, g) be a time-oriented spacetime. A smooth future-directed
causal curve c : (a, b) M (with possibly a = or b = +) is said
to be future-inextendible if lim
tb
c(t) does not exist. The denition of
a past-inextendible causal curve is analogous. The future domain of
dependence of S M is the set D
+
(S) of all events p M such that
any past-inextendible causal curve starting at p intersects S. Therefore any
causal inuence on an event p D
+
(S) had to register somewhere in S,
and one can expect that what happens at p can be predicted from data on
S. Similarly, the past domain of dependence of S is the set D

(S) of
all events p M such that any future-inextendible causal curve starting at
p intersects S. Therefore any causal inuence of an event p D

(S) will
register somewhere in S, and one can expect that what happened at p can
be retrodicted from data on S. The domain of dependence of S is simply
the set D(S) = D
+
(S) D

(S).
Let (M, g) be a stably causal spacetime with time function t : M
R. The level sets S
a
= t
1
(a) are said to be Cauchy hypersurfaces if
D(S
a
) = M. Spacetimes for which this happens have particularly good
causal properties.
Definition 7.9. A stably causal spacetime possessing a time function
whose level sets are Cauchy hypersurfaces is said to be globally hyper-
bolic.
Notice that the future and past domains of dependence of the Cauchy
hypersurfaces S
a
are D
+
(S
a
) = t
1
([a, +)) and D

(S
a
) = t
1
((, a]).
Exercises 7.10.
(1) (Time-orientable double covering) Using ideas similar to those of
Exercise 8.6.9 in Chapter 1, show that if (M, g) is a non-time-
orientable Lorentzian manifold then there exists a time-orientable
double covering, i.e. a time-orientable Lorentzian manifold (M, g)
and a local isometry : M M such that every point in M has
two preimages by . Use this to conclude that the only compact
surfaces which admit a Lorentzian metric are the torus T
2
and the
Klein bottle K
2
.
(2) Let (M, g) be a time oriented spacetime and p M. Show that:
(a) I
+
(p) is open;
(b) J
+
(p) is not necessarily closed;
(c) J
+
(p) I
+
(p);
(d) I
+
(p) = int J
+
(p)
(e) if r J
+
(p) and q I
+
(r) then q I
+
(p);
(f) if r I
+
(p) and q J
+
(r) then q I
+
(p);
(g) it may happen that I
+
(p) = M.
7. CAUSALITY 275
(3) Consider the 3-dimensional Minkowski spacetime (R
3
, g), where
g = dt dt +dx dx +dy dy.
Let c : R R
3
be the curve c(t) = (t, cos t, sin t). Show that
although c(t) is null for all t R we have c(t) I
+
(c(0)) for all
t > 0. What kind of motion does this curve represent?
(4) Let (M, g) be a stably causal spacetime and h an arbitrary sym-
metric (2, 0)-tensor eld with compact support. Show that for suf-
ciently small [[ the tensor eld g

:= g + h is still a Lorentzian
metric on M, and (M, g

) satises the chronology condition.

(5) Let (M, g) be the quotient of the 2-dimensional Minkowski space-
time by the discrete group of isometries generated by the map
f(t, x) = (t + 1, x + 1). Show that (M, g) satises the chronology
condition, but there exist arbitrarily small perturbations of (M, g)
(in the sense of Exercise 7.10.4) which do not.
(6) Let (M, g) be a time oriented spacetime and S M. Show that:
(a) S D
+
(S);
(b) D
+
(S) is not necessarily open;
(c) D
+
(S) is not necessarily closed.
(7) Let (M, g) be the 2-dimensional spacetime obtained by remov-
ing the positive x-semi-axis of Minkowski 2-dimensional spacetime
(cf. Figure 8). Show that:
(a) (M, g) is stably causal but not globally hyperbolic;
(b) there exist points p, q M such that J
+
(p) J

(q) is not
compact;
(c) there exist points p, q M with q I
+
(p) such that the
supremum of the lengths of timelike curves connecting p to q
is not attained by any timelike curve.
t t
x x
S
D(S)
p
J
+
(p)
Figure 8. Stably causal but not globally hyperbolic spacetime.
276 6. RELATIVITY
(8) Let (, h) be a 3-dimensional Riemannian manifold. Show that the
spacetime (M, g) = (R , dt dt + h) is globally hyperbolic if
and only if (, h) is complete.
(9) Show that the following spacetimes are globally hyperbolic:
(a) the Minkowski spacetime;
(b) the FLRW spacetimes;
(c) the region r > 2m of Schwarzschild spacetime;
(d) the region r < 2m of Schwarzschild spacetime.
(10) Let (M, g) be a global hyperbolic spacetime with Cauchy hyper-
surface S. Show that M is dieomorphic to R S.
8. Hawking Singularity Theorem
As we have seen in Sections 5 and 6, both the Schwarzschild solution and
the FLRW cosmological models display singularities, beyond which timelike
and null geodesics cannot be continued.
Definition 8.1. A spacetime (M, g) is said to be singular if it is not
geodesically complete.
It was once thought that the examples above were singular due to their
high degree of symmetry, and that more realistic spacetimes would be non-
singular. Following Hawking and Penrose [Pen65, Haw67, HP70], we will
show that this is not the case: any suciently small perturbation of these
solutions will still be singular.
The question of whether a given Riemannian manifold is geodesically
complete is settled by the Hopf-Rinow Theorem. Unfortunately, this theo-
rem does not hold in Lorentzian geometry (essentially because one cannot
use the metric to dene a distance function). For instance, compact man-
ifolds are not necessarily geodesically complete (cf. Exercise 8.12.1), and
the exponential map is not necessarily surjective in geodesically complete
manifolds (cf. Exercise 8.12.2).
Let (M, g) be a globally hyperbolic spacetime and S a Cauchy hypersur-
face with future-pointing unit normal vector eld n. Let c
p
be the timelike
geodesic with initial condition n
p
for each point p S. We dene a smooth
map exp : U M on an open set U R S containing 0 S as
exp(t, p) = c
p
(t).
Definition 8.2. The critical values of exp are said to be conjugate
points to S.
Loosely speaking, conjugate points are points where geodesics starting
orthogonally at nearby points of S intersect.
Let q = exp(t
0
, p) be a point not conjugate to S, and let (x
1
, x
2
, x
3
) be
local coordinates on S around p . Then (t, x
1
, x
2
, x
3
) are local coordinates
on some open set V q. Since

t
is the unit tangent eld to the geodesics
8. HAWKING SINGULARITY THEOREM 277
orthogonal to S, we have g
00
=

t
,

t
_
= 1. On the other hand, we have
g
0i
t
=

t
_

t
,

x
i
_
=
_

t
,
t

x
i
_
=
_

t
,
x
i

t
_
=
1
2

x
i
_

t
,

t
_
= 0
for i = 1, 2, 3, and, since g
0i
= 0 on S, we have g
0i
= 0 on V . Therefore
the surfaces of constant t are orthogonal to the geodesics tangent to

t
. For
this reason, (t, x
1
, x
2
, x
3
) is said to be a synchronized coordinate system.
On this coordinate system we have
g = dt dt +
3

i,j=1

ij
dx
i
dx
j
,
where the functions

ij
:=
_

x
i
,

x
j
_
form a positive denite matrix. Since the vector elds

x
i
can always be
dened along c
p
, the matrix (
ij
) is also well dened along c
p
, even at points
where the synchronized coordinate system breaks down, i.e. at points which
are conjugate to S. These are the points for which (t) := det (
ij
(t)) van-
ishes, since only then will
_

t
,

x
1
,

x
2
,

x
3
_
fail to be linearly independent.
(In fact the vector elds

x
i
are Jacobi elds along c
p
see Exercise 4.8.6
in Chapter 3).
It is easy to see that

0
00
=
i
00
= 0 and
i
0j
=
3

k=1

ik

kj
,
where (
ij
) = (
ij
)
1
and
ij
=
1
2

ij
t
(cf. Exercise 8.12.4). Consequently,
R
00
=
3

i=1
R
i
i00
=
3

i=1
_
_

i
00
x
i


i
i0
t
+
3

j=1

j
00

i
ij

3

j=1

j
i0

i
0j
_
_
=

t
_
_
3

i,j=1

ij

ij
_
_

i,j,k,l=1

jk

il

ki

lj
.
(cf. Chapter 4, Section 1). The quantity
:=
3

i,j=1

ij

ij
278 6. RELATIVITY
appearing in this expression is called the expansion of the synchronized
observers, and has an important geometric meaning:
=
1
2
tr
_
(
ij
)
1

t
(
ij
)
_
=
1
2

t
log =

t
log
1
2
.
Here we have used the formula
(log(det A))

= tr
_
A
1
A

_
which holds for any smooth matrix function A : R GL(n) (cf. Exam-
ple 7.1.4 in Chapter 1). Therefore the expansion yields the variation of the
3-dimensional volume element measured by synchronized observers. More
importantly for our purposes, we see that a singularity of the expansion
indicates a zero of , i.e. a conjugate point to S.
Definition 8.3. A spacetime (M, g) is said to satisfy the strong en-
ergy condition if Ric(V, V ) 0 for any timelike vector eld V X(M).
By the Einstein equation, this is equivalent to requiring that the reduced
energy-momentum tensor T satises T(V, V ) 0 for any timelike vector eld
V X(M). In the case of a pressureless uid with rest density function
C

(M) and unit velocity vector eld U X(M), this requirement

becomes

_
U, V
2
+
1
2
V, V
_
0,
or, since the term in brackets is always positive (cf. Exercise 8.12.5), simply
0. For more complicated matter models, the strong energy condition
produces equally reasonable restrictions.
Proposition 8.4. Let (M, g) be a globally hyperbolic spacetime satisfy-
ing the strong energy condition, S M a Cauchy hypersurface and p S
a point where =
0
< 0. Then the geodesic c
p
contains at least a point
conjugate to S, at a distance of at most
3

0
to the future of S (assuming
that it can be extended that far).
Proof. Since (M, g) satises the strong energy condition, we have R
00
=
Ric
_

t
,

t
_
0 on any synchronized frame. Consequently,

t
+
3

i,j,k,l=1

jk

il

ki

lj
0
on such a frame. Choosing an orthonormal basis (where
ij
=
ij
) and using
the inequality
(tr A)
2
ntr(A
t
A),
which holds for square nn matrices (as a simple consequence of the Cauchy-
Schwarz inequality), it is easy to show that
3

i,j,k,l=1

jk

il

ki

lj
=
3

i,j=1

ji

ij
= tr
_
(
ij
)(
ij
)
t
_

1
3

2
.
8. HAWKING SINGULARITY THEOREM 279
Consequently must satisfy

t
+
1
3

2
0.
Integrating this inequality yields
1

0
+
t
3
,
and hence must blow up at a value of t no greater than
3

0
.
Proposition 8.5. Let (M, g) be a globally hyperbolic spacetime, S a
Cauchy hypersurface, p M and c a timelike geodesic through p orthogonal
to S. If there exists a conjugate point between S and p then c does not
maximize length (among the timelike curves connecting S to p).
Proof. We will oer only a sketch of the proof. Let q be the rst con-
jugate point along c between S and p. Then we can use a synchronized
coordinate system around the portion of c between S and q. Since q is con-
jugate to S, there exists another geodesic c, orthogonal to S, with the same
length t(q), which (approximately) intersects c at q. Let V be a geodesically
convex neighborhood of q, let r V be a point along c between S and q,
and let s V be a point along c between q and p (cf. Figure 9). Then the
piecewise smooth timelike curve obtained by following c between S and r,
the unique geodesic in V between r and s, and c between s and p, connects S
to p and has strictly bigger length than c (by the generalized twin paradox).
This curve can be easily smoothed while retaining bigger length than c.
p
q
r
s
S
c
c
Figure 9. Proof of Proposition 8.5.
280 6. RELATIVITY
Proposition 8.6. Let (M, g) be a globally hyperbolic spacetime, S a
Cauchy hypersurface and p D
+
(S). Then D
+
(S) J

(p) is compact.
Proof. Let us dene a simple neighborhood U M to be a geodesi-
cally convex open set dieomorphic to an open ball whose boundary is a
compact submanifold of a larger geodesically convex open set (therefore U
is dieomorphic to S
3
and U is compact). It is clear that simple neighbor-
hoods form a basis for the topology of M. Also, it is easy to show that any
open cover V

A
has a countable, locally nite renement U
n

nN
by
simple neighborhoods (cf. Exercise 8.12.7).
If A = D
+
(S) J

(p) were not compact, there would exist a countable,

locally nite open cover U
n

nN
of A by simple neighborhoods not admit-
ting any nite subcover. Take q
n
A U
n
such that q
m
,= q
n
for m ,= n.
The sequence q
n

nN
cannot have accumulation points, since any point in
M has a neighborhood intersecting only nite simple neighborhoods U
n
. In
particular, each simple neighborhood U
n
contains only a nite number of
points in the sequence (as U
n
is compact).
Set p
1
= p. Since p
1
A, we have p
1
U
n
1
for some n
1
N. Let
q
n
, U
n
1
. Since q
n
J

(p
1
), there exists a future-directed causal curve c
n
connecting q
n
to p
1
. This curve will necessarily intersect U
n
1
. Let r
1,n
be an
intersection point. Since U
n
1
contains only a nite number of points in the
sequence q
n

nN
, there will exist innite intersection points r
1,n
. As U
n
1
is compact, these will accumulate to some point p
2
U
n
1
(cf. Figure 10).
Because U
n
1
is contained in a geodesically convex open set V , which can
be chosen so that v ((v), exp(v)) is a dieomorphism onto V V , we
have p
2
J

(p
1
): if
1,n
is the unique causal geodesic connecting p
1
to r
1,n
,
parameterized by the global time function t : M R, then the subsequence
of
1,n
corresponding to a convergent subsequence of r
1,n
will converge
to a causal geodesic
1
connecting p
1
to p
2
. If S = t
1
(0) then we have
t(r
1,n
) 0, implying that t(p
2
) 0 and hence p
2
A. Since p
2
, U
n
1
, there
must exist n
2
N such that p
2
U
n
2
.
Since U
n
2
contains only a nite number of points in the sequence q
n

nN
,
an innite number of curves c
n
must intersect U
n
2
to the past of r
1,n
. Let
r
2,n
be the intersection points. As U
n
2
is compact, r
2,n
must accumulate
to some point p
3
U
n
2
. Because U
n
2
is contained in a geodesically con-
vex open set, p
3
J

(p
2
): if
2,n
is the unique causal geodesic connecting
r
1,n
to r
2,n
, parameterized by the global time function, then the subse-
quence of
2,n
corresponding to convergent subsequences of both r
1,n

and r
2,n
will converge to a causal geodesic connecting p
2
to p
3
. Since
J

(p
2
) J

(p
1
) and t(r
2,n
) 0 t(p
3
) 0, we have p
3
A.
Iterating the procedure above, we can construct a sequence p
i

iN
of
points in A satisfying p
i
U
n
i
with n
i
,= n
j
if i ,= j, such that p
i
is connected
to p
i+1
by a causal geodesic
i
. It is clear that
i
cannot intersect S, for
t(p
i+1
) > t(p
i+2
) 0. On the other hand, the piecewise smooth causal curve
obtained by joining the curves
i
can easily be smoothed into a past-directed
8. HAWKING SINGULARITY THEOREM 281
causal curve starting at p
1
which does not intersect S. Finally, such curve is
inextendible: it cannot converge to any point, as p
i

iN
cannot accumulate.
But since p
1
D
+
(S), this curve would have to intersect S. Therefore A
must be compact.
p = p
1
p
2
p
3
U
n
1
U
n
2
Figure 10. Proof of Proposition 8.6.
Corollary 8.7. Let (M, g) be a globally hyperbolic spacetime and p, q
M. Then
(i) J
+
(p) is closed;
(ii) J
+
(p) J

(q) is compact.
We leave the proof of this corollary as an easy exercise. Proposition 8.6
is a key ingredient in establishing the following fundamental result.
Theorem 8.8. Let (M, g) be a globally hyperbolic spacetime with Cauchy
hypersurface S, and p D
+
(S). Then, among all timelike curves connecting
p to S, there exists a timelike curve with maximal length. This curve is a
timelike geodesic, orthogonal to S.
Proof. Consider the set T(S, p) of all timelike curves connecting S to p.
Since we can always use the global time function t : M R as a parameter,
these curves are determined by their images, which are compact subsets of
282 6. RELATIVITY
the compact set A = D
+
(S) J

(p). As it is well known (cf. [Mun00]),

the set C(A) of all compact subsets of A is a compact metric space for the
Hausdor metric d
H
, dened as follows: if d : M M R is a metric
yielding the topology of M,
d
H
(K, L) = inf > 0 [ K U

(L) and L U

(K),
where U

(K) is a -neighborhood of K for the metric d. Therefore, the

closure C(S, p) := T(S, p) is a compact subset of C(A). It is not dicult
to show that C(S, p) can be identied with the set of continuous causal
curves connecting S to p (a continuous curve c : [0, t(p)] M is said to be
causal if c(t
2
) J
+
(c(t
1
)) whenever t
2
> t
1
).
The length function : T(S, p) R is dened by
(c) :=
_
t(p)
0
[ c(t)[dt.
This function is upper semicontinuous, i.e. continuous for the topology
O = (, a) [ a +
in R. Indeed, let c T(S, p) be parameterized by its arclength u. For a suf-
ciently small > 0, the function u can be extended to the -neighborhood
U

(c) in such a way that its level hypersurfaces are spacelike and orthogonal
to c, that is, grad u is timelike and coincides with c on c (cf. Figure 11).
If T(S, p) is in the open ball B

(c) C(A) for the Hausdor metric d

H
then we can use u as a parameter, thus obtaining
du( ) = 1 , grad u = 1.
Therefore can be decomposed as
=
1
grad u, grad u
grad u +X,
where X is spacelike and orthogonal to grad u, and so
[ [ =

1
grad u, grad u
+X, X

1
2
.
Given > 0, we can choose > 0 suciently small so that

1
grad u, grad u
<
_
1 +

2(c)
_
2
on the -neighborhood U

(c) (as grad u, grad u = 1 along c). We have

() =
_
t(p)
0

d
dt

dt =
_
t(p)
0
[ [
du
dt
dt =
_
(c)
u(S)
[ [ du,
where we have to allow for the fact that c is not necessarily orthogonal to
S, and so the initial point of is not necessarily at u = 0 (cf. Figure 11).
8. HAWKING SINGULARITY THEOREM 283
Consequently,
() =
_
(c)
u(S)

1
grad u, grad u
X, X

1
2
du
<
_
(c)
u(S)
_
1 +

2(c)
_
du =
_
1 +

2(c)
_
((c) u( S)) .
Choosing suciently small so that
[u[ <
_
1
(c)
+
2

_
1
on S U

(c), we obtain () < (c) + , proving upper semicontinuity in

T(S, p). As a consequence, the length function can be extended to C(S, p)
through
(c) = lim
0
sup() [ B

(c) T(S, p)
(as for > 0 suciently small the supremum will be nite). Also, it is clear
that if c T(S, p) then the upper semicontinuity of the length forces the
two denitions of (c) to coincide. The extension of the length function to
C(S, p) is trivially upper semicontinuous: given c C(S, p) and > 0, let
> 0 be such that () < (c) +

2
for any B
2
(c) T(S, p). Then it is
clear that (c

) (c) +

2
< (c) + for any c

(c).
Finally, we notice that the compact sets of R for the topology O are the
sets with a maximum. Therefore, the length function attains a maximum
at some point c C(S, p). All that remains to be seen is that the maximum
is also attained at a smooth timelike curve . To do so, cover c with nitely
many geodesically convex neighborhoods and choose points p
1
, . . . , p
k
in c
such that p
1
S, p
k
= p and the portion of c between p
i1
and p
i
is contained
in a geodesically convex neighborhood for all i = 2, . . . , k. It is clear that
there exists a sequence c
n
T(S, p) such that c
n
c and (c
n
) (c).
Let t
i
= t(p
i
) and p
i,n
be the intersection of c
n
with t
1
(t
i
). Replace c
n
by the sectionally geodesic curve
n
obtained by joining p
i1,n
to p
i,n
in
the corresponding geodesically convex neighborhood. Then (
n
) (c
n
),
and therefore (
n
) (c). Since each sequence p
i,n
converges to p
i
,
n
converges to the sectionally geodesic curve obtained by joining p
i1
to p
i
(i = 2, . . . , k), and it is clear that (
n
) () = (c). Therefore is a
point of maximum for the length. Finally, we notice that must be smooth
at the points p
i
, for otherwise we could increase its length by using the
generalized twin paradox. Therefore must be a timelike geodesic. Using
a synchronized coordinate system around (0), it is clear that must be
orthogonal to S, for otherwise it would be possible to increase its length.
We have now all the necessary ingredients to prove the Hawking Singu-
larity Theorem:
284 6. RELATIVITY
p
c
S
u = 0
u = (c)
U

(c)
Figure 11. Proof of Theorem 8.8.
Theorem 8.9. (Hawking) Let (M, g) be a globally hyperbolic spacetime
satisfying the strong energy condition, and suppose that the expansion sat-
ises
0
< 0 on a Cauchy hypersurface S. Then (M, g) is singular.
Proof. We will show that no future-directed timelike geodesic orthog-
onal to S can be extended to proper time greater than
0
=
3

0
to the
future of S. Suppose that this was not so. Then there would exist a future-
directed timelike geodesic c orthogonal to S, parameterized by proper time,
dened in an interval [0,
0
+ ] for some > 0. Let p = c(
0
+ ). Accord-
ing to Theorem 8.8, there would exist a timelike geodesic with maximal
length connecting S to p, orthogonal to S. Because (c) =
0
+, we would
necessarily have ()
0
+ . Proposition 8.4 guarantees that would
develop a conjugate point at a distance of at most
0
to the future of S,
and Proposition 8.5 states that would cease to be maximizing beyond this
point. Therefore we arrive at a contradiction.
Remark 8.10. It should be clear that (M, g) is singular if the condition

0
< 0 on a Cauchy hypersurface S is replaced by the condition

0
> 0 on S. In this case, no past-directed timelike geodesic orthogonal
to S can be extended to proper time greater than
0
=
3

0
to the past of S.
Example 8.11.
(1) The FLRW models are globally hyperbolic (cf. Exercise 7.10.9),
and satisfy the strong energy condition (as > 0). Moreover,

ij
=
a
a

ij
=
3 a
a
.
8. HAWKING SINGULARITY THEOREM 285
Assume that the model is expanding at time t
0
. Then =
0
=
3 a(t
0
)
a(t
0
)
> 0 on the Cauchy hypersurface S = t = t
0
, and hence
Theorem 8.9 guarantees that this model is singular to the past of S
(i.e. there exists a Big Bang). Moreover, Theorem 8.9 implies that
this singularity is generic: any suciently small perturbation of an
expanding FLRW model satisfying the strong energy condition will
also be singular. Loosely speaking, any expanding universe must
have begun at a Big Bang.
(2) The region r < 2m of the Schwarzschild solution is globally hy-
perbolic (cf. Exercise 7.10.9), and satises the strong energy con-
dition (as Ric = 0). The metric can be written in this region as
g = d d +
_
2m
r
1
_
dt dt +r
2
d d +r
2
sin
2
dd,
where
=
_
2m
r
_
2m
u
1
_

1
2
du.
Therefore the inside of the black hole can be pictured as a cylinder
RS
2
whose shape is evolving in time. As r 0, the S
2
contracts
to a singularity, with the t-direction expanding. Since
3

i,j=1

ij
dx
i
dx
j
=
dr
d
_

m
r
2
dt dt +rd d +r sin
2
d d
_
,
we have
=
_
2m
r
1
_

1
2
_
2
r

3m
r
2
_
.
Therefore we have =
0
< 0 on any Cauchy hypersurface S =
r = r
0
with r
0
<
3m
2
, and hence Theorem 8.9 guarantees that
the Schwarzschild solution is singular to the future of S. More-
over, Theorem 8.9 implies that this singularity is generic: any suf-
ciently small perturbation of the Schwarzschild solution satisfying
the strong energy condition will also be singular. Loosely speaking,
once the collapse has advanced long enough, nothing can prevent
the formation of a singularity.
Exercises 8.12.
(1) (Clifton-Pohl torus) Consider the Lorentzian metric
g :=
1
u
2
+v
2
(du dv +dv du)
on M = R
2
0. The Lie group Z acts freely and properly on M
by isometries through
n (u, v) = (2
n
u, 2
n
v),
286 6. RELATIVITY
and this determines a Lorentzian metric g on M = M/Z

= T
2
.
Show that (M, g) is not geodesically complete (although M is com-
pact). (Hint: Look for null geodesics with v 0).
(2) (2-dimensional Anti-de Sitter universe) Consider R
3
with the pseudo-
Riemannian metric
g = du du dv dv +dw dw,
and let (M, g) be the universal covering of the submanifold
H = (u, v, w) R
3
[ u
2
+v
2
w
2
= 1)
with the induced metric. Show that:
(a) a model for (M, g) is M = R
_

2
,

2
_
and
g =
1
cos
2
x
(dt dt +dx dx)
(hence (M, g) is not globally hyperbolic);
(b) (M, g) is geodesically complete, but exp
p
is not surjective for
any p M (Hint: Notice that each isometry of (R
3
, g) determines an
isometry of (M, g));
(c) there exist points p, q M connected by arbitrarily long time-
like curves (cf. Exercise 10).
t
x
(, 0)
(, 0)

2
Figure 12. The exponential map is not surjective in the
2-dimensional Anti-de Sitter universe.
8. HAWKING SINGULARITY THEOREM 287
(3) By analogy with Exercise 3.3.5 in Chapter 3, we can dene a left-
invariant Lorentzian metric on the Lie group H = RR
+
of Exer-
cise 7.17.3 in Chapter 1 as
g :=
1
x
2
(dt dt +dx dx).
Show that this metric is not geodesically complete. (Remark: This
cannot happen in Riemannian geometry cf. Exercise 5.8.4 in Chapter 3).
(4) Show that the Christoel symbols for the metric
g = dt dt +
3

i,j=1

ij
dx
i
dx
j
,
satisfy

0
00
=
i
00
= 0 and
i
0j
=
3

k=1

ik

kj
,
where (
ij
) = (
ij
)
1
and
ij
=
1
2

ij
t
.
(5) Show that if U is a unit timelike vector eld and V is any timelike
vector eld then U, V
2
+
1
2
V, V is a positive function.
(6) Show that a spacetime (M, g) whose matter content is a pressure-
less uid with rest density function C

(M) and a cosmological

constant R (cf. Exercise 6.1.6) satises the strong energy con-
dition if and only if 0 and

4
.
(7) Let (M, g) be a spacetime. Show that any open cover V

A
has
a countable, locally nite renement U
n

nN
by simple neighbor-
hoods (i.e.,
nN
U
n
=
A
V

, for each n N there exists A

such that U
n
V

, and each point p M has a neighborhood

which intersects only nite simple neighborhoods U
n
).
(8) Prove Corollary 8.7.
(9) Let (M, g) be a globally hyperbolic spacetime, t : M R a global
time function, S = t
1
(0) a Cauchy hypersurface, p D
+
(S) and
A = D
+
(S) J

(p). Show that the closure C(S, p) := T(S, p)

in the space C(A) of all compact subsets of A with the Hausdor
metric can be identied with the set of continuous causal curves
connecting S to p (parameterized by t).
(10) Let (M, g) be a globally hyperbolic spacetime and p, q M with
q I
+
(p). Show that among all timelike curves connecting p to
q there exists a timelike curve with maximal length, which is a
timelike geodesic.
(11) Consider two events p and q on Schwarzschild spacetime corre-
sponding to the beginning and the end of a complete circular orbit
of radius r (cf. Exercise 5.1.2). Show that the corresponding time-
like geodesic is not maximal.
288 6. RELATIVITY
(12) Use ideas similar to those leading to the proof of Theorem 8.9 to
prove the Myers Theorem: if (M, , ) is a complete Riemannian
manifold whose Ricci curvature satises Ric(X, X) X, X for
some > 0 then M is compact. Can these ideas be used to prove
a singularity theorem in Riemannian geometry?
(13) Explain why the Hawking Singularity Theorem does not apply to
each of the following spacetimes:
(a) Minkowski spacetime;
(b) Einstein universe (cf. Exercise 6.1.7);
(c) de Sitter universe (cf. Exercise 6.1.7);
(d) 2-dimensional Anti-de Sitter universe (cf. Exercise 2).
9. Penrose Singularity Theorem
Let (M, g) be a globally hyperbolic spacetime, S a Cauchy hypersurface
with future-pointing unit normal vector eld n, and S a compact 2-
dimensional submanifold with unit normal vector eld in S. Let c
p
be the
null geodesic with initial condition n
p
+
p
for each point p . We dene
a smooth map exp : (, ) M for some > 0 as exp(r, p) = c
p
(r).
Definition 9.1. The critical values of exp are said to be conjugate
points to .
Loosely speaking, conjugate points are points where geodesics starting
orthogonally at nearby points of intersect (see also Exercise 4.8.6 in Chap-
ter 3).
Let q = exp(r
0
, p) be a point not conjugate to . If is a local parame-
terization of around p, then we can construct a system of local coordinates
(u, r, x
2
, x
3
) on some open set V q by using the map
(u, r, x
2
, x
3
) exp(r,
u
((x
2
, x
3
))),
where
u
is the ow along the timelike geodesics orthogonal to S and the
map exp : (, )
u
() M is dened as above.
Since

r
is tangent to null geodesics, we have g
rr
=

r
,

r
_
= 0. On
the other hand, we have
g
r
r
=

r
_

r
,

x

_
=
_

r
,
r

_
=
_

r
,
x

r
_
=
1
2

_

r
,

r
_
= 0,
for = 0, 1, 2, 3. Since g
ru
= 1 and g
r2
= g
r3
= 0 on
u
(), we have
g
ru
= 1 and g
r2
= g
r3
= 0 on V . Therefore the metric is written in this
coordinate system as
g = dudududrdrdu+
3

i=2

i
_
du dx
i
+dx
i
du
_
+
3

i,j=2

ij
dx
i
dx
j
.
9. PENROSE SINGULARITY THEOREM 289
Since
det
_
_
_
_
1
2

3
1 0 0 0

2
0
22

23

3
0
32

33
_
_
_
_
= det
_

22

23

32

33
_
,
we see that the functions

ij
:=
_

x
i
,

x
j
_
form a positive denite matrix, and so g induces a Riemannian metric on
the 2-dimensional surfaces exp(r,
u
()), which are then spacelike. Since
the vector elds

x
i
can always be dened along c
p
, the matrix (
ij
) is also
well dened along c
p
, even at points where the coordinate system breaks
down, i.e. at points which are conjugate to . These are the points for
which := det (
ij
) vanishes, since only then will
_

u
,

r
,

x
2
,

x
3
_
fail to
be linearly independent. (In fact the vector elds

x
i
are Jacobi elds along
c
p
see Exercise 4.8.6 in Chapter 3).
It is easy to see that

u
ur
=
u
rr
=
u
ri
=
r
rr
=
i
rr
= 0 and
i
rj
=
3

k=2

ik

kj
,
where (
ij
) = (
ij
)
1
and
ij
=
1
2

ij
r
(cf. Exercise 9.9.1). Consequently,
R
rr
= R
u
urr
+
3

i=2
R
i
irr
=
3

i=2
_
_

i
ir
r

3

j=2

j
ir

i
rj
_
_
=

r
_
_
3

i,j=2

ij

ij
_
_

i,j,k,l=2

jk

il

ki

lj
.
(cf. Chapter 4, Section 1). The quantity
:=
3

i,j=2

ij

ij
appearing in this expression is called the expansion of the null geodesics,
and has an important geometric meaning:
=
1
2
tr
_
(
ij
)
1

r
(
ij
)
_
=
1
2

r
log =

r
log
1
2
,
where := det (
ij
). Therefore the expansion yields the variation of the
area element of the spacelike 2-dimensional surfaces exp(r,
u
()). More
importantly for our purposes, we see that a singularity of the expansion
indicates a zero of , i.e. a conjugate point to
u
().
Definition 9.2. A spacetime (M, g) is said to satisfy the null energy
condition if Ric(V, V ) 0 for any null vector eld V X(M).
290 6. RELATIVITY
It is easily seen that this condition is implied by (but weaker than)
the strong energy condition. By the Einstein equation, it is equivalent to
requiring that the reduced energy-momentum tensor T satises T(V, V ) 0
for any null vector eld V X(M). In the case of a pressureless uid with
rest density function C

(M) and unit velocity vector eld U X(M),

this requirement becomes
U, V
2
0 0.
For more complicated matter models, the null energy condition produces
equally reasonable restrictions.
Proposition 9.3. Let (M, g) be a globally hyperbolic spacetime satis-
fying the null energy condition, S M a Cauchy hypersurface, S a
compact 2-dimensional submanifold with unit normal vector eld in S and
p a point where =
0
< 0. Then the null geodesic c
p
contains at least
a point conjugate to , at an ane parameter distance of at most
2

0
to
the future of (assuming that it can be extended that far).
Proof. Since (M, g) satises the null energy condition, we have R
rr
=
Ric
_

r
,

r
_
0. Consequently,

r
+
3

i,j,k,l=2

jk

il

ki

lj
0.
Choosing an orthonormal basis (where
ij
=
ij
), and using the inequality
(tr A)
2
ntr(A
t
A)
for square n n matrices, it is easy to show that
3

i,j,k,l=2

jk

il

ki

lj
=
3

i,j=2

ji

ij
= tr
_
(
ij
)(
ij
)
t
_

1
2

2
.
Consequently must satisfy

r
+
1
2

2
0.
Integrating this inequality yields
1

0
+
r
2
,
and hence must blow up at a value of r no greater than
2

0
.
We dene the chronological future and the causal future of the
compact surface as
I
+
() =
_
p
I
+
(p) and J
+
() =
_
p
J
+
(p)
(with similar denitions for the chronological past and the causal past
of ). It is clear that I
+
(), being the union of open sets, is itself open, and
9. PENROSE SINGULARITY THEOREM 291
also that J
+
() I
+
() and I
+
() = int J
+
(). On the other hand, it is
easy to generalize Proposition 8.6 (and consequently Corollary 8.7) to the
corresponding statements with compact surfaces replacing points (cf. Exer-
cise 9.9.2). In particular, J
+
() is closed. Therefore
J
+
() = I
+
() = J
+
() I
+
(),
and so, by a straightforward generalization of Corollary 7.5, every point in
this boundary can be reached from a point in by a future-directed null
geodesic. Moreover, this geodesic must be orthogonal to . Indeed, at we
have

u
= n and

r
= n +,
and so the metric takes the form
g = du du du dr dr du +
3

i,j=2

ij
dx
i
dx
j
.
If c : I R M is a future-directed null geodesic with c(0) , its initial
tangent vector
c(0) = u

u
+ r

r
+
3

i=2
x
i

x
i
= ( u + r)n + r +
3

i=2
x
i

x
i
must satisfy
u( u + 2 r) =
3

i,j=2

ij
x
i
x
j
.
Since c is future-directed we must have u + r > 0. On the other hand, by
choosing the unit normal to on S to be either or , we can assume
r 0. If c is not orthogonal to we then have
3

i,j=2

ij
x
i
x
j
> 0 u( u + 2 r) > 0 u > 0.
Now the region where u > 0 and r 0 is clearly a subset of I
+
(), since its
points can be reached from by a sectionally smooth curve composed of an
arc of timelike geodesic and an arc of null geodesic. Therefore, we see that
if c is not orthogonal to then c(t) I
+
() for all t > 0.
Even future-directed null geodesics orthogonal to may eventually enter
I
+
(). A sucient condition for this to happen is given in the following
result.
Proposition 9.4. Let (M, g) be a globally hyperbolic spacetime, S a
Cauchy hypersurface with future-pointing unit normal vector eld n, S
a compact 2-dimensional submanifold with unit normal vector eld in S,
p , c
p
the null geodesic through p with initial condition n
p
+
p
and
q = c
p
(r) for some r > 0. If c
p
has a conjugate point between p and q then
q I
+
().
292 6. RELATIVITY
Proof. We will oer only a sketch of the proof. Let s be the rst
conjugate point along c
p
between p and q. Since q is conjugate to p, there
exists another null geodesic starting at which (approximately) intersects
c
p
at s. The piecewise smooth null curve obtained by following between
and s, and c
p
between s and q is a causal curve but not a null geodesic.
This curve can be easily smoothed while remaining causal and nongeodesic,
and so by the generalization of Corollary 7.5 we have q I
+
().
p
q
s
S

c
p
Figure 13. Proof of Proposition 9.4.
Definition 9.5. Let (M, g) be a globally hyperbolic spacetime and S
a Cauchy hypersurface with future-pointing unit normal vector eld n. A
compact 2-dimensional submanifold S with unit normal vector eld
in S is said to be trapped if the expansions
+
and

of the null geodesics

with initial conditions n + and n are both negative everywhere on .
We have now all the necessary ingredients to prove the Penrose Singu-
larity Theorem.
Theorem 9.6. (Penrose) Let (M, g) be a connected globally hyperbolic
spacetime with a noncompact Cauchy hypersurface S, satisfying the null
energy condition. If S contains a trapped surface then (M, g) is singular.
Proof. Let t : M R be a global time function such that S = t
1
(0).
The integral curves of grad t, being timelike, intersect S exactly once, and
I
+
() at most once. This denes a continuous injective map : I
+
()
9. PENROSE SINGULARITY THEOREM 293
S, whose image is open. Indeed, if q = (p), then all points is some neigh-
borhood of q are images of points in I
+
(), as otherwise there would be a
sequence q
n
S with q
n
q such that the integral curves of grad t through
q
n
would not intersect I
+
(). Letting r
n
be the intersections of these curves
with the Cauchy hypersurface t
1
(t(r)), for some point r to the future of p
along the integral line of grad t, we would have r
n
r, and so r
n
I
+
()
for suciently large n (as I
+
() is open), leading to a contradiction.
Since is trapped (and compact), there exists
0
< 0 such that the
expansions
+
and

of the null geodesics orthogonal to both satisfy

+
,

0
. We will show that there exists a future-directed null geodesic
orthogonal to which cannot be extended to an ane parameter greater
than r
0
=
2

0
to the future of . Suppose that this was not so. Then,
according to Proposition 9.3, any null geodesic orthogonal to would have
a conjugate point at an ane parameter distance of at most r
0
to the future
of , after which it would be in I
+
(), by Proposition 9.4. Consequently,
I
+
() would be a (closed) subset of the compact set
exp
+
([0, r
0
] ) exp

([0, r
0
] )
(where exp
+
and exp

refer to the exponential map constructed using the

unit normals and ), hence compact. Therefore the image of would
also be compact, hence closed as well as open. Since M, and therefore S, are
connected, the image of would be S, which would then be homeomorphic to
I
+
(). But S is noncompact by hypothesis, and we reach a contradiction.

Remark 9.7. It should be clear that (M, g) is singular if the condition

of existence of a trapped surface is replaced by the condition of existence of
an anti-trapped surface, that is, a compact surface S such that the
expansions of null geodesics orthogonal to are both positive. In this case,
there exists a past-directed null geodesic orthogonal to which cannot be
extended to an ane parameter time greater than r
0
=
2

0
to the past of .
Example 9.8.
(1) The region r < 2m of the Schwarzschild solution is globally hy-
perbolic (cf. Exercise 7.10.9), and satises the null energy condition
(as Ric = 0). Since r (or r) is clearly a time function (depending
on the choice of time orientation), it must increase (or decrease)
along any future-pointing null geodesic, and therefore any sphere
of constant (t, r) is anti-trapped (or trapped). Since any Cauchy
hypersurface is dieomorphic to R S
2
, hence noncompact, we
conclude from Theorem 9.6 that the Schwarzschild solution is sin-
gular to past (or future) of . Moreover, Theorem 8.9 implies that
this singularity is generic: any suciently small perturbation of the
Schwarzschild solution satisfying the null energy condition will also
be singular. Loosely speaking, once the collapse has advanced long
enough, nothing can prevent the formation of a singularity.
294 6. RELATIVITY
(2) The FLRW models are globally hyperbolic (cf. Exercise 7.10.9),
and satisfy the null energy condition (as > 0). Moreover, radial
null geodesics satisfy
dr
dt
=
1
a
_
1 kr
2
.
Therefore, if we start with a sphere of constant (t, r) and follow
the orthogonal null geodesics along the direction of increasing or
decreasing r, we obtain spheres whose radii ar satisfy
d
dt
(ar) = ar +a r = ar
_
1 kr
2
.
Assume that the model is expanding, with the Big Bang at t = 0,
and spatially noncompact (in particular k ,= 1). Then, for su-
ciently small t > 0, the sphere is anti-trapped, and hence The-
orem 9.6 guarantees that this model is singular to the past of
(i.e. there exists a Big Bang). Moreover, Theorem 9.6 implies that
this singularity is generic: any suciently small perturbation of
an expanding, spatially noncompact FLRW model satisfying the
null energy condition will also be singular. Loosely speaking, any
expanding universe must have begun at a Big Bang.
Exercises 9.9.
(1) Show that the Christoel symbols for the metric
g = du du du dr dr du +
3

i=2

i
_
du dx
i
+dx
i
du
_
+
3

i,j=2

ij
dx
i
dx
j
satisfy

u
ur
=
u
rr
=
u
ri
=
r
rr
=
i
ur
=
i
rr
= 0 and
i
rj
=
3

k=2

ik

kj
,
where (
ij
) = (
ij
)
1
and
ij
=
1
2

ij
r
.
(2) Let (M, g) be a globally hyperbolic spacetime with Cauchy hyper-
surfaces S
0
and S
1
satisfying S
1
D
+
(S
0
), and S
1
a compact
surface. Show that:
(a) D
+
(S
0
) J

() is compact;
(b) J

() is closed.
(3) Explain why the Penrose Singularity Theorem does not apply to
each of the following spacetimes:
(a) Minkowski spacetime;
(b) Einstein universe (cf. Exercise 6.1.7);
(c) de Sitter universe (cf. Exercise 6.1.7);
10. NOTES ON CHAPTER 6 295
(d) 2-dimensional Anti-de Sitter universe (cf. Exercise 8.12.2).
10. Notes on Chapter 6
10.1. Bibliographical notes. There are many excellent texts on gen-
eral relativity, usually containing also the relevant dierential and Lorentzian
geometry. These range from introductory [Sch02] to more advanced [Wal84]
to encyclopedic [MTW73]. A more mathematically oriented treatment can
be found in [BEE96, ON83] ([GHL04] also contains a brief glance at
pseudo-Riemannian geometry). For more information on special relativity
and the Lorentz group see [Nab92, Oli02]. Causality and the singularity
theorems are treated in greater detail in [Pen87, HE95, Nab88], and in
the original papers [Pen65, Haw67, HP70].
Bibliography
[Ahl79] L. Ahlfors, Complex analysis, McGraw-Hill, 1979.
[AM78] R. Abraham and J. Marsden, Foundations of mechanics, Addison Wesley, 1978.
[Arn92] V. I. Arnold, Ordinary dierential equations, Springer, 1992.
[Arn97] V.I. Arnold, Mathematical methods of classical mechanics, Springer, 1997.
[Aud96] M. Audin, Spinning tops, Cambridge University Press, 1996.
[BEE96] J. Beem, P. Ehrlich, and K. Easley, Global lorentzian geometry, Marcel Dekker,
1996.
[BL05] F. Bullo and A. Lewis, Geometric control of mechanical systems, Springer, 2005.
[Blo96] E. Bloch, A rst course in geometric topology and dierential geometry,
Birk auser, 1996.
[Blo03] A. Bloch, Nonholonomic mechanics and control, Springer, 2003.
[Boo03] W. Boothby, An introduction to dierentiable manifolds and riemannian geom-
etry, Academic Press, 2003.
[BT82] R. Bott and L. Tu, Dierential forms in algebraic topology, Springer, 1982.
[BtD03] T. Brocker and T. tom Diek, Representations of compact lie groups, Springer,
2003.
[CB97] R. Cushmann and L. Bates, Global aspects of classical integrable systems,
Birkhauser, 1997.
[CCL00] S. Chern, W. Chen, and K. Lam, Lectures on dierential geometry, World
Scientic, 2000.
[dC76] M. do Carmo, Dierential geometry of curves and surfaces, Prentice-Hall, 1976.
[dC93] , Riemannian geometry, Birkhauser, 1993.
[dC94] , Dierential forms and applications, Springer, 1994.
[DK99] J. Duistermaat and J. Kolk, Lie groups, Springer, 1999.
[Fre82] M. Freedman, The topology of four-dimensional manifolds, J. Dierential
Geom. 17 (1982), 357453.
[GH02] J. Guckenheimer and P. Holmes, Nonlinear oscillations, dynamical systems,
and bifurcations of vector elds, Springer, 2002.
[GHL04] S. Gallot, D. Hulin, and J. Lafontaine, Riemannian geometry, Springer, 2004.
[Gom83] R. Gompf, Three exotic R
4
s and other anomalies, J. Dierential Geom. 18
(1983), 317328.
[GP73] V. Gillemin and A. Pollack, Dierential topology, Prentice-Hall, 1973.
[GPS02] H. Goldstein, C. Poole, and J. Safko, Classical mechanics, Addison Wesley,
2002.
[Gro70] M. Gromov, Isometric embeddings and immersions in riemannian geometry,
Usp. Mat. Nauk 25 (1970), 362.
[Haw67] S. Hawking, The occurrence of singularities in cosmology. iii. causality and
singularities, Proc. Roy. Soc. Lon. A 300 (1967), 187201.
[HE95] S. Hawking and G. Ellis, The large scale structure of space-time, Cambridge
University Press, 1995.
[Hel01] S. Helgasson, Dierential geometry, lie groups and symmetric spaces, American
Mathematical Society, 2001.
297
298 BIBLIOGRAPHY
[HP70] S. Hawking and R. Penrose, The singularities of gravitational collapse and cos-
mology, Proc. Roy. Soc. Lon. A 314 (1970), 529548.
[Jos02] J. Jost, Riemannian geometry and geometric analysis, Springer, 2002.
[Ker60] M. Kervaire, A manifold wich does not admit any dierentiable structure, Com-
ment. Math. Helv. 34 (1960), 257270.
[KM63] M. Kervaire and J. Milnor, Groups of homotopy spheres. i., Ann. of Math. (2)
77 (1963), 504537.
[KN96] S. Kobayashi and K. Nomizu, Foundations of dierential geometry, vol. I and
II, Wiley, 1996.
[Mil56] J. Milnor, On manifolds homeomorphic to the 7-sphere, Ann. of Math. 64
(1956), 399405.
[Mil59] , Dierentiable structures on spheres, Amer. J. Math. 81 (1959), 962
972.
[Mil97] , Topology from the dierentiable viewpoint, Princeton University Press,
1997.
[Mil07] , On the relationship between dierentiable manifolds and combinatorial
manifolds, Collected Papers III: Dierential Topology, American Mathematical
Society, 2007, pp. 1928.
[Mor98] F. Morgan, Riemannian geometry, A K Peters, 1998.
[MR99] J. Marsden and T. Ratiu, Introduction to mechanics and symmetry, Springer,
1999.
[MTW73] C. Misner, K. Thorne, and J. A. Wheeler, Gravitation, Freeman, 1973.
[Mun00] J. Munkres, Topology, Prentice-Hall, 2000.
[Nab88] G. Naber, Spacetime and singularities an introduction, Cambridge University
Press, 1988.
[Nab92] , The geometry of minkowski spacetime, Springer, 1992.
[Nas56] J. Nash, The imbedding problem for riemannian manifolds, Annals of Mathe-
matics 63 (1956), 2063.
[Nov65] S. P. Novikov, Topological invariance of rational pontrjagin classes, Soviet
Math. Dokl. 6 (1965), 921923.
[Oli02] W. Oliva, Geometric mechanics, Springer, 2002.
[ON83] B. ONeill, Semi-riemannian geometry, Academic Press, 1983.
[Pen65] R. Penrose, Gravitational collapse and space-time singularities, Phys. Rev. Lett.
14 (1965), 5759.
[Pen87] , Techniques of dierential topology in relativity, Society for Industrial
and Applied Mathematics, 1987.
[Rud86] W. Rudin, Real and complex analysis, McGraw-Hill, 1986.
[Sch02] B. Schutz, A rst course in general relativity, Cambridge University Press,
2002.
[Sma60] S. Smale, The generalized poincare conjecture in higher dimensions, Bull. AMS
66 (1960), 373375.
[TW92] E. Taylor and J. Wheeler, Spacetime physics, Freeman, 1992.
[Wal84] R. Wald, General relativity, University of Chicago Press, 1984.
[War83] F. Warner, Foundations of dierentiable manifolds and lie groups, Springer,
1983.
[Whi44a] H. Whitney, The selntersections of a smooth n-manifold in (2n 1)-space,
Annals of Mathematics 45 (1944), 247293.
[Whi44b] , The selntersections of a smooth n-manifold in 2n-space, Annals of
Mathematics 45 (1944), 220246.
[Wol78] J. A. Wolf, Spaces of constant curvature, Publish or Perish, 1978.
Index
1-parameter group
of dieomorphisms, 33
of isometries, 106
-algebra, 231
Aberration, 243
Acceleration
covariant, 133
proper, 244
Action
determined by a Lagrangian, 193
innitesimal, 197
of a discrete group, 43
of a group, 41
Hamiltonian, 222
Poisson, 222
Ane
connection, 98
map, 45, 104, 147
parameter, 102
Alternating tensor, 63
Angle, 93
hyperbolic, 243, 257
Angular
momentum, 163, 173, 198, 201
velocity, 176, 180
Anti-de Sitter universe, 286, 288, 295
Anti-trapped surface, 293
Antipodal map, 17
Arclength, 102
Arnold-Liouville Theorem, 212
Atlas, 12
equivalence, 12
maximal, 12
Ball
normal, 109
open, 118
Basis
associated to a parameterization, 20
change of, 85
dual, 84
equivalence, 48
of a ber bundle, 56
of a topology, 54
orthonormal, 241
orientation, 48
Bi-invariant metric, 113
curvature, 126
geodesic, 113
Levi-Civita connection, 113
Bianchi identity, 120
Biangle, 141
Big Bang, 265
Big Crunch, 266
Birkho Ergodicity Theorem, 213
discrete version, 216
Birkho Theorem, 262
Black hole, 256
Boundary
of a dierentiable manifold with
boundary, 52
of a topological manifold with
boundary, 9
Brachistochrone curve, 199
Brouwer Fixed Point Theorem, 84
Bump function, 87
Bundle
cotangent, 68
ber, 56
tangent, 20, 23, 50
Canonical
immersion, 24
symplectic form, 204
symplectic form with magnetic term,
226
symplectic potential, 204
Car and garage paradox, 242
Cartan connection, 246, 247
299
300 INDEX
Cartan formula, 76
Cartan structure equations, 130
Casimir function, 219
Cauchy hypersurface, 274, 276
Cauchy sequence, 118
Cauchy-Riemann equations, 155
Causal
continuous curve, 282
curve, 268
future, 268, 271, 290
past, 268, 271, 290
Center of mass, 170, 179, 200
Central eld, 162, 210, 214
Centrifugal force, 181
Chain rule, 23
Chandler precession, 181
Change of basis matrix, 85
Change of Variables Theorem, 86
Chart, 11
Chern, 156
Christoel symbols, 98
for the 2-sphere, 163
for the hyperbolic plane, 165
Chronological
future, 268, 271, 290
past, 268, 271, 290
spacetime, 273, 275
Chronology condition, 273, 275
Circle, 6
Circular orbit, 214, 215, 258, 261, 287
Clifton-Pohl torus, 285
Closed
form, 74
set, 54
Coframe, 127
orthonormal, 141
Collapse, 267
Commutator
of matrices, 38
of vector elds, 29
Compact
exhaustion, 88
subset, 54
topological space, 54
Compactly supported
form, 76
function, 83
Complete
integrability, 209
metric space, 118
Riemannian manifold, 115, 276
vector eld, 33
Conguration space, 157
Conformally related metrics, 159
Conjugate point, 114, 276, 288
Connected
subset, 55
sum, 8
topological space, 55
Connection
ane, 98
Cartan, 246, 247
compatible with the metric, 101
forms, 127
Levi-Civita, 101
symmetric, 100
Conservation of Energy Theorem, 159,
190, 195
Conservative
force, 158
mechanical system, 158
mechanical system with magnetic
term, 208
Constant curvature manifold, 124, 141,
263
Constraint
holonomic, 166
non-holonomic, 184
semi-holonomic, 186
true non-holonomic, 186
Continuity equation for an
incompressible uid, 182
Continuous
causal curve, 282
map, 54
Contractible manifold, 75
Contraction
of a tensor, 125
of a tensor by a vector, 67
Contravariant tensor, 62
Convergence of a sequence, 55
Coordinate
chart, 11
neighborhood, 11
Coordinate system, 11
normal, 112, 269
synchronized, 277
Copernican Principle, 263
Coriolis force, 181
Cosmological constant, 267
Cotangent
bundle, 68
space, 68
Covariant
acceleration, 133
tensor, 62
INDEX 301
Covariant derivative
of a 1-form, 101
of a tensor eld, 101
of a vector eld, 98
of a vector eld along a curve, 99
Covariant tensor, 69
Covering, 44
manifold, 44
map, 44
transformation, 44
universal, 44
Critical
density, 266
value, 27
Critical point, 27
nondegenerate, 140
of the action, 194
Curl, 107
Curvature
forms, 129
Gauss, 123, 134, 141, 149, 152, 154
geodesic, 133135, 154
mean, 149
nonpositive, 114
normal, 152
of a bi-invariant metric, 126
of a curve, 153
operator, 113, 119
principal, 149, 152
scalar, 125
sectional, 122, 154
tensor, 121
Curve
brachistochrone, 199
causal, 268
compatible with a holonomic
constraint, 166
compatible with a non-holonomic
constraint, 184
curvature, 153
dierentiable, 17
future-directed, 268
future-inextendible, 274
geodesic, 99
length, 93
past-inextendible, 274
piecewise dierentiable, 111
timelike, 238, 248
variation, 194
Cut locus, 114
Cycloid, 199
Cylinder, 8
DAlembert Principle, 167, 187
Darboux Theorem, 216
Dark energy, 267
De Rham cohomology, 74
De Sitter universe, 268, 288, 294
Deck transformation, 44
Deection of light, 262
Degree of a map, 97
Density
critical, 266
function, 171
of matter, 245
rest, 249
Derivative
covariant, 98, 99, 101
directional, 29, 34, 98
exterior, 72, 73
normal, 96
of a dierentiable map, 20
of a holomorphic function, 155
Dieomorphism, 16
group, 33
local, 16
Dierentiable
action, 42
curve, 17
distribution, 184, 190
form, 70
innitely, 15, 56
manifold, 11
manifold with boundary, 51
map, 15, 56
structure, 12
tensor eld, 69
vector eld, 29
Dierential
form, see also Form
of a function, 68
of a map, 22
Dirac measure, 231
Directional derivative, 29, 34, 98
Dirichlet condition, 97
Discrete group, 43
Distance
between simultaneous events, 233
on a connected Riemannian manifold,
115
Distribution, 184
dierentiable, 184, 190
integrable, 186, 191
orthogonal, 187
Divergence
of a vector eld, 84, 95, 106
302 INDEX
Theorem, 95
Domain of dependence, 274, 275
future, 274
past, 274
Doppler eect, 243
Double covering
orientable, 50
time-orientable, 274
Double pendulum, 170
Dual
basis, 84
space, 61, 84
Dumbbell, 170
Einstein eld equation, 248
Einstein universe, 267, 288, 294
Electric
eld, 200
potential, 199
Embedding, 25, 28
Energy
Conservation Theorem, 159
kinetic, 158
mechanical, 159
potential, 158
Energy-momentum tensor, 249
Enterprise, 245
Equations of structure, 130
Equilibrium point, 230
Equinox precession, 182
Equivalence
class, 55
of atlases, 12
of bases, 48
Principle, 246, 248
relation, 55
Euclidean
space, 92, 120, 130, 144
surface, 144, 147
Euler angles, 177
Euler characteristic, 11, 138, 139, 141
of a Lie group, 148
of the sphere, 139
of the torus, 139
Euler equations, 175
for an incompressible uid, 182
Euler force, 181
Euler top, 175
Euler-Lagrange equations, 194
Event, 233
horizon, 255
simultaneity, 233
Exact form, 74
Expansion, 278, 289
Exponential map
on a Lie group, 41, 113
on a Riemannian manifold, 107, 113
Extended Hamiltonian function, 202
Exterior derivative, 72, 73
External force, 158
conservative, 158
positional, 158
Fermat metric, 251
Fermi-Walker transport, 249
Fiber
bundle, 56
derivative, 195
Field
central, 162, 210, 214
electric, 200
electromagnetic, 199
magnetic, 200
of dual coframes, 127
of frames, 126
tensor, 68
vector, 29
First integral, 209
Fixed point, 41, 230
hyperbolic, 230
nondegenerate, 230
stable, 230
Flow
commuting, 34
geodesic, 112
Hamiltonian, 205
linear, 212
of a left-invariant vector eld, 40
of a vector eld, 32
Fluid
incompressible, 182
perfect, 249
Foliation, 185
leaf, 185
singular, 220
Force
centrifugal, 181
conservative, 158
Coriolis, 181
Euler, 181
external, 158
inertial, 181
positional, 158
reaction, 167, 187
Form, 70
closed, 74
INDEX 303
compactly supported, 76
connection, 127
covariant derivative, 101
curvature, 129
dierentiable, 70
exact, 74
harmonic, 96
Lie derivative, 75
local representation, 71
pull-back, 70
volume, 82
Foucault pendulum, 104
Frame
eld of, 126
inertial, 233
orthonormal, 129
rest, 244
rotating, 250
Free
action, 42
particle, 158, 234
Freedman, 16
Friedmann-Lematre-Robertson-Walker
model, 263, 266, 267, 276, 284, 294
with a cosmological constant, 267
Frobenius Theorem, 186
Fubini Theorem, 89
Full-measure set, 232
Function
bump, 87
compactly supported, 83
continuously dierentiable, 56
dierential, 68
extended Hamiltonian, 202
Hamiltonian, 195
innitely dierentiable, 56
Morse, 141
upper semicontinuous, 282
Casimir, 219
holomorphic, 155
measurable, 231
simple, 232
Fundamental group, 44, 58
Future
causal, 268, 271, 290
chronological, 268, 271, 290
domain of dependence, 274
Future-directed
causal curve, 268
timelike curve, 268
Future-inextendible causal curve, 274
Future-pointing vector, 238, 241
Galileo group, 234
Galileo spacetime, 234
Galileo transformation, 235
Gauss, 152
Gauss curvature, 123, 134, 141, 152, 154
of an isometric embedding, 149
Gauss map, 150
Gauss-Bonnet Theorem, 137
for manifolds with boundary, 140
for non-orientable manifolds, 140
General linear group, 36
General relativity, 248
Geodesic, 99
biangle, 141
completeness, 115
curvature, 133135, 154
ow, 112
homogeneity, 107
maximizing, 281, 287
minimizing, 111
null, 248, 272
of a bi-invariant metric, 113
of the hyperbolic plane, 105
of the hyperbolic space, 147
of the Schwarzschild spacetime, 260
precession, 258
reparameterized, 159, 199
spacelike, 248
timelike, 248
triangle, 104, 141, 257
Geodesically convex neighborhood, 268
Global time function, 273
Globally hyperbolic spacetime, 274, 276
Golfer dilemma, 193
Gompf, 16
Gradient, 94, 95
symplectic, 205
Grassmannian, 47
Gravitational
collapse, 267
potential, 245
redshift, 258
Green formula, 96
Gromov, 92
Group, 57
abelian, 57
action, 41
fundamental, 44, 58
Galileo, 234
general linear, 36
homomorphism, 57
isomorphism, 57
Lie, 36
304 INDEX
Lorentz, 238
orthogonal, 36
rotation, 37
special linear, 37
special orthogonal, 37, 171
special unitary, 38
unitary, 37
Half space, 8, 51
Hamilton equations, 203
Hamiltonian
completely integrable, 209
extended function, 202
ow, 205
function, 195
vector eld, 205, 218
Action, 222
Harmonic
form, 96
oscillator, 164
Hartman-Grobman Theorem, 230
Hausdor metric, 282
Hausdor space, 5, 54
Hawking, 257, 265, 276
Hawking Theorem, 284
Hessian, 140
Hilbert, 148
Hodge -operator, 95
Hodge decomposition, 96
Holomorphic function, 155
Holonomic constraint, 166
Homeomorphism, 54
Homogeneity of geodesics, 107
Homogeneous
Riemannian manifold, 117
space, 42
Homomorphism
of groups, 57
of Lie algebras, 31, 207
Homotopy, 58
invariance, 81
smooth, 75, 81
Hopf-Rinow Theorem, 115
Hubble constant, 265
Hyper-regular Lagrangian, 202
Hyperbolic
angle, 243, 257
plane, 105, 133, 145, 148, 165
space, 134, 142, 147, 249
surface, 145
xed point, 230
Hypersurface, 149
Cauchy, 274, 276
simultaneity, 233
Ice skate, 185, 187, 191
Immersion, 24, 28
canonical, 24
isometric, 148
Impact parameter, 262
Incompressible uid
continuity equation, 182
Euler equation, 182
Independence
of frequencies, 213
of functions, 209
Index of a singularity, 135
Induced
metric, 92
orientation, 53, 79
Inertia
ellipsoid, 181
moment of, 174, 175, 180
Newtons law of, 234
Inertial
force, 181
frame, 233
observer, 235, 243
Innitely dierentiable, 15
function, 56
Innitesimal action, 197
Inner product, 92
Instantaneous rest frame, 244
Integrable
distribution, 186, 191
Hamiltonian, 209
function, 232
Integral
curve, 31
of a compactly supported form, 77
of a compactly supported function, 83
submanifold, 186
Lebesgue, 232
Interior
of a set, 54
point on a manifold with boundary, 9
Inverse Function Theorem, 57
Involution of functions, 209
Isometric immersion, 148
Isometry, 93, 104
group, 144
of the Euclidean plane, 144, 147
of the hyperbolic plane, 145, 148
of the sphere, 104, 146, 148
subgroup, 144
Isomorphism
INDEX 305
of groups, 57
of Lie algebras, 31
Isotropic Riemannian manifold, 123,
126
Isotropy subgroup, 42, 211
Jacobi equation, 113
Jacobi eld, 113
Jacobi identity, 31
Jacobi metric, 159
Jacobi Theorem, 159
Jacobian matrix, 56
KAM Theorem, 213
Kepler problem, 165
Kernel of a group homomorphism, 57
Kervaire, 13, 16
Killing vector eld, 106
Killing-Hopf Theorem, 144
Kinetic energy, 158
Klein bottle, 7, 9, 147, 274
Koszul formula, 102
Kronecker symbol, 39
Kruskal extension, 256
Lagrange top, 178, 181, 215
Lagrangian, 193
G-invariant, 196
action determined by, 193
hyper-regular, 202
Laplace equation, 246
Laplacian operator, 96
Leaf
of a foliation, 185
symplectic, 220
Lebesgue integrable function, 232
Lebesgue integral, 232
Lebesgue measure, 231
Left-invariant
metric, 94, 171
vector eld, 38, 40
Legendre transformation, 202
Leibniz rule, 31, 209, 217
Length
contraction, 241
of a dierentiable curve, 93
of a piecewise dierentiable curve,
111
of a vector, 93
of a vector in Minkowski spacetime,
238
Levi-Civita connection, 101
of a bi-invariant metric, 113
of the hyperbolic plane, 105, 165
of the sphere, 104, 163
Levi-Civita Theorem, 102
Lie algebra, 31, 207
homomorphism, 31, 207
isomorphism, 31
of a Lie group, 38
of the general linear group, 38
of the orthogonal group, 39
of the special linear group, 40
of the special orthogonal group, 40,
174
of the special unitary group, 40
of the unitary group, 40
of vector elds on a manifold, 31
Lie bracket, 30
Lie derivative
of a form, 75
of a function, 35
of a tensor eld, 69
of a vector eld, 35
Lie group, 36
bi-invariant metric, 113
Euler characteristic, 148
exponential map, 41, 113
homomorphism, 40
left-invariant metric, 94, 171
left-invariant vector eld, 38, 40
of isometries, 144
Lie Theorem, 44
Light
cone, 238, 255
deection, 262
Lightlike, see also null
Linear
ow, 212
momentum, 200
linear fractional transformations, 156
Liouville Theorem, 206
Local
dieomorphism, 16
isometry, 93
Local Immersion Theorem, 24
Local representation
of a form, 71
of a map, 16
Loop, 58
Lorentz group, 238
Lorentz transformation, 240
Lorentzian manifold, 247
Lucas problem, 235
Mobius band, 8, 9, 50, 147
Mobius transformation, 17
306 INDEX
Mobius transformations, 156
Magnetic
eld, 200
vector potential, 199
Manifold
contractible, 75
covering, 44
dierentiable, 11
isotropic, 123, 126
Lorentzian, 247
of constant curvature, 124, 141, 263
orientable, 49, 82
orientation, 48, 83
oriented, 49
Poisson, 217
product, 14
pseudo-Riemannian, 247
Riemannian, 92
simply connected, 58
symplectic, 216
topological, 5
with boundary, 8, 51
Map
ane, 45, 104, 147
antipodal, 17
continuity, 54
covering, 44
degree, 97
derivative, 56
dierentiable, 15, 56
dierential, 22
exponential, 41, 107, 113
Gauss, 150
homotopy, 57
orientation preserving, 49
orientation reversing, 49
momentum, 223
Poisson, 221
Mass
center of, 170, 179, 200
distribution, 171
of the Schwarzschild solution, 255
operator, 157
Matrix
change of basis, 85
commutator, 38
exponential, 41
group, 41
Jacobian, 56
Matter density function, 245
Maximal atlas, 12
Mean curvature, 149
Measurable
function, 231
Lebesgue, 231
set, 231
Measure
Dirac, 231
nite, 231
full, 232
Lebesgue, 231
positive, 231
space, 231
zero, 232
Mechanical energy, 159
Mechanical system, 157
conservative, 158
conservative with magnetic term, 208
motion, 158
Metric
bi-invariant, 113
conformally related, 159
Fermat, 251
Hausdor, 282
induced, 92
Jacobi, 159
left-invariant, 94, 171
Minkowski, 236
pseudo-Riemannian, 247
Riemannian, 92
Metric space, 117
completeness, 118
topology, 118
Milnor, 13, 16
Minimal surface, 155
Minkowski metric, 236
Minkowski spacetime, 238, 276, 288, 294
Mixed tensor, 62
Moment of inertia
principal, 175, 180
tensor, 174, 175
Momentum
angular, 163, 173, 198, 201
linear, 200
map, 223
Morse function, 141
Morse Theorem, 140
Motion
of a mechanical system, 158
periodic, 164, 165
Myers Theorem, 288
Nash, 92
Neighborhood, 54
coordinate, 11
geodesically convex, 268
INDEX 307
normal, 107
simple, 280
Neumann condition, 97
Newton equation, 158, 160
for a conservative system, 160
generalized, 167, 187
Newtons law of inertia, 234
Newtons second law, 157
Nodal line, 178
Noether Theorem, 197, 223
Non-holonomic constraint, 184
Nondegenerate
2-tensor, 91
critical point, 140
xed point, 230
Normal
ball, 109
coordinates, 112, 269
curvature, 152
derivative, 96
modes, 170
neighborhood, 107
sphere, 109
subgroup, 57
Null
geodesic, 248, 272
vector, 238, 241
Observer
in a rotating frame, 250
inertial, 235, 243
stationary, 254
Open
ball, 118
cover, 54
equivalence relation, 55
set, 54
Orbit, 165
circular, 214, 215, 258, 261, 287
of a group action, 41
periodic, 164
space, 42
Orientable
double covering, 50
manifold, 49, 82
Orientation
induced on the boundary, 53, 79
number, 49
of a basis, 48
of a manifold, 48, 83
of a vector space, 48
time, 238
Orientation preserving
linear map, 48
map, 49
Orientation reversing
linear map, 48
map, 49
Oriented manifold, 49
Orthogonal
distribution, 187
group, 36
Orthonormal
basis, 241
eld of coframes, 141
eld of frames, 129
Painleve time coordinate, 255, 263
Paradox
car and garage, 242
twin, 242, 243, 250, 271
Parallel
postulate, 104, 105
transport, 100, 104, 105
Parameter
ane, 102
impact, 262
Parameterization, 11
Particle
free, 158, 234
in a central eld, 162, 210, 214
in an electromagnetic eld, 199
on a surface, 166, 167
Partition of unity, 77, 86
Past
causal, 268, 271, 290
chronological, 268, 271, 290
domain of dependence, 274
Past-inextendible causal curve, 274
Pathwise connected space, 186
Pendulum
double, 170
Foucault, 104
simple, 166, 167, 169
spherical, 169
Penrose, 257, 265, 276
Penrose Theorem, 292
Perfect
uid, 249
reaction force, 167, 168, 187, 188
Pericenter, 165
Perihelion precession, 261
Periodic
motion, 164, 165
orbit, 164
Picard-Lindelof Theorem, 31, 99
308 INDEX
Piecewise dierentiable curve, 111
Poincare Recurrence Theorem, 206, 209
Poincare Lemma, 75
Poinsot Theorem, 181
Poisson action, 222
Poisson bivector, 219
Poisson bracket, 207, 217
Poisson equation, 245
Poisson manifold, 217
Poisson map, 221
Polar coordinates, 162
Positional force, 158
Positive denite
2-tensor, 91
linear operator, 174
Potential
electric, 199
energy, 158
gravitational, 245
magnetic, 199
Precession
Chandler, 181
geodesic, 258
of Mercurys perihelion, 261
of the angular velocity, 180
of the equinoxes, 182
Thomas, 250
Principal
axis, 175, 180
curvature, 149, 152
direction, 149
moment of inertia, 175, 180
Principle
Copernican, 263
Equivalence, 246, 248
Relativity, 235
Product
manifold, 14
orientation, 50
topology, 55
Projection
bundle map, 20
stereographic, 14, 17
Projective
plane, 7, 9
space, 14, 43, 84
Proper
acceleration, 244
action, 42
map, 42, 57
time, 215, 239, 248
Pseudo-Riemannian
manifold, 247
metric, 247
Pseudo-rigid body, 182
Pseudosphere, 145
Pull-back
of a covariant tensor, 69
of a form, 70
Push-forward, 22
of a vector eld, 31
Quaternions, 47, 58
Quotient
metric, 94
space, 55
topology, 55
Rank Theorem, 27
Reaction force, 167, 187
perfect, 167, 168, 187, 188
Redshift
Doppler, 243
gravitational, 258
Regular
point, 27
value, 27
Relativity
general, 248
of simultaneity, 241
Principle, 235
special, 238
Reparameterization, 95, 159, 199
Resonant torus, 213
Rest
density, 249
frame, 244
Restricted 3-body problem, 200
Reversed triangle inequality, 243
Ricci tensor, 124, 247, 288
Riemann tensor, 119
Riemannian manifold, 92
complete, 115, 276
homogeneous, 117
isotropic, 123, 126
of constant curvature, 124, 141, 263
volume element, 93
Riemannian metric, 92
quotient, 94
Rigid body, 167, 170
general, 179
mass distribution, 171
symmetry, 180
with a xed point, 171
Rotating frame, 250
Rotation group, 37
INDEX 309
Sagnac eect, 251
Sard Theorem, 97
Scalar curvature, 125
Schur Theorem, 147
Schwarzschild solution, 252, 276, 285,
293
geodesic, 260
mass, 255
Second countability axiom, 5, 54
Second fundamental form, 149, 153
of a distribution, 188
along a normal vector, 149
Sectional curvature, 122, 154
Semi-holonomic constraint, 186
Sequence
Cauchy, 118
convergence, 55
Signature, 241, 247, 249
Simple
neighborhood, 280
pendulum, 166, 167, 169
Simply connected
covering manifold, 44
manifold, 58
Simultaneity
hypersurface, 233
of events, 233
relativity of, 241
Singular
point, 135
spacetime, 276
foliation, 220
Singularity
index, 135
isolated, 135
Smale, 13
Smooth, see also dierentiable
Spacelike
geodesic, 248
vector, 238, 241
Spacetime, 233, 248
chronological, 273, 275
Friedmann-Lematre-Robertson-
Walker, 263, 266, 267, 276, 284,
294
Galileo, 234
globally hyperbolic, 274, 276
Minkowski, 238, 276, 288, 294
Schwarzschild, 252, 276, 285, 293
singular, 276
stably causal, 273, 275
static, 251
time-orientable, 268
time-oriented, 268
Special linear group, 37
Special orthogonal group, 37, 171
Special relativity, 238
Special unitary group, 38
Sphere, 6, 13, 28, 133, 134
curvature, 144, 153
Euler characteristic, 139
isometry, 104, 146, 148
Levi-Civita connection, 104, 163
normal, 109
parallel transport, 104, 105
rolling without slipping, 192
standard dierentiable structure, 14
standard metric, 92, 104
symplectic structure, 226
Spherical pendulum, 169
Stabilizer, 42
Stable xed point, 230
Stably causal spacetime, 273, 275
Standard
dierentiable structure on RP
n
, 15
dierentiable structure on R
n
, 12
dierentiable structure on S
n
, 14
metric on RP
n
, 94
metric on S
2
, 104
metric on S
n
, 92
metric on T
n
, 94
Static spacetime, 251
Stationary
observer, 254
solution, 230
Stereographic projection, 14, 17
Stokes Theorem, 79
Strong energy condition, 278, 287, 289
Structure
dierentiable, 12
equations, 130
functions, 133
Subcover, 54
Subgroup, 57
isotropy, 42, 211
normal, 57
Submanifold, 26, 28
integral, 186
Submersion, 26
Subspace topology, 54
Support of a form, 76
Surface, 14
anti-trapped, 293
Euclidean, 144, 147
hyperbolic, 145
minimal, 155
310 INDEX
of revolution, 134, 154, 170, 214
trapped, 292
Symmetric
2-tensor, 91
connection, 100
Symmetry
of a distance function, 115
of a rigid body, 180
Symplectic
canonical potential, 204
gradient, 205
manifold, 216
leaf, 220
Symplectic form
canonical, 204
canonical with magnetic term, 226
on the sphere, 226
Synchronized coordinate system, 277
Tangent
bundle, 20, 23, 50
space, 18, 22
vector, 18
Tensor, 61
alternating, 63
contraction, 125
contraction by a vector, 67
contravariant, 62
covariant, 62
curvature, 121
energy-momentum, 249
mixed, 62
product, 61
Ricci, 124, 288
Riemann, 119
Poisson, 219
Tensor eld, 68
covariant derivative, 101
Lie derivative, 69
Theorema Egregium, 152
Thomas precession, 250
Time
average, 212
coordinate, 215
dilation, 240
function, 233, 273
orientation, 238
Painleve coordinate, 255, 263
proper, 215, 239, 248
Time-orientable
double covering, 274
spacetime, 268
Time-oriented spacetime, 268
Timelike
curve, 238, 248
geodesic, 248
vector, 238, 241
Topological
manifold, 5
manifold with boundary, 8
space, 54
Topology, 54
basis, 54
metric, 118
product, 55
subspace, 54
Torsion, 100
Torus, 7, 43, 274
Clifton-Pohl, 285
Euler characteristic, 139
of revolution, 7
resonant, 213
Totally geodesic submanifold, 153
Totally normal neighborhood, 111
Tractrix, 148
Tractroid, 148
Transitive action, 41
Transverse vector eld, 140
Trapped surface, 292
Triangle inequality, 115
reversed, 243
Triangulation, 10, 138
True non-holonomic constraint, 186
Twin paradox, 242
generalized, 243, 271
on a cylinder, 250
Unitary group, 37
Universal covering, 44
Universe, 263
anti-de Sitter, 286, 288, 295
de Sitter, 268, 288, 294
Einstein, 267, 288, 294
Upper semicontinuous function, 282
Variation of a curve, 194
Vector
future-pointing, 238, 241
length, 93
lightlike, see also null
null, 238, 241
spacelike, 238, 241
tangent, 18
timelike, 238, 241
Vector eld, 29
f-related, 34
INDEX 311
along a curve, 99
commutator, 29
commuting, 30, 34
compatible with a distribution, 186
complete, 33
covariant derivative, 98
divergence, 84, 95, 106
ow, 32
Hamiltonian, 205, 218
Jacobi, 113
Killing, 106
left-invariant, 38, 40
Lie algebra of, 31
Lie derivative, 35
on a submanifold, 34
parallel along a curve, 99
push-forward, 31
singular point, 135
transverse, 140
velocity, 249
Velocity
addition formula, 242
angular, 176
vector eld, 249
Volume
form, 82
of a compact manifold, 83
Volume element, 82
Riemannian, 93
Wedge product, 63
Wheel rolling without slipping, 184,
187, 189, 191
White hole, 256
Whitney Theorem, 28, 92
Zero measure set, 78, 232