Diffential Geometry: Lecture Notes

Dmitri Zaitsev
D. Zaitsev: School of Mathematics, Trinity College Dublin, Dublin 2, Ireland
E-mail address: [email protected]

Contents
Chapter 1. Introduction to Smooth Manifolds
1. Plain curves
2. Surfaces in R3
3. Abstract Manifolds
4. Topology of abstract manifolds
5. Submanifolds
6. Differentiable maps, immersions, submersions and embeddings

5
5
7
9
12
14
16

Chapter 2. Basic results from Differential Topology

1. Manifolds with countable bases
2. Partition of unity
3. Regular and critical points and Sards theorem
4. Whitney embedding theorem

19
19
20
22
25

Chapter 3. Tangent spaces and tensor calculus

1. Tangent spaces
2. Vector fields and Lie brackets
3. Frobenius Theorem
4. Lie groups and Lie algebras
5. Tensors and differential forms
6. Orientation and integration of differential forms
7. The exterior derivative and Stokes Theorem

27
27
31
33
34
36
37
38

Chapter 4. Riemannian geometry

1. Riemannian metric on a manifold
2. The Levi-Civita connection
3. Geodesics and the exponential map
4. Curvature and the Gauss equation

39
39
41
44
46

CHAPTER 1

Introduction to Smooth Manifolds

Even things that are true can be proved.
Oscar Wilde, The Picture of Dorian Gray
1. Plain curves
Definition 1.1. A regular arc or regular parametrized curve in the plain R 2 is any continuously
differentiable map f : I R2 , where I = (a, b) R is an open interval (bounded or unbounded:
a < b ) such that the R2 -valued derivative f 0 (t) is different from 0 = (0, 0) for all
t I. That is for every t I, f (t) = (f1 (t), f2 (t)) R2 and either f10 (t) 6= 0 or f20 (t) 6= 0.
The variable t I is called the parameter of the arc. One may also consider closed intervals,
in that case their endpoints require special treatment, well see them as boundary points.
Remark 1.2. There is a meaningful theory of nondifferentiable merely continuous arcs (including exotic examples such as Peano curves covering a whole square in R2 ) and of more restrictive
injective continuous arcs (called Jordan curves) that is beyond the scope of this course.
The assumption f 0 (t) 6= 0 roughly implies that the image of f looks smooth and can be
locally approximated by a line at each point. A map f with f 0 (t) 6= 0 for all t is also called
immersion.
Example 1.3. Without the assumption f 0 (t) 6= 0 the image of f may look quite unpleasant.
For instance, investigate the images of the following C maps:

1/t

t<0
(0, e )
2 3
f (t) = (t , t ), f (t) := (0, 0)
t=0

(e1/t , 0) t > 0.

The first curve is called Neil parabola or semicubical parabola. Both maps are not regular at
t = 0. Such a point is called a critical point or a singularity of the map f .
Definition 1.4. A regular curve is an equivalence class of regular arcs, where two arcs f : I
2
R and g : J R2 are said to be equivalent if there exists a bijective continuously differential
map : I J with 0 (t) > 0 for all t I (the inverse 1 is then exists and is automatically
continuously differentiable) such that f = g , i.e. f (t) = g((t)) for all t. Sometimes a finite (or
even countable) union of curves is also called a curve. A regular curve of class C k for 1 k
is an equivalence class of regular arcs of class C k (i.e. k times continuously differentiable), where
the equivalence is defined via maps that are also of class C k . Smooth usually stays for C .
5

1. INTRODUCTION TO SMOOTH MANIFOLDS

Recall that a map f = (f1 , . . . , fn ) from an open set in Rm into Rn is of class C k (k times
continuously differentiable) if all partial derivatives up to order k of every component f i exist and
are continuous everywhere on .
A map as above will be called orientation preserving diffeomorphism between I and J; this
will be defined later in a more general context.
Exercise 1.5. Give an example of a regular curve which is of class C 1 but not C 2 .
The map (t) := tan t defines a C diffeomorphism between the intervals (/2, /2) and
(, ). Hence any regular curve can be parametrized by an arc defined over a bounded interval
(why?).
Exercise 1.6. Give a parametrization of the line passing through 0 and a point (a, b) 6= 0
over the interval (0, 1).
Example 1.7. One of the most important curves is the unit circe. Its standard parametrization
is given by f (t) = (cos t, sin t), t R. Clearly f is not injective.
Exercise 1.8. Show that the circle cannot be parametrized by an injective regular arc.
Definition 1.9. If a regular curve C is parametrized by an arc f , a tangent vector to C at
a point p = f (t0 ) is any multiple (positive, negative or zero) of the derivative f 0 (t0 ). The tangent
line to C at p is parametrized by t 7 p + f 0 (t0 )t. (Sometimes t 7 f 0 (t0 )t is also called the tangent
line).
Exercise 1.10. Show that the tangent line define above is independent of the choice of the
parametrizing arc.
Exercise 1.11. In the above notation show that, if tn I is any sequence converging to
t0 , then f (tn ) 6= f (t0 ) for n sufficiently large and that the line passing through f (t0 ) and f (tn )
converges to the tangent line to C at p. Here convergence of lines can be defined as convergence
of their unital directional vectors.

2. SURFACES IN R3

2. Surfaces in R3
The main difference between curves and surfaces is that the latter in general cannot be
parametrized by a single regular map defined in an open set in R2 . The simplest example is
the unit sphere. Hence one may need different parametrizations for different points.
Definition 1.12. A parametrized (regular) surface element (or surface patch) is a C 1 map
f : U R3 , where U R2 is an open set, which is an immersion. The latter condition means
that, for every a U , the differential da f : R2 R3 is injective. Recall that
da f (u1 , u2 ) = u1

f
f
(a) + u2
(a).
t1
t2

The image of da f is called the tangent plane at f (a).

The condition that the differential da f is injective is equivalent to linear independence of the
f
f
partial derivative vectors t
(a) and t
(a). These span the tangent plane at p = f (a).
1
2
Example 1.13. Every open set U R2 and every C 1 function f on U gives a surface in R3
via its graph
{(t1 , t2 , f (t1 , t2 )) : (t1 , t2 ) U }.
p
In particular, a hemisphere is obtained for f (t1 , t2 ) = 1 t21 t22 with U the open disc given by
t21 + t22 < 1. One needs at least 2 surface elements to cover the sphere.
Exercise 1.14. Give two parametrized surface elements covering the unit sphere in R3 . (Hint:
Use stereographic projection).
Example 1.15. A torus (of revolution) is an important surface that admits a global (but not
injective) parametrization:


f (u, v) = (a + b cos u) cos v, (a + b cos u) sin v, b sin u , 0 < b < a, (u, v) R2 .
(2.1)

More generally, a surface of revolution is obtained by rotating a regular plane curve C

parametrized by t 7 (x(t), z(t)) (called the meridian or profile curve) in the (x, z)-plane around
the z-axis in R3 , where C is assumed not to intersect the z-axis (i.e. x(t) 6= 0 for all t). It admits
a parametrization of the form


(t, ) 7 x(t) cos , x(t) sin , z(t) .
(2.2)
Another important class of surfaces consists of ruled surfaces. A ruled surface is obtained by
moving a line in R3 and admits a parametrization of the form
(t, u) 7 p(u) + tv(u) R3 ,

t R,

u I,

(2.3)

where I is an interval in R, p, v : I R3 are C 1 maps with v nowhere vanishing.

Exercise 1.16. Show that the map (2.2) defines a regular surface element. Find a condition
on p and v in order that (2.3) define a regular surface element.
Exercise 1.17. Show that the hyperboloid given by x2 + y 2 z 2 = 1 is a ruled surface and
the hyperboloid given by x2 + y 2 z 2 = 1 is not a ruled surface.

1. INTRODUCTION TO SMOOTH MANIFOLDS

A general surface in the large is roughly defined by patching together surface elements. A
precise definition (as an immersed 2-dimensional submanifold) will be given later. Here we give a
definition of an embedded regular surface.
Definition 1.18. A subset S R3 is a embedded regular surface if, for each p S, there
exist an open neighborhood Vp of p in R3 and a parametrized regular surface element fp : Up
R2 Vp S which is a homeomorphism between Up and Vp S. The map fp : Up S is called a
parametrization of S around p.
The most important consequence of the above definition is the fact that the change of parameters is a diffeomorphism:
Theorem 1.19. If fp : Up S and fq : Uq S are two parametrizations as in Definition 1.18
such that fp (Up ) fq (Uq ) = W 6= , then the maps
(fq1 fp ) : fp1 (W ) fq1 (W ),

(fp1 fq ) : fq1 (W ) fp1 (W )

(2.4)

are continuously differentiable.

The proof is based on the Implicit Function Theorem (or the Inverse Function Theorem),
quoted here without proof:
Theorem 1.20 (Implicit Function Theorem). Consider an implicit equation
F (x, y) = 0

(2.5)

for a function y = f (x), where x = (x1 , . . . , xm ) Rm , y = (y1 , . . . , yn ) Rn , and F =

(F1 , . . . , Fn ) is a C 1 map from an open neighborhood U V of a point (a, b) in Rm Rn into
Rn . Suppose that F (a, b) = 0 and the square matrix
 F

i
(a, b)
(2.6)
yj
1i,jn

is invertible. Then (2.5) is uniquely solvable near (a, b), i.e. there exists a possibly smaller open
neighborhood U 0 V 0 U V of (a, b) in Rm Rn and a C 1 map f : U 0 V 0 such that, for
(x, y) U 0 V 0 , (2.5) is equivalent to y = f (x). If, moreover, F is of class C k , k > 1, then f is
also of class C k .
An immediate consequence is the Inverse Function Theorem (sometimes the Implicit Function
Theorem is deduced from the Inverse Function Theorem).
Corollary 1.21 (Inverse Function Theorem). Let G be a C 1 map from an open neighborhood
V of a point b in Rn into Rn with a := G(b). Assume that the differential of G at b is invertible.
Then G is also invertible near b, i.e. there exists an open neighborhood V 0 V of b in Rn such that
G(V 0 ) is open in Rn , G : V 0 G(V 0 ) is bijective onto and the inverse G1 is C 1 . If, moreover, G
is C k for k > 1, then also G1 is C k .
Proof of Theorem 1.19. Fix b fq1 (W ) and set a := fp1 (fq (b)). By Definition 1.12, the
differential da fp : R2 R3 is injective. After a possible permutation of coordinates in R3 , we may
assume that the differential da (fp1 , fp2 ) is invertible, where fp = (fp1 , fp2 , fp3 ). Then, by the Inverse

3. ABSTRACT MANIFOLDS

Function Theorem, the map (t1 , t2 ) 7 (fp1 (t1 , t2 ), fp2 (t1 , t2 )) has a C 1 local inverse defined in a
neighborhood of (fp1 (a), fp2 (a)) that we denote by . Then fp1 fq = (fq1 , fq2 ) near b proving
the conclusion of the theorem for the second map in (2.4). The proof for the first map is completely
analogous.

3. Abstract Manifolds
Definition 1.22. A n-manifold (or an n-dimensional differentiable manifold) of class C k is
a set M together with a family (U )A of subsets and injective maps : U Rn whith open
images (U ) Rn such that the union A U covers M and for any , A with U U 6= ,
the sets (U U ) and (U U ) are open in Rn and the composition map (called transition
map)
( 1
(3.1)
) : (U U ) (U U )
is of class C k .

A family (U , )A as in Definition 1.22 is called a C k atlas on M . Thus a C k manifold is a

set with a C k atlas. A C manifold is often called smooth manifold and a C 0 manifold a topological
manifold.
A comparison between Definition 1.22 and the defintion of an embedded regular surface (Definition 1.18) shows that the essential point (except for the change of dimension from 2 to n) was
to distinguish the fundamental peroperty of the transition maps (3.1) (which is Theorem 1.19 for
surfaces) and incorporate it as an axiom. This is the condition that will allow us to carry over
ideas of differential calculus in Rn to abstract differential manifolds.
The condition on the transition maps (3.1) is only nontrivial in case there are at least two
maps in the atlas. Hence, if the atlas consists of a single map, the manifold is C as in some of
the examples below.
Example 1.23. The most basic example of an n-manifold (of class C ) is the space Rn itself
with the atlas consisting of the identity map = id from U = Rn into Rn . More generally, every
open subset U Rn is an (C ) n-manifold with the atlas consisting again of the identity map.
Example 1.24. A regular curve in R2 as defined in Definition 1.4, parametrized by an injective
regular arc f : I R2 defines a 1-manifold structure on the image f (I) via the atlas consisting of
the inverse map := f 1 defined on U := f (I).
Example 1.25. The simplest example of a 1-manifold for which one needs at least two maps
in an atlas is the unit circle S 1 := {(x, y) R2 : x2 + y 2 = 1}. An atlas with four maps can
be given as follows. Let U1 , . . . , U4 be subsets of S 1 given by x < 0, x > 0, y < 0 and y > 0
respectively. Then a (C ) atlas is given by the maps : U R, = 1, . . . , 4, by 1 (x, y) := y,
2 (x, y) := y, 3 (x, y) := x and 4 (x, y) := x. Indeed, it is easy to see that each is injective
with open image (U ) = (1,
1) R, that the sets in (3.1) are open and the transition maps
1

are C . E.g., (2 4 )(t) = 1 t2 is smooth on 4 (U2 U4 ) = (0, 1) R.

Exercise 1.26. Give an atlas on S 1 consisting only of 2 different maps .

10

1. INTRODUCTION TO SMOOTH MANIFOLDS

Example 1.27. An embedded regular surface S in R3 as in defined in Definition 1.18 can be

given a structure of a 2-manifold with the atlas consisting of the inverse maps fp1 : Vp S R2 .
The required property for the transition maps follows from Theorem 1.19.
Example 1.28. Let M R3 be the torus obtained as the image of the map f : R2 R3
defined by (2.1). Then M has a natural structure of a smooth 2-manifold given by the atlas
(U , )A , where A := R2 , V := (1 , 2 + 1 ) (2 , 2 + 2 ) R2 for = (1 , 2 ) R2 ,
U := f (V ) M and := (f |V )1 . Each transition map 1
equals the identity.
The following is an exotic manifold known as the real line with double point.
Example 1.29. Take M := R {a}, where a is any point not in R, and define a C atlas
on M as follows. Set U1 := R and U2 := R \ {0} {a}, then clearly M = U1 U2 . Further define
i : Ui R, i = 1, 2, by 1 (x) = x and 2 (x) := x for x 6= a and 2 (a) := 0. Then all assumptions
of Definition 1.22 are satisfied (why?).
An easy way to obtain new manifolds is to take products. Let M be an n-manifold with an
atlas (U , )A and M 0 be an n0 -manifold with an atlas (U0 , 0 )B , both of class C k . Recall
that M M 0 is the set of all pairs (x, x0 ) with x M and x0 M 0 . Then the collection of maps
0

0 : U U Rn Rn = Rn+n ,

( 0 )(x, y) := ( (x), 0 (y)),

(, ) C := AB,

defines a C k atlas on M M 0 making it an (n + n0 )-manifold of the same differentiability class

(why?).
Any atlas as in Defintion 1.22 can be always completed to a maximal one by involving maps
more general than that play the role of coordinates. It is very convenient and important for
applications to have those general charts at our disposal.
Definition 1.30. An injective map from a subset U M into Rn with open image (U )
n
R is called a (coordinate) chart compatible with the atlas on M or simply a chart on M if, by
adding it to the given atlas one obtains a new (C k ) atlas on M , i.e. if, for every A, the images
1
(U U ) and (U U ) are open in Rn and the maps 1
are C k in their
and
domains of definition.
In particular, any map : U Rn in Definition 1.22 is a chart on M . The inverse
1 : (U ) U of a chart is called a (local) parametrizations of M and (U ) Rn the parameter
domain.
Lemma 1.31. The set of all charts on M is a maximal atlas, i.e. it is an atlas and it cannot
be extended to a larger family of maps satisfying the requirements of Definition 1.22.
Proof. Since each is a chart, the union of all charts covers M . In order to show that the
set of all charts is an atlas, consider two charts : U Rn and : V Rn with U V 6= .
We first show that (U V U ) is open (in Rn ) for every A. For this observe that both
(U U ) and (V U ) are open and hence so is their intersection W := (U V U ). Since
the map h := ( 1 ) : (U U ) (U U ) is continuous, the set (U V U ) = h1 (W )
is open. Interchanging the roles of and we conclude that (U V U ) is also open. Now,

3. ABSTRACT MANIFOLDS

11

since M = U , (U V ) = (U V U ) is open as a union of open sets and the same

holds for (U V ) = (U V U ).
In order to show that g := ( 1 ) : (U V ) (U V ) is of class C k , it is enough to show
that so is the restriction of g to (U V U ) for every A. The latter is C k as composition of the
C k maps ( 1 ) : (U V U ) (U V U ) and (1
) : (U V U ) (U V U ).
Thus the family of all charts satifies the requirements of Definition 1.22 and hence forms an atlas.
Maximality follows directly: if there is a larger family of maps as in Definition 1.22, each of
them must be a chart and hence the family is that of all charts.

Definition 1.32. The set of all charts on M is called a differentiable structure on M .
The same differentiable structure can be obtained through a different atlas. Indeed, it follows
from the proof of Lemma 1.31 that any covering of M by charts is an atlas defining the same
differentiable structure.
Exercise 1.33. Show that the atlas consisting of the map (x) = x3 defines a differentiable
structure on R which is different from the standard one (defined in Example 1.23). That is,
construct a chart for one atlas which is not a chart for the other atlas.
We shall use the convention that for a chart : U Rn on M , we shall write (x1 , . . . , xn ) for
the standard coordinates of Rn and identify them with corresponding functions on U as follows.
For p U , xi (p) is the ith coordinate of (p), i.e. (p) = (x1 (p), . . . , xn (p)) Rn . We call
(x1 (p), . . . , xn (p)) coordinates of p.

12

1. INTRODUCTION TO SMOOTH MANIFOLDS

4. Topology of abstract manifolds

Let M be an n-manifold with an atlas (U , )A as defined in previous section. Then M
carries canonical topology called manifold topology defined as follows. A set U M is said to be
open if for every A, the image (U U ) is open in Rn . Recall that a topological space is a
set X together with a collection of subsets of X called open sets such that (1) the empty set and
the whole set X are open; (2) union of any family of open sets is open; (3) intersection of any two
open sets is open.
Lemma 1.34. With open sets defined above M is a topological space.
Exercise 1.35. Prove the lemma.
Lemma 1.36. If : U Rn is any chart on M (see Definition 1.30), then U is open in M
and is a homeomorphism onto its image (U ) Rn . In particular, any is a homeomorphism
onto its image.
Proof. By Definition 1.30, : U (U ) is a bijection. To see that it is continuous we have
to prove that 1 (V ) is open in M for any open set V in Rn . According to the definition of open
sets this means that (1 (V ) U ) has to be open for any . The latter set coincides with
h1 (V ) where the transition map h := ( 1
) : (U U ) (U U ) is continuous. Hence
(1 (V ) U ) = h1 (V ) is open as desired.
To see that 1 is continuous we have to prove that (1 )1 (W ) = (W U ) is open in Rn for
any open set W in M . According to the definition of open sets, (W U ) is open for every .
Since (U, ) is a chart, (U U ) is open and hence so is the intersection (W U U ). Since
the transition map h := 1 is continuous, the preimage h1 ( (W U U )) = (W U U )
is open. Then also the union (W U U ) = (W U ) is open as desired.

As a consequence, the topology on M depends only on the differentiable structure but not on
the atlas defining this structure (why?).
Exercise 1.37. Check that the topology defined above is the only one for which every map
is a homeomorphism onto its image (U ).
The following observation can be used to obtain new examples of manifolds generalizing Example 1.23. Any open subset U in a manifold M is automatically a manifold with induced differentiable structure defined as follows. If (U , )A is an atlas on M , the restrictions (U
U, |U )A define an atlas on U .
Exercise 1.38. Check that the latter is indeed an atlas on U .
An important property of the topology of a manifold is that locally (or in the small) it is the
same as that of Rn . This is illustrated by the following statements.
Lemma 1.39. A manifold M is always locally compact and locally connected. That is, every
point p M has a fundamental system of open neighborhoods in M which are relatively compact
and connected.

4. TOPOLOGY OF ABSTRACT MANIFOLDS

13

Proof. For any p M , there exists a map : U Rn from the atlas with p U . Then
n
take Vp := 1
(B) with B being sufficiently small open ball in R with center (p). The desired
properties of Vp follow from the fact that is a homeomorphism onto its image.

A basic question about any topological space: Is it Hausdorff? Recall that a topological space
X is called Hausdorff (or satisfying T2 -axiom) if any two distinct points p, q X have disjoint
open neighborhoods, i.e. open sets Up , Uq X with p Up , q Uq and Up Uq = . A manifold
does not have to be Hausdorff in general (see e.g. Example 1.29 or Berger-Gostiaux, 2.2.10.4). In
most books (including Berger-Gostiaux) a manifold is always assumed to be Hausdorff. We shall
also make this assumption in the sequel.

14

1. INTRODUCTION TO SMOOTH MANIFOLDS

5. Submanifolds
We begin the discussion of submanifolds by introducing submanifolds of Rn which form an
important class of manifolds. This is a direct generalization of Definition 1.18 of an embedded
surface in R3 .
Definition 1.40. A subset S Rn is an m-dimensional submanifold of Rn of class C k
(m n), if, for each p S, there exist an open neighborhood Vp of p in Rn , an open set p Rm
and a homeomorphism fp : p Vp S which is C k and regular in the sense that for every a p ,
the differential da fp : Rm Rn is injective. The map fp : p S is called a (local) parametrization
of S around p.
An m-dimensional submanifold S of Rn of class C k carries a structure of an m-manifold of class
C k with the atlas (Up , p )pS given by Up := Vp S, p := fp1 , where Vp and fp are as the above
definition. A direct generalization of Theorem 1.19 (with the same proof based on the Inverse
Function Theorem) implies that the transition maps (3.1) are of class C k . Hence the collection
(Up , p )pS is indeed a C k atlas on M .
The following is an important characterization of submanifolds of Rn .
Theorem 1.41. Let S be a subset of Rn . The following properties are equivalent:
(i) S is an m-dimensional submanifold of Rn of class C k .
(ii) For every p S there exists an open neighborhood V Rn of p and a C k diffeomorphism
between V and an open set in Rn such that S V = 1 (Rm {0}), where Rm Rn
is the set of vectors whose last n m coordinates are 0.
(iii) For every p S there exists an open neighborhood V Rn of p and C k functions
gi : V R, i = 1, . . . , n m, such that the gradients gi (x) are linearly independent
and S V = {x V : g1 (x) = = gnm (x) = 0}.
(iv) For every p S there exists an open neighborhood V Rn and a submersion g : V
Rnm (i.e. the differential da g : Rn Rnm is surjective for each a V ) such that
S V = g 1 (0).
(v) S is locally a graph of a C k function from Rm to Rnm , i.e. for every p S having,
after possible reordering the coordinates, the form p = (p1 , . . . , pn ) Rn , there exist
neighborhoods V1 Rm of (p1 , . . . , pm ) and V2 Rnm of (pm+1 , . . . , pn ) and a C k map
h : V1 V2 such that S (V1 V2 ) = {(x, h(x)) : x V1 } (the graph of h).

In (i) we think of a submanifold as locally regularly parametrized, in (ii) as locally diffeomorphic

to an open set in Rm embedded into Rn in the standard way, in (iii) as given by a regular system
of equations, in (iv) as the zero-set of a submersion, and in (v) as locally a graph of a smooth
map.

Proof of Theorem 1.41. Given (i) and a local parametrization fp : p S, we embedd

p Rm into Rn in the standard way and extended fp to a neighborhood of a := fp1 (p) in Rn
as follows. After a permutation of coordinates, we first assume that da (fp1 , . . . , fpm ) : Rm Rm is

5. SUBMANIFOLDS

15

invertible. Then set

(u1 , . . . , un ) := fp (u1 , . . . , um ) + (0, . . . , 0, um+1 , . . . , un ).
Then it is easy to check that da : Rn Rn is invertible and hence, by the Inverse Function
Theorem, := 1 exists and satisfies (ii) for a possibly smaller neighborhood V of p.
Given (ii), we can take gi := xm+i 1 , i = 1, . . . , n m, to satisfy (iii). Given (iii), we obtain
(iv) with g being a restriction of (g1 , . . . , gnm ) to a suitable neighborhood of x.
Given (iv), observe that the n (n m) matrix associated with dp g has the maximal rank
n m. Therefore, after a permutation of coordinates of Rn , we can apply Theorem 1.20 (Implicit
Function Theorem) whose conclusion implies (v).
Finally, given (v), a local parametrization of S as in Definition 1.40 can be easily obtained by
setting p := V1 , fp (x) := (x, h(x)). Hence (v) implies (i).

An important special case of a submanifold of Rn is a hypersurface which is by definition, an
(n 1)-dimensional submanifold. According to property (iii) in Theorem 1.41 a subset S Rn is
a C k hypersurface if and only if it can be locally defined as the zero-set of a single function with
nontrivial gradient.
We now extend the notion of submanifolds in Rn by defining submanifolds of abstract manifolds
following the line of the property (ii) in Theorem 1.41 and using general charts as defined in
Definition 1.30.
Definition 1.42. Let M be an n-manifold of class C k . A subset S M is called an mdimensional submanifold of M (of the same class C k ) if, for every p S, there exists a chart
(U, ) on M with p U and S U = 1 (Rm ).
Remark 1.43. More generally, a C r submanifold of M can be defined for r k by viewing
M as a C r manifold.
It is straightforward to see that, for (U, ) ranging over all charts satisfying the conditions in
Definition 1.42, the pairs (S U, |SU ) form a C k atlas on S definining on S a structure of a C k
manifold.
Exercise 1.44. Show that the manifold topology on a submanifold S M coincides with
that induced by the manifold topology on M .
Exercise 1.45. Show that being submanifold is a transitive relation: If M1 is a submanifold
of M2 and M2 of M3 , then M1 is a submanifold of M3 .

16

1. INTRODUCTION TO SMOOTH MANIFOLDS

6. Differentiable maps, immersions, submersions and embeddings

0

Let Rn and 0 Rn be open sets. Recall that a map f : 0 is of class C k if all

partial derivatives of f up to order k exist and are continuous on .
Definition 1.46. Let M be an n-manifold with an atlas (U , )A of class C r and M 0 be
0
an n0 -manifold with an atlas (U0 , 0 )B of class C r . A continuous map f : M M 0 is called
differentiable of class C k with k min(r, r 0 ) if, for every x M and every A and B with
k
x U and f (x) U0 , the composition 0 f 1
is of class C in an open neighborhood of
(x).
Remark 1.47. The continuity of f is assumed with respect to the canonical topologies on M
and M 0 as defined in previous section. The continuity of f guarantees that f (y) U0 for y in a
neighborhood of x and hence the composition 0 f 1
is defined in a neighborhood of (x)
1
0
(by Lemma 1.36, both and are continuous on their domains of definition).
Remark 1.48. Given any x M , there always exist and as in Definition 1.46 because
M = A U and M 0 = B U0 .
Remark 1.49. The reason to choose k min(r, r 0 ) is to guarantee that the property of f to
0
be C k does not change when composing with C r maps on the right and C r maps on the left. This
choice is also illustrated by the following statements.
Example 1.50. If S M is a submanifold of a manifold M , the inclusion : S M is
automatically differentiable (why?)
Definition 1.51. A C k diffeomorphism between two manifolds M and M 0 of the same dimension is a bijection f : M M 0 such that both f and f 1 are of class C k . Two manifolds are said
to be C k diffeomorphic if there exists a C k diffeomorphism between them.
Being C k diffeomorphic is the basic equivalence relation in Differential Topology.
We mention without proof the following fundamental result (see M. Hirsch, Differential Topology, Chapter 2, Theorem 2.10 for a proof):
Theorem 1.52. For r 1, every C r manifold is C r diffeomorphic to a C manifold. If for
1 r < s , two C s manifolds are C r diffeomorphic, they are C s diffeomorphic.

Definition 1.53. In the notation of Definition 1.46, a C 1 map f is said to be an immersion

at a point x M if 0 f 1
is an immersion at (x), i.e. if the differential
n
n
d (x) (0 f 1
): R R

(6.1)

is injective. Similarly f is said to be a submersion at x M if the differential in (6.1) is surjective.

Without point specified, f is an immersion or submersion if it so at every point. Finally, f is called
an embedding if it is an injective immersion and is a homeomorphism onto its image f (M ).
An important property of being immersion or submersion at a point is that the latter properties
automatically hold in a neighborhood of that point.

6. DIFFERENTIABLE MAPS, IMMERSIONS, SUBMERSIONS AND EMBEDDINGS

17

Proposition 1.54. If S M is a submanifold, the inclusion : S M is an embedding.

Vice versa, if S and M are two manifolds and : S M is an embedding, then the image (S) is
a submanifold of M .
Exercise 1.55. Prove this proposition.
An important relation between immersions and embedding is the following statement which is
a consequence of the Inverse Function Theorem:
Proposition 1.56. If f : M M 0 is an immersion at x M , there exists an open neighborhood U of x in M such that the restriction f |U is an embedding of U into M 0 .
Example 1.57. Both regular arc and regular surface elements are examples of immersions.
Exercise 1.58. Prove that an injective immersion of a compact Hausdorff manifold is an
embedding. Can compact be dropped?
Exercise 1.59. Given two C k manifolds M1 and M2 , prove that the canonical projections
i : M1 M2 Mi , i = 1, 2, are C k submersions.
Exercise 1.60. For each k = 0, 1, . . ., define k : C C and k : C C (C := C \ {0}) by
k (z) = k (z) := z k . For which k are k and k immersions, submersions or embeddings?

CHAPTER 2

Basic results from Differential Topology

A topologist is one who doesnt know the difference between a doughnut and a
coffee cup.
John Kelley, In N. Rose, Mathematical Maxims and Minims
1. Manifolds with countable bases
From now one a manifold M will be assumed to be Hausdorff and to have a countable basis:
Definition 2.1. A topological space T is said to have countable basis (or base) or to be
separable if there exists a countable family (Ui ) of open sets which form a basis, that is, one (and
hence the other) of the following equivalent conditions holds:
(1) For any x M and any neighborhood V of x, there is with x Ui V ;
(2) Any open set in T is a union of some of the Ui .
Exercise 2.2. Show that (1) and (2) are indeed equivalent.
Example 2.3. Rn has a countable basis. The basis can be taken among open balls with rational
radius and center.
Exercise 2.4. Show that any submanifold of Rn has also countable basis.
On a manifold a basis can be chosen with additional properties:
Lemma 2.5. A manifold M has a countable basis (Ui ) whose elements are relatively compact
(i.e. their closures U i in M are compact).
Proof. Given a countable basis (V ), the subfamily consisting of relatively compact V s is
again a basis satisfying the conclusion.

The following simple but important statement asserts the existence of countable compact
exhaustions:
Lemma 2.6. A manifold M can be exhausted by countably many compact subsets in the sense
that there exists a sequence (Km ) of compact sets in M with Km IntKm+1 for every n and
m Km = M .
Proof. Given a countable relatively compact basis (Vm )mN as in Lemma 2.5, we construct
the sequence (Km ) inductively as follows. Set K1 := V 1 . Assuming Km is constructed, since it is
compact, it is covered by finitely many Vj s. We can choose j > m such that
Km V 1 V j
19

20

2. BASIC RESULTS FROM DIFFERENTIAL TOPOLOGY

and set Km+1 := V 1 V j . The sequence Km , inductively constructed, satisfies the desired
properties.

2. Partition of unity
Partition of unity is an important tool for glueing together functions (and other objects) defined
locally.
Definition 2.7. A C k partition of unity on a C k manifold M is a family (hi )iI of C k functions
hi : M R satisfying the following:
(1) 0 hi 1 for all i I;
(2) every point p M has a neighborhood intersecting only finitely many sets
P

supp hi := x M : h(x) 6= 0;

(2.1)

(3) iI hi 1 (locally this is always a finite sum in view of (2)).

In practice it is useful to have partitions of unity subordinate to a given covering of M in the
sense that each supp hi is contained in (at least one) single element of the covering:
Definition 2.8. A partition of unity (hi )iI is said to be subordinate to an open covering
(W ) of M if for every i I there is an with supp hi W .
Construction of a partition of unity is based on the existence of the following cut off functions:
Lemma 2.9. For every n, there exists a C function h : Rn R with h 0 such that h(x) = 1
for kxk 1 and h(x) = 0 for kxk 2.
Proof. Let
f (r) :=
Then f : R R is C and the function
g(r) :=

e1/r
0

r>0
r 0.

f (2 r)
f (2 r) + f (r 1)

is C with values in [0, 1] such that g(r) = 1 for r 1 and g(r) = 0 for r 2. It remains to set
h(x) := g(kxk).

We can now prove the main existence theorem for a partition of unity.
Theorem 2.10. Let (W ) be an open covering of a C k manifold M . Then there exists a C k
partition of unity subordinate to (W ).
Proof. Let (Km )mN be a compact exhaustion of M as in Lemma 2.6. For each fixed m,
consider the open sets
(IntKm+2 \ Km1 ) W .

2. PARTITION OF UNITY

21

These sets (for fixed m and varying ) cover the compact set Cm := Km+1 \ IntKm . For every
x Cm , choose a chart (Ux , x ) with
x Ux (IntKm+2 \ Km1 ) W

for some and such that x (x) = 0 and x (Ux ) contains the ball B3 (0) Rn with center 0
and radius 3. These properties are easy to obtain starting from any x and composing it with a
suitable affine transformation of Rn . Since Cm is compact, it is covered by finitely many open sets
of the form 1
x (B1 (0)). By collecting these sets for all m, we obtain a family of charts (U j , j )
satisfying the following properties:
(1) for every Uj there exists with Uj W ;
(2) (Uj ) is locally finite, i.e. for every x M , there exists a neighborhood of x that meets
only finitely many Uj s;
(3) j (Vj ) B3 (0) for all j;
(4) j 1
j (B1 (0)) = M .
With h as in Lemma 2.9, we now define functions fj : M R by
(
h(j (y)) y Uj
fj (y) =
0
y
/ Uj .
It follows from property (3) above and the properties of h that fj is a C k function on M . Furthermore, fj = 1 on 1
j (B1 (0)). Then it remains to set
X
g :=
gi , hi := gi /g
i

to obtain a partition of unity (hi ) as desired. Indeed, (2) implies that the sum is locally finite
and yields a well-defined C k function. Furthermore, (4) implies that g > 0 everywhere on M and
hence hi is well-defined. Since supp hj Uj , (hj ) is subordinate to (W ) in view of (1).


A partition of unity is used to glue together locally defined functions. For instance, we obtain
the following simple application of Theorem 2.10:
Lemma 2.11. Let M be a C k manifolds with p, q M , p 6= q. Then there exists a C k function
f : M R with 0 f 1, {p} = f 1 (0), {q} = f 1 (1).

Proof. Let (Up , p ) and (Uq , q ) be disjoint charts at p and q respectively with p (p) = 0,
p (q) = 0 and kp (x)k < 1, kq (y)k < 1 for x Up , y Uq . Consider the covering of M by Up , Uq
and U0 := M \ {p, q}. For each set in this covering we can solve the problem by taking fp := kp k2
on Up , fq := 1 kq k2 on Uq and f0 := 1/2 on U0 respectively. By Theorem 2.10, there exists a
partition of unity (hi ) subordinate to the covering {Up , Uq , U0 }. By summing hi s together we may
assume that i = p, q, 0 and supp hi Ui . Then
f := hp fp + hq fq + h0 f0

is a C k function on M satisfying the conclusion of the lemma.

22

2. BASIC RESULTS FROM DIFFERENTIAL TOPOLOGY

3. Regular and critical points and Sards theorem

We shall assume all manifolds to be Hausdorff and to have a countable topology basis.
In view of Property (iv) in Theorem 1.41, points where maps are submersive are of particular
interest.
Definition 2.12. Let M and M 0 be C k manifolds (k 1) and f : M M 0 a C k map. A
point x M is said to be regular for f if f is a submersion at x, and critical otherwise. A point
y M 0 is a regular value for f is every point x f 1 (y) is regular, and a critical value otherwise
(even if y
/ f (M )).

In case dim M 0 > dim M it is easy to see that all points of M are critical and all points of
f (M ) are critical values.
The following regular value theorem is often used to construct manifolds:

Theorem 2.13. Let f : M M 0 be a C k map between two C k manifolds. If y M 0 is a

regular value, then f 1 (y) is a C k submanifold of M .
The proof follows from Theorem 1.41.
Exercise 2.14. Give an example of a C map f : R R with infinitely many critical values.
A profound result known as Sards theorem says that the set of critical values cannot be too
large. To measure the largeness we have to introduce sets of measure zero.
Definition 2.15. An n-cube C Rn of edge > 0 is any set C obtained from [0, ]n by
euclidean motion. The measure (volume) of C is (C) = n . A subset S Rn is of measure zero
if for every > 0, it can be covered by a countable union of n-cubes, the sum of whose measures
is less than . A subset S in a manifold is said to be of measure zero if for every chart (U, ) on
M the set (S U ) Rn is of measure zero.
Exercise 2.16. Show that a countable set is of measure zero.
Note that we have not defined the measure of a subset of M but only a certain kind of
subsets that are of measure zero. A countable union of sets of measure zero is again a set of
measure zero (why?). On the other hand, it follows from elementary Measure Theory that a cube
is not of measure zero. The latter fact implies the following important property used in many
applications:
Lemma 2.17. The complement of a set of measure zero is dense.
The property to be of measure zero is a rather flexible notion invariant under diffeomorphisms
and even under general C 1 maps:
Lemma 2.18. Let U Rn be an open set and f : U Rn a C 1 map. If X U is of measure
zero, so is f (X).
Proof. The set U can be covered by countably many balls B U such that the norm kdx f k
is uniformly bounded on B by a constant C = C(B). Then
|f (x) f (y)| C|x y|

3. REGULAR AND CRITICAL POINTS AND SARDS THEOREM

23

for all x, y B. It follows that

if C B is an0 n-cube
of nedge , then f (C) is contained in an
0
n-cube C of edge less than nC. Hence (C ) ( nC) (C). The latter bound implies that
f (X B) is of measure zero. Then f (X) is also of measure zero as a countable union of sets of
measure zero.

We now state the following important theorem:
Theorem 2.19 (Sards Theorem). The set of critical values of a C k map f : M M 0 between
C manifolds (k 1) is of measure zero provided
k

k > max(0, dim M dim M 0 ).

(3.1)

In particular, the set of critical values of a C map between C manifolds is always of measure
zero.
With Lemma 2.17 we obtain the following important consequence:
Corollary 2.20. Under the assumptions of Theorem 2.19 the set of regular values of f is
dense in M 0 .
Example 2.21. A constant map has all points as critical but only one critical value.
In case of a map f : M M 0 with dim M 0 > dim M the set of critical points is exactly f (M )
and hence Sards Theorem asserts that f (M ) is of measure zero. This assertion is relatively easy
to prove.
Exercise 2.22. Give a proof of this assertion. Hint. Use Lemma 2.18.
We prove Sards theorem here only in the equidimensional case dim M = dim M 0 . The proof
for dim M > dim M 0 is more difficult and requires consideration of higher order partial derivatives.
Proof of Sards theorem for dim M = dim M 0 . Since M and M 0 have countable bases
consisting of charts, the general case is reduced to that where M and M 0 are relatively compact
open sets in Rn and f is C 1 map from a neighborhood of K := M Rn into Rn . We can further
assume that K is contained in the cube [0, 1]n .
If a point a M is critical, the image da f (Rn ) is a lower-dimensional plane which is of
measure zero in Rn . The idea of the proof is to compare the actual map f with its first order
Taylor expansion f (a) + da f (x a) by estimating the difference f (x) f (a) da f (x a).
Since the first order partial derivatives of f are continuous on the compact set K, we have
kda f k C for some C > 0 and all a K. Fix > 0. Since the first order partial derivatives of
f are also uniformly continuous on K, there exists > 0 such that kda f db f k whenever
ka bk for a, b K. We now estimate the difference mentioned above for kx ak :

Z 1


d

kf (x) f (a) da f (x a)k =
f
(a
+
t(x

a))dt

d
f
(x

a)
a


dt
0
Z 1




=
(da+t(xa) f da f )(x a)dt sup kda+t(xa) f da f k kx ak . (3.2)
0

0t1

24

2. BASIC RESULTS FROM DIFFERENTIAL TOPOLOGY

Let now C be an n-cube of edge / n containing a critical point a. Then the affine map ha (x) :=
f (a) + da f (x a) sends C into an (n 1)-cube of edge 2C inside a hyperplane H Rn (with
respect to some euclidean coordinates in H), where C > 0 is the bound for kdf k chosen above.
Now (3.2) implies that f (C ) is contained in a parallelepiped of measure 2C n which can be
covered by finitely many cubes whose sum of measures does not exceed 4C n .
Without loss of generality, = 1/m for some integer m. Then K [0, 1]n can be covered by
mn disjoint n-cubes of edge . Some of these cubes contain critical points. Above we have seen
that the image of each of those cubes is covered by cubes of total measure not greater than 4C n .
Summing over all cubes we see that the total set of critical values can be covered by cubes whose
total measure does not exceed 4C. It remains to observe that can be arbitrarily small.

As an application of Sards theorem we show that there is no retraction from the closed unit
ball in Rn onto the unit sphere. Recall that a retraction from a topological space X onto a subspace
Y is any continuous map r : X Y such that the restriction of r to X is the identity.
Theorem 2.23. There is no retraction from the ball
onto the sphere S n1 := {kxk = 1}.

B n := {kxk 1} Rn

Proof. Suppose there is a retraction r : B n S n1 . We first reduce the general case to the
case when r extends to a C map in a neighborhood of B n . For this, we find a new retraction
g : B n S n1 which is C in a neighborhood of S n1 by extending r continuously to a closed
ball of radius 1 + as constant in radial directions and scaling the ball with coefficient (1 + ) 1 .
Then we approximate g by a C map into S n1 which agrees with g on a neighborhood of S n1 .
Now, by Sards theorem, g defined in a neighborhood of B n , has a regular value y S n1 and
therefore the preimage V := g 1 (y) is an one-dimensional closed submanifold in a neighborhood
of B n . We have y V and the connected component of y in V (which is homeomorphic either to
a circle or to an interval) intersects the interior of B n and hence must contain another point of
S n1 . However the retraction condition implies S n1 V = {y} which is a contradiction.

Theorem 2.23 implies Brouwers fixed-point theorem:
Theorem 2.24 (Brouwers fixed-point theorem). Any continuous map f : B n B n has a
fixed point.
Indeed, if there is a continuous map as in Theorem 2.24 without fixed points, then one can
obtain a retraction r as in Theorem 2.23 by letting r(x) to be the intersection of S n1 with the
ray passing starting from f (x) and passing through x.

4. WHITNEY EMBEDDING THEOREM

25

4. Whitney embedding theorem

Theorem 2.25 (Whitney embedding theorem). Any n-manifold M of class C k (k 1) can
be C k embedded into R2n+1 .
We prove Theorem 2.25 in the easy case when M is compact and of class at least C 2 . The
proof is split into two steps. First we construct an embedding into Rm for some possibly large m.
Then we project the embedded copy of M to a smaller linear subspace of dimension (2n + 1) in
such a way that the projection is still an embedding.
Proposition 2.26. Let M be a compact manifold. Then there exists an integer m and an
embedding of M into Rm .
Proof. We use the proof of Theorem 2.10 on the existence of a partition of unity. There it
was constructed a covering of M by coordinate charts (Ui , i ) and smooth functions fi : M R,
0 fi 1, with suppfi Ui such that the interiors of the preimages fi1 (1) Ui for all i still
cover M . Since M is compact we may assume that the covering is finite, i.e. i = 1, . . . , l. The the
map
g(x) := (f1 (x)1 (x), . . . , fl (x)l (x), f1 , . . . , fl ) Rn Rn R R = Rnl+l ,

where each component fi (x)i (x) is extended by 0 for x

/ Ui , defines an embedding of M into
Rnl+l .

In view of Proposition 2.26 we may assume that M is a submanifold of Rm .
Proposition 2.27. Let M Rm be a compact submanifold of dimension n of class at least
C 2 such that m 2n + 1. Then, after a suitable linear change of coordinates in R m , the standard
projection from M to the subspace R2n+1 Rm is an embedding.

Proof. We may assume that m > 2n + 1. For every unit vector v S m1 Rm , denote by
fv the orthogonal projection along v from Rm onto the orthogonal complement of v (isomorphic
to Rm1 ). We want to choose v such that fv |M is an embedding.
The injectivity of fv |M means that v is not parallel to any secant of M , i.e.
xy
v 6=
, x 6= y M,
(4.1)
kx yk

or, equivalently, v is not in the image of the map c : (M M )\ S m1 , (x, y) 7 (xy)/kxyk,

where := {(x, x) : x M } is the diagonal. Since dim(M M ) = 2n < m 1, there is a dense
set of points v not in the image of c (e.g. by the easy case of Sards theorem).
The property of fv |M to be an immersion means that v is not tangent to M at any point.
Again, since v is a unit vector, the latter property means that v is not in the in the image of the
map
: S S m1 , S := {(x, v) M Rm : v Tx M, kvk = 1},
(4.2)
where Tx M denotes the space of all vectors tangent to M at x. Then S is a (2n 1)-dimensional
submanifold of M Rm . Since 2n 1 < m 1, it follows from the easy case of Sard that the
complement of the image (S) in S m1 is dense. Moreover, since S is compact, this complement

26

2. BASIC RESULTS FROM DIFFERENTIAL TOPOLOGY

is open. Hence we can choose v in both complements of and c guaranteeing that fv |M is an

injective immersion. Since M is compact, fv |M is in fact an embedding as desired.

Theorem 2.25 in case M is compact and k 2 follows now from Propositions 2.26 and 2.27.

CHAPTER 3

Tangent spaces and tensor calculus

1. Tangent spaces
We start by defining tangent vectors to submanifolds in Rn .
Definition 3.1. A vector X Rn is said to be tangent to a submanifold S Rn at a point
p S if there exists > 0 and a C 1 arc c : (, ) S with c(0) = p and c0 (0) = X.
Proposition 3.2. The set Tp S of all tangent vectors to an m-dimensional submanifold S Rn
at a fixed point p S is a vector subspace of Rn . Given a local parametrization fp : p Rm S
of S, one has Tp S = da f (Rm ) with a := fp1 (p).
Proof. Given fp : p Rn as above, C 1 arcs c : (, ) S with c(0) = p are in oneto-one correspondence with C 1 maps e
c : (, ) p with e
c(0) = a (as a consequence of the
Inverse Function Theorem). The correspondence is defined by composing with fp : c = fp e
c. Then
0
0
c (0) = da fp (e
c (0)) completes the proof since da fp is a linear map.


The various characterizations of submanifolds of Rn (Theorem 1.41) lead to equivalent characterizations of tangent spaces:
Theorem 3.3. Let S Rn be an m-dimensional C k submanifold and let gi : V R, i =
1, . . . , n m, be local defining functions of S in a neighborhood V of p with linear independent
gradients as in Theorem 1.41 (iii). Then X is tangent to S at p if and only if dp gi (X) = 0 for all
i.
We now give three equivalent definitions of a tangent vector to an abstract C k manifold M
(1 k ) at a point p M .
Definition 3.4 (Geometric Definition of tangent vectors as classes of equivalent
curves). A tangent vector at p is an equivalence class of C k arcs c : (, ) R M , > 0,
with c(0) = p where two arcs
c1 : (1 , 1 ) M and c2 : (2 , 2 ) M

are called equivalent if ( c1 )0 (0) = ( c2 )0 (0) for some (and hence for any) chart (U, ) on M
with U 3 p. The tangent space Tp M of M at p is the set of all tangent vectors at p.
Briefly: Tangent vectors are tangents to arcs lying on the manifold. Unfortunately there is
no privileged representative arc for a given tangent vector.
Example 3.5. If U Rn is an open set, a curve c : (, ) R U with c(0) = p is
equivalent to the affine curve 7 p + c0 (0). Then c 7 c0 (0) defines a one-to-one correspondence
27

28

3. TANGENT SPACES AND TENSOR CALCULUS

between the equivalence classes of curves and the space Rn . Hence the tangent space Tp U can be
identified with Rn .
Given a tangent vector X Tp M and a C k function f : U R defined in an open neighborhood
U of p in M , we can define the directional derivative of f in the direction of X by
DX f = Xf := (f c)0 (0),

(1.1)

where c is any curve representing X. It is easy to see that the right-hand side of (1.1) does not
depend on the particular representative c of X. Furthermore, DX f does not change if f is replaced
by its restriction to a smaller neighborhood of p. In other words, DX f = DX g if f and g coincide
in a neighborhood of p. This motivates the following
Definition 3.6. A germ at p of a C k function on M is an equivalence class of C k functions
defined in open neighborhoods of p such that two such functions f1 : U1 R and f2 : U2 R with
p U1 U2 are equivalent if their restrictions f1 |U and f2 |U coincide for some smaller neighborhood
U of p, U U1 U2 .
Denote by Fp (M ) the set of all germs of C k functions at p. Unlike the problem to add or
multiply functions with different domains of definitions, there is no problem to add or multiply
germs of functions. In particular, Fp (M ) has a natural structure of an R-algebra. We now have
the following equivalent definition of tangent vectors in the C case.
Definition 3.7 (Algebraic Definition of tangent vectors as derivations). A tangent
vector X on a C manifold M at a point p M is a derivation on the set Fp (M ), i.e a map
X : Fp (M ) R with following properties:
(1) linearity: X(af + bg) = aXf + bXg for a, b R, f, g Fp ;
(2) Leibnitz rule: X(f g) = (Xf )g + f (Xg) for f, g Fp .
Briefly: tangent vectors are derivations acting on germs of scalar functions.
In view of formula (1.1), every tangent vector X in the sense of Definition 3.4 defines a
tangent vector X = DX in the sense of Definition 3.7. Vice versa, we show that any derivation as
in Definition 3.7 arises in this way in the C case. The proof is based on the following lemma:
Lemma 3.8. Let B n Rn be the open unit ball and f : B n R a C function. Then there
exist C functions f1 , . . . , fn : B n R such that
f (x) = f (0) + x1 f1 (x) + + xn fn (x).
Proof. Since
f (x) f (0) =
it remains to set fi (x) :=

R1

1
0

X
d
f (tx1 , . . . , txn ) dt =
xi
dt
i=1

df
(tx1 , . . . , txn ) dt,
0 dxi

i = 1, . . . , n.

1
0

(1.2)

df
(tx1 , . . . , txn ) dt,
dxi


Note that if f is C k , fi is only C k1 , hence we may not find functions in the same class unless
k = .

1. TANGENT SPACES

29

Given a derivation X in the sense of Definition 3.7 and a local system of coordinates (or chart)
(x , . . . , xn ) on M at p, vanishing at p, we obtain a tangent vector in the sense of Defintion 3.4 as
follows. Denote by i R the value of X on the germ defined by the coordinate function xi . Then
Xf = (f c)0 (0) with c(t) := ( 1 , . . . , n )t. Indeed, X vanished on constant functions (as follows
from the Leibnitz rule) and hence, in view of formula (1.2),
X
X df
Xf =
i fi (0) =
i i (0) = (f c)0 (0).
dx
i
i
1

A basis of the tangent space to M at p is given in local coordinates (x1 , . . . , xn ) by

(p), . . . , n (p),
1
x
x

where /xi is the derivation given by f 7 f /xi . Other notation is also used frequently:




or

xi p
xi p

Finally we have the following analytic (also called physical) definition:

Definition 3.9 (Analytic Definition of tangent vectors via coordinates and their
transformation rules). A tangent vector at the point p is an association to every coordinate
chart (x1 , . . . , xn ) at p an n-tuple ( 1 , . . . , n ) of real numbers is such a way that, if (e1 , . . . , en ) is
associated with another coordinate system (e
x1 , . . . , x
en ), then it satisfies the transformation rule

e
xi
(1.3)
ei = j (p) j ,
x
where there is a summation over j (Einsteins convention).
Briefly: tangent vectors are represented by elements of Rn for each coordinate chart transforming via differentials of a coordinate change at the reference point.
Given a tangent vector represented by a curve c as in Definition 3.4, it is easy to define an
association in the sense of Definition 3.9 by taking ( 1 , . . . , n ) to be the derivative of c at t = 0
computed with respect to the given coordinate chart. Vice versa, given ( 1 , . . . , n ), a representing
curve c can be chosen t 7 ( 1 , . . . , n )t. A direct equivalence with Definition 3.7 is also easy to
obtain by letting 1 , . . . , n to be the values of a given derivation on the (germs of) coordinate
functions x1 , . . . , xn respectively. In other words, we have
= i

(p).
xi

(1.4)

Definition 3.10. Let f : M M 0 be a C k map. The differential of f or tangent map to f

at a point p M is the map dp f : Tx M Tf (x) M 0 (very often denoted by dfp or (df )p ) that
sends an equivalence class defined by a curve c : (, ) M to the equivalence class defined by
(f c) : (, ) M 0 .

30

3. TANGENT SPACES AND TENSOR CALCULUS

If (x1 , . . . , xn ) and (y 1 , . . . , y m ) are local coordinates on M and M 0 at p and p0 := f (p) respectively, we write f = (f 1 , . . . , f m ) and have the formula



f i

dp f
(p)
=
(p) i
(1.5)
j
j
x
x
y

2. VECTOR FIELDS AND LIE BRACKETS

31

2. Vector fields and Lie brackets

Definition 3.11. The tangent bundle T M of an n-manifold M is the disjoint union of all
tangent spaces Tx M for x M with the natural structure of a 2n-manifold. A vector field X on
a manifold M is a correspondence that associates to each point p M a vector X(p) T p M , i.e.
a mapping X : M T M . The field is differentiable if this mapping is differentiable.
Denote by C (M ) the algebra of C functions on M . Then, in view of the algebraic definition
of a tangent space in the previous section, a vector field X can be identified with a derivation of
C (M ) (or an operator acting there), i.e. a map
f C (M ) 7 Xf C (M )

satisfying both linearity and the Leibnitz rule as in Defintion 3.7.

It makes sense to iterate the above operators. If X and Y are two vector fields, X Y and Y X
are (in general different) operators involving higher order derivatives. However the difference of
them turns out to be a new vector field called their Lie bracket [X, Y ]:
X, Y ]f = X(Y f ) Y (Xf ).

(2.1)

Lemma 3.12. There exists a unique vector field Z such that, for all f C (M ), Zf =
X(Y f ) Y (Xf ).
P
P
Proof. Write X = i ai x i and Y = j bj x j , where ai and bi are smooth functions on M .
Then for all f ,
X f
X f
X bj
aj f
X(Y f ) Y (Xf ) = X(
bj j ) Y (
ai i ) =
(ai i bi i ) j ,
x
x
x
x x
j
i
ij

wherePwe have cancellation for terms involving second order derivatives of f . Then the formula
bj
i aj
Z = ij (ai x

i b xi ) shows both existence and uniqueness.
The basic properties of the Lie brackets are bilinearity, anticommutativity and the Jacobi
identity:
[[X, Y ], Z] + [[Y, Z], X] + [[Z, X], Y ] = 0.
Definition 3.13. A trajectory or integral curve of a vector field X on a manifold M is a curve
(arc) from an open interval (a, b) R into M safisfying the equation
d
(t0 ) = X((t0 ))
dt

for every t0 (a, b).

It follows from the local theory of Ordinary Differential Equations that a vector field (at least
of class C 1 ) locally has a unique integral curve p defined over an interval (a, b) containing 0 and
passing through any given point p M at t = 0. The uniqueness means that any two such curves
coincide on the intersection of their intervals of definition. There is a unique choice of the maximal

32

3. TANGENT SPACES AND TENSOR CALCULUS

possible intervale. Recall that the manifold M is assumed Hausdorff which is important for the
mentioned uniqueness.
It is a further consequence of the local theory of Ordinary Differential Equations that the
integral curve p also smoothly depends on p, i.e. the map (t, p) := p (t) is a smooth function
in (t, p) R M . The smoothness of is the same as that of the vector field. If for each p, the
interval for p is chosen to be the maximal one, the domain of definition of becomes an open
subset of R M . The map is called the local flow of the vector field X.
Definition 3.14. A vector field on M is called complete if its flow is defined on the whole
R M or, equivalently, for every p M , the integral curve p is defined on R.
Proposition 3.15. If M is compact, every vector field is complete.
If X is a complete vector field, by fixing t R we obtain a self-diffeomorphism (t, ) : M M
because its inverse is (t, ). In general we have
(t1 , (t2 , p)) = (t1 + t2 , p))

justifying the name one-parameter group of diffeomorphisms.

In general, the local flow (t, x) of a vector field X can be used to calculate Xf for a smooth
function f (exercise):
d
f ((t, )).
dt
Let (s, y) be the (local) flow of another vector field Y . Iterating (2.2) we obtain
Xf () =

(2.2)



d2 
f ((t, (s, ))) f ((s, (t, ))) ,
(2.3)

dt ds t=s=0
from where we see that, if the flows of X and Y commute (i.e. (t, (s, )) (s, (t, ))), then
[X, Y ] 0. The converse is also true (here without proof):
Theorem 3.16. Given vector fields X and Y on a manifold M , their flows commute if and
only if [X, Y ] 0.
[X, Y ]f () = X(Y f )() Y (Xf )() =

3. FROBENIUS THEOREM

33

3. Frobenius Theorem
The problem of finding integral curves for single vector fields naturally extends to the more
general problem of finding integral surfaces for systems of vector fields.
Definition 3.17. Given a system of vector fields X1 , . . . , Xd , on M , linearly independent at
every point, an integral submanifold S is a submanifold of M such that for every p S, the tangent
subspace Tp S Tp M is spanned by X1 (p), . . . , Xd (p).
In contrast to the integral curves of single vector fields, an integral submanifold may not exist
in general. A necessary condition comes from the following property:
Lemma 3.18. If S M is a submanifold and two vector fields X and Y on M are tangent to
S, then their Lie bracket is also tangent to S.
Here X is called tangent to S if X(p) Tp S for all p S.

Proof. By definition, Xf (p) = dtd |t=0 f (c(t)), where c : (, ) M is a curve representing

X(p) Tp M . If p S and X is tangent to S, the curve c can be chosen in S. Then
(Xf )|S = (X|S )(f |S ),

(3.1)

i.e. differentiating f along X and restricting to S yields the same result as differentiating the
restriction of f along the restriction of X. Then the Lie bracket (2.1) can be computed separately
for X, Y on M and for their restrictions to S. It follows from (3.1) that both brackets are equal
when restricted to S and hence are tangent to S.

Hence, if S is an integral submanifold of the system X1 , . . . , Xd , and p S, all Lie brackets
[Xi , Xj ](p) must be contained in the span of X1 (p), . . . , Xd (p). This condition turns out to be also
sufficient if assumed at every point:
Theorem 3.19 (Frobenius). Let X1 , . . . , Xd , be vector fields on a manifold M , linearly independent at every point. A necessary and sufficient condition for the existence, for every p M ,
of an integral submanifold S M passing through p is that [Xi , Xj ](p) belongs to the span of
X1 (p), . . . , Xd (p) for every p M .

34

3. TANGENT SPACES AND TENSOR CALCULUS

4. Lie groups and Lie algebras

Lie groups are important example of differentiable manifolds.
Definition 3.20. A Lie group is a set G which is both a C k manifold and a group such that
the maps
G G G, (x, y) 7 xy, G G, x 7 x1
are of class C k .
Exercise 3.21. Show that for every g G, the left and right translations
Lg , Rg : G G,

Lg (h) := gh,

Rg (h) := hg

are C k diffeomorphisms.
Remark 3.22. We mention without proof that a Lie group of class C 1 has an unique C
(even real-analytic) differentiable structure compatible with the group operation in the sense of
Definition 3.20. In particular, we may always assume a Lie group to be C .
Examples 3.23. (1) Any finite-dimensional real vector space is a Lie group with respect to
addition. (2) For K = R (real numbers), C (complex numbers) or H (quaternions), set K :=
K \ {0}. Then K is a Lie group with respect to multiplication.
The most important Lie groups, called classical Lie groups, are GL(n, R), GL(n, C), O(n),
U (n) and their corresponding subgroups SL(n, R), SL(n, C), SO(n), SU (n), defined as follows.
We write Rnn (resp. Cnn ) for the sets of all n n matrices with real (resp. complex) entries.
For a matrix A Cnn , denote by At the transpose of A and by A its conjugate. Then
GL(n, R) := {A Rnn : det A 6= 0},

GL(n, C) := {A Cnn : det A 6= 0},

O(n) := {A Rnn : AAt = id},

U (n) := {A Cnn : AAt = id},

SL(n, R) := {A Rnn : det A = 1},

SL(n, C) := {A Cnn : det A = 1},

SO(n) := O(n) SL(n, R),

SU (n) := U (n) SL(n, C).

The groups GL(n, R) and GL(n, C) are open (and dense) subsets of the matrix spaces Rnn
and Cnn respectively. The other groups are submanifolds. To see this it is sufficient to check the
submanifold property near the identity matrix id and use left (or right) multiplications to check
it near other points.
For every Lie group G there exists unique associated finite-dimensional Lie algebra g that as a
vector space can be identified with the tangent space Te G, where e G is the unit of the group.
In order to define an algebra operation (called the commutator) on g, identify a vector T e G

4. LIE GROUPS AND LIE ALGEBRAS

35

with the unique left-invariant vector field X on G with X(e) = . A vector field X on G is called
left-invariant if
da Lg (X(a)) = X(ga).
for all a, g G, where Lg (x) := gx is the left translation on G. The the usual Lie bracket induces
a Lie algebra structure on the (finite-dimensional) space of all left-invariant vector fields. In the
light of the above identification of g with left-invariant vector fields, this construction yields a Lie
algebra structure on g which is the Lie algebra structure induced by the group operation of G.
(Right-invariant vector fields would yield an isomorphic Lie algebra).
The Lie algebras corresponding to the above classical Lie groups are the classical matrix Lie
algebras gl(n, R), gl(n, C), o(n), u(n) and their corresponding subalgebras sl(n, R), sl(n, C), so(n),
su(n):
gl(n, R) := Rnn ,
gl(n, C) := Cnn ,
o(n) := {a Rnn : a + at = 0},

u(n) := {a Cnn : a + a
t = 0},

sl(n, R) := {a Rnn : tr a = 0},

sl(n, C) := {a Cnn : tr a = 0},

so(n) := o(n) sl(n, R),

su(n) := u(n) sl(n, C),

where tr stands for the trace of the matrix, i.e. the sum of all diagonal entries.
For every matrix group, its Lie algebra consists of matrices tangent to the group at the identity
matrix and their commutator is given by the formula [a, b] = ab ba.

36

3. TANGENT SPACES AND TENSOR CALCULUS

5. Tensors and differential forms

A tensor of type (m, n) on a manifold M is a correspondence that associates to each point
p M a multilinear map
p : Tp M Tp M Tp M Tp M R,

where the tangent space Tp M appears m times and the cotangent space Tp M (the dual of the
tangent) appears n times. For instance, a vector field X can be seen as a collection of linear maps
Xp : Tp M R.
An (exterior) m-form is a tensor of type (m, 0) which is skew symmetric in its arguments,
i.e.
p ((1) , . . . , (m) ) = (1) p (1 , . . . , m )
for any tangent vectors 1 , . . . , m Tp M and any permutation Sm . In any local coordinates
(x1 , . . . , xm ) it can be expressed as
= i1 ,...,im dxi1 dxim = I dxI ,

where the summation is taken over all sets of indices 1 i1 < < im n and I = (i1 , . . . , im ),
I = i1 ,...,im , dxI = dxi1 dxim . Here I is a function of the reference point and is said to
be smooth (or differentiable) if each I is smooth. The latter notion is independent of the choice
of coordinates.
The wedge product of an m-form and an m0 -form 0 is an (m + m0 )-form which is defined in
coordinates by
0
0
(I dxI ) (I0 0 dxI ) = I I0 0 dxI dxI ,
i.e. the coefficients are multiplied as functions and the wedge products of differentials are written
together with a wedge inbetween. A coordinate-free definition can be also given:
( 0 )p (1 , . . . , m+m0 ) =

Sm+m0

(1)
p ((1) , . . . , (m) ) p0 ((m+1) , . . . , (m+m0 ) ), p M.
m! m0 !

Another important operation with forms is their pullback. Let be an m-form on M and
: N M be any differentiable map from another manifold N to M . Then the pullback is
an m-form on N defined by
( )p (1 , . . . , m ) = (p) ( 1 , . . . , m ), p N,

for 1 , . . . , m Tp N , where = dp () T(p) M is the pushforward of Tp N .

6. ORIENTATION AND INTEGRATION OF DIFFERENTIAL FORMS

37

6. Orientation and integration of differential forms

Definition 3.24. A manifold M is orientable if it has an atlas (U , ) such that, if (x1 , . . . , xn )
0
0
and (x1 , . . . , xn ) are any coordinate systems defined by charts in this atlas, then the Jacobian
xi
determinant det( x
i0 ) is positive. An orientation of M is a choice of such an atlas. Any other chart
that can be added to this atlas without changing the above property is called positively oriented
with respect to the given orientation. If a manifold is orientable, it can be oriented in precisely
two different ways.
Remark 3.25. It can be shown that a compact surface (2-dimensional submanifold) in R3 is
always orientable.
The integral of an n-form on an orientable n-manifold M can be defined in three steps.
In the first step assume that M = U Rn is an open set and the support of is compact.
Writing = f (x)dx1 dxn , we set
Z
Z
=
f (x) dx1 dxn ,
U

where the right-hand side is the usual multiple integral in Rn .

In the second step assume that has compact support in an open set U M where a chart
: U Rn is defined and positively oriented. Then we set
Z
Z
Z
=
= (1 ) ,
M

where the right-hand side is an integral on an open set in Rn and is hence defined in the first
step. An important part of the justification of this definition involves checking that the integral is
in fact independent of the choice of the coordinate system which is based on the transformation
formula
 i
Z
Z
x
1
n
f (x) dx dx =
f (x(u)) det
du1 dun
uj
(U )
U
which holds for any orientation preserving diffeomorphism : U (U ) between open sets in
Rn .
Finally, in the third step, consider a covering of M by positively oriented coordinate charts
(U , ) and choose a partition of unity (fi )iI subordinate to the covering (U ) (see Section 2 of
Chapter 2). Then, for any i, the product i = fi is a new n-form on M with compact support
in some U and hence its integral over M is defined. We can now set
Z
XZ
=
i
M

assuming the sum is defined, i.e. it is either finite or an absolutely convergent series. The absolute
convergence is important because there is no distinguished order for the values of i I.

38

3. TANGENT SPACES AND TENSOR CALCULUS

7. The exterior derivative and Stokes Theorem

The exterior derivative of an m-form on M is an (m+1)-form on M defined in local cooordinates
by
d = d(I dxI ) = (dI ) dxI ,
where dI is the differential of the function I (seen as a 1-form).
A more geometric definition can be obtained from the formula
Z
1
d(1 , . . . , m+1 ) = lim m+1
,
h0 h
P (h1 ,...,hm+1 )

where P (h1 , . . . , hm+1 ) is the oriented boundary of the parallelorgram spanned by the vectors h1 , . . . , hm+1 (with respect to some local coordinates). The parallelogram is obtained (an
oriented) via the parametrization
(t1 , . . . , tm+1 ) [0, 1]m+1 7 t1 h 1 + + tm+1 h m+1 P (h1 , . . . , hm+1 ).

Stokes theorem relates the integral of over the boundary of a manifold to the integral of
d over the manifold itself. The standard setting here is that of a manifold with boundary. The
definition of a manifold M with boundary M is obtained from the definition of a manifold by
allowing charts : U Rn to be homeomorphisms onto open sets in the closed half-space
H n = {(x1 , . . . , xn ) Rn : xn 0}.

The boundary M consists precisely of the points of U (for some ) that correspond (under )
to boundary points of H n .
The following is easy to see (an exercise):
Lemma 3.26. Given an atlas (U , ) as above, the collection (V , |V ), where V :=
1
({xn = 0}), defines an (n 1)-manifold structure on the boundary M . If (U , ) defines an
orientation on M , the corresponding atlas (V , |V ) also defines an orientation on M that is
said to be induced by the orientation of M .
Theorem 3.27 (Stokes theorem). Let M be an n-manifold with boundary M and be a
differentiable (n 1)-form with compact support on M . Then
Z
Z
=
d.
M

CHAPTER 4

Riemannian geometry
1. Riemannian metric on a manifold
Definition 4.1. A (smooth) Riemannian metric on a manifold M is an association to every
p M a symmetric positive definite bilinear form gp : Tp M Tp M R (hence an inner product)
such that in every local coordinates (x1 , . . . , xn ), gp is given by
j
,b
) = gij (p)ai bj
(1.1)
i
j
x
x
with smooth coefficients gij (p). The pair (M, g) is called Riemannian manifold.
Formula (1.1) is often rewritten as g = gij dxi dxj . (Note that g is not a 2-form because it is
symmetric in contrast to skew-symmetric 2-forms.)
Example 4.2. The standard metric on Rn is defined by setting gij = ij , i.e.
X

a i bi
(1.2)
g(ai i , bj j ) =
x
x
i
gp (ai

which is the standard inner product (or scalar product) on Rn . More generally, if M is a submanifold in Rn , a metric g on M can be obtained by restricting the standard metric (1.2) to Tp M Tp M
for every p M . This induced metric is classically called the first fundamental form of M .
p Given a metric g on M , the norm (length) of a tangent vector Tp M is given by kk :=
gp (, ) and the angle between two tangen vectors , Tp M is given by


g(, )
1
= cos
.
kk kk
Rb
Furthermore, the length of a smooth arc c : [a, b] M is given by the integral a kc0 (t)k dt,
where the norm of the tangent vector c0 (t) Tc(t) M is calculated as explained above. In fact, it
follows from the transformation rule that the length of the arc c coincides with the length of any
reparametrized arc e
c = c with : [e
a, eb] [a, b] being a diffeomorphism. Hence one can talk
about the length of a curve (which is an equivalence class of arcs).
Using length of curves one can define the distance d(p, q) between two points p, q M to be
the infimum of lengths of curves connecting p and q. Without proof we quote here:
Proposition 4.3. The distance d associated to a Riemannian metric g satisfies metric space
axioms. The obtained metric induces the same topology as the manifold topology.
39

40

4. RIEMANNIAN GEOMETRY

Given a metric on an oriented manifold, one can also compute volumes (measures) of subsets
using the volume form which is defined in oriented local coordinates by
q
:= det(gij ) dx1 dxn .
(1.3)

It is a direct consequence of the transformation rule of gij and of the wedge product that the
above form is well-defined and independent of the choice of oriented local coordinates.
The following general statement is a consequence of the existence of partitions of unity:
Proposition 4.4. On every manifold M there exists a Riemannian metric g.

2. THE LEVI-CIVITA CONNECTION

41

2. The Levi-Civita connection

A connection on a manifold is an additional structure permitting to differentiate vector fields
in directions of tangent vectors.
Definition 4.5. An (linear) connection on a manifold M associates to every vector field
X on M and every tangent vector Tp M another tangent vector X Tp M , called the
covariant derivative of X in the direction of , such that X is bilinear in and X and satisfies
the following Leibnitz rule:
(f X) = df ()X + f X.
(2.1)

Choosing local coordinates (x1 , . . . , xn ) on M and writing = i xi and X = X j xj we obtain,

using the the linearity and (2.1):
j

X j

i j
i X
+

+ kij i X j k ,
i
j
j
i
j
i
x x
x x
x x
x

k
where ij (p) are coefficients of the vector i xj :

X = i

(2.2)

= kij (p) k .
(2.3)
j
x
x
The connection is said to be smooth if its coefficients kij , called Christoffel symbols are smooth
functions with respect to any choice of local coordinates. Given any connection and any choice
of local coordinates, the functions kij (p) are uniquely determined by (2.3). Vice versa, given any
collection of functions kij (p), a connection in the given coordinate chart can be defined by (2.2)
and is uniquely determined by that formula.
An important geometric interpretation of a connection is that of a parallel transport of tangent
vectors along curves. Given a differentiable (or piecewise differentiable) curve c : [a, b] M , a
vector field X on M is called parallel along c if c0 (t) X = 0 for all t. It follows from the theory of
ordinary differential equations that, given a curve c and a vector X0 Tc(t0 ) , there exists a parallel
vector field X near c(t0 ) with X(c(t0 )) = X0 , which is unique along c.
Definition 4.6. Given a connection , its torsion torsion is given by

xi

T (X, Y ) := X Y Y X [X, Y ]

for any vector fields X and Y . A connection is called symmetric or torsion free if T (X, Y ) 0 for
all X and Y .
A computation in local coordinates (x1 , . . . , xn ) gives



j
i
,Y
(2.4)
= X i Y j (kij kji ) k ,
T X
i
j
x
x
x

showing that the value T (X, Y )(p) Tp M depends only on the values of X(p), Y (p) Tp M and
hence defines, for every p, a bilinear skew-symmetric map Tp M Tp M Tp M . It follows that
is symmetric if and only if kij = kji explaining the name.
For connections on a Riemannian manifold, there is a natural compatibility condition.

42

4. RIEMANNIAN GEOMETRY

Definition 4.7. A connection on a Riemannian manifold (M, g) is called metric if

D g(X, Y ) = g( X, Y ) + g(X, Y )

(2.5)

for any tangent vector and vector fields X and Y , where on the left-hand side D denotes the
directional derivative of a smooth function.
A connection is metric if and only if for any parallel vector fields X, Y along any curve, g(X, Y )
is constant along that curve.
In local coordinates, choosing X = /xi , Y = /xj , = /xk , we can rewrite (2.5) as
Dk gij = gjl lki + gil lkj .
(2.6)
Theorem 4.8 (Levi-Civita). On a Riemannian manifold there exists an unique connection
that is both symmetric and metric, called the Levi-Civita connection.
Proof. Assuming is a Levi-Civita connection and permuting indices in (2.6) we have
Dk gij = gjl lki + gil lkj

(2.7)

Di gjk = gkl lij + gjl lik

(2.8)

Dj gki = gil ljk + gkl lji .

(2.9)

Adding the first two identities, subtracting the third and using the symmetry lij = lji we obtain:
2gjl lki = Dk gij + Di gjk Dj gki .
(2.10)
Denoting by (g ij ) the inverse matrix of (gij ) and solving (2.10) for Christoffel symbols, we
obtain the classical formula
g mj
(Dk gij + Di gjk Dj gki )
(2.11)
2
showing the uniqueness. To show the existence, use (2.11) to define m
ki and hence . Then
is obviously symmetric and (2.5) can be directly verified.

m
ki =

In the important case when M is a submanifold of Rn with its standard metric, the Levi-Civita
connection on M can be obtained from the following general statement.
f, ge) be a Riemannian manifold and M a submanifold of M
f with
Proposition 4.9. Let (M
e and the Levi-Civita connection of M
f and M respectively. Then,
induced metric. Denote by
for Tp M and X tangent to M ,
e X),
X = p (
(2.12)
f Tp M is the orthogonal projection.
where p : Tp M
Proof. One verifies directly that the connection given by the right-hand side of (2.12) satisfies
all the assumptions of the Levi-Civita connection. The statements follows then by the uniqueness
of the Levi-Civita connection.


2. THE LEVI-CIVITA CONNECTION

43

f T M the complementary orthogonal projection onto the orthogDenote now by p : Tp M

p

f.
onal complement Tp M of Tp M in Tp M
f at p M is given
Definition 4.10. The second fundamental form of the submanifold M M
by
e X Y (p)),
IIp (X, Y ) = p (
(2.13)
f tangent to M .
where X and Y are vector fields on M
Lemma 4.11. The second fundamental form IIp (X, Y ) depends only on the values X(p), Y (p)
and defines a symmetric bilinear map IIp : Tp M Tp M Tp M .
e is torsion-free, one has
Proof. Since
e XY
e Y X [X, Y ] = 0.
Te(X, Y ) =

For X and Y tangent to M , their Lie bracket [X, Y ] is also tangent to M . Hence, projecting on
Tp M , we obtain the symmetry IIp (X, Y ) = IIp (Y, X). The fact that, for a fixed Y , the value of
IIp (X, Y ) depends only on X(p) follows directly from the definition (2.13). Now the symmetry
implies that the roles of X and Y can be exchanged, completing the proof of the lemma.

Examples 4.12. If M is a surface in R3 (with the standard metric), II(X, Y ) is the classical
e XT )
second fundamental form of M obtained by differentiating the coefficients of Y along X (
2
and projecting to the normal direction. If M is a curve in R , its second fundamental form is
classically called the curvature of the curve. More generally, if M is a curve in a 2-dimensional
Riemannian manifold, its second fundamental form (more precisely, its value on the oriented unit
tangent vector) is called the geodesic curvature of M .

44

4. RIEMANNIAN GEOMETRY

3. Geodesics and the exponential map

A geodesic is a curve whose tangent vector is parallel with respect to the Levi-Civita connection:
Definition 4.13. A parametrized curve c : (a, b) R M in a Riemannian manifold M is a
geodesic if
c0 c0 = 0,

(3.1)

i.e. the vector field c0 (its local extension) is parallel along c.

The main fact about geodesics is given by the following statement.
Proposition 4.14. For every p M and Tp M there exists a unique geodesic c : I M
with 0 I, c(0) = p, c0 (0) = and I R maximal.
Proof. In local coordinates (x1 , . . . , xn ) the curve c(t) is given by its coordinate functions
(x1 (t), . . . , xn (t)). Then, using (2.2), the equation (3.1) can be rewritten as


i
j
d dxk

k dx dx
+ ij
=0
(3.2)
dt dt
dt dt xk
which is a system of second order ODEs. The statement now follows from the general existence
and uniqueness result for solutions of ODE systems.

Examples 4.15. In Rn with the standard metric we have kij = 0 and hence the geodesics are
the straight lines parametrized by t 7 at + b with a, b Rn . On the sphere S the geodesics are the
great circles parametrized by the arc length and by any scalar multiple of the arc length. Indeed,
the derivative c0 (t) for such a curve c in this case is orthogonal to T S and hence the covariant
derivative is 0 in view of (2.12).
We mention without proof the fundamental local distance minimizing property of geodesics.
Theorem 4.16. For any p M there exists a neighborhood U of p in M such that for any
x, y U , there exists unique geodesic connecting x and y whose length is equal to the distance
d(x, y) (defined in Section 1) and thus not greater than the length of any other curve connecting
x and y.
Given p and as in Proposition 4.14, denote by cp, (t) the corresponding unique geodesic
defined for t I. Obviously cp,0 (t) = p for all t and hence cp,0 (1) is defined. Furthermore, it
follows from the general facts about smooth dependence of solutions of ODE systems on the
initial data, that cp, (1) is defined for Tp M in a neighborhood U of 0 in Tp M and smoothly
depends on U .
Definition 4.17. The exponential map at p M is given by
expp () := cp, (1),

U Tp M M.

Lemma 4.18. The map expp is a local diffeomorphism at 0 with d0 expp = id.

(3.3)

3. GEODESICS AND THE EXPONENTIAL MAP

45

Proof. It follows directly from the definition of geodesics that with c(t) being geodesic, any
curve c(t) is also a geodesic for any R. Hence we obtain the homogeneity cp, (t) = cp, (t)
and hence
d 
d 
d 
d0 expp (v) =  expp (tv) =  cp,tv (1) =  cp,v (t) = v
dt t=0
dt t=0
dt t=0
as required (the last equality follows from the construction of cp,v (t). The fact that expp is a local
diffeomorphism at 0 follows from the inverse mapping theorem.

Using Lemma 4.18 one can defined local coordinates near p by taking as a chart the local
inverse of expp . These are the important normal coordinates at p in which every straight line
through the origin is a geodesic.

46

4. RIEMANNIAN GEOMETRY

4. Curvature and the Gauss equation

The curvature or the (Riemannian) curvature tensor of a Riemannian manifold M associates
to vector fields X, Y , Z the new vector field given by
R(X, Y )Z := X Y Z Y X Z [X,Y ] Z,

(4.1)

where is the Levi-Civita connection. More generally, given any connection , it curvature tensor
is given by (4.1). The name tensor reflects the fact that the value R(X, Y )Z at a point p depends
only on the values of X(p), Y (p) and Z(p). Indeed, a calculation in local coordinates shows that




i
k
s
j
R X
Z
= Rijk
X iY j Z k s
,Y
(4.2)
i
j
k
x
x
x
x
with
sjk sik

+ sil ljk sjl lik .

(4.3)
xi
xj
The curvature tensor R(X, Y )Z satisfies the Bianchi identity (similar to the Jacobi identity
for Lie brackets):
s
Rijk
=

R(X, Y )Z + R(Y, Z)X + R(Z, X)Y = 0.

(4.4)

Indeed, since the value of (R(X, Y )Z) at p depends only on X(p), Y (p) and Z(p), we may choose
X, Y, Z to have constant coefficients in some local coordinates and hence all Lie brackets to be
zero. Then the symmetry of implies X Y = Y X, Y Z = Z Y and Z X = X Z. Then
(4.4) is obtained directly from (4.1) using these relations.
It is convenient to associate with R(X, Y )Z the closely related 4-multilinear scalar form
R(X, Y, Z, V ) := g(R(X, Y )Z, V ).

(4.5)

Since g is positive definite, R(X, Y, Z, V ), in turn, uniquely determines R(X, Y )Z.

Lemma 4.19. The form R(X, Y, Z, V ) has the following symmetry properties:
R(X, Y, Z, V ) = R(Y, X, Z, V )

(4.6)

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

(4.7)

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

(4.8)

R(X, X, Z, V ) = R(X, Y, Z, Z) = 0.

(4.9)

In particular,
The proofs are standard and can be found in most books. (In case M is a submanifold in Rn
these properties follow immediately from the Gauss equation (4.10) below).
The following important relation known as Gauss equation generalizes the famous Theorema
Egregium of Gauss.

4. CURVATURE AND THE GAUSS EQUATION

47

Theorem 4.20 (The Gauss equation). Let M be a submanifold of a Riemannian manifold

f
e the corresponding curvature tensors and by II
(M , ge) with the induced metric. Denote by R and R
f. Then, for X, Y, Z, V tangent to M , one has the relation
the second fundamental form of M in M




e
R(X, Y, Z, V ) R(X,
Y, Z, V ) = ge II(Y, Z), II(X, V ) ge II(X, Z), II(Y, V ) .
(4.10)
f = Rn with the standard metric, R
e = 0 and hence R is given by the right-hand
In particular, if M
side of (4.10).
Proof. By defintion (4.1) we have

e
e X
e Y Z, V ge
eY
e X Z, V ge
e [X,Y ] Z, V .
R(X,
Y, Z, V ) = ge

(4.11)

Using the formula (2.12) and the definition (2.13) of the second fundamental form we write
e X
eY Z =
e X Y Z +
e X II(Y, Z).

(4.12)

e X
e Y Z = X Y Z + II(X, Y Z) +
e X II(Y, Z).

(4.13)

Since Y Z is also tangent to M , we can again apply (2.12) and (2.13) to the first term:
Since II(X, Y Z) is normal to T M , we have

ge II(X, Y Z), V = 0.

(4.14)

e is metric (2.5), we obtain

For the same reason ge(II(Y, Z), V ) = 0 and using the fact that






e X II(Y, Z), V + ge II(Y, Z),
e XV
0 = DX ge II(Y, Z), V = ge
(4.15)

and hence, using again that II(Y, Z) is normal to T M and II(X, V ) is the orthogonal projection of
e X V to the normal space,

e X II(Y, Z), V = e
e X V = e
ge
g II(Y, Z),
g II(Y, Z), II(X, V ) .
(4.16)
We can now compute the scalar product of (4.13) with V using (4.14) and (4.16):






e X
e Y Z, V = ge X Y Z, V ge II(Y, Z), II(X, V ) .
ge

(4.17)

Exchanging X and Y we obtain

eY
e X Z, V = ge Y X Z, V ge II(X, Z), II(Y, V ) .
ge

(4.18)

e [X,Y ] Z,
Finally, since [X, Y ] and V are tangent to M and [X,Y ] Z is the tangential component of
we have




e [X,Y ] Z, V = ge [X,Y ] Z, V .
ge
(4.19)
The required formula (4.10) follows directly from (4.17) (4.19).

48

4. RIEMANNIAN GEOMETRY

Example 4.21. In case M is a surface in R3 , the second fundamental form is scalar valued
and can be diagonalized in an orthonormal basis (E1 , E2 ) of vector fields, i.e.
g(Ei , Ej ) = ij ,

II(Ei , Ej ) = i ij .

(4.20)

The eigenvalues 1 , 2 are classically called the principal curvatures of M . The Gauss equation
(2.12) now yields
R(E1 , E2 , E1 , E2 ) = 1 2 .
(4.21)
The right-hand side is the classical Gaussian curvature of M which, in view of (4.21) depends only
on the intrinsic metric of M but not on the isometric embedding of M into R3 . The latter fact is
known as Theorema Egregium of Gauss. The formula (4.21) completely determines the tensor R
via the symmetry relations in Lemma 4.19:
R(E1 , E2 , E1 , E2 ) = R(E2 , E1 , E2 , E1 ) = R(E2 , E1 , E1 , E2 ) = R(E1 , E2 , E2 , E1 ) = 1 2
(4.22)
and all other values of R(Ei , Ej , Ek , El ) are zero.
Example 4.21 shows that the curvature tensor of the surface is essentially given by the number
1 2 in (4.21). For manifolds of higher dimension, one obtains a number called the sectional
curvature for every 2-dimensional subspace of the tangent space.
Definition 4.22. The sectional curvature of a Riemannian manifold (M, g) in the direction
of the plain defined by two linearly independent vectors , Tp M is defined by
K(, ) :=

R(, , , )
g(, )g(, ) g(, )2

(4.23)

The quantity in the denominator of (4.23) is the squared area of the parallelogram spanned
by and . The sectional curvature can be seen as the normalized value of R(, , , ). The
main fact about the ratio in (4.23) is that it depends only on the plain spanned in Tp M by
and but not on the actual choice of and . This can be seen by observing that any basis
change in a two-plain can be obtained as composition of the elementary changes (, ) 7 (, ),
(, ) 7 (, ) and (, ) 7 ( + , ) and verifying that the ratio K(, ) does not change under
these transformations. In particular, (, ) can be chosen to be an orthonormal basis in which case
we have K(, ) = R(, , , ).
Other important quantities obtained from R are Ricci tensor and the scalar curvature, both
obtained using traces or contractions.
The Ricci tensor at a point p M is given by the trace (contraction)
X
Ric(X, Y )(p) :=
R(X, Ei , Y, Ei )(p)
(4.24)
i

with the summation is taken over an orthonormal basis (E1 , . . . , En ) of Tp M . The right-hand side
is equal to the trace of the operator Z 7 R(X, Z)Y . The Ricci curvature is given by the ratio
Ric(X) :=

Ric(X, X)
g(X, X)

(4.25)

4. CURVATURE AND THE GAUSS EQUATION

49

and clearly depends only on the direction of X. It follows from Lemma 4.19 that Ric(X, Y ) is
symmetric in X and Y .
The scalar curvature is a (scalar) function on M obtained by taking the trace one more time:
X
S(p) :=
Ric(Ei , Ei )(p).
(4.26)
i