Special Mathematics Lecture

Differential geometry
Nagoya University, Fall 2024

Lecturer: Serge Richard

Teaching assistant: Fujii Moe

Goals of these Lectures notes:

Provide the necessary background information for understanding the main ideas of
differential geometry. These notes correspond to 14 lectures lasting 90 minutes each.

Website for this course:

http://www.math.nagoya-u.ac.jp/∼richard/SMLfall2024.html
Comments or corrections are welcome:
[email protected]
Contents

1 Manifolds 2
1.1 Elements of topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Topological manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Smooth manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Immersion, submersion, embedding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2 Tangent space and vector fields 13

2.1 Tangent space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Vector fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3 Vector fields and flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3 Tensors, tensor fields, differential forms 24

3.1 Tensors and exterior algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2 Dual basis and coordinate coframe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3 Tensor fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.4 Differential forms and exterior derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.5 Orientation of a manifold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

4 Integration on manifolds 35
4.1 Integration of m-forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2 Line integrals on manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

5 Riemannian manifolds 41
5.1 Definition and basic properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.2 The Frenet frame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.3 Differentiation on submanifolds of Rn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.4 Differentiation on Riemannian manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.5 Geodesics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

6 Curvature 57
6.1 Geometry of surfaces in R3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
6.2 Various curvatures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
6.3 Cartan structure equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
6.4 Holonomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

7 General relativity 67

b
 7.1 Pseudo-Riemannian manifolds and Einstein fields equations . . . . . . . . . . . . . . . . . . 67
7.2 Causality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

Index 73

Bibliography 76

c
Motivation

One can consider two approaches for differential geometry: the extrinsic approach and the intrinsic approach.
In the extrinsic approach, one looks at curves, surfaces or volumes in a bigger ambient space, typically Rn for
n large enough. This is the approach used in Calculus II and most of the representations are taking place in
this bigger space. Objects are more easily visualized in this representation, and the theory is usually simpler.
In the intrinsic approach, there is no ambient space, no representation of the objects in a bigger space, no
outside world. The 2-dimensional inhabitants of a flatland would have no idea of the third dimension. This
approach is less intuitive, but essential for some advanced theories, as for example for relativity. In this
context, what would be the outside ambient space of the 4-dimensional space-time space ?
Now, Nash embedding theorems (or imbedding theorems) state that every Riemannian manifold can be iso-
metrically embedded into some Euclidean space. It means that in good situations, the intrinsic approach can
be embedded into a larger space. However, this idea is not always useful, and the intrinsic approach has its
own value. In this course, we shall mainly develop the intrinsic point of view, but note that most pictures use
the extrinsic point of view.

1
Chapter 1

Manifolds

1.1 Elements of topology

In this section, we introduce a few notions of topology, which is a subject of the root of many mathematical
developments.

Definition 1.1.1 (Topological space). A topological space(M, T ) is a set M together with a family T of subsets
of M satisfying

(i) Ø and M belong to T ,

(ii) If Vα ∈ T , then ∪α Vα ∈ T , (T stable for arbitrary unions)

N V ∈ T,
(iii) If Vi ∈ T , then ∩i=1 (T stable for finite intersections).
i

The elements of T are usually called the open sets of M, and one says that T defines the topology of M. For
any V ∈ T , its complement V c := M \ V is called a closed set.

Be aware that the elements of T are called open sets by definition. In some examples, they do not coincide
with what could be intuitively thought as an open set.

Exercise 1.1.2. Exhibit a couple of topological spaces, by describing M and T and by checking the three
properties mentioned above.

Exercise 1.1.3 (Rn and the usual topology). Recall that a set U on Rn is “open” if for any x ∈ U there exists
a ball B(x, r) centred at x and of radius r > 0 such that B(x, r) ⊂ U. With this definition, check that the set of
all “open” sets on Rn defines a topology on Rn . This topology corresponds to the usual topology on Rn , and
justifies why the “open” sets can be called open, according to Definition 1.1.1. Note also that we simply write
Rn for the topological space, without mentioning explicitly the set of open sets.

Definition 1.1.4 (Neighbourhood). Let (M, T ) be a topological space, and let p ∈ M. A neighbourhood of p
is any element V ∈ T satisfying p ∈ V. The set of all neighbourhoods of p is denoted by N p .

In other words, a neighbourhood of p ∈ M is any open set containing p. Topological spaces are classified
according to some basic properties of open sets, called the separation axioms. In the sequel, we shall consider
only one rather essential property.

2
Definition 1.1.5 (Hausdorff property). A topological space (M, T ) is Hausdorff if for any p1 , p2 ∈ M with
p1 , p2 there exist V1 ∈ N p1 and V2 ∈ N p2 such that V1 ∩ V2 = Ø.
This property is also called the separability property. It is a necessary condition for the unicity of the limit of
any converging sequence. If this condition is not satisfied, surprising phenomena can take place.
Exercise 1.1.6. Study the separation axioms and exhibit some spaces which are not Hausdorff. Show that Rn
is Hausdorff.
Quite often, it is very difficult to describe all elements of T , namely all open sets of a topological space
(M, T ). The notion of basis is therefore very useful.
Definition 1.1.7 (Basis). A subset B := {Vα } ⊂ T is a basis for the topological space (M, T ) if for any p ∈ M
and any U ∈ N p , there exists V ∈ B such that p ∈ V ⊂ U.
In other words, a basis should contain enough elements such that it can be contained in any neighbourhood
of p, and contains p as well. Equivalently, B is a basis of (M, T ) if any element of T can be represented as a
union of elements of B. For examples, the set of all open intervals on R is a basis for the usual topology on
R, or the set {B(x, r) | x ∈ Rn , r > 0}, where B(x, r) denotes the open ball centred at x and of radius r, defines
a basis for the usual topology on Rn . Observe that these bases contain quite a lot of elements. Nevertheless,
they contain less elements than the corresponding set T . In some cases, basis can even be rather small:
Definition 1.1.8 (Second countable). A topological space (M, T ) is called second countable if it possesses a
basis which is countable (meaning that its elements can be indexes by N).
Exercise 1.1.9 (Rn is second countable). Check that Rn with the usual topology (see Exercise 1.1.3) is a
second countable topological space.
Exercise 1.1.10 (Discrete topology). Consider a set M endowed with the discrete topology Td :

Td := {V ⊂ M}.

It means that Td contains all subsets of M, or equivalently all subsets of M are open sets. We call (M, Td ) a
discrete space. Show that (M, Td ) is a Hausdorff space, and that (M, Td ) is second countable if and only if M
is a countable set.
Let us add one more characterization of M in terms of its topology.
Definition 1.1.11 (Connected space). The set M of a topological space (M, T ) is connected if it is not the
disjoint union of two non-empty open sets. If M is the disjoint union of two (or more) non-empty open sets,
then the set M is called a disconnected space.
When considering part of a topological space, a new notion appears very naturally.
Definition 1.1.12 (Subspace topology). Let (M, T ) be a topological space and let U ⊂ M. The subspace
topology consists in the family of sets TU := {V ∩ U | V ∈ T }, and this family makes (U, TU ) a new
topological space. The family TU is also called relative topology, induced topology, or trace topology.
Exercise 1.1.13. Check that (U, TU ) defines a topological space, and show that open sets of the subspace
topology might not be open sets of the original topology.
For the following definition, we shall need two more notions: Let (M, T ) be a topological space and U ⊂ M.
The family {Vα } ⊂ T is an open cover of U if U ⊂ ∪α Vα . A subcover of this open cover is a family {Vβ } ⊂ {Vα }
such that U ⊂ ∪β Vβ . It means that the subfamily {Vβ } is itself an open cover of U.

3
Definition 1.1.14 (Compact set). A subset U ⊂ M of a topological space (M, T ) is compact if any open cover
of A admits a finite subcover.
Be careful, the previous definition is very subtle.
Exercise 1.1.15. For any −∞ < a < b < ∞, show that the subset [a, b] of R (endowed with the usual topology)
is a compact set. Show also that the subsets (a, b), (a, b] or [a, b) are not compact, and similarly that (−∞, a]
or [b, ∞) are not compact.
Topological spaces are very much linked with the notion of continuity. More precisely:
Definition 1.1.16 (Continuous map). Let (M, T ) and (N, S) be topological spaces, and let f : M → N. The
map f is continuous if for any U ∈ S, its preimage f −1 (U) is an element of T , or more precisely

for any U ∈ S, one has f −1 (U) ∈ T ,

where f −1 (U) := {x ∈ M | f (x) ∈ U}.

Be aware that the previous definition does not mean that f is bijective. The notion of preimage of U by f is
well-defined for any f : M → N and any U ⊂ N.
Exercise 1.1.17. Consider a function f : Rn → R continuous in the sense defined above. Check that it is also
continuous in the sense provided in Calculus II. Show that the converse holds as well. Thus, these two notions
are equivalent, but Definition 1.1.16 holds in a wider context.
In the setting of the definition of a continuous map, if the map is bijective a more precise notion is very useful.
Definition 1.1.18 (Homeomorphism). Let (M, T ) and (N, S) be topological spaces, and let f : M → N be a
bijective map. If f and its inverse f −1 are continuous, then f is called a homeomorphism.
In the setting of topological spaces, a homeomorphism means that the two spaces are very similar. By con-
sidering two subsets U ⊂ M and V ⊂ N, one says that these two subsets are homeomorphic if there exists
a homeomorphism f : U → V when U is endowed with the subspace topology TU and V is endowed with
the subspace topology SV . In other words, f is a homeomorphism from the topological space (U, TU ) to the
topological space (V, SV ).

1.2 Topological manifolds

With the topological definitions of the previous section, it is now possible to define precisely the first version
of a manifold:
Definition 1.2.1 (Topological manifold). A m-dimensional topological manifold consists in a topological
space (M, T ) which is Hausdorff and second countable, and such that for any p ∈ M there exists U ∈ N p
homeomorphic to an open subset of Rm .
The figure on the cover of this course shows a topological manifold M with a subset U ⊂ M and a homeomor-
phism between U and its image φ(U) on R2 . Observe that φ is a homeomorphism if and only if it is bijective
and bicontinuous, when φ(U) is endowed with the subspace topology inherited from the usual topology on
R2 .
The unit circle
S := (x, y) ∈ R2 | x2 + y2 = 1


4
or the unit ball
B := (x, y, z) ∈ R3 | x2 + y2 + z2 = 1


are respectively 1-dimensional and 2-dimensional topological manifolds. Additional example of 2-dimensional
topological manifolds are given in Figure 1.1.

Figure 1.1: Manifolds from [2, p. 9]

Exercise 1.2.2. Provide a parametrization of the 2-torus and check that it is a 2-dimensional topological
manifold. For b > 0, check also that the following surface is a 2-dimensional topological manifold:

M = s cos(t), s sin(t), bt | (s, t) ∈ R2 .



Provide a picture of this manifold.

It turns out that any topological manifold (M, T ) is locally connected and locally compact, meaning that for
any p ∈ M there exists V ∈ N p such that (V, TV ) is connected and compact. The manifold M is also a
countable union of compact subsets. Additional properties also hold for any topological manifold, see [2,
Thm. 1.3.6]. These properties are inherited from the local homeomorphisms with subsets of Rm .
For compact topological manifolds of dimension 2, let us mention a very pleasant result, see [2, Thm. I.4.1]
and Figure 1.2.
Theorem 1.2.3. Every compact, connected, orientable1 2-dimensional topological manifold is homeomorphic
to a sphere with handles added. Two such manifolds with the same number of handles are homemorphic and
conversely, so that the number of handles (called the genus) is the only topological invariant.
1
We do not define this concept now and refer to [11].

5
 Figure 1.2: Homeomorphic topological manifolds, from [9]

Let us stress that all definitions so far do not require any visualization of these manifolds. In fact, visualizing
these 2-dimensional manifolds in R3 can generate some confusions. For example, the manifolds of Figure 1.3
are all homeomorphic, despite their different appearance in the Euclidean space R3 .
Let us now fix a point p ∈ M, and let U ∈ N p be a neighbourhood of p which is homeomorphic to an open
subset of Rm . We denote by φ : U → φ(U) ⊂ Rm this homeomorphism. Since for any q ∈ U the image φ(q)
is an element of Rm , it has m-coordinates, and it is thus natural to set

φ(q) =: x1 (q), x2 (q), . . . , xm (q) ∈ φ(U) ⊂ Rm .

We call this n-tuples the local coordinates of q ∈ U, or alternatively the map

U ∋ q 7→ x1 (q), x2 (q), . . . , xm (q) ∈ Rm

(1.2.1)

a local coordinate function. We also call the pair (U, φ) a coordinate neighbourhood, a coordinate system, or
a chart. Note that the initial point p does not play any role in this construction, it is valid for any other point
of U. By definition of the manifold, observe also that the local coordinate function is a continuous function.
Topological manifolds with boundaries can also be described. For that purpose, let us set

H n := (x1 , . . . , xn ) ∈ Rn | xn ≥ 0


which corresponds to a half-space in Rn . The set {(x1 , . . . , xn−1 , 0) ∈ Rn } is called the boundary of H n and is
denoted by ∂H n . As for Rn , we shall simply denote by Hn the set H n endowed with the subspace topology
inherited from Rn .
Definition 1.2.4 (Topological manifold with boundary). A m-dimensional topological manifold with boundary
consists in a topological space (M, T ) which is Hausdorff and second countable, and such that for any p ∈ M
there exists U ∈ N p homeomorphic either to an open subset of Rm or to an open subset of Hm with the image
of p in ∂H m .

6
 Figure 1.3: 2-dimensional topological manifolds of genus 3, all homeomorphic, from [2, p. 15]

Examples of topological manifolds with or without boundary are presented in Figure 1.4.
Example 1.2.5. Show that [0, 1] ⊂ R and that

{(x, y) ∈ R2 | x2 + y2 = 1 and y ≥ 0}

are topological manifolds with boundary. Describe the boundaries. What about [0, 1) ? What about [0, 1]2 ⊂
R2 ?

1.3 Smooth manifolds

For many applications, topological manifolds are not sufficient, a differential structure is needed. Observe
however that such a structure can only be borrowed from Rn , since it does not exist a priori on the manifold
M.
Definition 1.3.1 (Smooth manifold). A smooth manifold or C ∞ -manifold M of dimension m is a topological
manifold together with a family of coordinates neighbourhoods U := (Uα , φα )α with φα : Uα → Rm such that
(i) ∪α Uα = M,
(ii) For any α, β the maps
φα ◦ φ−1
β : φβ U α ∩ U β → R
m

and
φβ ◦ φ−1
α : φα U α ∩ U β → R
m

are smooth maps, or C ∞ -maps,

(iii) The family U is maximal, namely if (U, φ) is another coordinate neighbourhood on M which satisfy
condition (ii) with φα = φ and arbitrary φβ , then (U, φ) ∈ U.

7
 Figure 1.4: 2-dimensional topological manifolds without and with boundary

A family U satisfying conditions (i) and (ii) is called a smooth atlas or a C ∞ -atlas. It means that a smooth
manifold is a topological manifold endowed with a maximal smooth atlas. An atlas is also referred to as a
differential structure on M, and one sometimes considers C k -atlas or analytic atlas. Note that not all topolog-
ical manifolds admit such a differential structure, or that some manifolds admit more than one inequivalent
differential structure, see Figure 1.5. The existence and the unicity of such structures is a subtle question
which we will not consider. Fortunately, there exist many manifolds endowed with a smooth atlas.

Figure 1.5: Number of inequivalent smooth structures on the n-sphere Sn for the values of n from 1 up to 20,
from [10]

Exercise 1.3.2. Show that given a family U satisfying (i) and (ii), there always exists a maximal family
containing U, see also [2, Thm. III.1.3]. It means that any smooth atlas can always be completed into a
maximal smooth atlas.
Exercise 1.3.3 (r). Study some examples of smooth manifolds, as for example the n-dimensional Euclidean
space Rn , the n-sphere Sn , the real projective space Pn (R), or some group of matrices (Lie groups).
Let us now compare two smooth manifolds. For a given smooth manifold M, we simply call a chart any
element of its smooth atlas.
Definition 1.3.4 (Smooth map between smooth manifolds). Let M and N be smooth manifolds. A map F :
M → N is a smooth map if for any p ∈ M there exist a coordinate neighbourhood (U, φ) of p and a coordintate

8
neighbourhood (V, ψ) of F(p) with F(U) ⊂ V such that the map
ψ ◦ F ◦ φ−1 : φ(U) → ψ(V) (1.3.1)
is smooth.
In the sequel, whenever we write ψ ◦ F ◦ φ−1 : φ(U) → ψ(V) we shall implicitely assume that F(U) ⊂ V,
in order to have a well defined expression. The map (1.3.1) is called a local representation of F. The set of
all smooth maps from M to N is denoted by C ∞ (M, N). In the special case N = R we simply set C ∞ (M) for
C ∞ (M, R).
Definition 1.3.5 (Diffeomorphism). Any bijective map F : M → N satisfying F ∈ C ∞ (M, N) and F −1 ∈
C ∞ (N, M) is called a diffeomorphism.
We sometimes consider a local version of the previous definition, namely let F : M → N and let p ∈ M,
then F is a local diffeomorphism at p if there exists U ∈ N p and V ∈ NF(p) such that F|U : U → V is a
diffeomorphism. Here the notation F|U means the restriction of F to U.
Consider again the function provided in (1.3.1) between a smooth manifold M of dimension m and a smooth
manifold N of dimension n. Because this function is defined from a subset of Rm to a subset of Rn , the notion
of Jacobian of this function is well defined.
Definition 1.3.6 (Rank of a smooth map). Let F be a smooth map between a smooth manifold M of dimension
m and a smooth manifold N of dimension n. Let p ∈ M and (U, φ) and (V, ψ) be two charts of M and N
respectively such that p ∈ U and F(U) ⊂ V. The rank of F at p is defined by the rank of the Jacobian matrix
at p, namely
 ∂F 1
 ψ,φ (φ(p)) . . . ∂Fψ,φ (φ(p))
1 
 ∂x1 ∂x m
.. ..

rank(F) p := rank 
 n . ... .


 ∂Fψ,φ ∂Fψ,φ
n
(φ(p)) . . . ∂xm (φ(p))

∂x1
with Fψ,φ := ψ ◦ F ◦ φ−1 defined from φ(U) ⊂ Rm to ψ(V) ⊂ Rn .
Clearly, rank(F) p ≤ min{m, n}. It turns out that this rank at p is independent of the choice of the local charts,
which justifies why φ and ψ are not mentioned in the notation for the rank. The case of interest will be when
this rank is also independent of the choice of p, but this is not always the case. If the rank is independent of
p, which is denoted by rank(F), the following result holds [2, p. 70]:
Theorem 1.3.7. Let F be a smooth map between a smooth manifold M of dimension m and a smooth manifold
N of dimension n, and let us assume that rank(F) = k. Let p ∈ M. Then there exist two charts (U, φ) and
(V, ψ) of M and N respectively with F(U) ⊂ V such that p ∈ U, φ(p) = 0 ∈ Rm , ψ F(p) = 0 ∈ Rn , and

Fψ,φ (x1 , . . . , xm ) = (x1 , . . . , xk , 0, . . . , 0) ∈ Rn ,

with Fψ,φ := ψ ◦ F ◦ φ−1 defined from φ(U) ⊂ Rm to ψ(V) ⊂ Rn .
A direct consequence of this statement is that F can only be a diffeomorphism if m = n = k. If not, k would
be strictly smaller than m or than n, in which case it is not possible that both F and F −1 are injective.

1.4 Immersion, submersion, embedding

In this section we describe various ways of representing a smooth manifold in an ambient space. Note that
some of the following definitions depend on the authors.

9
Definition 1.4.1 (Immersion, submersion). Let F : M → N be a smooth map between smooth manifolds of
dimension m and n respectively. The map F is called an immersion if rank(F) = m. The map F is called a
submersion if rank(F) = n.

For example, consider M = (a, b), N = Rn , and F : M → N any smooth parametric curve c : (a, b) ∋ t 7→
c(t) ∈ Rn , with −∞ ≤ a < b ≤ ∞. Then the curve defined by c corresponds to an immersion if dc
dt (t) , 0 ∈ R
n

for any t ∈ (a, b).

Exercise 1.4.2. Exhibit a few smooth parametric curves in Rn , with some of them defining an immersion
and some of them not defining an immersion. Justify your statement. Inspiration can be obtained from [2,
p. 70–72].

Exercise 1.4.3. Exhibit a few smooth parametric surfaces in R3 , with some of them defining an immersion
and some of them not defining an immersion. Justify your statement.

Note that an immersion F needs not be injective, the condition is only on the Jacobian matrix associated with
F. However, if F defines an injective immersion, then its image F(M) can be endowed with the topology and
the differential structure of M. Note that this is not really interesting since the ambient manifold N does not
play any role in this construction. In such a situation, the image F(M) is called a submanifold or an immersed
submanifold2 , and F realizes a diffeomorphism between M and the image F(M).

When the topology of the ambient space N is taken into account, injective immersions become more interest-
ing, since various phenomena can take place. For example, Figure 1.6 represents an injective immersion with
a neighbourhood of the point p in the topology of M, and a neighbourhood of the point F(p) in the topology
of the ambient space R2 . When F(M) is endowed with the subspace topology inherited from R2 (obtained by

Figure 1.6: An injective immersion from [7, p. 111]

taking the intersections of F(M) with open sets of R2 ), it is clear that the neighbourhoods of p in M and the
neighbourhoods of F(p) in F(M) do not look similar: The former contains only two “parts” emanating from
p, while the latter contains an additional “part” converging to F(p).

Exercise 1.4.4. Study a similar situation in Exercise III.4.10 of [2] and which is related to Figure 1.7.

In the next definition, we provide the right concept for avoiding the situation exhibited in the previous example
and in the exercise.

Definition 1.4.5 (Embedding). Let M and N be smooth manifolds, and let F : M → N be an injective
immersion. The map F is called an embedding (or an imbedding) if F defines a homeomorphism between M
and F(M) when the latter is endowed with the relative topology inherited from N. In this case, F(M) is called
an embedded submanifold.
2
Some authors use this terminology even if the map F is not injective, and then use the terminology of injective immersed
submanifolds.

10
 Figure 1.7: An injective immersion from [2, p. 73]

Quite interestingly, the difference between immersion and embedding is of a global nature, it does not depend
on F locally. More precisely, following [2, Thm. III.4.12] one has:
Theorem 1.4.6. Let F : M → N be an immersion. Then each p ∈ M has a neighbourhood U such that F|U
is an embedding of U in N.
We close this section with one more type of submanifold. Its interest is for checking if a subset of an ambient
space is indeed a smooth manifold. In the following definition, a typical example is N = Rn . For ε > 0 we
shall also use the notation
Cεn (0) := (x1 , . . . , xn ) ∈ Rn | |x j | < ε for j ∈ {1, . . . , n}


which represents the open cube in Rn centred at 0 and of side 2ε.

Definition 1.4.7 (Submanifold property). A subset M of a smooth manifold N has the m-submanifold property
if for any p ∈ M there exists a chart (U, φ) of N with p ∈ U such that
(i) φ(p) = 0 ∈ Rn ,
(ii) φ(U) = Cεn (0) for some ε > 0,
(iii) φ(U ∩ M) = (x1 , . . . , xn ) ∈ Cεn (0) | xm+1 = xm+2 = · · · = xn = 0 .


Such a chart is called the preferred coordinates with respect to M.

We refer to Figure 1.8 for an illustration of the submanifold property. Based on this property, smooth mani-
folds can be obtained.
Lemma 1.4.8. Let M be a subset of a smooth manifold N having the m-submanifold property. Then, with
the subspace topology inherited by N, M is a m-dimensional topological manifold. All preferred coordinates
define a smooth atlas on M, once restricted to the first m coordinates. The corresponding maximal smooth
atlas makes M a smooth manifold, for which the inclusion map i : M → N is an embedding.
The proof of the previous statement is provided in [2, Lem. III.5.2] to which we refer for additional informa-
tion. Let’s finally give a name to the structure just obtained:

11
 Figure 1.8: The 2-manifold property, from [2, p. 76]

Definition 1.4.9 (Regular submanifold). Let M be a subset of a smooth manifold N having the m-submanifold
property. The corresponding smooth manifold M described in the previous lemma is called a regular subman-
ifold.

12
Chapter 2

Tangent space and vector fields

2.1 Tangent space

Let us firstly recall from Calculus II that if M is a parametric surface in R3 , if p ∈ M and if

c : (−ε, ε) ∋ t 7→ c(t) ∈ M
dc(t)
is a continuously differentiable parametric curve on M with c(0) = p, then the vector c′ (0) = dt t=0 is tangent
to M at p. The set of all such vectors generates the tangent space at p, see Figure 2.1.

Figure 2.1: A manifold M and the tangent spaces T p (M) in the extrinsic picture

Clearly, this definition uses the ambient space R3 . We shall now generalize it for any smooth manifold, and
make it intrinsic. The coming construction is a bit abstract, we shall come back to more concrete examples
subsequently.
Let M be a smooth manifold and let p ∈ M. We set C ∞ (p) for the set of all equivalent classes of smooth real-
valued functions on M defined on a neighbourhood of p. More precisely, two smooth functions are identified
if they coincide on a neighbourhood of p. Note that these functions need not be defined on the entire manifold

13
M, it is enough if they are defined on a open set of M containing p. The elements of C ∞ (p) are called germs of
C ∞ -functions at p. One easily observes that C ∞ (p) is an algebra (a vector space together with a multiplication
and some compatibility conditions).
Definition 2.1.1 (Tangent space and tangent vector). Let M be a smooth manifold and p ∈ M. The tangent
space T p (M) of M at p is the set of all maps X p : C ∞ (p) → R satisfying the following two conditions for
α, β ∈ R and f, g ∈ C ∞ (p) :
(i) X p (α f + βg) = αX p ( f ) + βX p (g), (linearity)
(ii) X p ( f g) = g(p) X p ( f ) + f (p) X p (g). (Leibniz rule)
The set T p (M) is endowed with the addition X p + Y p ( f ) = X p ( f ) + Y p ( f ) and with the scalar multiplication


αX p ( f ) = α X p ( f ), which turns T p (M) into a vector space. A tangent vector at p is any element X p ∈ T p (M).

Let us emphasize that this definition is independent of any coordinate system and does not depend on any
ambient space. In this sense, this definition is intrinsic. The link between this approach and the approach of
Calculus II is mentioned in Exercise 2.1.10.
Let us now look at this definition in the special case M = Rm , and more precisely for any p ∈ Rm , what is X p ?
For any v ∈ Rm and for any f ∈ C ∞ (p), let us set
m
X  ∂f 
X p ( f ) := vj (p) ≡ v · [∇ f ](p). (2.1.1)
j=1
∂x j

Recall that the r.h.s. represents the directional derivative in direction v, as introduced in Calculus II. Then,
one easily checks that the two conditions of Definition 2.1.1 are satisfied, and therefore X p defined by (2.1.1)
is an element of the tangent space T p (Rm ). In fact, it turns out that all elements of T p (Rm ) are of the form
introduced in (2.1.1). Thus, we have obtained a description of T p (Rm ) and can introduce a basis for this vector
space. This basis is usually denoted by
 ∂ ∂  m
,..., m

or alternatively E j,p (2.1.2)
∂x1 p ∂x p j=1

∂f
with E j,p ( f ) := (p) for any f ∈ C ∞ (p). It is clear that the above set is a basis of T p (Rm ) since the r.h.s. of
∂x j
(2.1.1) can be written as a linear combination of elements of this basis. Thus, the tangent space of Rm at p can
be identified with the m-dimensional space generated by the basis provided in (2.1.2).
Our next aim is to transport the basis constructed above for T p (Rm ) to T p (M), for any smooth manifold. For
that purpose, let us firstly state one fundamental abstract result [2, Thm IV.1.2]. Its proof is essentially a
routine computation. Recall that the notion of smooth maps between smooth manifolds has been introduced
in Definition 1.3.4. For any smooth map F : M → N between two smooth manifolds and for any p ∈ M one
sets
F ∗ : C ∞ F(p) → C ∞ (p) with F ∗ ( f ) := f ◦ F, ∀ f ∈ C ∞ F(p)



(2.1.3)
and

F∗ : T p (M) → T F(p) (N) with F∗ (X p ) ( f ) := X p F ∗ ( f ) , ∀X p ∈ T p (M) and f ∈ C ∞ F(p) .

 


(2.1.4)

A pictorial representation of F∗ is given in Figure 2.2. In the next statement, we provide some properties of
F ∗ and of F∗ . Note that a homomorphism is a map which preserves the structures.

14
 Figure 2.2: The tangent spaces T x (M) and T F(x) (N) and the map φ∗ ≡ dφ between them

Theorem 2.1.2. In the framework introduced above, the map F ∗ is a homomorphism of algebras while the
map F∗ is a homomorphism of vector spaces. In addition, if H = G ◦ F is the composition of two smooth maps
between manifolds, then H ∗ = F ∗ ◦ G∗ and H∗ = G∗ ◦ F∗ .
The map F∗ : T p (M) → T F(p) (N) is often called the differential of F, and various notations are used for it:
dF, DF or F ′ . Here and in the sequel, we shall use ∗ as a subscript when it goes in the direction of F, while
we shall use ∗ as the superscript when it goes in the direction opposite to F, as in F ∗ : C ∞ F(p) → C ∞ (p).

Exercise 2.1.3. Prove the above theorem, by getting inspiration from the proof of [2, Thm. IV.1.2].
In the special case of a diffeomorphism F : M → N, as introduced in Definition 1.3.5, the previous state-
ment can be strengthened. An isomorphism is a bijective map preserving the structures, with its inverse also
preserving the structures.
Theorem 2.1.4. Let F : M → N be a diffeomorphism, and let p ∈ M. Then F∗ : T p (M) → T F(p) (N) is an
isomorphism.
Exercise 2.1.5. Prove this statement, by using the last line in Theorem 2.1.2.
Let us now consider an application of the previous result in the special case N = Rm . Let p ∈ M and let
(U, φ) be a chart, with p ∈ U. The map φ : U 7→ φ(U) can be seen as a smooth map between two smooth
manifolds, which means that the above constructions apply. In particular, the map φ∗ : T p (M) → T φ(p) (Rm ) is
a homomorphism, and the map φ−1 m
∗ : T φ(p) (R ) → T p (M) is also a homomorphism (which means equivalently
that the maps φ∗ and φ−1 m
∗ are isomorphisms). Thus, we an look at the image of the basis of T φ (R ) introduced
in (2.1.2) through the map φ∗ .
−1

Definition 2.1.6 (Coordinate frame). For any coordinate system (U, φ), the family
  ∂ m
φ∗
−1
∂x j φ(p) j=1
determines at any p ∈ U a basis E1,p , . . . , Em,p of T p (M), called the coordinate frame.


15
Let us emphasize once more: The tangent space of M at p is independent of any ambient and of any coordinate
system. On the other hand, the choice of a coordinate system provides a natural basis for the tangent space,
namely the coordinate frame. By looking carefully at the expressions one has for any f ∈ C ∞ (p) and for
j ∈ {1, . . . , m}
 ∂   ∂( f ◦ φ−1 ) 
E j,p ( f ) = φ−1 = φ(p)


∗ ( f ) (2.1.5)
∂x j φ(p) ∂x j

Exercise 2.1.7. If (U, φ) and (V, ψ) are two coordinate systems at p ∈ M, write the expression for the change
of bases between the two corresponding coordinate frames.

Let us recall that in Definition 1.3.6, a smooth map F : M → N between smooth manifolds was considered,
and the rank of the corresponding Jacobian matrix was introduced. With the tangent spaces defined above,
more explicit formulas can be exhibited, and the rank of F gets a more intrinsic expression. We state the
following result, and leave it as an exercise to master these expressions and to understand its rather easy
proof.

Theorem 2.1.8. Let F be a smooth map between a smooth manifold M of dimension m and a smooth manifold
N of dimension n. Let p ∈ M and (U, φ) and (V, ψ) be two charts of M and N respectively such that p ∈ U
∂ −1 ∂
and F(U) ⊂ V. For j ∈ {1, . . . , m} set E j,p := φ−1
∗ ∂x j φ(p) , and for i ∈ {1, . . . , n} set Ẽ i,F(p) := ψ∗ ∂yi ψ(F(p))



which define the coordinate frames of T p (M) and of T F(p) (N), respectively.

(i) For j ∈ {1, . . . , m} one has

n  ∂ Fi
ψ,φ
X 
F∗ (E j,p ) = Ẽi,F(p) ,
i=1
∂x j φ(p)

with Fψ,φ := ψ ◦ F ◦ φ−1 and Fψ,φ

i the ith -component of this function,


(ii) The rank of F at p is given by the dimension of F∗ T p (M) ,

(iii) F∗ is an isomorphism into its image if and only if the rank of F is equal to m, while it is an isomorphism
onto T F(p) (N) if and only if the rank of F is equal to n.

Figure 2.3: The tangent space T p (N) generated by tangent vectors, from [2, p. 113]

16
Exercise 2.1.9 (r). Understand the previous statement, and prove it, getting inspiration from [2, Thm. IV.1.6
& Corol. IV.1.7].

Let us come back to curves and establish some relations, see Figure 2.3. Consider a smooth manifold N and
a smooth parametric curve c : (−ε, ε) → N with c(0) = p ∈ N. Note that c can be considered as a map
F between the manifold (−ε, ε) and the manifold N. We set C ⊂ N for the curve generated by c, namely
C := c (−ε, ε) and assume that c is a diffeomorphism between (−ε, ε) and C. Then, observe that (C, c−1 ) is a


coordinate system, and by Definition 2.1.6 the vector c∗ dtd 0 is a tangent vector to C at p. By (2.1.5), for any


f ∈ C ∞ (p) one has
d  d( f ◦ c)
c∗ (f) = (0).
dt 0 dt
Observe that this formula is valid for any f ∈ C ∞ (p) (defined only on C) or for f defined on N. Observe also
that this formula can be extended for any t ∈ (−ε, ε), namely c∗ dtd t represents the tangent vector to C at c(t).

Let us now consider a coordinate system (U, φ) of N with C ⊂ U, and let us set c j := (φ ◦ c) j for the local
coordinates of c. For any t ∈ (−ε, ε), meaning that c(t) ∈ U and for any f ∈ C ∞ c(t) one has

d  d( f ◦ φ−1 ◦ φ ◦ c)
c∗ (f) = (t)
dt t dt
d( f ◦ φ−1 ◦ (c1 , c2 , . . . , cn )
= (t)
dt
n
X ∂( f ◦ φ−1 ) 
 j
′
= φ c(t) c (t)
j=1
∂x j
n
X
= c j ′ (t) E j,c(t) ( f ),


j=1

where the third equality corresponds to the derivative of the composition of two functions, and where {E j,c(t) }nj=1
is the coordinate frame at c(t) ∈ N. Thus, a curve determines a tangent vector at any point of its image, namely
at any point c(t). In fact, the converse is also true, as we shall see in Section 2.3.

Exercise 2.1.10 (r). Report on the examples of a parametric curve and of a parametric surface as presented
in [2, p. 112–113].

Figure 2.4: A vector fields on the manifold R2

17
2.2 Vector fields
In this section, we introduce a new type of functions, defined on M and taking values on T p (M) at any p ∈ M,
see Figures 2.4 and 2.5. The difficulty is to consider some regularity condition on such functions, namely

Figure 2.5: 2-torus with vector fields [12]

continuity or differentiability. One solution is to consider the tangent bundle

G
T (M) := T p (M)
p∈M

and to endow it with a topology and a differential structure. A representation of the tangent bundle for a
2-sphere is provided in Figure 2.6. Then a map X : M → T (M) can be defined and some regularity classes

Figure 2.6: A 2-sphere and part of its tangent bundle

can be considered. Another solution is to use local charts and the coordinate frame introduced in the previous
section. We develop this approach below.
Recall that if (U, φ) is a chart with p ∈ U, the coordinate frame {E j,p } M
j=1 has been introduced in Definition
2.1.6. Then, any X p ∈ T p (M) can be expressed as
m
X
Xp = a j (p) E j,p
j=1

with a j (p) ∈ R. Thus, for defining a map X : M → T (M) we can look at the local map

Rm ⊃ φ(U) ∋ x 7→ [a ◦ φ−1 ](x) ∈ Rm (2.2.1)

and impose that this local representation of the function X is smooth. More precisely:

18
Definition 2.2.1 (Vector field). For any r ∈ {0, 1, 2, . . . , ∞}, a vector field of class C r on M is a map X : M →
T (M) whose components a in the coordinate frame of any chart (U, φ) are of class C r . In other words, the
map (2.2.1) is of class C r for any chart (U, φ). The set of all vector fields of class C ∞ is denoted by X(M).

Since vector fields of class C ∞ will be the most common one in the sequel, let us provide an equivalent
definition for them, whose proof is left as an exercise.

Lemma 2.2.2. The map X : M → T (M) belongs to X(M) if and only if for any C ∞ -function f defined on an
open set U of M, the function X f : U → R, defined at p ∈ U by [X f ](p) := X p ( f ), is a smooth function on U.

Let us now study some properties of the set X(M). It is clearly a vector space, with the addition defined by
(X + Y) p = X p + Y p for any X, Y ∈ X(M) and p ∈ M. It is also a module over C ∞ (M), namely if f ∈ C ∞ (M)
and X ∈ X(M), we can set ( f X) p := f (p)X p and check that f X is a new element of X(M). This new operation
consists in a multiplication of X(M) on the left by elements of C ∞ (M). The set X(M) has also the structure of
a Lie algebra, meaning that there exists a Lie bracket [ , ] : X(M) × X(M) → X(M) satisfying

(i) It is linear in each argument,

(ii) It is antisymmetric: [Y,X]=-[X,Y],

(iii) The Jacobi identity holds:

X, [Y, Z] + Y, [Z, X] + Z, [X, Y] = 0.
     

The existence of this additional structure on X(M) is quite surprising. Indeed, if X, Y ∈ X(M) and f ∈ C ∞ (p),
the map f 7→ X p (Y f ) does not define an element of T p (M) in general, and therefore XY does not in general
defines an element of X(M). On the other hand, by defining [X, Y] p ( f ) := X p (Y f ) − Y p (X f ), this verifies the
linearity property and the Leibniz rule. As a consequence, [X, Y] p is an element of T p (M) for each p ∈ M.
The details can be checked in the following exercise:

Exercise 2.2.3 (r). Check that X(M) endowed with the product [X, Y] is a Lie algebra, see [2, p. 152–153].

Exercise 2.2.4. For any X, Y ∈ X(M) and f, g ∈ C ∞ (M), check the following equality:

[ f X, gY] = f g[X, Y] + f (Xg)Y − g(Y f )X.

Exercise 2.2.5. State and study the hairy ball theorem on the sphere S2 . Look for applications of this state-
ment.

2.3 Vector fields and flows

In the last part of Section 2.1 it has been shown that any smooth parametric curve on a smooth manifold
defines a tangent vector at any point of its image. In the next theorem, we provide a stronger converse
statement. Note that such results are very much related to some existence theorems for solutions of ordinary
differential equations.

Exercise 2.3.1 (r). State as precisely as possible any quite general statement about the existence of ordinary
differential equations, and check how they apply in the present section.

Recall that the set of all smooth vector fields on M is denoted by X(M), see Definition 2.2.1.

19
Theorem 2.3.2 (Thm. IV.4.3 of [2]). Let M be a smooth manifold and let X ∈ X(M) be a smooth vector field
on M. Given any p ∈ M there exist ε > 0 and a unique parametric curve c p : (−ε, ε) → M with c p (0) = p
such that
d

(c p )∗ = Xc p (t) for any t ∈ (−ε, ε). (2.3.1)
dt t

Observe that this statement is stronger than the statement at the end of Section 2.1 since the curve provides the
tangent vector not only at t = 0 but for any t ∈ (−ε, ε). Clearly, the value ε > 0 depends on X and on p. The
curve c p is called the integral curve of X centred at p ∈ M or the orbit of p, see Figure 2.7. Relation (2.3.1) is
often denoted by
ċ p (t) = Xc p (t) for any t ∈ (−ε, ε).

Figure 2.7: A manifold, a vector field X and the integral curve of X centred at p

Remark 2.3.3. In Theorem 2.3.2 we only mentioned an interval (−ε, ε). In fact, there exists a maximal open
interval I p ⊂ R with 0 ∈ I p such that the function c p is defined on I p and has the properties mentioned above.
Clearly this interval depends on p, as indicated in the notation. Then, due to the uniqueness, the following
property holds, whenever it is well defined:

cc p (s) (t) = c p (s + t) and in particular cc p (s) (−s) = c p (0) = p. (2.3.2)

Observe also that c p is an immersion (see Definition 1.4.1) as soon as X p , 0, while this orbit reduces to a
single point if X p = 0.
So far, the integral curve constructed above depends only on a single point p ∈ M, can we do better ? Can
we consider the curves related to all points in a neighbourhood of p ? What about the maximal interval I p ?
Can we at least keep an minimal interval (−ε, ε) for all points in a neighbourhood of p ? For answering this
question, we shall firstly define a new set:
[
W := I p × {p} ⊂ R × M
p∈M

which is an open set. Since W is open and contains the subset {(0, p) | p ∈ M}, it follows (why ?) that for each
p, there exist a neighbourhood V of p and δ > 0 such that (−δ, δ) × V ⊂ W. This is the necessary uniformity
mentioned above. We shall also need the following concept:
Definition 2.3.4 (Flow). A flow or a one-parameter group action on M is a C ∞ -map θ : W → M which
satisfies the following conditions:

20
 (i) θ(0, p) = p for any p ∈ M,
(ii) min(Iθ(s,p) ) = min(I p ) − s and max(Iθ(s,p) ) = max(I p ) − s, and for any t satisfying ∈ min(I p ) − s < t <
max(I p ) − s one has
θ t, θ(s, p) = θ(s + t, p).


(2.3.3)

Observe that the function θ is a generalization of the functions c p since the dependence on p is becoming
explicit, namely θ(t, p) = c p (t). Also, (2.3.3) is equivalent to (2.3.2). The flow introduced in the previous
definition corresponds to a maximal flow in the sense that the domain W can not be extended. Quite often,
the flow is considered on a smaller domain O ⊂ W, which means that the domain is smaller, but the main
relation (2.3.3) still holds, wherever it is defined. One then speaks about a local flow. The main result in this
framework is:
Theorem 2.3.5. Let M be a smooth manifold, X ∈ X(M), and p ∈ M. Then there exists V ∈ N p and δ > 0
such that θ : (−δ, δ) × V → M is a local flow satisfying θ(0, q) = q for all q ∈ V and

θ̇(t, q) = Xθ(t,q) for all (t, q) ∈ (−δ, δ) × V. (2.3.4)

This flow is unique, in the sens that any other flow can have a different domain of definition (inside W), but
the two flows coincide on the intersection of their domain.

Figure 2.8: A vector field, the set V and some subsets θ(t, V) for different t, from [2, p. 133]

Figure 2.8 provides an illustration of the previous statement. In the sequel, we still provide some concepts
related to this local flow.
Definition 2.3.6 (Singular point). Let X ∈ X(M) and p ∈ M. The point p is called a singular point of X if
X p = 0, and is called a regular point otherwise.
Note that singular points play an important, and can reveal some properties of the manifold M itself. Some
critical points and the integral curves around them are presented in Figure 2.9.
Exercise 2.3.7 (r). Study the notion of critical points of a vector field and look at applications of this concept.
In the notion of local flow, the existence of δ > 0 and the existence of the flow on (−δ, δ) × V was sufficient. If
this domain can be maximally extended, then the vector field bares a special name:

21
 Figure 2.9: Some singular points and the integral curves around them, from [2, p. 140]

Definition 2.3.8 (Complete vector field). A smooth vector field X on M is complete if the corresponding flow
θ satisfying (2.3.4) is defined on R × M, or equivalently if W = R × M.
A very important result about completeness holds for compact manifold, see Definition 1.1.14 for the notion
of compactness.
Theorem 2.3.9 (Corollary IV.5.6 of [2]). If M is a compact smooth manifold, then any smooth vector field X
on M is complete.
Exercise 2.3.10. Study the examples of manifolds and flows proposed on pages 143–145 of [2].
We end this section with the use of a vector field for defining the notion of differentiation. We firstly consider
X ∈ X(M) and the integral curve c p introduced in Theorem 2.3.2. For f ∈ C ∞ (M) we then set



d f c p (t) f c p (t) − f (p)
(LX f ) p := = lim . (2.3.5)
dt t=0 t→0 t
This expression is also called the Lie derivative of f in the direction provided by X. It turns out that LX f = X f ,
where
M ∋ p 7→ [X f ](p) := X p ( f ) ∈ R
belongs to C ∞ (M), as mentioned in Lemma 2.2.2 and already used in the definition of the Lie bracket.
We shall now define the Lie derivative of a vector field. For that purpose, recall that a flow is a map θ : W ∋
(t, p) 7→ θ(t, p) ∈ M as introduced in Definition 2.3.4. Thus, for each fixed t the map M ∋ p 7→ θ(t, p) ∈ M

22
can be interpreted as a smooth map from M to M, see Definition 1.3.4. Consequently, the map (θt )∗ is a well
defined map between T p (M) and T θt (p) (M) ≡ T θ(t,p) (M), for any p ∈ M and t ∈ I p . In particular, by applying
(θ−t )∗ to an element of T θ(t,p) (M) one gets an element of T θ(−t,θ(t,p)) (M) = T p (M). Thus, for X, Y ∈ X(M) and
if θ is the flow related to X as described in Theorem 2.3.5, then we can set
1 
θ−t ∗ (Yθ(t,p) ) − Y p


(LX Y) p := lim
t→0 t
which represents the derivative of Y in the direction provided by X. In this framework, the following result
holds, whose proof is left as an exercise (see [2, Thm. IV.7.8] for inspiration).
Lemma 2.3.11. For any X, Y ∈ X(M), one has LX (Y) = [X, Y].
Exercise 2.3.12. Prove (and understand) the following equality, for any X, Y, Z ∈ X(M) and f ∈ C ∞ (M):
(i) LX [Y, Z] = [LX Y, Z] + [Y, LX Z],
(ii) LX ◦ LY − LY ◦ LX = L[X,Y] ,
(iii) LX ( f Y) = (LX f )Y + f (LX Y).
Exercise 2.3.13. With the notations introduced so far, provide a meaningful represention of the Lie derivative
of a vector field. You can get inspiration from similar pictures available on internet.

23
Chapter 3

Tensors, tensor fields, differential forms

3.1 Tensors and exterior algebra

Let us start by considering a finite dimensional real vector space V, and let V ∗ denotes its dual space, namely
the set of all linear maps from V to R. It is known that V ∗ is itself a vector space. If the dimension of V is n,
then the dimension of V ∗ is also n. The elements of V ∗ are also called the linear functionals on V.
Exercise 3.1.1. Prove the above statements on V ∗ .
We can now provide a rather simple definition for a tensor.
Definition 3.1.2 (Tensors). A tensor on a real vector space V consists in a multilinear map
ϕ : V × V × · · · × V × V ∗ × V ∗ × · · · × V ∗ → R.
If ϕ is defined on r copies of V and s copies of V ∗ , we say that ϕ is r-times covariant and s-times contravariant,
and we write ϕ ∈ T sr (V) or T r,s (V, V ∗ ).
For example T01 (V) is nothing by V ∗ . In the special case r = 0 or s = 0, we shall also use the shorter notation
T r (V) = T0r (V) and T s (V) = T s0 (V). Here and in the sequel, r and s are always elements of N := {0, 1, 2, . . . }.
Exercise 3.1.3. If V is of dimension n, show that T sr (V) is a vector space of dimension nr+s .
+r2
r
Observe that two tensors can be combined: for j ∈ {1, 2} and ϕ j ∈ T s jj (V), one sets ϕ1 ⊗ ϕ2 ∈ T sr11+s2
(V) by
ϕ1 ⊗ ϕ2 (v1 , . . . , vr1 +r2 , v∗1 , . . . , v∗s1 +s2 ) := ϕ1 (v1 , . . . , vr1 , v∗1 , . . . , v∗s1 ) ϕ2 (vr1 +1 , . . . , vr1 +r2 , v∗s1 +1 , . . . , v∗s1 +s2 ),
(3.1.1)
for any v1 , . . . , vr1 +r2 ∈ V and v∗1 , . . . , v∗s1 +s2 ∈ V ∗ . Clearly, the product ⊗ is not commutative, namely ϕ1 ⊗ ϕ2 ,
ϕ2 ⊗ ϕ1 in general.
Definition 3.1.4 (Symmetric tensor, alternating tensor). For ϕ ∈ T r (V) or for ϕ ∈ T s (V), we say that ϕ is
a symmetric tensor if it is invariant under the permutation of any of its arguments, while it is an alternating
tensor if it changes its sign under the permutation of any of its arguments. The set of symmetric tensors is
denoted by Σr (V) or Σ s (V), while the set of alternating tensors is denoted by r (V) or s (V).
V V

Let us now denote by S k the group of all permutations of k elements. Any element of S k is a bijective map σ
from {1, . . . , k} to {1, . . . , k}, written
σ : (1, . . . , k) 7→ σ(1), . . . , σ(k) .

24
We recall that a transposition consists in a permutation of two elements only, and that the sign sgn(σ) of
a permutation is 1 if the number of transpositions in a permutation is even, while it is −1 is the number of
transpositions in a permutation is odd.
We can now define two operations which are quite standard:
Definition 3.1.5 (Symmetrization, antisymmetrization). For any ϕ ∈ T r (V) the maps S : T r (V) → T r (V)
and A : T r (V) → T r (V) are defined for v1 , . . . , vr by

1 X
[Sϕ](v1 , . . . , vr ) := ϕ(vσ(1) , . . . , vσ(r) )
r! σ∈S
r

and
1 X
[Aϕ](v1 , . . . , vr ) := sgn(σ) ϕ(vσ(1) , . . . , vσ(r) ).
r! σ∈S
r

The notation S stands for symmetrization while the notation A stands for antisymmetization.
The following lemma can then be proved as an exercise:
Lemma 3.1.6. In the above framework one has
(i) S2 = S and A2 = A, which means that these operations are projections
(ii) ST r (V) = Σr (V) while AT r (V) = r (V),
V

(iii) ϕ ∈ Σr (V) if and only if Sϕ = ϕ and ϕ ∈ r (V) if and only if Aϕ = ϕ.

V

Remark 3.1.7. Note that the same definitions apply for T s (V) instead of T r (V), and that the same sym-
metrization and antisymmetrization map can be defined. Lemma 3.1.6 holds also in this context.
Let us now consider two vector spaces V and W and a linear map G : V → W. Then the map G∗ : T r (W) →
T r (V) defined on ϕ ∈ T r (W) and for v1 , . . . , vr by

[G∗ ϕ](v1 , . . . , vr ) := ϕ G(v1 ), . . . , G(vr )

is well-defined an corresponds to a linear map.

Exercise 3.1.8. Check the above statement.
One more object can be studied, the tensor algebra, as introduced in the following exercise:
Exercise 3.1.9. For any finite dimensional real vector space V Set T 0 (V) := R and define

T (V) := T 0 (V) ⊕ T 1 (V) ⊕ T 2 (V) ⊕ T 3 (V) ⊕ . . . (3.1.2)

endowed with the componentwise addition and with a multiplication extending the ⊗-operation (3.1.1) by
linearity. Check that these operations make the set T (V) an associative algebra, called the tensor algebra
over V. This algebra has a unit 1 = 1 ⊕ 0 ⊕ 0 ⊕ . . . , and its dimension is ∞.
Recall that r (V) consists of all alternating tensor with r variables in V. We can then define the subset (V)
V V
of T (V), see (3.1.2), given by
^ ^0 ^1 ^2 ^3
(V) := (V) ⊕ (V) ⊕ (V) ⊕ (V) ⊕ . . . .

25
 V
One easily observes that (V) is a subspace of T (V) but it is not a subalgebra (why ?). We shall then define
a new product: for ϕ ∈ (V) and ψ ∈ s (V) we set
Vr V

(r + s)!
ϕ ∧ ψ := A(ϕ ⊗ ψ) (3.1.3)
r! s!

and call it the wedge product of ϕ and ψ. By construction, ϕ ∧ ψ ∈ r+s (V). The main properties of the wedge
V
product is summarized in the following statement:
Lemma 3.1.10. The wedge product is bilinear and associative. When endowed with the wedge product and
V
extended by linearity, the set (V) becomes an associative algebra.
V
Definition 3.1.11 (Exterior algebra). The set (V) endowed with the wedge product ∧ is called the exterior
algebra or the Grassmann algebra over V.
We list a few additional property:
Lemma 3.1.12. For ϕ ∈ r (V) and ψ ∈ s (V) one has
V V

ψ ∧ ϕ = (−1)rs ϕ ∧ ψ.

Theorem 3.1.13. Assume that V is of dimension n, then

(i) If r > n, then r (V) = {0},
V

(ii) If 0 ≤ r ≤ n, then the dimension of r (V) is ( nr ) = n!

V
r!(n−r)! ,

(iii) The dimension of (V) is 2n .

V

Figure 3.1: An example of an exterior algebra, with the notation ΛV instead of (V), and Λr V instead of
V
Vr
(V). The vector space V is of dimension 2

In Figure 3.1 a representation of the exterior algebra for a vector space V of dimension 2 is provided. Finally,
we also have:

26
Lemma 3.1.14. If G : V → W is a linear map between vector spaces, then G∗ (ϕ ⊗ ψ) = G∗ (ϕ) ⊗ G∗ (ψ), and
by extension G∗ : T (W) → T (V) maps (W) → (V) and defines a homomorphism of these algebras (with
V V
the corresponding product).
Exercise 3.1.15. Prove some of the above lemmas or prove Theorem 3.1.13.

3.2 Dual basis and coordinate coframe

Let us now consider V endowed with a basis {E1 , . . . , En }. There exists then a unique basis of V ∗ , denoted by
{ζ1 , . . . , ζn } satisfying ζ j (Ek ) = δ jk . This basis of V ∗ is called the dual basis. As a consequence, for any v ∈ V
one has v = nj=1 ζ j (v)E j .
P

Exercise 3.2.1. Prove these statements.

Consider now a smooth manifold M and a local coordinate system (U, φ). For any p ∈ U, a basis of the
tangent space T p (M) is given by the coordinate frame introduced in Definition 2.1.6 with its elements given
for j ∈ {1, . . . , m} by
−1 ∂
 
E j,p := φ∗ .
∂x j φ(p)
Thus, if we consider the dual space T p (M)∗ of T p (M), simply denoted by T p∗ (M), there exists a dual basis for
this space, which is denoted by (dx j ) p mj=1 .


Let us “justify” this notation for the dual basis. For any f ∈ C ∞ (p) and for X p ∈ T p (M) let us set


d f p (X p ) := X p ( f ) ∈ R, (3.2.1)
inspired by the expression in (2.3.5), and observe in particular that
  ∂    ∂  ∂( f ◦ φ−1 )
d f p (E j,p ) = (d f ) p φ−1 = φ −1
= φ(p) .



∗ ∗ ( f ) (3.2.2)
∂x j φ(p) ∂x j φ(p) ∂x j
Thus, if we choose f = φi ∈ C ∞ (p) (component i of φ) then one has
∂ (φ ◦ φ−1 )i


dφ p (E j,p ) =
i

φ(p) = δi j .


(3.2.3)
∂x j

Since d f p , and in particular dφi p are linear functionals on T p (M), they belong to T p∗ (M). By the relation

(3.2.3) one then infers that dφi p m i=1 defines a basis of T p (M). By using the notation x instead of φ , as
∗ i i


 i
m
already introduced in (1.2.1) one gets that dx p i=1 corresponds to the dual basis of coordinate frame. This
new basis is called the coordinate coframe.
In the case M = Rm , we can choose (U, φ) = (Rm , id). Then for any f ∈ C ∞ (Rm ) and any p ∈ Rm we can
then decompose d f p on the basis dxi p m



i=1 , and the coefficients of this decomposition are then given by

∂f
d f p (Ei,p ) computed in (3.2.2). For shortness, we shall simply write ∂x i (p) for these coefficients, leading to
the equality
m m
X ∂( f ◦ φ−1 ) X ∂f
df p = φ(p) (p) dxi p .


i


dx ≡
i=1
∂x i p
i=1
∂x i

By eliminating the reference to p, one gets the formula

m
X ∂f i
df = dx
i=1
∂xi

27
which is called the differentialDifferential of f . Thus, at every point p of Rm , the expression d f


p is a linear
map, and more generally d f is a field of linear maps, as introduced below.
Exercise 3.2.2. In the special case M = Rm and f ∈ C ∞ (Rm ), explain the expression d f .

3.3 Tensor fields

Recall that a vector field X is a map from M to T (M) associating to any p ∈ M an element of T p (M). We shall
now generalize this notion.
Definition 3.3.1. For r, s ∈ N, a (r, s)-tensor field on M is a map

G r
ϕ : M ∋ p 7→ ϕ p ∈ T sr T p (M) ⊂ T s T p (M) .


p∈M

Figure 3.2: Pictorial representation of a tensor field

For example, a vector field corresponds to a (0, 1)-tensor field since T10 (V) is made of all linear maps from
V ∗ to R, which can be identified with V itself, or in more precise words (V ∗ )∗  V. Reciprocally, a (1, 0)-
tensor field is a map from M to T ∗ (M) := p∈M T p∗ (M) which is called a covector field. For f ∈ C ∞ (M), the
F
differential d f of f introduced in the previous section corresponds to a covector field. Finally, a (2, 0)-tensor
field is called a field of bilinear maps since it associates to any p ∈ M a bilinear map ϕ p : T p (M)×T p (M) → R.
There are several equivalent ways to impose smoothness to these tensor fields. For example, for a (r, s)-tensor
field, one can impose that for any smooth vector fields X1 , . . . , Xr and any smooth covector fields Y1 , . . . , Y s
the map
M ∋ p 7→ ϕ p X1,p , . . . , Xr,p , Y1,p , . . . , Y s,p ∈ R

is smooth. Alternatively, if (U, ϕ) is a local coordinate system we can expand ϕ p on a suitable basis of
T sr T p (M) expressed in terms of the coordinate frame {E j,p }mj=1 and of the coordinate coframe {(dx j ) p }mj=1 ,


and impose that these coefficients are smooth maps on M (as functions of p ∈ M). The resulting tensors fields
are called smooth tensor fields or C ∞ -tensor field.

28
Exercise 3.3.2. Write the details of the expansion of a (r, s)-tensor field on the basis indicated above.
By gathering the information obtained so far, we can define smooth tensor fields and list their main properties.
Definition 3.3.3 (Smooth tensor fields). For r, s ∈ N the set of all smooth (r, s)-tensor fields on M is denoted
by Trs (M).
Lemma 3.3.4. In the framework introduced before,
(i) Trs (M) is a vector space,
(ii) Tr0 (M) is a C ∞ (M)-module, with for f ∈ C ∞ (M)

ϕ(X1 , . . . , f Xq , . . . , Xr ) = f ϕ(X1 , . . . , Xq , . . . , Xr ),

(iii) If ϕi ∈ Trsii (M) for i ∈ {1, 2}, then ϕ1 ⊗ ϕ2 ∈ Trs11+r 2

+s2 (M), with a clear pointwise definition.

The proof of this lemma is left as an easy exercise. Similarly, the following properties can be proved in an
exercise:
Exercise 3.3.5. Let F : M → N be a smooth map between smooth manifolds, and let r ∈ N. In addition to the
map F ∗ and F∗ introduced in (2.1.3) and (2.1.4), let us set F ∗ acting on ϕ ∈ Tr0 (N) and for X1,p , . . . , Xr,p ∈
T p (M) as
[F ∗ (ϕ)] p X1,p , . . . , Xr,p := ϕF(p) F∗ (X1,p ), . . . , F∗ (Xr,p ) .

Show that F ∗ maps Tr0 (N) to Tr0 (M).

After smoothness, the next notion to be introduced on tensor fields is the notion of symmetric or alternating
tensor fields.
Definition 3.3.6 (Symmetric tensor field, alternating tensor field). For r ∈ N, a tensor field ϕ ∈ Tr (M) is
symmetric if ϕ p ∈ Σr T p (M) for any p ∈ M, and is alternating or skew symmetric if ϕ p ∈ r T p (M)

V

for any p ∈ M. The corresponding set of all smooth symmetric tensor field is denoted by Σr (M) while the
corresponding set of alternating tensor field is denoted by r (M).
V

For r = 2, a special case of a symmetric tensor field is going to play a very special role in the sequel. A
symmetric field ϕ of bilinear maps, i.e. ϕ ∈ Σ2 (M), is positive definite if ϕ p (X p , X p ) ≥ 0 for any p ∈ M and
any X p ∈ T p (M), with an equality if and only if X p = 0.
Definition 3.3.7 (Riemannian manifold). A manifold M endowed with a symmetric and positive definite field
ϕ of bilinear maps is called a Riemannian manifold, and ϕ is the Riemannian metric.
These manifolds are very convenient and will be further studied later, since they possess a notions of length
and of angles.

3.4 Differential forms and exterior derivative

In this section, we further study the set of alternating (or skew symmetric, or antisymmetric) tensor fields, as
introduced in Definition 3.3.6.
Definition 3.4.1 (Exterior differential forms, exterior algebra). Let M be a smooth manifold of dimension m.
For any r ∈ N a alternating tensor field ϕ ∈ r (M) is called an exterior differential formExterior differential
V

29
form of degree r, or simply a r-form. The set
m ^
^ M r
(M) := (M),
r=0

with 0 (M) := C ∞ (M), is called the algebra of differential forms or the exterior algebra, when endowed with
V
the componentwise addition and the wedge product (3.4.1).
Note it is natural to stop the summation at r = m, since r (M) = {0} for r > m, by Theorem 3.1.13. The main
V
Vr V
properties of (M) and (M) are gathered in the following statement, whose proof can be provided as an
exercise. For its statement, recall that the basis {(dx j ) p }mj=1 of T p∗ (M) (called dual basis) has been introduced
in Section 3.2.
Proposition 3.4.2. In the setting introduced above:
(i) If ϕ ∈ r (M) and ψ ∈ s (M), then ϕ ∧ ψ ∈ r+s (M), with for any p ∈ M,
V V V

(ϕ ∧ ψ) p = ϕ p ∧ ψ p (3.4.1)

and where the wedge product has been introduced in (3.1.3). Also ψ ∧ ϕ = (−1)rs ϕ ∧ ψ,
V
(ii) (M) is an algebra with the componentwise addition and the product defined above,
(iii) If (U, φ) is a chart on M and p ∈ U, then the set
n o
(dxi1 ) p ∧ · · · ∧ (dxir ) p
1≤i1 <i2 <···<ir ≤m
Vr
T p (M) , and accordingly the family dxi1 ∧· · ·∧dxir 1≤i1 <i2 <···<ir ≤m is a basis for r (U).

 V
is a basis for
Exercise 3.4.3. In the framework of Exercise 3.3.5, show that F ∗ maps r (N) into r (M).
V V

Let us now present the main result of this section, see [2, Thm. V.8.1].
V
Theorem 3.4.4. Let M be a smooth manifold, and let (M) be the exterior algebra on M. Then there exists
V V
a unique linear map d : (M) → (M) satisfying:
(i) If f ∈ 0 (M) ≡ C ∞ (M), then d f = d f ∈ 1 (M), with the r.h.s introduced in (3.2.1),
V V

(ii) If ϕ ∈ r (M) and ψ ∈ s (M), then

V V

d(ϕ ∧ ψ) = (dϕ) ∧ ψ + (−1)r ϕ ∧ (dψ),

(iii) d2 = d ◦ d = 0.
An explicit formula for the map d can be obtained in local coordinates. Namely, let (U, φ) be a chart and let
p ∈ U, let {E j,p }mj=1 be the coordinate frame, and let {(dx j ) p }mj=1 be the coordinate coframe. Let ϕ ∈ r (U) be
V
locally represented as X
ϕp = ai1 ,...,ir (p) (dxi1 ) p ∧ · · · ∧ (dxir ) p
1≤i1 <i2 <···<ir ≤m
with ai1 ,...,ir : U → R smooth maps. Then (dϕ) p is locally represented by
X
(dϕ) p = (dai1 ,...,ir ) p ∧ (dxi1 ) p ∧ · · · ∧ (dxir ) p (3.4.2)
1≤i1 <i2 <···<ir ≤m
V1
with dai1 ,...,ir ∈ (U).

30
Exercise 3.4.5. By taking (3.4.2) as a definition for dϕ, check that the map d satisfies the three properties
stated in Theorem 3.4.4.

Let us add two comments on the map d, which is called the exterior derivative. Clearly, this map is local,
meaning that if ϕ ∈ (U), then dϕ ∈ (U) for any open subset U of M. One also gets that if ϕ ∈ r (U), then
V V V
dϕ ∈ r+1 (U).
V

For the next exercise, observe that if ϕ ∈ 1 (M) and if X ∈ X(M), then ϕ(X) ∈ C ∞ (M), since the map
V
p 7→ ϕ(X) p = ϕ p (X p ) is smooth.

Exercise 3.4.6. For ϕ ∈

V1
(M) and if X, Y ∈ X(M), show that

[dϕ](X, Y) = X ϕ(Y) − Y ϕ(X) − ϕ([X, Y]).

The equality can be shown in local coordinates.

Let us now generalize the content of the previous exercise.

Proposition 3.4.7. Let ϕ ∈ r (M) and let X1 , . . . , Xr+1 ∈ X(M). Then the following equality holds:
V

r+1
X
[dϕ](X1 , . . . , Xr+1 ) = (−1)i−1 Xi ϕ(X1 , . . . , X̌i , . . . , Xr+1 )


i=1
X
+ (−1)i+ j ϕ([Xi , X j ], X1 , . . . , X̌i , . . . , X̌ j , . . . , Xr+1 ),
i< j

where X̌i means that this variable vector field is omitted, and where ϕ(Y1 , . . . , Yr ) corresponds to an element
of C ∞ (M).

Exercise 3.4.8. Prove the equality of the previous statement.

The next exercise is a continuation of Exercise 3.4.3.

Exercise 3.4.9. Let F : M → N be a smooth map between smooth manifolds, and let r ∈ N. Show that F ∗
and d commute, namely F ∗ ◦ d = d ◦ F ∗ .

Exercise 3.4.10 (r). Define the De Rham cohomology of a smooth manifold, and present a few results in this
direction, see for example [2, Sec. VI.7] and [4, p. 76], and also Figure 3.3.

Figure 3.3: Main idea of De Rham cohomology

31
3.5 Orientation of a manifold
Let us consider a vector space V of dimension n, and let {E j }nj=1 and {F j }nj=1 be two bases of V. Set A = (ai j ) ∈
Mn (R) by Fi = nj=1 ai j E j . The two bases {E j }nj=1 and {F j }nj=1 of V have the same orientation if det(A) > 0,
P
while they have the opposite orientation if det(A) < 0. Clearly, the set of all bases of V can be divided into two
classes, depending on their respective orientation. One class is usually said to contain the positively oriented
bases while the other class contain the negatively oriented bases. Clearly, this classification is relative and
depends on the choice of a distinguished basis.

Figure 3.4: The Moebius strip, a non-orientable manifold

Definition 3.5.1 (Orientable manifold). A smooth manifold is orientable if there exists a maximal atlas (Uα , φα )α
such that the Jacobian determinant of any transition function φα ◦φ−1
β : φβ (U α ∩U β ) → φα (U α ∩U β ) is positive.

Figure 3.5: The Klein bottle, a non-orientable manifold

Note that any transition function satisfying the above condition is also said to be orientation preserving,
since it maps a basis with one orientation in a basis with the same orientation. Two standard examples
of non-orientable manifolds are presented in Figures 3.4 and 3.5. The following statement is then a direct
consequence of the definition.
Lemma 3.5.2. Any connected orientable manifold of dimension n ≥ 1 has only two possible orientations.
In the special case of a manifold M = {p} of dimension 0, one can define an orientation as a map from p to
{−1, +1}.
The next statement plays a very important role for the notion of orientable manifold. For its statement, recall
that for a manifold of dimension m, the dimension of m T p (M) is equal to 1, for all p ∈ M.
V

32
Theorem 3.5.3. A manifold M of dimension m is orientable if and only if there exists ϕ ∈ (M) with ϕ p , 0
Vm
for all p ∈ M.
Exercise 3.5.4. Prove the previous statement, getting inspiration from pages 215–218 of [2].
Let us look at the above concept for smooth manifolds with boundary. For that purpose, consider any subset
Ω ⊂ Rm and a map f : Ω → Rn . Then f is said to be differentiable on Ω if there exist an open set U ⊂ Rm ,
with Ω ⊂ U, and a differentiable function f˜ : U → Rn such that f˜ Ω = f . With this definition at hand, a
smooth manifold with boundary is a topological manifold with boundary endowed with a maximal atlas for
which all transitions maps are smooth. Note that some of these transitions maps are defined on open subsets
of Hm instead of on open subsets of Rm . The boundary ∂M of a manifold with boundary is given by
[
∂M := φ−1
α φα (U α ) ∩ ∂H .
m

Figure 3.6: A smooth manifold with boundary, from [13]

Figure 3.6 represents a manifold with its boundary, with the spaces R2 and H2 clearly represented. A first
result about the boundary of a smooth manifold with boundary is provided in [4, Prop. I.9.2] :
Proposition 3.5.5. The boundary of a smooth manifold of dimension m is a smooth manifold of dimension
m−1
The final result, very useful in applications, is about orientability of a manifold with boundary. It reads:
Proposition 3.5.6. If M is a smooth orientable manifold with boudary, then its boundary ∂M is also ori-
entable.
We refer to [4, Prop. I.9.4] and its proof for a construction of an induced basis on ∂M from a basis on M. The
notion of outward pointing vector is also introduced (recall that any orientable manifold has only two possible
orientations).
Exercise 3.5.7. Study the proof of the previous proposition, and explain the notion of induced basis and of
outward pointing vector, see also Figure 3.7.

33
Figure 3.7: Manifold with boundary with outward and inward pointing vectors

34
Chapter 4

Integration on manifolds

4.1 Integration of m-forms

We consider an orientable manifold M of dimension m endowed with a maximal atlas (Uα , φα )α preserving
the orientation. We also consider ϕ ∈ m (M) with ϕ p = 0 if p < Uβ (we say that the support of ϕ is in Uβ ) for
V
some fixed β. We assume that the support supp(ϕ) is a compact subset of M. By the representation formula of
tensor fields, it follows that there exists a ∈ C ∞ (Uβ ) such that for any p ∈ Uβ

ϕ p ≡ ϕ(p) = a(p) (dx1 ) p ∧ · · · ∧ (dxm ) p . (4.1.1)

The function a is well-defined, since m T p (M) is of dimension 1, as mentioned in Theorem 3.1.13. We also
V

recall that φ−1
β : R ⊃ φβ (U β ) → U β ⊂ M generatesVthe pullback
m function (φ−1 )∗ : m (Uβ ) → m φβ (Uβ ) ,
V V

β
which extends trivialy to a pullback function from m (M) → m Rm ), see Exercise 3.4.3. It then follows
V
from (4.1.1) that
β ) (ϕ) = (a ◦ φβ ) dx ∧ · · · ∧ dx ,
(φ−1 ∗ −1 1 m

where {dx j }mj=1 denotes the dual basis of ∂x∂ j mj=1 of Rm . We are now ready to define the integral of ϕ. We shall

say speak about an oriented manifold when one orientation has been chosen for an orientable manifold.
Definition 4.1.1 (Integral of m-forms with support in one chart). Consider M a smooth oriented manifold of
dimension m. For the m-form ϕ defined in (4.1.1), the integral of ϕ is given by
Z Z Z
ϕ= ϕ := a ◦ φ−1
β (x , . . . , x ) dx . . . dx ,

1 m 1 m
(4.1.2)
M Uβ φβ (Uβ )

where the r.h.s. is a usual volume (or multiple) integral in Rm .

Clearly, this definition makes sense Ronly if the
R resulting integral does not depend on the choice of the chart,
namely if supp(ϕ) ⊂ Uα ∩ Uβ , then U ϕ = U ϕ, with the latter expression given by the r.h.s. of (4.1.2) with
β α
β replaced by α. We refer to [2, p. 237–238] for such a proof.
If the support of ϕ is not contained in a single chart, then a partition of unity has to be used. For that purpose,
let us consider ∪α Vα = M a cover of M by subsets. The cover is locally finite if for any p ∈ M there exists
U ∈ N p such that U ∩ Vα = Ø except for a finite number of α. We can then define the following useful
concept:

35
Definition 4.1.2 (Smooth partition of unity). A collection { fα }α of smooth functions on M is called a smooth
partition of unity of M if the following conditions are satisfied:
(i) fα ≥ 0 for all α,
(ii) {supp( fα )}α defines a locally finite cover of M,
(iii) α fα (p) = 1 for any p ∈ M.
P

Figure 4.1: Representation of three functions belonging to a smooth partition of unity for a subset of R

An illustration of (part of) a smooth partition is provided in Figure 4.1. Given an open cover {Vα }α of M,
we say that a smooth partition of unity { fβ }β is subordinated to this cover if for any β there exists Vα with
supp( fβ ) ⊂ Vα . Fortunately, subordinated smooth partitions of unity always exist, as mentioned in the next
statement, see [2, Thm. V.4.4].
Theorem 4.1.3. For any smooth manifold and for any open cover {Vα }α of M, there exists a smooth partition
of unity which is subordinated to it.
In Figure 4.2, a constructive approach for getting a smooth partition of unity is sketched. With this notion of
subordinated smooth partition of unity, the definition of the integral of an arbitrary m-form can be given:
Definition 4.1.4 (Integral of m-forms). Consider M a smooth oriented manifold of dimension m, let (Uα , φα )α
be a maximal atlas preserving the orientation, and let { fβ }β be a smooth partition of unity subordinated to an
open cover of M defined by the atlas. For a m-form ϕ with compact support we define the integral of ϕ on M
by Z Z X XZ
ϕ := fβ ϕ = fβ ϕ (4.1.3)
M M β β Uα

where fβ ϕ ∈
Vm
(Uα ) when supp( fβ ) ⊂ Uα .
In this definition, note that the sum is always locally finite, which justifies the exchanges of the sum and of
the integral. In addition, the expression U fβ ϕ has already been defined in (4.1.2). Let us also list additional
R
α
remarks about this definition.
(i) M ϕ is independent of the choice of the smooth partition of unity,
R

(ii) The map m (M) ∋ ϕ 7→ M ϕ ∈ R is a linear map,

V R

36
 Figure 4.2: Schematic construction of a smooth partition of unity

(iii) If F : M → N is a diffeomorphism, and ϕ ∈ m (N) is compactly supported, then M F ∗ (ϕ) = ± N ϕ,

V R R

with the sign depending on the choice of the orientation of each manifold,
(iv) The previous construction can be extended to m-forms with non-compact support, but one has then to
be careful about the convergence of the integrals.
Let us now come to the main use of the integral defined above. In Calculus II, several results linking integrals
of different natures are introduced, as recalled in Figure 4.3. These relations are called gradient theorem,
Stokes’ theorem, or divergence theorem. In fact, all these results can be encompassed in a single statement,
which is presented below.
Theorem 4.1.5 (Stokes’ theorem). Let M be an oritented m-dimensional manifold with boundary ∂M (en-
dowed with the induced orientation). Let ι : ∂M → M be the inclusion map, and let ϕ ∈ m−1 (M) be a m − 1
V
exterior differential form with compact support. Then the following equality holds:
Z Z
ι∗ (ϕ) = dϕ
∂M M

where ι∗ :
Vm−1 Vm−1
(∂M) was introduced in Exercise 3.4.3, and where dϕ belongs to m (M)
V
(M) →
One easily observes that if M has no boundary, namely if ∂M = Ø, then M dϕ = 0. In addition, if M
R

is a compact manifold, then the support of ϕ is automatically compact. The proof of Stokes’ theorem is
surprisingly not so complicated, and relies essentially on the fundamental theorem of calculus.
Exercise 4.1.6. Study the proof of Stokes’ theorem, as presented for example in [2, p. 260–261] and in [4,
p. 82–84].

37
 Figure 4.3: The formulation of various theorems of Calculus II

So far, we have only integrated m-forms on a m-dimensional manifold (or m − 1-forms on a m − 1-dimensional
manifold). What about smooth functions defined on M ? For this, we need a “measure” on M. Let us
recall from Theorem 3.5.3 that a smooth manifold M of dimension m is orientable if and only if there exists
ϕ ∈ m (M) with ϕ p , 0 for all p ∈ M. This m-form will be used for integrating. More precisely, let us fix
V
such a non-vanishing m-form ϕ, then for any f ∈ C ∞ (M) with compact support we set
Z Z
f := fϕ (4.1.4)
M M

where the r.h.s. is well defined thanks to the following exercise:

Exercise 4.1.7. Complement Lemma 3.3.4 by showing that r (M) is a C ∞ (M)-module. This corresponds to
V
showing that f ϕ ∈ (M) for any f ∈ C ∞ (M) and ϕ ∈ m (M).
Vm V

In particular, when M is compact, we define

Z Z
Vol(M) := 1= 1ϕ
M M

and call it the volume of M. It is clear that the above definitions depend on the choice of ϕ. In practice, we
usually fix such a non-vanishing m-form ϕ and define everything in terms of it.
Exercise 4.1.8 (r). Show that Green’s theorem, the divergence theorem, and the fundamental theorem of
calculus can be seen as special instances of Stokes’ theorem, as presented in Theorem 4.1.5, see for example
[2, p. 262–263]. Some identifications might be required.

38
4.2 Line integrals on manifolds
Let us now define the notion of line integrals. For that purpose, let c : [a, b] → M be a diffeomorphism


between [a, b] and its image C := c [a, b] , both being manifolds with boundaries. The map c corresponds to
a parametric curve, and the set C to the corresponding curve, an immersion of [a, b] in M, see also Definition
1.4.1. If ϕ is a 1-form on M, namely if ϕ ∈ 1 (M), we set
V

Z Z
ϕ := c∗ (ϕ), (4.2.1)
C [a,b]

where c∗ (ϕ) ∈ 1 [a, b] has a representation of the form f dt, for some f ∈ C ∞ [a, b] . The expression (4.2.1)
V


Rb
is called the line integral of ϕ along C and the r.h.s. takes the form a f (t) dt once the mentioned representation
is used.
Let us now mention a few results valid in this framework and which are generalizations of known result from
Calculus I & II.
Lemma 4.2.1. Let ϕ ∈ 0 (M) ≡ C ∞ (M) and let c be a parametric curve as introduced above, then
V

Z
dϕ = ϕ c(b) − ϕ c(a) .



C

Exercise 4.2.2. Show the previous lemma.

Consider now a smooth map H : [0, 1] × [a, b] → M with H(ϵ, a) =: p ∈ M and H(ϵ, b) =: q ∈ M for any
ϵ ∈ [0, 1]. We then set c0 : [a, b] → M and c1 : [a, b] → M with c0 (t) := H(0, t) and c1 (t) := H(1, t) for any
t ∈ [a, b]. In this setting, we say that the two corresponding curves C0 and C1 are homotopic curves from p to
q, see Figure 4.4 and also Figure 4.5.

Figure 4.4: Several homotopic curves, the parameter ϵ indexing the curves

R 4.2.3. In the framework introduced above and for any ϕ ∈ (M) satisfying dϕ = 0, when one has
V1
RTheorem
C
ϕ = C
ϕ.
0 1

39
Clearly, if ϕ = dψ for some ψ ∈ 0 (M), then the statement is a consequence of the previous lemma. However,
V
not all ϕ satisfying dϕ = 0 are of the form ϕ = dψ, which means that the theorem is not a consequence of the
previous lemma in general.
Exercise 4.2.4. Prove the previous statement, getting inspiration for example from [2, Thm.VI.6.6].
As a final remark in this chapter, let me observe that all statement hold with less regularity assumptions.
Smoothness is usually not required for any result related to integrals. However, the current framework has
been developed in a smooth setting, making the previous statements natural with smoothness assumptions.

Figure 4.5: Closed curves which are homotopic (left) and non-homotopic (right)

40
Chapter 5

Riemannian manifolds

5.1 Definition and basic properties

Let us recall from Definition 3.3.7 that a Riemannian manifold is a smooth manifold M endowed with a field
ϕ of symmetric and positive definite bilinear maps, or in other words, endowed with a map ϕ defined on M
and satisfying for any p ∈ M
ϕ p : T p (M) × T p (M) → R
which is linear in each argument, symmetric and verifying ϕ p (X p , X p ) ≥ 0 with the equality if and only if
X p = 0. The map ϕ is called the Riemannian metric.
Remark 5.1.1. On Rm any positive definite matrix A ∈ Mm (R) defines a bilinear map on Rm by the relation

Rm × Rm ∋ (X, Y) 7→ ⟨X, AY⟩ ∈ R

which is positive definite, where ⟨·, ·⟩ denotes the usual scalar product on Rm . Thus, if we identify T p (M) with
Rm , for every p ∈ M, we can think about the Riemannian metric as a function on M associating to any p ∈ M
a positive definite matrix in Mm (R).
Let us immediately state a simple lemma about the transfer of the Riemannian structure. We recall that
immersions have been introduced in Definition 1.4.1, and that the pullback of tensor fields have been studied
in Exercise 3.3.5.
Lemma 5.1.2. Let M, N be smooth manifolds, and let F : M → N be an immersion. If N is endowed with a
Riemannian metric ϕ, then F ∗ ϕ defines a Riemannian metric on M.
Exercise 5.1.3. Prove the above lemma, getting inspiration from [2, p. 185].
As a consequence of this lemma and of the smooth partition of unity introduced in Definition 4.1.2, the
following result can be inferred, see [2, Thm. V.4.5].
Theorem 5.1.4. Any smooth manifold can be endowed with a Riemannian metric.
Let us observe that any Riemannian metric ϕ on M defines an inner product on any tangent space T p (M).
Indeed, for any p ∈ M, the map ϕ p : T p (M) × T p (M) → R is linear in each argument, is symmetric, and is
positive definite. With this inner product, it is then possible to construct an orthonormal basis on each tangent
space by the Gram-Schmidt process. This basis will be denoted by {F1,p , . . . , Fm,p } but observe that there is a
priori no guarantee of any regularity as p moves on M.

41
Thanks to the existence of the orthonormal basis on each tangent space of M, a specific m-form ϕ can be
exhibited.
Theorem 5.1.5. Let M be a smooth Riemannian manifold of dimension m which is orientable. Correspond-
ing to each two orientations, there exists a unique m-form Ω satisfying Ω p F1,p , . . . , Fm,p = 1, whenever


{F1,p , . . . , Fm,p } is a positively oriented orthonormal basis of T p (M).
The m-form Ω is often called the volume form or volume element of the oriented Riemannian manifold M.
The proof of the existence of Ω is constructive and is based on the following ingredients: For any chart (U, φ)
on M, set {E j,p }mj=1 for the coordinate frame associated with p ∈ M, and define

gi j (p) := ϕ p Ei,p , E j,p

(5.1.1)
where ϕ is the Riemannian metric. Then gi j (p) m


i, j=1 defines a matrix, whose determinant is strictly positive.
The volume form Ω in local coordinates is then given by
q √
(φ−1 )∗ Ω := det (gi j ◦ φ−1 )m

1
i, j=1 dx ∧ · · · ∧ dx ≡
m
g dx1 ∧ · · · ∧ dxm

where g stands for det (gi j ◦ φ−1 )m

i, j=1 . In the special case M = R , one simply has Ω = dx ∧ · · · ∧ dx .
m 1 m

R formRΩ can be chosen for the integral of any smooth function f : Ω → R, as indicated
Note that the volume
in (4.1.4), namely M f := M f Ω. This choice provides a canonical way of integrating any f ∈ C ∞ (M).
Exercise 5.1.6 (r). Study Section V.7 of [2] and solve some of the corresponding exercises on [2, p. 219].
Let us now provide another use of the Riemannian metric ϕ for measuring length. Let c : [a, b] → M be a
smooth parametric curve on a Riemannian manifold M, and let C denote the corresponding curve. Recall that
the tangent vector at c(t) is given by
d 
ċ(t) := c∗ ∈ T c(t) (M).
dt t
We can then set:
Definition 5.1.7 (Length of a curve). The length of the curve C defined by c is given by
Z b

1/2
L := ϕc(t) ċ(t), ċ(t) dt.
a

It is then a standard exercise to show that this length does not depend on the parametrization but depends on
the curve C only. The arc length s : [a, b] → [0, L] on C can then be defined by
Z t

1/2
s(t) := ϕc(τ) ċ(τ), ċ(τ) dτ.
a
 2
ds
= ϕc ċ, ċ .


In this framework, the following notation is often used: dt

Let us still define one notion based on the length of a curve. Let p and q be two points on a Riemannian
manifold, and let us set
d(p, q) := inf L | L is the length of a smooth curve between p and q ,

(5.1.2)
where a smooth curve between p and q means any smooth parametric curve c : [a, b] → M with c(a) = p and
c(b) = q. It is rather clear that d(p, q) ≥ 0 and that d(p, q) = d(q, p).
In order to say more about the quantity d(p, q), let us recall the definition of a metric space. A pair (M, d)
composed of a set M and of a map d : M × M → R is called a metric space if d satisfies for any x, y, z ∈ M

42
 (i) d(x, y) ≥ 0,
(ii) d(x, y) = 0 if and only if x = y,
(iii) d(x, y) = d(y, x), (symmetry)
(iv) d(x, z) ≤ d(x, y) + d(y, z). (triangle inequality)
Note that a metric space is a topological space, since the metric induces a topology through the set of open
balls B(x, r) := {y ∈ M | d(x, y) < r} for any x ∈ M and any r > 0. We can now state the main properties of
the function d introduced in (5.1.2) and refer to [2, Thm. V.3.1] for the proof.
Theorem 5.1.8. A connected Riemannian manifold M defines a metric space (M, d) with the map d : M×M →
R given by (5.1.2). The topology defined by the metric d coincides with the initial topology on M.
Note that parametric curve of class C 1 would be sufficient, there is no need to consider only smooth parametric
curves. The resulting metric d would be the same.
We close this section with the notion of isometric Riemannian manifolds. Let M1 and M2 be two Riemannian
manifolds endowed with their respective Riemannian metric ϕ1 and ϕ2 . We say that these two manifolds are
isometric if there exists a smooth map F : M1 → M2 verifying F ∗ ϕ2 = ϕ1 . Note that this requirement is
equivalent to d1 (p, q) = d2 F(p), F(q) , where d1 and d2 are the respective metrics on M1 and M2 given by


(5.1.2). The isometric property plays a very important role in Nash embedding theorem which states that every
Riemannian manifold can be isometrically embedded into some Euclidean space Rn for n large enough.

5.2 The Frenet frame

In the sequel, we want to define the notion of differentiation on a manifold. In this short section, we start by
recalling an important construction in the framework of a curve in R3 . Let us emphasize that the following
developments heavily depend on the ambient space R3 , it is not yet an intrinsic construction. Firstly, in the
fixed canonical coordinate system (or frame) in R3 the motion of an object having a constant speed v ∈ R3 is
defined by the equation
ẋ(t) = v,

where x : R ∋ t 7→ R3 describes the position of the object. It is also possible to describe this evolution by using
a time dependent coordinate system. In particular, we can look at the evolution of this object by attaching the
reference system to it. This approach leads to the so-called Frenet frame and to the Frenet–Serret formulas,
as developed below.
We consider a parametric curve c : [a, b] ∋ t 7→ c(t) ∈ R3 which is assumed to be smooth. The curve is regular
if ċ(t) = dc
dt (t) , 0 for all t ∈ [a, b]. The arc length is then defined by
Z t
s ≡ s(t) := ∥ċ(τ)∥ dτ
a

where ∥ċ(τ)∥ denotes the Euclidean norm of ċ(τ) in R3 . The total length of the curve is given by L :=
Rb
a
∥ċ(τ)∥ dτ. An easy result of Calculus II then states:
Lemma 5.2.1. If c is regular, there exists a diffeomorphism ϕ : [0, L] → [a, b] such that ∥(c ◦ ϕ)′ (s)∥ = 1 for
all s ∈ [0, L].

43
In this framework, we say that the curve is parametrized by its arc length, and in this parametrization the
tangent vector T (s) := (c ◦ ϕ)′ (s) is of length 1. In the sequel, the letter s will be used for the parametrization
of a curve with its arc length.
Let us now observe that
d1 d∥T (s)∥2 d⟨T (s), T (s)⟩
0= = = = ⟨Ṫ (s), T (s)⟩ + ⟨T (s), Ṫ (s)⟩ = 2⟨T (s), Ṫ (s)⟩
ds ds ds
where Ṫ (s) = dT 3
ds (s). The standard notation ⟨·, ·⟩ for the scalar product in R has also been used. The previous
equation means in particular that
Ṫ (s) ⊥ T (s) for all s ∈ [0, L]
since the scalar product is 0. We can then set κ(s) := ∥Ṫ (s)∥ and call it the curvature of the curve. If κ(s) , 0
we can also define uniquely the vector N(s) ∈ R3 of norm 1 by the relation

Ṫ (s) = κ(s) N(s). (5.2.1)

If κ(s) , 0 we also define the vector B(s) ∈ R3 as the unique vector in R3 such that the triple {T (s), N(s), B(s)}
corresponds to a basis in R3 with the positive orientation. In other words, whenever κ(s) , 0 we have defined
an orthonormal basis, which can be thought as a field of orthonormal bases over the interval of time [0, L].
This family of orthonormal bases is called the Frenet frame. Let us emphasize that this frame depends only
on the curve, and does not use the canonical reference system.

Figure 5.1: Representation of the Frenet frame along a curve

Note that if κ(s) , 0 except for a finite number of points, then we can still define the above field of orthonormal
bases by continuity, while if κ(s) = 0 for s in an interval I, then the curve is a straight line on the interval I,
and one can still choose suitably orthonormal bases along the curve. For simplicity, we assume in the sequel
that κ(s) , 0 but the construction can be adapted otherwise.
The derivative of the three vector-valued functions T , N, and B can be expressed on the basis defined by
{T (s), N(s), B(s)} s∈[0,L] . The following equations encode these relations, and are called the Frenet-Serret for-
mulas: 



 Ṫ (s) = κ(s) N(s)
Ṅ(s) = −κ(s) T (s) + τ(s) B(s)

(5.2.2)




 Ḃ(s) = −τ(s) N(s)



44
where τ(s) is called the torsion. This system of equations determines the evolution of the tangent vector T (s),
of the normal vector N(s), and of the binormal vector B(s) along the curve defined by c.
Exercise 5.2.2. Derive the Frenet-Serret formulas. Show that τ(s) = 0 for any s if and only if the curve lies in
a plane, see [2, p. 303–304].
Exercise 5.2.3. Study the dynamics of an object evolving on the curve in terms of the Frenet frame {T, N, B},
see for example [2, p. 304] or other references.

5.3 Differentiation on submanifolds of Rn

In the previous section we introduced a frame adapted to a curve in R3 . This curve can be considered as
a 1-dimensional submanifold of R3 . In this section, we extend the framework by considering submanifolds
of higher dimensions in Rn . These preliminaries will help us understanding the notion of differentiation for
general Riemannian manifolds provided later. Let us again emphasize that the content of this section depends
on an ambient space, the constructions are not intrinsic.
Let us consider a smooth submanifold M of Rn , and consider a smooth vector field Z with Z p ∈ T p (Rn ) for
any p ∈ Rn . In particular, for each p ∈ M we can associate Z p ∈ T p (Rn ). By representing Z in the canonical
frames of T p (Rn ) we have Z p = nj=1 a j (p) ∂x∂ j | p with some a j ∈ C ∞ (Rn ). By restricting a = (a1 , . . . , an ) to
P

M one obtains a vector field along M, see Figure 5.2. If Z p ∈ T p (M) ⊂ T p (Rn ) for each p ∈ M we say that the
vector field Z is tangent to M. In this case, we simply write Z ∈ X(M).

Figure 5.2: The vector field Z along the manifold M

For p ∈ M, observe that T p (Rn ) is endowed with the canonical scalar product, which induces a scalar product
on T p (M), so that M becomes a Riemannian manifold. This allows us to decompose Z p in a unique way as
Z p = Z ′p +Z ′′ ′ ′′ ⊥ ⊥ n
p , with Z p ∈ T p (M) and with Z p ∈ T p (M) ≡ T p (M), the space perpendicular to T p (M) in T p (R ).
Clearly, the second space is called the normal space to M at p. The corresponding orthogonal projections are
denoted respectively by π′ and by π′′ , namely π′ (Z p ) = Z ′p and π′′ (Z p ) = Z ′′
p . These projections are linear
maps from T p (Rn ) onto the tangent subspace and onto the normal subspace at p, see Figure 5.3. With these
notations, a vector field Z is tangent to M if and only if π′ (Z p ) = Z p for any p ∈ M.
In the previous setting, let us also consider a smooth parametric curve c : t 7→ c(t) on M, and a vector field
Y ∈ X(M), i.e. Y is tangent to M. Then the map t → Yc(t) ∈ T c(t) (M) defines a vector field along the curve. By

45
 Figure 5.3: The vector field Z decomposed into two parts

dY
taking its derivative with respect to t, we get another vector field t 7→ dtc(·) (t). In general this new vector field
 dY 
is not tangent to M, but we may consider its projection onto the tangent space T c(t) (M), namely π′ dtc(·) (t) .
Definition 5.3.1 (Covariant derivative). Let Y ∈ X(M), and let c be a smooth parametric curve on M, and let
t 7→ Yc(t) ∈ T c(t) (M) be the corresponding vector field along the curve. Then
DY  dY 
c(·)
(t) := π′ (t) ∈ T c(t) (M)
dt dt
is called the covariant derivative of Y along the curve.

Figure 5.4: Representation of the covariant derivative of a vector field Y tangent to M

A representation of the covariant derivative is provided in Figure 5.4. By assumption and construction, the

46
two vector fields t 7→ Yc(t) and t 7→ DY
dt (t) are vector fields along the curve and tangent to M. However, the way
from which DY dt is obtained depends too much on the ambient space. Our final aim is to define the derivative of
a vector field along a curve on a Riemannian manifold which does not use any embedding of M into a larger
ambient space. Note also that for the above construction, we don’t need Y to be defined on all of M, it is
enough to consider a map with values in T c(·) (M). In the sequel, we shall just denote
 by t 7→ Y(t) ∈ T c(t) (M)
dY(·)
such vector field along the curve defined by c, and keep the notation DY
dt (t) := π′ dt (t) ∈ T c(t) (M).

Let us now mention some easy properties of the covariant derivative defined above.
Lemma 5.3.2. For i ∈ {1, 2} let t 7→ Yi (t) ∈ T c(t) (M) be vector field along a smooth parametric curve, and let
f ∈ C ∞ (R). Then the following equalities hold:
D(Y1 +Y2 )
(i) dt (t) = DY1
dt (t) + DY2
dt (t),
D( f Y1 ) df
(ii) dt (t) = dt (t) Y1 (t) + f (t) DY 1
dt (t),
d⟨Y1 ,Y2 ⟩
(iii) dt (t) = DY1
dt (t), Y2 (t) + Y1 (t), DY 2
dt (t) ,

where the scalar product corresponds to the one inherited from Rn .

Exercise 5.3.3. Prove the previous lemma.
Before moving to the abstract framework for the covariant derivative without an ambient space, let us mention
a key result. Assume that X p is a tangent vector to M at p ∈ M, and let c : (−ε, ε) → M be any smooth
parametric curve satisfying c(0) = p and ċ(0) = X p . Let also Y be a vector field tangent to M. Then, it turns
out that the covariant derivative DYdt (0) does not depend on c but on the tangent vector X p only. Since X p and
DY
dt (0) belong to T p (M), one gets a map

DY
T p (M) ∋ X p 7→ (0) ∈ T p (M). (5.3.1)
dt
We shall write
DY
(0) ∇X p Y := (5.3.2)
dt
computed on any smooth parametric curve c : (−ε, ε) → M satisfying c(0) = p and ċ(0) = X p . If we consider
the integral curve c : (−ε, ε) → M satisfying ċ(t) = Xc(t) , then the previous equation reads
DY
∇ dc = (5.3.3)
dt dt
which can be evaluated at any t ∈ (−ε, ε).
Exercise 5.3.4. Show that DY dt (0) does not depend on the curve satisfying c(0) = p and ċ(0) = X p , but depends
only on X p , see also [2, p. 312–315].
The following statement summarizes the various properties of the covariant derivative introduced above. Note
that if Y1 , Y2 ∈ X(M), we set ⟨Y1 , Y2 ⟩ for their inner product, namely for the smooth function on M defined by

⟨Y1 , Y2 ⟩ p := ⟨(Y1 ) p , (Y2 ) p ⟩.

For any p ∈ M the map

T p (M) ∋ X p 7→ ∇X p Y ∈ T p (M)
is defined by (5.3.1).

47
Theorem 5.3.5. Let M be a smooth submanifold in Rn . Then the following properties are satisfied:
(i) The map given by
T p (M) × X(M) ∋ (X p , Y) 7→ ∇X p Y ∈ T p (M)
is linear in both arguments. In addition, if f ∈ C ∞ (p) then one has

∇X p ( f Y) = f (p)∇X p Y + (X p f )Y p ,

(ii) If X, Y ∈ X(M), then ∇X Y defined by (∇X Y) p := ∇X p Y is a smooth vector field tangent to M,

(iii) If X, Y ∈ X(M), then one has [X, Y] = ∇X Y − ∇Y X, where [X, Y] was introduced in Section 2.2,
(iv) If Y1 , Y2 ∈ X(M) then one has X⟨Y1 , Y2 ⟩ = ⟨∇X Y1 , Y2 ⟩ + ⟨Y1 , ∇X Y2 ⟩.
We refer to [2, Thm. VII.2.11] for the proof of the above statement. Unfortunately, some concepts used in the
proof have not been introduced yet.
Exercise 5.3.6. For X, Y ∈ X(M) and for f ∈ C ∞ (M), show that

∇ f X Y = f (∇X Y), (5.3.4)

and show that

∇X ( f Y) = f ∇X Y + (X f )Y. (5.3.5)

5.4 Differentiation on Riemannian manifolds

Our aim in this section is to construct objects similar to the previous sections, but without having the manifold
embedded in a larger ambient space. The first definition is an abstract version of the operator ∇ which appeared
in (5.3.2) and (5.3.3). Properties (5.3.4) and (5.3.5) are the ones which have to be conserved in the abstract
framework.
Definition 5.4.1 (Affine connection). Let M be a Riemannian manifold. An affine connection ∇ on M corre-
sponds to a bilinear map ∇ : X(M) × X(M) → X(M), written ∇(X, Y) ≡ ∇X Y, satisfying for any f ∈ C ∞ (M)
and for X, Y ∈ X(M)
(i) ∇ f X Y = f (∇X Y),
(ii) ∇X ( f Y) = f ∇X Y + (X f )Y.
In order to check that the previous definition is meaningful, we recall that X(M) is a module over C ∞ (M), and
that X f defines an element of C ∞ (M), see Section 2.2. With these two facts, both sides of the equalities in (i)
and (ii) are well defined. Observe also that the affine connection is C ∞ (M)-linear in its first argument (the X
argument), but is only R-linear in the second argument (the Y argument), since for a constant function f one
has X f = 0.
So far, we don’t know if affine connections exist for Riemannian manifolds, but at least they exist for subman-
ifolds of Rn , as seen in the previous section. We can also wonder if the properties exhibited in (iii) and (iv) of
Theorem 5.3.5 also hold in this more general framework ?


Before answering these questions, let us introduce two more objects. We denote by End X(M) the set of all
maps from X(M) to itself which respect the properties of X(M), and call such a map an endomorphism of
X(M).

48
Definition 5.4.2 (Torsion and curvature). For a Riemannian manifold M and an affine connection ∇ one sets
for any X, Y ∈ X(M)
T (X, Y) := ∇X Y − ∇Y X − [X, Y] ∈ X(M)
and


R(X, Y) := ∇X ∇Y − ∇Y ∇X − ∇[X,Y] ∈ End X(M)
the former being called the torsion while the latter is called the curvature of the connection.
It is then interesting to observe that the map (X, Y) 7→ T (X, Y) is C ∞ (M)-linear in both arguments, and that the
map (X, Y, Z) 7→ R(X, Y)Z is C ∞ (M)-linear in the three arguments.
Exercise 5.4.3. Show the C ∞ (M)-linearity mentioned above, see also [8, Prop. 6.3].
With the definition of torsion introduced above, we say that a connection is torsion free or symmetric if
T (X, Y) = 0 for any vector field X and Y. In addition, if the connection satisfies the equality

X⟨Y1 , Y2 ⟩ = ⟨∇X Y1 , Y2 ⟩ + ⟨Y1 , ∇X Y2 ⟩, (5.4.1)

we say that the connection is a Riemannian connection or a Levi-Civita connection. Note that we use the
notations ⟨X, Y⟩ for the smooth function defined by M ∋ p 7→ ⟨X, Y⟩ p := ⟨X p , Y p ⟩ ∈ R, and also ⟨X p , Y p ⟩ for
ϕ(X p , Y p ), where ϕ denotes the Riemannian metric. It means that the equality (5.4.1) involves the Riemannian
metric. By the properties exhibited in (iii) and (iv) of Theorem 5.3.5, one infers that the connection introduced
in the previous section for submanifolds of Rn is a Riemannian connection. It turns out that the condition of
torsion free and of compatibility (5.4.1) with the Riemannian metric defines uniquely the connection:
Theorem 5.4.4. For any Riemannian manifold M, there exists a unique Riemannian connection on M.
We refer to [2, Thm. VII.3.3] or [8, Thm. 6.6] for the proof of this statement, which is rather involved. If not
stated otherwise, a Riemannian manifold is always endowed with its corresponding Riemannian connection.
Exercise 5.4.5 (Koszul formula). For X, Y, Z ∈ X(M) set

1 
⟨∇X Y, Z⟩ := X⟨Y, Z⟩ + Y⟨Z, X⟩ − Z⟨X, Y⟩ − ⟨X, [Y, Z]⟩ + ⟨Y, [Z, X]⟩ + ⟨Z, [X, Y]⟩ .
2
Check that the connection defined by this relation is torsion free and satisfies (5.4.1). This equality is called
Koszul formula and shows the existence of a Riemannian connection.
Let us now look at local expressions for the object introduced above. For that purpose, we consider a local
chart and the corresponding coordinate frame {E1,p , . . . , Em,p }, as introduced in Definition 2.1.6.
Lemma 5.4.6. Let M be Riemannian manifold of dimension m, and let ∇ be any affine connection on M. Let
(U, φ) be a chart. Then there exist m3 functions Γij,k : U → R which locally uniquely define ∇ (on U).

The set of functions {Γij,k }m

i, j,k=1 are called the Christoffel symbols.

Proof. Let us set E j for the vector field defined by (E j ) p := E j,p for any p ∈ U. Consider X, Y two vector fields
with support on U, and recall that X = mj=1 b j E j and Y = m k ∞ m
P P
k=1 a E k for two functions a, b ∈ C (U; R ). Let
us finally define
Xm
∇E j Ek =: Γij,k Ei . (5.4.2)
i=1

49
By the properties of the affine connection one infers that
m
X 
∇X Y = ∇Pmj=1 b j E j ak Ek
k=1
m
X
= b j ∇E j ak Ek
j,k=1
X m n o
= b j (E j ak )Ek + ak ∇E j Ek
j,k=1
m 
X m
X 
= Xai + Γij,k b j ak Ei , (5.4.3)
i=1 j,k=1

which means that the connection is entirely defined by the Christoffel symbols. Conversely, if we define ∇X Y
by the r.h.s. of (5.4.3), then one can check that it satisfies the two conditions of a connection, as mentioned in
Definition 5.4.1. □

We now infer an important consequence of the previous proof. With the same notation and from the r.h.s. of
(5.4.3) one deduces that the torsion satisfies
m
X
T (X, Y) = ∇X Y − ∇Y X − [X, Y] = Γij,k − Γik, j ak b j Ei .


i, j,k=1

Since ak and b j are arbitrary, and by the linear independence of the vectors Ei,p for i ∈ {1, . . . , m} it follows
that T (X, Y) = 0 for any X, Y ∈ X(M) if and only if Γij,k = Γik, j , for any i, j, k ∈ {1, . . . , m}. In other words, a
connection is torsion free if and only if the corresponding Christoffel symbols are symmetric in the two lower
indices.
Let us establish one more relation between the Christoffel symbols and the Riemannian metric. Recall that the
matrix element gi j (p) m


have been introduced in (5.1.1), and that the corresponding positive definite matrix

m i, j=1
is denoted by gi j (p) i, j=1 . By convention, the inverse of this matrix is denoted by gi j (p) i, j , with components

simply written gi j (p).

Lemma 5.4.7. Let M be Riemannian manifold of dimension m, let (U, φ) be a chart, and let { Γij,k }m
i, j,k=1 be
the Christoffel symbols for the Riemannian connection. Then Γ j,k satisfies
i

m
1 X iℓ  ∂gℓ j ∂gℓk ∂g jk 
Γij,k = g + − ,
2 ℓ=1 ∂xk ∂x j ∂xℓ

∂gℓ j ∂(gℓ j ◦φ−1 )

φ(p) .


where ∂xk
(p) means ∂xk
∂g
Let us emphasize that the notation ∂xℓkj is very commonly used without any explanation. However, one easily
realizes that an interpretation is necessary, since gℓ j is defined on M while partial derivatives can only be taken
on parts of Rm .
Exercise 5.4.8. Prove the previous lemma, getting inspiration from [4, Thm. 3.3.2].

50
With these new expressions and notations, let us look back at the covariant derivative introduced in Definition
5.3.1. We let c : [a, b] → M be a smooth parametric curve on M and let Y ∈ X(M). For simplicity, we
assume that there exists a chart (U, φ) with the curve contained in U and the support of the vector field also
contained in U. Then for any p ∈ U one has Y p = m k k ∞
P
k=1 b (p)E k,p for b ∈ C (U). Similarly, one has
ċ(t) = j=1 ċ (t)Ek,c(t) with ċ ∈ C ([a, b]). It then follows from (5.4.3) that
Pm j j ∞

m m
DY
X X 
(t) := ∇ċ(t) Y c(t) = ċ(t) bi + Γij,k c(t) b j c(t) ċ k (t) Ei,c(t) .




dt i=1 j,k=1

Note that we have defined DYdt in the previous line, since the former definition was only given for submanifolds
n
of R . Clearly, the two definitions coincide in this context, but the current definition holds without any ambient
space. Observe in addition that

d(bi ◦ c) d bi c(·)
d 

ċ(t) bi = c∗ bi = (t) = (t)
dt t dt dt
which leads to
m m

X  d bi c(·)


DY X
k 
(t) ≡ ∇ċ(t) Y c(t) = (t) + Γ j,k c(t) b c(t) ċ (t) Ei,c(t) .
i


j
dt i=1
dt j,k=1

We observe again on these expressions that it is sufficient for Y to be defined on the curve, it is not necessary
to have it defined outside of the curve. In addition, only the derivative ċ of c matters, as already mentioned in
the previous section.
We now discuss the use of the connection introduced above. First of all, a definition related to any covariant
derivative.
Definition 5.4.9 (Parallel along a curve). Let M be a smooth manifold with a connection ∇, let c : [a, b] → M
be a smooth parametric curve on M, and let Y ∈ X(M). The vector field Y is parallel along the curve C if
∇ċ(t) Y c(t) = 0 for all t ∈ [a, b]

The next statement is based on the existence of solutions for first order differential equations, see Exercise
2.3.1. Its proof can be found for example in [2, Thm. VII.3.12].
Theorem 5.4.10. Let M be a Riemannian manifold endowed with its Riemannian connection and let c :
[a, b] → M be a smooth parametric curve on M. Given X0 ∈ T c(a) (M), there exists a unique vector field
X ≡ Xc(·) defined on the curve C which is parallel along C and satisfies Xc(a) = X0 . If {F1,0 , . . . , Fm,0 } is an
orthonormal basis of T c(a) (M), then there exists a unique field of orthonormal bases {F1,c(·) , . . . , Fm,c(·) } with
{F1,c(·) , . . . , Fm,c(·) } a basis of T c(·) (M) and F j,c(a) = F j,0 for j ∈ {1, . . . , m}, such that each vector of the basis
is parallel along the curve C.
Let us emphasize that in this statement, the curve and the initial condition are given, and one looks for the
transport of the initial condition parallel along the curve. More generally, the notion of parallel transport is
a way of transporting geometrical data along smooth curves on a manifold. In Figure 5.5, the transport of
a vector along a curve is shown. Note that the curve has not to be smooth everywhere, it can have a finite
number of points where it is only continuous. Such a curve is called piecewise smooth.
Let us close this section with the main property of the parallel transport.

51
Figure 5.5: Representation of the parallel transport of a vector along a curve on a sphere. The curve is
continuous and made of three smooth parts

Proposition 5.4.11. The parallel transport on any Riemannian manifold preserves the scalar product. More
precisely, let M be a Riemannian manifold with metric ϕ and endowed with its Riemannian connection, and
let c : [a, b] → M be a smooth curve on M (or a piecewise smooth curve). Let t 7→ X(t) ∈ T c(t) (M) and
t 7→ Y(t) ∈ T c(t) (M) be two vector fields parallel along the curve. Then the maps t 7→ ϕc(t) X(t), Y(t) and

t 7→ ∥X(t)∥ := ϕc(t) X(t), X(t) 1/2 are constant for all t ∈ [a, b].

Exercise 5.4.12. Prove the previous statement, see for example [8, Prop. 14.16].

5.5 Geodesics
In the previous section, the curve was given and the transport was performed along the curve. In this section,
we look for curves which satisfy a certain optimal condition. Recall that giving a connection on a manifold
M of dimension m is equivalent to giving m3 Christoffel symbols {Γij,k } once a local chart is provided. In the
sequel, I denotes an interval in R.
Definition 5.5.1 (Geodesics). Let M be a Riemannian manifold and endowed with a connection ∇. A curve
c : I ∋ t 7→ c(t) ∈ M is a geodesic of the connection ∇ if ċ is parallel along c, or equivalently if it satisfies
dt (t) = 0 for any t ∈ I. In local coordinates, it means that for any i ∈ {1, . . . , m}
Dċ

m
X
c̈ +
i
Γij,k ċ j ċk = 0. (5.5.1)
j,k=1

The equation (5.5.1) is called the geodesic equation. It is a second order differential equation. For such an
equation, given an initial condition p ∈ M and X p ∈ T p (M), there exists a unique curve c : (−ε, ε) ∋ t 7→
c(t) ∈ M with c(0) = p and ċ(0) = X p such that c satisfies (5.5.1). Note that the notion of geodesic depends on
the curve, but also on its parametrization: for two parametrizations of the same curve, one might satisfy the
geodesic equation, while the other might not. In fact, if one of them satisfies the geodesic equation, the other
will also satisfy it if and only if the two parametrizations differ only by a shift, see [2, Lem. VII.5.2]. By using
a rescaling argument the following result can be obtained, see [2, Thm. VII.5.8] :

52
Theorem 5.5.2. Let M be a Riemannian manifold endowed with its Riemannian connection, and let (U, φ) be
a local chart. For any p ∈ U there exist V ∈ N p and ε > 0 such that for any q ∈ V and any Xq ∈ T q (M)
with ∥Xq ∥ < ε, there exists a unique geodesic c(·; q, Xq ) : (−2, 2) → U and satisfying c(0; q, Xq ) = q and
ċ(0; q, Xq ) = Xq . In addition, the map

(−2, 2) × V × Xq ∈ T q (M) | q ∈ V, ∥Xq ∥ < ε ∋ (t, q, Xq ) 7→ c(t; q, Xq ) ∈ U



is C ∞ .
Based on the previous result for q = p and for suitable X p let us set

exp(X p ) := c(1; p, X p ) ∈ U ⊂ M (5.5.2)

and call it the exponential map. More precisely, the point exp(X p ) is the point on the unique geodesic starting
at p and determined by X p and for which the time parameter is equal to 1. A representation of this construction
is provided in Figure 5.6.

Figure 5.6: Representation of the exponential map

By definition of a geodesic, the tangent vector ċ is parallel along the curve itself. As a consequence of
Proposition 5.4.11, one infers that for any geodesic the map t 7→ ∥ċ(t)∥ is constant. In the present situation
one has ∥ċ(t)∥ = ∥ċ(0)∥ = ∥X p ∥ and therefore the length of the geodesic between p and exp(X p ) is given by
Z 1 Z 1 Z 1

1/2
L= ϕc(t) ċ(t), ċ(t) dt = ∥ċ(t)∥ dt = ∥X p ∥ dt = ∥X p ∥.
0 0 0

Still in this framework, observe that the map t 7→ exp(tX p ) is well defined for any t ∈ [0, 1], and that this
map provides a smooth parametrization of the geodesic between p and exp(X p ), see also [2, Lem. VII.6.4] for
the details. By using this parametrization, the following result can be obtained. Note that a star domain of a
vector space is a subset Ω of the vector space with a distinguished point x0 ∈ Ω such that for any x ∈ Ω, all
points x0 + t(x − x0 ) with t ∈ [0, 1] also belong to Ω, see Figure 5.7a.
Theorem 5.5.3 (Normal neighbourhood theorem). Every point p of a Riemannian manifold M has a neigh-
bourhood U which is diffeomorphic under the exponential map to a star domain Ω of the vector space T p (M)
which is a neighbourhood of the distinguished point 0 p of T p (M).

53
 (b) Different tangent vectors at p leading to different
(a) A star domain in R2 geodesics on M

Figure 5.7: For the statement of Theorem 5.5.3

Since M is a Riemannian manifold, an orthonormal basis {F1,p , . . . , Fm,p } can be defined on the tangent space
T p (M). Then any X p ∈ T p (M) admits a unique decomposition X p = mj=1 x j F j,p for some x j ∈ R. Also, ∥X p ∥
P
is given by the Euclidean norm
m
X 1/2
∥x∥ = ∥(x1 , . . . , xm )∥ = (x j )2 .
j=1

By fixing ε > 0 such that Ωε := {X p ∈ T p (M) | ∥X p ∥ < ε} ⊂ Ω, with Ω given by Theorem 5.5.3, one infers
that the map
Xm 
φ : exp Ωε ∋ exp x j F j,p 7→ (x1 , . . . , xm ) ∈ B(0, ε) ⊂ Rm


(5.5.3)
j=1

is a diffeomeorphism. Note that the set exp Ωε , together with the map φ, is called a normal coordinate


neighbourhood of the point p, which is in fact nothing but a specific chart. An illustration of a normal
coordinate neighbourhood is given in Figure 5.8.

The normal coordinate neighbourhood has special features which makes it useful. For example, any point q
in this neighbourhood of p can be joined to p by a unique geodesic. Additional properties can be exhibited in
the following exercise:

Exercise 5.5.4. Describe some properties of the normal coordinate neighbourhood, as presented for example
in [2, Rem. VII.6.8] or here.

54
Figure 5.8: The tangent plane at p for a sphere, and a representation of the normal coordinate neighbourhood

Let us state the relation between geodesics and distance. The original statement and its proof can be found in
[2, Corol. VII.7.5].
Proposition 5.5.5. Let C be a piecewise smooth curve on a Riemannian manifold M, starting at p and ending
at q, and of length equal to d(p, q). If c corresponds to the parametrization of the curve with its arc length,
then c is smooth and a geodesic.
Note that the condition between the length and the distance is essential. For example, in R2 \ {0}, if we choose
p = (1, 0) and q = (−1, 0), one has d(p, q) = 2 but there is no piecewise smooth curve which realizes this
minimal length. If a geodesic from p to q has a length corresponding to the distance d(p, q), then we call this
geodesic a minimal or minimizing geodesic, see also Figure 5.9.

Figure 5.9: Representation of the shortest distance on earth

Let us close this chapter with a (not easy) statement related to geodesics. For any geodesic c : I → M, we say
that c can be extended on the interval I ′ if I ⊂ I ′ and if the new smooth parametric curve c : I ′ → M is again
a geodesic.

55
Theorem 5.5.6 (Hopf and Rinow). Let M be a connected Riemannian manifold. Then the following are
equivalent:
(i) M is geodesically complete, that is for any p in M, the exponential map (5.5.2) is defined on the entire
tangent plane T p (M),
(ii) All geodesics can be extended on the entire line R,
(iii) The bounded and closed subset of M are compact,
(iv) (M, d) is a complete3 metric space.
Furthermore, any one of the above implies that given any two points p, q ∈ M, there exists a minimal geodesic
connecting these two points.
As a direct consequence of the statement (iii) it turns out that any compact manifold is automatically geodesi-
cally complete.

3
meaning that every Cauchy sequence converges.

56
Chapter 6

Curvature

6.1 Geometry of surfaces in R3

In this preliminary section, we briefly skech some results about surfaces in R3 . Details can be provided as
exercises, following the approach of [2, Sec. 8.1 & 8.2].
Let M be a smooth 2-dimensional manifold in R3 , and let (U, φ) be a chart. We assume that Ξ := φ(U) is a
connected open subset of R2 , and we shall consider the map ξ : Ξ → M with ξ := φ−1 as a parametrization
of the manifold M restricted to U. Thus, for any p ∈ U ⊂ R3 there exists (u, v) ∈ Ξ such that  ξ(u,v) = p,
∂ ∂
and more concretely ξ(u, v) = (x1 , x2 , x3 ) ∈ R3 . The coordinate frame is given by E1 = φ−1 ∗ ∂u = ξ∗ ∂u and
∂ ∂
   
E2 = φ−1
∗ ∂v = ξ∗ ∂v . We also assume that M is oriented, and that (U, φ) belongs to the chosen atlas. As a
consequence, for any p ∈ U one can fix a unit normaln vector Nop such that {E1,p , E2,p , N p } defines a basis at
∂ ∂ ∂
p having the same orientation as the canonical basis ∂x , ∂y , ∂z of R3 . Clearly, N p is orthogonal to T p (M),
where the notion of orthogonality is computed according to the standard inner product of R3 .
Let us now consider a parametric curve c : (−ε, ε) → U with c(0) = p and and ċ(0) = X p ∈ T p (M). By
restricting N to the curve, one gets a vector field t 7→ N(t) := Nc(t) along the curve. Then, since N(t) is of
constant norm, one observes that
d⟨N(·), N(·)⟩
0= (t) = 2⟨Ṅ(t), N(t)⟩
dt
meaning that Ṅ(t) belongs to T c(t) (M) for any t ∈ (−ε, ε). As already observed several times, this quantity
depends only on X p and not on the curve itself, see Figure 6.1. More precisely one has:
Theorem 6.1.1. Let c : (−ε, ε) → M be a parametric curve with c(0) = p and and ċ(0) = X p ∈ T p (M). Let
N : (−ε, ε) ∋ t 7→ N(t) ∈ R3 be the unit normal vector field along the curve. Then the vector Ṅ(0) depends
only on X p , and the map
S : T p (M) ∋ X p 7→ S (X p ) := Ṅ(0) ∈ T p (M) (6.1.1)
is a linear map.
We refer to [2, Thm. VIII.1.1] for the proof of this statement. In addition, when R3 is endowed with the usual
inner product, it turns out that the following equality holds for any X p , Y p ∈ T p (M):

⟨S (X p ), Y p ⟩ = ⟨X p , S (Y p )⟩.

57
 Figure 6.1: The normal vector and its derivative

As a consequence, we can consider the map S : T p (M) ∋ X p 7→ S (X p ) ∈ T p (M) as a symmetric map (and
more precisely a self-adjoint or Hermitian operator) on T p (M). Looking at the dependence on p, it turns out
that for smooth vector fields X, Y on M, one can define the map

Ψ : X(M) × X(M) ∋ (X, Y) 7→ Ψ(X, Y) := ⟨S (X), Y⟩ ∈ C ∞ (M)

with ⟨S (X), Y⟩ p = ⟨S (X p ), Y p ⟩. Therefore, Ψ defines an element of Σ2 (M) (a smooth symmetric 2-covariant

tensor field). This tensor field is sometimes called the second fundamental form of the surface M. In that
respect, the first fundamental form corresponds simply to the inner product ⟨X, Y⟩ for X, Y ∈ X(M), See [2,
Sec. VIII.1] for more details.
As a self-adjoint operator on T p (M), which is a vector space of dimension 2, the operator S : X p 7→ S (X p )
possesses two real eigenvalues, which can be called k1 and k2 . These eigenvalues depend on p, obviously.
With no loss of generality, we shall always assume that k1 ≥ k2 . It turns out that k1 and k2 correspond to the
maximal and to the minimal values of Ψ(X p , X p ) = ⟨S (X p ), X p ⟩ over all X p ∈ T p (M) of length 1. Equivalently,
they correspond to the extremal values of the expression Ψ(X p , X p )/⟨X p , X p ⟩ over all X p ∈ T p (M) with X p , 0.
If k1 = k2 , 0 then the corresponding point p is called a umbilical point, while if k1 = k2 = 0 the corresponding
point is called a planar point. Note that for a sphere of radius R > 0, one has k1 = k2 = 1/R for any p, and for
a plane one has k1 = k2 = 0 for any p.
Let us now provide a geometric interpretation of the values k1 and k2 based on Figure 6.2. Consider a point
p ∈ M and let X p be a unit tangent vector at p. The vector X p and the unit normal vector N p determine a plane
passing through p and intersecting M along a curve, at least close to p. This plane is called the normal section
at p and determined by X p . Clearly, the vector N p is normal to the curve at p, and the vector X p is the unit
vector tangent to the curve at p. By parametrizing the curve with its arc-length, and denoting it by c, we get
c(0) = p and ċ(0) = X p . By considering the map ⟨N(t), ċ(t)⟩ = 0 for all t, we obtain by differentiation that

Ṅ(t), ċ(t)⟩ + ⟨N(t), c̈(t)⟩ = 0,

leading to Ṅ(t), ċ(t)⟩ = −⟨N(t), c̈(t)⟩ = −κ(t), where κ(t) is the curvature introduced in (5.2.1) in the frame-
work of the Frenet frame. Thus, by inserting this formula in the definition of S (X p ) given in (6.1.1) we gets

⟨S (X p ), X p ⟩ = Ṅ(0), ċ(0)⟩ = −κ(0).

By considering all possible X p ∈ T p (M) with ∥X p ∥ = 1 and by calculating the corresponding curvature, its
maximal and its minimal values correspond to the eigenvalues k1 and k2 mentioned above. These extreme

58
values are called the principal curvatures of M at p and the corresponding vectors X p the principal directions
at p.

Figure 6.2: Intersection of the manifold with a plane containing N p

Exercise 6.1.2. Study the concrete examples proposed in [2, p. 372–373].

In the framework introduced above, the product of the two eigenvalues, and their average, have a meaning.
More precisely, let us set
k1 + k2
K := k1 k2 and H := , (6.1.2)
2
which are called respectively the Gausssian curvature and the mean curvature. These quantities can be com-
puted at every point p of the manifold M. The interpretation of these quantities are important, for example the
manifold has a quite different shape at p depending if K > 0 or K < 0. In the special case K = 0, additional
investigations are necessary. Surfaces with 0 mean curvature also play an important role and correspond to
minimal surfaces.

Exercise 6.1.3. Study the notions of Gaussian curvature and mean curvature for 2-dimensional surfaces.

Let us finally state a fundamental result, the Theorema Egregium4 of Gauss. Recall that the notion of isometric
manifolds has been introduced at the end of Section 5.1.

Theorem 6.1.4. Let M, N be two 2-dimensional manifolds in R3 , and assume that F : M → N is a diffeomor-
phism which is also an isometry. Then the Gaussian curvature K is the same at p ∈ M and at F(p) ∈ N.

We refer to [2, Thm. VII.2.4] for a proof of this statement. We also quote the following interpretation of
this theorem provided by Wikipedia: The theorem says that Gaussian curvature can be determined entirely
by measuring angles, distances and their rates on a surface, without reference to the particular manner in
which the surface is embedded in the ambient 3-dimensional Euclidean space. In other words, the Gaussian
curvature of a surface does not change if one bends the surface without stretching it. Thus the Gaussian
curvature is an intrinsic invariant of a surface.
4
Latin for “Remarkable Theorem”.

59
6.2 Various curvatures
Let us now come back to the abstract setting: a smooth manifold endowed with a conection ∇. If the manifold
is a Riemannian manifold, then the Riemannian connection is chosen. Recall that the torsion and the curvature
of a connection have been introduced in Definition 5.4.2. If the connection is torsion free, a specific identity
has to be satisfied, which is called Bianchi identity:
Lemma 6.2.1. Assume that the connection on smooth manifold is torsion free, then the following identity
holds for the curvature: for any X, Y, Z ∈ X(M),

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

where 0 means the 0-vector field.

Exercise 6.2.2. Prove Bianchi identity getting inspiration from [4, Prop. 4.1.3].
In order to get some local formulas, we consider a chart (U, φ) and let {E j }mj=1 be the corresponding vector
fields already mentioned in the proof of Lemma 5.4.6. We then set
m
Ri, j,kℓ Eℓ .
X
R(Ei , E j )Ek =: (6.2.1)
ℓ=1

These coefficients can be expressed in terms of the Christoffel symbols, namely

∂Γℓj,k ∂Γℓi,k m m
Ri, j,kℓ = Γnj,k Γℓi,n − Γni,k Γℓj,n .
X X
− + (6.2.2)
∂xi ∂x j n=1 n=1

∂Γℓj,k ∂(Γℓj,k ◦φ−1 )

φ(p) for any p ∈ U, as already mention in Lemma 5.4.7.


where ∂xi
(p) means ∂xi

Exercise 6.2.3. Show the equality (6.2.2), see for example [4, p. 124–125]. Note that alternative formulas
also exist, because of some symmetries.
If M is a Riemannian manifold, with Riemannian metric ϕ we define the Riemannian curvature tensor field
R ∈ T 4 (M) given on X, Y, Z, W ∈ X(M) by

R(X, Y, Z, W) := ϕ R(X, Y)Z, W .

In local coordinates, this leads to m4 smooth functions on U defined by

m

X
Ri, j,k,ℓ := ϕ R(Ei , E j )Ek , Eℓ = Ri, j,kn gnℓ .
n=1

Some properties of this tensor field can be easily proved:

Lemma 6.2.4. For any Riemannian manifold (M, ϕ) the following equalities hold:
(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).

60
Exercise 6.2.5. Prove the previous Lemma. If necessary, look at [4, Prop. 4.1.4] or at [2, Thm. VIII.3.1].
Write the equalities in terms of the coefficients Ri, j,k,ℓ .
Remark 6.2.6. Let us now clarify a property which is rarely explicitly mentioned. As verified in Exercise 5.4.3,
the map (X, Y, Z) 7→ R(X, Y)Z is C ∞ (M)-linear in the three arguments, and consequently the Riemannian
curvature tensor field R(X, Y, Z, W) is C ∞ (M)-linear in its four arguments. Then it follows that there exists
a unique Riemannian curvature tensor (which is denoted by the same notation) which is R-linear in each
argument and which verifies the identity

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

for any p ∈ M. Observe that the r.h.s. is not defined with the material introduced above, since its expression
depends only on the values of the vector fields at p and not in a neighbourhood of p. Nevertheless, the
existence of this tensor (the Riemannian curvature tensor and not tensor field) is ensured by the C ∞ (M)-
linearity. Note that the same property holds for the map (X, Y, Z) 7→ R(X, Y)Z. We refer to [8, Sec. 7.7 & 7.8]
for additional information and for the proofs of similar statements, and encourage any interested reader to
look closely at this topic.
For any p ∈ M, we now consider a plane section π in T p (M), namely a 2-dimensional subspace π ⊂ T p (M).
We also assume that X p and Y p are two orthogonal unit vectors in T p (M) generating the plane π at p. By using
Remark 6.2.6 it makes sense to define the sectional curvature K(π) by

K(π) := −R(X p , Y p , X p , Y p ) = ϕ p R(X p , Y p )X p , Y p

It turns out that this expression does not depend on the explicit form of X p and Y p , but only on the plane
section. If X p and Y p still generates π but are not orthonormal, then one has to consider
R(X p , Y p , X p , Y p )
K(π) = −
ϕ p (X p , X p ) ϕ p (Y p , Y p ) − ϕ p (X p , Y p )2
which again is independent of the choice of the two vectors (as long as they generate π).
It turns out that the knowledge of the sectional curvatures leads to the full knowledge of the Riemannian
curvature tensor, as shown in [4, Prop. 4.1.6], see also [2, Thm. VIII.3.5].
Theorem 6.2.7. The Riemannian curvature tensor at p is uniquely determined by the values of all sectional
curvatures of plane sections of T p (M).
A Riemannian manifold M is said to be isotropic at p ∈ M if the sections curvature K(π) at p is independent
of the choice of the plane section π at p. A manifold is isotropic if it is isotropic at every point. Finally, an
isotropic manifold is said to be of constant curvature if K(π) takes the same value at all point p ∈ M. Note
that any 2-dimensional manifolds is isotropic at every points. In addition, the Gaussian curvature introduced
(6.1.2) for 2-dimensional surface in R3 corresponds to the sectional curvature for these manifolds, see [8,
Thm. 8.3].
Exercise 6.2.8. Study the notion of constant curvature and the corresponding geometry.
Let us introduce one more curvature, the Ricci curvature: for p ∈ M and any orthonormal basis {F1,p , . . . , Fm,p }
of T p (M), for any X p , Y p ∈ T p (M) we set
m
X m
X
Ricp (X p , Y p ) := R(F j,p , X p , Y p , F j,p ) = ϕ p R(F j,p , X p )Y p , F j,p .


j=1 j=1

61
Exercise 6.2.9. Show that this tensor is symmetric and independent of the chosen orthonormal bases of T p (M).
For X, Y ∈ X(M) it is then natural to define the symmetric tensor field Ric by Ric(X, Y) p := Ricp (X p , Y p ). In
local coordinates, one observes that
m
X
Ri j := Ric(Ei , E j ) = Rk,i, j k . (6.2.3)
k=1

This operation is called the contraction of a tensor. If we further contract the Ricci curvature one gets the
scalar curvature
Xm
R(p) := Ricp (F j,p , F j,p ). (6.2.4)
j=1

Exercise 6.2.10 (r). Study Gauss-Bonnet theorem, starting with manifolds without boundary and then for
manifolds with boundary.

6.3 Cartan structure equations

In this section we reformulate some of the results mentioned before in terms of differential forms. We firstly
consider a local smooth field of frames {X1 , . . . , Xm }, meaning that {X1,p , . . . , Xm,p } is a basis of T p (U) for some
open set U ⊂ M and that the dependence in p for each X j,p is smooth, as introduced in Definition 2.2.1. Given
a chart (U, φ) a possible choice for this field of frame is the coordinate frame, see Definition 2.1.6, but other
choices might also be useful. We then fix a smooth field of coframes, namely m fields of 1-forms {ω1 , . . . , ωm }
such that {(ω1 ) p , . . . , (ωm ) p } is a basis of T p∗ (M) and such that (ω j ) p (Xk,p ) = δ jk . Again the coordinate coframe
is a possible choice, but not the only one.
Given a connection ∇ on M, recall that this connections is fully determined by the Christoffel symbols, namely
by the relations
m
X
∇X j Xk = Γij,k Xi .
i=1

for j, k ∈ {1, . . . , m}. Note that the Christoffel symbols depend on the choice of the basis, and therefore these
symbols might be different from the ones introduced in (5.4.2) when the coordinate frame is chosen. For
j, k ∈ {1, . . . , m} we can then define the 1-forms
m
X
ωkj := Γki, j ωi (6.3.1)
i=1

called the connection forms. Indeed, the connection is then fully determined by these m2 1-forms since
Γki, j = ωkj (Xi ). In addition, for any X, Y ∈ X(M) one readily gets that ∇X Xk = mi=1 ωk (X) Xi , and
i
P

m 
X m
X 
∇X Y = Xbi + b j ωij (X) Xi
i=1 j=1

for Y =
Pm jX
j=1 b j. This expression can be compared with (5.4.3).

Exercise 6.3.1. Check the above equalities. Check also that ωij (X) = ωi ∇X X j for any X ∈ X(M).

62
If the Riemannian connection is chosen, the Christoffel symbols (expressed in the coordinate frame) satisfy
some specific relations, since the connection is torsion free and verifies the compatibility condition (5.4.1).
These properties can be translated on the connection forms, see [2, Thm. VII.4.6] or [4, Thm. 4.2.1].
Theorem 6.3.2 (Structure theorem of Cartan). Let M be a smooth manifold, let {X1 , . . . , Xm } be a smooth
field of frames, and let {ω1 , . . . , ωm } be the dual field of coframes. The connection forms of the Riemannian
connection are the unique solutions of the equations:
(i) dωi = mj=1 ω j ∧ ωij ,
P

(ii) dgi j = mk=1 gk j ωi + gki ω j

k k
P 

where gi j = ϕ(Xi , X j ).
Let us observe that if the bases {X1,p , . . . , Xm,p } are orthonormal, and accordingly the bases {(ω1 ) p , . . . , (ωm ) p }
are orthonormal, then g jk = δ jk and the second relation becomes simply ωki + ωkj = 0.
In the same framework as above, let us still introduce the curvature 2-forms: For any k, ℓ ∈ {1, . . . , m}
m
1X
Ωℓk := Ri, j,k ℓ ωi ∧ ω j (6.3.2)
2 i, j=1

where Ri, j,k ℓ are the coefficients of the curvature relative to these frames, see also (6.2.1), namely
m
Ri, j,k ℓ Xℓ .
X
R(Xi , X j )Xk =
ℓ=1

The 2-forms Ωℓk satisfy then the following statement:

Theorem 6.3.3. In the above framework and for any k, ℓ ∈ {1, . . . , m} the following equality holds:
(iii) Ωℓk = dωℓk − mj=1 ωk ∧ ωℓj .
P j

Exercise 6.3.4. Prove the previous statement, getting inspiration from [4, Prop. 4.2.3] or [2, Thm. VIII.4.1].
The equations (i) and (ii) of Theorem 6.3.2 together with equation (iii) of Theorem 6.3.3 are called the Cartan
structure equations for the connection forms defined in (6.3.1) and the curvature forms defined in (6.3.2).
With these definitions and equations, everything has been expressed in terms of forms. This formalism has
been developed by E. Cartan who made an extensive use of them. Let us stress once again that if one chooses
a particular field of orthonormal frames, then some of these relations become much simpler.

6.4 Holonomy
In this section we mention the basics of the notion of holonomy on a connected smooth manifold, and refer
to [1, Sec. 10] and [3] for additional information. Note that the theory can be developed in the more general
context of vector bundles or principal bundles.
Let c : [0, 1] → M be a piecewise smooth parametric curve on a connected Riemannian manifold M, with
c(0) = p. Let X p ∈ T p (M) be a tangent vector at p and set X(t) for its parallel transport along the curve, with
X(0) = X p , see Figure 6.3. Note that X(·) is uniquely defined, as mentioned in Theorem 5.4.10. We can then
look at the map
Pc : T p (M) ∋ X p ≡ X(0) 7→ X(1) ∈ T c(1) (M)

63
which gives the image of X p by its parallel transport along the curve defined by c. It is then easily checked
that
Pc2 Pc1 = Pc2 ◦c1 , (6.4.1)
where c1 and c2 are curves on M with c1 (1) = c2 (0), and where c2 ◦ c1 represents the curve obtained by
considering the 2 curves successively. It can also be checked that

Pc−1 = (Pc )−1 , (6.4.2)

where c−1 represents the parametric curve taken in the reverse direction, and (Pc )−1 denotes the inverse map
from T c(1) (M) to T c(0) (M) ≡ T p (M). If the curve is closed, namely if c(0) = c(1) = p, it turns out that


Pc ∈ GL T p (M) , the group of invertible linear maps on T p (M). In addition, since the parallel transport


preserves the norm, one has Pc ∈ O T p (M) , the orthogonal group on T p (M).

Figure 6.3: The parallel transport of a vector along a piecewise smooth closed curve on a sphere. The differ-
ence between X(0) and X(1) at P is clearly visible

Exercise 6.4.1. Prove the statements on the map Pc , getting inspiration from [3, Sec. 3.1] if necessary.
Let us now consider the set of all piecewise smooth closed curve starting at p, denoted by C p . Then, thanks to



(6.4.1) and (6.4.2) the set {Pc }c∈C p defines a group of O T p (M) . This group is denoted by Hol(p) ⊂ O T p (M) .
It is isomorphic for all p ∈ M, since the following equality holds:

Hol(p) = Pc−1 Hol(q) Pc ≡ (Pc )−1 Hol(q) Pc

where c is a piecewise smooth parametric curve from p to q. Note that this result holds since we have assumed
M to be connected, and that any smooth connected manifold is path connected. Based on these isomorphisms,
one can set:
Definition 6.4.2 (Holonomy group). For any connected Riemannian manifold M and any p ∈ M the holonomy


group Hol(M) (at the basepoint p) is defined as the subgroup Hol(p) ⊂ O T p (M) . For q , p, Hol(q) is
isomorphic to Hol(p), with an isomorphism depending on the choice of a piecewise smooth curve from p to q.
Note that this definition is not fully satisfactory since the representation of Hol(M) depends on the choice of
a basepoint p. Nevertheless for computations, one always refers to a specific basepoint, and therefore this
definition is the useful one.
A smaller subgroup can also be defined. Recall that the notion of homotopic curves has been introduced in
Section 4.2. With the same idea, a piecewise smooth closed curve is contractible if it can be continuously

64
deformed into the trivial curve (the curve c : [0, 1] → M with c(t) = p for all t ∈ [0, 1]). By letting C 0p denote
the set of all piecewise smooth contractible curves one is naturally led to the subgroup Hol0 (p) ⊂ Hol(p)
defined with the set {Pc }c∈C 0p . In addition, Hol0 (p) is contained in the special orthogonal group SO T p (M) on

T p (M). As above, one can define Hol0 (M) with contractible curves, and call it the restricted holonomy group.
It turns out that Hol0 (M) corresponds to the identity component of Hol(M), and that Hol0 (M) is a normal
subgroup of Hol(M), if the same basepoint is chosen for their definitions.
An interesting relation between orientation of the holomony group is mentioned in the next statement. Its
proof is proposed as an exercise in [1, Example 10.24].


Lemma 6.4.3. The connected Riemannian manifold M is orientable if and only if Hol(p) ⊂ SO T p (M) for
any p ∈ M.
The notion of holonomy is closely related to the curvature of the connection, as introduced in Definition
5.4.2. In fact the curvature of the connection can be seen as the generator of the holonomy. The relation
precise relation is rather involved, and is known under the name of Ambrose-Singer theorem. Its precise
formulation involves the Lie algebra related to the holonomy group. Nevertheless, in the restricted setting of
two commuting vector fields, a formal description can be understood as follow.
For a connected Riemannian manifold M and for p ∈ M, consider two smooth vector fields X and Y (defined
on neighbourhood of p) and which commute, namely [X, Y] = 0, see Section 2.2. Let θX and θY denote the
local flow generated by X and by Y, whose existence has been shown in Theorem 2.3.5. Consider then a
parallelogram, starting at p and defined by these flows, with the four sides of length t > 0, see Figure 6.4.
This parallelogram corresponds to a piecewise smooth closed curve, of length 4t. Consider an additional
Z ∈ X(M), with Z p ∈ T p (M) and its parallel transport along the parallelogram, denoted by t 7→ Z(t). Then
following formal relation holds:
dZ(·)
(0) = (∇Y ∇X − ∇X ∇Y )Z p = −R(X p , Y p )Z p ,


dt
where the content of Remark 6.2.6 has been taken into account for the right-hand side of the equality.

Figure 6.4: A parallelogram based on the flows of two commuting vector fields

For completeness, let us still provide the precise statement of Ambrose-Singer theorem, as stated in [1,
Thm. 10.58]. For that purpose, we denote by c an arbitrary piecewise smooth curve on M starting at p,
and by τ(c) the parallel transport of a tangent vector along c, namely τ(c)(X p ) is the parallel transport of
X p ∈ T p (M) along c. An illustration of the following statement is given in Figure 6.5.


Theorem 6.4.4. The Lie algebra hol(p) of Hol(p) corresponds to the subalgebra of so T p (M) generated by
elements of the form
τ(c) −1 R τ(c)(X p ), τ(c)(Y p ) ◦ τ(c)

where X p , Y p run through T p (M) and c through all piecewise smooth curves starting at p.

65
Figure 6.5: Illustration of the for Ambrose-Singer theorem

66
Chapter 7

General relativity

7.1 Pseudo-Riemannian manifolds and Einstein fields equations

In this final chapter, we just touch the surface of the subject, and refer to more specialized courses for a
real introduction to this topic, see for example [6]. One key difference with the previous chapters is that the
manifolds will not be considered Riemannian any more. Recall that the definition of a Riemannian manifold
is based on the existence of a symmetric and positive definite field of bilinear maps, see Definition 3.3.7. This
assumption will have to be modified.
Definition 7.1.1 (Non-degenerate bilinear map and tensor field). Let V be a finite dimensional real vector
space. A bilinear map ϕ : V × V → R (or equivalently ϕ ∈ T 2 (V) ) is non-degenerate if the condition
ϕ(u, v) = 0 for all v ∈ V implies that u = 0. A smooth tensor field ϕ ∈ T2 (M) on a smooth manifold M is
non-degenerate if ϕ p : T p (M) × T p (M) → R is non-degenerate for all p ∈ M.
Exercise 7.1.2 (Orthonormal basis and signature). Let ϕ be a symmetric and non-degenerate bilinear map
on a vector space V of dimension n. Show that there always exists an orthonormal basis {e1 , . . . , en } of V
satisfying ϕ(ei , e j ) = εi j with εi j ∈ {±1} if i = j and εi j = 0 if i , j. Moreover, show that the signature
s := ni=1 εii does not depend on the choice of the orthonormal basis.
P

A vector space V endowed with a symmetric and non-degenerate bilinear map is called a pseudo-Euclidean
space. Be aware that the notation for the signature of a pseudo-Euclidean space depends on the authors: it is
sometimes denoted (p, q), sometimes (q, p), with p the number of positive values for εii and q the number of
negative values for εii . In both case, the number s introduced in the previous exercise corresponds to p − q.
We can now define the family of manifolds involved in general relativity. Recall that Σ2 (M) denotes the field
of smooth symmetric bilinear maps on M.
Definition 7.1.3 (Pseudo-Riemannian manifold). A pseudo-Riemannian manifold, also called semi-Rieman-
nian manifold is a smooth manifold endowed with a non-degenerate tensor field ϕ ∈ Σ2 (M). The tensor field
ϕ is also called the metric tensor.
By definition each tangent space of a pseudo-Riemannian manifold is a pseudo-Euclidean space, all of them
with the same signature. An important family of pseudo-Rimannian manifolds is the family of Lorentzian
manifolds of dimension m and of signature s = m − 2. It means that εii = −1 for a single index i, while εii = 1
for all the other indices. A summary on Lorentzian manifolds can be found for example in the lecture notes
of F. Pfäffle. Even more concretely, the Minkowski space consists in the pseudo-Euclidean space R4 with a

67
specific metric tensor on each tangent space, leading to a signature s = 2, see below. Note that the alternative
choice of εii = −1 for three indices and εii = 1 for one index can also be taken, as done for example here. This
choice leads to a signature equas to −2.
In Theorem 5.4.4 it was stated that on any Riemannian manifold there exists only one symmetric connection
compatible with the metric ϕ. This connection can be defined by Koszul formula, as presented in Exercise
5.4.5. It turns out that the proof of uniqueness is based only on the non-degeneracy of the bilinear form, and not
on its positive-definiteness. As a consequence, a similar statement holds for pseudo-Riemannian manifold:
there exists a unique connection on any pseudo-Riemannian manifold which is symmetric (or equivalently
torsion free) and for which the map ϕc(t) X(t), Y(t) is independent of t, if X(t) and Y(t) are vectors transported


parallel along any smooth curve c. This condition corresponds to the preservation of the inner product in
the Riemannian setting. We refer to [6, Thm. 15.4] for the precise statement. Note that Koszul formula still
provides the expression for this connection. Also, Cartan structure equations, as introduced in Theorems 6.3.2
and 6.3.3 are still valid in this setting and for this connection, see [6, Sec. 15.7].
From now on, we shall consider only 4-dimensional Lorentzian manifolds endowed with the unique connec-
tion mentioned above, which is commonly called its Levi-Civita connection. These manifolds are commonly
called space-time manifolds, or simply space-times. With the connection, geodesics can be defined, and cor-
respond to parametric curves t 7→ c(t) ∈ M for which ċ is transported parallel to c, see Definition 5.5.1. This
implies that the map t 7→ ϕc(t) ċ(t), ċ(t) is constant, namely independent of t. By using the simpler notation5

⟨ċ, ċ⟩ := ϕc ċ, ċ

the following definition holds:

Definition 7.1.4 (timelike, null, spacelike geodesics). For a space-time manifold, a geodesic is timelike if
⟨ċ, ċ⟩ < 0, is null if ⟨ċ, ċ⟩ = 0, and is spacelike if ⟨ċ, ċ⟩ > 0.
Observe that this definition would not make sense for Riemannian manifolds. It is only meaningful for pseudo-
Riemannian manifold when the signature is different from ±m, with m the dimension of the manifold. This
terminology is borrowed from the Minkowski space in special relativity, for which the framework is the
manifold M := R4 endowed with the symmetric and non-degenerate tensor field ϕ satisfying

ϕ p (u1 , u2 , u3 , u4 ), (v1 , v2 , v3 , v4 ) = −u1 v1 + u2 v2 + u3 v3 + u4 v4

(7.1.1)

for any u = (u1 , u2 , u3 , u4 ) and v = (v1 , v2 , v3 , v4 ) belonging to T p (M) and any p ∈ M. Observe that (7.1.1)
can also be expressed by saying that the matrix gi j (p) 4i, j=1 introduced in (5.1.1) is the diagonal matrix with


diagonal entries (−1, 1, 1, 1). We refer to [4, Sec. 6.2] for an extremely quick introduction to special relativity
expressed with the tools of differential geometry.
In the framework of a space-time manifold (M, ϕ) the Einstein fields equations read
1
G jk + Λg jk ≡ R jk − Rg jk + Λg jk = κT jk (7.1.2)
2
with κ := 8πG
c4
is the Einstein gravitational constant (with G is the Newtonian constant of gravitation and
c is the speed of light in vacuum), and Λ the cosmological constant. On the l.h.s. of (7.1.2), the term R jk
corresponds to the Ricci curvature, as introduced in (6.2.3), while R denotes the scalar curvature, already seen
in (6.2.4). As before, the notation g jk stands for the metric tensor evaluated on a smooth field of frames,
5
Observe that we don’t write ∥ċ∥2 := ϕc ċ, ċ since the notation on the l.h.s. is usually used for a norm, while the r.h.s. does not


correspond to a norm.

68
as mentioned in (5.1.1) in the special case of the coordinate frame. Clearly, the Ricci curvature, the scalar
curvature, and the metric tensor have to be computed with the same field of frames. Altogether the terms
R jk − 12 Rg jk form the so-called Einstein tensor G jk . On the other hand, the r.h.s. of (7.1.2) corresponds to the
stress-energy tensor or energy-momentum tensor. It describes the density and flux of energy and momentum
in space-time, generalizing the stress tensor of Newtonian physics. In summary, the r.h.s. of (7.1.2) contains
the physics, while the l.h.s. is related to the geometry of the space-time. In other words, the flux of energy and
momentum dictates the geometry itself.
Let us now recall that the coefficients of the curvature tensor were expressed in terms of the Christoffel symbols
in (6.2.2), and that the Christoffel symbols could be expressed in terms of the metric tensor, as presented in
Lemma 5.4.7 in the framework of a Riemannian manifold. It turns out that these relations also hold for
pseudo-Riemannian manifolds. More precisely, one has
m ∂Γi m ∂Γi m X
m m X
m
Γℓj,k Γii,ℓ − Γℓi,k Γij,ℓ ,
j,k
X X i,k
X X
R jk = − +
i=1
∂xi i=1
∂x j i=1 ℓ=1 i=1 ℓ=1
m X
X m
R= giℓ Riℓ ,
i=1 ℓ=1

and
m
1 X iℓ  ∂gℓ j ∂gℓk ∂g jk 
Γij,k = g + − ,
2 ℓ=1 ∂xk ∂x j ∂xℓ
where we use again the convention about the derivatives mentioned in Lemma 5.4.7. By looking carefully
at these expressions one observes that the l.h.s. of (7.1.2) can be expressed in terms of the metric tensor
and its derivatives. More precisely, these equations correspond to 10 non-linear partial differential equations
containing the metric tensor and its derivatives up to order 2. Note that the number 10 is due to the symmetry
in the two indices.
Exercise 7.1.5. Work on some of the exercises proposed in [4] as for example Exercises 2.2 p. 260 or 4.3
p. 268. In particular, look at the twin paradox or at the car and garage paradox, and / or at other classical
exercises in this framework.
Another classical exercise in this setting is to derive the Schwarzschild solution of Einstein equations, see for
example [4, Sec. 6.5]. This solution has been the first discovered solution and is very much related to black
holes. Other exact solutions have then been found, like the Kerr solution, and later the Kerr-Newman solution.
Working on these solutions is part of any course on general relativity and the solutions can be found in several
textbooks. Note that Cartan structure equations provide a powerful starting point for deriving these solutions.

7.2 Causality
In this final section we mention a subject which has no analog for Riemannian manifolds, since for these
manifolds the metric is positive definite. A consequence of the non-positive definiteness has already been
mentioned in Definition 7.1.4, where timelike, null, spacelike geodesics were introduced. We now further
investigate this subject, focusing on 4-dimensional Lorentzian manifolds.
Let (M, ϕ) be a Lorentzian manifold and let p ∈ M. On any tangent space T p (M), the non-zero tangent vectors
can be divided into three families: X p is timelike if ϕ p (X p , X p ) < 0, X p is null or lightlike if ϕ p (X p , X p ) = 0,
and X p is spacelike if ϕ p (X p , X p ) > 0. We shall further divide the timelike tangent vectors, and divide them

69
into two equivalence classes. Namely, for two timelike elements X p , Y p ∈ T p (M), we say that X p and Y p
are equivalent if ϕ p (X p , Y p ) < 0, and write X p ∼ Y p in this situation. Usually, one of these classes is called
the class of future-directed vectors while the other class consists of past-directed vectors. Physically this
designation of the two classes of future- and past-directed timelike vectors corresponds to a choice of an
arrow of time at the point p. Note that the classes of future- and past-directed vectors can be extended to null
vectors by continuity.
Exercise 7.2.1. Show that the previous construction is correct, finding some inspiration here if necessary.
So far, this choice has been made independently in each tangent space. However. a more uniform approach is
necessary.
Definition 7.2.2 (Time-orientable manifold). A Lorentzian manifold (M, ϕ) is time-orientable if there exists
X ∈ X(M) (a smooth vector field on M) such that ϕ(X, X) < 0, or more precisely ϕ p (X p , X p ) < 0 for any
p ∈ M.
Clearly, this vector field takes values in the set of timelike vectors on each tangent space. By assuming the
existence of such a specific vector field X, then one can choose coherently a time orientation for timelike
vectors on each tangent space T p (M) by choosing all vectors which are in the component of X p . This choice
corresponds to choosing a time-orientation on a time-orientable manifold, one then speaks about a time-
oriented manifold.
Let us now extend the notion of timelike, null, spacelike geodesics of Definition 7.1.4 to arbitrary curves,
namely a curve c is timelike or chronological if ⟨ċ, ċ⟩ < 0, is null if ⟨ċ, ċ⟩ = 0, and is spacelike if ⟨ċ, ċ⟩ > 0.
We also say that the curve is causal or non-spacelike if ⟨ċ, ċ⟩ ≤ 0. Then, any timelike, null, or causal curve
c is said to be future-directed if ċ(t) belong to the set of future-directed vectors for all t, while the curve is
past-directed if ċ(t) belongs to the set of past-directed vectors for all t.
By using these notions on curves, causal relations can now be introduced between points of M. For any
p, q ∈ M,
(i) p chronologically precedes q (often denoted by p ≪ q) if there exists a future-directed timelike curve
from p to q,
(ii) p strictly causally precedes q (often denoted by p < q) if there exists a non-zero future-directed causal
curve from p to q,
(iii) p causally precedes q (often denoted by p ≺ q or p ≤ q) if p strictly causally precedes q or if p = q,
(iv) p horismos q (often denoted p → q or p ↗ q) if p = q or if there exists a future-directed null curve
from p to q.
Exercise 7.2.3. Exhibit the properties of these relations, see for example here or on page 12 of [5].
Given a point p ∈ M, it is then possible to describe very precisely what means its future and what means its
past. More precisely, the chronological future of p, denoted I + (p), is defined by I + (p) = {q ∈ M | p ≪ q},
while the chronological past of p, denoted I − (p), is defined by I − (p) = {q ∈ M | q ≪ p}. Similarly, the causal
future of p, denoted by J + (p), is defined by J + (p) = {q ∈ M | p ≺ q}, while the causal past of p, denoted by
J − (p), is defined by J − (p) = {q ∈ M | q ≺ p}.
In general, the structure of these sets are quite complicated. However, the causal properties are locally quite
similar to the simpler ones encountered in Minkowski space-time for special relativity. The following state-
ment is borrowed from [4, Prop. 7.1].

70
Proposition 7.2.4. Let (M, ϕ) be a time-oriented space-time, and let p0 ∈ M. Then there exists an open
neighbourhood U of p0 such that when restricted to U:
(i) U is geodesically convex, namely given p, q ∈ U there exists a unique geodesic (up to reparametriza-
tion) between p and q,
(ii) q ∈ I + (p) if and only if there exists a future-directed timelike geodesic connecting p to q,
(iii) J + (p) = I + (p),
(iv) q ∈ J + (p) \ I + (p) if and only if there exists a future-directed null geodesic connecting p to q.
Exercise 7.2.5. Based on the current physical knowledge (massive particle can not travel faster than or at the
speed of light), interpret the previous definitions and statements.
By using on the previous proposition, the global structure of the space-time can be partially described. In
particular, it can be shown that for any p ∈ M the set I + (p) is an open set in M, and that J + (p) ⊂ I + (p). A
general version of the twin paradox also exists in this general framework, as stated in [4, Prop. 7.3].
For the use of Lorentzian manifolds in physics, additional conditions are often imposed, as for example
reasonable causalities conditions. For example, these conditions exclude time travel, namely the possibility of
a particle returning to a point in its past history. The following two definitions take care of this requirement.
Definition 7.2.6 (Chronology condition). A space-time (M, ϕ) satisfies the chronology condition if it does not
contain closed timelike curves.
An interesting result related to this condition is contained in the next statement, borrowed from [4, Prop. 7.7].
Proposition 7.2.7. Any compact space time (M, ϕ) contains closed timelike curves.
The direct consequence of this statement is that a space-time satisfying the chronology condition can not be
compact ! An even stronger condition can be expressed in term of the gradient of a smooth function on M
(where M is any Riemannian or pseudo-Riemannian manifold). More precisely, for f ∈ C ∞ (M), we define
the vector field grad( f ) ∈ X(M) through a local chart (U, φ) and the corresponding coordinate frame:
X ∂( f ◦ φ−1 )
gi j (p) φ(p) Ei,p



(grad( f ) :=
p
i, j=1m
∂x j
Pm i jE
or more concisely i, j=1 g j ( f )E i .

Definition 7.2.8 (Stably causal). A space-time (M, ϕ) is said to be stably causal if there exists a global time
function, namely a smooth function t : M → R such that its gradient grad(t) is timelike, namely grad(t) p is
timelike for every p ∈ M.
Note that any stably causal space-time is time orientable and satisfies the chronology condition. In addi-
tion, any manifold obtained by small perturbations of a stably causal manifold also satisfies the chronology
condition.
Let us now introduce the last concept, Cauchy surfaces. For this, let us consider a smooth causal curve
c : (a, b) → M, with possibly a = −∞ and b = ∞. This curve is future-inextendible if it is future-directed
and limt↗b c(t) does not exist. The curve is past-inextendible if it is past-directed and limt↘a c(t) does not
exist. Given a subset S ⊂ M, the future Cauchy development of S or the future domain of dependence of
S , denoted by D+ (S ) consists of all p ∈ M such that any causal and past-inextensible curve starting at p
intersects S . Therefore “any causal influence on p ∈ D+ (S ) had to register somewhere on S , and one can

71
expect that what happens at p can be predicted from data on S ”. Similarly, the past Cauchy development
of S or the past domain of dependence of S , denoted by D− (S ) consists of all p ∈ M such that any causal
and future-inextensible curve starting at p intersects S . Therefore, “any causal influence on p ∈ D− (S ) will
register somewhere in S , and one can expect that what happened at p can be retrodicted from data on S ”. Then
the Cauchy development of the domain of dependence D(S ) of S is simply defined by D(S ) = D+ (S ) ∪ D− (S ).
A representation of these sets is provided in Figure 7.1.

Figure 7.1: Cauchy developments of a surface S

In order to be meaningful, the set S used in the previous definition should not be arbitrary, namely it should
be achronal: S ⊂ M is achronal if there are no p, q ∈ S with p ∈ I + (q).
Definition 7.2.9 (Cauchy surface). In the framework introduced above, a Cauchy surface is a closed achronal
set S whose Cauchy development is equal to M.
In the physics of general relativity, a Cauchy surface is usually interpreted as defining an “instant of time”.
Accordingly, Cauchy surfaces provide boundary conditions for the causal structure in which the Einstein
equations can be solved. For stably causal space-time, with global time function t : M → R, if all level sets
S τ := t−1 (τ) for τ ∈ R are Cauchy surfaces, then the space-time is said to be globally hyperbolic. Such spaces
have particularly good causal properties.
If D(S ) , M, then there exists a Cauchy horizon between D± (S ) and regions of the manifold are not com-
pletely determined by information on S . A physical example of a Cauchy horizon is the second horizon inside
a charged or rotating black hole. It is out of the scope of these notes to look beyond the Cauchy horizon, and
we leave this for your future studies and refer to more specialized books. Good luck for your future, look
behind your Cauchy horizon !

72
Index

Achronal, 72 Contraction of tensors, 62

Affine connection, 48 Contravariant, 24
Algebra of differential forms, 30 Coordinate coframe, 27
Alternating, 29 Coordinate frame, 15
Alternating tensor, 24 Coordinate neighbourhood, 6
Ambrose-Singer theorem, 65 Coordinate system, 6
Antisymmetrization, 25 Cosmological constant, 68
Arc length, 42, 43 Covariant, 24
Covariant derivative, 46
Basis, 3 Covector field, 28
Bianchi identity, 60 Curvature, 44, 49
Bilinear map, 28 Curvature 2-form, 63
Black hole, 69
Boundary, 6 Diffeomorphism, 9
Differential, 15
Cartan structure equations, 63 Discrete space, 3
Cauchy development, 72 Divergence theorem, 37
Cauchy horizon, 72 Domain of dependence, 72
Cauchy surface, 72 Dual basis, 27
Causal, 70
Causal future, 70 Einstein fields equations, 68
Causal past, 70 Einstein gravitational constant, 68
Causally precedes, 70 Einstein tensor, 69
Chart, 6 Embedded submanifold, 10
Christoffel symbols, 49 Embedding, 10
Chronological, 70 Endomorphism, 48
Chronological future, 70 Energy-momentum tensor, 69
Chronological past, 70 Exponential map, 53
Chronologically precedes, 70 Exterior algebra, 26, 30
Chronology condition, 71 Exterior derivative, 31
Closed set, 2 Extrinsinc, 1
Compact, 4
Connected, 3 First fundamental form, 58
Connection, 48 Flow, 20
Connection form, 62 Frame, 43
Constant curvature, 61 Frenet frame, 43
Continuous map, 4 Frenet-Serret formulas, 43
Contractible, 64 Future Caucy development, 71

73
Future domain of dependence, 71 Linear functional, 24
Future-directed, 70 Local coordinate function, 6
Future-inextendible, 71 Local coordinates, 6
Local flow, 21
Gauss curvature, 59 Local representation, 9
Genus, 5 Locally finite, 35
Geodesic, 52
Geodesic equation, 52 m-submanifold property, 11
Geodesically convex, 71 Maximal flow, 21
Germs, 14 Mean curvature, 59
Globally hyperbolic, 72 Metric space, 42
Gradient Theorem, 37 Metric tensor, 67
Gram-Schmidt process, 41 Minimal geodesic, 55
Grassmann algebra, 26 Minimal surface, 59
Module, 19
Hairy ball theorem, 19
Hausdorff property, 3 Nash embedding theorem, 43
Holonomy group, 64 Negatively oriented basis, 32
Homeomorphic, 4 Neighbourhood, 2
Homeomorphism, 4 Newtonian constant of gravitation, 68
Homomorphism, 14 Non-degenerate, 67
Homotopic curves, 39 Non-spacelike, 70
Horismos, 70 Normal, 45
Normal coordinate neighbourhood, 54
Identity component, 65 Normal section, 58
Imbedding, 10 Normal subgroup, 65
Immersed manifold, 10 Null, 68
Immersion, 10
Induced basis, 33 One-parameter group action, 20
Induced topology, 3 Open cover, 3
Inner product, 41 Open set, 2
Integral curve, 20 Orbit, 20
Integral of m-forms, 35 Orientation, 32
Intrinsic, 1 Orientation preserving, 32
Isometric, 43 Oriented manifold, 35
Isomorphism, 15 Outward pointing vector, 33
isotropic, 61
Parallel, 51
Kerr solution, 69 Partition of unity, 35
Koszul formula, 49 Past Cauchy development, 72
Past domain of dependence, 72
Length, 42 Past-directed, 70
Levi-Civita connection, 49, 68 Past-inextendible, 71
Lie algebra, 19, 65 Path connected, 64
Lie bracket, 19 Piecewise smooth, 51
Lie derivative, 22 Planar point, 58
Lightlike, 69 Plane section, 61
Line integral, 39 Positive definite, 29

74
Positively oriented basis, 32 Support, 35
Preferred coordinates, 11 Symmetric, 29, 49
Preimage, 4 Symmetric tensor, 24
Principal bundle, 63 Symmetrization, 25
Principal curvature, 59 Symmetry, 42
Principal direction, 59
Pseudo-Euclidean space, 67 Tangent bundle, 18
Pseudo-Riemannian manifold, 67 Tangent space, 14
Tangent vector, 14
Rank, 9 Tensor, 24
Regular, 43 Tensor algebra, 25
Regular manifold, 12 Theorema Egregium, 59
Regular point, 21 Time-orientable, 70
Relative topology, 3 Time-oriented, 70
Restricted holonomy group, 65 Timelike, 68
Ricci curvature, 61 Topological manifold, 4
Riemannian connection, 49 Topological manifold with boundary, 6
Riemannian curvature tensor, 61 Topological space, 2
Riemannian curvature tensor field, 60 Topology, 2
Rimannian manifold, 29 Torsion, 45, 49
Torsion free, 49
Scalar curvature, 62 Trace topology, 3
Schwarzschild solution, 69 Transposition, 25
Second countable, 3 Triangle inequality, 42
Second fundamental form, 58
Sectional curvature, 61 Umbilical point, 58
Semi-Riemannian manifold, 67 Unit normal vector, 57
Separation axioms, 2 Vector bundle, 63
Singular point, 21 Vector field, 19
Skew symmetric, 29 Volume element, 42
Smooth atlas, 8 Volume form, 42
Smooth manifold, 7
Smooth partition, 36 Wedge product, 26
Smooth tensor field, 28
Space-time, 68
Spacelike, 68
Speed of light, 68
Stably causal, 71
Star domain, 53
Stokes’ theorem, 37
Stress-energy tensor, 69
Strictly causally precedes, 70
Subcover, 3
Submanifold, 10
Submersion, 10
Subordinated, 36
Subspace topology, 3

75
Bibliography

[1] A. Besse, Einstein manifolds, Reprint of the 1987 edition, Springer, 2008.
[2] W. Boothby, An introduction to differentiable manifolds and Riemannian geometry, Second edition, Pure
Appl. Math. 120, Academic Press, Inc., Orlando, FL, 1986.
[3] A. Clarke and B. Santoro, Holonomy groups in Riemannian geometry, lecture notes, Preprint https:
//arxiv.org/abs/1206.3170.
[4] L. Godinho and J. Natario, An Introduction to Riemannian Geometry, with applications to mechanics
and relativity, Universitext, Springer, Cham, 2014.
[5] R. Penrose, Techniques of differential topology in relativity, Society for Industrial and Applied Mathe-
matics, Philadelphia, PA, 1972.
[6] N. Straumann, General relativity, second edition, Springer, 2013.
[7] L. Tu, An introduction to manifolds, Universitext, Springer, New York, 2008.
[8] L. Tu, Differential geometry, Grad. Texts in Math. 275, Springer, Cham, 2017.
[9] https://www.gmanetwork.com/news/scitech/science/583886/when-is-a-coffee-mug-a-donut-topology-explains-it/
story/
[10] https://en.wikipedia.org/wiki/Differential structure
[11] https://en.wikipedia.org/wiki/Orientability
[12] https://x.com/stefanhsommer/status/1767562471940718919
[13] https://en.wikipedia.org/wiki/Manifold

76