Differentiable Manifolds &

Riemannian Geometry

F REDERICK T SZ -H O F ONG

Hong Kong University of Science and Technology

(Version: January 18, 2021)
Contents

Preface ix

Part 1. Differentiable Manifolds

Chapter 1. Regular Surfaces 3

1.1. Local Parametrizations 3
1.2. Level Surfaces 10
1.3. Transition Maps 13
1.4. Maps and Functions from Surfaces 16
1.5. Tangent Planes and Tangent Maps 20

Chapter 2. Abstract Manifolds 25

2.1. Smooth Manifolds 25
2.2. Functions and Maps on Manifolds 34
2.3. Tangent Spaces and Tangent Maps 39
2.4. Inverse Function Theorem 46
2.5. Immersions and Submersions 51
2.6. Submanifolds 57

Chapter 3. Tensors and Differential Forms 63

3.1. Cotangent Spaces 63
3.2. Tangent and Cotangent Bundles 66
3.3. Tensor Products 77
3.4. Wedge Products 84
3.5. Exterior Derivatives 91

Chapter 4. Generalized Stokes’ Theorem 111

4.1. Manifolds with Boundary 111
4.2. Orientability 117
4.3. Integrations of Differential Forms 123
4.4. Generalized Stokes’ Theorem 132

vii
viii Contents

Chapter 5. De Rham Cohomology 141

5.1. De Rham Cohomology 142
5.2. Deformation Retracts 148
5.3. Mayer-Vietoris Theorem 153

Part 2. Euclidean Hypersurfaces

Chapter 6. Geometry of Curves 165
6.1. Curvature and Torsion 165
6.2. Fundamental Theorem of Space Curves 176
6.3. Plane Curves 181

Chapter 7. Geometry of Euclidean Hypersurfaces 185

7.1. Regular Hypersurfaces in Euclidean Spaces 185
7.2. Fundamental Forms and Curvatures 188
7.3. Theorema Egregium 201

Part 3. Riemannian Geometry

Chapter 8. Riemannian Manifolds 209
8.1. Riemannian Metrics 209
8.2. Levi-Civita Connections 214
Chapter 9. Parallel Transport and Geodesics 225
9.1. Parallel Transport 225
9.2. Geodesic Equations 230
9.3. Exponential Map 234
Chapter 10. Curvatures of Riemannian Manifolds 239
10.1. Riemann Curvature Tensor 239
10.2. Sectional Curvature 246
10.3. Ricci and Scalar Curvatures 253
10.4. Variation Formulae of Curvatures 260
Chapter 11. Curvatures and Topology 267
11.1. Jacobi Fields 268
11.2. Index Forms 274
11.3. Spaces of Constant Sectional Curvatures 279

Bibliography 285
Preface

This lecture note is written for courses MATH 4033 (Calculus on Manifolds) and MATH
6250I (Riemannian Geometry) taught by the author in the Hong Kong University of
Science and Technology.
The main goal of these courses is to introduce advanced undergraduates and first-
year graduate students the basic concepts of differentiable manifolds, tensor calculus,
cohomology, and Riemannian geometry. It presents some of the most essential knowledge
on differential geometry that is necessary for further studies or research in geometric
analysis, general relativity, string theory, and related fields. Before reading the lecture
notes and taking these courses, students are advised to have a solid conceptual background
of linear algebra (MATH 2131) and multivariable calculus.
The course MATH 4033 covers Chapters 1 to 5 in this lecture note. These chapters
are about the analytic, algebraic, and topological aspects of differentiable manifolds.
Chapters 6 and 7 form a crush course on differential geometry of hypersurfaces in
Euclidean spaces. The main purpose of these two chapters is to give some motivations
on why various abstract concepts in Riemannian geometry are introduced in the way
they are. The remaining Chapters 8 to 11 form an introduction course to Riemannian
geometry.
Students are very welcome to point out typographical errors of any part of the notes,
and contribute diagrams to the Riemannian geometry chapters. The author would like
to thank the following students for their diligent readings of the earlier version of the
lecture notes and for pointing out many typographical errors: Chow Ka-Wing, Alex
Chan Yan-Long, Aaron Chow Tsz-Kiu, Jimmy Choy Ka-Hei, Toby Cheung Hin-Wa, Poon
Wai-Tung, Cheng Chun-Kit, Chu Shek-Kit, Wan Jingbo, Nicholas Chin Cheng-Hoong, Tang
Tianchen, and Luk Hoi-Ping.
Frederick Tsz-Ho Fong
January 18, 2021
HKUST, Clear Water Bay, Hong Kong

ix
Part 1

Differentiable Manifolds
 Chapter 1

Regular Surfaces

“God made solids, but surfaces were

the work of the devil.”

Wolfgang Pauli

A manifold is a space which locally resembles an Euclidean space. Before we learn

about manifolds in the next chapter, we first introduce the notion of regular surfaces in
R3 which motivates the definition of abstract manifolds and related concepts in the next
chapter.

1.1. Local Parametrizations

In Multivariable Calculus, we expressed a surface in R3 in two major ways, namely using
a parametrization F (u, v) or by a level set f (x, y, z) = 0. In this section, let us first focus
on the former.
Recall that in MATH 2023 we can use a parametrization F (u, v) to describe a surface
in R3 and to calculate various geometric and physical quantities such as surface areas,
surface integrals and surface flux. To start the course, we first look into several technical
and analytical aspects concerning F (u, v), such as their domains and images, their
differentiability, etc. In the past, we can usually cover (or almost cover) a surface by a
single parametrization F (u, v). Take the unit sphere as an example. We learned that it
can be parametrized with the help of spherical coordinates:

F (θ, ϕ) = (sin ϕ cos θ, sin ϕ sin θ, cos ϕ)

where 0 < θ < 2π and 0 < ϕ < π. This parametrization covers almost every part
of the sphere (except the north and south poles, and a half great circle connecting
them). In order to cover the whole sphere, we need more parametrizations, such as
G(θ, ϕ) = (sin ϕ cos θ, sin ϕ sin θ, cos ϕ) with domain −π < θ < π and 0 < ϕ < π.
Since the image of either F or G does not cover the whole sphere (although almost),
from now on we call them local parametrizations.

3
4 1. Regular Surfaces

Definition 1.1 (Local Parametrizations of Class C k ). Consider a subset M ⊂ R3 . A

function F (u, v) : U → O from an open subset U ⊂ R2 onto an open subset O ⊂ M is
called a C k local parametrization (or a C k local coordinate chart) of M (where k ≥ 1) if
all of the following holds:
(1) F : U → R3 is C k when the codomain is regarded as R3 .
(2) F : U → O is a homeomorphism, meaning that F : U → O is bijective, and both
F : U → O and F −1 : O → U are continuous.
(3) For all (u, v) ∈ U, the cross product:
∂F ∂F
× 6= 0.
∂u ∂v
The coordinates (u, v) are called the local coordinates of M .
If F : U → M is of class C k for any integer k, then F is said to be a C ∞ (or smooth)
local parametrization.

Definition 1.2 (Surfaces of Class C k ). A subset M ⊂ R3 is called a C k surface, where

k ∈ N ∪ {∞}, in R3 if at every point p ∈ M , there exists an open subset U ⊂ R2 , an
open subset O ⊂ M containing p, and a C k local parametrization F : U → O which
satisfies all three conditions stated in Definition 1.1.
We say M is a regular surface in R3 if it is a C ∞ surface.

Figure 1.1. smooth local parametrization

To many students (myself included), the definition of regular surfaces looks obnox-
ious at the first glance. One way to make better sense of it is to look at some examples
and understand why each of the three conditions is needed in the definition.
The motivation behind condition (1) in the definition is that we are studying differen-
tial topology/geometry and so we want the parametrization to be differentiable as many
times as we like. Condition (2) rules out surfaces that have self-intersection such as the
Klein bottle (see Figure 1.2a). Finally, condition (3) guarantees the existence of a unique
tangent plane at every point on M (see Figure 1.2b for a non-example).
1.1. Local Parametrizations 5

(a) Klein Bottle has a self-intersection. (b) F (u, v) = (u3 , v 3 , uv) fails condition (3).

Figure 1.2. Examples of non-smooth parametrizations

Example 1.3 (Graph of a Function). Consider a smooth function f (u, v) : U → R

defined on an open subset U ⊂ R2 . The graph of f , denoted by Γf , is the subset
{(u, v, f (u, v)) : (u, v) ∈ U} of R3 . One can parametrize Γf by a global parametrization:
F (u, v) = (u, v, f (u, v)).
Condition (1) holds because f is given to be smooth. For condition (2), F is clearly
one-to-one, and the image of F is the whole graph Γf . Regarding it as a map F : U → Γf ,
the inverse map
F −1 (x, y, z) = (x, y)
is clearly continuous. Therefore, F : U → Γf is a homeomorphism. To verify condition
(3), we compute the cross product:
 
∂F ∂F ∂f ∂f
× = − , − , 1 6= 0
∂u ∂v ∂u ∂v
for all (u, v) ∈ U. Therefore, F is a smooth local parametrization of Γf . Since the image
of this single smooth local parametrization covers all of Γf , we conclude that Γf is a
regular surface. 

Figure 1.3. The graph of any smooth function is a regular surface.

6 1. Regular Surfaces

Exercise 1.1. Show that F (u, v) : (0, 2π) × (0, 1) → R3 defined by:
F (u, v) = (sin u, sin 2u, v)
satisfies conditions (1) and (3) in Definition 1.1, but not condition (2). [Hint: Try
to show F −1 is not continuous by finding a diverging sequence {(un , vn )} such that
{F (un , vn )} converges. See Figure 1.4 for reference.]

Figure 1.4. Plot of F (u, v) in Exercise 1.1

In Figure 1.3, one can observe that there are two families of curves on the surface.
These curves, often called coordinate curves, are obtained by varying one of the (u, v)-
variables while keeping the other constant. Precisely, they are the curves represented by
F (u, v0 ) and F (u0 , v) where u0 and v0 are fixed. As such, the partial derivatives ∂F
∂u (p)
∂F
and ∂v (p) give a pair of tangent vectors on the surface at point p. Therefore, their
cross product ∂F ∂F
∂u (p) × ∂v (p) is a normal vector to the surface at point p (see Figure
1.5). Here we have abused the notations for simplicity: ∂F ∂F
∂u (p) means ∂u evaluated at
(u, v) = F −1 (p). Similarly for ∂F
∂v (p).
Condition (3) requires that ∂F ∂F
∂u × ∂v is everywhere non-zero in the domain of F . An
equivalent statement is that the vectors ∂F ∂F
∂u (p), ∂v (p) are linearly independent for any
p ∈ F (U).

Figure 1.5. Tangent and normal vectors to a surface in R3

1.1. Local Parametrizations 7

Example 1.4 (Sphere). In R3 , the unit sphere S2 centered at the origin

p can be represented
by the equation x2 + y 2 + z 2 = 1, or in other words, z = ± 1 − x2 − y 2 . We can
parametrize the upper and lower hemisphere by two separate local maps:
p
F1 (u, v) = (u, v, 1 − u2 − v 2 ) : B1 (0) ⊂ R2 → S2+
p
F2 (u, v) = (u, v, − 1 − u2 − v 2 ) : B1 (0) ⊂ R2 → S2−
where B1 (0) = {(u, v) : u2 + v 2 < 1} is the open unit disk in R2 centered at the origin,
and S2+ and S2− are the upper and lower hemispheres of S2 respectively. Since B1 (0) is
√
open, the functions ± 1 − u2 − v 2 are smooth, and according to the previous example
both F1 and F2 are smooth local parametrizations.

Figure 1.6. A unit sphere covered by six parametrization charts

However, not all points on the sphere are covered by S2+ and S2− , since points on the
equator are not. In order to show that S2 is a regular surface, we need to write down
more smooth local parametrization(s) so that each point on the sphere can be covered by
at least one smooth local parametrization chart. One can construct four more smooth
local parametrizations (left, right, front and back) similar to F1 and F2 (see Figure 1.6).
It is left as an exercise for readers to write down the other four parametrizations. These
six parametrizations are all smooth and they cover the whole sphere. Therefore, it shows
the sphere is a regular surface. 

Exercise 1.2. Write down the left, right, front and back parametrizations Fi ’s
(i = 3, 4, 5, 6) of the sphere as shown in Figure 1.6. Indicate clearly the domain and
range of each Fi .
8 1. Regular Surfaces

Example 1.5 (Sphere: revisited). We can in fact cover the sphere by just two smooth
local parametrizations described below. Define F+ (u, v) : R2 → S2 \{(0, 0, 1)} where:
u2 + v 2 − 1
 
2u 2v
F+ (u, v) = , ,
u2 + v 2 + 1 u2 + v 2 + 1 u2 + v 2 + 1
It is called the stereographic parametrization of the sphere (see Figure 1.7) , which assigns
each point (u, v, 0) on the xy-plane of R3 to a point where the line segment joining
(u, v, 0) and the north pole (0, 0, 1) intersects the sphere. Clearly F+ is a smooth function.
We leave it as exercise for readers to verify that F+ satisfies condition (3) and that
F+−1 : S2 \{(0, 0, 1)} → R2 is given by:
 
−1 x y
F+ (x, y, z) = , .
1−z 1−z
As z 6= 1 for every (x, y, z) in the domain of F+−1 , it is a continuous function. Therefore,
F+ is a smooth local parametrization. The inverse map F+−1 is commonly called the
stereographic projection of the sphere.

Figure 1.7. Stereographic parametrization of the sphere

Note that the range of F+ does not include the point (0, 0, 1). In order to show
that the sphere is a regular surface, we need to cover it by another parametrization
F− : R2 → S2 \{(0, 0, −1)} which assigns each point (u, v, 0) on the xy-plane to a point
where the line segment joining (u, v, 0) and the south pole (0, 0, −1) intersects the sphere.
It is an exercise for readers to write down the explicit parametrization F− . 

Exercise 1.3. Verify that F+ in Example 1.4 satisfies condition (3) in Definition 1.1,
and that the inverse map F+−1 : S2 \{(0, 0, 1)} → R2 is given as stated. [Hint: Write
down F+ (u, v) = (x, y, z) and solve (u, v) in terms of (x, y, z). Begin by finding
u2 + v 2 in terms of z.]
Furthermore, write down explicitly the map F− described in Example 1.4, and
find its inverse map F−−1 .

Exercise 1.4. Find smooth local parametrizations which together cover the whole
ellipsoid:
x2 y2 z2
2
+ 2 + 2 =1
a b c
where a, b and c are positive constants.
1.1. Local Parametrizations 9

Exercise 1.5. Let M be the cylinder {(x, y, z) ∈ R3 : x2 + y 2 = 1}. The purpose

of this exercise is to construct a smooth local parametrization analogous to the
stereographic parametrization in Example 1.4:
Consider the unit circle x2 + y 2 = 1 on the xy-plane. For each point (u, 0) on
the x-axis, we construct a straight-line joining the point (0, 1) and (u, 0). This line
intersects the unit circle at a unique point p. Denote the xy-coordinates of p by
(x(u), y(u)).
(a) Find the coordinates (x(u), y(u)) in terms of u.
(b) Define:
F1 (u, v) = (x(u), y(u), v)
2
with R as its domain. Describe the image of F1 .
(c) Denote O1 to be the image of F1 . Verify that F1 : R2 → O1 is smooth local
parametrization of M .
(d) Construct another smooth local parametrization F2 such that the images of F1
and F2 cover the whole surface M (hence establish that M is a regular surface).

Let’s also look at a non-example of smooth local parametrizations. Consider the map:
G(u, v) = (u3 , v 3 , 0), (u, v) ∈ R × R.
It is a smooth, injective map from R onto the xy-plane Π of R3 , i.e. G : R2 → Π.
2

However, it can be computed that

∂G ∂G
(0, 0) = (0, 0) = 0
∂u ∂v
and so condition (3) in Definition 1.1 does not hold. The map G is not a smooth local
parametrization of Π. However, note that Π is a regular surface because F (u, v) = (u, v, 0)
is a smooth global parametrization of Π, even though G is not a “good” parametrization.
In order to show M is a regular surface, what we need is to show at every point
p ∈ M there is at least one smooth local parametrization F near p. However, to show
that M is not a regular surface, one then needs to come up with a point p ∈ M such that
there is no smooth local parametrization near that point p (which may not be easy).
10 1. Regular Surfaces

1.2. Level Surfaces

Many surfaces are defined using an equation such as x2 + y 2 + z 2 = 1, or x2 + y 2 = z 2 + 1.
They are level sets of a function g(x, y, z). In this section, we are going to prove a theorem
that allows us to show easily that some level sets g −1 (c) are regular surfaces.

Theorem 1.6. Let g(x, y, z) : R3 → R be a smooth function of three variables. Consider

a non-empty level set g −1 (c) where c is a constant. If ∇g(x0 , y0 , z0 ) 6= 0 at all points
(x0 , y0 , z0 ) ∈ g −1 (c), then the level set g −1 (c) is a regular surface.

Proof. The key idea of the proof is to use the Implicit Function Theorem. Given any
point p = (x0 , y0 , z0 ) ∈ g −1 (c), since ∇g(x0 , y0 , z0 ) 6= (0, 0, 0), at least one of the first
partials:
∂g ∂g ∂g
(p), (p), (p)
∂x ∂y ∂z
is non-zero. Without loss of generality, assume ∂g ∂z (p) 6= 0, then the Implicit Function
Theorem shows that locally around the point p, the level set g −1 (c) can be regarded as a
graph z = f (x, y) of some smooth function f of (x, y). To be precise, there exists an open
set O of g −1 (c) containing p such that there is a smooth function f (x, y) : U ⊂ R2 → R
from an open set U such that (x, y, f (x, y)) ∈ O ⊂ g −1 (c) for any (x, y) ∈ U. As such, the
smooth local parametrization F : U → O defined by:
F (u, v) = (u, v, f (u, v))
is a smooth local parametrization of g −1 (c).
∂g
In the case where ∂y (p) 6= 0, the above argument is similar as locally around p one
−1
can regard g (c) as a graph y = h(x, z) for some smooth function h. Similar in the case
∂g
∂x (p) 6= 0.
Since every point p can be covered by the image of a smooth local parametrization,
the level set g −1 (c) is a regular surface. 
Example 1.7. The unit sphere x2 +y 2 +z 2 = 1 is a level surface g −1 (1) where g(x, y, z) :=
x2 + y 2 + z 2 . The gradient vector ∇g = (2x, 2y, 2z) is zero only when (x, y, z) = (0, 0, 0).
Since the origin is not on the unit sphere, we have ∇g(x0 , y0 , z0 ) 6= (0, 0, 0) for any
(x0 , y0 , z0 ) ∈ g −1 (1). Therefore, the unit sphere is a regular surface.
Similarly, one can also check that the surface x2 + y 2 = z 2 + 1 is a regular surface.
It is a level set h−1 (1) where h(x, y, z) = x2 + y 2 − z 2 . Since ∇h = (2x, 2y, −2z), the
origin is the only point p at which ∇h(p) = (0, 0, 0) and it is not on the level set h−1 (1).
Therefore, h−1 (1) is a regular surface. 

However, the cone x2 + y 2 = z 2 cannot be shown to be a regular surface using

Theorem 1.6. It is a level surface h−1 (0) where h(x, y, z) := x2 + y 2 − z 2 . The origin
(0, 0, 0) is on the cone and ∇h(0, 0, 0) = (0, 0, 0). Theorem 1.6 fails to give any conclusion.
The converse of Theorem 1.6 is not true. Consider g(x, y, z) = z 2 , then g −1 (0) is the
xy-plane which is clearly a regular surface. However, ∇g = (0, 0, 2z) is zero at the origin
which is contained in the xy-plane.
1.2. Level Surfaces 11

Exercise 1.6. [dC76, P.66] Let f (x, y, z) = (x + y + z − 1)2 . For what values of c is
the set f −1 (c) a regular surface?

Exercise 1.7. A torus is defined by the equation:

p 2
z 2 = R2 − x2 + y 2 − r
where R > r > 0 are constants. Show that it is a regular surface.

The proof of Theorem 1.6 makes use of the Implicit Function Theorem which is an
existence result. It shows a certain level set is a regular surface, but it fails to give an
explicit smooth local parametrization around each point.
There is one practical use of Theorem 1.6 though. Suppose we are given F (u, v)
which satisfies conditions (1) and (3) in Definition 1.1 and that F is continuous and
F −1 exists. In order to verify that it is a smooth local parametrization, we need to prove
continuity of F −1 , which is sometimes difficult. Here is one example:
F (u, v) = (sin u cos v, sin u sin v, cos u), 0 < u < π, 0 < v < 2π
is a smooth local parametrization of a unit sphere. It is clearly a smooth map from
(0, π) × (0, 2π) ⊂ R2 to R3 , and it is quite straight-forward to verify condition (3) in
Definition 1.1 and that F is one-to-one. However, it is rather difficult to write down an
explicit F −1 , let alone to show it is continuous.
The following result tells us that if the surface is given by a level set satisfying
conditions stated in Theorem 1.6, and F satisfies conditions (1) and (3), then F −1 is
automatically continuous. Precisely, we have the following:

Proposition 1.8. Assume all given conditions stated in Theorem 1.6. Furthermore, suppose
F (u, v) is a bijective map from an open set U ⊂ R2 to an open set O ⊂ M := g −1 (c) which
satisfies conditions (1) and (3) in Definition 1.1. Then, F satisfies condition (2) as well
and hence is a smooth local parametrization of g −1 (c).

Proof. Given any point p ∈ g −1 (c), we can assume without loss of generality that
∂g
∂z (p) 6= 0. Recall from Multivariable Calculus that ∇g(p) is a normal vector to the level
surface g −1 (c) at point p. Furthermore, if F (u, v) is a map satisfying conditions (1) and
−1
(3) of Definition 1.1, then ∂F ∂F
∂u (p) × ∂v (p) is also a normal vector to g (c) at p.
Now that the k̂-component of ∇g(p) is non-zero since ∂g
∂z (p) 6= 0, so the k̂-component
∂F ∂F
of the cross product ∂u (p) × ∂v (p) is also non-zero. If we express F (u, v) as:
F (u, v) = (x(u, v), y(u, v), z(u, v)),
∂F ∂F
then the k̂-component of ∂u (p) ×(p) is given by:
∂v
∂x ∂x
 
∂x ∂y ∂y ∂x
− (p) = ∂u
∂y
∂v
∂y (p).
∂u ∂v ∂u ∂v ∂u ∂v
Define π : R3 → R2 by π(x, y, z) = (x, y). The above shows that the composition π ◦ F
given by
(π ◦ F )(u, v) = (x(u, v), y(u, v))
has non-zero Jacobian determinant at p. By the Inverse Function Theorem, π ◦ F has a
smooth local inverse near p. In particular, (π ◦ F )−1 is continuous near p.
Finally, by the fact that (π ◦ F ) ◦ F −1 = π and that (π ◦ F )−1 exists and is continuous
locally around p, we can argue that F −1 = (π ◦ F )−1 ◦ π is also continuous near p. It
completes the proof. 
12 1. Regular Surfaces

∂g
Exercise 1.8. Rewrite the proof of Proposition 1.8 by assuming ∂y (p) 6= 0 instead.

Example 1.9. We have already shown that the unit sphere x2 + y 2 + z 2 = 1 is a regular
surface using Theorem 1.6 by regarding it is the level set g −1 (1) where g(x, y, z) =
x2 + y 2 + z 2 . We also discussed that
F (u, v) = (sin u cos v, sin u sin v, cos u), 0 < u < π, 0 < v < 2π
is a possible smooth local parametrization. It is clearly smooth, and by direct computation,
one can show
∂F ∂F
× = sin u (sin u cos v, sin u sin v, cos u)
∂u ∂v
∂F ∂F
and so ∂u × ∂v = sin u 6= 0 for any (u, v) in the domain (0, π) × (0, 2π). We leave it as
an exercise for readers to verify that F is one-to-one (and so bijective when its codomain
is taken to be its image).
Condition (2) is not easy to verify because it is difficult to write down the inverse map
F −1 explicitly. However, thanks for Proposition 1.8, F is a smooth local parametrization
since it satisfies conditions (1) and (3), and it is one-to-one. 

Exercise 1.9. Consider that the Mercator projection of the unit sphere:
 
cos v sin v sinh u
F (u, v) = , ,
cosh u cosh u cosh u
where sinh u := 21 (eu − e−u ) and cosh u := 21 (eu + e−u ).
(a) What are the domain and range of F ?
(b) Show that F is a smooth local parametrization.

Exercise 1.10. Consider the following parametrization of a torus T2 :

F (u, v) = ((r cos u + R) cos v, (r cos u + R) sin v, r sin u)
where (u, v) ∈ (0, 2π) × (0, 2π), and R > r > 0 are constants. Show that F is a
smooth local parametrization.
1.3. Transition Maps 13

1.3. Transition Maps

Let M ⊂ R3 be a regular surface, and Fα (u1 , u2 ) : Uα → M and Fβ (v1 , v2 ) : Uβ → M be
two smooth local parametrizations of M with overlapping images, i.e. W := Fα (Uα ) ∩
Fβ (Uβ ) 6= ∅. Under this set-up, it makes sense to define the maps Fβ−1 ◦ Fα and Fα−1 ◦ Fβ .
However, we need to shrink their domains so as to guarantee they are well-defined.
Precisely:
(Fβ−1 ◦ Fα ) : Fα−1 (W) → Fβ−1 (W)
(Fα−1 ◦ Fβ ) : Fβ−1 (W) → Fα−1 (W)

Note that Fα−1 (W) and Fβ−1 (W) are open subsets of Uα and Uβ respectively. The
map Fβ−1 ◦ Fα describes a relation between two sets of coordinates (u1 , u2 ) and (v1 , v2 )
of M . In other words, one can regard Fβ−1 ◦ Fα as a change-of-coordinates, or transition
map and we can write:
Fβ−1 ◦ Fα (u1 , u2 ) = (v1 (u1 , u2 ), v2 (u1 , u2 )).

Figure 1.8. Transition maps

One goal of this section is to show that this transition map Fβ−1 ◦ Fα is smooth
provided that Fα and Fβ are two overlapping smooth local parametrizations. Before we
present the proof, let us look at some examples of transition maps.
14 1. Regular Surfaces

Example 1.10. The xy-plane Π in R3 is a regular surface which admits a global smooth
parametrization Fα (x, y) = (x, y, 0) : R2 → Π. Another way to locally parametrize Π is
by polar coordinates Fβ : (0, ∞) × (0, 2π) → Π
Fβ (r, θ) = (r cos θ, r sin θ, 0)
Readers should verify that they are smooth local parametrizations. The image of Fα
is the entire xy-plane Π, whereas the image of Fβ is the xy-plane with the origin and
positive x-axis removed. The transition map Fα−1 ◦ Fβ is given by:
Fα−1 ◦ Fβ : (0, ∞) × (0, 2π) → R2 \{(x, 0) : x ≥ 0}
(r, θ) 7→ (r cos θ, r sin θ)
To put it in a simpler form, we can say (x(r, θ), y(r, θ)) = (r cos θ, r sin θ). 

Exercise 1.11. Consider the stereographic parametrizations F+ and F− in Example

1.5. Compute the transition maps F+−1 ◦ F− and F−−1 ◦ F+ . State the maximum
possible domain for each map. Are they smooth on their domains?

Exercise 1.12. The unit cylinder Σ2 in R3 can be covered by two local parametriza-
tions:
F : (0, 2π) × R → Σ2 Fe : (−π, π) × R → Σ2
F (θ, z) := (cos θ, sin θ, z) Fe(θ,
e ze) := (cos θ,
e sin θ,
e ze)

Compute the transition maps F −1 ◦ Fe and Fe−1 ◦ F . State their maximum possible
domains. Are they smooth on their domains?

Exercise 1.13. The Möbius strip Σ2 in R3 can be covered by two local parametriza-
tions:
F : (−1, 1) × (0, 2π) → Σ2 Fe : (−1, 1) × (−π, π) → Σ2
  
θe
 θ


3 + u cos 2
cos θ
 3 + u
e cos 2 cos θe
  
F (u, θ) =  3 + u cos θ2 sin θ  Fe(e
u, θ)
e = 3+u
 e cos θ2 sin θe
e

u sin θ2 θe
u
e sin 2
Compute the transition maps, state their maximum possible domains and verify that
they are smooth.

The proposition below shows that the transition maps between any pair of smooth
local parametrizations are smooth:

Proposition 1.11. Let M ⊂ R3 be a regular surface, and Fα (u1 , u2 ) : Uα → M and

Fβ (v1 , v2 ) : Uβ → M be two smooth local parametrizations of M with overlapping images,
i.e. W := Fα (Uα ) ∩ Fβ (Uβ ) 6= ∅. Then, the transition maps defined below are also smooth
maps:
(Fβ−1 ◦ Fα ) : Fα−1 (W) → Fβ−1 (W)
(Fα−1 ◦ Fβ ) : Fβ−1 (W) → Fα−1 (W)

Proof. It suffices to show Fβ−1 ◦ Fα is smooth as the other one Fα−1 ◦ Fβ can be shown
by symmetry. Furthermore, since differentiability is a local property, we may fix a point
p ∈ W ⊂ M and show that Fβ−1 ◦ Fα is smooth at the point Fα−1 (p).
1.3. Transition Maps 15

By condition of (3) of smooth local parametrizations, we have:

∂Fα ∂Fα
(p) × (p) 6= 0
∂u1 ∂u2
By straight-forward computations, one can show that this cross product is given by:
 
∂Fα ∂Fα ∂(y, z) ∂(z, x) ∂(x, y)
× = det (p), det (p), det (p) .
∂u1 ∂u2 ∂(u1 , u2 ) ∂(u1 , u2 ) ∂(u1 , u2 )
Hence, at least one of the determinants is non-zero. Without loss of generality, assume
that:
∂(x, y)
det (p) 6= 0.
∂(u1 , u2 )
∂Fα ∂Fα ∂Fβ ∂Fβ
Both (p) × (p) and (p) × (p) are normal vectors to the surface at p.
∂u1 ∂u2 ∂v1 ∂v2
Given that the former has non-zero k̂-component, then so does the latter. Therefore, we
have:
∂(x, y)
det (p) 6= 0.
∂(v1 , v2 )
Then we proceed as in the proof of Proposition 1.8. Define π(x, y, z) = (x, y), then
π ◦ F β : U β → R2
(v1 , v2 ) 7→ (x(v1 , v2 ), y(v1 , v2 ))
∂(x, y)
has non-zero Jacobian determinant det at p. Therefore, by the Inverse Function
∂(v1 , v2 )
Theorem, (π ◦ Fβ )−1 exists and is smooth near p. Since Fβ−1 ◦ Fα = (π ◦ Fβ )−1 ◦ (π ◦ Fα ),
and all of (π ◦ Fβ )−1 , π and Fα are smooth maps, their composition is also a smooth
map. We have proved Fβ−1 ◦ Fα is smooth near p. Since p is arbitrary, Fβ−1 ◦ Fα is in fact
smooth on the domain Fα−1 (W). 

Exercise 1.14. Rewrite the proof of Proposition 1.11, mutatis mutandis, by assuming
∂(y, z)
det (p) 6= 0 instead.
∂(u1 , u2 )
16 1. Regular Surfaces

1.4. Maps and Functions from Surfaces

Let M be a regular surface in R3 with a smooth local parametrization F (u1 , u2 ) : U → M .
Then, for any p ∈ F (U), one can define the partial derivatives for a function f : M → R
at p as follows. The subtle issue is that the domain of f is the surface M , but by pre-
composing f with F , i.e. f ◦ F , one can regard it as a map from U ⊂ R2 to R. With a
little abuse of notations, we denote:
∂f ∂(f ◦ F )
(p) := (u1 , u2 )
∂uj ∂uj
where (u1 , u2 ) is the point corresponding to p, i.e. F (u1 , u2 ) = p.
∂f
Remark 1.12. Note that ∂u j
(p) is defined locally on F (U), and depends on the choice
of local parametrization F near p. 

Definition 1.13 (Functions of Class C k ). Let M be a regular surface in R3 , and f :

M → R be a function defined on M . We say f is C k at p ∈ M if for any smooth local
parametrization F : U → M with p ∈ F (U), the composition f ◦ F is C k at (u1 , u2 )
corresponding to p.
If f is C k at p for any p ∈ M , then we say that f is a C k function on M . Here k can
be taken to be ∞, and in such case we call f to be a C ∞ (or smooth) function.

Remark 1.14. Although we require f ◦F to be C k at p ∈ M for any local parametrization

F in order to say that f is C k , by Proposition 1.11 it suffices to show that f ◦ F is C k at
p for at least one F near p. It is because
f ◦ Fe = (f ◦ F ) ◦ (F −1 ◦ Fe)
and compositions of C k maps (between Euclidean spaces) are C k . 
Example 1.15. Let M be a regular surface in R3 , then each of the x, y and z coordinates
in R3 can be regarded as a function from M to R. For any smooth local parametrization
F : U → M around p given by
F (u1 , u2 ) = (x(u1 , u2 ), y(u1 , u2 ), z(u1 , u2 )),
we have x ◦ F (u1 , u2 ) = x(u1 , u2 ). Since F is C ∞ , we get x ◦ F is C ∞ as well. Therefore,
the coordinate functions x, y and z for any regular surface is smooth. 
Example 1.16. Let f : M → R be the function from a regular surface M in R3 defined
by:
2
f (p) := |p − p0 |
where p0 = (x0 , y0 , z0 ) is a fixed point of R3 . Suppose F (u, v) is a local parametrization
∂f
of M . We want to compute ∂u and ∂f
∂v .
Write (x, y, z) = F (u, v) so that x, y and z are functions of (u, v). Then
∂f ∂
:= (f ◦ F )
∂u ∂u
∂
= f (x(u, v), y(u, v), z(u, v))
∂u
∂
(x(u, v) − x0 )2 + (y(u, v) − y0 )2 + (z(u, v) − z0 )2


=
∂u
∂x ∂y ∂z
= 2(x − x0 ) + 2(y − y0 ) + 2(z − z0 )
∂u ∂u ∂u
1.4. Maps and Functions from Surfaces 17

Note that we can differentiate x, y and z by u because F (u, v) is smooth. Similarly, we

have:
∂f ∂x ∂y ∂z
= 2(x − x0 ) + 2(y − y0 ) + 2(z − z0 ) .
∂v ∂v ∂v ∂v
Again since F (u, v) (and hence x, y and z) is a smooth function of (u, v), we can
∂f
differentiate ∂u and ∂f
∂v as many times as we wish. This concludes that f is a smooth
funciton. 

Exercise 1.15. Let p0 (x0 , y0 , z0 ) be a point in R3 and let f (p) = |p − p0 | be the

Euclidean distance between p and p0 in R3 . Suppose M is a regular surface in R3 ,
one can then restrict the domain of f to M and consider it as a function:
f :M →R
p 7→ |p − p0 |
Under what condition is the function f : M → R smooth?

Now let M and N be two regular surfaces in R3 . Then, one can also talk about
mappings Φ : M → N between them. In this section, we will define the notion of smooth
maps between two surfaces.
Suppose F : UM → M and G : UN → N are two smooth local parametrizations of M
and N respectively. One can then consider the composition G−1 ◦ Φ ◦ F after shrinking
the domain. It is then a map between open subsets of R2 .
However, in order for this composition to be well-defined, we require the image
of Φ ◦ F to be contained in the image of G, which is not always guaranteed. Let
W := Φ(OM ) ∩ ON be the overlapping region on N of these two images. Then, provided
6 ∅, the composition G−1 ◦ Φ ◦ F becomes well-defined as a map on:
that W =

G−1 ◦ Φ ◦ F : (Φ ◦ F )−1 (W) → UN .

From now on, whenever we talk about this composition G−1 ◦ Φ ◦ F , we always implicitly
assume that W 6= ∅ and its domain is (Φ ◦ F )−1 (W).

Figure 1.9. maps between regular surfaces

18 1. Regular Surfaces

Definition 1.17 (Maps of Class C k ). Let M and N be two regular surfaces in R3 , and
Φ : M → N be a map between them. We say Φ is C k at p ∈ M if for any smooth local
parametrization F : UM → M with p ∈ F (UM ), and G : UN → N with Φ(p) ∈ G(UN ),
the composition G−1 ◦ Φ ◦ F is C k at F −1 (p) as a map between subsets of R2 .
If Φ is C k at p for any p ∈ M , then we say that Φ is C k on M . Here k can be taken
to be ∞, and in such case we call Φ to be C ∞ (or smooth) on M .

Remark 1.18. Although we require G−1 ◦ Φ ◦ F to be C k at p ∈ M for any local

parametrizations F and G in order to say that Φ is C k , by Proposition 1.11 it suffices to
show that G−1 ◦ Φ ◦ F is C k at p for at least one pair of F and G. It is because
Ge −1 ◦ Φ ◦ Fe = (G
e −1 ◦ G) ◦ (G−1 ◦ Φ ◦ F ) ◦ (F −1 ◦ Fe)
and compositions of C k maps (between Euclidean spaces) are C k . 
Example 1.19. Let S2 be the unit sphere in R3 . Consider the antipodal map Φ : S2 → S2
taking P to −P . In Example 1.4, two of the local parametrizations are given by:
q
F1 (u1 , u2 ) = (u1 , u2 , 1 − u21 − u22 ) : B1 (0) ⊂ R2 → S2+
q
F2 (v1 , v2 ) = (v1 , v2 , − 1 − v12 − v22 ) : B1 (0) ⊂ R2 → S2−
where B1 (0) is the open unit disk in R2 centered at the origin, and S2+ and S2− are the
upper and lower hemispheres of S2 respectively. One can compute that:
 q 
F2−1 ◦ Φ ◦ F1 (u1 , u2 ) = F2−1 ◦ Φ u1 , u2 , 1 − u21 − u22
 q 
= F2−2 −u1 , −u2 , − 1 − u21 − u22

= (−u1 , −u2 )
Clearly, the map (u1 , u2 ) 7→ (−u1 , −u2 ) is C ∞ . It shows the antipodal map Φ is C ∞ at
every point in F1 (B1 (0)). One can show in similar way using other local parametrizations
that Φ is C ∞ at points on S2 not covered by F1 .
Note that, for instance, the images of Φ ◦ F1 and F1 are disjoint, and so F1−1 ◦ Φ ◦ F1
is not well-defined. We don’t need to verify whether it is smooth. 

Exercise 1.16. Let Φ be the antipodal map considered in Example 1.19, and F+
and F− be the two stereographic parametrizations of S2 defined in Example 1.5.
Compute the maps F+−1 ◦ Φ ◦ F+ , F−−1 ◦ Φ ◦ F+ , F+−1 ◦ Φ ◦ F− and F−−1 ◦ Φ ◦ F− .
State their domains, and verify that they are smooth on their domains.

Exercise 1.17. Denote S2 to be the unit sphere x2 + y 2 + z 2 = 1. Let Φ : S2 → S2

be the rotation map about the z-axis defined by:
Φ(x, y, z) = (x cos α − y sin α, x sin α + y cos α, z)
where α is a fixed angle. Show that Φ is smooth.
1.4. Maps and Functions from Surfaces 19

Let M and N be two regular surfaces. If a map Φ : M → N is C ∞ , invertible, with

∞
C inverse map Φ−1 : N → M , then we say:

Definition 1.20 (Diffeomorphisms). A map Φ : M → N between two regular surfaces

M and N in R3 is said to be a diffeomorphism if Φ is C ∞ and invertible, and also the
inverse map Φ−1 is C ∞ . If such a map Φ exists between M and N , then we say the
surfaces M and N are diffeomorphic.

Example 1.21. The antipodal map Φ : S2 → S2 described in Example 1.19 is a diffeo-

morphism between S2 and itself. 
x2 y2 z2
Example 1.22. The sphere x2 + y 2 + z 2 = 1 and the ellipse + + = 1 are
a2 b2 c2
2
diffeomorphic, under the map Φ(x, y, z) = (ax, by, cz) restricted on S . 

Exercise 1.18. Given any pair of C ∞ functions f, g : R2 → R, show that the graphs
Γf and Γg are diffeomorphic.

Exercise 1.19. Show that Φ : S2 → S2 defined in Exercise 1.17 is a diffeomorphism.

20 1. Regular Surfaces

1.5. Tangent Planes and Tangent Maps

1.5.1. Tangent Planes of Regular Surfaces. The tangent plane is an important
geometric object associated to a regular surface. Condition (3) of a smooth local
parametrization F (u, v) requires that the cross-product ∂F ∂F
∂u × ∂v is non-zero for any (u, v)
in the domain, or equivalently, both tangent vectors ∂F ∂F
∂u and ∂v must be non-zero vectors
and they are non-parallel to each other.
Therefore, the two vectors ∂F ∂F 3
∂u and ∂v span a two-dimensional subspace in R . We
call this subspace the tangent plane, which is defined rigorously as follows:

Definition 1.23 (Tangent Plane). Let M be a regular surface in R3 and p be a point on

M . Suppose F (u, v) : U ⊂ R2 → M is a smooth local parametrization around p, then
the tangent plane at p, denoted by Tp M , is defined as follows:
   
∂F ∂F ∂F ∂F
Tp M := span (p), (p) = a (p) + b (p) : a, b ∈ R .
∂u ∂v ∂u ∂v
∂F ∂F
Here we have abused the notations for simplicity: ∂u (p) means ∂u evaluated at
(u, v) = F −1 (p). Similarly for ∂F
∂v (p).

Rigorously, Tp M is a plane passing through the origin while p + Tp M is the plane

tangent to the surface at p (see Figure 1.10). The difference between Tp M and p + Tp M
is very subtle, and we will almost neglect this difference.

Figure 1.10. Tangent plane p + Tp M at p ∈ M

Exercise 1.20. Show that the equation of the tangent plane p + Tp M of the graph
of a smooth function f (x, y) at p = (x0 , y0 , f (x0 , y0 )) is given by:
∂f ∂f
z = f (x0 , y0 ) + (x − x0 ) + (y − y0 )
∂x (x0 ,y0 ) ∂y (x0 ,y0 )

Exercise 1.21. [dC76, P.88] Consider the surface M given by z = xf (y/x), where
x 6= 0 and f is a smooth function. Show that the tangent planes p + Tp M must pass
through the origin (0, 0, 0).

1.5.2. Tangent Maps between Regular Surfaces. Given a smooth map Φ : M →

N between two regular surfaces M and N , there is a naturally defined map called the
tangent map, denoted by Φ∗ in this course, between the tangent planes Tp M and TΦ(p) N .
Let us consider a smooth local parametrization F (u1 , u2 ) : UM → M . The compo-
sition Φ ◦ F can be regarded as a map from UM to R3 , so one can talk about its partial
1.5. Tangent Planes and Tangent Maps 21

∂(Φ◦F )
derivatives ∂ui :

∂Φ ∂(Φ ◦ F ) d
(Φ(p)) := = Φ ◦ F ((u1 , u2 ) + têi )
∂ui ∂ui (u1 ,u2 ) dt t=0

where (u1 , u2 ) is a point in UM such that F (u1 , u2 ) = p. The curve F ((u1 , u2 ) + têi ) is a
curve on M with parameter t along the ui -direction. The curve Φ ◦ F ((u1 , u2 ) + têi ) is
then the image of the ui -curve of M under the map Φ (see Figure 1.11). It is a curve on
∂Φ
N so ∂u i
which is a tangent vector to the surface N .

Figure 1.11. Partial derivative of the map Φ : M → N

Exercise 1.22. Denote S2 to be the unit sphere x2 + y 2 + z 2 = 1. Let Φ : S2 → S2

be the rotation map about the z-axis defined by:
Φ(x, y, z) = (x cos α − y sin α, x sin α + y cos α, z)
where α is a fixed angle. Calculate the following partial derivatives under the given
local parametrizations:
∂Φ ∂Φ
(a) and under F (θ, ϕ) = (sin ϕ cos θ, sin ϕ sin θ, cos ϕ);
∂θ ∂ϕ
∂Φ ∂Φ
(b) and under F2 in Example 1.4;
∂u ∂v
∂Φ ∂Φ
(c) and under F+ in Example 1.5.
∂u ∂v
∂Φ
Next, we write the partial derivative ∂ui in a fancy way. Define:
 
∂F ∂Φ
Φ∗ := .
∂ui ∂ui

∂F
Then, one can regard Φ∗ as a map that takes the tangent vector ∂u i
in Tp M to another
n o
∂Φ ∂F
vector ∂uj in TΦ(p) N . Since ∂ui (p) is a basis of Tp M , one can then extend Φ∗ linearly
and define it as the tangent map of Φ. Precisely, we have:
22 1. Regular Surfaces

Definition 1.24 (Tangent Maps). Let Φ : M → N be a smooth map between two

regular surfaces M and N in R3 . Let F : UM → M and G : UN → N be two smooth
local parametrizations covering p and Φ(p) respectively. Then, the tangent map of Φ at
p ∈ M is denoted by (Φ∗ )p and is defined as:
(Φ∗ )p : Tp M → TΦ(p) N
2
! 2
X ∂F X ∂Φ
(Φ∗ )p ai (p) = ai (Φ(p))
i=1
∂u i i=1
∂ui

If the point p is clear from the context, (Φ∗ )p can be simply denoted by Φ∗ .

Remark 1.25. Some textbooks may use dΦp to denote the tangent map of Φ at p. 
Example 1.26. Consider the unit sphere S2 locally parametrized by
F (θ, ϕ) = (sin ϕ cos θ, sin ϕ sin θ, cos ϕ)
and the rotation map:
Φ(x, y, z) = (x cos α − y sin α, x sin α + y cos α, z)
From Exercise 1.22, one should have figured out that:
∂Φ
= (− sin ϕ sin(θ + α), sin ϕ cos(θ + α), 0)
∂θ
∂Φ
= (cos ϕ cos(θ + α), cos ϕ sin(θ + α), − sin ϕ)
∂ϕ
 
∂F ∂F
Next we want to write them in terms of the basis , . However, we should
∂θ ∂ϕ
be careful about the base points of these vectors. Consider a point p ∈ S2 with local
∂Φ ∂Φ
coordinates (θ, ϕ), the vectors and computed above are based at the point Φ(p)
∂θ ∂ϕ
with
 local coordinates (θ+ α, ϕ). Therefore, weshould express them in terms of the basis
∂F ∂F ∂F ∂F
(Φ(p)), (Φ(p)) , not (p), (p) !
∂θ ∂ϕ ∂θ ∂ϕ
At Φ(p), we have:
∂F ∂Φ
(Φ(p)) = (− sin ϕ sin(θ + α), sin ϕ cos(θ + α), 0) = (Φ(p))
∂θ ∂θ
∂F ∂Φ
(Φ(p)) = (cos ϕ cos(θ + α), cos ϕ sin(θ + α), − sin ϕ) = (Φ(p))
∂ϕ ∂ϕ
Therefore, the tangent map (Φ∗ )p acts on the basis vectors by:
 
∂F ∂F
(Φ∗ )p (p) = (Φ(p))
∂θ ∂θ
 
∂F ∂F
(Φ∗ )p (p) = (Φ(p))
∂ϕ ∂ϕ
In other words, the matrix representation [(Φ∗ )p ] with respect to the bases
   
∂F ∂F 2 ∂F ∂F
(p), (p) for Tp S (Φ(p)), (Φ(p)) for TΦ(p) S2
∂θ ∂ϕ ∂θ ∂ϕ
is the identity matrix. However, it is not perfectly correct to say (Φ∗ )p is an identity map,
since the domain and co-domain are different tangent planes. 
1.5. Tangent Planes and Tangent Maps 23

Exercise 1.23. Let Φ be as in Example 1.26. Consider the stereographic parametriza-

tion F+ (u, v) defined in Example 1.5. Suppose
 p ∈ S2 , express
 the matrixrepresen-
∂F+ ∂F+ ∂F+ ∂F+
tation [(Φ∗ )p ] with respect to the bases , and ,
∂u ∂v p ∂u ∂v Φ(p)

1.5.3. Tangent Maps and Jacobian Matrices. Let Φ : M → N be a smooth map

between two regular surfaces. Instead of computing the matrix representation of the
tangent map Φ∗ directly by taking partial derivatives (c.f. Example 1.26), one can also
find it out by computing a Jacobian matrix.
Suppose F (u1 , u2 ) : UM → M and G(v1 , v2 ) : UN → N are local parametrizations
of M and N . The composition G−1 ◦ Φ ◦ F can be regarded as a map between the
u1 u2 -plane to the v1 v2 -plane. As such, one can write
G−1 ◦ Φ ◦ F (u1 , u2 ) = (v1 (u1 , u2 ), v2 (u1 , u2 )).
By considering Φ ◦ F (u1 , u2 ) = G(v1 (u1 , u2 ), v2 (u1 , u2 )), one can differentiate both sides
with respect to ui :
2
∂ ∂ X ∂G ∂vk
(1.1) (Φ ◦ F ) = G(v1 (u1 , u2 ), v2 (u1 , u2 )) = .
∂ui ∂ui ∂vk ∂ui
k=1
n o
∂G
Here we used the chain rule. Note that ∂v k
is a basis for TΦ(p) N .
Using (1.1), one can see:
 
∂F ∂Φ ∂ ∂v1 ∂G ∂v2 ∂G
Φ∗ := = (Φ ◦ F ) = +
∂u1 ∂u1 ∂u1 ∂u1 ∂v1 ∂u1 ∂v2
 
∂F ∂Φ ∂ ∂v1 ∂G ∂v2 ∂G
Φ∗ := = (Φ ◦ F ) = +
∂u2 ∂u2 ∂u2 ∂u2 ∂v1 ∂u2 ∂v2
 
∂F
Hence the matrix representation of (Φ∗ )p with respect to the bases (p) and
  ∂ui
∂G
(Φ(p)) is the Jacobian matrix:
∂vi
" #
∂v1 ∂v1
∂(v1 , v2 ) ∂u1 ∂u2
= ∂v2 ∂v2
∂(u1 , u2 ) F −1 (p) ∂u1 ∂u2 −1 F (p)

Example 1.27. Let Φ : S2 → S2 be the rotation map as in Example 1.26. Consider again
the local parametrization:
F (θ, ϕ) = (sin ϕ cos θ, sin ϕ sin θ, cos ϕ).
By standard trigonometry, one can find out that Φ(F (θ, ϕ)) = F (θ + α, ϕ). Equivalently,
the map F −1 ◦ Φ ◦ F (in a suitable domain) is the map:
(θ, ϕ) 7→ (θ + α, ϕ).
As α is a constant, the Jacobian matrix of F −1n◦ Φ ◦ F ois the identity
n matrix,
o and
so the matrix [(Φ∗ )p ] with respect to the bases ∂F ,
∂θ ∂ϕ
∂F
and ∂F ∂F
,
∂θ ∂ϕ is the
p Φ(p)
identity matrix (which was also obtained by somewhat tedious computations in Example
1.26). 

Exercise 1.24. Do Exercise 1.23 by considering Jacobian matrices.

 Chapter 2

Abstract Manifolds

“Manifolds are a bit like pornography:

hard to define, but you know one when
you see one.”

Shmuel Weinberger

2.1. Smooth Manifolds

Intuitively, a manifold is a space which locally resembles an Euclidean space. Regular
surfaces are examples of manifolds. Being locally Euclidean, a manifold is equipped
with a local coordinate system around every point so that many concepts in Calculus on
Euclidean spaces can carry over to manifolds.
Unlike regular surfaces, we do not require a manifold to be a subset of Rn . A manifold
can just stand alone by itself like the Universe is regarded as a curved space-time sheet
with nothing “outside” in General Relativity. However, we do require that a manifold
satisfies certain topological conditions.

2.1.1. Point-Set Topology. In order to state the formal definition of a manifold,

there are some topological terms (such as Hausdorff, second countable, etc.) we will
briefly introduce. However, we will not take a long detour to go through every single
topological concept, otherwise we will not have time to cover the more interesting
material about smooth manifolds. Moreover, these topological conditions are very
common as long as the space we are looking at is not “strange”.
A topological space X is a set equipped with a collection T of subsets of X such that:
(a) ∅, X ∈ T ; and
[
(b) for any arbitrary sub-collection {Uα }α∈A ⊂ T , we have Uα ∈ T ; and
α∈A
N
\
(c) for any finite sub-collection {U1 , . . . , UN } ⊂ T , we have Uα ∈ T .
i=1

If T is such a collection, we call T a topology of X. Elements in T are called open sets of

X.

25
26 2. Abstract Manifolds

Example 2.1. The Euclidean space Rn equipped with the collection

T = collection of all open sets (in usual sense) in Rn
is an example of a topological space. The collection T is called the usual topology of
Rn . 
Example 2.2. Any subset S ⊂ Rn , equipped with the collection
TS = {S ∩ U : U is an open set (in usual sense) in Rn }
is an example of a topological space. The collection TS is called the subspace topology. 

Given two topological spaces (X, TX ) and (Y, TY ), one can talk about functions or
mapping between them. A map Φ : X → Y is said to be continuous with respect to TX
and TY if for any U ∈ TY , we have Φ−1 (U ) ∈ TX . This definition is a generalization of
continuous functions between Euclidean spaces equipped with the usual topologies. If
the map Φ : X → Y is one-to-one and onto, and both Φ and Φ−1 are continuous, then
we say Φ is a homeomorphism and the spaces (X, TX ) and (Y, TY ) are homeomorphic.
A topological space (X, T ) is said to be Hausdorff if for any pair of distinct points
p, q ∈ X, we have U1 , U2 ∈ T such that p ∈ U1 , q ∈ U2 and U1 ∩ U2 = ∅. In other words,
points of a Hausdorff space can be separated by open sets. It is intuitive that Rn with the
usual topology is a Hausdorff space. Any subset S ⊂ Rn with subspace topology is also a
Hausdorff space.
A topological space (X, TX ) is said to be second countable if there is a countable
sub-collection {Ui }∞ i=1 ⊂ T such that any set U ∈ T can be expressed as a union of some
of these Ui ’s. For instance, Rn with usual topology is second countable since by density
of rational numbers, any open set can be expressed as a countable union of open balls
with rational radii and centers.
This introduction to point-set topology is intended to be short. It may not make sense
to everybody, but it doesn’t hurt! Point-set topology is not the main dish of the course.
Many spaces we will look at are either Euclidean spaces, their subsets or sets derived
from Euclidean spaces. Most of them are Hausdorff and second countable. Readers
who want to learn more about point-set topology may consider taking MATH 4225. For
more thorough treatment on point-set topology, please consult [Mun00]. Meanwhile,
the take-home message of this introduction is that we don’t have to worry much about
point-set topology in this course!

2.1.2. Definitions and Examples. Now we are ready to learn what a manifold is.
We will first introduce topological manifolds, which are objects that locally look like
Euclidean space in certain continuous sense:

Definition 2.3 (Topological Manifolds). A Hausdorff, second countable topological

space M is said to be an n-dimensional topological manifold, or in short a topological
n-manifold, if for any point p ∈ M , there exists a homeomorphism F : U → O between
a non-empty open subset U ⊂ Rn and an open subset O ⊂ M containing p. This
homeomorphism F is called a local parametrization (or local coordinate chart) around p.

Example 2.4. Any regular surface is a topological manifold since its local parametriza-
tions are all homeomorphisms. Therefore, spheres, cylinders, torus, etc. are all topologi-
cal manifolds.
However, a double cone (see Figure 2.1) is not a topological manifold since the
vertex is a “bad” point. Any open set containing the vertex cannot be homeomorphic to
any open set in Euclidean space.
2.1. Smooth Manifolds 27

Figure 2.1. Double cone is not locally Euclidean near its vertex.

Remark 2.5. Note that around every p there may be more than one local parametriza-
tions. If Fα : Uα → Oα and Fβ : Uβ → Oβ are two local parametrizations around p, then
the composition:

(Fβ−1 ◦ Fα ) : Fα−1 (Oα ∩ Oβ ) → Fβ−1 (Oα ∩ Oβ )

(Fα−1 ◦ Fβ ) : Fβ−1 (Oα ∩ Oβ ) → Fα−1 (Oα ∩ Oβ )

are often called the transition maps between these local parametrizations. We need
to restrict their domains to smaller sets so as to guarantee the transition maps are
well-defined (c.f. Section 1.3). 

On a topological manifold, there is a coordinate system around every point. However,

many concepts in Calculus involve taking derivatives. In order to carry out differentiations
on manifolds, it is not sufficient to be merely locally homeomorphic to Euclidean spaces.
We need the local parametrization F to be differentiable in a certain sense.
For regular surfaces in R3 , the local parametrization F : U → R3 are maps between
Euclidean spaces, so it makes sense to take derivatives of F . However, an abstract
manifold may not be sitting in R3 or RN , and therefore it is difficult to make of sense
of differentiability of F : U → O. To get around this issue, we will not talk about the
differentiability of a local parametrization F , but instead talk about the differentiability
of transition maps.
In Proposition 1.11 of Chapter 1 we showed that any two overlapping local parametriza-
tions Fα and Fβ of a regular surface M have smooth transition maps Fβ−1 ◦ Fα and
Fα−1 ◦ Fβ . Now consider an abstract topological manifold. Although the local parametriza-
tions Fα and Fβ may not have a codomain sitting in Euclidean spaces, the transition
maps Fβ−1 ◦ Fα and Fα−1 ◦ Fβ are indeed maps between open subsets of Euclidean spaces!
While we cannot differentiate local parametrizations F : U → O ⊂ M for abstract
manifolds, we can do so for the transition maps Fβ−1 ◦ Fα and Fα−1 ◦ Fβ . This motivates
the definition of a smooth manifold:
28 2. Abstract Manifolds

Figure 2.2. transition maps of a manifold

Definition 2.6 (Smooth Manifolds). A n-dimensional topological manifold M is said

to be an n-dimensional smooth manifold, or in short a smooth n-manifold, if there is a
collection A of local parametrizations Fα : Uα → Oα such that
[
(1) Oα = M , i.e. these local parametrizations cover all of M ; and
α∈A
(2) all transition maps Fα−1 ◦ Fβ are smooth (i.e. C ∞ ) on their domains.

Remark 2.7. Two local parametrizations Fα and Fβ with smooth transition maps Fα−1 ◦Fβ
and Fβ−1 ◦ Fα are said to be compatible. 

Remark 2.8. We often use the superscript n, i.e. M n , to mean that the manifold M is
n-dimensional. 

Remark 2.9. A 2-dimensional manifold is sometimes called a surface. In this course, we

will use the term regular surfaces for those surfaces in R3 discussed in Chapter 1, while
we will use the term smooth surfaces to describe 2-dimensional smooth manifolds in the
sense of Definition 2.6. 

Example 2.10. Any topological manifold which can be covered by one global parametriza-
tion (i.e. image of F is all of M ) is a smooth manifold. Examples of which include Rn
which can be covered by one parametrization Id : Rn → Rn . The graph of Γf of any con-
tinuous function f : Rn → R is also a smooth manifold covered by one parametrization
F (x) = (x, f (x)) : Rn → Γf . Any regular curve γ(t) is a smooth manifold of dimension
1. 
2.1. Smooth Manifolds 29

Example 2.11. All regular surfaces in R3 are smooth manifolds by Proposition 1.11
(which we showed their transition maps are smooth). Therefore, spheres, cylinders, tori,
etc. are all smooth manifolds. 
Example 2.12 (Extended complex plane). Define M = C ∪ {∞}. One can show (omitted
here) that it is a Hausdroff, second countable topological space. Furthermore, one can
cover M by two local parametrizations:
F1 : R2 → C ⊂ M F2 : R2 → (C\{0}) ∪ {∞} ⊂ M
1
(x, y) 7→ x + yi (x, y) 7→
x + yi
The overlap part on M is given by C\{0}, corresponding to R2 \{(0, 0)} in R2 under
the parametrizations F1 and F2 . One can compute that the transition maps are given by:
 
−1 x y
F2 ◦ F1 (x, y) = ,− 2
x2 + y 2 x + y2
 
−1 x y
F1 ◦ F2 (x, y) = ,− 2
x2 + y 2 x + y2
Both are smooth maps on R2 \{(0, 0)}. Therefore, C ∪ {∞} is a smooth manifold. 

Exercise 2.1. Show that the n-dimensional sphere

Sn := {(x1 , . . . , xn+1 ) ∈ Rn+1 : x21 + . . . + x2n+1 = 1}
is a smooth n-manifold. [Hint: Generalize the stereographic projection to higher
dimensions]

Exercise 2.2. Discuss: According to Example 2.10, the graph of any continuous
function f : Rn → R is a smooth manifold as there is no transition map. However,
wouldn’t it imply the single cone:
n p o
(x, y, z) ∈ R3 : z = x2 + y 2
is a smooth manifold? It appears to have a “corner” point at the vertex, isn’t it?

2.1.3. Product and Quotient Manifolds. Given two smooth manifolds M m and
N , one can form an (m + n)-dimensional manifold M m × N n , which is defined by:
n

M m × N n := {(x, y) : x ∈ M m and y ∈ N n }.
Given a local parametrization F : UM → OM for M m , and a local parametrizaiton
G : UN → ON for N n , one can define a local parametrization:
F × G : UM × UN → OM × ON ⊂ M m × N n
(u, v) 7→ (F (u), G(v))
If {Fα } is a collection of local parametrizations of M m with smooth transition maps, and
{Gβ } is that of N n with smooth transition maps, then one can form a collection of local
parametrizations Fα × Gβ of the product M m × N n . It can be shown that these local
parametrizations of M m × N n also have smooth transition maps between open subsets
of Rm+n (see Exercise 2.3).

Exercise 2.3. Show that if Fα and Fαe are local parametrizations of M m with smooth
transition maps, and similarly for Gβ and Gβe for N n , then Fα × Gβ and Fαe × Gβe
have smooth transition maps.
30 2. Abstract Manifolds

The result from Exercise 2.3 showed that the product M m × N n of two smooth
manifolds M m and N n is a smooth manifold with dimension m + n. Inductively, the
product M1m1 × . . . Mkmk of k smooth manifolds M1m1 , . . . , Mkmk is a smooth manifold
with dimension m1 + . . . + mk .
Example 2.13. The cylinder x2 + y 2 = 1 in R3 can be regarded as R × S1 . The torus
can be regarded as S1 × S1 . They are both smooth manifolds. By taking products of
known smooth manifolds, one can generate a great deal of new smooth manifolds. The
n-dimensional cylinder can be easily seen to be a smooth manifold by regarding it as
R × Sn−1 . The n-dimensional torus S1 × . . . × S1 is also a smooth manifold. 
| {z }
n times

Another common way to produce a new manifold from an old one is to take quotients.
Take R as an example. Let us define an equivalence relation ∼ by declaring that x ∼ y if
and only if x − y is an integer. For instance, we have 3 ∼ 5 while 4 6∼ 92 . Then, we can
talk about equivalence classes [x] which is the following set:
[x] := {y ∈ R : y ∼ x}.
For instance, we have 5 ∈ [2] as 5 ∼ 2. Likewise −3 ∈ [2] as −3 ∼ 2 too. The set [2] is the
set of all integers. Similarly, one can also argue [−1] = [0] = [1] = [2] = [3] = . . . are all
equal to the set of all integers.
On the contrary, 1 6∈ [0.2] as 1 6∼ 0.2. Yet −1.8, −0.8, 0.2, . . . are all in the set [0.2].
The set [0.2] is simply the set of all numbers in the form of 0.2 + N where N is any integer.
One can also see that [−1.8] = [−0.8] = [0.2] = [1.2] = . . ..
Under such notations, we see that [1] = [2] while [1] 6= [0.2]. The notion of equiva-
lence classes provides us with a way to “decree” what elements in the “mother” set (R in
this case) are regarded as equal. This is how topologists and geometers interpret gluing.
In this example, we can think of 1, 2, 3, etc. are glued together, and also −1.8, −0.8, 0.2,
etc. are glued together. Formally, we denote
R/ ∼ := {[x] : x ∈ R}
which is the set of all equivalence classes under the relation ∼. This new set R/ ∼ is
called a quotient set of R by the equivalence relation ∼. By sketching the set, we can see
R/ ∼ is topologically a circle S1 (see Figure 2.3):

Figure 2.3. Quotient set R/ ∼

2.1. Smooth Manifolds 31

Exercise 2.4. Describe the set R2 / ∼ where we declare (x1 , y1 ) ∼ (x2 , y2 ) if and
only if x1 − x2 ∈ Z and y1 − y2 ∈ Z.

Example 2.14 (Real Projective Space). The real projective space RPn is the quotient
set of Rn+1 \{0} under the equivalence relation: (x0 , x1 , . . . , xn ) ∼ (y0 , y1 , . . . , yn ) if and
only if there exists λ ∈ R\{0} such that (x0 , x1 , . . . , xn ) = (λy0 , λy1 , . . . , λyn ). Each
equivalence class is commonly denoted by:
[x0 : x1 : · · · : xn ]
For instance, we have [0 : 1 : −1] = [0 : −π : π]. Under this notation, we can write:
RPn := [x0 : x1 : · · · : xn ] : (x0 , x1 , . . . , xn ) ∈ Rn+1 \{0}


It is important to note that [0 : 0 : · · · : 0] 6∈ RPn .

We are going to show that RPn is an n-dimensional smooth manifold. For each
i = 0, 1, . . . , n, we denote:
Oi := {[x0 : x1 : · · · : xn ] ∈ RPn : xi 6= 0} ⊂ RPn .
Define Fi : Rn → Oi by:
Fi (x1 , . . . , xn ) = [x1 : · · · : |{z}
1 : . . . xn ].
i
For instance, F0 (x1 , . . . , xn ) = [1 : x1 : · · · : xn ], F1 (x1 , . . . , xn ) = [x1 : 1 : x2 : · · · : xn ]
and Fn (x1 , . . . , xn ) = [x1 : · · · : xn : 1].
The overlap between images of F0 and F1 , for instance, is given by:
O0 ∩ O1 = {[x0 : x1 : x2 : · · · : xn ] : x0 , x1 6= 0}
F0−1 (O0 ∩ O1 ) = {(x1 , x2 , . . . , xn ) ∈ Rn : x1 6= 0}

One can compute that the transition map F1−1 ◦ F0 is given by:
 
−1 1 x2 xn
F1 ◦ F0 (x1 , . . . , xn ) = , ,...,
x1 x1 x1
which is smooth on the domain F0−1 (O0 ∩ O1 ). The smoothness of transition maps
between any other pairs can be verified in a similar way. 

Exercise 2.5. Express the transition map F3−1 ◦F1 of RP5 and verify that it is smooth
on its domain.
Example 2.15 (Complex Projective Space). The complex projective space CPn is an
important manifold in Complex Geometry (one of my research interests) and Algebraic
Geometry. It is defined similarly as RPn , with all R’s replaced by C’s. Precisely, we
declare for any two elements in (z0 , . . . , zn ), (w0 , . . . , wn ) ∈ Cn+1 \{(0, . . . , 0)}, we have
(z0 , . . . , zn ) ∼ (w0 , . . . , wn ) if and only if there exists λ ∈ C\{0} such that zi = λwi for
any i = 0, . . . n. Under this equivalence relation, the equivalence classes denoted by
[z0 : z1 : · · · : zn ] constitute the complex projective space:
CPn := {[z0 : z1 : · · · : zn ] : zi not all zero } .
It can be shown to be a smooth 2n-manifold in exactly the same way as in RPn . 

Exercise 2.6. Show that CPn is an 2n-dimensional smooth manifold by constructing

local parametrizations in a similar way as for RPn . For the transition maps, express
one or two of them explicitly and verify that they are smooth. What’s more can you
say about the transition maps apart from being smooth?
32 2. Abstract Manifolds

Exercise 2.7. Consider the equivalence relation of Rn defined as follows:

x ∼ y if and only if x − y ∈ Zn
Show that Rn / ∼ is a smooth n-manifold.

Example 2.16. The Klein Bottle K (see Figure 1.2a) cannot be put inside R3 without
self-intersection, but it can be done in R4 . It is covered by two local parametrizations
given below:

F1 : U1 → R4 where U1 := {(u1 , v1 ) : 0 < u1 < 2π and 0 < v1 < 2π}

 
(cos v1 + 2) cos u1
 (cos v1 + 2) sin u1 
 sin v1 cos u1  .
F1 (u1 , v1 ) =  
2
sin v1 sin u21
F2 : U2 → R4 where U2 := {(u2 , v2 ) : 0 < u2 < 2π and 0 < v2 < 2π}
 
−(cos v2 + 2) cos u2
 (cos v2 + 2) sin u2 
F2 (u2 , v2 ) = 
 sin v2 cos u2 + π  .


2 4

sin v2 sin u22 + π4
Geometrically speaking, the Klein bottle is generated by rotating the unit circle by
two independent rotations, one parallel to the xy-plane, another parallel to the zw-plane.
For geometric explanations for these parametrizations, see [dC94, P.36].
We leave it to readers to check that F1 and F2 are both injective and compatible with
each other. It will show that K is a 2-manifold. 

Exercise 2.8. Consider the Klein bottle K given in Example 2.16.

(a) Show that both F1 and F2 are injective.
(b) Let W = F1 (U1 ) ∩ F2 (U2 ). Find F1−1 (W) and F2−1 (W).
(c) Compute the transition maps F2−1 ◦ F1 and F1−1 ◦ F2 defined on the overlaps.

2.1.4. Differential Structures. A smooth manifold M n is equipped with a collection

smooth local parametrizations Fα : Uα ⊂ Rn → Oα ⊂ M n such that the images of these
Fα ’s cover the entire manifold, i.e.
[ [
M= Oα = Fα (Uα ).
all α’s all α’s

These local parametrizations need to be compatible with each other in a sense that any
overlapping parametrizations Fα and Fβ must have smooth transition maps Fα−1 ◦ Fβ
and Fβ−1 ◦ Fα . Such a collection of local parametrizations A = {Fα , Uα , Oα }α is called a
smooth atlas of M .
Given a smooth atlas A of M , we can enlarge the atlas by including more local
parametrizations Fnew : Unew → Onew that are compatible to all local parametrizations in
A. The differential structure generated by an atlas A is a gigantic atlas that contains all
local parametrizations which are compatible with every local parametrizations in A (for
more formal definition, please read [Lee09, Section 1.3]).
Let’s take the plane R2 as an example. It can be parametrized by at least three
different ways:
• the identity map F1 := id : R2 → R2 .
2.1. Smooth Manifolds 33

• the map F2 : R2 → (0, ∞) × (0, ∞) ⊂ R2 , defined as:

F2 (u, v) := (eu , ev ).
• and pathologically, by F3 : R2 → R2 defined as:
F3 (u, v) = (u, v + |u|).
F1−1
It is clear that ◦ F2 (u, v) = (eu , ev ) and F2−1 ◦ F1 (u, v) = (log u, log v) are smooth on
the domains at which they are defined. Therefore, we say that F1 and F2 are compatible,
and the differential structure generated by F1 will contain F2 .
On the other hand, F1−1 ◦ F3 (u, v) = (u, v + |u|) is not smooth, and so F1 and F3
are not compatible. Likewise, F2−1 ◦ F3 (u, v) = (log u, log(v + |u|)) is not smooth either.
Therefore, F3 does not belong to the differential structure generated by F1 and F2 .
As we can see from above, a manifold M can have many distinct differential struc-
tures. In this course, when we talk about manifolds, we usually only consider one
differential structure of the manifold, and very often we will only deal with the most
“natural” differential structure such as the one generated by F1 or F2 above for R3 , but
not like the pathological one such as F3 . Therefore, we usually will not specify the
differential structure when we talk about a manifold, unless it is necessary in some rare
occasions.

Exercise 2.9. Show that any smooth manifold has uncountably many distinct
differential structures. [Hint: Let B(1) := {x ∈ Rn : |x| < 1}, consider maps
s
Ψs : B(1) → B(1) defined by Ψs (x) = |x| x where s > 0.]
34 2. Abstract Manifolds

2.2. Functions and Maps on Manifolds

2.2.1. Definitions and Examples. Let’s first review how smooth functions f : M →
R and smooth maps Φ : M → N are defined for regular surfaces (Definitions 1.13 and
1.17). Given two and G : UN → ON ⊂ N , the compositions f ◦ F and G−1 ◦ Φ ◦ F are
functions or maps between Euclidean spaces. We say the function f is smooth if f ◦ F
is smooth for any local parametrizations F : UM → OM ⊂ M . We say Φ is smooth if
G−1 ◦ Φ ◦ F is smooth.
The definitions of differentiable functions and maps for regular surfaces carry over
to abstract manifolds in a natural way:

Definition 2.17 (Functions and Maps of Class C k ). Let M m and N n be two smooth
manifolds of dimensions m and n respectively. Then:
A scalar-valued function f : M → R is said to be C k at p ∈ M if for any smooth
local parametrization F : U → M with p ∈ F (U), the composition f ◦ F is C k at the
point F −1 (p) ∈ U as a function from subset from Rm to R. Furthermore, if f : M → R
is C k at every p ∈ M , then we say f is C k on M .
A map Φ : M → N is said to be C k at p ∈ M if for any smooth local parametrization
F : UM → OM ⊂ M with p ∈ F (UM ), and G : UN → ON ⊂ N with Φ(p) ∈ G(UN ),
the composition G−1 ◦ Φ ◦ F is C k at F −1 (p) as a map between subsets of Rm and Rn .
Furthermore, if Φ : M → N is C k at every p ∈ M , then Φ is said to be C k on M .
When k is ∞, we can also say that the function or map is smooth.

Figure 2.4. maps between two manifolds

Remark 2.18. By the definition of a smooth manifold (see condition (2) in Definition
2.6), transition maps are always smooth. Therefore, although we require f ◦ F and
G−1 ◦ Φ ◦ F to be smooth for any local parametrizations around p, it suffices to show
that they are smooth for at least one F covering p and at least one G covering Φ(p).

Exercise 2.10. Suppose Φ : M → N and Ψ : N → P are C k maps between smooth

manifolds M , N and P . Show that the composition Ψ ◦ Φ is also C k .

Example 2.19. Consider the 3-dimensional sphere

S3 = {(x1 , x2 , x3 , x4 ) ∈ R4 : x21 + x22 + x23 + x24 = 1 ∈ R}
2.2. Functions and Maps on Manifolds 35

and the complex projective plane

CP1 = {[z : w] : z 6= 0 or w 6= 0}.
Define a map Φ : S3 → CP1 by:
Φ(x1 , x2 , x3 , x4 ) = [x1 + ix2 , x3 + ix4 ].
Locally parametrize S3 by stereographic projection:
F : R3 → S3
−1 + k u2k
 P 
2u1 2u2 2u3
F (u1 , u2 , u3 ) = P 2, P 2, P 2, P 2
1 + k uk 1 + k uk 1 + k uk 1 + k uk
The image of F is S3 \{(0, 0, 0, 1)}. As usual, we locally parametrize CP1 by:
G : R2 → CP1
G(v1 , v2 ) = [1 : v1 + iv2 ]

The domain of G−1 ◦ Φ ◦ F is (Φ ◦ F )−1 Φ ◦ F (R3 ) ∩ G(R2 ) , and the map is explicitly

given by:
G−1 ◦ Φ ◦ F (u1 , u2 , u3 )
−1 + k u2k
 P 
2u1 2u2 2u3
= G−1 ◦ Φ P 2, P 2, P 2, P 2
1 + k uk 1 + k uk 1 + k uk 1 + k uk
−1 + k u2k
 P 
−1 2u1 2u2 2u3
=G +i : +i
1 + k u2k 1 + k u2k 1 + k u2k 1 + k u2k
P P P P
"
#
2u3 + i −1 + k u2k
P
= G−1 1 :
2u1 + 2iu2
2u1 u3 + u2 (−1 + k u2k ) −2u2 u3 + u1 (−1 + k u2k )
 P P 
= G−1 1 : + i
2(u21 + u22 ) 2(u21 + u22 )
2u1 u3 + u2 (−1 + k uk ) −2u2 u3 + u1 (−1 + k u2k )
 P 2 P 
= , .
2(u21 + u22 ) 2(u21 + u22 )
For any (u1 , u2 , u3 ) in the domain of G−1 ◦Φ◦F , which is (Φ◦F )−1 Φ ◦ F (R3 ) ∩ G(R2 ) ,

we have in particular Φ ◦ F (u1 , u2 , u3 ) ∈ G(R2 ), and so

2u1 2u2
+i 6= 0.
1 + k u2k 1 + k u2k
P P

Therefore, (u1 , u2 ) 6= (0, 0) whenever (u1 , u2 , u3 ) is in the domain of G−1 ◦ Φ ◦ F . From

the above computations, G−1 ◦ Φ ◦ F is smooth.
One can also check similarly that G e −1 ◦ Φ ◦ Fe is smooth for other combinations of
local parametrizations, concluding Φ is a smooth map.

Example 2.20. Let M × N be the product of two smooth manifolds M and N . Then,
the projection map πM and πN defined by:
πM : M × N → M
(p, q) 7→ p
πN :M ×N →N
(p, q) 7→ q
36 2. Abstract Manifolds

are both smooth manifolds. It can be shown by considering local parametrizations

F : UM → OM of M , and G : UN → ON of N . Then F × G : UM × UN → OM × ON is a
local parametrization of M × N . To show that πM is smooth, we compute:
F −1 ◦ πM ◦ (F × G) (u, v) = F −1 ◦ πM (F (u), G(v))
= F −1 (F (u))
=u
The map (u, v) 7→ u is clearly a smooth map between Euclidean spaces. Therefore, πM
is a smooth map between M × N and M . Similarly, πN is also a smooth map between
M × N and N . 

Exercise 2.11. Suppose Φ : Rn+1 \{0} → Rm+1 \{0} is a smooth map which satisfies
Φ(cx0 , cx1 , . . . , cxn ) = cd Φ(x0 , x1 , . . . , xn )
for any c ∈ R\{0} and (x0 , x1 , . . . , xn ) ∈ Rn+1 \{0}. Show that the induced map
e : RPn → RPm defined by:
Φ
Φe ([x0 : x1 : · · · : xn ]) = Φ(x0 , x1 , . . . , xn )
is well-defined and smooth. [Hint: To check Φ e is well-defined means to verify that
two equivalent inputs [x0 : x1 : · · · : xn ] = [y0 : y1 : · · · : yn ] will give the same
outputs Φ(x0 , x1 , . . . , xn ) and Φ(y0 , y1 , . . . , yn ).]

2 2
Exercise 2.12. Let M = {(w, z) ∈ C2 : |w| + |z| = 1}.
(a) Show that M is a 3-dimensional manifold.
(b) Define  
2 2
Φ(w, z) := z w̄ + wz̄, i(wz̄ − z w̄), |z| − |w|
for any (w, z) ∈ M . Show that Φ(w, z) ∈ R3 and it lies on the unit sphere S2 ,
and then verify that Φ : M → S2 is a smooth map.

2.2.2. Diffeomorphisms. Two smooth manifolds M and N are said to be diffeomor-

phic if they are in one-to-one correspondence with each other in smooth sense. Here is
the rigorous definition:

Definition 2.21 (Diffeomorphisms). A smooth map Φ : M → N between two smooth

manifolds M and N is said to be a diffeomorphism if Φ is a one-to-one and onto (i.e.
bijective), and that the inverse map Φ−1 : N → M is also smooth.
If such a map exists between M and N , the two manifolds M and N are said to be
diffeomorphic.
2.2. Functions and Maps on Manifolds 37

Example 2.22. Given a smooth function f : U → R from an open subset U ⊂ Rn . The

graph Γf defined as:
Γf := {(x, f (x)) ∈ Rn+1 : x ∈ U}
is a smooth manifold by Example 2.10. We claim that the projection map:
π : Γf → U
(x, f (x)) 7→ x
is a diffeomorphism. Both Γf and U are covered by one global parametrization. The
parametrization of U is simply the identity map idU on U. The parametrization of Γf is
given by:
F : U → Γf
x 7→ (x, f (x))

To show that π is smooth, we consider id−1

U ◦ π ◦ F , which is given by:

id−1 −1
U ◦ π ◦ F (x) = idU ◦ π(x, f (x))

= id−1
U (x)
= x.
Therefore, the composite id−1
◦ π ◦ F is simply the identity map on U, which is clearly
U
smooth.
π is one-to-one and onto with inverse map π −1 given by:
π −1 : U → Γf
x 7→ (x, f (x))
To show π −1 is smooth, we consider the composite F −1 ◦ π −1 ◦ idU :
F −1 ◦ π −1 ◦ idU (x) = F −1 ◦ π −1 (x)
= F −1 (x, f (x))
= x.
Therefore, the composite F −1 ◦ π −1 ◦ idU is also the identity map on U, which is again
smooth. 
Example 2.23. Let M be the cylinder x2 + y 2 = 1 in R3 . We are going to show that M
is diffeomorphic to R2 \{(0, 0)} via the diffeomorphism:
Φ : M → R2 \{(0, 0)}
(x, y, z) 7→ ez (x, y)
We leave it for readers to verify that Φ is one-to-one and onto, and hence Φ−1 exists. To
show it is a diffeomorphism, we first parametrize M by two local coordinate charts:
F1 : (0, 2π) × R → M F2 : (−π, π) × R → M
F1 (θ, z) = (cos θ, sin θ, z) F2 (θ,
e ze) = (cos θ,
e sin θ,
e ze)

The target space R2 \{(0, 0)} is an open set of R2 , and hence can be globally parametrized
by id : R2 \{(0, 0)} → R2 \{(0, 0)}.
We need to show Φ ◦ Fi and Fi−1 ◦ Φ−1 are smooth for any i = 1, 2. As an example,
we verify one of them only:
Φ ◦ F1 (θ, z) = Φ(cos θ, sin θ, z)
= (ez cos θ, ez sin θ).
38 2. Abstract Manifolds

To show F1−1 ◦ Φ−1 = (Φ ◦ F1 )−1 is smooth, we use Inverse Function Theorem. The
Jacobian of Φ ◦ F1 is given by:
 z
−e sin θ ez cos θ

D(Φ ◦ F1 ) = det z = −e2z 6= 0.
e cos θ ez sin θ
Therefore, Φ ◦ F1 has a C ∞ local inverse around every point in the domain. Since Φ ◦ F1
is one-to-one and onto, such a local inverse is a global inverse.
Similarly, one can show Φ ◦ F2 and F2−1 ◦ Φ−1 are smooth. All these show Φ and Φ−1
are smooth maps between M and R2 \{(0, 0)}, and hence are diffeomorphisms. 

Exercise 2.13. Show that the open square (−1, 1) × (−1, 1) ⊂ R2 is diffeomorphic
to R2 . [Hint: consider the trig functions tan or tan−1 .]

Exercise 2.14. Consider the map Φ : B1 (0) → Rn defined by:

|x|
Φ(x) = q
2
1 − |x|
where B1 (0) is the open unit ball {x ∈ Rn : |x| < 1}. Show that Φ is a diffeomor-
phism.

Exercise 2.15. Let M = R2 / ∼ where ∼ is the equivalence relation:

x∼y if and only if x − y ∈ Z2 .
From Exercise 2.7, we have already showed that M is a smooth manifold. Show
that M is diffeomorphic to a S1 × S1 .
2.3. Tangent Spaces and Tangent Maps 39

2.3. Tangent Spaces and Tangent Maps

3
 pon a regular surface M ⊂ R , the tangent plane Tp M at p is spanned by the
At a point
∂F ∂F
basis where F is a local parametrization around p. The basis are vectors
∂ui i=1,2 ∂ui
in R3 since F has an image in R3 . However, this definition of tangent plane can hardly
be generalized to abstract manifolds, as an abstract manifold M may not sit inside any
Euclidean space. Instead, we define the tangent space at a pointp ona smooth manifold
n
∂
as the vector space spanned by partial differential operators . Heuristically,
∂ui i=1
we generalize the concept of tangent planes of regular surfaces in R3 by “removing” the
∂F ∂
label F from the geometric vector , so that it becomes an abstract vector . For this
∂ui ∂ui
generalization, we first need to define partial derivatives on abstract manifolds.

2.3.1. Partial Derivatives and Tangent Vectors. Let M n be a smooth manifold

and F : U ⊂ Rn → O ⊂ M n be a smooth local parametrization. Then similar to
regular surfaces, for any p ∈ O, it makes sense to define partial derivative for a function
f : M → R at p by pre-composing f with F , i.e. f ◦ F , which is a map from U ⊂ Rn to
R. Let (u1 , . . . , un ) be the coordinates of U ⊂ Rn , then with a little abuse of notations,
we denote:
∂f ∂(f ◦ F )
(p) := (u)
∂uj ∂uj
where u is the point in U corresponding to p, i.e. F (u) = p.
∂f
Remark 2.24. Note that ∂u j
(p) is defined locally on O, and depends on the choice of
local parametrization F near p. 

∂
The partial derivative (p) can be thought as an operator:
∂uj
∂
(p) : C 1 (M, R) → R
∂uj
∂f
f 7→ (p).
∂uj
Here C 1 (M, R) denotes the set of all C 1 functions from M to R.
∂F ∂F
On regular surfaces (p) is a tangent vector at p. On an abstract manifold, (p)
∂uj ∂uj
cannot be defined since F may not be in an Euclidean space. Instead, the partial
∂ ∂F
differential operator (p) plays the role of (p), and we will call the operator
∂uj ∂uj
∂
(p) a tangent vector for an abstract manifold. It sounds strange to call a derivative
∂uj
operator a tangent vector. For beginners, you may try to get used to it by fantasizing the
∂
letter F whenever you see . “F” stands for “fantasize” (or in Hong Kong’s slang: try
∂uj
∂
to “FF” there were an F in )!
∂uj
Example 2.25. Let F (x, y, z) = (x, y, z) be the identity parametrization of R3 , and
G(ρ, θ, ϕ) = (ρ sin ϕ cos θ, ρ sin ϕ sin θ, ρ cos ϕ) be local parametrization of R3 by spherical
coordinates.
40 2. Abstract Manifolds

∂ ∂ ∂F ∂F
Figure 2.5. ∂u
and ∂v
are used in place of ∂u
and ∂v
on abstract manifolds.

∂ ∂ ∂
Then at any point p ∈ R3 , the vectors (p), (p), (p) are regarded as the
∂x ∂y ∂z
∂F ∂F ∂F
abstract form of the geometric vectors (p), (p), (p), which are respectively î, ĵ
∂x ∂y ∂z
and k̂ in standard notations.
∂ ∂ ∂
Also, the vectors (p), (p), (p) are regarded as the abstract form of the geo-
∂ρ ∂θ ∂ϕ
∂G ∂G ∂G
metric vectors (p), (p), (p), which are respectively the vectors at p tangent to
∂ρ ∂θ ∂ϕ
the ρ-, θ- and ϕ-directions on the sphere. 

Example 2.26. Take another example:

RP2 = {[x0 : x1 : x2 ] : at least one xi 6= 0}.

According to Example 2.14, one of its local parametrizations is given by:

F : R2 → RP2
(x1 , x2 ) 7→ [1 : x1 : x2 ]

∂F ∂F
Such a manifold is not assumed to be in RN , so we can’t define , as geometric
∂x1 ∂x2
∂ ∂
vectors in RN . However, as a substitute, we will regard the operators , as abstract
∂x1 ∂x2
tangent vectors along the directions of x1 and x2 respectively. 

2.3.2. Tangent Spaces. Having generalized the concept of partial derivatives to

abstract manifolds, we now ready to state the definition of tangent vectors for abstract
manifolds.

Definition 2.27 (Tangent Spaces). Let M be a smooth n-manifold, p ∈ M and F : U ⊂

Rn → O ⊂ M be a smooth local parametrization around p. The tangent space at p of
M , denoted by Tp M , is defined as:
 
∂ ∂
Tp M = span (p), . . . , (p) ,
∂u1 ∂un
∂
where ’s are partial differential operators with respect to the local parametrization
∂ui
F (u1 , . . . , un ).
2.3. Tangent Spaces and Tangent Maps 41

It seems that the definition of Tp M depends on the choiceof local 

parametrization
n
∂
F . However, we can show that it does not. We first show that (p) are linearly
∂ui i=1
independent, then we have dim Tp M = n = dim M .
Given a local parametrization F : U → O ⊂ M with local coordinates denoted
by (u1 , . . . , un ), then each coordinate ui can be regarded as a locally defined function
ui : O → R. Then we have:
(
∂uk 1 if i = j
(p) = δkj = .
∂uj 0 otherwise
 n
∂
Next we want to show are linearly independent. Suppose ai ’s are real
∂ui i=1
numbers such that
n
X ∂
ai = 0,
i=1
∂ui
n
X ∂f
meaning that ai = 0 for any differential function f (including the coordinate
i=1
∂ui
functions uk ’s). Therefore, we have:
n n
X ∂uk X
0= ai = ai δki = ak
i=1
∂ui i=1
 n
∂
for any k. This shows are linearly independent, show that dim Tp M = dim M .
∂ui i=1
Now we show Tp M does not depend on the choice of local coordinates. Suppose
F : U ⊂ Rn → O ⊂ M and Fe : Ue ⊂ Rn → O e be two local parametrizations. We use
(u1 , . . . , un ) to denote the Euclidean coordinates on U, and use (v1 , . . . , vn ) for U.e
∂f ∂f
The partial derivatives are are different. Via the transition map Fe−1 ◦ F ,
∂uj ∂vi
(v1 , . . . , vn ) can be regarded as functions of (u1 , . . . , un ), and therefore it makes sense of
∂vj
defining .
∂ui
Given a smooth function f : M → R, by the chain rule, one can write the partial
∂f ∂f
derivative in terms of as follows:
∂uj ∂vi
∂f ∂(f ◦ F )
(2.1) (p) :=
∂ui ∂ui F −1 (p)
∂
= (f ◦ Fe) ◦ (Fe−1 ◦ F )
∂ui F −1 (p)
n
X ∂(f ◦ Fe) ∂vj
=
j=1
∂vj ∂ui F −1 (p)
F−1 (p)
e
n
X ∂vj ∂f
= (p)
j=1
∂ui F −1 (p) ∂vj

In short, we can write:

n
∂ X ∂vj ∂
= .
∂ui j=1
∂ui ∂vj
42 2. Abstract Manifolds

∂ ∂ 0
In other words, can be expressed as a linear combination of s.
∂ui ∂vj
 n  n
∂ ∂
Therefore, span (p) ⊂ span (p) . Since both spans of vectors have
∂ui i=1 ∂vi i=1
equal dimension, their span must be equal. This shows Tp M is independent of choice of
local parametrizations. However, it is important to note that each individual basis vector
∂
(p) does depend on local parametrizations.
∂ui
Example 2.28. Consider again the real projective plane:
RP2 = {[x0 : x1 : x2 ] : at least one xi 6= 0}.
Consider the two local parametrizations:
F : R2 → RP2 G : R2 → RP2
F (x1 , x2 ) = [1 : x1 : x2 ] G(y0 , y2 ) = [y0 : 1 : y2 ]
Then, (y0 , y2 ) can be regarded as a function of (x1 , x2 ) via the transition map G−1 ◦ F ,
which is explicitly given by:
(y0 , y2 ) = G−1 ◦ F (x1 , x2 ) = G−1 ([1 : x1 : x2 ])
 
−1 −1 −1 1 x2
= G ([x1 : 1 : x1 x2 ]) = , .
x1 x1
∂ ∂ ∂ ∂
Using the chain rule, we can then express , in terms of , :
∂x1 ∂x2 ∂y0 ∂y2
∂ ∂y0 ∂ ∂y2 ∂
= +
∂x1 ∂x1 ∂y0 ∂x1 ∂y2
1 ∂ x2 ∂
=− 2 − 2
x1 ∂y0 x1 ∂y2
∂ ∂
= −y02 − y0 y2 .
∂y0 ∂y2
∂
We leave as an exercise. 
∂x2

∂ ∂ ∂
Exercise 2.16. Express as a linear combination , in Example 2.28.
∂x2 ∂y0 ∂y2
Leave the final answer in terms of y0 and y2 only.

Exercise 2.17. Consider the extended complex plane M := C ∪ {∞} (discussed in

Example 2.12) with local parametrizations:
F 1 : R2 → C ⊂ M F2 : R2 → (C\{0}) ∪ {∞} ⊂ M
1
(x1 , x2 ) 7→ x1 + x2 i (y1 , y2 ) 7→
y1 + y2 i
   
∂ ∂
Express the tangent space basis in terms of the basis .
∂xi ∂yj

Exercise 2.18. Given two smooth manifolds M m and N n , and a point (p, q) ∈
M × N , show that the tangent plane T(p,q) (M × N ) is isomorphic to Tp M ⊕ Tq N .
Recall that V ⊕ W is the direct sum of two vector spaces V and W , defined as:
V ⊕ W = {(v, w) : v ∈ V and w ∈ W }.
2.3. Tangent Spaces and Tangent Maps 43

2.3.3. Tangent Maps. Given a smooth map Φ between two regular surfaces in
R3 , we discussed in Section 1.5.2 on how to define its partial derivatives using local
parametrizations. To recap, suppose Φ : M → N and F (u1 , u2 ) : UM → OM ⊂ M and
G(v1 , v2 ) : UN → ON ⊂ N are local parametrizations of M and N respectively. Via Φ,
the local coordinates (v1 , v2 ) of N can be regarded as functions of (u1 , u2 ), i.e.
(v1 , v2 ) = G−1 ◦ Φ ◦ F (u1 , u2 ).
∂Φ
Then, according to (1.1), the partial derivative of the map Φ is given by:
∂ui
∂Φ ∂(Φ ◦ F ) ∂(G ◦ (G−1 ◦ Φ ◦ F ))
(Φ(p)) := =
∂ui ∂ui F −1 (p) ∂ui F −1 (p)
2
X ∂vj ∂G
= (Φ(p))
j=1
∂ui F −1 (p) ∂vj

which is a vector on the tangent plane TΦ(p) N .

Now, if we are given a smooth map Φ : M m → N n between two smooth abstract
manifolds M m and N n with local parametrizations F (u1 , . . . , um ) : UM ⊂ Rm → OM ⊂
M m around p ∈ M , and G(v1 , . . . , vn ) : U n
N ⊂ R → O
n
Nn ⊂ N around Φ(p) ∈ N , then
∂
the tangent space TΦ(p) N is spanned by (Φ(p)) . In view of (1.1), a natural
∂vj j=1
∂Φ
generalization of partial derivatives (p) to smooth maps between manifolds is:
∂ui

Definition 2.29 (Partial Derivatives of Maps between Manifolds). Let Φ : M m → N n

be a smooth map between two smooth manifolds M and N . Let F (u1 , . . . , um ) : UM →
OM ⊂ M be a smooth local parametrization around p and G(v1 , . . . , vn ) : UN → ON ⊂
N be a smooth local parametrization around Φ(p). Then, the partial derivative of Φ with
respect to ui at p is defined to be:
n
∂Φ X ∂vj ∂
(2.2) (p) := (Φ(p)).
∂ui j=1
∂ui −1
F (p) ∂v j

Here (v1 , . . . , vn ) are regarded as functions of (u1 , . . . , um ) in a sense that:

(v1 , . . . , vn ) = G−1 ◦ Φ ◦ F (u1 , . . . , um ).

∂Φ
Note that the partial derivative defined in (2.2) depends on the local parametriza-
∂ui
tion F . However, one can show that it does not depend on the choice of the local
parametrization G in the target space.
Suppose G(w
e 1 , . . . , wn ) is another local parametrization around Φ(p). Then by the
chain rule:
n n n
!
X ∂wj ∂ X ∂wj X ∂vk ∂
=
j=1
∂ui ∂wj j=1
∂ui ∂wj ∂vk
k=1
n
X ∂wj ∂vk ∂
=
∂ui ∂wj ∂vk
j,k=1
n
X ∂vk ∂
=
∂ui ∂vk
k=1
44 2. Abstract Manifolds

∂Φ
Therefore, the way to define in (2.2) is independent of choice of local parametrization
∂ui
G for the target manifold N .
Example 2.30. Consider the map Φ : RP1 × RP2 → RP5 defined by:
Φ([x0 : x1 ], [y0 : y1 : y2 ]) = [x0 y0 : x0 y1 : x0 y2 : x1 y0 : x1 y1 : x1 y2 ].
Under the standard local parametrizations F (u) = [1 : u] for RP1 , G(v1 , v2 ) = [1 : v1 : v2 ]
for RP2 , and H(w1 , . . . , w5 ) = [1 : w1 : · · · : w5 ] for RP5 , the local expression of Φ is
given by:
H −1 ◦ Φ ◦ (F × G)(u, v1 , v2 )
= H −1 ◦ Φ([1 : u], [1 : v1 : v2 ])
= H −1 ([1 : v1 : v2 : u : uv1 : uv2 ])
= (v1 , v2 , u, uv1 , uv2 ).
Via the map Φ, we can regard (w1 , w2 , w3 , w4 , w5 ) = (v1 , v2 , u, uv1 , uv2 ), and the partial
derivatives of Φ are given by:
∂Φ ∂w1 ∂ ∂w5 ∂ ∂ ∂ ∂
= + ... + = + v1 + v2
∂u ∂u ∂w1 ∂u ∂w5 ∂w3 ∂w4 ∂w5
∂Φ ∂w1 ∂ ∂w5 ∂ ∂ ∂
= + ... + = +u
∂v1 ∂v1 ∂w1 ∂v1 ∂w5 ∂w1 ∂w4
∂Φ ∂w1 ∂ ∂w5 ∂ ∂ ∂
= + ... + = +u
∂v2 ∂v2 ∂w1 ∂v2 ∂w5 ∂w2 ∂w5


Similar to tangent maps between regular surfaces in R3 , we define:

 
∂ ∂Φ
(Φ∗ )p (p) := (p)
∂ui ∂ui
and extend the map linearly to all vectors in Tp M . This is then a linear map between
Tp M and TΦ(p) N , and we call this map the tangent map.

Definition 2.31 (Tangent Maps). Under the same assumption stated in Definition 2.29,
the tangent map of Φ at p ∈ M denoted by (Φ∗ )p is defined as:
(Φ∗ )p : Tp M → TΦ(p) N
n
! n
X ∂ X ∂Φ
(Φ∗ )p ai (p) = ai (p)
i=1
∂u i i=1
∂ui

If the point p is clear from the context, (Φ∗ )p can be simply denoted by Φ∗ .

For brevity, we will from now on say “(u1 , . . . , um ) are local coordinates of M around
p” instead of saying in a clumsy way that “F : U → M is a local parametrization of M
around p and that (u1 , . . . , um ) are coordinates on U”.
Given a local coordinates (u1 , . . . , um ) around p, and local coordinates (v1 , . . . , vn )
around Φ(p), then from (2.2), the matrix representation of (Φ∗ )p with respect to bases
 m  n  j=1,...,n
∂ ∂ ∂vj
(p) and (Φ(p)) is given by where i stands for the
∂ui i=1 ∂vj j=1 ∂ui i=1,...,m
column, and j stands for the row. The matrix is nothing but the Jacobian matrix:
 
∂(v1 , . . . , vn )
= D(G−1 ◦ Φ ◦ F ) F −1 (p) .
 
(2.3) [(Φ∗ )p ] =
∂(u1 , . . . , um ) F −1 (p)
2.3. Tangent Spaces and Tangent Maps 45

Example 2.32. Consider again the map Φ : RP1 × RP2 → RP5 in Example 2.30. Under
the local parametrizations considered in that example, we then have (for instance):
 
∂ ∂Φ ∂ ∂ ∂
Φ∗ = = + v1 + v2 .
∂u ∂u ∂w3 ∂w4 ∂w5
Using the results computed in Example 2.30, the matrix representation of Φ∗ is given by:
 
0 1 0
 0 0 1
 
 1 0 0
[Φ∗ ] =  
v1 u 0 
v2 0 u
Hence, Φ∗ is injective. Remark: To be rigorous, we have only shown (Φ∗ )p is injective at
any p covered by the local coordinate charts we picked. The matrix [Φ∗ ] using other local
coordinate charts can be computed in a similar way (left as an exercise). 

Exercise 2.19. Consider the map Φ : RP1 × RP2 → RP5 defined as in Example 2.30.
This time, we use the local parametrizations
F (u) = [u : 1]
G(v0 , v2 ) = [v0 : 1 : v2 ]
H(w0 , w1 , w3 , w4 , w5 ) = [w0 : w1 : 1 : w3 : w4 : w5 ]
for RP1 , RP2 and RP5 respectively. Compute matrix representation of Φ∗ using these
local parametrizations.

Exercise 2.20. Note that in Definition 2.31 we defined Φ∗ using local coordinates.
Show that Φ∗ is independent of local coordinates. Precisely, show that if:
X ∂ X ∂
ai = bi
i
∂u i i
∂w i

where {ui } and {wi } are two local coordinates of M , then we have:
! !
X ∂Φ X ∂Φ X ∂ X ∂
ai = bi , which implies Φ∗ ai = Φ∗ bi .
i
∂ui i
∂wi i
∂ui i
∂wi

Exercise 2.21. The identity map idM of a smooth manifolds M takes any point
p ∈ M to itself, i.e. idM (p) = p. Show that its tangent map (idM )∗ at p is the identity
map on the tangent space Tp M .

Exercise 2.22. Consider two smooth manifolds M m and N n , and their product
M m × N n . Find the tangent maps (πM )∗ and (πN )∗ of projection maps:
πM : M × N → M
(x, y) 7→ x
πN : M × N → N
(x, y) 7→ y
46 2. Abstract Manifolds

2.4. Inverse Function Theorem

2.4.1. Chain Rule. Consider a smooth function Ψ(v1 , . . . , vk ) : Rk → Rm , another
smooth function Φ(u1 , . . . , un ) : Rn → Rk and the composition Ψ ◦ Φ. Under this
composition, (v1 , . . . , vk ) can be regarded as a function of (u1 , . . . , un ), and the output
(w1 , . . . , wm ) = Ψ(v1 , . . . , vk ) is ultimately a function of (u1 , . . . , un ). In Multivariable
Calculus, the chain rule is usually stated as:
∂wj X ∂wj ∂vl
=
∂ui ∂vl ∂ui
l

or equivalently in an elegant way using Jacobian matrices:

∂(w1 , . . . , wm ) ∂(w1 , . . . , wm ) ∂(v1 , . . . , vk )
= .
∂(u1 , . . . , un ) ∂(v1 , . . . , vk ) ∂(u1 , . . . , un )
Our goal here is to show that the chain rule can be generalized to maps between
smooth manifolds, and can be rewritten using tangent maps:

Theorem 2.33 (Chain Rule: smooth manifolds). Let Φ : M m → N n and Ψ : N n → P k

be two smooth maps between smooth manifolds M , N and P , then we have:
(Ψ ◦ Φ)∗ = Ψ∗ ◦ Φ∗

Proof. Suppose F (u1 , . . . , um ) is a smooth local parametrization of M , G(v1 , . . . , vn ) is

a smooth local parametrization of N and H(w1 , . . . , wk ) is a smooth parametrization
of P . Locally, (w1 , . . . , wk ) are then functions of (v1 , . . . , vn ) via Ψ; and (v1 , . . . , vn ) are
functions of (u1 , . . . , um ) via Φ, i.e.
(w1 , . . . , wk ) = H −1 ◦ Ψ ◦ G(v1 , . . . , vn )
(v1 , . . . , vn ) = G−1 ◦ Φ ◦ F (u1 , . . . , um )
Ultimately, we can regard (w1 , . . . , wk ) as functions of (u1 , . . . , um ) via the composition
Ψ ◦ Φ:
(w1 , . . . , wk ) = (H −1 ◦ Ψ ◦ G) ◦ (G−1 ◦ Φ ◦ F )(u1 , . . . , um )
= H −1 ◦ (Ψ ◦ Φ) ◦ F (u1 , . . . , um )

To find the tangent map (Ψ ◦ Φ)∗ , we need to figure out how it acts on the basis
∂
vectors , and recall that it is defined (see (2.2)) as follows:
∂ui
  k
∂ ∂(Ψ ◦ Φ) X ∂wj ∂
(Ψ ◦ Φ)∗ = = .
∂ui ∂ui j=1
∂ui ∂wj

Next, we use the (standard) chain rule for maps between Euclidean spaces:
k k Xn
X ∂wj ∂ X ∂wj ∂vl ∂
= .
j=1
∂ui ∂wj j=1
∂vl ∂ui ∂wj
l=1

Therefore, we get:
  k Xn
∂ X ∂wj ∂vl ∂
(Ψ ◦ Φ)∗ = .
∂ui j=1
∂vl ∂ui ∂wj
l=1
 
∂
Next, we verify that Ψ∗ ◦ Φ∗ will give the same output:
∂ui
2.4. Inverse Function Theorem 47

  n
∂ ∂Φ X ∂vl ∂
Φ∗ = =
∂ui ∂ui ∂ui ∂vl
l=1
  n
! n  
∂ X ∂vl ∂ X ∂vl ∂
Ψ∗ ◦ Φ∗ = Ψ∗ = Ψ∗
∂ui ∂ui ∂vl ∂ui ∂vl
l=1 l=1
 
n n k
X ∂vl ∂Ψ X ∂vl  X ∂wj ∂ 
= =
∂ui ∂vl ∂ui j=1 ∂vl ∂wj
l=1 l=1
k Xn
X ∂wj ∂vl ∂
= .
j=1
∂vl ∂ui ∂wj
l=1

Therefore, we have:
   
∂ ∂
(Ψ ◦ Φ)∗ = Ψ∗ ◦ Φ∗
∂ui ∂ui
for any i, and hence (Ψ ◦ Φ)∗ = Ψ∗ ◦ Φ∗ . 

Here is one immediate corollary of the chain rule:

Corollary 2.34. If Φ : M → N is a diffeomorphism between two smooth manifolds M

and N , then at each point p ∈ M the tangent map Φ∗ : Tp M → TΦ(p) N is invertible.

Proof. Given that Φ is a diffeomorphism, the inverse map Φ−1 : N → M exists. Since
Φ−1 ◦ Φ = idM , using the chain rule and Exercise 2.21, we get:

idT M = (idM )∗ = (Φ−1 ◦ Φ)∗ = (Φ−1 )∗ ◦ Φ∗ .

Similarly, one can also show Φ∗ ◦ Φ−1

∗ = idT N . Therefore, Φ∗ , and (Φ
−1
)∗ as well, are
invertible. 

Exercise 2.23. Given two diffeomorphic smooth manifolds M and N , what can you
say about dim M and dim N ?

Exercise 2.24. Let S2 = {(x, y, z) ∈ R3 : x2 + y 2 + z 2 = 1} be the unit sphere.

Consider the maps π : S2 → RP2 defined by
π(x, y, z) := [x : y : z]
2
and Φ : RP → R4 defined by:
Φ([x : y : z]) = (x2 − y 2 , xy, xz, yz).
Locally parametrize S2 stereographically:
u2 + v 2 − 1
 
2u 2v
F (u, v) = , ,
u2 + v 2 + 1 u2 + v 2 + 1 u2 + v 2 + 1
and RP2 by a standard parametrization:
G(w1 , w2 ) = [1 : w1 : w2 ].
Compute [Φ∗ ], [π∗ ] and [(Φ ◦ π)∗ ] directly, and verify that [(Φ ◦ π)∗ ] = [Φ∗ ][π∗ ].
48 2. Abstract Manifolds

2.4.2. Inverse Function Theorem. Given a diffeomorphism Φ : M → N , it is

necessary that Φ∗ is invertible. One natural question to ask is that if we are given Φ∗ is
invertible, can we conclude that Φ is a diffeomorphism?
Unfortunately, it is too good to be true. One easy counter-example is the map
Φ : R → S1 , defined as:
Φ(t) = (cos t, sin t).
As both R and S1 are one dimensional manifolds, to show that Φ∗ is invertible it suffices
to show that Φ∗ 6= 0, which can be verified by considering:
 
∂ ∂Φ
Φ∗ = = (− sin t, cos t) 6= 0.
∂t ∂t
However, it is clear that Φ is not even one-to-one, and hence Φ−1 does not exist.
Fortunately, the Inverse Function Theorem tells us that Φ is locally invertible near p
whenever (Φ∗ )p is invertible. In Multivariable Calculus/Analysis, the Inverse Function
Theorem asserts that if the Jacobian matrix of a smooth map Φ : Rn → Rn at p is
invertible, then there exists an open set U ⊂ Rn containing p, and an open set V ⊂ Rn
containing Φ(p) such that Φ|U : U → V is a diffeomorphism.
Now suppose Φ : M → N is a smooth map between two smooth manifolds M and N .
According to (2.2), the matrix representation of the tangent map Φ∗ is a Jacobian matrix.
Therefore, one can generalize the Inverse Function Theorem to smooth manifolds. To
start, we first define:

Definition 2.35 (Local Diffeomorphisms). Let Φ : M → N be a smooth map between

two smooth manifolds M and N . We say Φ is a local diffeomorphism near p if there
exists an open set OM ⊂ M containing p, and an open set ON ⊂ N containing Φ(p)
such that Φ|OM : OM → ON is a diffeomorphism.
If such a smooth map exists, we say M is locally diffeomorphic to N near p, or
equivalently, N is locally diffeomorphic to M near Φ(p). If M is locally diffeomorphic to
N near every point p ∈ M , then we say M is locally diffeomorphic to N .

Figure 2.6. A local diffeomorphism which is not injective.

Theorem 2.36 (Inverse Function Theorem). Let Φ : M → N be a smooth map between

two smooth manifolds M and N . If (Φ∗ )p : Tp M → TΦ(p) N is invertible, then M is locally
diffeomorphic to N near p.
2.4. Inverse Function Theorem 49

Proof. The proof to be presented uses the Inverse Function Theorem for Euclidean spaces
and then extends it to smooth manifolds. For the proof of the Euclidean case, readers
may consult the lecture notes of MATH 3033/3043.
Let F be a local parametrization of M near p, and G be a local parametrization
of N near Φ(p). Given that (Φ∗ )p is invertible, by (2.3) we know that the following
Jacobian matrix D(G−1 ◦ Φ ◦ F ) is invertible at F −1 (p). By Inverse Function Theorem
for Euclidean spaces, there exist an open set UM ⊂ Rdim M containing F −1 (p), and an
open set UN ⊂ Rdim N containing G−1 (Φ(p)) such that:
G−1 ◦ Φ ◦ F UM
: UM → UN
is a diffeomorphism, i.e. the inverse F −1 ◦ Φ−1 ◦ G exists when restricted to UN and is
smooth.
Denote OM = F (UM ) and ON = G(UN ). By the definition of smooth maps, this
shows Φ|OM and Φ−1 O are smooth. Hence Φ|OM is a local diffeomorphism near p.
N

Example 2.37. The helicoid Σ is defined to be the following surface in R3 :

Σ := {(r cos θ, r sin θ, θ) ∈ R3 : r > 0 and θ ∈ R}.

(a) A helicoid (b) Locally diffeomorphic to R2 \{0}

Figure 2.7. a helicoid is not globally diffeomorphic to R2 \{0}, but is locally diffeomor-
phic to R2 \{0}.

It can be parametrized by:

F : (0, ∞) × R → Σ
F (r, θ) = (r cos θ, r sin θ, θ)

Consider the map Φ : Σ → R2 \{0} defined as:

Φ(r cos θ, r sin θ, θ) = (r cos θ, r sin θ).
50 2. Abstract Manifolds

It is clear that Φ is not injective: for instance, Φ(cos 2π, sin 2π, 2π) = Φ(cos 0, sin 0, 0).
However, we can show that (Φ∗ )p is injective at each point p ∈ Σ.
The set R2 \{0} is open in R2 . The matrix [Φ∗ ] is the Jacobian matrix of Φ ◦ F :
Φ ◦ F (r, θ) = Φ(r cos θ, r sin θ, θ)
= (r cos θ, r sin θ)
 
cos θ −r sin θ
[Φ∗ ] = D(Φ ◦ F ) = .
sin θ r cos θ
As det[Φ∗ ] = r 6= 0, the linear map [Φ∗ ] is invertible. By Inverse Function Theorem, Φ is
a local diffeomorphism.

Exercise 2.25. Show that Sn and RPn are locally diffeomorphic via the map:
Φ(x0 , . . . , xn ) = [x0 : · · · : xn ].
2.5. Immersions and Submersions 51

2.5. Immersions and Submersions

2.5.1. Review of Linear Algebra: injectivity and surjectivity. Given a linear map
T : V → W between two finite dimensional vector spaces V and W , the following are
equivalent:
(a) T is injective;
(b) ker T = {0};
(c) The row reduced echelon form (RREF) of the matrix of T has no free column.
In each RREF of a matrix, we call the first non-zero entry (if exists) of each row to be a
pivot. A free column of an RREF is a column which does not have a pivot. For instance,
the following RREF:
 
1 3 0 0 0
R = 0 0 1 1 0
0 0 0 0 1
has three pivots, and two free columns (namely the second and fourth columns). Any
map with a matrix which can be row reduced to this R is not injective.
Surjectivity of a linear map T : V → W can also be stated in several equivalent ways:
(a) T is surjective;
(b) rank(T ) = dim W ;
(c) All rows in the RREF of the matrix of T are non-zero.
For instance, all rows of the matrix R above are non-zero. Hence any map with a matrix
which can be row reduced to R is surjective.

Exercise 2.26. Let T : V → W be a linear map between two finite dimensional

vector spaces V and W . Given that T is injective, what can you say about dim V and
dim W ? Explain. Now given that T is surjective, what can you say about dim V and
dim W ? Explain.

2.5.2. Immersions. Loosely speaking, an immersion from one smooth manifold to

another is a map that is “locally injective”. Here is the rigorous definition:

Definition 2.38 (Immersions). Let Φ : M → N be a smooth map between two smooth

manifolds M and N . We say Φ is an immersion at p ∈ M if the tangent map (Φ∗ )p :
Tp M → TΦ(p) N is injective. If Φ is an immersion at every point on M , then we simply
say Φ is an immersion.

Remark 2.39. As a linear map T : V → W between any two finite dimensional vector
spaces cannot be injective if dim V > dim W , an immersion Φ : M → N can only exist
when dim M ≤ dim N . 
Example 2.40. The map Φ : R → S1 defined by:
Φ(t) = (cos t, sin t)
( )
∂
is an immersion. The tangent space of R at any point t0 is simply span . The
∂t t=t0
tangent map (Φ∗ )t0 is given by:
 
∂ ∂Φ
(Φ∗ )t0 = = (− sin t0 , cos t0 ) 6= 0.
∂t ∂t t=t0
52 2. Abstract Manifolds

Therefore, the “matrix” of Φ∗ is a one-by-one matrix with a non-zero entry. Clearly, there
is no free column and so Φ∗ is injective at every t0 ∈ R. This shows Φ is an immersion.
This example tells us that an immersion Φ is not necessary injective. 
Example 2.41. Let M be a regular surface in R , then the inclusion map ι : M → R3 ,
2 3 2

defined as ι(p) = p ∈ R3 , is a smooth map, since for any local parametrization F (u1 , u2 )
of M 2 , we have ι ◦ F = F , which is smooth by definition (see p.4). We now show that ι
is an immersion:
 
∂F ∂(ι ◦ F ) ∂F
(ι∗ )p = = .
∂ui ∂ui F −1 (p) ∂ui F −1 (p)
Let F (u1 , u2 ) = (x1 (u1 , u2 ), x2 (u1 , u2 ), x3 (u1 , u2 )), then
3
∂F X ∂xj
= êi
∂ui j=1
∂ui

where {êi } is the standard basis of R3 . Therefore, the matrix of ι∗ is given by:
 
∂x1 ∂x1
 ∂u1
∂x2
∂u2
∂x2 
[ι∗ ] =  ∂u1 ∂u2  .
∂x3 ∂x3
∂u1 ∂u2

∂F ∂F ∂(x2 , x3 ) ∂(x3 , x1 ) ∂(x1 , x2 )

By the condition 0 6= × = ê1 + ê2 + ê3 , at each p ∈
∂u1 ∂u2 ∂(u1 , u2 ) ∂(u1 , u2 ) ∂(u1 , u2 )
M one least one of the following is invertible:
∂(x2 , x3 ) ∂(x3 , x1 ) ∂(x1 , x2 )
, , .
∂(u1 , u2 ) ∂(u1 , u2 ) ∂(u1 , u2 )
and hence has the 2 × 2 identity as its RREF. Using this fact, one can row reduce [ι∗ ] so
that it becomes:    
1 0 1 0
[ι∗ ] → . . . → 0 1 → 0 1
∗ ∗ 0 0
which has no free column. Therefore, [ι∗ ] is an injective linear map at every p ∈ M . This
shows ι is an immersion. 

Exercise 2.27. Define a map Φ : R2 → R4 by:

Φ(x, y) = (x3 , x2 y, xy 2 , y 3 ).
Show that Φ is an immersion at any (x, y) 6= (0, 0).

Exercise 2.28. Given two immersions Φ : M → N and Ψ : N → P between smooth

manifolds M , N and P , show that Ψ ◦ Φ : M → P is also an immersion.

Exercise 2.29. Consider two smooth maps Φ1 : M1 → N1 and Φ2 : M2 → N2

between smooth manifolds. Show that:
(a) If both Φ1 and Φ2 are immersions, then so does Φ1 × Φ2 : M1 × M2 → N1 × N2 .
(b) If Φ1 is not an immersion, then Φ1 × Φ2 cannot be an immersion.

A nice property of an immersion Φ : M m → N n is that for every Φ(p) ∈ N , one can

find a special local parametrization G of N such that G−1 ◦ Φ ◦ F is an inclusion map
from Rm to Rn , which is a map which takes (x1 , . . . , xm ) to (x1 , . . . , xm , 0, . . . , 0). Let’s
state the result in a precise way:
2.5. Immersions and Submersions 53

Theorem 2.42 (Immersion Theorem). Let Φ : M m → N m+k be an immersion at

p ∈ M between two smooth manifolds M m and N n+k with k ≥ 1. Given any local
parametrization F : UM → OM of M near p ∈ M , and any local parametrization
G : UN → ON of N near Φ(p) ∈ N , there exists a smooth reparametrization map
ψ : UeN → UN such that:
(G ◦ ψ)−1 ◦ Φ ◦ F (u1 , . . . , um ) = (u1 , . . . , um , 0, . . . , 0).
| {z }
k
See Figure 2.8 for an illustration.

Proof. The proof uses the Inverse Function Theorem. By translation, we may assume that
F (0) = p and G(0) = Φ(p). Given that (Φ∗ )p is injective, there are n linearly independent
rows in the matrix [(Φ∗ )p ]. WLOG we may assume that the first m rows of [(Φ∗ )p ] are
linearly independent. As such, the matrix can be decomposed into the form:
 
A
[(Φp )∗ ] =
∗
where A is an invertible m × m matrix, and ∗ denotes any k × m matrix.
Now define ψ : Rm+k → Rm+k as:
(*)
ψ(u1 , . . . , um , um+1 , . . . , um+k ) = G−1 ◦ Φ ◦ F (u1 , . . . , um ) + (0, . . . , 0, um+1 , . . . , um+k ).
We claim that this is the map ψ that we want. First note that ψ(0) = G−1 ◦ Φ(p) = 0
by our earlier assumption. Next we show that ψ has a smooth inverse near 0. The
Jacobian matrix of this map at 0 is given by:
 
A 0
[(Dψ)0 ] = .
∗ Ik
As rows of A are linearly independent, it is easy to see then all rows of [(Dψ)0 ] are
linearly independent, and hence [(Dψ)0 ] is invertible. By Inverse Function Theorem, ψ is
locally invertible near 0, i.e. there exists an open set UeN ⊂ Rm+k containing 0 such that
the restricted map:
ψ|UeN : UeN → ψ(UeN ) ⊂ UN .
has a smooth inverse.
Finally, we verify that this is the map ψ that we want. We compute:
(G ◦ ψ)−1 ◦ Φ ◦ F (u1 , . . . , um ) = ψ −1 (G−1 ◦ Φ ◦ F )(u1 , . . . , um ) .

By (*), we have ψ(u1 , . . . , um , 0, . . . , 0) = G−1 ◦ Φ ◦ F (u1 , . . . , um ), and hence:

ψ −1 (G−1 ◦ Φ ◦ F )(u1 , . . . , um ) = (u1 , . . . , um , 0, . . . , 0).

It completes our proof. 

54 2. Abstract Manifolds

Figure 2.8. Geometric illustration of the Immersion Theorem.

Example 2.43. Consider the map Φ : R → R2 defined by:

Φ(θ) = (cos θ, sin θ).
It is easy to see that [Φ∗ ] = (− sin θ, cos θ) 6= (0, 0) for any θ. Hence Φ is an immersion.
We can locally parametrize R2 near image of Φ by:
e r) := ((1 − r) cos θ, (1 − r) sin θ),
G(θ,
e −1 ◦Φ(θ) = G
then G e −1 (cos θ, sin θ) = (θ, 0). Note that the Immersion Theorem (Theorem
2.42) asserts that such G e exists, it fails to give an explicit form of such a G.
e 

Exercise 2.30. Consider the sphere S2 = {(x, y, z) : x2 + y 2 + z 2 = 1} in R3 . Find

local parametrizations F for S2 , and G for R3 such that the composition
G−1 ◦ ι ◦ F
takes (u1 , u2 ) to (u1 , u2 , 0). Here ι : S2 → R3 is the inclusion map.

2.5.3. Submersions. Another important type of smooth maps are submersions.

Loosely speaking, a submersion is a map that is “locally surjective”. Here is the rigorous
definition:

Definition 2.44 (Submersions). Let Φ : M → N be a smooth map between smooth

manifolds M and N . We say Φ is a submersion at p ∈ M if the tangent map (Φ∗ )p :
Tp M → TΦ(p) N is surjective. If Φ is a submersion at every point on M , then we simply
say Φ is a submersion.

Remark 2.45. Clearly, in order for Φ : M → N to be a submersion at any point p ∈ M ,

it is necessary that dim M ≥ dim N . 
Example 2.46. Given two smooth manifolds M and N , the projection maps πM :
M × N → M and πN : M × N → N are both submersions. To verify this, recall that
2.5. Immersions and Submersions 55

T(p,q) (M × N ) = Tp M ⊕ Tq N , and from Exercise 2.22 that (πM )∗ = πT M where πT M is

the projection map of the tangent space:
πT M (v, w) = v for any v ∈ Tp M and w ∈ Tq N .
 
The matrix [πT M ] is then of the form: I 0 where I is the identity matrix of dimension
dim M and 0 is the dim M × dim N zero matrix. There are pivots in every row so πT M is
surjective. Similarly we can also show (πN )∗ = πT N is also surjective. 
Example 2.47. Given a smooth function f : Rn → R, and at the point p ∈ Rn such
that ∇f (p) 6= 0, one can show f is a submersion at p. To show this, let {êi }ni=1 be the
standard basis of Rn , then
 
∂ ∂f
f∗ (êi ) = f∗ =
∂xi ∂xi
h i
∂f ∂f
and so the matrix of [f∗ ] is given by ∂x 1
· · · ∂x n
. At the point p, we have ∇f (p) 6= 0
which is equivalent to show [f∗ ] at p is a non-zero 1 × n matrix, which always have 1
pivot in its RREF. Therefore, (f∗ )p is surjective and f is a submersion at p. 

Exercise 2.31. Show that if M and N are two smooth manifolds of equal dimension,
then the following are equivalent:
(i) Φ : M → N is a local diffeomorphism.
(ii) Φ : M → N is an immersion.
(iii) Φ : M → N is a submersion.

Exercise 2.32. Find a smooth map Φ : R → R which is a submersion but is not

surjective.

Exercise 2.33. Show that the map Φ : Rn+1 \{0} → RPn defined as:
Φ(x0 , . . . , xn ) = [x0 : · · · : xn ]
is a submersion.
One nice property of a submersion Φ : M m → N n that locally around every p ∈ M ,
one can find special local parametrizations F of M near p, and G of N near Φ(p) such
that G−1 ◦ Φ ◦ F is a projection map. We will see later that this result will show any
level-set of Φ, if non-empty, must be a smooth manifold. Let’s state this result in a precise
way:

Theorem 2.48 (Submersion Theorem). Let Φ : M n+k → N n be a submersion at p ∈ M

between two smooth manifolds M n+k and N n with k ≥ 1. Given any local parametrization
F : UM → OM of M near p ∈ M , and any local parametrization G : UN → ON of N
near Φ(p) ∈ N , there exists a smooth reparametrization map ψ : UeM → UM such that:
G−1 ◦ Φ ◦ (F ◦ ψ)(u1 , . . . , un , un+1 , . . . , un+k ) = (u1 , . . . , un ).
See Figure 2.9 for illustration.

Proof. The proof uses again the Inverse Function Theorem. First by translation we may
assume that F (0) = p and G(0) = Φ(p). Given that (Φ∗ )p is surjective, there are n
linearly independent columns in the matrix [(Φ∗ )p ]. WLOG assume that they are the first
n columns, then [(Φ∗ )p ] is of the form:
[(Φ∗ )p ] = D(G−1 ◦ Φ ◦ F )0 = A ∗
   
56 2. Abstract Manifolds

where A is an n × n invertible matrix, and ∗ is any n × k matrix.

Now define φ : UM → Rn+k as:
(*) φ(u1 , . . . , un , un+1 , . . . , un+k ) = (G−1 ◦ Φ ◦ F (u1 , . . . , un+k ), un+1 , . . . , un+k ).
| {z }
∈Rn

This map is has an invertible Jacobian matrix at F −1 (p) since:

 
A ∗
[(Dφ)0 ] = .
0 I
By Inverse Function Theorem, there exists a local inverse φ−1 : UeM → φ−1 (UeM ) ⊂ UM .
Finally, we verify that:
G−1 ◦ Φ ◦ (F ◦ φ−1 )(u1 , . . . , un , un+1 , . . . , un+k )
= (G−1 ◦ Φ ◦ F ) φ−1 (u1 , . . . , un+k )

Let φ−1 (u1 , . . . , un+k ) = (v1 , . . . , vn+k ), then φ(v1 , . . . , vn+k ) = (u1 , . . . , un+k ). From
(*), we get:
(G−1 ◦ Φ ◦ F (v1 , . . . , vn+k ), vn+1 , . . . , vn+k ) = φ(v1 , . . . , vn+k )
= (u1 , . . . , un , un+1 , . . . , un+k )
−1
which implies G ◦ Φ ◦ F (v1 , . . . , vn+k ) = (u1 , . . . , un ). Combine with previous result,
we get:
(G−1 ◦ Φ ◦ F ) φ−1 (u1 , . . . , un+k )

= (G−1 ◦ Φ ◦ F )(v1 , . . . , vn+k )

= (u1 , . . . , un ).
Hence, G−1 ◦ Φ ◦ (F ◦ φ−1 ) is the projection from Rn+k onto Rn . It completes the proof
by taking ψ = φ−1 . 

Figure 2.9. Geometric illustration of the Submersion Theorem

2.6. Submanifolds 57

2.6. Submanifolds
In this section we talk about submanifolds. A subspace W of a vector space V is a subset
of V and is itself a vector space. A subgroup H of a group G is a subset of G and is
itself a group. It seems that a smooth submanifold N of a smooth manifold M might be
defined as a subset of M and is itself a smooth manifold. However, it is just one side of
the full story – we need more than that because we hope that the local coordinates of a
submanifold is in some sense compatible with the local coordinates of the manifold M .

Definition 2.49 (Submanifolds). Let M be a smooth n-manifold. A subset N ⊂ M is

said to be a smooth k-submanifold of M if N is a smooth k-manifold and the inclusion
map ι : N → M is an smooth immersion.

Example 2.50. Let Φ : M m → N n be a smooth map. Define ΓΦ to be the graph of Φ.

Precisely:
ΓΦ = {(p, Φ(p)) ∈ M × N : p ∈ M }.
We are going to show that the graph ΓΦ is a submanifold of M ×N , with dim ΓΦ = dim M .
To show this, consider an arbitrary point (p, Φ(p)) ∈ ΓΦ where p ∈ M . The product
manifold M × N can be locally parametrized by F × G where F is a smooth local
parametrization of M near p and G is a smooth local parametrization of N around Φ(p).
ΓΦ is locally parametrized around (p, Φ(p)) by:
Fe(u) := (F (u), Φ ◦ F (u)).
Here for simplicity, we denote u := (u1 , . . . , um ) where m = dim M . It can be verified
that if F1 and F2 are compatible (i.e. with smooth transition maps) parametrizations
of M around p, then the induced parametrizations Fe1 and Fe2 are also compatible (see
exercise below).
Recall that for any u = (u1 , . . . , um ) and v = (v1 , . . . , vn ), the product map F × G
defined as:
(F × G)(u, v) = (F (u), G(v))
is a local parametrization of M × N . Now we show that the inclusion ι : ΓΦ → M × N is
an immersion:
−1 −1
(F × G) ◦ ι ◦ Fe(u) = (F × G) ◦ ι (F (u), Φ(F (u)))
−1
= (F × G) (F (u), Φ(F (u)))
−1


= u, G ◦ Φ ◦ F (u) .
Its Jacobian matrix has the form:
 
I
[ι∗ ] =
[Φ∗ ]
which is injective since its RREF does not have any free column. This completes the proof
that ι is an immersion and so ΓΦ is a submanifold of M × N .


Exercise 2.34. Complete the exercise stated in Example 2.50 that if F1 and F2 are
compatible parametrizations of M around p, then the induced parametrizations Fe1
and Fe2 of ΓΦ are also compatible.
58 2. Abstract Manifolds

Exercise 2.35. Let M m be a smooth manifold. Consider the diagonal subset ∆M ⊂

M × M defined as:
∆M := {(x, x) ∈ M × M : x ∈ M }.
Show that ∆M is a submanifold of M × M .

Exercise 2.36. Show that if N is a submanifold of M , and P is a submanifold of N ,

then P is also a submanifold of M .

Exercise 2.37. Show that any non-empty open subset N of a smooth manifold M
is a submanifold of M with dim N = dim M .
We require a submanifold to have the inclusion map being smooth because we
want to rule out some pathological cases. Consider the graph of an absolute function,
i.e. Γ = {(x, |x|) : x ∈ R}, and R2 . The graph Γ can be parametrized by a single
parametrization F : R → Γ defined by:
F (t) = (t, |t|).
Then, since Γ equipped with this single parametrization, it is considered as a smooth
manifold (although quite difficult to accept) since there is essentially no transition map.
However, we (fortunately) can show that Γ is not a submanifold of R2 (with usual
differential structure, parametrized by the identity map). It is because the inclusion map
is not smooth:
id−1
R2 ◦ ι ◦ F (t) = (t, |t|)

Exercise 2.38. Show that if R2 is (pathologically) parametrized by

G : R2 → R2
(x, y) 7→ (x, y + |x|)
and Γ = {(x, |x|) : x ∈ R} is parametrized by F (t) = (t, |t|), then with differential
structures generated by these parametrizations, Γ becomes a submanifold of R2 .
That says: the “pathologically” smooth manifold Γ is a submanifold of this
“pathological” R2 .

We require the inclusion map is an immersion because we want a submanifold N of M to

be equipped with local coordinates “compatible” with that of M in the following sense:

Proposition 2.51. If N n is a submanifold of M m , then near every point p ∈ N , there

exists a smooth local parametrization G(u1 , . . . , um ) : U → O of M near p such that
G(0) = p and
N ∩ O = {G(u1 , . . . , un , 0, . . . , 0) : (u1 , . . . , un , 0, . . . , 0) ∈ U}.

Proof. By Theorem 2.42 (Immersion Theorem), given that ι : N → M is an immersion,

then around every point p ∈ N one can find a local parametrization F : UN → ON of N
near p, and another local parametrization G(u1 , . . . , um ) : UM → OM of M near ι(p) = p
such that:
G−1 ◦ ι ◦ F (u1 , . . . , un ) = (u1 , . . . , un , 0, . . . , 0)
| {z }
m−n

and so F (u1 , . . . , un ) = G(u1 , . . . , un , 0, . . . , 0). Note that in order for G−1 ◦ ι ◦ F to be

well-defined, we assume (by shrinking the domains if necessary) that ON = N ∩ OM .
2.6. Submanifolds 59

Therefore,
{G(u1 , . . . , un , 0, . . . , 0) : (u1 , . . . , un , 0, . . . , 0) ∈ U}
= {F (u1 , . . . , un ) : (u1 , . . . , un ) ∈ UN }
= ON = N ∩ OM
It completes our proof. 

We introduced submersions because the level-set of a submersion, if non-empty, can

in fact shown to be a submanifold! We will state and prove this result. Using this fact,
one can show a lot of sets are in fact manifolds.

Proposition 2.52. Let Φ : M m → N n be a smooth map between two smooth manifolds M

and N . Suppose q ∈ N such that Φ−1 (q) is non-empty, and that Φ is a submersion at any
p ∈ Φ−1 (q), then the level-set Φ−1 (q) is a submanifold of M with dim Φ−1 (q) = m − n.

Proof. Using Theorem 2.48 (Submersion Theorem), given any point p ∈ Φ−1 (q) ⊂ M ,
there exist a local parametrization F : UM → OM of M near p, and a local parametriza-
tion G of N near Φ(p) = q, such that:
G−1 ◦ Φ ◦ F (u1 , . . . , un , un+1 , . . . , um ) = (u1 , . . . , un )
and that F (0) = p, G(0) = q.
We first show that Φ−1 (q) is a smooth manifold. Note that we have:
Φ (F (0, . . . , 0, un+1 , . . . , um )) = G(0, . . . , 0) = q.
Therefore, F (0, . . . , 0, un+1 , . . . , um ) ∈ Φ−1 (q). Hence, Φ−1 (q) can be locally parametrized
by Fe(un+1 , . . . , um ) := F (0, . . . , 0, un+1 , . . . , um ). One can also verify that compatible
F ’s gives compatible Fe’s. This shows Φ−1 (q) is a smooth manifold of dimension m − n.
To show it is a submanifold of M , we need to compute the tangent map ι∗ . First
consider the composition:
F −1 ◦ ι ◦ Fe(un+1 , . . . , um ) = F −1 (F (0, . . . , 0, un+1 , . . . , um )) = (0, . . . , 0, un+1 , . . . , um ).
The matrix [ι∗ ] with respect to local parametrizations Fe of Φ−1 (q), and F of M is given
by the Jacobian:  
0
[ι∗ ] = [D(F −1 ◦ ι ◦ Fe)] =
I
which shows ι∗ is injective. Therefore, Φ−1 (q) is a submanifold of M . 

Using Proposition 2.52, one can produce a lot of examples of manifolds which are
level-sets of smooth functions.
Example 2.53. In R4 , the set Σ := {x3 + y 3 + z 3 + w3 = 1} is a smooth 3-manifold. It
can be shown by consider Φ : R4 → R defined by:
Φ(x, y, z, w) = x3 + y 3 + z 3 + w3 .
Then, Σ = Φ−1 (1). To show it is a manifold, we show Φ is a submersion at every p ∈ Σ.
By direct computation, we get:
[Φ∗ ] = [3x2 3y 2 3z 2 3w2 ]
Since [Φ∗ ] = 0 only when (x, y, z, w) = (0, 0, 0, 0) which is not contained in Σ, we have
shown (Φ∗ )p is injective for any p ∈ Σ. By Proposition 2.52, we have proved Σ = Φ−1 (1)
is a smooth manifold of dimension 4 − 1 = 3. 
Example 2.54. The set Mn×n (R) of all n × n real matrices can be regarded as R2n
equipped with the usual differential structure. Consider these subsets of Mn×n (R):
60 2. Abstract Manifolds

(a) GL(n, R) = the set of all invertible n × n matrices;

(b) Sym(n, R) = the set of all symmetric n × n matrices;
(c) O(n, R) = the set of all orthogonal matrices;
We are going to show that they are all submanifolds of Mn×n (R). Consider the determi-
nant function f : Mn×n (R) → R defined as:
f (A) := det(A).
Since f is a continuous function, the set GL(n, R) = f −1 (R\{0}) is an open subset of
Mn×n (R). Any (non-empty) open subset N of a smooth manifold M is a submanifold of
M with dim N = dim M (see Exercise 2.37).
n(n+1)
For Sym(n, R), we first label the coordinates of R 2 by (xij ) where and 1 ≤
n(n+1)
i ≤ j ≤ n. Then one can parametrize Sym(n, R) by F : R 2 → Sym(n, R) taking
n(n+1)
(xij )i≤j ∈ R 2 to the matrix A with (i, j)-th entry xij when i ≤ j, and xji when
n(n+1)
i > j. For instance, when n = 2, R 2 becomes R3 with coordinates labelled by
(x11 , x12 , x22 ). The parametrization F will take the point (x11 , x12 , x22 ) = (a, b, c) ∈ R3
to the matrix:  
a b
∈ Sym(n, R).
b c
Back to the general n, this F is a global parametrization and it makes Sym(n, R) a smooth
n(n+1)
2 -manifold. To show that it is a submanifold of Mn×n (R), we computed the tangent
map ι∗ of the inclusion map ι : Sym(n, R) → Mn×n (R):
 
∂ ∂ι ∂
ι∗ = = (ι ◦ F )
∂xij ∂xij ∂xij
1
= (Eij + Eji )
2
where Eij is the n × n matrix with 1 in the (i, j)-th entry, and 0 elsewhere.
 The tangent
∂
space T Sym(n, R) at each point is spanned by the basis . Its image
∂xij 1≤i≤j≤n
 
1
(Eij + Eji ) under the map ι∗ is linearly independent (why?). This shows
2 1≤i≤j≤n
ι∗ is injective, and hence Sym(n, R) is a submanifold of Mn×n (R). The image of the
inclusion map is the set of all symmetric matrices in Mn×n (R), hence we conclude that
TA0 Sym(n, R) ∼ = T Sym(n, R) for any A0 ∈ Sym(n, R).
The set of all orthogonal matrices O(n) can be regarded as the level-set Φ−1 (I) of
the following map:
Φ : Mn×n (R) → Sym(n, R)
A 7→ AT A
We are going to show that Φ is a submersion at every A0 ∈ Φ−1 (I), we compute its
tangent map:
 
∂ ∂
(Φ∗ ) = AT A = Eij
T
A + AT Eij = (AT Eij )T + AT Eij .
∂xij ∂xij
From now on we denote [A]ij to be the (i, j)-th entry of any matrix A (Be cautious that
Eij without the square brackets is a matrix, not the (i, j)-th entry of E). In fact, any
matrix A can be written as:
X n
A= [A]ij Eij .
i,j=1
2.6. Submanifolds 61

At A0 ∈ Φ−1 (I), we have AT0 A0 = I and so for any symmetric matrix B, we have:
 
∂
(Φ∗ )A0 = (AT0 Eij )T + AT0 Eij
∂xij
 
n n
1 X ∂ = 1
X
[A0 B]ij (AT0 Eij )T + AT0 Eij


(Φ∗ )A0  [A0 B]ij
2 i,j=1 ∂xij 2 i,j=1
 T  
n n
1 T X 1 X
= A0 [A0 B]ij Eij  + AT0 [A0 B]ij Eij 
2 i,j=1
2 i,j=1

1 T 1
= (A A0 B)T + (AT0 A0 B)
2 0 2
1 1
= B T + B = B.
2 2
Therefore, (Φ∗ )A0 is surjective. This shows Φ∗ is a submersion at every point A0 ∈
Φ−1 (I). This shows Sym(n, R) = Φ−1 (I) is a submanifold of Mn×n (R) of dimension
dim Mn×n (R) − dim Sym(n, R), which is n2 − n(n+1)
2 = n(n−1)
2 .


Exercise 2.39. Show that the subset Σ of R3 defined by the two equations below is
a 1-dimensional manifold:
x3 + y 3 + z 3 = 1
x+y+z =0

Exercise 2.40. Define

• SL(n, R) = the set of all n × n matrices with determinant 1
• sl(n, R) = the set of all n × n skew-symmetric matrices (i.e. the set of matrices
A ∈ Mn×n (R) such that AT = −A).
Show that they are both submanifolds of Mn×n (R), and find their dimensions.

Exercise 2.41. Consider the map Φ : S3 \{(0, 0)} → CP1 defined by:
Φ(x1 , x2 , x3 , x4 ) := [x1 + ix2 : x3 + ix4 ].
−1
Show that Φ ([1 : 0]) is a smooth manifold of (real) dimension 1, and show that
Φ−1 ([1 : 0]) is diffeomorphic to a circle.
 Chapter 3

Tensors and Differential

Forms

“In the beginning, God said, the

four-dimensional divergence of an
antisymmetric, second-rank tensor
equals zero, and there was light.”

Michio Kaku

In Multivariable Calculus, we learned about gradient, curl and divergence of a vector

field, and three important theorems associated to them, namely Green’s, Stokes’ and
Divergence Theorems. In this and the next chapters, we will generalize these theorems
to higher dimensional manifolds, and unify them into one single theorem (called the
Generalized Stokes’ Theorem). In order to carry out this generalization and unification, we
need to introduce tensors and differential forms. The reasons of doing so are many-folded.
We will explain it in detail. Meanwhile, one obvious reason is that the curl of a vector
field is only defined in R3 since it uses the cross product. In this chapter, we will develop
the language of differential forms which will be used in place of gradient, curl, divergence
and all that in Multivariable Calculus.

3.1. Cotangent Spaces

3.1.1. Review of Linear Algebra: dual spaces. Let V be an n-dimensional real
vector space, and B = {e1 , . . . , en } be a basis for V . The set of all linear maps T : V → R
from V to the scalar field R (they are commonly called linear functionals) forms a vector
space with dimension n. This space is called the dual space of V , denoted by V ∗ .
Associated to the basis B = {ei }ni=1 for V , there is a basis B ∗ = {ei }ni=1 for V ∗ :
(
1 if i = j
ei (ej ) =
0 if i 6= j
The basis B ∗ for V ∗ (do Exericse 3.1 to verify it is indeed a basis) is called the dual basis
of V ∗ with respect to B.

63
64 3. Tensors and Differential Forms

Exercise 3.1. Given that V is a finite-dimensional real vector space, show that:
(a) V ∗ is a vector space
(b) dim V ∗ = dim V
(c) If B = {ei }ni=1 is a basis for V , then B ∗ := {ei }ni=1 is a basis for V ∗ .
Given T ∈ V ∗ and that T (ei ) = ai , verify that:
Xn
T = ai ei .
i=1

3.1.2. Cotangent Spaces of Smooth Manifolds. Let M n be a smooth manifold.

Around p ∈ M , suppose there is a local parametrization F (u1 , . . . , un ). Recall that
the
 tangentn space Tp M at p is defined as the span of partial differential operators
∂
(p) . The cotangent space denoted by Tp∗ M is defined as follows:
∂ui i=1

Definition 3.1 (Cotangent Spaces). Let M n be a smooth manifold. At every p ∈ M ,

the cotangent space of M at p is the dual space of the tangent space Tp M , i.e.:
∗
Tp∗ M = (Tp M ) .
The elements in Tp∗ M are called cotangent vectors of M at p.

Remark 3.2. Some authors use Tp M ∗ to denote the cotangent space. 

 n
∂
Associated to the basis Bp = (p) of Tp M , there is a dual basis Bp∗ =
∂ui i=1
for Tp∗ M , which is defined as follows:

(du1 )p , . . . , (dun )p
  (
∂ 1 if i = j
(dui )p (p) = δij =
∂uj 0 if i 6= j

As (dui )p is a linear map from Tp M to R, from the above definition we have:

 
n n
X ∂ X
(dui )p  aj (p) = aj δij = ai .
j=1
∂uj j=1

Occasionally (just for aesthetic purpose), dui p can be denoted as dui p . Moreover,
whenever p is clear from the context (or not significant), we may simply write dui and
∂
∂ui .
Note that both Bp and Bp∗ depend on the choice of local coordinates. Suppose
(v1 , . . . , vn ) is another local coordinates around p, then by chain rule we have:
n
∂ X ∂uk ∂
=
∂vj ∂vj ∂uk
k=1
n
∂ X ∂vk ∂
= .
∂uj ∂uj ∂vk
k=1
3.1. Cotangent Spaces 65

We are going to express dv i in terms of duj ’s:

  n
!
i ∂ i
X ∂vk ∂
dv = dv
∂uj ∂uj ∂vk
k=1
n
X ∂vk  
i ∂
= dv
∂uj ∂vk
k=1
n
X ∂vk
= δik
∂uj
k=1
∂vi
= .
∂uj
This proves the transition formula for the cotangent basis:
n
X ∂vi k
(3.1) dv i = du .
∂uk
k=1
2
Example 3.3. Consider M = R which can be parametrized by
F1 (x, y) = (x, y)
F2 (r, θ) = (r cos θ, r sin θ).
From (3.1), the conversion between {dr, dθ} and {dx, dy} is given by:
∂x ∂x
dx = dr + dθ
∂r ∂θ
= (cos θ) dr − (r sin θ) dθ
∂y ∂y
dy = dr + dθ
∂r ∂θ
= (sin θ) dr + (r cos θ) dθ


Exercise 3.2. Consider M = R3 which can be parametrized by:

F1 (x, y, z) = (x, y, z)
F2 (r, θ, z) = (r cos θ, r sin θ, z)
F3 (ρ, φ, θ) = (ρ sin φ cos θ, ρ sin φ sin θ, ρ cos φ)
Express {dx, dy, dz} in terms of {dr, dθ, dz} and {dρ, dφ, dθ}.

Exercise 3.3. Suppose F (u1 , . . . , un ) and G(v1 , . . . , vn ) are two local parametriza-
tions of a smooth manifold M . Let ω : M → T M be a smooth differential 1-form
such that on the overlap of local coordinates we have:
X X
ω= aj duj = bi dv i .
j i

Find a conversion formula between aj ’s and bi ’s.

66 3. Tensors and Differential Forms

3.2. Tangent and Cotangent Bundles

3.2.1. Definitions. Let M be a smooth manifold. Loosely speaking, the tangent
bundle (resp. cotangent bundle) are defined as the disjoint union of all tangent (resp.
cotangent) spaces over the whole M . Precisely:

Definition 3.4 (Tangent and Cotangent Bundles). Let M be a smooth manifold. The
tangent bundle, denoted by T M , is defined to be:
[
TM = ({p} × Tp M ) .
p∈M

Elements in T M can be written as (p, V ) where V ∈ Tp M .

Similarly, the cotangent bundle, denoted by T ∗ M , is defined to be:
[
T ∗M = {p} × Tp∗ M .

p∈M

Elements in T M can be written as (p, ω) where ω ∈ Tp∗ M .

∗

Suppose F (u1 , . . . , un ) : U → M is a local parametrization of M , then a general

element of T M can be written as:
n
!
X ∂
p, Vi (p)
i=1
∂ui
and a general element of T ∗ M can be written as:
n
!
X
i
p, ai du p
.
i=1
We are going to explain why both T M and T ∗ M are smooth manifolds. The local
parametrization F (u1 , . . . , un ) of M induces a local parametrization Fe of T M defined
by:
(3.2) Fe : U × Rn → T M
n
!
i ∂
X
1 n
(u1 , . . . , un ; V , . . . , V ) 7→ F (u1 , . . . , un ), V .
i=1
∂ui F (u1 ,...,un )

Likewise, it also induces a local parametrization Fe∗ of T ∗ M defined by:

(3.3) Fe∗ : U × Rn → T ∗ M
n
!
X
i
(u1 , . . . , un ; a1 , . . . , an ) 7→ F (u1 , . . . , un ), ai du F (u1 ,...,un )
.
i=1
It suggests that T M and T ∗ M are both smooth manifolds of dimension 2 dim M . To do
so, we need to verify compatible F ’s induce compatible Fe and Fe∗ . Let’s state this as a
proposition and we leave the proof as an exercise for readers:

Proposition 3.5. Let M n be a smooth manifold. Suppose F and G are two overlapping
smooth local parametrizations of M , then their induced local parametrizations Fe and G
e
defined as in (3.2) on the tangent bundle T M are compatible, and also that F and G∗
e ∗ e
defined as in (3.3) on the cotangent bundle T ∗ M are also compatible.

Corollary 3.6. The tangent bundle T M and the cotangent bundle T ∗ M of a smooth
manifold M are both smooth manifolds of dimension 2 dim M .
3.2. Tangent and Cotangent Bundles 67

Exercise 3.4. Prove Proposition 3.5.

Exercise 3.5. Show that the bundle map π : T M → M taking (p, V ) ∈ T M to

p ∈ M is a submersion. Show also that the set:
Σ0 := {(p, 0) ∈ T M : p ∈ M }
is a submanifold of T M .

3.2.2. Vector Fields. Intuitively, a vector field V on a manifold M is an assignment

of a vector to each point on M . Therefore, it can be regarded as a map V : M → T M
such that V (p) ∈ {p} × Tp M . Since we have shown that the tangent bundle T M is also
a smooth manifold, one can also talk about C k and smooth vector fields.

Definition 3.7 (Vector Fields of Class C k ). Let M be a smooth manifold. A map

V : M → T M is said to be a vector field if for each p ∈ M , we have V (p) = (p, Vp ) ∈
{p} × Tp M .
If V is of class C k as a map between M and T M , then we say V is a C k vector field.
If V is of class C ∞ , then we say V is a smooth vector field.

Remark 3.8. In the above definition, we used V (p) to be denote the element (p, Vp ) in
T M , and Vp to denote the vector in Tp M . We will distinguish between them for a short
while. After getting used to the notations, we will abuse the notations and use Vp and
V (p) interchangeably. 

Remark 3.9. Note that a vector field can also be defined locally on an open set O ⊂ M .
In such case we say V is a C k on O if the map V : O → T M is C k . 

Under a local parametrization F (u1 , . . . , un ) : U → M of M , a vector field V : M →

T M can be expressed in terms of local coordinates as:
n
!
X
i ∂
V (p) = p, V (p) (p) .
i=1
∂ui

The functions V i : F (U) ⊂ M → R are all locally defined and are commonly called the
components of V with respect to local coordinates (u1 , . . . , un ).
Let Fe(u1 , . . . , un ; V 1 , . . . , V n ) be the induced local parametrization of T M defined
as in (3.2). Then, one can verify that:
n
!
−1 −1
X
i ∂
F ◦ V ◦ F (u1 , . . . , un ) = F
e e F (u1 , . . . , un ), V (F (u1 , . . . , un ))
i=1
∂ui F (u1 ,...,un )
1 n


= u1 , . . . , un ; V (F (u1 , . . . , un )), . . . , V (F (u1 , . . . , un )) .

Therefore, Fe−1 ◦ V ◦ F (u1 , . . . , un ) is smooth if and only if the components V i ’s are all
smooth. Similarly for class C k . In short, a vector field V is smooth if and only if the
components V i in every its local expression:
n
!
X
i ∂
V (p) = p, V (p) (p)
i=1
∂ui

are all smooth.

68 3. Tensors and Differential Forms

3.2.3. Differential 1-Forms. Differential 1-forms are the dual counterpart of vector
fields. It is essentially an assignment of a cotangent vector to each point on M . Precisely:

Definition 3.10 (Differential 1-Forms of Class C k ). Let M be a smooth manifold. A

map ω : M → T ∗ M is said to be a differential 1-form if for each p ∈ M , we have
ω(p) = (p, ωp ) ∈ {p} × Tp∗ M .
If ω is of class C k as a map between M and T ∗ M , then we say ω is a C k differential
1-form. If ω is of class C ∞ , then we say ω is a smooth differential 1-form.

Remark 3.11. At this moment we use ω(p) to denote an element in {p} × Tp∗ M , and
ωp to denote an element in Tp∗ M . We will abuse the notations later on and use them
interchangeably, since such a distinction is unnecessary for many practical purposes. 

Remark 3.12. If a differential 1-form ω is locally defined on an open set O ⊂ M , we

may say ω is C k on O to mean the map ω : O → T ∗ M is of class C k . 

Under a local parametrization F (u1 , . . . , un ) : U → M of M , a differential 1-form

ω : M → T ∗ M has a local coordinate expression given by:
n
!
X
i
ω(p) = p, ωi (p) du p
i=1

where ωi : F (U) ⊂ M → R are locally defined functions and are commonly called the
components of ω with respect to local coordinates (u1 , . . . , un ). Similarly to vector fields,
one can show that ω is a C ∞ differential 1-form if and only if all ωi ’s are smooth under
any local coordinates in the atlas of M (see Exercise 3.6).

Exercise 3.6. Show that a differential 1-form ω is C k on M if and only if all

components ωi ’s are C k under any local coordinates in the atlas of M .

Example 3.13. The differential 1-form:

y x
ω=− dx + 2 dy
x2 + y 2 x + y2

is smooth on R2 \{0}, but is not smooth on R2 . 

3.2.4. Push-Forward and Pull-Back. Consider a smooth map Φ : M → N between

two smooth manifolds M and N . The tangent map at p denoted by (Φ∗ )p is the induced
map between tangent spaces Tp M and TΦ(p) N . Apart from calling it the tangent map,
we often call Φ∗ to be the push-forward by Φ, since Φ and Φ∗ are both from the space M
to the space N .
The push-forward map Φ∗ takes tangent vectors on M to tangent vectors on N .
There is another induced map Φ∗ , called the pull-back by Φ, which is loosely defined as
follows:
(Φ∗ ω)(V ) = ω(Φ∗ V )
where ω is a cotangent vector and V is a tangent vector. In order for the above to make
sense, V has to be a tangent vector on M (say at p). Then, Φ∗ V is a tangent vector
in TΦ(p) N . Therefore, Φ∗ ω needs to act on V and hence is a cotangent vector in Tp∗ M ;
∗
whereas ω acts on Φ∗ V and so it should be a cotangent vector in TΦ(p) N . It is precisely
defined as follows:
3.2. Tangent and Cotangent Bundles 69

Definition 3.14 (Pull-Back of Cotangent Vectors). Let Φ : M → N be a smooth map

∗
between two smooth manifolds M and N . Given any cotangent vector ωΦ(p) ∈ TΦ(p) N,
∗ ∗
the pull-back of ω by Φ at p denoted by (Φ ω)p is an element in Tp M and is defined to
be the following linear functional on Tp M :
(Φ∗ ω)p : Tp M → R
(Φ∗ ω)p (Vp ) := ωΦ(p) ((Φ∗ )p (Vp ))

Therefore, one can think of Φ∗ is a map which takes a cotangent vector ωΦ(p) ∈ TΦ(p)
∗
N
to a cotangent vector (Φ∗ ω)p on Tp∗ M . As it is in the opposite direction to Φ : M → N ,
we call Φ∗ the pull-back whereas Φ∗ is called the push-forward.

Remark 3.15. In many situations, the points p and Φ(p) are clear from the context.
Therefore, we often omit the subscripts p and Φ(p) when dealing with pull-backs and
push-forwards. 

Example 3.16. Consider the map Φ : R → R2 \{0} defined by:

Φ(θ) = (cos θ, sin θ).

Let ω be the following 1-form on R2 \{0}:

y x
ω=− dx + 2 dy.
x2 +y 2 x + y2
First note that
x y
  z }| { z }| {
∂ ∂Φ ∂ (cos θ) ∂ ∂ (sin θ) ∂ ∂ ∂
Φ∗ = = + = −y +x .
∂θ ∂θ ∂θ ∂x ∂θ ∂y ∂x ∂y
Therefore, one can compute:
      
∂ ∂ ∂ ∂
(Φ∗ ω) = ω Φ∗ = ω −y +x
∂θ ∂θ ∂x ∂y
   
−y x
= −y +x
x2 + y 2 x2 + y 2
= 1.

Therefore, Φ∗ ω = dθ. 

Example 3.17. Let M := R2 \{(0, 0)} (equipped with polar (r, θ)-coordinates) and
N = R2 (with (x, y)-coordinates), and define:

Φ:M →N
Φ(r, θ) := (r cos θ, r sin θ)

One can verify that:

 
∂ ∂Φ ∂ ∂
Φ∗ = = (cos θ) + (sin θ)
∂r ∂r ∂x ∂y
 
∂ ∂Φ ∂ ∂
Φ∗ = = (−r sin θ) + (r cos θ)
∂θ ∂θ ∂x ∂y
∂ ∂
= −y +x
∂x ∂y
70 3. Tensors and Differential Forms

Hence, we have:
   
∗ ∂ ∂
(Φ dx) = dx Φ∗
∂r ∂r
 
∂ ∂
= dx (cos θ) + (sin θ)
∂x ∂y
= cos θ
    
∂ ∂
(Φ∗ dx) = dx Φ∗
∂θ ∂θ
 
∂ ∂
= dx −y +x
∂x ∂y
= −y = −r sin θ
We conclude:
Φ∗ dx = cos θ dr − r sin θ dθ.


Given a smooth map Φ : M m → N n , and local coordinates (u1 , . . . , um ) of M

around p and local coordinates (v1 , . . . , vn ) of N around Φ(p). One can compute a local
expression for Φ∗ :
n
∗
X
i ∂vi j
(3.4) Φ dv = du
j=1
∂uj

where (v1 , . . . , vn ) is regarded as a function of (u1 , . . . , um ) via the map Φ : M → N .

Exercise 3.7. Prove (3.4).

Exercise 3.8. Express Φ∗ dy in terms of dr and dθ in Example 3.17. Try computing

it directly and then verify that (3.4) gives the same result.

Exercise 3.9. Denote (x1 , x2 ) the coordinates for R2 and (y1 , y2 , y3 ) the coordinates
for R3 . Define the map Φ : R2 → R3 by:
Φ(x1 , x2 ) = (x1 x2 , x2 x3 , x3 x1 ).
Compute Φ (dy ), Φ (dy ) and Φ∗ (dy 3 ).
∗ 1 ∗ 2

Exercise 3.10. Consider the map Φ : R3 \{0} → RP2 defined by:

Φ(x, y, z) = [x : y : z].
Consider the local parametrization F (u1 , u2 ) = [1 : u1 : u2 ] of RP2 . Compute
Φ∗ (du1 ) and Φ∗ (du2 ).

3.2.5. Lie Derivatives. Derivatives of a function f or a vector field Y in Euclidean

spaces along a curve γ(t) : (−ε, ε) → Rn are defined as follows:
d f (γ(t + δ)) − f (γ(t))
Dγ 0 (t) f := (f ◦ γ)(t) = lim
dt δ→0 δ
d Y (γ(t + δ)) − Y (γ(t))
Dγ 0 (t) Y := (Y ◦ γ)(t) = lim
dt δ→0 δ
Now given any vector field X and any point p ∈ Rn , if one can find a curve
γ(t) : (−ε, ε) → M such that γ 0 (t) = X(γ(t)) for t ∈ (−ε, ε) and γ(0) = 0, then it is
3.2. Tangent and Cotangent Bundles 71

well-defined to denote:
(DX Y )p := Dγ 0 (t) Y, (DX f )p := Dγ 0 (t) f at t = 0
By the existence and uniqueness theorems of ODE, such a curve γ(t) exists uniquely
provided that the vector field X is C 1 .
Furthermore, it can also be checked that if γ1 , γ2 : (−ε, ε) → Rn are two curves
with γ1 (0) = γ2 (0) = p and with the same velocity vectors γ10 (0) = γ20 (0) at p, then it
is necessarily that Dγ10 f = Dγ20 f and Dγ10 Y = Dγ20 Y at p. Therefore, just the existence
theorem of ODE is sufficient to argue that DX Y and DX f are well-defined.

Exercise 3.11. Prove the above claim that Dγ10 f = Dγ20 f and Dγ10 Y = Dγ20 Y at p.

Remark 3.18. Consult any standard textbook about theory of ODEs for a proof of
existence and uniqueness of the curve γ(t) given any vector field X. Most standard
textbook uses contraction mapping to prove existence, and Gronwall’s inequality to prove
uniqueness. 

Now let M be a smooth manifold, and X be a smooth vector field on M . Then, one
can also extend the existence and uniqueness theorem of ODE to manifolds to prove that
for any point p ∈ M , there exists a smooth curve γ(t) : (−ε, ε) → M on M with γ(0) = p
such that:
d
γ(t) = X(γ(t)).
dt
d ∂


Recall that dt γ(t) is defined as γ∗ ∂t . This curve γ is called the integral curve of X
passing through p. This extension can be justified by applying standard ODE theorems
on the local coordinate chart covering p. Then one solves for the integral curve within
this chart until the curve approaches the boundary of the chart (say at point q). Since
the boundary of one chart must be the interior of another local coordinate chart, one can
then continue solving for the integral curve starting from q.

2.5 2.5

2.0 2.0

1.5 1.5

1.0 1.0

0.5 0.5

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

(a) vector field X (b) integral curves of X

Figure 3.1. a vector field and its integral curves

Now one can still talk about integral curves γ(t) given a vector field X on a manifold,
so one can define Dγ 0 (t) f and DX f in the same way as in Rn (as it makes perfect sense
to talk about f (γ(t + δ)) − f (γ(t)). However, it is not straight-forward how to generalize
the definitions of Dγ 0 (t) Y and DX Y where Y is a vector field on a manifold. The vectors
Y (γ(t + δ)) and Y (γ(t)) are at different based points, so one cannot make sense of
Y (γ(t + δ)) − Y (γ(t)).
72 3. Tensors and Differential Forms

One notion of differentiating a vector field by another one is called the Lie derivatives.
The key idea is to push-forward tangent vectors in a natural way so that they become
vectors at the same based point. Then, it makes sense to consider subtraction of vectors
and also derivatives.
To begin with, we denote the integral curves using a map. First fix a vector field X
on a manifold M . Then, given any point p ∈ M , as discussed before, one can find an
integral curve γ(t) so that γ(0) = p and γ 0 (t) = X(γ(t)). We denote this curve γ(t) by
Φt (p), indicating that it depends on p. Now, for any fixed t, we can view Φt : M → M as
a map. One nice way to interpret this map is to regard Φt (p) as the point on M reached
by flowing p along the vector field X for t unit time. As such, this map Φt is often called
the flow map.

Exercise 3.12. Consider the unit sphere S 2 parametrized by spherical coordinates.

∂
(θ, ϕ), and the vector field X = . Describe the flow map Φt for this vector field,
∂θ
i.e. state how Φt maps the point with coordinates (θ, ϕ).

There are many meaningful purposes of this interpretation of integral curves. Stan-
dard theory of ODE shows Φt is smooth as long as X is a smooth vector field. Moreover,
given s, t ∈ R and p ∈ M , we can regard Φs (Φt (p)) as the point obtained by flowing p
along X first for t unit time, then for s unit time. Naturally, one would expect that the
point obtained is exactly Φs+t (p). It is indeed true provided that X is independent of t.

Proposition 3.19. Given any smooth vector field X on a smooth manifold M , and denote
its flow map by Φt : M → M . Then, given any t, s ∈ R and p ∈ M , we have:
(3.5) Φt (Φs (p)) = Φt+s (p), or equivalently Φt ◦ Φs = Φt+s .
Consequently, for each fixed t ∈ R, the flow map Φt is a diffeomorphism with inverse Φ−t .

Proof. The proof is a direct consequence of the uniqueness theorem of ODE. Consider s
as fixed and t as the variable, then Φt (Φs (p)) and Φt+s (p) can be regarded as curves on
M . When t = 0, both curves pass through the point Φs (p). It remains to show that both
curves satisfy the same ODE, then uniqueness theorem of ODE guarantee that the two
curves must be the same. We leave the detail as an exercise for readers. 

Exercise 3.13. Complete the detail of the above proof that the curves {Φt (Φs (p))}t∈R
and {Φt+s (p)}t∈R both satisfy the same ODE.

Now we are ready to introduce Lie derivatives of vector fields. Given two vector
fields X and Y , we want to develop a notion of differentiating Y along X, i.e. the rate of
change of Y when moving along integral curves of X. Denote the flow map of X by Φt .
Fix a point p ∈ M , we want to compare YΦt (p) with Yp . However, they are at different
base points, so we push-forward YΦt (p) so that it becomes a vector based at p. To do so,
the natural way is to push it forward by the map Φ−t as it maps tangent vectors at Φt (p)
to tangent vectors at Φ−t (Φt (p)) = p.

Definition 3.20 (Lie Derivatives of Vector Fields). Let X and Y be smooth vector fields
on a manifold M . We define the Lie derivative of X along Y by:
d

(LX Y )p := (Φ−t )∗ YΦt (p)
dt t=0
where Φt denotes the flow map of X.
3.2. Tangent and Cotangent Bundles 73

It sounds like a very technical definition that is very difficult to compute! Fortunately,
we will prove that LX Y is simply the commutator [X, Y ], to be defined below. First recall
that a vector field on a manifold is a differential operator acting on scalar functions f .
After differentiating f by a vector field Y , we get another scalar function Y (f ). Then,
we can differentiate Y (f ) by another vector field X and obtaining X(Y (f )) (which for
simplicity we denote it by XY f . The commutator, or the Lie brackets, measure the
difference between XY f and Y Xf :

Definition 3.21 (Lie Brackets). Given two vector fields X and Y on a manifold M , we
define the Lie brackets [X, Y ] to be the vector field such that for any smooth function
f : M → R, we have:
[X, Y ]f := XY f − Y Xf.

Remark 3.22. Suppose under local coordinates (u1 , . . . , un ), the vector fields X and Y
can be written as:
n n
X ∂ X ∂
X= Xi Y = Yi ,
i=1
∂ui i=1
∂ui
then [X, Y ] has following the local expression:
n 
∂Y j ∂X j

X ∂
(3.6) [X, Y ] = Xi −Yi
i,j=1
∂ui ∂ui ∂u j

Exercise 3.14. Verify (3.6), i.e. show that for any smooth function f : M → R, we
have:
n  j j

i ∂Y i ∂X ∂f
X
XY f − Y Xf = X −Y
i,j=1
∂u i ∂ui ∂uj

Exercise 3.15. Let (u1 , . . . , un ) be a local coordinate of a manifold M , and define

∂ ∂
X= Y =
∂ui ∂uj
Then, what is [X, Y ]?

Exercise 3.16. Let X, Y, Z be vector fields on a manifold M , and ϕ : M → R be a

smooth scalar function. Show that:
(i) [X + Y, Z] = [X, Z] + [Y, Z]
(ii) [X, ϕY ] = (Xϕ)Y + ϕ[X, Y ]

It appears that [X, Y ] is more like an algebraic operation whereas Lie derivative LX Y
is a differential operation. Amazingly, they are indeed equal!

Proposition 3.23. Let X and Y be smooth vector fields on a manifold M . Then, we have:
LX Y = [X, Y ].

Proof. Denote Φt to be the flow map of the vector field X. Fix a point p ∈ M and let
F (u1 , . . . , un ) : U → M be a local parametrization covering p. In order to compute LX Y
at p, we may assume that t is sufficiently small so that Φt (p) is also covered by F . Denote
that coordinate representation of Φt by:
F −1 ◦ Φt ◦ F (u1 , . . . , un ) = (vt1 (u1 , . . . , un ), . . . , vtn (u1 , . . . , un )).
74 3. Tensors and Differential Forms

In local coordinates, the flow map Φt is then related to X under the relation:
n n
X ∂vti ∂ X ∂
i
= X i (Φt (p)) i .
i=1
∂t ∂u Φt (p) i=1 ∂u Φt (p)
| {z } | {z }
∂Φt XΦt (p)
∂t p

Equating the coefficients, we have

∂vti
(3.7) = X i (Φt (p))
∂t
for i = 1, . . . , n. Recall

that the Lie derivative (LX Y )p is the time derivative at t = 0 of
(Φ−t )∗ Y (Φt (p)) , which is given by:
n
!
X ∂
(Φ−t )∗ YΦt (p) = (Φt )−1 Y i (Φt (p))


∗ (Φt (p))
i=1
∂ui
n  
X ∂
= Y i (Φt (p)) · (Φt )−1
∗ (Φ t (p)) .
i=1
∂ui
It then follows that:
(3.8)
∂

(LX Y )p = (Φ−t )∗ YΦt (p)
∂t t=0
n  
X ∂ i −1 ∂
= Y (Φt (p)) · (Φt )∗ (Φt (p))
i=1
∂t t=0 ∂ui t=0
n  
X
i ∂ −1 ∂
+ Y (Φt (p)) · (Φt )∗ (Φt (p))
i=1 t=0 ∂t t=0 ∂ui
n n  
X ∂ ∂ X ∂ ∂
= Y i (Φt (p)) · (p) + Y i (p) · (Φt )−1
∗ (Φt (p)) .
i=1
∂t t=0 ∂ui i=1
∂t t=0 ∂ui
∂ i
To compute ∂t Y (Φt (p)), we use the chain rule:
n
X ∂Y i ∂v j
∂ i t
(3.9) Y (Φt (p)) =
∂t j=1
∂uj ∂t
n
X ∂Y i
= X j (Φt (p)) (from (3.7)).
j=1
∂u j
 
∂ −1 ∂
The term ∂t t=0
(Φ t )∗ ∂ui (Φ t (p)) is a bit more tricky. We first define Zik by the
coefficients of:
n   n
!
k ∂ ∂ ∂ ∂
X X
−1 k
Zi (p) = (Φt )∗ (Φt (p)) =⇒ Zi · (Φt )∗ = .
∂uk ∂ui ∂uk p ∂ui Φt (p)
k=1 k=1

Recall that the coordinate representation of Φt is given by vtk ’s, so we get:

n
X ∂vtj ∂ ∂
Zik = .
∂uk ∂uj Φt (p) ∂ui Φt (p)
j,k=1

By the linear independence of coordinate vectors, the coefficients Zik satisfy:

n
X ∂vtj
Zik = δij .
∂uk
k=1
3.2. Tangent and Cotangent Bundles 75

Differentiate both sides with respect to t, we get:

n
!
X ∂Zik ∂vtj ∂X j
+ Zik =0
∂t ∂uk ∂uk t=0
k=1

where we have used (3.7). At t = 0, we have vtj = uj , hence

n 
∂vtj ∂Z k ∂X j
X 
= δjk =⇒ Zij (p) = δij and i
δjk + δik = 0.
∂uk t=0 | {z } ∂t ∂uk t=0
k=1
from definition

It implies that:
∂Zij ∂X j
=− .
∂t t=0 ∂ui
Then, we can compute that:
  n
∂ ∂ ∂ X ∂
(3.10) (Φt )−1
∗ (Φt (p)) = Zik (p)
∂t t=0 ∂ui ∂t t=0 k=1 ∂uk
n
X ∂X k ∂
=− (p).
∂ui ∂uk
k=1

Finally, substitute (3.9) and (3.10) back into (3.8), we get:

n X
n n Xn
X ∂Y i ∂ X ∂X k ∂
(LX Y )p = X j (p) (p) − Y i (p) (p),
i=1 j=1
∂uj ∂ui i=1
∂ui ∂uk
k=1

which is exactly [X, Y ] at p according to (3.6). 

Now we know the geometric meaning of two commuting vector fields X and Y , i.e.
[X, Y ] = 0. According to Proposition 3.23, it is equivalent to saying LX Y = 0 at any
p ∈ M . Then by the definition of Lie derivatives, we can conclude that:
(Φ−t )∗ (YΦt (p) ) = (Φ−0 )∗ (YΦ0 (p) ) = Yp for any t.
In other words, we have YΦt (p) = (Φt )∗ (Yp ) for any t, meaning that pushing Y at p
forward by the flow map Φt of X will yield the vector field Y at the point Φt (p). This
result can further extends to show the flow maps of X and Y commute:

Exercise 3.17. Let X and Y be two vector fields on M such that [X, Y ] = 0. Denote
Φt and Ψt be the flow maps of X and Y respectively, show that for any s, t ∈ R, we
have:
Φs ◦ Ψt = Ψt ◦ Φs .
Sketch a diagram to illustrate its geometric meaning.

Lie derivatives on 1-forms can be defined similarly as on vector fields, except that we
uses pull-backs instead of push-forwards this time.

Definition 3.24 (Lie Derivatives of Differential 1-Forms). Let X and a smooth vector
field and α be a smooth 1-form on a manifold M . Denote the flow map of X by Φt ,
then we define the Lie derivative of α at p ∈ M along X by:
d
(LX α)p := (Φt )∗ αΦt (p) .
dt t=0
76 3. Tensors and Differential Forms

One can compute similarly as in Proposition 3.23 that the Lie derivative of a 1-form
Xn
α= αi dui can be locally expressed as:
i=1
n 
∂X i

X ∂αj
(3.11) LX α = Xi + αi duj
i,j=1
∂ui ∂uj

Exercise 3.18. Verify (3.11).

Exercise 3.19. Let X and Y be two vector fields on M , and α be a 1-form on M .

Show that:
X (α(Y )) = (LX α) (Y ) + α (LX Y ) .
3.3. Tensor Products 77

3.3. Tensor Products

In Differential Geometry, tensor products are often used to produce bilinear, or in
general multilinear, maps between tangent and cotangent spaces. The first and second
fundamental forms of a regular surface, the Riemann curvature, etc. can all be expressed
using tensor notations.

3.3.1. Tensor Products in Vector Spaces. Given two vector spaces V and W ,
their dual spaces V ∗ and W ∗ are vector spaces of all linear functionals T : V → R and
S : W → R respectively. Pick two linear functionals T ∈ V ∗ and S ∈ W ∗ , their tensor
product T ⊗ S is a map from V × W to R defined by:
T ⊗S :V ×W →R
(T ⊗ S)(X, Y ) := T (X) S(Y )
It is easy to verify that T ⊗ S is bilinear, meaning that it is linear at each slot:
(T ⊗ S) (a1 X1 + a2 X2 , b1 Y1 + b2 Y2 )
= a1 b1 (T ⊗ S)(X1 , Y1 ) + a2 b1 (T ⊗ S)(X2 , Y1 )
+ a1 b2 (T ⊗ S)(X1 , Y2 ) + a2 b2 (T ⊗ S)(X1 , Y2 )

Given three vector spaces U, V, W , and linear functionals TU ∈ U ∗ , TV ∈ V ∗ and

TW ∈ W ∗ , one can define a triple tensor product TU ⊗ (TV ⊗ TW ) by:
TU ⊗ (TV ⊗ TW ) : U × (V × W ) → R
(TU ⊗ (TV ⊗ TW ))(X, Y, Z) := TU (X) (TV ⊗ TW )(Y, Z)
= TU (X) TV (Y ) TW (Z)
One check easily that (TU ⊗ TV ) ⊗ TW = TU ⊗ (TV ⊗ TW ). Since there is no ambiguity, we
may simply write TU ⊗ TV ⊗ TW . Inductively, given finitely many vector spaces V1 , . . . , Vk ,
and linear functions Ti ∈ Vi∗ , we can define the tensor product T1 ⊗ · · · ⊗ Tk as a k-linear
map by:
T1 ⊗ · · · ⊗ Tk : V1 × · · · × Vk → R
(T1 ⊗ · · · ⊗ Tk )(X1 , . . . , Xk ) := T1 (X1 ) · · · Tk (Xk )

Given two tensor products T1 ⊗ S1 : V × W → R and T2 ⊗ S2 : V × W → R, one can

form a linear combination of them:
α1 (T1 ⊗ S1 ) + α2 (T2 ⊗ S2 ) : V × W → R
(α1 (T1 ⊗ S1 ) + α2 (T2 ⊗ S2 ))(X, Y ) := α1 (T1 ⊗ S1 )(X, Y ) + α2 (T2 ⊗ S2 )(X, Y )
The tensor products T ⊗ S with T ∈ V ∗ and S ∈ W ∗ generate a vector space. We denote
this vector space by:
V ∗ ⊗ W ∗ := span{T ⊗ S : T ∈ V ∗ and S ∈ W ∗ }.

Exercise 3.20. Verify that α (T ⊗ S) = (αT ) ⊗ S = T ⊗ (αS). Therefore, we can

simply write αT ⊗ S.

Exercise 3.21. Show that the tensor product is bilinear in a sense that:
T ⊗ (α1 S1 + α2 S2 ) = α1 T ⊗ S1 + α2 T ⊗ S2
and similar for the T slot.
78 3. Tensors and Differential Forms

Let’s take the dual basis as an example to showcase the use of tensor products.
Consider a vector space V with a basis {ei }ni=1 . Let {ei }ni=1 be its dual basis for V ∗ . Then,
one can check that:

(ei ⊗ ej )(ek , el ) = ei (ek ) ek (el )

= δik δjl
(
1 if i = k and j = l
=
0 otherwise
n
X
Generally, the sum Aij ei ⊗ ej will act on vectors in V by:
i,j=1
  !
n
X n
X n
X
i j
 Aij e ⊗ e αk ek , βl e l
i,j=1 k=1 l=1
n
X n
X n
X
= Aij αk βl (ei ⊗ ej )(ek , el ) = Aij αk βl δik δjl = Akl αk βl
i,j,k,l=1 i,j,k,l=1 k,l=1

n
X
In other words, the sum of tensor products Aij ei ⊗ ej is the inner product on V
i,j=1
represented by the matrix [Akl ] with respect to the basis {ei }ni=1 of V . For example, when
n
X Xn
Akl = δkl , then Aij ei ⊗ ej = ei ⊗ ei . It is the usual dot product on V .
i,j=1 i=1

Exercise 3.22. Show that {ei ⊗ej }ni,j=1 is a basis for V ∗ ⊗V ∗ . What is the dimension
of V ∗ ⊗ V ∗ ?

Exercise 3.23. Suppose dim V = 2. Let ω ∈ V ∗ ⊗ V ∗ satisfy:

ω(e1 , e1 ) = 0 ω(e1 , e2 ) = 3
ω(e2 , e1 ) = −3 ω(e2 , e2 ) = 0
Express ω in terms of ei ’s.

To describe linear or multilinear map between two vector spaces V and W (where
W is not necessarily the one-dimensional space R), one can also use tensor products.
Given a linear functional f ∈ V ∗ and a vector w ∈ W , we can form a tensor f ⊗ w, which
is regarded as a linear map f ⊗ w : V → W defined by:

(f ⊗ w)(v) := f (v)w.

Let {ei } be a basis for V , and {Fj } be a basis for W . Any linear map T : V → W can
be expressed in terms of these bases. Suppose:
X j
T (ei ) = Ai Fj .
j

Then, we claim that T can be expressed using the following tensor notations:
X j
T = Ai ei ⊗ Fj
i,j
3.3. Tensor Products 79

Let’s verify this. Note that a linear map is determined by its action on the basis {ei } for
V . It suffices to show:  
X j
 Ai ei ⊗ Fj  (ek ) = T (ek ).
i,j

Using the fact that:

(ei ⊗ Fj )(ek ) = ei (ek )Fj = δik Fj ,
one can compute:
 
X j X j
 Ai ei ⊗ Fj  (ek ) = Ai (ei ⊗ Fj )(ek )
i,j i,j
X X
= Aji δik Fj = Ajk Fj = T (ek )
i,j j

as desired.
Generally, if T1 , . . . , Tk ∈ V ∗ and X ∈ V , then
T1 ⊗ · · · ⊗ Tk ⊗ X
is regarded to be a k-linear map from V × . . . × V to V , defined by:
T1 ⊗ · · · ⊗ Tk ⊗ X : V × . . . × V → V
| {z }
k
(T1 ⊗ · · · ⊗ Tk ⊗ X)(Y1 , . . . , Yk ) := T1 (Y1 ) · · · Tk (Yk ) X
Example 3.25. One can write the cross-product in R3 using tensor notations. Think of
the cross product as a bilinear map ω : R3 × R3 → R3 that takes two input vectors u and
v, and outputs the vector u × v. Let {e1 , e2 , e3 } be the standard basis in R3 (i.e. {î, ĵ, k̂}).
Then one can write:
ω = e1 ⊗ e2 ⊗ e3 − e2 ⊗ e1 ⊗ e3
+ e2 ⊗ e3 ⊗ e1 − e3 ⊗ e2 ⊗ e1
+ e3 ⊗ e1 ⊗ e2 − e1 ⊗ e3 ⊗ e2
One can check that, for instance, ω(e1 , e2 ) = e3 , which is exactly e1 × e2 = e3 . 

3.3.2. Tensor Products on Smooth Manifolds. In the previous subsection we take

tensor products on a general abstract vector space V . In this course, we will mostly deal
with the case when V is the tangent or cotangent space of a smooth manifold M .
Recall that ifF (u1 , . .. , un ) is a local parametrization of M , then there is a local
n
∂
coordinate basis (p) for the tangent space Tp M at every p ∈ M covered by F .
∂ui j=1
on  
∗
n
j ∂
The cotangent space Tp M has a dual basis du p defined by duj = δij at
j=1 ∂ui
every p ∈ M .
∂
Then, one can take tensor products of dui ’s and ∂ui ’s
to express multilinear maps be-
Xn
tween tangent and cotangent spaces. For instance, the tensor product g = gij dui ⊗ duj ,
i,j=1
where gij ’s are scalar functions, means that it is a bilinear map at each point p ∈ M such
that:
Xn Xn
g(X, Y ) = gij (dui ⊗ duj )(X, Y ) = gij dui (X) duj (Y )
i,j=1 i,j=1
80 3. Tensors and Differential Forms

for any vector field X, Y ∈ T M . In particular, we have:

 
∂ ∂
g , = gij .
∂ui ∂uj

We can also express multilinear maps from Tp M × Tp M × Tp M to Tp M . For instance,

we let:
n
X
l ∂
Rm = Rijk dui ⊗ duj ⊗ duk ⊗ .
∂ul
i,j,k,l=1

Then, Rm is a mutlilinear map at each p ∈ M such that:

n
X
l ∂
Rm(X, Y, Z) = Rijk dui (X) duj (Y ) duk (Z) .
∂ul
i,j,k,l=1

It is a trilinear map such that:

  n
∂ ∂ ∂ X
l ∂
Rm , , = Rijk .
∂ui ∂uj ∂uk ∂ul
l=1

We call g a (2, 0)-tensor (meaning that it maps two vectors to a scalar), and Rm a
(3, 1)-tensor (meaning that it maps three vectors to one vector). In general, we can also
define (k, 0)-tensor ω on M which has the general form:

n
X
ωp = ωi1 i2 ···ik (p) dui1 p
⊗ · · · ⊗ duik p
i1 ,...,ik =1

 
∂
Here ωi1 i2 ···ik ’s are scalar functions. This tensor maps the tangent vectors ∂ui1 , . . . , ∂u∂i
k
to the scalar ωi1 i2 ...ik at the corresponding point.
Like the Rm-tensor, we can also generally define (k, 1)-tensor Ω on M which has the
general form:

n
X ∂
Ωp = Ωji1 i2 ···ik (p) dui1 p
⊗ · · · ⊗ duik p
⊗ (p)
i1 ,...,ik ,j=1
∂uj

 
where Ωji1 i2 ...ik ’s are scalar functions. This tensor maps the tangent vectors ∂
∂ui1 , . . . , ∂u∂i
k

to the tangent vector j Ωji1 i2 ...ik ∂u ∂

P
j
at the corresponding point.
l
Note that these gij , Rijk , ωi1 i2 ···ik and Ωji1 i2 ...ik are scalar functions locally defined
on the open set covered by the local parametrization F , so we can talk about whether
they are smooth or not:
3.3. Tensor Products 81

Definition 3.26 (Smooth Tensors on Manifolds). A smooth (k, 0)-tensor ω on M is a

k-linear map ωp : Tp M × . . . × Tp M → R at each p ∈ M such that under any local
| {z }
k
parametrization F (u1 , . . . , un ) : U → M , it can be written in the form:
Xn
ωp = ωi1 i2 ···ik (p) dui1 p ⊗ · · · ⊗ duik p
i1 ,...,ik =1

where ωi1 i2 ...ik ’s are smooth scalar functions locally defined on F (U).
A smooth (k, 1)-tensor Ω on M is a k-linear map Ωp : Tp M × . . . × Tp M → Tp M at
| {z }
k
each p ∈ M such that under any local parametrization F (u1 , . . . , un ) : U → M , it can
be written in the form:
n
X ∂
Ωp = Ωji1 i2 ···ik (p) dui1 p ⊗ · · · ⊗ duik p ⊗ (p)
i ,...,i ,j=1
∂uj
1 k

where Ωji1 i2 ...ik ’s are smooth scalar functions locally defined on F (U).

Remark 3.27. Since Tp M is finite dimensional, from Linear Algebra we know (Tp M )∗∗
∂
is isomorphic to Tp M . Therefore, a tangent vector ∂u i
(p) can be regarded as a linear
∗
functional on cotangent vectors in Tp M , meaning that:
∂  
duj p = δij .
∂ui p
Under this interpretation, one can also view a(k, 1)-tensor Ω as a(k + 1)-linear map Ωp :
Tp M × . . . × Tp M ×Tp∗ M → R, which maps ∂
∂ui1 , . . . , ∂u∂i , duj to Ωji1 i2 ...ik . However,
| {z } k

k
we will not view a (k, 1)-tensor this way in this course.
Generally, we can also talk about (k, s)-tensors,which is a (k + s)-linear map
 Ωp :
Tp M × . . . × Tp M × Tp∗ M × . . . × Tp∗ M → R taking ∂
∂ui1 , . . . , ∂u∂i , duj1 , . . . , dujs to a
| {z } | {z } k

k s
scalar. However, we seldom deal with these tensors in this course. 

Exercise 3.24. Let M be a smooth manifold with local coordinates (u1 , u2 ). Consider
the tensor products:
∂
T1 = du1 ⊗ du2 and T2 = du1 ⊗ .
∂u2
Which of the following is well-defined?
 
∂
(a) T1
∂u1
 
∂
(b) T2
∂u1
 
∂ ∂
(c) T1 ,
∂u1 ∂u2
 
∂ ∂
(d) T2 ,
∂u1 ∂u2
82 3. Tensors and Differential Forms

Exercise 3.25. let M be a smooth manifold with local coordinates (u1 , u2 ). The
linear map T : Tp M → Tp M satisfies:
 
∂ ∂ ∂
T = +
∂u1 ∂u1 ∂u2
 
∂ ∂ ∂
T = − .
∂u2 ∂u1 ∂u2
Express T using tensor products.

One advantage of using tensor notations, instead of using matrices, to denote a

multilinear map between tangent or cotangent spaces is that one can figure out the
conversion rule between local coordinate systems easily (when compared to using
matrices)

Example 3.28. Consider the extended complex plane M := C ∪ {∞} defined in Example
2.12. We cover M by two local parametrizations:

F1 : R2 → C ⊂ M F2 : R2 → (C\{0}) ∪ {∞} ⊂ M
1
(x, y) 7→ x + yi (u, v) 7→
u + vi
The transition maps on the overlap are given by:
 
x y
(u, v) = F2−1 ◦ F1 (x, y) = , −
x2 + y 2 x2 + y 2
 
u v
(x, y) = F1−1 ◦ F2 (u, v) = , −
u2 + v 2 u2 + v 2
Consider the (2, 0)-tensor ω defined using local coordinates (x, y) by:
2
+y 2 )
ω = e−(x dx ⊗ dy.

Using the chain rule, we can express dx and dy in terms of du and dv:
(u2 + v 2 ) du − u(2u du + 2v dv)
 
u
dx = d 2 2
=
u +v (u2 + v 2 )2
v 2 − u2 2uv
= 2 du − 2 dv
(u + v 2 )2 (u + v 2 )2
(u2 + v 2 )dv − v(2u du + 2v dv)
 
v
dy = −d = −
u2 + v 2 (u2 + v 2 )2
2 2
2uv v −u
=− 2 du + 2 dv
(u + v 2 )2 (u + v 2 )2
Therefore, we get:
2uv(u2 − v 2 ) (u2 − v 2 )2
dx ⊗ dy = du ⊗ du + du ⊗ dv
(u2 + v 2 )4 (u2 + v 2 )4
4u2 v 2 2uv(u2 − v 2 )
+ 2 dv ⊗ du + dv ⊗ dv
(u + v 2 )4 (u2 + v 2 )4
2
+y 2 )
Recall that ω = e−(x dx ⊗ dy, and in terms of (u, v), we have:
2 1
+y 2 ) − u2 +v
e−(x =e 2
.
3.3. Tensor Products 83

Hence, in terms of (u, v), ω is expressed as:

2uv(u2 − v 2 ) (u2 − v 2 )2

1
− u2 +v
ω=e 2
du ⊗ du + du ⊗ dv
(u2 + v 2 )4 (u2 + v 2 )4
4u2 v 2 2uv(u2 − v 2 )

+ 2 dv ⊗ du + dv ⊗ dv
(u + v 2 )4 (u2 + v 2 )4


Exercise 3.26. Consider the extended complex plane C ∪ {∞} as in Example 3.28,
and the (1, 1)-tensor of the form:
2 2 ∂
Ω = e−(x +y ) dx ⊗ .
∂y
Express Ω in terms of (u, v).

Generally, if (u1 , . . . , un ) and (v1 , . . . , vn ) are two overlapping local coordinates on a

smooth manifold M , then given a (2, 0)-tensor:
X
g= gij dui ⊗ duj
i,j

written using the ui ’s coordinates, one can convert it to vα ’s coordinates by the chain
rule:
!  
X X X ∂ui X ∂uj
g= gij dui ⊗ duj = gij dv α ⊗  dv β 
i,j i,j α
∂v α ∂v β
β
 
X X ∂ui ∂uj  α
=  gij dv ⊗ dv β
i,j
∂vα ∂vβ
α,β

Exercise 3.27. Given that ui ’s and vα ’s are overlapping local coordinates of a smooth
manifold M . Using these coordinates, one can express the following (3, 1)-tensor in
two ways:
X
l ∂ X
eη dv α ⊗ dv β ⊗ dv γ ⊗ ∂
Rm = Rijk dui ⊗ duj ⊗ duk ⊗ = R αβγ
∂ul ∂vη
i,j,k,l α,β,γ,η
l η
Express Rijk in terms of Rαβγ ’s.

Exercise 3.28. Given that ui ’s and vα ’s are overlapping local coordinates of a smooth
manifold M . Suppose g and h are two (2, 0)-tensors expressed in terms of local
coordinates as:
X X
g= gij dui ⊗ duj = geαβ dv α ⊗ dv β
i,j α,β
X X
i j
h= hij du ⊗ du = hαβ dv α ⊗ dv β .
e
i,j α,β

Let G be the matrix with gij as its (i, j)-th entry, and let g ij be the (i, j)-th entry of
G−1 . Similarly, define geαβ to be the inverse of geαβ . Show that:
X X
g ik hkj dui ⊗ duj = geαγ e
hγβ dv α ⊗ dv β .
i,j α,β
84 3. Tensors and Differential Forms

3.4. Wedge Products

Recall that in Multivariable Calculus, the cross product plays a crucial role in many
aspects. It is a bilinear map which takes two vectors to one vectors, and so it is a
(2, 1)-tensor on R3 .
The determinant is another important quantity in Multivariable Calculus and Linear
Algebra. Using tensor languages, an n × n determinant can be regarded as a n-linear
map taking n vectors in Rn to a scalar. For instance, for the 2 × 2 case, one can view:
 
a b
det
c d
as a bilinear map taking column vectors (a, c)T and (b, d)T in R2 to a number ad − bc.
Therefore, it is a (2, 0)-tensor on R2 ; and generally for n × n, the determinant is an
(n, 0)-tensor on Rn .
Both the cross product in R3 and determinant (n × n in general) are alternating, in a
sense that interchanging any pair of inputs will give a negative sign for the output. For
the cross product, we have a × b = −b × a; and for the determinant, switching any pair
of columns will give a negative sign:
a b b a
=− .
c d d c

In the previous section we have seen how to express k-linear maps over tangent
vectors using tensor notations. To deal with the above alternating tensors, it is more
elegant and concise to use alternating tensors, or wedge products that we are going to
learn in this section.

3.4.1. Wedge Product on Vector Spaces. Let’s start from the easiest case. Sup-
pose V is a finite dimensional vector space and V ∗ is the dual space of V . Given any two
elements T, S ∈ V ∗ , the tensor product T ⊗ S is a map given by:
(T ⊗ S)(X, Y ) = T (X) S(Y )
for any X, Y ∈ V . The wedge product T ∧ S, where T, S ∈ V ∗ , is a bilinear map defined
by:
T ∧ S := T ⊗ S − S ⊗ T
meaning that for any X, Y ∈ V , we have:
(T ∧ S)(X, Y ) = (T ⊗ S)(X, Y ) − (S ⊗ T )(X, Y )
= T (X) S(Y ) − S(X) T (Y )
It is easy to note that T ∧ S = −S ∧ T .
Take the cross product in R3 as an example. Write the cross product as a bilinear
map ω(a, b) := a × b. It is a (2, 1)-tensor on R3 which can be represented as:
ω = e1 ⊗ e2 ⊗ e3 − e2 ⊗ e1 ⊗ e3
+ e2 ⊗ e3 ⊗ e1 − e3 ⊗ e2 ⊗ e1
+ e3 ⊗ e1 ⊗ e2 − e1 ⊗ e3 ⊗ e2
Now using the wedge product notations, we can express ω as:
ω = (e1 ∧ e2 ) ⊗ e3 + (e2 ∧ e3 ) ⊗ e1 + (e3 ∧ e1 ) ⊗ e2
which is a half shorter than using tensor products alone.
3.4. Wedge Products 85

Now given three elements T1 , T2 , T3 ∈ V ∗ , one can also form a triple wedge product
T1 ∧ T2 ∧ T3 which is a (3, 0)-tensor so that switching any pair of Ti and Tj (with i 6= j)
will give a negative sign. For instance:
T1 ∧ T2 ∧ T3 = −T2 ∧ T1 ∧ T3 and T1 ∧ T2 ∧ T3 = −T3 ∧ T2 ∧ T1 .
It can be defined in a precise way as:
T1 ∧ T2 ∧ T3 := T1 ⊗ T2 ⊗ T3 − T1 ⊗ T3 ⊗ T2
+ T2 ⊗ T3 ⊗ T1 − T2 ⊗ T1 ⊗ T3
+ T3 ⊗ T1 ⊗ T2 − T3 ⊗ T2 ⊗ T1

Exercise 3.29. Verify that the above definition of triple wedge product will result in
T1 ∧ T2 ∧ T3 = −T3 ∧ T2 ∧ T1 .

Exercise 3.30. Propose the definition of T1 ∧ T2 ∧ T3 ∧ T4 . Do this exercise before

reading ahead.

We can also define T1 ∧ T2 ∧ T3 in a more systematic (yet equivalent) way using

symmetric groups. Let S3 be the permutation group of {1, 2, 3}. An element σ ∈ S3 is a
bijective map σ : {1, 2, 3} → {1, 2, 3}. For instance, a map satisfying σ(1) = 2, σ(2) = 3
and σ(3) = 1 is an example of an element in S3 . We can express this σ by:
 
1 2 3
or simply: (123)
2 3 1
A map τ ∈ S3 given by τ (1) = 2, τ (2) = 1 and τ (3) = 3 can be expressed as:
 
1 2 3
or simply: (12)
2 1 3
This element, which switches two of the elements in {1, 2, 3} and fixes the other one, is
called a transposition.
Multiplication of two elements σ1 , σ2 ∈ S3 is defined by composition. Precisely, σ1 σ2
is the composition σ1 ◦ σ2 . Note that this means the elements {1, 2, 3} are input into σ2
first, and then into σ1 . In general, σ1 σ2 6= σ2 σ1 . One can check easily that, for instance,
we have:
(12)(23) = (123)
(23)(12) = (132)

Elements in the permutation group Sn of n elements (usually denoted by {1, 2, . . . , n})

can be represented and mutliplied in a similar way.
Exercise 3.31. Convince yourself that in S5 , we have:
(12345)(31) = (32)(145) = (32)(15)(14)

The above exercise shows that we can decompose (12345)(31) into a product of three
transpositions (32), (15) and (14). In fact, any element in Sn can be decomposed this
way. Here we state a standard theorem in elementary group theory:

Theorem 3.29. Every element σ ∈ Sn can be expressed as a product of transpositions:

σ = τ1 τ2 . . . τr . Such a decomposition is not unique. However, if σ = τe1 τe2 . . . τek is another
decomposition of σ into transpositions, then we have (−1)k = (−1)r .

Proof. Consult any standard textbook on Abstract Algebra. 

86 3. Tensors and Differential Forms

In view of Theorem 3.29, given an element σ ∈ Sn which can be decomposed into

the product of r transpositions, we define:
sgn(σ) := (−1)r .
For instance, sgn(12345) = (−1)3 = −1, and sgn(123) = (−1)2 = 1. Certainly, if τ is a
transposition, we have sgn(στ ) = −sgn(σ).
Now we are ready to state an equivalent way to define triple wedge product using
the above notations:
X
T1 ∧ T2 ∧ T3 := sgn(σ)Tσ(1) ⊗ Tσ(2) ⊗ Tσ(3) .
σ∈S3

We can verify that it gives the same expression as before:

X
sgn(σ)Tσ(1) ⊗ Tσ(2) ⊗ Tσ(3)
σ∈S3
= T1 ⊗ T2 ⊗ T3 σ = id
− T2 ⊗ T1 ⊗ T3 σ = (12)
− T3 ⊗ T2 ⊗ T1 σ = (13)
− T1 ⊗ T3 ⊗ T2 σ = (23)
+ T2 ⊗ T3 ⊗ T1 σ = (123) = (13)(12)
+ T3 ⊗ T1 ⊗ T2 σ = (132) = (12)(13)
In general, we define:

Definition 3.30 (Wedge Product). Let V be a finite dimensional vector space, and V ∗
be the dual space of V . Then, given any T1 , . . . , Tk ∈ V ∗ , we define their k-th wedge
product by: X
T1 ∧ · · · ∧ Tk := sgn(σ) Tσ(1) ⊗ . . . ⊗ Tσ(k)
σ∈Sk
where Sk is the permutation group of {1, . . . , k}. The vector space spanned by T1 ∧
· · · ∧ Tk ’s (where T1 , . . . , Tk ∈ V ∗ ) is denoted by ∧k V ∗ .

Remark 3.31. It is a convention to define ∧0 V ∗ := R. 

If we switch any pair of the Ti ’s, then the wedge product differs by a minus sign.
To show this, let τ ∈ Sk be a transposition, then for any σ ∈ Sk , we have sgn(σ ◦ τ ) =
−sgn(σ). Therefore, we get:
X
Tτ (1) ∧ · · · ∧ Tτ (k) = sgn(σ)Tσ(τ (1)) ⊗ . . . ⊗ Tσ(τ (k))
σ∈Sk
X
=− sgn(σ ◦ τ )Tσ◦τ (1) ⊗ . . . ⊗ Tσ◦τ (k)
σ∈Sk
X
=− sgn(σ 0 )Tσ0 (1) ⊗ . . . ⊗ Tσ0 τ (k) (where σ 0 := σ ◦ τ )
σ∈Sk
= −T1 ∧ · · · ∧ Tk .
The last step follows from the fact that σ 7→ σ ◦ τ is a bijection between Sk and itself.

Exercise 3.32. Write down T1 ∧ T2 ∧ T3 ∧ T4 explicitly in terms of tensor products

(with no wedge and summation sign).
3.4. Wedge Products 87

Exercise 3.33. Show that dim ∧k V ∗ = Ckn , when n = dim V and 0 ≤ k ≤ n, by

writing a basis for ∧k V ∗ . Show also that ∧k V ∗ = {0} if k > dim V .

Exercise 3.34. Let {ei }ni=1 be a basis for a vector space V , and {ei }ni=1 be the
corresponding dual basis for V ∗ . Show that:
ei1 ∧ · · · ∧ eik (ej1 , . . . , ejk ) = δi1 j1 · · · δik jk .

Remark 3.32. The vector space ∧k V ∗ is spanned by T1 ∧ · · · ∧ Tk ’s where T1 , . . . , Tk ∈ V ∗ .

Note that not all elements in V ∗ can be expressed in the form of T1 ∧ · · · ∧ Tk . For instance
when V = R4 with standard basis {ei }4i=1 , the element σ = e1 ∧e2 +e3 ∧e4 ∈ ∧2 V ∗ cannot
be written in the form of T1 ∧ T2 where T1 , T2 ∈ V ∗ . It is because (T1 ∧ T2 ) ∧ (T1 ∧ T2 ) = 0
for any T1 , T2 ∈ V ∗ , while σ ∧ σ = 2e1 ∧ e2 ∧ e3 ∧ e4 6= 0. 

In the above remark, we take the wedge product between elements in ∧2 V ∗ . It is

defined in a natural way that for any T1 , . . . Tk , S1 , . . . , Sr ∈ V ∗ , we have:
(T1 ∧ · · · ∧ Tk ) ∧ (S1 ∧ · · · ∧ Sr ) = T1 ∧ · · · ∧ Tk ∧ S1 ∧ · · · ∧ Sr
| {z } | {z } | {z }
∈∧k V ∗ ∈∧r V ∗ ∈∧k+r V ∗

and extended linearly to other elements in ∧k V ∗ and ∧r V ∗ . For instance, we have:

(T1 ∧ T2 + S1 ∧ S2 ) ∧ |{z}
σ = T1 ∧ T2 ∧ σ + S1 ∧ S2 ∧ σ .
| {z } | {z }
∈∧2 V ∗ ∈∧k V ∗ ∈∧k+2 V ∗

While it is true that T1 ∧ T2 = −T2 ∧ T1 for any T1 , T2 ∈ V ∗ , it is generally not true

that σ ∧ η = −η ∧ σ where σ ∈ ∧k V ∗ and η ∈ ∧r V ∗ . For instance, let T1 , . . . , T5 ∈ V ∗
and consider σ = T1 ∧ T2 and η = T3 ∧ T4 ∧ T5 . Then we can see that:
σ ∧ η = T1 ∧ T2 ∧ T3 ∧ T4 ∧ T5
= −T1 ∧ T3 ∧ T4 ∧ T5 ∧ T2 (switching T2 subsequently with T3 , T4 , T5 )
= T3 ∧ T4 ∧ T5 ∧ T1 ∧ T2 (switching T1 subsequently with T3 , T4 , T5 )
= η ∧ σ.

Proposition 3.33. Let V be a finite dimensional vector space, and V ∗ be the dual space
of V . Given any σ ∈ ∧k V ∗ and η ∈ ∧r V ∗ , we have:
(3.12) σ ∧ η = (−1)kr η ∧ σ.
Clearly from (3.12), any ω ∈ ∧even V ∗ commutes with any σ ∈ ∧k V ∗ .

Proof. By linearity, it suffices to prove that case σ = T1 ∧ · · · ∧ Tk and η = S1 ∧ · · · ∧ Sr

where Ti , Sj ∈ V ∗ , in which we have:
σ ∧ η = T1 ∧ · · · ∧ Tk ∧ S1 ∧ · · · ∧ Sr
In order to switch all the Ti ’s with the Sj ’s, we can first switch Tk subsequently with each
of S1 , . . . , Sr and each switching contributes to a factor of (−1). Precisely, we have:
T1 ∧ · · · ∧ Tk ∧ S1 ∧ · · · ∧ Sr = (−1)r T1 ∧ · · · ∧ Tk−1 ∧ S1 ∧ · · · ∧ Sr ∧ Tk .
By repeating this sequence of switching on each of Tk−1 , Tk−2 , etc., we get a factor of
(−1)r for each set of switching, and so we finally get the following as desired:
k
T1 ∧ · · · ∧ Tk ∧ S1 ∧ · · · ∧ Sr = [(−1)r ] S1 ∧ · · · ∧ Sr ∧ T1 ∧ · · · ∧ Tk

88 3. Tensors and Differential Forms

From Exercise 3.33, we know that dim ∧n V ∗ = 1 if n = dim V . In fact, every

element σ ∈ dim ∧n V ∗ is a constant multiple of e1 ∧ · · · ∧ en , and it is interesting (and
important) to note that this constant multiple is related to a determinant! Precisely, for
each i = 1, . . . , n, we consider the elements:
Xn
ωi = aij ej ∈ V ∗
j=1

where aij are real constants. Then, the wedge product of all ωi ’s are given by:
     
n
X n
X n
X
ω1 ∧ · · · ∧ ωn =  a1j1 ej1  ∧  a2j2 ej2  ∧ · · · ∧  anjn ejn 
j1 =1 j2 =1 jn =1
X
j1 jn
= a1j1 a2j2 . . . anjn e ∧ · · · ∧ e
j1 ,...,jn distinct
X
= a1σ(1) a2σ(2) . . . anσ(n) eσ(1) ∧ · · · ∧ eσ(n)
σ∈Sn

Next we want to find a relation between eσ(1) ∧ · · · ∧ eσ(n) and e1 ∧ · · · ∧ en . σ ∈ Sn , by

decomposing it into transpositions σ = τ1 ◦ · · · ◦ τk , then we have:
eσ(1) ∧ · · · ∧ eσ(n) = eτ1 ◦···◦τk (1) ∧ · · · ∧ eτ1 ◦···◦τk (n)
= (−1)eτ2 ◦···◦τk (1) ∧ · · · ∧ eτ2 ◦···◦τk (n)
= (−1)2 eτ3 ◦···◦τk (1) ∧ · · · ∧ eτ3 ◦···◦τk (n)
= ...
= (−1)k−1 eτk (1) ∧ · · · ∧ eτk (n)
= (−1)k e1 ∧ · · · ∧ en
= sgn(σ)e1 ∧ · · · ∧ en .
Therefore, we have:
!
X
ω1 ∧ · · · ∧ ωn = sgn(σ)a1σ(1) a2σ(2) . . . anσ(n) e1 ∧ · · · ∧ en .
σ∈Sn

Note that the sum: X

sgn(σ)a1σ(1) a2σ(2) . . . anσ(n)
σ∈Sn
is exactly the determinant of the matrix A whose (i, j)-th entry is aij . To summarize, let’s
state it as a proposition:

Proposition 3.34. Let V ∗ be the dual space of a vector space V of dimension n, and let
{ei }ni=1 be a basis for V , and {ei }ni=1 be the corresponding dual basis for V ∗ . Given any n
Xn
elements ωi = aij ej ∈ V ∗ , we have:
j=1

ω1 ∧ · · · ∧ ωn = (det A) e1 ∧ · · · ∧ en ,
where A is the n × n matrix whose (i, j)-th entry is aij .

Exercise 3.35. Given an n-dimensional vector space V . Show that ω1 , . . . , ωn ∈ V ∗

are linearly independent if and only if ω1 ∧ · · · ∧ ωn 6= 0.
3.4. Wedge Products 89

Exercise 3.36. Generalize Proposition 3.34. Precisely, now given

Xn
ωi = aij ej ∈ V ∗
j=1

where 1 ≤ i ≤ k < dim V , express ω1 ∧ · · · ∧ ωk in terms of ei ’s.

Exercise 3.37. Regard det : Rn × Rn → R as a multilinear map:

| |
det(v1 , . . . , vn ) := v1 ··· vn .
| |
Denote {ei } the standard basis for Rn . Show that:
det = e1 ∧ · · · ∧ en .

3.4.2. Differential Forms on Smooth Manifolds. In the simplest term, differential

forms on a smooth manifold are wedge products of cotangent vectors in T ∗ M . At each
point p ∈ M , let (u1 ,n. . . , un ) be the local
o coordinates near p, then the cotangent space
Tp∗ M is spanned by du1 p , . . . , dun |p , and a smooth differential 1-form α is a map
from M to T ∗ M such that it can be locally expressed as:
n
!
X
i
α(p) = p, αi (p) du p
i=1

where αi are smooth functions locally defined near p. Since the based point p can usually
be understood from the context, we usually denote α by simply:
n
X
α= αi dui .
i=1

Since Tp∗ M is a finite dimensional vector space, we can consider the wedge products
of its elements. A differential k-form ω on a smooth manifold M is a map which assigns
each point p ∈ M to an element in ∧k Tp∗ M . Precisely:

Definition 3.35 (Smooth Differential k-Forms). Let M be a smooth manifold. A smooth

differential k-form ω on M is a map ωp : Tp M × . . . × Tp M → R at each p ∈ M such
| {z }
k times
that under any local parametrization F (u1 , . . . , un ) : U → M , it can be written in the
form:
Xn
ω= ωi1 i2 ···ik dui1 ∧ · · · ∧ duik
i1 ,...,ik =1
where ωi1 i2 ...ik ’s are smooth scalar functions locally defined in F (U), and they are
commonly called the local components of ω. The vector space of all smooth differential
k-forms on M is denoted by ∧k T ∗ M .

Remark 3.36. It is a convention to denote ∧0 T ∗ M := C ∞ (M, R), the vector space of all
smooth scalar functions defined on M . 

We will mostly deal with differential k-forms that are smooth. Therefore, we will
very often call a smooth differential k-form simply by a differential k-form, or even simpler,
a k-form. As we will see in the next section, the language of differential forms will unify
and generalize the curl, grad and div in Multivariable Calculus and Physics courses.
90 3. Tensors and Differential Forms

From algebraic viewpoint, the manipulations of differential k-forms on a manifold

are similar to those for wedge products of a finite-dimensional vector space. The major
difference is a manifold is usually covered by more than one local parametrizations,
hence there are conversion rules for differential k-forms from one local coordinate system
to another.
Example 3.37. Consider R2 with (x, y) and (r, θ) as its two local coordinates. Given a
2-form ω = dx ∧ dy, for instance, we can express it in terms of the polar coordinates
(r, θ):
∂x ∂x
dx = dr + dθ
∂r ∂θ
= (cos θ) dr − (r sin θ) dθ
∂y ∂y
dy = dr + dθ
∂r ∂θ
= (sin θ) dr + (r cos θ) dθ
Therefore, using dr ∧ dr = 0 and dθ ∧ dθ = 0, we get:
dx ∧ dy = (r cos2 θ)dr ∧ dθ − (r sin2 )dθ ∧ dr
= (r cos2 θ + r sin2 θ) dr ∧ dθ
= r dr ∧ dθ.


Exercise 3.38. Define a 2-form on R3 by:

ω = x dy ∧ dz + y dz ∧ dx + z dx ∧ dy.
Express ω in terms of spherical coordinates (ρ, θ, ϕ), defined by:
(x, y, z) = (ρ sin ϕ cos θ, ρ sin ϕ sin θ, ρ cos ϕ).

Exercise 3.39. Let ω be the 2-form on R2n given by:

ω = dx1 ∧ dx2 + dx3 ∧ dx4 + . . . + dx2n−1 ∧ dx2n .

| ∧ ·{z
Compute ω · · ∧ ω}.
n times

Exercise 3.40. Let (u1 , . . . , un ) and (v1 , . . . , vn ) be two local coordinates of a smooth
manifold M . Show that:
∂(u1 , . . . , un ) 1
du1 ∧ · · · ∧ dun = det dv ∧ · · · ∧ dv n .
∂(v1 , . . . , vn )

Exercise 3.41. Show that on R3 , a (2, 0)-tensor T is in ∧2 (R2 )∗ if and only if

T (v, v) = 0 for any v ∈ R3 .
3.5. Exterior Derivatives 91

3.5. Exterior Derivatives

Exterior differentiation is an important operations on differential forms. It not only
generalizes and unifies the curl, grad, div operators in Multivariable Calculus and Physics,
but also leads to the development of de Rham cohomology to be discussed in Chapter 5.

3.5.1. Definition of Exterior Derivatives. Exterior differentiation, commonly de-

noted by the symbol d, takes a k-form to a (k + 1)-form. To begin, let’s define it on scalar
functions first. Suppose (u1 , . . . , un ) are local coordinates of M n , then given any smooth
scalar function f ∈ C ∞ (M, R), we define:
n
X ∂f
(3.13) df := dui
i=1
∂u i

Although (3.13) involves local coordinates, it can be easily shown that df is independent
of local coordinates. Suppose (v1 , . . . , vn ) is another local coordinates of M which overlap
with (u1 , . . . , un ). By the chain rule, we have:
n
∂f X ∂f ∂vk
=
∂ui ∂vk ∂ui
k=1
n
X ∂vk i
dv k = du
i=1
∂ui
which combine to give:
n n Xn n
X ∂f X ∂f ∂vk i X ∂f
dui = du = dv k .
i=1
∂ui i=1
∂vk ∂ui ∂vk
k=1 k=1
Therefore, if f is smooth on M then df is a smooth 1-form on M . The components of df
∂f
are ∂u i
’s, and so df is analogous to ∇f in Multivariable Calculus. Note that as long as
f is C ∞ just in an open set U ⊂ M , we can also define df locally on U since (3.13) is a
local expression.
Exterior derivatives can also be defined on differential forms of higher degrees. Let
α ∈ ∧1 T ∗ M , which can be locally written as:
Xn
α= αi dui
i=1
where αi ’s are smooth functions locally defined in a local coordinate chart. Then, we
define:
n n X n
X X ∂αi j
(3.14) dα := dαi ∧ dui = du ∧ dui .
i=1 i=1 j=1
∂u j

Using the fact that duj ∧ dui = −dui ∧ duj and dui ∧ dui = 0, we can also express dα as:
X  ∂αi ∂αj

dα = − duj ∧ dui .
∂uj ∂ui
1≤j<i≤n
92 3. Tensors and Differential Forms

Example 3.38. Take M = R3 as an example, and let (x, y, z) be the (usual) coordinates
of R3 , then given any 1-form α = P dx + Q dy + R dz (which is analogous to the vector
field P î + Qĵ + Rk̂), we have:
dα = dP ∧ dx + dQ ∧ dy + dR ∧ dz
   
∂P ∂P ∂P ∂Q ∂Q ∂Q
= dx + dy + dz ∧ dx + dx + dy + dz ∧ dy
∂x ∂y ∂z ∂x ∂y ∂z
 
∂R ∂R ∂R
+ dx + dy + dz ∧ dz
∂x ∂y ∂z
∂P ∂P ∂Q ∂Q
= dy ∧ dx + dz ∧ dx + dx ∧ dy + dz ∧ dy
∂y ∂z ∂x ∂z
∂R ∂R
+ dx ∧ dz + dy ∧ dz
∂x ∂y
     
∂Q ∂P ∂R ∂P ∂R ∂Q
= − dx ∧ dy − − dz ∧ dx + − dy ∧ dz
∂x ∂y ∂x ∂z ∂y ∂z
 
which is analogous to ∇ × P î + Qĵ + Rk̂ by declaring the correspondence {î, ĵ, k̂}
with {dy ∧ dz, dz ∧ dx, dx ∧ dy}. 

One can check that the definition of dα stated in (3.14) is independent of local
coordinates. On general k-forms, the exterior derivatives are defined in a similar way as:

Definition 3.39 (Exterior Derivatives). Let M n be a smooth manifold and (u1 , . . . , un )

be local coordinates on M . Given any (smooth) k-form
Xn
ω= ωj1 ···jk duj1 ∧ · · · ∧ dujk ,
j1 ,...,jk =1

we define:
n
X
(3.15) dω := dωj1 ···jk ∧ duj1 ∧ · · · ∧ dujk
j1 ,··· ,jk =1
n n
X X ∂ωj1 ···jk i
= du ∧ duj1 ∧ · · · ∧ dujk
j1 ,··· ,jk =1 i=1
∂ui
In particular, if ω is an n-form (where n = dim V ), we have dω = 0.

Exercise 3.42. Show that dω defined as in (3.15) does not depend on the choice of
local coordinates.

Example 3.40. Consider R2 equipped with polar coordinates (r, θ). Consider the 1-form:
ω = (r sin θ) dr.
Then, we have
∂(r sin θ) ∂(r sin θ)
dω = dr ∧ dr + dθ ∧ dr
∂r ∂θ
= 0 + (r cos θ) dθ ∧ dr
= −(r cos θ) dr ∧ dθ.

3.5. Exterior Derivatives 93

Exercise 3.43. Let ω = F1 dy ∧ dz + F2 dz ∧ dx + F3 dx ∧ dy be a smooth 2-form on

R3 . Compute dω. What operator in Multivariable Calculus is the d analogous to in
this case?

Exercise 3.44. Let ω, η, θ be the following differential forms on R3 :

ω = x dx − y, dy
η = z dx ∧ dy + x dy ∧ dz
θ = z dy
Compute: ω ∧ η, ω ∧ η ∧ θ, dω, dη and dθ.

3.5.2. Properties of Exterior Derivatives. The exterior differentiation d can hence

be regarded as a chain of maps:
d d d d d
∧0 T ∗ M −→ ∧1 T ∗ M −→ ∧2 T ∗ M −→ · · · −→ ∧n−1 T ∗ M −→ ∧n T ∗ M.
Here we abuse the use of the symbol d a little bit – we use the same symbol d for all
d
the maps ∧k T ∗ M −→ ∧k+1 T ∗ M in the chain. The following properties about exterior
differentiation are not difficult to prove:

Proposition 3.41. For any k-forms ω and η, and any smooth scalar function f , we have
the following:
(1) d(ω + η) = dω + dη
(2) d(f ω) = df ∧ ω + f dω

Proof. (1) is easy to prove (left as an exercise for readers). To prove (2), we consider
n
X
local coordinates (u1 , . . . , un ) and let ω = ωj1 ···jk duj1 ∧ · · · ∧ dujk . Then, we
j1 ,...,jk =1
have:
n n
X X ∂
d(f ω) = (f ωj1 ···jk ) dui ∧ duj1 ∧ · · · ∧ dujk
j1 ,...,jk =1 i=1
∂u i
n n  
X X ∂f ∂ωj1 ···jk
= ωj1 ···jk + f dui ∧ duj1 ∧ · · · ∧ dujk
j1 ,...,jk =1 i=1
∂u i ∂u i

n
!  n

X ∂f X
= dui ∧  ωj1 ···jk duj1 ∧ · · · ∧ dujk 
i=1
∂u i j ,...,j =1
1 k

n n
X X ∂ωj1 ···jk i
+f du ∧ duj1 ∧ · · · ∧ dujk
j1 ,··· ,jk =1 i=1
∂ui

as desired. 

Identity (2) in Proposition 3.41 can be regarded as a kind of product rule. Given a
k-form α and a r-form β, the general product rule for exterior derivative is stated as:

Proposition 3.42. Let α ∈ ∧k T ∗ M and β ∈ ∧r T ∗ M be smooth differential forms on M ,

then we have:
d(α ∧ β) = dα ∧ β + (−1)k α ∧ dβ.
94 3. Tensors and Differential Forms

Exercise 3.45. Prove Proposition 3.42. Based on your proof, explain briefly why
the product rule does not involve any factor of (−1)r .

Exercise 3.46. Given three differential forms α, β and γ such that dα = 0, dβ = 0

and dγ = 0. Show that:
d(α ∧ β ∧ γ) = 0.

An crucial property of exterior derivatives is that the composition is zero. For instance,
given a smooth scalar function f (x, y, z) defined on R3 , we have:
∂f ∂f ∂f
df = dx + dy + dz.
∂x ∂y ∂z
Taking exterior derivative one more time, we get:
 
∂ ∂f ∂ ∂f ∂ ∂f
d(df ) = dx + dy + dz ∧ dx
∂x ∂x ∂y ∂x ∂z ∂x
 
∂ ∂f ∂ ∂f ∂ ∂f
+ dx + dy + dz ∧ dy
∂x ∂y ∂y ∂y ∂z ∂y
 
∂ ∂f ∂ ∂f ∂ ∂f
+ dx + dy + dz ∧ dz
∂x ∂z ∂y ∂z ∂z ∂z
   
∂ ∂f ∂ ∂f ∂ ∂f ∂ ∂f
= − dx ∧ dy + − dz ∧ dx
∂x ∂y ∂y ∂x ∂z ∂x ∂x ∂z
 
∂ ∂f ∂ ∂f
+ − dy ∧ dz
∂y ∂z ∂z ∂y
Since partial derivatives commute, we get d(df ) = 0, or in short d2 f = 0, for any scalar
function f . The fact that d2 = 0 is generally true on smooth differential forms, not only
for scalar functions. Precisely, we have:

Proposition 3.43. Let ω be a smooth k-form defined on a smooth manifold M , then we

have:
d2 ω := d(dω) = 0.

n
X
Proof. Let ω = ωj1 ···jk duj1 ∧ · · · ∧ dujk , then:
j1 ,...,jk =1
n n
X X ∂ωj 1 ···jk
dω = dui ∧ duj1 ∧ · · · ∧ dujk .
j1 ,··· ,jk =1 i=1
∂ui
 
n n
X X ∂ωj 1 ···jk
d2 ω = d  dui ∧ duj1 ∧ · · · ∧ dujk 
j1 ,··· ,jk =1 i=1
∂ui
n n X
n
X X ∂ 2 ωj 1 ...jk
= dul ∧ dui ∧ duj1 ∧ · · · ∧ dujk
j1 ,··· ,jk =1 i=1 l=1
∂ul ∂ui
n
X ∂ 2 ωj1 ...jk l
For each fixed k-tuple (j1 , . . . , jk ), the term du ∧ dui can be rewritten as:
∂ul ∂ui
i,l=1

X  ∂ 2 ωj ...j ∂ 2 ωj1 ...jk


1 k
− dul ∧ dui
∂ul ∂ui ∂ui ∂ul
1≤i<l≤n
3.5. Exterior Derivatives 95

which is zero since partial derivatives commute. It concludes that d2 ω = 0. 

Proposition 3.43 is a important fact that leads to the development of de Rham

cohomology in Chapter 5.
In Multivariable Calculus, we learned that given a vector field F = P î + Qĵ + Rk̂
and a scalar function f , we have:
∇ × ∇f = 0
∇ · (∇ × F ) = 0
These two formulae can be unified using the language of differential forms. The one-form
df corresponds to the vector field ∇f :
∂f ∂f ∂f
df = dx + dy + dz
∂x ∂y ∂z
∂f ∂f ∂f
∇f = î + ĵ + k̂
∂x ∂y ∂z
Define a one-form ω = P dx + Q dy + R dz on R3 , which corresponds to the vector field
F , then we have discussed that dω corresponds to taking curl of F :
     
∂Q ∂P ∂R ∂P ∂R ∂Q
dω = − dx ∧ dy − − dz ∧ dx + − dy ∧ dz
∂x ∂y ∂x ∂z ∂y ∂z
     
∂Q ∂P ∂R ∂P ∂R ∂Q
∇×F = − k̂ − − ĵ + − î
∂x ∂y ∂x ∂z ∂y ∂z
If one takes ω = df , and F = ∇f , then we have dω = d(df ) = 0, which corresponds to
the fact that ∇ × G = ∇ × ∇f = 0 in Multivariable Calculus.
Taking exterior derivative on a two-form β = A dy ∧ dz + B dz ∧ dx + C dx ∧ dy
corresponds to taking the divergence on the vector field G = Aî + B ĵ + C k̂ according to
Exercise 3.43:
 
∂A ∂B ∂C
dβ = + + dx ∧ dy ∧ dz
∂x ∂y ∂z
 
∂A ∂B ∂C
∇·G= + +
∂x ∂y ∂z
By taking β = dω, and G = ∇ × F , then we have dβ = d(dω) = 0 corresponding to
∇ · G = ∇ · (∇ × F ) = 0 in Multivariable Calculus.
Here is a summary of the correspondences:
Differential Form on R3 Multivariable Calculus
f (x, y, z) f (x, y, z)
ω = P dx + Q dy + R dz F = P î + Qĵ + Rk̂
β = A dy ∧ dz + B dz ∧ dx + C dx ∧ dy G = Aî + B ĵ + C k̂
df ∇f
dω ∇×F
dβ ∇·G
d2 f = 0 ∇ × ∇f = 0
d2 ω = 0 ∇ · (∇ × F ) = 0

3.5.3. Exact and Closed Forms. In Multivariable Calculus, we discussed vari-

ous concepts of vector fields including potential functions, conservative vector fields,
solenoidal vector fields, curl-less and divergence-less vector fields, etc. All these concepts
can be unified using the language of differential forms.
96 3. Tensors and Differential Forms

As a reminder, a conservative vector field F is one that can be expressed as F = ∇f

where f is a scalar function. It is equivalent to saying that the 1-form ω can be expressed
as ω = df . Moreover, a solenoidal vector field G is one that can be expressed as G = ∇×F
for some vector field F . It is equivalent to saying that the 2-form β can be expressed as
β = dω for some 1-form ω.
Likewise, a curl-less vector field F (i.e. ∇ × F = 0) corresponds to a 1-form ω
satisfying dω = 0; and a divergence-less vector field G (i.e. ∇ · G = 0) corresponds to a
2-form β satisfying dβ = 0.
In view of the above correspondence, we introduce two terminologies for differential
forms, namely exact-ness and closed-ness:

Definition 3.44 (Exact and Closed Forms). Let ω be a smooth k-form defined on a
smooth manifold M , then we say:
• ω is exact if there exists a (k − 1)-form η defined on M such that ω = dη;
• ω is closed if dω = 0.

Remark 3.45. By the fact that d2 = 0 (Proposition 3.43), it is clear that every exact form
is a closed form (but not vice versa). 
The list below showcases the corresponding concepts of exact/closed forms in Multi-
variable Calculus.
Differential Form on R3 Multivariable Calculus
exact 1-form conservative vector field
closed 1-form curl-less vector field
exact 2-form solenoidal vector field
closed 2-form divergence-less vector field
Example 3.46. On R3 , the 1-form:
α = yz dx + zx dy + xy dz
is exact since α = df where f (x, y, z) = xyz. By Proposition 3.43, we immediately get
dα = d(df ) = 0, so α is a closed form. One can also verify this directly:
dα = (z dy + y dz) ∧ dx + (z dx + x dz) ∧ dy + (y dx + x dy) ∧ dz
= (z − z) dx ∧ dy + (y − y) dz ∧ dx + (x − x) dy ∧ dz = 0.

Example 3.47. The 1-form:
y x
α := − dx + 2 dy
x2 + y 2 x + y2
defined on R2 \{(0, 0)} is closed:
   
∂ y ∂ x
dα = − 2 dy ∧ dx + dx ∧ dy
∂y x + y2 ∂x x2 + y 2
y 2 − x2 y 2 − x2
= 2 2 2
dy ∧ dx + 2 dx ∧ dy
(x + y ) (x + y 2 )2
=0
as dx ∧ dy = −dy ∧ dx. However, we will later see that α is not exact.
y
Note that even though we have α = df where f (x, y) = tan−1 , such an f is NOT
x
smooth on R2 \{(0, 0)}. In order to claim α is exact, we require such an f to be smooth
on the domain of α. 
3.5. Exterior Derivatives 97

Exercise 3.47. Consider the forms ω, η and θ on R3 defined in Exercise 3.44.

Determine whether each of them is closed and/or exact on R3 .

Exercise 3.48. The purpose of this exercise is to show that any closed 1-form ω on
R3 must be exact. Let
ω = P (x, y, z) dx + Q(x, y, z) dy + R(x, y, z) dz
be a closed 1-form on R3 . Define f : R3 → R by:
Z t=1
f (x, y, z) = (xP (tx, ty, tz) + yQ(tx, ty, tz) + zR(tx, ty, tz)) dt
t=0
Show that ω = df . Point out exactly where you have used the fact that dω = 0.

3.5.4. Pull-Back of Tensors. Let’s first begin by reviewing the push-forward and
pull-back of tangent and cotangent vectors. Given a smooth map Φ : M → N between
two smooth manifolds M m and N n , its tangent map Φ∗ takes a tangent vector in
Tp M to a tangent vector in TΦ(p) N . If we let F (u1 , . . . , um ) be local coordinates of M ,
G(v1 , . . . , vn ) be local coordinates of N and express the map Φ locally as:
(v1 , . . . , vn ) = G−1 ◦ Φ ◦ F (u1 , . . . , um ),
 
∂
then Φ∗ acts on the basis vectors by:
∂ui
 
∂ ∂Φ X ∂vj ∂
Φ∗ = = .
∂ui ∂ui j
∂ui ∂vj

The tangent map Φ∗ is also commonly called the push-forward map. It is important to
∂v
note that the vj ’s in the partial derivatves ∂uji can sometimes cause confusion if we talk
about the push-forwards of two different smooth maps Φ : M → N and Ψ : M → N .
Even with the same input (u1 , . . . , um ), the output Φ(u1 , . . . , um ) and Ψ(u1 , . . . , um ) are
generally different and have different vj -coordinates. To avoid this confusion, it is best
to write:
  X
∂ ∂(vj ◦ Φ) ∂
Φ∗ =
∂ui j
∂ui ∂vj
  X
∂ ∂(vj ◦ Ψ) ∂
Ψ∗ =
∂ui j
∂ui ∂vj
∂v
Here each vj in the partial derivatives ∂uji are considered to be a locally defined function
taking a point p ∈ N to its vj -coordinate.
For cotangent vectors (i.e. 1-forms), we talk about pull-back instead. According to
Definition 3.14, Φ∗ takes a cotangent vector in TΦ(p)
∗
N to a cotangent vector in Tp∗ M ,
defined as follows:
Φ∗ (dv i )(X) = dv i (Φ∗ X) for any X ∈ Tp M .
In terms of local coordinates, it is given by:
X ∂(vi ◦ Φ)
Φ∗ (dv i ) = duj .
j
∂u j

The pull-back action by a smooth Φ : M → N between manifolds can be extended to

(k, 0)-tensors (and hence to differential forms):
98 3. Tensors and Differential Forms

Definition 3.48 (Pull-Back on (k, 0)-Tensors). Let Φ : M → N be a smooth map

between two smooth manifolds. Given T a smooth (k, 0)-tensor on N , then we define:
(Φ∗ T )p (X1 , . . . , Xk ) = TΦ(p) (Φ∗ (X1 ), . . . , Φ∗ (Xk )) for any X1 , . . . , Xk ∈ Tp M

Remark 3.49. An equivalent way to state the definition is as follows: let T1 , . . . Tk ∈ T N

be 1-forms on N , then we define:
Φ∗ (T1 ⊗ · · · ⊗ Tk ) = (Φ∗ T1 ) ⊗ · · · ⊗ (Φ∗ Tk ).

∗
Remark 3.50. It is easy to verify that Φ is linear, in a sense that:
Φ∗ (aT + bS) = aΦ∗ T + bΦ∗ S
for any (k, 0)-tensors T and S, and scalars a and b. 
2 2 3
Example 3.51. Let’s start with an example on R . Let Φ : R → R be a map defined by:
Φ(x1 , x2 ) = ex1 +x2 , sin(x21 x2 ), x1 .

To avoid confusion, we use (x1 , x2 ) to label the coordinates of the domain R2 , and use
(y1 , y2 , y3 ) to denote the coordinates of the codomain R3 . Then, we have:
      
∗ 1 ∂ 1 ∂ 1 ∂Φ
Φ (dy ) = dy Φ∗ = dy
∂x1 ∂x1 ∂x1
 
1 ∂(y1 ◦ Φ) ∂ ∂(y2 ◦ Φ) ∂ ∂(y3 ◦ Φ) ∂
= dy + +
∂x1 ∂y1 ∂x1 ∂y2 ∂x1 ∂y3
∂(y1 ◦ Φ) ∂ x1 +x2
= = e = ex1 +x2 .
∂x1 ∂x1
Similarly, we have:
 
∂ ∂(y1 ◦ Φ) ∂ x1 +x2
Φ∗ (dy 1 ) = = e = ex1 +x2 .
∂x2 ∂x2 ∂x2
Therefore, Φ∗ (dy 1 ) = ex1 +x2 dx1 + ex1 +x2 dx2 = ex1 +x2 (dx1 + dx2 ). We leave it as an
exercise for readers to verify that:
Φ∗ (dy 2 ) = 2x1 x2 cos(x21 x2 ) dx1 + x21 cos(x21 x2 ) dx2
Φ∗ (dy 3 ) = dx1

Let f (y1 , y2 , y3 ) be a scalar function on R3 , and consider the (2, 0)-tensor on R3 :

T = f (y1 , y2 , y3 ) dy 1 ⊗ dy 2
The pull-back of T by Φ is given by:
Φ∗ T = f (y1 , y2 , y3 ) Φ∗ (dy 1 ) ⊗ Φ∗ (dy 2 )
= f (Φ(x1 , x2 )) ex1 +x2 (dx1 + dx2 ) ⊗ 2x1 x2 cos(x21 x2 ) dx1 + x21 cos(x21 x2 ) dx2

The purpose of writing f (y1 , y2 , y3 ) as f (Φ(x1 , x2 )) is to leave the final expression in

terms of functions and tensors in (x1 , x2 )-coordinates. 
Example 3.52. Let Σ be a regular surface in R3 . The standard dot product in R3 is given
by the following (2, 0)-tensor:
ω = dx ⊗ dx + dy ⊗ dy + dz ⊗ dz.
Consider the inclusion map ι : Σ → R3 . Although the input and output are the same
under the map ι, the cotangents dx and ι∗ (dx) are different! The former is a cotangent
vector on R3 , while ι∗ (dx) is a cotangent vector on the surface Σ. If (x, y, z) = F (u, v) is
3.5. Exterior Derivatives 99

a local parametrization of Σ, then ι∗ (dx) should be in terms of du and dv, but not dx, dy
and dz. Precisely, we have:

∂F ∂ι ∂(ι ◦ F ) ∂F
ι∗ = := =
∂u ∂u ∂u ∂u
      
∂F ∂F ∂F
ι∗ (dx) = dx ι∗ = dx
∂u ∂u ∂u
 
∂x ∂ ∂y ∂ ∂z ∂
= dx + +
∂u ∂x ∂u ∂y ∂u ∂z
∂x
= .
∂u
 
∂F ∂x
Similarly, we also have ι∗ (dx) = , and hence:
∂v ∂v

∂x ∂x
ι∗ (dx) = du + dv.
∂u ∂v

As a result, we have:

ι∗ ω = ι∗ (dx) ⊗ ι∗ (dx) + ι∗ (dy) ⊗ ι∗ (dy) + ι∗ (dz) ⊗ ι∗ (dz)

   
∂x ∂x ∂x ∂x
= du + dv ⊗ du + dv
∂u ∂v ∂u ∂v
   
∂y ∂y ∂y ∂y
+ du + dv ⊗ du + dv
∂u ∂v ∂u ∂v
   
∂z ∂z ∂z ∂z
+ du + dv ⊗ du + dv .
∂u ∂v ∂u ∂v

After expansion and simplification, one will get:

∂F ∂F ∂F ∂F ∂F ∂F ∂F ∂F
ι∗ ω = · du ⊗ du + · du ⊗ dv + · dv ⊗ du + · dv ⊗ dv,
∂u ∂u ∂u ∂v ∂v ∂u ∂v ∂v

which is the first fundamental form in Differential Geometry. 

Exercise 3.49. Let the unit sphere S2 be locally parametrized by spherical coordi-
nates (θ, ϕ). Consider the (2, 0)-tensor on R3 :
ω = x dy ⊗ dz
∗
Express the pull-back ι ω in terms of (θ, ϕ).

One can derive a general formula (which you do not need to remember in practice)
for the local expression of pull-backs. Consider local coordinates {ui } for M and {vi } for
N , and write (v1 , . . . , vn ) = Φ(u1 , . . . , um ) and

n
X
T = Ti1 ···ik (v1 , . . . , vn ) dv i1 ⊗ · · · ⊗ dv ik .
i1 ,...,ik =1
100 3. Tensors and Differential Forms

The pull-back Φ∗ T then has the following local expression:

(3.16)
n
X
Φ∗ T = Ti1 ···ik (v1 , . . . , vn ) Φ∗ (dv i1 ) ⊗ · · · ⊗ Φ∗ (dv ik )
i1 ,...,ik =1
   
n m m
X X ∂vi
X ∂vi
= Ti1 ···ik (Φ(u1 , . . . , um ))  1
duj1  ⊗ · · · ⊗  k
dujk 
i1 ,...,ik =1 j =1
∂u j1 j =1
∂u j k
1 k

n m
X X ∂vi1 ∂vik
= Ti1 ···ik (Φ(u1 , . . . , um )) ··· duj1 ⊗ · · · ⊗ dujk .
i1 ,...,ik =1 j1 ,...,jk =1
∂uj1 ∂ujk

In view of Ti1 ···ik (v1 , . . . , vn ) = Ti1 ···ik (Φ(u1 , . . . , um )) and the above local expression,
we define
Φ∗ f := f ◦ Φ
for any scalar function of f . Using this notation, we then have Φ∗ (f T ) = (Φ∗ f ) Φ∗ T for
any scalar function f and (k, 0)-tensor T .
Exercise 3.50. Let Φ : M → N be a smooth map between smooth manifolds M and
N , f be a smooth scalar function defined on N . Show that
Φ∗ (df ) = d(Φ∗ f ).
In particular, if (v1 , . . . , vn ) are local coordinates of N , we have Φ∗ (dv j ) = d(Φ∗ v j ).

Example 3.53. Using the result from Exercise 3.50, one can compute the pull-back
by inclusion map ι : Σ → R3 for regular surfaces Σ in R3 . Suppose F (u, v) is a local
parametrization of Σ, then:
ι∗ (dx) = d(ι∗ x) = d(x ◦ ι).
Although x ◦ ι and x (as a coordinate function) have the same output, their domains are
different! Namely, x ◦ ι : Σ → R while x : R3 → R. Therefore, when computing d(x ◦ ι),
one should express it in terms of local coordinates (u, v) of Σ:
∂(x ◦ ι) ∂(x ◦ ι) ∂x ∂x
d(x ◦ ι) = du + dv = du + dv.
∂u ∂v ∂u ∂v


Recall that the tangent maps (i.e. push-forwards) acting on tangent vectors satisfy
the chain rule: if Φ : M → N and Ψ : N → P are smooth maps between smooth
manifolds, then we have (Ψ ◦ Φ)∗ = Ψ∗ ◦ Φ∗ . It is easy to extend the chain rule to
(k, 0)-tensors:

Theorem 3.54 (Chain Rule for (k, 0)-tensors). Let Φ : M → N and Ψ : N → P be

smooth maps between smooth manifolds M , N and P , then the pull-back maps Φ∗ and
Ψ∗ acting on (k, 0)-tensors for any k ≥ 1 satisfy the following chain rule:
(3.17) (Ψ ◦ Φ)∗ = Φ∗ ◦ Ψ∗ .

Exercise 3.51. Prove Theorem 3.54.

3.5. Exterior Derivatives 101

Exercise 3.52. Denote idM and idT M to be the identity maps of a smooth manifold
∗
M and its tangent bundle respectively. Show that (idM ) = idT M . Hence, show that
if M and N are diffeomorphic, then for k ≥ 1 the vector spaces of (k, 0)-tensors
⊗k T ∗ M and ⊗k T ∗ N are isomorphic.

3.5.5. Pull-Back of Differential Forms. By linearity of the pull-back map, and the
fact that differential forms are linear combinations of tensors, the pull-back map acts on
differential forms by the following way:
Φ∗ (T1 ∧ · · · ∧ Tk ) = Φ∗ T1 ∧ · · · ∧ Φ∗ Tk
for any 1-forms T1 , . . . , Tk .
Example 3.55. Consider the map Φ : R2 → R2 given by:
Φ(x1 , x2 ) = (x21 − x2 , x32 ).
| {z }
(y1 ,y2 )

By straight-forward computations, we have:

Φ∗ (dy 1 ) = 2x1 dx1 − dx2
Φ∗ (dy 2 ) = 3x2 dx2
Therefore, we have:
Φ∗ (dy 1 ∧ dy 2 ) = Φ∗ (dy 1 ) ∧ Φ∗ (dy 2 ) = 6x1 x2 dx1 ∧ dx2 .
Note that 6x1 x2 is the Jacobian determinant det[Φ∗ ]. We will see soon that it is not a
coincident, and it holds true in general. 

Although the computation of pull-back on differential forms is not much different

from that on tensors, there are several distinctive features for pull-back on forms. One
feature is that the pull-back on forms is closely related to Jacobian determinants:

Proposition 3.56. Let Φ : M → N be a smooth map between two smooth manifolds.

Suppose (u1 , . . . , um ) are local coordinates of M , and (v1 , . . . , vn ) are local coordinates of
N , then for any 1 ≤ i1 , . . . , ik ≤ n, we have:
X ∂(vi1 , . . . , vik )
(3.18) Φ∗ (dv i1 ∧ · · · ∧ dv ik ) = det duj1 ∧ · · · ∧ dujk .
∂(uj1 , . . . , ujk )
1≤j1 <···<jk ≤m

In particular, if dim M = dim N = n, then we have:

(3.19) Φ∗ (dv 1 ∧ · · · ∧ dv n ) = det[Φ∗ ] du1 ∧ · · · ∧ dun
where [Φ∗ ] is the Jacobian matrix of Φ with respect to local coordinates {ui } and {vi }, i.e.
∂(v1 , . . . , vn )
[Φ∗ ] = .
∂(u1 , . . . , un )

Proof. Proceed as in the derivation of (3.16) by simply replacing all tensor products by
wedge products, we get:
m  
∗ i1 ik
X ∂vi1 ∂vik j1 jk
Φ (dv ∧ · · · ∧ dv ) = ··· du ∧ · · · ∧ du
j1 ,...,jk =1
∂uj1 ∂ujk
m  
X ∂vi1 ∂vik j1 jk
= ··· du ∧ · · · ∧ du
j ,...,j =1
∂uj1 ∂ujk
1 k
j1 ,...,jk distinct
102 3. Tensors and Differential Forms

The second equality follows from the fact that duj1 ∧ · · · ∧ dujk = 0 if {j1 , . . . , jk } are not
all distinct. Each k-tuples (j1 , . . . , jk ) with distinct ji ’s can be obtained by permuting a
strictly increasing sequence of j’s. Precisely, we have:
{(j1 , . . . , jk ) : 1 ≤ j1 , . . . , jk ≤ n and j1 , . . . , jk are all distinct}
[
= {(jσ(1) , . . . , jσ(k) ) : 1 ≤ j1 < j2 < . . . < jk ≤ n}
σ∈Sk

Therefore, we get:
Φ∗ (dv i1 ∧ · · · ∧ dv ik )
!
X X ∂vi1 ∂vik
= ··· dujσ(1) ∧ · · · ∧ dujσ(k)
∂ujσ(1) ∂ujσ(k)
1≤j1 <...<jk ≤m σ∈Sk
X X ∂vi1 ∂vik
= sgn(σ) ··· duj1 ∧ · · · ∧ dujk
∂ujσ(1) ∂ujσ(k)
1≤j1 <...<jk ≤m σ∈Sk
 
X ∂vi1 ∂vik ∂vip
By observing that sgn(σ) ··· is the determinant of ,
∂ujσ(1) ∂ujσ(k) ∂ujq 1≤p,q≤k
σ∈Sk
the desired result (3.18) follows easily.
The second result (3.19) follows directly from (3.18). In case of dim M = dim N = n
and k = n, the only possible strictly increasing sequence 1 ≤ j1 < . . . < jn ≤ n is
(j1 , . . . , jn ) = (1, 2, . . . , n). 

Proposition 3.57. Let Φ : M → N be a smooth map between two smooth manifolds. For
any ω ∈ ∧k T ∗ N , we have:
(3.20) Φ∗ (dω) = d(Φ∗ ω).
To be precise, we say Φ∗ (dN ω) = dM (Φ∗ ω), where dN : ∧k T ∗ N → ∧k+1 T ∗ N and
dM : ∧k T ∗ M → ∧k+1 T ∗ M are the exterior derivatives on N and M respectively.

Proof. Let {uj } and {vi } be local coordinates of M and N respectively. By linearity, it
suffices to prove (3.20) for the case ω = f dv i1 ∧ · · · ∧ dv ik where f is a locally defined
scalar function. The proof follows from computing both LHS and RHS of (3.20):
dω = df ∧ dv i1 ∧ · · · ∧ dv ik
Φ∗ (dω) = Φ∗ (df ) ∧ Φ∗ (dv i1 ) ∧ · · · ∧ Φ∗ (dv ik )
= d(Φ∗ f ) ∧ d(Φ∗ v j1 ) ∧ · · · ∧ d(Φ∗ v jk ).
Here we have used Exercise 3.50. On the other hand, we have:
Φ∗ ω = (Φ∗ f ) Φ∗ (dv j1 ) ∧ · · · ∧ Φ∗ (dv jk )
= (Φ∗ f ) d(Φ∗ v i1 ) ∧ · · · ∧ d(Φ∗ v ik )
d(Φ∗ ω) = d(Φ∗ f ) ∧ d(Φ∗ v i1 ) ∧ · · · ∧ d(Φ∗ v ik )
+ Φ∗ f d d(Φ∗ v i1 ) ∧ · · · ∧ d(Φ∗ v ik )

Since d2 = 0, each of d(Φ∗ v iq ) is a closed 1-form. By Proposition 3.42 (product rule) and
induction, we can conclude that:
d d(Φ∗ v i1 ) ∧ · · · ∧ d(Φ∗ v ik ) = 0

and so d(Φ∗ ω) = d(Φ∗ f ) ∧ d(Φ∗ v i1 ) ∧ · · · ∧ d(Φ∗ v ik ) as desired. 

3.5. Exterior Derivatives 103

Exercise 3.53. Show that the pull-back of any closed form is closed, and the pull-
back of any exact form is exact.

Exercise 3.54. Consider the unit sphere S2 locally parametrized by

F (θ, ϕ) = (sin ϕ cos θ, sin ϕ sin θ, cos ϕ).
Define a map Φ : S → R3 by Φ(x, y, z) = (xz, yz, z 2 ), and consider a 2-form
2

ω = z dx ∧ dy. Compute dω, Φ∗ (dω), Φ∗ ω and d(Φ∗ ω), and verify they satisfy
Proposition 3.57.

3.5.6. Unification of Green’s, Stokes’ and Divergence Theorems. Given a sub-

manifold M m in Rn , a differential form on Rn induces a differential form on M m . For
example, let C be a smooth regular curve in R3 parametrized by γ(t) = (x(t), y(t), z(t)).
The 1-form:
α = αx dx + αy dy + αz dz
is a priori defined on R3 , but we can regard the coordinates (x, y, z) as functions on the
dx
curve C parametrized by γ(t), then we have dx = dt and similarly for dy and dz. As
dt
such, dx can now be regarded as a 1-form on C. Therefore, the 1-form α on R3 induces
a 1-form α (abuse in notation) on C:
dx dy dz
α = αx (γ(t)) dt + αy (γ(t)) dt + αz (γ(t)) dt
 dt dt dt
dx dy dz
= αx (γ(t)) + αy (γ(t)) + αz (γ(t)) dt
dt dt dt
In practice, there is often no issue of using α to denote both the 1-form on R3 and its
induced 1-form on C. To be (overly) rigorous over notations, we can use the inclusion
map ι : C → R3 to distinguish them. The 1-form α on R3 is transformed into a 1-form
ι∗ α on C by the pull-back of ι. From the previous subsection, we learned that:
ι∗ (dx) = d(ι∗ x) = d(x ◦ ι).
Note that dx and d(x ◦ ι) are different in a sense that x ◦ ι : C → R has the curve C as its
domain, while x : R3 → R has R3 as its domain. Therefore, we have:
d(x ◦ ι) dx
d(x ◦ ι) = dt = dt.
dt dt
dx
In short, we may use ι∗ (dx) = dt to distinguish it from dx if necessary. Similarly,
∗
dt
we may use ι α to denote the induced 1-form of α on C:
 
dx dy dz
ι∗ α = αx (γ(t)) + αy (γ(t)) + αz (γ(t)) dt.
dt dt dt
An induced 1-form on a curve in R3 is related to line integrals in Multivariable
Calculus. Recall that the 1-form α = αx dx + αy dy + αz dz corresponds to the vector field
F = αx î + αy ĵ + αz k̂ on R3 . In Multivariable Calculus, we denote dl = dxî + dy ĵ + dz k̂
and    
F · dl = αx î + αy ĵ + αz k̂ · dxî + dy ĵ + dz k̂ = α.
Z
The line integral F · dl over the curve C ⊂ R3 can be written using differential form
C
notations: Z Z Z
F · dl = α or more rigorously: ι∗ α.
C C C
104 3. Tensors and Differential Forms

Now consider a regular surface M ⊂ R3 . Suppose F (u, v) = (x(u, v), y(u, v), z(u, v))
is a smooth local parametrization of M . Consider a vector G = βx î + βy ĵ + βz k̂ on R3
and its corresponding 2-form on R3 :
β = βx dy ∧ dz + βy dz ∧ dx + βz dx ∧ dy.
Denote ι : M → R the inclusion map. The induced 2-form ι∗ β on M is in fact related to
3

the surface flux of G through M . Let’s explain why:

ι∗ (dy ∧ dz) = (ι∗ dy) ∧ (ι∗ dz) = d(y ◦ ι) ∧ d(z ◦ ι)
   
∂y ∂y ∂z ∂z
= du + dv ∧ du + dv
∂u ∂v ∂u ∂v
 
∂y ∂z ∂z ∂y
= − du ∧ dv
∂u ∂v ∂u ∂v
∂(y, z)
= det du ∧ dv.
∂(u, v)
Similarly, we have:
∂(z, x)
ι∗ (dz ∧ dx) = det du ∧ dv
∂(u, v)
∂(x, y)
ι∗ (dx ∧ dy) = det du ∧ dv
∂(u, v)
All these show:
 
∂(y, z) ∂(z, x) ∂(x, y)
ι∗ β = βx det + βy det + βz det du ∧ dv
∂(u, v) ∂(u, v) ∂(u, v)
Compared with the flux element G · ν dS in Multivariable Calculus:
  ∂F × ∂F ∂F ∂F
∂u ∂v
G · ν dS = βx î + βy ĵ + βz k̂ · ∂F ∂F
× dudv
| {z ×
} | ∂u {z ∂v } | ∂u ∂v
{z }
G dS
ν
   ∂(y, z) ∂(z, x) ∂(x, y)

= βx î + βy ĵ + βz k̂ · det î + det ĵ + det k̂
∂(u, v) ∂(u, v) ∂(u, v)
 
∂(y, z) ∂(z, x) ∂(x, y)
= βx det + βy det + βz det dudv,
∂(u, v) ∂(u, v) ∂(u, v)
the only difference is that ι∗ β is in terms of the wedge product du ∧ dv while the flux
element G · ν dS is in terms of dudv. IgnoringZ Zthis minor difference (which will be
addressed in the next chapter), the surface flux G · ν dS can be expressed in terms
M
of differential forms in the following way:
ZZ ZZ ZZ
G · ν dS = β or more rigorously: ι∗ β
M M M

Recall that the classical Stokes’ Theorem is related to line integrals of a curve
and surface flux of a vector field. Based on the above discussion, we see that Stokes’
Theorem can be restated in terms of differential forms. Consider the 1-form α =
αx dx + αy dy + αz dz and its corresponding vector field F = αx î + αy ĵ + αz k̂. We have
already discussed that the 2-form dα corresponds to the vector field ∇×F . Therefore, the
surface flux of the vector field ∇ × F through M can be expressed in terms of differential
forms as: ZZ ZZ ZZ
∗
(∇ × F ) · ν dS = ι (dα) = d(ι∗ α).
M M M
3.5. Exterior Derivatives 105

If C is the boundary curve of M , then from our previous discussion we can write:
Z Z
F · dl = ι∗ α.
C C

The classical Stokes’ Theorem asserts that:

Z ZZ
F · dl = (∇ × F ) · ν dS
C M

which can be expressed in terms of differential form as:

Z ZZ Z ZZ
ι∗ α = d(ι∗ α) or simply: α= dα.
C M C M

Due to this elegant way (although not very practical for physicists and engineers) of
expressing Stokes’ Theorem, we often denote the boundary of a surface M as ∂M , then
the classical Stokes’ Theorem can be expressed as:
Z ZZ
α= dα.
∂M M

Using differential forms, one can also express Divergence Theorem in Multivariable
Calculus in a similar way as above. Let D be a solid region in R3 and ∂D be the boundary
surface of D. Divergence Theorem in MATH 2023 asserts that:
ZZ ZZZ
G · ν dS = ∇ · G dV,
∂D D
ZZ
where G = βx î + βy ĵ + βz k̂. As discussed before, the LHS is β where β =
∂D
βx dy ∧ dz + βy dz ∧ dx + βz dx ∧ dy. We have seen that:
 
∂βx ∂βy ∂βz
dβ = + + dx ∧ dy ∧ dz,
∂x ∂y ∂z
which is (almost) the same as:
 
∂βx ∂βy ∂βz
∇ · G dV = + + dxdydz.
∂x ∂y ∂z
ZZZ
Hence, the RHS of Divergence Theorem can be expressed as dβ; and therefore we
D
can rewrite Divergence Theorem as:
ZZ ZZZ
β= dβ.
∂D D

Again, the same expression! Stokes’ and Divergence Theorems can therefore be unified.
Green’s Theorem can also be unified with Stokes’ and Divergence Theorems as well. Try
the exercise below:

Exercise 3.55. Let C be a simple closed smooth curve in R2 and R be the region
enclosed by C in R2 . Given a smooth vector field F = P î + Qĵ on R2 , Green’s
Theorem asserts that:
Z ZZ  
∂Q ∂P
F · dl = − dxdy.
C R ∂x ∂y
Express Green’s Theorem using the language of differential forms.
106 3. Tensors and Differential Forms

3.5.7. Differential Forms and Maxwell’s Equations. The four Maxwell’s equa-
tions are a set of partial differential equations that form the foundation of electromag-
netism. Denote the components of the electric field e, magnetic field B, and current
density ĵ by

E = Ex î + Ey ĵ + Ez k̂
B = Bx î + By ĵ + Bz k̂
J = Jx î + Jy ĵ + Jz k̂

All components of E, B and J are considered to be time-dependent. Denote ρ to be the

charge density. The four Maxwell’s equations assert that:

∇·E=ρ ∇·B=0
∂B ∂E
∇×E=− ∇×B=J+
∂t ∂t

These four equations can be rewritten using differential forms in a very elegant way.
Consider R4 with coordinates (t, x, y, z), which is also denoted as (x0 , x1 , x2 , x3 ) in this
problem. First we introduce the Minkowski Hodge-star operator ∗ on R4 , which is a linear
map taking p-forms on R4 to (4 − p)-forms on R4 . In particular, for 2-forms ω = dxi ∧ dxj
(where i, j = 0, 1, 2, 3 and i 6= j), we define ∗ω to be the unique 2-form on R4 such that:
(
dt ∧ dx ∧ dy ∧ dz if i, j 6= 0
ω ∧ ∗ω =
−dt ∧ dx ∧ dy ∧ dz otherwise.

For instance, ∗(dx ∧ dy) = dt ∧ dz since dx ∧ dy ∧ dt ∧ dz = dt ∧ dx ∧ dy ∧ dz and

there is no dt term in dx ∧ dy. On the other hand, ∗(dt ∧ dx) = −dy ∧ dz since there is a
dt term in dt ∧ dx. The operator ∗ then extends linearly to all 2-forms on R4 .

Exercise 3.56. Compute each of the following:

∗(dt ∧ dx) ∗(dt ∧ dy) ∗(dt ∧ dz)
∗(dx ∧ dy) ∗(dy ∧ dz) ∗(dz ∧ dx)

To convert the Maxwell’s equations using the language of differential forms, we

define the following analogue of E, B, J and ρ using differential forms:

E = Ex dx + Ey dy + Ez dz
B = Bx dy ∧ dz + By dz ∧ dx + Bz dx ∧ dy
J = −(Jx dy ∧ dz + Jy dz ∧ dx + Jz dx ∧ dy) ∧ dt + ρ dx ∧ dy ∧ dz

Note that Ei ’s and Bj ’s may depend on t although there is no dt above. Define the
2-form:
F := B + E ∧ dt.

Exercise 3.57. Show that the four Maxwell’s equations can be rewritten in an
elegant way as:
dF = 0
d (∗F ) = J
where d is the exterior derivative on R4 .
3.5. Exterior Derivatives 107

3.5.8. Global Expressions of Exterior Derivatives. We defined exterior differen-

tiation using local coordinates. In fact, using Lie derivatives, one can derive a global
expression (i.e. without using local coordinates) of exterior derivatives on differential
forms.
We first introduce Lie derivatives on (p, 0)-tensors, which are similarly defined as
those on 1-forms. Let T be a (p, 0)-tensor and X be a vector field on M . Denote the flow
map of X by Φt , then the Lie derivative of T along X at p ∈ M is defined as:
d
Φ∗t TΦt (p) .


(LX T )p :=
dt t=0

Exercise 3.58. Guess the definition of Lie derivatives of a general (p, q)-tensor along
a vector field X. Check any standard textbook to see if your guess is right.

Remark 3.58. On a regular surface M in R3 with the first fundamental form denoted by
g, if X is a vector field on M such that LX g = 0, then we call X to be a Killing vector
field. The geometric meaning of such an X is that g is invariant when M moves along
the vector field X, or equivalently, g is symmetric in the direction of X. This concept
of Killing vector fields can be generalize to Riemannian manifolds and is important in
Differential Geometry and General Relativity, whenever symmetry plays an important
role. 

Since the pull-back of a tensor product satisfies Φ∗ (T ⊗ S) = Φ∗ T ⊗ Φ∗ S, it is easy

to show from definition that the Lie derivative satisfies the product rule:
(3.21) LX (T ⊗ S) = (LX T ) ⊗ S + T ⊗ (LX S).

Exercise 3.59. Prove (3.21).

Since differential forms are simply linear combinations of tensor products, the definition
of their Lie derivatives is the same as that for (k, 0)-tensors. One nice fact about Lie
derivatives on differential forms is so-called the Cartan’s magic formula, which relates Lie
derivatives and exterior derivatives. We first introduce the interior product:

Definition 3.59 (Interior Product). Let α be a k-form (where k ≥ 2) on a manifold M ,

and X be a vector field on M . Then, the interior product iX α is a (k − 1)-form defined
as follows. For any vector fields Y1 , . . . , Yk−1 on M , we define:
(iX α)(Y1 , . . . , Yk−1 ) := α(X, Y1 , . . . , Yk−1 ).
Pn ∂
Example 3.60. In local coordinates, if a vector field X can be written as X = i=1 X i ∂u i
,
j k
then iX (du ∧ du ) is an 1-form and we have:
   
j k

∂ j k ∂
iX (du ∧ du ) = (du ∧ du ) X, = X j δkl − X k δjl .
∂ul ∂ul
In other words, we have:
n
X
iX (duj ∧ duk ) = (X j δkl − X k δjl ) dul = X j duk − X k duj .
l=1

108 3. Tensors and Differential Forms

Pn ∂
Exercise 3.60. Let X = i=1 X i ∂u i
be a vector field on a manifold M with local
coordinates (u1 , . . . , un ). Derive the local expression of:
iX (duj1 ∧ duj2 ∧ · · · ∧ dujk )
where 1 ≤ j1 < j2 < · · · < jk ≤ n.

Now we are ready to present a beautiful and elegant formula due to Elie Cartan:

Proposition 3.61 (Cartan’s Magic Formula). Let X be a smooth vector field on a manifold
M , then for any differential k-form ω, we have:
(3.22) LX ω = iX (dω) + d(iX ω)

Proof. The proof is by induction on k, the degree of ω. We first show that (3.22) holds
for 1-forms.
Pn
Consider ω = j=1 ωj duj , we have already computed in (3.11) that:
n 
∂X i

X ∂ωj
LX ω = Xi + ωi duj .
i,j=1
∂ui ∂uj

Next we verify it is equal to the RHS of (3.22).

n
X ∂ωj i
dω = du ∧ duj
i,j=1
∂ui

From Example 3.60, we have

n n
X ∂ωj i j
X ∂ωj
iX (dω) = iX (du ∧ du ) = (X i duj − X j dui ).
i,j=1
∂ui i,j=1
∂ui

Moreover, we have
n
X
iX ω = ω(X) = X j ωj
j=1
n 
∂X j

X ∂ωj
d(iX ω) = ωj + X j dui
i,j=1
∂ui ∂ui

and it follows easily that:

n n
X ∂ωj j X ∂X j i
iX (dω) + d(iX ω) = Xi du + ωj du
i,j=1
∂ui i,j=1
∂ui

which is exactly LX ω after relabelling indices.

Now that (3.22) holds for 1-form. To complete the inductive proof, we just need to
show that if (3.22) holds for both differential forms ω and σ, then it also holds for ω ∧ σ.
It is left as an exercise for readers. 

Exercise 3.61. Complete the above inductive proof. [Note: the proof is somewhat
algebraic.]

Exercise 3.62. Show that if ω is closed, then LX ω is exact for any vector field X.
3.5. Exterior Derivatives 109

The purpose of introducing Cartan’s magic formula is it gives a coordinate-free

expression of exterior derivatives. Consider a 1-form ω, and two vector fields X and Y .
Then, from (3.22), we have:
(LX ω)(Y ) = (iX (dω)) (Y ) + (d(iX ω)) (Y ),
which, from the definition of iX and (3.19), can be simplified to:
X(ω(Y )) − ω(LX Y ) = (dω)(X, Y ) + d(ω(X))(Y ).
As ω(X) is a scalar function, we also have:
d(ω(X))(Y ) = Y (ω(X)) .
[Note that generally, (df )(Y ) = Y (f ) for any scalar function f .]
Finally, we get:
(3.23) (dω)(X, Y ) = X (ω(Y )) − Y (ω(X)) − ω([X, Y ])
for any vector fields X and Y . This is a global formula for dω as it does not involve any
local coordinates.
The expression (3.23) can be generalized to k-forms ω. The proof is by induction and
the Cartan’s magic formula again. For any k-form ω, and vector fields X0 , X1 , . . . , Xk ,
we have:
(dω)(X0 , X1 , · · · , Xk )
k
X
= (−1)i Xi (ω(X0 , · · · , X̂i , · · · , Xk ))
i=0
X  
+ (−1)i+j ω [Xi , Xj ], X0 , · · · , X̂i , · · · , X̂j , · · · , Xk .
0≤i<j≤k

Readers interested in the proof may consult [Lee13, P.370, Proposition 14.32].
 Chapter 4

Generalized Stokes’
Theorem

“It is very difficult for us, placed as we

have been from earliest childhood in a
condition of training, to say what
would have been our feelings had such
training never taken place.”

Sir George Stokes, 1st Baronet

4.1. Manifolds with Boundary

We have seen in the Chapter 3 that Green’s, Stokes’ and Divergence Theorem in Multi-
variable Calculus can be unified together using the language of differential forms. In
this chapter, we will generalize Stokes’ Theorem to higher dimensional and abstract
manifolds.
These classic theorems and their generalizations concern about an integral over a
manifold with an integral over its boundary. In this section, we will first rigorously define
the notion of a boundary for abstract manifolds. Heuristically, an interior point of a
manifold locally looks like a ball in Euclidean space, whereas a boundary point locally
looks like an upper-half space.

4.1.1. Smooth Functions on Upper-Half Spaces. From now on, we denote Rn+ :=
{(u1 , . . . , un ) ∈ Rn : un ≥ 0} which is the upper-half space of Rn . Under the subspace
topology, we say a subset V ⊂ Rn+ is open in Rn+ if there exists a set Ve ⊂ Rn open in Rn
such that V = Ve ∩ Rn+ . It is intuitively clear that if V ⊂ Rn+ is disjoint from the subspace
{un = 0} of Rn , then V is open in Rn+ if and only if V is open in Rn .
Now consider a set V ⊂ Rn+ which is open in Rn+ and that V ∩ {un = 0} 6= ∅. We
need to first develop a notion of differentiability for functions such an V as their domain.
Given a vector-valued function G : V → Rm , then near a point u ∈ V ∩ {un = 0}, we can
only approach u from one side only, namely from directions with positive un -coordinates.
The usual definition of differentiability does not apply at such a point, so we define:

111
112 4. Generalized Stokes’ Theorem

Definition 4.1 (Functions of Class C k on Rn+ ). Let V ⊂ Rn+ be open in Rn+ and that
V ∩ {un = 0} 6= ∅. Consider a vector-valued function G : V → Rm . We say G is C k
(resp. smooth) at u ∈ V ∩ {un = 0} if there exists a C k (resp. smooth) local extension
Ge : Bε (u) → Rm such that G(y)
e = G(y) for any y ∈ Bε (u) ∩ V . Here Bε (u) ⊂ Rn refers
n
to an open ball in R .
If G is C k (resp. smooth) at every u ∈ V (including those points with un > 0), then
we say G is C k (resp. smooth) on V .

Figure 4.1. G is C k at u if there exists a local extension G

e near u.

Example 4.2. Let V = {(x, y) : y ≥ 0 and x2 + y 2 < 1}, which is an open set in R2+
since V = {(x, y) : x2 + y 2 < 1} ∩ R2+ . Then f (x, y) : V → R defined by f (x, y) =
| {z }
open in R2
p p
x2 y2
1 − − is a smooth function on V since 1 − x2 − y 2 is smoothly on the whole
ball x2 + y 2 < 1.
√
However, the function g : V → R defined by g(x, y) = y is not smooth at every
∂g
point on the y-axis because ∂y → ∞ as y → 0+ . Any extension ge of g will agree with g
∂e
g
on the upper-half plane, and hence will also be true that ∂y → ∞ as y → 0+ , which is
sufficient to argue that such ge is not smooth. 

Exercise 4.1. Consider f : R2+ → R defined by f (x, y) = |x|. Is f smooth on R2+ ? If

not, at which point(s) in R2+ is f not smooth? Do the same for g : R2+ → R defined
by g(x, y) = |y|.

4.1.2. Boundary of Manifolds. After understanding the definition of a smooth

function when defined on subsets of the upper-half space, we are ready to introduce the
notion of manifolds with boundary:
4.1. Manifolds with Boundary 113

Definition 4.3 (Manifolds with Boundary). We say M is a smooth manifold with

boundary if there exist two families of local parametrizations Fα : Uα → M where Uα is
open in Rn , and Gβ : Vβ → M where Vβ is open in Rn+ such that every Fα and Gβ is a
homeomorphism between its domain and image, and that the transition functions of all
types:
Fα−1 ◦ Fα0 Fα−1 ◦ Gβ G−1
β ◦ Gβ
0 G−1
β ◦ Fα
0 0
are smooth on the overlapping domain for any α, α , β and β .
Moreover, we denote and define the boundary of M by:
[
∂M := {Gβ (u1 , . . . , un−1 , 0) : (u1 , . . . , un−1 , 0) ∈ Vβ }.
β

Remark 4.4. In this course, we will call these Fα ’s to be local parametrizations of interior
type, and these Gβ ’s to be local parametrizations of boundary type. 

Figure 4.2. A manifold with boundary

Example 4.5. Consider the solid ball B2 := {x ∈ R2 : |x| ≤ 1}. It can be locally
parametrized using polar coordinates by:
G : (0, 2π) × [0, 1) → B2
G(θ, r) := (1 − r)(cos θ, sin θ)
Note that the domain of G can be regarded as a subset
V := {(θ, r) : θ ∈ (0, 2π) and 0 ≤ r < 1} ⊂ R2+ .
Here we used 1 − r instead of r so that the boundary of B2 has zero r-coordinate, and
the interior of B2 has positive r-coordinate.
Note that the image of G does not cover the whole solid ball B2 . Precisely, the image
of G is B2 \{non-negative x-axis}. In order to complete the proof that B2 is a manifold
with boundary, we cover B2 by two more local parametrizations:
e : (−π, π) × [0, 1) → B2
G
e r) := (1 − r)(cos θ, sin θ)
G(θ,
and also the inclusion map ι : {u ∈ R2 : |u| < 1} → B2 . We need to show that the
transition maps are smooth. There are six possible transition maps:
e −1 ◦ G,
G G−1 ◦ G,
e ι−1 ◦ G, ι−1 ◦ G,
e G−1 ◦ ι, and e −1 ◦ ι.
G
114 4. Generalized Stokes’ Theorem

The first one is given by (we leave it as an exercise for computing these transition maps):
Ge −1 ◦ G : ((0, π) ∪ (π, 2π)) × [0, 1) → ((−π, 0) ∪ (0, π)) × [0, 1)
(
−1 (θ, r) if θ ∈ (0, π)
G ◦ G(θ, r) =
e
(θ − 2π, r) if θ ∈ (π, 2π)
which can be smoothly extended to the domain ((0, π) ∪ (π, 2π)) × (−1, 1). Therefore,
e −1 ◦ G is smooth. The second transition map G−1 ◦ G
G e can be computed and verified to
be smooth in a similar way.
For ι−1 ◦ G, by examining the overlap part of ι and G on B2 , we see that the domain
of the transition map is an open set (0, 2π) × (0, 1) in R2 . On this domain, ι−1 ◦ G is
essentially G, which is clearly smooth. Similar for ι−1 ◦ G.
e
To show G−1 ◦ ι is smooth, we use the Inverse Function Theorem. The domain of
ι ◦ G is (0, 2π) × (0, 1). By writing (x, y) = ι−1 ◦ G(θ, r) = (1 − r)(cos θ, sin θ), we check
−1

that on the domain of ι−1 ◦ G, we have:

∂(x, y)
det = 1 − r 6= 0.
∂(θ, r)
Therefore, the inverse G−1 ◦ ι is smooth. Similar for G
e −1 ◦ ι.
Combining all of the above verifications, we conclude that B2 is a 2-dimensional
manifold with boundary. The boundary ∂B2 is given by points with zero r-coordinates,
namely the unit circle {|x| = 1}. 

Exercise 4.2. Compute all transition maps

Ge −1 ◦ G, G−1 ◦ G,
e ι−1 ◦ G, ι−1 ◦ G,e G−1 ◦ ι, and e −1 ◦ ι
G
in Example 4.5. Indicate clearly their domains, and verify that they are smooth on
their domains.

Exercise 4.3. Let f : Rn → R be a smooth scalar function. The region in Rn+1

above the graph of f is given by:
Γf := {(u1 , . . . , un+1 ) ∈ Rn+1 : un+1 ≥ f (u1 , . . . , un )}.
Show that Γf is an n-dimensional manifold with boundary, and the boundary ∂Γf is
the graph of f in Rn+1 .

Exercise 4.4. Show that ∂M (assumed non-empty) of any n-dimensional manifold

M is an (n − 1)-dimensional manifold without boundary.

From the above example and exercise, we see that verifying a set is a manifold
with boundary may be cumbersome. The following proposition provides us with a very
efficient way to do so.

Proposition 4.6. Let f : M m → R be a smooth function from a smooth manifold M .

Suppose c ∈ R such that the set Σ := f −1 ([c, ∞)) is non-empty and that f is a submersion
at any p ∈ f −1 (c), then the set Σ is an m-dimensional manifold with boundary. The
boundary ∂Σ is given by f −1 (c).

Proof. We need to construct local parametrizations for the set Σ. Given any point p ∈ Σ,
then by the definition of Σ, we have f (p) > c or f (p) = c.
For the former case f (p) > c, we are going to show that near p there is a local
parametrization of Σ of interior type. Regarding p as a point in the manifold M , there
4.1. Manifolds with Boundary 115

exists a smooth local parametrization F : U ⊂ Rn → M of M covering p. We argue that

such a local parametrization of M induces naturally a local parametrization of Σ near
p. Note that f is continuous and so f −1 (c, ∞) is an open set of M containing p. Denote
O = f −1 (c, ∞), then F restricted to U ∩ F −1 (O) will have its image in O ⊂ Σ, and so is
a local parametrization of Σ near p.
For the later case f (p) = c, we are going to show that near p there is a local
parametrization of Σ of boundary type. Since f is a submersion at p, by the Submersion
Theorem (Theorem 2.48) there exist a local parametrization G : Ue → M of M near p,
and a local parametrization H of R near c such that G(0) = p and H(0) = c, and:
H −1 ◦ f ◦ G(u1 , . . . , um ) = um .
Without loss of generality, we assume that H is an increasing function near 0. We argue
that by restricting the domain of G to U ∩ {um ≥ 0}, which is an open set in Rm + , the
restricted G is a boundary-type local parametrization of Σ near p. To argue this, we note
that:
f (G(u1 , . . . , um )) = H(um ) ≥ H(0) = c whenever um ≥ 0.
Therefore, G(u1 , . . . , um ) ∈ f −1 ([c, ∞)) = Σ whenever um ≥ 0, and so G (when re-
stricted to U ∩ {um ≥ 0}) is a local parametrization of Σ.
Since all local parametrizations F and G of Σ constructed above are induced from
local parametrizations of M (whether it is of interior or boundary type), their transition
maps are all smooth. This shows Σ is an m-dimensional manifold with boundary. To
identify the boundary, we note that for any boundary-type local parametrization G
constructed above, we have:
H −1 ◦ f ◦ G(u1 , . . . , um−1 , 0) = 0
and so f (G(u1 , . . . , um−1 )) = H(0) = c, and therefore:
G(u1 , . . . , um−1 , 0) ∈ f −1 (c).
This show ∂Σ ⊂ f −1 (c). The other inclusion f −1 (c) ⊂ ∂Σ follows from the fact that
for any p ∈ f −1 (c), the boundary-type local parametrization G has the property that
G(0) = p (and hence p = G(0, . . . , 0, 0) ∈ ∂Σ). 
Remark 4.7. It is worthwhile to note that the above proof only requires that f is a
submersion at any p ∈ f −1 (c), and we do not require that it is a submersion at any
p ∈ Σ = f −1 ([c, ∞)). Furthermore, the codomain of f is R which has dimension 1, hence
f is a submersion at p if and only if the tangent map (f∗ )p at p is non-zero – and so it is
very easy to verify this condition. 

With the help of Proposition 4.6, one can show many sets are manifolds with
boundary by picking a suitable submersion f .
Example 4.8. The n-dimensional ball Bn = {x ∈ Rn : |x| ≤ 1} is an n-manifold with
boundary. To argue this, let f : Rn → R be the function:
2
f (x) = 1 − |x| .
Then Bn = f −1 ([0, ∞)).
The tangent map f∗ is represented by the matrix:
 
∂f ∂f
[f∗ ] = , ··· , = −2 [x1 , · · · , xn ]
∂x1 ∂xn
which is surjective if and only if (x1 , . . . , xn ) 6= (0, . . . , 0). For any x ∈ f −1 (0), we have
2
|x| = 1 and so in particular x 6= 0. Therefore, f is a submersion at every x ∈ f −1 (0).
By Proposition 4.6, we proved Bn = f −1 ([0, ∞)) is an n-dimensional manifold with
boundary, and the boundary is f −1 (0) = {x ∈ Rn : |x| = 1}, i.e. the unit circle. 
116 4. Generalized Stokes’ Theorem

Exercise 4.5. Suppose f : M m → R is a smooth function defined on a smooth

manifold M . Suppose a, b ∈ R such that Σ := f −1 ([a, b]) is non-empty, and that f is
a submersion at any p ∈ f −1 (a) and any q ∈ f −1 (b). Show that Σ is an m-manifold
with boundary, and ∂Σ = f −1 (a) ∪ f −1 (b).

4.1.3. Tangent Spaces at Boundary Points. On a manifold M n without ( boundary,

)n
∂
the tangent space Tp M at p is the span of partial differential operators ,
∂ui p
i=1
where (u1 , . . . , un ) are local coordinates of a parametrization F (u1 , . . . , un ) near p.
Now on a manifold M n with boundary, near any boundary point p ∈ ∂M n there
exists a local parametrization G(u1 , . . . , un ) : V ⊂ Rn+ → M of boundary type. Although
( )n
∂
G is only defined when un ≥ 0, we still define Tp M to be the span of .
∂ui p
i=1
Although such a definition of Tp M (when p ∈ ∂M ) is a bit counter-intuitive, the perk is
that Tp M is still a vector space. Given a vector V ∈ Tp M with coefficients:
n
X ∂
V = Vi .
i=1
∂ui p

We say that V is inward-pointing if V n > 0; and outward-pointing if V n < 0.

Furthermore, the tangent space Tp (∂M ) of the boundary manifold ∂M at p can be
regarded as a subspace of Tp M :
( )n−1
∂
Tp (∂M ) = span ⊂ Tp M.
∂ui p
i=1
4.2. Orientability 117

4.2. Orientability
In Multivariable Calculus, we learned (or was told) that Stokes’ Theorem requires the
surface to be orientable, meaning that the unit normal vector ν varies continuously on
the surface. The Möbius strip is an example of non-orientable surface.
Now we are talking about abstract manifolds which may not sit inside any Euclidean
space, and so it does not make sense to define normal vectors to the manifold. Even when
the manifold M is a subset of Rn , if the dimension of the manifold is dim M ≤ n − 2, the
manifold does not have a unique normal vector direction. As such, in order to generalize
the notion of orientability of abstract manifolds, we need to seek a reasonable definition
without using normal vectors.
In this section, we first show that for hypersurfaces M n in Rn+1 , the notion of
orientability using normal vectors is equivalent to another notion using transition maps.
Then, we extend the notion of orientability to abstract manifolds using transition maps.

4.2.1. Orientable Hypersurfaces. To begin, we first state the definition of ori-

entable hypersurfaces in Rn+1 :

Definition 4.9 (Orientable Hypersurfaces). A regular hypersurface M n in Rn+1 is said

to be orientable if there exists a continuous unit normal vector ν defined on the whole
Mn

Let’s explore the above definition a bit in the easy case n = 2. Given a regular surface
M 2 in R3 with a local parametrization (x, y, z) = F (u1 , u2 ) : U → M , one can find a
normal vector to the surface by taking cross product:
∂F ∂F ∂(y, z) ∂(z, x) ∂(x, y)
× = det î + det ĵ + det k̂
∂u1 ∂u2 ∂(u1 , u2 ) ∂(u1 , u2 ) ∂(u1 , u2 )
and hence the unit normal along this direction is given by:
∂(y,z) ∂(z,x) ∂(x,y)
det ∂(u1 ,u2 )
î + det ∂(u 1 ,u2 )
ĵ + det ∂(u1 ,u2 )
k̂
νF = on F (U).
∂(y,z) ∂(z,x) ∂(x,y)
det ∂(u1 ,u2 )
î + det ∂(u 1 ,u2 )
ĵ + det ∂(u1 ,u2 )
k̂

Note that the above ν is defined locally on the domain F (U).

Now given another local parametrization (x, y, z) = G(v1 , v2 ) : V → M , one can find
a unit normal using G as well:
∂(y,z) ∂(z,x) ∂(x,y)
det ∂(v 1 ,v2 )
î + det ∂(v 1 ,v2 )
ĵ + det ∂(v 1 ,v2 )
k̂
νG = on G(V).
∂(y,z) ∂(z,x) ∂(x,y)
det ∂(v 1 ,v2 )
î + det ∂(v 1 ,v2 )
ĵ + det ∂(v 1 ,v2 )
k̂

Using the chain rule, we have the following relation between the Jacobian determinants:
∂(∗, ∗∗) ∂(u1 , u2 ) ∂(∗, ∗∗)
det = det det
∂(v1 , v2 ) ∂(v1 , v2 ) ∂(u1 , u2 )
(here ∗ and ∗∗ mean any of the x, y and z) and therefore νF and νG are related by:
det ∂(u 1 ,u2 )
∂(v1 ,v2 )
νG = νF .
det ∂(u 1 ,u2 )
∂(v1 ,v2 )

Therefore, if there is an overlap between local coordinates (u1 , u2 ) and (v1 , v2 ), the unit
normal vectors νF and νG agree with each other on the overlap F (U) ∩ G(V) if and only
∂(u1 , u2 )
if det > 0 (equivalently, det D(F −1 ◦ G) > 0).
∂(v1 , v2 )
118 4. Generalized Stokes’ Theorem

From above, we see that consistency of unit normal vector on different local coor-
dinate charts is closely related to the positivity of the determinants of transition maps.
A consistence choice of unit normal vector ν exists if and only if it is possible to pick
a family of local parametrizations Fα : Uα → M 2 covering the whole M such that
det D(Fβ−1 ◦ Fα ) > 0 on Fα−1 (Fα (Uα ) ∩ Fβ (Uβ )) for any α and β in the family. The
notion of normal vectors makes sense only for hypersurfaces in Rn , while the notion of
transition maps can extend to any abstract manifold.
Note that given two local parametrizations F (u1 , u2 ) and G(v1 , v2 ), it is not always
∂(u1 , u2 )
possible to make sure det > 0 on the overlap even by switching v1 and v2 . It
∂(v1 , v2 )
is because it sometimes happens that the overlap F (U) ∩ G(V) is a disjoint union of
two open sets. If on one open set the determinant is positive, and on another one the
determinant is negative, then switching v1 and v2 cannot make the determinant positive
on both open sets. Let’s illustrate this issue through two contrasting examples: the
cylinder and the Möbius strip:

Example 4.10. The unit cylinder Σ2 in R3 can be covered by two local parametrizations:

F : (0, 2π) × R → Σ2 Fe : (−π, π) × R → Σ2

F (θ, z) := (cos θ, sin θ, z) Fe(θ,
e ze) := (cos θ,
e sin θ,
e ze)

Then, the transition map Fe−1 ◦ F is defined on a disconnected domain θ ∈ (0, π) ∪ (π, 2π)
and z ∈ R, and it is given by:
(
−1 (θ, z) if θ ∈ (0, π)
F ◦ F (θ, z) =
e
(θ − 2π, z) if θ ∈ (π, 2π)

By direct computations, the Jacobian of this transition map is given by:

D(Fe−1 ◦ F )(θ, z) = I

in either case θ ∈ (0, π) or θ ∈ (π, 2π). Therefore, det D(Fe−1 ◦ F ) > 0 on the overlap.
The unit normal vectors defined using these F and Fe:
∂F ∂F
∂r × ∂θ
νF = ∂F ∂F
on F ((0, 2π) × R)
∂r × ∂θ
∂F ∂F
×
e e
∂e
r ∂ θe
νFe = on Fe((−π, π) × R)
∂F ∂F
×
e e
∂e
r ∂ θe

will agree with each other on the overlap. Therefore, it defines a global continuous unit
normal vector across the whole cylinder. 

Example 4.11. The Möbius strip Σ2 in R3 can be covered by two local parametrizations:

F : (−1, 1) × (0, 2π) → Σ2 Fe : (−1, 1) × (−π, π) → Σ2

  
3+u e cos θ2 cos θe
e
3 + u cos θ2
cos θ


  
F (u, θ) =  3 + u cos θ2 sin θ  Fe(e
u, θ)
e = 3+u
 e cos θ2 sin θe
e

u sin θ2
e sin θ2
u
e
4.2. Orientability 119

In order to compute the transition map Fe−1 ◦ F (u, θ), we need to solve the system of
equations, i.e. find (e e in terms of (u, θ):
u, θ)
  !
θ θe
(4.1) 3 + u cos cos θ = 3 + u
e cos cos θe
2 2
  !
θ θe
(4.2) 3 + u cos sin θ = 3 + u
e cos sin θe
2 2
θ θe
(4.3) u sin =u
e sin
2 2
By considering (4.1)2 + (4.2)2 , we get:

θ θe
(4.4) u cos =u
e cos
2 2
We leave it as an exercise for readers to check that θ 6= π in order for the system to
be solvable. Therefore, θ ∈ (0, π) ∪ (π, 2π) and so the domain of overlap is a disjoint
union of two open sets.
When θ ∈ (0, π), from (4.3) and (4.4) we can conclude that u e = u and θe = θ.
When θ ∈ (π, 2π), we cannot have θe = θ since θe ∈ (−π, π). However, one can have
e = −u so that (4.3) and (4.4) become:
u
θ θe θ θe
sin = − sin and cos = − cos
2 2 2 2
which implies θe = θ − 2π.
To conclude, we have:
(
e−1 (u, θ) if θ ∈ (0, π)
F ◦ F (u, θ) =
(−u, θ − 2π) if θ ∈ (π, 2π)

By direct computations, we get:

(
e−1 1 if θ ∈ (0, π)
det D(F ◦ F )(u, θ) =
−1 if θ ∈ (π, 2π)

Therefore, no matter how we switch the order of u and θ, or u

e and θ,
e we can never allow
−1
det D(F ◦ F ) > 0 everywhere on the overlap. In other words, even if the unit normal
e
vectors νF and νFe agree with each other when θ ∈ (0, π), it would point in opposite
direction when θ ∈ (π, 2π). 

Next, we are back to hypersurfaces M n in Rn+1 and prove the equivalence between
consistency of unit normal and positivity of transition maps. To begin, we need the
following result about normal vectors (which is left as an exercise for readers):

Exercise 4.6. Let M n be a smooth hypersurface in Rn+1 whose coordinates are

denoted by (x1 , . . . , xn+1 ), and the unit vector along the xi -direction is denoted by
êi . Let F (u1 , . . . , un ) : U → M n be a local parametrization of M . Show that the
following vector defined on F (U) is normal to the hypersurface M n :
n+1
X ∂(xi+1 , . . . , xn+1 , x1 , . . . , xi−1 )
det êi .
i=1
∂(u1 , . . . , un )
120 4. Generalized Stokes’ Theorem

Proposition 4.12. Given a smooth hypersurface M n in Rn+1 , the following are equivalent:
(i) M n is orientable;
(ii) There exists a family of local parametrizations Fα : Uα → M covering M such that
for any Fα , Fβ in the family with Fβ (Uβ ) ∩ Fα (Uα ) 6= ∅, we have:
det D(Fα−1 ◦ Fβ ) > 0 on Fβ−1 (Fβ (Uβ ) ∩ Fα (Uα )).

Proof. We first prove (ii) =⇒ (i). Denote (uα α

1 , . . . , un ) to be the local coordinates of M
under the parametrization Fα . On every Fα (Uα ), using the result from Exercise 4.6, one
can construct a unit normal vector locally defined on Fα (Uα ):
Pn+1 ∂(xi+1 ,...,xn+1 ,x1 ,...,xi−1 )
i=1 det ∂(uα α
1 ,...,un )
êi
να = P
n+1 ∂(xi+1 ,...,xn+1 ,x1 ,...,xi−1 )
i=1 det ∂(uα ,...,uα ) 1 n
êi

Similarly, on Fβ (Uβ ), we have another locally defined unit normal vectors:

Pn+1 ∂(xi+1 ,...,xn+1 ,x1 ,...,xi−1 )
i=1 det ∂(uβ β êi
1 ,...,un )
νβ = P
n+1 ∂(xi+1 ,...,xn+1 ,x1 ,...,xi−1 )
i=1 det ∂(uβ ,...,uβ ) n
êi
1

Then on the overlap Fβ−1 (Fα (Uα ) ∩ Fβ (Uβ )), the chain rule asserts that:

∂(xi+1 , . . . , xn+1 , x1 , . . . , xi−1 )

det
∂(uβ1 , . . . , uβn )
∂(uα α
1 , . . . , un ) ∂(xi+1 , . . . , xn+1 , x1 , . . . , xi−1 )
= det det
β β
∂(u1 , . . . , un ) ∂(uα α
1 , . . . , un )
∂(xi+1 , . . . , xn+1 , x1 , . . . , xi−1 )
= det D(Fα−1 ◦ Fβ ) det
∂(uα α
1 , . . . , un )

and so the two unit normal vectors are related by:

det D(Fα−1 ◦ Fβ )
νβ = να .
det D(Fα−1 ◦ Fβ )

By the condition that det D(Fα−1 ◦ Fβ ) > 0, we have νβ = να on the overlap. Define
ν := να on every Fα (Uα ), it is then a continuous unit normal vector globally defined on
M . This proves (i).
Now we show (i) =⇒ (ii). Suppose ν is a continuous unit normal vector defined on
the whole M . Suppose Fα (uα α
1 , . . . , un ) : Uα → M is any family of local parametrizations
that cover the whole M . On every Fα (Uα ), we consider the locally defined unit normal
vector:
Pn+1 ∂(xi+1 ,...,xn+1 ,x1 ,...,xi−1 )
i=1 det ∂(uα α
1 ,...,un )
êi
να = P .
n+1 ∂(xi+1 ,...,xn+1 ,x1 ,...,xi−1 )
i=1 det α
∂(u ,...,u ) α ê i
1 n

n n+1
As a hypersurface M in R , there is only one direction of normal vectors, and so
we have either να = ν or να = −ν on Fα (Uα ). For the latter case, one can modify the
parametrization Fα by switching any pair of uα
i ’s such that να = ν.
After making suitable modification on every Fα , we can assume without loss of
generality that Fα ’s are local parametrizations such that να = ν on every Fα (Uα ). In
particular, on the overlap Fβ−1 (Fα (Uα ) ∩ Fβ (Uβ )), we have να = νβ .
4.2. Orientability 121

det D(Fα−1 ◦ Fβ )
By νβ = να , we conclude that det D(Fα−1 ◦ Fβ ) > 0, proving (ii).
det D(Fα−1 ◦ Fβ )

Remark 4.13. According to Proposition 4.12, the cylinder in Example 4.10 is orientable,
while the Möbius strip in Example 4.11 is not orientable. 

Exercise 4.7. Show that the unit sphere S2 in R3 is orientable.

Exercise 4.8. Let f : R3 → R be a smooth function. Suppose c ∈ R such that f −1 (c)

is non-empty and f is a submersion at every p ∈ f −1 (c). Show that f −1 (c) is an
orientable hypersurface in R3 .

4.2.2. Orientable Manifolds. On an abstract manifold M , it is not possible to

define normal vectors on M , and so the notion of orientability cannot be defined using
normal vectors. However, thanks to Proposition 4.12, the notion of orientability of
hypersurfaces is equivalent to positivity of Jacobians of transition maps, which we can
also talk about on abstract manifolds. Therefore, motivated by Proposition 4.12, we
define:

Definition 4.14 (Orientable Manifolds). A smooth manifold M is said to be orientable

if there exists a family of local parametrizations Fα : Uα → M covering M such that for
any Fα and Fβ in the family with Fβ (Uβ ) ∩ Fα (Uα ) 6= ∅, we have:
det D(Fα−1 ◦ Fβ ) > 0 on Fβ−1 (Fβ (Uβ ) ∩ Fα (Uα )).
In this case, we call the family A = {Fα : Uα → M } of local parametrizations to be an
oriented atlas of M .

Example 4.15. Recall that the real projective space RP2 consists of homogeneous triples
[x0 : x1 : x2 ] where (x0 , x1 , x2 ) 6= (0, 0, 0). The standard parametrizations are given by:
F0 (x1 , x2 ) = [1 : x1 : x2 ]
F1 (y0 , y2 ) = [y0 : 1 : y2 ]
F2 (z0 , z1 ) = [z0 : z1 : 1]
By the fact that [y0 : 1 : y2 ] = [1 : y0−1 : y2 y0−1 ], the transition map F0−1 ◦ F1 is defined on
{(y0 , y2 ) ∈ R2 : y0 6= 0}, and is given by: (x1 , x2 ) = (y0−1 , y2 y0−1 ). Hence,
−y0−2
 
∂(x1 , x2 ) 0
D(F0−1 ◦ F1 ) = =
∂(y0 , y2 ) −y2 y0−2 y0−1
1
det D(F0−1 ◦ F1 ) = − 3
y0
Therefore, it is impossible for det D(F0−1 ◦ F1 ) > 0 on the overlap domain {(y0 , y2 ) ∈ R2 :
y0 6= 0}.
At this stage, we have shown that this altas is not an oriented one. In order to prove
RP2 is non-orientable, we need to show any altas of RP2 is not oriented. We will prove
this using Proposition 4.25 later. 
122 4. Generalized Stokes’ Theorem

Exercise 4.9. Show that RP3 is orientable. Propose a conjecture about the ori-
entability of RPn .

Exercise 4.10. Show that for any smooth manifold M (whether or not it is ori-
entable), the tangent bundle T M must be orientable.

Exercise 4.11. Show that for a smooth orientable manifold M with boundary, the
boundary manifold ∂M must also be orientable.
4.3. Integrations of Differential Forms 123

4.3. Integrations of Differential Forms

Generalized Stokes’ Theorem concerns about integrals of differential forms. In this
section, we will give a rigorous definition of these integrals.

4.3.1. Single Parametrization. In the simplest case if a manifold M can be covered

by a single parametrization:
F (u1 , . . . , un ) : (α1 , β1 ) × · · · × (αn , βn ) → M,
then given an n-form ϕ(u1 , . . . , un ) du1 ∧ du2 ∧ · · · ∧ dun , the integral of ω over the
manifold M is given by:
Z Z βn Z β1
1 2 n
ϕ(u1 , . . . , un ) du ∧ du ∧ · · · ∧ du := ··· ϕ(u1 , . . . , un ) du1 du2 · · · dun
M αn α1
| {z } | {z }
integral of differential form ordinary integral in Multivariable Calculus

From the definition, we see that it only makes sense to integrate an n-form on an
n-dimensional manifold.
Very few manifolds can be covered by a single parametrization. Of course, Rn
is an example. One less trivial example is the graph of a smooth function. Suppose
f (x, y) : R2 → R is a smooth function. Consider its graph:
Γf := {(x, y, f (x, y)) ∈ R3 : (x, y) ∈ R2 }
which can be globally parametrized by F : R2 → Γf where
F (x, y) = (x, y, f (x, y)).
2 2
Let ω = e−x −y
dx ∧ dy be a 2-form on Γf , then its integral over Γf is given by:
Z Z Z ∞Z ∞
−x2 −y 2 2 2
ω= e dx ∧ dy = e−x −y dx dy = π.
Γf Γf −∞ −∞

Here we leave the computational detail as an exercise for readers.

It appears that integrating a differential form is just like “erasing the wedges”, yet
there are two subtle (but important) issues:
(1) In the above example, note that ω can also be written as:
2
−y 2
ω = −e−x dy ∧ dx.
It suggests that we also have:
Z Z ∞Z ∞
2
−y 2
ω= −e−x dy dx = −π,
Γf −∞ −∞

which is not consistent with the previous result. How shall we fix it?
(2) Even if a manifold can be covered by one single parametrization, such a parametriza-
tion may not be unique. If both (u1 , . . . , un ) and (v1 , . . . , vn ) are global coordinates
of M , then a differential form ω can be expressed in terms of either ui ’s or vi ’s. Is
the integral independent of the chosen coordinate system?
The first issue can be resolved easily. Whenever we talk about integration of differential
forms, we need to first fix the order of the coordinates. Say on R2 we fix the order to be
(x, y), then for any given 2-form we should express it in terms of dx ∧ dy before “erasing
the wedges”. For the 2-form ω above, we must first express it as:
2
−y 2
ω = e−x dx ∧ dy
before integrating it.
124 4. Generalized Stokes’ Theorem

For higher (say dim = 4) dimensional manifolds M 4 covered by a single parametriza-

tion F (u1 , . . . , u4 ) : U → M , if we choose (u1 , u2 , u3 , u4 ) to be the order of coordinates
and given a 4-form:
Ω = f (u1 , . . . , u4 ) du1 ∧ du3 ∧ du2 ∧ du4 + g(u1 , . . . , u4 ) du4 ∧ du3 ∧ du2 ∧ du1 ,
then we need to re-order the wedge product so that:
Ω = −f (u1 , . . . , u4 ) du1 ∧ du2 ∧ du3 ∧ du4 + g(u1 , . . . , u4 )du1 ∧ du2 ∧ du3 ∧ du4 .
The integral of ω over M 4 with respect to the order (u1 , u2 , u3 , u4 ) is given by:
Z Z
Ω= (−f (u1 , . . . , u4 ) + g(u1 , . . . , u4 )) du1 du2 du3 du4 .
M U

Let’s examine the second issue. Suppose M is an n-manifold with two different
global parametrizations F (u1 , . . . , un ) : U → M and G(v1 , . . . , vn ) : V → M . Given an
n-form ω which can be expressed as:
ω = ϕ du1 ∧ · · · ∧ dun ,
then from Proposition 3.56, ω can be expressed in terms of vi ’s by:
∂(u1 , . . . , un ) 1
ω = ϕ det dv ∧ · · · ∧ dv n .
∂(v1 , . . . , vn )
Recall that the change-of-variable formula in Multivariable Calculus asserts that:
Z Z
1 n ∂(u1 , . . . , un )
ϕ du · · · du = ϕ det dv 1 · · · dv n .
U V ∂(v 1 , . . . , v n )
Z
Therefore, in order for ω to be well-defined, we need
M
Z Z
∂(u1 , . . . , un ) 1
ϕ du1 ∧ · · · ∧ dun and ϕ det dv ∧ · · · ∧ dv n
F (U ) F (V) ∂(v1 , . . . , vn )
to be equal, and so we require:
∂(u1 , . . . , un )
det > 0.
∂(v1 , . . . , vn )
When defining an integral of a differential form, we not only need to choose a
convention on the order of coordinates, say (u1 , . . . , un ), but also we shall only consider
∂(u1 , . . . , un )
those coordinate systems (v1 , . . . , vn ) such that det > 0. Therefore, in
∂(v1 , . . . , vn )
order to integrate a differential form, we require the manifold to be orientable.

4.3.2. Multiple Parametrizations. A majority of smooth manifolds are covered by

more than one parametrizations. Integrating a differential form over such a manifold is
not as straight-forward as previously discussed.
In case M can be “almost” covered by a single parametrization F : U → M (i.e. the
set M \F (U)
Z has measure zero)
Z and the n-form ω is continuous, then it is still possible to
compute ω by computing ω. Let’s consider the example of a sphere:
M F (U )

Example 4.16. Let S2 be the unit sphere in R3 centered at the origin. Consider the
2-form ω on R3 defined as:
ω = dx ∧ dy.
2 3
Let ι : S → R be the inclusion
Z map, then ι∗ ω is a 2-form on S2 . We are interested in
the value of the integral ι∗ ω.
S2
4.3. Integrations of Differential Forms 125

Note that S2 can be covered almost everywhere by spherical coordinate parametriza-

tion F (ϕ, θ) : (0, π) × (0, 2π) → S2 given by:
F (ϕ, θ) = (sin ϕ cos θ, sin ϕ sin θ, cos ϕ).
Under the local coordinates (ϕ, θ), we have:
ι∗ (dx) = d(sin ϕ cos θ) = cos ϕ cos θ dϕ − sin ϕ sin θ dθ
ι∗ (dy) = d(sin ϕ sin θ) = cos ϕ sin θ dϕ + sin ϕ cos θ dθ
ι∗ ω = ι∗ (dx) ∧ ι∗ (dy)
= sin ϕ cos ϕ dϕ ∧ dθ.
Therefore,
Z Z Z 2π Z π
ι∗ ω = sin ϕ cos ϕ dϕ ∧ dθ = sin ϕ cos ϕ dϕ dθ = 0.
M M 0 0

Here we pick (ϕ, θ) as the order of coordinates. 

Exercise 4.12. Let ω = x dy ∧ dz + y dz ∧ dx + z dx ∧ dy. Compute

Z
ι∗ ω
S2
where S is the unit sphere in R centered at the origin, and ι : S2 → R3 is the
2 3

inclusion map.

Exercise 4.13. Let T2 be the torus in R4 defined as:

 
2 4 2 2 2 2 1
T := (x1 , x2 , x3 , x4 ) ∈ R : x1 + x2 = x3 + x4 = .
2
Let ι : T2 → R4 be the inclusion map. Compute the following integral:
Z
ι∗ x1 x2 x3 dx4 ∧ dx3 .


T2

FYI: Clifford Torus

2
The torus T in Exercise 4.13 is a well-known object in Differential Geometry called the
Clifford Torus. A famous conjecture called the Hsiang-Lawson’s Conjecture concerns about
this torus. One of the proposers Wu-Yi Hsiang is a retired faculty of HKUST Math. This
conjecture was recently solved by Simon Brendle in 2012.

Next, we will discuss how to define integrals of differential forms when M is covered
by multiple parametrizations none of which can almost cover the whole manifold. The
key idea is to break down the n-form into small pieces, so that each piece is completely
covered by one single parametrization. It will be done using partition of unity to be
discussed.
We first introduce the notion of support which appears often in the rest of the course
(as well as in advanced PDE courses).

Definition 4.17 (Support). Let M be a smooth manifold. Given a k-form ω (where

0 ≤ k ≤ n) defined on M , we denote and define the support of ω to be:
supp ω := {p ∈ M : ω(p) 6= 0},
i.e. the closure of the set {p ∈ M : ω(p) 6= 0}.
126 4. Generalized Stokes’ Theorem

Suppose M n is an oriented manifold with F (u1 , . . . , un ) : U → M as one of (many)

local parametrizations. If an n-form ω on M n only has “stuff” inside F (U), or precisely:
supp ω ⊂ F (U),
Z
then one can define ω as in the previous subsection. Namely, if on F (U) we have
M
ω = ϕ du1 ∧ · · · ∧ dun , then we define:
Z Z Z
ω= ω= ϕ du1 · · · dun .
M F (U ) U

Here we pick the order of coordinates to be (u1 , . . . , un ).

The following important tool called partitions of unity will “chop” a differential form
into “little pieces” such that each piece has support covered by a single parametrization.

Definition 4.18 (Partitions of Unity).

[ Let M be a smooth manifold with an atlas
A = {Fα : Uα → M } such that M = Fα (Uα ). A partition of unity subordinate to the
all α
atlas A is a family of smooth functions ρα : M → [0, 1] with the following properties:
(i) supp ρα ⊂ Fα (Uα ) for any α.
(ii) For any p ∈ M , there exists an open set O ⊂ M containing p such that
supp ρα ∩ O =
6 ∅
for finitely many α’s only.
X
(iii) ρα ≡ 1 on M .
all α

Remark 4.19. It can be shown that given any smooth manifold with any atlas, partitions
of unity subordinate to that given atlas must exist. The proof is very technical and is
not in the same spirit with other parts of the course, so we omit the proof here. It is
more important to know what partitions of unity are for, than to know the proof of
existence. 
Remark 4.20. Note that partitions of unity subordinate to a given atlas may not be
unique! 
Remark 4.21. Condition
P (ii) in Definition 4.18 is merely a technical analytic condition
to make sure the sum all α ρα (p) is a finite sum for each fixed p ∈ M , so that we do not
need to worry about convergence issues. If the manifold can be covered by finitely many
local parametrizations, then condition (ii) automatically holds (and we do not need to
worry about). 

Now, take an n-form ω defined on an orientable manifold M n , which is parametrized

by an oriented atlas A = {Fα : Uα → M }. Let {ρα : M → [0, 1]} be a partition of unity
subordinate to A, then by condition (iii) in Definition 4.18, we get:
!
X X
ω= ρα ω = ρα ω.
all α all α
| {z }
=1
Condition (i) says that supp ρα ⊂ Fα (Uα ), or heuristically speaking ρα vanishes outside
Fα (Uα ). Naturally, we have supp (ρα ω) ⊂ Fα (Uα ) for each α. Therefore, as previously
discussed, we can integrate ρα ω for each individual α:
Z Z
ρα ω := ρα ω.
M Fα (Uα )
4.3. Integrations of Differential Forms 127

Given that we can integrate each ρα ω, we define the integral of ω as:

Z XZ XZ
(4.5) ω := ρα ω = ρα ω.
M all α M all α Fα (Uα )

However, the sum involved in (4.5) is in general an infinite (possible uncountable!)

sum. To avoid convergence issue, from now on we will only consider n-forms ω which
have compact support, i.e.
supp ω is a compact set.

Recall that every open cover of a compact set has a finite sub-cover. Together with
condition (ii) in Definition 4.18, one can show that ρα ω are identically zero for all except
finitely many α’s. The argument goes as follows: at each p ∈ supp ω, by condition (ii) in
Definition 4.18, there exists an open set Op ⊂ M containing p such that the set:

Sp := {α : supp ρα ∩ Op 6= ∅}

is finite. Evidently, we have

[
supp ω ⊂ Op
p∈supp ω

and by compactness of supp ω, there exists p1 , . . . , pN ∈ supp ω such that

N
[
supp ω ⊂ Opi .
i=1

Since {q ∈ M : ρα (q)ω(q) 6= 0} ⊂ {q ∈ M : ρα (q) 6= 0} ∩ {q ∈ M : ω(q) 6= 0}, we have:

supp (ρα ω) = {q ∈ M : ρα (q)ω(q) 6= 0}

⊂ {q ∈ M : ρα (q) 6= 0} ∩ {q ∈ M : ω(q) 6= 0}
⊂ {q ∈ M : ρα (q) 6= 0} ∩ {q ∈ M : ω(q) 6= 0}
N
[
= supp ρα ∩ supp ω ⊂ (supp ρα ∩ Opi ) .
i=1

Therefore, if α is an index such that supp (ρα ω) 6= ∅, then there exists i ∈ {1, . . . , N }
such that supp ρα ∩ Opi 6= ∅, or in other words, α ∈ Spi for some i, and so:

N
[
{α : supp (ρα ω) 6= ∅} ⊂ Spi .
i=1

Since each Spi is a finite set, the setZ{α : supp (ρα ω) 6= ∅} is also finite. Therefore, there
are only finitely many α’s such that is non-zero, and so the sum stated in (4.5) is
Fα (Uα )
in fact a finite sum.
Now we have understood that there is no convergence issue for (4.5) provided that
ω has compact support (which is automatically true if the manifold M is itself compact).
There are still two well-definedness issues to resolve, namely whether the integral in (4.5)
is independent of oriented atlas A, and for each atlas whether the integral is independent
of the choice of partitions of unity.
128 4. Generalized Stokes’ Theorem

Proposition 4.22. Let M n be an orientable smooth manifold with two oriented atlas
A = {Fα : Uα → M } and B = {Gβ : Vβ → M }
such that det D(Fα−1
◦ Gβ ) > 0 on the overlap for any pair of α and β. Suppose {ρα :
M → [0, 1]} and {σβ : M → [0, 1]} are partitions of unity subordinate to A and B
respectively. Then, given any compactly supported differential n-form ω on M n , we have:
XZ XZ
ρα ω = σβ ω.
all α Fα (Uα ) all β Gβ (Vβ )

X
Proof. By the fact that σβ ≡ 1 on M , we have:
all β
 
XZ XZ X XXZ
ρα ω =  σβ ρα ω =
 ρα σβ ω.
all α Fα (Uα ) all α Fα (Uα ) all β all α all β Fα (Uα )∩Gβ (Vβ )

The last equality follows from the fact that supp σβ ⊂ Gβ (Vβ ).
One can similarly work out that
XZ XXZ
σβ ω = ρα σβ ω.
all β Gβ (Vβ ) all β all α Fα (Uα )∩Gβ (Vβ )
P P
Note that α β is a finite double sum and so there is no issue of switching them. It
completes the proof. 

By Proposition 4.22, we justified that (4.5) is independent of oriented atlas and the
choice of partitions of unity. We can now define:

Definition 4.23. Let M n be an orientable smooth manifold with an oriented atlas

A = {Fα (u1α , . . . , unα ) : Uα → M } where (u1α , . . . , unα ) is the chosen order of local
coordinates. Pick a partition of unity {ρα : M → [0, 1]} subordinate to the atlas A.
Then, given any n-form ω, we define its integral over M as:
Z XZ
ω := ρα ω.
M all α Fα (Uα )

If ω = ϕα du1α ∧ ··· ∧ dunα on each Fα (Uα ), then:

Z XZ
ω= ρα ϕα du1α · · · dunα .
M all α Uα

Remark 4.24. It is generally impossible to compute such an integral, as we know only

the existence of ρα ’s but typically not the exact expressions. Even if such a partition of
2
unity ρα ’s can be found, it often involves some terms such as e−1/x , which is almost
impossible to integrate. To conclude, we do not attempt compute such an integral, but
we will study the properties of it based on the definition. 

4.3.3. Orientation of Manifolds. Partition of unity is a powerful tool to construct

a smooth global item from local ones. For integrals of differential forms, we first defines
integral of forms with support contained in a single parametrization chart, then we uses
a partition of unity to glue each chart together. There are some other uses in this spirit.
The following beautiful statement can be proved using partitions of unity:

Proposition 4.25. A smooth n-dimensional manifold M is orientable if and only if there

exists a non-vanishing smooth n-form globally defined on M .
4.3. Integrations of Differential Forms 129

Proof. Suppose M is orientable, then by definition there exists an oriented atlas A =

{Fα : Uα → M } such that det D(Fβ−1 ◦ Fα ) > 0 for any α and β. For each local
parametrization Fα , we denote (u1α , . . . , unα ) to be its local coordinates, then the n-form:
ηα := du1α ∧ · · · ∧ dunα
is locally defined on Fα (Uα ).
Let {ρα : M → [0, 1]} be a partition of unity subordinate to A. We define:
X X
ω= ρα η α = ρα du1α ∧ · · · ∧ dunα .
all α all α

We claim ω(p) 6= 0 at every point p ∈ M . Suppose p ∈ Fβ (Uβ ) for some β in the atlas. By
(3.19), for each α, locally near p we have:
∂(u1α , . . . , unα ) 1
du1α ∧ · · · ∧ dunα = det du ∧ · · · ∧ dunβ ,
∂(u1β , . . . , unβ ) β
and so: !
X ∂(u1α , . . . , unα )
ω= ρα det du1β ∧ · · · ∧ dunβ .
∂(u1β , . . . , unβ )
all α
X ∂(u1α , . . . , unα )
Since ρα ≥ 0, ρα ≡ 1 and det > 0, we must have:
∂(u1β , . . . , unβ )
all α
X ∂(u1α , . . . , unα )
ρα det >0 near p.
∂(u1β , . . . , unβ )
all α

This shows ω is a non-vanishing n-form on M .

Conversely, suppose Ω is a non-vanishing n-form on M . Let C = {Gα : Vα → M }
be any atlas on M , and for each α we denote (vα1 , . . . , vαn ) to be its local coordinates.
Express Ω in terms of local coordinates:
Ω = ϕα dvα1 ∧ · · · ∧ dvαn .
Since Ω is non-vanishing, ϕα must be either positive on Vα , or negative on Vα . Re-define
the local coordinates by:
(
1 2 n (vα1 , vα2 , . . . , vαn ) if ϕα > 0
(e
vα , veα , . . . , veα ) :=
(−vα1 , vα2 , . . . , vαn ) if ϕα < 0
Then, under these new local coordinates, we have:
vα1 ∧ · · · ∧ de
Ω = |ϕα | de vαn .
From (3.19), we can deduce:
vα1 , . . . , veαn ) 1
∂(e
vα1 ∧ · · · ∧ de
Ω = |ϕα | de vαn = |ϕα | det vβn
v ∧ · · · ∧ de
de
vβ1 , . . . , veβn ) β
∂(e
vα1 , . . . , veαn ) and (e
on the overlap of any two local coordinates (e vβ1 , . . . , veβn ). On the other
hand, we have:
vβ1 ∧ · · · ∧ de
Ω = |ϕβ | de vβn .
This shows:
vα1 , . . . , veαn )
∂(e ϕβ
det 1 n = >0 for any α, β.
∂(e
vβ , . . . , veβ ) ϕα
Therefore, M is orientable. 
130 4. Generalized Stokes’ Theorem

The significance of Proposition 4.25 is that it relates the orientability of an n-manifold

(which was defined in a rather local way) with the existence of a non-vanishing n-form
(which is a global object). For abstract manifolds, unit normal vectors cannot be defined.
Here the non-vanishing global n-form plays a similar role as a continuous unit normal
does for hypersurfaces. In the rest of the course we will call:

Definition 4.26 (Orientation of Manifolds). Given an orientable manifold M n , a non-

vanishing global n-form Ω is called an orientation of M . A basis of tangent vectors
{T1 , . . . , Tn } ∈ Tp M is said to be Ω-oriented if 
Ω(T1 , . . . , Tn ) > 0. A local coordinate
∂ ∂
system (u1 , . . . , un ) is said to be Ω-oriented if Ω ∂u1 , . . . , ∂un > 0.

Recall that when we integrate an n-form, we need to first pick an order of local
coordinates (u1 , . . . , un ), then express the n-form according to this order, and locally
define the integral as:
Z Z
ϕ du1 ∧ · · · ∧ dun = ϕ du1 · · · dun .
F (U ) U

Note that picking the order of coordinates is a local notion. To rephrase it using global
terms, we can first pick an orientation Ω (which is a global object on M ), then we
require the order of any local coordinates (u1 , . . . , un ) to be Ω-oriented. Any pair of
local coordinate systems (u1 , . . . , un ) and (v1 , . . . , vn ) which are both Ω-oriented will
∂(u1 , . . . , un )
automatically satisfy det > 0 on the overlap.
∂(v1 , . . . , vn )
To summarize, given an orientable manifold M n with a chosen orientation Ω, then
for any local coordinate system F (u1 , . . . , un ) : U → M , we define:
(R
ϕ du1 · · · dun if (u1 , . . . , un ) is Ω-oriented
Z
ϕ du1 ∧ · · · ∧ dun = UR
1 n
F (U ) − U ϕ du · · · du if (u1 , . . . , un ) is not Ω-oriented
or to put it in a more elegant (yet equivalent) way:
Z    Z
∂ ∂
ϕ du1 ∧ · · · ∧ dun = sgn Ω , ..., ϕ du1 · · · dun .
F (U ) ∂u 1 ∂u n U

Exercise 4.14. Let Ω := dx∧dy ∧dz be the orientation of R3 . Which of the following
is Ω-oriented?
(a) local coordiantes (x, y, z)
(b) vectors {î, k̂, ĵ}
(c) vectors {u, v, u × v} where u and v are linearly independent vectors in R3 .

Exercise 4.15. Consider three linearly independent vectors {u, v, w} in R3 such that
u ⊥ w and v ⊥ w. Show that {u, v, w} has the same orientation as {î, ĵ, k̂} if and
only if w = cu × v for some positive constant c.

Proposition 4.25 can be used to complete the proof that RP2 is not orientable in Ex-
ample 4.15. In that example, we demonstrated that there are two local parametrizations
F0 (u1 , u2 ) and F1 (v1 , v2 ) with the properties that:
• the domain of each of Fi is connected; while
• their overlap, i.e. domain of F0−1 ◦ F1 , is not connected; and
• det D(F0−1 ◦ F1 ) is positive on one component U , but negative on another compo-
nent V .
4.3. Integrations of Differential Forms 131

To show that RP2 is not orientable, we argue by contradiction that there exists a global
∂
non-vanishing 2-form Ω. Then, if Ω( ∂u , ∂ ) > 0, then one has Ω( ∂v∂ 1 , ∂v∂ 2 ) > 0 on U
1 ∂u2
since det D(F0 ◦ F1 ) > 0 on U , and Ω( ∂v∂ 1 , ∂v∂ 2 ) < 0 on V since det D(F0−1 ◦ F1 ) < 0.
−1

However, since the domain of F1 (v1 , v2 ) is connected and Ω( ∂v∂ 1 , ∂v∂ 2 ) is a smooth (in
particular continuous) function on that domain, there must be a point p in the domain
of F1 such that Ω( ∂v∂ 1 , ∂v∂ 2 ) = 0 at p. It leads to a contradiction that Ω is non-vanishing.
∂
Similar for the case Ω( ∂u , ∂ ) < 0.
1 ∂u2
132 4. Generalized Stokes’ Theorem

4.4. Generalized Stokes’ Theorem

In this section, we (finally) state and give a proof of an elegant theorem, Generalized
Stokes’ Theorem. It not only unifies Green’s, Stokes’ and Divergence Theorems which we
learned in Multivariable Calculus, but also generalize it to higher dimensional abstract
manifolds.

4.4.1. Boundary Orientation. Since the statement of Generalized Stokes’ Theorem

involves integration on differential forms, we will assume all manifolds discussed in this
section to be orientable. Let’s fix an orientation Ω of M n , which is a non-vanishing n-form,
and this orientation determines how local coordinates on M are ordered as discussed in
the previous section.
Now we deal with the orientation of the boundary manifold ∂M . Given a local
parametrization G(u1 , . . . , un ) : V ⊂ Rn+ → M of boundary type. The tangent space
 n
∂
Tp M for points p ∈ ∂M is defined as the span of . As V is a subset of the upper
∂ui i=1
half-space {un ≥ 0}, the vector η := − ∂u∂ n in Tp M is often called an outward-pointing
“normal” vector to ∂M .
An orientation Ω of M n is a non-vanishing n-form. The boundary manifold ∂M n is
an (n−1)-manifold, and so an orientation of ∂M n should be a non-vanishing (n−1)-form.
Using the outward-pointing normal vector η, one can produce such an (n − 1)-form in a
natural way. Given any tangent vectors T1 , . . . , Tn−1 on T (∂M ), we consider the interior
product iη Ω, which is defined as:

(iη Ω)(T1 , . . . , Tn−1 ) := Ω(η, T1 , . . . , Tn−1 ).

Then iη Ω is an alternating multilinear map in ∧n−1 T ∗ (∂M ).

Locally, given a local coordinate system (u1 , . . . , un ), by recalling that η = − ∂u∂ n we
can compute:
   
∂ ∂ ∂ ∂
(iη Ω) ,..., = Ω η, ,...,
∂u1 ∂un−1 ∂u1 ∂un−1
 
∂ ∂ ∂
=Ω − , ,...,
∂un ∂u1 ∂un−1
 
n ∂ ∂ ∂
= (−1) Ω ,..., ,
∂u1 ∂un−1 ∂un

which is non-zero. Therefore, iη Ω is a non-vanishing (n − 1)-form on ∂M , and we can

take it as an orientation for ∂M . From now on, whenever we pick an orientation Ω for
M n , we will by-default pick iη Ω to be the orientation for ∂M .
Given an Ω-oriented local coordinate system G(u1 , . . . , un ) : V → M of boundary
type for M n , then (u1 , . . . , un−1 ) is iη Ω-oriented if n is even; and is not iη Ω-oriented if n
is odd. Therefore, when integrating an (n − 1)-form ϕ du1 ∧ · · · ∧ dun−1 on ∂M , we need
to take into account of the parity of n, i.e.
Z Z
(4.6) ϕ du1 ∧ · · · ∧ dun−1 = (−1)n ϕ du1 · · · dun−1 .
G(V)∩∂M V∩{un =0}

The “extra” factor of (−1)n does not look nice at the first glance, but as we will
see later, it will make Generalized Stokes’ Theorem nicer. We are now ready to state
Generalized Stokes’ Theorem in a precise way:
4.4. Generalized Stokes’ Theorem 133

Theorem 4.27 (Generalized Stokes’ Theorem). Let M be an orientable smooth n-

manifold, and let ω be a compactly supported smooth (n − 1)-form on M . Then, we
have:
Z Z
(4.7) dω = ω.
M ∂M
Here if Ω is a chosen to be an orientation of M , then we will take iη Ω to be the orientation
of ∂M where η is an outward-point normal vector of ∂M .
Z
In particular, if ∂M = ∅, then dω = 0.
M

4.4.2. Proof of Generalized Stokes’ Theorem. The proof consists of three steps:
Step 1: a special case where supp ω is contained inside a single parametrization chart
of interior type;
Step 2: another special case where supp ω is contained inside a single parametrization
chart of boundary type;
Step 3: use partitions of unity to deduce the general case.

Proof of Theorem 4.27. Throughout the proof, we will let Ω be the orientation of M ,
and iη Ω be the orientation of ∂M with η being an outward-point normal vector to ∂M .
All local coordinate system (u1 , . . . , un ) of M is assumed to be Ω-oriented.
Step 1: Suppose supp ω is contained in a single parametrization chart of interior type.
Let F (u1 , . . . , un ) : U ⊂ Rn → M be a local parametrization of interior type such
that supp ω ⊂ F (U). Denote:

du1 ∧ · · · ∧ du
ci ∧ · · · ∧ dun := du1 ∧ · · · ∧ dui−1 ∧ dui+1 ∧ · · · ∧ dun ,

or in other words, it means the form with dui removed.

In terms of local coordinates, the (n − 1)-form ω can be expressed as:
n
X
ω= ωi du1 ∧ · · · ∧ du
ci ∧ · · · ∧ dun .
i=1

Taking the exterior derivative, we get:

n X n
X ∂ωi j
dω = du ∧ du1 ∧ · · · ∧ du
ci ∧ · · · ∧ dun
i=1 j=1
∂uj

For each i, the wedge product duj ∧ du1 ∧ · · · ∧ du

ci ∧ · · · ∧ dun is zero if j 6= i. Therefore,
n
X ∂ωi
dω = dui ∧ du1 ∧ · · · ∧ du
ci ∧ · · · ∧ dun
i=1
∂ui
n
X ∂ωi 1
= (−1)i−1 du ∧ · · · ∧ dui ∧ · · · ∧ dun
i=1
∂ui

By definition of integrals of differential forms, we get:

Z Z X n
∂ωi 1
dω = (−1)i−1 du · · · dun .
M U i=1 ∂ui

Since supp ω ⊂ F (U), the functions ωi ’s are identically zero near and outside the bound-
ary of U ⊂ Rn . Therefore, we can replace the domain of integration U of the RHS integral
134 4. Generalized Stokes’ Theorem

by a rectangle [−R, R] × · · · × [−R, R] in Rn where R > 0 is a sufficiently large number.

The value of the integral is unchanged. Therefore, using the Fubini’s Theorem, we get:
Z Z R Z R X n
∂ωi 1
dω = ··· (−1)i du · · · dun
M −R −R i=1 ∂ui
n Z R Z R
X ∂ωi i 1
= (−1)i−1 ··· ci · · · dun
du du · · · du
i=1 −R −R ∂u i

X n Z R Z R
u =R
= (−1)i−1 ··· [ωi ]uii =−R du1 · · · du
ci · · · dun .
i=1 −R −R

n
Since ωi ’s vanish at the boundary of
Z the rectangle [−R, R] , we have ωi = 0 when
ui = ±R. As a result, we proved dω = 0. Since supp ω is contained in a single
M
parametrization
Z chart of interior type, we have ω = 0 on the boundary ∂M . Evidently,
we have ω = 0 in this case. Hence, we proved
∂M
Z Z
dω = ω=0
M ∂M

in this case.
Step 2: Suppose supp ω is contained inside a single parametrization chart of boundary type.
Let G(u1 , . . . , un ) : V ⊂ Rn+ → M be a local parametrization of boundary type such
that supp ω ⊂ G(V). As in Step 1, we express
n
X
ω= ωi du1 ∧ · · · ∧ du
ci ∧ · · · ∧ dun .
i=1

Proceed exactly in the same way as before, we arrive at:

Z Z X n
∂ωi 1
dω = (−1)i−1 du · · · dun .
M V i=1 ∂ui

Now V is an open set in Rn+ instead of Rn . Recall that the boundary is the set of
points with un = 0. Therefore, this time we replace V by the half-space rectangle
[−R, R] × · · · × [−R, R] × [0, R] where R > 0 again is a sufficiently large number.
One key difference from Step 1 is that even though ωi ’s has compact support in-
side V, it may not vanish on the boundary of M . Therefore, we can only guarantee
ωi (u1 , . . . , un ) = 0 when un = R, but we cannot claim ωi = 0 when un = 0. Some more
work needs to be done:
Z Z X n
∂ωi 1
dω = (−1)i−1 du · · · dun
M V i=1 ∂ui
Z RZ R Z R X n
∂ωi 1
= ··· (−1)i−1 du · · · dun
0 −R −R i=1 ∂ui
n−1 Z RZ R Z R
X ∂ωi 1
= (−1)i−1 ··· (−1)i−1 du · · · dun
i=1 0 −R −R ∂u i
Z RZ R Z R
∂ωn 1
+ (−1)n−1 ··· du · · · dun
0 −R −R ∂un
4.4. Generalized Stokes’ Theorem 135

One can proceed as in Step 1 to show that the first term:

n−1 Z RZ R Z R
X ∂ωi 1
(−1)i−1 ··· (−1)i−1 du · · · dun = 0,
i=1 0 −R −R ∂ui

which follows from the fact that whenever 1 ≤ i ≤ n − 1, we have ωi = 0 on ui = ±R.

The second term:
Z R Z R Z R
n−1 ∂ωn 1
(−1) ··· du · · · dun
0 −R −R ∂un
is handled in a different way:
Z RZ R Z R
∂ωn 1
(−1)n−1 ··· du · · · dun
0 −R −R ∂un
Z R Z R Z R
∂ωn n 1
= (−1)n−1 ··· du du · · · dun−1
−R −R 0 ∂un
Z R Z R
u =R
= (−1)n−1 ··· [ωn ]unn =0 du1 · · · dun−1
−R −R
Z R Z R
= (−1)n ··· ωn (u1 , . . . , un−1 , 0) du1 · · · dun−1
−R −R
where we have used the following fact:
u =R
[ωn (u1 , . . . , un )]unn =0 = ωn (u1 , . . . , un−1 , R) − ωi (u1 , . . . , un−1 , 0)
= 0 − ωn (u1 , . . . , un−1 , 0).
Combining all results proved so far, we have:
Z Z R Z R
dω = (−1)n ··· ωn (u1 , . . . , un−1 , 0) du1 · · · dun−1
M −R −R
Z Z
On the other hand, we compute ω and then compare it with dω. Note that
∂M M
the boundary ∂M are points with un = 0. Therefore, across the boundary ∂M , we have
dun ≡ 0, and so on ∂M we have:
Xn
ω= ωi (u1 , . . . , un−1 , 0) du1 ∧ · · · ∧ du
ci ∧ · · · ∧ dun
|{z}
i=1 =0
= ωn (u1 , . . . , un−1 , 0) du1 ∧ · · · ∧ dun−1
Z Z
ω= ωn (u1 , . . . , un−1 , 0) du1 ∧ · · · ∧ dun−1
∂M G(V)∩∂M
Z
= (−1)n ωn (u1 , . . . , un−1 , 0) du1 · · · dun−1
V∩{un =0}
Z R Z R
= (−1)n ··· ωn (u1 , . . . , un−1 , 0) du1 · · · dun−1
−R −R
Recall that we have a factor of (−1)n because the local coordinate system (u1 , . . . , un−1 )
for ∂M is iη Ω if and only if n is even, as discussed in the previous subsection.
Consequently, we have proved
Z Z
dω = ω
M ∂M
in this case.
136 4. Generalized Stokes’ Theorem

Step 3: Use partitions of unity to deduce the general case

Finally, we “glue” the previous two steps together and deduce the general case.
Let A = {Fα : Uα → M } be an atlas of M where all local coordinates are Ω-oriented.
Here A contain both interior and boundary types of local parametrizations. Suppose
{ρα : M → [0, 1]} is a partition of unity subordinate to A. Then, we have:

!
X X
ω= ρα ω= ρα ω
α α
| {z }
≡1
Z Z X XZ
ω= ρα ω = ρα ω.
∂M ∂M α α ∂M

For each α, the (n − 1)-form ρα ω is compactly supported in a single parametrization

chart (either of interior or boundary type). From Step 1 and Step 2, we have already
proved that Generalized Stokes’ Theorem is true for each ρα ω. Therefore, we have:

XZ XZ
ρα ω = d(ρα ω)
α ∂M α M
XZ
= (dρα ∧ ω + ρα dω)
α M
Z ! !
X X
= d ρα ∧ω+ ρα dω.
M α α

!
X X
Since ρα ≡ 1 and hence d ρα ≡ 0, we have proved:
α α

Z XZ Z Z
ω= ρα ω = 0 ∧ ω + 1 dω = dω.
∂M α ∂M M M

It completes the proof of Generalized Stokes’ Theorem. 

Remark 4.28. As we can see from that the proof (Step 2), if we simply choose an
orientation for ∂M such that (u1 , . . . , un−1 ) becomes the order of local coordinates for
∂M , then (4.7) would have a factor of (−1)n on the RHS, which does not look nice.
Moreover, if we pick i−η Ω to be the orientation of ∂M (here −η is then an inward-
pointing normal to ∂M ), then the RHS of (4.7) would have a minus sign, which is not
nice either. 

4.4.3. Fundamental Theorems in Vector Calculus. We briefly discussed at the

end of Chapter 3 how the three fundamental theorems in Vector Calculus, namely Green’s,
Stokes’ and Divergence Theorems, can be formulated using differential forms. Given that
we now have proved Generalized Stokes’ Theorem (Theorem 4.27), we are going to give
a formal proof of the three Vector Calculus theorems in MATH 2023 using the Theorem
4.27.
4.4. Generalized Stokes’ Theorem 137

Corollary 4.29 (Green’s Theorem). Let R be a closed and bounded smooth 2-submanifold
in R2 with boundary ∂R. Given any smooth vector field V = (P (x, y), Q(x, y)) defined in
R, then we have: I Z  
∂Q ∂P
V · dl = − dx dy,
∂R R ∂x ∂y
∂ ∂
The line integral on the LHS is oriented such that { ∂x , ∂y } has the same orientation as
{η, T } where η is the outward-pointing normal of R, and T is the velocity vector of the
curve ∂R. See Figure 4.3.

Proof. Consider the 1-form ω := P dx + Q dy defined on R, then we have:

 
∂Q ∂P
dω = − dx ∧ dy.
∂x ∂y
Suppose we fix an orientation Ω = dx ∧ dy for R so that the order of coordinates is (x, y),
then by generalized Stokes’ Theorem we get:
I Z   Z  
∂Q ∂P ∂Q ∂P
P dx + Q dy = − dx ∧ dy = − dx dy.
∂x ∂y ∂x ∂y
| ∂R H {z } |R {z } | R
{z }
R
∂R
ω R
dω (x,y) is the orientation

The only thing left to figure out is the orientation of the line integral. Locally parametrize
R by local coordinates (s, t) so that {t = 0} is the boundary ∂R and {t > 0} is the interior
of R (see Figure 4.3). By convention, the local coordinate s for ∂R must be chosen so
∂
that Ω(η, ∂s ) > 0 where η is a outward-pointing normal vector to ∂R. In other words,
∂ ∂ ∂
the pair {η, ∂s } should have the same orientation as { ∂x , ∂y }. According to Figure 4.3,
we must choose the local coordinate s for ∂R such that for the outer boundary, s goes
counter-clockwisely as it increases; whereas for each inner boundary, s goes clockwisely
as it increases. 

Figure 4.3. Orientation of Green’s Theorem

Next we show that Stokes’ Theorem in Multivariable Calculus is also a consequence

of Generalized Stokes’ Theorem. Recall that in MATH 2023 we learned about surface
integrals. If F (u, v) : U → Σ ⊂ R3 is a parametrization of the whole surface Σ, then we
138 4. Generalized Stokes’ Theorem

define the surface element as:

∂F ∂F
dS = × du dv,
∂u ∂v

and the surface integral of a scalar function ϕ is defined as:

Z Z
∂F ∂F
ϕ dS = ϕ(u, v) × du dv.
Σ U ∂u ∂v

However, not every surface can be covered (or almost covered) by a single parametriza-
tion chart. Generally, if A = {Fα (uα , vα ) : Uα → R3 } is an oriented atlas of Σ with a
partition of unity {ρα : Σ → [0, 1]} subordinate to A, we then define:
X ∂Fα ∂Fα
dS := ρα × duα dvα .
α
∂uα ∂vα

Corollary 4.30 (Stokes’ Theorem). Let Σ be a closed and bounded smooth 2-submanifold
in R3 with boundary ∂Σ, and V = (P (x, y, z), Q(x, y, z), R(x, y, z)) be a vector field
which is smooth on Σ, then we have:
I Z
V · dl = (∇ × V ) · ν dS.
∂Σ Σ

Here {î, ĵ, k̂} has the same orientation as {η, T, ν}, where η is the outward-point normal
vector of Σ at points of ∂Σ, T is the velocity vector of ∂Σ, and ν is the unit normal vector
to Σ in R3 . See Figure 4.4.

Proof. Define:
ω = P dx + Q dy + R dz
which is viewed as a 1-form on Σ. Then,
I I
(4.8) ω= V · dl.
∂Σ ∂Σ

By direct computation, the 2-form dω is given by:

     
∂Q ∂P ∂P ∂R ∂R ∂Q
dω = − dx ∧ dy + − dz ∧ dx + − dy ∧ dz.
∂x ∂y ∂z ∂x ∂y ∂z

Now consider an oriented atlas A = {Fα (uα , vα ) : Uα → R3 } of Σ with a partition of

unity {ρα : Σ → [0, 1]}, then according to the discussion near the end of Chapter 3, we
can express each of dx ∧ dy, dz ∧ dx and dy ∧ dz in terms of duα ∧ dvα , and obtain:
X
dω = ρα dω
α
      
X ∂Q ∂P ∂P ∂R ∂R ∂Q
= ρα − dx ∧ dy + − dz ∧ dx + − dy ∧ dz
α
∂x ∂y ∂z ∂x ∂y ∂z
   
X ∂Q ∂P ∂(y, z) ∂P ∂R ∂(z, x)
= ρα − det + − det
α
∂x ∂y ∂(uα , vα ) ∂z ∂x ∂(uα , vα )
  
∂R ∂Q ∂(x, y)
+ − det duα ∧ dvα .
∂y ∂z ∂(uα , vα )
4.4. Generalized Stokes’ Theorem 139

On each local coordinate chart Fα (Uα ), a normal vector to Σ in R3 can be found using
cross products:
∂Fα ∂Fα ∂(y, z) ∂(z, x) ∂(x, y)
× = det î + det ĵ + det k̂
∂uα ∂vα ∂(uα , vα ) ∂(uα , vα ) ∂(uα , vα )
     
∂Q ∂P ∂P ∂R ∂R ∂Q
∇×V = − î + − ĵ + − k̂.
∂x ∂y ∂z ∂x ∂y ∂z
Hence,  
X ∂Fα ∂Fα
dω = (∇ × V ) · × ρα duα ∧ dvα ,
α
∂uα ∂vα
and so Z  
XZ ∂Fα ∂Fα
dω = (∇ × V ) · × ρα duα dvα .
Σ α Uα ∂uα ∂vα
∂Fα ∂Fα
∂uα × ∂vα
X ∂Fα ∂Fα
Denote ν = , and recall the fact that dS := ρα × duα dvα ,
∂F
× ∂F
α
∂uα ∂vα
∂uα ∂vα
we get:
Z Z
(4.9) dω = (∇ × V ) · ν dS.
Σ Σ
Combining the results of (4.8) and (4.9), using Generalized Stokes’ Theorem (Theorem
4.7, we get: I Z
V · dl = (∇ × V ) · ν dS
∂Σ Σ
as desired.
To see the orientation of ∂Σ, we locally parametrize Σ by coordinates (s, t) such that
{t = 0} are points on ∂Σ, and so ∂Σ is locally parametrized by s. The outward-pointing
∂
 ∂
normal of is given by η := − ∂t
n ∂Σ in Σ o . By convention, the orientation of η, ∂s is the
∂ ∂
same as ∂uα , ∂vα , and hence:
   
∂ ∂ ∂
η, , ν has the same orientation as , ,ν .
∂s ∂uα ∂vα
∂Fα ∂Fα
∂uα × ∂vα
n
∂ ∂
o
As ν = , the set ∂uα , ∂vα , ν has the same orientation as {î, ĵ, k̂}. As
∂Fα ∂Fα
∂uα × ∂vα
∂
a result, the set {η, ∂s , ν} is oriented in the way as in Figure 4.4. 

Finally, we discuss how to use Generalized Stokes’ Theorem to prove Divergence

Theorem in Multivariable Calculus.

Corollary 4.31 (Divergence Theorem). Let D be a closed and bounded 3-submanifold

of R3 with boundary ∂D, and V = (P (x, y, z), Q(x, y, z), R(x, y, z)) be a smooth vector
field defined on D. Then, we have:
I Z
V · ν dS = ∇ · V dx dy dz.
∂D D
Here ν is the unit normal vector of ∂D in R3 which points away from D.

Proof. Let ω := P dy ∧ dz + Q dz ∧ dx + R dx ∧ dy. Then by direct computations, we get:

 
∂P ∂Q ∂R
dω = + + dx ∧ dy ∧ dz = ∇ · V dx ∧ dy ∧ dz.
∂x ∂y ∂z
140 4. Generalized Stokes’ Theorem

Figure 4.4. Orientation of Stokes’ Theorem

Using {î, ĵ, k̂} as the orientation for D, then it is clear that:
Z Z
(4.10) dω = ∇ · V dx dy dz.
D D
Consider an atlas A = {Fα (uα , vα , wα ) : Uα → R3 } of D such that for the local
parametrization of boundary type, the boundary points are given by {wα = 0}, and
interior points are {wα > 0}. Then, ∂D is locally parametrized by (uα , vα ).
∂
As a convention, the orientation of (uα , vα ) is chosen such that {− ∂w , ∂ , ∂ }
α ∂uα ∂vα

has the same orientation as {î, ĵ, k̂}, or equivalently, { ∂u∂α , ∂v∂α , − ∂w
∂
α
} has the same
orientation as {î, ĵ, k̂}.
∂Fα ∂Fα
∂uα × ∂vα
Furthermore, let ν be the unit normal of ∂D given by ν = . By the
∂Fα ∂Fα
∂uα × ∂vα
convention of cross products, { ∂F α ∂Fα
∂uα , ∂vα , ν} must have the same orientation as {î, ĵ, k̂}.
∂ ∂ ∂ ∂Fα ∂Fα ∂
Now that { ∂uα , ∂vα , − ∂wα } and { ∂uα , ∂vα , ν} have the same orientation, so ν and − ∂w α
are both pointing in the same direction. In other words, ν is the outward-point normal.
The rest of the proof goes by writing ω in terms of duα ∧ dvα on each local coordinate
chart:
X
ω= ρα ω
α
 
X ∂(y, z) ∂(z, x) ∂(x, y)
= ρα P det + Q det + R det duα ∧ dvα
α
∂(uα , vα ) ∂(uα , vα ) ∂(uα , vα )
 
X ∂Fα ∂Fα
= V · × ρα duα ∧ dvα
α
∂uα ∂vα
X ∂Fα ∂Fα
= V · ν ρα × duα ∧ dvα
α
∂uα ∂vα
Therefore, we get:
I I I
X ∂Fα ∂Fα
(4.11) ω= V · ν ρα × duα dvα = V · ν dS.
∂D ∂D α
∂uα ∂vα ∂D

Combining with (4.10), (4.11) and Generalized Stokes’ Theorem, the proof of this
corollary is completed. 
 Chapter 5

De Rham Cohomology

“ "Should you just be an algebraist or a

geometer?" is like saying "Would you
rather be deaf or blind?" ”

Michael Atiyah

In Chapter 3, we discussed closed and exact forms. As a reminder, a smooth k-form

ω on a smooth manifold M is closed if dω = 0 on M , and is exact if ω = dη for some
smooth (k − 1)-form η defined on the whole M .
By the fact that d2 = 0, an exact form must be closed. It is then natural to ask whether
every closed form is exact. The answer is no in general. Here is a counter-example. Let
M = R2 \{(0, 0)}, and define
y x
ω := − 2 2
dx + 2 dy.
x +y x + y2
It can be computed easily that dω = 0 on M , and so ω is closed.
However, we can show that ω is not exact. Consider the unit circle C parametrized
by (x, y) = (cos t, sin t) where 0 < t < 2π, and also the induced 1-form ι∗ ω (where
ι : C → M ). By direct computation, we get:
I Z 2π
sin t cos t
ι∗ ω = − 2 2 d(cos t) + 2 t + sin2 t
d(sin t) = 2π.
C 0 cos t + sin t cos
If ω were exact, then ω = df for some smooth function f : M → R. Then, we would
have: Z 2π
d(ι∗ f )
I I I
∗ ∗ ∗
ι ω= ι (df ) = d(ι f ) = dt.
C C C 0 dt
Since t = 0 and t = 2π represent the same point on C, by Fundamental Theorem of
Calculus, we finally get: I
ι∗ ω = 0
C
which is a contradiction! Therefore, ω is not exact on R2 \{(0, 0)}.
Heuristically, de Rham cohomology studies “how many” smooth k-forms defined on
a given manifold M are closed but not exact. We should refine the meaning of “how
many”. Certainly, if η is any (k − 1)-form on M , then ω + dη is also closed but not
exact. Therefore, when we “count” how many smooth k-forms on M which are closed

141
142 5. De Rham Cohomology

but not exact, it is fair to group ω and ω + dη’s together, and count them as one. In
formal mathematical language, equivalence classes are used as we will discuss in detail.
It turns out that the “number” of closed, not exact k-forms on a given M is a related to
the topology of M !
In this chapter, we will learn the basics of de Rham cohomology, which is a beautiful
topic to end the course MATH 4033.

5.1. De Rham Cohomology

Let M be a smooth manifold (with or without boundary). Recall that the exterior
derivative d is a linear map that takes a k-form to a (k + 1)-form, i.e. d : ∧k T ∗ M →
∧k+1 T ∗ M . We can then talk about the kernel and image of these maps. We define:
ker d : ∧k T ∗ M → ∧k+1 T ∗ M = {ω ∈ ∧k T ∗ M : dω = 0}

= {closed k-forms on M }
k−1
T M → ∧ T M = {ω ∈ ∧k T ∗ M : ω = dη for some η ∈ ∧k−1 T ∗ M }
∗ k ∗


Im d : ∧
= {exact k-forms on M }
In many occasions, we may simply denote the above kernel and image by ker(d) and
Im (d) whenever the value of k is clear from the context.
By d2 = 0, it is easy to see that:
Im d : ∧k−1 T ∗ M → ∧k T ∗ M ⊂ ker d : ∧k T ∗ M → ∧k+1 T ∗ M .

If all closed k-forms on a certain manifold are exact, then we have Im (d) = ker(d). How
“many” closed k-forms are exact is then measured by how Im (d) is “smaller” than ker(d),
which is precisely measured by the size of the quotient vector space ker(d)/Im (d). We
call this quotient the de Rham cohomology group1.

Definition 5.1 (de Rham Cohomology Group). Let M be a smooth manifold. For any
positive integer k, we define the k-th de Rham cohomology group of M to be the quotient
vector space:
ker d : ∧k T ∗ M → ∧k+1 T ∗ M


k
HdR (M ) := .
Im (d : ∧k−1 T ∗ M → ∧k T ∗ M )

Remark 5.2. When k = 0, then ∧k T ∗ M = ∧0 T ∗ M = C ∞ (M, R) and ∧k−1 T ∗ M is not

defined. Instead, we define
0
(M ) := ker d : C ∞ (M, R) → ∧1 T ∗ M = {f ∈ C ∞ (M, R) : df = 0},


HdR
which is the vector space of all locally constant functions on M . If M has N connected
components, then a locally constant function f is determined by its value on each of
the components. The space of functions {f : df = 0} is in one-to-one correspondence
an N -tuple (k1 , . . . , kN ) ∈ RN , where ki is the value of f on the i-th component of M .
0
Therefore, HdR (M ) ' RN where N is the number of connected components of M . 

5.1.1. Quotient Vector Spaces. Let’s first review the basics about quotient vector
spaces in Linear Algebra. Given a subspace W of a vector space V , we can define an
equivalence relation ∼ by declaring that v1 ∼ v2 if and only if v1 − v2 ∈ W . For example,
if W is the x-axis and V is the xy-plane, then two vector v1 and v2 are equivalent under
this relation if and only if they have the same ĵ-component.

1A vector space is also a group whose addition is the vector addition. Although it is more appropriate or precise to call the
quotient the “de Rham cohomology space”, we will follow the history to call it a group.
5.1. De Rham Cohomology 143

For each element v ∈ V (the bigger space), one can define an equivalence class:
[v] := {u ∈ V : u ∼ v} = {u ∈ V : u − v ∈ W }
which is the set of all vectors in V that are equivalent to v. For example, if W is the
x-axis and V is R2 , then the class [(2, 3)] is given by:
[(2, 3)] = {(x, 3) : x ∈ R}
which is the horizontal line {y = 3}. Similarly, one can figure out [(1, 3)] = [(2, 3)] =
[(3, 3)] = . . . as well, but [(2, 3)] 6= [(2, 2)], and the latter is the line {y = 2}.
The quotient space V /W is defined to be the set of all equivalence classes, i.e.
V /W := {[v] : v ∈ V }.
2
For example, if V is R and W is the x-axis, then V /W is the set of all horizontal lines in
R2 . For finite dimensional vector spaces, one can show (see Exercise 5.1) that
dim(V /W ) = dim V − dim W,
and so the “size” (precisely, the dimension) of the quotient V /W measures how small
W is when compared to V . In fact, if the bases of V and W are suitably chosen, we can
describe the basis of V /W in a precise way (see Exercise 5.1).

Exercise 5.1. Let W be a subspace of a finite dimensional vector space V . Suppose

{w1 , . . . , wk } is a basis for W , and {w1 , . . . , wk , v1 , . . . , vl } is a basis for V (Remark:
given any basis {w1 , . . . , wk } for the subspace W , one can always complete it to
form a basis for V ).
Pk Pl
(a) Show that given any vector i=1 αi wi + j=1 βj vj ∈ V , the equivalence class
represented by this vector is given by:
     
X k Xl X k Xl  Xl
 αi wi + βj vj  = γi wi + βj vj : γ i ∈ R =  β j vj  .
 
i=1 j=1 i=1 j=1 j=1

(b) Hence, show that {[v1 ], . . . , [vl ]} is a basis for V /W , and so

dim V /W = l = dim V − dim W.

Exercise 5.2. Given a subspace W of a vector space V , and define an equivalence

relation ∼ by declaring that v1 ∼ v2 if and only if v1 − v2 ∈ W . Show that the
following are equivalent:
(1) u ∈ [v]
(2) u − v ∈ W
(3) [u] = [v]

5.1.2. Cohomology Classes and Betti numbers. Recall that the k-th de Rham
k
cohomology group HdR (M ), where k ≥ 1, of a smooth manifold M is defined to be the
quotient vector space:
ker d : ∧k T ∗ M → ∧k+1 T ∗ M


k
HdR (M ) := .
Im (d : ∧k−1 T ∗ M → ∧k T ∗ M )
Given a closed k-form ω, we then define its equivalence class to be:
[ω] := {ω 0 : ω 0 − ω is exact}
= {ω 0 : ω 0 = ω + dη for some η ∈ ∧k−1 T ∗ M }
= {ω + dη : η ∈ ∧k−1 T ∗ M }.
144 5. De Rham Cohomology

An equivalence class [ω] is called the de Rham cohomology class represented by (or
containing) ω, and ω is said to be a representative of this de Rham cohomology class.
By Exercise 5.1, its dimension is given by
k
dim HdR (M )
= dim ker d : ∧k T ∗ M → ∧k+1 T ∗ M − dim Im d : ∧k−1 T ∗ M → ∧k T ∗ M

provided that both kernel and image are finite-dimensional.

k
Therefore, the dimension of HdR (M ) is a measure of “how many” closed k-forms on
M are not exact. Due to the importance of this dimension, we have a special name for it:

Definition 5.3 (Betti Numbers). Let M be a smooth manifold. The k-th Betti number
of M is defined to be:
k
bk (M ) := dim HdR (M ).

0
In particular, b0 (M ) = dim HdR (M ) is the number of connected components of M .
In case when M = R2 \{(0, 0)}, we discussed that there is a closed 1-form
−y dx + x dy
ω=
x2 + y 2
defined on M which is not
exact. Therefore, ω ∈ ker d : ∧1 T ∗ M → ∧2 T ∗ M yet ω 6∈

Im d : ∧0 T ∗ M → ∧1 T ∗ M , and so in HdR 1
(M ) we have [ω] 6= [0]. From here we can
1
conclude that HdR (M ) 6= {[0]} and b1 (M ) ≥ 1. We will later show that in fact b1 (M ) = 1
using some tools in later sections.
Exercise 5.3. If k > dim M , what can you say about bk (M )?

5.1.3. Poincaré Lemma. A star-shaped open set U in Rn is a region containing a

point p ∈ U (call it a base point) such that any line segment connecting a point x ∈ U
and the base point p must be contained inside U . Examples of star-shaped open sets
include convex open sets such an open ball {x ∈ Rn : |x| < 1}, and all of Rn . The
1
following Poincaré Lemma asserts that HdR (U ) = {[0]}.

1
Theorem 5.4 (Poincaré Lemma for HdR ). For any star-shaped open set U in Rn , we have
1
HdR (U ) = {[0]}. In other words, any closed 1-form defined on a star-shaped open set is
exact on that open set.

ωi dxi , we need to find a

P
Proof. Given a closed 1-form ω defined on U , given by ω = i
∂f
smooth function f : U → R such that ω = df . In other words, we need = ωi for any
∂xi
i.
Let p be the base point of U , then given any x ∈ U , we define:
Z
f (x) := ω
Lx

where Lx is the line segment joining p and x, which can be parametrized by:
γ(t) = (1 − t)p + tx, t ∈ [0, 1].
Write p = (p1 , . . . , pn ), x = (x1 , . . . , xn ), then f (x) can be expressed in terms of t by:
Z 1X n
f (x) = ωi (γ(t)) · (xi − pi ) dt.
0 i=1
5.1. De Rham Cohomology 145

Using the chain rule, we can directly verify that:

Z 1X n
∂f ∂
(x) = ωi (γ(t)) · (xi − pi ) dt
∂xj ∂xj 0 i=1
n Z 1 
X ∂ ∂
= ωi (γ(t)) · (xi − pi ) + ωi (γ(t)) · (xi − pi ) dt
i=1 0
∂xj ∂xj
xk ◦γ(t)
 
n Z 1 X n
z }| {
X ∂ωi ∂ ((1 − t)pk + txk ) 
= · (xi − pi ) + ωi (γ(t)) · δij  dt
 
∂x ∂x

i=1 0
 k j 
k=1

n Z n
!
1
X X ∂ωi
= t δjk · (xi − pi ) + ωj (γ(t)) dt
i=1 0
∂xk
k=1
n Z 1 
X ∂ωi
= t · (xi − pi ) + ωj (γ(t)) dt
i=1 0
∂xj

Since ω is closed, we have:

n  
X ∂ωi ∂ωj
0 = dω = − dxj ∧ dxi
i<j
∂xj ∂xi

∂ωi ∂ωj
and hence = for any i, j. Using this to proceed our calculation:
∂xj ∂xi
Z 1 
∂f ∂ωj
(x) = t · (xi − pi ) + ωj (γ(t)) dt
∂xj 0 ∂xi
Z 1
d
= (tωj (γ(t))) dt
0 dt
t=1
= [tωj (γ(t))]t=0 = ωj (γ(1)) = ωj (x).
In the second equality above, we have used the chain rule backward:
d ∂ωj
(tωj (γ(t))) = t · (xi − pi ) + ωj (γ(t)).
dt ∂xi
1
From this, we conclude that ω = df on U , and hence [ω] = [0] in HdR (U ). Since ω is
1
an arbitrary closed 1-form on U , we have HdR (U ) = {[0]}. 
k
Remark 5.5. Poincaré Lemma also holds for HdR , meaning that if U is a star-shaped
n k
open set in R , then HdR (U ) = {[0]} for any k ≥ 1. However, the proof involves the use
of Lie derivatives and a formula by Cartan, both of which are beyond the scope of this
0
course. Note also that HdR (U ) ' R since a star-shaped open set must be connected. 
Remark 5.6. We have discussed that the 1-form
−y dx + x dy
ω=
x2 + y 2
is closed but not exact. To be precise, it is not exact on R2 \{(0, 0)}. However, if we
regard the domain to be the first quadrant U := {(x, y) : x > 0 and y > 0}, which is a
star-shaped open set in R2 , then by Poinaré Lemma (Theorem 5.4), ω is indeed an exact
1-form on U . In fact, it is not difficult to verify that
 y
ω = d tan−1 on U .
x
146 5. De Rham Cohomology

Note that the scalar function tan−1 xy is smoothly defined on U . Whether a form is exact
or not depends on the choice of its domain! 

5.1.4. Diffeomorphic Invariance. By Proposition 3.57, we learned that the exte-

rior derivative d commutes with the pull-back of a smooth map between two manifolds.
An important consequence is that the de Rham cohomology group is invariant under
diffeomorphism.
Let Φ : M → N be any smooth map between two smooth manifolds. The pull-back
map Φ∗ : ∧k T ∗ N → ∧k T ∗ M induces a well-defined pull-back map (which is also denoted
by Φ∗ ) from HdRk k
(N ) to HdR (M ). Precisely, given any closed k-form ω on N , we define:
Φ∗ [ω] := [Φ∗ ω].
Φ∗ ω is a k-form on M . It is closed since d(Φ∗ ω) = Φ∗ (dω) = Φ∗ (0) = 0. To show it is
well-defined, we take another k-form ω 0 on N such that [ω 0 ] = [ω] in HdRk
(N ). Then,
there exists a (k − 1)-form η on N such that:
ω 0 − ω = dη on N .
∗ ∗
Using again d ◦ Φ = Φ ◦ d, we get:
Φ∗ ω 0 − Φ∗ ω = Φ∗ (dη) = d(Φ∗ η) on M
We conclude Φ∗ ω 0 − Φ∗ ω is exact and so
[Φ∗ ω 0 ] = [Φ∗ ω] in HdR
k
(M ).
This shows Φ∗ : HdR
k k
(N ) → HdR (M ) is a well-defined map.

k
Theorem 5.7 (Diffeomorphism Invariance of HdR ). If two smooth manifolds M and N
k k
are diffeomorphic, then HdR (M ) and HdR (N ) are isomorphic for any k ≥ 0.

Proof. Let Φ : M → N be a diffeomorphism, then Φ−1 : N → M exists and we have

Φ ◦ Φ−1 = idN and Φ−1 ◦ Φ = idM . By the chain rule for tensors (Theorem 3.54), we
have:
(Φ−1 )∗ ◦ Φ∗ = id∧k T ∗ N and Φ∗ ◦ (Φ−1 )∗ = id∧k T ∗ M .
k
Given any closed k-form ω on M , then in HdR (M ) we have:
Φ∗ ◦ (Φ−1 )∗ [ω] = Φ∗ [(Φ−1 )∗ ω] = [Φ∗ ◦ (Φ−1 )∗ ω] = [ω].
In other words, Φ∗ ◦ (Φ−1 )∗ is also the identity map of HdR
k
(M ). Similarly, one can also
−1 ∗ ∗ k k k
show (Φ ) ◦ Φ is the identity map of HdR (N ). Therefore, HdR (M ) and HdR (N ) are
isomorphic (as vector spaces). 

Corollary 5.8. Given any smooth manifold M which is diffeomorphic to a star-shaped

open set in Rn , we have HdR
1
(M ) ' {[0]}, or in other words, every closed 1-form ω on such
a manifold M is exact.

Proof. Combine the results of the Poincaré Lemma (Theorem 5.4) and the diffeomor-
1
phism invariance of HdR (Theorem 5.7). 

Consequently, a large class of open sets in Rn has trivial HdR

1
as long as it is dif-
2
feomorphic to a star-shaped manifold. For open sets in R , there is a celebrated result
called Riemann Mapping Theorem, which says any (non-empty) simply-connected open
bounded subset U in R2 is diffeomorphic to the unit open ball in R2 . In fact, the dif-
feomorphism can be chosen so that angles are preserved, but we don’t need this when
dealing with de Rham cohomology.
5.1. De Rham Cohomology 147

Under the assumption of Riemann Mapping Theorem (whose proof can be found in
1
advanced Complex Analysis textbooks), we can establish that HdR (U ) = {[0]} for any
2
(non-empty) simply-connected subset U in R . Consequently, any closed 1-form on such
a domain U is exact on U . Using the language in Multivariable Calculus (or Physics),
this means any curl-zero vector field defined on a (non-empty) simply-connected domain
U in R2 must be conservative on U . You might have learned this fact without proof in
MATH 2023.
148 5. De Rham Cohomology

5.2. Deformation Retracts

In the previous section, we learned that two diffeomorphic manifolds have isomorphic
de Rham cohomology groups. In short, we say de Rham cohomology is a diffeomorphic
invariance. In this section, we will discuss another type of invariance: deformation
retracts.
Let M be a smooth manifold (with or without boundary), and Σ is a submanifold
of M . Note that Σ can have lower dimension than M . Roughly speaking, we say Σ is a
deformation retract of M if one can continuously contract M onto Σ. Let’s make it more
precise:

Definition 5.9 (Deformation Retract). Let M be a smooth manifold, and Σ is a subman-

ifold of M . If there exists a C 1 family of smooth maps {Ψt : M → M }t∈[0,1] satisfying
all three conditions below:
• Ψ0 (x) = x for any x ∈ M , i.e. Ψ0 = idM ;
• Ψ1 (x) ∈ Σ for any x ∈ M , i.e. Ψ1 : M → Σ;
• Ψt (p) = p for any p ∈ Σ, t ∈ [0, 1], i.e. Ψt |Σ = idΣ for any t ∈ [0, 1],
then we say Σ is a deformation retract of M . Equivalently, we can also say M deforma-
tion retracts onto Σ.

One good way to think of a deformation retract is to regard t as the time, and Ψt
is a “movie” that demonstates how M collapses onto Σ. The condition Ψ0 = idM says
initially (at t = 0), the “movie” starts with the image M . At the final scene (at t = 1), the
condition Ψ1 : M → Σ says that the image eventually becomes Σ . The last condition
Ψt (p) = p for any p ∈ Σ means the points on Σ do not move throughout the movie.
Before we talk about the relation between cohomology and deformation retract, let’s first
look at some examples:
Example 5.10. The unit circle S1 defined by {(x, y) : x2 + y 2 = 1} is a deformation
retract of the annulus {(x, y) : 14 < x2 + y 2 < 4}. To describe such a retract, it’s best to
use polar coordinates:
Ψt (reiθ ) = (r + t(1 − r)) eiθ
For each t ∈ [0, 1], the map Ψt has image inside the annulus since r + t(1 − r) ∈ ( 21 , 2)
whenever r ∈ ( 12 , 2) and t ∈ [0, 1]. One can easily check that Ψ0 (reiθ ) = reiθ , Ψ1 (reiθ ) =
eiθ and Ψt (eiθ ) = eiθ for any (r, θ) and t ∈ [0, 1]. Hence Ψt fulfills all three conditions
stated in Definition 5.9. 
Example 5.11. Intuitively, we can see the letters E, F, H, K, L, M and N all deformation
retract onto the letter I. Also, the letter Q deformation retracts onto the letter O. The
explicit Ψt for each deformation retract is not easy to write down. 
Example 5.12. A two-dimensional torus with a point removed can deformation retract
onto two circles joined at one point. Try to visualize it! 

Exercise 5.4. Show that the unit circle x2 + y 2 = 1 in R2 is a deformation retract of

R2 \{(0, 0)}.

Exercise 5.5. Show that any star-shaped open set U in Rn deformation retracts
onto its base point.
5.2. Deformation Retracts 149

Exercise 5.6. Let M be a smooth manifold, and Σ0 be the zero section of the tangent
bundle, i.e. Σ0 consists of all pairs (p, 0p ) in T M where p ∈ M and 0p is the zero
vector in Tp M . Show that the zero section Σ0 is a deformation retract of the tangent
bundle T M .

Exercise 5.7. Define a relation ∼ of manifolds by declaring that M1 ∼ M2 if and

only if M1 is a deformation retract of M2 . Is ∼ an equivalence relation?

We next show an important result in de Rham theory, which asserts that deformation
retracts preserve the first de Rham cohomology group.

Theorem 5.13 (Invariance under Deformation Retracts). Let M be a smooth manifold,

1
and Σ be a submanifold of M . If Σ is a deformation retract of M , then HdR (M ) and
1
HdR (Σ) are isomorphic.

Proof. Let ι : Σ → M be the inclusion map, and {Ψt : M → M }t∈[0,1] be the family
of maps satisfying all conditions stated in Definition 5.9. Then, the pull-back map
ι∗ : ∧1 T ∗ M → ∧1 T ∗ Σ induces a map ι∗ : HdR
1 1
(M ) → HdR (Σ). Also, the map Ψ1 : M → Σ
∗ 1 1
induces a pull-back map Ψ1 : HdR (Σ) → HdR (M ). The key idea of the proof is to show
that Ψ∗1 and ι∗ are inverses of each other as maps between HdR 1 1
(M ) and HdR (Σ).
Let ω be an arbitrary closed 1-form defined on M . Similar to the proof of Poincaré
Lemma (Theorem 5.4), we consider the scalar function f : M → R defined by:
Z
f (x) = ω
Ψt (x)

Here, Ψt (x) is regarded as a curve with parameter t joining Ψ0 (x) = x and Ψ1 (x) ∈ Σ.
We will show the following result:

(5.1) Ψ∗1 ι∗ ω − ω = df

which will imply [ω] = Ψ∗1 ι∗ [ω], or in other words, Ψ∗t ◦ ι∗ = id on HdR
1
(M ).
To prove (5.1),
P we use local coordinates (u1 , . . . , un ), and express ω in terms of local
coordinates ω = i ωi dui . For simplicity, let’s assume that such a local coordinate chart
can cover the whole curve Ψt (x) for t ∈ [0, 1]. We will fix this issue later. For each
t ∈ [0, 1], we write Ψit (x) to be the ui -coordinate of Ψt (x), i.e. Ψit = ui ◦ Ψt . Then, one
can calculate df using local coordinates. The calculation is similar to the one we did in
the proof of Poincaré Lemma (Theorem 5.4):

1
∂Ψit
Z Z X
f (x) = ω= ωi (Ψt (x)) dt
Ψt (x) 0 i
∂t
( ! )
X ∂f X Z 1 ∂ X ∂Ψit
j
(df )(x) = du = ωi (Ψt (x)) dt duj
j
∂uj j 0 ∂u j i
∂t
   
X Z 1 X ∂ωi ∂Ψ k
∂Ψ i X ∂
 i
∂Ψ 
t t t 
=  + ωi (Ψt (x)) dt duj
j
 0 ∂uk Ψt (x) ∂uj ∂t i
∂t ∂uj 
i,k
150 5. De Rham Cohomology

∂ωi ∂ωk
Next, recall that ω is a closed 1-form, so we have = for any i, k. Using this on
∂uk ∂ui
the first term, and by switching indices of the second term in the integrand, we get:
   
X Z 1 X ∂ωk k
∂Ψt ∂Ψt i X  k
∂ ∂Ψt   j
(df )(x) =  + ωk (Ψt (x)) dt du
j
 0 ∂ui Ψt (x) ∂uj ∂t ∂t ∂uj 
i,k k
(Z
1
! ) t=1
X ∂ X ∂Ψkt j
X ∂Ψkt
= ωk (Ψt (x)) dt du = ωk (Ψt (x)) duj
j 0 ∂t ∂uj ∂uj t=0
k j,k

where the last equality follows from the (backward) chain rule.
Denote ιt : Ψt (M ) → M the inclusion map at time t, then one can check that
X
Ψ∗t ι∗t ω(x) = (ιt ◦ Ψt )∗ ω(x) = (ιt ◦ Ψt )∗ ωk duk
k
X
= ωk (ιt ◦ Ψt (x)) d(uk ◦ ιt ◦ Ψt (x))
k
X
= ωk (ιt ◦ Ψt (x)) dΨkt
k
X ∂Ψkt
= ωk (Ψt (x)) duj .
∂uj
j,k

Therefore, we get:
t=1
X ∂Ψkt t=1
df = ωk (Ψt (x)) duj = [Ψ∗t ι∗t ω]t=0 = Ψ∗1 ι∗1 ω − Ψ∗0 ι∗0 ω.
∂uj t=0
j,k

Since Ψ0 = idM and ι0 = idM , we have proved (5.1). In case Ψt (x) cannot be covered by
one single local coordinate chart, one can then modify the above proof a bit by covering
the curve Ψt (x) by finitely many local coordinate charts. It can be done because Ψt (x) is
compact. Suppose 0 = t0 < t1 < . . . < tN = 1 is a partition of [0, 1] such that for each
α, the curve Ψt (x) restricted to t ∈ [tα−1 , tα ] can be covered by a single local coordinate
chart, then we have:
N Z tα
X X ∂Ψit
f (x) = ωi (Ψt (x)) dt.
α=1 tα−1 i
∂t

Proceed as in the above proof, we can get:

N
X
Ψ∗tα ι∗tα ω − Ψ∗tα −1 ι∗tα −1 ω = Ψ∗1 ι∗1 ω − Ψ∗0 ι∗0 ω,


df =
α=1

which completes the proof of (5.1) in the general case.

To complete the proof of the theorem, we consider an arbitrary 1-form η on Σ. We
claim that

(5.2) ι∗ Ψ∗1 η = η.

We prove by direct verification using local coordinates (u1 , . . . , un ) on M such that:

(u1 , . . . , uk , 0, . . . , 0) ∈ Σ.
5.2. Deformation Retracts 151

Such a local coordinate system always exists near Σ by Immersion Theorem (Theorem
Pk
2.42). Locally, denote η = i=1 ηi dui , then
k
X k
X
(Ψ∗1 η)(x) = Ψ∗1 (ηi (x) dui ) = ηi (Ψ1 (x)) d(ui ◦ Ψ1 )
i=1 i=1
k X
k
X ∂Ψi1 (x) j
= ηi (Ψ1 (x)) du .
i=1 j=1
∂uj

Since Ψ1 (x) = x whenever x ∈ Σ, we have Ψi1 (x) = ui (x) where ui (x) is the i-th
∂Ψi1 (x) ∂ui
coordinate of x. Therefore, we get = = δij and so:
∂uj ∂uj
k
X k
X
(Ψ∗1 η)(x) = ηi (x)δij duj = ηi (x) dui = η(x)
i,j=1 i=1
∗
for any x ∈ Σ. In other words, ι Ψ∗1 η = η on Σ. This proves (5.2).
Combining (5.1) and (5.2), we get ι∗ ◦ Ψ∗1 = id on HdR1
(Σ), and Ψ∗1 ◦ ι∗ = id on
1
HdR (M ). As a result, Ψ∗1 and ι∗ are inverses
of each other in HdR1
. It completes the proof
1 1
that HdR (M ) and HdR (Σ) are isomorphic. 
1
Using Theorem 5.13, we see that HdR (R2 \{(0, 0)}) and HdR
1
(S1 ) are isomorphic, and
2 1
hence b1 (R \{(0, 0)}) = b1 (S ). At this moment, we still don’t know the exact value of
b1 (S1 ), but we will figure it out in the next section.
k
Note that Theorem 5.13 holds for HdR for any k ≥ 2 as well, but the proof again uses
some Lie derivatives and Cartan’s formula, which are beyond the scope of this course.
Another nice consequence of Theorem 5.13 is the 2-dimensional case of the following
celebrated theorem in topology:

Theorem 5.14 (Brouwer’s Fixed-Point Theorem on R2 ). Let B1 (0) be the closed ball
with radius 1 centered at origin in R2 . Suppose Φ : B1 (0) → B1 (0) is a smooth map
between B1 (0). Then, there exists a point x ∈ B1 (0) such that Φ(x) = x.

Proof. We prove by contradiction. Suppose Φ(x) 6= x for any x ∈ B1 (0). Then, we let
Ψt (x) be a point in B1 (0) defined in the following way:
(1) Consider the vector x − Φ(x) which is non-zero.
(2) Consider the straight ray emanating from x in the direction of x − Φ(x). This ray
will intersect the unit circle S1 at a unique point px .
(3) We then define Ψt (x) := (1 − t)x + tpx
We leave it as an exercise for readers to write down the explicit formula for Ψt (x), and
show that it is smooth for each t ∈ [0, 1].
Clearly, we have Ψ0 (x) = x for any x ∈ B1 (0); Ψ1 (x) = px ∈ S1 ; and if |x| = 1, then
px = x and so Ψt (x) = x.
Therefore, it shows S1 is a deformation retract of B1 (0), and by Theorem 5.13,
1 1 1
their HdR ’s are isomorphic. However, we know HdR (B1 (0)) ' {[0]}, while HdR (S1 ) '
1
HdR (R2 \{(0, 0)}) 6= {[0]}. It is a contradiction! It completes the proof that there is at
least a point x ∈ B1 (0) such that Φ(x) = x. 

Exercise 5.8. Write down an explicit expression of px in the above proof, and hence
show that Ψt is smooth for each fixed t.
152 5. De Rham Cohomology

Exercise 5.9. Generalize the Brouwer’s Fixed-Point Theorem in the following way:
given a manifold Ω which is diffeomorphic to B1 (0), and a smooth map Φ : Ω → Ω.
Using Theorem 5.14, show that there exists a point p ∈ Ω such that Φ(p) = p.

Exercise 5.10. What fact(s) are needed to be established in order to prove the
Brouwer’s Fixed-Point Theorem for general Rn using a similar way as in the proof of
Theorem 5.14?
5.3. Mayer-Vietoris Theorem 153

5.3. Mayer-Vietoris Theorem

In the previous section, we showed that if Σ is a deformation retract of M , then
1 1 1
HdR (Σ) and HdR (M ) are isomorphic. For instance, this shows HdR (S1 ) is isomorphic to
1 2 2 2
HdR (R \{(0, 0)}). Although we have discussed that HdR (R \{(0, 0)}) is non-trivial, we
still haven’t figured out what this group is. In this section, we introduce a useful tool,
called Mayer-Vietoris sequence, that we can use to compute the de Rham cohomology
groups of R2 \{(0, 0)}, as well as many other spaces.

5.3.1. Exact Sequences. Consider a sequence of homomorphism between abelian

groups:
Tk−1 k T Tk+1 Gk+1
· · · −−−→ Gk−1 −→ Gk −−−→ Gk+1 −−−→ · · ·
We say it is an exact sequence if the image of each homomorphism is equal to the kernel
of the next one, i.e.
Im Ti−1 = ker Ti for each i.

One can also talk about exact-ness for a finite sequence, say:
1 T 2 3 T T n Tn−1 T
G0 −→ G1 −→ G2 −→ · · · −−−→ Gn−1 −−→ Gn
However, such a T1 would not have a previous map, and such an Tn would not have the
next map. Therefore, whenever we talk about the exact-ness of a finite sequence of maps,
we will add two trivial maps at both ends, i.e.
0 1 T
2 3 T n T T 0
(5.3) 0−
→ G0 −→ G1 −→ G2 −→ · · · Gn−1 −−→ Gn −
→ 0.
0
The first map 0 − → G0 is the homomorphism taking the zero in the trivial group to the
0
zero in G0 . The last map Gn −→ 0 is the linear map that takes every element in Gn to the
zero in the trivial group. We say the finite sequence (5.3) an exact sequence if
0 0
Im (0 −
→ G0 ) = ker T1 , Im Tn = ker(Gn −
→ 0), and Im Ti = ker Ti+1 for any i.
0 0
Note that Im (0 −
→ G0 ) = {0} and ker(Gn −
→ 0) = Gn , so if (5.3) is an exact sequence, it
is necessary that
ker T1 = {0} and Im Tn = Gn
or equivalently, T1 is injective and Tn is surjective.
One classic example of a finite exact sequence is:
ι f
0→Z→
− C−
→ C\{0} → 0
where ι : Z → C is the inclusion map taking n ∈ Z to itself n ∈ C. The map f : C →
C\{0} is the map taking z ∈ C to e2πiz ∈ C\{0}.
It is clear that ι is injective and f is surjective (from Complex Analysis). Also, we have
Im ι = Z and ker f = Z as well (note that the identity of C\{0} is 1, not 0). Therefore,
this is an exact sequence.

Exercise 5.11. Given an exact sequence of group homomorphisms:

T S
0→A−
→B−
→ C → 0,
(a) If it is given that C = {0}, what can you say about A and B?
(b) If it is given that A = {0}, what can you say about B and C?
154 5. De Rham Cohomology

5.3.2. Mayer-Vietoris Sequences. We talk about exact sequences because there is

such a sequence concerning de Rham cohomology groups. This exact sequence, called the
k
Mayer-Vietoris sequence, is particularly useful for computing HdR for many manifolds.
The basic setup of a Mayer-Vietoris sequence is a smooth manifold (with or without
boundary) which can be expressed a union of two open sets U and V , i.e. M = U ∪ V .
Note that we do not require U and V are disjoint. The intersection U ∩ V is a subset of
both U and V ; and each of U and V is in turn a subset of M . To summarize, we have the
following relations of sets:
U
iU jU

U ∩V M

iV jV
V
where iU , iV , jU and jV are inclusion maps. Each inclusion map, say jU : U → M ,
induces a pull-back map jU∗ : ∧k T ∗ M → ∧k T ∗ U which takes any k-form ω on M , to the
k-form ω|U restricted on U , i.e. jU∗ (ω) = ω|U for any ω ∈ ∧k T ∗ M . In terms of local
expressions,P there is essentially no difference
P between ω and ω|U since U is open. If
locally ω = i ωi dui on M , then ω|U = i ωi dui as well. The only difference is the
domain: ω(p) is defined for every p ∈ M , while ω|U (p) is defined only when p ∈ U .
To summarize, we have the following diagram:
U
i∗
U
∗
jU

U ∩V M
∗
i∗
V
jV
V
Using the pull-backs of these four inclusions iU , iV , jU and jV , one can form a
sequence of linear maps for each integer k:
j ∗ ⊕j ∗ i∗ −i∗
(5.4) 0 → ∧k T ∗ M −−
U V
−−→ ∧k T ∗ U ⊕ ∧k T ∗ V −−
U V
−−→ ∧k T ∗ (U ∩ V ) → 0
Here, ∧k T ∗ U ⊕∧k T ∗ V is the direct sum of the vector spaces ∧k T ∗ U and ∧k T ∗ V , meaning
that:
∧k T ∗ U ⊕ ∧k T ∗ V = {(ω, η) : ω ∈ ∧k T ∗ U and η ∈ ∧k T ∗ V }.
The map jU∗ ⊕ jV∗ : ∧k T ∗ M → ∧k T ∗ U ⊕ ∧k T ∗ V is defined by:
(jU∗ ⊕ jV∗ )(ω) = (jU∗ ω, jV∗ ω) = ( ω|U , ω|V ).
i∗ −i∗
The map ∧k T ∗ U ⊕ ∧k T ∗ V −−
U V
−−→ ∧k T ∗ (U ∩ V ) is given by:
(i∗U − i∗V )(ω, η) = i∗U ω − i∗V η = ω|U ∩V − η|U ∩V .

We next show that the sequence (5.4) is exact. Let’s first try to understand the image
and kernel of each map involved.
Given (ω, η) ∈ ker(i∗U − i∗V ), we will have ω|U ∩V = η|U ∩V . Therefore, ker(i∗U − i∗V )
consists of pairs (ω, η) where ω and η agree on the intersection U ∩ V .
Now consider Im (jU∗ ⊕ jV∗ ), which consists of pairs of the form ( ω|U , ω|V ). Certainly,
the restrictions of both ω|U and ω|V on the intersection U ∩ V are the same, and hence
the pair is inside ker(i∗U − i∗V ). Therefore, we have Im (jU∗ ⊕ jV∗ ) ⊂ ker(i∗U − i∗V ).
5.3. Mayer-Vietoris Theorem 155

In order to show (5.4) is exact, we need further that:

(1) jU∗ ⊕ jV∗ is injective;
(2) i∗U − i∗V is surjective; and
(3) ker(i∗U − i∗V ) ⊂ Im (jU∗ ⊕ jV∗ )
We leave (1) as an exercises, and will give the proofs of (2) and (3).
Exercise 5.12. Show that jU∗ ⊕ jV∗ is injective in the sequence (5.4).

Proposition 5.15. Let M be a smooth manifold. Suppose there are two open subsets
U and V of M such that M = U ∪ V , and U ∩ V is non-empty, then the sequence of
maps (5.4) is exact.

Proof. So far we have proved that jU∗ ⊕ jV∗ is injective, and Im (jU∗ ⊕ jV∗ ) ⊂ ker(i∗U − i∗V ).
We next claim that ker(i∗U − i∗V ) ⊂ Im (jU∗ ⊕ jV∗ ):
Let (ω, η) ∈ ker(i∗U − i∗V ), meaning that ω is a k-form on U , η is a k-form on V , and
that ω|U ∩V = η|U ∩V . Define a k-form σ on M = U ∪ V by:
(
ω on U
σ=
η on V
Note that σ is well-defined on U ∩ V since ω and η agree on U ∩ V . Then, we have:
(ω, η) = ( σ|U , σ|V ) = (jU∗ σ, jV∗ σ) = (jU∗ ⊕ jV∗ )σ ∈ Im (jU∗ ⊕ jV∗ ).
Since (ω, η) is arbitrary in ker(i∗U − i∗V ), this shows:
ker(i∗U − i∗V ) ⊂ Im (jU∗ ⊕ jV∗ ).
Finally, we show i∗U − i∗V is surjective. Given any k-form θ ∈ ∧k T ∗ (U ∩ V ), we need
to find a k-form ω 0 on U , and a k-form η 0 on V such that ω 0 − η 0 = θ on U ∩ V . Let
{ρU , ρV } be a partition of unity subordinate to {U, V }. We define:
(
ρV θ on U ∩ V
ω0 =
0 on U \V
Note that ω 0 is smooth: If p ∈ supp ρV ⊂ V , then p ∈ V (which is open) and so ω 0 = ρV θ
in an open neighborhood of p. Note that ρV and θ are smooth at p, so ω 0 is also smooth
at p. On the other hand, if p 6∈ supp ρV , then ω 0 = 0 in an open neighborhood of p. In
particular, ω 0 is smooth at p.
Similarly, we define: (
0 −ρU θ on U ∩ V
η =
0 on V \U
which can be shown to be smooth in a similar way.
Then, when restricted to U ∩ V , we get:
ω 0 |U ∩V − η 0 |U ∩V = ρV θ + ρU θ = (ρV + ρU ) θ = θ.
In other words, we have (i∗U − i∗V )(ω 0 , η 0 ) = θ. Since θ is arbitrary, we proved i∗U − i∗V is
surjective. 
k
Recall that a pull-back map on k-forms induces a well-defined pull-back map on HdR .
The sequence of maps (5.4) between space of wedge products induces a sequence of
maps between de Rham cohomology groups:
k U Vj ∗ ⊕j ∗
k k U V k i∗ −i∗
(5.5) 0 → HdR (M ) −−−−→ HdR (U ) ⊕ HdR (V ) −−−−→ HdR (U ∩ V ) → 0.
156 5. De Rham Cohomology

Here, jU∗ ⊕ jV∗ and i∗U − i∗V are defined by:

(jU∗ ⊕ jV∗ )[ω] = (jU∗ [ω], jV∗ [ω]) = ([jU∗ ω], [jV∗ ω])
(i∗U − i∗V ) ([ω], [η]) = i∗U [ω] − i∗V [η] = [i∗U ω] − [i∗V η].

However, the sequence (5.5) is not exact because jU∗ ⊕ jV∗ may not be injective, and
i∗U − i∗V may not be surjective. For example, take M = R2 \{(0, 0)}, and define using
polar coordinates the open sets U = {reiθ : r > 0, θ ∈ (0, 2π)} and V = {reiθ : r > 0, θ ∈
1 1
(−π, π)}. Then, both U and V are star-shaped and hence both HdR (U ) and HdR (V ) are
trivial. Nonetheless we have exhibited that HdR (M ) is non-trivial. The map jU ⊕ jV∗ from
1 ∗

a non-trivial group to the trivial group can never be injective!

Exercise 5.13. Find an example of M , U and V such that the map i∗U − i∗V in (5.5)
is not surjective.

Nonetheless, it is still true that ker(i∗U − i∗V ) = Im (jU∗ ⊕ jV∗ ), and we will verify it in
the proof of Mayer-Vietoris Theorem (Theorem 5.16). Mayer-Vietoris Theorem asserts
that although (5.5) is not exact in general, but we can connect each short sequence
below:
0 U Vj ∗ ⊕j ∗
0 0 U V 0 i∗ −i∗
HdR (M ) −−−−→ HdR (U ) ⊕ HdR (V ) −−−−→ HdR (U ∩ V )
1 U Vj ∗ ⊕j ∗
1 1 U V 1 i∗ −i∗
HdR (M ) −−−−→ HdR (U ) ⊕ HdR (V ) −−−−→ HdR (U ∩ V )
2 U Vj ∗ ⊕j ∗
2 2 U V 2 i∗ −i∗
HdR (M ) −−−−→ HdR (U ) ⊕ HdR (V ) −−−−→ HdR (U ∩ V )
..
.
to produce a long exact sequence.

Theorem 5.16 (Mayer-Vietoris Theorem). Let M be a smooth manifold, and U and V

be open sets of M such that M = U ∪ V . Then, for each k ≥ 0 there is a homomorphism
k k+1
δ : HdR (U ∩ V ) → HdR (M ) such that the following sequence is exact:
δ k j ∗ ⊕j ∗
U V k k U V i∗ −i∗
k k+1 δ
··· −
→ HdR (M ) −−−−→ HdR (U ) ⊕ HdR (V ) −−−−→ HdR (U ∩ V ) −
→ HdR (M ) → · · ·
This long exact sequence is called the Mayer-Vietoris sequence.

The proof of Theorem 5.16 is purely algebraic. We will learn the proof after looking
at some examples.

5.3.3. Using Mayer-Vietoris Sequences. The Mayer-Vietoris sequence is particu-

larly useful for computing de Rham cohomology groups and Betti numbers using linear
algebraic methods. Suppose M can be expressed as a union U ∪ V of two open sets, such
k k
that the HdR ’s of U , V and U ∩ V can be computed easily, then HdR (M ) can be deduced
by “playing around” the kernels and images in the Mayer-Vietoris sequence. One useful
result in Linear (or Abstract) Algebra is the following:

Theorem 5.17 (First Isomorphism Theorem). Let T : V → W be a linear map between

two vector spaces V and W . Then, we have:
Im T ∼
= V / ker T.
In particular, if V and W are finite dimensional, we have:
dim ker T + dim Im T = dim V.
5.3. Mayer-Vietoris Theorem 157

Proof. Let Φ : Im T → V / ker T be the map defined by:

Φ(T (v)) = [v]
for any T (v) ∈ Im T . This map is well-defined since if T (v) = T (w) in Im T , then
v − w ∈ ker T , which implies [v] = [w] in the quotient vector space V / ker T . It is easy
(hence omitted) to verify that Φ is linear.
Φ is injective since whenever T (v) ∈ ker Φ, we have Φ(T (v)) = [0] which implies
[v] = [0] and hence v ∈ ker T (i.e. T (v) = 0). Also, Φ is surjective since given any
[v] ∈ V / ker T , we have Φ(T (v)) = [v] by the definition of Φ.
These show Φ is an isomorphism, hence completing the proof. 
1
Example 5.18. In this example, we use the Mayer-Vietoris sequence to compute HdR (S1 ).
Let:
M = S1 , U = M \{north pole}, V = M \{south pole}.
Then clearly M = U ∪ V , and U ∩ V consists of two disjoint arcs (each of which
deformation retracts to a point). Here are facts which we know and which we haven’t
yet known:
0
HdR (M ) ∼
=R 0
HdR (U ) ∼
=R 0
HdR (V ) ∼
=R 0
HdR (U ∩ V ) ∼
=R⊕R
1
HdR (M ) unknown 1
H (U ) = ∼0 1
H (V ) = ∼0 1
H (U ∩ V ) = ∼0
dR dR dR

By Theorem 5.16, we know that the following sequence is exact:

0 0 i∗ −i∗
U V 0 δ
1 U Vj ∗ ⊕j ∗
1 1
· · · → HdR (U ) ⊕ HdR (V ) −−−−→ HdR (U ∩ V ) −
→ HdR (M ) −−−−→ HdR (U ) ⊕ HdR (V )
| {z } | {z } | {z } | {z }
R⊕R R⊕R ? 0

Therefore, δ is surjective.
By First Isomorphism Theorem (Theorem 5.17), we know:
H 0 (U ∩ V )
1
HdR (M ) = Im δ ∼
= dR .
ker δ
0
Elements of HdR (U ∩ V ) are locally constant functions of the form:
(
a on left arc
fa,b =
b on right arc

Since the Mayer-Vietoris sequence is exact, we have ker δ = Im (i∗U − i∗V ). The space
0 0 0
HdR (U ),HdR (V ) and HdR (U ∩ V ) consist of locally constant functions on U , V and U ∩ V
respectively, and the maps i∗U − i∗V takes constant functions (k1 , k2 ) ∈ HdR0 0
(U ) ⊕ HdR (V )
to the constant function fk1 −k2 ,k1 −k2 on U ∩ V . Therefore, the first de Rham cohomology
group of M is given by:
{fa,b : a, b ∈ R} ∼ R2
1
HdR (M ) ∼
= = ,
{fa−b,a−b : a, b ∈ R} {(x, y) : x = y}
1
and hence b1 (M ) = dim HdR (M ) = 2 − 1 = 1. 
Example 5.19. Let’s discuss some consequences of the result proved in the previous
example. Recall that R2 \{(0, 0)} deformation retracts to S1 . By Theorem 5.13, we know
1
HdR (R2 \{(0, 0)}) ∼ 1
= HdR (S1 ).
This tells us b1 (R2 \{(0, 0)}) = 1 as well. Recall that the following 1-form:
−y dx + x dy
ω=
x2 + y 2
158 5. De Rham Cohomology

1
is closed but not exact. The class [ω] is then trivial in HdR (R2 \{(0, 0)}). In an one-
dimensional vector space, any non-zero vector spans that space. Therefore, we conclude:
1
HdR (R2 \{(0, 0)} = {c[ω] : c ∈ R}.
where ω is defined as in above.
As a result, if ω 0 is a closed 1-form on R2 \{(0, 0)}, then we must have
[ω 0 ] = c[ω]
for some c ∈ R, and so ω 0 = cω + df for some smooth function f : R2 \{(0, 0)} → R.
Using the language of vector fields, if V (x, y) : R2 \{(0, 0)} → R2 is a smooth
vector field with ∇ × V = 0, then there is a constant c ∈ R and a smooth function
f : R2 \{(0, 0)} → R such that:
!
−y î + xĵ
V=c + ∇f.
x2 + y 2


Exercise 5.14. Let T2 be the two-dimensional torus. Show that b1 (T2 ) = 2.

Exercise 5.15. Show that b1 (S2 ) = 0. Based on this result, show that any curl-zero
vector field defined on R3 \{(0, 0, 0)} must be conservative.

One good technique of using the Mayer-Vietoris sequence (as demonstrated in the
examples and exercises above) is to consider a segment of the sequence that starts and
ends with the trivial space, i.e.
0 → V1 → V2 → · · · → Vn → 0.
If all vector spaces Vi ’s except one of them are known, then the remaining one (at least its
dimension) can be deduced using First Isomorphism Theorem. Below is a useful lemma
which is particularly useful for finding the Betti number of a manifold:

Lemma 5.20. Let the following be an exact sequence of finite dimensional vector spaces:
1T 2 T Tn−1
0 → V1 −→ V2 −→ · · · −−−→ Vn → 0.
Then, we have:
dim V1 − dim V2 + dim V3 − · · · + (−1)n−1 dim Vn = 0

Proof. By exact-ness, the map Tn−1 : Vn−1 → Vn is surjective. By First Isomorphism

Theorem (Theorem 5.17), we get:
Vn = Im Tn−1 ∼
= Vn−1 / ker Tn−1 = Vn−1 /Im Tn−2 .
As a result, we have:
dim Vn = dim Vn−1 − dim Im Tn−2 .
Similarly, apply First Isomorphism Theorem on Tn−2 : Vn−2 → Vn−1 , we get:
dim Im Tn−2 = dim Vn−2 − dim Im Tn−3 ,
and combine with the previous result, we get:
dim Vn = dim Vn−1 − dim Vn−2 + dim Im Tn−3 .
Proceed similarly as the above, we finally get:
dim Vn = dim Vn−1 − dim Vn−2 + . . . + (−1)n dim V1 ,
as desired. 
5.3. Mayer-Vietoris Theorem 159

1
In Example 5.18 (about computing HdR (S1 )), the following exact sequence was used:
0
0 → HdR (S1 ) → HdR
0 0
(U ) ⊕ HdR 0
(V ) → HdR 1
(U ∩ V ) → HdR (S1 ) → HdR
1 1
(U ) ⊕ HdR (V )
| {z } | {z } | {z } | {z } | {z }
R R⊕R R⊕R ? 0

1
Using Lemma 5.20, the dimension of HdR (S1 ) can be computed easily:
1
dim R − dim R ⊕ R + dim R ⊕ R − dim HdR (S1 ) = 0
1
which implies dim HdR (S1 ) = 1 (or equivalently, b1 (S1 ) = 1). Although this method does
1
not give a precise description of HdR (S1 ) in terms of inclusion maps, it is no doubt much
easier to adopt.
In the forthcoming examples, we will assume the following facts stated below (which
we have only proved the case k = 1):
k
• HdR (U ) = 0, where k ≥ 1, for any star-shaped region U ⊂ Rn .
k
• If Σ is a deformation retract of M , then HdR ∼ H k (M ) for any k ≥ 1.
(Σ) = dR

Example 5.21. Consider R2 \{p1 , . . . , pn } where p1 , . . . , pn are n distinct points in R2 .

We want to find b1 of this open set.
Define U = R2 \{p1 , . . . , pn−1 }, V = R2 \{pn }, then U ∪ V = R2 and U ∩ V =
2
R \{p1 , . . . , pn }. Consider the Mayer-Vietoris sequence:
H 1 (U ∪ V ) → HdR
1 1
(U ) ⊕ HdR 1
(V ) → HdR 2
(U ∩ V ) → HdR (U ∪ V ) .
| dR {z } | {z }
0 0

Using Lemma 5.20, we know:

1 1 1
dim HdR (U ) ⊕ HdR (V ) − dim HdR (U ∩ V ) = 0
1
We have already figured out that dim HdR (V ) = 1. Therefore, we get:
1
dim HdR (R2 \{p1 , . . . , pn }) = dim HdR
1
(R2 \{p1 , . . . , pn−1 }) + 1.
By induction, we conclude:
b1 (R2 \{p1 , . . . , pn }) = dim HdR
1
(R2 \{p1 , . . . , pn }) = n.

n
Example 5.22. Consider the n-sphere S (where n ≥ 2). It can be written as U ∪V where
U := S n \{north pole} and V := S n \{south pole}. Using stereographic projections, one
can show both U and V are diffeomorphic to Rn . Furthermore, U ∩ V is diffeomorphic to
Rn \{0}, which deformation retracts to S n−1 . Hence HdR
k
(S n−1 ) = HdR
k
(U ∩ V ) for any k.
Now consider the Mayer-Vietoris sequence with these U and V , we have for each
k ≥ 2 an exact sequence:
k−1 k−1 k−1 k
HdR (U ) ⊕ HdR (V ) → HdR (U ∩ V ) → HdR (S n ) → HdR
k k
(U ) ⊕ HdR (V ) .
| {z } | {z }
0 0

This shows HdR k−1

(S n−1 ) ∼
= k
HdR (S n )
for any k ≥ 2. By induction, we conclude that
n
HdR (S n ) ∼ H
= dR1
(S 1 ∼
) = R for any n ≥ 2. 

5.3.4. Proof of Mayer-Vietoris Theorem. To end this chapter (and this course),
we present the proof of the Mayer-Vietoris’s Theorem (Theorem 5.16). As mentioned
before, the proof is purely algebraic. The key ingredient of the proof applies to many
other kinds of cohomologies as well (de Rham cohomology is only one kind of many
types of cohomology).
160 5. De Rham Cohomology

For simplicity, we denote:

X k := ∧k T ∗ M Y k := ∧k T ∗ U ⊕ ∧k T ∗ V Z k := ∧k T ∗ (U ∩ V )
H k (X) := HdR
k
(M ) H k (Y ) := HdR
k k
(U ) ⊕ HdR (V ) H k (Z) := HdR
k
(U ∩ V )
Furthermore, we denote the pull-back maps i∗U − i∗V and jU∗ ⊕ jV∗ by simply i and j
respectively. We then have the following commutative diagram between all these X, Y
and Z:
j i
0 Xk Yk Zk 0
d d d
j i
0 X k+1 Y k+1 Z k+1 0
d d d
j i
0 X k+2 Y k+2 Z k+2 0
The maps in the diagram commute because the exterior derivative d commute with
any pull-back map. The map d : Y k → Y k+1 takes (ω, η) to (dω, dη).
To give a proof of the Mayer-Vietoris Theorem, we first need to construct a linear
k k+1
map δ : HdR (Z) → HdR (Z). Then, we need to check that the connected sequence:
i δ j i δ
− H k (Z) −
··· → → H k+1 (X) −
→ H k+1 (Y ) →
− H k+1 (Z) −
→ ···
is exact. Most arguments involved are done by “chasing the commutative diagram”.
δ
Step 1: Construction of H k (Z) −
→ H k+1 (X)
Let [θ] ∈ H k (Z), where θ ∈ Z k is a closed k-form on U ∩ V . Recall from Proposition
5.15 that the sequence
j i
0 → Xk −
→Yk →
− Zk → 0
is exact, and in particular i is surjective. As a result, there exists ω ∈ Y k such that
i(ω) = θ.
From the commutative diagram, we know idω = diω = dθ = 0, and hence dω ∈ ker i.
By exact-ness, Im j = ker i and so there exists η ∈ X k+1 such that j(η) = dω.
Next we argue that such η must be closed: since j(dη) = d(jη) = d(dω) = 0, and j is
injective by exact-ness. We must have dη = 0, and so η represents a class in H k+1 (X).
To summarize, given [θ] ∈ H k (Z), ω and η are elements such that
i(ω) = θ and j(η) = dω.
k+1
We then define δ[θ] := [η] ∈ H (X).
Step 2: Verify that δ is a well-defined map
Suppose [θ] = [θ0 ] in HdR
k
(Z). Let ω 0 ∈ Y k and η 0 ∈ X k+1 be the corresponding
0
elements associated with θ , i.e.
i(ω 0 ) = θ0 and j(η 0 ) = dω 0 .
We need to show [η] = [η 0 ] in H k+1 (X).
From [θ] = [θ0 ], there exists a (k − 1)-form β in Z k−1 such that θ − θ0 = dβ, which
implies:
i(ω − ω 0 ) = θ − θ0 = dβ.
By surjectivity of i : Y k−1 → Z k−1 , there exists α ∈ Y k−1 such that iα = β. Then we get:
i(ω − ω 0 ) = d(iα) = idα
which implies (ω − ω 0 ) − dα ∈ ker i.
5.3. Mayer-Vietoris Theorem 161

By exact-ness, ker i = Im j and so there exists γ ∈ X k such that

jγ = (ω − ω 0 ) − dα.
Differentiating both sides, we arrive at:
djγ = d(ω − ω 0 ) − d2 α = j(η − η 0 ).
Therefore, jdγ = djγ = j(η − η 0 ), and by injectivity of j, we get:
η − η 0 = dγ
and so [η] = [η 0 ] in H k+1 (X).
Step 3: Verify that δ is a linear map
We leave this step as an exercise for readers.
i δ
Step 4: Check that H k (Y ) →
− H k (Z) −
→ H k+1 (X) is exact
To prove Im i ⊂ ker δ, we take an arbitrary [θ] ∈ Im i ⊂ H k (Z), there is [ω] ∈ H k (Y )
such that [θ] = i[ω], we will show δ[θ] = 0. Recall that δ[iω] is the element [η] in H k+1 (X)
such that jη = dω. Now that ω is closed, the injectivity of j implies η = 0. Therefore,
δ[θ] = δ[iω] = [0], proving [θ] ∈ ker δ.
Next we show ker δ ⊂ Im i. Suppose [θ] ∈ ker δ, and let ω and η be the forms such
that i(ω) = θ and j(η) = dω. Then [η] = δ[θ] = [0], so there exists γ ∈ X k−1 such that
η = dγ, which implies j(dγ) = dω, and so ω − jγ is closed. By exact-ness, i(jγ) = 0, and
so:
θ = i(ω) = i(ω − jγ).
For ω − jγ being closed, we conclude [θ] = i[ω − jγ] ∈ Im i in H k (Z).
δ j
Step 5: Check that H k (Z) −
→ H k+1 (X) −
→ H k+1 (Y ) is exact
First show Im δ ⊂ ker j. Let [θ] ∈ H k+1 (Z), then δ[θ] = [η] where
i(ω) = θ and j(η) = dω.
As a result, jδ[θ] = j[η] = [dω] = [0]. This shows δ[θ] ∈ ker j.
Next we show ker j ⊂ Im δ. Let j[ω] = [0], then jω = dα for some α ∈ Y k . Since:
i(α) = iα and j(ω) = dα
We conclude δ[iα] = [ω], or in other words, [ω] ∈ Im δ.
j i
Step 6: Check that H k+1 (X) −
→ H k+1 (Y ) →
− H k+1 (Z) is exact
The inclusion Im j ⊂ ker i follows from the fact that i(jη) = 0 for any closed
η ∈ X k+1 , and hence ij[η] = [0]. Finally, we show ker i ⊂ Im j: suppose [ω] ∈ ker i so
that iω = dβ for some β ∈ Z k . By surjectivity of i : Y k → Z k , there exists α ∈ Y k such
that β = iα. As a result, we get:
iω = diα = idα =⇒ ω − dα ∈ ker i.
k+1 k+1
Since ker i = Im j on the level of X →Y → Z k+1 , there exists γ ∈ X k+1 such
that jγ = ω − dα. One can easily show γ is closed by injectivity of j:
jdγ = djγ = d(ω − dα) = 0 =⇒ dγ = 0
k+1
and so [γ] ∈ H (X). Finally, we conclude:
j[γ] = [ω − dα] = [ω]
and so [ω] ∈ Im j.
* End of the proof of the Mayer-Vietoris Theorem *
** End of MATH 4033 **
*** I hope you enjoy it. ***
Part 2

Euclidean Hypersurfaces
 Chapter 6

Geometry of Curves

“You can’t criticize geometry. It’s never

wrong.”

Paul Rand

Riemannian geometry is a branch in differential geometry which studies intrinsic

geometric structure without referencing to the ambient space. It was first developed
by Gauss and Riemann, and was later adopted by Einstein to lay the mathematical
foundation of general relativity, which regards our space-time as an intrinsic manifold.
Despite the intrinsic nature of Riemannian geometry, many of its important notions
and concepts are stemmed from extrinsic geometry, namely hypersurfaces in Euclidean
spaces. In this and the next chapters, we will first explore ourselves to the basic differen-
tial geometry of Euclidean hypersurfaces.

6.1. Curvature and Torsion

6.1.1. Regular Curves. A curve in the Euclidean space Rn is regarded as a function
γ(t) from an interval I to Rn . The interval I can be finite, infinite, open, closed or
half-open. Denote the coordinates of Rn by (x1 , x2 , . . . , xn ), then a curve γ(t) in Rn can
be written in coordinate form as:

γ(t) = (γ 1 (t), . . . , γ n (t)).

One easy way to make sense of a curve is to regard it as the trajectory of a particle.
At any time t, the functions γ 1 (t), . . . , γ n (t) give the coordinates of the particle in Rn .
Assuming all xi (t), where 1 ≤ i ≤ n, are at least twice differentiable, then the first
derivative γ 0 (t) represents the velocity of the particle, its magnitude |γ 0 (t)| is the speed of
the particle, and the second derivative γ 00 (t) represents the acceleration of the particle.
We will mostly study those curves which are infinitely differentiable (i.e. C ∞ ). For
some technical purposes as we will explain later, we only study those curves γ(t) whose
velocity γ 0 (t) is never zero. We call those curves:

Definition 6.1 (Regular Curves). A regular curve is a C ∞ function γ(t) : I → Rn such

that γ 0 (t) 6= 0 for any t ∈ I.

165
166 6. Geometry of Curves

Example 6.2. The curve γ(t) = (cos(et ), sin(et )), where t ∈ (−∞, ∞), is a regular curve
since γ 0 (t) = (−et sin(et ), et cos(et )) and |γ 0 (t)| = et 6= 0 for any t.
However, γe(t) = (cos t2 , sin t2 ), where t ∈ (−∞, ∞), is not a regular curve since
0
e0 (0) = 0.
e (t) = (−2t sin t2 , 2t cos t2 ) and so γ
γ
Although both curves γ(t) and γ e(t) represent the unit circle centered at the origin
in R2 , one is regular but another is not. Therefore, the term regular refers to the
parametrization rather than the trajectory. 

6.1.2. Arc-Length Parametrization. From Calculus, the arc-length of a curve γ(t)

from t = t0 to t = t1 is given by:
Z t1
|γ 0 (t)| dt.
t0
Now suppose the curve γ(t) starts at t = 0 (call it the initial time). Then the following
quantity:
Z t
s(t) := |γ 0 (τ )| dτ
0
measures the distance traveled by the particle after t unit time since its initial time.
Given a curve γ(t) = (cos(et − 1), sin(et − 1)), we have
γ 0 (t) = (−et sin(et − 1), et cos(et − 1)),
|γ 0 (t)| = et 6= 0 for any t ∈ (−∞, ∞).
Therefore, γ(t) is a regular curve. By an easy computation, one can show s(t) = et −1 and
so, regarding t as a function of s, we have t(s) = log(s + 1). By substituting t = log(s + 1)
into γ(t), we get:
 
γ(t(s)) = γ(log(s + 1)) = cos(elog(s+1) − 1), sin(elog(s+1) − 1) = (cos s, sin s).

The curve γ(t(s)) is ultimately a function of s. With abuse of notations, we denote

γ(t(s)) simply by γ(s). Then, this γ(s) has the same trajectory as γ(t) and both curves
at C ∞ . The difference is that the former travels at a unit speed. The curve γ(s) is a
reparametrization of γ(t), and is often called an arc-length parametrization of the curve.
However, if we attempt to do find a reparametrization on a non-regular curve
e(t) = (cos(t2 ), sin(t2 )), in a similar way as the above, we can see that such the
say γ
reparametrization obtained will not be smooth. To see this, we first compute
Z t Z t (
0 t2 if t ≥ 0;
s(t) = |e
γ (τ )| dτ = 2|τ |dτ = 2
0 0 −t if t < 0.
Therefore, regarding t as a function of s, we have
(√
s if s ≥ 0;
t(s) = √
− −s if s < 0.
Then,
(
(cos(s), sin(s)) if s ≥ 0;
γ
e(s) := γ
e(t(s)) =
(cos(−s), sin(−s)) if s < 0,
e(s) = (cos(s), sin |s|), which is not differentiable at s = 0.
or in short, γ
It turns out the reason why the reparametrization by s works well for γ(t) but not
for γ
e(t) is that the former is regular but the later is not. In general, one can always
reparametrize a regular curve by its arc-length s. Let’s state it as a theorem:
6.1. Curvature and Torsion 167

Theorem 6.3. Given any regular curve γ(t) : I → Rn , one can always reparametrize it by
arc-length. Precisely, let t0 ∈ I be a fixed number and consider the following function of t:
Z t
s(t) := |γ 0 (τ )| dτ.
t0
Then, t can be regarded as a C ∞ function of s, and the reparametrized curve γ(s) := γ(t(s))
d
is a regular curve such that ds γ(s) = 1 for any s.

Proof. The Fundamental Theorem of Calculus shows

d t 0
Z
ds
= |γ (τ )| dτ = |γ 0 (t)| > 0.
dt dt t0
We have |γ 0 (t)| > 0 since γ(t) is a regular curve. Now s(t) is a strictly increasing function
of t, so one can regard t as a function of s by the Inverse Function Theorem. Since s(t) is
C ∞ (because γ(t) is C ∞ and |γ 0 (t)| 6= 0), by the Inverse Function Theorem t(s) is C ∞
too.
d
To verify that ds γ(s) = 1, we use the chain rule:
d dγ dt
γ(s) = ·
ds dt ds
1
= γ 0 (t) · ds
dt
d 1
γ(s) = |γ 0 (t)| · 0 = 1.
ds |γ (t)|


Exercise 6.1. Determine whether each of the following is a regular curve. If so,
reparametrize the curve by arc-length:
(a) γ(t) = (cos t, sin t, t), t ∈ (−∞, ∞)
(b) γ(t) = (t − sin t, 1 − cos t), t ∈ (−∞, ∞)

6.1.3. Definition of Curvature. Curvature is quantity that measures the sharpness

of a curve, and is closely related to the acceleration. Imagine you are driving a car along
a curved road. On a sharp turn, the force exerted on your body is proportional to the
acceleration according to the Newton’s Second Law. Therefore, given a parametric curve
γ(t), the magnitude of the acceleration |γ 00 (t)| somewhat reflects the sharpness of the
path – the sharper the turn, the larger the |γ 00 (t)|.
However, the magnitude |γ 00 (t)| is not only affected by the sharpness of the curve,
but also on how fast you drive. In order to give a fair and standardized measurement of
sharpness, we need to get an arc-length parametrization γ(s) so that the “car” travels at
unit speed.

Definition 6.4 (Curvature). Let γ(s) : I → Rn be an arc-length parametrization of a

path γ in Rn . The curvature of γ is a function κ : I → R defined by:
κ(s) = |γ 00 (s)| .

Remark 6.5. Since an arc-length parametrization is required in the definition, we talk

about curvature for only for regular curves. 
168 6. Geometry of Curves

Another way (which is less physical) to understand curvature is to regard γ 00 (s) as

d
ds T(s)where T(s) := γ 0 (s) is the unit tangent vector at γ(s). The curvature κ(s) is then
d
given by ds T(s) which measures how fast the unit tangents T(s) move or turn along
the curve (see Figure 6.1).

Figure 6.1. curvature measures how fast the unit tangents move

Example 6.6. The circle of radius R centered at the origin (0, 0) on the xy-plane can be
parametrized by γ(t) = (R cos t, R sin t). It can be easily verified that |γ 0 (t)| = R and so
γ(t) is not an arc-length parametrization.
To find an arc-length parametrization, we let:
Z t Z t
0
s(t) = |γ (τ )| dτ = R dτ = Rt.
0 0
s
Therefore, t(s) = Ras a function of s and so an arc-length parametrization of the circle
is:  s s
γ(s) := γ(t(s)) = R cos , R sin .
R R
To find its curvature, we compute:
d  s s
γ 0 (s) = R cos , R sin
ds s
R
s
R
= − sin , cos
 R R 
00 1 s 1 s
γ (s) = − cos , − sin
R R R R
1
κ(s) = |γ 00 (s)| = .
R
Thus the curvature of the circle is given by R1 , i.e. the larger the circle, the smaller the
curvature. 

Exercise 6.2. Find an arc-length parametrization of the helix:

γ(t) = (a cos t, a sin t, bt)
where a and b are positive constants. Hence compute its curvature.
6.1. Curvature and Torsion 169

Exercise 6.3. Prove that a regular curve γ(t) is a straight line if and only if its
curvature κ is identically zero.

6.1.4. Curvature Formula. Although the curvature is defined as κ(s) = |γ 00 (s)|

where γ(s) is an arc-length parametrization of the curve, it is very often impractical to
compute the curvature this way. The main reason is that the arc-length parametrizations
of many paths are very difficult to find explicitly. A “notorious” example is the ellipse:

γ(t) = (a cos t, b sin t)

where a and b are positive constants with a 6= b. The arc-length function is given by:
Z tp
s(t) = a2 sin2 τ + b2 cos2 τ dτ.
0

While it is very easy to compute the integral when a = b, there is no closed form or
explicit anti-derivative for the integrand if a 6= b. Although the arc-length parametrization
exists theoretically speaking (Theorem 6.3), it cannot be written down explicitly and so
the curvature cannot be computed from the definition.
The purpose of this section is to derive a formula for computing curvature without
the need of finding its arc-length parametrization. To begin, we first prove the following
important observation:

Lemma 6.7. Let γ(s) : I → Rn be a curve parametrized by arc-length, then the velocity
γ 0 (s) and the acceleration γ 00 (s) is always orthogonal for any s ∈ I.

Proof. Since γ(s) is parametrized by arc-length, we have |γ 0 (s)| = 1 for any s, and so:

d 0 2 d
|γ (s)| = 1=0
ds ds
d 0
(γ (s) · γ 0 (s)) = 0
ds
γ 00 (s) · γ 0 (s) + γ 0 (s) · γ 00 (s) = 0
2γ 00 (s) · γ 0 (s) = 0
γ 00 (s) · γ 0 (s) = 0

Therefore, γ 0 (s) is orthogonal to γ 00 (s) for any s. 

Proposition 6.8. Given any regular curve γ(t) in R3 , the curvature as a function of t can
be computed by the following formula:
|γ 0 (t) × γ 00 (t)|
κ(t) = 3 .
|γ 0 (t)|

Proof. Since γ(t) is a regular curve, there exists an arc-length parametrization γ(t(s)),
which for simplicity we denote it by γ(s). From now on, we denote γ 0 (t) as dγ(t) dt ,
0 dγ(s)
regarding t as the parameter of the curve, and γ (s) as ds regarding s as the parameter
of the curve.
170 6. Geometry of Curves

By the chain rule, we have:

dγ dγ ds ds
(6.1) = = γ 0 (s)
dt ds dt  dt
d2 γ
 
d dγ d 0 ds
(6.2) = = γ (s) (from (6.1))
dt2 dt dt dt dt
dγ 0 (s) ds d2 s
= + γ 0 (s) 2
dt dt dt
By the chain rule again, we get:
dγ 0 (s) dγ 0 (s) ds ds
= = γ 00 (s)
dt ds dt dt
Substitute this back to (6.2), we obtain:
 2
d2 γ 00 ds 0 d2 s
(6.3) = γ (s) + γ (s)
dt2 dt dt2
Taking the cross product of (6.1) and (6.3) yields:
 3  3
dγ d2 γ ds d2 s ds 0 ds
(6.4) × 2 = γ 0 (s)×γ 00 (s)+ 2 γ (s) × γ 0 (s) = γ 0 (s)×γ 00 (s).
dt dt dt dt
| dt {z } dt
=0
Since γ (s) and γ (s) are two orthogonal vectors by Lemma 6.7, we have |γ 0 (s) × γ 00 (s)| =
0 00

|γ 0 (s)| |γ 00 (s)| = κ(s). Taking the magnitude on both sides of (6.4), we get:
3
dγ d2 γ ds
× 2 =κ .
dt dt dt
Therefore, we get:
|γ 0 (t) × γ 00 (t)|
κ= .
ds 3
dt
The proof can be easily completed by the definition of s(t) and the Fundamental Theorem
of Calculus:
Z t
s= |γ 0 (τ )| dτ
0
ds
= |γ 0 (t)|
dt

Remark 6.9. Since the cross product is involved, Proposition 6.8 can only be used for
curves in R2 or R3 . To apply the result for curves in R2 , say γ(t) = (x(t), y(t)), one may
regard it as the curve γ(t) = (x(t), y(t), 0) in R3 . 

By Proposition 6.8, the curvature of the ellipse can be computed easily. See the
example below:
Example 6.10. Let γ(t) = (a cos t, b sin t, 0) be a parametrization of an ellipse on the
xy-plane where a and b are positive constants, then we have:
γ 0 (t) = (−a sin t, b cos t, 0)
γ 00 (t) = (−a cos t, −b sin t, 0)
γ 0 (t) × γ 00 (t) = (ab sin2 t + ab cos2 t) k̂ = ab k̂
Therefore, by Proposition 6.8, it’s curvature function is given by:
|γ 0 (t) × γ 00 (t)| ab
κ(t) = 3 = 2 2 .
|γ 0 (t)| (a sin t + b2 cos2 t)3/2
6.1. Curvature and Torsion 171

Exercise 6.4. Consider the graph of a smooth function y = f (x). Regarding the
graph as a curve in R3 , it can be parametrized using x as the parameter by γ(x) =
(x, f (x), 0). Show that the curvature of the graph is given by:
|f 00 (x)|
κ(x) = 3/2
.
(1 + f 0 (x)2 )

Exercise 6.5. For each of the following curves: (i) compute the curvature κ(t) using
Proposition 6.8; (ii) If it is easy to find an explicit arc-length parametrization of
the curve, compute also the curvature from the definition; (iii) find the (x, y, z)-
coordinates of the point(s) on the curve at which the curvature is the maximum.
(a) γ(t) = (3 cos t, 4 cos t, 5t).
(b) γ(t) = (t2 , 0, t).
(c) γ(t) = 2t, t2 , − 13 t3 .

6.1.5. Frenet-Serret Frame. For now on, we will concentrate on regular curves in
R3 . Furthermore, we consider mostly those curves whose curvature function κ is nowhere
vanishing. Therefore, straight-lines in R3 , or paths such as the graph of y = x3 , are
excluded in our discussion.

Definition 6.11 (Non-degenerate Curves). A regular curve γ(t) : I → R3 is said to be

non-degenerate if its curvature satisfies κ(t) 6= 0 for any t ∈ I.

We now introduce an important basis of R3 in the studies of space curves, the Frenet-
Serret Frame, or the TNB-frame. It is an orthonormal basis of R3 associated to each point
of a regular curve in R3 .

Definition 6.12 (Frenet-Serret Frame). Given a non-degenerate curve γ(s) : I → R3

parametrized by arc-length, we define:
T(s) := γ 0 (s) (tangent)
00
γ (s)
N(s) := (normal)
|γ 00 (s)|
B(s) := T(s) × N(s) (binormal)
The triple {T(s), N(s), B(s)} is called the Frenet-Serret Frame of R3 at the point γ(s) of
the curve. See Figure 6.2.

Remark 6.13. Note that T is a unit vector since γ(s) is arc-length parametrized. Recall
that κ(s) := |γ 00 (s)| and the curve γ(s) is assumed to be non-degenerate. Therefore, N
is well-defined for any s ∈ I and is a unit vector by its definition. From Lemma 6.7, T
and N are orthogonal to each other for any s ∈ I. Therefore, by the definition of cross
product, B is also a unit vector and is orthogonal to both T and N. To conclude, for each
fixed s ∈ I, the Frenet-Serret Frame is an orthonormal basis of R3 . 
 
Example 6.14. Let γ(s) = cos √s2 , sin √s2 , √s2 where s ∈ R. It can be verified easily
that it is arc-length parametrized, i.e. |γ 0 (s)| = 1 for any s ∈ R. The Frenet-Serret Frame
172 6. Geometry of Curves

Figure 6.2. Frenet-Serret frame

of this curve is given by:

 
1 s 1 s 1
T(s) = γ 0 (s) = − √ sin √ , √ cos √ , √
2 2 2 2 2
 
1 s 1 s
γ 00 (s) = − cos √ , − sin √ , 0
2 2 2 2
γ 00 (s)
 
s s
N(s) = 00 = − cos √ , − sin √ , 0
|γ (s)| 2 2
B(s) = T(s) × N(s)
î ĵ k̂
1 s 1 s 1
= − √2 sin √2 √2 cos √2 √2
s s
− cos √2 − sin √2 0
 
1 s 1 s 1
= √ sin √ , − √ cos √ , √
2 2 2 2 2


Definition 6.15 (Osculating Plane). Given a non-degenerate arc-length parametrized

curve γ(s) : I → R3 , the osculating plane Π(s) of the curve is a plane in R3 containing
the point represented by γ(s) and parallel to both T(s) and N, i.e.
Π(s) := γ(s) + span{T(s), N(s)}.
(See Figure 6.2)

Remark 6.16. By the definition of the Frenet-Serret Frame, the binormal vector B(s) is
a unit normal vector to the osculating plane Π(s). 

Exercise 6.6. Consider the curve γ(t) = (a cos t, a sin t, bt) where a and b positive
constants. First find its arc-length parametrization γ(s) := γ(t(s)), and then compute
its Frenet-Serret Frame.
6.1. Curvature and Torsion 173

Exercise 6.7. Show that if γ(s) : I → R3 is a non-degenerate arc-length parametrized

curve contained in the plane Ax + By + Cz = D where A, B, C and D are constants,
then T(s) and N(s) are parallel to the plane Ax + By + Cz = 0 for any s ∈ I, and
B(s) is a constant vector which is normal to the plane Ax + By + Cz = 0.

Exercise 6.8. [dC76, P.23] Let γ(s) be an arc-length parametrized curve in R3 . The
normal line at γ(s) is the infinite straight line parallel to N(s) passing through the
point represented by γ(s). Suppose the normal line at every γ(s) pass through a
fixed point p ∈ R3 . Show that γ(s) is a part of a circle.

6.1.6. Torsion. If the curve γ(s) : I → R3 is contained in a plane Π in R3 , then the

osculating plane Π(s) coincides the plane Π for any s ∈ I, and hence the binormal vector
B(s) is a unit normal vector to Π for any s ∈ I. By continuity, B(s) is a constant vector.
On the other hand, the helix considered in Example 6.14 is not planar since B(s) is
changing over s. As s increases, the osculating plane Π(s) not only translates but also
rotates. The magnitude of dBds is therefore a measurement of how much the osculating
plane rotates and how non-planar the curve γ(s) looks. It motivates the introduction of
torsion.
However, instead of defining the torsion of a curve to be dB ds , we hope to give a sign
for the torsion. Before we state the definition of torsion, we first prove the following fact:

Lemma 6.17. Given any non-degenerate, arc-length parametrized curve γ(s) : I → R3 ,

the vector dB
ds must be parallel to the normal N(s) for any s ∈ I.

Proof. First note that {T(s), N(s), B(s)} is an orthonormal basis of R3 for any s ∈ I.
Hence, we have:
dB(s)
= a(s)T(s) + b(s)N(s) + c(s)B(s)
ds
where a(s) = dB(s)
ds · T(s), b(s) =
dB(s)
ds · N(s) and c(s) = dB(s)
ds · B(s). It suffices to show
a(s) = c(s) = 0 for any s ∈ I.
d 2
Since B(s) is unit, one can easily see that c(s) ≡ 0 by considering ds |B| (c.f. Lemma
6.7). To show a(s) ≡ 0, we consider the fact that:

T(s) · B(s) = 0 for any s ∈ I.

Differentiate both sides with respect to s, we get:

dT dB
(6.5) ·B+T· = 0.
ds ds
d 0
Since dT
ds = ds γ (s) = γ 00 (s) = κN, we get dT
ds · B = 0 by the definition of B.
Combining this result with (6.5), we get a(s) = T· dB
ds = 0. Hence we have
dB
ds = b(s)N
and it completes the proof. 

Definition 6.18 (Torsion). Let γ(s) : I → R3 be an arc-length parametrized, non-

degenerate curve. The torsion of the curve is a function τ : I → R defined by:
dB
τ (s) := − · N.
ds
174 6. Geometry of Curves

dB
Remark 6.19. By Lemma 6.17, the vector ds and N are parallel. Combining with the
fact that N is unit, one can see easily that:
dB dB
|τ (s)| = |N| cos 0 = .
ds ds
Therefore, the torsion can be regarded as a signed dB
ds which measures the rate that the
osculating plane rotates as s increases (see Figure 6.3). The negative sign appeared in
the definition is a historical convention. 

Figure 6.3. Torsion measures how fast the osculating plane changes along a curve

 
Example 6.20. Consider the curve γ(s) = cos √s2 , sin √s2 , √s2 which is the helix
appeared in Example 6.14. The normal and binormal were already computed:
γ 00 (s)
 
s s
N(s) = 00 = − cos √ , − sin √ , 0
|γ (s)| 2 2
 
1 s 1 s 1
B(s) = √ sin √ , − √ cos √ , √ .
2 2 2 2 2
Taking the derivative, we get:
 
dB 1 s 1 s
= cos √ , sin √ , 0 .
ds 2 2 2 2
Therefore, the torsion of the curve is:
dB 1
τ (s) = − ·N= .
ds 2


Exercise 6.9. Consider the curve γ(t) = (a cos t, a sin t, bt) where a and b are
positive constants. Find its torsion τ (s) as a function of the arc-length parameter s.

Exercise 6.10. Let γ(s) : I → R3 be a non-degenerate, arc-length parametrized

curve. Prove that τ (s) = 0 for any s ∈ I if and only if γ(s) is contained in a plane.
[Hint: for the “only if” part, consider the dot product B · T.]
6.1. Curvature and Torsion 175

Exercise 6.11. [dC76, P.25] Suppose γ(s) : I → R3 is a non-degenerate, arc-length

parametrized curve such that τ (s) 6= 0 and κ0 (s) 6= 0 for any s ∈ I. Show that the
 2
1 d 1 1
curve lies on a sphere if and only if 2 + is a constant.
κ ds κ τ2

The torsion of a non-degenerate curve γ(t) can be difficult to compute from the
definition since it involves finding an explicit arc-length parametrization. Fortunately,
just like the curvature, there is a formula for computing torsion.

Proposition 6.21. Let γ(t) : I → R3 be a non-degenerate curve, then the torsion of the
curve is given by:
(γ 0 (t) × γ 00 (t)) · γ 000 (t)
τ (t) = 2 .
|γ 0 (t) × γ 00 (t)|

Proof. See Exercise #6.12. 

Exercise 6.12. The purpose of this exercise is to give a proof of Proposition 6.21. As
γ(t) is a (regular) non-degenerate curve, there exist an arc-length parametrization
γ(s) := γ(t(s)) and a Frenet-Serret Frame {T(s), N(s), B(s)} at every point on the
curve. With a little abuse of notations, we denote κ(s) := κ(t(s)) and τ (s) := τ (t(s)).
(a) Show that
(γ 0 (s) × γ 00 (s)) · γ 000 (s)
τ (s) = .
κ(s)2
(b) Using (6.3) in the proof of Proposition 6.8, show that
 3
ds
γ 000 (t) = γ 000 (s) + v(s)
dt
where v(s) is a linear combination of γ 0 (s) and γ 00 (s) for any s.
(c) Hence, show that
(γ 0 (t) × γ 00 (t)) · γ 000 (t)
(γ 0 (s) × γ 00 (s)) · γ 000 (s) = 6 .
|γ 0 (t)|
[Hint: use (6.4) in the proof of Proposition 6.8.]
(d) Finally, complete the proof of Proposition 6.21. You may use the curvature
formula proved in Proposition 6.8.

Exercise 6.13. Compute the torsion τ (t) for the ellipsoidal helix:
γ(t) = (a cos t, b sin t, ct)
where a and b are positive and c is non-zero.
176 6. Geometry of Curves

6.2. Fundamental Theorem of Space Curves

In this section, we discuss a deep result about non-degenerate curves in R3 . Given an
arc-length parametrized, non-degenerate curve γ(s), one can define its curvature κ(s)
and torsion τ (s) as discussed in the previous section. They are scalar-valued functions
of s. The former must be positive-valued, while the latter can take any real value. Both
functions are smooth.
Now we ask the following questions:

Existence: If we are given a pair of smooth real-valued functions α(s)

and β(s) defined on s ∈ I where α(s) > 0 for any s ∈ I, does there
exist a regular curve γ(s) : I → R3 such that its curvature κ(s) is
identically equal to α(s), and its torsion τ (s) is identically equal to
β(s)?
Uniqueness: Furthermore, if there are two curves γ(s) and γ̄(s) in
R3 whose curvature are both identical to α(s), and torsion are both
identical to β(s), then is it necessary that γ(s) ≡ γ̄(s)?

The Fundamental Theorem of Space Curves answers both questions above. Using the
classic existence and uniqueness theorems in Ordinary Differential Equations (ODEs), one
can give an affirmative answer to the above existence question – yes, such a curve exists –
and an “almost” affirmative answer to the uniqueness question – that is, although the
curves γ(s) and γ̄(s) may not be identical, one can be transformed from another by a rigid
body motion in R3 . The proof of this theorem is a good illustration of how Differential
Equations interact with Differential Geometry – nowadays a field called Geometric Analysis.

FYI: Geometric Analysis

Geometric Analysis is a modern field in mathematics which uses Differential Equations to
study Differential Geometry. In the past few decades, there are several crowning achieve-
ments in this area. Just to name a few, these include Yau’s solution to the Calabi Conjecture
(1976), and Hamilton–Perelman’s solution to the Poincaré Conjecture (2003), and Brendle–
Schoen’s solution to the Differentiable Sphere Theorem (2007).

6.2.1. Existence and Uniqueness of ODEs. A system of ODEs (or ODE system) is
a set of one or more ODEs. The general form of an ODE system is:

x01 (t) = f1 (x1 , x2 , . . . , xn , t)

x02 (t) = f2 (x1 , x2 , . . . , xn , t)
..
.
x0n (t) = fn (x1 , x2 , . . . , xn , t)

where t is the independent variable, xi (t)’s are unknown functions, and fj ’s are prescribed
functions of (x1 , . . . , xn , t) from Rn × I → R.
An ODE system with a given initial condition, such as (x1 (0), . . . , xn (0)) =
(a1 , . . . , an ) where ai ’s are constants, is called an initial-value problem (IVP).
We first state a fundamental existence and uniqueness theorem for ODE systems:
6.2. Fundamental Theorem of Space Curves 177

Theorem 6.22 (Existence and Uniqueness Theorem of ODEs). Given functions fi ’s

(1 ≤ i ≤ n) defined on Rn × I, we consider the initial-value problem:
x0i (t) = fi (x1 , . . . , xn , t) for 1 ≤ i ≤ n
with initial condition (x1 (0), . . . , xn (0)) = (a1 , . . . , an ). Suppose for every 1 ≤ i, j ≤ n,
∂fi
the first partial derivative ∂x j
exists and is continuous on Rn × I , then there exists a
unique solution (x1 (t), . . . , xn (t)), defined at least on a short-time interval t ∈ (−ε, ε),
to the initial-value problem. Furthermore, as long as the solution remains bounded, the
solution exists for all t ∈ I.

Proof. MATH 4051. 

6.2.2. Frenet-Serret System. Given an arc-length parametrized and non-degenerate

curve γ(s) : I → R3 , recall that tangent and binormal satisfy:
T0 (s) = κ(s)N(s)
B0 (s) = −τ (s)N(s).
Using the fact that N = B × T, one can also compute:
N0 (s) = B0 (s) × T(s) + B(s) × T0 (s)
= −τ (s)N(s) × T(s) + B(s) × κ(s)N(s)
= −κ(s)T(s) + τ (s)B(s).
The Frenet-Serret System is an ODE system for the Frenet-Serret Frame of a non-
degenerate curve γ(s):
T0 = κN
N0 = −κT +τ B
B0 = −τ N
or equivalently in matrix form:
 0   
T 0 κ 0 T
(6.6) N = −κ 0 τ  N
B 0 −τ 0 B
Since each vector of the {T, N, B} frame has three components, therefore the Frenet-
Serret System (6.6) is an ODE system of 9 equations with 9 unknown functions.

6.2.3. Fundamental Theorem. We now state the main theorem of this section:

Theorem 6.23 (Fundamental Theorem of Space Curves). Given any smooth positive
function α(s) : I → (0, ∞), and a smooth real-valued function β(s) : I → R, there exists
an arc-length parametrized, non-degenerate curve γ(s) : I → R3 such that its curvature
κ(s) ≡ α(s) and its torsion τ (s) ≡ β(s).
Moreover, if γ̄(s) : I → R3 is another arc-length parametrized, non-degenerate curve
whose curvature κ̄(s) ≡ α(s) and torsion τ̄ (s) ≡ β(s), then there exists a 3 × 3 constant
matrix A with AT A = I, and a constant vector p, such that γ̄(s) = Aγ(s) + p for any
s ∈ I.

Proof. The existence part consists of three major steps.

Step 1: Use the existence theorem of ODEs (Theorem 6.22) to show there exists a
moving orthonormal frame {ê1 (s), ê2 (s), ê3 (s)} which satisfies an ODE system (see
(6.7) below) analogous to the Frenet-Serret System (6.6).
178 6. Geometry of Curves

Step 2: Show that there exists a curve γ(s) whose Frenet-Serret Frame is given by
T(s) = ê1 (s), N(s) = ê2 (s) and B(s) = ê3 (s). Consequently, from the system (6.7),
one can claim γ(s) is a curve that satisfies the required conditions.
Step 3: Prove the uniqueness part of the theorem.
Step 1: To begin, let’s consider the ODE system with unknowns ê1 , ê2 and ê3 :
 0   
ê1 (s) 0 α(s) 0 ê1 (s)
(6.7) ê2 (s) = −α(s) 0 β(s) ê2 (s)
ê3 (s) 0 −β(s) 0 ê3 (s)
which is an analogous system to the Frenet-Serret System (6.6). Impose the initial
conditions:
ê1 (0) = î, ê2 (0) = ĵ, ê3 (0) = k̂.
Recall that α(s) and β(s) are given to be smooth (in particular, continuously differ-
entiable). By Theorem 6.22, there exists a solution {ê1 (s), ê2 (s), ê3 (s)} defined on a
maximal interval s ∈ (T− , T+ ) that satisfies the system with the above initial conditions.
Note that {ê1 (s), ê2 (s), ê3 (s)} is orthonormal initially at s = 0, we claim it remains
so as long as solution exists. To prove this, we first derive (see Exercise 6.14):
    
ê1 · ê1 0 0 0 2α 0 0 ê1 · ê1
ê2 · ê2   0 0 −2α 2β
0 0  ê2 · ê2 
 
  
d ê3 · ê3   0
   0 0 0 −2β 0  ê3 · ê3 
 
(6.8) = 
ds ê1 · ê2  −α α 0
  0 0 β  ê1 · ê2 
 
ê2 · ê3   0 −β β 0 0 −α ê2 · ê3 
ê3 · ê1 0 0 0 −β α 0 ê3 · ê1

Exercise 6.14. Verify (6.8).

Regarding êi · êj ’s are unknowns, (6.8) is a linear ODE system of 6 equations with
initial conditions:
(ê1 · ê1 , ê2 · ê2 , ê3 · ê3 , ê1 · ê2 , ê2 · ê3 , ê3 · ê1 )s=0 = (1, 1, 1, 0, 0, 0)
It can be verified easily that the constant solution (1, 1, 1, 0, 0, 0) is indeed a solution
to (6.8). Therefore, by the uniqueness part of Theorem 6.22, we must have
(ê1 · ê1 , ê2 · ê2 , ê3 · ê3 , ê1 · ê2 , ê2 · ê3 , ê3 · ê1 ) = (1, 1, 1, 0, 0, 0)
for any s ∈ (T− , T+ ). In other words, the frame {ê1 (s), ê2 (s), ê3 (s)} is orthonormal as
long as solution exists.
Consequently, each of {ê1 (s), ê2 (s), ê3 (s)} remains bounded, and by last statement
of Theorem 6.22, this orthonormal frame can be extended so that it is defined for all
s ∈ I.
Step 2: Using the frame ê1 (s) : I → R3 obtained in Step 1, we define:
Z s
γ(s) = ê1 (s) ds.
0
Evidently, γ(s) is a curve starting from the origin at s = 0. Since ê1 (s) is continuous, γ(s)
is well-defined on I and by the Fundamental Theorem of Calculus, we get:
T(s) := γ 0 (s) = ê1 (s)
which is a unit vector for any s ∈ I. Therefore, γ(s) is arc-length parametrized.
Next we verify that γ(s) is the curve require by computing its curvature and torsion.
By (6.8),
γ 00 (s) = ê01 (s) = α(s)ê2 (s)
6.2. Fundamental Theorem of Space Curves 179

By the fact that ê2 (s) is unit, we conclude that:

κ(s) = |γ 00 (s)| = α(s)
and so N(s) = κ(s) 1
γ 00 (s) = ê2 (s). For the binormal, we observe that ê3 = ê1 × ê2
initially at s = 0 and that the frame {ê1 (s), ê2 (s), ê3 (s)} remains to be orthonormal
for all s ∈ I, we must have ê3 = ê1 × ê2 for all s ∈ I by continuity. Therefore,
B(s) = T(s) × N(s) = ê1 (s) × ê2 (s) = ê3 (s) for any s ∈ I. By (6.8), we have:
B0 (s) = ê03 (s) = −β(s)ê2 (s) = −β(s)N.
Therefore, τ (s) = −B0 (s) · N(s) = β(s).
Step 3: Now suppose there exists another curve γ̄(s) : I → R3 with the same curvature
and torsion as γ(s). Let {T̄(s), N̄(s), B̄(s)} be the Frenet-Serret Frame of γ̄(s). We define
the matrix:  
A = T̄(0) N̄(0) B̄(0) .
By orthonormality, one can check that AT A = I. We claim that γ̄(s) = Aγ(s) + γ̄(0) for
any s ∈ I using again the uniqueness theorem of ODEs (Theorem 6.22).
First note that A is an orthogonal matrix, so the Frenet-Serret Frame of the trans-
formed curve Aγ(s) + γ̄(0) is given by {AT(s), AN(s), AB(s)} and the frame satisfies
the ODE system:
 0   
AT(s) 0 α(s) 0 AT(s)
AN(s) = −α(s) 0 β(s) AN(s)
AB(s) 0 −β(s) 0 AB(s)
since the Frenet-Serret Frame {T(s), N(s), B(s)} does.
Furthermore, the curve γ̄(s) also has curvature α(s) and torsion β(s), so its Frenet-
Serret Frame {T̄(s), N̄(s), B̄(s)} also satisfies the ODE system:
 0   
T̄(s) 0 α(s) 0 T̄(s)
N̄(s) = −α(s) 0 β(s) N̄(s) .
B̄(s) 0 −β(s) 0 B̄(s)
Initially at s = 0, the two Frenet-Serret Frames are equal by the definition of A and
choice of êi (0)’s in Step 1:
 
AT(0) = T̄(0) N̄(0) B̄(0) î = T̄(0)
 
AN(0) = T̄(0) N̄(0) B̄(0) ĵ = N̄(0)
 
AB(0) = T̄(0) N̄(0) B̄(0) k̂ = B̄(0)
By the uniqueness part of Theorem 6.22, the two frames are equal for all s ∈ I. In
particular, we have:
AT(s) ≡ T̄(s).
Finally, to show that γ̄(s) ≡ Aγ(s) + γ̄(0), we consider the function
2
f (s) := |γ̄(s) − (Aγ(s) − γ̄(0))| .
Taking its derivative, we get:
f 0 (s) = 2 (γ̄ 0 (s) − Aγ 0 (s)) · (γ̄(s) − (Aγ(s) − γ̄(0)))


= 2 T̄(s) − AT(s) · (γ̄(s) − (Aγ(s) − γ̄(0)))
| {z }
=0
=0
for any s ∈ I. Since f (0) = 0 initially by the fact that γ(0) = 0, we have f (s) ≡ 0 and so
γ̄(s) ≡ Aγ(s) + γ̄(0), completing the proof of the theorem. 
180 6. Geometry of Curves

The existence part of Theorem 6.23 only shows a curve with prescribed curvature
and torsion exists, but it is in general difficult to find such a curve explicitly. While the
existence part does not have much practical use, the uniqueness part has some nice
corollaries.
First recall that a helix is a curve of the form γa,b (t) = (a cos t, a sin t, bt) where
a 6= 0 and b can be any real number. It’s arc-length parametrization is given by:
 
s s bs
γa,b (s) = a cos √ , a sin √ , √ .
a2 + b2 a2 + b2 a2 + b2
It can be computed that its curvature and torsion are both constants:
a
κa,b (s) ≡ 2
a + b2
b
τa,b (s) ≡ 2 .
a + b2
Conversely, given two constants κ0 > 0 and τ0 ∈ R, by taking a = κ2κ+τ 0
2 and
0 0
τ0
b = κ2 +τ 2 , the helix γa,b (s) with this pair of a and b has curvature κ0 and torsion τ0 .
0 0
Hence, the uniqueness part of Theorem 6.23 asserts that:

Corollary 6.24. A non-degenerate curve γ(s) has constant curvature and torsion if and
only if γ(s) is congruent to one of the helices γa,b (s).

Remark 6.25. Two space curves γ(s) and γ e(s) are said to be congruent if there exists a
3 × 3 orthogonal matrix A and a constant vector p ∈ R3 such that γ e(s) = Aγ(s) + p. In
simpler terms, one can obtain γ
e(s) by rotating and translating γ(s). 
6.3. Plane Curves 181

6.3. Plane Curves

A plane curve γ(s) is an arc-length parametrized curve in R2 . While it can be considered
as a space curve by identifying R2 and the xy-plane in R3 , there are several aspects of
plane curves that make them distinguished from space curves.

6.3.1. Signed Curvature. Given an arc-length parametrized curve γ(s) : I → R2 ,

we define the tangent frame T(s) as in space curves, i.e.
T(s) = γ 0 (s).
1 0
However, instead of defining the normal frame N(s) = κ(s) T (s), we use the frame JT(s)
where J is the counter-clockwise rotation by π2 , i.e.
 
0 −1
J= .
1 0

One can easily check that {T(s), JT(s)} is an orthonormal frame of R2 for any s ∈ I.
Let’s call this the TN-Frame of the curve. We will work with the TN-Frame in place of the
Frenet-Serret Frame for plane curves. The reasons for doing so are two-folded. For one
thing, the normal frame. JT(s) is well-defined for any s ∈ I even though κ(s) is zero for
some s ∈ I. Hence, one can relax the non-degeneracy assumption here. For another, we
can introduce the signed curvature k(s):

Definition 6.26 (Signed Curvature). Given an arc-length parametrized plane curve

γ(s) : I → R2 , the signed curvature k(s) : I → R is defined as:
k(s) := T0 (s) · JT(s).

Note that T(s) is unit, so by Lemma 6.7 we know T(s) and T0 (s) are always or-
thogonal and hence it is either in or against the direction of JT(s). Therefore, we
have
|k(s)| = |T0 (s)| |JT(s)| = |γ 00 (s)| = κ(s).
The sign of k(s) is determined by whether T0 and ĵT are along or against each other (see
Figure 6.4).

Figure 6.4. Signed curvature

Example 6.27. Let’s compute the signed curvature of the unit circle
γ(s) = (cos εs, sin εs)
182 6. Geometry of Curves

where ε = ±1. The curve is counter-clockwise orientable when ε = 1, and is clockwise

orientable when ε = −1. Clearly it is arc-length parametrized, and
T(s) = γ 0 (s) = (−ε sin εs, ε cos εs)
JT(s) = (−ε cos εs, −ε sin εs)
k(s) = T0 (s) · JT(s)
= (−ε2 cos εs, −ε2 sin εs) · (−ε cos εs, −ε sin εs)
= ε3 = ε.


Exercise 6.15. Consider a plane curve γ(s) parametrized by arc-length. Let θ(s) be
the angle between the x-axis and the unit tangent vector T(s). Show that:
T0 (s) = θ0 (s)JT(s) and k(s) = θ0 (s).

Exercise 6.16. [dC76, P.25] Consider a plane curve γ(s) parametrized by arc-length.
Suppose |γ(s)| is maximum at s = s0 . Show that:
1
|k(s0 )| ≥ .
|γ(s0 )|

Given a regular plane curve γ(t) : I → R2 , not necessarily arc-length parametrized.

Denote the components of the curve by γ(t) = (x1 (t), x2 (t)).
(a) Show that its signed curvature (as a function of t) is given by:
γ 00 (t) · Jγ 0 (t) x01 (t)x002 (t) − x02 (t)x001 (t)
k(t) = 3 = 3/2
.
|γ 0 (t)| (x01 (t)2 + x02 (t)2 )
(b) Hence, show that the graph of a smooth function y = f (x), when considered as a
curve parametrized by x, has signed curvature given by:
f 00 (x)
k(x) = .
(1 + f 0 (x)2 )3/2
The signed curvature characterizes regular plane curves, as like curvature and torsion
characterize non-degenerate space curves.

Theorem 6.28 (Fundamental Theorem of Plane Curves). Given any smooth real-valued
function α(s) : I → R, there exists a regular plane curve γ(s) : I → R2 such that its signed
curvature k(s) ≡ α(s). Moreover, if γ̄(s) : I → R2 is another regular plane curve such
that its signed curvature k̄(s) ≡ α(s), then there exists a 2 × 2 orthogonal matrix A and a
constant vector p ∈ R2 such that γ̄(s) ≡ Aγ(s) + p.

Proof. See Exercise #6.17. 

Exercise 6.17. Prove Theorem 6.28. Although the proof is similar to that of Theorem
6.23 for non-degenerate space curves, please do not use the latter to prove the former
in this exercise. Here is a hint on how to begin the proof: Consider the initial-value
problem
ê0 (s) = α(s) J ê(s)
ê(0) = î
6.3. Plane Curves 183

Exercise 6.18. Using Theorem 6.28, show that a regular plane curve has constant
signed curvature if and only if it is a straight line or a circle

Exercise 6.19. [Küh05, P.50] Find an explicit plane curve γ(s) such that the signed
curvature is given by k(s) = √1s .

6.3.2. Total Curvature. In this subsection, we explore an interesting result concern-

ing the signed curvature of a plane curve. We first introduce:

Definition 6.29 (Closed Curves). An arc-length parametrized plane curve γ(s) :

[0, L] → R2 is said to be closed if γ(0) = γ(L). It is said to be simple closed if γ(s)
is closed and if γ(s1 ) = γ(s2 ) for some si ∈ [0, L] then one must have s1 , s2 = 0 or L.

The following is a celebrated result that relates the local property (i.e. signed
curvature) to the global property (topology) of simple closed curves:

Theorem 6.30 (Hopf). For any arc-length parametrized, simple closed curve γ(s) :
[0, L] → R2 such that γ 0 (0) = γ 0 (L), we must have:
Z L
k(s) ds = ±2π.
0

The original proof was due to Hopf. We will not discuss Hopf’s original proof in this
course, but we will prove a weaker result, under the same assumption as Theorem 6.30,
that Z L
k(s) ds = 2πn
0
for some integer n.
Let {T(s), JT(s)} be the TN-frame of γ(s). Since T(s) is unit for any s ∈ [0, L], one
can find a smooth function θ(s) : [0, L] → R such that
T(s) = (cos θ(s), sin θ(s))
for any s ∈ [0, L]. Here θ(s) can be regarded as 2kπ + angle between T(s) and î. In
order to ensure continuity, we allow θ(s) to take values beyond [0, 2π].
Then JT(s) = (− sin θ(s), cos θ(s)), and so by the definition of k(s), we get:
k(s) = T0 (s) · JT(s)
= (−θ0 (s) sin θ(s), θ0 (s) cos θ(s)) · (− sin θ(s), cos θ(s))
= θ0 (s).
Therefore, the total curvature is given by:
Z L Z L
k(s) ds = θ0 (s) ds = θ(L) − θ(0).
0 0
Since it is assumed that T(0) = T(L) in Theorem 6.30, we have
θ(L) ≡ θ(0) (mod 2π)
and so we have: Z L
k(s) ds = 2πn
0
for some integer n.
 Chapter 7

Geometry of Euclidean
Hypersurfaces

“Geometry is the art of correct

reasoning from incorrectly drawn
figures.”

Henri Poincaré

Riemannian geometry is a branch in differential geometry which studies intrinsic

geometric structure without referencing to the ambient space. First developed by Gauss,
Riemann, et. al, it later became the mathematical foundation in Einstein’s general
relativity. Many important notions and concepts of Riemannian geometry are, however,
stemmed from extrinsic geometry. In this chapter, we will first explore ourselves to the
basic differential geometry of Euclidean hypersurfaces.

7.1. Regular Hypersurfaces in Euclidean Spaces

The prefix hyper- in the word “hypersurface” means the manifold is one dimensional
lower than the ambient space. A hypersurface in Rn+1 is a n-dimensional subset of
Rn+1 . As in the case of regular surfaces in R3 , we want to impose some conditions of a
hypersurface so that it becomes regular in the way we desire.

Definition 7.1 (Regular Hypersurfaces in Rn+1 ). Let Σn be a non-empty subset of

Rn+1 . Suppose Σn can be covered
[ by the image of a family of local parametrizations
A = {Fα : Uα → Σ}, i.e. Σ = Fα (Uα ), such that each Uα is an open set in Rn and
α
each Fα : Uα → Σ satisfies all three conditions below:
(1) Fα (u1α , · · · , unα ) is a C ∞ map from Uα to Rn+1 ;
(2) Fα is a homeomorphism between Uα and its image Oα := Fα (Uα ); and
n o
(3) The vectors ∂F
∂u
α
1 , · · · , ∂Fα
∂u n in Rn+1 are linearly independent on Uα .
α α

Then, we say that Σ is a regular hypersurface in Rn+1 .

n

185
186 7. Geometry of Euclidean Hypersurfaces

∂Fα
Each vector is a tangent to Σn at the based point p = Fα (u1α , · · · , unα ).
∂uiα (u1 ,··· ,un )
α α
With a bit abuse of notations and for simplicity, we will from now on denote
∂Fα ∂Fα
i
(p) :=
∂uα ∂uiα (u1α ,··· ,un
α)

where p = Fα (u1α , · · · , unα ), to emphasize that p is the based point of the tangent vector.
By condition (3) in Definition 7.1, we know that the following vector space
 
∂Fα ∂Fα
Tp Σn := span (p), · · · , (p)
∂u1α ∂unα
has dimension n. This is called the tangent space at p ∈ Σn .
Example 7.2. The graph Σf of any smooth function f : Rn → R is a regular hypersurface
in Rn+1 . One can parametrize Σf by a single parametrization:
F : Rn → Σf


(x1 , · · · , xn ) 7→ x1 , · · · , xn , f (x1 , · · · , xn )
By straight-forward computations, we have:
∂F ∂f
= êi + ên+1
∂xi ∂xi
n on
∂F
where {ê1 , · · · , ên+1 } is the standard basis vectors in Rn+1 . It is clear that ∂xi are
i=1
linearly independent. 
Example 7.3. The n-dimensional unit sphere:
( n+1
)
X
n n+1
S := (x1 , · · · , xn+1 ) ∈ R : x2i =1
i=1

is a regular hypersurface in Rn=1 . It can be parametrized by a pair of (inverse) stere-

ographic projections F+ : Rn → Sn \{(0, · · · , 0, 1)} and F− : Rn → Sn \{(0, · · · , 0, −1)}
with F+ given by
n 2
X 2ui êi |u| − 1
F+ (u1 , · · · , un ) = 2 + 2 ên+1
j=1 |u| + 1 |u| + 1
2
where |u| = u21 + · · · + u2n . We leave it as an exercise for readers to verify F+ satisfies
the conditions in Definition 7.1 and write down the south-pole map F− . 

Exercise 7.1. Fill in the omitted detail in Examples 7.2 and 7.3.

Exercise 7.2. Show that any regular hypersurface in Rn+1 is a smooth manifold,
and also a smooth submanifold of Rn+1 .
Regular hypersurfaces are the higher dimensional generalization of regular surfaces
discussed in Chapter 1. As such many important results about regular surfaces can carry
over naturally to regular hypersurfaces. We state the results below and leave the proofs
as exercises.

Theorem 7.4 (c.f. Theorem 1.6). Let g(x1 , · · · , xn+1 ) : Rn+1 → R be a smooth function.
Consider a non-empty level set g −1 (c) where c is a constant. If ∇g(p) 6= 0 at all points
p ∈ g −1 (c), then the level set g −1 (c) is a regular hypersurface in Rn+1 .
7.1. Regular Hypersurfaces in Euclidean Spaces 187

Proposition 7.5 (c.f. Proposition 1.8). Assume all given conditions stated in Theorem
7.4. Furthermore, suppose F is a bijective map from an open set U ⊂ Rn to an open set
O ⊂ Σ := g −1 (c) which satisfies conditions (1) and (3) in Definition 7.1. Then, F satisfies
condition (2) as well and hence is a smooth local parametrization of g −1 (c).

Proposition 7.6 (c.f. Proposition 1.11). Let Σ ⊂ Rn+1 be a regular surface, and
Fα : Uα → M and Fβ : Uβ → M be two smooth local parametrizations of Σ with
overlapping images, i.e. W := Fα (Uα ) ∩ Fβ (Uβ ) 6= ∅. Then, the transition maps defined
below are also smooth maps:
(Fβ−1 ◦ Fα ) : Fα−1 (W) → Fβ−1 (W)
(Fα−1 ◦ Fβ ) : Fβ−1 (W) → Fα−1 (W)

Exercise 7.3. Prove Theorem 7.4, Proposition 7.5, and Proposition 7.6.
188 7. Geometry of Euclidean Hypersurfaces

7.2. Fundamental Forms and Curvatures

7.2.1. First Fundamental Form. In this subsection, we introduce an important
concept in differential geometry – the first fundamental form. Loosely speaking, it is the
dot product of the tangent vectors of a regular surface. It captures and encodes intrinsic
geometric information (such as curvature) about the hypersurface.

Definition 7.7 (First Fundamental Form). The first fundamental form of a regular
hypersurface Σn ⊂ Rn+1 is a 2-tensor g on Σ with local expression g = gij dui ⊗ duj
where gij is given by:  
∂F ∂F
gij = , ,
∂ui ∂uj
where h·, ·i denotes the usual dot product on Rn+1 .

Exercise 7.4. Show that gij dui ⊗ duj is independent of local coordinates, i.e. if
G(vα ) is another local parametrization and denote
 
∂G ∂G
geαβ = , ,
∂vα ∂vβ
then
geαβ dv α ⊗ dv β = gij dui ⊗ duj .

Exercise 7.5. Let ι : Σ → Rn+1 be an inclusion map from a regular hypersurface Σ.

Show that the first fundamental form of Σ can be expressed as:
n+1
!
X
∗ α α
g=ι dx ⊗ dx
α=1

where (xα ) is the standard coordinates of Rn+1 .

Example 7.8. Let S2 be the unit 2-sphere and F be the following local parametrization:
F (u, v) = (sin u cos v, sin u sin v, cos u), (u, v) ∈ (0, π) × (0, 2π)
By direct computations, we have:
∂F
= (cos u cos v, cos u sin v, − sin u)
∂u
∂F
= (− sin u sin v, sin u cos v, 0)
∂v
   
∂F ∂F ∂F ∂F ∂F ∂F ∂F ∂F
g , = · =1 g , = · =0
∂u ∂u ∂u ∂u ∂u ∂v ∂u ∂v
   
∂F ∂F ∂F ∂F ∂F ∂F ∂F ∂F
g , = · =0 g , = · = sin2 u
∂v ∂u ∂v ∂u ∂v ∂v ∂v ∂v
Therefore, we get its first fundamental equals
g = du ⊗ du + sin2 u dv ⊗ dv.

7.2. Fundamental Forms and Curvatures 189

Exercise 7.6. Show that the first fundamental form of the graph Σf in Example 7.2
is given by:
 
∂f ∂f
g = δij + dxi ⊗ dxj = δij dxi ⊗ dxj + df ⊗ df.
∂xi ∂xj

As gij = gji , the tensor g is symmetric. As such the tensor notation gij dui ⊗ duj is
often written simply as gij dui duj . For instance, the first fundamental form of the unit
sphere in Example 7.8 can be expressed as:
g = du2 + sin2 u dv 2
where du2 is interpreted as (du)2 = du du, not d(u2 ).
Another way to represent the first fundamental form is by the matrix:
 
g11 g12 · · · g1n
 g21 g22 · · · g2n 
[g] :=  . .. ..  .
 
 .. . . 
gn1 gn2 · · · gnn
   
∂F ∂F ∂F ∂F
It is a symmetric matrix since g ∂u i
, ∂uj = g ∂uj ∂ui .
,
∂F ∂F
Given two tangent vectors Y = Y i ∂u i
and Z = Z i ∂u i
in Tp Σ, the value of g(Y, Z) is
related to the entries of [g] in the following way:
 
∂F ∂F
g(Y, Z) = g Y i , Zj = Y i Z j gij
∂ui ∂uj
 
g11 g12 · · · g1n  1 
Z
 1   g21 g22 · · · g2n 
n    .. 
= Y ··· Y  . .. ..   . 
 .. . . 
Zn
gn1 gn2 · · · gnn
Note that the matrix [g] depends on local coordinates although the tensor g does not.
As computed in Example 7.8, the matrix [g] of the unit sphere (with respect the
parametrization F used in the example) is given by:
 
1 0
[g] := 2
0 sin u
Evidently, it is a diagonal matrix. Try to think about the geometric meaning of [g] being
diagonal.
We will see in subsequent sections that g “encodes” crucial geometric information
such as curvatures. There are also some familiar geometric quantities, such as length
and area, which are related to the first fundamental form g.
Consider a curve γ on a regular hypersurface Σn ⊂ Rn+1 parametrized by F (ui ).
Suppose the curve can be parametrized by γ(t), a < t < b, then from calculus we know
the arc-length of the curve is given by:
Z b
L(γ) = |γ 0 (t)| dt
a
In fact one can express this above in terms of g. The argument is as follows:
Suppose γ(t) has local coordinates coordinates (γ i (t)) such that F (γ i (t)) = γ(t).
Using the chain rule, we then have:
∂F dγ i
γ 0 (t) = .
∂ui dt
190 7. Geometry of Euclidean Hypersurfaces

p
Recall that |γ 0 (t)| = hγ 0 (t), γ 0 (t)i and that γ 0 (t) lies on Tp M , we then have:
p
|γ 0 (t)| = g (γ 0 (t), γ 0 (t))
We can then express it in terms of the matrix components gij ’s:
dγ i dγ j
(7.1) g (γ 0 (t), γ 0 (t)) = gij
dt dt
where gij ’s are evaluated at the point γ(t). Therefore, the arc-length can be expressed in
terms of the first fundamental form by:
Z bp Z br
dγ i dγ j
L(γ) = g (γ 0 (t), γ 0 (t)) dt = gij dt
a a dt dt
Another familiar geometric quantity which is also related to g is the area of a surface.
For simplicity, we focus on dimension 2 first. Suppose a regular surface Σ can be almost
everywhere parametrized by F (u, v) with (u, v) ∈ D ⊂ R2 where D is a bounded domain,
the area of this surface is given by:
ZZ
∂F ∂F
A(M ) = × dudv
D ∂u ∂v

It is also possible to express ∂F ∂F

∂u × ∂v in terms of the first fundamental form g. Let θ
be the angle between the two vectors ∂u and ∂F
∂F
∂v , then from elementary vector geometry,
we have:
2 2 2
∂F ∂F ∂F ∂F
× = sin2 θ
∂u ∂v ∂u ∂v
2 2 2 2
∂F ∂F ∂F ∂F
= − cos2 θ
∂u ∂v ∂u ∂v
2 2  2
∂F ∂F ∂F ∂F
= − ·
∂u ∂v ∂u ∂v
    2
∂F ∂F ∂F ∂F ∂F ∂F
= · · − ·
∂u ∂u ∂v ∂v ∂u ∂v
= g11 g22 − (g12 )2
= det[g].
Therefore,
ZZ p
(7.2) A(M ) = det[g] dudv
D

In fact, for higher dimensional regular hypersurfaces parametrized almost everywhere

by F (ui ) : U → Σ, we also have its area equals
Z p
det[g] du1 · · · dun .
U

Example 7.9. Let Σf be the graph of a smooth function f : U → R defined on an open

subset U of Rn , then Σf has a globally defined smooth parametrization:
F (u1 , · · · , un ) = (u1 , · · · , un , f (u1 , · · · , un )).
By straight-forward computations, we can get:
∂f ∂f
gij = δij + ,
∂ui ∂uj
7.2. Fundamental Forms and Curvatures 191

Therefore, the length the curve γ(t) := F (γ 1 (t), · · · , γ n (t)) can be computed by
integrating the square root of:
n  2 n
dγ i dγ j X dγ i X ∂f dγ i ∂f dγ j
gij = +
dt dt i=1
dt i,j=1
∂ui dt ∂uj dt
n 2 2
dγ i
  
X d
= + f (γ(t)) .
i=1
dt dt

To compute the surface area of a region F (Ω) ⊂ Σf where Ω is a bounded domain

on U, we first compute det[g]. The matrix [g] can be written as
[g] = In×n + (∇f )(∇f )T .
By standard linear algebra1, we know that its eigenvalues are
2
1 + |∇f | , 1, · · · , 1,
2
and so det[g] = 1 + |∇f | . According to (7.2), we have:
ZZ q
2
A(F (Ω)) = 1 + |∇f | du1 · · · dun
Ω
which is exactly the same as what you have seen in multivariable calculus. 

7.2.2. Second Fundamental Form. In this subsection we introduce another im-

portant 2-tensor on Tp Σ, the second fundamental form h. While the first fundamental
form g encodes information about angle, length and area, the second fundamental form
encodes information about various curvatures.
We will see in subsequent sections that curvatures of a regular hypersurface are,
roughly speaking, determined by rate of changes of tangent and normal vectors just like
the case for regular curves. Let’s first talk about the normal vector – or in differential
geometry jargon – the Gauss Map.
Given a regular hypersurface Σ in Rn+1 with F (ui ) : U ⊂ R2 → Σ as one of its
local parametrization. Let p ∈ M , the orthogonal complement of the tangent space Tp Σ
in Rn+1 is a 1-dimensional vector space denoted by Np Σ. There are exactly two unit
vectors in Np Σ. In dimension 2 (i.e. regular surfaces), ∂F ∂F
∂u (p) and ∂v (p) are two linearly
independent tangents at p, and so the two unit normals are given by
∂F ∂F ∂F ∂F
∂u (p) × ∂v (p) ∂u (p) × ∂v (p)
ν(p) = ∂F ∂F
or − ∂F ∂F
∂u (p) × ∂v (p) ∂u (p) × ∂v (p)
See the result from Exercise 4.6 for higher dimensional hypersurfaces.
Example 7.10. Consider the unit sphere S2 (1) with smooth local parametrization:
F (u, v) = (sin u cos v, sin u sin v, cos u), (u, v) ∈ (0, π) × (0, 2π)
It is straight-forward to compute that:
∂F ∂F
= sin2 u cos v, sin2 u sin v, sin u cos u


×
∂u ∂v
∂F ∂F
× = sin u
∂u ∂v
ν(u, v) = (sin u cos v, sin u sin v, cos u) = F (u, v)
This unit normal vector ν is outward-pointing. 

1Note that ∇f is an eigenvector with eigenvalue 1 + |∇f |2 , and (∇f )(∇f )T has rank 1.
192 7. Geometry of Euclidean Hypersurfaces

Given a regular hypersurface, there are always two choices of normal vector at each
point. For a sphere, once the normal vector direction is chosen, it is always consistent
with your choice when we move the normal vector across the sphere. That is, when you
draw a closed path on the sphere and see how the unit normal vector varies along the
path, you will find that the unit normal remains the same when you come back to the
original point. We call it an orientable hypersurface if there exists a continuous choice
of unit normal vector across the whole hypersurface. See Chapter 4 for more thorough
discussions on orientability.
When Σ is an orientable regular hypersurface, the chosen unit normal vector ν can
then be regarded as a map. The domain of ν is Σ. Since ν is unit, the codomain can be
taken to be the unit sphere Sn . We call this map as:

Definition 7.11 (Gauss Map). Suppose Σ is an orientable regular hypersurface. The

Gauss map of Σ is a smooth function ν : Σ → Sn such that for any p ∈ Σ, the output
ν(p) is a unit normal vector of Σ at p. Here Sn is the unit n-sphere in Rn+1 .

As computed in Example 7.10, the Gauss map ν for the unit sphere S2 is given by F
(assuming the outward-pointing convention is observed). It is not difficult to see that
the Gauss map ν for a sphere with radius R centered the origin in R3 is given by R1 F .
Readers should verify this as an exercise.
For a plane Π, the unit normal vector at each point is the same. Therefore, the Gauss
map ν(p) is a constant vector independent of p.
A unit cylinder with z-axis as its central axis can be parametrized by:
F (u, v) = (cos u, sin u, v), (u, v) ∈ (0, 2π) × R.
By straight-forward computations, one can get:
∂F ∂F
× = (cos u, sin u, 0)
∂u ∂v
which is already unit. Therefore, the Gauss map of the cylinder is given by:
ν(u, v) = (cos u, sin u, 0).
2
The image of ν in S is the equator.
It is not difficult to see that the image of the Gauss map ν, which is a subset of S2 , is
related to how “spherical” or “planar” the surface looks. The smaller the image, the more
planar it is.
The curvature of a regular curve is a scalar function κ(p). Since a curve is one
dimensional, we can simply use one single value to measure the curvature at each point.
However, a regular hypersurface has arbitrary dimensions and hence has higher degree
of freedom than curves. It may bend by a different extent along different direction. As
such, there are various notions of curvatures for regular hypersurfaces. We first talk
about the normal curvature, which is fundamental to many other notions of curvatures.
Let Σ be an orientable regular hypersurface with its Gauss map denoted by ν. At each
point p ∈ Σ, we pick a unit tangent vector ê in Tp Σ. Heuristically, the normal curvature
at p measures the curvature of the surface along a direction ê. Precisely, we define:

Definition 7.12 (Normal Curvature). Let Σ be an orientable regular hypersurface with

Gauss map ν. For each point p ∈ Σ, and any unit tangent vector ê ∈ Tp Σ, we let Πp (ν, ê)
be the plane in Rn+1 passing through p and parallel to both ν(p) and ê (see Figure 7.1).
Let γ be the curve of intersection of Σ and Πp (ν, ê). The normal curvature at p in
the direction of ê of Σ, denoted by kn (p, ê), is defined to be the signed curvature k(p) of
the curve γ at p with respect to the Gauss map ν.
7.2. Fundamental Forms and Curvatures 193

Remark 7.13. Since kn (p, ê) is defined using the Gauss map ν, which always comes with
two possible choice for any orientable regular surface, the normal curvature depends on
the choice of the Gauss map ν. If the opposite unit normal is chosen to be the Gauss map,
the normal curvature will differ by a sign.

Figure 7.1. normal curvature at p in a given direction ê

We will first make sense of normal curvatures through elementary examples, then
we will prove a general formula for computing normal curvatures.
Example 7.14. Let P be any plane in R3 . For any point p ∈ P and unit tangent ê ∈ Tp P ,
the plane Πp (ν, ê) must cut through P along a straight-line γ. Since γ has curvature 0,
we have:
kn (p, ê) = 0
for any p ∈ P and ê ∈ Tp P . See Figure 7.2a. 
Example 7.15. Let S2 (R) be the sphere with radius R centered at the origin in R3 with
Gauss map ν taken to be inward-pointing. For any point p ∈ S2 and ê ∈ Tp S2 (R), the
plane Πp (ν, ê) cuts S2 (R) along a great circle (with radius R). Since a circle with radius
R has constant curvature R1 , we have:
1
kn (p, ê) =
R
for any p ∈ S2 (R) and ê ∈ Tp S2 (R). See Figure 7.2b. 
194 7. Geometry of Euclidean Hypersurfaces

(a) plane (b) sphere

(c) cylinder

Figure 7.2. Normal curvatures of various surfaces

Example 7.16. Let M be the (infinite) cylinder of radius R with x-axis as the central axis
with outward-pointing Gauss map ν. Given any p ∈ M , if êx is the unit tangent vector
at p parallel to the x-axis, then the Πp (ν, êx ) cuts the cylinder M along a straight-line.
Therefore, we have:
kn (p, êx ) = 0
for any p ∈ M . See the blue curve in Figure 7.2c.
On the other hand, if êyz is a horizontal unit tangent vector at p, then Πp (ν, êyz ) cuts
M along a circle with radius R. Therefore, we have:
1
kn (p, êyz ) = −
R
for any p ∈ M . See the red curve in Figure 7.2c. Note that the tangent vector of the
curve is moving away from the outward-pointing ν. It explains the negative sign above.
For any other choice of unit tangent ê at p, the plane Πp (ν, ê) cuts the cylinder along
an ellipse, so the normal curvature along ê may vary between 0 and − R1 . 

In the above examples, the normal curvatures kn (p, ê) are easy to find since the
curve of intersection between Πp (ν, ê) and the surface is either a straight line or a circle.
Generally speaking, the curve of intersection may be of arbitrary shape such as an ellipse,
and sometimes it is not even easy to identify what curve it is. Fortunately, it is possible
to compute kn (p, ê) for any given unit tangent ê in a systematic way. Next we derive an
expression of kn (p, ê), which will motivate the definition of the second fundamental form
and Weingarten’s map (also known as the shape operator).
7.2. Fundamental Forms and Curvatures 195

Proposition 7.17. Let Σ be the regular hypersurface in Rn+1 with Gauss map ν. Fix
p ∈ Σ and a unit vector ê ∈ Tp Σ, then the normal curvature of Σ at p along ê is given by:
kn (p, ê) = − hê, Dê νi .
Furthermore, suppose ê can be locally expressed as
∂F
ê = X i ,
∂ui
then we have
   2 
∂F ∂ν i j ∂ F
kn (p, ê) = − , X X = , ν X iX j .
∂ui ∂uj ∂ui ∂uj

Proof. Let γ be the intersection curve between the plane Πp (ν, ê) and Σ. We parametrize
γ by arc-length s ∈ (−ε, ε) such that γ(0) = p and |γ 0 (s)| ≡ 1 on (−ε, ε). Then, γ 0 (s) is
orthogonal to ν(γ(s)), and so

kn (p, ê) = hγ 00 (s), ν(γ(s))is=0

 
0 d
= − γ (s), ν(γ(s))
ds s=0
= −hê, Dê νi.

The last step follows from the fact that γ 0 (0) = ê.
The local expression follows immediately from the fact that
∂ν
DX i ∂F ν = X i D ∂F ν = X i .
∂ui ∂ui ∂ui

shows thatEkn (p, ê) can be written as a quadratic form on X i ’s with

Proposition 7.17 D
∂2F
coefficients given by ∂ui ∂uj , ν . This motivates the second fundamental form:

Definition 7.18 (Second Fundamental Form). Given a regular hypersurface in Rn+1

with Gauss map ν. We define the second fundamental form of Σ at p to be the 2-tensor:
h(X, Y ) := −hX, DY νi.
Under a local coordinate system (u1 , · · · , un ), we denote its component as
     2 
∂F ∂F ∂F ∂ν ∂ F
hij := h , =− , = ,ν .
∂ui ∂uj ∂ui ∂uj ∂ui ∂uj

Remark 7.19. As such, the normal curvature of Σ at p along ê is given by

kn (p, ê) = h(ê, ê).

Remark 7.20. h(X, Y ) is tensorial by the fact that Df Y ν = f DY ν and hf X, Y i =

f hX, Y i.

Example 7.21. Let Σf be the graph of a smooth function f (u1 , u2 ) : U → R defined on

an open subset U of R2 , then Σf has a globally defined smooth parametrization:

F (u1 , u2 ) = (u1 , u2 , f (u1 , u2 )).

196 7. Geometry of Euclidean Hypersurfaces

By straight-forward computations, we can get:

   
∂F ∂f ∂F ∂f
= 1, 0, = 0, 1,
∂u1 ∂u1 ∂u2 ∂u2
2
∂2f 2
∂2f
   
∂ F ∂ F
= 0, 0, = 0, 0,
∂u21 ∂u21 ∂u1 ∂u2 ∂u1 ∂u2
2
∂2f 2
∂2f
   
∂ F ∂ F
= 0, 0, = 0, 0,
∂u2 ∂u1 ∂u2 ∂u1 ∂u22 ∂u22
In short, we have
∂2F ∂2f
 
= 0, 0, .
∂ui ∂uj ∂ui ∂uj
Let’s take the Gauss map ν to be:
 
∂f ∂f
∂F
× ∂F − ∂u , ∂u2 , 1
∂u1 ∂u2 1
ν= =r 2  2 .
∂F ∂F
×

∂f ∂f
∂u1 ∂u2 1 + ∂u 1
+ ∂u2

Then, the second fundamental form is given by:

 2  ∂2f
∂ F ∂ui ∂uj
hij = ,ν = q .
∂ui ∂uj 2
1 + |∇f |
2
The matrix whose (i, j)-th entry given by ∂u∂i ∂u
f
j
is commonly called the Hessian of
f , denoted by ∇∇f or Hess(f ). Using this notation, the matrix of second fundamental
form of Σf is given by:
Hess(f )
[h] = q .
2
1 + |∇f |


Exercise 7.7. Generalize Example 7.21 to higher dimensional graph in Rn+1 :

xn+1 = f (x1 , · · · , xn ).

Given that kn (p, ê) depends on the unit direction ê ∈ Tp Σ, it is then natural to ask
when kn (p, ê) achieves the maximum and minimum among all unit vectors ê in Tp Σ. It is
a simple optimization problem of critical points of the h(ê, ê) (as a function of ê) subject
to the condition g(ê, ê) = 1. We call the critical values of h(ê, ê) subject to g(ê, ê) = 1 to
be principal curvatures and we have the following important result:

Proposition 7.22. Given a regular hypersurface Σ in Rn+1 and fix a point p ∈ Σ, the
principal curvatures are eigenvalues of the linear map:
Sp : Tp Σ → Tp Σ
∂F
∂F
7→ g jk hki .
∂ui ∂uj
The map Sp is called the Weingarten’s map, or the shape operator.

∂F 2
Proof. Denote ê = X i ∂u i
, then kn (p, ê) = h(ê, ê) = hij X i X j and |ê| = gij X i X j . To
determine the principal curvatures, we use Lagrange’s multiplier to find the critical
values of hij X i X j under the constraint gij X i X j = 1. Here we treat (X 1 , · · · , X n ) as
the variables, while gij and hij are regarded as constants since they do not depend on ê.
7.2. Fundamental Forms and Curvatures 197

We need to solve the system:

∂ ∂
hij X i X j = λ gij X i X j



k k
k = 1, 2, · · · , n
∂X ∂X
gij X i X j = 1
∂X i
Using ∂X k
= δik , one can easily obtain
hij δik X j + X i δjk = λgij δik X j + δjk X i

(7.3) =⇒ hkj X j = λgkj X j

for any k = 1, 2, · · · , n. Multiplying g ik (which is the (i, k)-entry of g −1 ) on both sides,
we then get:
g ik hkj X j = λg ik gkj X j =⇒ g ik hkj X j = λX i
for any i = 1, 2, · · · , n. In other words, we have
[g]−1 [h](X 1 , · · · , X n )T = λ(X 1 , · · · , X n )T ,
and so λ is the eigenvalue of the matrix [g]−1 [h]. From (7.3), we have:
hkj X j X k = λ gkj X j X k = λ.
| {z } | {z }
kn (p,ê) =1

Therefore, if kn (p, ê) achieves its maximum and minimum among all ê ∈ Tp Σ, then
kn (p, ê) is an eigenvalue of [g]−1 [h], which is the matrix representation of S with respect
to local coordinates (u1 , · · · , un ) as desired 

From now on we denote

 
∂F ∂F
hji jk
:= g hki =⇒ S = hji .
∂ui ∂uj
Remark 7.23. It is also interesting to note that locally
 
∂ν ∂F ∂F
(7.4) = −g jk hki = −S .
∂ui ∂uj ∂uj
To show this, we let
∂ν ∂F
= Aji .
∂ui ∂uj
Then we consider
   
∂F ∂ν ∂F ∂F
hik = − , =− , Aj = −gij Ajk .
∂ui ∂uk ∂ui k ∂uj
Taking g li on both sides, we get Alk = −g li hik = −hlk as desired. In other words, the
shape operator can be regarded as the minus of the tangent map of ν.
Denote the eigenvalues of Sp , i.e. the principal curvatures, by λ1 (p), · · · , λn (p). The
maximum and minimum possible normal curvatures at p among all unit directions are
two of the principal curvatures. We further define:

Definition 7.24 (Mean Curvature and Gauss Curvature).

H(p) := λ1 (p) + · · · + λn (p) = tr[g]−1 [h] (mean curvature)
det[h]
K(p) := λ1 (p) · · · λn (p) = det[g]−1 [h] = (Gauss curvature)
det[g]

It turns out that when dim Σ = 2, the Gauss curvature K depends only on the first
fundamental form even though it is defined using the second fundamental form as well.
It is a famous theorem by Gauss commonly known as Theorema Egregium.
198 7. Geometry of Euclidean Hypersurfaces

Exercise 7.8 (Rigid-Body Motion). A map Φ : Rn+1 → Rn+1 is said to be a rigid-

body motion if there exist an (n + 1) × (n + 1) orthogonal matrix A and a constant
vector p ∈ Rn+1 such that Φ(x) = Ax + p for all x ∈ Rn+1 . Consider a regular
Euclidean hypersurface Σ in Rn+1 , and its image Σ e := Φ(Σ). Show that first and
second fundamental forms (and hence all curvatures we have discussed) of Σ and Σ e
are the same up to a sign, i.e.
Φ∗ ge = g and Φ∗ e h = h.
Here ge and e
h are first and second fundamental forms of Σ
e respectively.

7.2.3. Curvatures of Graphs. This subsection assumes Σ is a two dimensional

regular surface in R3 . In Examples 7.9 and 7.21, we computed the first and second
fundamental forms of the graph Σf of a function f . Using these, it is not difficult to
compute various curvatures of the graph. In this subsection, we are going to discuss the
geometric meaning of each curvature in this context, especially at the point where the
tangent plane is horizontal.

Proposition 7.25. Let Σf be the graph of a two-variable function f (u1 , u2 ) : U ⊂ R2 → R.

Suppose p is a point on Σf such that the tangent plane Tp Σf is horizontal, i.e. p is a
critical point of f , and the Gauss map ν is taken to be upward-pointing, then
• K(p) > 0 and H(p) > 0 =⇒ p is a local minimum of f
• K(p) > 0 and H(p) < 0 =⇒ p is a local maximum of f
• K(p) < 0 =⇒ p is a saddle of f

Proof. At a critical p of f , we have ∇f (p) = 0. From Examples 7.9 and 7.21, we have
computed:
∂f ∂f
gij (p) = δij + (p) (p) = δij
∂ui ∂uj
| {z }
=0
∂2f
∂ui ∂uj (p) ∂2f
hij (p) = q = (p)
2 ∂ui ∂uj
1 + |∇f (p)|
Note that the Gauss map ν was taken to be upward-pointing in Example 7.21, as required
in this proposition.
Therefore, we have:
 2 
det[h] ∂ f 2


K(p) = (p) = det (p) = f11 f22 − f12 (p)
det[g] ∂ui ∂uj
1 ij 1 ∂2f 1
H(p) = g (p)hij (p) = δ ij (p) = (f11 + f22 ) (p)
2 2 ∂ui ∂uj 2
From the second derivative test in multivariable calculus, given a critical point p, if
2
f11 f22 − f12 > 0 and f11 + f22 > 0 at p, then p is a local minimum of f . The other cases
can be proved similarly using the second derivative test. 

Given any regular surface Σ (not necessarily the graph of a function) and any point
p ∈ Σ, one can apply a rigid-motion motion Φ : R3 → R3 so that Tp Σ is transformed into
a horizontal plane. Then, the new surface Φ(Σ) becomes locally a graph of a function f
near the point Φ(p). Recall that the Gauss curvatures of p and Φ(p) are the same as given
by Exercise 7.8. If K(p) > 0 (and hence K(Φ(p)) > 0), then Proposition 7.25 asserts that
7.2. Fundamental Forms and Curvatures 199

Φ(p) is a local maximum or minimum of the function f and so the surface Φ(Σ) is locally
above or below the tangent plane at Φ(p). In other words, near p the surface Σ is locally
on one side of the the tangent plane Tp Σ. On the other hand, if K(p) < 0 then no matter
how close to p the surface Σ would intersect Tp Σ at points other than p.

7.2.4. Surfaces of Revolution. Surfaces of revolution are surfaces obtained by

revolving a plane curve about a central axis. They are important class of surfaces,
examples of which include spheres, torus, and many others. In this subsection, we will
study the fundamental forms and curvatures of these surfaces.
For simplicity, we assume that the z-axis is the central axis. A surface of revolution
(about the z-axis) is defined as follows.

Definition 7.26 (Surfaces of Revolution). Consider the curve γ(t) = (x(t), 0, z(t)),
where t ∈ (a, b), on the xz-plane such that x(t) ≥ 0 for any t ∈ (a, b). The surface of
revolution generated by γ is obtained by revolving γ about the z-axis, and it can be
parametrized by:
F (t, θ) = (x(t) cos θ, x(t) sin θ, z(t)), (t, θ) ∈ (a, b) × [0, 2π].

It is a straight-forward computation to verify that:

∂F
(7.5) = (x0 (t) cos θ, x0 (t) sin θ, z 0 (t))
∂t
∂F
(7.6) = (−x(t) sin θ, x(t) cos θ, 0)
∂θ
∂F ∂F
(7.7) × = (−x(t) z 0 (t) cos θ, −x(t) z 0 (t) sin θ, x(t) x0 (t))
∂t ∂θ

Exercise 7.9. Verify (7.5)-(7.7) and show that:

∂F ∂F
× = |x(t) γ 0 (t)| .
∂t ∂θ
Under what condition(s) will F be a smooth local parametrization when (t, θ) is
restricted to (a, b) × (0, 2π)?

Under the condition on γ(t) = (x(t), 0, z(t)) that its surface of revolution is smooth,
one can easily compute that the first fundamental form is given by:
 0 2 2

(x ) + (z 0 ) 0
(7.8) [g] = (matrix notation)
0 x2
h i
2 2
g = (x0 ) + (z 0 ) dt2 + x2 dθ2 (tensor notation)
∂F
× ∂F
and the second fundamental form with respect to the Gauss map ν := ∂t ∂θ
is given
| ∂F
∂t ∂θ |
× ∂F
by:
 0 00
x z − x00 z 0

1 0
(7.9) [h] = q (matrix notation)
2 2 0 x z0
(x0 ) + (z 0 )
1  0 00
(x z − x00 z 0 ) dt2 + x z 0 dθ2

h= q (tensor notation)
2 2
(x0 ) + (z 0 )
200 7. Geometry of Euclidean Hypersurfaces

Exercise 7.10. Verify that the first and second fundamental forms of a surface of
revolution with parametrization
F (t, θ) = (x(t) cos θ, x(t) sin θ, z(t)), (t, θ) ∈ (a, b) × [0, 2π]
are given as in (7.8) and (7.9).

As both [g] and [h] are diagonal matrices, it is evident that the principal curvatures,
i.e. the eigenvalues of [g]−1 [h], are:
x0 z 00 − x00 z 0 x0 z 00 − x00 z 0
k1 = h i3/2 = 3
2
(x0 ) + (z 0 )
2 |γ 0 |

z0 z0
k2 = q =
2 2 x |γ 0 |
x (x0 ) + (z 0 )
Note that here we are not using the convention that k1 ≤ k2 as in before, since there is
no clear way to tell which eigenvalue is larger.
Therefore, the Gauss and mean curvatures are given by:
(x0 z 00 − x00 z 0 ) z 0
(7.10) K = k1 k2 = 4
x |γ 0 |
!
1 1 x0 z 00 − x00 z 0 z0
(7.11) H = (k1 + k2 ) = 3 +
2 2 |γ 0 | x |γ 0 |

Example 7.27. Let’s verify that the round sphere (of any radius) has indeed constant
Gauss and mean curvatures. Parametrize the sphere by:
F (t, θ) = (R sin t cos θ, R sin t sin θ, R cos t),
i.e. taking x(t) = R sin t and z(t) = R cos t. By (7.10), then the Gauss curvature is clearly
given by:
 
(R cos t)(−R cos t) − (−R sin t)(−R sin t) (−R sin t) 1
K= = 2.
(R sin t) · R4 R
By (7.11), the mean curvature is given by:
1 −R2
 
−R sin t 1
H= + =− .
2 R3 (R sin t) · R R


Exercise 7.11. Compute the Gauss and mean curvatures of a round torus using
(7.10) and (7.11).
7.3. Theorema Egregium 201

7.3. Theorema Egregium

The goal of this section is to give the proof of a celebrated theorem due to Gauss, known
in Latin as Theorema Egregium (a surprising/remarkable theorem). The theorem asserts
that although the Gauss curvature of a two-dimensional regular surface in R3 was defined
using both the first and second fundamental forms, it indeed depends only on the first
fundamental form g.
It is remarkable in a sense that to define a notion of curvature, we no longer require
the surface to have an ambient Euclidean space. So long as one can declare an appropriate
2-tensor g to act in lieu as the “first fundamental form”, then one can still make sense
of curvatures. Einstein’s theory of general relativity relies very much on Riemannian
geometry because it regards the Universe as an intrinsic 4-dimensional manifold without
the ambient space (there is nothing outside the Universe). Riemannian geometry is a
good fit mathematical language to formulate the theory of general relativity in a rigorous
way.
To prove the Theorema Egregium, we first need to introduce covariant derivatives,
which depend only on the first fundamental form. Then, the theorem can be proved by
showing det[h] depends only on the covariant derivatives (and hence only on the first
fundamental form).

7.3.1. Covariant Derivatives. Let’s first recall directional derivatives in multivari-

able calculus. Let Σ be a regular hypersurface, and γ(t) : (a, b) → Σ be a smooth curve
on Σ. Given a vector field X on Σ, the directional derivative of X at p along γ is given by
d
Dγ 0 X(p) := X(γ(t))
dt t=t0

where t0 is a time such that γ(t0 ) = p.

When γ(t) is a ui -coordinate curve and X is any vector field, then
∂X
Dγ 0 X = .
∂ui
∂F
In particular, if X = ∂uj , then we have:

∂ ∂F ∂2F
Dγ 0 X = = .
∂ui ∂uj ∂ui ∂uj
∂F
In general, suppose X = X i ∂u i
and given any curve γ(t) on Σ locally expressed as
(u1 (t), · · · , un (t)), then by the chain rule we have:
d ∂X dui
Dγ 0 X = X(γ(t)) =
dt ∂u dt
  i
∂ ∂F dui
= Xj
∂ui ∂uj dt
∂X j ∂F ∂2F
 
dui
= + Xj
∂ui ∂uj ∂ui ∂uj dt
∂X j ∂F
Note that under a fixed local coordinate system (u1 , · · · , un ), the quantities ∂ui ∂uj +
j ∂2F dui
X are uniquely determined by the vector field X whereas
∂ui ∂uj are uniquely dt
determined by the tangent vector γ 0 of the curve. Now given another vector field
∂F
Y = Y i ∂ui
, then one can define
DY X(p) := Dγ 0 X(p)
202 7. Geometry of Euclidean Hypersurfaces

where γ is any curve on Σ which solves the ODE γ 0 (t) = Y (γ(t)) and γ(0) = p. The
Existence Theorem of ODEs guarantees there is such a curve γ that flows along Y . Locally,
DY X can be expressed as:
∂X j ∂F ∂2F
 
(7.12) DY X = + Xj Y i.
∂ui ∂uj ∂ui ∂uj

In short, covariant derivatives on hypersurfaces are projection of directional deriva-

tives onto the tangent space. They play an important role in differential geometry as
it can be shown to be depending only on the first fundamental form, in contrast to
directional derivatives which sit in the ambient space Rn+1 . Therefore, we are able to
generalize the notion of covariant derivatives to Riemannian manifolds.

Definition 7.28 (Covariant Derivatives). Let Σ be a regular surface with Gauss map ν,
and γ(t) be a smooth curve on Σ. Given two vector fields X and Y on Σ, we define the
covariant derivative of X at p ∈ M along Y to be
T

∇Y X(p) := (DY X(p)) = DY X − hDY X, νi ν (p).
T
Here (DY X(p)) represents the projection of DY X(p) onto the tangent space Tp Σ.

Using the fact that hDY X, νi = −hX, DY νi = h(X, Y ) = hij X i Y j and (7.12), we
can derive the local expression for ∇Y X:

 
X  ∂X j ∂F 2
∂ F
 X
(7.13) ∇Y X = + Xj Y i − hij X i Y j  ν
i,j
∂ui ∂uj ∂ui ∂uj i,j
| {z } | {z }
DY X hDY X,νiν

Suppose X, X, e Y and Ye are vector fields and ϕ is a smooth scalar functions. One
can verify that the following properties hold:
(a) ∇ϕY X = ϕ∇Y X
(b) ∇Y (ϕX) = (∇Y ϕ) X + ϕ∇Y X
(c) ∇Y +Ye X = ∇Y X + ∇Ye X
(d) ∇Y (X + X)
e = ∇Y X + ∇Y X
e
∂F ∂F
According to (7.12) and (7.13), given any vector fields X = X i ∂ui
and Y = Y i ∂u i
,
∂2F
the second derivatives determine both DY X and ∇Y X. We are going to express
∂ui ∂uj n o
∂2F ∂F ∂F
in terms of this tangent-normal basis ∂u
∂ui ∂uj 1
, · · · , ∂un
, ν of Rn+1 . From (7.13),
we have:
∂2F ∂F ∂F
= D ∂F = ∇ ∂F + hij ν.
∂ui ∂uj ∂u j ∂ui ∂u j ∂ui

∂F
We denote the coefficients of the tangent vector ∇ ∂F ∂ui by the following Christoffel
∂uj

symbols:
 
∂F ∂F
(7.14) ∇ ∂F = Γkij .
∂uj ∂ui ∂uk

The Christoffel symbols can be shown to be depending only on the first fundamental
form g. We will use it to prove that Gauss curvature depends also only on g but not on h.
7.3. Theorema Egregium 203

Proposition 7.29. In a local coordinate system (u1 , · · · , un ), the Christoffel symbols Γkij ’s
can be locally expressed in terms of the first fundamental form as:
 
k 1 kl ∂gjl ∂gil ∂gij
(7.15) Γij = g + − .
2 ∂ui ∂uj ∂ul

D E
∂F ∂F
Proof. First recall that gij = ∂u ,
i ∂uj
. By differentiating both sides respect to ul , we
get:
 2   2 
∂gij ∂ F ∂F ∂ F ∂F
= , + , .
∂ul ∂ul ∂ui ∂uj ∂ul ∂uj ∂ui
Using (7.14), we get:
* + * +
∂gij ∂F ∂F ∂F ∂F
(*) = Γkli + hli ν , + Γklj + hlj ν ,
∂ul ∂uk ∂uj ∂uk ∂ui
| {z } | {z }
∂2 F ∂2 F
∂ul ∂ui ∂ul ∂uj

= Γkil gkj + Γklj gki

By cyclic permutation of indices {i, j, l}, we also get:
∂gil
= Γkij gkl + Γkjl gki


(**)
∂uj
∂gjl
= Γkji gkl + Γkil gkj


(***)
∂ui
Recall that Γkij = Γkji and hij = hji for any i, j and k. By considering (**)+(***)-(*), we
get:
∂gil ∂gjl ∂gij
(7.16) + − = 2Γkij gkl
∂uj ∂ui ∂ul
Finally, by multiplying g lq on both sides of (7.16) and summing up over all l, we get:
 
∂gil ∂gjl ∂gij
g lq + − = 2g lq Γkij gkl
∂uj ∂ui ∂ul
= 2Γkij δkq
= 2Γqij .
Relabelling the index q by k, we complete the proof of (7.15). 

7.3.2. The Proof of Theorema Egregium. The key ingredient of Gauss’s Theorema
Egregium is the following Gauss-Codazzi’s equations, which hold for any regular hyper-
surface in any dimension. When the dimension is two, the RHS of the Gauss’s equation
is similar to det[h] while the LHS depends only g. This gives Theorema Egregium as a
direct consequence of the Gauss’s equation.

Theorem 7.30 (Gauss-Codazzi’s Equations). On any regular hypersurface Σ in Rn+1 ,

the following equations hold:
∂Γqjk ∂Γqik
− + Γljk Γqil − Γlik Γqjl = g ql (hjk hli − hik hlj ) (Gauss)
∂ui ∂uj
∂hjk ∂hik
− + Γljk hil − Γlik hjl = 0 (Codazzi)
∂ui ∂uj
204 7. Geometry of Euclidean Hypersurfaces

Proof. The key step of the proof is to start with the fact that:
∂3F ∂3F
=
∂ui ∂uj ∂uk ∂uj ∂ui ∂uk
for any i, j, k, and then rewrite both sides in terms of the tangent-normal basis of Rn+1 .
Gauss’s equation follows from equating the tangent coefficients, and Codazzi’s equation
is obtained by equating the normal coefficient.
By (7.14), we have:
∂2F ∂F
= Γljk + hjk ν.
∂uj ∂uk ∂ul
Differentiating both sides with respect to ui , we get:
∂3F
 
∂ ∂F
= Γljk + hjk ν
∂ui ∂uj ∂uk ∂ui ∂ul
∂Γljk ∂F ∂2F ∂hjk ∂ν
= + Γljk + ν + hjk
∂ui ∂ul ∂ui ∂ul ∂ui ∂ui
∂Γljk ∂F
 
∂F ∂h jk ∂F
= + Γljk Γqil + hil ν + ν − hjk hqi
∂ui ∂ul ∂uq ∂ui ∂uq
q
∂Γjk ∂F
 
∂F ∂hjk ∂F
= +Γljk Γqil + + Γljk hil ν − hjk hql gil
∂ui ∂uq ∂uq ∂ui ∂uq
| {z }
l7→q

∂Γqjk
!  
∂F ∂hjk
= + Γljk Γqil ql
− g hjk hil + + Γljk hil ν
∂ui ∂uq ∂ui

By switching i and j, we get:

 q
∂3F
  
∂Γik ∂F ∂hik
= + Γlik Γqjl − g ql hik hlj + + Γlik hjl ν
∂uj ∂ui ∂uk ∂uj ∂uq ∂uj

The Gauss-Codazzi’s equations can be obtained by equating the coefficients of each

tangent and normal component. 

We derived the Gauss-Codazzi’s equations (Theorem 7.30). It is worthwhile the

note that the LHS of the Gauss’s equation involves only Christoffel’s symbols and their
derivatives:
∂Γqjk ∂Γqik
− + Γljk Γqil − Γlik Γqjl = g ql (hjk hli − hik hlj )
∂ui ∂uj
| {z }
depends only on Γk
ij ’s

From (7.15) we also know that the Christoffel symbols depend only on the first funda-
mental form g but not on h. For simplicity, we denote:

q ∂Γqjk ∂Γqik
Rijk := − + Γljk Γqil − Γlik Γqjl .
∂ui ∂uj
q
The lower and upper indices for Rijk are chosen so as to preserve their positions in the
q
RHS expression (q being upper, and i, j, k being lower). We will see later that Rijk ’s are
the components of the Riemann curvature tensor.
We are now in a position to give a proof of Gauss’s Theorema Egregium as a direct
consequence of the Gauss’s equation.
7.3. Theorema Egregium 205

Theorem 7.31 (Theorema Egregium, Gauss). On any regular surface Σ2 in R3 , the

Gauss curvature K depends only on its first fundamental form g. In other words, K is
intrinsic for regular surfaces.

Proof. Consider the Gauss’s equation, which asserts that for any i, j, k and q:
q
Rijk = g ql (hjk hli − hik hlj ) .
Multiplying both sides by gpq , and summing up all q’s, we get:
q
gpq Rijk = gpq g ql (hjk hli − hik hlj ) = hjk hpi − hik hpj .
The above result is true for any i, j, k and p. In particular, when (i, j, k, p) = (1, 2, 2, 1),
we get:
q
g1q R122 = h22 h11 − h212 = det[h].
q
This shows det[h] depends only on g since R122 does so.
Finally, recall that the Gauss curvature is given by:
det[h]
K= .
det[g]
Therefore, we have completed the proof that K depends only g. 
Remark 7.32. It is important to note that Theorem 7.31 holds for surfaces (i.e. dim Σ =
2).
q
The long Rijk -term can be interpreted in a nicer way using covariant derivatives. For
simplicity, we denote
∂F
∂i :=
∂ui
∇i := ∇∂i = ∇ ∂F
∂ui

where (u1 , · · · , un ) is a local coordinate system of regular hypersurface Σ in Rn+1 . By

direct computations, we can verify that:
∇j ∂k = Γljk ∂l
∇i (∇j ∂k ) = ∇i Γljk ∂l

∂Γljk
= ∂l + Γljk ∇i ∂l
∂ui
∂Γqjk
= ∂q + Γljk Γqil ∂q
∂ui
Similarly, we also have:
∂Γqik
∇j (∇i ∂k ) = ∂q + Γlik Γqjl ∂q .
∂uj
q
Hence, the term Rijk is the commutator of ∇i and ∇j :
[∇i , ∇j ]∂k = ∇i (∇j ∂k ) − ∇j (∇i ∂k )
∂Γqjk
!
∂Γqik
= − + Γljk Γqil − Γlik Γqjl ∂q
∂ui ∂uj
q
= Rijk ∂q .
In summary, we see that once gij is fixed, the covariant derivative ∇ and hence the
q
curvature term Rijk are uniquely determined. It motivates the idea that if we can
206 7. Geometry of Euclidean Hypersurfaces

declare the gij ’s on an abstract manifold M , then we can define its curvatures using the
prescribed gij even if M is not a submanifold of an Euclidean space. This motivates the
development of intrinsic geometry, a branch of geometry that plays no reference to the
ambient space but only the manifold itself. This branch is called Riemannian Geometry,
which is what this course is about!
Part 3

Riemannian Geometry
 Chapter 8

Riemannian Manifolds

“When there is matter, there is

geometry.”

Johannes Kepler

8.1. Riemannian Metrics

On a regular hypersurface in a Euclidean space, the first fundamental form g encodes
many of its geometric properties such as length, area, and in two-dimensional case, the
Gauss curvature. Now we want to extend all these geometric notions to abstract manifold
which needs not have an ambient Euclidean space.
Without the ambient space, the “first fundamental form” g (which will be renamed
as the Riemannian metric) is now not being induced from the dot product of the ambient
space, but instead is being defined. It is analogous to the development of topological
spaces from metric spaces. Open sets in a topological space are no longer characterized
using metric balls, but are instead being declared as a collection of subsets which is called
the topology of the space.

Definition 8.1 (Riemannian Metrics). On a smooth manifold M , a Riemannian metric

g is a C ∞ (2, 0)-tensor on M such that
symmetry: g(X, Y ) = g(Y, X) for any X, Y ∈ Tp M ;
positivity: g(X, X) ≥ 0 for any X ∈ Tp M , with equality holds if and only if X = 0.
The pair (M, g) is called a Riemannian manifold.

We denote the local components of g by:

 
∂ ∂
gij = g , ,
∂ui ∂uj
so that g = gij dui ⊗ duj .
Example 8.2. The Euclidean space Rn is a Riemannian manifold with a Riemannian
metric
Xn
δ := dxi ⊗ dxi .
i=1

209
210 8. Riemannian Manifolds

It is called the flat metric on Rn . 

Example 8.3 (Hyperbolic Spaces - Poincaré Disc Model). The n-dimensional hyperbolic
space Hn (under the Poincaré model) is topologically an open unit ball in Rn :
Hn := {x ∈ Rn : |x| < 1}
equipped with the Poincaré metric
n
X 4 dxi ⊗ dxi
g=  2 .
2
i=1 1 − |x|


Example 8.4. Any regular hypersurface surface Σn in Rn+1 is a Riemannian manifold
with Riemannian metric g given by the first fundamental form. The symmetry and strict
positivity conditions are inherited from the flat metric on Rn+1 .
For instance, the round sphere S2 with radius R has a Riemannian metric given by:
g = R2 dϕ2 + R2 sin2 ϕ dθ2 .

Example 8.5. Given any smooth immersion Φ : Σ → M between two C ∞ smooths, and
suppose M is a Riemannian manifold with metric g. Then, Σ has an induced Riemannian
metric given by ge := Φ∗ g. The symmetry condition holds trivially. To verify the strict
positivity condition, we consider X ∈ Tp Σ, then
ge(X, X) = g(Φ∗ X, Φ∗ X) ≥ 0,
and by strict positivity of g, we have equality holds if and only if Φ∗ X = 0. Since Φ∗ is
injective (as Φ is an immersion), we must have X = 0.
In particular, any submanifold of a Riemannian manifold is itself a Riemannian
manifold with induced metric defined using the pull-back of the inclusion map. 
Example 8.6 (Product Manifolds). Suppose (M, gM ) and (N, gN ) are two Riemannian
manifolds, then the product M × N is also a Riemannian manifold with metric given by:
∗ ∗
gM ⊕ gN := πM gM + πN gN
where πM : M × N → M and πN : M × N → N are projection maps. A tangent vector
X ∈ Tp (M × N ) can be expressed as (XM , XN ) ∈ Tp M ⊕ Tp N . The product metric acts
on tangent vectors by:


(gM ⊕ gN ) (XM , XN ), (YM , YN ) = gM (XM , YM ) + gN (XN , YN ).

Example 8.7 (Conformal Metrics). Given any Riemannian manifold (M, g) and a smooth
function f : M → R, one can define another Riemannian metric ge on M by conformal
rescaling:
ge = ef g.

Example 8.8 (Fubini-Study Metric on CPn ). On CPn , there is an important Riemannian
metric called the Fubini-Study metric, which is more convenient to be expressed using
complex coordinates. Parametrize CPn using standard local coordinates
  h i
(j) (j) (j) (j) (j) (j)
F (j) z0 , · · · , zj−1 , zj+1 , · · · , zn(j) = z0 : · · · : zj−1 : 1 : zj+1 : · · · : zn(j) .
8.1. Riemannian Metrics 211

(j)
For simplicity, we let zj = 1. The Fubini-Study metric is defined to be:
Pn+1 (j) 2
 
2
∂ log k=1 z k
gFS := 2 Re  dzp(j) ⊗ dz̄q(j)  .
 
(j) (j)
∂zp ∂ z̄q

We leave it as an exercise for readers to show that gFS is independent of j, and is a

Riemannian metric. 

Given a Riemannian metric, one can then define the length of curve and volume of a
region of a manifold. Suppose γ(t) : [a, b] → M is a C 1 curve on a Riemannian manifold
(M, g), then we define the length of γ by (with respect to g) by:
Z bq


Lg (γ) := g γ 0 (t), γ 0 (t) dt
a
0 ∂
where γ (t) := γ∗ ∂t . Note that Lg (γ) depends on the metric g.
Given an open subset O of Riemannian manifold (M, g) such that O can be covered
by one single parametrization chart (u1 , · · · , un ), the volume of O (with respect to g) is
defined to be: Z q
Volg (O) = det[gij ] du1 ∧ · · · ∧ dun
O
∂
where gij = g( ∂u , ∂ ). The total volume of an orientable Riemannian manifold (M, g)
i ∂uj
is defined by:
XZ q
Volg (M ) := α ] du1 ∧ · · · ∧ dun
ρα det[gij α α
α Oα

where A := {Fα (uiα ) : Uα → Oα } is an oriented atlas of M and ρα is a partition of unity

subordinate to A.

8.1.1. Isometries. Recall that two smooth manifolds M and N are considered to
be the same in topological sense if there exists a diffeomorphism between them. Assume
further that they are Riemannian manifolds with metrics gM and gN . Heuristically, we
consider M and N to be geometrically the same if their Riemannian metrics are the “same”.
Precisely, take a diffeomorphism Φ : M → N , which sets up a one-one correspondence
between p ∈ M and Φ(p) ∈ N , and a tangent vector X ∈ Tp M corresponds to Φ∗ X ∈
TΦ(p) N . We consider gM and gN are “the same” if for any X, Y ∈ Tp M , we have
gM (X, Y ) = gN (Φ∗ X, Φ∗ Y ) meaning that inner product between X and Y is the same
as that between their correspondences Φ∗ X and Φ∗ Y . In this case, we then call Φ is an
isometry, and (M, gM ) and (N, gN ) are said to be isometric. Using the pull-back map, we
can rephrase these definitions in the following way:

Definition 8.9 (Isometries). Two Riemannian manifolds (M, gM ) and (N, gN ) are said
to be isometric if there exists a diffeomorphism Φ : M → N such that Φ∗ gN = gM . Such
a diffeomorphism Φ is called an isometry between (M, gM ) and (N, gN ).

Example 8.10. Consider the unit disk model D = {(x, y) ∈ R2 : x2 + y 2 < 1} of the
hyperbolic 2-space with the Poincaré’s metric introduced in Example 8.3:
4(dx2 + dy 2 )
g= .
1 − x2 − y 2
It is well-known that the hyperbolic plane can be equivalently modeled by the upper-half
plane U = {(u, v) : v > 0} with metric
du2 + dv 2
ge = .
v2
212 8. Riemannian Manifolds

These two models can be shown to be isometric. The isometry Φ : D → U can be

described using complex coordinates in an elegant way:
1−z
Φ(z) = i
1+z
where z = x + yi. 

Exercise 8.1. Complete the calculations in Example 8.10 to show that Φ is indeed
an isometry.

Without the ambient space, we now make sense of symmetries in an intrinsic way.
A C ∞ vector field V on (M, g) generates a 1-parameter subgroup of diffeomorphisms
Φt : M → M such that
∂Φt
(p) = V (Φt (p)), Φ0 = idM .
∂t
We then say g is symmetric along V if Φ∗t gΦt (p) = gp for any t ∈ R and p ∈ M .
Generally, such a diffeomorphism family Φt is hard to compute. Fortunately, one can
use the Lie derivative to check if g is symmetric along a given vector field without actually
computing Φt . The condition Φ∗t gΦt (p) = gp for any t ∈ R and p ∈ M is equivalent to
LV g = 0.

8.1.2. Covering Spaces and Quotient Manifolds. A surjective smooth map π :

M → M between two smooth manifolds is called a covering map if for any p ∈ M , there
f
exists an open neighborhood U containing p such that π −1 (U ) is a disjoint union of open
sets {Vα } in M
f each of which is diffeomorphic to U via π . The manifold M
Vα
f is then
said to be a covering space of M .
The quotient map π : Rn → Rn /Zn taking x to its equivalent class [x] is an example
of a covering map. Similar for the quotient map π : Sn → RPn (where RPn is regarded
as Sn with all antipodal points identified).
Some smooth manifolds are constructed using quotients. The torus Rn /Zn and RPn
are good examples. Subject to certain conditions, one can define a Riemannian metric of
the quotient manifold using that of the covering manifold, and vice versa. The crucial
condition is that M
f has to be symmetric enough. Precisely, we have the following results:

Proposition 8.11. Suppose π : M f → M is a covering map. Consider the following Deck

transformation group defined by
Deck(π) := {Φ ∈ Diff(M
f) : π ◦ Φ = π}.

Then, any Riemannian metric g of M induces a Riemannian metric π ∗ g on M

f which is
invariant under any Φ ∈ Deck(π), i.e. Φ∗ (π ∗ g) = π ∗ g.
Conversely, suppose the Deck transformation group acts transitively on π −1 (p) for
any p ∈ M (which is true when M f is simply-connected). Then, given any Riemannian
metric ge of M such that any Deck transformation Φ : M
f f→M f is an isometry of (M
f, ge),
∗ ∗
i.e. Φ ge = ge, then there exists a Riemannian metric g on M such that π g = ge.

Proof. Suppose g is a Riemannian metric of M . Since π is a covering map, in particular

it is an immersion (and submersion too although it is not needed). This implies π ∗ g is a
Riemannian metric of M f→M
f (see Example 8.5). For any Deck transformation Φ : M f,
∗ ∗ ∗ ∗ ∗ ∗
we have π ◦ Φ = π and so Φ ◦ π = π , this shows Φ (π g) = π g, completing the proof
of the first part of the proposition.
8.1. Riemannian Metrics 213

Conversely, given ge a Riemannian metric on M f such that Φ∗ ge = ge for any Φ ∈

Deck(π), we need to construct a Riemannian metric g on M such that Φ∗ g = ge. Given any
p ∈ M , by the covering map condition, one can pick q ∈ π −1 (p) and a neighborhood V
f → Tp M
of q such that π V is a diffeomorphism onto its image. In particular, (π∗ )q : Tq M
is invertible. We define gp : Tp M × Tp M → R as follows:
gp (X, Y ) := geq (π∗ )−1 −1


q (X), (π∗ )q (Y ) for any X, Y ∈ Tp M.
It is smooth as π V : V → π(V ) is a diffeomorphism (so is its inverse). We need to justify
that such a definition of gp is independent of the choice of q in π −1 (p). It thanks to
the invariant condition of ge under the Deck transformation group. Given another point
q 0 ∈ π −1 (p), by standard topology theory one can find a Deck transformation Φ such that
Φ(q) = q 0 . Then, we have
geq0 (π∗ )−1 −1 −1 ∗
) geq (π∗ )−1 −1




q 0 (X), (π∗ )q 0 (Y ) = (Φ q 0 (X), (π∗ )q 0 (Y )
−1 −1
= geΦ(q0 ) Φ−1 −1


∗ (π∗ )q 0 X, Φ∗ (π∗ )q 0 Y

= geq (π∗ )−1 −1

q X, (π∗ )q Y .

The last steps follows from π ◦ Φ = π. Finally, we check that π ∗ g = ge:

(π ∗ g)(X, Y ) = g(π∗ X, π∗ Y )
= ge π∗−1 (π∗ X), π∗−1 (π∗ Y )

= ge(X, Y ).
∗
This shows π g = ge. 
Example 8.12 (Flat Torus). A n-dimensional torus Tn can be regarded as the quotient
manifold Rn /Zn whose elements are equivalent classes [(x1 , · · · , xn )] under the relation
(x1 , · · · , xn ) ∼ (y1 , · · · , yn ) ⇐⇒ xi − yi ∈ Z for any i.
The quotient map π : Rn → Tn is then a covering map. Any Deck transformation
Φ : Rn → Rn is a diffeomorphism of Rn such that for any (x1 , · · · , xn ) ∈ Rn we have
[Φ(x1 , · · · , x1 )] = [(x1 , · · · , xn )]. In other words, there exists integers m1 , · · · , mn such
that
Φ(x1 , · · · , xn ) = (x1 , · · · , xn ) + (m1 , · · · , mn ).
These integers does not depend on xi ’s by continuity. Hence, the Deck transformation
group of the covering π : Rn → Tn is the set of translations by integer points.
It is clear that Φ∗ δ = δ where δ is the Euclidean metric of Rn . Hence by Proposition
8.11 (the converse part), there exists a metric g on Tn such that π ∗ g = δ. 
Example 8.13 (Real Projective Space). Consider RPn constructed by identifying an-
tipodal points on Sn . The quotient map π : Sn → RPn is a covering map, and Deck
transformations Φ : Sn → Sn of this covering are either Φ = id or Φ = −id. Both are
isometries of the round metric ground of Sn , so it induces a Riemannian metric g such that
π ∗ g = ground . 
214 8. Riemannian Manifolds

8.2. Levi-Civita Connections

Our next goal is to extend the notion of covariant derivatives of Euclidean hypersurfaces
to its analogue on Riemannian manifolds. On Euclidean hypersurfaces, the covariant
derivatives are inherited from the directional derivatives on the Euclidean ambient
space (which no longer exists for Riemannian manifolds. However, since the first
fundamental form g determines the covariant derivative (see Proposition 7.29) for
Euclidean hypersurfaces, it motivates us to define “covariant derivatives” (renamed as
Levi-Civita connections or Riemannian connections) using Riemannian metrics for an
abstract manifold.
 under local coordinates {ui }
Let g be a Riemannian metric on a manifoldM . Suppose
∂ ∂
on M the metric has local components gij = g ∂ui , ∂uj , then we define the Christoffel
symbol as:  
k 1 kl ∂gjl ∂gil ∂gij
Γij := g + −
2 ∂ui ∂uj ∂ul
where g kl is the (k, l)-th entry of the matrix [g]−1 . From now on we will denote
Γ∞ (T M ) = the space of C ∞ vector fields on M .
We then define an operator
∇ : Γ∞ (T M ) × Γ∞ (T M ) → Γ∞ (T M )
by declaring that it acts on coordinate vectors by
∂ ∂
∇ ∂ := Γkij
∂ui ∂uj ∂uk
and then is extended to general vector fields by the rules
(8.1) ∇X (Y1 + Y2 ) = ∇X Y1 + ∇X Y2
(8.2) ∇X1 +X2 Y = ∇X1 Y + ∇X2 Y
(8.3) ∇ϕX (f Y ) = [X(f ) · Y + f ∇X Y ] ϕ

∂ ∂
Exercise 8.2. Let X = X i ∂u i
and Y = Y j ∂u j
. Show that ∇X Y has the following
local expression:
∂Y k
 
∂
(8.4) ∇X Y = X i + X i Y j Γkij .
∂ui ∂uk
Check also that (8.4) is independent of local coordinates, i.e. given another local
coordinates {vα } so that X = X e α ∂ , Y = Ye β ∂ , and geαβ = g( ∂ , ∂ ), verify
∂vα ∂vα ∂vα ∂vβ
that: !
eγ  k

e α ∂Y + X
X eγ
e α Ye β Γ ∂
= X i ∂Y
+ X i j k
Y Γ ij
∂
αβ
∂vα ∂vγ ∂ui ∂uk
e γ in the new coordinate system are defined as
where the Christoffel symbols Γ αβ
 
e γ = 1 geγη ∂e
Γ
gβη
+
∂e
gαη
−
∂e
gαβ
.
αβ
2 ∂e
vα ∂e
vβ ∂evη

The operator ∇ described above is called the Levi-Civita connection (or Riemannian
connection) of (M, g). Note that the Christoffel symbols Γkij , and hence the Levi-Civita
connection, depend on g.
The term connection has a broader meaning. Any operator D : Γ∞ (T M )×Γ∞ (T M ) →
∞
Γ (T M ) satisfying all of (8.1)-(8.3) is called a connection on M .
8.2. Levi-Civita Connections 215

Exercise 8.3. Given a connection D : Γ∞ (T M ) × Γ∞ (T M ) → Γ∞ (T M ) (not

necessarily the Levi-Civita connection), show that if X1 (p) = X2 (p) at a point p ∈ M ,
then we have DX1 Y (p) = DX2 Y (p).

Exercise 8.4. Consider a connection D : Γ∞ (T M ) × Γ∞ (T M ) → Γ∞ (T M ) on M

k
with local coordinates (ui ). Let γij be the coefficients when D acts on coordinate
vectors, i.e.
∂ k ∂
D ∂ = γij .
∂ui ∂u ∂uk
j
∂
Show that DX Y is uniquely determined by X i ’s Y i ’s, and γij
k
’s where X = X i ∂u i
∂
and Y = Y i ∂u i
.

It is straight-forward to verify that Γkij and gij satisfy the following relations:

(8.5) Γkij = Γkji

∂gij
(8.6) = Γlik gjl + Γljk gil
∂uk
for any i, j, k. Using invariant notations (i.e. without using local coordinates), these
relations can be written in equivalent form as:
(8.7) ∇X Y − ∇Y X = [X, Y ]
(8.8) Z (g(X, Y )) = g(∇Z X, Y ) + g(X, ∇Z Y )
for any vector fields X, Y, Z. Here [X, Y ] denotes the Lie brackets between X and Y .
∂ ∂
Let us verify that (8.6) and (8.8) are equivalent. Write X = X i ∂u i
, Y = Y i ∂u i
, and
∂
Z = Z i ∂ui , then

∂X i ∂Y j
 
∂ ∂gij i j
Z (g(X, Y )) = Z k gij X i Y j = Z k X Y + gij Y j + gij X i


.
∂uk ∂uk ∂uk ∂uk
∂
On the other hand, we have (for simplicity, denote ∂k := ∂uk and ∇k := ∇∂k ):

g(∇Z X, Y ) + g(X, ∇Z Y )
= g Z k ∇k (X i ∂i ), Y j ∂j + g X i ∂i , Z k ∇k (Y j ∂j )

i
 
k ∂X ∂ k i l ∂ j ∂
=g Z + Z X Γki ,Y
∂uk ∂ui ∂ul ∂uj
j
 
∂ ∂Y ∂ ∂
+ g Xi , Zk + Z k Y j Γlkj
∂ui ∂uk ∂uj ∂ul
∂X i j ∂Y j i
= Z k glj Γlki X i Y j + gil Γlkj X i Y j +gij Z k Y + gij Z k


X .
| {z } ∂uk ∂uk
∗

From the above calculations, it is clear that if (8.6) holds, then

∂gij i j
∗ = Zk X Y ,
∂uk
∂
and so we have proved (8.8). To prove (8.8) implies (8.6), one may simply take Z = ∂uk ,
∂ ∂
X = ∂u i
, and Y = ∂u j
and substitute them into (8.8).

Exercise 8.5. Show that (8.5) and (8.7) are equivalent.

216 8. Riemannian Manifolds

While we define Levi-Civita connections using local coordinates, it is possible to give

a more global definition, since we have the following uniqueness result:

Proposition 8.14. Any connection D : Γ∞ (T M ) × Γ∞ (T M ) → Γ∞ (T M ) on a Riemann-

ian manifold (M, g) satisfying both:
• DX Y − DY X = [X, Y ] for any X, Y ∈ Γ∞ (T M )
• Z (g(X, Y )) = g(DZ X, Y ) + g(X, DZ Y ) for any X, Y, Z ∈ Γ∞ (T M )
must be equal to the Levi-Civita connection.

k
Proof. Let (ui ) be local coordinates of M and define γij by
∂ k ∂
D ∂ = γij .
∂ui ∂uj ∂uk
Then, the given two conditions of this proposition are equivalent to
k k
γij = γji
∂gij l l
= γik gjl + γjk gil
∂uk
for any i, j, k. The proof is identical to the one that shows (8.5)-(8.6) and (8.7)-(8.8) are
equivalent (see page 215). By cyclic permutation of indices, we can then show:
∂gjk ∂gik ∂gij l
+ − = 2γij gkl
∂ui ∂uj ∂uk
for any i, j, k. Multiplying g −1 on both sides, we get:
 
l 1 kl ∂gjk ∂gik ∂gij
γij = g + − ,
2 ∂ui ∂uj ∂uk
which is exactly the Christoffel symbols Γkij for the Levi-Civita connection of g. Since the
action of D is uniquely determined by its action on coordinate vectors, we must have
D = ∇.


Therefore, one can also define the Levi-Civita connection with respect to (M, g) as
the unique operator that satisfies all of (8.1)-(8.3) and (8.7)-(8.8).

Exercise 8.6. Let (M, g) be a Riemannian manifold, and Σ be a submanifold of M .

Denote ι : Σ → M to be the inclusion map. Then, ḡ := ι∗ g is a Riemannian metric
¯ is given by:
on Σ. Show that the Levi-Civita connection of (Σ, ḡ), denoted by ∇,
¯ X Y = (∇X Y )T := projection of ∇X Y onto T Σ
∇
for any X, Y ∈ Γ∞ (T Σ).

8.2.1. Tensorial quantities. A linear operator T acting on vector fields and 1-forms
(X1 , · · · , Xr , ω1 , · · · , ωs ) ∈ Γ∞ (T M )r × Γ∞ (T ∗ M )s is said to be tensorial if one can
factor out a scalar function in each slot, i.e.
T (ϕ1 X1 , · · · , ϕr Xr , ψ1 ω1 , · · · , ψs ωs ) = ϕ1 · · · ϕr ψ1 · · · ψs T (X1 , · · · , Xr , ω1 , · · · , ωs )
for any scalar functions ϕi ’s and ψi ’s.
A Riemannian metric g is tensorial as it is required to be. For Euclidean hypersurfaces,
the second fundamental form h is tensorial since Df X ν = f DX ν. However, the operator
8.2. Levi-Civita Connections 217

S(X, Y ) := DX Y − DY X, where D is any connection, is not tensorial. One can compute

that
S(f X, Y ) = Df X Y − DY (f X) = f DX Y − (Y f )X − f DY X = f S(X, Y ) − (Y f )X.
However, one can modify it a little to make it to make it tensorial. Define
T (X, Y ) = DX Y − DY X − [X, Y ].
Then, we have
T (f X, Y ) = f DX Y − f DY X − (Y f )X −[f X, Y ].
| {z }
Df X Y −DY (f X)

For any scalar function ϕ, we have

[f X, Y ]ϕ = f X(Y (ϕ)) − Y (f X(ϕ))
= f XY (ϕ) − Y (f )X(ϕ) − f Y X(ϕ).
Therefore, [f X, Y ] = f [X, Y ] − (Y f )X. By cancellations, we get
T (f X, Y ) = f (DX Y − DY X) − f [X, Y ] = f T (X, Y ).
One can also check T (X, f Y ) = f T (X, Y ) in a similar way.
It is important to note that the a connection, written as D(X, Y ) := DX Y , is NOT
tensorial! While it is still true that D(f X, Y ) = f D(X, Y ), it fails to be tensorial in the
second slot:
D(X, f Y ) = DX (f Y ) = (Xf )Y + f DX Y 6= f D(X, Y ).
However, it is interesting (also an important fact) that the difference between two
connections is nonetheless tensorial! Suppose D and D e are two connections on a
manifold. Then, one can easily check that
(D − D)(X,
e f Y ) = DX (f Y ) − D
e X (f Y )

= (Xf )Y + f DX Y − (Xf )Y − f D
eX Y

= f (D − D)(X,
e Y ).

If an operator T is tensorial, then at any fixed point p ∈ M the action of T on

tangents and cotangents at p is uniquely determined by its action on a basis for Tp M and
Tp∗ M . Take a 2-tensor for example, if {ei } is a basis of Tp M for a fixed p ∈ M , then any
vectors Xp , Yp ∈ Tp M can be expressed as Xp = X i ei and Yp = Y i ei , and so
T (Xp , Yp ) = X i Y j T (ei , ej )
which depends only on the point p. However, if we consider a connection D instead
(which is not tensorial), we will see that
D(Xp , Yp ) = DX i ei (Y j ej ) = X i (Dei Y j ) + X i Y j Dei ej .
The quantity Dei Y j is a derivative of Y j which depends on a neighborhood of p (not just
p itself).

8.2.2. Levi-Civita connection on tensor bundles. From now on we will denote

∇ as the Levi-Civita connection of a given Riemannian metric g, and may sometimes
write ∇g to specify the metric. We will use D for an unspecified connection of a manifold.
The Levi-Civita connection ∇ (in fact any connection D) extends to an operator on
tensor bundles T r,s (M ) := T ∗ M ⊗r ⊗ T M ⊗s . First, the operator ∇X is extended to act
on 1-forms by the following rule.
218 8. Riemannian Manifolds

Given any vector field X and 1-form α, the output ∇X α is a 1-form such that given
any vector field Y the product rule holds formally:
X (α(Y )) = (∇X α)(Y ) + α(∇X Y )
or equivalently,
(∇X α)(Y ) := X(α(Y )) − α(∇X Y ).
Locally, we have:
      
∂ ∂ ∂ ∂
(∇i duj ) = duj − duj ∇i
∂uk ∂ui ∂uk ∂uk
 
∂ j ∂
= δ − duj Γlik
∂ui k ∂ul
= 0 − Γlik δlj = −Γjik .
In other words, we have
∇i duj = −Γjik duk .
Then, given any vector fields X, Y1 , · · · , Ys and 1-forms α1 , · · · , αr , we define the
operator ∇X by:
∇X (Y1 ⊗ Y2 ) := (∇X Y1 ) ⊗ Y2 + Y1 ⊗ (∇X Y2 )
and more generally,
r

X
∇X (⊗ri=1 ωi ) ⊗ (⊗sj=1 Yj ) =


ω1 ⊗ · · · ⊗ ∇X ωi ⊗ · · · ⊗ ωr ⊗ Y1 ⊗ · · · ⊗ Ys
i=1
s
X

+ ω1 ⊗ · · · ⊗ ωr ⊗ Y1 ⊗ · · · ⊗ ∇X Yj ⊗ · · · ⊗ Ys .
j=1

For instance, for a 2-tensor T = Tij dui ⊗ duj , we have

∇k T = ∇k Tij dui ⊗ duj

∂Tij i
= du ⊗ duj + Tij (∇k dui ) ⊗ duj + Tij dui ⊗ (∇k duj )
∂uk
∂Tij i
= du ⊗ duj − Tij Γikl dul ⊗ duj − Tij Γjkl dui ⊗ dul
∂uk
 
∂Tij
= − Tlj Γlki − Til Γlkj dui ⊗ duj .
∂uk
We denote ∇k Tij to be the local expression such that
∇k Tij dui ⊗ duj =: (∇k Tij ) dui ⊗ duj ,

so from the above calculation, we have:

∂Tij
∇k Tij = − Tlj Γlki − Til Γlkj .
∂uk
∂f
While ∇k f := ∂u k
for a scalar function f , note that ∇k Tij does not mean the partial
derivative of the scalar Tij with respect to uk .
∂
Likewise, given a vector field Y = Y i ∂u i
, we denote ∇j Y i to be the local expression
such that  
i ∂ ∂
∇j Y =: (∇j Y i ) .
∂ui ∂ui
8.2. Levi-Civita Connections 219

By direct computatios, we have:

∂Y i ∂
 
i ∂ ∂
∇j Y = + Y i ∇j
∂ui ∂uj ∂ui ∂ui
∂Y i ∂ ∂
= + Y i Γkji
∂uj ∂ui ∂uk
∂Y i
 
k i ∂
= + Y Γjk .
∂uj ∂ui

Therefore, we conclude that

∂Y i
∇j Y i = + Y k Γijk .
∂uj

Given a (r, s)-tensor T locally expressed as

···js ∂ ∂
T = Tij11···i dui1 ⊗ · · · ⊗ duir ⊗ ⊗ ··· ⊗ ,
r
∂uj1 ∂ujs

···js
we define ∇k Tij11···i r
to be the local expression such that
 
···js ∂ ∂
∇k Tij11···i dui1 ⊗ · · · ⊗ duir ⊗ ⊗ ··· ⊗
r
∂uj1 ∂ujs

···js
 ∂ ∂
= ∇k Tij11···i dui1 ⊗ · · · ⊗ duir ⊗ ⊗ ··· ⊗ .
r
∂uj1 ∂ujs

Exercise 8.7. Show that:

···js
j1 ···js ∂Tij11···i Xr
j1 ···jp−1 ljp+1 ···js jp
Xs
···js
∇k Ti1 ···ir = r
+ Ti1 ···ir Γkl − Tij11···i Γl .
q−1 liq+1 ···ir kiq
∂uk p=1 q=1

Exercise 8.8. Show that (8.6) can be rephrased as:

∇k gij = 0 for any i, j, k.

Exercise 8.9. Show that the Codazzi equation (Theorem 7.30) is equivalent to
∇i hjk = ∇j hik
for any i, j, k.

Since ∇X is tensorial in the X-slot, given a (r, s)-tensor T , one can then view X as
an input and define a new (r + 1, s)-tensor ∇T as

(∇T )(X, Y1 , · · · , Yr , α1 , · · · , αs ) := (∇X T )(Y1 , · · · , Yr , α1 , · · · , αs ).

∂
Locally, given a vector field Y = Y i ∂u i
, we have:

∂ ∂
∇j Y = (∇j Y i ) =⇒ ∇Y := duj ⊗ (∇j Y ) = (∇j Y i ) duj ⊗ .
∂ui ∂ui
220 8. Riemannian Manifolds

Now that ∇Y is a (1, 1)-tensor, one can then differentiate it again:

∇l (∇Y )
 
i j ∂
= ∇l (∇j Y ) du ⊗
∂ui
∂(∇j Y i ) j ∂ ∂ ∂
= du ⊗ − (∇j Y i )Γjlp dup ⊗ + (∇j Y i )Γpli duj ⊗
∂ul ∂ui ∂ui ∂up
i
 
∂(∇j Y ) ∂
= − (∇p Y i )Γplj + (∇j Y p )Γilp duj ⊗ .
∂ul ∂ui
Similarly, we will define ∇l ∇j Y i to be the local component such that
 
∂ ∂
∇l (∇j Y i ) duj ⊗ =: (∇l ∇j Y i ) duj ⊗
∂ui ∂ui
so that
∂(∇j Y i )
∇l ∇j Y i = − (∇p Y i )Γplj + (∇j Y p )Γilp
∂ul
∂Y i ∂Y i
   
∂
= + Y k Γijk − + Y k Γipk Γplj
∂ul ∂uj ∂up
p
 
∂Y
+ + Y k Γpjk Γilp .
∂uj

Exercise 8.10. Let α = αi dui be a 1-form, compute the local expression of ∇j αi

and ∇l ∇j αi .

It is important to note that ∇i and ∇i are different! The former was discussed above,
but we define
∇i := g ij ∇j .
For instance, when acting on scalar functions, we have
∂f ∂f
∇i f = whereas ∇i f = g ij ∇j f = g ij .
∂ui ∂uj
∂
For a vector field Y = Y i ∂u i
, we have:
∂Y i
 
∇j Y i = g jk ∇k Y i = g jk + Y l Γikl .
∂uk
∂
While we can define ∇Y as a (1, 1)-tensor (∇i Y j ) dui ⊗ ∂u j
, it also makes sense to regard
it as a (0, 2)-tensor:
∂ ∂ ∂ ∂
(∇i Y j ) ⊗ = (g ik ∇k Y j ) ⊗ .
∂ui ∂uj ∂ui ∂uj
Very often, it is not difficult to judge whether ∇Y means a (1, 1)-tensor or a (0, 2)-tensor
according to the context. Similar notations apply to other higher-rank tensors.
However, when acting on scalar functions f , it is a convention to regard ∇f as a
vector field but not a 1-form, that is:
∂ ∂f ∂
∇f := (∇i f ) = g ij
∂ui ∂uj ∂ui
∂f
instead of (∇i f ) dui = ∂u i
dui . We have this convention because we already have another
∂f
symbol to denote ∂u i
dui , namely the exterior derivative df . The vector field ∇f is called
8.2. Levi-Civita Connections 221

the gradient of f with respect to g. When g is the Euclidean metric on Rn , the gradient
∇f is the standard gradient in multivariable calculus.

Exercise 8.11. Given f : M → R is a scalar function on a smooth manifold M .

Check that for any vector field X on M , we have
g(∇f, X) = df (X).
Suppose f (c) is a submanifold of M . Show that for any p ∈ f −1 (c), the vector
−1

∇f (p) is a normal to the level set f −1 (c), i.e.

g(∇f, T ) = 0
−1
for any T ∈ Tp f (c) ⊂ Tp M .

Using the Levi-Civita connection, one can also define divergence and Laplacian
∂
operators. Given a vector field X = X i ∂u i
on a Riemannian manifold (M, g), we define
its divergence with respect to g by
divg X := ∇i X i .
The Laplacian operator ∆ acting on functions is the divergence of the gradient with
respect to g:
∆g f := divg (∇f ) = ∇i ∇i f.
Since ∇i = g ij ∇j and g is constant under ∇, we have
 2 
ij ij ij ∂ f k ∂f
∆g f = ∇i (g ∇j f ) = g ∇i ∇j f = g − Γij .
∂ui ∂uj ∂uk

8.2.3. Contraction of tensors. Given two tensors of different types, one can “cook
them up” to form new tensors. For instance, consider the first fundamental form gij (and
its inverse g ij ), and the second fundamental form hij of a regular Euclidean hypersurface.
They are both (2, 0)-tensors. We can define “cook up” two new tensors by summing up
indices:
g ij hij or g ik hkj .
They are different tensors. The first one g ij hij sums over both i and j, so there is no
free component and it is a scalar function (this function is the mean curvature). For the
second one g ik hkj , we sum them up over k leaving i and j to be free. It gives a new
tensor denoted by, say, A with one lower component and one upper component (i.e. a
∂
(1, 1)-tensor), that if we input ∂u j
into A, it will output
 
∂ ∂ ∂
A = g ik hkj , or equivalently, A = g ik hkj duj ⊗ .
∂uj ∂ui ∂ui
In this case, the operator A is simply the shape operator discussed in Section 7.2. We
often denote the components of A by hij := g ik hkj .
For more complicated tensors, we can “cook them up” in similar ways. For example,
k
given S = Sij dui ⊗ duj ⊗ ∂u∂ k and T = Tji duj ⊗ ∂u
∂
i
, we can form many combinations:

component type full form

∂
Wijl := Sij
k l
Tk (2, 1)-tensor W = Sij Tk dui ⊗ duj ⊗ ∂u
k l
l
Uik := Sij Tkj
l l l j i k
(2, 1)-tensor U = Sij Tk du ⊗ du ⊗ ∂ul∂

k j k j
αi := Sij Tk (1, 0)-tensor α = Sij Tk dui

It is interesting to note that if each summations is over exactly one upper index and
one lower index, then the new tensor produced is independent of local coordinates. Take
222 8. Riemannian Manifolds

S and T above as an example, if we write them in two different coordinate systems:

k ∂ γ ∂
S = Sij dui ⊗ duj ⊗ = Sαβ dv α ⊗ dv β ⊗
∂uk ∂vγ
∂ ∂
T = Tji duj ⊗ = Tβα dv β ⊗ ,
∂ui ∂vα
then it can be easily checked that
k l ∂ γ ∂
(8.9) Sij Tk dui ⊗ duj ⊗ = Sαβ Tγη dv α ⊗ dv β ⊗ .
∂ul ∂vη

Exercise 8.12. Verify (8.9).

k l
However, it is generally not true that Sij Tj gives a well-defined (1, 2)-tensor, since
both j’s are lower indices.
It is interesting and very useful to note that the product rule holds for contractions
of tensors when taking derivatives by the Levi-Civita connection. For example, we have
∇i (Sjk Tqk ) = (∇i Sjk )Tqk + Sjk (∇i Tqk ),
which precisely means
Sjk Tqk duj ⊗ duq = (∇i Sjk )Tqk + Sjk (∇i Tqk ) duj ⊗ duq .



∇ ∂
∂ui

Let’s verify this by direct computations. The LHS equals

∇ ∂ Sjk Tqk duj ⊗ duq


∂ui

∂
= (Sjk Tqk ) duj ⊗ duq + Sjk Tqk (∇i duj ) ⊗ duq + Sjk Tqk duj ⊗ (∇i duq )
∂ui
!
∂Sjk k ∂Tqk
= T + Sjk duj ⊗ duq − Sjk Tqk Γjip dup ⊗ duq − Sjk Tqk duj ⊗ (Γqip dup )
∂ui q ∂ui
!
∂Sjk k ∂Tqk
= T + Sjk duj ⊗ duq − Spk Tqk Γpij duj ⊗ duq − Sjk Tpk Γpiq duj ⊗ duq .
∂ui q ∂ui

The last step follows from relabelling indices of the last two terms. By taking out the
common factor duj ⊗ duq , we have proved
∂Sjk k ∂Tqk
∇i (Sjk Tqk ) = Tq + Sjk − Spk Tqk Γpij − Sjk Tpk Γpiq .
∂ui ∂ui
For the RHS, we recall that:
∂Sjk
∇i Sjk = − Γpij Spk − Γpik Sjp ,
∂ui
∂Tqk
∇i Tqk = − Γpiq Tpk + Γkip Tqp .
∂ui
Combining these results, we can easily get that
(∇i Sjk )Tqk + Sjk (∇i Tqk ) = ∇i (Sjk Tqk )
by cancellations and relabelling of indices.
A similar product rule holds for all other legitimate contractions (i.e. one “up” paired
with one “down”). The proof is similar to the above special case although it is quite
tedious. We omit the proof here.
8.2. Levi-Civita Connections 223

Exercise 8.13. Prove that ∇i (S kl Tkj ) = (∇i S kl )Tkj + S kl ∇i Tkj .

In particular, as we have ∇i gjk = 0 for any i, j, k, one can apply the product rule to
see that gij can be treated as a constant when we differentiate it by ∇. For example:
∇i (gjk Tlk ) = gjk ∇i Tlk .
Recall also that gik g kj = δij . Applying the product rule one can get
∇p (gik g kj ) = 0 =⇒ gik ∇p g kj = 0.
Taking g −1 on both sides, we can get ∇p g kj = 0 as well.
 Chapter 9

Parallel Transport and

Geodesics

“Be grateful for all ordeals, they are the

shortest way to the Divine.”

The Mother

Let (M, g) be a Riemannian manifold. From now on unless otherwise is said, we

will denote ∇ to be the Levi-Civita connection with respect to g. Furthermore, as ∇
coincides with the covariant derivative in case when M is an Euclidean hypersurface and
g is the first fundamental form, we will use also use the term covariant derivative for the
Levi-Civita connection.

9.1. Parallel Transport

9.1.1. Parallel Transport Equation. On an Euclidean space, we can make good
sense of translations of vectors as Tp Rn is naturally identified with Tq Rn for any other
point q. However, on an abstract manifold M , the tangent spaces Tp M and Tq M may
not be naturally related to each other, so it is non-trivial to make sense of translating a
vector V ∈ Tp M to a vector in Tq M .
If one translates a vector V at p ∈ Rn along a path γ(t) connecting p and q, then it is
d
ease to see that dt V (γ(t)) = 0 along the path. In other words, Dγ 0 (t) V = 0 where D is
the directional derivative of Rn . Now on a Riemannian manifold we have the notion of
the covariant derivative ∇. Using ∇ in place of D in Rn , one can still define the notion
of translations as follows:

Definition 9.1 (Parallel Transport). Given a curve γ : [0, T ] → M and a vector V0 ∈

Tγ(0) M , we define the parallel transport of V0 along γ(t) to be the unique solution V (t)
to the ODE:
∇γ 0 (t) V (t) = 0
for any t ∈ [0, T ] with the initial condition V (0) = V0 .

225
226 9. Parallel Transport and Geodesics

In terms of local coordinates F (u1 , · · · , un ), the parallel transport equation can be

∂
expressed as follows. Let γ(t) = F (γ 1 (t), · · · , γ n (t)) and V (t) = V i (t) ∂u i
, then

dγ i ∂
γ 0 (t) =
dt ∂ui
 
∂
∇γ 0 (t) V (t) = ∇ dγ i ∂ V j (t)
dt ∂ui ∂uj
i j
 
dγ ∂V ∂ j k ∂
= + V Γij
dt ∂ui ∂uj ∂uk
dV j ∂ dγ i ∂
= + V j Γkij .
dt ∂uj dt ∂uk
Therefore, the parallel transport equation ∇γ 0 (t) V (t) = 0 is equivalent to

dV k dγ i
+ V j Γkij = 0 for any k.
dt dt
Notably, the equation depends on γ 0 (t) only but not on γ(t) itself, so it also makes sense
of parallel transporting V0 along a vector field X. It simply means parallel transport of
V0 along the integral curve γ(t) such that γ 0 (t) = X(γ(t)).
Even though the parallel transport equation is a first-order ODE whose existence
and uniqueness of solutions are guaranteed, it is often impossible to solve it explicitly
(and often not necessary to). Nonetheless, there are many remarkable uses of parallel
transports. When we have an orthonormal basis {ei }ni=1 of Tp M at a fixed point p ∈
M , one can naturally extend it along a curve to become a moving orthonormal basis
{ei (t)}ni=1 along that curve. The reason is that parallel transport preserves the angle
between vectors.
Given a curve γ(t) on a Riemannian manifold (M, g), and let V (t) be the parallel
transport of V0 ∈ Tp M along γ, and W (t) be that of W0 ∈ Tp M . Then one can check
easily that
d



g V (t), W (t) = g ∇γ 0 V, W + g(V, ∇γ 0 W = 0
dt
by the parallel transport equations ∇γ 0 V = ∇γ 0 W = 0. Therefore, we have



g V (t), W (t) = g V0 , W0
for any t. In particular, if V0 and W0 are orthogonal, then so are V (t) and W (t) for any t.
If we take W0 = V0 , by uniqueness theorem p of ODE we have W (t) ≡ V (t). The above
result shows the length of V (t), given by g(V (t), V (t)), is also a constant. Hence, the
parallel transport of an orthonormal basis remains to be orthonormal along the curve.

9.1.2. Holonomy and de Rham Splitting Theorem. Another remarkable conse-

quence of the above observation is that parallel transport can be used to define an O(n)-
action on Tp M . Consider a closed piecewise smooth curve γ(t) where γ(0) = p ∈ M . The
curve needs not to be smooth at the closing point p. Given a vector V ∈ Tp M , we parallel
transport it subsequentially along each smooth segment of γ. Precisely, suppose γ is
defined on [t0 , t1 ] ∪ [t1 , t2 ] ∪ · · · ∪ [tk−1 , tk ] with γ(t0 ) = γ(tk ) = p and that γ is smooth
on each (ti , ti1 ) and is continuous on [t0 , tk ]. We solve the parallel transport equation
to get V (t) which is continuous on [t0 , tk ] and satisfies ∇γ 0 V = 0 on each (ti , ti−1 ). We
denote the final vector V (tk ) by Pγ (V ) ∈ Tp M . It then defines a map
Pγ : Tp M → Tp M.
9.1. Parallel Transport 227

By linearity of ∇γ 0 , it is easy to show that Pγ is a linear map. Moreover, as Pγ (V ) and

V have the same length, we have in fact Pγ ∈ O(Tp M ), the orthogonal group acting on
Tp M .
The set of all Pγ ’s, where γ is any closed piecewise smooth curve based at p, is in
fact a group with multiplication given by compositions, and inverse given by parallel
transporting vectors along the curve backward. We call this:

Definition 9.2 (Holonomy Group). The holonomy group of (M, g) based at p ∈ M is

given by:
Holp (M, g) := {Pγ : γ is a closed piecewise smooth curve on M based at p}.

Exercise 9.1. Let (M, g) be a connected complete Riemannian manifold, and p and
q be two distinct points on M . Show that Holp (M, g) and Holq (M, g) are related by
conjugations (hence are isomorphic).

When (M, g) is the flat Euclidean space Rn , parallel transport is simply translations.
Any vector will end up being the same vector after transporting back to its based point.
Therefore, Holp (Rn , δ) is the trivial group for any p.
For the round sphere S2 in R3 , we pick two points P , Q on the equator, and mark
N to be the north pole. Construct a piecewise great circle path P → N → Q → P , then
one can see that what parallel transport along this path does is a rotation on vectors in
TP S2 by an angle depending on the distance between P and Q. For instance, when P
and Q are antipodal, the parallel transport map is rotation by π. When P = Q, then the
parallel transport map is simply the identity map. By varying the position of Q, one can
obtain all possible angles from 0 to 2π. To explain this rigorously, one way is to solve the
parallel transport equation in Definition 9.1. There is a more elegant to explain this after
we learn about geodesics.
It is remarkable that the holonomy group reveals a lot about the topological structure
about a Riemannian manifold, and is a extremely useful tool for classification problems
of manifolds. There is a famous theorem due to Ambrose-Singer that the Lie algebra of
Holp (M, g) is related to how curved (M, g) is around p.
Another beautiful theorem which demonstrates the usefulness of parallel transport is
the following one due to de Rham. It is widely used in Ricci flow to classify the topology
of certain class of manifolds by decomposing them into lower dimensional ones.

Theorem 9.3 (de Rham Splitting Theorem - local version). Suppose the tangent bundle
T M of a Riemannian manifold (M, g) can be decomposed orthogonally into T M = E1 ⊕E2
such that each of Ei ’s is invariant under parallel transport, i.e. whenever V ∈ Ei , any
parallel transport of V stays in Ei . Then, M is locally a product manifold (N1 , h) × (N2 , k)
such that T Ni = Ei , and g is a locally product metric π1∗ h + π2∗ k.

Proof.
 ∂The proof begins byconstructing a local coordinate system {xi , yα } such that
∂
span ∂x i
= E 1 and span ∂yα = E2 . This is done by Frobenius’ Theorem which
asserts that such that such local coordinate system would exist if each Ei is closed under
the Lie brackets, i.e. whenever X, Y ∈ Ei , we have [X, Y ] ∈ Ei .
Let’s prove E1 is closed under the Lie brackets (and the proof for E2 is exactly the
same). The key idea is to use (8.7) that:
[X, Y ] = ∇X Y − ∇Y X.

First pick an orthonormal basis {ei , eα } at a fixed point p ∈ M , such that span ei = E1
and span eα = E2 . Extend them locally around p using parallel transport, then one
228 9. Parallel Transport and Geodesics

would have ∇ei = ∇eα = 0 for any i and α. Now suppose X, Y ∈ E1 , we want to show
∇X Y ∈ E1 . Since g(Y, eα ) = 0 and ∇X eα = 0, we have by (8.8):


0 = X g(Y, eα ) = g(∇X Y, eα ).
This shows ∇X Y ⊥ eα for any α, so ∇X Y ∈ E1 . The same argument shows ∇Y X ∈ E1 ,
and so does [X, Y ].
Frobenius’
 ∂  ∂ that there exists a local coordinate system {xi , yα } with
Theorem asserts
span ∂x i
= E1 and span ∂yα = E2 . Next, we argue that the metric g is locally
expressed as:
g = gij dxi ⊗ dxj + gαβ dy α ⊗ dy β ,
and that gij depends only on {xi }, and gαβ depends only on {yα }. To show this, we
j β
argue that Christoffel symbols of type Γα i
ij , Γiα , Γiα , and Γαβ are all zero.
∂ ∂
Consider V0 := ∂xi ∈ Tp M , and let V (t) be its parallel transport along ∂xj . Since
V (t) ∈ E1 , it can be locally expressed as V (t) = V k
(t) ∂x∂ k .
We then have

∂V k ∂
 
k l ∂ α ∂
0 = ∇j V (t) = +V Γjk + Γjk .
∂xj ∂xk ∂xl ∂yα
By equating coefficients, we get V k Γα α
jk = 0. In particular, it implies Γik = 0 at p
∂
since V (0) = ∂x i
at p. Apply the same argument at other points covered by the local
coordinates {xi , yα }, we have proved Christoffel symbols of type Γα
ik all vanish. Using
a similar argument by parallel transporting V0 along ∂y∂α instead, one can also show
Γβαi = 0. The other combinations Γjiα and Γiαβ follow from symmetry argument.
Finally, we conclude that
∂
gij = Γkαi gkj + Γkαj gki = 0,
∂yα
so gij depends only on {xi }. Similarly, gαβ depends only on {yα }. It completes our
proof. 

Remark 9.4. There is a global version of de Rham splitting theorem. If one further
assumes that M is simply-connected, then the splitting would be in fact global. Generally
if (M, g) satisfies the hypothesis of the de Rham splitting theorem, one can consider
its universal cover (Mf, π ∗ g) which is simply-connected. Applying the theorem one can
f, π ∗ g) splits isometrically as a product manifold.
assert that (M 

9.1.3. Parallel Vectors and Tensors. A vector field X satisfying ∇X = 0 is called

a parallel vector field. It has the property that for any curve γ(t) on M with γ(0) = p, if
we parallel transport X(p) ∈ Tp M along γ one would get the vectors X(γ(t)) for any
t. If such a vector field exists and is non-vanishing, then by de Rham splitting theorem
applied to T M = span{X} ⊕ span{X}⊥ , one can assert that locally the manifold M
splits off as a product of a 1-dimensional manifold and an (dim M − 1)-manifold.
As one can take covariant derivatives on tensors of any type, we also have a notion
of parallel tensors. A tensor T is said to be parallel if ∇T = 0. Notably, the Riemannian
metric g itself is a parallel (2, 0)-tensor, as ∇i gjk = 0.
Now consider a parallel (1, 1)-tensor T , which operates on vector fields and output
vector fields. Given λ ∈ R, the λ-eigenspace Eλ of T turns out is invariant under parallel
transport. To see this, consider V ∈ Eλ at p such that T (V ) = λV , or in terms of local
coordinates:
Tij V i = λV j .
9.1. Parallel Transport 229

Now parallel transport V along an arbitrarily given curve γ starting from P , then we
have:
∇γ 0 Tij V i − λV j = (∇γ 0 Tij )V i + Tij ∇γ 0 V i − λ∇γ 0 V j = 0

by the fact that T is parallel. Therefore, T (V ) − λV satisfies the parallel transport

equation along γ. As it equals 0 at the starting point p, it remains so along the curve.
This shows V (t) is also an eigenvector of T with eigenvalue λ. If we assume further that
T is self-adjoint respect to g, i.e. g(T (X), Y ) = g(X, T (Y )) for any vector fields X and
Y , then the eigenspaces of T with distinct eigenvalues would be orthogonal to each other
(by freshmen linear algebra). The de Rham splitting theorem can be applied to show the
manifold splits according to the eigenspaces of T .
Exercise 9.2. Consider a Riemannian manifold (M, g) and a parallel (1, 1)-tensor J
such that J 2 = −id. Suppose further that
g(JX, JY ) = g(X, Y )
for any vector fields X and Y . Consider the two-form defined by
ω(X, Y ) := g(JX, Y ).
(1) Verify that ω is indeed a two-form, i.e. ω(X, Y ) = −ω(Y, X).
(2) Show that ∇ω = 0 and dω = 0.
230 9. Parallel Transport and Geodesics

9.2. Geodesic Equations

One classic problem in differential geometry is to find the shortest path on a surface
connecting two distinct points. Such a path is commonly called a geodesic, or more
accurately minimizing geodesic. To find such a path is a problem in calculus of variations.
In this section, we will derive an equation for us to find geodesics connecting two
given points. The technique for deriving such an equation is very common in calculus of
variations. We consider a family of curves γs (t) connecting the same given points with
γ0 (t) being the candidate curve for the geodesic. Then, we compute the first derivative
d
ds L(γs ) of the length L(γs ), and see under what condition on γ0 would guarantee that
the first derivative at s = 0 equals to 0.

Proposition 9.5. Let γs (t) : (−ε, ε) × [a, b] → M be a 1-parameter family of curves with
γs (a) = p and γs (b) = q for any s ∈ (−ε, ε). Here s is the parameter of the family, and t
is the parameter of each curve γs . Then, the first variation of the arc-length is given by:
Z b   0 
d ∂γ γ0 (t)
(9.1) L(γs ) = − g , ∇γ00 (t) dt.
ds s=0 a ∂s s=0 |γ00 (t)|
Therefore, if γ0 minimizes the length among all variations γs , it is necessarily that
 0 
γ0 (t)
(9.2) ∇γ0 (t)
0 = 0 for any t ∈ [a, b].
|γ00 (t)|

Proof. For simplicity, we denote

∂γ ∂γ
S := , T := .
∂s ∂t
Z bp
d d
L(γs ) = g(T, T ) dt
ds ds a
Z b
1 d
= g(T, T ) dt
a 2 |T | ds
Z b
1
= · 2g (∇S T, T ) dt
a 2 |T |
Z b
1
= g(∇T S, T ) dt.
a |T |
Here we have used (8.7) so that ∇S T = ∇T S + [S, T ] and
      
∂ ∂ ∂ ∂
[S, T ] = γ∗ , γ∗ = γ∗ , = γ∗ (0) = 0.
∂s ∂t ∂s ∂t
Now evaluate the above derivative at s = 0. We get
Z b   Z b   
d T T
L(γs ) = g ∇T S, dt =− g S, ∇T dt
ds s=0 a |T | s=0 a |T | s=0
as desired. We have used integration by parts in the last step. The boundary term
vanishes because γs (a) and γs (b) are both independent of s, and hence ∂γ
∂s = 0 at both
t = a and t = b.

9.2. Geodesic Equations 231

Exercise 9.3. Compute the first variation of the energy functional E(γs ) of a 1-
parameter family of curves γs (t) : (−ε, ε) × [a, b] → M with γs (a) = p and γs (b) = q
for any s. The energy functional is given by
1 b
Z  
∂γs ∂γs
E(γs ) := g , dt.
2 a ∂t ∂t
Note that it is different from arc-lengths: there is not square root in the integrand.
Furthermore, show that if γ0 minimizes E(γs ), then one has ∇γ00 (t) γ00 (t) = 0 for
any t ∈ [a, b].

Since every (regular) curve can be parametrized by constant speed, we can assume
without loss of generality that |γ00 (t)| ≡ C, so that (9.2) can be rewritten as
(9.3) ∇γ00 γ00 = 0.
We call (9.3) the geodesic equation and any constant speed curve γ0 (t) that satisfies (9.3)
is called a geodesic. For simplicity, let’s assume from now on that every geodesic has a
constant speed. We can express (9.3) using local coordinates. Suppose under a local
coordinate system F (u1 , · · · , un ), the coordinate representation of γ0 is given by:
F −1 ◦ γ0 (t) = (γ 1 (t), · · · , γ n (t)).
Then γ00 (t) is given by
dγ i ∂
γ00 := ,
dt ∂ui
so we have
dγ j ∂
 
∇γ00 γ00 = ∇γ00
dt ∂uj
d2 γ j ∂ dγ j ∂
= 2
+ ∇ dγ i ∂
dt ∂uj dt dt ∂ui ∂uj

d2 γ j ∂ dγ i dγ j k ∂
= + Γ
dt2 ∂uj dt dt ij ∂uk
 2 k
dγ i dγ j k

d γ ∂
= 2
+ Γij
dt dt dt ∂uk
Hence, the geodesic equation (9.3) is locally a second-order ODE (assuming γ0 is arc-
length parametrized):
d2 γ k dγ i dγ j k
(9.4) + Γ = 0 for any k.
dt 2 dt dt ij
Example 9.6. Consider the round sphere S2 in R3 which, under spherical coordinates,
has its Riemannian metric given by
g = sin2 θ dϕ2 + dθ2 .
By direct computations using (7.15), we get:
Γθϕϕ = − sin θ cos θ
Γϕ ϕ
ϕθ = Γϕθ = cot θ

and all other Christoffel symbols are zero. For instance,

   
1 ∂ ∂ ∂ 1 ∂ ∂ ∂
Γθϕϕ = g θθ gϕθ + gϕθ − gϕϕ + g θϕ gϕϕ + gϕϕ − gϕϕ .
2 ∂ϕ ∂ϕ ∂θ 2 ∂ϕ ∂ϕ ∂ϕ
Note that the second term vanishes since [g] is diagonal under this local coordinate
system.
232 9. Parallel Transport and Geodesics

Write γ(t) locally as (γ ϕ (t), γ θ (t)), then the geodesic equations are given by

d2 γ ϕ dγ ϕ dγ θ ϕ dγ θ dγ ϕ ϕ
+ Γ + Γ =0
dt2 dt dt ϕθ dt dt θϕ
d2 γ θ dγ ϕ dγ ϕ θ
+ Γ = 0,
dt2 dt dt ϕϕ
that is
d2 γ ϕ dγ ϕ dγ θ
2
+2 cot θ = 0
dt dt dt
 ϕ 2
d2 γ θ dγ
2
− sin θ cos θ = 0.
dt dt

Clearly, the path with (γ θ (t), γ ϕ (t)) = (t, c), where c is a constant, is a solution to the
system. This path is a great circle. 

Exercise 9.4. Show that geodesic is an isometric property, i.e. if Φ : (M, g) → (M

f, ge)
is an isometry, then γ is a geodesic of (M, g) if and only if Φ(γ) is a geodesic of
(M
f, ge).

Exercise 9.5. Consider the surface of revolution F (u, θ) = (x(u) cos θ, x(u) sin θ, z(u))
given by a profile curve (x(u), 0, z(u)) on the xz-plane. Show that the profile curve
itself is a geodesic of the surface.

Exercise 9.6. Consider the hyperbolic space with the upper-half plane model, i.e.
dx2 + dy 2
H2 = {(x, y) ∈ R2 : y > 0}, g = .
y2
Show that straight lines normal to the x-axis, as well as semi-circles intersecting the
x-axis orthogonally, are geodesics of the hyperbolic space.

Since (9.4) is a second-order ODE system, it has local existence and uniqueness when
given both initial position and velocity. That is, given p ∈ M and V ∈ Tp M , there exists a
unique (constant speed) geodesic γV : (−ε, ε) → M such that γV (0) = p and γV0 (0) = V .
Let c > 0, the geodesic γcV (starting from p) is one that the speed is c |V |, and
so it travels c times faster than γV does. One should then expect that γcV (t) = γV (ct)
provided ct is in the domain of γV . It can be easily shown to be true using uniqueness
theorem ODE, since γV (ct) t=0 = γV (0) = p and

d
γV (ct) = cγV0 (ct) t=0
= cγV0 (0) = cV.
dt t=0

Therefore, γV (ct) is a curve initiating from p with initial velocity cV . It can be shown
using the chain rule that γV (ct) also satisfies (9.4). Hence, by uniqueness theorem of
ODE, we must have γV (ct) = γcV (t).
Note that a geodesic γ(t) may not be defined globally on (−∞, ∞). Easy counter-
examples are straight-lines ax + by = 0 on R2 \{0}. If every geodesic γ passing through p
on a Riemannian manifold (M, g) can be extended so that its domain becomes (−∞, ∞),
then we say (M, g) is geodesically complete at p. If (M, g) is geodesically complete is every
point p ∈ M , then we say (M, g) is geodesically complete.
9.2. Geodesic Equations 233

On a connected Riemannian manifold (M, g) there is a natural metric space structure,

with the metric1 d : M × M → R defined as:
d(p, q)
:= inf{L(γ)|γ : [0, τ ] → M is a piecewise-C ∞ curve such that γ(0) = p, γ(τ ) = q}.

Exercise 9.7. Check that d is a metric on M (in the sense of a metric space).

On a metric space, we can talk about Cauchy completeness, meaning that every
Cauchy sequence in M with respect to d converges to a limit in M . Interestingly, the two
notions of completeness are equivalent! It thanks to:

Theorem 9.7 (Hopf-Rinow). Let (M, g) be a connected Riemannian manifold, and let
d : M × M → R be the distance function induced by g. Then the following are equivalent:
(1) There exists p ∈ M such that (M, g) is geodesically complete at p
(2) (M, g) is geodesically complete
(3) (M, d) is Cauchy complete

We omit the proof in this note as it “tastes” differently from other parts of the course.
Interested readers may consult any standard reference of Riemannian geometry to learn
about the proof.

1Here metric means the distance function of a metric space, not a Riemannian metric on a manifold!
234 9. Parallel Transport and Geodesics

9.3. Exponential Map

9.3.1. Definition of the Exponential Map. Thanks to the Hopf-Rinow’s Theorem,
we will simply call geodesically complete Riemannian manifold (M, g) to be a complete
Riemannian manifold. Such a metric g is called a complete Riemannian metric. On a
complete Riemannian manifold (M, g), given any p ∈ M and V ∈ Tp M , the unique
geodesic γV (t) with γV (0) = p and γV0 (0) = V is defined for any t ∈ (−∞, ∞), and in
particular, γV (1) is well-defined. This is a point on M which is |V | unit away from p
along the geodesic in the direction of V . The map V 7→ γV (1) is an important map called:

Definition 9.8 (Exponential Map). Let (M, g) be a complete Riemannian manifold. Fix
a point p ∈ M and given any tangent vector V ∈ Tp M , we consider the unique geodesic
γV with initial conditions γV (0) = p and γV0 (0) = V . We define expp : Tp M → M by
expp (V ) := γV (1).
The map expp is called the exponential map at p.

Remark 9.9. Standard ODE theory shows expp is a smooth map. For detail, see e.g. p.74
of John M. Lee’s book. 

Lemma 9.10. Let p ∈ M which is a complete Riemannian manifold and V ∈ Tp M be

a fixed tangent. The push-forward of expp at 0 ∈ Tp M , (expp )∗0 : T0 (Tp M ) → Tp M , is
given by (expp )∗0 (V ) = V for any V ∈ T0 (Tp M ). Here we identify T0 (Tp M ) with Tp M .

Proof. The key observation is the rescaling property γcV (t) = γV (ct), which implies
expp (cV ) = γcV (1) = γV (c), and hence
d d
(expp )(0 + tV ) = γV (t) = γV0 (t) = V.
dt t=0 dt t=0
In other words, (expp )∗0 (V ) = V . 

Clearly, (expp )∗0 is invertible, and by inverse function theorem we can deduce that
expp is locally a diffeomorphism near 0 ∈ Tp M :

Corollary 9.11. There exists an open ball B(0, ε) on Tp M such that expp B(0,ε)
is a
diffeomorphism onto its image.

To sum up, expp (V ) is well-defined for all V ∈ Tp M provided that (M, g) is complete.
However, it may fail to be a diffeomorphism if the length of V is too large. The maximum
possible length is called:

Definition 9.12 (Injectivity Radius). Let (M, g) be a complete Riemannian manifold,

and let p ∈ M . The injectivity radius of (M, g) at p is defined to be:
inj(p) := sup{r > 0 : expp B(0,r)
is a diffeomorphism onto its image}.
The injectivity radius of (M, g) is the minimum possible injectivity radius over all points
on M , i.e.
inj(M, g) := inf inj(p).
p∈M
9.3. Exponential Map 235

Example 9.13. Injectivity radii may not be easy to be computed, but we have the
following intuitive examples:
• For the round sphere S2 with radius R, the injectivity radius at every point is given
by πR.
• For the flat Euclidean space, the injectivity radius at every point is +∞.
• For a round torus obtained by rotating a circle with radius r, the injectivity radius
at every point is πr.


9.3.2. Geodesic Normal Coordinates. A consequence of Corollary 9.11 is that

expp gives a local parametrization of M covering p. With a suitable modification, one
can construct an important local coordinate system {ui }, called the geodesic normal
coordinates:

Proposition 9.14 (Existence of Geodesic Normal Coordinates). Let (M, g) be a Riemann-

ian manifold. Then at every p ∈ M , there exists a local parametrization G(u1 , · · · , un )
covering p, such that all of the following hold:
 
∂ ∂
(1) gij := g , = δij for any i, j at the point p;
∂ui ∂uj
(2) Γkij = 0 for any i, j, k at p; and
∂gij
(3) = 0 for any i, j, k at p
∂uk
This local coordinate system is called the geodesic normal coordinates at p.

Proof. The key idea is to slightly modify the map expp . By Gram-Schmidt’s orthogonal-
ization, one can take a basis {êi }ni=1 for Tp M which are orthonormal with respect to g,
i.e. g(êi , êj ) = δij . Then we define an isomorphism E : Rn → Tp M by:
E(u1 , · · · , un ) := ui êi .
Next we define the parametrization G : E −1 (B(0, ε)) → M by G := expp ◦E. Here ε > 0
is sufficiently small so that expp B(0,ε) is a diffeomorphism onto its image.
Next we claim such G satisfies all three required conditions. To prove (1), we
compute that:
   
∂ ∂G ∂ ∂
(p) := = G∗ (0) = (expp )∗ ◦ E∗ (0) .
∂ui ∂ui 0 ∂ui ∂ui
| {z } | {z }
∈Tp M ∈T0 Rn
∂
From the definition of E, we have E∗ ( ∂u i
) = êi , and also at p, we have:
(expp )∗0 (êi ) = êi
∂
according to Lemma 9.10. Now we have proved ∂u i
(p) = êi . Hence, (1) follows directly
from the fact that {êi } is an orthonormal basis with respect to g:
 
∂ ∂
g (p), (p) = g(êi , êj ) = δij .
∂ui ∂uj
Next we claim that (2) is an immediate consequence of the geodesic equation (9.4).
Consider the curve γ(t) = expp (t(êi + êj )), which is a geodesic passing through p. Then,
the local expression of γ(t) is given by:
G−1 ◦ γ(t) = (γ 1 (t), · · · , γ n (t))
236 9. Parallel Transport and Geodesics

where (
t if k = i or j
γ k (t) = .
0 otherwise
By (9.4), we have:
d2 γ k i
k dγ dγ
j

2
+Γij = 0.
|dt
{z } |dt{zdt}
=0 =1
This shows Γkij = 0 along γ(t) (see footnote2), and in particular Γkij = 0 at p = γ(0).
Finally, (3) is an immediate consequence of (8.6).


The geodesic normal coordinates are one of the “gifts” to Riemannian geometry. In
the later part of the course when we discuss geometric flows, we will see that it simplifies
many tedious tensor computations. It is important to note that different points give
rise to different geodesic normal coordinate systems! Note also that a geodesic normal
coordinate system satisfies the three properties in Proposition 9.14 at one point p only.
It is not always possible to pick a local coordinate system such that gij = δij on the whole
chart, unless the Riemannian manifold is locally flat.
Here is one demonstration of how useful geodesic normal coordinates are. Suppose
M is a Riemannian manifold with a smooth family of Riemannian metrics g(t), t ∈ [0, T ),
such that g(t) evolves in the direction of a family of symmetric 2-tensor v(t), i.e.
∂
g(t) = v(t).
∂t
∂
In terms of local components, we may write ∂t gij (t) = vij (t) where
g(t) = gij (t) dui ⊗ duj , v(t) = vij (t) dui ⊗ duj .
We want to derive the rate of change of the Christoffel symbols Γkij (t)’s, which change
over time.
∂ k
First we fix a time and a point (t0 , p). Note that ∂t Γij is tensorial since the difference
k k k
Γij (t) − Γij (t0 ) is tensorial (even though Γij (t) itself isn’t). Therefore, if one can express
∂ k
∂t Γij at (t0 , p) as another tensorial quantity using one particular local coordinate system,
then this expression holds true under all other local coordinate systems.
Let’s choose the geodesic normal coordinates {ui } at p with respect to the metric
g(t0 ). Then, we have
∂gij
gij (t0 , p) = δij , Γkij (t0 , p), (t0 , p) = 0.
∂uk
Recall that from (7.15) we have
 
k 1 kl ∂gjl ∂gil ∂gij
Γij = g + − .
2 ∂ui ∂uj ∂ul
Taking time derivatives, we have:
    
∂ k 1 ∂ kl ∂gjl ∂gil ∂gij 1 kl ∂ ∂gjl ∂ ∂gil ∂ ∂gij
Γ = g + − + g + − .
∂t ij 2 ∂t ∂ui ∂uj ∂ul 2 ∂ui ∂t ∂uj ∂t ∂ul ∂t

2Note that Γk here is in fact Γk (γ(t)), so one can only claim Γk = 0 along γ(t) but not at other points. Also, γ(t)
ij ij ij
itself depends on the choice of i and j. This argument does not imply there is a geodesic on which all Christoffel’s symbols
vanish.
9.3. Exponential Map 237

∂g
However, when evaluating at (t0 , p), the space derivatives ∂ujli ’s all become zero, so we
can ignore the first term above and obtain
∂Γkij
 
1 ∂vjl ∂vil ∂vij
(t0 , p) = g kl + − .
∂t 2 ∂ui ∂uj ∂ul (t0 ,p)
∂v
Note that ∂ujli ’s are not tensorial, but at (t0 , p) all Christoffel symbols are zero, so by
Exercise 8.7 we have
∂vjl
= ∇i vjl at (t0 , p),
∂ui
and so
∂Γkij 1
= g kl (∇i vjl + ∇j vil − ∇l vij ) at (t0 , p).
∂t 2
Note that now both sides are tensorial, so given that the above equation holds for one
particular coordinate system, it holds for all other coordinate systems. Say {yα }, we will
still have
∂Γγαβ 1
= g γη ∇α vβη + ∇β vαη − ∇η vαβ at (t0 , p).


∂t 2
Furthermore, even though this derivative expression holds at (t0 , p) only, we can repeat
the same argument using geodesic normal coordinates with respect g(t) at other time
and at other points, so that we can conclude under any local coordinates on M , we have:
∂Γkij 1
= g kl (∇i vjl + ∇j vil − ∇l vij ) .
∂t 2

Exercise 9.8. Given g(t) is a smooth family of Riemannian metrics on M satisfying

∂gij
= vij
∂t
where v(t) is a smooth family of symmetric 2-tensors on M . Recall that the Laplacian
of a scalar function f with respect to g is defined to be
∆g f = g ij ∇i ∇j f,
so it depends on t if g(t) is time-dependent. Compute the evolution formula for:
∂
∆g(t) f
∂t
where f is a fixed (time-indepedent) scalar function.
Hint: first show that
∂ ij ∂
g = −g ip g jq gpq .
∂t ∂t
 Chapter 10

Curvatures of Riemannian
Manifolds

“Arc, amplitude, and curvature sustain

a similar relation to each other as time,
motion, and velocity, or as volume,
mass, and density.”

Carl Friedrich Gauss

10.1. Riemann Curvature Tensor

10.1.1. Motivations and Definitions. Recall that Gauss’s Theorema Egregium is an
immediate consequence of the Gauss’s equation, which can be written using covariant
derivatives as:
∂ ∂ ∂
∇i ∇j − ∇ j ∇i = g ql (hjk hli − hik hlj ) .
∂uk ∂uk ∂uq
Taking the inner product with ∂u∂ p on both sides, we get
 
∂ ∂ ∂
g ∇i ∇j − ∇ j ∇i , = g ql (hjk hli − hik hlj ) gpq
∂uk ∂uk ∂up
= δpl (hjk hli − hik hlj )
= hjk hpi − hik hpj .
In dimension 2, we get
det[h] = h11 h22 − h212 = g(∇1 ∇2 ∂2 − ∇2 ∇1 ∂2 , ∂1 ).
It motivates us to define a (3, 1)-tensor T and a (4, 0)-tensor S so that
T (∂1 , ∂2 , ∂2 ) = ∇1 ∇2 ∂2 − ∇2 ∇1 ∂2 and S(∂1 , ∂2 , ∂2 , ∂1 ) = g(∇1 ∇2 ∂2 − ∇2 ∇1 ∂2 , ∂1 ),
then the Gauss curvature in the 2-dimension case can be expressed in tensor notations as
q
T122 g1q S1221
K= 2 = g g 2 .
g11 g22 − g12 11 22 − g12

Naturally, one may attempt:

T (X, Y, Z) = ∇X (∇Y Z) − ∇Y (∇X Z)

239
240 10. Curvatures of Riemannian Manifolds

for the (3, 1)-tensor. However, even though it indeed gives T122 = ∇1 ∇2 ∂2 − ∇2 ∇1 ∂2 ,
one can easily verify that such T is not tensorial. We modify such an T a bit and define
the following very important tensors in Riemannian geometry:

Definition 10.1 (Riemann Curvature Tensors). Let (M, g) be a Riemannian manifold,

then its Riemann curvature (3, 1)-tensor is defined as
Rm(3,1) (X, Y )Z := ∇X (∇Y Z) − ∇Y (∇X Z) − ∇[X,Y ] Z,
and the Riemann curvature (4, 0)-tensor is defined as:
Rm(4,0) (X, Y, Z, W ) := g Rm(3,1) (X, Y )Z, W .

If the tensor type is clear from the context, we may simply call it the Riemann curvature
tensor and simply denote it by Rm. Alternative notations include R, R, etc.

∂ ∂
When X = ∂ui and Y = ∂uj , we have [X, Y ] = 0, so that it still gives

Rm(3,1) (∂1 , ∂2 )∂2 = ∇1 ∇2 ∂2 − ∇2 ∇1 ∂2 ,

and at the same time Rm(3,1) is tensorial.
Exercise 10.1. Verify that Rm(3,1) is tensorial, i.e.
Rm(3,1) (f X, Y )Z = Rm(3,1) (X, f Y )Z = Rm(3,1) (X, Y )f Z = f Rm(3,1) (X, Y )Z.

It is clear from the definition that Rm(3,1) (X, Y )Z = −Rm(3,1) (Y, X)Z. That explains
why we intentionally write its input vectors by (X, Y )Z instead of (X, Y, Z), so as to
emphasize that X and Y are alternating.
We express the local components of Rm(3,1) by Rijk
l
, so that
∂
Rm(3,1) = Rijk
l
dui ⊗ duj ⊗ duk ⊗
∂ul
where
 
∂ ∂ ∂ ∂
l
Rijk = Rm(3,1) ,
∂ul ∂ui ∂uj ∂uk



= ∇i ∇j ∂k − ∇j ∇i ∂k − ∇[∂i ,∂j ] ∂k
!
∂Γljk ∂Γlik p l p l ∂
= − + Γjk Γip − Γik Γjp .
∂ui ∂uj ∂ul

Therefore, the local components of Rm(3,1) are given by

∂Γljk ∂Γlik
l
Rijk = − + Γpjk Γlip − Γpik Γljp .
∂ui ∂uj
For the (4, 0)-tensor, the local components are given by:
p p
Rijkl = Rm(4,0) (∂i , ∂j , ∂k , ∂l ) = g Rm(3,1) (∂i , ∂j )∂k , ∂l = g(Rijk



∂p , ∂l = gpl Rijk .

10.1.2. Geometric Meaning of the Riemann Curvature (3, 1)-Tensor. While

l
Rijk and Rijkl are related to the Gauss curvature in the two dimension case, their
geometric meanings in higher dimensions are not obvious at the first glance. In fact, the
Riemann curvature (3, 1)-tensor has a non-trivial relation with parallel transports. Given
tangent vectors X1 , X2 , Y at p, the vector Rm(3,1) (X1 , X2 )Y at p in fact measures the
defect between the parallel transports of Y along two different paths.
For simplicity, consider local coordinates {ui } such that p has coordinates (0, · · · , 0),
and let’s only consider the form Rm(∂1 , ∂2 )Y .
10.1. Riemann Curvature Tensor 241

Take an arbitrary Y (p) ∈ Tp M . We are going to transport Y (p) in two different

paths. One path is first along the u1 -direction, then along the u2 -direction; another is the
opposite: along u2 -direction first, followed by the u1 -direction.
Let’s consider the former path: first u1 , then u2 . Let Y = Y k ∂u∂ k be the transported
vector, then along the first u1 -segment (u1 , u2 ) = (u1 , 0), the components Y k satisfy the
equation:
∂Y k
∇∂1 Y = 0 =⇒ + Y j Γk1j = 0 for any k.
∂u1
Provided that we are sufficiently close to p, we have the Taylor expansion:

∂Y k 1 ∂2Y k
Y k (u1 , 0) = Y k (0, 0) + (u1 − 0) + (u1 − 0)2 + O(u31 ).
∂u1 (0,0) 2 ∂u21 (0,0)

Using the parallel transport equation, both the first and second derivatives of Y k can be
expressed in terms of Christoffel’s symbols.

∂Y k
= −Y j Γk1j
∂u1
∂2Y k ∂Y j k j
∂Γk1j
= − Γ − Y
∂u21 ∂u1 1j ∂u1
∂Γk1j
= Y l Γj1l Γk1j − Y j
∂u1

These give the local expression of Y k along the first segment (u1 , 0):

∂Γk1j
 
1
Y k (u1 , 0) = Y k (0, 0) − Y j Γk1j u1 + Y l Γj1l Γk1j − Y j u21 + O(u31 ).
(0,0) 2 ∂u1 (0,0)

Next, we consider the second segment: transporting Y (u1 , 0) along the u2 -curve. By a
similar Taylor expansion, we get:

∂Y k 1 ∂2Y k
Y k (u1 , u2 ) = Y k (u1 , 0) + u2 + u22 + O(u32 ).
∂u2 (u1 ,0) 2 ∂u22 (u1 ,0)

As before, we next rewrite the derivative terms using Christoffel’s symbols. The parallel
transport equation along the u2 -curve gives:

∂Y k
= −Y j Γk2j
∂u2
∂2Y k l j k j
∂Γk2j
= Y Γ Γ
2l 2j − Y .
∂u22 ∂u2

Plugging these in, we get:

∂Γk2j
 
k k j 1 l j k
Y (u1 , u2 ) = Y (u1 , 0) − Y Γk2j u2 + Y Γ2l Γ2j − Y j
u22 + O(u32 ).
(u1 ,0) 2 ∂u2 (u1 ,0)

Each term Y k (u1 , 0) above can then be expressed in terms of Y k (0, 0) and Γkij (0, 0)
using our previous calculations. It seems a bit tedious, but if we just keep terms up to
242 10. Curvatures of Riemannian Manifolds

second-order, then we can easily see that:

∂Γk1j
 
k k j 1 l j k
Y (u1 , u2 ) = Y (0, 0) − Y Γk1ju1 + Y Γ1l Γ1j − Y j
u2
(0,0) 2 ∂u1 (0,0) 1
 
− Y j (0, 0) − Y l (0, 0)Γj1l (0, 0)u1 Γk2j (u1 , 0) u2
!
1 l j k j
∂Γk2j 3

+ Y Γ2l Γ2j − Y u22 + O |(u1 , u2 )|
2 ∂u2 (u1 ,0)

∂Γk1j
 
k j k 1 l j k j
= Y (0, 0) − Y Γ1j u1 + Y Γ1l Γ1j − Y u2
(0,0) 2 ∂u1 (0,0) 1
∂Γk2j
 
j
− Y (0, 0) − Y (0, 0)Γ1l (0, 0)u1 Γk2j (0, 0) u2 − Y j (0, 0)
j l
u1 u2
∂u1 (0,0)
!
k
1 ∂Γ2j 3

+ Y l Γj2l Γk2j − Y j u22 + O |(u1 , u2 )| .
2 ∂u2 (0,0)

Next, we consider the parallel transport in another path, first along u2 , then along
u1 . Let Ye = Ye k ∂u∂ k be the local expression of the transported vector, then the local
expression of Ye k at (u1 , u2 ) can be obtained by switching 1 and 2 of that of Y k (u1 , u2 ):
∂Γk2j
 
1 el j k
Ye k (u1 , u2 ) = Ye k (0, 0) − Ye j Γk2j u2 + Y Γ2l Γ2j − Ye j
u2
(0,0) 2 ∂u2 (0,0) 2
∂Γk1j
 
j
e j e l
− Y (0, 0) − Y (0, 0)Γ2l (0, 0)u2 Γk1j (0, 0) u1 − Ye j (0, 0) u1 u2
∂u2 (0,0)
!
1 el j k ∂Γk1j 3

+ Y Γ1l Γ1j − Ye j u21 + O |(u1 , u2 )| .
2 ∂u1 (0,0)

Recall that Y (0, 0) = Ye (0, 0). By comparing Y (u1 , u2 ) and Ye (u1 , u2 ), we get:

Y k (u1 , u2 ) − Ye k (u1 , u2 )
∂Γk1l ∂Γk2l
 
3
= Y l Γj1l Γk2j − Γj2l Γk1j +


− u1 u2 + O |(u1 , u2 )|
∂u2 ∂u1 (0,0)

k 3

= R21l Yl u1 u2 + O |(u1 , u2 )| .
(0,0)

In other words, we have:

3

Y (u1 , u2 ) − Ye (u1 , u2 ) = Rm(∂2 , ∂1 )Y u1 u2 + O |(u1 , u2 )| .
(0,0)

Therefore, the vector Rm(∂2 , ∂1 )Y at p measures the difference between the parallel
transports of those two different paths.

10.1.3. Symmetric properties and Bianchi identities. We will explore more about
the geometric meaning of the Rm(4,0) -tensor in the next section. Meanwhile, let’s discuss
some nice algebraic properties of this tensor. The Riemann curvature (4, 0)-tensor satisfies
some nice symmetric properties. The first two indices ij, and the last two indices kl
of the components Rijkl , and is symmetric if one swap the whole ij with the whole kl.
Precisely, we have:
10.1. Riemann Curvature Tensor 243

Proposition 10.2 (Symmetric Properties of Rm(4,0) ). The local components of Rm(4,0)

satisfy the following properties:
• Rijkl = −Rjikl = −Rijlk , and
• Rijkl = Rklij .

Proof. The fact that Rijkl = −Rjikl follows immediately from the definition of Rm(3,1) .
It only remains to show Rijkl = Rklij , then Rijkl = −Rijlk would follow immediately.
We prove it by picking geodesic normal coordinates at a fixed point p, so that
gij = δij , ∂k gij = 0, and Γkij = 0 at p.
Then at p, we have:
Rijkl − Rklij
p p
= gpl Rijk − gpj Rkli
∂Γpjk
!
∂Γpik
= gpl − + Γqjk Γpiq − Γqik Γpjq
∂ui ∂uj
 p
∂Γpki

∂Γli q p q p
− gpj − + Γli Γkq − Γki Γlq
∂uk ∂ul
∂Γljk ∂Γlik ∂Γjli ∂Γjki
= − − + .
∂ui ∂uj ∂uk ∂ul
Note that in geodesic normal coordinates we only warrant Γkij = 0 at a point, which does
not imply its derivatives vanish at that point! Next we recall that
 
1 ∂gkp ∂gjp ∂gjk
Γljk = g lp + − ,
2 ∂uj ∂uk ∂up
and so at p we have:
∂Γljk 1 lp ∂ 2 gkp ∂ 2 gjp ∂ 2 gjk
 2
∂ 2 gjl ∂ 2 gjk
  
1 ∂ gkl
= g + − = + − .
∂ui 2 ∂ui ∂uj ∂ui ∂uk ∂ui ∂up 2 ∂ui ∂uj ∂ui ∂uk ∂ui ∂ul
By permutating the indices, one can find out similar expressions for
∂Γlik ∂Γjli ∂Γjki
, , and .
∂uj ∂uk ∂ul
Then by cancellations one can verify that Rijkl − Rklij = 0 at p. Since Rijkl − Rklij is
tensorial, it holds true under any local coordinate system and at every point. 

Another nice property of Rm(4,0) is the following pair of Bianchi identities, which
assert that the indices of the tensor exhibit some cyclic relations.

Proposition 10.3 (Bianchi Identities). The local components of Rm(4,0) satisfy the fol-
lowing: for any i, j, k, l, p, we have
R(ijk)l := Rijkl + Rjkil + Rkijl = 0.
∇(i Rjk)lp := ∇i Rjklp + ∇j Rkilp + ∇k Rijlp = 0.

Sketch of proof. Both can be proved using geodesic normal coordinates. Consider the
geodesic normal coordinates {ui } at a point p, then as in the proof of Proposition 10.2,
we have
∂Γljk ∂Γlik
Rijkl = − at p.
∂ui ∂uj
244 10. Curvatures of Riemannian Manifolds

By relabelling indices, we can write down the local expressions of Rjkil and Rkijl . By
summing them up, one can prove the first Bianchi identity by cancellations.
The second Bianchi identity can be proved in similar way. However, we have to be
careful not to apply Γkij (p) = 0 too early! Even though one has Γkij = 0 at a point p, it
does not warrant its derivatives equal 0 at p. The same for ∂k gij , which are all zero at p,
but it does not imply ∂l ∂k gij = 0 at p.
We sketch the proof and leave the detail for readers to fill in. Under geodesic normal
coordinates at p, we have
∇i Rjklp = ∂i Rjklp , at p.
Now consider
∂Γqjl
( !)
∂ ∂ ∂Γqkl m q m q
Rjklp = gqp − + Γkl Γjm − Γjl Γkm .
∂ui ∂ui ∂uj ∂uk
Use (7.15) again to write Γkij ’s in terms of derivatives of gij ’s. Be caution that to evaluate
∂
at p only after differentiation by ∂u i
. Get similar expressions for ∇j Rkilp and ∇k Rijlp at
p. One then should see all terms got cancelled when summing them up. 

Exercise 10.2. Prove the second Bianchi identity using geodesic normal coordinates.

All of the above-mentioned symmetric properties and Bianchi identities can be

written using invariant notations instead of local coordinates as follows:
• Rm(X, Y, Z, W ) = −Rm(Y, X, Z, W ) = −Rm(X, Y, W, Z)
• Rm(X, Y, Z, W ) = Rm(Z, W, X, Y )
• Rm(X, Y, Z, W ) + Rm(Y, Z, X, W ) + Rm(Z, X, Y, W ) = 0
• (∇X Rm)(Y, Z, W, U ) + (∇Y Rm)(Z, X, W, U ) + (∇Z Rm)(X, Y, W, U ) = 0
for any vector fields X, Y, Z, W, U, V .

10.1.4. Isometric invariance. The Riemann curvature tensors (both types) can
be shown to be invariant under isometries, meaning that if Φ : (M, g) → (M f, ge) is an
∗ ∗g
isometry (i.e. Φ ge = g), then we have Φ Rm = Rm. To prove this, we first show that the
Levi-Civita connection is also isometric invariant.

Proposition 10.4. Suppose Φ : (M, g) → (M

f, ge) is an isometry, then we have
Φ∗ (∇X Y ) = ∇
e Φ X (Φ∗ Y )
∗

∞
for any X, Y ∈ Γ (T M ). Here ∇ and ∇
e are the Levi-Civita connections for g and ge
respectively.

Idea of Proof. While it is possible to give a proof by direct computations using local
coordinates, there is a much smarter way of doing it. Here we outline the idea of proof
and leave the detail for readers to fill in. The desired result is equivalent to:
∇X Y = (Φ−1 )∗ ∇


e Φ X (Φ∗ Y ) .
∗

We define an operator D : Γ∞ (T M ) × Γ∞ (T M ) → Γ∞ (T M ) by:

D(X, Y ) = (Φ−1 )∗ ∇


e Φ X (Φ∗ Y ) .
∗

Then, one can verify that such D is a connection on M satisfying conditions (8.7) and
(8.8). This shows D must be the Levi-Civita connection of M by Proposition (8.14),
completing the proof. 
10.1. Riemann Curvature Tensor 245

Exercise 10.3. Complete the detail of the proof of Proposition 10.4.

Now given that the Levi-Civita connection is isometric invariant, it follows easily that
Rm is too.

Proposition 10.5. Suppose Φ : (M, g) → (M

f, ge) is an isometry, then we have

Φ∗ Rm
g = Rm
where both Riemann curvature tensors are of (4, 0)-type.

Proof. Let X, Y, Z, W be four tangent vectors on M , then we have:

(Φ∗ Rm)(X,
g Y, Z, W )
= Rm(Φ
g ∗ X, Φ∗ Y, Φ∗ Z, Φ∗ W )
 
= ge ∇Φ∗ X ∇Φ∗ Y (Φ∗ Z) − ∇Φ∗ Y ∇Φ∗ X (Φ∗ Z) − ∇[Φ∗ X,Φ∗ Y ] (Φ∗ Z), Φ∗ W
e e e e e
 
= ge ∇Φ∗ X Φ∗ (∇Y Z) − ∇Φ∗ Y Φ∗ (∇X Z) − ∇Φ∗ [X,Y ] (Φ∗ Z), Φ∗ W
e e e
 
= ge Φ∗ (∇X ∇Y Z − ∇Y ∇X Z − ∇[X,Y ] Z), Φ∗ W

= (Φ∗ ge) Rm(X, Y )Z, W = g Rm(X, Y )Z, W = Rm(X, Y, Z, W ).


Example 10.6. As derivatives (hence curvatures) are local properties, the above results
also hold if Φ is just an isometry locally between two open subsets of two Riemannian
manifolds. The flat metric δ on Rn certainly has Rm = 0. Now for the torus Tn := Rn /Zn ,
it has an induced metric g from the covering map π : Rn → Tn (see Proposition 8.11)
which is a local isometry since π ∗ g = δ. This metric also has a zero Riemann curvature
tensor by Proposition 10.5. As such, we call this torus with such an induced metric the
flat torus. 
Example 10.7. Consider the round sphere Sn , defined as x21 + · · · + x2n+1 = 1 in Rn+1 ,
with Riemannian metric given by the first fundamental form g = ι∗ δ. From standard Lie
theory, one can show the symmetry group SO(n + 1) acts transitively on Sn , meaning
that given any p, q ∈ Sn , there exists Φ : Rn+1 → Rn+1 such that Φ(p) = q.

g(Φ∗ X, Φ∗ Y ) = δ(ι∗ Φ∗ X, ι∗ Φ∗ Y ) = Φ(X) · Φ(Y ) = X · Y = g(X, Y ).

Hence Φ∗ g = g, and consequently Φ∗ Rm = Rm. Therefore, if one can find out the
Riemann curvature tensor at one point p ∈ Sn , then we can determine this tensor at all
other points on Sn . In the next section, we will focus on the input type Rm(X, Y, Y, X),
known as sectional curvatures, and will use it to determine the exact form of Rijkl of the
sphere. 
246 10. Curvatures of Riemannian Manifolds

10.2. Sectional Curvature

The Riemann curvature tensor is defined in a way so that in the two dimensional case
it (essentially) gives the Gauss curvature. The Rm(3,1) -tensor essentially measures the
defect of parallel transports along two different paths. Here we will exploit more about
the geometric meanings of the Rm(4,0) -tensor in higher dimensions.
In this section, we will explain two geometric quantities associated with the Riemann
curvature tensor, namely sectional curvatures and the curvature operator. The former is
inspired from the formula of Gauss curvature two dimensional case. The latter makes
good use of the symmetric properties of Rijkl and has deep connections with holonomy
groups and parallel transports which we discussed earlier.

10.2.1. Definition of Sectional Curvature. Recall that for a regular surface Σ2 ∈

3
R , its Gauss curvature equals to

R1221 Rm(∂1 , ∂2 , ∂2 , ∂1 )
K= 2 = 2 2 .
g11 g22 − g12 |∂1 | |∂2 | − g(∂1 , ∂2 )2

Inspired by this, we define the sectional curvatures as follows:

Definition 10.8 (Sectional Curvature). On a Riemannian manifold (M, g), consider

two linearly independent tangent vectors X and Y in Tp M . We define the sectional
curvature at p associated to {X, Y } by
Rm(X, Y, Y, X)
Kp (X, Y ) := 2 2
|X| |Y | − g(X, Y )2

Although it appears that Kp (X, Y ) depends on both X and Y , we can show it is in

fact independent of the choice of basis in span{X, Y }:

Proposition 10.9. Let Πp be a 2-dimensional subspace in Tp M , then given any bases

{X1 , Y1 } and {X2 , Y2 } for Πp we have:
Kp (X1 , Y1 ) = Kp (X2 , Y2 ).

Proof. Let a, b, c, d be real constants such that

X1 = aX2 + bY2
Y2 = cX2 + dY2

Then, we have

Rm(X1 , Y1 , Y1 , X1 )
= Rm(aX2 + bY2 , cX2 + dY2 , cX2 + dY2 , aX2 + bY2 )
= ad Rm(X2 , Y2 , cX2 + dY2 , aX2 + bY2 )
+ bc Rm(Y2 , X2 , cX2 + dY2 , aX2 + bY2 )


= ad ad Rm(X2 , Y2 , Y2 , X2 ) + bc Rm(X2 , Y2 , X2 , Y2 )


+ bc ad Rm(Y2 , X2 , Y2 , X2 ) + bc Rm(Y2 , X2 , X2 , Y2 )
= (ad − bc)2 Rm(X2 , Y2 , Y2 , X2 ).
10.2. Sectional Curvature 247

The last step follows from symmetric properties of Rm.

2 2
|X1 | |Y1 | − g(X1 , Y1 )2
2 2 2 2
= a2 |X2 | + 2abg(X2 , Y2 ) + b2 |Y2 | c2 |X2 | + 2cdg(X2 , Y2 ) + d2 |Y2 |

2 2
2
− ac |X2 | + (ad + bc)g(X2 , Y2 ) + bd |Y2 |
 
2 2
= (ad − bc)2 |X2 | |Y2 | − g(X2 , Y2 )2 .

The last step follows from direct computations.

It follows easily from the definition that Kp (X1 , Y1 ) = Kp (X2 , Y2 ). 

Remark 10.10. Therefore, one can also define the sectional curvature at p associated to
a plane (i.e. 2-dimensional subspace) Π in Tp M :
Kp (Π) := Kp (X, Y )
where {X, Y } is any basis for Π. 
Remark 10.11. Note that Kp (X, Y ) is not a (2, 0)-tensor! It is evident from the above
result that K(2X, Y ) = K(X, Y ) 6= 2K(X, Y ). 

Although the sectional curvature is essentially the Riemann curvature tensor re-
stricted on inputs of type (X, Y, Y, X), it is interesting that Rm(X, Y, Z, W ) itself can also
be expressed in terms of sectional curvatures:

Proposition 10.12. The Riemann curvature tensor Rm(4,0) is uniquely determined by its
sectional curvatures. Precisely, given any tangent vectors X, Y, Z, W ∈ Tp M , we have:
Rm(X, Y, Z, W )
= Rm(X + W, Y + Z, Y + Z, X + W ) − Rm(X + W, Y, Y, X + W )
− Rm(X + W, Z, Z, X + W ) − Rm(X, Y + Z, Y + Z, X)
− Rm(W, Y + Z, Y + Z, W ) + Rm(X, Z, Z, X) + Rm(W, Y, Y, W )
− Rm(Y + W, X + Z, X + Z, Y + W ) + Rm(Y + W, X, X, Y + W )
+ Rm(Y + W, Z, Z, Y + W ) + Rm(Y, X + Z, X + Z, Y )
+ Rm(W, X + Z, X + Z, W ) − Rm(Y, Z, Z, Y ) − Rm(W, X, X, W ).

Proof. Omitted. We leave it as an exercise for readers who need to wait for half an hour
for a morning minibus back to HKUST.
It is worthwhile to note that the result holds true when Rm is replaced by any
(4, 0)-tensor T satisfying all of the following:
• Tijkl = −Tjikl = −Tijlk = Tklij
• Tijkl + Tjkil + Tkijl = 0


10.2.2. Geometric Meaning of Sectional Curvature. We are going to explain

the geometric meaning of sectional curvatures. Let X, Y be two linearly independent
tangent vectors in Tp M of a Riemannian manifold (M, g). Consider the 2-dimensional
surface Σp (X, Y ) obtained by spreading out geodesics from p along all directions in the
plane span{X, Y }. The sectional curvature Kp (X, Y ) can then be shown to be the Gauss
curvature of the surface Σp (X, Y ) with the induced metric. Precisely, we have
248 10. Curvatures of Riemannian Manifolds

Proposition 10.13. Let (M, g) be a Riemannian manifold, p ∈ M , and {X, Y } be two

linearly independent tangents in Tp M . Consider the surface:
Σp (X, Y ) := {expp (uX + vY ) : u, v ∈ R and |uX + vY | < ε}
where inj(p) > ε > 0. Denote ḡ to be the induced Riemannian metric ι∗ g where ι :
Σp (X, Y ) → M is the inclusion map. Then, we have:
Kp (X, Y ) = Gauss curvature of Σp (X, Y ) at p with respect to the metric ḡ.

Proof. Denote ∇ and Rm the Levi-Civita connection and the Riemann curvature tensor
of (M, g), and denote ∇ and Rm to be those of (Σ, g). Here we denote Σ := Σp (X, Y )
for simplicity.
The Gauss curvature of (Σ, g) is given by
Rm(X, Y, Y, X)
.
g(X, X)g(Y, Y ) − g(X, Y )2
As we have g = g when restricted on Tp Σ, it suffices to show:
Rm(X, Y, Y, X) = Rm(X, Y, Y, X).

By Exercise 8.6, we know that for any vector fields X, Y on Σ, we have

∇X Y = (∇X Y )T
where T denotes the projection onto Tp Σ. Hence, one can decompose ∇X Y as:
∇X Y = ∇X Y + h(X, Y )

N
where h(X, Y ) = ∇X Y , the projection onto the normal space of Σ.
Consider the local parametrization G(u, v) = expp (uX + vY ) of Σ. As in the proof
of Proposition 9.14, we have
∂G ∂G
(p) = X, (p) = Y,
∂u ∂v
and also that Γ∗uu , Γ∗uv , and Γ∗vv (here ∗ means any index) all vanish at p, or equivalently,
we have:
∇u ∂u = ∇u ∂v = ∇v ∂v = 0 at p.
In particular, it implies h(∂u , ∂u ) = h(∂u , ∂v ) = h(∂v , ∂v ) = 0 at p.
Now consider the relation between the two Riemann curvature tensors:
Rm(∂u , ∂v )∂v = ∇u ∇v ∂v − ∇v ∇u ∂v



= ∇u ∇v ∂v + h(∂v , ∂v ) − ∇v ∇u ∂v + h(∂u , ∂v )


= ∇u ∇v ∂v + h(∂u , ∇v ∂v ) + ∇u h(∂v , ∂v )


− ∇v ∇u ∂v − h(∂v , ∇u ∂v ) − ∇v h(∂u , ∂v )
= Rm(∂u , ∂v )∂v + h(∂u , ∇v ∂v ) − h(∂v , ∇u ∂v )



+ ∇u h(∂v , ∂v ) − ∇v h(∂u , ∂v )
Recall that h(X, Y ) ⊥ T Σ for any X, Y ∈ Γ∞ (T Σ), so we have


Rm(∂u , ∂v , ∂v , ∂u ) = g Rm(∂u , ∂v )∂v , ∂v
 




= g Rm(∂u , ∂v )∂v , ∂u + g ∇u h(∂v , ∂v ) − ∇v h(∂u , ∂v ) , ∂u .
10.2. Sectional Curvature 249

We are only left to show the terms involving h vanish at p. Take an orthonormal frame1
{eα } of normal vectors to the surface Σ. Write h(∂u , ∂v ) = hα α
uv eα (and similarly for huu
α
and hvv ). Then, one can compute that:
   
g ∇u h(∂v , ∂v ) , ∂u = g ∇u hα = g hα




e
vv α , ∂ u vv ∇u eα , ∂u .

Here we have used the fact that eα ⊥ ∂u . Using this fact again, one can also show that


g ∇u eα , ∂u = −g(eα , ∇u ∂u ) = 0 at p.
Similarly, one can also show
 


g ∇v h(∂u , ∂v ) , ∂u = 0 at p.

It completes our proof that

Rm(∂u , ∂v , ∂v , ∂u ) = Rm(∂u , ∂v , ∂v , ∂u ) at p,
and the desired result follows from the fact that ∂u = X and ∂v = Y at p. 

As we have seen in the proof above, the key idea is to relate Rm and Rm. In the
above proof it suffices to consider inputs of type (X, Y, Y, X), yet it is not difficult to
generalize the calculation to relate Rm(X, Y, Z, W ) and Rm(X, Y, Z, W ) of any input
types. This in fact gives a generalized Gauss’s equation, which we leaves it as an exercise
for readers.
Exercise 10.4. Let (M, g) be a Riemannian manifold with Levi-Civita connection ∇,
and (Σ, g) be a submanifold of M with induced metric g and Levi-Civita connection
∇. Given any vector field X, Y ∈ T Σ, we denote
h(X, Y ) := (∇X Y )N = normal projection of ∇X Y onto the normal space N Σ.
Prove that for any X, Y, Z, W ∈ T Σ, we have
Rm(X, Y, Z, W )
= Rm(X, Y, Z, W ) + h(X, Z)h(Y, W ) − h(X, W )h(Y, Z).

10.2.3. Constant Sectional Curvature Metrics. A Riemannian metric is said to

have constant sectional curvature if Kp (Π) is independent of both p and the choice of
plane Π ⊂ Tp M . We will show that such metric has local components of Rm(4,0) given
by:
Rijkl = C(gil gjk − gik gjl )
where C is a real constant. To begin, we first introduce a special product for a pair of
2-tensors, commonly known as the Kulkarni-Nomizu’s product. Given two symmetric
2-tensors h and k, we define h ∧ k as a 4-tensor given by:
(h ∧ k)(X, Y, Z, W )
:= h(X, Z)k(Y, W ) + h(Y, W )k(X, Z) − h(X, W )k(Y, Z) − h(Y, Z)k(X, W ).
It is straight-forward to show that whenever h and k are both symmetric, then the
Kulkarni-Nomizu’s product h ∧ k satisfies symmetric properties and the first Bianchi
identity like the Rm(4,0) -tensor:
• (h ∧ k)ijkl = −(h ∧ k)jikl = −(h ∧ k)ijlk = (h ∧ k)klij ; and
• (h ∧ k)ijkl + (h ∧ k)jkil + (h ∧ k)kijl = 0.

1“Frame” typically means a set of locally defined vector fields which are linearly independent.
250 10. Curvatures of Riemannian Manifolds

Exercise 10.5. Prove the above symmetric properties and the first Bianchi identity
for h ∧ k.

Therefore, the (4, 0)-tensor h ∧ k is completely determined by its values of type

(h ∧ k)(X, Y, Y, X)
(see Proposition 10.12). Now we consider the product g ∧ g, which is given by
(g ∧ g)(X, Y, Z, W ) = 2g(X, Z)g(Y, W ) − 2g(X, W )g(Y, Z),
or locally, (g ∧ g)ijkl = 2gik gjl − 2gil gjk .
One can easily see that
(g ∧ g)(X, Y, Y, X) = 2 g(X, Y )2 − g(X, X)g(Y, Y ) .

Now if (M, g) has constant sectional curvature, then there exists a constant C such that
Kp (X, Y ) = C for any p ∈ M and any linearly independent vectors {X, Y } ⊂ Tp M . In
other words, we have
C
Rm(X, Y, Y, X) = C g(X, X)g(Y, Y ) − g(X, Y )2 = − (g ∧ g)(X, Y, Y, X).


2
Since Rm + C2 (g ∧ g) is a (4, 0)-tensor satisfying symmetric properties and the first
Bianchi’s identity, and it equals zero when acting on any (X, Y, Y, X), we can conclude
that Rm + C2 (g ∧ g) ≡ 0. To conclude, we have proved that if Kp (X, Y ) = C for any
p ∈ M and any linearly independent vectors {X, Y } ⊂ Tp M , then


Rm(X, Y, Z, W ) = C g(X, W )g(Y, Z) − g(X, Z)g(Y, W ) ,
Rijkl = C(gil gjk − gik gjl ).

Example 10.14. The most straight-forward example of metrics with constant sectional
curvature is the Euclidean space Rn with the flat metric δ. We have Rm ≡ 0. Consequently,
the n-torus Tn equipped with the quotient metric given by the covering π : Rn → Tn
also has 0 sectional curvature. 
Example 10.15. The round sphere Sn of radius r, given by x21 + · · · + x2n+1 = r2 , has
constant sectional curvature r12 . The key reason is that geodesics are great circles of
radius r. Given any linearly independent vectors X, Y ∈ Tp Sn , the sectional surface
Σp (X, Y ) considered in Proposition 10.13 is formed by spreading out geodesics from p.
As each geodesic has the same curvature 1r , the principal curvatures of Σp (X, Y ) at p are
{ 1r , 1r }, showing that Σp (X, Y ) has Gauss curvature r12 at p.
This shows Rijkl = C(gil gjk − gik gjl ) for some constant C. Next we try to find out
what this C is. Consider the sectional curvature associated to {∂i , ∂j }, then we have:
1 Rijji
= K(∂i , ∂j ) = .
r2 gii gjj − gij gji
By the fact that Rijji = C(gii gjj − gij gji ) for constant sectional curvature metrics, we
must have C = r12 , and we conclude that:
1
Rijkl = (gil gjk − gik gjl ).
r2

Example 10.16. The hyperbolic space Hn under the upper-half space model:
Hn := {(x1 , · · · , xn ) ∈ Rn : x1 > 0}
10.2. Sectional Curvature 251

can be easily shown to have constant sectional curvatures by direct computations. Recall
that its Riemannian metric is given by:
Pn
dxi ⊗ dxi
g = i=1 2 .
x1
1
Under the global coordinates {xi }, the metric components gij = δ
x21 ij
forms a diagonal
matrix, so g ij
= x21 δij .
It follows that ∂k gij = −2x−3
1 δk1 δij .
1
1
Γkij = g kl ∂i gjl + ∂j gil − ∂l gij = x21 ∂i gjk + ∂j gik − ∂k gij


2 2
1
= x21 − 2x−3 −3 −3


1 δi1 δjk − 2x1 δj1 δik + 2x1 δk1 δij
2
1

= δk1 δij − δi1 δjk − δj1 δik .
x1
By the local expression:
∂Γljk ∂Γlik
 
Rijkq = gql − + Γpjk Γlip − Γpik Γljp ,
∂ui ∂uj
it is straight-forward to verify that Rijkl = C(g ∧ g)ijkl for some constant C > 0. This is
equivalent to saying that the metric has negative constant sectional curvature. We leave
it as an exercise for readers to complete the computations. 

Exercise 10.6. Complete the computations of Rijkq for the hyperbolic space under
the upper-half plane model. Find out the C > 0 and the sectional curvature explicitly.
Also, consider the rescaled hyperbolic metric ge := αg, where α > 0 is a constant.
What is the sectional curvature of ge?

The above examples of metrics on Rn , Sn , and Hn , and their quotients (such as Tn ,

n
RP , and higher genus tori) in fact form a complete list of geodesically complete constant
sectional curvature metrics (certainly, up to isometry). This is one major goal of the next
chapter, which discusses second variations of arc-lengths, index form, etc. in order to
establish such a classification result.

Brendle-Schoen’s Differentiable Sphere Theorem

A typical type of questions that geometers and topologists would like to ask is given some
conditions on curvatures, then what can we say about the topology of the manifold? In 1951,
H.E. Rauch posted a question of whether a compact, simply-connected Riemannian manifold
whose sectional curvatures are all bounded in ( 14 , 1] must be topologically a sphere. In 1960,
M. Berger and W. Klingenberg gave an affirmative answer to the question. The result is sharp
in a sense that CPn with Fubini-Study metric has sectional curvature 14 along holomorphic
planes.
The results of Berger and Klingenberg are topological – they showed such a manifold
must be homeomorphic to Sn , but higher dimensional spheres can have many exotic differen-
tial structures (see Milnor’s exotic spheres). It was a long-standing conjecture of whether
the 14 -pinched condition would warrant the sphere must have the standard differential
structure. In 2007, Simon Brendle and Richard Schoen (both at Stanford at that time) gave
an affirmative answer to this conjecture using the Ricci flow.

10.2.4. Curvature Operator. Another interesting notion about the Rm(4,0) -tensor
is the curvature operator. Recall that the tensor satisfies symmetric properties Rijkl =
−Rjikl = Rklij . If we group together i with j, and k with l, then one can regard Rm(4,0)
as a symmetric operator on ∧2 T M . Precisely, we define
252 10. Curvatures of Riemannian Manifolds

R : ∧2 T M → ∧2 T M
R(X ∧ Y ) := −Rm X, Y, g pk ∂p , g lq ∂q ∂k ∧ ∂l

and extend tensorially to all of ∧2 T M . The minus sign is to make sure that positive
curvature operator (to be defined later) would imply positive sectional curvature.
Furthermore, given the metric g, we can define an induced metric, still denoted by g,
on ⊗2 T M and ∧2 T M by the following way:
g(X ⊗ Y, Z ⊗ W ) := g(X, Z)g(Y, W ).
As X ∧ Y := X ⊗ Y − Y ⊗ X, one can easily verify that
g(X ∧ Y, Z ∧ W ) = (g ∧ g)(X, Y, Z, W ).
Then, one can also verify easily that R is a self-adjoint operator with respect to this
induced metric g on ∧2 T M , meaning that



g R(X ∧ Y ), Z ∧ W = g X ∧ Y, R(Z ∧ W ) .


Indeed, we have g R(X ∧ Y ), Z ∧ W = −Rm(X, Y, Z, W ). By standard linear algebra,
such an operator would be diagonalizable and that all its eigenvalues are real.
Exercise 10.7. Verify all the above claims, including
• R is well-defined;
• g(X ∧ Y, Z ∧ W ) = (g ∧ g)(X, Y, Z, W );


• g R(X ∧ Y ), Z ∧ W = Rm(X, Y, Z, W ); and
• R is self-adjoint with respect the inner product g on ∧2 T M .

We say (M, g) has positive curvature operator if all eigenvalues of R are positive at
every point in M . The curvature operator is related to sectional curvatures in a sense
that the later are the “diagonals” of R. As a matrix is positive-definite implies all its
diagonal entries are positive (NOT vice versa), so positive curvature operator implies
positive sectional curvatures. The converse is not true: CPn with the Fubini-Study metric
can be shown to have positive sectional curvatures (in fact K(X, Y ) is either 1 or 4), but
does not have positive curvature operator.
It had been a long-standing conjecture (recently solved in 2006) that what topology
a compact Riemannian manifold (M, g) must have if it has positive curvature operator. In
dimension 3, positive curvature operator implies positive Ricci curvature (to be defined
in the next section). In 1982, Richard Hamilton introduced the Ricci flow to show that
such a 3-manifold must be diffeomorphic to S3 with round metric or its quotient. The
key idea (modulo intensive technical detail) is to show that such a metric would evolve
under the heat equation to the one which constant sectional curvature (compared with
heat diffusion distributes temperature evenly in the long run), which warrants that such
a manifold must be diffeomorphic to S3 or its quotient.
Several years after the celebrated 3-manifold paper, Hamilton used the Ricci flow
again to show that if a compact 4-manifold (M, g) has positive curvature operator, then
it must be diffeomorphic to S4 or its quotient. He conjectured that the result holds for
any dimension. It was until 2006 that Christoph Boehm and Burkhard Wilking resolved
this conjecture completely, again using the Ricci flow (combined with some Lie algebraic
techniques).
10.3. Ricci and Scalar Curvatures 253

10.3. Ricci and Scalar Curvatures

In this section we introduce the Ricci curvature and the scalar curvature. The former is
essentially the trace of the Riemann curvature tensor, and the later is the trace of the
former.

10.3.1. Ricci Curvature. We will define the Ricci curvature using both invariant
(global) notations and local coordinates. Let’s start with the invariant notations first
(which give clearer geometric meaning):

Definition 10.17 (Ricci Curvature). The Ricci curvature of a Riemannian manifold

(M, g) is a 2-tensor defined as:
n
X
Ric(X, Y ) := Rm(ei , X, Y, ei )
i=1
where {ei } is an orthonormal frame of the tangent space.

Exercise 10.8. Show that Ric(X, Y ) is independent of the choice of orthonormal

basis {ei } of the tangent space.

By the symmetry properties of Rm, it is clear that Ric is a symmetric tensor, i.e.
Ric(X, Y ) = Ric(Y, X).

Given a unit vector X ∈ Tp M , one can pick an orthonormal basis {ei } for Tp M such
that e1 = X, then we have:
n
X n
X
Ric(X, X) = Rm(ei , e1 , e1 , ei ) = K(ei , e1 ).
i=2 i=2

Therefore, the quantity Ric(X, X) is essentially the sum of sectional curvatures associated
to all orthogonal planes containing X.
We say that (M, g) has positive Ricci curvature at p ∈ M if Ric is positive-definite
at p, namely Ric(X, X) ≥ 0 for any X ∈ Tp M with equality holds if and only if X = 0.
If it holds true at every p ∈ M , we may simply say (M, g) has positive Ricci curvature.
Clearly, positive sectional curvature implies positive Ricci curvature. If one only has
Ric(X, X) ≥ 0 for any X ∈ Tp M without the equality conditions, we say (M, g) has
semi-positive (or non-negative) Ricci curvature at p.

Exercise 10.9. Suppose (M, g) has non-negative sectional curvature. Show that if
Ric ≡ 0, then Rm ≡ 0.

In terms of local coordinates {ui }, we denote

 
∂ ∂
Rij := Ric ,
∂ui ∂uj
so that Ric = Rij dui ⊗ duj . We next show that Rij has the following local expression:
X
k
(10.1) Rij = Rkij = g kl Rkijl .
k

One nice way of proving it is to use geodesic normal coordinates. Fix a point p ∈ M ,
and pick an orthonormal basis {ei } ∈ Tp M , then there exists a local coordinate system
∂
{ui } such that ∂u i
= ei at p, and gij = 0 at p. By Exercise 10.8, the definition of Ric is
independent of such an orthonormal basis.
254 10. Curvatures of Riemannian Manifolds

Now we consider:
X
Rij := Ric(∂i , ∂j ) = Rm(ek , ∂i , ∂j , ek )
k
X X
= Rm(∂k , ∂i , ∂j , ∂k ) = Rkijk
k k
= δlk Rkijl = g lk Rkijl .
P
The last line was to convert k Rkijk into a tensorial quantity. Now that both Rij and
g lk Rkijl are tensorial (the latter is so because the summation indices k and l appear
as top-bottom pairs), so the identity Rij = g lk Rkijl holds under any local coordinates
covering p. Again since p is arbitrary, we have proved (10.1) holds globally on M .
It is possible to give a less “magical” (but more transparent) proof of (10.1). Write
∂
ek = Aik ∂u i
where {ek } is an orthonormal basis of Tp M and {ui } is any local coordinate
system. By orthogonality, we have
δij = g(ei , ej ) = Aki gkl Alj .
In matrix notations, we have I = [A][g][A]T , where Aki is the (i, k)-th entry of [A], then
we get [g] = [A]−1 ([A]T )−1 and [g]−1 = [A]T [A], i.e.
X
g ij = Aik Ajk .
k

Now consider
X
Rij = Rm(ek , ∂i , ∂j , ek )
k
X
= Rm(Apk ∂p , ∂i , ∂j , Aqk ∂q )
k,p,q
X X
= Apk Aqk Rpijq = g pq Rpijq
k,p,q p,q

as desired.

10.3.2. Scalar Cuvature. The scalar curvature R is the trace of the Ricci tensor, in
a sense that
Xn
R := Ric(ei , ei )
i=1
where {ei } is an orthonormal frame of the tangent space. It can be shown that R
is independent of the choice of the orthonormal frame, and has the following local
expression
(10.2) R = g ij Rij

Exercise 10.10. Show that the definition of R does not depend on the choice of the
orthonormal frame {ei }ni=1 , and prove (10.2).

Using the second Bianchi identity, one can prove the following identity relating the
Ricci curvature and scalar curvature.

Proposition 10.18 (Contracted Bianchi Identity). On a Riemannian manifold (M, g),

we have the following identity:
1 1 ∂R
(10.3) divg Ric = dR, or in local coordinates: ∇i Rij = .
2 2 ∂uj
10.3. Ricci and Scalar Curvatures 255

Proof. The second Bianchi identity asserts that

∇i Rjklp + ∇j Rkilp + ∇k Rijlp = 0.
Multiplying both sides by g jp and keeping in mind that ∇g = 0, we have
∇i Rkl + ∇p Rkilp − ∇k Ril = 0.
Here we have used the fact that Rijlp = −Rjilp in the last term.
Next, multiply both sides by g kl and using again the fact that ∇g = 0, we get
∇i R − ∇p Rip − ∇l Ril = 0.
By rearrangement and change-of-indices, we get:
1 1 ∂R
∇i Rij = ∇i R =
2 2 ∂ui
as desired.


One immediate consequence of the contracted Bianchi identity is that the 2-tensor
Ric − R2 g is divergence-free:
 
R 1 1 1
∇i Rij − gij = ∇i Rij − gij ∇i R = ∂i R − ∂i R = 0.
2 2 2 2
R
The 2-tensor G := Ric − 2g is called the Einstein tensor, named after Albert Einstein’s
famous equation:
R
Rµν − gµν = Tµν
2
where the tensor Tµν described stress and energy. Here we use µ and ν for indices
instead of i and j as physicists use the former. We will learn in the next subsection
R p that a
metric satisfying the Einstein’s equation is a critical metric of the functional M R det[g],
known as the Einstein-Hilbert’s functional.
Inspired by the (vacuum) Einstein’s equation (where T = 0), mathematicians called
a Riemannian metric g satisfying the equation below an Einstein metric:
Rij = cgij , where c is a real constant.
A metric g with constant sectional curvature must be an Einstein metric. To see this, let’s
suppose
Rijkl = C(gil gjk − gik gjl ).
Then by tracing both sides, we get:
Rjk = g il Rijkl = C(g il gil gjk − g il gik gjl ) = C(ngjk − δlk gjl ) = C(n − 1)gjk .
When dim M ≥ 3, a connected Riemannian metric satisfying Rij = f gij , where f is
a smooth scalar function, must be an Einstein metric. To show this, we first observe that
f=R n:
R = g ij Rij = g ij f gij = nf.
Then, the contracted Bianchi identity shows
 
1 R 1 1
∂j R = ∇i Rij = ∇i gij = gij ∇i R = ∂j R.
2 n n n
When n > 2, it implies ∂j R = 0 for any j, and hence R is a constant.
The proof obviously fails in dimension 2. In contrast, it is always true that Rij = R2 gij
when n = 2, for any Riemannian metric g. It can be seen easily by combinatoric
inspections:
256 10. Curvatures of Riemannian Manifolds

R11 = g ij Ri11j = g 22 R2112 = g 22 R1221

R12 = g 21 R2121 = −g 21 R1221
R21 = g 12 R1212 = −g 12 R1221
R22 = g 11 R1221 .
Hence, the scalar curvature is given by
2R1221
R = g ij Rij = 2(g 11 g 22 − g 12 g 21 )R1221 = 2 det[g]−1 R1221 = .
det[g]
In dimension 2, we have
 11
g 12
  
g 1 g22 −g12
= .
g 21 g 22 det[g] −g 21 g11
R
It is then clear that Rij = 2 gij for any i, j = 1, 2.
One bonus result we have obtained from above is that R = 2K in dimension 2, where
K is the Gauss curvature. Recall from the proof of Gauss’s Theorema Egregium that we
have shown det[h] = R1221 .

Exercise 10.11. Let (M, g) be a connected Riemannian manifold with dim M ≥ 3.

Suppose for any fixed p ∈ M , the sectional curvature Kp (Π) is independent of any
2-plane Π ⊂ Tp M , but not assumed to be independent of p. Show that in fact Kp (Π)
is independent of p as well.

An Einstein metric must have constant scalar curvature. The proof is easy: if
Rij = cgij , then R = g ij Rij = nc. It was a fundamental question of what manifolds
admit a metric with constant scalar curvature. In dimension 2, it was the Uniformization
Theorem. Any simply connected Riemann surface (i.e. complex manifold with real
dimension 2) must be biholomorphic (i.e. conformal) to one of the standard models with
constant curvature: disc, plane, sphere (corresponding to negative, zero, and positive
curvatures respectively).
In higher dimensions (i.e. dim M ≥ 3), the positive curvature case was known as the
Yamabe problem, named after Hidehiko Yamabe. The problem was progressively resolved
by Neil Trudinger, Thierry Aubin, and finally by Richard Schoen in 1984, concluding that
on any compact manifold (M, g) of dimension n ≥ 3 with positive scalar curvature, there
exists a smooth function f on M such that the conformally rescaled metric ge = ef g has
constant positive scalar curvature.

10.3.3. Decomposition of Riemann Curvature Tensor. On a Riemannian mani-

fold with dimension n ≥ 3, the Riemann curvature (4, 0)-tensor admits an orthogonal
decomposition into scalar part, Ricci part, and Weyl part:
 
R 1 R
Rm = g ∧ g+ Ric − g ∧ g + W.
2n(n − 1) n−2 n
The Weyl tensor is defined as
 
R 1 R
W := Rm − g ∧ g− Ric − g ∧ g.
2n(n − 1) n−2 n
We will soon see that W = 0 when n = 3. When n ≥ 4, the Weyl tensor is not necessarily
0, but if it is so, then one can show the metric g is locally conformal to a flat metric.
Let’s first discuss what motivates such a definition of the Weyl tensor. Given any pair
of tensor fields of the same type, it is possible to find an inner product between using
10.3. Ricci and Scalar Curvatures 257

k ∂
the given Riemannian metric g. For instance, consider S = Sij dui ⊗ duj ⊗ ∂uk , and
T = Tijk dui ⊗ duj ⊗ ∂u∂ k , then we define:
g(S, T ) := g ip g jq gkr Sij
k r
Tpq .
It is a globally defined scalar function on M , and such a definition is independent of local
k k
P
coordinates. In the Euclidean case where gij = δij , then δ(S, T ) = i,j,k Sij Tij which
appears like the dot product.
We can then make sense of two tensor fields (of the same type) being orthogonal to
each other. For example, one can check:
   
R R R R
g Ric − g, g = g ik g jl Rij − gij gkl = g ij Rij − g ij gij = R − · n = 0,
n n n n
R
so Ric − ng and g are orthogonal.

Exercise 10.12. Show that g ∧ g and (Ric − R n g) ∧ g are orthogonal to each other
(with respect to the inner produced inherited from g).

Recall that for a constant sectional curvature metric, we have

C
Rijkl = C(gil gjk − gik gjl ) = (g ∧ g)ijkl ,
2
R
and this implies Rij = C(n − 1)gij , and hence it is necessary that C = n(n−1) . Therefore,
the difference
R
Rijkl − (g ∧ g)ijkl
2n(n − 1)
measures how close is the metric from having constant sectional curvature.
The tensor Ric − Rn g measures how close the metric is from being Einstein. Since an
Einstein metric may not have constant sectional curvature, we want to append a term
above such that it captures how much the metric deviate from being Einstein:
 
R R
Rm − (g ∧ g) − C Ric − g ∧ g
2n(n − 1) n
where C is a constant to be determined. Denote
 
R R
W := Rm − (g ∧ g) − C Ric − g ∧ g,
2n(n − 1) n
then one can find there exists a unique constant C such that W has zero trace, i.e.
g il Wijkl = 0,
1
and this constant C is in fact n−2 .
We are going to show that W ≡ 0 in dimension 3 using some elementary linear
algebra (dimension counting). We present the proof this way because it is easier for us
to see why dimension 3 is special.
By the symmetric properties of Rm, this (4, 0)-tensor can be regarded as a section
of the bundle ∧2 T ∗ M ⊗S ∧2 T ∗ M , where ⊗S denotes the symmetric tensor product. We
denote
E := ∧2 T ∗ M ⊗S ∧2 T ∗ M,

B := T ∈ E : Tijkl + Tjkil + Tkijl = 0 for any i, j, k, l. .
That is, B is a sub-bundle of E consisting of (4, 0)-tensors satisfying the Bianchi identity
in addition to the symmetric properties like Rm. We begin by finding the dimensions of
E and B:
258 10. Curvatures of Riemannian Manifolds

Lemma 10.19. On a Riemannian manifold (M, g) of dimension n ≥ 2, we have

n(n − 1)(n2 − n + 2)
dim E =
8
1 2 2
dim B = n (n − 1).
12

Proof. Each two-form is spanned by dui ∧ duj with i < j. Hence we have
n(n − 1)
dim ∧2 T ∗ M = C2n = .
2
Similarly, given a vector space V of dimension N , we have
N (N + 1)
dim(V ⊗S V ) = C2N + N = .
2
Hence, we have
n(n−1) n(n−1)

Cn 2 2 +1
dim E = C2 2 = ,
2
as desired.
The dimension of B can be found in a trickier way. We consider a map known Bianchi
symmetrization:

b : E → ⊗4 T ∗ M
1
b(T )ijkl := (Tijkl + Tjkil + Tkijl ).
3
Then, B = ker(b), hence dim B = dim E − dim Im (b) by elementary linear algebra.
One can show that the image Im (b) is in fact ∧4 T ∗ M . To show b(T ) ∈ ∧4 T ∗ M , we
observe that Tikjl + Tkjil = −Tjkil − Tkijl , so that

3b(T )jikl = Tjikl + Tikjl + Tkjil = −Tijkl − Tjkil − Tkijl = −3b(T )ijkl ,

and similarly for other switching of indices. Conversely, to show ∧4 T ∗ M ⊂ Im (b), we

observe by direct computations that:

1
b (dui ∧ duj ) ⊗S (duk ∧ dul ) = dui ∧ duj ∧ duk ∧ dul .
3

By rank-nullity theorem, we conclude that

n(n − 1)(n2 − n + 2) 1 2 2
dim B = dim E − dim ∧4 T ∗ M = − C4n = n (n − 1).
8 12


Next, recall that h ∧ k ∈ B whenever h and k are symmetric 2-tensors. We define

the following linear map:

L : S 2 (T ∗ M ) → B
h 7→ h ∧ g

Here S 2 (T ∗ M ) denotes the space of symmetric 2-tensors.

10.3. Ricci and Scalar Curvatures 259

This map can be shown to be injective whenever dim M ≥ 3. Whenever we have

h ∧ g = 0, one can check that in geodesic normal coordinates:
0 = g(h ∧ g, h ∧ g)
X
= (hil δjk + hjk δil − hik δjl − hjl δik )(hil δjk + hjk δil − hik δjl − hjl δik )
i,j,k,l
!2
X X
= 4(n − 2) h2ij +4 hii .
i,j i

Hence, we must have hij = 0 for any i, j. As h is tensorial, this holds true under any
local coordinate system.
In fact, the map L is adjoint to the trace map. Define a map:
Tr : B → S 2 (T ∗ M )
Tijkl (dui ∧ duj ) ⊗S (duk ⊗ dul ) 7→ g il Tijkl duj ⊗S duk
Then one can prove the following useful identity (whose proof is left as an exercise):
Exercise 10.13. Given any symmetric 2-tensor h, and a 4-tensor T ∈ B, we have:
g(T, L(h)) = g(4 Tr(T ), h).
As a result, the Weyl tensor is orthogonal to the image of L.

Now the key observation is that L is in fact bijective when dim M = 3. It is because
S 2 (T ∗ M ) and B both have the same dimension (equal to 6). Since W ∈ B, there exists a
symmetric 2-tensor h such that L(h) = W . Then, the fact that W = 0 follows easily from
Exercise 10.13:
g(W, W ) = g(W, L(h)) = 4g(Tr(W ), h) = 0
by the fact that Tr(W ) = 0. This clearly shows W = 0.
Consequently, in dimension 3, the Riemann curvature tensor can be expressed as:
 
R R
Rm = g ∧ g + Ric − g ∧ g.
12 3
An immediate corollary is that in dimension 3, the Ricci tensor Ric determines Rm. Also,
any Einstein metric (in dimension 3 only!) must have constant sectional curvature.
In dimension 4 or above, the Weyl tensor may not be 0. But being trace-free, it is
orthogonal to all 4-tensors of type h ∧ g (where h is symmetric 2-tensor). In particular, it
is orthogonal to both
 
R 1 R
(g ∧ g) and Ric − g ∧ g,
2n(n − 1) n−2 n
making the following an orthogonal decomposition:
 
R 1 R
Rm = g ∧ g+ Ric − g ∧ g + W.
2n(n − 1) n−2 n
The Weyl tensor can also be a conformal invariant: consider ge = ef g where f is a smooth
scalar function. It can be shown that the Weyl tensors are related by W (e g ) = ef W (g).
The proof is somewhat computational (either using geodesic normal coordinates or
Cartan’s notations). One needs to find out the how the Rm, Ric and R of ge are related to
those of g.
260 10. Curvatures of Riemannian Manifolds

10.4. Variation Formulae of Curvatures

The ultimate goal of this section is to show that the Euler-Lagrange’s equation of the
following Einstein-Hilbert’s functional:
Z q
LEH (g) := R det[gij ] du1 ∧ · · · ∧ dun
M
is in fact the vacuum Einstein equation
R
Ric −g = 0.
2
We will assume throughout that g is Riemannian, i.e. positive-definite, although almost
all parts of the proof remains valid with a Lorzentian g (having (−, +, +, +) signature).
We need to understand what metric g minimizes the functional LEH . Mathematically,
it means that if g(t) is a 1-parameter family of Riemannian metrics with g(0) = g and
∂
∂t g(t) = v(t) where v(t) is a 1-parameter family of symmetric 2-tensors, then LEH (g(t))
has achieves minimum at t = 0. It is then necessary that
d
LEH (g(t)) = 0
dt t=0
for any variation directions v(t).
In order to differentiate LEH , we need to know the variation formulae for R and
det[gij ]. We have already derived using geodesic normal coordinates the variation
∂
formula in Section 9.3: given that ∂t g(t) = v(t), then we have
∂ k 1
Γ = g kl (∇i vjl + ∇j vil − ∇l vij ) .
∂t ij 2
We will use this result to derive the variation formulae for curvatures.

10.4.1. Variation Formula for Ricci Tensor. Let’s first start with the Ricci tensor.
∂ ∂
Given that ∂t gij (t) = vij (t). We want to find a nice formula for ∂t Rij where Rij :=
Ricg(t) (∂i , ∂j ). Again we use geodesic normal coordinates at a fixed point p. Recall that
k
Rij = Rkij = ∂k Γkij − ∂i Γkkj + quadratic terms of Γkij ’s.
Although Γkij (p) = 0 does not warrant its derivative is 0 at p, we still have
∂
(quadratic terms of Γkij ’s) = 0
∂t
at p by the product rule. Hence, we may focus on the the first two terms when computing
∂
∂t Rij . We consider:

∂ ∂ ∂ k
∂k Γkij = Γ
∂t ∂uk ∂t ij
 
∂ 1 kl
= g (∇i vjl + ∇j vil − ∇l vij )
∂uk 2
1 ∂
= g kl (∇i vjl + ∇j vil − ∇l vij ) .
2 ∂uk
Here we have use the fact that ∂k g kl = 0 at p for geodesic normal coordinates. Next note
that
∇k ∇i vjl = ∂k (∇i vjl ) + terms with Christoffel symbols,
so under geodesic normal coordinates we have
∇k ∇i vjl = ∂k (∇i vjl ).
10.4. Variation Formulae of Curvatures 261

This shows at p under geodesic normal coordinates we have:

∂ ∂ k 1
Γ = g kl (∇k ∇i vjl + ∇k ∇j vil − ∇k ∇l vij ).
∂t ∂uk ij 2
Similarly, we can also get at p:
∂ ∂ k 1
Γkj = g kl (∇i ∇k vjl + ∇i ∇j vkl − ∇i ∇l vkj ).
∂t ∂ui 2
Combining these, we get:
∂ 1
Rij = g kl (∇k ∇i vjl + ∇k ∇j vil − ∇k ∇l vij − ∇i ∇k vjl − ∇i ∇j vkl + ∇i ∇l vkj ).
∂t 2
Although the above holds true for one particular coordinate system and at one point only,
yet since both sides are tensorial, it holds for all local coordinate systems and at every
point.
We can further simplify the expression a bit. For instance, we can write g kl ∇k = ∇l ,
and g kl ∇k ∇l = ∆ (the tensor Laplacian). Noting that ∇g = 0, we can also write
g kl ∇i ∇j vkl = ∇i ∇j (g kl vkl ). It is common to denote the trace of v (with respect to g)
by Trg v := g kl vkl . Note that Trg v = g(v, g) where g(·, ·) is the induced inner product on
2-tensors. After all these make-over, we can simplify the variation formula for Ric as:
∂ 1 1 1 1
Rij = − ∆vij − ∇i ∇j (Trg v) + ∇l ∇i vjl + ∇l ∇j vil .
∂t 2 2 2 2
10.4.2. Variation Formula for Scalar Curvature. Next we derive the variation
formula for the scalar curvature. Recall that R = g ij Rij , and we already have the
variation formula for Rij . We need the formula for g ij . Using the fact that
gij g jk = δik ,
by taking the time derivative on both sides we get
   
∂ jk ∂ jk
gij g + gij g = 0.
∂t ∂t
By rearrangement, we get
∂ jk ∂
g = −g jp g kq gpq .
∂t ∂t
∂
Given that ∂t gij = vij , we conclude that
∂ ij
g = −g ip g jq vpq
∂t
Now we are ready to derive the variation formula for R:
∂ ∂
R = (g ij Rij )
∂t ∂t  
1 1 1 1
= −g ip g jq vpq Rij + g ij − ∆vij − ∇i ∇j (Trg v) + ∇l ∇i vjl + ∇l ∇j vil
2 2 2 2
1 1 1 1
= −g(v, Ric) − ∆(g ij vij ) − ∆(Trg v) + ∇l ∇j vjl + ∇l ∇i gil
2 2 2 2
l i
= −∆(Trg v) − g(v, Ric) + ∇ ∇ vil .

In particular, if the metric g(t) deforms along the direction of −2Ric(g(t)), i.e.
∂
gij = −2Rij ,
∂t
then the scalar curvature evolves by
∂ 2
R = −∆(−2R) − g(−2Ric, Ric) − 2∇i ∇j Rij = ∆R + 2 |Ric|
∂t
262 10. Curvatures of Riemannian Manifolds

where we have used the contracted Bianchi identity ∇j Rij = 12 ∂i R. Using parabolic
maximum principle, one can then show on a compact manifold if R ≥ C at t = 0, then
∂
it remains so for all t > 0 under the evolution ∂t gij = −2Rij . This deformation of the
metric g(t) is called the Ricci flow, which is the major tool for resolving the Poincaré
conjecture.

10.4.3. Variation Formula for Determinants. Next we want to compute the varia-
tion formula for det[gij ]. The following general result is very useful:

Lemma 10.20. Let A(t) be a time-dependent invertible n × n matrix with entries denoted
by Aij , then we have:
 
∂ −1 ∂A
(10.4) det A(t) = Tr A det A(t),
∂t ∂t
where Aij is the (i, j)-th entry of A−1 .

Proof. It is best be done using differential forms. Let {ei }ni=1 be the standard basis of
Rn . Then we have:

Ae1 ∧ · · · ∧ Aen = det(A) e1 ∧ · · · ∧ en .

We then differentiate both sides by t using the product rule:

n
X ∂A ∂
Ae1 ∧ · · · ∧ Aei−1 ∧ ei ∧ Aei+1 ∧ · · · ∧ Aen = det(A) e1 ∧ · · · ∧ en .
i=1
∂t ∂t

Next we prove the following general result then the proof is completed. For any invertible
n × n matrix A, and any n × n matrix B, we have:

n
X
Ae1 ∧ · · · ∧ Aei−1 ∧ Bei ∧ Aei+1 ∧ · · · ∧ Aen = Tr(A−1 B) det(A) e1 ∧ · · · ∧ en .
i=1

To show this, we first prove the above holds for diagonalizable matrices A, which is
dense in the set of invertible matrices. Let P = [Pij ] be an invertible matrix such that
A = P DP −1 where D = diag(λ1 , · · · , λn ). Then, we have

n
X
Ae1 ∧ · · · ∧ Aei−1 ∧ Bei ∧ Aei+1 ∧ · · · ∧ Aen
i=1
n
X
= P DP −1 e1 ∧ · · · ∧ P (P −1 BP )P −1 ei ∧ · · · ∧ P DP −1 en
i=1
n
X
= det(P ) DP −1 e1 ∧ · · · ∧ (P −1 BP )P −1 ei ∧ · · · ∧ DP −1 en
i=1
n
X
= det(P ) det(D) P −1 e1 ∧ · · · ∧ D−1 (P −1 BP )P −1 ei ∧ · · · ∧ P −1 en
i=1
10.4. Variation Formulae of Curvatures 263

Write fi := P −1 ei , then we have

n
X
P −1 e1 ∧ · · · ∧ D−1 (P −1 BP )P −1 ei ∧ · · · ∧ P −1 en
i=1
n
X
= f1 ∧ · · · ∧ D−1 P −1 BP fi ∧ · · · ∧ fn
i=1
Xn
= f1 ∧ · · · ∧ (D−1 P −1 BP )ji fj ∧ · · · ∧ fn
i,j=1
n
X
= f1 ∧ · · · ∧ (D−1 P −1 BP )ii fi ∧ · · · ∧ fn
i=1
= Tr(D−1 P −1 BP ) f1 ∧ · · · ∧ fn
= Tr(D−1 P −1 BP ) det(P −1 ) e1 ∧ · · · ∧ en .
Recall that A = P DP −1 , so we have D−1 P −1 BP = P −1 A−1 BP , and so
Tr(D−1 P −1 BP ) = Tr(A−1 B).
Combining with the results above, we get
n
X
Ae1 ∧ · · · ∧ Aei−1 ∧ Bei ∧ Aei+1 ∧ · · · ∧ Aen = det(D)Tr(A−1 B) e1 ∧ · · · ∧ en
i=1

as desired. Note that det(D) = det(A).



∂
As a corollary, given that g(t) satisfies ∂t gij = vij , then we have
∂
det[gij ] = Tr([g]−1 [v]) det[g] = g ij vij det[g] = (Trg v) det[g].
∂t
One immediate consequence of the above variation formula is that the mean cur-
vature H of a Euclidean hypersurface Σ (with boundary C) must be 0 if it minimizes
the area among all smooth hypersurface with the boundary C. To see this, we let
Σ0 be such a hypersurface, and Σt is a 1-parameter family of hypersurfaces such that
Area(Σ0 ) ≤ Area(Σt ) for any t.
Denote Ft (u1 , · · · , un ) to be the local parametrization of Σt , and gij (t) be the first
fundamental form. Suppose ∂F ∂t = f ν where ν := νt is the unit normal for Σt , then one
can check that
     
∂gij ∂ ∂F ∂F ∂ν ∂F ∂F ∂ν
= , = f , + ,f .
∂t ∂t ∂ui ∂uj ∂ui ∂uj ∂ui ∂uj
∂
Here we have used the fact that ν and ∂ui are orthogonal. Recall that
∂ν ∂F
= −hki ,
∂ui ∂uk
we conclude that
∂gij
= −2f hij ,
∂t
and hence
∂p 1 p
det[g] = p (−2f g ij hij ) det[g] = −f H det[g].
∂t 2 det[g]
264 10. Curvatures of Riemannian Manifolds

Therefore, the variation formula for the area is given by:

Z Z
d ∂p p
Area(Σt ) = det[g] du1 ∧ · · · ∧ dun = − f H det[g] du1 ∧ · · · ∧ dun .
dt Σt ∂t Σt
Consequently, if we need Σ0 to minimize the area, then we must have
Z p
f H det[g] du1 ∧ · · · ∧ dun = 0
Σ0
for any normal variation f ν. It is then necessary that H = 0 on Σ0 .

10.4.4. Deriving the vacuum Einstein equation. Finally (after a short detour to
minimal surfaces), we get back on p
deriving the Einstein equation. The essential task is to
∂
derive the variation formula for R det[g]. Suppose ∂t gij = vij , then
∂ p p 1 p
R det[g] = −∆Trg v − g(v, Ric) + ∇i ∇j vij


det[g] + R · Trg v · det[g]
∂t   2
R p
= −∆Trg v − g(v, Ric) + ∇i ∇j vij + g(v, g) det[g]
2
 
R
p
= −∆Trg v − g v, Ric − g + ∇i ∇j vij det[g].
2
Consequently, we have
Z  
d R
p
LEH (g(t)) = −∆Trg v − g v, Ric − g + ∇i ∇j vij det[g] du1 ∧ · · · ∧ dun .
dt M 2
Assuming M has no boundary, any divergence terms such as ∇i Tij or ∇i T ij , etc.
must have integral 0 by the Stokes’ Theorem. This is can be shown by the following nice
observation:
Exercise 10.14. Suppose X is a vector field on a Riemannian manifold (M, g)
without boundary. Denote the volume element by
q
dµg := det[gij ] du1 ∧ · · · ∧ dun .
Show that
d(iX dµg ) = ∇i X i dµg .
Hence, by Stokes’ Theorem, we get
Z Z
i
∇i X dµg = d(something) = 0.
M M

One can then write ∇i ∇j vij = ∇k (g ik ∇j vij ) and letting

∂ ∂
X = g ik ∇j vij =: X k ,
∂uk ∂uk
then ∇i ∇j vij = ∇i X i , and so
Z
∇i ∇j vij dµg = 0.
M
Similarly, for any scalar function f , we have
∆f = g ij ∇i ∇j f = ∇i (g ij ∇j f ) = ∇i ∇i f.
By letting X = ∇f = ∇i f ∂i , one can also show
Z
∆f dµg = 0
M
provided that M has no boundary.
10.4. Variation Formulae of Curvatures 265

Therefore, the variation formula for the LEH -functional can be simplified as:
Z  
d R
LEH (g(t)) = − g v, Ric − g dµg .
dt M 2
If we need g(0) to minimize LEH under any variations v, then it is necessary that
R
Ric − g ≡ 0
2
which is exactly the vacuum Einstein’s equation.
By taking the trace with respect to g on both sides, we get:
R
R − n = 0.
2
If n > 2, we then have R = 0, and so Ric ≡ 0. Therefore, the solution to the vacuum
Einstein’s equation is necessarily Ricci-flat (but may not be Riemann-flat).
Exercise 10.15. Consider the following modified Einstein-Hilbert’s functional:
Z


LeEH (g) := R − 2Λ + ϕ(g) dµg
M
where Λ is a real constant, and ϕ(g) is a scalar function depending on the metric g.
Show that is g minimizes this functional among all variations of g, then it is
necessary that
R
Ric − g + Λg = T
2
where T is a symmetric 2-tensor depending on ϕ and its derivatives.
 Chapter 11

Curvatures and Topology

「隻手換乾坤，拓撲知曲率。」

丘成桐

A typical kind of questions that geometers and topologists would like to address is
that given some curvature conditions (note that curvatures are local properties), what
can one say about the global properties of the manifold (such as diameters, compactness,
topological type, etc.)?
There are two fundamental analytical tools for addressing these kind of questions,
namely the existence of Jacobi fields and the second variations of arc-lengths. The
Riemann curvature tensors play an important role in these two tools. They are used to
transfer local information about curvatures to some more global information (such as
diameter of a manifold).
One neat result is the following theorem due to Bonnet and Myers: which says that
if (M n , g) is a complete Riemannian manifold with Ric ≥ k(n − 1)g on M where k > 0 is
a positive constant, then one has diam(M, g(t)) ≤ √πk , and consequently M is compact
and has finite fundamental group π1 (M ). We will prove the theorem in Section 11.2.
The equality case of Bonnet-Myers’ Theorem was addressed by S.-Y. Cheng (Professor
Emeritus of HKUST Math) in 1975, who proved the equality holds if and only if (M n , g)
is a isometric to the round sphere with constant sectional curvature k.
In laymen terms, a Jacobi field J is a vector field whose integral curves are all
geodesics. If there are infinitely many geodesics connecting p and q, then there would
exist a Jacobi field J such that J(p) = J(q) = 0. An intuitive example is the round
2-sphere with p and q being the north and south poles. There is a family of great
∂ ∂
semi-circles connecting them and the Jacobi field is ∂ϕ (or ∂θ in PHYS convention).
Another fundamental result relating Jacobi fields and geodesics is that the existence
question of non-trivial Jacobi field J with J(p) = J(q) = 0 is (roughly) equivalent to
whether (expp )∗ is singular at the vector tX corresponding to q. In this case, if one
connects p and q by a unit-speed geodesic γ(t) (say γ(0) = p and γ(L) = q), then one can
use the above-mentioned equivalence to show that {γ(t)}t>0 is no longer a minimizing
geodesic when t > L. This can give an upper bound L on the distance d(p, q), and by
repeating the argument on arbitrary p and q, we can estimate the diameter (maximal
distance) of the manifold (M, g).

267
268 11. Curvatures and Topology

11.1. Jacobi Fields

11.1.1. Definition. As mentioned in the introduction, a Jacobi field is the variation
field of a family of geodesics. Let γs (t) : (−ε, ε) × [0, T ] → M be such a family, with s as
the parameter of the family, and t as the arc-length parameter of each curve. Denote
∂ ∂
V := γs (t) and T := γs (t).
∂s ∂t
We want to derive an equation for the variation vector field V . First, observe that since
each γs is a geodesic, the tangent vector field T satisfies

∇T T = 0 for any (s, t).

Consider the Riemann curvature tensor:

Rm(V, T )T = ∇V ∇T T − ∇T ∇V T − ∇[V,T ] T.
∂
We have ∇V ∇T T = ∇ ∂γ ∇T T = 0 since ∇T T = 0 holds for any s. Also, by V = γ∗ ∂s
∂s
∂
and T = γ∗ ∂t , we can see that [V, T ] = 0. The only survival term is the second one, and
further observe that ∇V T − ∇T V = [V, T ] = 0, so we get

Rm(V, T )T = −∇T ∇T V.

This is so-called the Jacobi field equation.

Definition 11.1 (Jacobi Fields). Let (M n , g) be a complete Riemannian manifold, and

γ : [0, T ] → M be an arc-length parametrized geodesic. Then, a vector field V defined
on γ is said to be a Jacobi field along γ if the following holds:
(11.1) ∇γ̇ ∇γ̇ V + Rm(V, γ̇)γ̇ = 0.

Note that (11.1) is a second-order ODE. Once we prescribed V (0) and ∇γ̇ V at t = 0,
then there exists a unique Jacobi field along γ with these initial data. The solution
space is a vector space by the linearity of the equation (11.1). The solution space is
parametrized by V (0) and ∇γ̇ V t=0 , hence is 2n-dimensional.

Example 11.2. Two easy examples of Jacobi fields are V = γ̇, and W = tγ̇. The first
one is obvious by the geodesic equation ∇γ̇ γ̇ = 0, and Rm(γ̇, γ̇)γ̇ = 0 by the alternating
property of Rm. For W = tγ̇, we also have Rm(W, γ̇)γ̇ = 0 by the same reason. We can
then show
∇γ̇ ∇γ̇ (tγ̇) = ∇γ̇ (γ̇ + t∇γ̇ γ̇ ) = ∇γ̇ γ̇ = 0.
| {z }
=0

The Jacobi fields γ̇ and tγ̇ are merely tangent vector fields so that their integral
curves are just reparametrization of the geodesic. The shape of the geodesic along these
variations is unchanged, and so they are not interesting examples. However, they are
essentially the “only” tangential Jacobi fields, since we can show any Jacobi field V along
γ can be decomposed into:
V = V ⊥ + aγ̇ + btγ̇
for some constants a and b, V ⊥ ⊥ γ̇, and still we have V ⊥ being a Jacobi field. The
argument is as follows:
11.1. Jacobi Fields 269

Consider the inner product g V (t), γ̇(t) , we will show that it is a linear polynomial
of t using the fact that ∇γ̇ γ̇ = 0:
d2

2
g V (t), γ̇(t)
dt

= g ∇γ̇ ∇γ̇ V, γ̇
= −g Rm(V, γ̇)γ̇, γ̇)
= −Rm(V, γ̇, γ̇, γ̇) = 0.
This shows g(V (t), γ̇(t)) = a + bt for some constants a and b. Then, it is easy to show
g V − aγ̇ − btγ̇, γ̇) = 0
and hence V − aγ̇ − btγ̇ ⊥ γ̇. Here we assume for simplicity that γ is arc-length
parametrized. By linearity of (11.1), V ⊥ is still a Jacobi field along γ, and it is normal to
the curve γ.
Consequently, the space of Jacobi fields has an orthogonal decomposition into
tangential and normal subspaces. The tangential subspace has dimension 2 spanned by γ̇
and tγ̇ (note that they are linearly independent as functions of t, even thought they are
parallel vectors pointwise), and the normal subspace has dimension 2n − 2.

Exercise 11.1. Suppose a Jacobi field V along a geodesic γ is normal to the curve
at two distinct points. What can you say about V ?

11.1.2. Jacobi Fields on Spaces with Constant Curvatures. Since the Jacobi
field equation (11.1) involves the curvature term, it is expected that one can find some
explicit solutions if the Riemann curvature term is nice. Consider a complete Riemannian
manifold (M, g) with constant sectional curvature C, i.e.
2 2
Rm(X, Y, Y, X) = C |X| |Y | − g(X, Y )2

for any X, Y ∈ Tp M .
Now given an arc-parametrized geodesic γ(t) : [0, T ] → M , and fix a unit vector
E0 ∈ Tγ(0) M such that E0 ⊥ γ̇(0), and extend this vector by parallel transport along γ,
i.e.
∇γ̇ E(t) = 0 and E(0) = E0 .
We want to find a Jacobi field along γ that is the form of V (t) = u(t)E(t) where u(t) is
a scalar function. It turns out that u(t) will satisfy a hand-solvable ODE if (M, g) has
constant sectional curvature.
Consider the Jacobi equation:
0 = ∇γ̇ ∇γ̇ (uE) + Rm(uE, γ̇)γ̇


= ∇γ̇ u̇E + uRm(E, γ̇)γ̇
= üE + uRm(E, γ̇)γ̇.
Here we have used the fact that ∇γ̇ E = 0. Taking inner product with E on both sides,
we get


0 = ü + ug Rm(E, γ̇)γ̇, E .
2
Here we have used the fact that |E(t)| = 1 by the property of parallel transport. Note
also that g(E(t), γ̇(t)) = 0 since it is so at t = 0, we have Rm(E, γ̇, γ̇, E) = C and we get
that
ü + Cu = 0.
270 11. Curvatures and Topology

It is a hand-solvable ODE, and the general solution is given by:

 a √ b
√
 √C sin(t C) + √C cos(t C)
 if C > 0
u(t) = at + b if C = 0
 √a
 √ b
√
−C
sinh(t −C) + −C cosh(t −C) if C < 0
√

where a and b are any constants.

Now each E0 ∈ Tγ(0) M with E0 ⊥ γ̇(0) gives a 2-dimensional subspace of Jacobi
fields. On the other hand, there are (n − 1) dimensions of vectors at γ(0) that are
orthogonal to γ̇(0). The above solutions of Jacobi fields form a (2n − 2) dimensional
subspace, and therefore these form all possible Jacobi fields normal to γ̇.

11.1.3. Comparison Theorem of Jacobi Fields. In the method of solving for

Jacobi fields for constant sectional curvature metrics, the factor u(t) is equal to |V (t)|.
Hence, |V (t)| satisfies the second-order ODE:
d2
|V (t)| + C |V (t)| = 0.
dt2
Generally speaking, on an arbitrary complete Riemannian manifold (M, g), it is possible
to derive a differential inequality of a similar form.
Suppose the sectional curvature of (M, g) is bounded above by a constant C (which
can be positive, zero, or negative). Given a unit speed geodesic γ(t) and a normal Jacobi
field V (t) on γ(t), we can compute that
d2 2 d

2
|V | = 2 g ∇γ̇ V, V
dt dt


= 2g ∇γ̇ ∇γ̇ V, V + 2g ∇γ̇ V, ∇γ̇ V

2
= −2g Rm(V, γ̇)γ̇, V + 2 |∇γ̇ V |
2 2
≥ −2C |V | + 2 |∇γ̇ V | .
Hence, we have
2
d2

d 2 2
2 |V | 2 |V | + 2 |V | ≥ −2C |V | + 2 |∇γ̇ V | .
dt dt
| {z }
d2
dt2
|V |2

Note that  2  2
2 d 2 V
|∇γ̇ V | − |V | = |∇γ̇ V | − g ∇γ̇ V, ≥0
dt |V |
by Cauchy-Schwarz’s inequality. We can conclude that as long as |V | > 0, we have
d2
|V | ≥ −C |V | .
dt2
Therefore, we can then derive comparison inequality between Jacobi fields of a
constant sectional curvature metric and a general metric with sectional curvature bounded
from above.
The key idea is that for any non-negative function f (t) that satisfies the inequalities
f 00 (t) + Cf (t) ≥ 0 for t > 0 , f (0) = 0, and f 0 (0) > 0,
then one can claim that
 f 0 (0) √ π
 √C sin(t C) for any t ∈ [0, √C ]
 if C > 0
f (t) ≥ f 0 (0)t for any t ≥ 0 if C = 0 .
 f 0 (0) √
sinh(t −C) for any t ≥ 0 if C < 0
√
−C
11.1. Jacobi Fields 271

To prove this, let u(t) be the solution to ü + Cu = 0 with u(0) = 0 and u0 (0) = f 0 (0), i.e.
 f 0 (0) √
 √C sin(t C)
 if C > 0
u(t) = f 0 (0)t if C = 0 .
 f 0 (0) √
sinh(t −C) if C < 0
√
−C
Sturm-Liouville’s comparison theorem of ODEs then asserts that f (t) ≥ u(t) as long as
u(t) > 0. Now let f (t) = |V (t)| where V (t) is a Jacobi field along γ such that V (0) = 0,
if one can argue that f 0 (0) > 0, then we conclude that
 f 0 (0) √ π
 √C sin(t C) for any t ∈ [0, √C ]
 if C > 0
(11.2) 0
|V (t)| ≥ f (0)t for any t ≥ 0 if C = 0 .
 f 0 (0) √
sinh(t −C) for any t ≥ 0 if C < 0
√
−C
As a corollary, if the sectional curvature is non-positive, any non-trivial Jacobi field V (t)
along γ(t) with V (0) = 0 would not vanish again for any t > 0.
We are left to show f 0 (0) > 0. Since V (0) = 0, we cannot compute f 0 (0) directly.
Consider the limit quotient:
|V (t)| − |V (0)| V (t)
f 0 (0) = lim+ = lim+ .
t→0 t t→0 t
We leave it as an exercise

for readers to show that for t > 0 sufficiently small, we in fact
have V (t) = t ∇γ̇ V γ(0) , then it completes the proof.

Exercise 11.2. Suppose V is a Jacobi field along an arc-length parametrized curve

γ0 (t) with V (0) = 0. Denote exp := expγ(0) and W0 := ∇γ̇ V γ(0) . Show that the
family of curves

γs (t) := exp t(γ̇0 (0) + sW0 )
is a geodesic family that gives the prescribed Jacobi field V (t). Furthermore, show
∂
that, under geodesic normal coordinates (u1 , · · · , un ) at p with ∂u 1
= γ̇0 (0), we
have
∂
V (t) = tW i
∂ui γ0 (t)
∂
for t > 0 sufficiently small, where W0 = W i ∂u i
at Tγ(0) M .

11.1.4. Conjugate Points and Jacobi Fields. One fundamental theorem about
Jacobi fields and the exponential maps is that the existence of Jacobi field V (t) along a
curve γ(t) : [0, T ] → M with V (0) = V (T ) = 0, is equivalent to (expγ(0) )∗ being singular
at the point corresponding to γ(T ). In this subsection, let’s make this relation more
precise. We first introduce:

Definition 11.3 (Conjugate Points). Let γ be a geodesic connecting points p and q

on a Riemannian manifold (M, g). We say q is conjugate to p along γ if there exists a
non-zero Jacobi field V along γ such that V (p) = V (q) = 0.

∂
Example 11.4. The vector field (sin ϕ) ∂θ (in MATH spherical coordinates) on the stan-
dard 2-sphere is a Jacobi field vanishing at the north and south poles. Therefore, the
north and south poles are conjugate point to each other along any great circle passing
through them. 
272 11. Curvatures and Topology

The relation between exponential maps and conjugate points is as follows:

Proposition 11.5. Consider an arc-parametrized geodesic γ on a Riemannian manifold

(M, g). Denote p = γ(0), T = γ̇, and q = expp (t0 T ) where t0 > 0. Then, q is conjugate to
p along γ if and only if the tangent map (expp )∗t0 T : Tt0 T (Tp M ) → Tq M is singular.

Proof. Before giving the proof, we first observe the following key fact. Let γs (t) be a
family of geodesic defined by:


γs (t) := expp t(T + sW )
where W is any vector in Tp M .
Then, γ0 = γ. The Jacobi field generated by this family is given by:
∂γs (t) ∂

(11.3) V (t) := = expp tT + s(tW )) = (expp )∗tT (tW ).
∂s s=0 ∂s s=0

The last step follows from the fact that the derivative is the directional derivative of expp
at tT along tW . In particular, V (q) = (expp )∗t0 T (t0 W ).
To show the (=⇒)-part, we consider a
basis {Ei }ni=1 of Tp M , and consider the
family of geodesics γsi (t) := expp t(T + sEi ) which generates the Jacobi fields Vi (t) =
(expp )∗tT (tEi ) according to the above computation. These Jacobi fields are linearly
independent because (expp )∗tT is invertible for small t ≥ 0: suppose there are constants
λi ’s such that
Xn
λi Vi (t) ≡ 0,
i=1
then we have
 Xn 
(expp )∗tT t λi Ei ≡ 0.
i=1
Pick a small τ > 0 such that (expp )∗τ T is invertible, we have
n
X n
X
τ λi Ei = 0 =⇒ λi Ei = 0.
i=1 i=1

By linear independence of Ei ’s, we get λi = 0 for any i. Therefore, {Vi (t)}ni=1 is a basis
of Jacobi fields that vanish at p (which is an n-dimensional vector space).
To prove the desired claim, we suppose otherwise that there exists a non-zero Jacobi
field V along γ such that V (p) = V (q) = 0 but (expp )∗t0 T is invertible. From above, V (t)
must be spanned by {Vi (t)}ni=1 :
n
X
V (t) ≡ ci Vi (t).
i=1

However, it would show

n
X n
X 
0 = V (q) = ci Vi (t0 ) = (expp )∗t0 T t 0 ci E i ,
i=1 i=1

and consequently
n
X
t0 ci Ei = 0 =⇒ ci = 0 for any i.
i=1
It is a contradiction to the fact that V (t) 6≡ 0.
11.1. Jacobi Fields 273

The (⇐=)-part is easier: suppose (expp )∗t0 T is singular, and in particular, there exists
W 6= 0 in Tp M such that
(expp )∗t0 T (W ) = 0.
Then, the Jacobi field V (t) defined in (11.3) satisfies:
V (q) = (expp )∗t0 T = 0.
By invertibility of (expp )∗tT for t > 0 small, V (t) is non-zero Jacobi field. It completes
the proof. 
274 11. Curvatures and Topology

11.2. Index Forms

11.2.1. Second Variations of Arc-Length. In Chapter 9, we have computed the
first variation formula of the arc-length functional. If γs (t) : (−ε, ε) × [a, b] is a 1-
parameter family of curves with the same end-points γs (a) = p and γs (b) = q for any
s ∈ (−ε, ε), then we have shown that
Z b   Z b  
d T T
L(γs ) = − g S, ∇T dt = g ∇T S, dt
ds a |T | a |T |
where S = ∂γ ∂γs
∂s and T = ∂t . One necessary condition for γ0 to minimize the arc-length
s

T
is that ∇T |T | = 0, and if we assume γ0 has constant speed, then ∇T T = 0 along γ.
However, to determine whether it is a local minimum, we need to consider the second
derivative. In this section, we will derive the second variation formula of arc-length and
discuss its applications.

Proposition 11.6 (Second Variations of Arc-Length). Let γs (t) : (−ε, ε) × [a, b] be a

1-parameter family of curves with the same end-points γs (a) = p and γs (b) = q for any
s ∈ (−ε, ε), γ0 is a unit-speed geodesic. Then, we have:
Z b
d2 2
(11.4) L(γ s ) = ∇T S N − Rm(S, T, T, S) dt
ds2 s=0 a
∂γs ∂γs
where S = ∂s , T = ∂t , and S N = S − g(S, T )T is the normal projection of S.

Proof. We differentiate the first derivative of L:

d2
L(γs )
ds2
Z b  
d T
= g ∇T S, dt
ds a |T |
Z b    
T T
= g ∇S ∇T S, + g ∇T S, ∇S dt
a |T | |T |
|T | ∇S T − g ∇S T, |TT | T
Z b  
!
T
= g ∇S ∇T S, + g ∇T S, 2 dt.
a |T | |T |
Now take s = 0, we have |T | = 1 and ∇T T = 0 (since γ0 is a geodesic), then we have
d2
L(γs )
ds2 s=0
Z b
= g(∇S ∇T S, T ) + g(∇T S, ∇S T ) − g(∇T S, T )g(∇S T, T ) dt
a
Z b
2
= g(∇S ∇T S, T ) + |∇T S| − g(∇T S, T )2 dt
a
Z b
2
= g(∇T ∇S S, T ) + g(Rm(S, T )S, T ) + |∇T S| − g(∇T S, T )2 dt
a

where we have used the fact that [S, T ] = 0. By Pythagoreas’ Theorem, we have
2 2
|∇T S| − g(∇T S, T )2 = (∇T S)N ,
and as ∇T T = 0, we can also show
∇T S N = ∇T (S − g(S, T )T ) = ∇T S − g(∇T S, T )T = (∇T S)N .
11.2. Index Forms 275

For the first term in the integrand, we have

Z b Z b
d
g(∇T ∇S S, T ) dt = g(∇S S, T ) − g(∇S S, ∇T T ) dt = 0
a a dt
since S(a) = S(b) = 0, and ∇T T ≡ 0 along γ.
Summing all up, we have proved the desired formula (11.4). 

Remark 11.7. The Rm(S, T, T, S) term in (11.4) also equals Rm(S N , T, T, S N ) since
Rm(T, T, T, T ) = 0. 

Inspired by (11.4), we define:

Definition 11.8 (Index Forms). Let γ : [a, b] → M be a geodesic on a Riemannian

manifold (M, g). The index form I : Vγ × Vγ → R of γ is a bilinear form on the following
space of vector fields:
Vγ := {V (t) | V is a vector field on γ such that V (a) = V (b) = 0 and g(V, γ̇) ≡ 0},
and is defined as
Z b
I(V, W ) := g(∇T V, ∇T W ) − Rm(V, T, T, W ) dt
a
where T := γ̇.

In particular, for any V ∈ Vγ , the second variation of L(γ) along the variation field V
is given by I(V, V ). When γ is minimizing geodesic, it is necessary that I(V, V ) ≥ 0 for
any V ∈ Vγ . In other words, if we want to show a certain geodesic γ is not minimizing,
one needs to construct a vector field V ∈ Vγ such that I(V, V ) < 0.
When V ∈ Vγ is a C ∞ Jacobi field and W ∈ Vγ is any C ∞ vector field, one can use
integration by parts and show:
Z b
 t=b−

I(V, W ) = g(∇T V, W ) t=a+ − g ∇T ∇T V + Rm(V, T )T, W dt = 0.
a

where we have used the fact that W (a) = W (b) = 0 and ∇T ∇T V + Rm(V, T )T = 0.
Note that if V and W are merely piecewise smooth on [a, b], say they are smooth on
[a, c) and (c, b], then we have
 t=c−  t=b−
I(V, W ) = g(∇T V, W ) t=a+ + g(∇T V, W ) t=c+ .

11.2.2. Geodesics Beyond Conjugate Points. One fundamental fact about geodesics
is that it is never minimizing if γ has an interior point conjugate to the starting point.

Proposition 11.9. Suppose γ : [a, b] → M is a unit-speed geodesic and there exists

c ∈ (a, b) such that γ(c) is conjugate to γ(a), then γ is not a minimizing geodesic.

Proof. By Proposition 11.5, the given condition implies there exists a non-zero Jacobi
field V (t) defined on [a, c] such that V (a) = V (c) = 0. We can assume that V is normal
to γ̇.
Then, we extend V to the whole curve γ : [a, b] → M as follows:
(
V (t) if t ∈ [a, c]
V (t) :=
e .
0 if t ∈ (c, b]
276 11. Curvatures and Topology

Note that Ve (t) is not smooth. We are going to search for a C ∞ vector field W defined on
[a, b] and a small ε > 0 such that
I(Ve + εW, Ve + εW ) < 0.

Consider that
I(Ve + εW, Ve + εW ) = I(Ve , Ve ) + 2εI(Ve , W ) + ε2 I(W, W ).
Let’s compute each term one-by-one:
 t=c−  t=b−
I(Ve , Ve ) = g(∇T Ve , Ve ) t=a+ + g(∇T Ve , Ve ) t=c+
= g(∇T V (c), V (c)) − g(∇T V (a), V (a)) + g(∇T 0, 0) − g(∇T 0, 0)
=0
 t=c−  t=b−
I(Ve , W ) = g(∇T Ve , W ) t=a+ + g(∇T Ve , W ) t=c+
= g(∇T V (c), W (c)) − g(∇T W (a), W (a))
In order for W ∈ Vγ , we need W (a) = W (b) = 0. If we can construct W ∈ Vγ such that
W (c) = −∇T V (c), then we will have I(Ve , W ) < 0. Regardless of the sign of I(W, W ),
we will have
I(Ve + εW, Ve + εW ) = 2εI(Ve , W ) + ε2 I(W, W ) < 0
for sufficiently small ε > 0.
One can use a bump function to construct such a vector field W . Let ρ : [a, b] →
[0, 1] be a C ∞ function such that ρ(a) = ρ(b) = 0 and ρ(c) = 1. Take any paral-
lel unit vector field E(t) which is normal to γ̇(t). This vector field exists by parallel
transporting −∇T V (c) along the curve γ. Let X(t) be this parallel transport, then
W (t) := ρ(t)X(t)N is a smooth vector field in Vγ such that W (c) = −∇T V (c)N and so
2
I(Ve , W ) = − ∇T V (c)N < 0, completing the proof. 

11.2.3. Bonnet-Myers’ Theorem. We are ready to prove the Bonnet-Myers’ The-

orem mentioned in the introduction. The key ingredient is the use of index forms to
show that any minimizing geodesic cannot exceed certain length. The theorem was first
due to Bonnet who assumed a lower bound on the sectional curvature. It was then later
extended by Myers who only required a Ricci lower bound.

Theorem 11.10 (Bonnet-Myers). Let (M n , g) be a complete Riemannian manifold of

dimension n ≥ 2 such that there exists a constant k > 0 such that
Ric ≥ (n − 1)kg,
π
then the diameter of (M, g) is bounded above by √
k
.
As a corollary, M is compact and has finite fundamental group.

π
Proof. We claim that no minimizing geodesic has length greater than √
k
. Suppose
otherwise and γ(t) : [0, L] → M is a unit-speed geodesic with L > √πk .
Construct a parallel orthonormal frame {Ei (t)}ni=1 such that E1 (t) = γ̇(t), then for
each i we define  π 
Vi (t) := sin t Ei (t),
L
and consider its index form I(Vi , Vi ). Since each Ei (t) is parallel, we have
π  2 
π  π π
∇γ̇ Vi = cos t Ei (t) and ∇γ̇ ∇γ̇ Vi = − sin t Ei (t).
L L L2 L
11.2. Index Forms 277

Hence, we have
Z L
I(Vi , Vi ) = − g(∇γ̇ ∇γ̇ Vi , Vi ) + Rm(Vi , γ̇, γ̇, Vi ) dt
0
Z L 2
π 2 π

2 π

= 2
sin t − sin t Rm(Ei , E1 , E1 , Ei ) dt.
0 L L L
Summing up i over 2 to n, we get
n Z L n
!
 (n − 1)π 2 X
2 π
X
I(Vi , Vi ) = sin t − Rm(Ei , E1 , E1 , Ei ) dt.
i=2 0 L L2 i=2
Note that by the given condition about the Ricci curvature, we have
X n
Rm(Ei , E1 , E1 , Ei ) = Ric(E1 , E1 ) ≥ (n − 1)kg(E1 , E1 ) = (n − 1)k.
i=2
Hence, we have
n L
(n − 1)π 2
Z  
X 
2 π 
I(Vi , Vi ) ≤ sin t − (n − 1)k dt < 0
i=2 0 L L2
since we assumed L > √πk . At least one of the Vi ’s gives I(Vi , Vi ) < 0, and this shows γ
cannot be a minimizing geodesic. It leads to a contradiction.
√
(M, g) is compact because expp now maps B(0, π/ k) onto M . To argue that it has
finite fundamental group, we consider its universal cover π : M f → M . Recall that M
f
∗ ∗
admits a Riemannian metric π g which is locally isometric to g, so (M , π g) also satisfies
f
the same Ricci curvature lower bound. This shows M f is compact too, and consequently
π −1 (p) is a finite set for any p ∈ M . This shows π1 (M ) is a finite group. 
Remark 11.11. The Ricci condition cannot be relaxed to Ric > 0. Any non-compact
regular surface in R3 with positive Gauss curvature serves as a counter-example, and there
are plenty of them! Recall that for regular surfaces, we have K = 2R and Ric = R2 g. 
Remark 11.12. As the circle S1 has an infinite π1 (isomorphic to Z) and so is S1 × M for
any complete Riemannian manifold M , by the Bonnet-Myers’ Theorem it is impossible
for S1 × M to admit a Riemannian metric whose Ricci curvature has a uniform positive
lower bound. If M is compact, it is even impossible for S1 × M to admit a Riemannian
metric whose positive Ricci curvature. 
The equality of Bonnet-Myers’ Theorem was proved by Cheng:

Theorem 11.13 (Cheng, 1975). Let (M n , g) be a complete Riemannian manifold of

dimension n ≥ 2. Suppose there exists a constant k > 0 such that Ric ≥ (n − 1)kg on M
and the diameter of (M, g) equals √πk , then (M n , g) is isometric to the round sphere of
radius √1k .

11.2.4. Synge’s Theorem. Another application of index forms is the proof of

Synge’s Theorem, which is about the dimension, orientability, and simply-connectedness
of a Riemannian manifold with positive sectional curvature.

Theorem 11.14 (Synge). Let (M, g) be a compact Riemannian manifold with positive
sectional curvature. Then,
• if dim M is even and M orientable, then M is simply-connected;
• if dim M is odd, then it must be orientable.
278 11. Curvatures and Topology

The proof of the theorem is based on two lemmas, one analytic and another linear
algebraic.

Lemma 11.15. Let (M, g) be a compact Riemannian manifold, then in every free homo-
topy class [γ] of smooth closed curves , there exists a closed smooth geodesic γ
e (smooth at
based point too) that minimizes the arc-length among all curves in [γ].

The proof of the lemma is by some convergence and compactness argument. We

omit the proof here. Interested readers may consult J. Jost’s book Theorem 1.5.1.
Another lemma is the following observation on orthogonal matrices.

Lemma 11.16. Any orthogonal matrix A ∈ O(n) with det(A) = (−1)n−1 must have 1
as one of its eigenvalues.

Proof. Any real eigenvalue of A ∈ O(n) is either 1 or −1. It can be shown by considering
Av = λv, so that
2 2 2
kvk = v T AT Av = (Av)T (Av) = Av · Av = kAvk = λ2 kvk
If n is even, then det(A) = −1. Since complex eigenvalues occur as conjugate pairs,
the product of all complex (non-real) eigenvalues is positive, and hence the product
of real eigenvalues must be negative. There are even many real eigenvalues counting
multiplicity, so at least one of the real eigenvalue is 1. If n is odd, then det(A) = 1. There
are odd many real eigenvalues counting multiplicity. We have at least one of the real
eigenvalue must be 1. 

Proof of Theorem 11.14. We first prove the first statement. Suppose dim M is even and
M is orientable, but M is not simply-connected. Take a non-trivial homotopy class [γ]
of closed curves with γ being the minimizer of arc-length within the class (such γ exists
thanks to Lemma 11.15). Consider the parallel transport map Pγ : Tγ(0) M → Tγ(0) M
along γ which has determinant 1 by orientability. Note that Pγ (γ̇(0)) = γ̇(0) as γ is a
smooth closed geodesic. Consider the orthogonal complement E of span{γ̇(0)} in Tγ̇(0) M
so that E is invariant under Pγ and det(Pγ E ) = 1. Note that E has odd dimension, by
Lemma 11.16, there exists an eigenvector X0 ∈ E such that
Pγ (X0 ) = X0 .
Extend X0 by parallel transport along γ so that ∇γ̇ X(t) = 0 and X(0) = X0 , then we
get: Z
2
I(X, X) = |∇γ̇ X| − Rm(X, γ̇, γ̇, X) < 0,
γ | {z } | {z }
=0 >0
1
hence γ is not a minimizing geodesic . It leads to a contradiction, completing the proof
of the first statement.
The second statement can be proved in a similar way. Suppose dim M is odd, but M
is not orientable. Then, one can find a closed curve γe such that det Pγe = −1. Let γ ∈ [e γ]
be a smooth closed minimizing geodesic in the free homotopy class [e γ ], then we still have
det Pγ = −1 by continuity. Since Pγ (γ̇(0)) = γ̇(0), and the orthogonal complement E of
span{γ̇(0)} has odd dimension. By Lemma 11.16, there exists an eigenvector X0 ∈ E
such that Pγ (X0 ) = X0 . The rest of the proof goes exactly as in the even dimension case.

1Note that the index form we have discussed before require the variation vector field to vanish at end points, which is not
the case here. However, as the curve γ and vector field X(t) is smooth everywhere including the base point, the boundary
terms of integration by parts also vanish. Therefore, the second variation formula of L(γ) along X is also given by the index
form I(X, X).
11.3. Spaces of Constant Sectional Curvatures 279

11.3. Spaces of Constant Sectional Curvatures

The goal of this last section is to classify all simply-connected, complete Riemannian
manifolds with constant sectional curvatures. We will show that they are either Sn , Rn ,
or Hn , depending on whether the sectional curvature is positive, zero, or negative.
We will first prove a topological result, called the Cartan-Hadamard’s Theorem,
concerning spaces with non-positive sectional curvatures (not necessarily constant). The
theorem pinpoints the topological type of such a manifold. Then, we will use Jacobi
fields to express these metrics explicitly, and show that they are isometric to one of the
standard metrics of Sn , Rn , and Hn .

11.3.1. Cartan-Hadamard’s Theorem. We first determine the topological type of

manifolds with non-positive sectional curvatures. Here, non-positive sectional curvature
means Kp (X, Y ) ≤ 0 for any p ∈ M and any linearly independent vectors X, Y ∈ Tp M .

Theorem 11.17 (Cartan-Hadamard). Let (M n , g) be a complete, connceted Riemannian

manifold with non-positive sectional curvatures. Then, for any p ∈ M , the exponential
map expp : Tp M → M is a covering map.
Consequently, M n is diffeomorphic to a quotient manifold of Rn . If in addition M n is
simply-connected, then M n is diffeomorphic to Rn .

Proof. One key ingredient of the proof is that the existence of a Jacobi field V (t) along
a geodesic γ(t) : [a, b] → M with V (a) = V (b) = 0 is equivalent to the non-invertibility
of (expγ(a) )∗ at the point corresponding to γ(b) (see Proposition 11.5). On the other
hand, such a Jacobi field does not exist if the metric has non-positive sectional curvature
according to (11.2). Hence, (expp )∗ is always invertible, and by the inverse function
theorem, it is a local diffeomorphism everywhere on Tp M . In particular, ge := (expp )∗ g
defines a Riemannian metric on Tp M , and (Tp M, ge) and (M, g) are locally isometric
through the map expp .
We first show that expp is surjective. Given a point q ∈ M , we let γ be the minimizing
unit-speed geodesic from p to q. Suppose d(p, q) = r > 0, then the geodesic γ is given by
γ(t) = expp (tγ̇(0)).
Hence, we have q = γ(r) = expp (rγ̇(0)). This shows expp is surjective.
Next, we show that for each q ∈ M , there exists ε > 0 such that
{Bε (Q)}Q∈expp−1 (q)
is a disjoint collection of open sets in Tp M . We pick such an ε > 0 so that Bε (q) is a
geodesic ball such that all geodesics from q leaves the ball through ∂Bε (q) (before they
come back to the ball, if ever). Index the set exp−1 p (q) by {Qα }. For any distinct pair
of Qα and Qβ , we connect them through a minimizing geodesic γ e (with respect to the
metric ge). Then, the curve γ := expp ◦eγ is a geodesic on M from q to q. However, by our
choice of ε, such a geodesic γ must go outside the geodesic ball Bε (q), and hence has
length > 2ε. This shows Qα and Qβ must be more than 2ε apart, so Bε (Qα ) and Bε (qβ )
are disjoint.
Then we argue that

G
exp−1
p Bε (q) = Bε (Qα ).
α
The ⊃-part is easy. Suppose X ∈ Bε (Qα ) for some Qα . As expp is a local isometry and
d(X, Qα ) < ε, so we have d expp (X), expp (Qα )) < ε too. It shows X ∈ exp−1


p Bε (q) .
280 11. Curvatures and Topology

Conversely, if Y ∈ exp−1


p Bε (q) , then consider the points y := expp (Y ) and q in
M . Let γ be a minimizing geodesic from y to q, and γ e be the geodesic lifted on Tp M ,
i.e. γ
e is geodesic on (Tp M, ge) such that γ e˙ (0)) = γ̇(0). Let r be
e(0) = Y , and (expp )∗ (γ


the length of γ so that d(y, q) = r < ε, then we have expp γ e(r) = γ(r) = q. Hence
Q := γ e(r) ∈ exp−1
p (q). Note that d(Q, Y ) = d(q, y) = r < ε by isometry, so Y ∈ Bε (Q).
This proves ⊂-part of our claim. It completes the proof that expp is a covering map.


11.3.2. Gauss Lemma. We next prove a fundamental result about radial tangent
vectors from 0 in Tp M and its image curve under the exponential map.

Lemma 11.18 (Gauss). Let (M, g) be a complete Riemannian manifold, p ∈ M , and

consider the exponential map expp : Tp M → M . Denote the radial vector in Tp M
∂
by r ∂r = r∂r (where r is the distance from the origin). Then, for any tangent vector
X ∈ Tr∂r (∂Br (0)) ⊂ Tr∂r (Tp M ) ∼
= Tp M , we have


g (expp )∗r∂r (r∂r ), (expp )∗r∂r (X) = 0.
Here we identify Tr∂r (Tp M ) with Tp M , so that the r∂r in blue is regarded as in Tr∂r (Tp M ).

∂
Proof. Consider a curve σ(s) ⊂ ∂Br (0) ⊂ Tp M such that σ(0) = r ∂r and σ 0 (0) = X,
and define a family of geodesics

γs (t) := expp (tσ(s)).

For each fixed s, the curve γs (t) is a geodesic by the definition of expp . Next we compute

∂ d
γs (t) = expp (0 · σ(s)) = 0
∂s (s,t)=(0,0) ds s=0
∂ d
γs (t) = expp (tσ(0))
∂t (s,t)=(0,0) dt t=0
= (expp )∗0 (σ(0)) = σ(0)
∂ d
γs (t) = expp (σ(s))
∂s (s,t)=(0,1) ds s=0
= (expp )∗σ(0) (σ 0 (0))
= (expp )∗r∂r (X)
∂ d
γs (t) = expp (tσ(0))
∂t (s,t)=(0,1) dt t=1
d
= expp (σ(0) + tσ(0))
dt t=0
= (expp )∗σ(0) (σ(0))
= (expp )∗r∂r (r∂r)

Therefore, to prove our claim, it suffices to show

 
∂γs (t) ∂γs (t)
g , = 0.
∂s ∂t (s,t)=(0,1)
11.3. Spaces of Constant Sectional Curvatures 281

As before, we denote S = ∂γ∂s s (t)

and T = ∂γ∂t
s (t)
for simplicity. Since we already
know g(S, T ) = 0 when (s, t) = (0, 0), we then calculate:
d
g(S, T ) = g(∇T S, T ) + g(S, ∇T T )
dt | {z }
=0
= g(∇S T, T )
1 d
= g(T, T ).
2 ds
2
We argue that |T | = g(T, T ) is independent of s as follows. With a fixed s, the length of
the curve segment γs (t) on [t, t + τ ] is
Z t+τ


L γs [t,t+τ ] = |T | .
t
Recall that the geodesic γs (t) is given by expp (tσ(s)), so that


L γs [t,t+τ ] = τ |σ(s)|
by the definition of expp . This shows
Z t+τ
d d

|T | = |T | = L γs [t,t+τ ]
= |σ(s)| = r
dτ τ =0 t dτ τ =0
since the curve σ(s) ⊂ ∂Br (0) ⊂ Tp M . This completes proof that
d 1 d 2
g(S, T ) = |T | = 0,
dt 2 ds
and conclude that g(S, T ) = 0 when (s, t) = (0, 1) as desired. 

In short, the Gauss’s Lemma asserts that radial lines and round spheres in Tp M
remain to be orthogonal under the image of expp . We will use this lemma in the next
subsection to classify spaces with constant sectional curvatures.

11.3.3. Classification of Space Forms. Now we are ready to give a complete

classification of (simply-connected) spaces of constant sectional curvatures. The key
ingredient is to make use of the Jacobi fields (which are explicitly solvable) to give a
fairly explicit expression of the Riemannian metric. We will see that such a Riemannian
metric is uniquely determined by the sectional curvature, hence proving uniqueness of
such a metric up to isometry.

Theorem 11.19 (Uniformization Theorem of Constant Sectional Curvature Spaces).

Any simply-connected, complete Riemannian manifold of constant sectional curvatures
must be isometric to one of the standard models: Sn , Rn , or Hn .
Consequently, any complete Riemannian manifold of constant sectional curvatures
must be isometric to a quotient manifold of either Sn , Rn , or Hn .

Proof. Let (M n , g) be a simply-connected, complete Riemannian manifold with constant

sectional curvature C. By Cartan-Hadamard’s Theorem, we already know such (M, g) is
diffeomorphic to Rn in case of C = 0 or C < 0, and the exponential map expp : Tp M →
M is a diffeomorphism.
We will first deal with the cases C = 0 and C < 0. Let (M fn , ge) be Rn with the
n
Euclidean metric (in case of of C = 0), or H with the hyperbolic metric of constant
sectional curvature C (in case of C < 0). Pick any point p ∈ M and pe ∈ M f, and
we identify Tp M and TpeM f. Denote the exponential maps by expp : Tp M → M and
xpp : TpeM → M .
eg f f
282 11. Curvatures and Topology

We will show that

g (expp )∗ (X), (expp )∗ (X) = ge (g
expp )∗ (X), (g
expp )∗ (X)
∗
for any X ∈ Tr∂r (Tp M ) ∼
= Tp M . This is implies exp∗p g = eg xpp ◦ exp−1
xpp ge so that eg p is
an isometry from M to M . Note that expp is invertible in case of C = 0 and C < 0 by
f
Cartan-Hadamard’s Theorem, and simply-connectedness of M .
We first assume that X ∈ Tr∂r (∂Br (0)) ⊂ Tr∂r (Tp M ), and will then handle an
arbitrary X ∈ Tr∂r (Tp M ) using Gauss’s Lemma. Consider the family of geodesics:
  
∂ X
γs (t) := expp t +s .
∂r r
∂ ∂
Note that use ∂r for the radial vector instead of r ∂r so that γ0 (t) is a unit-speed geodesic.
Many results about Jacobi fields that we derived required γ0 to be unit-speed.
The family of geodesics generates a Jacobi field
 
∂ tX
V (t) := γs (t) = (expp )∗t∂r .
∂s s=0 r
In particular, we have V (r) = (expp )∗r∂r (X). From the above discussion, we need to
find out |V (r)|g = (expp )∗r∂r (X) g . Since V (0) = 0, such a Jacobi field has been solved
explicitly in page 269, in which we found:
 u0 (0) √
 √C sin(t C)
 if C > 0
0
|V (t)| = u(t) = u (0)t if C = 0 .
 u0 (0) √
sinh(t −C) if C < 0
√
−C

To find u0 (0), we consider the results from Exercise 11.2, which shows
Xi ∂
V (t) = t
r ∂ui γ(t)

under geodesic normal coordinates (u1 , · · · , un ) at p. Hence, we have

d Xi ∂ |X|g
u0 (0) = |V (t)| = = .
dt t=0 r ∂ui γ(0) r

Under geodesic normal coordinates at p, we also have |X|g = kXk, where k·k is the
standard Euclidean norm of Tp M ∼
= Rn . Therefore, we conclude that |V (r)|g depends
only on r, C, and kXk.
By exactly the same argument, we can get the same result for (g expp )∗ (X) ge. There-
fore, we have:
 kXk √
 r√C sin(r C)
 if C > 0
expp )∗r∂r (X) ge = (expp )∗r∂r (X) g = |V (r)| = kXk
(g if C = 0
 kXk
 √
√
r −C
sinh(r −C) if C < 0
for any X ∈ Tr∂r (∂Br (0)).
Now given an arbitrary X ∈ Tr∂r (Tp M ), one can decompose it into
X = Xrad + Xsph
where Xrad is the radial component, and Xsph is tangential to ∂Br (0). By Gauss’s Lemma,
we have


g (expp )∗r∂r (Xrad ), (expp )∗r∂r (Xsph ) = 0.
11.3. Spaces of Constant Sectional Curvatures 283

Therefore, one can conclude that

2 2 2
(expp )∗r∂r (X) g
= (expp )∗r∂r (Xrad ) g
+ (expp )∗r∂r (Xsph ) g .
By the definition of expp (and a similar arc-length argument as in the proof of the Gauss’s
Lemma), one can argument easily that
2 2
(expp )∗r∂r (Xrad ) g
= kXrad k .
2
By the previous computations, we also have that (expp )∗r∂r (Xsph ) g
depends only on r,
C, and Xsph . One can repeat these arguments on eg
xpp and yield the same result.
Therefore, we conclude that
(g
expp )∗r∂r (X) g
e
= (expp )∗r∂r (X) g

for any r > 0, X ∈ Tr∂r (Tp M ) in the case of C = 0 and C < 0. Hence, g and ge are
isometric in these cases.
Finally, we deal with the case C > 0. Let (Sn , ge) be the round sphere with sectional
curvature C. With the same notations as the above, one can also argument in the same
way that g is isometric to ge locally (in the region on which expp is a diffeomorphism).
Therefore, by compactness of M (guaranteed by Bonnet-Myers’ Theorem), one can cover
M by finitely many open geodesic balls {Bα } each of which is isometric to another
geodesic ball {Beα } on Sn via the map, say, ϕα : Bα ⊂ M → B eα ⊂ Sn .
We want to glue these local isometries ϕα ’s to form a global isometry. However, it is
not a priori true that any pair ϕα and ϕβ of local isometries must agree on the overlap.
However, by the transitive action of SO(n) acting on Sn , one can compose an isometry
Φαβ : Sn → Sn that maps ϕα (Bα ∩ Bβ ) to ϕβ (Bα ∩ Bβ ) isometrically and Φαβ ◦ ϕα agrees
with ϕβ on the overlap. Replace ϕα by Φαβ ◦ ϕα . Repeat this replacement process for
each overlap (there are finitely many), one can construct a global isometry ϕ : M → Sn .
It completes the proof of the case C > 0. 

By the above uniformization theorem, we say that Rn , Sn , and Hn are standard

models in Riemannian geometry. These three models and their quotient manifolds are
called space forms.
Therefore, to prove a certain Riemannian manifold is diffeomorphic to a sphere (or
its quotient), one can show that it admits a Riemannian metric with a constant positive
sectional curvature. The Ricci flow, and also other geometric flows as well, is a very
effective tool to produce such a metric by “distributing” curvatures uniformly across the
manifold like heat diffusion.

** This is the end of this lecture notes. **

* The course MATH 6250I will continue on the introduction to the Ricci flow. *
Bibliography

[dC76] Manfredo P. do Carmo, Differential geometry of curves and surfaces, Prentice-Hall, Inc.,
Englewood Cliffs, N.J., 1976, Translated from the Portuguese. MR 0394451
[dC94] , Differential forms and applications, Universitext, Springer-Verlag, Berlin, 1994,
Translated from the 1971 Portuguese original. MR 1301070
[Küh05] Wolfgang Kühnel, Differential geometry, curves-surfaces-manifolds, Student Mathematical
Library, vol. 16, American Mathematical Society, Providence, RI, 2005. MR 3443721
[Lee09] Jeffrey M. Lee, Manifolds and differential geometry, Graduate Studies in Mathematics,
vol. 107, American Mathematical Society, Providence, RI, 2009. MR 2572292
[Lee13] John M. Lee, Introduction to smooth manifolds, second ed., Graduate Texts in Mathemat-
ics, vol. 218, Springer, New York, 2013. MR 2954043
[Mun00] James Munkres, Topology, 2 ed., Prentice Hall, 2000.

285