You are on page 1of 141

Lectures on differential geometry

Ben Andrews

Mathematical Sciences Institute, Australian National Univer-


sity; and Yau Mathematics Center, Tsinghua University
E-mail address: Ben.Andrews@anu.edu.au
Abstract. These notes were produced to accompany the course MATH3342
(Advanced Differential Geometry) offered at ANU in the first semester
of 2017.
Contents

Chapter 1. Review of calculus and the definition of manifolds 1


Chapter 2. Submanifolds 9
Chapter 3. Examples of manifolds, and smooth maps 15
Chapter 4. Tangent vectors 21
Chapter 5. Partitions of unity and applications 31
Chapter 6. The tangent bundle and other vector bundles 41
Chapter 7. Sard’s theorem and Embedding theorems 51
Chapter 8. Vector fields, Lie brackets, and sections 59
Chapter 9. ODE, Lie derivatives and the Frobenius theorem 69
Chapter 10. Differentiating vector fields 79
Chapter 11. Riemannian metrics and the Levi-Civita connection 91
Chapter 12. Riemannian geodesics and completeness 99
Chapter 13. Tensors and tensor fields 105
Chapter 14. Hypersurfaces and submanifolds 123
Bibliography 137

v
CHAPTER 1

Review of calculus and the definition of manifolds

Contents
1.1. Introduction 1
1.2. Review of differential calculus of several variables 1
1.2.1. The derivative 1
1.2.2. The Chain Rule 2
1.2.3. Higher derivatives and Taylor approximations 3
1.2.4. Diffeomorphisms and local diffeomorphisms 4
1.2.5. The inverse function theorem 5
1.2.6. Immersions, Embeddings and Submersions 5
1.3. Differentiable manifolds 7
1.3.1. Manifolds 7
1.3.2. Charts, transition maps and atlases 7

1.1. Introduction

The objects of study in differential geometry (or differential topology) are


smooth manifolds. Roughly speaking these are spaces which are endowed
with a structure designed to make sense of notions of smoothness and dif-
ferentiability. Our starting point is differential calculus in several variables,
which provides notions of differentiability for functions between open sub-
sets of Euclidean spaces. We will briefly review these concepts, and then
proceed to the definition of manifold which involves identifying regions of
a space with regions in Euclidean space, in a way that ensures the ideas of
smoothness are well-defined.

1.2. Review of differential calculus of several variables

1.2.1. The derivative.


Definition 1.2.1. A function f from an open subset U of Rm to Rn is
differentiable at a point x ∈ U if it is well approximated by a linear function
L : Rm → Rn near x, in the sense that
f (y) = f (x) + L(y − x) + o(ky − xkRm ) as y → x,
1
2 1. REVIEW OF CALCULUS AND THE DEFINITION OF MANIFOLDS

or equivalently if
kf (y) − f (x) − L(y − x)kRn
lim = 0.
y→x ky − xkRm
In this case the linear map L (which is unique) is called the derivative of f
at x and denoted by Df |x .

Differentiability at a point x implies the existence of directional derivatives


in all directions at x, since we have
f (x + sv) − f (x) L(sv) + o(ksvk)
Dv f (x) = lim = lim = L(v) = Df |x (v).
s→0 s s→0 s
In particular the partial derivatives exist and we have
∂f
= Df |x (ei )
∂xi
where ei is the ith standard basis vector in Rm . This allows us to recon-
struct the derivative from the partial derivatives: Writing f and v ∈ Rm in
components as  1   1 
f v
 ..   .. 
f =  . , v= . 
fn vm
we have
∂f 1 ∂f 1
  
∂x1
··· ∂xm
v1
.. .. ..   .. 
Df |x (v) =   . .

. . .
∂f n ∂f n vm
∂x1
··· ∂xm
Remark. While differentiability implies the existence of partial derivatives,
the converse is not true: The function f (x, y) = √ xy
2 2
(and f (0, 0) = 0)
x +y
has partial derivatives at every point, and in fact directional derivatives at
every point, but is not differentiable at the origin. However, if a function has
partial derivatives at each point of an open set U , and each partial derivative
is a continuous function, then the function is necessarily differentiable.

1.2.2. The Chain Rule.


Theorem 1.2.2. Let U ⊂ Rk and V ⊂ Rm be open sets, and f : U → Rm
and g : V → Rn maps, and x ∈ U with f (x) ∈ V . If f is differentiable at x
and g is differentiable at f (x), then g ◦ f is differentiable at x and we have
D(g ◦ f )|x = Dg|f (x) ◦ Df |x .
Remark 1.2.3. There is an interesting construction which makes the state-
ment of the above theorem simpler (we will see later that this is a prototype
for the notion of the tangent bundle and the tangent map which makes sense
in the context of manifolds): For a point x ∈ Rm , by a ‘vector based at x’
we mean a pair (x, v) with v ∈ Rn . The set Rm x of vectors based at x is of
1.2. REVIEW OF DIFFERENTIAL CALCULUS OF SEVERAL VARIABLES 3

course just a copy of Rm , but the collection of all vectors based at points of
U ⊂ Rm corresponds to U × Rm .
The derivative makes sense as a map taking based vectors to based vectors,
by setting T f (x, v) := (f (x), Df |x (v)). That is, T f takes a vector based at
a point x to an image vector based at the image point.
Then the chain rule has the very simple statement
T (g ◦ f ) = T g ◦ T f.

1.2.3. Higher derivatives and Taylor approximations. We can


similarly define higher derivatives inductively as best polynomial approxi-
mations: Recall that a map L : V × · · · × V → W is called k-multilinear
| {z }
k times
if it is separately linear in each argument. Such a map is called symmet-
ric if it is independent of the order of the arguments: L(v1 , v2 , · · · , vk ) =
L(vσ(1) , · · · , vσ(k) ) where σ is any permutation of {1, · · · , k}.
The kth derivative is a symmetric k-multilinear map from Rm to Rn for each
k: If we have already defined such maps Df |x , D2 f |x , · · · Dk−1 f |x , then we
say f is k times differentiable at x if there is a symmetric k-multilinear map
L : Rm × · · · × Rm → Rn such that
1 1
f (x + v) = f (x) + Df |x (v) + D2 f |x (v, v) + · · · + Dk−1 f |x (v, · · · , v)
2 (k − 1)!
1
+ L(v, · · · , v) + o(kvkk ) as v → 0.
k!
In this case the map L is unique, and we denote it by Dk f |x .
Exercise 1.2.4. Show that the map L is unique if it exists [If a symmetric
k-multilinear map L satisfies L(v, · · · , v) = 0 for all v, why is it true that
L(v1 , · · · , v)k) = 0 for all vectors v1 , · · · , vk ?]

If the kth derivative Dk f |x exists, then the kth partial derivatives also exist,
and are given by
∂kf
= Dk f |x (ei1 , · · · , eik ).
∂xi1 · · · ∂xik
Since a k-multilinear map is determined by its action on basis elements by
multilinearity, this means that we can recover the kth derivative entirely
from the kth order partial derivatives.
Definition 1.2.5. Let U be an open set in Rm . For k ≥ 1, we define
C k (U, Rn ) to be the set of maps f from U to Rn such that Dj f |x exists for
every x ∈ U and 1 ≤ j ≤ k, and the map x 7→ Dk f (x) is continuous on U .
We say f is C ∞ (U, Rn ) if f is C k for every k ≥ 1.
4 1. REVIEW OF CALCULUS AND THE DEFINITION OF MANIFOLDS

In most situations we will work with C ∞ functions. In particular, I will use


the word ‘smooth’ to mean C ∞ .

Exercise 1.2.6. The ‘tangent map’ T f gives another way to define higher
derivatives: If f is a map from U to Rn , then T f is a map from U × Rm to
Rn × Rn , which is again a map between open sets of Euclidean spaces. So
we can make sense of T f being differentiable, and define it’s tangent map
T 2 f = T (T f ). How does this relate to the second derivative D2 f |x ? The
chain rule then says T 2 (g ◦ f ) = T 2 g ◦ T 2 f . What does this say about the
second derivatives of the composition? What about higher derivatives?

1.2.4. Diffeomorphisms and local diffeomorphisms.

Definition 1.2.7. If U and V are open sets in Rn , we say that f is a


diffeomorphism from U to V if f is a smooth bijection from U to V with
smooth inverse.

Definition 1.2.8. A map f : U ⊂ Rn → Rn is a local diffeomorphism near


x ∈ U if there exists a neighbourhood A of x such that the restriction of f
to A is a diffeomorphism to an open subset B of Rn .

Example 1.2.9. Here is an example of a map which is a local diffeomor-


phism everywhere, but not a diffeomorphism: Let U = (0, ∞) × R and
V = R2 \ {0}, and define f (r, θ) = (r cos θ, r sin θ). Then the restriction of
f to (0, ∞) × (a, a + 2π) is a diffeomorphism for any a ∈ R (so that f is a
local diffeomorphism about every point of U ) but f is not injective on U ,
hence not a diffeomorphism. Of course an injective map which is a local
diffeomorphism about every point is a diffeomorphism to some open set.

Diffeomorphisms are important because they provide a natural equivalence


relation between open sets (and later, we will see that this generalises to
an equivalence relation between manifolds): Two open sets A and B are
equivalent (up to diffeomorphism) if there exists a diffeomorphism ϕ from
A to B.

Remark. There are of course purely topological notions of equivalence: In


point-set topology the natural equivalence relation is homeomorphism (i.e.
existence of a continuous bijection with continuous inverse). A homeomor-
phism equivalence ϕ : A → B means that the two sets have the ‘same’ open
and closed sets (in the sense that a subset Ω ⊂ A is open/closed in A if and
only if ϕ(Ω) is open/closed in B) and the same continuous functions (in the
sense that a map f : B → X to a topological space X is continuous on B if
and only if f ◦ ϕ : A → X is continuous on A). Diffeomorphism equivalence
of course implies homeomorphism equivalence, but adds to this the fact that
the two spaces have the same smooth functions (f : B → Rn is smooth if
and only if f ◦ ϕ : A → Rn is smooth).
1.2. REVIEW OF DIFFERENTIAL CALCULUS OF SEVERAL VARIABLES 5

1.2.5. The inverse function theorem. The inverse function theorem


is a fundamental tool in differential geometry, and will apply it many times
— in particular next lecture when we discuss submanifolds.
Theorem 1.2.10. A smooth (or C k , k ≥ 1) map f : U ⊂ Rn → Rn for
which Df |x is invertible is a local (smooth, or C k ) diffeomorphism about x.

1.2.6. Immersions, Embeddings and Submersions. We make some


definitions of particular classes of smooth maps:
Definition 1.2.11. A smooth (or C k ) map f : U ⊂ Rm → Rn is called an
immersion if the derivative Df |x is injective for each x ∈ U (note that this
implies that n ≥ m). If f is also injective, and is a homeomorphism to its
image, then we say that f is an embedding.

A point of terminology: In topology, an embedding is a map which is a


homeomorphism to its image. The definition above corresponds to what
is called a smooth embedding (when the distinction is necessary). Thus a
smooth embedding is a map which is both a topological embedding and an
immersion.
The difference between an immersion and an embedding is slightly subtle:
An injective immersion is not necessarily an embedding, since we also have
to check the homeomorphism condition. To interpret this, we regard the
image set as a (metric/topological) subspace of the target Euclidean space
Rn , so that the open sets are the restrictions to the image set of open sets
in Rn . An immersion is smooth, so certainly continuous, and we know that
the inverse map exists since f is injective. The remaining condition is that
the f −1 is continuous, which means that images of open sets in U are open
in f (U ). That is, we require that for every open set A ⊂ U , there exists an
open set B ⊂ Rn such that f (A) = B ∩ f (U ).
A special case is useful to observe:
Definition 1.2.12. A map f is called proper if f −1 (K) is compact whenever
K is compact. A topological space is called locally compact if every point
has a neighbourhood with compact closure.

Note that any open subset of a finite dimensional Euclidean space is locally
compact.
Proposition 1.2.13. A continuous proper injective map f from a topological
space X to a locally compact topological space Y is a topological embedding.

Proof. As above, it suffices to show that images under f of open sets


in X are open, or equivalently (since f is injective) that images of closed
sets in X are closed. Let K be a closed set in X, and let y be a point
6 1. REVIEW OF CALCULUS AND THE DEFINITION OF MANIFOLDS

in the closure of the image f (K) ⊂ Y . Since Y is locally compact, we


can find a neighbourhood U of y which has compact closure. Since f is
proper, f −1 (Ū ) is compact. Since y ∈ f (K) there exists a sequence of
points (yk ) in f (K) which converges to y. Since yk ∈ f (K), we can write
yk = f (xk ) for some xk ∈ K. For k large enough we have yk ∈ U and hence
xk ∈ K ∩f −1 (Ū ) which is compact, so there exists a convergent subsequence
(xk0 ) converging to some point x ∈ K ∩ f −1 (Ū ) ⊂ K. Since f is continuous,
we have f (x) = limk0 →∞ f (xk0 ) = y, so y ∈ f (K) and f (K) is closed. 

From this we can conclude the following:


Corollary 1.2.14. An injective proper immersion is an embedding.
Definition 1.2.15. A smooth (or C k ) map f : U ⊂ Rm → Rn is called
a submersion if the derivative Df |x is surjective for each x ∈ U (note that
this implies that n ≤ m).

Immersions and submersions are therefore maps for which the derivative has
maximal rank. These maps have nice properties, captured by the implicit
function theorem:
Theorem 1.2.16 (The implicit function theorem). Let f : U ⊂ Rn+k → Rn
be a smooth map, and x0 = (y0 , z0 ) ∈ U ⊂ Rn × Rk such that the first n
columns of Df |x are linearly independent. There there is a neighbourhood
A of z in Rk , and neighbourhood B of f (y0 , z0 ) in Rn , a neighbourhood C of
x0 in Rn+k , and a smooth map F : B × A → Rn such that y = F (z, f (y, z))
for all points (y, z) ∈ C.

To understand this theorem, observe that if we fix the value of f (y, z) in B,


then y is given by a smooth function of z — that is, the equation f (y, z) =
c (which is really n equations applied to the n + k variables) implicitly
determines the n variables y as a function of the remaining k variables z.
The theorem also gives smooth dependence on the value c that we choose.
We will not prove this theorem, but the proof will essentially be covered
when we discuss submanifolds in the next lecture.
The following result generalises the implicit function theorem to a more
generally situation maps of constant rank:
Theorem 1.2.17 (The Rank Theorem). Let f : U ⊂ Rm → Rn be a smooth
map, such that the rank k of the derivative Df |x is constant on U . Then
for any x ∈ U there exist neighbourhoods A of x in Rm and B of f (x) in
Rn , and diffeomorphisms ϕ : A → C ⊂ Rm and η : B → D ⊂ Rn such that
η ◦ f ◦ ϕ−1 (x1 , · · · , xm ) = (x1 , · · · , xk , 0, · · · , 0)
for all (x1 , · · · , xm ) ∈ C.
1.3. DIFFERENTIABLE MANIFOLDS 7

To summarise: Up to diffeomorphism equivalence, constant rank maps are


linear.

1.3. Differentiable manifolds

1.3.1. Manifolds. A manifold is a space which is ‘locally Euclidean’.


More precisely, the definition is as follows:
Definition 1.3.1. A (topological) manifold M of dimension n is a (Haus-
dorff1, second countable2) topological space such that for every x ∈ M there
is a neighbourhood U of x and a homeomorphism ϕ : U → V , where V is
an open subset of Rn .

For students who have not studied point-set topology, it is sufficient to take
M to be a separable metric space3.

1.3.2. Charts, transition maps and atlases. A useful notion from


the definition of manifolds is the following:
Definition 1.3.2. A chart for a manifold M is a homeomorphism from an
open subset of M to an open subset of Rn .

Thus a topological manifold is a space equipped with a collection of charts


with domains covering M . The idea is that a chart gives a ‘local picture’ of
the space around any given point. This local picture can be used to decide
topological concepts such as continuity, convergence of sequences, etc in the
space: For example, a function f defined on M is continuous if and only if
f ◦ ϕ−1 is continuous for every chart ϕ of M . We want to use the same point
of view to define the concepts of differential calculus — the idea is that
we decide whether a function f is smooth by looking at the composition
f ◦ ϕ−1 . The difficulty is that different charts can give different answers: If
we change from a chart ϕ to another chart η, then we are replacing f ◦ ϕ−1
by f ◦ η −1 = (f ◦ ϕ−1 ) ◦ (ϕ ◦ η −1 ). The map ϕ ◦ η −1 is called the transition
map from the chart η to the chart ϕ. To make the definition of smoothness
independent of the choice of chart, we need the following condition:
1A topological space M is Hausdorff if for every two points x and y in M there exist
disjoint open subsets U and V with x ∈ U and y ∈ V .
2
A topological space is second countable if the topology of M has a countable basis.
That is, there is a countable collection B of open subsets of M such that (1) If x ∈ M and
U is a neighbourhood of x, then there exists A ∈ B with x ∈ A ⊂ U ; and (2) If A1 , A2 ∈ B
and x ∈ A1 ∩ A2 then there exists A ∈ B with x ∈ A ⊂ A1 ∩ A2 .
3 Recall that a metric space is separable if it has a countable dense subset. Separability
implies second-countability since we can take B to be the collection of balls of rational
radius about points in a countable dense subset of M . Not also that a metric space is
always Hausdorff.
8 1. REVIEW OF CALCULUS AND THE DEFINITION OF MANIFOLDS

Definition 1.3.3. Two charts ϕ and η for a manifold M are smoothly com-
patible if the transition map ϕ ◦ η −1 is a diffeomorphism on its domain of
definition.
Definition 1.3.4. A smooth atlas for a manifold M is a collection of charts
{ϕα : Uα → Vα }α∈A for M , such that any pair is smoothly compatible:
ϕα ◦ ϕ−1
β is a diffeomorphism from ϕβ (Uα ∩ Uβ ) to ϕβ (Uα ∩ Uβ ); and the
S
domains of definition cover M : α∈A Uα = M .

Now we can make sense of a smooth manifold :


Definition 1.3.5. A smooth manifold is a manifold M equipped with a
smooth atlas.
Example 1.3.6. Let M be an open subset of a Euclidean space Rn , and
take the atlas consisting of a single chart which is the identity map on U .
There is no compatibility condition to check, so this is a smooth atlas and
U is a smooth manifold. In particular Rn is a smooth manifold.
Definition 1.3.7. Two smooth atlases for a manifold M are compatible if
the union of the two is a smooth atlas. This defines an equivalence rela-
tion on smooth atlases. A differentiable structure is an equivalence class of
smooth atlases on M .

Given this, it would be better to define a smooth manifold to be a manifold


together with a differentiable structure, since we don’t want to consider the
particular choice of charts as important — replacing a given atlas of charts
with an atlas of compatible charts gives the same notions of differentiability,
and so should be considered as producing the same smooth manifold.4
Example 1.3.8. A given topological space can carry many different differen-
tiable structures: For example, if M = R we can take two charts ϕ1 (x) = x
and ϕ2 (x) = x3 . Then both {ϕ1 } and {ϕ2 } are smooth atlases for M , but
they define different smooth structures: The transition map between the two
charts is not a diffeomorphism, so the two atlases are not compatible and
define different differentiable structures. We will see later that these two
structures are equivalent in a different sense (diffeomorphism equivalence).

4This is a point where it is also more convenient to work with topological spaces
rather than metric spaces: In the definition, if we take M to be a metric space, then
we could replace the distance function by another equivalent one (i.e. one which has the
same open sets, or the same convergent sequences) without changing the manifold in any
substantial way. Thus we should take the metric space itself to have a metric structure
only up to this equivalence.
CHAPTER 2

Submanifolds

Contents
2.1. Introduction 9
2.2. Review of manifolds 9
2.3. Submanifolds 9
2.3.1. Equivalent descriptions of submanifolds 11
2.3.2. Examples 13

2.1. Introduction

Our aim today is to explore examples of smooth manifolds, having reached


the definition of manifolds at the end of last lecture. The first example
is submanifolds, which are those manifolds which are naturally included in
a background Euclidean space. We will explore these, including various
different ways they can arise, and consider some example.

2.2. Review of manifolds

Recall that a manifold is a (separable metric) space M , equipped with a


differentiable structure — the latter is defined by a smooth atlas, which is a
collection of charts, each of which is a homeomorphism from an open set of
M to an open set of Rn . The charts are required to be smoothly compatible
in the sense that the transition maps between them are diffeomorphisms,
every point of M must be in the domain of at least one chart.

2.3. Submanifolds

Definition 2.3.1. A subset Σ ⊂ Rn is a k-submanifold if every point x ∈ Σ


has a neighbourhood U on which there is a diffeomorphism Ψ : U → V ⊂ Rn
such that Ψ(Σ ∩ U ) = (Rk × {0}) ∩ V . Such a map Ψ is called a submanifold
chart for Σ.

Here Rk × {0} is the k-dimensional linear subspace of Rn given by


{(x1 , · · · , xk , 0, · · · , 0) : x1 , · · · , xk ∈ R}.
9
10 2. SUBMANIFOLDS

Thus the definition say that Σ can be ‘straightened’ (i.e. mapped to a


k-dimensional linear subspace) in a neighbourhood of any given point by
applying a suitable diffeomorphism to the ambient space Rn .
Proposition 2.3.2. A k-submanifold Σ ⊂ Rn is a manifold, if equipped
with the subspace topology inherited from Rn and the atlas of charts given
as follows:
S If Ψα : Uα → Vα are a collection of submanifold
charts for Σ
with α Uα ⊃ Σ, then we take A = {ϕα := π ◦ Ψα Σ∩Uα }, where π is the
projection from Rl × {0} to Rk .

Proof. The given maps are homeomorphisms to their images: Ψα is


a diffeomorphism, so the restriction to Σ ∩ Uα is a homeomorphism to its
image which is a subset of Rk × {0}. π is a homeomorphism from Rk × {0}
to Rk since it is smooth and has a smooth right-inverse π −1 (x) = (x, 0).
Therefore ϕα is a homeomorphism to its image.
Next we check smoothness: Note that ϕ−1 −1 −1

α = Ψα (Rk ×{0})∩Vα ◦π , which is
smooth as a map from Rk into Rn . Therefore the transition map ϕβ ◦ ϕ−1 α =
π ◦ Ψβ ◦ Ψ−1
α ◦ π −1 is a smooth map into Rk . Since this is also true with α

and β exchanged, the inverse map is also smooth and the transition map is
therefore a diffeomorphism. 

A large family of examples is provided by the following:


Example 2.3.3. Graphs of smooth functions are submanifolds: Let A ⊂ Rk
be an open set, and u : A → Rn−k a smooth map. Then the graph of u is
the set
Graph(u) = {(x, u(x)) : x ∈ A} ⊂ Rk × Rn−k ' Rn .
To show that Graph(u) is a submanifold of dimension k in Rn , consider the
following map Ψ : A × Rn−k → A × Rn−k :
Ψ(x, y) = (x, y − u(x)).
This map is certainly smooth (since u is smooth), and the inverse map is
given explicitly by Ψ−1 (x, y) = (x, y +u(x)) which is also smooth. Therefore
Ψ is a diffeomorphism from A × Rn−k to itself, and we have
(x, y) ∈ Graph(u) ⇐⇒ x ∈ A, y = u(x)
⇐⇒ x ∈ A, Ψ(x, y) = (x, y − u(x)) = 0
⇐⇒ Ψ(x, y) ∈ (Rk × {0}) ∩ (A × Rn−k ).
Thus Ψ is a submanifold chart for Graph(u) covering the entire set, and
Graph(u) is a k-dimensional submanifold.

More generally, essentially the same argument show that a set which can be
written as a graph in a neighbourhood of each point is a submanifold.
2.3. SUBMANIFOLDS 11

2.3.1. Equivalent descriptions of submanifolds. The definition of


submanifolds is a very natural one, but in some circumstances it can be
difficult to work with. Depending on the context it can be useful to have
other alternative ways of defining or working with submanifolds, and some
of these are provided by the following result:

Proposition 2.3.4. Let Σ be a subset of Rn . The following are equivalent:

(1). Σ is a k-dimensional submanifold;


(2). Σ is locally the graph of a smooth function. That is, for every x ∈ Σ
there is a neighbourhood U of x in Rn , a linear injection ι : Rk →
Rn and a complementary linear injection ι⊥ : Rn−k → Rn , an open
subset A ⊂ Rk , and a smooth map u : A → Rn−k such that

Σ ∩ U = {ι(z) + ι⊥ (u(z)) : z ∈ A}.

(3). Σ is locally the level set of a submersion. That is, for every x ∈ Σ
there is a neighbourhood U of x in Rn and a submersion G : U →
Rn−k such that Σ ∩ U = G−1 (0) = {y ∈ U : G(y) = 0}.
(4). Σ is locally the image of an embedding. That is, for every x ∈ Σ
there exists a neighbourhood U of x in Rn , an open set A ⊂ Rk ,
and a smooth embedding F : A → U such that Σ ∩ U = F (A).

Remark. In part (2) the choice of injections ι and ι⊥ amount to choosing


axes over which to write the submanifold as a graph. In fact one can choose
these to be coordinate inclusions ι(x1 , · · · , xk ) = (x1 , · · · , xk , 0, · · · , 0) and
ι⊥ (y 1 , · · · , y n−k ) = (0, · · · , 0, y 1 , · · · , y n−k ), after a suitable re-ordering of
the coordinates on Rn — that is, we can take Σ to be locally the graph over
some k-dimensional coordinate plane in Rn .

Proof. Some implications are very easy: The implication (2) =⇒ (1) is
essentially that given in Example 2.3.3; the implication
 (1) =⇒ (3) is proved
H(x)
by writing a submanifold chart Ψ in the form Ψ(x) = ∈ Rk ×Rn−k ,
G(x)
and then checking that G is a submersion with zero set equal to Σ ∩ U ; the
implication (1) =⇒ (4) is similarly proved by defining F to be the restriction
of Ψ−1 to Rk ' Rk × {0} ⊂ Rn .
To complete the argument we will prove (3) =⇒ (1) and (4) =⇒ (2).
(3) =⇒ (1): Let G : U → Rn−k be a submersion with Σ ∩ U = G−1 (0), and
write the components of G as G1 , · · · , Gn−k . Then for any fixed x ∈ Σ ∩ U ,
DG|x is surjective, and so has (n − k) linearly independent columns. Re-
order the coordinates on Rn to make these the last (n−k) columns, and note
that it follows that the last (n−k) columns of DG|x form an (n−k)×(n−k)
12 2. SUBMANIFOLDS

invertible matrix. Now define


   
x1 x1
 ..   .. 
 .  
    . 

 xk   x k 
Ψ 
 x k+1
 = 
  G (x , · · · , xn )
1 1
.

..
   
 ..   
 .   . 
xn Gn−k (x1 , · · · , xn )
Computing the derivative at x, we find
 
1 ··· 0 0 ··· 0
 .. .. .. .. .. .. 

 . . . . . . 

 0 ··· 1 0 ··· 0 
DΨ|x =  ∂1 G1
.
··· ∂ k G1 ∂k+1 G1 ··· ∂n G 1 
.. .. .. ..
 
 .. .. 
 . . . . . . 
∂1 Gn−k · · · ∂k Gn−k ∂k+1 Gn−k · · · ∂n Gn−k
In this matrix, the upper left (k × k) block is the identity matrix which is
invertible; and the lower right (n − k) × (n − k) block is the last (n − k)
columns of DG|x , which is also invertible by assumption. It follows that
DΨ|x is invertible, and therefore (by the inverse function theorem) Ψ is a
diffeomorphism from a sufficiently small neighbourhood Ũ of x to an open
set Ṽ in Rn . Furthermore we have y ∈ Σ ∩ Ũ ⇐⇒ G(y) = 0 and y ∈
Ũ ⇐⇒ Ψ(y) ∈ (Rk × {0}) ∩ Ṽ . Thus Ψ is a submanifold chart for Σ, and
since x is arbitrary Σ is a k-dimensional submanifold in Rn .
(4) =⇒ (2): Suppose F : A ⊂ Rk → Rn is an embedding with F (A) = Σ∩U
for some open set U in Rn , and let z ∈ A. Then DF |x is injective, so k of
the rows of DF |x are linearly independent. Re-ordering the coordinates on
Rn , we can assume that the first
 k rows
 of DF |x are linearly independent.
F1
For convenience we write F = ∈ Rk × Rn−k , so we have that DF1 |x
F2
is a k × k invertible matrix.
We augment F to make a map from A × Rn−k ⊂ Rk × Rn−k ' Rn to Rn , by
taking    
y F1 (y)
F̂ = .
z F2 (y) + z
The derivative of F̂ is then given by
 
DF1 |x 0
DF̂ |x =
DF2 |x In−k
where In−k is the (n − k) × (n − k) identity matrix. It follows that DF̂ |x is
invertible, and therefore that F̂ is a diffeomorphism from a neighbourhood
A × B of (x, 0) in Rk × Rn−k to a neighbourhood Ũ of F (x) in Rn . F̂ maps
2.3. SUBMANIFOLDS 13

A × {0} to Σ ∩ Ũ , so F̂ −1 is a submanifold chart for Σ, and since x = F (z)


is arbitrary, Σ is a k-dimensional submanifold in Rn . 

2.3.2. Examples. Depending on the context, one or another of the


equivalent descriptions of submanifolds given by Proposition 2.3.4 may be
more convenient. In particular, it is often the case that a submanifold
is naturally presented as the solution of a set of equations, so that the
description as a level set of a submersion is convenient. Here are some
examples:
Example 2.3.5. Show that the sphere S n = {x : |x| = 1} in Rn+1 is an
n-dimensional submanifold.

To do this, we will show that the smooth map G : x 7→ |x|2 − 1 is a


submersion on an open set containing S n . Since G is a map to R, in order
to show that DG|x is surjective, we need only show that it does not vanish:
If we can show DG|x (v) = c 6= 0 for some v, then we have DG|x (λv/c) = λ
for each λ ∈ R, so DG|x is surjective. We choose to compute DG|x in the
radial direction:
d d
(1 + s)2 |x|2 − 1 s=0 = 2|x|2 .

DG|x (x) = (G(x + sx)) s=0 =
ds ds
This is not zero provided x 6= 0, and therefore G is a submersion on the
open set U = Rn+1 \ {0} containing S n .
Example 2.3.6. Show that the set SL(n) of n × n matrices with unit de-
terminant is a submanifold in the space Mn of n × n matrices.

As in the previous example, the set SL(n) is defined by the vanishing of a


2
single smooth function G(M ) = det M − 1 on the vector space Mn ' Rn .
We need to show that the derivative of G does not vanish on some open set
containing SL(n). We use the same trick as in the last example, and choose
a convenient (radial) direction in which to differentiate:
d
DG|M (M ) = (det(M + sM ) − 1| s=0
ds
d
= (det((1 + s)M ) − 1| s=0
ds
d
((1 + s)n det M − 1| s=0

=
ds
= n det M.
Therefore G is a submersion on the open set GL(n) = {M : det M 6=
0} ⊂ Mn [Note that the latter is open since it is the inverse image by the
(continuous) determinant function of the open set R \ {0}].
Example 2.3.7. Show that the orthogonal group O(n) = {M ∈ Mn :
M T M = In } is a submanifold.
14 2. SUBMANIFOLDS

This one is a little more complicated: O(n) is indeed defined as the set of
matrices where a certain equation is satisfied, so it makes sense to define
G(M ) = M T M − In and try to show that G is a submersion. The map G
is matrix valued, but it is important to observe that the image actually lies
in a lower-dimensional subspace of Mn , namely the linear subspace Sym(n)
consisting of the symmetric n × n matrices (since (M T M )T = M T (M T )T =
M T M ). We will show that DG|M is surjective as a map to Sym(n) for any
M ∈ GL(n):
We need to show that for any symmetric matrix C, we can find a direction
A ∈ Mn such that DG|M (A) = C. We compute:
d
DG|M (A) = (G(M + sA)) s=0
ds
d
(M T + sAT )(M + sA) − I s=0

=
ds
= M T A + AT M.
Given C, if M ∈ GL(n) we can choose A = 21 (M T )−1 C. Then we have
M T A = 21 M T (M T )−1 C = 12 C, and AT M = 12 C T M −1 M = 21 C, so DG|M (A) =
C as required. Therefore DG|M is surjective, and G is a submersion to the
space Sym(n) which as dimension n(n+1) 2 . It follows that O(n) is a subman-
2 n(n+1) n(n−1)
ifold of Mn of dimension n − 2 = 2 .
CHAPTER 3

Examples of manifolds, and smooth maps

Contents
3.1. Introduction 15
3.2. Manifold construction 15
3.2.1. Example: Real projective spaces 15
3.2.2. The smooth chart lemma 17
3.2.3. Further examples 18
3.3. Smooth maps 19
3.3.1. Smooth maps and submanifolds 20
3.3.2. Examples 20

3.1. Introduction

Last lecture we discussed submanifolds, which are manifolds which come


equipped with an ambient Euclidean space. Today I want to discuss some
further examples of manifolds which arise naturally in other ways, and then
move on to define notions of smoothness for maps between manifolds.

3.2. Manifold construction

Let us begin with an important example:

3.2.1. Example: Real projective spaces. The real projective space


is the space of lines through the origin in (n + 1)-dimensional Euclidean
space. More formally, we define an equivalence relation on the non-zero
point in Rn+1 by calling two points equivalent if they lie on the same line
through the origin, or equivalently if they are related by scalar multiplication
by a non-zero real number:
x ∼ y ⇐⇒ x = λy for some λ ∈ R \ {0}.
The set of equivalence classes under this equivalence relation is the real
projective space RPn . We write
Rn+1 \ {0}
RPn = .
R \ {0}
15
16 3. EXAMPLES OF MANIFOLDS, AND SMOOTH MAPS

Given x ∈ Rn+1 \ {0} we denote by [x] the equivalence class of x.


In order to make RPn into a manifold, we need to define charts. First set
U = {[x] : x = (x1 , · · · , xn+1 ), xn+1 6= 0}. This is a well-defined subset
of RPn , since the condition is independent of the choice of representative
from
 a given equivalence
 class [x]. On U we define ϕ([x1 , · · · , xn+1 ]) =
x1 xn n
xn+1 , · · · , xn+1 . Note that ϕ is a bijection from U to R , since we can
write down the inverse map ϕ−1 (z1 , · · · , zn ) = [z1 , · · · , zn , 1].
More generally, for any L ∈ GL(n + 1), there is a bijection from RPn to
itself given by
L([x]) := [L(x)].
This is well-defined, since [L(λx)] = [λL(x)] = [L(x)], and has inverse given
by L−1 ([x]) = [L−1 (x)]. For each such L we set UL = L−1 U = {L−1 ([x]) :
[x] ∈ U } = {[L−1 (x)] : [x] ∈ U } = {[z] : (Lz)n+1 6= 0}. On this set we
define ϕL := ϕ ◦ L.
We will construct a smooth structure on RPn in which the collection of maps
{ϕL : UL → Rn }L∈GL(n+1) form a smooth atlas of charts1. Let us check the
condition on the transition maps: If P, L ∈ GL(n + 1) then
   
−1 z n
UL = L : z∈R ,
1
and
n+1
 −1  n+1
( ) ( )
X X
UP = [x] : i
Pn+1 xi 6= 0 = L x : (P L−1 )in+1 xi 6= 0 .
i=1 i=1
Therefore
n
(   X )
z
UP ∩ UL = L−1 : (P L−1 )in+1 zi + (P L−1 )n+1
n+1 6= 0 .
1
i=1
The image under ϕL is then given by
n
( )
X
n −1 i −1 n+1
ϕL (UL ∩ UP ) = z ∈ R : (P L )n+1 zi + (P L )n+1 6= 0 .
i=1
This set is open, since it is the set where a certain continuous function is
non-zero. On this set kth component of the transition map can be computed
as follows:
  
−1
k k −1 z
ϕP ◦ ϕL (z) = ϕ PL
1
Pn
(P L )i zi + (P L−1 )n+1
−1
= Pn i=1 −1 i k k
−1 )n+1
.
i=1 (P L ) z
n+1 i + (P L n+1

1In fact multiplying L by a scalar does not affect the transformation of RPn , so we
could restrict to L ∈ SL(n + 1) with no loss of generality.
3.2. MANIFOLD CONSTRUCTION 17

Since the denominator does not vanish on the domain of definition, the
transition map is smooth, and hence (since we can interchange P and L) is
a diffeomorphism.
Clearly every point of RPn is in the domain of some chart ϕL . It remains
to check that the maps are homeomorphisms, and for this we need to define
a topology on RPn . There is a natural way to do this (called the quotient
topology), but we will take a different approach and use the maps above to
construct a topology on M — this is a useful way to construct manifolds,
particularly in situations like this where the topology is not given in advance.

3.2.2. The smooth chart lemma.


Lemma 3.2.1. Let M be a set, and A = {ϕα : Uα → Vα }α∈A a collection
of maps, where each ϕα is a map from a subset Uα to an open subset Vα of
Rn . Suppose that the following conditions hold:

(i). Each ϕα is a bijection from Uα to Vα ;


(ii). For each α, β ∈ A the set ϕα (Uα ∩ Uβ ) is open in Rn ;
(iii). For each α, β ∈ A, the transition map ϕβ ◦ ϕ−1 α : ϕα (Uα ∩ Uβ ) →
ϕβ (Uα ∩ Uβ ) is smooth; S
(iv). There is a countable subset A0 ⊂ A such that α∈A0 Uα = M ;
(v). For any distinct points p and q in M , either there exists α ∈ A with
p, q ∈ Uα , or there exist α, β ∈ A with Uα ∩ Uβ = ∅ and p ∈ Uα ,
q ∈ Uβ .

Then there is a unique differentiable structure on M in which A is a smooth


atlas.

Proof. We need to construct a topology on M which is Hausdorff and


second countable (or a metric which is separable) with respect to which the
sets Uα are open and the maps ϕα are homeomorphisms.
A topological proof: Let B = {ϕ−1 (A) : A ⊂ Vα open in Rn , α ∈ A},
the collection of all inverse image of open sets by the maps ϕα . We will
show that B is a basis for a topology on M — precisely, that the collection
T of sets given by arbitrary unions of sets in B is a topology. T is closed
under arbitrary unions. It remains to checkSthat T is closed under finite
intersections: An element of T has the form α∈A ϕ−1 α (Wα ) where Wα is an
open subset of Vα for each α ∈ A. The intersection of two such sets then
has the form [
−1
ϕ−1
α (Wα ) ∩ ϕβ (Zβ )
α,β∈A
where Wα and Zα are open in Vα for each α. But now we observe that
−1 −1
ϕ−1 −1
α (Wα ) ∩ ϕβ (Zβ ) = ϕα (Wα ∩ ϕα (Uα ∩ Uβ )) ∩ ϕβ (Zβ ).
18 3. EXAMPLES OF MANIFOLDS, AND SMOOTH MAPS

Now Wα ∩ ϕα (Uα ∩ Uβ ) is open in Vα by (ii). By (iii) we have


−1
ϕ−1 (ϕβ ◦ ϕ−1

α (Wα ∩ ϕα (Uα ∩ Uβ )) = ϕβ α )(Wα ∩ ϕα (Uα ∩ Uβ ))

and the set (ϕβ ◦ ϕ−1


α )(Wα ∩ ϕα (Uα ∩ Uβ ) is open in Vβ . Therefore we have
−1 −1
ϕ−1 (ϕβ ◦ ϕ−1

α (Wα ) ∩ ϕβ (Zβ ) = ϕβ α )(Wα ∩ ϕα (Uα ∩ Uβ ) ∩ Zβ ) ∈ B.

Therefore the intersection of any two elements of T is a union of elements


of B, hence an element of T . This proves that T is a topology. With
respect to this topology ϕα and ϕ−1
α take open sets to open sets, so ϕα is a
homeomorphism.
It remains to check that the topology is Hausdorff and second countable.
The Hausdorff condition is immediate from (v): If p and q are distinct
points in M , then either there is α ∈ A with p, q ∈ Uα — in which case we
can certainly find disjoint small open balls in Vα about ϕα (p) and ϕα (q),
and then the inverse images of these are disjoint open sets in M separating
p and q — or there are α and β with Uα ∩ Uβ disjoint, and p ∈ Uα and
q ∈ Uβ — so Uα and Uβ are disjoint open sets separating p and q.
To show the topology is second countable we use (iv): Let B0 be the collec-
tion of sets of the form ϕ−1
α (B) where B ⊂ Vα is a ball of rational radius
about a point with rational coordinates, and α ∈ A0 . Then B0 is a countable
set, and every element of B is a union of elements of B0 . That is, B0 is a
countable basis for the topology of M , so the topology is second countable.
Exercise 3.2.2. A metric space proof would involve putting a distance
function on M with respect to which the charts are homeomorphisms. How
could this be done?

Now let us apply the smooth chart lemma to show that our charts for RPn
can form the atlas for a smooth structure: We have already checked (i) and
(ii) and (iii). Part (iv) is clear, since we can cover RPn with the finitely many
charts corresponding to the maps L which exchange the ith and (n + 1)st
coordinates on Rn+1 , for i = 1, · · · , n + 1. Finally, (v) holds since for any
nonzero x and y with [x] 6= [y] we can choose a basis {ẽ1 , · · · , ẽn+1 } with
ẽn+1 = x+y and ẽn = x−y, and choose L to be the GL(n+1) transformation
which takes ei to ẽi for each i. Then x = 21 ẽn + 12 ẽn+1 and y = − 12 ẽn + 21 ẽn+1 ,
so both have non-zero (n + 1)st component under L and so lie in UL .

3.2.3. Further examples. Some further important examples related


to the previous one are the following:
Example 3.2.3 (Stereographic atlas for the sphere). We consider the sphere
S n as a submanifold of Rn+1 ' Rn × R. The stereographic projection from
3.3. SMOOTH MAPS 19

the north pole N = (0, 1) is the map which takes (x, y) ∈ S n \ {N } to the
point in Rn × {0} along the same line through N . Show that the two maps
given by stereographic projections from the north and south poles form a
smooth atlas for S n .
Example 3.2.4 (Complex projective spaces). The non-zero complex num-
bers act by scalar multiplication on the non-zero elements of Cn+1 \ {0}.
n+1 \{0}
The quotient by this action is the complex projective space CPn = C C\{0} .
Show that this can be given the structure of a smooth manifold of dimension
2n by taking the charts
 
z1 zi−1 zi+1 zn+1
ϕi ([z1 , · · · , zn+1 ]) = ,··· , , ,··· , ∈ Cn ' R2n .
zi zi zi zi
Example 3.2.5 (Grassmannians). The Grassmannian Gk,n is the space of
k-dimensional linear subspaces of Rn . Charts for Gk,n can be defined as
follows: For any basis O = {e1 , · · · , en } for Rn , define a map FO from
Mk,n−k to Gk,n by
 
 n−k
X j 
FO (M ) = span ei + Mi ek+j : i = 1, · · · , k .
 
j=1

Show that the inverses of these maps form a collection of such charts defining
the structure of a differentiable manifold of dimension k(n − k) on Gk,n .

3.3. Smooth maps

Next we formalise the definition of smoothness which motivated our defini-


tion of compatibility for charts:
Definition 3.3.1. Let M be a smooth manifold. Then a function f : M →
R is smooth (and we write f ∈ C ∞ (M )) if for every smooth chart ϕ : U →
V ⊂ Rn for M , the composition f ◦ ϕ−1 is a smooth function on V .

More generally, we can make sense of smoothness for maps from one manifold
to another:
Definition 3.3.2. If M m and N n are smooth manifolds, then a map f :
M → N is smooth (and we write f ∈ C ∞ (M, N )) if for every chart ϕ for
M and η for N , the composition η ◦ f ◦ ϕ−1 is smooth (as a function from
an open subset of Rm to Rn ).

Note that it is sufficient to check this for one chart about each point of M ,
since changes of chart are given by transition maps which are smooth.
Proposition 3.3.3. Any composition of smooth maps between manifolds is
smooth.
20 3. EXAMPLES OF MANIFOLDS, AND SMOOTH MAPS

Proof. Suppose f ∈ C ∞ (M, N ) and g ∈ C ∞ (N, K). To check that


g ◦ f ∈ C ∞ (M, K), we let x ∈ M be arbitrary and choose a chart ϕ for M
near x, a chart η for N near f (x), and a chart ψ for K near g ◦ f (x). Then
ψ ◦ (g ◦ f ) ◦ ϕ−1 = (ψ ◦ g ◦ η −1 ) ◦ (η ◦ f ◦ ϕ−1 )
in a a neighbourhood of x. The map ψ ◦g ◦η −1 is smooth since g is smooth as
a map between manifolds, and the map η◦f ◦ϕ−1 is smooth since f is smooth
as a map between manifolds. Therefore by the chain rule, the composition
is smooth and hence g ◦ f is smooth as a map between manifolds. 

3.3.1. Smooth maps and submanifolds. An important special case


to consider is that of submanifolds. There are two useful results which apply
frequently in this case:
Proposition 3.3.4. Suppose that M is a submanifold in Rk , and f : U →
N is a smooth map, where U is an open subset of Rk containing M . Then
the restriction f |M of f to M is in C ∞ (M, N ).

Proof. Let Ψ : A ⊂ Rk → Rk be a submanifold chart for M near


a point x ∈ M , so that ϕ = π ◦ Ψ|M ∩U is a chart for M , where π is the
projection from Rk to Rm . Also let η be a chart for N neat f (x). Then
η ◦ f |M ◦ ϕ−1 = η ◦ f ◦ Ψ−1 ◦ π −1
where π −1 (x1 , · · · , xm ) = (x1 , · · · , xm , 0, · · · , 0). η ◦ f is smooth since f is
smooth from Rk into N , so this is a composition of smooth maps, hence
smooth. 
Proposition 3.3.5. Suppose that M is a manifold, and N is a submanifold
of Rk . If f ∈ C ∞ (M, Rk ) with f (M ) ⊂ N , then f ∈ C ∞ (M, N ).

Proof. Let Ψ be a submanifold chart for N near f (x) ∈ N , so that


π◦Ψ|N is a chart for N . Since f (M ) ⊂ M and f is continuous, η◦f = π◦Ψ◦f
is well-defined on a neighbourhood of x ∈ M . Also let ϕ be a chart for M
near x. Then η ◦ f ◦ ϕ−1 = π ◦ (Ψ ◦ f ◦ ϕ−1 ), and f ◦ ϕ−1 is smooth since f
is smooth as a map into Rk . 
Exercise 3.3.6. Show that any orthogonal transformation L ∈ O(n + 1)
restricts to the sphere S n to give a diffeomorphism of S n .
Exercise 3.3.7. Show that the inversion map from GL(n) to GL(n) is a
diffeomorphism, and that it restricts to SL(n) to give a diffeomorphism of
SL(n) and to O(n) to give a diffeomorphism of O(n).

3.3.2. Examples.
Example 3.3.8. Show that the projection x 7→ [x] from Rn+1 \ {0} to RPn
is smooth.
CHAPTER 4

Tangent vectors

Contents
4.1. Introduction 21
4.2. Examples and constructions 21
4.2.1. Open subsets 21
4.2.2. Product manifolds 21
4.3. Tangent vectors and covectors 22
4.3.1. The derivative for functions between Euclidean spaces 22
4.3.2. Smooth curves through a point 23
4.3.3. Germs of smooth functions 23
4.3.4. The directional derivative pairing 23
4.3.5. Tangent vectors 24
4.3.6. Covectors 26
4.3.7. Coordinate tangent vectors and coordinate differentials 27
4.3.8. Tangent vectors as derivations 27
4.4. The derivative of a smooth map between manifolds 29

4.1. Introduction

We will first discuss some natural examples and constructions on manifolds,


and then introduce the notion of tangent vectors and covectors.

4.2. Examples and constructions

4.2.1. Open subsets. Let M be a manifold, with atlas given by {ϕα :


Uα → Vα }α∈A . Then any open subset U of M is also a manifold, with atlas
given by {ϕα |Uα ∩U }α∈A . This is clear, since the transition maps for these
are simply restrictions of the transition maps for the original atlas.
This gives many examples of manifolds, and shows that in general they can
be very ‘wild’ — any open subset of a Euclidean space in a manifold, for
example.

4.2.2. Product manifolds. Let M m and N n be smooth manifolds,


with atlases {ϕα : Uα → Vα }α∈A and {ηβ : Wβ → Zβ }β∈B respectively.
21
22 4. TANGENT VECTORS

Then we can make the product M × N into a smooth manifold of dimension


m + n, by taking the atlas {(ϕα , ηβ ) : α ∈ A, β ∈ B}, where
(ϕα , ηβ ) : Uα × Wβ → V α × Zβ ⊂ Rm × Rn ' Rn+m
is given by
(ϕα , ηβ )(x, y) := (ϕα (x), ηβ (y)).
Exercise 4.2.1. Check the above statement by finding the transition maps
for the above atlas.
Exercise 4.2.2. If M and N are manifolds and x ∈ N , show that M ×{x} =
{(y, x) : y ∈ M } is a submanifold of M × N of dimension equal to that of
M.
Exercise 4.2.3. If M is a manifold, then the diagonal ∆ = {(x, x) : x ∈ M }
is a submanifold of M × M with dimension equal to that of M .
Exercise 4.2.4. Show that the projection π : M × N → M given by
π(x, y) = x is smooth, as is the inclusion ιy : M → M × N given by
x 7→ (x, y) for y ∈ N fixed.

4.3. Tangent vectors and covectors

In this section we develop the notion of tangent vectors, which we will use
to make sense of the derivative of a smooth map.

4.3.1. The derivative for functions between Euclidean spaces.


It is useful to first recall a convenient geometric meaning of the derivative
for a smooth map between Euclidean spaces: Given that a function f : U ⊂
Rm → Rn is smooth, we can compute the action of the derivative at a point
x ∈ U in any direction v ∈ Rm as a ‘directional derivative’: That is, we
consider moving along the straight line from x in direction v (i.e. along the
curve s 7→ x + sv) and differentiating at s = 0:
d
Df |x (v) = (f (x + sv)) s=0 .
ds
That is, the derivative tells us the direction that the image moves as we
move in a given direction in the domain from x.
One can formalise this a little more: Given x ∈ U , we think of a direction
at x to be a pair (x, v) where v ∈ Rm . Then Df |x is a linear map from the
space of directions at x to the space of directions at the image point f (x),
by the map
Df : (x, v) 7→ (f (x), Df |x (v)).

To make sense of this on a manifold, we need to make sense of the space


of directions at a given point x ∈ M . There are many ways to define this
4.3. TANGENT VECTORS AND COVECTORS 23

space, but we will adopt a definition which takes smooth curves in M passing
through x as the starting point:

4.3.2. Smooth curves through a point. By a smooth curve through


x we mean a smooth map γ from an open neighbourhood I of the origin to
M , such that γ(0) = x. The set of all such curves will be denoted by Cx .
We will make sense of what it means for two such curves to have ‘the same
tangent vector’, by considering the derivatives of smooth functions along the
curve.

4.3.3. Germs of smooth functions. By a smooth function defined


near x, we mean a function f ∈ C ∞ (U ) where U is an open neighbourhood
of x. In what follows we will only be concerned with the function on an
arbitrarily small neighbourhood of U , so it is convenient to consider two
functions as being ‘the same’ if they agree on some neighbourhood of x.
More formally, we define an equivalence relation on the space of locally
defined smooth functions near x, by setting f ∼ g if and only if there exists
an open neighbourhood W of x on which f and g are both defined and are
equal. An equivalence class of locally defined functions is referred to as a
smooth function germ near x, and the space of smooth function germs near
x is denoted by Cx∞ (M ).

The space Cx∞ (M ) has a well-defined vector space structure: If f and f˜


are locally defined functions representing the same germ, and g and g̃ are
locally defined functions near x representing another germ, then we can
unambiguously define the germ af + bg for real numbers a and b, since the
functions af + bg and af˜+ bg̃ coincide on a sufficiently small neighbourhood
of x, and so represent the same germ.
Note that we could also define ‘germs of smooth curves’ by taking two curves
in Cx to be ‘the same’ if they coincide on some neighbourhood of the origin.

4.3.4. The directional derivative pairing. There is a natural pair-


ing between smooth function germs and curves through x, given by comput-
ing the directional derivative of the function f along the curve γ:
d
γ ∈ Cx , f ∈ Cx∞ (M ) 7→ (f (γ(s))|s=0 ∈ R.
ds
This is of course independent of the choice of representative of the function
germ. We will use this below in our definition of the space of tangent vectors.
We observe that each curve γ ∈ Cx induces a linear functional on the vector
space Cx∞ (M ), since
d d d d
((af + bg)(γ(s))|s=0 = a (f (γ(s))|s=0 + b (g(γ(s))|s=0
ds ds ds ds
for any f, g ∈ Cx∞ (M ) and a, b ∈ R.
24 4. TANGENT VECTORS

4.3.5. Tangent vectors. Now we define an equivalence relation on the


space Cx of smooth curves in M through x: We say that two such curves γ1
and γ2 have the same direction if
d d
(f ◦ γ1 (s)) s=0 = (f ◦ γ2 (s)) s=0
ds ds
for every smooth function germ f at x. We denote the equivalence class of
γ the tangent vector of γ and denote it by γ 0 . The space of such tangent
vectors is denoted by Tx M , so we have Tx M = {γ 0 : γ ∈ Cx }.
We note that there is still a well-defined pairing between Tx M and Cx∞ (M ),
given by
d
γ 0 (f ) =

(f ◦ γ(s)) s=0
ds
for any γ with tangent vector γ 0 — this is the derivative of f in direction
γ 0 . For each γ 0 this defines a linear functional on Cx∞ (M ), which is different
for different tangent vectors, and so gives an injective map from Tx M into
the vector space of linear functionals on Cx∞ (M ).
The equivalence relation defining γ 0 looks strange, but we can re-interpret
it in a simpler way once we have chosen a chart near x:
Proposition 4.3.1. Let ϕ be a chart for M about x. Then two curves γ1
and γ2 in Cx have the same direction if and only their images under ϕ have
the same tangent vector:
d d
(ϕ ◦ γ2 (s)) s=0 ∈ Rn .

(ϕ ◦ γ1 (s)) s=0 =
ds ds
d

Furthermore the map γ 7→ ds (ϕ ◦ γ(s)) s=0 induces a bijection from Tx M
to Rn .

Proof. First suppose that γ1 and γ2 have the same direction. If ϕ :


U ⊂ M to Rn is a smooth chart with components ϕ1 , · · · , ϕn , then we
observe that each function ϕi is smooth near x: To check this we need to
choose a chart (let’s choose ϕ!) and show that ϕi ◦ ϕ−1 is a smooth function.
But this is immediate, since
ϕi ◦ ϕ−1 (x1 , · · · , xn ) = xi
which is certainly a smooth function on Rn . Therefore, since γ1 and γ2 have
the same direction, we have
d d
ϕi (γ1 (s)) s=0 = ϕi (γ2 (s)) s=0
 
ds ds
for i = 1, · · · , n, and therefore
d d
(ϕ(γ1 (s))) s=0 = (ϕ(γ2 (s))) s=0 .
ds ds
4.3. TANGENT VECTORS AND COVECTORS 25

d d

Conversely, if ds (ϕ ◦ γ1 (s)) s=0 = ds (ϕ ◦ γ2 (s)) s=0 , and f is any locally
defined smooth function near x, then
d d
(f ◦ ϕ−1 ) ◦ (ϕ ◦ γ1 ) s=0

(f ◦ γ1 (s)) s=0 =
ds ds  
−1
d
= D(f ◦ ϕ ) ϕ(x) (ϕ ◦ γ1 (s)) s=0

ds
 
d
= D(f ◦ ϕ−1 ) ϕ(x)

(ϕ ◦ γ2 (s)) s=0
ds
d
(f ◦ ϕ−1 ) ◦ (ϕ ◦ γ2 ) s=0

=
ds
d
= (f ◦ γ2 (s)) s=0 .
ds

Since f ∈ Cx (M ) is arbitrary, we conclude that γ1 and γ2 have the same
direction.
d

This shows that the map from γ to ds (ϕ ◦ γ(s)) s=0 is well-defined on Tx M
(since it gives the same result for any two curves with the same direction)
and is injective (since it gives the same result only if the two curves have
the same direction). It remains to show that the map is surjective, and this
is immediate since for any v ∈ Rn the smooth curve
γ(s) = ϕ−1 (ϕ(s) + sv)
is sent by this map to
d d d
ϕ ◦ ϕ−1 (ϕ(s) + sv) s=0 =

(ϕ ◦ γ(s)) s=0 = (ϕ(s) + sv) s=0 = v.
ds ds ds


A further piece of useful information is the following:


Proposition 4.3.2. The bijection between Tx M and Rn induced by a chart ϕ
induces a vector space structure on Tx M , which is independent of the choice
of ϕ. With respect to this vector space structure the inclusion of Tx M into
the vector space of linear functionals on Cx∞ (M ) is linear.

Proof. We must check that the vector space operations (scalar multi-
plication and addition) induced by above identification are unchanged if we
change from one chart ϕ to another chart η near x. To see this, consider two
charts ϕ and η about x. The correspondence above takes a vector v ∈ Rn
through the chart ϕ to the tangent vector of the curve s 7→ ϕ−1 (ϕ(x) + sv),
and this is then identified with the vector in Rn given by
d
η ◦ ϕ−1 (ϕ(x) + sv) |s=0 = D(η ◦ ϕ−1 )|ϕ(x) (v).

ds
Since this is a linear transformation of v, it respects scalar multiplication and
addition, so these are well-defined independent of the choice of chart. 
26 4. TANGENT VECTORS

4.3.6. Covectors. The constructions above also lead to a natural fur-


ther equivalence relation on the space of smooth function germs near x: We
say that to such function germs f and g have the same differential if their
pairings with any smooth curve agree:
d d
f ∼ g ⇐⇒ (f ◦ γ(s)) s=0 = (g ◦ γ(s)) s=0 for all γ ∈ Cx .
ds ds
The space of equivalence classes (which we call covectors as x) is denoted by
Tx∗ M , and inherits a vector space structure from the vector space structure
on Cx∞ (M ). The equivalence class of a function f is denoted by df and
called the differential of f at x.
Proposition 4.3.3. Let ϕ be a local chart for M near x. Then the map
from Cx∞ (M ) to (Rn )∗ given by
f 7→ D(f ◦ ϕ−1 )|ϕ(x)
induces a linear isomorphism between Tx∗ M and (Rn )∗ .

Proof. We check that the map is well-defined: If f ∼ g then for any


vector v ∈ Rn ,
d
D(f ◦ ϕ−1 )|ϕ(x) (v) = f ◦ ϕ−1 (ϕ(x) + sv) s=0

ds
d
= (f ◦ γ(s)) s=0
ds
−1
where γ(s) = ϕ (ϕ(x) + sv) ∈ Cx . Since f ∼ g, this is unchanged if f
is replaced by g, so the resulting linear map on Rn depends only on the
equivalence class of f in Tx∗ M . The map is clearly linear on Cx∞ (M ), hence
also on Tx∗ M .
The map to Rn is injective: If D(f ◦ ϕ−1 )|ϕ(x) = D(g ◦ ϕ−1 )|ϕ(x) then for
any curve γ ∈ Cx we have
d d
f ◦ ϕ−1 ◦ ϕ ◦ γ(s) s=0

(f ◦ γ(s)) s=0 =
ds ds  
−1
d
= D(f ◦ ϕ ) ϕ(x) (ϕ ◦ γ(s)) s=0

ds
 
−1
d
= D(g ◦ ϕ ) ϕ(x) (ϕ ◦ γ(s)) s=0

ds
d
= (g ◦ γ(s)) s=0
ds
so that df = dg in Tx∗ M . Finally, the map is surjective since for any map
L ∈ (Rn )∗ we can define f (y) = L ◦ ϕ(y), in which case
d d
D(f ◦ϕ−1 )|ϕ(x) (v) = L ◦ ϕ ◦ ϕ−1 (ϕ(x) + sv) s=0 =

(L(ϕ(x) + sv)) s=0 = L(v)
ds ds
for every v ∈ R , and so L is in the image of the map from Tx∗ M .
n 
4.3. TANGENT VECTORS AND COVECTORS 27

By the definition, there is a well-defined pairing between smooth curves


through x and covectors at x, and this descends to a well-defined pairing
between tangent vectors at x and covectors at x, given by
d
df (v) = (f ◦ γ(s)) s=0
ds
where γ is any smooth curve through x with tangent vector γ 0 = v. Each
tangent vector acts as a linear functional on Tx∗ M , and this gives a linear
isomorphism between Tx M and the dual space of Tx∗ M (hence also between
Tx∗ M and the dual space of Tx M ). This follows from the next result about
coordinate bases:

4.3.7. Coordinate tangent vectors and coordinate differentials.


The identifications above between Tx M and Rn (and between Tx∗ M and
(Rn )∗ ) provide useful bases for Tx M and Tx∗ M (depending of course on the
choice of local chart ϕ): We define ∂i to be the tangent vector correspond-
ing to the standard ith basis element ei of Rn : This is the tangent vector
corresponding to the smooth curve γ(s) = ϕ−1 (ϕ(x) + sei ), and acts on a
smooth function f by ‘coordinate differentiation in the chart’:
d ∂
f ◦ ϕ−1 (ϕ(x) + sei ) s=0 = −1
 
∂i f = f ◦ ϕ ϕ(x)
.
ds ∂xi
The basis {∂1 , · · · , ∂n } is called the coordinate tangent basis for Tx M .
The corresponding dual basis for Tx∗ M also has a nice description: If we
write ϕ in components as x1 , · · · , xn , then each function xi is a smooth
function near x, and we write dxi for the corresponding differential (equiv-
alence class). This acts on a coordinate tangent vector ∂j according to
d
dxi (∂j ) = xi ◦ ϕ−1 (ϕ(x) + sej ) s=0

ds
(
d
(xi (x) + s), j = i;
= dsd i
ds (x (x)), j 6= i
(
1, j = i;
= .
0, j 6= i

It is convenient to write this using the Kronecker delta δji , which is defined
to be 1 if i = j and 0 if i 6= j.

4.3.8. Tangent vectors as derivations. There is an alternative way


of thinking about tangent vectors which is more abstract but sometimes
convenient: One can think of a tangent vector as a ‘differential operator’
which acts on functions:
28 4. TANGENT VECTORS

Definition 4.3.4. Let M be a smooth manifold, and x a point of M . A


derivation at x is an R-linear map from Cx∞ (M ) to R, taking a function f
to vf ∈ R, which satisfies the ‘product rule’
v(f g) = f (x)vg + vf g(x).

We will (temporarily) denote the space of derivations at x by Dx M (since we


will identify these with tangent vectors, we will eventually simply identify
Dx M with Tx M ). The space Dx M has a natural vector space structure,
defined by taking
(av + bw)(f ) := av(f ) + bw(f )

for all f ∈ Cx (M ).
Proposition 4.3.5. The map from Tx M to Dx M defined by γ 0 ∈ Tx M 7→
Dγ ∈ Dx M where
d
Dγ f := (f ◦ γ(s)) s=0
ds
is a well-defined linear isomorphism.

Proof. First we check that Dγ is a derivation: It is clearly R-linear on


Cx∞ (M ), and if f, g ∈ Cx∞ (M ) then we have
d
Dγ (f g) = (f ◦ γ(s)g ◦ γ(s)) s=0
ds
d d
= f (x) (g ◦ γ(s)) s=0 + g(x) (f ◦ γ(s)) s=0
ds ds
= f (x)Dγ g + g(x)Dγ f
as required. Next we check that the map is well-defined and injective on
d

Tx M : If γ and σ are in Cx , then γ ∼ σ if and only if ds (f ◦ γ(s)) s=0 =
d

ds (f ◦ σ(s)) s=0 for every s, which means Dγ (f ) = Dσ (f ) for every f , which

means Dγ = Dσ .
The next requirement is to check that the map is surjective, so that every
derivation arises as Dγ for some γ. To see this, fix v ∈ Dx M , and fix a chart
ϕ = (x1 , · · · , xn ) for M near x, so that xi ∈ Cx∞ (M ) for each i. Without
loss of generality we can assume that ϕ(x) = 0. Define v i = v(xi ), and set
γ(s) = ϕ−1 (s(v 1 , · · · , v n )). We will prove that v = Dγ .
We need two simple results:
Lemma 4.3.6. If v is a derivation at x and c is a constant function, then
v(c) = 0.

Proof. We have
v(1) = v(1) = v(1.1) = 1.v(1) + v(1).1 = 2v(1)
so v(1) = 0. For c ∈ R, R-linearity gives v(c) = v(c.1) = cv(1) = 0. 
4.4. THE DERIVATIVE OF A SMOOTH MAP BETWEEN MANIFOLDS 29

Lemma 4.3.7. If f, g ∈ Cx∞ with f (x) = g(x) = 0, then v(f g) = 0.

Proof. Immediate from the product rule. 

Note that
n
d X ∂
f ◦ ϕ−1 (s(v 1 , · · · , v n )) s=0 = v i i (f ◦ ϕ−1 )|0 .

Dγ (f ) =
ds ∂x
i=1

Let f˜ = f ◦ ϕ−1 , which is smooth in a neighbourhood of 0 in Rn . Then


applying the Taylor theorem with remainder to the function g(s) = f˜(sy)
on the interval [0, 1] gives
n n
∂ f˜ i X i j 1 ∂ 2 f˜
X Z
f˜(y) = f˜(0) + i
y + y y (1 − s) i j sy ds.
∂x 0 ∂x ∂x
i=1 i,j=1

Bringing this back onto M using the chart ϕ gives for p close to x with
ϕ(p) = y (and noting y j ◦ ϕ = xj ) this becomes
n n
X ∂ f˜ i X i j
f (p) = f (x) + x + x x g(p)
∂xi
i=1 i,j=1

where g ∈ Cx∞ (x). Note that each term in the last tsum is the product of
two smooth function xi and xj g which both vanish at x. Applying v gives
(by Lemmas 4.3.6 and 4.3.7)
n n
X ∂ f˜ i
X
i ∂f
˜
vf = v(x ) = v = Dγ (f )
∂xi ∂xi
i=1 i=1
as required. 

4.4. The derivative of a smooth map between manifolds

Now we are in a position to define the derivative of a smooth map. As in


the case of maps between Euclidean space, the derivative is a linear map
taking the tangent space at a point in the domain to the tangent space at
the image point.
Definition 4.4.1. Let M and N be smooth manifolds, and F ∈ C ∞ (M, N ).
Then DF |x is the linear map from Tx M to TF (x) N defined by
DF |x (γ 0 ) = (F ◦ γ)0
for any γ ∈ Cx .

To understand this definition, note that if γ is a smooth curve in M through


x, then F ◦ γ defines a smooth map in N through F (x). We should check
that the resulting map is well-defined, so that if γ ∼ σ then the image curves
30 4. TANGENT VECTORS

F ◦γ and F ◦σ have the same direction: This is true because if f is a smooth


function on N near F (x), then f ◦ F ∈ Cx∞ (M ) and so
d
(F ◦ σ)0 (f ) = (f ◦ F ◦ σ(s)) s=0 = σ 0 (f ◦ F )

ds
which clearly depends only on the equivalence class of σ in Tx M . The latter
expression also tells us how to think about the derivative when we interpret
a tangent vector as a derivation:
Proposition 4.4.2. If v is a derivation on Cx∞ (M ), then DF |x (v) is the
derivation on CF∞(x) (N ) defined by
(DF |x (v))(f ) = v(f ◦ F )
for any f ∈ CF∞(x) (N ).

This is immediate from the computation above and the correspondence in


Proposition 4.3.5.
Exercise 4.4.3. If F ∈ C ∞ (M m , N n ) and x ∈ M , and ϕ is a smooth chart
for M about x and η is a smooth chart for N about F (x), show that
n
DF |x (∂iϕ ) = Λji ∂jη
X

j=1

where {∂1ϕ , · · ·
, ∂m ϕ
}
is the coordinate tangent basis for Tx M arising from
the chart ϕ, , ∂nη } is the coordinate tangent basis for TF (x) N arising
{∂1η , · · ·
from the chart η, and Λ = D(η ◦ F ◦ ϕ−1 )ϕ(x) ∈ L(Rm , Rn ).
Exercise 4.4.4. If F ∈ C ∞ (M, N ) and G ∈ C ∞ (N, K) where M , N and
K are smooth manifolds, and x ∈ M , prove that
D(G ◦ F )|x = DG|F (x) ◦ DF |x .
[Hint: This can be easily using the definition of the derivative acting on
equivalence classes of curves, or using the action on derivations from Propo-
sition 4.4.2, or using the result of the previous problem.
Exercise 4.4.5. Show that any smooth chart ϕ : U → V for M is a
diffeomorphism from U ⊂ M to V ⊂ Rn .
CHAPTER 5

Partitions of unity and applications

Contents
5.1. Introduction 31
5.2. Partitions of unity 31
5.2.1. Definition 31
5.2.2. Existence 32
5.2.3. The compact case 35
5.3. Applications of partitions of unity 36
5.3.1. Extension of locally defined smooth functions 36
5.3.2. Smooth functions on submanifolds are restrictions 37
5.3.3. Exhaustion functions 38
5.4. Level sets of smooth functions 39

5.1. Introduction

The next topic we will discuss is partitions of unity, a very useful tool for
turning local constructions into global ones. We will construct partitions
of unity on any manifold, and use these in various applications such as
constructing embeddings into Euclidean spaces.

5.2. Partitions of unity

5.2.1. Definition.
Definition 5.2.1. Let M be a manifold, and A = {Uα }α∈A an open cover of
M . A (smooth) partition of unity subordinate to the cover A is a collection
of functions {ρα }α∈A ⊂ C ∞ (M ) satisfying

(i). For every x ∈ M and every α ∈ A, 0 ≤ ρα (x) ≤ 1;


(ii). supp(ρα ) = {x ∈ M : ρα (x) 6= 0} is contained in Uα for each α ∈
A;
(iii). The set of supports {supp(ρα ) : α ∈ A} is locally finite. That is,
for every point x ∈ M there exists an open neighbourhood U of x
in M such that {αP: supp(ρα ) ∩ U 6= ∅} is finite.
(iv). For each x ∈ M , α∈A ρα (x) = 1.
31
32 5. PARTITIONS OF UNITY AND APPLICATIONS

Note that although A may be infinite, the sum in (iv) is in fact a finite sum
for each x ∈ M in view of the local finiteness of (iii).

5.2.2. Existence. We will see some examples of the use of partitions


of unity in the next section. First we will prove the basic existence result
which holds in great generality:
Proposition 5.2.2. Let M be a manifold and A an open cover. Then there
exists a smooth partition of unity subordinate to A.

Proof. We will accomplish this in several stages: First, we will develop


the basic building blocks for the construction, which are smooth compactly
supported functions on Euclidean space:
Lemma 5.2.3. If 0 < r1 < r2 then there exists Hr1 ,r2 ∈ C ∞ (Rn ) such that
H(x) = 1 for x ∈ Br1 (0), 0 < H(x) < 1 for x ∈ Br2 (0) \ Br1 (0), and
H(x) = 0 for x ∈
/ Br2 (0).

Proof. The function


( 1
e− t , t > 0;
f (t) =
0, t≤0
is in C ∞ (R), with f (t) > 0 if and only t > 0. [Proof: Exercise — you
can do this by showing by induction that the kth derivative has the form
P (1/k)e−1/t where P is a polynomial. It follows that the kth derivative has
limit equal to zero as t approaches zero from above; you also need to check
that the k derivative exists and equals zero at k = 0].
f (r2 −t)
Now define h(t) = f (r2 −t)+f (t−r1 ) . Since every t is either less than r2 or
greater than r1 , the denominator is strictly positive, and we have h(t) = 0
for t ≥ r2 and h(t) = 1 for t ≤ r1 and 0 < h(t) < 1 for r1 < t < r2 . Finally,
define H(x) = h(|x|). Since |x| is smooth on Rn \ {0}, we have that H is
smooth on Rn \ {0}, while H is also constant and hence smooth on Br1 (0),
ad so is smooth everywhere on Rn . 

Next we need to construct a suitable collection of ‘coordinate balls’ (i.e.


images of balls in Rn by smooth charts for M ) on which to transplant the
functions constructed above. We will later need to address the issues of
local finiteness and that of being subordinate to a given open cover, but for
now let us simply construct a countable basis for the topology consisting of
coordinate balls:
Definition 5.2.4. A collection on open subsets B = {Vβ : β ∈ B} of M
is a basis for the topology of M if for every x ∈ M and every open set U
containing x, there exists an element Vβ ∈ B such that x ∈ Vβ ⊂ U .
5.2. PARTITIONS OF UNITY 33

Recall that our assumption on the topology of M is that it is second count-


able, which means that there exists a countable basis for the topology.
Exercise 5.2.5. If M is a separable metric space (so that there exists a
countable dense subset of M ) then the topology of M is second-countable
[Hint: Consider the collection of distance balls of rational radius about
points in a countable dense subset of M ].
Lemma 5.2.6. Any smooth atlas for M has a countable sub-atlas (that is, a
countable subset which is also a smooth atlas for M ).

Proof. Let {Ai : i ∈ N} be a countable basis for the topology of M .


If C = {ϕγ : Uγ → Vγ }γ∈C is a smooth atlas for M , then let J = {i ∈
N : Ai ⊂ Uγ for some γ ∈ C}. For each i ∈ J choose γ(i) ∈ C such that
Ai ⊂ Uγ(i) . Then for each x ∈ M we have that x ∈ Uγ for some γ ∈ C
since C is an atlas, so by the definition of basis there exists i ∈ N such that
S∈ Ai ⊂ Uγ for some 0γ, and therefore i ∈ J and x ∈ Ai ⊂ Uγ(i) . Therefore
x
i∈J Uγ(i) = M and C = {ϕγ(i) : i ∈ J} is a countable sub-atlas of C. 

From this we can construct a countable basis for the topology of M consisting
of coordinate balls:
Lemma 5.2.7. Let M be a manifold. Then there exists a countable basis
D = {Bi : i ∈ N} for the topology of M consisting of precompact coordinate
balls. That is, each Bi has the form ϕ−1 (Br (p)), where ϕ : U → V is a
smooth chart for M , and Br (p) ⊂ V .

Proof. Let C 0 = {ϕi : Ui → Vi }i∈N be a countable smooth atlas for


M , and define
D = {ϕ−1 n
i (Br (p)) : r ∈ Q, p ∈ Q , Br (p) ⊂ Vi }.
Then D is certainly countable, and forms a basis for the topology of M :
Given any x ∈ M and a neighbourhood U of x in M we can choose i such
that x ∈ Ui , and then U ∩ Ui is an open neighbourhood of x in M contained
in Vi , so ϕ(U ∩ Ui ) is an open subset of Vi containing ϕi (x). Therefore there
exists r ∈ Q and p ∈ Qn such that ϕi (x) ⊂ Br (p) and Br (p) ⊂ ϕi (U ∩ Ui ),
so that we have x ∈ ϕ−1 −1 −1
i (Br (p)) and ϕi (Br (p)) = ϕ (Br (p)) ⊂ U is
compact. 

Next we will address the local finiteness. This will be done by appealing to
compactness of certain subsets of M , which we will construct as elements of
an exhaustion:
Definition 5.2.8. An exhaustion of a manifold M is a sequence of subsets
{Ki : i ∈ N} such that each Ki is compact and is contained in the interior
of Ki+1 , and every point of M is contained in Ki for some i ∈ N.
34 5. PARTITIONS OF UNITY AND APPLICATIONS

Lemma 5.2.9. If M is a manifold, then there exists an exhaustion {Ki }i∈N


of M .

Proof. We define a sequence of compact sets inductively as follows:


Let D = {Bi : i ∈ N} be a countable basis for the topology of M consisting
of precompact sets
S i(as constructed in the previous lemma). Inductively, we
construct Ki = m j=1 Bj by choosing mi as follows: Choose m1 = 1. Then if
we have chosen m2 , · · · , mi such that Kj ⊂ Int(Kj+1 ) for j = 1, · · · , i − 1,
we observe that the collection D S is an open cover of Ki , so there exists a
finite subcover, and hence Ki ⊂ m i=1 Bj for sufficiently large m. Choose
mi+1 to be the smallest value of m > i for which this is true. Then we have
 
mi+1 mi+1
[ [
Ki ⊂ Bj = Int  Bj  = Int(Ki+1 ),
j=1 j=1
S∞ S∞
as required. Furthermore mi ≥ i implies i=1 Ki ⊃ i=1 Bi = M. 

Using this exhaustion we can produce a locally finite collection of balls which
is subordinate to a given open cover of M .
Definition 5.2.10. Let A and A0 be open covers of a topological space M .
We say that A0 is a refinement of A if every element of A0 is a subset of an
element of A (i.e. each of the open set in the cover A0 is contained in one
of the open sets in the cover A).
Definition 5.2.11. An open cover A of M is locally finite if for every x ∈ M
there exists a neighbourhood U of x in M such that only finite many of the
open sets in A intersect U .
Lemma 5.2.12. Let M be a manifold, and A = {Uα }α∈A an open cover of
M . Then there exists a countable locally finite refinement of A consisting of
precompact coordinate balls.

Proof. Let {Ki : i ∈ N} be an exhaustion of M (as constructed in


Lemma 5.2.9). Also let D = {Bi }i∈N be a countable basis for the topology
of M consisting of coordinate balls as constructed in Lemma 5.2.7.
Let V1 = K1 , and for i > 1 let Vi = Ki \ Int(Ki−1 ). Also let W1 = Int(K2 ),
W2 = Int(K3 ), and Wi = Int(Ki+1 ) \ Ki−2 for i > 2. Then each Vi is
compact and contained in the open set Wi , and we have Wi ∩ Wj = ∅ unless
|i − j| ≤ 2.
Now we observe that since A is an open cover and D is a basis for the
topology, for each x ∈ Vi we can find α(x) ∈ A such that x ∈ Uα (x), and
then find j(x) ∈ N such that x ∈ Bj(x) and Bj(x) ⊂ Uα(x) ∩ Wi . Vi is
compact, and {Bj(x) : x ∈ Vi } is an open cover of Vi , so by compactness
5.2. PARTITIONS OF UNITY 35

there is an open subcover {Bj(x(i,k)) : 1 ≤ k ≤ n(i)}. Also we have that


Bj(x(i,k)) ⊂ Wi ∩ Uα(x(i,k)) for each i and k.
The collection {Bj(x(i,k) : i ∈ N, 1 ≤ k ≤ n(i)} is now a countable collec-
tion of precompact coordinate balls which covers M , since every x ∈ M is
contained in some Vi . Furthermore, this collection is locally finite, since if
x ∈ Vi and |i0 − i| > 2 then Bj(x(i0 ,k)) ⊂ Wi0 , and Wi0 ∩ Wi = ∅, so only
finite many of the balls Bj(x(i0 ,k)) intersect the open neighbourhood Wi of x.
Finally, the collection is a refinement of A since Bj(x(i,k)) ⊂ Uα(x(i,k)) . 

Now we can complete the proof of the existence of a partition of unity: Given
an open cover A = {Uα : α ∈ A} for M , let D = {Bi = ϕ−1 i (Bri (pi )) : i ∈
N} be the countable locally finite refinement of A consisting of precompact
coordinate balls, where ϕi : Ui → Vi is a smooth chart for M . For each i
we have Bi ⊂ Uα(i) for some α(i), so there exists r̃i > ri such that Br̃i (pi ) ⊂
Vi ∩ Uα(i) . Define fi ∈ C ∞ (M ) by
(
Hri ,r̃i (ϕi (x) − pi ), x ∈ Ui ;
fi (x) =
0, otherwise,
where Hri ,r̃i is the smooth cut-off function constructed in Lemma 5.2.3.
Next, define for each α ∈ A
X
Fα (x) = fi (x).
i: α(i)=α

Since the collection {Bi } is locally finite, the sum defining Fα is finite in
a neighbourhood of any point x, and so ρα is a smooth function. Finally,
define X X
F (x) = Fα (x) = fi (x),
α i
which is a smooth function for the same reason, and is strictly positive since
every fi is non-negative, and every point x ∈ M is in Bi for some i, and
hence fi (x) = 1. Finally, define
Fα (x)
ρα (x) := .
F (x)
This is a smooth function since Fα and F are both smooth and F is positive,
and we have satisfied (iv). 

5.2.3. The compact case. To help understand the previous construc-


tion, we consider the special case where the manifold M is compact: In this
case the argument is quite a lot simpler since we do not have to construct
locally finite refinements and exhaustions.
Let M be a compact manifold and A = {Uα : α ∈ A} an open cover of
M . For each x ∈ M we can choose α(x) such that x ∈ Uα(x) , and a smooth
36 5. PARTITIONS OF UNITY AND APPLICATIONS

chart ϕx : Ux → Vx for M with x ∈ Ux , and then choose rx > 0 such


that B3rx (ϕx (x)) ⊂ ϕx (Uα(x) ∩ Ux ) ⊂ Vx . Let Bx = ϕ−1
x (Brx (ϕx (x))) and
−1
B̂x = ϕx (B2rx (ϕx (x))). Then the collection {Bx }x∈M is an open cover of
M , and so has a finite subcover {Bxi }ki=1 .
Define fi (y) = Hrxi ,2rxi (ϕxi (y) − ϕxi (xi )). Then fi ∈ C ∞ (M ) since it is
smooth in Uxi and zero outside B̂xi , and each fi is non-negative and equal
to 1 on Bi . Then let X
Fα (y) = fi (y)
i: α(xi )=α
and
X k
X
F (y) = Fα (y) = fi (y).
α∈A i=1
Then F is smooth since it is a finite sum of smooth functions, and F (y) ≥ 1
for every y since y ∈ Bi for some i. Finally, let

ρα = .
F
Then {ρα : α ∈ A} is the desired partition of unity. Note that the lo-
cal finiteness of the supports is automatic since only finitely many of the
functions ρα are different from the zero function.

5.3. Applications of partitions of unity

Now let us see some examples of the application of partitions of unity:

5.3.1. Extension of locally defined smooth functions. The first


application is to the problem of extending locally defined smooth functions
to global ones:
Proposition 5.3.1. Let M be a manifold, A a closed subset of M and U
an open subset of M containing A. Suppose f ∈ C ∞ (U ). Then there exists
a function f˜ ∈ C ∞ (M ) such that f˜(x) = f (x) for x ∈ A and f˜(x) = 0 for
x∈/ U.

Proof. Consider the open cover of M consisting of the two sets U and
M \ K. Since A ⊂ U this is a cover of M , so there exists a partition of unity
{ρ1 , ρ2 } for M with supp(ρ1 ) ⊂ U and supp(ρ2 ) ⊂ M \ K. Define
(
f (x)ρ1 (x), x ∈ U ;
f˜(x) =
0, x∈/ U.
Then f˜ is smooth: It is clearly smooth on U ; if x ∈
/ U then x ∈ M \ supp(ρ1 )
which is open, so there is a neighbourhood of x on which f˜ ≡ 0 and hence
f˜ is smooth at x. 
5.3. APPLICATIONS OF PARTITIONS OF UNITY 37

Exercise 5.3.2. Give a counterexample to the proposition above if we drop


the assumption that A is closed.
Exercise 5.3.3. We defined tangent vectors at x as derivations at x acting
on the vector space of smooth function germs Cx∞ (M ). One could instead
define derivations at x acting on C ∞ (M ) (i.e. on smooth functions defined
on the whole manifold). Show that the space of derivations at x acting on
Cx∞ (M ) is isomorphic to the space of derivations at x acting on C ∞ (M ).

5.3.2. Smooth functions on submanifolds are restrictions. Next


we consider extension of smooth function defined on submanifolds. We
proved earlier that if Σ is a submanifold of a manifold M , and f is a smooth
function on M , then the restriction of f to Σ is a smooth function on Σ.
Here we prove the converse:
Proposition 5.3.4. Let Σk be a closed submanifold of a manifold M m.
∞ ∞

Then for every f ∈ C (Σ) there exists g ∈ C (M ) such that g Σ = f .

Proof. We use the definition of submanifold to first construct local


extensions of f into M : For each x ∈ M there is a smooth chart ϕx : Ux →
Vx ⊂ Rm for M such that ϕx (Σ ∩ Ux ) = (Rk × {0}) ∩ Vx , and we can assume
that Vx = Ax × Bx ⊂ Rk × Rm−k ' Rm On the set Ux we can define an
extension f˜x ∈ C ∞ (Ux ) by
f˜x (p) = f ◦ ϕ−1 ◦ π ◦ ϕ
where π(x, y) = (x, 0) projects Vx smoothly to (Rk × {0}) ∩ Vx . If y ∈ Σ ∩ Ux
then this gives f˜x (y) = f (y).
Next we use a partition of unity to combine these local extensions into a
smooth function on M : The collection A = {Ux : x ∈ Σ} ∪ {M \ Σ} is an
open cover of M , so there exists a partition of unity {ρx : x ∈ Σ} ∪ {ρΣ }
subordinate to this cover. Define
X
g(y) = ρx (y)f˜x (y)
x∈Σ

where we interpret ρx (y)f˜x (y) to be zero if y ∈/ Ux . Then each function f˜x


is smooth on M , and the sum is locally finite, so g is a smooth function on
M . Further, for y ∈ Σ we have f˜x (y) = f (y) whenever ρx (y) 6= 0, so
X
g(y) = ( ρx (y))f (y) = (1 − ρΣ (y))f (y).
x∈Σ

But supp(ρΣ ) ⊂ M \ Σ, so ρΣ (y) = 0 for y ∈ Σ, and therefore


g(y) = f (y)
for y ∈ Σ, as required. 
38 5. PARTITIONS OF UNITY AND APPLICATIONS

We could also insist that the extension vanish outside a specified neigh-
bourhood of Σ (this follows by combining the result just proved with the
extension result of the previous section).

Exercise 5.3.5. Let M be a manifold, and A ⊂ M a closed subset. We say


that a function f on A is smooth (and write f ∈ C ∞ (A)) if for every x ∈ A
there exists an open set U of M containing x and a function f ∈ C ∞ (U ) such
that f˜(x) = f (x) for each x ∈ A∩U . Prove the following extension theorem:
If A is a closed subset of a manifold M , and U is an open neighbourhood
of A, then every function f ∈ C ∞ (A) is the restriction to A of a function
g ∈ C ∞ (M ) which vanishes outside U .

5.3.3. Exhaustion functions.

Definition 5.3.6. An exhaustion function on a manifold M is a smooth


function f ∈ C ∞ (M ) such that f −1 ((−∞, c]) is a compact subset of M for
every c ∈ R.

Thus an exhaustion function can be thought of as a function which ap-


proaches infinity along any sequence which ‘approaches infinity’ in M —
here we understand that a sequence (xn ) in M ‘approaches infinity’ if it
eventually leaves each compact subset K of M , i.e. for any compact subset
K of M there exists N such that n ≥ N implies xn ∈ / K.

Proposition 5.3.7. Let M be a manifold. Then there exists a smooth ex-


haustion function f on M .

Proof. Let {Ki }i∈N be an exhaustionSof M by compact sets, so that


each Ki is compact, Ki ⊂ Int(Ki+1 ) and i Ki = M . Consider the open
cover of M given by the open sets Wi = Int(Ki+1 ) \ Ki−2 for i ≥ 2, W1 =
Int(K2 ), and W2 = Int(K3 ). Let {ρi }i∈N be a partition of unity subordinate
to this cover, and define
X∞
f (x) = iρi (x).
i=1

Then f ∈ C ∞ (M ) since the sum is locally finite. Furthermore, if x ∈


/ Ki
then x ∈ Wj , which implies that j ≥ i. Therefore

X ∞
X ∞
X
f (x) = jρj (x) ≥ iρj (x) = i ρj (x) = i.
j=1 j=1 j=1

Conversely, if f (x) < i then x ∈ Ki . Then for any c ∈ R we have that


f −1 ((−∞, c]) is a closed subset of the compact set Ki for any i > c, and is
therefore compact. 
5.4. LEVEL SETS OF SMOOTH FUNCTIONS 39

5.4. Level sets of smooth functions

We know that level sets of smooth submersions, but if we drop the sub-
mersion condition (that the derivative is surjective) then things can be very
much worse:

Proposition 5.4.1. Let M be a manifold and A any closed subset of M .


Then there exists g ∈ C ∞ (M ) such that A = f −1 (0).

Proof. First let C be a closed subset of Rn . Then the complement of


C is open, and hence is the union of a countable number of open balls of
radius at most 1 (for example, for each point xi with rational coordinates
in Rn \ C take the largest ri ≤ 1 for which Bri (xi ) ∩ C = ∅).
Let h ∈ C ∞ (Rn ) be a smooth bump function which is equal to 1 on B 1/2 (0)

and nonzero precisely in B1 (0). Let Ci = sup{max0≤j≤i D(j) h x : x ∈
B 1 (0)} for each i ≥ 0. Then define
∞  
X 1  ri i x − xi
f (x) = h .
Ci 2 ri
i=1

We will check that f is a smooth function which is zero precisely on C: First,


since Ci ≥ sup h for every i, and ri ≤ 1, the ith term in the sum defining
f is no greater than 2−i at every point x ∈ Rn , and so the series converges
uniformly to a continuous limit function f .
Now consider the derivatives: The kth derivative of the ith term (thought
of as a k-multilinear map at each point) in the sum defining f is given by

(ri )i−k
 
x − xi
D k h ,
2 i Ci ri

which for i ≥ k is bounded by 2−i by the choice of Ci . Therefore the sum of


these terms converges uniformly on Rn to a continuous k-multilinear map,
and it follows (from convergence in the Taylor expansion about any point)
that the limit is the kth derivative of f . Therefore f ∈ C ∞ (Rn ), and we
have f (x) = 0 for x ∈ C (since all the terms are zero), but f (x) > 0 for
x∈/ C (since x ∈ Bri (xi ) for some i, so the ith term in the sum is strictly
positive).
Now let A be a closed set in an n-dimensional manifold M . Then for each
x ∈ A we can choose a coordinate chart ϕx : Ux → Vx ⊂ Rn with x ∈ Ux .
Then by the previous argument there exists fx ∈ C ∞ (Rn ) which is non-
negative and vanishes precisely on ϕx (Ux ∩ A) (in particular, the set in Vx
where fx vanishes is ϕx (Ux ∩ A)). Now {Ux : x ∈ A} ∪ {M \ A} is an open
cover, so there is a partition of unity {ρx : x ∈ A} ∪ {ρM } subordinate to
40 5. PARTITIONS OF UNITY AND APPLICATIONS

this cover. Define


X
g(y) = fx (ϕx (y))ρx (y) + ρM (y).
x∈A
This is locally a finite sum of non-negative smooth functions, hence is smooth
and non-negative.
If g(x) = 0Pthen every term in the sum must be zero. In particular ρM (y) =
0. Since x ρx (y) + ρM (y) = 1 this implies that ρx (y) > 0 for some x.
It follows that y ∈ Ux . We must then have fx (ϕx (y)) = 0, implying that
ϕx (y) ∈ ϕx (Ux ∩ A) and therefore y ∈ A.
Conversely, if y ∈ A then ρM (y) = 0, and fx (ϕx (y)) = 0 for all x such that
ρx (y) > 0, and so we have g(y) = 0. 
CHAPTER 6

The tangent bundle and other vector bundles

Contents
6.1. The tangent bundle 41
6.2. Vector bundles 43
6.2.1. Bundle maps 45
6.2.2. Sub-bundles 46
6.2.3. Pull-back bundles 46
6.2.4. The tangent bundle of a submanifold 46
6.3. Lie Groups 49

6.1. The tangent bundle

Recall that at each point of a manifold M n we have an n-dimensional vector


space Tx M consisting of the tangent vectors at x, which we can think of as
differential operators (‘derivations’) acting on (germs of) smooth functions,
or as tangent vectors γ 0 to smooth curves γ through x. As we have seen,
any smooth chart ϕ : U → V for M near x (which we write in components
as ϕ = (x1 , · · · , xn ) provides a convenient isomorphism between Tx M and
Rn , given by
v ∈ Tx M 7→ vϕ := (vx1 , · · · vxn )
as derivations, or alternatively by
d
γ 0 7→ (ϕ ◦ γ(s))

ds s=0
for any smooth curve γ ∈ Cx .
The tangent bundle T M is the total space of all tangent vectors to points
of M . As a set, we can write this as
T M = {(x, v) : x ∈ M, v ∈ Tx M }.
It is important to observe that the tangent bundle T M can itself be given
the structure of a smooth manifold:
Proposition 6.1.1. If M is a smooth manifold of dimension 2n with a
smooth atlas A = {ϕα : Uα → Vα }α∈A , then T M can be given the structure
41
42 6. THE TANGENT BUNDLE AND OTHER VECTOR BUNDLES

of a smooth manifold of dimension 2n, with smooth atlas given by the charts
ϕ̃α : T Uα := {(x, v) : x ∈ Uα , v ∈ Tx M } → Vα × Rn defined by
ϕ̃α (x, v) = (ϕα (x), vϕα ) ∈ Vα × Rn .

Proof. We first check that the transition maps are diffeomorphisms: If


ϕ and η are two charts, then η̃ − takes (p, v) to (η −1 (p), (s 7→ η −1 (p + sv))0 ).
Then ϕ̃ takes this to (ϕ ◦ η −1 (p), ds ϕ ◦ η −1 (p + sv) s=0 ). Thus we have
d


ϕ̃ ◦ η̃ −1 (p, v) = ϕ ◦ η −1 (p), D(ϕ ◦ η −1 )|p (v) .




Since ϕ ◦ η −1 is smooth, clearly ϕ̃ ◦ η̃ −1 is smooth (in both p and v). Since


we can interchange ϕ and η it follows that the transition maps are diffeo-
morphisms.
For completeness we need to consider the topological aspects: Since M is a
manifold we can cover M by countably many charts (as proved in Lemma
5.2.6), so T M is also covered by countably many charts.
Since M is Hausdorff, given any pair of distinct points x and y in M there
exist disjoint open sets A and B in M with x ∈ M and y ∈ N . Therefore, if
(x, v) and (y, w) are distinct points in T M and x 6= y then we can choose a
chart ϕ about x for M with domain contained in A and a chart η about y for
M with domain contained in B, where A and B are disjoint open subsets of
M . Then (x, v) is contained in the domain of ϕ̃, and (y, w) is contained in
the domain of η̃, and these domains are disjoint. Otherwise we have x = y
and so both (x, v) and (y, w) are in the domain of a single chart ϕ̃ for T M
corresponding to a chart ϕ for M about x.
By the smooth chart lemma, M has a unique topological structure which
makes the charts ϕ̃ homeomorphisms (generated by the open sets ϕ̃−1 (B)
where B is an open subset of V × Rn ), and with respect to this topology
T M is a smooth manifold with the atlas à = {ϕ̃α : α ∈ A}. 
Proposition 6.1.2. The projection π from T M to M given by π(x, v) = x
is a smooth map, and the derivative of π is surjective at every point.

Proof. To check smoothness of π at (x, v) we choose a chart ϕ for M


about x, so that ϕ̃ is a chart for T M about (x, v). With respect to these
charts we have
ϕ ◦ π ◦ ϕ̃−1 (p, v) = ϕ ◦ π(ϕ−1 (p), (s 7→ ϕ−1 (p + sv))0 ) = ϕ(ϕ−1 (p)) = p
which is certainly a smooth function of p. To check the surjectivity of the
derivative at (x, v) we need to find for any γ 0 ∈ Tx M a smooth curve γ̃
through (x, v) in T M such that π ◦ γ has tangent vector γ 0 . Choose a chart
ϕ as above, and take
v i ∂iϕ )
X
γ̃(s) = (γ(s)),
i
6.2. VECTOR BUNDLES 43
P i
where v ∈ i v ∂i ∈ Tx M . Then we have γ̃(0) = (γ(0), v) = (x, v) as
required, while Dπ (x,v) (γ̃ 0 ) = (π ◦ γ̃)0 = γ 0 . 

Also important is the following observation: If F : M → N is a smooth


map between manifolds, then the derivative defines a natural map from T M
to T N , defined by

DF (x, v) := (F (x), DF |x (v)) ∈ T N

for each (x, v) in T M .

Proposition 6.1.3. If F ∈ C ∞ (M, N ) then DF ∈ C ∞ (T M, T N ).

Proof. We must show that given charts ϕ̃ for T M about (x, v) (which
we can take as corresponding to a chart ϕ for M about x) and η̃ for T N
about (F (x), DF |x (v)) (which can be chosen as arising from a chart η for
N about F (x)), the composition η̃ ◦ DF ◦ ϕ̃−1 is smooth.
We can compute this map explicitly: Under ϕ̃−1 , the point (p, v) maps to
(ϕ−1 (p), γ 0 ) where γ is the smooth curve defined by γ(s) = ϕ−1 (p + sv).
The map DF takes this to (F (ϕ−1 (p)), (F ◦γ)0 ), and then η takes this to
η ◦ F ◦ ϕ−1 (p), ds
d
η ◦ F ◦ ϕ−1 (p + sv) |s=0 . This gives

η̃ ◦ DF ◦ ϕ̃−1 (p, v) = η ◦ F ◦ ϕ−1 (p), D η ◦ F ◦ ϕ−1 |p (v) .


  

By assumption F is smooth, so η ◦ F ◦ ϕ−1 is smooth, and therefore η̃ ◦ DF ◦


ϕ̃−1 is smooth in both p and v, as required. 

6.2. Vector bundles

The tangent bundle has a special structure, which arises frequently in dif-
ferential geometry: It is an instance of a vector bundle.

Definition 6.2.1. A vector bundle (E, M, π) of rank k is a manifold E


(called the total space, a manifold M (called the base space), and a smooth
surjective map π from E to M such that

(i). For each x ∈ M , the fibre Ex = π −1 (x) of E over x has the structure
of a vector space of dimension k; and
(ii). For each x ∈ M there exists a neighbourhood U of x in M and a
diffeomorphism Ψ : π −1 (U ) → U × Rk (called a local trivialisation
such that Ψ(ξ) = (π(ξ), L(ξ)), where the restriction of L to each
fibre Ex is a linear isomorphism from Ex to Rk .
44 6. THE TANGENT BUNDLE AND OTHER VECTOR BUNDLES

The tangent bundle is a particular example: For the local trivialisations we


can choose a smooth chart ϕ : U → V for M , and then define
n
!
X
Φ x, v i ∂i = (x, (v 1 , · · · , v n )),
i=1
or equivalently
Φ(x, v) = (x, vϕ).

Another obvious example is the product bundle M × Rk .


Here is another interesting example:
Example 6.2.2 (The Möbius bundle). We take E to be R2 / ∼, where
(x1 , y1 ) ∼ (x2 , y2 ) if and only if x1 − x2 ∈ Z and y1 = (−1)x1 −x2 y2 . We
will write this is a bundle over the base space M = S 1 = R/Z, which is
the set of equivalence classes of point in R modulo the equivalence relation
x2 ∼ x1 if and only if x2 − x1 ∈ Z.
To give E and M manifold structures, we observe that the projection from
R2 to E, when restricted to a subset of the form (a, a + 1) × R, defines a
bijection ψa−1 to its image, so we can take {ϕa : a ∈ R} to be an atlas
of charts for E. Similarly, the projection from R to R/Z, restricted to
(a, a + 1), defines a bijection ϕ−1
a to its image in R/Z. The transition maps
between these charts are bijections ϕb ◦ ϕ−1
a : (a, a + 1) \ {a + frac(b − a)} →
(b, b + 1) \ {b + frac(a − b)} (where frac(x) = x − bxc is the fractional part
of x) given explicitly by
ϕb ◦ ϕ−1
a (x) = x − bb − ac.
This map is a translation when restricted to each of the subintervals of the
domain, and so is smooth. Similarly, ψb ◦ ψa−1 is a bijection from ((a, a +
1) \ {a + frac(b − a)}) × R to ((b, b + 1) \ {b + frac(a − b)}) × R defined by
 
ψb ◦ ψa−1 (x, y) = x − bb − ac, (−1)bb−ac y .
On each connected component of the domain this is either a translation in
x, or a translation in x composed with a reflection in y, and so is smooth
on its domain.
The fact that this defines a smooth manifold structure can be checked eas-
ily using the smooth chart lemma (Lemma 3.2.1), and the vector bundle
structure is also straightforward.
Example 6.2.3. Recall that we defined the cotangent space Tx∗ M of M at x
to be the space of equivalence classes of germs of smooth functions under the
d d

equivalence relation f ∼ g if and only if ds f ◦γ(s) s=0 = ds g ◦γ(s) s=0 for all
smooth curves γ in M through x. The equivalence class of f is denoted by
df (x). T ∗ M has the structure of a vector bundle over M , with local charts
given as follows: If ϕ : U → V is a smooth chart for M , then any co-vector
6.2. VECTOR BUNDLES 45

at x ∈ U can be written in terms P of the basis of coordinate differential


dx1 , · · · , dxn , explicitly by df = ni=1 df (∂i )dxi . Define ϕ̃ : T ∗ U → U × Rn
by
ϕ̃(df (x)) = (ϕ(x), (df (∂i ), · · · , df (∂n ))).

6.2.1. Bundle maps.

Definition 6.2.4. If E is a vector bundle over a manifold M , and Ẽ is a


vector bundle over a manifold M̃ , then a bundle map G from E to Ẽ is a
smooth map which is linear on each fibre: Precisely, G is a smooth map
E to Ẽ such that for each x ∈ M , there exists F (x) ∈ M̃ such that
from
G Ex is a linear map from Ex to ẼF (x) .

It follows from the definition that F is a smooth map from M to M̃ : For


−1
if ϕ : πE (U ) → U × Rk is a local trivialisation for E with x ∈ U , and
−1
η : πẼ (V ) → V × R` is a local trivialisation for Ẽ with F (x) ∈ V , then we
have
η ◦ G ◦ ϕ−1 (x, v) = (F (x), L(x)v)
for some linear map L(x) ∈ L(Rk , R` ), and since this map is smooth we have
that F is smooth and L is smooth on U . We say that G covers the map
F ∈ C ∞ (M, M̃ ).
A bundle map is also sometimes referred to as a bundle homomorphism. If
there exists an inverse bundle map then we say that G is a bundle isomor-
phism. If a bundle map G is injective on each fibre Ex
If E and Ẽ are vector bundles over the same manifold M , and G : E → Ẽ
is a bundle map which covers the identity map on M , we say that G is a
bundle map over M . This defines a natural isomorphism on the set of vector
bundles over a manifold M , where we say two such bundles are isomorphic if
there is a bundle isomorphism over M from one to the other. In particular,
a bundle E over M is called trivial is it is isomorphic to a product bundle
M × Rk .

Exercise 6.2.5. Suppose F ∈ C ∞ (M, N ). Show that DF ∈ C ∞ (T M, T N )


is a bundle map covering F .

Exercise 6.2.6. Let E be a vector bundle over M , and Σ a submanifold


in M . Show that π −1 (Σ) is a vector bundle over Σ, and the inclusion of
π −1 (Σ) into Σ is a bundle map covering the inclusion of Σ in M .

Exercise 6.2.7. Let E be a vector bundle over M , and f ∈ C ∞ (M ). Show


that the map ξ ∈ Ex 7→ f (x)ξ ∈ Ex defines a smooth bundle map trom E
to itself over M . Under what conditions is this map a bundle isomorphism?
46 6. THE TANGENT BUNDLE AND OTHER VECTOR BUNDLES

6.2.2. Sub-bundles.
Definition 6.2.8. Let E be a vector bundle of rank n over a manifold M m .
Then a subset Ẽ of E is a sub-bundle of E if for every x ∈ M there exists a
neighbourhood U of x and a local trivialisation ψ : π −1 (U ) → U × Rn for
E such that
ψ(Ẽ ∩ π −1 (U )) = π −1 (U ) × (Rk × {0}).

The definition implies that for each x, Ex ∩ Ẽ is a k-dimensional vector


subspace of the n-dimensional fibre Ex . Note that Ẽ is also a submanifold
of E.

6.2.3. Pull-back bundles.


Definition 6.2.9. Let f ∈ C ∞ (M, N ), and suppose E is a vector bundle
over N . Then the pull-back bundle of E by f , denoted by f ∗ E, is the vector
bundle over M defined by
f ∗ E = {(x, v) : x ∈ M, v ∈ Ef (x) },
−1
with the projection π(x, v) = x. If ψ : πE (U ) × U × Rk is a local triviali-
sation of E, given by Ψ(v) = (πE (x), L(v)) then define
Ψ(x, v) = (x, L(v))
Then Ψ : π −1 (f −1 (U )) → f −1 (U ) × Rk is a local trivialisation of E.

There is a natural smooth bundle map from f ∗ E to E covering f , given by


(x, v) ∈ f ∗ E 7→ v ∈ Ef (x) .

6.2.4. The tangent bundle of a submanifold.


Proposition 6.2.10. If F : M m → N n is an embedding, then DF (T M ) is
a sub-bundle of F ∗ (T N ), and DF defines a bundle isomorphism from T M
to DF (T M ) over M .

Proof. The image F (N ) is a submanifold of N . Given x ∈ M , choose


a smooth chart ϕ for M near x and a local chart η for N near F (x) which is
a submanifold chart for F (N ): That is, η(F (M ) ∩ W ) = (RmP × {0}) ∩ Z ⊂
Rn . Then a local trivialisation for F ∗ T N is given by (x, ni=1 v i ∂iη ) 7→
(x, (v 1 , · · · , v n )). It follows that DF |x (Tx M ) = span(∂1η , · · · , ∂nη ) ⊂ TF (x) N =
(F ∗ (T N ))x , so this local trivialisation shows that DF (T M ) is a sub-bundle
of F ∗ T N . 

We call the bundle DF (T M ) the space of ‘geometric tangent vectors’, and


we will routinely identify this with the tangent bundle itself when we are
working with submanifolds.
6.2. VECTOR BUNDLES 47

A particular case is useful to think about here: If Σk is a submanifold in


Euclidean space Rn , then the tangent bundle to Rn is just the product bundle
Rn × Rn . The pull-back bundle (by the inclusion of Σ into Rn ) is again just
the product bundle Σ × Rn . So the proposition says that we can identify
the space of tangent vectors to Σ at x with a k-dimensional subspace of Rn ,
given by the image of the derivative of the inclusion map. Equivalently, we
can identify Tx M with the subspace of Rn consisting of the tangent vectors
to smooth curves through x which lie in Σ.
Proposition 6.2.11. If Σ is a submanifold of M , and ι is the inclusion of
Σ in M , then the space of geometric tangent vectors Dι(Tx Σ) is given by

(i). The space of tangent vectors γ 0 (0) to smooth curves γ in M through


x with image in Σ;
(ii). The kernel of DG|x , where G is a local defining function for Σ, i.e.
G : U → Rn−k is a submersion from a neighbourhood U of x with
Σ ∩ U = G−1 (0);
(iii). The image of DF |z where F is a local parametrisation of Σ, i.e.
F : A ⊂ Rk → M is an embedding with F (z) = x and Σ ∩ U =
F (A) for some open neighbourhood U of x in M ;
(iv). The space DΨ|−1 k
x (R × {0}) where Ψ is a submanifold chart for Σ
near x, i.e. Ψ : U ⊂ M → V ⊂ Rn is a diffeomorphism with
Ψ(Σ ∩ U ) = (Rk × {0}) ∩ V .
(v). The space of derivations v acting on Cx∞ (M ) such that vf = 0
whenever f |Σ = 0.

Proof. (i). Recall that Tx Σ is the quotient of CxΣ (the space of


smooth curves in Σ through x) by the equivalence relation γ1 ∼
γ2 ⇐⇒ ds f ◦ γ1 (s) s=0 = ds f ◦ γ2 (s) s=0 for all f ∈ Cx∞ (Σ).
d d


Such a tangent vector γ 0 maps to the tangent vector Dι|x (γ 0 ) =
(ι ◦ γ)0 ∈ Tx M , i.e. the equivalence class of the same curve γ under
the equivalence relation in which Cx∞ (M ) replaces Cx∞ (Σ). We
observe that two curves γ1 and γ2 in Σ are equivalent in Σ if and
only if they are equivalent in M , since every smooth function on
M restricts to one on Σ, and every smooth function on Σ extends
to one on M .
(ii). For every smooth curve γ in Σ passing through x we have G◦γ(s) =
0 for all s, and hence DG|x (γ 0 ) = dsd
(G ◦ γ(s))|s=0 = 0. Therefore
Dι(Tx Σ) is contained in the kernel of DG|x , and since both have
the same dimension they must coincide.
(iii). If F is a local parametrisation of Σ with F (z) = x, then DF |z (v) =
(F (z + sv))0 is the tangent vector to a smooth curve in Σ passing
through x, and so is contained in Dι(Tx Σ). Therefore DF |z (Rk ) ⊂
Dι(Tx Σ). Since the dimension of DF |z (Rk ) and of Dι(Tx Σ) are
both k, the two spaces coincide.
48 6. THE TANGENT BUNDLE AND OTHER VECTOR BUNDLES

(iv). This is immediate from (iv) since the restriction of Ψ−1 to Rk × {0}
is a local parametrisation of Σ.
(v). A tangent vector in Tx Σ (as a derivation) maps by Dι to the deriva-
tion on Cx∞ (M ) defined by (Dι(v))f
= v(f ◦ ι). Note f ◦ ι is just
the restriction of f to Σ. If f Σ = 0 then f ◦ ι = 0 and hence
(Dι(v))f = 0. Conversely,
if ξ is a derivation on Cx∞ (M ) for which
ξ(f ) = 0 whenever f Σ = 0, then we can define a derivation v act-
ing on Cx∞ (Σ) as follows: If f ∈ Cx∞ (Σ) then choose g ∈ Cx∞ (M )
with g Σ = f , and define v(f ) = ξ(g) (note an extension always
exists by Proposition 5.3.4). This is well-defined
since if g1 and g2
are two such extensions of f , then (g1 − g2 ) Σ = 0, and therefore
0 = ξ(g1 −g2 ) = ξ(g1 )−ξ(g2 ). The fact that v is a derivation follows
directly, and we have by construction (Dι(v))g = v(g|Σ ) = ξ(g).


Example 6.2.12 (The tangent space of the sphere). The sphere S n ⊂ Rn+1
is the zero set of the submersion G : Rn+1 \{0} → R given by G(x) = |x|2 −1.
Therefore the tangent space Tx S n at a point x ∈ S n is given by

Tx S n = ker(DG|x )
= {v ∈ Rn+1 : DG|x (v) = 0}
d
= {v ∈ Rn+1 :

G(x + sv) s=0 = 0}
ds
n+1 d
|x + sv|2 s=0 }

= {v ∈ R :
ds
d
= {v ∈ Rn+1 : |x|2 + 2sx · v + s2 |v|2 s=0 = 0}

ds
n+1
= {v ∈ R : x · v = 0}.

That is, Tx S n is the orthogonal complement of Rx in Rn+1 .

Example 6.2.13 (The tangent space to SL(n)). We proved previously that


SL(n) is a submnaifold of Mn of codimension one (hence dimension n2 − 1)
and is given by the zero set of the submersion G : GL(n) → R defined by
G(M ) = det M − 1. The tangent space is then given by the kernel of the
6.3. LIE GROUPS 49

derivative of G:
TM SL(n) = kerDG|M
d
= {A : (det(M + sA)) s=0 = 0}
ds
d
det(M (I + sM −1 A)) s=0 = 0}

= {A :
ds
d
det I + sM −1 A s=0 = 0}

= {A :
ds
= {A : trace(M −1 A) = 0} [Why?]
= {M B : trace(B) = 0}.
In particular the tangent space at the identity matrix In is the space of
trace-free matrices.
Example 6.2.14 (The tangent space to O(n)). The group O(n) of orthogo-
nal n × n matrices is a submanifold of dimension n(n−1)
2 in the space of n × n
matrices, given by the zero set of the submersion G(M ) = M T M − I from
GL(n) to Sym(n). The tangent space is therefore given as follows:
TM O(n) = ker(DG|M )
d
M + sA)T (M + sA) − I s=0 = 0}

= {A ∈ Mn :
ds
= {A ∈ Mn : AT M + M T A = 0}
= {M B : B ∈ Mn , B T + B = 0}.
In particular, the tangent space at the identity matrix is the space of anti-
symmetric matrices.

6.3. Lie Groups

Definition 6.3.1. A Lie group G is a smooth manifold which has a group


structure, such that the group operators (the group multiplication mult :
G × G → G and the inversion inv : G → G) are smooth.
Example 6.3.2. The group GL(n) of invertible n × n matrices is an open
subset of Mn , hence a manifold. The group operation is matrix multiplica-
tion, which is a bilinear (hence smooth) map from Mn × Mn → Mn , hence
also from GL(n) × GL(n) → GL(n). The inversion is given by
cofactor(M )
M −1 =
det M
where cofactor(M ) is the matrix of cofactors of M , for which each component
is a polynomial in the components of M , hence a smooth function on Mn ,
and det M is a polynomial (hence smooth) function on Mn which is non-zero
on det M . Therefore inv(M ) is a smooth function on GL(n), and GL(n) is
a Lie group.
50 6. THE TANGENT BUNDLE AND OTHER VECTOR BUNDLES

Proposition 6.3.3. If G is a Lie group, and Σ is a subgroup of G which is


also a submanifold of G, then Σ is a Lie group.

Proof. Since Σ is a subgroup, the restriction of mult to Σ × Σ (which


is a submanifold of G × G) has values in Σ, and the restriction of inv to
Σ has values in Σ. Since mult ∈ C ∞ (G × G, G) and inv ∈ C ∞ (G, G), the
restriction to Σ × Σ is in C ∞ (Σ × Σ, Σ), and the restriction of inv is in
C ∞ (Σ, Σ) by Propositions 3.3.4 and 3.3.5. 
Example 6.3.4. The groups SL(n) and O(n) are matrix groups (that is,
subgroups of GL(n)) which are also submanifolds in GL(n). It follows that
SL(n) and O(n) are also examples of Lie groups.

On a Lie group G we have a natural family of diffeomorphisms, defined as


follows: Fix g ∈ G, and define `g : G → G by `g (h) = gh. This is called the
left transltion by g.
Note that `g = mult ◦ ιg , where ιg (gh) = (g, h) is a smooth map from G to
G × G. Therefore `g is a composition of smooth maps, hence smooth.
We also have that `g−1 ◦ `g (h) = `g−1 (gh) = g −1 (gh) = (g −1 g)h = h, so that
`g−1 = (`g )−1 , and it follows that each `g is a diffeomorphism. Note that
`g ◦ `h = `gh for any g and h in G.
Since `g (e) = g, the derivative map D`g |e is a linear isomorphism from Te G
to Tg G for each g. This gives a natural way to identify tangent spaces of G
at different point.
In the examples of the matrix groups GL(n), SL(n) and O(n), the left-
translation `M is simply matrix multiplication on the left by the matrix M ,
which is a linear map of A ∈ Mn . Since the derivative of a linear map
is the map itself, this means that D`M |I (A) = M A. This agrees with our
computations of the tangent spaces above: The tangent space to the identity
of the special linear groups SL(n) is given by
TI SL(n) = {B : Trace(B) = 0},
and the tangent space at an arbitrary point M ∈ SL(n) is given by left-
translation:
TM SL(n) = D`M |I ({B : Trace(B) = 0} = {M B : Trace(B) = 0}.
The case of O(n) is similar.
CHAPTER 7

Sard’s theorem and Embedding theorems

Contents
7.1. Embedding compact manifolds in high dimension 51
7.2. Sard’s theorem and sets of measure zero 52
7.3. Reducing dimension of embeddings 53
7.4. Embedding non-compact manifolds 55

It is an important question whether an abstract manifold can always be


realised as a submanifold of some Euclidean space. This was answered by
Hassler Whitney in the mid-1900s. We begin with a simple case of his result:

7.1. Embedding compact manifolds in high dimension

Proposition 7.1.1. Let M be compact manifold. Then for sufficiently large


N there exists a smooth embedding F : M → RN .

Proof. The proof uses smooth cut-off functions, which we can construct
from cut-off functions on Rn from Lemma 5.2.3:
Let M be a compact n-dimensional manifold. For each x ∈ M we can find
a chart ϕx : Ux → Vx ⊂ Rn about x. For each x we can choose rx > 0
such that B3rx (ϕx (x)) ⊂ Vx , and then define Bx = ϕ−1
x (Brx (ϕx (x))) and
B̂x = ϕ−1
x (B 2rx (ϕx (x))).
Since M is compact, we can find a finite sub-cover corresponding to points
x1 , · · · , xk . Label the corresponding charts ϕi and the balls Bi and B̂i . For
each i we can find a smooth function ρi which is equal to 1 on Bi (and only
on Bi ) and equal to zero outside B̂i , by taking ρi (y) = Hr(xi ),2r(xi ) (ϕi (y) −
ϕi (xi )), and ρi (y) = 0 otherwise.
Then define F : M → Rk(n+1) by
F (x) = (ρ1 (x)ϕ1 (x), ρ1 (x), ρ2 (x)ϕ2 (x), ρ2 (x), · · · , ρk (x)ϕk (x), ρk (x)).
This is clearly smooth. We will show that F is an injective immersion, hence
an embedding by Corollary 1.2.14.
51
52 7. SARD’S THEOREM AND EMBEDDING THEOREMS

First, F is injective: If F (y) = F (z) then we can choose i such that y ∈


Bi and hence ρi (y) = 1. But then ρi (z) = 1 also since ρi is one of the
components of F , so z ∈ Bi also. We also have ρi (y)ϕi (y) = ρi (z)ϕi (z), so
ϕi (y) = ϕi (z) and hence y = z since ϕi is a homeomorphism on B̂i ⊃ Bi .
Next, DF |x is injective for each x ∈ M : If DF |x (v) = 0, then choose i
such that x ∈ Bi , so that ρi = 1 on a neighbourhood of x. It follows that
0 = D(ρi ϕi )(v) = Dϕi (v) and hence v = 0 since ϕi is a diffeomorphism on
Bi (compare Proposition 4.3.1). 

We remark that the same conclusion (with ‘injective immersion’ replacing


‘embedding’ holds if M is not compact, provided we can express M as a
finite union of coordinate charts of the form used in the proof.

7.2. Sard’s theorem and sets of measure zero

To improve the result (and to extend it to the non-compact setting) we need


to use Sard’s Theorem, which concerns sets of measure zero in a manifold.
We first make some definitions:
Definition 7.2.1. Let M be a manifold. A subset A ⊂ M is of measure
zero in an n-dimensional manifold M if ϕ(A ∩ U ) has measure zero in Rn
for every smooth chart ϕ : U → V for M .

Recall that a subset C of Rn has measure zero if for any δ > 0, C can be
covered by a countable union of balls such that the sum of the volumes of the
balls is less than δ. One can check that countable unions of sets of measure
zero have measure zero, and that submanifolds of dimension less than n in
Rn have measure zero.
Making sense of this on a manifold requires the following:
Lemma 7.2.2. If A ⊂ Rn has measure zero, and F : U → Rn is a smooth
map defined on a neighbourhood U of A, then F (A) has measure zero.

Proof. Since U is open, we can write U is a countable union of open


balls Ui each of which has closure contained in U . On each Ui the derivative
of F is bounded (since Ui is compact and contained in U ) so we have
|F (x) − F (y)| ≤ Ci |x − y|
for all x, y ∈ Ui . Since A has measure zero, A ∩ Ui also has measure zero, so
P δ > 0 we can find a collection of balls {Bj } which cover A ∩ Ui such
for any
that j |Bj | < δ where |Bj | is the volume of Bj in Rn . The estimate above
shows that F (Bj ∩ Ui ) has diameter bounded by Ci times the diameter of
Bj , and hence is contained in a ball of volume at most (2Ci )n times |Bj |.
Therefore F (A ∩ Ui ) is covered by a union of balls with total volume at most
7.3. REDUCING DIMENSION OF EMBEDDINGS 53

(2Ci )n δ. Since δ > 0, A ∩ Ui has measure zero. Therefore A is a union of


countable many sets of measure zero, hence has measure zero. 

The Lemma implies in particular that it suffices to check that a subset A of


M has measure zero with respect to some collection of charts covering A,
since the transition maps between charts are diffeomorphisms.
Here is a useful consequence (a special case of Sard’s theorem):
Proposition 7.2.3. If F : M m → N n is a smooth map, and m < n, then
F (M ) has measure zero in N .

Proof. Define F̂ : M × Rn−m → N by F̂ (x, y) = F (x). Then F̂ is


smooth. We observe that M × {0} is a submanifold of dimension less than
n in M × Rn−m , and so has measure zero in M × Rn−m [Exercise. Hint:
First show that a subspace Rk ⊂ Rn has measure zero if k < n, then use
charts to prove that a submanifold must have measure zero]. Therefore
F (M ) = F̂ (M × {0}) has measure zero in N by Lemma 7.2.2. 

More generally, if F : M m → N n is a smooth map, we say that x ∈ M is a


critical point of F if DF |x is not surjective (so in particular, if m < n then
every point is a critical point). We say that y ∈ N is a critical value of F
if y = F (x) for some critical point x of F . Then the full version of Sard’s
theorem is as follows:
Theorem 7.2.4. The set of critical values of a smooth function F : M → N
has measure zero in N .

We will not prove the full version here.

7.3. Reducing dimension of embeddings

We can use the result of Proposition 7.2.3 (or Theorem 7.2.4) to show that
compact manifolds of a given dimension n can be embedded into a Euclidean
space of a dimension depending only on n:
Proposition 7.3.1. If M is a manifold of dimension n and there exists an
injective immersion F of M into some Euclidean space RN , then there also
exists an injective immersion of M into R2n+1 and an immersion of M into
R2n .

Proof. The idea is to show that if F is an injective immersion of M


into RN with N > 2n + 1, then there is a subspace V of dimension N − 1 in
RN such that π ◦ F is also an injective immersion, where π is the orthogonal
projection onto V . Applying this repeatedly gives the desired result, since
54 7. SARD’S THEOREM AND EMBEDDING THEOREMS

we can always reduce the dimension of the ambient space if it is larger than
2n + 1.
Let us examine more closely the conditions required for such a projection to
be an injective immersion: There are two conditions: We need π ◦ F to be
injective, and we need the derivative of π ◦ F to be injective at each point.
The injectivity of π ◦ F is the statement that if x and y are distinct points
in M , then we never have π(F (y)) − π(F (x)) = π(F (y) − F (x))0. The latter
is equivalent to the statement that F (y) − F (x) does not lie on the line Rv,
where v is a unit vector orthogonal to the subspace V we are projecting to.
Since we know F (y) − F (x) 6= 0, this is equivalent to the statement that
[F (y) − F (x)] 6= [v] in RPN −1 for each y 6= x.

To be more precise, we define a map F̂ from M × M \ ∆ to RPN −1 by

F̂ (x, y) = [F (y) − F (x)],

where ∆ = {(x, x) : x ∈ M } which is an n-dimensional submanifold in the


2n-dimensional manifold M ×M . Thus M ×M \∆ is itself a smooth manifold
of dimension 2n, and F̂ ∈ C ∞ (M × M \ ∆, RPN −1 ) (it is the composition
of a smooth map to RN \ {0} with the smooth projection from RN \ {0} to
RPN −1 ). The injectivity of F becomes the requirement that F̂ never takes
the value [v].
Now consider the injectivity of the derivative of π ◦ F : We require that for
w 6= 0 in Tx M , the image D(π◦F )|x (w) = π(DF |x (w)) is not zero. Since the
kernel of π is just Rv, this is equivalent to the statement that DF |x (w) ∈
/ Rv,
and since we know that DF |x is injective and hence DF |x (w) 6= 0, this is
equivalent to the statement that [DF |x (w)] 6= [v]. For all (x, w) ∈ T M .
We can simplify this a little further, since [DF |x (λw)] = [DF |x (w)] for any
λ ∈ R \ {0}, so in fact the map F̃ : (x, [w]) 7→ [DF |x (w)] is well-defined and
smooth on the projectivised tangent bundle P T M = {[w] ∈ PTx M, x ∈ M },
which is a smooth manifold of dimension 2n−1. The condition for injectivity
if the derivative is then that F̃ never takes the value [v].
Putting these two conditions together, in order to choose v such that π ◦ F
is an injective immersion, we simply need to find a point [v] ∈ RPN −1 which
is not in the images of either F̂ or F̃ . Finally, we observe that if N > 2n + 1
then the dimension of M × M \ ∆ is 2n which is less than the dimension
N − 1 of RPN −1 , so the image of F̂ has measure zero in RPN −1 . Also, the
dimension of P T M is 2n − 1, which is also less than N − 1, so the image
of F̃ also has measure zero. Hence the union of these two images is a set of
measure zero in RPN −1 , and hence is not all of RPN −1 . Therefore we can
choose v such that [v] is not in the image of either map, and this guarantees
that π ◦ F is an injective immersion, as required.
7.4. EMBEDDING NON-COMPACT MANIFOLDS 55

The result for (possibly non-injective) immersions is similar, except that


we only need to avoid the image of the map F̃ , so the condition N > 2n
suffices. 

Of course, we would actually like an embedding, but the result we just


proved is often enough in view of Proposition 1.2.13). In particular we have
the following useful corollaries:
Corollary 7.3.2. If M is a compact manifold of dimension n, then there
exists a smooth embedding of M in R2n+1 and an immersion into R2n .

Proof. The immersion result is immediate from the previous propo-


sition. The embedding result follows since an injective immersion from a
compact manifold is an embedding by Proposition 1.2.13. 

In our later argument we will make use of the following variation of the
result:
Corollary 7.3.3. Let K be a compact subset of a manifold M n . Then there
exists an open set U in M containing K and a smooth injective immersion
of U into R2n+1 and an immersion of U into R2n .

Proof. In the proof of Proposition 7.1.1 we choose the charts to cover


K, and take U to be the union of the sets in finite subcover {B2ri (xi )}ki=1
and proceed as before to produce an injective immersion of U . The result
then follows directly by the Proposition. 

7.4. Embedding non-compact manifolds

Now we proceed to the problem of constructing embeddings for non-compact


manifolds.
Theorem 7.4.1 (The ‘easy’ Whitney embedding theorem). Let M be a
smooth manifold of dimension n. Then there exists a proper embedding
of M into R2n+1 and a proper immersion of M into R2n .

Recall that a map between topological spaces is proper if inverse images of


compact sets are compact, which in our setting is equivalent to the statement
that images of sequences which ‘escape to infinity’ (in the sense that the
terms eventually leave any compact subset) also escape to infinity.

Proof. We will begin by dividing the manifold into suitable compact


pieces. To do this, let {Ki }∞
i=1 be an exhaustion of M by compact subsets,
S that each Ki is compact, we have Ki ⊂ Int(Ki+1 ) for each i ≥ i, and
so
i≥1 Ki = M .
56 7. SARD’S THEOREM AND EMBEDDING THEOREMS

It is convenient to consider the compact sets given by V1 = K1 , and Vi = Ki \


Int(Ki−1 ) for i > 1. We also define open sets W1 = Int(K2 ), W2 = Int(K3 ),
and Wi = Int(Ki+1 ) \ Ki−2 for i > 2, so that Vi ⊂ Wi for each i. It is
important for what follows that Vi ∩ Vj = ∅ for |j − i| > 1, and Wi ∩ Wj = ∅
for |j − i| > 2.
For each i we can construct a smooth cut-off function ρi which is equal
to 1 on Ki and vanishes outside Int(Ki+1 ). Let χ1 = ρ1 , χ2 = ρ2 , and
χi = ρi (1 − ρi−2 ) for i > 2, so that χi is a smooth cut-off function which is
equal to 1 on Vi and equal to zero outside Wi .
By Corollary 7.3.3 we can find a smooth injective immersion Fi of a neigh-
bourhood of the compact set Wi ⊂ M into R2n+1 . We can assume without
loss of generality (by adding a constant to Fi and multiplying by a suitable
positive constant) that the image of the map Fi is contained in the unit ball
about the origin in R2n+1 .
Now we can construct an injective immersion into a Euclidean space of
dimension 4n + 4, as follows: Define
!
X X X X
F (x) = χi (x)Fi (x), χi (x)Fi (x), (1 − ρi (x)), (1 − ρi (x)) .
i odd i even i odd i even

Here the first two entries have values in R2n+1 , and the last two are real
valued. Note that the sums is each term are finite sums at each point, since
for any x we have x ∈ Ki and hence 1 − ρi (x) = 0 for large i. In particular
the function defined by the third sum is equal to zero in V1 , between 0 and 1
on Int(V2 ), equal to 1 on V3 , and in general equal to k on V2k+1 and between
k − 1 and k on Int(V2k ). Similarly, the function defined by the last sum is
equal to zero on K2 , has values between 0 and 1 on Int(V3 ), is equal to 1 on
V4 . More generally it is equal to k on V2k and takes values between k and
k + 1 on Int(V2k+1 ) for k ≥ 1.
We prove that F is a proper injective immersion of M , hence a proper
embedding of M :
First we prove injectivity: If F (x) = F (y), then we can choose j such that
x ∈ Vj . If j = 2k + 1 is odd then we have
X
(1 − ρi (x)) = k,
i odd

and hence the sameP is true for y, from which weP deduce that y ∈ Vi also.
This implies that i odd χi (x)Fi (x) = Fj (x) and i odd χi (y)Fi (y) = Fj (y)
and hence Fj (x) = Fj (y) and x = y. The case where j is even is similar.
Next we prove that F is an immersion: Given x ∈ M , choose j such that
x ∈ Vj . Suppose that j is odd. Then if DF |x (v) = 0 for some v ∈ Tx M , we
7.4. EMBEDDING NON-COMPACT MANIFOLDS 57

also have !
X
0=D χi Fi |x (v) = DFj |x (v)
i odd
since χj is equal to 1 on Vj . Since Fj is an embedding, DFj |x is injective
and hence v = 0. The case where j is even is similar.
Next we observe that F is proper: If x ∈ / Ki then x ∈ Vj for some j > i,
and hence either the second-last or last component in the definition of F
has magnitude greater than bj/2c. In particular for any sequence xj which
escapes to infinity, |F (xj )| approaches infinity.
The final step in the proof is to show that the dimension of the ambient
space can be reduced from 4n + 4 to 2n + 1: The fact that we can find an
injective immersion into R2n+1 follows directly from Proposition 7.3.1. It
only remains to show that we can arrange that the resulting map into R2n+1
is proper: The map F constructed above has image lying within bounded
distance of the two-dimensional plane P spanned by the last to components,
since the embeddings Fi all have image in the unit ball, and the image of any
sequence which escapes to infinity in M must have projection onto P ap-
proaching infinity. Provided we choose the projections to be onto subspaces
with normal direction not contained in this two-dimensional subspace, the
projection is non-singular on P , and hence the image of these sequences
still approach infinity and the projection map is proper. The fact that we
can always choose the directions to avoid the subspace P follows from the
fact that the elements of P form a set of measure zero in RPN −1 since the
dimension of the latter is N − 1 > 2n ≥ 2, while the set we must avoid,
corresponding to the projection of the two dimensional subspace P onto
RPN −1 , has dimension 1 and so has measure zero in RPN −1 . 

The result we just proved is called the ‘easy’ Whitney embedding theorem,
to distinguish it from the ‘hard’ Whitney embedding theorem. The latter
combines the argument above with more sophisticated methods from al-
gebraic topology to prove that any manifold of dimension n has a proper
embedding into R2n and a proper immersion into R2n−1 . In particular, any
surface can be properly embedded into four-dimensional space, and properly
immersed into R3 . This is sharp in general: A non-orientable compact sur-
face such as the Klein bottle or RP2 cannot be embedded into R3 (since any
compact embedded surface in R3 bounds a region and so can be oriented).
In higher dimensions the hard Whitney embedding theorem is still not sharp
in general.
CHAPTER 8

Vector fields, Lie brackets, and sections

Contents
8.1. Vector fields 59
8.2. Lie brackets 61
8.2.1. The Lie algebra of vector fields 62
8.2.2. F -related vector fields 63
8.2.3. The Lie algebra of a Lie group 64
8.3. Sections of vector bundles 67
8.4. Bases of smooth sections and local trivialisations 67

8.1. Vector fields

A vector field assigns to each point x of a manifold a tangent vector at


that point. in such a way that it varies smoothly. A precise definition is as
follows:
Definition 8.1.1. Let M be a manifold. A smooth vector field V on M is
a smooth map from M to T M such that π ◦ V is the identity map on M .

Unravelling the definition, this means that for each x ∈ M , V (x) is a point
in T M such that π(V (x)) = x — which means that V (x) ∈ Tx M for each
x. Thus V associates to each x ∈ M a tangent vector V (x) ∈ Tx M at that
point. It is an important part of the definition that V is smooth as a map
from T M to M , since this captures our requirement that the vector vary
smoothly.
Proposition 8.1.2. If V (x) ∈ Tx M for each x ∈ M , then V defines a
smooth vector field if and only if the coefficients of V are smooth in any
chart: That is, if ϕ : U → W is a smooth chart for M , and ∂P 1 , · · · , ∂n are
the coordinate tangent vectors in the chart ϕ, then V (x) = ni=1 V i (x)∂i
with V i ∈ C ∞ (U ) for each i.

Proof. Recall that a chart for T M is given by


n
X
ϕ̃(x, v i ∂i ) = (ϕ(x), (v 1 , · · · , v n )).
i=1
59
60 8. VECTOR FIELDS, LIE BRACKETS, AND SECTIONS

Using this chart we have ϕ̃ ◦ V ◦ ϕ−1 (p) = ϕ̃((ϕ−1 (p), i V i (ϕ−1 (p)∂i )) =
P
(p, (V 1 ◦ ϕ−1 (p), · · · , V n ◦ ϕ−1 (p))). The requirement that V is smooth is
therefore equivalent to the smoothness of the maps V i ◦ ϕ−1 on W for each
i, which is equivalent to the smoothness of V i on U for each i. 

We use the notation X (M ) to denote the vector space of smooth vector fields
on M . Note that multiplying a smooth vector field pointwise by a smooth
function gives a smooth vector field, so the space of smooth vector fields is
in fact a module over the ring (or algebra) of smooth functions on M .
Example 8.1.3. If G is a Lie group and v is a vector in Te G (where e is the
identity element in G), then we can use the maps D`g ||e to define a natural
vector field on G: We set
V (g) = D`g |e (v).
The resulting vector field is called a left-invariant vector field. To check that
this defines a smooth vector field, let us unravel the definition as follows:
Since mult is smooth from G × G to G, the derivative Dmult is smooth from
T G × T G to T G. The left-invariant vector field is then given by
V (g) = Dmult((e, v), (g, 0)).
We observe that the map g ∈ G 7→ ((e, v), (g, 0)) ∈ T G × G is a smooth map
from G to T G × T G, and hence V is a smooth map from G to T G with
π ◦ V = IdG .
Example 8.1.4. Consider the group C \ {0} ' R2 \ {0} under complex
multiplication. The left-invariant vector fields are as follows: Fix v ∈ C, and
consider this as an element of the tangent space T1 C ' C. The corresponding
left-invariant vector field is given by
d
V (z) = D`z |1 (v) = (z(1 + sv)) s=0 = zv.
ds
These form a two-dimensional√vector space, spanned by the vector fields
z 7→ z and z 7→ zi, where i = −1. The plot below shows the vector fields
corresponding to v = 1, v = i and v = (1 + i).

Another useful way to think about smooth vector fields is using derivations:
Definition 8.1.5. Let M be a smooth manifold. Then a derivation on
C ∞ (M ) is an R-linear map V from C ∞ (M ) to C ∞ (M ) such that
V (f g) = (V (f ))g + f (V (g))
for all f, g ∈ C ∞ (M ).
8.2. LIE BRACKETS 61

Proposition 8.1.6. There is a natural isomorphism (as C ∞ (M )-modules)


between smooth vector fields on M and derivations on C ∞ (M ), given by the
correspondence

(V (f ))(x) := (V (x))f,

where V (x) acts as a derivation at x on f for each x ∈ M .

Proof. Suppose that V ∈ X (M ). The only difficult point is to check


that the function V (f ) defined this way is in C ∞ (MP ): This follows since we
can write in the domain of a smooth chart ϕ V (x) = ni=1 V i (x)∂i (x), where
∂1 (x), · · · , ∂n (x) ∈ Tx M are the basis of coordinate tangent P vectors for Tx M ,
and each V i is a smooth function. But then we have V f = ni=1 V i ∂i f , and
∂i f ◦ ϕ−1 = ∂x ∂
i (f ◦ ϕ
−1 ) is a smooth function for each i, so V f is a smooth

function on the domain of the chart.


The R-linearity and derivation properties follow directly from the derivation
property for derivations at x.
Conversely, if V is a derivation on C ∞ (M ), then define a derivation V (x) at
x by V (x)f := (V f )(x). This defines an element of Tx M for each x. Writing
the vector in a chart ϕ, we have V i (x) = (V ϕi )(x) for each i, where ϕi is
the ith component of ϕ (extended to a smooth function on M outside some
neighbourhood). Each component function V i is therefore smooth on the
domain of the chart, and so this defines a smooth vector field. 

8.2. Lie brackets

The characterisation of vector fields as derivations acting on smooth func-


tions gives rise to the following natural definition:

Definition 8.2.1. Let X, Y ∈ X (M ). Then the Lie bracket [X, Y ] of X


and Y is the vector field defined (as a derivation) by

[X, Y ]f := X(Y f ) − Y (Xf ).

The Lie bracket is thus the commutator of X and Y as operators on smooth


functions. To check that this definition makes sense, we need to confirm that
[X, Y ] is in fact a derivation. The R-linearity is immediate since it is linear
combination of compositions of R-linear maps. The derivation property can
62 8. VECTOR FIELDS, LIE BRACKETS, AND SECTIONS

be confirmed as follows:
[X, Y ](f g) = X(Y (f g)) − Y (X(f g))
= X((Y f )g + f (Y g)) − Y ((Xf )g + f (Xg))
= (X(Y f ))g + (Y f )(Xg) + (Xf )(Y g) + f (X(Y g))
− (Y (Xf ))g − (Xf )(Y g) − (Y f )(Xg) − f (Y (Xg))
= (X(Y f ) − Y (Xf ))g + f (Y (Xg) − X(Y g))
= ([X, Y ]f )g + f ([X, Y ]g)
as required.
It is instructive to compute the Lie bracket of two vector fields in Rn : De-
noting by ∂i the ith partial derivative, thought of as Pa derivation (hence a
tangent vector) we have for X = i X i ∂i and Y = i Y i ∂i
P
 
X X
[X, Y ]f = X i ∂i  Y j ∂j f  − (X ↔ Y )
i j
X
= X i ∂i Y j ∂j f + X i Y j ∂i ∂j f
i,j

−Y j ∂j X i ∂i f − X i Y j ∂j ∂i f

X
X i ∂i Y j − Y i ∂i X j ∂j f

=
i,j
2 2
where we used the fact that ∂i ∂j f = ∂x∂i ∂x
f ∂ f
j = ∂xj ∂xi = ∂j ∂i f . That is, since
f is arbitrary,
X
X i ∂i Y j − Y i ∂i X j ∂j .

[X, Y ] =
i,j
Note that the sameP expression holds for vector fields in a manifold written
i
in the form X = i X ∂i with respect to the coordinate tangent vectors of
any chart (since these act simply by partial differentiation in Rn through the
chart), and in particular the Lie bracket of any pair of coordinate tangent
vectors is zero in any chart.

8.2.1. The Lie algebra of vector fields. The Lie bracket operation
on vector fields gives X (M ) the algebraic structure of a Lie algebra:
Definition 8.2.2. A (real) Lie algebra is a real vector space E endowed
with an operation [., .] : E × E → E which is antisymmetric, linear in each
argument and satisfies the Jacobi identity:
[a, [b, c]] + [b, [c, a]] + [c, [a, b]] = 0 for all a, b, c ∈ E.
Proposition 8.2.3. If M is a smooth manifold, then X (M ) is a Lie algebra
under the Lie bracket operation.
8.2. LIE BRACKETS 63

Proof. We check the Jacobi identity: If X, Y and Z are smooth vector


fields on M , and f is a smooth function on M , then
[X, [Y, Z]]f = X([Y, Z]f ) − [Y, Z](Xf )
= X(Y (Zf )) − X(Z(Y f )) − Y (Z(Xf )) − Z(Y (Xf )).
Therefore we have
[X, [Y, Z]]f + [Y, [Z, X]]f + [Z, [X, Y ]]f
= X(Y (Zf )) − X(Z(Y f )) − Y (Z(Xf )) + Z(Y (Xf ))
+ Y (Z(Xf )) − Y (X(Zf )) − Z(X(Y f )) + X(Z(Y f ))
+ Z(X(Y f )) − Z(Y (Xf )) − X(Y (Zf )) + Y (X(Zf )
= 0.


8.2.2. F -related vector fields. It is perhaps not surprising that the


Lie bracket is invariant by diffeomorphisms (since these are simply reparametri-
sations). In fact a stronger statement holds, for which we need the following
definition:
Definition 8.2.4. Let M and N be smooth manifolds, and F ∈ C ∞ (M, N ).
Vector fields X ∈ X (M ) and Y ∈ X (N ) are called F -related if for all x ∈ M ,
DF |x (Xx ) = YF (x) .

In particular, if F is a diffeomorphism, then there is a unique F -related vec-


tor field Y ∈ X (N ) for each X ∈ X (M ), given by Y (z) = DF |F −1 (z) (XF −1 (z) ).
Other examples include the following:
• If Σ is a submanifold in M , then Y ∈ X (M ) is ι-related to X ∈ X (Σ)
(where ι is the inclusion) if Yx = Xx for each x ∈ Σ (that is, Y is a smooth
extension of X from Σ to M ).
• If G is a submersion and Xx ∈ ker(DG|x ) for each x, then the zero vector
field is G-related to X.
The Lie bracket behaves nicely with respect to F -related vector fields, as
shown by the following:
Proposition 8.2.5. Let F ∈ C ∞ (M, N ), and suppose W ∈ X (N ) is F -
related to X ∈ X (M ) and Z ∈ X (N ) is F -related to Y ∈ X (M ). Then
[W, Z] is F -related to [X, Y ].

Proof. F -relatedness can be expressed as follows:


Lemma 8.2.6. If F ∈ C ∞ (M, N ), a vector field X on M is F -related to a
vector field W on N if and only if
X(f ◦ F ) = (W (f )) ◦ F
64 8. VECTOR FIELDS, LIE BRACKETS, AND SECTIONS

for all f ∈ C ∞ (N ).

Proof. If f is a smooth function on N , then f ◦ F is a smooth function


on M , and by Proposition 4.4.2 we have
X|x (f ◦ F ) = (DF |x (X|x ))(f ) = (W |F (x) )(f ) = (W (f )) ◦ F |x ,
so X(f ◦ F ) = (W (f )) ◦ F .
Conversely, if X(f ◦ F ) = (W (f )) ◦ F for all f ∈ C ∞ (N ), then
(DF |x (X|x ))(f ) = X|x (f ◦ F ) = (W (f ))|F (x) = (W |F (x) (f )
for all f ∈ C ∞ (M ), so DF |x (X|x ) = W |F (x) as a derivation for each x, and
therefore W is F -related to X. 

To prove the proposition, we must show that DF ([X, Y ]|x ) = [W, Z]|F (x) for
each x ∈ M . If f is a smooth function on N , then
DF ([X, Y ]f = [X, Y ](f ◦ F )
= X(Y (f ◦ F )) − Y (X(f ◦ F ))
= X((Zf ) ◦ F ) − Y ((W f ) ◦ F ))
= (W (Zf )) ◦ F − (Z(W f )) ◦ F
= ([W, Z]f ) ◦ F
which is equivalent to the statement that [W, Z] is F -related to [X, Y ], by
the Lemma. 

8.2.3. The Lie algebra of a Lie group. An important application of


the previous result on F -related vector fields is in the context of Lie groups:
If G is a Lie group, recall that for each g ∈ G we have a diffeomorphism of
G given by the left translation `g , and the derivative of the left-translation
map is therefore an isomorphism between tangent spaces at different points
of G, which we used to construct left-invariant vector fields by setting
V (g) = D`g |e (v)
for some fixed v ∈ Te G.
Proposition 8.2.7. A vector field X ∈ X (G) is left-invariant if and only if
X is `g -related to itself for every g ∈ G.

Proof. X is `g -related to itself if D`g |h (Xh ) = X`g (h) = Xgh for all
h ∈ G. In particular this implies Xg = D`g |e (Xe ), so if X is `g -related
to itself for all g ∈ G then X is certainly left-invariant. Conversely, if
X is left-invariant then for each h ∈ G we have Xh = D`h |e (Xe ) and so
D`g |h (Xh ) = D`g |h (D`h |e (Xe )) = D(`g ◦ `h )|e (Xe ) = D`gh |e (Xe ) = Xgh , so
X is `g -related to itself for every g. 
8.2. LIE BRACKETS 65

Corollary 8.2.8. If X and Y are left-invariant vector fields on a Lie group


G, then [X, Y ] is also left-invariant.

Proof. If X and Y are left-invariant vector fields, then by Proposition


8.2.7 X is `g -related to X and Y is `g -related to Y , so by Proposition 8.2.5
[X, Y ] is `g -related to [X, Y ] for each g ∈ G. Applying Proposition 8.2.7
again, we deduce that [X, Y ] is left-invariant. 

The space of left-invariant vector fields is isomorphic to the tangent space


Te G, and so is a vector subspace of X (G) of dimension equal to the dimension
of G. Corollary 8.2.8 shows that this subspace is closed under the Lie bracket
operation, and so forms a Lie sub-algebra of the Lie algebra of vector fields.
This is called the Lie algebra of G. This is conventionally written in fraktur
font, i.e. the Lie algebra of G is denoted by g.
Example 8.2.9. If G = GL(n), the group of nonsingular n × n matrix
under matrix multiplication, we can compute the left-invariant vector fields
easily: The tangent space to any point is Mn , the space of n × n matrices,
and the left-translations are simply matrix multiplication on the left. Since
these maps are linear, we have D`M |I (A) = M A for each M ∈ GL(n) and
A ∈ Mn . Thus the left-invariant vector fields are given by
VA (M ) = M A.
The Lie brackets of two such left-invariant vector fields can be computed
as follows: If f is a smooth function onP Mn , then the action of a vector
A = (Aij ) at M on f is given by Af = ni,j=1 Aij ∂M ∂f
ij
. So we have for a
P ∂
left-invariant vector field VA (M ) = i,j,k Mik Akj ∂Mij where Aij is constant

for each i, j, and so (denoting ∂ij := ∂Mij ) we have
[VA , VB ] = Mik Akj ∂ij (Mpq Bqr ) ∂pr − (A ↔ B)
= Mik Akj δip δjq Bqr ∂pr − (A ↔ B)
= Mik Akj Bjr ∂ir − (A ↔ B)
= Mik (Akj Bjr − Bkj Ajr ) ∂ir
= Mik [A, B]kr ∂ir
which is the left-invariant vector field corresponding to V[A,B] where [A, B] =
AB − BA is the matrix commutator of A and B. Thus the Lie bracket on
left-invariant vector fields agrees with the Lie bracket on Mn given by matrix
commutation.
Since the left-invariant vector fields on SL(n) and O(n) are simply the re-
strictions of left-invariant vector fields for GL(n) to the submanifolds given
by SL(n) and O(n), the same is true for the Lie algebras: The Lie algebra
sl(n) of left-invariant vector fields on SL(n) is isomorphic to the tangent
space TI SL(n) (which is the space of trace-free matrices) with Lie bracket
66 8. VECTOR FIELDS, LIE BRACKETS, AND SECTIONS

given by the matrix commutator. We can check directly that the matrix
commutator of two trace-free matrices is trace-free: We have
X X
Trace([A, B]) = (Aik Bki − Bik Aki ) = (Aik Bki − Bki Aik ) = 0,
i,k i,k

where we interchanged i and k in the second sum. Similarly, the Lie algebra
o(n) to O(n) is isomorphic to the space of antisymmetric matrices (which
is the tangent space to the identity TI O(n)) with the Lie bracket given by
the matrix commutator. Again, we can check that the commutator of two
antisymmetric matrices gives an antisymmetric matrix:
[A, B]ij + [A, B]ji = Aik Bkj − Bik Akj + Ajk Bki − Bjk Aki
= Aik Bkj − Bik Akj + (−Akj )(−Bik ) − (−Bkj )(−Aik )
= 0.
Proposition 8.2.10. Let F : G → H be a smooth group homomorphism.
Then DF induces a Lie algebra homomorphism from the Lie algebra g of G
to the Lie algebra h of H.

Note that a map f between Lie algebras g and h is a Lie algebra homomor-
phism if it is linear and preserves Lie brackets: f ([a, b]) = [f (a), f (b)] for all
a, b ∈ g.

Proof. Recall that g is linearly isomorphic to Te G with the corre-


spondence given by V ∈ g 7→ V (e) ∈ Te G, or in the reverse direction by
v ∈ Te G 7→ (g 7→ D`g |e (v)). We can therefore define a linear map f from g
to h by
V ∈ g 7→ V |e ∈ Te G 7→ DF |e (V |e ) ∈ Te H 7→ (ξ 7→ D`ξ |e (DF |e (V |e ))) ∈ h.
We will show that f is a Lie algebra homomorphism.
The group homomorphism property of F says that F (gh) = F (g)F (h) for
all g, h ∈ G. That is, F ◦ `g = `F (g) ◦ F for all g ∈ G. Let X and Y be
left-invariant vector fields on G. Then we have
DF |g (Xg ) = DF |g D`g |e (Xe )
= D(F ◦ `g )|e (Xe )
= D(`F (g) ◦ F )|e (Xe )
= D`F (g) |e DF |e (Xe ).

That is, the left-invariant vector field f(X) on H given by f(X) ξ = D`ξ (DF |e (Xe ))
is F -related to the left-invariant vector field X on G, and is the only left-
invariant vector field on H for which this is true.
If X and Y are two left-invariant vector fields on G (i.e. two elements of
g) it follows that [f(X), f(Y )] is F -related to [X, Y ], and so equals f([X, Y ]),
and we have proved that f is a Lie agebra homomorphism. 
8.4. BASES OF SMOOTH SECTIONS AND LOCAL TRIVIALISATIONS 67

8.3. Sections of vector bundles

The idea of a vector field is a special case of a more general concept:


Definition 8.3.1. Suppose that E is a vector bundle over a manifold M ,
with projection π. Then a section of E is a smooth map ψ from M to E
such that π ◦ ψ = IdM .

That is, a section of E associates to each x ∈ M an element ψ(x) of the


fibre Ex , in such a way that it varies smoothly in x. Note that since we
can multiply a section pointwise by a smooth function to produce another
section, the space of sections of E is a real vector space and a C ∞ (M )-
module. We denote the space of smooth sections of E by Γ(E). In particular,
Γ(T M ) = X (M ). The space Γ(T ∗ M ) of smooth sections of the cotangent
bundle of M is also denoted by Ω1 (M ) and called the space of 1-forms on
M . In particular, every smooth function f ∈ C ∞ (M ) gives rise to a 1-form
df , the differential of f , defined by
df (v) := v(f )
for all v ∈ T M . A trivial example is the space of smooth sections of the
trivial rank 1 vector bundle M × R, which is the same as C ∞ (M ).

8.4. Bases of smooth sections and local trivialisations

In the case of the tangent bundle and vector fields, the smoothness condition
on a vector field V reduces to theP smoothness of the coefficient functions V i
arising in the expression V = i V i ∂i in terms of the basis of coordinate
tangent vectors. For a general vector bundle E a similar criterion arises
from any local trivialisation Ψ for E: Recall that Ψ : π −1 (U ) → U × Rk
is a diffeomorphism which is also a linear isomorphism on each fibre Ex
over a point x ∈ U . From this we can construct a basis for each fibre, by
ξi (x) = Ψ−1 (x, ei ) where ei is the standard basis element for Rk . We call
this a smooth local basis for E. In this notation we have the following:
Proposition 8.4.1. A section ψ of E is smooth if and only if for any local
trivialisation Ψ for E, ψ = ki=1 ψ i ξi where ψ 1 , · · · , ψ k ∈ C ∞ (U ).
P

A smooth local basis determines a local trivialisation of E, and vice versa:


Proposition 8.4.2. Let E be a vector bundle over a manifold M , and sup-
pose that ξ1 , · · · , ξk are smooth sections of E over an open set U in M , such
that for each x ∈ U , {ξ1 (x), · · · , ξk (x)} are a basis for Ex . Then the map
k
X
µ= µi ξi ∈ π −1 (U ) 7→ (π(µ), (µ1 , · · · , µk ) ∈ U × Rk
i=1
is a local trivialisation of E.
68 8. VECTOR FIELDS, LIE BRACKETS, AND SECTIONS

Proof. Exercise. 

In particular we have the following immediate corollary:


Proposition 8.4.3. A vector bundle E over M is trivial if and only if there
exists a global smooth basis of sections of E, i.e. ξ1 , · · · , ξk ∈ Γ(E) such that
ξ1 (x), · · · , ξk (x) is a basis for Ex for each x ∈ M .

In particular it follows that the Möbius bundle is not trivial: Since this is a
rank 1 vector bundle, the proposition shows that triviality is equivalent to
the existence of a non-vanishing section, but we can prove that there are no
non-zero sections: Suppose ψ is such a section. Then we can define a function
on R by [(x, u(x))] = ψ([x]), and u is smooth since ψ is smooth. Since we
assume ψ(0) 6= 0, we have [(1, u(1))] = [(0, u(0)] so that u(1) = −u(0) 6= 0.
But then it follows from the intermediate value theorem that there exists
s ∈ (0, 1) with u(s) = 0 and hence ψ([s]) = 0.
Example 8.4.4 (The hairy ball theorem). The tangent bundle of the two-
dimensional sphere is non-trivial — in fact there does not exist any non-
vanishing vector field on S 2 . To see this, suppose that V is a non-vanishing
vector field on S 2 . We can write V with respect to the charts given by
stereographic projections ϕ± from the north and south poles, for which the
transition map is
z
ϕ+ ◦ ϕ−1
− (z) =
|z|2
for z ∈ R2 \ {0}. In the ϕ− chart we can write Dϕ− (V ) = ρ(z)(cos θ, sin θ)
with ρ 6= 0, and observe that as z → 0 both ρ and θ converge to a constant.
The map θ : R2 → R/Z is well-defined, and in particular since V is contin-
uous the winding number of θ around a circle of radius r is integer-valued,
continuous in r, and is zero for small r, hence is zero for every r.
V z
Now in the chart ϕ+ we have Dϕ+ (V ) = |z| 2 − 2 |z|4 z · V In particular

along the equator (corresponding to |z| = 1 in either chart) the vector


V = ρ(cos θ, sin θ) transforms to ρ(− cos(θ + 2φ), sin(θ + 2φ)) which has
winding number −2. This is impossible if V is non-zero in te chart ϕ+ ,
since then we must also have winding number zero around any circle.

On the other hand we have


Proposition 8.4.5. If G is a Lie group, then T G is trivial.

Proof. If {e1 , · · · , en } is any basis for Te G, the corresponding left-


invariant vector fields form a basis for each tangent space of G. 

It follows that there is no Lie group structure possible on S 2 .


CHAPTER 9

ODE, Lie derivatives and the Frobenius theorem

Contents
9.1. ODE on manifolds 69
9.1.1. Initial value problems and the local flow 69
9.1.2. Sketch of local existence, uniqueness and smoothness 70
9.1.3. The local group property 71
9.1.4. The compact case 72
9.1.5. Non-compact cases: The maximal domain of existence 72
9.2. The flow of a left-invariant vector field 73
9.3. Lie derivatives 74
9.4. Commuting vector fields 75
9.5. The Frobenius theorem 76

9.1. ODE on manifolds

9.1.1. Initial value problems and the local flow. Let V be a


smooth vector field on a manifold, and x0 a point in M . The initial value
problem for V with initial value x0 is the problem of finding a smooth curve
γ : I → M such that γ(0) = x0 and such that the curve ‘follows the vector
field V ’, in the sense that γ 0 (t) = V (γ(t)) for each t ∈ I.
The initial value problem can be understood in more familiar terms: If ϕ
is a smooth chart −1
−1
Pn fori M about the point x0 . Writing γ(t) = ϕ (z(t)) and
V (ϕ (z)) = i=1 V (z)∂i , the initial value problem becomes
d i
(9.1.1) z (t) = V i (z 1 (t), · · · , z n (t)), z i (0) = ϕi (x0 )
dt
for i = 1, · · · , n. This is a system of first order ODE, and applying standard
results gives the following:
Theorem 9.1.1. For any V ∈ X (M ) and any x0 ∈ M there exists a unique
solution of the initial value problem for V with initial condition x0 defined
on some interval containing 0.

We will sketch the argument in the next section. We also know that the
solution depends smoothly on the initial data x0 as well as on the elapsed
time t.
69
70 9. ODE, LIE DERIVATIVES AND THE FROBENIUS THEOREM

We use the following notation: Given a vector field V , we denote the solution
of the initial value problem with initial condition x0 by γ(t) =: ΨV (x0 , t).
The map Ψ is the ‘solution operator’ of the initial value problem, and we
will refer to it as the floe of the vector field V . ΨV is a map defined on
an open subset U of M × R (of the form t− (x0 ) < t < t+ (x0 ) for some
t− (x0 ) ∈ [−∞, 0) and t+ (x0 ) ∈ (0, ∞] for each x0 ∈ M ) and taking values in
M . The smooth dependence of the solution on the initial data and on the
elapsed time t implies that Ψ ∈ C ∞ (U, M ).
We remark that one cannot in general expect to have solutions defined for
all times: The following two examples illustrate how this may fail:
Example 9.1.2. Let M = R, and let V (x) = x2 . The corresponding initial
value problem is
d
x(t) = x(t)2 , x(0) = x0 .
dt
The solution is given by
x0
x(t) = ,
1 − x0 t
for t in the interval (−∞, x−1 −1
0 ) if x0 > 0, t in R if x0 = 0, and t in (x0 , ∞)
if x0 < 0. Clearly there is no way that this interval of existence can be
extended to all of R for x0 6= 0.
Example 9.1.3. Let M = (0, 1), and V (x) = 1 for all x. The solution of
the ODE dxdt = 1 with initial data x(0) = x0 is simply x(t) = x0 + t, but we
can only take the interval of existence to be −x0 < t < 1 − x0 .

9.1.2. Sketch of local existence, uniqueness and smoothness. It


suffices to prove local existence for the solution of the system (9.1.1) in Rn .
We assume that V ∈ C ∞ (U, Rn ) where U is an open set containing z0 .
The standard proof uses Picard iteration on a suitable subset of C 0 (I, Rn ),
where I = (−δ, δ) for some δ > 0 to be chosen, using the iteration
Z t
(T z)(t) = z0 + V (z(s)) ds.
0
Choose R > 0 such that BR (z0 ) ⊂ U , and then choose M such that |F (y)| +
|DF |y | ≤ M for all y ∈ BR (z0 ). Then choosing δ < M/R guarantees that
(T z) ∈ C(I, BR (z0 )) whenever z ∈ C(I, BR (z0 )). Also we have
Z δ
kT z − T wk∞ ≤ |V (z(s)) − V (w(s))| ds ≤ M δkz − wk∞ ,
0
so T is a contraction mapping in C(I, BR (z0 )) provided δM < 1. It follows
that T has a unique fixed point in C(I, BR (z0 )), and hence that the IVP
has a unique solution on I = [−δ, δ] provided δ < min{1, R}/M .
To prove that the solution depends continuously on z0 , we can consider an
R
iteration in C(I × BR/2 (z0 ), BR (z0 )), choosing δ < 2M (guaranteeing that T
9.1. ODE ON MANIFOLDS 71

is well-defined as a map from C(I ×BR/2 (z0 ), BR (z0 )) to itself, and δ < 1/M
(guaranteeing that T is a contraction).
It remains to prove that the solution depends smoothly on z0 . To prove this,
it is enough to prove that the iterates zk := T k z0 in the Picard iteration
remain bounded in C p for each p, since a sequence of function which is
bounded in C p and converges in C 0 is also convergent in C q for 0 < q < p.
If we assume that the Picard iteration begins from z0 (t) equal to the identity
map, then we have
Z t
Dz0 = I and Dzk (t) = I + DV |zk (s) Dzk (s) ds,
0

from which it follows by induction on k that |Dzk (t)| ≤ eM |t| for every
k. Similar k-independent bounds on Dj zk can be obtained by successively
differentiating the equation.

9.1.3. The local group property. A consequence of the uniqueness


of solutions is the following ‘group property’:
Lemma 9.1.4. Let V ∈ X (M ) and x0 ∈ M , and let I be the interval of
existence of the flow of V from x0 . Then for any t in I, the interval of
existence of the flow from ΨV,t (x0 ) is I − t, and for s in this interval we
have
ΨV,s (ΨV,t (x0 )) = ΨV,s+t (x0 ).

Proof. We check that the left-hand side γ(s) := ΨV,s (ΨV,t (x0 )) is a
solution of the same initial value problem as the right-hand side σ(s) :=
ΨV,s+t (x0 ): When s = 0 the two are clearly the same since ΨV,0 is the
identity map on M . Differentiating with respect to s we find that
d
γ 0 (s) = (ΨV,s (ΨV,t (x0 )) = V (ΨV,s (ΨV,t (x0 )) = V (γ(s)).
ds
Similarly we have
d
σ 0 (s) = (ΨV,s+t (x0 )) = V (ΨV,s+t (x0 )) = V (σ(s)).
ds
The uniqueness for the initial value problem then implies that γ(s) = σ(s).

Corollary 9.1.5. For each t ∈ R, the flow ΨV,t of V for time t is a
diffeomorphism between open subsets of M .

Proof. If ΨV,t (x0 ) exists, then by the smooth dependence on initial


conditions ΨV,t is also defined on a neighbourhood of x0 , so the domain U of
ΨV,t is open. Uniqueness of solutions implies that ΨV,t is injective for each
t, and has the smooth inverse ΨV,−t . 
72 9. ODE, LIE DERIVATIVES AND THE FROBENIUS THEOREM

Definition 9.1.6. Let M be a manifold. Then the diffeomorphism group


of M , denoted by Diff(M ), is the space of diffeomorphisms from M to M ,
equipped with the group structure given by the operation of composition.
Corollary 9.1.7. In the case that the flow of V is defined on all of M × R,
the map t 7→ ΨV,t defines a group homomorphism from R (with addition
as the group operation) to Diff(M ) — that is, this defines a one-parameter
subgroup of Diff(M ).

9.1.4. The compact case. An important special case is where the


manifold M is compact, or more generally where the support of the vector
field V is compact: In this case we have a globally defined flow:
Theorem 9.1.8. Suppose that V ∈ X (M ) has compact support (that is,
the set supp(V ) = {x ∈ M : V (x) 6= 0} is compact). Then the flow ΨV is
defined on all of M × R.

Proof. For x ∈
/ supp(V ), the flow of V is constant and hence defined
for all times.
By the local existence theorem, we can cover M by a collection of open sets
{Uα }α∈A and choose positive numbers δα such that there is a local flow Ψα
of V defined on Uα ×(−δα , δα ). Now choose a finite subcover for the support
of V , say {Uα (i)}i=1···k , and let δ = min{δα(i) : i = 1, · · · , k} > 0. Then
define
Ψ(x, t) = Ψα(i) (x, t)
whenever x ∈ Uα(i) and |t| < δ, and Ψx,t = x if x ∈ / supp(V ). This is well-
defined, since if x ∈ Uα(i) ∩ Uα(j) then Ψα(i) (x, t) = Ψα(j) (x, t) since both
are solutions of the same initial value problem. This defined a smooth map
Ψ : M × (−δ, δ) → M which is a flow of V .
Finally, we can extend Ψ to M × R using the local group property: The map
Ψ̃ : M × (−2δ, 2δ) → M defined by Ψ̃(x, t) = Ψ(Ψ(x, t/2), t/2) is a flow
of V defined on M × (−2δ, 2δ), and hence agrees with Ψ on M × (−δ, δ).
Iterating this construction extends Ψ to M × (−2N δ, 2N δ) for any N , hence
to all of M × R. 

The result above is not exhaustive: There are many examples of vector
fields which are not compactly supported, but for which the flow is globally
defined. Examples include vector fields V on Rn with length satisfying
|V (x)| ≤ C(1 + |x|) for some C.

9.1.5. Non-compact cases: The maximal domain of existence.


In the general situation, if the solution of the initial value problem does not
exist for all times, we can characterise the maximal interval of existence:
9.2. THE FLOW OF A LEFT-INVARIANT VECTOR FIELD 73

Proposition 9.1.9. If V ∈ X (M ) and x0 ∈ M is such that the maxi-


mal interval of existence (t− (x0 ), t+ (x0 )) of ΨV,t (x0 ) is bounded above, then
ΨV,t (x0 ) escapes M as t approaches t+ (x0 ), in the sense that for any com-
pact subset K of M , there exists t̄(K, x0 ) < t+ (x0 ) such that t(K, x0 ) < t <
t+ (x0 ) implies ΨV,t (x0 ) ∈
/ K.

There is a similar characterisation of t− (x0 ) if this is finite.

Proof. Suppose that ΨV,t (x0 ) does not escape from M . That is, there
exists a compact subset K in M and a sequence of times tk approaching
t+ (x0 ) from below, such that ΨV,tk (x0 ) is in K for every k. By compactness,
there is a subsequence t0k such that ΨV,t0k (x0 ) converges to a point x̄ ∈ K.
In particular ΨV,t0k (x0 ) is in the domain of a local flow of V defined on
U × (−δ, δ), for sufficiently large k, and therefore ΨV,t = ΨV,t−t0k ◦ ΨV,tk is
defined on [0, t0k + δ). For k large the maximal time of existence is therefore
bigger than t+ (x0 ), which is impossible. 

9.2. The flow of a left-invariant vector field

In this section we will consider the special case of the flow of left-invariant
vector fields on Lie groups, where the left-invariance will allow us to deduce
that a global flow always exists, even where the group is non-compact. We
first observe that the left-invariance of the vector field result in an important
simplification:
Proposition 9.2.1. If V ∈ g, then ΨV,t (g) = `g ◦ ΨV,t (e) for any g ∈ G,
where e is the identity element of G.

Proof. The left-invariance of the vector field means that D`g (V ) = V


for any g ∈ G. Therefore we have
d
(`g ◦ ΨV,t (e)) = D`g (V (ΨV,t (e))) = V (`g ◦ ΨV,t (e)),
dt
so `g ◦ ΨV,t (e) is a solution of the initial value problem for V with initial
value g, hence equals ΨV,t (g). 
Corollary 9.2.2. If G is a Lie group and V ∈ g, then the flow ΨV exists
on all of G × R.

Proof. By the local existence theorem, ΨV,t (e) is defined for |t| < δ for
some δ > 0. By the previous result, the flow of V exists on G × (−δ, δ) and
is given by ΨV,t (g) = `g ◦ ΨV,t (e). As in the compact case, the flow therefore
extends to G × R using the local group property. 
Definition 9.2.3. If G is a Lie group and V ∈ g, then we define a smooth
map exp from g to G by exp(V ) := ΨV,1 (e).
74 9. ODE, LIE DERIVATIVES AND THE FROBENIUS THEOREM

Proposition 9.2.4. The exponential map restricts to one-dimensional sub-


spaces of g to give all one-parameter subgroups of G. That is, for any V ∈ g,
the map t 7→ exp(tV ) is a one-parameter subgroup of G, and every one-
parameter subgroup of G arises in this way. Furthermore, if V ∈ g then the
flow of V is given by ΨV,t (g) = g exp(tV ).

Proof. If V ∈ g, so that V is a left-invariant vector field, then exp(tV ) =


ΨtV,1 (e) = ΨV,t (e). By the group property of the flow we have exp((t +
s)V ) = ΨV,t+s (e) = ΨV,t ◦ ΨV,s (e) = ΨV,t (exp(sV )) = `exp(sV ) ◦ Ψt,V (e) =
`exp(sV ) (exp(tV ) = exp(sV ) exp(tV ), so the map t 7→ exp(tV ) is a one-
parameter subgroup of G.
Conversely, a one-parameter subgroup is a (smooth) group homomorphism
σ from R to G, i.e. a smooth map from R to G such that σ(t+s) = σ(t) σ(s)
for all s, t ∈ R. Differentiating with respect to s at s = 0 for any t gives
σ 0 (t) = D`σ(t) (σ 0 (0)),
so that σ 0 (t) = V (σ(t)) where V is the left-invariant vector field given by
Vg = D`g (σ 0 (0)) for all g. Since σ(0) = e it follows that σ(t) = ΨV,t (e) =
exp(tV ) as claimed.
The result on the flow of the left-invariant vector field V follows directly
from Proposition 9.2.1. 

9.3. Lie derivatives

Since the flow of a vector field produces a family of diffeomorphisms, we have


a natural way to identify tangent spaces at points on a given trajectory of
the flow. This leads to a natural notion of differentiation along the flow of
a vector field, which we can now define:
Definition 9.3.1. Let M be a smooth manifold and X, Y ∈ X (M ). Then
the Lie derivative of Y along the flow of X, denoted by LX Y , is the vector
field defined as follows:
d   
LX Y |x = (DΨX,t |x )−1 YΨX,t (x) t=0 .
dt

 X,t |x is
To make sense of this, we recall that ΨX,t is a diffeomorphism, so DΨ 
−1
an isomorphism between Tx M and TΨX,t (x) M . Thus (DΨX,t |x ) YΨX,t (x)
is an element of Tx M for each t, and we can differentiate this at t = 0 to
produce the Lie derivative.
We will see later a generalisation of this idea to differentiating more general
objects (tensors) along the flow of a vector field.
The next result relates the Lie derivative to the Lie bracket:
9.4. COMMUTING VECTOR FIELDS 75

Proposition 9.3.2. If X, Y ∈ X (M ), then


LX Y = [X, Y ].

Proof. We consider the action of the vector field LX Y on a smooth


function f : We note that by the chain rule, (DΨX,t |x )−1 = DΨ−X,t |ΨX,t (x) ,
and so
(DΨ−X,t |x (Y )) f = Y |ΨX,t (x) (f ◦ Ψ−X,t ),
and so
d
(LX Y )f = ((Y (f ◦ Ψ−X,t )) ◦ ΨX,t ) t=0
dt
= Y (−Xf ) + X(Y f )
= [X, Y ]f.
d
Here we used the fact that dt (f ◦ΨX,t ) = (Xf )◦ΨX,t for any smooth function
f , by definition of the flow. 

9.4. Commuting vector fields

Definition 9.4.1. Two vector fields X and Y on M are said to commute if


[X, Y ] = 0 everywhere (equivalently, if the corresponding differential opera-
tors acting on smooth functions commute).

The main result of this section is that the flows of commuting vector fields
also commute:
Proposition 9.4.2. Let X, Y ∈ X (M ) such that [X, Y ] = 0 everywhere on
M . Then
ΨX,t ◦ ΨY,s = ΨY,s ◦ ΨX,t
whenever both sides are defined.

Proof. For any s, the identity holds for t = 0 (since ΨX,t is the identity
map. We will fix s and x ∈ M , and show that the two sides (evaluated at
x) satisfy the same initial value problem as functions of t. Accordingly, let
F (t) = ΨX,t ◦ΨY,s (x) and G(t) = ΨY,s ◦ΨX,t (x). Then we have F (0) = G(0),
and
F 0 (t) = X(F (t))
by the definition of ΨX . On the other hand we have
G0 (t) = DΨY,s (X(ΨX,t (x))).
To establish the result we need to show that this is equal to X(G(t)) for
each t.
Lemma 9.4.3. If [X, Y ] = 0, then X = ΨY,s -related to itself for every s
(where defined).
76 9. ODE, LIE DERIVATIVES AND THE FROBENIUS THEOREM

Proof. We must show that X(f ◦ ΨY,s ) = (Xf ) ◦ ΨY,s for each s. This
is certainly true for s = 0. Differentiating with respect to s gives
d
(X(f ◦ ΨY,s ) = (Xf ) ◦ ΨY,s ) = (X(Y f )−Y (Xf ))◦ΨY,s = ([X, Y ]f )◦ΨY,s = 0.
ds
The result follows. 

By the Lemma, we have G0 (t) = DΨY,s (X(ΨX,t (x))) = X(ΨY,s ◦ ΨX,t (x)) =
X(G(t)). Therefore F and G satisfy the same initial value problem, and
hence agree. 

This has an important consequence, which tells us when a family of ordinary


differential equations can be simultaneously solved:
Corollary 9.4.4. Let M be a smooth manifold, and X1 , · · · , Xk ∈ X (M ).
Then [Xi , Xj ] = 0 for all i and j if and only if for every x ∈ M there exists
a smooth map F : A → M with F (0) = x and DF |z (ei ) = Xi (F (z)) for all
x ∈ A and i = 1, · · · , k, where A is an open neighbourhood of the origin in
Rk .

Proof. (⇒) If [Xi , Xj ] = 0 for all i and j, then define F (z 1 , · · · , z k ) =


ΨX1 ,z1 ◦ ΨX2 ,z2 ◦ · · · ◦ ΨXk ,zk (x). Then we have (since the flows commute)
∂  
DF |z (ei ) = i ΨXi ,zi ◦ ΨX1 ,z1 ◦ · · · ◦ Ψ \Xi ,zi ◦ · · · ◦ ΨXk ,zk (x) = Xi (F (z))
∂z
for each i.
(⇐) If there exists such a map F then X1 , · · · , Xk are F -related to e1 , · · · , ek ,
ad so [Xi , Xj ] is F -related to [ei , ej ] = 0, and hence equals zero. 

9.5. The Frobenius theorem

Let M be a smooth manifold. Then by a distribution of k-planes on M we


mean a (smoothly varying) choice of a k-dimensional subspace of Tx M for
each point x ∈ M . More formally, at each point x ∈ M the space of k-
dimensional subspaces of the n-dimensional vector space Tx M is a manifold
of dimension k(n − k) which we call the Grassmannian of k-planes in Tx M
and denote by Gk (Tx M ). The union of these over all x ∈ M is a manifold of
dimension k(n − k) + n which we denote by Gk (T M ), and this is a fibration
over M with fibre diffeomorphic to Gk,n . A distribution of k-planes on M
is then a smooth map D from M to Gk (T M ) such that D(x) ∈ Gk (Tx M )
for each x ∈ M .
Definition 9.5.1. A distribution D of k-planes on M is called integrable if
for every x ∈ M there exists a submanifold Σ in M of dimension k containing
x, such that Tz Σ = D(z) for each z ∈ Σ.
9.5. THE FROBENIUS THEOREM 77

Definition 9.5.2. If D is a distribution of k-planes on M , then a vector


field V ∈ X (M ) lies along D if V (x) ∈ D(x) for each x ∈ M . The space of
such vector fields is a vector subspace (and a C ∞ (M )-sub-module) of X (M ),
which we denote by X (D).
Theorem 9.5.3 (The Frobenius theorem). A distribution D of k-planes on
M is integrable if and only if the space X (D) of vector fields along D is
closed under Lie brackets: X, Y ∈ X (D) =⇒ [X, Y ] ∈ X (D).

Proof. (⇐) If D is integrable, then for any x ∈ M there exists a


submanifold Σ with x ∈ Σ such that Tz Σ = D(z) for each z ∈ Σ. Any
X ∈ X (D) is ι-related to the restriction X|Σ , where ι is the inclusion of Σ in
M . Therefore if X, Y ∈ X (D) then [X, Y ] is ι-related to [X|Σ , Y |Σ ] ∈ X (Σ),
and hence [X, Y ](x) ∈ D(x). SInce x is arbitrary, [X, Y ] ∈ X (D). That is,
X (D) is closed under Lie brackets.
(⇒) Suppose X (D) is closed under Lie brackets. Fix x ∈ M and choose
a basis e1 , · · · , en for Tx M such that D(x) = span(e1 , · · · , ek ). Choose a
chart for M about x such that ei = ∂i for each i at the point x. For y in
a neighbourhoodP of x, it follows that there is a basis for D(y) of the form
{Ei = ∂i + np=k+1 Λpi ∂p : i = 1, · · · , k}. We compute the Lie brackets
[Ei , Ej ]:
h i
[Ei , Ej ] = ∂i + Λpi ∂p , ∂j + Λqj ∂q
= (Ei Λqj )∂q − (Ej Λpi )∂p
∈ span(∂k+1 , · · · , ∂n ).
But by assumption [Ei , Ej ] ∈ D = span(E1 , · · · , Ek ), so we must have
[Ei , Ej ] = 0. By Corollary 9.4.4 there exists F : A ⊂ Rk → M with
F (0) = x and DF |z (ei ) = Ei for each i ∈ {1, · · · , k} and z ∈ A. Since
E1 , · · · , Ek are linearly independent, the map F is an immersion, hence
locally an embedding, and the image of F (on a sufficiently small neigh-
bourhood of 0) is a submanifold of M with tagent space given by D. That
is, D is integrable. 
CHAPTER 10

Differentiating vector fields

Contents
10.1. Connections 79
10.1.1. Differentiating vector fields in a chart 79
10.1.2. New structure: Connections 80
10.1.3. Connection coefficients 81
10.1.4. Local nature of connections 82
10.2. Connections on vector bundles 83
10.3. Pullback connection 84
10.4. Connection on vector bundles over intervals 84
10.5. Geodesics 86
10.5.1. The geodesic flow on the tangent bundle 87
10.5.2. The exponential map 88

10.1. Connections

Vector calculus in Rn naturally involves differentiation of vector fields —


for example one is often interested in computing the divergence or curl of
a vector field in R3 . In this section we introduce a new structure on a
manifold: connections on tangent bundles. This is introduced to provide a
way of differentiating vector fields on manifolds, and will later be extended
to more general contexts.

10.1.1. Differentiating vector fields in a chart. In previous situ-


ations where we needed to make sense of differentiation on a manifold, we
simply used charts for the manifold to reduce to the more familiar situation
of calculus on Rn . However this does not work for differentiation of vector
fields: The result is dependent on the chart.
To see this, suppose we have two charts ϕ and η for M neat a point x, ad
let V be a vector field on M . Then we can write V in each chart:
n
Vϕi ∂iϕ = Vηi ∂iη .
X X
V =
i=1 i
79
80 10. DIFFERENTIATING VECTOR FIELDS

These are related as follows: We have ∂iϕ = ◦ ϕ−1 ))ji ∂jη , and so
P
j (D(η

Vηi = Vϕj Λij


where Λij = (D(η ◦ ϕ−1 ))ij .
Differentiating the vector field V in the chart ϕ along a direction v =
j ϕ
P
j v ∂j gives
 
 (v j ∂ ϕ Vϕi ) ∂ ϕ .
X X
Dvϕ V = j i
i j

Converting this to the chart η gives


 

v j Λ`j ∂`η Vϕi  Λki ∂kη = v j Λ`j ∂`η Vϕi Λki ∂kη .
X X X
Dvϕ V = 
k i,j i,j,k

v j Λ`j ∂`η gives


P
On the other hand, differentiating V in the direction v = j
the following:
 

v j Λ`j ∂`η  Vϕi Λki  ∂kη


X X
Dvη V =
j i,k

v j Λ`j ∂`η Vϕi Λki ∂kη


X
=
i,j,k

v j Λ`j Vϕi (∂`η Λki )∂kη .


X
+
i,j,k

The two expressions differ by the second term, which will vanish in general
only if the derivatives of the coefficients Λji vanish — that is, if the transition
maps are affine transformations (i.e. a combination of a linear tranformation
and a translation). This is certainly not true for general smooth charts for
M , so differentiation of vector fields in charts does not make sense indepen-
dent of charts.

10.1.2. New structure: Connections. In order to overcome this dif-


ficulty, we need to introduce extra structure to the manifold. The structure
required to make sense of differentiation of vector fields is called a connec-
tion.
Definition 10.1.1. A connection ∇ is a map from X (M ) × X (M ) to X M ,
denoted (X, Y ) 7→ ∇X Y , satisfying

(i). ∇X Y is C ∞ (M )-linear in X, so that ∇f1 X1 +f2 X2 Y = f1 ∇X1 Y +


f2 ∇X2 Y for any X1 , X2 , Y ∈ X (M ) and f1 , f2 ∈ C ∞ (M );
(ii). ∇X Y is R-linear in Y , so that ∇X (a1 Y1 +a2 Y2 ) = a1 ∇X Y1 +a2 ∇x Y2
for all a1 , a2 ∈ R, X, Y1 , Y2 ∈ X (M );
10.1. CONNECTIONS 81

(iii). ∇ satisfies a Leibniz rule, so that


∇X (f Y ) = (Xf )Y + f ∇X Y
for any X, Y ∈ X (M ) and f ∈ C ∞ (M ).
Example 10.1.2. The best-known Pexample of a connection is the usual one
on Rn , where we define for X = i X i ei and Y = j Y j ej
P
X X X
∇X Y = X i ∂i ( Y j ej ) := X i ∂i Y j ej .
i j i,j
That is, we differentiate the coefficients and set the derivative of the ‘con-
stant’ vector fields ei to be zero.

10.1.3. Connection coefficients. In general, if ∇ is a connection on


a manifold M , then we can completely determine the connection in a chart
once we know what it does to the coordinate tangent vectors {∂i }, since
every smooth vector field is a C ∞ -linear combination of these: We have
∇X Y = ∇X i ∂i (Y j ∂j ) = X i ∂i Y j ∂j + X i Y j ∇∂i ∂j .
In the chart we can define smooth functions Γij k by setting ∇∂i ∂j = k Γij k ∂k .
P

The fact that Γij k is smooth follows since ∇∂i ∂j is a smooth vector field (by
the definition of connection). These are called the connection coefficients or
sometimes the Christoffel symbols of the connection. Once these are known,
we have the general expression
!
X X
i j i k j
∇X Y = X ∂i Y + X Y Γik ∂j
i,j k
for the connection applied to an arbitrary pair of vector fields.
Conversely, if Γij k are smooth functions on the domain of a chart for each
i, j, k, then the expression above defines a connection.
Exercise 10.1.3. Use a partition of unity, to show that every manifold has
a connection (in fact a space of connection of uncountable dimension).
Example 10.1.4. The standard connection on Rn is a particular exam-
ple of a natural construction of a connection on any Lie group, called
the left-invariant connection: Let G be a Lie group. Then given a basis
{e1 , · · · , en } for the tangent space Te G at the identity, we have correspond-
ing left-invariant vector fields Ei |g = D`g |e (ei ) which form a basis for the
Lie algebra g of left-invariant vector fields of G. Thus any vector field
can be P written as a C ∞ (G)-linear combination of E1 , · · · , En : We have
V = i V Ei where V i ∈ C ∞ (G) for each i. We can then define
i
X X
∇v V = ∇v ( V i Ei ) = (v(V i ))Ei .
i i
That, is, we take the connection applied to any of the left-invariant vector
fields to be zero.
82 10. DIFFERENTIATING VECTOR FIELDS

Exercise 10.1.5. Show that the left-invariant connection is a connection.


Example 10.1.6. Another important example arises for submanifolds of
Euclidean space: If Σ is a submanifold of Rn , then for each x ∈ Σ let πx
be the orthogonal projection of Rn onto the tangent space Tx Σ (which we
identify as usual with a subspace of Rn ). A vector field Y on Σ is then a
smooth n
Pnmap αfrom Σ to R which lies in Tx Σ for each x, and we can write
Y = α=1 Y eα , and each Y α ∈ C ∞ (Σ). We define
!
X
∇X Y |x = πx X(Y α )eα .
α
We check that the Leibniz rule holds:
!
X
∇X (f Y ) = πx X(f Y α )eα
α
!
X
α α
= πx (f X(Y ) + X(f )Y )eα
α
!
X
α
= πx f XY eα + X(f )Y
α
= f ∇X Y + X(f )Y
where we used the fact that πx is linear, and that πx (Y ) = Y since Y is
tangential.

10.1.4. Local nature of connections. Although we defined a con-


nection as an operator acting on two vector fields (producing the derivative
of the second in the direction of the first at each point), we most often em-
ploy it as an operator which differentiates a vector field in the direction of
a vector, giving a vector at the same point. To construct this, we simply
define for X ∈ X (M ) and v ∈ Tx M
∇v X := (∇V X) |x
where V is any smooth vector field on M with V (x) = x. This is well-
defined, since if Ṽ is any other such vector field, we have
 
(∇V X) |x − ∇Ṽ X |x = ∇V −Ṽ X |x .
Writing V and Ṽ in a local chat as i V i ∂i and i Ṽ i ∂i , this becomes
P P
!
  X X
∇P (V i −Ṽ i )∂i X = (V i − Ṽ i )∇∂i X |x = (V i (x)−Ṽ i (x))(∇∂i X)|x = 0,
i x
i i

since V (x) = Ṽ (x) = v. Indeed we could have defined a connection as being


a map from T M × X (M ) to T M which preserves fibres and satisfies the
Leibniz rule, but then we would also have needed to include the prescription
that the result depends smoothly on the point.
10.2. CONNECTIONS ON VECTOR BUNDLES 83

From now on we will denote by ∇v X ∈ Tx M the derivative of a vector field


X in the direction of a vector v ∈ Tx M , as defined above.

10.2. Connections on vector bundles

The definition of connection extends naturally from, the tangent bundle to


more general vector bundles:
Definition 10.2.1. Let E be a vector bundle over a manifold k. A con-
nection on E is a map from X (M ) × Γ(E) to Γ(E), taking a vector field
X on M and a smooth section φ of E to a section ∇X φ of E, which is
C ∞ (M )-linear in the first argument, R-linear in the second argument, and
satisfies a Leibniz rule:
∇X (f φ) = X(f )φ + f ∇X φ
for any X ∈ X (M ), φ ∈ Γ(E), and f ∈ C ∞ (M ).

If we have a local chart for M (with coordinate tangent vectors ∂1 , · · · , ∂n )


and a local trivialisation of E (equivalently, a basis of smooth local sections
Ψ1 , · · · , Ψk of E over some open set U of M ), then the connection ∇ is de-
termined by the connection coefficients with respect to his basis, determined
by
Xk
∇∂i Ψα = Γiα β Ψβ .
β=1
αΨ
P
Explicitly, we have for an arbitrary section φ = αφ α the formula
!
X
α
∇ X φ = ∇ i X i ∂i
P φ Ψα
α
X X
α
= X(φ )Ψα + X i φα ∇∂i Ψα
α i,α
 
X X
= X(φα )Ψα + X i φβ Γiβ α  Ψα .
α i,β

Conversely, and collection of smooth functions Γiα β defines a connection on


E on the domain of the chart and the local trivialisation of E.
As before, the C ∞ -linearity in the first argument means that the result is
well-defined as a map from Tx M × Γ(E) to Ex for each x ∈ M , so we can
make sense of differentiating a smooth section Ψ of E at a point x ∈ M in
a direction v ∈ Tx M to obtain an element ∇v Ψ of the fibre Ex of E over x.
Definition 10.2.2. A section φ of E is called parallel (with respect to a
connection ∇) if ∇v φ = 0 for all v ∈ T M .
84 10. DIFFERENTIATING VECTOR FIELDS

10.3. Pullback connection

A special situation (which we will apply in a moment) is that of pull-back


bundles: Recall that if E is a vector bundle over a manifold M , and F is a
smooth map from a manifold Σ to M , then F ∗ E is the vector bundle over
Σ which has fibre at z ∈ Σ equal to the fibre of E at F (z).
Note that if Ψ is a smooth section of E over M (i.e. a smooth map from M
to E preserving fibres), then F Ψ : Σ → F ∗ E defined by F Ψ(z) = Ψ(F (z)) ∈
EF (z) = F ∗ Ez is a smooth section of F ∗ E over Σ.
Recall that if {Ψα } is a basis of local section of E over an open set U of
M , then {F Ψα } is a smooth basis of local section for F ∗ E over the open set
F −1 (U ) in Σ.
Proposition 10.3.1. Let E be a vector bundle over M with connection ∇,
F ∈ C ∞ (Σ, M ), and F E the pull-back bundle over Σ. Then there is a unique
connection F ∇ on F ∗ E such that
F
∇v (Ψ ◦ F ) = ∇DF (v) Ψ.

Proof. Since a basis of smooth local sections for F ∗ E is given by {F Ψα }


where {Ψα } is a basis of smooth local sections forPE over some open set U ,
any section φ of F ∗ E can be written in the form α φαF Ψα , where each of
the functions φα is smooth on F −1 (U )
X
F
∇X φ = F ∇X ( φαF Ψα )
α
X X
= X(φα )F Ψα + φαF ∇X F Ψα
α α
X X
α F
= X(φ ) Ψα + φα ∇DF (X) Ψα .
α α
This proves uniqueness. On the other hand, this formula defines a connection
on F ∗ E, so we also have existence. 

We call the connection F ∇ on F E provided by this proposition the pull-back


connection.

10.4. Connection on vector bundles over intervals

Let I be an interval in R, and E a vector bundle over I with a connection


∇. In this situation T I is one-dimensional, and we have a canonical choice
∂s of basis for the tangent space. Let Ψα (s) ∈ Es , α = 1, · · · , k be a smooth
basis of local section for E. Then the connection is defined by
∇s Ψα = Γα β Ψβ .
10.4. CONNECTION ON VECTOR BUNDLES OVER INTERVALS 85

(we can suppress the third index on the connection coefficients since there
is one a one-dimensional space of tangent vectors).
Using this we can construct a convenient basis of sections for E:
Proposition 10.4.1. For any s0 ∈ I and any Ψ0 ∈ Es0 there exists a unique
parallel section Ψ ∈ Γ(E) such that Ψ(s0 ) = Ψ0 . The map Ps0 ,s : Ψ0 7→ Ψ(s)
is a linear isomorphism from Es0 to Es for each s ∈ I.

We call the map Ψs0 ,s the parallel transport from s0 to s along I (corre-
sponding to the connection ∇).

Proof. The equation ∇s Ψ = 0 can be written as follows:


0 = ∇s Ψ
X
= ∇s ( φα Ψα )
α
 
X X
= ∂s φα + Γαβ  Ψα ,
α β

so we must solve the initial value problem


d α X
φ =− Γβ α φβ , φα (s0 )φα0 .
ds
β

This is a system of linear ODE with smooth coefficients, and so has a unique
solution. The solution Ps0 ,s (φ0 ) depends linearly on the initial data, and has
no kernel (since the unique solution which is zero at s is the zero solution),
and so is a linear isomorphism for each s0 . 

An important instance of this is when E is a vector bundle over a manifold


M , and σ : I → M is a smooth curve. Then the we can parallel-transport
a section of E along the curve σ by applying the construction above in the
pull-back bundle σ ∗ E over I.
Example 10.4.2 (Parallel transport on a submanifold). For a curve σ on
a submanifold Σ in Rn equipped with the submanifold connection, we can
make sense of parallel transport as follows: Given an initial vector V0 ∈
Tσ(0) Σ ⊂ Rn , the requirement that V (s) ∈ Tσ(s) Σ be parallel along σ is that
 
dV
0 = ∇s V = πσ(s) ,
ds
dV
which means that ds is in the normal space to Σ at σ(s) for each s.

Let us consider the example where Σ = S 2 ⊂ R3 : Then the tangent space


Tz S 2 at z ∈ S 2 is given by {v ∈ R3 : v · z = 0}, and the normal space is Rz.
86 10. DIFFERENTIATING VECTOR FIELDS

A vector field V (s) ∈ Tz S 2 is parallel provided


 
dV dV d
= ·z z = (V · z) z − V · σ 0 (s)z = −V (s) · σ 0 (s)z.
ds ds ds
Suppose σ is the great circle around the equator in S 2 , traversed at constant
speed. Then the tangent space at a point σ(s) is spanned by e3 and σ 0 (s),
and the parallel transport takes e3 to e3 and σ 0 (s) to σ 0 (s): V = e3 is
constant in R3 , so dV /ds = 0 which is certainly in the normal space. V = σ 0
gives
V 0 = σ 00 = −σ
which is in the normal space Rσ(s).
Observe that parallel transport around a closed loop does not always bring us
back to where we started in E: For example, consider the parallel transport
of a tangent vector in S 2 around the closed loop given by three quarter
great circles at right angles (i.e. the boundary curve of the region given
by intersecting the positive octant in R3 with S 2 ): Suppose we begin with
V equal to the tangent vector at the start of the first arc. Then parallel
transport keeps this equal to the tangent vector until the end of that arc.
This is then orthogonal to the tangent vector at the start of the second arc,
so parallel transport keeps the vector constant and normal to the tangent
vector until the end of that arc. At that point, V is (minus) the tangent
vector of the third arc, and so remains equal to minus the tangent vector
until the end of the third arc, which is the same as the original starting
point, but the vector V is now rotated by π/2 from the original value.
This example shows that on S 2 there do not exist non-zero parallel vector
fields, since if there was such a vector field if would also be parallel when
restricted to any curve in S 2 , and so would agree with the parallel transport
along any curve.

10.5. Geodesics

If σ : I → M is a smooth curve, and ∇ is a connection on T M , then the


construction above gives a natural connection σ ∇ on the pull-back bundle
σ ∗ T M over I, the sections of which are vector fields along σ, assigning to
each s ∈ I a tangent vector V (s) to M at σ(s).
In particular, for any curve σ, the tangent vector σ 0 (s) defines a section of
σ ∗ T M , and we can compute the derivative ∇s σ 0 of this vector field along σ.

Definition 10.5.1. A geodesic σ in M (with respect to the connection


∇) is a smooth curve for which the tangent vector σ 0 is parallel along σ:
σ ∇ σ 0 = 0.
s
10.5. GEODESICS 87

If we choose a local chart ϕ for M , we can write the geodesic equation in


the coordinates of the chart as follows: If we write σ(s) = ϕ−1 (x(s)), then
P i
we have σ 0 = i dxds ∂i . Differentiating gives
!
X dxi
0 = σ ∇s ∂i
ds
i
X d2 xi X dxi
σ
= ∂i + ∇s ∂i
ds2 ds
i i
X d 2 xi X dxi dxj
= ∂i + ∇∂j ∂i
ds2 ds ds
i i,j
 
X d2 xiX dxj dxk
Γjk i x(t)  ∂i ,

= 
2
+
ds ds ds
i j,k

where i Γjk i ∂i = ∇∂j ∂k defines the connection coefficients of the connec-


P
tion ∇ in the chart ϕ. The geodesic equation is therefore a system of second
order ordinary differential equations for the functions xi (t).
It follows that there exists a unique geodesic σ in M (defined for |s| < δ for
some δ > 0) with initial conditions σ(0) = p ∈ M and σ 0 (0) = v ∈ Tp M .

10.5.1. The geodesic flow on the tangent bundle. The geodesic


equation, as a system of second order differential equations, does not seem
to fit into the machinery of flows of vector fields — this always corresponds
to a system of first order ODE. However, recall that a system of second
order ODE can be rewritten as a system of first order ODE by including the
first derivatives as additional variables: In the case of the geodesic equation
we define new variables ẋi , i = 1, ·, n, and consider the augmented system
d i
x = ẋi ;
ds
d i X
ẋj ẋk Γjk i x(s) .

ẋ = −
ds
j,k

This gives the derivatives of the 2n variables x1 , · · · , xn , ẋ1 , · · · , ẋn as func-


tions of the same variables, and so defines a system of 2n first order ODE.
Geometrically, the variables x1 , · · · , xn , ẋ1 , · · · , ẋn give the coordinates in
T M (in the standard chart ϕ̃ for T M ) of the curve (x(s), x0 (s)). We can
therefore interpret the system of first order ODE as the flow of a vector field
defined on T M , which we call the geodesic flow. Explicitly, we define
X X
ẋj ẋk Γjk i x ∂˙i ,

G|(x,ẋ) = ẋ∂i −
i i,j,k
88 10. DIFFERENTIATING VECTOR FIELDS

where ∂1 , · · · , ∂n , ∂˙1 , · · · , ∂˙n are the coordinate tangent vectors on T M (i.e.


a basis for T (T M P)) corresponding to the chart ϕ̃ for T M (recall that this is
defined by ϕ̃(p, i v ∂i (p)) = (ϕ(p), (v 1 , · · · , v n ))).
i

Exercise 10.5.2. Show that the formula above defines a vector field G ∈
Γ(T (T M )) (that is, show that it is well defined independent of the chart ϕ).

It follows from our results on flows of vector fields that there exists a smooth
local flow of G defined on an open neighbourhood of the zero section in
T (T M ), and taking values in T M . By construction, the trajectories of the
flow are of the form (σ(t), σ 0 (t)) where σ is a geodesic of the connection ∇.

10.5.2. The exponential map. The exponential map (corresponding


to a connection ∇ on M ) is the map from T M to M defined by
exp(p, v) = π ◦ ΨG,1 (p, v).
Since ΨG,t (p, v) = (σ(t), σ 0 (t)) where σ is the geodesic with initial point
σ(0) = p and initial direction σ 0 (0) = v, the expential map gives geodesics
directly:
σ(t) = exp(p, tv)
is the geodesic which starts at p with direction v.
Exercise 10.5.3. Show that if G is a Lie group and ∇ the left-invariant
connection, then
exp∇ (p, D`p (v)) = p expG (tv)
where exp∇ is the exponential map of the connection ∇ as defined above,
and expG is the exponential map of the Lie group.
Proposition 10.5.4. Let p ∈ M , and define expp : Tp M → M by expp (v) :=
exp(p, v) (where this is defined). Then expp is a diffeomorphism from a
neighbourhood of the origin in Tp M to a neighbourhood of p in M .

Proof. expp is the restriction of the smooth map exp to the fibre Tp M ,
which is a submanifold of T M , so is smooth. We compute the derivative at
the origin:
d d
D expp |0 (v) = expp (sv) = σ(s)|s=0
ds ds
where σ(s) = exp(sv) is the geodesic starting from p with σ 0 (0) = v. That
is,
D expp |0(v) = σ 0 (0) = v,
so D expp |0 is the identity map from Tp M to Tp M , which is an isomorphism.
By the inverse function theorem, expp is a local diffeomorphism. 
10.5. GEODESICS 89

Since Tp M is a vector space of dimension n, hence isomorphic to Rn , expp


(restricted to a neighbourhood of the origin) gives a diffeomorphism from
an open set in Rn to a neighbourhood of p in M . That is, exp−1 p is a
smooth chart for M . The corresponding coordinates are called exponential
coordinates for M about p (corresponding to the connection ∇).
CHAPTER 11

Riemannian metrics and the Levi-Civita


connection

Contents
11.1. Riemannian metrics 91
11.2. Left-invariant metrics 92
11.3. The induced metric on a submanifold 92
11.4. Metrics on vector bundles 93
11.5. Compatibility of metric and connection 93
11.6. Symmetry 95
11.7. The Levi-Civita Theorem 97

11.1. Riemannian metrics

In this section we begin the study of Riemannian geometry, which concerns


the properties of manifolds equipped with a Riemannian metric, defined as
follows:
Definition 11.1.1. A Riemannian metric on a manifold M is a map which
assigns to each point x ∈ M an inner product gx on the tangent space Tx M ,
varying smoothly in the sense that gx (Ux , Vx ) is smooth whenever U and V
are smooth vector fields on M .

A Riemannian metric provides a way of measuring the lengths of tangent


vectors and angles between tangent vectors, and this leads to a wealth of
geometric concepts: For example we can define the length of a smooth path
σ ∈ C ∞ ([0, 1], M ) as follows:
Z 1
L[σ] = gσ(s) (σ 0 (s), σ 0 (s))1/2 ds.
0
This in turn allows us to define a distance function on M (provided M is
path-connected), called the Riemannian distance function:
d(p, q) = inf{L[σ] : σ ∈ C ∞ ([0, 1], M ), σ(0) = p, σ(1) = q}.
Exercise 11.1.2. Show that the Riemannian distance is unchanged if we
replace ‘σ ∈ C ∞ ([0, 1], M )’ with the requirement that σ is piecewise smooth,
91
92 11. RIEMANNIAN METRICS AND THE LEVI-CIVITA CONNECTION

i.e. σ is continuous, and there are finitely many points 0 = t0 < t1 < · · · <
tN = 1 such that the restriction of σ to each subinterval [ti=1 , ti ] is smooth.
Exercise 11.1.3. Verify that d is a distance function (i.e. it is symmetric,
non-negative and zero only when p = q, and satisfies the triangle inequality).

11.2. Left-invariant metrics

An important example of a Riemannian metric is a left-invariant metric on


a Lie group G, defined as follows: Let ge be an inner product on the vector
space Te G, and define for each h ∈ G
gh (u, v) = ge D`−1 −1

h (u), D`h (v)
for all u, v ∈ Th G.
A special case of this is an inner product on a vector space, where the group
operation is addition.
Left-invariant metrics have the special property that the inner product of
a pair of left-invariant vector fields is constant over G, and that the left
translations `h are isometries according to the following definition:
Definition 11.2.1. A diffeomorphism T : (M, g) → (N, h) between Rie-
mannian manifolds is an isometry if
hT (z) (DT |z (u), DT |z (v)) = gz (u, v)
for all z ∈ M , and u, v ∈ Tz M .

An isometry is thus a map which preserves the smooth structure of the


manifold and the geometric structure provided by the Riemannian metric.
The natural notion of equivalence in Riemannian geometry is equivalence
up to isometry.

11.3. The induced metric on a submanifold

Another important example of a Riemannian metric is the induced metric


on a submanifold in Rn : If Σ is a submanifold in Rn , then tangent vectors
to Σ are vectors in Rn , and we can compute their inner product using the
usual Euclidean inner product on Rn :
gx (u, v) = u · v
for all x ∈ Σ and u, v ∈ Tx M ⊂ Rn .
More generally, suppose that F : M → N is an immersion, and h is a
Riemannian metric on N . Then we can induce a Riemannian metric on M
by defining
gx (u, v) = hF (x) (DF |x (u), DF |x (v))
11.5. COMPATIBILITY OF METRIC AND CONNECTION 93

for each x ∈ M and u, v ∈ Tx M .

11.4. Metrics on vector bundles

The notion of a Riemannian metric is a special case of a more general notion,


that of a metric on a vector bundle:
Suppose that E is a vector bundle over a manifold E. A metric g on E is
defined by an inner product gx on each fibre Ex of E, varying smoothly in
the sense that the function g(ξ, η) is smooth whenever ξ and η are smooth
section of E.

11.5. Compatibility of metric and connection

Definition 11.5.1. Let E be a vector bundle over a manifold M , with a


metric g and a connection ∇. We say ∇ and g are compatible if for any
smooth section ξ and η of E and any X ∈ T M ,

X(g(ξ, η)) = g(∇X ξ, η) + g(ξ, ∇X η).

Lemma 11.5.2. Suppose that g and ∇ are a compatible metric and connection
on a vector bundle E over a manifold N , and let F ∈ C ∞ (M, N ). Define
a metric F g on F ∗ E by F g|x = g|F (x) . Then F g and F ∇ are compatible on
F ∗ E.

Proof. Let {Ψα }α=1,··· ,k be smooth local sections giving a basis for
E at each point in some neighbourhood U of N . Then {Ψα ◦ F }α=1,··· ,k
provides a basis of smooth local sections of the pull-back bundle F ∗ E, and
any smooth section can be written in the form

X
ξ= ξ α Ψα ◦ F.
α
94 11. RIEMANNIAN METRICS AND THE LEVI-CIVITA CONNECTION

Given two such sections ξ and η, we compute for X ∈ T M


X F g(ξ, η) − F g(F ∇X ξ, η) − F g(ξ, F ∇X η)
 
X
=X ξ α η β F g(Ψα ◦ F, Ψβ ◦ F 
α,β
 ! 
X X
− F g F ∇X ξ α Ψα ◦ F , η β Ψβ ◦ F 
α β
  
X X
− Fg  ξ α Ψα ◦ F, F ∇X  η β Ψβ ◦ F 
α β
X   
= X(ξ α )η β + ξ α X(η β ) g(Ψα , Ψβ ) ◦ F + ξ α η β DF (X)(g(Ψα , Ψβ ))
α,β
X  
F
− g X(ξ α )Ψα ◦ F + ξ α (∇DF (X) Ψα ) ◦ F, η β Ψβ ◦ F
α,β
X  
F
− g ξ α Ψα ◦ F, X(η β )Ψβ ◦ F + η β (∇DF (X) Ψβ ) ◦ F
α,β
X
ξ α η β DF (X)(g(Ψα , Ψβ )) − g(∇DF (X) Ψα , Ψβ ) − g(Ψα , ∇DF (X) Ψβ ) ◦ F

=
α,β
= 0,
since the last bracket vanishes by the compatibility of g and ∇. 

Compatibility is a very useful property: For example it gives the following


consequence:
Proposition 11.5.3. If g and ∇ are a compatible metric and connection
on a vector bundle E over M , then parallel sections have constant inner
products: If ξ and η are parallel sections of E, then the function g(ξ, eta) is
locally constant. In particular, parallel transport along a curve with respect to
a connection ∇ preserves inner products with respect to a compatible metric
g.

Proof. We have by compatibility that


X(g(ξ, η) = g(∇X ξ.η) + g(ξ, ∇X η) = 0
since ξ and η are parallel, so ∇X ξ = ∇X η = 0. 

An important example of a compatible metric and connection arises in the


submanifold case, which we now consider in slightly greater generality: Sup-
pose that F : Σ → M is an immersion (this includes the case where Σ is a
submanifold and F is the inclusion map), and suppose we have a connection
11.6. SYMMETRY 95

¯ on T M and a compatible Riemannian metric ḡ Then we define induced



metric and connection on T Σ by

gx (u, v) = ḡF (x) (DF |x (u), DF |x (v))

and
F ¯ u (DF (V ))

DF |x (∇u V ) = πx ∇
where πx is the orthogonal projection (with respect to ḡ) of TF (x) M onto
the tangent subspace DF |x (Tx Σ). This includes as a special case the sub-
manifold connection we wrote down earlier for submanifolds in Euclidean
space.

Proposition 11.5.4. The induced metric g and connection ∇ on T Σ are


¯ are compatible.
compatible if ḡ and ∇

Proof. We compute

X(g(Y, Z)) = X(F ḡ(DF (Y ), DF (Z)))


¯ X DF (Y ), DF (Z)) + F ḡ(DF (Y ), F ∇
= F ḡ(F ∇ ¯ X DF (Z))
= F ḡ(πx F ∇¯ X DF (Y ) , DF (Z)) + F ḡ(DF (Y ), πx (F ∇
¯ X DF (Z)))


= F ḡ(DF (∇X Y ), DF (Z)) + F ḡ(DF (Y ), DF (∇X Z))


= g(∇X Y, Z) + g(Y, ∇X Z).

Exercise 11.5.5. Show that the left-invariant connection is compatible with


any left-invariant metric on a Lie group.

11.6. Symmetry

Another special property of connections arises only in the case of connections


on the tangent bundle: A connection ∇ on T M is called symmetric if

∇X Y − ∇Y X = [X, Y ]

for all smooth vector fields X and Y on M .

Proposition 11.6.1. The symmetry of ∇ is equivalent to the statement that


in any chart,
Γij k = Γji k
where Γij k are the connection coefficients defined by ∇∂i ∂j = Γij k ∂k .
P
k
96 11. RIEMANNIAN METRICS AND THE LEVI-CIVITA CONNECTION

Proof. This condition P


is clearly necessaryPsince [∂i , ∂j ] = 0. To prove
the converse, we write X = i X i ∂i and Y = j Y j ∂j and compute
X 
∇X i ∂i Y j ∂j − ∇Y j ∂j X i ∂i − [X i ∂i , Y j ∂j ]
 
∇X Y − ∇Y X − [X, Y ] =
i,j
!
X X
i j j k
= X ∂i Y ∂j + Y Γij ∂k
i,j k
!
X
−Y j ∂j X i ∂i + X i Γji k ∂k
k

i j j i

− X ∂i Y ∂j − Y ∂j X ∂i
X
= X i Y j (Γij k − Γji k )∂k
i,j,k
= 0.

The usual connection on Rn of course satisfies the symmetry condition (since


Γij k = 0 in the usual coordinates).

Exercise 11.6.2. If G is any Lie group such that the associated Lie al-
gebra has a non-trivial Lie bracket operation, show that the left-invariant
connection is not symmetric.
¯ is a symmetric connection on M , and F : Σ →
Proposition 11.6.3. If ∇
M is an immersion, then the induced connection on T Σ is symmetric.

Proof. If X and Y are tangent vector fields on Σ, then we can (locally)


find vector fields X̃ and Ỹ on M such that X̃ is F -related to X and Ỹ is
F -related to Ỹ . Then we have

DF (∇X Y − ∇Y X) = π F ∇X DF (Y ) − F ∇Y DF (X)

 
= π ∇X̃ Ỹ − ∇Ỹ X̃
 
= π [X̃, Ỹ ]
= π ◦ DF ([X, Y ])
= DF ([X, Y ]),

where we used the fact that [X̃, Ỹ ] is F -related to [X, Y ], and that π leaves
tangent vectors unchanged. 
11.7. THE LEVI-CIVITA THEOREM 97

11.7. The Levi-Civita Theorem

In the submanifold case (for submanifolds of Euclidean space), we therefore


have an induced Riemannian metric g, and an induced connection ∇ which
is both compatible with g and symmetric. This turns out to be an important
characterisation:

Theorem 11.7.1 (The Levi-Civita Theorem). Let g be a Riemannian metric


on a manifold M . Then there exists a unique connection ∇ on T M which
is symmetric and compatible with g.

We call this connection the Riemannian connection or the Levi-Civita con-


nection.

Proof. We will prove uniqueness by finding a formula for the connec-


tion. More precisely, let X, Y and Z be smooth vector fields on M . Then
we will find a formula for g(∇X Y, Z): By assumption, we have that ∇ is
compatible with g, so that

Zg(X, Y ) = g(∇Z X, Y ) + g(X, ∇Z Y );

This is also true when we interchange any two of the vector fields. Further-
more, we have the symmetry condition which says that

g(∇X Y − ∇Y X, Z) = g([X, Y ], Z).

Again, this holds if we interchange any two of the vector fields. Therefore
we have
1 1 1
g(∇X Y, Z) = g(∇X Y, Z) + g(∇Y X, Z) + g([X, Y ], Z)
2 2 2
1 1 1 1 1
= Xg(Y, Z) − g(Y, ∇X Z) + Y g(X, Z) − g(X, ∇Y Z) + g([X, Y ], Z)
2 2 2 2 2
1 1 1
= Xg(Y, Z) + Y g(X, Z) + g([X, Y ], Z)
2 2 2
1 1
− g(Y, ∇Z X + [X, Z]) − g(X, ∇Z Y + [Y, Z])
2 2
1 1 1
= Xg(Y, Z) + Y g(X, Z) − Zg(X, Y )
2 2 2
1 1 1
+ g([X, Y ], Z) − g([X, Z], Y ) − g([Y, Z], X).
2 2 2
Note that the expression on the right-hand side depends only on g and not
on ∇, so we have a formula to compute the connection. It remains to prove
that this formula defines a connection: That is, we need to check that this
expression is C ∞ (M )-linear in X and Z, and satisfies a Leibniz rule in Y .
98 11. RIEMANNIAN METRICS AND THE LEVI-CIVITA CONNECTION

We check the Leibniz rule (the others are similar):


1 1
g(∇X (f Y ), Z) = f g(∇X Y, Z) + (Xf )g(Y, Z) − (Zf )g(X, Y )
2 2
1 1
+ (Xf )g(Y, Z) + (Zf )g(Z, X)
2 2
= f g(∇X Y, Z) + (Xf )g(Y, Z)
= g(f ∇X Y + (Xf )Y, Z).
Finally, the fact that the connection defined by this formula is symmetric
and compatible with g can be checked directly. 

The Levi-Civita provides a canonical choice of connection associated to any


Riemannian metric. In particular this means that we can make sense of
parallel transport, geodesics, and the exponential map for Riemannian man-
ifolds. We will explore these further in the next few lectures.
CHAPTER 12

Riemannian geodesics and completeness

Contents
12.1. Riemannian geodesics 99
12.2. The Gauss Lemma and length-minimizing curves 100
12.3. The metric in exponential coordinates 101
12.4. Completeness and the Hopf-Rinow Theorem 102

12.1. Riemannian geodesics

In the last lecture we proved the Levi-Civita theorem, which provides a


unique symmetric connection compatible with any given Riemannian metric.
It follows that a Riemannian metric on a manifold M leads to a well-defined
parallel transport operator along any curve, and also to unique geodesics
starting at any point with any given initial velocity. The corresponding
exponential map is well-defined on some open neighbourhood of the zero
section in T M , and the restriction of this to any tangent space Tp M defines
coordinates on a neighbourhood of p which are called Riemannian normal
coordinates.
EXAMPLE: Submanifolds, geodesics on the plane and the sphere
Riemannian geodesics are curves with zero acceleration as measured by the
Riemannian connection (equivalently, for which the tangent vector is parallel
along the curve). These have some properties which are particularly nice
(compared to those of geodesics for an arbitrary connection):
Proposition 12.1.1. Let γ : I → M be a Riemannian geodesic for a
Riemannian metric g. Then the length (measured by g) of the tangent vector
γ 0 is constant along γ.

Proof. We compute
d
g(γ 0 (s), γ 0 (s)) = 2g(γ 0 (s), ∇s γ 0 (s)) = 0,
ds
since ∇s γ 0 = 0 by the geodesic equation. 
99
100 12. RIEMANNIAN GEODESICS AND COMPLETENESS

12.2. The Gauss Lemma and length-minimizing curves

The Gauss Lemma gives some further very important properties of Rie-
mannian geodesics:
Lemma 12.2.1. Let F (s, t) = expp (sv(t)), where v ∈ C ∞ (I, Tp M ) with
gp (v(t), v(t)) = 1 for each t. Then
gF (s,t) (∂s F, ∂t F ) = 0
for all t ∈ I and s in the domain of definition of F .

Proof. First, we note that ∂t F (0, t) = D expp |0 (0) = 0, so we certainly


have gF (0,t) (∂s F, ∂t F ) = 0. Then we compute

g(∂s F, ∂t F ) = g(∇s ∂s , ∂t ) + g(∂s , ∇s ∂t )
∂s
= 0 + g(∂s , ∇t ∂s ) by symmetry
= ∂t g(∂s , ∂s ) − g(∇s ∂s , ∂t )
= 0,
since we know by Proposition 12.1.1 that g(∂s , ∂s ) = 1 for all s and t. 

An important application of the Gauss Lemma is the following result, which


relates geodesics to curves which achieve the distance between their end-
points:
Proposition 12.2.2. Let M be a manifold with a metric g, and fix p ∈ M .
Suppose R > 0 is such that the exponential map expp is a diffeomorphism
on BR (0) ⊂ Tp M . Then for any v ∈ Tp M with v| = 1 and any r ∈ (0, R),
d(p, expp (rv)) = r,
and the distance from p to expp (rv) is achieved by the geodesic s 7→ expp (sv),
and only by this curve (up to reparametrisation).

Proof. We will prove that any other piecewise smooth curve joining p
to q = expp (rv) has length greater than r. Let σ : [0, 1] → M be such a
curve.
Let [0, s̄] be the connected component of 0 in the closed set σ −1 (expp (B̄r (0)),
so that s̄ is the first time at which σ(s) is in the set expp (Sr (0)). Thus on
the interval (0, s̄) we can write uniquely σ(s) = expp (r(s)v(s)) where v(s) ∈
S1 (0) ⊂ Tp M . Define F (s, t) = expp (sv(t)) for s ∈ (0, r̄) and t ∈ (0, r̄).
Then we have σ(s) = F (r(s), s) and so
σ 0 = r0 ∂s F + ∂t F,
and so by the Gauss Lemma we have
gσ(s) (σ 0 (s), σ 0 (s)) = (r0 )2 g(∂s , ∂s ) + g(∂t , ∂t ) ≥ (r0 )2 ,
12.3. THE METRIC IN EXPONENTIAL COORDINATES 101

since we have g(∂s , ∂s ) = 1 by Proposition 12.1.1 and g(∂s , ∂t ) = 0 by the


Gauss Lemma. It follows that
Z r̄ Z r̄
0 0 1/2
L[σ] ≥ g(σ , σ ) ds ≥ |r0 | ds ≥ r(r̄) − r(0) = r.
0 0
Equality holds only if r̄ = 1 and |r0 | = r0 for each s and ∂t F = 0 for each s.
The latter implies that v(s) is constant since expp is a diffeomorphism (so
D expp is injective). Therefore equality holds if and only if σ(s) = expp (r(s)v
where r(s) is an increasing function — equivalently, σ is a reparametrisation
of the radial geodesic s 7→ expp (sv). 

It follows from Proposition 12.2.2 that any curve which attains the distance
between its endpoints must be a geodesic (up to reparametrisation), since
(by the triangle inequality) it must also achieve the distance between its
endpoints when restricted to any subinterval, and for short subintervals the
Proposition applies.

12.3. The metric in exponential coordinates

The Riemannian exponential coordinates have very nice properties:


Proposition 12.3.1. Let (M, g) be a Riemannian manifold, and p ∈ M .
Choose R > 0 such that expp is a diffeomorphism on BR (0) ⊂ Tp M .
Choose an orthonormal basis {e1 , · · · , en } for Tp M . Then define a chart
ϕ : expp (BR (0)) → BR (0) ⊂ Rn by ϕi (q) = gp (exp−1 p (q), ei ). In this chart
we have ϕ(p) = 0, ∂i (p) = ei for each i, gij (p) = δij , and
∂k gij (p) = 0
for each i, j, k. Furthermore the connection coefficients satisfy Γij k (p) = 0
for each i, j, k.

Proof. We have that expp (tz) is a geodesic for each z ∈ Tp M , so


0 = ∇t ∂t |t=0 = ∇z i ∂i z j ∂j |p = z i z j Γij k (p)∂k .


Choosing z = ei gives 0 = Γii k (p) for each i and k. If i 6= j then choosing


z = ei + ej gives
0 = Γii k (p) + Γjj k (p) + Γij k (p) + Γji k (p) = 2Γij k (p),
where we used Γii k (p) = Γjj k (p) = 0 and Γij k = Γji k (by symmetry of
the connection). This shows that all connection coefficients vanish at p. It
follows that
∂k gij = g(∇k ∂i , ∂j ) + g(∂i , ∇k ∂j ) = Γki ` g`j + Γkj ` g`i = 0
for any i, j, k at the point p. 
102 12. RIEMANNIAN GEODESICS AND COMPLETENESS

12.4. Completeness and the Hopf-Rinow Theorem

In this section we will prove a fundamental result about the Riemannian


distance function: Provided the space is complete as a metric space, we can
always find a shortest path (a length-minimizing geodesic) between any pair
of points. This is a particular consequence of the following result:

Theorem 12.4.1 (The Hopf-Rinow Theorem). Let (M, g) be a connected


Riemannian manifold. Then the following are equivalent:

(i) M is complete with respect to the Riemannian distance d;


(ii) The exponential map is defined on all of T M ;
(iii) There exists p ∈ M such that expp is defined on Tp M .

Furthermore, (iii) implies:


(∗)p : For every q ∈ M there exists a unit speed geodesic γ : [0, d(p, q)] → M
with γ(0) = p and γ(d(p, q)) = q (in particular, γ attains the distance from
p to q).

Proof. (i) =⇒ (ii): Fix any v ∈ T M with g(v, v) = 1 and consider


the geodesic γ : s 7→ exp(sv). We show that this is defined for all s > 0:
If not, then the maximal time of existence s̄ is finite. Note that γ(s) is
Cauchy with respect to the Riemannian distance d as s approaches s̄, since
R s0 p
d(γ(s), γ(s0 )) ≤ L[γ|[s,s0 ] = s g(γ 0 , γ 0 ) = |s−s0 |. Since d is complete, γ(s)
converges to a point q ∈ M as s → s̄. In exponential coordinates about q,
we have Γij k (q) = 0 and therefore |Γij k (expq (z))| ≤ C|z| for |z| < R small,
for some C. By the geodesic equation we have γ 0 = i ẋi ∂i , where
P


X i j k
Γjk ẋ ẋ ≤ C|s − s̄||ẋ|2 ≤ C|s − s̄|

|∂s ẋ| = −
j,k

since |ṡ| ∼ |γ 0 | = 1. Therefore ẋ(s) is Cauchy as s → s̄, and so γ 0 (s)


converges to v ∈ Tq M as s → s̄. But now applying the local existence
theorem for geodesics at (q, v) extends the geodesic to a larger time interval,
contradicting the maximality of s̄.
(ii) =⇒ (iii): This is immediate.
((iii) and (∗)p ) =⇒ (i): We must show that M is complete with respect to
the RIemannian distance. Let (qn ) be a Cauchy sequence with respect to d.
Then in particular (qn ) is bounded, so there exists R such that d(p, qn ) ≤ R
for all n. By (∗)p , for each n there exists vn ∈ S1 (0) ⊂ Tp M such that
qn = expp (dn vn ) where dn = d(p, qn ) ≤ R. That is, qn ∈ expp (B̄R (0)). Now
expp is continuous (since it is defined on Tp M ), and B̄R (0) is compact, so
12.4. COMPLETENESS AND THE HOPF-RINOW THEOREM 103

expp (BR (0)) is compact, hence complete. Therefore qn converges to some


q ∈ expp (BR (0)), and M is complete with respect to d.
There is one step remaining, which takes most of the work:
(iii) =⇒ (∗)p : First choose r > 0 such that expp is a diffeomorphism
on B2r (0) ⊂ Tp M . Then the sphere Sr (0) is compact in Tp M , and the
function rz 7→ d(expp (rz), q) is continuous (since expp is continuous and d is
continuous), so there exists z0 ∈ S1 (0) ⊂ Tp M such that d(expp (rz0 ), q) =
d(expp (rS1 (0), q). Define γ(s) = expp (sz0 ), and note that γ(s) is defined for
all s by the assumption (iii).
Lemma 12.4.2. d(p, q) = r + d(γ(r), q).

Proof. Every piecewise smooth path σ from p to q passes through


expp (Sr (0)), so we have
L[σ] ≥ d(p, expp (Sr (0)) + d(expp (Sr (0)) + d(Sr (0), q) = r + d(γ(r), q).
Taking the infimum over σ gives d(p, q) ≥ r + d(γ(r), q). The reverse in-
equality r + d(γ(r), q) ≥ d(p, q) follows from the triangle inequality since
d(p, γ(r)) = r. 

We will prove that d(p, q) = s + d(γ(s), q) for each s ∈ [0, d(p, q)]. This im-
plies in particular (choosing s = d(p, q)) that d(p, q) = d(p, q)+d(γ(d(p, q)), q),
which implies that d(γ(d(p, q)), q) = 0 and so γ(d(p, q)) = q, proving (∗)p .
To prove this, let S = {s : d(p, q) = s + d(γ(s), q)} ⊂ [0, d(p, q)]. The
Lemma above shows that S is non-empty.
Lemma 12.4.3. S is an interval containing 0.

Proof. If t < s and s ∈ S then


d(p, q) = s + d(γ(s), q) = t + ((s − t) + d(γ(s), q)) ≥ t + d(γ(t), q) ≥ d(p, q),
so equality holds throughout, and in particular t ∈ S from the last equality.

Lemma 12.4.4. S is closed.

Proof. This is immediate since d and γ are continuous. 


Lemma 12.4.5. S is open in [0, d(p, q)].

Proof. Suppose s ∈ S with s < d(p, q), and let p̃ = γ(s) Choose ε > 0
such that expp̃ is a diffeomorphism on B2ε (0) ⊂ Tp̃ M , and choose z of unit
104 12. RIEMANNIAN GEODESICS AND COMPLETENESS

length in Tp̃ (M ) such that d(expp̃ (εz), q) = d(expp̃ (Sε (0)), q). Then since
every curve from p̃ to q passes through expp̃ (Sε (0)), we have
d(p̃, q) = d(p̃, expp̃ (Sε (0))) + d(expp̃ (Sε (0)), q)
= ε + d(expp̃ (εz), q).
Let σ(t) = γ(t) for t ≤ s, and σ(t) = expp̃ ((t − s)z) for s ≤ t ≤ s + ε.Then
we have
d(p, q) = s + d(p̃, q)
= s + ε + d(σ(s + ε), q)
= L[σ] + d(σ(s + ε), q)
≥ d(p, σ(s + ε)) + d(σ(s + ε), q)
≥ d(p, q),
and we conclude that equality holds throughout. In particular we have
L[σ] = d(p, σ(s + ε), so σ is a length-minimizing curve, hence a geodesic. By
uniqueness of geodesics, σ = γ and s + ε ∈ S. It follows that [0, s + ε] ⊂ S,
and hence S is open. 

It follows that S = [0, d(p, q)], as claimed. This completes the proof of (∗)p ,
and the proof of the Theorem. 
CHAPTER 13

Tensors and tensor fields

Contents
13.1. Equivalence of Riemannian manifolds 105
13.2. Multilinear algebra 106
13.2.1. dual spaces and dual bases 106
13.2.2. Tensor products of vector spaces 107
13.2.3. Multilinear maps 108
13.2.4. Contractions 109
13.2.5. Inner products 110
13.3. Tensor bundles 111
13.4. Connections on tensor bundles 112
13.5. A test for tensors 114
13.6. The curvature of a connection 116
13.7. Symmetries of the curvature 118
13.8. Sectional curvatures 120
13.9. The flag curvature 121
13.10. The Ricci curvature 121
13.11. The scalar curvature 121
13.12. The curvature operator 121
13.13. Three-manifolds 122

13.1. Equivalence of Riemannian manifolds

When studying smooth manifolds, the natural notion of equivalence is equiv-


alence up to diffeomorphism — that is, two smooth manifolds are considered
equivalent if there is a diffeomorphism between them. It is interesting to
note that there are no local invariants of smooth manifolds up to diffeomor-
phism, except for their dimension: Given any two manifolds of the same
dimension, and any points x and y in these manifolds, there are neighbour-
hoods of x and y which are diffeomorphic to each other (for example, take
the inverse images of a small ball in charts about each point). The study of
smooth manifolds is usually called differential topology.
However, in the study of Riemannian manifolds, we have additional struc-
ture, and the natural notion of equivalence is more restrictive: We should
105
106 13. TENSORS AND TENSOR FIELDS

only consider those diffeomorphisms which also preserve the RIemannian


structure:

Definition 13.1.1. Let M be a smooth manifold with Riemannian metric


g, and let N be a smooth manifold with Riemannian metric h. Then a
diffeomorphism F : M → N is called an isometry if

hF (x) (DF |x (u), DF |x (v)) = gx (u, v)

for all x ∈ M and u, v ∈ Tx M .

The natural notion of equivalence of Riemannian manifolds is equivalence up


to isometry. This extra structure changes the situation a great deal, and we
will soon see that Riemannian manifolds do have local invariants — by this
we mean quantities that can be computed at each point, which distinguish
manifolds which are not isometric (i.e. which must agree at corresponding
points if there exists an isometry between the two manifolds).
In order to derive and work with these invariants, we first need a framework
to describe them, and this is provided by the machinery of tensors. We will
start by constructing tensor spaces associated to any vector space, and then
install this structure onto manifolds by constructing tensor bundles. These
are vector bundles over the manifold, and their sections are called tensor
fields.

13.2. Multilinear algebra

13.2.1. dual spaces and dual bases. Let V be a (finite dimensional)


vector space, and V ∗ the dual vector space (that is, the space of real linear
functions on V ). Let {e1 , · · · , en } be a basis for V . Then we can define a
‘dual basis’ for V ∗ as follows: We set
(
1, i = j;
ei∗ (ej ) = δji =
6 j.
0, i =

That is, for v = i v i ei ∈ V we have ei∗ (v) = v i . Each ei∗ is a linear map on
P
V ∗ , hence an element of V ∗ . We can check that these form a basis for V ∗ :
First {e1∗ , · · · , en∗ } are linearly independent, since if j ωj ej∗ = 0, then for
P

each i we have 0 = ( j ωj ej∗ )(ei ) = j ωj δij = ωi . Also, this set spans V ∗ ,


P P

since if ω is any linear function on V then we can define ωi := ω(ei ), and


13.2. MULTILINEAR ALGEBRA 107

ωj ej∗ . Then for any vector v = i v i ei ∈ V we have


P P
then let w̃ = j
X X
w̃(v) = ωj ej∗ ( v i ei )
j i

ωj v i δij
X
=
i,j
X
= ωi v i
i
X
= v i ω(ei )
i
X
= ω( v i ei )
i
= ω(v).
Since v is arbitrary, this shows that ω = ω̃ and hence ω is in the linear span
of {e1∗ , · · · , en∗ }.

13.2.2. Tensor products of vector spaces. Suppose V and W are


two finite-dimensional vector spaces. Then the tensor product space V ⊗ W
is the vector space generated by formal products v ⊗ w where v ∈ V and
w ∈ W , and we impose the conditions
(a1 v1 + a2 v2 ) ⊗ w = a1 (v1 ⊗ w) + a2 (v2 ⊗ w)
and
v ⊗ (b1 w1 + b2 w2 ) = b1 (v ⊗ w1 ) + b2 (v ⊗ w2 ).
That is, we impose that the operation ⊗ : V × W → V ⊗ W is bilinear.
The general object in V ⊗ W is then a linear combination of the form
N
X
ai (vi ⊗ wi ),
i=1

where ai ∈ R, vi ∈ V for each i ∈ {1, · · · , N } and wi ∈ W for each i ∈


{1, · · · , N }. If we choose a basis {e1 , · · · en } for V and a basis ẽ1 , · · · , ẽm for
W , then we have vi = j vij ej for each i, and wi = j wij ẽj for each i, and
P P
so we have
N n X m N
!
ai vij wik ej ⊗ ẽk ,
X X X
ai (vi ⊗ wi ) =
i=1 j=1 k=1 i=1

from which we see that {ej ⊗ek : 1 ≤ j ≤ n, 1 ≤ k ≤ m} is a basis for V ⊗W ,


and in particular V ⊗ W has dimension equal to mn = dim(V ) · dim(W ).
n m
Pn Pmif V = R and W = R , then an element of V ⊗ W has the
In particular,
form i=1 j=1 Bij ei ⊗ ej , which we can identify with the n × m matrix
(Bij ).
108 13. TENSORS AND TENSOR FIELDS

The tensor product operation ⊗ is a bilinear map from V × W to V ⊗ W .


It is associative, in the sense that there is a well-defined isomorphism from
(U ⊗V )⊗W to U ⊗(V ⊗W ), and with this identification we have (u⊗v)⊗w =
u ⊗ (v ⊗ w). This allows us to unambiguously define the space V1 ⊗ · · · ⊗ Vk
and the element v1 ⊗ · · · ⊗ vk within it.

13.2.3. Multilinear maps. The definition above is somewhat abstract,


but a more concrete realisation of the tensor product is the following:
If V1 , · · · , Vk are vector spaces, then a map T : V1 × · · · × Vk → R is called
multilinear if it is linear in each argument: That is,
X X
T (v1 , · · · , vi−1 , aj wj , vi+1 , · · · , vk ) = aj T (v1 , · · · , vi−1 , wj , vi+1 , · · · , vk )
j j

for any i ∈ {1, · · · , k}, v` ∈ V` for ` 6= i, and aj ∈ R, wj ∈ Vi .


The space of such maps is a vector space, which we denote by V1∗ ⊗ · · · ⊗ Vk∗
(we will show below that this is isomorphic to the abstract tensor space
defined previously). There is a natural operation (the tensor product) on
such multilinear functions, defined as follows: If S ∈ V1∗ ⊗ · · · ⊗ Vk∗ and
∗ ⊗ · · · ⊗ V ∗ , then S ⊗ T is the element of V ∗ ⊗ cdots ⊗ V ∗ defined
T ∈ Vk+1 k+` 1 k+`
by
(S ⊗ T )(v1 , · · · , vk+` ) = S(v1 , · · · , vk )T (vk+1 , · · · , vk+` ),
for any vi ∈ Vi , i = 1, · · · , k + `.
The correspondence with the abstract tensor product space
V1∗ ⊗ · · · ⊗ Vk∗
(i) (i)
is given as follows: If e1 , · · · , eni is a basis for Vi , and e1(i) , · · · , en(i)i if the
dual basis, then
ei(1)
1
⊗ · · · ⊗ ei(k)
k

for 1 ≤ ij ≤ nj for j = 1, · · · , k, forms a basis for the space of multilinear


maps from V1 × · · · × Vk to R, where this element acts on basis vectors as
follows:  
(1) (k)
ei(1)
1
⊗ · · · ⊗ ei(k)
k
ej1 , · · · , ejk = δji11 · · · δjikk .

Since any finite dimensional vector space is a dual space, this gives us a
convenient interpretation of the tensor space V × W for any vector spaces
V and W , as the space of bilinear functions on V ∗ × W ∗ . In what follows
we will mostly use this manifestation of tensors as multilinear maps.
A special case with special notation is the following: If V1 = · · · = Vk = V
and Vk+1 = · · · = Vk+` = V ∗ , then we denote by T k,` (V ) the space of
multilinear maps from V k × (V ∗ )` to R. In the case where V = Tx M we will
also write this as Txk,` M .
13.2. MULTILINEAR ALGEBRA 109

It is important to note that the tensor spaces we describe can have several
interpretations: For example, the space V ∗ ⊗ W can be interpreted as the
space of bilinear maps acting on a vector in V and a covector of W , but
(perhaps more naturally) also as the space of linear maps from V to W : If
T ∈ L(V, W ) then we can define T̃ ∈ V ∗ ⊗ W by
T̃ (v, ω) := ω(T (v)).
Exercise 13.2.1. Show that the map from T to T̃ is an isomorphism.

The same space can also be interpreted as the space of linear maps from
W ∗ to V ∗ since there is a natural isomorphism from V ∗ ⊗ W to W ⊗ V ∗
(the isomorphism from L(V, W ) to L(W ∗ , V ∗ ) is the map which associates
to each linear map it’s adjoint map).

13.2.4. Contractions. Contractions are natural operations which gen-


eralise the trace of a linear map: Suppose T ∈ V ⊗V ∗ ⊗W (which we identify
with the space of linear maps from V ⊗ V ∗ to W ) for some vector spaces V
and W . Then the contraction of T , often denoted by Tr(T ), is the element
of W defined by
n
X
Tr(T ) = T (ei , ei∗ ),
i=1
where {e1 , · · · , en } is any basis for V , and {e1∗ , · · · , en∗ } is the dual basis.
This is defined independent of the choice of basis, since any other basis is of
the form ẽi = Λji ej for some Λ ∈ GL(n), and then the dual basis is defined
by
δji = ẽi∗ (ẽj ) = ẽi∗ (Λkj ek ) = Λkj ẽi∗ (ek ),
i i
implying that ẽi∗ (ek ) = Λ−1 k and hence ẽi∗ = Λ−1 k ei∗ . This gives

T (Λpi ep , (Λ−1 )iq eq∗ ) =


X X X p X
T (ẽi , ẽi∗ ) = Λi (Λ−1 )iq T (ep , eq∗ ) = T (ep , ep∗ ).
i i i p

In particular, we have many contractions from Tk,` (V ) to Tk−1,`−1 (V ) for


each choice of vector and co-vector arguments to contract: A tensor T ∈
T k,` (V ) can be written as
T = Ti1 ···ik j1 ···j` dxi1 ⊗ · · · ⊗ dxik ⊗ ∂j1 ⊗ · · · ⊗ ∂j` ,
and a contraction (say on the last vector argument and the last co-vector
argument) would be
Tr(T ) = Ti1 ···ik−1 p j1 ···j`−1 p dxi1 ⊗ · · · ⊗ dxik−1 ⊗ ∂j1 ⊗ · · · ⊗ ∂j`−1 .
It is sometimes convenient to denote a tensor simply by its components
Ti1 ···ik j1 ···j` . A contraction then amounts to performing sum over a pair of
indices (one raised, corresponding to a vector argument of the tensor, and
the other lowered, corresponding to a co-vector argument). This justified our
110 13. TENSORS AND TENSOR FIELDS

‘summation convention’ which says that we should sum over repeated indices
in this situation — this summation is a contraction, and the argument above
shows that this is a well-defined and geometrically invariant operation.

13.2.5. Inner products. Suppose that we have an inner product g on


the vector space V . We will show that there are natural inner products
induced on each of the tensor spaces T k,` (V ) constructed from V .
The first step is to construct an inner product on the dual space V ∗ , which
we do using the following observation:
Lemma 13.2.2. An inner product g on V induces an isomorphism ] from V
to V ∗ , defined by the relation
(](u))(v) = g(u, v).

Proof. Choose a basis {e1 , · · · en } for V . Then the relation implies that
(](ei ))(ej ) = gij = gik ek∗ (ej ),
for each j. It follows that ](ei ) = gik ek∗ for each i. To show this is a
isomorphism it suffices to check that ] has no kernel: If ](u) = 0, then
g(u, v) = (](u))v = 0 for all v, and in particular choosing v = u gives
g(u, u) = 0 and hence u = 0 since g is an inner product. 

The inverse map is denoted by [ : V ∗ → V . We have [(ek∗ ) = (g −1 )k` e` ,


where (g −1 )kl is the inverse matrix of (gij ), defined by (g −1 )ik gkj = δji .
We can now define an inner product on V ∗ by setting g(ξ, η) = g([(ξ), [(η))
for all ξ, η ∈ V ∗ . This is the unique inner product on V ∗ which make the
isomorphism ] an isometry. In a basis {ei } for V with associated dual basis
{ei∗ } for V ∗ , we then have
g ij := g(ei∗ , ej∗ ) = g((g −1 )ik ek , (g −1 )j` e` ) = (g −1 )ik (g −1 )j` gkl = (g −1 )ij .
That is, the inner product on V ∗ , written with respect to the dual basis, is
simply the inverse matrix of the inner product on V . We will continue to
use the convention that g ij (with raised indices) is the inner product on V ∗ ,
so corresponds to the (i, j) component of the inverse matrix of g.
Now we can define inner products on more general tensor spaces:
Proposition 13.2.3. Given an inner product g on V , and the induced inner
product on V ∗ , there is a unique inner product on each of the tensor spaces
T k,` (V ) such that
hS ⊗ T, S 0 ⊗ T 0 i = hS, S 0 ihT, T 0 i
for any tensors S, S 0 ∈ T k,` (V ) and T, T 0 ∈ T p,q (V ).
13.3. TENSOR BUNDLES 111

Proof. We first derive an expression for this inner product: Any ele-
ment of T k,` (V ) can be written in the form
Ti1 ···ik j1 ···j` ei∗1 ⊗ · · · ⊗ ei∗k ⊗ ej1 ⊗ · · · ⊗ ej` ,
so we must have for two such tensors T and S
hT, Si = Ti1 ···ik j1 ···j` Sp1 ···pk q1 ···q` hei∗1 ⊗ · · · ⊗ ej` , ep∗1 ⊗ · · · ⊗ eq` i
= Ti1 ···ik j1 ···j` Sp1 ···pk q1 ···q` hei∗1 , ep∗1 i · · · hej` , eq` i
= Ti1 ···ik j1 ···j` Sp1 ···pk q1 ···q` g i1 p1 · · · g ik pk gj1 q1 · · · gj` q` .
One can now check directly that this formula defines an inner product with
the required properties, independent of the choice of basis for V . 

We note that if {e1 , · · · , en } is an orthonormal basis for V , then the dual


basis {e1∗ , · · · , en∗ } is an orthonormal basis for V ∗ , and {ei∗1 ⊗ · · · ⊗ ei∗k ⊗ ej1 ⊗
· · · ⊗ ej` : 1 ≤ i1 , · · · , j` ≤ n} is an orthonormal basis for T k,` (V ).
Exercise 13.2.4. The inner product g on V is itself an element of T 2,0 (V ),
so we can measure it’s length 2,0
√ in the induced inner product on T (V ). Show
that hg, gi = n, so |g| = n.
Exercise 13.2.5. Show that tensor products commute with taking duals,
in the sense that there is a natural isomorphism from (E ⊗ Ẽ)∗ to E ∗ ⊗ Ẽ ∗ .
Exercise 13.2.6. The induced inner product on T 2,0 (V ) = V ∗ ⊗ V ∗ is an
element of T 2,0 (T 2,0 (V )) = (T 2,0 (V ))∗ ⊗ (T 2,0 (V ))∗ = (V ∗ ⊗ V ∗ )∗ ⊗ (V ∗ ⊗
V ∗ )∗ = V ⊗ V ⊗ V ⊗ V = T 0,4 (V ). Show that it has length n in the induced
inner product on T 0,4 (V ).

13.3. Tensor bundles

Our next task is to install this machinery on a manifold:


Suppose we have two vector bundles E and Ẽ of rank k and ` respectively
over a manifold M . Then the vector bundle E ⊗ Ẽ is the vector bundle with
fibre at x ∈ M given by Ex ⊗ Ẽx . Local bases of smooth sections (which
define corresponding local trivialisations which determine the vector bundle
structure) are given by {ψα ⊗ ηβ : 1 ≤ α ≤ k, 1 ≤ β ≤ `}, where {ψα : 1 ≤
α ≤ k} is a basis of smooth local sections of E, and {ηβ : 1 ≤ β ≤ `} is a
basis of smooth local sections for Ẽ. Thus E ⊗ Ẽ is a vector bundle of rank
k` over M .
Iterating this construction gives more general vector bundles E1 ⊗ · · · ⊗ Ek
constructed from any vector bundles E1 , · · · , Ek over M . In particular, the
bundle T ∗ M ⊗ · · · ⊗ T ∗ M ⊗ T M ⊗ · · · ⊗ T M is denoted by T k,` (M ).
112 13. TENSORS AND TENSOR FIELDS

Sections of E1 ⊗ · · · ⊗ Ek have the form


T = T α1 ···αk ψα(1)
1
⊗ · · · ⊗ ψα(k)
k

(i) (i)
where {ψ1 , · · · , ψni } is a basis of smooth local sections for Ei for each
i ∈ {1, · · · , k}, the coefficient functions T α1 ···αk are each smooth functions
on the common domain of the local trivialisations of E1 , · · · , Ek , and we
understand that the index αi should be summed from 1 to ni (which is the
rank of the vector bundle Ei ).
In particular, sections of T k,` (M ) have the following form in any chart for
M:
T = Ti1 ···ik j1 ···j` dxi1 ⊗ · · · ⊗ dxik ⊗ ∂j1 ⊗ · · · ⊗ ∂j` ,
where {dxi } is the basis of coordinate differentials for T ∗ M in the chart.
∂xi
Note that dxi (∂j ) = ∂x i
j = δj , so this is the dual basis to the coordinate
tangent basis {∂i } of T M .

13.4. Connections on tensor bundles

Our next task is to establish the machinery for differentiating tensor fields
— that is, we define connections on each of the tensor spaces. The basic
result is the following:
Proposition 13.4.1. Suppose ∇(i) is a connection on Ei for each i ∈
{1, · · · , k}. Then there exists a unique connection ∇ on E1∗ ⊗ · · · ⊗ Ek∗ (the
bundle of multilinear maps on E1 × · · · × Ek ) satisfying
X(T (ξ1 , · · · , ξk )) = (∇X T )(ξ1 , · · · , ξk )
(1) (k)
+ T (∇X ξ1 , · · · , ξk ) + · · · + T (ξ1 , · · · , ∇X ξk )
for each ξi ∈ Γ(Ei ), i = 1, · · · , k, X ∈ X (M ), and T ∈ Γ(E1∗ ⊗ · · · ⊗ Ek∗ ).

Proof. The given equation can be re-arranged to give a formula for the
connection:
(1)
(∇X T )(ξ1 , · · · , ξk ) = X(T (ξ1 , · · · , ξk )) − T (∇X ξ1 , · · · , ξk ) − · · ·
(k)
· · · − T (ξ1 , · · · , ∇X ξk ).
We must check that this does indeed defined a connection. For this, we
need to check that the right-hand side, when evaluated at any point x in M ,
depends only on the values of X, ξ1 , · · · , ξk at x (for this it suffices to check
that the right-hand side is C ∞ (M )-linear in these arguments) and that the
Leibniz rule holds.
The C ∞ -linearity in X is immediate, since each of the connections ∇(i) is
C ∞ -linear in X, and T is multilinear. We check the C ∞ -linearity in ξ1 (the
13.4. CONNECTIONS ON TENSOR BUNDLES 113

other arguments are similar):


(1)
(∇X T )(f ξ1 , · · · , ξk ) = X(T (f ξ1 , · · · , ξk )) − T (∇X (f ξ1 ), · · · , ξk ) − · · ·
(k)
− T (f ξ1 , · · · , ∇X ξk )
(1)
= X(f T (ξ1 , · · · , ξk )) − T (f ∇X ξ1 + Xf ξ1 , · · · , ξk ) − · · ·
(k)
− f T (ξ1 , · · · , ∇X ξk )
(( (1)
=(
Xf 1 , · · · , ξk ) + f X(T (ξ1 , · · · , ξk )) − f T (∇X ξ1 , · · · , ξk )
(T((ξ( ((
(( (k)
−(
Xf 1 , · · · , ξk ) − · · · − f T (ξ1 , · · · , ∇X ξk )
(T((ξ( ((
 
(1) (k)
= f X(T (ξ1 , · · · , ξk )) − T (∇X ξ1 , · · · , ξk ) − · · · − T (ξ1 , · · · , ∇X ξk )
= f (∇X T )(ξ1 , · · · , ξk )
as required.
Finally, we check the Leibniz rule for the connection: Multiplying T by a
function f gives
(1)
(∇X (f T ))(ξ1 , · · · , ξk ) = X(f T (ξ1 , · · · , ξk )) − f T (∇X ξ1 , · · · , ξk ) − · · ·
(k)
− f T (ξ1 , · · · , ∇X ξk )
(1)
= Xf T (ξ1 , · · · , ξk ) + f X(T (ξ1 , · · · , ξk ) − f T (∇X ξ1 , · · · , ξk )
(k)
− · · · − f T (ξ1 , · · · , ∇X ξk )
= Xf T (ξ1 , · · · , ξk ) + f (∇X T )(ξ1 , · · · , ξk )
= (Xf T + f ∇X T )(ξ1 , · · · , ξk )
so the required Leibniz rule holds, and ∇ is a connection. 

In particular, taking k = 1 the proposition defines a connection on E ∗ from


a connection on E according to the formula
(∇X ω)(Y ) = X(ω(Y )) − ω(∇X Y )
for any sections ω of E ∗ , Y of E, and X of T M . Given this, the proposition
also defines a canonical connection on each of the tensor spaces T k,l (E).
If ∇ is a connection on E, then given a basis of smooth local section {psiα }
for E the connection is determined by the connection coefficients Γiα β de-
fined by the formula
∇∂i ψα = Γiα β ψβ .
Using these we can also write the connection on the other tensor bundles:
First, consider the connection on E ∗ : If {ψ∗α } is the dual basis for E ∗ , then
we have
(∇i ψ∗α )(ψβ ) = ∂i (ψ∗α (ψβ )) − ψ∗α (∇i ψβ ) = −ψ∗α (Γiβ γ ψγ ) = −Γiβ α .
It follows that ∇i ψ∗α = −Γiβ α ψ∗α .
114 13. TENSORS AND TENSOR FIELDS

Using this, we can find an expression for the connection applied to a general
section T of T k,ell (E) in terms of a local basis: If T = Ti1 ···ik j1 ···j` ψ∗i1 ⊗ · · · ⊗
ψ∗ik ⊗ ψj1 ⊗ · · · ⊗ ψj` , then we have according to the proposition that the
components of ∇T are given by
(∇i T )i1 ···ik j1 ···j` = ∂i Ti1 ···ik j1 ···j`
− Γii1 p Tpi2 ···ik j1 ···j` − · · · − Γiik p Ti1 ···ik−1 p j1 ···j`
+ Γip j1 Ti1 ···ik pj2 ···j` + · · · + Γip j` Ti1 ···ik j1 ···j`−1 p .
Exercise 13.4.2. Show that if {e1 , · · · , en } are a basis for E which is par-
allel, then the dual basis {e1∗ , · · · , en∗ } for E ∗ is also parallel, and the basis
{ei∗1 ⊗ · · · ei∗k ⊗ ej1 ⊗ · · · ⊗ ej` : 1 ≤ i1 , · · · , j` ≤ n} for T k,l (E) is parallel.
Exercise 13.4.3. Show that if ∇ and g are a compatible connection and
metric on a vector bundle E, then the induced metrics and connections on
E ∗ are also compatible.
Exercise 13.4.4. If ∇(i) and gi are a compatible connection and metric on
Ei for i = 1, 2, show that the induced connection and metric on E1∗ ⊗ E2∗ are
compatible.
Exercise 13.4.5. If ∇ and g are compatible on E, show that the induced
metric and connection on each of the tensor spaces T k,` (E) are compatible.

13.5. A test for tensors

Now we will formalise an observation which we have used several times


previously:
If T is a tensor field (say, a section of a vector bundle E1∗ ⊗ · · · ⊗ Ek∗ ),
then T (x) acts at each point x on elements ξi (x) ∈ (Ei )x of the fibres of
the vector bundles at x, in a way which is linear in each argument at each
point, and which depends smoothly on x. In particular, this implies that
for any smooth sections ξ ∈ Γ(Ei ), the function T (ξ1 , · · · , ξk ) is a smooth
function on M . Thus we can think of the section T as a map which takes
Γ(E1 ) × · · · × Γ(Ek ) to C ∞ (M ). The linearity at each point implies that
this map is multilinear over the ring C ∞ (M ) of smooth functions on M ,
so that multiplying any one of the arguments by a smooth function simply
multiplies the result by the same smooth function.
Very useful for us is the converse statement:
Proposition 13.5.1. Let T̃ : Γ(E1 ) × · · · × Γ(Ek ) → C ∞ (M ). Then there
exists a tensor T ∈ Γ(E1∗ ⊗ · · · ⊗ Ek∗ ) such that
(T̃ (ξ1 , · · · , ξk ))(x) = Tx (ξ1 (x), · · · , ξk (x))
for all ξi ∈ Γ(Ei ), i = 1, · · · , k, and x ∈ M , if and only if T̃ is multilinear
over C ∞ (M ).
13.5. A TEST FOR TENSORS 115

Proof. If there is such a tensor field T , hen the action on sections


is certainly C ∞ (M )-multilinear. We prove the converse: Fix x ∈ M and
(j)
suppose that we have bases of smooth local sections {ψi : 1 ≤ i ≤ ni } of
Ei for each j, defined in a neighbourhood of x. Denote the dual bases for
Ei∗ by ψ(j)
i . If there is such a tensor T , then we can write

i1 ik
T = Ti1 ···ik ψ(1) ⊗ ψ(k) ,
and in particular we have
(1) (k) (1) (k)
Ti1 ···ik (x) = Tx (ψi1 (x), · · · , ψik (x)) = T̃ (ψi1 , · · · , ψik )(x).
It remains to check that the tensor defined in this way gives the correct result
when applied to arbitrary sections and evaluated at x: Arbitrary sections ξi
(i)
of Ei have the form ξi = ξij ψj , where ξij is smooth, and so by C ∞ -linearity
we have
(1) (k)
T̃ (ξ1 , · · · , ξk )(x) = ξ1i1 (x) · · · ξkik (x)T̃ (ψi1 , · · · , ψik )(x)
= ξ1i1 (x) · · · ξkik (x)Ti1 ···ik (x)
= Tx (ξ1 (x), · · · , ξk (x))
as required. 
Exercise 13.5.2. We cheated in the proof by applying T̃ to local smooth
sections, while the proposition refers to global sections of Ei for each i. How
can this be fixed?
Example 13.5.3 (Connections are not tensors). Let E be a vector bundle,
and ∇ a connection on E. Then ∇ takes a vector field and a section of E
and gives a section of E, or equivalently takes sections of T M , E and E ∗
and gives a smooth function:
∇ : X ∈ X (M ), ξ ∈ Γ(E), ω ∈ Γ(E ∗ ) 7→ ω(∇X ξ) ∈ C ∞ (M ).
We will show that ∇ is however not a tensor: It is tensorial in ω and X
(that is, the result evaluated at a point x depends only on the values if these
arguments at x) but not in ξ: We have
∇(X, f ξ, ω) = ω(∇X (f ξ)) = ω(Xf ξ + f ∇X ξ) = f ∇(X, ξ, ω) + Xf ω(ξ).
The last term is certainly not always zero: We can choose f , X, ω and ξ
arbitrarily and so can certainly choose these to make the last term non-zero.
This makes sense conceptually: A tensor is supposed to depend only on
the value of the sections at a point, while the connection is a differential
operator and so should also depend on the derivative at the point.
Example 13.5.4 (Differences of connections are tensors). Now suppose that
∇(2) and ∇(1) are two connections on a vector bundle E. We will show that
the difference ∇(2) − ∇(1) taking X ∈ X (M ), ξ ∈ Γ(E) and ω ∈ Γ(E ∗ ) to
116 13. TENSORS AND TENSOR FIELDS

(2) (1)
ω(∇X ξ − ∇X ξ) is a tensor (i.e. a section of the bundle T ∗ M ⊗ E ∗ ⊗ E):
As before the tensoriality in X and ω is immediate. In ξ we have
(2) (1) (2) (1)
∇X (f ξ) − ∇X (f ξ) = (Xf ξ + f ∇X ξ) − (Xf ξ + f ∇X ξ)
(2) (1)
= f (∇X ξ − ∇X ξ),
so the terms involving the derivative Xf cancel, and the result is a tensor.
Example 13.5.5 (The Lie bracket is not a tensor). The Lie bracket operation
takes two vector fields X and Y and gives a new vector field [X, Y ]. However
this does not define a (2, 1)-tensor: We have
[f X, Y ]g = f X(Y g) − Y (f Xg)
= f X(Y g) − f Y (Xg) − Y f Xg
= (f [X, Y ] − Y f X)g
for any smooth function g, and so [f X, Y ] = f [X, Y ] − Y f X. The last term
is not in general zero, so the Lie bracket is not a tensor.

However, we can put this together with the previous examples to give the
following:
Example 13.5.6 (The torsion tensor). Let ∇ be a connection on T M . Then
the operator T : X (M ) × X (M ) → X (M ) defined by
T (X, Y ) = ∇X Y − ∇Y X − [X, Y ]
defines a (2, 1)-tensor on M , called the torsion of ∇: This is antisymmetric
in X and Y , so we only need to check that it is tensorial in X. We have
T (f X, Y ) = ∇f X Y − ∇Y (f X) − [f X, Y ]
= f ∇X Y − f ∇Y X − Y f X − f [X, Y ] + Y f X
= f (∇X Y − ∇Y X − [X, Y ])
= f T (X, Y ).
The torsion vanishes if and only if ∇ is a symmetric connection.

13.6. The curvature of a connection

A very important example of a tensor constructed in the way described in


the previous section is the following:
Definition 13.6.1. If ∇ is a connection on a vector bundle E over M , then
the curvature tensor R of ∇ is the section of T ∗ M ⊗ T ∗ M ⊗ E ∗ ⊗ E defined
by
R(X, Y, ξ) = ∇Y (∇X ξ) − ∇X (∇Y ξ) − ∇[Y,X] ξ ∈ Γ(E).
13.6. THE CURVATURE OF A CONNECTION 117

To make sense of this definition we should first verify that R is tensorial


in all arguments: In X (and similarly Y since R is antisymmetric in these
arguments) we have
R(f X, Y, ξ) = ∇Y (f ∇X ξ) − f ∇X (∇Y ξ) − ∇f [Y,X]+Y f X ξ
= Y f ∇X ξ + f (∇Y (∇X ξ) − ∇X (∇Y ξ) − ∇[Y,X] ξ) − Y f ∇X ξ
= f R(X, Y, ξ).
In ξ we have:
R(X, Y, f ξ) = ∇Y (∇X (f ξ)) − ∇X (∇Y (f ξ)) − ∇[Y,X] (f ξ)
= ∇Y (Xf ξ + f ∇X ξ) − ∇X (Y f ξ + f ∇Y ξ) − ([Y, X]f )ξ − f ∇[Y,X] ξ
= Y (Xf )ξ + Xf ∇Y ξ + Y f ∇X ξ + f ∇Y (∇X ξ)
− X(Y f )ξ − Y f ∇X ξ − Xf ∇Y ξ − f ∇X (∇Y ξ)
− ([Y, X]f )ξ − f ∇[Y,X] ξ

= f ∇Y (∇X ξ) − ∇X (∇Y ξ) − ∇[Y,X] ξ
= f R(X, Y, ξ).

Some idea of the meaning of the curvature tensor can be obtained from the
following:
Proposition 13.6.2. Let ∇ be a connection on a vector bundle E over M .
Then the curvature tensor of ∇ vanishes if and only if there exists a basis
of parallel sections of E in a neighbourhood of any point of M .

Proof. If there exists a parallel basis {ξ1 , · · · , ξk } of local sections of


E on some domain, then we have
R(X, Y, ξi ) = ∇Y (∇X ξi ) − ∇X (∇Y ξi ) − ∇[Y,X] ξi = 0
since ∇ξi = 0 by assumption. It follows that the tensor R vanishes since the
ξi form a basis.
Conversely, suppose that R = 0. Choose a chart ϕ for M , such that Q =
[0, 1]n is in the range of the chart. Choose a basis ξi for the fibre of E
at ϕ−1 (0), and parallel transport along the ∂1 direction to define ξi at all
points of the form ϕ−1 (x1 , 0, · · · , 0) for x1 ∈ [0, 1]. Note that by construction
∇∂1 ξi = 0 along this line. Then parallel-transport along the ∂2 direction to
define ξi at the points ϕ−1 (x1 , x2 , 0, · · · , 0), and observe that ∇2 ξi = 0 at
all of these points. We claim that ∇1 ξi = 0 also at all points of the form
ϕ−1 ((x1 , x2 , 0, · · · , 0)): To see this we note that ∇1 ξi |(x1 ,0,··· ,0) = 0, and that
0 = R(∂1 , ∂2 , ξi ) = ∇2 (∇1 ξi ) − ∇1 (∇2 ξi ) = ∇2 (∇1 ξi )
since [∂1 , ∂2 ] = 0, and ∇2 ξi = 0 everywhere. That is, ∇1 ξi is given by
parallel transport in the ∂2 direction of the zero vector, and is therefore zero
everywhere.
118 13. TENSORS AND TENSOR FIELDS

Continuing in this way, after parallel transporting in the ∂j direction we


have ξi defined for points of the form ϕ−1 (x1 , · · · , xj , 0, · · · , 0) and satisfying
∇p ξi = 0 on this set for 1 ≤ p ≤ j. In particular when j = n we have ξi
defined on all of ϕ−1 (Q) and satisfying ∇∂p ξi = 0 for all p and I, as required.
Since parallel transport along any curve is a linear isomorphism, the sections
ξi define a basis for the fibre of E at each point. 

In the case of the Riemannian connection, we have a stronger conclusion:


Proposition 13.6.3. Suppose that g is a Riemannian metric on a manifold
M , and the curvature tensor of the Riemannian connection ∇ vanishes.
Then near any point there exists a chart for M for which gij = δij , so that
M is locally isometric to Euclidean space.

Proof. By the previous proposition, if the curvature vanishes then


there is a neighbourhood of any point on which there is a basis of parallel
local sections E1 , · · · , En of T M . If we choose the basis to be orthonor-
mal at one point, then by compatability of the connection with the metric
the inner products are constant on this neighbourhood, so the vectors are
orthonormal everywhere. Also, since the connection is symmetric we have
[Ei , Ej ] = ∇Ei Ej − ∇Ej Ei = 0
since all of the vectors Ei are parallel. By the Frobenius theorem, there
exists a map F from Rn to M with DF (ei ) = Ei for each i everywhere.
It follows that F is a local diffeomorphism, hence defined a chart for M in
which ∂i = Ei . Then we have gij = g(Ei , Ej ) = δij as claimed. 

Thus the vanishing of the Riemann curvature tensor is equivalent to the


manifold being locally flat — that is, locally isometric to Euclidean space
with the Euclidean inner product.

13.7. Symmetries of the curvature

The curvature of a connection on a vector bundle E is a section of T ∗ M ⊗


T ∗ M ⊗ E ∗ ⊗ E — in particular, if E = T M then this defines a (3, 1)-tensor
field on M . If we have a metric on E then it is sometimes convenient to
convert this to a section of T ∗ M ⊗ T ∗ M ⊗ E ∗ ⊗ E ∗ (in the case E = T M ,
a (4, 0)-tensor) as follows:
R(X, Y, ξ, η) := g(R(X, Y, ξ), η).
Proposition 13.7.1. (i). The curvature tensor is antisymmetric in the
first two arguments: R(X, Y, ξ) = −R(Y, X, ξ);
(ii). If ∇ is compatible with the metric g on E, then R is antisymmetric
in the last two arguments: R(X, Y, ξ, η) = −R(X, Y, η, ξ);
13.7. SYMMETRIES OF THE CURVATURE 119

(iii). If E = T M and ∇ is symmetric, then


R(X, Y, Z) + R(Y, Z, X) + R(Z, X, Y ) = 0.
(iv). If ∇ is the Riemannian connection of the metric g on T M , then
R(X, Y, W, Z) = R(W, Z, X, Y ).

Proof. The first symmetry is immediate from the definition. For the
second we use the definition of the Lie bracket to write
0 = Y (X(g(ξ, η)) − X(Y (g(ξ, η)) − [Y, X]g(ξ, η)
= Y (g(∇X ξ, η) + g(ξ, ∇X η)) − X(g(∇Y ξ, η) + g(ξ, ∇Y η))
− g(∇[Y,X] ξ, η) − g(ξ, ∇[Y,X] η)
= g(∇Y (∇X ξ, η) + g(∇X ξ, ∇Y η) + g(∇Y ξ, ∇X η) + g(ξ, ∇Y (∇X η))
− g(∇X (∇Y ξ), η) − g(∇Y ξ, ∇X η) − g(∇X ξ, ∇Y η) − g(ξ, ∇X (∇Y η))
− g(∇[Y,X] ξ, η) − g(ξ, ∇[Y,X] η)
= g(R(X, Y, ξ), η) + g(ξ, R(X, Y, η))
= R(X, Y, ξ, η) + R(X, Y, η, ξ).
For the third symmetry (called the first Bianchi identity) we argue as follows:
R(X, Y, Z) + R(Y, Z, X) + R(Z, X, Y ) = ∇Y (∇X Z) − ∇X (∇Y Z) − ∇[Y,X] Z
+ ∇Z (∇Y X) − ∇Y (∇Z X) − ∇[Z,Y ] X
+ ∇X (∇Z Y ) − ∇Z (∇X Y ) − ∇[X,Z] Y
= ∇Y (∇X Z − ∇Z X) − ∇[X,Z] Y
+ ∇Z (∇Y X − ∇X Y ) − ∇[Y,X] Z
+ ∇X (∇Z Y − ∇Y Z) − ∇[Z,Y ] X
= ∇Y [X, Z] − ∇[X,Z] Y
+ ∇Z [Y, X] − ∇[Y,X] Z
+ ∇X [Y, Z] − ∇[Y,Z] X
= [Y, [X, Z]] + [Z, [Y, X]] + [X, [Y, Z]]
=0
by the Jacobi identity, see Proposition 8.2.3.
The fourth identity follows from the first three: We have
2R(X, Y, W, Z) = R(X, Y, W, Z) + R(Y, X, Z, W ) by (i),(ii);
= −R(Y, W, X, Z) − R(W, X, Y, Z) − R(X, Z, Y, W ) − R(Z, Y, X, W ) by (iii);
= −R(W, Y, Z, X) − R(X, W, Z, Y ) − R(Z, X, W, Y ) − R(Y, Z, W, X) by (i),(ii);
= R(Z, W, Y, X) + R(W, Z, X, Y ) by (iii);
= 2R(W, Z, X, Y ) by (i),(ii).

120 13. TENSORS AND TENSOR FIELDS

The curvature tensor of the Riemannian connection defines the fundamental


local invariant of Riemannian geometry. Unfortunately, it is algebraically
a very complicated object: In an n-dimensional manifold, the number of
independent components of a (4, 0)-tensor grows like n4 , while the degrees
of freedom available to simplify this by choosing a good basis grows like n2 ,
so we are left with O(n4 ) components (in contrast, for a symmetric (2, 0)-
tensor we can rotate to reduce to diagonal form, so most of the information
is contained in a relatively manageable n eigenvalues).
If n = 2, however, the symmetries of the curvature tensor give us the follow-
ing: There is only one independent component of the Riemann curvature
tensor. To see this, we note that the curvature is determined by it’s action
on basis elements ∂i , so we need only find Rijk` = R(∂i , ∂j , ∂k , ∂` ), where
each of the indices must be either 1 or 2. The antisymmetry (i) implies that
the first two indices must be different, and so {i, j} = {1, 2}. Similarly, the
antisymmetry (ii) implies that {k, `} = {1, 2}. Thus the symmetries implies
that the only non-zero components are R1212 = −R2112 = R2121 = −R1221 .

Proposition 13.7.2. If g is a Riemannian metric on a 2-dimensional man-


ifold, then the curvature tensor R satisfies

Rijkl = K(gik gjl − gjk gil ),

where K is the Gauss curvature, defined by K = 21 g ik g jl Rijkl .

Proof. Since both sides are tensors, it suffices to check the equation
holds in an orthonormal basis. The right-hand side is antisymmetric in i and
j and also in k and l, so as above the only non-zero components are the ones
listed above. Set K = R1212 . Then we have R1212 = K(g11 g22 − g12 g21 ) = K
as required, and by symmetry equality also holds for all of the other non-zero
components. Finally we check that g ik g jl Rijkl = Kg ik g jl (gik gjl − gjk gil ) =
K(n2 − n) = 2K, as claimed (here we used g ij gij = n). 

13.8. Sectional curvatures

In higher dimensions, the sectional curvatures are quantities computed from


the curvature tensor which often arise naturally. The sectional curvature
σ(P ) is defined for each two-dimensional subspace P of Tx M , and is com-
puted as follows: For any orthonormal basis {e1 , e2 } of P , we set

σ(P ) := R(e1 , e2 , e1 , e2 ).

For the same reasons as the n = 2 case, this number is independent of


the choice of basis, and so is a well-defined function of P . This will arise
naturally when we discuss variations of geodesics in the next chapter.
13.12. THE CURVATURE OPERATOR 121

13.9. The flag curvature

Another quite natural construction from the curvature tensor is the flag
curvature, which is defined for any (unit length) vector v as follows: The
flag curvature Rv is a bilinear form on Tx M defined by
Rv (u, w) := R(u, v, w, v).
We note: Rv is a symmetric bilinear form, by the symmetry (iv) of Propo-
sition 13.7.1. The vector v is a null eigenvector, by the antisymmetry (i), so
we can consider Rv as a symmetric bilinear form acting on the orthogonal
subspace v ⊥ in Tx M . The value of Rv acting on a unit vector w orthogonal
to v is R(w, v, w, v), which is the sectional curvature of the plane spanned
by v and w.

13.10. The Ricci curvature

The Ricci curvature Ric is a trace of the curvature tensor, defined by Rij =
g kl Rikjl . The symmetry (iv) of the curvature tensor implies that the Ricci
tensor is a symmetric bilinear form acting on Tx M . If we fix a unit vector v,
then Ric(v, v) is the trace of the flag curvature Rv : Choose an orthonormal
basis {e1 , · · · , en } for TM with v = en . Then we have
X X
Rc(v, v) = g kl Rknln = Rknkn = Rv (ek , ek ).
k<n k<n
Thus Rc(v, v) can be interpreted (up to a factor depending on n) as an
average of sectional curvatures over all two-dimensional subspaces containing
v (which also equals the sum of the n − 1 sectional curvatures of the planes
generated by en and ek for k = 1, · · · , n − 1).

13.11. The scalar curvature

The scalar curvature R is the trace of the Ricci curvature: R = g ij Rij =


g ij g kl Rikjl . Up to a factor depending only on n, this is just the average of
the sectional curvatures over all two-dimensional subspaces of Tx M .

13.12. The curvature operator

Finally, I want to discuss a useful manifestation of the curvature tensor


which makes use of the symmetries we derived above: For an n-dimensional
vector space V , we can construct a new vector space Λ2 (V ) (called the space
of algebraic 2-planes over V ) as follows: Λ2 (V ) is a subspace of the tensor
space V ⊗ V (which we identify with the space of bilinear maps on V ∗ )
consisting of those elements which are antisymmetric: S(ξ, η) = −S(η, ξ)
for all ξ, η ∈ V ∗ . If {ei } is a basis for V , then a basis for Λ2 (V ) is given by
122 13. TENSORS AND TENSOR FIELDS

{ei ∧ ej := ei ⊗ ej − ej ⊗ ei : 1 ≤ i < j ≤ n}, so Λ2 V is a vector space of


dimension n(n−1)
2 .

The curvature operator is the symmetric bilinear form R acting on Λ2 (T M )


defined by
R(S ij (ei ∧ ej , T kl ek ∧ el ) := S ij T kl Rijkl .
The antisymmetries (i) and (ii) of the curvature tensor in the first and second
pairs of indices implies that the entire curvature tensor can be recovered from
R, and the symmetry (iv)between the first pair and the last pair implies that
R is a symmetric bilinear form on Λ2 (Tx M ). The converse is not quite true:
Not every symmetric bilinear form on Λ2 (T M ) could arise as a curvature
operator, since we must also satisfy the first Bianchi identity (iii).

13.13. Three-manifolds

The case n = 3 has some nice special structure: In this case the space
Λ2 (Tx M ) is three-dimensional: If {e1 , e2 , e3 } is an orthonormal basis for
Tx M , then we can take {e2 ∧ e3 , e3 ∧ e1 , e1 ∧ e2 } to be an orthonormal basis
for Λ2 . This gives a well-defined isomorphism L from Tx M to Λ2 (Tx M ):
L(v i ∂i ) = v i µijk g jp g kq ∂p ∧ ∂q ,
where µ is the volume form of g, defined to be the totally antisymmetric
(3, 0)-tensor which takes any three vectors to the (signed) volume of the
parallelopiped they generate (this requires that M be oriented, otherwise
the isomorphism is defined only up to sign).
This isomorphism allows us to define a new symmetric bilinear form on
Tx M , called the Einstein tensor, by
E(u, v) := R(L(u), L(v)).
Since L is an isomorphism, the tensor E determines the curvature tensor
completely, so in three dimensions the curvature tensor reduces to a sym-
metric bilinear form. In particular we can understand this rather completely,
by observing that there exists an orthonormal basis of eigenvectors for E.
The eigenvalues of E are called the principal sectional curvatures of g.
If v is a unit vector, then we can choose an orthonormal basis with e3 = v,
and then compute
Ric(v, v) = R1313 +R2323 ; E(v, v) = R1212 ; R = 2(R1212 +R1313 +R2323 ),
1
so E = 2 Rg − Rc. Since R = Tr(Ric) it follows that the Einstein tensor,
hence the full curvature tensor, is completely determined by the Ricci tensor.
In four and higher dimensions, it is no longer true that the Ricci tensor
determines the full curvature tensor, and the algebra of curvature becomes
much more difficult.
CHAPTER 14

Hypersurfaces and submanifolds

Contents
14.1. Hypersurfaces 123
14.1.1. Induced metric and connection 124
14.1.2. The Gauss map 124
14.1.3. The second fundamental form and the Weingarten map 124
14.1.4. The Weingarten relation 125
14.1.5. The Gauss and Codazzi equations 126
14.1.6. Embedding theorem 127
14.2. Graphical hypersurfaces 127
14.3. Principal curvatures and sectional curvatures 128
14.4. The sphere 129
14.5. Minkowski space and spacelike hypersurfaces 130
14.6. Hyperbolic space 132
14.7. Submanifolds 133
14.7.1. Gauss, Codazzi and Ricci identities 135

14.1. Hypersurfaces

Next we return to look at submanifolds in the context of Riemannian ge-


ometry, and in particular the simplest situation of hypersurfaces. These are
submanifolds M of dimension n in a Euclidean space of dimension n + 1, so
the dimension of the normal space Nx M at each point is just 1. We will make
the simplifying sssumption that the normal bundle is trivial, which means
that there exists a global unit length section ν of N M , i.e. a smoothly de-
fined unit normal vector ν(x) at each point x ∈ M . Note that this amounts
simply to a choice of sign at each point, since there are only two unit length
normal vectors at each point. Locally, such a choice can always be made,
so the description of the local geometry of a hypersurface which follows also
makes sense without this global assumption, but one has to keep in mind
that there can be sign changes if we change the choice of direction of the
normal.
123
124 14. HYPERSURFACES AND SUBMANIFOLDS

14.1.1. Induced metric and connection. We will consider the slightly


more general situation of an immersed submanifold, given by a smooth im-
mersion F : M n → Rn+1 . We use the submanifold metric and connection
defined previously: The Riemannian metric is therefore defined by
gx (u, v) = uF · vF
for all x ∈ M and u, v ∈ Tx M ⊂ Rn+1 . The Riemannian connection then
agrees with the submanifold connection defined by
(∇X Y )F = πx (X(Y F )))
where X ∈ Tx M and πx is the orthogonal projection onto the tangent sub-
space Tx M .

14.1.2. The Gauss map. The map x 7→ ν(x) ∈ S n ⊂ Rn+1 is a


smooth map to the sphere, which is called the Gauss map. We can use this
to make more explicit the projection onto the tangent space:
πx (v) = v − (v · ν(x))ν(x).

14.1.3. The second fundamental form and the Weingarten map.


The connection ∇ is defined by the tangential part of the derivative of a tan-
gent vector field. The second fundamental form is defined by the normal
part: We define it by the following equation:
X(Y F ) = (∇X Y )F − h(X, Y )ν.
That is,
h(X, Y ) = −X(Y F ) · ν.
Proposition 14.1.1. h is a symmetric (2, 0)-tensor field on M .

Proof. We apply the test for tensors: h is certainly tensorial in X; in


Y we have
h(X, f Y ) = −X(f Y F ) · ν
= −(Xf Y F + f X(Y F )) · ν
= −f X(Y F ) · ν
= f h(X, Y ),
since Y F · ν = 0 because Y F is tangential while ν is normal.
To prove that h is symmetric, we argue as follows:
h(X, Y ) − h(Y, X) = −X(Y F ) · ν + Y (XF ) · ν = [Y, X]F · ν = 0,
since [X, Y ] is a tangent vector field, so [X, Y ]F is in the tangent subspace
and hence orthogonal to ν. 
14.1. HYPERSURFACES 125

The second tensor we introduce is called the Weingarten map, which is a


linear map on Tx M (hence a section of T ∗ M × T M ) defined by
(W(v))F = Dv ν.
To justify this we must show that the right-hand side is in the tangential
space, and this is true because |ν|2 = 1, which implies
0 = Dv |ν|2 = 2ν · Dv ν,
so that Dv ν has no normal component and so is in the tangent space, hence
equal to wF for some w ∈ Tx M (which we denote by W(v)).
Thus the second fundamental form tells us how to differentiate tangent vec-
tor fields, while the Weingarten map tells us how to differentiate the normal
vector.

14.1.4. The Weingarten relation. The Weingarten relation is an


equation relating the second fundamental form to the Weingarten map,
showing that they are really different manifestations of the same tensor:
Proposition 14.1.2. For any x ∈ M and any X, Y ∈ Tx M ,
h(X, Y ) = g(W(X), Y ).

Proof. We differentiate the orthogonality relation


YF ·ν =0
to obtain
0 = X(Y F ) · ν + Y F · DX ν
= −h(X, Y ) + Y F · (W(X))F
= −h(X, Y ) + g(Y, W(X)).

2
In local coordinates, the components of h are given by h(∂i , ∂j ) = − ∂x∂i ∂x
F
j ·ν,
∂ν p ∂F
while those of W are given by ∂xi = Wi ∂xp . The Weingarten relation then
says
hij = Wip gpj ,
or equivalently
Wij = hip g pj .
There is a convention commonly adopted here, which is that ‘raising an
index’ means contracting with the inverse of the metric (i.e. converting
one of the arguments of the tensor from a vector argument to a co-vector
argument using the isomorphism between vectors and co-vectors provided
by the metric). Thus we understand
hji = hip g pj = Wij .
126 14. HYPERSURFACES AND SUBMANIFOLDS

Similarly, the curvature (4, 0)-tensor is obtained from the curvature (3, 1)-
tensor by ‘lowering an index’ by contracting with the metric:

Rijkl = Rijk p gpl .

In this sense the second fundamental form and the Weingarten map can be
seen to be essentially the same geometric object.

14.1.5. The Gauss and Codazzi equations. There are two impor-
tant identities satisfied by the second fundamental form, both reflecting the
symmetry of third derivatives of the immersion: Since X(Y f ) − Y (Xf ) −
[X, Y ]f = 0 for any function f (to a vector space) and for any smooth vector
fields X and Y on M , choosing f = ZF for some vector field Z gives

0 = X(Y (ZF ))) − Y (X(ZF )) − [X, Y ](ZF )


= X((∇Y Z)F − h(Y, Z)ν) − Y ((∇X Z)F − h(X, Z)ν)
− (∇[X,Y ] Z)F + h([X, Y ], Z)ν
= (∇X (∇Y Z))F − h(X, ∇Y Z)ν − X(h(Y, Z))ν − h(Y, Z)(W(X))F
− (∇Y (∇X Z))F + h(Y, ∇X Z)ν − Y (h(X, Z))ν + h(X, Z)(W(Y ))F
− (∇[X,Y ] Z)F + h([X, Y ], Z)ν.

Since this vanishes, each of the tangential and normal components must
vanish. The tangential part gives

R(Y, X, Z) = h(Y, Z)W(X) − h(X, Z)W(Y ).

This is called the Gauss identity, and related the curvature tensor of the
metric g to the second fundamental form. Taking an inner product with
another vector field yields the more symmetric form

R(Y, X, Z, W ) = h(Y, Z)h(Z, W ) − h(X, Z)h(Y, W ),

where we used the Weingarten relation on the right hand side.


The normal component gives

0 = Y (h(X, Z)) − X(h(Y, Z)) − h(X, ∇Y Z) + h(Y, ∇X Z) + h([X, Y ], Z)


= (∇Y h)(X, Z) + h(∇Y X, Z) + h(X, ∇Y Z)
− (∇X h)(Y, Z) − h(∇X Y, Z) − h(Y, ∇X Z)
− h(X, ∇Y Z) + h(Y, ∇X Z) + h([X, Y ], Z)
= (∇Y h)(X, Z) − (∇X h)(Y, Z).

Since we already know that h is a symmetric (2, 0)-tensor, this identity says
that the (3, 0)-tensor ∇h is totally symmetric. This is called the Codazzi
identity.
14.2. GRAPHICAL HYPERSURFACES 127

14.1.6. Embedding theorem. The Gauss and Codazzi identities are


the fundamental identities which must be satisfied by the second fundamen-
tal form. The importance of these is reflected in the following result:
Theorem 14.1.3. Let (M, g) be a Riemannian n-manifold. Suppose that
h ∈ Γ(T ∗ M ⊗ T ∗ M ) is a symmetric (2, 0)-tensor field on M satisfying the
equations
h(X, W )h(Y, Z) − h(X, Z)h(Y, W ) = R(X, Y, W, Z)
and
(∇X h)(Y, Z) = (∇Y h)(X, Z)
for all X, Y, W, Z ∈ X (M ). Then for any x ∈ M there exists an immersion
F of a neighbourhood of x in M into Rn+1 with second fundamental form
equal to h. Furthermore, this immersion is unique up to rigid motions of
Rn+1 .

Proof. 

We remark that if M is simply connected then F can be defined on all of M ,


and the immersion F is unique up rigid motions applied to each connected
component of M .

14.2. Graphical hypersurfaces

The second fundamental form describes the ‘bending’ of the hypersurface in


the background space, as measured by the change in direction of the tangent
plane as we move along the hypersurface.
A convenient way to understand the second fundamental form is to consider
writing the hypersurface as a graph, so that the embedding F is given locally
by
X
F (x1 , · · · , xn ) = (x1 , · · · , xn , u(x1 , · · · , xn )) = xi ei + u(x)en+1 .
i
for some smooth function u. In this parametrisation we can compute the sec-
ond fundamental form as follows: We first compute the coordinate tangent
vectors ∂i F :
∂i F = ei + ∂i uen+1 .
From this we can compute the components of the induced Riemannian met-
ric:
gij = ∂i F · ∂j F = δij + ∂i u∂j u.
We can also find the unit normal vector, which we choose to be downward
pointing: This is given by
P
i ∂i uei − en+1
ν= p ,
1 + |Du|2
128 14. HYPERSURFACES AND SUBMANIFOLDS

where |Du|2 = i (∂i u)2 . To check this we need only observe that the given
P
vector is of unit length and is orthogonal to ∂i F for each i. Next we compute
the second fundamental form:
P
i ∂i uei − en+1 ∂i ∂j u
hij = −∂i ∂j F · ν = −∂i ∂j uen+1 · p =p .
1 + |Du| 2 1 + |Du|2

Note that the sign of h(v, v) is simply the sign of D2 u(v, v). In particular,
h(v, v) > 0 means that the surface bends upwards (i.e. away from the unit
normal direction) as we move along the direction v in the hypersurface,
while h(v, v) < 0 means that the surface bends downwards (towards the
unit normal).

Definition 14.2.1. An immersed hypersurface is called locally convex if the


second fundamental form is everywhere positive definite (i.e. the surface
bends away from the unit normal in all directions).

14.3. Principal curvatures and sectional curvatures

Let us now look briefly at the algebra of curvature for a hypersurface: The
Gauss equation gives an expression for the curvature tensor in terms of
the second fundamental form h, which is a symmetric bilinear form. The
second fundamental form can be understood relatively easily: There exists
an orthonormal basis {ei } for Tx M for which h(ei , ei ) = κi , and h(ei , ej ) = 0
for i 6= j. The eigenvalues κ1 , · · · , κn are called the principal curvatures, and
the directions e1 , · · · , en are called the principal directions.
In the frame of principal directions, the Gauss equation gives the following:
If i 6= j then

κi κj
 if i = k and j = l;
Rijkl = −κi κj if i = l and j = k;

0 otherwise.

This says in particular than in the basis {ei ∧ ej : 1 ≤ i < j ≤ n} for


Λ2 (T M ), the curvature operator R is diagonalised, with eigenvalue κi κj in
the direction ei ∧ ej .
The sectional curvature of the plane generated by the principal directions
ei and ej is just κi κj . In particular this is positive if the surface bends in
the same direction in along the lines in directions ei and ej (that is, either
the hypersurface bends towards the normal in both of these directions, or
it bends away from the normal in both of these directions). The sectional
curvature is negative if the surface bends in opposite directions along the
two directions in the hypersurface. In particular, the principal curvatures
and sectional curvatures of a locally convex hypersurface are all positive.
14.4. THE SPHERE 129

Note that a hypersurface for n = 3 cannot have all sectional curvatures


negative, since this would imply that all of the principal curvatures have
different signs.
There are several other curvature quantities which can be conveniently de-
fined for hypersurfaces: The mean curvature H is the trace of the second
fundamental form:
h = g ij hij .
In particular, computing in the basis of principal directions shows that the
mean curvature is the sum of the principal curvatures H = κ1 + · · · + κn .
The Gauss curvature K is the determinant of the Weingarten map:
n
det(h) Y
K = det(W) = = κi .
det(g)
i=1

14.4. The sphere

Some important cases which are easy to compute are the following:
If M is a hyperplane, so that F (x)·z = c for some unit vector z and constant
c, then tangent vector satisfy XF · z = 0 for all X ∈ T M , so we can take
ν = z. Since z is constant, we have W(X) = 0 for all X, and so also
h(X, Y ) = 0. Thus hyperplanes have all principal curvatures (and mean
and Gauss curvatures) equal to zero.
If M is a sphere, then we have |F (x) − p|2 = r2 for some centre p and radius
r. Differentiating gives
0 = 2(F − p) · XF
for all X, and so we can choose ν(x) = F (x)−p
r . Differentiating gives
 
F (x) − p 1
W(X)F = Xν = X = ( X)F
r r
and we conclude that W = 1r Id. By the Weingarten relation we have h =
1 1
r g. The principal curvatures are therefore identically equal to r , the mean
curvature is equal to nr , and the Gauss curvature equals r−n . By the Gauss
equation, every sectional curvature is equal to r−2 , and the curvature tensor
is given by
1
Rijkl = 2 (gik gjl − gjk gil ) .
r
We remark that a locally convex hypersurface (with κi > 0) has positive
sectional curvatures. However it is not possible to find a hypersurface (for
n ≥ 3) with negative sectional curvatures, since this requires all principal
curvatures to have different signs. In the next section we will construct
examples of Riemannian manifolds with negative sectional curvature as hy-
persurfaces in Minkowski space.
130 14. HYPERSURFACES AND SUBMANIFOLDS

14.5. Minkowski space and spacelike hypersurfaces

Minkowski space Rn,1 (the space-time of special relativity) is the set Rn+1 =
{x = (x0 , x1 , · · · , xn ) : xi ∈ R, i = 0, · · · , n} equipped
P with the non-
degenerate symmetric bilinear form hx, yi = −x0 y 0 + ni=1 xi y i . Here x0
represents the ‘time’ direction, and x1 , · · · , xn represent ‘space’ directions.
Vectors in Rn,1 can be classified as follows:
• x is called space-like if hx, xi > 0;
• x is called time-like if hx, xi < 0;
• x is called light-like or null if hx, xi = 0.
The space of light-like vectors is the right circular cone {(x0 , ~x) ∈ R × Rn :
|x0 | = |~x|}. The light-like vectors with |x0 | > |~x| are those ‘inside’ the
cone, which separate into ‘future’ and ‘past’ according to whether x0 > 0 or
x0 < 0. The spacelike vectors are those ‘outside’ the cone, where |~x| > |x0 |.
Definition 14.5.1. An immersion F : M n → Rn,1 is called spacelike (and
the image is called a spacelike hypersurface if DF (X) is spacelike for every
X ∈ T M . Equivalently, the induced bilinear form
g(X, Y ) = hDF (X), DF (Y )i
is a Riemannian metric on M .
Example
Pn 14.5.2 (Graphical hypersurfaces). Consider the case where F (x1 , · · · , xn ) =
i 1 n
i=1 x ei + u(x , · · · , x )e0 . Then the tangent vectors are

∂i F = ei + (∂i u)e0 ,
so we have induced bilinear form
gij = h∂i F, ∂j F i = δij − (∂i u)(∂j u).
To determine when this is positive definite, fix x and choose an orthonor-
Du(x)
mal basis {ei } for Rn such that e1 = |Du(x)| (or arbitrary if Du(x) = 0).
Therefore ∂i u(x) = 0 for i > 1 and ∂i u = |Du|. Then we have
g11 = 1 − |Du|2 ; gii = 1 for i > 1,
and gij = 0 for i 6= j. Thus the eigenvalues of g are equal to 1 and 1 − |Du|2 ,
and g is positive definite if and only if |Du(x)| < 1. That is, a graphical
hypersurface is spacelike if and only if it is the graph of a function with
gradient less than 1 everywhere.

We can develop the geometry of spacelike hypersurfaces in Minkowski space


in a way which is closely analogous to what we did for Euclidean hypersur-
faces:
14.5. MINKOWSKI SPACE AND SPACELIKE HYPERSURFACES 131

We define the induced inner product by

g(X, Y ) = hXF, Y F i.

At each point we can find a normal vector ν(x) such that hν(x), XF i = 0
for each X ∈ Tx M . This can be done by starting with the vector e0 and
then using Gram-Schmidt orthogonalisation to produce a vector orthogonal
to the tangent vectors ∂1 F, · · · , ∂n F :

n = e0 − g ij h∂i F, e0 i∂j F.

By construction this is orthogonal to ∂i F for each i. We have

hn, e0 i = −1 − g ij h∂i F, e0 ih∂j F, e0 i < 0,

and

hn, ni = −1 − g ij h∂i F, e0 ih∂j F, e0 i < 0,


p
so n is a future time-like vector. Dividing by 1 + g ij h∂i F, e0 ih∂j F, e0 i gives
a future timelike normal vector ν with hν, νi = −1.
We now have a decomposition of Rn,1 at each point F (x) into a tangential
subspace DF (Tx M ) and the normal subspace generated by ν(x). In partic-
ular the derivatives of tangent and normal vectors can be expressed using
this decomposition as follows:

X(Y F ) = (∇X Y )F + h(X, Y )ν;

and

X(ν) = (W(X))F.

This defines the second fundamental form h, the Weingarten map W, and
the connection ∇ on the tangent bundle. By almost identical computations
to the Euclidean case we deduce that h is a symmetric (2, 0)-tensor, W is
a (1, 1)-tensor, and ∇ is the Riemannian (Levi-Civita) connection of the
Riemannian metric g. Note the change of sign compared to the Euclidean
hypersurface case. We note however that the Weingarten relation holds as
before: Differentiating the orthogonality relation hY F, νi we find

0 = hX(Y F ), νi + hY F, (W(X))F i = h(X, Y )hnu, nui + g(W(X), Y ).

Since hν, νi = −1 we deduce that h(X, Y ) = g(W(X), Y i just as in the


Euclidean case.
132 14. HYPERSURFACES AND SUBMANIFOLDS

The sign change in the expression for X(Y F ) results in a sign change in the
Gauss equation: As before we derive this using the identity
0 = X(Y (ZF ))) − Y (X(ZF )) − [X, Y ](ZF )
= X((∇Y Z)F + h(Y, Z)ν) − Y ((∇X Z)F + h(X, Z)ν)
− (∇[X,Y ] Z)F − h([X, Y ], Z)ν
= (∇X (∇Y Z))F + h(X, ∇Y Z)ν + X(h(Y, Z))ν + h(Y, Z)(W(X))F
− (∇Y (∇X Z))F − h(Y, ∇X Z)ν + Y (h(X, Z))ν − h(X, Z)(W(Y ))F
− (∇[X,Y ] Z)F − h([X, Y ], Z)ν.
The tangential part of this (and taking an inner product using the Wein-
garten relation) gives the Gauss identity
R(X, Y, W, Z) = h(Y, W )h(X, Z) − h(X, W )h(Y, Z).
The normal part gives the Codazzi identity which is unchanged from the
Euclidean case.
In particular, if {e1 , · · · , en } is an orthonormal frame of principal directions
for h with eigenvalues κi , then the sectional curvature of the plane generated
by ei and ej is equal to −κi κj . In the case of a graphical hypersurface where
the function u is convex (with gradient bounded by 1) all of the principal
curvatures κi are positive, and hence all of the sectional curvatures are
negative, in direct contrast to the Euclidean case.

14.6. Hyperbolic space

A very natural example of a spacelike hypersurface is the future timelike


unit sphere Hn , which we also call hyperbolic space. This is the hypersurface
{x : hx, xi = −1, x0 > 0}, which is the graph of the function u(x) =
1 + |x|2 . Since the gradient is ∂i u = √ x 2 with length √ |x| 2 < 1, this
p i

1+|x| 1+|x|
is a spacelike hypersurface.
We can compute the geometric invariants of hyperbolic space is direct anal-
ogy with the computations we carried out for the sphere in Euclidean space:
First we observe that differentiating the equation hF (x), F (x)i = −1 gives
hF, XF i = 0, so that the future timelike unit normal is given by ν = F .
Differentiating gives
W(X)F = Xν = XF
so that W is the identity map and h = g. This implies that all principal
curvatures are equal to 1 and all sectional curvatures are equal to −1. Thus
the hyperbolic space is a space of constant sectional curvatures equal to
−1. An analogue of the embedding theorem implies conversely that any
Riemannian manifold with constant sectional curvatures equal to −1 can be
14.7. SUBMANIFOLDS 133

embedded as a spacelike hypersurface in Minkowski space Rn,1 with second


fundamental form h = g, and therefore is locally isometric to Hn .
There are other descriptions of hyperbolic space which are sometimes easier
to work with. We will describe two of these, which arise from particular
parametrisations of hyperbolic space:
First, consider the stereographic projection of Hn from the ‘pole’ (1, 0, · · · , 0)
of the past timelike unit sphere onto the unit ball B 1 (0) ⊂ Rn ' Rn × {0} ⊂
Rn,1 . This map is given explicitly as follows: We have F (x) = (−1, 0) +
λ(1, x) where λ > 0 is chosen so that F (x) ∈ Hn . This means
−1 = hF (x), F (x)i
= h(λ − 1, λx), (λ − 1, λx)i
= −(λ − 1)2 + λ2 |x|2
= −1 + 2λ + λ2 (|x|2 − 1).
2
This means we must have λ = 1−|x|2
. Computing the induced metric on
B 1 (0), we have
2 4xi
∂i F = (0, ei ) + (1, x),
1 − |x|2 (1 − |x|2 )2
and so
gij = h∂i F, ∂j F i
4 16xi xj 2 16xi xj
= δij + (|x| − 1) +
(1 − |x|2 )2 (1 − |x|2 )4 (1 − |x|2 )3
4
= δij .
(1 − |x|2 )2
4
The metric gij = (1−|x| 2 )2 δij is called the Poincaré metric. Since it is pro-

duced by a diffeomorphism to the metric on hyperbolic space, this is a


Riemannian metric of constant sectional curvature equal to −1.

14.7. Submanifolds

Now consider the more general situation of a submanifold of arbitrary codi-


mension. The most important difference is that the normal bundle N M is
now of dimensions greater than 1, so it no longer suffices to consider a single
unit normal vector field. However, we can proceed in a suitably modified
way: As before, we define the induced Riemannian metric of an immersion
F : M → RN by g(X, Y ) = XF · Y F for each X, Y ∈ Tx M . Then as
before the connection is defined by the tangential part of the derivative of a
tangent vector field:
(∇X Y )F = π (X(Y F ))
134 14. HYPERSURFACES AND SUBMANIFOLDS

where π is the orthogonal projection onto the tangential subspace DF (Tx M )


of RN . We will denote the complementary projection onto the normal space
Nx M by π ⊥ . Then the second fundamental form is the tensor h ∈ Γ(T ∗ M ⊗
T ∗ M ⊗ N M ) defined by
h(X, Y ) = −π ⊥ (X(Y F )) ∈ Nx M.
That is, the connection and the second fundamental form give the decompo-
sition of the derivative of a tangent vector field into tangential and normal
components:
(14.7.1) X(Y F ) = (∇X Y )F − h(X, Y ).
The only difference is that instead of being real-valued, the second funda-
mental form of a submanifold has values in the normal space.
We can do a similar decomposition for sections of the normal bundle: If
ξ ∈ Γ(N M ), then we can decompose the derivative of ξ into tangential and
normal components, which we write as follows:
(14.7.2) X(ξ) = (W(X, ξ))F + ∇⊥
X ξ,
where ∇⊥ ⊥
X ξ = π (X(ξ)).

Proposition 14.7.1. W ∈ Γ(T ∗ M ⊗ N ∗ M ⊗ T M ), and ∇⊥ is a connec-


tion on the vector bundle N M , compatible with the induced metric given by
g ⊥ (ξ, η) = ξ · η.

Proof. We have
(W(X, f ξ))F + ∇⊥
X (f ξ) = X(f ξ)
= Xf ξ + f X(ξ)
= (f W(X, ξ))F + (f ∇⊥
X ξ + Xf ξ)
where we uses the derivation rule for vector fields in the second equality, and
equation (14.7.2) in the other two. Separating into tangential and normal
components, we find
W(X, f ξ) = f W(X, ξ)
and
∇⊥ ⊥
X (f ξ) = Xf ξ + f ∇X ξ,
showing as required that W is a tensor and ∇⊥ is a connection.
The compatibility of the metric and connection on N M is given by the
following (very similar to the argument for the tangent bundle):
Xg ⊥ (ξ, η) = X(ξ · η)
= Xξ · η + ξ · Xη
= π ⊥ (Xξ) · η + ξ · π ⊥ (Xη)
= g ⊥ (∇⊥ ⊥ ⊥
X ξ, η) + g (ξ, ∇X η)
14.7. SUBMANIFOLDS 135

as required. Here we used the fact that π(Xξ) · η = 0 since π(Xξ) is


tangential while η is normal. 

We also note the submanifold version of the Weingarten relation: Differen-


tiating Y F · ξ = 0 in direction X gives

0 = X(Y F ) · ξ + Y F · X(ξ)
= −h(X, Y ) · ξ + Y F · W(X, ξ)F
= −g ⊥ (h(X, Y ), ξ) + g(Y, W(X, ξ)).

14.7.1. Gauss, Codazzi and Ricci identities. We derive identities


from the symmetry of second derivatives of tangent vector fields exactly as
in the hypersurface case. In the higher-codimension case we have further
identities arising from the symmetry of second derivatives of normal vector
fields also.
First the tangential case: We can make the computation a little simpler by
assuming that we are working with vector fields X, Y, Z such that at some
chosen point we have ∇X = ∇Y = ∇Z = 0 (for example, this is true at
x if X, Y, Z are coordinate tangent vector fields for exponential coordinates
about x). Then we also have [X, Y ] = 0 at this point, and so


0 = X(Y (ZF )) − Y (X(ZF ) − 
[X,
Y](ZF )


= X((∇Y Z)F − h(Y, Z)) − Y ((∇X Z)F − h(X, Z))


 ⊥
= (∇X (∇Y Z))F −  ∇
h(X, Y Z) − ∇X (h(Y, Z)) − W(X, h(Y, Z))F
 ⊥
− (∇Y (∇X Z))F +  ∇X Z) + ∇Y (h(X, Z)) + W(Y, h(X, Z))F

h(Y,

= (R(Y, X, Z) − W(X, h(Y, Z)) + W(Y, h(X, Z)))F + (∇Y h)(X, Z) − (∇X h)(Y, Z),

where we used ∇⊥
 
∇
 
X (h(Y, Z)) = (∇X h)(Y, Z)+ h(∇ X Y, Z)+
h(Y, X Z). This
  
gives the Gauss equation as the tangential part:

R(X, Y, W, Z) = g ⊥ (h(X, W ), h(Y, Z)) − g ⊥ (h(Y, W ), h(X, Z));

and the Codazzi identity as the normal part:

(∇X h)(Y, Z) = (∇Y h)(X, Z).


136 14. HYPERSURFACES AND SUBMANIFOLDS

Next we compute the normal identities, using the same idea (here also choos-
ing ξ such that ∇⊥ ξ = 0 at the given point):
0 = X(Y ξ) − Y (Xξ) − 
[X,
Y

= X(∇⊥ ⊥
Y ξ + (W(Y, ξ))F ) − Y (∇X ξ + (W(X, ξ))F )

= ∇⊥ ⊥
X (∇Y ξ) + (∇X W)(Y, ξ))F − h(X, W(Y, ξ))

− ∇⊥ ⊥
Y (∇X ξ) − (∇Y W)(X, ξ))F + h(Y, W(X, ξ))

= R⊥ (Y, X, ξ) − h(X, W(Y, ξ) + h(Y, W(X, ξ))


+ ((∇X W)(Y, ξ) − (∇Y W)(X, ξ)) F.
The normal component of this identity gives the curvature R ⊥ of the con-
nection ∇⊥ on N M , in terms of the second fundamental form and Wein-
garten map. This is called the Ricci identity. Taking an inner product with
another normal vector η, it becomes
R⊥ (X, Y, ξ, η) = g(W(X, ξ), W(Y, η)) − g(W(X, η), W(Y, ξ)).
The tangential component gives the identity
(∇X W)(Y, ξ) = (∇Y W)(X, ξ).
This is (by virtue of the Weingarten relation and the compatability of the
connections and metrics — which imply that ∇g = 0 and ∇g ⊥ = 0) equiv-
alent to the Codazzi identity.
Bibliography

137

You might also like