Riemannian geometry

Martin Taylor∗
Spring 2022

Contents
1 Introduction 2
1.1 What is Riemannian geometry? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Why the need for abstraction? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Outline of the course . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Books . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Smooth manifolds 5
2.1 Topological and smooth manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Examples of smooth manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Smooth maps and diffeomorphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3 Tangent spaces and the tangent bundle 8

3.1 Submanifolds of Rn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2 Tangent spaces and the differential of a smooth map . . . . . . . . . . . . . . . . . . . . . . . 9
3.3 The tangent bundle and vector fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.4 The Lie bracket . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

4 Tensors, vector bundles and tensor fields 14

4.1 Cotangent spaces and the cotangent bundle . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.2 Multi-linear algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.3 Vector bundles and tensor bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.4 *Other important vector bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

5 Riemannian metrics 23
5.1 Expression in local coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.2 Lengths and angles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.3 *Volume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.4 Examples of Riemannian metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.5 Isometries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.6 Conformal maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.7 Existence of Riemannian metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.8 Musical isomorphisms and the inner product of tensor fields . . . . . . . . . . . . . . . . . . . 29

6 Flows of vector fields and the Lie derivative 31

6.1 Derivatives of vector fields in Rn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
6.2 The flow of a vector field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
6.3 The Lie derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
∗ [email protected]

1
7 Affine connections and the Levi-Civita connection 33
7.1 Connections and the covariant derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
7.2 Connections along curves and parallel transport . . . . . . . . . . . . . . . . . . . . . . . . . . 36
7.3 The Levi-Civita connection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
7.4 Divergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

8 Geodesics and Riemannian distance 42

8.1 Geodesics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
8.2 The exponential map and normal coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
8.3 Riemannian distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
8.4 Variations of curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
8.5 Geodesics are locally length minimising . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
8.6 Completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

9 Curvature 51
9.1 The Riemann curvature tensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
9.2 Sectional curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
9.3 Model spaces of constant curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
9.4 Ricci curvature and scalar curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

10 Submanifolds 59
10.1 Second fundamental form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
10.2 The Gauss equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

11 Jacobi fields 62
11.1 Definition and basic properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
11.2 The second variation formula, the index form, and conjugate points . . . . . . . . . . . . . . . 65

12 Classical theorems in Riemannian geometry 66

12.1 The Bonnet–Myers Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
12.2 The Cartan–Hadamard Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

13 *Lorentzian geometry and Penrose’s incompleteness theorem 68

13.1 Timelike, null, and spacelike vectors, curves, and submanifolds . . . . . . . . . . . . . . . . . 69
13.2 Time orientations and causal structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
13.3 Global hyperbolicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
13.4 Closed trapped 2-surfaces and black holes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
13.5 The Penrose Incompleteness Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
13.6 The Schwarzschild and Kerr families of black holes . . . . . . . . . . . . . . . . . . . . . . . . 71
13.7 Sketch proof of the Penrose Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

1 Introduction
These notes accompany the Riemannian geometry course at Imperial College London. Please email any
typos or suggestions (of which I expect there to be many) to [email protected]. Check the
blackboard page for the most up to date version of the notes. Sections of the notes marked with a ∗ contain
non-examinable material.
We begin with an informal introduction to the course.

2
1.1 What is Riemannian geometry?
The objects of concern in Riemannian geometry are manifolds. Informally (see Section 2.1 for a precise
definition), a topological manifold is a topological space which, moreover, “looks locally like Rn ”. Examples
are Rn itself, the sphere S n , products of these manifolds, appropriate quotients, etc.
If M is a topological manifold, then since M is, in particular, a topological space, one has a notion of
what it means for a function f : M → R to be continuous. In order to make sense of the geometric notions
which are introduced in this course it becomes necessary to perform calculus on such functions. In order to
entertain the notion of the differentiability of such a function f : M → R, one requires M to have additional
structure. A differentiable manifold is a topological manifold equipped with, in addition, a differentiable
structure. See Section 2.1 for a precise definition of a differentiable structure but, for now, one can think of
a differentiable structure as the necessary structure required in order to perform classical calculus on M.
The subject of differential topology is the study of differentiable manifolds and maps between them. But
what defines geometry?
In the world of differential topology one cannot tell the difference between donuts and coffee cups. Roughly
speaking, geometry is the study of properties, such as lengths, distances, angles, volume, etc., which allow
one to distinguish donuts from coffee cups.
Consider, for example, the notion of length. What structure is required on M in order to make sense of
the length of a curve γ : (a, b) → M?
Suppose first that M ⊂ Rn is an embedded manifold (and note that the embedding of M into Rn gives
M a “shape” or “geometry”). Given a curve in M, γ : (a, b) → M ⊂ Rn , one can simply forget about M
and consider γ merely as a curve in Rn , γ : (a, b) → Rn , for which one has the familiar formula for its length
Z bp
L(γ) = γ̇(s) · γ̇(s)ds. (1)
a
n
Here γ̇(s) ∈ R is the tangent vector to the curve γ at time s ∈ (a, b), and · denotes the Euclidean dot
product on Rn .
How does the definition (1) fail when M is no longer embedded in Rn ? If M is no longer embedded
then the tangent vector γ̇(s) is no longer an element of Rn , but now lives in the tangent space to M at γ(s),
denoted Tγ(s) M (see Section 3.2 for a precise definition of tangent space). The definition (1) then fails due
to the fact that the Euclidean dot product is not (canonically) defined on the tangent space Tγ(s) M.
One then sees that, in order to abstract the definition (1), the relevant structure one requires is an analogue
of the Euclidean dot product on each tangent space of M. Such a structure is known as a Riemannian metric.
To be (slightly) more precise (see Section 5 for a concrete definition), a Riemannian metric on M is a “smooth
assignment” of an inner product to each tangent space Tp M of M. Given such a Riemannian metric on M,
the length of γ : (a, b) → M is then defined as
Z bq
L(γ) = gγ(s) (γ̇(s), γ̇(s))ds. (2)
a

It turns out that a Riemannian metric gives rise not only to a notion of length, but also other geometric
notions such as distances, angles, volume, etc.
A differentiable manifold M together with a Riemannian metric g is known as a Riemannian manifold
(M, g). The subject of Riemannian geometry is the study of Riemannian manifolds.

1.2 Why the need for abstraction?

The whole motivating discussion of the previous section was driven by the desire to define geometric notions
on M, without invoking an embedding of M into Rn . It is worth taking a moment to address the obvious
question — why the need for abstraction? Why not define a Riemannian manifold to be a manifold M along
with an embedding ι : M → Rn , for some large n? Why are we so reluctant to invoke an embedding of M
when doing so would absolve us of many of our difficulties? These questions become even more pertinent in
view of the following two theorems (whose proofs are outside of the scope of the course).

3
Theorem 1.1 (Whitney 1936). If M is a smooth manifold of dimension n, there exists an embedding (an
immersion which is homeomorphic onto its image) F : M → R2n+1 .
In other words, in studying smooth manifolds one loses no generality in restricting to submanifolds of
Rn .

Theorem 1.2 (Nash 1956). If (M, g) is a Riemannian manifold of dimension n, there exists an isometric
(n+2)(n+3)
embedding F : M → R 2 .1
An isometric embedding will be defined later in the course. Until then you should view Theorem 1.2 as the
statement that, in the study of Riemannian manifolds, one loses no generality by restricting to submanifolds
of Rn .
It should be obvious by now that the reason for considering abstract manifolds is not because one loses
any generality in only considering embedded manifolds. The point is — it is typically the case that the
embeddings of Theorem 1.1 and Theorem 1.2 are very unnatural. This is perhaps difficult to appreciate
fully at the moment, when most of the manifolds one has encountered (such as Rn itself, or S n ) arise so
naturally as submanifolds of Rn . Most of the manifolds we are typically interested in, however, do not embed
so naturally. (One such example of a smooth manifold, which arises very naturally and will be defined in
Section 3.3, is the tangent bundle of the sphere T S n . Other examples are quotients of familiar spaces, such
as the projective plane. Hyperbolic space is a Riemannian manifold which does not embed naturally, neither
do many very naturally arising Riemannian manifolds in branches of physics such as classical mechanics
or general relativity.) One should also add that the proof of Theorem 1.2 is rather involved. It would be
very strange if one required such a heavy result in order just to make sense of the basic geometric notions
introduced in this course.
After a bit of (what can at first seem like, very) hard work in giving, and becoming familiar with, the
abstract definition of a Riemannian manifold, there is a huge payoff later when many of the objects which we
will introduce are defined in a way that is much more natural than they would be had we insisted on using
an embedding into Rn from the beginning. In other words, were one to insist on only considering embedded
manifolds, far from simplifying matters, the embeddings would become a major hinderance.

1.3 Outline of the course

The following is a brief outline of the topics covered in the course.

• Review of notions from differential topology: topological and smooth manifolds, tangent and cotangent
spaces, vector bundles, tensor bundles, Lie bracket, Lie derivative.
• Riemannian metrics, lengths, angles, existence of metrics.

• Derivatives of tensor fields, Lie derivative, affine connections, the Levi-Civita connection, parallel
transport.
• Geodesics, Riemannian distance, exponential map, Hopf–Rinow Theorem.
• Curvature: Riemann and Ricci curvature tensors, scalar curvature, sectional curvatures.

• Submanifolds: the second fundamental form and the Gauss equation.

• Jacobi fields and second variation of curves.
• Classical theorems in Riemannian geometry: Bonnet–Myers Theorem, Cartan–Hadamard Theorem.
• Time permitting: Lorentzian geometry, Penrose and Hawking singularity theorems.
1 In the original theorem of Nash, the dimension of the target, RN , of the embedding was much larger.

4
1.4 Books
Though the course will not strictly follow a textbook, the most relevant textbooks are [6] and [10]. The
relevant background on differential manifolds can be found in the book [11]. The books [2] and [14] were very
helpful in preparing the course, though they cover far more material than will be considered here. There
is no shortage of other good books on Riemannian geometry. See, for example, [3], [7], [9], [12], [13]. The
lecture notes [1], [5], [4] and the book [8] were helpful in preparing Section 13.

2 Smooth manifolds
The basic objects of study in this course are smooth manifolds.

2.1 Topological and smooth manifolds

Recall the following notions from topology.

Definition 2.1 (Hausdorff, second countable, and locally Euclidean topological spaces). A topological space
X is called:
• Hausdorff if, for all x, y ∈ X with x 6= y there exists open sets U, V ⊂ X such that x ∈ U , y ∈ V and
U ∩ V = ∅;
• Second countable if there exists a countable basis for the topology of X;

• Locally Euclidean of dimension n for some n ∈ N if, for all x ∈ X, there exists an open set U ⊂ X
such that x ∈ U , an open set V ⊂ Rn , and a homeomorphism φ : U → V .
These notions are used to define a topological manifold.
Definition 2.2 (Topological manifolds). Given n ∈ N, a topological manifold of dimension n is defined to
be a topological space M which is Hausdorff, second countable, and locally Euclidean of dimension n.
By far the most important component of Definition 2.2 is that M be locally Euclidean. The other two
conditions are imposed to rule out certain pathologies.
Exercise 2.3 (Non-Hausdorff and non-second countable spaces).

1. Give an example of a topological space X and n ∈ N such that X is second countable and locally
Euclidean of dimension n, but is not Hausdorff.
2. Give an example of a topological space X and n ∈ N such that X is Hausdorff and locally Euclidean of
dimension n, but is not second countable.

Given a topological manifold M, one has a notion of what it means for a function f : M → R to be
continuous. There is, however, no way to entertain the notion of such a function being smooth without
introducing additional structure.
Definition 2.4 (Charts and smooth atlases). Given n ∈ N, a smooth n-atlas of a topological space X is
a collection {(Uα , φα )}α∈A where {Uα }α∈A is an open cover of X, there exist open sets Vα ⊂ Rn such that
φα : Uα → Vα are homeomorphisms, and for all α, β ∈ A such that Uα ∩ Uβ 6= ∅, the map

φα ◦ φ−1
β : φβ (Uα ∩ Uβ ) → φα (Uα ∩ Uβ ),

is a smooth diffeomorphism.
Each element (Uα , φα ) of the atlas is called a chart.

5
 Throughout this course, smooth will always mean C ∞ . Though one can, of course, introduce less (or
more) regular notions of smooth atlases and smooth manifolds, for simplicity we will not do so here. We
often refer to a “smooth atlas”, by which we mean a “smooth n-atlas for some n ∈ N”.
Remark 2.5 (Remarks on Definition 2.4).

1. Note that there is no sense in which the maps φα and φβ of Definition 2.4 are individually smooth.
The composition φα ◦ φ−1 n
β , however, is a bonafide map between two open subsets of R , and we all know
n
very well what it means for a map between open subsets of R to be smooth.
2. There is no restriction on the size of the index set A in Definition 2.4. It could have one element or
could be uncountably infinite.

Definition 2.6 (Maximal atlases). A smooth atlas {(Uα , φα )}α∈A of a topological space X is called maximal
if, for any (U, φ) such that U ⊂ X is open and φ : U → V is a homeomorphism, for some V ⊂ Rn open, and
the map
φ ◦ φ−1
α : φα (Uα ∩ U ) → φ(Uα ∩ U ),

is a smooth diffeomorphism for all α ∈ A such that Uα ∩ U 6= ∅, then (U, φ) ∈ {(Uα , φα )}α∈A .
Given any atlas, there clearly exists a unique maximal atlas containing it. Indeed, take the union of
the atlas with all such (U, φ) as in Definition 2.6. Given a smooth atlas A, let max(A) denote the unique
maximal smooth atlas containing A.
Definition 2.7 (Smooth structures and smooth manifolds). A smooth structure on a topological manifold
is a maximal atlas. A smooth manifold of dimension n is a topological manifold of dimension n together
with a smooth structure.

The manifolds considered in this course will always be assumed to be connected. From now on, “Mn is
a manifold” means “M is a connected smooth manifold of dimension n”.
For each i = 1, . . . , n, let π i : Rn → R denote the projection onto the i-th component:

π i (x1 , . . . , xn ) := xi .

Definition 2.8 (Local coordinates). If Mn is a manifold and (U, φ) is a chart, the collection {π i ◦ φ | i =
1, . . . , n} is called a system of local coordinates of M. We typically abuse notation and write (x1 , . . . , xn )
as the local coordinates, with the understanding that xi = π i ◦ φ.
For brevity, one often says “let (U, φ, xi ) be a chart of M” to mean that (U, φ) is a chart of M and
x , . . . , xn are the associated system of local coordinates.
1

2.2 Examples of smooth manifolds

The most basic example of a smooth manifold is Rn with the Euclidean topology. The identity map Id : Rn →
Rn is a global chart. Hence {(Rn , Id)} is an atlas. The standard smooth structure on Rn is the maximal
atlas containing {(Rn , Id)}.
Consider now the topological manifold

S 1 = {(x, y) ∈ R2 | x2 + y 2 = 1},

with the subspace topology. Let A1 = {(U, φ), (V, ψ)}, where

U = {(cos θ, sin θ) | θ ∈ (0, 2π)}, V = {(cos θ, sin θ) | θ ∈ (−π, π)},

and φ : U → (0, 2π) ⊂ R is defined by φ(cos θ, sin θ) = θ (meaning the unique value of θ ∈ (0, 2π)) and
ψ : V → (−π, π) ⊂ R is defined by ψ(cos θ, sin θ) = θ (meaning the unique value of θ ∈ (−π, π)). Clearly

6
{U, V } is an open cover of S 1 and φ, ψ are homeomorphisms onto (0, 2π) and (−π, π) respectively. Now
φ(U ∩ V ) = (0, π) ∪ (π, 2π) and ψ(U ∩ V ) = (−π, 0) ∪ (0, π), and the transition map

−1 θ if θ ∈ (0, π),
ψ ◦ φ (θ) =
θ − 2π if θ ∈ (π, 2π),

is a smooth diffeomorphism. It follows that A1 is an atlas on S 1 , which induces a unique smooth structure
on S 1 .
Remark 2.9 (Slight abuse of notation). In accordance with the abuse of notation of Definition 2.8, when
it is clear from the context which chart is under consideration, we write θ ∈ S 1 when in fact we really mean
θ ∈ (0, 2π) and the point in S 1 under consideration is φ−1 (θ) ∈ S 1 . Though this abuse may seem strange at
first, it becomes very cumbersome to always include explicit reference to the chart φ.
Consider now A2 = {(UN , φN ), (US , φS )} where UN = S 1 r {(0, 1)}, US = S 1 r {(0, −1)} and φN , φS
denote stereographic projection from (0, 1) and (0, −1) respectively,

φN : UN → R, φS : US → R.

Exercise 2.10 (Atlas defined by stereographic projection).

1. Derive expressions for φN (x, y) and φS (x, y).
2. Show that A2 is an atlas.
One can similarly show that

S n = {(x1 , . . . , xn+1 ) ∈ Rn+1 | (x1 )2 + . . . + (xn+1 )2 = 1},

naturally admits a smooth n-manifold structure for all n ∈ N.

2.3 Smooth maps and diffeomorphisms

Let Mm and N n be manifolds. It is now possible to define what it means for a function f : M → R to be
smooth.
Definition 2.11 (Smooth functions). A function f : M → R is smooth at p ∈ M if there exists a chart
(U, φ) of M such that p ∈ U and f ◦ φ−1 : φ(U ) → R is smooth at φ(p) ∈ Rm .
The function f is smooth if it is smooth at every p ∈ M.
More generally, it is now possible to define what it means for a function F : M → N to be smooth.
Definition 2.12 (Smooth functions between smooth manifolds). A function F : M → N is smooth at
p ∈ M if there exist charts (U, φ) and (V, ψ) of M and N respectively such that p ∈ U , F (p) ∈ N ,
F (U ) ⊂ V ,
ψ ◦ F ◦ φ−1 : φ(U ) → ψ(V ),
is smooth at φ(p) ∈ Rm .
The function F is smooth if it is smooth at every p ∈ M.
Note that φ(U ) ⊂ Rm and ψ(V ) ⊂ Rn are open subsets and so the compositions f ◦ φ−1 and ψ ◦ F ◦ φ−1
of Definitions 2.11 and 2.12 respectively are maps between open subsets of Euclidean space and so it indeed
makes sense to talk about their smoothness.
Exercise 2.13 (Smoothness of a function is well defined). Show that smoothness of F is well defined, i.e.
that the definition does not depend on the choice of charts: if (U1 , φ1 ), (U2 , φ2 ) are charts of M and (V1 , ψ1 ),
(V2 , ψ2 ) are appropriate charts of N , then ψ1 ◦ F ◦ φ−1
1 is smooth at φ1 (p) if and only if ψ2 ◦ F ◦ φ−1 2 is
smooth at φ2 (p).

7
Exercise 2.14 (Composition of smooth functions is a smooth function). Suppose that M, N and P are
smooth manifolds and F : M → N and G : N → P are smooth. Show that G ◦ F : M → P is smooth.
Definition 2.15 (Smooth diffeomorphisms). A map F : M → N is called a smooth diffeomorphism if it is
a smooth bijection and F −1 : N → M is smooth.
Define
C ∞ (M, N ) = {smooth maps from M to N },
and
C ∞ (M) = C ∞ (M, R).
As a brief aside, one can ask the following question:
How may smooth structures are there, up to diffeomorphism, on Rn ?
As long as n 6= 4 the answer, perhaps unsurprisingly, is only one — the standard smooth structure defined
in Section 2.2. Surprisingly, on R4 there are uncountably many non-diffeomorphic smooth structures (the
names associated with these facts are Donaldson, Freedman, Taubes).
One can ask a similar question about S n . In 1956 Milnor constructed non-diffeomorphic smooth structures
on S 7 , a work for which he was awarded the Fields Medal in 1962. For most values of n it is still unknown
whether there exists a non-standard smooth structure on S n . Such spheres with non-standard smooth
structures are known as exotic spheres.
Note that the words “up to diffeomorphism” in the above question are important.
Exercise 2.16 (Two atlases on R such that the identity map is not smooth). Find two atlases A1 , A2 on
R such that (R, max(A1 )) and (R, max(A2 )) are diffeomorphic, but such that the identity map

Id : (R, max(A1 )) → (R, max(A2 )),

is not a smooth map. Here max(A1 ) and max(A2 ) denote the unique maximal atlases containing A1 and A2
repsectively.
Smooth curves are particular examples of smooth maps.
Definition 2.17 (Curves). For a, b ∈ R, a < b, a smooth map γ : (a, b) → M is call a (smooth) curve.

3 Tangent spaces and the tangent bundle

In the previous section we defined what it means for a function between two manifolds to be differentiable. In
this section we discuss what type of object the derivative of such a function is. Understanding this requires
first introducing tangent spaces.

3.1 Submanifolds of Rn
If M ⊂ Rn is an embedded manifold, one pictures the tangent space to M at the point p ∈ M as “the set of
directions at p” or “the best linear approximation to M at p”. Such definitions do not lend themselves well
to abstract manifolds. In this informal section we motivate the definition of tangent space given in Section
3.2 by examining more closely the situation for embedded manifolds.
Note instead that there is a correspondence between “directions at p ∈ M” and “directional derivatives
at p”. Indeed, one associates the direction v with the directional derivative in the direction v and vice versa.
For example, consider a smooth function f : Rn → R, a point p ∈ Rn and a direction v ∈ Rn (note that
the tangent space to Rn is canonically identified with Rn itself). Define the directional derivative of f at p
in the direction v as
d
Dv f |p = f (p + tv)|t=0 .
dt

8
The directional derivative Dv satisfies the product rule

Dv (f g)|p = Dv f |p g(p) + f (p)Dv g|p ,

for all smooth functions f and g.

Define a derivation at p ∈ Rn to be a linear map

X|p : {smooth functions at p} → R,

which satisfies the product rule

X|p (f g) = X|p (f ) · g(p) + f (p) · X|p (g),

for all smooth functions f, g.

Exercise 3.1 (Derivation/direction correspondence for submanifolds of Rn ). Show that, in Rn , for every
derivation X|p at p there exists a vector v ∈ Rn whose associated directional derivative at p is X|p , i.e. such
that Dv |p = X|p .
Show, moreover, that the set of derivations at p admits a natural vector space structure and is vector
space isomorphic to Rn .
The above discussion suggests that, rather than thinking of the tangent space to an embedded manifold
at a point p as being the “set of all directions at p”, one could equally view the tangent space as the “set of
all derivations at p”. The latter lends itself much better to generalisation to abstract manifolds.

3.2 Tangent spaces and the differential of a smooth map

Consider now a smooth manifold M. Given p ∈ M, define the space

C ∞ (p) = {functions f defined on an open neighbourhood U of p such that f : U → R is smooth at p}.

Note that the space C ∞ (p) is an algebra with respect to pointwise multiplication, i.e. if f, g ∈ C ∞ (p) then
f g ∈ C ∞ (p).
Definition 3.2 (Tangent vectors and tangent space).
• A derivation or tangent vector X at p ∈ M is a linear functional

X : C ∞ (p) → R,

which satisfies the Leibniz rule

X(f g) = Xf · g(p) + f (p) · Xg, (3)
for all f, g ∈ C ∞ (p).
• The tangent space to M at p, denoted Tp M, is the set of all tangent vectors to M at p.
Note the reason for including the Leibniz rule in the definition of a derivation.
Exercise 3.3 (Why the Leibniz rule?). Why is the Leibniz rule (3) included in Definition 3.2? More
precisely, where is the Leibniz rule used in Exercise 3.1? What additional types of objects would be tangent
vectors to Rn if (3) was not included in the definition?
Given X, Y ∈ Tp M and λ ∈ R, define

(X + Y )f = Xf + Y f, (λX)f = λXf,

for all f ∈ C ∞ (p).

9
Exercise 3.4 (The tangent space is a vector space). Show that the tangent space Tp M is a vector space
with the above addition and scalar multiplication operations.
Any curve γ : (−ε, ε) → M such that γ(0) = p ∈ M defines a derivation
γ̇(0) : C ∞ (p) → R,
by
d  df (γ(s)) 
γ̇(0)f :=  f ◦γ =  .
ds s=0 ds s=0
d
(Note that f ◦ γ is a map f ◦ γ : (−ε, ε) → R, and so we indeed know how to define ds of such a composition.)
∞
Indeed, if f, g ∈ C (p) then f ◦ γ, g ◦ γ : (−ε, ε) → R and
df (γ(s))g(γ(s)) 
γ̇(0)f g = = γ̇(0)f · g(p) + f (p) · γ̇(0)g,
ds

s=0

by the product rule on R.

Recall now that a chart (U, φ) of M defines a local coordinate system (x1 , . . . , xn ) (see Definition 2.8).
Definition 3.5 (Coordinate vectors). Given a chart (U, φ, xi ) of M and p ∈ U , define the coordinate vectors
∂  ∂ 
, . . . ,  ∈ Tp M,
∂x1 p ∂xn p


by
∂  ∂f ◦ φ−1 
f := , (4)
∂xi p ∂xi
 
φ(p)

for all f ∈ C ∞ (p).

Note that on the right hand side of (4) xi denotes the i-th standard coordinate on Rn , and that f ◦
−1
φ : φ(U ) → Rn , where φ(U ) ⊂ Rn is an open set.
Let ei denote the standard i-th basis vector and define the curves αi : (−ε, ε) → U ⊂ M for i = 1, . . . , n
by
αi (s) = φ−1 φ(p) + sei .

Exercise 3.6 (Coordinate vectors as tangent vectors to curves). Show that the tangent vectors to the curves
α1 , . . . , αn at s = 0 are the coordinate vectors of Definition 3.5,
∂ 
α̇i (0) =  .
∂xi p
∂ 
 
Exercise 3.7 (Coordinate vectors form a basis). Show that the coordinate vectors ∂x 1
p
, . . . , ∂x∂n p form a
basis of Tp M.
Exercise 3.7 in particular implies that any X ∈ Tp M can be uniquely written as
n
X ∂ 
X= Xi  , (5)
i=1
∂xi p

for some X 1 , . . . , X n ∈ R.
Throughout this course Einstein’s greatest contribution to differential
P geometry, namely the Einstein
summation convention, will be adopted, whereby the summation sign is omitted and repeated indices
indicate a summation.2 For example (5) is simply written as
∂ 
X = Xi i  .
∂x p
2 Einstein once joked to a friend “I have made a great discovery in mathematics; I have suppressed the summation sign every

time that the summation must be made over an index which occurs twice . . . .”

10
 Note in particular that, for any vector X ∈ Tp M, the curve γX : (−ε, ε) → M defined by
γX (t) = φ−1 (φ(p) + tX i ei ),
satisfies γ̇X (0) = X, i.e. every tangent vector arises as a tangent vector to some curve.
Exercise 3.8 (Change of coordinate vectors under change of coordinates). Consider two charts (U, φ) and
(U e with associated local coordinates (x1 , . . . , xn ) and (e
e , φ) x1 , . . . , x
en ) respectively. If p ∈ U ∩ U
e then

∂  ∂xj ∂ 
=  ,
∂exi p xi ∂xj p
∂e


where
∂xj
(p) := ∂i (π j ◦ φ ◦ φe−1 )(φ(p)).
e
xi
∂e
In particular, if p ∈ U ∩ U
e and X ∈ Tp M, then

∂  e i ∂  ,

X = Xi = X
∂xi p xi p
∂e


where
i
Xi = X e j ∂x (p).
xj
∂e
Given a smooth function F : M → N , the derivative of F at p ∈ M is defined to be a linear map from
Tp M to TF (p) N .
Definition 3.9 (Pushforward). Given p ∈ M, the map F∗p : Tp M → TF (p) N , also denoted dF |p , defined by
(F∗p X)(h) := X(h ◦ F ),
for all X ∈ Tp M and h ∈ C ∞ (F (p)), is called the differential or derivative of F at p. The vector F∗p X is
called the pushforward of X by F .
Exercise 3.10 (Pushforward of tangent vector of a curve). Show that F∗p is a linear map and that, if
γ : (−ε, ε) → M is such that γ̇(0) = X, then
dh(F (γ(s))) 
F∗p X(h) =  ,
ds s=0

for all h ∈ C ∞ (F (p)).

Exercise 3.11 (Properties of pushforward). Suppose F ∈ C ∞ (M, N ). Show the following properties of the
differential of F .
1. If F is constant then F∗p ≡ 0 for all p ∈ M.
2. The differential satisfies the chain rule: if G ∈ C ∞ (N , P) then
(G ◦ F )∗p = G∗F (p) ◦ F∗p : Tp M → TG◦F (p) P,
for all p ∈ M.
3. The identity map satisfies (IdM )∗p = IdTp M : Tp M → Tp M for all p ∈ M.
4. If F is a diffeomorphism then F∗p is a vector space isomorphism.
5. Given charts (U, φ) and (V, ψ) of Mm and N n respectively with coordinate systems (x1 , . . . , xm ) and
(y 1 , . . . , y n ) respectively, with p ∈ U , F (p) ∈ V , then
 ∂   ∂(ψ ◦ F ◦ φ−1 )j ∂ 
F∗p = (φ(p)) .

∂xi p ∂xi ∂y j F (p)
 

11
3.3 The tangent bundle and vector fields
The goal of this section is to introduce vector fields. One should view a vector field as a “smooth assignment
of a vector at each p ∈ M”. Note then that a vector field should in particular be a map
[
X: M → {p} × Tp M,
p∈M

which has the property that Xp ∈ Tp M for all p ∈ M.

S In order to call such a map “smooth” it is necessary
to define a smooth manifold structure on the space p∈M {p} × Tp M.
Definition 3.12 (Tangent bundle). The tangent bundle of a manifold M, denoted T M, is defined to be
the disjoint union of the tangent spaces at each point in M,
[
TM = {p} × Tp M.
p∈M

Points in T M are typically denoted (p, v) or vp ∈ T M. The projection map π : T M → M defined by

π(p, v) = p is surjective and π −1 (p) = {p} × Tp M for all p ∈ M.
The smooth structure on M can be used to define a topology and smooth structure on T M so that π
is continuous and smooth. If dimM = n then dimT M = 2n. Indeed, given a chart (U, φ) for M, define
VU := ∪p∈U {p} × Tp M and define Φφ : VU → φ(U ) × Rn ⊂ R2n by

Φφ (p, v i ∂xi |p ) := (φ(p), v 1 , . . . , v n ).

Clearly Φφ is a bijection onto its image. The topology of T M is defined so that each Φφ is a homeomor-
phism onto its image.
If {(Uα , φα )}α∈A is an atlas for M then {(VUα , Φφα )}α∈A is an atlas for T M. Indeed, {VUα }α∈A is
clearly an open cover of T M since {Uα }α∈A is an open cover of M. Let U , U e be such that U ∩ U
e 6= ∅ and
1 n n
consider the corresponding Φφ , Φφe. If p ∈ U ∩ U and (v , . . . , v ) ∈ R then, by Exercise 3.8,
e

∂xj
v i ∂xei |p = v i ∂ j |p ,
xi x
∂e
and so
1 n
Φφ ◦ Φ−1 1 n e−1 (q), v i ∂x (p), . . . , v i ∂x (p)),
e (q, v , . . . , v ) = (φ ◦ φ
φ xi
∂e xi
∂e
with q = φ(p).
e The smoothness of this map then follows from the smoothness of φ ◦ φe−1 .
When one refers to the tangent bundle T M of M, one typically does not just mean the set T M as
defined in Definition 3.12, but rather the set T M equipped with this smooth manifold structure.
Definition 3.13 (Vector fields). Let M be a manifold. A vector field on M is a smooth map X : M → T M
such that π ◦ X = IdM .
Denote the set of all vector fields on M by

X(M) = {vector fields on M}.

The space X(M) is a vector space with the operations

(aX + bY )|p = aX|p + bY |p ,

for all p ∈ M, X, Y ∈ X(M) and a, b ∈ R. Moreover, given a smooth function f ∈ C ∞ (M), define
f X : M → T M by f X|p := f (p)X|p , so that f X ∈ X(M). The space X(M) is then linear over C ∞ (M)
(note that C ∞ (M) is a ring, but not a field).

12
Example 3.14 (Coordinate vector fields). Given a chart (U, φ) of M with local coordinates (x1 , . . . , xn ),
the maps
∂
: U → π −1 (U ),
∂xi
∂ ∂
for i = 1, . . . , n, defined by ∂xi (p) = ∂xi |p ∈ Tp M (see Definition 3.5) are vector fields on U called coordinate
vector fields.
∂ ∂
Given X ∈ X(M) and a chart (U, φ), since ∂x 1 |p , . . . , ∂xn |p form a basis of Tp M for each p ∈ U , one can

write in U
∂ 
X = X i (p) i  ,
∂x p
for some smooth functions X i : U → R, for i = 1, . . . , n, called the components of X with respect to (U, φ).
∂f
Given a chart (U, φ) of M and f ∈ C ∞ (U ), note that ∂x i ∈ C
∞
(U ) for i = 1, . . . , n, where

∂f ∂ 
(p) :=  f.
∂xi ∂xi p

3.4 The Lie bracket

Finally, there is a natural way of “multiplying” two vector fields to produce a third. This “multiplication”
operation is known as the Lie bracket.

Definition 3.15 (Lie bracket). The Lie bracket of X, Y ∈ X(M) is defined as

[X, Y ]f := X(Y f ) − Y (Xf ),

for all f ∈ C ∞ (M).

Exercise 3.16 (Properties of Lie bracket).

1. Show that [X, Y ] ∈ X(M).
∂ i ∂
2. Show that, if X = X i ∂x i and Y = Y ∂xi in some local coordinates then

 ∂Y j ∂X j  ∂f
[X, Y ]f = X i i − Y i .
∂x ∂xi ∂xj

Exercise 3.17 (Further properties of Lie bracket). Suppose X, Y, Z ∈ X(M), a, b ∈ R and f, g ∈ C ∞ (M).
Show that the Lie bracket has the following properties.

• [X, Y ] = −[Y, X] (anti-commutitivity);

• [aX + bY, Z] = a[X, Z] + b[Y, Z] (R linearity);
• [[X, Y ], Z] + [[Z, X], Y ] + [[Y, Z], X] = 0 (Jacobi identity);
• [f X, gY ] = f g[X, Y ] + f X(g)Y − gY (f )X.

One in particular sees from the latter property that [·, ·] is not linear over the ring C ∞ (M).
Though introduced as way of “multiplying” two vector fields, one should really view the Lie bracket as
a way of differentiating vector fields. We will return to this comment in Section 6.3.

13
4 Tensors, vector bundles and tensor fields
Recall from Section 1.1 that a Riemannian metric, i.e. the relevant object for entertaining geometric concepts,
on a manifold M should be, for each p ∈ M, a map Tp M × Tp M → R, which moreover depends smoothly
on p. In order to fulfil the latter requirement, it is necessary to introduce a smooth manifold structure on
the space [
{p} × {Bilinear maps : Tp M × Tp M → R}. (6)
p∈M

Such a space with such a smooth structure is an example of a vector bundle. For fixed p, such an element of
such a vector space is an example of a tensor. Such a smooth assignment of an element of each such space
is an example of a tensor field.

4.1 Cotangent spaces and the cotangent bundle

A simpler example of a vector bundle than (6) is the cotangent bundle which is introduced in this section.3
Given a finite dimensional vector space V , define the dual space of V as
V ∗ = {L : V → R | L is linear}.
The dual space V ∗ is a vector space with the obvious addition and scalar multiplication operations. Elements
of V ∗ are called covectors.
Throughout Section 4, vector spaces which are denoted V are always assumed to be finite dimensional.
Exercise 4.1 (Identification of a vector space with its bidual). Given a (finite dimensional) vector space V
and an element v ∈ V , define δv : V ∗ → R by δv (L) := L(v). Show that the map Φ : V → (V ∗ )∗ defined by
Φ(v) := δv ,
is a vector space isomorphism.
Remark 4.2 (Another abuse of notation). Exercise 4.1 implies that V ∼ = (V ∗ )∗ in a canonical way. It is
worth emphasising the latter point. We know very well that any two finite dimensional vector spaces of the
same dimension are isomorphic. In fact there are infinitely many isomorphisms and, for two general vector
spaces of the same dimension, none of these isomorphisms is preferred. By Exercise 4.1 there is one preferred
isomorphism between V and (V ∗ )∗ . As such, we view V and (V ∗ )∗ as the same space and write
V = (V ∗ )∗ ,
with this slight abuse being understood as an identification of the two spaces using the canonical isomorphism
Φ of Exercise 4.1.
Remark 4.3 (Yet another abuse of notation). In the spirit of Remark 4.2, here is a natural place to make the
following observation regarding another slight abuse of notation which will be used frequently. Given p ∈ Rn ,
the existence of the canonical (global) Cartesian coordinate system on Rn means that there is a canonical
isomorphism between the tangent space Tp Rn and Rn itself. Indeed, any X ∈ Tp Rn can be uniquely written
as
∂ 
X = Xi i  ,
∂x p
where (x , . . . , x ) are the standard Cartesian coordinates on Rn . One then defines a canonical isomorphism
1 n

Φ : Tp Rn → Rn by
 ∂  
Φ X i i  := (X 1 , . . . , X n ).
∂x p
One easily checks that Φ is a vector space isomorphism. In view of this canonical isomorphism, we again
abuse notation slightly and view Tp Rn and Rn as the same space, implicitly meaning they are identified via
Φ.
3 An even simpler example is the tangent bundle, introduced in Section 3.3.

14
Definition 4.4 (Cotangent space). Given a manifold M and a point p ∈ M, define the cotangent space at
p, denoted Tp∗ M, to be the dual of the tangent space,
Tp∗ M := (Tp M)∗ .
Elements of Tp∗ M are called cotangent vectors or covectors at p.
Given a smooth function f : M → R and p ∈ M, its differential at p, f∗p , is also denoted dfp and, using
the abuse of Remark 4.3, is a map
dfp : Tp M → R.
Exercise 4.5 (Coordinate covectors). Given a chart (U, φ) of M with local coordinates xi : U → R, for
i = 1, . . . , n, show that, if p ∈ U then
dx1p , . . . , dxnp ,
 
forms a basis of Tp∗ M. Moreover, dx1p , . . . , dxnp is the dual basis of the basis ∂ 1  , . . . , ∂n  of Tp M, i.e.
∂x p ∂x p
 ∂  
dxip  = δji .

∂xj p
Exercise 4.6 (Differential of a function in local coordinates). If f : M → R is a smooth function and
(U, φ, xi ) is a chart for M, show that
∂   ∂  
f = df  ,

p
∂xi p ∂xi p

and hence
∂f 
dfp =  dxi ,
∂xi p p
where
∂f  ∂ 
:=  f.
∂xi p ∂xi p


Definition 4.7 (Cotangent bundle). The cotangent bundle of a manifold M, denoted T ∗ M, is the disjoint
union of the cotangent space at each point in M:
[
T ∗ M := {p} × Tp∗ M.
p∈M

The map π : T M → M defined, for p ∈ M, ω ∈ Tp∗ M, by π(p, ω) = p is called the natural projection.
∗

Similar to T M, the cotangent bundle T ∗ M admits a topology and smooth structure (of a 2n dimensional
manifold) such that π is smooth.
Exercise 4.8 (Topology and smooth structure of cotangent bundle). Fill in the details of the definition of
the topology and smooth structure of T ∗ M.
The analogue of a vector field (see Definition 3.13) for the cotangent bundle is called a one form.
Definition 4.9 (One forms). A one form on M is a smooth map ξ : M → T ∗ M such that π ◦ ξ = IdM .
If f : M → R is a smooth function, then df , defined by
df (p) := dfp ,
is a one form on M.
Define the space of one forms on M
Γ(T ∗ M) = {one forms on M}.
Clearly Γ(T ∗ M) is a vector space, with the obvious operations.
Note that d can be viewed as a map
d : C ∞ (M) → Γ(T ∗ M). (7)
This comment will be briefly returned to in Section 4.4.

15
4.2 Multi-linear algebra
Before defining more general vector bundles, it is convenient to first recall some notions from (multi-) linear
algebra.
Definition 4.10 (Multi-linear maps). If V1 , . . . , Vk are (finite dimensional) vector spaces, a function T : V1 ×
. . . × Vk → R is called multi-linear if

T (v1 , . . . , vi−1 , avi + bṽi , vi+1 , . . . , vk ) = aT (v1 , . . . , vi , . . . , vk ) + bT (v1 , . . . , ṽi , . . . , vk )

for all v1 ∈ V1 , . . . , vi , ṽi ∈ Vi , . . . , vk ∈ Vk and a, b ∈ R.

Definition 4.11 (Tensor product of vector spaces). The tensor product of vector spaces V1 , . . . , Vk is defined
as
V1 ⊗ . . . ⊗ Vk = {T : V1∗ × . . . × Vk∗ → R | T is multi-linear}.
Note that if k = 1 then {T : V1∗ → R | T is linear} = (V1∗ )∗ and so the notation of Definition 4.11 is
consistent, provided one uses the slight abuse of notation of Remark 4.2.
Definition 4.12 ((r, s) tensors). Given a vector space V , elements of the tensor product

Vsr := V ⊗ . . . ⊗ V ⊗ V ∗ ⊗ . . . ⊗ V ∗ ,
| {z } | {z }
r times s times

are called (r, s) tensors, or r-contravariant, s-covariant tensors.

The space Vsr is a vector space with the obvious operations.
Definition 4.13 (Tensor product of tensors). Given a vector space V and T ∈ Vsr , T̃ ∈ Vs̃r̃ , for some
r+r̃
r, r̃, s, s̃ ≥ 0, define the tensor product of T and T̃ , denoted T ⊗ T̃ , to be the element of Vs+s̃ defined by

(T ⊗ T̃ )(L1 , . . . , Lr+r̃ , v1 , . . . , vs+s̃ ) := T (L1 , . . . , Lr , v1 , . . . , vs )T̃ (Lr+1 , . . . , Lr+r̃ , vs+1 , . . . , vs+s̃ ),

for all L1 , . . . , Lr+r̃ ∈ V ∗ , v1 , . . . , vs+s̃ ∈ V .

The tensor product operation is
• Linear:

(T1 + T2 ) ⊗ T3 = T1 ⊗ T3 + T2 ⊗ T3 ,
T1 ⊗ (T2 + T3 ) = T1 ⊗ T2 + T1 ⊗ T3 ,
(λT1 ) ⊗ T2 = T1 ⊗ (λT2 ) = λ(T1 ⊗ T2 ),

for all λ ∈ R;
• Associative:
(T1 ⊗ T2 ) ⊗ T3 = T1 ⊗ (T2 ⊗ T3 ),

• Not commutative:
T1 ⊗ T2 6= T2 ⊗ T1 in general,

for all appropriate tensors T1 , T2 , T3 .

Remark 4.14 (Index convention). Given a vector space V , we adopt the convention that subscripts are used
for the indices labelling elements of a basis of V and superscripts are used for the indices labelling elements
of a basis of V ∗ . Given bases {ei } of V and {θi } of V ∗ , we can then write any v ∈ V as v = v i ei and any
L ∈ V ∗ as Li θi .

16
Exercise 4.15 (Basis for Vsr ). If V is a vector space, {ei }ni=1 is a basis for V and {θi }ni=1 is its dual basis
for V ∗ , then
{ei1 ⊗ . . . ⊗ eir ⊗ θj1 ⊗ . . . ⊗ θjs } 1≤i1 ,...,ir ≤n ,
1≤j1 ,...,js ≤n

is a basis for Vsr . In particular, dimVsr =n r+s

.
Given any such bases for V and V ∗ , any (r, s) tensor T can uniquely written as

T = T i1 ...ir j1 ...js ei1 ⊗ . . . ⊗ eir ⊗ θj1 ⊗ . . . ⊗ θjs .

The constants T i1 ...ir j1 ...js are called the components of T with respect to the basis {ei1 ⊗ . . . ⊗ eir ⊗ θj1 ⊗
. . . ⊗ θjs }.
When defining certain operations on tensors, it is often convenient to first define them on reducible
tensors, and then extend to all tensors by linearity.
Definition 4.16 (Reducible tensor). If V is a vector space, a tensor T ∈ Vsr is called reducible if it can be
written as
T = v1 ⊗ . . . ⊗ vr ⊗ L1 ⊗ . . . ⊗ Ls , (8)
for some v1 , . . . , vr ∈ V , L1 , . . . , Ls ∈ V ∗ .
Example 4.17 (Examples of reducible tensors). Given a vector space V and v1 , v2 , v3 , v4 ∈ V , in the space
V02 the tensor v1 ⊗ v2 is reducible, v1 ⊗ v2 + v1 ⊗ v3 is reducible to v1 ⊗ (v2 + v3 ), but v1 ⊗ v2 + v3 ⊗ v4 is
not, in general, reducible.
Clearly any tensor can be written as a linear combination of reducible tensors.
Definition 4.18 (Tensor contraction). Given a vector space V , r, s ≥ 1, and 1 ≤ i ≤ r, 1 ≤ j ≤ s, for any
r−1
reducible tensor T ∈ Vsr of the form (8) define the element of Vs−1

ci j (T ) := Lj (vi )v1 ⊗ . . . ⊗ vi−1 ⊗ vi+1 ⊗ . . . ⊗ vr ⊗ L1 ⊗ . . . ⊗ Lj−1 ⊗ Lj+1 ⊗ . . . ⊗ Ls ,

r−1
and then extend ci j to act on the general T ∈ Vsr by linearity. The map ci j : Vsr → Vs−1 is called tensor
contraction.
For any vector space V define

V 0 := R, (V ∗ )0 := R, V00 := R ⊗ R = R.

Example 4.19 (Example of a tensor contraction). If V is a vector space and T ∈ V11 = V ⊗ V ∗ , and if {ei }
a basis for V with dual basis {θi } for V ∗ , then T = T i j ei ⊗ θj and

c1 1 (T ) = θj (ei )T i j = δij T i j = T i i ∈ R.

4.3 Vector bundles and tensor bundles

A vector bundle over a manifold M is, roughly speaking, a smooth assignment of isomorphic vector spaces
to each point p ∈ M.
Definition 4.20 (Vector bundles). Consider some k ≥ 0. A vector bundle of rank k is a triple (E, M, π)
where M is a manifold of dimension n, E is a manifold of dimension n + k and π : E → M is a smooth
surjection such that π −1 (p) is a vector space for all p ∈ M, with the following properties.
1. There exists an open cover {Vα }α∈A of M and a family of diffeomorphisms {ψα }α∈A ,

ψα : π −1 (Vα ) → Vα × Rk .

Here Vα × Rk is equipped with the product smooth structure.

17
 2. For any p ∈ M and any α ∈ A,
ψα π −1 (p) = {p} × Rk ,

and
ψα π−1 (p) : π −1 (p) → {p} × Rk ,


is a vector space isomorphism.

3. If α, β ∈ A are such that Vα ∩ Vβ 6= ∅, then the diffeomorphisms

ψα ◦ ψβ−1 : (Vα ∩ Vβ ) × Rk → (Vα ∩ Vβ ) × Rk ,

take the form

ψα ◦ ψβ−1 (p, a) = (p, Aαβ (p)a),
where Aαβ : Vα ∩ Vβ → GL(k, R) is smooth.
The manifold M is called the base space, the manifold E is called the total space, and the map π is called
the projection map. For each p ∈ M, the vector space Ep := π −1 (p) is called the fibre of E at p. The pairs
(Vα , ψα ) are called local trivialisations. The maps Aαβ are called transition maps.
We often say “let π : E → M be a vector bundle” to mean “let (E, M, π) be a vector bundle”.
Remark 4.21 (Smoothness of transition maps). In order to make sense of the smoothness of Aαβ : Vα ∩Vβ →
GL(k, R) in Definition 4.20, it is necessary to introduce a smooth manifold structure on the general linear
group GL(k, R).
Exercise 4.22 (Smooth structure on GL(n, R)). Given a smooth manifold M and an open subset Σ ⊂ M ,
show that Σ is itself a smooth manifold. Deduce that the space GL(n, R) of real valued n × n invertible
matrices can be given the structure of a smooth n2 dimensional manifold.
Remark 4.23 (Local trivialisations vs charts). The local trivialisations of a vector bundle are not to be
confused with the charts of E!
Example 4.24 (Examples of vector bundles).
1. Given a manifold M and k ≥ 0, the trivial bundle over M is the vector bundle (M × Rk , M, π), where
π : M × Rk → M is defined by π(p, a) = p for p ∈ M, a ∈ Rk . There is only one local trivialisation
(M, IdM×Rk ), and so the compatibility condition trivially holds.
2. Given a manifold M, the tangent bundle (T M, M, π), where π : T M → M is the natural projection, is
a vector bundle. Indeed, let {(Vα , φα )}α∈A be an atlas of M. Define local trivialisations {(Vα , ψα )}α∈A
by
 ∂  
ψα : π −1 (Vα ) → Vα × Rn , ψα p, ai i  = (p, a1 , . . . , an ).
∂x p
To check the compatibility condition, consider α, β ∈ A with Vα ∩ Vβ 6= ∅. Let {x̃i } denote the
coordinates of (Vα , φα ) and {xi } denote the coordinates of (Vβ , φβ ). Given
 ∂    ∂  
p, ai i  = p, ãi i  ∈ π −1 (Vα ∩ Vβ ),
∂x p ∂ x̃ p
it follows from Exercise 3.8 that
∂ x̃i 
ãi = aj  .
∂xj p
Hence the transition map
 ∂ x̃i  n
Aαβ (p) = ,

∂xj p i,j=1


is smooth.

18
Exercise 4.25 (The cotangent bundle is a vector bundle). Show that the cotangent bundle (T ∗ M, M, π) of
a manifold M, where π : T ∗ M → M is the natural projection, is a vector bundle.
Note that the local trivialisations of a vector bundle make it look locally like the trivial bundle of Example
4.24. Hence the nomenclature.
Given a vector bundle (E, M, π) over a manifold of M, a smooth assignment to each point of M of an
element of each fibre of E is called a section.
Definition 4.26 (Sections of vector bundles). A section of a vector bundle (E, M, π) is a smooth map
s : M → E such that π ◦ s = IdM , i.e. s(p) ∈ Ep for all p ∈ M.
The space of all sections of a vector bundle (E, M, π) is denoted Γ(E). Clearly Γ(E) is a vector space
with the obvious operations.

Example 4.27 (Vector fields are sections of T M). Vector fields are sections of the tangent bundle, and
Γ(T M) = X(M).
Definition 4.28 (Local and global frames). Suppose a vector bundle (E, M, π) of rank k admits k sections
which are everywhere linearly independent. Such a collection of sections is called a global frame of (E, M, π).
If these k sections are only locally defined, we call such a collection a local frame.
Remark 4.29 (Existence and non-existence of local and global frames). Global frames of vector bundles
do not always exist. For example, there is no nowhere vanishing vector field on S 2 . This statement is the
content of the Hairy Ball Theorem.
Local frames do always exist. If (E, M, π) is a vector bundle with local trivialisations {(Vα , ψα )}α∈A
then, for each α, {ψα−1 (p, ei )}ki=1 is a local frame, where ei is the i-th standard basis vector of Rk .
∂ n
If M is a manifold and (U, φ) is a chart with local coordinates {xi }, then { ∂x i }i=1 is a local frame for

T M.
Given a vector bundle, the operations of the following two definitions provide ways of generating many
more examples of vector bundles.
S
Definition 4.30 (Dual bundle). Given a vector bundle (E, M, π), recall that E = p∈M Ep . Define the
dual bundle of (E, M, π), denoted (E ∗ , M, π ∗ ), by
[
E∗ = {p} × (Ep )∗ , π ∗ : E ∗ → M, π ∗ (p, θ) = p.
p∈M

Exercise 4.31 (Dual bundle is a vector bundle). Show that (E ∗ , M, π ∗ ) is a vector bundle. Use the local
trivialisations (Vα , ψα ) for (E, M, π) to define local trivialisations for (E ∗ , M, π ∗ ). The transition maps for
(E ∗ , M, π ∗ ) are
(A∗ )αβ = (A−1 T
αβ ) .

Example 4.32 (Cotangent bundle). The dual bundle of the tangent bundle of a manifold is the cotangent
bundle.
Definition 4.33 (Tensor product bundles). Given vector bundles (E, M, π) and (Ẽ, M, π̃) of ranks k and
l respectively, define the tensor product bundle
[
E ⊗ Ẽ := {p} × (Ep ⊗ Ẽp ),
p∈M

with the projection map π ⊗ π̃ : E ⊗ Ẽ → M defined in the obvious way.

19
Exercise 4.34 (Tensor product bundle is a vector bundle). Show that the tensor product bundle of (E, M, π)
and (Ẽ, M, π̃) (of ranks k and l respectively) is a vector bundle of rank kl. If Aαβ and Ãαβ are transition
maps for E and Ẽ respectively, then

(AE⊗Ẽ )αβ (a ⊗ ã) := Aαβ a ⊗ Ãαβ ã ∈ Rk ⊗ Rl ∼

= Rkl ,

for a ∈ Rk , ã ∈ Rl , are transition maps for E ⊗ Ẽ.

Definition 4.35 (The (r, s) tensor bundles of a smooth manifold). Let M be a manifold. The (r, s) tensor
bundle of M is the vector bundle

Tsr M := T M ⊗ . . . ⊗ T M ⊗ T ∗ M ⊗ . . . ⊗ T ∗ M .
| {z } | {z }
r times s times

An (r, s) tensor field T is a section of Tsr M: T ∈ Γ(Tsr M).

Given local coordinates {xi }ni=1 of a chart (U, φ) of M, any (r, s) tensor field T can be written locally as

∂ ∂
T = T i1 ...ir j1 ...js ⊗ ... ⊗ ⊗ dxj1 ⊗ . . . ⊗ dxjs ,
∂xi1 ∂xir
for some smooth functions T i1 ...ir j1 ...js : U → R.
Example 4.36 (Tangent and cotangent bundles are (r, s) tensor bundles). For any smooth manifold M the
tangent and cotangent bundles are tensor bundles. Indeed, T01 M = T M and T10 M = T ∗ M.
Recall that, given a function F : M → N , there is a way of “pushing forward” vectors on M to vectors
on N (see Definition 3.9). There is similarly a way of “pulling back” one forms, and more generally (0, k)
tensor fields on N to one forms or (0, k) tensor fields on M.
Definition 4.37 (Pullback of a (0, k) tensor field). Let M and N be smooth manifolds, F : M → N a
smooth function, and ω ∈ Γ(Tk0 N ) a (0, k) tensor field on N . The pullback of ω with respect to F is a (0, k)
tensor field on M, denoted F ∗ ω ∈ Γ(Tk0 M), defined by

(F ∗ ω)p (v1 , . . . , vk ) = ωF (p) (F∗p v1 , . . . F∗p vk ),

for all v1 , . . . , vk ∈ Tp M and all p ∈ M.

If f ∈ C ∞ (N ), define
F ∗ f = f ◦ F.
Exercise 4.38 (Properties of pullback). Show that pullback has the following properties. Let M, N , Q be
manifolds, F : M → N and G : N → Q be smooth, and consider ω ∈ Γ(Tk0 N ), η ∈ Γ(Tl0 N ). Then
1. (G ◦ F )∗ = F ∗ ◦ G∗ .
2. F ∗ (ω ⊗ η) = F ∗ ω ⊗ F ∗ η. In particular, if f ∈ C ∞ (N ) = Γ(T00 N ), then F ∗ (f ω) = f ◦ F · F ∗ ω.
3. If f ∈ C ∞ (N ) then F ∗ (df ) = d(F ∗ f ) = d(f ◦ F ).

4. If p ∈ M and {y i } are the local coordinates of a chart for N containing F (p) ∈ N , then

F ∗ ωj1 ...jk dy j1 ⊗ . . . ⊗ dy jk = ωj1 ...jk ◦ F · d(y j1 ◦ F ) ⊗ . . . ⊗ d(y jk ◦ F ).

20
4.4 *Other important vector bundles
There are many other important examples of vector bundles, which do not form part of the examinable
material of this course. Most notable among these examples are the vector bundles of differential forms.
Before introducing the vector bundle of differential forms, consider first the following notions from linear
algebra.
Definition 4.39 (Alternating tensors). Let V be a vector space. A (0, k) tensor T ∈ Vk0 is called antisym-
metric or alternating if

T (v1 , . . . , vi , . . . , vj , . . . , vk ) = −T (v1 , . . . , vj , . . . , vi , . . . , vk ),

for all 1 ≤ i < j ≤ k. The space of antisymmetric T ∈ Vk0 is denoted Λk V ∗ ,

Λk V ∗ := {T ∈ Vk0 | T is antisymmetric}.

Two antisymmetric tensors can be multiplied to produce another antisymmetric tensor using an operation
known as wedge product.
Definition 4.40 (The alternating map). Let V be a vector space. Define the alternating map Alt : Vk0 →
Λk V ∗ by
1 X
Alt(T ) := sign(σ)σ(T ),
k!
σ∈Sk

where Sk is the symmetric group of order k (the group of permutations of the set {1, . . . , k}), where

σ(T )(v1 , . . . , vk ) := T (vσ(1) , . . . , vσ(k) ),

for all T ∈ Vk0 .

Definition 4.41 (The wedge product). Let V be a vector space. For ω ∈ Λk V ∗ and η ∈ Λl V ∗ , define the
wedge product or exterior product of ω and η to be the antisymmetric (0, k + l) tensor,

(k + l)!
ω ∧ η := Alt(ω ⊗ η).
k! l!
Explicitly,
1 X
ω ∧ η(v1 , . . . , vk+l ) = sign(σ)ω(vσ(1) , . . . , vσ(k) )η(vσ(k+1) , . . . , vσ(k+l) ).
k! l!
σ∈Sk+l

Example 4.42 (Wedge product of two covectors). If V is a vector space and ω, η ∈ Λ1 V ∗ = V ∗ , then

ω ∧ η = ω ⊗ η − η ⊗ ω.

Exercise 4.43 (Properties of wedge product). Let V be a vector space. The wedge product satisfies the
following properties.
1. The wedge product is bilinear and associative:

(aw1 + bw2 ) ∧ w3 = a(w1 ∧ w3 ) + b(w2 ∧ w3 ),

for all w1 , w2 ∈ Λk V ∗ , w3 ∈ Λl V ∗ and a, b ∈ R, and

(w1 ∧ w2 ) ∧ w3 = w1 ∧ (w2 ∧ w3 ),

for all w1 ∈ Λk1 V ∗ , w2 ∈ Λk2 V ∗ , w3 ∈ Λk3 V ∗ and a, b ∈ R.

21
 2.
ω ∧ η = (−1)kl η ∧ ω,
for all ω ∈ Λk V ∗ , η ∈ Λl V ∗ .
3. Given ω 1 , . . . , ω k ∈ V ∗ and v1 , . . . , vk ∈ V ,

(ω 1 ∧ . . . ∧ ω k )(v1 , . . . , vk ) = det ω i (vj ) .

4. If {ei }ni=1 is a basis of V and {θj }nj=1 is its dual basis, then {θi1 ∧ . . . ∧ θik }1≤i1 <...<ik ≤n is a basis of
Λk V ∗ . Any α ∈ Λk V ∗ can therefore be written as
X
α= α(ei1 , . . . , eik )θi1 ∧ . . . ∧ θik .
1≤i1 <...<ik ≤n

5. Covectors ω1 , . . . , ωk ∈ V ∗ are linearly independent if and only if

ω1 ∧ . . . ∧ ωk 6= 0.
n
6. dim(Λk V ∗ ) = . In particular dim(Λn V ∗ ) = 1, and dim(Λk V ∗ ) = 0 for all k ≥ n + 1.


k

Consider now a smooth manifold M.

Definition 4.44 (Differential forms). A (differential) k-form on a manifold M is a section of the vector
bundle Λk M := Λk T ∗ M. The space of k-forms is denoted Ωk (M),

Ωk (M) := Γ(Λk M).

In a local coordinate system {xi } for M, any ω ∈ Ωk (M) can locally be written as

ω = ωi1 ...ik dxi1 ∧ . . . ∧ dxik .

We adopt the convention that Ω0 (M) = C ∞ (M). Recall the map d : Ω0 (M) → Ω1 (M) defined in (7).
This map can be extended to a map d : Ωk (M) → Ωk+1 (M) for any k ≥ 0.
Definition 4.45 (The exterior derivative). The exterior derivative d : Ωk (M) → Ωk+1 (M) is defined as
follows. If (U, φ, xi ) is a local chart of M then any ω ∈ Ωk (M) can be written as

ω = ωi1 ...ik dxi1 ∧ . . . ∧ dxik .

The exterior derivative of ω is defined in these local coordinates by

dω = d(ωi1 ...ik ) ∧ dxi1 ∧ . . . ∧ dxik ,

where d(ωi1 ...ik ) is the usual differential of ωi1 ...ik ∈ C ∞ (M).

One can check that the definition of d is globally well defined and independent of the choice of local chart.
Exercise 4.46 (Properties of exterior derivative). The exterior derivative has the following properties.
1. If f ∈ C ∞ (M) = Ω0 (M) then df is the usual differential of f .
2. If ω ∈ Ωk (M) then d(dω) = 0.
3. d(aω + bη) = adω + bdη for all ω, η ∈ Ωk (M), a, b ∈ R.
4. d(ω ∧ η) = dω ∧ η + (−1)k ω ∧ dη for all ω ∈ Ωk (M), η ∈ Ωl (M).

R If M is an orientable manifold of dimension n and ω ∈ Ωn (M), then one can define the integral of ω,
M
ω. See, for example, [11].

22
5 Riemannian metrics
We are now in a position to define a Riemannian metric, and some immediate geometric notions such an
object gives rise to. First, recall the definition of an inner product.
Definition 5.1 (Inner product). An inner product on a vector space V is a function h·, ·i : V × V → R
which is
• Symmetric: hu, vi = hv, ui for all u, v ∈ V ;

• Bilinear: hau + bv, wi = ahu, wi + bhv, wi;

• Positive definite: hu, ui > 0 for all u 6= 0.
Example 5.2 (Euclidean dot product). The Euclidean dot product is an inner product on Rn .

A Riemannian metric on a manifold M associates to each point p ∈ M, in a smooth way, an inner

product on the tangent space Tp M.
Definition 5.3 (Riemannian metrics). A Riemannian metric on a manifold M is a (0, 2) tensor field
g ∈ Γ(T20 M) which is

• Symmetric: gp (Xp , Yp ) = gp (Yp , Xp ) for all Xp , Yp ∈ Tp M and all p ∈ M;

• Positive definite gp (Xp , Xp ) > 0 for all Xp ∈ Tp M with Xp 6= 0, and all p ∈ M.
Definition 5.4 (Riemannian manifolds). A pair (M, g), where M is a smooth manifold and g is a Rieman-
nian metric on M, is called a Riemannian manifold.

More generally, one can define a semi-Riemannian metric by relaxing the positive definiteness condition.
Definition 5.5 (Semi-Riemannian metrics). A semi-Riemannian metric or pseudo-Riemannian metric on
a manifold M is a (0, 2) tensor field g ∈ Γ(T20 M) which is
• Symmetric: gp (Xp , Yp ) = gp (Yp , Xp ) for all Xp , Yp ∈ Tp M and all p ∈ M;

• Non-degenerate: for all p ∈ M, if Xp ∈ Tp M is such that g(Xp , Yp ) = 0 for all Yp ∈ Tp M, then

Xp = 0.
Of particular note, for the role they play in Einstein’s general theory of relativity, are Lorentizian metrics.
A Lorentzian metric is a semi-Riemannian metric which everywhere has signature (−, +, . . . , +) (i.e. the

n+1
matrix gij (p) i,j=1 , defined below, has one negative eigenvalue and n positive eigenvalues for all p ∈ M).
In this course we will mostly consider Riemannian metrics, but many of the definitions and results apply
also to semi-Riemannian metrics. See the book [12] for more on semi-Riemannian geometry.

5.1 Expression in local coordinates

Given a Riemannian manifold (M, g) and a chart (U, φ, xi ), consider the functions
 ∂  ∂  
gij : U → R, gij (p) := gp ,  .

∂xi p ∂xj p


n
For each p ∈ U , gij (p) i,j=1 is a symmetric positive definite n × n matrix. The functions gij are called the
components of g with respect to {xi }.
Recall g ∈ Γ(T20 M) means that, using a local coordinate frame {dxi } for T ∗ M, we can write

g = gij dxi ⊗ dxj .

23
Since g is symmetric we often omit the ⊗ sign. To be more precise, define the symmetric product of one
forms ω, η ∈ Γ(T ∗ M) by
1

ωη := ω ⊗ η + η ⊗ ω .
2
Since (gij ) is symmetric, it follows that
1
g = gij dxi ⊗ dxj = gij dxi ⊗ dxj + gji dxi ⊗ dxj = gij dxi dxj .


2
More generally, if {ei } is a local frame for T M and {θi } is its dual frame for T ∗ M (i.e. θj (ei ) = δij for
all p) then
g = gij θi ⊗ θj ,
where now gij = g(ei , ej ).
Note that, given two vector fields X, Y ∈ X(M), the inner product g(X, Y ) is a function on M: g(X, Y ) ∈
C ∞ (M) where
g(X, Y )(p) := gp (Xp , Yp ).

5.2 Lengths and angles

A Riemannian metric on a manifold immediately gives rise to notions of lengths of vectors and angles between
vectors, and hence to notions of orthogonality and orthonormality of vectors.
Definition 5.6 (Lengths, angles, orthogonality and orthonormality of vectors). Let (M, g) be a Riemannian
manifold and p ∈ M.
• The length of v ∈ Tp M is defined to be
q
|v|g := gp (v, v).

If it is clear which metric g is meant, we often write |v| for |v|g .

• The angle between two non-zero vectors v, w ∈ Tp M is the unique θ ∈ [0, π] such that
gp (v, w)
cos θ = .
|v|g |w|g

• Two vectors v, w ∈ Tp M are called orthogonal if gp (v, w) = 0.

• Vectors e1 , . . . , ek ∈ Tp M are called orthonormal if
gp (ei , ej ) = δij ,
for all i, j = 1, . . . , k.
Using the above, one can define the length of a curve and angles between two intersecting curves.
Definition 5.7 (The length of a curve and the angle between two intersecting curves). Let (M, g) be a
Riemannian manifold. The length of a curve γ : (a, b) → M is defined to be
Z bq
L(γ) := gγ(s) (γ̇(s), γ̇(s))ds.
a

If γ and γ̃ are curves in M such that γ(c1 ) = γ̃(c2 ) = p ∈ M for some c1 , c2 ∈ R, and γ̇(c1 ) 6= 0,
˙ 2 ) 6= 0, define the angle between γ and γ̃ at p as the unique θ ∈ [0, π] such that
γ̃(c
˙ 2 ))
gp (γ̇(c1 ), γ̃(c
cos θ = ,
˙ 2 )|g
|γ̇(c1 )|g |γ̃(c
˙ 2 ).
i.e. the angle between γ̇(c1 ) and γ̃(c

24
 One has the following expressions in local coordinates for the inner product of two vectors and of the
length of a vector.
Exercise 5.8 (Inner product of two vectors in local coordinates). Let (U, φ, xi ) be a chart of (M, g) and
consider p ∈ U . Show that, if
∂  ∂ 
X = X i i  , Y = Y i i  ∈ Tp M,
∂x p ∂x p
then
gp (X, Y ) = gij (p)X i Y j ,
where
g = gij dxi dxj .
In particular, the length of X takes the form
q
|X|g = gij (p)X i X j .

5.3 *Volume
A Riemannian metric also give rise to a notion of volume. Since volume is defined using differential forms,
this section is non-examinable.
If (Mn , g) is an oriented Riemannian manifold, there exists a unique n-form dVg such that

dVg |p (E1 , . . . , En ) = 1,

whenever E1 , . . . , En is an oriented orthonormal basis for Tp M.

Exercise 5.9 (Volume form in local coordinates). Show that, in local coordinates {xi },
q
dVg = det(gij )dx1 ∧ . . . ∧ dxn .

The volume of M is defined to be Z

Vol(M ) := dVg .
M

More generally, if f ∈ C ∞ (M) is compactly supported, then f dVg is an n-form and so one can define the
integral of f to be Z
f dVg .
M

5.4 Examples of Riemannian metrics

The most basic example of a Riemannian metric is the Euclidean metric on Rn .

• The Euclidean metric on Rn in Cartesian coordinates takes the form

gEucl = δij dxi dxj = (dx1 )2 + . . . + (dxn )2 .

The following proposition gives rise to many other examples.

Proposition 5.10 (Pullback of a metric by an immersion). Let (M, g) be a Riemannian manifold, N a
smooth manifold, and let f : N → M be an immersion (a smooth map such that f∗p : Tp N → Tf (p) M is
injective for all p ∈ N ). The pullback f ∗ g ∈ Γ(T20 N ) defines a metric on N .

25
Proof. Recall that
(f ∗ g)p (u, v) = gf (p) (f∗p u, f∗p v),
for all u, v ∈ Tp N and all p ∈ N . Clearly (f ∗ g)p is symmetric. Since f∗p is injective, if f∗p u = 0 then it
follows that u = 0, and hence f ∗ g is positive definite.

• If N is a submanifold of a Riemannian manifold (M, g), the inclusion map ι : N ,→ M defines a metric
on N , ι∗ g, called the induced metric.

In particular, the Euclidean metric induces a metric on every submanifold of Rn .

• Consider the n sphere S n := {(x1 )2 + . . . + (xn+1 )2 = 1} ⊂ Rn+1 and the inclusion map ι : S n ,→ Rn+1 .
The induced Euclidean metric ι∗ gEucl on S n is called the round metric (or standard metric) on S n .
Proposition 5.11 (The round metric on S 2 in polar coordinates). In (θ, φ) polar coordinates for S 2 , the
round metric takes the form
gS 2 = dθ2 + sin2 θdφ2 .
Proof. The inclusion map ι : S 2 ,→ R3 in polar coordinates takes the form

ι(θ, φ) = (sin θ cos φ, sin θ sin φ, cos θ).

First one computes the push forwards ι∗(θ,φ) ∂θ and ι∗(θ,φ) ∂φ . Given f ∈ C ∞ (R3 ),



ι∗(θ,φ) (∂θ )f = ∂θ f ◦ ι(θ, φ) = ∂θ f (sin θ cos θ, sin θ sin φ, cos θ) = cos θ cos φ∂x f + cos θ sin φ∂y f − sin θ∂z f.

It follows that
ι∗(θ,φ) ∂θ = cos θ cos φ∂x + cos θ sin φ∂y − sin θ∂z ,
and similarly
ι∗(θ,φ) ∂φ = − sin θ sin φ∂x + sin θ cos φ∂y .
Now

gS 2 |(θ,φ) (∂θ , ∂θ ) = gEucl |ι(θ,φ) (ι∗(θ,φ) ∂θ , ι∗(θ,φ) ∂θ ) = cos2 θ cos2 φ + cos2 θ sin2 φ + sin2 θ = 1,

gS 2 |(θ,φ) (∂θ , ∂φ ) = − cos θ cos φ sin θ sin φ + cos θ sin φ sin θ cos φ = 0,
and
gS 2 |(θ,φ) (∂φ , ∂φ ) = sin2 θ sin2 φ + sin2 θ cos2 φ = sin2 θ,
and so it follows that
gS 2 = dθ2 + sin2 θdφ2 .

• The n ball B n = {(x1 , . . . , xn ) ∈ Rn | (x1 )2 + . . . + (xn )2 < 1} with the metric

 2
2
(dx1 )2 + . . . + (dxn )2 ,


g= 1 2 n 2


1 − (x ) + . . . + (x )

is called hyperbolic space.

• If (Mn1 1 , g1 ) and (Mn2 2 , g2 ) are Riemannian manifolds, recall that the product M1 × M2 admits the
smooth structure of an n1 + n2 dimensional smooth manifold (the product smooth structure). The
product manifold M1 × M2 admits a Riemannian metric g = g1 × g2 called the product metric


g1 × g2 |(p1 ,p2 ) (u1 , u2 ), (v1 , v2 ) = g1 |p1 (u1 , v1 ) + g2 |p2 (u2 , v2 ),

for (u1 , u2 ), (v1 , v2 ) ∈ T(p1 ,p2 ) (M1 × M2 ) ∼

= Tp1 M1 × Tp2 M2 , u1 , v1 ∈ Tp1 M1 , u2 , v2 ∈ Tp2 M2 .

26
 • Given a smooth function f : M1 → (0, ∞), define the warped product metric of (M1 , g1 ) and (M2 , g2 )
with respect to f on the product manifold M1 × M2 as


g1 ×f g2 |(p1 ,p2 ) (u1 , u2 ), (v1 , v2 ) = g1 |p1 (u1 , v1 ) + f (p1 )g2 |p2 (u2 , v2 ).

The warped product manifold is denoted (M1 , g1 ) ×f (M2 , g2 ).

Exercise 5.12 (Euclidean metric on R2 in polar coordinates). Show that, in polar coordinates, the Euclidean
metric on R2 r {0} takes the form
gEucl = dr2 + r2 dθ2 ,
and hence (R2 r {0}, gEucl ) is isometric (see Definition 5.14 below) to the warped product

(R+ , gEucl ) ×r2 (S 1 , gRound ).

• If (M, g) is a Riemannian manifold and λ : M → (0, ∞) is a smooth function, then (M, λg) is also a
Riemannian manifold.

Exercise 5.13 (Angles for conformal metrics). For such a smooth function λ : M → (0, ∞), show that the
g angle between two vectors or curves is equal to the λg angle.
The manifolds (M, g) and (M, λg) are said to be conformal. If λ is a constant then (M, g) and (M, λg)
are said to be homothetic.

5.5 Isometries
In Riemannian geometry, the notion of “sameness” is that of isometry. One can call Riemannian geometry
the study of properties which are invariant under isometry.
Definition 5.14 (Isometries). Two Riemannian manifolds (M, g) and (N , h) are called isometric if there
exists a diffeomorphism f : M → N such that f ∗ h = g. Such a diffeomorphism f is called an isometry.
An isometry f : (M, g) → (M, g) is called an isometry of (M, g). Think of such an f as describing a
symmetry of (M, g).
Definition 5.15 (Local isometries). Two Riemannian manifolds (M, g) and (N , h) are called locally iso-
metric if, for each p ∈ M, there exists open sets U ⊂ M and V ⊂ N with p ∈ U and an isometry f : U → V .
A Riemannian manifold (M, g) is called flat if it is locally isometric to (Rn , gEucl ).
Note that flat manifolds can have different topologies to Rn .
One can ask the following naive question: is every Riemannian manifold flat? It is not yet obvious that
the answer to this question is no. It will become clear after we discuss curvature that the answer is indeed
no. Much of our efforts for the coming lectures will be directed towards introducing a notion of curvature.
For now, one can take this question as a motivation for introducing curvature. The answer to this question
is yes, however, when n = 1.
Exercise 5.16 (One dimensional Riemannian manifolds). Show that every one dimensional Riemannian
manifold is flat.
Recall that an immersion is a map between manifolds f : M → N such that f∗p is injective for all p ∈ M.
When a manifold (M, g) can be immersed into a larger manifold (N , h) in such a way that the geometry of
(M, g) arises as the induced geometry of (N , h), we call such an immersion an isometric immersion.
Definition 5.17 (Isometric immersions). An immersion f : (M, g) → (N , h) is called isometric if f ∗ h = g.
Recall the Nash Embedding Theorem, Theorem 1.2, which guarantees that every Riemannian manifold
can be isometrically immersed (in fact, embedded) into RN for some N .

27
5.6 Conformal maps
More generally, one can consider conformal maps.
Definition 5.18 (Conformal maps). A diffeomorphism f : (M, g) → (N , h) between two Riemannian man-
ifolds is called a conformal map with conformal factor λ : M → (0, ∞) if

f ∗ h = λ2 g.

Two manifolds (M, g) and (N , h) are called conformal if there is a conformal diffeomorphism between them.
Exercise 5.19 (Conformal maps preserve angles). Show that conformal maps preserve angles between vectors
and curves.

Example 5.20 (Stereographic projection). Stereographic projection from the north pole N of S n , Φ : S n r
{N } → Rn , is a conformal map.
Definition 5.21 (Locally conformally flat manifolds). A Riemannian manifold (M, g) is called locally
conformally flat if, for all p ∈ M there exists an open set U ⊂ M with p ∈ U , an open set V ⊂ Rn and a
conformal diffeomorphism f : U → V .
Example 5.22 (The round sphere is locally conformally flat). The round sphere (S n , gRound ) is locally
conformally flat. Stereographic projection is a local conformal diffeomorphism to Rn .

5.7 Existence of Riemannian metrics

Every smooth manifold M admits a Riemannian metric. One easy way to establish this fact is to appeal to the
Whitney Embedding Theorem, Theorem 1.1, which guarantees the existence of an embedding F : M → RN ,
for some large N . The pullback of the Euclidean metric, F ∗ gEucl , is then a metric on M. Such a proof is
overkill, however. This section concerns a more direct proof, using a partition of unity.

Definition 5.23 (Partitions of unity). Let M be a smooth manifold.

• An open cover {Vα }α∈A of M is called locally finite if, for every p ∈ M, there exists a neighbourhood
W of p such that the set
{α ∈ A | W ∩ Vα 6= ∅},
is finite.
• The support of a function f : M → Ris defined to be

supp(f ) = {p ∈ M | f (p) 6= 0}.

• Given a locally finite open cover of M, {Vα }α∈A , a partition of unity subordinate to {Vα }α∈A is a
collection {fα }α∈A of smooth functions fα : Vα → [0, 1] such that supp(fα ) ⊂ Vα for all α ∈ A and
X
fα (p) = 1, (9)
α∈A

for all p ∈ M.
Note that the summation (9) is finite for all p ∈ M since the cover {Vα }α∈A is locally finite.

Proposition 5.24 (Existence of partitions of unity). Every smooth manifold admits a sub-atlas {(Vα , φα )}α∈A
such that {Vα }α∈A is locally finite, and a partition of unity {fα }α∈A subordinate to {Vα }α∈A .
Proof. The proof is non-examinable and not given here. See, for example, [11].

28
 The following proof follows a standard construction in Riemannian geometry of first constructing in each
chart using the local homeomorphism to an open subset of Rn and the Euclidean metric on Rn , and then
patching the construction in each chart together using a partition of unity.
Theorem 5.25 (Existence of Riemannian metrics). Every manifold M admits a Riemannian metric.
Proof. Let {Vα }α∈A be a locally finite open cover of M with charts φα : Vα → Wα ⊂ Rn , and let {fα }α∈A
be a subordinate partition of unity. For p ∈ M, define
X
gp := fα (p)φ∗α gEucl |p .
α∈A

Note that the summation involves only finitely many terms for each p ∈ M. Clearly g is symmetric. The
positive definiteness followsPfrom the positive definiteness of each φ∗α gEucl |p , together with the fact that
0 ≤ fα ≤ 1 for each α and α∈A fα (p) = 1.

5.8 Musical isomorphisms and the inner product of tensor fields

Let (M, g) be a Riemannian manifold. For each p ∈ M, the metric g defines a canonical isomorphism
between Tp M and Tp∗ M.
Definition 5.26 (The flat isomorphism). Define the map “flat” [ : Tp M → Tp∗ M by

X [ (Y ) := gp (X, Y ),

for all X, Y ∈ Tp M.
Exercise 5.27 (Flat is an isomorphism). Show that [ : Tp M → Tp∗ M is a vector space isomorphism.
Definition 5.28 (The sharp isomorphism). Define the map “sharp” ] : Tp∗ M → Tp M to be the inverse of
the isomorphism [: ] = [−1 . For ξ ∈ Tp∗ M, write ξ ] for ](ξ).

In local coordinates, if X = X i ei ∈ Tp M for some basis {ei } of Tp M then

X [ = Xi θi ,

where {θi } is the dual basis to {ei } of Tp∗ M and

Xi = gij X j ,

and gij = gp (ei , ej ). The operation of applying the map [ to a vector is often called “lowering an index”.
Recall that (gij )ni,j=1 is a positive definite symmetric matrix, and hence is invertible. The components of
the inverse of this matrix is denoted g ij , so that

g ij gjk = δki ,

for i, k = 1, . . . , n.
If ξ = ξi θi ∈ Tp∗ M, then
ξ ] = ξ i ei ,
where
ξ i = g ij ξj .
The operation of applying the map ] to a covector is often called “raising an index”.
The isomorphisms [, ] also extend to isomorphisms

[ : X(M) → Γ(T ∗ M), ] : Γ(T ∗ M) → X(M).

29
Definition 5.29 (Gradient of a function). Given a function f ∈ C ∞ (M), define the gradient of f , grad(f ) ∈
X(M), as
grad(f ) := (df )] .
The operation ] is used to extend g to an inner product on Tp∗ M. Given ω, η ∈ Tp∗ M, define
gp (ω, η) := gp (ω ] , η ] ).
In coordinates
gp (ω, η) = g ij ωi ηj .
Similarly one can defined the norm of a covector ω ∈ Tp∗ M,
q
|ω|g := |ω ] |g = gp (ω ] , ω ] ).

Exercise 5.30 (Length of a covector in local coordinates). Let (U, φ, xi ) be a chart of (M, g) and consider
p ∈ U . Show that, if
ω = ωi dxi p ∈ Tp∗ M,


then q
|ω|g = g ij (p)ωi ωj ,
In particular, for any function f : M → R,
p
|df |g = g ij ∂xi f ∂xj f .
The operations [ and ] are extended in the obvious way to act on higher order tensor fields. For example,
if T is a (1, 2) tensor field, T ∈ Γ(T21 M), then for each p ∈ M, Tp is a map
Tp : Tp∗ M × Tp M × Tp M → R.
Define
Tp[ : Tp M × Tp M × Tp M → R,
by
Tp[ (X, Y, Z) = Tp (X [ , Y, Z).
∂
In coordinates, if T = T i jk ∂x j k
i ⊗ dx ⊗ dx , then

T [ = Tijk dxi ⊗ dxj ⊗ dxk , Tijk = gil T l jk .

The operations [ and ] are defined similarly for higher order tensor fields, though one typically has to explain
which argument the operation is being applied to.
The metric g can moreover be extended to act on pairs of general (r, s) tensor fields as follows. First,
extend g to act on reducible (r, s) tensors by
gp (X1 ⊗. . .⊗Xr ⊗ξ 1 ⊗. . .⊗ξ s , X̃1 ⊗. . .⊗ X̃r ⊗ ξ˜1 ⊗. . .⊗ ξ˜s ) := gp (X1 , X̃1 ) . . . gp (Xr , X̃r )gp (ξ 1 , ξ˜1 ) . . . gp (ξ s , ξ˜s ),
for all X1 , X̃1 , . . . , Xr , X̃r ∈ Tp M, ξ 1 , ξ˜1 . . . , ξ s , ξ˜s ∈ Tp∗ M. The metric g is then extended to general (r, s)
tensors by linearity.
Similarly, one defines the length of an (r, s) tensor T as
p
|T |g = g(T, T ).
Exercise 5.31 (Length of a tensor in local coordinates). Suppose, in a local chart (U, φ, xi ) of (M, g),
T ∈ Γ(Trs M) takes the form
∂ ∂
T = T i1 ...ir j1 ...js i
⊗ ... ⊗ ⊗ dxj1 ⊗ . . . ⊗ dxjs .
∂x 1 ∂xir
Show that
|T |2g = gi1 k1 . . . gir kr g j1 l1 . . . g js ls T i1 ...ir j1 ...js T k1 ...kr l1 ...ls .

30
6 Flows of vector fields and the Lie derivative
Recall (see Section 3) that we have a well defined way of taking derivatives of smooth functions on a manifold.
We would like to have a notion of differentiation for vector fields and higher order tensor fields. There is
a notion of differentiation intrinsic to a smooth manifold, know as the Lie derivative, introduced in this
section. The Lie derivative, however, has some undesirable properties. A better notion of differentiation
requires an additional structure, known as a connection, and is discussed in Section 7.

6.1 Derivatives of vector fields in Rn

The existence of the canonical Cartesian coordinate system means that there is a canonical way of taking
derivatives of vector fields on Rn . It is worth taking a moment to recall this notion of differentiation and to
examine the obstructions to generalising it to abstract manifolds.
In Rn we can, using the slight abuse of Remark 4.3, view a vector field X as a smooth function X : Rn →
R . Given a point p ∈ Rn and a direction v ∈ Rn , the directional derivative of X at p in the direction v is
n

given by
X(p + tv) − X(p)
Dv X|p := lim . (10)
t→0 t
Consider now a smooth manifold M and a vector field X ∈ X(M). There are two immediate issues in
generalising the definition (10) of the directional derivative of X:

1. Since now p ∈ M and v ∈ Tp M, there is no definition of p + tv. One could replace v ∈ Tp M with a
vector field V ∈ X(M) and replace p + tv with a curve γ with γ(0) = p and γ̇(0) = v. The existence
of such a curve is discussed in Section 6.2.
2. Suppose one can find such a γ. Then Xγ(t) ∈ Tγ(t) M and Xp ∈ Tp M lie in two different vector
spaces. In order to subtract Xp from Xγ(t) one needs a way to identify these spaces. Different ways of
identifying these spaces give rise to different notions of differentiation. An intrinsic way of identifying
these spaces, which gives rise to the Lie derivative, is discussed in Section 6.3.

6.2 The flow of a vector field

Recall the local existence theorem for an ordinary differential equation in Rn .
Theorem 6.1 (Picard/Cauchy–Lipschitz). Given an open set U ⊂ Rn , a smooth vector field V : U → Rn
and a point x0 ∈ U , there exists a unique maximal interval (T− , T+ ) ⊂ R with T− < 0, T+ > 0 and a unique
solution x : (T− , T+ ) → Rn of the initial value problem

ẋ(t) = V (x(t)), x(0) = x0 .

The solution depends smoothly on x0 . Moreover, if T+ < ∞ then, for any compact set K ⊂ U , there exists
tK ∈ (0, T+ ) such that x(tK , T+ ) ∩ K = ∅.
This local existence theorem yields the following theorem on the existence of integral curves of a vector
field on a manifold.
Theorem 6.2 (Existence and uniqueness of integral curves of a vector field). Given a smooth manifold M,
a vector field V ∈ X(M) and a point p ∈ M, there exists a unique maximal interval (T− , T+ ) ⊂ R with
T− < 0, T+ > 0 and a unique curve γ : (T− , T+ ) → M such that

γ̇(t) = V |γ(t) , γ(0) = p,

for all t ∈ (T− , T+ ). Moreover, if T+ < ∞ then, for any compact set K ⊂ M, there exists tK ∈ (0, T+ ) such
that γ(tK , T+ ) ∩ K = ∅. In particular, if M is compact then T− = −∞, T+ = ∞.

31
 The curve γ of Theorem 6.2 is called the integral curve of V through p.
Definition 6.3 (Complete vector fields). Let M be a smooth manifold and V ∈ X(M) a vector field. If, for
every p ∈ M, T− = −∞, T+ = ∞, where T± are as in Theorem 6.2, then V is called complete.
Exercise 6.4 (Examples of complete and incomplete vector fields). Give an example of a manifold M and
a vector field on M which is not complete. Give an example of a non-compact manifold M and a nontrivial
complete vector field on M.
In fact, Theorem 6.1 tells us more.
Theorem 6.5 (Flow of a vector field). Given a smooth manifold M, a vector field V ∈ X(M) and a point
p ∈ M, there exists U ⊂ M open with p ∈ U , ε > 0 and a unique smooth map
Φ : (−ε, ε) × U → M,
such that
∂Φ
(s, p) = V |Φ(s,p) , Φ(0, p) = p,
∂s
for all s ∈ (−ε, ε), p ∈ U . For each s ∈ (−ε, ε), the map Φs := Φ(s, ·) : U → M is a diffeomorphism onto its
image. Moreover, if s, t, s + t ∈ (−ε, ε), then Φs ◦ Φt = Φs+t .
The map Φ of Theorem 6.5 is called the flow of the vector field V . The curve s 7→ Φ(s, p) is called the
integral curve, or flow line, of V through p.

6.3 The Lie derivative

Given a vector field V , the Lie derivative involves using the pullback of the flow Φt of V to identify the
tangent space TΦt (p) M with the tangent space Tp M.
Recall that, for a given function f : M → N , there is only a way of pushing forward vectors, and pulling
back (0, k) tensors (or, more generally, (0, k) tensor fields). If f is a diffeomorphism then these operations
can be extended using the inverse f −1 .
Definition 6.6 (Pullback of a vector by a diffeomorphism). If f : M → N is a diffeomorphism between
smooth manifolds and X ∈ Tf (p) N , define
−1
f ∗ X|p := f∗f (p) X ∈ Tp M.

Definition 6.7 (Lie derivative of a vector field). Let M be a smooth manifold and X, V ∈ X(M) be vector
fields on M. Define the Lie derivative of X along V to be the vector field LV X ∈ X(M) defined by
(Φ∗t X)p − Xp (Φ−t )∗Φt (p) X − Xp
LV X|p := lim = lim ,
t→0 t t→0 t
where Φ is the flow of V .
Exercise 6.8 (Lie derivative and Lie bracket). If X, V ∈ X(M) are vector fields, then the Lie derivative
takes the form
LV X = [V, X],
where [·, ·] is the Lie bracket (see Definition 3.15).
More generally, (r, s) tensor fields can be pulled back by diffeomorphisms.
Definition 6.9 (Pullback of a general tensor field by a diffeomorphism). If f : M → N is a diffeomorphism
between smooth manifolds and T ∈ Γ(Tsr N ) is an (r, s) tensor field, define the pullback of T by f , f ∗ T ∈
Γ(Tsr M), by
(f ∗ T )p (ξ1 , . . . , ξr , X1 , . . . , Xs ) := Tf (p) ((f −1 )∗ ξ1 )f (p) , . . . , ((f −1 )∗ ξr )f (p) , f∗p X1 , . . . , f∗p Xs ,

for all ξ1 , . . . , ξr ∈ Tp∗ M and X1 , . . . , Xs ∈ Tp M.

32
 The definition of Lie derivative can then be extended to general (r, s) tensor fields.
Definition 6.10 (Lie derivative of a tensor field). If M is a smooth manifold, V ∈ X(M) is a vector field,
and T ∈ Γ(Tsr M) is an (r, s) tensor field, the Lie derivative of T in the direction V is the (r, s) tensor field
LV T ∈ Γ(Tsr M) defined by
(Φ∗ T )p − Tp
LV T |p := lim t ,
t→0 t
where Φ is the flow of V .

Exercise 6.11 (Properties of Lie derivative). Let M be a smooth manifold, V ∈ X(M) be a vector field,
and T1 , T2 , T3 ∈ Γ(Tsr M) be (r, s) tensor fields. Show that the Lie derivative has the following properties.
1. LV f = V f for all f ∈ C ∞ (M) = Γ(T00 M).
2. LV (aT1 + bT2 ) = aLV T1 + bLV T2 for all a, b ∈ R.

3. LV (T1 ⊗ T2 ) = (LV T1 ) ⊗ T2 + T1 ⊗ (LV T2 ).

4. LV (f T ) = V f · T + f LV T for all f ∈ C ∞ (M).
5. LV (ci j T ) = ci j (LV T ) for all 1 ≤ j ≤ r, 1 ≤ i ≤ s.
6. In local coordinates,

(LV T )i1 ...ir j1 ...js = V (T i1 ...ir j1 ...js ) − ∂k V i1 T ki2 ...ir j1 ...js − . . . − ∂k V ir T i1 ...ir−1 k j1 ...js
+ ∂j1 V k T i1 ...ir kj2 ...js + . . . + ∂js V k T i1 ...ir j1 ...js−1 k .

Continuous symmetries of Riemannian manifolds are conveniently described using the Lie derivative.

Definition 6.12 (Killing vectors). On a Riemannian manifold (M, g), a vector field K ∈ X(M) is called a
Killing vector if
LK g = 0.
Killing vector fields generate continuous symmetries of (M, g).

Remark 6.13 (Properties of Lie derivative).

1. The Lie derivative only requires a smooth manifold structure. It does not, for example, require a
Riemannian metric.
2. The Lie derivative has the following undesirable property: if f ∈ C ∞ (M), then Lf V T 6= f LV T in
general (which can be easily seen using Exercise 3.17 and Exercise 6.8), i.e. LV T |p depends not just
on V at p ∈ M, but on V in a full neighbourhood of p.
A better notion of derivative requires a better way of “connecting” nearby tangent spaces.

7 Affine connections and the Levi-Civita connection

Such a way of “connecting tangent spaces” requires an additional structure, called a connection. In Section
7.1 affine connections are introduced. The link with “connecting tangent spaces” will become apparent in
Exercise 7.19. It will then be seen that a Riemannian metric on a manifold induces a canonical connection,
known as the Levi-Civita connection.

33
7.1 Connections and the covariant derivative
Of primary concern to us will be affine connections. First, a connection on a general vector bundle is
introduced.
Definition 7.1 (Connections and affine connections). A connection on a vector bundle π : E → M is a map

∇ : X(M) × Γ(E) → Γ(E),

where ∇(V, T ) is denoted ∇V T , such that:

• For each T ∈ Γ(E), the map V 7→ ∇V T is linear over C ∞ (M), i.e.

∇f V +hW T = f ∇V X + h∇W T,

for all V, W ∈ X(M), T ∈ Γ(E), f, h ∈ C ∞ (M).

• For each V ∈ X(M), the map T 7→ ∇V T is linear over R, i.e.

∇V (aS + bT ) = a∇V S + b∇V T,

for all V ∈ X(M), S, T ∈ Γ(E), a, b ∈ R.

• ∇ satisfies the following product rule

∇V (f T ) = V f · T + f ∇V T,

for all V ∈ X(M), T ∈ Γ(E), f ∈ C ∞ (M).

An affine connection, or linear connection on a smooth manifold M is a connection on the tangent bundle
T M.
Given an affine connection ∇ and vector fields V, X ∈ X(M), the vector field ∇V X is called the covariant
derivative of X in the direction V . We will see shortly that an affine connection on a manifold induces a
canonical connection on each (r, s) tensor bundle.
Example 7.2 (Euclidean connection). The canonical connection, or Euclidean connection, on Rn defined
by
X|p+tVp − X|p
(∇V X)p := lim ,
t→0 t
for all X, V ∈ X(Rn ) is the usual directional derivative.
Remark 7.3. 1. An affine connection ∇ is not a tensor field. For fixed V the map X 7→ ∇V X is, in
particular, not linear over C ∞ (M).
2. The notation D is also commonly used for connections.
Definition 7.4 (Christoffel symbols). Given an affine connection ∇ on a smooth manifold M and a local
frame {ei }ni=1 defined on U ⊂ M, define smooth functions Γkij : U → R, for i, j, k = 1, . . . , n by

∇ei ej = Γkij ek .

The functions {Γkij } are called the Christoffel symbols (or connection coefficients) of ∇ with respect to the
frame {ei }.
Exercise 7.5 (Covariant derivative in a local frame). Given a local frame {ei } for a smooth manifold M
and X, Y ∈ X(M), show that
∇X Y = X(Y k ) + X i Y j Γkij ek ,

where X = X i ei , Y = Y i ei .

34
 Note in particular that ∇X Y |p depends only on X at p, and on Y in a neighbourhood of p.
Example 7.6 (Euclidean connection in Cartesian coordinates). If {xi } are Cartesian coordinates on Rn ,
then the canonical connection satisfies
i
∂ j ∂Y ∂
∇X Y = X(Y i ) i
= X ,
∂x ∂x ∂xi
j

for all X, Y ∈ X(Rn ). The Christoffel symbols on ∇ in Cartesian coordinates therefore all identically vanish.
Exercise 7.7 (Euclidean connection in polar coordinates). Compute the Christoffel symbols of the canonical
connection on R2 in polar coordinates.
Proposition 7.8 (Existence of connections). Every smooth manifold admits an affine connection.
Proof. The proof can be established following the strategy of the proof of Theorem 5.25, by pulling back the
Euclidean connection in each chart and stitching each chart together using a partition of unity, and is left
as an exercise.
There are two alternative ways of proving Proposition 7.8:
1. Wait for the existence of the Levi-Civita connection (see Theorem 7.26) and use the existence of a
Riemannian metric on any manifold, Theorem 5.25.
2. The second alternative way is the “overkill proof” of appealing to the Whitney Embedding Theorem,
Theorem 1.1, and considering the induced Euclidean connection.
An affine connection ∇ on a smooth manifold M can be extended to act on tensor fields,
∇ : X(M) × Γ(Tsr M) → Γ(Tsr M).
Indeed, given an affine connection ∇ on M, there is a unique connection on Tsr M satisfying the following
three properties.
• For any function f ∈ C ∞ (M) = Γ(T00 M),
∇V f := V f,
for all V ∈ X(M).
• The product rule
∇V (T1 ⊗ T2 ) = (∇V T1 ) ⊗ T2 + T1 ⊗ (∇V T2 )
holds for all T1 ∈ Γ(Tsr11 M), T2 ∈ Γ(Tsr22 M), V ∈ X(M).
• ∇ commutes with contractions
∇V ci j T = ci j ∇V T ,

for all V ∈ X(M), T ∈ Γ(Tsr M), 1 ≤ i ≤ s, 1 ≤ j ≤ r.

Explicitly, for any one form ω ∈ Γ(T ∗ M) = Γ(T10 M), one computes, for any X, Y ∈ X(M),
X(ω(Y )) = ∇X c1 1 (Y ⊗ ω) = c1 1 ∇X (Y ⊗ ω) = c1 1 ∇X Y ⊗ ω + Y ⊗ ∇X ω ,

by the first, third, and second properties above respectively, and hence
(∇V ω)(X) = V (ω(X)) − ω(∇V X),
for all V, X ∈ X(M). In components,
∇V ω = V (ωi ) − V k ωj Γjki θi ,

where Γjki are the Christoffel symbols of ∇ with respect to a local frame {ei }, and {θi } is its dual frame.

35
Exercise 7.9 (Covariant derivative of a tensor field). Show, in a similar way, that, for T ∈ Γ(Tsr M),


(∇X T )(ξ1 , . . . , ξr , Y1 , . . . , Ys ) = X T (ξ1 , . . . , ξr , Y1 , . . . , Ys )
− T (∇X ξ1 , ξ2 , . . . , ξr , Y1 , . . . , Ys ) − . . . − T (ξ1 , . . . , ξr−1 , ∇X ξr , Y1 , . . . , Ys )
− T (ξ1 , . . . , ξr , ∇X Y1 , Y2 , . . . , Ys ) − . . . − T (ξ1 , . . . , ξr , Y1 , . . . , Ys−1 , ∇X Ys ),

for all Y1 , . . . , Ys ∈ X(M), ξ1 , . . . , ξr ∈ Γ(T ∗ M), and hence, in components,

∇X T = (∇X T )i1 ...ir j1 ...js ei1 ⊗ . . . ⊗ eir ⊗ θj1 ⊗ . . . ⊗ θjs ,

where

(∇X T )i1 ...ir j1 ...js = X T i1 ...ir j1 ...js + X k Γik1l T l i2 ...ir j1 ...js + . . . + X k Γikrl T i1 ...ir−1 l j1 ...js

− X k Γlk j1 T i1 ...ir l j2 ...js − . . . − X k Γlk jr T i1 ...ir j1 ...js−1 l .

Definition 7.10 (Total covariant derivative). If ∇ is an affine connection on M and T ∈ Γ(Tsr M), define
r
the total covariant derivative of T , ∇T ∈ Γ(Ts+1 M) by

(∇T )(ξ1 , . . . , ξr , Y1 , . . . , Ys , X) := (∇X T )(ξ1 , . . . , ξr , Y1 , . . . , Ys ).

Definition 7.11 (Parallel tensor fields). A tensor field T ∈ Γ(Tsr M) is called parallel if ∇T ≡ 0.
For k ≥ 2, one can inductively define the k-th total covariant derivative

∇k : Γ(Tsr M) → Γ(Ts+k
r
M),

by
∇k := ∇(∇k−1 ).
Remark 7.12 (Second covariant derivative of a vector field). If X, V1 , V2 ∈ X(M) then

(∇2 X)(V1 , V2 ) 6= ∇V2 ∇V1 X,

in general. In fact, one computes using Exercise 7.9,

(∇2 X)(V1 , V2 ) = ∇V2 ∇V1 X − ∇∇V2 V1 X.

This computation will be relevant when we discuss the Riemann curvature tensor in Section 9.

7.2 Connections along curves and parallel transport

We are typically interested not only in taking covariant derivative of vector fields on a manifold M, but
vector fields along curves γ (such as the tangent vector to γ). If a curve γ intersects itself then a given vector
field along γ may not extend to a vector field in M.
Definition 7.13 (Vector fields along curves). Let γ : (a, b) → M be a curve. A smooth map V : (a, b) → T M
is called a vector field along γ if V (t) ∈ Tγ(t) M (i.e. if π ◦ V (t) = γ(t)) for all t ∈ (a, b).
Denote
X(γ) := {vector fields along γ}.
A connection ∇ defines a “connection along γ”, which can be used to differentiate vector fields along γ.
If γ : (a, b) → M is a curve, t0 ∈ (a, b), {ei } is a local frame around γ(t0 ), and V ∈ X(γ), then we can
write
V (t) = V i (t)ei |γ(t) ,
for functions V i : (t0 − ε, t0 + ε) → R.

36
Definition 7.14 (Covariant derivative along a curve). Given a curve γ : (a, b) → M and a connection ∇
on M, the covariant derivative along γ is the map

Dt : X(γ) → X(γ),

defined locally by
Dt V |t0 = (V i )0 (t0 )ei |γ(t0 ) + V i (t0 )∇γ̇(t0 ) ei |γ(t0 ) .
The covariant derivative along γ has the following properties
1. The definition of Dt V is independent of the local frame {ei }.
2. Covariant derivative along γ is linear over R:

Dt (aV + bW ) = aDt V + bDt W,

for all a, b ∈ R, V, W ∈ X(γ).

3. Covariant derivative along γ satisfies the following product rule:

Dt (f V ) = f 0 · V + f · Dt V,

for all V ∈ X(γ), f ∈ C ∞ (a, b).

4. If V ∈ X(γ) and there exists Ṽ ∈ X(M) such that

Ṽ |γ(t) = V (t),

for all t ∈ (a, b), then Dt V = ∇γ̇ Ṽ .

In view of the latter property, given V ∈ X(γ) we often abuse notation and write ∇γ̇ V for Dt V , even
when V does not extend to a vector field on M.
Definition 7.15 (Parallel vector fields along a curve). Given a curve γ, a vector field along γ, V ∈ X(γ),
is called parallel along γ if Dt V ≡ 0.
In Section 8.1, a geodesic will be defined to be a curve γ whose tangent vector γ̇ is parallel along γ.
Note that, given p ∈ Rn , a vector v ∈ Tp Rn ∼ = Rn can be uniquely extended to a parallel vector field
n
on R by taking it to be constant (one easily checks that such a vector field is parallel with respect to the
Euclidean connection). On a general manifold M there is no such procedure (indeed the notion of a vector
field being “constant” is special to Rn , relying on the canonical identification of each tangent space to Rn
with Rn itself). Remarkably, there is still a way of extending a vector at a point to a parallel vector field
along a curve.

Proposition 7.16 (Existence and uniqueness of parallel transport). Given an affine connection ∇ on a
manifold M, a curve γ : (a, b) → M, t0 ∈ (a, b) and v0 ∈ Tγ(t0 ) M, there exists a unique parallel vector field
V (t) along γ such that V (t0 ) = v0 .
Proof. The main content of the proof is in the Picard/Cauchy–Lipschitz Theorem, Theorem 6.1. Indeed, let
{ei } be a local frame around γ(t0 ), so that

γ̇(t) = γ̇ i (t)ei |γ(t) , V (t) = V i (t)ei |γ(t) ,

for appropriate functions γ̇ i and V i . Now V satisfies

Dt V ≡ 0, V (t0 ) = v0 ,

37
if and only if
 dV i 
(t) + γ̇ j (t)V k (t)Γijk (γ(t)) ei |γ(t) = 0, V i (t0 )ei |γ(t0 ) = v0i ei |γ(t0 ) ,
dt
if and only if
dV i
(t) + γ̇ j (t)V k (t)Γijk (γ(t)) = 0, V i (t0 ) = v0i , (11)
dt
for i = 1, . . . , n. Note that (11) is an initial value problem for a linear system of ordinary differential
equations. Theorem 6.1 guarantees there exists a unique solution. Since the system is linear, the solution
exists in the entire domain of the local frame {ei }. Indeed, the Grönwall inequality a priori implies that, for
any T ≥ t0 such that γ(t) lies in the domain of the local frame for all t0 ≤ t ≤ T ,
n
X n
X n
X n
X 
sup |V i (t)| ≤ |v0i | exp sup |γ̇ i (t)| sup |Γijk (γ(t))| ,
i=1 t0 ≤t≤T i=1 i=1 t0 ≤t≤T i,j,k=1
t0 ≤t≤T

and so there exists a compact set which V i (t) cannot leave for all t0 ≤ t ≤ T , i = 1, . . . , n. If {ei } extends
to a frame around γ(t) for all t ∈ (a, b) then we are done.
Otherwise, let α denote the supremum over all times s ∈ [t0 , b] such that a unique parallel field V (t)
exists for all t ∈ [t0 , s]. Clearly α > t0 by the above. Assume α < b. Take ε small and a local frame
on γ(α − ε, α + ε). By the above there exists a unique parallel vector field Ṽ on (α − ε, α + ε) such that
Ṽ (α − ε/2) = V (α − ε/2). By uniqueness V and Ṽ agree on their common domain. Hence Ṽ is a unique
extension of V , which contradicts α < b.

Definition 7.17 (Parallel transport). If γ : (a, b) → M is a curve in M and t0 , t ∈ (a, b), the map

Pt0 ,t : Tγ(t0 ) M → Tγ(t) M,

defined by Pt0 ,t (v) := V (t), where V (t) is the unique parallel vector field along γ with V (t0 ) = v (see
Proposition 7.16), is called parallel transport along γ.

Remark 7.18 (Parallel transport is an isomorphism between (r, s) tensor spaces). Parallel transport Pt0 ,t
is a vector space isomorphism between Tγ(t0 ) M and Tγ(t) M. It can moreover be extended, in the obvious
way, to a vector space isomorphism

Pt0 ,t : Tsr M|γ(t0 ) → Tsr M|γ(t) ,

for all r, s.
The following exercise make the link between an affine connection and a way of “connecting nearby
tangent spaces”.
Exercise 7.19 (“Connecting nearby tangent spaces”). If ∇ is an affine connection on M and V, X ∈ X(M),
show that
−1
(P0,t X − X)|p
∇V X|p = lim ,
t→0 t
where P0,t denotes parallel transport along the integral curve γ of V with γ(0) = p.
Similarly for ∇V T with T ∈ Γ(Tsr M).

38
7.3 The Levi-Civita connection
A Riemannian metric g on a manifold M chooses one particular connection which interacts with g in a
particularly nice way.
Definition 7.20 (Compatible connections). An affine connection ∇ on a Riemannian manifold (M, g) is
called compatible with g if g is parallel with respect to ∇, i.e. if ∇g ≡ 0.
Exercise 7.21 (Compatibility of a connection with a metric). Show that ∇ is compatible with g if and only
if
X(g(Y, Z)) = g(∇X Y, Z) + g(Y, ∇X Z),
for all X, Y, Z ∈ X(M).
Note that if γ : (a, b) → M is a curve and X and Y are two parallel vector fields along γ, it follows from
Exercise 7.21 that the fact that ∇ is compatible with g implies that
d

gγ(t) (X(t), Y (t)) = 0,
dt
i.e. if ∇ is compatible with g then lengths of vectors and angles between vectors are preserved under parallel
transport. It is still possible, however, for parallel vectors along γ to “twist around γ”.

Definition 7.22 (Torsion tensor field). Given a connection ∇ on M, define the torsion tensor to be the
map
τ : X(M) × X(M) → X(M),
defined by
τ (X, Y ) = ∇X Y − ∇Y X − [X, Y ].
Exercise 7.23 (The torsion tensor field is a tensor field). Show that, if ∇ is a connection, then its torsion
tensor τ is a (1, 2) tensor field, τ ∈ Γ(T21 M).

Definition 7.24 (Torsion free connections). An affine connection ∇ on a manifold M is called symmetric
or torsion free if its torsion tensor τ identically vanishes, i.e. if

∇X Y − ∇Y X = [X, Y ],

for all X, Y ∈ X(M).

The name “symmetric” is explained by the following exercise.

Exercise 7.25 (Symmetric Christoffel symbols). Let {xi } be a local coordinate system of a smooth manifold
M. Show that a connection ∇ is symmetric if and only if its Christoffel symbols with respect to the coordinate
∂
frame { ∂x i } satisfy

Γkij = Γkji ,
for all i, j, k = 1, . . . , n.
The following theorem is often called the fundamental theorem of Riemannian geometry.

Theorem 7.26 (Existence and uniqueness of a compatible torsion free connection). Given a Riemannian
manifold (M, g), there exists a unique affine connection ∇ on M which is torsion free and compatible with
g. This affine connection ∇ moreover satisfies




2g(∇X Y, Z) = X g(Y, Z) + Y g(X, Z) − Z g(X, Y ) − g(X, [Y, Z]) − g(Y, [X, Z]) + g(Z, [X, Y ]). (12)

39
Proof. The proof proceeds by showing that any such connection must satisfy (12), and then checking that
(12) indeed defines a torsion free compatible connection. Indeed, suppose ∇ is such a connection. Then,
since ∇ is compatible with g,

g(∇X Y, Z) = X g(Y, Z) − g(Y, ∇X Z),
(see Exercise 7.21) and, since ∇ is torsion free,


g(∇X Y, Z) = X g(Y, Z) − g(Y, [X, Z]) − g(Y, ∇Z X). (13)

Similarly, interchanging the roles of X, Y and Z,

g(∇Y Z, X) = Y g(X, Z) + g(Z, [X, Y ]) − g(Z, ∇X Y ), (14)

and

g(∇Z X, Y ) = Z g(X, Y ) − g(X, [Y, Z]) − g(X, ∇Y Z). (15)
Adding (13) to (14) and subtracting (15) then yields (12). The uniqueness of such a connection is then
evident. For existence, it is left as an exercise to check that (12) indeed defines a torsion free compatible
connection. E.g., to check that (12) defines a connection, let {ei } be a local orthonormal frame, set
n
X
∇X Y = g(∇X Y, ek )ek ,
k=1

and replace g(∇X Y, ek ) using (12). Check that this ∇ satisfies the properties of Definition 7.1, and then
check that ∇ is torsion free and compatible with g.
Definition 7.27 (The Levi-Civita connection of a Riemannian manifold). Given a Riemannian manifold
(M, g), the connection ∇ of Theorem 7.26 is called the Levi-Civita connection of (M, g).
In a given chart (U, φ) of a Riemannian manifold (M, g) with local coordinates {xi }, (12) gives an
∂
expression for the Christoffel symbols Γijk of ∇ with respect to the coordinate frame { ∂x i } in terms of the
∂ ∂ ∂
components of g. Indeed, setting X = ∂xi , Y = ∂xj and Z = ∂xk , (12) gives

∂gjk ∂gik ∂gij

2gkl Γlij = + − ,
∂xi ∂xj ∂xk
from which it follows that
g kl  ∂gjl ∂gil ∂gij 
Γkij = + − . (16)
2 ∂xi ∂xj ∂xl
The Levi-Civita connection is a geometric object.
Exercise 7.28 (The Levi-Civita connection is a geometric object). Suppose ϕ : (M, g) → (M̃, g̃) is a local
isometry. Show that ϕ∗ (∇X Y ) = ∇ ˜ ϕ X ϕ∗ Y for all X, Y ∈ X(M), where ∇ and ∇˜ are the Levi-Civita
∗

connections of (M, g) and (M̃, g̃) respectively.

7.4 Divergence
The Levi-Civita connection on a Riemannian manifold (M, g) defines the divergence operator.
If V is a vector space, recall the contraction map c1 1 : V ⊗ V ∗ → R. This map is also called trace, and is
denoted tr := c1 1 .
Definition 7.29 (Divergence). Let (M, g) be a Riemannian manifold with Levi-Civita connection ∇. Define
the divergence of X ∈ X(M) by

divX := tr ∇X .
If ξ ∈ Γ(T ∗ M) is a one form, define
divξ := divξ ] .

40
 ∂
In local coordinates, if X = X i ∂x i , then

divX = ∇i X i .
∂
(Recall that ∇i X j := ∇X(dxj , ∂x j i
i ) 6= ∂xi (X ) in general.) If ξ = ξi dx , then

divξ = g ij ∇i ξj .

Exercise 7.30 (Divergence in local coordinates). In a local coordinate system {xi } for (M, g), show that,
for any vector field X ∈ X(M),
1 p
divX = √ ∂xi ( det gX i ),
det g
where

det g = det gij .

Unless specified otherwise, we adopt the (not completely standard) convention that the divergence of
T ∈ Γ(T0r M), for some r ≥ 1, is defined to be divT ∈ Γ(T0r−1 M) by taking the divergence with respect to
the first argument,

(divT )(ξ1 , . . . , ξr−1 ) := tr ∇T (·, ξ1 , . . . , ξr−1 ) .
In components, if T = T i1 ...ir ∂x∂i1 ⊗ . . . ⊗ ∂
∂xir then divT = (divT )i1 ...ir−1 ∂x∂i1 ⊗ . . . ⊗ ∂
∂xir−1
, where

(divT )i1 ...ir−1 = ∇j T ji1 ...ir−1 .

Similarly, unless specified otherwise, if T ∈ Γ(Ts0 M) for some s ≥ 1, define divT ∈ Γ(Ts−1 0
M) by raising the
index of the first argument and taking the divergence,

divT (X1 , . . . , Xs−1 ) := tr ∇T (·, X1 , . . . , Xs−1 )] .

In components, if T = Ti1 ...ir dxi1 ⊗ . . . ⊗ dxir then divT = (divT )i1 ...ir−1 dxi1 ⊗ . . . ⊗ dxir−1 , where

(divT )i1 ...ir−1 = g jk ∇j Tki1 ...ir−1 .

To take the divergence of a general T ∈ Γ(Tsr M), one has to specify which argument the divergence is being
taken with respect to.

Definition 7.31 (Laplace–Beltrami operator). Given a Riemannian manifold (M, g), define the Laplace–
Beltrami operator ∆g : C ∞ (M) → C ∞ (M) (or more generally ∆g : Γ(Tsr M) → Γ(Tsr M)) by

∆g f := div∇f,

for all f ∈ C ∞ (M).

One has the following expression for the Laplace–Beltrami operator.

Exercise 7.32 (Laplace–Beltrami operator in local coordinates). In a local coordinate system {xi } for
(M, g), show that the Laplace–Beltrami operator takes the form
1 p
∂xi g ij det g∂xj f ,


∆g f = √
det g
where

det g = det gij .

41
8 Geodesics and Riemannian distance
Each Riemmanian manifold has a notion of geodesic, or “curve which is as straight as possible”. Moreover,
there is a notion of Riemannian distance, with the distance between two points defined to be the infimum
of the length over all curves which join these two points. This notion of distance gives each Riemannian
manifold a metric space structure. When this metric space is complete, this infimum is always attained by
a geodesic.
From now on it will be assumed that (M, g) is a Riemmanian manifold with Levi-Civita connection ∇.

8.1 Geodesics
A geodesic in (M, g) should be viewed as a curve which is “as straight as possible” and thus is a generalisation
of a straight line in Rn . Note that straight lines in Rn have the property that they minimise the distance
between any two points, and moreover that, when traversed at constant speed, they are the unique curves
γ with “zero acceleration” in the sense that γ̈ ≡ 0. One could define geodesics in (M, g) to be curves which
locally minimise length, but it turns out to be much simpler to take the latter property of straight lines in
Rn as a definition and deduce this local length minimising property.
Note that, if γ : (a, b) → M, then γ̇ ∈ X(γ).
Definition 8.1 (Geodesics). A curve γ : (a, b) → M is called a geodesic if Dt γ̇(t) = 0 for all t ∈ (a, b), i.e.
if γ̇ is parallel along γ.
Recall the definition 7.14 of covariant derivative along γ and hence, given a local coordinate system {xi },
γ is a geodesic if and only if
  ∂ 
γ̈ i (t) + γ̇ j (t)γ̇ k (t)Γijk (γ(t)) = 0,

∂xi γ(t)


if and only if
γ̈ i (t) + γ̇ j (t)γ̇ k (t)Γijk (γ(t)) = 0, (17)
for i = 1, . . . , n. The equations (17) are called the geodesic equations.
The following theorem guarantees that, for a given initial position and velocity, there locally exists a
unique geodesic.
Theorem 8.2 (Local existence and uniqueness of geodesics with given initial point and velocity). Given
p ∈ M and v ∈ Tp M, there exists a unique maximal interval (T− , T+ ) ⊂ R, with T− < 0 and T+ > 0, and a
unique geodesic γ : (T− , T+ ) → M such that γ(0) = p and γ̇(0) = v.
Proof. Let (U, φ) be a chart with p ∈ U , and let {xi } be the associated local coordinates. The vector v can
∂
then be written v = v i ∂x i |p . By Theorem 6.1
4
there exists ε > 0 and γ 1 , . . . , γ n : (−ε, ε) → M satisfying
the system of nonlinear ordinary differential equations (17) together with the initial conditions

γ i (0) = φ(p)i , γ̇ i (0) = v i .

The maximality part of the statement follows from the maximality part of Theorem 6.1, being careful that
γ may leave the domain of the chart (U, φ) (as in the proof of Proposition 7.16).
Alternative proof of Theorem 8.2 (which is really just a more geometric take on the previous proof ). Recall that
a chart (U, φ), with local coordinates x1 , . . . , xn , for M defines a chart (π −1 (U ), Φ) for T M, where Φ : π −1 (U ) →
R2n is defined by
 ∂  
Φ p, v i i  = (x1 (p), . . . , xn (p), v 1 , . . . , v n ).
∂x p
4 To bring the system (17) to the correct form to directly apply Theorem 6.1, one has to perform the well known trick of

converting a second order ordinary differential equation on Rn to a first order ordinary differential equation on R2n . See also
the “alternative proof” below.

42
The associated local coordinates {xi } for M are thus extended to a local coordinate system {xi , v i } for T M.
The second order geodesic equations (17) can be rewritten in first order form
γ̇ i (t) = v i (t), v̇ i (t) = −v j (t)v k (t)Γijk (γ(t)).
Solutions of this system are exactly the integral curves of the vector field
G = v i ∂xi − v j v k Γijk ∂vi .
Note that G ∈ X(T M) = Γ(T T M) is a vector field not on M but on the tangent bundle T M. The existence
and uniqueness then follows from Theorem 6.2 (with T M in place of M).
Definition 8.3 (Geodesic spray and geodesic flow). Given a Riemannian manifold (M, g), the vector field
G ∈ X(T M) = Γ(T T M) defined locally by
G = v i ∂xi − v j v k Γijk ∂vi ,
is called the geodesic spray of M. The flow of G is called the geodesic flow of M.
The maximal geodesic of Theorem 8.2 is often denoted γp,v or sometimes just γv .
Example 8.4 (Examples of geodesics).
1. Consider (Rn , gEucl ). In Cartesian coordinates gij = δij and so Γijk ≡ 0 for all i, j, k = 1, . . . , n. The
geodesic equations then take the form
γ̈ i (t) = 0,
for i = 1, . . . , n. All solutions take the form,
γ(t) = a + tb,
for some a, b ∈ Rn , i.e. all geodesics in Rn are straight lines.
2. Consider now (S 2 , gRound ). Recall that, in spherical polar coordinates,
gRound = dθ2 + sin2 θ dφ2 .
Using the expression (16) one computes
cos θ
Γθφφ = − sin θ cos θ, Γφθφ = ,
sin θ
and
Γθθθ = Γθθφ = Γφφφ = Γφθθ = 0.
The geodesic equations for γ(t) = (θ(t), φ(t)) then take the form

2 cos θ(t)
θ̈(t) − φ̇(t) sin θ(t) cos θ(t) = 0, φ̈(t) + 2θ̇(t)φ̇(t) = 0.
sin θ(t)
One sees, for example, that θ(t) = t, φ(t) = φ0 is a solution for any φ0 ∈ (0, 2π). The image of
(θ(t), φ(t)) is a “great circle” i.e. the intersection of S 2 with a plane through the origin in R3 . In fact,
the image of any geodesic in (S 2 , gRound ) is a great circle.
Definition 8.5 (Geodesic completeness). A Riemannian manifold (M, g) is called geodesically complete if,
for all (p, v) ∈ T M, the maximal geodesic γp,v is defined for all t ∈ R, i.e. if for all (p, v) ∈ T M, the T± of
Theorem 8.2 satisfy T± = ±∞.
Example 8.6 (Examples of geodesic completeness). The Riemannian manifold (Rn , gEucl ) is geodesically
complete. The Riemannian manifold (Rn r {0}, gEucl ) is not.
Exercise 8.7 (Compact manifolds are geodesically complete). Show that every compact Riemannian man-
ifold is geodesically complete.

43
8.2 The exponential map and normal coordinates
The exponential map on the tangent bundle of a Riemannian manifold is defined by following the geodesic
corresponding to a given element of T M for time 1.
Definition 8.8 (The exponential map). Let (M, g) be a Riemannian manifold and, given p ∈ M, let
Ep ⊂ Tp M be the set of vectors v ∈ Tp M such that γp,v is defined on an interval containing [0, 1]. Define
the exponential map at p, expp : Ep → M, by
expp (v) := γp,v (1).
S
Set E = p∈M {p} × Ep ⊂ T M and define the exponential map, exp : E → M, by
exp(p, v) = expp (v) = γp,v (1).
Note that exp(p, v) = π ◦ ΦG (p, v), where π : T M → M is the natural projection and ΦG is the geodesic
flow (see Definition 8.3). In particular, exp is smooth.
Lemma 8.9 (Rescaling geodesic directions). Given (p, v) ∈ T M and λ ∈ R, γp,λv (t) = γp,v (λt) whenever
either side is defined.
Proof. Suppose λ is such that γp,v (λt) is defined. Set γ̃(t) = γp,v (λt). Clearly γ̃(0) = p and, in local
coordinates,
d 
γ̃ 0 (0)i =  γp,vi 0
(λt) = λγp,v (0)i = λv i ,
dt t=0
so that γ̃ 0 (0) = λv. Since γ̃ is also a geodesic (check that it solves the geodesic equation in local coordinates)
it follows by uniqueness that γ̃ = γp,λv , i.e.
γp,λv (t) = γp,v (λt).

In particular, Lemma 8.9 implies that

exp(p, λv) = γp,v (λ),
whenever either side is defined and, if exp(p, v) is defined then so is exp(p, λv) for all 0 ≤ λ ≤ 1. Hence Ep
is star shaped for all p ∈ M.
Lemma 8.10 (The exponential map is a local diffeomorphism around the origin). For each p ∈ M there
exists an open neighbourhood U ⊂ Tp M with 0 ∈ U and an open neighbourhood V ⊂ M with p ∈ V such
that
expp : U → V,
is a diffeomorphism.
Proof. The proof follows from the Inverse Function Theorem after establishing that the differential of the
exponential map at the origin, (expp )∗0 is invertible. Since Tp M is a vector space there is a canonical
identification between T0 Tp M and Tp M and so, abusing notation slightly, we write,
(expp )∗0 : T0 Tp M = Tp M → Tp M.
We will show that (expp )∗0 is the identity map.
Recall that, if F : M → N is a smooth map between manifolds, p ∈ M and β : (−ε, ε) → M satisfies
β(0) = p, β 0 (0) = v ∈ Tp M, then F∗p v = (F ◦ β)0 (0).
Given v ∈ Tp M, it follows that
d expp (tv)  dγp,v (t) 
(expp )∗0 v = = = v,
dt dt
 
t=0 t=0

by Lemma 8.9, i.e. (expp )∗0 = IdTp M .

44
Definition 8.11 (Normal neighbourhoods). Such a neighbourhood V of p, as in Lemma 8.10, is called a
normal neighbourhood of p.
Lemma 8.10 in particular implies that, for all p ∈ M, there exists ε > 0 such that expp : B(0, ε) ⊂
Tp M → M is a diffeomorphism onto its image. The image expp (B(0, ε)) ⊂ M is called a geodesic ball.
For each point p ∈ M, the exponential map defines a particularly nice system of local coordinates on a
normal neighbourhood of p, known as normal coordinates.
Definition 8.12 (Normal coordinates). Consider p ∈ M and an orthonormal basis {ei } of Tp M, define the
isomorphism
E : Rn → Tp M, E(x1 , . . . , xn ) = xi ei .
If V is a normal neighbourhood of p, define a chart

φ : V → W ⊂ Rn , φ := E −1 ◦ (expp )−1 .

The local coordinates {xi } associated to this chart (V, φ) are called normal coordinates centred at p.
Normal coordinates centred at p have the following nice properties.
Exercise 8.13 (Properties of normal coordinates). Given p ∈ M, show the following properties of normal
coordinates centred at p:
1. The coordinates of p are (0, . . . , 0).
2. Given v = v i ei ∈ Tp M, the associated geodesic satisfies γp,v (t) = (tv 1 , . . . , tv n ).
3. At the point p the components of the metric take the form

gij (p) = δij .

4. At the point p,
∂gij
Γkij (p) = 0, (p) = 0,
∂xk
for all i, j, k = 1, . . . , n.

8.3 Riemannian distance

A Riemannian metric g on M defines a metric (in the metric spaces sense) d : M × M → [0, ∞) called
Riemannian distance.
Definition 8.14 (Smooth curves on closed intervals). A map from a closed interval c : [a, b] → M is called a
smooth curve if c|(a,b) : (a, b) → M is a smooth curve in the usual sense, c is continuous, and all derivatives
of c with respect to any chart containing c(a) extends continuously to c(a) and all derivatives of c with respect
to any chart containing c(b) extends continuously to c(b).
Riemannian distance is most conveniently defined not using smooth curves, but rather piecewise smooth
curves.
Definition 8.15 (Piecewise smooth curves). A continuous map c : [a, b] → M is called a piecewise smooth
curve if there exists a finite subdivision a = t0 < t1 < . . . < tk = b of [a, b] such that c|[ti ,ti+1 ] is a smooth
curve for all i = 0, 1, . . . , k − 1.
If c : [a, b] → M is such a piecewise smooth curve, define the length of c by
k−1
X k−1
X Z ti+1
|c0 (t)|g dt.


L(c) := L c|(ti ,ti+1 ) =
i=0 i=1 ti

Such a curve c is said to join c(a) and c(b).

45
 Given p, q ∈ M, define the set

Cp,q = {c : [a, b] → M piecewise smooth | c(a) = p, c(b) = q}.

Recall that we always assume M is connected, and hence path connected, and so Cp,q 6= ∅ for all p, q ∈ M.

Definition 8.16 (Riemannian distance). The map d : M × M → [0, ∞) defined by

d(p, q) := inf{L(c) | c ∈ Cp,q },

is called the Riemannian distance of (M, g).

A curve γ which joins p and q is called length minimising if L(γ) = d(p, q).

Proposition 8.17 (Riemannian manifolds are metric spaces). With d defined as above, (M, d) is a metric
space.
The proof uses the following exercise.
Exercise 8.18 (Lengths of curves are invariant under reparameterisation). Show that lengths of curves are
invariant under reparameterisation.
Proof of Proposition 8.17. It follows from the definition of L(c) that d(p, q) = d(q, p) ≥ 0 and d(p, p) = 0 for
all p, q ∈ M.
Consider p, q, r ∈ M. If γ1 : [a, b] → M satisfies γ1 (a) = p, γ1 (b) = q and γ2 : [c, d] → M satisfies
γ2 (c) = q, γ2 (d) = r, define γ : [0, 2] → M, by


γ1 (b − a)s + a
if s ∈ [0, 1],
γ(s) =
γ2 (d − c)(s − 1) + c if s ∈ (1, 2].

By Exercise 8.18
L(γ) = L(γ1 ) + L(γ2 ).
Taking the infimum over all such γ1 , γ2 gives the triangle inequality

d(p, r) ≤ d(p, q) + d(q, r).

It remains to show that d(p, q) > 0 for all p 6= q. Consider p ∈ M, ε > 0 small and normal coordinates
on the geodesic ball Vε := expp (B(0, ε)) centred at p. If ε is sufficiently small there exist C, c > 0 such that
q q
c δij v i v j ≤ |v|g(r) ≤ C δij v i v j ,
p
for all r ∈ Vε , v ∈ Tr M. (Indeed |v|g(p) = δij v i v j for all v ∈ Tp M by Exercise 8.13. The claim then
follows by continuity.) Take ε sufficiently small so that q ∈
/ Vε and take any curve γ joining p and q. Let t0
be the first time such that γ(t0 ) ∈ ∂Vε . Then
Z t0 Z t0 q
|γ 0 (t)|g dt ≥ c


L(γ) ≥ L(γ|[a,t0 ] ) = δij γ 0 (t)i γ 0 (t)j dt ≥ cdEucl φ(γ(t0 )), φ(γ(a)) = cε.
a a

Taking the infimum over all such curves γ gives

d(p, q) ≥ cε > 0.

46
8.4 Variations of curves
The main result of this section is the statement that the geodesic equation is the Euler–Lagrange equation
of the length functional, a statement known as the first variation formula. See Proposition 8.23. It follows
that every length minimising curve is a geodesic.
Definition 8.19 (Admissible curves and admissible vector fields). An admissible curve is a continuous
piecewise smooth curve γ : [a, b] → M such that γ̇ is nonvanishing and left and right limits of γ̇ exist (in
every chart) everywhere.
If γ is an admissible curve, a vector field V ∈ X(γ) is called admissible if it is continuous and piecewise
smooth.
Such an admissible vector field V is called proper if V (a) = V (b) = 0.
Definition 8.20 (Variation of an admissible curve). A variation of an admissible curve γ : [a, b] → M is a
continuous map
A : (−ε, ε) × [a, b] → M,
such that
1. A(0, t) = γ(t) for all t ∈ [a, b].
2. There exists a subdivision a = t0 < t1 < . . . < tk = b such that A is smooth on each (−ε, ε) × (ti , ti+1 ).
3. For each s ∈ (−ε, ε) the curve t 7→ As (t) := A(s, t) is an admissible curve.
A variation A of γ is called proper if A(s, a) = γ(a), A(s, b) = γ(b) for all s ∈ (−ε, ε).
If A is a variation of an admissible curve γ then the vector fields along γ,
d d
∂t A(0, t) := A(s, t)|s=0 , ∂s A(0, t) := A(s, t)|s=0 ,
dt ds
are admissible vector fields along γ.
Exercise 8.21 (Every admissible vector field arises from a variation). Show that every admissible vector
field V along γ arises as V = ∂s A(0, t) for some variation A of γ.
If V is a vector field on the image of a variation A, let Dt denote the covariant derivative along the curves
t 7→ As (t) := A(s, t), and let Ds denote the covariant derivative along the curves s 7→ At (s) := A(s, t).
Lemma 8.22 (Covariant derivatives of a variation commute). If A : (−ε, ε) × [a, b] → M is a variation of
an admissible curve γ : [a, b] → M, then
Ds ∂t A = Dt ∂s A,
on any rectangle (−ε, ε) × (ti , ti+1 ) on which A is smooth.
Proof. Consider local coordinates {xi } around a point A(s0 , t0 ) at which A is smooth. Write A in coordinates
as A(s, t) = (A1 (s, t), . . . , An (s, t)), so that
∂Ai ∂Ai
∂t A = ∂ i, ∂s A = ∂ i,
∂t x ∂s x
and
∂ 2 Ai ∂Ai ∂ 2 Ai ∂Ai ∂Aj k
Ds ∂t A = ∂xi + ∇ ∂A ∂xi = ∂xi + Γ ∂ k.
∂s∂t ∂t ∂s ∂s∂t ∂t ∂s ij x
Similarly,
∂ 2 Ai ∂Ai ∂Aj k
Dt ∂s A = ∂xi + Γ ∂ k.
∂t∂s ∂s ∂t ij x
The proof follows from the fact that Γkij = Γkji with respect to any coordinate frame since ∇ is torsion
free.

47
Proposition 8.23 (First variation formula). If γ : [a, b] → M is an admissible curve, smooth on (ti , ti+1 )
for i = 0, . . . , k, parameterised by unit speed (i.e. |γ̇(t)|g = 1 for all t), if A is a proper variation of γ then
Z b k−1
d  X
−
gγ(ti ) V |ti , γ̇(t+


 L(As ) = − gγ(t) (V, Dt γ̇)dt − i ) − γ̇(ti ) ,
ds s=0 a i=1

where V (t) = ∂s A(0, t) and γ̇(t±

i ) denote the left and right limits of γ̇ at ti ,

γ̇(t−
i ) = lim γ̇(t), γ̇(t+
i ) = lim γ̇(t),
t↑ti t↓ti

(defined with respect to appropriate local coordinate charts) i.e. the geodesic equation is the Euler–Lagrange
equation of the length functional.
Proof. First note that




∂s g(∂t A(s, t), ∂t A(s, t)) = 2g Ds ∂t A(s, t), ∂t A(s, t) = 2g Dt ∂s A(s, t), ∂t A(s, t)




= 2∂t g ∂s A(s, t), ∂t A(s, t) − 2g ∂s A(s, t), Dt ∂t A(s, t) ,
by Lemma 8.22 and the fact that ∇ is compatible with g. Hence,
k−1
X Z ti+1
d  p

L(A s ) = ∂s g(∂t A(s, t), ∂t A(s, t)) dt
ds s=0

s=0
i=0 ti
k−1
X Z ti+1 


= ∂t g(∂s A(s, t), ∂t A(s, t)) − g(∂s A(s, t), Dt ∂t A(s, t))dt

ti s=0
i=0
k−1
X Z ti+1 
gγ(ti+1 ) V |ti+1 , γ̇(t− +



= i+1 ) − gγ(ti ) V |ti , γ̇(ti ) − gγ(t) (V, Dt γ̇(t))dt ,
i=0 ti

since γ is parameterised by unit speed. The proof follows from the fact that V |a = V |b = 0 since A is a
proper variation.
The following proposition in particular implies that every length minimising curve is a geodesic.
Proposition 8.24 (Geodesics and critical points of the length functional). A unit speed admissible curve γ
is a critical point of the length functional L if and only if it is a geodesic.
d
Proof. If γ is a geodesic then the first variation formula implies that ds |s=0 L(As ) = 0 for every variation A
of γ, i.e. γ is a critical point of L.
Suppose now that γ is a unit speed admissible curve which is a critical point of L. For each i = 0, . . . , k−1,
let ϕ be a smooth bump function such that ϕ(t) = 0 for t ≤ ti , t ≥ ti+1 and ϕ(t) > 0 for t ∈ (ti , ti+1 ).
Setting V (t) = ϕ(t)Dt γ̇(t) in the first variation formula (using Exercise 8.21) gives
Z ti+1
ϕ(t)|Dt γ̇(t)|2g dt = 0.
ti

It follows that Dt γ̇(t) = 0 for all t ∈ (ti , ti+1 ) for each i = 0, . . . , k − 1, and hence γ is a geodesic on each
interval on which it is smooth. To see that γ is actually a geodesic, for each i take V such that
−
V (ti ) = γ̇(t+
i ) − γ̇(ti ), V (tj ) = 0 for i 6= j.
The first variation formula then implies that
− 2
0 = |γ̇(t+
i ) − γ̇(ti )|g ,
−
for each i, i.e. γ̇(t+
i ) = γ̇(ti ) for each i and so γ is a geodesic.

48
8.5 Geodesics are locally length minimising
In the previous section it was shown that all length minimising curves are geodesics. The converse is not
true in general (consider great circles on the sphere). In this section a partial converse is shown: geodesics
are locally length minimising.
The main content is in the following.
Proposition 8.25 (Gauss Lemma). Given p ∈ M and normal coordinates {xi } centred at p on Vε =
i
expp (B(0, ε)), the vector field xr ∂xi on Vε r {p} is orthogonal to the geodesic spheres expp (∂B(0, δ)) for all
p
0 < δ < ε. Here r(x) = (x1 )2 + . . . + (xn )2 .
Exercise 8.26 (Scaling vector has unit length).
1. Show that if γ : (a, b) → M is a geodesic, then |γ̇(t)|g is conserved along γ.
xi
2. Show that, on Vε r {p}, the vector field r ∂x
i has unit length
 xi 
 ∂xi  = 1.
 
r g

xi i
Hint: use the fact that r ∂x
i = γ̇(1), where γ(t) = exp(p, t xr ei ) for some orthonormal basis {ei } for
Tp M.
Proof of Proposition 8.25. Consider q ∈ expp (∂B(0, δ)) and X ∈ Tq M such that X is tangential to expp (∂B(0, δ)).
There exists v ∈ Tp M such that q = exp(p, v) and w ∈ Tv Tp M = Tp M such that X = (expp )∗v w. Note
that V = xi (q)ei where (x1 (q), . . . , xn (q)) are the coordinates of q. Set γ(t) = exp(p, tv). Then γ(0) = p,
γ(1) = q and
∂  ∂  xi ∂ 
γ̇(1) = v i i  = xi (q) i  = δ  ,
∂x γ(1) ∂x q r ∂xi q
by the definition of normal coordinates. The proposition will follow from the fact that
gq (X, γ̇(1)) = 0.
Let σ : (−ε, ε) → ∂B(0, δ) ⊂ Tp M be a curve such that σ(0) = v, σ̇(0) = w. Define the smooth variation
A(s, t) = exp(p, tσ(s)).
Note that
d 
∂s A(0, 0) = exp(p, 0) = 0,
ds s=0

d 
∂s A(0, 1) = exp(p, σ(s)) = (expp )∗v w = X,
ds s=0

d 
∂t A(0, 1) =  exp(p, tv) = γ̇(1),
dt t=1
so it suffices to show that
d 
g(∂s A(s, t), ∂t A(s, t))  = 0.

dt s=0
Now,
d 
g(∂s A(s, t), ∂t A(s, t)) = g(Dt ∂s A(s, t), ∂t A(s, t)) + g(∂s A(s, t), Dt ∂t A(s, t))
dt
= g(Ds ∂t A(s, t), ∂t A(s, t))
1 d 
= g(∂t A(s, t), ∂t A(s, t)) = 0,
2 ds
by Lemma 8.22, the fact that the curves t 7→ A(s, t) are geodesics and so |∂t A(s, t)|g = |∂t A(s, 0)|g =
|σ(s)|g = δ for all s, by Exercise 8.26.

49
 Given p ∈ M and q in a normal neighbourhood of p, q ∈ Vε = expp (B(0, ε)) with q 6= p, there exists
0 < δ < ε and v ∈ Tp M with |v|g = 1 such that q ∈ expp (∂B(0, δ)) and q = exp(p, δv).
Proposition 8.27 (Geodesics are locally length minimising). The geodesic γ : [0, δ] → M defined by γ(t) =
exp(p, tv) is the unique (up to reparameterisation) length minimising curve joining p and q in M.

Proof. Note that |γ̇(t)|g = 1 for all t ∈ [0, δ], hence L(γ) = δ. Let σ : [a, b] → M be another curve joining p
and q, so that σ(a) = p, σ(b) = q. Assume that σ is parameterised by arc length, so that |σ̇(t)|g = 1 for all
t ∈ [a, b]. Suppose moreover that σ(t) 6= p for all t > a. Define

b0 = min{t ∈ [a, b] | σ(t) ∈ expp (∂B(0, δ))}.

For any t ∈ (a, b0 ] the tangent vector to σ can be written

xi
σ̇(t) = α(t) ∂ i + X(t),
r x
for some α(t) and some vector field X(t) which is tangent to the geodesic sphere through σ(t), expp (∂B(0, r(σ(t)))).
Note that X(t)r = 0 for all t since X(t) is tangential to the level sets of r, and
n
xi X xi xi
∂xi r = = 1,
r i=1
r2

and so σ̇(t)r = α(t). By the Gauss Lemma, Proposition 8.25,

 xi 
g ∂xi , X(t) = 0,
r
for all t, and so
|σ̇(t)|2g = (α(t))2 + |X(t)|2g ≥ (α(t))2 ,
by Exercise 8.26. Hence,
Z b0 Z b0 Z b0
L(σ) ≥ L(σ|(a,b0 ) ) = |σ̇(t)|g dt ≥ α(t)dt = σ̇(t)rdt
a a a
Z b0
dr(σ(t))
= dt = r(σ(b0 )) − r(σ(a)) = δ = L(γ).
a dt

Suppose now that L(σ) = L(γ). Then, by the above, L(σ) = L(σ|(a,b0 ) ) so it must be the case that b0 = b
and Z b Z b
|σ̇(t)|g dt = α(t)dt.
a a

It must then be the case that X(t) = 0 for all t and, since |σ̇(t)|g = 1, it must be the case that α(t) = 1 for
all t, and so
xi
σ̇(t) = ∂xi .
r
i
Since it is also the case that γ̇(t) = xr ∂xi it follows that σ and γ are both integral curves of the same vector
field passing through p, and therefore must be equal.

50
8.6 Completeness
Recall that a Riemannian manifold comes with a notion of geodesic completeness (see Definition 8.5). Since
M together with the Riemannian distance function d is a metric space, there is also a notion of completeness
of (M, g) as a metric space. (Recall that a metric space (M, d) is complete if every Cauchy sequence converges
to a limit in M .) These two notions of completeness coincide.
Theorem 8.28 (Hopf–Rinow). Let (M, g) be a (connected) Riemannian manifold. The following are equiv-
alent.
1. (M, d) is a complete metric space.

2. (M, g) is geodesically complete.

3. There exists p ∈ M such that Ep = Tp M.
4. The closed bounded subsets of M are compact.
Moreover, any one of the above conditions imply that, for any points p, q ∈ M, there exists a minimising
geodesic joining p and q.
The proof of the Hopf–Rinow Theorem is non-examinable.
Definition 8.29 (Complete Riemannian manifolds). A Riemannian manifold (M, g) is called complete if
either (M, g) is geodesically complete or, equivalently, if (M, d) is a complete metric space.

9 Curvature
One simple way to show that two Riemannian manifolds are not locally isometric is to show that they have
different curvature, which is the subject of this section. (Note that we still have not ruled out the possibility
that every Riemannian manifold is locally isometric to Euclidean space.)

9.1 The Riemann curvature tensor

In order to motivate the definition, recall that, in Euclidean space, every vector Zp ∈ Tp Rn can be extended
to a globally parallel vector field, i.e. to a vector field Z ∈ X(Rn ) such that ∇Eucl Z ≡ 0. Indeed, one simply
takes Z = Zpi ∂xi , where (x1 , . . . , xn ) are Cartesian coordinates on Rn . This property, in a sense, is a local
characterisation of Euclidean space and is taken as the basis for the definition of curvature.
On a Riemannian manifold (M, g), what is the obstruction to extending Zp ∈ Tp M to a parallel vector
field? For simplicity, suppose n = 2. Consider local coordinates (x, y) around p ∈ M. Take Zp ∈ Tp M and
extend to Z along {y = 0} by parallel transport

∇∂x Z|{y=0} = 0.

Now extend along {x = const} by parallel transport

∇∂y Z = 0.

Is it the case that Z is parallel? I.e. is it the case that ∇∂x Z|{y=c} = 0 for c 6= 0? It would be the case (by
uniqueness of parallel transport) if ∇∂y ∇∂x Z = 0, so what is the obstruction to this? If it were the case that
covariant derivatives commute (as they do in Euclidean space) then it would be the case that

∇∂y ∇∂x Z = ∇∂x ∇∂y Z = 0.

Curvature is therefore defined to be the failure of covariant derivatives to commute.

51
Definition 9.1 (Riemann curvature tensor). Given a Riemannian manifold (M, g), define the Riemann
curvature tensor of (M, g) to be the map R : X(M) × X(M) × X(M) → X(M) defined by

R(X, Y )Z := ∇X ∇Y Z − ∇Y ∇X Z − ∇[X,Y ] Z,

for X, Y, Z ∈ X(M), where ∇ is the Levi-Civita connection of g.

We also refer to R simply as the curvature tensor.
Exercise 9.2 (Failure of covariant derivatives to commute). Show that

R(X, Y )Z = ∇2 Z(Y, X) − ∇2 Z(X, Y ),

for all X, Y, Z ∈ X(M).

Remark 9.3 (Properties of the Riemann curvature tensor).
1. The definition of the curvature of a Riemannian manifold involves only the Levi-Civita connection ∇,
and not the metric g directly. One can therefore define similarly the curvature of any connection on a
vector bundle.
2. The curvature tensor R is anti-symmetric in the first two arguments, i.e.

R(X, Y )Z = −R(Y, X)Z,

for all X, Y, Z ∈ X(M).

3. There is no agreed upon sign convention for the Riemann curvature tensor. Many books define the
Riemann curvature tensor to be the negative of the above R.
Proposition 9.4 (The Riemann curvature tensor is a tensor field). For any X, Y, Z ∈ X(M), R(X, Y )Z|p
depends only on X|p , Y |p , Z|p , i.e. R is a (1, 3) tensor field R ∈ Γ(T31 M).
Proof. Given f ∈ C ∞ (M), it suffices to show that,

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

It follows from Exercise 9.2 that R(f X, Y )Z = R(X, f Y )Z = f R(X, Y )Z. Now

∇X ∇Y (f Z) = ∇X ((Y f )Z + f ∇Y Z) = XY f · Z + Y f · ∇X Z + Xf · ∇Y Z + f · ∇X ∇Y Z,

and similarly
∇Y ∇X (f Z) = Y Xf · Z + Xf · ∇Y Z + Y f · ∇X Z + f · ∇Y ∇X Z,
and
∇[X,Y ] (f Z) = [X, Y ]f Z + f ∇[X,Y ] Z,
from which it follows that R(X, Y )f Z = f R(X, Y )Z.
Define the (0, 4) Riemann curvature tensor of a Riemannian manifold (M, g), by

Rm(X, Y, Z, W ) = g(R(X, Y )Z, W ),

for all X, Y, Z, W ∈ X(M), i.e. Rm = R[ , so that Rm ∈ Γ(T40 M). We often still write R for Rm. In local
coordinates
∂
R = Rijk l dxi ⊗ dxj ⊗ dxk ⊗ , Rm = Rijkl dxi ⊗ dxj ⊗ dxk ⊗ dxl ,
∂xl
where Rijkl = glm Rijk m .

52
Exercise 9.5 (Riemann curvature tensor in local coordinates). Show that

Rijk l = ∂xi Γljk − ∂xj Γlik + Γm l m l

jk Γim − Γik Γjm .

Exercise 9.6 (Expansion for metric in normal coordinates). Let {xi } be normal coordinates centred at
p ∈ M. Show that
1
gij (x) = δij − Riklj (p)xk xl + O(|x|3 ),
3
around p = (0, . . . , 0). (Hint: attempt this problem after reading Section 11.)
It seems a priori that it requires n4 functions to locally describe R. In fact, R possesses many symmetries.
Proposition 9.7 (Algebraic symmetries of the curvature tensor). For all X, Y, Z, W ∈ X(M),
(i) Rm(X, Y, Z, W ) = −Rm(Y, X, Z, W ),
(ii) Rm(X, Y, Z, W ) = −Rm(X, Y, W, Z),
(iii) R(X, Y )Z + R(Z, X)Y + R(Y, Z)X = 0 (first Bianchi identity),
(iv) Rm(X, Y, Z, W ) = Rm(Z, W, X, Y ).
Proof. (i) is immediate from the definition.
For (ii), it follows from the fact that ∇ is compatible with g that

XY (g(W, Z)) = g(∇X ∇Y W, Z) + g(∇Y W, ∇X Z) + g(∇X W, ∇Y Z) + g(W, ∇X ∇Y Z), (18)

and similarly

Y X(g(W, Z)) = g(∇Y ∇X W, Z) + g(∇X W, ∇Y Z) + g(∇Y W, ∇X Z) + g(W, ∇Y ∇X Z), (19)

Subtracting (19) from (18) gives

[X, Y ](g(W, Z)) = g(∇X ∇Y W − ∇Y ∇X W, Z) + g(W, ∇X ∇Y Z − ∇Y ∇X Z),

and (ii) then follows from the fact that

[X, Y ](g(W, Z)) = g(∇[X,Y ] W, Z) + g(W, ∇[X,Y ] Z).

(iii) follows from a direct computation, using the torsion free property of ∇.
Finally, for (iv), note that (iii) implies that

Rm(X, Y, Z, W ) + Rm(Z, X, Y, W ) + Rm(Y, Z, X, W ) = 0,

Rm(Y, Z, W, X) + Rm(W, Y, Z, X) + Rm(Z, W, Y, X) = 0,
Rm(Z, W, X, Y ) + Rm(X, Z, W, Y ) + Rm(W, X, Z, Y ) = 0,
Rm(W, X, Y, Z) + Rm(Y, W, X, Z) + Rm(X, Y, W, Z) = 0.

Summing, and using (i) and (ii), gives

2Rm(Z, X, Y, W ) − 2Rm(Y, W, Z, X) = 0.

1 2 2
Accounting for the symmetries of Proposition 9.7, R can in fact be described locally by 12 n (n − 1)
functions.
The total covariant derivative ∇R possesses an additional symmetry.

53
Proposition 9.8 (Second Bianchi identity). For all V, W, X, Y, Z ∈ X(M),

∇X Rm(Y, Z, V, W ) + ∇Z Rm(X, Y, V, W ) + ∇Y Rm(Z, X, V, W ) = 0.

Exercise 9.9. Prove Proposition 9.8. Hint: use normal coordinates.

The Riemann curvature tensor is invariant under local isometries.
Exercise 9.10 (The Riemann curvature tensor is a geometric object). Suppose ϕ : (M, g) → (M̃, g̃) is a
local isometry. Show that ϕ∗ (R(X, Y )Z) = R̃(ϕ∗ X, ϕ∗ Y )ϕ∗ Z for all X, Y, Z ∈ X(M).
Vanishing of the Riemann curvature tensor is a local characterisation of Euclidean space.
Theorem 9.11 (Riemann curvature characterises flatness). A Riemannian manifold is flat if and only if
its curvature tensor vanishes identically.
Proof. A flat Riemannian manifold clearly has vanishing curvature tensor, so consider the converse. Compare
the proof with the motivating discussion at the beginning of the section.
We will use the following non-examinable fact: If {ei } is a local frame around p ∈ M such that [ei , ej ] = 0
for all i, j = 1, . . . , n, then there exists local coordinates {y i } around p such that ei = ∂yi for i = 1, . . . , n.
Using this fact it suffices, given p ∈ M, to find an orthonormal frame {ei } around p such that [ei , ej ] = 0
for all i, j, for then
gij = g(∂yi , ∂yj ) = g(ei , ej ) = δij ,
and so the map q 7→ (y 1 (q), . . . , y n (q) is a local isometry to Euclidean space.
∂
Let {xi } be normal coordinates around p and set ei |p = ∂x 2
i |p . Extend each ei along {x = . . . = x
n
= 0}
by parallel transport,
∇∂x1 ei |{x2 =...=xn =0} = 0.
Now extend along {x3 = . . . = xn = 0} by parallel transport,

∇∂x2 ei |{x3 =...=xn =0} = 0,

etc. Let Mk = {xk+1 = . . . = xn = 0}. We will inductively show that, for each k = 1, . . . , n,

∇∂xj ei |Mk = 0, for all j = 1, . . . , k. (20)

Indeed, for k = 1 (20) holds by definition. Assume then that (20) holds for some 1 ≤ k ≤ n − 1. Then

∇∂xk+1 ei |Mk+1 = 0,

by definition, and it remains to check that

∇∂xj ei |Mk+1 = 0,

for j = 1, . . . , k. Since R vanishes, it follows that

∇∂xk+1 ∇∂xj ei = ∇∂xj ∇∂xk+1 ei ,

and so
∇∂xk+1 ∇∂xj ei |Mk = 0.
By the inductive hypothesis ∇∂xj ei = 0 on Mk and so it follows that ∇∂xj ei = 0 on Mk+1 by uniqueness of
parallel transport.
Now ∇ei = 0 for each i = 1, . . . , n and so

[ei , ej ] = ∇ei ej − ∇ej ei = 0,

and the proof follows.

54
 The Riemann curvature tensor R can be extended to a map R : X(M) × X(M) × Γ(Tsr M) → Γ(Tsr M),

R(X, Y )T := ∇X ∇Y T − ∇Y ∇X T − ∇[X,Y ] T,

for all r, s ≥ 0.

Exercise 9.12 (Properties of the Riemann curvature tensor). Show that the Riemann curvature tensor
satisfies the following properties, for all X, Y ∈ X(M), T ∈ Γ(Tsr11 M), S ∈ Γ(Tsr22 M).



1. R(X, Y )(T ⊗ S) = R(X, Y )T ⊗ S + T ⊗ R(X, Y )S .
2. R(X, Y )f = 0 for all f ∈ C ∞ (M) = Γ(T00 M).


3. R(X, Y )(c(T )) = c R(X, Y )T for any contraction c.

9.2 Sectional curvature

Since the Riemann curvature tensor is a (1, 3) tensor field, it is difficult to make statements of the form “this
manifold has positive curvature”. Such statements are most easily made in terms of sectional curvature.
Definition 9.13 (Sectional curvature). Consider a Riemannian manifold (M, g), p ∈ M and a 2 dimen-
sional subspace (a 2-plane) Πp ⊂ Tp M. If U, V is a basis of Πp , the sectional curvature of Πp is defined
by
Rmp (U, V, V, U )
κ(Πp ) :=
2 .
gp (U, U )gp (V, V ) − gp (U, V )
If X, Y ∈ Tp M are linearly independent then we often write

κ(X, Y ) := κ(span{X, Y }).

Remark 9.14 (Properties of sectional curvature).

1. If the basis {U, V } is orthonormal, then

κp (Πp ) = Rmp (U, V, V, U ).

2. The sectional curvature κ(Πp ) does note depend on the choice of basis for Πp .

3. Recall the Gauss curvature of a surface. Given a 2-plane Πp ⊂ Tp M, define the local sub-surface of
M by S := expp (Πp ⊂ W ), where W ⊂ Tp M is a open neighbourhood of 0 such that such that expp |W
is a diffeomorphism onto its image. Then Tp S = Πp and κ(Πp ) is the Gauss curvature of S at p.
Clearly the Riemann curvature tensor determines all sectional curvatures. The converse is also true.
The following proposition guarantees that the Riemann curvature tensor can be recovered from all of the
sectional curvatures.
Proposition 9.15 (Sectional curvature determines the Riemann curvature tensor). Suppose two functions
R1 , R2 : Tp M × Tp M × Tp M × Tp M → R satisfy the algebraic symmetries of the Riemann curvature tensor
Rm and
R1 (X, Y, Y, X) R2 (X, Y, Y, X)

2 =
2 ,
2 2
|X|g |Y |g − gp (X, Y ) 2
|X|g |Y |2g − gp (X, Y )
for all X, Y ∈ Tp M. Then R1 = R2 .

55
Proof. Setting R3 = R1 − R2 , it follows that R3 (X, Y, Y, X) = 0 for all X, Y ∈ Tp M and it suffices to show
that
R3 (X, Y, Z, W ) = 0,
for all X, Y, Z, W ∈ Tp M. Now, for any X, Y, Z ∈ Tp M,

0 = R3 (X + Y, Z, Z, X + Y ) = R3 (X, Z, Z, Y ) + R3 (Y, Z, Z, X) + R3 (X, Z, Z, Y ) + R3 (Y, Z, Z, Y )

= 2R3 (X, Z, Z, Y ).
Hence, for all X, Y, Z, W ∈ Tp M,
0 = R3 (X, Z + W, Z + W, Y ) = R3 (X, Z, W, Y ) + R3 (X, W, Z, Y ).
The first Bianchi identity then implies
0 = R3 (X, Y, Z, W ) + R3 (Z, X, Y, W ) + R3 (Y, Z, X, W ) = 3R3 (X, Y, Z, W ).

The following exercise gives a geometric interpretation to sectional curvature.

Exercise 9.16 (Geometric interpretation of sectional curvature). Given p ∈ M and a 2-plane Πp ⊂ Tp M,
for r > 0 sufficiently small let Cr0 ⊂ Πp be the circle of radius r centred at the origin,
Cr0 = {v ∈ Πp | |v|g = r},
and let Cr := expp (Cr0 ) be its image under the exponential map. Show that
2πr − L(Cr ) π
lim = κ(Πp ),
r→0 r3 3
where L(Cr ) is the length of Cr .
Sectional curvature therefore measures the infinitesimal deviation of L(Cr ) with the length of a circle of
radius r in Euclidean space.
Definition 9.17 (Constant curvature manifolds). A Riemannian manifold (M, g) is called a constant cur-
vature manifold if there exists c ∈ R such that κ(Πp ) = c for all 2-planes Πp ⊂ Tp M for all p ∈ M.
Proposition 9.18 (Riemann curvature tensor of constant curvature manifolds). Given p ∈ M, if c ∈ R
then κ(Πp ) = c for all 2-planes Πp ⊂ Tp M if and only if


Rp (X, Y )Z = c gp (Y, Z)X − gp (X, Z)Y , (21)
for all X, Y, Z ∈ Tp M.
Proof. If (21) holds then it is straightforward to check that κ(Πp ) = c for all 2-planes Πp ⊂ Tp M. Suppose
then that κ(Πp ) = c for all 2-planes Πp ⊂ Tp M. Define


S(X, Y, Z, W ) = c gp (Y, Z)gp (X, W ) − gp (X, Z)gp (Y, W ) ,
and note that S has the same symmetries as the Riemann curvature tensor Rm. For any 2-plane Πp =
span{U, V },
S(U, V, V, U )
= c = κ(Πp ),
2
|U |g |V |2g − (gp (U, V ))2
and so S = Rmp by Proposition 9.15.
It follows that (M, g) is a constant curvature manifold if and only if


R(X, Y )Z = c g(Y, Z)X − g(X, Z)Y ,
for all X, Y, Z ∈ X(M).

56
9.3 Model spaces of constant curvature
Euclidean space (Rn , gEucl ) has constant curvature 0.
Exercise 9.19 (The round sphere is a constant curvature manifold). Show that the unit round sphere
(S n , gRound ) has constant curvature 1.

Recall the Poincaré ball model of hyperbolic space

4
B n = {(x1 , . . . , xn ) ∈ Rn | (x1 )2 + . . . + (xn )2 < 1}, (dx1 )2 + . . . + (dxn )2 .


gB n = 2 2
(1 − |x| )

Exercise 9.20 (Poincaré ball and Poincaré half space models of hyperbolic space are isometric). Show that
(B n , gB n ) is isometric to the Poincaré half space model of hyperbolic space
1
H n {(x1 , . . . , xn ) ∈ Rn | xn > 0}, (dx1 )2 + . . . + (dxn )2 .


gH n =
(xn )2

Hint: define f : B n → H n by
1
f (x1 , . . . , xn ) = (2x1 , . . . , 2xn−1 , 1 − |x|2 ).
(x1 )2 + ... + (xn−1 )2 + (xn − 1)2

Check that
1
f −1 (x1 , . . . , xn ) = (2x1 , . . . , 2xn−1 , |x|2 − 1),
(x1 )2 + . . . + (xn−1 )2 + (xn + 1)2

so that f is a diffeomorphism, and then check that f ∗ gH n = gB n .

Exercise 9.21 (Hyperbolic space is a constant curvature manifold). Show that (B n , gBn ) (and hence also
(H n , gH n )) has constant curvature −1.

9.4 Ricci curvature and scalar curvature

The Riemann curvature tensor is a complicated object and carries a lot of information. The Ricci curvature
and scalar curvature are simpler, but less definitive, measures of curvature.
If V is a vector space, recall the contraction map c1 1 : V ⊗ V ∗ → R. This map is also called trace, and is
denoted tr := c1 1 . Recall that R ∈ Γ(T31 M) and so, for p ∈ M and X, Y ∈ Tp M, one can view Rp (·, X)Y
as a map Rp (·, X)Y : Tp∗ M × Tp M → R given by Rp (·, X)Y (ξ, W ) = ξ(Rp (W, X)Y ).

Definition 9.22 (Ricci curvature tensor). The Ricci curvature tensor of a Riemannian manifold (M, g) is
the (0, 2) tensor field Ric ∈ Γ(T20 M) defined by


Ric(X, Y ) := tr R(·, X)Y .

We often write Ric(g) for Ric when it is necessary to emphasise the dependence on g.

Remark 9.23 (Properties of Ricci curvature).

1. In local coordinates the Ricci curvature tensor takes the form,

Ric = Ricij dxi ⊗ dxj ,

where Ricij = Rkij k = g kl Rkijl .

2. Due to the symmetries of R, taking the trace over any other two arguments gives either −Ric or 0.

57
 3. If {ei } is an orthonormal basis for Tp M, then
n
X
Ric(X, Y ) = Rm(ei , X, Y, ei ).
i=1

4. The Ricci curvature tensor is symmetric,

Ric(X, Y ) = Ric(Y, X),

for all X, Y ∈ X(M).

5. The Ricci curvature tensor is described locally by 21 n(n + 1) functions (compare with the 1 2 2
12 n (n − 1)
functions required to locally describe the Riemann curvature tensor).
6. The Ricci curvature tensor features in several famous geometric equations such as the Einstein equations
of general relativity, and the Ricci flow, which was famously used to resolve the Poincaré conjecture.
7. Since Ric is symmetric and bilinear, it is completely determined by the associated quadratic form
X 7→ Ric(X, X).
8. Given V ∈ Tp M with |V |g = 1, if V is extended to an orthonormal basis {V = e1 , e2 , . . . , en } then
n
X n
X
Ric(V, V ) = Rm(ek , e1 , e1 , ek ) = κ(V, ek ).
k=1 k=2

The following exercise gives a geometric interpretation to Ricci curvature.

Exercise 9.24 (Geometric interpretation of Ricci curvature). Let {xi } be normal coordinates centred at
p ∈ M. Show that
1
q
det gij (x) = 1 − Rickl (p)xk xl + O(|x|3 ),
6
around p = (0, . . . , 0). (Hint: use Exercise 9.6.)
It follows from Exercise 9.24 that, if γ is a unit speed geodesic such that γ(0) = p, then
p
det gij (γ(t)) − 1 1
lim = − Ric(γ̇(0), γ̇(0)),
t→0 t2 6
i.e., recalling Exercise 5.9, if v ∈ Tp M is a unit vector then Ric(v, v) measures the infinitesimal deviation
between the volume of a small wedge in the direction γp,v with the volume of the corresponding wedge in
Euclidean space.
If one views the Riemann curvature tensor as a “twisted hessian” of the metric g, then it is appropriate
to view the Ricci curvature tensor as some sort of Laplacian of g. Recall that, appropriately viewed, the
Laplacian of a function contains much of the information of the whole hessian of that function.
Exercise 9.25 (Ricci curvature in harmonic coordinates). Recall the Laplace–Beltrami operator (see Def-
inition 7.31). Show that, in local coordinates xi which satisfy ∆g (xi ) = 0, the Ricci curvature takes the
form
1
Ric(g)ij = − ∆g (gij ) + Nij (g, ∂g),
2
for appropriate functions Nij .
Definition 9.26 (Scalar curvature). The scalar curvature of a Riemannian manifold (M, g) is the function
S : M → R defined by
S = tr(Ric] ).

58
 Again, we often write S(g) for S to emphasise the dependence on g.
Remark 9.27 (Properties of scalar curvature).
1. In local coordinates
S = g ij Ricij = g ij Rkij k .

2. If p ∈ M and {ei } is an orthonormal basis of Tp M, then

n
X n
X X
S= Ric(ei , ei ) = Rm(ej , ei , ei , ej ) = κ(ei , ej ).
i=1 i,j=1 j6=i

The following exercise gives a geometric interpretation of scalar curvature. Since it involves volume, the
exercise is non-examinable.

Exercise 9.28 (*Geometric interpretation of scalar curvature). Given p ∈ M, show that

Vol(B(p, r)) − VolEucl (BEucl (0, r)) ωn−1

lim n+2
=− S(p),
r→0 r 6n(n + 2)

where Vol(B(p, r)) is the Riemannian volume of the geodesic ball B(p, r) ⊂ M,
R VolEucl (BEucl (0, r)) is the
Euclidean volume of BEucl (0, r) = {x ∈ Rn | |x| < r}, n = dimM, and ωn−1 = S n−1 dσS n−1 is the Euclidean
volume of S n−1 .
Exercise 9.28 in particular implies that scalar curvature S(p) is an infinitesimal measure of the deviation
of the volume of a geodesic ball centred at p with the volume of a Euclidean ball.

Definition 9.29 (Einstein manifolds). A Riemannian manifold (M, g) is called an Einstein manifold if
there exists a constant λ ∈ R such that
Ric(g) = λg.
Exercise 9.30 (Curvature and Einstein manifolds). Show that the scalar curvature of an Einstein manifold
is constant. Show that, if (Mn , g) has constant curvature c, then (Mn , g) is an Einstein manifold with
constant λ = (n − 1)c.
Remark 9.31 (Riemann curvature tensor and Ricci and scalar curvatures in low dimensions).
1. In dimension 2 the Riemann curvature tensor is determined by the scalar curvature and g,
1
κ(Tp M) = S(p).
2

2. In dimension 3, the Riemann curvature tensor is determined by the Ricci curvature and g,
1
Rijkl = Ricil gjk − Ricik gjl + Ricjk gil − Ricjl gik − S(gil gjk − gik gjl ).
2

3. A 3 dimensional Riemannian manifold is an Einstein manifold if and only if it has constant curvature.

10 Submanifolds
If M ⊂ M is a submanifold of a Riemannian manifold (M, g), many of the geometric properties of M with
the induced metric g can be related to the geometric properties of (M, g).

59
10.1 Second fundamental form
Let M ⊂ M be a submanifold of a Riemannian manifold (M, g) and let g denote the induced metric. Given
p ∈ M, the tangent space to M at p can be decomposed as

Tp M = Tp M ⊕ Np M,

where Tp M is the tangent space to M at p and Np M = (Tp M)⊥ is the orthogonal complement of Tp M ⊂
Tp M with respect to g, called the normal space. The disjoint union
[
NM = {p} × Np M,
p∈M

together with the natural projection map π : N M → M, π(p, n) = p is a vector bundle over M, called the
normal bundle.
For all p ∈ M, any v ∈ Tp M can be uniquely decomposed as

v = v> + v⊥ ,

where v > ∈ Tp M, v ⊥ ∈ Np M . Similarly, any vector field V ∈ X(M) when restricted to M can be uniquely
decomposed
V |M = V > + V ⊥ ,
with Vp> ∈ Tp M, Vp⊥ ∈ Np M for all p ∈ M.

Definition 10.1 (Second fundamental form). Define the second fundamental form of M to be the map
Π : X(M) × X(M) → Γ(N M) given by

Π(X, Y ) := (∇X Y )⊥ .

Proposition 10.2 (Properties of second fundamental form). The second fundamental form Π is symmetric
and bilinear over C ∞ (M).
Proof. Consider vector fields X, Y ∈ X(M) and recall that [X, Y ] ∈ X(M). Now

⊥
⊥
Π(X, Y ) − Π(Y, X) = ∇X Y − ∇Y X = [X, Y ] = 0,

since ∇ is torsion free, i.e. Π is symmetric. Clearly Π is linear over C ∞ (M) in the first argument. The
C ∞ (M) linearity over the second argument follows by the fact that it is symmetric.
In the case that M is an oriented hypersurface in M, i.e. dimM = dimM − 1, we often write

Π(X, Y ) = k(X, Y )n,

where n is the oriented unit normal to M. By a slight abuse, we also refer to k as the second fundamental
form. It follows from Proposition 10.2 that k is a symmetric (0, 2) tensor field, k ∈ Γ(T20 M). Since k is
symmetric and bilinear, it is determined by the associated quadratic form X 7→ k(X, X).
Exercise 10.3 (Second fundamental form of a hypersurface). Show that, for all X, Y ∈ X(M),

k(X, Y ) = −g(∇X n, Y ).

Proposition 10.4 (Covariant derivative of submanifold). If M is a submanifold of (M, g) then, for any
vector fields X, Y ∈ X(M),
∇X Y = ∇X Y + Π(X, Y ),
where ∇ is the Levi-Civita connection of the induced metric g.

60
Proof. The proof follows from checking that the map ∇> : X(M) × X(M) → X(M), defined by ∇> X Y :=
(∇X Y )> , is a connection on M which is symmetric and compatible with the induced metric g (Exercise).
The second fundamental form should be viewed as a measure of extrinsic curvature: it describes how the
submanifold curves within the ambient manifold.

Exercise 10.5 (Second fundamental form measures failure of (M, g) geodesics to be (M, g) geodesics). Let
M be a submanifold of (M, g) with the induced metric g. Show that any geodesic γ : [a, b] → M is also a
geodesic in (M, g) if and only if the second fundamental form Π vanishes identically.
If γ is a geodesic in (M, g), then k(γ̇, γ̇) measures the failure of γ to be a geodesic in (M, g).
Exercise 10.6 (Second fundamental form of cylinder in R3 ). Show that the cylinder

C := {(x, y, z) ∈ R3 | x2 + y 2 = 1} ⊂ R3 ,

with the induced Euclidean metric is flat. Consider the coordinate chart U = {(x, y, z) ∈ C | y > 0} and
φ : U → R2 defined by φ(x, y, z) = (x, z). Show that, in the associated (x, z) local coordinate system, the
second fundamental form of C is given by
1
k= dx2 .
x2 − 1

When M is an oriented hypersurface in M, one can consider the trace of the second fundamental form k.
The trace of the second fundamental form measures the change in volume as the hypersurface is deformed
in the normal direction.
Exercise 10.7 (Mean curvature measures change in volume under deformations in normal direction). Let
M be an oriented compact hypersurface of a compact Riemannian manifold (M, g). Consider the oriented
unit normal vector field n ∈ Γ(N M) and define

Mt := {Φt (p) | p ∈ M} ⊂ M,

where Φt : M → M is defined by Φt (p) := expp (tn(p)). For sufficiently small t, Mt is a hypersurface of M.

Show that, in any local coordinate system,
d 
q
q

det (g t )ij (p) = −trk(p) det (g t )ij (p),
dt t=0


where g t is the pullback by Φt of the induced metric on Mt .

The trace of the second fundamental form of a hypersurface, trk, is called the mean curvature.

10.2 The Gauss equation

The curvature of a submanifold can be related to the curvature of the ambient manifold using the second
fundamental form.

Theorem 10.8 (The Gauss equation). Let M is a submanifold of (M, g) with the induced metric g. The
Riemann curvature tensor of (M, g) satisfies

Rm(X, Y, Z, W ) = Rm(X, Y, Z, W ) + g(Π(X, W ), Π(Y, Z)) − g(Π(X, Z), Π(Y, W )),

for all X, Y, Z, W ∈ X(M).

61
Proof. Proposition 10.4 implies that



∇X ∇Y Z = ∇X ∇Y Z + Π(Y, Z) = ∇X ∇Y Z + Π(X, ∇Y Z) + ∇X Π(Y, Z) ,

where ∇ is the Levi-Civita connection of the induced metric g. Since g(W, Π(A, B)) = 0 for all A, B ∈ X(M),
it follows that





g(Π(X, ∇Y Z), W ) = 0, g ∇X Π(Y, Z) , W = −g Π(Y, Z), ∇X W = −g Π(Y, Z), Π(X, W ) .

Hence

g(∇X ∇Y Z, W ) = g(∇X ∇Y Z, W ) − g Π(Y, Z), Π(X, W ) ,
and similarly

g(∇Y ∇X Z, W ) = g(∇Y ∇X Z, W ) − g Π(X, Z), Π(Y, W ) ,
and

g(∇[X,Y ] Z, W ) = g ∇[X,Y ] Z + Π([X, Y ], Z), W = g(∇[X,Y ] Z, W ).
The proof follows.
If (M, g) is an embedded surface in Euclidean space, Proposition 10.8 in particular implies that

Rm(X, Y, Z, W ) = k(X, W )k(Y, Z) − k(X, Z)k(W, Y ).

This is the celebrated Theorema Egregium (“Remarkable Theorem”) of Gauss.

11 Jacobi fields
Jacobi fields are vector fields along a geodesic which measure the effect of curvature on neighbouring geodesics.

11.1 Definition and basic properties

The definition of a Jacobi field can immediately be given.
Definition 11.1 (Jacobi fields and the Jacobi equation). If γ : [a, b] → M is a geodesic, a vector field J
along γ, J ∈ X(γ), is called a Jacobi field if it satisfies the Jacobi equation,

Dt2 J = R(γ̇, J)γ̇. (22)

Equation (22) is called the Jacobi equation.

We now proceed to uncover the link between Jacobi fields and properties of neighbouring geodesics.
Definition 11.2 (Variation through geodesics). Let γ : [a, b] → M be a geodesic. A variation of γ through
geodesics is a smooth map A : (−ε, ε) × [a, b] → M such that A(0, t) = γ(t) for all t ∈ [a, b] and, for each
s ∈ (−ε, ε), the curve t 7→ A(s, t) is a geodesic.
Recall that a variation of γ gives rise to a variational vector field along γ,

∂A(s, t) 
∈ X(γ).
∂s

s=0

Theorem 11.3 (Variations through geodesics and Jacobi fields). If γ is a geodesic and A is a variation of γ
through geodesics, then J(t) := ∂A(s,t)
∂s |s=0 ∈ X(γ) is a Jacobi field, i.e. J satisfies the Jacobi equation (22).

62
Proof. It suffices to show that, for any vector field V ∈ X(A(s, ·)) for all s,
∂A ∂A

Ds Dt V − Dt Ds V = R , V. (23)
∂s ∂t
Indeed, if (23) holds, then, since t 7→ A(s, t) is a geodesic for all s ∈ (−ε, ε),
∂A ∂A ∂A ∂A
∂A
0 = Ds Dt = Dt Ds + , ,
∂t ∂t ∂s ∂t ∂t
by Lemma 8.22. Setting s = 0 gives (22).
To see that (23) holds, consider local coordinates {xi } and write V (s, t) = V i (s, t)∂xi . Then

∂2V i ∂V i ∂V i
Ds Dt V = ∂xi + Ds ∂xi + Dt ∂xi + V i Ds Dt ∂xi ,
∂s∂t ∂t ∂s
and similarly,
∂2V i ∂V i ∂V i
Dt Ds V = ∂xi + Dt ∂xi + Ds ∂xi + V i Dt Ds ∂xi ,
∂t∂s ∂s ∂t
and so
Ds Dt V − Dt Ds V = V i Ds Dt ∂xi − Dt Ds ∂xi .

Now
∂ 2 Aj ∂Aj ∂Ak
Ds Dt ∂xi = ∇∂xj ∂xi + ∇∂xk ∇∂xj ∂xi ,
∂s∂t ∂t ∂s
and
∂ 2 Aj ∂Aj ∂Ak
Dt Ds ∂xi = ∇∂xj ∂xi + ∇∂xk ∇∂xj ∂xi ,
∂t∂s ∂s ∂t
and so  ∂A ∂A 
Ds Dt ∂xi − Dt Ds ∂xi = R , ∂ i,
∂s ∂t x
and (23) follows.
The Jacobi equation (22) should be viewed as a second order ordinary differential equation.
Proposition 11.4 (Initial value problem for Jacobi equation). If γ : [a, b] → M is a geodesic then, given
X, Y ∈ Tγ(a) M, there exists a Jacobi field J along γ such that

J(a) = X, Dt J(a) = Y.

Proof. Let {ei |p } be an orthonormal basis of Tp M. Extend to an orthonormal frame {ei (t)} along γ by
parallel transport Dt ei = 0 (recall that parallel transport preserves lengths and angles). Writing J(t) =
J i (t)ei , the Jacobi equation becomes

J¨i (t) = Rjkl i (γ(t))γ̇ j (t)J k (t)γ̇ l (t),

for i = 1, . . . , n. The proof reduces to an initial value problem for this linear system of ODE with initial
conditions
J i (a) = X i , J˙i (a) = Y i ,
for which existence and uniqueness follows from Theorem 6.1. (As in the proof of Proposition 7.16, since
the equation is linear, the solution J i (t) lives for all t ∈ [a, b], for all i = 1, . . . , n.)
Exercise 11.5 (Variations through geodesics and Jacobi fields). Show that every Jacobi field J along a
geodesic γ : [a, b] → M arises as the variational vector field of some variation of γ through geodesics. Hint:
let σ : (−ε, ε) → M be a curve in M such that σ(0) = γ(a), σ̇(0) = J(a) and let W ∈ X(σ) be a vector field
such that W (0) = γ̇(a), Ds W (0) = Dt J(a), then define A(s, t) = exp(σ(s), tW (s)).

63
 Given a geodesic γ, define the space of Jacobi fields

J (γ) = {Jacobi fields along γ}.

Since the Jacobi equation (22) is linear, it follows that J (γ) is a vector space. It follows from Proposition
11.4 that J (γ) is 2n dimensional, where n = dimM.
Along any geodesic γ : [a, b] → M there are two trivial Jacobi fields γ̇(t) and tγ̇(t). These two Jacobi
fields arise from the variations A(s, t) = γ(s + t) and A(s, t) = γ(es t) which reparameterise γ. These two
Jacobi fields span a two dimensional subspace of J (γ), and so the space of nontrivial Jacobi fields along γ
has dimension 2n − 2.
Exercise 11.6 (Normal Jacobi fields). Let γ : [a, b] → M be a geodesic. Suppose X, Y ∈ Tγ(a) M are normal
to γ, i.e. g(X, γ̇) = g(Y, γ̇) = 0. Show that the Jacobi field J arising from Proposition 11.4 is normal to γ,
i.e. g(J(t), γ̇(t)) = 0 for all t ∈ [a, b].
The nontrivial Jacobi fields along γ can be determined from the trivial Jacobi fields and those normal to
γ.
The Jacobi fields of constant curvature manifolds which vanish initially can be computed explicitly.

Proposition 11.7 (Jacobi fields of constant curvature manifolds). Suppose (M, g) has constant curvature
c and let γ : [a, b] → M be a unit speed geodesic. If J(t) is a normal Jacobi field along γ which vanishes
initially, J(a) = 0, then  √
sin(t c)

 √
c
W (t) if c > 0,
J(t) = √
tW (t) if c = 0,
 sinh(t
 −c)
√
−c
W (t) if c < 0,
where W (t) is parallel along γ, Dt W (t) = 0.
Proof. Recall that (M, g) has constant curvature c if and only if


R(X, Y )Z = c g(Y, Z)X − g(X, Z)Y .

Hence
Dt2 J = c g(γ̇, J)γ̇ − g(γ̇, γ̇)J ,

and so
Dt2 J + cJ = 0.
Suppose J(t) = f (t)W (t), where W is normal to γ and parallel along γ. Then

f 00 (t) + cf (t) W (t) = 0,

and so, if J is nontrivial,

f 00 (t) + cf (t) = 0.
The unique solution of this ordinary differential equation with f (0) = 0 is a constant multiple of
 √
sin(t c)

 √
c
if c > 0,
f (t) = √
t if c = 0,
 sinh(t
 −c)
√
−c
if c < 0.

The proof then follows from the fact that the dimension of the space of normal Jacobi fields which vanish
initially is n − 1.

64
11.2 The second variation formula, the index form, and conjugate points
In this section we are concerned with the question of when geodesics fail to be length minimising with respect
to “nearby” curves.
If A is a variation of a geodesic γ : [a, b] → M then, if γ is length minimising amongst all nearby curves,
it must be the case that
d2 
 L(A(s, ·)) ≥ 0,
ds2 s=0
i.e. γ is a local minimum of the length functional.

Proposition 11.8 (Second variation formula). Let γ : [a, b] → M be a unit speed geodesic and let A : (−ε, ε)×
[a, b] → M be a proper variation of γ. Then
b
d2 
Z
 L(A(s, ·)) = |Dt V ⊥ (t)|2g − Rm(V ⊥ (t), γ̇(t), γ̇(t), V ⊥ (t))dt,
ds2 s=0 a

where V ⊥ is the normal component of V = ∂A

∂s |s=0 :

V ⊥ = V − g(V, γ̇)γ̇.

Proof. Assume, for simplicity, that A is smooth. In the proof of the first variation formula, Proposition 8.23,
we showed that Z b
d g(Dt ∂s A, ∂t A)
L(A(s, ·)) = p dt.
ds a g(∂t A, ∂t A)
Now
d  g(Dt ∂s A, ∂t A)  g(Ds Dt ∂s A, ∂t A) + g(Dt ∂s A, Ds ∂t A) g(Dt ∂s A, ∂t A)g(Ds ∂t A, ∂t A)
= −
|∂t A|3
p
ds g(∂t A, ∂t A) |∂t A|
g(Dt Ds ∂s A + R(∂s A, ∂t A)∂s A, ∂t A) + |Dt ∂s A|2 g(Dt ∂s A, ∂t A)2
= − ,
|∂t A| |∂t A|3

by Lemma 8.22 and equation (23). Hence

b
d2 
Z
 L(A(s, ·)) = g(Dt Ds ∂s A|s=0 , γ̇) − Rm(V (t), γ̇(t), γ̇(t), V (t)) + |Dt V |2g − g(Dt V, γ̇)2 dt.
ds2 s=0 a

Clearly,
Rm(V (t), γ̇(t), γ̇(t), V (t)) = Rm(V (t)⊥ , γ̇(t), γ̇(t), V (t)⊥ ),
and
Dt V = Dt V ⊥ + g(Dt V, γ̇)γ̇,
so that
|Dt V |2g = |Dt V ⊥ |2g + g(Dt V, γ̇)2 .
The proof follows from the fact that
d

g(Dt Ds ∂s A, γ̇) = g(Ds ∂s A, γ̇) ,
dt
and so Z b
g(Dt Ds ∂s A, γ̇)dt = 0,
a
since A is a proper variation.

65
Definition 11.9 (Index form). Let γ : [a, b] → M be a geodesic. For vector fields X, Y ∈ X(M) which are
normal to γ, the symmetric bilinear map defined by
Z b
I(X, Y ) := g(Dt X, Dt Y ) − Rm(X, γ̇, γ̇, Y )dt,
a

is called the index form of γ.

The second variation formula, Proposition 11.8, implies that, for any variation A of γ,

d2 
 L(A(s, ·)) < 0,
ds2 s=0
if and only if
∂A 
I(V, V ) < 0, V =  .
∂s s=0
If X, Y ∈ X(γ) are admissible vector fields then
Z b k
X
−
I(X, Y ) = − g(Dt2 X − R(γ̇, X)γ̇, Y )dt − g(Dt X(t+
i ) − Dt X(ti ), Y (ti )),
a i=0

where ti ∈ (a, b), for i = 0, . . . , k are the points at which X is not smooth.

Definition 11.10 (Conjugate points). If p ∈ M and γ : [a, b] → M is a geodesic such that γ(a) = p, a point
q = γ(t∗ ) for some t∗ ∈ (a, b] is called a conjugate point to p along γ if there exists a nontrivial Jacobi field
J along γ such that J(a) = J(b) = 0.
Exercise 11.11 (Normal Jacobi fields). Let J be a Jacobi field along γ : [a, b] → M such that J(a) = J(b) =
0. Show that J is normal to γ.

The following theorem implies that geodesics are not length minimising past conjugate points.
Theorem 11.12 (Geodesics are not length minimising past conjugate points). If γ : [a, b] → M is a unit
speed geodesic and there exists c ∈ (a, b) such that γ(c) is a conjugate point to γ(a), then there exists a proper
normal vector field V ∈ X(γ) such that I(V, V ) < 0.

Proof. There exists a nontrivial Jacobi field J along γ such that J(a) = J(c) = 0. Define

J(t) if t ∈ [a, c],
X(t) :=
0 if t ∈ (c, b].

It follows from Exercise 11.11 that X is a normal proper vector field along γ, which is smooth everywhere
except at c. Let Y ∈ X(γ) be such that Y (c) = −Dt X(c− ). Note that Y 6= 0 as J is nontrivial. Now, given
ε > 0, set V = X + εY . Then, since I(X, X) = 0 and I(X, Y ) = −|Dt X(c− )|2 , it follows that

I(V, V ) = −2ε|Dt X(c− )|2 + ε2 I(Y, Y ) < 0,

if ε is sufficiently small.

12 Classical theorems in Riemannian geometry

This section concerns two classical theorems in Riemannian geometry, the Bonnet–Myers Theorem and the
Cartan–Hadamard Theorem, which both involve deducing topological properties of Riemannian manifolds
based on some geometric assumptions satisfied by the Riemannian metric.

66
12.1 The Bonnet–Myers Theorem
The Bonnet–Myers Theorem guarantees that, if the Ricci curvature of a Riemannian manifold obeys an
appropriate lower bound, then the manifold is compact and satisfies a diameter bound.
Definition 12.1 (Diameter of a Riemannian manifold). If (M, g) is a Riemannian manifold, the diameter
of (M, g) is defined by
diam(M) := sup{dg (p, q) | p, q ∈ M},
where dg is the Riemannian distance.
Theorem 12.2 (Bonnet–Myers). Let (Mn , g) be a complete (and connected) Riemannian manifold and
suppose ρ > 0 is such that
n−1
Ric(X, X) ≥ g(X, X),
ρ2
for all X ∈ X(M). Then M is compact and diam(M) ≤ πρ.
Remark 12.3 (Remarks on Bonnet–Myers Theorem).
1. Think of the Bonnet–Myers Theorem as a “weak topological assumption” (completeness) plus a “curva-
ture assumption” (the lower bound on the Ricci curvature) guaranteeing a “strong topological statement”
(the diameter bound) and “topological information” (compactness).
2. The round sphere (S n , gRound ) satisfies Ric = (n − 1)g and diam(S n ) = π, so the theorem is sharp is
this sense.
3. In fact, the Bonnet–Myers Theorem involves “comparing with (S n , gRound )”. The assumption involves
comparing the Ricci curvature of (M, g) with that of (S n , gRound ). The Vi (t) considered in the proof
are exactly the Jacobi fields of the round sphere which vanish at t = 0 (see Proposition 11.7).
4. The lower bound on the Ricci curvature cannot be relaxed to Ric(X, X) > 0 for all X ∈ X(M) as the
the paraboloid
{(x, y, z) ∈ R3 | z = x2 + y 2 },
with the induced Euclidean metric satisfies Ric(X, X) > 0 (Exercise) but is complete and not compact.
The theorem is then also sharp in this sense.
Proof of Bonnet–Myers Theorem. Suppose that the diameter bound fails. Then, there exists p, q ∈ M and
a length minimising unit speed geodesic γ : [0, L] → M such that γ(0) = p, γ(L) = q, and L(γ) > ρπ. Then,
for every proper vector field V ∈ X(γ),
I(V, V ) ≥ 0. (24)
We will construct a V which contradicts this fact.
Extend e1 := γ̇(0) to an orthonormal basis {ek }nk=1 for Tp M and extend to an orthonormal frame along
γ by parallel transport. Set  
πt
Vi (t) := sin ei (t),
L
for i = 1, . . . , n, where L = L(γ). Then
n n Z L n Z L
π2
   
X X X πt
πt
I(Vi , Vi ) = −g(Dt2 Vi − R(γ̇, Vi )γ̇, Vi )dt = sin2 − sin2
Rm(ei , γ̇, γ̇, ei )dt
i=2 i=2 0 i=2 0
L2 LL
Z L   2 
πt π 1
≤ (n − 1) sin2 − dt < 0,
0 L L2 ρ2
and so there exists i = 2, . . . , n such that I(Vi , Vi ) < 0, which contradicts (24). Hence diam(M) ≤ ρπ. Given
p ∈ M, it follows that expp : B(0, 2ρ) → M is a surjection, hence M is compact.

67
12.2 The Cartan–Hadamard Theorem
The Cartan–Hadamard Theorem guarantees that a simply connected Riemannian manifold with non-positive
curvature is diffeomorphic to Rn .
Definition 12.4 (Simply connected topological space). A topological space M is called simply connected if,
for all pairs of curves γ1 , γ2 : [0, 1] → M such that γ1 (0) = γ1 (1) = γ2 (0) = γ2 (1), there exists a continuous
map A : [0, 1]2 → M such that A(0, t) = γ1 (t) and A(1, t) = γ2 (t) for all t ∈ [0, 1].
Theorem 12.5 (Cartan–Hadamard). If (Mn , g) is a complete simply connected Riemannian manifold of
non-positive curvature (i.e. κ(Πp ) ≤ 0 for all 2-planes Πp ⊂ Tp M for all p ∈ M), then M is diffeomorphic
to Rn . More precisely, for all p ∈ M the exponential map expp : Tp M → M is a diffeomorphism.

The proof of the Cartan–Hadamard Theorem relies on the following.

Lemma 12.6 (Conjugate points and the exponential map). Given p ∈ M and v ∈ Tp M, the exponential
map expp : Tp M → M is a local diffeomorphism around v if and only if q = expp (v) is not conjugate to p
along the geodesic t 7→ expp (tv) =: γ(t).

Proof. By the inverse function theorem expp is a local diffeomorphism near v if and only if (expp )∗v : Tv Tp M =
Tp M → Tq M is an isomorphism if and only if (expp )∗v is injective. Consider the variation of geodesics
A : (−ε, ε) × [0, 1] → M defined by A(s, t) = expp (t(v + sw)). The variational vector field J(t) = ∂s A(0, t)
is a Jacobi field along γ satisfying J(0) = 0, J(1) = (expp )∗v w. Since the dimension of the space of Jacobi
fields with J(0) = 0 is n, all such Jacobi fields with J(0) = 0 take this form for some w ∈ Tp M. Hence
(expp )∗v fails to be injective if and only if there exists w 6= 0 such that (expp )∗v w = 0 if and only if there
exists a nontrivial Jacobi field J satisfying J(0) = J(1) = 0.
Proof of the Cartan–Hadamard Theorem. We will only show that, for all p ∈ M, the exponential map
expp : Tp M → M is a local diffeomorphism. This step contains most of the content of the proof.
Consider p ∈ M and v ∈ Tp M, and let J be a Jacobi field along t 7→ expp (tv) such that J(0) = 0. Then

d2
(g(J, J)) = −2Rm(J, γ̇, γ̇, J) + 2|Dt J|2 = −2κ(J, γ̇) |J|2 |γ̇|2 − (g(γ̇, J))2 + 2|Dt J|2 ≥ 0,


dt2
by the Cauchy–Schwarz inequality. It follows that
d
(g(J, J)) ≥ 0,
dt
i.e. |J(t)|2g is non-decreasing. Since J(0) = 0, if J(t∗ ) = 0 for some t∗ > 0 then it must be the case that
J ≡ 0. Hence there are no conjugate points to p. Lemma 12.6 then implies that expp : Tp M → M is a local
diffeomorphism.

13 *Lorentzian geometry and Penrose’s incompleteness theorem

Recall that, for n ≥ 1, a semi Riemannian metric g on an n+1 dimensional manifold M is called a Lorentzian
metric if it has signature (−, +, . . . , +), i.e. if, for each p ∈ M, there exists e0 , e1 , . . . , en ∈ Tp M such that

gp (e0 , e0 ) = −1, gp (ei , e0 ) = 0 for i = 1, . . . , n, gp (ei , ej ) = δij , for i, j = 1, . . . , n.

A pair (M, g) is called a Lorentzian manifold. We call such a manifold n + 1 dimensional to emphasise the
Lorentzian signature of the metric. Much (but certainly not all) of the theory of Riemannian manifolds carries
over to Lorentzian manifolds, for example a Lorentzian metric on a manifold defines a unique Levi-Civita
connection, which in turn defines parallel transport and a Riemann curvature tensor.

68
 Lorentzian manifolds are of particular interest in general relativity – a theory of gravity postulated by
Einstein in 1915, which concerns Lorentzian manifolds satisfying the Einstein equations. In the absence of
matter, the vacuum Einstein equations take the remarkably simple form
Ric(g) = 0. (25)
Appropriately viewed, (25) are a system of nonlinear hyperbolic partial differential equations for g.
The goal of this section is to sketch a proof of the Penrose Incompleteness Theorem – a seminal result in
Lorentzian geometry, which profoundly shaped the way we view black holes and general relativity, and was
accordingly celebrated with the award of the 2020 Nobel prize in physics to Roger Penrose. The theorem
can, in a sense, be viewed as the Lorentzian analogue of the Bonnet–Myers Theorem and is thus a nice
application of many of the ideas developed in this course. Before stating the theorem there are several
notions in Lorentzian geometry which must first be introduced.
For the rest of this section it will be assumed that (M, g) is a 3 + 1 dimensional Lorentzian manifold.
The prototype of such a space is Minkowski space, namely R3+1 equipped with the Minkowski metric
g = −dt2 + (dx1 )2 + (dx2 )2 + (dx3 )2 .

13.1 Timelike, null, and spacelike vectors, curves, and submanifolds

The failure of the metric g to be positive definite means that a Lorentzian manifold does not retain the
metric space structure of a Riemannian manifold. A Lorentzian metric, however, retains some convexity
properties when restricted to certain directions.
Definition 13.1 (Timelike, spacelike, and null vectors). Given p ∈ M, a vector v ∈ Tp M is called timelike
if gp (v, v) < 0, spacelike if gp (v, v) > 0, or null if gp (v, v) = 0. A vector v ∈ Tp M is called causal if it is
either timelike or null.
Definition 13.2 (Timelike, spacelike, and null curves). A curve γ : (a, b) → M is called timelike if
g(γ̇(t), γ̇(t)) < 0 for all t ∈ (a, b), spacelike if g(γ̇(t), γ̇(t)) > 0 for all t ∈ (a, b), or null if if g(γ̇(t), γ̇(t)) = 0
for all t ∈ (a, b).
A symmetric bilinear map h on a vector space V is called degenerate if there exists X ∈ V with X 6= 0
such that h(X, Y ) = 0 for all Y ∈ V .
Definition 13.3 (Timelike, spacelike, and null submanifolds). A submanifold N ⊂ M is called timelike
if the induced metric g|N is Lorentzian, spacelike if the induced metric g|N is Riemannain, and null if the
induced metric g|N is degenerate.

13.2 Time orientations and causal structure

One new feature of Lorentzian geometry, not present in Riemannian geometry, is the notion of causality.
Definition 13.4 (Time orientations). A time orientation on a Lorentzian manifold (M, g) is a global
(nowhere vanishing) timelike vector field T . A causal vector v in a time oriented Lorentzian manifold
(M, g, T ) is called future directed if g(v, T ) < 0, and past directed if g(v, T ) > 0.
From now on, Lorentzian manifolds will be assumed to be time oriented.
Definition 13.5 (Causal and chronological futures). Given an Lorentzian manifold (M, g) and a set S ⊂ M,
the causal future of S is the set
J + (S) := {p ∈ M | there exist a causal curve γ : [a, b] → M with γ(a) ∈ S, γ(b) = p}.
The chronological future of S is the set
I + (S) := {p ∈ M | there exist a timelike curve γ : [a, b] → M with γ(a) ∈ S, γ(b) = p}.
There are similar definitions of causal past and chronological past but, as we are progressive, let us not
concern ourselves with these here.

69
13.3 Global hyperbolicity
Recall that a curve γ : (a, b) → M in a manifold M is called inextendible if it is not the restriction of a curve
γ̃ defined on a strictly larger domain than (a, b).
Definition 13.6 (Cauchy hypersurfaces and global hyperbolicity). Let (M, g) be a Lorentzian manifold.
A spacelike hypersurface H ⊂ M is called a Cauchy hypersurface if every inextendible causal curve in M
intersects H exactly once. A Lorentzian manifold (M, g) is called globally hyperbolic if it admits a Cauchy
hypersurface.
When g is a Lorentzian metric, the associated Laplace–Beltrami operator (see Definition 7.31) is a
hyperbolic operator and hence is denoted g (and the notation ∆g is reserved for when g is Riemannian and
hence the associated Laplace–Beltrami operator is of elliptic type). For example, when g is the Minkowski
metric, the associated Laplace–Beltrami operator is the standard wave operator −∂t2 + ∆x .
The significance of globally hyperbolic manifolds is that solutions of hyperbolic equations, whose symbol
is related to that of the wave equation
g ψ = 0,
such as the wave equation itself, or the vacuum Einstein equations (25) (recall Exercise 9.25), are uniquely
determined by their Cauchy data on a Cauchy hypersurface.

13.4 Closed trapped 2-surfaces and black holes

If S ⊂ M is a spacelike 2-surface then, for each p ∈ S, there are two associated future directed null normal
directions, spanned by future directed null vectors Lp and Lp normalised so that g(Lp , Lp ) = −2.5 If S is a
closed surface then we can designate the vector field L on S to be incoming and the vector field L on S to be
outgoing. To each null normal L and L one can associate a second fundamental form χ and χ respectively,
χ(X, Y ) = g(∇X L, Y ), χ(X, Y ) = g(∇X L, Y ),

for all X, Y ∈ Tp S.6

Definition 13.7 (Closed trapped surfaces). A closed spacelike 2-surface S is called a closed trapped surface
if the traces of the two associated null second fundamental forms are negative,
trχ < 0, trχ < 0,
on all of S.
In view of Exercise 10.7, one should view a closed trapped surface as a closed surface whose area de-
creases when deformed in each of the two null directions. Contrast with the standard spacelike 2-spheres in
Minkowski space.
The notion of a closed trapped surface is closely tied to the notion of a black hole. The definition of
a black hole is more difficult to give than that of a closed trapped surface due to the global aspect of the
definition. Informally, points are said to be in a black hole region if their causal future does not ever meet
“distant observers”. In order to make the definition more precise one needs to make this notion of “distant
observers” more precise.
Definition 13.8 (Informal definition of a black hole). A Lorentzian manifold (M, g) is said to contain a
black hole region if there is an appropriate asymptotic boundary, called future null infinity and denoted I + ,
representing “distant observers”, and a point p ∈ M such that
J + (p) ∩ I + = ∅.
5 Note that these vectors are not unique, as any positive multiples of L and L also describe these directions. Since these
p p
vectors are null, and hence have zero length, there is no way to fix them by, for example, insisting that they each have unit
length. Any rescaling Lp 7→ aLp , Lp 7→ a−1 Lp will preserve the condition that g(Lp , Lp ) = −2, but the choice of Lp and Lp
are otherwise unique.
6 Note the difference in sign convention from that of Section 10.

70
The set of all such p is called the black hole region, denoted B.
Again, note the global aspect of Definition 13.8. In contrast, compare with the definition of a closed
trapped surface, which is purely local. One can easily show that, if a Lorentzian manifold (M, g) contains a
closed trapped surface and obeys an appropriate Ricci curvature assumption (for example, if (M, g) solves
the vacuum Einstein equations (25)), then, if (M, g) admits an appropriate notion of future null infinity
I + , (M, g) contains a black hole region (in the sense of Definition 13.8) which contains the closed trapped
surface.

13.5 The Penrose Incompleteness Theorem

The Penrose theorem can now be stated.
Theorem 13.9 (Penrose, 1965). Let (M, g) be a 3 + 1 dimensional globally hyperbolic Lorentzian manifold,
with a non-compact Cauchy hypersurface, whose Ricci curvature satsifies

Ric(V, V ) ≥ 0, (26)

for all null vectors V . Suppose moreover that (M, g) contains a closed trapped surface, in the sense of
Definition 13.7. Then (M, g) is future geodesically incomplete.
Note that the assumption (26) is trivially satisfied by solutions of the vacuum Einstein equations (25).
Theorem 13.9 is often called a singularity theorem. Note however that the conclusion of the theorem
involves geodesic incompleteness (hence the name “incompleteness theorem”) rather than any singular be-
haviour (which one usually associates with something becoming infinite). One of the most remarkable aspects
of Theorem 13.9 is that it applies in different cases where the nature of the incompleteness is very different.
Two primary examples are discussed in Section 13.6 below. The fact that one can make such a soft con-
clusion as geodesic incompleteness of solutions of a highly nonlinear system of partial differential equations
such as (25) without in fact understanding the nature of that incompleteness is another remarkable aspect
of Theorem 13.9.
Theorem 13.9 can be viewed as a Lorentzian analogue of the Bonnet–Myers Theorem, Theorem 12.2.
Indeed, following [5], the Bonnet–Myers Theorem can be phrased

Completeness,
⇒ Compactness.
Lower bound on Ricci curvature,
The Penrose Theorem, stated in contrapositive form, can similarly be phrased,
Completeness,
Lower bound on Ricci curvature, ⇒ Compactness of every Cauchy hypersurface.
Existence of closed trapped surface,

13.6 The Schwarzschild and Kerr families of black holes

In order to properly understand the significance of Theorem 13.9 it is helpful to view it in the light of some
examples of solutions of the vacuum Einstein equations (25).
The most famous family of solutions of (25) is the one parameter Schwarzschild family. Discovered in
1915, very shortly after Einstein arrived at the final form of the field equations of general relativity (which
reduce to (25) in the absence of matter), the Schwarzschild metrics take the most familiar form, for fixed
M > 0,
   −1
2M 2M
gM = − 1 − dt2 + 1 − dr2 + r2 (dθ2 + sin2 θdφ2 ). (27)
r r
Compare with the Minkowski metric which, in polar coordinates on R3 , takes the form

g = −dt2 + dr2 + r2 (dθ2 + sin2 θdφ2 ).

71
There was much confusion surrounding the metric (27) for many years after its discovery, in particular due to
its apparently singular nature at r = 0 and r = 2M . Nowadays, we understand that the (t, r, θ, φ) coordinate
chart in which (27) is written covers only (appropriate subsets of) the regions r > 2M and 0 < r < 2M of
the maximally extended Schwarzschild solution
8M 3 − r
gM = − e 2M dU dV + r2 γ̊, (28)
r
defined on the manifold M = W × S 2 , where W = {(U, V ) ⊂ R2 | U V < 1} and γ̊ is the unit round metric
on S 2 . In the expression (28), the function r = r(U, V ) is defined implicitly by the relation
 r   r 
− 1 exp = −U V.
2M 2M
The Lorentzian manifold (M, gM ) is inextendible as a suitably regular Lorentzian manifold, is globally
hyperbolic with non-compact Cauchy hypersurface {U + V = 0}, and each spacelike 2-sphere defined by
U = U0 , V = V0 is a closed trapped surface if U0 > 0 and V0 > 0. The Penrose Theorem thus applies to
(M, gM ). The nature of the incompleteness is a curvature singularity at r = 0 (or equivalently at U V = 1),
as can be seen by observing that |Rm|gM ∼ M r 3 as r → 0. This singular behaviour, in an otherwise perfectly
reasonably solution of (25), disturbed people at first. A common opinion was that this singular behaviour
was a pathology, resulting from the high degrees of symmetry possessed by the Schwarzschild solution. Given
that any small perturbation of the Schwarzschild solution will also possess a closed trapped surface (noting
that the conditions trχ < 0, trχ < 0 are open conditions), the Penrose Theorem guarantees that, on the
contrary, this apparently pathological behaviour, provided one is willing to weaken “singularity” to “geodesic
incompleteness”, is a generic feature of solutions of (25).
What could the nature of the geodesic incompleteness predicted by the Penrose Theorem be if not a
singularity? The answer to this question is best addressed by way of another example, in the form of the
Kerr family. For given parameters |a| ≤ M , the Kerr metric takes the most familiar form

∆ 2 2 %2 2 2 2 sin2 θ
ga,M = − (dt − a sin θdφ) + dr + % dθ + (adt − (r2 + a2 )dφ)2 ,
%2 ∆ %2
where
∆ = r2 − 2M r + a2 , %2 = r2 + a2 cos2 θ.
Note that when a = 0 this expression reduces to (27). As with the Schwarzschild metric in the form (27), the
Kerr metric in the above form arises most naturally in a coordinate chart of a maximally extended globally
hyperbolic Lorentzian Kerr manifold. In contrast to the case of Schwarzschild, however, the maximally
extended globally hyperbolic Kerr manifold is smoothly extendible as a Lorentzian manifold, provided a 6= 0.
The geodesic incompleteness guaranteed by Theorem 13.9 thus has nothing to do in Kerr with any
singular behaviour, but rather a breakdown of global hyperbolicity. In view of the discussion in Section 13.3,
this means that the incompleteness in Kerr involves a breakdown in deterministic properties of equations
such as the vacuum Einstein equations (25). One may view this breakdown as being even worse for a theory
than the presence of the type of singular behaviour exhibited by the Schwarzschild solution.
It is conjectured that the situation in Kerr with a 6= 0 is not typical, and that the incompleteness
guaranteed by the Penrose Theorem is generically a result of some form of singularity (such as the singular
nature of the incompleteness in Schwarzschild). This conjecture, first formulated by Penrose, goes by the
name of the Strong Cosmic Censorship Conjecture, and constitutes one of the most important open problems
in classical general relativity.

13.7 Sketch proof of the Penrose Theorem

This sketch proof follows [4].
Let S be a closed trapped surface, in the sense of Definition 13.7. Recall the future directed null vector
fields L and L on S. At each p ∈ S, there is a unique geodesic maximal geodesic γp : (T− (p), T+ (p)) → M such

72
that γp (0) = p, γ̇p (0) = Lp . Similarly, there is a unique geodesic maximal geodesic γ p : (T − (p), T + (p)) → M
such that γ p (0) = p, γ̇ p (0) = Lp . For fixed s ≥ 0, define the sets

Ss = {γp (s) | p ∈ S}, S s = {γ p (s) | p ∈ S}.

Define also the sets

C(s) = {γp (s0 ) | p ∈ S, 0 ≤ s0 ≤ s}, C(s) = {γ p (s0 ) | p ∈ S, 0 ≤ s0 ≤ s},

and
C = {γp (s) | p ∈ S, 0 ≤ s ≤ T+ (p)}, C = {γ p (s) | p ∈ S, 0 ≤ s ≤ T + (p)},

Lemma 13.10 (The boundary of the causal future of a spacelike 2-sphere). If s is sufficiently small then
Ss and S s are spacelike 2-spheres and the geodesic congruences C(s) and C(s) are null hypersurfaces (with
boundary). Moreover,
∂J + (S) ⊂ C ∪ C. (29)
Note that the reverse inclusion of (29) is not true in general. Consider, for example, a standard 2-sphere
in Minkowski space.
The curves γp and γ p are called the null generators of C and C respectively. If s is such that C(s) and
C(s) are smooth null hypersurfaces, then the vector fields L and L extend to vector fields along C(s) and
C(s) respectively by
Lγp (s) = γ̇p (s), Lγ (s) = γ̇ p (s),
p

and therefore χ and χ extend to (0, 2) tensor fields on C and C respectively. Note that

∇L L = 0 on C, ∇L L = 0 on C.

Moreover there exists a unique extension of L to a future directed null vector field L on C(s), normal to the
spheres Ss , satisfying g(L, L) = −2, and a unique extension of L to a future directed null vector field L on
C(s), normal to the spheres S s , satisfying g(L, L) = −2.
The proof of Theorem 13.9 is based on understanding behaviour of certain Jacobi fields along the geodesics
γp and γ p . As opposed to conjugate points, which, for example, feature in the proof of Theorem 12.5, it is
slightly more convenient to introduce a related notion of focal points.7
Definition 13.11 (Focal points). Given p ∈ S, a point q = γp (s∗ ) for some s∗ ∈ (0, T+ (p)) is called a
focal point to p if there exists a nontrivial (non-identically vanishing) normal vector field J ∈ X(γp ) along
γp such that [L(s), J(s)] = 0 for all s ∈ [0, T+ (p)) and J(s∗ ) = 0. Similarly, a point q = γ p (s∗ ) for some
s∗ ∈ (0, T + (p)) is called a focal point to p if there exists a nontrivial normal vector field J ∈ X(γ p ) along γ p
such that [L(s), J(s)] = 0 for all s ∈ [0, T + (p)) and J(s∗ ) = 0.
Note that such a J is a Jacobi field along γp . Indeed, since ∇L J − ∇J L = [L, J] = 0,

∇L ∇L J = ∇L ∇J L = R(L, J)L,

since ∇L L = 0. Similarly for J.

Compare the following lemma to Theorem 11.12.
Lemma 13.12 (Null generators beyond focal points). Consider p ∈ S and let s∗ > 0 be such that γp (s∗ ) is
a focal point to p along γp . Then γp (s∗ ) ∈ I + (S). Similarly, if s∗ > 0 is such that γ p (s∗ ) is a focal point to
p along γ p . Then γ p (s∗ ) ∈ I + (S).
7 The distinction between conjugate points and focal points made here is non-standard. The terms are often used inter-

changeably to mean either of the two notions introduced here.

73
Proof. See Proposition 2 of [4].
As a consequence of Lemma 13.12, it in particular follows that, if s∗ > 0 is such that γp and γ p each
contain a focal point before time s∗ , for all p ∈ S, then (29) can be improved to

∂J + (S) ⊂ C(s∗ ) ∪ C(s∗ ). (30)

Let g/ denote the induced (Riemannian) metric on the spacelike 2-spheres Ss and S s , and define

1 1
χ̂ = χ − trχg/, χ̂ = χ − trχg/,
2 2
to be the trace free parts of the second fundamental forms χ and χ respectively.
Lemma 13.13 (Null mean curvatures satisfy Raychaudhuri equaions). The quantities trχ and trχ satisfy
the Raychaudhuri equations
1 1
L(trχ) = − (trχ)2 − |χ̂|2g/ − Ric(L, L), L(trχ) = − (trχ)2 − |χ̂|2g/ − Ric(L, L).
2 2
Proof. Consider some p ∈ S and a local frame e1 , e2 for S around p. For convenience, extend e1 , e2 along γp
by solving [L, eA ] = 0 for A = 1, 2. One then computes

L(χ(eA , eB )) = g(∇L ∇eA L, eB ) + g(∇eA L, ∇L eB ) = g(∇eA ∇L L, eB ) + g(∇eA L, ∇L eB ) + Rm(L, eA , L, eB ),

and so, since ∇L L = 0 and ∇eA L = χA C eC ,

(∇L χ)(eA , eB ) = L(χ(eA , eB )) − χ(∇L eA , eB ) − χ(eA , ∇L eB )

= g(∇eA L, ∇L eB )−Rm(eA , L, L, eB )−g(∇∇L eA L, eB )−g(∇eA L, ∇L eB ) = −χA C χCB −Rm(eA , L, L, eB ).

The first equation then follows by using the fact that g is compatible with ∇ and so
1
L(trχ) = g AB (∇L χ)(eA , eB ) = −χAB χAB − g AB Rm(eA , L, L, eB ) = − (trχ)2 − |χ̂|2g/ − Ric(L, L).
2
The second equation follows similarly.

Lemma 13.13 and the curvature assumption in Theorem 13.9 imply that
1
L(trχ) ≤ − (trχ)2 . (31)
2
In particular, trχ satisfies the remarkable monotonicity property

L(trχ) ≤ 0.

Recall that one views the assumption that trχ < 0 on S as meaning that the area of S can only decrease
when infinitesimally deformed in the direction of L. This monotonicity property, together with the fact that
S is a closed trapped surface, in particular implies that the area must continue to decrease as one deforms
further in the L direction.
The main analytic content of the proof of Theorem 13.9 is contained in the following proposition, which
is an easy application of the quantitative form (31) of the monotonicity property of trχ.
Proposition 13.14 (Occurrence of focal points). Under the assumptions of Theorem 13.9, if, for all p ∈ S,
T+ (p) and T + (p) are sufficiently large then, for all p ∈ S, the null generators γp and γ p both contain a focal
point to p.

74
Proof. Consider first the null generators γp . Since S is compact, there exists k > 0 such that

sup trχ(p) ≤ −k.

p∈S

The inequality (31) implies that

1
L(−(trχ)−1 ) ≤ − ,
2
and so, integrating along the integral curves of L,
1 s
−(trχ(γp (s)))−1 ≤ − ,
k 2
i.e.,
 −1
s 1
trχ(γp (s)) ≤ − , (32)
2 k
for all s ≥ 0. It follows that trχ(γp (s)) → −∞ before time s = k2 . The inequality (32) forms the main
content of the proof. Some extra work is required to check that this indeed leads to the existence of a focal
point.
Consider some p ∈ S. Let e1 , e2 be an orthonormal frame for Tp S and extend to an orthonormal frame
for Tγp (s) Ss for all s by solving
1
∇L eA = − g(∇eA L, L)L, A = 1, 2.
2
This choice ensures that L(g(eA , K)) = 0 for K = eB , L, L, so that e1 , e2 is an orthonormal frame for each
Tγp (s) Ss . Consider now an arbitrary vector v ∈ Tp S. Define a vector field J along γp by solving [L, J] = 0,
J(0) = v. One can check that
L(g(J, L)) = L(g(J, L)) = 0,
and so J(s) ∈ Tγp (s) Ss for all s. By the linearity of the equation [L, J] = 0 in J over R we can write
J A (s) = M A B (s)v B for some matrix M (s) (independent of v). Clearly M (0) = Id and so det M (0) = 1 and
det M (s) 6= 0 for small s. If there exists s such that det M (s) = 0 then there exists v ∈ Tp M with v 6= 0
such that J(s) = 0, i.e. to find a focal point it suffices to find a zero of det M . Now

L(J A ) = L(g(J, eA )) = g(∇L J, eA ) + g(J, ∇L eA ) = g(∇J L, eA ) − g(∇eA L, L)g(L, eA ) = χ(J, eA ),

and so
L(M A B (s))v B = χ(J(s), eA ) = χ(eA , eC )M C B v B .
Since v was arbitrary, it follows that

L(M A B (s)) = χ(eA , eC )M C B .

Using the well known identity

d det M  dM 
= det M tr M −1 ,
ds ds
it follows that
d log det M
= trχ.
ds
The inequality (32), together with the fact that log det M (0) = 0, then implies that
 2
s 1
det M (s) ≤ − ,
2 k

i.e. there exists a focal point along γp by time s = k2 .

The proof for the null generators γ p is identical, using now the Raychaudhuri equation for trχ.

75
 A set K ⊂ M is called a future set if p ∈ K implies that I + (p) ⊂ K. A set A ⊂ M is called achronal if
there is no pair p, q ∈ A such that q ∈ I + (q).
Lemma 13.15 (Topological boundary of a future set is achronal). If K ⊂ M is a future set, then its topo-
logical boundary ∂K ⊂ M is a closed achronal three dimensional embedded Lipschitz submanifold (without
boundary).

Proof. See Proposition 6.3.1 of [8].

The remaining ingredients of the proof of Theorem 13.9 consist of the following topological facts.
Proposition 13.16 (Topological facts).

1. If A is a compact topological space, B is a Hausdorff topological space and h : A → B is a continuous

bijection, then h is a homeomorphism.
2. If M is a (connected) topological manifold and N ⊂ M is a compact submanifold (without boundary)
such that dimN = dimM, then M is compact.
Proof. See Section 5.5 of [1].

The proof of Theorem 13.9 can now be completed.

Proof of Theorem 13.9. Suppose, for the sake of contradiction, that (M, g) is future geodesically complete.
The null generators γp and γ p are therefore future complete (i.e. T+ (p) = T + (p) = ∞ for all p ∈ S). By
Proposition 13.14 there therefore exists s∗ > 0 such that, for all p ∈ S, the null generators γp and γ p contain
a focal point to p by time s∗ . Lemma 13.12 then implies (see (30)) that

∂J + (S) ⊂ C(s∗ ) ∪ C(s∗ ).

Since C(s∗ ) ∪ C(s∗ ) is compact and ∂J + (S) is closed (the topological boundary of a set always being closed),
it follows that ∂J + (S) is compact.
Note that J + (S) is a future set, and hence ∂J + (S) is achronal by Lemma 13.15. Since (M, g) is time
oriented there exists a globally timelike vector field T . By assumption (M, g) admits a non-compact Cauchy
hypersurface H. For each q ∈ ∂J + (S) the integral curve of T through q intersects H exactly once and, since
∂J + (S) is achronal, does not intersect ∂J + (S) again. The map F : ∂J + (S) → H, where F (q) is defined
to be the unique point where the integral curve of T intersects H, is therefore a continuous injection. By
Proposition 13.16, ∂J + (S) is homeomorphic to F (∂J + (S)). Since ∂J + (S) is a compact three dimensional
submanifold of M (see Lemma 13.15) it follows that F (∂J + (S)) is a compact three dimensional topological
submanifold of the Cauchy hypersurface H. Proposition 13.16 then implies that H is compact, which
contradicts the assumption that H is non-compact. Hence (M, g) is geodesically incomplete.

References
[1] Aretakis, S., Lecture notes on general relativity, Columbia University.
[2] Berger, M., A panoramic view of Riemannian geometry.
[3] Boothby, W., An introduction to differentiable manifolds and Riemannian geometry.

[4] Christodoulou, D., Mathematical problems in general relativity II.

[5] Dafermos, M., Part III Differential geometry lecture notes.
[6] Do Carmo, M., Riemannian geometry.

76
 [7] Gallot, S., Hulin, D., Lafontaine, J., Riemannian geometry.
[8] Hawking, S., Ellis, G., The large scale structure of space-time.
[9] Jost, J., Riemannian geometry and geometric analysis.
[10] Lee, J., Riemannian manifolds.

[11] Lee, J., Introduction to smooth manifolds.

[12] O’Neill, B., Semi–Riemannian geometry with applications to relativity.
[13] Petersen, P., Riemannian geometry.

[14] Spivak, M., A comprehensive introduction to differential geometry, volumes 1–5.

77