Introduction to Riemannian Geometry

Course module NWI-WB045B

Radboud University Nijmegen

Spring semester 2022/23
(last updated on June 7, 2023)

Annegret Burtscher
 Contents

Preface vii

Chapter 1. Introduction 1
1.1. Classfication and local-to-global theorems 1
1.2. Intrinsic vs. extrinsic properties 1
1.3. Abstract Riemannian manifolds in higher dimensions 1
1.4. Applications 2

Chapter 2. Riemannian manifolds 5

2.1. Basic definitions 5
2.1.1. Inner products 5
2.1.2. Riemannian metrics 6
2.1.3. Semi-Riemannian metrics 8
2.1.4. Local representation of metrics 10
2.1.5. Riemannian submanifolds 11
2.2. Basic constructions 14
2.2.1. Raising and lowering indices 15
2.2.2. Volume form and integration 16
2.2.3. Differential operators 18
2.3. Lengths and Riemannian distance 20
2.3.1. Lengths of curves 20
2.3.2. Riemannian distance function 21
2.4. Symmetries of Riemannian manifolds 23
2.4.1. Isometries 24
2.4.2. Homogeneity and isotropy 25
2.4.3. Conformal transformations 28

Chapter 3. Connections and covariant differentation 31

3.1. Affine connections 31
3.2. Connections in vector bundles 35
3.3. Covariant differentiation of tensor fields 35
3.4. Levi-Civita connection 38
3.4.1. Metric connections 38
3.4.2. Symmetric connections 39
3.4.3. Fundamental Theorem of Riemannian Geometry 40
3.5. Parallel transport 42
3.5.1. Parallel vector and tensor fields 42
3.5.2. Vector and tensor fields along curves 43
3.5.3. Covariant derivative along curves 44
iii
iv CONTENTS

3.5.4. Parallel transport along curves 45

Chapter 4. Geodesics 51
4.1. Geodesics 51
4.1.1. Basic definitions and existence 51
4.1.2. Geodesic flow 54
4.1.3. Geodesics on submanifolds 55
4.2. Exponential map 56
4.2.1. Normal neighborhoods and normal coordinates 61
4.3. Minimizing curves 64
4.3.1. Minimizing curves are geodesics 65
4.3.2. Geodesics are locally minimizing 70
4.4. Completeness (not covered in the course) 76

Chapter 5. Curvature 79
5.1. Riemann curvature tensor 80
5.1.1. Motivation 80
5.1.2. Definition 81
5.1.3. Flatness 83
5.1.4. Symmetries 85
5.1.5. Expansion in normal coordinates 87
5.2. Sectional curvature 87
5.2.1. Definition 88
5.2.2. Relation to the Riemann curvature tensor 89
5.2.3. Constant sectional curvature 90
5.3. Ricci and scalar curvature 91
5.3.1. Definition and symmetries 91
5.3.2. Geometric interpretation of Ricci and scalar curvature 93
5.3.3. Traceless Ricci curvature and Einstein metrics 94
5.3.4. Curvature in General Relativity (not covered in the course) 96
5.4. Weyl curvature (not covered in the course) 98
5.4.1. Definition 98
5.4.2. Conformal flatness 100

Chapter 6. Extrinsic curvature of submanifolds 103

6.1. Second fundamental form 103
6.1.1. Definition 103
6.1.2. Gauss formula 104
6.1.3. Geometric interpretation of the second fundamental form 105
6.1.4. Gauss and Codazzi equations 106
6.2. Sectional curvature revisited 108
6.2.1. Geometric interpretation of sectional curvature 108
6.2.2. Constant sectional curvature and model spaces 109
6.3. Hypersurfaces 110
6.3.1. Scalar second fundamental form and shape operator 110
6.3.2. Principal curvatures 112
6.3.3. Minimal hypersurfaces 114
 CONTENTS v

Appendix A. Tensors and tensor fields 121

A.1. Vector bundles 121
A.2. Tensors on a vector space 123
A.3. Tensor fields on a manifold 124
Appendix B. Exterior derivatives 127
B.1. Differential of a function 127
B.2. Exterior derivative of a form 127
Bibliography 129
Index 133
 Preface

These lecture notes have been composed for the 3rd-year Bachelor course “Riemannian
Geometry” (NWI-WB045B) at Radboud University Nijmegen.
The lecture notes closely follow the structure of the book on Riemannian Geometry by
John Lee [36], which builds upon his earlier book [35] on smooth manifolds. The book [35]
was also used for the course “Manifolds” (NWI-WB079C) which is a prerequisite for this
course on Riemannian Geometry. Besides, there are many other excellent introductory books
(and lecture notes) about Riemannian Geometry that can be used as well, such as the classic
book by do Carmo [23], the comprehensive book by Petersen [45], the book by O’Neill [43]
with a broader perspective also on semi-Riemannian Geometry and the vast panorama by
Berger [7].
If you continue with a Master, a good follow-up course will be the MasterMath course on
Differential Geometry and certain courses in Mathematical Physics.
If you encounter any mistakes or typos please send me an email to [email protected].
Comments and suggestions are also welcome.
Nijmegen, January 2022
Annegret Burtscher

I am grateful to Max van Horssen for pointing out several typos.

Nijmegen, July 2022
Annegret Burtscher

vii
 CHAPTER 1

Introduction

Riemannian geometry is the branch of differential geometry, where the ancient geometric
objects such as length, angles, areas, volumes and curvature come back to life in a modern
reincarnation on smooth manifolds.
Read [36, Ch. 1] for an intuitive understanding of curvature and to understanding where
we are aiming at with this course. To catch a glimpse of typical results and properties we will
encounter in this course it helps to recall what we have learned about curves and surfaces in
the Euclidean plane and space (this is covered in the “Curves and surfaces” course).

1.1. Classfication and local-to-global theorems

The machinery used in smooth manifold theory is very technical (as you may know from
the “Manifolds” course) but Riemannian geometry is also about intuition. If you have at-
tended the course “Curves and Surfaces” you have a good idea of where we are going. In
short, we are extending the notion of curvature from two-dimensional surfaces in R3 to ar-
bitrary dimensions. This requires a great deal of machinery that we need to develop, but at
the end we will be rewarded with a deep understanding between local and global concepts
(local-to-global theorems) such as the relation of curvature to topology in the Gauss–Bonnet
Theorem, as well as a characterization of rigid shapes via curvature (classification theorems).

1.2. Intrinsic vs. extrinsic properties

When dealing with submanifolds it is crucial to destinguish between intrinsic and extrinsic
properties. Intrinsic properties do not depend on the ambient space (Gauss’s Theorema
Egregium tells us that the Gaussian curvature of surfaces in R3 is intrinsic) while extrinsic
ones do (such as the second fundamental form or mean curvature). See also Figure 1.1.

1.3. Abstract Riemannian manifolds in higher dimensions

Abstract Riemannian manifolds as we will encounter in this course are not per se embed-
ded in a higher-dimensional Euclidean space (although they can be, via Nash’s embedding
theorem). Instead, the geometry has to be described entirely by intrinsic tools. The most
important tool here are curves, and particularly geodesics, along which one can move effort-
lessly. Curves help us to “feel” curvature, more precisely, sectional curvature which is derived
from 2-planes cutting through a point, similar to what one does for surfaces in R3 . The man-
ifolds of constant sectional curvature are the usual Euclidean space, sphere, and hyperbolic
space, see Figure 1.2. Comparing these model spaces to other manifolds with lower or upper
(sectional or Ricci) curvature bounds will again reveal a lot about their global geometry and
topology.
1
2 1. INTRODUCTION

Figure 1.1. Plot of intrinsic Gaussian curvature kG versus extrinsic mean curvature
kM for surfaces in space (taken from [12, Fig. 1], where it is used for the classification
of folds in geology). Note that, for instance, a basin/valley is intrinsically the same as
a dome/mountain but from an extrinsic perspective they are fundamentally different.

1.4. Applications
Riemannian geometry appears in many areas of pure and applied mathematics (e.g.,
minimal surfaces, curvature flows optimizing shapes, learning) as well as in mathematical
physics (e.g., Einstein’s theory of relativity, noncommutative geometry). On the other hand,
it also makes use of knowledge from various fields, not only geometry but also calculus,
ordinary/partial differential equations and so on. It is beyond the scope of this introductory
course to deal with applications.
See Figure 1.3 that displays Eddington’s experiment carried in 1919 to verify Einstein’s
general theory of relativity which models the universe as a Lorentzian manifold, where matter
induces curvature and curvature determines how light and particles travel. Lorentzian geom-
etry is the indefinite twin of Riemannian geometry, where the scalar product is no longer
positive definite (a bit more about it in Section 2.1.3 below).
 1.4. APPLICATIONS 3

Figure 1.2. Model spaces of zero curvature (plane), constant positive (sphere), and
constant negative (hyperbolic space) curvature (taken from https://www.mu6.com/
riemann_space.html).

Figure 1.3. A sketch of Eddington’s experiment that, roughly put, measured that
light bends in the universe. It was spectacular news at the time (in 1919) and one
of the first experimental proofs of the general theory of relativity. Read more at
https://en.wikipedia.org/wiki/Eddington_experiment.
 CHAPTER 2

Riemannian manifolds

2.1. Basic definitions

We always consider manifolds are smooth, Hausdorff, second countable and connected.
Riemannian manifolds are smooth manifolds which are equipped with inner products on their
tangent spaces that vary smoothly. Tangent spaces are vector spaces. In this Section we
introduce the basic notions of inner product, Riemannian and semi-Riemannian metrics, and
discuss their local repentations, isometries and submanifolds.
2.1.1. Inner products. Inner products are crucial for Euclidean Geometry. On Rn the
dot product is given by
n
X
v·w = v i wi , v = (v 1 , . . . , v n ), w = (w1 , . . . , wn ) ∈ Rn , (2.1)
i=1
More generally, in Linear Algebra one considers inner products on vector spaces.
Definition 2.1. An inner product on V is a map ⟨., .⟩ : V ×V → R satisfying the following
properties for all v, w, x ∈ V and α, β ∈ R:
(i) symmetry: ⟨v, w⟩ = ⟨w, v⟩,
(ii) bilinearity: ⟨αv + βw, x⟩ = α⟨v, x⟩ + β⟨w, x⟩,
(iii) positive definiteness: ⟨v, v⟩ ≥ 0, and ⟨v, v⟩ = 0 ⇔ v = 0.
A vector space V endowed with an inner product ⟨., .⟩ is called an inner product space.
With an inner product at hand, one can define the length (or norm) of vector v by
1
|v| := ⟨v, v⟩ 2 .
The norm completely determines the inner product via the polarization identity
1
|v + w|2 − |v − w|2 ,


⟨v, w⟩ = v, w ∈ V. (2.2)
4
Exercise 2.2. Prove the polarisation identity (2.2).
The angle between two nonzero vectors v, w ∈ V is the unique θ ∈ [0, π] for which
⟨v, w⟩
cos θ = .
|v||w|
Two vectors v, w ∈ V are called orthogonal if
⟨v, w⟩ = 0,
π
that is, if θ = 2 or one of them is zero. If S is a linear subspace of V then
S ⊥ := {v ∈ V ; ⟨v, s⟩ = 0 for all s ∈ S}
is the orthogonal complement of S.
5
6 2. RIEMANNIAN MANIFOLDS

Exercise 2.3. Show that S ⊥ is also a linear subspace of V .

The vectors v1 , . . . , vk are orthonormal if they are of unit length and pairwise orthogonal,
that is, if
⟨vi , vj ⟩ = δij ,
where δij denotes the Kronecker delta. Recall that one can always construct an orthonormal
basis using the Gram–Schmidt orthogonalization.
Proposition 2.4 (Gram–Schmidt orthogonalization). Let V be an n-dimensional vec-
tor space with inner product ⟨., .⟩ and with ordered basis (v1 , . . . , vn ). Then there exists an
orthonormal basis (b1 , . . . , bn ) of V such that
span(v1 , . . . , vk ) = span(b1 , . . . , bk ), k = 1, . . . , n.
Proof. It is easy to see that the recursively defined vectors
v1
b1 := ,
|v1 |
vj − j−1
P
i=1 ⟨vj , bi ⟩bi
bj := Pj−1 , 2 ≤ j ≤ n,
|vj − i=1 ⟨vj , bi ⟩bi |
have the desired properties. □
Definition 2.5. Let V and W be two vector spaces equipped with inner products ⟨., .⟩V
and ⟨., .⟩W . A map F : V → W is called a linear isometry if it preserves the inner product,
that is,
⟨F (v), F (w)⟩W = ⟨v, w⟩V , v, w ∈ V.

Exercise 2.6. Show that a linear isometry F : V → W is a linear map.

By mapping one orthonormal basis to another, we immediately obtain the following.
Proposition 2.7. All inner product spaces of the same finite dimension are linearly
isometric. □
2.1.2. Riemannian metrics. On a manifold we define inner products in each tangent
space. We assume that a manifold is always Hausdorff and second countable.
Definition 2.8. Let M be a smooth manifold. A Riemannian metrics on M is a smooth
covariant 2-tensor field g ∈ T 2 (M ) whose value gp at every p ∈ M is an inner product on
Tp M .
The pair (M, g) is called a Riemannian manifold .
Example 2.9 (Euclidean space). The simplest and most important Riemannian manifold
is Euclidean space En = (Rn , ḡ). Due to the tangent bundle identification Tp Rn ∼
= Rn × Rn
n
the standard dot product (2.1) induces a Riemannian metric on R pointwise by
ḡp (v, w) = v · w, v, w ∈ Tp Rn ∼
= Rn , p ∈ Rn .
The following fundamental result distinguishes Riemannian structures from semi-Riemannian
ones.
Proposition 2.10. Every smooth manifold admits a Riemannian metric.
 2.1. BASIC DEFINITIONS 7

Proof. Cover M with chart neighborhoods and let {χα }α∈A be a subordinate partition
of unity. For each α ∈ A chose some chart ((x1 , . . . , xn ), U ) such that supp χα ⊆ U . On U
define X
gα := dxi ⊗ dxi .
i
Since a positive linear combination of inner products is again a (positive definite) inner prod-
uct, the expression X
g := χα gα
α∈A
is a Riemannian metric on M . □
If M is a smooth manifold with boundary then (M, g) is called a Riemannian manifold
with boundary. Most proofs will work as well also for manifolds with boundary, but occassional
difficulties may arise in the proof. In this course we are not so much interested in this case,
so we will not specifically look at these situations.
As in Section 2.1.1 we can define the lengths of tangent vectors v ∈ Tp M via |v|g :=
1/2
⟨v, v⟩g , and also angles and orthogonality in the same way. If the Riemannian metric g is
fixed we occassionally also just write ⟨v, w⟩ and |v| without the index g.
Isometries. A Riemannian structure is preserved by isometries.
Definition 2.11. Let (M, g) and (M f, ge) be Riemannian manifolds. An isometry from
f, ge) is a diffeomorphism φ : M → M
(M, g) to (M f such that1 φ∗ ge = g, that is,
gp (v, w) = geφ(p) (dφp (v), dφp (w)), v, w ∈ Tp M, p ∈ M.
We say that (M, g) and (M
f, ge) are isometric if there exists an isometry between them.

Being an isometry φ : (M, g) → (M f, ge) is equivalent to φ being a smooth bijection and

each differential dφp : Tp M → Tφ(p) Mf being a linear isometry. Global isometries are crucial
for Riemannian Geometry, but also local isometries can provide important information.
Definition 2.12. A map φ : M → M f is a local isometry between Riemannian manifold
(M, g) and (M , ge) if for each point p ∈ M there exists a neighborhood U such that φ|U is an
f
isometry onto an open subset in M f.

Exercise 2.13. Prove that if (M, g) and (M

f, ge) are Riemannian manifolds of the same
dimension, then a smooth map φ : M → Mf is a local isometry if and only if φ∗ ge = g.
A Riemannian manifold is said to be flat if it is locally isometric to Euclidean space.
Problem 2.14. Show that every Riemannian 1-manifold is flat.
An important result in connection with isometries is the Nash Embedding Theorem from
1954 [40, 41] which states that every Riemannian manifold can be isometrically embedded
into some (higher dimensional) Euclidean space. Recall that every n-dimensional smooth
manifold can be embedded into R2n already by the Whitney Embedding Theorem, but this
result tells us nothing about Riemannian structures.
The set of isometries of a Riemannian manifold to itself forms a group with respect to
composition. More about this group and important examples in Section 2.4.1.
1Recall that φ∗ g
e = g denotes the pullback of a tensor field by φ.
8 2. RIEMANNIAN MANIFOLDS

2.1.3. Semi-Riemannian metrics. Semi-Riemannian (in particular, Lorentzian) man-

ifolds are crucial in Mathematical Physics, specifically in General Relativity. While we will
not study semi-Riemannian manifolds in this course explicitely, it is good to know that many
basic constructions, in fact, carry over verbatim to this setting. For a short overview of the
basics see the book of Lee [36, p. 40–45], for an extensive treatment look at the books of
O’Neill [43] and Beem, Ehrlich and Easley [6].
In place of inner products we use certain scalar products.
Definition 2.15. Let V be a finite-dimensional vector space over R. A symmetric bilinear
form b : V × V → R that is nondegenerate, i.e.,
b(v, w) = 0 for all w ⇒ v = 0,
is called a scalar product.
A symmetric bilinear form b is called positive (or negative) semidefinite if b(v, v) ≥ 0
(or ≤ 0) for all v ∈ V . The bilinear form b is neither definite or semidefinite (positive or
negative), it is called indefinite.
Exercise 2.16. Show that a symmetric bilinear form is positive definite if and only if it
is positive semi-definite and nondegenerate.

Exercise 2.17. Show that g(v, w) = −v 1 w1 + v 2 w2 is a scalar product on R2 .

Clearly, if b is symmetric bilinear form on V and W is a linear subspace of V , then also
b|W is a symmetric bilinear form. Similarly, if b is (semi-)definite, then so is b|W . This gives
rise to the following definition.
Definition 2.18. The index of a symmetric bilinear form b on V is
ind V := ν := max{dim S; S is a subspace of V such that b|S is negative definite}.
Thus 0 ≤ ν ≤ dim V , where ν = 0 if and only if b is positive semi-definite.
Orthogonality is defined in the same way, but note that for indefinite scalar products
orthogonal vectors no longer need to have right angles. Moreover, the orthogonal subspace
S ⊥ need not be a complement of V since, in general, S + S ⊥ ̸= V .
Exercise 2.19. Show that in Exercise 2.17 we have for w = (1, 1) and S = span({w})
that S = S ⊥ ⊊ R2 .

Exercise 2.20. Let S be a subspace of V . Then

(i) dim S + dim S ⊥ = dim V ,
(ii) (S ⊥ )⊥ = S.
For indefinite scalar products there always exist degenerate subspaces (e.g., the span of
a null vector as in Exercise 2.19), while for inner products every subspace is again an inner
product space. Using basic Linear Algebra one can prove the following result.
Lemma 2.21. Let g a scalar product on V and let S be a subspace of V . The following
are equivalent:
(i) S is nondegenerate, meaning that g|S is nondegenerate.
(ii) S ⊥ is nondegenerate.
 2.1. BASIC DEFINITIONS 9

(iii) V = S ⊕ S ⊥ (direct sum).

Problem 2.22. Prove Lemma 2.21.

As in Section 2.1.1 one can show that a nontrivial vector space with scalar product
possesses an orthonormal basis. This implies that for a nondegenerate subspace S of V we
have
ind V = ind S + ind S ⊥ .
Note that in order to have a linear isometry between vector spaces with scalar products
both dimension and index have to agree!
A semi-Riemannian manifold is defined in analogy to a Riemannian manifold with scalar
products on the tangent spaces.
Definition 2.23. Let M be a smooth manifold. A semi-Riemannian metric on M is a
symmetric nondegenerate covariant 2-tensor field g ∈ T 2 (M ) with constant index ν (called
the index of M ).
A semi-Riemannian manifold (or pseudo-Riemannian manifold) (M, g) is a smooth man-
ifold M endowed with a semi-Riemannian metric g.
We have 0 ≤ ν ≤ n = dim M . If ν = 0 then (M, g) is a Riemannian manifold. If ν = 1 and
n ≥ 2 then (M, g) is called a Lorentzian 2 manifold and g is a Lorentzian metric. Lorentzian
Geometry is the geometric framework needed to understand General Relativity.
Example 2.24 (Minkowski space). Let 0 ≤ ν ≤ n. Then
ν
X n
X
q̄(v, w) := ⟨v, w⟩ := − v i wi + v i wi
i=1 i=ν+1

defines a metric tensor of index ν on Rn , often denoted by Rnν or Rν,n−ν . If ν = 1 then R1,n−1
is called the Lorentzian vector space (or Minkowski space) of dimension n and the scalar
product denoted by η̄. If n = 4, R1,3 it is the physically relevant Minkowski space, which is
also a spacetime (time-oriented Lorentzian manifold) and the simplest solution of the vacuum
Einstein equations in General Relativity.
One of the most notable differences between Riemannian and Lorentzian metrics is that
while every manifold admits a Riemannian metric (see Proposition 2.10) not every manifold
admits a semi-Riemannian or even Lorentzian metric. A problem occurs, in particular, for
compact manifolds.
Theorem 2.25 ([43, p. 149]). For a smooth manifold M the following are equivalent:
(i) existence of a Lorentzian metric,
(ii) existence of a continuous line field,
(iii) existence of a nonvanishing continuous vector field,
(iv) existence of a spacetime structure,
(v) either M is noncompact, or M is compact and has zero Euler characteristic.
We will not prove this entire result, but you can look at the following part.
2Hendrik Lorentz (1853–1928, born in Arnhem) was an eminent and influental Dutch physicist who won
the nobel prize in physics in 1902 with Pieter Zeeman for the Zeemann effect. He also worked extensively in
Special and General Relativity.
10 2. RIEMANNIAN MANIFOLDS

Problem 2.26. A smooth manifold M admits a Lorentzian metric if and only if it admits
a rank-1 distribution, i.e., a rank-1 subbundle of T M (see hints in the book of Lee [36, p. 54,
Problem 2-34]).

2.1.4. Local representation of metrics. We discuss the expression of a Riemannian

metric in coordinates and frames. Let (M, g) be a Riemannian manifold (with or without
boundary).
Coordinates. If (x1 , . . . , xn ) are coordinates on an upen set U ⊆ M , then we can write g
locally in U as3
g = gij dxi ⊗ dxj ,
where the gij are n2 smooth functions given by
gij (p) = ⟨∂i |p , ∂j |p ⟩, p ∈ U,
where ∂i = ∂/∂xi is the i-th coordinate vector field. The matrix (gij (p))ij is nonsingular and
symmetric in i and j, which is why we can also drop the ⊗ and simply write
g = gij dxi dxj .
Example 2.27. The Euclidean metric on Rn in standard coordinates can be expressed in
the following ways X X
ḡ = dxi dxi = (dxi )2 = δij dxi dxj .
i

Frames. A local frame for an n-dimensional smooth manifold M is an ordered n-tuple of

vector fields (E1 , . . . , En ) defined on an open set U ⊆ M such that for each p ∈ U the vectors
(E1 |p , . . . , En |p) are linearly independent and span Tp M . For example, the coordinate vector
∂
fields ( ∂x i ) define a smooth frame on a coordinate nieghborhood U , but often frames that
are adapted to a particular problem at hand are more useful. On Riemannian manifolds we
usually use orthonormal frames.
Generally, if (E1 , . . . , En ) is any smooth local frame for T M on an open set U ⊆ M and
(ε1 , . . . , εn ) is its dual coframe, then we can locally write the metric tensor g as
g = gij εi εj ,
where gij (p) = ⟨Ei |p , Ej |p ⟩. Again, the functions (gij ) are smooth and symmetric.
For smooth vector fields X, Y ∈ X(M ), we can thus locally write ⟨X, Y ⟩ = gij X i Y j and
|X| = ⟨X, X⟩1/2 , which is continuous and smooth on the open set {X ̸= 0}.
A particularly useful frame is an orthonormal frame, which is a local frame (Ei ) on U ⊆ M
such that E1 |p , . . . En |p form an orthonormal basis on Tp M at each p ∈ U , that is,
⟨Ei , Ej ⟩ = δij .
In this frame g is locally
g = (ε1 )2 + . . . (εn )2 ,
where (εi )2 = εi εi = εi ⊗ εi as before.
Thanks to the Gram–Schmidt algorithm there exists an orthonormal frame locally around
every point [36, Prop. 2.8]. Note that while one can use any coordinate frame (∂i ) to construct
3Using the Einstein summation convention, which drops the P sign and replaces it by upper and lower
Pn Pn
indices, automatically indicating the summation. Here, gij dxi ⊗ dxj actually means i=1 j=1 gij dxi ⊗ dxj .
 2.1. BASIC DEFINITIONS 11

an orthonormal frame, one can only find an orthonormal coordinate frame if the metric is
flat, i.e., locally isometric to Euclidean space (see [36, Ch. 7]).

2.1.5. Riemannian submanifolds. Generally, submanifolds can be obtained via im-

mersions4 (or embeddings). For Riemannian manifolds we also want to preserve the Riemann-
ian structure.
f, ge) be a Riemannian manifold, M be a smooth manifold and F : M →
Lemma 2.28. Let (M
f be a smooth map. The smooth 2-tensor field g := F ∗ ge, i.e.,
M
gp (v, w) := (F ∗ ge)p (v, w) = geF (p) (dFp (v), dFp (w)), v, w ∈ Tp M, p ∈ M
is a Riemannian metric on M if and only if F is an immersion.

Exercise 2.29. Prove Lemma 2.28.

An immersion (or embedding) F : M → M f between two Riemannian manifolds (M, g)
f, ge) that satisfies g = F ∗ ge is called an isometric immersion (or isometric embedding).
and (M
Because of Lemma 2.28 we can equip submanifolds with a natural Riemannian structure
coming from the ambient Riemannian manifold.
Definition 2.30. Let (M f, ge) be a Riemannian manifold. Suppose M ⊆ M
f is an (im-
mersed or embedded) submanifold. The induced Riemannian metric on M is the pullback
metric g = ι∗ ge induced by the inclusion ι : M ,→ M
f.

We automatically assume that submanifolds are endowed with the induced metrics. Many
examples are of this kind.
Example 2.31 (Sphere). An important example of a submanifold is the n-dimensional
sphere of radius R > 0, defined by
Sn (R) := {x ∈ Rn+1 ; |x| = R}
The metric induced from the embedding Sn (R) ,→ (Rn+1 , ḡ) is the canonical metric on Sn (R).
It is called the round metric of radius R and denoted by g̊R . The unit sphere is Sn := Sn (1)
with round metric g̊.
By Hilbert’s Theorem from 1901 the hyperbolic plane H2 cannot be isometrically im-
mersed in E3 , but one could do it in E4 . We can, however, in analogy to the sphere introduce
hyperbolic space as a codimension-1 submanifold of Minkowski space R1,n .
Example 2.32 (Hyperbolic space). Let n > 1 and fix R > 0. The hyperbolic space 5 Hn (R)
of radius R is the submanifold of the Minkowski space (R1,n , η̄) (see Example 2.24), defined as
the “upper sheet” (by using the standard unit timelike vector e0 = (1, 0, . . . , 0), or in standard

4Recall that a map F : M → M

f is an immersion if its differential dFp : Tp M → TF (p) M
f is injective
at every point p ∈ M (equivalently, rank F = dim M ). An embedding F : M ,→ M f is an immersion that
is homeomorphic to its image. Therefore, immersed submanifolds may have selfintersections but embedded
submanifolds don’t.
5This definition of hyperbolic space, called the “hyperboloidal model”, is particularly useful to understand
its symmetries, see Section 2.4.1. There are several isometric models of hyperbolic space [36, Thm. 3.7] which
are used in other situations.
12 2. RIEMANNIAN MANIFOLDS

coordinates (τ, ξ 1 , . . . , ξ n ) in terms of the sign of τ ) of the two-sheeted hyperboloid in R1,n ,

that is,
Hn (R) := {x ∈ R1,n ; η̄(x, x) = −R2 and η̄(x, e0 ) < 0}
= {(τ, ξ) ∈ R1,n ; τ 2 − |ξ|2 = R2 } ∩ {τ > 0},
with the pullback metric ğR induced by Minkowski space (R1,n , η̄), called the hyperbolic met-
ric. Hyperbolic space is Hn := Hn (1) with metric ğ. If n = 2 we call H2 the hyperbolic
plane.
Many computations on n-dimensional submanifolds M ,→ M f are carried out using an
adapted orthonormal frame (an orthonormal frame (E1 , . . . , En , En+1 , . . . , Em ) such that the
first n vector fields are tangent to M , which exists by [36, Prop. 2.14]) or smooth local
parametrizations 6 (the inverse of charts).
If M is an immersed n-dimensional submanifold of Rm and X : U → Rm is a local
parametrization of M , then the induced metric on U is
m m n m n
∗
X
i 2
X X ∂X i j
2 X X ∂X i ∂X i j k
g = X ḡ = (dX ) = du = du du . (2.3)
∂uj ∂uj ∂uk
i=1 i=1 j=1 i=1 j,k=1

Figure 2.1. An immersed 2-dimensional submanifold M of R3 , given by as local

parametrization X : U → R3 . In fact, M is a graph, and X could be the graph
parametrization X = (u1 , u2 , f ).

Example 2.33 (Surfaces in R3 ). For surfaces in R3 the expression (2.3) is the first
∂X 2
fundamental form (or line element), usually given in terms of the coefficients E = | ∂u1| ,
∂X ∂X ∂X 2
F = ∂u1 · ∂u2 , and G = | ∂u2 | (studied in the “Curves and Surfaces” course).
6A smooth local parametrization is a smooth map X : U → M
f, where U is an open subset of Rn and X is
a diffeomorphism onto its image X(U ) ⊆ M . Note that (V = X(U ), φ = X −1 ) is a smooth coordinate chart
on M .
 2.1. BASIC DEFINITIONS 13

We can use the expression (2.3) to compute metrics induced on submanifolds of Rm in

various coordinates. Important examples are graphs and surfaces of revolution for which
special parametrizations and coordinates adapted to the setting in hand to are used.
Example 2.34 (Graphs). For U an open subset of Rn and f : U → R a smooth function
the graph of f ,
Γ(f ) := {(y, f (y)); y ∈ U } ⊆ Rn+1
is an embedded submanifold of dimension n. The global parametrization X : U → Rn+1 is
called a graph parametrization and the corresponding coordinates (u1 , . . . , un ) on Γ(f ) are
the graph coordinates. In graph coordinates, the induced metric on Γ(f ) is
X ∗ ḡ = X ∗ ((dx1 )2 + (dxn+1 )2 ) = (du1 )2 + . . . + (dun )2 + df 2 .

Exercise 2.35. Compute the round metric of the upper hemisphere √ of S2 with respect
to the graph parametrization X : B2 → S3 , given by X(u, v) = (u, v, 1 − u2 − v 2 ).

Example 2.36 (Surfaces of revolution). Let C be an embedded 1-dimensional submanifold

in the half plane H = {(r, z); r > 0}. The set C is used as generating curve for the surface of
revolution p
SC := {(x, y, z) ∈ R3 ; ( x2 + y 2 , z) ∈ C}.
Every smooth local parametrization γ(t) = (a(t), b(t)) of C yields a smooth local parametriza-
tion
X(t, θ) = (a(t) sin θ, a(t) cos θ, b(t)).
for SC . The induced metric on SC is given by (2.3) and thus
X ∗ ḡ = d(a(t) cos θ)2 + d(a(t) sin θ)2 + d(b(t))2
= (a′ (t)2 + b′ (t)2 )dt2 + a(t)2 dθ2 .

Figure 2.2. Generating curve C in the half plane H for a surface of revolution in R3 .
14 2. RIEMANNIAN MANIFOLDS

Exercise 2.37. Consider a torus T2 of revolution obtained by rotating the circle C =

{(r, z); (r − 2)2 + z 2 = 1}. Assume that C is parametrized by γ(t) = (2 + cos t, sin t) and show
that the induced metric is dt2 + (2 + cos t)2 dθ2 .
Alternatively, the Riemannian metric induced on the torus Tn = S 1 × . . . × S 1 can also
be computed via its product structure. We immediately treat a generalization of Riemannian
products.
Definition 2.38. Let (M1 , g1 ) and (M2 , g2 ) be two Riemannian manifolds, and f : M1 →
(0, ∞) be a smooth function. The warped product M1 ×f M2 is the product manifold M1 × M2
endowed with the Riemannian metric g = g1 ⊕ f 2 g2 , defined by
g(p1 ,p2 ) ((v1 , v2 ), (w1 , w2 )) = g1 |p1 (v1 , w1 ) + f (p1 )2 g2 |p2 (v2 , w2 ),
where (v1 , v2 ), (w1 , w2 ) ∈ Tp1 M1 ⊕ Tp2 M2 .
Many examples can be constructed using warped products.
Problem 2.39. Show that metrics induced on surfaces of revolution are isometric to
warped products metrics (see [36, Ex. 2.24(b) and Problem 2-3]).

Exercise 2.40. Let g0 and g1 be two Riemannian metrics on M . Then, for any number
λ ∈ [0, 1], the convex combination
g := λg0 + (1 − λ)g1
is also a Riemannian metric on M .
While it is straightforward to induce Riemannian metrics on submanifolds and products
this is, for instance, not the case for quotients. Submersions yield Riemannian structures only
in very special cases.

Definition 2.41. Suppose that (M f→

f, ge) and (M, g) are Riemannian manifolds and π : M
M is a smooth submersion7. Then π is said to be a Riemannian submersion if for every p ∈ M f
⊥
the differential dπp : Hx → Tπ(p) M is a linear isometry, where Hx = (Ker dπp ) denotes the
horizontal tangent space at p. In other words, ge = π ∗ g on Hx .
Example 2.42. If M ×f N is a warped product, then the projection πM : M × N → M
is a Riemannian submersion but πN generally not.
More about submersions in [36, p. 21–24].

2.2. Basic constructions

Riemannian manifolds come with canonical notions such as an associated volume form,
various differential operators, and an induced metric space structure. First we need to un-
derstand how tensors transform in a way that respects the Riemannian structure (see also
[45, Sec. 1.5]).

7Recall that a smooth map F : M f → M is called a smooth submersion if its differential dF is surjective
at every point (equivalently, rank F = dim M ).
 2.2. BASIC CONSTRUCTIONS 15

2.2.1. Raising and lowering indices. An (s, t)-tensor T is a section on the bundle
. . ⊗ T M} ⊗ |T ∗ M ⊗ .{z
| M ⊗ .{z
T . . ⊗ T ∗ M}
s times t times
Given a Riemannian metric g on M , we can turn T into an (s − k, t + k)-tensor in a canonical
way because T M is naturally isomorphic to T ∗ M . In what follows we make this connection
precise.
Let us make this idea more precise. For smooth vector fields X, Y ∈ X(M ) the map
ĝ(X)(Y ) := g(X, Y )
is C ∞ (M )-linear
in Y and thus ĝ(X) a smooth 1-form8 by the Tensor Characterization
Lemma A.17. Moreover, ĝ(X) is also C ∞ (M )-linear in X and thus ĝ a smooth bundle
homomorphism (often just denoted by g)
ĝ : T M → T ∗ M.

Suppose (Ei ) is smooth local frame on M and (εi ) its dual coframe. Then X = X i Ei ∈
X(M ) and
ĝ(X) = (gij X i )εj = Xj εj ∈ Ω1 (M ),
where we have obtained the components Xj = gij X i of ĝ(X) from X by lowering an index
(sometimes ĝ(X) is called X flat and denoted by X ♭ ).
Since ĝ = (gij ) is invertible at every p the matrix of ĝ −1 is the inverse matrix of (gij ). We
write ĝ −1 = (g ij ) so that g ij gjk = δki . Thus
ĝ −1 (ω) = ω i Ei ∈ X(M )
for ω i = g ij ωj . Thus the vector field ĝ −1 (ω) is obtained from the 1-form ω ∈ Ω1 (M ) by
raising an index (sometimes ĝ −1 (ω) is called ω sharp and denoted by ω ♯ ).
The mutually inverse isomorphisms ♯ = ĝ −1 : T ∗ M → T M and ♭ = ĝ : T M → T ∗ M
are also called the musical isomorphisms, and we call the (co)vectors obtained in this way
metrically equivalent. An important application is the following definition of the gradient9.
Definition 2.43. Let (M, g) be a Riemannian manifold and f : M → R smooth. The
gradient of f is the vector field
grad f := (df )♯ .

Exercise 2.44. Show that grad f is characterized by

dfp (w) = ⟨grad f |p , w⟩, p ∈ M, w ∈ Tp M.
If (Ei ) is a local smooth frame then grad f = (g ij Ei f )Ej (in particular, if (Ei ) is an orthonor-
mal frame then the components of grad f and df agree).

Problem 2.45. A level set f −1 (c) of a smooth function f : M → R is called regular if

every point p ∈ f −1 (c) satisfies dfp ̸= 0. Prove that the gradient is orthogonal to regular level
sets (see [36, Prop. 2.37]).

8Lee [35, 36] prefers to call 1-forms covector fields.

9Clearly, the gradient therefore depends on the choice of a Riemannian metric, but without the use of
a Riemannian metric (and the corresponding musical isomorphism) there really is no way to define gradient
vector field in an invariant way. See the discussion in [45, Sec. 2.1.1] for why this is.
16 2. RIEMANNIAN MANIFOLDS

Problem 2.46. Suppose (M, g) is a Riemannian manifold, f ∈ C ∞ (M ), and X ∈ X(M )

is a nowhere-vanishing vector field. Prove that X = grad f if and only if Xf ≡ |X|2g and X
is orthogonal to the level sets of f at all regular points.
We initially set out to apply the flat and sharp operators to tensors T which are locally
of the form
T = Tji11...j
...is
t
Ei1 ⊗ . . . ⊗ Eis ⊗ εj1 ⊗ . . . ⊗ εjt
It is now clear how to do this using the musical isomorphisms, but it is important to keep
track of the index that is lowered or raised (see [36, p. 27]).
Example 2.47. A 3-tensor A of type (2, 1) given in a local frame by
A = Ai j k εi ⊗ Ej ⊗ εk
can be turned into the covariant 3-tensor A♭ with components
Aijk = gjl Ai l k .
One important application of raising and lowering indices are contractions, which lowers
the rank of a tensor by 2. Contractions are traces of tensors. Given a covariant 2-tensor field
h = hij εi ⊗ εj , we first obtain the (1, 1)-tensor field h♯ and define the trace of h with respect
to g as
trg h := tr(h♯ ) = tr(g jk hij εi ⊗ Ek ) = g ij hij = hi i .
On a Riemannian manifold (M, g) the musical isomorphisms can be further used to carry
the inner product on Tp M over to covectors in Tp∗ M via
⟨ω, η⟩g := ⟨ω ♯ , η ♯ ⟩g , ω, η ∈ Ω1 (M ).
One can extend this construction to tensor bundles of any rank in the obvious way.
Exercise 2.48. Show that in coordinates ⟨ω, η⟩ = g ij ωi ηj = ω j ηj .

Exercise 2.49. Let (M, g) be a Riemannian manifold with or without boundary, let
(Ei ) be a local frame for M , and let (εi ) be its dual coframe. Show that the following are
equivalent:
(i) (Ei ) are orthonormal.
(ii) (εi ) are orthonormal.
(iii) (εi )♯ = Ei for each i.

2.2.2. Volume form and integration. A Riemannian metric g induces a canonical

volume form on an oriented manifold M . This follows from the usual signed volume of
the volume of a parallelepiped in Rn spanned by vectors (v1 , . . . , vn ). If (e1 , . . . , en ) is the
canonical basis then the volume of the parallelepiped is
Vol(v1 , . . . , vn ) = det[vi · ej ] = det([v1 , . . . , vn ][e1 , . . . , en ]T ) = det[v1 , . . . , vn ].
The same formula also holds for any orthonormal basis, and via the tangent space Tp M ∼
= Rn
it directly translates to M .
Proposition 2.50 (Riemannian volume form). Let (M, g) be an oriented Riemannian
manifold. There is a unique n-form dVg on M , called the Riemannian volume form, given
with respect to an local oriented orthonormal frame (E1 , . . . , En ) for T M by
dVg (E1 , . . . , En ) = 1. (2.4)
 2.2. BASIC CONSTRUCTIONS 17

Proof. Suppose first that the volume form dVg exists. For an local oriented orthonormal
frame (E1 , . . . , En ) and (ε1 , . . . , εn ) its dual coframe the volume form10 is
dVg = f ε1 ∧ . . . ∧ εn .
Because of (2.4) we must have f = 1 (the volume of the parallelepiped is 1), and hence
uniqueness.
To prove existence, we simply define dVg in a neighborhood of each point by
dVg := ε1 ∧ . . . ∧ εn .
It remains to check that this definition is independent of the choice of oriented coframe. Let
(E en ) is another oriented orthonormal frame and and (ε̃1 , . . . , ε̃n ) its dual coframe. We
e1 , . . . , E
can write
ei = Aj Ej
E i
for some transition matrix A = (Aji ) of smooth functions, and hence
ε1 ∧ . . . ∧ εn = det A ε̃1 ∧ . . . ∧ ε̃n
Since both frames are orthonormal we have that A(p) ∈ O(n) for each p, hence det A = ±1.
A consistent orientation forces a positive sign, hence det A = 1 and
dVg (E
e1 , . . . , E ei )) = det(Aj ) = 1.
en ) = det(εj (E □
i

Exercise 2.51. Prove that, with respect to any local oriented coordinates (x1 , . . . , xn )
we have q
dVg = det(gij )dx1 ∧ . . . ∧ dxn .

The Riemannian volume form allows to integrate functions over oriented Riemannian
manifolds.
Definition 2.52. Let (M, g) be an oriented Riemannian manifold and let f : M → R be
continuous and compactly supported function, then
ˆ
f dVg
M
is called the integral of f over M .
Definition 2.53. If (M, g) is a compact oriented Riemannian manifold, then the volume
of M is defined by ˆ ˆ
Vol(M ) := dVg = 1 dVg .
M M

If D ⊆ M is a regular domain (a closed, embedded codimension-0 submanifold with

boundary), the same definitions apply with respect to the induced Riemannian metric on D.

Exercise 2.54. Suppose (M, g) and (M f, ge) are oriented Riemannian manifolds, and
f is an orientation-preserving isometry. Prove that φ∗ dVge = dVg .
φ: M → M

10The notation dV with the d is commonly used but slightly misleading, because dV is not necessarily
g g
an exact form.
18 2. RIEMANNIAN MANIFOLDS

Problem 2.55. Suppose M is a hypersurface in an oriented Riemannian manifold (M f, ge)

and g is the induced metric on M . Then M is orientable if and only if there exists a global
unit normal vector field N for M , and in that case the volume form of (M, g) is given by
dVg = (N ⌟dVge)|M
(see page 18 for the definition of the interior multiplication ⌟).
Problem 2.56. Suppose (M1 , g1 ) and (M2 , g2 ) are oriented Riemannian manifolds of
dimensions k1 and k2 , respectively. Let f : M1 → (0, ∞) be a smooth function, and let
g = g1 ⊕ f 2 g2 be the corresponding warped product metric on M1 ×f M2 . Prove that the
Riemannian volume form of g is given by
dVg = f k2 dVg1 ∧ dVg2 ,
where f , dVg1 and dVg2 are understood to be pulled back to M1 × M2 by the projection maps.
Recall that on nonorientable manifolds we can compute integrals of functions using den-
sities instead of differential forms (for more background about densities see [35, p. 427–434]).
We obtain a canonical Riemannian density as well.
Proposition 2.57 (Riemannian density). If (M, g) is a Riemannian manifold, then there
exists a unique smooth positive density µg on M , called the Riemannian density, with the
property that
µg (E1 , . . . , En ) = 1
for every local orthonormal frame (E1 , . . . , En ).
Exercise 2.58. Prove Proposition 2.57 by showing that µg can be defined in terms of
any local orthonormal frame by
µg = |ε1 ∧ . . . ∧ εn |,
where |ω|(v1 , . . . , vn ) := |ω(v1 , . . . , vn )| for an n-form ω ∈ Ωn (Tp∗ M ).

Exercise 2.59. Show that if (M, g) is an oriented Riemannian manifold then µg = |dVg |.
For any compactly suppported continuous function f : M → R we have
ˆ ˆ
f µg = f dVg
M M
(this is why the Riemannnian density is often also denoted by dVg ).

2.2.3. Differential operators. The most important differential operators on Riemann-

ian manifolds are the gradient, divergence and Laplacian. We have already introduced the
gradient in Definition 2.43 using the ♯ isomorphism which raises an index. The divergence
and Laplacian can be introduced using the Levi–Civita connection (which we will introduce
later, but see also [43, p. 85–87]) or via the volume form on (oriented) Riemannian manifolds,
as we will see now. On Euclidean space they reduce to the usual definitions.
Definition 2.60. Let (M, g) be an (oriented) Riemannian manifold with volume form
dVg . Let X be a smooth vector field on M . The exterior derivative of the (n − 1)-form11
X⌟ dVg can be expressed by a smooth function, called the divergence of X, multiplied by dVg ,
d(X⌟ dVg ) = (div X)dVg .
11The symbol ⌟ denotes the interior multiplication of ω by X, that is, X⌟ ω(Y , . . . , Y ) :=
2 n
ω(X, Y2 , . . . , Yn ).
 2.2. BASIC CONSTRUCTIONS 19

Remark 2.61. Note that div X is well-defined also on nonorientable manifolds because
we can still choose an orientation locally around every point to define the divergence as above,
and because reversing the orientation changes the sign of dVg on both sides of the equation
leaving div X independent of the orientation.

Problem 2.62. Let (M, g) be a Riemannian manifold and let (xi ) be smooth coordinates
on an open set U ⊆ M . Show that
1 p
div(X i ∂i ) = √ ∂i (X i det g),
det g
∂
where ∂i = ∂x i and det g = det(gkl ) is the determinant of the component matrix of g in these
coordinates.
In particular, on Rn with the Euclidean metric and the standard coordinates we recover
i
the usual formula div(X) = ni=1 ∂X
P
∂xi
.
Using the divergence operator one can prove the following important result (which holds
also in the nonorientable case by using the Riemannian density [35, Thm. 16.48]).
Problem 2.63 (Divergence Theorem). Suppose (M, g) is a compact orientable Riemann-
ian manifold with boundary.
(i) Prove the following divergence theorem for X ∈ X(M ):
ˆ ˆ
(div X)dVg = ⟨X, N ⟩g dVg̃ ,
M ∂M
where N is the outward unit normal to ∂M and g̃ is the induced metric on ∂M .
(ii) Show that the divergence operator satisfies the following product rule for u ∈
C ∞ (M ) and X ∈ X(M ):
div(uX) = u div X + ⟨grad u, X⟩g ,
and deduce the following “integration by parts” formula:
ˆ ˆ ˆ
⟨grad u, X⟩g dVg = u⟨X, N ⟩g dVg̃ − u div X dVg .
M ∂M M
What does this formula say when M is a compact interval in R?
Using the divergence we can define the Laplace–Beltrami operator (with the convention
that the eigenvalues are positive).
Definition 2.64. Let (M, g) be an (orientable) Riemannian manifold. The Laplace–
Beltrami operator (or Laplacian) is the linear operator ∆ : C ∞ (M ) → C ∞ (M ) defined by
∆u := div(grad u).

Problem 2.65. Let (M, g) be a Riemannian manifold and let (xi ) be smooth coordinates
on an open set U ⊆ M . Show that
1 p
∆u = √ ∂i (g ij det g ∂j u),
det g
∂
where ∂i = ∂x i and det g = det(gkl ) is the determinant of the component matrix of g in these
coordinates.
In particular, on Rn with the Euclidean metric and the standard coordinates we recover
∂2u
the usual formula ∆u = ni=1 (∂x
P
i )2 .
20 2. RIEMANNIAN MANIFOLDS

Remark 2.66. While on a Riemannian manifold the Laplace–Beltrami operator is ellip-

tic, it is a hyperbolic operator on a Lorentzian manifold (usually called wave operator or
d’Alembertian then, and denoted by □).
Another important second order operator is the Hessian (see later or [36]).

2.3. Lengths and Riemannian distance

According to Marcel Berger [7, Sec. 14.6] “Gromov’s mm spaces are [...] the geometry
of the future, redefining what we mean by a geometric space, to unify the subjects of prob-
ability and metric geometry”. Here, mm stands for metric measure, the latter connecting
the geometric theory to probability. Gromov’s theory [11, 27] extends the geometry of Rie-
mannian manifolds to length metric spaces equipped with measures on which one can define
upper and/or lower curvature bounds. We have already seen in Section 2.2.2 that there is a
canonical notion of volumes on Riemannian manifolds. In this section we will establish the
connection to metric spaces12.
2.3.1. Lengths of curves. The basic notions of curves are the following.
Definition 2.67. A (parametrized) curve on a smooth manifold M is a continuous map
γ : I → M , where I ⊆ R is an interval.
A curve segment is curve defined on a compact interval I.
We say that γ is a smooth curve if it is smooth as a map from the manifold (with boundary)
I to M .
A regular curve is a smooth curve satisfying γ ′ (t) ̸= 0 for all t ∈ I.
In what follows with work with piecewise smooth curves in order to allow for “corners” or
“kinks”. To define these pieces we use a partition of [a, b], that is, a finite sequence (a0 , . . . , ak )
of real numbers such that a = a0 < a1 < . . . < ak = b.
Definition 2.68. A (continuous) curve segment γ : [a, b] → M on a smooth manifold M
is called piecewise regular (or admissible) if there exists a partition (a0 , . . . , ak ) of [a, b] such
that γ|[ai−1 ,ai ] is a regular curve segment for all i = 1, . . . , k.

Figure 2.3. A piecewise regular curve.

At the partition points a1 , . . . , ak−1 there are two one-sided velocity vectors
γ ′ (a− ′
i ) := lim γ (t), γ ′ (a+ ′
i ) := lim γ (t),
t↗ai t↘ai

12In what follows we reserve the term metric for the Riemannian metric (a tensor) and will instead use
distance for the induced metric (in the usual topological sense).
 2.3. LENGTHS AND RIEMANNIAN DISTANCE 21

that are nonzero, but not necessarily equal.

Generally, the parametrization of a curve is not relevant, but in order to remain admissible,
only a certain class of reparametrizations is allowed.
Definition 2.69. Let γ : [a, b] → M be an admissible curve. A reparametrization of γ is
a curve γ̃ = γ ◦ φ where φ : [c, d] → [a, b] is a homeomorphism and for a partition (c0 , . . . , ck )
of [c, d] the restrictions γ|[ci−1 ,ci ] are diffeomorphisms on their image.
A reparametrization is called forward if it is increasing, and backward if it is decreasing.
Definition 2.70. The length of an admissible curve γ : [a, b] → M is defined as
ˆ b
Lg (γ) := |γ ′ (t)|g dt.
a
Note that due to the regularity of γ this integral is well-defined.
The following result shows that a length structure is obtained from these definitions.
Proposition 2.71 (Properties of lengths). Suppose (M, g) is a Riemmanian manifold
and γ : [a, b] → M an admissible curve. The following hold:
(i) Additivity of length: If a < c < b, then Lg (γ) = Lg (γ|[a,c] ) + Lg (γ|[c,b] ).
(ii) Parameter independence: If γ e is a reparametrization of γ, then Lg (γ) = Lg (e γ ).
(iii) Isometry invariance: If (M, g) and (M , ge) are Riemannian manifolds and φ : M →
f
f is a local isometry, then Lg (γ) = Lge(φ ◦ γ).
M
(iv) Regularity of the arc-length function: The function s : [a, b] → M , defined by
ˆ t
s(t) := Lg (γ|[a,t] ) = |γ ′ (u)|g du,
a
is continuous everyhwere and smooth wherever γ is, with derivative s′ (t) = |γ ′ (t)|
called the speed of γ.

Exercise 2.72. Prove Proposition 2.71.

If γ : [a, b] → M is a unit-speed admissible curve, i.e., |γ ′ (t)| = 1 wherever γ is smooth,
then the arc-length function simply reads s(t) = t − a. The following results shows that such
a parametrization by arc length is always possible.
Proposition 2.73. Suppose (M, g) is a Riemannian manifold.
(i) Every regular curve in M has a unit-speed parametrization.
(ii) Every admissible curve in M has a unique forward reparametrization by arc length.
Sketch of proof. For a regular curve show that s is a strictly increasing local diffeo-
morphism, and thus φ = s−1 the desired reparametrization (use the chain rule). For an
admissible curve it follows from induction on its pieces. (For further details see [36, Prop.
2.49].) □
2.3.2. Riemannian distance function. We extend the most important concept from
classical geometry to the Riemannian setting.
Definition 2.74. Suppose (M, g) is a connected Riemannian manifold. The Riemannian
distance between each pair of points p, q ∈ M is defined as
dg (p, q) := inf{Lg (γ); γ admissible curve between p and q}.
22 2. RIEMANNIAN MANIFOLDS

The following result guarantees that this definition yields a well-defined map d : M ×M →
[0, ∞).
Proposition 2.75. If M is a connected smooth manifold, then any two points can be
joined by an admissible curve.

Figure 2.4. Proof of Prop. 2.75, namely that there exists a piecewise regular curve c
between any two points p and q on a connected manifold.

Proof. Let p, q ∈ M be arbitrary points. Since M is connected it is also path-connected.

Hence there exists a continuous path γ : [a, b] → M . By compactness of γ([a, b]), there are
finitely many curve segments γ([ai−1 , ai ]) each of which are contained in a single coordinate
ball. Therefore, they can be replaced by a straight line in coordinates, which in M yields a
piecewise regular curve between the same points. □

Exercise 2.76. Show that the infimum of curve lengths dḡ (p, q) fails to be realized on
the punctured plane R2 \ {(0, 0)}.

Proposition 2.77 (Isometry invariance of the Riemannian distance function). Suppose

f, ge) are connected Riemannian manifolds and φ : M → M
(M, g) and (M f is an isometry. Then

dge(φ(p), φ(q)) = dg (p, q), p, q ∈ M.

Exercise 2.78. Prove Proposition 2.77.

Note that unlike lengths of curves, distances are not preserved by local isometries.
Problem 2.79. Suppose (M, g) and (M f, ge) are connected Riemannian manifolds, and
φ: M → M
f is a local isometry. Show that

dge(φ(p), φ(q)) ≤ dg (p, q), p, q ∈ M.

Give an example to show that equality need not hold.
The following important results holds.
 2.4. SYMMETRIES OF RIEMANNIAN MANIFOLDS 23

Theorem 2.80 (Riemannian manifolds as metric spaces). Let (M, g) be a connected Rie-
mannian manifold. The distance function dg is a metric on M whose metric topology induces
the manifold topology.

Exercise 2.81. Show dg satisfies the triangle inequality and is symmetric.

We will only prove the rest of the statement later, because the proof of positive definiteness
is much easier to do with normal neighborhoods. Nevertheless, it is also possible (but much
more involved) to show this result in an elementary way (see [36, p. 37–39] or [13, Sec. 4.1]).
In fact, all metric space structures induced by (continuous) Riemannian metrics are equivalent
on compact sets [13, Thm. 4.5].
Using this metric space structure on (M, g), all notions and results from the analysis
and geometry of metric spaces carry over. For instance, we say that (M, dg ) is metrically
complete if every Cauchy sequences in M converges. A subset A ⊆ M is bounded if there
exists a constant C > 0 such that the diameter satisfies
diam(A) = sup{dg (p, q); p, q ∈ A} ≤ C.
Since every compact metric space is bounded, every compact connnected Riemannian mani-
fold has finite diameter.
Later we will prove more important properties related to dg , such as when the Heine–Borel
property holds and the Hopf–Rinow Theorem relating metric and geodesic completeness.
Remark 2.82. On Lorentzian manifolds one can consider the lengths of timelike curves
in much the same fashion and even introduce a Lorentzian distance (based on the maximiza-
tion of curve lengths). While these constructions are still very important and provide some
information about the underlying Lorentzian manifold, they are of much more limited use
compared to Riemannian manifolds, because they do not induce a canonical metric space
structure (see [6, Ch. 4] for more details). Recent alternative approaches are the null dis-
tance [2, 15, 50, 53] and Lorentzian length spaces [34], both rooted in Lorentzian causality
theory (which studies the order structures induced by timelike and causal curves, a unique
feature of time-oriented Lorentzian manifolds).

2.4. Symmetries of Riemannian manifolds

Before introducing the technical notions of connections and curvature on Riemannian
manifolds, we pause for some considerations related to symmetries of M . Our most important
model spaces13 are (again) Euclidean space, sphere, and hyperbolic space. We will briefly also
discuss some other important examples such as Lie groups, homogeneous and isotropic spaces.
All of these spaces are highly symmetric and thoroughly discussed in [36, Ch. 3] (and more).

13Euclidean space, sphere, and hyperbolic space are called model spaces, because they have high degree of
symmetry and–as we will later see—are spaces of constant curvature 0, 1 and −1, see Figure 1.2 on page 3. The
Lorentzian analogues Minkowski space, de Sitter space and anti-de Sitter space (not discussed here) appear as
important and simplest solutions to the Einstein equations in General Relativity.
By the way, Willem de Sitter (1872–1934) was a Dutch mathematician, physicist and astronomer who
made major contributions to physical cosmology.
24 2. RIEMANNIAN MANIFOLDS

2.4.1. Isometries. We have already considered isometries between two Riemannian

manifolds in Section 2.1.2. In this section we fix the Riemannian manifold to see what
kind of information isometries encode about it.
Definition 2.83. An isometry of M is an isometry of (M, g) to itself. The set of all
isometries of (M, g) forms a group under composition, called the isometry group of (M, g)
and denoted by Iso(M, g).
The Myers–Steenrod Theorem [38] from 1939 states that if (M, g) is a connected Rie-
mannian manifold then Iso(M, g) is a Lie group acting smoothly on M . Let us observe this
in our most important examples where the isometry groups are (compact) Lie groups.
Example 2.84 (Euclidean group). The Euclidean group is the semidirect product14
E(n) :=Rn ⋊ O(n)
={F : Rn → Rn ; F (x) = b + Ax for b ∈ Rn and A ∈ O(n)}
The translational part and rotational part are uniquely determined. It is clear that the maps
F preserve the Euclidean metric ḡ and hence are isometries, that is, Iso(En ) ⊇ Rn ⋊ O(n).
The converse is based on a uniqueness result for Riemannian isometries which states that
any two (local) isometries that agree at one point p (here, the origin) and whose differentials
agrees at p are, in fact, equal (we will prove this statement later). Thus
Iso(En ) = E(n) = Rn ⋊ O(n).
For the sphere and hyperbolic space we make use of their description as codimension-1
submanifolds of Rn+1 and R1,n , respectively (see Section 2.1.5).
Example 2.85 (Rotation group). The linear action of the orthogonal group O(n + 1) of
Rn+1 preserves Sn (R) (the origin remains fixed) and the Euclidean metric ḡ, and thus acts
isometrically on S(R) and preserves g̊R . One can also show the converse (again later), thus
spheres (of any Radius R > 0) have isometry group
Iso(Sn (R)) = O(n + 1).
Example 2.86 (Lorentz group). The (n + 1)-dimensional Lorentz group O(1, n) denotes
the group of linear maps from R1,n to itself that preserve the Minkowski metric15 η, i.e.,
O(1, n) := {L : Rn+1 → Rn+1 ; η̄(Lv, Lv) = η̄(v, v)}.
One can show (see Ex. 2.87 below) that O(1, n), just like in the case of the sphere, preserves
the two-sheeted hyperboloid. Since hyperbolic space, however, only consists of the upper
hyperboloid, we have to restrict to a subgroup “preserving the forward time direction”. This
subgroup
O+ (1, n) := {L ∈ O(1, n); L0 0 > 0}
14If G is a group, H < G a subgroup and N ◁ G a normal subgroup (a subgroup invariant under conjugation,
i.e., gng −1 ∈ N for g ∈ G, n ∈ N ) such that G = N H and N ∩ H = {1}, then we say G = N ⋊ H
is a semidirect product of H extended by N . Equivalently, it means that we can write every g ∈ G as a
product nh with a unique n ∈ N and h ∈ H (or the other way round). The group operation is given by
(n1 , h1 )(n2 , h2 ) := (n1 h1 n2 , h1 h2 ) for n1 , n2 ∈ N , h1 , h2 ∈ H. Note that in case of the Euclidean group the
subgroup T(n) ∼ = Rn of translations is the normal subgroup (Why?).
15Just like in the Euclidean case, the isometry group of Minkowski space itself is actually larger, namely
the Poincaré group R1,n ⋊ O(1, n).
 2.4. SYMMETRIES OF RIEMANNIAN MANIFOLDS 25

is called the orthochronous Lorentz group. Since O+ (1, n) preserves Hn (R), and because it
preserves the Minkowski metric it acts isometrically on Hn (R), i.e., preserves ğR . One can
show (later) that it is, in fact, the whole isometry group, that is,
Iso(Hn (R)) = O+ (1, n).

Exercise 2.87. Show that each element in O(1, n) preserves the two-sheeted hyperboloid
{τ 2 − |ξ|2 = R2 }.
We have seen that Lie groups appear as isometry groups. Conversely, we can ask when a
Lie group can be turned into a group of isometries.
Example 2.88 (Lie groups). For a Lie group16 G the tangent bundle can be trivialized,
i.e.,
TG ∼
= G × Te G
by using left (or right) translations on G. Thus fixing an inner product on Te G ∼
= g induces a
left-invariant17 Riemannian metric g on G (implying, at the same time, that left translations
are Riemannian isometries), i.e.,
L∗φ g = g, φ ∈ G,
where Lφ (φ′ ) = φφ′ denotes a left translation [36, Lem. 3.10].

Problem 2.89. Let o(n) denote the Lie algebra of O(n), identified with the algebra of
skew-symmetric18 n × n matrices, and define a bilinear form on o(n) by
⟨A, B⟩ := tr(AT B).
Show that this determines a bi-invariant Riemannian metric on O(n) (show that it is Ad-
invariant, using [36, Prop. 3.12]).

2.4.2. Homogeneity and isotropy. Depending on particular properties of the isometry

group various “highly symmetric” Riemannian metrics can be identified. We distinguish be-
tween homogeneous Riemannian manifolds that “look the same at every point” and isotropic
ones that “look the same in every direction”.
Definition 2.90. Let (M, g) be a Riemannian manifold. We say that (M, g) is a homo-
geneous Riemannian manifold if Iso(M, g) acts transitively on M , that is, for all p, q ∈ M
exists φ ∈ Iso(M, g) such that φ(p) = q.
Example 2.91. Since a Lie group acts transitively on itself by left translations, every
left-invariant metric is homogeneous.
16A Lie group is a smooth manifold and a group such that the group operations (multiplication and
inversion) are smooth. For more background see [36, App. C].
17From a Riemannian point of view bi-invariant metrics are much more interesting because their curvatures
are intimately tied to the Lie algebra structure. Bi-invariant metrics are much harder to find though, but they
always exist for compact Lie groups [36, Cor. 3.15]. More generally, John Milnor [39] showed that a Lie group
admits a bi-invariant Riemannian metric if and only if it is isomorphic to the product G × H with G compact
and H abelian.
18A matrix A is skew-symmetric if AT = −A.
26 2. RIEMANNIAN MANIFOLDS

Remark 2.92 (Poincaré conjecture). Locally homogeneous Riemannian metrics (where

homogeneity holds in a neighborhood) play a key role in the classification of compact 3-
manifolds. In 1904 Poincaré conjectured that every simply connected, closed topological
3-manifold is homeomorphic to the 3-sphere. This was called the Poincaré conjecture and
made it as one of the ten biggest unsolved mathematics problems on the Clay Millenium List19
compiled in 2000. Later in 1982, William Thurston made a more general conjecture about the
classification of 3-manifolds, called the geometrization conjecture, stating that every compact
orientable 3-manifold can be expressed as the connected sum of compact manifolds, each of
which admits a Riemannian covering by a homogenous Riemannian manifold or can be cut
along embedded tori so that each piece admits a finite-volume locally homogeneous Riemann-
ian metric. The classification of all simply connected homogeneous Riemannian 3-manifolds
that admit finite-volume Riemannian quotients (of which there are 8) played a key role. Even-
tually, Thurston’s geometrization conjecture was shown to be true by Grigori Pereleman in
2003, building upon earlier work by Richard Hamilton, using Ricci flow. Perelman published
his work in preprints only. He was offered (and declined) the Fields Medal in 2006. He was
also awarded (and refused; you may want to read the novel [59] for potential further insight)
the Clay Millenium Prize worth 1 million US dollars in 2010 for the solution of the Poincaré
conjecture, so far the only solved Millenium Prize Problem. In any case, a true win and
milestone for Riemannian Geometry!
Definition 2.93. The isotropy subgroup Isop (M, g) of p ∈ M (also called stabilizer of p)
consists of all isometries that fix p.
Example 2.94. The isometries of Euclidean space and Minkowski space generally contain
translations, however, isotropies cannot. Therefore,
Iso0 (En ) = O(n), Iso0 (R1,n ) = O(1, n).
Example 2.95. The isotropy groups of Sn (R) are isomorphic to O(n), i.e., the elements
of O(n + 1) fixing a 1-dimensional subspace of Rn+1 (the rotation axis).
Example 2.96. The isotropy group of hyperbolic space Hn (R) that preserves the origin
(in the hyperboloidal model it can be identified with the timelike vector Re0 ) is O(n).
Note that Rn ∼
= E(n)/O(n) and Sn (R) ∼= O(n + 1)/O(n). In fact, one can show that any
homogeneous space (M, g) can be written as a quotient Iso(M, g)/ Isop (M, g).
Generally, the differential dφ maps T M to itself, and every φ ∈ Iso(M, g) restricts to
linear isometries dφp : Tp M → Tφ(p) M for all p ∈ M . For isotropies, dφp : Tp M → Tp M is
even more special, and gives rise to the following definition.
Definition 2.97. A Riemannian manifold (M, g) is called isotropic at p if the isotropy
representation
Ip : Isop (M, g) → GL(Tp M ), Ip (φ) = dφp ,
acts transitively on the set of unit vectors in Tp M . In other words, for any two unit vectors
v, w ∈ Tp M there exists and isotropy φ ∈ Isop (M, g) such that20 φ∗ (v) = w.
If M is isotropic at every p we say that M is isotropic.
19See https://en.wikipedia.org/wiki/Millennium_Prize_Problems for the complete list.
20Recall that the pushforward of X ∈ X(M ) is the unique vector field φ X ∈ X(M
f) that satisfies
∗
dφp (Xp ) = (φ∗ X)φ(p) for all p ∈ M .
 2.4. SYMMETRIES OF RIEMANNIAN MANIFOLDS 27

The following result shows that for homogeneous spaces we only need to have isotropy at
one point and can otherwise push it around.
Proposition 2.98. Let (M, g) be a Riemannian manifold. If M is homogeneous and
isotropic at one point, then it is isotropic everywhere.
Proof. Suppose M is isotropic at the point q. Let p be any other point in M . Suppose
v and w are (different) unit vectors in Tp M . By homogeneity there exists an isometry ψ ∈
Iso(M, g) such that ψ(p) = q. Thus dψp (v), dψp (w) are two unit vectors in Tq M , and because
M is isotropic at q there exists an isotropy φ ∈ Isoq (M, g) that transforms dψp (v) into dψp (w),
i.e.,
d(φ ◦ ψ)p (v) = dφψ(p) (dψp (v)) = dφq (dψp (v)) = dψp (w).
e := ψ −1 ◦ φ ◦ ψ ∈ Isop (M, g) satisfies
Hence φ
−1
ep (v) = dψq−1 ◦ dφq ◦ dψp (v) = dψψ(p)
dφ (dψp (w)) = w.
Hence (M, g) is also isotropic at p, and since p was arbitrary, M is isotropic everywhere. □
Remark 2.99. One can also show that (everywhere!) isotropic Riemannian manifolds are
automatically homogeneous, but there are counterexamples for the converse statement. The
Berger metrics on S3 are homogeneous without being isotropic anywhere, see [36, Problem
3-10].
There is an even more special class of symmetries, based on the transformation not only
of the unit sphere in Tp M as in isotropy but based on the transformation of orthonormal
bases, therefore called frame-homogeneous (sometimes, isotropy is defined like that, see [23]).
Frame-homogeneous Riemannian manifolds are both homogeneous and isotropic [36, p. 56].
Example 2.100. Assuming that we have shown Iso(En ) = E(n) (see Ex. 2.84) immediately
implies that En is frame-homogeneous.
By explicitely constructing the rotation moving the orthonormal basis of the north pole (or
origin) to one at another point one obtains the same result for Sn (R) and Hn (R) [36, Props.
3.2 and 3.9].

We have by now observed several properties of the isometry group on model spaces and
Lie groups with regards to homogeneity and isotropy. These results (and more) can be
summarized in the following table (symmetric spaces are Riemannian manifolds with point
reflections at every point; more details and more spaces can be found in [36, Ch. 3]).

En Sn Hn Lie groups symmetric spaces

homogeneous x x x x x
isotropic x x x
frame-homogeneous x x x
Remark 2.101 (Cosmology). Statistically our universe looks “on the large scale” very bor-
ing. In Cosmology one therefore studies homogeneous and isotropic solutions of the Einstein
equations with cosmological constant (following the cosmological principle). The prototypi-
cal example are the Friedmann–Lemaı̂tre–Robertson–Walker (FLRW) spacetimes, which are
nothing else but Lorentzian metrics that are warped products I ×a Σ, where the first part
I is 1-dimensional and negative definite, the second space Σ is a 3-dimensional Riemannian
28 2. RIEMANNIAN MANIFOLDS

model space (that is, E3 , S3 or H3 ), and the warping function a the scale factor. The Ein-
stein equations with a perfect fluid source reduce to a system of two second-order ordinary
differential equations for a, called the Friedmann equations (derived by him in the 1920s).
2.4.3. Conformal transformations. An important generalization of an isometry is
that of a conformal transformation. It does not preserve the lengths of vectors but the angle
between them. In two dimensions (or one complex dimension), conformal geometry is precisely
the study of Riemann surfaces. More generally one refers to a manifold equipped with an
equivalence class of metrics up to a conformal factor as conformal manifolds. Conformal
transformations are particularly important in Lorentzian Geometry and General Relativity
because they leave the light cone structure, and hence the basic geometric concept of causality,
intact.
Definition 2.102. Let M be a manifold. We say that two metrics g1 and g2 are conformal
(or conformally related) to each other if there exists a smooth function f : M → (0, ∞) such
that g2 = f g1 .
Definition 2.103. Let (M, g) and (M f, ge) be two Riemannian manifolds. A diffeomor-
phism φ : M → M f is called a conformal transformation if the pullback of ge is conformal to
g, i.e.,
φ∗ ge = f g for some positive f ∈ C ∞ (M ).
Two Riemannian manifolds are said to be conformally equivalent if there exists a confor-
mal transformation between them.

Exercise 2.104. (i) Show that for every smooth manifold M , conformality is an
equivalence relation on the set of Riemannian metrics on M .
(ii) Show that conformal equivalence is an equivalence relation on the class of all Rie-
mannian metrics.

Exercise 2.105. Suppose g1 and g2 = f g1 are conformally related metrics on an oriented

n
n-manifold. Show that their volume forms are related by dVg2 = f 2 dVg1 .
Besides global conformal transformation one can also study important local version. In
this context, a Riemannian manifold (M, g) is said to be locally conformally flat if around ever
point in M there is a neighborhood that is conformally equivalent to an open set in (Rn , ḡ).
Both sphere and hyperbolic space are locally conformally flat (but they are not locally flat!).
Proposition 2.106. Each sphere with a round metric is locally conformally flat.
Sketch of proof. A local conformal equivalence between En and Sn (R) (local, because
the north pole N is excluded) is given by the stereographic projection
σ : Sn (R) \ {N } → Rn ,
Rξ
(ξ 1 , . . . , ξ n , τ ) 7→ u :=
,
R−τ
which can be shown to be a conformal tranformation between Sn (R) \ {N } and Rn . In
particular, it is a diffeomorphism and satisfies
4R4
(σ −1 )∗ g̊R = ḡ. □
(|u|2 + R2 )2
(For the full proof and pictures see [36, p. 59–61].)
 2.4. SYMMETRIES OF RIEMANNIAN MANIFOLDS 29

Using the (isometric) Poincaré ball or half space models one can also prove that hyperbolic
space Hn (R) is locally conformally flat [36, p. 62–66].
Remark 2.107 (Rigidity of conformal flatness). Conformal flatness is directly related
to different curvature components vanishing or being constant. In dimension 2 all spaces
with constant sectional curvature are conformally flat. Liouville’s Theorem from 1850 states
that metrics conformal to the flat metric ḡ on En , for n ≥ 3, can only be obtained through
translation, similarity, orthogonal transformation and inversion. Hence the class of conformal
mappings in dimension n ≥ 3 is much poorer than in dimension 2, adding additional rigidity
to metrics that are conformally flat. The Weyl–Schouten Theorem states that a Riemannian
manifold of dimension n ≥ 3 is conformally flat if and only if the Schouten21 tensor is a
Codazzi tensor (if n = 3) or if the Weyl22 tensor vanishes (if n ≥ 4). More about this in
Section 5.4.
Remark 2.108 (Conformal transformations in General Relativity). In Lorentzian Ge-
ometry, conformal transformations are of great importance because they preserve the causal
structure. Already in 1921 Kasner [33] essentially proved that the a vacuum spacetime that
is locally conformally flat is, in fact, flat. Around the same time Brinkmann [10] investigated
conformal transformations between two Einstein spaces as well as conditions for a space to
be conformal to an Einstein space. As a byproduct he found plane waves, which reappear in
an important paper of Penrose23 that also makes use of conformal transformations: In 1976
Penrose [44] introduced adapted coordinates around a null geodesic in order to associate with
any spacetime metric g a family of conformally equivalent metrics gλ = λ−2 g. The Penrose
limit is then the metric g0 = limλ→0 gλ , and it can be shown that this is a plane wave metric.
It follows, for instance, that the Penrose limit of an Einstein manifold is Ricci flat.

21Jan Schouten (1883–1971) was a Dutch mathematician contributing to tensor calculus and Ricci calculus.
22Hermann Weyl (1885–1955) was one of the most influential mathematicians of the 20th century, worked
in many different areas of mathematics and made major contributions everywhere, also related to theoretical
physics.
23Sir Roger Penrose (1931–) is a British mathematican and mathematical physicist. In 2020 he won the
Nobel Prize for physics for his singularity theorem in General Relativity (from 1965) describing, vaguely put,
black hole formulation.
 CHAPTER 3

Connections and covariant differentation

In the “Manifolds” course you may have already heard about Lie derivatives and exte-
rior derivative on manifolds [35, Ch. 9 and 14]. The main goal of this chapter is to introduce
connections and their use for covariant differentiation, which immediately relates to the direc-
tional derivative (on functions), the Lie derivative (on vector fields) and the exterior derivative
(on forms) by extending them to tensors.
A nice introduction and motivation for connections via parallelism can be found in the
book of do Carmo [23, Ch. 2]. Also in the book of Lee [36, p. 85–88] there is a good
recollection of the concept of directional differentiation of vector fields for submanifolds M of
Rn . In this setting the projection onto the part that is tangential to M is used. This is clearly
not an intrinsic geometric concept and hence highlights that there is something to resolve
in the setting of abstract manifolds. In this section we will therefore introduce covariant
differentiation using the basic idea of parallelism, which indeed only depends on the intrinsic
geometry of a manifold (since it is invariant by isometry).
Note that connections can be defined on any manifold, however, a Riemannian manifold
(and also every semi-Riemannian manifold) can be equipped with a canonical connection,
called the Levi-Civita connection, which is fundamental for the modern definition of curvature
via parallel transport along vector fields and curves. Riemann himself did not know what
a connection was. He computed the Riemann curvature tensor as a second-order correction
term in the Taylor expansion of the Riemannian metric around a point (in coordinates). The
notion of a connection postdates Riemann and was developed by the Italian school around
Christoffel and Levi-Civita, Ricci, Bianchi etc. in the context of tensor analysis.

3.1. Affine connections

We define a connection ∇ (pronounced “del” or “nabla”) as a way of differentiating vector
fields and later apply it to vector fields along curves.
Definition 3.1. Let M be a smooth manifold. An affine connection on M is a map
∇ : X(M ) × X(M ) → X(M ),
(X, Y ) 7→ ∇X Y,
satisfying the following properties:
(i) ∇X Y is C ∞ (M )-linear in X: for f1 , f2 ∈ C ∞ (M ), X1 , X2 ∈ X(M ),
∇f1 X1 +f2 X2 Y = f1 ∇X1 Y + f2 ∇X2 Y.
(ii) ∇X Y is R-linear in Y : for a1 , a2 ∈ R, Y1 , Y2 ∈ X(M ),
∇X (a1 Y1 + a2 Y2 ) = a1 ∇X Y1 + a2 ∇X Y2 .
31
32 3. CONNECTIONS AND COVARIANT DIFFERENTATION

(iii) ∇ satiesfies the product rule: for f ∈ C ∞ (M ),

∇X (f Y ) = f ∇X Y + X(f )Y.
The vector field ∇X Y is called the covariant derivative of Y in direction X.
Remark 3.2. Beware, the map (X, Y ) 7→ ∇X Y does not define a tensor field on M . This
is because in place of being C ∞ (M )-linear in Y it is only R-linear in Y and satisfies the Leibniz
product rule (this follows form the Tensor Characterization Lemma A.17). Nevertheless, ∇ is
a well-defined coordinate-independent map. This explains also the use of the word “covariant”
in the covariant derivative which just means that it transforms covariantly (that is, correctly).
Remark 3.3. While (X, Y ) 7→ ∇X Y is not a tensor field on M , we will later see that the
following modification
τ (X, Y ) := ∇X Y − ∇Y X − [X, Y ],
is a (1, 2)-tensor field, called the torsion tensor of ∇. Torsion (or actually, torsion-freeness)
will again become relevant when we consider special connections on Riemannian manifolds.
While the connection appears to be a global concept we can, in fact, see that it is a local
operator because ∇X Y |p only depends on the values of X at p and the values of Y in a
neighborhood of p.
Lemma 3.4 (Locality). Suppose ∇ is an affine connection on a smooth manifold M . If
X, X, Y, Ye ∈ X(M ) such that for a point p ∈ M we have that X|p = X|
e e p and Y = Ye on a
neighborhood of p, then ∇X Y |p = ∇Xe Ye |p .
Proof. First consider X. It follows from the definition of ∇ that X 7→ ∇X Y is a tensor,
hence we even have that
e p =⇒ (∇X Y )|p = ∇X| Y = (∇ e Y )|p .
X|p = X| (3.1)
p X
Second consider Y . It suffices to show that on an open set U we have Y |U = 0 implies
that (∇X Y )|U = 0. Let q ∈ U and choose a bump function χ ∈ C ∞ (M ) with support in U
and such that χ ≡ 1 in a neighborhood of q. Hence χY ≡ 0 on M , and therefore by (iii) of
Definition 3.1 we have for every X ∈ X(M ) that
(iii)
0 = (∇X ( χY ))|q = X(χ)|q Y |q + χ(q)(∇X Y )|q ,
|{z} | {z } |{z}
≡0 =0 =1

hence ∇X Y |U = 0. □
This result implies that the restriction ∇U of an affine connection ∇ on an open set U of
M defines a unique connection on T M |U (see [36, Prop. 4.2]).
For computations we want to express a connection in terms of a local frame (or local
coordinates). Let (Ei ) be a smooth local frame on T M on an open set U ⊆ M . For every i
and j, ∇Ei Ej is a vector field that we can expand in the same frame, so that
∇Ei Ej =: Γkij Ek . (3.2)
The n3 smooth functions Γkij : U → R are called the connection coefficients of ∇ with respect
to the given frame.
 3.1. AFFINE CONNECTIONS 33

The connection on U is then completely determined by the connection coefficients {Γkij }:

We can write two X, Y ∈ X(U ) in the given frame, i.e., X = X i Ei , Y = Y j Ej , and using
properties (i)–(iii) of ∇ compute
∇X Y = ∇X (Y j Ej )
= X(Y j )Ej + Y j ∇X i Ei Ej
= X(Y j )Ej + X i Y j ∇Ei Ej
= X(Y j )Ej + X i Y j Γkij Ek .
Hence locally on U we have we obtain the following result (by relabeling).
Proposition 3.5. Let M be a smooth manifold, and let ∇ be an affice connection on
M . Suppose (Ei ) is a smooth local frame over an open subset U ⊆ M , and let {Γkij } be the
connection coefficients of ∇ with respect to this frame. For smooth vector fields X, Y ∈ X(M ),
written in terms of the frame as X = X i Ei , Y = Y j Ej one has
∇X Y = (X(Y k ) + X i Y j Γkij )Ek . (3.3)
□
Example 3.6 (Euclidean connection). In T Rn we define the Euclidean connection ∇
pointwise by
∇X Y := X(Y 1 )∂1 + . . . + X(Y n )∂n , X, Y ∈ X(Rn ), (3.4)
∂
where ∂i := ∂xi
with respect to the standard coordinates (x1 , . . . , xn ) in Rn .

Exercise 3.7. Check that the Euclidean connection, defined in (3.4), is indeed a con-
nection (satisfies the required properties) and that its connection coefficients in the standard
coordinate frame are all zero.
There is a natural way to define an affine connection on a submanifold M of Rn by
orthogonally projecting onto T M (the same idea can be used if Rn is replaced by another
ambient semi-Riemannian manifold M f). Here enter some ideas from considering connections
of vector bundles, which we briefly describe in Section 3.2. For now we introduce some
notation to make this idea in the context of submanifolds precise.
Definition 3.8. Let M be a submanifold embedded in a Riemannian manifold (M f, ge).
At every point p ∈ M we can defined the normal space at p as
Np M := (Tp M )⊥ := {v ∈ Tp M
f; gep (v, w) = 0 for all w ∈ Tp M },
and the normal bundle of M as G
N M := Np M.
p∈M

One can show that N M is a smooth rank-(m − n) vector subbundle (see Section 3.2) of
f|M , and that there are smooth bundle homomorphisms
the ambient tanget bundle T M
π⊤ : T M
f|M → T M, π⊥ : T M
f|M → N M,
called the tangential and normal projection, that for each p restrict to orthogonal projections
from Tp Mf to Tp M and Np M , respectively (the proof uses an adapted orthonormal frame, see
[36, Prop. 2.16]).
34 3. CONNECTIONS AND COVARIANT DIFFERENTATION

Example 3.9 (Tangential connection on a submanifold of Rn ). Let M be an embedded

submanifold in Rn and X, Y ∈ X(M ). By using smooth extensions X e and Ye of X and Y to an
n
open set of R (one can show that such extensions exist, but they are not unique, see [36, Ex.
A.23]) and the orthogonal projection π ⊤ onto T M we can define the tangential connection
∇⊤ on T M by
 
∇⊤X Y := π ⊤
∇ X
e Y
e | M . (3.5)

Problem 3.10. Show that the tangential connection ∇⊤ defined in (3.5) for submanifolds
M of Rn is indeed a connection, that is, verify that
(i) ∇⊤ is well-defined (independent of the extensions X e and Ye ), and
⊤
(ii) ∇ satisfies the axioms of an affine connection on M .
(Hint: (i) Note that the value of ∇Xe Ye at a point p ∈ M only depends on X ep = Xp and
is hence independent of the extension. Show then that the tangential directional derivative
∇⊤ ⊤ ⊤
Xp Y := π (∇Xp Y ) is also independent of the choice of extension Y , so ∇ is well-defined.
e e
Smoothness of ∇Xe Ye holds by using an adapted orthonormal frame [36, Prop. 2.14]. (ii) It
is straightforward to check that (i)–(iii) of Definition 3.1 hold. For verifying the product rule
extend f ∈ C ∞ (M ) to a smooth function fe on a neighborhood of M .)
Now that we have seen some examples of important connections, the question remains
whether there always exists a connection in the tangent bundle of a manifold. Let us start
with a simplified assumption, and work our way up.
On any manifold admitting a global frame (Ei ) a connection can easily be obtained by
picking n3 smooth real-valued functions {Γkij } and then defining ∇ via (3.3), i.e.,
∇X Y := (X(Y k ) + X i Y j Γkij )Ek
(see Exercise 3.12 below). Using a partition of unity {φα } subordinate to an atlas {Uα } then
allows one to obtain a connection on any manifold M . More precisely, on each neighborhood
Uα a connection ∇α exists by the above, and those connections can be put together to a
connection X
∇X Y := φα ∇αX Y, X, Y ∈ X(M ),
α
on M (for the details see the proof of [36, Prop. 4.12]). Thus we argued that connections
always exist.
Proposition 3.11. The tangent bundle of every smooth manifold admits a connection.
□

Exercise 3.12. Suppose M is a smooth n-manifold admitting a global frame (Ei ). Then
(3.3) gives a one-to-once correspondence between connections on T M and sets of n3 functions
Γkij ∈ C ∞ (M ).
(Hint: We have already seen that every connection determines functions {Γkij } by (3.3) in
a unique way. On the other hand, one can easily check that ∇X Y defined by (3.3) using given
smooth real-valued functions {Γkij } yield a expression that is smooth in X and Y , R-linear in
Y , and C ∞ (M )-linear in X. Checking the product rule is a straightforward computation.)
 3.3. COVARIANT DIFFERENTIATION OF TENSOR FIELDS 35

3.2. Connections in vector bundles

Recall that a vector field is just a section of the tangent bundle T M , i.e., X(M ) =
Γ(T M ) = C ∞ (M, T M ). In this spirit one can extend Definition 3.1 to sections Γ(E) of
vector bundles. For a short intro to vector bundles see Section A.1 of Appendix A as well as
[36, p. 382–384], and [35, Ch. 10] for a more complete treatment.
Definition 3.13. Let π : E → M be a smooth vector bundle over a smooth manifold M .
A connection in E is a map
∇ : X(M ) × Γ(E) → Γ(E),
(X, Y ) 7→ ∇X Y,
for which (i)–(iii) of Definition 3.1 hold for sections Γ(E) in place of X(M ).
The expression ∇X Y is called the covariant derivative of Y in direction X.

Exercise 3.14 (Trivial connection). Show that any connection in the trivial line bundle
prM : M × R → M is given by the exterior derivative d : C ∞ (M ) → Ω1 (M ) (recall that by
Example A.6 the sections are Γ(M ×R) = C ∞ (M ); for the differential of a function see [35, p.
280–284] and, more generall, for the exterior derivative see [35, p. 362–373]) with covariant
derivative along X being of the form ∇X f = df (X) + f ω(X) for any 1-form ω (the special
case ∇X f = df (X) = X(f ) is the flat connection, for which the curvature tensor vanishes).
How can we extend this definition to a connection in the trivial bundle M × Rk of rank
k?

Exercise 3.15. Show that there is a connection in any vector bundle π : E → M over a
smooth manifold M .
Considering connections on vector bundles is a far-reaching concept that can also be used
in connection with, for instance, Lie group actions on manifolds. These ideas are relevant,
in particular, in Gauge Theory and Algebraic Geometry. We will later use connections in
tensor bundles and the normal bundle, and one can either view those as extensions of affine
connections or in this wider framework of vector bundles. Vector bundles themselves are
widely used in Differential Geometry, and if you follow a Master course in Differential Geom-
etry, you will learn a lot more about them. Since the main focus of this course is to provide
a short introduction to Riemannian Geometry, we keep the notions brief and focus on the
implications and use of connections in this context. See [36, p. 88–91] and [29, Ch. 2] for
more details.

3.3. Covariant differentiation of tensor fields

An affine connection on M is a connection in the tangent bundle T M and gives us a way
to compute the covariant derivatives of vector fields. We will see that every such connection
induces a canonical connection in all tensor bundles over M , which can also be interpreted
in the sense of Section 3.2 because every tensor bundle T (k,l) T M is a vector bundle with
sections being the tensor fields T (k,l) (M ) = Γ(T (k,l) T M ) (recall Section A.3 for the definition
of tensor fields).
To indeed obtain a unique tensor derivation on T (k,l) (M ) for a given connection in T M ,
we extend it via the exterior derivative d on C ∞ (M ) to a tensor derivation (essentially using
the idea of Exercise 3.14 in (ii) of Proposition 3.16 below). In Section 3.4 we will furthermore
see that on a Riemannian manifold we can already pick a canonical connection in T M .
36 3. CONNECTIONS AND COVARIANT DIFFERENTATION

Proposition 3.16. Let M be a smooth manifold and ∇ be an affine connection on T M .

Then there exists a unique connection on each tensor bundle T (k,l) T M (denoted also by ∇),
∇ : X(M ) × T (k,l) (M ) → T (k,l) (M ),
such that the following four properties are satisfied:
(i) In T (1,0) T M = T M , ∇ agrees with the given connection.
(ii) In T (0,0) T M = M × R, ∇ is given by the ordinary differentiation of functions, i.e.,
∇X f = X(f ), f ∈ C ∞ (M ).
(iii) ∇ obeys the product rule with respect to tensor products, i.e.,
∇X (F ⊗ G) = (∇X F ) ⊗ G + F ⊗ (∇X G).
(iv) ∇ commutes with all contractions, i.e., if trace acts on any pair of covariant and
contravariant indices, then
∇X (tr F ) = tr(∇X F ).
In addition, ∇ satisfies:
(a) ∇ obeys the product rule with respect to the natural pairing of a 1-form and a vector
field, i.e.,
∇X (ω(Y )) = (∇X ω)(Y ) + ω(∇X Y ).
(b) For all F ∈ T (k,l) (M ) with ω 1 , . . . , ω k ∈ Ω1 (M ), Y1 , . . . , Yl ∈ X(M ) we have
(∇X F )(ω 1 , . . . , ω k , Y1 , . . . , Yl ) = X(F (ω 1 , . . . , ω k , Y1 , . . . , Yl ))
k
X
− F (ω 1 , . . . , ∇X ω i , . . . , ω k , Y1 , . . . , Yl ) (3.6)
i=1
Xl
− F (ω 1 , . . . , ω k , Y1 , . . . , ∇X Yj , . . . , Yl ).
j=1

Proof. Step 1. (i)–(iv) ⇒ (a)–(b). This is an easy computation, since for (a) we have
ω(Y ) = tr(ω ⊗ Y ) (prove this in coordinates, where both are ωi Y i ) and by (i)–(iv)
(iv)
∇X (ω(Y )) = ∇X (tr(ω ⊗ Y )) = tr(∇X (ω ⊗ Y ))
(iii)
= tr(∇X ω ⊗ Y + ω ⊗ ∇X Y ) = (∇X ω)(Y ) + ω(∇X Y ),
and (b) is obtained by using induction applied to
F (ω 1 , . . . , ω k , Y1 , . . . , Yl ) = |tr ◦ .{z
. . ◦ tr}(F ⊗ ω 1 ⊗ . . . ⊗ ω k ⊗ Y1 ⊗ . . . ⊗ Yl ).
k+l

Step 2. Uniqueness. Assume ∇ is a connection satisfying (i)–(iv), then by the above also
(a)–(b) holds. For every 1-form ω we obtain
(a) (ii)
(∇X ω)(Y ) = ∇X (ω(Y )) − ω(∇X Y ) = X(ω(Y )) − ω(∇X Y ), (3.7)
hence the connection on 1-forms is uniquely determined by the original affine connection on
T M . Similarly, (b) gives a formula that determines the covariant derivative of an arbitrary
tensor field F in terms of the covariant derivatives of vector fields (using (i)) and 1-forms
(using (3.7)), thus the connection in every tangent bundle is also unique.
 3.3. COVARIANT DIFFERENTIATION OF TENSOR FIELDS 37

Step 3. Existence. The covariant derivatives on 1-forms is defined by (3.7), then define ∇
on all tensor bundles via (3.6). We first note that ∇X F is a tensor field since it is multilinear
over C ∞ (M ), i.e., for f ∈ C ∞ (M ) and all i and j a computation (cancellation of two terms
involving f ) yields, for instance,
(∇X F )(ω 1 , . . . , f ω i + ω
e i , . . . , ω k , Y1 , . . . , Yl ) = . . .
= f (∇X F )(ω 1 , . . . , ω i , . . . , ω k , Y1 , . . . , Yl ) + (∇X F )(ω 1 , . . . , ω
e i , . . . , ω k , Y1 , . . . , Yl ).
Thus indeed we obtain a map ∇ : X(M )×T (k,l) (M ) → T (k,l) (M ), and it remains to check that
∇ satisfies the properties (i)–(iii) of a connection from Definition 3.1. Here, C ∞ (M )-linearity
in X and R-linearity in F follow from (3.6) and (3.7), while the product rule follows from the
usual differention of functions by X since
X(f g) = X(f )g + f X(g)
applied to F ∈ T (k1 ,l1 ) (M ) and G ∈ T (k2 ,l2 ) (M ) “pointwise” via the smooth functions
F ⊗ G(ω 1 , . . . , ω k1 +k2 , Y1 , . . . , Yl1 +l2 )
= F (ω 1 , . . . , ω k1 , Y1 , . . . , Yl1 )G(ω k1 +1 , . . . , ω k2 , Yl1 +1 , . . . , Yl1 +l2 ). □
In coordinates we obtain the following expressions.
Problem 3.17. Let M be a smooth manifold and ∇ an affine connection on T M . Suppose
(Ei ) is a local frame for M , (εj ) its dual coframe, and {Γkij } the corresponding connection
coefficients of ∇. Let X = X i Ei be a smooth vector field. Show that
(i) the covariant deriviative of a 1-form ω = ωi εi is locally given by
∇X ω = (X(ωk ) − X j ωi Γijk )εk ,
(ii) if F ∈ T (k,l) (M ), locally given by F = Fji11...j
...ik
l
Ei1 ⊗ . . . ⊗ Eik ⊗ εj1 ⊗ εjl , then the
covariant derivative of F is locally given by
k l
!
m i1 ...p...ik is p
i1 ...ik
X X
m i1 ...ik
∇X F = X(Fj1 ...jl ) + X Fj1 ...jl Γmp − X Fj1 ...p...jl Γmjs Ei1 ⊗ . . . ⊗ εjl . (3.8)
s=1 s=1

The covariant derivative ∇X F of a (k, l)-tensor field F is C ∞ (M )-linear in X and (as can
be shown) C ∞ (M )-linear in all its k + l arguments, hence by the Tensor Characterization
Lemma ∇X F can be seen as a (k, l + 1)-tensor field on M . This gives rise to the following
definition.
Definition 3.18. Let M be a smooth manifold and let ∇ be an affine connection in T M .
Let F ∈ T (k,l) M . Then the total covariant derivative of F is ∇F ∈ T (k,l+1) (M ) given by
(∇F )(ω 1 , . . . , ω k , Y1 , . . . , Yl , X) := (∇X F )(ω 1 , . . . , ω k , Y1 , . . . , Yl ). (3.9)
In a local frame (Ei ) we denote the total covariant derivative with ; (semicolon): For
instance, for a vector field Y = Y i Ei the (1, 1)-tensor field ∇Y reads
∇Y = Y i ;j Ei ⊗ εj , with Y i ;j = Ej Y i + Y k Γijk ,
following (3.8). Similarly, for 1-form ω we obtain
∇ω = ωi;j εi ⊗ εj , with ωi;j = Ej ωi − ωk Γkji .
38 3. CONNECTIONS AND COVARIANT DIFFERENTATION

Exercise 3.19. Let M be a smooth manifold and let ∇ be an affine connection on T M .

Suppose (Ei ) is a smooth local frame for T M and {Γkij } are the corresponding connection
coefficients. Show that the components of the total covariant deriviative of a (k, l)-tensor field
F with respect to this frame are given by
  Xk l
i1 ...p...ik is
Γp .
X
Fji11...j
...ik
l ;m
= Em F i1 ...ik
j1 ...jl + F j1 ...jl Γmp − Fji11...p...i
...ik
l mjs
i=1 i=1

3.4. Levi-Civita connection

Everything that has been said so far (and much more [36, Ch. 4]) holds for any connection
in any vector bundle on any manifold. Since we want to use the covariant derivative (and
parallel transport) to study Riemannian manifolds, it is important to pick a connection that
goes well with its special structure. In this section we single out connections by requiring two
additional properties, namely compatibility with the metric and symmetry (also called torsion-
freeness). Both of these properties are inspired by properties of the tangential connection ∇⊥
for submanifolds of Rn (see Example 3.9). In fact, it turns out that these two additional
properties already determine a unique affine connection on a Riemannian manifold (and,
similarly, on a semi-Riemannian manifold), called the Levi-Civita connection, named after the
Italian geometer Tullio Levi-Civita who defined it in 1917 (though several other mathematics
such as Christoffel, Brouwer, Schouten and Weyl were also involved and obtained similar
results independently).

3.4.1. Metric connections. The Euclidean connection ∇ on En satisfies

∇X ⟨Y, Z⟩ = ⟨∇X Y, Z⟩ + ⟨Y, ∇X Z⟩, (3.10)
as well as the tangential connection ∇⊤ on an embedded submanifold of En (same in the
semi-Riemannian setting).
Exercise 3.20. Prove (3.10) by computing in terms of the standard basis (the left hand
side is simply X⟨Y, Z⟩ by Prop. 3.16 (ii)).

Exercise 3.21. Show that the tangential connection on an embedded submanifold of En

satisfies the product rule (3.10) (with ∇ replaced by ∇⊤ , and with respect to the induced
metric and tangential vector fields).
This concept can be extended to abstract manifolds in the same way.
Definition 3.22. Let (M, g) be a Riemannian manifold. An affine connection ∇ on T M
is said to be compatible with g (or a metric connection) if for all X, Y, Z ∈ X(M ) we have the
product rule1
∇X ⟨Y, Z⟩ = ⟨∇X Y, Z⟩ + ⟨Y, ∇X Z⟩. (3.11)
This condition can be characterized as follows.
Proposition 3.23 (Characterization of metric connections). Let (M, g) be a Riemannian
manifold and ∇ an affine connection on M . Then ∇ is compatible with g (as in Def. 3.22) if
and only if g is parallel with respect to ∇, i.e., ∇g ≡ 0.
1Recall that, if g is fixed, we often simply write ⟨., .⟩ in place of g(., .).
 3.4. LEVI-CIVITA CONNECTION 39

Proof. By (3.9) and (3.6) the total covariant derivative of the symmetric 2-tensor g is
given by
(3.9) (3.6)
(∇g)(Y, Z, X) = (∇X g)(Y, Z) = X(g(Y, Z)) − g(∇X Y, Z) − g(Y, ∇X Z).
This expression vanishes for all X, Y, Z if and only if (3.11) holds. □
One can show that the condition of metric compatibility is not sufficient to fix a unique
connection on a Riemannian manifold, since any other affine connection whose difference
tensor is in some variables antisymmetric is also compatible to g [36, Prob. 5-1]. In fact, the
space of metric connections on a Riemannian manifold (of dimension at least 2) is an infinite-
dimensional affine space. This is why we have to include more properties of the tangential
connection.
3.4.2. Symmetric connections. On Euclidean space En the Lie bracket2 [X, Y ] of two
smooth vector fields X, Y can be written in the standard coordinates as
(3.4)
[X, Y ] = X(Y i )∂i − Y (X i )∂i = ∇X Y − ∇Y X.
Due to the coordinate-independence of the expression in terms of the Euclidean connection,
this property can be defined and studied for general affine connections. Importantly, we do
not need a Riemannian metric for this concept (in contrast to metric connections).
Definition 3.24. Let M be a smooth manifold and let ∇ be a connection on M . The
map τ : X(M ) × X(M ) → X(M ), defined by
τ (X, Y ) := ∇X Y − ∇Y X − [X, Y ], (3.12)
is called the torsion tensor of ∇.
We say that ∇ is symmetric (or torsion-free) if τ ≡ 0.

Problem 3.25. Let M be a smooth manifold and let ∇ be a connection on M , and let τ
be given by (3.12).
(i) Show that τ is a (1, 2)-tensor field (thus justifying the name torsion tensor).
(ii) Show that ∇ is symmetric if and only if Γkij = Γkji with respect to every coordinate
frame (not necessarily with respect to other frames!).
(iii) Show that ∇ is symmetric if and only if the covariant Hessian ∇2 u := ∇(du) (this
definition is inspired by the fact that ∇u(X) = ∇X u = X(u) = du(X)) of every
u ∈ C ∞ (M ) is a symmetric 2-tensor field. (See [36, Ex. 4.22] for a formula of the
covariant Hessian.)
Not only is the Euclidean connection ∇ symmetric, also the tangential connections ∇⊤
of submanifolds embedded in Rn is.
Exercise 3.26. Let M be an embedded submanifold of Euclidean space. Then the tan-
gential connection ∇⊤ is symmetric. (Hint: Extend the vector fields X, Y to the ambient
space as X,
e Ye , and use that the Lie bracket is natural [36, Prop. A.39]; in particular [X,
e Ye ]
is tangent to M and restricted to M equals [X, Y ].)

2Given two vector fields X, Y ∈ X(M ) the map [X, Y ] : C ∞ (M ) → C ∞ (M ), defined by [X, Y ](f ) :=
X(Y (f )) − Y (X(f )), is a derivation an thus [X, Y ] ∈ X(M ), called the Lie bracket of X and Y . Conceptually,
the Lie bracket [X, Y ] is a sort commutator of vector fields, and also the derivative of Y along the flow generated
by X, and the Lie derivative LX Y of Y along X.
40 3. CONNECTIONS AND COVARIANT DIFFERENTATION

3.4.3. Fundamental Theorem of Riemannian Geometry. We have started out this

section by proposing that on Riemannian manifolds we want to mimick the properties of
the tangential connection ∇⊥ of embedded submanifolds. We have found that ∇⊥ is both
compatible with the induced Riemannian metric as well as symmetric. The following result
states that both of these properties combined already uniquely determine a connection on a
Riemannian manifold.
Theorem 3.27 (Fundamental Theorem of Riemannian Geometry). Let (M, g) be a Rie-
mannian manifold. There exists a unique affine connection ∇, called the Levi-Civita connec-
tion, that in addition to (i)–(iii) of Definition 3.1 satisfies
(iv) ∇ is symmetric: for X, Y ∈ X(M ),
[X, Y ] = ∇X Y − ∇Y X.
(v) ∇ is compatible with g: for X, Y, Z ∈ X(M );
∇X ⟨Y, Z⟩ = ⟨∇X Y, Z⟩ + ⟨Y, ∇X Z⟩.
Moreover, the Levi-Civita connection ∇ satisfies the Koszul formula, that is,
1
⟨∇X Y, Z⟩ = X⟨Y, Z⟩ + Y ⟨Z, X⟩ − Z⟨X, Y ⟩
2

− ⟨Y, [X, Z]⟩ − ⟨Z, [Y, X]⟩ + ⟨X, [Z, Y ]⟩ , (3.13)
for smooth vector fields X, Y, Z ∈ X(M ).
Proof. Step 1. Uniqueness and Koszul formula. We prove uniqueness by deriving an
explicit formula for ∇. For X, Y, Z ∈ X(M ) we have
(v) (iv)
X⟨Y, Z⟩ = ⟨∇X Y, Z⟩ + ⟨Y, ∇X Z⟩ = ⟨∇X Y, Z⟩ + ⟨Y, ∇Z X⟩ + ⟨Y, [X, Z]⟩.
Cyclic permuation and adding up/substracting implies
X⟨Y, Z⟩ + Y ⟨Z, X⟩ − Z⟨X, Y ⟩ = ⟨∇X Y, Z⟩ + ⟨Y, ∇Z X⟩ + ⟨Y, [X, Z]⟩
+ ⟨∇Y Z, X⟩ + ⟨Z, ∇X Y ⟩ + ⟨Z, [Y, X]⟩
− ⟨∇Z X, Y ⟩ − ⟨X, ∇Y Z⟩ − ⟨X, [Z, Y ]⟩
= 2⟨∇X Y, Z⟩ + ⟨Y, [X, Z]⟩ + ⟨Z, [Y, X]⟩ − ⟨X, [Z, Y ]⟩.
Hence we can solve for ⟨∇X Y, Z⟩, and obtain the Koszul formula
1
⟨∇X Y, Z⟩ = (X⟨Y, Z⟩ + Y ⟨Z, X⟩ − Z⟨X, Y ⟩ − ⟨Y, [X, Z]⟩ − ⟨Z, [Y, X]⟩ + ⟨X, [Z, Y ]⟩) ,
2
with a right hand side that does not depend on the connection ∇. Hence for any two connec-
tions ∇1 , ∇2 satisfying (i)–(v), and all vector fields X, Y, Z ∈ X(M ), we have
⟨∇1X Y − ∇2X Y, Z⟩ = 0,
thus (by injectivity of the musical isomorphism ♭) ∇1X Y = ∇2X Y and hence ∇1 = ∇2 is
uniquely determined.
Step 2. Existence. Let F (X, Y, Z) be the right hand side of (3.13), i.e.,
1

F (X, Y, Z) := X⟨Y, Z⟩ + Y ⟨Z, X⟩ − Z⟨X, Y ⟩ − ⟨Y, [X, Z]⟩ − ⟨Z, [Y, X]⟩ + ⟨X, [Z, Y ]⟩ .
2
For fixed X, Y ∈ X(M ) the map Z 7→ F (X, Y, Z) is C ∞ (M )-linear (check!), hence an element
ω(X,Y ) ∈ Ω1 (M ). By the musical isomorphism ♯ there exists a unique metrically equivalent
 3.4. LEVI-CIVITA CONNECTION 41

♯ ♯
vector field ω(X,Y ) ∈ X(M ), i.e., ⟨ω(X,Y ) , Z⟩ = F (X, Y, Z) for all Z ∈ X(M ). We denote
♯
∇X Y := ω(X,Y ) and check that it satisfies all desired properties (i)–(v) of a symmetric metric
connection. □

Problem 3.28. Finish the proof of Theorem 3.27, that is, prove that Z 7→ F (X, Y, Z)
is a 1-form and that the expression ∇X Y defined in Step 2 of the proof indeed satisfies all
properties (i)–(v) from Definition 3.1 and Theorem 3.27.
Whenever we are in the setting of Riemannian Geometry (thus, almost always in this
course), we assume that ∇ is the Levi-Civita connection of the Riemannian metric g. As
desired, we have managed to single out the natural connections in Euclidean space.
Example 3.29 (Euclidean connection). The Levi-Civita connection on the Euclidean
space En is the Euclidean connection ∇.
Example 3.30 (Tangential connection). Let M be an embedded submanifold of En . By
Exercise 3.26 the tangential connection ∇⊤ (as defined in Example 3.9) is symmetric, and by
Exercise 3.21 it is compatible with the induced metric. Hence, due to uniqueness, ∇⊤ is the
Levi-Civita connection on M .
Similar to the local computations necessary for the tangential connection in Exercise 3.9
we can use Lemma 3.4 (and [36, Prop. 4.2]) to restrict ∇ to a (coordinate) neighborhood
U (of p) and obtain a well-defined element ∇X Y ∈ X(U ). The connection coefficients of
the Levi-Civita connection (which, by default, is the connection we use on a Riemannian
manifold) in coordinates have a particular name.
Definition 3.31. Let (U, (xi )) be a coordinate chart of a Riemannian manifold (M, g).
The Christoffel symbols of g with respect to this chart are the smooth functions Γkij : U → R
∂
such that (with coordinate vector fields ∂i = ∂x i)

∇∂i ∂j =: Γkij ∂k . (3.14)

Since, the Lie bracket vanishes for coordinate vector fields, i.e.,
[∂i , ∂j ] = 0,
it immediately follows from the symmetric condition (iv) (see also Problem 3.25) that
Γkij = Γkji .
Recall, however, that ∇ is not a tensor field and so the Christoffel symbols do not transform
in the usual way. Nonetheless, they are easy compute via the metric coefficients
gij = ⟨∂i , ∂j ⟩
and their inverses.
Corollary 3.32. Let (U, (xi )) be a coordinate chart on a Riemannian manifold (M, g).
The Christoffel symbols are given by
1
Γkij = g kl (∂i gjl + ∂j gil − ∂l gij ). (3.15)
2
42 3. CONNECTIONS AND COVARIANT DIFFERENTATION

Proof. By the Koszul formula (3.13), taking into account that [∂i , ∂j ] = 0, we obtain
1
⟨∇∂i ∂j , ∂l ⟩ = (∂i ⟨∂j , ∂l ⟩ + ∂j ⟨∂l , ∂i ⟩ − ∂l ⟨∂i , ∂j ⟩).
2
In terms of the coefficients this reads
1
Γmij gml = (∂i gjl + ∂j gil − ∂l gij ),
2
and multiplying both formulas with the inverse matrix g kl , that is, gml g kl = δlk , yields (3.15).
□
Of course, all general formulas that we have derived about affine connections and their
coefficients still hold. For instance, we can can compute the Levi-Civita covariant derivative
∇X Y of Y in direction X via the local formula (3.3). Similarly, the Levi-Civita connection in
T M extends to a unique tensor covariant derivative ∇X in T (k,l) T M via ∇X f = X(f ) etc.
(as shown in Proposition 3.16) and allows us to define a total covariant differential ∇ of a
tensor field (via Definition 3.18).
An important consequence of the coordinate-independent definition of the Levi-Civita
connection is that it respects isometries (via pullbacks).
Proposition 3.33 (Naturality of the Levi-Civita connection). Suppose (M, g) and (M f, ge)
are Riemannian (or semi-Riemannian) manifolds, and let ∇ and ∇ e denote the Levi-Civita
f is an isometry, then φ∗ ∇
connection of g and ge, respectively. If φ : M → M e = ∇, where φ∗ ∇e
is the pullback of ∇ by φ defined by
e
(φ∗ ∇)
e X Y := (φ−1 )∗ (∇
e φ∗ X (φ∗ Y )).

The proof is straight-forward and can be found in [36, Prop. 5.13] based on the earlier
result [36, Lem. 4.37] that the pullback connection is indeed a connection.

3.5. Parallel transport

The basic idea behind using connections is that it provides us with a tool to “parallel
transport” vectors, meaning that we want to effordlessly move a tangent vector from one
point to another along curves on a manifold. This idea will be crucial for the definition of
curvature, and we will explore it in detail now. Note that one can study parallel transport
with respect to any connection (even in vector bundles) but towards the end of this section
we will already see that on Riemannian manifolds there is more to gain.
3.5.1. Parallel vector and tensor fields. Before we discuss how to transport some-
thing parallel along a curve, we first want to understand what it means for a vector field (or
tensor field) to be parallel on all of M . Afterwards we study this notion on curves.
Definition 3.34. Let M be a smooth manifold with affine connection ∇. A vector field
X on M is called parallel if ∇Y X = 0 for all Y ∈ X(M ).
Example 3.35. The coordinate vector fields on Rn are parallel (with respect to the
Euclidean connection) because for any Y = Y j ∂j we obtain ∇Y ∂i = Y j ∇∂j ∂i = 0. In
fact, precisely all constant vector fields are parallel on Rn . (The same result holds for semi-
Euclidean space Rν,n−ν .)

Problem 3.36. Suppose G is a Lie group.

 3.5. PARALLEL TRANSPORT 43

(i) Show that there is a unique connection3 ∇ in T G with the property that every
left-invariant vector field is parallel.
(ii) Show that the torsion tensor of ∇ is zero if and only if G is abelian.
The following result shows that one can use the total covariant derivative ∇ of Defini-
tion 3.18 to characterize parallel vector and tensor fields. In fact, we have already mentioned
this when characterizing metric connections in Proposition 3.23.
Proposition 3.37. Let M be a smooth manifold with affine connection ∇. Then a smooth
vector (or tensor) field A is parallel on M if and only if ∇A ≡ 0.

Problem 3.38. Prove Propostion 3.37.

3.5.2. Vector and tensor fields along curves. Our initial motivation for introducing
connections was that we want to make sense of the derivative of a vector field along a curve,
and we want to study vector fields that are parallel along curves. Thus we first have to say
what we even mean by a vector field along a curve, and how to use the covariant derivative
in this context. We first pin down what we mean by such a vector field.
Definition 3.39. Let M be a smooth manifold and let γ : I → M be a smooth curve. A
smooth 4 vector field along γ is a smooth map V : I → T M such that V (t) ∈ Tγ(t) M .
By X(γ) we denote the set of all smooth vector fields along γ.
The set X(γ) is real vector space with respect to pointwise addition scalar multiplication.
In fact, it is a module over C ∞ (M ) with respect to pointwise multiplication, i.e., (f X)(t) :=
f (t)X(t).
Remark 3.40 (Vector fields along functions). Apart from curves, on can also consider
vector fields along functions between arbitrary smooth manifolds M and N . A vector field
along f ∈ C ∞ (N, M ) is a smooth map Z : N → T M such that π ◦ Z = f , where π : T M → M
is the natural projection. Similarly, we write X(f ) for vector fields along f , which also is a
C ∞ (N )-module. Vector fields along curves appear as the special case where N is an interval
in R.
Example 3.41 (Velocity field of a curve). If γ is a smooth curve on M , then γ ′ (t) ∈ Tγ(t) M
and smooth because locally the velocity γ ′ (t) is of the form
γ ′ (t) = γ̇ 1 (t)∂1 |γ(t) + . . . + γ̇ n (t)∂n |γ(t) , (3.16)
hence γ′ ∈ X(γ).
Example 3.42 (Normal field of a curve). If γ is a curve in R2 , consider N (t) := Rγ ′ (t),
where R is the counterclockwise rotation by π2 , i.e., N (t) = (−γ̇ 2 (t), γ̇ 1 (t)). Thus N is smooth
and hence N ∈ X(γ).
Example 3.43 (Restriction of vector field to curve). If γ : I → M is a smooth curve on M
and Ve a smooth vector field defined on an open subset containing γ(I). Then the restriction
V := Ve ◦ γ ∈ X(γ).
3This is usually not the Levi-Civita connection.
4Often continuity is sufficient, but we will restrict ourselves to the smooth setting.
44 3. CONNECTIONS AND COVARIANT DIFFERENTATION

While every vector field can be restricted to a smooth curve, not every vector field V
along a curve can be extended to a vector field Ve in a neighborhood of the curve (as seen in
Figure 3.1, this is clearly impossible if there exists γ(t1 ) = γ(t2 ) but γ ′ (t1 ) ̸= γ ′ (t2 )). This
prompts the following definition.
Definition 3.44. A smooth vector field V along a curve γ on M is extendible if there
exists a smooth vector field Ve on a neighoorhood of γ(I) such that V = Ve ◦ γ.
Tensor fields along curves, and their extendibility, are defined in the same way.

Figure 3.1. A restricted/extendible and nonextendible vector field along a curve.

Problem 3.45. Consider the figure eight curve γ : (−π, π) → R2 , defined by

γ(t) := (sin 2t, sin t).
Show that γ is an injective smooth immersion, but that the velocity vector field γ ′ is not
extendible.

3.5.3. Covariant derivative along curves. The following result gives a bit more in-
sight into why and how an affine connection gives rise to a derivative of vector fields along
curves, thereby also giving meaning to the acceleration of a curve in M (relevant in Chapter 4).
The same result holds also for tensor fields along curves.
Theorem 3.46. Let M be a smooth manifold with an affine connection ∇. Then each
smooth curve γ : I → M gives rise to a unique operator
Dt : X(γ) → X(γ),
called the covariant derivative along γ, satisfying the following properties for V, W ∈ X(γ),
a, b ∈ R, and f ∈ C ∞ (I):
(i) linearity over R: Dt (aV + bW ) = aDt (V ) + bDt (W ),
(ii) product rule: Dt (f V ) = f ′ V + f Dt V ,
(iii) If V ∈ X(γ) is extendible, then for every extension Ve of V ,
Dt V = ∇γ ′ (t) Ve .
 3.5. PARALLEL TRANSPORT 45

Remark 3.47. Note that property (iii) makes sense because ∇X Y (p) only depends on
the value of Xp and the value Y along a curve tangent to X at p.
Proof. Step 1. Uniqueness. Suppose Dt is such an operator satisfying (i)–(iii). Let
t0 ∈ I arbitrary. Similar to the proof of Lemma 3.4 one can see that the value Dt V |t0 only
depends on the value of V in a small interval (t0 − ε, t0 + ε) (if t0 is an endpoint extend the
coordinate representation of γ beyond t0 , show the result, and restrict back to I). For smooth
coordinates (xi ) on M in a neighborhood of γ(t0 ) we can write
V (t) = V j (t)∂j |γ(t)
for t near t0 . By the properties of Dt and ∇ we have
(ii),(iii)
Dt V (t) = V̇ j (t)∂j |γ(t) + V j (t)∇γ ′ (t) ∂j |γ(t)
(3.2),(3.16)
 
= V̇ k (t) + γ̇ i (t)V j (t)Γkij (γ(t)) ∂k |γ(t) , (3.17)

and hence uniqueness (if it exists).

Step 2. Existence. If γ(I) is contained in a single coordinate chart, simply define Dt by
(3.17) and check that (i)–(iii) hold. In the general case we can cover γ(I) by coordinate charts
and check that the definitions of Dt V agree when the charts overlap. □

Exercise 3.48. Complete the proof of Theorem 3.46 by showing that Dt as defined in
(3.17) indeed satisfies properties (i)–(iii).

Corollary 3.49. Suppose γ is a smooth curve on a Riemannian manifold (M, g) and

Dt the induced covariant derivative of Theorem 3.46. Then, in addition to (i)–(iii), Dt also
satisfies
d
(iv) dt ⟨V, W ⟩ = ⟨Dt V, W ⟩ + ⟨V, Dt W ⟩.

Problem 3.50. Prove Corollary 3.49. (Hint: Use (3.17) and the compatibility of the
Levi-Civita connection with g.)

3.5.4. Parallel transport along curves. The idea of parallel transport is that of mov-
ing around tangent vectors along a curve in an effordless way. The notion is made precise
by restricting the initial Definition 3.34 of parallel vector fields to a curve with help of the
induced covariant derivative Dt .
Definition 3.51. Let M be a smooth manifold with affine connection ∇. A smooth
vector field V along a smooth curve γ is said to be parallel along γ (with respect to ∇) if
Dt V ≡ 0.

Exercise 3.52. Let γ : I → Rn be a smooth curve, and let V be a smooth vector field
along γ. Show that V is parallel along γ (with respect to the Euclidean connection ∇) if and
only if its component functions (with respect to the standard basis) are constants.
In (3.17) we have derived a local formula for Dt V in terms of the connection coefficients
Γkij and the coefficients V j of V . Based on this local differential identity Dt V (t) = 0 we can
prove that we can indeed parallel transport vectors along curves.
46 3. CONNECTIONS AND COVARIANT DIFFERENTATION

Figure 3.2. A vector field that is parallel along a curve in R2 .

Theorem 3.53 (Existence and uniqueness of parallel fields). Suppose γ : I → M is a

smooth curve on a smooth manifold M with affine connection ∇ (and induced covariant
derivative Dt along γ). If t0 ∈ I and v ∈ Tγ(t0 ) M , then there is a unique parallel vector field
V ∈ X(γ), called the parallel transport of v along γ, defined on all of I with V (t0 ) = v.
Proof. In local coordinates V has to satisfy (3.17)= 0, i.e.,
 
Dt V (t) = V̇ k (t) + γ̇ i (t)V j (t)Γkij (γ(t)) ∂k |γ(t) = 0.
This is a system of first-order linear ordinary differential equations for the coefficients V 1 , . . . , V n ,
which possesses a unique solution on the entire interval for given initial data V 1 (t0 ), . . . , V n (t0 )
(see [36, Thm. 4.31] to recall the required result from the theory of ordinary differential equa-
tions). The claim thus follows by covering γ(I) with chart domains. □
Theorem 3.53 gives rise to the following definition.
Definition 3.54. Let M be a smooth manifold with affine connection ∇. Let γ : I → M
be a smooth curve. For each t0 , t1 ∈ I the map
P γ = Ptγ0 t1 : Tγ(t0 ) M → Tγ(t1 ) M,
v 7→ V (t1 ),
where V is the unique vector field obtained in Theorem 3.53, is called the parallel transport
map.
It is straightforward to extend this result and definition to admissible curves γ : [a, b] →
M and thus piecewise smooth vector fields V ∈ X(γ): In [36, Cor. 4.33] it is shown that
Theorem 3.53 immediately yields a parallel transport along piecewise smooth curves, the
only difference being that the vector field obtained is continuous everywhere but smooth only
where γ is.
In fact, one can obtain a parallel frame along γ in the following way: By transporting
any basis (b1 , . . . , bn ) for Tγ(t0 ) M along γ, one obtains an n-tuple of parallel vector fields
(E1 , . . . , En ) along γ. By Proposition 3.55 below these vector fields again form a basis (Ei (t))
for Tγ(t) M at each point γ(t). Then each vector field V ∈ X(γ) in such a frame reads
V (t) = V i (t)Ei (t), and since the Ei ’s are parallel it is easy to compute the covariant derivative
along γ, i.e.,
Dt V (t) = V̇ i (t)Ei (t), (3.18)
 3.5. PARALLEL TRANSPORT 47

where γ and V are smooth.

Proposition 3.55. Let M be a smooth manifold with affine connection ∇. Suppose
γ : [a, b] → M is a smooth curve on M . Then the parallel transport P γ : Tγ(a) M → Tγ(b) M is
an invertible linear map.
Corollary 3.56. If (M, g) is a Riemannian manifold, then P γ is a linear isometry.
Corollary 3.56 implies that any orthonormal basis at a point on γ can be extended to a
parallel orthonormal frame along γ (see Figure 3.3 and [36, Prop. 5.5]).

Figure 3.3. A parallel orthonormal frame along a curve in R2 .

Problem 3.57. (i) Prove Proposition 3.55. (Hint: Consider v1 , v2 ∈ Tp M and

corresponding parallel vector fields V1 , V2 ∈ X(γ). Then V1 + V2 is also parallel and
P γ (v1 + v2 ) = P γ (v1 ) + P γ (v2 ) by unique solvability of the ordinary differential
equation. Analogously for av1 , a ∈ R. Show that P γ is injective and use that
dim Tp M = dim Tq M to argue that it is bijective.)
(ii) Prove Corollary 3.56. (Hint: Show, in addition, that ⟨P γ (v1 ), P γ (v2 )⟩ = ⟨v1 , v2 ⟩.)
We have now seen that an affine connection really determines a way of differentiation
and a notion for parallel transporting vectors along curves, which really “connects” nearby
tangent spaces. It is natural to ask whether parallel transport also uniquely determines the
covariant derivative along a curve and ultimately the connection. This is indeed the case.
Theorem 3.58 (Parallel transport determines covariant differentiation). Let M be a
smooth manifold with affine connection ∇. Suppose γ : I → M is a smooth curve and
V ∈ X(γ). Then
Ptγ1 t0 V (t1 ) − V (t0 )
Dt V (t0 ) := lim , t0 ∈ I. (3.19)
t1 →t0 t1 − t0
Proof. If (Ei ) is a parallel frame along γ, then V = V j Ej and by (3.18) thus, on the
one hand, Dt V (t0 ) = V̇ i (t0 )Ei (t0 ).
On the other hand, the parallel vector field of V (t1 ) along γ is the constant coeffi-
cient field V i (t1 )Ei (t), and thus Ptγ0 t1 V (t1 ) = V i (t1 )Ei (t0 ). Taking the limit yields also
Pγ V (t1 )−V (t0 )
limt1 →t0 t1 t0 t1 −t0 = V̇ i (t0 )Ei (t0 ). Hence both expressions on the left and right of (3.19)
are equal. □

By restricting a vector field Y to γ, Theorem 3.46(iii) implies that if γ(0) = p and

γ ′ (0) = Xp then ∇X Y |p = Dt Y (γ(0)). Hence Theorem 3.53 immediately implies a result for
the connection everywhere.
48 3. CONNECTIONS AND COVARIANT DIFFERENTATION

Corollary 3.59 (Parallel transport detMermines the connection). Let M be a smooth

manifold with affine connection ∇. Suppose X, Y ∈ X(M ). Then, for every p ∈ M ,
γ
Ph0 Yγ(h) − Yp
∇X Y |p = lim , (3.20)
h→0 h
where γ : I → M is a smooth curve through γ(0) = p with γ ′ (0) = Xp . □

Exercise 3.60. In Definition 3.34 we said that a vector field X is parallel if ∇Y X = 0

for all Y ∈ X(M ). Show that this is equivalent to X being parallel along every smooth curve
in M . (Remark : This is how Lee [36, p. 110] defines parallel vector fields.)
We mostly deal with parallel vector fields, however, also parallel differential forms natu-
rally occur as the following exercise shows.
Exercise 3.61. Let (M, g) be an oriented Riemannian manifold. The Riemannian volume
form dVg is parallel with respect to the Levi-Civita connection. (Use Exercise 3.60 and choose
a parallel oriented orthonormal frame along each curve.)
It is crucial to note that while it is always possible to extend a vector at a point on a
curve to a parallel vector field along it by Theorem 3.53, it is generally not possible to extend
it to a parallel vector field on an open subset of a point. In fact, the impossibility of this is
intimately tied to curvature, as we will soon see.
One immediate consequence of the failure of a connection “to commute”, even without
defining curvature explictely, can be seen by studying admissible loops, i.e., piecewise smooth
curves γ : [a, b] → M with γ(a) = γ(b) = p. This is based on the fact that the corresponding
parallel transport defines an invertible linear map on Tp M via Proposition 3.55, i.e., P γ ∈
GL(Tp M ).
Definition 3.62. Let M be a smooth manifold. The holonomy group of an affine con-
nection ∇ at a point p ∈ M is defined by
Hol(p) := {P γ ∈ GL(Tp M ); γ is an admissible loop at p}.
The restricted holonomy group Hol0 (p) based at p is the subgroup of Hol(p) coming from
contractible loops γ, i.e., loops that are path-homotopic to the constant loop.

Problem 3.63. Let (M, g) be a connected Riemannian manifold and p ∈ M .

(i) Show that Hol(p) is a subgroup of O(Tp M ) (the set of all linear isometries on Tp M ).
(ii) Show that Hol0 (p) is a normal subgroup of Hol(p).
(iii) Show that the holonomy group depends on the base point only up to conjugation
with an element in GL(Rn ). (Hint: Use that M is connected and parallel transport
along an admissible path γ from p to q, i.e., explicitely Hol(q) = P γ Hol(p)(P γ )−1 .
With this understanding one often drops the base point and just writes Hol(M ).)
(iv) Show that M is orientable if and only if Hol(p) ⊆ SO(Tp M ) (the set of all isometries
with determinant +1) for some p ∈ M .
(v) Show that g is flat (locally isometric to Euclidean space) if and only if Hol0 (p) is
the trivial group for some p ∈ M .
 3.5. PARALLEL TRANSPORT 49

Remark 3.64 (Berger’s list). In 1955, Marcel Berger classified all possible holonomy
groups for simply connected n-dimensional Riemannian manifolds which are irreducible (not
locally a product space) and nonsymmetric (not locally a Riemannian symmetric space).
There are only 7 of those: SO(n), U( n2 ), SU( n2 ), Sp( n4 ), Sp( n4 )Sp(1), G2 , and Spin(7). Several
of those are Ricci flat. See Berger’s own summary in [7, Sec. 13.4] for more background.
Riemannian manifolds with special holonomy play an important role in string theory com-
pactifications, for instance, Calabi–Yau manifolds.
Holonomy groups are important in all kinds of settings, also in connection with vector
bundles. The Ambrose–Singer Theorem from 1953, for instance, relates the holonomy of a
connection in a principal bundle (equipped with a group action) with the curvature form of
the connection.
 CHAPTER 4

Geodesics

By equipping a smooth manifold with an affine connection we obtained a tool to parallel

transport vectors along curves, from one tangent space to another. We have seen that on
Riemannian manifolds equipped with the canonical Levi-Civita connection parallel transport
is actually an isometry.
Given that the velocity vector of a curve γ : I → M is a special vector field along this
curve (recall Example 3.41) it is natural to ask when this velocity vector field γ ′ is itself
parallel along γ. Such curves with “zero acceleration” are called geodesics. In principle, one
can study them on any manifold with affine connection, but on Riemannian manifolds (M, g)
they provide us with a fundamental tool to understand the underlying geometry. For instance,
geodesics are also “locally length-minimizing” curves with respect to the induced Riemannian
distance dg (recall Definition 2.74). Strictly speaking, we will only now prove that dg indeed
gives rise to a metric space structure on M .
Essentially all of the local behavior of geodesics is encoded in the exponential map (with
corresponding special coordinates and neighborhoods), which provides us with a powerful
theoretical as well as computational tool. It is the aim of this section to rigorously introduce
and analyze geodesics and all their related standard concepts.

4.1. Geodesics
4.1.1. Basic definitions and existence. We have already defined the velocity of a
curve γ in Example 3.41. By covariant differentiation along γ we can also define the acceler-
ation.
Definition 4.1. Let M be a smooth manifold with affine connection ∇ and let γ : I → M
be a smooth curve. The acceleration of γ is the vector field Dt γ ′ along γ.
Curves where the velocity γ ′ is parallel along γ have a special name.
Definition 4.2. Let M be a smooth manifold with affine connection ∇. A smooth curve
γ : I → M is called geodesic (with respect to ∇) if its acceleration is zero, i.e., Dt γ ′ ≡ 0.

Exercise 4.3. Show that for a geodesic γ the length of the tangent vector |γ ′ (t)| is
constant.
In smooth coordinates (x1 , . . . , xn ), if we write the component functions of γ (by abuse
of notation1) as γ(t) = x(t) = (x1 (t), . . . , xn (t)), it follows immediately from (3.17) that γ is
a geodesic if and only if the component functions satisfy the geodesic equation
ẍk (t) + ẋi (t)ẋj (t)Γkij (x(t)) = 0, 1 ≤ k ≤ n. (4.1)
1More carefully one should actually write (x ◦ γ)(t) = ((x1 ◦ γ)(t), . . . , (xn ◦ γ)(t)) or = (γ 1 (t), . . . , γ n (t)).

51
52 4. GEODESICS

This is a system of n second-order ordinary differential equations and the standard Picard–
Lindelöf Theorem from ODE theory implies existence and uniqueness of solutions to the
following initial value problem (this result is more involved for manifolds with boundary).
Theorem 4.4 (Existence and uniqueness of geodesics). Let M be a smooth manifold with
affine connection ∇. For every p ∈ M , w ∈ Tp M , and t0 ∈ R, there exists an open interval
I ⊆ R containing t0 and a geodesic γ : I → M satisfying γ(t0 ) = p and γ ′ (t0 ) = w. Any two
such geodesics agree on their common domain.
The proof follows a standard argument, where the only additional step needed is the
rewriting of (4.1) as a system of 2n first-order equations in local charts.
Proof. Step 1. Local existence and uniqueness. Let (xi ) be smooth coordinates on some
neighborhood U of p. A smooth curve γ(t) = (x1 (t), . . . , xn (t)) in U is a geodesics if and only
if it satisfies the geodesic equation (4.1). By introducing auxiliary variables v i = ẋi the n
second-order equations (4.1) can be written as a system of 2n first-order equations,
ẋk (t) = v k (t), (4.2)
v̇ k (t) = −v i (t)v j (t)Γkij (x(t)), (4.3)
for the variables (x1 , . . . , xn , v 1 , . . . , v n ) on U × Rn . In other words, the equations (4.2)–(4.3)
are the equations of flow for the vector field G ∈ X(U × Rn ) given by
∂ ∂
G(x,v) = v k − v i (t)v j (t)Γkij (x(t)) . (4.4)
∂xk (x,v) ∂v k (x,v)

By the fundamental theorem of flows [35, Thm. 9.12], for each (p, w) ∈ U × Rn and t0 ∈ R,
there exists an open interval I0 containing t0 and a unique smooth solution ζ : I0 → U × Rn ,
ζ(t) = (xi (t), v i (t)), to this system satisfying the initial conditions ζ(t0 ) = (p, w). The curve
γ(t) := (x1 (t), . . . , xn (t)) in U is thus the desired solution.
Step 2. Global uniqueness. If there are two geodesics γ, γ̃ : I → M with the same initial
conditions γ(t0 ) = γ̃(t0 ) and γ ′ (t0 ) = γ̃ ′ (t0 ), then due to the uniqueness of ODE solutions,
they agree on an interval (t0 − ε, t0 + ε). Suppose
β := inf{b ∈ I; γ(b) ̸= γ̃(b)}.
Clearly, β > t0 , and by continuity, γ(β) = γ̃(β) as well as γ ′ (β) = γ̃ ′ (β). Thus by local
uniqueness we must have equality also in a neighborhood of β, a contradiction to it being the
infimum. □
Definition 4.5. Let M be a smooth manifold with affine connection ∇. A geodesic
γ : I → M is called maximal if there exists no geodesic γ
e : Ie → M with I ⫋ Ie and γ
e|I = γ.
A geodesic segment is a geodesic with compact domain.
Definition 4.6. A smooth manifold with affine connection is called geodesically complete
if each maximal geodesic is defined on all of R.
Theorem 4.4 immediately implies the existence of maximal geodesics.
Corollary 4.7. Let M be a smooth manifold with affine connection ∇. For each p ∈ M
and v ∈ Tp M , there exists a unique maximal geodesic γ : I → M with γ(0) = p and γ ′ (0) = v,
defined on some open interval I containing 0. □
 4.1. GEODESICS 53

The unique maximal geodesic of Corollary 4.7 is often called the geodesic with initial point
p and initial velocity v, and denoted by γv . The point p is omitted because it can be recovered
from the natural projection π : T M → M via p = π(v).
Exercise 4.8 (Geodesics in Rn ). Show that the maximal geodesics on Rn with the Eu-
clidean connection ∇ (see Example 3.6) are exactly the constant curves and the straight lines
with constant-speed parametrizations. See Figure 4.1 (Same in Rν,n−ν .)

Figure 4.1. The maximal geodesics in Rn are straight lines with constant speed w.

Exercise 4.9 (Geodesics on a cylinder). Let M be cylinder in R3 with radius 1 and

consider the chart
(cos θ, sin θ, z) 7→ (θ, z), θ ∈ (0, 2π), z ∈ R.
Show that the geodesics on the cylinder are generally helixes, with special cases circles of
latitute and generating straight lines. (Hint: Either compute and solve the geodesic equation
directly in these coordinates, or “unwinde” M in R2 via an isometry then use the solutions
of Exercise 4.8).

Problem 4.10. See [36, Problem 5-4] about geodesics on surfaces of revolution.

Problem 4.11. Suppose (M1 , g1 ) and (M2 , g2 ) are Riemannian manifolds

(i) Prove that if M1 × M2 is endowed with the product metric, then a curve γ : I →
M1 × M2 of the form γ(t) = (γ1 (t), γ2 (t)) is a geodesic if and only if γi is a geodesic
in (Mi , gi ) for i = 1, 2.
(ii) Suppose f : M1 → R∗ is a strictly positive smooth function, and M1 ×f M2 is the
resulting warped product. Let γ1 : I → M1 be a smooth curve and q0 a point in M2 ,
and define γ : I → M1 ×f M2 by γ(t) := (γ1 (t), q0 ). Prove that γ is a geodesic with
respect to the warped product metric if and only if γ1 is a geodesic with respect to
g1 .
54 4. GEODESICS

4.1.2. Geodesic flow. By using the projection π : T M → M we can reformulate the

approach via the vector field G used in the proof of Theorem 4.4 in a more invariant way.
The importance of G is that it actually defines a global vector field on T M .
Theorem 4.12. Let M be a smooth manifold with affine connection ∇. Then there is a
unique vector field G ∈ X(T M ), called the geodesic vector field, with the property that the
projection π : T M → M provides a one-to-one correspondence between (maximal) integral
curves of G and (maximal) geodesics on M .
Proof. Step 1. Existence and properties of G in a chart. Let (xi ) be any smooth
local coordinates on an open set U ⊆ M , and let (xi , v i ) be the associated natural coordi-
nates on π −1 (U ), that is, by following the local trivialization Φ : U × Rn → Rn × Rn with
(p, (v 1 , . . . , v n )) 7→ (xi (p), v i (p)). Thus G as defined in (4.4) is a smooth vector field on
π −1 (U ) ⊆ T M , and the integral curves η(t) = (xi (t), v i (t)) of G satisfy (4.2)–(4.3), which
was equivalent to the geodesic equation under the substitution v k = ẋk (as seen in Theo-
rem 4.4). In other words: Every integral curve of G on π −1 (U ) projects to a geodesic under
π : T M → M (just π(x, v) = x); conversely, every geodesic γ(t) = (x1 (t), . . . , xn (t)) in U lifts
to an integral curve of G in π −1 (U ) by setting v i (t) = ẋi (t).
Step 2. Global existence, properties, and uniqueness of G. To obtain a global vector field
G we can simply patch together the chart versions of Step 1. That this is possible rests on
the fact that G can be shown to be coordinate independent. One (tedious) way to prove this
is to show that G behaves well where the charts overlap. Instead, we show that G acts on a
function f ∈ C ∞ (T M ) by
d
Gf (p, v) = f (γv (t), γv′ (t)). (4.5)
dt t=0

Due to the coordinate independence of this formula, the various coordinate dependend defi-
nitions of G given by (4.4) must then agree (and G is unique).
To prove (4.5) we use the usual notation for the components xi (t) of a geodesic γv and its
velocity as v i (t) = ẋi (t). The chain rule and geodesic equation (4.2)–(4.3) allow us to rewrite
the right hand side of (4.5) as
 
d ′ ∂f k ∂f k
f (γv (t), γv (t)) = (x(t), v(t))ẋ (t) + k (x(t), v(t))v̇ (t)
dt t=0 ∂xk ∂v t=0
∂f ∂f
= (p, v)v k − k (p, v)v i v j Γkij (p) = Gf (p, v).
∂xk ∂v
By the fundamental theorem of flows [35, Thm. 9.12] there exists an open neighborhood
D ⊆ R × T M containing {0} × T M and a smooth map θ : D → T M such that each curve

θ(p,v) (t) = θ(t, (p, v))

is the unique maximal integral curve of G starting at (p, v), defined on the open interval
containing 0. By the local description (4.4) of G and Theorem 4.4 these are precisely the
(maximal) geodesics. □

Based on the properties of the pullback connection in Proposition 3.33, together with the
fact that being a geodesic is a local property one can directly show the following.
 4.1. GEODESICS 55

Proposition 4.13 (Naturality of geodesics). Suppose (M, g) and (M

f, ge) are Riemannian
manifolds, and φ : M → Mf is a local isometry. If γ is a geodesic in M , then φ ◦ γ is a
geodesics in M .
f □
4.1.3. Geodesics on submanifolds. To compute geodesics of a submanifold M em-
bedded in Rn one uses the covariant derivative along a curve γ : I → M with respect to the
tangential connection ∇⊤ , which by a standard calculation in [36, Prop. 5.1] is shown to be
given by
Dt⊤ V (t) = π ⊤ (Dt V (t)). (4.6)
The geodesics on M then satisfy the following constraints.
Proposition 4.14. Suppose M ⊆ Rn is an embedded submanifold. A smooth curve
γ : I → Rn is a geodesic with respect to the tangential connection ∇⊤ on M if and only if its
ordinary acceleration γ ′′ (t) is orthogonal to Tγ(t) M for all t ∈ I.
Proof. The Christoffel symbols of the Euclidean connection are all zero. Thus by (3.17)
Dt γ ′ (t) = γ ′′ (t).
The orthogonality of γ ′′ (t) to Tγ(t) M follows from (4.6) together with the condition Dt⊤ γ ′ (t) ≡
0 for a geodesic. □
Recall that the model spaces En , Sn (R), and Hn (R) are highly symmetric. While it is
possible to compute their geodesics via the geodesic equation, it is more insightful to use their
symmetries and Proposition 4.14 for embedded submanifolds. We will give a short overview
here. More details can be found in [36, p. 136–145].
Example 4.15 (Euclidean space). We have seen that the Euclidean connection ∇ is the
Levi–Civita connection of Euclidean space (Example 3.29). Therefore, constant-coefficient
vector fields are parallel (Example 3.35), and the Euclidean geodesics are straight lines with
constant-speed parametrizations (Exercise 4.8). Every Euclidean space En is geodesically
complete.
Example 4.16 (Spheres). Using Proposition 4.14 the geodesics on a sphere can be ob-
tained from its normal space. One can show that a nonconstant curve on Sn (R) is a maximal
geodesics if and only if it is a periodic constant-speed curve whose image is a great circle, that
is, a subset of the form Sn (R) ∩ Π, where Π ⊆ Rn+1 is a 2-dimensional linear subspace:
Firstly, one shows that for any p ∈ Sn (R) the set of vectors orthogonal to p is exactly the
tangent space Tp Sn (R) (by checking for the defining function f (x) = |x|2 = R2 of the sphere
that dfp (v) = 2⟨v, p⟩ = 0).
Secondly, for any nonzero v ∈ Tp Sn (R) one considers the smooth curve γ : R → Rn+1 ,
with a = |v|/R, defined by
v
γ(t) := (cos at)p + (sin at) ,
a
and checks that γ(t) ∈ Sn (R) and that γ ′′ (t) is proportional to γ(t). Hence it is ḡ-orthogonal
to Tγ(t) Sn (R) and thus by Proposition 4.14 a geodesic in Sn (R), namely γv . The image of γv
is the great circle formed by Π = span{p, av } with period 2π a and constant speed (check!).
Conversely, if Π is given by an orthonormal basis {v, w}, then a great circle C is the image
of a geodesic with initial point p = Rw and initial velocity v.
It follows, in particular, that every sphere is geodesically complete.
56 4. GEODESICS

Figure 4.2. The images of maximal geodesics of Sn and Hn are great circles and great
hyperbolas, respectively, and can be obtained by intersecting 2-dimensional linear
subspaces with Sn and Hn (see Examples 4.16 and 4.17).

Example 4.17 (Hyperbolic spaces). The geodesics on Hn (R) can be obtained in the same
way using the hyperboloidal model. The intersection of Hn (R) with a 2-dimensional linear
subspace of R1,n is called a great hyperbola. One can show that the geodesics are of the form
v
γ(t) = (cosh at)p + (sinh at) .
a
For this computation and to see the behavior of geodesics also in the other models of hyperbolic
space, see [36, p. 138–142].

Exercise 4.18. Prove the claim of Example 4.17. More precisely, show that the maximal
geodesics of Hn (R) are the constant-speed embeddings of R whose image is a great hyperbola.
(Hint: Use Proposition 4.14, following essentially the proof used in Example 4.16 but with
defining function f (x) = η̄(x, x) = −R2 given by the Minkowski metric η̄ and grad f = 2p ∈
R1,n .)

4.2. Exponential map

Since we want to understand the behavior of a Riemannian manifold locally around a
point it is not enough to know about the behavior geodesics in certain directions, but it is
important to study the collective behavior of all geodesics emanating from this point in all
possible directions. This information is conveniently encoded in the exponential map2. Note
that everything in this section works verbatim for semi-Riemannian manifolds and, in fact,
even for any other affine connection.
Since γ ′ is parallely propagated along a geodesic γ, and parallel transport P γ is an isometry
by Corollary 3.56, geodesics always have constant velocity |γ ′ (t)| = |γ ′ (t0 )|. Therefore, only
linear reparametrizations are allowed without loosing the geodesic property.

2Note that, a priori, this exponential map (denoted by exp) has nothing to with the exponential map of
a Lie group G (denoted by expG ), however, they do agree for Lie groups with bi-invariant metrics.
 4.2. EXPONENTIAL MAP 57

Lemma 4.19 (Rescaling Lemma). Let M be a smooth manifold with affine connection ∇.
For every p ∈ M , v ∈ Tp M and c, t ∈ R we have that
γcv (t) = γv (ct), (4.7)
whenever both sides are defined.
Proof. If c = 0 both sides define the constant curve p, so assume that c ̸= 0. It suffices
to show that γcv (t) exists and (4.7) holds whenever the right hand side is defined. (The other
way round follows in the same way.)
Suppose γ = γv with maximal domain I ⊆ R. Define a new curve
γ̃ : c−1 I → M,
t 7→ γ̃(t) := γ(ct).
Clearly, γ̃(0) = γ(0) = p and by the chain rule γ̃ ′ (t) = cγ ′ (ct). In particular, γ̃ ′ (0) = cγ ′ (t) =
cv. Hence initial point and velocity of both sides of (4.7) are equal. It remains to show
that γ̃ is a geodesic, then the statement follows by uniqueness and maximality of geodesics
(Corollary 4.7). It follows by the chain and product rule of Theorem 3.46 (and the fact that
(ct)′′ = 0) that
 
′ d ˙k k ˙ i ˙ j
Dt γ̃ (t) =
e γ̃ (t) + Γij (γ̃(t))γ̃ (t)γ̃ (t) ∂k
dt
 
= c2 γ̈ k (ct) + c2 Γkij (γ(ct))γ̇ i (ct)γ̇ j (ct) ∂k = c2 Dt γ ′ (ct) = 0,
hence γ̃ is a geodesic, and so γ̃ = γcv . □
The Rescaling Lemma allows us define a map v 7→ γv in a sensible way.
Definition 4.20. Let M be a smooth manifold with affine connection. The domain of
the exponential map is defined by
E := {v ∈ T M ; γv is defined on [0, 1]},
and the exponential map on M by
exp : E → M,
v 7→ exp(v) := γv (1).
For each p ∈ M the exponential map of M at p is the restriction
expp : Ep := E ∩ Tp M → M.
Clearly, by definition, the domain E is maximal. In general, exp is only locally defined
because it is based on the local result about the existence and uniqueness of solutions to
ordinary differential equations. If M is geodesically complete, however, then E = T M and
exp well-defined globally.
By the Rescaling Lemma it follows immediately that
expp (tv) = γtv (1) = γv (t).
In other words, the straight lines t 7→ tv through the origin in Tp M are mapped by expp to
geodesics through p on M (see Figure 4.3).
Many more nice properties hold.
58 4. GEODESICS

Figure 4.3. The exponential map at a point p maps a straight line t 7→ tv in Tp M to

the geodesic γv in M .

Proposition 4.21 (Properties of the exponential map). Let (M, g) be a Riemannian

manifold and exp : E → M the exponential map. The following hold:
(i) E is an open subset of T M containing the image of the zero section3 T M0 , and the
set Ep ⊆ Tp M is star-shaped4 with respect to 0.
(ii) For each v ∈ T M , the geodesic γv is given by
γv (t) = exp(tv)
for all t such that either side is defined.
(iii) The exponential map is smooth.
(iv) For each point p ∈ M , the differential
d(expp )0 : T0 (Tp M ) ∼
= Tp M → Tp M
is the identity map of Tp M .
Proof. Suppose n = dim M .
(ii) As discussed above, this follows immediately from the Rescaling Lemma.
(i) We first prove that Ep is star-shaped. If v ∈ Ep , then by definition of γv it is defined
on at least [0, 1], and thus by the rescaling lemma
expp (tv) = γtv (1) = γv (t)
is defined for t ∈ [0, 1]. Hence Ep is star-shaped.
To prove that E is open we use the geodesic vector field G ∈ X(T M ) from Theorem 4.12.
By the fundamental theorem of flows [35, Thm. 9.12] there exists an open set D ⊆ R × T M
containing {0} × T M and a unique smooth maximal flow θ : D → T M whose infinitesimal
generator is G, i.e., each curve t 7→ θ(t, (p, v)) is the unique maximal integral curve of G
starting at (p, v), defined on an open interval containing 0.
3The zero section of a vector bundle is the submanifold of the bundle that consists of all the zero vectors,
i.e., T M0 := {0p ; p ∈ M }.
4A subset S of a vector space is called star-shaped around 0 if v ∈ S and t ∈ [0, 1] imply that tv ∈ S. It
follows that every star-shaped set is convex.
 4.2. EXPONENTIAL MAP 59

Suppose (p, v) ∈ E. Thus the geodesic γv is defined at least on [0, 1], and so is the integral
curve of G starting at (p, v) ∈ T M by Theorem 4.12. Hence (1, (p, v)) ∈ D, and therefore
there exists a neighborhood of (1, (p, v)) in R × T M on which the flow of G is defined. In
particular, there is a neighborhood of (p, v) on which the flow exists for t ∈ [0, 1], on which
therefore also the exponential map is defined. Hence E is open.
(iii) By Theorem 4.12 the geodesics are projections of the integral curves of G. Hence the
exponential map can be expressed as
expp (v) = γv (1) = π ◦ θ(1, (p, v)),
wherever it is defined. Since π and θ are smooth, expp (v) depends smoothly on (p, v).
(iv) For any v ∈ T0 (Tp M ) ∼ = Tp M chose a curve τ in Tp M starting at 0 with initial velocity
v = τ ′ (0) = d0 τ . For convenience we use the curve τ (t) = tv. Then
d d d
d(expp )0 (v) = (expp ◦τ )(t) = expp (tv) = γv (t) = v.
dt t=0 dt t=0 dt t=0
Thus d(expp )0 = IdTp M . □
Properties (iii) and (iv) together with the inverse function theorem (see, for instance,
[35, Thm. 4.5]) thus immediately imply the following.
Theorem 4.22. The exponential map expp at every point p ∈ M is a local diffeomorphism,
that is, there exist a neighborhood V of 0 in Tp M and a neighborhood U of p in M such that
expp : V → U is a diffeomorphism. □
Example 4.23 (Exponential map for En ). If v ∈ Tp Rn ∼
= Rn , then the geodesic through
p with initial velocity v is t 7→ p + tv. Hence
expp (v) = p + v,
which is a global diffeomorphism (even an isometry). The same holds for Rν,n−ν .

Figure 4.4. The exponential map at a point p in Rn in direction v ∈ Tp Rn ∼

= Rn . It
n
is a global diffeomorphism, i.e., Ep = Tp R .

In this spirit we will later prove the Gauss Lemma 6.3 which essentially states that expp
is always a “partial isometry”.
Example 4.24 (Exponential map for Sn ). In Example 4.16 we have seen that the geodsics
of Sn are great circles parametrized proportionally to arc length. Given p ∈ Sn and v ∈ Tp Sn
60 4. GEODESICS

the point expp v ∈ Sn is obtained by running along the geodesic γv/|v| a length equal to |v|,
starting from p.
It is clear that expp is defined over the entire tangent space and transforms Bπ (0) injec-
tively into Sn \ {q}, where q is the antipodal point to p. The boundary ∂Bπ (0) is transformed
to q and the open annulus B2π (0) \ Bπ (0) is transformed injectively onto Sn \ {p, q}, and
∂B2π (0) collapes to p, etc.

Figure 4.5. The exponential map at a point p in direction v ∈ Tp Sn of the sphere. It

is a diffeomorphism on Ep = Bπ (0) ⊆ Tp Sn ∼
= Rn but no longer even injective once it
hits the antipodal point q of p.

If, instead, we consider the Riemannian manifold Sn \{q}, expp is defined only on Bπ (0) ⫋
Tp (Sn \ {q}).

Remark 4.25. Theorem 4.22 can easily be extended to the full exponential map. One
can think of points in T M as given by p ∈ M and v ∈ Tp M and thus

exp(p, v) = (p, expp (v)).

Then one can show (see [45, Prop. 5.5.1] or [43] for the full details) that for each p ∈ M and
0 ∈ Tp M the map
d exp : T(p,0) (T M ) → T(p,p) (M × M )
is of the form
 
Id 0
,
∗ Id
hence also nonsingular, and by the inverse function theorem the map exp is therefore a
(local) diffeomorphism of an open neighborhood of the zero section T M0 in T M to an open
neighborhood of the diagonal ∆M in M × M .

Proposition 4.13 on the naturality of geodesics translates to the same property of the
exponential map.

Proposition 4.26 (Naturality of the exponential map). Suppose (M, g) and (M

f, ge) are
Riemannian manifolds and φ : M → M f is a local isometry. Then for every p ∈ M the
 4.2. EXPONENTIAL MAP 61

following diagram commutes:

dφp
Ep Eeφ(p)
expp expφ(p)

φ
M M
f.

Exercise 4.27. Prove Proposition 4.26.

An important consequence of this result is that local isometries (on connected manifolds)
are completely determined by their values and differentials at a single point.
Proposition 4.28. Let (M, g) and (M f, ge) be Riemannian manifolds, and M connected.
Suppose ψ, φ : M → M
f are local isometries such that for some p ∈ M we have φ(p) = ψ(p)
and dφp = dψp . Then φ ≡ ψ.

Problem 4.29. Prove Proposition 4.28.

We can use Proposition 4.28 to finish up an old problem related to the isometry groups
of the modal spaces.
Problem 4.30. Recall the groups E(n), O(n + 1), and O+ (1, n) defined in Section 2.4.1
which act isometrically on the model Riemannian manifolds En , Sn (R), and Hn (R), respec-
tively. Show that, in fact,
Iso(En ) = E(n),
Iso(Sn (R)) = O(n + 1),
Iso(Hn (R)) = O+ (1, n).

4.2.1. Normal neighborhoods and normal coordinates. In Proposition 4.21 we

have shown that Ep is a star-shaped region around 0 in Tp M . Moreover, by Theorem 4.22
expp is a local diffeomorphism. This gives rise to the following definition.
Definition 4.31. Let (M, g) be a Riemannian manifold. A neighborhood U of p ∈ M
is that is the diffeomorphic image under expp of a star-shaped neighborhood of 0 ∈ Tp M is
called a normal neighborhood of p.
On any normal neighborhood there exist distinguished coordinate systems, which are very
helpful for tensor calculations.
Definition 4.32. Let (M, g) be a Riemannian manifold, p ∈ M and U = expp (V ) be a
normal neighborhood of p. Suppose (bi ) is an orthonormal basis for Tp M , which determines
a basis isomorphism B : Rn → Tp M via B(x1 , . . . , xn ) = xi bi . The (Riemannian) normal
coordinates (U, φ = (xi )) centered at p induced by (bi ) are obtained by combining B with
expp to get φ = B −1 ◦ (expp |V )−1 : U → Rn as
B −1
Tp M Rn
φ
(expp |V )−1

U
62 4. GEODESICS

which assigns to each q ∈ U the coordinates of exp−1

p (q) ∈ Tp M with respect to (bi ), i.e.,

(expp |V )−1 (q) = xi (q)bi , q ∈ U.

One can show that the normal coordinate chart associated with a given orthonormal basis
is unique and satisfies
∂i |p = dφ−1
p (∂i |0 ) = d(expp )0 ◦ |{z}
dB0 (∂i |0 ) = B(∂i |0 ) = bi , (4.8)
| {z }
=IdTp M =B

since d(expp )0 is the identity [36, Prop. 5.23]. Moreover, any two normal coordinate charts
(x̃j ) and (xi ) are related by some (constant) orthogonal matrix (Aji ) ∈ O(n), i.e.,
x̃j = Aji xi .
Proposition 4.33 (Properties of normal coordinates). Let (M, g) be a Riemannian man-
ifold, and let (U, (xi )) be any normal coordinate chart centered at p ∈ M .
(i) The coordinates of p are (0, . . . , 0).
(ii) The components of the metric are gij (p) = δij .
(iii) For every v = v i ∂i |p ∈ Tp M , the geodesic γv is represented in normal coordinates
by the line
γv (t) = (tv 1 , . . . , tv n ), (4.9)
as long as t is in some interval I containing 0 with γv (I) ⊆ U .
(iv) The Christoffel symbols vanish at p, i.e., Γkij (p) = 0.
Proof. (i) follows from the definition of normal coordinates.
(ii) follows immediately from (4.8) since
gij (p) = ⟨∂i |p , ∂j |p ⟩ = ⟨bi , bj ⟩ = δij .
(iii) follows from Proposition 4.21(ii).
(iv) Let v = v i ∂i |p ∈ Tp M . The geodesic equation (4.1) for γv (t) = (tv 1 , . . . , tv n ) is
Γkij (γv (t))v i v j = 0.
At t = 0 we see that for all k the quadratic form (Γkij (p))ij is zero. Hence by the polarisation
identity Γkij (p)v i wj = 0 for all v, w, and therefore identitically zero. □

Problem 4.34. Suppose (M, g) is a Riemannian manifold and (U, φ) is a smooth coor-
dinate chart on a neighborhood of p ∈ M such that φ(p) = 0 and φ(U ) is star-shaped with
respect to 0. Prove that this chart is a normal coordinate chart for g if and only if gij (p) = δij
and xi xj Γkij (x) ≡ 0 is satisfied on U .

Problem 4.35. Suppose (M, g) is a Riemannian manifold, and let div and ∆ be the
divergence and Laplace operators definde in Section 2.2.3.
(i) Show that for every vector field X ∈ X(M ), div X can be written in terms of the
total covariant derivative as5
div X = trg (∇X),
5Based on this result, one can then define the divergence of any smooth k-tensor field F by div F :=
trg (∇F ), where the trace is taken on the last two indices of the (k + 1)-tensor field ∇F .
 4.2. EXPONENTIAL MAP 63

and that if X = X i Ei in terms of some local frame, then div X = X i ;i . (Hint: Show
that it suffices to prove the formulas at the origin in normal coordinates.)
(ii) Show that the Laplace operator acting on a smooth function u can be expressed as
∆u = trg (∇2 u),
and in terms of any local frame,
∆u = g ij u;ij = u;i i .

Problem 4.36. Let (M, g) be a Riemannian or pseudo-Riemannian manifold and p ∈ M .

Show that for every orthonormal basis (b1 , . . . , bn ) for Tp M , there is a smooth orthonormal
frame (Ei ) on a neighborhood of p such that Ei |p = bi and (∇Ei )p = 0 for each i.
The geodesics starting at p and lying in a normal neighborhood of p have the very simple
form (4.9), and a special name.
Proposition 4.37. Let U be a normal neighborhood of p ∈ M . Then for each q ∈ U there
exists a unique geodesic σ : [0, 1] → U connecting p and q, called radial geodesic. It satisfies
σ ′ (0) = exp−1
p (q) ∈ V .

Problem 4.38. Prove Proposition 4.37. (Hint: Existence is clear. To prove uniqueness
assume that τ is any such geodesic with initial velcity w and use Theorem 4.4 and the
properties of the exponential map to obtain w = v and thus τ = σ.)
While normal neighborhoods are crucial, we will learn about the even more special convex
neighborhoods C in Section 4.3.2. These are neighborhoods that are normal neighborhoods
for all of their points. By Proposition 4.37 we then know that there exists a unique geodesic
σ in C between any two points p and q in C. But beware, there may still be other geodesics
between p and q that leave C (see Figure 4.6)! This ambiguity will be resolved in Section 4.3
when we bring the Riemannian distance dg back into play to further reduce the size of these
convex neighborhoods.

Figure 4.6. Within the convex neighborhood C there is a unique geodesic between p
and q along the great circle connecting it. Globally, however, there are two geodesics
between p and q (also the long arc).
64 4. GEODESICS

Problem 4.39. Show that (M, g) is connected if and only if any two points in M can
be connected by a broken geodesic, which is a piecewise smooth curve whose smooth parts
are geodesics. (Hint: One direction is clearly sufficient. Conversely, show that the set {p ∈
M ; q can be connected to p by a broken geodesic} is nonempty, open and closed – and hence
equal to M .)

Remark 4.40 (Tubular neighborhoods). One can generalize the construction of normal
neighborhoods to embedded submanifolds P of a Riemannian manifold (M, g). If E is the
domain of the exponential map of M , then one calls the restriction of the exponential map
to the normal bundle N P , i.e., E : EP → M for EP = E ∩ N P , the normal exponential map
of P in M . Normal neighborhoods are diffeomorphic images of open subsets V ⊆ EP whose
intersection with each fiber Nx P is star-shaped with respect to 0. A particularly important
type is that of a tubular neighborhood, when V ⊆ EP has the special form

V := {(x, v) ∈ N P ; |v|g < δ(x)},

for some continuous function δ : P → (0, ∞). If δ(x) ≡ ε it is called an ε-tubular neigh-
borhood. It is not difficult to prove that every embedded submanifold of M has a tubular
neighborhood in M , and that every every compact submanifold has a uniform ε-tubular
neighborhood [36, Thm. 5.25]. In 1939 Hermann Weyl [56] showed (based on earlier work
of Harold Hotelling [31] for curves) that the volume of such a tubular ε-neighborhood of an
embedded compact submanifold P in Euclidean space is an intrinsic quantity, more precisely,
that it is a polynomial with coefficients that only depend on certain integrals of the intrinsic
curvature of P . This result can be generalized to a much larger class of normal neighborhoods
whose sections satisfy certain symmetry conditions based on the dimension and codimension
of P in M [16]. For instance, in the case P is an embedded curve it is sufficient that V in
each fiber has the center of mass in the origin, that is, on the curve. For many more insights
about tubes and their relation to curvature see the book of Gray [25].

4.3. Minimizing curves

There is an important link between geodesics and length-minimizing curves. In this Sec-
tion we will show that every length-minimizing curve is a geodesic and that, almost conversely,
every geodesic is locally length-minizmizing.
Recall that the definition of an admissible curve γ : [a, b] → M as a continuous and
piecewise smooth curve, and its length on a connected Riemannian manifold (M, g) given by
ˆ b
Lg (γ) := |γ ′ (t)|g dt
a

as introduced in Section 2.3.

Definition 4.41. Let (M, g) be a Riemannian manifold. An admissible curve γ in M is

said to be a minimizing curve if Lg (γ) ≤ Lg (e
γ ) for every admissible curve γ
e with the same
endpoints.

From the definition of the Riemannian distance dg it follows that a curve γ between two
points p, q ∈ M is minimizing if and only if dg (p, q) = Lg (γ).
 4.3. MINIMIZING CURVES 65

4.3.1. Minimizing curves are geodesics. Our first goal is to prove that every mini-
mizing curve is a geodesic. The best approach to this problem is via the calculus of variations
although for our purposes we will not (and cannot, in this short time) introduce this theory
in full generality6. The idea is to consider an admissible class of functions (in our case the
admissible curves between given end points) together with a functional that should be mini-
mized or maximized (in this case the length functional). Analogous to multivariable calculus
on then uses differentiation to compute extremals and show convexity/concavity properties.
In this context, the admissible class of functions is infinite-dimensional, and so instead of
derivatives one has to use the first variation to compute potential extrema (this translates
then to a set of ordinary or partial differential equations, called the Euler–Lagrange equation)
and the second variation or other arguments to conclude that the critical point obtained in
the first step is not a saddle point but indeed a minimum or maximum. In our situation the
Euler–Lagrange equation turns out to be the geodesic equation. Many important equations
in geometry and almost all equations in physics arise from the same variational approach
(think of the Einstein equations, the Yamabe equation, or the minimal surface equation).
Since the first variation requires a local computation, it is sufficient to restrict to the
following one-parameter family of curves. Thanks to the exponential map we can restrict to
a smooth family.
Definition 4.42. Let M be a smooth manifold and let I, J ⊆ R be intervals. A (contin-
uous) one-parameter family Γ : J × I → M is called an admissible family of curves if
(i) the domain is of the form J × [a, b] for J an open interval,
(ii) there is a partition (a0 , . . . , ak ) of [a, b] such that Γ is smooth on every J × [ai−1 , ai ]
(called admissible partition), and
(iii) Γs (t) := Γ(s, t) is an admissible curve for every s ∈ J.
Such a family defines two collections of curves in M (see Figure 4.7): the main curves
Γs (t) := Γ(s, t), and the transverse curves Γ(t) (s) := Γ(s, t).

Figure 4.7. An admissible family of curves Γ : J × [a, b] → M with main curves Γs

and transverse curves Γ(t) .

6See my Bachelor course “Optimization in Geometry and Physics” [14] for the classical one-dimensional
theory and Master courses for the full modern theory via direct methods.
66 4. GEODESICS

The velocity vectors of the main and transverse curves (they exist where Γ is sufficiently
smooth) are denoted by
′
∂t Γ(s, t) = (Γs )′ (t) ∈ TΓ(s,t) M, and ∂s Γ(s, t) = Γ(t) (s) ∈ TΓ(s,t) M,
respectively. Note that for an admissible family the transverse curves are smooth on J but
the main curves are generally only piecewise regular. Thus the vector fields ∂s Γ and ∂t Γ are
smooth on each rectangle J × [ai−1 , ai ], but not necessarily everywhere.
Exercise 4.43. Suppose Γ is an admissible family of curves.
(i) Show that ∂s Γ is a piecewise smooth vector field along Γ, that is, that it is a
continuous vector field along Γ whose restrictions to each J × [ai−1 , ai ] are smooth
for an admissible partition.
(ii) Argue that ∂t Γ is generally not continuous at t = ai .
Based on such the notion of such an admissible family of curves we define the variation
of an admissible curve.
Definition 4.44. Let M be a smooth manifold and let γ : [a, b] → M be an admissible
curve. A variation of γ is an admissible family of curves Γ : J × [a, b] → M such that J is an
open interval containing 0 and Γ0 = γ.
The variation Γ is called a proper variation if all the main curves have the same starting
and end point, i.e., Γs (a) = γ(a) and Γs (b) = γ(b) for all s ∈ J.

Figure 4.8. A proper variation Γ of γ with variation field V (t) = ∂s Γ(0, t) of Γ. Every
vector field V along γ is a variation field of some variation of γ (see Exercise 4.45).

If Γ is a variation of γ, then the piecewise smooth vector field V (t) = ∂s Γ(0, t) along γ
(see Exercise 4.43(i)) is called the variation field of Γ. It is called proper if V (a) = V (b) = 0.
Clearly, the variation field of a proper variation is proper. One can also prove the converse
in the following sense.
Exercise 4.45. Suppose γ is an admissible curve and V is a piecewise smooth vector field
along γ. Show that V is the variation field of some variation of γ. Moreover, if V is proper
also the variation of γ can be chosen proper. (Hint: Define Γ(s, t) := expγ(t) (sV (t)).)
In order to relate the minimization property of curves to the geodesic equation, we need
to compute the covariant derivative along the main curves and use the symmetry of the
Levi-Civita connection7.
7The following proof works, in fact, for any symmetric connection.
 4.3. MINIMIZING CURVES 67

Lemma 4.46 (Symmetric Lemma). Let Γ : J ×[a, b] → M be an admissible family of curves

in a Riemannian manifold. On every rectangle J × [ai−1 , ai ] where Γ is smooth we have

Ds ∂t Γ = Dt ∂s Γ.

Proof. Since this is a local property we can fix local coordinates (xi ) around a point
Γ(s0 , t0 ). Writing the components of Γ as Γ(s, t) = (x1 (s, t), . . . , xn (s, t)) we have

∂xk ∂xk
∂t Γ = ∂k , ∂s Γ = ∂k ,
∂t ∂s
and by the coordinate formula (3.17) for Dt we obtain
 2 k
∂xi ∂xj k

∂ x
Ds ∂t Γ = + Γij ∂k
|∂s∂t
{z } ∂t ∂s |{z}
2 k =Γkji
= ∂∂t∂s
x

∂ 2 xk ∂xi ∂xj k
 
= + Γ ∂k = Dt ∂s Γ,
∂t∂s ∂s ∂t ij

where we used the symmetry condition Γkij = Γkji and the fact that Γ is smooth on each
rectangle (and hence the derivatives commute). □

In the next step we compute a differential equation that every length-minimizing curve
has to satisfy that the first variation vanishes. The first variation is essentially the first
derivative of a functional on a function space. In our case it is somewhat simpler because it
is the length functional on a one-parameter family (alternatively, one can and often does use
the energy functional which then also determines the curve with a unit speed; we assume this
conditions holds).

Theorem 4.47 (First variation). Let (M, g) be a Riemannian manifold. Suppose γ : [a, b] →
M is a unit-speed admissible curve, Γ : J × [a, b] → M is a proper variation of γ and
V = ∂s Γ(0, .) is its variation field. Then Lg (Γs ) is a smooth function of s, and
ˆ b k−1
d X
Lg (Γs ) = − ⟨V, Dt γ ′ ⟩dt − ⟨V (ai ), ∆i γ ′ ⟩, (4.10)
ds s=0 a i=1

where (a0 , . . . , ak ) is an admissible partition of V , and for each i = 1, . . . , k − 1, ∆i γ ′ =

γ ′ (a+ ′ − ′
i ) − γ (ai ) is the “jump” in the velocity field γ at ai .

Proof. On each compact [ai−1 , ai ] the integrand of Lg (Γs ) is smooth, and so we can
exchange differentiation with integration and obtain by the chain rule and the Symmetric
Lemma 4.46 that
ˆ ai
d ∂
Lg (Γs |[ai−1 ,ai ] ) = ⟨∂t Γ, ∂t Γ⟩1/2 dt
ds a ∂s
ˆ i−1
ai ˆ ai
1 1
= ⟨∂t Γ, ∂t Γ⟩−1/2 2⟨Ds ∂t Γ, ∂t Γ⟩dt = ⟨Dt ∂s Γ, ∂t Γ⟩dt.
ai−1 2 ai−1 |∂ t Γ|
68 4. GEODESICS

At s = 0 we have ∂s Γ(0, t) = V (t) and ∂t Γ(0, t) = γ ′ (t) (which has unit length), and thus
ˆ ai ˆ ai  
d (iv) d
Lg (Γs |[ai−1 ,ai ] ) = ⟨Dt V, γ ′ ⟩dt = ⟨V, γ ′ ⟩ − ⟨V, Dt γ ′ ⟩ dt
ds s=0 ai−1 ai−1 dt
ˆ ai
′ − ′ +
= ⟨V (ai ), γ (ai )⟩ − ⟨V (ai−1 ), γ (ai−1 )⟩ − ⟨V, Dt γ ′ ⟩dt,
ai−1

where we used property (iv) of the Levi-Civita connection (compatibility with g) as derived
for Dt in Corollary 3.49. Summing over i yields
ˆ b k−1
d ′
X
Lg (Γs ) = − ⟨V, Dt γ ⟩dt − ⟨V (ai ), ∆i γ ′ ⟩ + ⟨V (b), γ ′ (b)⟩ − ⟨V (a), γ ′ (a)⟩. (4.11)
ds s=0 a i=1

Since the variation is proper, we have V (a) = V (b) = 0 and thus (4.10). □

Setting the first variation to zero yields critical points of the length functional via the
Euler–Lagrange equation, which turn out to be the geodesic equation.
Theorem 4.48. In a Riemannian manifold, every minimizing curve which is parametrized
by unit speed is a geodesic.
Proof. Suppose γ : [a, b] → M is a minimizing unit-speed curve with admissible partition
(a0 , . . . , ak ). Thus for every proper variation Γ of γ, Lg (Γs ) is smooth and by elementary
calculus the minimization property of γ in this one-parameter family implies
d
Lg (Γs ) = 0.
ds s=0

Thus for the corresponding proper variation field V , by Theorem 4.47 we obtain that
ˆ b k−1
X
− ⟨V, Dt γ ′ ⟩dt − ⟨V (ai ), ∆i γ ′ ⟩ = 0. (4.12)
a i=1

We show that, in fact, all of these summands vanish. First we show that Dt γ ′ = 0 on each
subinterval [ai−1 , ai ], which means that γ is a “broken geodesic”: Suppose φ ∈ C ∞ (R) is a
bump function with φ > 0 on (ai−1 , ai ) and else zero. For the vector field V = φDt γ ′ we
obtain from (4.12) that
ˆ ai
0=− φ|Dt γ ′ |2 dt,
ai−1

and since φ > 0 on this interval we must have Dt γ ′ = 0 on (ai−1 , ai ).

Second we show that ∆i γ ′ = 0 for each i = 1, . . . , k − 1, which means that γ has no
corners: For each i there is a piecewise smooth vector field V along γ such that V (ai ) = ∆i γ ′
and 0 at all other points aj , j ̸= i. Thus (4.12) implies that

−|∆i γ ′ |2 = 0,
and so ∆i γ ′ = 0 for each i.
Since the two one-sided velocity vectors agree at each ai , the uniqueness of geodesics
implies that γ|[ai−1 ,ai ] can be continued by γ|[ai ,ai+1 ] as geodesic. □
 4.3. MINIMIZING CURVES 69

Figure 4.9. Minimizing the length of a curve γ can be achieved by deforming it in the
direction of the accelartion (left) and by rounding the corners (right).

Remark 4.49. Note that the geometric interpretion of what is being optimized in a
length-minimizing curve that appears for the choice of the vector fields V used in the proof
of Theorem 4.48: When we use V = φDt γ ′ we deform in the direction of the acceleration,
and for V with V (ai ) = ∆i γ ′ we round the corners.
Actually we did not use the minimization feature of a curve but rather that it is a critical
d
point of the length functional, i.e., ds |s=0 Lg (Γs ) = 0.
Corollary 4.50. A unit-speed admissible curve γ is a critical point for Lg if and only if
it is a geodesic.
Proof. If γ is a critical point, then the proof of Theorem 4.48 goes through verbatim.
Conversely, if γ is a geodesic, we know that is has constant-speed parametrization, is smooth
and satisfies Dt γ ′ = 0, hence the right hand side of (4.10) vanishes. □

Problem 4.51. Instead of minimizing the length functional it is often more convenient
to consider the energy functional for an admissible curve γ : [a, b] → M , defined by
ˆ
1 b ′ 2
E(γ) := |γ (t)| dt.
2 a
(i) Prove that an admissible curve is a critical point of E with respect to proper varia-
tions) if and only if it is a geodesic (which means, in particular, that it has constant
speed).
(ii) Prove that if γ is an admissible curve that minimizes the energy among admissible
curves with the same endpoints, then it also minimizes the length.
(iii) Prove that if γ is an admissible curve that minimizes the length among admissible
curves with the same endpoints, then it minimizes the energy if and only if it has
constant speed.
(Remark : Since the integrand of the energy functional is smooth (no square root), its critical
points have automatically constant-speed parametrizations. Hence it is sometimes more useful
to prove the existence of geodesics with certain properties via the energy.)

Remark 4.52. Using the Gauss Lemma (see Section 4.3.2 below) and the more special
convex neighborhoods one can prove that minimizing curves are geodesics also in a different
way (see [36, p. 165 f]) without the use of variations.
Remark 4.53. In Lorentzian geometry it does not make sense to consider length-minimizing
curves because those are simply the null curves. Instead, the class of admissible curves con-
sists only of timelike curves and one seeks to maximize the length functional between two
points that are chronologically related. The result is quite similar to the Riemannian situa-
tion though. In flat Minkowski space, for instance, the curves connecting such points are the
straight lines, and those with constant-speed parametrization are again geodesics.
70 4. GEODESICS

4.3.2. Geodesics are locally minimizing. It is clear that the full converse of Theo-
rem 4.48 does not hold. We can see this by picking two (not antipodal) points on a great
circle of the unit sphere Sn . Then only one of the two geodesic segments is length-minimizing
(see Figure 4.6). We will see, however, that if the two points on a geodesic are not too far
apart (in the case of the sphere less than π, that is, half the length of a great circle) then the
length-minimizing property still holds. Formally, we say these curves are locally minimizing.
How small local really needs to be could be made precise with the notion of the injectivity
radius (for the sphere, inj(Sn ) = π).
Definition 4.54. Let (M, g) be a Riemannian manifold. An admissible curve γ : I → M
is said to be locally minimizing if for every t0 ∈ I there exists a neighborhood I0 ⊆ I containing
t0 such that γ|[a,b] is minimizing for every [a, b] ⊆ I0 .
Lemma 4.55. Every minimizing admissible curve on a Riemannian manifold is locally
minimizing.

Exercise 4.56. Prove Lemma 4.55.

The key ingredient in the proof that geodesics are locally minimizing is the fundamental
Gauss Lemma. It states that any sufficiently small sphere centered at a point in a Riemannian
manifold is perpendicular to every geodesic through the point (see, e.g., [36, Thm. 6.9] and
[45, Lem. 5.5.5]). One can also understand it in the way that the exponential map is a local
radial isometry (see [23, Lem. 3.5] or [43, p. 127]). We start with this more general approach,
because it holds verbatim also in the semi-Riemannian setting (and is the formulation of the
Gauss Lemma used on Wikipedia!), and subsequently prove that the radial vector field is
orthogonal to small geodesic spheres. In order to formulate the first result, we identify for
p ∈ M and x ∈ Tp M the spaces Tx (Tp M ) ∼ = Tp M and say that vx ∈ Tx (Tp M ) is radial if it is
a scalar multiple of x.
Theorem 4.57 (Gauss Lemma). Let (M, g) be a Riemannian manifold, p ∈ M and
x ∈ Ep ⊆ Tp M (so that expp x is defined, see Definition 4.20). Let vx , wx ∈ Tx (Tp M ) ∼
= Tp M
with vx radial. Then
⟨d(expp )x (vx ), d(expp )x (wx )⟩ = ⟨vx , wx ⟩. (4.13)
Proof. Since vx is radial, we may assume without loss of generality that vx = x (since
(4.13) is linear in vx ), and write v instead of x. We also write w in place of wx .
First observe that since v ∈ Ep the expression expp v is defined. Consider the two-
parameter map
u(s, t) := t(v + sw)
in Tp M . Since v ∈ Ep , Ep is open and the curve v(s) := v + sw is continuous (with8 v(0) = v,
v ′ (0) = w, and constant |v(s)|), we also have that v + sw ∈ Ep for s ∈ (−ε, ε). Therefore,
since [0, 1] is compact,
u(s, t) ∈ Ep , (s, t) ∈ (−ε, ε) × [0, 1].
We can thuse define a parametrized surface in M by
Γ : (−ε, ε) × [0, 1] → M,
Γ(s, t) = expp (tv(s)) = expp (t(v + sw)).
8Note that any other curve v with these properties would also work for the proof. See also Figure 4.10.
 4.3. MINIMIZING CURVES 71

Note that ∂t Γ(0, 1) = d(expp )x (v) and ∂s Γ(0, 1) = d(expp )x (w), hence the left hand side of
(4.13) is now expressed in terms of partial derivatives ∂t Γ and ∂s Γ, i.e.,
⟨∂s Γ, ∂t Γ⟩(0, 1) = ⟨d(expp )x (w), d(expp )x (v)⟩. (4.14)
For fixed s, let us consider the main curves Γs = Γ(s, .) : t 7→ Γ(s, t). They are geodesics with
initial velocity ∂t Γ(s, 0) = v + sw and constant speed |∂t Γ(s, .)|. In particular, Dt ∂t Γ = 0.
(see Figure 4.10).

Figure 4.10. The map Γ is a variation of the curve t 7→ expp (tv) which is proper
on one side since Γs (0) = Γ(s, 0) = p. The main curves Γs (t) = expp (tv(s)) are all
geodesics, thus Dt ∂t Γ = 0. (Note that in this picture the curve s 7→ v(s) in Tp M is
not a straight line as chosen for convenience in the proof of Theorem 6.3. This is to
emphasize that the particular shape of the curve is largely irrelevant. Verify that all
we need for the proof is that v(0) = v, v ′ (0) = w, and that the curve has constant
speed.)

By the Symmetric Lemma 4.46 we furthermore have that Ds ∂t Γ = Dt ∂s Γ. Thus from the
metric compatibility (iv) of Dt from Corollary 3.49 it follows that for all (s, t)
(iv)
∂t ⟨∂s Γ, ∂t Γ⟩ = ⟨Dt ∂s Γ , ∂t Γ⟩ + ⟨∂s Γ, Dt ∂t Γ⟩
| {z } | {z }
=Ds ∂t Γ =0
(iv) 1
= ⟨Ds ∂t Γ, ∂t Γ⟩ = ∂s ⟨∂t Γ, ∂t Γ⟩ = ⟨v, w⟩ + s⟨w, w⟩.
2 | {z }
=|v+sw|2

For s = 0 we already know that ∂t Γ(0, 0) = v, and furthermore obtain

∂s Γ(0, 0) = lim ∂s Γ(0, t) = lim d(expp )tv (tw) = IdTp M (0) = 0.
t→0 t→0
Thus by the above and the fundamental theorem of calculus
(4.14)
⟨d(expp )x (w), d(expp )x (v)⟩ = ⟨∂s Γ, ∂t Γ⟩(0, 1)
ˆ 1
= ⟨∂s Γ, ∂t Γ⟩(0, 0) + ⟨v, w⟩dt = ⟨v, w⟩,
| {z } 0
=⟨0,v⟩=0

which is the desired equality. □

72 4. GEODESICS

We can reinterpret the Gauss Lemma in terms of geodesic balls and the radial vector field,
which we now make precise.
Definition 4.58. Let (M, g) be a Riemannian manifold and p ∈ M . Suppose ε > 0 is
such that expp is a diffeomorphism from the ball Bε (0) ⊆ Tp M (with respect to gp ), then the
image expp (Bε (0)) is a normal neighborhood of p and called a geodesic ball in M .
If Bε (0) is contained in an open neighborhood V ⊆ Tp M where expp is a diffeomorphism
onto its image, then expp (Bε (0)) is called a closed geodesic ball, and expp (∂Bε (0)) a geodesic
sphere.
In Riemannian normal coordinates centered at p the open and closed geodesic balls, and
the geodesic sphere are just the coordinate balls and spheres.
Definition 4.59. Let (M, g) be a Riemanian manifold and U be a normal neighborhood
of p ∈ M . The radial distance function r : U → [0, ∞) is defined by
r(q) := | exp−1
p (q)|, q ∈ U,
and ∂r on U \ {p} is the radial vector field .
Clearly, both r and ∂r are smooth on U \ {p} and r2 is smooth on U , because expp is a
√
diffeomorphism on U and g is smooth (but . is only smooth away from 0).
In other words, the radial function is simply the Euclidean distance function from the
origin in Tp M in exponential coordinates. To be more precise, one can show9 that in any
normal coordinates (xi ) on U centered at p we have
v
u n n
uX X xi ∂
r(x) = t (xi )2 , and ∂r (x) = . (4.15)
r(x) ∂xi
i=1 i=1

Exercise 4.60. Prove (4.15).

Example 4.61. In En , r(x) is the distance of x to the origin, and ∂r is the unit vector
field point radially outward from the origin.

Exercise 4.62. Suppose σ is the radial geodesic from p to some q in the (maximal)
normal neighborhood of p. Then Lg (σ) = r(q).
Another way to formulate the Gauss Lemma is the following.
Theorem 4.63 (Gauss Lemma). Let (M, g) be a Riemannian manifold and let U be a
geodesic ball centered at p ∈ M . Then the radial vector field ∂r is a unit vector field orthogonal
to the geodesic spheres in U \ {p}.

Problem 4.64. Use the first version of the Gauss Lemma (Theorem 6.3) to prove the
second version (Theorem 4.63). (Hint: In normal coordinates the geodesic spheres Σδ :=
expp (∂Bδ (0)) are given by ni=1 (xi )2 = δ 2 . For any q ∈ U \ {p} with r(q) = δ show that
P
⟨∂r |q , w⟩g = 0, for all w ∈ Tq M that are tangent to Σδ at q, using the Gauss Lemma.)
9This is actually how Lee [36, p. 158] define the radial distance function r and corresponding radial vector
field ∂r . This definition is more intuitive, but the downside is that one then still has to show that r and ∂r are
well-defined, i.e., that they are independent of the choice of normal coordinates.
 4.3. MINIMIZING CURVES 73

Petersen [45, Lemma 5.5.5] actually calls the following formulation the Gauss Lemma.
Corollary 4.65. Let (M, g) be a Riemannian manifold and let U be a geodesic ball
centered at p ∈ M . Then grad r = ∂r on U \ {p}.
Proof. By Problem 2.46 this is equivalent to ∂r being orthogonal to the level sets of r
(follows from the Gauss Lemma) and ∂r (r) ≡ |∂r |2g (a direct computation in normal coordi-
nates yields ∂r (r) ≡ 1 and this is equal to |∂r |2g by the Gauss Lemma). □
Remark 4.66. In semi-Riemannian Geometry it is also possible to equip Tx (Tp M ) with a
scalar product and show that the exponential map is a radial isometry in this sense (see, e.g.,
[6, Thm. 10.18] and [43, p. 126 ff]). Instead of geodesics balls and radial distance functions,
however, one has to use the quadratic form q̃(x) := ⟨x, x⟩ on Tp M and q := q̃ ◦ exp−1 p on
normal neighborhoods. Instead of geodesic spheres, the level sets of q are local hyperquadrics
on semi-Riemannian manifolds.
We set out to prove that geodesics are locally length-minimizing. In a first step, we will
prove this result for radial geodesics from Proposition 4.37.
Proposition 4.67. Let (M, g) be a Riemannian manifold. Suppose p ∈ M and q con-
tained in a geodesic ball around p. Then (up to reparametrization) the radial geodesic σ from
p to q is the unique minimizing curve in M from p to q.
Moreover, the radial distance function is equal to the Riemannian distance function within
every geodesic ball, that is,
r(q) = Lg (σ) = dg (p, q).
Proof. Choose ε > 0 such that U = expp (Bε (0)) is a geodesic ball containing q. Let
σ : [0, c] → M be the radial geodesic from p to q (see Proposition 4.37), and assume that it is
parametrized by unit speed, i.e., σ(t) = expp (tv) for v ∈ Tp M , |v| = 1. Then the velocity is
equal to ∂r and Lg (σ) = c = r(q) (see also Exercise 4.62).
To show that σ is minimizing, we need to consider an arbitrary admissible curve γ : [0, b] →
M from p to q (without loss of generality we assume it is parametrized by arc length, and
that γ(t′ ) ̸= p for t′ ∈ (0, b]) and show that Lg (γ) ≥ c with equality if and only if γ ≡ σ.
Step 1. Assume that γ([0, b]) ⊆ U . Then the composition r ◦ γ with the radial function r
is continuous on [0, b] and piecewise smooth on (0, b). By the fundamental theorem of calculus
ˆ b ˆ b ˆ b
d ′
c = r(γ(b)) − r(γ(0)) = r(γ(t))dt = dr(γ (t))dt = ⟨grad r|γ(t) , γ ′ (t)⟩dt,
| {z } 0 dt 0 0
=0

and by the Cauchy–Schwarz inequality,

ˆ b
Lg (σ) = c ≤ | grad r|γ(t) | |γ ′ (t)|dt = Lg (γ),
0 | {z }
=|∂r |=1

so σ is minimizing.
Next, assume that Lg (γ) = c. Then b = c, and the Cauchy–Schwarz inequality must be an
equality almost everywhere, which is the case if and only if γ ′ (t) and grad r|γ(t) are collinear
for all t. Since we assumed that γ has unit speed, in fact,
γ ′ (t) = grad r|γ(t) = ∂r |γ(t) .
74 4. GEODESICS

Therefore, both γ and σ are integral curves of ∂r , passing through q at time t = c, and so by
uniqueness γ = σ.
Step 2. Assume that γ leaves U . If γ([0, b]) is not contained in U = expp (Bε (0)), then let
b0 ∈ (0, b] be the first parameter point for which γ(b0 ) belongs to ∂U . Then
Lg (γ) ≥ Lg (γ|[0,b0 ] ) ≥ ε > c = Lg (σ),
and equality cannot be achieved. □
Remark 4.68 (Normal neighborhoods are not good enough). Note that geodesic balls B
at a point p are in many ways better than arbitrary normal neighborhoods U of p. Although
there exists a unique radial geodesic between p and q ∈ U , this geodesic need not be the
shortest curve in M . If, however, q ∈ B, then the radial geodesic between p and q is the
unique shortest curve in M .
The following example illustrates the problem clearly: Suppose M = S1 × R is a cylinder
in R3 and L a vertical line. Then U = M \ L is a normal neighborhood of any point p ∈ U but
in general radial geodesics from p to some point q ∈ U need not be minimizing in M. On the
other hand, within geodesic balls B centered at p they are (see Figure 4.11 and Exercise 4.9
for a description of the geodesics in M ).

Figure 4.11. (a) The set U = M \ L is a normal neighborhood of p and the radial
geodesic σ connecting p and q ∈ M \ L is the unique shortest curve in U , however,
there is a shorter curve τ connecting p and q in M . (b) On the other hand, the largest
possible geodesic ball B centered at p has radius π. If q is in this geodesic ball, then
the radial geodesic is indeed the unique shortest cuve in M connecting p and q. If w
is a point in M \ B then there are still shortest curves from p to w but uniqueness is
no longer guaranteed (it fails when w is a point on the vertical line through −p).

It is important to note that Proposition 4.67 relates the local radial distance function r
defined in (4.15) to the global Riemannian distance function dg from Definition 2.74. We have
already seen, however, that geodesics can cease to be minimizing after a while. In this sense,
Proposition 4.67 is not a global result. It may be even worse, namely that there doesn’t even
 4.3. MINIMIZING CURVES 75

exist a shortest curve between two (even arbitrarily close) points on a Riemannian manifold:
think about the punctured plane R2 \ {(0, 0)} with the induced Euclidean metric and consider
the points p = (ε, 0) and q = (−ε, 0).
Since the radial distance function and Riemannian distance function are equal on geodesic
balls by Proposition 4.67, one can show the following corollary relating the geodesic and metric
balls globally.
Corollary 4.69. In a connected Riemannian manifold the geodesic balls (and spheres)
are also metric balls (and spheres) of the same radius.
Sketch of proof. For V = expp (Bε (0)), it follows from Proposition 4.67 that if q ∈ V
then dg (p, q) < ε. If q ̸∈ V , then one can argue by contradiction (similar to the proof of 4.67)
that dg (p, q) > ε. For the detailled proof see [36, Cor. 6.13]. □
We can now also prove the remaining property of the Riemannian distance function,
namely that dg indeed is a metric on M (see Theorem 2.80), and show some other important
relations between g and dg which are crucial from the metric geometric point of view (see also
Section 4.4).
Problem 4.70. Let (M, g) be a connected Riemannian manifold.
(i) Show that dg (p, q) = 0 if and only if p = q.
(ii) Argue that dg induces the manifold topology.

Problem 4.71. Let (M, g) be a connected Riemannian manifold. Show that the distance
function on a smooth manifold determines the Riemannian metric:
(i) Show that if γ : (−ε, ε) → M is any smooth curve, then
dg (γ(0), γ(t))
|γ ′ (0)|g = lim .
t↘0 t
(ii) Show that if g and ḡ are two Riemannian metrics on M such that dg (p, q) = dḡ (p, q)
for all p, q ∈ M , then g = ḡ.
Finally, we sketch how to extend Proposition 4.67 to show that every geodesic is locally
minimizing. For this we use that we can find not just normal but actually convex neighbor-
hoods around each point.
Definition 4.72. An open set in a Riemannian manifold is called (geodesically) convex
provided it is a normal neighborhood of each of its points.
Remark 4.73. While geodesic balls are clearly related to Euclidean balls in the tangent
space and normal neighborhoods are related to star-shaped domains, the same is not true for
convex neighborhoods: In the cylinder example used in Remark 4.68 the set U = M \ L is
clearly a convex set, but not a convex neighborhood.
Lemma 4.74. Every point in a Riemannian manifold possesses a base of convex neighbor-
hoods.
We will not prove this Lemma here but see, for instance, [43, p. 130] and [36, Problem
6-5]. The proof is based on the fact that the full exponential map on M is a local diffeomor-
phism (which was discussed in Remark 4.25), so that we can remove the dependence on the
center point p of normal neighborhoods. The convex neighborhoods constructed are, in fact,
sufficiently small geodesic balls.
76 4. GEODESICS

Theorem 4.75. Every geodesic in a Riemannian manifold is locally minimizing.

Proof. Suppose γ : I → M is a geodesic with I an open interval. Let t0 ∈ I and let C be
a convex neighborhood of γ(t0 ). Consider the connected component I0 of γ −1 (C) containing
t0 . If a, b ∈ I0 , a < b, then γ(b) is in C and hence, in particular, in a geodesic ball of γ(a).
Thus by Proposition 4.67, the radial geodesic σ from γ(a) to γ(b) is the unique minimizing
curve between these points. On the other hand, by construction, the restriction γ|[a,b] is also a
geodesic that is contained entirely in C. Thus by the uniqueness of radial geodesics obtained
in Proposition 4.37, the geodesics γ|[a,b] and σ must coincide. □
Remark 4.76. The proof shows that one does not need the full generality of a convex
neighborhood. It is sufficient to use uniformly normal neighborhoods, that is, neighborhoods
W such that for some δ > 0 the set W is contained in a geodesic ball of radius δ around each
of its points (see [36, p. 163], and [23, p. 72] who calls them totally normal neighborhoods).

4.4. Completeness (not covered in the course)

In the previous Section we have seen how geodesic and metric balls are related. In what
follows we sketch that also geodesic and metric completeness are intimately tied to each other.
This is the Hopf–Rinow Theorem.
Definition 4.77. A Riemannian manifold is said to be geodesically complete (at p) if
every maximal geodesic (through p) is defined on all of R.
Definition 4.78. A Riemannian manifold (M, g) is said to be metrically complete if it
is complete as metric space with respect to the Riemannian distance function dg in the sense
that every Cauchy sequence converges.
The following result [30] of Heinz Hopf and Willi Rinow from 1931 is a crucial feature of
Riemannian manifolds (it is false for Lorentzian manifolds, see Clifton–Pohl torus in [43, p.
193]). One can prove it with the methods we have developed so far, but since we do not
want to dwelve further in the metric direction, we only state the result. For a proof see, for
instance, [45, Theorem 5.7.1], [36, p. 166 ff], or [43, p. 138 ff].
Theorem 4.79 (Hopf–Rinow). Let (M, g) be a (connected) Riemannian manifold. The
following statements are equivalent:
(i) M is geodesically complete.
(ii) M is geodesically complete at p.
(iii) M satisfies the Heine–Borel property, i.e., every closed bounded set is compact.
(iv) M is metrically complete.
An essential step of the proof is the following observation.
Lemma 4.80. Suppose (M, g) is a connected Riemannian manifold and that expp is defined
on all of Tp M . Then for every q ∈ M there exists a minimizing geodesic from p to q.
Sketch of proof of Lemma 4.80. Choose ε > 0 such that the closed geodesic ball
Bε (p) around p does not contain q. By continuity of dg and compactness of the geodesic
sphere Sε (p) = ∂Bε (p), the function y 7→ dg (y, q) attains a minimum at x ∈ Sε (p). Suppose
γ is the maximal unit speed geodesic whose initial segment γ|[0,ε] is the radial geodesic from
p to x. By assumption, γ is defined on R.
 4.4. COMPLETENESS 77

One then shows that

dg (p, q) = dg (p, x) + dg (x, q).
Then let T := dg (p, q) and consider the set
A := {b ∈ [0, T ]; γ|[0,b] is minimizing and satisfies b + dg (γ(b), q) = T }.
Since ε ∈ A, A := sup A ≥ ε > 0. By continuity of the distance function, A is closed and
hence A ∈ A. If T = A we are done. One assumes A < T , then extends it a bit a minimizing
curve (without additional corner) that still satisfies the condition of A and so arrives at a
contradiction. □
Sketch of proof of Theorem 4.79. (i)⇒(ii) is trival.
For (ii)⇒(iv) suppose M is geodesically complete at p. Then expp is defined on all of
Tp M . Let (qi ) be a Cauchy sequence in M . For each i, let γi (t) = expp (tvi ) be a unit-speed
minimizing geodesic from p to qi . Setting di = dg (p, qi ) yields qi = expp (di vi ). Since Cauchy
sequences are bounded, so is (di ) and since |vi | = 1 also the sequence (di vi ) of vectors in
Tp M is bounded. Thus a subsequence (dik vik ) converges to some v ∈ Tp M , and by continuity
of the exponential map limk→∞ expp (dik vik ) = expp v. Since the original sequence (qi ) was
already Cauchy, it converges as a whole to the limit q = expp v in M .
To show (iv)⇒(i) assume that M is metrically complete but not geodesically complete.
Then there exists a unit-speed geodesic γ : [0, b) → M with no extension to b. Then for a
sequence (ti ) approaching b one can show that qi = γ(ti ) satisfies
dg (qi , qj ) ≤ |ti − tj |,
hence (qi ) is a Cauchy sequence and hence converges to some point q in M . By considering
a convex neighborhood C(q) around q one can construct geodesic extensions γ e of γ for all
e′ (tj ) = γ ′ (tj ) but γ
qj ∈ C(q) with j sufficiently large such that γ e is defined for tj + δ > b, a
contradiction.
Hence (i)⇔(ii)⇔(iv). In addition, one can prove (ii)⇒(iii) using Lemma 4.80 by picking a
point q in a closed and bounded set A and considering the minimizing geodesic σ : [0, 1] → M
from p to q. From Theorem 4.67 we know that |σ ′ (0)| = L(σ) = dg (p, q). The triangle
inequality together with the boundedness of A implies that each such σ ′ (0) for all q ∈ A is in
a compact ball Br (0) ⊆ Tp M . Since A ⊆ expp (Br (0)) and expp (Br (0)) is compact, also A is.
Finally, (iii)⇒(iv) follows because every Cauchy sequence is bounded, and hence the
closure is compact. Thus the sequence contains a convergent subsequence, and since it is
actually Cauchy, it converges itself.¸ □
In addition, some more useful and important implications follow.
Corollary 4.81. On a connected complete Riemannian manifold any two points can be
joined by a minimizing geodesic. □
In some situations the Heine–Borel property holds (almost) trivially. For instance, we get
the following results.
Corollary 4.82. Every compact Riemannian manifold is complete. □
Corollary 4.83. If a Riemannian manifold (M, g) admits a proper Lipschitz function
f : M → R, then M is complete.

Exercise 4.84. Prove Corollary 4.83. How can this result be used for warped products?
78 4. GEODESICS

Problem 4.85. A curve γ : [0, b) → M (with 0 < b ≤ ∞) is said to diverge to infinity if

for every compact set K ⊆ M , there is a time T ∈ [0, b) such that γ(t) ̸∈ K for all t > T .
Prove that a connected Riemannian manifold is complete if any only if every regular curve
that diverges to infinity has infinite length. (The length of a curve whose domain is not
compact is just the supremum of the length of its restrictions to compact submanifolds.)

Remark 4.86. The Hopf–Rinow Theorem does not make sense on semi-Riemannian man-
ifolds (the Clifton–Pohl torus [43, p. 193] is a counterexample to Corollary 4.82), and it also
false in infinite dimensions [5]. It does, however, generalize to locally compact length metric
spaces. This result is due to Stefan Cohn–Vossen, see [11, Theorem 2.5.28], and an important
cornerstone of Metric Geometry. In Lorentzian geometry one can still characterize globally
hyperbolic spacetimes (which in many ways resemble complete Riemannian manifolds) by
metric completeness of a null distance [15, Theorem 1.4].
 CHAPTER 5

Curvature

The notion of curvature is due to Bernhard Riemann and was presented in his habilitation
lecture “Über die Hypothesen, welche der Geometrie zu Grunde liegen” (German for “On the
hypotheses which lie at the bases of geometry”) in 1854, but was published1 only in 1868 after
his death. For some historical background and the original spirit from a modern point of view
see, for instance, the recent book by Jürgen Jost [47] (or the much older re-edition [48] in
German and with comments by Hermann Weyl). It is important to know that Riemann’s
mentor Gauss had at the time already extensively studied surfaces in the Euclidean 3-space. In
his Theorema Egregium 2 (Latin for “Remarkable Theorem”) Gauss showed that the curvature
of a surface can be expressed solely in terms of the first fundamental form and that the way
the surface is embedded does not matter. In order words, Gaussian curvature is an intrinsic
local invariant of a surface.

In Gauss’s spirit, Riemann introduced curvature at a point p on an abstract Riemannian

manifold3 M by considering 2-dimensional subspaces Π ⊆ Tp M and using the Gaussian
curvature K(p, Π) of the corresponding 2-dimensional submanifold S generated by geodesics
starting at p and tangent to Π (essentially the image of Π under expp ). Nowadays we call
K(p, Π) the sectional curvature of M at p with respect to Π. To be more precise, Riemann
computes an expansion of the metric in normal coordinates to second order and recovers so
the Gaussian curvature. This approach thus allows for the desired geometric interpretation.

Moreover, Riemann considers constant curvature spaces. Clearly, if M = En we have

K ≡ 0. In addition, Riemann computed the sectional curvature of a sphere of radius R and
showed that it is everywhere K ≡ R12 . But he did not stop there, he also considered the case
when K is negative. At the time, many mathematicians tried to find geometries that satisfied
the first four of Euclid’s axioms but not the fifth “parallel postulate”4. This is how hyperbolic
space was discovered independently by Nikolai Lobachevsky, János Bolyai and Carl Friedrich
Gauss. Riemann essentially introduced it as constant curvature space with K ≡ − R12 (see, for
instance, [7, Section 4.3.2] for some more details). In retrospect, the reason why hyperbolic
space was not found earlier is due to Hilbert’s Theorem: the hyperbolic plane H2 is not
isometrically embeddable in E3 (we already know from Example 2.32 that hyperbolic space

1See https://www.emis.de/classics/Riemann for Riemann’s original (transcripted) texts.

2This theorem is covered in the 2nd Year Bachelor course “Curves and Surfaces”.
3Note that even the notion of an abstract manifold wasn’t yet available at this point and had to be
introduced by Riemann, which shows just how pioneering and influential his work was (see [37, p. 295 ff] for
a wider discussion).
4Euclid’s fifth postulate states: If a straight line falling on two straight lines makes two interior angles on
the same side with sum less than two right angles, the two straight lines, if produced indefinitely, meet on that
side on which the angles are less than two right angles.
79
80 5. CURVATURE

is a codimension-one submanifold of Minkowski space though, and from Nash’s Embedding

Theorem we know we can still isometrically embed it in some En for n ≥ 4).
Note that Riemann did not make use of parallel transport and connections (which we
have already heard in Chapter 3 were introduced by the Italian school only several decades
later), in fact, he didn’t even use tensors. We are going to go a very different, modern, and
rather abstract route but try to relay some of Riemann’s original observations as well.
The notion of curvature is indispensible in mathematics and physics today. Note that
everything we learn in this chapter works verbatim for semi-Riemannian manifolds (see [43],
modulo different sign conventions that different authors use). As an important application,
we will discuss briefly how curvature is used to formulate the theory of General Relativity.
Exercise 5.1. Look at Riemann’s original paper and try to find evidence that he uses
normal coordinates and discusses constant curvature spaces. (Hint: Mannigfaltigkeit = man-
ifold, Krümmung = curvature)

5.1. Riemann curvature tensor

So far we have studied Riemannian manifolds individually, occassionally also by looking
at (local) isometries between two Riemannian manifolds. Our goal now is to fully capture via
local invariants when two Riemannian manifolds are locally equivalent (or not). We start by
analyzing, once more, the specific properties of Euclidean space and describe its flatness in an
abstract way. As an extension of this observation we then introduce the Riemann curvature
tensor as a failure of the covariant derivatives to commute.
5.1.1. Motivation. Suppose M is a 2-dimensional Riemannian manifold and z ∈ Tp M .
If M = E2 we can extend z trivially to a smooth parallel (constant) vector field on all of
M . For a general manifold M (such as the 2-sphere) assume we are given local coordinates
(x1 , x2 ) at p. Then we can extend z at least locally to a vector field Z, first by parallel
transport along the x1 -axis and then by parallel transport along x2 -axis (see Figure 5.1).
By construction, ∇∂2 Z ≡ 0, but it is unclear whether Z is still parallel with respect to
any other x1 -line, i.e., whether ∇∂1 Z ≡ 0 (by contruction it is true at x2 = 0). Because
parallel transport is unique and because the parallel transport of zero is zero, we can ask,
equivalently, whether
∇∂2 ∇∂1 Z = 0. (5.1)
Suppose we know that this expression commutes, i.e., that
∇∂2 ∇∂1 Z = ∇∂1 ∇∂2 Z, (5.2)
then we obtain immediately from ∇∂2 Z ≡ 0 that also ∇∂1 Z ≡ 0. But can we assume (5.2)?
The formula (5.2) holds for En with the Euclidean connection because
∇∂i ∇∂j Z = (∂i ∂j Z k )∂k = (∂j ∂i Z k )∂k = ∇∂j ∇∂i Z,
but may not hold for general Riemannian manifolds. Moreover, in order to check whether
the commutativity condition (5.2) is well-defined (and satisfied), it should be formulated in
an coordinate-independent manner. Since on En we know that ∇Y ∇X Z = Y X(Z k )∂k we can
use the Lie bracket [X, Y ] = XY − Y X to write
∇X ∇Y Z − ∇Y ∇X Z = ∇[X,Y ] Z.
 5.1. RIEMANN CURVATURE TENSOR 81

Figure 5.1. The vector z ∈ Tp M is parallel transported along x1 and x2 . The different
tangent vectors on the north pole hint that commutativity is not satisfied on S2 .

The above observations immediately translate into a necessary condition for Riemannian
manifolds that are flat, that is, locally isometric to Euclidean space.
Proposition 5.2. If (M, g) is a flat Riemannian manifold, then its Levi-Civita connection
satisfies the flatness criterion, that is, for any smooth vector fields X, Y, Z defined on an open
subset of M we have
∇X ∇Y Z − ∇Y ∇X Z = ∇[X,Y ] Z. (5.3)

Exercise 5.3. Fill in the details in the above exposition and prove Proposition 5.2.

Problem 5.4. Show that the round 2-sphere is not locally isometric to Euclidean space
via Proposition 5.2. (Hint: Let X(φ, θ) = (sin φ cos θ, sin φ sin θ, cos φ) be the spherical coor-
dinate parametrization of an open set U of S2 , and let Xθ = X∗ (∂θ ) and Xφ = X∗ (∂φ ) denote
the corresponding coordinate vector fields. Compute ∇Xθ (Xφ ) and ∇Xφ (Xφ ) and note that
Xφ is parallel along the equator φ = π2 and along each meridian θ = θ0 .)

5.1.2. Definition. In general, as Figure 5.1 shows for the sphere, parallel transport along
closed curves in an arbitrary Riemannian manifold does not commute. Curvature measures
precisely this noncommutativity.
Definition 5.5. Let (M, g) be a Riemannian manifold. The map R : X(M )3 → X(M ),
defined by5
R(X, Y )Z := ∇X ∇Y Z − ∇Y ∇X Z − ∇[X,Y ] Z, (5.4)
5Warning! Different authors use different conventions, e.g., doCarmo [23] and O’Neill [43] define the
Riemann curvature tensor with the opposite sign, but their sectional curvatures have the same sign. We use
Lee’s book [36] and thus his convention. Note that also the index notation discussed later varies considerably,
and it is important to be aware of the convention in use when reading a book or article.
82 5. CURVATURE

is called the Riemann curvature tensor .

The following calculation verifies that R is indeed a (1, 3)-tensor field via the tensor
characterization lemma, hence justifying the name a posteriori.
Proposition 5.6. The map R defined in (5.4) is multilinear over C ∞ (M ), and thus a
(1, 3)-tensor field.
Proof. For f ∈ C ∞ (M ) it follows from the properties (i) and (iii) of affine connections
and the Lie bracket that
R(X, f Y )Z = ∇X ∇f Y Z − ∇f Y ∇X Z − ∇[X,f Y ] Z
= ∇X (f ∇Y Z) − f ∇Y ∇X Z − ∇f [X,Y ]+(Xf )Y Z
= (Xf )∇Y Z + f ∇X ∇Y Z − f ∇Y ∇X Z − f ∇[X,Y ] Z − (Xf )∇Y Z
= f R(X, Y )Z.
Since R(X, Y )Z = −R(Y, X)Z the same is true in X. The linearity in Z is an exercise. □

Problem 5.7. Complete the proof of Proposition 5.6 by showing that R(X, Y )(f Z) =
f R(X, Y )Z for all X, Y, Z ∈ X(M ), f ∈ C ∞ (M ).
Due to R ∈ T (1,3) (M ) we can write it in local coordinates (xi ) as
R = Rijk l dxi ⊗ dxj ⊗ dxk ⊗ ∂l ,
where the coefficients Rijk l are given implicitely by
R(∂i , ∂j )∂k = Rijk l ∂l .
One can easily compute the coefficients via the Christoffel symbols which in turn, namely by
(3.15), follows directly from the metric coefficients gij .
Proposition 5.8. Let (M, g) be a Riemannian manifold. In local coordinates, the com-
ponents of the Riemann curvature tensor are given by
Rijk l = ∂i Γljk − ∂j Γlik + Γm l m l
jk Γim − Γik Γjm . (5.5)

Problem 5.9. Prove Proposition 5.8.

The endomorphism Z 7→ R(X, Y )Z is also called the (directional) curvature operator.
It is, however, often more covenient to describe the Riemann curvature tensor in its
equivalent covariant form, namely as a (0, 4)-tensor field Rm = R♭ (recall that ♭ denotes a
musical isomorphism introduced in Section 2.2.1 to lower an index). In terms of vector fields
we thus have
Rm(X, Y, Z, W ) = ⟨R(X, Y )Z, W ⟩, (5.6)
and in local coordinates
Rijkl = glm Rijk m .

Problem 5.10. Let G be a Lie group with a bi-invariant metric g. Let X, Y, Z ∈ X(G)
be left invariant vector fields on G. Show that
1
R(X, Y )Z = [Z, [X, Y ]].
4
 5.1. RIEMANN CURVATURE TENSOR 83

(Hint: First show that ∇X Y = 12 [X, Y ] by using the symmetry of the connection and the fact
that ∇X X = 0. This is [36, Prob. 5-8(b)]. See also Ex 3 (on p. 80 f) and Exercise 1 (on p.
103) in [23].)

Remark 5.11 (Curvature and curves). So far we have not said how curvature is related
to (variations of) curves. This connection is made precise in [36, Prop. 7.5], where it is shown
that for a smooth one-parameter family Γ of curves in (M, g) and a smooth vector field V
along Γ one obtains
Ds Dt V − Dt Ds V = R(∂s Γ, ∂t Γ)V. (5.7)
Before introducing other curvature notions, we first establish some basic properties of the
Riemann curvature tensor.

5.1.3. Flatness. By defining the curvature tensor we have indeed achieved what we
hoped for, namely to obtain a local invariant of a Riemannian manifold.
Proposition 5.12. The curvature tensor is a local isometry invariant: if (M, g) and
f f is a local isometry, then φ∗ Rm
(M , ge) are Riemannian manifolds and φ : M → M g = Rm.

Exercise 5.13. Prove Proposition 5.12.

In what follows we show the converse in the flat case, which yields a more qualitative
description of curvature as a geometric obstruction of a Riemannian manifold to being locally
isometric to Euclidean space. The proof rests one the construction of a local parallel vector
field, which should not come as a surprise after the motivating discussion in Section 5.1.1
and our earlier introduced notion of a holonomy group (see Definition 3.62), both essentially
related to the noncommutativity of parallel transport along closed curves.
Lemma 5.14. Suppose M is a smooth manifold with affine connection ∇ satisfying the
flatness condition (5.3). Then for any p ∈ M and v ∈ Tp M there exists a parallel vector field
V in a neighborhood of p such that Vp = v.
Proof. Let p ∈ M , v ∈ Tp M and (xi ) smooth coordinates on M centered at p. Without
loss of generality we assume that the image of the coordinate domain is an open cube Cε with
side length 2ε (same notation used for the domain).
We parallel transport v along the x1 -axis, then from each point of the x1 -axis along the
x2 -axis etc. through to the xn -axis (see Figure 5.2). The result is a vector field V on Cε .
It is smooth because we essentially solve linear ODEs when parallel transporting and thus
have smooth dependence on the initial data. By construction, ∇∂1 V = 0 on the x1 -axis, and
generally ∇∂k V = 0 on the set Mk := {xk+1 = . . . = xn = 0} ⊆ Cε .
We will show by induction on k that
∇ ∂1 V = . . . = ∇ ∂k V = 0 on Mk . (5.8)
For k = 1 it is true by construction. Assume that (5.8) is true for some k < n. By construction,
∇∂k+1 V = 0 on Mk+1 ⊇ Mk . Consider i ≤ k. Since [∂k+1 , ∂i ] = 0 the flatness condition (5.3)
translates to
∇∂k+1 (∇∂i V ) = ∇∂i (∇∂k+1 V ) = 0 on Mk+1 .
| {z }
=0
84 5. CURVATURE

Figure 5.2. Construction of the vector field V in the proof of Lemma 5.14 via parallel
transport of v along the coordinate axes.

Thus ∇∂i V is parallel along the xk+1 -curves starting on Mk . Moreover, ∇∂i V vanishes on Mk
by construction, hence by uniqueness of the parallel transport, ∇∂i V = 0 on each xk+1 -curve,
and hence on all of Mk+1 . This proves (5.8) for k + 1. □
We can now characterize when a Riemannian manifold is locally isometric to Euclidean
space. For the proof we will use another distinctive feature of Euclidean space, namely that
it admits a parallel orthonormal frame globally (see Example 3.35). Using Lemma 5.14 we
will show that flat Riemannian manifolds share this feature locally.
Theorem 5.15. A Riemannian manifold is flat if and only if its curvature tensor vanishes
identically.
Proof. We have already seen in Proposition 5.2 that flatness of the Levi-Civita connec-
tion implies (5.3), which means that the curvature tensor (5.4) vanishes identically.
Conversely, suppose that the curvature tensor of (M, g) vanishes. We first show that g
then admids a parallel orthonormal frame in a neighborhood of each point: For any p ∈ M
and an orthonormal basis (b1 , . . . , bn ) for Tp M there exist by Lemma 5.14 parallel vector fields
E1 , . . . , En on a neighborhood U of p such that
Ei |p = bi .
Since parallel transport is an isometry by Corollary 3.56, the vector fields (Ej ) are also
orthonormal on U .
By the symmetry of the Levi-Civita connection (property (iv) in Theorem 3.27),
[Ei , Ej ] = ∇Ei Ej − ∇Ej Ei = 0,
and so (Ej ) is a commuting orthonormal frame on U .
Thus by the canonical form theorem for commuting vector fields (see [35, Theorem 9.46])
there exist coordinates (y i ) on a (possibly smaller) neighborhood V of p such that
Ei = ∂i for i = 1, . . . , n.
 5.1. RIEMANN CURVATURE TENSOR 85

In these coordinates, we therefore have that on the whole neighborhood6

gij = g(∂i , ∂j ) = g(Ei , Ej ) = δij ,
and hence the chart y = (y 1 , . . . , y n ) is an isometry from the neighborhood V of p to an open
subset of the n-dimensional Euclidean space. □

5.1.4. Symmetries. We have already used that R(X, Y )Z = −R(Y, X)Z in the proof
of Proposition 5.6, but besides this trivial symmetry of R there are several others that follow
from properties of the definition or the Levi-Civita connection, and the Bianchi symmetries
obtained by cyclic permutation.
Proposition 5.16 (Symmetries of the curvature tensor). Let (M, g) be a Riemannian
manifold. The (0, 4)-curvature tensor has the following symmetries for all X, Y, Z, W ∈
X(M ):
(i) Rm(W, X, Y, Z) = −Rm(X, W, Y, Z).
(ii) Rm(W, X, Y, Z) = −Rm(W, X, Z, Y ).
(iii) Rm(W, X, Y, Z) = Rm(Y, Z, W, X).
(iv) Rm(W, X, Y, Z) + Rm(X, Y, W, Z) + Rm(Y, W, X, Z) = 0.
Sketch of proof. (i) is immediate from the definition of the curvature tensor as endo-
morphism because R(W, X)Y = −R(X, W )Y .
(ii) follows from the compatibility of the Levi-Civita connection with the metric g. More
precisely, one needs to show Rm(W, X, Y, Y ) = 0 for all Y (see [36, p. 203] for details) and
then expand Rm(W, X, Y + Z, Y + Z) = 0.
(iv) follows from the symmetry of the connection. In particular, we notice that Z remains
at the same position, so it is sufficient to prove
R(W, X)Y + R(X, Y )W + R(Y, W )X = 0.
By writing everything out one notices that this is just the Jacobi identity
[W, [X, Y ]] + [X, [Y, W ]] + [Y, [W, X]] = 0.
(iii) follows form (i), (ii) and (iv) by adding up four cyclically permuted Bianchi identities.
□

The curvature identity (iv) is called the first Bianchi identity (or algebraic Bianchi in-
dentity). We next prove the second Bianchi identity (or differential Bianchi identity). It has
vast impliciations for General Relativity, which we discuss later.
Proposition 5.17 (Differential Bianchi Identity). The total covariant derivative7 of the
curvature tensor of a Riemannian manifold satisfies
∇Rm(X, Y, Z, V, W ) + ∇Rm(X, Y, V, W, Z) + ∇Rm(X, Y, W, Z, V ) = 0. (5.9)

6Recall that by using normal coordinates we can always have this property g (p) = δ (see Proposi-
ij ij
tion 4.33(ii)), but generally not in a whole neighborhood. This proof thus also explains why curvature has to
become “visible” in higher order terms when the metric components are expressed in normal coordinates.
7Recall the unique construction of a covariant derivative ∇F of a tensor field F ∈ T (k,l) (M ) from Sec-
tion 3.3.
86 5. CURVATURE

Proof. Note that (5.9) is a pointwise condition, and so it suffices to assume that X, Y, Z, V, W
are coordinate vector fields. This way the commutators vanish, i.e., [∂i , ∂j ] ≡ 0. If we use
normal coordinates at p we know, in addition, that the covariant derivatives vanish at p, i.e.,
Γkij (p) = 0 (see Proposition 4.33(iv)).
By Proposition 5.16 the symmetries of Rm imply that (5.9) is equivalent to

∇Rm(Z, V, X, Y, W ) + ∇Rm(V, W, X, Y, Z) + ∇Rm(W, Z, X, Y, V ) = 0. (5.10)

Then by definition of the total derivative ∇ and the compatibility condition (iv) of ∇ with g
we have

∇Rm(Z, V, X, Y, W ) = (∇W Rm)(Z, V, X, Y ) = ∇W ⟨R(Z, V )X, Y ⟩

= ∇W ⟨∇Z ∇V X − ∇V ∇Z X − ∇[Z,V ] X, Y ⟩
(iv)
= ⟨∇W ∇Z ∇V X − ∇W ∇V ∇Z X, Y ⟩.

Cyclic permutation and adding up the three expressions in (5.10) yields

∇Rm(Z, V, X, Y, W ) + ∇Rm(V, W, X, Y, Z) + ∇Rm(W, Z, X, Y, V )

= ⟨R(W, Z)(∇V X) + R(Z, V )(∇W X) + R(V, W )(∇Z X), Y ⟩.

Due to the choice of normal coordinates at p we have ∇V X = ∇W X = ∇Z X = 0 at p, and

so we are done. □

In local coordinates, the following expressions hold.

Corollary 5.18. In terms of the components of the Riemann curvature tensor we have

Rijkl = −Rjikl = −Rijlk = Rklij ,

as well as the algebraic Bianchi identity

Rijkl + Rjkil + Rkijl = 0,

and the differential Bianchi identity

Rijkl;m + Rijlm;k + Rijmk;l = 0. (5.11)

Exercise 5.19. Prove Corollary 5.18.

Remark 5.20 (Ricci identities). One can also prove that the Riemann curvature appears
directly as the obstruction to commutation of the total second covariant derivatives (which
we did not define but are introduced in [36, p. 99 f]). It can be shown that

∇2X,Y Z − ∇2Y,X Z = R(X, Y )Z, (5.12)

for smooth vector fields X, Y, Z ∈ X(M ). This statement can also be translated to 1-forms
and (k, l)-tensor fields (see [36, Thm. 7.14] for more details).
 5.2. SECTIONAL CURVATURE 87

5.1.5. Expansion in normal coordinates. We conclude this section about the local
curvature invariants of a Riemannian manifold by making a connection to Riemann’s original
approach to curvature via an expansion in normal coordinates. Surprisingly, these results
are hard to find and not really covered in standard text books on Riemannian Geometry.
One can find them without proofs in Riemann’s habilitation lecture and Berger’s overview
book [7, p. 202]. Almost complete proofs in the analytic setting can be found in Heckman’s
lecture notes [29, Sec. 3.5], Gray’s book [25, Ch. 9] and Weyl’s comments [48, p. 25 ff] to
Riemann’s original article. Apparently a derivation is also contained in Spivak’s volumes [54].
We will sketch the approach used in [25, Ch. 9] (unfortunately Gray also uses the opposite
sign convention but below it is converted).
Our aim is to observe how curvature describes the deviation of the Riemannian metric
to flat space in a very geometric way. We can also understand Ricci and scalar curvature in
this way. In my opinion, this geometric understanding of curvature is crucial and naturally
resurfaces in the context of smooth curvature comparison theorems (see, for instance, [18]).
Subsequently, such geometric interpretations of curvature reappeared as definitions in what is
now called Synthetic Geometry, that is, as sectional curvature bounds in length metric spaces
(see, for instance, [1, 11, 27]) and as Ricci curvature bounds in metric measure spaces (used
in geometric measure theory mixed with optimal transport theory, see [58]). These ideas are
the forefront of research in Riemannian Geometry today.
Assume that (M, g) is an analytic Riemannian manifold, that is, a manifold with an atlas
whose transition maps and the Riemannian metric (and thus the exponential map) are real
analytic and thus have locally converging power series expansions.
Let p ∈ M and (x1 , . . . , xn ) be normal coordinates at p. The normal coordinate vector
fields ∂i are orthonormal at p, and the components of a covariant tensor field W ∈ T (0,l) (M )
satisfy
∞
X 1
W (∂a1 , . . . , ∂al ) = Wa1 ...al = (∂i . . . ∂ik Wa1 ...al )(p)xi1 . . . xik .
k! 1
k=0
One can express the right hand side in terms of covariant derivatives of W and R [27, Thm.
9.6], which in the case that W is parallel is of the form
l
1X
Wa1 ...al = Wa1 ...al (p) + Riak j s Wa1 ...ak−1 sak+1 ...al (p)xi xj + higher order terms.
6
k=1
The expansion of g in normal coordinates about p can then be obtain as a consequence (we
will not prove it, but see [25, Cor. 9-8]).
Theorem 5.21. Let (M, g) be an analytic Riemannian manifold and (x1 , . . . , xn ) be nor-
mal coordinates about a point p ∈ M . Then the metric tensor is given by
1 1
gij = δij + Rkilj (p)xk xl + ∇k Rlimj (p)xk xl xm + higher order terms.
3 6
5.2. Sectional curvature
At this point we can readily define the sectional curvature of a point p in the plane Π ⊆
Tp M and show that it contains the same information as the Riemann curvature tensor. This
is indeed how many authors of text books on Riemannian Geometry proceed (see, for instance,
[23, 29, 43, 45]). The relation to Riemann’s original approach via the Gaussian curvature of a
88 5. CURVATURE

surface, however, is then lost and needs to be established a posteriori. Indeed, this connection
can only be made precise once we understand the instrinsic (and extrinsic) curvature of
submanifolds. We will therefore return to this geometric interpretation of sectional curvature
in Section 6.2.2.

5.2.1. Definition. As already mentioned, the sectional curvature at a point p depends

on two-dimensional subspaces Π ⊆ Tp M . We think of Π as being spanned by a pair of (linearly
independent) vectors v, w in an inner product space V . By
p
|v ∧ w| := |v|2 |w|2 − ⟨v, w⟩2
we denote the area of the parallelogram spanned by v and w (compare this to the definition
of the canonical Riemannian volumes in Section 2.2.2!).
Recall that, by the Cauchy–Schwarz inequality, |v ∧ w| ≥ 0 with equality if and only if
v and w are linearly dependent. If v and w are orthonormal then |v ∧ w| = 1. Hence the
following definition makes sense.
Definition 5.22. Let (M, g) be a Riemannian manifold of dimension n ≥ 2, and p ∈ M .
If v, w are two linearly independent vectors in Tp M , then the sectional curvature of the plane
Π spanned by v and w is defined by
Rmp (v, w, w, v)
sec(v, w) := . (5.13)
|v ∧ w|2
Note that the sectional curvature can be defined equally well for vector fields X, Y ∈ X(M )
(at least on a maximal open subset of M where Xp and Yp are linearly independent, in order
to guarantee that the denominator is nonzero). For sec(v, w) to be a sensible definition it
remains to observe that sec(v, w) only depends on Π = span{v, w} but not the particular
choice of v and w.
Lemma 5.23. Let (M, g) be a Riemannian manifold. Suppose Π ⊆ Tp M is a nondegenerate
tangent plane at p. Then
sec(Π) = sec(v, w) (5.14)
is independent of the choice of basis v, w for Π.

Exercise 5.24. Prove Lemma 5.23. (Hint: Avoid long calculations by passing from
the basis (v, w) to another basis (ṽ, w̃) and iterating the elementary transformations (a)
(v, w) → (w, v), (b) (v, w) → (λv, w), and (c) (v, w) → (v + λw, w). Observe that sec(v, w) is
invariant under these transformations, which completes the proof.)
Note that the sectional curvature is often denoted by K in place of sec (see [23, 29, 43]).
This is because the sectional curvature of the plane Π ⊆ Tp M is directly related to the
Gaussian curvature K at p of the embedded surface SΠ = expp (Π ∩ V ) in U ⊆ M (where
V is a star-shaped region around 0 in Tp M , which via expp is diffeomorphic to an open
neighborhood U of p in M ), called the plane section determined by Π. We will make this
connection of the sectional and Gaussian curvature precise in Section 6.1, but the following
example already hints that we are on the right track.
 5.2. SECTIONAL CURVATURE 89

Example 5.25 (Curvature of surfaces in R3 ). Suppose M is a smooth surface in E3

with local parametrization X : R2 ⊇ U → R3 and first fundamental form g with coefficients
E = | ∂X 2 ∂X ∂X ∂X 2
∂u | , F = ∂u · ∂v , and G = | ∂v | (see Example 2.33), i.e.,
g = Edu2 + 2F dudv + Gdv 2 .
Let K denote the Gaussian curvature and recall that it is given by (see [45, Sec. 8.1] or the
“Curves and Surfaces” course [28])
LN − M 2
K= ,
EG − F 2
2 2 2
where L = ∂∂ 2Xu
∂ X
· n, M = ∂u∂v · n and N = ∂∂ 2Xv · n with respect to the unit normal vector
Xu ×Xv
n = |Xu ×Xv | are the coefficients of the second fundamental form. In terms of the Riemann
curvature tensor (via the Christoffel symbols, see (5.5)) the Gauss equations read
EK = R211 2 , F K = R212 2 = R121 1 , GK = R122 1 ,
which in turn implies that
R1221 = R122 k gk1 = GKE − F KF = K(EG − F 2 )
and hence
R1221
sec(∂u , ∂v ) = = K.
EG − F 2
Exercise 5.26. Fill in the details in Example 5.25 (in particular, the computation of
EK, F K, and GK, using the Gauss equations).

5.2.2. Relation to the Riemann curvature tensor. Besides the (upcoming) simpler
geometric interpretation of the sectional curvature, we can still show that sec(Π), for all
nondegenerate planes Π, completely determines the Riemann curvature tensor R. While this
result is of great importance, with the above definition of sectional curvature the proof is
purely algebraic (using the algebraic symmetries of R obtained in Proposition 5.16) and no
deep insights are needed.
Theorem 5.27. The Riemann curvature of a Riemannian manifold is completely deter-
mined by its sectional curvature, and vice versa.
Proof. It follows immediately from Definition 5.22 (together with Lemma 5.23¸) that
the sectional curvature is fully determined by the Riemann curvature.
To prove the converse observe that, since both expressions are given pointwise, it suffices to
solve the associated linear problem on a real vector space V of dimension n ≥ 2: Suppose Rm
is a quadrilinear map V × V × V × V → R that satisfies properties (i)–(iv) of Proposition 5.16
(sometimes called form of curvature type on V or algebraic curvature tensor ). Then one can
show that Rm is completely determined by the knowledge of the values Rm(v, w, v, w) for all
v, w ∈ V as follows. By assumption we know Rm(v, w, v, w) for all v, w ∈ V . By replacing v
with v + x in Rm(v, w, v, w) and using the multilinearity and symmetry (iii) we have
Rm(v + x, w, v + x, w) = Rm(v, w, v, w) + 2Rm(v, w, x, w) + Rm(x, w, x, w),
and thus also know Rm(v, w, x, w) for all v, w, x ∈ V . In the next step we replace w with
w + y, and hence obtain
Rm(v, w + y, x, w + y) = Rm(v, w, x, w) + Rm(v, w, x, y) + Rm(v, y, x, w) + Rm(v, y, x, y),
90 5. CURVATURE

and therefore A(v, w, x, y) := Rm(v, w, x, y) + Rm(v, y, x, w) is known for all v, w, x, y ∈ V .

By (iii) and (i) furthermore
A(v, w, x, y) := Rm(v, w, x, y) + Rm(v, y, x, w)
= Rm(v, w, x, y) + Rm(x, w, v, y)
= Rm(v, w, x, y) − Rm(w, x, v, y).

Thus if Rmg is another such quadrilateral form of curvature type with sec = sec,
f then for all
v, w, x, y ∈ V
A(v, w, x, y) = A(v,
e w, x, y),

and thus
Rm(v, w, x, y) − Rm(v,
g w, x, y) = Rm(w, x, v, y) − Rm(w,
g x, v, y).

In other words, the expression Rm(v, w, x, y) − Rm(v,

g w, x, y) is invariant by cyclic permuta-
tions in the first three slots. Applying the algebraic Bianchi identitiy (iv) of Proposition 5.16
thus yields
3[Rm(v, w, x, y) − Rm(v,
g w, x, y)] = 0,

hence Rm is uniquely determined by the sectional curvature. □

5.2.3. Constant sectional curvature. We consider some important special cases re-
lated to sectional curvature.
Definition 5.28. A Riemannian manifold/metric is said to have constant sectional cur-
vature if the sectional curvatures are the same for all planes at all points.
Let us first understand when the sectional curvature is independent of the section in
terms of the Riemann curvature tensor. We already know that this is possible, in principle,
by Theorem 5.27.
Proposition 5.29. Let (M, g) be a Riemannian manifold. Then sec(Π) = c for all planes
Π ⊆ Tp M if and only if
R(v, w)x = c(⟨w, x⟩v − ⟨v, x⟩w) (5.15)
for all v, w, x ∈ Tp M .
For the proof (and more characterizations) we refer to [45, Prop. 3.1.3]. A neat charater-
ization of constant curvature spaces can be given by using the Kulkarni–Nomizu product ?
which is defined in (5.29) below (see [36, Prop. 8.36] for more details).
Proposition 5.30. Let M be an n-dimensional manifold. A Riemannian metric g on M
has constant curvature c if and only if
1
Rm = cg ? g.
2
In this case, the Ricci tensor and scalar curvature of g are given by
Rc = (n − 1)cg, S = n(n − 1)c.
 5.3. RICCI AND SCALAR CURVATURE 91

Once we have understand sectional curvature via the curvature of submanifolds SΠ by

the planes Π we will see in Section 6.2.2 that our standard examples En , Sn , and Hn have
constant sectional curvatures 0, 1 and −1, respectively. This is not surprising, because we
have seen that they are highly symmetric and we know that (local) isometries preserve the
curvature tensor by Proposition 5.12. By using this very argument it is easy to see that all
frame-homogeneous Riemannian manifolds have constant sectional curvature.
Lemma 5.31. If a Riemannian manifold (M, g) is frame-homogenous, then it has constant
sectional curvature.
Proof. Recall that a frame-homogeneous means that we can move one orthonormal basis
to another via an isometry. Thus given any two 2-planes at the same or different points, there
is an isometry taking one to the other. Thus the result follows from the isometry invariance
of the curvature tensor (Proposition 5.12). □

5.3. Ricci and scalar curvature

While we have tried to motivate why the Riemann curvature tensor is defined as it is,
it remains difficult to interpret it geometrically or even grasp what a general Riemannian
(sub)manifold with a particular curvature tensor looks like. As often, it is most insightful to
restrict ourselves to special (but not too special) classes of examples that are easier to grasp.
In the context of sectional curvature we have already mentioned space forms. But there
are also other ways to obtain geometric insight. In what follows we construct well-known
simpler curvature tensors that contain a lot (but in dimensions ≥ 4 not all) information of
the Riemann curvature tensor. The Ricci and scalar curvature are naturally obtained via
metric contractions (briefly already mentioned in Section 2.2.1) but can geometrically be
better interpreted by using the sectional curvature. Both are of immense importance in both
Physics and Geometry.

5.3.1. Definition and symmetries. Note that the definitions of the Ricci and scalar
curvature are different for the opposite sign convention of the Riemann curvature tensor.
Thus, independent of the convention used, the sign of Rc and S are eventually the same. In
our case they read as follows.
Definition 5.32. Let (M, g) be a Riemannian manifold. The Ricci curvature tensor
Rc : X(M )2 → R (or Ric) is a covariant 2-tensor field defined as the trace of the curvature
operator, i.e.,
Rc(X, Y ) := tr(Z 7→ R(Z, X)Y ), X, Y ∈ X(M ).
The components of Rc are denoted by
Rij := Rkij k = g km Rkijm .
The symmetries of the Riemann curvature tensor obtained in Section 5.1.4 naturally descend
to its contractions.
Lemma 5.33. The Ricci curvature is a symmetric 2-tensor field. Componentwise it can
be expressed in any of the following ways:
Rij = Rkij k = Rik k j = −Rki k j = −Rikj k .
92 5. CURVATURE

Exercise 5.34. Prove Lemma 5.33 by applying the symmetries of the Riemann curvature
tensor.

Definition 5.35. Let (M, g) be a Riemannian manifold. The scalar curvature S : M → R

is the function defined as the trace of the Ricci tensor, i.e.,
S = trg Rc = Ri i = g ij Rij .
Also the differential Bianchi identity can be contracted. To formulate it in a clean way
we use the exterior 8 covariant derivative DT ∈ T (0,3) (M ) of a 2-tensor field T , defined by
(DT )(X, Y, Z) := −(∇T )(X, Y, Z) + (∇T )(X, Z, Y ).
In components we have
(DT )ijk = −Tij;k + Tik;j .
Proposition 5.36 (Contracted Bianchi identites). Let (M, g) be a Riemannian manifold.
The covariant derivatives of Riemann, Ricci, and scalar curvature satisfy (the trace being on
the first and last indices)
trg (∇Rm) = −D(Rc), (5.16)
1
trg (∇Rc) = dS, (5.17)
2
which in components reads
Rijkl; i = Rjk;l − Rjl;k , (5.18)
1
Ril; i = S;l . (5.19)
2
Proof. The differential Bianchi identity for R was obtained in (5.11) and reads
Rijkl;m + Rijlm;k + Rijmk;l = 0.
Raising the index m and contracting on i and m yields, by Corollary 5.18,
0 = Rijkl; i + Rijl i ;k + Rij i k;l = Rijkl; i + Rjl;k − Rjk;l ,
| {z }
=−Rijk i ;l

because ∇ commutes with both trace (Proposition 3.16(iv)) and the musical isomorphism9.
This establishes the first equation (5.18).
Contraction on j and k in (5.18) furthermore yield
(5.18)
Ril; i = Rkil k ; i = −Rikl k ; i = −Rkl; k + Rk k ; l,
| {z }
=S;l

hence (5.19). These are the components of (5.16)–(5.17), and so we are done. □

8It is called exterior covariant derivative because it is a generalization of the ordinary exterior derivative
of a 1-form, since (dη)(Y, Z) = −(∇η)(Y, Z) + (∇η)(Z, Y ) (see [36, Problem 5-13]).
9We did not prove this earlier, but it follows immediately from the properties of the covariant derivative for
tensor fields because F ♭ = tr(F ⊗g), g is parallel and therefore ∇X (F ⊗g) = (∇X F )⊗g by Proposition 3.16(iii),
and again because ∇ commutes with tr by Proposition 3.16(iv) therefore
∇X (F ♭ ) = ∇X (tr(F ⊗ g)) = tr((∇X F ) ⊗ g) = (∇X F )♭
(substituting F = G also yields (∇X G)♯ = ∇G♯ ). See also [36, Prop. 5.17].
♯
 5.3. RICCI AND SCALAR CURVATURE 93

Problem 5.37. Suppose (M, g) is a Riemannian manifold and u ∈ C ∞ (M ). Prove

Bochner’s formula:
1
∆( | grad u|2 ) = |∇2 u|2 + ⟨grad(∆u), grad u⟩ + Rc(grad u, grad u).
2
(Hint: Use that the Laplace operator in any local frame reads ∆u = g ij u;ij = u;i i (see
Problem 4.35(ii)) and the Ricci identity βj;pq − βj;qp = Rpqj m βm for β ∈ Ω1 (M ) (you can
assume this without proof, the details are in [36, Thm. 7.14]).)

5.3.2. Geometric interpretation of Ricci and scalar curvature. In Section 5.2

we have already shown that the full curvature information of the Riemann curvature is also
contained in the sectional curvature. So it is only natural to ask how the translates to the
Ricci and scalar curvature. Since the Ricci tensor is symmetric and bilinear it is sufficient to
know the values of Rc(v, v) for unit vectors v, analogous to the polarization identity (2.2).
Proposition 5.38. Let (M, g) be an n-dimensional Riemannian manifold and p ∈ M .
For every unit vector v ∈ Tp M and any orthonormal basis (b1 = v, . . . , bn ) for Tp M we have
n
X
Rcp (v, v) = sec(v, bk ), (5.20)
k=2
X
S(p) = sec(bj , bk ). (5.21)
j̸=k

Proof. Given v ∈ Tp M and (b1 , . . . , bn ), as in the hypothesis we have |bj ∧ bk | = δjk , and
thus it follows immediately from the definitions that
n
X n
X
Rcp (v, v) = R11 (p) = Rk11 k (p) = Rmp (bk , b1 , b1 , bk ) = sec(b1 , bk )
k=1 k=2

and
n
X n
X X
S(p) = Rj j (p) = Rcp (bj , bj ) = Rmp (bk , bj , bj , bk ) = sec(bj , bk ). □
j=1 j,k=1 j̸=k

Remark 5.39 (More averaging). One can also geometrically describe the Ricci curvature
in terms of an integral over the sectional curvature that does not refer to a basis (see [36,
Problem 8-22]). For each v ∈ Tp M
ˆ
n−1
Rcp (v, v) = sec(v, w)dVĝ ,
Vol(Sn−2 ) w∈Sv⊥
where Sv⊥ denotes the set of unit vectors in Tp M that are orthogonal to v and ĝ denotes the
Riemannian metric on Sv⊥ induced from the flat metric gp on Tp M . Similarly, one can obtain
the scalar curvature as an integral over the Ricci curvature [23, Ex. 9 on p. 107].
Furthermore, in Proposition 4.33 and Theorem 5.21 we have seen that in normal coor-
dinates gij (p) = δij + O(r2 ) and that curvature appears in the second order terms of this
expansion. This expansion immediately also yields geometric interpretations of the Ricci and
scalar curvature in terms of volumes.
94 5. CURVATURE

Corollary 5.40. The expansion of the volume form dVg of an analytic Riemannian
manifold (M, g) in normal coordinates about p is given by
dVg (∂1 , . . . , ∂n ) = (dVg )1...n
1 1
= 1 − Rcij (p)xi xj − ∇i Rcjk (p)xi xj xk + higher order terms.
6 12

Exercise 5.41. Compute the second order coefficient in the expansion ofpCorollary 5.40
by assuming Theorem 5.21. (Hint: Recall that in local coordinates dVg = det(gij )dx1 ∧
. . . ∧ dxn .)
We already know that on surfaces only scalar curvature if relevant. Hence the following
result from 1848 should not come as a surprise.
Corollary 5.42 (Bertrand–Diguet–Puiseux Theorem). Let (M, g) be a surface. Then
the area of a geodesic ball at a point p with small radius r > 0 is given by
 
2 1 2 4
Vol(p, r) = πr 1 − Sr + O(r ) .
24

Exercise 5.43. Prove Corollary 5.42

Originally the above theorem was used as another argument to show that the Gaussian
curvature K is intrinsic and does not depend on the embedding (simply because Vol(p, r)
does not). Closely related to is the following conjecture from 1979.
Conjecture 5.44 (Gray–Vanhecke [26]). Let (M, g) be a Riemannian manifold. Suppose
the volume of a geodesic balls in M coincides with the volume of the corresponding Euclidean
ball, that is,
n
(πr2 ) 2
Vol(p, r) = ,
( n2 )!
for all p and all sufficiently small r > 0. Then (M, g) is flat.
The converse is obviously true. By Corollary 5.42 the conjecture holds for n = 2 because
the Gaussian curvature is the only relevant curvature. The conjecture also holds for n = 3.
One can prove it by including higher order terms and using that the Ricci curvature is
sufficient. In dimensions n ≥ 4 the conjecture is still open. It has only been resolved in some
special cases (such as locally symmetric spaces).
Problem 5.45. Let (M, g) be a connected Riemannian manifold. Show that if (M, g) is
homogeneous, then it has constant scalar curvature.

5.3.3. Traceless Ricci curvature and Einstein metrics. The scalar curvature is the
trace of the Ricci tensor, but the Ricci tensor generally contains information that is not
already encoded in its trace. In this section we investigate what information this traceless
part of the Ricci tensor contains.
Definition 5.46. Let (M, g) be an n-dimensional Riemannian manifold. The traceless
Ricci tensor of g is the symmetric 2-tensor
1
R̊c := Rc − Sg.
n
 5.3. RICCI AND SCALAR CURVATURE 95

Proposition 5.47. Let (M, g) be a Riemannian manifold. Then trg R̊c ≡ 0, and the
Ricci tensor decomposes orthogonally as Rc = R̊c + n1 Sg. Therefore,
1 2
|Rc|2g = |R̊c|2g + S . (5.22)
n
Proof. Since in every local frame trg g = gij g ij = δii = n, it follows immediately from
the definition that trg R̊c ≡ 0.
Orthogonality is evident because for any symmetric 2-tensor h (and thus also R̊c) we have
⟨h, g⟩ = g ik g jl hij gkl = g ij hij = trg h,
and thus, in particular, ⟨R̊c, g⟩ = 0.
Finally, (5.22) holds because ⟨g, g⟩ = trg g = n. □

We know establish a connection of the traceless Ricci curvature to the (Riemannian)

Einstein equations. More about the Lorentzian context and application in General Relativity
in Section 5.3.4 below.
Definition 5.48. A Riemannian metric g is called an Einstein metric if
Rc = λg for some constant λ.
The following result is a consequence of the twice-contracted second Bianchi identity and
shows that we do not actually need to assume that λ is a constant.
Proposition 5.49 (Schur’s Lemma). Suppose (M, g) is a connected Riemannian manifold
of dimension n ≥ 3 whose Ricci tensor satisfies
Rc = f g
for some f ∈ C ∞ (M ). Then f is constant and g is an Einstein metric.
Proof. Assume that Rc = f g, then taking the trace on both sides yields
nf = trg (f g) = trg (Rc) = S.
Hence R̊c = Rc − n1 Sg ≡ 0, and therefore ∇R̊c ≡ 0. Since also ∇g ≡ 0 by Proposition 3.23
we have
1 1
∇Rc = (∇S)g = (dS)g, (5.23)
n n
which in coordinates reads Rij;k = n1 S;k gij . Taking the trace in the first and last components
i and k (the last one corresponds to the covariant derivative, not both components of g) yields
by the contracted second Bianchi identity from Proposition 5.36
1 (5.17) (5.23) 1
dS = trg (∇Rc) = dS.
2 n
Since n ≥ 3 we have that dS ≡ 0, and thus S (and therefore f ) is constant on the connected
components of M . □
Corollary 5.50. If (M, g) is a connected Riemannian manifold of dimension n ≥ 3,
then g is Einstein if and only if R̊c ≡ 0.
96 5. CURVATURE

Proof. If g is an Einstein metric with Rc = λg, then taking the trace on both sides
yields λ = n1 S, and therefore
1
R̊c = Rc − Sg = Rc − λg.
n
Conversely, if R̊c = 0, then Rc = n1 Sg, and by Schur’s Lemma g is Einstein. □

Schur’s Lemma is often employed to prove roundness of geometric objects, for instance, to
characterize the limits of convergent geometric flows (such as Ricci flow and mean curvature
flow). So-called almost rigidity results are also true: Camillo De Lellis and Peter Topping
[22] have recently shown that if the traceless Ricci tensor is approximately zero then the
scalar curvature is approximately constant. See Wikipedia for more information (but note
that there is also an unrelated Schur Lemma in Representation Theory).
Problem 5.51. Show that a Riemannian 3-manifold is Einstein if and only if it has
constant sectional curvature.

5.3.4. Curvature in General Relativity (not covered in the course). The name
“Einstein metric” in Section 5.3.3 originates from the general theory of Relativity10 obtained
by Albert Einstein in 1915. In this theory graviation is no longer described as a force but by
the geometry of a 4-dimensional (time-oriented) Lorentzian manifold. More precisely, General
Relativity is the study of solutions (M, g) of the Einstein equation 11
1
Rc − Sg = T, (5.24)
2
where T is the stress-energy tensor describing the matter content of the universe. The reason
for formulating the Einstein equation not just with the Ricci tensor is due to the fact that12
div Rc ̸= 0 but from a physical perspective some kind of local energy conservation in the form
of
div T = 0,
in coordinates T αβ ;β = 0, is desired.
Exercise 5.52. Check that the Einstein tensor G := Rc − 12 Sg is indeed divergence-free.
The special case T ≡ 0 describes the vacuum Einstein equation. Note that taking the
trace on both sides of (5.24) then yields S = 2S (since trg g = dim M = 4), which implies
S = 0. Hence the vacuum Einstein equation is equivalent to
Rc = 0, (5.25)

10Note that Riemann has not made any direct connections to Physics in his habilitation lecture. In fact,
after mentioning the work of Newton he concludes with “Es führt dies hinüber in das Gebiet einer andern
Wissenschaft, in das Gebiet der Physik, welches wohl die Natur der heutigen Veranlassung nicht zu betreten
erlaubt.” (German for “This leads us into the domain of another science, of physics, into which the object of
this work does not allow us to go to-day.”). It is nice to see that today there is actually a very elegant, deep,
and fruitful connection of Riemann’s Geometry to Physics.
11Here (5.24) is normalized so that the discussion of the physical constants can be neglected.
12We have defined the divergence of a vector field in Definition 2.60, but by using the result of Problem 4.35
one generalizes it by div F = trg (∇F ) to any tensor field F .
 5.3. RICCI AND SCALAR CURVATURE 97

which means that the metric g is indeed Einstein in the sense of Definition 5.48. Minkowski
space (described in Example 2.24) is the simplest solution to equation 5.25, the Schwarzschild–
Droste and Kerr metrics describing static and rotating black holes, respectively, are other
prominent solutions.
In 1917 Einstein added a cosmological constant Λ to (5.24) in the hope to obtain static
solutions to (5.24) (a then prevelant point of view) yielding
1
Rc − Sg + Λg = T, (5.26)
2
then removed it again in 1931 after Hubble’s observation of the expanding universe. In the
1990s a positive cosmological constant was added again by the physics community after the
discovery of the accelerating expansion of the universe, which also offers the currently simplest
explanation to dark energy. In any case, in the vacuum setting this would mean that
Rc = Λg, (5.27)
which coincides precisely with the Riemannian definition of Einstein metrics.
Exercise 5.53. Verify that (5.26) with T ≡ 0 indeed implies (5.27)

In the quest for finding a variational formulation of the Einstein equations (which is highly
relevant from a physics perspective), David Hilbert was searching for a suitable action. He
showed that the Einstein equations appear as critical points of the Einstein–Hilbert action
ˆ
S(g) := SdVg . (5.28)
M
In the language of the calculus of variations, the Einstein equation is thus simply the Euler–
Lagrange equation δS(g) = 0 of the functional 5.28.
However, unlike in other physical theories there is no sensible total energy13. In 1915
Hilbert suggested to Emmy Noether to investigate this problem of energy and conservation
of energy. This is how she obtained not only an answer to the problem in General Relativity
but also her first theorem (now called Noether’s Theorem) [42] (see also [17, 46]) determining
the conserved quantities for every system of physical laws that possesses some continuous
symmetry (think of conservation of energy, momentum, center of mass etc.). Unfortunately,
for General Relativity with its much larger infinite symmetry group (the Einstein equations
are invariant under coordinate change, thus diffeomorphism invariant) her first result cannot
be used. Even Noether’s second theorem which is applicable in this setting yields nothing
new in General Relativity, only the contracted second Bianchi identity that we have already
seen holds on any semi-Riemannian manifold anyway.
Nonetheless, conservation laws for energies/masses are still very relevant in General Rel-
ativity. The ADM mass, for instance, is well-defined and nonnegative for asympotically flat
Riemannian initial data slices with nonnegative scalar curvature. This is the celebrated Pos-
itive Mass Theorem of Richard Schoen and Shing-Tung Yau [51, 52] (for dimension n < 8)
and Edward Witten [57] (valid for all dimensions on spin manifolds), both derived around
1980, and for which both Yau and Witten in part obtained their Fields medals.
If you would like to learn more about the interplay between Geometry and General Rela-
tivity I recommend the Master course “Singularities and Black Holes” as a follow-up to this
13See also the discussion of Michael Weiss and John Baez trying to answer Is energy conserved in General
Relativity?
98 5. CURVATURE

course, and if you like it very much the Gravity+ synergy track in the Mathematical Physics
specialization. But even if physics is not for you, let me assure you that the Einstein manifolds
that show up as critical points of the total scalar curvature functional (5.28) are relevant also
in pure Riemannian Geometry today (but note that there are even smooth compact manifolds
that do not admit any Einstein metrics at all, see [8, Ch. 6]).

5.4. Weyl curvature (not covered in the course)

The Ricci and scalar curvature introduced in Section 5.3 contain a lot of important in-
formation of the Riemann curvature tensor but clearly not all. The Weyl curvature tensor
encodes all the rest. We will see that it is closely related to the conformal structure of mani-
fold (recall Section 2.4.3 about conformal transformations) and as such, for instance, also of
great importance in General Relativity. We will only be sketchy in this section, further details
can be found in Lee’s book [36, p. 212–222] and in Petersen’s book [45, Ex. 3.4.23–26].
5.4.1. Definition. The first question to ask is: How much information exactly is con-
tained in the Ricci curvature? A rough answer can be obtained by counting dimensions. The
approach is a bit similar to what we have done in Theorem 5.27 for sectional curvature. As-
suming we are in a real-vector space V , then the space R(V ∗ ) ⊆ T 4 (V ∗ ) of covariant 4-tensors
that have the symmetries (i)–(iv) of Proposition 5.16 of the Riemann curvature tensor turns
out to be
n2 (n2 − 1)
dim R(V ∗ ) =
12
(see [36, Prop. 7.21]). The elements of R(V ∗ ) are called algebraic curvature tensors. The
Ricci tensor, on the other hand, due to being symmetric contains the information of order
n(n − 1), so clearly some information is lost in the process of taking the trace as soon as
n ≥ 4. Our approach is to recover as much as possible of the Riemann curvature tensor, and
express the remaining part by suitable new tensor that is trace-free.
If we endow V with a scalar product g ∈ Σ2 (V ∗ ) (denoting the space of symmetric
tensors) and consider the trace operation trg : R(V ∗ ) → Σ2 (V ∗ ) with respect to the first and
last indices (as in Rc = trg (Rm)), then we would like to know how far from being bijective
the trg -operator is. Assuming surjectivity, injectivity can be studied by constructing a right
inverse to the trace, i.e., a map
G : Σ2 (V ∗ ) → R(V ∗ )
such that
trg (G(h)) = h, h ∈ Σ2 (V ∗ ).
It turns out that this is possible with the help of the Kulkarni–Nomizu product, which for
h, k ∈ Σ2 (V ∗ ) is defined by
h ? k(w, x, y, z) := h(w, z)k(x, y) + h(x, y)k(w, z) − h(w, y)k(x, z) − h(x, z)k(w, y), (5.29)
and in components reads
(h ? k)ijlm = him kjl + hjl kim − hil kjm − hjm kil .
The right inverse for trg is given by
 
1 trg h
G(h) := h− g ? g,
n−2 2(n − 1)
and furthermore
Im G = Ker(trg )⊥
 5.4. WEYL CURVATURE (NOT COVERED IN THE COURSE) 99

(see [36, Prop. 7.23]).

Definition 5.54. Let (M, g) be a Riemannian manifold. The Schouten tensor of g is the
symmetric 2-tensor field
 
1 S
P := Rc − g ,
n−2 2(n − 1)
and the Weyl tensor of g is the algebraic curvature tensor field
1 S
W := Rm − P ? g = Rm − Rc ? g + g ? g.
n−2 2(n − 1)(n − 2)
The Weyl tensor is trace-free.

Proposition 5.55. For every Riemannian manifold (M, g) of dimension n ≥ 3,

trg W = 0,

and
Rm = W + P ? g
is the orthogonal decomposition of Rm corresponding to R(V ∗ ) = Ker(trg ) ⊕ Ker(trg )⊥ .

Exercise 5.56. Prove Proposition 5.55 by using the formula for G above and P ? g =
G(Rc) = G(trg Rm) (show this).

This result simplifies R(V ∗ ) is lower dimensions, and one can show that for dimension
n = 3 the map G is an isomorphism [36, Cor. 7.25]. Therefore, in dimension n = 3 the
composition trg ◦G is the identity and thus also trg an isomorphism. Since trg W = 0 by
Proposition 5.55 it follows that W itself is always zero. Hence in dimension 3 the Ricci
curvature determines the entire curvature tensor. More generally, in dimension n ≥ 3 one
can prove the following formula.

Proposition 5.57 (Ricci decomposition of the curvature tensor). Let (M, g) be a Rie-
mannian manifold of dimension n ≥ 3. Then the (0, 4)-curvature tensor of g has the following
orthogonal decomposition:
1 1
Rm = W + R̊c ? g + Sg ? g. (5.30)
n−2 2n(n − 1)

Exercise 5.58. Prove the formula (5.30) in Proposition 5.57. (Hint: Substitute the
formula for the traceless Ricci curvature, that is, Rc = R̊c + n1 Sg into (5.30) and simplify.
The orthogonality follows from [36, Lemma 7.22].)

Problem 5.59. Derive the formulas of the Riemann curvature tensor Rm in terms of Rc
and S in dimension n = 2 and n = 3. (Hint: The case n = 3 follows from above. See [36, p.
215–216] for more info on n = 2.)
100 5. CURVATURE

5.4.2. Conformal flatness. The strength of the Weyl and Schouten tensors lie in their
remarkable behavior with respect to conformal transformations. Recall that, by Defini-
tion 2.102, conformal metrics g and ge on a manifold M are always related by
ge = e2f g for some f ∈ C ∞ (M ).
One can show that the Levi–Civita connections of g and ge are related by
∇
e X Y = ∇X Y + (Xf )Y + (Y f )X − ⟨X, Y ⟩g grad f. (5.31)

Exercise 5.60. Show that Γ e k = Γk + f;i δ k + f;j δ k − g kl f;l gij by using (3.15). Then derive
ij ij j i
(5.31) by expanding it in coordinates.
By using the properties of the Kulkarni–Nomizu product one can prove the following
result (for the proof see [36, Thm. 7.30]).
Theorem 5.61 (Conformal transformation of curvature). Let (M, g) be a Riemannian
manifold of dimension n, f ∈ C ∞ (M ), and ge = e2f g. Then the curvature tensors of ge are
related to those g by
g = e2f (Rm − (∇2 f ) ? g + (df ⊗ df ) ? g − 1 |df |2 (g ? g)),
Rm (5.32)
g
2
f = Rc − (n − 2)(∇2 f ) + (n − 2)(df ⊗ df ) − (∆f + (n − 2)|df |2g )g,
Rc (5.33)
Se = e−2f (S − 2(n − 1)∆f − (n − 1)(n − 2)|df |2g ), (5.34)
where ∆f = div(grad f ) is the Laplace–Beltrami operator with respect to g as defined in
Section 2.2.3.
If, in addition, dimension n ≥ 3, then
1
Pe = P − ∇2 f + (df ⊗ df ) − |df |2g g, (5.35)
2
f = e2f W.
W (5.36)

Exercise 5.62. Suppose (M, g) is a Riemannian manifold and ge = λg for some λ > 0.
Use Theorem 5.61 to prove that for every p ∈ M and plane Π ⊆ Tp M , the sectional curvatures
of Π with respect to ge and g are related by sec(Π)
f = λ−1 sec(Π).
We see from (5.36) that the conformal curvature structure of a metric tensor is encoded
in the Weyl tensor. In fact, there is an important characterization of local conformal flatness
that was already mentioned earlier (for the proof see [36, p. 216–222]).
Definition 5.63. A Riemannian manifold is said to be locally conformally flat if every
point has an open neighborhood that is conformally equivalent to an open subset in Euclidean
space.
Theorem 5.64 (Weyl–Schouten Theorem). Suppose (M, g) is a Riemannian manifold of
dimension ≥ 3.
(i) If n ≥ 4, then (M, g) is locally conformally flat if and only if the Weyl tensor W is
identically zero.
(ii) If n = 3, then (M, g) is locally conformally flat if and only if the Cotton tensor14
C = −DP is identically zero.
14As before, D denotes the exterior covariant derivative.
 5.4. WEYL CURVATURE (NOT COVERED IN THE COURSE) 101

Problem 5.65. Show that, as long as n ≥ 3, the Weyl tensor vanishes for locally confor-
mally flat g. (Hint: Use that for an embedding φ : U → Rn we have φ∗ ḡ = e2f g =: ge.)
The only other nontrivial case n = 2 is handled using isothermal coordinates. In Sec-
tion 2.4.3 we have sketched that the round sphere S2 and hyperbolic plane H2 are locally
conformally flat. It actually turns out that every Riemannian 2-manifold is locally confor-
mally flat (see, for instance, [19]).
All results above translate verbatim to semi-Riemannian manifolds. The 2-dimensional
Lorentzian result is obtained as follows.
Problem 5.66. Suppose (M, g) is a 2-dimensional Lorentzian manifold, and p ∈ M .
(i) Show that there is a smooth local frame (E1 , E2 ) in a neighborhood of p such that
g(E1 , E1 ) = g(E2 , E2 ) = 0.
(ii) Show that there are smooth coordinates (x, y) in a neighborhood of p such that
(dx)2 − (dy)2 = f g for positive function f ∈ C ∞ (M ). (Hint: Use [36, Prop. A.45]
to show that there exist coordinates (t, u) in which E1 = ∂t , and coordinates (v, w)
in which E2 = ∂v , and set x = u + w, y = u − w.)
(iii) Show that (M, g) is locally conformally flat.

Problem 5.67. Let (M, g) be a Riemannian manifold of dimension n ≥ 3. We define the

conformal Laplacian L : C ∞ (M ) → C ∞ (M ) by
4(n − 1)
Lu := − ∆u + Su,
n−2
where ∆ is the Laplace–Beltrami operator of g and S is its scalar curvature.
(i) Suppose ge = e2f g for some f ∈ C ∞ (M ), and L
e denotes the conformal Laplacian
 n−2 
n+2
with respect to ge. Show that e 2 f Lu
e = L e 2 fu .
(ii) Conclude that ge conformal to g has constant scalar curvature λ if and only if ge =
4 n+2
φ n−2 g, where φ is a smooth positive solution to the Yamabe equation Lφ = λφ n−2 .
 CHAPTER 6

Extrinsic curvature of submanifolds

Until now we have looked at the intrinsic curvature of a Riemannian manifold, and we
have seen that it is a local invariant and such independent of the embedding. In order to
understand the Riemann curvature tensor better as a generalization of Gaussian curvature
of surfaces in R3 to higher dimensions (which means, ultimately, to understand sectional
curvature geometrically) we also need to understand the interplay of curvature and its extrinsic
counterparts that do depend on the embedding.
Many constructions from the “Curves and Surfaces” course will thus reappair here in a
broader context, since we consider embeddings in arbitrary Riemannian manifolds and not
just in Euclidean space.
Some results we present here, in particular, the Gauss and Codazzi equations also hold
in semi-Riemannian geometry and are prominently used in the context of the initial value
formulation of the Einstein equations in General Relativity. Moreover, we will will define
minimal (hyper)surfaces and show some basic properties relating to their curvature. Besides
geodesics, minimal surfaces are another prominent example of variational problems studied
in pure Riemannian Geometry and are highly relevant to this day.

6.1. Second fundamental form

6.1.1. Definition. Let (M, g) be a Riemannian submanifold of a Riemannian manifold
(M
f, ge), that is, with respect to an inclusion ιM : M ,→ M
f we are concerned with the pullback
∗
metric g = ιM ge. In what follows we want to better understand the relationship between the
geometric of M and that of M f. We call M f the ambient manifold.
f|M = T M ⊕ N M
Recall that (see page 33) the orthogonal direct sum decomposition T M
yields the orthogonal tangential and normal projections, denoted by
π⊤ : T M
f|M → T M,

π⊥ : T M
f|M → N M.

We often write X ⊤ := π ⊤ X and X ⊥ := π ⊥ X for a section X of T M f|M .

⊤
Via the tangential projection we defined the tangential connection ∇ which we have seen
is the Levi–Civita connection of the submanifold in the Euclidean case. Now we are more
concerned with the normal component, but basically proceed in the same way: We extend
vector fields X, Y ∈ X(M ) to vector fields on an open subset of M f (still denoted by X and
Y ), apply the ambient covariant derivative operator ∇ and then decompose on M into
e

∇ e X Y )⊤ + (∇
e X Y = (∇ e X Y )⊥ . (6.1)
103
104 6. EXTRINSIC CURVATURE OF SUBMANIFOLDS

Definition 6.1. Let (M, g) be a Riemannian manifold embedded in a Riemannian man-

f, ge). The second fundamental form is the map II : X(M ) × X(M ) → Γ(N M ) given
ifold (M
by
e X Y )⊥ ,
II(X, Y ) := (∇
where X and Y are arbitrarily extended to an open subset of M
f.

Note that II is indeed well-defined since π ⊥ maps smooth sections to smooth sections.
Moreover, the following properties hold.
Proposition 6.2 (Properties of the second fundamental form). Suppose (M, g) is an em-
bedded Riemannnian submanifold of the Riemannian (or semi-Riemannian) manifold (M f, ge),
and let X, Y ∈ X(M ).
(i) II(X, Y ) is independent of the extensions of X and Y .
(ii) II(X, Y ) is C ∞ (M )-bilinear in X and Y .
(iii) II(X, Y ) is symmetric in X and Y .
(iv) The value II(X, Y )|p depends only on Xp and Yp .

Proof. (iii) By symmetry of the connection ∇

e we have

II(X, Y ) − II(Y, X) = (∇ e Y X)⊥ = [X, Y ]⊥ .

e XY − ∇

Since X and Y are tangent to M , so is [X, Y ] [36, Cor. A.40]. Thus [X, Y ]⊥ = 0, and II is
symmetric.
(i and iv) Recall that ∇
e X Y |p only depends on Xp by (3.1), and is thus independent of
the extension.
(ii) Since around every point p a function f ∈ C ∞ (M ) can be extended to a smooth
function in a neighborhood of M in M f linearity of II(X, Y ) in X and Y follows. □

6.1.2. Gauss formula. Let us now turn to the tangential component. In the case that
(Mf, ge) is Euclidean space we have discussed in Example 3.30 that ∇e ⊤ coincides with the
induced Levi–Civita connection of (M, g) which follows from the Fundamental Theorem 3.27
of Riemannian Geometry. The same procedure can be applied in this general setting. Again
it suffices to prove that the map ∇⊤ : X(M ) × X(M ) → X(M ) defined by

∇⊤
X Y := (∇X Y )
e ⊤

is a symmetric connection on M that is compatible with g.

Theorem 6.3 (Gauss Formula). Suppose (M, g) is an embedded Riemannian manifold of
f, ge). For X, Y ∈ X(M ) we have
a Riemannian manifold (M

∇
e X Y = ∇X Y + II(X, Y ). (6.2)

Proof. Due to the orthogonal decomposition ∇ e X Y )⊤ + (∇

e X Y = (∇ e X Y )⊥ it is sufficient
to prove that ∇⊤ X Y := (∇X Y )
e ⊤ = ∇ Y for all points in M . It is easy to check that
X
⊤
∇ : X(M ) × X(M ) → X(M ) is an affine connection on M .
Compatibility with respect to g follows directly from the compatibility of ∇ e with respect
to ge: For X, Y, Z ∈ X(M ) and smooth extensions X, e Y, Z to an open subset of M in M f we
 6.1. SECOND FUNDAMENTAL FORM 105

obtain
∇⊤
X ⟨Y, Z⟩ = X⟨Y, Z⟩ = X⟨Y , Z⟩ = ∇X
e e e e e ⟨Ye , Z⟩
e

= ⟨∇ e + ⟨Ye , ∇
e e Ye , Z⟩ e = ⟨π ⊤ (∇
e e Z⟩ e + ⟨Ye , π ⊤ (∇
e e Ye ), Z⟩ e e Z)⟩
e
X X X X
= ⟨∇⊤ ⊤
X Y, Z⟩ + ⟨Y, ∇X Z⟩,

since Ye and Ze are tangent to M .

We prove symmetry: The vector fields X, Y ∈ X(M ) are ι-related to the smooth extensions
Xe and Ye . Thus by the naturality of the Lie bracket [36, Prop. A.39] also [X, Y ] is ι-related
to [X,
e Ye ]. In particular, [X,
e Ye ] is tangent to M and equal to [X, Y ]. Thus
∇⊤ ⊤
X Y − ∇ Y X = π (∇X
⊤ e e
e Y |M − ∇Ye X|M ) = [X, Y ]|M = [X, Y ].
e e e e
| {z }
=[X,
e Ye ]|M

By the Fundamental Theorem of Riemannian Geometry, it follows that ∇⊤ is the unique

Levi–Civita connection of (M, g) and therefore equal to ∇. □
6.1.3. Geometric interpretation of the second fundamental form. One can use
the Gauss formula to compare intrinsic and extrinsic covariant derivatives along curves. Sup-
pose (M, g) is an embedded Riemannian manifold of a Riemannian manifold (M f, ge), and
γ : I → M is a smooth curve. For X ∈ X(γ) tangent to M we have (see [36, Cor. 8.3])
e t X = Dt X + II(γ ′ , X).
D (6.3)
If γ has unit speed, we define the intrinsic (geodesic) curvature of γ by the map κ : I → R,
κ(t) := |Dt γ ′ (t)|.
Thus γ has vanishing geodesic curvature (in M ) if and only if it is a g-geodesic. However, γ
also has an extrinsic curvature κ
e as curve in M f. By (6.3),
e t γ ′ = II(γ ′ , γ ′ ).
D
One can thus interpret the second fundamental form II(v, v) for v ∈ Tp M as the ge-acceleration
of the g-geodesic γv at p. This observation about curves (and geodesics) can be applied also
to submanifolds.
Definition 6.4. An embedded Riemannian submanifold (M, g) in (M f, ge) is called totally
geodesic if every geodesic in M remains in M forever.
Thus II can be viewed as measuring the failure of M being totally geodsic in M
f (see also
Figure 6.1).
Proposition 6.5. Suppose (M, g) is an embedded Riemannian submanifold in a Riemann-
ian manifold (M
f, ge). Then M is totally geodesic in M
f if and only if the second fundamental
form II vanishes identically.

Problem 6.6. Prove Proposition 6.5.

Problem 6.7. Let G be a Lie group with bi-invariant metric.

(i) Suppose X and Y are orthonormal elements of Lie(G), the Lie algebra of left-
invariant vector fiels on G. Show that sec(Xp , Yp ) = 41 |[X, Y ]|2 for each p ∈ G, and
conclude that the sectional curvature of (G, g) are all nonnegative.
106 6. EXTRINSIC CURVATURE OF SUBMANIFOLDS

Figure 6.1. In the Euclidean space E3 the sphere S2 is not totally geodesic but E2 is
(because ḡ-geodesics are simply straight lines, indicated in blue).

(ii) Show that every Lie subgroup of G is totally geodesic in G.

(iii) Now suppose G is connected. Show that G is flat if and only if it is abelian.

6.1.4. Gauss and Codazzi equations. The second fundamental form is also crucial
in describing the difference in the curvatures of M and M
f.

Theorem 6.8 (Gauss Equation). Suppose (M, g) is an embedded Riemannian submanifold

f, ge). For W, X, Y, Z ∈ X(M ) the following equation holds:
in a Riemannian manifold (M
Rm(W,
g X, Y, Z) = Rm(W, X, Y, Z) − ⟨II(W, Z), II(X, Y )⟩ + ⟨II(W, Y ), II(X, Z)⟩. (6.4)

Proof. Extend W, X, Y, Z arbitrarily to an open neighborhood of Mf. Then by definition

of Rm
g
Rm(W,
g X, Y, Z) = ⟨∇
eW∇ e X Y, Z⟩ − ⟨∇
e X∇
e W Y, Z⟩ − ⟨∇
e [W,X] Y, Z⟩.

By the Gauss formula (6.2)

⟨∇
eW∇
e X Y, Z⟩ = ⟨∇
e W ∇X Y, Z⟩ + ⟨∇
e W II(X, Y ), Z⟩.

Since ⟨II(X, Y ), Z⟩ = 0 on M , the metric compatibility of ∇

e (and again the Gauss formula)
implies
⟨∇
e W II(X, Y ), Z⟩ = −⟨II(X, Y ), ∇
e W Z⟩ = −⟨II(X, Y ), ∇W Z + II(W, Z)⟩
= −⟨II(X, Y ), II(W, Z)⟩.

The same decomposition can be obtained for the second coefficient in Rm(W,
g X, Y, Z). Thus
Rm(W,
g X, Y, Z) = ⟨∇
e W ∇X Y, Z⟩ − ⟨II(X, Y ), II(W, Z)⟩

− ⟨∇
e X ∇W Y, Z⟩ + ⟨II(W, Y ), II(X, Z)⟩ − ⟨∇
e [W,X] Y, Z⟩.

In each term involving ∇e only the tangential component survives (because Z is tangential
to M ), thus we can use g in terms of ge and replace the terms by Rm, and so the Gauss
equation (6.4) remains. □
 6.1. SECOND FUNDAMENTAL FORM 107

Note that the Gauss equation is a scalar equation. In a similar fashion, one can obtain
another vectorial equation in the normal direction. In order to formulate it, one uses the
normal connection ∇⊥ : X(M ) × Γ(N M ) → Γ(N M ), defined by
∇⊥ ⊥
X N := (∇X N ) ,
e

where N is extended to a smooth vector field on a neighborhood M of M

f.

Exercise 6.9. Show that ∇⊥ is a well-defined connection on N M which is compatible

with ge, that is,
X⟨N1 , N2 ⟩ = ⟨∇⊥ ⊥
X N1 , N2 ⟩ + ⟨N1 , ∇X N2 ⟩.

Analogous to the Gauss formula we obtain the Weingarten equation for the Weingarten
map.
Definition 6.10. Let (M, g) an embedded Riemannian submanifold of a Riemannian
f, ge). For each normal vector field N ∈ Γ(N M ), the self-adjoint linear map
manifold (M
WN : X(M ) → X(M ), characterized by
⟨WN (X), Y ⟩ := ⟨N, II(X, Y )⟩,
is called the Weingarten map in direction of N .
Since Π is bilinear over C ∞ (M ), also WN is C ∞ (M )-linear and thus defines a smooth
bundle homomorphism from T M onto itself. It can be used to measure the change in the
normal direction. For the proof see [36, Prop. 8.4] (basically analogous to the Gauss formula).
Proposition 6.11 (Weingarten equation). Suppose (M, g) is an embedded Riemannian
submanifold in a Riemannian manifold (M f, ge). For every X ∈ X(M ) and N ∈ Γ(N M ) we
have the orthogonal decomposition
∇e X N = ∇⊥
X N − WN (X).

Just like the Gauss formula implies Gauss equation, does the Weingarten equation imply
the following vector-valued constraint equation, independently discovered by Karl M. Pe-
tersen (1853), Gaspare Mainardi (1856), and Delfio Codazzi (1868–1869). This equation is
formulated using a combined connection of the tangential connection ∇⊤ and normal connec-
tion ∇⊥ to form a new connection ∇F on any tensor product of copies of T M and N M . In
particular, for any smooth section B : X(M ) × X(M ) → Γ(N M ), we have
(∇FX B)(X, Y ) := ∇⊥
X (B(Y, Z)) − B(∇X Y, Z) − B(Y, ∇X Z).

Exercise 6.12. Prove that ∇F is a connection on F , where F → M denotes the bundle

whose fiber at each point p ∈ M is a set of bilinear maps Tp M × Tp M → Np M (this is a
smooth vector bundle).

Theorem 6.13 (Petersen–Codazzi–Mainardi Equation). Suppose (M, g) is an embedded

Riemannian submanifold in a Riemannian manifold (M f, ge). For all W, X, Y ∈ X(M ) the
following equation holds:
(R(W,
e X)Y )⊥ = (∇FW II)(X, Y ) − (∇FX II)(W, Y ). (6.5)

The proof is similar to that of the Gauss equation, by starting to compute Rm(W,
g X, Y, N )
for N a normal vector field along M using the Gauss equation (see [36, Thm. 8.9] for details).
108 6. EXTRINSIC CURVATURE OF SUBMANIFOLDS

6.2. Sectional curvature revisited

6.2.1. Geometric interpretation of sectional curvature. By using the Gauss equa-
tion we can finally make a connection to Riemann’s original definition of curvature. Let us
recall some notation first. In Section 5.2 we have defined the sectional curvature of a plane
Π ⊆ Tp M spanned by linearly independent vectors v and w (where the choice does not matter)
via the Riemann curvature tensor as
Rmp (v, w, w, v)
sec(Π) := sec(v, w) = .
|v ∧ w|2
Suppose expp : Tp M ⊇ V → U ⊆ M is a local diffeomorphism. Then we call
SΠ := expp (Π ∩ V )
the plane section determined by Π. It is a 2-dimensional embedded submanifold of M , and
swept out by geodesics with initial velocities in Π = Tp SΠ .
From the “Curves and Surfaces” course we know what the Gaussian curvature K of an
embedded Riemannian submanifold of R3 is (see also Example 5.25), and we know that it is an
intrinsic quantity by the Theorema Egregium. How to define it for an abstract Riemannian 2-
manifold (M, g), not necessarily embedded it R3 ? It turns out that for surfaces in R3 we have
that the scalar curvature satisfies S = 2K (see Exercise 6.14 below and also Proposition 5.30),
so we simply define
1
K := S (6.6)
2
also in the general case.
Exercise 6.14. Suppose (M, g) is an embedded Riemannian surface in E3 . Show that
2S = K. (Hint: Use the Gauss equation and express the scalar second fundamental form
with respect to an orthonormal basis.)

Theorem 6.15. Let (M, g) be a Riemannian manifold, p ∈ M and Π a plane in Tp M .

Then
sec(Π) = K(SΠ )(p), (6.7)
where K(SΠ )(p) denotes the Gaussian curvature of SΠ at p.
Proof. Let ĝ denote the induced Riemannian metric on SΠ , Rm d the associated Riemann
curvature tensor, and K(p) its Gaussian curvature at p. In the first step we show how K
b b can
be computed directly from Rmd p . In the second step we show that the second fundamental
form of SΠ vanishes at p. Thus by the Gauss equation the curvature tensor of M and SΠ agree
at p. Since Rm is used for the definition of the sectional curvature, we obtain equation 6.7.
Step 1. Expression of K(p).
b We already know by Lemma 5.23 that the sec(Π) is inde-
pendent of the choice of basis vectors of Π. Assume that (b1 , b2 ) is an orthonormal basis of
Π. Then, by Proposition 5.38 we have that
2
1b 1 X d
K(p)
b = S(p) = Rmp (bi , bj , bj , bi )
2 2
i,j=1
1 d
= (Rmp (b1 , b2 , b2 , b1 ) + Rmp (b2 , b1 , b1 , b2 )) = Rmp (b1 , b2 , b2 , b1 )
d d
2
 6.2. SECTIONAL CURVATURE REVISITED 109

by the symmetries of the Riemann curvature tensor.

Step 2. II = 0 at p. Let z ∈ Π = Tp (SΠ ) be arbitrary, and let γ = γz be the g-geodesic with
initial velocity z. Then γz (t) ∈ SΠ for t sufficiently close to 0, and by the Gauss formula (6.3)
for vector fields along curves therefore
0 = Dt γ ′ = Db t γ ′ + II(γ ′ , γ ′ ).
Due to the right hand side being an orthogonal direct sum both terms must vanish identically.
Thus at t = 0 we have II(z, z) = 0. Since z was arbitrary and Π is symmetric, by the
polarization identity Π = 0 at p.
Step 3. Relation of Rm and Rm. b By the Gauss equation, since II = 0 at p by Step 2, we
have
Rmp (b1 , b2 , b2 , b1 ) = Rm
d p (b1 , b2 , b2 , b2 )
for an orthonormal basis (b1 , b2 ) of Π ⊆ Tp M . Thus by Step 1 it follows immediately from
the definition of sectional curvature in Section 5.2 that K(p) = sec(b1 , b2 ) = sec(Π). □
6.2.2. Constant sectional curvature and model spaces. We now return to complete
manifolds with constant sectional curvature, called space forms. Let us understand the special
cases that we are well acquainted with.
Theorem 6.16. The following Riemannian manifolds have the indicated constant sec-
tional curvature:
(i) (Rn , ḡ) has constant sectional curvature 0.
(ii) (Sn (R),g̊) has consant sectional curvature 1/R2 .
(iii) (Hn (R), ğ) has constant sectional curvature −1/R2 .
For the proof we use Lemma 5.31 for highly symmetric Riemannian manifold that already
implies that we are dealing with constant curvature spaces.
Proof. By Lemma 5.31 is suffices to compute the sectional curvature for one plane at
one point.
(i) Since Rm ≡ 0 on Rn it follows from the definition1 that sec ≡ 0.
(ii) Consider the plane Π spannes by (∂1 , ∂2 ) at the point (0, . . . , 0, R). The geodesics are
great circles with initial velocities in Π in the (x1 , x2 , xn+1 ) subspace. Thus SΠ is isometric
to the 2-sphere of radius R embedded in R3 . Since S2 (R) has Gaussian curvature 1/R2 it
follows that Sn (R) has constant sectional curvature 1/R2 .
(iii) This is Problem 6.17 below. □
For every real number c we thus have a model space with constant sectional curvature c.
These spaces thus play a crucial role in comparison results and local-to-global theorems.
Problem 6.17. Complete the proof of Theorem 6.16 by showing in that the hyperbolic
space of radius R has constant sectional curvature equal to −1/R2 in the hyperboloidal model:
Compute the sectional curvature form of Hn (R) ⊆ R1,n at the point (0, . . . , 0, R) and use the
Gauss equation.

Exercise 6.18. Show that the metric on real projective space RPn has constant sectional
curvature.
1If we were to use the more geometric definition of sectional curvature via the Gaussian curvature of
generated submanifolds, it also follows immediately, because the planes are actually planes and thus have
Gaussian curvature zero.
110 6. EXTRINSIC CURVATURE OF SUBMANIFOLDS

Riemannian manifolds of constant sectional curvature have been widely studied and can
be seen to be special also in other ways: In Lemma 5.31 we have mentioned that a frame-
homogeneous Riemannian manifold has constant sectional curvature, and in Problem 5.51
Riemannian Einstein 3-manifolds are characterized by having constant sectional curvature.
A related result is the following (also taken from [36, Ch. 8]).
Problem 6.19. Suppose (M, g) is a 3-dimensional Riemannian manifold is homogeneous
and isotropic. Show that g has constant sectional curvature. (Remark : An analogous result
in dimension 4 is not true, see [36, Prob. 8-13].)

6.3. Hypersurfaces
We consider now the special situation that M is a hypersurface, that is, a codimension-1
Riemannian submanifold, of (M f, ge). Thus at each point there are exactly two unit normal
vectors, and due to local orientability we can always choose a smooth unit normal vector field
along M in a small neighborhood of each point. If both M and M f are orientable, we can
do so globally. In what follows we only do local computations to ensure existence of a unit
normal field, but it is still important to realize that several concepts and formulas depend on
the chosen orientation.
6.3.1. Scalar second fundamental form and shape operator. Assume we have
chosen a unit normal vector field N on a hypersurface M ⊆ Mf and X, Y ∈ X(M ). We
define the scalar second fundamental form of M as the symmetric covariant 2-tensor field
h ∈ Γ(Σ2 T ∗ M ) given by
(6.2)
h(X, Y ) := ⟨N, II(X, Y )⟩ = ⟨N, ∇
e X Y ⟩,
where the second equality follows from the Gauss formula (6.2) together with the fact that
∇X Y ⊥N . Furthermore, since N is unital and spans N M ,
II(X, Y ) = h(X, Y )N
(note that the sign of h depends on the choice of N , see Figure 6.2 for curves).
The choice of unit normal field N also determines the Weingarten map s := WN : X(M ) →
X(M ), in this situation called the shape operator of M . By the above it is given by
⟨sX, Y ⟩ := ⟨N, II(X, Y )⟩ = h(X, Y ).
Since h is symmetric, s is a self-adjoint endomorphism of T M , that is,
⟨sX, Y ⟩ = ⟨X, sY ⟩, X, Y ∈ X(M ).

The fundamental equations of submanifolds derived earlier, thus simplify for hypersur-
faces.
Problem 6.20. Prove the following equations for a Riemannian hypersurface (M, g) in a
Riemannian manifold (M f, ge), and N a smooth unit normal vector field along M :
(i) Gauss formula: ∇
e X Y = ∇X Y + h(X, Y )N ,
(ii) Weingarten equation: ∇ e X N = −sX,
(iii) Gauss equation: Rm(W, X, Y, Z) = Rm(W, X, Y, Z) − 21 (h ? h)(W, X, Y, Z),
g
(iv) Codazzi equation: Rm(W,
g X, Y, N ) = (Dh)(Y, W, X).
 6.3. HYPERSURFACES 111

Figure 6.2. The scalar second fundamental form h satisfies II(X, Y ) = h(X, Y )N ,
here indicated along a curve γ (with “natural” acceleration γ ′′ ).

Problem 6.21. Suppose (M, g) is a Riemannian hypersurface in an (n + 1)-dimensional

Lorentzian manifold (M f, ge), and N is a smooth unit normal vector field along M . Define
the scalar second fundamental form h and the shape operator s by requiring that II(X, Y ) =
h(X, Y )N and ⟨sX, Y ⟩ = h(X, Y ) for all X, Y ∈ X(M ). Prove the following Lorentzian
analogues of the formulas of Problem 6.20:
(i) Gauss formula: ∇e X Y = ∇X Y + h(X, Y )N ,
(ii) Weingarten equation: ∇ e X N = sX,
(iii) Gauss equation: Rm(W,
g X, Y, Z) = Rm(W, X, Y, Z) + 21 (h ? h)(W, X, Y, Z),
(iv) Codazzi equation: Rm(W,
g X, Y, N ) = −(Dh)(Y, W, X).

Remark 6.22 (Constraint equations in general relativity). Suppose (M f, ge) is an (n + 1)-

dimensional Lorentzian manifold, and assume that ge satisfies the Einstein equation (5.26)
with a cosmological constant Λ,
1
Rc − Sg + Λg = T.
2
Suppose (M, g) is a Riemannian hypersurface in M f with scalar second fundamental form h
as defined in Problem 6.21. One can use the results of Problem 6.21 to show that g and h
have to satisfy the Einstein constraint equations on M , i.e.,
S − 2Λ − |h|2g + (trg h)2 = 2ρ,
div h − d(trg h) = J,
where ρ := T (N, N ), and J := T (N, X). They are the Lorentzian versions of the scalar Gauss
equation and vector-valued Codazzi equation, and after some rewriting can be seen to form a
system on nonlinear elliptic partial differential equations. Solving these constraint equations
is highly nontrivial and several approaches are in use. It is a crucial step in understanding
the initial value formulation of the Einstein equations in General Relativity. In 1952 Yvonne
112 6. EXTRINSIC CURVATURE OF SUBMANIFOLDS

Choquet-Bruhat [24] showed (in the vacuum case) that these conditions are sufficient to obtain
a well-posed initial value problem for the vacuum Einstein equations by suitable transforming
the Einstein equations to a second-order quasilinear wave equation. The step from local
existence to the existence of a maximal globally hyperbolic development is again non-trivial.
The uniqueness of such a maximal development was only settled in 1969 in joint work with
Robert Geroch [20] (for more details on the initial value problem of the Einstein equations
see the excellent and self-contained book by Hans Ringström [49], and his list of errata).

Problem 6.23. Suppose (M, g) is a Riemannian hypersurface in a Riemannian manifold

(M
f, ge), and N is a unit normal vector field along M . We say that M is convex (with respect
to N ) if its scalar second fundamental form satisfies h(v, v) ≤ 0 for all v ∈ T M . Show that if
M is convex and M f has sectional curvatures bounded below by a constant c, then all sectional
curvatures of M are bounded below by c.

6.3.2. Principal curvatures. Recall that the shape operator s is a self-adjoint linear en-
domorphism of the tangent space Tp M (or generally, any such operator on a finite-dimensional
inner product space V ). Consider the smooth map
f (v) := ⟨sv, v⟩.
On the compact set C = {v; q(v) := ⟨v, v⟩ = 1} the maximum is achieved for some v0 ∈ C.
By the Lagrange multiplier rule there is a real number λ0 such that dfv = λ0 dqv , hence by
linearity and selfadjointness of s it follows that for all w ∈ V ,
⟨sv0 , w⟩ = λ0 ⟨v0 , w⟩,
and thus all such maximal v0 are s-eigenvectors with real eigenvalues λ0 due to the nonde-
generacy of the inner product. Consider b1 = v0 to be the first unit basis vector, and set
B := span{b1 }. Clearly, s(B) ⊆ B, and thus by self-adjointness s(B ⊥ ) ⊆ B ⊥ . We can apply
induction to obtain the following result.
Proposition 6.24 (Finite-Dimensional Spectral Theorem). Suppose V is a finite-dimensional
inner product space and s : V → V is a self-adjoint linear endomorphism. Then V has an
orthonormal basis of s-eigenvectors, and all eigenvalues are real.
Suppose (M, g) is a Riemannian hypersurface of a Riemannian manifold (M f, ge) and p ∈
M . Then the shape operator s : Tp M → Tp M has real eigenvalues κ1 , . . . , κn and there is an
orthonormal basis (b1 , . . . , bn ) for Tp M consisting of s-eigenvectors with
sbi = κi bi , i = 1, . . . , n.
Hence s (and h) are in this basis represented by diagonal matrices, and thus
n
X
h(v, w) = ⟨sv, w⟩ = κi v i wi .
i=1

Definition 6.25. Let (M, g) be a Riemannian hypersurface of a Riemannian manifold

(M , ge). The eigenvalues of the shape operator at a point p ∈ M are called the principal
f
curvatures of M at p, and the corresponding eigenspaces are called the principal directions
The following combinations are particularly important.
 6.3. HYPERSURFACES 113

Definition 6.26. Let (M, g) be a Riemannian hypersurface of an (n + 1)-dimensional

Riemannian manifold (M
f, ge). The Gaussian curvature is defined as
K := det(s),
and the mean curvature as
1 1
H := tr(s) = trg (h).
n n
Since the determinant and trace of a linear endomorphism are independent of the basis,
they are well-defined for a chosen unit normal. In terms of the principal curvatures
1
K = κ1 κ2 · · · κn , H = (κ1 + . . . + κn ). (6.8)
n
Exercise 6.27. Prove the formulas (6.8) for the mean and Gaussian curvature. Do these
quantities depend on the choice of normal vector N ? If yes, how?

Exercise 6.28. At this point we have two definitions of the Gaussian curvature for hy-
persurfaces embedded in a Riemannian manifold (via the scalar curvature in (6.6) and via
the shape operator in Definition 6.26). Do they coincide?

Problem 6.29. Suppose U is an open set in Rn and f ∈ C ∞ (U ). Let Γ(f ) = {(x, f (x)); x ∈
U } ⊆ Rn+1 be the graph of f , endowed with the Riemannian metric and upward unit normal.
(i) Compute the components of the shape operator in graph coordinates, in terms of f
and its partial derivatives.
(ii) Let Γ(f ) ⊆ Rn+1 be the n-dimensional paraboloid defined as the graph of f (x) =
|x|2 . Compute the principal curvatures of Γ(f ).

Problem 6.30. Let (M, g) be an embedded Riemannian hypersurface in a Riemannian

manifold (Mf, ge), let F be the local defining function for M and let N = grad F/| grad F |.
(i) Show that the scalar second fundamental form of M with respect to the unit normal
N is given by
−∇ e 2 F (X, Y )
h(X, Y ) = − , X, Y ∈ X(M ).
| grad F |
(ii) Show that the mean curvature of M is given by
 
1 grad F
H = − divge ,
n | grad F |
where n = dim M and divge is the divergence operator of ge as introduced in Prob-
lem 4.35. (Hint: First prove the following linear algebra lemma: If V is a finite-
dimensional inner product space, w ∈ V is a unit vector, and A : V → V is a linear
map that takes w⊥ to w⊥ , then tr(A|w⊥ ) = tr B, where B : V → V is defined by
Bx = Ax − ⟨x, w⟩Aw.)

Problem 6.31. Let M ⊆ Rn+1 be a Riemannian hypersurface, and let N be a smooth

unit normal vector field along M . At each point p ∈ M , Np ∈ Tp Rn+1 ca ben thought
of as unit vector in Rn+1 and therefore as a point in Sn . Thus each choice of unit normal
vector field defines a smooth map ν : M → Sn , called the Gauss map of M . Show that
ν ∗ dVg̊ = (−1)n KdVg , where K is the Gaussian curvature of M .
114 6. EXTRINSIC CURVATURE OF SUBMANIFOLDS

6.3.3. Minimal hypersurfaces. Let us present some important application in the field
of Geometric Analysis of the above concepts, namely that of minimal (hyper)surfaces, which
is a paramount research problem that has stimulated new mathematics since ancient times.
In R3 the question can be phrased as follows (see Figure 6.3): Suppose C is a simple
closed curve in R3 . Is there an embedded (or immersed) surface M that hast least area among
all surfaces with boundary ∂M = C? If so, what is it?

Figure 6.3. A simple closed curve C in R3 , and different surfaces Mi with the same
boundary ∂Mi = C. Which one has least area?

Note that there is a close analogy to the study of length-minimizing curves of Section 4.3,
which is a similar (but simpler) variational problem in one less dimension.
We will treat the above question in the general case where Rn is replaced by an abstract
Riemannian manifold (M f, ge) in which M is embedded as hypersurface. Let us make the
problem precise by giving some definitions.
Definition 6.32. Suppose M is a compact codimension-1 submanifold with nonempty
boundary in an (n + 1)-dimensional Riemannian manifold (M f, ge). We say that M is area-
2
minimizing if it has the smallest area among all compact embedded hypersurfaces in M
f with
the same boundary.
Before fixing the boundary, we derive a necessary condition of area-minimizing compact
hypersurfaces. For the proof, note that most of results we have shown earlier are true for
manifolds with boundary, although we did not mention it explicitely. Moreover, some details
are only sketched, since we did not cover Fermi coordinates (a generalization of normal coor-
dinates) and will simply take their existence for grated and refer to literature for proofs (see,
e.g., [25, 36]).
Theorem 6.33. Let M be a compact codimension-1 submanifold with nonempty boundary
in an (n + 1)-dimensional Riemannian manifold (M
f, ge). If M is area-minimizing, then its
mean curvature is identically zero.
Proof. Let g be the induced metric on M . By assumption M minimizes the area among
hypersurfaces with the same boundary. Thus, in particular, M minimizes the area among
small perturbations of M in a neighborhood of an interior point p ∈ Int(M ). We use this
idea to set up a one-dimensional variational problem and show that M must have zero mean
curvature H.
2The term area is only used due to it’s analogy with the 2-dimensional question above. Strictly speaking
it is the n-dimensional volume of M with respect to its induced Riemannian metric.
 6.3. HYPERSURFACES 115

Step 1. Write ge in Fermi coordinates adapted to M . For each p ∈ Int(M ) one can
construct so-called Fermi coordinates3 (x1 , . . . , xn , v) on a neighborhood U e ⊆Mf of p ∈ M
1 n
[36, p. 135 f], that is, coordinates such that (x , . . . , x ) are coordinates on M and ge is of the
form
ge = dv 2 + gαβ (x, v)dxα dxβ (6.9)
[36, Prop. 6.37, Ex. 6.43, p. 238 f]. Set U := Ue ∩ M and assume without loss of generality
e ∩ ∂M = ∅.
that it is a regular coordinate ball in M and U
Step 2. Construct perturbations Mt of M . Take an arbitrary funciton φ ∈ C ∞ (M ) with
compact support in U . For sufficiently small t we can define a set (see Figure 6.4)
e ; v(z) = tφ(x1 (z), . . . , xn (z))} ⊆ M
Mt := (M \ U ) ∪ {z ∈ U f.

Figure 6.4. The submanifold Mt is a perturbation of M and only differs in a neigh-

borhood U of p.

Then Mt is an embedded hypersurface in M f which largely agrees with M but is a

graph v = tφ in Ue . One can then construct two new Riemannian metrics using the graph
parametrization ft (x) = (x, tφ(x)) and resulting diffeomorphism Ft : M → Mt , given by,
(
z, z ∈ M \ supp φ,
Ft (z) =
ft (z), z ∈ U,
with the following properties:
(i) the induced Riemannian metric ĝt := ι∗Mt ge on Mt , and
(ii) the pullback metric gt := Ft∗ ĝt = Ft∗ ge on M .
Note that, for t = 0, M0 = M and g0 = ĝ0 = g. It follows from (6.9) that
(
g, on M \ U,
gt =  2  α β
t ∂α φ(x)∂β φ(x) + gαβ (x, tφ(x)) dx dx , in U,
By assumption and construction we have thus obtained a smooth hypersurface Mt with same
boundary ∂Mt = ∂M such that Area(Mt , ĝt ) is minimal for t = 0.
Step 3. Compute the area of (Mt , ĝt ). By definition of the parametrized metrics, Ft : (M, gt ) →
(Mt , ĝt ) is an isometry. Thus
Area(Mt , ĝt ) = Area(M, gt ) = Area(M \ U, g) + Area(U, gt ),
3Fermi coordinates, named after the Italian mathematician Enrique Fermi (1922), are a generalization
of normal coordinates when one moves from (a geodesic ball around) a point p to (a neighborhood of) a
submanifold M by making use of properties of the exponential map expN M : N M → M defined on the normal
bundle of M in M
f (namely that it is a local diffeomorphism of the zero section of N M ). Many results about
normal coordinates can be generalized to this setting as well. See [25, Ch. 2] for full details.
116 6. EXTRINSIC CURVATURE OF SUBMANIFOLDS

and it remains to compute the second term on U . In the chosen coordinates,

ˆ p
Area(U, gt ) = det gt dx1 · · · dxn ,
U
where det gt = det((gt )αβ ). Note that the integrand depends smoothly on t and (x1 , . . . , xn ).
Step 4. Vary the area. By Step 3, t 7→ Area(Mt , ĝt ) is a smooth function and we can
compute the variation by
ˆ
d ∂ p
Area(Mt , ĝt ) = det gt dx1 · · · dxn
dt t=0 U ∂t t=0
ˆ
1 ∂
= (det g)−1/2 (det gt ) dx1 · · · dxn ,
U 2 ∂t t=0
since the integrand has compact support on U . One can show that
∂ ∂gαβ
(det gt ) = (det g)g αβ φ = (det g)(−2nH)φ,
∂t t=0 ∂v
where H is the mean curvature of (M, g) (see [36, p. 241] and [36, Prop. 8.17] for the
computation of the mean curvature in these coordinates, see also Exercise 6.34 below). Thus
ˆ
d
Area(Mt , ĝt ) = −n HφdVg . (6.10)
dt t=0 U
Step 5. Conclude that mean curvature H ≡ 0. By assumption, Area(Mt , ĝt ) attains a
minimum at t = 0, thus (6.10) implies that for every such φ,
ˆ
HφdVg = 0. (6.11)
U
Suppose, for contradiction, that H(p) ̸= 0. Suppose, without loss of generality, that H(p) > 0.
Then
´ we can pick a nonnegative bump function φ with φ(p) > 0 and supp φ ⊆ {H > 0} Then
U HφdV g > 0, which contradicts (6.11). Thus H ≡ 0 on U , and since p was arbitrary, H ≡ 0
on Int(M ). Thus by continuity, H ≡ 0 on all of M . □
1 αβ
Exercise 6.34. Verify that H = − 2n g ∂v gαβ |v=0 used in Step 4 indeed holds. (Hint:
Use (6.9) and look into [36, Prop. 8.17].)
Because of Theorem 6.33 the following terminology has been established.
Definition 6.35. Let M be a compact codimension-1 submanifold with ∂M ̸= ∅ in
Riemannian manifold (Mf, ge). If the mean curvature of M vanished identically, i.e., H ≡ 0,
then we call M a minimal (hyper)surface.
Note that, strictly speaking, the condition H ≡ 0 only implies that M is a critical point
of the area. But one can, just as in the case of geodesics, show that minimal hypersurfaces
are locally area-minimizing.
Remark 6.36 (Minimal surfaces in General Relativity). Certain extensions and modifica-
tions of the notion of minimal surface are significant in General Relativity. They are known as
apparent horizons, and provide a curvature-based approach to understanding the boundary
of black holes. More precisely, they mark the boundary between outward-moving light that
moves outward or inward. Unlike the event horizon (which is a global boundary), however,
apparent horizons are local and depend on the slicing. An apparent horizon is often also
 6.3. HYPERSURFACES 117

called marginally outer trapped surface (MOTS) because it is the outermost of all trapped
surfaces.

In the language of Calculus of Variations, Theorem 6.33 is a variational problem without

contraints. There, we fixed the boundary, but now we want to require that the interior has a
certain fixed volume instead. We can turn our initial question into a problem with contraints
analogous to Dido’s problem in R2 : Queen Dido considered a regular domain D in R2 which is
enclosed by a curve C. Her aim was to find the domain D with maximal area for fixed length
L(∂D) = L(C) (but arbitrary shape). Clearly, the solution is a ball D of a certain radius,
and the boundary C is the corresponding circle (see [14] for several proofs and a presentation
to the general public with an elementary geometric proof). Instead of fixing the length of
C and maximizing the area of D one can, equivalently, fix the area of D and minimize the
length of C. This is what we do in more generality now. The role of the curve C is played
by a hypersurface M , and (Mf, ge) replaces E2 .

Theorem 6.37. Suppose (M f, ge) is an (n + 1)-dimensional Riemannian manifold, D ⊆

f is a compact regular domain4, and M = ∂D. If M is surface area-minizing among all
M
boundaries of compact regular domains with the same volume as D, then M has constant
mean curvature (computed with respect to the outward unit normal).
The proof is somewhat analogous (but more involved, because 2-paramater perturbations
of D are needed) than that of Theorem 6.33.
Proof. Let g be the induced metric on M . Assume, for contradiction, that the mean
curvature of M is not constant, that is, that there are points p, q ∈ M with H(p) < H(q).
Step 1. Write ge in Fermi coordinates adapted to M . Since M is compact it has an ε-
tubular neighborhood for some ε > 0 (see Remark 4.40). As in the proof of Theorem 6.33
there are Fermi coordinates (x1 , . . . , xn , v) for M on an open set U e ⊆M f containing p. Let
U := Ue ∩ M and assume, without loss of generality, that U is a regular coordinate ball in M
and that the image of the chart is a cylindrical set of the form U b × (−ε, ε) for some open set
Ub ⊆ Rn . Similarly, let (y 1 , . . . , y n ) be Fermi coordinates for M on W
f⊆M f containing q, and
let W := Wf ∩ M (with same properties as U ).
By potentially changing the signs of v and/or w, we have that
D∩U
e = {v ≤ 0}, D∩W
f = {w ≤ 0}.
By further shrinking the domains U and W we can assume that we have barriers, i.e., constants
H1 and H2 with H(p) < H1 < H2 < H(q), so that in the whole domains still
H|U ≤ H1 < H2 ≤ H|W . (6.12)
Step 2. Construct perturbations Ds,t of D. Consider φ, ψ ∈ C ∞ (M ) with compact sup-
ports in U and W , respectively, and normalization
ˆ ˆ
φdVg = ψdVg = 1. (6.13)
U W

For s, t ∈ R sufficiently small, define the sets Ds,t ⊆ M

f by (see Figure 6.5)

Ds,t := {z ∈ U
e ; v(z) ≤ sφ(x(z))} ∪ {z ∈ W
f ; w(z) ≤ tψ(y(z))} ∪ (D \ (U
e ∪W
f ),

4Recall that a regular domain is a closed, embedded codimension-0 submanifold with boundary.
118 6. EXTRINSIC CURVATURE OF SUBMANIFOLDS

and set
Ms,t := ∂Ds,t
Clearly, for s = t = 0, we obtain D0,0 = D and M0,0 = M . For sufficiently small s and t, Ds,t
is a regular domain and Ms,t a compact smooth hypersurface, and the functions
(s, t) 7→ V (s, t) := Vol(Ds,t ), (s, t) 7→ A(s, t) := Area(Ms,t )
are smooth.

Figure 6.5. The regular domain Ds,t is a perturbation of D and only differs in the
disjoint neighborhoods U of p and W of q.

Step 3. Vary the volumes of Ds,t and areas of Ms,t = ∂Ds,t . As in (6.10) one can show
that
ˆ ˆ
∂A ∂A
(0, 0) = −n HφdVg , (0, 0) = −n HψdVg . (6.14)
∂s U ∂t W
It remains to compute the variation of the volume at (s, t) = (0, 0). If we hold t = 0 fixed and
vary s, the only volume change of Ds,t takes place in U e . Thus by the fundamental theorem
of calculus,
∂V d
(0, 0) = Vol(Ds,0 ∩ U
e)
∂s ds s=0
ˆ ˆ sφ(x) p !
d
= det ge(x, v)dv dx1 · · · dxn
ds s=0 U −ε
ˆ ˆ sφ(x) p !
d
= det ge(x, v)dv dx1 · · · dxn
U ds s=0 −ε
ˆ p
= φ(x) det ge(x, 0)dx1 · · · dxn
U
 6.3. HYPERSURFACES 119

Since in the chosen coordinates gαβ (x) = geαβ (x, 0),

ˆ
∂V
(0, 0) = φ dVg = 1. (6.15)
∂s U
Similarly, ∂V∂t (0, 0) = 1.
Step 4. Conclude that H is constant. Step 3 allows us to combine s and t as follows.
Since ∂V∂t (0, 0) ̸= 0, the implicit function theorem guarantees that there is a smooth function
λ : (−δ, δ) → R for some small δ > 0 so that
V (s, λ(s)) ≡ V (0, 0) = Vol(D).
Differentiation with respect to s yields by the chain rule and Step 3
d ∂V ∂V (6.15)
0= V (s, λ(s)) = (0, 0) + λ′ (0) (0, 0) = 1 + λ′ (0),
ds s=0 ∂s ∂t
i.e., λ′ (0) = −1.
By assumption, M minimizes the area, hence also
d
0= A(s, λ(s)),
ds s=0
and, similarly, by the chain rule, Step 3, and λ′ (0) = −1,
ˆ ˆ
∂A ∂A (6.14)
0= (0, 0) + λ′ (0) (0, 0) = −n HφdVg + n HψdVg .
∂s ∂t U W
Thus ˆ ˆ
HφdVg = HψdVg .
U W
On the other hand, due to our restrictions of U and W so that (6.12) holds, and the normal-
ization (6.13) of φ and ψ, imply that
ˆ ˆ
HφdVg ≤ H1 < H2 ≤ HψdVg ,
U W
a contradiction. □
Theorem 6.37 suggest the following definition.
Definition 6.38. An immersed hypersurface M in a Riemannian manifold (M f, ge) is said
to be a constant-mean-curvature (CMC) surface if its mean curvature is constant.
Figure 6.6 summarizes the results of Theorems 6.33 and 6.37 visually for (M f, ge) = E3 . In
the first setting, we showed that a area-minimizing surface M with fixed boundary must have
zero mean curvature (here, a flat soap film). In the second setting, we only fixed the volume
of the interior domain D and searched for the surface M enclosing it with least surface area.
We concluded that M must have constant mean curvature (here, a round soap bubble).
In the spirit of our original question we can further ask ourselves: What are the com-
pact CMC surfaces in R3 ? Certainly, spheres S2 (R) have constant mean curvature and are
compact, but are they the only ones?
The answer is far from trivial. Already in 1841 Charles-Eugène Delaunay showed that the
only surfaces of revolution with constant mean curvature are surfaces obtained by rotating
the roulettes of the conics. These are the plane, cylinder, sphere, the catenoid, the unduloid
and nodoid. All of those, except the sphere, however, are not compact. On the other hand,
120 6. EXTRINSIC CURVATURE OF SUBMANIFOLDS

Figure 6.6. The results of Theorems 6.33 (left) and 6.37 (right) visualized in Euclidean
3-space. On the left hand side the boundary ∂M is fixed and the surface M with
minimal area must have zero mean curvature, and is a minimal (hyper)surface. On
the right hand side, instead, only the volume of a compact regular domain D is given,
and the surface M which encloses this domain with minimal area must have constant
mean curvature, and is a CMC surface.

there may be other compact CMC surfaces that cannot be obtained by rotation. In 1853
J. H. Jellet showed that also in the class of compact star-shaped surfaces only the sphere
remains. Finally, A. D. Alexandrov proved in a series of six papers in 1956–1960 that a
compact embedded surface in R3 with constant mean curvature H ̸= 0 must indeed be a
sphere. At this point one would expect, and Heinz Hopf indeed conjectured it in 1956, that
this result can be extended to higher dimensions and to immersions as well, namely that
any immersed compact orientable CMC hypersurface in Rn+1 must be a standard embedded
n-sphere. But this is false! In 1982 Wu-Yi Hsiang constructed a counterexample in R4 . It
is also false in R3 if one replaces embedded by immersed, an example being the Wente torus
constructed in 1984.
Generally, even in the non-compact setting, up until then only a few examples of general
CMC surfaces in R3 were known. This has changed with more modern glueing constructions
in the 1990s. See [9] for an up-to-date (May 2022) and thourough survey of current research
on CMC (hyper)surfaces using conservation laws and glueing.
Remark 6.39 (CMC slices in General Relativity). When studying the initial value prob-
lem of the Einstein equation it is necessary to pick a Cauchy surface, a suitable Riemannian
3-manifold representing an instance of “time”. One way is to pick/assume a CMC slicing of
the spacetime. In this case, the time is exactly the constant mean curvature of the hyper-
surface Σt , i.e., H(Σt ) = t. Global existence for the Einstein vacuum evolution equations in
CMC time is a long standing conjecture in relativity. Not all spacetimes have CMC slices,
and so a crucial question concerning CMC slices is also their generality. See, for instance,
[3, 4, 21] for some discussion of these issues.
 APPENDIX A

Tensors and tensor fields

The most important constructions on smooth manifolds for Riemannian Geometry are
tensor fields. They are sections of tensor bundles, which is a special kind of vector bundle.
We briefly review all these constructions here. Most are already covered in the ”Manifolds”
course. See [36, Appendix] and [35] for more details and exercises. Here we just repeat the
basic definitions. We assume a background in basic Linear Algebra.
Note that there is a shorter equivalent way to introduce the set of tensor fields as a
C ∞ (M )-module (see [43, Chapter 2]) but then the relation to vector bundles is still open,
and this vector bundle viewpoint is sometimes beneficial (in this course, for example, in the
context of connections). For us this interpretation as a C ∞ (M )-module is a consequence and
stated in Lemma A.17.

A.1. Vector bundles

Definition A.1. A (real) smooth 1 vector bundle of rank k is a pair of smooth manifolds
E and M (with or without boundary), together with a smooth surjective map π : E → M
satisfying
(i) For each p ∈ M , the set Ep = π −1 (p) is a k-dimensional vector real vector space.
(ii) For each p ∈ M , there exists a neighborhood U of p and a diffeomorphism ΦU : π −1 (U ) →
U × Rk such that
• the diagram
ΦU
EU = π −1 (U ) U × Rk
π
prU
U
commutes, where prU : U × Rk → U is the projection onto the first factor; and
• for each q ∈ U , ΦU restricts to a linear isomorphisms Eq → {q} × Rk ∼ = Rk .
The space M is called the base, E is called the total space, and π : E → M its projection. Each
set Ep = π −1 (p) is called the fiber of E over p, and each diffeomorphism ΦU : π −1 (U ) → U ×Rk
as above a smooth local trivialization of E over U .
There are certain special cases of vector bundles. For instance, a rank-1 vector bundle
is called a line bundle. If there exists a global trivialization of E over all of M , i.e., E is
diffeomorphic to M × Rk , then E is called a trivial bundle (or smoothly trivial ).
The simplest example of a rank-k bundle is therefore the following.

1Generally, smoothness is not required, only continuity for π (and Φ being a homeomorphisms), which
U
is why we call it a smooth vector bundle. We only consider smooth vector bundles.
121
122 A. TENSORS AND TENSOR FIELDS

Example A.2 (Product bundle). The product space E = M ×Rk with π = pr1 : M ×Rk →
M as projection is a product bundle. The identity map is a global trivialization, hence it is
trivial.
Most vector bundles are not trivial, hence they require more than one local trivialization.
Example A.3 (Möbius strip). On S1 × R define the equivalence relation
(θ, t) ∼ (θ′ , t′ ) :⇐⇒ (θ′ , t′ ) = (θ + π, −t),
and let E = (S1 × R)/Z2 be the quotient space with quotient map q : S1 × R → E (see
Figure A.1). One can show that the projection pr1 : S1 × R → S1 onto the first factor decends
to a (smooth) surjective map π : E → S1 with the desired properties, turning π : E → S1 into a
smooth real line bundle over S1 , called the Möbius strip. (For more details see [35, Ex. 10.3].)

Figure A.1. Rough sketch of the Möbius strip (you can find nice vidoes online, craft
one your self, or read a bit more about its math and relevance for the world).

The most important examples of vector bundles are tangent bundle T M , cotangent bundle
T ∗M , and tensor bundles T (k,l) T M (for the latter see Section A.3).
Exercise A.4 (Tangent bundle). Let M be a smooth n-dimensional manifold and let T M
be its tangent bundle. Show that with its standard projection map, its natural vector space
structure on each fiber, and the natural smooth structure of a 2n-dimensional manifold (see,
e.g., [35, Prop. 3.18]), T M is a smooth vector bundle of rank n over M . (Hint: Locally on
∂
any smooth chart (U, φ = (x1 , . . . , xn )) define ΦU (v i ∂x 1 n
i |p ) = (p, dφp (v)) = (p, (v , . . . , v )),
and verify that it is a smooth local trivialization.)
We now turn to define sections on vector bundles E → M in an attempt to generalize
vector fields X(M ) on the tangent bundle T M . We will use it to define tensor fields in
Section A.3.
 A.2. TENSORS ON A VECTOR SPACE 123

Definition A.5. A smooth 2 section of a vector bundle π : E → M is a smooth map

σ : M → E such that π ◦ σ = IdM (equivalently, σ(p) = Ep for all p ∈ M ).
The set of all sections of E is denoted by Γ(E).
Example A.6 (Smooth functions). The sections of the trivial line bundle M × R are
simply the smooth real-valued functions, i.e., Γ(M × R) = C ∞ (M ).
Example A.7 (Vector fields and 1-forms). The sections of the tangent bundle T M are
precisely the vector fields, i.e., Γ(T M ) = X(M ). The sections of the cotangent bundle T ∗ M
are the 1-forms, i.e., Γ(T ∗ M ) = Ω1 (M ). For smooth k-forms we have Ωk (M ) = Γ(Λk T ∗ M ).

A.2. Tensors on a vector space

Let V be an n-dimensional real vector space and V ∗ the dual space of V , i.e., the set of
linear maps V → R. While the elements of V are called vectors, the elements of V ∗ are called
covectors.
Definition A.8. Let k, l ∈ N ∪ {0}. A multilinear map
∗
F: V · · × V }∗ × |V × ·{z
| × ·{z · · × V} → R
k copies l copies

is called a (mixed) (k, l)-tensor , or k-contravariant, l-covariant tensor.

The spaces of tensors on V are denoted by
T l (V ∗ ) := {covariant l-tensors on V },
T k (V ) := {contravariant k-tensors on V },
T (k,l) (V ) := {mixed (k, l)-tensors on V }.
The rank of a tensor is the number or arguments it takes, i.e., k + l.
By convention, T (0,0) V = R. Note that T l (V ∗ ) = V ∗ and T k (V ) = V . One can show
(see, for instance, [36, Prop. B.1]) that
T (1,1) (V ) ∼
= End(V ).
Natural and important operations on tensors are tensor products and trace (or contrac-
tion).
Definition A.9. Let F ∈ T (k,l) (V ), G ∈ T (p,q) (V ) be tensors. The tensor product
F ⊗ G ∈ T (k+p,l+q) (V ) is defined by
F ⊗ G(ω 1 , . . . , ω k+p , v1 , . . . , vl+q )
= F (ω 1 , . . . , ω k , v1 , . . . , vl )G(ω k+1 , . . . , ω k+p , vl+1 , . . . , vl+q ).
The tensor product is associate because multiplication in R is.
Exercise A.10. Suppose (bi ) is a basis for the n-dimensional vector space V , and (β j ) is
the corresponding dual basis for V ∗ . Show that the set of all tensors
bi1 ⊗ . . . ⊗ bik ⊗ β j1 ⊗ . . . ⊗ β jl

2Again, we only consider smooth sections, while generally one can also study continuous sections.
124 A. TENSORS AND TENSOR FIELDS

is a basis for T (k,l) (V ), i.e., every tensor F ∈ T (k,l) (V ) is of the form (using the Einstein
summation convention)
F = Fji11...j
...ik
l
bi1 ⊗ . . . ⊗ bik ⊗ β j1 ⊗ . . . ⊗ β jl ,
which components Fji11...j
...ik
l
= F (β i1 , . . . , β ik , bj1 , . . . , bjl ). Compute dim T (k,l) (V ).

Definition A.11. The trace (or contraction) is the operator tr : T (k+1,l+1) (V ) → T (k,l) (V ),
which for F ∈ T k+1,l+1 (V ) is defined by
(tr F )(ω 1 , . . . , ω k , v1 , . . . , vl ) := tr(F (ω 1 , . . . , ω k , ·, v1 , . . . , vl , ·)).
| {z }
∈T (1,1) (V )

In terms of a basis, the componontes of tr F are (tr F )ij11...i k i1 ...ik m

...jl = Fj1 ...jl m .

Exercise A.12. Show that the trace on any pair of indices (one upper and one lower) is
a well-defined linear map T (k+1,l+1) (V ) → T (k,l) (V ).
To define differential forms one needs alternating tensors, for Riemannian geometry sym-
metric tensors are more important.
Definition A.13. A covariant tensor F ∈ T l (V ∗ ) is said to be symmetric if for any pair
of arguments 1 ≤ i < j ≤ k
F (v1 , . . . , vi , . . . , vj , . . . , vk ) = F (v1 , . . . , vj , . . . , vi , . . . , vk ).
The set of symmetric l-tensors on V is a linear subspace, denoted by Σl (V ∗ ).

A.3. Tensor fields on a manifold

On a smooth manifold M , for each p ∈ M , the tangent space Tp M is a vector space and
thus one can consider tensors on Tp M . The disjoint union of tensors of the same type defines
a tensor bundle of this type.
Definition A.14. The bundle of (k, l)-tensors on M is given by
G
T (k,l) T M := T (k,l) (Tp M ).
p∈M

Tp M and the contangent bundle T ∗ M = Tp∗ M

F F
The tangent bundle T M = p∈M p∈M
are special cases.
Exercise A.15. Show that each tensor bundle is a smooth vector bundle (see Defini-
tion A.1) over M , with a local trivialization over every open subset that admits a smooth
local frame for T M .
Recall that we have seen in Example A.7 that vector fields and 1-forms are sections of
the tangent and cotangent bundle, respectively. They are also special cases of tensor fields
(with (k, l) being (1, 0) and (0, 1)). Given Exercise A.4, we now define a tensor field on M as
a section (see Definition A.5) of a smooth tensor bundle over M .
Definition A.16. The sections of the tensor bundle T (k,l) T M are the (k, l)-tensor fields
3

T (k,l) (M ) := Γ(T (k,l) T M ).

3There are also other ways to denote (k, l)-tensor fields, for example, as T k (M ) [43, p. 35].
l
 A.3. TENSOR FIELDS ON A MANIFOLD 125

Note that T (1,0) (M ) = X(M ) and T (0,1) (M ) = Ω1 (M ). For smooth covariant l-tensor
fields we write T l (M ) := Γ(T l T ∗ M ) := Γ(T (0,l) T M ).
Multilinearity over the space of smooth functions characterizes tensor fields [35, Lemma
12.24] as a module. O’Neill [43, p. 55 ff] actually defines tensor fields in this way.
Lemma A.17 (Tensor Characterization Lemma). A map
F : Ω1 (M ) × . . . × Ω1 (M ) × X(M ) × . . . × X(M ) → C ∞ (M )
| {z } | {z }
k copies l copies

is induced by a smooth (k, l)-tensor field if and only if is multilinear over C ∞ (M ).

The idea for why this result is true can be seen by considering for F ∈ T (k,l) (M ), ω i ∈
Ω1 (M ), Xj ∈ X(M ) the real-valued (smooth) function F, pointwise defined by
F(ω 1 , . . . , ω k , X1 , . . . , Xl )(p) := Fp (ω 1 |p , . . . , ω k |p , X1 |p , . . . , Xl |p )
(how to establish smoothness is explained in [36, Ex. B.5]).
Analogous to the tensor case (see Exercise A.10), given a smooth local frame (Ei ) with
dual coframe (εi ), it follows that the tensor fields Ei1 ⊗ . . . Eik ⊗ εj1 ⊗ . . . εjl form a smooth
local frame for T (k,l) (M ). In particular, for local coordinates (xi ) on U , a (k, l)-tensor field
F has the local expression
F = Fji11...j
...ik
l
∂i1 ⊗ . . . ⊗ ∂ik ⊗ dxj1 ⊗ . . . ⊗ dxjl ,
with coefficients Fji11...j
...ik
l
∈ C ∞ (U ).
Just like vectors one also pull back tensor fields in the natural way.
Definition A.18. Suppose M and N are manifolds, F : M → N is a smooth map and
A ∈ T (0,l) (N ). For every p ∈ M the pointwise pullback of A by F is the tensor dFp∗ (A) ∈
T l (Tp∗ M ), defined by
dFp∗ (A)(v1 , . . . , vl ) := A(dFp (v1 ), . . . , dFp (vl )).
The pullback of A by F is the tensor field F ∗ A ∈ T (0,l) (M ) defined by
(F ∗ A)p := dFp∗ (AF (p) ).
 APPENDIX B

Exterior derivatives

For more details see Lee [35, p. 280–284, 362–372].

B.1. Differential of a function

Let M be a smooth manifold. For every smooth real-valued function f ∈ C ∞ (M ) we
define the differential of f , piontwise by
dfp (v) := vf, v ∈ Tp M.
It is clear that df is a smooth 1-form, because dfp it depends linearly on v ∈ Tp M for each
p ∈ M and because df (X) = Xf is smooth for every smooth X ∈ X(M ).
Example B.1. The coordinate 1-forms dxj associated with the coordinate functions
xj : U → R on an open set U ⊆ M .
In coordinates, we can therefore write
∂f i
df = dx .
∂xi

Exercise B.2. For f (x, y) = x2 y cos x on R2 compute df in terms of dx and dy.

The differential satisfies the usual rules of differentiation (linearity, product rule, chain
rule etc.), see [35, Prop. 11.20]. In particular, df = 0 if and only if f is constant.

B.2. Exterior derivative of a form

On forms one can define a natural generalization of the differential on functions, called
the exterior derivative.
Every f ∈ C ∞ (M ) defines a 1-form df , but in order for a 1-form ω to satisfy ω = df it
needs to be closed [35, Prop. 11.44], i.e., in every coordinate chart (but independent of the
choice coordinates)
∂ωj ∂ωi
− j = 0. (B.1)
∂xi ∂x
Note that (B.1) is antisymmetric in i and j, and thus can be interpreted as components of
an alternating tensor field, i.e., a 2-form, locally given by
X  ∂ωj ∂ωi

dω = − dxi ∧ dxj .
∂xi ∂xj
i<j

In other words, ω is closed if and only if dω = 0 in all charts.

In fact, dω can be defined globally, and for differential forms of any degree. One can show
the following result [35, Thm. 14.24].
127
128 B. EXTERIOR DERIVATIVES

Theorem B.3 (Existence and uniqueness of exterior differentation). Let M be a smooth

manifold. For all k, there are unique operators
d : Ωk (M ) → Ωk+1 (M ),
called exterior differentation, satisfying
(i) d is linear over R.
(ii) If ω ∈ Ωk (M ) and η ∈ Ωl (M ), then
d(ω ∧ η) = dω ∧ η + (−1)k ω ∧ dη.
(iii) d ◦ d ≡ 0.
(iv) For f ∈ Ω0 (M ) = C ∞ (M ), df is the differential of f , given by df (X) = Xf .
In any smooth i
P′ coordinate Pchart (x ), d is given as follows: For a smooth k-form locally
given by ω = J ωJ dx = {J:j1 <...<jk } ωJ dxJ (the primed summation indicates that the
J

multi-indices are increasing) we have

!
X X
′ J ′
d ωJ dx = dωJ ∧ dxJ ,
J J
where dωJ is the differential of the function ωJ (see [35, p. 363] for more details).
Problem B.4. Show that for any ω ∈ Ω1 (M ) and X, Y ∈ X(M ) we have
dω(X, Y ) = X(ω(Y )) − Y (ω(X)) − ω([X, Y ]),
where [, ] is the Lie bracket.
One can furthermore show that the pullback commutes with d [35, Prop. 14.26], i.e., for
F : M → N , F ∗ : Ωk (N ) → Ωk (M ) for any k, and all ω ∈ Ωk (M )
F ∗ (dω) = d(F ∗ ω).
Also the Lie derivative commutes with d [35, Cor. 14.36], i.e., for V ∈ X(M ) and ω ∈ Ωk (M ),
LV (dω) = d(LV ω).
The latter follows from Cartan’s Magic Formula [35, Thm. 14.35], which states that for any
V ∈ X(M ), ω ∈ Ω1 (M )
LV ω = V ⌟(dω) + d(V ⌟ ω).
Using differential forms and the exterior derivative one can generalize the fundamental
theorem of calculus to manifolds in a very elegant way.
Theorem B.5 (Stokes’ Theorem). Let M be an oriented smooth n-manifold with bound-
ary, and let ω be a compactly supported (n − 1)-form on M . Then
ˆ ˆ
dω = ω.
M ∂M
For the proof see [35, p. 411–415].
 Bibliography

[1] Stephanie Alexander, Vitali Kapovitch, and Anton Petrunin, Alexandrov geometry: foundations, arXiv,
2019. https://arxiv.org/abs/1903.08539.
[2] Brian Allen and Annegret Burtscher, Properties of the null distance and spacetime convergence, Int. Math.
Res. Not. IMRN 2022, no. 10, 7729–7808.
[3] Lars Andersson, The global existence problem in general relativity, The Einstein equations and the large
scale behavior of gravitational fields, Birkhäuser, Basel, 2004, pp. 71–120.
[4] Lars Andersson, Thierry Barbot, François Béguin, and Abdelghani Zeghib, Cosmological time versus CMC
time in spacetimes of constant curvature, Asian J. Math. 16 (2012), no. 1, 37–87.
[5] C. J. Atkin, The Hopf-Rinow theorem is false in infinite dimensions, Bull. London Math. Soc. 7 (1975),
no. 3, 261–266.
[6] John K. Beem, Paul E. Ehrlich, and Kevin L. Easley, Global Lorentzian geometry, 2nd ed., Monographs
and Textbooks in Pure and Applied Mathematics, vol. 202, Marcel Dekker, Inc., New York, 1996.
[7] Marcel Berger, A panoramic view of Riemannian geometry, Springer-Verlag, Berlin, 2003.
[8] Arthur L. Besse, Einstein manifolds, Classics in Mathematics, Springer-Verlag, Berlin, 2008. Reprint of
the 1987 edition.
[9] Christine Breiner, Nikolaos Kapouleas, and Stephen Kleene, Conservation laws and gluing constructions
for constant mean curvature (hyper)surfaces, Notices Amer. Math. Soc. 69 (2022), no. 5, 762–773.
[10] H. W. Brinkmann, Einstein spaces which are mapped conformally on each other, Math. Ann. 94 (1925),
no. 1, 119–145.
[11] Dmitri Burago, Yuri Burago, and Sergei Ivanov, A course in metric geometry, Graduate Studies in Math-
ematics, vol. 33, American Mathematical Society, Providence, RI, 2001.
[12] Annegret Burtscher, Marcel Frehner, and Bernhard Grasemann, Tectonic geomorphological investigations
of antiforms using differential geometry: Permam anticline, northern Iraq, AAPG Bulletin 96 (2012),
no. 2, 301–314.
[13] Annegret Y. Burtscher, Length structures on manifolds with continuous Riemannian metrics, New York
J. Math. 21 (2015), 273–296.
[14] Annegret Burtscher, Optimization in Geometry and Physics, 2021. Lecture notes, https://www.math.ru.
nl/~burtscher/lecturenotes/2021OGPnotes.pdf.
[15] Annegret Burtscher and Leonardo Garcı́a-Heveling, Global hyperbolicity through the eyes of the null dis-
tance, arXiv, 2022. https://doi.org/10.48550/arxiv.2209.15610.
[16] Annegret Burtscher and Gert Heckman, Variations of Weyl’s tube formula, J. Geom. Anal. 31 (2021),
no. 12, 11952–11970.
[17] Nina Byers, E. Noether’s discovery of the deep connection between symmetries and conservation laws, The
heritage of Emmy Noether (Ramat-Gan, 1996), Israel Math. Conf. Proc., vol. 12, Bar-Ilan Univ., Ramat
Gan, 1999, pp. 67–81.
[18] Jeff Cheeger and David G. Ebin, Comparison theorems in Riemannian geometry, AMS Chelsea Publishing,
Providence, RI, 2008. Revised reprint of the 1975 original.
[19] Shiing-shen Chern, An elementary proof of the existence of isothermal parameters on a surface, Proc.
Amer. Math. Soc. 6 (1955), 771–782.
[20] Yvonne Choquet-Bruhat and Robert Geroch, Global aspects of the Cauchy problem in general relativity,
Comm. Math. Phys. 14 (1969).
[21] Alan A. Coley, Mathematical general relativity, Gen. Relativity Gravitation 51 (2019), no. 6, Paper No.
78, 37.
[22] Camillo De Lellis and Peter M. Topping, Almost-Schur lemma, Calc. Var. Partial Differential Equations
43 (2012), no. 3-4, 347–354.

129
130 BIBLIOGRAPHY

[23] Manfredo Perdigão do Carmo, Riemannian geometry, Mathematics: Theory & Applications, Birkhäuser
Boston, Inc., Boston, MA, 1992.
[24] Y. Fourès-Bruhat, Théorème d’existence pour certains systèmes d’équations aux dérivées partielles non
linéaires, Acta Math. 88 (1952), 141–225 (French).
[25] Alfred Gray, Tubes, 2nd ed., Progress in Mathematics, vol. 221, Birkhäuser Verlag, Basel, 2004.
[26] A. Gray and L. Vanhecke, Riemannian geometry as determined by the volumes of small geodesic balls,
Acta Math. 142 (1979), no. 3-4, 157–198.
[27] Misha Gromov, Metric structures for Riemannian and non-Riemannian spaces, Progress in Mathematics,
vol. 152, Birkhäuser Boston, Inc., Boston, MA, 1999.
[28] Gert Heckman, Classical Differential Geometry, 2019. Lecture notes, https://www.math.ru.nl/
~heckman/DiffGeom.pdf.
[29] , Introduction to Riemannian Geometry, 2019. Lecture notes, https://www.math.ru.nl/~heckman/
DiffGeom.pdf.
[30] H. Hopf and W. Rinow, Ueber den Begriff der vollständigen differentialgeometrischen Fläche, Comment.
Math. Helv. 3 (1931), no. 1, 209–225 (German).
[31] Harold Hotelling, Tubes and Spheres in n-Spaces, and a Class of Statistical Problems, Amer. J. Math. 61
(1939), no. 2, 440–460.
[32] Jürgen Jost, Riemannian geometry and geometric analysis, 7th ed., Universitext, Springer, Cham, 2017.
[33] Edward Kasner, Einstein’s Theory of Gravitation: Determination of the Field by Light Signals, Amer. J.
Math. 43 (1921), no. 1, 20–28.
[34] Michael Kunzinger and Clemens Sämann, Lorentzian length spaces, Ann. Global Anal. Geom. 54 (2018),
no. 3, 399–447.
[35] John M. Lee, Introduction to smooth manifolds, 2nd ed., Graduate Texts in Mathematics, vol. 218,
Springer, New York, 2013.
[36] , Introduction to Riemannian manifolds, Graduate Texts in Mathematics, vol. 176, Springer, Cham,
2018. Second edition.
[37] Lizhen Ji, Athanase Papadopoulos, and Sumio Yamada (eds.), From Riemann to differential geometry
and relativity, Springer, Cham, 2017.
[38] S. B. Myers and N. E. Steenrod, The group of isometries of a Riemannian manifold, Ann. of Math. (2)
40 (1939), no. 2, 400–416.
[39] John Milnor, Curvatures of left invariant metrics on Lie groups, Advances in Math. 21 (1976), no. 3,
293–329.
[40] John Nash, C 1 isometric imbeddings, Ann. of Math. (2) 60 (1954), 383–396.
[41] , The imbedding problem for Riemannian manifolds, Ann. of Math. (2) 63 (1956), 20–63.
[42] E. Noether, Invariante Variationsprobleme, Nachrichten von der Gesellschaft der Wissenschaften zu
Göttingen, Mathematisch-Physikalische Klasse 1918 (1918), 235-257 (German).
[43] Barrett O’Neill, Semi-Riemannian geometry, Pure and Applied Mathematics, vol. 103, Academic Press,
Inc., 1983. With applications to relativity.
[44] Roger Penrose, Any space-time has a plane wave as a limit, Differential geometry and relativity, Reidel,
Dordrecht, 1976, pp. 271–275. Mathematical Phys. and Appl. Math., Vol. 3.
[45] Peter Petersen, Riemannian geometry, 3rd ed., Graduate Texts in Mathematics, vol. 171, Springer, Cham,
2016.
[46] J. Read and N. Teh (eds.), The Philosophy and Physics of Noether’s Theorems: A Centenary Volume,
Cambridge University Press, Cambridge, 2022.
[47] Bernhard Riemann, On the hypotheses which lie at the bases of geometry, Classic Texts in the Sciences,
Birkhäuser/Springer, Cham, 2016. Edited and with commentary by Jürgen Jost; Expanded English trans-
lation of the German original.
[48] B. Riemann, Über die Hypothesen, welche der Geometrie zugrunde liegen, Springer, Berlin, 1919 (German).
Edited and with commentary by H. Weyl.
[49] Hans Ringström, The Cauchy problem in general relativity, ESI Lectures in Mathematics and Physics,
European Mathematical Society (EMS), Zürich, 2009.
[50] A. Sakovich and C. Sormani, The null distance encodes causality, Journal of Mathematical Physics 64
(2023), no. 1, 012502.
[51] Richard Schoen and Shing Tung Yau, On the proof of the positive mass conjecture in general relativity,
Comm. Math. Phys. 65 (1979), no. 1, 45–76.
 BIBLIOGRAPHY 131

[52] , The energy and the linear momentum of space-times in general relativity, Comm. Math. Phys.
79 (1981), no. 1, 47–51.
[53] Christina Sormani and Carlos Vega, Null distance on a spacetime, Classical Quantum Gravity 33 (2016),
no. 8, 085001, 29.
[54] Michael Spivak, A comprehensive introduction to differential geometry. Vol. I-V, 2nd ed., Publish or
Perish, Inc., Wilmington, Del., 1979.
[55] Hermann Weyl, Mathematische Analyse des Raumproblems, Wissenschaftliche Buchgesellschaft, Darm-
stadt, 1977 (German). Vorlesungen gehalten in Barcelona und Madrid; Reprinting of the 1923 original.
[56] , On the Volume of Tubes, Amer. J. Math. 61 (1939), no. 2, 461–472.
[57] Edward Witten, A new proof of the positive energy theorem, Comm. Math. Phys. 80 (1981), no. 3, 381–402.
[58] Cédric Villani, Optimal transport, Grundlehren der mathematischen Wissenschaften, vol. 338, Springer-
Verlag, Berlin, 2009. Old and new.
[59] Philippe Zaouati, Perelman’s refusal: a novel, American Mathematical Society, Providence, RI, 2021.
 Index

act transitively, 25 admissible, 20

affine connection, 31 arc-length, 21
existence, 34 length, 21
symmetric, 39 piecewise regular, 20
angle, 5 regular, 20
anti-de Sitter space, 23 reparametrization, 21
arc-length, 21 segment, 20
smooth, 20
bilinear, 5 speed, 21
unit-speed, 21
Cartan’s Magic Formula, 50
Christoffel symbols, 41 density, 18
classification theorem, 1 derivative
complete covariant, see covariant derivative
metrically, 23 diameter, 23
conformal, 28 differential
equivalence, 28 of a function, 49
flatness, 28 differential form, 45
manifold, 28 distance, 21
transformation, 28 divergence, 18
connection theorem, 19
affine, see affine connection dot product, 5
coefficients, 32, 37
Einstein
Euclidean, see Euclidean, connection
summation convention, 10
flat, 35
embedding, 7, 11
in vector bundle, 35
isometric, 7, 11
Levi-Civita, see Levi-Civita connection
Euclidean
pullback, 42
connection, 33, 41
tangential, see tangential, connection
group, 24
contraction, 16
metric, 5, 10
convex
exterior
combination, 14
derivative, 35
cosmological
exterior derivative, 50
principle, 27
covariant first fundamental form, 12
Hessian, 39 FLRW spacetimes, 27
covariant derivative, 32 frame, 10
of section, 35 frame-homogeneous, 27
total, 37
curvature geometrization conjecture, 26
bound, 20 gradient, 15
curve, 20 Gram–Schmidt algorithm, 6

133
134 INDEX

graph conformally flat, 28

coordinates, 13 Lorentz group, 24
of a function, 13 orthochronous, 25
parametrization, 13 Lorentzian
manifold, 9, 23
Hessian metric, 9
covariant, 39 vector space, 9
homogeneous Riemannian manifold, 25
Hopf–Rinow Theorem, 23 Möbius strip, 44
hyperbolic manifold
metric, 12 topology, 23
plane, 12 metric
space, 11 connection, 38
hyperboloidal model, 11 equivalence, 15
equivalence on compact sets, 23
immersion, 11
geometry, 20
isometric, 11
parallel, 38
indefinite, 8
topology, 23
index
Minkowski space, 9, 11, 23, 24
lowering, 15
musical isomorphism, 15, 40
of a semi-Riemannian manifold, 9
Myers–Steenrod Theorem, 24
of a symmetric bilinear form, 8
raising, 15
Nash Embedding Theorem, 7
inner product, 5
nondegenerate, 8
on tensors, 16
norm, 5
space, 5
normal
integral, 17
bundle, 33
intrinsic, 31
space, 33
isometry, 7
group, 24
orthogonal, 5, 7
linear, 6, 7
complement, 5
metric, 22
orthonormal, 6
isotropic, 26
basis, 6
at a point, 26
frame, 10
isotropy
representation, 26
parallel, 38
subgroup, 26
parallelism, 31
Koszul formula, 40 parametrization, 12
by arc length, 21
Laplace–Beltrami operator, 19 partition
length of an interval, 20
of tangent vector, 7 Penrose limit, 29
of vector, 5 Poincaré
Levi-Civita connection, 40 conjecture, 26
Lie group, 24
algebra polarization identity, 5
orthogonal, 25 positive
bracket, 39 definite, 5
derivative, 39, 50 semidefinite, 8
group, 25, 25 product bundle, 44
line bundle, 43 projection
Liouville Theorem, 29 normal, 33
local isometry, 7 tangential, 33
local-to-global theorem, 1 pullback connection, 42
locally pushforward, 26
 INDEX 135

regular domain, 17 metric, 23

regular level set, 15 torsion tensor, 32, 39
Ricci torsion-free, see symmetric, connection
flow, 26 torus, 14
Riemann trace, 16, 36
surface, 28 triangle inequality, 23
Riemannian trivial bundle, 43
density, 18
distance, 21 unit sphere, 11
manifold, 6 vector
homogeneous, see homogeneous bundle, 43
with boundary, 7 section, 45
metric, 6 trivial, 43
volume form, 16 volume, 17
round metric, 11 form, 16, 28

scalar warped product, 14

product, 8 Weyl tensor, 29
Schouten tensor, 29 Weyl–Schouten Theorem, 29
semi-Riemannian Whitney Embedding Theorem, 7
manifold, 9
metric, 9
semidirect product, 24
skew-symmetric, 25
spacetime, 9
stabilizer, see isotropy subgroup
stereographic projection, 28
Stokes’ Theorem, 50
submersion, 14
Riemannian, 14
surface
of revolution, 13
symmetric, 5
bilinear form, 8
index, 8
connection, 39, 41
distance, 23
space, 27

tangential
connection, 34, 41
derivative, 31
tensor, 45
bundle, 46
characterization lemma, 47
contraction, 46
field, 46
product, 45
pullback, 47
rank of, 45
symmetric, 46
trace, 46
tensor field, 6
topology
manifold, 23