Hu Zhang Zhao

Lectures on Riemannian Geometry
Shiping Liu
e-mail: [email protected]
Department of Mathematics, USTC
typed by Yirong Hu, Shiyu Zhang, Qizhi Zhao.

Contents
Introduction 3
1 Riemannian Metric 5
1.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 When are two Riemannian manifolds “equivalent”? . . . . . . . . . . 8
1.4 Existencce of Riemannian Metrics . . . . . . . . . . . . . . . . . . . 8
1.5 The Metric Structure . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.6 Riemannian Measure, Volume . . . . . . . . . . . . . . . . . . . . . 14
1.7 Divergence Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2 Geodesics 23
2.1 Geodesic equations and Christoffel symbols . . . . . . . . . . . . . . 23
2.2 Minimizing Properties of Geodesics . . . . . . . . . . . . . . . . . . 28
2.3 Global Properties:Hopf-Rinow Theorem . . . . . . . . . . . . . . . . 37
2.4 Existence of geodesics in given homotopy class . . . . . . . . . . . . 43
3 Connections, Parallelism, and covariant Derivatives. 49

3.1 Affine Connections. . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.2 Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.3 Covariant derivatives of a tensor field . . . . . . . . . . . . . . . . . 57
3.4 Levi-Civita Riemannian Connections . . . . . . . . . . . . . . . . . . 59
3.5 The First variation of Arc Length and Energy . . . . . . . . . . . . . 66
3.6 Covariant differentiation, Hessian, and Laplacian . . . . . . . . . . . 69
3.7 Appendix: The technical lemma. . . . . . . . . . . . . . . . . . . . . 71
4 Curvatures 73
4.1 The Second Variation . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.2 Properties of Curvature tensor : Geometric meaning and Symmetries . 77
4.2.1 Ricci Identity . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.2.2 Geometric meaning :A test case [Spivak 2.Chap6.Thm 10] . . 78
4.2.3 Bianchi Identities . . . . . . . . . . . . . . . . . . . . . . . . 80
4.2.4 Riemannian curvature tensor . . . . . . . . . . . . . . . . . . 82
4.3 Sectional Curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3
CONTENTS 1
4.4 Ricii Curvature and Scalar curvature . . . . . . . . . . . . . . . . . 89

4.5 The Second Variation : Revisited [JJ,4.1] [WSY,chap 6] . . . . . . . . 95
5 Space forms and Jacobi fields 113

5.1 Isometries and totally geodesic submanifold . . . . . . . . . . . . . . 113
5.2 Space forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
5.3 Geodesics in Rn , S n , Hn . . . . . . . . . . . . . . . . . . . . . . . . 120
5.4 What is a Jacobi field? . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.5 Conjugate Points and ,Minimizing Geodesics . . . . . . . . . . . . . 128
5.6 Index forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
5.7 The proof of Morse index theorem . . . . . . . . . . . . . . . . . . . 139
5.8 Cartan-Hadamard Theorem . . . . . . . . . . . . . . . . . . . . . . . 142
5.9 Uniqueness of simply-connected space forms . . . . . . . . . . . . . 147
5.10 Convexity: Another application of Cartan-Hadamard Theorem . . . . 151
6 Comparison Theorem 157

6.1 Sturm Comparison Theorem . . . . . . . . . . . . . . . . . . . . . . 157
6.2 Morse-Schoenberg Comparison and Rauch Comparison Theorems . . 160
6.3 Cut Point and Cut Locus . . . . . . . . . . . . . . . . . . . . . . . . 170
6.4 Hessian Comparison Theorem . . . . . . . . . . . . . . . . . . . . . 176
6.5 Laplacian Comparison Theorem . . . . . . . . . . . . . . . . . . . . 180
6.6 Comments on Injectivity Radius Estimate and Sphere Theorems . . . 183
6.7 Volume Coparison Theorems . . . . . . . . . . . . . . . . . . . . . . 184
6.7.1 Maximal Diameter Theorem . . . . . . . . . . . . . . . . . . 192
6.7.2 Volume Growth Rate Estimate . . . . . . . . . . . . . . . . . 192
7 Candidates for Synthetic Curvature Conditions 195

7.1 Nonpositive Sectional Curvature and Convexity . . . . . . . . . . . . 195
7.2 Bochner technique and Bakry-Émery Γ-calculus . . . . . . . . . . . . 199
7.2.1 A computation trick . . . . . . . . . . . . . . . . . . . . . . 199
7.2.2 The Hodge Laplacian . . . . . . . . . . . . . . . . . . . . . . 200
7.2.3 Weitzenböck formula . . . . . . . . . . . . . . . . . . . . . . 206
7.2.4 Bakry-Émery Γ-calculus:A systematic way of understanding
the Bochner formula . . . . . . . . . . . . . . . . . . . . . . 211
2 CONTENTS
Introduction
On June 10, 1854, Georg Friedrich Bernhard Riemann delivered a lecture entitled “über
die Hypothesen, welche der Greometrie zu liegen”(On the Hypotheses which lie at the
Foundations of Geometry) to the faculty of Göttingen University. This lecture was
published later in 1866, and gives birth to Riemannian geometry.
This lecture was given by Riemann as a probationrary inaugural lecture for seeking
the position of “Privatdocent”. Privatdocent is a position in the German university
system. It is a lecturer who receive no salary, but is merely forwarded fees paid by those
students who elected to attend his lectures. To attain such a position, one has to submit
an inaugural paper(Habilitationsschrift) and give a probationary inaugural lecture on a
topic chosen by the faculty, from a list of 3 proposed by the coordidates. The first 2
topics which Riemann submitted were ones on which he has already worked; The 3rd
topic ha chose was the foundations of geometry. Usually, the faculty chooses the first
topic proposed by the candidate. However, contrary to all traditions, Gauss passed over
the first two and chose instead the 3rd of Riemann’s topics. So Riemann has to prepare
a lecture on a topic that he had not worked on before. In the end, Riemann finished his
lecture in about seven more weeks.
Why did Gauss choose the 3rd topic? In fact, that is topic in which Gauss had been
interested in for many years.
The single most important work in the history of differential geometry is Gauss’
paper, in Latin, of 1827: “Disquisitiones generales circa super ficies curvas”(General
Investigations of Curved Surfaces). The most influential result in Gauss’ paper is the
so-called ”Theorem Egregium”. Roughly speaking, this theorem asserts that the Gauss
cvrvature of a surface is determined by its first fundamental form. This opens the door
to “Intrinsic geometry” and provides possibility of studying more abstract spaces other
than surfaces in E 3 . For example, one can study the geometry of a “flat torus”, which
is a topological torus associated with a “flat metric”.
What Riemann did in his lecture is developing higher-dim intrinsic geometry. Ge-
ometry presupposes the concept of space. In this course of Riemannian geometry,
the space we study is a C ∞ -manifold M(Hausdorff and second countable) associated
a Riemannian metric g. A vital step is to understand what is the extension of Gauss
curvature in higher dimensional manifolds. The original definition of Gauss curvature
using Gauss map is not available in higher dimensional manifolds. The expression
of Gauss curvature in the Gauss equation is possibly extended to higher dimensional
spaces.
Actually, before discussing “curvature”, we can already see a lot of information of
3
4 Introduction
the geometry of the underlying spaces from the Riemannian metric.

We will roughly follow the scheme below:
(I). Riemannian metric
From the Riemannian metric, we can calculate the length of curves, and moreover,
we obtain a natural volume measure.
(II). Geodesics
It’s natural to look for shortest curves connecting two points. Geodesics are curves
which are “locally” shortest. They can be obtained from the First variation of the
length functional. In order to explore the problem whether a geodesic is shortest, we
will discuss exponential maps, normal coordinates, Hopf-Rinow theorem(1931), etc.
(III).Connections, parallelism, and Covariant Derivatives
We reinterpret the geodesic equations in terms of connections and covariant deriva-
tives, or parallelism. In this process, we develop (Abstract) calculus on Riemmanian
manifolds.
(IV).Curvature
When we consider the Second Variation of the length functional. In this variational
formula, a “curvature” term will appear, which is a generalization of Gaussian curva-
ture of surfaces. We will discuss properties of Riemannian curvature tensor and various
curvature notions.
(V).Spaces forms and Jacobi fields
We will discuss the complete Riemannian manifolds with constant curvature, which
are referred to as space forms. Those will be model spaces when we study the geometry
of a general Riemannian manifold. In this process, we discuss the theory of Jacobi
fieldsvariational vector fields of family of geodesics.
(V I).Comparison theorems
We explore geometry of Riemannian manifolds with curvature bounds via compar-
ing them with spaces forms”.
Chapter 1
Riemannian Metric
1.1 Definition
Recall in the theory of surfaces in E 3 : For a surface S ⊂ E 3 , ∀ p ∈ S , and any two
tangent vectors X,Y∈ T p S , we have the inner product hX, Yi p . (hX, Yi p is the inner
product of E 3 )
This inner product hX, Yi p corresponds to the first fundenmental forms of S at p.

Based on this inner product, one can compute the lengths of a curve in S, the area of a
domain in S, etc.
Now, let us consider a C ∞ -manifold M n (dimM = n).
Definition 1.1 (Riemannian Metric). A Riemannian metric g on M is a “C ∞ assign-
ment”: For each tangent vector space T p M(p ∈ M) of M, we assign an inner product
g p (·, ·)=h·, ·i p , which is smoothly dependent on p in the following sense: f(p):=hX p , Y p i p =
g p (X p , Y p ) is a smooth function on U ⊂ M for any smooth tangent vector fields X,Y on
U ⊂ M.
Remark 1.1. Recall that by inner product, we mean g p (·, ·) is symmetric, positive
definite and bilinear.
What it looks like in local coordinates: Given p ∈ M. For any coordinate neigh-
borhood U 3 p, let its coordinate functions be x1 , x2 , · · · , xn . Then the tangent vector
5
6 CHAPTER 1. RIEMANNIAN METRIC
space T p M is spanned by ∂x∂1 , · · · , ∂x∂n . Its dual spacethe cotangent vector spaces T p∗ M
is spanned by dx1 , · · · , dxn . Then we denote
∂ ∂ ∂ ∂
h , j i p = g( i , j )(p) := gi j (p),
∂x ∂x
i ∂x ∂x
and for any smooth tangent vector fields
∂ ∂
X = Xi ,Y = Yj j
∂xi ∂x
we have hX p , Y p i p = X i (p)Y j (p) ∂x∂ i , ∂x∂ j

D E
= gi j (p)X i (p)Y j (p) . Here we use the
p
Einstein summation: an index occuring twice in a product is to be summed from 1 up to
the space dimension. Therefore, in local coordinates, we can consider the Riemannian
metric g as
g = gi j dxi ⊗ dx j ,
where (1) “g p (·, ·) depends smoothly on p” is equivalent to say “gi j (p) is smooth on
U 3 p, ∀i, j”.
(2) “g p (·, ·) is an inner product” is equivalent to say
· gi j = g ji , i.e the matrix (gi j (p)) is symmetric at any p ∈ U.
· The matrix (gi j (p)) is positive definite at any p ∈ U.
Hence, we can reformulate the definition of a Riemannian metric g as belows.
Definition 1.2 (Riemannian Metric). A Riemannian metric g on a C ∞ -manifold M is a

smooth (0,2)-tensor satisfying
g(X, Y) = g(Y, X), g(X, X) = 0 and g p (X, X) = 0 ⇔ X(p) = 0
for any smooth tangent vector fields X,Y.
Definition 1.3. A Riemannian manifold is a differentiable(we will always assume C ∞ )

manifold equipped with a Riemannian metric.
Remark 1.2. a couple (M,g).
1.2 Examples
(1)M = Rn , ∀p ∈ M, T p Rn = Rn . The standard inner product on Rn gives a standard
Riemannian metric g0 : X
g0 (X, Y) = X i Y i = X T Y.
i
Another way to see it: R is covered by a single coordinate (x1 , · · · , xn ).

n
Matrix : (gi j ) = (δi j ) = In×n .

T ensor : g = dx1 ⊗ dx1 + · · · + dxn ⊗ dxn .
1.2. EXAMPLES 7
More generally, (gi j ) can be any positive definite and symmetric n × n matrix A = (ai j ).
Then g = ai j dxi ⊗ dx j and g(X, Y) = X T AY.
(2) Induced Metric: Let f : M n → N n+k be a smooth immersion(i.e. d f p :
T p M → T f (p) N is injective for any p ∈ M). Let (N, gN ) be a Riemannian manifold(e.g.
(N, gN ) = (Rn+k , g0 )). We can define the pull-back metric f ∗ gN on M as belows
( f ∗ gN ) p (X p , Y p ) = (gN ) f (p) (d f p (X p ), d f p (Y p )).
One can verify that f ∗ gN is a Riemannian metric on M. (( f ∗ gN ) p (X p , X p ) ⇔ d f p (X p ) =

d f p is in jective
0 ⇐⇒ X p = 0)
Definition 1.4. We call f ∗ gN an induced metric on M with respect to the smooth im-
mersion f : M n → N n+k .
A special case: M ⊂ N is an immersed submanifold. Then the inclusion map
i : M → N is an immersion. In this case, the induced metric (i∗ g) p is just the restriction
of (gN ) p on T p M ⊂ T p N.
Example 1.1. Let M = S 1 be the unit circle in R2 . Choose a coordinate neighborhood
{θ : 0 < θ < 2π}. Then the inclusion map is given by
 x = cos θ


 y = sin θ

gS 1 = (dx ⊗ dx + dy ⊗ dy)|S 1 = dθ ⊗ dθ. Then we have dx = − sin θdθ, dy = cos θdθ

and gS 1 = (dx ⊗ dx + dy ⊗ dy)|S 1 = dθ ⊗ dθ.
Example 1.2. Let M = S 2 be the unit sphere in R3 . Choose a coordinate neighborhood
{(θ, z) : 0 < θ < 2π, −1 < z < 1}
 p



 x = 1 − z2 cos θ

 p
y = 1 − z2 sin θ





z = z


Then the induced metric on S 2 is gS 2 = 1

1−z2
dz ⊗ dz + (1 − z2 )dθ ⊗ dθ. (Exercise!)
(3)Product metric: Let (M, g M ), (N, gN ) be two Riemannian manifolds. Let M × N

be the Cartesian product. Let π1 : M × N → M, π2 : M × N → N be the natural
projections, Then M × N has the following product metric g:
g p,q (X, Y) = (g M ) p (dπ1 (X), dπ1 (Y)) + (gN ) p (dπ2 (X), dπ2 (Y))
∀(p, q) ∈ M × N, ∀X, Y ∈ T (p,q) (M × N). For example, the torus T n = S 1 × · · · × S 1 has
a product metric based on the induced metric on S 1 . Such tori are flat tori.
(4)If g1 , g2 are two Riemannian metrics of M, so does ag1 + bg2 , ∀a, b > 0.
1.3 When are two Riemannian manifolds “equivalent”?

Definition 1.5 (Isometry). Let (M, g M ), (N, gN ) be two Riemannian manifolds. Let
φ : M → N be a diffeomorphism(i.e. φ is bijective, C ∞ and π−1 is also C ∞ ). If
φ∗ gN = g M . (That is, (g M ) p (X, Y) = (gN )φ(p) (dφ p (X), dφ p (Y)), ∀p ∈ M, ∀X, Y ∈ T p M),
then we call φ an isometry.
1.4 Existencce of Riemannian Metrics

Theorem 1.1. A C ∞ -manifold M (Hausdorff,second countable) has a Riemannian met-
ric.
Proof. Let {Uα } be a locally finite covering of M by coordinate neighborhoods.That is
,any point p of has a neighborhood U such that U ∩ Uα , ∅ at most for a finite number
of indices.
Let {φα } be a C ∞ partition of unity on M subordinate to the covering {Uα } That is
(1) φα ≥ 0, φα = 0 on M \Uα .
(2) φα (p) = 1,∀p ∈M.
P
α
On each Uα we can define a Riemannian metric gα (·, ·) induced by the local coor-
dinates.(e.g. (Uα , xαi ) → gα = dxαi ⊗ dxαi )
P
i
Let us set
X
g p (X, Y) := φα (p)(gα ) p (X, Y), ∀p ∈ M, ∀X, Y ∈ T p M. (1.4.1)
α
It is direct to verify that this construction defines a Riemannian metirc on M .(In

fact ,the main point is to check that g is positive definite.This is because the summation
1.5. THE METRIC STRUCTURE 9
φα α = 1 =⇒ ∃β s.t.φβ (p) > 0. Hence g p ≥

P
in (1.4.1) is actually a finite sum,and
α
φβ (p)gβ > 0.

Remark 1.3. Whitney (1936) showed that any n-dimensional C ∞ manifold M n can
be embedded into R2n+1 .Thus M n always has a Riemannian metric induced from the
standard Riemannian metric g0 of R2n+1 .
On the other hand ,given a Riemannian manifold (M n , g M ) ,the Riemannian metric
g M is usually different from the metric induced from g0 of R2n+1 .In fact ,Nash’s em-
bedding theorem tells that for any Riemannian manifold (M n , g M ) ,there exists a N,s.t.
(M n , g M ) can be underlineisometrically embedded into (RN , g0 ). In other words,there
exists an embedding ϕ : M n → RN s.t. g M = ϕ∗ g0 . Nevertheless, the intrisic point
of view in the above proof offers great conceptual and technical advantages over the
approach of submanifold geometry of Euclidean space.
1.5 The Metric Structure

The Riemannian metirc g on M induces a natural distance function d. That is a function
d : M × M → R satisfying for any p, q, r ∈ M
(1)d(p, q) ≥ 0, and d(p, q) = 0 ⇐⇒ p = q.
(2)d(p, q) = d(q, p).
(3)d(p, q) ≤ d(p, r) + d(r, q).
To show this fact,let’s consider the length of curves in M.
Let γ : [a, b] → M be a smooth (parametrized) curve in M. For any t ∈ [a, b], we
have the tangent vector
d
γ̇(t) := dγ( ) ∈ T γ(t) M.
dt
We always assume the parametrization is regular ,i.e. γ̇(t) , 0, ∀t. Then the length
of γ is
Z b
Length(γ) := |γ̇(t)|dt
a
Z bq
= hγ̇(t), γ̇(t)iγ(t) dt.
a
Lemma 1.1. The quantity Length(γ) does not depend on the choice of parametirzation.
Proof. Let γ1 : [c, d] → M (c < d) is another regular parametrization of the same

curve. Then there exist a smooth function t1 = t1 (t) : [a, b] → [c, d] s.t.
γ1 (t1 (t)) = γ(t).
Since both parametrization are regular, we haveγ˙1 (t1 ) , 0, γ0(t)

˙ ,0
Obseve that
dt1
γ̇(t) = γ˙1 (t1 ) · . (1.5.1)
dt
By definition, (1.5.1) can be checked as belows:for any smooh function f ,
d d d
γ̇(t) · f = dγ( ) · f = f (γ(t)) = f (γ1 (t1 (t)))
dt dt dt
d dt1 d dt1
= f (γ1 (t1 )) = dγ1 ( ) f ·
dt1 dt dt1 dt
Hence we have dtdt1 , 0, which means either dtdt1 > 0 or dtdt1 < 0. As we assume
a ≤ b, c ≤ d, we have here dtdt1 > 0.
Now we caculate
Z d q
Length(γ1 ) = hγ˙1 (t), γ̇1 (t)iγ1 (t1 ) dt1
c
Z b r
dt dt dt1
= h γ̇(t), γ̇(t)iγ(t) dt
a dt1 dt1 dt
Z bq
dt dt1
= hγ̇(t), γ̇(t)iγ(t) dt
a dt1 dt
Z bq
= hγ̇(t), γ̇(t)iγ(t) dt = Length(γ).
a

Exercise 1.1. Let ϕ : (M, g M ) → (N, gN ) be an isometry,γ be a smooth curve in M.

Show that Length M (γ) = LengthN (ϕ(γ)).
Arclength parametrization. Now we look for a ”standard” parametrization for a

given curve. Consider smooth curve γ : [a, b] → M. We can define the arclength
function of γ:
Z t q
s(t) := hγ̇(t), γ̇(t)iγ(t) dt
a
Then s = s(t) : [a, b] → [0, Length(γ)] is a strictly increasing function. Denote byt =
t(s) its inverse function. Then we can reparametrize γ(t) as
γ1 (t) = γ(t(s)), s ∈ [0, Length(γ)]
Proposition 1.1. hγ˙1 (s), γ˙1 (s) ≡ 1
Proof. We caculate
dt dt
hγ˙1 (s), γ˙1 (s)iγ1 (s) = hγ̇(t) , γ̇(t) iγ(t(s))
ds ds
dt 2
=( ) · hγ̇(t), γ̇(t)γ(t)
ds
=1

Remark 1.4. The length of a (continuous and) piecewise smooth curve is defined as
the sum of the smooth pieces.
On a Riemannian manifold (M, g),the distance between two points p, q can be de-
fined:
d(p, q) := inf{Length(γ) : γ ∈ C p,q }
where C p,q := {γ : [a, b] → M : γ piecewise smooth curve with γ(a) = p, γ(b) = q}

It can be checked immediately that d : M × M → R satisfies
(1)d(p, p) = 0, d(p, q) ≥ 0
(2)d(p, q) = d(q, p)
(3)d(p, q) ≤ d(p, r) + d(r, q) (By definition)
To show that d is indeed a distance function , it remains to prove d(p, q) = 0 =⇒
p = q or equivalently, p , q =⇒ d(p, q) > 0
Theorem 1.2. (M, d) is a metric space.
Proof. It remains to show p , q =⇒ d(p, q) > 0

There exists a coordinate neighborhood U of q, with coordinate map ϕ such that

 ϕ(q) = 0 ∈ Bδ (0) := V ⊂ Rn



p<U

Denote ϕ−1 : Bδ (0) → U. Then on Bδ (0),we have the pull-back metric (ϕ−1 )∗ g := h
Let γ : [a, b] → Bδ (0) be a curve connecting 0 = ϕ(q) and a point on ∂B 2δ (0). Let c
be the smallest number with γ(c) ∈ ∂B 2δ (0). Then we have
Z c
Lengthh (γ) ≥ Lengthh (γ|[a,c] ) = hγ̇, γ̇ih dt.
p
a
Obeserve that there exists a positive constant s.t.
hγ̇, γ̇ih ≥ hγ̇, γ̇ig0 on Bδ (0)
(Exercise: Let K be an open set of Rn , [gi j (x)] be n2 continuous functions on K

such that the matrix [gi j (x)] is symmetric for any x ∈ K.
(1)Denote by λ(x), Λ(x) the smallest, largest eigenvalues of [gi j (x)]. Show that both
λ and Λ are continuous functions on K.
(2)Suppose K is compact and [gi j (x)] is positive definite for any x ∈ K. Show that
there exist positive constants 1 , 2 such that
X
1 |v|2 ≤ gi j (x)vi v j ≤ 2 |v|2
i, j
for any x ∈ K and any (v1 , · · · , vn ) = v ∈ Rn .)

Therefore, we have
√
δ
Z c
√ √
Lengthh (γ) ≥ hγ̇, γ̇ig0 dt = Lengthg0 (γ|[a,c] ) ≥ .
a 2
Any curve connecting p, q ∈ M must intersect with φ−1 (∂B 2δ (0)) at some point. Hence,
we have √
δ
d(p, q) ≥ > 0.
2

Moreover, we have the following property.

Proposition 1.2. Let (M,g) be a Riemannian manifold. For given p ∈ M, the function
f (·) := d(·, p) : M → R is continuous.
Proof. We need to show f (qi ) → f (q) while qi tends to q (in the sense of the mani-
fold topology). By triangle inequality, | f (qi ) − f (q)| ≤ d(qi , q). So it suffices to prove
d(p, q) → 0 as i → ∞. Pick a coordinate neighborhood U of q such that
∃ φ : U → Bδ (0) =: V ⊂ Rn , φ(q) = 0.
Without loss of generality, we assume φ(qi ) ∈ B 1i (0). Denote by h := (φ−1 )∗ g.

Choose γei : [0, 1] → V be the curve γei (t) = tφ(qi ). (Note γei (0) = φ(q), γei (1) = φ(qi )).
Then ∃2 > 0 such that
√

Z 1p Z 1
√
γi ) =
Lengthh (e hγ̇, γ̇ih dt ≤ 2 hγ̇, γ̇ig0 dt ≤ .
0 0 i
Therefore, we have
√
2
d(pi , qi ) ≤ Lengthh (e
γi ) ≤ → 0, as i → ∞.
i

Corollary 1.1. The topology on M induced by the distance function d coincides with
original manifold topology of M.
Proof. The continuity of f (·) = d(·, p) tells every open set of the topology induced by
d is again open of the manifold topology. By the proof of 1.5.2, one can see the other
way around every open set of the manifold topology is open in the topology induced
by d.
Remark 1.5. It actually suffice to show that the topology induced by d coincides with
the one in Rn in each coordinate neighberhood, which is induced by the Euclidean
distance. We know for any point in this coordinate, exist positive constant 1 , 2 with
1 |v|2 ≤ gi j vi v j ≤ 2 |v|2 , ∀v ∈ Rn .
By our precious argument, this implies the Riemannian distance d and the Euclidean
distance control each other. Hence the two topology coincides.
1.6 Riemannian Measure, Volume

For any p ∈ M, (T p M, g) is an vector space with inner products. Consider an or-
thonormal basis {e1 , · · · , en } of (T p M, g). The volume of the parallelepiped spanned by
e1 , · · · , en , vol(e1 , · · · , en ) = 1.
Now we hope to develop a natural notion of intergration on a Riemannian manifold.

Locally, such an intergration should be the intergration over on Euclidean subset V:
Z
()dx1 · · · dxn .
V
So we need to know the length of the tangent vecters { ∂x∂ i (p), i = 1, 2, · · · , n} and
the volume of the parallelepiped spanned by them.
We can write ∂x∂ i (p) = aij e j .
Then gik (p) = ∂x∂ i (p), ∂x∂ k (p) = aij e j , alk el (p) = aij alk δlj = ali alk .
D E D E P
l
 1
a1 a21 · · · an1 

a12 a22 · · · an2 
Matrix form:[gik ] = AAT , with A =  . .. . . ..  .

 .. .
 1 .2 . 
an an · · · ann
Therefore, we have
∂ ∂
vol( (p), · · · , n (p)) = det(aij )(p)vol(e1 , · · · , en )
∂x1 ∂x
q
= det(gi j )(p).
Volume of “small” compact domain.

∀p ∈ M, let (U, x1 , · · · , xn ) be a coordinate neighberhood: x : U → Rn . Consider
compact set K ⊂ U such that x(K) is measurable(in Rn ). Then we define its volume as
Z q
vol(K) := det(gi j ) ◦ x−1 dx1 · · · dxn (dx1 · · · dxn : Lebesgue measure on Rn ).
x(K)
(1.6.1)
Well-definedness?
Proposition 1.3. The difinition (1.6.1) does not depends on the choice of the coordi-
nate.
1.6. RIEMANNIAN MEASURE, VOLUME 15
Proof. Suppose we have another coordinate neighborhood (V, y1 , · · · , yn ) containing

K: y : V → Rn .
y◦x−1 : x(U) → y(V) is a diffeomorphism. Observe that ∂x∂ i (p) = ∂(y∂x◦xi ) (x(p)) ∂y∂k (p).
k −1
We have
∂ ∂
* +
gi j (p) : =
x
(p), j (p)
∂xi ∂x
∂ ∂ ∂(yk ◦ x−1 ) ∂(yl ◦ x−1 )
* +
= (p), (p) (x(p)) (x(p))
∂yk ∂yl ∂xi ∂x j
∂(yk ◦ x−1 ) ∂(yl ◦ x−1 )
= (x(p)) (x(p))gykl (p).
∂x i ∂x j
Denote the jacobian matrix of the map y ◦ x−1 by

 ∂(y1 ◦x−1 ) ∂(y1 ◦x−1 ) ∂(y1 ◦x−1 ) 
 ∂x1 ∂x2
··· ∂xn 

J(x(p)) =  ... .. ..
 
 .

 n −1 . .
∂(y ◦x ) ∂(yn ◦x−1 ) ∂(yn ◦x−1 )

∂x1 ∂x2
··· ∂xn
We obtain [gixj (p)] = J T (x(p))[gykl (p)]J(x(p)).

q q
Hence det(gixj )(p) = |det(J(x(p)))| det(gykl )(x(p)).
Therefore we have
Z q −1
Z q
y n y=y◦x
det(gi j ) ◦ y dy · · · dy =
−1 1
det(gyij ) ◦ y−1 (y ◦ x−1 )|det(J(x(p)))|dx1 · · · dxn
y(K) x(K)
Z q
= det(gixj ) ◦ x−1 dx1 · · · dxn
x(K)

Volume of “larger” conmpact domain.

Now let us consider the case when K can not be contained in a single coordinate
nerghborhood. Let Uα , xα1 , · · · , xαn be a locally finite covering of M by coordinate neigh-
borhoods. Let {φα }α be a C ∞ partition of unity on M subordinate to the covering {Uα }.
Then we define
X Z q
vol(K) := (φα ◦ xα−1 ) det(gixjα ) ◦ xα−1 dxα1 · · · dxαn .
α T
xα (K Uα )
Proposition 1.4. This definition does not depend on the choice of the covering of co-
ordinate neighborhoods and partiton of unity.
Proof. Let {Vβ , y1β , · · · , ynβ }β be another locally finite covering of M by coordinate neigh-
borhood and {ψβ } be a partition of unity of M subordinate to {Vβ }. Then we compute
X Z q
y
(ψβ ◦ y−1
β ) det(yi jβ ) ◦ y−1 1 n
β dyβ · · · dyβ
β T
yβ (K Vβ )
X Z X q
y
= (φα ◦ y−1
β )(ψ β β dyβ · · · dyβ .
det(yi jβ )) ◦ y−1 1 n
β T α
yβ (K Vβ )
We can exchange the order of the two summations, since each is a finite sum.
X Z X q
yβ
(φα ◦ y−1
β )(ψ −1 1 n
β det(yi j )) ◦ yβ dyβ · · · dyβ
β T α
yβ (K Vβ )
X Z X q
y
= (ψβ ◦ y−1
β )(φα det(yi jβ )) ◦ y−1 1 n
β dyβ · · · dyβ
α T β
yβ (K Vβ )
Z q
change o f variables
X X
= (ψβ ◦ xα−1 )(φα det(yixjα )) ◦ xα−1 dxα1 · · · dxαn
α T β
xα (K Uα )
X Z q
= (φα ◦ xα−1 ) det(yixjα ) ◦ xα−1 dxα1 · · · dxαn
α T
xα (K Uα )

Let us denote by C00 (M) the vector space of continuous functions on M with com-
pact support. For any f ∈ C00 (M), we define
Z X Z q
f := (φα f ) ◦ xα−1 det(gixjα ) ◦ xα−1 dxα1 · · · dxαn .
α
M xα (Uα )
From the above discussion, we know this is well-defined. Moreover, since φα ≥ 0,

we know Z
f ≥0⇒ f ≥ 0.
M
1.6. RIEMANNIAN MEASURE, VOLUME 17
R
Therefore, we obtain a positive linear functional k : C00 (M) → R, k( f ) := f.
M
By riesz representation theorem, these exists a unique Borel measure dvol such that
Z Z
k( f ) = f = f dvol
M M
for any f ∈ C00 (M).
Remark 1.6. In each coordinate neighborhood, the integration with respect to dvol
can be considered an the integration with respect to the n-form
q
Ω0 = det(tgi j )dx1 ∧ · · · ∧ dxn .
Notice that when we change the coordinate, we have
dy1 ∧ · · · ∧ dyn = det(J(x(p)))dx1 ∧ · · · ∧ dxn .
That is, Ω0 may change sign when we change from one coordinate to the other one.
Particularly, we can have a globally defined n-form Ω0 when M is orientable. In
this case, Z Z
f dvol = f Ω0 .
M M
(Recall: On an orientable manifold M, let {e1 , · · · , en } be an orthonormal frame fields,

and {ω1 , · · · , ωn } be its dual. Then
Ω0 = ω1 ∧ · · · ∧ ωn .
Once the measure dvol is obtained, the machine of measure theory is intiated. We
define the L p (1 ≤ p ≤ ∞) norm of f ∈ C00 (M) as
Z
1
k f kL p := ( | f | p dvol) p .
M
We can take the completion of C0∞ (M) with respect to L p -norm, the resulting space
is called L p (M).
·In particular for p = 2, we can define inner product:
Z
h f1 , f2 iL2 := f1 f2 dvol, ∀ f1 , f2 ∈ L2 (M).
M
This verifies L2 (M) to be a Hilbert Space.

1.7 Divergence Theorem

Let X be a smooth tangent vector fields of M. The divergence of X is defined as a C ∞
function div(X) on M as below: Let (U, x1 , · · · , xn ) be a coordinate neighborhood, we
have the volume element
q
Ω0 = det(gi j )dx1 ∧ · · · ∧ dxn .
The divergence divX : M → R is a C ∞ function on M such that

(divX)Ω0 = d(i(X)Ω0 )
where i(X) is the interior product with respect to X(i.e. the contraction of a differential
form with the vector field X).
That is, for any vector fields Y1 , · · · , Yn−1 , we have
i(X)Ω0 (Y1, · · · , Yn−1 ) := Ω0 (X, Y1 , · · · , Yn−1 ).
Remark 1.7. (1)When we change coordinates, Ω0 may change sign but this does not
matter for the definition of divX. The global definition of divX does not require the
orientability of M.
(2)Let us consider the expression of divX in local coordinate. Let X = X i ∂x∂ i ,
∂ p
i(X)Ω0 = i(X i ) det(gkl )dx1 ∧ · · · ∧ dxn .
∂xi
Lemma 1.2. i( ∂x∂ i (dx1 ∧ · · · ∧ dxn ) = (−1)i−1 dx1 ∧ · · · ∧ dx
ci ∧ · · · ∧ dxn .
Proof. Let Y1 , · · · , Yn−1 be any (n-1) smooth vector fields. We compute

∂
i( (dx1 ∧ · · · ∧ dxn )(Y1 , · · · , Yn−1 )
∂xi
∂
=dx1 ∧ · · · ∧ dxn ( i , Y1 , · · · , Yn−1 )
∂x
X
σ(1) ∂
= (sgnσ)dx ⊗ · · · ⊗ dxσ(n) ( i , Y1 , · · · , Yn−1 )
σ∈S (n)
∂x
X
= (sgnσ)dxσ(2) ⊗ · · · ⊗ dxσ(n) (Y1 , · · · , Yn−1 )
σ∈S (n)
σ(1)=i

√
Hence i(X)Ω0 = i X det(gkl )(−1)i−1 dx1 ∧ · · · ∧ dxn . Let G = det(gkl ).
P i
p p
Furthermore, we obtain
X X ∂ √
d(i(X)Ω0 ) = (−1)i−1 ( GX i )dxk ∧ dx1 ∧ · · · ∧ dx
ci · · · ∧ dxn
i k
∂x k
X ∂ √
= ( GX i )dx1 ∧ · · · ∧ dxn
i
∂x i
= (divX)Ω0 .
1.7. DIVERGENCE THEOREM 19
√ i
Hence divX = √1 ∂ i ( GX ).
G ∂x
Notice that in particular, when M = Rn , (gi j ) = (δi j ), we have for X = X i ∂x∂ i ,
n n
X ∂ i X ∂X i i
divX = X = X.
i=1
∂xi i=1
∂xi
This reduces to the classical divergence.

(3)By Cartan’s magical formula
LX ω = i(X)dω + div(X)ω. (LX ω : Lie derivative o f di f f erential f orms)
we have LX Ω0 = i(X)dΩ0 + div(X)Ω = div(X)Ω0 .

This tells us that the divergence of a vector field is “infinitesional” changing rate of
the volume element along the vecter field.
Theorem 1.3. [Divergence Theorem] Let X be a smooth vector fields on (M,g).Then
Z
div(X) dvol = 0
M
Proof. Let {Uα } be a locally finite covering of M by coordinate neighborhood, {φα } be

a partition of unity subordinate to {Uα }. Then we have
X
X= φα X
α
Since X has compact support and the summation above is finite, we have
Z X XZ
div( φα X) dvol = div(φα X) dvol.
α α
M Uα
div(φα X) dvol holds for each α. Without loss of generality,

R
So it is enough to show
Uα
we assume the support of X is contained in a coordinate neighborhood (U, x1 , · · · , xn ),
and X = X i ∂x∂ i . By definition, we have
1 ∂ √
Z Z
div(X) dvol = √ (X i G) dvol
G ∂x i
M U
1 ∂ √ √
Z
= (√ (X i G) G) ◦ x−1 dx1 · · · dxn
G ∂x i
x(U)
∂ √
Z
= (X i G ◦ x−1 )dx1 · · · dxn
∂xi
x(U)
=0

Gradient vector fields of a function

Let f ∈ C ∞ (M). The gradient vector field of f, grad f, is defined as a smooth vector
field such that
hgrad f, Yi (= g(grad f, Y)) = Y( f ).
expressions in local coordinates: (U, x1 , · · · , xn ), Y = Y j ∂x∂ j , suppose grad f =
X i ∂x∂ i .
Then by definition
∂f
hgrad f, Yi = gi j X i Y j = Y( f ) = Y k .
∂xk
That is
∂f k ∂f
(gi j X i )Y j = Y , ∀Y ⇒ gi j X i = .
∂x k ∂x j
Recall [gi j ] is a positive definite matrix. Denote by [gi j ] its inverse matrix, i.e.

0, i f i , j
g gk j = δ j = 
ik i

1, i f i = j
Next, we compute
∂f ∂f
gi j X i = ⇒ gk j gi j X i = gk j j
∂x j ∂x
∂f
⇒ X k = δki X i = gk j gi j X i = gk j .
∂x j
∂f ∂
Hence grad f = gk j ∂x j ∂xk .
Remark 1.8. (1)For the case M = Rn , (gi j ) = (δi j ), we have

X ∂f ∂ ∂ ∂
grad f = = ( 1 , · · · , n ).
i
∂x i ∂xi ∂x ∂x
.
(2)“The gradient vector field is vertical to the level set of a function.”
Proposition 1.5. Let f ∈ C ∞ (M), c be a regular value of f. Then the vector field grad f
is vertical to the level set f −1 (c).
Proof. Since c is a regular value of f, f −1 (c) is a submanifold of M. Let X ∈ T f −1 (c) ⊂
T M. We know
X( f ) = 0 on f −1 (c).
Therefore, hgrad f, Xi = X( f ) = 0 on f −1 (c).
(3)We say a one form η is dual to a vector field X if
hX, Yi = η(Y), f or any vector f ield Y.
In particular, we see
hgrad f, Yi = Y( f ) = d f (Y), ∀Y.
That is, d f is dual to grad f .
1.7. DIVERGENCE THEOREM 21
div(grad f ) dvol = 0.
R
Corollary 1.2. Let f ∈ C0∞ (M). Then
M
Definition 1.6. The Laplacian of a smooth function f is ∆ f := div(grad f ).
Remark 1.9. (1)In local coordinate (U, x1 , · · · , xn ), we have

1 ∂ √
∆f = √ ((grad f )i G)
G ∂x i
1 ∂ ij ∂ f √
= √ (g G)
G ∂xi ∂x j
1 ∂ √ ij ∂ f
= √ ( Gg ).
G ∂xi ∂x j
∂2 f
In particular, for the case M = Rn , gi j = δi j , we have ∆ f = .
P
(∂xi )2
i
(2). ∆ f dvol = 0, ∀ ∈ C0∞ (M).
R
M
(3).For any smooth function f and vector field X, we have
div( f X) = f div(X) + hgrad f, XXi .
Proof.
1 ∂ √
div( f X) = √ ( f X i G)
G ∂x i
1 ∂ i√ 1 ∂ √
= √ X G+ f √ (X i G)
G ∂x i
G ∂x i
= X( f ) + f div(X)
= hgrad f, Xi + f divX.

Theorem 1.4. [Green’s formula] Let f,h be two smooth functions,, at least one of which
has ccompact support. Then
Z Z
f ∆h dvol = − hgrad f, grad hi dvol
M M
Z
= (∆ f )h dvol.
M
Proof. Applying div( f X) = f divX + hgrad f, Xi to X = grad h, we have
div( f (grad h)) = f div(grad h) + hgrad f, grad hi = f ∆h + hgrad f, grad hi .
Since f · grad h has compact support, we can apply divergence theorem to derive the
Green’s formula.
Remark 1.10. (1)∆ is called the Laplace-Beltrami operator. On a compact manifold

(M,g), we have
h∆ f, hiL2 = h f, ∆hiL2 , ∀ f, h ∈ C0∞ (M).(i.e.∆ is sel f − ad joint)

Z
h∆ f, f iL2 = − hgrad f, grad f ig = − |grad f |2 dvol ≤ 0.(i.e. − ∆ is positive)
M
(2)Divergence theorem and Green’s formulas can be extended to compact Rieman-

nian manifolds with boundary.
Theorem 1.5. Let M be a compact Riemannian manifolds with C ∞ boundary ∂M. Let
ν be the outward normal vector field on ∂M, X be a smooth vector field on M. Then
Z Z
(divX) dvol M = g(X, ν) dvol∂M .
M ∂M
Remark 1.11. Let (M, g) be a compact submanifold of (N, gN ). Then ∂M has a Rie-
mannian metric induced from gN . Therefore we have a natural dvol∂M . As a corallary,
we have
Z Z Z
f ∆h dvol M = − g hgrad f, grad hi dvol M + g(grad h, ν) f dvol∂M .
M M ∂M
Last lecture: More explanation on the calculation of

X
sgn(σ)dxσ(2) ⊗ · · · ⊗ dxσ(n) (Y1 , · · · , Yn−1 ).
σ∈S (n)
σ(1)=i
Notice that {σ(2), · · · , σ(n)} = {1, 2, · · · ,bi, · · · , n}. So σ(2), · · · , σ(n) is produced
by a permutation τ of {1, 2, · · · ,bi, · · · , n}. Moreover, sgn(σ) = (−1)i−1 sgn(τ).
Hence
X
sgn(σ)dxσ(2) ⊗ · · · ⊗ dxσ(n) (Y1 , · · · , Yn−1 )
σ∈S (n)
σ(1)=i
X
= (−1)i−1 sgn(τ)dxτ(1) ⊗ · · · ⊗ dxτ(i−1) ⊗ dxτ(i+1) ⊗ · · · ⊗ dxτ(n) (Y1 , · · · , Yn−1 )
τ∈S (n−1)
=(−1)i−1 dx1 ∧ · · · ∧ dx
ci ∧ · · · ∧ dxn (Y1 , · · · , Yn−1 ).
Chapter 2
Geodesics
2.1 Geodesic equations and Christoffel symbols

Let γ : (a, b) → (M, g) be a rugular smooth curve(i.e.ṙ(t) , 0, ∀t ∈ (a, b)). Recall its
length is defined as
Z b q
L(γ) := Length(γ) := h˙(γ)(t), ˙(γ)(t)iγ(t) dt.
a
In a local coordinate neignborhood (U, x1 , · · · , xn ). The curve can be written as
(x1 (γ(t)), · · · , xn (γ(t))).
When γ|(a,b) is contained in U, we have γ̇(t) = xi (γ(t)) ∂x∂ i and

Z b q
L(γ) = gi j (x(γ(t))) ẋi (t) ẋ j (t)dt.
a
Example 2.1. Consider the unit sphere S 2 ⊂ R3 .
23
24 CHAPTER 2. GEODESICS
Consider the coordinate neighborhood (φ, θ) : φ ∈ (− pi2 , pi

2 ), θ ∈ (0, 2π). We have
the induced Riemannian metric
g = dφ ⊗ dφ + cos2 φdθ ⊗ dθ.
Consider a smooth curve γ(t), t ∈ (a, b) on S n with the spherical coordinate (φ(t), θ(t)).
Then Z q b
L(γ) = φ̇2 (t) + cos2 φ(t)θ̇2 (t)dt.
a
Observe that Z b Z b
L(γ) ≥ |φ̇(t)|dt ≥ | φ̇(t)dt| = |φ(b) − φ(a)|.
a a
where “=” holds iff θ̇(t) = 0(⇔ θ(t) ≡ const) and φ is monotonic.Therefore, where
γ(a) and γ(b) has the same coordinate θ, the shortest curve connecting them is the
great circle passing through them.
A natural question is then: given two points p, q ∈ M,
(1)does there exist a shortest curve connecting p,q?
(2)if it exists, is it unique.
In order to find the shortest curve, we consider the critical point of the Length
functional, Length(γ). Note that Length(γ) is a bit massy to bundle with since it has a
√
· term as the integrand. In fact, we can consider the Energy functional instead:
1 b 1 b
Z Z
E(γ) := hγ̇(t), γ̇(t)i = gi j (x(γ(t))) ẋi (t) ẋ j (t)dt.
2 a 2 a
(In phisics, E(γ) is usually called “action of γ” where γ is contained as the orbit of a
mass point. In physics, we have the so-called “least action principle”).
In the following, we will explain why we can consider the critical value of E(γ)
instead of that of L(γ).
Lemma 2.1. For each smooth curve γ : (a, b) → M, we have
L2 (γ) ≤ 2(b − a)E(γ)
hγ̇(t), γ̇(t)i =: kγ̇(t)k ≡ const.
p
and “=” holds if and only if
Proof. By Hölder’s inequality,
Z b Z b
1
Z b
1 √
L(γ) = kγ̇(t)k2 dt) 2 =
2
p
kγ̇(t)kdt ≤ ( 1 dt) ( 2 b − a 2E(γ).
a a a
with equality precisely if kγ̇(t)k ≡ const.

2.1. GEODESIC EQUATIONS AND CHRISTOFFEL SYMBOLS 25
Recall the length of a curve does not depend on the choice of the parametrized by
arc length, i.e. kγ̇k = 1, in order to find shortest curves. In this case, we have
b − a = L(γ), and L(γ)2 = 2(b − a)E(γ) ⇒ L(γ) = 2E(γ).
Hence, after parametrizing curves by arc length, it is enough to minimize E(γ).

Moreover, observe that if γ ∈ C p,q is a shortest curve from p to q. γ : (a, b) → M.
Then for any a ≤ a1 ≤ b1 ≤ b, γ is also a shortest curve from γ(a1 ) to γ(b1 ),
otherwise, we can shorten γk(a,b) further.

So we can localize our problem, and consider the case when p, q ∈ U.
Lemma 2.2. The Euler-Lagrange equations for the energy E are
ẍi (t) + Γijk (x(t)) x˙j (t) x˙k (t) = 0, i = 1, 2, · · · , n (2.1.1)
with
1 il
Γijk = g (glk, j + g jl,k − g jk,l ),
2
and
∂
g jl,k = g jl .
∂xk
Definition 2.1. (geodesics)A smooth curve γ : [a, b] → M which satisfies(with ẋi (t) =
d i
dt x (γ(t)))
ẍi (t)) + Γijk (x(t)) x˙j (t) x˙k (t) = 0, i = 1, 2, · · · , n
is called a geodesics.
Remark 2.1. Christoffel is a German mathematician. He studied in Berlin, and worked
in ETH Zūrich, Strasburg. Riemann’s 1854 lecture was only published in 1868. Christof-
fel published in Crelles Journal(Journal für die reine and and ) in 1869 an article
discussing the necessary condition when two quadratic differential forms
X X
F= ωi,k dxi dxk , F 0 = w0i,k dx0 i dx0 k .
i,k i,k
can be transfirmed into eaach other via independent variable changes. It was there be
introduced the “Christoffel symbols”.
Influced by Christoffel’s work, italian mathematician Gregorio Ricci-Curbasto pub-
lished 6 articles during 1883-1888 on Christoffel’s method, and introduced a new cal-
culus: He interpreted Christoffel’s algorithm into “covariant differentiations”. Ricci(1893)
called it “absolute differential calculus”.
Later in 1901, Ricci and his student Levi-Civita published Ricci’s calculus in French
in Klein’s journal(Mathematische Annalen). It is nowed called “tensor analysis”.
Einstein(1914) derives the geodesic equation using Christoffel symbols in his Berlin
lecture.
Levi-Civita(1916/17) realized the geometric meaning of Christoffel symbols: it de-
termines the “parallel transport” of vectors. This pull Christoffel and Ricci’s discus-
sion back to the track of geometry.
Proof. Let us first look at a general functional
Z a
I(x) = f (t, x(t), ẋ(t))dt
b
where x(t) := (x (t), · · · , x (t)).

1 n
Claim: the Euler-Lagrange equation of I(x) is

d ∂f ∂f
− = 0, i = 1, · · · , n.
dt ∂ ẋi ∂xi
proof of claim: Consider y(t) = (y1 (t), · · · , yn (t)) with y(a) = y(b) = 0. Solving
d |=0 I(x + y) = 0, we have
d
∂f ∂f
Z b
0= ( i yi (t) + i yi (t))dt
a ∂x ∂ ẋ
∂f d ∂f i
Z b
= ( i− )y (t)dt
a ∂x dt ∂ ẋi
By the fundamental lemma of calculus of variations, we have
d ∂f ∂f
− , i = 1, · · · , n.
dt ∂ ẋi ∂xi
This is the E-L equation of I.
For our energy functional
Z b
E(γ) = gi j (x(t)) ẋ j (t) ẋk (t)dt,
a
where f (t, x(t), ẋ(t)) = gi j (x(t)) ẋ j (t) ẋk (t).

We have
d
[gik (x(t)) ẋk (t) + g ji (x(t)) ẋ j (t)] − g jk,i (x(t)) ẋ j (t) ẋk (t) = 0, i = 1, 2, · · · , n.
dt
Hence, gik,l ẋl ẋk + gik ẍk + g ji,l ẋl ẋ j + g ji ẍ j − g jk,i ẋ j ẋk = 0, i = 1, 2, · · · , n.
⇒
2gim ẍm + (gik, j + g ji,k − g jk,l ) ẋ j ẋk = 0, l = 1, · · · , n. (2.1.2)
il
Multiply both sides by g and sum up over i, we have
1
ẍl + gil (gik, j + g ji,k − g jk,l ) ẋ j ẋk = 0, l = 1, · · · , n.
2

2.1. GEODESIC EQUATIONS AND CHRISTOFFEL SYMBOLS 27
Remark 2.2. :
1.When we calculated the term Γijk ẋ j ẋ, . pay attention to the fact that
1 il
Γijk ẋ j ẋk = g (g jl,k + gkl, j − glk, j ) ẋ j ẋk
2
1
= gil (g jl,k + gkl, j − glk, j ) ẋ j ẋk .
2
2.As mentioned before, we only need to consider curves parametrized by arc length
when looking for shortest curves. Now, we explain the other aspect: The solution of
the E-L equation(2.1.1) on page 25 i.e. every geodesic, is parametrized proportionally
to arc length.
Explanation:
d
(gi j (x(t)) ẋi (t) ẋ j (t))
dt
=gi j ẍi ẋ j + gi j ẋi ẍ j + gi j,k ẋi ẋ j ẋk
=2gi j ẍi ẋ j + gi j,k ẋi ẋ j ẋk
=2glm ẍm ẋl + gl j,k ẋl ẋ j ẋk (change the indexes)
= − (glk, j + g jl,k − g jk,l ẋl ẋi ẋk + gl j,k ẋl ẋ j ẋk (use (2.1.2))
=(g jk,l − glk, j ) ẋl ẋ j ẋk
=0
Hence h ẋ, ẋi ≡ const.

That is, every geodesic is parametrized proportionally arc length.
3(Curves in TM).We explain another viewpoint about the geodesic. First, any
smooth curve in M gives a cvrve in its tangent bundle TM.
(1)Systems of coordinates. The toral space of TM is the set of pairs (q, v), q ∈ M,
v ∈ T q M. Let (U, x1 , · · · , xn ) be a coordinate neighborhood of M. ∀q ∈ U, any vector
in T q M can be written as
∂
yi i
∂x
Recall locally we have T U=U × Rn . Then (U × Rn , x1 , · · · , xn , y1 , · · · , yn ) is a coor-
dinate neighborhood of (q, v) ∈ T M. Then one can show that we obtain a differentiable
structure for TM.
(2)Let t → γ(t) be a C in f ty curve in M, then it determines curve t → (γ(t), γ̇(t)) ∈
T M. If, moreover, γ is a geodesic in M, then the curve t → (γ(t), γ̇(t)) in terms of
coordinates (x1 , · · · , xn , y1 , · · · , yn )
t → (x1 (t), · · · , xn (t), y1 (t), · · · , yn (t))
satisfies
 ẋ (t) = y (t)
 k k

k = 1, 2, · · · , n.

(2.1.3)
 ẏk (t) + Γk (x(t))yi y j = 0


ij
Local Existence and uniqueness of geodesics

From ODE theory: (See proposition 2.5 in do carmo, chapter 3 and discussions
before that proposition.)
Theorem 2.1. For any p ∈ M, there exists
an open set V ⊂ M, p ∈ V
numbers δ > 0 and > 0
a C ∞ mapping: γ : (−, ) × U → M. U = (q, v) : q ∈ V, v ∈ T q M, kvk < δ.
Remark 2.3. Let’s have a closer look at the relations between the domain (−, ), and
the length of the velocity kvk < δ. Fix q ∈ M, let’s denote γv (t) as the geodesics with
γv (0) = q, γ̇v (0) = v.
Then we can claim γλv (t) = γv (λt).

This is because: in local coordinates, γv (t) is written as (x1 (t), · · · , xn (t)).
Then satisfy
(x (t), · · · , x (t)) = v
 1 n


(2.1.4)
 ẍk (t) + Γk (x(t)) ẋi ẋ j = 0


ij
(x (λt), · · · , x (λt)) = λv

 1 n


⇒ (2.1.5)
˙ i x(λt)
 ẍk (λt) + Γk (x(λt)) x(λt)

ij
˙ j = λ2 ( ẍi + Γkij (x) ẋi ẋ j )|λt = 0
Hence γλv (t) = (x1 (λt), · · · , xn (λt)) = γv (λt).

⇒Lemma: If γ(t, q, v) is defined for t ∈ (−, ) and kvk < δ, then γ(t, q, λv) is
defined for t ∈ (− λ , λ ) and kvk < δ.
Corollary 2.1. Let p ∈ M, v ∈ T p M. Then ∃ > 0 and a unique geodesic γ : [0, ] →
M with γ(0) = p, γ̇(0) = v.
δ
Proof. Assign s = 2kvk , then ksvk < δ. By theorem 2.1, ∃ 0 > 0, and a unique geodesic
γ sv : [0, 0 ] → M with γ sv (0) = p, γ̇ sv (0) = sv. Hence γv (t) = γ sv ( st ) is a geodesic
defined on [0, s0 ]. Hence = s0 , by the uniqueness of Thm 2.1, we show this corol-
lary.
Exercise 2.1. Compute the geodesic equation of S 2 in sphererical coordinates.
Exercise 2.2. What is the transformation behavior of the Christofell symbols? Do they
define a tensor
2.2 Minimizing Properties of Geodesics

Next, we explain that a geodesic is “locally” shortest curve. For that purpose, we first
discuss the important concept Exponential map.
Let (M, g) be a Riemannian manifold, p ∈ M. Roughly speaking, the exponential
map of M at p maps v ∈ T p M, with g p (v, v) = kvk, to a point q on the geodesic
2.2. MINIMIZING PROPERTIES OF GEODESICS 29
γv : [0, b] → M with γ(0) = p, γ̇(0) = v, such that the arc length c

pq = kvk. This means,
we should pick q = γv (1).(Since as a geodesic,kγ̇tk = kγ̇(0)k = kvk.)
Definition 2.2. (Exponential Map) Let (M, g) be a Riemannian manifold, p ∈ M.

Denote V p := {v ∈ T p M : the geodesic γv with γv (0) = p, γ̇v = v is de f ined on [0, 1]}.
exp p : V p → M, v 7→ γv (1) is called the exponential map of M at p.
In the following we use γ p,v to denote the geodesic with γ p,v = p, γ̇ p,v = v.(Oftenly,
p is omitted.)
What does V p look like?
(1)Star-sharped around 0 ∈ T p M.
If v ∈ V p , i.e. γv is defined on [0, 1], then γλv (0 < λ < 1) is defined on [0, λ1 ], and,
in particular, on[0, 1]. Hence v ∈ V p ⇒ λv(0 ≤ λ ≤ 1) ∈ V p .
(2)∀p ∈ M, ∃ 0 s.t. B(0, 0 ) ⊂ V p i.e. ∀ ω ∈ T p M, kωk ≤ 0 , we have γ p,ω is
defined on [0, 1].
By Theorem ??, ∃ , δ > 0, s.t. ∀v ∈ T p M, kvk < δ, γ p,v is defined on [0, ], hence
γ p,v is defined on [0, 1].
⇒ ∀ω ∈ T p M with kωk ≤ kvk < δ, we have γ p,ω is defined on [0, 1]. That is,
ω ∈ Vp.
Example 2.2.
(1) M = Rn , gi j = δi j .
The geodesic equation is ẍi (t) = 0. The geodesics are straight lines parametrized pro-
portionally to arc length. ∀p ∈ Rn , v ∈ T p Rn . exp p (v) = p + v. V p = T p Rn = Rn .
(2) Circle (S 1 , dθ ⊗ dθ).
∀p ∈ M, T p S 1 can be identified with R. Then exp p (v) = eiv , V p = T p S 1 = R for
p = ei0 = 1.(In local coordinates, exp p : v → v.). This is the simplest example explain-
ing why the terminology “exponential map” is used. It actually comes from Lie theory.
(3) Open disc in R2 : D0 = {(x2 + y2 ) ∈ R2 |x2 + y2 = 1} with a Riemannian metric

induced from the canonical Euclidean metric on R2 .
exp0 (v) = 0 + v = v. But V0 , R2 , V0 = D0 (we identify T 0 D0 with R2 ).
Theorem 2.2. The exponential map exp p maps a neighborhood of 0 ∈ T p M diffeomorphically

onto a neighborhood of p ∈ M.
Remark 2.4. Reason for restricting to a neighborhood:

(1)exp p may not be defined on the whole T p M.
(2)even if exp p is defined on the whole T p M, it is not necessarilly a diffeomor-
phism.(Example of (S 1 , dθ ⊗ dθ), exp p is not injective.)
Proof. 0 ∈ T p M, decp p (0) : T 0 (T p M) → T p M. Since T p M is a vector space, we may

identify T 0 (T p M) with T p M.
⇒ dexp p (0) : T p M → T p M. Now we can calculate dexp p (0)(v) for a v ∈
T 0 (T p M) = T p M.
recall: for a c∞ map f : M → N, x 7→ y. One way to calculate d f : T x M → T y N is
the following :
For any v ∈ T x M, consider a curve γ with γ(0) = x, γ̇(0) = v. Then ξ = f (γ) is a
curve in N, and d f (v) = ξ̇(0).
Here, exp p : V p ⊂ T p M → M, 0 7→ p. For v ∈ T 0 (T p M) = T p M, consider γ(t) = tv.
We have
d
dexp p (0)(v) = exp p (tv)|t=0
dt
d d
= γtv (1)|t=0 = γv (t)|t=0 = γ̇v (0) = v.
dt dt
That is dexp p (0) = id|T p M .

In particular, dexp p (0) has maximal range, and by “inverse function theorem”, there
exists a neighborhood of 0 ∈ T p M, which is mapped diffeomorphically onto a neigh-
borhood of p ∈ M.
Example 2.3.
S 2 ⊂ R3 . q is a antipodalpoint of p. exp p is defined on the whole T p S 2 . Let
B(0, π) ⊂ T p (S 2 ) be the open ball around 0 in T p (S 2 ).(with the scalar product given by
the Riemannian metric of S 2 ). exp p : B(0, π) → S 2 \{q}, B(0, π)\B(0, π) = ∂B(0, π) →
{q} diffeomorphically. However, exp p (B(0, 2π)\B(0, π)) = S 2 \{p, q} and exp p (B(0, 2π)\B(0, 2π)) =
{p}.
And we can identify T p M with Rn via Φ : T p M → Rn , v = vi ei ⇒ (v1 , · · · , vn ).

Thm 2,2,1(page 28) tells that ∃ a neighborhood U 3 p, such that exp−1
p map U diffeo-
=
morphically onto a neighborhood of 0 ∈ T p M viaΦ Rn . In particular, p 7→ 0.
Definition 2.3. (Normal coordinates) The local coordinates defined by (U, exp−1
p ) are
called (Riemannian) normal coordinates with center p.
The advantage of such a choice of coordinates is presented in the following result:
Theorem 2.3. In normal coordinates, we have
gi j (0) = δi j (2.2.1)
Γijk (0) = 0. (2.2.2)
for the Riemannian metric, and all i,j,k.
Proof. (2.2.1) follows from the identification Φ of T p M and Rn (Recall g p (ei , e j ) = δi j ).

Next, we show (2.2.2), recall in the local coordinate (U, exp−1p ) = (U, v , · · · , v ),
1 n
the geodesic equation is
v̈i + Γijk (v(t))v̇ j (t)v̇k (t) = 0, i = 1, 2, · · · , n. (2.2.3)
On the other hand, in exp−1 p (U) ⊂ R , the line tv, t ∈ R, v ∈ R is exp p (γtv (1)) =
d d −1
−1
exp p (γv (t)), i.e. is the image of a geodesic in M via the coordinate map.
Remark 2.5. Even if exp p is defined on the whole T p M, it may be not a global diffeo-
morphism. Suppose exp p : B(0, ρ) → exp p (B(0, ρ)) is diffeomorphic, how large can ρ
be?
Here we mention the following concept of injectivity radius.
Definition 2.4. Let M be a Riemannian manifold,p ∈ M. The injectivity radius of p is
i(p) := supρ > 0 : exp p is a di f f eomorphism on B(0, ρ) ⊂ T p M.
The injectivity radius of M is
i(M) := inf i(p).

p∈M
The above example shows that i(S 2 ) = π.
Normal coordinates
T p M has an inner product defined by g. Let e1 , · · · , en (n = dim M) be an orthonor-

mal basis of T p M(w.r.t. the inner of product given by g). Then for each v ∈ T p M, we
can write v = vi ei . Therefore, v(t) = tv satisfies (2.2.3). This implies Γijk (tv)v j vk =
0, i =, · · · , n, ∀v ∈ Rd .
In particular, for t=0
Γijk (0)v j vk = 0, i = 1, 2, · · · , n, ∀v ∈ Rd . (2.2.4)
For any indices l and m, pick v = el + em , we have
Γilm (0) = 0, i = 1, 2, · · · , n.
That is Γijk (0) = 0, ∀i, j, k.

Recall the definition of Γijk (0), we obtain at 0 ∈ Rd : gil (g jl,k + glk, j − g jk,l ) =
0, ∀i, j, k.
⇒ g jl,k + glk, j − g jk,l = 0, ∀ j, k, l.
By a cyclic permutation of the indices: j → k, k → l, l → j, we have gk j,l,k + g jl,k −
gkl, j = 0. Notice that glk, j = gkl, j , g jk,l = gk j,l , we get 2g jl,k = 0 at 0 ∈ Rd .
Remark 2.6. In general, the second derivatives of the metric cannot be made to vanish
at a given point by a suitable choice of local coordinates. The obstruction is given by
the “curvature tensor”.
On Rn we can introduce the standard polar coordinates (r, ϕ1 , · · · , ϕn−1 ) where

ϕ = (ϕ1 , · · · , ϕn−1 ) parametrizes the unit sphere S n−1 .
Then via Φ, we obtain polar coordinate on T p M. We can write the metric g in polar
coordinate:
∂ ∂
grr := g11 = g( , ), grϕ := (g1,l )l=2,··· ,n , gϕϕ = (gkl )k,l=2,··· ,n
∂r ∂r
In particular at 0 ∈ T p M, we have
grr (0) = 1, grϕ (0) = 0. (2.2.5)
(The reason as (2.2.1) in Theorem 2.2.2 on page 29.)
The point is that in this case we can show (2.2.5) holds true not only at 0 ∈ T p M(=
Rn ), but in the whole coordinate neighborhood.
Theorem 2.4. For the polar coordinates, obtained by transforming the Euclidean co-
ordinates of Rd , on which the normal coordinates with centre p are based, into polar
coordinates, we have
1 0 ···0 
 
0 
gi j =  .

gϕϕ (r, ϕ)

 ..

 
0
where gϕϕ (r, ϕ) is the (n − 1) × (n − 1) matrix of the components of the matrix w.r.t
angular variables (ϕ1 , · · · , ϕn−1 ) ∈ S n−1 .
Proof. In this case, tv, v ∈ Rn is transformed to be ϕ ≡ const. That is, ϕ ≡ const are
geodesic when parametrized by arc length in the local coordinates. They are given by
x(t) = (t, ϕ0 ). ϕ0 f ixed.
Geodesic equation gives
Γirr (x(t))t˙t˙ = Γirr (x(t)) = 0, ∀i (ϕ˙0 = 0).
Compare with the situation in (2.2.4) on page 30. Hence at x(t) ∈ T p M = Rn ,
gil (grl,r + glr,r − grr,l ) = 0, ∀i
⇒
2grm,r − grr,m = 0, ∀m. (2.2.6)
For m = r, we have grr,r = 0, combining with grr (0) = 1 ⇒ grr ≡ 1.
Hence grr,m ≡ 0, ∀m ⇒ grϕ,r = 0, ∀ϕ ⇒ grϕ ≡ 0.
Remark 2.7. dr ⊗ dr + gϕϕ (r, ϕ)dϕ ⊗ dϕ is not a proguct metric, since gϕϕ may depend
on r.
Remark 2.8. Since g is positive definite, we have
(gϕϕ (r, ϕ))
is also positive definite.

Corollary 2.2.
(1)For any p ∈ M, ∃ρ > 0 s.t. the (Riemannian) polar coordinates may be introduced
on B(p, ρ) := {q ∈ M, d(ϕ, q) = ρ}.
(2)For any such that ρ, and q ∈ ∂B(p, ρ), there is a unique normal geodesic whose
length(=ρ) is the shortest one among all curves that belongs to C p,q .
Proof.
(1).By theorem 2.3, polar coordinates can be introduced on a neighborhood U of p.
Since manifold topology and metric topology coincides, such a ρ can be found.
(2).Consider any c ∈ C p,q . Without loss of generality, let’s assume it’s smooth. c
may leave our polar coordinate neighborhood. Let t0 := inf{t ∈ τ : d(c(t), p) ≥ ρ}, then
c|[0,t0 ] ⊂ B(p, ρ). In polar coordinates, write c(t) = (r(t), ϕ(t)), c(t0 ) = (ρ, ϕ(t0 )). We
calculate
Z t0 q
L(c|[0,t0 ] ) = gi j (c(t))ċi (t)ċ j (t)dt
0
Z t0 q
= grr (c(t))ṙṙ + gϕϕ(c(t))ϕ̇ϕ̇ dt
0
Z t0 Z t0
≥ |ṙ|dt ≥ | ṙdt| = |r(t0 ) − r(0)| = ρ
0 0
Moreover, “=” holds iff. gϕϕ ϕ̇ϕ̇ ≡ 0(⇔ ϕ̇ = 0 ⇒ ϕ ≡ const) and ṙ ≥ 0 or ṙ ≤ 0. Hence
the “=” holds iff c(t) = (t, ϕ0 ), where q = (ρ, ϕ0 ). Recall (t, ϕ0 ) is the geodesic in the
polar coordinates.
Remark 2.9.
(1)From the proof, we see ∀c ∈ C p.q , L(c) ≥ L(γ), where γ is the radical geodesic. And
“=” holds iff c is a monotone reparametrization of γ.
(2)There may exists other geodesics from p to q, whose length is longer. That is the
“shortestness” property of a geodesic is not global! From Corollary 2.2.1, we see
∀p, q ∈ M, where they are close “enough” to each other, then there exists precisely one
geodesic of shortest length. Can we have a uniform description of the “closedness”
which ensure the existence of shortest geodesics, at least when M is compact? For this
surpose, we first discuss a refinement of Theorem 2.2.1(page 28)(dexp p (0) = I) and
the “totally normal neighborhood”.
Theorem 2.5. (totally normal neighborhood) For any point p ∈ M, there exists a
neighborhood M of p and a number δ > 0, such that, for every q ∈ W,(in other words,
injectivity radius i(q) ≥ δ), and expq (B(0, δ)) ⊃ W.
Remark 2.10. (Terminologies) If exp p is a diffeomorphism of a neighborhood V of
the origin in T p M. Then we call exp p (V) := U a normal neighborhood of p. Theorem
2.2.4 tells, ∃ a neighborhood of p such that W is a normal neighborhood of each q of
W. W is then called a totally normal neighborhood of p ∈ M.
If B(0, ) is such that B(0, ) ⊂ V, we call exp p (B(0, )) the normal ball with cen-
ter p and radius . The geodesic in exp p (B(0, )) that begin at p are referred to as
radical geodesics.
Remark 2.11. By Corollary 2.2.1, any 2 point in W can be connected by a cunique
minimizing geodesic. For the proof, we first discuss a revision.
Theorem 2.6. There exists a neighborhood U of p, U := {(q, v) ∈ T M : q ∈ V, v ∈
T q M, kvk < } such tvhat exp : U → M, (q, v) 7→ expq v is well defined. Consider the
following map.
F : U → M × M, (q, v) 7→ (q, expq )
In particular, we see F(p, 0) = (p, p). Then dF(p.0) : T (p,0) (T M) → (M × M).
Lemma 2.3. For each p ∈ M and with it the zero vector 0 ∈ T p M, dF(p, 0) is nonsin-
gular.
Proof.
Proof of lemma 2.2.1:
First note that, we can identify the tangent space T (p,p) (M × M) to T p M × T p M,
T (p,0) (T M) to T p M × T 0 (T p M) T p M × T p M. F : U → M × M. In local coordinates,
this map centre considered as (x1 , · · · , xn , v1 , · · · , vn ) → (x11 , · · · , x1n , x21 , · · · , x2n ).
We consider dF(p,0) as a linear map T p M × T p M → T p M × T p M,
varying p, F is identity in the first coordinate. Hence on the first factor to the
factor dF(p,0) is identity.
fix p and vary v in T p M, the first coordinate of F is fixed and the second coordinate
is exp p v.
Hence, dF(0,p) is identically 0 from the second factor to the first factor and identity
from the second factor to the second factor.(Theorem 2.2.1, page 28)
1 1
! !
dFv=0 dFq=p I 0
2 2 = (2.2.7)
dFv=0 dFq=p I 0

proof of theorem 2.6:
By lemma and the inverse function theorem, we know that F is a local diffeomor-
phism. This means that ∃ a neighborhood U 0 ⊂ U of (p, 0) ∈ T M s.t. F maps U 0
diffeomorphically onto a neighborhood W 0 of (p, exp p 0 p = p) ∈ M × M.
By shrinking U 0 if necessary, we can take U 0 to be the form
U 0 = {(q, v) : q ∈ V 0 , v ∈ T q M, kvk < δ}
where V 0 ⊂ V is a neighborhood of p ∈ M.
Now choose a neighborhood W of p in M so that W × W ⊂ W 0 . Then form the
definition of F, we see exp p (B(0, δ)) ⊃ W.
Now, we have some immediate consequence.
Corollary 2.3. Let Ω be a compact subset of a Riemannian manifold M. There ex-

ists ρ0 > 0 with the propert for any p ∈ Ω, Riemmanian polar coordinates may be
introduced on B(p, ρ0 )
Proof. ∀p ∈ Ω, we can find a totally normal neighborhood W p of p. By compactness,

N
we have a finite subcover {W pi }i=1 of {W p } p∈Ω of Ω. Since for each W p , ∃ a ρ p s.t.
Riemmanian polar coordinates may be defined on B(q.ρ p ), ∀q ∈ W p . We pick ρ0 =
min {δ pi }
i=1,··· ,N
Corollary 2.4. Let Ω be a compact subset of a Riemannian manifold M. Then there

exists ρ0 > 0 with the property that for any two points p, q ∈ M with d(p, q) ≤ ρ0 can
be connected by precisely one geodesic of shortest path.
The geodesic depends continuously on (p, q):
Proof. ρ0 from Corollary 2.2.2 satisfies the first claim by Corollary 2.2.1. Morever,
given (p, q), d(p, q) ≤ ρ0 , there exists a unique v ∈ T p M(given by F −1 (p, q) = (p, v))
that depends continuously on (p, q) and is s.t. γ(v) = v.
Corollary 2.5. Let M be a compact Riemannian manifold, i(M) > 0.
Local isometries map geodesics to be geodesics

Recall that a C ∞ differentiable map h : M → N is a local isometry, if ∀p ∈ M, ∃ a
neighborhood U for which h|U : U → h(U) is an isometry and h(U) is open in N and
g|U = h∗ (γ|h(U) ) where (gi j (p), (γαβ (h(p))) are the metrics on U, h(U) respectively. In
α
∂α (p)
fact, gi j (p) = γαβ (h(p)) ∂h∂x(p)
i ∂x j
.
A local isometry has the same effect as a coordinate change. We have already see
in the Homework Exercise 2, that the geodesic equations
∂x
i
ẍ + Γijk ẋ j ẋk = (ẏα + e
Γαηγ ẏη ẏγ ) α .
∂y
i
Hence geodesic is mapped to be geodesic(det( ∂yx α ) , 0.). Intuitively, Isometries
leave the lengths of tangent vectors and therefore also the lengths and energies of curves
invariant. Thus, critical points, i.e. geodesics, are mapped geodesics.
This observation has interesting consequences.
2.3. GLOBAL PROPERTIES:HOPF-RINOW THEOREM 37
Example 2.4. (geodesics of S n )

The orthogonal group O(n + 1) operates isomotrically on Rn+1 , and since it maps
S into S n , it also operates isometricaly on S n .
n
Let now p ∈ S n , v ∈ T p (S n ). Let E be the two dimensions plane through the origin
of Rn+1 containing v.
Claim:the geodesic γv pass through p with tangent vector v is the great circle
through p with tangent vector v(parametrized proportionally to arc length), i.e. the
intersection of S n and E.
Proof. Let S ∈ O(n + 1) be the reflection acress E, then S v = v, S p = p.

γv is a geodesic⇒ S γv is also a geodesic through p with tangent vector v. By
uniqueness result, γv = S γv .
Hence image of γv is the great circle.
Example 2.5. (geodesics on T2 )

ω1 = (1, 0) ∈ R2 , ω2 = (0, 1) ∈ R2 . Consider z1 , z2 ∈ R2 as equivalent if ∃ m1 , m2 ∈
Z s.t. z1 − z2 = m1 ω1 + m2 ω2 .
The covering map π : R2 → T2 , z 7→ [z], differentiable structure:
∆α ⊂ is open and does not contain equivalent points, then Uα := π(∆α ), zα = (π|−1 ∆α ).
For each chart (U, (π|U )−1 ) we use the Euclidean metric on π−1 (U). Since the trans-
lations
z 7→ z + m1 ω1 + m2 ω2 , m1 , m2 ∈ Z
are Euclidean isometries, the Euclidean metrics on the different components of π−1 (U)(which
are obtained from each other by translations.) yield tyhe same metric on U. Hence the
Riemmanian metric on T2 is well defiend, and π : R2 → T2 is a local isometry. There-
fore, Euclidean geodesics of R2 are mapped onto geodesics of T 2 .
2.3 Global Properties:Hopf-Rinow Theorem

In the last section, we know when two points p, q ∈ M are close enough to each other,
there exists precisely one geodesic with the shortest length. Naturally, one would ask
the following questions.
Question 1: If a curve γ is of shortest lengths, is γ a geodesic?
Question 2: Let γ : [0, 1] → M be a geodesic, is it the shortest curve from γ(0) to
γ(1)?
Question3: Given p, q ∈ M, does there exists a curve from p to q with the shortest
length?
Recall that a geodesic γ has to be parametrized proportionally to its arc length.

Hence, a proper way to formulate Question 1 is:
Question 1’: Let γ : [0, 1] → M, kγ̇k = 1 and ∀ξ ∈ C p,q (piecewise C ∞ curve from
p to q), Length(γ) ≤ Length(ξ). Is γ necessarily a geodesic ?
Proposition 2.1. If a piecewise C ∞ γ : [a, b] → M with parameter proportional to

arc length, has length less than or equal to the length of any other piecewise C ∞ curve
from γ(a) to γ(b), then γ is a geodesic.
Proof. Let t ∈ [a, b], and let W be a totally normal neighborhood of γ(t). There exists
a closed interval I ⊂ [a, b] with nonempty interior, t ∈ I s.t. γ(I) ⊂ W. By the global
“shortestness” property of γ, we know γ(I) is a piecewise smooth curve connecting
two points in W which the shortest length. By corollary 2.2.1, noticing further γ is
parametrized proportionally to arc length, we konw γ|I is a geodesic.
Concerning Question 2, we have found several counter-examples, like
So the answer to Question 2 is “No!” Then one may ask when is a geodesic γ also
a minimizing curve? We will discuss this issue in later lecture.
The answer to Question 3 is also “No”. If a curve γ from p to q is of the shortest
curve, after choosing the parameter proportional to arc length, Proposition tells that γ
must be a geodesic.
Then there is no curve from p to q with the shortest length.(but you have a mini-
mizing sequence of curves)
When is the answer to Question 3 “Yes”? It turns out, one have to require M to be
compact!!
Given a Riemannian manifold (M, g), recall that (M, g) with the distance function
d derived from g is a metric space (M, d). And the topology of (M, d) coincides with
original topology of M. Therefore (M, d) is a complete metric space iff M is complete
as a topological space with regard to its original topologies. So we do not need to
distinguish this two Completeness.
Hopf-Rinow theorem tells completeness implies the existence of minimizing geodesic
between any two points. Morover, H-R thm also gives two several equivalent descrip-
tions of completeness.
Theorem 2.7. (Hopf-Rinow 1931).(Über den Begriff der vollständigen differential ge-
ometrischen Flaäche, Commentarii Mathematici Helvetici, 1931)
Let M be a Riemannian manifold. The following statements are equivalent:
(i) M is a complete metric space.
(ii) The closed and bounded subsets of M is compact.
(iii) ∃p ∈ M for which exp p is defined on all of T p M.
(iv) ∀p ∈ M, exp p is defined on all of T p M.
Each of the statement (1)-(4) implies
(v) Any two points p, q ∈ M can be joined by a geodesic of length d(p, q), i.e. by a
geodesic of shortest length.
Digest of the theorem:

(1) (i)⇒(v) but bot vise versa. Conuterexample: open disc is not complete, but
satisfies property (v).
(2) Definition 2.3.1(geodesically complete): A Riemannian manifold M is geodesically complete
if for all p ∈ M, exp p is defined on all of T p M, or, in other words, if any geodesic
gamma(t) with γ(0) = p is defined for all t ∈ R.
H-R theorem tells completeness(manifold topology, apriori independent of the metric)⇔geodesical completeness.(depends
on the Riemannian metric)
(3) (M, g) as a complete metric space is a very special one as shown in (ii).
Consider a countable set A = {ai ; i = 1, 2, · · · } with a discrete metric, i.e. d :
A × A → [0, ∞) s.t. d(ai , a j ) = δi j . Then (A, d) is complete and bounded. But A is not
compact.
(4) Corollary 2.3.2: Let (M, g) be a complete Riemannian manifold. Then exp p :
T p M → M is surjective for any p ∈ M.
Proof. The “core” of result is (iv)⇒(v):

Given p, q ∈ M, we hope to find a shortest geodesic γ from p to q. We know
γ(0) = p, but how to decide γ̇(0)?
Consider a normal ball B(p, q). Since ∂B(p, ρ) is compact, and d(p, ) is a contin-
uous function, there exists p0 ∈ ∂B(p, ρ) s.t. d(p, ) attain its minimum on ∂B(p, ρ) at
p0 .
Now the idea is the following:
At p0 , consider the normal ball B(p0 , ρ1 ), find p1 to be the point at which d(p, )
attain its minimum on ∂B(p0 , ρ1 ). And, we continue this procedure, and hopefully we
arrive at q.
Two issues in this arguement:
(1) Can the piecewise geodesics(or broken geodesic) be a single geodesic?
(2) Can we arrive at q eventually?
To solve this two issues, we argue as below:
In B(p, ρ), we know the radical geodesic from p to p0 is
c(t) = exp p tV, f or some V ∈ T p M
with p0 = exp p ρV.(That is c is parametrized by arc length.)

We consider the curve
c(t) = exp p tV, t ∈ [0, ∞)(by (iv) we can do this)
we hope to show c(r) = q, where r = d(p, q). If this was shown to be true, we know
c(r) is the shortest one and we are done.
In other words, we hope to prove
d(c(r), q) = 0 (2.3.1)
Next we know
d(c(0), q) = d(p, q) = r (2.3.2)
Consider the set

I := {t ∈ [0, r], d(c(t), q) = r − t}.
(ii)⇔ 0 ∈ I. Hence I , Φ. Moreover, since f (t) := d(c(t), q) − r + t is continuous, and

I = f −1 (0) [0, r], I is closed. Let T = sup t. Since I is closed, we see T ∈ I. If T = r,
T
t∈I
then we are done.
Suppose T < r, consider the normal ball B(c(T ), ρ1 )(Without loss of generality, we
can assume ρ1 < r − T .).
Let p ∈ ∂B(c(T ), ρ1 ) be the point at which d(q, ) attain its minimum on ∂B(c(T ), ρ1 ).
Consider the three points: c(T ), p1 , q. By definition, d(q, c(T )) = r − T . We have
d(c(T ), p1 ) + d(p1 , q) ≥ d(c(T ), q) (2.3.3)
by using triangle inequation.

On the other hand, think of any curve γ from c(T ) to q. There exists t, s.t. γ(t) ∈
∂B(c(T ), ρ1 ).
Length(γ) ≥d(c(T ), γ(t)) + d(γ(t), q)

≥d(c(T ), p1 ) + d(γ(t), q)
≥d(c(T ), p1 ) + d(p1 , q)
⇒ d(c(T ), q) ≥ d(c(T ), p1 ) + d(p1 , q).

Combining with inequation (2.3.3), we get euqality:
d(c(T ), q) = d(c(T ), p1 ) + d(p, q)

⇒ d(p1 , q) = d(c(T ), q) − d(c(T ), p1 )
= r − T − ρ1
= r − (T + ρ1 ).
Now if we show
p1 = exp p (T + ρ1 )V = c(T + ρ1 ), (2.3.4)
we have T + ρ1 ∈ I, which contradicts to the definition of T .
It remains to show (2.3.4). We use Proposition 2.1 to prove it. That is, we show the
curve
c|[0,T ] and the radical geodesic f rom c(T ) to p1 (2.3.5)
is the shortest curve from p to p1 . Note the length of the curve is T + ρ1 .
Consider the three points p, p1 and q to figure out d(p, p1 ).
d(p, q) ≤d(p, p1 ) + d(p1 , q)

⇒ T + ρ1 ≤d(p, p1 )
Hence the “broken” curve in (2.3.5) is the shortest curve from p to q. Then proposition
tells that it is a smooth geodesic when parametrized with arc length.
By uniqueness of geodesics with given initial values, it has to coincide with c.
Therefore
p1 = exp p (T + ρ1 )V = c(T + ρ1 ).
Then we finish the proof of (iv)⇒(v).

Next, we prove the euivalence of (i)-(iv).
(iv)⇒(iii) is trivial.
(iii)⇒(ii): Let K ⊂ M be closed and bounded.
“boundness”⇒K ⊂ B(p, r) for some r > 0.
Since exp p is defined on all of T p M, from the proof for (iv)⇒(v), we know any
q ∈ B(p, r) can be connected to p via c(t) = exp p (tV), with c(d(p, q)) = exp p (d(p, q)V)
for some V.
Hence B(p, r) is the image of the compact ball in T p M of radius r under the con-
tinuous map exp p . Hence, B(p, r) is compact. Since K is assumed to be closed, and
shown to be contained in a compact set, it must to be compact itself.
(ii)⇒(i): Let {pn }n∈N ⊂ M be a cauthy sequence and p0 ∈ M. Since {pn }n∈M is a
cauthy sequence, we have ∀ > 0, ∃ N, when n, l > N,
|d(pn , p0 ) − d(pl , p0 )| ≤ d(pn , pl ) < .
That is {d(pn , p0 )}n∈N is a cauthy sequence in R, then lim d(pn , p0 ) exists. If {pn }n has
n→∞
an accumulation point a0 , i.e. ∃ subsequence {pnk } s.t. pik → a0 as k → ∞.
Pick p0 = a0 , we have
lim d(pn , p0 ) = lim d(pnk , p0 ) = 0.

n→∞ k→∞
That is, pn → p0 as n → ∞.
Otherwise, if {pn } has no accumulate point, then {pn } is closed. Note {pn } is
bounded since it is cauthy.
By assumeption (ii), {pn } is conpact. But each pn is not an accumulate point, we
have pn ∈ Un , pi < Un , ∀i , n. Hence {Un } is an open cover of {pn } without any finite
subcover. This contradicts to the conpactness of {pn }.
(i)⇒(iv): Let γ be a geodesic in M, parametrized by arc length, and being defined
on a maximal interval I. Then I is not empty. Moreover, by the “local existence and
uniqueness of geodesics”(ODE theory), we know I is an open interval. Next we show
I is closed. Then I has to be (−∞. + ∞).
Let {tn }n∈N ⊂ I be converging to t. Notice that
d(γ(tn ), γ(tm )) ≤ |tn − tm | = length o f γ(t[

n )γ(tm ).
We know {γ(tn )}n∈N is a cauthy sequence in M.

M is compact⇒ ∃ p0 ∈ M, γ(tn ) → p0 as n → ∞. And, ∃ δ > 0, ∃ W which is a
totally normal neighborhood of p0 , s.t. ∀q ∈ W, expq (B(0, δ)) ⊃ W.
There exists N, s.t. when n, m ≥ N, we have
|tn − tm | < δ (a)

and γ(tn ), γ(tm ) ∈ W (b)
By (a) and the property of W, there exist a unique geodesic c from γ(tn ) to γ(tm ) less
than δ. Therefore c has to be a subarc of γ. Since expγ(tn ) is a diffeomorphic on
B(0, δ) ⊂ T γ(tn ) M and expγ(tn ) (B(0, δ)) ⊃ W, c extends γ to p0 .
2.4. EXISTENCE OF GEODESICS IN GIVEN HOMOTOPY CLASS 43
Corollary 2.6. Compact Riemannian manifold is complete.
Proof. closed subset of a compact space is compact.
Corollary 2.7. A closed submanifold of a complete Riemannian manifold is complete

in the induced metric. In particular, the closed submanifold of Euclidean space are
complete.
2.4 Existence of geodesics in given homotopy class

In complete Riemannian manifold, any two points p, q can be connected by a shortest
geodesic. In this part, we discuss the existence of such geodesics in given homotopy
class.
Definition 2.5. Two curves γ0 , γ1 on a manifold M with commen initial and end points
p and q, i.e. two continuous maps
γ0 , γ1 : I = [0, 1] → M
with γ0 (0) = γ1 (0) = p, γ0 (1) = γ( 1) = q, are called homotopic if there exists a

continuous map Γ : I × I → M with
Γ(0, s) = p, Γ(1, s) = q ∀s ∈ I
Γ(t, 0 = γ0 (t), Γ(t, 1) = γ1 (t) ∀t ∈ I.
Two closed curves c0 , c1 in M, i.e. two continuous maps c0 , c1 : S 1 → M are called

homotopic if there exists a continuous map c : S 1 × I → M with
c(t, 0) = c0 (t), c(t, 1) = c1 (t) for all t ∈ S 1 .(S 1 , as usual, is the unit circle
parametrized by [0, 2π))
Remark 2.12. The concept of homotopy defines an equivalence relation on the set of
all curves in M with fixed initial and each points as well as on the set of all closed
curves in M.
With the examples of torus in mind, let’s first consider the existence of closed
geodesic on a compact Riemannian manifold.
Theorem 2.8. Let M be a compact Riemannian manifold. Then every homotopy class
of closed curves in M contains a curve which is a shortest curve in its homotopy class
and a geodesic.
As a preparation, we first show

Lemma 2.4. Let M be a compact Riemannian manifold. Let ρ0 > 0 be the constants
with the following property: any two points p, q ∈ M with d(p, q) ≤ ρ0 can be con-
nected by precisely one geodesic of shortest path. Let γ0 , γ1 : S 1 → M be closed
curves with
d(γ0 (t), γ1 (t)) ≤ ρ0 ∀ t ∈ S 1 .
Then γ0 and γ1 are homotopic.
Remark 2.13. The existence of such ρ0 on a compact Riemannian manifold has been
proven in corollary 2.2.3. In fact, we know moreover, this geodesic depends continu-
ously on (p, q).
Proof of lemma 2.4.1:
∀t ∈ S 1 , let ct (s) : I → M be the unique shortest curve(which is therefore, a
geodesic) from γ0 (t) to γ1 (t), as usual parametrized proportionally to arc length. Recall
that ct depends continuously on its end points, hence on t, Γ(t, s) = ct (s) is continuously
and yields the desired homotopy.
Proof of Theorem 2.4.1:
Consider the lengths of the curvesin a given homotopy class: they are numbers in
[0, +∞). Let {γn }n∈N be minimizing sequence for arc length in this homotopy class. All
{γn }n∈N are parametrized proportionally to arc length.
We may assume each γn is piecewise geodesic: for each γn , we may find 0 = t0 <
t1 < · · · < tm < tm+1 = 2π with the property that
ρ0
L(γn |[t j−1 ,t j ] ) < , j = 1, · · · , m + 1.
2
Replacing γn |[t j−1,t j ] by the shortest geodesic arc between γn (t j−1 ) and γn (t j ), we obtain
a new closed curve γen .
By using triangle inequality, we have d(γn (t), γen (y)) ≤ ρ0 . As a consequence of
Theorem 2.4.1, γen is homotopic to γn and the lengths of γen is no longer than γn .
We may thus assume that for any γn , ∃ points p0,n , · · · , pm,n for which d(p j−1,n , p j,n ) ≤
ρ0
2 (pm+1,n := p0,n , j = 1, · · · , m + 1).
Observe that the lengths of γn are bounded as they constitute a minimizing se-
quence. Therefore, we may assume that m is independent of n.
p0,1 p1,1 p2,1 ··· pm,1

p0,2 p1,2 p2,2 ··· pm,2
.. .. .. ..
. . . ··· .
p0,n p1,n p2,n ··· pm,n
.. .. .. ..
. . . ··· .
↓ ↓ ↓ ↓
p0 p1 p2 pm
Recall that the geodesic between p j−1,n and p j,n depends continuously on its endpoints,
and hence converges to the shortest arc between p j−1 and p j .(shortest arc is geodesic)
These shortest geodesic arcs yields a closed curve γ. By Lemma 2.4.1, γ is ho-
motopic to γn , and Length(γ) = lim Length(γn ), Recall {γn } is a minimizing sequence
n→∞
for the length in their homotopy class. Therefore, γ is shortest curve in this homotopy
class.
Claim: γ is a geodesic.
Otherwise, there would exist points p and q on γ for which one of the two arcs of γ
between p and q would have length at most ρ20 , but would not be geodesic.
Then this arc cpq can be shortened by replacing it by the shortest geodesic between
p and q. Denote this new curve as eγ. We have
γ(t)) ≤ ρ0 , ∀ t ∈ S 1 .
d(γ(t),e
However, γ and eγ is homotopic as a consequence of Lemma 2.4.1. This contradicts to

the minimizing property of γ. Therefore, γ is desired closed geodesic.
Remark 2.14. If the compact Riemannian manifold M is simply-connected, the above
arguement leads to the trivial closed geodesic: a point.
Now, we discuss the existence of shortest geodesics in a given homotopic class of
curves with fixed initial values and end points in a complete Riemannian manifold.
Theorem 2.9. Let (M, g M ) be a complete and connected Riemannian manifold, p, q ∈
M. Every homotopy class of paths from p to q contains a geodesic γ that minimized
length among all admissible curves in the same homotopy class.
The idea to prove Theorem 2.4.2 is first going to the universal covering manifold
of M. In M,e curves connecting corresponding points e p and e
q only have one honotopy
class. For that purpose, we need first show that M is complete ⇒ M e is complete.
Recall: A covering map is a surjective continuous map π : M e → M between con-
nected and locally-path-connected topological spaces, for which each points of M has
connected neighborhood U that is evenly covered, meaning that each connected com-
ponent of π−1 (U) is mapped homeomorphically onto U by π.
It is called a smooth covering map if M e and M are smooth manifolds and each
component π (U) is mapped diffeomorphically onto U.
−1
Any Riemannian metric on M induceds a Riemannian metric on M. e This makes π

into a Riemannian covering. In particular, π is a local isometry.
Lemma 2.5. Suppose M e and M are connected Riemannian manifolds, and π : M

e→M
is a Riemannian covering map. If M is complete, then M
e is also complete.
Proof. Let e p∈M e and e e be arbitrary, and let p = π(e

v ∈ T ep M p), and v = dπ(e
p)(ev).
Completeness of M implies that the geodesic γ with γ(0) = p and γ̇ = v is defined
for all t ∈ R.(Recall a fundamental property of covering map is the path-lifting property:
If π : Me → M is covering map, then every continuous map γ : I → M lifts to a path e γ
in M s.t. π ◦ e
e γ = γ.)
γ of γ starting at e
Here the lifts e v. Since π is a local
p with the initial tangent vector e
isometry, we know e γ is a geodesic. Since γ is defined for all t ∈ R, so does e γ. This
proves the completeness of M. e
Prood of Theorem 2.4.2: Consider the universal convering π : M e → M cof M,
endowed with the induced metric e g = π∗ g. Given p, q ∈ M and a path σ : [0, 1] → M
from p to q. Choose a e p ∈ π−1 (p), and let σe : [0, 1] → M e be the lift of a starting with
e q=σ
p, and set e e(1).
By Hopf-Rinow, and the fact M e is complete, there exists a minimizing e g-geodesic
γ from e
e p to e
q.
Because π is a local isometry, γ = π ◦ e γ is a geodesic from p to q in M.
Since Me is simply connected, we have σ e and e γ are homotopic. Hence σ and γ
are also homotopic. That is γ is a geodesic in the homotopy class [σ]. If σ1 is any
other admissible curve from p to q in the homotopy class [σ], then by the monodromy
theorem, its lifts γe1 starting at e
p also ends at qe1 ,(γe1 and σ
e are homotopic, which is trivial
in a simply connected space.) In M, e we know Length(e γ) ≤ Length(γe1 ). Therefore,
Length(γ) ≤ Length(γ1 )
We’d like take this chance to discuss further about Riemannian covering map.
Theorem 2.10. Let ( M,e
e g) and (M, g) are connected Riemannian manifolds with M e
complete, and π : M
e → M is a local isometry. Then M is complete and π is a Rieman-
nian covering map.
e and M are connected Riemannian manifolds and π : M
Corollary 2.8. Suppose M e→
M is a Riemannian covering map. Then M is complete iff M
e is complete.
Proof. Combination of Theorem 2.4.3 and Lemma 2.4.2.

Path-lifting property for geodesic of a local isometry π.
Let p ∈ π( M),e and e p ∈ π−1 (p). Let γ : I → M be a geodesic with p = γ(0), v = γ̇(0).
Let e v := (dπ(e −1
p)) (v) ∈ T ep M.(π
e induces dπ(e p) is a linear isometry.) Let eγ be the
geodesic in M with initial point e
e p and initial tangent vector e v. Since Me is complete,
π◦e γ is a geodesic with initial point p and initial tangent vector v. Hence, π◦e γ = γ on I.
So e γ|I is a lift of γ starting at ep.
M is complete: Let p ∈ π(M), γ : I → M be any geodesic starting at p, then
γ has a lift e γ : I → M. e Since M e is complete, π ◦ e γ is a geodesic defined on all of
R and coincides with γ on I. That is γ extends to all of R. Thus M is complete by
Hopf-Rinow Theorem.
π is suijective.
⇒
∀ep ∈ M, e write p = π(e p). Let q ∈ M be arbitrary. M is completeH − R ∃ a
minimizing geodesic from p to q.
Let e γ be the lift of γ starting at e p, and r = d(p, q), we have π(e γ(r)) = γ(r) = q. So
q ∈ π( M).
e
Every point of M has a neighborhood U that is evenly covered.
Let p ∈ M, let U = B (p) be a geodesic ball(normal ball) centered at p, < in j(p).
Write π−1 (p) = {f pα }α∈A . For each α write U fα be the metric ball of radius around
pα .
f
Claim: U fα T U fβ = ∅, ∀α , β.
Proof: ∀α , β, there exists a minimizing geodesic e γ from fpα to peβ because M
e is
complete. The projective curve γ := π ◦ e γ is a geodesic that starts and end at p, whose
length is the same as that of e γ. Such a geodesic must leave U and reenter it.
γ) =
Since all geodesics passing through p are radial geodesics, we have Length(e
pα , peβ ) ≥ 2 ⇒ Uα Uβ = ∅
T
Length(γ) ≥ 2 ⇒ deg (f
Claim:π−1 (U) = U .
Sf
α
α
Proof: ∀e q ∈ Ufα for some α, there is a geodesic eγ of length¡ from p] a lpha to e
q.
Then π ◦ e γ is a geodesic of the same length from p to π(eq), showing that π(e q) ∈ U =
α ⊆ π (U). ∀e q ∈ π−1 (U), we get q = π(e
Sf −1
B (p). i.e. U q). That is, q ∈ U. So that is a
α
minimizing radial geodesic. γ in U from p to q, and r = dg (p, q) < . Let e γ be the lift
of γ starting at e
q.
It follows that π(eγ(r)) = γ(r) = p. Therefore e
γ(r) = f pα for some α, and deg (eq, f
pα ) ≤
Length(e γ) = r < . So e
q ∈ Uα .
f
It remains to show that π : Ufα → U is a diffeomorphism for each α.
It is certaining a local diffeomorphism. It is bijective: we can construct the inverse
pα .
explicitly. It sends each radial geodesic starting at p to its lift starting at f
This completes the proof.
Chapter 3
Connections, Parallelism, and

covariant Derivatives.
Consider the geodesic equation again. In (U, x),
ẍi (t) + Γijk (x(t)) ẋ j (t) ẋk (t) = 0, i = 1, · · · , n. (*)
Recall under coordinate change (xi ) → (yα ), the christoffel symbols behave as
α ∂yη ∂yγ ∂xi ∂2 yα ∂xi

Γijk = Gamma
] ηγ (y(x))
α
+ j k α.
∂x ∂x ∂y
j k ∂x ∂x ∂y
Therefore Γijk is not coefficients of a tensor!! This fact suggest in particular that we
should pay more attention to taking derivetive in local coordinates. It would be nice if
we have “the derivetive of a tensor is again a tensor”. This will be solved by so-called
“covariant derivatives”.
On the other hand, the LHS of (*) behaves under coordinate change (xi ) → (yα ) as
∂x i
( ẍi (t) + Γijk (x) ẋ j ẋk ) = (ÿα + e
Γαηγ (y(x))ẏη ẏγ ) α .
∂y
That is, it behaves like a (1,0)-tensor(i.e. vector field). Recall, in local coordinatesm if
X = X i ∂x∂ i = Y α ∂y∂α . Then X i = Y α yxα . This suggest that ẍi (t) + Γijk (x) ẋ j ẋk is coefficients
i
of a (1,0)-tensor. This leads to the concepts of connections, and parallelism.
3.1 Affine Connections.

Refenrence:[WSY Chap.1][do Carmo, 2.2]
On Rn , let v be a vector at p ∈ Rn , f ∈ C ∞ (U), p ∈ U ⊂ Rn , we hava the following
“directional derivative” of f at p along v:
f (p + tv) − f (p)
Dv f = lim .
t→0 t
49
50CHAPTER 3. CONNECTIONS, PARALLELISM, AND COVARIANT DERIVATIVES.
Let X be a C ∞ vector field. In local coordinates, (U, X), we have
X = (X 1 , · · · , X n ), X i ∈ C ∞ (U)
where X = X i ∂x∂ i . Then the directional derivative of X along v is defined as Dv X =

(Dv X 1 , · · · , Dv X n ). That is Dv X = (Dv X i ) ∂x∂ i .
P
i
It is direct to check the following properties:
(a) Dαv X = αDv X, ∀α ∈ R
(b) Dv ( f X) = (Dv f )X + f Dv X, ∀ f
(c) Dv (X1 + X2 ) = Dv X 1 + Dv X 2 , ∀X 1 , X 2
(d) Dv1 +v2 X = Dv1 X + Dv2 X, ∀v1 , v2 .
In fact, we also have Dv ∂x∂ i = 0. But this property can not be extended to manifold
case. In general, we can define the following concept:
Definition 3.1. (Affine connection). An affine connection O on a smooth manifold M is

a map
O : Γ(T M) × Γ(T M) → Γ(T M).
(Γ(T M) is the set of all smooth vector fields on M.)
O
This map is denoted by (X, Y) → OX Y, and which satisfies the following properties:
(i) O f X+gY Z = f OX Z + gOY Z (linear over the C ∞ functions in the arguement X.)
(ii) OX (Y + Z) = OX Y + OX Z
(iii) OX f Y = f OX Y + X( f )Y
in which X, Y, Z ∈ Γ(T M) anf f,g are any real-valued C ∞ functions on M. The vector
field OX Y is called the covariant derivative of Y along X(with respect to the connection
O).
Digest:
1 On Rn , the “directional derivatives” provides an affine connection. For X, Y ∈
Γ(T Rn ), define (OX Y)(p) = DX(p) Y, p ∈ R. Then one can check 4 satisfies (i)-(iii).
2 Let X, Y ∈ Γ(T M). In a local coordinates (U, x1 , · · · , xn ), X, Y can be considered
as vector fields on x(U) ⊂ Rn . In (U, x), we can define OX Y as the directional derive-
tive DX Y. A natural question is: can we obtain an affine connection by defining it as
directional derivatives in every local coordinates?
The answer is No! Suppose we have two coordinates (U, x1 , · · · , xn ) and (V, y1 , · · · , yn ).
S
When U V , ∅, we have
X ∂ ∂
DX Y = (DX f i ) where Y = f i i in U
i
∂x i ∂x
X ∂ ∂
= (DX gi ) where Y = gi i in V
i
∂yi ∂y
X ∂xk ∂
= (DX gi ) .
i
∂yi ∂xk
3.1. AFFINE CONNECTIONS. 51
Need
∂xi
DX f i = DX G j
∂y j
∂y j ∂xi
= DX ( f k k ) j
∂x ∂y
∂y j ∂xi
= DX f k δik + f k DX ( )
∂xk ∂y j
∂y j ∂xi
= DX f i + f k [DX (δik ) − k DX k ]
∂x ∂y
∂y ∂x
i i
= DX f i − f k k j
∂x ∂y
∂xi
= DX f i − g j DX j
∂y
That is we need
∂xi
g j DX = 0, ∀ i. (*)
∂y j
We can find examples that (*) does not hold.

3 Existence: Many “trivial” connections: Fix a coordinate neighborhood U, define
a “local” connection OU on U via directional derivatives on Rn . This can be extended
“trivially” to a connection on M.
Lemma 3.1. The set of all affine connections on M form a convex set. Namely, if
O(1) , · · · , O(k) are affine connection on M, and f1 , · · · , fk ∈ C ∞ (M), s.t. fi = 1. Then
P
P (i) i
fi O is also an affine connection on M.
i
Proof. Properties (i),(ii) of an affine connection can be checked directly.

For (iii), we check for X, Y ∈ Γ(T M), f ∈ C ∞ (M).
X X
( fi O(i) )X (gY) = f i (O(i)
X (gY))
i i
X
= fi (X(g)Y + gO(i)
X Y)
i
X X
=( fi )X(g)Y + g( fi O(i) )X Y.
i i
f i = 1.
P
Here we need the property that
i
Exercise 3.1. Find a nontrivial connection on M via “partition of unity”.
4 Locality: “OX Y depends only on local information of X and Y”

Proposition 3.1. For any open subset U ⊂ M, if

X|U = X|
e U and Y|U = Y|
e U,
then OX Y|U = OXeY|

e e.
U
(1) (2)
Proof. We will show OX Y|U = OXeY|U = OXeY|
e e.
U
For (1), it’s enough to show X|U = 0 ⇒ OX Y|U (a).
For (2), it’s enough to show Y|U = 0 ⇒ OX Y|U (b).
Proof of (a): ∀p ∈ U, ∃ open V ⊂ U and a function f ∈ C0∞ (U) s.t. f = 1 on V. We
i
check (1 − f )X = X since X|U = 0. Then OX Y = O(1− f )X Y = (1 − f )OX Y. In particular,
OX Y(p) = (1 − f (p))OX Y = 0. Therefore OX Y|U = 0.
Exercise 3.2. Show that Y|U = 0 implies OX Y|U = 0.
Proposition 3.2. If X(p) = X(p),
e then OX Y(p) = OXeY(p).
Proof. Again, it’s enough to show X(p) = 0 ⇒ OX Y(p) = 0 (∗).
By proposition 3.1, we only need to show (*) for X supported in an coordinate
neighborhood (U, x), with x(p) = 0(the origin of Rn ). Now we can write X = X I ∂x∂ i
with X i (0) = 0. By taylor’s theorem, ∃ functions Xki s.t.
X i (x1 , · · · , cn ) = X i (0) + xk Xki = xk Xki .
So OX Y = O xk X i ∂ Y = xk O X i ∂ Y. In particular at p,
k ∂xi k ∂xi
OX Y(p) = xk (p)OX i ∂ Y(p) = 0.

k ∂xi

Consequently, for v ∈ T p M, and Y ∈ Γ(T M), we can define Ov Y(p) := OX Y(p),
where X is any vector field with X(p) = v.(This is like a “directional derivative”of Y at
p along v.)
But, it is not true that Y(p) = Y(p)
e ⇒ OX Y(p) = OX Y(p).
e It is not hard to construct
counterexamples.
Proposition 3.3. Let γ : (−, ) → M be a smooth curve on M, with γ(0) = p and
γ̇(0) = v. Suppose X, Y, Y
e are vector fields on M s.t.
X(p) = v, Y(γ(t)) = Y(γ(t)),

e − < t,
then OX Y(p) = OX Y(p).
e
Proof. It’s enough to show Y = 0 along γ ⇒ Ov Y(p) = 0.
Let (U, x1 , · · · , xn ), p ∈ U be a coordinate neighborhood around p with x(p) = 0.
Let Y = f i ∂x∂ i . Then
∂ ∂ ∂
Ov Y(p) = Ov ( f i )(p) = (v( f i ) i + f i Ov i )(p)
∂xi ∂x ∂x
d ∂ ∂
= |t=0 f i ◦ γ(t) i + f i (p)Ov i
dt ∂x ∂x
Since f i ◦ γ(t) = 0, t ∈ (−, ), we have Ov Y(p) = 0.
3.2. PARALLELISM 53
3.2 Parallelism
What is going an geometrically? [Spivak II, chapter 6]
Consider a curve c : [a, b] → M. By a vector field V along c, we mean
t ∈ [a, b] 7→ V(t) ∈ T γ(t) M.
In a coordinate neighborhood (U, x1 , · · · , xn ), we can write

n
X ∂
V(t) = vi (t) |c(t) .
i=1
∂xi
We call V a C ∞ vector field along c if the functions vi are C ∞ on [a, b]. This is
equivalent to saying that
t 7→ V(t) f
is C ∞ for every C ∞ function f on M.
Notice that a vector field V along c may not be extended to a vector field on M.
When c is an embedding, V(t) can be extended to a vector field V

e on M. We have
V(c(t))
e = V(t), ∀t ∈ [a, b].
e is a C ∞ vector field along c.

Then O dc V
dt
By locality, we know O dc V e We call O dc V

e does not depend on the extension V. e the
dt dt
covariant derivative of V along c, we denote it by the convenient symbolism DV dt .
We would like to generalize this covariant derivative along c to any curve c.(This is
actually the concept of “induced connections” for which we will discuss later.)
Proposition 3.4. Let M be a differential manifold with an affine connection O. There

exists a unique correspondence from C ∞ vector fields V along the smooth curve c :
[a, b] → M to C ∞ vector fields along c : V → DV dt , called the covariant derivative of V
along c, such that
D
(a) dt (V + W) = DV dt + dt .
DW
df
(b) dt ( f V) = dt V + f dt , for F ∈ C ∞ ([a, b]).
D DV
(c) If V(s) = Y(c(s)) for some C ∞ vector field Y defined in a neighborhood of c(t),
then DVdt = O Y.
dc
dt
Proof. Let us suppose initially that there exists a correspondence satisfying (a),(b) and
(c).
Let p = c(t0 ) ∈ M, and (U, x1 , · · · , xn ) is a coordinate neighborhood of p. For t
n
sufficiently to t0 , we can express V locally as V(t) = v j (t) ∂x∂ j |c(t) . By (a),(b),(c), we
P
j=1
have
n
DV (a) X D j ∂
= (v (t) j |c(t) )
dt j=1
dt ∂x
n
(b)
X dv j (t) ∂ D ∂
== [ | + v j (t) ( j |c(t) )]
j c(t)
j=1
dt ∂x dt ∂x
n
(c)
X dv j (t) ∂ ∂ dc ∂ dci ∂
= [ | c(t) + v j
(t)O dc ] ( = dc( ) = |c(t) )
j=1
dt ∂x j dt ∂x j dt ∂t dt ∂xi
n
X dv j (t) ∂ dci ∂
= [ |c(t) + v j (t) O ∂i |c(t) j ]
j=1
dt ∂x j dt ∂x ∂x
∂ ∂
(c(t)) ∂x∂ k |c(t)
n o n o
Note O ∂
| is a C ∞ vector field along c. Hence ∃
∂x j
k
ij s.t. O ∂
| ∂x j
= k
ij
∂xi c(t) ∂xi c(t)
n k Pn k o dci j ∂
DV
= ( dvdt +
P
⇒ dt i j (c(t)) dt v (t)) ∂xk |c(t) (∗)
k=1 i, j
The expression (*) show us that if there is a correspondence satisfying (a),(b),(c),
then such a correspondence is unique.
To show existence, define DV dt in (U, x) by (*). We can verify that (*)Tpossesses the
desired properties. If (V, y) is another coordinate neighborhood with U V , ∅, then
we define DV V by the uniqueness of DV
T
dt in (V, y) by (*), the definition agree in U dt in
U. Therefore, the definition can be extended over all of M.
Remark 3.1. Even at points where dc dt = 0, dt is not necessarily 0!! If c is a constant

DV
curve, c(t) = p ∈ M, ∀t. Then a vector field V along c is just a curve in T p M, and DV dt
is just the ordinary derivative of this curve.
Definition 3.2. (Parallelism) Let M be a differentiable manifold with an affine con-

nection O. A vector field V along a curve c : [a, b] → M is called parallelism when
dt = 0, ∀t ∈ [a, b]. When M = R , O be the directional derivative, we obtain the
DV n
standard picture of a parallel vector field.
Proposition 3.5. [do Carmo, Prop 2.6] Let M be a differentiable manifold with an
affine connection O. Let c : I → M be a smooth curve in M, and let V0 ∈ T c(t0 ) M, t0 ∈ I.
Then there exists a unique parallel vector field V along c, such that V(t0 ) = V0 .
Remark 3.2. V(t) is called the parallel transport of V(t0 ) along c.
Proof. First consider the case when c(I) is certained in a coordinate neighborhood
(U, x1 , · · · , xn ). Then V0 can be expressed as:V0 = v0j ∂x∂ j |c(t0 ) .
P
j
3.2. PARALLELISM 55
Suppose there exists a vector field V in U which is parallel along c, with V(t0 ) = V0 .
Then V = v j (t) ∂x∂ j |c(t) satisfies
P
DV X dvk X n k o dci j ∂
0= = { + i j (c(t)) v (t)} k |c(t) .
dt k
dt i, j
dt ∂x
The equations
dvk X n k o dci j
+ i j (c(t)) v (t) = 0, k = 1, · · · , n
dt i, j
dt
are linear differential equations. So there is a unique solution satisfying the initial
condition
vk (t0 ) = vk0 , k = 1, · · · , n.
Due it’s linearity, the solution is defined for all t ∈ I, this proves the existence and
uniqueness of V in this case. In general, for any t1 ∈ I, there is a finite cover of
c([t0 , t1 ]) by coordinate neighborhood.
In each of those coordinate neighborhood , V is defined. By uniqueness, the defini-

tions coincide when the intersections are not empty, there allowing the definition of V
along all of [t0 , t1 ]
Now consider c : [a, b] → M, Va ∈ T c(0) M. Then there is a unique V(t) ∈ T c(t) M
s.t. Vt is the parallel transport of Va along c.
It’s clear from the definition that

(V + W)t = Vt + Wt , (λV)t = λVt .
That is, we have a linear transformation Pc,a,t = Pt : T c(a) M → T c(t) M, Va 7→ Vt .
Moreover, Pt is one-to-one. Its inverse is given by the parallel transport along the
reversed portion of c from t to a.
ϕ : [a, b] → M, t 7→ c(a + b − t).
dϕ dc
ϕ(a) = c(b), ϕ(b) = c(a). (t) = − (a + b − t).
dt dt
Therefore, Vb ∈ T c(b) M is the parallel transportation of Va ∈ T c(a) M along c iff Va
is the parallel transportation of Vb along ϕ.(When c is embedding, this is seen from
O dϕ Ve = O− dc (a+b−t) V
e = −O dc (a+b−t) V.)
e
dt (t) dt dt
Hence Pt is an isomorphism between two vector space T c(0) M and T c(t) M.
Remark 3.3. (justification of the term “connection”) A connection O gives the possi-
bility of comparing, or “connecting”, tangent spaces at different points.
Note the isomorphism between two tangent spaces given by the parallel transport
depends on the choice of curves connecting the two points.
The parallel transport Pt is defined in terms of O, but we can also reverse the pro-
cess.
Proposition 3.6. [spivak II, Chapter 6.Prop 3] Let c be a curve with c(0) = p and
ċ(0) = X p . Let Y ∈ Γ(T M), then
1
OX p Y = lim (P−1
h Yc(h) − Y p )
h→0 h
Remark 3.4. Parallel transport enables us to use the idea of “directional derivative”
to define OX p Y.
Proof. Let V1 , c . . . , Vn be parallel vector fields along c which are linearly independent
at c(0), and(since parallel transports are isomorphisms), hence at all points of c. Set
n
Y(c(t)) = f i (t)Vi (t). Then
P
i=1
1
lim (P−1 h Yc(h) − Y p )
h→0 h
1 X i X
= lim ( f (h)P−1 h Vi (h) − f i (0)Vi (0))
h→0 h
i i
1X i
= lim ( f (t)V i (0) − f i (0)Vi (0))
h→0 h
i
X X dfi
= lim( f i (t) − f i (0))Vi (0) = |t=0 Vi (0)
i
h→0
i
dt
D X
= |t=0 f i (t)Vi (t)
dt i
=OX p Y.

Remark 3.5. Recall (*) on page 73, and geodesic equation in last Chapter. If γ :
[a, b] → M is a geodesic, then we have Dγ̇(t)dt = 0 where dt is determined by a connec-
D
∂ k ∂
tion O on M, for which in (U, x), O ∂i ∂x j = Γi j ∂xk .
∂x
3.3. COVARIANT DERIVATIVES OF A TENSOR FIELD 57
3.3 Covariant derivatives of a tensor field

In this section, we extend the covariant derivative of a vector field Y along X to that
of a tensor field along X. Similar as previous cases, we can do this via pure algebraic
discussions, or via parallel transport.
For (0,0)-tensor(=functions), we have a nice derivative:
OX : C ∞ (M) → C ∞ (M), f 7→ cOX f = X( f ) = d f (X).
We can check that this derivative satisfies (i)-(iii) in Definition 3.1.1. The following
property enables us to define the covariant derivative 4X A of (r,s)-tensor A(via an al-
gebraic discussions).
In fact we can define a connection on (r,s)-tensor fields
O : Γ(T M) × Γ(⊗r,s T M) → Γ(⊗r,s T M), (X, A) 7→ OX A.
Proposition 3.7. Let M be a diffentiable manifold with an affine connection O. There
is a unique connection on all tensor fields O : Γ(T M) × Γ(⊗r,s T M) → Γ(⊗r,s T M) that
satisfies
(i) O f X+gY A = f OX A + gOX A.
(ii) OX (A1 + A2 ) = OX A1 + OX A2 .
(iii) OX ( f A) = X( f )A + f OX A.
and
(iv) O coincide with the given connections on Γ(T M) and C ∞ (M).
(v) OX (T 1 ⊗ T 2 ) = (OX T 1 ) ⊗ T 2 + T 1 ⊗ (OX T 2 )
(vi) C(OX T ) = OX C(T ), where C : Γ(⊗r,s T M) → Γ(⊗r−1,s−1 T M) is the contraction
map that pairs the first vector with the first covector.
Remark 3.6. (i)-(iii) is the properties for a connection, (iv)-(vi) provides a unique
extension to all tensor fields.
Proof. First, we derive the formula of O on 1-forms.
Let ω ∈ Ω1 (M) = Γ(T ∗ M) be any 1-form, then
(iv)
X(ω(Y)) = OX (ω(Y)) = OX (C(ω ⊗ Y))
(vi) (v)
= COX (ω ⊗ Y) = = C(OX ω ⊗ Y + ω ⊗ OX Y)
= (OX ω)(Y) + ω(OX Y)
So we conclude
1 (OX ω)(Y) = X(ω(Y)) − ω(OX Y)

Next, we can use (v) iteratively to show that for any (r,s)-tensor field A,
2 (OX A)(ω1 , · · · , ωr , Y1 , · · · , Y s )

X X
=X(A(ω1 , · · · , ωr , Y1 , · · · , Y s )) − A(ω1 , · · · , OX ωi , · · · , ωr , Y1 , · · · , Y s ) − A(ω1 , · · · , ωr , Y1 , · · · , OX Y j , · · · , Y s )
i j
This shows the uniqueness.

1 and
For the existence, one need to check that the connections defined by 2
satisfies all conditions (i)-(vi).
Remark 3.7. OX A is called the covariant derivative of the (r,s)-tensor fields A along
X.
The properties (iv)-(vi) are very natural. To elaborate this point, we briefly discuss
another way of defining OX A, via parallel transport.
Recall for an isomorphism ϕ : V → W between two vector spaces Vand W, there
is an induced isomorphism ϕ∗ : W ∗ → V ∗ between their dual spaces W ∗ , V ∗ defined by
f or α ∈ W ∗ : ϕ∗ (α)(v) := α(ϕ(v)), ∀v ∈ V.
Then for any vi ∈ V, α j ∈ V ∗ , define
ϕ(v1 ⊗ · · · ⊗ vr ⊗ α1 ⊗ · · · ⊗ α s )
e
:=ϕ(v1 ) ⊗ · · · ⊗ ϕ(vr ) ⊗ (ϕ∗ )−1 (α1 ) ⊗ · · · ⊗ (ϕ∗ )−1 (α s )
Using linearity, we can extend e ϕ to ⊗r,s V all (r,s)-tensor over V! This defines an iso-
r,s r,s
morphism between ⊗ V → ⊗ W.
Recall the parallel transport along c. Pc,t : T c(0) M → T c(t) M is an isomorphism. We
can extend it to be an isomorphism P ec,t : ⊗r,s T c(0) M → ⊗r,s T c(t) M. As in proposition
3.4, we define
1 e−1
OX p A := lim (P c,h Ac(h) − A p ) (**)
h→0 h
where c is a curve with c(0) = p, ċ(0) = X p .

Clearly if A ∈ Γ(⊗r,s T M), then OX p ∈ Γ(⊗r,s T M). We also check that OX p A given in
(**) satisfies prop.(iv)-(vi).
Exercise 3.3. Let Y ∈ Γ(T M), ω, η ∈ Γ(T ∗ M). Consider the tensor field K = Y ⊗ ω ⊗ η.
Let X p ∈ T p M, and OX p K be defined in (**).
(i) Show OX p K = OX p Y ⊗ ω ⊗ η + Y ⊗ OX p ω ⊗ η + Y ⊗ ω ⊗ OX p η.
(ii) Let CK = ω(Y)η. Show OX p (CK) = C(OX p K).
Remark 3.8. The definition (**) leads to be dependent on X p and the curve c. How-
ever, (**) does not depend on choice of c. Recall OX p Y depends only on X p . We only
need to show for any η ∈ Γ(T ∗ M), OX p η also depends only on X p . We need show
(OX p η)(Y), ∀Y ∈ Γ(T M), depends only on X p , not on c.
Consider Y ⊗ η, we have
OX p (Y ⊗ η) = (OX p Y) ⊗ η + Y ⊗ OX p η
exchangewithcontraction
⇒ X p (η(Y)) = η(OX p Y) + (OX p η)Y
⇔(OX p η)(Y) = X p (η(Y)) − η(OX p Y)
RHS only depends on X p , not on c.
Now, for any tensor field A, and a field X, we can define (OX A)(p) = OX p A, ∀p ∈
M.
3.4. LEVI-CIVITA RIEMANNIAN CONNECTIONS 59
3.4 Levi-Civita Riemannian Connections

There are too many connections on a given smooth manifold. Let X, Y ∈ Γ(T M). In a
coordinate neighborhood (U, x), write X = X i (x) ∂x∂ i , Y = Y j (x)
f rac∂∂x j . By definition we have
∂ ∂Y j ∂ ∂
∇X Y = ∇X i ∂ ) = Xi i
(Y j + X i Y j ∇ ∂i j .
∂xi∂x j ∂x ∂x j ∂x ∂x
∂
∈ Γ(T M), there exists functions ikj s.t. ∇ ∂i |c(t) ∂x∂ j = ikj , k = 1, 2, · · · , n
n o n o
Since ∇ ∂i ∂x j
∂x ∂x
s.t.
∂ n o
∇ ∂ = k
.
∂xi ∂x j ij
∂ n o ∂ n o ∂
⇒ ∇X Y = X(Y j ) + XiY j k
= (X(Y k ) + X i Y j
k
ij ) .
∂x j ij
∂xk ∂xk
n o
That is, the connection ∇ is determined by the n s smooth functions ikj .
Let c : [a, b] → M be a curve such that the relocity vector field ċ(t)(along c) is
parallel. Then locally, we can write c(t) = (x1 (t), · · · , xn (t)) and
Dċ(t) d ∂ dxi n k o ∂
0= = ẋk (t) k |c(t) + ẋ j (t) i j (x(t)) |c(t)
dt dt ∂x dt ∂xk
n o ∂
= ( ẍk (t) + ẋi (t) ẋ j (t) ikj (c(t))) k |c(t) .
∂x
n o
⇒ ẍk (t) + ikj (c(t)) ẋi (t) ẋ j (t) = 0, k = 1, · · · , n.
Recall the geodesic equation of a Riemannian manifold (M, g) are
ẍk (t) + Γkij (c(t)) ẋi (t) ẋ j (t), k = 1, · · · , n.
We hope to find a connection, under which a geodesic is a curve whose velocity vector
field is parallel along it. That is, we are looking for a connection ∇, s.t.
n o 1 kl
k
ij = g (gl j,i + gil, j − gi j,l ).
2
From this aim, we see the connection has to be “compatible” with the Riemannian
metric.
Recall that along a geodesic γ, we have hγ̇, gammai
˙ g ≡ const. It is natural to require
g, as a (0,2)-tensor, is parallel w.r.t ∇. i.e.
∇X g(Y, Z) = X(g(Y, Z)) − g(∇X Y, Z) − g(Y, ∇X Z) = 0, ∀X, Y, Z ∈ Γ(T M).
Definition 3.3. We say ∇ is compatible with g if the Riemannian metric g is paralell.

In other words, ∇ is conpatible with g if for all X, Y, Z ∈ Γ(T M),
X(g(Y, Z)) = g(∇X Y, Z) + g(Y, ∇X Z).

n o
k
Let us calculate ij of such connections.
∂ ∂ ∂ ∂ ∂ ∂ ∂
gi j,k = (g( i , j )) = g(∇ ∂k i , j ) + g( k , ∇ ∂k j )
∂x k ∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x
n o ∂ ∂ ∂ nlo ∂
= g( ki
l , ) + g( i , k j ).
∂xl ∂x j ∂x ∂xl
That is
n o n o
gi j,k = gl j l
ki + gil l
kj (1)
Permutation indius, we obtain

n o n o
gki, j = gli l
jk + gkl l
ji (2)
n o n o
g jk,i = glk l
jk + g jl l
ik (3)
(1),(2),(3) give us
n o n o n o n o n o n o
gi j,k + gki, j − g jk,i = g jl ( kil − ikl ) + gkl ( jil − ilj ) + gil ( klj + jkl )
Now if we further have the symmetry

n o n o
l
ki = l
ik , ∀i, l, k (3.4.1)
then
n o
gi j,k + gki, j − g jk,i = 2gil l
kj
1 n o
⇒ g pi (gi j,k + gki, j − g jk,i ) = 2g pi gil klj
2
npo 1 p
k j = g (gi j,k + gki, j − g jk,i = Γk j .)
pi
2
Then we obtain the christoffel symbols!!(That is, under such connections, a geodesic
is a curve whose velocity v.f. is parallel)
Express the condition (3.4.1) in global terms:
∂ ∂
∇ ∂ − ∇ ∂i k . (3.4.2)
∂xk ∂xi ∂x ∂x
For X, Y ∈ Γ(T M), only ∇X Y − ∇Y X is not a tensor. The global expression of LHS of
(3.4.2) is as follows. For X, Y ∈ Γ(T M)
T (X, Y) := ∇X Y − ∇Y X − [X, Y].
Proposition 3.8. T is a (1,2)-tensor.

Proof. T gives the multilinear map
T : Γ(T M) × Γ(T M) → Γ(T M), (X, Y) 7→ T (X, Y).
Moreover T ( f X, Y) = T (X, f Y) = f T (X, Y).

Example 3.1.
T ( f X, Y) = ∇ f X Y − ∇Y ( f X) − [ f X, Y]
= f ∇X Y − Y( f )X − f ∇Y X − f XY + Y( f )X + f Y X
= f T (X, Y)
Hence T is a tensor. It is a (1,2)-tensor in the sense. T (ω, X, Y) = ω(T (X, Y)).
Definition 3.4. (Torsion free) We call T the torsion tensor of ∇. If T = 0, we call ∇

torsion free(or symmetric) connection.
So, our calculations tell us: A torsion free connection ∇ which is compatible with
g has in each coordinate neighborhood.
n o
l
jk = Γijk .
Definition 3.5. A connection ∇ on (M, g) is called a Levi-Civita connection(also called

a Riemannian connection), if it is torsion free, and it is compatible with g.
In this language, our previous calculations tell that if a Levi-Civita connection ex-
ists on (M, g), it is uniquely determined by the Christoffel symbols.
Conversly, we can define a connection ∇ as follows: in each coordinate neighbor-
hood (U, x),
∂ ∂Y k ∂
∇X Y := ∇X i ∂ (Y j ) := (X i i + X i Y j Γkij ) k .
∂xi ∂x j ∂X ∂x
We can check this is well-defined, and ∇ is torsion free and is compatible with g. This
shows the existence of a Levi-Civita connections on (M, g).
Actually, we prove the following important result.
Theorem 3.1. (The fundamental theorem of Riemannian geometry) On any Rieman-

nian manifold (M, g), there exists a unique Levi-Civita connection,
Remark 3.9. This is a remarkable point to note the following observation: On a

smooth manifold, once we fix a Riemannian metric g, then we get:
a canonical distance function
a canonical measure
a canonical affine connection
We in fact already show a proof via local coordinate calculations for Theorem 3.4.1.
We provide a coordinate free proof below.
Assume the Levi-Civita connection ∇ exist, then we calculate for all X, Y, Z ∈
Γ(T M),
∇g=0
g(∇X Y, Z) : = h∇X Y, Zi = X(hY, Zi) − hY, ∇X Zi (3.4.3)
torsion f ree
= X(hY, Zi) − hY, ∇Z X + [X, Z]i
= X(hY, Zi) − hY, ∇Z Xi − hY, [X, Z]i
∇g=0
= X(hY, Zi) − (ZhY, Zi − h∇Z Y, Xi) − hY, [X, Z]i
= X(hY, Zi) − Z(hY, Xi) + h∇Z Y, Xi − hY, [X, Z]i
torsion− f ree
= X(hY, Zi) − Z(hY, Xi) + h∇Y Z, Xi + h[Z, Y], Xi − hY, [X, Z]i
∇g=0
= X(hY, Zi) − Z(hY, Xi) + Y(hZ, Xi) − hZ, ∇Y Xi + h[Z, Y], Xi − hY, [X, Z]i
torsion f ree
= X(hY, Zi) − Z(hY, Xi) + Y(hZ, Xi) − hZ, ∇X Yi − hZ, [Y, X]i + h[Z, Y], Xi − hY, [X, Z]i.
⇒ 2h∇X Y, Zi = XhY, Zi + YhZ, Xi − ZhX, Yi − hrX, [Y, Z]i + hY, [Z, X]i + hZ, [X, Y]i
(3.4.4)
⇒
2h∇X Y, Zi = XhY, Zi + YhZ, Xi − ZhX, Yi − hX, [Y, Z]i + hY, [Z, X]i + hZ, [X, Y]i
The RHS is determined by the metric g. So the uniqueness is proved.

For existence, check the ∇X Y defined by (3.4.3) satisfies all conditions of Levi-
Civita connections.
Remark 3.10. The formula (3.4.3) is called the Koszul formula. In local coordinate
(U, x), let X, Y, Z be ∂x∂ i , ∂x∂ j , ∂x∂ k , we will derive the formula for Christoffel symbols Γijk .
Sometimes, using (3.4.3) is more important than using Γkij . If in an open subset U ⊂
M, there exists an orthonormal frame field E1 , · · · , En ,(i.e. hEi , E j i(p) = δi j , ∀ p ∈ U),
(3.4.3) gives
2h∇Ei E j , Ek i = −hEi , [E j , Ek ]i + hE j , [Ek , Ei ]i + hEk , [Ei , E j ]i.
Exercise 3.4. Suppose we know the following fact: There exists three vector field on
S 3 ⊂ R4 , i, j, k, which are linearly independent at any point of S 3 , such that
[i, j] = k, [ j, k] = i, [k, i] = j.
Assign to S 3 a Riemannian metric g s.t. i, j, k are orthonormal at any point of S 3 .

Calculate the Levi-Civita connection ∇ of g on S 3 .
Next, we give more geometric interpretations for the two properties of the Levi-
Civita connection.
(a) ∇ is compatible with the metric.
(b) ∇ is torsion free.
Proposition 3.9. (geometric meaning of (a)) Let M be a smooth manifold with an affine
connection ∇. Then ∇ is conpatible with the g iff any parallel transport is an isometry.
Proof. ⇒: Let c : [a, b] → M be a smooth curve with p = c(a). The parallel transport
Pc,a,t : T c(a) M → T c(t) M, t ∈ [a, b]
is an isomorphism.
⇐: i.e. any parallel transport is an isometry⇒ ∇ is compatible with g.
For any X, Y, Z ∈ Γ(T M). Look at X(p), X(hY, Zi) = X(p)hY, Zi. Let c : [0, 1] → M
with c(0) = p, ċ(0) = X(p). We have XhY, Zi = dtd |t=0 hY(c(t)), Z(c(t))i.
Let {E1 , · · · , En } is an orthonormal basis of T p M, and {E1 (t), · · · , En (t)} is given by
Ei (t) = Pc,t Ei . Since Pc,t is isomotry, {Ei (t)} is orthonormal(∀t).
⇒ hY(c(t)), Z(c(t))i = hY i (t)Ei (t), Z i (t)Ei (t)i = Y i (t)Z j (t)δi j = Y i (t)Z i (t).
P
i
X d X dY i X dZ i
⇒ XhY, Zi = |t=0 (Y i (t)Z i (t)) = (0)Z i (0) + Y i (0) (0)
i
dt i
dt i
dt
DY DZ
=h |t=0 , Zi + hY, |t=0 i
dt dt
=h∇X(p) Y, Zi + hY, ∇X(p) Zi
This shows ∇ is compatible with g
Let Va , Wa ∈ T c(a) M, and Vt := Pc,a,t Va , Wt := Pc,a,t Wa . Then Vt , Wt are two C ∞

vector fields along c.
If Vt , Wt can be extended to two C ∞ vector fields on M, we have
dt =0
dc
metric compatibility
∇ dc hVt , Wt i = h∇ dc Vt , Wt i + hVt , ∇ dc Wt i = 0.
dt dt dt
That is Pc,a,t preserves the norms of vectors and angles between vectors. ⇒ Pc,a,t is an
isometry.
In general, we have to use the following property of induced connection:
d DVt DWt
hVt , Wt i = h , Wt i + hVt , i. (3.4.5)
dt dt dt
Proof. In a coordinate neighborhood (U, x1 , · · · , xn ). c(t) := (x1 (t), · · · , xn (t)), Vt :=

V i (t) ∂x∂ i |c(t) , Wt = W i (t) ∂x∂ i |c(t) .
We calculate
d
hVt , Wt i
dt
d ∂ ∂
= (V i (t)W j (t)h i |c(t) , j |c(t) i)
dt ∂x ∂x
d ∂ ∂ d ∂ ∂
= (V i (t)W j (t))h i |c(t) , j |c(t) i + V i (t)W j (t) h i |c(t) , j |c(t) i
dt ∂x ∂x dt ∂x ∂x
∂ ∂ d ∂ ∂
= (V̇ i (t)W j (t) + V i (t)Ẇ j (t))h i |c(t) , j |c(t) i + V i (t)W j (t)dc( )h i |c(t) , j |c(t) i
∂x ∂x dt ∂x ∂x
∂ ∂
= (V̇ i (t)W j (t) + V i (t)Ẇ j (t))h i |c(t) , j |c(t) i
∂x ∂x
∂ ∂ ∂ ∂
+ V (t)W (t)(h∇dc( d ) i , j i + h i |c(t) , ∇dc( d ) j |c(t) i)
i j
dt ∂x ∂x ∂x dt ∂x
∂ ∂ ∂
= hV̇ (t) i |c(t) + V (t)∇dc( d ) i , W (t) j |c(t) i
i i j
∂x dt ∂x ∂x
∂ ∂ ∂
+ hV (t) i |c(t) , Ẇ (t) j |c(t) + W (t)∇dc( d ) j i
i j j
∂x ∂x dt ∂x
DVt DWt
=h , Wt i + hVt , i.
dt dt
∇ of ∇ com-
Remark 1: In fact (3.4.5) is a general property of the induced connection e
patible with g. Let ϕ : N → M be a C ∞ map, u ∈ T x N, V, W are two smooth vector
fielda along ϕ, then
∇u V.Wi + hV, e
he ∇u Wi = uhV, Wi.
Proposition 3.10. (geometric meaning of (b)) Let ∇ be a torsion-free connection of M.

Let s : R2 → M be a C ∞ map.(a “parametrized surface” in M. Let V(x, y) ∈ T s(x,y) M
∂ ∂s ∂ ∂s
be a vector field along s. For convinience, let us denote ds( ∂x ) := ∂x , ds( ∂y ) := ∂y .
∇,
Then for the induced connection e
DV
∇ ∂ V(x, y) = (
e )(x,y)
∂x ∂x
can be considered as the covariant derivative along c(t) := s(t, y) of the vector field
t 7→ V(t, y) along c, evalued at t = x. Similarly, we have e
∇ ∂ V = DV
∂y . Then, we have
∂y
D ∂s D ∂s
= (3.4.6)
∂x ∂y ∂y ∂x
∂
∇ ∂ ds( ∂y
Remark 3.11. In symbols of induced connection, (3.4.5) can be written as e )=
∂x
∂ ∂ ∂
∇ ∂ ds( ). In case ds( ), ds( ) are both vector fields on M, (e.g. when s is an em-
e
∂x ∂x ∂y
∂y
bedding), is equivalent to say
∂ ∂
∇ds( ∂ ) ds( ) = ∇ds( ∂ ) ds( ).
∂x ∂y ∂y ∂x
This equivalent to the torsion free property since
∂ ∂ ∂ ∂
[ds( ), ds( )] = ds([ , ]) = 0.
∂x ∂y ∂x ∂y
Proof. Express both sides in a coordinate neighborhood as (U, x1 , · · · , xn ), s = (s1 , · · · , sn )
∂s ∂ ∂si ∂ ∂s ∂ ∂si ∂
= ds( ) , = ds( ) = .
∂y ∂y ∂y ∂si ∂x ∂x ∂x ∂si
⇒
D ∂s ∂2 si ∂ ∂si ∂
= + ∇ ∂
∂x ∂y ∂x∂y ∂s i ∂y ds( ∂x ) ∂si
∂si ∂ ∂si ∂s j ∂
= + ∇∂
∂x∂y ∂s i ∂y ∂x ∂s j ∂si
Similarly,
D ∂s ∂2 s i ∂si ∂s j ∂
= + ∇ ∂i j .
∂y ∂x ∂y∂x ∂y ∂x ∂s ∂s
Then the proposition follows from the fact that
∂2 s i ∂ ∂ ∂ ∂
and ∇ ∂ j i − ∇ ∂i j = [ i , j ] = 0
∂y∂x ∂s ∂s ∂s ∂s ∂s ∂s

Remark 2: (3.4.6) is also a general property of an induced connection e ∇ of a torsion

free connection ∇. Let ϕ : N → M be a C ∞ map, X, Y be two C ∞ vector fields on N.
Then dϕ(X), dϕ(Y) are C ∞ vector fields along ϕ, then e ∇Y dϕ(X) = dϕ([X, Y]).
∇X dϕ(Y)− e
By Remark 1 and the above Remark 2, when doing calculations, we can assume
∇ and proceed formally as if vector fields along ϕ were actually defined
the notation e
on M.
Exercise 3.5. (Variation of the energy functional: A coordinate free calculation).

Let γ : [a, b] → M be a C ∞ curve, and α : (−, ) × [a, b] → M be a variation.
Recall
1 b
Z
E(γ) := hγ̇(t), γ̇(t)idt
2 a
1 b
Z
d d
= hdγ( ), dγ( )idt.
2 a dt dt
Show that
b
∂ ∂α ∂α
Z
dE(α(s)) Dγ̇(t) Dγ̇ Dγ̇
| s=0 = − hdα( )(0, t), idt − h (0, a), (a)i + h (0, b), (b)i
ds a ∂s dt ∂s dt ∂s dt
Hint: using Proposition 3.10.
Recall a calculation in coordinates has be carried out in our disscussions in Chapter 2, Geodesics.
Remark 3.12. Recall that any Riemannian manifold M can be embedded to the stan-
∇
dard Euclidean space E isometrically. In the Euclidean space the Levi-Civita Connections e
is given by the directional derivatives. So for any X, Y ∈ Γ(T M), X, Y can also be ex-
tended to vecter fields X, Y on E(at least locally around M.) But usually ∇X Y(p) ∈
T p E is not lie in T p M any more, The orthogonal projection π : T p E → T p M gives
π(∇X Y(p)) ∈ T p M. One can check that π(∇X Y(p)) gives a Levi-Civita connection on M
w.r.t. the induced metric from E.
3.5 The First variation of Arc Length and Energy

In chapter 1, we derive the geodesic equations as the Euler-Langrange equations of
the Length and Energy functionals .via local coordinates computations. Now, with the
convenient notion of (Levi-Civita) Connections, we can carry out an easier computa-
tion(intrinstic).
Recall for any smooth curve c : [a, b] → M, we have
Z b p
L(c) : = hċ(t), ċ(t)idt
a
r
∂ ∂
Z b
= hdc( ).dc( )idt
a ∂t ∂t
1 b ∂ ∂
Z
E(c) = hdc( ), dc( )idt.
2 a ∂t ∂t
Definition 3.6. Let c : [a, b] → M be a smooth curve, ∀i0. A variation of c is a map

F : [a, b] × (−, ) → M with F(t, 0) = c(t), ∀t ∈ [a, b]. The variation is called proper
if the endpoints stay fixed, i.e. F(a, s) = c(a).F(b, s) = c(b), ∀s ∈ (−, ).
3.5. THE FIRST VARIATION OF ARC LENGTH AND ENERGY 67
For simplicity, we will denote
∂F ∂ ∂F ∂
= dF( ), = dF( ).
∂s ∂s ∂t ∂t
We also denote c s (t) = F(t, s).
∂F ∂F
Definition 3.7. We call V(t) = ∂s (t, 0) = ∂s (c(t)) the variation field of f along c.(It is
a vector field along c)
Theorem 3.2. (The First Variation Formula) Let F(t, s) be a variation of a smooth
curve c. Let us write L(s) := L(c s ), E(s) := E(c s ) for simplicity. Then
Z b
d 1 d
| s=0 L(s) := L0 (0) = ( hV(t), ċ(t)i − hV(t), ∇ ∂ ċ(t)i)dt
ds a |ċ(t)| dt ∂t
(∇ ∂ ċ(t)i is oftenly written as ∇dotc ċ.)

∂t
Z b
d
| s=0 E(s) := E 0 (0) = hV(b), ċ(b)i − hV(a), ċ(a)i − −hV(t), ∇ ∂ ċ(t)idt.
ds a
∂t
Proof.
d ∂F
b
∂F
Z
d 1
L(s) = h (t, s), (t, s)i 2 dt
ds a ds ∂t ∂t
d ∂F ∂F
Z b
1
= h (t, s), (t, s)idt
a 2h
∂F ∂F 1
ds ∂t ∂t
∂t (t, s), ∂t (t, s)i
2
∂F ∂F
Z b
(3.4.5) 1
= he∇∂ (t, s), (t, s)idt
∂F ∂F 1 ∂s ∂t ∂t
a h
∂t (t, s), ∂t (t, s)i
2
∂F ∂F
Z b
(3.4.6) 1
= he∇∂ (t, s), (t, s)idt
∂F ∂F 1 ∂t ∂s ∂t
a h
∂t (t, s), ∂t (t, s)i
2
d ∂F ∂F ∂F ∂F
Z b
(3.4.5) 1
= ( h (t, s), (t, s)i − h (t, s), e
∇∂ (t, s)i)
a h
∂F ∂F 1
dt ∂t ∂t ∂t ∂t ∂t
∂t (t, s), ∂t (t, s)i
2
Z b
d 1 d
⇒ | s=0 L(s) = ( hV(t), ċ(t)i − hV(t), e ∇ ∂ ċ(t)i)dt.
ds a |ċ(t)| dt ∂t
Similarly, we obtain the formula for E 0 (0).
Observe that when c is parametrized proportionally to arc length i.e. |ċ(t)| ≡ const,
the variations of L and E leads to the same critical point.(We observed this fact using
Holder inequality in Chapter 2.)
A smooth curve c : [a, b] → M is a critical point of the energy E for all proper
∇ ∂ ċ(t)i = 0.(i.e. c is a geodesic.)
variations iff e
∂t
Note, by property of parallel transport e ∇ ∂ ċ(t)i = 0 ⇒ |ċ(t)| ≡ const ⇒ c is

∂t
parametrized proportionally to arc length.
More generally, we consider piecewise smooth curve c : [a, b] → M. That is, we

have a subdivision a = t0 < t1 < t2 < · · · < tk < tk+1 = b s.t. c is smooth on each
interval [ti , ti+1 ].
Correspondingly, we consider “piecewise smooth variations” of c, which are con-
tinuous functions F : [a, b]×(−, ) → M such that F is smooth on each [ti , ti+1 ]× ×,
and ∂F 0
∂s is well defined even at ti s.
Then, as a direct consequence of Theorem 3.5.1, we have
Corollary 3.1. Let c be a piecewise smooth curve and F be a correponding piecewise
smooth variation. Then
E 0 (0)
k Z b
d X
= | s=0 E(c s ) = hV(b), ċ(b)i − hV(a), ċ(a)i − hV(ti ), γ̇(ti+ ) − γ̇(ti− )i − hV(t), ∇ ∂ ċ(t)idt.
ds i=1 a
∂t
It turns out the first variation formulas are also very useful for non-proper varia-
tions. We discuss here Gauss’ lemma. Recall in a normal neighborhood U p of a point
p ∈ M, we can introduce a polar coordinate such that g = dr ⊗ dr + gϕϕ (r, ϕ)dϕ ⊗ dϕ.
Here the fact gr,ϕ ≡ 0 on the whole U p is also called Gauss’ lemma.
Lemma 3.2. (Gauss’ Lemma). In U p , the geodesics through p are perpendicular to
the hypersurfaces {exp p (v) : kvk = const < δ}.(Piecely, let v ∈ T p M, ρ(t) = tv is a ray
through 0 ∈ T p M. Let ω ∈ T ρ(r) (T p M) is perpendicular to ρ0 (r). Then
h(dexp p )(ρ(r))(ω), (dexp p )(ρ(r))(ρ0 (r))i (3.5.1)
)
Proof. Let v(s) : (−, ) → T p M be a curve with v(0) = rv = ρ(r), v̇(0) = ω, and
kv(s)k = r.
Then we have a variation F(t, s) = exp p (tv(s)), t ∈ [0, r], s ∈ (−, ). with F(t, 0) =
exp p (tv) = c(t). Rr
Notice that E(c s ) = 21 0 hv(s), v(s)idt = 12 r3 ≡ const.
Theorem 2⇒
Z r
0 = E (0) = hV(r), ċ(r)i − hV(0), ċ(0)i −
0
hV(t), e
∇ ∂ ċ(t)idt.
∂t
0
∂F
∇ ∂ ċ(t) = 0, and V(0) =
Since e ∂s |t=0,s=0 = 0, we conclude
∂t
hV(r), ċ(r)i = 0. (3.5.2)
Recall
∂F ∂
V(r) = |t=r,s=0 = | s=0 exp p (rv(s)) = (dexp p )(ρ(r))(ω).
∂s ∂s
ċ(r) = (dexp p )(ρ(r))(ρ0 (r)).
We see (3.4.2) implies (3.4.1).

3.6. COVARIANT DIFFERENTIATION, HESSIAN, AND LAPLACIAN 69
3.6 Covariant differentiation, Hessian, and Laplacian

Recall the covariant derivative of a (r,s)-tensor A satisfies ∇ f X+gY A = f ∇X A + g∇Y A.
That is ∇X A is linear over C ∞ functions for the arguement X. Therefore we can define
a (r,s+1)-tensor ∇A for each (r,s)-tensor A as
∇A(ω1 , · · · , ωr , X1 , · · · , X s , X) := ∇X A(ω1 , · · · , ωr , X1 , · · · , X s ), ∀ωi ∈ Γ(T ∗ M), X j , X ∈ Γ(T M).
We call A the covariant differention of A.

Portionally, consider a (0,0)-tensor, i.e. a function f . The covariant differention ∇ f
of f is given as
∀X ∈ Γ(T M) : ∇ f (X) := ∇X f = X( f ) = d f (X)
⇒ ∇ f = d f is a (0,1)-tensor.
We can then discuss iteratively
∇2 f := ∇(∇ f ), ∇3 f = ∇(∇2 f ), · · ·
generally, ∇2 A, ∇3 A, · · ·
Warning: ∇2 A(· · · , X, Y) , ∇Y ∇X A(· · · )!!
For X, Y ∈ Γ(T M), we have=
∇ f (X, Y) = ∇(∇ f )(X, Y) = ∇Y (∇ f )(X)

= Y(∇ f (X)) − (∇ f )(∇Y X)
= Y(X f ) − ∇Y X( f ).
Proposition 3.11. ??∇2 f (X, Y) = ∇2 f (X, Y) = T (X, Y)( f ).

Proof.
∇2 f (X, Y) − ∇2 f (Y, X)
=Y X f − (∇Y X) f − XY f + (∇X Y) f
=[Y, X] f − (∇Y X − ∇X Y) f
=T (X, Y) f.

That is, when the connection ∇ is torsion-free, we have
∇2 f (X, Y) = ∇2 f (Y, X), ∀X, Y ∈ Γ(T M)
i.e. ∇2 f is a symmetric (0,2)-tensor field.

We call ∇2 f the Hessian of f .
Example 3.2. On Rn , given the canonical connection, we have
 1
 x 
∂ 2
∇ f (X, Y) = (Y , · · · , Y )( i j )  ... 
2 1 n f  
∂x ∂x  n 
x
where X = X i ∂x∂ i , Y = Y j ∂x∂ j .

Since ∇2 f ( ∂x∂ i , ∂x∂ j ) = ∂
( ∂
∂xi ∂x j
f) = ∂2
∂xi ∂x j
.
The trace of the Hessian is the Laplacian

For any X ∈ Γ(T M), we have a linear map ∇X : Γ(T M) → Γ(T M), Y 7→ ∇Y X.
At a point p, ∇X = T p M → T p M is a linear transformation between two vector
space. Hence, it make sense to talk about the trace of ∇X at each p, which gives us a
function on M.
Lemma 3.3. ∀X ∈ Γ(T M), we have div(X) = tr(∇X).

√
Recall from Chapter 1, divX = √1G ∂X∂ i (X i G), G = det(gi j ) in local coordinate.
Proof. We only need to prove it at one point p ∈ M. Pick a coordinate neighborhood

p ∈ U, (U, x), we have
∇X = (∇ ∂ X)dxi
∂xi
∂X k ∂ ∂
= ⊗ dxi + X j Γkij k ⊗ dxi
∂xi ∂xk ∂x
∂X i
Therefore tr(∇X) = ∂xi
+ X j γii j .
√
proposition: Let ∇ be the Levi-Civita connection on (M, g). Then Γ jji = √1 ∂ i G.
G ∂x
(see Appendix) √ √
∂X i
+ X j γii j = ∂X + X i √1G ∂x∂ i G = √1G ∂x∂ i (X i G) = div(X).
i
∂xi ∂xi

Recall since g is non-degenerate and bilinear on T p M, we have isomorphisms be-

tween T M and T ∗ M:
b : T M → T ∗ M, X 7→ φ(X). b(X)(Y) := g(X, Y)
and
# : T ∗ M → T M, ω 7→ #(ω). g(#(ω), Y) = ω(Y).
In local coordinate, b(X i ∂i ) = gi j X i dx j , #(wi dxi ) = gi j ωi ∂x∂ j .

Now we define the trace of S as the trace of the linear map: X 7→ #S (X, ). Note
g(#S (X, ), Y) = S (X, Y).
In local coordinate, S = S i j dxi dx j
∂
#S (X, ) = #(S i j X i dx j ) = S i j X i g jk .
∂xk
Hence tr(S ) := tr(X 7→ #S (X, )) = gi j S i j = gi j S ( ∂x∂ i , ∂x∂ j ).

Let us come back to Hess f .
Lemma 3.4. ∀X, Y ∈ Γ(T M), Hess f (X, Y) = g(∇X (grad f ), Y)

3.7. APPENDIX: THE TECHNICAL LEMMA. 71
Proof.
RHS = ∇X (g(( gard) f, Y)) − g(gard f, ∇X Y)

= ∇X (Y f ) − (∇X Y) f
= X(Y f ) − (∇X Y) f = Hess f (X, Y).

Hence we have
tr(Hess f ) = tr(X 7→ ∇X (grad f ))

= tr(∇] grad f )
= div(grad f )
= ∆ f (Laplace − Beltrami operator)
Recall from Chapter 1. that
1 ∂ ik √ ∂ f
∆f = √ (g G) i .
G ∂xk ∂x
3.7 Appendix: The technical lemma.

√
Proposition 3.12. Let ∇ be the Levi-Civita connection on (M, g). Then Γiji = √1 ∂ i ( G), G :=
G ∂x
det(gi j ).
Proof. Recall
1 jk
Γ jji =g (g jk,i + gki, j − g ji,k )
2
1 ∂
= g jk i (gik )
2 ∂x
1 ∂
= tr[(grs )n×n i (g jk )n×n ]
2 ∂x
Note moreover, (grs )n×n is the inverse matrix of (g jk )n×n .
We need the following result:
Lemma: Let A = A(t) be a family of nonsingular matrices that depends smoothly on t,
then
d d
tr(A−1 A−1 ) = ln det A.
dt dt
Sketch of proof: Observe the Lemma is obvious, when A is 1×1. For a diagonal matrix
A.
 −1  0
A1 (t)  A1 (t)


= tr[
 ..   .. 
 .   . ]

A−1 A0n (t)

n (t)
 −1
A1 (t)A01 (t)

 X n
= tr 
 .. 
= A−1 0
. 
 i (t)Ai (t)

A−1 0 i=1

n (t)An (t)
A1 (t)
 
n n 
d X d Y d ..
= ln Ai (t) = Ai (t) =
 
dt i=1
ln
dt i=1 dt
ln det 
 . 

An (t)
Recall both trace and det are invariants under similar transformation. For diago-
nalizable A, we have A = P−1 DP−1 . Then det A = det D and tr(A−1 dtd A) = tr(D−1 dtd D)
since
tr(P−1 DP(P−1 DP)0 ) = tr(P−1 D−1 P(P−1 )0 DP + P−1 D−1 PP−1 D0 P + P−1 D−1 PP−1 DP0 )
= tr(P(P−1 )0 + P−1 P0 ) + tr(D−1 D0 ) = tr(D−1 D0 )
Hence Lemma is true for diagonalizable metrices.

By standard permutation trick, one can prove Lemma in its full generality.
Let us continue:
1 ∂
Γiji = tr[(grs )n×n i (g jk )n×n ]
2 ∂x
1 ∂ 1 ∂
= ln det(g jk )n×n = (ln G)
2 ∂xi 2 ∂xi
∂ √ 1 ∂ √
= i (ln G) = √ ( G)
∂x G ∂xi
.
Chapter 4
Curvatures
The Riemannian curvature tensor was introduced by Riemann in his 1854 lecture as a
natural invariant for what is called the equivalence problem in Riemannian geometry.
This problem, comes out of the problem one faces when writing the same metric in two
different coordinates. Namely, how is one to know that they are the same or equivalent.
The idea is to find invariants of the metric that can be computed in coordinates and
then try to show that two metrics are equivalent if their invariant expressions are equal.
This problem was furthur classified by Christoffel.
Our previous discussions on geodesics and connections follow roughly the histori-
cal development. However, we will have a discussion on the curvature tensor different
from it’s historical development.(Notice that the idea of a connection postdates Rie-
mann’s introduction of the curvature tensor).
On retrospection of our previous discussions, roughly speaking ,”the first varia-
tion (of length or energy) gives the connection”. In this chapter, we will see, roughly
speaking, ”the second variation gives the curvation”!
4.1 The Second Variation

We already know that a geodesic is not necessarily minimizing. Is a geodesic a ”local
minima”? One way to explore this properly is to calculate the second variation of
length or energy. (Recall from 3.5 and Exercise 6.2 ,among curves ∈ C p,q (piecewise
smooth curves from p to q), a geodesic is characterized as the critical point of the
energy functional).
Let γ : [a, b] → M be a normal geodesic, i.e. γ̇(t) ≡ 1. We consider a 2-parameter
variation F of γ . That is , a smooth map
F : [a, b] × (−, ) × (−δ, δ) → M
such that F(t, 0, 0) = γ(t)

Let E(v, w) be the energy of the curve γv,w (t) := F(t, v, w). And
∂F ∂F
V(t) = (t, 0, 0), W(t) = (t, 0, 0)
∂v ∂w
73
74 CHAPTER 4. CURVATURES
are the two corresponding variation fields.

Recall
∂E 1 b ∂ ∂F ∂F
Z
(v, w) = h (t, v, w), (t, v, w)idt
∂w 2 a ∂w ∂t ∂t
∂F ∂F
Z b
= he
5∂ , idt (compatibility)
∂v ∂t ∂t
a
∂F ∂F
Z b
= he
5∂ , idt (torsion − f ree)
∂t ∂v ∂t
a
Then
∂2 b
∂ ∂F ∂F
Z
E(v, w) = he
5∂ , idt
∂w∂v a ∂w ∂t ∂v ∂t
b
∂F ∂F ∂F ∂F
Z !
= h∇ ∂ ∇ ∂ , i + h∇ ∂ ,∇ ∂ i dt (compatibility)
∂w ∂t ∂v ∂t ∂t ∂v ∂w ∂t
a
b
∂F ∂F ∂F ∂F
Z !
= h∇ ∂ ∇ ∂ , i + h∇ ∂ ,∇∂ i dt (torsion − f ree)
∂w ∂t ∂v ∂t ∂t ∂v ∂t ∂w
a
Restrcting the above equation to the curve γ ,i.e. to the case where v = w = 0:
∂2
Z b
v=w=0 E(v, w) = h∇ ∂ ∇ ∂ V(t), γ̇(t)i + h∇ ∂ V(t), ∇ ∂ W(t)i dt
∂w∂v a
∂w ∂t ∂t ∂t
Now ,we hope to make use of the fact that γ is geodesic ,i.e.,∇ ∂ γ̇(t) = 0. For this
∂t
purpose,we hope to interchange the order of the covariant derivative ∇ ∂ , ∇ ∂ . Hence
∂w ∂t
we proceed:
∂2
Z b
v=w=0 E(v, w) = h∇ ∂ ∇ ∂ V(t) − ∇ ∂ ∇ ∂ V(t), γ̇(t)idt
∂w∂v a
∂w ∂t ∂t ∂w
Z b
+ (h∇ ∂ ∇ ∂ V(t), γ̇(t)i + h∇ ∂ V(t), ∇ ∂ W(t)i)dt
∂t ∂w ∂t ∂t
a
=: I + II
Then the second term becomes
b
∂
Z !
II = h∇ ∂ V(t), γ̇(t)i − h∇ ∂ V(t), ∇ ∂ γ̇(t)i + h∇ ∂ V, ∇ ∂ Wi dt
a ∂t ∂w ∂w ∂t ∂t ∂t
b Z b
= h∇ ∂ V(t), γ̇(t)i + h∇ ∂ V(t), ∇ ∂ W(t)idt.
∂w a ∂t ∂t
a
Therefore,we obtain the following Second Variation Formula:
∂2 b Z b
E(v, w) = h∇ ∂ V(t), γ̇(t)i + h∇ ∂ V(t), ∇ ∂ W(t)i + h∇ ∂ ∇ ∂ V(t) − ∇ ∂ ∇ ∂ V(t), γ̇(t)i dt
∂w∂v ∂w ∂t ∂t ∂w ∂t ∂t ∂w
v=w=0 a a
(SVF)
4.1. THE SECOND VARIATION 75
Remark 4.1. Usually ,we will suppose the notation e 5 and proceed formula , as b if
∂2
vectors along γ were actually defined on M and write ∂w∂v |v=w=0 E(v, w) = h∇w V, γ̇i +
Rb a
a
h∇ j V, ∇ j Wi + h∇ ∇
w j V − ∇ ∇
j w V, γ̇i dt
Remark 4.2. In particular ,(SVF) tells
d2 b Z b
E(v) =: E 00
(0) = h∇ ∂ V(t), γ̇(t)i +
h∇ ∂ V(t), ∇ ∂ V(t)i+h∇ ∂ ∇ ∂ V(t)−∇ ∂ ∇ ∂ V(t), γ̇(t)idt
dv2 v=0 ∂v ∂t ∂t ∂v ∂t ∂t ∂v
a a
For proper variations,(i.e. V(a) = V(b) = 0), or generally,when∇ ∂ V(t) = 0 at t = a

∂v
and t = b ,we have
Z b
E”(0) = h∇ ∂ V(t), ∇ ∂ V(t)i + h∇ ∂ ∇ ∂ V(t) − ∇ ∂ ∇ ∂ V(t), γ̇(t)i dt
∂t ∂t ∂v ∂t ∂t ∂v
a
Now we see the sign of the second term is very importan to decide the sign of E”(0),
which is useful to decide whther a geodesic has a locally minimal energy functional
(for curves parametrized proportionally to arclength, equicalent to a locally minimal
arc length).
In particular ,if the second term vanishes,(or ≥ 0), we have E”(0) ≥ 0 and the local
minimum is guaranted.
In Rn (a flat case), any geodesic is minimizing. From that sense , te term
h∇ ∂ ∇ ∂ V(t) − ∇ ∂ ∇ ∂ V(t), γ̇(t)i

∂v ∂t ∂t ∂v
play the role of ”curvature”. b

Consider variations with the property h∇ ∂ V(t), γ̇(t)i = 0 (e.g., proper variations),
∂w a
we have
∂2
Z b
v=w=0 E(v, w) = h∇ ∂ V(t), ∇ ∂ Wi + h∇ ∂ ∇ ∂ V(t) − ∇ ∂ ∇ ∂ V(t), γ̇(t)i dt
∂w∂v a
∂t ∂t ∂w ∂t ∂t ∂w
:= I(V(t), W(t))
We will see this quality plays a central role in our subsequent discussions about curvature-
related geometries.
The second term in I(V, W) suggests to define for X, Y, Z ∈ Γ(T M),
R̄(X, Y)Z := ∇X ∇Y Z − ∇Y ∇X Z ∈ Γ(T M)
But
R̄(X, f Y)Z = ∇X ( f ∇Y Z) − f ∇Y ∇X Z
= X( f )∇Y Z + f [∇X ∇Y Z − ∇Y ∇X Z]
i.e. R̄ is not a tensor!!

We define for X, Y, Z ∈ Γ(T M), R(X, Y)Z := ∇X ∇Y Z − ∇Y ∇X Z − ∇[X,Y]Z It gives a
multilinear map
R : Γ(T M) × Γ(T M) × Γ(T M) → Γ(T M)
Proposition 4.1. R is a (1,3) tensor.

Proof. Notice that this time
R(X, f Y)Z = ∇X ( f (∇Y Z)) − f ∇Y ∇X Z − ∇[X, f Y] Z

= X( f )∇Y Z + f (∇X ∇Y Z − ∇Y ∇X Z) − ∇X( f )Y+ f [X,Y] Z
= X( f )∇Y Z + f (∇X ∇Y Z − ∇Y ∇X Z) − X( f )∇Y Z − f ∇[X,Y] Z
= f R(X, Y)Z
One can furthur check that
R( f X, Y)Z = R(X, f Y)Z = R(X, Y)( f Z) = f R(X, Y)Z
Hence R is a tensor . We say it is a (1,3) tensor . We actually mean R(W, X, Y, Z) :=

W(R(X, Y)Z).

We will call R the curvature tensor.
Remark 4.3. (1) The curvature tensor is well-defined for any affine connection on M
(2) Notice that X,Y appear skew-symmetrically in R(X, Y)Z while Z plays its own
role on top of the line ,hence we use the usual notation R(X, Y)Z instead of R(X, Y, Z).
(3) Some textbooks adopt a different sign in the definition of R. One should always
first check the author’s notation for curvature tensor when reading works on Rieman-
nian geometry , unfortunately.
(4)(Locality): At p ∈ M, R(X, Y)Z(p) only depends on X(p), Y(p), Z(p) ∈ T p M.
This is due to the tensorial property
∂ ∂ ∂
R(X, Y)Z = X i Y j Z k R( , )
∂X i ∂X j ∂X k
Now let’s come back to the SVF:
Z b
I(V, W) == (h∇ ∂ V(t), ∇ ∂ Wi + h∇ ∂ ∇ ∂ V(t) − ∇ ∂ ∇ ∂ V(t), γ̇(t)i)dt
∂t ∂t ∂w ∂t ∂t ∂w
a
Proposition 4.2. Let s : R2 → M be a paramatrized surface , and V a C ∞ vector field

along s .Then
∂s ∂s
!
D D D D
V− V=R , V (*)
∂x ∂y ∂y ∂x ∂x ∂y
∂s ∂s
(or,in another notation ,∇ ∂ ∇ ∂ V − ∇ ∂ ∇ ∂ V = R( ∂x , ∂y )V)
∂x ∂y ∂y ∂x
Sketch of proof: First by the locality remark above ,at each point s(x, y) ∈ M, the
RHS is well-defined since
∂s ∂ ∂s
!
= ds , , V ∈ T s(x,y) M.
∂x ∂x ∂y
Then (*) can be proved by computing in a coordinate neighborhood.
4.2. PROPERTIES OF CURVATURE TENSOR : GEOMETRIC MEANING AND SYMMETRIES77
Definition of both sides are clear: pick (u, x1 , x2 ....xn ), s(x, y) = (s1 (x, y), ..., sn (x, y))
∂s ∂s ∂si ∂s j k ∂ ∂ ∂
! !
R , V= V R , j
∂x ∂y ∂x ∂y ∂x ∂x ∂xk
i
while in LHS
D ∂V i (x, y) ∂ ∂
!
D D
V= + V i ∇ds( ∂ ) k
∂x ∂y ∂x ∂y ∂xk ∂y ∂x
D ∂V i ∂ i ∂s
j
∂
!
= +V ∇∂
∂x ∂y ∂xi ∂y ∂x j ∂xk
∂2 V i ∂ ∂V i ∂s j ∂ ∂ i ∂s
j
∂ i ∂s ∂s
j l
∂
!
= + ∇ ∂ + V ∇ ∂ + V ∇ ∂l ∇ ∂ j k
∂x∂y ∂x i ∂y ∂x ∂x j ∂x i ∂x ∂y ∂x j ∂x k ∂y ∂x ∂x ∂x ∂x
So the meaning of each term is clear. The equality (*) then just follows from direct
computation
In particular , 6.4.2 tells
∇ ∂ ∇ ∂ V(t) − ∇ ∂ ∇ ∂ V(t) = R(w(t), γ̇(t))V(t)

∂w ∂t ∂t ∂w
and hence
Z b
I(V, W) = h∇ ∂ V(t), ∇ ∂ W(t)i + hR(W(t), γ̇(t))V(t), γ̇(t)idt
∂t ∂t
a
Z b
= h∇γ̇ (t)V(t), ∇γ̇(t) W(t)i + hR(W, γ̇), γ̇i(t)dt
a
If we furthur denote T = γ̇, then we have

Z b
I(V, W) = (h∇T V, ∇T Wi + hR(W, T )V, T i) dt
a
Rb
In particular ,I(V, V) = a
(h∇T V, ∇T Vi + hR(V, T )V, T i)dt
4.2 Properties of Curvature tensor : Geometric mean-

ing and Symmetries
Curvature tensor measures the non-commutatively of the convariant derivatives.
4.2.1 Ricci Identity

Recall for f ∈ C ∞ (M), its Hessian ∇2 f is symmetric (for torsion free connection )
∇2 f (X, Y) = ∇2 f (Y, X)
Nr,s
For any tensor field Φ ∈ Γ( T M), we can define
R(X, Y)Φ = ∇X ∇Y Φ − ∇Y ∇X Φ − ∇[X,Y] Φ

It is obvious that
R(X, Y) f = X(Y f ) − Y(X f ) − [X, Y] f = 0
So we can write (for torsion-free connection ∇)
∇2 f (X, Y) − ∇2 f (Y, X) = R(Y, X) f = −R(X, Y) f
We can furthur check the case Φ = Z ∈ Γ(T M)
∇2 Z(X, Y) = ∇Y (∇Z)X = ∇Y (∇X Z) − ∇Z(∇Y X) = ∇Y ∇X Z − ∇∇Y X Z
Hence
∇2 Z(X, Y) − ∇2 Z(Y, X) = ∇Y ∇X Z − ∇X ∇Y Z − ∇∇Y X Z + ∇∇X Y Z

= ∇Y ∇X Z − ∇X ∇Y Z − ∇[Y,X] Z
= R(Y, X)Z = −R(X, Y)Z
It is direct to check the gneral case.

Nr,s
Proposition 4.3. (Ricci Identity) ∀X, Y ∈ Γ(T M), Φ ∈ Γ( T M) we have
∇2 Φ(..., X, Y) − ∇2 Φ(..., Y, X) = R(Y, X)Φ(...) = −R(X, Y)Φ(...)
Remark 4.4. In Euclidean space Rn , pick the directional derivative as the covariant
derivative .One easily check that R(X, Y) vanishes. In Rn , we can interchange the order
of taking derivatives freely . However this is not true anymore when R is nontrivial.
4.2.2 Geometric meaning :A test case [Spivak 2.Chap6.Thm 10]

The Ricci identity from last subsection provides an explanation of the curvature tensor
from a viewpoint of analysis : it is a term measuring the non-commutativity of taking
cavariant derivatives.
We persue for a geometric meaning of the curvature tensor. Back to Riemann’s
”equicalence problem”: If we know a Riemannian metric g = gi j dxi ⊗ dx j gives R = 0,
is there a coordinate change x → y s.t. g = i dyi ⊗ dyi ? (Or ,does R = 0 implies
P
locally isometric to (Rn , h, i)?)
The answer is yes!
Theorem 4.1. Let (M, g) be an n-dim Riemannian manifold for which the curvature
tensor R (for the Levi-Civita connection) is 0. Then M is locally isometric to Rn with
its canonical Riemannian metric .
Proof. Let p ∈ M, pick a coordinate neighborhood (U, y1 ...., yn ) Let g = gi j dyi ⊗ dy j

To prove this theorem, it is equivalent to show there exist open setV ⊂ U, and a
coordinate change x : V → Rn , s.t. g = i dxi ⊗ dxi
P
So without loss of generality, we can assume we are in Rn ,with y1 ..., yn the standard
coordinate system ,with a metric g = gi j dyi ⊗ dy j and ∇ be the corresponding Levi-
Civita connection
Step 1:We claim that we can find vector fields X, with arbitrary initial values X(0) ∈
T 0 Rn , satisfying
∇ ∂i X = 0, f or all i
∂y
and hence ∇Z X = 0 for all Z.

To do this ,we first choose the curve y 7→ (y, 0, ...0)
Then for each fixed y1 ,we choose the curve y 7→ (y1 , y, 0..., 0) with X(y1 , 0..., 0) as
the initial value, we obtain X(y1 , y, 0..., 0) via parallel transport along y 7→ (y1 , y, 0..., 0)
Now we have a vector field X defined on the surface
s(y1 , y2 ) = (y1 , y2 , 0, ..., 0)
∇
By construction ,we have e ∂ X = 0 along s
∂y2
∇
while e ∂ X = 0 along {s(y, 0)}
∂y1
∇ ∂1 X vanish along s?
Question: Does e
∂y
Now we use
∂s ∂s
!
∇ ∂1 ∇ ∂2 X − ∇ ∂2 ∇ ∂1 X = R , X = 0 ⇐⇒ e ∇ ∂1 X) = 0
∇ ∂2 (e (*)
∂y ∂y ∂y ∂y ∂y1 ∂y2 ∂y ∂y

∇ ∂1 X
Since e = 0,i.e., e
∇ ∂1 X is parallel along {s(y, 0)},
∂y y2 =0 ∂y
∇
We have by (*) e ∂ X = 0 along s .
∂y1
We can continue in this way to obtain the desired X . This proves the claim.
Now at 0,we can choose X1(0) , ..., Xn(0) as orthonormal w.r.t the metric g . And con-
struct X1 , ..., Xn in the above way. By property of parallel transport, they are orthonor-
mal everywhere.
Step 2: Since ∇ is torsion free , we have
0 = ∇Xi X j − ∇X j Xi − [Xi , X j ]
By construction ,∇Xi X j = ∇X j Xi = 0, therefore , we obtain [Xi , X j ] = 0, ∀i, j

This means that thereis a coordinate system x1 , ..., xn with Xi = ∂x∂ i .(Frobenius
theorem in ”differential manifold” course . [xi , x j ] = 0 means intergrability.)
Step 3: Since ∂x∂1 , ..., ∂x∂n are orthonormal everywhere , we have
X
g= dxi ⊗ dxi
i
Remark 4.5. In some sense , the flatness R = 0 is a kind of integrability condition.

It is not true that R = 0 implies M is globally isometric to Rn .
Example 4.1. S 1 × R1 cylinder is the product of a unit circle S 1 and R1 .

∀p = (x, y, z) ∈ S 1 × R1 , we can write x = cos θ, y = sin θ, z = z 0 ≤ θh2π
Therefore in the coordinate neighborhood {(θ, z)|0 < θ < 2π, z ∈ R1 }
We have the induced metric g = dθ ⊗ dθ + dz ⊗ dz
Hence cylinder has R = 0
Corollary 4.1. If we find n everywhere linearly independent vector fields X1 , ..., Xn ,which
are parallel (i.e. ∇Z Xi = 0, ∀Z)
then the manifold is flat.
Parallel translation of a vector along a closed curve generally bring it back to a

different vector.
4.2.3 Bianchi Identities

Before continuing the duscussions of the geometric aspect of the curvature tensor , we
prepare symmetry properties of the curvature tensor in this section . We will work on
a smooth manifold with a symmetric (i.e. torsion-free) connection ∇.
Proposition 4.4. The curvature tensor satisfies the following identities : For any
X, Y, Z, W ∈ Γ(T M),
(1)R(X, Y)Z = −R(Y, X)Z
(2)R(X, Y)Z + R(Y, Z)X + R(Z, X)Y = 0 (The first Bianchi identity)
(3)(∇X R) (Y, Z)W + (∇Y R) (Z, X)W + (∇Z R) (X, Y)W = 0 (The Second Bianchi iden-
tity)
Remark 4.6. Let T be any mapping with 3 vector field variables and values that canbe
added. Summing over cyclic permutations (denote by symbol ”S”)of the cariables
gives us a new map.
S T (X, Y, Z) = T (X, Y, Z) + T (Y, Z, X) + T (Z, X, Y)
For example , the Jacobi identity for vector fields can be written as S [X, [Y, Z]] = 0
In this way, the First Bianchi identity is S R(X, Y)Z = 0 while the second Bianchi
identity is S (∇X R)(Y, Z)W = 0
Proof. (1) is obvious from the definition
(2):
S R(X, Y)Z = S ∇X ∇Y Z − ∇Y ∇X Z − ∇[X,Y] Z

= S (∇X ∇Y Z) − S (∇Y ∇X Z) − S ∇[X,Y] Z

= S (∇Z ∇X Y) − S (∇Z ∇Y X) − S ∇[X,Y] Z

= S (∇Z (∇X Y − ∇Y X)) − S ∇[X,Y] Z

= S (∇Z [X, Y]) − S ∇[X,Y] Z

= S ([Z, [X, Y]])

=0
(3)Denote
R(X, Y)Z = ∇X ∇Y Z − ∇Y ∇X Z − ∇[X,Y] Z
= [∇X , ∇Y ] Z − ∇[X,Y] Z
then
(∇Z R) (X, Y)W = ∇Z (R(X, Y)W) − R (∇Z X, Y) W − R(X, ∇Z Y)W − R(X, Y) (∇Z W)
= [∇Z , R(X, Y)] W − R (∇Z X, Y) W − R(X, ∇Z Y)W
Keeping in mind that we only do cyclic sums over X, Y, Z and that we have the
Jacobi identity for oprators:
S [∇X , [∇Y , ∇Z ]] = 0
We have
S (∇X R) (Y, Z) W = S [∇X , R (Y, Z)] W − S R (∇X Y, Z) W − S R (Y, ∇X Z) W
= S [∇X , [∇Y , ∇Z ]] W − S ∇X , ∇[Y,Z] W − S R (∇X Y, Z) W − S R (Y, ∇X Z) W

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

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

= −S ∇X , ∇[Y,Z] W − S R([X, Y], Z)W

= −S ∇X , ∇[Y,Z] W − S [∇[X,Y] , ∇Z ]W + S ∇[[X,Y],Z] W

= S [∇[Y,Z] , ∇X ]W − S [∇[X,Y] , ∇Z ]W
=0

In local coordinates , we write

∂ ∂ ∂ ∂
!
R , = Rkli j k .
∂xi ∂x j ∂xl ∂x
∂ ∂
=∇ ∂i ∇ ∂ j l − ∇ ∂ j ∇ ∂i l
∂x ∂x ∂x ∂x ∂x ∂x
γ ∂ ∂
! !
=∇ ∂i Γ jl γ − ∇ ∂ j Γγil γ
∂x ∂x ∂x ∂x
γ
∂Γ jl ∂ ∂
= i γ + Γγjl Γµiγ µ
∂x ∂x ∂x
∂Γγil ∂ ∂
− γ
− Γγil Γµjγ µ
∂x ∂x
j ∂x
 ∂Γ jl ∂Γil
 k k

γ k γ k 
 ∂
=  i − + Γ jl Γiγ − Γil Γ jγ  k
∂x ∂x j ∂x
That is ,
∂Γkjl ∂Γkil
Rkli j = − + Γγjl Γkiγ − Γγil Γkjγ
∂xi ∂x j
We see Rkli j = −Rklji , and Rkli j + Rkijl + Rkjli = 0
4.2.4 Riemannian curvature tensor

Now we consider a Riemannian manifold (M, g) with a Levi-Civita connection ∇ . We
can use g to convert the (1,3)-tensor R to be a (0,4)-tensor:
hR(X, Y)Z, Wig := R(W, Z, X, Y)
In local coordinates
∂ ∂ ∂ ∂ ∂ ∂ ∂ ∂
! * ! +
Rkli j =R , , , = R , ,
∂xk ∂xl ∂xi ∂x j ∂xi ∂x j ∂xl ∂xk
= gkm Rm li j
 ∂Γ jl ∂Γm
 m 
= gkm  i − ilj + Γγjl Γm − Γγ m
Γ

∂x ∂x iγ il jγ 
∂Γmjl∂ ∂gkm
gkm =
gkm Γmjl − Γmjl
∂xi ∂x i ∂xi
1 ∂ p p

= g jk,l + gkl, j − g jl,k − Γ m
gmp Γ + g kp Γ
2 ∂xi jl ik im
∂ 2
∂ gkl
2 ∂ 2 !
1 g jk g jl
= + − − gmp Γmjl Γikp − gkp Γmjl Γim
p
2 ∂xi ∂xl ∂xi ∂x j ∂xi ∂xk
∂Γm 1 ∂2 gik ∂2 gkl ∂2 gil

!
p m p
gkm ilj = + − − gmp Γm il Γ jk − gkp Γil Γ jm
∂x 2 ∂x j ∂xl ∂x j ∂xi ∂x j ∂xk
1 ∂2 g jk ∂2 gik ∂2 g jl ∂2 gil
!
m p m p
⇒ Rkli j = − − + + gmp Γ Γ − Γ Γ
2 ∂xi ∂xl ∂x j ∂xl ∂xi ∂xk ∂x j ∂xk il jk jl ik
Proposition 4.5. We have the following identities

(1) hR(X, Y)Z, Wi = − hR(Y, X)Z, Wi, i.e. Rkli j = −Rkl ji
(2)hR(X, Y)Z, Wi = − hR(X, Y)W, Zi,i,e, Rlki j = −Rlki j
(3)hR(X, Y)Z, Wi + hR(Y, Z)X, Wi + hR(Z, X)Y, Wi
(4)hR(X, Y)Z, Wi = hR(Z, W)X, Yi, i,e, Rkli j = Ri jkl
(5)∇R(W, Z, X, Y, V) + ∇R(W, Z, Y, V, X) + ∇R(W, Z, V, X, Y) = 0
Proof. (1) is obvious.

(2)can be seen from its expression in local coordinates. One can also use the com-
patibility of ∇ with g ,to derive directly
hR(X, Y)Z, Wi = −hZ, R(XhY)Wi
(3) follows directly from the First Bianchi Identity.

(5) follows from the second Bianchi Identity once we observe
∇V R(W, Z, X, Y) = h∇V R(X, Y)Z, Wi
(4) is a consequence of properties (1)–(3).

Although one can also see (4) directly from its expressions in local coordinats, it is
deserved to have a look at the proof in [Spivak 2 ,Chap 4D,Proposition 11]. A clever
diagram proof taken from Milnor’s Morse Theory book is presented there.

There are interesting consequence derived from these symmetries.

The Proposition 4.5 (1) (2) , that is ,hR(X, Y)Z, Wi is skew-symmetric in both (X, Y)
and (Z, W) , tells
Corollary 4.2. For two vector fields

! !
a b X
(aX + bY, cX + dY) =
c d Y
We have
hR(aX + bY, cX + dY)(cX + dY), cX + dYi

" !#2
a b
= det hR(X, Y)Y, Xi
c d
Proof. Exercise

Proposition 4.5 (1)(2)(3) tells the curvature tensor R is completely determined by

the values of hR(X, Y)Y, Xi
Corollary 4.3. Suppose hR1 (X, Y)Y, Xi = R2 (X, Y)Y, Xi, ∀X, Y , Then hR1 (X, Y)Z, Wi =
R2 (X, Y)Z, Wi, ∀X, Y, Z, W
Proof. It is clearly suffice to prove that if
hR(X, Y)Y, Xi = 0, ∀X, Y ⇒ hR(X, Y)Z, Wi = 0
Now we have
0 = hR(X, Y + W)(Y + W), Xi = hR(X, Y)Y, Xi + hR(X, Y)W, Xi + hR(X, W)Y, Xi + hR(X, W)W, Xi
= 2hR(X, Y)W, Xi, ∀X, Y, W(U se(1)(2)(4))
Moreover
0 = hR(X + Z, Y)W, X + Zi
= hR(X, Y)W, Xi + hR(X, Y)W, Zi + hR(Z, Y)W, Xi + hR(Z, Y)W, Xi
0 = hR(X, Y)W, Zi + hR(Z, Y)W, Xi

= −hR(Y, W)X, Zi − hR(W, X)Y, Zi + hR(Z, Y)W, Xi(using First Bianchi Identity)
⇒ 2hR(Z, Y)W, Xi = hR(Y, W)X, Zi (1)

By a similiar argument starting from
0 = hR(X + W, Y)Y, X + Wi
We will obtain
2hR(X, Z)Y, Wi = hR(Y, Z)X, Wi (2)

Using symmetries , we can rewrite (1) and (2) as
2hR(Y, Z)X, Wi = hR(X, Z)Y, Wi

and 2hR(X, Z)Y, Wi = hR(Y, Z)X, Wi
which implies hR(X, Z)Y, Wi = 0, ∀X, Y, Z, W

4.3 Sectional Curvature

Consider another (0,4)-tensor : for X, Y, Z, W ∈ ΓT M
G(X, Y, Z, W) = hX, Zig hY, Wig − hX, Wig − hY, Zig

It is not hard to check G satisfies the following properties
4.3. SECTIONAL CURVATURE 85
(1)G(X, Y, Z, W) = −G(Y, X, Z, W)
(2)G(X, Y, W, Z) = −G(X, Y, Z, W)
(3)G(X, Y, Z, W) + G(Y, Z, X, W) + G(Z, X, Y, W) = 0
(4)G(X, Y, Z, W) = G(Z, W, X, Y)
Recall from last section that (4) is actually a consequence of properties (1)–(3)
Hence G behaves very similar to the Riemannian curvature tensor hR(X, Y)Z, Wi.
In particular ,for the linearly independent vector X p , Y p ∈ T p M,
G(X p , Y p , X p , Y p ) = hX p , X p ihY p , Y p i − hX p , Y p i2
= hX p , X p ihY p , Y p i − hX p , X p ihY p , Y p i cos2 θ
= hX p , X p ihY p , Y p i sin2 θ
equals the area of the parallelogram spanned by X p , Y p ,

By the proof of Corollary 4.2, we have

G aX p + bY p , cX p + dY p , aX p + bY p , cX p + dY p
" !#2
a b
= det G Xp, Yp, Xp, Yp
c d
Therefore, we have
Proposition 4.6. The quantity
D E
R Xp, Yp Yp, Xp R Xp, Yp, Xp, Yp
K Xp, Yp := =
G Xp, Yp, Xp, Yp G Xp, Yp, Xp, Yp

R Xp, Yp, Xp, Yp
=D ED E D E2
Xp, Xp Yp, Yp − Xp, Yp
depends only on the two dimensional subspace
π p = span(X p , Y p ) ⊂ T p M
That is , it is independent of the choice of basis {X p , Y p } of π p
Definition 4.1. (sectional curvature) We will call K(π p ) = K(X p , Y p )

the sectional curvature of (M, g) at p with respect to the plane π p = span(X p , Y p )
Remark 4.7. Note that Proposition

(1) The sectional curvature K is Not a function on M except when dimM=2.
(2) K(Ag) = A1 K(g)
Proposition 4.7. Let (M, g) be a 2-dim Riemannian manifold , and let X p , Y p ∈ T p M
be linearly independent . Then
D E
R Xp, Yp Yp, Xp
K(p) = K X p , Y p = D ED E D E2
Xp, Xp Yp, Yp − Xp, Yp
is the same as the Gaussian curvature at p.
Sketch of proof :Let (u, x1 , x2 ) be a coordinate neighborhood of p ∈ M
By Proposition 4.6 ,it suffices to verify the proposition when
∂ ∂
X p = 1 , Y p = 2
∂x p ∂x p
In this case
∂ ∂ ∂ ∂
* ! +
R , , = R1212 (p)
∂x1 p ∂x2 p ∂x2 p ∂x1 p

while G(X p , Y p , X p , Y p ) = g11 g12 − g212

(p)
Hence K(p) = g R1212 2
11 g12 −g12
Recall that the Gaussian curvature can be expressed via the first fundamental form
Edx1 ⊗ dx1 + Fdx1 ⊗ dx2 + Fdx2 ⊗ dx1 + Gdx2 ⊗ dx2

where in our case E = g11 , F = g12 = g21 , G = g22
Remark 4.8. Note that Proposition 4.6 and Proposition 4.7 together implies that Gaus-
sian curvature is indeed independent of the choice of coordinates.
Or equivalently, if g1 , g2 are locally isometric , then they lead to the same Gauss
curvature . This is the celebrated ” Theorem Egregium”
Remark 4.9. We see in Exercise 5(2)(This exercise is in that isometries preserve Levi-
Civita connctions. That is , given (M1 , g1 , ∇(1) ), (M2 , g2 , ∇(2) ) ,and ϕ : M1 −→ M2 be
an isometry. Then for any X, Y ∈ Γ(T M) we habe
x Y) = ∇dϕ(x) dϕ(Y)
(2)
dϕ(∇(1)
As direct consequences, we see if
R(i) (X, Y)Z := ∇(i) (i) (i) (i) (i)

X ∇Y Z − ∇Y ∇X Z − ∇[X,Y] Z
then
dϕ(R(1) (X, Y)Z) = R(2) (dϕ(X), dϕ(Y))dϕ(Z) and g1 (R(1) (X, Y)Z, W) = g2 (R(2) (dϕ(X), dϕ(Y))dϕ(Z), dϕ(W))◦ϕ
In particular , if ϕ : M1 −→ M2 is an isometry s.t. dϕ(π p ) = π0ϕ(p) ⊂ T ϕ(p) M2

We have the sectional curvature of π p and that of π0ϕ(p) are the same .
4.3. SECTIONAL CURVATURE 87
Proposition 4.8. Let (m, g) be a Riemannian manifold , and let π p be a 2-dim subspave
of T p M , spanned by X p , Y p ∈ T p M . Let O ⊂ π p be a neighborhood of 0 ∈ T p M on
which exp p is a diffeomorphism , let i : exp p (O) ,→ M be the inclusion , and let R̄ be
the curvature tensor for exp p (O) with the induced Riemannian metric i∗ g. Then
D E D E
R̄(X p , Y p )Y p , X p = R(X p , Y p )Y p , X p
hR(X p ,Y p )Y p ,X p i
Consequently ,K(π p ) = G(X p ,Y p ,X p ,Y p ) is the Gaussian curvature at p of the surface
exp p (O)
Proposition 4.9. The Riemannian curvature tensor at p is determined by all the sec-
tional curvature at p.
Proof. By Corollary 4.3

Definition 4.2. A Riemannian manifold (M, g) is said to have constant (sectional) curvature
if its sectional curature K(π p ) is a constant,i.e. is independent of p and is independent
of π p ∈ T p M.
Proposition 4.10. A Riemannian manifold (M, g) has constant curvature k if and only
if
hR(X, Y)W, Zi = kG(X, Y, Z, W), ∀X, Y, Z, W ∈ Γ(T M)

i.e.
R(Z, W, X, Y) = R(X, Y, Z, W) = kG(X, Y, Z, W), orR = kG
Proof. Recall both hR(X, Y)W, Zi and G(X, Y, Z, W) satisfy the symmetries (1)–(3) .Hence
S (X, Y, Z, W) := hR(X, Y)W, Zi − kG(X, Y, Z, W)

satisfies
(1) S (X, Y, Z, W) = −S (Y, X, Z, W)
(2) S (X, Y, Z, W) = −S (X, Y, W, Z)
(3) S (X, Y, Z, W) + S (Y, Z, X, W) + S (Z, X, Y, W) = 0
(4) S (X, Y, Z, W) = S (Z, W, X, Y)
Notice our assumption implies S (X, Y, X, Y) = 0
By the proof of Corollary 4.3 , we have S (X, Y, Z, W) = 0

Up to now ,we haven’t made use of the second Bianchi identity . Proposition 4.5
(5)
∇R(W, Z, X, Y, V) + ∇R(W, Z, Y, V, X) + ∇R(W, Z, V, X, Y) = 0

In fact , it leads to the following Schur’s theorem.
Theorem 4.2. (Schur) Let (M, g) be a connected Riemannian manifold of dimension

n ≥ 3. If
K(π p ) = f (p) (*)
depends only on p ,then (M, g) is of constant curvature.
Remark 4.10. (1) Thm 4.2 is obviously not true for (M, g) with dim =2. We know in
that case (*) always holds, but M need not be of constant curvature.
(2) Thm4.2 says that the isometry of a Riemannian manifold ,i.e., the property that
at each point all directions are geometrically indistinguishable, implies the homogene-
ity , i.e., that all points are geometrically indistinguishable. In particular, a pointwise
propert implies a global one .
Before proving Thm 4.2 , we prepare the following lemma.
Lemma 4.1. The tensor G is parallel ,i.e. ∇G = 0 .
Proof. For any X, Y, Z, W, V ∈ Γ(T M), we have
(∇V G) (X, Y, Z, W)= V(hX, ZihY, Wi − hX, WihY, Zi)

− h∇V X, Zi hY, Wi − hX, ∇V Zi hY, Wi
−hX, Zi h∇V Y, Wi − hX, Zi hY, ∇V Wi
− h∇V X, Wi hY, Zi − hX, ∇V Wi h Y, Zi
− hX, Wi h∇V Y, Zi − hX, Wi hY, ∇V Zi
By
V(hX, Zi hY, Wi) = V(hX, Zi) hY, Wi + hX, Zi · V hY, Wi
and compatibility of ∇ with g ,we conclude
(∇V G)(X, Y, Z, W) = 0
Proof. Proof of Thm 2:(An application of the second Bianchi Identity).

By assumption and Proposition 4.10 ,we have
R = f G, when f : M −→ R
Lemma 4.1 above tells ∇G = 0 Hence for all V ∈ Γ(T M),
we have ∇V R = ∇V ( f G) = V( f )G
By the second Bianchi Identity ,we have
0 = ∇V R(W, Z, X, Y) + ∇X R(W, Z, Y, V) + ∇Y R(W, Z, V, X) (*)

= V( f )G(W, Z, X, Y) + X( f )G(W, Z, Y, V) + Y( f )G(W, Z, V, X)
for any X, Y, Z, W, V ∈ Γ(T M)

4.4. RICII CURVATURE AND SCALAR CURVATURE 89
Since it is a tensor identity , the RHS only depends on X p , Y p , Z p , W p , V p ∈ T p M.

Since dim(M) ≥ 3, we can pick X p , Y p , V p ∈ T p M such that .
D E D E D E
Xp, Yp = Xp, Vp = Yp, Vp = 0
and X p , 0, Y p , 0, |V p | = 1
then (*) implies

O = V p ( f ) hW p , X p ihZ p , Y p i − hW p , Y p ihZ p , X p i

+ X p ( f ) hW p , Y p ihZ p , V p i − hW p , V p ihZ p , Y p i

+ Y p ( f ) hW p , V p ihZ p , X p i − hW p , X p ihZ p , V p i
Recall , we still have freedom for the choice of W p , Z p .

Let us set Z p = V p ,then
D E D E
0 = Xp( f ) Wp, Yp − Yp( f ) Wp, Xp
for ∀W p ∈ T p M
Hence 0 = X p ( f )Y p − Y p ( f )X p
D E
However , X p , Y p = 0. That is
X p ( f ) = Y p ( f ) = 0, ∀X p , 0, Y p , 0
So f must be a constant function on M.

4.4 Ricii Curvature and Scalar curvature

The Ricci curvature tensor is defined to be
Ric(Y, Z) := tr(X 7→ R(X, Y)Z)
Notice that at p
R(·, Y)Z : T p M −→ T p M
is a linear map between vector spaces.

In local coordinate (u, x1 , x2 , ..., xn ),we have
∂ ∂ ∂ ∂
! ! ! X
Ric pq := Ric , = tr X →
7 R X, = Rqj jp
∂x p ∂xq ∂x p ∂xq j
Moreover,we have
X X
Rqj jp = gi j gil Rlq jp
j i, j
X
= gi j Riq jp
j
∂ ∂ ∂ ∂
X * ! +
= g R ij
, ,
i, j
∂x j ∂x p ∂xq ∂xi
∂ ∂
* +
= tr R(·, p ) q , ·
∂x ∂x
∂ ∂
!
= trR ·, q , ·, p
∂x ∂x
Therefore
Ric(Y, Z) = trR(,̇Y, ·, Z)
(as a (0,2)-tensor)
Recall the trace of a (0,2)-tensor from the section on Hessian .
In particular ,we observe that
Ric(Y, Z) = Ric(Z, Y)
That is ,Ric is a symmetric (0,2)-tensor fiels on M.
Definition 4.3. (Ricci curvature ) The Ricci curvature at p in the direction X p ∈ T p M

is defined as
Ric(X p ) := Ric(X p , X p )
Remark 4.11. Ricci curvature is again NOT a function on M . We can think of the
Ricci curvature as a function defined on one-dimensional subspace of T p M.
We can ask similar questions about Ricci curvature as in the case of sectional
curvature : What information do we lose when restricting the Ricci curvature tensor to
Ric(X, X ? The answer is again that we do not lose anything .
Lemma 4.2. Let T be a symmetric 2-tensor ,then for any X, Y ,we have
1
T (X, Y) = (T (X + Y, X + Y) − T (X, X) − T (Y, Y))
2
Hence
1
Ric(X p , Y p ) = (Ric(X p + Y p ) − Ric(X p ) − Ric(Y p ))
2
We actually should have normalized the length of the vector along which the Ricci
curvature is calculated. After that , Ricci curvature is defined on ”tangent directions”
!
X Ric(X) Ric(X, X)
Ric := =
kXk g(X, X) g(X, X)
Definition 4.4. The Ric manifold is called an Einstein manifold with Einstein constant
k , if
Ric(X) = kg(X, X), ∀X ∈ Γ(T M)
i,e., M has ”constant Ricci curvature”.
Remark 4.12. Let X p ∈ T p M be an unit tangent vector . Extend it to be an othornormal

basis {X p , e2 , ..., en } of T p M . Then
Ric(X p ) = trR(·, X p , ·, X p )
Xn
= R(ei , X p , ei , X p )
i=2
n
X
= K(ei , X p )
i=2
In particular , if (M, g) has constant curvaturek,then (M, g) is Einstein with Einstein

constant (n − 1)k
Proposition 4.11. A Riemannian manifold is an Einstein constant k if and only if
Ric = kg
Proof. Define T (X, Y) = Ric(X, Y) − kg(X, Y) . Hence T is symmetric. By assumption

T (X, X) = 0 Lemma 4.2 tells T (X, Y) = 0 i.e. Ric = kg
We also have the following version of Schur’s theorem.
Theorem 4.3. (Schur) Let (M, g) be a connected Riemannian manifold of dimesion ≥ 3.

If Ric(X p ) = f (p)g(X p , X p ), ∀X p ∈ T p M, where f (p) depends only on p ,then (M, g) is
Einstein.
Proof. Apply the second Bianchi identity in the same manner as in Theorem 4.2
Step 1: By Proposition 4.11 , the assumption implies
Ric = f g
Note for Levi-Civita connection, we have automatically
∇g = 0
Hence ∀V ∈ Γ(T M), we have
∇V Ric = V( f )g
Step 2: Apply 2nd Bianchi Identity . At p ∈ M ,pick as normal coordinate (u, x1 , ..., xn ),
we have for X p , Y p , V p ∈ T p M

∇V Ric X p , Y p = V Ric X p , Y p − Ric ∇V p X p , Y p − Ric X p , ∇X Y p
 n !
X ∂ ∂
= V p  R , X p , i , Y p 

∂x ∂x

i
i=1
n
∂ ∂
X !
− R , ∇V p X p , i , Y p
i=1
∂xi ∂x
n
∂ ∂
X !
− R , X p , , ∇ Y
Vp p
i=1
∂xi ∂xi
n ∂ ∂
X !!
= ∇V p R , X p , , Y p
i=1
∂xi ∂xi
(We used ∇V p ∂x∂ i = 0 since in normal coordinate)

Second Bianchi identity implies
n "
∂ ∂ ∂ ∂ ∂
X #
0= ∇V p R( , X p , , Y p ) + ∇ ∂ R( , X p , Y p , V p ) + ∇Yp ( , X p , V p , )
i=1
∂xi ∂xi ∂xi ∂xi ∂xi ∂xi
n
X ∂
= ∇V p Ric(XP , Y p ) − ∇Y p Ric(X p , V p ) + ∇ ∂i R( , Xp, Yp, Vp)
i=1
∂x ∂xi
n
X ∂
= V p ( f )g(X p , Y p ) − Y p ( f )g(X p , V p ) + ∇ ∂i R( , Xp, Yp, Vp)
i=1
∂x ∂xi
Step 3: Pick special X p , Y p , V p
Let X p = Y p = ∂x∂ j , V p = ∂x∂h
We have
n
∂f ∂f X ∂ ∂ ∂ ∂
0= − δ jh + ∇ ∂i R( i , j , j , h )
∂x h ∂x i
i=1
∂x ∂x ∂x ∂x ∂x
Summing j from 1 to n,
n n
∂f ∂f XX ∂ ∂ ∂ ∂
0=n· − h− ∇ ∂i R( j , i , j , h )
∂x h ∂x i=1 j=1
∂x ∂x ∂x ∂x ∂x
n
∂f X ∂ ∂
= (n − 1) − (∇ ∂i Ric)( i , h )
∂x h
i=1
∂x ∂x ∂x
n
∂f X ∂f ∂ ∂
= (n − 1) − g( i , h )
∂x h
i=1
∂x ∂x ∂x
i
∂f
= (n − 2)
∂xh
∂f
Hence, when n ≥ 3, we have ∂x h = 0,∀h = 1, 2..., n
This implies that f ≡ constant .

Remark 4.13. (1): In fact ,Theorem 4.3 implies 4.2 . Notice that K(π p ) = f (p) depends
only on p implies
Ric(X p ) Ric(X p )
= f (p), depends only on p =⇒ ≡ constant(By T heorem 4.3)
g(X p , X p ) g(X p , X p )
(a)
Ric(X p ) Ric(X p )
= f (p), depends only on p implies = (n − 1) f (p) (b)
g(X p , X p ) g(X p , X p )
(a)+(b) =⇒ K(π p ) = f (p) ≡ constant
(2) In [BSSG] ,(M, g) is called Einstein if
Ric(X p ) = f (p)g(X p , X p )
when f (p) depends only on p . By Theorem 4.3 ,there is no difference from our
definition in case dim ≥ 3
[BSSG] ’s notation has the preperty that ”any 2-dim Riemannian manifold in Ein-
stein”.
Definition 4.5. (Scalar curvature) The scalar curvature S is defined as the trace of the
Ricci curvature tensor (which is a symmetric (0,2)-tensor ),i.e.
S = gi j Rici j = trRic(·, ·)
Remark 4.14. (1) S is indeed a function on M.
(2) Let {e1 , ..., en } be an orthonormal basis of T p M , we have
n
X n
X
S (p) = tr(Ric)(p) = Ric(ei , ei ) = Ric(ei )
i=1 i=1
n
X
= trR(·, ei , ·, ei )
i=1
Xn Xn
= R(e j , ei , e j , ei )
i=1 j=1
n X
X n
= K(e j , ei )
i=1 j=1
X
=2 K(ei , e j )
i< j
(3) If (M, g) is of constant curvature k, we have
Ric = (n − 1)kg, and S = n(n − 1)k

If (M, g) is Einstein with Einstein constant k , we have
S = nk
Proposition 4.12. An n(≥ 3) -dimensional Riemannian manifold (M, g) is Einstein iff

S
Ric = g
n
Proof. =⇒ By definition .
⇐=Ric = Sn g where Sn (p) depends only on p. Schur’s Theorem (n ≥ 3) =⇒ S
n ≡
constant

In particular , Ricci curvature provides less information than sectional curvature,

and scalar curvature provides even less information than Ricci curvature . But in di-
mension 2 or 3 , something special happens.
Ric(X )
n = 2 , K(π p ) = g(X p ,Xpp ) = 2S (p)
There is no difference from an information point of view in knowing K, Ric, orS .
Let {e1 , e2 , e3 } be an orthonormal basis for T p M , then
K (e1 , e2 ) + K (e1 , e3 ) = Ric (e1 )




(e1 , e2 ) + K (e2 , e3 ) = Ric (e2 )

K


 K (e , e ) + K (e , e ) = Ric (e )


1 3 2 3 3
In other words,
  K (e1 , e2 )

    
 1 0 1   Ric (e1 ) 
 1

1 0   K (e2 , e3 )  =  Ric (e2 )
  
 (**)
K (e1 , e3 )
 
0 1 1 Ric (e3 )
Notice that  
 1 0 1 
det  1 1 0  = 2 , 0
 
0 1 1

Therefore any sectional curvature can be computed from Ric .
Proposition 4.13. (M 3 , g) is Einstein iff (M 3 , g) has constant sectional curvature .
Proof. ⇐= By definition
=⇒ Solving (**) for the case Ric(e1 ) = Ric(e2 ) = Ric(e3 ) we obtain K(e1 , e2 ) =
K(e2 , e3 ) = K(e1 , e3 )

But for scalar curvature , when n = 3 , there are metrics with constant scalar curva-
ture that are not Einstein.
We will see whether the (sectional, Ricci , scalar ) curvatures of Riemannian man-
ifolds are constant , or more generally although not constant but still bounded by some
inequalities have much implications to the analysis, geometry and topology of (M, g).
4.5. THE SECOND VARIATION : REVISITED [JJ,4.1] [WSY,CHAP 6] 95
In particular , we explain the terminology ”Ric ≥ k”, (Ricci curvature is lower

bounded), this means more precisely that
Ric(X) = Ric(X, X) ≥ kg(X, X), ∀X
We have discussed several times that a symmetric (0,2)-tensor has a ”correspond-

ing” linear transformation ,since
X
Ric X p , X p = R ei , X p , ei , X p
i
XD E
= R(ei , X p )X p , ei
i
XD E
= R(X p , ei )ei , X p
i
D E
= #Ric X p , · , X p
X
=⇒ #Ric(X p , ·) = R(X p , ei )ei
i
X
X p 7→ R(X p , ei )ei
i
is a linear transformation between T p M and T p M. The condition ”Ric ≥ k” is equiva-

lent to say all eigenvalues of X p 7→ i R(X p , ei )ei are ≥ k .
P
Let us mentions the following theorem of Lohkamp.
Theorem (Lohkamp , Annuls of Math . 140(1994),655-683) For each manifold M n , n ≥
3, there is a complete metric g M with
−a(n)g M < Ric(g M ) < −b(n)g M

with constants a(n) > b(n) > 0 depending only on the dimension n.
Theorem 4.4. (Lohkamp)For each manifold M n , n ≥ 3, there is a complete metric

g M with negative Ricci curvature and finite volume. That is , there are NO topological
obstructions for negative Ricci curvature metrics.
4.5 The Second Variation : Revisited [JJ,4.1] [WSY,chap

6]
Recall from section 1 that the curvature tensor is closely related to the second variation
of the energy functional (and the length functional ) of a normal geodesic . In this
section , we will discuss some geometric and topological implications when assuming
curvature restrictions via applying SVF .
Let γ be a normal geodesic , i.e. ,|γ̇| = 1 . Consider a variation
F :[a, b] × (−, ) −→ M
(t, v) 7→ F(t, v)
(i.e. F is smooth and F(t, 0) = γ(t))

The variational field V(t) = ∂F
∂v (t, 0) is a vector field along γ.
Definition 4.6. (geodesic variation ) . The variation F is called a geodesic variation

if each curve γv (t) := F(t, v) is a geodesic
Next , we recall briefly the second variation formula from section 1. For the one-
parameter family of curvs {γv }v∈(−,) , we have E(v) := E(γv ) be a function on (−, )
.Since γ0 = γ is a geodesic , we have E 0 (0) = 0.
∂2 ∂F ∂F ∂F ∂F
Z b * + * +!
E(v) = ∇∂ ∇∂ , + ∇∂ ,∇∂ dt
∂v2 ∂v ∂t ∂v ∂t ∂t ∂v ∂t ∂v
a
∂F ∂F ∂F ∂F ∂F ∂F ∂F ∂F
Z b * ! + * + * +
= R , , + ∇∂∇∂ , + ∇∂ ,∇∂ dt
a ∂v ∂t ∂v ∂t ∂t ∂v ∂v ∂t ∂t ∂v ∂t ∂v
∂F ∂F ∂F ∂F ∂ ∂F ∂F ∂F ∂F
Z b* ! + Z b * + * +
=− R , , dt + ∇∂ , − ∇∂ ,∇∂ dt
a ∂v ∂t ∂t ∂v a ∂t
∂v ∂v ∂t ∂v ∂v ∂t ∂t
∂F ∂F
Z b* +
+ ∇∂ ,∇∂ dt
∂t ∂v ∂t ∂v
a
Proposition 4.14. Let F be a geodesic variation of a curve γ : [a, b] −→ M

Then
∂2 ∂F ∂F ∂F ∂F ∂F ∂F
Z b * + * ! +!
E(v) = ∇∂ ,∇∂ − R , , dt
∂v2 ∂t ∂v ∂t ∂v ∂v ∂t ∂t ∂v
a
∂F
Proof. Use the fact ∇ ∂ ∂v = 0 since F is a geodesic variation
∂v

In particular , for a geodesic variation of a normal geodesic γ : [a, b] −→ M ,we
have
∂2
Z b
E(v) := E”(0) = (h∇T V, ∇T Vi − hR(V, T )T, Vi) dt
∂v2 v=0

a
Observe that when M has nonpositive sectional curvature , we have
∂F ∂F ∂F ∂F ∂F ∂F ∂F ∂F ∂F ∂F
* ! + ! !
− R , , = −K , G , , , ≥0
∂v ∂t ∂t ∂v ∂v ∂t ∂v ∂t ∂v ∂t
∂2
Hence ∂v2
E(v) ≥ 0,for v ∈ (−, ). This tells immediately:
Corollary 4.4. On a Riemannian manifold with nonpositive sectional curvature , geodesics
with fixed endpoints are always locally minimizing
Remark 4.15. Here a geodesic γ is locally minimizing means that for rhis γ : [a, b] −→
M there exists some δ > 0,such that for any smooth curve c : [a, b] −→ M with
c(a) = γ(a), c(b) = γ(b) , d(γ(t), c(t)) ≤ δ, ∀t ∈ [a, b] we have E(c) ≥ E(γ)
Proof. For each t ∈ [a, b] ,Let δt be the parameter of the totally normal neighborhood
Wt of γ(t) (That is ,∀q ∈ Wt ,expq ) Since γ([a, b]) is compact , we can find a finite
subcover of the cover {expγ(t) B(0, δt )}t∈[a,b] . Hence we can find a positive number
δ > 0 for γ : [a, b] −→ M such that δt ≥ δ, ∀t ∈ [a, b]
Let c : [a, b] −→ M be another curve s.t. d(γ(t), c(t)) ≤ δ,∀t ∈ [a, b]

Construct the variation as
F(t, s) = expγ(t) s · exp−1
γ(t) (c(t)), t ∈ [a, b].s ∈ [−1, 1]
Notice that F(t, 0) = γ(t), F(t, 1) = c(t)
F is a geodesic variation (and proper )
F is proper and γ is a geodesic =⇒ E 0 (0) = 0
F is a geodesic variation =⇒ E”(s) ≥ 0, s ∈ [−1, 1]
Recall Taylor’s expansion of an one-variablw smooth functional, we have
Z 1
E(1) = E(0) + E (0) +
0
(1 − t)E”(t)dt ≥ E(0)
0
That is E(c) ≥ E(γ)

Remark 4.16. (1) Note that the ”locally minimizing energy ” also implies ”locally
minimizing length” From the proof above , for any curve c : [a, b] −→ M close to the
normal geodesic γ(t), we can reparametrize c : [a, b] −→ M,s.t.
Exercise 4.1. Let γ : [a, b] −→ M be a smooth curve , and
F :[a, b] × (−, ) × (−δ, δ) −→ M

(t, v, w) 7→ F(t, v, w)
be a 2-parameters variation of γ . Denote by
∂F ∂F
V(t) = (t, 0, 0), W(t) = (t, 0, 0)
∂v ∂w
the two corresponding variational field.
(1) Show that

∂2 b
∂F ∂F ∂F ∂F ∂F ∂F
Z * + * +
1 n
L(v, w) = ∇∂ ,∇∂ − R( , ) ,
∂w∂v ∂F
a k ∂t k
∂t ∂v ∂t ∂w ∂w ∂t ∂t ∂v
∂F ∂F ∂F ∂F ∂F ∂F o
* + * +* +
1
+ ∇∂∇ ∂ , − ∇ ∂ , ∇ ∂ , dt
∂w ∂v ∂t 2 ∂t ∂v ∂t ∂v ∂w ∂t
k ∂F
∂t
∂t k
E 21
where k ∂F ∂F ∂F
D
∂t k = ∂t , ∂t
(2) Let γ be a mpr,a; geodesic , i.e.,kγ̇k = 1. Show that
∂2 b
Z b
L(v, w) = (h∇T V, ∇T Wi − hR(W, T )T, Vi − T hV, T i T hW, T i) dt+h∇W V, T i
∂w∂v w=v=0
a
a
where T (t) := γ̇(t) is the velocity field along γ
(3) Consider the orthogonal component V, e We of V, W with respect to T , that is
V ⊥ := V − hV, T i T
W ⊥ := W − hW, T i T
Show that
∂2 b
Z D E D E b
L(v, w) = ∇T V ⊥ , ∇T W ⊥ − R(W ⊥ , T )T, V ⊥ dt + h∇W V, T i
∂w∂v (0,0) a a
Remark 4.17. Observe in the above proof, the ”properness ” of the variation F is only
used to conclude E 0 (0) = 0. When we consider variation of closed geodesics, i.e. a
geodesic
γ : S 1 −→ M
When S 1 is the unit circle parametrized by [0, 2π).
(in fact , γ : S 1 −→ M, γ(0) = γ(2π), γ̇(0) = γ̇(2π) ), the argument in the proof
still works.
Corollary 4.5. On a Riemannian manifold with nonpositive (negative,resp) sectional
curvature , closed geodesics are locally minimiing .(strict local minima,resp)
Notice that on a manifold with vanishing curvature , closed geodesics are still lo-
cally minimizing , but not necessarily strictly so any more .On a manifold with positive
curvature , closed geodesics in general do not minimize anymore.(?) The following
picture is very imspiring .
We will derive a global consequence of this fact (?).

We give a general remark about how (SVF) implies minimizing property of geodesics.
Let γ : [a, b] −→ M be a normal geodesic , F be a variation of γ , we have
d2 b Z b
v=0 E(v) = h∇V V, T i a +

2
h∇T V, ∇T Vi − hR(V, T )T, Vi dt
dv a
where V is the variational field and T is the velocity field along γ.

Geometrically speaking , when F is a proper variation , or γ : [a, b] −→ M is a
closed geodesic ,(=⇒ γ(a) = γ(b), T (a) = T (b)),
d2
Rb
E(v) = a h∇T V, ∇T Vi − hR(V, T )T, Vi dt

we have dv 2
v=0
(1) If M has nonpositive (negative ,resp) curvature , E”(0) ≥ 0 (E”(0) > 0,resp)
=⇒ γ is (strictly) locally minimizing .
(2) If M has positive curvature , -hR(V, T )T, Vi < 0.
If
hR(V, T )T, Vi > h∇T V, ∇T Vi (?)
then E”(0) < 0, and hence γ cannot be (locally) minimizing .
The philosophy if (2) leads to the applications of (SVF) we will discuss soon.
Synge Theorem (⇐=) When M is compact , orientable , even-dimension, of positive
curvature , for any nontrivial closed geodesic γ , we can find V ,s.t. (?) holds.
That is , under the assumptions, any nontrivial (not homotopic to constant curve)
geodesic can not be locally minimizing.
Bonnet-Myers Theorem (⇐=) When M is of sectional curvature ≥ k > 0, geodesics
of length > √πK can not be (locally) minimizing .
Next, let us discuss this two applications in more detail.
Synge Theorem [WSY, Chap 6][JJ, Chap 4,4.1][dC,Chap 9,3]
Lemma 4.3. Let (M, g) be an , orientable, even-dimensional Riemannian manifold
with positive sectional curvature. Then any closed geodesic which are not homotopic to a constant curve
can not be minimizing in its (free) homotopy class.
Lemma 4.4. Let (M, g) be a compact Riemannian manifold . Then every (free) homo-
topy class of closed curve in M contains a shortest one (which is , therefore, a closed
geodesic) [JJ,Theorem 1.5.1]
Remark 4.18. (1) A closed curve c can be considered as a continuous map c : S 1 −→
M , where S 1 is the unit circle.
Recall that two contimuous maps
c0 , c1 : S 1 −→ M
are called homotopic if there exists a continuous map
F : S 1 × [0, 1] −→ M
with F(t, 0) = c0 (t).F(t, 1) = c1 (t),∀t ∈ S 1
And the concept of homotopy defines an equivalence relation of all closed curves
in M.
(2) Suppose (M, g) satisfy both the assumptions of Lemma 4.3 and that of Lemma
4.4 , then every homotopy class of closed curves in M contains the constant curve .
That is , M us simply connected , i.e.,π1 (M) = {1}. This is exactly what Synge Theorem
says.
Theorem 4.5. [Synge,1936,On the connectivity of spaces of positive curvature ,Quar-
tely Journal of Mathematics (Oxford Series),7,316-320] Any compact, orientable, even-dimensional
Riemannian manifold,positive curvature is simply connected.
Now we start to prove Lemma 4.3 . In fact the restricitions ”orientable , even-
dimensional” guarantee the existence a parallel normal vector field along a nontrivial
closed geodesic.
If γ : [a, b] −→ M is not a closed one , it is not hard to find a parallel normal vector
field along it . Just pick a vector V(a) ∈ T γ(a) M, hV(a), γ̇(a)i = 0 and V(t) is given by
the parallel transport along γ.
But for a closed one , V(b) = Pγ.a.b V(a) is not necessarily coincide with V(a).
Note for the velocity field along a closed geodesic γ , we have
Pγ,a,b γ̇(a) = γ̇(b) = γ̇(a)
That is , the orthogonal linear transformation
Pγ,a,b : T p M −→ T p M (p = γ(a) = γ(b))

has an eigenvalue +1 with eigenvector γ̇(a)
If the multiplicity of eigenvalue +1 ≥ 2, then we have a vector V(p) ∈ T p M lying
in the orthogonal complement of γ̇(a) s.t. Pγ,a,b V(p) = V(p) .
Hence V(t) := Pγ,a,t V(a) gives a parallel normal vector field along γ.
Next , we explain ”orientable , even-dimensional ” guarantee that the multiplicity
of eigenvalue +1 of Pγ,a,b ≥ 2.
Since Pγ,a,b is orthogonal, we have det(V)=+1 or -1
Lemma 4.5. If det(Pγ,a,b =)=+1,and M is even-dimensional then the multiplicity of

the eigenvalue +1 ≥ 2
Proof. Since Pγ,a,b : T p M −→ T p M is orthogonal , its eigenvalues can be listed as
λ1 , λ̄1 , ..., λ j , λ̄ j , −1, −1, · · · , −1, 1, 1, · · · , 1

| {z } | {z }
k l
where λi , i = 1, 2, ..., j are complex numbers with |λi | = 1,
M even − dimensional =⇒ T p M even − dimensional

=⇒ k + l is even
Since Pγ,a,b : γ̇(a) 7→ γ̇(b) = γ̇(a), l ≥ 1(i.e. l , 0)

Hence l is even and l , 0. That is l ≥ 2
In fact , det(Pγ,a,b )=+1 is guaranteed by ” orientability ” of M . Let us recall briefly
the concept of orientability.
Given a vector space V, let {ei }ni=1 and { f j }nj=1 be two basis , and f j = aij ei . Then
det(aij ) is either positive or negative. If det(aij )¿0, we say the two basis have the same
orientation . This defines an equivalence relation for all basis of V. There exactly 2
equivalant classes. We call each of them an orientation of V .
Alternatively , the orientation of V can be described as below : Consider the dual
space V ∗ of V . Then we have
dim Λn (V ∗ ) = 1
and let Ω( f1 , ..., fn ) = det(aij )Ω(e1 , e..., en ) .
That is , given a non-zero Ω ∈ Λn (V ∗ ), two basis {ei },{ f j } have the same orientation
iff Ω( f1 , ..., fn ) and Ω(e1 , ..., en ) have the same sign . In this sense , a nonzero Ω ∈
Λn (V ∗ ) determines an orientation of V.
The second way of description can be generalized to the setting of a manifold. M
is orientable if there exists a C ∞ nowhere vanishing n-form ω. At each p ∈ M, the
basis of T p M are divided into two classes, those with ω(e1 , ..., en ) > 0 and those with
ω(e1 , ..., en ) < 0 . The first class is called the basis coherent with the orientation .
Lemma 4.6. Let (M, g) be an orientable Riemannian manifold and γ : [a, b] −→ M be
a closed curve . Then the parallel transport Pγ,a,b : T p M −→ T p M has determinant 1.
Proof. Since Pγ,a,b is orthogonal , we only need to show
det(Pγ,a,b ) > 0
Let ω be a C ∞ nowhere vanishing n-form ω on M , whose existence is guaranteed
by orientability . Let {ei } be a basis of T p M with ω(e1 , ..., en ) > 0.
Let {ei (t)} := {Pγ,a,b (ei )} be the parallel transport of {ei } along γ . Then t 7→
ω(e1 (t), ..., en (t)) is a nowhere vanishing c∞ function on [a,b] . In particular , ω(e1 (b), ..., en (b)) >
0.
Note {ei (b)ni=1 } is also a basis of T p M , and
ω(e1 (b), ..., en (b)) = det(Pγ.a.b )ω(e1 (b), ..., en ).

Therefore, we have
det(Pγ.a.b ) > 0

Proof of Lemma 1
Let γ : [a, b] −→ M be a nontrivial closed geodesic in M (let p = γ(a) = γ(b)).
By lemma 4.5 and 4.6 , there exists V(p) ∈ T p M,hV(p), γ̇(a)i 0 and Pγ,a,b V(p) =
V(p)
Therefore V(t) := Pγ,a,b V(p) is a parallel normal vector field along γ .
Since γ([a, b]) is compact , there exists δ > 0,s.t.
F :[a, b] × (−δ, δ) −→ M
(t, v) 7→ expγ(t) vV(t)
is a (geodesic) variation of γ. (existence of δ is shown by the argument we used in

the proof of Corollary 4.4)
Since γ is a geodesic, we have E 0 (0) = 0. Moreover,
Z b
E”(0) =

a
(∇T V = 0 since V is parallel)
Z b
=− hR(V, T )T, Vi dt < 0 since sectional curvature > 0
a
Therefore, for v , 0 small enough, γv : [a, b] −→ M is a closed curve homotopic

to γ but with E(γv ) < E(γ).
That is γ is not minimizing in its homotopy class. (In fact , for length we also have
l(γv )2 ≤ 2(b − a)E(γv ) < 2(b − a)E(γ) = l(γ)2 (since kγ̇k ≡ constant))
Lemma 4.4 is a general result for compact Riemannian manifold ( no curvature
restriction is needed).
Proof of Lemma 4.4 Recall from Corollary (2.3) of Chapter 2 , that for a compact
Riemannian manifold M , there exists a ρ0 > 0, s.t. any p, q ∈ M with d(p, q) ≤ ρ0 can
be conneted by precisely one geodesic of shortest path.(Recall this is proved by using
the concept of totally normal neighborhood).
Moreover,the geodesic depends continuously on (p, q). This implies immediately.
Claim. Let (M, g) be a compact Riemannian manifold, and ρ0 > 0 be chosen as
above. Let γ0 , γ1 : S 1 −→ M be closed curves with
d(γ0 (t), γ1 (t)) ≤ ρ0 , ∀t ∈ S 1

Then γ0 , γ1 are homotopic.
For any t ∈ S 1 , let ct (s) : [0, 1] be the unique shortest geodesic from γ0 (t) to γ1 (t)
. (paramatrized proportionally arclength). Since ct depends continuously on its end
points. The map
F(t, s) := ct (s)
is contimuous and yields the desired homotopy.
Next we find the shortest curve in a given homotopy class by method of minimizing
sequence .
Let {γn }n∈N be a minimizing sequence for length in the given homotopy class.
Here and in the sequal , all curves are parametrized proportionally to arc length. γn :
[0, 2π] −→ M .
We may assume each γn us piecewise geodesic. This is because: there exists m,
and
0 = t0 < t1 < ... < tm < tm+1 = 2π
s.t.,L(γn |[ t j−1 , t j ]) ≤ ρ0 /2
(This is realizable since one can equally divide [0,2π],s.t.
ρ0 ρ0
|t j − t j−1 | = , j = 1, ..., m, |tm+1 − tm | <
2|γ̇| 2|γ̇|
l m
and m = 2π
ρ0 )
2|γ̇|
Then replacing γn by the shortest geodesic arc from γn (t j−1 ) to γn (t j ). By the
[t j−1 ,t j ]
claim , this will not change the homotopy class of the curve.
Equivalently to say, we have a minimizing sequence {γn }n such that for each γn ,
there exists p0,n , ..., pm,n for which d(p j−1 , p j ) ≤ ρ0 /2, j = 1, ..., m+1 with pm+1,n = p0,n
and for which γn contains the shortest geodesic from p j−1 to p j .
Since {γn }n is minimizing , the length of γn are bounded, say L(γn ) ≤ C. Then we
can assume that m is independent of n. (This is because L(γn ) ≤ C =⇒ |γ̇n | ≤ 2π C
=⇒
4π|γ|
m ≤ ρ0 + 1 ≤ ρ0 + 1)
2C
Since M id compact , after selection of a subsequence , the points p0,n , ..., pm,n
converge to points p0 , ..., pm as n −→ ∞. The segment of γn from p j−1,n to p j,n then
converges to the shortest geodesic from p j−1 to p j (Recall such geodesic depends con-
tinuously on its end points).
The union of these geodesic segments yields a closed curve γ . By the claim ,γ is
still in the given homotopy class and L(γ) = limn−→∞ L(γn ),i.e. ,γ is the shortest one
in its homotopy class. Therefore ,γ has to be geodesic.(Otherwise, there exists p and q
on γ on which one of the two segments of γ) from p to q has length ≤ ρ0 /2, but is not
geodesic. Then replace this segment by the unique shortest geodesic from p to q. The
resulting curve lies still in the same homotopy class but with a shorter length , which is
a contradiction.)
Proof of Thoerem 4.5 Suppose M is not simply connected. Then there is a homo-
topy class of closed curves which are not homotopic to a constant curve. By Lemma
4.4 , there is a shortest closed geodesic γ in this given homotopy class . By Lemma 4.3
,γ cannot be minimizing , this is a contradiction.
Remark 4.19. (1) Synge theorem tells that any compact , orientable, even-dimensional
manifold which is not simply connected does not admit a metric of positive curvature.
(2) Examples: the real projective spave P2 (R) of dimension two, is compact, non-
orientable. Recall in Excercise (3),2(Homework 3 for 2017) ,we checked there is a
Riemannian metric on P2 (R),s.t. the covering map π : S 2 −→ P2 (R) is a local isometry.
Hence P2 (R) has sectional curvature 1¿0.
But we know π1 (P2 (R)) = Z2 . Hence ”orientability” in the assumption of Synge
Theorem is necessary.
Similarly, ”evem-dimensional” assumption is also necessary. P3 (R) is orientable,compact,
odd-dimensional , of positive curvature , but π1 (P3 (R)) = Z2 .
The above examples are inspiring and it is natural to ask what we can say when
(M, g) is not orientable or not even-dimensional.
Corollary 4.6. Let (M, g) be a compact,non-orientable,even-dimensional Riemannian
manifold of positive sectional curvature, then π1 (M) = Z2 .
Theorem 4.6. (Synge 1936) Let (M, g) be a compact , odd-dimensional Riemannian
manifold of positive sectional curvature, then M is orientable.
Remark 4.20. In particular, Corollary 4.6 gives a geometric proof of the fact π1 (Pn (R)) =
Z2 , when n even (knowing Pn (R) is non-orientable for n is even . Although π1 (Pn (R)) =
Z2 , ∀n).Theorem 4.6 gives a geometric proof of the fact Pn (R) is orientable for n odd.
But we can not say too much about the fundamental group π1 (M) for a compact,odd-
dimensional manifold admitting a metric of positive curvature.
The proofs use property of the orientable double cover of a non-orientable mani-
fold. In order not to interupt our current topic too much , we postpone the proofs.
Bonnet-Myers Theorem :[PP,Chap 6,4.1]
The following lemma was first priven by Bonnet for surfaces and later by Synge
for general Riemannian manifolds as an application of his (SVF).
Lemma 4.7. (Bonnet 1855 and Synge 1926) Let(M, g) be a Riemannian manifold with
sectional curvature ≤ k > 0 . Then geodesics of length ¿ √πk cannot be (locally) mini-
mizing
Proof. Let γ : [0, l] −→ M be a normal (i.e. |γ̇| = 1) geodesic of length l > √π .
k
Let E(0) be a unit vector in T γ(0) M with hE(0), γ̇(0)i = 0.
Then we obtain E(t) := Pγ,0,t E(0) a parallel (vector field along γ).
Consider the following vector field along γ
V(t) := sin(πt/l)E(t)
It corresponds to a proper variation since V(0) = V(l) = 0
By (SVF):
b
d2
Z
E(v) = E”(0) =

dv2 v=0 a
π
Observe ∇T V = sin0 (πt/l)E(t) = l cos(πt/l)E(t) and hence h∇T V, ∇T Vi = ( πl )2 cos2 (πt/l)
and
hR(V, T )T, Vi = sin2 (πt/l) hR(E, T )T, Ei = sin2 (πt/l)K(E, T )
π 2 l
Z Z l
=⇒ E”(0) = 2
( ) cos (πt/l)dt − sin2 (πt/l)K(E, T )dt
0 l 0
π 2 l
Z Z l
2
≤( ) cos (πt/l)dt − k sin2 (πt/l)dt
l 0 0
π π
Z lh i
<k cos2 (πt/l) − sin2 (πt/l) dt (S ince l > √ =⇒ ( )2 < k)
0 k l
Z l
=k cos(2πt/l)dt = 0
0
Hence all nearby curves in the variation are shorter than γ . (by the same argument
as in the end of the proof for Lemma 4.3)
In the above , we see that we actually has (n-1) choices of the parallel normal vector
fields along γ. When sectional curvature ≥ k > 0 , our above argument works for each
of these (n − 1) parallel vector field along γ. On the other hand , for our purpose here,
it’s enough to know that our above argument works for at least one of those (n − 1)
vector fields along γ. This leads to the following extension due to Myers.
Lemma 4.8. (Myers 1941). Let (M, g) be a Riemannian manifold with Ricci curvature
Ric≥ (n − 1)k > 0. Then geodesics of length > √πk cannot be minimizing.
Proof. Similarly as in the proof of Lemma 4.7. Let γ : [0, l] −→ M be a normal

geodesic with l > √πk .
Choose E2 , ..., En ∈ T γ(0) M s.t. γ̇(0), E2 , ..., En form an orthonormal basis for
T γ(0) M. Then Ei (t) := Pγ,0,t Ei and γ̇(t) form an orthonormal basis for T γ(t) M.
Consider n − 1 variational fields along γ
Vi (t) = sin(πt/l)Ei (t), i = 2, 3, ..., n
We have for each i,
π 2 l
Z l
d2
Z
E(vi ) = ( ) cos 2
(πt/l)dt − sin2 (πt/l)K(Ei , T )dt
dv2i vi =0 l

0 0
Z l Z l
<k cos2 (πt/l)dt − sin2 (πt/l)K(E, T )dt
0 0
Taking the summation,

n Z l Z l
X d2 X
E(vi ) < (n − 1)k cos2
(πt/l)dt − sin2
(πt/l) K(Ei , T ) dt
dv2i vi =0

i=2 0 0 i
| {z }
Ric(T )
Z l Z l
≤ (n − 1)k cos2 (πt/l)dt − (n − 1)k sin2 (πt/l)dt
0 0
=0
Hence there exists an i0 ∈ {2, · · · , n}, s.t.

n
X d2
E(vi0 ) < 0
dv2i0 vi0 =0

i=2
And hence γ is not (locally) minimizing .

If we assume further that (M, g) is complete, the above lemma implies and upper
bound of the diameter of (M, g). This seems to have first been pointed out by Hopf-
Rinow(1931) for surfaces in their famous paper on completeness and then by Myers
for general Riemannian manifolds.
(1935. Duke J. Math for sectional curvature restriction
1941. Duke J. Math for Ricci curvature restriction)
Corollary 4.7. Suppose (M, g) is a complete Riemannian manifold with Ricci curva-
ture Ric ≥ (n − 1)k > 0. Then
π
diam(M, g) ≤ √
k
Further more, (M, g) has finite fundamental group.
Remark 4.21. Corollary 4.7 is aftenly referred to as Bonnet-Myers Theorem.
Proof. Lemma 4.8 tells no geodesic can realize distance between any p, q ∈ M with
d(p, q) > √πk . Hopf-Rinow Theorem tells that completeness implies any p, q ∈ M can
be connected by a shortest geodesic. Hence d(p, q) ≤ √πk , ∀p, q ∈ M
Covering spaces and Fundaental groups

A continuous map π : X −→ M is called a covering map if each p ∈ M has a
neighborhood U with the property that each connected component of π−1 (U) is mapped
homeomorphically onto U.
FACT 1: Let M be a differential manifold .X has a natural differentiable structure
s.t. π : X −→ M is a C ∞ and locally diffeomorphism.
Let {(Uα , xα )} be a differentiable structure of M . Uα small s.t. π−1 (Uα ) are disjoint
open sets of X,each connected component Uαi ⊂ X , we assign coordinate map xα ◦ π
(note π : Uαi −→ Uα is homeomorphism)
This leads to a differentiable structure for X, under which π is C ∞ and locally
diffeomorphism.
π−1 (Uα )
xα ◦π
π
#
Uα / Rn
xα
FACT 2 Let (M, g) be a Riemannian manifold . Note π is surjective . We can assign

g = π∗ g a Riemannian metric for X . Then π : (X,e
by e g −→ (M, g)) is a locally isometry
.
FACT 3. If (M, g) is complete , then (X,e
g) is also complete,
Proof. Suppose γ be a normal geodesic on (X,e g) with the maximal interval [0, b), b <
∞
Since π is locally isometry, we have π(γ) : [0, b) −→ M is a geodesic of (M, g).
Since (M, g) is complete , we have the geodesic π(γ) in (M, g) can be extended to
π(γ)(b) := o ∈ M
Pick
a small normal neighborhood U of p in M. Then ∃a < b s.t. π(γ)(a) ∈ U and
π(γ) ∈ U. Let Ui be the connected component of π−1 (U) containing γ(a).
[a,b]
Then the isometry π−1 : U −→ Ui maps a geodesic to a geodesic. Hence the
geodesic γ van be extended over b in Ui . This contradicts to the maximality of b
The equivalene or homotopy classes of closed curves with fixed base point p ∈ M
form a group π1 (M, p), the fundamental group of M with base point p.
π1 (M, p) and π1 (M, q) are isomorphic for any p, q ∈ M. Hence , it make sense to
speak of the fundamental group π1 (M) of M without reference to a base point.
Let π : X −→ M be a covering map . A deck transformation is a homeomorphism
ϕ : X −→ X with π = π ◦ ϕ
FACT 4 A deck transformation ϕ of (X,e

g) is an isometry .
Proof. Since π is locally isometry, and π = π ◦ ϕ , we know ϕ is locally isometry. Since

ϕ us homeomorphic , we have ϕ is an isometry.
A deck transformation which has a fixed point is the identity.

If π : M
e −→ M be the universal covering of M . π(x0 ) = p0 ∈ M
(1)π1 (M, p0 ) is in 1-1 correspondance with π−1 (p0 ).
x1 ∈ π−1 (p0 ) corresponds to the homotopy class of π(γ x1 ) where γ x1 (0) = x0 , γ x1 (1) =
x1 .
(2) The set D of all deck transformation is in 1-1 correspondance with π1 (M, p0 ).
Associate each deck transformation ϕ with ϕ(p0 ) ∈ π−1 (p0 ) So muc for the general
facts. Let’s come back to our discussion about Bonnet-Myers Theorem and Synge
Theorem .
We have shown if (M, g) is a complete Riemannian manifold with Ricci curvature
Ric ≥ (n − 1)k > 0.Then
π
diam(M, g) ≤ √
k
(Corolarry 4.7) and hence , in particular (M, g) is compact .
(The last assertion follows from Hopf-Rinow (The whole manifold is a bounded
closed set))
Moreover . (M, g) has finite fundamental group .
Proof for the last statement:
Let π : M e −→ M be the universal covering . From our previous discussion, ( M,e
e g)
is a C Riemannian manifold , and π : ( M,e
∞ e g) −→ (M, g) is a locally isometry. Hence
e g) is also bounded from below by (n − 1)k
the Ricci curvature of ( M,e
Moreoveer , (M, g) is complete =⇒ ( M,e e g) ≤ √π
e g) is complete. Hence diam( M,e
k
π
and ( √k ) is compact.
Then ∀p ∈ M, π−1 (p) is finite . Since otherwise , π−1 (p) has an accumulated point
p∈M
e e , and π is not a locally diffeomorphism . Therefore , the fundamental group is
finite.
Extension: Cheeger-Gromoll [JDG,1971] Ric ≥ 0,positive at one point ,thn π1 (M)
is finite. Next , we discuss Synge Thoerem further.
Firstly , recall the remarks on 4.3 that ”orientable”,”even-dimensional” are all nec-
essary.
By Bonnet-Myers, the ”compactness” can be replaced by the assumption that M is
complete and has sectional curatures bounded away from 0 .
In fact , the ”compactness” alone , due to a theorem of Gromoll-Meyer.
Theorem 4.7. (Gromoll-Meyer On complete open manifolds of positive curvature ,

Ann of Math,go(1639),75-90)
If M is a connected , complete , non-compact n -dimensional manifold with all
sectional curvatures positive, then M is diffeomorphic to Rn .
What happens if M is not orientable?

We give a proof of Corollary 4.6
Proof. Every non-orientable differential manifold M has an orientable double cover M̄

:
A brief description : At each point p ∈ M , the T p M can be separated into two
disjoint sets : Recall the orientation of a vector space, two bases are equivalent if their
transformation matrix has determinant > 0 . This is an equivalente relation . Let O p be
the quotient space of T p M w.r.t. the equivalence relation. O p ∈ O p will be called an
orientation of T p M
M̄ = {(p, O p ) : p ∈ M, O p ∈ O p }
M̄ has a natural differentiable structure s.t. π : M̄ −→ M is C ∞ and surjective .

∀p ∈ M, ∃U ∈ M, p ∈ U s.t.
π−1 (U) = V1 t V2
π : Vi −→ U is a diffeomorphism [dC. Chap 0, Ex12 ]

Example:S 2 is the orientable double cover if P2 (R) .
Then by our previous discussion , M̄ is orientable , and compact , even-dimensional
nd positive sectional curvature. Hence M̄ is simply connected. Therefore π1 (M) = Z2

What happens if not ”even-dimensional”

Theorem 4.6
For that purpose, we prove a more general result
Theorem 4.8. (Weinstein 1968) Let f be an isometry of a compact orientable Rieman-

nian manifold M n . Suppose that M has positive sectional curvature , and f reverses
the orientatiion of M and n is odd. Then f has a fixed point .
Proof. Suppose to the contrary , f (q) , q, ∀q ∈ M

Let p ∈ M such that d(p, f (p)) attains the minimum
d(p, f (p)) = inf d(q, f (q))(We use M is compact)

q∈M
Since M is compact =⇒ complete, ∃ a normal minimizing geodesic
γ : [0, l] −→ M,
γ(0) = p, γ(l) = f (p) and l = d(p, f (p))

Claim: ( f ◦˙ γ)(0) = γ̇(l)

proof of claim:
Let p0 = γ(t0 ), t0 , 0, t0 , l, f (p0 ) = f ◦ γ(t0 )
We have
d(p0 , f (p0 )) ≤ d(p0 , f (p0 )) + d( f (p), f (p0 ))

= d(p0 , f (p0 )) + d(p, p0 ) (S ince f is isometry)
= d(p, f (p)) (S ince γ is minimal)
Then by d(p, f (p)) = inf q∈M d(q, f (q)), we know the ”≤ ” is an ”=” i.e.,d(p0 , f (p0 )) =
d(p0 , f (p)) + d( f (p), f (p0 ))
That is the curve γ 0 ∪ f ◦ γ 0 is a shortest curve and hence a geodesic.

[t ,l] [0,t ]
In particular , this implies ( f ◦˙ γ)(0) = γ̇(l).
Next consider P−1 γ,0,l ◦ d f p : T p M −→ T p M
Then it is an isometry and hence , an orthogonal transformation.
Note d f p (γ̇(0)) = ( f ◦˙ γ)(0), (since f (p) = f ◦ γ(0))
We have
γ,0,l ◦ d f p )(γ̇(0)) = Pγ,0,l (( f ◦ γ)(0))

(P−1 −1 ˙
= P−1
γ,0,l (γ̇(l)) = γ̇(0)
That is P−1γ,0,l ◦ d f p has eigenvalue +1 with multiplicity ge1.

SInce Pγ,0,t preserves orientation and f reverse the orientation , we have det(P−1
γ,0,l ◦
d f p ) = −1
List all its eigenvalues as
λ1 , λ̄1 , ..., λ j , λ̄ j , −1, −1, · · · , −1, 1, 1, · · · , 1

| {z } | {z }
l k
n odd ⇒ k + l odd
) )
k even
We have ⇒ ⇒k>2
det = −1 ⇒ l odd k>1

⇒ ∃V ∈ T p M, < V, γ̇(0) >= 0, and P−1
γ,0,l ◦ d f p (V) = V
i.e. Pγ,0,l V = d f p (V)
Define V(t) = Pγ,0,t V,

We have F(t, s) = expγ(t) (sV(t)), s ∈ (−, ), t ∈ [0, l] is a variation of γ ,
with
F(t, 0) = γ(t)
F(0, s) = β(s)
F(l, s) = f ◦ β(s)
∂F
and = V(t)
∂s s=0
By the (SVF), we have
Z b
d2
= h∇T V, ∇T Vi − hR(V, T )T, Vi dt < 0

E(s)
ds2 s=0

a
This shows that ∃ small enough s,s.t. the curve γ s has the property L(γ s )2 ≤
2lE(γ s ) < 2lE(γ) = L(γ)2
Hence let p s = γ s (0), we have
d(p s , f (p s )) ≤ L(γ s ) < L(γ) = d(p, f (p))

which contradicts to the minimality of d(p, f (p))

Proof of Theorem 4.6
Suppose M is not orientable , let M̄ be the orientable double cover of M. Then
( M̄, π∗ g) is a compact orientable manifold with positve sectional curvature . Let ϕ be a
deck transformation of M̄ with ϕ , id.
Because M is not orientable ,ϕ is an isometry which reverse the orientation of M̄
. Since n is odd, we can apply Weinstein’s theorem to conclude ϕ has a fixed point .
Therefore ϕ = id, which is a contradicition.
Exercise 4.2. (1)Prove Weinstein Theorem for even-dimensional case: Let f be an

isometry of a compact orientable Riemannian manifold M n . Suppose M has positive
sectional curvature , and f preserve the orientation of M and n is even. Then f has a
fixed point.
(2) Prove Synge Theorem (even-dimensional) as a corollary.
Chapter 5
Space forms and Jacobi fields
We start our further investigation on geometry and topology of Riemannian manifolds

by studying the simplest cases: complete Riemannian manifolds with constant sec-
tional curvature, which are called space forms. We again will study the behavior of
geodesics in order to reveal the underlying geometry and topology.
The first problem we’re concerned about space forms is the existence. Recall if
a Riemannian manifold (M, g) has constant sectional curvature k, then (M, λg) has
constant sectional curvature λk for λ > 0. Therefore, we only need to consider space
forms with sectional curvature 0,+1,-1.
Obviousely, Rn with the Euclidean metric has 0 sectional curvature.(For example,
by the formula in local coordinate:
∂ ∂ ∂ ∂ 1 ∂2 g jk ∂2 gik ∂2 g jl ∂2 gil p m p
hR( , ) , i = ( − − + ) + gmp (Γm
il Γ jk − Γ jl Γik )
∂xi ∂x j ∂xl ∂xk 2 ∂xi ∂xl ∂x j ∂xl ∂xi ∂xk ∂x j ∂xk
recall from the previous discussion. For Rn , Rkli j = 0, ∀i, j, k, l.) Hence Rn is a space
form with sectional curvature 0. In fact, we have the following result:
Theorem 5.1. For any c ∈ R and any n ∈ Z+ , there exists a unique(upto isometries)
simply connected n-dimensional space form with constant sectional curvature c.
In order to discuss the existence for the other two cases c = +1 or −1, we first
introduce same useful ideas.
5.1 Isometries and totally geodesic submanifold

Let (M, g), (M, g) be two Riemannian manifolds, and f : M → M be an immersion. If
f ∗ g = g, then we say f is a isometric immersion and M is called the Riemannian embedded submanifold,
or regular submanifold.
Let dim M = n, dim M = n + k, we say M has codimension k in M. In particular, if
k = 1, M is called a hypersurface in M.
113
114 CHAPTER 5. SPACE FORMS AND JACOBI FIELDS
Definition 5.1. (totally geodesic submanifold). Let M L

be a submanifold of M. We
identify p ∈ M with f (p) ∈ M. Then T p M = T p M T p⊥ M, where T p⊥ M is the
orthonormal complement of T p M in T p M.
M is called a totally geodesic submanifold if ∀ geodesic γ in M with γ(0) ∈ M, γ̇(0) ∈

T p M, we have γ ⊂ M.
Remark 5.1. Recall from the Final Remark of our discussions about Levi-Civita con-
nection, we know for the Levi-Civita connection ∇ and ∇ for M and M respectively, we
heve
∇γ̇ γ̇ = 0 ⇒ ∇γ̇ γ̇ = 0.
That is, γ is also a geodesic in M.
There is a characterization of totally geodesic submanifold by the second fundamental form.
In fact, the decomposition T p M = T p M
L ⊥
T p M is differentiable, and, hence, the
tangent bundle T M = T M
L
N M where N M is the normal bundle.
For any X, Y ∈ Γ(T M), define
B(X, Y) = ∇X Y − ∇X Y ∈ Γ(N M).
First, we observe that ∀function f on M, we have
B( f X, Y) = f B(X, Y) (easy)

(5.1.1)

B(X, f Y) = f B(X, Y)


We also have B(X, Y) = B(Y, X).(usingtorsion− f reeproperty.) B is called the second fundamental form
of the submanifold M in M.
Theorem 5.2. M is a totally geodesic submanifold of M if and only if B ≡ 0.
Proof. Due to the property (5.1.1), we can speak of the map for all p
B : T p M × T p M → Np M
which is bilinear and symmetric. Let M be a totally geodesic submanifold of M, then

∀V ∈ T p M, let γ be the geodesic in M with γ(0) = p, γ̇(0) = V. ∇γ̇ γ̇ = 0, then we have
∇γ̇ γ̇ = 0.
That is
∇γ̇ γ̇ − ∇γ̇ γ̇ = ∇v v − ∇v v = B(v, v) = 0.
Since B is bilinear and symmetric, we have
B(v, v) = 0, ∀v ∈ T p M,
Conversely, suppose B ≡ 0. Then ∀p ∈ M, V ∈ T p M, let γ be a geodesic in M with

γ(0) = p, γ̇(0) = V. Let ξ be the geodesic in M. Due to the uniqueness of the geodesic
with initial data γ(0), γ̇(0), we conclude ξ = γ. Hence γ ⊂ M.
5.1. ISOMETRIES AND TOTALLY GEODESIC SUBMANIFOLD 115
Remark 5.2. Notice that by ∇γ̇ γ̇, ∇γ̇ γ̇ we are actually using the induced connection,
One can check B ≡ 0 ⇒ ∇ξ̇(t) ξ̇
Totally geodesic submanifold can be considered as a generalization of geodesics.
Rk ,→ Rk+1 is a totally geodesic submanifold but S 2 ⊂ R3 is not.
we have
Proposition 5.1. Let M be a totally geodesic submanifold of M, denote by K and K

for their sectional curvature respectively. Then every 2-dim sections π p T p M and any
p ∈ M, we have K(π p ) = π p .
Proof. By definition and Theorem 6.2.9.
Next, we give a relation between isometry and totally geodesic submanifold.
Theorem 5.3. [WSY, p59, Lemma 3] Let f : (M, g) → (M, g) be an isometry. Then
every connected component of the fixed point set M = Fix( f ) = {p ∈ M| f (p) = p} is
totally geodesic submanifold.
Proof. Observe that Fix( f ) is a closed subset: It is the preimage of the diagonal in
M × M under the differentiable mapping p 7→ (p, f (p)) ∈ M × M. Let p ∈ Fix( f ). If p
is not isolated, consider H = {v ∈ T p M.
Let δ be small enough s.t.
exp : B(0, δ) ⊂ T p M → B p (δ) ⊂ M

p
is a diffeomorphism.
Claim: exp p (H B(0, δ)) = M B p (δ). This claims implies immediately that M
T T
is submanifold of M.
Proof of the claim:
(1) ∀q ∈ M B p (δ), choose V ∈ B(0, δ) ⊂ T p M, s.t. exp p V = q and γ : [0, 1] →
T
M, γ(t) = exp p tV is the unique shortest geodesic. By our previous discussion, V ∈ H.

Hence M B p (δ) ⊂ exp p (H B(0, δ)).
T T
(2) Let V ∈ H B(0, δ), let q = exp p V. Let γ : [0, 1] → M be the geodesic
T
γ(t) = exp p tV, then γ(0) = p, γ(1) = q. Then f ◦ γ is also a geodesic with ( f ◦ γ)(0) =
f (p) = p. Moreover ( f ◦˙ γ)(0) = d f p (γ̇(0)) = d f p (V) = V = γ̇(0), then by uniqueness,
f ◦ γ = γ, and in particular
f (q) = f ◦ γ(1) = γ(1) = q.
Hence exp PV ⊂ B p (δ) M i.e. exp(H B(0, δ)) = B p (δ) M. This complete the
T T T
proof of the claim.
The above arguement (2) also tells that any geodesic γ in M with γ(0) ∈ M, γ̇(0) ∈
T γ(0) M satisfies f (γ) = γ. Hence M is a totally geodesic submanifold M.
5.2 Space forms

We continue the discussion about the existence of sapce forms with sectional curvature
+1 or -1.
Example 5.1. S 2 ⊂ R3 with the induced metric of the Euclidean metric of R3 . Since
x = r cos ϕ cos θ




y = r cos ϕ sin θ

(5.2.1)




 z = r sin ϕ

g|S 2 = (dx2 + dy2 + dz2 )|S 2 = dϕ2 + cos2 ϕdθ2 .(r = 1) Then
∂ ∂ ∂ ∂ ∂ ∂ 2
h , ih , i − h i, = cos2 ϕ.
∂θ ∂θ ∂ϕ ∂ϕ ∂θ ∂ϕ
∂ ∂ ∂ ∂
hR( , ) , i
∂ϕ ∂θ ∂θ ∂ϕ
1 p m p
= (gθϕ,ϕθ − gϕϕ,θθ − gθθ,ϕϕ + gϕθ,θϕ ) + gmp (Γm
ϕθ Γθϕ − Γθθ Γϕϕ )
2
cos ϕ sin ϕ 1
(check Γϕϕθ = Γθθθ = Γϕϕϕ = 0, Γθϕθ = − .) = − gθθ,ϕϕ + gθθ Γ( ϕθ)θ Γθθϕ = cos2 ϕ
cos2 ϕ 2
⇒ sectionalcurvature K ≡ 1.
Proposition 5.2. The unit sphere S n ⊂ Rn+1 (n ≥ 2) has constant sectional curvature
+1.
Proof. n = 2 has been checked in Example 1.
When n ≥ 3, define an isometry f : Rn+1 → Rn+1 as below
f : (x1 , x2 , x3 , −x4 , · · · , xn+1 ) = (x1 , x2 , x3 − x4 , · · · , −xn+1 ).
It induce an isometry f : S n → S n .
3 2
Observe the set of fixed point of f : S n → S n = {(x1 , x2 , x3 , 0, · · · , c)| xi = 1} =
P
i=1
S 2 . Therefore, S 2 is a totally geodesic submanifold of S n . Since sectional curvature of
S 2 is 1, we have S n has sectional curvature K(π p ) = 1 for some π p ⊂ T p S 2 . For any
π0q ⊂ T q S n . Suppose π p = {e1 , e2 }, the positive vector of p be en+1 , π0q = span{e01 , e02 },
the positive vector of q be e0n+1 . First let rotate φ in span{en+1 , e0n+1 } be s.t. φ(en+1 ) =
e0n+1 and dφ(π p ) = πq . Then let φ0 be the ratation which fix q and send πq to π0q . Then
the isometry φ0 ◦ φ send p to q, and π p ⊂ T p S n to π0q ⊂ T q S n . Hence K(π0q ) = 1.
Proposition 5.3. The unit ball Bn = {x ∈ Rn |kxkh1} ⊂ Rn with the hyperbolic metric
4 X
g= P i22 dxi ⊗ dxi
(1 − (x ) ) i
i
is a space form with constant sectional curvature -1.

5.2. SPACE FORMS 117
s
Proof. First, we show (Bn , g) := H n is complete. Consider the curve γ(s) := ( ees −1
+1 , 0, · · · , 0).
We compute
4 ∂ es − 1 2
hγ̇(s), γ̇(s)ig = ( ( ))
(1 − ( e s −1
e s +1)2 )2 ∂s e s + 1
4 e s (e s + 1) − (e s − 1)e s 2
= 4es ( )
( (es +1)2 )2 (e s + 1)2
(e s + 1)4 (2e s )2
= =1
4e2s (e s + 1)4
That is γ is paramertrized by arc length.

Observation: Any orthonormal transformation of Rn induces an isometry (Bn , g) →
(B , g).
n
Let f : Bn → Bn be the isometry induced by (x1 , x2 , · · · , xn ) 7→ (x1 , −x2 , · · · , −xn ).

Note Fix( f ) = γ((−∞, ∞)). By Theorem 6.2.10, γ is a geodesic. Use the Observation
againm A(γ) is a geodesic for any isometry A induced by a orthonormal transformation
Rn → Rn . That is all geodesic starting from 0 can be defined on [0.∞). We conclude
the completeness by Hopf-Rinow Theorem.
Next,we show Hn has constant sectional curvature -1: i.e. ∀p ∈ Bn , ∀ 2-dim section
π p ⊂ T p Bn , we have to show K(π p ) = −1.
Let ~p be the position vector of p. Identifying T p Bn with Rn . Let E be a 3-dimensional
linear subspace of Rn containing ~p and π p . If ~p ∈ π p , or ~p = 0, the choice of E is not
unique. The reason we have to consider such a 3 − dim subspace: There is no ob-
vious way to say Rn is homogeneous, i.e. ∀p, q ∈ Bn , ∃ isometry f : Bn → Bn s.t.
f (p) = f (q). However, we will show this later.
Let Rn = E E , let f : Bn → Bn be the isometry induced by the orthonormal
L ⊥
transformation
(e, e1 ) 7→ (e, −e1 ), e ∈ E, e⊥ ∈ E ⊥ .
Observe Fix( f ) = E Bn . Use the Observation again, choose orthonormal transfor-

T
mation A s.t. A(E) = {(x1 , x2 , x3 , 0, · · · , 0)} ⊂ Rn . A induce an isometry Bn → Bn .
Hence, it remains to show B3 with the hyperbolic metric has constant sectional curva-
ture -1.
Use the spherical coordinate {ρ, ϕ, θ} on B3 \ {0}, the hyperbolic metric can be writ-
ten as
4
(dρ2 + ρ2 dθ2 + ρ2 cos2 θdϕ2 )
(1 − ρ2 )2
where dρ2 := dρ ⊗ dρ and, similarly, dθ2 , dϕ2 .

Consider vector fields
1 − ρ2 ∂ 1 − ρ2 ∂ 1 − ρ2 ∂
X1 = , X2 = , X3 = .
2 ∂ρ 2ρ ∂θ 2ρ cos θ ∂ϕ
Then we have hXi , X j i = δi j . We calculate
1 − ρ2 ∂ 1 − ρ2 ∂ f 1 − ρ2 ∂ 1 − ρ2 ∂ f
[X1 , X2 ]( f ) = ( )− ( )
2 ∂ρ 2ρ ∂θ 2ρ ∂θ 2 ∂ρ
1 − ρ2 ∂ 1 − ρ2 ∂ f 1 − ρ2 1 − ρ2 ∂2 f 1 − ρ2 1 − ρ2 ∂2 f
= ( ) + −
2 ∂ρ 2ρ ∂θ 2 2ρ ∂ρ∂θ 2ρ 2 ∂θ∂ρ
1 − ρ2 ∂ 1 − ρ2 ∂
= ( ) f
2 ∂ρ 2ρ ∂θ
1 − ρ2 ∂ 1 − ρ2 ∂ 1 − ρ2 −(ρ2 + 1) ∂
⇒ [X1 , X2 ] = ( ) = .
2 ∂ρ 2ρ ∂θ 2 2ρ2 ∂θ
∂ 1 − ρ2 −2ρ(2ρ) − 2(1 − ρ2 ) −(ρ2 + 1)
( ( )= = )
∂ρ 2ρ 4ρ2 2ρ2
1 + ρ2
⇒ [X1 , X2 ] = − X2 . (1)
2ρ
Similarly,
1 − ρ2
[X2 , X3 ] = + tan θX 3 (2)
2ρ
−(ρ2 + 1)
[X1 , X3 ] = − X3 (3)
2ρ2
Recall for orthonormal vector fields X, Y, Z, we have by koszul formula
2h∇X Y, Zi = −hX, [Y, Z]i + hY, [Z, X]i + hZ, [X, Y]i.
Employing (1),(2), and (3), we have
∇X1 X1 = ∇X1 X2 = ∇X1 X3 = 0

ρ2 + 1
∇X2 X2 = − X1 , ∇X2 X3 = 0
2ρ
ρ2 + 1 1 − ρ2
∇X3 X3 = − X1 + tan θX2 .
2ρ 2ρ
By torsion-free property (∇X Y − ∇Y X) = [X, Y], the above information is enough to

calculate the sectional curvature.
K(X1 , X2 ) = hR(X1 , X2 )X2 , X1 i





K(X2 , X3 ) = hR(X2 , X3 )X3 , X2 i

(5.2.2)




K(X , X ) = hR(X , X )X , X i

1 3 1 3 3 1
5.2. SPACE FORMS 119
R(X1 , X2 )X2 =∇X1 ∇X2 X2 − ∇X2 ∇X1 X2 − ∇[X1 ,X2 ] X2

ρ2 + 1
=∇X1 (− X1 ) − ∇− 1+ρ2 X X2
2ρ 2ρ 2
1 − rho ∂ ρ + 1
2 2
1 + ρ2 ρ2 + 1
= (− )X1 + (− )X1
2 ∂ρ 2ρ 2ρ 2ρ
∂ ρ2 + 1 2ρ 2ρ − 2(ρ2 + 1) 2ρ2 − 2 ρ2 − 1
( )= = =
∂ρ 2ρ 4ρ2 4ρ2 2ρ2
1−ρ ρ −1
2 2
(ρ + 1)
2 2
⇒ R(X1 , X2 )X2 = − X1 − X1
2 2ρ2 4ρ2
(ρ2 − 1)2 − (ρ2 + 1)2
= X1 = −X1 .
4ρ2
Hence K(X1 , X2 ) = −1.
R(X1 , X3 )X3 = ∇X1 ∇X3 X3 − ∇X3 ∇X1 X3 − ∇[X1 ,X3 ] X3

ρ2 + 1 1 − ρ2
=∇X1 (− X1 + tan θX2 ) − ∇ 1+ρ2 X X3
2ρ 2ρ 2ρ 3
ρ2 + 1 1 − ρ2 ρ2 + 1
=X1 (− )X1 + X1 ( tan θ)X2 + ∇X3 X3
2ρ 2ρ 2ρ
1 − ρ2 ρ2 − 1 1 − ρ2 ρ2 + 1 1 + ρ2 ρ2 + 1 1 + rho2 1 − ρ2
=− 2
X1 − 2
tan θX2 − X1 + tan θX2
2 2ρ 2 2ρ 2ρ 2ρ 2ρ 2ρ
(ρ2 − 1)2 − (ρ2 + 1)2
= X1 = −X1
4ρ2
Hence K(X1 , X3 ) = −1.
R(X2 , X3 )X3 = ∇X2 ∇X3 X3 − ∇X3 ∇X2 X3 − ∇[X2 ,X3 ] X3

ρ2 + 1 1 − ρ2
=∇X2 (− X1 + tan θX2 ) − ∇ 1−ρ2 tan θX X3
2ρ 2ρ 2ρ 3
1−ρ ∂ ρ +1
2 2
1−ρ ∂ 1−ρ
2 2
= (− )∇X2 X1 + ( tan θ)X2
2ρ ∂θ 2ρ 2ρ ∂θ 2ρ
1 − ρ2 1 − ρ2
+ tan θ∇X2 X2 − tan θ∇X3 X3
2ρ 2ρ
ρ2 + 1 ρ2 + 1 (1 − ρ)2 1
=− X2 + X2
2ρ 2ρ 4ρ2 cos2 θ
1 − ρ2 ρ2 + 1 1 − ρ2 ρ2 + 1
− tan θ X1 + tan θ X1
2ρ 2ρ 2ρ 2ρ
ρ2 + 1 ρ2 + 1
− tan θ tan θX2
2ρ 2ρ
(ρ2 + 1)2 (1 − ρ2 )2 1
=− X 2 + [ 2 − tan2 θ]X2
4ρ2 4ρ2 cos θ
(ρ2 − 1)2 − (ρ2 + 1)2
= X2 = −X2 .
4ρ2
Hence K(X2 , X3 ) = −1
5.3 Geodesics in Rn , S n , Hn
Since R2 , S 2 , H2 are totally geodesic submanifold of Rn , S n , Hn , respectively, we
only need to consider geodesics in R2 , S 2 .H2 .
Let us measure the convergence/divergence properties of geodesic emanating from

a reference point 0 by the length of the circle
c(r) := {x ∈ M : d(0, x) = r.
When r is small eougn, c(r) is the image of S (0, r) ⊂ T 0 M under the diffeomorphism
exp0 , i.e. c(r) = exp0 S (0, r).
5.3. GEODESICS IN RN , S N , HN 121
Let c0 (r), c+ (r), c− (r) be the length of c(r) in R2 , S 2 , H 2 , respectively.

(1) c0 (r) = 2πr is linear in r.
(2) c+ (r), i.e. M = S 2 .
(3) c− (r), i.e. M = H 2 .(Here, we choose 0 to be the centre of the disc. Later,
we will see c− (r) does not depend on the choice of 0.) Recall the normal geodesic
emanating from 0 is given by
es − 1
γ : [0, ∞) → B2 , s 7→ p, ∀p ∈ ∂B2 .
es + 1
Then
Z 2π
2
c− (r) = ρdθ|ρ= err −1
0 (1 − ρ2 ) e +1
er −1
2 er +1 e2r − 1 er − e−r
= r 2π = 2π = 2π
1− ( eer −1
+1 )
2 2er 2
⇒ c− (r) = 2π sinh r.
We see that c− (r) grows much faster than c0 (r).
In the above 3 particular cases, we see the sign of the curvature is closely related to
the behavior of geodesic. What happens in genernal?
In order to answer this question, we consider the quantity c(r) for a Riemannian
manifold (M, g). Let 0 ∈ M, and δi0 be a small number such that exp0 is a diffeomor-
phism on B(0, δ) ⊂ T 0 M.
Consider the polar coordinate (ρ, θ) in T 0 M. Then for any fixed r, er(θ) = (r, θ) is a
d
curve in T 0 M. dθ (r, θ) is the velocity field along e
r(θ).
Let r < δ, we have
Z 2π
d d
c(r) = hd exp( (r, θ)), d exp( (r, θ)) > dθ
0 0 dθ 0 dθ
d
So, for our purpose, we have to explore the interaction between the norm of d exp0 ( dθ (r, θ))
and the curvature of (M, g). Note that R T 0 M 3 ~p = (r, θ p ), if we write
n
d d
d exp( (r, θ)) = d exp(~p)( ).
0 dθ 0 dθ
In order to calculate its norm, we first observe it can be extended to be a variational

field of a geodesic variation of
t
γ(t) = exp ~p, t ∈ [0, r].
0 r
In fact, we pick
t d
F : [0, r] × (−, ) → M, (t, s) 7→ exp (~p + s ).
0 r dθ
∂F ∂
We observe that F(t, 0) = γ(t), and ∂s (t, 0) = ∂s | s=0 exp0 rt (~p + s dθ
d
) is the variational
field along γ. In particular
∂F ∂ d
(r, 0) = | s=0 exp(~p + s )
∂s ∂s 0 dθ
d
=d exp(~p)( )
0 dθ
d
=d exp( (r, θ)).
0 dθ
In order to calculate ∂F ∂F
∂s (r, 0), we calculate the whole variational field. V(t) = ∂s (t, 0), t ∈
[0, r].
Here, we can be slightly more general: consider a general vector X ∈ T ~p (T 0 M) and
the variation
t
F(t, s) = exp (~p + sX), t ∈ [0, r], s ∈ (−, ).
0 r
∂F
Let V(t) := ∂s (t, 0) be the geodesic variational field along γ.
To calculate V(t), t ∈ [0, r], we derive the equations it satisfies:
∂F e e ∂F
∇∂e
e ∇∂ =∇ ∂ ∇ ∂
∂t ∂t ∂s ∂t ∂s ∂t
∂F e e ∂F e e ∂F
=e
∇∂e ∇∂ +∇∂∇∂ −∇∂ ∇∂
∂s ∂t ∂t ∂t ∂s ∂t ∂s ∂t ∂t
∂F ∂F ∂F
=R( , ) .
∂t ∂s ∂t
Restricting to the normal geodesic γ, we have
∇T ∇T V = R(T, V)T, or ∇T ∇T V + R(V, T )T = 0 (*)
Definition 5.2. (Jacobi field), Let γ : [a, b] → M be a geodesic, and T be the velocity
field along γ. If a vector field V along γ satisfies
∇T ∇T V + R(V, T )T = 0, (5.3.1)
we call V a Jacobi field(along γ). The equation (5.3.1) is called the Jacobiequation.
5.3. GEODESICS IN RN , S N , HN 123
Choose parallel vector fields Y1 , · · · , Yn along γ which are orthonormal at γ(a), and
hence orthonormal everywhere along γ, then ∃ f i (t) s.t. V(t) = f i (t)Yi (t), and
d2 f i (t)
∇T ∇T V + R(V, T )T = Yi + f i R(Yi , T )T = 0
dt2
d2 f i (t)
⇔h Yi , Y j i + h f i R(Yi , T )T, Y j i = 0, ∀ j = 1, · · · , n.
dt2
d2 f j
⇔ 2 + f i (R(Y)i, T )T, Y j ) = 0, ∀ j = 1, · · · , n.
dt
Hence, V(t) the solution of the above system of second order linear ODE. It will be
∇T V ∈ T 0 T γ(0) M. Recall
determined by its initial conditions V(0) and V̇(0) := e
∂
V(0) = | s=0 F(t, s)|t=0
∂s
∂ ∂
= | s=0 F(0, s) = | s=0 exp 0 = 0
∂s ∂s 0
∂F ∂F
V̇(0) =∇T V(0) = e ∇∂ | s=0 (0, s) = e
∇∂ (0, 0)
∂t ∂s ∂s ∂t
1
=e
∇ ∂ (~p + sX).
∂t r
Note 1r (~p + sX) is a vector field along the constant curve. By definition of induced
connection, we have
1 X
V̇(0) =e ∇ ∂ ( (~p + sX)) = ∈ T 0 (T 0 M)
∂s r r
X X
i.e. f˙i (0)Yi (0) = = h , Yi (0)iYi (0)
r r
X
i.e. f˙i (0) = h , Yi (0)i.
r
So, in order to solve V(t), where V(r) = d exp0 (~p)(X) is what we need, we have solve
f¨ + hR(Yi , T )T, Y j i f i = 0, j = 1, · · · , n
 j




 f j (0) = 0




 (5.3.2)

 1
 f˙ j (0) = hX, Y j (0)i



r
Next, we come back to the calculation of c(r):
Z 2π
d d 1
c(r) = hd exp( (r, θ)), d exp( (r, θ))i 2 dθ
0 0 dθ 0 dθ
where
f¨ + hR(Yi , T )T, Y j i f i = 0, j = 1, · · · , n
 j




 f j (0) = 0




 (5.3.3)

 1 d
 f˙ j (0) = h (r, θ), Y j (0)i



r dθ
In particular, for the cases of R2 , S 2 , H2 . Let Y(t) be a unit parallel vector field along
γ s.t. hY(t), T (t)i = 0, ∀t.
By Gauss’s lemma, dθ d
(t, θ), t ∈ [0, r] is vertical to the radial geodesic at everywhere
t. In fact, the variational field V(t) along γ is perpendicular to γ everywhere.(by First
variational formula.) So, we can write
V(t) = f (t)Y(t).
Then the equation (5.3.2) become
f + hR(Y, T )T, Yi f = 0
¨



 f (0) = 0



 (5.3.4)
1 d 1 d


 f˙(0) = h (r, θ), Y(0)i = | (r, θ)| = 1



r dθ r dθ
Recall we have constant sectional curvature in R2 , S 2 , H2 . Therefore, we need solve
f¨ + K f (t) = 0





f (0) = 0

(5.3.5)




 f˙(0) = 1


The solution is given by

t, K = 0




f (t) =  sin t, K = +1

(5.3.6)



 sinh t, K = −1

Therefore, we recover the results:
c0 (r) = 2πr




c+ (r) = 2π sin r

(5.3.7)




c (r) = 2π sinh r

−
Therefore, (**) establish the relations between c(r) and the curvature.
5.4 What is a Jacobi field?

We have already seen the definition of Jacobi fields . Now, we want to understand this
concept further.
As we have explained, it is a solution of a system of second order ODE:
d2 f j
+ f i hR(Yi , T )T, Y j i = 0, j = 1, · · · , n.
dt2
and Y1 , · · · , Yn are parallel orthonormal vector fields along γ. And the Jacobi field is
given by V(t) = f i (t)Yi (t).
5.4. WHAT IS A JACOBI FIELD? 125
Proposition 5.4. Let γ : [a, b] → M be any geodesic.

(1) Given V, W ∈ T γ(a) M, there exists a unique Jacobi field U(t), t ∈ [a, b] such that
U(0) = V, e
∇ ∂ U(t) := U̇(0) = W.
∂t
(2) The linear space of all Jacobi fields along γ is of 2n dim’l.
(3) The zero points of a Jacobi field U along γ are discrete, if U is not identically 0
along γ.
Proof. (1),(2) follows directly from the theory of 2nd linear ODEs. Given U(0), U̇(0),
the 2nd order linear ODE has a unique solution.
For (3), assume the zero points are not discrete. Then there is an accermulated point
γ(t0 ). Then U(t0 ) = 0, and
∂ dfi ∂
U̇(0) = e
∇ ∂ ( f i (t) i ) = (t0 ) i + f i (t0 )e
∇ ∂ (t0 )
∂t ∂x dt ∂x ∂t
pick (xi ) to be the normal coordinate around t0 , then
dfi ∂
U̇(t0 ) = (t0 ) i = 0.
dt ∂x
Then U is identically zero along γ
From our discussion in section 5.3, we have seen that the variational dield odf a
geodesic variation of a geodesic γ is a Jacobi field along γ. In fact, the converse also
hold.
Proposition 5.5. Let γ : [a, b] → M be a geodesic and U be a vector field along γ.

Then U is a Jacobi field if and only if U is the variational field of a geodesic variation
of X.
Proof. (⇐) The calculations in the end of !!! proves this directions.
(⇒) Let U be a Jacobi field along γ. Let β : (−, ) → M be the geodesic with
β(0) = γ(0)


β̇(0) = U(0)
 (5.4.1)
We put
F : [0, b] × (−, ) → M, (t, s) 7→ exp t(V(s) + sW(s)).

β(s)
where V,W are parallel vector fields along β with
V(0) = γ̇(0), W(0) = U̇(0) = e

∇ ∂ U(0).
∂t
Then
F(t, 0) = exp tV(0) = exp tγ̇(0) = γ
β(0) γ(0)
and F(t, s) = expβ(s) t(V(s) + sW(s)) are all geodesics for s ∈ (−, ). That is F is a
geodesic variation of γ. Therefore its variational field
∂F ∂F
Y(t) =: (t, 0) = (t, s)| s=0
∂s ∂s
is a Jacobi field. Meanwhile, we have
∂ ∂
Y(0) = | s=0 F(0, s) = | s=0 exp 0 = β̇(0) = U(0)
∂s ∂s β(s)
and
∂
Ẏ(0) =e
∇ ∂ F(t, s)| s=0
∂t ∂s
∂
=e
∇ ∂ F(t, s)| s=0
∂s ∂t
=e
∇ ∂ (V(s) + sW(s))| s=0
∂s
=W(0) = U̇(0).
Then Proposition 5.4(1) implies that U = Y. thqat is U is the variational field of

F.
Remark 5.3. We summarize what we learned about Jacobi fields up to now:
(1) Let β : (−, ) → M be a curve, V(s), W(s) are parellel vector fields along β.
Then the family of geodesics
γ s (t) := exp t(V(s) + sW(s))

β(s)
leads to a geodesic variation F(t, s) := γ s (t) whose variational field along γ0 (t) is a
Jacobi field U(t) with
U(0) = β̇(0), U̇(0) = W(0).
In particular, when β(s) = p ∈ M is a constant curve, we have.
(2) The 1-parameter family of geodesics
γ s (t) = exp t(V + sW), V, W ∈ T p M

0
gives the Jacobi field U(t) along γ0 with
U(0) = β̇(0) = 0, U̇(0) = W.

5.4. WHAT IS A JACOBI FIELD? 127
Since T p M is an inner product space, we can restrict hV, Wi = 0. Then, we have

(3) The 1-parameter family of geodesics
γ s (t) = exp t(V + sW), hV, Wi = 0

p
givea a “normal Jacobi field” U(t) with U(0) = 0, U̇(0) = W, and hU(t), γt0 (t)i =
0, ∀t.
Observe htW, tVi = 0. Recall the Gauss lemma we derived from the First variation
formula, we have, since U(t) is the variational fields, hU(t), γ̇0 (t)i = 0.
We will se later that normal Jacobi fields with U(0) = 0 are very important for any
further investigation.
Relations with the SVF:
Recall for proper variations of a geodesic γ, we have
∂2 b
Z
|(0,0) E(v, w) = I(V, W) = h∇T W, ∇T Wi − hR(W, T )T, Vidt.
∂v∂w a
Observe that T (V, W) is bilinear. I(V, W) is also symmetric since E is C ∞ in (u, w).
Resturning to the original problem: to determine whether a geodesic is (locally)
∂2
minimizing. For that purpose, we hope to decide whether det( ∂v∂w )|(0,0) E(v, w) is
positive or not. We will see the existance of Jacobi field(vanishing at the two ends
γ(a), γ(b)) will be an obstruction.

Then U is a Jacobi field if and only if
I(U, Y) = 0
for all vector fields Y along γ with Y(a) = Y(b) = 0.
Proof.
Z b
I(U, Y) = h∇T U, ∇T Yi − hR(U, T )T, Yidt
a
Z b
= h−∇T ∇T U, Yi − hR(U, T )T, Yidt
a
since ∇ is compatible with g and Y(a) = Y(b) = 0.
Z b
= h−∇T ∇T U − R(U, T )T, Yidt
a
for all Y with Y(a) = Y(b) = 0.

Therefore ∇T ∇T U + R(U, T )T = 0 holds by the fundamental lemma of the calculus
of variatiuons.

Then U is a Jacobi field if and only if it is a critical point of I(X, X) w.r.t. all variations
with fixed endpoints, i.e.
d
| s=0 I(X + sY, X + sY) = 0
ds
for all vector fields Y along γ with Y(a) = Y(b) = 0
Proof. We compute
d
| s=0 I(X + sY, X + sY) == 2I(X, Y).
ds
Then Proposition 5.6 follows directly from Proposition 5.5.
Remark 5.4. The Jacobi equation is the Euler-Lagrange equation for I(X) = I(X, X).
In fact, one can consider the second variation for each critibal point os a variational
problem. The second variation then is a quadratic integral in the variation fields,
and the second variation may be considered as a new variational problem. This new
variational problem is called accessary variational problem of the original one.
5.5 Conjugate Points and ,Minimizing Geodesics

From Proposition 5.5, we see that if there exists nonzero Jacobi field U along the
geodesic γ : [a, b] → M with U(a) = U(b) = 0, then I(U, U) = 0, i.e. I is not positive
definite, and hence γ|[a,b] may not be strictly local minimizing. This phenomena can
be observed explicitly. For any semicircle from the north pole p to the south pole q, ∃
nonzero Jacobi field U along it with U(a) = U(b) = 0, each semicircle has the same
length π.
Definition 5.3. (Conjugate points) Let γ : [a, b] → M be a geodesic. For t0 , t1 ∈ [a, b],
if there exists a Jacobi field U(t) along γ that does not vanish identically, but satisfies
U(t0 ) = U(t1 ) = 0
then t0 , t1 are called conjudate values along γ. The multiplicity of t0 and t1 as con-
jugate values is defined as the dimensions of the vector space consisting of all such
Jacobi fields. We also say γ(t0 ), γ(t1 ) are conjugate points of γ.(This terminology is
ambiguous when γ has self-intersetions).
Recall a Jacobi field U is determined by its initial values U(t0 ), U̇(t0 ) at any point
to. Hence, the multiplicity of two conjugate values t0 , t1 is clearly ≤ n. Actually, it is
≤ n − 1. This is because a Jacobi field which is tangent to γ and vanish at to will not
vanish at t1 .
Proposition 5.8. Let γ : [a, b] → M be a geodesic with velocity field T (t) = γ̇(t).
(1) The vector field f T along γ is a Jacobi field if and only if f is linear.
5.5. CONJUGATE POINTS AND ,MINIMIZING GEODESICS 129
(2) Every Jacobi field U along γ can be written uniquely as
f T + U ⊥,
where f is linear and U ⊥ is a Jacobi field prependicular to γ.

(3) If a Jacobi field U along γ is prependicular to γ at two points t0 and t1 , then
U is prependicular to γ everywhere. In particular, if U(t0 ) = U(t1 ) = 0, then U is
perpendicular to γ everywhere.
Proof. (1) f T is a Jacobi field⇒
0 = ∇T ∇T ( f T ) = R( f T, T )T = f 00 (t)T.
Hence f is linear.
(2) Let U be a Jacobi field along γ, we can write U = f T + U ⊥ for some f and
some U ⊥ with hU ⊥ , T i = 0.
U is Jacobi⇒ 0 = ∇T ∇T ( f T + U ⊥ ) + R( f T + U ⊥ , T )T = f 00 T + ∇T ∇T U ⊥ +
R(U ⊥ , T )T.
In particular, we have
0 = f 00 T + h∇T ∇T U ⊥ , T i + hR(U ⊥ , T )T, T i.
By symmetry, hR(U ⊥ , T )T, T i = 0.

d ⊥
0 = hU ⊥ , T i ⇒0 = hU , T i = h∇T U ⊥ , T i
dt
d
⇒0 = h∇T U ⊥ , T i = h∇T ∇T U ⊥ , T i
dt
Hence 0 = f 00 and
∇T ∇T U ⊥ + R(U ⊥ , T )T = 0,
i.e. U ⊥ is a Jacobi field. Uniqueness is obvious.
(3) Write U = f T + U ⊥ . When hU(t0 ), T i = hU(t1 ), T i = 0 implies f (t0 ) = f (t1 ) =
0. Recall f is linear, we have f ≡ 0. Therefore U = U ⊥ .
Proposition 5.9(3) shows that for the purpose of investigating conjugate values, we
need consider only normal Jacobi fields.
Conjugate points play an important role in the study of local minima for length. A
geodesic γ : [a, b] → M can not locally minimizi length if ∃τ ∈ (a, b) conjugate to a.
Intuitive arguement:
γ(τ) conjugate to γ(a), ∃ a geodesic η from γ(a) to γ(τ) with nearly the same length
as γ|[0,τ] .
Then η followed by γ|[τ,b] has nearly the same length as γ. By the first curve has a
corner, and can be shorten by replacing the corner with a minimal geodesic. Therefore
γ is not a minimizing curve.
In fact we have the following theorem of Jacobi.
Theorem 5.4. (Jacobi) Let γ : [a, b] → M be a geodesic from p = γ(a) to q = γ(b).

(1) If there is no conjugate points of p along γ, then there exists > 0 so that for
any piecewise smooth curve c : [a, b] → M from p to q satisfying d(γ(t), c(t))h, we
have
L(c) ≥ L(γ)
with equality holds if and only if c is a reparametrization of γ.

(2) If there exists t ∈ (a, b) so that q = γ(t) is a conjugate point of p, then there is a
proper variation of γ so that
L(γ s ) < L(γ)
for any 0 < |s| < .
The above results are direct consequences of the corresponding properties of index
forms, which will be discussed in the next subsection,
Next, we derive a characterization of the conjugate points in terms of critical point
of the exponential map.
Theorem 5.5. Let γ : [0, 1] → M be a geodesic with γ(0) = p ∈ M and γ̇ = V ∈ T p M,

so that γ can be described as
t 7→ exp tV.
p
Then 0 and 1 are conjugate values for γ if and only if V is a critical points of exp p .
Moreoverm the multiplicity of the conjugate values 0 and 1 is the dimension of the
kernel of d exp p : T V (T p M) → T γ(1) M.
Proof. “⇐” Supposs that V ∈ T p M is a critical point for exp p . That is 0 = d exp p (V̇)(X)
for some nonzero X ∈ T V (T p M). Let c be a path in T p M with c(0) = V, ċ(0) = X.
We put
F(t, s) = exp t(c(s)), t ∈ [0, 1].
p
Then F(t, 0) = exp p tV = γ, and γ s (t) = exp p tc(s) is a geodesic. That is, F is a
geodesic variation of γ. So the variational field
∂
U(t) := | s=0 exp tc(s)
∂s p
∂ ∂
is a Jacobi field along γ. We compute U(0) = ∂s | s=0 exp p 0 = 0, and U(1) = ∂s | s=0 exp p c(s) =
d exp p (c(s))(ċ(0)) = d exp p (V̇)(X) = 0. Next, we hope to show U is not identically
5.6. INDEX FORMS 131
zero. This is because

∂
∇ ∂ U(t)|t=0 = e
U̇(0) =e ∇ ∂ |t=0 | s=0 exp t(c(s))
∂t ∂t ∂s p
∂
∇ ∂ | s=0 |t=0 exp tc(s) = e
=e ∇ ∂ | s=0 c(s)
∂s ∂t p ∂s
(the covariant derivative o f the vector f ield s 7→ c(s) along the constant curve s 7→ p)
=ċ(0) = X , 0.
Therefore, we show 0 and 1 are conjugate values for γ.

(“⇒”) We argue by contridiction. Suppose V is not a critical point for exp p . If
X1 , · · · , Xn ∈ T V (T p M) are m linearly independent vectors, then d exp p (V)(X1 ), · · · , d exp p (V)(Xn ) ∈
T γ(1) M are linearly independent. Choose paths c1 , · · · , cn in T p M with
ci (0) = V


ċ (0) = X , i = 1, · · · , n,
 (5.5.1)
i i
And F(t, s) := d exp p tc1 (s) are geodesic variation of γ with variational fields Vi (t). The
Vi are Jacobi fields along γ which vanish at 0. Moreover, the Vi (1) := d exp p (V)(Xi )
are independent, so no nontrivial linear combination of the Vi can vanish at 1. Since
the vector space of Jacobi fields along γ which vanish at 0 has dimension exactly n, it
follows that no nonzero Jacobi field along γ vanishes at 0 and also at 1.
5.6 Index forms

In this section, we discuss the minimizing property of a geodesic via Index forms: For
that purpose, we need consider a piecewise C ∞ proper variation of a geodesic γ. That
is, we compare the length of a geodesic γ : [a, b] → M with the length of any piecewise
C ∞ curve from γ(a) to γ(b). The corresponding variational field of γ is then a piecewise
C ∞ vector field along γ. Recall our calculations for the second variation formula(SVF),
the result is the same as the case of smooth variation:
∂2
Z b
|(v,w)=(0,0) E(v, w) = h∇W V, T i|ba + h∇T V, ∇T Wi − hR(W, T )T, Vidt,
∂v∂w a
Rb
where h∇W V, T i a = 0 when the variation is proper.
Definition 5.4. (Index form) The index form of a geodesic γ is
Z b
I(V, W) = (h∇T V, ∇T Wi − hR(W, T )T, Vi)dt
a
where V, W are two piecewise smooth vector fields along γ.

Remark 5.5. (1) If V, W are C ∞ on each [ti , ti+1 ] where
a = t0 < t1 < · · · < tn < tn+1 = b

is a subdivision of [a, b]. Then by integration by parts,

Z b
I(V, W) = h∇T V, ∇T Wi − hR(W, T )T, Vidt
a
k
X Z b
= hV, ∇T Wi|tti+1
i
+ h−∇T ∇T W, Vi − hR(W, T )T, Vidt
i=0 a
Z b
⇒ I(V, W) = − h∇T ∇T W + R(W, T )T, Vidt + h∇T W, Vi|ba
a
k
X
− h∇T (t+j ) W − ∇T (t−j ) W, Vi. (5.6.1)
j=1
(2) Note for a proper variation
∂2
|(v,w)=(0,0) E(v, w) = I(V, W).
∂v∂w
Let V := the set of all piecewise smooth vector fields along γ : [a, b] → M, and
V0 = {x ∈ V|X(a) = 0, X(b) = 0}.
We need extend Proposition 5.6 to piecewise smooth vector fields.
Proposition 5.9. Let γ : [a, b] → M be a geodesic and U ∈ V. Then U is a Jacobi

field if and only if I(U, Y) = 0, ∀Y ∈ V0 .
Proof. Note that, comparing with Proposition 5.6, we here have U ∈ V may be piece-
wise smooth, and so does Y. However, a Jacobi field is smooth. (The result and proof
here is very much similar in sprit to the characterization of geodesic.(see previous exer-
cise); A piecewise smooth curve c is a geodesic if and only if, for every proper variation
Fof c, we have E 0 (0) = 0.)
(⇒) If U is a Jacobi field, then I(U, Y) = 0, ∀Y ∈ V0
Z b
I(U, Y) = (h∇T U, ∇T Yi − hR(Y, T )T, Ui)dt
a
Z b
Remark(1)
= − h∇T ∇T U + R(U, T )T, Yidt + h∇T U, Yi|ba
a
k
X
− h∇T (t+j ) U − ∇T (t−j ) U, Yi.
j=1
Y ∈ V ⇒ h∇T U, Yi|ba = 0
k
X
U is smooth ⇒ h∇T (t+j ) U − ∇T (t−j ) U, Yi = 0
j=1
U is Jacobi
= 0
(⇐) Assume I(U, Y) = 0, ∀Y ∈ V0 . Let f : [a, b] → R be a smooth function s.t.
f (ti ) = 0, i = 0, · · · , k + 1
and f i0 otherwise. Set W = U, V = Y = f (∇T ∇T U + R(U, T )T ). Note that Y is

well-defined and Y ∈ V0 .
Therefore
X Z ti+1
0 = T (U, Y) = − f (t)|∇T ∇T U + R(U, T )T |2 dt.
i ti
Hence, we have ∇T ∇T U + R(U, T )T = 0 on each [ti , ti+1 ].

That is, “piecewisely”, U is a Jacobi field. (∗)
Next, for any j = 1, · · · , k, let Y ∈ V0 s.t.
Y(t j ) = 0, ∀i , j



(5.6.2)
Y(t j ) = ∇T (t+j ) U − ∇T (t−j ) U


Then 0 = I(U, Y) = |∇T (t+j ) U − ∇T (t−j ) U|2 .

Hence ∇T (t+j ) U = ∇T (t−j ) U.
Therefore U is a C 1 vector field along γ. Combining with the fact (*) and using the
uniqueness of Jacobi fields with given initial data, we conclude U is the Jacobi field on
[a, b].
Recall our previous discussions about SVF, we say the property “γ is loccaly min-
imizing” is equivalent to “I(V, V) = 0, ∀0 , V ∈ V0 ”. Since I(V, W) is a biliear,
symmetric from on the vector space V0 , the later condition is equivalent to say “I is
positive definite on V0 ”.
To illustrate the idea, we can compare the index form with the Hessian of a function
f : Rn → R. Consider a curve ξ in Rn , with ξ(0) ∈ Rn . Then the second order derivative
d2
of f along ξ is ds 2 f (ξ(s)). Hessian of f valued at the vector ξ̇(0) is
d2
f (ξ(s))| s=0 = Hess f (ξ̇(0), ξ̇(0)).
dt2
2
d2
2 f (ξ(s))| s=0 only depends on ξ̇(0). Once we know f (ξ(s))| s=0 , ∀ξ,
d
In particular ds ds2
then we have Hess f (v, v) for any v, and hence
1
Hess f (v, w) = (Hess f (v + w, v + w) − Hess f (v, v) − Hess f (w, w)).
2
Analogously, we replace Rn by the space ℘ of all curves c : [a, b] → M. Given a “point”
of ℘, i.e. a curve γ ∈ ℘, consider a “curve” through it, i.e. a 1-parameter family of
curves {γ s }. Let E be a function E : ℘ → R. The restriction R ◦ γ s := E(s), and
d2 d2
E(γ s )| s=0 = E(s)| s=0 = “ Hess EhV(t), V(t)i00 .
ds2 ds2
By polaritation, one have “Hess E(V, W)”, the Hessian of E on the “Hilbert space of
curves”. All formal discussion here can ve made rigorous,
In particular, when considering ℘0 of all curves c : [a, b] → M s.t. c(a) =
γ(a), c(b) = γ(b), the “Hessian of E” is given by the index form.
Next, our aim is discuss the relation between Algeraic properties of the index form
of the geodesic γ and Minimizing properties of the geodesic γ. Given a normal geodesic
γ : [a, b] → M, we can imagine the end point γ(b) move from γ(a) slowly to γ(b).
When |b − a| is small enough, γ|[a.b] is minimizing, hence we can expect I is positive
definite on V0 .
By the rough idea we explained before Theorem 5.4, when |b − a| is large, s.t. there
is a conjugate value of a in (a, b), γ|[a,b] is not(locally) minimizing, then we can expect
∃ X s.t. I(X, X) < 0.
In the case of a and b are conjugate values of γ, we have from Proposition 5.6, for
any Jacobi field U along γ with U(a) = U(b) = 0, we have I(U, U) = 0.
Theorem 5.6. Let γ : [a, b] → M be a geodesic from p = γ(a) to q = γ(b).

(1) p = γ(a) has no conjugate point along γ⇔ the index form I is positive definite
on V0 .
(2) q = γ(b) is a conjugate point of p along γ, and ∀t ∈ (a, b), γ(a) and γ(t) are
not conjugate point.(i.e. q is the first conjugate point of p)⇔I is positive semidefinite
but not positive definite on V0 .
(3) ∃ t ∈ (a, b), s.t. p = γ(a) and q = γ(t) are conjugate points⇔I(X, X) < 0 for
some X ∈ V0
Remark 5.6. Theorem refz.23 tells if γ(a) has no conjugate point along γ|[a,b] , then
for any [a, β] ⊂ [a, b], (α < β), γ(α) also has no conjugate point along γ|[α,β] . Since
otherwise, let Je be a nonzero Jacobi field along γ|[α,β] with J(α) e = 0 = J(β).
e Let Irs be
the index form of γ|[r,s] . Then let J|[a,α] ≡ 0 ≡ J|[β,b] , J|[α,β] = J.
e
Iab (J, J) = Iaα (0, 0) + Iαβ ( J, e + Iβb (0, 0) = Iαβ ( J,

e J) e = 0.
e J)
Hence (1) tells, p = γ(a) does have a conjugate along γ|[a,b] .
To show Theorem 5.6(1), we first prove the following useful Lemma.
Lemma 5.1. Let γ : [a, b] → M n be a geodesic, and γ(1) has no conjugate point along
γ. Then for any Va ∈ T γ(a) M and Vb ∈ T γ(b) M, there exists a unique Jacobi field U such
that
U(a) = Va , U(b) = Vb .
Proof. By proposition 5.4(2), the vector space of all Jacobi fields along γ is of dimen-
sion 2n. Let `0 be the subspace of Jacobi fields U with U(a) = Va . Then dim `0 = n.
Note that T γ(b) M is also a vector space with dim T γ(b) M = n. In fact, the linear trans-
formation
A : `0 → T γ(b) M, U 7→ U(b)
is injective. This is because if we have U1 , U2 ∈ `0 s.t. U1 (b) = U2 (b).
Then U1 − U2 is again a Jacobi field along γ, we check
U1 − U2 (a) − 0, U1 − U2 (b) = 0.
Since γ(a) and γ(b) are not conjugate points, we have U1 − U2 ≡ 0. Therefore A is
injective, and hence, an isomorphism,
Proof of (1):
(⇒) Let {γ̇(b), E2 , · · · , En } be an orthonormal basis of T γ(b) M. From Lemma 1, ∃!
Jacobi field Ji along γ s.t.
Ji (0) = 0, Ji (b) = Ei , i = 2, · · · , n.
Moreover, Proposition 5.9(3) tells hJi (t), γ̇(t)i = 0, ∀t ∈ [a, b]. By the arguement in the
proof of Theorem ??, {Ji (t)} are linearly independent in any T γ(t) M.
For any U ∈ V0 , note
Z b
I( f γ̇(t), f γ̇(t)) = h f˙(t)γ̇(t), f˙(t)γ̇(t)idt ≥ 0,
a
and “ = ⇒ f ≡ 0. Only consider U = f i Ji for some functions f i s.t. f i (a) = f i (b) = 0.

00
Next, we compute
Z b
I(U, U) = h∇T ( f i Ji ), ∇T ( f j J j )i − hR( f i Ji , T )T, f j J j idt
a
Z b Z b Z b
= h f˙i Ji , f˙ j J j idt + h f˙i Ji , f j ∇T J j i + h f i ∇T Ji , f˙ j J j i
a a a
Z b Z b
+ f i f j h∇T Ji , ∇T J j idt − f i f j hR(Ji , T )T, J j idt.
a a
Z b Z b
S uppose C = h f ∇T Ji , f˙ J j i, D =
i j
f i f j h∇T Ji , ∇T J j idt
a a
Z b
E= f f hR(Ji , T )T, J j idt
i j
a
Observe that
Z b
D= f i f j h∇T Ji , ∇T J j idt
a
Z b
d
= { ( f i f j h∇T Ji , J j i) − f˙i f j h∇T Ji , J j i − f i f˙ j h∇T Ji , J j i − f i f j h∇T ∇T Ji , J j i}dt
a dt
Z b
= f f h∇T Ji , J j i|a −
i j b
f˙i f j h∇T Ji , J j i − C + E
a
Z b
=− f˙i f j h∇T Ji , J j i − C + E
a
In fact, h∇T Ji , J j i = hJi , ∇T J j i. This is because
h∇T Ji , J j i − hJi , ∇T J j i|t=0 = 0 (since Ji (0) = J j (0) = 0)

and
d
(h∇T Ji , J j i − hJi , ∇T J j i) = h∇T ∇T Ji , J j i − hJi , ∇T ∇J j i
dt
= − hR(Ji , T )T, J j i + hR(Ji , T )T, Ji i = 0, ∀t.
Therefore D = −B − C + E, and hence

Z b
I(U, U) = h f˙i Ji , f˙ j J j idt ≥ 0.
a
Moreover
f˙i = 0


“= 00
⇔ f i = 0 ⇔ U = 0.

holds ⇔ i (5.6.3)

f (0) = 0 = f (b)
i 

This proves the positive definiteness of I on V0 .
Proof of (2):
(⇒) Choose any c ∈ (a, b). Pick a parallel orthonormal vector fields {γ̇(t), E2 (t), · · · , En (t)}.
Then any U ∈ V0 = V0 (a, b),
n
X
U(t) = f i (t)Ei (t)
i=2
for some functions f i with f i (a) = f i (b) = 0.

Define τ : V0 (a, b) → V0 (a, c) by
n
X b−a b−a
τ(V)(t) = f i (a + (t − a))Ei (a + (t − a)) = .
i=2
c−a c−a
By Theorem ??(1), we know I c (τ(V), τ(V)) > 0. We can check by definition that
Z b
lim I c (τ(V), τ(V)) = I(V, V) = ( f˙i )2 − f i f j hR(Ei , T )T, E j idt ≥ 0.
c→b a
Hence I is positive sedefinite, since for any nonzero Jacobi field U with U(a) = U(b) =
0, we have I(U, U) = 0.
Proof of (3):
(“⇒”) Let t is conjugate to a along γ, and there is a nonzero Jacobi field J along γ
s.t. J(a) = J(t) = 0. Let Jebe the vector field along γ with
e = J(t), a ≤ t ≤ t
J(t)
e = 0, t ≤ t ≤ b
J(t)
∇ ∂ Je− ∇T Jesince
Notice that the discontinuity of e
∂t
∇T (t+ ) Je− ∇T (t− ) Je = −∇T (t− ) Je , 0.
e = 0, this implies Je ≡ 0.)

(Otherwise, together with J(t)
Choose a vector field along γ which satisfies
U(a) = 0 = U(b), hU(t), ∇T (t+ ) Je− ∇T (t− ) Ji

e = −1.
Define the vector field along γ

1e
X := J − cU,
c
where c is a small number.
Then I(X, X) = c12 I( J, e U) + c2 I(U, U), where
e − 2I( J,
e J)
I( J, e = 0 since Je ∈ V0 (a, b).

e J)
e U) = −hU(t, ∇ + Je− ∇ − J)i
I( J, e = 1.
T (t ) T (t )
Hence I(X, X) = −2 + c2 I(U, U).

For sufficiently small c, this is < 0.
(1)⇐)follows from (2)⇒(3)⇒.
Similarly, (2)(⇐), (3)⇐ are proved.
Let us mention a very useful lemma. Recall in Proposition 5.7, we have shown a
Jacobi field U is the critical point of I(X, X).
Lemma 5.2. (Minimizing property of Jacobi field) Let γ : [a, b] → M be a geodesic
without conjugate points, U be a Jacobi field along γ, and X a piecewise C ∞ vector
field along γ with
X(a) = U(a), X(b) = U(b).
Then I(U, U) ≤ I(X, X) where “=” holds iff X = U.
Proof. From (5.6.1), we see for any piecewise C ∞ vector field W along γ, we have
I(U, W) = h∇T U, Wi|ba .
Since X − U ∈ V0 (a, b), Theorem ??(1)(⇒) tells
0 ≤I(X − U, X − U) = I(X, X) + I(U, U) − 2I(X, U)

=I(X, X) + h∇T U, Ui|ba − 2h∇T U, Xi|ba
=I(X, X) − h∇T U, Ui|ba
=I(X, X) − I(U, U)
If I(X, X) = I(U, U), we have I(X − U, X − U) = 0.

Therefore Theorem ??(1) tells X − U = 0.
Remark 5.7. In the case, we derive Lemma 5.6.2 from Theorem ??(1)(⇒). In face, the
converse is also true.
In fact, the results in T heorem?? can be pushed forward much further to the cele-
brated Morse index Theorem.
We particularly observe that for a geodesic γ : [a, b] → M γ(a) has a conjugate
point in (a, b)⇔ ∃X ∈ V0 (a, b), I(X, X) < 0.
Definition 5.5. (index and nullity of γ). We call for a geodesic γ : [a, b] → M
ind(γ) = max dim{A ⊂ V0 |I is negatively de f inite on the subspace A}
the index of γ.
We call
N(γ) = dim{X ∈ V0 |I(X, Y) = 0, ∀Y ∈ V0 }
the nullity of γ.
Remark 5.8. In fact, N(γ) is equal to the multiplicity of a and b as conjugate values.
If γ(b) is not conjugate to γ(a), then N(γ) = 0.
In this language, Theorem ??(3) can be restated as
∃ t s.t. N(γ|[a,t] ) ≥ 1 ⇔ index(γ) ≥ 1.
A far-reaching generalization is the following celebrated Theorem.
Theorem 5.7. (Morse index Theorem:) The index of γ : [a, b] → M is the number
of t ∈ (a, b) which are conjugate to a, each conjugate value being counted with its
multiplicity. The index is always finite. That is
X
ind(γ) = N(γ|[a,t] ) < ∞.
a<t<b
In particular, γ(a) has only finite many conjugate points along γ.
For the proof, one need to show the index of γ increases by at least v as t passes
a conjugate value t with multiplicity v. This can be handled by essentially the same
trick which was used in the proof of Theorem ??(3). We refer to [WSY, Chapter 9] for
details of proof.(see [JJ, section 4.3] for an analytic proof!!)
It is a good point to reflect our proof of Bonet-Myers Theorem. We show that if
sectionalcurvature ≥ ki0, for a geodesic γ of length l > √πk , we have for
π
V(t) = sin( t)E(t)
l
where E(t) is a parallel vector field along γ,
I(V, V) = 0.
Note when l = √π , sectional curvature=k > 0

k
√
V(t) = sin( kt)E(t)
is a Jacobi field along γ.(Exercise)

In particular, when sectional curvature = k > 0, a geodesic γ of length l > √πk
contains at least a conjugate point of γ(0). Hence ind(γ) ≥ 1, and γ is not (locally)
minimizing.
5.7. THE PROOF OF MORSE INDEX THEOREM 139
The proof of Bonnet-Myers tells when sectional curvature ≥ k, a geodesic of

length l > √πk also contains at least a conjugate point, and ind(γ) ≥ 1.
On the other hand, if sectional curvature = 0 or sectional curvature = −k, k >
0, the Jacobi field along γ with J(0) = 0 are linearly combinations of dtE(t) and
d sinh(kt)E(t) which will never vanish anywhere other than 0. Hence γ does not contain
conjugate points. In fact, this is true for the case sectional curvature ≤0. This is the
our next topic.
5.7 The proof of Morse index theorem

Rb
Recall we have I( f T, f T ) = a
( f˙)2 dt ≥ 0, “ =00 ⇔ f ≡ 0. And
I( f T, U) = 0, ∀U ∈ V0 (a, b), hU, T i = 0.
So we can restrict ourself to the subspace
V0⊥ (a, b) := {X ∈ V0 (a, b)|hX, T i = 0}
when studying index and nullity of γ.

For simplicity, let’s take (a, b) = (0, 1). Firstly,we show ind(γ) < ∞:
We first explain that we can find a finite-dim subspace T − 1 of V0⊥ (0, 1) s.t. the
index, nullity of I do not change when restricting to T 1 .
By considering the open covers of γ|[0,1] by the totally normal neighborhood of each
γ(t), t ∈ [0, 1], we can find a finite subdivision, 0 = t0 < t1 < · · · < tk < tk+1 = 1 such
that γ|[ti ,ti+1 ] lies in a totally normal neighborhood Ui . In particular, γ|[ti ,ti+1 ] contains no
conjugate point for each i.
Define:
T 1 := T 1 (1) := {X ∈ V0 : X is Jacobian along each γ|[ti ,ti+1 ] , ∀ i = 0, · · · , k}.

T 2 := T 2 (1) := {X ∈ V0 : X(ti ) = 0, ∀ i = 0, · · · , k + 1}
Lemma 5.3. we have

M
(i) V0⊥ (0, 1) = T 1 T2
(ii) I(T 1 , T 2 ) = 0
(iii) I|T2 is positive de f inite.
Proof. Consider the map

M M
ϕ : T 1 → T γ(t1 ) M ··· T γ(tk ) M, J 7→ (J(t1 ), · · · , J(tk )).
Clearly, this is a linear map, and 1-1.(Since on γ|[ti ,ti+1 ] , J is uniquely determined by
J(ti ) and J(ti+1 ).)
Therefore ϕ is a linear isometry, In particular, dim T 1 = nk < ∞.
Given any X ∈ V0⊥ (0, 1). Let JX := ϕ−1 (X(t1 ), · · · , X(tk )). Then we have JX ∈
T 1 , X − JX ∈ T 2 .
Moreover, T 1 T 2 = {0} sinceL J(ti ) = 0 = J(ti+1 ) ⇒ J ≡ 0 on γ|[ti ,ti+1 ] .

T
This shows (i) V0⊥ (0, 1) = T 1 T2.
For (ii), we have ∀ X1 ∈ T 1 , X2 ∈ T 2
k
X
I(X1 , X2 ) = h∇T X1 , X2 i|10 − h∇T (t+j ) X1 − ∇T (t−j ) X1 , X2 i = 0.
j=1
For (iii), any X ∈ T 2

k
X
I(X, X) = Ittii+1 (Xi , Xi )i0
i=0

By Lemma 5.7.1, we obtain immediately
ind(γ) ≤ dim(T 1 ) ≤ ∞
and the index, nullity of I|T1 equal to ind(γ), N(γ) respectively.
Lemma 5.4. (i) ∀ τ ∈ (0, 1], ∃ δ > 0, s.t. ∀ ∈ [0, δ], we have ind(τ − ) = ind(τ).
(ii) ∀ τ ∈ [0, 1), ∃ δ > 0, s.t. ∀ ∈ [0, δ], we have ind(τ + ) = ind(τ) + N(τ).
Proof. For given τ, we can assume the division we choose previously has the property
that τ ∈ (t j , t j+1 ).
Define T 1 (τ) := {X ∈ V0⊥ (0, 1), X|[ti ,ti+1 ] is Jacobian, i = 0, · · · , j−1.X|[t j ,τ] is also Jacobian.}.
Similarly, consider
M M
ϕτ : T 1 (τ) → T γ(t1 ) M ··· T γ(t j ) M, X 7→ (X(t1 ), · · · , X(t j ))
is a linear isometry.
I τ |T1 (τ) can be considered as a quadratic form over T γ(t1 ) M
L L
··· T γ(t j ) M in the
sense
I0τ (x, y) := I0τ ((ϕτ)−1 (x), (ϕτ)−1 (y)).
Note ind(τ) is the index of I0τ |T1 (τ) .
Hence ∀ X, Y ∈ T 1 (τ), denote Xi = X|[ti ,ti+1 ] , X j = X|[t j ,τ] .
j−1
X
I0τ (x, y) = h∇T Xi , Yi|tti+1
i
− h∇T X j (t j ), Y(t j )i.
i=0
For given division 0 = t0 < t1 < · · · < t j < t j+1 < · · · < tk+1 = 1,
M M
T 1 (τ) T γ(t1 ) M ··· T γ(t j ) M
is a fixed vector space. So when τ changes, only h∇T Ẋ j (t j ), Y(t j )i change in I0τ (x, y),
j
where x = (x1 , · · · , x j ), y = (y1 , · · · , y j ) ∈
L
T γ(ti ) M.
i=1
5.7. THE PROOF OF MORSE INDEX THEOREM 141
X j and Y are Jacobi field on [t j , τ) with
X j (t j ) = x j , X j (τ) = 0
Y(t j ) = y j , Y(τ) = 0
By construction, γ|[t j ,t j+1 ] ⊂ U j a totally normal neighborhood.

Hence geodesics lying inside U j depends smoothly on their endpoints.
Since Jacobi fields are variational vector fields geodesic variations, we have X j , Y|[t j ,τ]
also depends continuouly on endpoint γ(t j ) and γ(τ).
Therefore −h∇T X j (t j ), Y(t j )i are smooth w.r.t. τ. That is I0τ (x, y) is a continuous w.r.t
j
τ for given x, y. So for x ∈ T γ(ti ) M, I0τ (x, y) < 0(> 0) implies
L
i=1
∃ δ > 0 s.t. I0τ± (x, x) < 0, ∀ ∈ [0, δ].
This tells
ind(τ ± ) ≥ ind(τ) (5.7.1)

ind+ (τ ± ) ≥ ind+ (τ) (5.7.2)
By linear algebraic theory, the linear space T 1 (τ) can be decomposed into:
maximal positive definite subspace
maximal negative definite subspace
null space {X ∈ T 1 (τ)|I(X, Y) = 0, ∀ Y ∈ T 1 (τ)}
j
n j = dim( T γ(ti ) M) = ind+ (τ) + ind(τ) + N(τ).
L
i=1
Using ind+ (τ) = n j − ind(τ) − N(τ), we derive from (5.7.2) that
n j − ind(τ ± ) − N(τ ± ) ≥ n j − ind(τ) − N(τ)
i.e.
ind(τ ± ) ≤ ind(τ) + N(τ) − N(τ + ) ≤ ind(τ) + N(τ) (5.7.3)
Combining (5.7.1) and (5.7.3) gives
ind(τ) ≤ ind(τ ± ) ≤ ind(τ) + N(τ) (5.7.4)

j
For any ξ, t j hξ < τ < t j+1 , ∀x ∈
L
T γ(ti ) M, we have
i=1
I0ξ (x, x) − I0τ (x, x) = Itξj (X j,ξ , X j,ξ ) − Itτj (X j,τ , X j,τ )
where X j,ξ is the Jacobi field with X j,ξ (t j ) = x j , X j,ξ (ξ) = 0, X j,τ is the Jacobi field with
X j,τ (t j ) = x j , X j,τ (τ) = 0.
By minimizing property of Jacobian field, we have
I(X j,τ , X j,τ ) ≤ I(X j,ξ , X j,ξ )

and “=” holds iff X j,τ = X j,ξ ⇔ X j,τ = 0.

That is,
I0τ (x, x) ≤ I0ξ (x, x),
and “ =00 holds iff X j,τ = 0
Hence
(i) I0ξ (x, x) < 0 ⇒ I0τ (x, x) < 0

(ii)Let x be in the null space o f I0ξ . T hen (ϕξ )−1 (x) ∈ V0⊥ (0, ξ) is a Jacobi f ield vanishing at 0 and ξ.
Observe that x j = (ϕξ )−1 (x)(t j ) , 0. Since otherwise, we have (ϕξ )−1 (x)|[t j ,ξ] ≡ 0. (γ|[t j ,ξ]
contains no conjugate point) and therefore (ϕξ )−1 (x)(x) ≡ 0 ⇒ x ≡ 0, contradicting to
x , 0.
Therefore, we have
I0τ (x, x) < I0ξ (x, x) = 0, ∀ x ∈ the null space o f I0ξ .
In conclusion, (i)+(iiimplies
ind(τ) ≥ ind(ξ) + N(ξ).
We have
ind(τ) ≤ ind(τ − ) ≤ ind(τ) − N(τ − ) ≤ ind(τ).
⇒ ind(τ) = ind(τ − ).
ind(τ + ) ≤ ind(τ) + N(τ) ≤ ind(τ + ).
⇒ ind(τ + ) = ind(τ) + N(τ).

Proof pf Morse:
Lemma 4.7.2⇒if γ(τ) is not a conjugate point of γ(0), ∃ δ > 0 s.t. ind(t)|(τ−δ,τ+δ) is
constant.
If γ(τ) is conjugate to γ(0), ∃ δ > 0 s.t. ind(t)|(τ−δ,τ) is constant and ind(t)|(τ,τ+δ) is
also constant.
And the jump size of ind(t) at t = τ is N(τ).
So when t changes from 0 to 1, ind(t) changes from 0, and jump where τ is con-
jugate value. Since ind(1) < ∞, we know this jump can only happen finitely many
times.
⇒ ind(1) =
P
N(τ).
0<τ<1
5.8 Cartan-Hadamard Theorem

Recall in (IV) §5 we have shown that when sec ≤ 0, every geodesic is locally minimiz-
ing. This indicates that no conjugate points exist in this setting.
Proposition 5.10. If all sectional curvature of (M, g) are ≤ 0, the no two points of M
are conjugate along any geodesic.
5.8. CARTAN-HADAMARD THEOREM 143
Proof. Let γ be a geodesic with velocity field along velocity field T (t) = γ̇(t) it. Let
U(t) be a Jacobi field along γ. Then
∇T ∇T U + R(U, T )T = 0.
So h∇T ∇T U, Ui = −hR(U, T )T, Ui ≥ 0.

Therefore, dtd h∇T U, Ui = h∇T ∇T U, Ui + h∇T U, ∇T Ui ≥ 0, that is, h∇T U, Ui is
non-decreasing.
2
Note that dtd h∇T U, Ui = 12 dtd 2 hU, Ui.
d2
That is, we have shown dt2
|U(t)|2 ≥ 0, i.e. |U(t)|2 is convex. So U(t0 ) = 0 = U(t1 )
⇒ U ≡ 0.

Remark: In fact, for a normal Jacobi field U(t) with U(0) = 0, define f : (0, ∞) →
1
R by f (t) = |U(t)| = hU(t), U(t)i 2 . At the values t with U(t) , 0, we compute
d
d hU(t), U(t)i hU̇(t), U(t)i
f˙ = f (t) = dt 1
=
dt 2hU(t), U(t)i 2 |U(t)|

hÜ(t), U(t)i + hU̇(t), U̇(t)i |U(t)| − hU̇(t), U(t)i hU̇(t),U(t)i
|U(t)|
f¨(t) =
|U(t)|2
hU̇(t), U(t)i2 |U̇(t)|2 1
=− + − hR(U, T )T, Ui (By Cauchy − Schwarz)
|U(t)|3 |U(t)| |U(t)|
|U̇(t)|2 |U̇(t)|2 1
≥− + − K(U, T ) hU, UihT, T i − hU, T i2
|U(t)| |U(t)| |U(t)|
= − K(U, T )|U(t)|.
2
That is, dtd 2 f (t) ≥ −K(U, T ) f (t), where f (t) = |U(t)|, and f (0) = 0.
A comparison result:
f (t) ≥ − β f (t)
 00




f (0) =0





 f˙(0) =1


and
g (t) = − βg(t)
 00



g(0) =0





 ġ(0) =1

Then f (t) ≥ g(t). (Use ( f − g)00 ≥ 0 and ( f − g)(0) = 0 = dtd ( f − g)(0).)

This is a very useful principle to investigate the geometry of a Riemannian manifold
via its Jacobi field and that of the space form. (Lecture 20. 2017.05.02)
Theorem 5.8 (Cartan-Hadamard). A complete, simply-connected, n-dimensional Rie-

mannian manifold (M, g) with all sectional curvature ≤ 0 is diffeomorphic to Rn , more
precisely,
exp p : T p M → M
is a diffeomorphism.
Remark: In 1898, Hadamard proved such properties for a complete, simply-connected
surface with non-positive Gauss curvature. In 1928, E.Cartan extended it to n-dimensional
Riemannian manifolds. In fact, Hadamard’s result has been proved by von Mongoldt
in 1881.
The assumption of ’simply-connectivity’ is necesary. For example, the cylinder
C ≡ {(x, y, z) ∈ R3 : x2 = z2 = 1, y ∈ R}
is complete and with secitonal curvature zero. Its exponential map exp p : T p M → M
is a non-trival covering map.
An important feature of theorem 5.8 is that it not only asserts that M and Rn
are diffeomorphic, but also gives the diffeomorphism map explicitly. Recall we have
mentioned Gromall − Meyer(1969) Theorem: any non-compact complete Riemannian
manifold (M n , g) with positive sectional curvature is diffeomorphic to Rn . But in this
case, the diffeomorphism map is not necessarily given explicitly by exp p . This is a big
difference between our understanding about nonpositively curved complete simply-
connnected Riemannian manifold and positively curved non-compact complete Rie-
mannian manifold, although their topology are both trival.
Theorem 5.8 is a direct consequence of Proposition 5.10 and the following general
result.
Theorem 5.9. Let (M, g) be a complete, connected, n-dimensional Riemannian man-
ifold, and let p be a point of M such that no point of M is conjugate to p along any
geodesic. Then
exp p : T p M → M
is a covering map. In particular, if M is simply-connected, then M is diffeomorphic to
Rn .
Proof. We first make it clear what the assuption ”p∗has no conjugate point” tell us:
For exp p : T p M → M, we have the tensor exp p g on T p M which is defined as
∗
∀V, W ∈ T X (T p M), ∀X ∈ T p M, exp p g(V, W) = g (d exp p )X V, (d exp p )X W .

”p has no conjugate point” ⇒ d exp p : T X (T p M) → T exp p M is 1-1. There-
X ∗
fore (exp p )∗ g(V, V) = 0 ⇔ d exp p (V) = 0 ⇔ V = 0. And, hence, exp p g is a
X
Riemannian metric on T p M. That is,
∗
exp p : (T p M, exp p g) → (m, g)
5.8. CARTAN-HADAMARD THEOREM 145
is a local isometry.
∗
Moreover, we claim (T p M, exp p g) is complete.
∗
That is beacuase, all straight lines throgh 0 ∈ T p M are geodesics of (T p M, exp p g)
since their images under the local isometry exp p : T p M → M are geodeisc in M. That
is, all geodesics through 0 ∈ T p M can be defined for all t. It follows that T p M is
geodesic complete and, hence, complete, by Hopf-Rinow.
In concluision, ”p has no conjugate point” ⇒ ”exp p is a local isometry and T p M is
complete”. Then theorem 5.9 is a consequence of the following lemma:
Lemma 5.5. Let M and N be connected Riemannian manifolds with M complete and
let φ : M → N be a local isometry. Then N is complete1 and φ a covering map2 onto3
N.
Remark: Lemma 5.5 has 3 conclusions. Note that the completeness of M is needed.
For example the inclusion map i : B(0, 1) ⊂ R2 → R2 is a local isometry but the open
disk B(0, 1) is not complete. i is not a covering map.
Proof of Lemma 5.5.
1, N is complete: We will show any geodesic on N is depend for all t. For any
geodesic γ : (−, ) → N, there exists a p0 ∈ M s.t. γ(0) = φ(p0 ).
−1
Then we find the geodesic c in M with c(0) = p0 , ċ(0) = dφ p0 γ̇(0). This is
possible since φ is a local isometry.
One can check the curve φ ◦ c is a geodesic (since φ is a local isometry which
preserves geodesic), and φ ◦ c(0) = φ(p0 ), (φ ◦˙ c)(0) = dφ p0 ċ(0) = γ̇(0).
Hence γ = φ ◦ c.
M is complete ⇒ c is defined for all t ⇒ γ is defined for all t . Hopf-Rinow tells N
is complete.
2, φ is onto N: That is, we have to show φ(M) = N. Since φ is everywhere regular
(i.e. ∀p ∈ M, dφ p is 1-1), by iverse function theorem, φ is open and, in particular, φ(M)
is open.
In fact, φ(M) is also closed, and, hence, φ(M) = N. Following is the reason.
Let q ∈ φ(M), and let V be a totally normal neighborhood of q. There is a q0 ∈ V
of the form q0 = φ(p0 ) for p0 ∈ M. Let γ be the geodesic in V s.t. γ(0) = q0 , γ(1) = q.
−1
Consider the geodesic c in M s.t. c(0) = p0 , ċ(0) = dφ p0 γ̇(0). Then γ = φ ◦ c.
Define p = c(1), then φ(p) = φ(c(1)) = phi ◦ c(1) = γ(1) = q. Therefore q ∈ φ(M).
That is φ(M) ⊂ φ(M), so φ(M) is closed.
3, φ is a covering map:
For fixed q ∈ N, let B(0, 2) = {Y ∈ T q N : ||Y|| < 2} ⊂ T q N, where > 0 is small
enough such that exp p is a diffeomorphism. Suppose p ∈ φ−1 (q). We have
dφ p
B(0, 2) ⊂ T p M / B(0, 2) ⊂ T q N
exp p expq (♥)
φ
B p (2) ⊂ M / Bq (2) ⊂ N
where φ ◦ exp p = expq ◦dφ p .

This is because, ∀V ∈ B(0, 1) ⊂ T p M, γ(t) = φ ◦ exp p (tV), t ∈ [0, 2) is a geodesic
in N s.t. γ(0) = φ(p) = q, γ̇(0) = dφ p (V). On the other hand, c(t) = exp p ◦dφ p (tV) =

exp p tdφ p (V) is a geodesic in N with c(0) = expq O = q, ċ(0) = dφ p (V). Hence
expq ◦dφ p (tV) = φ ◦ exp p (tV), ∀V ∈ B(0, 1) ⊂ T p M, ∀t ∈ [0, 2) ⇒
expq ◦dφ p |B(0,2)⊂T p M = φ ◦ exp p (tV)|B(0,2)⊂T p M (5.8.1)
That is, the diagram (♥) commutes.

Therefore, we have (using the fact expq : B(0, 2) ⊂ T p M → Bq (2) ⊂ N is a
diffeomorphisms) the LHS of (5.8.1) is a diffeomorphism. Since exp p : B(0, 2) ⊂
T p M → B p (2) ⊂ M is surjective, and φ ◦ exp p is a diffeomorphism by (5.8.1), we
have B(0, 2) ⊂ T p M → B p (2) ⊂ M is a diffeomorphism. Therefore, (5.8.1) ⇒
φ = expq ◦dφ p ◦ exp−1 p is also a diffeomorphism.
Now let, W = expq (B(0, )) ⊂ N and ∀p ∈ M, W p = exp(B(0, 2)) ⊂ M. We claim
that
(1): φ−1 (W) = ∪ p∈φ−1 (q) W p .
Now that φ : W p → W is a diffeomorphism by our previous argument. So in order
to show φ : M → N is a covering map, we only need to show the claim and
(2): W pi ∩ W p j , φ, ∀pi , p j ∈ φ−1 (q), pi , p j .
Proof of (1)
φ : W p → W is a diffeomorphism tells ∪ p∈φ−1 (q) W p ⊂ φ−1 (W). Now ∀p0 ∈ φ−1 (W),
5.9. UNIQUENESS OF SIMPLY-CONNECTED SPACE FORMS 147
let γ be the geodesic in W with γ(0) = φ(p0 ), γ(1) = q and of length d(φ(p0 ), q).

Let c be the geodesic in M with c(0) = p0 , ċ(0) = dφ−1 p0 γ̇(0). Then γ = φ ◦ c. M
is complete ⇒ q = γ(1) = φ ◦ c(1) is well defined. In particular, c(1) ∈ φ−1 (q) and
p0 = c(0) ∈ Wc(1) . This implies φ−1 (W) ⊂ ∪ p∈φ−1 (q) W p . Hence, we prove the claim (1).
Proof of (2)
Suppose ∃p1 , p2 ∈ φ−1 (q), p1 , p2 s.t. W p1 ∩ W p2 , φ, then we have p2 ∈
exp p1 (B(0, 2)).
But φ is a diffeomorphism on the B(0, 2) ⊂ T p1 M → B p1 (2) ⊂ M. Hence
φ(p1 ) = φ(p2 ) → p1 = p2 .

5.9 Uniqueness of simply-connected space forms

Now we can prove the ’uniqueness part’ of Theorem (5.1) which we started in the very
beginning of this Chapter. In fact, we can prove.
Theorem 5.10 (Uniqueness). Let (M, g) and (M, g) be two n-dimensional simply-
connected space form with sectional curvature c ∈ R. Let p ∈ M, p ∈ M, {e1 , . . . , en },
{e1 , . . . , en } be orthonormal basis of T p M, T p M, respectively. Then there exists a unique
isometry φ : M → M such that φ(p) = p, dφ p (ei ) = ei , ∀i.
Proof. Since K(Ag) = A1 K(g), we only need consider the cases c = 0, +1, −1. We first
show the existence of such an isometry.
Case 1. c = 0, −1. By Cartan-Hadamard, the maps exp p : T p M → M and exp p :
T p M → M are both diffeomorphisms. Let Φ be the unique isometry from T p M to T p M
(as inner product) such that Φ(ei ) = ei , i = 1, . . . , n.
Φ / TpM
TpM
exp p exp p

φ
M /M
This leads to φ : M → M where φ = exp p ◦Φ ◦ (exp p )−1 . Notice that φ is a

diffeomorphism. So it remains to show φ∗ g = g.
It is enough to show ∀q ∈ M ∀X ∈ T q M, φ∗ g(X, X) = g(X, X) i.e. the lenghth of
dφ(X) equals the length of X, ∀X ∈ T p M, ∀q.
Recall Lemma ... tells, there is a unique Jacobi field U along the geodesic from p
to q s.t. U(0) = 0 and U(1) = X. So we can calculate |X|2 = g(X, X) by calculate the
whole Jacobi field U(t). In fact we can construct U(t) explicitly.
Let V p ∈ T p M be such that exp p V p = q. Let W ∈ T V p (T p M) be such that

d exp p (W) = X. Consider the variation F(t, s) = exp p t V p + sW . We know
Vp
∂
the variational field U(t) = ∂s | F(t, s) is the Jacobi field with U(0) = p, U(1) =
s=0
∂

∂s exp p V p + sW = d exp p V p (W) = X, U̇(0) = X, U̇(0) = W.

Next, we show g dφ p (X), dφ p (X) can also be calculated by computing a whole
Jacobi field.
Consider F(t, s) = exp p t Φ(V p ) + sΦ(W) (we identify T V p (T p M) with T p M).
˙
Similarly, the variation field ∂ | F(t, s) = U(t) is a Jacobi field with U(0) = 0, U(t).
∂s s=0
We claim U(1) = dφq (X). This is seen from
φ ◦ F(t, s) = exp p ◦Φ ◦ (exp p )−1 ◦ exp p t(V p + sW)

= exp p ◦Φ(t(V p + sW))
= exp p t(Φ(V p ) + sΦ(W))
=F(t, s)
Hence
∂
dφF(t,s) ((U(t)) =dφ ◦ | s=0 F(t, s)
∂s
∂
= | s=0 (φ ◦ F(t, s))
∂s
∂
= | s=0 F(t, s)
∂s
=U(t)
In particular, U(1) = dφF(1,0) ((U(1)) = dφq (X). Hence, it remains to show |U(1)| =
|U(1)|.
5.9. UNIQUENESS OF SIMPLY-CONNECTED SPACE FORMS 149
Pick parallel orthonormal vector fields {e1 (t), . . . , en (t)}, {e1 (t), . . . , en (t)} along γ0 ,
i
γ0 respectively such that ei (0) = ei , ei (0) = ei . Then U(t) = f i (t)ei (t), U(t) = f (t)ei (t)
i
for some functions f i , f .
Solving Jacobi equation in M, M respectively:
in M: ∇T ∇T U + R(U, T ) = 0, T = γ̇0 (t) ⇔ fï (t)ei (t) + R( f j e j , T )T = 0 ⇔
fï (t) + f j hR(e j , T )T, ei i = 0, i = 1, . . . , n. (sectional curvature=c ⇒ hR(e j , T )T, ei i =
c(δi j hT, T i − hT, ei ihT, e j i) ⇔ hV p , V p i = hγ̇0 (0), γ̇0 (0)i = 0, i = 1, . . . , n.). Recall
we have further U(0) = 0, U̇(0) = W. Hence f i , i = 1, . . . , n satisfing the following
equation.

f (t) + c f j (t) δi j hV p , V p i − hV p , ei ihV p , e j i = 0, i = 1, . . . , n
 ï





f (0) = 0
 i




 f˙i (0) = hW, e i

i
i
in M: f , i = 1, . . . , n satisfies
ï

j
f (t) + c f (t) δi j hΦ(V p ), Φ(V p )i − hΦ(V p ), ei ihΦ(V p ), e j i = 0, i = 1, . . . , n







 i



 f (0) = 0

 f˙i (0) = hΦ(W), ei i




Since Φ : T p M → T p M is an isometry, we have hV p , V p i = hΦ(V p ), Φ(V p )i,

hV p , ei i = hΦ(V p ), Φ(ei )i = hΦ(V p ), ei i, hW, ei i = hΦ(W), Φ(ei )i = hΦ(W), ei i. By
i
the uniqueness of the solution, we have f i (t) = f (t), ∀t, ∀i ∈ {1, . . . , n}. In particular,
P i
|U(1)|2 = i f i (1)2 = i f (1)2 = |U(1)|2 . This proves the existence of an isometry
P
claimed in Theorem for the case c = 0 or t. The uniqueness of the isometry φ follows
from the following lemma.
Lemma 5.6. Let M be a connected Riemannian manifold and N be a Riemannian

manifold. Let φ1 , φ2 : M → N be two locally isometry such that ∃x ∈ M, φ1 (x) =
φ2 (x) = x0 ∈ N, (dφ1 ) x = (dφ2 ) x : T x M → T x0 N. Then φ1 = φ2 .
Proof. Define A ⊂ M to be A = {z ∈ M : φ1 (z) = φ2 (z), (dφ1 )z = (dφ2 )z }. By

assumption, x in A, i.e. A , φ. From the definition, A is closed.
Next, we show A is open. Thus since M is connected, we have A = M. Suppose
z ∈ A, then z0 = φ1 (z) = φ2 (z) ∈ N. Choose δ > 0 small enough, such that expz :
B(0, δ) ⊂ T z M → Bz (δ) ⊂ M is a diffeomorphism and expz0 : B(0, δ) ⊂ T z0 N is defined.
B(0, δ) ⊂ T z M
(dφi )z
/ B(0, δ) ⊂ T z0 N
expz expz0
φi
Bz (δ) ⊂ M / Bz0 (δ) ⊂ N
By similar argument in the proof of Lemma 5.5, we have φi ◦ expz = expz0 ◦(dφi )z ,
i = 1, 2. Notice that expz |Bz (δ)⊂M is invertible, we have φi = expz0 ◦(dφi )z ◦ (expz )−1 .
Now we check ∀y ∈ Bz (δ),
φ1 (y) = expz0 ◦(dφ1 )z ◦ (expz )−1 (y) = expz0 ◦(dφ2 )z ◦ (expz )−1 (y) = φ2 (y)
and (dφ1 ) = (dφ2 ). Therefore, we have Bz (δ) ⊂ A ⇒ A is open.

Case 2. c = +1. We can suppose M = Sn . ∀p ∈ Sn , any two geodesic from p will
together at its antipodal point p0 . Therefore, exp−1 n 0 n
p : S \{p } → T p S is a well-defined
smooth map.
Φ / TpM
T p Sn
O
exp−1
p
exp p

Sn \{p0 } M
where Φ is the isometry (of inner product spaces) with Φ(ei ) = ei .
Then φ : exp p ◦ Φ ◦ exp−1
p is a local isometry by the same argument as in the first
case.
Next, we extend φ to be defiend on the whole Sn . Pick any z ∈ Sn \{p0 }, z , p. Let
z = −z is the antipodal point of z. Let φ(z) = z ∈ M, then (dφ)z : T z Sn → T z M. Define
1
ψ : Sn \{z0 } → M as ψ = expz ◦(dφ)z ◦ exp−1

z .
Similar arguments tell that ψ is also a locally isometry. Consider the connected
Riemannian manifold W = Sn \{p0 , z0 }. We have two local isometries φ, ψ : W → M.
Oberserve that ψ(z) = expz ◦(dφ)z ◦ exp−1z (z) = z = φ(z), (dψ)z = (dφ)z . By lemma
5.6, we have φ = ψ|W . Now define θ : S → M by
n
φ(y), if y ∈ S \{p }
 n 0
θ(y) = 

 ψ(y), if y ∈ Sn \{z0 }
This is a well-defined c∞ map on Sn , and θ is a local isometry. By lemma 5.5, we

have θ is a covering map. Since M is simply-connected, θ is a diffeomorphism and
hence an isometry. Moreover dθ(ei ) = ei . This proves the existence. The uniqueness
follows again from lemma 5.6.

Theorem 5.10 has ery interesting consequnces. When M = M, we have
Corollary 5.1. Let M be a n-dimensional complete simply-connected Riemannian
manifold. Then M is a space-form iff ∀p, p ∈ M, and any orthonormal basis {e1 , . . . , en },
{e1 , . . . , en } of T p M, T p M, respectively, there exists an isometry φ : M → M s.t.
φ(p) = p, dφ(ei ) = ei , ∀i.
Definition 5.6 (Homogenous Riemannian manifolds). A Riemannian manifold (M, g)
is called homogenous is ∀p, q ∈ M, there exists an isometry
φ:M→M
5.10. CONVEXITY: ANOTHER APPLICATION OF CARTAN-HADAMARD THEOREM151
such that φ(p) = q

(M, g) is called two-point homogenous, if for any two pairs of points p1 , p2 and
q1 , q2 ∈ M with d(p1 , p2 ) = d(q1 , q2 ), there exists an isometry φ : M → M s.t. φ(pi ) =
qi , i = 1, 2.
Corollary 5.2. All simply-connected space forms are two-point homogenous.
Proof.
Let d(p1 , p2 ) = d(q1 , q2 ) = d. Let η, ξ : [0, α] → M be two normal geodesics

with ξ(0) = p1 , ξ(α) = p2 , η(0) = q1 , η(α) = q2 . (The existence is guaranted by
completeness via Hopf-Rinow).
Pick orhtonormal basis {e1 , . . . , en } and {e01 , . . . , e0n } of T p1 M and T q1 M, respectively,
where e1 = ξ̇(0), e01 = η̇(0). Then Theorem 5.10 ⇒ ∃ an isometry φ : M → M with
φ(p1 ) = q1 , dφ(ei ) = e0i ∀i. So φ ◦ ξ is a geodesic with φ ◦ ξ(0) = φ(p1 ) = q1 ,
(φ ◦˙ ξ) = dφ(ξ̇(0)) = dφ(e1 ) = e01 = η(0).˙
Therefore φ ◦ ξ = η. In particular φ(p2 ) = φ(ξ(α)) = η(α) = q2 .

5.10 Convexity: Another application of Cartan-Hadamard

Theorem
Convex functions and convex (sub)sets are important and useful concepts in analysis.
We discuss these topics on Riemannian manifolds in this section.
What is a convex function?
Recall that we call a function f : [a, b] → R to be convex if f (λx1 + (1 − λ)x2 ) ≤
λ f (x1 ) + (1 − λ) f (x2 ), ∀x1 , x2 ∈ [a, b], λ ∈ [0, 1]. One can prove that a convex function
must be Lipschitz continuous.
Recall for a C ∞ function f : [a, b] → R, it is convex iff f 00 ≥ 0 on [a, b]. This can be
shown via its Taylor expansion. That is
f 00 (x0 )
f (x) = f (x0 ) + f 0 (x0 )(x − x0 ) + (x − x0 )2 + . . .
2
f 00 (x∗ )
= f (x0 ) + f 0 (x0 )(x − x0 ) + (x − x0 )2
2
for some x∗ lying between x0 and x.

(⇒) Apply to the case x = x + h, x0 = x and x = x − h, x0 = x, we
have f 00 (x) =
f (x+h)+ f (x−h)−2 f (x) f (x+h)+ f (x−h)
limh→0 h2
. Also convexity implies f (x) = f 2 + 2 ≤
x+h x−h
2 .
Hence f 00 (x) ≥ 0.
(⇐) Apply to x0 = λx1 + (1 − λ)x2 , x = x1 gives f (x1 ) ≥ f (λx1 + (1 − λ)x2 ) +
f 0 (x0 )(1 − λ)(x1 − x2 ). Apply to x0 = λx1 + (1 − λ)x2 , x = x2 gives f (x2 ) ≥ f (λx1 +
(1 − λ)x2 ) + f 0 (x0 )λ(x2 − x1 ). Multiply the first by λ, multiply the second by (1 − λ),
and add them up providing λ f (x1 ) + (1 − λ) f (x2 ) ≥ f (λx1 + (1 − λ)x2 ).
We say a C ∞ function f is strictly convex, if f 00 > 0. Now consider a function
f : M → R where M is a Riemannian manifold. A suitable definition of convexing is:
Definition 5.7 (convex functions). We call a function f : M → R a convex function if
for any geodesic γ : [a, b] → M, f ◦ γ is convex, i.e. if ∀t1 , t2 ∈ [a, b], ∀λ ∈ [0, 1], it
holds that f (λ(λt1 + (1 − λ)t2 )) ≤ λ f (γ(t1 )) + (1 − λ) f (γ(t2 )).
Proposition 5.11. A C ∞ fucntion f : M → R is convex iff ( f ◦ γ)00 ≥ 0 for all geodesic
γ, which is further equivalent to Hess f ≥ 0, i.e. Hess f is positive semidefinite.
Proof. C ∞ f : M → R is convex iff ( f ◦ γ)00 ≥ 0, ∀γ follows from our previous
discussions.
Notice further that for any p ∈ M, any V p ∈ T p M, letting γ(t) be the geodesic with
γ(0) = p, γ̇(0) = V p , we have
Hess f (V p , V p ) = Hess f (γ̇(t), γ̇(t))|t=0

=∇2 f (γ̇, γ̇)|t=0 = ∇(∇ f )(γ̇(t), γ̇(t))|t=0
=∇γ̇(0) (∇γ̇(t) f ) − ∇∇γ̇ γ̇ f = ( f ◦ γ)00 .
Hence ( f ◦ γ)00 ≥ 0, ∀γ ⇔ Hess f (V p , V p ) ≥ 0, ∀p, ∀V p ∈ T p M.

We say a C ∞ function f : M → R to be a strictly convex funciton if Hess f > 0.
Next, let us consider a particular function on M. Given a fixed point O ∈ M,
consider the funtion %(·) = d(·, O) : M → R.
Theorem 5.11. Let M be a complete, simply-connected Riemanian manifold with non-
positive sectional curvature. Let O ∈ M. Then the function %2 is C ∞ and strictly
convex.
Example: In Rn with the canonical Euclidean metric, let O = 0 ∈ Rn , we compute
∂2 ∂2 X k 2
Hess%2 (X, X) =X i X j %2 (x) = X i X j i j (x )
∂x ∂x
i j ∂x ∂x k
X X
=X i X j 2 δk j δki = 2 (xk )2 = 2|X|2
k k
First, we observe, without any curvature restriction, %2 is always ”locally” strictly

convex.
Lemma 5.7. Let M be a Riemannian manifold, and O ∈ M. Then there exists a

neighbor U0 of O s.t. %2 is smooth and strictly convex in U0 .
Proof. Let (U, x1 , . . . , xn ) be a normal coordinate neighborhood of O ∈ M, such that
xi (O) = 0. Then %2 (0) = ni=1 (xi )2 , ∀x ∈ U.
P
Recall any geodesic γ with γ(0) = O, γ̇(0) = V(= V 1 , . . . , V n ) ∈ T O M, can be
written as γ(t) = (x1 (t), . . . , xn (t)) where xi (t) = V i t, %2 ◦ γ = ni=1 t2 (V i )2 . Hence
P
Hess%2 (V, V) = (%2 ◦ γ)00 = 2 ni=1 (V i )2 > 0. Therefore, there exists a neighborhood of
P
O, UO ⊂ U, s.t. Hess%2 is positive definite on UO .

But for ”global” results, we need curvature restriction. Let us recall the first(second)
variation formula for length functions.
Lemma 5.8. Let γ : [a, b] → M be a normal geodeisc, and F : [a, b] × (−, ) → M
be a variation of γ with variational field V(t), t ∈ [a, b]. Then
Z b
d
L0 (0) = | s=0 L(s) = hV(t), γ̇(t)i|ba − hV(t), ∇γ̇ γ̇idt
ds a
Z b
d2
L00 (0) = 2 | s=0 L(s) = h∇V V, γ̇i|ba + h∇γ̇ V ⊥ , ∇γ̇ V ⊥ i − hR(V ⊥ , γ̇)γ̇, V ⊥ idt
ds a
where V ⊥ = V − hV, γ̇iγ̇.
Proof. of Theorem 5.11.
By Cartan-Hadamard
Theorem and the definition of expO , we have x ∈ M, %2 (x) =
g expO (x), expO (x) = | exp−1
−1 −1 2 ∞
O (x)| is a C function on M. By Lemma 5.7, it remains
to show that Hess%2 (V, V) > 0 for any x ∈ M, any 0 , V ∈ T x M.
Let ξ : [0, ] → M be the geodesic of M with ξ(0) = x, ξ̇(0) = V. Let γ s , s ∈ [0, ]
be the geodesic from O to ξ(s) . Let us parametrize γ s to be γ s : [0, r] → M where
r = %(x). Hence γ = γ0 is a normal geodesic.
Hence, we have the following variation F : [0, r] × [0, ] → M, where F(t, s) =

γ s (t). Notice that the corresponding variational field V(t), satisfies V(0) = O, and
∂ ∂
V(r) = | s=0 F|t=r = | s=0 F(r, s)
∂s ∂s
∂
= | s=0 ξ(s) = ξ̇(0) = V
∂s
Now we compute
Hess%2 (V, V) =(%2 ◦ ξ)00 (O)

d2 2
= (% ◦ ξ(s))| s=0
ds2
!2
d2 d
=2%(ξ(s)) 2 %(ξ(s))| s=0 + 2 %(ξ(s))| s=0
ds ds
!2
d2 d
=2r 2 %(ξ(s))| s=0 + 2 %(ξ(s))| s=0
ds ds
Recall by Cartan-Hadamard Theorem, %(ξ(s)) = d(ξ(s), O) = L(γ s ). Therefore

Lemma 5.8 tells us Hess%2 (V, V) = 2rL00 (0) + 2(L0 (0))2 = 2rL00 (0) + 2(hV, γ̇(r)i)2 .
Notice that we have ∇V V|t=0orr = 0, and hence
Z r
L00 (0) = h∇T V ⊥ , ∇T V ⊥ i − hR(V ⊥ , γ̇)γ̇, V ⊥ i dt
Z0 r
≥ h∇T V ⊥ , ∇T V ⊥ idt
0
Rb
That is, Hess%2 (V, V) ≥ 2r a
h∇T V ⊥ , ∇T V ⊥ idt + 2(hV, γ̇(r)i)2 .
1. If hV, γ̇i , 0, we obtain Hess%2 (V, V) ≥ 2hV, γ̇(r)i2 > 0.
2. Otherwise if hV, γ̇i = 0. Since V is a Jacobi field and hV, γ̇(0)i = h0, γ̇(0)i = 0,
proposition (5.8) tells hV(T ), γ̇(t)i = 0 ∀t ∈ [0, r]. Therefore, V(t) = V ⊥ (t).
We observe that ∇T V . 0. Since otherwise V(t) is parallel along γ, which
contradicts
Rr to the fact V(0) = 0, V(r) = V , 0. That is, Hess%2 (V, V) ≥
0
h∇T V , ∇T V ⊥ idt > 0.
⊥
Definition 5.8 (Convex and totally convex subsets of M). Let M be a Riemannian
manifold. A subset Ω ⊂ M is called convex, if whenever p, q ∈ Ω and γ is a minimizing
geodesic from p to q, then γ ⊂ Ω. Ω is called totally convex if whenever p, q ∈ Ω and
γ is a geodesic from p to q, then γ ⊂ Ω.
Recall by Cartan-Hadamard Theorem, on a complete simply-connected Rieman-

nian manifold with nonpositive sectional curvature, and geodesic is minimizing, and,
hence, any convex subset is totally convex. However these two concepts do have dif-
ference.
Example. On S2 ⊂ R3 the unit sphere {p ∈ S2 |d(p, O) < r} where r ≤ π2 , is convex,
but is not totally convvex. {p ∈ S2 |d(p, O) < π2 } is not convex.
Convex functions and convex subsets are related by the following result.
Proposition 5.12. Let τ : M → R be a convex function on a complete Riemannian
manifod M. Then the sub-level set
Mc = {x ∈ M : τ(x) < c}
is totally convex.
Proof. ∀p, q ∈ Mc , and any geodesic γ : [a, b] → M from p to q, we have (τ ◦ γ)00 ≥ 0.
Therefore τ ◦ γ : [a, b] → R : [a, b] → R attains its maximum at the two ends. Hence
τ ◦ γ(t) ≤ max{τ ◦ γ((a), τ ◦ γ(b)} = max{τ(p), τ(q)} < c
.
This is γ ⊂ Mc .

Therefore, Theorem 5.11 tells that any (open or closed) geodesic balls
{x ∈ M : d(x, O) < (≤)r}
is totally convex on a complete simply-connected Riemannian maniflod with nonposi-

tive curvature. In particular, every point is totallly convex (i.e. no nontrivial geodesic
γ : [a, b] → M with γ(a) = γ(B) = x exists).
Proper totally convex sets (Ω , M) do not exists in many manifolds. Existence of
such kind of subsets has significant to pological implications.
Theorem 5.12 (The Soul Theorem, Cheeger-Gromoll 1972). If (M, g) is a complete
non-compact Riemannian manifold with nonnegative sectional curvature, then M con-
tains a closed totally convex submanifold S , such that M is diffeomorphic to the normal
bundle over S .
S is called a soul of M.
{(x, y, z)|z = c} is a soul of the cylinder.

We also explain the geodesic meaning of the local result Lemma 5.7.
Theorem 5.13 (Whitehead 1932). Let (M, g) be a Riemannian manifold. Any p ∈ M

has a convex neighborhood.
Proof. Recall for any p ∈ M, there exists a totally normal neighborhood, that is, a
neighborhood p ∈ W and a number δ > 0 such that any two q1 , q2 ∈ W can be joined
by a unique minimizing geodesic. However, such a geodesic may not lie completely in
W.
By Lemma 5.7, there exists a neighborhood U p s.t. d2 (·, p) is strictly convex in
U p . Pick r small enough s.t. B p (r) = {q ∈ M, d(q, p) < r} ⊂ U p ∩ W. The proof of
proporsition 5.12 tells B p (r) is convex.

Chapter 6
Comparison Theorem
Recall in our discusions about Cartan-Hadamard and Bonnet-Myers Theorems, we, in

fact, have model spaces in mind.
1. Cartan-Hadamard: use Rn as a model, and replace the ”zero curvature” of ”Rn ”
by ”cur ≤ 0”.
2. Bonnet-Myers: use Sn as a model, and replace ”Ricci curvature = n − 1” of S n
by ”Ricci cur≥ (n − 1)”.
In this chapter, we aim at estabilshing quantitative comparison result with model spaces.
6.1 Sturm Comparison Theorem1

We start from a pure analysis result of Sturm.
Theorem 6.1 (Sturm). Let f and h be two continuous functions satisfying f (t) ≤ h(t)
for all t in an interval I, and let φ and η be two functions satisfying the differential
equations:
φ + f φ = 0
 00

 η00 + hη = 0 on I.
 (6.1.1)
Assume that φ is not the zero function and let a, b ∈ I be two consecutive zeros of
φ. Then:
1 Spivak IV, Chap 8, 15-17
157
158 CHAPTER 6. COMPARISON THEOREM
1. The function η must have a zero in (a, b), unless f = h everywhere on [a, b] and
η is a constant multiple of φ on [a, b],
2. Suppose that η(a) = 0, and also η0 (a) = φ0 (a) > 0. If τ is the smallest zero of η
in (a, b), then
φ(t) ≥ η(t) f or a ≤ t ≤ τ
and equality holds for some t only if f = h on [a, t].
Remark: The restriction η0 (a) = φ0 (a) > 0 in the theorem 6.1 case 2 can be achieved
by choosing a suitable multiple of η, and changing φ to −φ if necessary.
Proof. (6.1.1) gives
ηφ00 − φη00 = (h − f )φη. (6.1.2)

Suppose that η were nowhere zero on (a, b). W.o.l.g., we can assume
η, φ > 0 on (a, b). (6.1.3)

Thus (6.1.2) gives
ηφ00 − φη00 = (h − f )φη ≥ 0.

Therefore,
Z b Z b
0≤ (ηφ − φη ) =
00 00
(ηφ0 − φη0 )0
a a
. (6.1.4)
= η(b)φ0 (b) − η(a)φ0 (b) − φ(b)η0 (b) + φ(a)η0 (a)
= η(b)φ0 (b) − η(a)φ0 (a)
On the other hand, (6.1.3) implies
φ0 (a) > 0,φ0 (b) < 0


 ⇒ η(b)φ (b) − η(a)φ (a) ≤ 0.
0 0
(6.1.5)

η(a) ≥ 0,η(b) ≥ 0 
If f , h, we have
Z b
0< (ηφ00 − φη00 ).
a
which is a contradiction to (6.1.5). Hence η must have a zero on [a, b).
If f = h, then by (6.1.4) and (6.1.5)
0 = η(b)φ0 (b) − η(a)φ0 (a) ⇒ η(a) = η(b) = 0.

Now φ and η satisfy the same equation
φ + f φ = 0
 00

 η00 + f η = 0 on [a, b].

and φ(a) = φ(b) = 0. The solution η must be a constant multiple of φ on [a, b].
6.1. STURM COMPARISON THEOREM 159
Next suppose η(a) = 0, η0 (a) = φ0 (a) > 0. (recall φ(a) = 0). Let τ be the smallest
zero of η in [a, b]. Then φ > 0, η > 0 on (a, τ). Hence
(φ0 η − η0 φ)0 = φ00 η − η00 φ = (h − f )φη ≥ 0
on (a, τ). Recall φ0 (a)η(a) − η0 (a)φ(a) = 0, this implies φ0 η − η0 φ ≥ 0 on (a, τ).

Since η > 0 on (a, τ), we obtain
!0
φ0 η − η0 φ φ
= ≥0
η2 η
on (a, τ).
But by L’Hôpital’s Rule and our assumption, we have
φ(t) φ0 (t)
lim = lim 0 = 1.
t→a η(t) t→a η (t)
φ
Therefore, η ≥ 1 on (a, τ).
This proves φ(t) ≥ η(t) for a ≤ t ≤ τ. If φ(t) = η(t) for some t, then φη = 1 on (a, t).
0
Hence φη = 0 ⇒ φ0 η − η0 φ = 0 on (a, t) ⇒ φ00 η − η00 φ = 0 on (a, t) ⇒ f = h on (a, t).
By continuity, f = h on [a, t].

Geometric translations
Theorem 6.2 (Bonnet 1855). Let M be a surface, and γ : [0, L] → M be a normal
geodesic. Let k > 0.
1. If K(p) ≤ k for all p = γ(t), and γ has length L < √π then γ contains no
k
conjugate points.
2. If K(p) ≥ k for all p = γ(t), and γ has length L > √πk then there is a point
τ ∈ (0, L) conjugate to 0, and therefore γ is not of minimal length.
Proof. Let Y be a unit parallel vector field along γ with hY, γ̇i = 0, ∀t. Any normal
Jacobi field U can be written as U = φY for some function φ.
Jacobi equation ∇T ∇T U + R(U, T )T = 0 implies
φ00 + K(Y, T )φ(t) = 0. (6.1.6)

The corresponding discussion on constant curved surfaes (model spaces) gives
η00 (t) + kη(t) = 0 (6.1.7)

√ π
with a solution η(t) = sin( kt). Note 0 and √k are two consecutive zeros of η.
√ cannot have a solution φ

1. K(Y, T ) ≤ k, ∀t. Theorem 6.1 case (1) implies (6.1.6)
vanishing at 0 and at L < √πk . Since otherwise, sin( kt) has to vanish at some
point on (0, L), which is false.
Theorem 6.1 case (1) implies any Jacobi field φY must has a
2. K(Y, T ) ≥ k, ∀t.
π
zero on 0, √k ⊂ (0, L). So if we choose any nonzero Jacobi field Y along γ with
η(0) = 0, this Jacobi field will also vanish at some τ ∈ (0, L). Thus τ is conjugate
to 0.
Theorem 6.2 case (2) is the result which Bonnet used to show his diameter estimate.
From the above proof, we observe the following facts: The Jacobi field U = φY
where Y is a unit parallel vector field along γ with hY, γ̇i = 0, we have
φ00 (t) + K(γ(t))φ(t) = 0,

φ(a) = 0 ⇔ U(a) = 0,

φ0 (a) = U̇(a) ,
|φ(a)| = |U(a)| .
So Sturm comparison theorem case(2) can be translated as:

Given two surfaces M and M̄. Let γ : [a, b] → M and γ̄ : [a, b] → M̄ be two
normal geodesics such that
K(γ(t)) ≤ K̄(γ̄(t)). (6.1.8)
Let τ ∈ (a, b] such that γ, γ̄ have no point in [a, τ] conjugate to γ(a), γ̄(a) respec-
tively.
Let U, Ū be normal Jacobi fields along γ, γ̄ respectively with U(a) = Ū(a) = 0 and

U̇(a) = Ū˙ (a). Then U̇(a) ≥ Ū˙ (a), for a ≤ t ≤ τ. And ’=’ holds for some t only if

K ◦ γ = K̄ ◦ γ̄ on [a, t].
Remark The above ”comparison of Jacobi fields” implies Bonnet Theorem 6.2 by
choose one of M, M̄ to be the sphere S2 √1 . In fact in theorem 6.1, case (2) ⇒ (1)
k
when η(a) = 0 is the case.
6.2 Morse-Schoenberg Comparison and Rauch Com-

parison Theorems2
It is natural to ask for higher-dimensional generalizations of the geometric translation
of Theorem (6.1). Let (M, g) be an n-dimensional Riemann manifold γ : [a, b] → M
2 Spivak IV, Chap 8, 18-23

6.2. MORSE-SCHOENBERG COMPARISON AND RAUCH COMPARISON THEOREMS161
be a normal geodesic.
Now a normal Jacobi field U along γ cannot always be written as φY where Y is a unit
normal parallel vector field along γ. In fact, let {Y1 , . . . , Yn } be an orthornormal parallel
vector field along γ with γ̇(t) = Y1 (t). Then a normal Jacobi field U along γ can be
written as
X n
U(t) = φi (t)Yi (t).
i=2
n n
φ00i (t)Yi (t) + φi (t)R(Yi (t), T )T = 0
P P
Jacobi equation ⇒
i=2 i=2
n
⇒ φ00j (t) + φi (t)hR(Yi , T )T, Y j i = 0 ∀2 ≤ j ≤ n.
P
i=2
This system of equations do not involve the sectional curvature directly.
hR(Y2 , T )T, Y2 i . . . hR(Y2 , T )T, Yn i

 
d2 .. .. ..

(φ2 (t), . . . , φn (t)) + (φ2 (t), . . . , φn (t))   = 0.

dt 2  . . . 
hR(Yn , T )T, Y2 i . . . hR(Yn , T )T, Yn i
Recall the space of normal Jacobi fields along γ vanishing at t = a is of dimension
n − 1. We actually have to solve the following to solve the equation to compute Jacobi
fields:
 2
 dtd 2 A + AR = 0


(6.2.1)
A(0) = 0 dA (0) = Idn−1


dt

where R = hR(Yi , T )T, Y j i is symmetric.
ij
We will not discuss the generalization of Theorem (6.1) to the equation (6.2.1),
but instead, will discuss the generalizaion of its geometric translations. These two are
different aspect of the same result.(See[WSY, Chap8, Appendix])
From the geometirc viewpoint, we are going to compare the Jacobi fields along
geodesics in two Riemann manifolds, whose sectional curvatures satisfy certain com-
parison estimate. For that purpose, we need ”move” a vector field along a geodesic γ
to a geodesic γ̄ in another Riemann manifold M̄.
Lemma 6.1. Let (M, g), ( M̄, ḡ) be two Riemann manifolds of the same dimension n,
and let γ (γ̄) : [a, b] → M ( M̄) be a normal geodesic in M ( M̄). Then there is a vector
space isomorphism
Φ : {piecewise C ∞ vector f ields along γ̄} → {piecewise C ∞ vector f ields along γ}
such that for all t ∈ [a, b], we have for any piecewise C ∞ vector field X along γ
∇ ∂ X is continuous at t, then ∇T̄ Φ(X) B e

1. If ∇T X B e ∇ ∂ Φ(X) is continuous at t.
∂t ∂t
2. hX(t), T (t)ig = hΦ(X)(t), T̄ (t)iḡ .
3. |X(t)|g = |Φ(X)(t)|ḡ , where |X(t)|g =

p
g(X(t), X(t)).
4. |∇T X|g = |∇T̄ Φ(X)|ḡ , it being understood that this equation refers to left and right
hand limit at discontinuous points.
where T = γ̇ and T̄ = γ̄˙ .
Proof.
What could be a natural choice of such a Φ?

An isomorphism between T γ(t0 ) M and T γ̄(t0 ) M̄ for a fixed point is easy: Pick φt0 :
T γ(t0 ) M → T γ̄(t0 ) M̄ be one isomorphism which preserve the inner products given by g
and ḡ respectively.
How to extend it?
For any t ∈ [a, b], we define φt : T γ(t) M → T γ̄ (t) M̄, which is given by
parallel transport along γ toγ(t0 ) φt0

Vt −−−−−−−−−−−−−−−−−−−−−−−−→ Pγ,t,t0 (Vt ) −−→
φt0 (Pγ,t,t0 (Vt )) 7−→ Pγ̄,t0 ,t (φt0 (Pγ,t,t0 (Vt )))

Then define Φ(X)(t) = φt (X(t)) .
Next, we give Φ an explicit expression. Let Y1 , . . . , Yn be parallel, everywhere
orthonormal vector fields along γ with Y1 (t0 ) = γ̇(t0 ). Let Ȳ1 , . . . , Ȳn be parallel, every-
where orthonormal vector fields along γ with Ȳ1 (t0 ) = γ̄˙ (t0 ).
A piecewise C ∞ vector field X(t) along γ can be written as
n
X
X(t) = fi (t)Yi (t)
i=1
for certain functions fi : [a, b] → R.

φt0 is chosen such that φt0 (Yi (t0 )) = Ȳi (t0 ), for i = 1, . . . , n. Then
n
X
Φ(X)(t) = φt (X(t)) = fi (t)Ȳi (t).
i=1
This shows Φ(X) is C ∞ everywhere that X is, and that

hX(t), γ̇(t)ig = f1 (t) = hΦ(X)(t), γ̄˙ (t)iḡ ,
Xn
|X(t)|g = fi2 (t) = |Φ(X)(t)|ḡ ,
i=1
n
X
|∇T X|g = fi0 (t) 2 = |∇T̄ Φ(X)|ḡ .

i=1

Theorem 6.3. Let M and M̄ be two Riemann manifold of the same dimension n, and
let γ (γ̄) : [a, b] → M ( M̄) be a normal geodesic in M ( M̄). For each t ∈ [a, b], sup-
pose that for all 2-dimensional sections Πγ(t) ⊂ T γ(t) M, and all 2-dimensional sections
Π̄γ̄(t) ⊂ T γ̄(t) M̄, the sectional curvatures satisfy K(Πγ(t) ) ≤ K̄(Π̄γ̄(t) ).
Then we have
ind(γ) ≤ ind(γ̄).
In particular, if I(W, W) < 0 for some W ∈ V0 (a, b), then also I( ¯ W̄, W̄) < 0 for some
W̄ ∈ V̄0 (a, b), where V0 (a, b) is the set of the piecewise C ∞ vector fields along γ with
W(a) = W(b) = 0.
Proof. Let W ∈ V0 (a, b), recall that
Z b
I(W, W) = {h∇T W, ∇T Wi − hR(W, T )T, Wi} dt.
a
Let Φ be constructed as in Lemma (6.1). Then Φ(W) ∈ V̄0 (a, b) and h∇T W, ∇T Wig =
h∇T̄ Φ(W), ∇T̄ Φ(W)iḡ . Also we have

hR(W, T )T, Wig = K(W, T ) hW, WihT, T i − hW, T i2

≤ K̄(Φ(W), T̄ ) hΦ(W), Φ(W)ihT̄ , T̄ i − hΦ(W), T̄ i2
= hR̄(Φ(W), T̄ )T̄ , Φ(W)i.
¯
That is, I(W, W) ≥ I(Φ(W), Φ(W)).
So if A ⊂ V0 (a, b) is a subspace on which I is negative definite, then Φ(A) ⊂
V̄0 (a, b) is a subspace of the same dimension on which I¯ is again negative definite. By
definition, this means ind(γ) ≤ ind(γ̄).

Corollary 6.1 (The Morse-Schoenberg Comparison Theorem). Let (M, g) be a Rie-
mann manifold of dimension n, and let γ : [0, L] → M be a normal geodesic. Let
k > 0.
1. If K(Πγ(t) ) ≤ k for all Πγ(t) ⊂ T γ(t) M, and γ has lenghth L < √π , then ind(γ) = 0
k
and γ contains no cojugate point.
2. f K(Πγ(t) ) ≥ k for all Πγ(t) ⊂ T γ(t) M, and γ has lenghth L > √πk , then there is a
point τ ∈ (0, L) conjugate to 0, and γ is not of minimal length.
Remark
1. This is a high-dimensional generalization of Bonnet Theorem (6.2).
2. Case 1 above is a generalization of proposition 5.10, which asserts M contains

no conjugate points if sec ≤ 0.

Proof. For case 1, we apply Theorem 6.3 to M = M, M̄ = S n √1k . Choosing γ :

[0, L] → M, γ̄ : [0, L] → S n √1k to be normal geodesic. We have ind(γ) ≤ ind(γ̄).
Now ind(γ̄) = 0 since γ̄ contains no conjugate points (The Morse Index Theorem).
Therefore ind(γ) = 0 which implies γ contains no conjugate point.
For case 2, similar argument. Recall case 2 has already been proved when we
discussed Bonnet-Myers Theorem. Now it is a good chance to understand the proof
π

there in a more structural way: We choose V(t) = sin L t E(t) along γ in M and show
I(V, V) < 0. Here V(t) is the image of a Jacobi field on Sn via the isomorphism map Φ
defined in Lemma 6.1.

Remark: Recall in the proof of Bonnet-Myers Theorem, we already show that

Corollary 1 case 2 can be improved by weakening the sectional curvature restriction
to Ricci curvature restriction.
We still miss the generalization of the 2nd part of Sturm comparison theorem: we
have not compared |Φ(W)|ḡ with |W|g up to the first zero of Φ(W). Such information is
providede by
Theorem 6.4 (Rauch Comparison Theorem). Let M, M̄ be two Riemann manifolds of

the same dimension n, and let γ : [a, b] → M, γ̄ : [a, b] → M̄ be normal geodesics.
Let U, Ū be normal Jacobi fields along γ, γ̄ respectively with U(a) = Ū(a) = 0, and
|∇T U(a)|g = ∇T̄ Ū(a)ḡ . Suppose:
1. γ̄ has no conjugate point on [a, b].
2. K(Πγ(t) ) ≤ K̄(Π̄γ̄(t) ) for all t ∈ [a, b], all 2-dimensional sections Πγ(t) ⊂ T γ(t) M,
Π̄γ̄(t) ⊂ T γ̄(t) M̄.

Then we have |U(t)|g ≥ Ū(t)ḡ , for all t ∈ [a, b].
Remark
1. This is a generalization of the second part of Sturm Comparison Theorem. No-

tice that the Morse-Schoenberg Comparison Theorem (Cor 6.1) is also a direct
consequence of theorem 6.4
2. ”U, Ū be normal Jacobi fields”: In fact, we only need require hU̇(a), γ̇(a)i =
hŪ˙ (a), γ̄˙ (a)i = 0. This is because, from proposition (5.8), U = f T + U ⊥ where
f is linear. And f (a) = 0, f 0 (a) = 0 forces f ≡ 0. Hence U = U ⊥ .
Proof. If Ū ≡ 0, it is trival.
If Ū . 0, then Ū(t) , 0 for all t ∈ (a, b], since γ̄ has no conjugate points.
It suffices to prove that
|U(t)|g
lim = 1, (6.2.2)
t→a Ū(t)
ḡ
d |U(t)|g
≥ 0 ∀t ∈ (a, b]. (6.2.3)
dt Ū(t)
ḡ
It turns out it is equivalent (but much easier) to consider the norm suquare.
hU(t), U(t)ig
lim = 1, (6.2.4)
t→a hŪ(t), Ū(t)iḡ
d hU(t), U(t)ig
≥ 0 ∀t ∈ (a, b]. (6.2.5)
dt hŪ(t), Ū(t)iḡ
To prove (6.2.4), we note that:
hU(t), U(t)ig hU(t), ∇T U(t)ig

lim = lim
t→a hŪ(t), Ū(t)iḡ t→a hŪ(t), ∇T̄ Ū(t)iḡ
h∇T U(t), ∇T U(t)ig + hU(t), ∇T ∇T U(t)ig
= lim
t→a h∇ Ū(t), ∇ Ū(t)iḡ + hŪ(t), ∇ ∇ Ū(t)iḡ
T̄ T̄ T̄ T̄
=1.
To prove (6.2.5), we note that
d hU(t), U(t)ig 2hU(t), ∇T U(t)ihŪ(t), Ū(t)i − 2hU(t), U(t)ihŪ(t), ∇T̄ Ū(t)i

= .
dt hŪ(t), Ū(t)iḡ hŪ(t), Ū(t)i2
(6.2.6)
Hence (6.2.5) ⇔
hU(t), ∇T U(t)ihŪ(t), Ū(t)i ≥ hU(t), U(t)ihŪ(t), ∇T̄ Ū(t)i. (6.2.7)
So for each t0 ∈ [a, b], it suffices to show that
hU, Ui(t0 )
hU, ∇T Ui(t0 ) ≥ hŪ, ∇T̄ Ūi(t0 ). (6.2.8)
hŪ, Ūi(t0 )
We denote hU,Ui(t0 )
hŪ,Ūi(t0 )
= c2 . For any piecewise C ∞ vector field W along γ, the index
form
Z b
I(W, W) = {h∇T W, ∇T Wi − h(R(W, T )T, Wi}dt
a
Z b b
=− h(R(W, T )T + ∇T ∇T W, Widt + h∇T W, Wi
a a
k
X
− h∇T (t+j ) W − ∇T (t−j ) W, Wi.
j=1
Since U, Ū are Jacobi fields, with U(a) = Ū(a) = 0, we have
hU, ∇T Ui(t0 ) = Iat0 (U, U),

hŪ, ∇T̄ Ūi(t0 ) = I¯at0 (Ū, Ū).

So, it remains to show Iat0 (U, U) ≥ c2 I¯at0 (Ū, Ū), or Iat0 |U(t0 )| , |U(t0 )| ≥ I¯at0 , Ū
U U Ū
|Ū(t0 )| |Ū(t0 )|
Consider the map Φ constructed in Lemma 6.1, as the input, at t0 , we choose φt0
U(t )
such that φt0 |U(t 0
= Ū(t 0)
.
0 )|
|Ū(t0 )|
Then Φ |U(t0 )| is a smooth vector field along γ̄, such that Φ |U(t
U U
(t0 ) = Ū(t 0)
.
0 )| |Ū(t0 )|
As in the proof of Theorem 6.3, we see
! ! !!
U U U U
Iat0 , ≥ I¯at0 Φ ,Φ . (6.2.9)
|U(t0 )| |U(t0 )| |U(t0 )| |U(t0 )|
Now using the minimizing property of Jacobi field in lemma (5.2), we have
! !!  
U U  Ū Ū
I¯a0 Φ ,Φ ,  .
t t

≥ I¯a0  (6.2.10)
|U(t0 )| |U(t0 )| Ū(t0 ) Ū(t0 )

This is applicable since Φ U
(a) = Ū(a)
= 0, and Φ U
(t0 ) = Ū(t0 )
=0
|U(t0 )| |Ū(t0 )| |U(t0 )| |Ū(t0 )|
Comnining (6.2.9) and (6.2.10) yields
!  
U U  Ū Ū 
Iat0 , ¯t0 
≥ Ia 
 ,  . (6.2.11)
|U(t0 )| |U(t0 )| Ū(t0 ) Ū(t0 ) 

Recall from the proof of uniqueness of simply-connected space forms (Theorem

5.10), we have used the idea of comparing the norm of Jacobi field. Therefore, we have
the same sectional curvatures, and the corresponding Jacobi field has the same norm.
(It definitely deserves to read through that proof again with this new perspective).
Recall for given t ∈ [a, b], the Jacobi field U(t) at t can be expressed as U(t) =
d expγ(a) (W) for some W.
(V)
Hence we have the following equivalent form of Rauch Comparison Theorem.
Theorem 6.5. Let M, M̄ be two Riemann manifolds of the same dimension n, and let
p ∈ M, p̄ ∈ M̄, φ : T p M → T p̄ M̄ be an isometry (of inner product spaces), V ∈ T p M,
V̄ = φ(V).
Let γ(t) = exp p tV, t ∈ [0, 1], γ̄ = exp p̄ tV̄, t ∈ [0, 1] be geodesics in M, M̄

respectively. Let X ∈ T V T p M , φ(X) ∈ T V̄ T p̄ M̄ . Suppose
1. γ̄ has no conjugate point.
2. K(Πγ(t) ) ≤ K̄(Π̄γ̄(t) ) for all t ∈ [0, 1], all 2-dimensional sections Πγ(t) ⊂ T γ(t) M,
Π̄γ̄(t) ⊂ T γ̄(t) M̄.
Then we have
d exp p (X) ≥ d exp p̄ (X̄).
(V) (V̄)
Proof. The geodesic variation
F(t, s) = exp p t(V + sX)
has variational field U(t) which is a Jacobi field such that
U(0) = 0,
U̇(0) = X,

U(1) = d exp p (X).
(V)
Similarly F̄(t, s) = exp p̄ t(V̄ + sX̄), gives Jacobi field Ū(t) such that
Ū(0) = 0,
Ū˙ (0) = φ(X) = X̄,

Ū(1) = d exp p (X̄).
(V̄)
Recall from Gauss lemma,

D E
hX, Vi = (d exp p )(V) (V), (d exp p )(V) (X)
and d exp p is an isometry along the radical direction, we only need to consider the
case hX, Vi = 0. (and, hence, hΦ(X), Φ(V)i = 0).
Therefore, theorem 6.5 follows from theorem 6.4.

A particular interesting case:
Corollary 6.2. Let (M, g) be a complete Riemann manifold with nonpositive sectional
curvature. Then ∀p ∈ M, exp p : T p M → M satisfies

d exp p (X) ≥ |X|
(V)
the right hand side means the norm of the flat metric on T p M. ∀V ∈ T p M, ∀X ∈
where
T V T p M ' T p M. In particular, for any curve γ ⊂ T p M, one have L(γ) ≤ L(exp p ◦γ).
Remark: This strengthen the result Proposition 5.10 where we show exp p has no
critical points.
Corollary 6.3. Let (M, g) be a complete simply-connected Riemann manifold with non-
positive sectional curvature. Consider a geodesic triangle in M(i.e. each side of the
triangle is a minimizing geodesic). Let the side lengths are a, b, c with opposite angles
A, B, C respectively.
Then
1. a2 + b2 − 2ab cos C ≤ c,
2. A + B + C ≤ π.
Moreover, if M has negative sectional curvature, then the inequalities are strict.
Proof. Denote the vertex at the angle C by p.

In T p M, draw a triangle 4OPQ,
−−→
where O is the origin, such that |OP| = a, |OQ| = b, ∠O = C. In particular, exp p OP,
−−→
exp p OQ is the other two verties of the geodesic triangle. Let ξ be the preimage of the
geodesic c in T p M. Then
|PQ| ≤ L(ξ) ≤ c. (6.2.12)
The Euclidean cosine law tells |PQ|2 = a2 + b2 − 2ab cos C.

Since a, b, c satisfy triangle inequalities, we can construct a triangle in R2 with side
length a, b, c. Denote the corresponding opposite angles by A0 , B0 , C 0 . Then.
(6.2.12) ⇒ a2 + b2 − 2ab cos C 0 = c2 ≥ a2 + b2 − 2ab cos C

⇒ C 0 ≥ C.
Similarly, we show A0 ≥ A, B0 ≥ B. Hence π = A0 + B0 + C 0 ≥ A + B + C.

When sec < 0, the inequality in Rauch’s theorem is also strict. And hence inequal-
ity in Cor (6.2) is strict. the last conclusion of this corallary then follows.

Final Remark: ”γ̄ hs no conjugate point” in Rauch Comparison theorem is nec-

essary. For example, let us consider two spheres M = S 2 (2), M̄ = S 2 (3). Let
γ : [0, 3π] → S 2 (2), γ̄ : [0, 3π] → S 2 (3) are normal geodesics. Let W, W̄ be unit
parallel normal vector fields along γ, γ̄ respectively. Then
t
U(t) = 2 sin W(t),
2
t
Ū(t) = 3 sin W̄(t).
3

are Jacobi fields such that U(0) = Ū(0) = 0, U̇(0) = Ū˙ (0) = 1.

Recall sec S 2 (r) = r12 . sec S 2 (2) = 41 > sec S 2 (3) = 19 . but Ū(3π) = 0 <

|U(3π)| = 2 sin 3π
2 = 2.

6.3 Cut Point and Cut Locus

Next, we are going to discuss two more important comparison theorems:
Hessian and Laplacian comparison theorems. That is, we compare Hess% of the
distance function %(x) B d(O, x) on different Riemannian manifolds whose sectional
curvatures can be ”well-compared”. Laplacian is the trace of Hessian, so it is nat-
ural to expect a Laplacian comparison result based on Ricci curvature-comparison-
assumption.
Hess% is closely related to the SVF. This has been shown when we discussed
d2
the convexity of %2 in (V)§(9). Recall Hess% (V, V) = ds 2 % (ξ(s)), where ξ is the
s=0
2
geodesic with ξ(0) = X, ξ̇(0) = V. And we hope to calculate ds d
2 % (ξ(s)) via the sec-
s=0
ond variation formula of length. For that purpose, we consider the family of geodesics
γ s : [0, %(x)] → M from 0 to ξ(s), we hope
1. %(ξ(s)) = Length(γ s ),
2. F(s, t) B γ s (t) gives a variation.
For this purpose, we have to require that γ does not contain ”cut point”! Recall in
the comparison of Jacobi fields (Rauch), we have to do the comparison before the first
conjugate point.
Another motivation is from the Bonnet/ Morse-Schoenberg comparison theorem.
K ≥ k and geodesic γ of length > √πk contains conjugate point, and hence γ is not
minimizing ⇒ diam≤ √πk .
A counterexample is the projective space Pn (Rn ), with constant sectional curvature
=1 and diameter only π2 . Hence, we may add the hypothesis of simply-connectivity.
Question: Should a complete simply-connected manifold with all sectional curva-
ture ≤ k have diameter ≥ √πk ?
One might expect to prove as follows: Pick p, q ∈ M maximal distance apart, and
consider a minimizing geodesic
γ : [0, L] → M
from p to q. If L < √πk , then γ contains no conjugate point on (0, L). Extend γ to
γ0 : [0, L0 ] → M where L < L0 < √πk . Thus γ0 contains no conjugate point on (0, L0 ),
and hence γ is a local minimum for length.
Notice, if we can conclude γ is a global minimum for length, we will get a contra-
dictionand conclude diam≥ √πk .
Example: consider S 1 × R
6.3. CUT POINT AND CUT LOCUS 171
choose γ to be a horizontal circle. curvature=0, Cartan-Hadamard ⇒ no conjugate

point, but γ is not global minimizing, although it is loally minimizing.
There are NO general criterion to decide whether a given geodesic is minimizing.
For simplicity, we will deal only with the case of a complete Riemannian manifold
(M, g). Let γ : [0, ∞) → M be a normal geodesic starting at a point O = γ(0) in M.
Then γ|[0,t0 ] is minimizing iff d (O, γ(t0 )) = t0 = Length γ|[0,t0 ] .

By triangle inequality, we see if γ|[0,t0 ] is minimizing so does γ|[0,t] for any t < t0 .
On the other hand, we know for small enough t, γ|[0,t] is minimizing. So the set
A = {t > 0 : d(O, γ(t)) = t}

= {t > 0 : γ|[0,t] is a minimizing geodesic}
is either (0, ∞) or (0, a] for some a > 0.(Notice that, the latter case means γ|[0,t] is
not minimizing for all t > a).
1. If A = (0, a], we say that γ(a) is the cut point of O along the geodesic γ.
2. If A = (0, ∞), we say that O has no cut point along γ.
3. The cut locus C(O) ⊂ M of O is the set of all points which are cut points of O
along some normal geodesic starting from O.
4. The cut locus C̃(O) of O in T 0 M is the set of all vectors aX ∈ T 0 M for which
X is a unit vector and expO aX is the cut point of O along the geodesic γX (t) =
expO tX.
Thus we have
expO C̃(O) = C(O).
The first conjugate locus of O in T 0 M is the set of all vectors aX ∈ T 0 M. For which
X is a unit vector and a is the first conjugate value of O along γX .
For a geodesic γ : [a, b] → M, recall that if ∃τ ∈ (a, b) conjugate to a, then γ
is not local minimizing. Hence γ also contains a cut point γ(t0 ), with t0 ≤ t. Briefly
expressed, the cut point comes before or at the first conjugate point.
What happens if t0 , τ?
In the example of a cylinder S 1 × R, there are no conjugate points. But when a cut
point happens,
we see there are two minimizing geodesic from γ(a) to γ(t0 ). In fact, if there are
two distinct minimizing geodesics γ, c from γ(a) to γ(t0 ), we see γ|[a,t0 +] can not be
minimizing for any > 0, This is because the path ξ : [a, t0 + ] → M,

c(t), t ∈ [a, t + ]
 0
ξ(t) = 

 γ(t), t ∈ [t0 t0 + ]
is of the same length as γ|[a,t0 +] and is not smooth at t0 , and hence it can be made
shorter.
In other words, if there are two distinct minimizing geodesics from γ(a) to γ(t0 ),
and t0 is the minimal value such that this happens, γ(t0 ) is a cut point of γ(a) along γ.
Indeed, this is the only possible case.
Theorem 6.6 (cut points). Let (M, g) be complete and let γ : [0, ∞) → M be a normal
geodesic with cut point γ(a). Then at least one of the following holds:
1. The number a is first conjugate value of O along γ,
2. These are at least two minimal geodesic from p = γ(0) to q = γ(a). And a is the
minimal value such that this case happens.
Proof. Choose a sequence a1 > a2 > a3 such that lim ai = a.
t→∞
Since γ(a) is a cut point, bi = d(p, γ(ai )) < ai . Let Xi ∈ T p M be the unit vectors such
that t → exp p tXi , 0 ≤ t ≤ bi is a minimal geodesic from p to γ(ai ). Let X = γ̇(0).
Since γ is normal, we have γ(t) = exp p tX, 0 ≤ t ≤ a. All Xi are distinct from X. We
have
lim bi = lim d(p, γ(ai )) = d(p, γ(a)) = a.
t→∞ t→∞
Therefore, the vectors bi Xi are contained in a compact subset of T p M. Choosing a
subsequence if necessary, we have
lim bi Xi = aY
t→∞
for some unit vector Y ∈ T p M.

Since exp p (aY) = lim exp p (bi Yi ) = lim γ(ai ) = γ(a), the geodesic t 7→ exp p (tY),
i→∞ i→∞
0 ≤ t ≤ a is a minimizing geodesic from p to q.
If X , Y, we have two minimizing geodesics from p to q. If ∃0 < t0 < a, such that
there is another geodesic c from p to γ(t0 ),
which is of the same length as γ|[0,t0 ] , we observe the path ξ : [0, a] → M,


 c(t), t ∈ [0, t0 ]
ξ(t) = 

γ(t), t ∈ [t , a]
0
can be made shorter. But Length ξ|[0,a] = Length γ|[0,a] . This contradicts to the

fact that γ(a) is a cut point.
If X = Y, we have lim bi Xi = aY = aX = lim ai X. But exp p (bi Xi ) = γ(ai ) =
i→∞ i→∞
exp p (ai X). Moreover, ai X and bi Xi are distinct for each i, since bi < ai , Xi , X. So
exp p is not 1-1 in any small neighborhood of aX in T p M, i.e. aX is a critical point of
exp p . By theorem (5.5, we conclude that a is a conjugate value of O along γ.

Recall along a geodesic γ : [a, b] → M, p is a conjugate to q implies q is also

conjugate to p. It turns out that this is also true for cut points.
Corollary 6.4. In a complete manifold M, if q is the cut point of p along a geodesic γ

from p to q, then p is the cut point of q along the geodesic γ obtained by traversing γ
in the opposite direction.
Proof. By assumption, we have that γ is a minimizing geodesic from p to q. So γ is

also minimizing from p to q. So the cut point of q along γ, if exists, occurs past or at
p.
If q is conjugate to p along γ, then p is conjugate to q along γ. The cut point of q

along γ must occur before or at p. So if must occur at p.
If there is another minimizing geodesic from p to q, then again p must be the cut
point.

Let p ∈ M. Denote by S p the unit sphere of T p M. Let R∗ = R ∪ {∞} be the real

numbers together with some other set ”∞”. Consider the following topology of R∗ :
a basis consists of all sets of the form (a, b) ⊂ R together with all sets of the form
(a, ∞] = (a, ∞) ∪ {∞}.
We now define a function τ : S p → R∗ by
a > 0,i f exp p (aX) is the cut point o f p along the geodesic γX = exp p tX


τ(x) = 

 ∞,i f γX has no cut point

Theorem 6.7. If M is a complete manifold, and p ∈ M. Then the function τ : S p → R∗

is continuous.
Proof. Let X1 , X2 , . . . be a sequence of unit vectors in S p converging to X ∈ S p . We

have to show {τ(Xi )} converge to τ(X). Since the values of τ lie in the compact set
{α ∈ R∗ : α ≥ 0}, we can assume, by chooosing a subsequence, that τ(Xi ) converges to
some α ∈ R∗ .
If α = ∞, then for any given t, there exsits N s.t. τ(xi ) > t, ∀i ≥ N. hence
d(p, exp p tX) = d p, lim exp p tXi = t. By definition, τ(X) ≥ t.
i→∞
If α < ∞, then τ(Xi )Xi → αX. Therefore d(p, exp p αX) = α. Hence τ(X) ≥ α. So
limi→∞ τ(Xi ) ≤ τ(X)(♥).
Next, if τ(X) > α, then exp p (αX) is not a conjugate point of p along exp p tX since a
conjugate point cannot come before a cut point. So the map exp p is a diffeomorphism
on some neighborhood U around αX in T p M. W.l.o.g., we assume all τ(Xi )Xi lie in U.
Therefore exp p τ(Xi )Xi is not a conjugate point of p along exp p tXi . then theorem
6.6 implies that there exists another minimizing geodesic from p to exp p τ(Xi )Xi , i.e.
∃Yi ∈ S p s,t, exp p τ(Xi )Xi = exp p τ(X)Yi .
Since exp p |U is 1-1, we have τ(Xi )Yi < U. By choosing a subsequence, we can
assume Yi → Y. Then αY also lies outside U. So
exp p (αY) = lim exp p (τ(Xi )Yi ) = lim (τ(Xi )Xi ) = exp p (αX).
i→∞ i→∞
That is, exp p (tX), t ∈ [0, α] and exp p (tY), t ∈ [0, α] are two distinct minimizing
geodesics from p to exp p (αX). This contradicts to α < τ(X). Hence τ(X) ≤ α.
Combinig with (♥), we obtain
τ ≤ lim inf τ(Xi ) ≤ lim sup ≤ τ(X).

i→∞ i→∞

Corollary 6.5. The cut locus C(p) of p ∈ M and the cut locus C̃(p) of p in T p M are
closed subset of M and T p M respectively.
Proof. Let q ∈ M s.t. ∃pi ∈ C(p) with pi → q. Let γi (t) be the minimizing normal
geodesic from p to pi with γi (ti ) = pi . Then ti = τ(γ̇i (0)). W.l.o.g., we can assume
γ̇i (0) converges to Y ∈ S p . Then
q = lim pi = lim γ(ti ) = lim exp p (τ(γ̇i (0))γ̇i (0)) = exp p (τ(Y)Y) ∈ C(p).
i→∞ i→∞ i→∞
C̃(p) is the preimage of C(p) under exp p , and hence also closed.

Theorem 6.8. Let M be complete, p ∈ M. Define
E(p) = {tV : V ∈ S p and 0 ≤ t < τ(V)}.
Then exp p maps E(p) differmorphically onto an open subset of M, and M is the disjoint
union of exp E(p) and C(p).
Proof. Clearly, d exp p is 1-1 on E(p), since the first conjugate point can not occur
before the cut point. Next we show exp p is 1-1 on E(p). Suppose not, ∃ω1 , ω2 ∈ E(p)
with ||ω1 || ≤ ||ω2 ||, say, such that exp p tω1 = exp p tω2 = q.
Then the geodesic exp p tω1 , t ∈ [0, 1], has length from p to q less than or equal to
that of the geodesic exp p tω2 , t ∈ [0, 1]. But these contradicts to the minimal property
of exp p tω2 , t ∈ [0, 1 + ], since q comes before the cut point of γ. Therefore, exp p is
diffeomorphic on E(p) onto an open subset of M(since E(p) open).
We next show that exp E(p) and C(p) are disjiont. If not, ∃ω ∈ E(p) and V ∈ C̃(p)
with exp p ω = exp p V = q
If ||V|| ≤ ||ω||, similar arguments as above, we have a contradiction.
If, otherwise, ||V|| > ||ω||, then the geodesic exp p tV, t ∈ [0, 1] is longer than the
geodesic exp p tω, t ∈ [0, 1] from p to q. This contradicts to the minimizing property of
exp p tV, t ∈ [0, 1](since V ∈ C̃(p)).
Clearly, expE(p)∪C(p) ⊂ M. On the other hand, ∀q ∈ M completeness implies
that there is a normal minimizing geodesic γ(t) = exp p tV from p = γ(0) to q = γ(a).
Clearly a ≤ τ(V) so aV ∈ E(p) or aV ∈ C̃(p).

Remark(injectivity raius): Recall the injectivity radius of p ∈ M is defined as
i(p) = sup{ρ > 0 : exp p is a di f f eomorphism on B(O, ρ) ⊂ T p M}
Then Theorem 6.6 and theorem 6.8 together implies
i(p) = sup{ρ > 0 : B(O, ρ) ⊂ E(p)}.
Notice each ray tX, X ∈ S p intersect with C̃(p) at at most one point. Hence, we
have
Proposition 6.1. The cut locus C̃(p) of p in T p M is of zero measure for any p ∈ M.
Observe that for any compact Riemannian manifold M given p ∈ M, each ray tX
must intersect with c̃(p). In fact, we have
Proposition 6.2. A complete Riemannian manifold M is compact i f f for any p ∈ M,

the cut locus C̃(p) is homeomorphisc to the unit sphere S p ⊂ T p M.
Proof. If M is compact, Let diam(M)=δ. Then any normal geodesic from p of length>
δ is not minimizing. Hence τ(X) < δ + 1, ∀X ∈ S p . Define the map
β : S p → C̃(p)
X 7→ τ(X)X
So β is a homeomorphism.
On the other hand, if C̃(p) is homeomorphic to S p ⊂ T p M, we have in particular
C̃(p) is compact. Let A ⊂ S p be the set of X ∈ S p s.t. τ(X) < ∞, i.e. A = τ−1 ([0, ∞]).
Then the map β : A → C̃(p) is a homeomorphism. This tells further that A is compact.
So A is closed subset. Theorem 6.6 implies that A is open: ∀X ∈ A, ∃U, s,t, X ∈ U ⊂
A.(Since otherwise, ∃Xi → X s.t. τ(Xi ) → ∞. Continuity of τ tells τ(X) = ∞. This
contradicts to X ∈ A.)
Therefore A = S p . That is, every geodesic from p has a cut point. In other words,
∀X ∈ S p , τ(X) < ∞. Hence max < ∞. ∀p, q ∈ M, ∃ a minimizing geodesic γ from p
X∈S p
to q, s.t. d(p, q) = L(γ) ≤ max < ∞. ∀p, q ∈ M which implies that M is compact.
X∈S p

6.4 Hessian Comparison Theorem

Theorem 6.9 (Hessian Comparison). Let M, M be two Riemannian manifolds of the
same dimension n and let γ : [a, b] → M, γ : [a, b] → M be two normal geodesics.
Denote p = γ(a), p = γ(a), and % = d(p, .), % = d(p, .) be the distance function resp.
Suppose:
• γ|[a,b] and γ|[a,b] are minimizing and contain no cut point.
• K(Πγ(t) ) ≤ K(Πγ ) for all t ∈ [a, b], all 2-dim sections.
Then we have %, % are C ∞ in a neighborhood of γ, γ resp(except p,p). and Hess% ≥

Hess% along γ, γ.
Remark
1. ”Hess% ≥ Hess% along γ, γ” means for anyDt ∈ (a, Eb], and for any X ∈ T γ(t) M, X ∈
T γ(t) M satisfying |X|g = |X|g , hX, γ̇(t)ig = X, γ̇(t) , and we have Hess%(X, X) ≥
g
Hess%(X.X).
2. Since γ(b) is not a cut point, we have (b − a)γ̇(0) < C̃(p). By Corollary 6.5, C̃(p)
is closed, hence ∃ an open neighborhood U = {V ∈ T p M : |V − (b − a)γ̇(0)| < }
s.t. U ∩ C̃(p) = φ.
6.4. HESSIAN COMPARISON THEOREM 177

Notice we can write U = {(b − a)V ∈ T p M : |V − γ̇(0)| < b−a }. Let U = {(t −

a)V ∈ T p M : t ∈ [a, b], |V − γ̇(0)| < b−a }. Theorem 6.8 tells that exp p on U is a
diffeomorphism, and %(exp p W) = |W|, ∀W ∈ U. So % is C ∞ on exp p U\{p}. This
proves the smoothness of %, % claimed in the theorem.
E any X ∈ T γ(t) M for some t, we can decompose X = aγ̇(t) + X , where

⊥
Proof.
D For
X , γ(t) = 0.
⊥ ˙
Observe that

Hess% γ(t),
˙ γ(t)
˙ = ∇2 % γ(t),
˙ γ(t)
˙ = ∇(∇%) γ(t),
˙ γ(t)
˙
= ∇γ̇ ∇γ̇ % − ∇∇γ̇ γ̇ % = γ̇γ̇% − ∇γ̇ γ̇% = 0.

Hess% γ̇, X ⊥ = ∇X ⊥ ∇γ̇ % − ∇∇X⊥ γ(t)
˙ %
= − (∇X ⊥ γ̇) (%) = − h∇X ⊥ γ̇, grad%i .
(Recall γ̇% = 1 = hγ̇, grad%i, and hgrad%, Ei = 0 for any hE, γ̇i = 0.)
We have
1
grad% = γ̇ = − h∇X ⊥ γ̇, γ̇i = − X ⊥ hγ̇, γ̇i = 0.
2
That is

Hess%(X, X) = Hess% aγ̇(t) + X ⊥ , aγ̇(t) + X ⊥ = Hess%(X ⊥ , X ⊥ ).
So we only need to consider X ∈ T γ(t) M, X ∈ T γ(t) M, which are perpendic to γ̇(t)

and γ̇(t) resp..
W.l.o.g., we let t = b. X ∈ T γ(b) M, X ∈ T γ M, hX, γ̇(b)i = 0, hX, γ̇(t)i = 0.
Let ξ be the geodesic s.t. ξ(0) = γ(b), ξ̇(0) = X. Since γ(b) is not a cut point, by
the same argument as in Remark (2), there exists a neighborhood U s.t. exp p : U →
exp p U is a diffeomorphism, and ∀W ∈ U, exp p tW is minimizing. Therefore ∃ > 0
s.t. the minimizing geodesic γ s : [a, b] → M, s ∈ [0, ] from p to ξ(s) forms a well-
defined variation F(t, s) = γ s (t), t ∈ [a, b], s ∈ [0, ]. The corresponding variational
∂
field U(t) = ∂s | s=0 F(t, s) is a Jacobi field with U(0) = 0, U(b) = X. And U is also a
normal field.
Notice that %(ξ(s)) = Length(γ s ) = L(γ s ) = L(s). We have Hess%(X, X) = L00 (s).
Recall from lemma 5.8.
b Z b
L (0) = h∇U U, γ̇ia +
00
h∇T U ⊥ , ∇T U ⊥ i − hR(U ⊥ , T )T, U ⊥ idt = I(U, U).
a
Similarly, Hess%(X, X) = I(U, U), where U is a normal Jacobi field along γ with
U(0) = 0, U(b) = X. It remains to show I(U, U) ≥ I(U, U).
Using the construction of Φ in Lemma 6.1. (firstly, choose φb : T γ(b) M → T γ(b) M be

the isometry of inner product spaces which sends X to X, γ̇(b) to γ̇. This is possible
since |X| = |X|, hX, γ̇(b)i = hX, γ̇(b)i, |γ̇(b)| = |γ̇(b)| = 1).
As in Theorem 6.3, we have I(U, U) ≥ I(Φ(U), Φ(U)). By minimizing property of
Jacobi field I(Φ(U), Φ(U)) ≥ I(U, U). So I(U, U) ≥ I(U, U).

Corollary 6.6. Under the same assumption of Thm6.9, and let f : [0, b − a] → R be a
C ∞ function which satisfies f 0 ≥ 0. Then we have Hess f (%) ≥ Hess f (%) along γ, γ.
Proof. Note for any X ∈ T γ(t) M for any t,
d2
Hess f (%)(X, X) = | s=0 f (%(ξ(s)))
ds2 !
d 0 d
= f (%(ξ(s))) %(ξ(s)) | s=0
ds ds
!2
d d2
= f (%(t))
00
| s=0 %(ξ(s)) + f 0 (%(t)) 2 | s=0 %(ξ(s))
ds ds
= f 00 (%) (X%)2 + f 0 (%)Hess%(X, X)
= f 00 (%)hX, grad%i2 + f 0 (%)Hess%(X, X)
= f 00 (%)hX, γ̇(t)i2 + f 0 (%)Hess%(X, X).
Hence we have Hess f (%)(X, X) ≥ Hess f (%)(X, X) for X, X s.t. |X| = |X|, hX, γ̇(t)i =
hX, γ̇(t)i.
Example(Hess% on manifolds with constant sectional curvature)

6.4. HESSIAN COMPARISON THEOREM 179
Let γ : [0, b] → M be a normal minimizing geodesic. From the proof of thm6.9,

we see for X ∈ T γ(b) M, hX, γ̇(b)i = 0,
Hess%(X, X) = I(U, U)
where U is a normal Jacobi field along γ with U(0) = 0, U(b) = X. Recall

Z b
I(U, U) = h∇T U, ∇T Ui − hR(U, T )T, Uidt
a
Z b
=− h∇T ∇T U + R(U, T )T, Uidt + h∇T U, Ui|ba
a
= h∇T U(b), U(b)i.
That is Hess%(X, X) = h∇T U(b), U(b)i.

Let γ̇(t), E2 (t), . . . , En (T ) be parallel orthonormal vector fields along γ. If M has
constant sectional curvature K, the Jacobi fields U with U(0) = 0 is given by U(t) =
f (t) ni=2 Ei (t), where
P
 f (t) + K f (t) = 0
 00

 f (0) = 0

ct, K = 0


√



⇒ f (t) = c sin Kt, K > 0

f or some constant c.

c sinh √−Kt, K < 0




Hence ∇T U(b) = f 0 (b) Ei (b), h∇T U(b), U(b)i = (n − 1) f 0 (b) f (b) and |X|2 =
P
(n − 1) f (b)2 . Therefore
Hess%(X, X) = (n − 1) f 0 (b) f (b)

f 0 (b) 2
= |X| .
f (b)
In particular, if K = 0, Hess%(X, X) = 1% g(X, X) for < X, γ̇ >= 0(FACT1).

In general, for X ∈ T γ(b) M, we have
f 0 (b) ⊥ 2
Hess%(X, X) = Hess%(X ⊥ , X ⊥ ) = |X |
f (b)
f 0 (b)
= hX − hX, γ̇(b)iγ̇(b), X − hX, γ̇(b)iγ̇(b)i
f (b)
f 0 (b)
= hX, Xi − hX, γ̇(b)i2 .
f (b)
Again when K = 0, we have
Hess%2 (X, X) = 2(X%)2 + 2%Hess%(X, X)

= 2hX, γ̇(b)i2 + 2g(X ⊥ , X ⊥ )
= 2g(X, X) (FACT2).
Corollary 6.7. Let M be a complete simply-connected Riemannian manifold with non-

positive sectional curvature, % = d(p, .), p ∈ M. Then on M \ {p}, we have
Hess%2 ≥ 2g, (6.4.1)
n−1
∆% ≥ . (6.4.2)
%
Proof. By Cartan-Hadamard, there is no cut point. Then apply Thm 6.9 and FACT2,
we get (6.4.1), and apply FACT1, we get (6.4.2).
Remark:
1. Corollary(6.7) is a stengthen result of the convexity (Theorem 5.11). We dis-
cussed in the end of last chapter.
2. Under the same assumption of thm(6.9), by taking trace, we obtain ∆%(γ(t)) ≥
∆%(γ(t)) (F).
A natural question is: In order to obtain the comparison (F), is it enough to assume
Ricci curvature comparison instead of sectional curvature comparison?
6.5 Laplacian Comparison Theorem

Theorem 6.10 (Laplacian Comparison). Let M, M be two Riemannian manifolds of
the same dimension n and let γ : [a, b] → M, γ : [a, b] → M be two normal geodesics.
Denote p = γ(a), p = γ(a), and % = d(p, .), % = d(p, .) be the distance function resp.
Ric, Ric, ∆, ∆ be the Ricci curvature tensor and Laplacian of M, M resp.
Suppose:
• 1- γ|[a,b] and γ|[a,b] are minimizing and contain no cut point.

• 2- Ric(γ̇, γ̇)(t) ≤ Ric γ̇, γ̇ (t), ∀t ∈ [a, b].
• 3- M is a space form of sectional curvature k.
Thus we have ∆%(γ(t)) ≥ ∆%((t)), ∀t ∈ [a, b]. Moreover , ∆%(γ(b)) = ∆%((b)) iff
∀t ∈ [0, b], any section in tangent bundle of M containing γ̇ has sectional curvature
k, and any normal Jacobi field U(t) along γ with U(a) = 0 can be represented as
U(t) = f (t)E(t), where E(t) is parallel along γ and f : [a, b] → R is a solution of
f + kf = 0
 00

 f (0) = 0

Remark:
0
(t)
1. On a space form of sectional curvature k, we have ∆% (γ(t)) = (n − 1) ff (t) where
ct, k = 0


√



f (t) =  c sin( kt), k > 0


c sinh( √−kt), k < 0




6.5. LAPLACIAN COMPARISON THEOREM 181
2. We will explain why we have to add the assumption (3) during the proof.
Proof. W.l.o.g., we consider t = b, and an orthonormal basis {e1 , . . . , en } of T γ(b) M

with e1 = γ̇(b). Then Delta%(γ(b)) = ni=2 Hess%(ei , ei ). Same argument as in the proof
P
of Theorem 6.9, we obtain Hess%(ei , ei ) = I(Ui , Ui ), where Ui is a normal Jacobi field
along γ with Ui (0) = 0, Ui (b) = ei .
Do similar thing on M: let {e1 , . . . , en } be a orthonormal basis of T γ(b) M with e1 =
γ̇(b). We have Hess%(ei , ei ) = I(U i , U i ). Then let φb : T γ(b) M → T γ(b) M be the isometry
with φb (ei ) = ei , i = 1, . . . , n. Then construct Φ as in Lemma 6.1 optinal ransport. Then
we obtain (n − 1) vector fields Φ(Ui ), i = 1, . . . , n along γ with
Φ(Ui )(a) =0 = U i (a),

Φ(Ui )(b) = φb (Ui (b)) =φb (ei ) = ei = U i (b).
By the minimizing property of Jacobi fields, we have
I (Φ(Ui ), Φ(Ui )) ≥ I(U i , U i ) (6.5.1)
for i = 2, . . . , n
The Laplacian Comparison is reduced to show
n
X n
X
I (Ui , Ui ) ≥ I(U i , U i ). (6.5.2)
i=2 i=2
Since
n
X n
X
I (Φ(Ui ), Φ(Ui )) ≥ I(U i , U i ). (6.5.3)
i=2 i=2
it is enough to show
n
X n
X
I (Ui , Ui ) ≥ I(U i , U i ). (6.5.4)
i=2 i=2
Recall, if we have ”sectional-curvature comparison”, we conclude the above in-

equlity immediately. If we only have ”Ricci-curvature comparison”, it turns out, we
have to assume assumption 3 in order to get (6.5.4).
(6.5.4) ⇔
n Z
X b
(h∇T Ui , ∇T Ui i − hR(Ui , T )T, Ui i) dt
i=2 a
Xn Z b
≥ h∇T Φ(Ui ), ∇T Φ(Ui )i − hR(Φ(Ui ), T )T , Φ(Ui )i dt.
i=2 a
Lemma 6.1 ⇒ |∇T Ui | = |∇T Φ(Ui )|. Hence

(6.5.4) ⇔
Z b n
X
hR(Ui , T )T, Ui idt
a i=2
Z b n
X
≤ hR(Φ(Ui ), T )T , Φ(Ui )idt.
a i=2
Notice that at γ(b), T = e1 , Ui = ei is an orthornomal basisand ni=2 hR(Ui , T )T, Ui i(b) =

P
Ric (γ̇, γ̇) (b). Similarly, ni=2 hR(Φ(Ui ), T )T , Φ(Ui )i(b) = Ric γ̇, γ̇ (b). But in general,
P
{T, U2 , . . . , Un } is not orthonormal any more at t , b. A solution is to add the assump-
tion. In that case, M is a space form of constant curvature k. And we know the Jacobi
field Ui (t) = f (t)ei (t), where {ei (t)} is a parallel orthononormal vector fields along γ
with ei (b) = ei and f is a solution of
f + kf = 0
 00

 f (a) = 0, f (b) = 1

Hence ni=2 hR(Ui , T )T, Ui i(t) = f 2 (t) ni=2 hR(ei , T )T, ei i = f 2 (t)Ric(T, T ) = f 2 (t)Ric(γ̇, γ̇)(t).
P P
of Φ, we know Φ(Ui ) = f (t)ei (t). So ni=2 hR(Φ(Ui ),T )T, Φ(Ui )i =

P
And by the construction

f 2 (t)Ric γ̇, γ̇ . The assumption (2) implies f 2 (t)Ric(γ̇, γ̇)(t) ≤ f 2 (t)Ric γ̇, γ̇ , ∀t ∈
[a, b]. We then prove (6.5.4) and hence the Laplacian comparison.
If ∆%(γ(b)) = ∆%(γ(b)), we have ”=” holds in (6.5.1). Recall the minimizing prop-
erty of Jacobi field, this can only happen when ∀i = 2, . . . , n, Φ(Ui ) = U i . By the
construction, we know Φ(Ui ) = f (t)ei (t). Any normal Jacobi field U(t) along γ with
U(0) = 0 cab be expressed as a linear combination of Φ(Ui ), i = 2, . . . , n.
n
X
U= ci f (t)ei (t), ci ∈ R
i=2
n
X
= f (t) ci ei (t).
i=2
By Jacobi equation, Φ(Ui ) = f (t)ei (t) is a Jacobi field implies hR(ei , T )T, ei i = k.

Corollary 6.8. Under the same assumption of thm (6.10), and let f : [a, b] → R be a
smooth function with f 0 ≥ 0. Then
∆ f (%)(γ(t)) ≥ ∆ f (%)((t))
∀t ∈ [a, b]
Proof.
∆ f (%) = f 00 (%) + f 0 (%)∆%,
∆ f (%) = f 00 (%) + f 0 (%)∆%.

6.6. COMMENTS ON INJECTIVITY RADIUS ESTIMATE AND SPHERE THEOREMS183
6.6 Comments on Injectivity Radius Estimate and Sphere

Theorems
Is it always true for a simply-connected complete two dimensional Riemannian mani-
fold (M, g) with Gauss curvature ≤ β, β > 0 that expO : T O M → M is a diffeomorphism
on B(0, √πβ ).
This is not always true unfortunately.

By comparison theorem, Gauss curvature ≤ β implies expO B(0, √πβ ) contains no
conjugate point of O. But this is not enough to conclude that expO is a diffeomorphism.
Actually, in other words, we are asked to show in j M ≥ √πβ . For that result, we need
further restrict Gauss curvature > 0. This is actually Klingenberg’s injectivity radius
estimate.
Theorem 6.11 (Klingenberg,1959). Suppose (M, g) is an orientable even-dimensional

manifold with 0 <sectional curvature≤ β. Then in j M ≥ √πβ . If M is not orientable, then
in j M ≥ 2 √π β .
For the proof, we refer to [Spivak IV, Chap8 34-36] or [PP,§6.2]. Here, we explain
the rough ideas

Proof. We need to show expO B(0, √πβ ) has no cut point. If there is a cut point of q
of O along γ, since q can not be a conjugate point of O, we have by 6.6, there exists
exactly two minimal geodesics from O to q. (denoted by γ,γ1 )
In fact, one can further argue that when q is the closest one to O in C(0), γ̇|q =
−γ̇1 |q . Moreover, when O is a point such that it is the minimum point of the function
d(p, C(p)), we have the two geodesics γ, γ1 give a close geodesic.
Recall in the proof of Synge theorem, under the assumption ”orientable, even-
dimensional”, any closed geodesic has a variation F(t, s) such that the curves γ s (t) =
F(t, s) has length L(γ s (t)) < L(γ0 (t)), with 0 < s small. Hence the whole curve γ s lie
in the interior of the cut locus. So there exists a minimal geodesic σ s from γ s (0) to the
fartherst point γ s (t s ) along γ s .
Choose subsequence if necessary, those geodesics σ s converge to a minimal geodesic
σ from O to q which is different from γ and γ1 . This contradicts to the assumption that
q is a cut point and γ, γ1 are the only minimizing geodesic from O to q. So there is no
cut point in expO B(0, √π ) .
β
A much deeper result by Klingenberg asserts that if a simply-connected manifold

has all its sectional curvature in the interval ( 41 β, β], then in j M ≥ πβ . There are further
improvement on the left end of the interval.
Actually, those injectivity radius estimate and Rauch comparison theorem are cru-
cial tools to establish fascinating Sphere Theorems
Theorem 6.12 (Topological Sphere Theorem). Let (M, g) be a simply-connected com-

plete Riemannian manifold. Suppose M has all its sectional curvatures in the interval
( 14 β, β], β > 0. Then M is homeomorphic to the sphere.
This result is due to Rauch (prove in the case sec ∈ ( 43 β, β]), Klingenberg, Berger.
Very brief explanation: In topology, Brown theorem tells that: if a compact mani-
fold M is the union of two open sets, each of which is diffeomorphic to Rn , then M is
diffeomorphic to Sn .
So let p, q ∈ M, s.t. d(p, q) = diam(M, g) ≤ √2πβ . exp p : T p M → M, expq : T q M →
M are diffeomorphisms on B(0, δ p ), B(0, δq ) at least when δ p , δq are small enough.
exp p (B(0, δ p )) ∪ expq (B(0, δq )) ⊂ M
On the other hand, if we have δ p , δq large enough, we have exp p (B(0, δ p ))∪expq (B(0, δq )) =
M.
Scaling β to be 1, diam(M, g) ≤ 2π, then Klingenberg ⇒ in j M ≥ π. It remain to
show
M ⊂ exp p (B(0, δ p )) ∪ expq (B(0, δq ))
That is for any x ∈ M, if d(p, x) ≥ in j M ≥ π, then we need show d(q, x) < in j M . For
this purpose, we need a global version of the Rauch theorem:
Toponogov triangle comparison theorem. (This has been discussed in the tutorial)
6.7 Volume Coparison Theorems

Now let us come back to investigate a geometric quantity which we have discussed at
the very beginning of this course: the volume.
Recall E(p) = {tV : V ∈ S p and 0 ≤ t < τ(V)} from 6.8. Let us denote by
E p = exp p E(p). We have shown that
exp p : E(p) → E p
is a diffeomorhism, and E(p) is diffeomorphic to an open ball. Since the cut locus is of
zero measure, we have
Z Z
Vol(M) = dvol = dvol.
M Ep
Note E p ⊂ M can be considered as a coordinate neighborhood!

Hence
6.7. VOLUME COPARISON THEOREMS 185
Z q
Vol(M) = det(gi j )dx1 . . . dxn
Ep
Z q
= det(gi j )dx1 ∧ · · · ∧ dxn .
Ep
(This is another way to see the definition of volume does not depend on orientabil-
ity).
For the ball B p (r) = {q ∈ M|d(p, q) < r} we have Vol(B p (r)) = B (r) dvol =
R
R p
p 1 n
E p ∩B p (r)
det(gi j )dx ∧ · · · ∧ dx .
How to calculate vol(B p (r)) in a Riemannian manifold?
We first prepare some algebraic results: Let f : Ṽ → V be a linear transformation
between two n-dimensional inner product spaces. Let {ẽ1 , . . . , ẽn }, {e1 , . . . , en } be their
orthonormal basis, and {ω̃1 , . . . , ω̃n }, {ω1 , . . . , ωn } be their dual basis. Let Ω̃ = ω̃1 ∧
· · · ∧ ω̃n , Ω = ω1 ∧ · · · ∧ ωn . Then f ∗ Ω is defined as

f ∗ Ω(X1 , . . . , Xn ) = Ω f (X̃1 ), . . . , f (X̃n ) .
Therefore, ∃a0 ∈ R s.t. f ∗ ω = a0 Ω̃.

On the other hand, let f (ẽi ) = nj=1 αij e j , i = 1, . . . , n. Then a0 = det[aij ] = det( f ).
P
Claim: Let {Ã1 , . . . , Ãn } be a basis of Ṽ. Then
f ∗Ω | f (Ã1 ) ∧ · · · ∧ f (Ãn )|
|a0 | = |det( f )| = | |= .
Ω̃ |Ã1 ∧ · · · ∧ Ãn |
Proof. Recall ∀Xi ∈ V, i = 1, . . . , n, we have
1
|X1 ∧ · · · ∧ Xn | = hX1 ∧ · · · ∧ Xn , X1 ∧ · · · ∧ Xn i 2
q
= det[< Xi , X j >].
In particular, if {Xi } are orthonormal, |X1 ∧ · · · ∧ Xn | = 1. Therefore, letting Ãi =

βij ẽ j , i = 1, . . . , n. We have
Pn
j=1
| f (Ã1 ) ∧ · · · ∧ f (Ãn )| = |det[βij ]| × | f (ẽ1 ) ∧ · · · ∧ f (ẽ j )|

= |det[βij ]| × |det[αij ]| × |e1 ∧ · · · ∧ en |
= |det[βij ]| × |det( f )|.
On the other hand,
|Ã1 ∧ · · · ∧ Ãn | = |det[βij ]| |ẽ1 ∧ · · · ∧ ẽn | = |det[βij ]|.
This proves the claim.

Now, let γ : [a, b] → M be a normal geodesic with γ(0) = p. Let Ω̃ be a volume

n-form of the Euclidean space T p M, and Ω(t) be a volume n-form at γ(t) ∈ M. Then
we have
exp p : T p M → M,

(d exp p )tγ̇(0) : T tγ̇(0) T p M → T γ(t) M.
Define a function φ : [0, b] → R to be

∗
(d exp p )tγ̇(0) (Ω(t))

φ(t) = .
Ω̃
Note that Ω(t), Ω̃ depend on the choice of different orientations, but φ does not. We
can always choose proper orientation s.t.
(d exp p )∗ (Ω(t))
φ(t) = .
Ω̃
(we omit the subscript: tγ̇(0)).
(because we are working in the single coordinate neighborhood E p !!)
We in fact can define a function. φ : E(p) → R as below: ∀ỹ ∈ E(p),
∗ ∗
(d exp p ) (Ω(y)) (d exp p ) (Ω(y))
φ(ỹ) = =
Ω̃(ỹ) Ω̃(ỹ)

where y = exp p ỹ.

Then we can rewrite the volume of M as
Z Z
Vol(M) = Vol(E p ) = Ω= Ω
Ep exp p (E(p))
Z Z
= (d exp p )∗ Ω = φΩ̃
E(p) E(p)
Z
= φdvolT p M
E(p)
and Vol(B p (r)) = E(p)∩B(0,r) φdvolT p M .

R
So the key point is to complete the funtion φ, for which we need employ results
about Jacobi fields again.
Lemma 6.2. Let γ : [0, b] → M be a normal geodesic containing no conjugate point.

Let J1 , . . . , Jn−1 be (n-1) linearly independent normal Jacobi fields along γ with Ji (0) =
0, i = 1, 2, . . . , n − 1. Thus we have
|J1 (t) ∧ · · · ∧ Jn−1 (t)|
φ(t) =
t | J˙1 (0) ∧ · · · ∧ J˙n−1 (0)|
n−1
t ∈ (0, b]
Proof.
d exp p : T γ(t) T p M → T γ(t) M.
Note that J1 (t), . . . , Jn−1 (t), γ̇ are n linearly independent vectors in T γ (t)M, and
hence form a basis of T γ(t) M. By ”dimension consideration”,, this also implies J˙1 (0) . . . J˙n−1 (0)
are linearly independent.
hJi (t), γ̇(t)i = 0, t ∈ [0, b] implies h J˙i (0), γ̇(0)i = 0. Hence J˙1 (0) . . . J˙n−1 (0), γ̇(0)
form a basis of T p M. Recall for the variation F(t, s) = exp p t(γ̇(0) + sW), we have
its
variational
field U(t) is a Jacobi field with U(0) = 0, U̇(0) = W, and U(t) =
d exp p (tW). Pick W = J˙i (0), we then obtain Ji (t) = d exp p (t J˙i (0)). Notice
tγ̇(0) tγ̇(0)
that for any t ∈ (0, b], t J1 (0) . . . t Jn−1 (0), γ̇(0) also form a basis of T p M.
˙ ˙
Hence
|γ̇(t) ∧ J1 (t) ∧ · · · ∧ Jn−1 (t)|
φ(t) =
γ̇(0) ∧ t J˙1 (0) ∧ · · · ∧ t J˙n−1 (0)
|J1 (t) ∧ · · · ∧ Jn−1 (t)|
= .
t J˙1 (0) ∧ · · · ∧ J˙n−1 (0)
n−1
Theorem 6.13 (Bishop). Let M be a Riemannian maninfold with Ric ≥ (n−1)k. Let γ :
[0, b] → M be a normal geodesic containing no cut point, then φφ(t)
k (t)
is nonincreasing,
∀t ∈ (0, b]. (φk is the funcion on the simply-connected space form M k of sectional
curvature k).
f (t) n−1
By lemma 6.2, we can check φk (t) = k
t , where
f, k = 0






 1 √
 √ sin kt, k > 0



fk (t) = 


 k

 1 √
 √ sinh −kt, k < 0





−k
We will show this result by reduce it to the Laplacian comparison via the following
lemma.
Lemma 6.3. Let γ : [0, b] → M be a normal geodesic with no cut point of γ(0). Let
%(x) = d(x, γ(0)). Then
φ0
!
n−1
(t) = ∆% − (γ(t)) (6.7.1)
φ %
t ∈ (0, b]
Proof. We only need prove (6.7.1) at γ(b). Let J1 , . . . , Jn−1 be Jacobi fields along γ
with Ji (0) = 0, i = 1, . . . , n − 1 s.t. hJi (b), J j (b)i = δi j , 1 ≤ i, j ≤ n − 1. Recall
Hessρ(γ̇, γ̇) = 0. Hence we have

n−1
X
∆%(γ(b)) = Hess%(Ji (b), Ji (b)) (6.7.2)
i=1
n−1
X n−1
X
= I(Ji (b), Ji (b)) = h∇T J˙i (b), Ji (b)i. (6.7.3)
i=1 i=1
On the other hand,

φ0 φ
d 2
(b) = dt 2 (b)
φ 2φ
where
|J1 (b) ∧ · · · ∧ Jn−1 (b)|
φ(b) =
bn−1 J˙1 (0) ∧ · · · ∧ J˙n−1 (0)
1
= .
b J1 (0) ∧ · · · ∧ J˙n−1 (0)
n−1 ˙
We have
d 2 d hJ1 (t) ∧ · · · ∧ Jn−1 (t), J1 (t) ∧ · · · ∧ Jn−1 (t)i
φ (b) = |t=b 2
dt dt

t2n−2 J˙1 (0) ∧ · · · ∧ J˙n−1 (0)
2 n−1 ˙ |J1 (t) ∧ · · · ∧ Jn−1 (t)|2
P
i=1 hJ1 (t) ∧ · · · ∧ Ji (t) ∧ · · · ∧ Jn−1 (t), J1 (t) ∧ · · · ∧ Jn−1 (t)i
= 2 − 2(n − 1) 2 |t=
t2n−2 J˙1 (0) ∧ · · · ∧ J˙n−1 (0) t2n−1 J˙1 (0) ∧ · · · ∧ J˙n−1 (0)
 Pn−1
 i=1 h J˙i (b), Ji (b)i n − 1 

2
=  − 2n−1  .
J˙ (0) ∧ · · · ∧ J˙ (0)2  b2n−2 b
1 n−1
Hecne, we calculate
φ0 (b) φ
d 2 n−1
X n−1
= dt 2 (b) = h J˙i (b), Ji (b)i − . (6.7.4)
φ(b) 2φ i=1
b
Combining (6.7.2), (6.7.3) and (6.7.4), we obtain
φ0 n−1
(b) = ∆%(γ(b)) −
φ %(γ(b))
!
n−1
= ∆% − (γ(b)).
%

Proof. of Theorem 6.13. Let γ : [0, b] → M k be a normal geodesic in the simply-

connected space form M k of sectional curvature k.
Note that γ has no cut point iff b < √πk ( √πk = ∞ if k < 0). By our assumption
γ : [0, b] → M is a normal geodesic in M without cut point, hence b < √πk , by Bonnet-
Myers Theorem. Therefore, the above assertion implies γ : [0, b] → M k has no cut
point too. So the Laplacian comparison Theorem is applicable. Hence Lemma 6.3 tells
φ0 φ0
(t) ≥ k (t)
φ φk
0 0
Hence
φ 0 (lnφ) (t) ≥ (lnφk ) (t), which implies (lnφ − lnφk )0 (t) ≥ 0. This means
φk (t) ≥ 0. This completes the proof.

Theorem 6.14 (Bishop-Gromov). If (M, g) is a complete Riemannian manifold with
Ric ≥ (n − 1)k, k ∈ R. Let p ∈ M be an arbitrary point. Then the function
vol(B p (r))
r 7→
vol(Bk (r))
is nondecreasing, where Bk (r) is a geodesic ball of radius r in the simply-connected
space form M k .
g) is a complete Riemannian manifold with Ric ≥ (n −

Corollary 6.9 (Bishop). If (M,
1)k > 0. Then Vol(M) ≤ Vol S n ( √1k ) . The equality holds iff M is isometric to S k ( √1k ).
What is Vol(B p (r))? First note that when B p (r) ⊂ E p , we have
Z
Vol(B p (r)) = dvol M
B p (r)
Z
= φdvolT p M
B(0,r)
Z rZ
= φ(t, θ)tn−1 dtdθ.
0 Sn−1
The assumption B p (r) ⊂ E p means rθ ∈ E(p) for any θ ∈ Sn−1 .

How to go beyond cut point?
Let χ be the characteristic function of E(p) ⊂ T M , i.e.
1, when(r, θ) ∈ E(p)


χ(r, θ) = 

0, otherwise
Then for any B p (r) ⊂ M,
Z
vol(B p (r)) = φdvolT p M
B(0,r)∩E p
Z
= χφdvolT p M
B(0,r)
Z rZ
= χ(t, θ)φ(t, θ)tn−1 dtdθ.
0 Sn−1
Remark: Recall that for (t, 0) ∈ E(p), φ(t, θ)tn−1 = |J1 (t)∧···∧Jn−1 (t)|
. On the simply-
| J˙1 (0)∧···∧ J˙n−1 (0)|
connected space form M k , we have φk (t, θ)tn−1 = φk (t)t = ( fk (t)) , where
n−1 n−1
f, k = 0






 1 √
 √ sin kt, k > 0



fk (t) = 


 k

 1 √
 √ sinh −kt, k < 0





−k
We also define the characteristi funtion χk on M k . Recall, when k ≤ 0, χk ≡ 1, when
k > 0, χk = 0 only at the one point. Since ∀p ∈ M k , C(p) = −p and d (p, C(p)) = √πk .
Therefore we have χ(r, θ) ≤ χk (r, θ) = χk (r). That is, the function
χ(r, θ)
r 7→
χk (r)
is non-increasing (where we use 00 = 0).

Recall Theorem 6.13 tells r 7→ φ(r,θ)
φk (r) is non-increasing (r < τ(0)). Hence
χ(r, θ)φ(r, θ)
r 7→ (6.7.5)
χk (r)φk (r)
is non-creasing.
Consider the function
Z
a(t) = χφ(t, θ)tn−1 dθ,
Sn−1
Z
ak (t) = χk φk (t, θ)tn−1 dθ.
Sn−1
χφ(t,θ)
First observe ak (t) = χk φk (t)tn−1 vol(Sn−1 ). Hence we have =
a(t) 1
R
ak (t) vol(Sn−1 ) Sn−1 χk φk (t)
dθ.
Vol(B p (r))
(6.7.5) implies immediately that t 7→ a(t)
ak (t) is non-increasing. This tells r 7→ Vol(Bk (r)) =
Rr
a(t)dt
R 0r is non-increasing due to the following Lemma 6.4.
0
ak (t)dt

Lemma 6.4. Let f, g : [0, ∞) → (0, ∞) be two positive function and the funcion
R t
f (t) f
t 7→ g(t) is non-increasing. Then the funcion t 7→ R0t is also non-increasing.
0
g
Proof. Let us denote h = gf . For t1 ≤ t2 , we hope to show

R t1 R t2
0
f f
R t1 ≥ R0t2
0
g 0
g
Z t1 Z t2 Z t2 Z t1
i.e. f g≥ f g.
0 0 0 0
We observe
Z t1 Z t2 Z t1 Z t1 Z t1 Z t2
f g= f g+ f g,
0 0 0 0 0 t1
Z t2 Z t1 Z t2 Z t1 Z t1 Z t1
f g= f g+ f g.
0 0 t1 0 0 0
R t1 R t2 R t2 R t1
Hence it remains to show 0 f t g ≥ t f 0 g.
1 1
This follows from the calculation:
Z t1 Z t2 Z t1 Z t2 Z t1 ! Z t2
f g= gh g≥ g h(t1 ) g
0 t1 0 t1 0 t1
Z t1 ! Z t2 ! Z t2 Z t1
≥ g hg = f g.
0 t1 t1 0
Remark: Recall we actually have φ(r, θ) = det(gi j ) ◦ x−1 (r, θ), r < τ(θ), and
p
f (r) n−1
φk (r) = kr . limr→0 φ(r, θ) = 1 ⇒ limr→0 φ(r,θ)
φk (r) = 1. Hence Theorem 6.13 ⇒
φ(r, θ) ≤ φk (r), when r < τ(θ) ⇒ χφ(r, θ) ≤ χk φk (r) ⇒
π
!
Vol B p ( √ ) = Vol(M) ≤ Vol(Mk ). (6.7.6)
k
Vol(B p (r))
In fact, noe have limr→0 Vol(Bk (r))
= 1. So (6.7.6) can be derived from Theorem
6.14.
Proof. of Corollary 6.9.

Recall the sphere of radius √1k has constant sectional curvature k. Hence Theorem

6.13 ⇒ Vol(M) ≤ Vol Sn ( √1k ) . If ”=” holds, then all inequalities in the proof of
Theorem 6.13 should be ”=”. Particularly, ∆%(γ(t)) = ∆%k (γ(t)) for any t s.t. tγ̇ ∈
E(γ(0)) ⊂ T γ(0) M ∀γ.
Recall from the Laplacian comparison theorem. this means any secion in T γ(t) M
containing γ̇(t) has sectional curvature k. Since γ is arbitrary, we have M hasconstant
√1
sectional curvature k. So its universal covering space is isometric to S (by the
k
uniqueness of simply-connected space forms). But since Vol(M) = Vol S( √1k ) , we

have M is isometric to S √1k .

Next we explore two applications of Volume Comparison Theorem.

6.7.1 Maximal Diameter Theorem

Theorem 6.15 (Maximal Diameter Theorem, Shiu-Yuen Cheng 1975). Let M be a
complete Riemannian manifold with Ric ≥ (n − 1)k > 0 and diam M = √πk . Then M is

isometric to S √1k .
Remark
1. This is a good complement of Bonnet-Myers

Diameter Estimate: the ”=” holds
in Bonnet-Myers iff M is isometric to S √1 .
k
2. When assuming sec ≥ k > 0, this result has been proved by Toponogov in
1959. Cheng’s original proof uses his comparison theorems for first eigenvalues.
Shioham (Trans. AMS. 1983) gives a much more elementary proof using the
Volume Comparison .

By scaling, we only need deal with the case k = 1. Let p, q ∈ M be two points
such that d(p, q) = diam M = π. Then B p (r) ∩ Bq (π − r) = φ, ∀r ∈ [0.π]. Hence
Vol(B p (r)) + Vol(Bq (π − r)) ≤ Vol(M), ∀r ∈ [0.π].
Using Theorem 6.14, we have
Vol(M) ≥ Vol(B p (r)) + Vol(Bq (π − r))

Vol(B p (r)) Vol(Bq (π − r))
= Vol(B0 (r)) + Vol(B0 (π − r))
Vol(B0 (r)) Vol(B0 (π − r))
Vol(B p (π)) Vol(Bq (π))
≥ 0
Vol(B0 (r)) + Vol(B0 (π − r))
Vol(B (π)) Vol(B0 (π))
Vol(M)
= Vol(B0 (r)) + Vol(B0 (π − r))

Vol(B0 (π))
= Vol(M).
Hence all ”≤” are ”=”. Particulary,
Vol(B p (π)) Vol(B p (r))

=
Vol(B0 (π)) Vol(B0 (r))
∀r ∈ (0, π].
Vol(B p (π))
Let r → 0, we have 1 = Vol(B0 (π)) = Vol(M)
Vol(Sn (1)) . Then corollary 6.9 implies that M is
isometric to Sn (1).

6.7.2 Volume Growth Rate Estimate

Theorem 6.16. Let (M n , g) be a complete Riemannian manifold with Ric ≥ 0.
1. we have Vol(B p (r)) ≤ Vol(B0 (r)) = ωn rn and = holds iff M is isometric to Rn .

2. If, furthermore, M n is non-compact, then there exists a positive constant c de-

pending only on p and n such that Vol(B p (r) ≥ cr, for any r > 2. (Calabi,
Notices AMS 1975, ST Yau, Indiana Univ. Math J 1976. independently.)
Vol(B (r))
Proof. (1). follows directly from Theorem 6.13, and limr→0 Vol(B0p (r)) = 1. Then ”=”
case again following from that in Laplacian Comparison Theorem.
Vol(B (r))
(2). (Following Gromov). From the following proof, we see again that Vol(B0p (r))
decreases tells much more than only Vol(B p (r)) ≤ Vol(B0 (r)) !!
Since M is non-compact complete, for any p ∈ M, there exists a ray, i.e. a geodesic
γ : [0, ∞) → M with γ(0) = p, and d(p, γ(t)) = t, ∀t ≥ 0.
Let t > 23 , Theorem 6.13 tells

Vol Bγ(t) (t + 1) ωn (t + 1)n (t + 1)n
≤ = . (6.7.7)
ωn (t − 1)n (t − 1)n

Vol Bγ(t) (t − 1)
On the other hand,

Vol B p (1) Vol Bγ(t) (t + 1) − Vol Bγ(t) (t − 1) (t + 1)n − (t − 1)n
≤ ≤ .
(t − 1)n

Vol Bγ(t) (t − 1) Vol Bγ(t) (t − 1)
(6.7.8)
n
i.e. Vol Bγ(t) (t − 1) ≥ 1t (t+1)(t−1)
n −(t−1)n Vol(B p (t))t.
(t−1)n

Observe that ∃Cn > 0 s.t. (t+1)n −(t−1)n ≥ Cn on [ 23 , ∞). Since Bγ( 2r + 12 ) 2r − 21 ⊂
B p (r),
∀r > 2. ⇒
!!
r 1
Vol(B p (r)) ≥ Vol Bγ( 2r + 12 ) −
2 2
r 1
≥ Cn Vol(B p (1))( + ).
2 2

Final Remark: Recall we used Lapkacian Comparison to prove the volume com-
parison: Ric ≥ (n − 1)k ⇒ Vol(B p (r)) ≤ Vol(Bk (r)). (Inside the cut locus, this is due
to Bishop. Gromov made the crucial step to show it for any r, and to a ”full” use of
Vol(B (r))
the fact that Vol(Bkp(r)) decrease.) It is natural to ask the ”other direction”. Recall we do
not have the ”other direction” in Laplacian Comparison, but we do have it in Hessian
Laplacian.
Excercise(Gunther, 1960) Let (M, g) be a complete Riemannian manifold, with
sectional curvature ≤ k. Let B p (r) be a ball in M which does not meet the cut locus of
p. Then Vol(B p (r)) ≥ Vol(Bk (r)).
Chapter 7
Candidates for Synthetic

Curvature Conditions
In the last part of our course , we discuss properties of a Riemannian manifold which
does not necessarily depend on the smooth structure of the underlying spaces. Those
properties may be taken to be definition of a general (metric, measure) space with
aurvature restrictions.
7.1 Nonpositive Sectional Curvature and Convexity

Theorem 7.1. Let (M, g) be a complete, simply-connected Riemannian manifold with
nonpositive curvature. Let p ∈ M, γ : [0, 1] → M be a geodesic. Then
d2 (p, γ(t)) ≤ (1 − t)d2 (p, γ(0)) + td2 (p, γ(1)) − t(1 − t)d2 (γ(0), γ(1)). (7.1.1)
Remark: Actually, on a complete Riemannian manifold with sec ≤ 0, 7.1.1 holds

whenever γ(t) ⊂ E p . For simplicity, we suppose M is simply-connected, then E p = φ
and exp p : T p M → M is a diffeomorphism.
Proof. Let k0 : [0, 1] → R be given by
k0 (t) = (1 − t)d2 (p, γ(0)) + td2 (p, γ(1)) − t(1 − t)d2 (γ(0), γ(1)).
195
196 CHAPTER 7. CANDIDATES FOR SYNTHETIC CURVATURE CONDITIONS
We have
k0 (0) = d2 (p, γ(0)),

k0 (1) = d2 (p, γ(1)),
k000 (t) = −2d2 (γ(0), γ(1)) = −2|γ̇(t)|2 .
Let %(x) = d2 (x, p), then % ◦ γ(t) satisfies
% ◦ γ(0) = d2 (p, γ(0)) = k0 (0),

% ◦ γ(1) = d2 (p, γ(1)) = k0 (1),
% ◦ γ00 (t) = Hess% (γ̇(t), γ̇(t)) ≥ 2|γ̇(t)|2 = k000 (t).
Therefore the function h : [0, 1] → R given by h(t) = (% ◦ γ − k0 ) (t) satisfies that

h(0) = h(1) = 0, h00 (t) ≥ 0, ∀t ∈ [0, 1]. Therefore h(t) ≤ 0. (Convex functions attains
maximum on the boundary). That is % ◦ γ(t) ≤ k0 (t), ∀t ∈ [0, 1].
Corollary 7.1. Let (M, g) be a compact simply-connected Riemannian manifold with
sec ≤ 0. Let γ1 , γ2 : [0, 1] → M be geodesics with γ1 (0) = p = γ2 (0).
Then for 0 ≤ t ≤ 1,
d(γ1 (t), γ2 (t)) ≤ td(γ1 (1), γ2 (1)).
Proof. Appling Theorem 7.1 twice:
d2 (γ1 (1), γ2 (t)) ≤ td2 (γ1 (1), γ2 (1)) + (1 − t)d2 (γ1 (1), p) − t(1 − t)d2 (p, γ2 (1)), (7.1.2)
d2 (γ2 (1), γ1 (t)) ≤ td2 (γ1 (1), γ2 (t)) + (1 − t)d2 (γ2 (t), p) − t(1 − t)d2 (p, γ1 (1)). (7.1.3)
Inserting 7.1.2 into 7.1.3, and observing d2 (γ2 (t), p) = t2 d2 (γ1 (t), p), we complete
the proof.
Remark: The property (7.1.1) for all p and all geodesic γ is actually equivalent to
the nonpositive sectional curvature of M. Namely, if the sectional curvature ≥ k > 0 in
a neighborhood of p, then locally (% ◦ γ)00 (t) = Hess%(γ̇(t), γ̇(t)) ≤ Hess%(γ̇(t), γ̇(t)).
Then one can show ”>” in (7.1.1). In fact, this is taken to be the definition of a
length space with nonpositive sectional curvature in the sense of Alexandrow.
Corollary 7.1 is also equivalent to nonpositive sectional curvature, and is taken as a
general curvature bound notion by Busemann. see[Chap2, Jost, Nonpositive Curvature:
Geometric and Analytic Aspects, Birkhāuser].
7.1. NONPOSITIVE SECTIONAL CURVATURE AND CONVEXITY 197
Theorem 7.2 (Reshetnyak’s quadrilateral comparison theorem). Let (M, g) be a com-

pact simply-connected Riemannian manifold with sec ≤ 0. Let γ1 , γ2 : [0, 1] → M.
Then
d2 (γ1 (0), γ2 (1)) + d2 (γ2 (0), γ1 (1))

≤d2 (γ1 (0), γ2 (0)) + d2 (γ1 (1), γ2 (1)) + d2 (γ1 (0), γ1 (1))
+d2 (γ2 (0), γ2 (1)) − (d(γ1 (0), γ2 (0)) − d(γ1 (1), γ2 (1)))2 .
Proof. For simplicity, denote the distance as a1 , a2 , b1 , b2 , d1 , d2 as in the above figure.

Let δ be the geodesic from γ1 (0) to γ2 (1) whose length is d2 . Consider
dλ2,0 = d(γ2 (0), δ(λ)),

dλ1,1 = d(γ1 (1), δ(λ)).
Then theorem 7.1 ⇒
(dλ2,0 )2 ≤ (1 − λ)b21 + λa22 − λ(1 − λ)d22 ,

(dλ1,1 )2 ≤ (1 − λ)a21 + λb22 − λ(1 − λ)d22 .
Therefore, for > 0,

2 2 1 2
d12 ≤ dλ0,2 + dλ1,1 ≤ (1 + ) dλ2,0 + (1 + ) dλ1,1

1
≤ (1 + )(1 − λ)b1 + (1 + )λa2 + (1 + )(1 − λ)a21
2 2

1 2 1
+ (1 + )λb2 − (2 + + )λ(1 − λ)d2 . 2

λ

Set = 1−λλ so that the coefficent of d2 becomes 2 + λ + 1−λ λ(1 − λ) = 2λ(1 −
2 1−λ
λ 2
λ) + (1 − λ)2 + λ2 = 1. This yeilds d12 + d22 ≤ 1−λ λ b1 + a2 + a1 + 1−λ b2 . Set λ = b1 +b2 ⇒
2 2 2 b1
d12 + d22 ≤ a21 + a22 + 2b1 b2 = a21 + a22 + b21 + b22 − (b1 − b2 )2 .
Corollary 7.2. Let (M, g) be as in Theorem 7.2, and γ1 , γ2 : [0, 1] → M be geodesics.

Then ∀0 ≤ t ≤ 1, 0 ≤ s ≤ 1
d2 (γ1 (0), γ2 (t)) + d2 (γ1 (1), γ2 (1 − t)) ≤ d2 (γ1 (0), γ2 (0))

+ d2 (γ1 (1), γ2 (1)) + 2t2 d2 (γ2 (0), γ2 (1)) + t d2 (γ1 (0), γ1 (1)) − d2 (γ2 (0), γ2 (1))
− ts (d(γ1 (0), γ1 (1)) − d(γ2 (0), γ2 (1)))2
− t(1 − s) (d(γ1 (0), γ2 (0)) − d(γ1 (1), γ2 (1)))2 .
Proof. Notice that Theorem 7.2 is the case t = 1, s = 0. By symmetry, we also have
the above inequation holds for t = 1, s = 1. i.e.
d12 + d22 ≤ a21 + a21 + b21 + b22 − (a1 − a2 )2 .
Taking convex combinations yields the inequlity for t = 1, 0 ≤ s ≤ 1,
d12 + d22 ≤ a21 + a21 + b21 + b22 − s(a1 − a2 )2 − (1 − s)(b1 − b2 )2 .
Therefore, for 0 ≤ t ≤ 1, Theorem 7.1 implies
d2 (γ1 (0), γ2 (t)) + d2 (γ(1), γ2 (1 − t)) ≤ (1 − t)b21 + td22 − t(1 − t)a22 + td12 + (1 − t)b22 − t(1 − t)a22

≤ t a21 + a22 + b21 + b22 − s(a1 − a2 )2 − (1 − s)(b1 − b2 )2 + (1 − t) b21 + b22 − 2t(1 − t)a22

= b21 + b22 + 2t2 a22 + t a21 − a22 − ts(a1 − a2 )2 − t(1 − s)(b1 − b2 )2 .

Exercise: Let (M, g) be as above, and γ1 , γ2 : [0, 1] → M be geodesics. Then we
have ∀0 ≤ t ≤ 1, 0 ≤ s ≤ 1.
d2 (γ1 (t), γ2 (t)) ≤ (1 − t)d2 (γ1 (0), γ2 (0)) + td2 (γ1 (1), γ2 (1))
n o
− t(1 − t) s[d(γ1 (0), γ1 (1)) − d(γ2 (0), γ2 (1))]2 + (1 − s)[d(γ1 (0), γ2 (0)) − d(γ1 (1), γ2 (1))]2 .
Hint: Using the above corollary.

Remark: The above exercise tells particular that d2 : M × M → R where M is
a compact simply-connected Riemannian manifold with sec≤ 0, is a convex function.
This is because a geodesic γ on M × M is given as (γ1 , γ2 ) where γ1 , γ2 are geodesics
in M. Exercise tells that d2 ◦ γ = d2 (γ1 (t), γ2 (t)) is a convex function. ([JJ,Corollary
4.8.2]).
7.2. BOCHNER TECHNIQUE AND BAKRY-ÉMERY Γ-CALCULUS 199
7.2 Bochner technique and Bakry-Émery Γ-calculus

7.2.1 A computation trick
When we verify tensor equalities or tensor inequalities, we can pick a proper local
coordinate or a proper local frame at a point, and check the equality or inequality at
two points.
1. normal coordinate system around p ∈ M x : U → Rn such that
x (p) = 0
 i



∂ ∂



gi j (p) = g( i , j )(p) = δi j


∂x ∂x





Γi (p) = 0


jk
2. (local) coordinate frame: { ∂x∂ i }
3. (local) orthonormal frame: {e1 } with hei , e j i = δi j .
4. local normal frame at x: {Ei } with ∇Ei E j (x) = 0, ∀i, j.
Exercise: Pick an orthonormal basis {e1 , . . . , en } for the vector space T x M.
Choose any p ∈ M near to x, let γ be the geodesic from x to p. Let Ei (p) be the vector
in T p M which is obtained by transport ei parallelly along γ. The local frame obtained
in this way is what we want.
Let us discuss an exomple.
Lemma 7.1. Choosing any local frame {Vi }ni=1 on M and its dual coframe {ωi }ni=1 , then
we have X
d= ωi ∧ ∇Vi . (7.2.1)
i
Recall d : A p (M) → A p+1 (M) is the exterior derivative where A p (M) stands for the
vector space of smooth p-forms on M.
Proof. Let us denote by d = i ωi ∧ ∇Vi . Notice this does not depend on the choice
P
of different frames. Indeed, for the frame Xk = cik Vi and its dual ηk = dik ωi , where
i k
= δij . Hence
P
k ck d j
X X
η k ∧ ∇ Xk = dkj ω j ∧ ∇cik Vi
k k, j,i
X X X
= dkj cik ω j ∧ ∇ Vi = δij ω j ∧ ∇Vi = ωi ∧ ∇Vi .
k, j,i j,i i
As both sides of (7.2.1) are independent of the choice of frames, we can prove it
pointwise, and choose the normal coordinate {xi } around a fixed point p ∈ M, and
consider the local coordinate frame { ∂x∂ i } and its dual {dxi }.
First, we observe
∂ !
j ∂ ∂
! !
∇ ∂i dxi = ∇ ∂ dx ( ) − dx j
∇ ∂
∂x ∂xk ∂xi ∂xk ∂xi ∂xk
∂
!
= − dx j Γlik l = −Γlik δkj = −Γikj .
∂x
This means that ∇ ∂i dx j = −Γikj dxk . Hence at p, we have

∂x

∇ ∂i dx j (p) = 0. (7.2.2)
∂x
By the linear property of d and RHS of (7.2.1), we only need to verify (7.2.1) when
applying to any q-form η = f dx1 ∧ · · · ∧ dxq . Then
 
X i X
 η =

 dx ∧ ∇ ∂ dxi ∧ ∇ ∂ f dx1 ∧ · · · ∧ dxq
∂xi ∂xi
i i
X ∂f X ∂f
= dx ∧ i dx1 ∧ · · · ∧ dxq =
i
dxi ∧ dx1 ∧ · · · ∧ dxq
i
∂x i
∂x i
=dη.
7.2.2 The Hodge Laplacian

On a Riemannian manifold (M, g), g induces an inner product on T x M for each x ∈ M:
Let us denote g(X, Y) = hX, Yi, ∀X, Y ∈ T x M. Choose a local orthonomal frames {Ei }ni=1
on M, i.e. hEi , E j i = δi j . Let {ωi }ni=1 be the dual coframe of {Ei }ni=1 , i.e. ωi (X j ) = δij .
We stipulate that these 1-forms are orthonormal pairwise, i.e. we define hωi , ω j i = δi j .
Extend it by linearly: ∀φ, ψ ∈ A1 (M), suppose φ = i φi ωi , ψ = j ψ j ω j . Then
P P
hφ, ψi = i φi ψ. We can check that the above definition is independent of the choice of
P
frames. In this way, each cotangent space T x∗ M becomes an inner product space.
We can continue to assign a natural inner proudct on the space ∧ p T x∗ M: we stipulate
that the p-forms are {ωi1 ∧· · ·∧ωi p : i1 < · · · < i p } are orthonormal. For ∀φ, ψ ∈ A p (M),
suppose
X
φ= φi1 ...i p ωi1 ∧ · · · ∧ ωi p ,
i1 <···<i p
X
ψ= ψi1 ...i p ωi1 ∧ · · · ∧ ωi p .
i1 <···<i p
Then define
X
hφ, ψi(x) = φi1 ...i p ψi1 ...i p (x).
i1 <···<i p
We can check that this definition is independent of the choice of frame.

Now let M be an n-dimensional orientable manifold and we suppose {ωi }ni=1 is a
locally orthonormal coframe such that
ω1 ∧ · · · ∧ ωn
is the volume form Ω on M. Recall Ω determines an orientaion on M. Then we

define the Hodge star operator ? : A p (M) → An−p (M) by pairing φ ∧ ?ψ = hφ, ψiΩ,
∀φ, ψ ∈ A p (M).
For that purpose, we obeserve that we have to define

? ω1 ∧ · · · ∧ ω p = ω p+1 ∧ · · · ∧ ωn . (7.2.3)
If 1 ≤ i1 < · · · < i p ≤ n, 1 ≤ i p+1 < · · · < in ≤ n and {i p+1 , . . . , in } = {1, . . . , n} \

{i1 , . . . , i p }, then {ωi1 , . . . , ωi p , ωi p+1 , . . . , ωin−1 , i1 ...in ωin } is also an orthonormal coframe
determining the same orientaion, where i1 ,...,in is the sign of the permutaion
(1, . . . , n) → (i1 , . . . , ın )
ω ∧ · · · ∧ ωi p ∧ ωi p+1 ∧ · · · ∧ ωin−1 ∧ i1 ...in ωin 

 i1 
 
=ω1 ∧ · · · ∧n = Ω
Hence, by the definition (7.2.3), we have

? ωi1 ∧ · · · ∧ ωi p = i1 ,...,in ωi p+1 ∧ · · · ∧ ωin .
Extendind ? as an A0 (M)-linear operator, for f = i1 <···<i p fi1 ...i p ωi1 ∧ · · · ∧ ωi p we

P
have
X
?f = fi1 ...i p ? ωi1 ∧ · · · ∧ ωi p
i1 <···<i p
X
= i1 ...in fi1 ...i p ωi p+1 ∧ · · · ∧ ωin .
i1 <···<i p ,i p+1 <···<in
We can check that this definition is independent of the choice of an orthonormal

frame. Moreover we have for
X
φ= φi1 ...i p ωi1 ∧ · · · ∧ ωi p ,
i1 <···<i p
X
ψ= ψi1 ...i p ωi1 ∧ · · · ∧ ωi p .
i1 <···<i p
that
   
 X   X 
φ ∧ ?ψ =   φi1 ...i p ω 1 ∧ · · · ∧ ω p  ∧ 
i i   j1 ... jn ψ j1 ... j p ω p+1 ∧ · · · ∧ ω n 
j j
i1 <···<i p j1 <···< j p , j p+1 <···< jn

X
= i1 ...in φi1 ...i p ψi1 ...i p ωi1 ∧ · · · ∧ ωi p ∧ ω j p+1 ∧ · · · ∧ ω jn
i1 <···<i p ,i p+1 <···<in
 
 X 
=  φi1 ...i p ψi1 ...i p  ω1 ∧ · · · ∧ ωn = hφ, ψiΩ.
i1 <···<i p
Proposition 7.1. We have

?Ω = 1, ?1 = Ω,
? ? φ = (−1) p(n−p) , ∀φ ∈ A p (M),
hφ, ψi = h?φ, ?ψi, ∀φ, ψ ∈ A p (M).
Proof. We only show that the last:

 
 X 
? ? φ = ?   i1 ...in φi1 ...i p ω ∧ · · · ∧ ω 
i p+1 i n
i1 <···<i p ,i p+1 <···<in

X
= i1 ...in i p+1 ...in i1 ...i p φi1 ...i p ωi1 ∧ · · · ∧ ωi p
i1 <···<i p ,i p+1 <···<in
X
= φi1 ...i p (−1) p(n−p) ωi1 ∧ · · · ∧ ωi p
i1 <···<i p
= (−1) p(n−p) φ.
Then h?φ, ?ψiΩ = ?φ∧?(?ψ) = (−1) p(n−p) ?φ∧ψ = φ∧?ψ = hψ, φiΩ = hφ, ψiΩ.
Hence h?φ, ?ψi = hφ, ψi.

Definition 7.1 (Hodge Laplacian). We define the operator
δ : A p (M) → A p−1 (M)
by δ = (−1)np+n+1 ? d?. Then Hodge Laplacian ∆ is defined as ∆ = δd + dδ : A p (M) →

A p (M).
Let us explain why we define δ like that, especially with such a complicated sign.
Let M be an n-dimensional, closed orientable Riemannian manifold. We can introduce
an inner product in the space of the whole exteror Algebra A∗ (M) = ⊕np=0 A p (M) as
below:
∀φ, ψ ∈ A p (M), set
Z Z
(φ, ψ) = hφ, ψiΩ = φ ∧ ?ψ
M M
p q
∀φ ∈ A (M) ∀ψ ∈ A (M) with p , q set
(φ, ψ) = 0.
Excercise: Check this definition gives an inner product in A∗ .
Proposition 7.1 implies (?φ, ?ψ) = (ψ, φ). That is, ? is an isometry tansformation
beween A∗ and itself. The definition of δ is carefully given to ensure the following
property:
Proposition 7.2. ∀α ∈ A p−1 (M), β ∈ A p (M), we have (dα, β) = (α, δβ).
Proof.
Z
(dα, β) = dα ∧ ?β
ZM
= d(α ∧ ?β) − (−1) p−1 α ∧ d ? β
M
Z
= −(−1) p−1 (−1)(n−p+1)(p−1) α ∧ ? ? (d ? β)
M
Z
= (−1) (n−p+2)(p−1)+1
α ∧ ?(?d ? β)
M
Z
= (−1)np+n+1 α ∧ ?(?d ? β).
M
Hence (dα, β) = α ∧ ?(δβ) = (α, δβ).

R
M

Now the definition of δ is justified. Notice that δ2 = 0 following from d2 = 0.
Hence ∆ = (δ + d)(δ + d).
Proposition 7.3. The Hodge Laplacian ∆ is a self-adjoint operator.
Proof. ∀α, β ∈ A∗ , we have
(∆α, β) = ((δ + d)(δ + d)α, β)

= (d(δ + d)α, β) + (δ(δ + d)α, β)
= ((δ + d)α, δβ) + ((δ + d)α, dβ)
= ((δ + d)α, (δ + d)β)
= (α, (δ + d)(δ + d)β) = (α, ∆β)
.

As a direct corollary, ∆ is a positive operator on A∗ (M). That is, ∀ f ∈ A∗ , (∆ f, f ) =

((δ + d) f, (δ + d) f ) = (d f, d f ) + (δ f, δ f ) ≥ 0.
Suppose λ is any eigenvalue of ∆, i.e., ∆ f = λ f for some nontrival f , then λ( f, f ) =
(∆ f, f ) ≥ 0. ⇒ λ ≥ 0. Furthermore, f ∈ A∗ is harmonic (∆ f = 0) iff d f = δ f = 0.
Next, we aim at establishing an expression for δ similar to lemma 7.1 for d. Recall
for any vector field X on M, the inerior product
i(X) : A p (M) → A p−1 (M)
by (i(X)φ)(Y1 , . . . , Y p−1 ) = φ(X, Y1 , . . . , Y p−1 ), ∀φ ∈ A p (M), and any vector fields

Y1 , . . . , Y p−1 .
Proposition 7.4. For any φ ∈ A p (M), ψ ∈ Aq (M), we have
1. i(X)(φ ∧ ψ) = (i(X)φ) ∧ ψ + (−1) p φ ∧ (i(X)ψ),
2. i(X)( f φ) = f (i(X)φ),
3. i(X) ◦ i(X) = 0.
Proof. Direct proof. We check 3. ∀φ ∈ A p (M), we have
((i(X) ◦ i(X) = 0)φ)(Y1 , . . . , Y p−2 ) = φ(X, X, Y1 , . . . , Y p−2 ) = 0.
Lemma 7.2. Choosing any local orthonomal frame {Ei }ni=1 on M, we have
n
X
δ=− i(E j )∇E j . (7.2.4)
j=1
Proof. Denote δ = − nj=1 i(E j )∇E j . We can check δ is independent of the choice of
P
local orthonormal frames. So we only need to prove (7.2.4) at a fixed point x ∈ M.
We pick a local normal frame {Ei } at x. Let {ωi } be the dual coframe. Since ∇Ei E j =
k Γi j E k , we have
P k

∇Ei ω j (Ek ) = ∇Ei ω j (Ek ) − ω j ∇Ei E j
= −ω j (Γlik El ) = −Γikj .
⇒ ∇Ei ω j = −Γikj ωk . Therefore at x ∈ M, we have ∇Ei ω j (x) = 0, 1 ≤ i, j ≤ n. By

linearity, we only need to verify δη = δη for η = f ω1 ∧ · · · ∧ ω p . We compute
p
X
δη = − i(E j )∇E j ( f ω1 ∧ · · · ∧ ω p )
j=1
p
X
=− i(E j ) E j ( f ) ω1 ∧ · · · ∧ ω p
j=1
p 
X p
X

=− E j ( f ) (−1) ω ∧ · · · ∧ ω ∧ i(E j )ω ∧ · · · ∧ ω 
 k−1 1 k−1 k p


j=1 k=1
p
X
=− E j ( f )(−1) j−1 ω1 ∧ · · · ∧ b
ω j ∧ · · · ∧ ωp
j=1
p
X
= ω j ∧ · · · ∧ ωp.
(−1) j E j ( f )ω1 ∧ · · · ∧ b
j=1
On the other hand, note that for j = 1, . . . , p.

? ω j ∧ ω p+1 ∧ · · · ∧ ωn = (−1)(n−p)(p−1) × (−1) j−1 ω1 ∧ · · · ∧ b
ω j ∧ · · · ∧ ωp
= (−1)np+n+1+ j ω1 ∧ · · · ∧ b
ω j ∧ · · · ∧ ωp.
Therefore
(δη)(p) = (−1)np+n+n+ ? d ? ( f ω1 ∧ · · · ∧ ω p )
= (−1)np+n+1 ? d( f ω p+1 ∧ · · · ∧ ωn )
 n 
X 
= (−1) np+n+1
?  ω ∧ ∇E j ( f ω ∧ · · · ∧ ω )
 j p+1 n
j=1
 n 
X 
= (−1) np+n+1
?   E j ( f )ω ∧ ω ∧ · · · ∧ ω 
j p+1 n
j=1
n
X
= (−1)np+n+1 E j ( f ) ? (ω j ∧ ω p+1 ∧ · · · ∧ ωn )
j=1
n
X
= ω j ∧ · · · ∧ ωp
(−1) j E j ( f )ω1 ∧ · · · ∧ b
j1
= (δη)(p).

Observation: for f ∈ C ∞ (M), we have

 
X X i
∆ f = −δd f = − i(E j )∇E j  ω ∧ ∇Ei f 


j i
X
=− i(E j )∇E j Ei ( f )ωi
j
X X
=− E j (Ei ( f )) ωi (E j ) = − Ei ((E j )( f ))
i, j i
= trHess f.
So Hodge Lapalacian is the negative of the Laplace-Beltrani operator we defined

before.
7.2.3 Weitzenböck formula

For ω ∈ A p (M), which can be considered as a (0, p) tensor, recall ∇2 ω is a (0, p + 2)-
tensor. Let {ei } be a local orthonormal frame. We define
X
tr(∇2 ω)(. . . ) = ∇2 ω(. . . , ei , e j ).
i
One can check this definition is independent of the choice of an orthonormal frame,
and
X
tr(∇2 ω)(. . . ) = ∇ei ∇ei ω − ∇∇ei ei ω .
i
Theorem 7.3 (Weitzenböck formula). For any ω ∈ A p (M), let {ei } be a local orthonor-
mal frame and {ωi } its dual, then
∆ω = −tr(∇2 ω) − ωi ∧ i(e j )R(ei , e j )ω. (7.2.5)
Proof. We can check that the RHS is independent of the choice of othonormal frame.
So we will prove the W-formula at a point x ∈ M, and pick a local normal frame {Ei }.
We again use {ωi } for its dual.
Recall the covariant derivative is commutable with the contraction, i.e.

∇Ei C(E j ⊗ ∇E j ω) = C∇Ei (E j ⊗ ∇E j ω)

= C E j ⊗ ∇Ei ∇E j ω .

Hence ∇Ei i(E j )∇E j ω = i(E j )∇Ei ∇E j ω. Recall ∇Ei E j = 0, ∀i, j ⇒ ∇E j ω j = 0,
∀i, j. Therefore, we compute

X X
∆ω = dδω + δdω = ωi ∧ ∇Ei (δω) − i(E j )∇E j (dω)
i j
   
X X  X X i
=− ω ∧ ∇Ei  i(Ei )∇E j ω − i(E j )∇E j  ω ∧ ∇Ei ω
i


 
 
i j j i
X X
=− ωi ∧ ∇Ei i(Ei )∇E j ω − i(E j )∇E j ωi ∧ ∇Ei ω
ij j,i
X X
=− ωi ∧ i(Ei )∇Ei ∇E j ω − i(E j ) ωi ∧ ∇E j ∇Ei ω
ij i, j
X X X
=− ωi ∧ i(Ei )∇Ei ∇E j ω − δij ∇E j ∇Ei ω + ωi ∧ i(E j )∇E j ∇Ei ω
ij i, j i, j
X X
=− ∇Ei ∇Ei ω − ωi ∧ i(E j ) ∇Ei ∇E j ω − ∇E j ∇Ei ω
i i, j
X
= −tr(∇ ω) −
2
ω ∧ i(E j )R(Ei , E j )ω.
i
i, j
Corollary 7.3 (Bochner). For any ω ∈ A p (M), we have

1
− ∆|ω|2 = −h∆ω, ωi + |∇ω|2 + F(ω) (7.2.6)
2
where F(ω) = −hωi ∧ i(E j )R(ei , e j )ω, ωi
Remark: Here we are using Hodge Laplacian.
Proof. We complete in local normal frame {Ei }, we have
−h∆ω, ωi + F(ω) = h−∆ω − ωi ∧ i(E j )R(Ei , E j )ω, ωi

= htr(∇2 ω), ωi
X
= h ∇Ei ∇Ei ω, ωi
i
X
= ∇Ei h∇Ei ω, ωi − h∇Ei ω, ∇Ei ωi

i
1X
= ∇Ei ∇Ei hω, ωi − |∇ω|2
2 i
1
= − ∆|ω|2 − |∇ω|2 .
2

In particular, when ω ∈ A1 (M), the curvature term becomes simpler. Let ]ω =

hω, ωi iEi , we have
F(ω) = −hωi ∧ i(E j )R(Ei , E j )ω, ωi

= −i(E j )R(Ei , E j )ωhωi , ωi

= − R(Ei , E j )ω (E j )hωi , ωi.
Claim:
(R(X, Y)(ω)) (Z) = −ω ( R(X, Y)Z )

= ∇X ∇Y ω − ∇Y ∇X ω − ∇[X,Y] ω (Z),

(∇X ∇Y ω) (Z) = (∇X (∇Y ω)) (Z) = X ((∇Y ω)(Z)) − (∇Y ω)(∇X Z)
= X{Y(ω(Z)) − ω(∇Y Z)} − Y(ω(∇X Z)) + ω(∇Y ∇X Z)
= X(Y(ω(Z))) − X(ω(∇Y Z)) − Y(ω(∇X Z)) + ω(∇Y ∇X Z),
∇[X,Y] ω (Z) = [X, Y](ω(Z)) − ω(∇[X,Y] (Z)).

Hence (R(X, Y)(ω)) (Z) = ω(∇Y ∇X Z) − ω(∇X ∇Y Z) + ω(∇[X,Y] (Z)) = −ω(R(X, Y)Z).
Now we continue our computation:

F(ω) = − R(Ei , E j )ω (E j )hωi , ωi

= ω R(Ei , E j )E j hωi , ωi
= h]ω, R(Ei , E j )ωE j ihωi , ωi
= h]ω, R(hωi , ωiEi , E j )E j i
= h]ω, R(]ω, E j )E j i = hR(]ω, E j )E j , ]ωi
X X
= R(]ω, E j , ]ω, E j ) = R(E j , ]ω, E j , ]ω)
j j
= trR(, ]ω, , ]ω) = Ric(]ω, ]ω).
Corollary 7.4. For any ω ∈ A1 (M), we have

1
− ∆|ω|2 = −h∆ω, ωi + |∇ω|2 + Ric(]ω, ]ω). (7.2.7)
2
Theorem 7.4 (Bochner). Let (M, g) be a closed oriented Riemannian manifold.
1. If Ric ≥ 0, then any harmonic 1-form ω is parallel, i.e. ∇ω = 0.
2. If Ric ≥ 0 on M and Ric > 0 at one point, then there is no non-trival harmonic
1-form.
Proof. Recall M −∆|ω|dvol M = 0. Hence we have
R
Z
0=− h∆ω, ωi + |∇ω|2 + Ric(]ω, ]ω)dvol M
M
Z
= |∇ω|2 + Ric(]ω, ]ω)dvol M ≥ 0.
M
⇒ ∇ω = 0, i.e. ω is parallel. If Ric > 0 at some point, we must have ]ω = 0,

i.e.ω = 0.

Corollary 7.5. For any f ∈ C ∞ (M), we have
1
∆LB |grad f |2 = |Hess f |2 + hgrad(∆LB f ), grad f i + Ric(grad f, grad f ). (7.2.8)
2
Proof. We see d f ∈ A1 (M), and |d f |2 , ](d f ) = grad f .
−h∆d f, d f i = −h(dδ + δd)d f, d f i = −hdδd f, d f i

= −hd(δd f ), d f i = −hd(∆ f ), d f i
= −hgrad(∆ f ), grad f i = hgrad(∆LB f ), d f i.
X
|∇d f |2 = h∇Ei d f, ∇Ei d f i
i
X
= h∇Ei grad f, ∇Ei grad f i
i
X *X X +
= h∇Ei grad f, E j iE j , h∇Ei grad f, Ek iEk
i j k
X X
= h∇Ei grad f, E j i =
2
Hess f (Ei , E j )2
ij ij
= |Hess f | . 2
Let (M, g) be a closed Riemannian manifold. We say λ ∈ R is an eigenvalue of ∆LB

if ∃ a smooth function u , 0 such that ∆LB u + λu = 0. It is konwm that the eigenvalues
can be listed as 0 = λ0 < λ1 ≤ λ2 ≤ · · · ≤ λk ≤ · · · % ∞.
Theorem 7.5 (Lichnerowicz). Let (M n , g) be a closed Riemannian manifold with Ric ≥

k > 0. Then we have λ1 ≥ n−1
n
k.
Proof. Integrate the Bochner formula in corallary 7.5, we have
Z
0= |Hess f |2 + hgrad(∆LB f ), grad f i + Ric(grad f, grad f )
ZM Z Z
= |Hess f |2 − λ1 hgrad f, grad f i + Ric(grad f, grad f )
ZM ZM Z M
≥ |Hess f |2 − λ1 |grad f |2 + |grad f |2 .

M M M
⇒ λ1 ≥ k.
We can make more use of |Hess f |2 term.

X X
|Hess f |2 = Hess f (Ei , E j )2 ≥ Hess f (Ei , Ei )2
ij i
 2
1 X 1 1
≥  Hess f (Ei , Ei ) = (∆LB f ) = λ21 f 2

n i n n
λ1
Z Z Z
1
⇒ |Hess f |2 ≥ λ1 λ1 f 2 = h f, −∆LB f i
M n M n M
λ1
Z
= hgrad f, grad f i
n M
λ Z
1
⇒0≥ − λ1 + k |grad f |2
n M
nk
⇒ λ1 ≥ .
n−1

Like for the Bonnet-Myers Theorem, we have the following RIGIDITY result due
to Obata.
Theorem 7.6 (Obata). Let (M n , g) be a closed Riemannian manifold

with Ric ≥ (n −
1)k, k > 0. Then λ1 = nk iff (M n , g) is isometric to the space S √1
k
.
Proof. W.l.o.g., we can suppose k = 1. If λ1 = n, then the proof of 7.5 implies
Ric(grad f, grad f ) = (n − 1)|grad f |2 .
Since ∆LB u2 = 2|grad f |2 + 2u∆LB u.
1 1
⇒ ∆LB |∇ f |2 + f 2 = ∆LB |∇ f |2 + |grad f |2 + f ∆LB f
2 2
λ1
≥ h f, −∆LB f i − n|grad f |2 + (n − 1)|grad f |2
n
+ |grad f |2 + f ∆LB f
=0.

∆ |∇ f |2 + f 2 = 0. Hence ∆LB |∇ f |2 + f 2 = 0. That is,
R 1
Recall M 2 LB
|grad f |2 + f 2 ≡ const.
Normalize f so that max M f 2 = 1. Since at the maximum/minimum points of f , we

have grad f = 0. Therefore we have |grad f |2 + f 2 = 1 and max M f = − min M f = 1.
Let p, q ∈ M be points s.t. f (p) = −1, f (q) = 1.
Let γ be a normal minimizing geodesic from p to q. Note that

d
f ◦ γ(t) |grad f (γ(t))|2
dt
≤ p = 1.
1 − ( f ◦ γ(t))2 1 − ( f ◦ γ(t))2
p
Integrating over t,
Z
d(p,q) dtd f ◦ γ(t) Z 1 1
= √ dx = π
1 − f ◦ γ(t)
p
0 2 −1 1 − x2
Z d(p,q) d
dt f ◦ γ(t)

≤ ≤ d(p, q).
1 − f ◦ γ(t)2
p
0
On the other hand, Ric ≥ (n − 1) ⇒ d(p, q) ≤ π ⇒ diam = π ⇒ M is isometric to

S(1).

7.2.4 Bakry-Émery Γ-calculus:A systematic way of understanding

the Bochner formula
For any f, g ∈ C ∞ (M), define
1
Γ( f, g) = (∆LB ( f g) − f ∆LB g − g∆LB f )
2
Observe Γ( f, f ) = |grad f |2 .
Interatively, define Γ2 ( f, g) = 12 (∆(Γ( f, g)) − Γ( f, ∆LB g) − Γ(∆LB f, g)).
Observe that Γ2 ( f, f ) = 12 ∆LB |grad f |2 − hgrad f, grad(∆LB f )i. So the Bochner for-
mula in Corollary 7.5 implies
Γ2 ( f, f ) =|Hess f |2 + Ric(grad f, grad f )

1
≥ (∆LB f )2 + Ric(grad f, grad f ).
n
Moreover Ric ≥ k implies Γ2 ( f, f ) ≥ 1n (∆LB f )2 + kΓ( f, f ), ∀ f ∈ C ∞ (M).
The above property enables us to define ”Ricci curvature lower bound” for a general
operator, which can be operators on more general spaces. Here we discuss possibilites
on discrete metric spaces: a combinatorial graph G = (V, E), where
1. V is the set of vertices (points),
2. E is the set of edges,
3. metric: combinatorial distance (length of shortest path).
For example, a discrete set {p1 , . . . , pn } with the metric d(pi , p j ) = δi j can be rep-
resented by a complete graph Kn . Define the degree of a vertex p to be deg(p) =
P
q∈V,d(p,q)=1 1.
We can consider the graph Laplacian ∆ defined via ∆ f (x) = y∈V,d(y,x)=1 ( f (y) −
P
f (x)),
for f : V → R. We say λ is eigenvalue of ∆ if ∃ f , 0 s.t. ∆ f + λ f = 0. We can list

0 = λ0 < λ1 ≤ λ2 ≤ · · · ≤ λ|V|−1 .
Definition 7.2. A graph G = (V, E) is said to satisfy the curvature dimension inequality
CD(K, n) for some K ∈ R, n ∈ (0, ∞] if for all f : V → R, it holds
1
Γ2 ( f, f )(x) ≥ (∆ f )2 (x) + KΓ( f, f )(x)
n
∀x ∈ V.
Since we do not have a proper understanding about the ”dimension” of a graph,

quite oftenly we assume CD(K, ∞) conditions.
Theorem 7.7 (L.-Münch-Peyerimhoff,arXiv:1608.09998,arXiv:1705.08119). Let G =

(V, E) be a connected graph satisfying CD(K, ∞), and degmax < ∞. Then
2 degmax
diamd (G) ≤ .
K
Moreover, ”=” holds iff G is a degmax -dimensional hypercube. Under the same assump-
tion, (By standard argument, we have λ1 ≥ K) λdegmax = K iff G is a degmax -dimensional
hypercube.
Remark: (1)Hypercube:
(2) λ1 = K is not strong enough to conclude the Rigidity Theorem. counterexam-

ple:
Open question: Let G = (V, E) be a connected graph satisgying CD(0, ∞). What is
the volume growth rate? Polynomial? This is equivalent to ask for the (non-)existence
of a family of expanders in the class of Graphs satisfying CD(0, ∞).
Bibliography
[1] Jürgen Jost, Riemannian geometry and geometric analysis, Springer.
[2] Manfredo Perdigão do Carmo, Riemannian geometry, Birkhäuser

[3] Michael Spivak. A comprehensive inrtoduction to differential geometry,
volume I − V, third edition, Publish or Perish.
[4] Peter Petersen, Riemannian geometry, GTM171, Springer
215

Hu Zhang Zhao

Uploaded by

Copyright:

Available Formats

Hu Zhang Zhao

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Hu Zhang Zhao

Uploaded by

Copyright:

Available Formats

Lectures on Riemannian Geometry

typed by Yirong Hu, Shiyu Zhang, Qizhi Zhao.

3 Connections, Parallelism, and covariant Derivatives. 49

4.4 Ricii Curvature and Scalar curvature . . . . . . . . . . . . . . . . . 89

5 Space forms and Jacobi fields 113

6 Comparison Theorem 157

7 Candidates for Synthetic Curvature Conditions 195

the geometry of the underlying spaces from the Riemannian metric.

This inner product hX, Yi p corresponds to the first fundenmental forms of S at p.

we have hX p , Y p i p = X i (p)Y j (p) ∂x∂ i , ∂x∂ j

Definition 1.2 (Riemannian Metric). A Riemannian metric g on a C ∞ -manifold M is a

g(X, Y) = g(Y, X), g(X, X) = 0 and g p (X, X) = 0 ⇔ X(p) = 0

for any smooth tangent vector fields X,Y.

Definition 1.3. A Riemannian manifold is a differentiable(we will always assume C ∞ )

Remark 1.2. a couple (M,g).

Another way to see it: R is covered by a single coordinate (x1 , · · · , xn ).

Matrix : (gi j ) = (δi j ) = In×n .

( f ∗ gN ) p (X p , Y p ) = (gN ) f (p) (d f p (X p ), d f p (Y p )).

One can verify that f ∗ gN is a Riemannian metric on M. (( f ∗ gN ) p (X p , X p ) ⇔ d f p (X p ) =

gS 1 = (dx ⊗ dx + dy ⊗ dy)|S 1 = dθ ⊗ dθ. Then we have dx = − sin θdθ, dy = cos θdθ

Then the induced metric on S 2 is gS 2 = 1

(3)Product metric: Let (M, g M ), (N, gN ) be two Riemannian manifolds. Let M × N

1.3 When are two Riemannian manifolds “equivalent”?

1.4 Existencce of Riemannian Metrics

It is direct to verify that this construction defines a Riemannian metirc on M .(In

φα α = 1 =⇒ ∃β s.t.φβ (p) > 0. Hence g p ≥

1.5 The Metric Structure

Proof. Let γ1 : [c, d] → M (c < d) is another regular parametrization of the same

γ1 (t1 (t)) = γ(t).

Since both parametrization are regular, we haveγ˙1 (t1 ) , 0, γ0(t)

Exercise 1.1. Let ϕ : (M, g M ) → (N, gN ) be an isometry,γ be a smooth curve in M.

Arclength parametrization. Now we look for a ”standard” parametrization for a

γ1 (t) = γ(t(s)), s ∈ [0, Length(γ)]

Proposition 1.1. hγ˙1 (s), γ˙1 (s) ≡ 1

where C p,q := {γ : [a, b] → M : γ piecewise smooth curve with γ(a) = p, γ(b) = q}

Theorem 1.2. (M, d) is a metric space.

Proof. It remains to show p , q =⇒ d(p, q) > 0

Obeserve that there exists a positive constant  s.t.

hγ̇, γ̇ih ≥ hγ̇, γ̇ig0 on Bδ (0)

(Exercise: Let K be an open set of Rn , [gi j (x)] be n2 continuous functions on K

for any x ∈ K and any (v1 , · · · , vn ) = v ∈ Rn .)

Moreover, we have the following property.

Without loss of generality, we assume φ(qi ) ∈ B 1i (0). Denote by h := (φ−1 )∗ g.

1.6 Riemannian Measure, Volume

Now we hope to develop a natural notion of intergration on a Riemannian manifold.

Volume of “small” compact domain.

Proof. Suppose we have another coordinate neighborhood (V, y1 , · · · , yn ) containing

Denote the jacobian matrix of the map y ◦ x−1 by

We obtain [gixj (p)] = J T (x(p))[gykl (p)]J(x(p)).

Volume of “larger” conmpact domain.

From the above discussion, we know this is well-defined. Moreover, since φα ≥ 0,

for any f ∈ C00 (M).

Notice that when we change the coordinate, we have

dy1 ∧ · · · ∧ dyn = det(J(x(p)))dx1 ∧ · · · ∧ dxn .

(Recall: On an orientable manifold M, let {e1 , · · · , en } be an orthonormal frame fields,

This verifies L2 (M) to be a Hilbert Space.

1.7 Divergence Theorem

The divergence divX : M → R is a C ∞ function on M such that

Proof. Let Y1 , · · · , Yn−1 be any (n-1) smooth vector fields. We compute

Obeserve that there exists a positive constant s.t.

hγ̇, γ̇ih ≥ hγ̇, γ̇ig0 on Bδ (0)