Math 6455 Nov 1, 20061

Differential Geometry I
Fall 2006, Georgia Tech

Lecture Notes 14
Connections
Suppose that we have a vector field X on a Riemannian manifold M . How can we
measure how much X is changing at a point p ∈ M in the direction Yp ∈ Tp M ?
The main problem here is that there exists no canonical way to compare a vector
in some tangent space of a manifold to a vector in another tangent space. Hence
we need to impose a new kind of structure on a manifold. To gain some insight, we
first study the case where M = Rn .

0.1 Differentiation of vector fields in Rn

Since each tangent space Tp Rn is canonically isomorphic to Rn , any vector field on
Rn may be identified as a mapping X : Rn → Rn . Then for any Yp ∈ Tp Rn we
define the covariant derivative of X with respect to Yp as

∇Yp X := Yp (X 1 ), . . . , Yp (X n ) .

Recall that Yp (X i ) is the directional derivative of X i at p in the direction of Y , i.e.,

if γ : (−, ) → M is any smooth curve with γ(0) = p and γ 0 (0) = Y , then

Yp (X i ) = (X i ◦ γ)0 (0) = grad X i (p), Y .



The last equality is an easy consequence of the chain rule. Now suppose that
Y
Y : Rn → Rn is a vector field on Rn , p 7−→ Yp , then we may define a new vec-
tor field on Rn by
(∇Y X)p := ∇Yp X.
∇
Then the operation (X, Y ) 7−→ ∇X Y may be thought of as a mapping ∇ : X (Rn ) ×
X (Rn ) → X (Rn ), where X denotes the space of vector fields on Rn .
Next note that if X ∈ X (Rn ) is any vector field and f : M → R is a function,
then we may define a new vector field f X ∈ (Rn ) by setting (f X)p := f (p)Xp (do
not confuse f X, which is a vector field, with Xf which is a function defined by
Xf (p) := Xp (f )). Now we observe that the covariant differentiation of vector fields
on Rn satisfies the following properties:
1. ∇Y (X1 + X2 ) = ∇Y X1 + ∇Y X2
1
Last revised: March 9, 2012

1
 2. ∇Y (f X) = (Y f )∇Y X + f ∇Y X

3. ∇Y1 +Y2 X = ∇Y1 X + ∇Y2 X

4. ∇f Y X = f ∇Y X

It is an easy exercise to check the above properties. Another good exercise to

write down the pointwise versions of the above expressions. For instance note that
item (2) implies that

∇Yp (f X) = (Yp f )∇Yp X + f (p)∇Yp X,

for all p ∈ M .

0.2 Definition of connection and Christoffel symbols

Motivated by the Euclidean case, we define a connection ∇ on a manifold M as any
mapping
∇ : X (M ) × X (M ) → X (M )
which satisfies the four properties mentioned above. We say that ∇ is smooth if
whenever X and Y are smooth vector fields on M , then ∇Y X is a smooth vector
field as well. Note that any manifold admits the trivial connection ∇ ≡ 0. In the
next sections we study some nontrivial examples.
Here we describe how to express a connection in local charts. Let Ei be a basis
for the tangent space of M in a neighborhood of a point p. For instance, choose a
local chart (U, φ) centered at p and set Ei (q) := dφ−1
φ(q) (ei ) for all q ∈ U . Then if X
and Y are any vector fields on M , we may write X = i X i Ei , and Y = i Y i Ei
P P
on U . Consequently, if ∇ is a connection on M we have
X  X
X i Ei = Y (X i )Ei + X i ∇Y Ei .


∇Y X = ∇Y
i i

Now note that since (∇Ej Ei )p ∈ Tp M , for all p ∈ U , then it is a linear combination
of the basis elements of Tp M . So we may write
X
∇Ej Ei = Γkji Ek
k

for some functions Γkji on U which are known as the Christoffel symbols. Thus
X X X 
∇Y X = Y (X i )Ei + X i Yj Γkji Ek
i j k
X X 
= Y (X k ) + Y i X j Γkij Ek
k ij

2
Conversely note that, a choice of the functions Γkij on any local neighborhood of M
defines a connection on that neighborhood by the above expression. Thus we may
define a connection on any manifold, by an arbitrary choice of Christoffel symbols
in each local chart of some atlas of M and then using a partition of unity.
Next note that for every p ∈ U we have:
X X 
(∇Y X)p = Yp (X k ) + Y i (p)X j (p)Γkij (p) Ek (p). (1)
k ij

This immediately shows that

Theorem 0.1. For any point p ∈ M , (∇Y X)p depends only on the value of X
at p and the restriction of Y to any curve γ : (−, ) → M which belongs to the
equivalence class of curves determined by Xp .
Thus if p ∈ M , Yp ∈ Tp M and X is any vector field which is defined on an open
neighborhood of p, then we may define

∇Yp X := (∇Y X)p

where Y is any extension of Yp to a vector field in a neighborhoodP ofi p. Note that

such an extension may always be found: for instance, if Yp = Yp Ei (p), where
Ei are some local
P i basis for tangent spaces in a neighborhood U of p, then we may
set Yq := Yp Ei (q) for all q ∈ U . By the previous proposition, (∇Y X)p does not
depend on the choice of the local extension Y , so ∇Yp X is well defined.

0.3 Induced connection on submanifolds

As we have already seen M admits a standard connection when M = Rn . To give
other examples of manifolds with a distinguished connection, we use the following
observation.
Lemma 0.2. Let M be a manifold, M be an embedded submanifold of M , and X be
a vector field of M . Then for every point p ∈ M there exists an open neighborhood
U of p in M and a vector filed X defined on U such that X p = Xp for all p ∈ M .
Proof. Recall that, by the rank theorem, there exists a local chart (U , φ) of M
centered at p such that φ(U ∩ M ) = Rn−k where k = dim(M ) − dim(M ). Now,
note that dφ(X) is a vector field on Rn−k and let Y be an extension of dφ(X) to
Rn (any vector field on a subspace of Rn may be extended to all of Rn ). Then set
−1
X := dφ (Y ).

Now if M is a Riemannian manifold with connection ∇, and M is any subman-

ifold of M , we may define a connection on M as follows. First note that for any
p ∈ M,
Tp M = Tp M ⊕ (Tp M )⊥ ,

3
that is any vector X ∈ Tp M may written as sum of a vector X > ∈ Tp M (which is
tangent to M and vector X ⊥ := X − X > (which is normal to M ). So for any vector
fields X and Y on M we define a new vector field on M by setting, for each p ∈ M ,

(∇Y X)p := (∇Y X)>

p

where Y and X are local extensions of X and Y to vector fields on a neighborhood

of p in M . Note (∇Y X)p is well-defined, because it is independent of the choice of
local extensions X and Y by Theorem 0.1.

0.4 Covariant derivative

We now describe how to differentiate a vector field along a curve in a manifold M
with a connection ∇. Let γ : I → M be a smooth immersion, i.e., dγt 6= 0 for all
t ∈ I, where I ⊂ R is an open interval. By a vector filed along γ we mean a mapping
X : I → T M such that X(t) ∈ Tγ(t) M for all t ∈ I. Let X (γ) denote the space of
vector fields along γ.
For any vector field X ∈ X (γ), we define another vector field Dγ X ∈ X (γ),
called the covariant derivative of X along γ, as follows. First recall that γ is locally
one-to-one by the inverse function theorem. Thus, by the previous lemma on the
existence of local extensions of vector fields on embedded submanifolds, there exists
an open neighborhood U of γ(t0 ) and a vector field X defined on U such that
X γ(t) = X(t) for all t ∈ (t0 − , t0 + ). Set

Dγ X(t0 ) := ∇γ 0 (t0 ) X.

Recall that γ 0 (t0 ) := dγt0 (1) ∈ Tγ(t0 ) M. By Theorem 0.1, Dγ X(t0 ) is well defined,
i.e., it does not depend on the choice of the local extension X. Thus we obtain a
mapping Dγ : X (γ) → X (γ). Note that if X, Y ∈ X (γ), then (X + Y )(t) := X(t) +
Y (t) ∈ X (γ). Further, if f : I → R is any function then (f X)(t) := f (t)X(t) ∈
X (γ). It is easy to check that

Dγ (X + Y ) = Dγ (X) + Dγ (Y ) and Dγ (f X) = f Dγ (X).

Proposition 0.3. If γ : I → Rn , and X ∈ X (γ), then Dγ X = X 0 . In particular,

Dγ γ 0 = γ 00 .
Proof. Let X be a vector field on an open neighborhood of γ(t0 ) such that

X(γ(t)) = X(t),

for all t ∈ (t0 − , t0 + ). Then

Dγ X(t0 ) = ∇γ 0 (t0 ) X = (X ◦ γ)0 (t0 ) = X 0 (t0 ).

4
Corollary 0.4. Let M be an immersed submanifold of Rn with the induced con-
nection ∇, and corresponding covariant derivative D. Suppose γ : I → M is an
immersed curve, and X ∈ XM (γ) is a vector field along γ in M . Then Dγ X =
(X 0 )> .

0.5 Geodesics
Note that, by the last exercise, the only curves γ : I → Rn with the property that

Dγ γ 0 ≡ 0

are given by γ(t) = at + b, which trace straight lines. With this motivation, we
define a geodesic (which is meant to be a generalization of the concept of lines) as
an immersed curve γ : I → M which satisfies the above equality for all t ∈ I. A nice
supply of examples of geodesics are provided by the following observation:

Proposition 0.5. Let M ⊂ Rn be an immersed submanifold, and γ : I → M an

immersed curve. Then γ is a geodesic of M (with respect to the induced connection
from Rn ) if and only if γ 00> ≡ 0. In particular, if γ : I → M is a geodesic, then
kγ 0 k = const.

Proof. The first claim is an immediate consequence of the last two results. The last
sentence follows from the leibnitz rule for differentiating inner products in Euclidean
space: hγ 0 , γ 0 i0 = 2hγ 00 , γ 0 i. Thus if γ 00> ≡ 0, then kγ 0 k2 = const.

As an application of the last result, we can show that the geodesics on the sphere
S2 are those curves which trace a great circle with constant speed:

Example 0.6 (Geodesics on S2 ). A C 2 immersion γ : I → S2 is a geodesic if and

only if γ has constant speed and lies on a plane which passes through the center of
the sphere, i.e., it traces a segment of a great circle.
First suppose that γ : I → S2 has constant speed, i.e. kγ 0 k = const., and that
γ traces a part of a great circe, i.e., hγ, ui = 0 for some fixed vector u ∈ S2 (which
is the vector orthogonal to the plane in which γ lies). Since hγ 0 , γ 0 i = kγ 0 k2 is
constant, it follows from the Leibnitz rule for differentiating the innerproduct that
hγ 00 , γ 0 i = 0. Furthermore, differentiating hγ, ui = 0 yields that hγ 00 , ui = 0. So, γ 00
lies in the plane of γ, and is orthogonal to γ. So, since γ traces a circle, γ 00 must be
parallel to γ. This in turn implies that γ 00 must be orthogonal to Tγ S2 , since γ is
orthogonal to Tγ S2 . So we conclude that (γ 00 )> = 0.
Conversely, suppose that (γ 00 )> = 0. Then γ 00 is parallel to γ. So if u := γ × γ 0 ,
then u0 = γ 0 × γ 0 + γ × γ 00 = 0 + 0 = 0. So u is constant. But γ is orthogonal to u, so
γ lies in the plane which passes through the origin and is orthogonal to u. Finally,
γ has constant speed by the last proposition.

5
0.6 Ordinary differential equations
In order to prove an existence and uniqueness result for geodesic in the next section
we need to develop first a basic result about differential equations:

Theorem 0.7. Let U ⊂ Rn be an open set and F : U → Rn be C 1 , then for every

x0 ∈ U , there exists an  > 0 such that for every 0 <  <  there exists a unique
curve x : (−, ) → U with x(0) = x0 and x0 (t) = F (x(t)).

Note that, from the geometric point of view the above theorem states that there
passes an integral curve through every point of a vector field. To prove this result
we need a number of preliminary results. Let I ⊂ R be an interval, (X, d) be a
compact metric space, and Γ(I, X) be the space of maps γ : I → X. For every pair
of curves γ1 , γ ∈ Γ(I, X) set


δ(γ1 , γ2 ) := sup d γ1 (t), γ2 (t) .
t∈I

It is easy to check that (Γ, δ) is a metric space. Now let C(I, X) ⊂ Γ(I, X) be the
subspace of consisting of continuous curves.

Lemma 0.8. (C, δ) is a complete metric space.

Proof. Let γi ∈ C be a Cauchy sequence. Then, for every t ∈ I, γi (t) is a Cauchy

sequence in X. So γi (t) converges to a point γ(t) ∈ X (since every compact metric
space is complete). Thus we obtain a mapping γ : I → X. We claim that γ is
continuous which would complete the proof. By the triangular inequality,

d(γ(s), γ(t)) ≤ d(γ(s), γi (s)) + d(γi (s), γi (t)) + d(γi (t), γ(t))
≤ 2δ(γ, γi ) + d(γi (s), γi (t)).

So, since γi is continuous,

lim d(γ(s), γ(t)) ≤ 2δ(γ, γi ).

t→s

All we need then is to check that limi→∞ δ(γ, γi ) = 0: Given  > 0, choose i
sufficiently large so that δ(γi , γj ) <  for all j ≥ i. Then, for all t ∈ I, d(γi (t), γj (t)) ≤
 , which in turn yields that d(γi (t), γ(t)) ≤ . So δ(γi , γ) ≤ .

Now we are ready to prove the main result of this section:

Proof of Theorem 0.7. Let B = Brn (x0 ) denote a ball of radius r centered at x0 .
Choose r > 0 so small that that B ⊂ U . For any continuous curve α ∈ C((−, ), B)
we may define another continuous curve s(α) ∈ ((−, ), Rn ) by
Z t
s(α)(t) := x0 + F (a(u))du.
0

6
We claim that if  is small enough, then s(α) ∈ C((−, ), B). To see this note that
Z t Z t

F (α(u))du ≤
ks(α)(t) − x0 k = kF (α(u))kdu ≤  sup kF k.
0 0 B

So setting  ≤ r/ supB kF k, we may then assume that

s : C((−, ), B) → C((−, ), B).

Next note that for every α, β ∈ C((−, ), B) , we have

Z t Z t

δ(s(α), s(β)) = sup F (α(u)) − F (β(u))du ≤ sup kF (α(u)) − F (β(u))kdu.
t t

0 0

Further recall that, since F is C 1 , by the mean value theorem there is a constant K
such that
kF (x) − F (y)k ≤ Kkx − yk,
√
for all x, y ∈ B (in particular recall that we may set K := n supB |Dj F i |). Thus
Z t Z t
kF (α(u)) − F (β(u))kdu ≤ K kα(u) − β(u)kdu ≤ Kδ(α, β).
0 0

So we conclude that
δ(s(α), s(β)) ≤ Kδ(α, β).
Now assume that  < 1/K (in addition to the earlier assumption that  ≤ r/ supB kF k),
then, s must have a unique fixed point since it is a contraction mapping. So for
every 0 <  <  where
 
r 1
 := min ,√
supB kF k n supB |Dj F i |

there exists a unique curve x : (−, ) → B such that x(0) = s(x)(0) = x0 , and
x0 (t) = s(x)0 (t) = F (x(t)).
It only remains to show that x : (−, ) → U is also the unique curve with
x(0) = x0 and x0 (t) = F (x(t)), i.e., we have to show that if y : (−, ) → U is any
curve with y(0) = x0 and y 0 (t) = F (y(t)), then y = x (so far we have proved this
only for y : (−, ) → B). To see this recall that  ≤ r/ supB kF k where r is the
radius of B. Thus
Z t Z t
0
ky(t) − x0 k ≤ ky (u)kdu = kF (y(u))kdu ≤  sup kF k ≤ r.
0 0 B

So the image of y lies in B, and therefore we must have y = x.

7
0.7 Existence and uniqueness of geodesics
Note that for every point p ∈ Rn and and vector X ∈ Tp Rn ' Rn , we may find
a geodesic through p and with velocity vector X at p, which is given simply by
γ(t) = p + Xt. Here we show that all manifolds with a connection share this
property:
Theorem 0.9. Let M be a manifold with a connection. Then for every p ∈ M and
X ∈ Tp M there exists an  > 0 such that for every 0 <  <  there is a unique
geodesic γ : (−, ) → M with γ(0) = p and γ 0 (0) = X.
To prove this theorem, we need to record some preliminary observations. Let
M and M f be manifolds with connections ∇ and ∇ e respectively. We say that a
diffeomorphism f : M → M is connection preserving provided that
f


(∇Y X)p = ∇ e df (Y ) df (X)
f (p)

for all p ∈ M and all vector fields X, Y ∈ X (M ). It is an immediate consequence

of the definitions that
Lemma 0.10. Let f : M → M f be a connection preserving diffeomorphism. Then
γ : I → M is a geodesic if and only of f ◦ γ is a geodesic.
Note that if f : M → Mf is a diffeomorphism, and M has a connection ∇, then
f induces a connection ∇e on Mf by

e pe := ∇ −1 e df −1 (Ye ) −1 .


(∇
e e X)
Y df (X) f (e
p)

It is clear that then f : M → M

f will be connection preserving. So we may conclude
that
Lemma 0.11. Let (U, φ) be a local chart of M , then γ : I → U is a geodesic if and
only of φ ◦ γ is a geodesic with respect to the connection induced on Rn by φ.
Now we are ready to prove the main result of this section:

Proof of Theorem 0.9. Let (U, φ) be a local chart of M centered at p and let ∇ be
the connection which is induced on φ(U ) = Rn by φ. We will show that there exists
an  > 0 such that for every 0 <  <  there is a unique geodesic c : (−, ) → Rn ,
with respect to the induced connection, which satisfies the initial conditions

c(0) = φ(p) and c0 (0) = dφp (X).

Then, by a previous lemma, γ := φ−1 ◦ c : (−, ) → M will be a geodesic on M

with γ(0) = p and γ 0 (0) = X. Furthermore, γ will be unique. To see this suppose
that γ : (−, ) → M is another geodesic with γ(0) = p and γ 0 (0) = X. Let 0
be the supremum of t ∈ [0, ] such that γ(−t, t) ⊂ U , and set c := φ ◦ γ|(−0 ,0 ) .

8
Then, by Theorem 0.7, c = c on (−0 , 0 ), because 0 < . So it follows that γ = γ on
(−0 , 0 ), and we are done if (−0 , 0 ) = (−, ). This is indeed the case, for otherwise,
(−0 − δ, 0 + δ) ⊂ (−, ), for some δ > 0. Further γ(±0 ) = γ(±0 ) ∈ U . So if δ is
sufficiently small, then γ(−0 − δ, 0 + δ) ⊂ U , which contradicts the definition of 0 .
So all we need is to establish the existence and uniqueness of the geodesic
c : (−, ) → Rn mentioned above. For c to be a geodesic we must have
Dc c0 ≡ 0.
We will show that this may be written as a system of ordinary differential equations.
To see this first recall that
Dc ċ(t) = ∇ċ(t) ċ
where ċ is a vector filed in a neighborhood of c(t) which is a local extension of ċ,
i.e.,
ċ(c(t)) = ċ(t).
By (1) we have
X k
X 
∇ċ(t) ċ = ċ(t)(ċ ) + ċi (t)ċj (t)Γkij (c(t)) ek ,
k ij

where ei are the standard basis of Rn and Γkij (p) = h(∇ei ej )p , ek i. But
k k
ċ(t)(ċ ) = (ċ ◦ c)0 (t) = (ċk )0 (t) = c̈k (t).
So Dc c0 ≡ 0 if and only if
X
c̈k (t) + ċi (t)ċj (t)Γkij (c(t)) = 0
ij

for all t ∈ I and all k. This is a system of n second order ordinary differential
equations (ODEs), which we may rewrite as a system of 2n first order ODEs, via
substitution ċ = v. Then we have
ċk (t) = v k (t)
X
v̇ k (t) = − v i (t)v j (t)Γkij (c(t)).
ij

Now let α(t) := (c(t), v(t)), and define F : R2n → R2n , F = (F 1 , . . . , F 2n ) by

X
F ` (x, y) = y` , and F `+n (x, y) = − y i y j Γ`ij (x)
ij

for ` = 1, . . . , n. Then the system of 2n ODEs mentioned above may be rewritten

as
α0 (t) = F (α(t)),
which has a unique solution with initial conditions α(0) = (φ(p), dφ(X)).

9
0.8 Parallel translation
Let M be a manifold with a connection, and γ : I → M be an immersed curve.
Then we say that a vector field X ∈ X (γ) is parallel along γ if

Dγ X ≡ 0.

Thus, in this terminology, γ is a geodesic if its velocity vector field is parallel. Further
note that if M is a submanifold of Rn , the, by the earlier results in this section, X
is parallel along γ if and only (X 0 )> ≡ 0.

Example 0.12. Let M be a two dimensional manifold immersed in Rn , γ : I → M

be a geodesic of M , and X ∈ XM (γ) be a vector field along γ in M . Then X is
parallel along γ if and only if X has constant length and the angle between X(t)
and γ 0 (t) is constant as well. To see this note that (γ 00 )> ≡ 0 since γ is a geodesic;
therefore,
hX, γ 0 i0 = hX 0 , γ 0 i + hX, γ 00 i = hX 0 , γ 0 i.
So, if (X 0 )> = 0, then it follows that hX, γ 0 i is constant which since γ 0 and X
have both constant lengths, implies that the angle between X and γ 0 is constant.
Conversely, suppose that X has constant length and makes a constant angle with γ 0 .
Then hX, γ 0 i is constant, and the displayed expression above implies that hX, γ 0 i = 0
is constant. Furthermore, 0 = hX, Xi0 = 2hX, X 0 i. So X 0 (t) is orthogonal to both
X(t) and γ 0 (t). If X(t) and γ 0 (t) are linearly dependent, then this implies that X 0 (t)
is orthogonal to Tγ(t) M , i.e., (X 0 )> ≡ 0. If X(t) and γ 0 (t) are linearly dependent,
then (X 0 )> = Dγ (X) = Dγ (f γ 0 ) = f Dγ (γ 0 ) ≡ 0.

Example 0.13 (Foucault’s Pendulum). Here we explicitly compute the parallel

translation of a vector along a meridian of the sphere. To this end let

X(θ, φ) := (cos(θ) sin(φ), sin(θ) sin(φ), cos(φ))

be the standard parametrization or local coordinates for S2 − {(0, 0, ±1)}. Suppose

that we want to parallel transport a given unit vector V0 ∈ TX(θ0 ,φ0 ) S2 along the
meridian X(θ, φ0 ), where we identify tangent space of S2 with subspaces of R3 .
So we need to find a mapping V : [0, 2π] → S2 such that V (0) = V0 and V 0 (θ) ⊥
TX(θ,φ0 ) S2 . The latter condition is equivalent to the requirement that

V 0 (θ) = λ(θ)X(θ, φ0 ), (2)

since the normal to S2 at the point X(θ, φ) is just X(θ, φ) itself. To solve the above
differential equation, let

∂X/∂θ(θ, φ0 )
E1 (θ) := = (− sin(θ), cos(θ), 0),
k∂X/∂θ(θ, φ0 k

10
and
∂X/∂φ(θ, φ0 )
E2 (θ) := = (cos(θ) cos(φ0 ), sin(θ) cos(φ0 ), − sin(φ0 )).
k∂X/∂φ(θ, φ0 k
Now note that {E1 (θ), E2 (θ)} forms an orthonormal basis for TX(θ0 ,φ0 ) S2 . Thus (2)
is equivalent to

hV 0 (θ), E1 (θ)i = 0 and hV 0 (θ), E2 (θ)i = 0. (3)

So it remains to solve this differential equation. To this end first recall that since
V0 has unit length, and parallel translation preserves length, we may write

V (θ) = cos(α(θ))E1 (θ) + sin(α(θ))E2 (θ).

So differentiation yields that

V 0 = E10 cos(α) − sin(α)α0 E1 + sin(α)E20 + cos(α)α0 E2 .

Further, it is easy to compute that

E10 = − cos(φ0 )E2 − sin(φ0 )E3 and E20 = cos(φ0 )E1 ,

where E3 (θ) := X(θ, φ0 ). Thus we obtain:

V 0 = sin(α)(cos(φ0 ) − α0 )E1 + cos(α)(α0 − cos(φ0 ))E2 + (∗)E3 .

So for (3) to be satisfied, we must have α0 = cos(φ0 ) or

α(θ) = cos(φ0 )t + α(0),

which in turns determines V . Note in particular that the total rotation of V with
respect to the meridian X(θ, φ0 ) is given by
Z 2π
α(2π) − α(0) = α0 dθ = 2π cos(φ0 ).
0

Thus  
−1 α(2π) − α(0)
φ0 = cos .
2π
The last equation gives the relation between the precession of the swing plane of
a pendulum during a 24 hour period, and the longitude of the location of that
pendulum on earth, as first observed by the French Physicist Leon Foucault in
1851.
Lemma 0.14. Let I ⊂ R and U ⊂ Rn be open subsets and F : I × U → Rn , be
C 1 . Then for every t0 ∈ I and x0 ∈ U there exists an  > such that for every
0 <  <  there is a unique curve x : (t0 − , t0 + ) → Rn with x(t0 ) = x0 and
x0 (t) = F (t, x(t)).

11
Proof. Define F : I × U → Rn+1 by F (t, x) := (1, F (t, x)). Then, by Theorem 0.7,
there exists an  > 0 and a unique curve x : (t0 −, t0 +) → Rn+1 , for every 0 <  < ,
such that x(t0 ) = (1, x0 ) and x0 (t) = F (x(t)). It follows then that x(t) = (t, x(t)),
for some unique curve x : (t0 − , t0 + ) → Rn . Thus F (x(t)) = (1, F (t, x(t))), and
it follows that x0 (t) = F (t, x(t)).

Lemma 0.15. Let A(t), t ∈ I, be a C 1 one-parameter family of matrices. Then for

every x0 ∈ Rn and t0 ∈ I, there exists a unique curve x : I → Rn with x(t0 ) = x0
such that x0 (t) = A(t) · x(t).
Proof. Define F : I × Rn → Rn by Ft (x) = A(t) · x. By the previous lemma,
there exists a unique curve x : (t0 − , t0 + ) → Rn with x(t0 ) = x0 such that
Ft (x(t)) = x0 (t) for all t ∈ (t0 − , t0 + ).
Now let J ⊂ I be the union of all open intervals in I which contains t0 and such
that x0 (t) = F (x(t)) for all t in those intervals. Then J is open in I and nonempty.
All we need then is to show that J is closed, for then it would follow that J = I.
Suppose that t is a limit point of J in I. Just as we argued in the first paragraph,
there exists a curve y : (t − , t + ) → Rn such that y 0 (t) = F (y(t)) and y 0 (t) 6= 0.
Thus we may assume that y 0 6= 0 on (t − , t + ), after replacing  by a smaller
number. In particular y 0 (e t) 6= 0 for some et ∈ (t − , t + ) ∩ J, and there exists a
matrix B such that B · y (e 0 t) = x0 (e
t).
Now let y(t) := B · y(t). Since F (y(t)) = y 0 (t), we have F (y(t)) = y 0 (t). Further,
by construction y(e t) = x(e t), so by uniqueness part of the previous result we must
have y = x on (t − , t + ) ∩ J. Thus x is defined on J ∪ (t − , t + ). But J was
assumed to be maximal. So (t − , t + ) ⊂ J. In particular t ∈ J, which completes
the proof that J is closed in I.

Theorem 0.16. Let X : I → M be a C 1 immersion. For every t0 ∈ I and X0 ∈

Tγ(t0 ) M , there exists a unique parallel vector field X ∈ X (γ) such that X(t0 ) = X0 .
Proof. First suppose that there exists a local chart (U, φ) such that γ : I → U is an
embedding. Let X be a vector field on U and set X(t) := X(γ(t)). By (1),
X k X 
Dγ (X)(t) = ∇γ 0 (t) X = γ 0 (t)(X ) + γ i (t)X j (t)Γkij (γ(t)) Ek (γ(t)).
k ij

Further note that

γ 0 (t)X = (X ◦ γ)0 (t) = X 0 (t).
So, in order for X to be parallel along γ we need to have
X
Ẋ k + γ i (t)Γkij (γ(t))X j (t) = 0,
ij

for k = 1, . . . , n. This is a linear system of ODE’s in terms of X i , and therefore by

the previous lemma it has a unique solution on I satisfying the initial conditions
X i (t0 ) = X0i .

12
 Now let J ⊂ I be a compact interval which contains t0 . There exists a finite
number of local charts of M which cover γ(J). Consequently there exist subintervals
J1 , . . . , Jn of J such that γ embeds each Ji into a local chart of M . Suppose that
t0 ∈ J` , then, by the previous paragraph, we may extend X0 to a parallel vector
field defined on J` . Take an element of this extension which lies in a subinterval J`0
intersecting J` and apply the previous paragraph to J`0 . Repeating this procedure,
we obtain a parallel vector field on each Ji . By the uniqueness of each local extension
mentioned above, these vector fields coincide on the overlaps of Ji . Thus we obtain
a well-defined vector filed X on J which is a parallel extension of X0 . Note that
if J is any other compact subinterval of I which contains t0 , and X is the parallel
extension of X0 on J, then X and X coincide on J ∩ J, by the uniqueness of local
parallel extensions. Thus, since each point of I is contained in a compact subinterval
containing t0 , we may consistently define X on all of I.
Finally let X be another parallel extension of X0 defined on I. Let A ⊂ I be the
set of points where X = X. Then A is closed, by continuity of X and X. Further
A is open by the uniqueness of local extensions. Furthermore, A is nonempty since
t0 ∈ A. So A = I and we conclude that X is unique.

Using the previous result we now define, for every X0 ∈ Tγ(t0 ) M ,

Pγ,t0 ,t (X0 ) := X(t)

as the parallel transport of X0 along γ to Tγ(t) M . Thus we obtain a mapping from

Tγ(t0 ) M to Tγ(t) M .

Exercise 0.17. Show that Pγ,t0 ,t : Tγ(t0 ) M → Tγ(t) M is an isomorphism (Hint: Use
the fact hat Dγ : X (γ) → X (γ) is linear). Also show that Pγ,t0 ,t depends on the
choice of γ.

Exercise 0.18. Show that

−1
Xγ(t0 ) − Pγ,t 0 ,t
(Xγ(t) )
∇γ 0 (t0 ) X = lim .
t→t0 t
(Hint: Use a parallel frame along γ.)

13