Riemannian Geometry

Eckhard Meinrenken

Lecture Notes, University of Toronto, Spring 2002

 Contents

1. Manifolds 4
2. Examples of manifolds 6
3. Submanifolds 7
4. Tangent spaces 9
5. Tangent map 11
6. Tangent bundle 13
7. Vector fields as derivations 15
8. Flows of vector fields 17
9. Geometric interpretation of the Lie bracket 21
10. Lie groups and Lie algebras 23
11. Frobenius’ theorem 27
12. Riemannian metrics 33
13. Existence of Riemannian metrics 35
14. Length of curves 36
15. Connections and parallel transport 38
16. Geodesics 43
17. The Hopf-Rinow Theorem 47
18. The curvature tensor 51
19. Connections on vector bundles 53

3
4 CONTENTS

1. Manifolds
1.1. Motivation. One of Riemann’s key ideas was to develop the notion of a “manifold”
independent from an embedding into an ambient Euclidean space. Roughly, an n-dimensional
manifold is a space that locally looks like Rn . More precisely, a manifold is a space that can be
covered by coordinate charts, in such a way that the change of coordinates between any two
charts is a smooth map. The following examples should give an idea what we have in mind.
Example 1.1. Let S 2 ⊂ R3 be the unit sphere defined by the equation (x0 )2 +(x1 )2 +(x2 )2 .
For j = 0, 1, 2 let Uj+ ⊂ S 2 be the subset defined by xj > 0 and Uj− the subset defined by
xj < 0. Let φ± ± 2
j : Uj → R be the maps omitting the jth coordinate. Then all transition maps
± ± −1
φj ◦ (φk ) are smooth. For instance,
φ+ − − − + + −
2 ◦ φ1 : φ1 (U1 ∩ U2 ) → φ2 (U1 ∩ U2 )
+
√
is the map (u, v) 7→ (u, − 1 − u2 − v 2 ), and this is smooth since u2 + v 2 < 1 on the image of
φ−1.
Example 1.2. The real projective plane RP (2) is the set of all lines (=1-dimensional
subspaces) in R3 . Any such line is determined by its two points of intersection{x, −x} with S 2 .
Thus RP (2) may be identified with the quotient of S 2 by the equivalence relation, x ∼ −x.
Let π : S 2 → RP (2) be the quotient map. To get a picture of RP (2), note that for 0 < ǫ < 1,
the subset {(x0 , x1 , x2 ) ∈ S 2 | x2 ≥ ǫ} is a 2-disk, containing at most one element of each
equivalence class. Hence its image under π is again a 2-disk. On the other hand, the strip
{(x0 , x1 , x2 ) ∈ S 2 | − ǫ ≤ x2 ≥ ǫ} contains, with any x, also the point −x. Its image under π
looks like a Moebius strip. Thus RP (2) looks like a union of a Moebius strip and a disk, glued
along their boundary circles. This is still somewhat hard to imagine, since we cannot perform
this gluing in such a way that RP (2) would become a surface in R3 . Nonetheless, it “should
be” a surface: Using the coordinate charts from S 2 , let Uj = π(Uj+ ), and let φj : Uj → R2
be the unique maps such that π ◦ φj = φ+ j . Then the Uj cover RP (2), and the “change of
coordinate” maps are again smooth.
It is indeed possible to embed RP (2) into R4 : One possibility is the map,
(1) [(x0 , x1 , x2 )] 7→ (x1 x2 , x0 x2 , x0 x1 , t0 x20 + t1 x21 + t2 x22 )
where t0 , t1 , t2 ∈ R are distinct (e.g. t0 = 1, t1 = 2, t2 = 3). However, these embedding do not
induce the “natural” metric on projective space, i.e. the metric induced from the 2-sphere.
1.2. Topological spaces. To develop the concept of a manifold as a “space that locally
looks like Rn ”, our space first of all has to come equipped with some topology (so that the word
“local” makes sense). Recall that a topological space is a set M , together with a collection of
subsets of M , called open subsets, satisfying the following three axioms: (i) the empty set ∅ and
the space M itself are both open, (ii) the intersection of any finite collection of open subsets is
open, (iii) the union of any collection of open subsets is open. The collection of open subsets of
M is also called the topology of M . A map f : M1 → M2 between topological spaces is called
continuous if the pre-image of any open subset in M2 is open in M1 . A continuous map with
a continuous inverse is called a homeomorphism.
One basic ingredient in the definition of a manifold is that our topological space comes
equipped with a covering by open sets which are homeomorphic to open subsets of Rn .
 1. MANIFOLDS 5

Definition 1.3. Let M be a topological space. An n-dimensional chart for M is a pair

(U, φ) consisting of an open subset U ⊂ Rn and a continuous map φ : U → Rn such that φ is a
homeomorphism onto its image φ(U ). Two such charts (Uα , φα ), (Uβ , φβ ) are C ∞ -compatible
if the transition map
φβ ◦ φ−1
α : φα (Uα ∩ Uβ ) → φβ (Uα ∩ Uβ )
is a diffeomorphism (a smooth map with smooth inverse). A covering A = (Uα )α∈A of M by
pairwise C ∞ -compatible charts is called a C ∞ -atlas.
Example 1.4. Let X ⊂ R2 be the union of lines R × {1} ∪ R × {−1}. Let M = X/ ∼ be
its quotient by the equivalence relation, (u, 1) ∼ (u, −1) for u < 0. Let π : X → M be the
quotient map.
Thus M is obtained by gluing to copies of the real line along the negative axis. It is
somewhat hard to picture this space, since π(0, ±1) are distinct points in M . Nonetheless, M
admits a C ∞ -atlas: Let U+ = π(R × {1}) and U− = π(R × {−1}), and define φ± : U± → R by
φ± (u, ±1) = u. Then (U± , φ± ) defines an atlas with two charts (the transition map is just the
identity map).
The example just given shows that existence of an atlas does not imply that our space
looks “nice”. The problem with the example is that the points π(0, ±1) in M do not admit
disjoint open neighborhoods. Recall that a topological space is called Hausdorff if any two
distinct points in the space admit disjoint open neighborhoods. Thus, we require manifolds to
be Hausdorff.
We will impose another restriction on the topology. Recall that A basis for a topological
space M is a collection B of open subsets of M such that every open subset of M is a union
of open subsets in the collection B. For example, the collection of open balls Bǫ (x) in Rn
define a basis. But one already has a basis if one takes only all balls Bǫ (x) with x ∈ Qn and
ǫ ∈ Q>0 ; this then defines a countable basis. A topological space with countable basis is also
called second countable. We will require manifolds to admit a countable basis. This will imply,
among other things, that the manifold admits a countable atlas, a fact that is useful for certain
inductive arguments.
Two atlases on a topological space are called equivalent if their union is again an atlas. It
is not hard to check that this is indeed an equivalence relation. An equivalence class of atlases
is called a C ∞ -structure on M .
Definition 1.5 (Manifolds). A C ∞ -manifold is a Hausdorff topological space M , with
countable basis, together with a C ∞ -structure.
It is perhaps somewhat surprising that the two topological restrictions (Hausdorff and
countable basis) rule out any further “accidents”: The topological properties of manifolds are
just as nice as those of Euclidean Rn .
Definition 1.6. A map F : N → M between manifolds is called smooth (or C ∞ ) if for
all charts (U, φ) of N and (V, ψ) of M , with F (U ) ⊂ V , the composite map
ψ ◦ F ◦ φ−1 : φ(U ) → ψ(V )
is smooth. The space of smooth maps from N to M is denoted C ∞ (N, M ). A smooth map
F : N → M with smooth inverse F −1 : M → N is called a diffeomorphism.
6 CONTENTS

Problems 1.7. 1. Review point set topology: Continuous maps, coverings, neigh-
borhoods, Hausdorff property, compactness, ...
2. Show that equivalence of C ∞ -atlases is an equivalence relation. Warning: C ∞ -compa-
tibility of charts on a topological space is not an equivalence relation. (Why?)
3. Given a manifold M with C ∞ -atlas A, let A′ be the collection of all C ∞ -charts (U, φ)
on M that are compatible with all charts in A. Show that A′ is again an atlas, and
that A′ contains any atlas equivalent to A.
4. Verify that the map (1) is 1-1.

2. Examples of manifolds

Spheres. The unit sphere S n ⊂ Rn+1 is a manifold of dimension n, with charts Uj± constructed
similar to S 2 . Another choice of atlas, with only two charts, is given by “stereographic projec-
tion” from the north and south pole.

Projective spaces. Let RP (n) be the quotient S n / ∼ under the equivalence relation x ∼ −x.
It is easy to check that this is Hausdorff and has countable basis. Let π : S n → RP (n) be the
quotient map. Just as for n = 2, the charts Uj = π(Uj+ ), with map φj induced from Uj+ , form
an atlas.

Products. If Mj are aPfinite collection of manifolds of dimensions nj , their direct product is a

manifold of dimension nj . For instance, the n-torus is defined as the n-fold product of S 1 ’s.

Lens spaces. Identify R4 with C2 , thus S 3 = {(z, w) : |z|2 + |w|2 = 1}. Given natural
numbers q > p ≥ 1 introduce an equivalence relation, by declaring that (z, w) ∼ (z ′ , w′ ) if
2πi kq 2πi kp
(z ′ , w′ ) = (e z, e q w)
for some k ∈ {0, . . . , q − 1}. Let L(p, q) = S 3 / ∼ be the lens space. Note that L(1, 2) = RP (3).
If p, q are relatively prime, L(p, q) is a manifold. Indeed, if p, q are relatively prime then for all
(z, w) ∈ S 3 , the only solution of
2πi kq 2πi kp
(z, w) = Φk (z, w) := (e z, e q w)
is k = 0. Let fk (z, w) = ||(z, w) − Φk (z, w)||. Then fk > 0 for k = 1, . . . , q − 1. Since S 3 is
compact, each fk takes on its minimum on S 3 . Let ǫ > 0 be sufficiently small so that fk > ǫ
for all k = 1, . . . , q − 1.
Then if U is an open subset of S 3 that is contained in some open ball of radius ǫ in R3 ,
then U contains at most one element of each equivalence class. Let (U, φ) be a coordinate chart
for S 3 , with U sufficiently small in this sense. Let V = π(U ), and ψ : V → R3 the unique
map such that ψ ◦ π = φ. Then (V, ψ) is a coordinate chart for L(p, q), and the collection of
coordinate charts constructed this way defines an atlas.

Grassmannians. The set Gr(k, n) of all k-dimensional subspaces of Rn is called the Grass-
mannian of k-planes in Rn . A C ∞ -atlas may be constructed as follows. For any subset
 3. SUBMANIFOLDS 7

I ⊂ {1, . . . , n} let I ′ = {1, . . . , n}\I be its complement. Let RI ⊂ Rn be the subspace consisting
of all x ∈ Rn with xi = 0 for i 6∈ I.
′
If I has cardinality k, then RI ∈ Gr(k, n). Note that RI = (RI )⊥ . Let UI = {E ∈
′
Gr(k, n)| E ∩ RI = 0}. Each E ∈ UI is described as the graph of a unique linear map
′
A : RI → RI , that E = {x + A(x)| x ∈ RI . This gives a bijection
φI : UI → L(RI , RI ) ∼
′
= Rk(n−k) .
We can use this to define the topology on the Grassmannian: It is the smallest topology for
which all maps φI are continuous. To check that the charts are compatible, suppose E ∈ UI ∩UI˜,
and let AI and AI˜ be the linear maps describing E in the two charts. We have to show that
the map taking AI to AI˜ is smooth. Let ΠI denote orthogonal projection Rn → RI . The map
AI is determined by the equations
AI (xI ) = (1 − ΠI )x, xI = ΠI x
for x ∈ E, and x = xI + AI xI . Thus
AI˜(xI˜) = (I − ΠI˜)(AI + 1)xI , xI˜ = ΠI˜(AI + 1)xI .
˜
The map S(AI ) : ΠI˜(AI + 1) : RI → RI is an isomorphism, since it is the composition of two
˜
isomorphisms (AI + 1) : RI → E and ΠI˜|E : E → RI . The above equations show,
AI˜ = (I − ΠI˜)(AI + 1)S(AI )−1 .
The dependence of S on the matrix entries of AI is smooth, by Cramer’s formula for the inverse
matrix. It follows that the collection of all φI : UI → Rk(n−k) defines on Gr(k, n) the structure
of a manifold of dimension k(n − k).

Rotation groups. Let Matn ∼

2
= Rn be the set of n × n-matrices. The subset SO(n) = {A ∈
Matn | At A = I, det(A) = 1 is the group of rotations in Rn . Let so(n) = {B ∈ Matn | B t + B =
0} ∼
= Rn(n−1)/2 . Then exp(B) ∈ SO(n) for all B ∈ so(n). For ǫ sufficiently small, exp restricts to
a bijection from V = {B ∈ so(n)| ||B|| < ǫ}. For any A0 , let U = {A ∈ SO(n)| A = A0 exp(B)}.
Let φ be the map taking A to B = log(AA−1 0 ). Then the set of all (U, φ) constructed this way
define an atlas, and give SO(n) the structure of a manifold of dimension n(n − 1)/2.

3. Submanifolds
Let M be a manifold of dimension m.
Definition 3.1. A subset S ⊂ M is called an embedded submanifold of dimension k ≤ m,
if S can be covered by coordinate charts (U, φ) for M with the property φ(U ∩ S) = Φ(U ) ∩ Rk .
Charts (U, φ) of M with this property are called submanifold charts for S.
Thus S becomes a k-dimensional manifold in its own right, with atlas consisting of charts
(U ∩ S, φ|U ∩S ).
Example 3.2. S n is a submanifold of Rn+1 : A typical submanifold chart is
P q P
V = {x ∈ Rn+1 | x0 > 0, i>0 x2i < 1}, φ(x) = (x1 , . . . , xn , 1 − i>0 x2i − x0 ).
8 CONTENTS

Example 3.3. Similarly, if f : U → Rn−k is any smooth function on an open subset

U ⊂ Rk , the graph Γf = {(x, f (x))| x ∈ U } is an embedded submanifold of U × Rk , with
submanifold chart
φ : U × Rk → Rn , (y ′ , y ′′ ) 7→ (y ′ , f (y ′ ) − y ′′ )
Recall that for any smooth function F : V → Rm on an open subset V ⊂ Rn , a point
a ∈ Rm is a regular value if for all x ∈ F −1 (a), the Jacobian DF (x) : Rn → Rm is onto. If
m = 1, this means that the gradient ∇F does not vanish on F −1 (a). (We do not require that
a is in the image of F – thus regular value is a bit misleading.)
Proposition 3.4. Let V ⊂ Rn open, and F : V → Rm smooth. For any regular value
a ∈ Rm of F , the inverse image F −1 (a) is an embedded submanifold of dimension k = n − m.
In fact, there exists a submanifold chart (U, φ) around any given x ∈ F −1 (a) such that
F (φ−1 (y ′ , y ′′ )) = a + y ′′
for all (y ′ , y ′′ ) ∈ φ(U ) ⊂ Rk × Rm .
Proof. This is really just a version of the implicit function theorem from multivari-
able calculus. A more familiar version of this theorem states that for all x ∈ F −1 (a), af-
ter possibly renumbering the coordinates in Rn , the equation F (y) = a can be “solved” for
y ′′ = (yk+1 , . . . , yn ) as a function of y ′ = (y1 , . . . , yk ). That is, there exists a unique function ga
from a neighborhood of x′ ∈ Rn−m to Rm , such that on a sufficiently small neighborhood U of
x
F −1 (a) ∩ U = {(y ′ , ga (y ′ ))} ∩ U.
This means that on U , the level set F −1 (a) is the graph of the function ga , and therefore an
embedded submanifold.
But in fact, ga depends smoothly on the value a = F (x). That is, taking U sufficiently
small, we have
F −1 (F (y)) ∩ U = {(y ′ , gF (y) (y ′ ))} ∩ U.
for all y = (y ′ , y ′′ ) ∈ U . Then φ(y) = (y ′ , gF (y) (y ′ ) − y ′′ ) is a submanifold chart with the desired
property. 

Manifolds are often described as level sets for regular values:

Example 3.5. For 0 < r < R, the 2-torus can be identified with the embedded submanifold
F −1 (r2 ) where
q
F (x1 , x2 , x3 ) = ( x21 + x22 − R)2 + x23
is a smooth function on the complement of the x3 -axis, x21 + x22 > 0. One checks that indeed,
a = r2 is a regular value of this function.
The proposition generalizes to maps between manifolds: If F ∈ C ∞ (M, N ), a point a ∈ N
is called a regular value of F if it is a regular value “in local coordinates”: That is, for all
p ∈ F −1 (a), and all charts (U, φ) around p and (V, ψ) around a, with F (U ) ⊂ V , the Jacobian
D(ψ ◦ F ◦ φ−1 ) at φ(p) is onto.
 4. TANGENT SPACES 9

Theorem 3.6. If a ∈ N is a regular value of F ∈ C ∞ (M, N ), then F −1 (a) is an embedded

submanifold of M . In fact, given a coordinate chart (V, ψ) around a, with ψ(a) = 0, each
p ∈ F −1 (a) admits a submanifold chart (U, φ) with
ψ ◦ F ◦ φ−1 (y ′ , y ′′ ) = y ′′
for y = (y ′ , y ′′ ) ∈ φ(U ) ⊂ Rn .
Proof. Choose any coordinate chart (U ′ , φ′ ) around x. The Proposition, applied to the
map ψ ◦ F ◦ (φ′ )−1 , gives a change of coordinates with the desired properties. 
Problems 3.7.
1. Show that conversely, every submanifold
p is locally the graph of a function.
2. Let S ⊂ R3 be the 2-torus ( x21 + x22 − R)2 + x23 = r2 , and F : S → R the function
(x1 , x2 , x3 ) 7→ x2 . What are the critical points for this function? What is the shape of the level
sets F −1 (a) for a a singular value?
3. Let Symn ⊂ Matn be the subspace of symmetric matrices, and F : Matn → Symn the map
A 7→ At A. Show that the identity matrix I is a regular value of this map. This proves that
the orthogonal group O(n) is an embedded submanifold of Matn , of dimension n(n − 1)/2. (In
fact, by a theorem of E. Cartan, every closed subset G ⊂ Matn , with the property that G is a
group under matrix multiplication, is an embedded submanifold of Matn .)

4. Tangent spaces
For embedded submanifolds M ⊂ Rn , the tangent space Tp M at p ∈ M can be defined as
the set of all velocity vectors v = γ̇(0), where γ : R → M is a smooth curve with γ(0) = p.
Thus Tp M becomes a vector subspace of Rn . To extend this idea to general manifolds, note
that the vector v = γ̇(0) defines a “directional derivative” C ∞ (M ) → R:
d
v : f 7→ dt |t=0 f (γ(t)).

We will define Tp M as a set of directional derivatives.

Definition 4.1. Let M be a manifold, p ∈ M . The tangent space Tp M is the space of all
linear maps v : C ∞ (M ) → R of the form
d
v(f ) = dt |t=0 f (γ(t))

for some smooth curve γ ∈ C ∞ (R, M ) with γ(0) = p.

The following alternative description of Tp M makes it clear that Tp M is a vector subspace
of the space of linear maps C ∞ (M ) → R, of dimension dim Tp M = dim M .
Proposition 4.2. Let (U, φ) be a coordinate chart around p, with φ(p) = 0. A linear map
v : C ∞ (M ) → R is in Tp M if and only if it has the form,
m
X ∂(f ◦ φ−1 )
v(f ) = ai |x=0
∂xi
i=1

for some a = (a1 , . . . , am ) ∈ Rm .

10 CONTENTS

Proof. Given a linear map v of this form, let γ(t) be any smooth curve with φ(γ(t)) = ta
for |t| sufficiently small 1. Then
d d −1
Pm ∂(f ◦φ−1 )
dt |t=0 f (γ(t)) = dt |t=0 (f ◦ φ )(ta) = i=1 ai ∂xi |x=0 ,
by the chain rule. Conversely, given any curve γ with γ(0) = p, let γ̃ = φ◦γ be the corresponding
curve in φ(U ) (defined for small |t|). Then
d ∂ −1
Pm ∂(f ◦φ−1 )
dt |t=0 f (γ(t)) = ∂t (f ◦ φ )(γ̃(t)) = i=1 ai ∂xi |x=0 ,
dγ̃
where a = dt |t=0 . 
Corollary 4.3. If U ⊂ Rm is an open subset, the tangent space Tp U is canonically
identified with Rm .
We now describe a third definition of Tp M which characterizes “directional derivatives”
in a coordinate-free way, without reference to curves γ. Note first that every tangent vector
v ∈ Tp M satisfies a product rule,
(2) v(f1 f2 ) = f1 (p) v(f2 ) + v(f1 ) f2 (p)
for all fj ∈ C ∞ (M ). Indeed, in local coordinates (U, φ), this just follows from the product rule
from calculus,
∂ ˜ ˜ ∂ f˜2 ∂ f˜1 ˜
(f1 f2 ) = f˜1 (x) + f2 (x)
∂xi ∂xi ∂xi
where f˜j = fj ◦φ−1 . It turns out that the product rule completely characterizes tangent vectors:
Proposition 4.4. A linear map v : C ∞ (M ) → R is a tangent vector if and only if it
satisfies the product rule (2).
Proof. Let v : C ∞ (M ) → R be a linear map satisfying the product rule (2). To show
that v ∈ Tp M , we use the second definition of Tp M in terms of local coordinates.
We first note that by the product rule applied to the constant function 1 = 1 · 1 we have
v(1) = 0. Thus v vanishes on constants. Next we show that v(f1 ) = v(f2 ) if f1 = f2 near p.
Equivalently, we show that v(f ) = 0 if f = 0 near p. Choose χ ∈ C ∞ (M ) with χ(p) = 1, zero
outside a small neighborhood of p so that f χ = 0. The product rule tells us that
0 = v(f χ) = v(f )χ(p) + v(χ)f (p) = v(f ).
Thus v(f ) depends only on the behavior of f in an arbitrarily small neighborhood of p. In
particular, letting (U, φ) be a coordinate chart around p, with φ(p) = 0, we may assume that
supp(f ) ⊂ U .2 Consider the Taylor expansion of f˜ = f ◦ φ−1 near x = 0:
X ∂
f˜(x) = f˜(0) + xi |x=0 f˜ + r(x)
∂xi
i
The remainder term r is a smooth function that vanishes at x = 0 together with P its first
derivatives. This means that it can be written (non-uniquely) in the form r(x) = i xi ri (x)
1More precisely, choose any function χ : R → R with χ(t) = t for |t| < ǫ/2 and χ̇(t) = 0 for |t| ≥ ǫ.
Choose ǫ sufficiently small, so that the ball of radius ǫ||a|| is contained in φ(U ). Then χ(t)a ∈ φ(U ) for all t,
and γ(t) = φ−1 (χ(t)a) is a well-defined curve with the desired properties.
2The support supp(f ) of a function f on M is the closure of the set of all points where it is non-zero.
 5. TANGENT MAP 11

where ri are smooth functions that vanish at 0.3 By the product rule, v vanishes on r ◦ φ−1
(since it is a sum of products of functions that vanish at p). It also vanishes on the constant
f˜(0) = f (p). Thus
X ∂
v(f ) = v(f˜ ◦ φ−1 ) = ai |x=0 f˜
∂xi
i
with ai = v(xi ◦ φ−1 ). (Here the coordinates xi are viewed as functions on Rn , x 7→ xi .) 
Remark 4.5. There is a fourth definition of Tp M , as follows. Let Cp∞ (M )
denote the
∞ 2
P
subspace of functions vanishing at p, and let Cp (M ) consist of finite sums i fi gi where
fi , gi ∈ Cp∞ (M ). Since any tangent vector vanishes v : C ∞ (M ) → R vanishes on constants, v
is effectively a map v : Cp∞ (M ) → R. Since tangent vectors vanish on products, v vanishes on
the subspace Cp∞ (M )2 ⊂ Cp∞ (M ). Thus v descends to a linear map Cp∞ (M )/Cp∞ (M )2 → R,
i.e. an element of the dual space (Cp∞ (M )/Cp∞ (M )2 )∗ . The map
Tp M → (Cp∞ (M )/Cp∞ (M )2 )∗
just defined is an isomorphism, and can therefore be used as a definition of Tp M . This may
appear very fancy on first sight, but really just says that a tangent vector is a linear functional
on C ∞ (M ) that vanishes on constants and depends only on the first order Taylor expansion of
the function at p.

5. Tangent map
Definition 5.1. For any smooth map F ∈ C ∞ (M, N ) and any p ∈ M , the tangent map
Tp F : Tp M → TF (p) N is defined by the equation
Tp F (v)(f ) = v(f ◦ F )
It is easy to check (using any of the definitions of tangent space) that Tp F (v) is indeed a
tangent vector. For example, if γ : R → M is a curve on M representing v, we have
d
Tp F (v)(f ) = v(f ◦ F ) = |t=0 f (F (γ(t))
dt
which shows that Tp F (v) is the tangent vector at F (p) represented by the curve F ◦γ : R → N .
Similarly, it is easily verified that under composition of functions,
Tp (F2 ◦ F1 ) = TF1 (p) F2 ◦ Tp F1 .
In particular, if F is a diffeomorphism, Tp F is invertible and we have
TF (p) F −1 = (Tp F )−1 .
It is instructive to work out the expression for Tp F in local coordinates. We had seen
that any chart (U, φ) around p defines an isomorphism Tp M → Rm . This is the same as the
isomorphism given by the tangent map,
Tp φ : Tp U = Tp M → Tφ(p) φ(U ) = Rm .
Similarly, a chart (V, ψ) around F (p) gives an identification TF (p) ψ : TF (p) V ∼
= Rn . Suppose
F (U ) ⊂ V .
3Exercise: Show that if h is any function on Rn with h(0) = 0, then h can be written in the form h = P x h
i i
where all hi are smooth. Show that if the first derivatives of h vanish at 0, then hi (0) = 0.
12 CONTENTS

Theorem 5.2. In local charts (U, φ) and (V, ψ) as above, the map

TF (p) ψ ◦ Tp F ◦ (Tp φ)−1 : Rm → Rn

is the Jacobian of the map F̃ = ψ ◦ F ◦ φ−1 : φ(U ) → ψ(V ).

Proof. Let a ∈ Rm represent v ∈ Tp M in the chart (U, φ), and let b ∈ Rn represent its
image under Tp F . We denote the coordinates on φ(U ) by x1 , . . . , xm and the coordinates on
ψ(V ) by y1 , . . . , yn . Let f˜ = f ◦ ψ −1 ∈ C ∞ (ψ(V )). Then
m
X ∂
v(f ◦ F ) = ai | f (F (φ−1 (x)))
∂xi x=φ(p)
i=1
Xm
∂
= ai | f˜(F̃ (x))
∂xi x=φ(p)
i=1
m Xn
X ∂ F̃j ∂ f˜
= ai |x=φ(p) |
∂xi ∂yj y=ψ(F (p))
i=1 j=1
n
X ∂ f˜
= bj |
∂yj y=ψ(F (p))
j=1

Pm ∂ F̃j
where bj = i=1 ∂xi |x=φ(p) ai = (DF̃ )ji ai . 

Thus Tp F is just the Jacobian expressed in a coordinate-free way. As an immediate appli-

cation, we can characterize regular values in a coordinate-free way:

Definition 5.3. A point q ∈ N is a regular value of F ∈ C ∞ (M, N ) if and only if the

tangent map Tp F is onto for all p ∈ F −1 (q).

This is clearly equivalent to our earlier definition in local charts.

Definition 5.4. Let γ ∈ C ∞ (J, M ) be a smooth curve (J ⊂ R an open interval). The

tangent (or velocity) vector to γ at time t0 ∈ J is the vector
d
γ̇(t0 ) := Tt0 γ( |t=t0 ) ∈ Tγ(t0 ) M.
dt
dγ
We will also use the notation dt (t0 ) to denote the velocity vector.
d(F ◦γ)
Problems 5.5. 1. Show that if F ∈ C ∞ (M, N ), Tγ(t) F ( dγdt ) = dt for all t ∈ J.
2. Suppose that S ⊂ M is an embedded submanifold, and let ι : S → M, p 7→ p be the
inclusion map. Show that ι is smooth and that the tangent map Tp ι is 1-1 for all p ∈ S. Show
that if M is an open subset of Rm , this becomes the identification of Tp S as a subspace of Rm ,
as described at the beginning of this section.
3. Suppose F ∈ C ∞ (M, N ) has q ∈ N as a regular value. Let S = F −1 (q) ֒→ M be the
level set. For p ∈ S, show that Tp S is the kernel of the tangent map Tp F .
 6. TANGENT BUNDLE 13

6. Tangent bundle
Let M be a manifold of dimension m. If M is an embedded submanifold of Rn , the tangent
bundle T M is the subset of R2n = Rn × Rn given by
T M = {(p, v) ∈ Rn × Rn | p ∈ M, v ∈ Tp M }
where each Tp M is identified as a vector subspace of Rn . It is not hard to see that T M
is, in fact, a smooth embedded submanifold of dimension 2m. Moreover, the natural map
π : T M → M, (p, v) 7→ p is smooth, and its “fibers” π −1 (p) = Tp M carry the structure of
vector spaces.
Definition 6.1. A vector bundle of rank k over a manifold M is a manifold E, together
with a smooth map π : E → M , and a structure of a vector space on each fiber Ep := π −1 (p),
satisfying the following local triviality condition: Each point in M admits an open neighborhood
U , and a smooth map
ψ : π −1 (U ) → U × Rk ,
such that ψ restricts to linear isomorphisms Ep → Rk for all p ∈ U .
The map ψ : EU ≡ π −1 (U ) → U × Rk is called a (local) trivialization of E over U . In
general, there need not be a trivialization over U = M .
Definition 6.2. A vector bundle chart for a vector bundle π : E → M is a chart (U, φ)
for M , together with a chart (π −1 (U ), φ̂) for EU = π −1 (U ), such that φ̂ : π −1 (U ) → Rm × Rk
restricts to linear isomorphisms from each fiber Ep onto {φ(p)} × Rk .
Every vector bundle chart defines a local trivialization. Conversely, if ψ : E|U → U × Rk
is a trivialization of EU , where U is the domain of a chart (U, φ), one obtains a vector bundle
chart (π −1 (U ), φ̂) for E.
Example 6.3. (Vector bundles over the Grassmannnian) For any p ∈ Gr(k, n), let Ep ⊂ Rn
be the k-plane it represents. Then E = ∪p∈Gr(k,n) Ep is a vector bundle over Gr(k, n), called
′
the tautological vector bundle. Recall the definition of charts φI : UI → L(RI , RI ) for the
Grassmannian, where any p = {E} = UI is identified with the linear map A having E as its
graph. Let
′
φ̂I : π −1 (UI ) → L(RI , RI ) × RI
be the map φ̂I (v) = (φ(π(v)), πI (v)) where πI : Rn → RI is orthogonal projection. The φ̂I
serve as bundle charts for the tautological vector bundle. There is another natural vector
bundle E ′ over Gr(k, n), with fiber Ep′ := Ep⊥ the orthogonal complement of Ep . A special case
is k = 1, where Gr(k, n) = RP (n − 1). In this case E is called the tautological line bundle, and
E ′ the hyperplane bundle.
At this stage, we are mainly interested in tangent bundles of manifolds.
Theorem 6.4. For any manifold M , the disjoint union T M = ∪p∈M Tp M carries the
structure of a vector bundle over M , where π takes v ∈ Tp M to the base point p.
Proof. Recall that any chart (U, φ) for M gives identifications Tp φ : Tp M → Rm for all
p ∈ U . Taking all these maps together, we obtain a bijection,
T φ : π −1 (U ) → U × Rm .
14 CONTENTS

We take the collection of (π −1 (U ), T φ) as vector bundle charts for T M . We need to check that
the transition maps are smooth. If (V, ψ) is another coordinate chart with U ∩ V 6= ∅, the
transition map for π −1 (U ∩ V ) is given by,
T ψ ◦ (T φ)−1 : (U ∩ V ) × Rm → (U ∩ V ) × Rm .
But Tp ψ ◦ (Tp φ)−1 = Tφ(p) (ψ ◦ φ−1 ) is just the Jacobian for the change of coordinates ψ ◦ φ−1 ,
and as such depends smoothly on x = φ(p). 
Definition 6.5. A (smooth) section of a vector bundle π : E → M is a smooth map
σ : M → E with the property π ◦ σ = idM . The space of sections of E is denoted Γ∞ (M, E).
Thus, a section is a family of vectors σp ∈ Ep depending smoothly on p.
Examples 6.6. (a) Every vector bundle has a distinguished section, the zero section
p 7→ σp = 0.
(b) A section of the trivial bundle M × Rk is the same thing as a smooth function from
M to Rk . In particular, if ψ : EU → U × Rk is a local trivialization of a vector bundle
E, the section σ (restricted to U ) becomes a smooth function ψ ◦ σ|U : U → Rk .
(c) Let π : E → M be a rank k vector bundle. A frame for E over U ⊂ M is a collection
of sections σ1 , . . . , σk of EU , such that (σj )p are linearly independent at each point
p ∈ U . Any frame over U defines a P local trivialization ψ : EU → U × Rk , given in
−1
terms of its inverse map ψ (p, a) = j aj (σj )p . Conversely, each local trivialization
gives rise to a frame.
The space Γ∞ (M, E) is a vector space under pointwise addition: (σ1 + σ2 )p = (σ1 )p + (σ2 )p .
Moreover, it is a C ∞ (M )-module under multiplication4: (f σ)p = fp σp .
Definition 6.7. A section of the tangent bundle T M is called a vector field on M . The
space of vector fields is denoted
X(M ) = Γ∞ (M, T M ).
Thus, a vector field X ∈ X(M ) is a family of tangent vectors Xp ∈ Tp M depending smoothly
on the base point.
In the next section, we will discuss the space of vector fields in more detail.
Problems 6.8. 1. Let S ⊂ M be an embedded submanifold. Show that for any
vector bundle π : E → M , the restriction E|S → S is a vector bundle over S. In
particular, T M |S is defined; its sections are called “vector fields along S”. The bundle
T M |S contains the tangent bundle T S as a sub-bundle: For all p ∈ S, Tp S is a vector
subspace of Tp M . The normal bundle of S in M is defined as a “quotient bundle”
νS = T M |S /T S with fibers,
(νS )p = Tp M/Tp S
Show that this is again a vector bundle.

4Here and from now on, we will often write f or f | for the value f (p).
p p
 7. VECTOR FIELDS AS DERIVATIONS 15

2. Let F : M → N be a smooth map, and π : E → N a vector bundle. Show that

F ∗ E := ∪p∈M EF (p)
is a vector bundle over M . It is called the pull-back bundle. Sections of F ∗ (T N ) are
called vector fields along (the map) F . For instance, if γ : J → M is a smooth curve,
the set of velocity vectors γ̇(t) becomes a vector field along γ.
3. Let E, E ′ be two vector bundles over M . Show that
E ⊕ E ′ := ∪p∈M Ep ⊕ Ep′
is again a vector bundle over M . It is called the direct sum (or Whitney sum) of
E and E ′ . For instance, the direct sum of the two natural bundles E, E ′ over the
Grassmannian has fibers Ep ⊕Ep′ = Rn , hence E ⊕E ′ is the trivial bundle Gr(k, n)×Rn .
4. Show that for any vector bundle E → M ,
E ∗ = ∪p∈M Ep∗
(where Ep∗ = L(Ep , R) is the dual space to Ep ) is again a vector bundle. It is called
the dual bundle to E. In particular, one defines T ∗ M := (T M )∗ , called the cotangent
bundle. The sections of T ∗ M are called covector fields or “1-forms”.

7. Vector fields as derivations

Let X ∈ X(M ) be a vector field on M . Each Xp ∈ Tp M defines a linear map Xp :
C ∞ (M ) → R. Letting p vary, this gives a linear map
X : C ∞ (M ) → C ∞ (M ), (X(f ))p = Xp (f ).
Note that the right hand side really does define a smooth function on M . Indeed, this follows
from the expression in local coordinates (U, φ). Let a ∈ C ∞ (φ(U ), Rk ) be the expression of X
in the local trivialization, that is, (T φ)(Xp ) = (φ(p), a(φ(p))). Thus
a(φ(p)) = (a1 (φ(p)), . . . , am (φ(p)))
are simply the components of Xp in the coordinate chart (U, φ):
m
X ∂
Xp (f ) = ai (φ(p)) | (f ◦ φ−1 ).
∂xi φ(p)
i=1
for p ∈ U . The formula shows that
m
X ∂
X(f )|U ◦ φ−1 = ai (f ◦ φ−1 ).
∂xi
i=1
That is, in local coordinates X is represented by the vector field
Xm
∂
ai : C ∞ (φ(U )) → C ∞ (φ(U )).
∂xi
i=1

Theorem 7.1. A linear map X : C ∞ (M ) → C ∞ (M ) is a vector field if and only if it is a

derivation of the algebra C ∞ (M ): That is,
X(f1 f2 ) = f2 X(f1 ) + f1 X(f2 )
for all f1 , f2 ∈ C ∞ (M ).
16 CONTENTS

Proof. For all p ∈ M , X defines a tangent vector Xp by Xp (f ) = X(f )p . We have to

show that p 7→ Xp defines a smooth section of T M . Choosing local coordinates (U, φ) around
p, taking p to x = φ(p), the tangent vector X(f )p is represented by a(x) = (a1 (x), . . . , am (x)) ∈
Rm . That is,
m
X ∂
X(f )p (f ) = aj (x) (f ◦ φ−1 ).
∂xi
j=1
Taking for f any function fj with fj ◦ −1 = xj on some open neighborhood V ⊂ φ(U ) of
φ (x)
the given point φ(p), we see that
aj = X(fj ) ◦ φ−1
on V . Since X(fj ) ∈ C ∞ (M ), it follows that aj is smooth. This proves that p 7→ Xp is a
smooth section of T M over U . 
If X, Y are vector fields (viewed as linear maps C ∞ (M ) → C ∞ (M )), the composition X ◦Y
is not a vector field. However, the Lie bracket (commutator)
[X, Y ] := X ◦ Y − Y ◦ X
is a vector field. Indeed, it is easily checked that the right hand side defines a derivation.
Alternatively, the calculation
P can be carried out
Pmin local coordinates (U, φ): One finds that if
XU is represented by m a ∂
i=1 i ∂xi and Y | U by ∂
i=1 i ∂xi then [X, Y ]|U is represented by
b ,
m X
X m
∂bi ∂ai ∂
(aj − bj ) .
∂xj ∂xj ∂xi
i=1 j=1

Let F ∈ C ∞ (M, N ) be a smooth map. Then the collection of tangent maps Tp F : Tp M →

TF (p) N defines a map T F : T M → T N which is easily seen to be smooth. The map T F is an
example of a vector bundle map: It takes fibers to fibers, and the restriction to each fiber is a
linear map. For instance, local trivializations ψ : E|U → U × Rk are vector bundle maps.
Definition 7.2. Let F ∈ C ∞ (M, N ). Two vector fields X ∈ X(M ) and Y ∈ X(N ) are
called F -related if for all p ∈ M , Tp F (Xp ) = YF (p) . One writes X ∼F Y .
For example, if S ⊂ M is an embedded submanifold and ι : S → M is the inclusion, vector
fields X on S and Y on M are ι-related if and only if Y is tangent to S, and X is its restriction.
Theorem 7.3. a) One has X ∼F Y if and only if for all f ∈ C ∞ (N ), X(f ◦ F ) = Y (f ).
b) If X1 ∼F Y1 and X2 ∼F Y2 then [X1 , X2 ] ∼F [Y1 , Y2 ].
Proof. At any p ∈ M , the condition X(f ◦ F ) = F ◦ Y (f ) says that
(Tp F (Xp ))(f ) = Y (f )F (p) = YF (p) (f ).
This proves (a). Part (b) follows from (a):
[X1 , X2 ](f ◦ F ) = X1 (X2 (f ◦ F )) − X2 (X1 (f ◦ F ))
= X1 (Y2 (f ) ◦ F ) − X2 (Y1 (f ) ◦ F )
= Y1 (Y2 (f )) ◦ F − Y2 (Y1 (f )) ◦ F
= [Y1 , Y2 ](f ) ◦ F.

 8. FLOWS OF VECTOR FIELDS 17

Part (b) shows, for instance, that if two vector fields are tangent to a submanifold S ⊂ M
then their bracket is again tangent to S. (Alternatively, one can see this in coordinates, using
submanifold charts for S.)
Problems 7.4. 1. Given an example of vector fields X, Y ∈ X(R3 ) such that X, Y, [X, Y ]
are linearly independent at any point p ∈ R3 . Thus, there is no 2-dimensional submanifold S
with the property that X, Y are tangent to S everywhere.
2. For any n, give an example of vector field X, Y on Rn such that X, Y together with
iterated Lie brackets [X, Y ], [[X, Y ], Y ], . . . span Tp Rn = Rn everywhere.

8. Flows of vector fields

Suppose X is a vector field on a manifold M . A smooth curve γ ∈ C ∞ (J, M ), where J is
an open neighborhood of 0 ∈ R, is called a solution curve to X if for all t ∈ J,
γ̇(t) = Xγ(t) .
In local coordinates (U, φ) around a given point p = γ(t0 ), write
φ ◦ γ(t) = x(t) = (x1 (t), . . . , xn (t))
(defined for t sufficiently close to t0 ). Then
d d X dxi ∂
γ̇(f ) = f (γ(t)) = (f ◦ φ−1 )(x(t)) = (f ◦ φ−1 ) x(t)
.
dt dt dt ∂xi
i
P ∂
Let i ai ∂xi represent X in the chart, that is
X ∂
Xγ(t) (f ) = ai (x(t)) (f ◦ φ−1 ) x(t)
.
∂xi
i

Hence, the equation for a solution curve corresponds to the following equation in local coordi-
nates:
ẋi = ai (x(t)),
for i = 1, . . . , n. This is a first order system of ordinary differential equations (ODE’s). One of
the main results from the theory of ODE’s reads:
Theorem 8.1 (Existence and uniqueness theorem for ODE’s). Let V ⊂ Rm be an open
subset, and a ∈ C ∞ (V, Rm ). For any given x0 ∈ V , there exists an open interval J ⊂ R around
0, and a solution x : J → V of the ODE
dxi
= ai (x(t)).
dt
with initial condition x(0) = x0 . In fact, there is a unique maximal solution of this initial value
problem, defined on some interval Jx0 , such that any other solution is obtained by restriction
to some subinterval J ⊂ Jx0 .
The solution depends smoothly on initial conditions, in the following sense:
18 CONTENTS

Theorem 8.2 (Dependence on initial conditions for ODE’s). Let a ∈ C ∞ (V, Rm ) as above.
For any x0 ∈ V , let Φ(t, x0 ) := γx0 (t) : Jx0 → V be the maximal solution with initial value
γx0 (0) = x0 . Let
[
J = Jx0 × {x0 } ⊂ R × V.
x0 ∈V
Then J is an open neighborhood of {0} × V , and the map Φ : J → V is smooth.
Example 8.3. If V = (0, 1) ⊂ R and a(x) = 1, the solution curves to ẋ = a(x(t)) = 1 with
initial condition x0 ∈ V are x(t) = x0 + t, defined for −x0 < t < 1 − x0 . Thus
J = {(t, x)| 0 < t + x < 1}, Φ(t, x) = t + x
in this case. One can construct a similar example with V = R, where solution curves escape to
1
infinity in finite time: For instance, a(x) = x2 has solution curves, x(t) = − t−c , these escape
2
to infinity for t → c. Similarly, a(x) = 1 + x has solution curves x(t) = tan(t − c), these reach
infinity for t → c ± π2 .
If a = (a1 , . . . , am ) : φ(U ) → Rm corresponds to X in a local chart (U, φ), then any solution
curve x : J → φ(U ) for a defines a solution curve γ(t) = φ−1 (x(t)) for X. The existence and
uniqueness theorem for ODE’s extends to manifolds, as follows:
Theorem 8.4. Let X ∈ X(M ) be a vector field on a manifold M . For any given p ∈ M ,
there exists a solution curve γ : J → M of with initial condition γ(0) = p. Any two solutions
to the same initial value problem agree on their common domain of definition.
Proof. Existence and uniqueness of solutions for small times t follows from the existence
and uniqueness theorem for ODE’s, by considering the vector field in local charts. To prove
uniqueness even for large times t, let γt : J1 → M and γ2 : J2 → M be two solutions to the
IVP. We have to show that γ1 = γ2 on J1 ∩ J2 . Suppose not. Let b > 0 be the infimum of all
t ∈ J1 ∩ J2 with γ1 (t) 6= γ2 (t). If γ1 (b) = γ2 (b), the uniqueness part for solutions of ODE’s, in
a chart around γj (b), would show that the γj (t) coincide for |t − b| sufficiently close to b. This
contradiction shows that γ1 (b) 6= γ2 (b). But then we can choose disjoint open neighborhoods
Uj of γj (b). For |t − b| sufficiently small, γj (t) ∈ Uj . In particular, γ1 (t) 6= γ2 (t) for small |t − b|,
again in contradiction to the definition of b. 
Note that the uniqueness part uses the Hausdorff property in the definition of manifolds.
Indeed, the uniqueness part may fail for non-Hausdorff manifolds.
Example 8.5. A counter-example is the non-Hausdorff manifold Y = R×{1}∪R×{−1}/ ∼,
where ∼ glues two copies of the real line along the strictly negative real axis. Let U± denote
∂
the charts obtained as images of R × {±1}. Let X be the vector field on Y , given by ∂x
in both charts. It is well-defined, since the transition map is just the identity map. Then
γ+ (t) = π(t, 1) and γ− (t) = π(t, −1) are both solution curves, and they agree for negative t
but not for positive t.
Theorem 8.6. Let X ∈ X(M ) be a vector field on a manifold M . For eachSp ∈ M , let γp :
Jp → M be the maximal solution curve with initial value γp (0) = p. Let J = p∈M {p} × Jp ,
and let
Φ : J → M, Φ(t, p) ≡ Φt (p) := γp (t).
 8. FLOWS OF VECTOR FIELDS 19

Then J is an open neighborhood of {0} × M in R × M , and the map Φ is smooth. If

(t1 , Φt2 (p)), (t2 , p) ∈ J then also (t1 + t2 , p) ∈ J , and one has
Φt1 (Φt2 (p)) = Φt1 +t2 (p).
The map Φ is called the flow of the vector field X.
Proof. Define J = ∪p∈M {p}×Jp where Jp is the largest interval around 0 for which there
is a solution curve γp (t) with initial value γp (0) = p. Let Φ(t, p) := γp (t). We first establish
the property Φt1 (Φt2 (p)) = Φt1 +t2 (p). Given (t2 , p) ∈ J X , consider the two curves
γ(t) = Φt (Φt2 (p)), λ(t) = Φt+t2 (p).
By definition of Φ, the curve γ is a solution curve with initial value γ(0) = Φt2 (p), defined for
the set of all t with (t, Φt2 (p)) ∈ J X .
We claim that λ is also a solution curve, hence coincides with γ by uniqueness of solution
curves. We calculate
d d
λ(t) = Φt+t2 (p)
dt dt
d
= |u=t+t2 Φu (p)
du
= XΦu (p) |u=t+t2 = Xλ(t)
It remains to show that J is open and Φ is smooth. We will use the property Φt1 +t2 = Φt2 ◦ Φt1
of the flow to write the flow for large times t as a composition of flows for small times, where
we can use the result for ODE’s in local charts.
Let (t, p) ∈ J , say t > 0. Since the interval [0, t] is compact, we can choose ti > 0 with
t = t1 + t2 + . . . + tN such that the curve Φs (p) stays in a fixed coordinate chart V0 for
0 ≤ s ≤ t1 , the curve Φs (Φt1 (p)) stays in a fixed coordinate chart V1 for 0 ≤ s ≤ t2 , and so on.
Also, let ǫ > 0 be sufficiently small, such that Φs (Φt (p)) is defined and stays in VN := VN −1
for −ǫ ≤ s ≤ ǫ.
Inductively define p0 , . . . , pN by let pk+1 = Φtk+1 (pk ) where p0 = p. Thus pN = Φt (p).
Choose open neighborhoods Uk of pk , with the property that
Φs (U0 ) ⊂ V0 for 0 ≤ s ≤ t1
Φs (U1 ) ⊂ V1 for 0 ≤ s ≤ t2
···
and Φs (UN ) ⊂ VN for −ǫ < s < ǫ. Let U be the set of all points q ∈ M such that
q ∈ U0 , Φt1 (q) ∈ U1 , Φt1 +t2 (q) ∈ U2 , . . . , Φt (q) ∈ UN .
Then U is an open neighborhood of p. The composition Φs+t = Φs ◦ΦtN ◦· · ·◦Φt1 is well-defined
on U , for all −ǫ < s < ǫ. Thus
(t − ǫ, t + ǫ) × U ⊂ J .
The map Φ, restricted to this set, is smooth, since it is written as a composition of smooth
maps:
Φ(t + s, ·) = Φ(s, ΦtN ◦ · · · Φt1 (·)).

20 CONTENTS

Let X be a vector field, and J = J X be the domain of definition for the flow Φ = ΦX .
Definition 8.7. A vector field X ∈ X(M ) is called complete if J X = M × R.
Thus X is complete if and only if all solution curves exist for all time. Above, we had
seen some examples of incomplete vector fields on M = R. In these examples, the vector field
increases “too fast towards infinity”. Conversely, we expect that vector fields X are complete
if they vanish outside a compact set. This is indeed the case. The support supp(X) is defined
to be the smallest closed subset outside of which X is zero. That is,
supp(X) = {p ∈ M |Xp 6= 0}.
Proposition 8.8. Every vector field of compact support is complete. In particular, this is
the case if M is compact.
Proof. By compactness, there exists ǫ > 0 such that the flow for any point p exists for
times |t| ≤ ǫ. But this implies that any integral curve can be extended indefinitely. 
Theorem 8.9. If X is a complete vector field, the flow Φt defines a 1-parameter group of
diffeomorphisms. That is, each Φt is a diffeomorphism and
Φ0 = idM , Φt1 ◦ Φt2 = Φt1 +t2 .
Conversely, if Φt is a 1-parameter group of diffeomorphisms such that the map (t, p) 7→ Φt (p)
is smooth, the equation
d
Xp (f ) = |t=0 f (Φt (p))
dt
defines a smooth vector field on M , with flow Φt .
Proof. It remains to show the second statement. Clearly, Xp is a tangent vector at p ∈ M .
Using local coordinates, one can show that Xp depends smoothly on p, hence it defines a vector
field. Given p ∈ M we have to show that γ(t) = Φt (p) is an integral curve of X. Indeed,
d d d
Φt (p) = |s=0 Φt+s (p) = |s=0 Φs (Φt (p)) = XΦt (p) .
dt ds ds

By a similar argument, one establishes the identity
d ∗ d
Φt (f ) = Φ∗t |s=0 Φ∗s (f ) = Φ∗t X(f )
dt ds
which we will use later on. In fact, this identity may be viewed as a definition of the flow.
Example 8.10. Let X be a complete vector field, with flow Φt . For each t ∈ R, the tangent
map T Φt : T M → T M has the flow property,
T Φt1 ◦ T Φt2 = T (Φt1 ◦ Φt2 ) = T (Φt1 +t2 ),
and the map R × T M → T M, (t, v) 7→ Φt (v) is smooth (since it is just the restriction of the
map T Φ : T (R × M ) → T M to the submanifold R × T M ). Hence, T Φt is a flow on T M , and
b ∈ X(T M ). This is called the tangent lift of X.
therefore corresponds to a vector field X
 9. GEOMETRIC INTERPRETATION OF THE LIE BRACKET 21

Example 8.11. Given A ∈ Matm (R) let Φt : Rm → Rm be multiplication by the matrix

P j
etA = j tj! Aj (exponential map of matrices). Since e(t1 +t2 )A = et1 A et2 A , and since (t, x) 7→
etA x is a smooth map, Φt defines a flow. What is the corresponding vector field X? For any
function f ∈ C ∞ (Rm ) we calculate,
d
Xx (f ) = f (etA x)
dt t=0
X ∂f
= (Ax)j
∂xj
j
X ∂f
= Ajk xk
∂xj
jk
P ∂
showing that X = jk Ajk xk ∂xj .

Problems 8.12. 1. Let X ∈ X(N ), Y ∈ X(M ) be vector fields and F ∈ C ∞ (N, M ) a

smooth map. Show that X ∼F Y if and only if it intertwines the flows ΦX Y
t , Φt : That is,

F ◦ ΦX Y
t = Φt ◦ F.
P ∂
2. Let X be a vector field on U ⊂ M , given in local coordinates by i ai ∂xi . Let
(x1 , . . . , xm , v1 , . . . , vm ) be the corresponding coordinates on T U ⊂ T M . Show that the tangent
lift X̂ is given by
X ∂ X ∂ai ∂
ai + vj
∂xi ∂xj ∂vi
i ij
3. Show that for any vector field X ∈ X(M ) and any x ∈ M with Xx 6= 0, there exists a
local chart around x in which X is given by the constant vector field ∂x∂ 1 . Hint: Show that if S is
an embedded codimension 1 submanifold, with x ∈ S and Xx 6∈ Tx S, the map U × (−ǫ, ǫ) → M
is a diffeomorphisms onto its image, for some open neighborhood U of x in S. Use the time
parameter t and a chart around x ∈ U to define a chart near x.

9. Geometric interpretation of the Lie bracket

If f ∈ C ∞ (N ) and F ∈ C ∞ (M, N ) we define the pull-back F ∗ (f ) = f ◦ F ∈ C ∞ (M ). Thus
pull-back is a linear map,
F ∗ : C ∞ (N ) → C ∞ (M ).
Using pull-backs, the definition of a tangent map reads
Tp F (v) = v ◦ F ∗ : C ∞ (N ) → R.
For instance, the definition of F -related vector fields X ∼F Y can be re-phrased as, X ◦ F ∗ =
F ∗ ◦ Y . For any vector field X ∈ X(N ) and any diffeomorphism F ∈ C ∞ (M, N ), we define
F ∗ X ∈ X(M ) by
F ∗ X(F ∗ f ) = F ∗ (X(f )).
That is,
F ∗ X = F ∗ ◦ X ◦ (F ∗ )−1 .
Lemma 9.1. If X, Y are vector fields on N , F ∗ [X, Y ] = [F ∗ X, F ∗ Y ].
22 CONTENTS

Any complete vector field X ∈ X(M ) with flow Φt gives rise to a family of maps Φ∗t :
X(M ) → X(M ). One defines the Lie derivative LX of a vector field Y ∈ X(M ) by
d
LX (Y ) = Φ∗t Y ∈ X(M ).
dt t=0

The definition of Lie derivative also works for incomplete vector fields, since the definition only
involves derivatives at t = 0.
Theorem 9.2. For any X, Y ∈ X(M ), the Lie derivative LX Y is just the Lie bracket
[X, Y ]. One has the identity
[LX , LY ] = L[X,Y ] .

Proof. Let Φt = ΦX ∞
t be the flow of X. For f ∈ C (M ) we calculate,

d
(LX Y )(f ) = |t=0 (Φ∗t Y )(f )
dt
d
= |t=0 Φ∗t (Y (Φ∗−t (f )))
dt
d
= |t=0 (Φ∗t (Y (f )) − Y (Φ∗t (f ))
dt
= X(Y (f )) − Y (X(f ))
= [X, Y ](f ).
The identity [LX , LY ] = L[X,Y ] just rephrases the Jacobi identity for the Lie bracket. 

Again, let X be a complete vector field with flow Φ. Let us work out the Taylor expansion
of the map Φ∗t at t = 0. That is, for any function f ∈ C ∞ (M ), consider the Taylor expansion
(pointwise, i.e. at any point of M ) of the function
Φ∗t f = f ◦ Φt ∈ C ∞ (M )
around t = 0. We have,
d ∗ d d
Φt f = |s=0 Φ∗t+s f = |s=0 Φ∗t Φ∗s f = Φ∗t X(f ).
dt ds ds
By induction, this shows
dk ∗
Φ f = Φ∗t X k (f ),
dtk t
where X k = X ◦ · · · ◦ X (k times). Hence, the Taylor expansion reads
∞ k
X t
Φ∗t f = X k (f ).
k!
k=0

One often writes the right hand side as exp(tX)(f ). Suppose now that Y is another vector
field, with flow Ψs . In general, Φt ◦ Ψs need not equal Ψs ◦ Φt , that is, the flows need not
commute. Let us compare the Taylor expansions of Φ∗t Ψ∗s f and Ψ∗s Φ∗t f . We have, in second
 10. LIE GROUPS AND LIE ALGEBRAS 23

order,
s2 2
Φ∗t Ψ∗s f = Φ∗t (f + sY (f ) + Y (f ) + · · · )
2
s2 2 t2
= f + sY (f ) + Y (f ) + tX(f ) + stX(Y (f )) + X 2 (f ) + · · ·
2 2
where the dots indicate cubic or higher terms in the expansion. Interchanging the roles of X, Y ,
and subtracting, we find,
(Φ∗t Ψ∗s − Ψ∗s Φ∗t )f = st[X, Y ](f ) + . . .
This shows that [X, Y ] measures the extent to which the flows fail to commute (up to second
order to the Taylor expansion). In fact,
Theorem 9.3. Let X, Y be complete vector fields. Then [X, Y ] = 0 if and only if the flows
of X and Y commute.
Proof. Let Φt be the flow of X and Ψs the flow of Y . Suppose [X, Y ] = 0. Then
d
(Φt )∗ Y = (Φt )∗ LX Y = (Φt )∗ [X, Y ] = 0
dt
for all t. Integrating from 0 to t, this shows (Φt )∗ Y = Y for all t, which means that Y is
Φt -related to itself. It follows that Φt takes the flow Ψs of Y to itself, which is just the desired
equation Φt ◦ Ψs = Ψs ◦ Φt . Conversely, by differentiating the equation Φt ◦ Ψs = Ψs ◦ Φt with
respect to s, t, we find that [X, Y ] = 0. 

10. Lie groups and Lie algebras

10.1. Definition of Lie groups.
Definition 10.1. A Lie group is a group G, equipped with a manifold structure, such that
group multiplication g1 , g2 7→ g1 g2 is a smooth map G × G → G.
Examples of Lie groups include: The general linear group GL(n, R) (invertible matrices in
Matn (R)), the special linear group SL(n, R) (those with determinant 1), the orthogonal group
O(n) and special orthogonal group SO(n), the unitary group U(n) and the special unitary
group SU(n) and the complex general linear or special linear groups GL(n, C) and SL(n, C).
A important (and not very easy) theorem of E. Cartan says that any subgroup H of a Lie
group G that is closed as a subset of G, is in fact an embedded submanifold, and hence is a
Lie group in its own right. Thanks to Cartan, we don’t actually have to check in any of these
examples of matrix groups that they are embedded submanifolds: It is automatic from the fact
that they are groups and closed subsets. Most examples of Lie groups encountered in practice
(for instance, all compact groups) are matrix Lie groups. (An example of a Lie group that is
not isomorphic to a matrix Lie group is the double covering of SL(2, R).) Any a ∈ G defines
two maps la , ra : G → G with
la (g) = ag, ra (g) = ga.
The maps la , ra are called left-translation and right-translation, respectively. They are diffeo-
morphisms of G, with inverse maps la−1 and ra−1 .
24 CONTENTS

Proposition 10.2. For any Lie group, inversion g 7→ g −1 is a smooth map (hence a
diffeomorphism).
Proof. Consider the map F : G × G → G × G, (g, h) 7→ (g, gh). We claim that F is a
diffeomorphism. Once this is shown, smoothness of the inversion map follows since it can be
written as a composition
G −→ G × G −→ G × G −→ G
where the first map is the inclusion g 7→ (g, e), the second maps is F −1 (g, h) = (g, g −1 h), and
the last map is projection to the second factor. Clearly F is a bijection, with inverse map
F −1 (a, b) = (a, a−1 b). To show that F is a diffeomorphism, it suffices to show that all elements
of G × G are regular values of F , i.e. that the tangent map is a bijection everywhere. 5 Let us
calculate the tangent map to F at (g, h) ∈ G × G. Suppose the path γ(t) = (gt , ht ) represents
a vector (v, w) in the tangent space, with g0 = g and h0 = h. To calculate
dγ d d
|t=0 ) = |t=0 F (γ(t)) = |t=0 (gt , gt ht ),
T(g,h) F (v, w) = T(g,h) F (
dt dt dt
we have to calculate the tangent vector to the curve t 7→ gt ht ∈ G. We have
d d d
|t=0 (gt ht ) = |t=0 (ght ) + |t=0 (gt h)
dt dt dt
d d
= Th lg ( |t=0 (ht )) + Tg rh ( |t=0 (gt ))
dt dt
= Th lg (w) + Tg rh (v).
This shows
T(g,h) F (v, w) = (v, Th lg (w) + Tg rh (v))
which is 1-1 and therefore a bijection. 

For matrix Lie groups, smoothness of the inversion map also follows from Cramer’s rule for
the inverse matrix.

10.2. Definition of Lie algebras, the Lie algebra of a Lie group.

Definition 10.3. A Lie algebra is a vector space g, together with a bilinear map [·, ·] :
g × g → g satisfying anti-symmetry
[ξ, η] = −[η, ξ] for all ξ, η ∈ g,
and the Jacobi identity,
[ξ, [η, ζ]] + [η, [ζ, ξ]] + [ζ, [ξ, η]] = 0 for all ξ, η, ζ ∈ g.
The map [·, ·] is called the Lie bracket.

5We are using the following corollary of the regular value theorem: If F ∈ C ∞ (M, N ) has bijective tangent
map at any point p ∈ M , then F restricts to a diffeomorphism from a neighborhood U of p onto F (U ). Thus,
if F is a bijection it must be a diffeomorphism. (Smooth bijections need not be diffeomorphisms in general, the
map F : R → R, t 7→ t3 is a counter-example.)
 10. LIE GROUPS AND LIE ALGEBRAS 25

Any associative algebra is a Lie algebra, with bracket the commutator. The space of vector
fields X(M ) on a manifolds is a Lie algebra, with bracket what we’ve already called the Lie
bracket of vector fields.
For any Lie group G, one defines a Lie algebra structure on the tangent space to the identity
element, g := Te G in the following way. Let X(G)L denote the space of left-invariant vector
fields on G. Thus X ∈ X(G)L if and only if la∗ (X) = X for all a ∈ G. Evaluation at the identity
element gives a linear map
X(G)L → g, X 7→ ξ := Xe .
This map is an isomorphism: Given ξ ∈ g, one defines a left-invariant vector field X by
Xg = Te Lg (ξ). (Exercise: Check that X is indeed smooth!) The Lie bracket of two vector
fields is again left-invariant:
la∗ [X, Y ] = [la∗ X, la∗ Y ] = [X, Y ].
Thus X(G)L is a Lie subalgebra of the Lie algebra of all vector fields on G. Using the isomor-
phism X(G)L ∼ = g, this gives a Lie algebra structure on g. That is, if we denote by X = ξ L the
left-invariant vector field on G generated by ξ, we have,
[ξ L , η L ] = [ξ, η]L .
Problems 10.4. We defined the Lie bracket on g = Te G by its identification with left-
invariant vector fields. A second Lie algebra structure on g is defined by identifying Te G with
the space of right-invariant vector fields. How are the two brackets related? (Answer: One has
[ξ R , η R ] = −[ξ, η]R , so the two brackets differ by sign.)
10.3. Matrix Lie groups. Let G = GL(n, R). Since GL(n, R) is an open subset of the set
MatR (n) of n × n-matrices, all tangent spaces are identified with MatR (n) itself. In particular
g = gl(n, R) ∼
= MatR (n). Let us confirm the obvious guess that the Lie bracket on g is simply
the commutator of matrices. The left-invariant vector field corresponding to ξ ∈ g is
d
ξgL = |t=0 (g exp(tξ)) = gξ
dt
(matrix multiplication). Its action on functions f ∈ C ∞ (G) is,
d X ∂f
ξ L (f )g = |t=0 (g exp(tξ)) = (gξ)ij |g .
dt ∂gij
ij

Hence,
d X ∂f
ξ L η L (f )g = |t=0 (g exp(tξ)η)ij |
dt ∂gij g exp(tξ)
ij
X ∂f
= (gξη)ij | + ...,
∂gij g exp(ξ)
ij

where . . . involves second derivatives of the function f . (When we calculate Lie brackets, the
second derivatives drop out so we need not care about . . ..) We find,
X ∂f
(ξ L η L − η L ξ L )(f )g = (g(ξη − ηξ))ij | .
∂gij g exp(ξ)
ij
26 CONTENTS

Comparing to
X ∂f
[ξ, η]L (f )g = (g([ξ, η]))ij |
∂gij g exp(ξ)
ij
this confirms that the Lie bracket is indeed just the commutator. 6 We obtain similar results
for other matrix Lie groups: For instance, the Lie algebra of O(n) = {A| At A = I} is the space
o(n) = {B| B + B t = 0},
with bracket the commutator, while the Lie algebra of SL(n, R) is
sl(n, R) = {B| tr(B) = 0},
with bracket the commutator. In all such cases, this follows from the result for the general
linear group, once we observe that the exponential map for matrices takes g ⊂ gl(n, R) to the
corresponding subgroup G ⊂ GL(n, R).
10.4. The exponential map for Lie groups. There is an alternative characterization
of the Lie algebra in terms of 1-parameter subgroups. A 1-parameter subgroup of a Lie group
G is a smooth group homomorphism φ : R → G, that is, φ(0) = e and φ(t1 + t2 ) = φ(t1 )φ(t2 ).
For any such φ, the velocity vector at t = 0 defines an element ξ ∈ Te G = g. Let ξ L be the
corresponding left-invariant vector field. Then φ(t) is an integral curve for ξ L :
d d d d L
φ(t) = |s=0 φ(t + s) = |s=0 φ(t)φ(s) = Te lφ(t) |s=0 φ(s) = Te lφ(t) ξ = ξφ(t) .
dt ds ds ds
More generally, a similar calculation shows that for all g ∈ G, the curve γ(t) = gφ(t) is an
integral curve through g. That is, the flow of ξ L is Φ(t, g) = gφ(t).
Suppose conversely that X is a left-invariant vector field. If γ(t) is an integral curve, then
so is its left translate gγ(t) for any g. It follows that X is complete and has a left-invariant flow.
Let φ(t) = Φ(t, e), then φ(t) is a 1-parameter subgroup, and X = ξ L for the corresponding
ξ ∈ g. To summarize, elements of the Lie algebra are in 1-1 correspondence with 1-parameter
subgroups. Let φξ (t) denote the 1-parameter subgroup corresponding to ξ ∈ g.
Definition 10.5. For any Lie group G, with Lie algebra g, one defines the exponential map
exp : g → G, exp(ξ) := φξ (1).
Note that this generalizes the exponential map for matrices. Indeed, suppose G ⊆ GL(n, R)
is a matrix Lie group, with Lie algebra g ⊆ gl(n, R). Then the flow of the left-invariant vector
field corresponding to ξ ∈ gl(n, R) is just Φt (g) = g exp(tξ) (using the exponential map for
matrices).
Theorem 10.6. The exponential map is smooth, and defines a diffeomorphism from some
open neighborhood U of 0 to exp(U ).
Proof. We leave smoothness as an exercise. For the second part, it suffices to show that
the tangent map at 0 is bijective. Since g is a vector space, the tangent space at 0 is identified
with g itself. Note that
φtξ (1) = φξ (t).
6This motivates why we used left-invariant vector fields in the definition of Lie bracket: Otherwise we would
have found minus the commutator at this point.
 11. FROBENIUS’ THEOREM 27

Hence
d d
(T0 exp)(ξ) = |t=0 exp(tξ) = |t=0 φξ (t) = ξ
dt dt
thus T0 exp is simply the identity map. 
For matrix Lie groups, exp coincides with the exponential map for matrices (hence its
name).
10.5. Group actions. Lie groups often arise as transformation groups, by some “action”
on a manifold M .
Definition 10.7. An action of a Lie group G on a manifold M is a group homomorphism
G → Diff(M ), g 7→ Φg such that the action map Φ : G × M → M, (g.p) 7→ Φg (p) is smooth.
Note that an action of G = R is the same thing as a flow. Every matrix Lie group G ⊂
GL(n, R) acts on Rn in the obvious way. Any Lie group G acts on itself by multiplication from
the left g 7→ lg , multiplication from the right g 7→ rg−1 , and also by the adjoint (=conjugation)
action g 7→ lg rg−1 .
Definition 10.8. An action of a finite dimensional Lie algebra g on a manifold M is a Lie
algebra homomorphism g → X(M ), ξ 7→ ξM such that the action map g × M → T M, (ξ, p) 7→
ξM (p) is smooth.
Theorem 10.9. Given an action of a Lie group G on a manifold M , one obtains an action
of the corresponding Lie algebra g, by setting
d
ξM (p) = |t=0 Φ(exp(−tξ)).
dt
The vector field ξM is called the generating vector field corresponding to ξ.
Exercise 10.10. Prove this theorem. Hints: First verify the theorem for the left-action
of a group on itself. (Show that ξM equals −ξ R in this case.) Then, use that the action map
Φ : G × M → M is equivariant, i.e. Φ ◦ (la × id) = Φa ◦ Φ. Finally, show that (−ξ R , 0) ∼Φ ξM .
This implies
(−[ξ, η]R , 0) = ([ξ R , η R ], 0) ∼Φ [ξM , ηM ].
Deduce [ξM , ηM ] = [ξ, η]M .
Note: Many people omit the minus sign in the definition of the generating vector field ξM .
But then ξ 7→ ξM is not a Lie algebra homomorphism but an “anti-homomorphism”. We prefer
to avoid “anti” whenever possible.

11. Frobenius’ theorem

11.1. Submanifolds. We defined embedded submanifolds as subsets of manifolds admit-
ting submanifold charts. One often encounters more general submanifolds, in the following
sense.
Definition 11.1. Let S, M be manifolds of dimensions dim S ≤ dim M . A smooth map
F ∈ C ∞ (S, M ) is an immersion if for all p ∈ S, the tangent map Tp F is 1-1. It is called a
submanifold if in addition F is 1-1.
28 CONTENTS

Locally, the image of any immersion looks like an embedded submanifold, by the following
result:
Theorem 11.2. Let F ∈ C ∞ (S, M ) be an immersion. Then every point p in U has an
open neighborhood U ⊂ S such that F (U ) is an embedded submanifold.
Proof. Using local coordinates, it suffices to prove this for the case that S is an open
subset of Rs and M an open subset of Rm . Given p, we may renumber the coordinates such
that Tp F (Rs ) ∩ Rm−s = {0}, where we view Rm−s as the subspace where the first s coordinates
are 0. Define a map,
Fe : S × Rm−s → Rm , (q, y) 7→ F (q) + y.
It is easily checked that T(p,0) Fe is a bijection. Hence, by the regular value theorem, there
exists an open neighborhood U of p and an open ball Bǫ (0) around 0 ∈ Rm−s such that Fe
restricts to a diffeomorphism from U × Bǫ (0) onto its image in M ⊂ Rm . This gives the desired
submanifold chart. 
Example 11.3. An smooth immersion γ : J → M from an open interval J ⊂ R is the
same thing as a regular curve: For all t ∈ J, γ̇(t) 6= 0.
In general, submanifolds need not be embedded submanifolds: For instance, the integral
curves of a complete vector field define submanifolds R → M , but usually their images are
not embedded. (Note that some authors use “submanifold” to denote embedded submanifolds,
while others use the same terminology for immersions! We follow the conventions from F.
Warner’s book.)
11.2. Integral submanifolds. Let X1 , . . . , Xk be a collection of vector fields on a man-
ifold M such that the Xi are pointwise linearly independent. That is, at every p ∈ M the
values (Xi )p of the vector fields span a k-dimensional subspace of the tangent space Tp M . A
k-dimensional submanifold ι : S ֒→ M is called an integral submanifold for X1 , . . . , Xk , if each
Xj is tangent to S, that is (Xj )ι(p) ∈ Tp ι(Tp S) ⊂ Tι(p) M for all p ∈ S. We had seen above that
the Lie bracket of any two vector fields tangent to S is again tangent to S. Hence, a necessary
condition for the existence of integral submanifolds through every given point p ∈ S is that the
Xj are in involution: That is,
k
X
(3) [Xi , Xj ] = clij Xl
l=1

for some functions clij . Frobenius’ theorem (see below) asserts that this condition is also
sufficient.
Example 11.4. On M = R3 \{x2 = 0} consider the vector fields,
∂ ∂ ∂ ∂
X = x3 − x2 , Z = x1 − x2
∂x2 ∂x3 ∂x2 ∂x1
We have,
∂ ∂
[X, Z] = x1 − x3 =: Y.
∂x3 ∂x1
Using x1 X + x2 Y + x3 Z = 0, we see that X, Z are in involution.
 11. FROBENIUS’ THEOREM 29

As stated, the scope of Frobenius’ theorem is limited since in general, manifolds need not
admit pointwise linearly independent vector fields – often they don’t even admit any vector
field without zeroes. It is convenient to shift attention to the subbundle of T M spanned by
the vector fields, rather than the vector fields themselves:

Definition 11.5. A k-dimensional distribution on a manifold M is a rank k vector sub-

bundle E of the tangent bundle T M . That is, M can be covered by open subsets U ⊂ M
such that over each U , there are k vector fields X1 , . . . , Xk spanning E. The distribution is
called integrable if any such local basis is in involution. An submanifold ι : S → M is called
an integral submanifold for a (possibly non-integrable) distribution E if Tp ι(Tp S) = Eι(p) for
all p ∈ S.

Exercise 11.6. Show P that the condition of being in involution does not depend on the
choice of Xi ’s: If Xi′ = j aij Xj and the Xj are in involution, then so are the Xi′ .

Example 11.7. On M = R3 \{0} consider the three vector fields, X, Y, Z introduced above.
They are pointwise linearly dependent: x1 X +x2 Y +x3 Z = 0. It follows that the vector bundle
E spanned by X, Y, Z has rank 2. The above local calculation shows that E is integrable. The
spheres x21 + x22 + x23 = r2 are integral submanifolds.

11.3. Frobenius’ theorem.

Theorem 11.8 (Frobenius). A rank k distribution E on a manifold M is integrable, if and

only if there exists an integral submanifold through every point p ∈ M . In this case, every point
p ∈ M admits a coordinate neighborhood (U, φ) in which E is spanned by the first k coordinate
vector fields, ∂x∂ 1 , . . . , ∂x∂ k .

Proof. We have seen that if there exists an integral submanifold through every point,
then E must be integrable. Suppose conversely that E is integrable. It suffices to construct
the coordinate charts (U, φ) described in the theorem: In such coordinates, it is clear that the
integral submanifolds are given by setting the coordinates xk+1 , . . . , xm equal to constants.
Choose an arbitrary chart around p, with coordinates y1 , . . . , ym , where p corresponds to
y = 0. Using the chart, we may assume that M is an open subset of Rm . Consider the k-
dimensional subspace Ep = E0 . Renumbering the coordinates if necessary, we may assume
that E0 ∩ Rm−k = {0}, where Rm−k is identified with the subspace of Rm where the first k
coordinates are 0. Passing to a small neighborhood of p = 0 if necessary, we may assume that
Eq ∩ Rm−k = {0} for all q, or equivalently that orthogonal projection from Eq to Rk is an
isomorphism. That is, E is spanned by vector fields of the form,
m
X
∂ ∂
Xi = + air .
∂yi ∂yr
r=k+1

It turns out that we got very lucky: the Xi commute! Indeed, by definition of the Lie bracket
we have,
m
X ajr air ∂
[Xi , Xj ] = ( − ) ,
∂yi ∂yj ∂yr
r=k+1
30 CONTENTS

but since the Xi are in involution, we also have

X k
X m
X k
X
∂
∂
[Xi , Xj ] = clij Xl = clij + cνij aνl .
∂yl ∂yl
l l=1 l=k+1 ν=1
∂
Comparing the coefficients of ∂y l
for l ≤ k, we find that clij = 0, showing that the Xi commute.
Hence their flow Φiti commute (wherever defined). Choose ǫ > 0 sufficiently small, and let U ′
be a small open neighborhood of p such that for all t = (t1 , . . . , tk ) ∈ Bǫ (0) and all q ∈ U the
“joint flow”
Φ(t1 , . . . , tk , q) = Φ1t1 ◦ · · · ◦ Φktk (q)
is defined. Since the flows commute, we have
∂
|s=0 Φ(t1 , . . . , tj + s, . . . , tk , q) = (Xj )Φ(t,q) .
∂s
Define a map
F : Bǫ (0) × (U ′ ∩ Rm−k ) → U, (t, q) 7→ Φt (q).
By construction,
∂
TF( ) = Xj , j = 1, . . . , k
∂tj
and T F ( ∂y∂ j ) = ∂
∂yj ) for j > k. In particular, T(0,0) F is invertible, hence F restricts to a
diffeomorphism on some open subset U ⊂ Bǫ (0) × (U ′ ∩ Rm−k ). The inverse map F −1 gives
the required change of coordinates. 

One has the following addendum to Frobenius’ theorem:

Theorem 11.9. Suppose E ⊂ T M is an integrable distribution. For each p ∈ M , there
is a unique maximal connected integral submanifold ι : S → M passing through p. That is,
if ι′ : S ′ → M is any other integral submanifold through p, then there exists a smooth map
F : S ′ → S such that F is a diffeomorphism onto its image and ι′ = ι ◦ F .
This is analogous to the fact that every vector field has a unique maximal integral curve
through every given point of M . The idea of proof is to “patch together” the local solutions.
Again, the theorem fails for non-Hausdorff manifolds.
The maximal integral submanifolds are called the leaves of the integrable distribution, and
the decomposition of M into leaves is called a foliation.

11.4. Applications to Lie groups. A homomorphism of Lie groups is a smooth group

homomorphism F : H → G. The tangent map at the identity T0 F : h → g is then a
homomorphism of Lie algebras, i.e. takes brackets to brackets. (To see this, one proves that
the left-invariant vector fields corresponding to ξ and to T0 F (ξ) are F -related.) A 1-1 Lie group
homomorphism is called a Lie subgroup, in this case T0 F : h → g is a Lie subalgebra.
Theorem 11.10. Let G be a Lie group, with Lie algebra g, and j : h ⊂ g a Lie subalgebra.
Then there exists a unique Lie subgroup F : H → G having h as its Lie algebra: That is,
j = T0 F .
 11. FROBENIUS’ THEOREM 31

Proof. Let E ⊂ T G be the distribution spanned by the left-invariant vector fields, ξ L

with ξ ∈ h. Since [ξ L , η L ] = [ξ, η]L , this distribution is integrable. Let Lg denote the leaf
through g ∈ G. The distribution is left-invariant: That is, for all a ∈ G the tangent map to
left translation, T la : T G → T G takes E to itself. Hence, for any a ∈ G the left translate
la (Lg ) = Lag is again a leaf. Let H := Le . If h1 , h2 ∈ H we have Lh1 = Lh2 = H, and acting
by a = h−12 we get
Lh−1 h1 = Le = H,
2

proving that h−1

2 h1 ∈ H. This shows that H is a group. Smoothness of the group operations
follows from that for G. 
We next describe an application of Frobenius’ theorem to actions of Lie groups and Lie
algebras on manifolds.
Definition 11.11. An action of a Lie group G on a manifold M is a group homomorphism
G → Diff(M ), g 7→ Φg such that the action map
Φ : G × M → M, (g.p) 7→ Φg (p)
is smooth. An action of a finite dimensional Lie algebra g on a manifold M is a Lie algebra
homomorphism g → X(M ), ξ 7→ ξM such that the action map g × M → T M, (ξ, p) 7→ ξM (p)
is smooth.
Examples 11.12. 1) Note that an action of the (additive) Lie group G = R is the same
thing as a global flow, while an action of the Lie algebra g = R (with zero bracket) is the same
thing as a vector field.
2) Every matrix Lie group G ⊂ GL(n, R), and every matrix Lie algebra acts on Rn by
multiplication.
3) The rotation action of SO(n) on Rn restricts to an action on the sphere, S n−1 ⊂ Rn .
4) Any Lie group G acts on itself by multiplication from the left, la (g) = ag, multiplication
from the right ra−1 (g) = ga−1 , and also by the adjoint (=conjugation) action
Ada (g) := la ra−1 (g) = aga−1 .
The maps
ξ 7→ ξ L , ξ 7→ −ξ R , ξ 7→ ξ L − ξ R
are all Lie algebra actions of g on G.
Theorem 11.13. Given an action of a Lie group G on a manifold M , one obtains an action
of the corresponding Lie algebra g, by setting
d
ξM (p) = |t=0 Φ(exp(−tξ), p).
dt
The vector field ξM is called the generating vector field corresponding to ξ.
Proof. Let us first note that if G acts on manifolds M1 , M2 , and if F : M1 → M2 is a
G-equivariant map, i.e.
F (g.p) = g.F (p) ∀p ∈ M1
then ξM1 ∼F ξM2 . This follows because F takes integral curves for ξM1 to integral curves for
ξM2 . Thus, if we can show [ξ, η]M1 = [ξM1 , ηM1 ]. then a similar property holds for M2 .
32 CONTENTS

We apply this to the special case

F : G × M → M, (g, p) 7→ Φ(g −1 , p)
Φ : G × M → M , with G-acting on G × M by the right-action on G and trivial action on M ,
and on M by the given action. The map is equivariant.
This reduces the problem to the special case M = G with action the right-action a 7→ ra−1
of G on itself. We claim that
ξM = ξ L
in this case. Indeed, the flow g 7→ g exp(−tξ)−1 = g exp(tξ) commutes with left translations,
hence it is the flow of a left invariant vector field. Taking the derivative at g = e, t = 0 we see
that this vector field is ξ L , as claimed. But [ξ L , η L ] = [ξ, η]L . 
Exercise 11.14. Show that the generating vector field for the left action of G on itself is
−ξ R , and the generating vector field for the adjoint action is ξ L − ξ R .
Note: Many people omit the minus sign in the definition of the generating vector field ξM .
But then ξ 7→ ξM is not a Lie algebra homomorphism but an “anti-homomorphism”. We prefer
to avoid “anti” whenever possible.
Let us now consider the inverse problem: Try to integrate a given Lie algebra action to an
action of the corresponding group!
Suppose G is a connected Lie group, with Lie algebra g. We assume that G is also simply
connected: That is, every loop in G can be contracted to a point. For instance, G = SU(n) is
simply connected. If G is a compact Lie group with finite center, one also knows that some
finite cover of G is simply connected.
Theorem 11.15. Every Lie algebra action ξ 7→ ξM of g on a compact manifold M “ex-
ponentiates” uniquely to a Lie group action of the simply connected Lie group G, that is, an
action for which ξM are the generating vector fields.
Sketch of proof. Every G-action on M decomposes G × M into submanifolds Lp =
{(g −1 , g.p)| g ∈ G}, and the action may be recovered from this decomposition. The idea of
proof, given a g-action, is to construct Lp as leafs of a foliation. Let E ⊂ T (G × M ) be the
distribution, of rank equal to dim G, spanned by all vector fields (ξ L , ξM ) ∈ X(G × M ) as ξ
ranges over the Lie algebra. Since
[(ξ L , ξM ), (η L , ηM )] = ([ξ, η]L , [ξ, η]M ),
the distribution is involutive. Hence it defines a foliation of G × M into submanifolds of
dimension dim G.
Given p ∈ M , let Lp ֒→ G × M be the unique leaf containing the point (e, p). Projection
to the first factor induces a smooth map Lp → G, with tangent map taking (ξ L , ξM ) to ξ L .
Since the tangent map is an isomorphism, the map Lp → G is a local diffeomorphism (that
is, every point in Lp has an open neighborhood over which the map is a diffeomorphism onto
its image). We claim that this map is surjective. Proof: By the Lemma given below, and
since the exponential map exp : g → G is a diffeomorphism on some neighborhood of 0, every
g ∈ G can be written as a product g1 . . . gN of elements gj = exp(ξj ) where ξj ∈ g. The curve
t 7→ g1 . . . gj−1 exp(tξj ) is an integral curve of the left-invariant vector field ξjL . Taking all this
curves together defines a piecewise smooth curve γ connecting e to g. This curve lifts to Lp :
 12. RIEMANNIAN METRICS 33

Since M is compact, each ξM is complete, hence each smooth segment of γ to an integral curve
of (ξ L , ξM ).
We have shown at this stage that the map Lp → G is a local diffeomorphism onto its image.
Since G is simply connected by assumption, it follows that the map is in fact a diffeomorphism.
Hence, for every g, Lp contains a unique point of the form (g −1 , p′ ). Define g.p = Φ(g, p) := p′ .
We leave it as an exercise to check that this map defines a smooth G-action. 
Lemma 11.16. Let G be a connected Lie group, and U ⊂ G an open neighborhood of the
group unit e ∈ G. Then every g ∈ G can be written as a finite product g = g1 · · · gN of elements
gj ∈ U .
Proof. We may assume
S∞ that g −1 ∈ U whenever g ∈ U . For each N , let U N = {g1 · · · gN | gj ∈
U }. WeShave to show N =0 U = N G. Each U N is open,Shence their union is open as well. If
∞ N , then gU ∈ G\
S ∞ N (for if gh ∈ ∞ N with h ∈ U we would have
g ∈ G\ N =0 US N =0 U S N =0 U
−1 ∞ N ∞ N
g = (gh)h ∈ N =0 U .) This showsSthat G\ N =0 U is also open. Since G is connected,
it follows that the open and closed set ∞ N =0 U
N is all of G. 

12. Riemannian metrics

Let us quickly recall some linear algebra. A bilinear form on a vector space V is a bilinear
map g : V × V → R. Such a bilinear form is called symmetric if g(v, w) = g(w, v) for all
v, w, and in this case is completely determined by the associated quadratic form q(v) = g(v, v).
g is called an inner product if it is positive definite, i.e. g(v, v) > 0 for all v ∈ V . More
generally, a symmetric form g is called non-degenerate if g(v, w) = 0 for all w implies v = 0.
Non-degenerate symmetric bilinear forms are also called indefinite inner products.
Given a basis e1 , . . . , en of V , one can describe any bilinear form in terms of the matrix
gij = g(ei , ej ). The bilinear form g is symmetric if and only if the matrix gij is symmetric, and
in this case one can always choose the basis such that gij is diagonal. In fact, one can choose
the basis in such a way that only +1, 0, −1 arise as diagonal entries. Let d+ , d0 , d− the number
of +1, 0, −1 diagonal entries. Then g is non-degenerate if d0 = 0, and is an inner product if
and only if d0 = d− = 0, i.e. if there exists a basis such that gij = δij .
Exercise 12.1. Show that one can split V = V+ ⊕ V− where dim V± = d± and g is positive
definite on V+ , negative definite on V− . However, looking at the case (d+ , d− ) = (1, 1), observe
that this splitting is not unique.
Definition 12.2. A Riemannian metric on a manifold M is a family of inner products
gp : Tp M × Tp M → R, depending smoothly on p in the sense that the quadratic form
q : T M → R, v 7→ gp (v, v) for v ∈ Tp M
is a smooth map q ∈ C ∞ (T M ). More generally, a pseudo-Riemannian metric of signature
(d+ , d− ) is defined by letting the gp be indefinite inner products of signature (d+ , d− ).
The case of signature (3, 1) is relevant to general relativity, with 3 space dimensions and 1
time dimension. Again, there is no distinguished splitting into “space” and “time” directions.
Lemma 12.3. Any pseudo-Riemannian metric defines a symmetric C ∞ (M )-bilinear map
g : X(M ) × X(M ) → C ∞ (M ), g(X, Y )p = gp (Xp , Yp ).
34 CONTENTS

Conversely, every symmetric C ∞ (M )-bilinear map g : X(M ) × X(M ) → C ∞ (M ), with the

property that g(X, Y )p = 0 for all Y implies Xp = 0, defines a pseudo-Riemannian metric.
Proof. Let g be a pseudo-Riemannian metric, with quadratic form q. View vector fields
as smooth sections, X ∈ Γ∞ (M, T M ). Then
g(X, Y ) = 21 (g(X + Y, X + Y ) − g(X, X) − g(Y, Y )) = 12 (q ◦ (X + Y ) − q ◦ X − q ◦ Y )
is smooth, while C ∞ (M )-bilinearity is obvious. Conversely, suppose we are given a C ∞ (M )-
bilinear map g : X(M ) × X(M ) → C ∞ (M ) with the property that g(X, Y )p = 0 for all Y
implies Xp = 0. The following Lemma shows that g(X, Y )p depends only on Xp , Yp . Hence we
can define
gp (Xp , Yp ) := g(X, Y )p .
If gp (v, w) = 0 for all v, choose Y with Yp = w. Then g(X, Y )p = g(Xp , Yp ) = 0 for all
X, Y , which by assumption implies Yp = 0. Hence gp is non-degenerate. Using the formula
q ◦ X = g(X, X), and passing to local coordinates, one sees that gp depends smoothly on p,
hence it defines a Riemannian metric. 
Lemma 12.4. If A : X(M ) × · · · × X(M ) → C ∞ (M ) is a C ∞ (M )-multilinear map, then
the value of A(X1 , . . . , Xr ) at p ∈ M depends only on (X1 )p , . . . , (Xr )p . More generally, this
Lemma holds true for any C ∞ (M )-multilinear map from X(M ) × · · · × X(M ) to a C ∞ (M )-
module.
Proof. It suffices to consider the case r = 1. We have to show that if X vanishes at p,
then A(X) vanishes
P at p. But if Xp = 0, we can write (using local coordinates, and the Taylor
∞
expansion) X = i fi Xi where Xi ∈ X(M ) and where fi ∈ C (M ) vanish at p. Hence,
X X
A(X)p = A( fi Xi )p = fi (p)A(Xi )p = 0,
i i

by C ∞ (M )-linearity. 
Definition 12.5. A (pseudo)-Riemannian manifold (M, g) is a manifold M together with
a (pseudo)-Riemannian metric. An isometry between (pseudo)-Riemannian manifold (M1 , g1 )
and (M2 , g2 ) is a diffeomorphism F : M1 → M2 such that for all p ∈ M1 , the tangent map
Tp F : Tp M1 → TF (p) M2 is an isometry, i.e. preserves inner products.
In local coordinates x1 , . . . , xm on U ⊂ M , any pseudo-Riemannian metric is determined
by smooth functions
∂ ∂
gij (x) = g( , ).
∂xi ∂xj
Indeed, one recovers g from the gij by
X ∂ X ∂ X
g( ai , bj )= gij ai bi .
∂xi ∂xj
i j ij

Conversely, every collection of smooth functions gij , such that each (gij (x)) is a non-degenerate
symmetric bilinear form, defines a Riemannian metric. In particular, to gij = δij defines the
 13. EXISTENCE OF RIEMANNIAN METRICS 35

standard metric on Rn . How does gij depend on the choice of coordinates? Let y = φ(x) be a
coordinate change, and let g̃ij (y) denote the matrix in y-coordinates. We have,
∂ X ∂xa ∂
= .
∂yi a
∂y i ∂x a

Hence,
∂ ∂ X ∂xa ∂xb
g̃ij (y) = g( , )= gab (φ(y))
∂yi ∂yj ∂yi ∂yj
ab

Lemma 12.6. Let S ⊂ M be an embedded submanifold, and g a Riemannian metric on M .

Then the restriction of g to the tangent spaces Tp S ⊂ Tp M defines a Riemannian metric on
M . More generally, if ι : S → M is an immersion, there is a unique Riemannian metric on S
such that each tangent map Tp ι : Tp S → Tι(p) M is an isometry onto its image.
In particular, every embedded submanifold of Rm inherits a Riemannian metric from the
standard Riemannian metric on Rm .
Example 12.7. The 2-torus T 2 can be defined as a direct product of the circle S 1 ⊂ R2
with itself. Correspondingly we have an embedding T 2 → R4 and the corresponding induced
metric g on T 2 . The resulting metric on T 2 is simply the product of the metrics on the S 1
factors, and in particular is flat: T 2 is locally isometric to R2 . It follows that there is no
embedding of T 2 into R3 , inducing the same metric g: We had seen in our curves and surfaces
course that any connected surface in R3 with vanishing first fundamental form is an open subset
of a plane, hence cannot be a compact surface.

13. Existence of Riemannian metrics

To show that every manifold admits a Riemannian metric, we need an important technical
tool called partitions of unity.
Theorem 13.1 (Partitions of unity). Let M be a manifold.
a) Any open cover {Uα } of M has a locally finite refinement {Vβ }: That is, {Vβ } is an
open cover, each Vβ is contained in some Uα , and the cover is locally finite in the sense that
each point in M has an open neighborhood meeting only finitely many Vβ ’s.
b) For any locally finite cover Uα of M , there exists a partition of unity, that is a collection
of functions χα ∈ C ∞ (M ) with supp(χα ) ⊂ Uα , such that 0 ≤ χα ≤ 1 and
X
χα = 1.
α
P
Note that the sum α χα is well-defined, since only finitely many χα ’s are non-zero near any
given point. We will omit the somewhat technical proof of this result. The proof is contained
in most books on differential geometry (e.g. Helgason), and can also be found in the lecture
notes from my “manifolds” course. The main steps for part (b) are as follows:
(i) One constructs a “shrinking” of the open cover Uα to a new cover Vα , such that
V α ⊂ Uα . The new cover is still locally finite.
(ii) One constructs functions
P fα ∈ C ∞ (M ) supported on Uα , such that fα > 0 on Vα ,
(iii) One defines f = α fα > 0, and sets χα = fα /f .
36 CONTENTS

Theorem 13.2. Every manifold M admits a Riemannian metric.

Proof. Choose an atlas {(Uα , φα )} of M . Passing to a refinement, we may assume that
the atlas is locally finite. Choose a partition of unity χα for the cover {Uα }. Since φα identifies
Uα with an open subset of Rm , we obtain Riemannian metrics P gα on Uα from the standard
m
Riemannian metrics on R . For all p ∈ M , the sum gp = χα (p)(gα )p is well-defined. Since
all χα (p) ≥ 0, with at least one strictly positive, gp is an inner product with clearly a smooth
dependence on p. Thus g is a Riemannian metric on M . 
It is not true that every manifold admits a pseudo-Riemannian metric of given signature
(d+ , d− ), where both d± 6= 0.

14. Length of curves

Suppose (M, g) is a Riemannian manifold (that is, a manifold with a Riemannian metric).7
For any tangent vector v ∈ Tp M , we define its length as ||v|| = gp (v, v)1/2 .
Definition 14.1. Let γ : [a, b] → M be a smooth curve in M .8 One defines the length of
γ to be the integral
Z b
L(γ) = ||γ̇(t)|| dt.
a
The length functional is invariant under reparametrizations of the curve γ. Somewhat more
generally, we have:
Proposition 14.2. Let σ : [a, b] → R be a smooth function, with the property that σ(t1 ) ≤
σ(t2 ) for t1 ≤ t2 . Let γ : [a, b] → M be a smooth curve of the form γ = γ̃ ◦ σ. Then
L(γ) = L(γ̃).
Proof. By substitution of variables t̃ = σ(t), 9
Z b
d
L(γ) = || (γ̃ ◦ σ)||dt
a dt
Z b
dγ̃ dσ
= || (σ(t))|| | | dt
a dt̃ dt
Z σ(b)
dγ̃
= || ||dt̃
σ(a) dt̃
= L(γ̃)

The definition of L(γ) applies to piecewise smooth curves: That is, continuous curves
γ : [a, b] → M such that there exists a subdivision a = t0 ≤ · · · ≤ tN = b of the interval, with
each γ|[ti ,ti+1 ] a smooth curve.
7For the following discussion, see chapter 1.4 in Jost’s book.
8Here smooth means that γ extends to a smooth curve on an open interval J containing [a, b].
9If σ : [a, b] → R is a continuous, piecewise smooth map, which is weakly increasing in the sense that
Rb R σ(b)
σ(t1 ) ≤ σ(t2 ) for t1 ≤ t2 , then a
f (σ(t))| dσ
dt
|dt = σ(a)
f (t̃)dt̃.
 14. LENGTH OF CURVES 37

If the curve γ is contained in a fixed coordinate chart (U, φ), and (x1 (t), . . . , xm (t)) describes
the curve in local coordinates, we have
Z b sX
L(γ) = gij (x(t)) ẋi x˙j dt.
a ij

Definition 14.3 (Distance function). Let (M, g) be a connected Riemannian manifold.

For p, q ∈ M , the distance d(p, q) between any two points on M is infimum of L(γ), as γ varies
over all piecewise smooth curves γ : [0, 1] → M with γ(0) = p and γ(1) = q. (If no such path
exists, we set d(p, q) = ∞.)
Problems 14.4. 1. Show that for any manifold M , the following are equivalent: (i) M is
connected, (ii) any two points p, q can be joined by a continuous path, (iii) any two points p, q
can be joined by a piecewise smooth path, (iv) any two points p, q can be joined by a smooth
path. Hence d(p, q) < ∞ for a connected manifold.
2. Show that in the definition of distance function, one can replace piecewise smooth paths
by smooth paths. In fact, any piecewise smooth path is of the form γ = λ ◦ σ, where σ is
weakly increasing and piecewise smooth, and λ is smooth.
Lemma 14.5. Let (U, φ) be a coordinate chart in which g is given by gij (x), and K ⊂ φ(U )
a compact subset. Then there exist λ ≥ µ > 0 with
sX sX sX
(4) µ ξi ξi ≥ gij (x)ξi ξj ≥ λ ξi ξi .
i ij i

for x ∈ K, ξ ∈ Rn .
2n
P
P The set of all (x, ξ) ∈ R with x ∈ K and i ξi ξi = 1 is compact. Hence the
Proof.
function ij gij (x)ξi ξj takes on its maximum λ and minimum µ on this set. By definition of
a Riemannian metric, µ > 0. 
Theorem 14.6. For any connected manifold M , the distance function d defines a metric
on M . That is, d(p, q) ≥ 0 with equality if and only if p = q, and for any three points p, q, r,
one has the triangle inequality
d(p, q) + d(q, r) ≥ d(p, r).
Proof. The triangle inequality is immediate from the definition. Suppose p 6= q. We have
to show d(p, q) > 0. Choose a chart (U, φ) around p, with φ(p) = 0, and let ǫ > 0 be sufficiently
small, such that the closed ball Bǫ is contained in φ(U ) and φ−1 (Bǫ ) does not contain q. Let
gij represent the metric in the chart (U, φ).
Given a curve γ from p to q, let t1 < b is such that γ(t) ∈ φ(U ) for a ≤ t ≤ t1 and
γ(t1 ) ∈ B ǫ \Bǫ . Write φ(γ(t)) = x(t) for a ≤ t ≤ t1 . Using the Lemma,
Z t1 sX Z t1 sX
L(γ) ≥ gij (x(t))ẋi ẋj dt ≥ λ ẋi ẋi dt ≥ λǫ,
a ij a i

since the length of the path from φ(p) = 0 to x(t1 ) must be at least the Euclidean distance ǫ.
Hence also
d(p, q) = inf γ L(γ) ≥ λǫ > 0.
38 CONTENTS


Theorem 14.7. For any manifold M , the topology defined by the metric coincides with the
manifold topology.
Proof. This follows from the Lemma: In local charts, ǫ-balls for the metric d contain
sufficiently small Euclidean δ-balls, and vice versa. 

15. Connections and parallel transport

In this section, we will define the parallel transport of tangent vectors on any Riemannian
manifold (M, g). If M ⊂ Rm is an embedded submanifold of Rm , with metric induced from
Rm , we can follow the strategy from the “curves and surfaces” course: At any p ∈ M we have
an orthogonal projection Πp : Rm → Tp M . If γ(t) is a curve and X(t) ∈ Tγ(t) M a vector field
along γ, we say that X is parallel along γ if the covariant derivative
∇X dX
:= Πγ(t)
dt dt
vanishes for all t. Here we have used that X(t) can be viewed as an Rm -valued function of t.
Using the existence and uniqueness theorem for ODE’s, one finds that any parallel vector field
along γ is determined by its value X(t0 ) at any fixed time t0 .
For a general Riemannian manifold (M, g), we don’t have “orthogonal projection” at our
disposal. It is remarkable that there exists, nevertheless, a well-defined concept of parallel
transport on any Riemannian manifold (M, g). That is, parallel transport is really an intrinsic
property. Our starting point for defining parallel transport is to define generalized covariant
derivatives, called affine connections. We will then show that any Riemannian manifold carries
a distinguished affine connection.

15.1. Affine connections. Let M be a manifold.

Definition 15.1. An affine connection on M is a bi-linear map
∇ : X(M ) × X(M ) → X(M ), (X, Y ) → ∇X (Y )
such that
∇X (f Y ) = f ∇X (Y ) + X(f )Y
∇f X (Y ) = f ∇X (Y ).
for all f ∈ C ∞ (M ), X, Y ∈ X(M ).
The second condition says that the operator ∇X is C ∞ -linear in the X variable. One calls
∇X (Y ) the covariant derivative of Y in the direction of X. Y is called covariant constant in
the direction of X if ∇X (Y ) = 0.
If M = U is an open subset of Rm , any affine connection ∇ is determined by its values on
coordinate vector fields. The functions Γijk ∈ C ∞ (U ) defined by
∂ X ∂
(5) ∇ ∂ ( )= Γijk
∂xj ∂xk ∂xi
i
 15. CONNECTIONS AND PARALLEL TRANSPORT 39

are called the Christoffel symbols of ∇. The full connection is given in terms of Christoffel
symbols by the formula,
 
X ∂ X X ∂bk ∂ X ∂ 
∇P a j ∂ ( bk )= aj  + Γijk bk .
j ∂xj ∂xk ∂xj ∂xk ∂xi
k j k k,i

Conversely, it is easily checked that any collection of smooth functions Γijk defines an affine
connection by this formula. In particular, open subsets U ⊂ Rm have the standard affine
connection, given by Γijk = 0.
More generally, for affine connections on manifolds one defines Christoffel symbols of a
connection with respect to a given chart. First we note that if U ⊂ M is an open subset, affine
connections ∇ have a unique restriction ∇|U with the property
(∇|U )X|U (Y |U ) = ∇X (Y )|U .
Moreover, every connection is determined by its restrictions to elements of an open cover of
M . Hence we may define:
Definition 15.2. Let ∇ be an affine connection on a manifold M . If (U, φ) is a chart,
defining local coordinates x1 , . . . , xm , one defines the Christoffel symbols Γijk of ∇|U in the
given chart to be the functions defined by (5).
Problems 15.3. 1. Calculate the Christoffel symbols of the standard connection on R2 in
polar coordinates. The solution shows that Christoffel symbols may vanish in one coordinate
system but be non-zero in another. 2. Work out the transformation property of Christoffel
symbols under change of coordinates.
Proposition 15.4. For any affine connection ∇ on M , the map T : X(M ) × X(M ) →
X(M ) given by
T (X, Y ) = ∇X (Y ) − ∇Y (X) − [X, Y ]
is C ∞ (M )-linear in both X and Y . It is called the torsion of ∇.
Proof. For all f ∈ C ∞ (M ),
T (X, f Y ) − f T (X, Y ) = ∇X (f Y ) − f ∇X (Y ) − [X, f Y ] + f [X, Y ]
= X(f ) Y − X(f )Y = 0.
Similarly T (f X, Y ) − f T (X, Y ) = 0. 

In local coordinates we have, in terms of the Christoffel symbols,

∂ ∂ X ∂
T( , )= (Γijk − Γikj ) .
∂xj ∂xk ∂xi
jk

Hence, the connection is torsion-free if and only if the Christoffel symbols Γijk are symmetric
in j, k. In particular, if the Christoffel symbols have this symmetry property in one system of
coordinates, then also in every other system.
40 CONTENTS

15.2. The Levi-Civita connection.

Proposition 15.5. Let (M, g) be a (pseudo-)Riemannian manifold. For any affine con-
nection ∇ on M , and any Z ∈ X(M ) the map ∇Z g : X(M ) × X(M ) → C ∞ (M ) given by
(∇Z g)(X, Y ) = Z g(X, Y ) − g(∇Z (X), Y ) − g(X, ∇Z (Y ))
is C ∞ (M )-linear
in both X, Y . It is called the covariant derivative of g in the direction of Z.
The connection ∇ is called a metric connection if ∇X g = 0.
The proof is straightforward. In local coordinates and the corresponding Christoffel symbols
for ∇, we have
∂ ∂ ∂gij X l
(∇g)kij := (∇ ∂ g)( , )= − Γki glj − Γlkj gli .
∂xk ∂xi ∂xj ∂xk
l
and ∇ is a metric connection if and only of the right hand side vanishes.
Theorem 15.6 (Fundamental Theorem of Riemannian Geometry). Suppose (M, g) is a
pseudo-Riemannian manifold. There exists a unique torsion-free metric connection ∇ on M .
It is called the Levi-Civita connection.
Proof. Suppose ∇ is a torsion-free metric connection. Since ∇ is metric, we have
Z g(X, Y ) = g(∇Z (X), Y ) + g(X, ∇Z (Y )).
Using the torsion free condition ∇Z (X) = ∇X (Y ) + [Z, X] this gives,
Z g(X, Y ) = g(Y, ∇X (Z)) + g(X, ∇Z (Y )) + g([Z, X], Y ).
Permuting letters we also have
X g(Y, Z) = g(Z, ∇Y (X)) + g(Y, ∇X (Z)) + g([X, Y ], Z),
Y g(Z, X)) = g(X, ∇Z (Y )) + g(Z, ∇Y (X)) + g([Y, Z], X)
Use these equations to eliminate ∇X and ∇Y , and obtain
Z g(X, Y )−X g(Y, Z)+Y g(Z, X)) = 2g(X, ∇Z (Y ))+g([Z, X], Y )−g([X, Y ], Z)+g([Y, Z], X),
that is,
2g(X, ∇Z (Y ))
(6) = Z g(X, Y ) − X g(Y, Z) + Y g(Z, X)) − g([Z, X], Y ) + g([X, Y ], Z) − g([Y, Z], X).
Since g is non-degenerate, any vector field W is completely determined by its parings g(X, W )
with all vector fields X. In particular, (6) specifies the vector field W = ∇Z (Y ). This shows
that a torsion-free metric connection ∇ is determined by the metric g. Conversely, it is straight-
forward to check that formula (6) defines a torsion-free metric connection. For instance, if we
replace Y by f Y for some function f , we find
2(g(X, ∇Z (f Y )) − f g(X, ∇Z (Y ))
= Z(f ) g(X, Y ) − X(f ) g(Y, Z) + g(X(f )Y, Z) + g(Z(f )Y, X)
= 2 Z(f ) g(X, Y )
which shows that ∇Z (f Y )−f ∇Z (Y ) = Z(f ) Y . The other properties are checked similarly. 
 15. CONNECTIONS AND PARALLEL TRANSPORT 41

Exercise 15.7. Try to re-derive the explicit formula (6) for the Levi-Civita connection
without looking at the notes. Fill in the details of showing that this formula defines a torsion-
free metric connection.
Corollary 15.8. Every manifold admits a torsion-free affine connection ∇.
Proof. We have seen that every manifold admits a Riemannian metric g. Thus one can
take the Levi-Civita connection with respect to g. 
∂ ∂ ∂
Taking X = ∂xl , Y =
∂xk , Z = ∂xj to be coordinate vector fields in (6), we obtain a formula
for the Christoffel symbols Γijk of the Levi-Civita connection:
X ∂gkl ∂gjl ∂gjk
2 Γijk gil = + − .
∂xj ∂xk ∂xl
i
Letting (g −1 )ij denote the inverse matrix to gij , this gives:
Theorem 15.9. In local coordinates, the Christoffel symbols for the Levi-Civita connection
are given by  
i 1 X −1 ∂gkl ∂gjl ∂gjk
Γjk = (g )il + − .
2 ∂xj ∂xk ∂xl
l

We had seen a similar formula in the curves and surfaces course. In fact, we could have
used this formula to define a connection in local coordinates, and then check that the local
definitions patch together. However, the significance of this rather complicated formula would
remain obscure from such an approach.
It is immediate from this formula that ∇ is torsion-free, since the Christoffel symbols are
symmetric in j, k.
15.3. Parallel transport. Let ∇X (Y ) be an affine connection on a manifold M . Since
∇X (Y ) is C ∞ -linear in the X-variable, the value of ∇X (Y ) at p depend only on Xp . Thus if
v ∈ Tp M one can define ∇v (Y ) ∈ Tp M by ∇v (Y ) := ∇X (Y )p where X is any vector field with
Xp = v. If γ : J → M is any curve, one can therefore define
∇γ̇(t) Y ∈ Tγ(t) M.
P ∂
If x(t) is the description of the curve γ in local coordinates x1 , . . . , xm , so that γ̇ = i ẋi ∂xi ,
P
and Y = k bk ∂x∂ k ,
!
X ∂bi X i ∂
∇γ̇(t) Y = ẋj + Γjk bk .
∂xj ∂xi
ij k
 
X dbi X ∂
=  + Γijk x˙j bk  .
dt ∂xi
i jk

Here db d
dt = dt bi (x(t)). Note that this formula depends only on the “restriction” of Y to γ,
i

or more precisely on the section of the pull-back bundle γ ∗ (T M ) → J defined by Y . In fact,

∗
Pof γ (T ∂M ) → J. In
the formula makes sense for any vector field along γ, that is, any section
local coordinates, vector fields along γ are given by expressions Y = i bk (t) ∂xk ∈ Tγ(t) M
42 CONTENTS

depending smoothly on t, and the above formula in local coordinates defines a new vector field
along γ,10
DY
≡ ∇γ̇(t) Y.
dt
Definition 15.10. A vector field Y along a curve γ : J → M , is called parallel along γ if
the covariant derivative DY
dt vanishes everywhere.
Theorem 15.11. Let ∇ be a metric connection on a manifold M . Let γ : J → M be
a smooth curve, X0 ∈ Tγ(t0 ) M where t0 ∈ J. Then there is a unique parallel vector field
X(t) ∈ Tγ(t) along γ, with the property X(t0 ) = X0 . The linear map
Tγ(t0 ) M → Tγ(t) M, X0 7→ X(t)
is called parallel transport along γ, with respect to the connection ∇.
Proof. In local coordinates as above, parallel vector fields are the solutions of the first
order ordinary differential equations,
dbi X i
+ Γjk x˙j bk = 0.
dt
jk
Hence, for “short times” the theorem follows from the existence and uniqueness theorem for
ODE’s, and for “long times” by patching together local solutions. 
Proposition 15.12. Let (M, g) be a pseudo-Riemannian manifold, and ∇ an affine con-
nection on M . Then
d DX DY
g(X(t), Y (t)) = (∇γ̇ g)(X(t), Y (t)) + g( , Y ) + g(X, ).
dt dt dt
P ∂ P
Proof. In local coordinates, write X(t) = i ai (t) ∂x i
and Y (t) = j bj (t) ∂x∂ j , and let
x(t) = (x1 (t), . . . , xm (t)) be the coordinate expression for the curve γ. Then
d d X
g(X(t), Y (t)) = gij ai bj
dt dt
ij
X ∂gij X X
= ẋk + gij ȧi bj + gij ai ḃj
∂xk
ijk ij ij

and
DX X X
g( ,Y ) = (ȧi + Γilm ẋl am )bj ,
dt
ij lm
DY X X
g(X, ) = ai (ḃj + Γjlm ẋl bm ).
dt
ij lm

10Of course, it would be better to give a coordinate free definition. For this, one has to generalize the notion
of an affine connection, and introduce connections ∇ on vector bundles E → M . For any X ∈ X(M ), ∇X is an
endomorphism of the space of sections Γ∞ (E). For any smooth map F : N → M one then obtains a connection
F ∗ ∇ on the pull-back bundle F ∗ E. In our case, we obtain a connection γ ∗ ∇ on γ ∗ (T M ). One then defines
DY
:= (γ ∗ ∇) ∂ Y (t).
dt ∂t
 16. GEODESICS 43

Taking this three equations together, and using

∂ ∂ ∂gij X l
(∇g)kij = (∇ ∂ g)( , )= − Γki glj − Γlkj gli ,
∂xk ∂xi ∂xj ∂xk
l

the Proposition follows.11 

As an immediate consequence, we have:
Proposition 15.13. An affine connection ∇ on a pseudo-Riemannian manifold (M, g) is
a metric connection if and only if parallel transport along curves preserves inner products.

16. Geodesics
Let ∇ be an affine connection on a manifold M .
Definition 16.1. A smooth curve γ : J → M is called a geodesic for the connection ∇, if
and only if the velocity vector field γ̇ is parallel along γ.
Exercise 16.2. Show that if γ : J → M is a geodesic, and φ : J˜ → J is a diffeomorphism
(change of parameters), then
γ̃(t̃) = γ(φ(t̃))
dφ
is a geodesic if and only if = const, i.e. if and only if φ(t̃) = at̃ + b for some a 6= 0, b.
dt̃
P ∂
As a special case of the differential equation for a parallel vector field X(t) = i bi (t) ∂x i
,
here X(t) = γ̇ i.e. bi = ẋi , we find:
Theorem 16.3. In local coordinates, geodesics are the solutions of the second order ordinary
differential equation,
d2 xi X i
+ Γjk ẋj ẋk = 0.
dt2
jk

Notice that only the symmetric part Γijk +Γikj , that is the torsion-free part of ∇, contributes
to the geodesic equation. Thus, if one is interested in the geodesic flow of a metric connection
∇, one might as well assume that ∇ is the Levi-Civita connection. On Rm with the standard
Riemannian metric, geodesics are straight lines with constant speed parametrization.
It is a standard trick in ODE theory to reduce higher order ODE’s to a system of first
order ODE’s, by introducing derivatives as parameters. In our case, if we introduce ẋi =: ξi ,
the geodesic equation becomes a system,
dxi
= ξi
dt
dξi X
= − Γijk ξj ξk .
dt
jk

11We had to resort to this terrible proof since we defined the covariant derivative along curves in coordinates
only. In the coordinate free definition, the Proposition is almost a triviality because it is essentially just the
definition of ∇g!
44 CONTENTS

Notice that xi , ξi are just the standard local coordinates on T M induced by the local coordinates
xi on M . Hence, the above first order system defines a vector field S on T M , given in local
coordinates by
X ∂ X ∂
S= ξi − Γijk ξj ξk
∂xi ∂ξi
i ijk

Definition 16.4. The vector field S is called the geodesic spray of ∇, and its flow is called
the geodesic flow.
Theorem 16.5. For any p ∈ M, v ∈ Tp M there exists a unique maximal geodesic γv : J →
M , where γv (0) = p, γ̇v (0) = v.
Proof. Let Φt denote the geodesic flow, and π : T M → M the base point projection.
The geodesics on M are just the projections of solution curves of the geodesic spray S. In
particular, γv is given by
γv (t) = π(Φt (v)).

Notice that the geodesic flow has the property
d
(π(Φt (v))) = Φt (v).
dt
This is the coordinate free reformulation of ẋi = ξi . Furthermore, it has the property
Φt (av) = Φat (v)
for a ∈ R; this just says that if γ(t) is a geodesic, with γ̇(0) = v, then t 7→ γ(at) is also a
geodesic, but with initial velocity av.
Exercise 16.6. Show that every non-constant geodesic is regular, i.e. γ̇ 6= 0 everywhere.
Definition 16.7. The manifold M with affine connection ∇ is called geodesically complete
if the geodesics spray is a complete vector field. A (pseudo-)Riemannian manifold (M, g) is
called geodesically complete if it is geodesically complete for the Levi-Civita connection.
Thus geodesic completeness means that all geodesics exist for all time.
The property γav (t) = γv (at) for all a ∈ R is reminiscent of a property of 1-parameter
subgroups of Lie groups. Similar to the Lie groups case we define:
Definition 16.8. Suppose (M, ∇) is geodesically complete. The map
Expp : Tp M → M, v 7→ γv (1)
is called the exponential map based at p.
Compare with the very similar definition of exponential maps for Lie groups – the curves
γv play the role of 1-parameter subgroups! In terms of the exponential map, we have
γv (t) = Expp (tv).
Theorem 16.9. The exponential map Expp is smooth. It defines a diffeomorphism from a
neighborhood of 0 ∈ Tp M onto a neighborhood of p ∈ M .
 16. GEODESICS 45

Proof. Let Φ : R×T M → T M denote the flow of the geodesic spray, S for the connection
∇, and let π : T M → M be the base point projection. Then Expp is just the restriction of
the map π ◦ Φ to the submanifold {1} × Tp M , and hence is smooth. Compute T0 Expp : For
v ∈ Tp M we have,
d d
T0 Expp (v) = |t=0 Expp (tv) = |t=0 γv (t) = v,
dt dt
so T0 Expp is just the identity map Tp M → Tp M . From the inverse function theorem, it then
follows that Expp is a diffeomorphism on some small neighborhood of 0 ∈ Tp M . 

If one chooses a basis in Tp M , thus identifying Tp M = ∼ Rm , the exponential map gives a

system of local coordinates x1 , . . . , xm on a neighborhood of p. These coordinates are called
normal coordinates at p, and have very nice properties:
Theorem 16.10. In normal coordinates x1 , . . . , xm based at p ∈ M , the geodesics through
p are given by straight lines,
xi (t) = tai ai ∈ R.
Moreover, all Christoffel symbols Γijk vanish at 0.

Proof. By definition of the exponential map, Expp (ta) for a ∈ Rm ∼ = Tp M is the geodesic
with initial velocity v = a. Inserting xi (t) = tai into the geodesic equation, we obtain
X
Γijk (ta) aj ak = 0
jk

for all a. Setting t = 0, it follows that Γijk (0) = 0. 

We now specialize to the case that ∇ is the Levi-Civita connection corresponding to a

(pseudo-)Riemannian metric g on M . Define the energy function
E ∈ C ∞ (T M ), E(v) = 12 gp (v, v), v ∈ Tp M.
(Thus, the energy function is just the quadratic function associated to g, up to the factor 12 .)
Since parallel transport for a metric connection preserves inner products, the geodesic flow
preserves the energy: That is, S(E) = 0. It follows that S is tangent to the level surfaces of
the energy functional.
Geodesics for the Levi-Civita connection have an important alternative characterization,
as critical points of the action functional.
Definition 16.11. Let γ : [a, b] → M be a smooth curve in M .12 One defines the action
of γ by
Z b Z b
A(γ) = 1
E(γ̇(t)) dt = 2 ||γ̇(t)||2 dt.
a a
R
1 bP
In local coordinates, A(γ) = 2 a ij gij (x(t)) ẋi x˙j dt. The action functional is closely
related to the length functional (assuming that g is positive definite, so that L(γ) is defined):

12Here smooth means that γ extends to a smooth curve on an open interval J containing [a, b].
46 CONTENTS

Lemma 16.12. Let γ : [a, b] → M be a smooth curve. Then

L(γ)2 ≤ 2(b − a)A(γ).
Equality holds if and only if γ has constant speed, that is ||γ̇|| is constant.
Proof. The Cauchy-Schwartz inequality13 implies
Z b Z b
2
( f (t)dt) ≤ (b − a) f (t)2 dt
a a
with equality if and only if f is constant. 
Suppose γ : [a, b] → M is a smooth curve. A 1-parameter variation of γ is a family of
curves γs : [a, b] → M defined for −ǫ < s < ǫ, with γs (a) = γ(a) and γs (b) = γ(b) for all t,
γ0 = γ, and such that the map (s, t) 7→ γs (t) is smooth.
Theorem 16.13. A smooth curve γ : [a, b] → M is a geodesic if and only if for all 1-
parameter variations γs of γ,
d
|s=0 A(γs ) = 0.
ds
Proof. Let γs (t) be a 1-parameter variation. We can view γs (t) as a curve with parameter
t, depending on s as a parameter, or vice versa. Let a ′ indicate s-derivatives. Since the Levi-
Civita connection ∇ is torsion-free,
Dγ ′ Dγ̇
= .
dt ds
′ P  d2 xi P 
(In local coordinates, the left hand side is given by Dγ dt = i dsdt + Γ i ẋ x′
jk jk j
∂
k ∂xi , while
the right hand side is given by a similar expression with s, t-derivatives in opposite order. The
two expressions are the same since the Christoffel symbols for a torsion free connection are
symmetric in j, k.) Since ∇ is a metric connection, we can therefore compute,
Z b
d 1 ∂
A(γs ) = 2 g(γ̇, γ̇) dt
ds a ∂s
Z b
Dγ̇
= g( , γ̇) dt
a ds
Z b
Dγ ′
= g( , γ̇) dt
a dt
Z b Z b
d Dγ̇
= g(γ ′ , γ̇) dt − g(γ ′ , ) dt
a dt a dt
Z b
Dγ̇
= − g(γ ′ , ) dt.
a dt
13The Cauchy-Schwartz inequality for integrals says that
Z b Z b Z b
( f (t)g(t)dt)2 ≤ ( f (t)2 dt) ( g(t)2 dt)
a a a

with equality if and only if f, g are linearly dependent (i.e. proportional). The desired inequality follows by
setting g = 1.
 17. THE HOPF-RINOW THEOREM 47

Here we have used that γ ′ (a) = γ ′ (b) = 0. The resulting expression vanishes at s = 0, for all
variations, if and only if Ddtγ̇ = 0, i.e. if and only if γ is a geodesic. 
In particular, if γ : [a, b] → M minimizes the action, in the sense that
A(γ) ≤ A(γ̃)
for all paths γ̃ : [a, b] → M (defined on the same interval [a, b]) with γ̃(a) = γ(a), γ̃(b) = γ(b),
then γ is a geodesic. (However, it is not necessary for a geodesic to minimize the action.)
Theorem 16.14. A curve γ : [a, b] → M with ||γ̇(t)|| = const is a geodesic if and only if,
for all 1-parameter variations γs of γ,
∂
|s=0 L(γs ) = 0.
∂s
We leave the proof as an exercise. We have to put in by hand the assumption that γ
has constant speed, since the length functional is invariant under reparametrizations. (The 1-
parameter variations γs need not have finite speed.) In particular, length minimizing, constant
speed curves are always geodesics.

17. The Hopf-Rinow Theorem

The Hopf-Rinow Theorem says that a Riemannian manifold (M, g) is geodesically complete
(the geodesic flow is complete, i.e. all geodesics exists for all time) if and only if it is completene
as a metric space (every Cauchy sequence converges). To prepare for the proof, we need some
more facts on normal coordinates and the exponential map.
Definition 17.1. Let (M, g) be a Riemannian manifold. The injectivity radius ip (M ) > 0
of p ∈ M is the supremum of the set of all r > 0 such that the exponential map Expp is defined
on the open ball Br (0) and is injective. The injectivity radius i(M ) ≥ 0 of M is the infimum
of all ip (M ) with p ∈ M .
Example 17.2. For the unit circle S 1 ⊂ R2 with the standard Riemannian metric, each
point has injectivity radius π. Similarly, for the sphere M = S m−1 ⊂ Rm , the injectivity radius
of any point is ip (M ) = π. For M = Rm , ip (M ) = ∞.
Theorem 17.3. For all 0 < r < ip (M ), the radial geodesics Expp (tv) intersect the spheres
Expp (Sr (0)) orthogonally. For any v ∈ Sr (0), the point q = Expp (v) has distance d(p, q) = r
from p, and the geodesic Expp (tv) is the unique (up to reparametrization) curve of length d(p, q)
connecting p, q. In particular,
Expp (Sr (0)) = Sr (p)
for any 0 < r < ip (M ).
We will obtain this result as a consequence of the following Lemma on “geodesic polar
coordinates” around p. Let x1 , . . . , xm denote the normal coordinates on a neighborhood U of
p, obtained by choosing an orthornormal basis in Tp M . In this coordinates,
gij (0) = δij , Γijk (0) = 0.
P 2
Introduce polar coordinates (ρ, φ1 , . . . , φm−1 ) on Tp M , thus ρ2 = xi and φ1 , . . . , φm−1 are
local coordinates on the unit sphere S m−1 ⊂ Tp M . (The particular choice of coordinates on
48 CONTENTS

S m−1 will be irrelevant.) Using Expp , we can view these as coordinates on (suitable open
∂
subsets of) Expp (Br (0)) for r < ip (M ). In particular, the coordinate vector field ∂ρ given as

m
∂ 1 X ∂
= xi
∂ρ ||x|| ∂xi
i=1

is a well-defined vector fiel on Expp (Br (0)\{0}). Note that its integral curves are exactly the
unit speed radial geodesics.

Lemma 17.4 (Geodesic polar coordinates). In geodesic polar coordinates around p,

∂ ∂ ∂ ∂
gρρ ≡ g( , ) = 1, gρφj ≡ g( , ) = 0.
∂ρ ∂ρ ∂ρ ∂φj

That is, the radial geodesics Expp (tv) are orthogonal to the spheres Expp (Sr (0)), for all 0 <
r < ip (M ).
∂ ∂
Proof. Thus ∇ ∂ ∂ρ = 0. In particular, the length of ∂ρ is constant along radial geodesics.
∂ρ
But
∂ ∂
lim g( , )|tx = 1,
t→0 ∂ρ ∂ρ
∂ ∂ ∂
since gij (0) = δij and since ∂ρ has length one in the Euclidean metric. It follows that g( ∂ρ , ∂ρ )=
1 everywhere. Furthermore, using that the connections is torsion-free,

∂ ∂ ∂ ∂ ∂ ∂ ∂
g( , ) = g(∇ ∂ , ) + g( , ∇ ∂ )
∂ρ ∂ρ ∂φj ∂ρ ∂ρ ∂φj ∂ρ ∂ρ ∂φj
∂ ∂
= g( , ∇ ∂ )
∂ρ ∂φj ∂ρ
∂ ∂ ∂
= 21 g( , )
∂φj ∂ρ ∂ρ
∂
= 12 1 = 0.
∂φj

∂ ∂ ∂ ∂
Thus g( ∂ρ , ∂φ j
) is constant in radial directions. But limt→0 g( ∂ρ , ∂φ j
)|tx = 0, again since
∂ ∂
gij (0) = δij . Thus g( ∂ρ , ∂φ j
) = 0 everywhere. 

Proof of Theorem 17.3. Let γ(t) (0 ≤ t ≤ 1) be any curve with γ(0) = p and γ(1) = q.
Suppose first that γ(t) ∈ Expp (Br (0)\{0}) for 0 < t < 1. In geodesic polar coordinates

∂ X ∂
γ̇ = ρ̇ + φ˙j
∂ρ ∂φj
j
 17. THE HOPF-RINOW THEOREM 49

thus g(γ̇, γ̇) ≥ |ρ̇|2 with equality if and only if φj = const. It follows that
Z 1
L(γ) = g(γ̇, γ̇)1/2 dt
0
Z 1
≥ |ρ̇| dt
0
Z 1
≥ ρ̇ dt
0
= ρ(1) = r,
with equality if and only if φj = const and ρ̇ ≥ 0 for all t. Clearly, curves leaving the set
Expp (Br (0)) for some time t ∈ (0, 1) will be even longer. 
Corollary 17.5. Let p, q ∈ M . Suppose there exists a piecewise smooth curve γ : [0, 1] →
M of length d(p, q) from p to q. Then γ is a reparametrization of a smooth (!) geodesic of
length d(p, q).
Proof. Since γ([0, 1]) ⊂ M is compact, the infimum of the set of all injectivity radii
iγ(t) (M ) is strictly positive. Let ǫ > 0 be smaller than this infimum. Then for any two points
on the curve, of distance less than ǫ, the unique shortest curve connecting these points is the
geodesic given by the exponential map. In particular, γ must coincide with that geodesic up
to reparametrization. 
We are now ready to prove the Hopf-Rinow theorem. We recall that a sequence xn , n =
1, . . . , ∞ in a metric space (X, d) (where d is the metric = distance function) is a Cauchy
sequence if for all ǫ > 0, there exists N > 0 such that d(xn , xm ) < ǫ for n, m ≥ N . In
particular, every convergent sequence is a Cauchy sequence. A metric space is called complete
if every Cauchy sequence in X converges. For instance, every compact metric space is complete,
while e.g. bounded open subsets of Rm (with induced metric) are incomplete.
Exercise 17.6. Show that every Cauchy sequence is bounded. That is, there exists p ∈ X
and R > 0 such that xn ∈ BR (p) for all n.
Theorem 17.7 (Hopf-Rinow). A Riemannian manifold (M, g) is geodesically complete, if
and only if it is complete as a metric space. In this case, any two points p, q may be joined by
a smooth geodesic of length d(p, q).
Proof. We may assume that M is connected.
Suppose M is geodesically incomplete. That is, there exists a maximal unit speed geodesic
γ : (a, b) → M with b < ∞. Since d(γ(ti ), γ(tj )) ≤ |tj − ti |, it follows that the sequence γ(ti )
for ti → b is a Cauchy sequence. On the other hand, this sequence cannot converge since γ(t)
leaves every given compact set14 for t → b.
Thus we have found a non-convergent Cauchy sequence, showing that M is incomplete as
a metric space.
The other direction is a bit harder: Suppose M is geodesically complete. Pick p ∈ M . We
will show that every closed metric ball Br (p) is compact, which implies that M is metrically
14For any compact set K, there exists ǫ > 0 less than the injectivity radius of any point in K. Hence, unit
speed geodesics for points starting in K exist at least for time ǫ.
50 CONTENTS

complete (Any Cauchy sequence is bounded, hence is contained in Br (p) for r sufficiently large.
Any Cauchy sequence in a compact set converges.) By geodesic completeness, the exponential
map
Expp : Tp M → M
is defined. It suffices to show that for all r > 0,
Expp (Br (0)) = Br (p),
where Br (0) ⊂ Tp M is the ball of radius r for the inner product gp . Indeed, Br (0) is compact,
and images of compact sets under continuous maps are again compact. The inclusion ⊆ is
clear; the harder part is the opposite inclusion Br (p) ⊆ Expp (Br (0)). Let
H = {r > 0 | Expp (Br (0)) = Br (p)}.
We have to show H = [0, ∞). We first show that H is closed. Let rn ∈ H with limn→∞ rn = r.
We have to show r ∈ H. Given q ∈ Br (p), choose qn ∈ Brn (p) with qn → q. Choose
vn ∈ Brn (0) with Expp (vn ) = qn . Since Brn (0) ⊂ Br (0) which is compact, there exists a
convergent subsequence. Let v ∈ Br (0) be the limit point; then Expp (v) = q. Since q was
arbitrary, this shows r ∈ H.
We next show that if r ∈ H then r + ǫ ∈ H for ǫ > 0 sufficiently small. Since H is closed,
this will finish the proof H = [0, ∞). Let q ∈ Br+ǫ (p).
Choose 0 < ǫ < iK (M ) := inf p∈K ip (M ), where K is the compact subset K = Br (p). Thus,
for any x ∈ Br (p) with d(x, q) ≤ ǫ, there exists a unique geodesic in M joining x, q, of length
d(x, q). To find such a point x, choose a sequence of curves γn : [0, 1] → M connecting p, q, of
length ≤ d(p, q) + n1 . Let tn ∈ [0, 1] be the smallest value such that xn := γn (tn ) ∈ ∂Br (p). We
have
1
d(p, q) ≤ d(p, xn ) + d(xn , q) ≤ L(γn ) ≤ d(p, q) + .
n
Since Br (p) is compact some subsequence of the sequence xn converges to a limit point x ∈
∂Br (p), with
d(p, x) + d(x, q) = d(p, q).
Since d(p, q) ≤ r + ǫ and d(p, x) = r, this implies d(x, q) ≤ ǫ. Choose v ∈ Br (0) with
Expp (v) = p.
Since d(x, q) ≤ ǫ, there exists a unique unit speed geodesic of length d(x, q) from x to
q. Together with the unit speed geodesic Expp (tv/r), we obtain a piecewise smooth curve of
length d(p, q) from p to q. As observed above, it is automatic that this curve is smooth, hence
a geodesic. It hence coincides with the unique continuation of the geodesic Expp (tv/r). It
follows that Expp (ṽ) = q for
r+ǫ
ṽ = v = (1 + ǫ/r)v ∈ Br+ǫ (0).
r

Note that we didn’t quite use geodesic completeness in the proof: We only used that Expp
is defined on all of Tp M . One might call this geodesic completeness at p. What we’ve shown is
that geodesic completeness at any point p implies geodesic completeness everywhere.
 18. THE CURVATURE TENSOR 51

18. The curvature tensor

An affine connection ∇ on a manifold M is called flat if, around any point, there exist
local coordinates in which all Christoffel symbols of ∇ vanish. A (pseudo-)Riemannian metric
is called flat if the corrsponding Levi-Civita connection is flat.
Flatness of a connection implies that parallel transport along a path does not change
under 1-parameter variations of the path. In practice, the definition is not always easy to
verify, mainly because Christoffel symbols may vanish in one coordinate system and be non-
zero in another. One is therefore interested in invariants of a connection: That is, quantities
constructed from the connection whose vanishing does not depend on a choice of coordinates.
One such example is the torsion T (X, Y ) = ∇X (Y ) − ∇Y (X) − [X, Y ] of a connection: Recall
that by C ∞ -linearity, it defines a bi-linear map T : Tp M × Tp M → Tp M , and clearly T has
to vanish for any flat connection. A second invariant is the curvature operator to be discussed
now.
Definition 18.1. For vector fields X, Y one defines the curvature operator R(X, Y ) :
X(M ) → X(M ) by
R(X, Y )(Z) = ∇X ∇Y Z − ∇Y ∇X Z − ∇[X,Y ] Z.
In short, R(X, Y ) = [∇X , ∇Y ] − ∇[X,Y ] .
Theorem 18.2. The map (X, Y, Z) 7→ R(X, Y )(Z) is C ∞ (M )-linear in X, Y, Z. It follows
that for u, v ∈ Tp M , there is a well-defined linear map Rp (u, v) : Tp M → Tp M such that
Rp (u, v)(w) = (R(X, Y )(Z))p
whenever Xp = u, Yp = v, Zp = w.
Proof. R is C ∞ (M )-linear in Z: For all f ,
∇X ∇Y (f Z) = f ∇X ∇Y Z + X(f )∇Y Z + Y (f )∇X Z + X(Y (f ))Z,
∇Y ∇X (f Z) = f ∇Y ∇X Z + Y (f )∇X Z + X(f )∇Y Z + Y (X(f ))Z,
∇[X,Y ] (f Z) = f ∇[X,Y ] Z + [X, Y ](f ) Z.
Subtracting the last two equations from the first, we find R(X, Y )(f Z) = f R(X, Y )(Z) as
desired. Similarly one checks C ∞ (M )-linearity in X, Y . 
C ∞ -linearity of the curvature operator implies that in local charts, R is determined by its
values on coordinate vector fields. We can thus introduce components Rijk l of the curvature
tensor, defined by
∂ ∂ ∂ X ∂
l
R( , )( )= Rijk .
∂xi ∂xj ∂xk ∂xl
l
These can be expressed in terms of Christoffel symbols: We find, after short calculation,

l
∂Γljk ∂Γlik X  r l 
Rijk = − + Γjk Γir − Γrik Γljr .
∂xi ∂xj r

Recall that this complicated expression appeared in the proof of Gauss’ theorem egregium in
the curves and surfaces course, but it was somewhat unmotivated back then!
52 CONTENTS

Since R has four indices, the curvature tensor seems to give (dim M )4 invariants of a
connection. In reality, the number is much smaller, due to symmetry properties of the curvature
tensor. First of all, it is of course anti-symmetric in X, Y . More interesting is:
Theorem 18.3 (Bianchi identity). Suppose ∇ has vanishing torsion. Then
R(X, Y )Z + R(Y, Z)X + R(Z, X)Y = 0.
l + Rl + Rl
That is, in local coordinates, Rijk jki kij = 0.

Proof. We show that the left hand side vanishes at any given p ∈ M . Let Expp :
Br (0) → M be the exponential map, where r < ip (M ). Introduce normal coordinates on
U = Expp (Br (0)) = Br (p). Then all Christoffel symbols Γkij vanish at 0, and we have (at 0)

l l l
∂Γljk ∂Γlik ∂Γlki ∂Γlji ∂Γlij ∂Γlkj
Rijk + Rjki + Rkij = − + − + − .
∂xi ∂xj ∂xj ∂xk ∂xk ∂xi
In the torsion-free case, this vanishes since the Christoffel symbols are symmetric in the lower
indices. 
Exercise 18.4. Give a coordinate-free proof of the Bianchi identity.
Suppose now that g is a (pseudo-)Riemannian metric and ∇ the corresponding Levi-Civita
connection. For vector fields X, Y, Z, W define the curvature tensor of g by
R(X, Y, Z, W ) = g(R(X, Y )Z, W ).
∂
In components Rijkl = R( ∂x , ∂ , ∂ , ∂ ) we have,
i ∂xj ∂xk ∂xl
X
r
Rijkl = Rijk grl .
r

Theorem 18.5. The curvature tensor has the symmetry properties,

R(X, Y, Z, W ) = −R(Y, X, Z, W ) = −R(X, Y, W, Z) = R(Z, W, X, Y )
and
R(X, Y, Z, W ) + R(Y, Z, X, W ) + R(Z, X, Y, W ) = 0.
Proof. The last identity is just re-stating the Bianchi identity. Anti-symmetry of R(X, Y, Z, W )
in the first two entries X, Y is obvious from the definition. To prove anti-symmetry in the last
two entries, it is enough to show that R(X, Y, Z, Z) = 0 for all X, Y, Z. We have,
R(X, Y, Z, Z) = g(∇X ∇Y (Z), Z) − g(∇Y ∇X (Z), Z) − g(∇[X,Y ] Z, Z)
= X g(∇Y (Z), Z) − g(∇Y (Z), ∇X (Z)) − Y g(∇X (Z), Z) + g(∇X (Z), ∇Y (Z))
− 21 [X, Y ] g(Z, Z)
= X g(∇Y (Z), Z) − Y g(∇X (Z), Z) − 12 [X, Y ] g(Z, Z)


= 12 X(Y g(Z, Z)) − Y (X g(Z, Z)) − [X, Y ] g(Z, Z)
= 0.
It remains to prove R(X, Y, Z, W ) = R(Z, W, X, Y ). In fact, this is a consequence of the other
symmetry properties, although in a rather non-obvious way. First, one adds the four equations
 19. CONNECTIONS ON VECTOR BUNDLES 53

obtained from the Bianchi identity R(X, Y, Z, W ) + R(Y, Z, X, W ) + R(Z, X, Y, W ) = 0 by

interchaning W with X, Y, Z. This gives
R(X, Y, Z, W ) + R(Y, Z, X, W ) + R(Z, X, Y, W )
+ R(W, Y, Z, X) + R(Y, Z, W, X) + R(Z, W, Y, X)
+ R(X, W, Z, Y ) + R(W, Z, X, Y ) + R(Z, X, W, Y )
+ R(X, Y, W, Z) + R(Y, W, X, Z) + R(W, X, Y, Z) = 0.
Using anti-symmetry in the first two entries and the last two entries, we obtain some cancella-
tions and find,
R(W, X, Y, Z) + R(W, Y, Z, X) + R(W, Z, X, Y ) = 0.
Using the Bianchi identity again, we have
R(Z, X, W, Y ) = −R(W, Z, X, Y ) − R(X, W, Z, Y )
= −R(W, Z, X, Y ) − R(W, X, Y, Z)
= R(W, Y, Z, X).
Exercise 18.6. Prove anti-symmetry of R(X, Y, Z, W ) in Z, W in local (normal) coordi-
nates, similar to our proof of the Bianchi identity.


19. Connections on vector bundles

In this section we define connections on vector bundles E → M be a vector bundle. (We
are mainly interested in E = T M , but other bundles will appear as well.)
Definition 19.1. A connection (covariant derivative) on E is a bi-linear map
∇ : X(M ) × Γ∞ (E) → Γ∞ (E), (X, σ) 7→ ∇X σ,
such that ∇ is C ∞ -linear in the X variable and
∇X (f σ) = f ∇X σ + X(f )σ
for all f ∈ C ∞ (M ), X ∈ X(M ), σ ∈ Γ∞ (E).
Definition 19.2. The curvature operator corresponding to the connection ∇ is the linear
map R(X, Y ) : Γ∞ (E) → Γ∞ (E)
R(X, Y ) = ∇X ∇Y − ∇Y ∇X − ∇[X,Y ] .
As for affine connections, the curvature operator R(X, Y ) is in fact a C ∞ -linear map, and
moreover is C ∞ -linear in X, Y also.

19.1. Connections on trivial bundles. Let us first consider the case of a trivial vector
bundle, E = M × Rk . Let e1 , . . . , ek be the standard basis of Rk . These define “constant”
sections ǫ1 , . . . , ǫk of M × Rk , and the most general section has the form,
X
σ= σa ǫa .
a
54 CONTENTS

where the σa are functions. It is immediate that

X
∇0X σ := X(σa )ǫa
a

defines a connection. This is called the trivial connection on the trivial bundle E = M × Rk .
Now let ∇X be any connection. Define a map
A : X(M ) → C ∞ (M, End(Rk )), X 7→ A(X)
by
∇X σ = ∇0X σ + A(X)σ.
Thus A(X) is a matrix-valued function on M , measuring the difference from the trivial con-
nection. Letting Aab (X) be its components, we have
X X
∇X σ = X(σa )ǫa + Aab (X)σb ǫa .
a ab
That is, X
(∇X σ)a = X(σa ) + Aab (X)σb .
b
Notice that the map X 7→ A(X) is C ∞ (M )-linear. Conversely, every C ∞ (M )-linear map
of this form defines a connection. That is:
Proposition 19.3. The space of connections on a trivial bundle E = M × Rk is in 1-1
correspondence with the space of C ∞ (M )-linear maps, X(M ) → C ∞ (M, End(Rk )), X 7→ A(X).
Under this correspondence, the map A defines the connection
∇X = ∇0X + A(X).
One calls A the connection 1-form of the connection ∇.
Suppose now that ǫ′a ∈ Γ∞ (M, Rk ) is a new basis of the space of sections. That is,
ǫ′a = gba ǫb
where the matrix-valued function g with coefficients gab ∈ C ∞ (M ) is invertible everywhere.
Let σa′ denote the components of σ in the new basis, i.e.
σa′ = gab σb .
Define the connection 1-form A′ of ∇ in the new basis by
X X
∇X σ = (X(σa′ ) + A′ (X)ab σb′ )ǫ′a .
a b
P
We have ǫ′a = −1
b (g )ba ǫb ,
therefore
X X X
X(σc ) + A(X)cb σb = (g −1 )ca (X(σa′ ) + A′ (X)ab σb′ )
b a b
X X X X
−1
= gca ( gab X(σb ) + X(gab )σb + A′ (X)ab gbc σc )
a b b b
X X
−1 −1 ′
= X(σc ) + (g )ca X(gab ) + gca A (X)ab gbd σd .
ab abd
 19. CONNECTIONS ON VECTOR BUNDLES 55

Comparing, we read off, using matric notation,

A(X) = g −1 A′ (X)g + g −1 X(g),
or equivalently,
A′ (X) = gA(X)g −1 − X(g)g −1 .
In the theoretical physics literature, connections are called gauge fields, and sections of (possibly
trivial) bundles E are called particle fields. The change of basis using g is called a gauge
transformation, and the above formula is called the gauge group action of C ∞ (M, Gl(k, R)).
Exercise 19.4. Show that the curvature operator R(X, Y ) on E ×Rk transforms according
to
R′ (X, Y ) = g R(X, Y ) g −1 .
Give a formula for R in terms of connection 1-forms.
19.2. Connections on non-trivial vector bundles. This discussion carries over to
more general vector bundles, as follows. Let End(E) → M be the endomorphism bundle of E,
with fibers End(E)p = End(Ep ) the vector space15 of endomorphisms Ep → Ep . The space of
sections of End(E) is isomorphic to the space of C ∞ (M )-linear endomorphisms of the vector
space Γ∞ (E):
Γ∞ (End(E)) = EndC ∞ (M ) (Γ∞ (E)).
A connection gives a linear map ∇X ∈ End(Γ∞ (E)), which is not C ∞ (M )-linear. However,
the difference between any two connections is:
A(X) = ∇′X − ∇X ∈ EndC ∞ (M ) (Γ∞ (E)) = Γ∞ (End(E)).
Conversely, if ∇X is any connection, and is
X 7→ A(X) ∈ Γ∞ (End(E)) is C ∞ (M )-linear, then ∇′X = ∇X + A(X) defines a new
connection. This proves half of:
Proposition 19.5. Every vector bundle E admits a connection ∇. The most general
connection is ∇′X = ∇X + A(X) for some C ∞ (M )-linear map, A : X(M ) → Γ∞ (End(E)).
Proof. Any local trivializations E|U ∼ = U × Rk defines a connection ∇U on E|U coming
from the trivial connection on U × Rk . Let Uα be a locally finite open cover of M , with local
trivializations of E|Uα , and let ∇α be the corresponding local connections. Let χα be a partition
of unity, and define X
∇X (σ) = χα ∇αX (σ|Uα ).
α
This has all the properties of a connection. 
Let E|Uα ∼
= Uα × Rk be a local trivialization. Thus ∇ becomes a connection on Uα × Rk ,
described by some Aα (X) ∈ C ∞ (Uα , End(Rk )). The maps X 7→ Aα are called the local
connection 1-forms for ∇. If Uα is a coordinate chart, with local coordinates x1 . . . , xm , Aα is
described by m matrix valued functions
∂
Aα ( ) ∈ C ∞ (Uα , End(Rk )).
∂xi
15In fact, each fiber End(E ) is an algebra, and accordingly End(E) is an example of an algebra bundle.
p
56 CONTENTS

The components
∂
Γbia := (Aα )ab (
)
∂xi
are also called the Christoffel symbols of the connection with respect to the given local coordi-
nates.
19.3. Constructions with connections. Given a vector bundle E, let E ∗ be its dual
bundle. There is a natural pairing of the spaces of sections,
h·, ·i : Γ∞ (E ∗ ) × Γ∞ (E) → C ∞ (M ), hτ, σip := hτp , σp i ≡ τp (σp ).
In other words, Γ∞ (E ∗ ) is identified with the space of C ∞ -linear maps Γ∞ (E) → C ∞ (M ).
Proposition 19.6 (Duals). For any connection ∇ on E, there is a unique connection ∇∗
on E ∗ with property,
Xhτ, σi = h∇∗X τ, σi + hτ, ∇X σi.
Proof. Try to define ∇∗ by this equation:
h∇∗X τ, σi = Xhτ, σi − hτ, ∇X σi.
For f ∈ C ∞ (M ) we have,
Xhf τ, σi − hf τ, ∇X σi = hX(f )τ, σi + f (Xhτ, σi − hτ, ∇X σi)
Showing that ∇∗X (f τ ) = X(f )τ + f ∇∗X τ as desired. 
If E, E ′ are two vector bundles over M , we can form the direct sum E ⊕ E ′ , with
Γ∞ (E ⊕ E ′ ) = Γ∞ (E) ⊕ Γ∞ (E ′ ).
Proposition 19.7 (Direct sums). If ∇ is a connection on E and ∇′ a connection on E ′ ,
there is a unique connection ∇ ⊕ ∇′ on E ⊕ E ′ such that
(∇ ⊕ ∇′ )X (σ ⊕ σ ′ ) = ∇X σ ⊕ ∇′X σ ′ .
Finally, recall that if E is a vector bundle over M , and F ∈ C ∞ (N, M ) a smooth map
from a manifold N , we define a pull-back bundle F ∗ E with fibers (F ∗ E)q = EF (q) . Its space
of sections Γ∞ (F ∗ E) is generated (as a C ∞ (N )-module) by the subspace F ∗ Γ∞ (E).
Proposition 19.8. Let E → M be a vector bundle with connection ∇, and F ∈ C ∞ (N, M ).
Then there is a unique connection F ∗ ∇ such that for all σ ∈ Γ∞ (E), q ∈ N , w ∈ Tq N
(F ∗ ∇)w (F ∗ σ) = ∇Tq F (w) σ.
Proof. Exercise. 
The pull-back connection F ∗ ∇ can be desribed in terms of connection 1-forms: If E|U ∼ =
U × Rk is a local trivialization of E, and X 7→ Aab (X) the connection 1-form of ∇ in terms
of this local trivialization. Then we obtain a local trivialization F ∗ E|F −1 (U ) ∼
= F −1 (U ) × Rk ,
∗ 16
with connection 1-forms given by the pull-back forms, F Aab .
16Recall that C ∞ (M )- linear maps X(M ) → C ∞ (M ) are identified with sections of T ∗ M , i.e. 1-forms, and
that there is a natural pull-back map F ∗ Γ∞ (T ∗ M ) → Γ∞ (T ∗ N ) given by (F ∗ α)q = (Tq F )∗ αF (q) .
 19. CONNECTIONS ON VECTOR BUNDLES 57

19.4. Parallel transport. Suppose E is a vector bundle over M with connection ∇, and
γ : J → M is any smooth curve. Sections of the pull-back bundle γ ∗ E are called sections of E
along γ. A connection ∇ on E induces a pull-back connection γ ∗ ∇ on γ ∗ E, and one can define
a covariant derivative along γ by
D Dσ
: Γ∞ (γ ∗ E) → Γ∞ (γ ∗ E), := (γ ∗ ∇) ∂ σ.
Dt Dt ∂t

A section σ along γ is called parallel if Dσ Dt = 0.

Suppose E|U = U × Rk is a local trivialization of E with γ(t) ∈ U , given by a basis
ǫ1 , . . . , ǫk ∈ Γ∞ (U, E|U ) of the space of sections. Then we can write
X
σ(t) = σa (t)(ǫa )γ(t) ∈ Eγ(t) ,
a
and the components of the covariant derivative is given by the formula,
 
Dσ dσa X
= + Aab (γ̇)σb (t).
Dt a dt
b
Furthermore, if U is the domain of a coordinate chart, defining local coordinates x1 , . . . , xk ,
and
∂
Γbia = Aab ( )
∂xi
are the corresponding Christoffel symbols, the formula becomes,
 
Dσ dσa X b
= + Γia ẋi σb (t).
Dt a dt
ib
As for affine connections, one shows that for any given σt0 ∈ Eγ(t0 ) , there is a unique parallel
section σ(t) along γ with initial value σ(t0 ) = σt0 . In this way, connections ∇ define parallel
transport in vector bundles.
There is a more geometric way of understanding paralel transport on a vector bundle
π : E → M . Consider the tangent map T π : T E → T M . Its kernel at u ∈ Ep is
ker(Tu π) = Tu (Ep ),
the tangent space at u to the fiber Ep . It is called the vertical subspace
Vu E = Tu (Ep ) ⊂ Tu E.
Note that since Ep is a vector space, Tu (Ep ) ∼= Ep . This means that we have a natural
isomorphism, V E = π ∗ E (the pull-back of E to a vector bundle over E.).
Since the map Tu π : Tu E → Tp M is clearly onto, it we have
Tu E/Vu E = Tp M.
It turns out that every connection ∇ defines a complementary horizontal subspace Hu E ⊂ Tu E,
with
Tu E = Vu E ⊕ Hu E.
In fact, HE is a vector subbundle of T E, called the horizontal bundle, and T E = V E ⊕ HE.
In a coordinate free way, one may define Hu E as follows:
58 CONTENTS

Theorem 19.9. Let ∇ be a connection on E. Given u ∈ Ep , there exists a section σ ∈

Γ∞ (E) with σp = e and (∇σ)p = 0. The image
Hu E := Tp σ(Tp M )
is independent of the choice of σ.
Exercise 19.10. 1. Proof this theorem.
2. Given an alternative definition of Hu E in local coordinates and using a local trivialization
of E: Show that the horizontal space in E = U × Rk is spanned by all
∂ X
− Γbia ub ǫa (p)
∂xi a

Note that it is impossible, in general, to choose σ with σp = u and ∇σ = 0 everywhere:

This is related to the problem that in general, distributions of rank ≥ 2 need not be integrable.
One can characterize parallel transport in terms of the horizontal bundle HE ⊂ T E as
follows: For any curve γ(t) in M , and any given u ∈ Eγ(t0 ) , there is a unique curve σ(t) in E
such that σ(t0 ) = u and
π(σ(t)) = γ(t), σ̇ ∈ Hσ(t) E
for all t. The curve σ(t) is called the horizontal lift of γ. Note that a curve σ(t) projecting to
γ(t) is the same thing as a section of E along γ.
The splitting T E = HE ⊕ V E = π ∗ T M ⊕ V E given by ∇ defines a horizontal lift of vector
fields:
Lift∇ : X(M ) → X(E).
Here Lift∇ (X)u is the unique tangent vector in Hu E projecting to Xp . The horizontal lifts of
integral curves of X are integrla curves of its horizontal lift Lift∇ (X). Note that by construction,
Lift∇ (X) ∼π X.
Hence if X, Y are two vector fields,
[Lift∇ (X), Lift∇ (Y )] ∼π [X, Y ].
This shows that the vector field [Lift∇ (X), Lift∇ (Y )] − Lift∇ ([X, Y ]) must be vertical. That is,
it is a section of V E = π ∗ E. What is this section?
Theorem 19.11. For any u ∈ Ep , and any X, Y , we have
[Lift∇ (X), Lift∇ (Y )]u − Lift∇ ([X, Y ])u = R(X, Y )p u.
Theorem 19.12. The following are equivalent:
(a) The curvature R of ∇ vanishes.
(b) Horizontal lift X(M ) → X(E) is a Lie algebra homomorphism.
(c) The horizontal distribution HE ⊂ T E is integrable.
(d) Parallel transport along paths is invariant under homotopies leaving the end points
fixed.
We leave the proofs as exercises, or to be looked up in textbooks.