Chapter 3

Connections

Contents
3.1 Parallel transport . . . . . . . . . . . . . . . . . 69
3.2 Fiber bundles . . . . . . . . . . . . . . . . . . . . 72
3.3 Vector bundles . . . . . . . . . . . . . . . . . . . 75
3.3.1 Three definitions . . . . . . . . . . . . . . . . . 75
3.3.2 Christoffel symbols . . . . . . . . . . . . . . . . 80
3.3.3 Connection 1-forms . . . . . . . . . . . . . . . . 82
3.3.4 Linearization of a section at a zero . . . . . . . 84
3.4 Principal bundles . . . . . . . . . . . . . . . . . 88
3.4.1 Definition . . . . . . . . . . . . . . . . . . . . . 88
3.4.2 Global connection 1-forms . . . . . . . . . . . . 89
3.4.3 Frame bundles and linear connections . . . . . 91

3.1 The idea of parallel transport

A connection is essentially a way of identifying the points in nearby fibers
of a bundle. One can see the need for such a notion by considering the
following question:

Given a vector bundle π : E → M , a section s : M → E and a vector

X ∈ Tx M , what is meant by the directional derivative ds(x)X?

If we regard a section merely as a map between the manifolds M and

E, then one answer to the question is provided by the tangent map T s :
T M → T E. But this ignores most of the structure that makes a vector

69
70 CHAPTER 3. CONNECTIONS

bundle interesting. We prefer to think of sections as “vector valued” maps

on M , which can be added and multiplied by scalars, and we’d like to think
of the directional derivative ds(x)X as something which respects this linear
structure.
From this perspective the answer is ambiguous, unless E happens to
be the trivial bundle M × Fm → M . In that case, it makes sense to think
of the section s simply as a map M → Fm , and the directional derivative
is then 
d  s(γ(t)) − s(γ(0))
ds(x)X = s(γ(t)) = lim (3.1)
dt t=0
t→0 t
for any smooth path with γ̇(0) = X, thus defining a linear map

ds(x) : Tx M → Ex = Fm .

If E → M is a nontrivial bundle, (3.1) doesn’t immediately make sense

because s(γ(t)) and s(γ(0)) may be in different fibers and cannot be added.
Yet in this case, we’d like to think of different fibers as being equivalent, so
that ds(x) can still be defined as a linear map Tx M → Ex . The problem
is that there is no natural isomorphism between Ex and Ey for x 6= y; we
need an extra piece of structure to “connect” these fibers in some way, at
least if x and y are sufficiently close.
This leads to the idea of parallel transport, or parallel translation. The
easiest example to think of is the tangent bundle of a submanifold M ⊂ Rm ;
for simplicity, picture M as a surface embedded in R3 (Figure 3.1). For
any vector X ∈ Tx M at x ∈ M and a path γ(t) ∈ M with γ(0) = x, it’s
not hard to imagine that the vector X is “pushed” along the path γ in a
natural way, forming a smooth family of tangent vectors X(t) ∈ Tγ(t) M .
In fact, this gives a smooth family of isomorphisms

Pγt : Tx M → Tγ(t) M

with Pγ0 = Id, where “smooth” in this context means that in any local
trivialization of T M → M near x, Pγt is represented by a smooth path of
matrices. The reason this is well defined is that M has a natural connec-
tion determined by its embedding in R3 ; this is known as the Levi-Cività
connection, and is uniquely defined for any manifold with a Riemannian
metric (see Chapter 4).
Notice that in general if γ(0) = x and γ(1) = y, the isomorphism
1
Pγ : Tx M → Ty M may well depend on the path γ, not just its endpoints.
One can easily see this for the example of M = S 2 by starting X as a
vector along the equator and translating it along a path that moves first
along the equator, say 90 degrees of longitude, but then makes a sharp
turn and moves straight up to the north pole. The resulting vector Y
at the north pole is different from the vector obtained by transporting X
PSfrag replacements
3.1. PARALLEL TRANSPORT
γ 71
X(p0 )
M
p0
v0
R4
C
x
y
Ex
Ey

Figure 3.1: Parallel transport of two tangent vectors along a path in a

surface.

along the most direct northward path from x to y. Equivalently, one can
translate a vector along a closed path and find that it returns to a different
place in the tangent space than where it began (see Figure 3.2). As we will
see in Chapter 5, these are symptoms of the fact that S 2 has nontrivial
curvature.
For a general vector bundle π : E → M , we now wish to associate with
any path γ(t) ∈ M a smooth family of parallel transport isomorphisms
Pγt : Eγ(0) → Eγ(t) ,
with Pγ0 = Id. These are far from unique, and as the example of S 2 shows,
we must expect that they will depend on more than just the endpoints of
the path. But the isomorphisms are also not arbitrary; we now determine
what conditions are needed to make this a useful concept.
The primary utility of the family Pγt is that, once chosen, it enables
us to differentiate sections along paths. Namely, suppose γ(t) ∈ M is a
smooth path through γ(0) = x, and we are given a smooth section along γ,
i.e. a map s(t) with values in E such that π ◦ s(t) = γ(t) (equivalently, this
is a section of the pullback bundle γ ∗ E, cf. §2.3). We define the covariant
derivative of s along γ,
 
D  d  t −1 

s(t) := ∇t s(0) := (Pγ ) ◦ s(t)  . (3.2)
dt t=0 dt t=0

This is well defined since each of the vectors (Pγt )−1

◦ s(t) belongs to the
same fiber Ex , thus ∇t s(0) ∈ Ex . Defining ∇t s(t) similarly for all t gives
another smooth section of E along γ.
For a smooth section s : M → E, we are also now in a position to define
directional derivatives. Reasoning by analogy, we choose a path γ(t) ∈ M
with γ(0) = x and γ̇(t) = X ∈ Tx M , and define the covariant derivative

d  t −1 
∇s(x)X := ∇X s := (Pγ ) ◦ s(γ(t))  . (3.3)
dt t=0
 72 CHAPTER 3. CONNECTIONS

PSfrag replacements
γ
X(p0 )
M

v0
R4 p0
C
x
y
Ex
Ey

Figure 3.2: Parallel transport of a vector along a closed path in S 2 ⊂ R3

leads to a different vector upon return.

Once again we are differentiating a path of vectors in the same fiber Ex ,

so ∇s(x)X ∈ Ex . We can now begin to deduce what conditions must be
imposed on the isomorphisms Pγt : first, we must ensure that the expression
(3.3) depends only on s and X = γ̇(0), not on the chosen path γ. Let us
assume this for the moment. Then for any vector field X ∈ Vec(M ), the
covariant derivative defines another section ∇X s : M → E. Any sensible
use of the word “derivative” should require that the resulting map
∇s(x) : Tx M → Ex
be linear for all x. This is not automatic; it imposes another nontrivial
condition on our definition of parallel transport. It turns out that from
these two requirements, we will be able to deduce the most elegant and
useful definition of a connection for vector bundles.

3.2 Connections on fiber bundles

Before doing that, it helps to generalize slightly and consider an arbitrary
fiber bundle π : E → M , with standard fiber F . Now parallel transport
along a path γ(t) ∈ M will be defined by a smooth family of diffeomor-
phisms Pγt : Eγ(0) → Eγ(t) , and we define covariant derivatives again by for-
mulas (3.2) and (3.3). Now however, we are differentiating paths through
the fiber Ex ∼= F , which is generally not a vector space, so ∇s(x)X is
not in the fiber itself but rather in its tangent space Ts(x) (Ex ) ⊂ Ts(x) E.
Remember that the total space E is itself a smooth manifold, and has its
own tangent bundle T E → E.
3.2. FIBER BUNDLES 73

Definition 3.1. Let π : E → M be a fiber bundle. The vertical bundle

V E → E is the subbundle of T E → E defined by

V E = {ξ ∈ T E | π∗ ξ = 0}.

Its fibers Vp E := (V E)p ⊂ Tp E are called vertical subspaces.

Then Vp E = Tp (Eπ(p) ), so the vertical subbundle is the set of all vectors

in T E that are tangent to any fiber.

Exercise 3.2. Show that V E → E is a smooth real vector bundle if

π : E → M is a smooth fiber bundle, and the rank of V E → E is the
dimension of the standard fiber F .

By the above definition, the covariant derivative defines for each section
s : M → E and x ∈ M a map

∇s(x) : Tx M → Vs(x) E. (3.4)

We shall require the definition of parallel transport in fiber bundles to

satisfy two (not quite independent) conditions:

(i) The definition of ∇X s in (3.3) is independent of γ except for the

tangent vector γ̇(0) = X.

(ii) The map ∇s(x) : Tx M → Vs(x) E is linear.

Proposition 3.3. Suppose π : E → M is a fiber bundle and for every path

γ(t) ∈ M there is a smooth family of diffeomorphisms Pγt : Eγ(0) → Eγ(t)
satisfying Pγ0 = Id and conditions (i) and (ii). Then for each x ∈ M and
p ∈ Ex , there is a unique linear injection

Horp : Tx M → Tp E

such that Horp (γ̇(0)) = dtd Pγt (p)t=0 for all paths with γ(0) = x. Moreover,
the image of Horp is complementary to Vp E in Tp E.

Proof. Fix x0 ∈ M and p0 ∈ Ex0 . Then for any path γ(t) ∈ M with
γ(0) = x0 , the family of diffeomorphisms Pγt : Ex0 → Eγ(t) is the flow of
some vector field Y (t, p) on the total space of the pullback bundle γ ∗ E.
Choosing any section s : M → E with s(x0 ) = p0 and writing F (t, p) =
(Pγt )−1 (p), we have

d  ∂F
∇γ̇(0) s = F (t, s(γ(t))) = (0, p0 ) + D2 F (0, p0 ) ◦ T s(γ̇(0))
dt t=0 ∂t
= −Y (0, p0 ) + T s(γ̇(0)),
74 CHAPTER 3. CONNECTIONS

thus

d t 
Horp0 (γ̇(0)) = Pγ (p0 ) = Y (0, p0 ) = T s(γ̇(0)) − ∇γ̇(0) s. (3.5)
dt t=0

This expression is clearly a linear function of γ̇(0). It is also injective since

∇γ̇(0) s ∈ Vp0 E, and T s(γ̇(0)) ∈ Vp0 E if and only if γ̇(0) = 0, as we can see
by applying π∗ . The same argument shows (im Horp0 ) ∩ Vp0 E = {0}, and
since any non-vertical vector ξ ∈ Tp0 E \ Vp0 E can be written as T s(γ̇(0))
for some path γ and section s, clearly

im Horp0 ⊕Vp0 E = Tp0 E.

The moral is that parallel transport, if defined properly, determines

for every p ∈ E a horizontal subspace Hp E := im Horp complementary to
the vertical subspace Vp E. Conversely, it’s easy to see that choosing such
complimentary subspaces Hp E determines Pγt uniquely. This should be
sufficient motivation for the following definition.

Definition 3.4. A connection on the fiber bundle π : E → M is a smooth

distribution HE on the total space such that HE ⊕ V E = T E. For any
p ∈ E, the fiber Hp E ⊂ Tp E is called the horizontal subspace at p.

We can now recast all of the previous concepts in terms of horizontal

subspaces. Assume a connection (i.e. a horizontal subbundle) has been
chosen. Then for each x ∈ M and p ∈ Ex , the linear map π∗ : Tp E →
Tx M restricts to an isomorphism Hp E → Tx M . Its inverse is called the
horizontal lift
Horp : Tx M → Hp E.
A path through the total space E is called horizontal if it is everywhere
tangent to HE. Then given x0 ∈ M and p0 ∈ Ex0 , any path γ(t) ∈ M
with γ(0) = x0 lifts uniquely to a horizontal path γ̃(t) ∈ E with γ̃(0) = p0 .
This path is similarly called the horizontal lift of γ, and its tangent vectors
satisfy
d
γ̃(t) = Horγ̃(t) (γ̇(t)).
dt
By considering horizontal lifts for all possible p ∈ Ex0 , we obtain naturally
the parallel transport diffeomorphisms Pγt : Ex0 → Eγ(t) . Finally, (3.5)
yields a convenient formula for the covariant derivative with respect to any
vector X ∈ Tx M ,
∇X s = T s(X) − Hors(x) (X).
3.3. VECTOR BUNDLES 75

Note that since π∗ T s(X) = X, the second term on the right is simply the
projection of T s(X) to the horizontal subspace. We can express this more
simply by defining the vertical projection

K : T E → V E,

which maps each Tp E to the vertical subspace Vp E by projecting along

Hp E. Then
∇X s = K ◦ T s(X), (3.6)
so the covariant derivative is literally the “vertical part” of the tangent
map. For a section s(t) ∈ E along a path γ(t) ∈ M , we have the analogous
formula
∇t s(t) = K(ṡ(t)). (3.7)
As one would expect, it is clear from this formula that s(t) is a horizontal
lift of γ(t) if and only if ∇t s ≡ 0.
The projection K : T E → V E is called a connection map, and it gives
an equivalent definition for connections on fiber bundles.

Definition 3.5. A connection on the fiber bundle π : E → M is a smooth

fiberwise linear map K : T E → V E such that K(ξ) = ξ for all ξ ∈ V E.

The two definitions are related by setting HE = ker K.

Exercise 3.6. Show that every smooth fiber bundle admits a connection.
Hint: any local trivialization defines a natural connection in its neighbor-
hood. Use a partition of unity to piece together the connection maps. (See
the proof of Theorem 3.37 if you need more hints.)

Remark 3.7. The existence of connections for bundles on infinite dimen-

sional manifolds is a far more intricate problem, because such manifolds
do not generally admit smooth partitions of unity. However, more direct
constructions of connections succeed in many interesting cases, such as for
the “manifolds of maps” defined in [Elı̆67].

3.3 Connections on vector bundles

3.3.1 Three definitions
For a vector bundle π : E → M , some minor changes in the previous
discussion are appropriate in order to exploit the linear structure on the
fibers. Most importantly, it is no longer enough for the parallel transport
maps Pγt : Eγ(0) → Eγ(t) to be diffeomorphisms; they should be linear
isomorphisms. We thus define a linear connection to be any connection
on the fiber bundle E → M for which the induced parallel transport is
76 CHAPTER 3. CONNECTIONS

linear. It will always be assumed that a connection on a vector bundle is

a linear connection unless otherwise noted. We will prove the existence of
such objects later, in the context of principal bundles.
To see more concretely what linearity entails, observe that for any scalar
λ ∈ F, there is a fiberwise linear map

mλ : E → E : v 7→ λv,

which is a diffeomorphism if λ 6= 0. Choose a path γ(t) ∈ M , label

x = γ(0), X = γ̇(0) ∈ Tx M , and choose v ∈ Ex . Then for any linear
connection, we have Pγt (λv) = mλ (Pγt (v)), and differentiating at t = 0,

Horλv (X) = (mλ )∗ Horv (X).

This implies Hλv E = (mλ )∗ Hv E. Though it may not be obvious just yet,
this is enough of a criterion to identify linear connections. We shall prove
this below, after giving two new equivalent definitions.

Definition 3.8. A connection on the vector bundle π : E → M is a smooth

distribution HE on the total space such that HE ⊕ V E = T E and for any
scalar λ ∈ F and v ∈ E,

Hλv E = (mλ )∗ Hv E.

Observe that the vector space structure on each fiber Ex gives natural
isomorphisms

d 
Vertv : Ex → Vv E : w 7→ (v + tw)
dt t=0

for each v ∈ Ex . It is thus appropriate to rewrite the projection K :

T E → V E as a map K : T E → E that takes Tv E to Eπ(v) and satisfies
K(Vertv (w)) = w for all w ∈ Eπ(v) . This will be called a connection map
for the vector bundle π : E → M . Setting ker K = HE, it is an easy
exercise to verify that the following is now equivalent to Definition 3.8.

Definition 3.9. A connection on the vector bundle π : E → M is a smooth

map K : T E → E such that

1. For each v ∈ E, K defines a real linear map Tv E → Eπ(v) .

2. K(Vertv (w)) = w for all w ∈ Eπ(v) .

3. For all scalars λ ∈ F, K ◦ (mλ )∗ = mλ ◦ K.

We now show that these new definitions are equivalent to the notion of
a linear connection defined above. The following lemma will be of use.
3.3. VECTOR BUNDLES 77

Lemma 3.10. Let V and W be real, normed vector spaces. Then any map
F : V → W that is differentiable at 0 and satisfies F (λv) = λF (v) for all
scalars λ ∈ R and all v ∈ V is linear.

Proof. The key is to show that under this assumption, F is actually equal
to its derivative at zero, dF (0) : V → W . Clearly F (0) = 0, so we can
write
F (v) = dF (0)v + |v|η(v)
for some function η : V → W such that limv→0 η(v) = 0. Then

1 dF (0)λv + λ|v|η(λv)
F (v) = lim+ F (λv) = lim+
λ→0 λ λ→0 λ
= dF (0)v + lim+ |v|η(λv) = dF (0)v.
λ→0

Remark 3.11. The vector spaces V and W need not be finite dimensional—
in particular, they could be Banach spaces.

Proposition 3.12. If π : E → M is a vector bundle and K : T E → E

is a connection as defined above, then the induced parallel transport maps
Pγt : Eγ(0) → Eγ(t) are linear (with respect to F).

Proof. For any path γ(t) ∈ M , denote γ(0) = x, γ̇(0) = X ∈ Tx M , and

choose any v ∈ Ex , λ ∈ F. Denote by γ̃(t) ∈ E the horizontal lift of γ
with γ̃(0) = v, and similarly let γ̃λ (t) ∈ E denote the horizontal lift with
γ̃λ (0) = λv. We have,

d d
mλ (γ̃(t)) = (mλ )∗ γ̃(t) = (mλ )∗ Horγ̃(t) (γ̇(t))
dt dt
= Horλγ̃(t) (γ̇(t))
d
= γ̃λ (t),
dt
hence γ̃λ (t) ≡ λγ̃(t). This proves that the diffeomorphisms Pγt : Eγ(0) →
Eγ(t) satisfy
Pγt (λv) = λPγt (v) (3.8)
for all λ ∈ F, and by Lemma 3.10, Pγt is real linear. If F = C, it is clearly
also complex linear since (3.8) holds for λ ∈ C.
As before, the covariant derivative of a section s : M → E in the
direction X ∈ Tx M is defined by

d  t −1 
∇X s = (Pγ ) ◦ s(γ(t))  , (3.9)
dt t=0
78 CHAPTER 3. CONNECTIONS

where γ̇(0) = X, and ∇X s can now be regarded as a vector in Ex . In light

of the new definition for the connection map K : T E → E, we have also
∇X s = K ◦ T s(X). (3.10)
Similar remarks apply to sections along paths.
For any smooth section s : M → E and vector field X ∈ Vec(M ), ∇X s
now defines another section of E, while ∇s itself defines a smooth section
of the bundle of real linear maps HomR (T M, E) = T ∗ M ⊗R E. Using (3.9)
and the fact that Pγt is linear, one sees that the resulting map
∇ : Γ(E) → Γ(HomR (T M, E))
is also linear (with respect to F).
In this context, we have the following version of the product rule:
Proposition 3.13. For any section s : M → E and smooth function
f : M → F,
∇(f s) = df (·)s + f ∇s, (3.11)
where both sides are regarded as sections of HomR (T M, E).
This also follows easily from (3.9), using the linearity of parallel trans-
port. Formula (3.11) is called a Leibnitz rule for the operator ∇; such
relations appear naturally in any context that involves derivatives of bi-
linear products. We’ll see more examples in Chapter 4 when we define
connections on the tensor bundles associated with E.
Prop. 3.13 has a converse of sorts, which leads to a third equivalent
definition for linear connections. Suppose we have a vector bundle π :
E → M and an F-linear operator
D : Γ(E) → Γ(HomR (T M, E))
satisfying the Leibnitz rule D(f s) = df · s + f Ds. We denote DX s :=
Ds(x)X for X ∈ Tx M .
Proposition 3.14. Given the map D above, there is a unique connection
on π : E → M such that D = ∇.
The proof is based on the observation that any two operators satisfying
the same Leibnitz rule must differ by an operator which is tensorial :
Lemma 3.15. Suppose D, D 0 : Γ(E) → Γ(HomR (T M, E)) are two F-
linear operators that satisfy the Leibnitz rule (3.11) for all smooth functions
f : M → F and sections s ∈ Γ(E). Then the operator L : Vec(M )×Γ(E) →
Γ(E) defined by
0
L(X, s) = DX s − DX s
determines a bilinear bundle map T M ⊕ E → E. The map is real linear
in T M and F-linear in E.
3.3. VECTOR BUNDLES 79

Proof. We must verify that L : Vec(M ) × Γ(E) → Γ(E) is C ∞ -linear in s,

i.e. that L(X, f s) = f L(X, s) for all f ∈ C ∞ (M, F). Indeed,

0 0
L(X, f s) = DX (f s) − DX (f s) = df (X)s + f DX s − df (X)s − f DX s
0
= f (DX s − DX s) = f L(X, s).

Clearly also L(f X, s) = f L(X, s) for all smooth real-valued functions f .

This shows that for each x0 ∈ M , the value of L(X, s)(x0 ) depends only
on X(x0 ) and s(x0 ).

Proof of Prop. 3.14. Uniqueness is easy: if there is such a connection, then

the resulting horizontal lift maps Horv : Tx M → Tv E for v ∈ Ex must
satisfy
Horv (X) = T s(X) − Vertv (DX s)
for all X ∈ Tx M and s ∈ Γ(E) with s(x) = v. To prove existence, we must
verify that the right hand side of this expression gives a well defined linear
map Tx M → Tv E, regardless of the choice of section with s(x) = v.
We show this by choosing another connection K, e which induces a covari-
e
ant derivative operator ∇ : Γ(E) → Γ(HomR (T M, E)), satisfying the Leib-
nitz rule (3.11). By Lemma 3.15, there is a bundle map L : T M ⊕ E → E
such that L(X, s) ≡ ∇ e X s − DX s, and for X ∈ Tx M , we have

e X s) + Vertv (L(X, v))

T s(X) − Vertv (DX s) = T s(X) − Vertv (∇
g v (X) + Vertv (L(X, v)),
= Hor

g v : Tx M → Tv E denotes the horizontal lift map defined by K.

where Hor e
The right hand side now depends only on X and v, and gives a well defined
linear injection Tx M → Tv E. There is a unique connection K such that
this map is Horv for each v ∈ Ex , and by construction, ∇ = D.

We are now ready for the “quick and dirty” definition of linear connec-
tions that is most commonly found in modern introductions to differential
or Riemannian geometry. Prop. 3.14 shows that it is equivalent to our
previous two definitions.

Definition 3.16. A connection on the vector bundle π : E → M is an

F-linear operator
∇ : Γ(E) → Γ(HomR (T M, E))
satisfying the Leibnitz rule

∇(f s) = df (·)s + f ∇s

for all functions f ∈ C ∞ (M, F) and sections s ∈ Γ(E).

80 CHAPTER 3. CONNECTIONS

3.3.2 Christoffel symbols

Let π : E → M denote a vector bundle, and choose a local trivialization

Φ : E|U → U × Fm

over some open subset U ⊂ M . Using the trivialization we can identify

sections s : U → E|U with maps U → Fm . Then Φ determines a natural
connection on E|U , for which the covariant derivative acts on sections s :
U → Fm by s 7→ ds.
Given another connection ∇ : Γ(E) → Γ(HomR (T M, E)), Lemma 3.15
implies that ∇ differs from this natural connection on E|U by a bilinear
bundle map
ΓΦ : (T M ⊕ E)|U → EU ;
that is, for any section s ∈ Γ(E|U ) expressed as a map U → Fm , we have

∇s(x)X = ds(x)X + ΓΦ (X, s(x)) for x ∈ U . (3.12)

Note that ΓΦ is real linear in the first factor and F-linear in the second.
It must be emphasized that ΓΦ is not globally defined, and it depends on
the choice of trivialization. Of course we’re being somewhat sloppy with
notation; one can think of ΓΦ either as a bundle map on (T M ⊕ E)|U , or—
since we’re really working in a trivialization—as a bilinear map T M |U ×
Fm → F m .
One more often sees ΓΦ expressed in local coordinates as a set of locally
defined functions with three indices. Assume U admits a coordinate system
(x1 , . . . , xn ); this then determines a framing (∂1 , . . . , ∂n ) of the tangent
bundle T M |U , i.e. a set of linearly independent vector fields that span
the tangent space at each point. There is similarly a canonical framing
(e(1) , . . . , e(m) ) of E|U determined by Φ. Then the functions Γaib : U → F
are defined by
ΓΦ (∂i , e(b) ) = Γaib e(a) ,
so that for any X = X i ∂i ∈ Tx M and v = v b e(b) ∈ Ex , we have

ΓΦ (X, v) = ΓΦ (X i ∂i , v b e(b) ) = X i v b ΓΦ (∂i , e(b) ) = Γaib X i v b e(a) ,

i.e. (ΓΦ (X, v))a = Γaib X i v b . A section s : U → E|U can now be expressed
as s = sa e(a) for a set of m functions sa : U → F, and writing ∇i := ∇∂i ,
equation (3.12) becomes

(∇i s)a = ∂i sa + Γaib sb . (3.13)

In this context the functions Γaib are called Christoffel symbols. Standard
treatments of general relativity usually define connections purely in terms
3.3. VECTOR BUNDLES 81

of the Christoffel symbols, while the covariant derivative is defined by

(3.13).
In order for this definition of a connection to make sense, one must
have the right notion of how the symbols Γaib “transform” with respect to
changes in coordinates and local trivializations.
Exercise 3.17. Given a bundle π : E → M and a sufficiently small
open set U ⊂ M , use a coordinate system (x1 , . . . , xn ) and a framing
(e(1) , . . . , e(m) ) to identify E|U with the trivial bundle V × Fm , where V is
an open subset of Rn . Suppose Γaib are the corresponding Christoffel sym-
bols for some connection ∇ on E. Then another choice of coordinates and
framing can be expressed by smooth functions
V → Rn : (x1 , . . . , xn ) 7→ (x̃1 , . . . , x̃n )
V → Rm : (x1 , . . . , xn ) 7→ ẽ(1) = (ẽ1(1) , . . . , ẽm
(1) )
..
.
V → R : (x , . . . , xn ) 7→ ẽ(m) = (ẽ1(m) , . . . , ẽm
m 1
(m) )

e a denote the Christoffel symbols of ∇ with respect to (x̃1 , . . . , x̃n )

Let Γ ib
and (ẽ1 , . . . , ẽm ). Derive the transformation formula
j j
e a = ∂x ẽc Γa + ∂x ∂ ẽa .
Γ ib
∂ x̃i (b) jc ∂ x̃i ∂xj (b)
As a special case when E = T M , show that this becomes
 i
e i ∂xp ∂xq i ∂xp ∂ ∂x
Γjk = j k
Γpq + j p .
∂ x̃ ∂ x̃ ∂ x̃ ∂x ∂ x̃k
Remark 3.18. The definitions we’ve stated are not quite strict enough to
define connections on an infinite dimensional Banach space bundle. Sup-
pose for instance that the base M is a Banach manifold that looks locally
like the Banach space X, and E → M is a bundle with fibers isomorphic
to a Banach space Y. One must now explicitly require that for any choice
of smooth chart ϕ : U → X and local trivialization Φ : E|U → U × Y,
there is a smooth Christoffel map
ΓΦ : U → L(X ⊗R Y, Y).
In the infinite dimensional case, this is stricter than simply asking for the
map (x, X, v) 7→ ΓΦ (x, X, v) to be smooth; we saw an analogous situation
in the definition of a Banach space bundle (see Definition 2.60). This tech-
nical requirement is needed in order that the covariant derivative should
define a continuous linear map
∇ : Γ(E) → Γ(HomR (T M , E)).
We will continue to assume for the remainder of this discussion that all
objects are finite dimensional unless otherwise noted.
82 CHAPTER 3. CONNECTIONS

3.3.3 Connection 1-forms

As an alternative to the Christoffel symbols, one can express covariant
derivatives in local trivializations via matrix-valued 1-forms. If Uα ⊂ M is
an open subset and Φα : E|Uα → Uα × Fm is a trivialization, we can write
any smooth section s ∈ Γ(E) over Uα via the smooth map sα : Uα → Fm
such that
Φα ◦ s(x) = (x, sα (x)). (3.14)
Then for x ∈ Uα and X ∈ Tx M , the covariant derivative of s in the direction
of X can always be written in the form

(∇X s)α = dsα (X) + Aα (X)sα (x), (3.15)

where Aα is an m-by-m matrix-valued 1-form. The existence of such a

1-form is another easy consequence of Lemma 3.15; in fact, it’s not hard
to express Aα directly in terms of the Christoffel symbols:

Exercise 3.19. Choosing coordinates (x1 , . . . , xn ), let Ai = Aα (∂i ) and

denote the entries of this m-by-m matrix by (Ai )ab . Show that (Ai )ab = Γaib .

We call Aα the connection 1-form for ∇ with respect to the trivialization

Φα . This leads to yet another definition of connections that is somewhat
untidy but popular in the physics world: a connection is a choice of m-by-m
matrix-valued 1-forms Aα over Uα corresponding to each local trivialization
Φα and satisfying the appropriate transformation property with respect to
change of trivialization (see the exercise below).

Exercise 3.20. If g = gβα : Uα ∩ Uβ → GL(m, F) is the transition map

relating two trivializations Φα and Φβ , prove the transformation formula

Aα (X) = g −1 Aβ (X)g + g −1 dg(X). (3.16)

Physicists refer to (3.16) as a gauge transformation, alluding to the

important role that connection 1-forms play in quantum field theory: in
that context they are called gauge fields, and they serve to model elemen-
tary particles such as photons and other “gauge bosons” that mediate the
fundamental forces of nature. The choice of the letter A to denote a con-
nection form is in fact motivated by physics, where the vector potential of
classical electromagnetic field theory (conventionally denoted by A) can
be interpreted as a connection form for a trivial Hermitian line bundle.
There is another reason to use connection 1-forms rather than Christof-
fel symbols when the vector bundle has extra structure. In this case it’s
appropriate to restrict attention to a particular class of connections, and it
turns out that this restriction can be expressed elegantly via the connection
forms.
3.3. VECTOR BUNDLES 83

Definition 3.21. Let π : E → M be a vector bundle with a G-structure,

for some Lie group G. Then a connection ∇ on E is called G-compatible
if all parallel transport isomorphisms respect the G-structure: this means
that for any sufficiently short path γ(t) ∈ M , the maps Pγt : Eγ(0) → Eγ(t)
can be written in a G-compatible trivialization as

Pγt : Fm → Fm : v 7→ g(t)v

for some smooth map g(t) ∈ G with g(0) = .

The definition seems less abstract when we apply it to particular struc-
tures: e.g. for G = O(m) or U(m), the structure in question is a bundle
metric, and ∇ is called a metric connection if all parallel transport maps
are isometries. The terms complex connection and symplectic linear con-
nection can be defined analogously.
The existence of G-compatible connections will follow from our discus-
sion of principal bundles below. For now, we take existence as a given and
examine the consequences for connection 1-forms. Suppose in particular
that π : E → M is a vector bundle of rank m with a G-structure, for some
Lie subgroup G ⊂ GL(m, F)—as a concrete example to keep in mind, the
reader may assume G = O(m) and the G-structure is a bundle metric.
Assume Φα : E|Uα → Uα × Fm is a G-compatible trivialization (e.g. an
orthonormal frame), ∇ is a G-compatible connection, and γ(t) is a smooth
path in M with γ(0) = x ∈ Uα . Let s(t) ∈ Eγ(t) be a section along γ which
is parallel, i.e. ∇t s ≡ 0. Then (3.15) gives

ṡα (t) + Aα (γ̇(t))s(t) = 0. (3.17)

Since the trivialization and connection are both G-compatible, the fact
that s(t) is parallel also implies we can write sα (t) = g(t)sα (0) for some
smooth path of matrices g(t) ∈ G with g(0) = , thus Aα (γ̇(0)) = −ġ(0).
This cannot be just any arbitrary m-by-m matrix: the tangent space T G

is generally a proper subspace of the space of all matrices, called the Lie
algebra
T G = g ⊂ Fm×m

of the group G (see Appendix B). For example if G = O(m), then g =

o(m) is the space of antisymmetric matrices. This leads to a convenient
characterization of G-compatible connections.
Proposition 3.22. If E → M is a vector bundle with a G-structure and
∇ is a connection on E, then ∇ is G-compatible if and only if for every
G-compatible trivialization Φα , the corresponding connection 1-form

Aα ∈ Ω1 (Uα , Fm×m )

takes values in the Lie algebra g ⊂ Fm×m of G.

84 CHAPTER 3. CONNECTIONS

Proof. The argument above proves that G-compatibility implies Aα ∈

Ω1 (Uα , g). Conversely if the latter is true and v(t) ∈ Eγ(t) is any parallel
section along a smooth path γ(t) ∈ Uα , (3.17) implies

v̇α (t) = −Aα (γ̇(t))vα (t).

The result now follows from Exercise 3.23 below.

Exercise 3.23. For any matrix Lie group G ⊂ GL(m, F) and a smooth
path of matrices A(t) ∈ g, show that the unique solution Φ(t) ∈ Fm×m to
the initial value problem
(
Φ̇(t) = A(t)Φ(t),
Φ(0) =

satisfies Φ(t) ∈ G. Hint: show first that for any A ∈ g and B ∈ G,

AB ∈ TB G; then Φ(t) is an orbit of a time-dependent vector field on G.

3.3.4 Linearization of a section at a zero

The following observation is trivial but useful:
Proposition 3.24. Suppose π : E → M is a smooth vector bundle and
s : M → E is a smooth section with s(x) = 0. Then the linear map

∇s(x) : Tx M → Ex

is independent of the choice of connection.

The proof is easy if we view it in the right context: recall that every vec-
tor bundle has a preferred embedding M ,→ E, the zero section. One sees
easily from the condition Hλv E = (mλ )∗ Hv E that horizontal subspaces in
T E are always tangent to M —thus every linear connection looks identical
along the zero section. Put another way, at any point x ∈ M ⊂ E, which
we view as lying either in M or in the zero section of E, there is a natural
isomorphism
Tx E = Ex ⊕ Tx M,
and any connection map K : T E → E defines the projection to the first
factor. Thus the expression ∇s(x) = K ◦ ds(x) is invariantly defined as
long as s(x) = 0.
We call ∇s(x) : Tx M → Ex the linearization of s at x ∈ s−1 (0), and
since it doesn’t depend on ∇, we may as well denote

ds(x) := ∇s(x).

The definition of ds(x) via a connection is often convenient, but not nec-
essary, as the next result shows.
3.3. VECTOR BUNDLES 85

Proposition 3.25. Suppose the section s : M → E has a zero at x, and we

choose a trivialization on some neighborhood x ∈ U ⊂ M so as to identify
sections with maps f : U → Fm . Then the linearization at x is expressed in
this trivialization as df (x) : Tx M → Fm , and the resulting map Tx M → Ex
is independent of the chosen trivialization.

Proof. Differentiation in a trivialization can be viewed simply as covariant

differentiation with respect to a connection determined by the trivializa-
tion. The result then follows from Proposition 3.24. (Alternatively, one
can prove this by a direct computation without mentioning connections at
all. Try it if you have a moment to spare.)
In the special case E = T M , there is a nice way to write the lineariza-
tion without any arbitrary choices. A section in this case is a vector field
X ∈ Vec(M ), and it determines a flow ϕt : M → M ; these diffeomor-
phisms may not be globally defined if M is noncompact, but they are at
least defined for t close to 0 and x in a neighborhood of any point x0 with
X(x0 ) = 0. In particular, ϕt (x0 ) = x0 for all t, and there is a corresponding
smooth family of linear maps

dϕt (x0 ) : Tx0 M → Tx0 M.

These determine a differentiable path through the identity in the vector

space End(Tx0 M ), and as it turns out,

d t 
dϕ (x0 ) = dX(x0 ). (3.18)
dt t=0

Exercise 3.26. Use a coordinate system around x0 to prove (3.18).

Another example of this construction is the Hessian of a smooth func-

tion f : M → R at a critical point. In general, the first derivatives of
f are easily characterized via the differential df ∈ Ω1 (M ), but without a
connection there is no such simple coordinate-invariant construction that
describes its second derivatives. If a connection ∇ is chosen on T ∗ M , then
this information is of course contained in the tensor field ∇df ∈ Γ(T20 M ).
We observe now that for all critical points x ∈ M of f , i.e. points where
dfx : Tx M → R is the zero map, the bilinear form

∇dfx : Tx M × Tx M → R

doesn’t depend on the choice of connection. We call this bilinear map the
Hessian of f at x.

Proposition 3.27. For any critical point x of f ∈ C ∞ (M ), the Hessian

∇dfx is symmetric, i.e. ∇dfx (X, Y ) = ∇dfx (Y, X) for all X, Y ∈ Tx M .
86 CHAPTER 3. CONNECTIONS

Proof. Since the bilinear form doesn’t depend on ∇, we can choose co-
ordinates (x1 , . . . , xn ) on a neighborhood U of x and define ∇ on T ∗ M |U
so that the sections dxj are parallel. Then identifying Tx M with Rn via
the coordinates, we have ∇fx (v, w) = vT Hw where H is the symmetric
n-by-n matrix
 ∂2f ∂2f

∂x1 ∂x1
· · · ∂xn ∂x1
 ..  .
H =  ... ..
. . 
∂2f ∂2f
∂x1 ∂xn
··· ∂xn ∂xn

Remark 3.28. For every proof that uses coordinates there is a cleverer proof
that avoids them: we’ll see in Chapter 4 how to give a coordinate-free proof
of Prop. 3.27 using symmetric connections.
To see why the linearization is useful, we need some basic facts about
transversality; more details can be found in [Hir94]. Suppose M is a smooth
manifold with two smooth submanifolds N1 and N2 . We say that N1 and
N2 are transverse in M , written N1 t N2 , if for every intersection point
x ∈ N1 ∩ N2 ,
Tx N1 + Tx N2 = Tx M.
The expression on the left means all vectors that can be written as X + Y
for X ∈ Tx N1 and Y ∈ Tx N2 . At any point of transverse intersection,
the subspace Tx N1 ∩ Tx N2 ⊂ Tx M has the smallest possible dimension,
determined by the simple formula

codim(Tx N1 ∩ Tx N2 ) = codim Tx N1 + codim Tx N2 ,

where the codimension of a subspace V ⊂ Tx M is defined as codim V =

dim Tx M − dim V .
Similarly, we define the codimension of a smooth submanifold N ⊂ M
by codim N = dim M − dim N . The key fact about transversality is the
following result, which can be proved via the implicit function theorem:

Proposition 3.29. Suppose M is a manifold without boundary, contain-

ing two transverse submanifolds N1 and N2 without boundary. Then the
intersection N1 ∩ N2 is a smooth submanifold of M , with

codim(N1 ∩ N2 ) = codim N1 + codim N2 .

Thus transversely intersecting submanifolds induce a very nice struc-

ture on their intersection. Not much can be said in general about subman-
ifolds that are not transverse, but the following basic result of differential
topology permits us to avoid worrying about it much of the time.
3.3. VECTOR BUNDLES 87

Proposition 3.30. Given any two smooth submanifolds N1 ⊂ M and

N2 ⊂ M without boundary, one can move N1 by an arbitrarily small per-
turbation so that N1 t N2 .

One often abbreviates this by saying that generic submanifolds intersect

transversely.1 We refer to [Hir94] for the proof, but mention a special case
in which the result is easy to visualize: suppose N1 and N2 are smooth
curves in the plane M = R2 . Then transversality can only fail if N1 and N2
have a point of tangent intersection. One can always change the direction
of N1 just slightly to kill the tangency.
Notice that transverse intersection points are impossible unless dim N1 +
dim N2 ≥ dim M ; in this case Proposition 3.30 says that one can perturb
N1 so that the two submanifolds have no intersection at all. This is the
case for instance with a pair of smooth curves in R3 . It is also the reason
why airplanes generically do not crash into each other.
Now, suppose π : E → M is a smooth vector bundle of rank m over
an n-dimensional manifold. Any smooth section s : M → E then defines
a submanifold s(M ) ⊂ E of codimension m, and in particular there is
the special submanifold M ⊂ E defined by the zero section. Using this
notation, we say that a section s ∈ Γ(E) is transverse to the zero section
if s(M ) t M .

Theorem 3.31. A smooth section s : M → E is transverse to the zero

section if and only if for every zero x ∈ s−1 (0), the linearization ds(x) :
Tx M → Ex is a surjective map. In this case the zero set s−1 (0) ⊂ M
is a smooth submanifold, with dimension equal to dimR ker ds(x) for any
x ∈ s−1 (0).

Proof. Identify M with the zero section so that we can treat it as a sub-
manifold of E. Then if s(x) = 0, transversality is achieved if and only
if
im T s|Tx M + Tx M = Tx E.
Identifying Tx E with Ex ⊕Tx M in the canonical way, it’s equivalent that the
projection K : Tx E → Ex should map im T s|Tx M onto Ex , which means
K ◦ T s|Tx M = ∇s(x) is surjective. The dimension formula is a simple
exercise.

Note that by Proposition 3.30, sections s : M → E are generically

transverse to the zero section.
1
The word “generic” has a variety of precise meanings in different mathematical
contexts, but it always refers to a situation in which one expects things to appear a
certain way, and the alternative is somehow an exceptional case. For example, a generic
pair of vectors in R2 is linearly independent, just as a generic matrix is invertible. Given
a non-generic situation, one can always achieve the generic case by a small perturbation.
88 CHAPTER 3. CONNECTIONS

Remark 3.32. One can alternatively use a more straightforward implicit

function theorem argument to show that s−1 (0) ⊂ M is a smooth subman-
ifold if and only if ds(x) is surjective for all x ∈ s−1 (0). A similar statement
is true for smooth maps f : Rn → RN ; this case can be reduced to that
one by choosing local coordinates and trivializations.

3.4 Connections on principal bundles

Before plunging into the definitions for principal connections, let us provide
some motivation. As we saw in Section 2.8, one can always use a principal
bundle to encode the essential data of a vector bundle and any additional
structure that’s attached to it. Thus instead of dealing with the assorted
variety of structures such as bundle metrics or symplectic structures that
might be associated with a vector bundle, one can describe all of these
in a unified way by identifying the structure group and defining the cor-
responding frame bundle. A connection on the frame bundle will induce
a connection on the original bundle, with parallel transport that respects
any special structure that may be present. Indeed, the same is true for
more general fiber bundles with finite dimensional structure groups. Thus
the construction of connections on principal bundles has implications that
go well beyond the study of principal bundles themselves.

3.4.1 Definition
Let G be a Lie group, and π : E → M a principal G-bundle. Recall
that the fibers of E have intrinsic structure in the form of a smooth fiber
preserving right group action

E × G → E : (p, g) 7→ pg,

which is free and transitive on each fiber. A connection on E → M is

called a principal connection if the parallel transport diffeomorphisms Pγt :
Eγ(0) → Eγ(t) are G-equivariant, that is,

Pγt (pg) = Pγt (p)g

for all p ∈ Eγ(0) and g ∈ G.

As usual, this first definition is conceptually simple but hard to work
with in practice, so we’ll give some equivalent definitions. The first step
is to determine the implications of G-equivariance for the horizontal sub-
bundle HE ⊂ T E. Each g ∈ G determines a fiber preserving diffeomor-
phism Rg : E → E : p 7→ pg. Choose a path γ(t) ∈ M with γ(0) = x,
γ̇(0) = X ∈ Tx M , and let p ∈ Ex . Given a principal connection, the
3.4. PRINCIPAL BUNDLES 89

horizontal lift isomorphisms have the property

 
d t  d 
Horpg (X) = Pγ (pg) = Rg ◦ Pγ (p) t
dt t=0 dt t=0 
d t 
= (Rg )∗ Pγ (p) = (Rg )∗ Horp (X).
dt t=0

We conclude Hpg E = (Rg )∗ Hp E. Conversely, any fiber bundle connection

with this property defines G-equivariant parallel transport; one can prove
this along the same lines as for vector bundles. The following is therefore
an equivalent definition for principal connections.
Definition 3.33. A connection on the principal fiber bundle π : E → M
with structure group G is a smooth distribution HE on the total space
such that HE ⊕ V E = T E and for any g ∈ G and p ∈ E,
Hpg E = (Rg )∗ Hp E.

3.4.2 Global connection 1-forms

The connection map K : T E → V E for a principal G-bundle E → M
can be put in a simplified form by noting that in this situation the vertical
bundle V E → E is necessarily trivial; in fact, we will now show that it has
a preferred trivialization. Recall that every Lie group G has an associated
Lie algebra g, which is the tangent space Te G with a bracket operation
[ , ] : g × g → g induced by the group multiplication on G. There is also
a so-called exponential map
exp : g → G
defined such that for each X ∈ g, t 7→ exp(tX) is the unique Lie group
homomorphism R → G with dtd exp(tX)t=0 = X. For matrix groups G ⊂
GL(m, F) in particular, where g is naturally a subspace of Fm×m , it turns
out that [A, B] = AB − BA and exp(tA) = etA is defined by the familiar
power series expansion. (These concepts are reviewed in Appendix B). Now
if E → M is a principle G-bundle, the right action E × G → E defines a
natural bundle isomorphism

d 
E × g → V E : (p, X) 7→ X(p) = (p exp(tX)) . (3.19)
dt t=0

The vertical vector field on E defined by X(p) is called the fundamental

vector field determined by X ∈ g.
If HE ⊂ T E is a principal connection, we can use the isomorphism
(3.19) to rewrite the connection map K : T E → V E as a g-valued 1-form
A ∈ Ω(E, g) := Γ(Hom(T E, E × g)),
90 CHAPTER 3. CONNECTIONS

giving for each p ∈ E a linear map Ap : Tp E → g, such that ker Ap =

Hp E. Since K is a projection onto V E, A has the property A(X(p)) = X
for all p ∈ E and X ∈ g. Recall also that HE is required to satisfy
Hpg E = (Rg )∗ Hp E for all p ∈ E and g ∈ G. To express the consequence of
this condition for A, we need the following fact about fundamental vector
fields. Recall from Appendix B the definition of the adjoint representation
Ad : G → Aut(g) : g 7→ Adg , which for matrix groups takes the form

AdB (A) = BAB−1 .

Lemma 3.34. If g ∈ G and X ∈ g, then

(Rg )∗ X = Adg−1 (X).

Proof. For p ∈ E, compute

 
d  d 
((Rg )∗ X)(pg) = Rg (p exp(tX)) = p exp(tX)g 

dt t=0 dt t=0 
d
 d 
= pg g exp(tX)g  = pg exp (t Adg−1 (X))
−1
dt t=0 dt t=0
= Adg−1 (X)(pg).

Any vector ξ ∈ Tp E can be written uniquely as ξ = ξh + X(p) where

ξh ∈ Hp E and X ∈ g. Then A(ξ) = X, and Hpg E = (Rg )∗ Hp E implies

Rg∗ A(ξ) = A((Rg )∗ (ξh + X(p))) = A(Adg−1 (X)(pg)) = Adg−1 (X)

= Adg−1 ◦A(ξ).

Conversely, it’s not hard to show (Exercise 3.36 below) that any g-valued
1-form satisfying these conditions defines a principal connection by HE =
ker A. We therefore have a useful new definition for principal connections.
Definition 3.35. A connection on the principal fiber bundle π : E → M
with structure group G is a smooth g-valued 1-form A ∈ Ω(E, g) such that:
(i) A(X(p)) = X for all X ∈ g and p ∈ E,
(ii) (Rg )∗ A = Adg−1 ◦A for all g ∈ G.
Exercise 3.36. Show that if A ∈ Ω(E, g) satisfies the conditions in Def-
inition 3.35, then the distribution HE = ker g satisfies the conditions in
Definition 3.33.
Defining connections in terms of global 1-forms has several advantages,
the first of which is that proving existence is now a simple exercise with
partitions of unity.
3.4. PRINCIPAL BUNDLES 91

Theorem 3.37. Every principal fiber bundle admits a connection.

Proof. Assume π : E → M is a principal G-bundle and {(Uα , Φα )} is a
system of local trivializations. Since M is paracompact, we can replace
{Uα } with a locally finite refinement and choose a smooth partition of
unity {ϕα }. Each trivialization Φα : π −1 (Uα ) → Uα × G defines an obvious
notion of G-equivariant parallel transport within Uα , and thus a connection
Aα ∈ Ω(Uα , g) over Uα . We use these to define a global g-valued 1-form
X
A= ϕα A α .
α

We claim that A is a connection. For X ∈ g, x ∈ M and p ∈ Ex , we have

X X
A(X(p)) = ϕα (x)Aα (X(p)) = ϕα (x)X = X.
α α

Likewise for g ∈ G,
X X
(Rg )∗ A = ϕα · (Rg )∗ Aα = ϕα · Adg−1 Aα
α α
X
= Adg−1 ϕα Aα = Adg−1 A,
α

proving the claim.

3.4.3 Frame bundles and linear connections

Theorem 3.37 is more than an existence result for principal connections:
it also implies the existence of compatible connections on any fiber bundle
with a finite dimensional structure group. We now prove this in particular
for linear connections on vector bundles.
Recall from Example 2.81 in Chapter 2 that for every vector bundle
E → M with structure group G, there is an associated frame bundle
F G E → M , a principal G-bundle whose fibers are spaces of preferred
bases for the fibers of E. Each element of the fiber F G Ex thus defines
a unique isomorphism Fm → Ex , so there is a smooth, fiber preserving
inclusion map
Ψ : F G E → Hom(M × Fm , E)
taking each p ∈ F G Ex to the corresponding isomorphism Ψ(p) : Fm → Ex .
Treating any g ∈ G as a linear map on Fm , we have

Ψ(pg) = Ψ(p) ◦ g.

There is also a a “vertical tangent map”

V Ψ : V F G E → Hom(M × Fm , E)
92 CHAPTER 3. CONNECTIONS

such that for any smooth path p(t) ∈ F G Ex ,

d
V Ψ(ṗ(t)) = Ψ(p(t)),
dt
where the right hand side is interpreted as the derivative of a smooth path
in the vector space Hom(Fm , Ex ).

Theorem 3.38. Suppose π : E → M is a smooth vector bundle, G is

a Lie group and E → M is equipped with a G-structure. Then the set of
G-compatible linear connections on E is in one-to-one correspondence with
the set of principal connections on the frame bundle F G E → M .
In particular, given a connection on F G E there is a unique connection
on E such that for every smooth path γ(t) in M , the parallel sections of
E along γ are of the form v(t) = Ψ(s(t))v, where s(t) is a parallel section
of F G E along γ and v ∈ Fm is constant. This connection on E is G-
compatible, and for any smooth section s(t) ∈ F G Eγ(t) and smooth map
v(t) ∈ Fm , the section Ψ(s)v ∈ Eγ(t) satisfies

∇t (Ψ(s)v) = V Ψ(∇t s)v + Ψ(s)∂t v. (3.20)

Proof. If a connection on F G E is given, then for a smooth path γ(t) ∈ M

with γ(0) = x and γ̇(t) = X ∈ Tx M , we pick any frame p ∈ F G Ex and
define a horizontal lift map by

d 
HorΨ(p)v : Tx M → TΨ(p)v E : X 7→ Ψ(Pγ (p))v
t
dt t=0

for every v ∈ Fm . Any other choice of p ∈ F G Ex gives the same result

due to the G-equivariance of Pγt . The images of all the horizontal lift maps
define a distribution HE, which is clearly a fiber bundle connection on E.
The fact that it is G-compatible follows from the observation that each
p ∈ F G Ex defines a G-compatible trivialization of E along γ by
 
−1 
m t
Eγ(t) → {γ(t)} × F : v 7→ γ(t), Ψ Pγ (p) (v) ;

in this trivialization, parallel transport in E along γ is the identity. Note

that since G acts linearly on Fm , the resulting connection on E is auto-
matically linear.
Conversely if E has a G-compatible connection with parallel transport
Qtγ : Eγ(0) → Eγ(t) along γ, there is a unique principal connection on F G E
such that for each p ∈ F Ex , the expression

Ψ(p(t))v = Qtγ ◦ Ψ(p)v

extends p to a parallel section p(t) of F G E along γ.

3.4. PRINCIPAL BUNDLES 93

Equation (3.20) now follows from the definition of covariant differenti-

ation along a path: we have

d t −1 
∇t (Ψ(s)v)|t=0 = (Qγ ) (Ψ (s(t)) v(t))
dt t=0

d 
 
= Ψ [Pγt ]−1 (s(t)) v(t) 
dt
   t=0 
d t −1  d 
=VΨ (Pγ ) (s(t))  v(0) + Ψ(s(0)) v(t)
dt t=0 dt t=0
= V Ψ(∇t s)|t=0 · v(0) + Ψ(s(0))v̇(0).

Corollary 3.39. Every smooth vector bundle with a G-structure admits a

G-compatible connection.
This proves the existence of metric connections, complex connections
and linear symplectic connections on every bundle with the corresponding
structure. The beauty of the principal bundle formalism is that all of these
existence results follow from a single construction. We obtain also a conve-
nient relation between the global connection 1-form A ∈ Ω1 (F G E, g) and
the local connection forms Aα ∈ Ω1 (Uα , g) corresponding to trivializations.
Proposition 3.40. Suppose A ∈ Ω1 (F G E, g) is a connection on F G E. Let
Φα : E|Uα → Uα × Fm denote any local trivialization of E, with correspond-
ing local connection 1-form Aα ∈ Ω1 (Uα , g), and denote by sα : Uα → F G E
the unique local section of F G E such that
Φ−1
α (x, v) = Ψ(sα (x))v.

Then
Aα = s∗α A
Proof. Let x ∈ Uα and X ∈ Tx M . The form Aα is defined such that for
any smooth map v : Uα → Fm ,
∇X (Ψ(sα )v) = Ψ(sα ) (dv(X) + Aα (X)v) ,
whereas (3.20) gives
∇X (Ψ(sα )v) = V Ψ(∇X sα )v + Ψ(sα )dv(X),
implying V Ψ(∇X sα )v = Ψ(sα )Aα (X)v. Recalling from (3.19) the defini-
tion of the fundamental vector field, we have
∇X sα = K(T sα (X)) = A(T sα (X))(sα (x)) = s∗α A(X)(sα (x))

d 
= [sα (x) exp(tsα A(X))] ,
∗
dt t=0
94 CHAPTER 3. CONNECTIONS

and thus

d 
V Ψ(∇X sα )v = Ψ (sα (x) exp(tsα A(X))) v
∗
dt
 t=0
d 
= Ψ(sα (x)) exp(ts∗α A(X))v = Ψ(sα (x)) ◦ s∗α A(X)v.
dt t=0

We conclude s∗α A(X) = Aα (X).

Remark 3.41. One can define frame bundles not just for vector bundles
but for any smooth fiber bundle E → M with a finite dimensional struc-
ture group G: the frame bundle is then a principal G-bundle whose fibers
consist of preferred diffeomorphisms between fibers of E and the standard
fiber F . With this notion, the existence result above generalizes nicely
to a construction of G-compatible connections on any such fiber bundle.
The only limitation is that this argument applies only to fiber bundles
with finite dimensional structure groups—extending it beyond this setting
would require considerably more analytical effort. Fortunately one can of-
ten prove existence by more direct means in particular cases of interest,
e.g. symplectic fibrations (cf. [MS98]).

References
[Elı̆67] H. I. Elı̆asson, Geometry of manifolds of maps, J. Differential Geometry 1
(1967), 169–194.
[Hir94] M. W. Hirsch, Differential topology, Springer-Verlag, New York, 1994.
[MS98] D. McDuff and D. Salamon, Introduction to symplectic topology, The Clarendon
Press Oxford University Press, New York, 1998.