Chapter 2

The Tangent Structure

In this chapter we introduce the notions of tangent space and cotangent

space of a smooth manifold. The union of the tangent spaces of a given
manifold will be given a smooth structure making this union a manifold
in its own right, called the tangent bundle. Similarly we introduce the
cotangent bundle of a smooth manifold. We then discuss vector fields and
their integral curves together with the associated dynamic notions of Lie
derivative and Lie bracket. Finally, we define and discuss the notion of a
1-form (or covector field), which is the notion dual to the notion of a vector
field. One can integrate 1-forms along curves. Such an integration is called a
line integral. We explore the concept of exact 1-forms and nonexact 1-forms
and their relation to the question of path independence of line integrals.

2.1. The Tangent Space

If c : (−, ) → RN is a smooth curve, then it is common to visualize the
“velocity vector” ċ(0) as being based at the point p = c(0). It is often
desirable to explicitly form a separate N -dimensional vector space for each
point p, whose elements are to be thought of as being based at p. One
way to do this is to use {p} × RN so that a tangent vector based at p is
taken to be a pair (p, v) where v ∈ RN . The set {p} × RN inherits a vector
space structure from RN in the obvious way. In this context, we provisionally
denote {p}×RN by Tp RN and refer to it as the tangent space at p. If we write
c(t) = (x1 (t), . . . , xN (t)), then the velocity vector of a curve c at time1 t = 0
1 dxN
is (p, dx
dt (0), . . . , dt (0)), which is based at p = c(0). Ambiguously, both

1 It is common to refer to the parameter t for a curve as “time”, although it may have nothing

to do with physical time in a given situation.

55
56 2. The Tangent Structure

1 N 1 N
(p, dx dx dx dx
dt (0), . . . , dt (0)) and ( dt (0), . . . , dt (0)) are often denoted by ċ(0) or
c (0). A bit more generally, if V is a finite-dimensional vector space, then V
is a smooth manifold and the tangent space at p ∈ V can be provisionally
taken to be the set {p} × V. We use the notation vp := (p, v). If vp := (p, v)
is a tangent vector at p, then v is called the principal part of vp .
We have a natural isomorphism between RN and Tp RN given by v →
(p, v), for any p. Of course we also have a natural isomorphism Tp RN ∼ =
Tq RN for any pair of points given by (p, v) → (q, v). This is sometimes
referred to as distant parallelism. Here we see the reason that in the
context of calculus on RN , the explicit construction of vectors based at a
point is often deemed unnecessary. However, from the point of view of
manifold theory, the tangent space at a point is a fundamental construction.
We will define the notion of a tangent space at a point of a differentiable
manifold, and it will be seen that there is, in general, no canonical way to
identify tangent spaces at different points.
Actually, we shall give several (ultimately equivalent) definitions of the
tangent space. Let us start with the special case of a submanifold of RN . A
tangent vector at p can be variously thought of as the velocity of a curve,
as a direction for a directional derivative, and also as a geometric object
which has components that depend in a special way on the coordinates
used. Let us explore these aspects in the case of a submanifold of RN . If
M is an n-dimensional regular submanifold of RN , then a smooth curve
c : (−, ) → M is also a smooth curve into RN and ċ(0) is normally thought
of as a vector based at the point p = c(0). This vector is tangent to M . The
set of all vectors obtained in this way from curves into M is an n-dimensional
subspace of the tangent space of RN at p (described above). In this special
case, this subspace could play the role of the tangent space of M at p. Let
us tentatively accept this definition of the tangent space at p and denote it
by Tp M .
Let vp := (p, v) ∈ Tp M . There are three things we should notice about
vp . First, there are many different curves c : (−, ) → M with c(0) = p
which all give the same tangent vector vp , and there is an obvious equivalence
relation among these curves: two curves passing through p at t = 0 are
equivalent if they have the same velocity vector. Already one can see that
perhaps this could be turned around so that we can think of a tangent vector
as an equivalence class of curves. Curves would be equivalent if they agree
infinitesimally in some appropriate sense.
The second thing that we wish to bring out is that a tangent vector can
be used to construct a directional derivative operator. If vp = (p, v) is a
tangent vector in Tp RN , then we have a directional derivative operator at p
which is a map C ∞ (RN ) → R given by f → Df (p)v. Now if vp is tangent
2.1. The Tangent Space 57

to M , we would like a similar map C ∞ (M ) → R. If f is only defined on M ,

then we do not have Df to work with but we can just take our directional
derivative to be the map given by
Dvp : f → (f ◦ c) (0),
where c : I → M is any curve whose velocity at t = 0 is vp . Later we use
the abstract properties of such a directional derivative to actually define the
notion of a tangent vector.
Finally, notice how vp relates to charts for the submanifold. If (U, y) is
a chart on M with p ∈ U , then by inverting we obtain a map y−1 : V → M ,
which we may then think of as a map into the ambient space RN . The map
y−1 parameterizes a portion of M . For convenience, let us suppose that
y−1 (0) = p. Then we have the “coordinate curves” y i → y−1 (0, . . . , y i , . . . , 0)
for i = 1, . . . , n. The resulting tangent vectors Ei at p have principal parts
given by the partial derivatives so that
 
∂y−1
Ei := p, (0) .
∂y i
It can be shown that (E1 , . . . , En ) is a basis for Tp M . For another coordinate
system ȳ withȳ−1 (0) = p, we similarly define a basis (Ē1 , . . . , Ēn ). If vp =
 n i n i 1 n 1 n
i=1 a Ei = i=1 ā Ēi , then letting a = (a , . . . , a ) and ā = (ā , . . . , ā ),
the chain rule can be used to show that
ā = D(ȳ ◦ y−1 ) y(p)
a,
which is classically written as


n
∂ ȳ i j
āi = a .
∂y j
j=1

Both (a1 , . . . , an )
and (ā1 , . . . , ān )
represent the tangent vector vp , but with
respect to different charts. This is a simple example of a transformation
law.
The various definitions for the notion of a tangent vector given below in
the general setting will be based in turn on the following three ideas: (1)
Equivalence classes of curves through a point. (2) Transformation laws for
the components of a tangent vector with respect to various charts. (3) The
idea of a “derivation” which is a kind of abstract directional derivative. Of
course we will also have to show how to relate these various definitions to
see that they are really equivalent.

2.1.1. Tangent space via curves. Let p be a point in a smooth n-

manifold M . Suppose that we have smooth curves c1 and c2 mapping into M ,
each with open interval domains containing 0 ∈ R and with c1 (0) = c2 (0) =
p. We say that c1 is tangent to c2 at p if for all smooth real-valued functions
58 2. The Tangent Structure

f defined on an open neighborhood of p, we have (f ◦ c1 ) (0) = (f ◦ c2 ) (0).

This is an equivalence relation on the set of all such curves. The reader
should check that this really is an equivalence relation and also do so when
we introduce other simple equivalence relations later. Define a tangent
vector at p to be an equivalence class under this relation.
Notation 2.1. The equivalence class of c will be denoted by [c], but we also
denote tangent vectors by notation such as vp or Xp , etc. Eventually we will
often denote tangent vectors simply as v, w, etc., but for the discussion to
follow we reserve these letters without the subscript for elements of Rn for
some n.

If vp = [c] then we will also write ċ(0) = vp . The tangent space Tp M is

defined to be the set of all tangent vectors at p ∈ M . A simple cut-off func-
tion argument shows that c1 is equivalent to c2 if and only if (f ◦ c1 ) (0) =
(f ◦ c2 ) (0) for all globally defined smooth functions f : M → R.
Lemma 2.2. c1 is tangent to c2 at p if and only if (f ◦ c1 ) (0) = (f ◦ c2 ) (0)
for all Rk -valued functions f defined on an open neighborhood of p.

Proof. If f = (f 1 , . . . , f n ), then (f ◦ c1 ) (0) = (f ◦ c2 ) (0) if and only if

 i   
f ◦ c1 (0) = f i ◦ c2 (0) for i = 1, . . . , k. Thus (f ◦ c1 ) (0) = (f ◦ c2 ) (0)
if c1 is tangent to c2 at p. Conversely, let g be a smooth real-valued function
defined on an open neighborhood of p and consider the map f = (g, 0, . . . , 0).
Then the equality (f ◦ c1 ) (0) = (f ◦ c2 ) (0) implies that (g ◦ c1 ) (0) =
(g ◦ c2 ) (0).

The definition of tangent space just given is very geometric, but it has
one disadvantage. Namely, it is not immediately obvious that Tp M is a
vector space in a natural way. The following principle is used to obtain a
vector space structure:
Proposition 2.3 (Consistent transfer of linear structure). Suppose that S
is a set and {Vα }α∈A is a family of n-dimensional vector spaces. Suppose
that for each α we have a bijection bα : Vα → S. If for every α, β ∈ A the
map b−1
β ◦bα : Vα → Vβ is a linear isomorphism, then there is a unique vector
space structure on the set S such that each bα is a linear isomorphism.

Proof. Define addition in S by s1 + s2 := bα (b−1 −1

α (s1 ) + bα (s2 )). This
definition is independent of the choice of α. Indeed,
−1 −1
bα (b−1 −1 −1 −1
α (s1 ) + bα (s2 )) = bα [bα ◦ bβ ◦ bβ (s1 ) + bα ◦ bβ ◦ bβ (s2 )]
 
−1 −1
= bα ◦ b−1
α ◦ b β bβ (s1 ) + bβ (s2 )
 
= bβ b−1 −1
β (s1 ) + bβ (s2 ) .
2.1. The Tangent Space 59

The definition of scalar multiplication is a · s := bα (a b−1

α (s)), and this is
shown to be independent of α in a similar way. The axioms of a vector
space are satisfied precisely because they are satisfied by each Vα .

We will use the above proposition to show that there is a natural vector
space structure on Tp M . For every chart (xα , Uα ) with p ∈ U , we have a
map bα : Rn → Tp M given by v → [γv ], where γv : t → x−1
α (xα (p) + tv) for t
in a sufficiently small but otherwise irrelevant interval containing 0.
Lemma 2.4. Foreach chart
 (Uα , xα ), the map bα : Rn → Tp M is a bijection
and b−1 −1 (x (p)).
β ◦ bα = D xβ ◦ xα α

Proof. We have
d
(xα ◦ γv ) (0) = xα ◦ x−1
α (xα (p) + tv)
dt t=0
d
= (xα (p) + tv) = v.
dt t=0
Suppose that [γv ] = [γw ] for v, w ∈ Rn . Then by Lemma 2.2 we have
v = (xα ◦ γv ) (0) = (xα ◦ γw ) (0) = w.


This means that bα is injective.

Next we show that bα is surjective. Let [c] ∈ Tp M be represented by
c : (−, ) → M . Let v := (xα ◦ c) (0) ∈ Rn . Then we have bα (v) = [γv ],
where γv : t → x−1α (xα (p)+tv). But [γv ] = [c] since for any smooth f defined
near p we have
d  
(f ◦ γv ) (0) = f ◦ x−1
α (xα (p) + tv) = D f ◦ xα
−1
(xα (p)) · v
dt t=0
 
= D f ◦ x−1
α (xα (p)) · (xα ◦ c) (0) = (f ◦ c) (0).
Thus bα is surjective. From Lemma 2.2 we see that the map [c] → (xα ◦ c) (0)
is well-defined, and from the above we see that this map is exactly b−1
α . Thus
d  
b−1
β ◦ bα (v) = xβ ◦ x−1 −1
α (xα (p) + tv)) = D xβ ◦ xα (xα (p))v.

dt t=0

The above lemma and proposition combine to provide a vector space

structure on the set of tangent vectors. Let us temporarily call the tan-
gent space defined above, the kinematic tangent space and denote it by
(Tp M )kin . Thus, if Cp is the set of smooth curves c defined on some open
interval containing 0 such that c(0) = p, then
(Tp M )kin = Cp /∼,
where the equivalence is as described above.
60 2. The Tangent Structure

Exercise 2.5. Let c1 and c2 be smooth curves mapping into a smooth

manifold M , each with open interval domains containing 0 ∈ R and with
c1 (0) = c2 (0) = p. Show that
(f ◦ c1 ) (0) = (f ◦ c2 ) (0)
for all smooth f if and only if the curves x ◦ c1 and x ◦ c2 have the same
velocity vector in Rn for some and hence any chart (U, x).

2.1.2. Tangent space via charts. Let A be the maximal atlas for an n-
manifold M . For fixed p ∈ M , consider the set Γp of all triples (p, v, (U, x)) ∈
{p} × Rn × A such that p ∈ U . Define an equivalence relation on Γp by
requiring that (p, v, (U, x)) ∼ (p, w, (V, y)) if and only if
(2.1) w = D(y ◦ x−1 ) x(p)
· v.
In other words, the derivative at x(p) of the coordinate change y ◦ x−1
“identifies” v with w. The set Γp /∼ of equivalence classes can be given a
vector space structure as follows: For each chart (U, x) containing p, we have
a map b(U,x) : Rn → Γp /∼ given by v → [p, v, (U, x)], where [p, v, (U, x)]
denotes the equivalence class of (p, v, (U, x)). To see that this map is a
bijection, notice that if [p, v, (U, x)] = [p, w, (U, x)], then
v = D(x ◦ x−1 ) x(p)
·v =w
by definition. By Proposition 2.3 we obtain a vector space structure on
Γp /∼ whose elements are tangent vectors. This is another version of the
tangent space at p, and we shall (temporarily) denote this by (Tp M )phys .
The subscript “phys” refers to the fact that this version of the tangent space
is based on a “transformation law” and corresponds to a way of looking
at things that has traditionally been popular among physicists. If vp =
[p, v, (U, x)] ∈ (Tp M )phys , then we say that v ∈ Rn represents vp with respect
to the chart (U, x).
This viewpoint takes on a more familiar appearance if we use a more
classical notation. Let (U, x) and (V, y) be two charts containing p in their
domains. If an n-tuple (v 1 , . . . , v n ) represents a tangent vector at p from the
point of view of (U, x), and if the n-tuple (w1 , . . . , wn ) represents the same
vector from the point of view of (V, y), then (2.1) is expressed in the form


n
∂y i
(2.2) wi = vj ,
∂xj x(p)
j=1

where we write the change of coordinates as y i = y i (x1 , . . . , xn ) with 1 ≤

i ≤ n.
Notation 2.6. It is sometimes convenient to index the maximal atlas: A =
{(Uα , xα )}α∈A . Then we would consider triples of the form (p, v, α) and let
2.1. The Tangent Space 61

the defining equivalence relation for (Tp M )phys be (p, v, α) ∼ (p, w, β) if and
only if
D(xβ ◦ x−1
α ) xα (p) · v = w.

2.1.3. Tangent space via derivations. We abstract the notion of di-

rectional derivative for our next approach to the tangent space. There are
actually at least two common versions of this approach, and we explain both.
Let M be a smooth manifold of dimension n. A tangent vector vp at p is
a linear map vp : C ∞ (M ) → R with the property that for f, g ∈ C ∞ (M ),
vp (f g) = g(p)vp (f ) + f (p)vp (g) .
This is the Leibniz law. We may say that a tangent vector at p is a
derivation of the algebra C ∞ (M ) with respect to the evaluation map evp
at p defined by evp (f ) := f (p). Alternatively, we say that vp is a deriva-
tion at p. The set of such derivations at p is easily seen to be a vector
space which is called the tangent space at p and is denoted by Tp M . We
temporarily distinguish this version of the tangent space from (Tp M )kin and
(Tp M )phys defined previously by denoting it (Tp M )alg and referring to it as
the algebraic tangent space. We could also consider the vector space of
derivations of C r (M ) at a point for r < ∞, but this would not give a finite-
dimensional vector space and so is not a good candidate for the definition
of the tangent space (see Problem 18). Recall that if (U, x) is a chart on an
n-manifold M , we have defined ∂f /∂xi by
∂f  
i
(p) := Di f ◦ x−1 (x(p))
∂x
(see Definition 1.57).
∂
Definition 2.7. Given (U, x) and p as above, define the operator ∂xi p
:
C ∞ (M ) → R by
∂ ∂f
i
f := (p).
∂x p ∂xi

It is often helpful to use the easily verified fact that if ci : (−, ) → M

is the curve defined for sufficiently small  by
ci (t) := x−1 (x(p) + ei ),
where ei is the i-th member of the standard basis of Rn , then
∂ f (ci (h)) − f (p)
f = lim .
∂xi p h→0 h
∂
From the usual product rule it follows that ∂xi p
is a derivation at p and so
∂
is an element of (Tp M )alg . We will show that ( ∂x1 p
, . . . , ∂x∂n p ) is a basis
for the vector space (Tp M )alg .
62 2. The Tangent Structure

Lemma 2.8. Let vp ∈ (Tp M )alg . Then

(i) if f, g ∈ C ∞ (M ) are equal on some neighborhood of p, then vp (f ) =
vp (g);
(ii) if h ∈ C ∞ (M ) is constant on some neighborhood of p, then vp (h) =
0.

Proof. (i) Since vp is a linear map, it suffices to show that if f = 0 on a

neighborhood U of p, then vp (f ) = 0. Of course vp (0) = 0. Let β be a
cut-off function with support in U and β(p) = 1. Then we have that βf is
identically zero and so
0 = vp (βf ) = f (p)vp (β) + β(p)vp (f )
= vp (f ) (since β(p) = 1 and f (p) = 0).

(ii) From what we have just shown, it suffices to assume that h is equal
to a constant c globally on M . In the special case c = 1, we have
vp (1) = vp (1 · 1) = 1 · vp (1) + 1 · vp (1) = 2vp (1) ,
so that vp (1) = 0. Finally we have vp (c) = vp (1c) = c (vp (1)) = 0.

Notation 2.9. We shall often write vp f or vp · f in place of vp (f ).

We must now deal with a technical issue. We anticipate that the action
of a derivation is really a differentiation and so it seems that a derivation at
p should be able to act on a function defined only in some neighborhood U of
∂
p. It is pretty easy to see how this would work for ∂x i p . But the domain of
∞ ∞
a derivation as defined is the ring C (M ) and not C (U ). There is nothing
in the definition that immediately allows an element of (Tp M )alg to act on
C ∞ (U ) unless U = M . It turns out that we can in fact identify (Tp U )alg
with (Tp M )alg , and the following discussion shows how this is done. Once
we reach a fuller understanding of the tangent space, this identification will
be natural and automatic. So, let p ∈ U ⊂ M with U open. We construct
a rather obvious map Φ : (Tp U )alg → (Tp M )alg by using the restriction
map C ∞ (M ) → C ∞ (U ). For each wp ∈ Tp U , we define w p : C ∞ (M ) → R
by wp (f ) := wp ( f |U ). It is easy to show that w p is a derivation of the
appropriate type and so w p ∈ (Tp M )alg . Thus we get a linear map Φ :
(Tp U )alg → (Tp M )alg . We want to show that this map is an isomorphism,
but notice that we have not yet established the finite-dimensionality of either
(Tp U )alg or (Tp M )alg . First we show that Φ : wp → w p has trivial kernel. So
suppose that w p = 0, i.e. w p (f ) = 0 for all f ∈ C (M ). Let h ∈ C ∞ (U ).
∞

Pick a cut-off function β with support in U so that βh extends by zero to

a smooth function f on all of M that agrees with h on a neighborhood of
2.1. The Tangent Space 63

p. Then by the above lemma, wp (h) = wp ( f |U ) = w

p (f ) = 0. Thus, since h
was arbitrary, we see that wp = 0 and so Φ has trivial kernel.
Next we show that Φ is onto. Let vp ∈ (Tp M )alg . We wish to define wp ∈
(Tp U )alg by wp (h) := vp (βh), where β is as above and βh is extended by zero
to a function in C ∞ (M ). If β1 is another similar choice of cut-off function,
then βh and β1 h (both extended to all of M ) agree on a neighborhood of p,
and so by Lemma 2.8, vp (βh) = vp (β1 h). Thus wp is well-defined. Thinking
of β ( f |U ) as defined on M , we have wp (f ) := wp ( f |U ) = vp ( βf |U ) = vp (f )
since βf |U and f agree on a neighborhood of p. Thus Φ : (Tp U )alg →
(Tp M )alg is an isomorphism.
Because of this isomorphism, we tend to identify (Tp U )alg with (Tp M )alg
∂
and in particular, if (U, x) is a chart, we think of the derivations ∂x i p,

1 ≤ i ≤ n as being simultaneously elements of both (Tp U )alg and (Tp M )alg .

∂ ∂ (f ◦x−1 )
In either case the formula is the same: ∂x i pf = ∂ui
(x(p)).
Notice that agreeing on a neighborhood of a point is an important rela-
tion here and this provides motivation for employing the notion of a germ
of a function (Definition 1.68). First we establish the basis theorem:

Theorem 2.10. Let M be an n-manifold and (U, x) a chart  with p ∈ U .

∂ ∂
Then the n-tuple of vectors (derivations) ∂x1 p , . . . , ∂xn p is a basis for
(Tp M )alg . Furthermore, for each vp ∈ (Tp M )alg we have

n
∂
vp = vp (xi ) .
∂xi p
i=1

Proof. From our discussion above we may assume that x(U ) is a convex
set such as a ball of radius ε in Rn . By composing with a translation we
assume that x(p) = 0. This makes no difference for what we wish to prove
since vp applied to a constant is 0. For any smooth function g defined on
the convex set x(U ) let
 1
∂g
gi (u) := (tu) dt for all u ∈ x(U ).
0 ∂ui

The fundamental theorem of calculus can be used to show that g = g(0) +


gi ui . We see that gi (0) = ∂u ∂g
i . For a function f ∈ C ∞ (U ), we let
0 
g := f ◦x−1 . Using the above, we arrive at the expression f = f (p) + fi xi ,
∂ ∂f
and applying ∂x i p we get fi (p) = ∂xi . Now apply the derivation vp to
p
64 2. The Tangent Structure


f = f (p) + fi xi to obtain


vp f = 0 + vp (fi xi )



= vp (xi )fi (p) + 0 vp fi

∂f
= vp (xi ) .
∂xi p
 ∂
This shows that vp = vp (xi ) ∂x i p and thus we have a spanning set.
∂ ∂
To see that ( ∂x i p , . . . , ∂xi p ) is a linearly independent set, let us assume
 i ∂
that a ∂xi p = 0 (the zero derivation). Applying this to xj gives 0 =
 i ∂xj  i j
a ∂xi = a δi = aj , and since j was arbitrary, we get the result.

p
Remark 2.11. On the manifold Rn , we have the identity map id : Rn → Rn
which gives the standard chart. As is often the case, the simplest situations
have the most confusing notation because of the various identifications that
may exist. On R, there is one coordinate function, which we often denote
by either u or t. This single function is just idR . The basis vector at t0 ∈ R
∂ ∂
associated to this coordinate is ∂u t0
(or ∂t t0
). If we think of the tangent
space at t0 ∈ R as being {t0 }×R, then ∂u t0 is just (t0 , 1). It is also common
∂

∂
to denote ∂u t0
by “1” regardless of the point t0 .
Above, we used the notion of a derivation as one way to define a tangent
vector. There is a slight variation of this approach that allows us to worry
a bit less about the relation between (Tp U )alg and (Tp M )alg . Let Fp =
Cp∞ (M, R) be the algebra of germs of functions defined near p. Recall that
if f is a representative for the equivalence class [f ] ∈ Fp , then we can
unambiguously define the value of [f ] at p by [f ](p) = f (p). Thus we have
an evaluation map evp : Fp → R.
Definition 2.12. A derivation (with respect to the evaluation map evp )
of the algebra Fp is a map Dp : Fp → R such that Dp ([f ][g]) = f (p)Dp [g] +
g(p)Dp [f ] for all [f ], [g] ∈ Fp .
The set of all these derivations on Fp is easily seen to be a real vector
space and is sometimes denoted by Der(Fp ).
Remark 2.13. The notational distinction between a function and its germ
at a point is not always maintained; Dp f is taken to mean Dp [f ].
Let M be a smooth manifold of dimension n. Consider the set of all
germs of C ∞ functions Fp at p ∈ M. The vector space Der(Fp ) of derivations
of Fp with respect to the evaluation map evp could also be taken as the
definition of the tangent space at p. This would be a slight variation of
what we have called the algebraic tangent space.
2.2. Interpretations 65

2.2. Interpretations
We will now show how to move from one definition of tangent vector to the
next. Let M be a (smooth) n-manifold. Consider a tangent vector vp as an
equivalence class of curves represented by c : I → M with c(0) = p. We
obtain a derivation by defining
d
vp f := f ◦ c.
dt t=0
This gives a map (Tp M )kin → (Tp M )alg which can be shown to be an iso-
morphism. We also have a natural isomorphism (Tp M )kin → (Tp M )phys .
Given [c] ∈ (Tp M )kin , we obtain an element vp ∈ (Tp M )phys by letting vp be
d
the equivalence class of the triple (p, v, (U, x)), where v i := dt t=0
xi ◦ c for
a chart (U, x) with p ∈ U .
If vp is a derivation at p and (U, x) an admissible chart with domain
containing p, then vp , as a tangent vector in the sense of Definition 2.1.2, is
represented by the triple (p, v, (U, x)), where v = (v 1 , . . . , v n ) is given by
v i = vp xi (vp is acting as a derivation).
This gives us an isomorphism (Tp M )alg → (Tp M )phys .
Next we exhibit the inverse isomorphism (Tp M )phys → (Tp M )alg . Sup-
pose that [(p, v, (U, x))] ∈ (Tp M )phys where v ∈ Rn . We obtain a derivation
by defining
vp f = D(f ◦ x−1 ) x(p) · v.
In other words,


n
∂
vp f = vi f
∂xi p
i=1
for v = (v 1 , . . . , v n ). It is an easy exercise that vp defined in this way is
independent of the representative triple (p, v, (U, x)).
We now adopt the explicitly flexible attitude of interpreting
a tangent vector in any of the ways we have described above de-
pending on the situation. Thus we effectively identify the spaces
(Tp M )kin , (Tp M )phys and (Tp M )alg . Henceforth we use the notation Tp M
for the tangent space of a manifold M at a point p.
Definition 2.14. The dual space to a tangent space Tp M is called the
cotangent space and is denoted by Tp∗ M . An element of Tp∗ M is referred
to as a covector.

The basis for Tp∗ M that is dual to the coordinate basis ( ∂x∂ 1 ,..., ∂
∂xnp )
p 
described above is denoted ( dx1 p , . . . , dxn |p ). By definition dxi p
∂
∂xj p
=
δji . (Note that δji = 1 if i = j and δji = 0 if i = j. The symbols δij and
66 2. The Tangent Structure

δ ij are defined similarly.) The reason for the differential notation dxi will
be explained below. Sometimes one abbreviates ∂x∂ j p and dxi p to ∂x∂ j and
dxi respectively, but there is some risk of confusion since later ∂x∂ j and dxi
will more properly denote not elements of the vector spaces Tp M and Tp∗ M ,
but rather fields defined over a chart domain. More on this shortly.

2.2.1. Tangent space of a vector space. Our provisional definition of

the tangent space at point p in a vector space V was the set {p}×V, but this
set does not immediately fit any of the definitions of tangent space just given.
This is remedied by finding a natural isomorphism {p}×V ∼ = Tp V. One may
pick a version of the tangent space and then exhibit a natural isomorphism
directly, but we take a slightly different approach. Namely, we first define
a natural map jp : V → Tp V. We think in terms of equivalence classes of
curves. For each v ∈ V, let cp,v : R →V be the curve cp,v (t) := p + tv. Then
jp (v) := [cp,v ] ∈ Tp V.
As a derivation, jp (v) acts according to
d
f −→ f (p + tv) .
dt 0

On the other hand, we have the obvious projection pr2 : {p}×V → V. Then
our natural isomorphism {p} × V ∼ = Tp V is just jp ◦ pr2 . The isomorphism
between the vector spaces {p} × V and Tp V is so natural that they are often
identified. Of course, since Tp V itself has various manifestations ((Tp V)a lg ,
(Tp V)phys , and (Tp V)kin , we now have a multitude of spaces which are po-
tentially being identified in the case of a vector space.
Note: Because of the identification of {p} × V with Tp V , we shall often
denote by pr2 the map Tp V → V. Furthermore, in certain contexts, Tp V
is identified with V itself. The potential identifications introduced here are
often referred to as “canonical” or “natural” and jp : V → Tp V is often
called the canonical or natural isomorphism. The inverse map Tp V → V
is also referred to as the canonical or natural isomorphism. Context will
keep things straight.

2.3. The Tangent Map

The first definition given below of the tangent map at p ∈ M of a smooth
map f : M → N will be considered our main definition, but the others are
actually equivalent. Given f and p as above, we wish to define a linear map
Tp f : Tp M → Tf (p) N . Since we have several definitions of tangent space, we
expect to see several equivalent definitions of the tangent map. For the first
definition we think of Tp M as (Tp M )kin .
 Chapter 12

Connections and
Covariant Derivatives

The terms “covariant derivative” and “connection” are sometimes treated

as synonymous. In fact, a covariant derivative is sometimes called a Koszul
connection. From one point of view, the central idea is that of measuring
the rate of change of sections of bundles in the direction of a vector or vector
field on the base manifold. Here the derivative viewpoint is prominent. From
another related point of view, a connection provides an extra structure that
gives a principled way of lifting curves from the base to the total space. The
lifts are parallel sections along the curve. In this chapter, we will always
take the typical fiber of an F-vector bundle of rank k to be Fk . We let
(e1 , . . . , ek ) be the standard basis of Fk . Thus every vector bundle chart
(U, φ) is associated with a frame field (e1 , . . . , ek ) where ei (x) := φ−1 (x, ei ).

12.1. Definitions
Let π : E → M be a smooth F-vector bundle of rank k over a manifold M . A
covariant derivative can either be defined as a map ∇ : X(M ) × Γ(M, E) →
Γ(M, E) with certain properties from which one derives a well-defined map
∇ : T M ×Γ(M, E) → Γ(M, E) with nice properties or the other way around.
We rather arbitrarily start with the first of these definitions.

Definition 12.1. A covariant derivative or Koszul connection on a

smooth F-vector bundle E → M is a map ∇ : X(M ) × Γ(M, E) → Γ(M, E)
(where ∇(X, s) is written as ∇X s) satisfying the following four properties:
(i) ∇f X (s) = f ∇X s for all f ∈ C ∞ (M ), X ∈ X(M ) and s ∈ Γ(M, E);

501
502 12. Connections and Covariant Derivatives

(ii) ∇X1 +X2 s = ∇X1 s + ∇X2 s for all X1 , X2 ∈ X(M ) and s ∈ Γ(M, E);
(iii) ∇X (s1 + s2 ) = ∇X s1 + ∇X s2 for all X ∈ X(M ) and s1 , s2 ∈
Γ(M, E);
(iv) ∇X (f s) = (Xf )s + f ∇X s for all f ∈ C ∞ (M ; F), X ∈ X(M ) and
s ∈ Γ(M, E).
For a fixed X ∈ X(M ), the map ∇X : Γ(M, E) → Γ(M, E) is called the
covariant derivative with respect to X.

As we will see below, this definition is enough to imply that ∇ induces

maps ∇U : X(U ) × Γ(U, E) → Γ(U, E), one for each open U ⊂ M , that are
naturally related in a sense we make precise below (this is not necessarily
true for infinite-dimensional manifolds). Furthermore, we also prove that
for a fixed p ∈ M , the value (∇X s)(p) depends only on the value of X at
p and only on the values of s along any smooth curve c representing Xp .
Thus we get a well-defined map ∇ : T M × Γ(M, E) → Γ(M, E) such that
∇v s = (∇X s) (p) for any extension of v ∈ Tp M to a vector field X with
Xp = v. The resulting properties are
(i ) ∇av (s) = a∇v s for all a ∈ R, v ∈ T M and s ∈ Γ(M, E);
(ii ) for all p ∈ M we have ∇v1 +v2 s = ∇v1 s + ∇v2 s for all v1 , v2 ∈ Tp M ,
and s ∈ Γ(M, E);
(iii ) ∇v (s1 + s2 ) = ∇v s1 + ∇v s2 for all v ∈ T M and s1 , s2 ∈ Γ(M, E);
(iv ) for all p ∈ M we have ∇v (f s) = (vf )s(p)+f (p)∇v s for all v ∈ Tp M ,
s ∈ Γ(M, E) and f ∈ C ∞ (M ; F);
(v ) p → ∇Xp s is smooth for all smooth vector fields X.
A map satisfying these properties is also called a covariant derivative (or
Koszul connection). Note that it is easy to obtain a Koszul connection
in the first sense since we just let (∇X s) (p) := ∇Xp s.
Definition 12.2. A covariant derivative on the tangent bundle T M of an
n-manifold M is usually referred to as a covariant derivative on M .

In Chapter 4 have already met the Levi-Civita connection ∇ on Rn ,

which is, from the current point of view, a Koszul connection on the tangent
bundle of Rn . The definition of this connection takes advantage of the
natural identification of tangent spaces which makes taking the difference
quotient possible:
X(p + tv) − X(p)
∇v X = lim .
t→0 t
In that same chapter we obtained, by a projection, a covariant derivative on
(the tangent bundle of) any hypersurface in Rn . A covariant derivative on
12.1. Definitions 503

a submanifold of arbitrary codimension can be obtained in the same way.

Let M be a submanifold of Rn and let X ∈ X(M ) and v ∈ Tp M . We have
 
d
∇v X := X ◦c ∈ Tp M .
dt 0
One may easily verify that ∇, so defined, is a covariant derivative (in the
second sense above).
Returning to the case of a general vector bundle, let us consider how
covariant differentiation behaves with respect to restriction to open subsets
of our manifold. Recall the restriction map rVU : Γ(U, E) → Γ(V, E) given
by rVU : σ → σ|V and where V ⊂ U .
Definition 12.3. A natural covariant derivative ∇ on a smooth F-
vector bundle E → M is an assignment to each open set U ⊂ M of a map
∇U : X(U ) × Γ(U, E) → Γ(U, E) written as ∇U : (X, σ) → ∇U
X σ such that
the following assertions hold:
(i) For every open U ⊂ M , the map ∇U is a Koszul connection on the
restricted bundle E|U → U .
X σ) = ∇r U X rV σ (nat-
(ii) For nested open sets V ⊂ U , we have rVU (∇U V U
V
urality with respect to restrictions).
(iii) For X ∈ X(U ) and σ ∈ Γ(U, E) the value (∇U
X σ)(p) only depends
on the value of X at p ∈ U .

Here ∇U X σ is called the covariant derivative of σ with respect to X.

We will denote all of the maps ∇U by the single symbol ∇ when there is no
chance of confusion. We have explicitly worked the naturality conditions (ii)
and (iii) into the definition of a natural covariant derivative, so this defini-
tion is also appropriate for infinite-dimensional manifolds. The definition of
Koszul connection did not specifically include these naturality features and
was only defined for global sections. We shall now see that, in the case of
finite-dimensional manifolds, a Koszul connection gives a natural covariant
derivative for free.
Lemma 12.4. Suppose ∇ : X(M ) × Γ(E) → Γ(E) is a Koszul connection
for the vector bundle E → M . Then if for some open U either X|U = 0 or
σ|U = 0, then
(∇X σ)(p) = 0 for all p ∈ U .

Proof. We prove the case of σ|U = 0 and leave the case of X|U = 0 to the
reader.
Let q ∈ U . Then there is some relatively compact open set V with
q ∈ V ⊂ U and a smooth function f that is identically one on V and zero
504 12. Connections and Covariant Derivatives

outside of U . Thus f σ ≡ 0 on M and so since ∇ is linear, we have ∇(f σ) ≡ 0

on M . Thus since (iv) of Definition 12.1 holds for global fields, we have
∇X (f σ)(q) = f (p)(∇X σ)(q) + (Xq f )σ(q)
= (∇X σ)(q) = 0.
Since q ∈ U was arbitrary, we have the result.

We now define a natural covariant derivative derived from a given Koszul

connection ∇. Given any open set U ⊂ M , we define ∇U : X(U )×Γ( E|U ) →
Γ( E|U ) by
 U 
(12.1) ) (p) , p ∈ U,
∇X σ (p) := (∇X σ
 ∈ X(M ) and σ
for X  ∈ Γ(E) chosen to be any sections which agree with
X and σ on some open V with p ∈ V ⊂ V̄ ⊂ U . By the above lemma this
 and σ
definition does not depend on the choices of X .
Proposition 12.5. Let E → M be a rank k vector bundle and suppose that
∇ : X(M ) × Γ(E) → Γ(E) is a Koszul connection. If for each open U we
define ∇U as in ( 12.1) above, then the assignment U → ∇U is a natural
covariant derivative as in Definition 12.3.

Proof. We must show that (i), (ii) and (iii) of Definition 12.3 hold. It is
easily checked that (i) holds, that is, that ∇U is a Koszul connection for each
U . The demonstration that (ii) holds is also easy and we leave it for the
reader to check. Now since X → ∇X σ is linear over C ∞ (U ), (iii) follows by
familiar arguments (∇X σ is linear over functions in the argument X).

Because of this last lemma, we may define ∇vp σ for vp ∈ Tp M by

∇vp σ := (∇X σ) (p),
where X is any vector field with X(p) = vp ∈ Tp M . We say that ∇X σ is
“tensorial” in the variable X. The result can be seen as a special case of
Proposition 6.55. We now see that it is safe to write expressions not directly
justified by the definition of Koszul connection. For example, if X ∈ X(U )
and σ ∈ Γ(V, E), where U ∩ V = ∅, then ∇X σ is taken to be an element of
Γ(U ∩ V, E) defined by
(∇X σ) (p) = ∇Xp σ := ∇U
Xp σ for all p ∈ U ∩ V .

This is a particularly useful convention when U is the domain of a chart

∂
(U, x) and X = ∂x i and also when σ is a member of a frame field of the
vector bundle defined on some open set.
In the same way that one extends a derivation on vector fields to a tensor
derivation, one may show that a covariant derivative on a vector bundle
induces naturally related covariant derivatives on all the multilinear bundles.
12.1. Definitions 505

In particular, if π ∗ : E ∗ → M denotes the dual bundle to π : E → M we

may define connections on π ∗ : E ∗ → M and on π ⊗ π ∗ : E ⊗ E ∗ → M . We
do this in such a way that for s ∈ Γ(M, E) and s∗ ∈ Γ(M, E ∗ ) we have
∗ ∗
∇E⊗E
X (s ⊗ s∗ ) = ∇X s ⊗ s∗ + s ⊗ ∇E ∗
X s ,

and
∗ ∗ ∗ ∗
X s )(s) = X(s (s)) − s (∇X s).
(∇E
Of course, the second formula follows from the requirement that covariant
differentiation commutes with contraction:
∗
X(s∗ (s)) = (∇X C(s ⊗ s∗ )) = C(∇E⊗E
X (s ⊗ s∗ ))
 ∗ ∗
 ∗ ∗
= C ∇X s ⊗ s∗ + s ⊗ ∇E X s = s∗ (∇X s) + (∇E
X s )(s),

where C denotes the contraction given by s ⊗ α → α(s). All this works like
the tensor derivation extension procedure discussed previously, and we often
write all of these covariant derivatives with the single symbol ∇.
The bundle E ⊗ E ∗ → M is naturally isomorphic to End(E), and by
this isomorphism we get a connection on End(E). If we identify elements
of Γ (End(E)) with End(ΓE) (see Proposition 6.55), then we may use the
following formula for the definition of the connection on End(E):
(∇X A)(s) := ∇X (A(s)) − A(∇X s).
Indeed, since C : s ⊗ A → A(s) is a contraction, we must have
∇X (A(s)) = C (∇X s ⊗ A + s ⊗ ∇X A)
= A(∇X s) + (∇X A)(s).

Notice that if we fix s ∈ Γ(E), then for each p ∈ M , we have an element

(∇s) (p) of L(Tp M, Ep ) ∼
= E ⊗ T ∗ M given by
(∇s) (p) : vp → ∇vp s for all vp ∈ Tp M .
Thus we obtain a section ∇s of E ⊗ T ∗ M given by p → (∇s) (p), which is
easily shown to be smooth. In this way, we can also think of a connection
as giving a map
∇ : Γ(E) → Γ(E ⊗ T ∗ M )
with the property that
∇f s = f ∇s + s ⊗ df
for all s ∈ Γ(E) and f ∈ C ∞ (M ). Since, by definition, Γ(E) = Ω0 (E) and
Γ(E ⊗ T ∗ M ) = Ω1 (E), we really have a map Ω0 (E) → Ω1 (E). Later we will
extend to a map Ωk (E) → Ωk+1 (E) for all integral k ≥ 0. Now if X is a
smooth vector field, then X → ∇X s ∈ Γ(E), so we may also view ∇s as an
element of Hom(X(M ), Γ(E)).
506 12. Connections and Covariant Derivatives

12.2. Connection Forms

Let π : E → M be a rank r vector bundle with a connection ∇. Recall that
a choice of a local frame field over an open set U ⊂ M is equivalent to a
trivialization of the restriction πU : E|U → U . Namely, if φ = (π, Φ) is such
a trivialization over U , then defining ei (x) = φ−1 (x, ei ), where (ei ) is the
standard basis of Fn , we have a frame field (e1 , . . . , ek ). We now examine
the expression for a given covariant derivative from the viewpoint of such a
local frame. It is not hard to see that for every such frame field there must
be a matrix of 1-forms ω = (ωji )1≤i,j≤r such that for X ∈ Γ(U, E) we may
write

k
∇X ej = ωji (X)ei .
i=1

The forms ωji

are called connection forms.

Proposition 12.6. If s = i si ei is the local expression of a section s ∈
Γ(E) in terms of a local frame field (e1 , . . . , ek ), then the following local
expression holds:


∇X s = (Xsi + ωri (X)sr )ei .
i r

Proof. We simply compute:

 
∇X s = ∇X si ei
i
 
= (Xsi )ei + si ∇X ei
i i
 
= i
(Xs )ei + si ωij (X)ej
i i,j
 
= (Xsi )ei + sr ωri (X)ei
i i,r

  
= Xsi + ωri (X)sr ei .

i r

So the i-th component of ∇X s is

(12.2) (∇X s)i = Xsi + ωri (X)sr .
r

We may surely choose U small enough that it is also the domain of a coor-
dinate frame {∂µ } for M . Thus we have


(12.3) ∇∂µ ej = k
ωµj ek ,
k
12.3. Differentiation Along a Map 507

k = ω j (∂ ). We now have the local formula

where ωµj i µ

k 


n

n

k 
(12.4) ∇X s = X µ ∂µ si + X µ ωµr
i r
s ei .
i=1 µ=1 µ=1 r=1

Or, using the summation convention,

 
∇X s = X µ ∂µ si + X µ ωµr
i r
s ei .

Now suppose that we have two moving frames whose domains overlap,
say u = (e1 , . . . , ek ) and u = (e1 , . . . , ek ). Let us examine how the matrix of
1-forms ω = (ωij ) is related to the corresponding ω  = (ωij ) defined in terms
of the frame u . The change of frame is

j
ei = gi ej ,
j

which in matrix notation is

u = ug
for some smooth g : U ∩ U  → GL(n). (We treat u as a row vector of fields.)
For a given moving frame, let ∇u := [∇e1 , . . . , ∇ek ]. Differentiating both
sides of u = ug and using matrix notation we have
u = ug,
∇u = ∇(ug),
u ω  = (∇u)g + udg,
u ω  = ugg −1 ωg + ugg −1 dg,
u ω  = u g −1 ωg + u g −1 dg,
and so we obtain the transformation law for connection forms:
ω  = g −1 ωg + g −1 dg.

12.3. Differentiation Along a Map

Once again let π : E → M be a vector bundle with a Koszul connection
∇. Consider a smooth map f : N → M and a section σ : N → E along f .
Let e1 , . . . , ek be a frame field defined over U ⊂ M . Since f is continuous,
O = f −1 (U ) is open and e1 ◦ f, . . . , ek ◦ f are fields along f defined on O.

We may write σ = ka=1 σ a ea ◦ f for some functions σ a : O ⊂ N → F.
For any p ∈ O and v ∈ Tp N , we define
% &

k
k
(12.5) ∇v σ :=
f
(dσ · v) +
a a r
ωr (T f (v)) σ (p) ea (f (p)).
a=1 r=1