10 Construction of the tangent bundle

10.1 Cotangent spaces and the differential

Since the tangent space is a vector space, we can do all the constructions we saw previously
in the abstract vector space setting.
Definition. Let M be a manifold and p M . The cotangent space to M at p is

Tp M := (Tp M ) .

Since dim Tp M is finite, we have Tp M =vec T M . If { a } is the basis of Tp M

p x p
induced by some chart (U, x), then the dual basis is denoted as {(dxa )p }. We have, by
definition  !

(dxa )p = ba .
xb p
Once we have the cotangent space, we can define the tensor spaces.
Definition. Let M be a manifold and p M . The tensor space (Tsr )p M is defined as

(Tsr )p M := Tsr (Tp M ) = Tp M Tp M Tp M Tp M .

| {z } | {z }
r copies s copies

Definition. Let M and N be manifolds and let : M N be smooth. The differential

(or derivative) of at p M is the linear map

dp : Tp M
T(p) N
X 7 dp (X)

where dp (X) is the tangent vector to N at (p)

dp (X) : C (N )
R
g 7 (dp (X))(g) := X(g ).

If this definition looks confusing, it is worth it to pause and think about what it is
saying. Intuitively, if takes us from M to N , then dp takes us from Tp M to T(p) N . The
way in which it does so, is the following.

M N C (M ) C (N )

g X
g dp (X)

R R
Given X Tp M , we want to construct dp (X) T(p) N , i.e. a derivation on N at f (p).
Derivations act on functions. So, given g : N R, we want to construct a real number
by using and X. There is really only one way to do is. If we precompose g with , we
obtain g : M R, which is an element of C (M ). We can then happily apply X to
this function to obtain a real number. You should check that dp (X) is indeed a tangent
vector to N .

1
Remark 10.1. Note that, to be careful, we should replace C (M ) and C (N ) above with
C (U ) and C (V ), where U M and V N are open and contain p and (p), respec-
tively.
0 0
Example 10.2. If M = Rd and N = Rd , then the differential of f : Rd Rd at p Rd
0 0
dp f : Tp Rd
=vec Rd Tf (p) Rd =vec Rd

is none other than the Jacobian of f at p.

A special case of the differential is the gradient of a function in C (M ).

Definition. Let M be a manifold and let f : M R be smooth. The gradient of f at

p M is the covector

Tf (p) R
dp f : Tp M =vec R
X 7 dp f (X) := X(f ).

In fact, we can define the gradient operator at p M as the R-linear map

dp : C (U )
Tp M
f 7 dp f,

with p U M .

Remark 10.3. Note that, by writing dp f (X) := X(f ), we have committed a slight (but
nonetheless real) abuse of notation. Since dp f (X) Tf (p) R, it takes in a function and
return a real number, but X(f ) is already a real number! This is due to the fact that we
have implicitly employed the isomorphism

d : Tp Rd Rd
X 7 (X(proj1 ), . . . , X(projd )),

which, when d = 1, reads

1 : Tp R R
X 7 X(idR ).

In our case, we have

dp f (X) := X( f ) 7 X(idR f ) = X(f ).

This notwithstanding, the best way to think of dp f is as a covector, i.e. dp f takes in a

tangent vector X and returns the real number X(f ), in a linear fashion.
Recall that if (U, x) is a chart on M , then the co-ordinate maps xa : U x(U ) Rdim M
are smooth functions on U . We can thus apply the gradient operator dp (with p U ) to
each of them to obtain (dim M )-many elements of Tp M .

2
Proposition 10.4. Let (U, x) be a chart on M , with p U . The set B = {dp xa | 1 a
dim M } forms a basis of Tp M .

Proof. We already know that Tp M = dim M , since it is the dual space to Tp M . As

|B| = dim M by construction, it suffices to show that it is linearly independent. Suppose
that
a dp xa = 0,
for some a R. Applying the left hand side to the basis element x b p yields

 !  

a dp x a
= a (xa ) (definition of dp xa )
xb p xb p
= a b (xa x1 )(x(p)) (definition of
)


xb p
= a b (proja )(x(p))
= a ba
= b .

Therefore, B is linearly independent and hence a basis of Tp M . Moreover, since we have

shown that  !

dp xa = ba ,
xb p

this basis is, in fact, the dual basis to { x a p }.

Remark 10.5. Note a slight subtlety. Given a chart (U, x) and the induced basis { x a p }

of Tp M , the dual basis to { x a p } exists simply by virtue of Tp M being the dual space to

Tp M . What we have shown above is that the elements of this dual basis are given explicitly
by the gradients of the co-ordinate maps of (U, x). In our notation, we have

(dxa )p = dp xa , 1 a dim M.

10.2 Push-forward and pull-back

The push-forward of a smooth map : M N at p M is just another name for the
differential of at p. We give the definition again in order to establish the new notation.

Definition. Let : M N be a smooth map between smooth manifolds. The push-

forward of at p M is the linear map:

( )p : Tp M
T(p) N
X 7 ( )p (X) := X( ).

If : R M is a smooth curve on M and : M N is smooth, then : R N

is a smooth curve on N . Informally, we say that we can push a curve on M forward to
the curve on N .

3
Proposition 10.6. Let : M N be smooth. The tangent vector X,p Tp M is pushed
forward to the tangent vector X,(p) T(p) N , i.e.

( )p (X,p ) = X,(p) .

Proof. Let f C (V ), with (V, x) a chart on N and (p) V . By applying the definitions,
we have

( )p (X,p )(f ) = (X,p )(f ) (definition of ( )p )

0
= ((f ) ) (0) (definition of X,p )
0
= (f ( )) (0) (associativity of )
= X,(p) (f ) (definition of X,(p) )

Since f was arbitrary, we have ( )p (X,p ) = X,(p) .

Related to the push-forward, there is the notion of pull-back of a smooth map.

Definition. Let : M N be a smooth map between smooth manifolds. The pull-back

of at p M is the linear map:

( )p : T(p)

Tp M
N
7 ( )p (),

where ( )p () is defined as

( )p () : Tp M
R
X 7 (( )p (X)),

In words, if is a covector on N , its pull-back ( )p () is a covector on M . It acts

on tangent vectors on M by first pushing them forward to tangent vectors on N , and then
applying to them to produce a real number.
Remark 10.7. If you dont see it immediately, then you should spend some time proving
that all the maps that we have defined so far and claimed to be linear are, in fact, linear.
Remark 10.8. We have seen that, given a smooth : M N , we can push a vector
X Tp M forward to a vector ( )p (X) T(p) N , and pull a covector T(p)
N back to

a covector ( )p () Tp M .

( )p
C (M ) C (N ) Tp M T(p) N

X
( )p (X) ( )p ()

R R

4
 However, if : M N is a diffeomorphism, then we can also pull a vector Y T(p) N
back to a vector ( )p (Y ) Tp M , and push a covector Tp M forward to a covector
N , by using 1 as follows:
( )p () T(p)

( )p (Y ) := ((1 ) )(p) (Y )
( )p () := ((1 ) )(p) ().

1 ((1 ) )(p)
C (M ) C (N ) Tp M T(p) N

Y
( )p (Y ) ( )p ()

R R
Again, this is only possible if is a diffeomorphism. In general, you should keep in mind
that

Vectors are pushed forward,

covectors are pulled back.

10.3 Immersions and embeddings

We will now consider the question of under which circumstances a smooth manifold can sit
in Rd , for some d N. There are, in fact, two notions of sitting inside another manifold,
called immersion and embedding.

Definition. A smooth map : M N is said to be an immersion of M into N if the

derivative

dp ( )p : Tp M
T(p) N
is injective, for all p M . The manifold M is said to be an immersed submanifold of N .

Definition. A smooth map : M N is said to be an embedding of M into N if

: M N is an immersion;

M
=top (M ) N , where (M ) carries the subset topology inherited from N .

The manifold M is said to be an embedded submanifold of N .

In the early days of differential geometry there were two approaches to study manifolds.
One was the extrinsic view, within which manifolds are defined as special subsets of Rd ,
and the other was the intrinsic view, which is the view that we have adopted here.
Whitneys theorem, which we will state without proof, states that these two approaches
are, in fact, equivalent.

Theorem 10.9 (Whitney). Any smooth manifold M can be

embedded in R2 dim M ;

5
 immersed in R2 dim M 1 .

Example 10.10. The Klein bottle can be embedded in R4 but not in R3 . It can, however,
be immersed in R3 .
What we have presented above is referred to as the strong version of Whitneys theorem.
There is a weak version as well, but there are also even stronger versions of this result, such
as the following.

Theorem 10.11. Any smooth manifold can be immersed in R2 dim M a(dim M ) , where a(n)
is the number of 1s in a binary expansion of n N.

Example 10.12. If dim M = 3, then as

310 = (1 21 + 1 20 )10 = 112 ,

we have a(dim M ) = 2, and thus every 3-dimensional manifold can be immersed into R4 .
Note that even the strong version of Whitneys theorem only tells us that we can immerse
M into R5 .

10.4 The tangent bundle

We would like to define a vector field on a manifold M as a smooth map that assigns
to each p M a tangent vector in Tp M . However, since this would then be a map to a
different space at each point, it is unclear how to define its smoothness.
The simplest solution is to merge all the tangent spaces into a unique set and equip it
with a smooth structure, so that we can then define a vector field as a smooth map between
smooth manifolds.

Definition. Given a smooth manifold M , the tangent bundle of M is the disjoint union of
all the tangent spaces to M , i.e.
a
T M := Tp M,
pM

equipped with the canonical projection map

: TM M
X 7 p,

where p is the unique p M such that X Tp M .

We now need to equip T M with the structure of a smooth manifold. We can achieve
this by constructing a smooth atlas for T M from a smooth atlas on M , as follows.
Let AM be a smooth atlas on M and let (U, x) AM . If X preim (U ) T M , then
X T(X) M , by definition of . Moreover, since (X) U , we can expand X in terms of
the basis induced by the chart (U, x):
 
a
X=X ,
xa (X)

6
where X 1 , . . . , X dim M R. We can then define the map

: preim (U ) x(U ) Rdim M

=set R2 dim M
X 7 (x((X)), X 1 , . . . , X dim M ).

Assuming that T M is equipped with a suitable topology, for instance the initial topol-
ogy (i.e. the coarsest topology on T M that makes continuous), we claim that the pair
(preim (U ), ) is a chart on T M and

AT M := {(preim (U ), ) | (U, x) AM }

is a smooth atlas on T M . Note that, from its definition, it is clear that is a bijection. We
will not show that (preim (U ), ) is a chart here, but we will show that AT M is a smooth
atlas.

Proposition 10.13. Any two charts (preim (U ), ), (preim (U e AT M are C -compatible.

e ), )

Proof. Let (U, x) and (U e) be the two charts on M giving rise to (preim (U ), ) and
e, x
(preim (U ), ), respectively. We need to show that the map
e e

e 1 : x(U U
e ) Rdim M x e ) Rdim M
e(U U

is smooth, as a map between open subsets of R2 dim M . Recall that such a map is smooth
if, and only if, it is smooth componentwise. On the first dim M components, e 1 acts as

e x1 : x(U U
x e) x
e(U U
e)
x(p) 7 x
e(p),

while on the remaining dim M components it acts as the change of vector components we
met previously, i.e.
e a = b (y a x1 )(x(p)) X b .
X a 7 X
Hence, we have

e 1 : x(U U e ) Rdim M x e ) Rdim M

e(U U
(x((X)), X 1 , . . . , X dim M ) 7 (e
x((X)), Xe 1, . . . , X
e dim M ),

which is smooth in each component, and hence smooth.

The tangent bundle of a smooth manifold M is therefore itself a smooth manifold

of dimension 2 dim M , and the projection : T M M is smooth with respect to this
structure.
Similarly, one can construct the cotangent bundle T M to M by defining
a
T M := Tp M
pM

and going through the above again, using the dual basis {(dxa )p } instead of {


xa p }.