This action might not be possible to undo. Are you sure you want to continue?
BACKGROUND
The prerequisites for this course are a background in general topology, advanced calculus
(analysis), and linear algebra.
The background in general topology is a knowledge of the notions of a topology as given
in a one semester or quarter course: the relative topology, the quotient topology, compact
sets and topological spaces, second countable topological spaces, and Hausdorﬀ spaces.
At several points in the notes connections are made to algebraic topology, e.g., the
fundamental group, but algebraic topology is not required in the logical dependence of
these notes.
The assumed analysis background is a course in Calculus of several variables that in
cludes the The Inverse Function Theorem, The Implicit Function Theorem, The Change
of Variables Formula from integration, and The Mean Value Theorem for Integrals. In
addition, we will also use the theorem of Existence and Uniqueness of solutions to ﬁrst
order diﬀerential equations. We now state and discuss these Theorems as well as setting
some notation.
Suppose O ⊂ R
m
and f : O → R
n
. If p ∈ O and v ∈ R
m
, then the directional
derivative of f at p in the direction v is
D
v
f(p) = lim
h→0
f(p +hv) −f(p)
h
, if the limit exists.
Some authors require that v is a unit vector. This assumption is almost universal in
calculus books, but not in analysis books or literature. We do not assume that v is a unit
vector.
If f is diﬀerentiable, then the derivative at p is the linear map Df(p) : R
m
→ R
n
deﬁned by
Df(p)(v) = D
v
f(p).
The matrix for Df(p), in the standard basis, is the Jacobian matrix Jf(p) and
Jf(p) =
∂f
1
∂x
1
· · ·
∂f
1
∂x
m
.
.
.
.
.
.
∂f
n
∂x
1
· · ·
∂f
n
∂x
m
The basic theorems that we assume from a background of analysis are: The Inverse
Function Theorem, The Implicit Function Theorem, The Change of Variables Formula
from integration, and The Mean Value Theorem for Integrals. In addition, we will also use
the theorem of existence and uniqueness of solutions to ﬁrst order diﬀerential equations.
We now state and discuss these Theorems.
copyright c 2002
Typeset by A
M
ST
E
X
1
2 CHAPTER 1 BACKGROUND
Inverse Function Theorem.
Implicit Function Theorem. Let f : R
n+m
→ R
m
be a C
∞
function. Write points
in R
n+m
= R
n
× R
m
as (t
1
, · · · , t
n
, x
1
, · · · , x
m
) = (t; x) and f = (f
1
, · · · , f
m
). Let
f(t
0
; x
0
) = p for p ∈ R
m
. Suppose that det(
∂f
j
∂x
i
(t
0
;x
0
)
) = det( Df
R
m
(t
0
; x
0
)) = 0.
Then there is a ndimensional open set T
0
containing t
0
, an mdimensional open set X
0
containing x
0
, and a unique function g : T
0
→ R
m
such that
(1) g is C
∞
(2) g(t
0
) = x
0
, and
(3) for (t; x) ∈ T
0
×X
0
, f(t; x) = p if and only if g(t) = x.
There is a corollary of the Implicit Function Theorem that will, for us, be the useful
form of the Implicit Function Theorem. It is sometimes called the Rank Theorem.
The Rank Theorem. Suppose O ⊂ R
n+m
is an open set and f : O → R
m
is C
∞
.
Suppose p ∈ R
m
, q ∈ O and f(q) = p. If Df(q) has rank m, then there is an open set
U ⊂ O, q ∈ U; open sets U
1
⊂ R
n
, and U
2
⊂ R
m
;and, a diﬀeomorphism H : U → U
1
×U
2
such that y ∈ U ∩ f
−1
(p) if and only if H(y) ∈ U
1
×{0}.
Proof. The kernel of Df(q) is ndimensional, so let Ker(Df(q)) = span{v
1
, · · · , v
n
} = R
n
and let span{v
n+1
, · · · , v
n+m
} = R
m
, so that v
n+1
, · · · , v
n+m
is an ordered basis of
R
n+m
= R
n
× R
m
. Let K : R
n+m
→ R
n+m
be deﬁned by K(y) =
n+m
i=1
y
i
v
i
+ q
for each y = (y
1
, · · · , y
n+m
) ∈ R
n+m
. The map K is a C
∞
diﬀeomorphism, it is
an isomorphism and a translation. Also Df ◦ K(0)(R
m
) = Df(q) ◦ DK(0)(R
m
) =
Df(q)(span{v
n+1
, · · · , v
n+m
}) = R
m
, and so det( Df ◦ K(0)
R
m
) = 0. We have that
K
−1
(O) is an open set, f ◦ K : K
−1
(O) → R
m
is a C
∞
function, f ◦ K(0) = p, and
det( Df ◦ K(0)
R
m
) = 0, and hence we can apply the Implicit Function Theorem. We con
clude that there are open sets T ⊂ R
n
, X ⊂ R
m
, T × X ⊂ K
−1
(O), and a C
∞
function
g : T → X such that f ◦ K(t; x) = p if and only if g(t) = x, i.e., (f ◦ K)
−1
(q) is the graph
of g.
Take an open set U
1
⊂ T with 0 ∈ U
1
and an open set U
2
⊂ X such that O
1
=
{(t, x)  x ∈ g(t) + U
2
} ⊂ T × X. It is an exercise in general topology that this can
always be arranged (one ﬁrst takes U
1
so that its closure is a compact subset of T). Let
h : O
1
→ U
1
×U
2
by h(t, x) = (t, x−g(t)). The map h is a C
∞
diﬀeomorphism, its inverse
is (t, x) → (t, x +g(t)). Under this diﬀeomorphism, G
g
∩ O
1
maps to U
1
×{0}.
To complete the proof, take H = h◦K
−1
and U = h◦K
−1
(O
1
). Then H : U → U
1
×U
2
is a C
∞
diﬀeomorphism and x ∈ U ∩ f
−1
(p) if and only if H(x) ∈ U
1
×{0}.
Change of Variables Formula.
Mean Value Theorem for Integrals.
CHAPTER 2
MANIFOLDS
In this chapter, we address the basic notions: What is a manifold and what is a map
between manifolds. Several examples are given.
An n dimensional manifold is a topological space that appears to be R
n
near a point,
i.e., locally like R
n
. Since these topological spaces appear to be locally like R
n
, we may
hope to develop tools similar to those used to study R
n
in order to study manifolds. The
term manifold comes from “many fold,” and it refers to the many dimensions of space that
a manifold may describe.
Later in this section we will also need structure to discuss C
∞
functions on manifolds.
Our notion of smooth will be C
∞
, i.e., continuous partial derivatives of all orders. The
terms C
∞
and smooth are usually synonymous, however, in this chapter and the next,
we will use C
∞
to describe maps between real vector spaces (as in advanced Calculus)
and smooth for maps between manifolds. We use this distinction since many confusing
compositions occur and the reader is assumed to be familiar with C
∞
maps between real
vector spaces while results on smooth maps between manifolds must be proven. We begin
with the notion of a topological manifold.
Deﬁnition 2.1***. A topological manifold of dimension n is a second countable Haus
dorﬀ space M for which each point has a neighborhood homeomorphic to an open set in
R
n
.
This notion appears to capture the topological ideal of locally looking like R
n
, but in
order to do Calculus we will need more structure. The term nmanifold is usually written
for n dimensional manifold, and the dimension n is often suppressed.
Deﬁnition 2.2***. Let U ⊂ M be a connected open set in a topological nmanifold M,
and φ : U → R
n
be a homeomorphism to its image φ(U), an open set in R
n
.
The pair (U, φ) is called a coordinate system or chart. If x
0
∈ U and φ(x
0
) = 0 ∈ R
n
,
then the coordinate system is centered at x
0
. Call the map φ a coordinate map.
Deﬁnition 2.3***. Suppose M is a topological manifold. For k ∈ N, k = 0, or k = ∞,
a C
k
atlas is a set of charts {(U
i
, φ
i
)  i ∈ I} such that
(1)
i∈I
U
i
= M, and
(2) φ
i
◦ φ
−1
j
is C
k
on its domain for all i, j ∈ I.
The k denotes the degree of diﬀerentiability. If k = 0, then maps are just continuous.
Deﬁnition 2.4***. If A = {(U
i
, φ
i
)i ∈ I} is a C
k
atlas for an nmanifold M
n
and
f : U → R
n
is a homeomorphism onto its image with U ⊂ M open, then (U, f) is
compatible with A if φ
i
◦ f
−1
: f(U ∩U
i
) → φ(U ∩U
i
) is C
k
and f ◦ φ
−1
i
is C
k
for all i ∈ I.
copyright c 2002 Larry Smolinsky
Typeset by A
M
ST
E
X
1
2 CHAPTER 2 MANIFOLDS
Theorem 2.5***. If A = {(U
i
, φ
i
)i ∈ I} is a C
k
atlas for M
n
then A is contained in a
unique maximal atlas for M where atlases are ordered as sets by containment.
Proof. Let
M ={(V
α
, f
α
)V
α
⊂ M is open, f
α
: V
α
→ R
n
is a homeomorphism onto its image, and (V
α
, f
α
) is compatible with A}.
Since M contains all compatible charts it is the unique maximal atlas if it is an atlas. We
now show that it is an atlas. If (V
α
, f
α
) and (V
β
, f
β
) are in M then we must show that
f
β
◦ f
−1
α
: f
α
(V
α
∩ V
β
) → f
β
(V
α
∩ V
β
) is C
k
. Suppose m ∈ V
α
∩ V
β
. Take i ∈ I such that
m ∈ U
i
. Then on f
α
(V
α
∩V
β
∩U
i
), f
β
◦ f
−1
α
= (f
β
◦ φ
−1
i
) ◦ (φ
i
◦ f
−1
α
). Since f
β
and f
α
are
compatible with A, f
β
◦ φ
−1
i
and φ
i
◦ f
−1
α
are C
k
on open sets of R
n
to open sets of R
n
.
Therefore f
β
◦ f
−1
α
is C
k
on its domain.
Deﬁnition 2.6***. A maximal C
k
atlas is called a C
k
diﬀerential structure.
Deﬁnition 2.7***. A C
k
nmanifold is a topological nmanifold M along with a C
k
diﬀerential structure S. By Theorem 2.5***, a single atlas is enough to determine the
diﬀerential structure.
The reader should note that this deﬁnition for a C
0
structure agrees with the deﬁnition
of a topological manifold. A C
∞
nmanifold is also called a smooth manifold. The word
“manifold,” without other adjectives, will denote a smooth manifold as these
will be the subject of the remainder of this manuscript.
Usually the notation for the structure S is suppressed. However, the phrase “a manifold
M” supposes that there is an unnamed diﬀerential structure S in the background. In
particular this means that if A is an atlas, then A ⊂ S; and, if (U, φ) is a chart, then
(U, φ) ∈ S. The diﬀerential structure contains all compatible charts. For example, if
(U, φ) ∈ S and V ⊂ U is open, then (V, φ
V
) ∈ S.
Example 2.8.a***. A real ndimensional vector space is an nmanifold.
Let V be an ndimensional vector space. Pick an ordered basis v
1
, · · · , v
n
and deﬁne
charts (V, f
v
1
,··· ,v
n
), where f
v
1
,··· ,v
n
(
n
i=1
a
i
v
i
) = (a
1
, · · · , a
n
) is an isomorphism of V to R
n
.
The image is all of R
n
, which is an open subset of itself. These charts are compatible since
f
w
1
,··· ,w
n
f
−1
v
1
,··· ,v
n
is a linear automorphism on R
n
.
Recall that every real vector space is isomorphic to R
n
, however R
n
comes equiped with
a standard ordered basis or set of coordinates. A real vector space may appear in several
guises. We now mention a few instances of vector spaces that relate to matrices.
Example 2.8.b***. Let the set of n × n real matrices be denoted Mat
n×n
. The set
Mat
n×n
is a real vector space of dimension n
2
.
Let e
1
, e
2
, · · · , e
m
be the standard ordered basis for R
m
. If e
ij
is the matrix with 1
in the ith row and jth column and zero elsewhere, then {e
ij
 1 ≤ i, j ≤ n} is a basis.
The isomorphism of Mat
n×n
to R
n
2
is determined by the linear map f(e
ij
) = e
n(i−1)+j
for 1 ≤ i, j ≤ n, and allows one to think of the matrix entries as coordinates in R
n
2
. In
coordinates, if A = (a
ij
), then f(A) = (a
11
, a
12
, · · · , a
1n
, a
21
, · · · , a
nn
).
CHAPTER 2 MANIFOLDS 3
Example 2.8.c***. The symmetric matrices Sym
n×n
= {A ∈ Mat
n×n
 A
t
= A} form a
vector subspace of dimension
n(n + 1)
2
.
The entries on and above the diagonal can be arbitrary. Below the diagonal, the entries
are determined by symmetry, i.e., a
ij
= a
ji
.
Example 2.8.d***. The skew symmetric matrices Skew
n×n
= {A ∈ Mat
n×n
 A
t
= −A}
form a vector subspace of dimension
n(n −1)
2
.
The entries on the diagonal must be zero, since a
ij
= −a
ji
. The entries above the
diagonal can be arbitrary and the entries below are determined by the antisymmetry, i.e.,
a
ij
= −a
ji
.
Example 2.9a***. The sphere S
n
= {x ∈ R
n+1
 x = 1} is an nmanifold.
We construct an atlas {(U
1
, φ
1
), (U
2
, φ
2
)} with the aid of a standard wellknown map
called stereographic projection. Let U
1
= S
n
\{(0, · · · , 0, 1)} and U
2
= S
n
\{(0, · · · , 0, −1)}.
Note that U
1
∪ U
2
= S
n
. Let φ
1
(x
1
, x
2
, · · · , x
n+1
) = (
x
1
1−x
n+1
, · · · ,
x
n
1−x
n+1
). The map
φ
1
: U
1
→ R
n
is called stereographic projection. The inverse map φ
−1
1
: R
n
→ U
1
is
deﬁned by
φ
−1
1
(y
1
, · · · , y
n
) =
2y
1
n
i=1
y
2
i
+ 1
,
2y
2
n
i=1
y
2
i
+ 1
, · · · ,
2y
n
n
i=1
y
2
i
+ 1
, 1 −
2
n
i=1
y
2
i
+ 1
.
Both φ
1
and φ
−1
1
are continuous and hence φ
1
is a homeomorphism.
The second coordinate chart (U
2
, φ
2
), stereographic projection from the south pole,
is given by φ
2
= −φ
1
◦ (−1) where (1) is multiplication by −1 on the sphere. Since
multiplication by 1 is a homeomorphism of the sphere to itself (its inverse is itself), the
map φ
2
: U
2
→ R
n
is a homeomorphism.
Checking the compatability conditions, we have φ
2
◦φ
−1
1
(y
1
, · · · , y
n
) =
1
n
i=1
y
2
i
(y
1
, · · · , y
n
)
and φ
2
◦ φ
−1
1
= φ
1
◦ φ
−1
2
. Hence, S
n
is shown to be an nmanifold.
Example 2.9b***. Another atlas for the sphere S
n
.
We use 2(n + 1) coordinate charts to construct this atlas. For each i ∈ {1, · · · , n + 1}
let U
i,+
= {(x
1
, · · · , x
n+1
)  x
i
> 0} and U
i,−
= {(x
1
, · · · , x
n+1
)  x
i
> 0}. Deﬁne φ
i,+
:
U
i,+
→ R
n
by φ
i,+
(x
1
, · · · , x
n+1
) = (x
1
, · · · , x
i−1
, x
i+1
, · · · , x
n+1
) and φ
i,+
: U
i,−
→ R
n
by φ
i,−
(x
1
, · · · , x
n+1
) = (x
1
, · · · , x
i−1
, x
i+1
, · · · , x
n+1
). The coordinate x
i
is a function of
x
1
, · · · , x
i−1
, x
i+1
, · · · , x
n+1
on the sets U
i,+
and U
i,−
.
The atlases in Examples 2.9a*** and 2.9b*** are compatible and give the same diﬀer
ential structure. See Exercise 1***.
Example 2.10***. Suppose U
1
⊂ R
n
and U
2
⊂ R
m
are open sets. If f : U
1
→ U
2
is a
C
∞
function, then the graph of f,
G
f
= {(x, y) ∈ R
n+m
 y = f(x)}
4 CHAPTER 2 MANIFOLDS
is a manifold.
There is only one coordinate neighborhood required. Let π : U
1
× U
2
→ U
1
be the
projection π(x, y) = x and let i
f
: U
1
→ U
1
×U
2
be deﬁned by i
f
(x) = (x, f(x)). The one
coordinate neighborhood is (G
f
, π
G
f
). Both π and i
f
are C
∞
maps. The composites are
π
G
f
◦ i
f
= I
U
1
and i
f
◦ π
G
f
= I
G
f
. Hence, π
G
f
is a homeomorphism.
Proposition 2.11***. An open subset of an nmanifold is an nmanifold.
Proof. Suppose M is an nmanifold and U ⊂ M is an open subset. If (φ, V ) is a chart of
M, then (φ
V ∩U
, V ∩ U) is a chart for U.
While the next example uses Proposition 2.11***, it is just an open subset of R
n
2
.
Example 2.12***. GL(n, R) = {n ×n nonsingular matrices} is an n
2
manifold.
Consider the function det : Mat
n×n
→ R. We use the usual coordinates on Mat
n×n
, the
entries as was described in the ﬁrst example. In these coordinates,
det(A) =
σ∈S
n
(−1)
signσ
a
1σ(1)
a
2σ(2)
· · · a
nσ(n)
.
This function is an nth degree polynomial. Hence, it is a continuous map. The set
det
−1
(R\ 0) is an open set, the set of nonsingular matrices, GL(n, R).
If a vector space V does not have a natural choice of basis or we do not wish to focus on
the choice of basis, then we write its set of automorphisms as GL(V ). By picking a basis
for V , it becomes GL(n, R).
It is useful to see that the basic operations in GL(n, R) are C
∞
.
Proposition 2.13***. The following maps are C
∞
:
(1) GL(n, R) ×GL(n, R) → GL(n, R) by (A, B) → AB
(2) GL(n, R) → GL(n, R) by A → A
T
(3) GL(n, R) → GL(n, R) by A → A
−1
Proof. Recall that GL(n, R) is an open subset of R
n
2
and that GL(n, R) × GL(n, R) is
an open subset of R
n
2
×R
n
2
. The notion of C
∞
is from advanced Calculus.
The ﬁrst map, multiplication, is a quadratic polynomial in each coordinate as the i, j
entry of AB is
n
k=1
a
ik
b
kj
.
The second map, transpose, is just a reordering of coordinates. In fact, transpose is a
linear map.
The third map, inverse, is a rational function of the entries of A. The numerator is the
determinant of a minor of A and the denominator is detA, a polynomial that is nonzero
on GL(n, R).
Proposition 2.14***. The product of an nmanifold and an mmanifold is an (n +m)
manifold.
Proof. Suppose M is an mmanifold with atlas A
M
= {(U
i
, φ
i
)  i ∈ I} and N is an
nmanifold with atlas A
N
= {(V
j
, ψ
j
)  j ∈ J}. An atlas for M ×N is
A = {(U
i
×V
j
, φ
i
×ψ
j
)  i ∈ I, j ∈ J},
CHAPTER 2 MANIFOLDS 5
where φ
i
×ψ
j
(x, y) = (φ
i
(x), ψ
j
(y)) ∈ R
n
×R
m
for (x, y) ∈ U
i
×V
j
.
It is easy to see that A is an atlas, since the union
(i,j)∈I×J
U
i
×V
j
= M×N, and it is
easy to check compatibility. If (U
i
1
, φ
i
1
), (U
i
2
, φ
i
2
) ∈ A
M
and (V
j
1
, ψ
j
1
), (V
j
2
, ψ
j
2
) ∈ A
N
,
then
(1) (φ
i
1
×ψ
j
1
) ◦ (φ
i
2
×ψ
j
2
)
−1
= (φ
i
1
◦ (φ
i
2
)
−1
) ×(ψ
i
1
◦ (ψ
i
2
)
−1
)
on the set (φ
i
2
×ψ
j
2
)((U
i
1
×V
j
1
) ∩ (U
i
2
×V
j
2
)) = φ
i
2
(U
i
1
∩ U
i
2
) ×ψ
i
2
(V
j
1
∩ V
j
2
) an open
set in R
m
×R
n
. Since A
M
and A
N
are atlases, the right side of (1) is C
∞
, hence the left
side is also, and A is an atlas.
When can a manifold be pieced together from abstract charts? The question is of
philosophical and practical interest.
Theorem 2.15***. Let X be a set. Suppose A = {(U
i
, φ
i
)i ∈ I} satisﬁes
(1) U
i
⊂ X for each i ∈ I.
(2)
i∈I
U
i
= X.
(3) φ
i
: U
i
→ φ
i
(U
i
) ⊂ R
n
is a bijection for all i.
(4) φ
i
(U
i
), φ
i
(U
i
∩ U
j
) ⊂ R
n
are open for all i, j ∈ I.
(5) φ
j
φ
−1
i
: φ
i
(U
i
∩ U
j
) → φ
j
(U
i
∩ U
j
) is C
∞
for all i, j ∈ I.
Then there is a unique topology on X such that each U
i
is open and each φ
i
is a home
omorphism. If the topology is second countable and Hausdorﬀ then X is an nmanifold
and A is an atlas.
Remark 2.16***. If for every x, y ∈ X there are i, j ∈ I with x ∈ U
i
, y ∈ U
j
and
U
i
∩ U
j
= ∅ or there is an i ∈ I with x, y ∈ U
i
, then X is Hausdorﬀ.
If the open cover {U
i
i ∈ I} has a countable subcover, then X is second countable since
the countable union of second countable spaces is a second countable space.
Proof of Theorem 2.15***. For each x ∈ M we give a neighborhood basis. Induce a
topology on U
i
by taking O ⊂ U
i
is open if and only if O = φ
−1
i
(W) for W ⊂ R
n
open.
The map φ
i
is then a homeormorphism. We must check that this gives a well deﬁned
neighborhood basis for a topology. Suppose x ∈ U
i
∩ U
j
. The neighborhood basis is then
deﬁned by both φ
i
and φ
j
. Since φ
j
φ
−1
i
is a homeomorphism of open sets, φ
j
= (φ
j
φ
−1
i
)◦φ
i
deﬁnes a neighborhood basis of x in a manner consistent with φ
i
. This construction deﬁnes
a neighborhood basis of each point and so a topology on X. This is the only topology with
each φ
i
a homeomorphism as the basis determines the topology. We now see that Asatisﬁes
the conditions for an atlas: {U
i
i ∈ I} is an open cover by 2, φ
i
is a homeomorphism by
3 and the construction above, and the compatibility condition is 5. If X is Hausdorﬀ and
second countable, then X is an nmanifold.
Now that we know the deﬁnition of a manifold, the next basic concept is a map between
manifolds.
Deﬁnition 2.17***. Suppose f : M
m
→ N
n
is a function between manifolds. If for all
charts (U, φ) and (W, ψ) in the diﬀerential structures of M and N respectively, ψ◦ f ◦ φ
−1
is C
∞
on its domain, then f is a smooth map or function.
6 CHAPTER 2 MANIFOLDS
The reader should note that Deﬁnition 2.17*** agrees with the notion of a C
∞
map of
functions from Calculus, i.e., if O ⊂ R
m
, then f : O → R
n
is C
∞
as deﬁned in Calculus
if and only if it is smooth as a map between manifolds.
In order to check if a function between manifolds is smooth, one does not have to check
every chart in a diﬀerential structure. It is enough to check one chart about each point as
is proved in the following theorem.
Proposition 2.18***. Let f : M
m
→ N
n
be a function between manifolds. Further
suppose that for each x ∈ M there are charts (U
x
, φ
x
) about x for M and (W
f(x)
, ψ
f(x)
)
about f(x) for N such that ψ
f(x)
◦ f ◦ φ
−1
x
is C
∞
on its domain. Then f is smooth.
Proof. Suppose that (U, φ) and (W, ψ) are in the diﬀerential structure for M and N re
spectively. We wish to show that ψ ◦ f ◦ φ
−1
is C
∞
on its domain, φ(f
−1
(W) ∩ U).
Take a point in the domain of ψ ◦ f ◦ φ
−1
, say φ(x) for x ∈ f
−1
(W) ∩ U. Then
φ(f
−1
(W) ∩ U ∩ f
−1
(W
f(x)
) ∩ U
x
) is an open neighborhood of φ(x) in R
m
, and on this
open set,
ψ ◦ f ◦ φ
−1
= (ψ ◦ ψ
−1
f(x)
) ◦ (ψ
f(x)
◦ f ◦ φ
−1
x
) ◦ (φ
x
◦ φ
−1
)
The compositions in parentheses are maps between real spaces. The ﬁrst and third are
C
∞
since (W
f(x)
, ψ
f(x)
), (W, ψ) and (U
x
, φ
x
), (U, φ) are compatible pairs of charts. The
second composition is C
∞
by the hypothesis of the theorem. Hence, ψ ◦ f ◦ φ
−1
is C
∞
, so
by Deﬁnition 2.17***, f is smooth.
Proposition 2.19***. The composition of smooth functions is a smooth function. Sup
pose f : M
m
→ N
n
and g : N
n
→ K
k
are smooth functions between manifolds. Then
g ◦ f : M
m
→ K
k
is a smooth function.
Proof. Suppose p ∈ M. Suppose that (U, φ), (V, ψ), and (W, ϕ) are chart on M, N, K
respectively; and p ∈ U, f(p) ∈ V , and g(f(p)) ∈ W. The composite ϕ ◦ g ◦ f ◦ φ
−1
is
deﬁned on a neighborhood of p, and we need to show that this composite is C
∞
on some
neighborhood of p. The composite function
ϕ ◦ g ◦ ψ
−1
◦ ψ ◦ f ◦ φ
−1
= ϕ ◦ g ◦ f ◦ φ
−1
on some neighborhood of p. Note that (ϕ ◦ g ◦ ψ
−1
) ◦ (ψ ◦ f ◦ φ
−1
) is the same function
and it is C
∞
since each of the functions in parentheses is C
∞
because both f and g are
smooth.
The notion of equivalence between diﬀerential manifolds is diﬀeomorphism.
Deﬁnition 2.20***. Suppose M and N are diﬀerential manifolds. If there is a smooth
map f : M → N with a smooth inverse f
−1
: N → M, then
(1) f is called a diﬀeomorphism, and
(2) M and N are diﬀeomorphic.
We make the following simple but useful observation.
CHAPTER 2 MANIFOLDS 7
Proposition 2.21***. Suppose that M is an nmanifold, U ⊂ M, φ : U → R
n
, and
φ(U) is open in R
n
. The pair (U, φ) is a chart of the manifold M if and only if φ is a
diﬀeomorphism.
Proposition 2.22***. If M is an nmanifold, then the set of diﬀeomorphisms of M is a
group under composition.
Let Diﬀ(M) denote the group of diﬀeomorphisms of M.
Proof. The identity map on M, the map I
M
, is a smooth map, which we easily check. If
(U.ψ) is any coordinate chart for M, then φ ◦ I
M
◦ φ
−1
is just the identity on φ(U) ⊂ R
n
.
By Proposition 2.18***, I
M
is smooth.
A diﬀeomorphism f has a smooth inverse by its deﬁnition.
The composition of two diﬀeomorphisms is again a diﬀeomorphism. If f and g are
diﬀeomorphisms, the composition is smooth by Proposition 2.19***. The inverse of f ◦ g
is g
−1
◦ f
−1
.
Finally note that the composition of functions is associative. Hence, Diﬀ(M) is a
group.
We now consider a method of constructing manifolds. Those students who have al
ready learned of covering spaces will recognize the construction, although no background
is assumed in these notes.
The construction involves a subgroup G of the diﬀeomorphisms of an nmanifold M.
We can consider the quotient space M/G deﬁned by the equivalence relation on M that
x ∼ y if and only if y = g(x) for some g ∈ G. Notice that ∼ is an equivalence relation so
that M/G makes sense as a topological space:
(1) If x ∼ x since x = I
M
(x)
(2) If x ∼ y then y = g(x) and x = g
−1
(y), so y ∼ x
(3) If x ∼ y and y ∼ z, then y = g(x) and z = h(y), so z = h ◦ g(x) and x ∼ z
We know that M/G is a topological space under the quotient topology and that the
quotient map π : M → M/G is continuous. Can we guarantee it is a manifold? In general,
the answer is no, but we do have the following theorem.
Theorem 2.23***. Suppose M is an nmanifold, and G is a ﬁnite subgroup of Diﬀ(M).
Suppose that G satisﬁes one of the following, either (a) or (b):
(a) If for some x ∈ M and g ∈ G, g(x) = x, then g is the identity, or
(b) There is atlas A for M such that if (U, φ) ∈ A,
(1) then (g(U), φ ◦ g
−1
) ∈ A, and
(2) h(U) ∩ g(U) = ∅ for all g, h ∈ G, g = h.
Then M/G is an nmanifold and the quotient map π : M → M/G is a smooth map.
The condition “if for some x ∈ M and g ∈ G, g(x) = x, then g is the identity,” says that
no g, other that the identity can ﬁx a point. The group G is said to operate without ﬁxed
points. Before proving Theorem 2.23***, we ﬁrst prove the following lemma.
Lemma 2.24***. Suppose M is an nmanifold, and G is a ﬁnite subgroup of Diﬀ(M).
Suppose that G satisﬁes the following property: If for some x ∈ M and g ∈ G, g(x) = x,
then g is the identity. Then there is atlas A for M such that if (U, φ) ∈ A,
(1) then (g(U), φ ◦ g
−1
) ∈ A, and
8 CHAPTER 2 MANIFOLDS
(2) h(U) ∩ g(U) = ∅ for all g, h ∈ G, g = h.
Proof. We ﬁrst show that for every g ∈ G and x ∈ M there is an open set O
g
, a neighbor
hood of g(x), such that
(1) if g, h ∈ G and g = h then O
g
∩ O
h
= ∅.
Fix x ∈ M. For each pair g(x), h(x) there are open sets O
gh
and neighborhood of g(x)
and O
hg
and neighborhood of h(x) such that O
gh
∩ O
hg
= ∅. These sets exist since M is
Hausdorﬀ. Let O
g
=
h∈G
O
gh
. The intersection is ﬁnite, so O
g
is open. Given g, h ∈ G,
g = h, then O
g
⊂ O
gh
and O
h
⊂ O
hg
, so
O
g
∩ O
h
⊂ O
gh
∩ O
hg
= ∅.
The next step is to show there is a neighborhood of x, call it O, such that
g(O) ∩ h(O) = ∅ for any g = h.
We have g(x) ∈ O
g
, so g
−1
(O
g
) is a neighborhood of x. Let
O =
g∈G
g
−1
(O
g
).
Now, g(O) ⊂ g(g
−1
(O
g
)) = O
g
, and therefore, g(O) ∩ h(O) ⊂ O
g
∩ O
h
= ∅.
We now produce the atlas. Take a chart (U
x
, φ
x
) about x with U
x
⊂ O. Then (g(U
x
), φ
x
◦
g
−1
) is a chart about g(x). Note that h(U) ∩ g(U) ⊂ h(O) ∩ g(O) = ∅. Hence, the atlas
{(g(U
x
), φ
x
◦ g
−1
)  x ∈ M, g ∈ G} satisﬁes the required properties.
Proof of Theorem 2.23***. We ﬁrst construct the coordinate charts. Take A an atlas that
satisﬁes the two items in Lemma 2.24*** and take (U, φ) ∈ A. The quotient map π is a
continuous and open map. By the second item in Lemma 2.24***, π is also onetoone
on U. Therefore π
U
is a homeomorphism. Denote its inverse by i
U
: π(U) → U. Let
Φ = φ ◦ i
U
, then the pair (π(U), Φ) is a coordinate chart for M/G.
We have constructed charts about any point and it remains to show that these charts
are compatible. Suppose z ∈ M/G and z ∈ π(V ) ∩ π(U) where (V, ψ) ∈ A and (π(V ), Ψ)
is a chart of M/G constructed as in the previous paragraph. We wish to show that
Ψ◦ Φ
−1
: Φ(π(U) ∩ π(V )) → Ψ(π(U) ∩ π(V ))
is a diﬀeomorphism. Since it as an inverse (of the same form) we only have to show it is
C
∞
. Now, for some y ∈ M, π
−1
(z) = {g(y)  g ∈ G} a set of G points. There are two
group elements g, h ∈ G with g(y) ∈ V and h(y) ∈ U. Let O be a neighborhood of h(y) in
hg
−1
(V ) ∩ U. A neighborhood of g(y) in V ∩ gh
−1
(U) is gh
−1
(O). The map Ψ ◦ Φ
−1
on
the open set π(O) is
(1)
Ψ◦ Φ
−1
= (ψ ◦ i
V
)(φ ◦ i
U
)
−1
= ψ ◦ (i
V
◦ π
U
) ◦ φ
−1
= ψ ◦ (gh
−1
) ◦ φ
−1
CHAPTER 2 MANIFOLDS 9
since gh
−1
: O → gh
−1
(O) is a diﬀeomorphism, so is the composite (1). This completes
the proof that M/G is a manifold.
That π is a smooth map is almost a tautology. Given x ∈ M take (π(U), i
U
) a chart in
A about x and (π(U), Φ) the chart about π(x) constructed above in the ﬁrst paragraph.
The map which we are required to show is C
∞
is Φ◦ π ◦ φ
−1
= φ ◦ (i
U
◦ π
U
) ◦ φ
−1
which
is just the identity on φ(U). Therefore π is smooth.
We give two examples.
Example 2.25***. Real Projective Space RP
n
.
The nsphere, S
n
⊂ R
n+1
admits the action of the group Z
2
= {1, −1}. Multiplication
by 1 is the identity which is smooth, Theorem 2.22***. We check 1 using the charts from
Example 2.9b***. Multiplication by 1 is a map
−1 : U
i,+
→ U
i,−
and −1 : U
i,−
→ U
i,+
.
Now, both φ
i,+
◦(−1)◦φ
−1
i,−
and φ
i,−
◦(−1)◦φ
−1
i,+
are the same as maps from B
1
(0) → B
1
(0).
They are (y
1
, · · · , y
n
) → (−y
1
, · · · , −y
n
) which is C
∞
on R
n
. Hence, {1, −1} ⊂ Diﬀ(M).
Therefore, by Theorem 2.23***, S
n
/Z
2
is an nmanifold. It is called real projective space.
Denote it RP
n
.
Let q : S
n
→ RP
n
be the quotient map. Notice that if U ⊂ S
n
that is entirely in some
hemisphere of S
n
then q(U) can serve as a neighborhood for a chart since U ∩ −U = ∅.
Example 2.26***. Conﬁguration Space.
Imagine that nparticles are moving in R
m
. These particles are ideal particles in that
they are points, i.e., they have no diameter. What are the possible arrangements? If the
particles are labeled, they we can write the arrangement as an ntuple, (x
1
, x
2
, · · · , x
n
).
Since no two particles can occupy the same point in space, the manifold that describes such
arrangements is {(x
1
, x
2
, · · · , x
n
)  x
i
= x
j
for i = j}, i.e., R
mn
with several subspaces
removed. It is a manifold since it is an open subset of R
mn
. It is ordered conﬁguration
space. Denote it C
nm
If the nparticles are not labeled, then we can only know the location of the n pari
ciles and not which particle occupies which location, e.g., we cannot distinguish between
(x, y, z) and (y, x, z). The symmetric group S
n
acts on C
nm
by σ · (x
1
, x
2
, · · · , x
n
) =
(x
σ
−1
1
, x
σ
−1
2
, · · · , x
σ
−1
n
). The space C
nm
/S
n
is conﬁguration space.
We show that C
nm
/S
n
is an mnmanifold. We use Theorem 2.23***. The manifold
M is C
nm
and the subgroup of the diﬀeomorphisms is S
n
the permutation group. Each
permutation is the restriction of a linear map on R
mn
to C
nm
, and so each permutation
is C
∞
, i.e., smooth. The inverse of a permutation is again a permutation. Hence, S
n
is
a subgroup of Diﬀ(C
nm
). If σ ﬁxes x = (x
1
, · · · , x
n
), then x
i
= x
σ(i)
for i = 1, · · · , n. If
σ(i) = j and j = i, then x
i
= x
j
for an i = j. Therefore x is not in C
nm
. If σ(i) = i for
all i = 1, · · · , n, then σ is the identity. We can now apply Theorem 2.23***. By Theorem
2.23***, C
nm
/S
n
is an nmmanifold.
In physics, the term conﬁguration space is used for to describe the space of physical
conﬁgurations of a mechanical system. If m = 3, then Example 2.26*** is the conﬁguration
space of n particles in ordinary 3dimensional space.
10 CHAPTER 2 MANIFOLDS
Exercises
Exercise 1***. Verify the calculations of Example 2.9a***. Show the two atlases given
for S
n
in Example 2.9a*** and Example 2.9b*** give the same diﬀerential structure and
so may be merged.
Exercise 2***. S
1
× S
1
is a 2manifold, S
2
× S
1
is a 3manifold,and S
2
× S
1
× S
1
is a
4manifold.
Of course these all follow from Proposition 2.13***. The reader should note, however,
that there is an ambiguity in S
2
× S
1
× S
1
, is it (S
2
× S
1
) ×S
1
or S
2
× (S
1
× S
1
)? The
reader should show that the atlases are compatible and so these are the same manifold.
There is also a second approach that is sometimes used to deﬁne smooth functions. In
this approach, one ﬁrst deﬁnes a smooth function for f : M → R only. The statement of
the next exercise would be a deﬁntion in some textbooks, e.g., Warner and Helgason, but
for us, it is a proposition.
Exercise 3***. Show that a function f : M
m
→ N
n
between manifolds is smooth if and
only if for all open sets U ⊂ N and all smooth functions g : U → R, g ◦ f is smooth on its
domain.
Exercise 4***. Consider R with the following three atlases:
(1) A
1
= {f  f(x) = x}
(2) A
2
= {f  f(x) = x
3
}
(3) A
3
= {f  f(x) = x
3
+x}
Which of these atlases determines the same diﬀerential structure. Which of the manifolds
are diﬀeomorphic?
Exercise 5***. Let M, N, and Q be manifolds.
(1) Show that the projections π
1
: M ×N → M and π
2
: M ×N → N are smooth.
(2) Show that f : Q → M ×N is smooth iﬀ π
i
f is smooth for i = 1, 2.
(3) Show for b ∈ N that the inclusion x → (x, b) : M → M ×N is smooth.
The following is a diﬃcult exercise.
Exercise 6***. Prove that the set of all n × n matrices of rank k (where k < n) is a
smooth manifold. What is its dimension?
If this is too hard, then prove that the set of all n × n matrices of rank 1 is a smooth
manifold of dimension 2n −1.
CHAPTER 2 MANIFOLDS 11
Warmup Exercises, Chapter 2
Exercise 1*. Suppose that {(U
i
, φ) : i ∈ I} is an atlas for M. Argue that a chart (V, ψ)
is compatible with the atlas if for each x ∈ V , there exists an open set W, x ∈ W ⊆ V
and an i
x
∈ I such that φ
i
x
◦ (ψ
W
)
−1
and ψ
W
◦ φ
−1
i
x
are C
∞
.
Exercise 2*. Suppose that {(U
i
, φ) : i ∈ I} is an atlas for M, J ⊆ I and
i∈J
U
i
= M.
Argue that {(U
i
, φ) : i ∈ I} is an atlas that generates the same diﬀerentiable structure on
M.
Exercise 3*. Let φ : U → R
n
be a chart for a smooth manifold M and let V be a
nonempty open subset of U. Argue that φ
V
: V → R
n
is also a chart in the diﬀerentiable
structure of M.
Exercise 4*. Let U be a nonempty open subset of a manifold M. Show that the charts
of M with domain contained in U form a diﬀerentiable structure on U. Show that the
restriction of any chart on M to U belongs to this diﬀerentiable structure.
Exercise 5*. Prove Proposition 2.21***.
CHAPTER 3
SUBMANIFOLDS
One basic notion that was not addressed in Chapter 1 is the concept of containment:
When one manifold is a subset on a second manifold. This notion is subtle in a manner
the reader may have experienced in general topology. We give an example.
Example 3.1. Let Q
d
be the rational numbers with the discrete topology and R the
usual real numbers. The inclusion ι : Q
d
→ R is a onetoone continuous map.
Is Q
d
a subspace of R? The usual answer is no, it is only a subspace if ι is a homeomor
phism to its image, ι(Q
d
) = Q. This map is not a homeomorphism to its image since the
topology on Q, the subspace topology, is not discrete. the same issue arises in manifold
theory. In fact Q
d
is a 0dimensional manifold and ι is a smooth map. We, however, will
not refer to Q
d
as a submanifold of R, i.e., submanifold is not quite the correct relationship
of Q
d
to R. This relationship will be studied in Chapter 6*** (immersed submanifold). For
the notion “N is a submanifold of M,” we require that N inherits its diﬀerential structure
from M. Some authors refer to this relationship as “embedded submanifold.”
We give the deﬁnition of submanifold in Deﬁnition 3.2***. At ﬁrst glance, it may appear
to be overly restrictive, however, it turns out not to be the case. It is analogous to the
notion of subspace topology. This assertion is justiﬁed in a later chapter, Theorem 6.3***.
Deﬁnition 3.2***. Suppose m > n and write R
m
= R
n
× R
m−n
. Let M be an m
manifold and N ⊂ M. Suppose that for each x ∈ N there is a chart of M
(1) (U, φ) centered at x such that φ
−1
(R
n
×{0}) = U ∩ N
Then N is an ndimensional submanifold of M.
Charts that satisfy property (1) from Deﬁnition 3.2*** are called slice charts or slice
coordinate neighborhoods.
We next observe that if N ⊂ M is an ndimensional submanifold, then N is an n
manifold.
Proposition 3.3***. Suppose that N is an ndimensional submanifold of the mmanifold
M. Then N is an nmanifold and
A ={(U
x
∩ N, φ
U∩N
)  (U
x
, φ) is a chart of M centered at x ∈ N such
that φ
−1
(R
n
×{0}) = U ∩ N}
is an atlas for N.
Proof. We ﬁrst show that N is a topological manifold in the induced topology on N. The
topology on N is the subspace topology. Therefore, N is Hausdorﬀ and second countable
copyright c 2002
Typeset by A
M
ST
E
X
1
2 CHAPTER 3 SUBMANIFOLDS
since M is Hausdorﬀ and second countable. Also since the topology on N is the subspace
topology, U
x
∩ N is an open set in N. The set φ(U
x
∩ N) = φ(U
x
) ∩ (R
n
× {0}), since
(U
x
, φ) is a slice neighborhood. The map φ is a homeomorphism to its image, so φ
U
x
∩N
is a homeomorphism to its image. Now, since there is a neighborhood U
x
∩N around any
x ∈ N, the requirements of Deﬁnition 2.1*** are shown and N is a topological manifold.
To show that N is an nmanifold, we must show that the charts satisfy the compatibility
condition of Deﬁnition 2.3*** item 2. Suppose (U
y
, ψ) is another coordinate neighborhood
in the potential atlas A. Then
φ
U
x
∩N
(ψ
U
y
∩N
))
−1

ψ(U
x
∩U
y
∩N)
= (φψ
−1
)
ψ(U
x
∩U
y
∩N)
is C
∞
since it is the composition of the inclusion of R
n
into R
m
composed with φψ
−1
,
two C
∞
functions. This completes the proof.
Theorem 3.4***. Suppose O ⊂ R
n+m
is an open set and f : O → R
m
is a C
∞
map. Let
q ∈ R
m
and M = f
−1
(q). If Df(x) has rank m for all x ∈ M, then M is an ndimensional
submanifold of R
n+m
.
Proof. For each x ∈ M, there is a neighborhood of x, U
x
⊂ R
n+m
; an open set U
1
⊂ R
n
;
an open set U
2
⊂ R
m
; and a diﬀeomorphism H : U
x
→ U
1
× U
2
such that H(M ∩ U
x
) =
U
1
×{0} as was laid out in the Rank Theorem. The pair (U
x
, H) is a chart for the smooth
manifold R
n+m
since H is a diﬀeormorphism, Proposition 2.21***. This chart satisﬁes the
required property given in Deﬁnition 3.2***. Hence, M is an ndimensional submanifold
of R
n+m
.
Example 3.5***. S
n
is a submanifold of R
n+1
.
Let l : R
n+1
→ R by l(x
1
, x
2
, · · · , x
n+1
) =
x
2
1
+· · · x
2
n+1
. Then S
n
= l
−1
(1). The
map l is C
∞
on R
n+1
\ 0. In fact, each partial derivative is a rational function of
x
1
, · · · , x
n+1
, and l, i.e.,
∂l
∂x
i
=
x
i
l
. Therefore the partial derivative of a rational func
tion of x
1
, · · · , x
n+1
, l is another such function and l is C
∞
. To check that the rank of
D
v
l(x) is one, it is enough to show that some directional derivative is not zero. Hence, for
x ∈ S
n
, we compute,
D
x
l(x) =
d
dt
l(x +tx)
t=0
=
d
dt
(1 +t)
x
2
1
+· · · x
2
n+1
t=0
=
d
dt
(1 +t)
t=0
= 1.
Since l has rank 1 on S
n
, Theorem 3.4*** applies.
Proposition 3.3*** along with the Rank Theorem*** gives instructions for computing
an atlas. The atlas essentially comes from the Implicit Function Theorem. The charts
include the atlas in Example 2.9b***. The reader should check this fact.
CHAPTER 3 SUBMANIFOLDS 3
Example 3.6***. Let SL(n, R) = {A ∈ Mat
n×n
 detA = 1}. Then
(1) If A ∈ SL(n, R), then Ddet(A) has rank 1.
(2) SL(n, R) is an (n
2
−1)manifold.
First, det : R
n
2
→ R is a polynomial (as shown in Example 2.12***) and so C
∞
. To
show that Ddet(A) has rank 1 it is only necessary to show that some directional derivative
D
B
det(A) is nonzero. We compute for A ∈ SL(n, R),
D
A
det(A) =
d
dt
det(A+tA)
t=0
=
d
dt
(1 +t)
n
det(A)
t=0
=
d
dt
(1 +t)
n

t=0
= n.
Hence item 1 is shown.
To see item 2, we use Theorem 3.4***. The set GL(n, R) ⊂ R
n
2
−1
×Ris an open set and
det : GL(n, R) → R is a C
∞
map. The set det
−1
(1) = SL(n, R), and Ddet(A) has rank 1
for each A ∈ SL(n, R). Therefore, by Theorem 3.4***, SL(n, R) is an (n
2
−1)manifold.
Example 3.7***. Let O(n, R) = {A ∈ Mat
n×n
 AA
T
= I} and let SO(n, R) = {A ∈
Mat
n×n
 AA
T
= I and detA = 1}. Also let f : Mat
n×n
→ Mat
n×n
by f(A) = AA
T
(1) f(A) ∈ Sym
n×n
and f is C
∞
.
(2) If A ∈ O(n, R), then Df(A) has rank
n(n + 1)
2
.
(3) O(n, R) and SO(n, R) are
n(n −1)
2
manifolds.
First note that AA
T
= I implies that (det(A))
2
= 1 so A is invertible.
The map f is a composition the identity cross the transpose and multiplication. These
maps are C
∞
maps by Proposition 2.13***. Since f is the composition of C
∞
maps, it
is C
∞
. Since f(A) = AA
T
and (AA
T
)
T
= AA
T
, f(A) ∈ Sym
n×n
. This argument shows
item 1.
To show item 2, it is enough to show that Df(A) is surjective since dimSym
n×n
=
n(n + 1)
2
. Now,
Df(A)(M) =
d
dt
(A+tM)(A+tM)
T
t=0
= AM
T
+MA
T
= AM
T
+ (AM
T
)
T
Hence Df(A) is a composition of two maps
Mat
n×n
→ Mat
n×n
M → MA
T
and
Mat
n×n
→ Sym
n×n
X → X +X
T
4 CHAPTER 3 SUBMANIFOLDS
The ﬁrst map is onto if A
T
is invertible, which it is if A ∈ f
−1
(I). The second map is
onto, since, if Y ∈ Sym
n×n
then Y = Y
T
and Y =
1
2
Y + (
1
2
Y )
T
, i.e., if X =
1
2
Y , then
Y = X +X
T
. Therefore, Df(A) is a surjection and has rank
n(n + 1)
2
.
The third item follows from Theorem 3.4***. Since f
−1
(I) = O(n, R), O(n, R) is a
manifold of dimension n
2
−
n(n + 1)
2
=
n(n −1)
2
. The determinant function is continuous
on Mat
n×n
and O(n, R) ⊂ Mat
n×n
. If AA
T
= I, then detA = ±1. Therefore SO(n, R)
is an open subset of O(n, R) and hence a manifold of dimension
n(n −1)
2
, by Proposition
2.11***.
The next example requires notation and a lemma.
Lemma 3.8***. Suppose that n is even and let J be the n ×n matrix
J =
0
n/2×n/2
−I
n/2×n/2
I
n/2×n/2
0
n/2×n/2
Then J
T
= −J = J
−1
.
Proof. Just compute the transpose, the negative, and the inverse.
.
Example 3.9***. Suppose that n is even and let Sp(n, R) = {A ∈ Mat
n×n
 AJA
T
= J}.
Also let f : Mat
n×n
→ Mat
n×n
by f(A) = AJA
T
(1) f(A) ∈ Skew
n×n
and f is C
∞
.
(2) If A ∈ Sp(n, R), then Df(A) has rank
n(n −1)
2
.
(3) Sp(n, R) is an
n(n + 1)
2
manifold.
First note that AJA
T
= J implies that (det(A))
2
= 1 so A is invertible.
The map f is a composition the identity cross the transpose and multiplication. These
maps are C
∞
maps by Proposition 2.13***. Since f is the composition of C
∞
maps, it
is C
∞
. Since f(A) = AJA
T
and (AJA
T
)
T
= AJ
T
A
T
= −AJA
T
by Lemma 3.8***,
f(A) ∈ Skew
n×n
. This argument shows item 1.
To show item 2, it is enough to show that Df(A) is surjective since dimSkew
n×n
=
n(n −1)
2
. Now,
Df(A)(M) =
d
dt
(A+tM)J(A+tM)
T
t=0
= AJM
T
+MJA
T
= AJM
T
−(AJM
T
)
T
Hence Df(A) is a composition of two maps
Mat
n×n
→ Mat
n×n
M → MJA
T
and
Mat
n×n
→ Sym
n×n
X → X −X
T
CHAPTER 3 SUBMANIFOLDS 5
The ﬁrst map is onto if JA
T
is invertible, which it is if A ∈ f
−1
(J). The second map is
onto, since, if Y ∈ Skew
n×n
then Y = −Y
T
and Y =
1
2
Y − (
1
2
Y )
T
, i.e., if X =
1
2
Y , then
Y = X −X
T
. Therefore, Df(A) is a surjection and has rank
n(n −1)
2
.
The third item follows from Theorem 3.4***. Since f
−1
(I) = Sp(n, R), Sp(n, R) is a
manifold of dimension n
2
−
n(n −1)
2
=
n(n + 1)
2
.
Proposition 3.10***. Suppose that N is an ndimensional submanifold of the mmanifold
M. Suppose that U is an open neighborhood of N in M and g : U → P is a smooth map
to a manifold P. Then, g
N
: N → P is a smooth map.
Proof. Suppose x ∈ N is an arbitrary point and (O, η) is a chart of P about g(x). By
Proposition 2.18***, it is enough to check that there is a chart of N, (V, ψ), with x ∈ V
and η◦g
M
◦ψ
−1
is C
∞
. By the deﬁnition of submanifold, Deﬁnition 3.2***, we can always
ﬁnd a product chart (of M) centered at x. Suppose the chart is (W, φ), so we can take
V = W ∩ N and ψ = φ
W∩N
. Then
η ◦ g
M
◦ ψ
−1
= η ◦ g
M
◦ (φ
−1
)
R
n
×{0}∩φ(W)
,
the restriction of a C
∞
function of m variables to its ﬁrst n variables by setting the last
m−n variables to zero. This map is C
∞
.
Example 3.11***.
Suppose that M is either of O(n), SO(n), SL(n, R), or Sp(n, R). Then multiplication
M ×M → M
(A, B) → AB
and inverse
M → M
A → A
−1
are smooth maps.
Just as M is a submanifold of GL(n, R), M × M is also a submanifold of GL(n, R) ×
GL(n, R). Both GL(n, R) and GL(n, R) ×GL(n, R) are open subsets of real spaces. Mul
tiplication and inverse are C
∞
functions as explained in Example 2.13***. By Proposition
3.10***, these maps are smooth maps between manifolds.
6 CHAPTER 3 SUBMANIFOLDS
Exercises
Exercise 1***. Suppose O ⊂ R
n+m
is an open set, f : O → R
m
is a C
∞
map, q ∈ R
m
,
and M = f
−1
(q). Further suppose U ⊂ R
s+k
is an open set, g : U → R
k
is a C
∞
map,
p ∈ R
k
, and n = g
−1
(p). If Df(x) has rank m for all x ∈ M and Dg(x) has rank k for
all x ∈ N, then show that M and N are manifolds; and, (f × g)
−1
(q, p) is the manifold
M ×N.
Exercise 2***. Show that the atlas for S
n
constructed in Example 3.5*** does include
the 2(n + 1) charts constructed in Example 2.9b***.
Exercise 3***. Let M be deﬁned by
M = {(w, x, y, z) ∈ R
4
 w
3
+x
2
+y
3
+z
2
= 0 and ye
z
= we
x
+ 2}.
Show that M is a two dimensional submanifold of R
4
.
The reader may wish to compare the following Exercise to Example 2.10***.
Exercise 4***. Suppose U
1
⊂ R
n
and U
2
⊂ R
m
are open sets. If f : U
1
→ U
2
is a C
∞
function, then show the graph of f,
G
f
= {(x, y) ∈ R
n+m
 y = f(x)}
is a submanifold of R
n+m
.
In the exercise below, we use that R
4
∼
= C
2
under the isomorphism of real vector spaces,
(w, x, y, z) → (w +xi, y +zi).
Exercise 5***. a. Suppose p and q are relatively prime integers. Let ω = e
2πi
p
. Show
that τ : S
3
→ S
3
by τ(z
1
, z
2
) = (ωz
1
, ω
q
z
2
) is a smooth map. Let G = {τ, τ
2
, · · · , τ
p
}.
Show that G is a subgroup of Diﬀ(S
3
).
b. Show that S
3
is {(z
1
, z
2
) ∈ C
2
 z
1
¯ z
1
+z
2
¯ z
2
= 1}, and a submanifold of C
2
.
c. Show that S
3
/G is a 3manifold. It is denoted L(p, q) and is called a lens space.
CHAPTER 3 SUBMANIFOLDS 7
Warmup Exercises, Chapter 3
Exercise 1*. Let F : R
2
→ R be deﬁned by
F(x, y) = x
3
+xy +y
3
.
Which level sets are embedded submanifolds of R.
Exercise 2*. Consider the map F : R
4
→ R
2
deﬁned by
F(x
1
, x
2
, x
3
, x
4
) = (x
2
1
+x
2
, x
2
1
+x
2
2
+x
2
3
+x
2
4
+x
2
).
Show that F restricted to M = F
−1
(0, 1) has rank 2 at every point of M.
Exercise 3*. Which level sets of
f(x
1
, . . . , x
n+
1
) = x
1
x
2
· · · x
n+1
+ 1
are submanifolds, according to Theorem 3.4***.
CHAPTER 4
PARTITIONS OF UNITY
AND SMOOTH FUNCTIONS
In this section, we construct a technical device for extending some local constructions
to global constructions. It is called a partition of unity. We also use the opportunity to
discuss C
∞
functions. We begin with examples of C
∞
functions on R and R
n
. Some of
these are required for the construction of partitions of unity. At the end of the chapter, we
return to C
∞
funcitons and apply our new techniques. The ﬁrst examples we construct,
Examples 4.1abc***, are standard and we follow Warner.
Example 4.1a***. A function on R which is C
∞
but not analytic. Let
f(t) =
1
e
1/t
if t > 0
0 if t ≤ 0.
We show that for each n = 0, 1, 2, 3, · · · there is a polynomial p
n
such that
f
(n)
(t) =
p
n
(1/t)
1
e
1/t
if t > 0
0 if t ≤ 0.
We use mathematical induction to show the claim. The claim is true for n = 0 as f
(0)
= f.
Here p
0
(x) = 1. For k > 0 we separately handle t > 0, t < 0 and t = 0. If f
(k)
(t) =
p
k
(1/t)
1
e
1/t
for t ∈ (0, ∞), then
f
(k+1)
(t) = p
k
(1/t)(−
1
t
2
)
1
e
1/t
+p
k
(1/t)(
1
t
2
)
1
e
1/t
and p
k+1
(x) = −x
2
p
k
(x) + x
2
p
k
(x). For t ∈ (−∞, 0), f is zero and so are its derivatives
of all orders. At t = 0 we use the deﬁnition of the derivative.
lim
h→0
+
f
(k)
(h) −f
(k)
(0)
h
= lim
h→0
+
p
k
(1/h)
1
e
1/h
h
= lim
x→∞
xp
k
(x)
e
x
= 0.
The substitution was x = 1/h, and the last limit was a polynomial divided by an exponen
tial: a standard L’Hopitals Theorem example from Calculus. The limit, lim
h→0
−
f
(k)
(h)−f
(k)
(0)
h
is zero as the numerator is identically zero.
copyright c 2002
Typeset by A
M
ST
E
X
1
2 CHAPTER 4 PARTITIONS OF UNITY AND SMOOTH FUNCTIONS
Example 4.1b***. Let
g(t) =
f(t)
f(t) +f(1 −t)
=
0 if t ≤ 0
1
1 +
e
1/t
e
1/(1−t)
if 0 < t < 1
1 if t ≥ 1.
Then g is C
∞
on R and g is strictly increasing on [0, 1].
The function g is C
∞
since it is a quotient of C
∞
functions and the denominator is
never 0. If t ≤ 0, then f(t) = 0 so g(t) = 0. If t ≥ 1, then f(1 −t) = 0 so g(t) =
f(t)
f(t)
= 1.
Recall that f
(t) =
1
t
2
f(t) for t > 0, and so,
d f(t)
dt
= −f
(1 −t). Now, for t ∈ (0, 1),
g
(t) =
f
(t)(f(t) +f(1 −t)) −f(t)(f
(t) −f
(1 −t))
(f(t) +f(1 −t))
2
=
f
(t)f(1 −t) +f(t)f
(1 −t)
(f(t) +f(1 −t))
2
=
f(t)f(1 −t)
(f(t) +f(1 −t))
2
1
t
2
+
1
(1 −t)
2
This expression is positive on (0, 1). Hence g is strictly increasing on [0, 1]. The graph of
g is as follows.
Example 4.1c***. The bump function on (−2, 2). Let
h(t) = g(t + 2)g(2 −t) =
0 if t ≤ −2
g(t + 2) if −2 < t < −1
1 if −1 ≤ t ≤ 1
g(2 −t) if 1 < t < 2
0 if 2 ≤ t
The function h is C
∞
on all of R, h(x) = 0 if x ∈ (−∞, −2] ∪ [2, ∞), h(x) = 1 if
x ∈ [−1, 1], h is strictly increasing on [−2, −1], and strictly decreasing on [1, 2].
Its graph is shown below.
Let C(r) denote the open cube in R
n
which is {(x
1
, · · · , x
n
)x
i
∈ (−r, r) for i = 1, 2, 3, · · · , n}
and let
¯
C(r) be its closure.
Example 4.2***. The bump function on C(2) ⊂ R
n
. Let
b(x
1
, · · · , x
n
) = Π
n
i=1
b(x
i
).
Then b(x
1
, · · · , x
n
) = 0 on the complement of C(2), b(x
1
, · · · , x
n
) = 1 for (x
1
, · · · , x
n
) in
the closure of C(1), and 0 < b(x
1
, · · · , x
n
) < 1 otherwise.
CHAPTER 4 PARTITIONS OF UNITY AND SMOOTH FUNCTIONS 3
Deﬁnition 4.3***. If U = {U
α
α ∈ A} is an open cover of a manifold M, then a subset
of U which is also a cover is called a subcover.
Deﬁnition 4.4***. If U = {U
α
α ∈ A} is an open cover of a manifold M, then the open
cover V = {V
γ
γ ∈ Γ} is a reﬁnement if for all γ ∈ Γ there is an α ∈ A such that V
γ
⊂ U
α
.
Deﬁnition 4.5***. A collection of subsets U = {U
α
α ∈ A} of a manifold M is called
locally ﬁnite, if for all m ∈ M there is an neighborhood O of m with U
α
∩ O = ∅ for only
a ﬁnite subset of A.
Deﬁnition 4.6***. A partition of unity on a manifold M is a collection of smooth func
tions {φ
i
: M → R  i ∈ I} such that
(1) { the support of φ
i
 i ∈ I} is locally ﬁnite
(2) φ
i
(p) ≥ 0 for all p ∈ M, i ∈ I, and,
(3)
i∈I
φ
i
(p) = 1 for all p ∈ M.
Note that the sum is ﬁnite for each p.
Deﬁnition 4.7***. The partition of unity on a manifold M {φ
i
 i ∈ I} is subordinate
to the open cover U = {U
α
α ∈ A} if for all i ∈ I there is an α ∈ A such that the support
of φ
i
is in U
α
.
Lemma 4.8***. Suppose M is a connected manifold. Then there is a sequence of open
sets O
i
such that
(1)
¯
O
i
is compact
(2)
¯
O
i
⊂ O
i+1
(3)
∞
i=1
O
i
= M
Proof. Take a countable basis for the topology of M (as M is second countable) and for each
x ∈ M pick a compact set K
x
that contains x in its interior (as M is locally compact). Since
M is Hausdorﬀ, we obtain another basis for the topology of M by keeping only those basis
elements which are in some K
x
. We now have a countable basis U = {U
i
i = 1, 2, 3, · · · }
such that if U
i
∈ U then
¯
U
i
is compact.
Let O
1
= U
1
. Each of the other open sets will have the form O
k
=
j
k
i=1
U
i
. Suppose we
have constructed O
k
. We show how to construct O
k+1
. Since
¯
O
k
is compact let j
k+1
be
the smallest counting number with
¯
O
k
⊂
j
k+1
i=1
U
i
. We establish the required properties.
Since
¯
O
k
=
j
k
i=1
¯
U
i
is a ﬁnite union of compact spaces, it is compact. By construction,
¯
O
k
⊂ O
k+1
. If
¯
O
k
⊂ O
k
then O
k
= M as O
k
is open and closed, otherwise j
k+1
> j
k
.
Therefore
k
i=1
U
i
⊂ O
k
so
∞
i=1
O
i
= M, and (3) follows.
Proposition 4.9***. Suppose U = {U
α
α ∈ A} is a basis for the manifold M with
¯
U
α
compact for all α ∈ A. Suppose W = {W
β
β ∈ B} is any open cover. Then there is a
countable locally ﬁnite reﬁnement of W, V = {V
i
i ∈ I}, with V
i
∈ U for all i ∈ I.
Proof. If the manifold has more than one component then we separately handle each
component, hence we assume that M is connected. By the previous lemma, there is a
4 CHAPTER 4 PARTITIONS OF UNITY AND SMOOTH FUNCTIONS
collection of open sets {O
i
i = 1, 2, 3, · · · } which satisfy the conditions of the lemma.
Now
¯
O
i+1
\ O
i
is compact and contained in the open set O
i+2
\
¯
O
i−1
. Note that O =
{O
i+2
\
¯
O
i−1
, O
4
 i = 3, 4, · · · } is an open cover of M.
We construct V by reducing U in two steps. Let U
be the set of all U
α
∈ U such that
there is a β with U
α
⊂ W
β
and that U
α
⊂ O
i+2
\
¯
O
i−1
or O
4
. The set U
is a basis since
O and W are open covers and M is Hausdorﬀ, i.e., for each x ∈ M, U
contains a nbhd
basis. Take a ﬁnite subset of U
each of which is in O
4
and covers
¯
O
3
, a compact set. For
each i > 2 take ﬁnite subsets of U
each of which is in O
i+2
\
¯
O
i−1
and covers
¯
O
i+1
\ O
i
, a
compact set. The union of these various ﬁnite collections V = {V
j
j ∈ I} is locally ﬁnite
since an open set in the i
th
collection can only meet open sets from the (i −2)
nd
collection
up through the (i +2)
nd
. These are each ﬁnite collections. The set V is a countable union
of ﬁnite sets and so countable. The cover V is subordinate to W since U
is subordinate
to W.
Lemma 4.10***. Suppose M is a manifold. Then there is a basis {U
α
α ∈ A} such that
(1)
¯
U
α
is compact and
(2) For each α ∈ A there is a smooth function ϕ
α
: M → R such that ϕ
α
(x) = 0 if
x / ∈ U
α
and ϕ
α
(x) > 0 if x ∈ U
α
.
Notice that the function guaranteed in the lemma cannot be analytic but must be C
∞
.
For example on R the support of ϕ
α
is compact and so the function is zero on (m, ∞) for
some m. Hence if x > m, then ϕ
(n)
α
(x) = 0, for n = 1, 2, 3, · · · .
Proof. First note that R(m) = {φ
−1
(C(2))(U, φ)is a chart centered at m and C(3) ⊂
φ(U)} is a neighborhood basis at m. If fact, if (U, φ) is any chart centered at m the
charts (U, kφ) for k large suﬃce. Now, we produce the function. If R ∈ R(m), then there
is a C
∞
function ϕ
R
: M → R with ϕ
R
(x) = 0 if x / ∈ R and ϕ
R
(x) > 0 if x ∈ R. If (U, φ)
is a chart with C(3) ⊂ φ(U) and R = φ
−1
(C(2)) ∈ R(m), then take
ϕ
R
(x) =
b ◦ φ(x) if x ∈ U
0 if x / ∈
¯
R
Here b is the C
∞
function we produced on R
n
with b(x) > 0 for x ∈ C(2) and b(x) = 0
for x / ∈ C(2). The function ϕ
R
is smooth since b ◦ φ and 0 agree on the overlap of their
domains, U \
¯
R an open set.
Theorem 4.12***. If M is a manifold and W is any open cover, then M admits a
countable partition of unity subordinate to the cover W with the support of each function
compact.
Proof. Apply Proposition 4.9*** to the basis constructed in Lemma 4.10***. We obtain
a locally ﬁnite collection {U
i
i = 1, 2, 3, · · · } with ϕ
i
: M → R as in the lemma. Let
ϕ(x) =
∞
i=0
ϕ
i
(x). Recall that for each x there is an open set O with x ∈ O and ϕ
i
= 0 on
O for all but ﬁnitely many i. The sum is ﬁnite for each x. It is the fact that locally ﬁnite
requires an open set about each x that meets only a ﬁnite number of U
i
(say i = 1, · · · , m)
CHAPTER 4 PARTITIONS OF UNITY AND SMOOTH FUNCTIONS 5
that gives the C
∞
diﬀerentiablity, since on the open set O the function ϕ is a ﬁnite sum
of smooth functions: the composition of
M
(ϕ
1
,··· ,ϕ
m
)
−−−−−−−→ R
m
Σ
−→ R.
If only each x were in a ﬁnite number of U
i
, then we could only guarantee welldeﬁned
but not smooth. Let ψ
i
=
ϕ
i
ϕ
. Then {ψ
i
i = 1, 2, 3, · · · } form a partition of unity with the
support of ψ
i
being U
i
.
It is interesting to note that there is no special ﬁeld to study the zeros of C
∞
functions.
Algebraic geometry studies the sets which are zeros of polynomials, and the nature of the
zeros of analytic functions is also studied, but the sets which are zeros of smooth functions
does not give a new area of study. The reason is:
Theorem 4.13***. If X ⊂ M is closed, then there is a smooth function f : M → R with
f(x) = 0 if and only if x ∈ X.
Proof. Let N = M \ X. N is an open subset of M and so a manifold. Let {U
α
α ∈ A}
be the basis of M produced in the lemma. Let U = {U
α
α ∈ A and U
α
⊂ N}. The
collection U is a basis for N with each
¯
U
α
compact for U
α
∈ U. We apply Proposition
4.9*** to the manifold N with W = {N} to get a countable locally ﬁnite subcollection of
U, {U
i
i = 1, 2, 3, · · · }. Each U
i
is equipped with φ
i
: M → R whose support is exactly
U
i
, as in Lemma 4.10***. Let f =
∞
i=1
φ
i
. The function f is C
∞
and is nonzero exacty on
∪
∞
i=1
U
i
= N
Theorem 4.13*** demonstrates a dramatic array of possible behavior of C
∞
functions
compared to analytic functions. To observe the restricted nature of the zeros of analytic
fuctions, we consider functions on the real line.
Theorem 4.14***. Suppose f : R → R and X = {x  f(x) = 0}.
(1) If f is a polynomial, then X is a ﬁnite set.
(2) If f is an analytic function, then X is a discrete set.
(3) If f is a C
∞
function, then X can be an arbitrary closed set.
Proof. An nth degree polynomial has at most n zeros, which shows the ﬁrst item. The
third item follows from Theorem 4.13***.
To show the second item we suppose that the zeros of f are not discrete and show that
implies f is identically zero. Suppose that p
m
∈ X for m = 1, 2, 3, · · · and lim
m→∞
p
m
= p.
Now,
f(x) =
∞
n=0
1
n!
f
(n)
(p)(x −p)
n
.
In order to show f is the zero function, it is enough to show that f
(n)
(p) = 0 for all whole
numbers n.
We show that f
(n)
(p) = 0 for all whole numbers n by mathematical induction. We
ﬁrst observe it is true for n = 0, f
(0)
= f. By continuity, lim
m→∞
f(p
m
) = f(p) and each
6 CHAPTER 4 PARTITIONS OF UNITY AND SMOOTH FUNCTIONS
f(p
m
) = 0, and therefore, f(p) = 0. We now assume that f
(k)
(p) = 0 for k < n and show
this implies f
(n)
(p) = 0. First note that
lim
m→∞
n!
f(p
m
)
(p
m
−p)
n
= lim
m→∞
n!
0
(p −p
m
)
n
= 0.
Now compute the same limit using L’Hospital’s rule,
0 = lim
m→∞
n!
f(p
m
)
(p
m
−p)
n
= lim
x→p
n!
f(x)
((x −p)
n
= lim
x→p
n!
f
(x)
n(x −p)
(
n −1)
.
.
.
= lim
x→p
n!
f
(n−1)
(x)
n!(x −p)
.
By the induction hypothesis, these limits are all of the indeterminate form
0
0
. One last
application of L’ Hopitals rule yields,
lim
m→∞
n!
f(p
m
)
(p
m
−p)
n
= lim
x→p
n!
f
n
(x)
n!
= f
(n)
(p)
the last equality by continuity of the nth dervative. Therefore f
(n)
(p) = 0 for all whole
numbers n, and f is identically zero.
Exercises
Exercise 1***. Show that the bump function b on R
n
has the following property: if
b(x) = 0, then all of the derivatives of b at x are also zero. Show that if X ⊂ R
n
is a
closed set, then there is a function f : R
n
→ R such that f(x) = 0 if and only if x ∈ X
and all partials of f vanish on X.
Exercise 2***. Assume the following version of the StoneWeirerstrass Theorem: If K ⊂
R
n
is compact and f : K → R is continuous, then given any > 0 there is a polynomial
function g such that  g
K
(x) −f(x) < for all x ∈ K.
Prove that is M is a smooth manifold and f : M → R is continuous, then given any
> 0 there is a C
∞
function g : M → R
n
such that f(x) −g(x) < for all x ∈ M.
CHAPTER 5
TANGENT VECTORS
In R
n
tangent vectors can be viewed from two perspectives
(1) they capture the inﬁnitesimal movement along a path, the direction, and
(2) they operate on functions by directional derivatives.
The ﬁrst viewpoint is more familiar as a conceptual viewpoint from Calculus. If a point
moves so that its position at time t is ρ(t), then its velocity vector at ρ(0) is ρ
(0), a tangent
vector. Because of the conceptual familiarity, we will begin with the ﬁrst viewpoint,
although there are technical diﬃculties to overcome. The second interpretation will be
derived as a theorem. The second viewpoint is easier to generalize to a manifold. For
instance, operators already form a vector space. It is the second viewpoint that ultimately
plays the more important role.
Suppose M is an nmanifold. If m ∈ M, then we deﬁne a tangent vector at m as
an equivalence class of paths through m. Equivalent paths will have the same derivative
vector at m and so represent a tangent vector. The set of all tangent vectors at m forms the
tangent space. The description and notation of tangent vectors in R
n
from the advanced
Calculus setting and in the present setting is discussed in Remark 5.9***.
Deﬁnition 5.1***. Suppose M is a manifold. A path is a smooth map ρ : (−, ) → M,
where > 0.
As was mentioned, if M = R
n
, then ρ
(0) is the velocity vector at ρ(0). We also recall,
from advanced Calculus, the relationship between the derivative map and the directional
derivative,
(1) Dρ(0)(1) = D
1
ρ(0) = ρ
(0)
Deﬁnition 5.2***. Suppose M is a manifold and m ∈ M A tangent vector at m is an
equivalence class of paths α with α(0) = m. Let (U, φ) be a coordinate chart centered at
m, two paths α and β are equivalent if
dφ ◦ α(t)
dt
t=0
=
dφ ◦ β(t)
dt
t=0
.
Denote the equivalence class of a path α by [α]. We can picture [α] as the velocity vector
at α(0).
We next observe that the equivalence class doesn’t depend on the speciﬁc choice of
a coordinate chart. If (W, ψ) is another coordinate neighborhood centered at m, then
ψ ◦ α = ψ ◦ φ
−1
◦ φ ◦ α, and, we use formula (1),
dψ ◦ α(t)
dt
t=0
= D(ψ ◦ φ
−1
)(φ(m)) ◦ D(φ ◦ α)(0)(1).
copyright c 2002
Typeset by A
M
ST
E
X
1
2 CHAPTER 5 TANGENT VECTORS
The diﬀeomorphisms ψ ◦ φ
−1
and φ ◦ α are maps between neighborhoods in real vector
spaces, so
dφ ◦ α(t)
dt
t=0
=
dφ ◦ β(t)
dt
t=0
if and only if
dψ ◦ α(t)
dt
t=0
=
dψ ◦ β(t)
dt
t=0
.
Therefore the notion of tangent vector is independent of the coordinate neighborhood. If
ρ : (−, ) → M is a path in M with ρ(0) = m, then [ρ] is a tangent vector to M at m and
is represented by the path ρ. Consistent with the notation for R
n
, we can denote [ρ] by
ρ
(0).
Let TM
m
denote the set of tangent vectors to M at m. Other common notations are
M
m
and T
m
M.
Theorem 5.3***. Suppose M, N, and R are manifolds.
(1) If φ : M → N is a smooth map between manifolds and m ∈ M then there is an
induced map φ
∗m
: TM
m
→ TN
φ(m)
.
(2) If ψ : N → R is another smooth map between manifolds then (ψ◦ φ)
∗m
= ψ
∗φ(m)
◦
φ
∗m
. This formula is called the chain rule.
(3) If φ : M → M is the identity then φ
∗m
: TM
m
→ TM
m
is the identity. If
φ : M → N is a diﬀeomorphism and m ∈ M then φ
∗m
is 11 and onto.
(4) TM
m
is a vector space of dimension n, the dimension of M, and the induced maps
are linear.
The induced map φ
∗m
is deﬁned by
φ
∗m
([α]) = [φ ◦ α].
Notice that if M = R
m
, N = R
n
, then we have a natural way to identify the tangent
space and the map φ
∗
. We have coordinates on the tangent space so that
[φ ◦ α] =
dφ ◦ α(t)
dt
t=0
and
φ
∗m
([α]) = Dφ(m)(α
(0)).
The induced map φ
∗m
is also commonly denoted Tφ or dφ. These results follow for
neighborhoods in manifolds since these are manifolds too. Also note that if there is a
neighborhood U of m ∈ M and φ
U
is a diﬀeomorphism onto a neighborhood of φ(m) then
φ
∗m
is an isomorphism.
Proof.
(1) If φ : M → N is a smooth map and m ∈ M then there is an induced map
φ
∗m
: TM
m
→ TN
φ(m)
deﬁned by φ
∗m
([α]) = [φ◦ α]. We need to show this map is
CHAPTER 5 TANGENT VECTORS 3
welldeﬁned. Take charts (U, θ) on N centered on φ(m) and (W, ψ) on M centered
on m. If [α] = [β], then
dψ ◦ α(t)
dt
t=0
=
dψ ◦ β(t)
dt
t=0
(θ ◦ φ ◦ ψ
−1
)
∗
(
dψ ◦ α(t)
dt
t=0
) = (θ ◦ φ ◦ ψ
−1
)
∗
(
dψ ◦ β(t)
dt
t=0
)
d
dt
(θ ◦ φ ◦ ψ
−1
◦ ψ ◦ α)(t)
t=0
=
d
dt
(θ ◦ φ ◦ ψ
−1
◦ ψ ◦ β)(t)
t=0
d
dt
(θ ◦ φ ◦ α)(t)
t=0
=
d
dt
(θ ◦ φ ◦ β)(t)
t=0
so φ
∗m
is well deﬁned on equivalence classes.
(2) If φ : M → N and ψ : N → R are smooth maps, then (ψ◦φ)
∗m
([α]) = [ψ◦φ◦α] =
ψ
∗φ(m)
([φ ◦ α]) = ψ
∗φ(m)
◦ φ
∗m
([α]).
(3) I
M∗m
([α]) = [I
M
◦ α] = [α]. If φ ◦ φ
−1
= I
M
then φ
∗
◦ (φ
−1
)
∗
= I
M∗
= I
TM
m
.
Also, if φ
−1
◦ φ = I
M
, then (φ
−1
)
∗
◦ φ
∗
= I
TM
m
. Therefore φ
∗
is a bijection and
(φ
∗
)
−1
= (φ
−1
)
∗
.
(4) Let (U, φ) be a coordinate neighborhood centered at m. We ﬁrst show that TR
n
0
is an ndimensional vector space. Since R
n
requires no coordinate neighborhood
(i.e., it is itself), [α] is equivalent to [β] if and only if α
(0) = β
(0): two paths
are equivalent if they have the same derivative vector in R
n
. Every vector v is
realized by a path α
v
, α
v
(t) = tv. This identiﬁcation gives TR
n
0
the vector space
structure. We show that the linear structure is well deﬁned on TM
m
. The linear
structure on TM
m
is induced by the structure on TR
n
0
(where [α] +k[β] = [α+kβ]
and induced maps are linear) via the coordinate maps. If (V, ψ) is another chart
centered at m, then the structure deﬁned by ψ and φ agree since (φ ◦ ψ
−1
)
∗
is an
isomorphism and (φ ◦ ψ
−1
)
∗
◦ ψ
∗
= φ
∗
.
We can give explicit representatives for linear combinations of paths in the tangent space
TM
m
. In the notation of the proof of Theorem 5.3*** part 4,
k[α] + c[β] = [φ
−1
(kφ ◦ α + cφ ◦ β)]
Note that the coordinate chart serves to move the paths into R
n
where addition and
multiplication makes sense.
Before we turn to the second interpretation of a tangent vector as a directional derivative,
we pause for a philosophical comment. We ﬁrst learn of functions in our grade school
education . We learn to speak of the function as a whole or its value at particular points.
Nevertheless, the derivative at a point does not depend on the whole function nor is it
determined by the value at a single point. The derivative requires some open set about a
point but any open set will do. If M is a manifold and m ∈ M, then let G
m
be the set of
functions deﬁned on some open neighborhood of m.
4 CHAPTER 5 TANGENT VECTORS
Deﬁnition 5.6***. A function : G
m
→ R is called a derivation if for every f, g ∈ G
m
and a, b ∈ R,
(1) (af + bg) = a(f) +b(g) and
(2) (fg) = (f)g(m) + f(m)(g)
Denote the space of derivations by D. The product rule which occurs in the deﬁnition
is called the Leibniz rule, just as it is in Calculus.
Proposition 5.7***. Elements of TM
m
operate as derivations on G
m
. In fact there is a
linear map : TM
m
→ D given by v →
v
.
The theorem is straightforward if the manifold is R
n
. If v ∈ TR
n
x
, then the derivation
v
is the directional derivative in the direction v, i.e.,
v
(f) = Df(x)(v). On a manifold
the argument is really the same, but more technical as the directions are more diﬃcult to
represent. We will see in Theorem 5.8*** that the derivations are exactly the directional
derivatives.
Proof. If α : ((−, ), {0}) → (M, {m}) represents v then deﬁne
v
(f) =
df ◦ α(t)
dt
t=0
.
The fact that
v
is a linear functional and the Leibniz rule follow from these properties of
the derivative.
To show that is a linear map requires calculation. Suppose (U, φ) is a coordinate chart
centered at m. If [α] and [β] are equivalence classes that represent tangent vectors in TM
m
and c, k ∈ R, then φ
−1
((kφα(t) + cφβ(t))) represents k[α] + c[β]. Hence,
k[α]+c[β]
(f) =
df(φ
−1
((kφα(t) +cφβ(t))))
dt
t=0
= f
∗
φ
−1
∗
d(kφα(t) +cφβ(t))
dt
t=0
= f
∗
φ
−1
∗
k
d(φα(t)
dt
t=0
+c
d(φβ(t))
dt
t=0
= kf
∗
φ
−1
∗
dφα(t)
dt
t=0
+cf
∗
φ
−1
∗
dφβ(t)
dt
t=0
= k
df(φ
−1
(φ(α(t))))
dt
t=0
+c
df(φ
−1
(φ(β(t))))
dt
t=0
= k
df((α(t)))
dt
t=0
+c
df((β(t)))
dt
t=0
= k
[α]
(f) + c
[β]
(f)
Lines 3 and 4 respectively follow from the linearity of the derivative and the total derivative
map. Therefore is linear.
The second interpretation of tangent vectors is given in the following Theorem.
CHAPTER 5 TANGENT VECTORS 5
Theorem 5.8***. The linear map : TM
m
→ D given by v →
v
is an isomorphism.
The elements of TM
m
are the derivations on G
m
.
Proof. We ﬁrst note two properties on derivations.
(1) If f(m) = g(m) = 0, then (fg) = 0
Since (fg) = f(m)(g) + g(m)(f) = 0 + 0.
(2) If k is a constant, then (k) = 0
Since (k) = k(1) = k((1) + (1)) = 2k(1), (k) = 2(k) and (k) = 0.
We now observe that is onetoone. Let (U, φ) be a coordinate chart centered at m.
Suppose v = 0 is a tangent vector. We will show that
v
= 0. Let φ
∗
(v) = w
1
∈ R
n
. Note
that w
1
= 0. Then [φ
−1
(tw
1
)] = v where t is the real variable. Let w
1
, · · · , w
n
be a basis
for R
n
and π(
n
i=1
a
i
w
i
) = a
1
. Then
v
(π ◦ φ) =
[φ
−1
(tw
1
)]
(π ◦ φ)
=
dπ(φ(φ
−1
(tw
1
)))
dt
t=0
=
dtw
1
dt
t=0
= w
1
.
Next we argue that is onto. Let (U, φ) be a coordinate chart centered at m and let
e
i
for i = 1, · · · , n be the standard basis for R
n
. We consider the path t → φ
−1
(te
i
) and
compute some useful values of , i.e., the partial derivatives.
[φ
−1
(te
i
)]
(f) =
dfφ
−1
(te
i
)
dt
t=0
=
∂fφ
−1
∂x
i
0
Let x
i
(a
1
, · · · , a
n
) = a
i
. Suppose d is any derivation. We will need to name certain values.
Let d(x
i
◦ φ) = a
i
. These are just ﬁxed numbers. Suppose f is C
∞
on a neighborhood of
m. Taylor’s Theorem says that for p in a neighborhood of
0 ∈ R
n
,
f ◦ φ
−1
(p) = f ◦ φ
−1
(
0) +
n
i=1
∂f ◦ φ
−1
∂x
i
0
x
i
(p) +
n
i,j=1
R
ij
(p)x
i
(p)x
j
(p)
where R
ij
(p) =
1
0
(t −1)
∂
2
f◦φ
−1
∂x
i
∂x
j
tp
dt are C
∞
functions. So,
f = f(m) +
n
i=1
∂f ◦ φ
−1
∂x
i
0
x
i
◦ φ +
n
i,j=1
(R
ij
◦ φ) · (x
i
◦ φ) · (x
j
◦ φ).
We now apply d. By (2), d(f(m)) = 0. Since x
j
◦φ(m) = 0, the terms d((R
ij
◦φ)· (x
i
◦φ)·
(x
j
◦φ)) = 0 by (1). Also, d(
∂f◦φ
−1
∂x
i
0
x
i
◦φ) = a
i
[φ
−1
(te
i
)]
(f). Hence, d =
P
n
i=1
a
i
[φ
−1
(te
i
)]
,
and is onto.
6 CHAPTER 5 TANGENT VECTORS
Remark 5.9***. Tangent vectors to points in R
n
.
The usual coordinates on R
n
give rise to standard coordinates on T
p
R
n
. Let e
i
=
(0, · · · , 0, 1, 0, · · · , 0) with the only nonzero entry in the ith spot. The path in R
n
deﬁned
by α
i
(t) = te
i
+ p is a path with α
i
(0) = p. Its equivalence class [α
i
] is a vector in T
p
R
n
and we denote it
∂
∂x
i
p
. In Advanced Calculus, the ordered basis
∂
∂x
1
p
, · · · ,
∂
∂x
n
p
is
the usual basis in which the Jacobian matrix is usually written and sets up a natural
isomorphism T
p
R
n
∼
= R
n
. The reader should notice that the isomorphism is only natural
because R
n
has a natual basis and is not just an abstract ndimensional vector space.
If ρ is a path in R
n
, then ρ
(0) ∈ T
ρ(0)
R
n
via this isomorphism. This notation is also
consistant with the operator notation (the second interpretation) since,
∂
∂x
i
p
(f) = [f ◦ α
i
]
=
d
dt
f(te
i
+ p)
t=0
=
∂f
∂x
i
x=p
∈ R
n
∼
= T
p
R
n
In the ﬁrst line, the tangent vector
∂
∂x
i
p
operates via the second interpretation on the
function f.
Example 5.10***. TM
x
for M an ndimensional submanifold of R
k
.
Suppose M ⊂ R
k
is a submanifold and i : M → R
k
is the inclusion. Take (U
x
, φ) a
slice coordinate neighborhood system for R
k
centered at x as speciﬁed in the deﬁnition
of a submanifold, Deﬁnition 3.2***, φ : U
x
→ U
1
× U
2
. Under the natural coordinates of
TR
k
x
∼
= R
k
, TM
x
= φ(U
1
×{0}) ⊂ R
k
and i
∗x
has rank n.
To see these facts, note that φ ◦ i ◦ (φ
U
x
∩M
)
−1
: U
1
× {0} → U
1
× U
2
is the inclusion.
So, rank(i
∗
) = rank((φ◦i ◦φ
U
x
∩M
)
∗
) = n. Under the identiﬁcation TR
k
x
∼
= R
k
, φ
∗x
(R
n
×
{0}) = Dφ(x)(R
n
×{0}) ⊂ R
k
. This is the usual picture of the tangent space as a subspace
of R
k
(i.e., shifted to the origin) that is taught in advanced Calculus.
Example 5.11***. TS
n
x
for S
n
⊂ R
n+1
, the nsphere.
This is a special case of Example 5.10***. Suppose (x
1
, · · · , x
n+1
) ∈ S
n
, i.e.,
n+1
i=1
x
2
i
=
1. One of the x
i
must be nonzero, we assume that x
n+1
> 0. The other cases are
analogous. The inclusion from the Implicit Function Theorem is φ
R
n(x
1
, · · · , x
n
) =
(x
1
, · · · , x
n
,
1 −
n
i=1
x
2
i
) so
Dφ
R
n(x
1
, · · · , x
n
)(v
1
, · · · , v
n
) = (v
1
, · · · , v
n
,
n
i=1
−x
i
v
i
1 −
n
i=1
x
2
i
)
).
CHAPTER 5 TANGENT VECTORS 7
Since x
n+1
> 0, x
n+1
=
1 −
n
i=1
x
2
i
) and the tangent space is
T
(x
1
,··· ,x
n+1
)
S
n
= {(v
1
, · · · , v
n
,
n
i=1
−x
i
v
i
x
n+1
)  v
i
∈ R}
= {(w
1
, · · · , w
n+1
) 
n+1
i=1
w
i
x
i
= 0}
Example 5.12***. Recall that O(n) ⊂ Mat
n×n
= R
n
2
is a submanifold of dimension
n(n−1)
2
which was shown in Example 3.7***. Then, we claim,
X ∈ T
A
O(n) ⊂ Mat
n×n
if and only if XA
−1
is skew.
This computation is a continuation of Example 3.7***. Suppose A ∈ O(n). Since
O(n) = f
−1
(I), T
A
O(n) ⊂ Ker(Df(A)). The dimension of the kernel and the dimension
of T
A
O(n) are both
n(n−1)
2
. Therefore T
A
O(n) = Ker(Df(A)). It is enough to show that
Ker(Df(A)) ⊂ {X  XA
−1
is skew} since the dimension of {X  XA
−1
is skew} is the
dimension of Skew
n×n
=
n(n−1)
2
(from Example 2.8d***). So it is enough to show that
XA
−1
is skew.
Again, from Example 3.7***, Df(A)(X) = AX
T
+ XA
T
. If Df(A)(X) = 0, then
AX
T
= −XA
T
. Since A ∈ O(n), A
−1
= A
T
. So,
(XA
−1
)
T
= (XA
T
)
T
= AX
T
= −XA
T
= −XA
−1
Therefore XA
−1
is skew.
Example 5.13***. Recall that Sp(n, R) ⊂ Mat
n×n
= R
n
2
is a submanifold of dimension
n(n+1)
2
which was shown in Example 3.9***. Then, we claim,
X ∈ T
A
Sp(n, R) ⊂ Mat
n×n
if and only if JXA
−1
is symmetric.
This computation is a continuation of Example 3.9***. Suppose A ∈ Sp(n, R).
Since Sp(n, R) = f
−1
(J), T
A
Sp(n, R) ⊂ Ker(Df(A)). The dimension of the kernel and
the dimension of T
A
Sp(n, R) are both
n(n+1)
2
. Therefore T
A
Sp(n, R) = Ker(Df(A)). It
is enough to show that Ker(Df(A)) ⊂ {X  JXA
−1
is symmetric} since the dimension
of {X  JXA
−1
is symmetric } is the dimension of Sym
n×n
=
n(n+1)
2
(from Example
2.8c***). So it is enough to show that JXA
−1
is symmetric.
Again, from Example 3.9***, Df(A)(X) = AJX
T
+ XJA
T
. If Df(A)(X) = 0, then
−AJX
T
= XJA
T
. Since A ∈ Sp(n, R), A
−1
= JA
T
J
T
. So,
(JXA
−1
)
T
= (JXJA
T
J
T
)
T
= (−JAJX
T
J
T
)
T
= −JXJ
T
A
T
J
T
= JXJA
T
J
T
as J
T
= −J by Lemma 3.8***
= JXA
−1
Therefore XA
−1
is symmetric.
8 CHAPTER 5 TANGENT VECTORS
Remark 5.14***. Notation for Tangent vectors
The space R
n
comes equipped with a canonical basis e
1
, · · · , e
n
which allows us to pick
a canonical basis for TR
n
x
. For an nmanifold M, TM
p
doesn’t have a natural basis. We
can give coordinates on TM
p
in terms of a chart. Suppose that (U, φ) is a chart for a
neighborhood of p ∈ U ⊂ M. Write φ = (φ
1
, · · · , φ
n
) in terms of the coordinates on R
n
.
Hence, φ
i
= x
i
◦ φ. We can import the coordinates TR
n
φ(p)
. Let
∂
∂φ
i
p
= φ
−1
∗
(
∂
∂x
i
φ(p)
)
As a path
∂
∂φ
i
p
is the equivalence class of φ
−1
(te
i
+φ(p)). As an operator,
∂
∂φ
i
p
(f) =
∂f ◦ φ
−1
∂x
i
φ(p)
.
CHAPTER 5 TANGENT VECTORS 9
Exercises
Exercise 1***. Suppose F : R
4
→ R
2
by
F((w, x, y, z)) = (wxyz, x
2
y
2
).
Compute F
∗
and be explicit in exhibiting the bases in the notation used in Remark 5.9***.
Where is F singular?
The reader may wish to review Example 2.10** and Exercise 4*** from chapter 3 for
the following exercise.
Exercise 2***. Let g((x, y)) = x
2
+ y
2
and h((x, y)) = x
3
+ y
2
. Denote by G
g
and G
h
the graphs of g and h which are submanifolds of R
3
. Let F : G
g
→ G
h
by
F : ((x, y, z)) = (x
3
, xyz, x
9
+ x
2
y).
The reader may wish to review Example 2.10** and Exercise *** from chapter 3.
a. Explicitly compute the derivative F
∗
and be clear with your notation and bases.
b. Find the points of G
g
where F is singular. What is the rank of F
∗p
for the various
singular points p ∈ G
g
.
Exercise 3***. Let F : R
3
→ S
3
be deﬁned by
F((θ, φ, η)) = (sin η sin φcos θ, sinη sin φsin θ, sinη cos φ, cos η).
Use the charts from stereographic projection to compute F
∗
in terms of the bases discussed
in Remark 5.9*** and Remark 5.14***.
CHAPTER 6
IMMERSIONS AND EMBEDDINGS
In this chapter we turn to inclusion maps of one manifold to another. If f : N → M is
an inclusion, then the image should also be a manifold. In chapter 3, we saw one situation
where a subset of f(N) ⊂ M inherited the structure of a manifold: when each point of
f(N) had a slice coordinate neighborhood of M. In this chapter, we show that is the only
way it can happen if f(N) is to inherit its structure from M.
We ﬁrst review the situation for functions f : R
n
→ R
m
for n ≤ m. The inﬁnitesimal
condition of a function to be onetoone is that the derivative is onetoone. That the
derivative is onetoone is not required for the function to be onetoone, but it is suﬃcient
to guarantee the function is onetoone in some neighborhood (by the Inverse Function
Theorem). On the other hand, if f(y
0
) = f(z
0
), then there is a point x
0
on the segment
between y
0
and z
0
where Df(x
0
) is not onetoone. This last statement is a consequence of
Rolle’s Theorem. This discussion, perhaps, serves as some motivation to study functions
whose derivative is injective. A second justiﬁcation is that if f is to be a diﬀeomorphism
to its image, then the derivative must be invertible as a linear map.
While the phrase “f(N) inherits manifold structure form M” is vague, it certainly
includes that “f(N) inherits its topology from M” which is precise.
Deﬁnition 6.1***. Suppose f : N → M is a smooth map between manifolds. The map
f is called an immersion if f
∗x
: T
x
N → T
f(x)
M is injective for all x ∈ N.
The derivative is injective at each point is not enough to guarantee that the func
tion is onetoone, as very simple example illustrate. Take f : R → R
2
by f(x) =
(sin(2πx), cos(2πx)). This function is inﬁnitetoone as f(x + 1) = f(x), but Df(x) is
injective for all x ∈ R. Hence it is clear that we will need some other condition to obtain
an inclusion. An obvious ﬁrst guess, that turns out to be inadequate, is that f is also
onetoone.
Example 6.2***. A onetoone immersion f : N → M in which f(N) is not a topological
manifold.
Let N = (
−π
4
,
3π
4
), M = R
2
, and f(x) = (cos(x) cos(2x), sin(x) cos(2x)). The image
f(
−π
4
,
3π
4
) is two petals of a four leafed rose. The map is onetoone: only x =
π
4
maps
to (0, 0). Note that if
−π
4
or
3π
4
were in the domain, then they would also map to (0, 0).
Df(x) is rank one, so f is a onetoone immersion. However, f(N) is not a topological
manifold. Suppose U ⊂ B
1/2
((0, 0)), then U ∩ f(N) cannot be homeomorphic to an open
interval. An interval with one point removed has two components, by U ∩f(N) \(0, 0) has
copyright c 2002
Typeset by A
M
ST
E
X
1
2 CHAPTER 6 IMMERSIONS AND EMBEDDINGS
at least four components. Hence no neighborhood of (0, 0) ∈ f(N) is homeomorphic to an
open set in R.
Deﬁnition 6.3***. Suppose f : N → M is a smooth map between manifolds. The map
f is called an embedding if f is an immersion which is a homeomorphism to its image.
This extra topological condition is enough to guarantee that f(N) is a submanifold in
the strong sense of Deﬁnition 3.2***.
Theorem 6.4***. Suppose N
n
and M
m
are manifolds and f : N → M is a smooth map
of rank n. If f is a homeomorphism to its image, then f(N) is a submanifold of M and f
is a diﬀeomorphism to its image.
Proof. To show that f(N) is a submanifold of M, we suppose x
0
∈ N and we must produce
a slice neighborhood of f(x
0
) ∈ f(N) ⊂ M. We produce this neighborhood in three steps.
The ﬁrst step is to clean up the local picture by producing coordinate neighborhoods of x
0
and f(x
0
) that properly align. The second step is to produce a coordinate neighborhood
of f(x
0
) in M in which f(N) looks like the graph of a function. The graph of a function
was already seen to be a submanifold, and we have virtually completed the construction.
The third step is to construct the slice neighborhood.
As a ﬁrst step, we produce coordinate neighborhoods:
(1) (O
2
, ψ) a coordinate neighborhood in N centered at x
0
(2) (U
2
, τ) a coordinate neighborhood in M centered at f(x
0
)
with f
−1
(U
2
∩ N) = O
2
and (τ ◦ f)
∗x
0
(TN
x
0
) = R
n
× {0}
Take O
1
⊂ N a coordinate neighborhood in N centered at x
0
, and U
1
⊂ M such that
f
−1
(U
1
∩ N) = O
1
. Such a U
1
exists since f is a homeomorphism to its image, and f(N)
has the subspace topology. Take (U
2
, φ) a coordinate neighborhood of M centered at f(x
0
)
with U
2
⊂ U
1
. Let O
2
⊂ f
−1
(U
2
), x
0
∈ O
2
. Then (O
2
, ψ) is a coordinate neighborhood
of N centered at x
0
. Let v
1
, · · · , v
n
span (φ ◦ f)
∗x
0
(TN
x
0
) ⊂ R
m
, and let v
1
, · · · , v
m
be a
basis of R
m
. Let H : R
m
→ R
m
be the isomorphism H(
m
i=1
a
i
v
i
) = (a
1
, · · · , a
m
). Then
(U
2
, H◦φ) is a coordinate neighborhood in M centered at f(x
0
) and (H◦φ◦f)
∗x
0
(TN
x
0
) =
R
n
× {0}. Let τ = H ◦ φ and the coordinate neighborhoods are constructed.
The second step is to cut down the neighborhood of f(x
0
) so that f(N) looks like
the graph of a function. This step requires the inverse function theorem. We produce
coordinate neighborhoods:
(1) (O
3
, ψ) a coordinate neighborhood in N centered at x
0
(2) (U
3
, τ) a coordinate neighborhood in M centered at f(x
0
), τ : U
3
→ W
3
× W
2
⊂
R
n
×R
m−n
(3) a C
∞
function g : W
3
→ W
2
such that τ(f(N) ∩ U
4
) is the graph of g.
Let W
2
⊂ R
n
and W
4
⊂ R
m−n
be open sets such that W
4
× W
2
is a neighborhood of
0 ∈ τ(U
2
) ∈ R
m
. Now deﬁne U
4
= τ
−1
(W
4
× W
2
) and O
4
= f
−1
(U
4
). Then O
4
⊂ O
2
,
(O
4
, ψ) is a chart centered at x
0
, and U
4
⊂ U
2
, (U
4
, τ) is a chart centered at f(x
0
). Let
p
1
: R
m
→ R
n
be the projection onto the ﬁrst n coordinates and p
2
: R
m
→ R
m−n
be the
projection onto the last m − n coordinates. The function p
1
◦ τ ◦ f ◦ ψ
−1
maps the open
set ψ(O
4
) to W
4
. Since (τ ◦ f)
∗x
0
(TN
x
0
) = R
n
× {0}, D(p
1
◦ τ ◦ f ◦ ψ
−1
)(0) has rank n,
CHAPTER 6 IMMERSIONS AND EMBEDDINGS 3
i.e., it is an isomorphism. By the Inverse Function Theorem, there is a neighborhood V of
0 ∈ R
n
, V ⊂ ψ(O
4
) and a neighborhood W
3
of 0 ∈ R
n
, W
3
⊂ W
4
such that
p
1
◦ τ ◦ f ◦ ψ
−1
: V → W
3
is a diﬀeomorphism. Let O
3
= ψ
−1
(V ) and U
3
= τ
−1
(W
3
× W
2
). Then (U
3
, τ) is a
coordinate chart centered at f(x
0
), τ : U
3
→ W
3
× W
2
, and (O
3
, ψ) is a coordinate chart
centered at x
0
. Let g be the composition
W
3
(p
1
◦τ◦f◦ψ
−1
)
−1
−−−−−−−−−−−→ ψ
−1
(O
3
)
(p
2
◦τ◦f◦ψ
−1
)
−−−−−−−−−→ W
2
The function g is the composition of two C
∞
functions. We now observe that the graph
of g is τ(f(N) ∩ U
3
). The points in τ(f(N) ∩ U
3
) are τ ◦ f ◦ ψ
−1
(ψ(O
3
)). If x ∈ ψ(O
3
),
then its coordinates in W
3
×W
2
is (p
1
◦ τ ◦ f ◦ ψ
−1
(x), p
2
◦ τ ◦ f ◦ ψ
−1
)(x)) which agrees
with the graph of g. The second step is established.
The third step is to produce the slice neighborhood. Take W
1
an open set with compact
closure and
¯
W
1
⊂ W
3
. Let be such that 0 < < max{g(x) −y  x ∈
¯
W
1
, y ∈ R
n
\ W
2
}.
Let V
1
⊂ W
1
× W
2
be the open set {(x, y) ∈ W
1
× W
2
 g(x) − y < }. Let γ : V
1
→
W
1
× B
(0) by γ(x, y) = (x, y − g(x)). The map γ is a diﬀeomorphims with inverse
(x, y) → (x, y +g(x)). The image of the graph of g under γ is W
1
×{0}. Let U = τ
−1
(V
1
),
then (U, γ ◦ τ) is the slice neighborhood: y = g(x) if and only if γ(x, y) = (x, 0).
It remains to show that f is a diﬀeomorphism. Since f is a homeomorphism to its
image, f has a continuous inverse. We need to see that f is smooth as is its inverse. We
use Proposition 2.18***. Given x ∈ N, there is a chart of f(N) about x that arises from
a slice chart about x in M, Proposition 3.3***. Let (U, φ), φ : U → W
1
× W
2
be the
slice chart and (U ∩ f(N), p
1
◦ φ) the chart for f(N). The map f is a diﬀeomorphism if
p
1
◦φ◦f ◦ψ
−1
and its inverse are C
∞
in a neighborhood of ψ(x) and φ(f(x)), respectively.
Now, since (U, φ) is a slice neighborhood,
p
1
◦ φ ◦ f ◦ ψ
−1
= φ ◦ f ◦ ψ
−1
.
The derivative D(φ◦ f ◦ ψ
−1
)(x) has rank n since φ and ψ are diﬀeomorphisms, and f has
rank n. By the Inverse Function Theorem, p
1
◦ φ ◦ f ◦ ψ
−1
is C
∞
on a neighborhood of
φ(f(x)). By Proposition 2.18***, f and f
−1
are smooth functions.
Some authors use the terminology that the image of a manifold under an immersion is
a submanifold, but this usage is less common. Furthermore it requires the immersion in
the deﬁnition. We will use the term immersed submanifold.
Deﬁnition. Suppose N and M are manifolds and f : N → M is an immersion. Then
(N, f) is an immersed submanifold.
This terminology is suggested by Exercise 1.***
Proposition 6.5***. Suppose N and M are manifolds and f : N → M is a onetoone
immersion. If N is compact, then f is an embedding.
Proof. We just need to show that f is a homeomorphism to its image. It is a onetoone
continuous map from a compact space to a Hausdorﬀ space. By a standard result in general
topology, f is a homeomorphism.
4 CHAPTER 6 IMMERSIONS AND EMBEDDINGS
Exercises
Exercise 1***. Suppose that f : N → M is a onetoone immersion. Show that for every
x ∈ N there is a neighborhood U of x such that f 
U
: U → M is an embedding. Show
that the result holds even if f is not onetoone.
The next exercise is a diﬃcult and interesting exercise.
Exercise 2***. Every compact nmanifold embeds in R
N
for some N.
This result is true without the hypothesis of compactness.
The dimension N can be taken to be 2n. That every nmanifold embedds in R
2n
is a
result by H. Whitney. It is also interesting to note that every compact nmanifold immerses
in R
2n−α(n)
where α(n) is the number of ones in the dyadic expansion of n. This result
was proven by Ralph Cohen. The connection to the dyadic expansion and that this result
is the best possible arose in work by William S. Massey.
Exercise 3***. Let f : RP
2
→ R
3
by f([x, y, z]) = (xy, xz, yz). Show that f is a
welldeﬁned smooth function. Is f onetoone? Is f an immersion?
Let g : RP
2
→ R
4
by g([x, y, z]) = (xy, xz, yz, x
4
). Is g an embedding or an immersion?
CHAPTER 7
VECTOR BUNDLES
We next begin addressing the question: how do we assemble the tangent spaces at various
points of a manifold into a coherent whole? In order to guide the decision, consider the
case of U ⊂ R
n
an open subset. We reﬂect on two aspects.
The ﬁrst aspect is that the total derivative of a C
∞
function should change in a C
∞
manner from point to point. Consider the C
∞
map f : U → R
m
. Each point x ∈ U gives
a linear map f
∗x
= Df(x) : R
n
→ R
m
the total derivative which is represented by an
m × n matrix, the Jacobian matrix of Df(x). The Jacobian matrix is a matrix of C
∞
functions in x. While for each x ∈ U there is a linear map
Df(x) : TU
x
= R
n
→ R
m
= TR
m
f(x)
these ﬁt together to give a C
∞
map on the product
U ×R
n
→ R
m
×R
m
(x, v) → (f(x), Df(x)(v)).
The second aspect is that we wish to deﬁne vector ﬁelds. A vector ﬁeld is a choice of
tangent vector at each point. For an open subset U of R
n
, a vector ﬁeld is just given by a
function g : U → R
n
(as the reader probably learned in Advanced Calculus). In order to
keep track of the tail, we write the vector ﬁeld as
V : U → U ×R
n
x → (x, g(x)).
Any C
∞
function g gives a vector ﬁeld. The complication on a manifold M is that the
vector with tail at x ∈ M must be in the vector space TM
x
and these vector spaces change
with x. In this chapter, we study the required concepts to assemble the tangent spaces of
a manifold into a coherent whole and construct the tangent bundle. The tangent bundle
is an example of an object called a vector bundle.
Deﬁnition 7.1***. Suppose M
n
is a manifold. A real vector bundle over M consists of
a topological space E, a continuous map π : E → M and a real vector space V (called
the ﬁber) such that for each m ∈ M, π
−1
(m) is a vector space isomorphic to V , and there
exists an open neighborhood U of m, and a homeomorphism
µ
U
: π
−1
(U) → U ×V
copyright c 2002
Typeset by A
M
ST
E
X
1
2 CHAPTER 7 VECTOR BUNDLES
such that µ
−1
U
(m, −) : {m} ×V → π
−1
(m) is a linear isomorphism.
The bundle is smooth if E is a smooth manifold, π is smooth, and µ
U
is a diﬀeomorphism.
In these notes, all vector bundles will be smooth. We may denote a vector bundle by
π : E → M (and suppress the vector space) or as E. If the dimension of the vector space
is m then the bundle is often called an mplane bundle. A 1plane bundle is also called
a line bundle. A bundle over a manifold is trivial if it is simply the Cartesian product of
the manifold and a vector space. The neighborhoods U over which the vector bundle looks
like a product are called trivializing neighborhoods.
Note that µ
W
◦ µ
−1
U
: {m} × V → {m} × V is a linear isomorphism. Denote this map
h
WU
(m).
Deﬁnition 7.2***. If µ
U
: π
−1
(U) → U × V and µ
W
: π
−1
(W) → W × V are trivial
neighborhoods of a vector bundle then
µ
W
◦ µ
−1
U
: (W ∩ U) ×V → (W ∩ U) ×V
(x, v) → (x, h
WU
(x)v)
where h
WU
: W ∩ U → GL(V ). The h
WU
are associated to each pair of trivial neighbor
hoods (U, µ
U
) and (W, µ
W
). They are called transition functions.
Theorem 7.3***. Every smooth vector bundle has smooth transition functions, i.e.,
h
WU
: W ∩ U → GL(V ) is smooth.
Proof. The map µ
W
◦ µ
−1
U
deﬁnes h
WU
so the issue is to see that h
WU
is smooth. Let
h
WU
(x) be the matrix (h
ij
(x))
ij
in a ﬁxed basis for V . Then, µ
W
◦ µ
−1
U
(x, (r
1
, · · · , r
n
)) =
(x, (
j
h
1j
(x)r
j
, · · · ,
j
h
nj
(x)r
j
)). To see that each h
ij
(x) is smooth let r vary over e
i
for i = 1, · · · n. Since µ
W
◦ µ
−1
U
is smooth, so are its coordinate functions.
Example 7.4***. Line Bundles Over S
1
.
We take the circle to be S
1
= {e
θi
 θ ∈ R} the unit circle in the complex plane
{(cos θ, sinθ)  θ ∈ R
2
}.
One line bundle over the circle is
1
S
1
, the trivial bundle π
: S
1
×R → S
1
by π
((e
θi
, r)) =
e
θi
. For the trivialization neighborhoods, only one is needed: take U = S
1
.
There is another, more interesting line bundle over S
1
. Let E = {(e
θi
, re
θ
2
i
)  r, θ ∈ R}
and π
γ
: E → S
1
by π
γ
((e
θi
, re
θ
2
i
)) = e
θi
. Denote this bundle γ
1
S
1
Notice that π
−1
γ
(e
θi
)
is a real line in the complex plane. Two values of θ that diﬀer by 2π determine the same
point, so
θ
2
is not welldeﬁned. Nevertheless, the line in the complex plane through e
θ
2
i
is
well deﬁned since e
2π
2
i
= −1.
We now construct the trivializing neighborhoods. Let U = S
1
\ {1} = {(e
θi
, re
θ
2
i
)  θ ∈
(0, 2π)} and W = S
1
\ {−1} = {(e
θi
, re
θ
2
i
)  θ ∈ (π, 3π)}. Now,
(1)
µ
U
: π
−1
γ
(U) → U ×R
(e
θi
, re
θ
2
i
) → (e
θi
, r).
CHAPTER 7 VECTOR BUNDLES 3
This map is well deﬁned since θ ∈ (0, 2π), a restricted domain which allows us to determine
θ from e
θ
2
i
. We similarly deﬁne
(2)
µ
W
: π
−1
γ
(U) → U ×R
(e
θi
, re
θ
2
i
) → (e
θi
, r)
for θ ∈ (π, 3π).
We next check the compatibility condition. The set U ∩W is S
1
\{1, −1} = {(e
θi
, re
θ
2
i
) 
θ ∈ (0, π) ∪ (π, 2π)}. Suppose θ ∈ (0, π) then
µ
W
µ
−1
U
((e
θi
, r)) = µ
W
((e
θi
, re
θ
2
i
))
= µ
W
((e
(θ+2π)i
, −re
θ+2π
2
i
))
= (e
(θ+2π)i
, −r)
= (e
θi
, −r)
Notice that we had to change the expression for the second coordinate because formulas
(1) and (2) require diﬀerent domains. We have that h
UW
(e
θi
)(r) = −r for θ ∈ (0, π). Now,
suppose θ ∈ (π, 2π), then
µ
W
µ
−1
U
((e
θi
, r)) = µ
W
((e
θi
, re
θ
2
i
))
= (e
θi
, r)
We have that h
UW
(e
θi
)(r) = r for θ ∈ (π, 2π). Therefore the transition function h
UW
:
U ∩ W → Gl(1, R) is
h
UW
(x) =
_
−1 if Im(x) > 0
1 if Im(x) < 0
.
Example 7.5***. The Tautological Line Bundle Over RP
n
Deﬁne a Z
2
action on S
n
× R by (−1) · (x, r) = (−x, −r). We show that this action
satisﬁes the hypotheses of Theorem 2.23***. Suppose (x, r) ∈ S
n
× R. Take U an open
neighborhood of x in S
n
that is entirely in one hemisphere. Then it follows that U∩−U = ∅,
and U × R and −U × R are disjoint neighborhoods of (x, r) and (−1) · (x, r). Let E =
S
n
×R/Z
2
. By 2.23***, E is a smooth manifold and the quotient map ˜ q : S
n
×R → E is
a local diﬀeomorphism. Let π
E
: E → RP
n
by π
E
(x, r) = [(x, r)]. The following diagram
is a commutative diagram of smooth maps,
S
n
×R
˜ q
−−−−→ E
π
¸
¸
_
¸
¸
_
π
E
S
n
q
−−−−→ RP
n
where π(x, r) = x and q is the quotient map from Example 2.25***, RP
n
. Let U be an
open set in S
n
that is entirely in one hemisphere so that U ∩ −U = ∅. Then ˜ q
U×R
:
4 CHAPTER 7 VECTOR BUNDLES
U × R → π
−1
E
(q(U)) is a diﬀeomorphism and linear on each ﬁber. If V is another such
open subset of S
n
then
(˜ q
V ×R
)
−1
◦ ˜ q
U×R
: U ∩ (V ∪ −V ) ×R → U ∩ (V ∪ −V ) ×R
(x, r) → (x, h(x)r)
where
h(x) =
_
1 if x ∈ U ∩ V
−1 if x ∈ U ∩ −V
Hence, π
E
: E → RP
n
is a vector bundle.
Proposition 7.6***. Transition functions satisfy the following property:
h
WU
(x) = h
WO
(x) ◦ h
OU
(x) for x ∈ W ∩ O ∩ U.
Proof. This property follows from the deﬁnition of transition functions and is the equation
of the last coordinate of the equation below.
(µ
W
◦ µ
−1
O
) ◦ (µ
O
◦ µ
−1
U
) = µ
W
◦ µ
−1
U
The next theorem shows that a choice of transition functions consistent with the prop
erties of the last proposition will determine the vector bundle. A bundle can be deﬁned
by the gluing (transition) functions. First a lemma.
Lemma 7.7***. Suppose X is a set and {U
i
 i ∈ I} is a collection of subsets. If
h
ij
: U
i
∩ U
j
→ GL(n, R) for all (i, j) ∈ I × I satisﬁes h
ij
(x) = h
ik
(x) ◦ h
kj
(x) for
x ∈ U
i
∩ U
j
∩ U
k
, then they also satisfy
(1) h
ii
(x) = I
V
for x ∈ U
i
(2) h
ij
(x) = (h
ji
(x))
−1
for x ∈ U
i
∩ U
j
Proof. Since h
ii
(x) = h
ii
(x) ◦ h
ii
(x) and h
ii
(x) has an inverse, multiply both sides by
h
ii
(x)
−1
to get (1). Since h
ij
(x)h
ji
(x) = h
ii
(x) = I
V
by the hypothesis and (1), h
ij
(x) =
h
ji
(x)
−1
.
Theorem 7.8***. Suppose M is a manifold and {(U
i
, ψ
i
)i ∈ I} is a countable atlas for
M. Suppose V = R
m
is a vector space and for all (i, j) ∈ I × I there is a C
∞
function
f
ij
: U
i
∩ U
j
→ GL(V ) such that f
kj
(x) ◦ f
ji
(x) = f
ki
(x) for all x ∈ U
i
∩ U
j
∩ U
k
.
On {(u, v, i)i ∈ I, u ∈ U
i
, v ∈ V } let (u, v, i) (u
, v
, i
) if and only if u = u
, f
i
i
(u)(v) =
v
.
Then is an equivalence relation. Furthermore, if E =
{(u, v, i)i ∈ I, u ∈ U
i
, v ∈ V }
and π([u, v, i]) = u, then π : E → M is a smooth vector bundle with ﬁber V .
Proof. We ﬁrst show that is an equivalence relation, The functions f
ij
also satisfy the
hypotheses of Lemma 7.7***. The relation is an equivalence relation as reﬂexivity,
symmetry and transitivity are guaranteed by conditions 1 and 2 of Lemma 7.7***, and
condition 3 is the hypothesis on the f
ij
’s.
We must verify the various requirements of the deﬁnition of a vector bundle for π : E →
M. The longest part is done ﬁrst, showing that E is a manifold.
CHAPTER 7 VECTOR BUNDLES 5
We ﬁrst introduce the map µ. Let µ
i
: π
−1
(U
i
) → U
i
× V by µ
i
([u, v, i]) = (u, v). This
map is a bijection: if u ∈ U
i
then the class [u, v, i] has (u, v, i) as the unique representative
with the last coordinate i by the deﬁnition of as f
ii
(u) = I
V
.
We verify that E is a smooth manifold by using Theorem 2.15*** Let φ
i
= (ψ
i
×I
V
) ◦µ
i
so φ
i
: π
−1
(U
i
) → ψ
i
(U
i
) ×V ⊂ R
n
×V = R
n+m
. We claim that E is a smooth manifold
with atlas A = {(π
−1
(U
i
), φ
i
)i ∈ I}. We verify the conditions of Theorem 2.15*** with
X = E and A. The ﬁrst two conditions are true, π
−1
(U
i
) ⊂ E and
i∈I
π
−1
(U
i
) = E
by construction. Next, the map φ
i
is a bijection since both µ
i
: π
−1
(U
i
) → U
i
× V and
ψ
i
× I
V
: U
i
× V → ψ
i
(U
i
) × V are bijections. The fourth condition is also immediate:
φ
i
: π
−1
(U
i
) → ψ
i
(U
i
)×V ⊂ R
n
×V and φ
i
: π
−1
(U
i
)∩π
−1
(U
j
) → ψ
i
(U
i
∩U
j
)×V ⊂ R
n
×V
are both open. The last condition is that φ
j
φ
−1
i
: φ
i
(π
−1
(U
i
) ∩ π
−1
(U
j
)) → φ
j
(π
−1
(U
i
) ∩
π
−1
(U
j
)) is smooth. We now show that the map is smooth. Suppose x ∈ ψ
i
(U
i
∩ U
j
) and
v ∈ V . Then,
φ
j
φ
−1
i
(x, v) = (ψ
j
×I
V
) ◦ µ
j
◦ µ
−1
i
◦ (ψ
−1
i
×I
V
)(x, v)
= (ψ
j
×I
V
) ◦ µ
j
◦ µ
−1
i
(ψ
−1
i
(x), v)
= (ψ
j
×I
V
) ◦ µ
j
([ψ
−1
i
(x), v, i])
= ψ
j
×I
V
◦ µ
j
([ψ
−1
i
(x), f
ji
(ψ
−1
i
(x))(v), j]) by the equivalence relation
= ψ
j
×I
V
(ψ
−1
i
(x), f
ji
(ψ
−1
i
(x))(v))
= (ψ
j
ψ
−1
i
(x), f
ji
(ψ
−1
i
(x))(v))
We have shown that φ
j
φ
−1
i
(x, v) = (ψ
j
ψ
−1
i
(x), f
ji
(ψ
−1
i
(x))(v)). The ﬁrst coordinate is
smooth since M is a manifold. The second coordinate map is smooth since it is a compo
sition of the following smooth maps:
ψ(U
i
∩ U
j
) ×R
m
f
1
−→ (U
i
∩ U
j
) ×R
m
f
2
−→ GL(n, R) ×R
m
f
3
−→ R
m
where f
1
(y, v) = (ψ
−1
(y), v) is smooth since ψ
−1
is smooth, f
2
(x, v) = (f
ji
(x), v) is
smooth by Theorem 7.3***, and f
3
is smooth since it is evaluation of a linear function at
a vector, i.e., matrix multiplication. The hypotheses of Theorem 2.15*** are veriﬁed. We
show in the next two paragraphs that E is second countable and Hausdorﬀ, and hence a
manifold.
Since I is countable, {π
−1
(U
i
)i ∈ I} is countable and so E is a second countable space.
Suppose [u, v, i], [u
, v
, j] ∈ E. If u = u
, then there are open sets O
1
and O
2
that
separate u and u
in M. Hence π
−1
(O
1
) and π
−1
(O
1
) separate [u, v, i] and [u
, v
, j] in E.
If u = u
, then [u, v, i], [u
, v
, j] ∈ π
−1
(U
i
) which is Hausdorﬀ. Therefore E is Hausdorﬀ
and E is a manifold.
We now show that π : E → M is a vector bundle. The vector space structure on
π
−1
(m) is deﬁned by the structure on V: a[m, v, i] + b[m, v
, i] = [m, av + bv
, i]. This is
well deﬁned since if [m, v, i] = [m, w, j] and [m, v
, i] = [m, w
, j] then f
ij
(m)(w) = v and
6 CHAPTER 7 VECTOR BUNDLES
f
ij
(m)(w
) = v
so
a[m, w, j] +b[m, w
, j] = a[m, f
ij
(m)(w), i] +b[m, f
ij
(m)(w
), i]
= [m, af
ij
(m)(w) +bf
ij
(m)(w
), i]
= [m, f
ij
(m)(aw +bw
), i]
= [m, av +bv
, i]
The map µ
i
: π
−1
(U
i
) → U
i
×V is a diﬀeomorphism. This was shown above as these maps
are part of the manifold structure for E (Theorem 2.21***).
µ
−1
i
(m, −) : V → π
−1
(m) is an isomorphism as µ
−1
i
(m, −)(v) = µ
−1
i
(m, v) = [m, v, i].
π is smooth since if m is in the coordinate neighborhood given by (U
i
, ψ
i
), then [m, v, i]
is in the coordinate neighborhood given by (π
−1
(U
i
), φ
i
) and ψ
i
◦ π ◦ φ
−1
i
(x, v) = x for
(x, v) ∈ φ
i
(π
−1
(U
i
)) ⊂ R
n
×V .
Deﬁnition 7.9***. A bundle map between two vector bundles E
1
π
1
−→ M
1
and E
2
π
2
−→ M
2
is a pair of smooth maps f : E
1
→ E
2
and g : M
1
→ M
2
such that π
2
◦ f = g ◦ π
1
and
f
x
: π
−1
1
(x) → π
−1
2
(g(x)) is linear.
Deﬁnition 7.10***. Suppose that E
1
π
1
−→ M and E
2
π
2
−→ M are bundles over a manifold
M. A bundle equivalence between these bundles is a a bundle map (f, I
M
) over the identity
is an isomorphism on each ﬁber.
Please note the result Exercise 2***: if E
1
an E
2
are vector bundles over the manifold M
and if E
1
is bundle equivalent to E
2
by a bundle equivalence f, then E
2
is bundle equivalent
to E
1
by f
−1
which is also a bundle equivalence.
Deﬁnition 7.11***. A section of a vector bundle, E
π
−→ M is a smooth map s : M → E
such that π ◦ s = I
M
.
Notation 7.12***. Suppose E is E
π
−→ M a vector bundle with ﬁber V . Let Γ(E) or
Γ(E) denote the sections of the vector bundle E.
Example 7.13***. The sections of the trivial bundle.
The sections of
n
M
are smooth maps s : M → M ×R
n
with π(s(x)) = x for all x ∈ M.
Hence, the ﬁrst coordinate of s(x) is x and s(x) = (x, f(x)) for any smooth function
f : M → R
n
. Therefore, C
∞
(M, R
n
) = Γ(
n
M
) by the correspondence: if f ∈ C
∞
(M, R
n
)
then s ∈ Γ(
n
M
) with s(x) = (x, f(x)).
Example 7.14***. The sections of line bundles over S
1
We ﬁrst note that the function p : R → S
1
by p(θ) = (cos θ, sinθ) ∈ R
2
or e
θi
∈ C
is a local diﬀeomorphism. An inverse of p : (θ
0
−
π
2
, θ
0
+
π
2
) → p(θ
0
−
π
2
, θ
0
+
π
2
) is
p
−1
(x, y) = arcsin(y cos θ
0
− x sinθ
0
) + θ
0
. Of course, p
−1
isn’t globally well deﬁned as
p
−1
(cos θ, sinθ) = {θ + 2kπ  k ∈ Z}
In Example 7.13*** it was shown that the sections of
1
S
1
were given by the smooth
functions f : S
1
→ R. We also note that these sections can be given by the functions
{g : R → R  g(θ +2π) = g(θ)}, i.e., periodic functions of the angle rather than of a point
CHAPTER 7 VECTOR BUNDLES 7
on the circle. Given a section s ∈ Γ(
1
S
1
) with s(cos θ, sinθ) = ((cos θ, sinθ), f(cos θ, sinθ),
the function g is g(θ) = f(cos θ, sinθ). Conversely, given a periodic function g let the
section s be s(cos θ, sinθ) = ((cos θ, sinθ), g(θ)). The section is welldeﬁned since g is
periodic and it is smooth since p has a smooth local inverse.
We now describe the sections of γ
1
S
1
. The bundle was described in Example 7.4***. We
show that Γ(γ
1
S
1
) correspond to the functions {g : R → R  g(θ + 2π) = −g(θ)}. Given a
function g take s to be the section s(e
θi
) = (e
θi
, g(θ)e
θ
2
i
). Conversely, given a section s,
it can be written as s(e
θi
) = (e
θi
, h(e
θi
)), where by the deﬁnition of γ
1
S
1
, h(e
θi
) is a real
multiple of e
θ
2
i
. Then let g(θ) = h(e
θi
)e
−
θ
2
i
.
Remark 7.15***. A section of a vector bundle is a way of choosing an element in each
ﬁber that varies in a smooth manner. One speaks of an “element of a vector space” and
the appropriate generalization to a vector bundle usually is a “section of a vector bundle.”
Example 7.16***. The zero section
Every vector space has a distinguished element: zero, the additive identity. If E
π
−→ M
is a vector bundle, let 0
x
∈ π
−1
(x) be the zero. Let z : M → E be deﬁned by z(x) = 0
x
for all x ∈ M. The section z is called the zero section.
We check that this map is a smooth section. If x ∈ M take an open neighborhood that
is both part of a coordinate neighborhood (U, φ) and a trivializing neighborhood for E,
µ : U → U × R
m
. The map z is smooth on U if and only if Z = (φ × I
V
) ◦ µ ◦ z ◦ φ
−1
:
φ(U) → φ(U) ×R
m
is smooth. The map Z is Z(x) = (x, 0) which is a smooth map of an
open subset of R
n
to R
n
×R
m
. Therefore, z is a smooth section.
Proposition 7.17***. A bundle π : E → M with an ndimensional ﬁber is a trivial
bundle if and only if it has n sections {s
1
, · · · , s
n
} such that s
1
(x), · · · , s
n
(x) are a basis
for π
−1
(x) for each x ∈ M.
Proof. If f : M × R
n
→ E is a bundle equivalence over M, then let s
i
be deﬁned by
s
i
(x) = f(x, e
i
). The sections {s
1
, · · · , s
n
} are the required set of sections. Note that
s
i
(x) = (x, e
i
) are the required sections for the trivial bundle.
If π : E → M is a bundle with sections {s
1
, · · · , s
n
} such that s
1
(x), · · · , s
n
(x) are a
basis for π
−1
(x) for each x ∈ M, then
f : M ×R
n
→ E
deﬁned by f(x, (a
1
, · · · , a
n
)) =
n
i=1
a
i
s
i
(x) is a smooth map and isomorphism on each
ﬁber. Therefore f is a bundle equivalence.
Theorem 7.18***. If N is an nmanifold, then there is a vector bundle π
N
: TN → N
such that π
−1
N
(x) = TM
x
. Furthermore, if M is an mmanifold, and f : N → M is a
smooth map, then there is a bundle map f
∗
: TN → TM deﬁne by f
∗

TN
x
= f
∗x
for each
x ∈ N, i.e., the bundle map on TN
x
is the derivative map.
The ﬁrst part of the theorem asserts the existence of a vector bundle. The substance,
the justiﬁcation to call the bundle the tangent bundle is in the second paragraph. The
derivative map at each point combines to form a smooth map. It indicates that the
8 CHAPTER 7 VECTOR BUNDLES
construction of the tangent bundle is the philosophically correct manner to combine the
tangent spaces into a coherent whole.
Proof of Theorem 7.18***. For U an open set in R
n
, the tangent bundle is a simple cross
product U ×R
n
. A basis in the ﬁber over a point p is
n
i=1
a
i
∂
∂x
i
¸
¸
¸
¸
p
where
∂
∂x
i
¸
¸
¸
p
represents
the equivalence class of the path t → te
i
+p. Observe that TU = U ×R
n
agrees with the
usual notions from Calculus as discussed in the beginning of this chapter.
We now construct the tangent bundle for a smooth nmanifold M. Let {(U
i
, φ
i
)i ∈ I}
be a countable atlas, and let f
ji
: U
i
∩ U
j
→ GL(R
n
) be deﬁned by
f
ji
(u) = (φ
j
◦ φ
−1
i
)
∗φ
i
(u)
= D(φ
j
◦ φ
−1
i
)(φ
i
(u)).
First note that since φ
j
◦ φ
−1
i
is a C
∞
function on an open set of R
n
into R
n
, each matrix
entry of the Jacobian of φ
j
◦φ
−1
i
is a C
∞
function. Therefore, D(φ
j
◦φ
−1
i
)(φ
i
(u)) = f
ji
(u)
is a smooth function.
We next show that these maps satisfy the conditions of Theorem 7.8*** and so deﬁne
a vector bundle, TM → M. The above functions f
ji
(u) = (φ
j
◦ φ
−1
i
)
∗φ
i
(u)
satisfy the
condition for transition functions. Suppose u ∈ U
i
∩ U
j
∩ U
k
, then (φ
j
◦ φ
−1
k
)
∗φ
k
(u)
(φ
k
◦
φ
−1
i
)
∗φ
i
(u)
= (φ
j
◦ φ
−1
i
)
∗φ
i
(u)
by the chain rule. Hence f
jk
(u) ◦ f
ki
(u) = f
ji
(u). The
hyposthesis of Theorem 7.8*** are satisﬁed and so give a welldeﬁned bundle, the tangent
bundle of M.
The next task is to show that if f : N → M is a smooth map, then there is a bundle
map f
∗
: TN → TM deﬁne by f
∗

TN
x
= f
∗x
for each x ∈ N.
The map f
∗
is well deﬁned and linear on each tangent space TN
x
by the given formula.
It must be shown that it is a smooth map from TN to TM. Let π
N
: TN → N and
π
M
: TM → M be the projections for the tangent bundles. Take any point in TN. It is
in the ﬁber over some point in N, say x ∈ N.
Let (W, ψ) be a chart on M with f(x) ∈ W. By construction of the tangent bundle,
π
−1
M
(W) is open in TM and
Ψ : π
−1
M
(W) → ψ(W) ×R
m
v → (ψ(π
M
(v)), ψ
∗π
M
(v)
(v))
is a coordinate chart.
Let (U, φ) be a chart on N with x ∈ U and f(U) ⊂ W. Again, by construction of the
tangent bundle, π
−1
N
(U) is open in TN and
Φ : π
−1
N
(U) → φ(U) ×R
n
v → (φ(π
N
(v)), φ
∗π
N
(v)
(v))
is a coordinate chart. We check smoothness by using Proposition 2.18***. Suppose (x, v) ∈
Φ(π
−1
(U) ⊂ R
n
×R
n
. Then compute
Ψ◦ f
∗
◦ Φ
−1
((x, v)) = (ψ ◦ f ◦ φ
−1
(x), ψ
∗f(φ
−1
(x)
◦ f
∗φ
−1
(x)
◦ φ
−1
∗x
(v))
= (ψ ◦ f ◦ φ
−1
(x), (ψ ◦ f ◦ φ
−1
)
∗x
(v)).
CHAPTER 7 VECTOR BUNDLES 9
The last line follows by the chain rule, Theorem 5.3*** part 2. This map is C
∞
in (x, v)
since it the derivative map for ψ ◦ f ◦ φ
−1
a C
∞
function between open subsets of R
n
and
R
m
.
Deﬁnition 7.19***. A section of the tangent bundle of a manifold M is called a vector
ﬁeld on M.
Example 7.20***. The tangent bundle to the 2sphere TS
2
.
Let S
2
= {(x, y, z) ∈ R
3
 x
2
+ y
2
+ z
2
= 1}. Recall from Calculus that the tangent
plane to S
2
translated to (0, 0, 0) is P
(x,y,z)
= {(v
1
, v
2
, v
3
) ∈ R
3
 (x, y, z) · (v
1
, v
2
, v
3
) = 0}.
Let E = {(x, v) ∈ R
3
× R
3
 x = 1 and v ∈ P
x
} and π : E → S
2
by π((x, v)) = x is a
vector bundle which is bundle equivalent to the tangent bundle.
We ﬁrst examine the tangent bundle of S
2
using the atlas constructed in Example
2.9A*** for S
2
. Let U
1
= S
2
\ {(0, 0, 1)} and U
2
= S
2
\ {(0, 0, −1)}. Let φ
i
: U
i
→ R
2
for i = 1 and 2 be φ
1
(x, y, z) = (
2x
1−z
,
2y
1−z
) and φ
2
(x, y, z) = (
2x
1+z
,
2y
1+z
). Then φ
−1
1
(x, y) =
(
4x
x
2
+y
2
+4
,
4y
x
2
+y
2
+4
,
x
2
+y
2
−4
x
2
+y
2
+4
) and φ
2
◦ φ
−1
1
(x, y) = (
4x
x
2
+y
2
,
4y
x
2
+y
2
). The transition function
is then g
21
: U
1
∩ U
2
→ Gl(2, R) deﬁned by
(Eq 1***) g
21
(x) = D(φ
2
◦ φ
−1
1
)(φ
1
(x))
and
D(φ
2
◦ φ
−1
1
)((x, y)) =
_
4(y
2
−x
2
)
(x
2
+y
2
)
2
−8xy
(x
2
+y
2
)
2
−8xy
(x
2
+y
2
)
2
4(x
2
−y
2
)
(x
2
+y
2
)
2
_
It is C
∞
since φ
1
and D(φ
2
◦ φ
−1
1
) are C
∞
.
We now turn to π : E → S
2
. For each x ∈ S
2
, we note that π
−1
(x) = P
x
is a vector space,
a two dimensional subspace of R
3
. We deﬁne homeomorphisms h
i
: U
i
× R
2
→ π
−1
(U
i
)
for i = 1 and 2. These maps will be used to show E is a 4manifold. They will commute
with the projections and be linear on each ﬁber. Their inverses will be the trivialization
maps for the bundle. The compatibility condition for the charts is the smoothness of the
transition functions.
Let
h
1
: U
1
×R
n
→ π
−1
(U
1
)
(x, v) → (x, Dφ
−1
1
(φ
1
(x))v)
and
h
2
: U
2
×R
n
→ π
−1
(U
2
)
(x, v) → (x, Dφ
−1
2
(φ
2
(x))v)
The inverses are
µ
1
: π
−1
(U
1
) → U
1
×R
n
(x, v) → (x, Dφ
1
(x)v)
and
µ
2
: π
−1
(U
2
) → U
2
×R
n
(x, v) → (x, Dφ
2
(x)v)
These maps are homeomorphisms as
Dφ
−1
i
(φ
i
(x))Dφ
i
(x) = I
P
x
and Dφ
i
(x)Dφ
−1
i
(φ
i
(x)) = I
R
2
10 CHAPTER 7 VECTOR BUNDLES
by the chain rule. Note that the transition function for the bundle E is
g
21
(x) = (Dφ
2
(x)Dφ
−1
1
(φ
1
(x)) = D(φ
2
◦ φ
−1
1
)(φ
1
(x))
which is the same as we computed for the tangent bundle.
We further pursue Example 7.20*** and prove that TS
2
is not trivial. We actually show
more. We prove that S
2
doesn’t have any nowhere zero section.
The proof of this theorem requires that the reader knows some of the basics of homotopy.
We will require
(1) The deﬁnition of a homotopy class.
(2) The fact that the map f : S
1
→ R
2
\ {(0, 0)} deﬁned by f(e
θi
) = e
2θi
or
f((cos θ, sinθ)) = (cos 2θ, sin2θ) is not homotopic to a constant map.
We state the speciﬁc elements we will use in the following lemma.
Lemma 7.21***. Suppose g : S
1
→ GL(2, R) is a continuous function. For any continu
ous function f : S
1
→ R
2
\{(0, 0)}, let G(f) : S
1
→ R
2
\{(0, 0)} by G(f)(p) = g(p)(f(p)).
We then have
(1) If f : S
1
→ R
2
\ {(0, 0)} be any continuous map, then the map f is homotopic to
−f.
(2) If f, h : S
1
→ R
2
\ {(0, 0)} and f is homotopic to h, then G(f) is homotopic to
G(h).
(3) Let S
1
⊂ S
2
be the equator and g the transition function that from TS
2
deﬁned
in (Eq I***) of Example 7.20***. Furthermore, let ι : S
1
→ R
2
\ {(0, 0)} be the
constant map ι(p) = (1, 0). Then the map G(ι) is not homotopic to a constant
map.
Proof. To show item 1 we give a homotopy from f to −f. Let H : S
1
×[0, 1] → R
2
\{(0, 0)}
by
H(p, t) =
_
cos πt sin πt
−sin πt cos πt
_
f(p).
Item 2 is also easy to see by the appropriate homotopy. Suppose that H : S
1
×[0, 1] →
R
2
\ {(0, 0) is the homotopy between f and h, then (p, t) → g(p)(H(p, t)) is the homotopy
between G(f) and G(h).
To show item 3, we ﬁrst write p = (x, y, 0) for p on the equator of S
2
and recall that
φ
1
(p) = (2x, 2y) for the map φ, stereographic projection (see Example 7.20***). Since the
transition function g(p) is the matrix given in (Eq I***) of Example 7.20***,
G(ι)(p) = D(φ
2
◦ φ
−1
1
)(φ
1
(p))(ι(p))
=
_
4((2y)
2
−(2x)
2
)
((2x)
2
+(2y)
2
)
2
−8(2x2y)
((2x)
2
+(2y)
2
)
2
−8(2x2y)
((2x)
2
+(2y)
2
)
2
4((2x)
2
−(2y)
2
)
((2x)
2
+(2y)
2
)
2
_
_
0
1
_
=
_
y
2
−x
2
−2xy
_
The last line follows since x
2
+y
2
= 1. Writing x = cos θ and y = sin θ, G(ι)(cos θ, sinθ) =
−(cos 2θ, sin2θ) using the double angle identities. By item 1, G(ι) is homotopic to the
map (cos θ, sinθ) → (cos 2θ, sin2θ), which is not homotopic to a constant map.
CHAPTER 7 VECTOR BUNDLES 11
Theorem 7.22***. There is no nonzero vector ﬁeld on S
2
.
As a consequence of this theorem and Proposition 7.17***, TS
2
is not equivalent to a
trivial vector bundle.
Proof. We use proof by contradition. Suppose V is a nowhere zero section of TS
2
, i.e., V is
a nowhere zero vector ﬁeld. We examine the two maps that characterize the vector ﬁeld in
the trivializing neighborhoods for TS
2
as given in Example 7.20***. We use the maps and
notation from Example 7.20***. For each p ∈ S
2
\ {(0, 0, 1)}, φ
1∗
(V
p
) = (φ
1
(p), f(p)) ∈
R
2
×R
2
\ {(0, 0)} and for each p ∈ S
2
\ {(0, 0, −1)}, φ
2∗
(V
p
) = (φ
2
(p), h(p)) ∈ R
2
×R
2
\
{(0, 0)}. The images of the maps f and g avoid (0, 0) precisely because the vector ﬁeld is
nowhere zero. It is easy to see that f and g are homotopic to constant maps. The null
homotopy for f is H : S
1
×[0, 1] → R
2
\ {(0, 0)} by H(p, t) = f(φ
−1
1
(tφ
1
(p))) and for h is
K : S
1
× [0, 1] → R
2
\ {(0, 0)} by K(p, t) = h(φ
−1
1
(tφ
1
(p))). The maps f and h are also
related by the transition function on the sphere by h(p) = g
21
(p)f(p). In other words, if f
is null homotopic then G(f) is also null homotopic. This conclusion contradicts item 3 of
the lemma. Therefore TS
2
doesn’t have a nonzero section.
Duality and Dual Bundles.
We review the notion of duality from linear algebra. Suppose V is a vector space. Let
V
∗
be the dual vector space Hom(V, R), the vector space of linear functions from V to R.
The relationship between V and V
∗
is usually expressed in the evalutation pairing:
V
∗
×V → R
(f, v) → f(v)
We give the basic properties of V
∗
in the next two theorems. The ﬁrst theorem is usually
covered in a linear algebra course. The second isn’t diﬃcult, but has a ﬂavor diﬀerent from
a ﬁrst course in linear algebra.
Theorem 7.23. For any ﬁnite dimensional vector space V , let V
∗
be the dual space of
linear functionals.
(1) V
∗
is a vector space of the same dimension as V .
(2) Suppose that e
1
, · · · , e
n
is a basis for V . Let e
∗
i
(
n
j=1
a
j
e
j
) = a
i
. Then e
∗
1
, · · · , e
∗
n
is a basis for V
∗
called the dual basis.
(3) Let V and W be vector spaces and f : V → W be a linear map. Let f
∗
: W
∗
→ V
∗
by f
∗
(g) = g ◦ f.
Proof. We can add linear functionals and multiply be a constant. If fact V
∗
= Hom(V, R)
is isomorphic to the vector space of n by 1 matrices. The isomorphism is given by a choice
of basis. If the choice of basis is e
1
, · · · , e
n
, then e
∗
i
is the matrix with a 1 in the i
th
row
and zeros elsewhere. This shows items 1 and 2.
Item 3 is a short computation. Suppose g
1
, g
2
∈ V
∗
. Then,
f
∗
(ag
1
+bg
2
) = (ag
1
+bg
2
) ◦ f
= ag
1
◦ f +bg
2
◦ f
= f
∗
(ag
1
)
= af
∗
(g
1
) +bf
∗
(ag
2
)
12 CHAPTER 7 VECTOR BUNDLES
Theorem 7.24***. The following are properties of ∗. Suppose V , W and X are ﬁnite
dimensional vector spaces.
(1) If h : V → W and f : W → X are linear, then (f ◦ h)
∗
= h
∗
◦ f
∗
(2) I
∗
V
= I
V
∗
(3) If f : W → V is linear and onetoone, then f
∗
is onto.
(4) If f : W → V is linear and onto, then f
∗
is onetoone.
(5) If f : W → V is an isomorphism, then f
∗
is an isomorphism and (f
−1
)
∗
= (f
∗
)
−1
.
(6) The map F : GL(V ) → GL(V
∗
) by F(f) = f
∗
is a C
∞
map.
Proof. We check each item.
Proof of item 1:
(f ◦ h)
∗
(α) = α ◦ f ◦ h
= h
∗
(α ◦ f)
= f
∗
◦ h
∗
(α)
Proof of item 2: I
∗
V
(α) = α ◦ I
V
= α = I
V
∗
Proof of item 3: The function f is onetoone if and only if there is a function h such
that h ◦ f = I
V
. Hence (h ◦ f)
∗
= I
∗
V
. By 1 and 2, f
∗
◦ h
∗
= I
V
∗. Since f
∗
has a right
inverse, f
∗
is onto.
Proof of item 4: The function f is onto if and only if there is a function h such that
f ◦ h = I
V
. Hence (f ◦ h)
∗
= I
∗
V
. By 1 and 2, h
∗
◦ f
∗
= I
V
∗. Since f
∗
has a left inverse,
f
∗
is onetoone.
Proof of item 5: Since f is onetoone and onto, so is f
∗
by 3 and 4. Now, f ◦ f
−1
= I
V
.
Items 1 and 2 imply, (f
−1
)
∗
◦ f
∗
= I
V
∗, therefore (f
−1
)
∗
= (f
∗
)
−1
Proof of item 6: We pick charts on GL(V ) and GL(V
∗
) by picking bases for V and V
∗
.
Suppose {e
i
i = 1 · · · n} is a basis for V . Let e
∗
i
∈ V
∗
, where e
∗
i
(
n
j=1
b
j
e
j
) = b
i
. Take
{e
∗
i
i = 1 · · · n} as a basis for V
∗
. If f is represented by the matrix M in the e basis then
f
∗
is represented by M
T
in the e
∗
basis. We check this fact.
f
∗
(
n
i=1
a
i
e
∗
i
)(
n
j=1
b
j
e
j
) = (
n
i=1
a
i
e
∗
i
)
n
j=1
n
k=1
m
kj
b
j
e
k
=
n
i,j=1
a
i
m
ij
b
j
= (
n
i,j=1
m
ij
a
i
e
∗
j
)(
n
k=1
b
k
e
k
)
So, f
∗
(
n
i=1
a
i
e
∗
i
) =
n
i,j=1
m
ij
a
i
e
∗
j
and M
T
represents f
∗
in the e
∗
basis. If we consider
GL(V ) and GL(V
∗
) as open sets in R
n
2
via matrix coordinates, then F is M → M
T
. Since
transposing is just a permutation of the entries, the map is C
∞
and in fact, linear.
If π : E → M is a vector bundle E with ﬁber the vector space V , then we can then
construct E
∗
, a dual bundle ρ : E
∗
→ M. The ﬁber should be V
∗
, but how should the
ﬁbers be assembled? Consider the following: each point in V
∗
gives a map from V to R, so
CHAPTER 7 VECTOR BUNDLES 13
each point in ρ
−1
(x) should give a map from π
−1
(x) to R. A section of ρ : E
∗
→ M picks
out a point in each ﬁber and varies in a smooth manner. Therefore, we want a section of
ρ : E
∗
→ M to give a map to R which is linear on each ﬁber. Another way to think of
the above description is that a section of the dual bundle should give a bundle map of E
to
1
M
, the trivial line bundle over M.
Before proceeding to the construction of the cotangent bundle, we ﬁrst undertake a
discussion of the meaning of a dual of a bundle, i.e., the co in cotangent. For a vector
space the meaning is expressed in the evaluation pairing,
V
∗
×V → R
(f, v) → f(v).
For vector bundles, there should be a pairing for each x ∈ M and this pairing should vary
smoothly in x. Therefore there should be an evaluation pairing,
(2) Γ(E
∗
) ×Γ(E) → Γ(
1
M
)
with (f, s)(x) = (x, f(x)(s(x))). Note that this condition is guided by Remark 7.15***.
Deﬁnition 7.25***. Suppose E is a vector bundle π : E → M. The vector bundle
ρ : E
∗
→ M is the dual bundle to E if ρ
−1
(x) = (π
−1
(x))
∗
for all x ∈ M and if E
∗
satisﬁes
the following property: There is an evaluation pairing
(2) F : Γ(E
∗
) ×Γ(E) → Γ(
1
M
)
deﬁned by F(f, s) = f(x)(s(x)) for all x ∈ M.
The pairing is the usual pairing on each ﬁber, but the ﬁbers ﬁt together smoothly so the
evaluation varies smoothly over M.
The reader should notice that the dual bundle to the trivial bundle is again a trivial
bundle since the dual to the bundle π : M ×V → M is the bundle π : M ×V
∗
→ M.
Theorem 7.26***. Suppose E is a vector bundle π : E → M with ﬁber V , then there is
a dual vector bundle E
∗
, π : E
∗
→ M with ﬁber V
∗
.
Furthermore, if U = {(U
i
, ϕ
i
)  i ∈ I} is a countable atlas for M with each U
i
also a
trivializing neighborhood for E and g
ij
: U
i
∩U
j
→ GL(V ) are the transition functions for
E, then f
ij
: U
i
∩U
j
→ GL(V
∗
) deﬁned by f
ij
(x) = (g
ji
(x))
∗
for each x ∈ U
i
∩U
j
forms a
set of transition functions for E
∗
.
Proof. The cover U, the vector space V
∗
and the functions f
ij
satisfy Theorem 7.8*** and
so deﬁne a vector bundle E
∗
. To see this fact, we must check the condition the f
ij
must
satisfy.
f
ik
(x)f
kj
(x) = (g
ki
(x))
∗
(g
jk
(x))
∗
= (g
jk
(x)g
ki
(x))
∗
by properties of *, Theorem 7.24***
= (g
ji
(x))
∗
by properties of transition functions, Proposition 7.6***
= f
ij
(x)
14 CHAPTER 7 VECTOR BUNDLES
which is the required property. It remains to check the property of the pairing (2). This
condition is a local condition and so we check it in the trivializations. Suppose that
µ : π
−1
(U
i
) → U
i
× V by µ(e) = (π(e), h
i
(e)) and γ : π
∗−1
(U
i
) → U
i
× V
∗
by γ(e) =
(π(e), k
i
(e)). If s
1
and s
2
are sections then for x ∈ U
i
the pairing (2) is
(s
2
(x), s
1
(x)) → (x, k
i
(s
2
(x))(h
i
(s
1
(x)))).
Since s
1
, s
2
, h
i
, and k
i
are smooth, the pairing is also smooth, but we must check that the
pairing is welldeﬁned. If x ∈ U
j
, then f
ij
(x)(k
j
(s
2
(x))) = g
∗
ji
(x)(k
j
(s
2
(x))) = k
i
(s
2
(x))
and g
ij
(x)(h
j
(s
1
(x))) = h
i
(s
1
(x)). Hence,
k
i
(s
2
(x))(h
i
(s
1
(x)) = f
ij
(x)(k
j
(s
2
(x)))(g
ij
(x)(h
j
(s
1
(x))))
= g
∗
ji
(x)(k
j
(s
2
(x)))(g
ij
(x)(h
j
(s
1
(x))))
= (k
j
(s
2
(x)))(g
ji
(x)g
ij
(x)(h
j
(s
1
(x))))
= k
j
(s
2
(x))(h
j
(s
1
(x)))
So the pairing is welldeﬁned.
Example 7.27***. The Cotangent Bundle
Suppose M is a smooth nmanifold. The dual bundle to the tangent bundle is called the
cotangent bundle. The cotangent space at a point x is (TM
x
)
∗
which we denote T
∗
M
x
.
The bundle dual to the tangent bundle is (TM)
∗
in the duality notation, but is usually
denoted T
∗
M.
Let f
ij
(m) = (ϕ
j
ϕ
−1
i
)
∗
ϕ
j
(m)
. Note that if g
ij
are the transition functions from which we
constructed the tangent bundle above, then f
ij
(x) = (g
ji
(x))
∗
. This is the construction
speciﬁed by Theorem 7.25*** to construct the bundle dual to the tangent bundle. Call it
the cotangent bundle and denote it T
∗
M.
If φ : N → M is a smooth map between smooth manifolds, then there is an induced
map φ
∗
φ(x)
: T
∗
M
φ(x)
→ T
∗
N
x
deﬁne by φ
∗
φ(x)
(γ) = γ ◦ φ
∗x
.
Example 7.28***. A section of the cotangent bundle, df.
Suppose f : M → R, then let df
x
= π ◦ f
∗
where π : TR = R×R → R is projection in
the ﬁber direction, i.e., π(a
∂
∂x
i
¸
¸
¸
p
) = a. Suppose that φ : M → N is a smooth map and an
onto map. Then,
φ
∗
(df
x
) = df
x
◦ φ
∗
= π ◦ f
∗
◦ φ
∗
= d(f ◦ φ).
Suppose f : M → R then we will show that df deﬁnes a section of T
∗
M. If x ∈ U
i
,
then x → d(f ◦ ϕ
−1
i
)
ϕ
i
(x)
is a smooth map since x → (f ◦ ϕ
−1
i
)
∗ϕ
i
(x)
is smooth. In the
notation of Theorem 7.8***, df
x
is represented by [x, d(f ◦ ϕ
−1
i
)
ϕ
i
(x)
, i]. We have that if
x ∈ U
i
∩ U
j
, then
[x, d(f ◦ ϕ
−1
i
)
ϕ
i
(x)
, i] = [x, d(f ◦ ϕ
−1
j
)
ϕ
j
(x)
, j]
as f
ij
(x)((d(f ◦ ϕ
−1
j
)
ϕ
j
(x)
) = (ϕ
j
◦ ϕ
−1
i
)
∗
d(f ◦ ϕ
−1
j
)
ϕ
j
(x)
= d(f ◦ ϕ
−1
j
◦ ϕ
j
◦ ϕ
−1
i
)
ϕ
i
(x)
=
d(f ◦ ϕ
−1
i
)
ϕ
i
(x)
using the equivalence relation of Theorem 7.8***.
CHAPTER 7 VECTOR BUNDLES 15
Remark 7.29. Sections of the tangent and cotangent bundle are both used to generalize
types of integration.
Sections of the tangent bundle are used to integrate, solve a diﬀerential equation. Sec
tions of the cotangent bundle are used to compute deﬁnite integrals.
Exercises
Exercise 1***. Let π : E → M be a vector bundle. Show that it has sections that are
not identically zero.
This next exercise justiﬁes the terminology bundle equivalence.
Exercise 2***. If E
1
an E
2
are vector bundles over the manifold M and if E
1
is bundle
equivalent to E
2
by a bundle equivalence f, then E
2
is bundle equivalent to E
1
by f
−1
which is also a bundle equivalence.
Show that bundle equivalence is an equivalence relation.
Exercise 3***. Show that every vector bundle over an interval is trivial.
Exercise 4***. Show that up to bundle equivalence, there are exactly two distinct line
bundles over the circle.
This next exercise introduces the notion of a Reimannian metric.
Exercise 5***. Suppose that E is a vector bundle π : E → M. A Riemannian metric on
E is a choice of inner product <, >
x
for each ﬁber π
−1
(x) such that there is an induced
map on sections
<, >: Γ(E) ×Γ(E) → Γ(
1
M
)
deﬁned by < s
1
, s
2
> (x) =< s
1
(x), s
2
(x) >
x
.
a. Show that every vector bundle has a Riemannian metric. This argument will require a
partition of unity.
b. Show that E is bundle equivalent to E
∗
.
CHAPTER 9
MULTILINEAR ALGEBRA
In this chapter we study multilinear algebra, functions of several variables that are linear
in each variable separately. Multilinear algebra is a generalization of linear algebra since a
linear function is also multilinear in one variable. If V
1
, V
2
, · · · , V
k
and W are vector spaces,
then we wish to understand what are all the multilinear maps g : V
1
×V
2
×· · · ×V
k
→W
and notation to systematically express them. This may seem like a diﬃcult and involved
problem. After all the reader has probably taken considerable eﬀort to learn linear algebra
and multilinear algebra must be more complicated. The method employed is to convert g
into a linear map ˜ g on a diﬀerent vector space, a vector space called the tensor product of
V
1
, V
2
, · · · , V
k
. Since ˜ g is a linear map on a vector space, we are now in the realm of linear
algebra again. The beneﬁt is that we know about linear maps and how to represent all of
them. The cost is that the new space is a complicated space.
Deﬁnition 9.1***. Suppose V
1
, V
2
, · · · , V
k
and W are vector spaces. A function f :
V
1
×V
2
×· · · ×V
k
→W is called multilinear if it is linear in each of its variables, i.e.,
f(v
1
, · · · , v
i−1
,av
i
+ bv
i
, v
i+1
, · · · , v
k
)
= af(v
1
, · · · , v
i−1
, v
i
, v
i+1
, · · · , v
k
) + bf(v
1
, · · · , v
i−1
, v
i
, v
i+1
, · · · , v
k
)
for all a, b ∈ R, v
j
∈ V
j
for j = 1, · · · , k and v
i
∈ V
i
for i = 1, · · · , k.
Our objective is to reduce the study of multilinear maps to the study of linear maps. We
use F(V
1
, V
2
, · · · , V
k
) to denote the vector space having its basis {(v
1
, · · · , v
k
) ∈ V
1
×· · · ×
V
k
} = V
1
×· · · ×V
k
. Each element of V
1
×· · · ×V
k
is a basis element of F(V
1
, V
2
, · · · , V
k
).
For example, if V = R, then F(V ) is an inﬁnite dimesional vector space in which each
r ∈ R is a basis element. This vector space is enormous, but it is just an intermediate
stage. It has the following important property:
Lemma 9.2***. If V
1
, V
2
, · · · , V
k
and W are vector spaces, then linear maps from F(V
1
, V
2
, · · · , V
k
)
to W are in one to one correspondence with set maps from V
1
×· · · ×V
k
to W.
Proof. This property follows since a linear map is exactly given by specifying where a
basis should map, and V
1
× · · · × V
k
is a basis of F(V
1
, V
2
, · · · , V
k
). Given any set map
g : V
1
×· · · ×V
k
→W we obtain a linear map
˜
˜ g : F(V
1
, V
2
, · · · , V
k
) →W.
We next improve upon the construction of F by forming a quotient of F to make a
smaller space. We can do this improvement since we are not interested in set maps from
copyright c 2002 Larry Smolinsky
Typeset by A
M
ST
E
X
1
2 CHAPTER 9 MULTILINEAR ALGEBRA
V
1
×· · · ×V
k
to W but only in multilinear maps. Let R ⊂ F be the vector subspace of F
spanned by the following vectors
(1***)
(v
1
, · · · , v
i−1
,av
i
+ bv
i
, v
i+1
, · · · , v
k
)
−a(v
1
, · · · , v
i−1
, v
i
, v
i+1
, · · · , v
k
) −b(v
1
, · · · , v
i−1
, v
i
, v
i+1
, · · · , v
k
)
for each a, b ∈ R, v
j
∈ V
j
for j = 1, · · · , k and v
i
∈ V
i
for i = 1, · · · , k. The vector given in
(1)*** is a single vector expressed as a sum of three basis elements, and each basis element
is an ntuple in V
1
×· · · ×V
k
. The subspace R has the following important property
Lemma 9.3***. If V
1
, V
2
, · · · , V
k
and W are vector spaces, then linear maps from F(V
1
, V
2
, · · · , V
k
)
to W which vanish on R are in one to one correspondence with multilinear maps from
V
1
×· · · ×V
k
to W.
Proof. The correspondence is the same correspondence as is given in Lemma 9.2***. Using
the same notation as in the proof of Lemma 9.2***, we must show that g is multilinear if
and only if
˜
˜ g vanishes on R.
Suppose g : V
1
×· · · ×V
k
→W is a multilinear map. Then
g(v
1
, · · · , v
i−1
,av
i
+ bv
i
, v
i+1
, · · · , v
k
)
= ag(v
1
, · · · , v
i−1
, v
i
, v
i+1
, · · · , v
k
) + bg(v
1
, · · · , v
i−1
, v
i
, v
i+1
, · · · , v
k
)
which is true if and only if
˜
˜ g(v
1
, · · · , v
i−1
,av
i
+ bv
i
, v
i+1
, · · · , v
k
)
= a
˜
˜ g(v
1
, · · · , v
i−1
, v
i
, v
i+1
, · · · , v
k
) + b
˜
˜ g(v
1
, · · · , v
i−1
, v
i
, v
i+1
, · · · , v
k
)
=
˜
˜ g(a(v
1
, · · · , v
i−1
, v
i
, v
i+1
, · · · , v
k
)) +
˜
˜ g(b(v
1
, · · · , v
i−1
, v
i
, v
i+1
, · · · , v
k
))
which is true if and only if
˜
˜ g((v
1
, · · · , v
i−1
,av
i
+ bv
i
, v
i+1
, · · · , v
k
)
−a(v
1
, · · · , v
i−1
, v
i
, v
i+1
, · · · , v
k
) −b(v
1
, · · · , v
i−1
, v
i
, v
i+1
, · · · , v
k
)) = 0.
In the computation above,
˜
˜ g is a linear map and each ntuple is a basis vector in the
vector space. The last line states that
˜
˜ g vanishes on R and the ﬁrst line states that g is
multilinear. Hence g is multilinear if and only if
˜
˜ g vanishes on R.
We are now ready to deﬁne the vector space discussed in the beginning of the chapter.
Deﬁnition 9.4***. Suppose that V
1
, V
2
, · · · , V
k
are vector spaces. Then the vector space
F(V
1
, · · · , V
k
)/R along with the map φ : V
1
× · · · × V
k
→ F(V
1
, · · · , V
k
)/R is call the
tensor product of V
1
, V
2
, · · · , V
k
. The vector space F/R is denoted V
1
⊗ · · · ⊗ V
k
. The
image φ((v
1
, · · · , v
k
)) is denoted v
1
⊗ · · · ⊗ v
k
. Usually the map φ is supressed, but it is
understood to be present.
Usually the map φ is suppressed, but it is understood to be present. Although the vector
space F is inﬁnite dimensional, we will soon show that V
1
⊗· · · ⊗V
k
is ﬁnite dimensional
(Proposition 9.8***). We ﬁrst show that
CHAPTER 9 MULTILINEAR ALGEBRA 3
Proposition 9.5***. The map φ in the deﬁnition of the tensor product is a multilinear
map.
Proof. We must show that
φ(v
1
, · · · , v
i−1
,av
i
+ bv
i
, v
i+1
, · · · v
k
)
= aφ(v
1
, · · · , v
i−1
, v
i
, v
i+1
, · · · v
k
) + bφ(v
1
, · · · , v
i−1
, v
i
, v
i+1
, · · · v
k
)
or, using the notation of Deﬁnition 9.4***,
(v
1
⊗· · · ⊗v
i−1
⊗av
i
+ bv
i
⊗v
i+1
⊗· · · ⊗v
k
)
−a(v
1
⊗· · · ⊗v
i−1
⊗v
i
⊗v
i+1
· · · ⊗v
k
)
−b(v
1
⊗· · · ⊗v
i−1
⊗v
i
⊗v
i+1
⊗· · · ⊗v
k
) = 0
This equation is equivalent to the following statement in F,
(v
1
, · · · , v
i−1
,av
i
+ bv
i
, v
i+1
, · · · v
k
)
−a(v
1
, · · · , v
i−1
, v
i
, v
i+1
, · · · v
k
) −b(v
1
, · · · , v
i−1
, v
i
, v
i+1
, · · · v
k
) ∈ R
The vector on the left is in R since it is the element in expression (1)***.
The main property of the tensor product is the universal mapping property for multi
linear maps. It is stated in the following theorem.
Proposition 9.6***. Suppose V
1
, V
2
, · · · , V
k
are vector spaces. The tensor product φ :
V
1
×· · ·×V
k
→V
1
⊗· · ·⊗V
k
satisﬁes the following property, the universal mapping property
for multilinear maps:
If W is a vector space and g : V
1
×· · · × V
k
→ W is a multilinear map, then there is a
unique linear map ˜ g : V
1
⊗· · · ⊗V
k
→W such that ˜ g ◦ φ = g.
Proof. Given the multilinear map g, there is a unique linear map
˜
˜ g : F(V
1
, · · · , V
k
) → W
by Lemma 9.2***. Since g is multilinear, the map
˜
˜ g vanishes on R by Lemma 9.3***.
Hence there is a unique welldeﬁned map induce by
˜
˜ g, call it ˜ g : F/R →W.
The ability of the tensor product to convert multilinear maps into linear maps is an
immediate consequence of Proposition 9.6***.
Theorem 9.7***. Suppose V
1
, V
2
, · · · , V
k
and W are vector spaces. Linear maps ˜ g : V
1
⊗
· · ·⊗V
k
→W are in one to one correspondence with multilinear maps g : V
1
×· · ·×V
k
→W.
Proof. Given a multilinear map g, Proposition 9.6*** produces the unique linear map ˜ g.
Given a linear map ˜ g let g = ˜ g ◦ φ. The map g is a composition of a linear map and a
multilinear map, Proposition 9.5***. The composition of a linear map and a multilinear
map is a multilinear linear map. The reader should check this fact.
4 CHAPTER 9 MULTILINEAR ALGEBRA
Theorem 9.8***. Suppose V
1
, V
2
, · · · , V
k
are vector spaces and dimV
i
= n
i
. Let {e
i
j
 j =
1, · · · , n
i
} be a basis for V
i
. Then dimV
1
⊗· · · ⊗V
k
= n
1
n
2
· · · n
k
and {e
1
j
1
⊗e
2
j
2
⊗· · · ⊗e
k
j
k

j
i
= 1, · · · , n
i
, i = 1, · · · , k} is a basis for the tensor product V
1
⊗· · · ⊗V
k
.
Proof. We ﬁrst show that dimV
1
⊗ · · · ⊗ V
k
≥ n
1
n
2
· · · n
k
. Let W be the vector space of
dimension n
1
n
2
· · · n
k
and label a basis E
j
1
,··· ,j
k
for j
i
= 1, · · · , n
k
. Deﬁne L : V
1
× · · · ×
V
k
→W by
L(
n
1
j
1
=1
a
1j
1
e
1
j
1
, · · · ,
n
k
j
k
=1
a
kj
k
e
k
j
k
) =
n
1
j
1
=1
· · ·
n
k
j
k
=1
a
1j
1
· · · a
kj
k
E
j
1
,··· ,j
k
The map L maps onto a basis of W since L(e
1
j
1
, · · · , e
k
j
k
) = E
j
1
,··· ,j
k
.
We next observe that L is multilinear. Let v
r
=
n
r
j
r
=1
a
rj
r
e
r
j
r
for r = 1, · · · , k and
v
i
=
n
i
j
i
=1
a
ij
i
e
i
j
i
so that av
i
+ bv
i
=
n
i
j
i
=1
(aa
ij
i
+ ba
ij
i
)e
i
j
i
and
L(v
1
, · · · , v
i−1
, av
i
+ bv
i
, v
i+1
· · · , v
k
)
=
n
1
j
1
=1
· · ·
n
k
j
k
=1
a
ij
1
· · · a
i−1j
i−1
(aa
ij
i
+ ba
ij
i
)a
i+1j
i+1
· · · a
kj
k
E
j
1
,··· ,j
k
= a
n
1
j
1
=1
· · ·
n
k
j
k
=1
a
ij
1
· · · a
i−1j
i−1
(a
ij
i
)a
i+1j
i+1
· · · a
kj
k
E
j
1
,··· ,j
k
+ b
n
1
j
1
=1
· · ·
n
k
j
k
=1
a
ij
1
· · · a
i−1j
i−1
(a
ij
i
)a
i+1j
i+1
· · · a
kj
k
E
j
1
,··· ,j
k
= aL(v
1
, · · · , v
i−1
, v
i
, v
i+1
· · · , v
k
) + bL(v
1
, · · · , v
i−1
, v
i
, v
i+1
· · · , v
k
)
By Proposition 9.6***, there is an induced linear map
˜
L : V
1
⊗ · · · ⊗ V
k
→ W. This
map hits a basis since L maps onto a basis. Since
˜
L is linear, it is onto. Therefore
dimV
1
⊗· · · ⊗V
k
> dimW = n
1
n
2
· · · n
k
.
We show that {e
i
j
 j = 1, · · · , n
i
} is a spanning set. The set V
1
× · · · V
k
is a basis
of F(V
1
, · · · , V
k
) and F(V
1
, · · · , V
k
) maps onto the tensor product V
1
⊗ · · · ⊗ V
k
= F/R.
Therefore the elements of the form v
1
⊗ · · · ⊗ v
k
span the tensor product. Let v
i
=
n
i
j
i
=1
a
ij
i
e
i
j
i
for each i = 1, · · · , k. Then
v
1
⊗· · · ⊗v
k
=
n
1
j
1
=1
a
1j
1
e
1
j
1
⊗· · · ⊗
n
k
j
k
=1
a
kj
i
e
k
j
k
=
n
1
j
1
=1
· · ·
n
k
j
k
=1
a
1j
1
· · · a
kj
i
e
1
j
1
⊗· · · ⊗e
k
j
k
Since {e
i
j
 j = 1, · · · , n
i
} is a spanning set and its cardinality is n
1
n
2
· · · n
k
, it is
basis.
CHAPTER 9 MULTILINEAR ALGEBRA 5
Example 9.9***. A multilinear map g : R
3
×R
3
×R
3
→R.
Suppose that e
1
, e
2
, e
3
is the standard basis for R
3
. The vector space R
3
⊗R
3
⊗R
3
is 27
dimensional and has as a basis {e
i
⊗e
j
⊗e
k
 i, j, k = 1, 2, 3} This basis is usually represented
as a triple index, indexed by the basis elements of R
3
. The linear map ˜ g : R
3
⊗R
3
⊗R
3
→R
can be represented as a 1 ×27 matrix, but this is not the usual way to represent tensors.
Suppose that ˜ g(e
i
⊗ e
j
⊗ e
k
) = a
ijk
using the triple index to write the basis elements of
the tensor product. Then, using the e
i
basis for R
3
, we have
˜ g
x
1
x
2
x
3
y
1
y
2
y
3
z
1
z
2
z
3
=
3
i=1
3
j=1
3
k=1
a
ijk
x
i
y
j
z
k
Example 9.10***. R⊗· · · ⊗R is isomorphic to R, R⊗W is isomorphic to W.
A basis for R ⊗ · · · ⊗ R is 1 ⊗ · · · ⊗ 1 and a
1
⊗ · · · ⊗ a
k
= (a
1
· · · a
k
)1 ⊗ · · · ⊗ 1. The
isomorphism is to multiply the entries.
Similarly, R⊗W →W by r ×w →rw induces the second isomorphism.
These isomorphisms are natural and standard. They are natural because R is not an
abstract 1dimensional vector space, but R has a distinguished multiplicative unit 1. These
isomorphisms are used as an identiﬁcation in these notes and in physics literature.
The following theorem deﬁnes and gives properites of induced maps between tensor
products.
Theorem 9.11***. Suppose V
1
, · · · , V
k
and W
1
, · · · , W
k
are vector spaces. Further sup
pose that f
i
: V
i
→W
i
is a linear map for each i = 1, · · · , k. Then there is an induced linear
map f
1
⊗· · · ⊗f
k
: V
1
⊗· · · ⊗V
k
→W
1
⊗· · · ⊗W
k
deﬁned by f
1
⊗· · · ⊗f
k
(v
1
⊗· · · ⊗v
k
) =
f
1
(v
1
) ⊗· · · ⊗f
k
(v
k
). These induced maps satisfy the following two properties
(1) If Q
1
, · · · , Q
k
are another collection of vector spaces and g
i
: W
i
→ Q
i
is a linear
map for each i −1, · · · , k, then g
1
⊗· · · ⊗g
k
◦f
1
⊗· · · ⊗f
k
= (g
1
◦f
1
) ⊗· · · ⊗(g
k
◦f
k
)
(2) If I
X
denotes the identity on X, then I
V
1
⊗· · · ⊗I
V
k
= I
V
1
⊗···⊗V
k
Proof. Let φ : W
1
×· · · ×W
k
→W
1
⊗· · · ⊗W
k
be the map from the deﬁnition of the tensor
product. It is multilinear by Proposition 9.5***. Let L : V
1
× · · · × V
k
→ W
1
⊗ · · · ⊗ W
k
be deﬁned by L = φ ◦ (f
1
, · · · , f
k
). We show that L is multilinear since each f
i
is linear
and φ is multilinear.
L(v
1
, · · · , v
i−1
,av
i
+ bv
i
, v
i+1
, v
k
)
= f
1
(v
1
) ⊗· · · ⊗f
i−1
(v
i−1
) ⊗f
i
(av
i
+ bv
i
) ⊗f
i+1
(v
i+1
) ⊗· · · ⊗f(v
k
)
= f
1
(v
1
) ⊗· · · ⊗f
i−1
(v
i−1
) ⊗(af
i
(v
i
) + bf
i
(v
i
)) ⊗f
i+1
(v
i+1
) ⊗· · · ⊗f(v
k
)
= af
1
(v
1
) ⊗· · · ⊗f
i−1
(v
i−1
) ⊗(f
i
(v
i
)) ⊗f
i+1
(v
i+1
) ⊗· · · ⊗f(v
k
)
+ bf
1
(v
1
) ⊗· · · ⊗f
i−1
(v
i−1
) ⊗(f
i
(v
i
)) ⊗f
i+1
(v
i+1
) ⊗· · · ⊗f(v
k
)
= aL(v
1
, · · · , v
i−1
, v
i
, v
i+1
, · · · , v
k
) + bL(v
1
, · · · , v
i−1
, v
i
, v
i+1
, · · · , v
k
)
There is an induced linear map
˜
L by Theorem 9.7*** and this map is f
1
⊗· · · ⊗f
k
.
6 CHAPTER 9 MULTILINEAR ALGEBRA
The ﬁrst property follows from the deﬁnition of the induced map. We check the second
property,
I
V
1
⊗· · · ⊗I
V
k
(v
1
⊗· · · ⊗v
k
) = I
V
1
(v
1
) ⊗· · · ⊗I
V
k
(v
k
)
= v
1
⊗· · · ⊗v
k
,
which veriﬁes the second property.
We have the following corollary.
Corollary 9.12***. Suppose V
1
, V
2
, · · · , V
k
are vector spaces. Let {e
i
j
 j = 1, · · · , n
i
}
be a basis for V
i
and {(e
i
j
)
∗
 j = 1, · · · , n
i
} the dual basis for V
∗
i
. Using the isomorphism
from Example 9.10***, R⊗· · · ⊗R
∼
= R, we have that
(e
1
j
1
)
∗
⊗· · · ⊗(e
k
j
k
)
∗
= (e
1
j
1
⊗· · · ⊗e
k
j
k
)
∗
,
which gives an isomorphism V
∗
⊗· · · ⊗V
∗
∼
= (V ⊗· · · ⊗V )
∗
.
Proof. We check this formula on the basis given in Theorem 9.8***.
(e
1
j
1
)
∗
⊗· · · ⊗(e
k
j
k
)
∗
(e
1
i
1
⊗· · · ⊗e
k
i
k
) =
1 if i
1
= j
1
, · · · , i
k
= j
k
0 if i
r
= j
r
for some r
This formula follows from the deﬁnition of the induced map in Theorem 9.11*** and
the deﬁnition of the dual basis. It shows that {e
1
j
1
)
∗
⊗ · · · ⊗ (e
k
j
k
)
∗
} is the basis dual to
{e
1
j
1
⊗· · · ⊗(e
k
j
k
}.
Example 9.13***. Hom(V, W) = V
∗
⊗W. If {e
i
 i = 1, · · · n} is a basis for V , {e
i
 i =
1, · · · m} is a basis for W, and f ∈ Hom(V, W), then f =
m
i=1
n
j=1
a
ij
e
∗
j
⊗e
i
where (a
ij
)
is the matrix representing f in the given bases.
We check the formula:
(
n
i=1
n
j=1
a
ij
e
∗
j
⊗e
i
)(e
r
) =
n
i=1
n
j=1
a
ij
(e
∗
j
(e
r
)) ⊗e
i
=
n
i=1
a
ir
1 ⊗e
i
=
n
i=1
a
ir
e
i
Which is the rth column vector of the matrix. Also notice the last line used the ideniﬁ
cation from Example 9.10***.
CHAPTER 9 MULTILINEAR ALGEBRA 7
Example 9.14***. Bilinear Maps
The usual dot product in R
n
is a bilinear map. In fact, any bilinear map <, >: R
n
×
R
n
→R induces a linear map
˜
<, > : R
n
⊗R
n
→R.
Now,
˜
<, > ∈ Hom(R
n
⊗R
n
, R) and
Hom(R
n
⊗R
n
, R) = (R
n
⊗R
n
)
∗
by deﬁnition of the dual
= (R
n
)
∗
⊗(R
n
)
∗
using the isomorphism in 9.12***.
Using the standard basis {e
i
 i = 1, · · · , n} for R
n
, we can write
˜
<, > =
n
i=1
n
j=1
g
ij
e
∗
i
⊗
e
∗
j
. Readers who are familar with representing a bilinear form as a matrix should note that
(g
ij
) is the matrix for <, >. In practice the tilde is not used and we only use it here for
clariﬁcation.
Example 9.15***. The Cross Product in R
3
The vector cross product in R
3
is a bilinear map and so induces a map
−×− : R
3
⊗R
3
→R
3
.
If we use the identiﬁcations Hom(R
3
⊗ R
3
, R
3
) = (R
3
)
∗
⊗ (R
3
)
∗
⊗ R
3
, then the cross
product is
e
∗
2
⊗e
∗
3
⊗e
1
−e
∗
3
⊗e
∗
2
⊗e
1
+ e
∗
3
⊗e
∗
1
⊗e
2
−e
∗
1
⊗e
∗
3
⊗e
2
+ e
∗
1
⊗e
∗
2
⊗e
3
−e
∗
2
⊗e
∗
1
⊗e
3
.
Example 9.16***. The Stress Tensor
Consider a solid object and put it in the usual coordinate system so that the origin 0 is
at an interior point. One problem is to understand and describe the stresses at a point (the
origin). Since the interior point, the origin, isn’t moving, all the forces must be in balance.
This fact is a consequece of Newton’s laws of motion. However, we can ask about the stress
(force per unit area) on a surface through the origin. There may be forces perpendicular
to the surface, e.g., from compression and forces along the surface, e.g., shearing forces
from twisting.
Given a vector u ∈ R
3
, let S be the square with area u that is perpendicular to u. Let
F(u) be the force on the square S from the side that u points so that F(u) is a vector in
R
3
. Deﬁne a bilinear function τ by
τ(v, w) = v · DF(0)(w)
where DF(0) is the best linear approximation to F at the origin, i.e., the derivative and
the dot is the usual dot product. The number τ(e
i
, e
j
) is the i −th component of the force
per unit area acting on the surface perpendicular to e
j
. The stress tensor τ is a map
τ : R
3
⊗R
3
→R.
In terms of the identiﬁcation Hom(R
3
⊗ R
3
, R) = (R
3
⊗ R
3
)
∗
= (R
3
)
∗
⊗ (R
3
)
∗
we have
that τ =
3
i=1
3
j=1
τ
ij
e
∗
i
⊗e
∗
j
.
8 CHAPTER 9 MULTILINEAR ALGEBRA
Deﬁnition 9.17***. Suppose V is a vector space. Elements of the space V ⊗· · · ⊗V ⊗
V
∗
⊗· · ·⊗V
∗
, the rfold tensor of V and the sfold tensor of V
∗
are tensors of type (r, s)***.
Remark 9.18***. A tensor is an object that transforms ...
Since tensors are elements of a vector space, they are vectors. However, diﬀerent ter
minology is used emphasizing a conceptual diﬀerence. The primary object of study is
usually associated to the vector space V and a tensor of type (r, s) is then an auxilary
object. This perspective is apparent when one changes coordinates. If V is an n di
mensional vector space then the space of tensor of type (r, s) is an n(r + s)dimensional
vector space. A change of coordinates in V is given by a map in GL(V )
∼
= GL(n, R)
and a change in coordinates in the space of tensor of type (r, s) is given by a map in
GL(V ⊗· · · ⊗V
∗
)
∼
= GL(n(r + s), R). However the only change in coordinates allowed in
the tensor product are those that are induced from a change in coordinates V via Theorem
9.11***. If f : V →V , then f
∗
: V
∗
→V
∗
and
f ⊗· · · ⊗f ⊗f
∗
⊗· · · ⊗f
∗
: V ⊗· · · ⊗V ⊗V
∗
⊗· · · ⊗V
∗
→V ⊗· · · ⊗V ⊗V
∗
⊗· · · ⊗V
∗
are the induced change of coordinates.
We give the speciﬁc formula. Suppose e
1
, · · · , e
n
is a basis for V and the map f is
represented by the matrix (a
ij
). Then
f(
n
i=1
x
i
e
i
) =
n
i=1
n
j=1
a
ij
x
j
e
i
f(
n
i=1
y
i
e
∗
i
) =
n
i=1
j=1
a
ji
y
j
e
∗
i
If
n
i
1
,··· ,i
r
,j
1
,··· ,j
s
=1
T
j
1
,··· ,j
s
i
1
,··· ,i
r
e
i
1
⊗· · · ⊗e
i
r
⊗e
∗
j
1
⊗· · · ⊗e
∗
j
s
is a tensor of type (r, s)*** then
its image under the induced map is
(2***)
n
i
1
, · · · , i
r
, j
1
, · · · , j
s
p
1
, · · · , p
r
, q
1
, · · · , q
s
= 1
T
q
1
,··· ,q
s
p
1
,··· ,p
r
a
i
1
p
1
· · · a
i
r
p
r
a
q
1
j
1
· · · a
q
r
j
s
e
i
1
⊗· · · ⊗e
i
r
⊗e
∗
j
1
⊗· · · ⊗e
∗
j
s
Formula (2)*** is often expressed in physics by saying that under the linear change of
coordinates f, T
q
1
,··· ,q
s
p
1
,··· ,p
r
transforms to
n
p
1
,··· ,p
r
,q
1
,··· ,q
s
=1
T
q
1
,··· ,q
s
p
1
,··· ,p
r
a
i
1
p
1
· · · a
i
r
p
r
a
q
1
j
1
· · · a
q
r
j
s
.
Furthermore, the summation sign is often supressed.
We now turn our attention to functions that are alternating and multilinear.
CHAPTER 9 MULTILINEAR ALGEBRA 9
Deﬁnition 9.19***. Suppose V and W are vector spaces. A function from the kfold
cross product to W g : V ×V ×· · · ×V →W is called alternating if it is multilinear and
if for all v
1
, · · · , v
k
∈ V , f(v
1
, · · · , v
k
) = 0 whenever v
i
= v
j
for some i = j.
There is a common equivalent deﬁnition which we give as a proposition.
Proposition 9.20***. Suppose V and W are vector spaces. A function g : V ×V ×· · · ×
V → W is alternating if and only if it is multilinear and it satisﬁes the following: for all
v
1
, · · · , v
k
∈ V and permutations σ ∈ Σ
k
,
(3***) f(v
1
, · · · , v
k
) = (−1)
sign σ
f(v
σ(1)
, · · · , v
σ(k)
).
Proof. Suppose f is an alternating function. We ﬁrst show that (3)*** holds for permu
tations. For any v
1
, · · · , v
k
∈ V and i < j we have that
f(v
1
, · · · , v
i
+ v
j
, · · · , v
j
+ v
i
, · · · , v
k
) =f(v
1
, · · · , v
i
, · · · , v
i
, · · · , v
k
)
+f(v
1
, · · · , v
j
, · · · , v
j
, · · · , v
k
)
+f(v
1
, · · · , v
i
, · · · , v
j
, · · · , v
k
)
+f(v
1
, · · · , v
j
, · · · , v
i
, · · · , v
k
)
since f is multilinear. The ﬁrst two terms on the right side are zero and the left side is
zero, because the function is alternating. Therefore
f(v
1
, · · · , v
i
, · · · , v
j
, · · · , v
k
) = (−1)
sign τ
f(v
τ(1)
, · · · , v
τ(i)
, · · · , v
τ(j)
, · · · , v
τ(k)
)
for τ the permutation (i, j). Now suppose that σ is an arbitrary permutation. Then σ can
be written as a composition of transpositions
σ = τ
m
· · · τ
1
where we have written this product as a composition of functions in functional notation
(not group multiplication). Thus we have that
f(v
1
, · · · , v
k
) =(−1)
sign τ
1
f(v
τ
1
(1)
, · · · , v
τ
1
(k)
)
=(−1)
sign τ
2
(−1)
sign τ
1
f(v
τ
2
τ
1
(1)
, · · · , v
τ
2
τ
1
(k)
)
.
.
.
=(−1)
sign τ
m
· · · (−1)
sign τ
2
(−1)
sign τ
1
f(v
τ
m
···τ
2
τ
1
(1)
, · · · , v
τ
m
···τ
2
τ
1
(k)
)
=(−1)
sign σ
f(v
σ(1)
, · · · , v
σ(k)
)
Therefore, (3)*** is satisﬁed.
The converse is easy to show. Suppose that f is a multilinear function which satisﬁes
(3)***. Suppose that v
1
, · · · , v
k
∈ V , and v
i
= v
j
for some i = j. Let τ be the transposition
τ = (i, j). Then v
τ(i)
= v
τ(j)
so, (v
1
, · · · , v
k
) = (v
τ(1)
, · · · , v
τ(k)
) and
f(v
1
, · · · , v
k
) =−1f(v
τ(1)
, · · · , v
τ(k)
)
=−f(v
1
, · · · , v
k
).
Therefore f(v
1
, · · · , v
k
) = 0 and f is alternating.
10 CHAPTER 9 MULTILINEAR ALGEBRA
Example 9.21***. The determinant function using either row or column vectors.
Suppose V is an mdimensional vector space and v
1
, · · · , v
m
are vectors in V . Let
det(v
1
, · · · , v
m
) be the determinant of the matrix whose i −th column vector is v
i
. Then
det : V ×· · · ×V →R
is an alternating map. The determinant is linear in each column vector and if two column
vectors are the same then the determinant vanishes.
We can do the same with row vectors. Let det
v
1
.
.
.
v
m
be the determinant of the matrix
whose ith row vector is v
i
. Then
det : V ×· · · ×V →R
is an alternating map. The determinant is linear in each row vector and if two row vectors
are the same then the determinant vanishes.
Example 9.22***. The cross product in R
3
and R
m
The cross product in R
3
is familiar as a map
−×− : R
3
⊗R
3
→R
3
from Calculus and Example 9.15***. It is multilinear and it is also alternating since
v × v = 0. Recall that the direction of v × w is perpendicular to the plane containing v
and w, and the direction is determined by the right hand rule.
Also recall from Calculus that the cross product v ×w can be computed as
det
e
1
e
2
e
3
v
w
which is formally interpreted as expansion in the ﬁrst row and e
1
, e
2
, e
3
is the standard
basis in R
3
often denoted i, j, k in Calculus and physics books. The vectors v and w are
row vectors. If u is another vector in R
3
, then u · (v ×w) = det
u
v
w
In R
m
we can deﬁne the cross product of m−1 vectors.
−×· · · ×− : R
m
⊗R
m
→R
m
.
The cross product v
1
×v
2
×· · · ×v
m−2
×v
m−1
can be formally computed as
det
e
1
· · · e
m
v
1
.
.
.
v
m−1
CHAPTER 9 MULTILINEAR ALGEBRA 11
by expansion along the ﬁrst row. If u is another vector in R
m
, then
u · (v
1
×v
2
×· · · ×v
m−2
×v
m−1
) = det
u
v
1
.
.
.
v
m−1
.
To check this formula, expand the right side along the ﬁrst row. Also notice that if
u ∈ span{v
1
, · · · , v
m−1
}, then the determinant vanishes. Hence v
1
×v
2
×· · · ×v
m−2
×v
m−1
is perpendicular to the m−1 dimensional subspace spanned by {v
1
, · · · , v
m−1
}.
The cross product is a multilinear and alternating function since the determinant is
multilinear and alternating.
Example 9.23***. v
m
· (v
1
×· · · ×v
m−1
) for v
1
, · · · , v
m
∈ R
m
This function on the mfold cross product of R
m
to R is multilinear and alternating
since it was shown to be the determinant
det
v
m
v
1
.
.
.
v
m−1
in the last example.
Example 9.24***. Signed Volume in R
3
Given three linearly independent vectors v
1
, v
2
, v
3
∈ R
3
, we may form a parallelepiped
with these three vectors as sides. The parallelepiped is {av
1
+ bv
2
+ cv
3
 a, b, c ∈ [0, 1]}.
We wish to compute the volume of the parallelepiped as a function of the three vectors
and obtain a multilinear function. Note that volume is always a positive number and the
value of a multilinear function must include negative values (since constants always pull
out). We will try to ﬁnd a notion of signed volume whose absolute value gives the actual
volume. Denote the signed volume of the parallelepiped by Vol
3
(v
1
, v
2
, v
3
). The volume
of a parallelepiped is the area of the base times the height, so
Vol
3
(v
1
, v
2
, v
3
) = Vol
2
(v
1
, v
2
)(v
3
· n)
where Vol
2
(v
1
, v
2
) is the area of the parallelogram determined by v
1
, v
2
, and n =
v
1
×v
2
v
1
×v
2

is
a unit vector perpendicular to the plane containing the parallelogram. The signed volume
is
Vol
3
(v
1
, v
2
, v
3
) = Vol
2
(v
1
, v
2
)(v
3
· n).
Note that the function is clearly linear in the third variable since the dot product is linear
in its ﬁrst variable. Permuting the variables keeps the correct formula up to the sign (so
you can move any variable to the third spot). To check the sign under a permutation,
one must observe that the cross product dotted with the last vector is alternating by the
previous example.
12 CHAPTER 9 MULTILINEAR ALGEBRA
Example 9.25***. Signed Volume in R
m
.
This example is essentially the same as Example 9.24***. Given m linearly independent
vectors v
1
, · · · , v
m
∈ R
m
, we may form a parallelepiped with these m vectors as sides. The
parallelepiped is {a
1
v
1
+· · · + a
m
v
m
 a
i
∈ [0, 1], i = 1, · · · m}. Again we wish to compute
the volume of the parallelepiped as a function of the m vectors. We show that there is a
multilinear alternating function that gives a signed volume. Denote the signed volume of
the parallelepiped by Vol
m
(v
1
, · · · , v
m
). The volume of a parallelepiped is the area of the
base times the height, so
Vol
m
(v
1
, · · · , v
m
) = Vol
m−1
(v
1
, · · · , v
m−1
)(v
m
· n)
where Vol
m−1
(v
1
, · · · , v
m−1
) is the m−1 dimensional volume of the parallelepiped deter
mined by v
1
, · · · , v
m−1
, and n =
v
1
×···×v
m−1
v
1
×···×v
m−1

is a unit vector perpendicular to the plane
containing the parallelepiped (Example 9.22***). The signed volume is
Vol
m
(v
1
, · · · , v
m
) = Vol
m−1
(v
1
, · · · , v
m−1
)(v
m
· n).
The linearity check is the same as in the last example. The function is clearly linear in the
last variable since the dot product is linear in its ﬁrst variable. Permuting the variables
keeps the correct formula up to the sign (so you can move any variable to the last spot).
To check the sign under a permutation, one must observe that the cross product dotted
with the last vector is alternating by Example 9.23***. Notice that if two of the entries in
Vol
m
(v
1
, · · · , v
m
) are the same then the parallelepiped is in an m−1 dimensional subspace
and so has zero volume.
Example 9.26***. Minors
Suppose that W is an mdimensional vector space, V is an ndimensional vector space,
f : W → V is linear, and det : V × · · · × V → R is the determinant in terms of column
vectors from the n fold product V . Then the map
F(w
1
, · · · , w
n
) = det(f(w
1
), · · · , f(w
n
)),
i.e., F = det◦(f×· · ·×f) from the n fold product of W to Ris an alternating map. Suppose
that W has e
1
, · · · , e
m
as an ordered basis and m ≥ n. Fix numbers i
1
< i
2
< · · · < i
n
in
{1, · · · , m}. Let p(
m
i=1
a
i
e
i
) =
n
j=1
a
i
j
e
i
j
, or p =
n
j=1
e
∗
i
j
. Let
M
i
1
,··· ,i
n
(w
1
, · · · , w
n
) = det ◦ (p ×· · · ×p)(w
1
, · · · , w
n
).
Now express the vector w
i
∈ W in terms of the basis: w
i
=
m
j=1
w
ji
e
j
. The function
M
i
1
,··· ,i
n
is the minor of the matrix
w
11
w
12
· · · w
1n
w
21
w
22
· · · w
2n
.
.
.
.
.
.
.
.
.
w
m1
w
m2
· · · w
mn
obtained by choosing the rows i
1
, · · · , i
n
, M
i
1
,··· ,i
n
(w
1
, · · · , w
n
) = det
w
i
1
1
· · · w
i
1
n
.
.
.
.
.
.
w
i
n
1
· · · w
i
n
n
.
We now ﬁnd the appropriate vector space in which to describe alternating multilinear
maps.
CHAPTER 9 MULTILINEAR ALGEBRA 13
Deﬁnition 9.27***. Suppose that V is a vector space and let A ⊂ V ⊗ · · · ⊗ V be the
subspace spanned by
{v
1
⊗· · · ⊗v
k
 v
1
, · · · , v
k
∈ V, v
i
= v
j
for some i = j}.
The vector space V ⊗ · · · ⊗ V/A along with the map ψ : V × · · · × V → V ⊗ · · · ⊗ V/A
is called the exterior product. The vector space V ⊗· · · ⊗ V/A is denote V ∧ · · · ∧ V and
ψ(v
1
, · · · , v
n
) is denoted v
1
∧ · · · ∧ v
n
.
Recall that there is a unique induced linear map
˜
ψ : V ⊗ · · · ⊗ V → V ∧ · · · ∧ V such
that
˜
ψ ◦ φ = ψ given by Proposition 9.6***.
Proposition 9.28***. The induce map ψ : V × · · · × V → V ∧ · · · ∧ V is a multilinear
alternating map.
Proof. The induced map is a composition of the multilinear map φ from Proposition 9.5***
and the linear quotient map
˜
ψ. Hence ψ =
˜
ψ ◦ φ is multilinear. The map is alternating
since φ(v
1
, · · · , v
n
) ∈ A if v
i
= v
j
for some i = j.
The main property of the exterior product is the universal mapping property for multi
linear alternating maps. It is stated in the following theorem.
Proposition 9.29***. Suppose V is vector a space. The exterior product ψ : V ×· · · ×
V → V ∧ · · · ∧ V satisﬁes the following property, the universal mapping property for
multilinear alternating maps:
If W is a vector space and g : V ×· · · ×V →W is an alternating multilinear map, then
there is a unique linear map g
: V
1
∧ · · · ∧ V →W such that g
◦ ψ = g.
Proof. There is a unique linear map ˜ g : V ⊗ · · · ⊗ V → W. The map g is alternating so
g(v
1
, · · · , v
k
) = 0 if v
i
= v
j
for some i = j. From Proposition 9.6***, g = ˜ g ◦ φ If v
i
= v
j
for some i = j, then
0 = ˜ g ◦ φ(v
1
, · · · , v
k
) = ˜ g(v
1
⊗· · · ⊗v
k
).
Therefore A ⊂ Ker ˜ g and the unique map ˜ g determines a unique map g
: V ⊗· · · ⊗V/A →
W.
The ability of the exterior product to convert alternating multilinear maps into linear
maps is an immediate consequence of Proposition 9.29***.
Theorem 9.30***. Suppose V and W are vector spaces. Linear maps ˜ g : V ∧· · ·∧V →W
are in one to one correspondence with alternating multilinear maps g : V ×· · · ×V →W.
Proof. Given a multilinear map g, Proposition 9.29*** produces the unique linear map g
.
Given a linear map g
let g = g
◦ ψ. The map g
is a composition of a linear map and an
alternating multilinear map, Proposition 9.28***. The composition of a linear map and
an alternating multilinear map is an multilinear alternating map. The reader should check
this fact.
14 CHAPTER 9 MULTILINEAR ALGEBRA
Theorem 9.31***. Suppose V are vector spaces and dim V = n. Let {e
i
 i = 1, · · · , n}
be a basis for V . Then the dimension of the kfold wedge product dimV ∧ · · · ∧ V =
n
k
and {e
j
1
∧ e
j
2
∧ · · · ∧ e
j
k
 1 ≤ j
1
< j
2
< · · · < j
k
≤ n} is a basis for the exterior product
V ∧ · · · ∧ V .
Proof. We ﬁrst show that dimV ∧ · · · ∧ V ≥
n
k
. Let W be the vector space with basis
{E
j
1
,··· ,j
k
 1 ≤ j
1
< j
2
< · · · < j
k
≤ n}. Let L : V × · · · × V → W be the map from the
kfold cross product of V to W deﬁned by
L(v
1
, · · · , v
k
) =
1≤j
1
<j
2
<···<j
k
≤n
M
j
1
,··· ,j
k
(v
1
, · · · , v
k
)E
j
1
,··· ,j
k
where M
j
1
,··· ,j
k
is the multilinear alternating function from Example 9.26***. The function
L is multilinear and alternating since each M
j
1
,··· ,j
k
is multilinear and alternating. The
function L hits a whole basis to W since L(e
j
1
, · · · , e
j
k
) = E
j
1
,··· ,j
k
. There is an induced
linear map L
: V ∧ · · · ∧ V → W which is onto since it hits a basis to W. Therefore
dimV ∧ · · · ∧ V ≥ dim W =
n
k
.
We next show that {e
j
1
∧ e
j
2
∧ · · · ∧ e
j
k
 1 ≤ j
1
< j
2
< · · · < j
k
≤ n} is a spanning set
(the same size as the dimension of W). Hence it is a basis. To observe it is a spanning set
ﬁrst note that {e
i
1
⊗· · ·⊗e
i
k
 i
1
, · · · , i
k
∈ {1, · · · , n}} is a spanning set for the kfold tensor
product V ⊗· · · ⊗V by Theorem 9.8***. Therefore {e
i
1
∧· · · ∧e
i
k
 i
1
, · · · , i
k
∈ {1, · · · , n}}
is a spanning set for the exterior product. Finally note that if σ is the permutation or
{1, · · · , k} so that i
σ(1)
< · · · < i
σ(k)
then e
i
1
∧· · ·∧e
i
k
= ±e
i
σ(1)
∧· · ·∧e
i
σ(k)
= ±e
j
1
∧· · ·∧e
j
k
where the last wedge product is in the desired set. See Exercise 2***
The reader should note that the proof of Theorem 9.31*** actually shows that v
1
∧· · · ∧
v
k
=
1≤j
1
<j
2
<···<j
k
≤n
M
j
1
,··· ,j
k
(v
1
, · · · , v
k
)e
j
1
∧ · · · ∧ e
j
k
.
We now establish the existence of induced maps and two properties of induced maps.
Theorem 9.32***. Suppose V and W are vector spaces and f : V → W is a linear
map. Then there is an induced linear map ∧
k
f : V ∧ · · · ∧ V → W ∧ · · · ∧ W deﬁned by
∧
k
f(v
1
∧ · · · ∧ v
k
) = f(v
1
) ∧ · · · ∧ f(v
k
). These induced maps satisfy the following two
properties
(1) If Q is another vector space and g : W →Q is another linear map, then ∧
k
g◦∧
k
f =
∧
k
(g ◦ f)
(2) If I
X
denotes the identity on X, then I
V
∧ · · · ∧ I
V
= I
V ∧···∧V
Proof. We show that ∧
k
f exists and is deﬁned by the formula in the statement of the
theorem. Let ψ : W × · · · × W → W ∧ · · · ∧ W be the multilinear alternating map in
from Deﬁnition 9.27*** an Lemma 9.28***. The function ψ ◦ (f ×· · · ×f) is a multilinear
alternating map from V ×· · · ×V to W ∧· · · ∧W since ψ is multilinear alternating and f
is linear. Therefore there is a unique induced linear map by Proposition 9.29***. Call the
map ∧
k
f : V ∧ · · · ∧ V →W ∧ · · · ∧ W. Since
ψ ◦ (f ×· · · ×f)(v
1
, · · · , v
k
) = ψ(f(v
1
), · · · , f(v
k
)) = f(v
1
) ∧ · · · ∧ f(v
k
),
we have that
∧
k
f(v
1
∧ · · · ∧ v
k
) = f(v
1
) ∧ · · · ∧ f(v
k
).
The two properties follow from the formula for ∧
k
f.
CHAPTER 9 MULTILINEAR ALGEBRA 15
Theorem 9.***. Suppose V is an n dimensional vector space, e
1
, · · · , e
n
is a basis for
V , and f : V → V is a linear map that has matrix A = (a
ij
) in the basis e
1
, · · · , e
n
. Let
c
1
, · · · , c
n
be the column vectors of A so that c
j
=
n
i=1
a
ij
e
i
. Then the induced linear
map ∧
k
f : V ∧ · · · ∧ V →V ∧ · · · ∧ V on the kfold wedge product is
∧
k
f(e
j
1
∧ · · · ∧ e
j
k
) =
1≤i
1
<···<i
k
≤n
M
j
1
,··· ,j
k
(c
i
1
, · · · , c
i
k
)e
i
1
∧ · · · ∧ e
i
k
where M
j
1
,··· ,j
k
(c
i
1
, · · · , c
i
k
) is the minor obtained from A by choosing the rows numbered
j
1
, · · · , j
k
and the columns numbered i
1
, · · · , i
k
, see Example 9.26***. The elements e
j
1
∧
· · · ∧ e
j
k
such that 1 ≤ j
1
< · · · < j
k
≤ n form a basis for the wedge product by Theorem
9.31***.
As a special case, if k = n, then ∧
n
f(e
1
∧ · · · ∧ e
n
) = det(f)e
1
∧ · · · ∧ e
n
.
Proof. The proof is a short computation.
∧
k
f(e
j
1
∧ · · · ∧ e
j
k
) =
n
i
1
=1
a
i
1
j
1
e
i
1
∧ · · · ∧
n
i
k
=1
a
i
k
j
k
e
i
k
=
n
j
1
,··· ,j
k
=1
a
i
1
j
1
· · · a
i
k
j
k
e
i
1
∧ · · · ∧ e
i
k
=
1≤j
1
<···<j
k
≤n
σ∈Σ
k
(−1)
sign σ
a
i
σ(1)
j
1
· · · a
i
σ(k)
j
k
e
i
1
∧ · · · ∧ e
i
k
The last line follows since the wedge is zero if the vectors are not distinct and by use of
Exercise 9.2***. To complete the proof, we observe that
M
j
1
,··· ,j
k
(c
i
1
, · · · , c
i
k
) =
σ∈Σ
k
(−1)
sign σ
a
i
σ(1)
j
1
· · · a
i
σ(k)
j
k
.
According to Theorem 9.30*** the multilinear alternating functions from V × · · · × V
to R can be interpreted as (V ∧ · · · ∧ V )
∗
(See Exercise 9.7***. We wish to construct
an isomorphism between ∧
k
(V
∗
) and (∧
k
V )
∗
analogous to the identiﬁcation for tensor
products. Unfortunately, there are two common choices in use. We discuss both. One is
used when the author is primarily interested in using both tensor and exterior products.
The other is used when the author is primarily interested in integration.
The notation used below is now given. We use ∧
k
V and ⊗
k
V for the k fold wedge and
tensor products. The symbols φ and ψ are used for the maps in Deﬁnitions 9.4*** and
9.27*** as well as the maps they induce. The subscript indicates to the reader the vector
used to produce the tensor or exterior product.
Isomorphism I. If we are required to be consistent with the maps already deﬁned, then
there would be no choice. We already have the following maps:
(∧
k
V )
∗
(ψ
V
)
∗
−−−−→ (⊗
k
V )
∗
(∧
k
V
∗
) ←−−−−
ψ
V
∗
(⊗
k
V
∗
)
16 CHAPTER 9 MULTILINEAR ALGEBRA
The top map is the interesting map to observe. Take
(ψ
V
)
∗
((e
i
1
∧ · · · ∧ e
i
k
)
∗
)(e
i
1
⊗· · · ⊗e
i
k
) = (e
j
1
∧ · · · ∧ e
j
k
)
∗
(ψ
V
(e
i
1
⊗· · · ⊗e
i
k
))
=
±1 if i
1
, · · · , i
k
is a permutation of j
1
, · · · , j
k
0 otherwise
since e
i
1
∧ · · · ∧ e
i
k
= (−1)
signτ
e
i
τ(1)
∧ · · · ∧ e
i
τ(k)
.
Hence, we have (φ
V
)
∗
((e
i
1
∧ · · · ∧ e
i
k
)
∗
) = (
σ∈Σ
k
(−1)
σ
e
i
σ(1)
⊗ · · · ⊗ e
i
σ(k)
)
∗
, and the
induced isomorphism (∧
k
V )
∗
→(∧
k
V
∗
) is
(e
i
1
∧ · · · ∧ e
i
k
)
∗
→(k!)e
∗
i
1
∧ · · · ∧ e
∗
i
k
Isomorphism II. The second map is convenient for area. The second map is
(e
i
1
∧ · · · ∧ e
i
k
)
∗
→e
∗
i
1
∧ · · · ∧ e
∗
i
k
Note that (
∂
∂x
∧
∂
∂y
)
∗
(
a
b
,
c
d
) is the area of the parallelogram determined by
a
b
and
c
d
. If we write (
∂
∂x
)
∗
= dx and (
∂
∂y
)
∗
= dy, then under the ﬁrst isomorphism
dx ∧ dy does not determine the area of the parallelogram but
1
2!
times the area, however,
the second pairing dx ∧ dy does determine the area. Since the measuring of area is the
central feature of integration, the second pairing lends itself to integration.
CHAPTER 9 MULTILINEAR ALGEBRA 17
Exercise 9.1***. Write out the details to demonstrate the validity of Formula (2)***
Exercise 9.2***. Show that A ⊂ V ⊗ · · · ⊗ V deﬁned in Deﬁnition 9.*** contains the
elements
v
1
⊗· · · ⊗v
k
−(−1)
sign σ
v
σ(1)
⊗· · · ⊗v
σ(k)
for all permutations σ ∈ Σ
k
. Show that v
1
∧ · · · ∧ v
k
= (−1)
sign σ
v
σ(1)
∧ · · · ∧ v
σ(k)
.
Exercise 9.3***. Suppose that V
1
, V
2
, and V
3
are vector spaces. Show that there are
welldeﬁned isomorphisms
f : V
1
⊗V
2
⊗V
3
→(V
1
⊗V
2
) ⊗V
3
such that f(u ⊗v ⊗w) = (u ⊗v) ⊗w and
g : V
1
⊗V
2
⊗V
3
→V
1
⊗(V
2
⊗V
3
)
such that g(u ⊗v ⊗w) = u ⊗(v ⊗w).
Exercise 9.4***. Suppose that V is a vector space. Show that there are welldeﬁned
isomorphisms
f : V ∧ V ∧ V →(V ∧ V ) ∧ V
such that f(u ∧ v ∧ w) = (u ∧ v) ∧ w and
g : V ∧ V ∧ V →V ∧ (V ∧ V )
such that g(u ∧ v ∧ w) = u ∧ (v ∧ w).
Exercise 9.5***. Suppose that f : R
3
→R
3
is represented by the matrix
2 1 3
−1 0 1
0 4 0
in the ordered basis e
1
, e
2
, e
3
and that g : R
3
→ R
3
is represented by the matrix
1 2 3
4 0 −2
1 0 0
in the same ordered basis. Find the matrix for f ⊗ g in the ordered
basis e
1
⊗e
1
, e
2
⊗e
1
, e
3
⊗e
1
, e
1
⊗e
2
, e
2
⊗e
2
, e
3
⊗e
2
, e
1
⊗e
3
, e
2
⊗e
3
, e
3
⊗e
3
.
Exercise 9.6***. Suppose that V
1
, · · · , V
k
are vector spaces and let Mult(V
1
×· · ·×V
k
, R)
denote the multilinear functions f : V
1
×· · · ×V
k
→R. Show that Mult(V
1
×· · · ×V
k
, R)
is a vector space under the usual addition an multiplication in R. Show that there is a
canonical isomorphism
Mult(V
1
×· · · ×V
k
, R)
∼
= (V
1
⊗· · · ⊗V
k
)
∗
.
18 CHAPTER 9 MULTILINEAR ALGEBRA
Exercise 9.7***.
a. Suppose that V is a vector space and let Alt(V × · · · × V, R) denote the k variable
multilinear alternating functions f : V × · · · × V → R. Show that Alt(V × · · · × V, R)
is a vector space under the usual addition an multiplication in R. Show that there is a
canonical isomorphism
Alt(V ×· · · ×V, R)
∼
= (V ∧ · · · ∧ V )
∗
.
b. If dimV = n, then show that any multilinear alternating function in nvariables
f : V ×· · · ×V →R is a multiple of det. Show that
Vol
n
(v
1
, · · · , v
n
) = det(v
1
, · · · , v
n
).
Exercise 9.8***. Suppose that V
1
, · · · , V
k
and Q are vector spaces and ρ : V
1
×· · ·×V
k
→
Q is a multilinear function. Show that if ρ : V
1
× · · · × V
k
→ Q satisﬁes the universal
property for multilinear functions (Proposition 9.6***), then there is a unique isomorphism
ι : V
1
⊗· · · ⊗V
k
→Q such that ι ◦ ρ = φ.
Exercise 9.8***. Suppose that V and Q are vector spaces and ρ : V × · · · × V → Q
is a multilinear function in k variables. Show that if ρ : V × · · · × V → Q satisﬁes the
universal property for altermating multilinear functions (Proposition 9.29***), then there
is a unique isomorphism ι : V ∧ · · · ∧ V →Q such that ι ◦ ρ = ψ.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.