Foundations in Analysis and Geometry on

Manifolds (V3D3/F4D1)

Matthias Lesch
Mathematisches Institut, Universität Bonn, Endenicher Allee 60,
53115 Bonn, Germany
Email address: [email protected]
URL: www.math.uni-bonn.de/people/lesch
Abstract. Winter Semester 2022/2023
Predecessors: Global Analysis 2004-2016, c Matthias Lesch
 Contents

List of Figures 5
Preface 6
Chapter 1. Differentiable manifolds 7
1. From submanifolds to differentiable manifolds 7
1.1. Submanifolds 7
1.2. Differentiable manifolds 7
1.3. Examples of manifolds 9
1.4. Smooth Maps 10
2. Partition of unity 10
2.1. Paracompactness 10
2.2. Bump functions and partition of unity 11
Addendum 12
3. The tangent space 13
3.1. Motivation and Definition 13
3.2. Simple properties 14
3.3. Dimension 15
3.4. The physicists definition of tangent vectors 17
3.5. Curves 17
4. Local properties of differentiable maps, submanifolds 18
5. The Morse-Sard Theorem 19
6. Applications of Sard’s Theorem 22
6.1. Embedding 22
6.2. Transversality 24
Appendix A. 25
1. Structure theorems for differentiable maps 25
1.1. Implicit Function Theorem 25
1.2. Inverse Function Theorem 25
1.3. Rank Theorem 25
2. Submanifolds of Rn 26
2.1. Immersion Theorem 26
Bibliography 27

3
 List of Figures

1 The separation axiom for topological spaces. 7

2 Simple equivalence relation on R leading to a non-separated “manifold”. 8
3 The function f1 (x) = e−1/x for x > 0 and 0 for x ≤ 0. 11
f1 (x)
4 The function f2 (x) = f1 (x)+f1 (1−x)
. 12
5 The function f3 (x) = f2 (2 + x) · f2 (2 − x). 12
6 Given a closed set A ⊂ M and an open neighborhood G ⊃ A there is a
smooth function which is identically 1 on A and identically 0 outside G. 14

5
6 LIST OF FIGURES

Preface
Here goes the preface.
 CHAPTER 1

Differentiable manifolds
Lecture 01,
1. From submanifolds to differentiable manifolds 12.10.2022,
Lennart
1.1. Submanifolds. We recall from Analysis II/III the notion of a differen- Ronge
tiable submanifold of Rn :
M ⊂ Rn is called a differentiable
 submanifold if it is locally of
the form M ∩ U = F = 0 , where F : U −→ Rn−k is a C∞ –
submersion.
Recall that a C∞ –submersion F : U −→ Rn−k is a C∞ –map such that at every
point p ∈ U the derivative DF(p) has rank n − k.
Our first goal is to define manifolds independently of an embedding into an
ambient Rn .

1.2. Differentiable manifolds.

Definition 1. 1. A topological space X is called separated (resp. a Hausdorff space),
if for each pair of distinct points x, y ∈ X, x 6= y there exist open disjoint neighborhoods
Ux 3 x, Uy 3 y (i. e. Ux ∩ Uy = ∅). Cf. Figure 1.
2. X is said to have a countable basis if there exists a countable family (Un )n∈N of
open subsets with the following property: given x ∈ X and an open neighborhood U of x
there exists an n ∈ N such that x ∈ Un ⊂ U.

Definition 2. Let X be a separated topological space.

1. An n–dimensional chart is a pair (U, ϕ) consisting of an open subset U ⊂ X and
a homeomorphism ϕ : U −→ V ⊂ Rn onto an open subset V ∈ R .
n

2. A family of charts A = (Uα , ϕα )  α ∈ I is called an atlas, if X = α∈I Uα .

S

3. The atlas A is called C∞ (Ck , analytic, continuous, algebraic, linear,. . . ), if ϕα ◦

∞
ϕ−1
β is C (Ck , analytic, . . . ) for all α, β ∈ I. Note that ϕα ◦ ϕ−1
β maps ϕβ (Uα ∩ Uβ )
onto ϕα (Uα ∩ Uβ ).
4. A chart (U, ϕ) is C∞ (Ck , . . . )–compatible with the atlas A if ϕα ◦ ϕ−1 and
∞
ϕ ◦ ϕ−1α are C (. . . ) for all α ∈ I.

Ux ∩ Uy = ∅

Ux x• y• Uy

Figure 1. The separation axiom for topological spaces.

7
8 1. DIFFERENTIABLE MANIFOLDS

∞ 1
0
−1 0 1 −∞ −1

Figure 2. Simple equivalence relation on R leading to a non-

separated “manifold”.

Amax = (U, ϕ)  (U, ϕ) compatible with A is the maximal atlas associated to


A.
5. A C∞ (Ck , . . . ) manifold is a pair (M, A) consisting of a separated topological
space M with a countable basis and a maximal C∞ (Ck , . . . ) atlas A. We use the word
“smooth” as synonym for C∞ .
Remark 3. 1. To determine a C∞ –structure it suffices in view of Def. 2, 4., to
provide some (as small as possible) atlas.
2. Whitney showed that every Ck –structure, k ≥ 1, already contains a C∞ –
structure. Therefore we only consider smooth (C∞ –)manifolds.
3. The case C0 is more complicated, though from the point of view of differ-
ential topology it is not that interesting.
There exist C0 –manifolds which do not admit a differentiable structure, e. g.
in dimension 4.
4. A C0 –manifold may admit several non-equivalent C∞ –structures, i. e. there
exist C∞ –manifolds which are homeomorphic but not diffeomorphic.
For example a famous result due to Milnor says that the 7–sphere S7 admits
exactly 28 different C∞ –structures.
5. The assumption that a manifold should have a countable basis is made for
convenience. It guaranties paracompactness and the existence of a partition of
unity (see Section 2 below). “Manifolds” which do not have a countable basis
can only be constructed using set theoretic tools like Zorn’s Lemma. Therefore,
from a pragmatic point of view they are not of much use.
6. The condition that a manifold should be separated excludes pathologies,
as illustrated by the next example.

Example 4 (cf. Figure 2). On the one-dimensional manifold R consider the

equivalence relation
x ∼ y ⇐⇒ x = y or |x| = |y| > 1,
and let X = R/ ∼ be the quotient space with the quotient topology. Equivalence
classes are denoted by [·].
1 : (−∞, 1) → X, x 7→ [x] and U2 =
We define the charts U1 = (−∞, 1)/ ∼, ϕ−1
(−1, ∞)/ ∼, ϕ2 : (−1, ∞) → X, x 7→ [x]. The ϕj , j = 1, 2, are homeomorphisms,
−1

X \ U1 = {[1]}, X \ U2 = {[−1]} and thus X = U1 ∪ U2 . Furthermore, U1 ∩ U2 =

X \ {[±1]} and

x, |x| < 1,
1 : (−∞, 1) \ {−1} −→ (−1, ∞) \ {1},
ϕ2 ◦ ϕ−1 x 7→
−x, |x| > 1,
 1. FROM SUBMANIFOLDS TO DIFFERENTIABLE MANIFOLDS 9
 
showing that (Uj , ϕj )  j = 1, 2 is a smooth atlas. However, X is not separated
 points {[±1]}
since the two cannot be separated
∞
by disjoint open neighborhoods.
In sum {(Uj , ϕj )} j = 1, 2 is a C –atlas of the non-separated manifold X.


Some authors deal with non-separated manifolds. We do not want to deal with
such pathologies and therefore in this text X is not considered to be a manifold.
partly in Lec-
1.3. Examples of manifolds. 1. M = Rn , A = (Rn , id). Thus we have the ture 02,
reassuring fact that the Euclidean standard space with the atlas consisting of the 17.10.2022
only chart id is a smooth manifold.
2. Every C∞ –submanifold of Rn (in the sense of Analysis II) with the at-
las consisting of all pairs (ϕ(U), ϕ−1 ), where ϕ : U −→ M runs through the
parametrizations of M, is a C∞ –manifold.
3. Projective spaces. Let K = R or C and denote by KPn the set of all lines in
K , that is KPn = (Kn+1 \ {0})/ ∼ is the quotient with respect to the equivalence
n+1

relation
x ∼ y : ⇐⇒ x = λ · y for some λ ∈ K \ {0}
⇐⇒ x, y linearly dependent over K.
Furthermore, since every line contains a point on the sphere we have
RPn = Sn /(x ∼ −x) = Sn /(Z/2Z), CPn = S2n+1 / ∼ = S2n+1 /S1 .
This shows that the sphere is a two-fold covering of the real projective space and
that the complex projective space is the homogeneous space obtained from the
natural S1 action on odd spheres which sit inside Cn+1 .
Points in KPn are usually written in the form [x] = [x0 : . . . : xn ] with
(x0 , . . . , xn ) ∈ Kn+1 \ {0}. These are called homogeneous coordinates since they
are unique only up to multiplication by a nonzero number in K. We define a
smooth atlas for KPn as follows:
 
Uj := [x0 : . . . : xn ] ∈ KPn  xj 6= 0 ,
xi

ϕj : Uj −→ Kn , [x0 : . . . : xn ] 7→ .
xj i=0,...,n; i6=j
ϕj is certainly a homeomorphism. We compute the change of coordinates,
w. l. o. g. let 0 ≤ i < j ≤ n. Then
 
ϕi (Ui ∩ Uj ) = y ∈ Kn  yj 6= 0 ,
and
ϕj ◦ϕ−1
i (y1 , . . . , yn )


= ϕj [y1 : . . . : yi−1 : 1 : yi : . . . : yj : . . . : yn ]
y yi−1 1 yi yj−1 yj+1 yn 
1
= ,..., , , ,..., , ,..., .
yj yj yj yj yj yj yj
 
Thus (Uj , ϕj )  0 ≤ j ≤ n is even a rational (algebraic) atlas of KPn .
4. Abstract tori. Let ω1 , . . . , ωn ∈ Rn be an R–basis. Then ÜÜ
Γ := Z · ω1 + · · · + Z · ωn
 10 1. DIFFERENTIABLE MANIFOLDS

is called a lattice in Rn . Construct a smooth atlas for the quotient space Rn /Γ and
show that it is diffeomorphic to the n–torus T n := S1 × . . . × S1 (n factors).
cn = Rn ∪ {∞}. Topologically this is the Alexandroff(=one-point) compact-
5. R
ification of Rn . We define an atlas consisting of two charts:
cn \ {∞} = Rn , ϕ∞ (x) = x,
U∞ = R

cn \ {0}, ϕ0 = |x|2 , x 6= ∞,
x
U0 = R
0, x = ∞.
We have U0 ∩ U∞ = Rn \ {0} and ϕ0 ◦ ϕ−1
∞ (x) =
x
|x|2
, which is certainly smooth.

∼ Sn (diffeomorphism).
cn =
Exercise 5. Show that R
1.4. Smooth Maps. In the sequel we denote (C∞ )–manifolds just by one let-
ter, e. g. M. An atlas is always tacitly assumed to be given.
Definition 6. f : M −→ N is called differentiable, if for each p ∈ M there exist charts
(Up , ϕp ), (Uq , ϕq ) 1 , q = f(p), such that ϕq ◦ f ◦ ϕ−1
p is differentiable in a neighborhood
of ϕp (p). Note that ϕq ◦ f ◦ ϕp is defined on the (open) set ϕp (f−1 (Uq ) ∩ Up ).
−1

If the condition of the definition is satisfied then for arbitrary charts (U, ϕ)
Lecture 02, of M and (V, ψ) of N the composition ψ ◦ ϕ−1 is differentiable on ϕ(f−1 (V) ∩ U).
17.10.2022
Remark 7. 1. In Def. 6 the requirement that ϕq ◦ f ◦ ϕ−1 p is defined in a neigh-
borhood of ϕp (p) ensures that smooth maps are continuous; hence f−1 (Uq ) ∩ Up
must indeed be open.
2. The composition of differentiable maps is differentiable.
3. f ∈ C∞ (M, N) is called a diffeomorphism if f is bijective and f−1 ∈ C∞ (N, M).
By the Inverse Function Theorem this is equivalent to the fact that f is bi-
jective and for all p ∈ M its derivative ( cf. Section 3 below) Dp f is invertible.
Charts are important examples of diffeomorphisms.

2. Partition of unity
In this section we discuss in detail the consequences of the assumption that
manifolds are second countable (i. e. have a countable basis).

2.1. Paracompactness.
Definition 8 (Paracompact). A topological space X is called paracompact, if each
open covering (Ui )i∈I of X has a locally-finite refinement (Vj )j∈J .
(Vj )j∈J is a refinement of Ui , if for each j ∈ J there exists an i ∈ I such that Vj ⊂ Ui .
(Vj )j∈Jis locally-finite,
 each x ∈ X there exists an open neighborhood U 3 x
if for
such that j ∈ J U ∩ Uj 6= ∅ is finite.


Proposition 9. Let (M, A) be a C∞ –manifold in the sense of Definition 2. Then M is

paracompact.
1
This notation implies p ∈ Up .
 2. PARTITION OF UNITY 11

1 2 3

Figure 3. The function f1 (x) = e−1/x for x > 0 and 0 for x ≤ 0.

In more detail, if (Uα )α∈A is an open covering then there exists a (at most) countable
locally-finite refinement (Vβ )β∈B . Moreover, Vβ can be chosen in such a way that there
exist charts 
ψβ : Vβ −→ B(0, 3) = x ∈ Rn  |x| < 3


with [
ψ−1


M= β B(0, 1) .
β∈B

Proof. M is locally compact and second countable. Thus there exists an ex-
haustion by compacta K1 ⊂ K̊2 ⊂ K2 ⊂ K̊3 ⊂ . . . such that M = ∞
S
j=1 Kj .
Kj+1 \ K̊j is compact and hence for p ∈ Kj+1 r K̊j there exists a chart (Vp , ϕp )
with ϕp (Vp ) = B(0, 3), Vp ⊂ K̊j+2 \ Kj−1 , and Vp ⊂ Uα for some α ∈ A.
By compactness there exist finitely many points pj1 , . . . , pjrj , such that the
compact set
rj
[
Kj+1 \ K̊j ⊂ ϕ−1
jl (B(0, 1)), ϕjl := ϕpjl .
l=1

Since Vp ⊂ K̊j+2 \ Kj−1 the family Vjl , ϕjl

j=1,2,... is a countable locally-finite re-
1≤l≤rj
finement of (Uα )α∈A . 
2.2. Bump functions and partition of unity. We first construct a few smooth
helper functions.

e−1/t , t > 0, f1 (t)
f1 (t) := f2 (t) := , (2.1)
0, t ≤ 0, f1 (t) + f1 (1 − t)
f3 (t) := f2 (2 + t)f2 (2 − t), f4 (x) := f3 (|x|), x ∈ Rn . (2.2)
One learns in Analysis I that f1 (Figure 3) and hence f2 , f3 , f4 are indeed smooth. Lecture 03,
19.10.2022
Proposition 10 (Partition of Unity). Let M be a C∞ –manifold and (Uα )α∈A an open
covering. Then there exist ϕn ∈ C∞ (M), n ∈ N, with the following properties:
(1) 0 ≤ ϕn ≤ 1,
(2) (supp ϕn )n is locally-finite,
(3) For each n the support of ϕn is contained in some Uα ,
P∞
(4) ϕn (p) = 1 for all p ∈ M.
n=1
Because of (3) (ϕn ) is called a partition of unity subordinated to the open covering.
12 1. DIFFERENTIABLE MANIFOLDS

0
1

f1 (x)
Figure 4. The function f2 (x) = f1 (x)+f1 (1−x)
.

−2 −1 1 2

Figure 5. The function f3 (x) = f2 (2 + x) · f2 (2 − x).

Addendum. If A ⊂ N then A can be chosen as index set for the partition. More
precisely, one finds (ϕα )α∈A satisfying (1), (2), (4) and (3’): supp(ϕα ) ⊂ Uα .

Proof. Let Vn , ψn : Vn −→ B(0, 3) as in Prop. 9. Put


f4 (ψn (x)), x ∈ Vn ,
ϕ(x)
e :=
0, x∈/ Vn ,

and
X
∞
ϕ
e n ∈ C∞ (M),
fn
ϕ
e := ϕ ϕn := .
n=1
ϕ
e

Since ψn−1 (B(0, 1)) covers M the function ϕ e is everywhere positive. By the locally-
finiteness of the Vn the sum is also locally-finite and hence ϕ e is indeed smooth.
The family (ϕn ) therefore satisfies (1)–(4). Note that the functions ϕn are by
construction even compactly supported.
Now consider A ⊂ N and let the ϕn be given as above. First, we enlarge the
index set of the definition of the covering (Uα ) by putting Uk := ∅ if k ∈ N \ A.
Then define inductively
 
J0 := i ∈ N  supp ϕi ⊂ U0 ,
 
Jk := i ∈ N  i 6∈ J0 ∪ . . . ∪ Jk−1 , supp ϕi ⊂ Uk .
 3. THE TANGENT SPACE 13
S
Then, Ji ∩ Jj = ∅ for i 6= j and N = i∈N Ji is a partition of the natural numbers
and we may put
X
χk := ϕi .
i∈Jk

By construction χk = 0 if k 6∈ A and by locally-finiteness of the supports of the

ϕn the functions χk are smooth. Furthermore, we have
(1) 0 ≤ χα ≤ 1,
(2) (supp
P χα )α∈A is locally-finite,
(4) χα (p) = 1 for all p ∈ M.
α∈A
It remains to show that supp χα ⊂ Uα . It is a priori not obvious that supp χα =
∪i∈Jα supp ϕi as the union might be infinite and an infinite union of closed sets
is not necessarily closed. Here, this is however the case as the family (supp ϕi )i
is locally finite. We prove the a priori weaker but sufficient statement supp χα ⊂
Uα :
To this end let p 6∈ Uα . Since (supp ϕi )i is locally-finite we may choose an
open neighborhood U 3 p such that
 
B = i ∈ N  U ∩ supp ϕi 6= ∅
is finite. Pick i ∈ B. If supp ϕi ⊂ Uα then since, p 6∈ supp ϕi we find a smaller
open neighborhood U 0 3 p such that U ∩ supp ϕi = ∅ and hence be may re-
move i from B. Since B is finite after finitely many steps we arrive at an open
neighborhood U e 3 p such that Ue ∩ supp ϕi 6= ∅ implies supp ϕi 6⊂ Uα . But this
means that χα vanishes on U and hence p 6∈ supp χα . We have thus shown that
supp χα ⊂ Uα and we are done. 
Remark 11. Note that the functions ϕn constructed in the proof of Proposition 10
are by construction even compactly supported. In the situation of the addendum,
however, this is not always true. The functions χα are supported in Uα but in
general the support will just be a closed set.
Proposition 12. Let A ⊂ M be a closed subset and G ⊂ M open with A ⊂ G. Then
there exists f ∈ C∞ (M) with 0 ≤ f ≤ 1 and

1, p ∈ A,
f(p) =
0, p ∈
/ G.
See Figure 6.
Proof. {M \ A, G} is a finite open covering. Let {ϕ0 , ϕ1 } be a subordinated
partition of unity. Then f = ϕ1 does the job. 

3. The tangent space

3.1. Motivation and Definition. Let F : Rn −→ R be a submersion and M =
{F = 0}. Then the tangent space to M at p is given by

⊥
Tp M = γ 0 (0) ∈ Rn  γ : I → M ⊂ Rn curve , γ(0) = p = grad F(p) .

14 1. DIFFERENTIABLE MANIFOLDS

A
 

Figure 6. Given a closed set A ⊂ M and an open neighborhood

G ⊃ A there is a smooth function which is identically 1 on A and
identically 0 outside G.

Here, by slight abuse of language under a “curve”, unless otherwise said, we will
understand a smooth curve. This definition of tangent space uses the embedding
M ,→ Rn and hence it cannot be used for general manifolds.
However, every v ∈ Tp M defines a directional derivative v : C∞ (M) −→ R as
follows: choose a curve γ : I −→ M with γ(0) = p, γ 0 (0) = v, and put
d 
vf :=  f(γ(t)).
dt t=0
vf depends only on v (i.e. on γ(0) and γ 0 (0) but not any further on γ). Moreover,
f 7→ vf is a derivation, that is it is linear and it satisfies Leibniz’ rule
v(f · g) = v(f) · g(p) + f(p) · v(g).
Definition 13. Let M be a C∞ –manifold. A linear map Xp : C∞ (M) −→ R is called a
derivation at p if it satisfies Leibniz’ rule
Xp (f · g) = f(p) · Xp g + g(p) · Xp f.
The tangent space at p, Tp M, is by definition the vector space of all derivations at p.
Note that a priori Tp M is a subset of the dual space C∞ (M)∗ of C∞ (M) and
since linear combinations of derivations are again derivations, Tp M is indeed a
subspace of C∞ (M)∗ .

3.2. Simple properties.

3.2.1. Let ϕ ∈ C∞ (M) be a smooth function which is constant in a neigh-
borhood of p. Then Xp ϕ = 0 for every Xp ∈ Tp M.
Proof. W. l. o. g. let ϕ = 1 in a neighborhood of p. Using Prop. 12 we may
choose a χ ∈ C∞ (M)
 which is constant 1 in a neighborhood of p and such that
supp χ ⊂ q ∈ M ϕ(q) = 1 . Then χ · ϕ = χ and hence


Xp (χ) = Xp (χ · ϕ) = ϕ(p) · Xp (χ) + χ(p) · Xp (ϕ) = Xp (χ) + Xp (ϕ),

thus Xp ϕ = 0. 
3.2.2. Tp M ∩ Tq M = {0} if p 6= q. Thus, if X ∈ Tp M ∩ Tq M then X = 0
or p = q. In other words each nonzero tangent vector X 6= 0 determines its
basepoint uniquely.
 3. THE TANGENT SPACE 15

Proof. Let X ∈ Tp M ∩ Tq M and p 6= q. Let f ∈ C∞ (M) be given. Then choose

ϕ ∈ C∞ (M) such that ϕ ≡ 1 in a neighborhood of p and ϕ ≡ 0 in a neighborhood
of q. Then by 3.2.1 we have Xϕ = Xq ϕ = 0. Furthermore, f(1 − ϕ) ≡ 0 in a
neighborhood of p and hence by 3.2.1 we have Xp (f(1 − ϕ)) = 0. Thus we infer
X(f) = Xp (f) = Xp (fϕ) = Xq (fϕ) = f(q)Xq ϕ + ϕ(q)Xq f = 0. 
3.2.3. Although a priori a derivation X ∈ Tp M is defined only on C∞ (M) it
can be defined on C∞ (U) for any open neighborhood U of p as follows: given
f ∈ C∞ (U) choose fe ∈ C∞ (M) with f ≡ fe in a neighborhood of p. Then put
e Using the previous observations it is clear that Xf is well-defined and
Xf := Xf.
that the so defined X satisfies Leibniz’ rule on C∞ (U) as well.
3.2.4. Let (U, ϕ) be a chart with coordinate functions x1 , . . . , xm . Then
∂  ∂ 
f ◦ ϕ−1 (r) = D(f ◦ ϕ−1 )(ϕ(p))[ei ].


 f :=
∂xi p ∂ri ϕ(p)


Here, ei denotes the i-th canonical base vector in Rm .

3.3. Dimension.
Lemma 14. For p ∈ M there exists a chart (U, ϕ) centered2 at p with the following
property: For each f ∈ C∞ (U) there exist functions f1 , . . . , fm ∈ C∞ (U) such that
X
m
f = f(p) + fj · xj .
j=1

Here, x1 , . . . , xm are the coordinate functions of the chart, m is the dimension of the
chart.
Proof. Let first (U, ϕ) be a chart centered at p with range (−ε, ε)m . Then in
terms of the coordinates we have
X
m
f(x) = f(p) + f(x1 , . . . , xj , 0, . . . , 0) − f(x1 , . . . , xj−1 , 0, . . . , 0)
j=1
m Z1
X
= f(p) + (∂j f)(x1 , . . . , xj−1 , txj , 0, . . . , 0)dt · xj ,
j=1 0

R1
so fj (x) := 0 (∂j f)(x1 , . . . , xj−1 , txj , 0, . . . , 0)dt does the job. Lecture 04,
If (U, ϕ) is a general chart centered at p then there exists an ε > 0 and an 24.10.2022

open subset U ⊃ V 3 p with V ⊂ U and ϕ(V) = (−ε, ε)m . Hence by the first part
of this proof we have
 Xm
fV = f(p) + fej · xj ,
j=1

with fej ∈ C∞ (V). Now choose an open neighborhood W of p with W ⊂ V. By

∞

Proposition 12 there exists χ ∈ C (M) with χ W = 1 and χM\V = 0. Then fej χ


2
“centered at p” means ϕ(p) = 0.
 16 1. DIFFERENTIABLE MANIFOLDS

extends naturally by 0 to a smooth function on M and with these functions we

have

f = fχ + f(1 − χ)
Xm 
(f − f(p))(1 − χ)xj 
= f(p) + fej χ + Pm 2 · xj
j=1 j=1 xj

f(1−χ)xj
Note that since 1 − χ vanishes in a neighborhood of p the function Pm 2 is
j=1 xj
smooth in a neighborhood of p. 

Proposition 15. Let M be a C∞ –manifold and (U, ϕ) an m–dimensional chart centered

at p. Then dim Tp M = m. Moreover,
∂  ∂ 
 ,...,
∂x1 p ∂xm p


is a basis of Tp M.

Remark 16. Because of this proposition we call dim Tp M the dimension of M at p.

Clearly, the map M 3 p 7→ dim Tp M ∈ Z is a locally constant function on M. If
M is connected then it is constant and hence it will be called the dimension of M.
In the sequel we will mostly consider manifolds of “pure dimension”, notation
Mm . This is not a serious loss of generality since every manifold is a disjoint
union of connected components of pure dimension.
Formally, the manifold definition also makes sense in dimension 0. R0 = {0}
is the one-point space and hence a 0–dimensional manifold is nothing but an at
most countable discrete set.

Proof. Let V ⊂ U be a neighborhood of p with ϕ(V) = (−ε, ε)m . Then by the

P e
proof of the previous Lemma we have for X ∈ Tp M and f = mj=1 fj · xj + f(p) ∈
C∞ (V), since xj (p) = 0,
X
m X
m

∂ 
Xf = fj (p)Xxj = Xxj  f.
j=1 j=1
∂xj p
Pm

∂ 

∂ 

∂ 
This shows that X = j=1 Xxj ∂xj 
. Since ∂xj 
xi = δij the vectors ∂xj 
,j =
p p p
1, . . . , m, are linearly independent and the claim is proved. 

Examples 17. 1. Rn . At p ∈ Rn the usual partial derivatives ∂x∂ j  form a canonical

p
basis of Tp Rn .
2. Polar coordinates in R3 . Consider
Φ : (0, ∞) × (0, 2π) × (0, π) −→ R3 ,
(r, ϕ, ϑ) 7→ (r cos ϕ sin ϑ, r sin ϕ sin ϑ, r cos ϑ).

ÜÜ Φ is a diffeomorphism onto an open dense subset of R3 (which one? ). For r = 1

Φ(1, . . . , . . . ) is a parametrization of S2 up to a meridian.
 3. THE TANGENT SPACE 17

The chain rule yields

∂  ∂ 
 f = f(r cos ϕ sin ϑ, . . . , . . . )

∂r p ∂r r(p),ϕ(p),ϑ(p)
∂ ∂  ∂ 
= cos ϕ sin ϑ  f + sin ϕ sin ϑ  f + cos ϑ  f.
∂x p ∂y p ∂z p
We leave the calculation of the partial derivatives with respect to ϕ, ϑ as an
exercise. ÜÜ
n
3. In general let Φ = (y1 , . . . , yn ) be a chart of an open subset U of R . Then
∂  ∂ X n
∂Φj
−1
 f= (f ◦ Φ ◦ Φ)(p) = ∂j (f ◦ Φ−1 )(Φ(p)) (p)
∂xi p ∂xi j=1
∂xi

X
n
∂yj ∂ 
= (p)  f.
j=1
∂xi ∂yj p

Definition 18 (Tangent map). Let M and N be C∞ –manifolds, f ∈ C∞ (M, N), p ∈

M. The tangent map of f at p is linear map defined by
Tp f = Dp f : Tp M −→ Tf(p) N,
Tp f(X) · ϕ = X(ϕ ◦ f), ϕ ∈ C∞ (N).
Remark 19. 1. You should check for yourself that for open subsets M ⊂ Rm , N ⊂
Rn , Tp f is in the standard basis just given by the Jacobi matrix of f.
2. One has the chain rule Tp (f ◦ g) = Tg(p) f · Tp g.
Each chart (U, ϕ) yields an isomorphism
Tp ϕ : Tp M −→ Tϕ(p) Rn = ∼ Rn .

If ψ is another chart then

Tp ψ = Tp (ψ ◦ ϕ−1 ◦ ϕ) = Dϕ(p) (ψ ◦ ϕ−1 )Tp ϕ.
3.4. The physicists definition of tangent vectors. For a physicist a vector is
given by a family of coordinate vectors in Rm (the coordinate representations of
the vectors) together with a transformation rule. Accordingly, a tangent vector to
p ∈ M is a family (ξϕ )ϕ , where ξϕ ∈ Rm and ϕ runs through all charts centered
at p, with the transformation rule
ξψ = D0=ϕ(p) (ψ ◦ ϕ−1 )[ξϕ ].
Clearly, in our language, for v ∈ Tp M the vector ξϕ is nothing but Tp ϕ[v].
3.5. Curves. Let I ∈ R be an interval. A curve in Mm is a C∞ –map c : I −→
M.
d 

ċ(t0 ) := Tt0 c  ∈ Tp M.
dt t0
is called the velocity vector of c.
Each v ∈ Tp M is the velocity vector
Pm of a∂ curve
 through p: to see this let ϕ be
a chart which is centered at p, v = j=1 vj ∂xj p (see Prop. 15), and


c(t) := ϕ−1 (tv1 , . . . , tvm ).

 18 1. DIFFERENTIABLE MANIFOLDS

For f ∈ C∞ (M) we have

d 
d  Xm
∂(f ◦ ϕ−1 )
ċ(0)f = T0 c  f =  f(c(t)) = vj (0) = vf,
dt 0 dt t=0 j=1
∂x j

thus ċ(0) = v. Furthermore, by Definition 18 we have for h ∈ C∞ (Rm ) that

X
m
∂ (h ◦ ϕ ◦ ϕ−1 ) X ∂h
m
Tp ϕ(v)h = v(h ◦ ϕ) = vj (0) = vj (0),
j=1
∂xj j=1
∂x j

proving Tp ϕ(v) = (v1 , . . . , vm )t .

4. Local properties of differentiable maps, submanifolds

In the sequel, we will consider only manifolds of “pure” dimension, notation
Mm . Since any manifoldis a disjoint union of connected components of pure
Lecture 05, dimension this is not a serious restriction.
26.10.2022
Definition 20. Let Mm , Nn be C∞ –manifolds and let f : M −→ N be a C∞ –map.
1. p ∈ M is called a critical point of f, if rank Tp f < n = dim N. f(p) is then
called a critical value.
2. q ∈ M is called a regular value of f, if rank Tp f = n for all p ∈ f−1 ({q}). We
mention explicity that this includes that q is also considered a regular value if f−1 ({q}) =
∅!
3. f is called a submersion, if rank Tp f = n for all p ∈ M.
4. f is called an immersion, if rank Tp f = m = dim M for all p ∈ M.
5. f is called a subimmersion, if M 3 p 7→ rank Tp f is constant.
All local structural theorems for differentiable maps (e. g. the Inverse and Im-
plicit Function Theorems, local triviality of submersions, Rank Theorem, cf. Sec-
tion 1) hold verbatim for maps between manifolds. Using charts all these results,
due to their local nature, can be reduced to the corresponding Rm –result. There-
fore, we will refrain from reformulating these results here for manifolds.
Definition 21. N ⊂ Mm is called a (C∞ )–submanifold of M, if for each chart (U, φ)
of M the image φ(N ∩ U) ⊂ Rm is a submanifold of Rm . Needless to say that in this
case N itself is again a manifold. The usual equivalent characterizations of submanifolds
from Analysis II hold verbatim as well for submanifolds of manifolds.
In particular N ⊂ M is a submanifold iff for each p ∈ N there is a chart
(U, ϕ) of M centered at p such that ϕ(U ∩ N) = ϕ(U) ∩ Rn × {0Rm−n }. This
characterization does not refer to any previous notion of submanifold of Rm .
Definition 22. f ∈ C∞ (M, N) is called an embedding, if f is an injective immersion
and f : M −→ f(M) ⊂ N is a homeomorphism. Here, f(M) carries the relative topology
w.r.t. N.
In view of the Immersion Theorem 2.1, the range f(M) ⊂ N is then indeed a
submanifold.
 5. THE MORSE-SARD THEOREM 19

Proposition 23. Let Mm , Nn be C∞ –manifolds and f : M −→ N a subimmersion of

rank k. Then
1. For q ∈ N the inverse image f−1 ({q}) ⊂ M is a submanifold of dimension m − k
or ∅.
2. For p ∈ M, q = f(p), there exist neighborhoods U of p and V of q such that
S = f(U) ⊂ V is a k–dimensional submanifold of N.

Proof. Let f(p) = q. By the Rank Theorem 1.3 we may choose charts (U, ϕ =
(x1 , . . . , xm )) centered at p and (V, ψ = (y1 , . . . , yn )) centered at q in such a way
that f(U) ⊂ V and such that ψ ◦ f ◦ ϕ−1 is given by
(x1 , . . . , xm ) 7→ (x1 , . . . , xk , 0, . . . , 0).
Then

 
ϕ f−1 ({q}) ∩ U = (x1 . . . , xm ) ∈ ϕ(U)  x1 = · · · = xk = 0
= {0} × Rm−k ∩ ϕ(U),
proving that f−1 ({q}) ⊂ M is a submanifold of dimension m − k or ∅.
Furthermore,
ψ(f(U)) = ψ(V) ∩ Rk × {0},
proving the second claim. 
Corollary 24. Let Mm , Nn be C∞ –manifolds and f : M −→ N a smooth map. Then
for each regular value q of f the inverse image f−1 ({q}) is a submanifold of dimension
m − n or ∅.
Proof. We cannot immediately apply the previous proposition as f is not as-
sumed to be a subimmersion. However, if f(p) = q then, since q is regular and
hence Tp f has full rank, there is an open neighborhood U of p on which f is
a submersion. Recall that having full rank is an open condition. It ultimately
follows from the linear algebra fact that in m × n matrices of rank n are open in
the space of all rectangular m × n matrices. 
Thus f−1 ({q}) has an open neighborhood U on which fU is a submersion.
Now apply the proposition with M replaced by U. 

5. The Morse-Sard Theorem

We investigate the existence of regular values. Unless otherwise said,
M, N, . . . denote smooth manifolds.
Definition 25 (Null set). A ⊂ Mm is called a set of measure zero (short: a null set),
if for each chart (U, ϕ) the set ϕ(A ∩ U) ∈ Rm is a null set with respect to the Lebesgue
measure λm in Rm .
Remark 26. One easily checks the following facts:
1. Since diffeomorphisms of open subsets of Rm map Lebesgue null sets
to null sets this definition of null sets in manifolds is independent of the chart
chosen and therefore null sets are well-defined.
2. One point sets {p} are null sets if dim M > 0. Recall that a manifold of
dimension 0 is a finite or countable discrete set. By convention, the Lebesgue
20 1. DIFFERENTIABLE MANIFOLDS

measure on a discrete set is the counting measure. Therefore, if dim M = 0 then

there are no null sets except ∅.
3. Countable unions of null sets are null sets.
4. If A ⊂ M is a null set then M \ A is dense in M. Note that this is also true
if dim M = 0.
Theorem 27 (Morse-Sard). Let f : Mm −→ Nn , n ≥ 1, be a C∞ –map. The set of
critical values of f is a null set in N.
Remark 28. There is a version of the Morse-Sard Theorem for Cl –maps between
Ck –manifolds. In that case, however, for the validity of the Theorem k, l depend
on m, n. This can, in principle, be deduced by analyzing the proof given below
with regard to the values of m, n, k, l (which we won’t do).
Proof. Put  
Cf := p ∈ M  p critical point of f .
By Remark 26.3 and since manifolds are second countable it suffices to show that
for each p ∈ M there exists a neighborhood p ∈ U ⊂ M, such that f(Cf ∩ U) ⊂ N
is a null set. Therefore we may w. l. o. g. assume that M = U ⊂ Rm is an
open subset of Rm and that N = Rn . We then proceed by induction over m, the
dimension of M.
m = 0. Then ran f is at most countable and hence a null set in N since
n = dim N ≥ 1.
Next assume that the claim is proved for all dimensions m 0 < m. Put for
1 ≤ l < ∞:
  ∂|α| 
Cl := x ∈ U  ∀|α| ≤ l : α fx = 0
  ∂x
= x ∈ U all derivatives of f up to order l vanish in x .

We will now prove:
(1) f(Cf \ C1 ) is a null set.
(2) f(Cl \ Cl+1 ) is a null set.
(3) f(Ck ) is a null set for k ≥ m/n.
Then of course the claim follows as then
bm/nc
[
f(Cf ) = f(Cf \ C1 ) ∪ f(Cbm/nc+1 ) ∪ f(Cl \ Cl+1 )
l=1

is a null set.
f(Cf \ C1 ) is a null set. Let ξ ∈ Cf r C1 . Then we find indices i, j such that
∂fj
∂xi
(ξ) 6= 0. W. l. o. g. we may assume i = j = 1. Define h : U −→ Rm , h(x) =
(f1 (x), x2 , . . . , xm ) (if m = 1 then h(x) = f1 (x)). The derivative of h at ξ has the
form
 ∂f 
∂x
1
(ξ) 6
= 0 ∗ . . . ∗
 1 . . . .. 
0 1 .
(resp. Dh(ξ) = f10 (ξ) if m = 1).

Dh(ξ) =  
..
,
. . . ∗

 .
0 ... 0 1
 5. THE MORSE-SARD THEOREM 21

Hence Dh(ξ) is invertible and by the Inverse Function Theorem ( cf. 1.2) there
exist neighborhoods V of ξ and V 0 of h(ξ), such that hV : V −→ V 0 is a diffeo-
morphism. Consider

g := f ◦ h−1 : V 0 −→ Rn ,
g(t, x) = (t, g
e(t, x)).

The critical points of g are exactly the elements of h(Cf ∩V) and hence the critical
values of g coincide with the critical values of f restricted to V. Furthermore,
since
 
1 0 . . . 0
Dg(t, x) =
∗ Deg(t, ·)x
(t, x) is critical for g if and only if x is critical for g
et = g e(t, ·). In particular, if this does not
n = 1 then g has no critical points. mean it has
no critical val-
After making V 0 a little smaller we may assume that it is of product form ues
V 0 = I × W 0 with an open interval I ⊂ R and W 0 ⊂ Rm−1 . Then we apply
Fubini’s Theorem and obtain


λn (f(Cf ∩ V)) = λn g(t, x)  (t, x) ∈ V 0 critical

Z
 

= λn−1 g e(t, x)  x critical for g
et dt = 0,
I

since the integrand of the last integral vanishes by induction hypothesis. (We
leave the simple corner case n = 1 to the reader). Lecture 06,
f(Cl \ Cl+1 ) is a null set for 1 ≤ l < ∞. Let ξ ∈ Cl \ Cl+1 . Then there exist 31.10.2022
k1 . . . , kl+1 such that
∂l+1 f
(ξ) 6= 0.
∂xk1 . . . ∂xkl+1
W. l. o. g. let k1 = 1. Similarly as before put h : U → Rm ,
 ∂l f 
h(x) = , x2 , . . . , xm ,
∂xk2 . . . xkl+1

and as before h is a diffeomorphism of a neighborhood V of ξ onto an open set

V 0 = I × W 0 ⊂ Rm . Certainly,

h(Cl ∩ U) ⊂ V 0 ∩ ({0} × Rm−1 ).

: V 0 → Rn and g

Put g := f ◦ h−1 e := gV 0 ∩{0}×Rm−1 = g|W. Then
 
x ∈ V 0  x critical for g = {0} × e
 
x∈W  e x critical for g
e ,


f(Cl ∩ U) = g x ∈ V 0  x critical for g



e {0} × e

=g x∈W  ex critical for g
e ,

and by induction hypothesis this is a null set.

22 1. DIFFERENTIABLE MANIFOLDS

f(Ck ) is a null set for k ≥ m/n. Let d > 0 and let W be a cube of side length
d. By the Taylor formula there exists a constant L = L(k, W, f) depending on k, f,
and W such that for x ∈ Ck ∩ W, y ∈ W
|f(x) − f(y)| ≤ L|x − y|k+1 .
We subdivide W into rm cubes of side length dr . For x ∈ Ck ∩ Wj , y ∈ Wj we then
have
√ d
|x − y| ≤ m ,
r
hence
 √md k+1
|f(x) − f(y)| ≤ L ,
r
 √ k+1
that is f(Ck ∩Wj ) is contained in a cube of side length 2L md r
. Consequently,
√
m m m
 md
k+1 n
λ (f(Ck ∩ W)) ≤ r · 2 2L
r
m−n(k+1)
= const · r −→ 0,
if k ≥ m
n
and r → ∞.
Thus if k ≥ m
n
then λm (f(Ck ∩ W)) = 0. 

6. Applications of Sard’s Theorem

6.1. Embedding. We start with a basic Lemma from Topology.
Lemma 29. Let X, Y be Hausdorff spaces and f : X → Y a bijective continuous map. If
X is compact then f is a homeomorphism.
Proof. Let g := f−1 be the inverse map and consider a closed set A ⊂ X. As a
closed subset of a compact Hausdorff space, A is compact and hence g−1 (A) =
f(A) is, as image of a compact set under a continuous map, compact and thus
closed. In sum we have shown that inverse images of closed sets under g are
closed and hence g is continuous. 
Corollary 30. Let f : Mm → Nn be an injective immersion. If M is compact then f is
an embedding.
Proof. f(M) ⊂ N is compact and f is a continuous bijection onto f(M). By
the previous Lemma f : M → f(M) ⊂ N is a homeomorphism and the claim
follows 
Proposition 31 (H. Whitney). Let Mm be an m–dimensional differentiable manifold.
Then there exists an embedding Mm −→ R2m .
We are lacking all the means to prove the Theorem in this generality. It is
much more easy (and it would be possible with the methods available) to prove
an embedding into R2m+1 . However, to prove the result for compact M is still
easier and since we will not make much use of the embedding Theorem will
restrict ourselves to this case.
 6. APPLICATIONS OF SARD’S THEOREM 23

ref for non-

compact case Only for M compact and into R2m+1 . 1. We first prove that there is an em-
bedding into some RN for N large enough. Choose charts (Uj , ϕj )j=1,...,r with
ran ϕj ⊃ B(0, 3) = {|x| < 3} and
r
[
M= ϕ−1 (B(0, 1));
j=1

this is possible since M is compact. Then choose g ∈ C∞ (Rm ) with


1, |x| ≤ 4/3,
g(x) =
0, |x| ≥ 5/3,
and put 
g(ϕj (p)) · ϕj (p), p ∈ Uj ,
fj (p) =
0, otherwise.
fj ∈ C∞ (M, Rm ) und fj  ϕ−1
j (B(0, 1)) is a diffeomorphism. Hence

(f1 , . . . , fr , g ◦ ϕ1 , . . . , g ◦ ϕr ) : M −→ R(m+1)r
is an injective immersion and by Corollary 30 it is an embedding.
2. Now consider a compact submanifold M ⊂ RN . As long as N > 2m + 1
we will show that there exists a projection onto a hyperplane whose restriction
to M is an embedding.
For ω ∈ SN−1 let πω be the orthogonal projection onto hωi⊥ ' RN−1 , πω (x) =
x − hω, xiω. πω ist linear.
We first show that πω is injective for almost all ω. For p, q ∈ M we have
πω p = πω q ⇔ πω (p − q) = 0 ⇔ p − q || ω.
Consider

Φ : M × M \ (p, p)  p ∈ M −→ SN−1


(p, q) 7→ (p − q)/|p − q|.

If 2m < N − 1 then for dimensional reasons and Sard’s Theorem we conclude
that 
ω ∈ SN−1  ∃p, q ∈ M, p 6= q : p − q || ω = ran Φ ∪ − ran Φ


is a null set.
πω is an immersion if for p ∈ M, v ∈ Tp M \ {0} we have πω (v) 6= 0. Thus πω is
an immersion if and only if ω 6∈ Tp M for all p ∈ M. Putting
v
σ : TM \ {0} −→ SN−1 , v 7→
|v|
this is equivalent to ω 6∈ ran σ. If 2m < N − 1 again by Sard’s Theorem ran σ is
a null set. 3
3
This uses the fact that TM can be turned into a 2m-dimensional manifold. We will see this
below in the section on vector bundles. Furthermore TM \ {0} means that the zero section (i.e. all
the 0’s in all tangent spaces Tp M, p ∈ M is removed). Also this will be better understood when
we will have discussed vector bundles.
24 1. DIFFERENTIABLE MANIFOLDS

In sum for 2m + 1 < N the set ran Φ ∪ − ran Φ ∪ ran σ is a null set and for ω
outside this set πω is an injective immersion and hence, since M is compact, is
an embedding. 
6.2. Transversality. Skipped
 APPENDIX A

1. Structure theorems for differentiable maps

Lecture 01,
To understand manifold theory the following structure theorems for differ- 12.10.2022,
entiable maps from Analysis II will be crucial and therefore they should be Lennart
Ronge
recalled.

1.1. Implicit Function Theorem. Let U ⊂ Rpx , V ⊂ Rqy be open subsets and
let F : U × V −→ Rqy be a Ck –map, k ≥ 1. Let a fixed solution (a, b) ∈ U × V of
the equation F(a, b) = 0 be given such that the matrix
 ∂F 
i
D2 F(a, b) := (a, b)
∂yj 1≤i,j≤q

is invertible.
Then there exist open neighborhoods a ∈ U1 ⊂ U, b ∈ V1 ⊂ V and a Ck –map
ϕ : U1 −→ V1 , such that
   
(x, y) ∈ U1 × V1  F(x, y) = 0 = (x, y) ∈ U1 × V1  y = ϕ(x) ,
i. e. in U1 × V1 the solutions of the equation F(x, y) = 0 are given as the graph of
the smooth map ϕ. Furthermore, the derivative of ϕ can be expressed in terms
of the derivatives of F as follows:

−1  ∂F 
i
Dϕ(x) = − (D2 F)(x, ϕ(x) (D1 F)(x, ϕ(x)), D1 F(a, b) := (a, b) .
∂xj 1≤i,j≤p

1.2. Inverse Function Theorem. Let U ⊂ Rp be an open subset and f : U −→

Rp a Ck –map, k ≥ 1. For a fixed a ∈ U assume that the derivative Df(a) is
invertible.
 Then there exist open neighborhoods a ∈ U0 ⊂ U, f(a) ∈ V, such that
fU :−→ V is a Ck –diffeomorphism.
0

1.3. Rank Theorem. Let U ⊂ Rpx be an open subset and f : U −→ Rq a

Ck –map, k ≥ 1. Consider a fixed a ∈ U, b := f(a). Then we have:
1. If rank Df(a) = r then there exist local diffeomorphisms ψ : Uψ ⊂ U −→
Vψ ⊂ Rp with ψ(a) = 0, ϕ : Uϕ ⊂ Rq −→ Vϕ ⊂ Rq mit ϕ(b) = 0, such that
ϕ ◦ f ◦ ψ−1 (x = (x1 , . . . , xp )) = (x1 , . . . , xr , f(x)),
e

with a Ck –map f.
e
2. If rank Df(x) = r is constant in a neighborhood of a, then one can find ϕ
and ψ in such a way that fe = 0.

25
 26 A.

2. Submanifolds of Rn
In my course on Analysis II we studied, as an application of the structure
theorems for differentiable maps, submanifolds of Rn . We recall here the impor-
tant
2.1. Immersion Theorem. Let U ⊂ Rk open and let ϕ : U → Rn be an
immersion, that is rank Dϕ(x) = k for all x ∈ U. Then for any x0 ∈ U there exists
an open neighborhoodU ⊃ U 0 3 x0 such that ϕ(U 0 ) ⊂ Rn is a k–dimensional
submanifold and ϕ : U 0 → ϕ(U 0 ) is a homeomorphism.
As a consequence, if ϕ : U → Rn is an injective immersion such that ϕ : U →
ϕ(U) is a homeomorphism (with ϕ(U) having the relative topology from Rn )
ÜÜ then ϕ(U) is a submanifold of Rn .
 Bibliography

[Trè06] F. Trèves, Topological vector spaces, distributions and kernels, Dover Publications Inc., Mine-
ola, NY, 2006, Unabridged republication of the 1967 original. MR 2296978 (2007k:46002)

27