Differentiable Manifolds

Elden Elmanto

Department of Mathematics, University of Toronto, 40 St. George St.,

Toronto, ON, M5S 2E4, Canada
E-mail address: [email protected]
URL: https://www.eldenelmanto.com/
 Contents

Chapter 1. Introduction 5

Chapter 2. What is differential topology? 7

1. Motivation: exotic spheres 7
2. Some recollections on analysis 8
3. Some recollections on topology 13
Chapter 3. Examples of smooth manifolds 17
1. Spheres 17
2. Projective spaces 19
3. Tori 19
Chapter 4. Morphsims of manifolds, tangent spaces and derivatives 21
1. Smooth maps of manifolds 21
2. The tangent space and the derivative 24
3. Application: globalizing calculus 32
Chapter 5. The Whitney embedding theorem 33
Appendix A. 35

Appendix. Bibliography 37

3
 CHAPTER 1

Introduction

These are notes for MATD67H3, the UTSC version of a first class in differentiable manifolds.

5
 CHAPTER 2

What is differential topology?

1. Motivation: exotic spheres

One of the most important and beautiful collection of spaces in mathematics are the spheres.
We are all familiar with the circle
S1 := {(x, y) : x2 + y 2 = 1} ⊂ R2 .
This is an example of a differentiable manifold, the main object of study in this class. What
makes the circle “nice” (or, really, what makes it a differentiable manifold) is that around each
point x ∈ S1 there exists a neighborhood around x which looks like the interval (a, b) ⊂ R. To
make this more precise let us recall the trigonometric identity
cos(t)2 + sin(t)2 = 1.
Then, roughly speaking, we can define a function
θ+ : (0, 2π) → R2 t 7→ (cos(t), sin(t))
whose image almost hits all S1 . The main property of this map is that it is a bijection onto
its image; even better it is a something called a diffeomorphism. This is a concept which we
will introduce later on in class but it is a bijection that preserves “all possible differentiable
information” between the two spaces. Now, note, that (cos(0), sin(0)) = (cos(2π), sin(2π)) =
(1, 0) and hence we are missing the point (1, 0). The extension of θ+ to [0, 2π) is a bijective
map, but its inverse is not continuous: if it were, one can show that its image had to be open
which is not the case. Hence we cannot parametrize S1 fully using any one open subset of R.
To resolve this problem, consider another function
θ− : (0, 2π) → R2 t 7→ (cos(t + π), sin(t + π))
which does contain the point (1, 0) in the image. The union of the images of θ− and θ+ turns
out to cover all of S1 . This is a key difference in analysis versus differential topology: whereas
analysis studies intervals (a, b) ⊂ R and functions on them, spaces that one encounters in
differential topology will require “gluing” together these intervals. This is where the rubber
meets the road: how intervals and their generalizations (open subsets U ⊂ Rn ) are glued
together is what makes differential topology both interesting and difficult.
To make all of this precise, we will introduce the notion of a differentiable structure on
topological spaces. Roughly speaking a differentiable structure introduces to each point x ∈ S1
the concept of an infinitely differentiable function f : S1 → R around x which are consistent as
x varies. For example: if x 6= (1, 0) then x is in the image of θ+ and we can compose f with θ+
to get
f ◦ θ+ : (0, 2π) → R.
Now we open up a book in analysis and there is a notion (to be reviewed today) of what
it means for this composite to be differentiable, continuously differentiable or even infinitely
differentiable (smooth)! In this sense, θ+ is used to “coordinatize” the function f and we are
happy if the resulting coordinatization of f is smooth. To explain “consistency” will take us
further afield, but it concerns how different coordinatization varies (roughly, the overlaps of the
images of θ− and θ+ ).
7
8 2. WHAT IS DIFFERENTIAL TOPOLOGY?

For our informal discussion, we assert that anytime we have a set Z ⊂ Rn defined by
some polynomial equation (in the above case Z is defined by x2 + y 2 = 1) then Z inherits a
standard differentiable structure. Roughly: for any x ∈ Z and any open subset U ⊂ Z
(these are of the form U ∩ Z where U ⊂ Rn is a union of open balls in Rn ), then we declare
differentiable functions to be those that come from infinitely differentiable functions f : U → R.
Of fundamental interest are the spheres:
Sn = {(x1 , x2 , · · · , xn+1 ) : x21 + x22 + · · · x2n+1 = 1} ⊂ Rn+1 ,
with its standard differentiable structure.
Now, when n = 1, we have the standard differentiable structure, and also one defined in the
previous paragraph using θ+ and θ− . It turns out that they are diffeomorphic. In topology,
we have the notion of a homeomorphism: two spaces X and Y are homeomorphic if there are
continuous functions
f :X→Y g:Y→X
such that g ◦ f, f ◦ g are the identities. A diffeomorphism is a homeomorphism that preserves
differentiable structure. For example, f, g could both be the identity map on underlying sets,
but one can imagine equipping X and Y with different differentiable structures and X and Y
would not be a diffeomorphic. In the case of the sphere, we say that a differentiable structure
on Sn is exotic if it is not diffeomorphic to the standard structure.
Question 1.0.1. Is there any exotic differentiable structure on the sphere?
It turns out that when n = 1, 2, 3, 5, 6 (Note that 4 is skipped) one cannot find any exotic
differentiable structure on Sn . Milnor shocked mathematics when he proved:
Theorem 1.0.2 (Milnor 1956, [Mil56]). There exists an exotic differentiable structure on
S7 .
The construction is not easy and even proving that his sphere is not the standard one
requires technology outside the scope of this class. Roughly, however, Milnor’s exotic 7 sphere
is constructed out of arranging S3 and S4 to “glued” in a strange way. He then proves (and in
fact develops an entire technology for proving) that it is not the boundary of any 8-dimensional
differentiable manifold: note that Sn is the boundary of
Dn+1 := {(x1 , x2 , · · · , xn+1 ) : x21 + x22 + · · · x2n+1 6 1}.
The concept of a manifold with boundary is something we will encounter in this class. In fact,
this result comes as a surprise to Milnor himself. At this point in history, we have understood
smooth structures better than ever before. As a sampler: there are 27 exotic smooth structures
on S7 , while there are 523263 on S19 and we know that S61 is the last odd-dimension where no
exotic smooth structure can exist (we only knew this in 2017). But a question remains:
Question 1.0.3. Is there an exotic 4-sphere?

2. Some recollections on analysis

The first order of business is to define what a differentiable manifold is. We say the words
and then fill in the meaning.
Definition 2.0.1. A space X is a topological manifold of dimension n if it is a second
countable Hausdorff space such that for any x ∈ X there exists an open Vx ⊂ X for which Vx is
homeomorphic to an open subset Ux of Rn . A chart on X is a triple (U, V, ϕ) where U ⊂ Rn is
an open, V ⊂ X is an open and ϕ :S U → V is a homeomorphism. An atlas for X is a collection
of charts {(Uα , Vα , ϕα )} such that α Vα = X. An atlas is said to be smooth if for any indices
α, β the map
ψαβ := ϕ−1 −1 −1
β ϕα : ϕα (Vα ∩ Vβ ) → ϕβ (Vα ∩ Vβ )
is a smooth.
 2. SOME RECOLLECTIONS ON ANALYSIS 9

Definition 2.0.1 is a combination of both analytic and topological definitions. So we begin

by reviewing some analysis and some topology. In particular, our first order of business is to
explain the following concepts:
(1) a topological space;
(2) a second countable space;
(3) a Hausdorff space;
(4) smooth maps.
We will start from the last concept, which is analytic in nature. We will give references for
the proofs of these statements but may not necessarily give them in full; we refer to [Spi65] for
a textbook reference. A key theme in this class is that we are trying to globalize analysis with
multiple variables.
2.1. Reminders on differential analysis in one variable. Let us recall the following
basic definitions in one variable:
Definition 2.1.1. A function f : R → R is continuous at x0 if for all  > 0, there exists
a δ > 0 such that for any
y ∈ (f (x0 ) − , f (x0 ) + ),
there exists x ∈ (x0 − δ, x0 + δ) such that f (x) = y.
We have seen that this is the key definition in analysis and arguably the main object of
study. It says that for any -neighborhood of f (x0 ), there exists a δ-neighborhood around x0
such that any y in the -neighborhood, it is realized as a value of f . This definition is local
in nature: it makes no claim on whatever happens outside of a small neighborhood of x0 . In
particular if (a, b) ⊂ R is an open interval we can ask that f is continuous on every point of
(a, b). Though making “small neighborhood” a precise concept is not easy. In particular, even
if f is only defined on a subset A ⊂ R we know how to make sense of continuity: we ask that
for any x0 ∈ A and any -neighborhood of f (x0 ), there exists a δ > 0 such that for any y in the
-neighborhood, it can be realized as a a value of f on A ∩ (x0 − δ, x0 + δ).
Continuous functions are often pathological — the functions that appear in daily life, or
even mathematics, are of the following form
(1) trigonometric functions like sin, cos, tan,
(2) exponential function exp,
(3) polynomials: f (x) = 27x5 + 6x2 + 7 and so on.
These functions are differentiable, which is a kind of “niceness” or “regularity” condition that
one can put on functions.
Definition 2.1.2. A function f : R → R is differentiable at x0 if the limit
f (x0 +h)−f (x0 )
lim h
h→0
exists; in which case we denote its value by f 0 (x0 ).
Again, the notion of a differentiable function is local in nature: it only demands “good
behavior” around x0 . We recall that if f is differentiable at x0 , then f must be continuous
at x0 although the converse is false. Now, suppose that U ⊂ R is an open subset and f is a
function f : (a, b) → R, then we can ask that f is differentiable everywhere on U, i.e., for any
x0 ∈ (a, b), f 0 (x0 ) exists. Then we can ask that the function
A→R x0 7→ f 0 (x0 )
is continuous. In this case, we say that f is continuously differentiable. More generally, if
A ⊂ R is a subset, then f is continuously differentiable if for all x ∈ A, there exists a small
open interval (a, b) around x and an continuously differentiable map F : (a, b) → R such that
F|(a,b)∩A = f |(a,b)∩A . The notation for this is:
f ∈ C1 (A)
10 2. WHAT IS DIFFERENTIAL TOPOLOGY?

and we read “f is a C1 -function.” If we set C0 (A) to be the collection of continuous function,

then f is C1 if and only if f 0 is in C0 . So we say that f ∈ Ck (A) if f 0 ∈ Ck−1 (A). This means
that f is k-times differentiable with all its derivatives continuous. The collection C∞ (A) is then
defined to be those functions on f which is in Ck (A) for all k. This means that all derivatives of
f exists and are continuous; in this case we say that f is a smooth function. All the functions
listed above are indeed smooth. We now generalize the definition of a smooth function to the
multivariate context.

2.2. Reminders on differential analysis in multiple variables. Now, we want to do

all of this with multiple variables. Presumably you have already done this in your analysis class
and maybe found this intimidating. Let me remark that in this class we will only be performing
“soft analysis” — this means that we will hardly ever find an explicit  − δ estimate and content
ourselves with very very rough estimates (or no estimate at all).
Example 2.2.1 (Soft analysis). Most arguments in your first analysis class will involve
choosing suitable values/estimates for  and δ. One of the things that this is good for is that
it keep us honest and not make mistakes, an important task for any mathematician. In this
class, we assume some familiarity with analysis and thus we allow ourselves to perform soft
arguments. For example, recall that we have the intermediate value theorem:
Theorem 2.2.2 (Intermediate value theorem). Let f : [a, b] → R be a continuous function,
then for any x in between f (a) and f (b) there exists an element c ∈ [a, b] such that f (c) = x.
A soft analysis proof goes like this: since f is continuous and [a, b] is connected, f ([a, b]) is
connected and thus any element in between the values f (a) and f (b) must be hit. Presumably
the reader can fill in the  − δ details but we will not demand that in this class.
Now, a function f : Rn → Rm can be represented as
f (x1 , · · · , xn ) = (f1 (x1 , · · · , xn ), · · · , fm (x1 , · · · , xn )).
Spelling this out, this means that we are given an m-tuple of functions which has n-variables
as input. At this point you should be somewhat familiar with the notation
B (x0 ) := {x ∈ Rm : |x − x0 | < } ⊂ Rm .
This is called the “ball of radius ” and | − | refers to the norm in m-space. This is the
generalization of an -interval in one variable. The notion of a continuous function is easy
enough to generalize:
Definition 2.2.3. Let U ⊂ Rn be an open subset. A function f : U → Rm is continuous
at x0 ∈ U if for all  > 0, there exists a δ > 0 such that for any y ∈ B (f (x0 )), there exists
x ∈ Bδ (x0 ) ⊂ U such that f (x) = y. More generally, if A is any subset of Rn , it is continous
at x0 ∈ A if there exists an open subset U of Rn containing x0 and a continuous function
F : U → Rn such that F|U∩A = f |U∩A .
This is all good, but in the interest of only performing soft analysis arguments we note the
following formulation of continuity.
Lemma 2.2.4. Let A ⊂ Rn . A function f : A → Rm is continuous if and only if for any
U ⊂ Rm , there exists an open V ⊂ Rn such that f −1 (U) = V ∩ A.
k
S Here’s the key point: a subset W ⊂ R is open if and only if it can be written as
Proof.
a union w∈W B (w) = W. 

Lemma 2.2.4 says that continuity is the preservation of a topological property of subsets,
namely of openness. It is more abstract than the  − δ argument but also conceptually simpler.
As an application, we give a soft-analysis argument of the following result which might be
familiar in an analysis class.
 2. SOME RECOLLECTIONS ON ANALYSIS 11

Corollary 2.2.5 (Maximum value theorem). Let A ⊂ Rn be a compact subset and f :

A → Rm continuous. Then f (A) is compact. In particular, it contains the supremum.
Proof. Let O = {Vα }α∈A be an open cover of f (A). Since f is continuous, for each open
Vα , we have that f −1 (Vα ) is open. In particular {f −1 (Vα )} is an open cover of A. But now,
A is compact, hence it refines to a finite cover f −1 (Vj ), j = 1, · · · , n. Then {Vj }j=1,··· ,n is an
subcover of O which covers f (A). 
We will revisit these ideas in a more abstract context in the next lecture. But we now
discuss differentiability. To motivate this definition we reformulate Definition 2.1.2. In the
notation of that definition, let us consider the linear function:
Dfx0 : R → R λ(h) = f 0 (x0 ) · h.
This is a linear function that simply scales vectors by the quantity f 0 (x0 ). This might seem
puzzling at first since Dfx0 is basically the same information as f 0 (x0 ); the latter is the value
of the former at 1. It turns out that, for our purposes, the right way to view the derivative at
a point is not as a number, in other words we don’t really want to think of the derivative as a
function x 7→ f 0 (x). But we want to think of as a function x 7→ Dfx where the target is a linear
function. Do not worry about this for now, but note that a function f : R → R is differentiable
if and only if
lim f (x0 +h)−fh(x0 )−λ(h) = 0.
h→0
Hence we can reformulate the notion of a function being differentiable at x0 by saying that
there exists a linear function Dfx0 : R → R such that
f (x0 +h)−f (x0 )−Dfx0 (h)
lim h = 0.
h→0

One of the first results that one proves about this definition is that Dfx0 is, in fact, unique and
so the notation is justified; we also note that differentiable functions are continuous.
In any case, the higher dimension generalization of being differentiable is as follows.
Definition 2.2.6. A function f : Rn → Rm is said to be differentiable at x0 ∈ Rn if
there exists a linear transformation
Dfx0 : Rn → Rm
such that
|f (x0 +h)−f (x0 )−Dfx0 (h)|
lim |h| = 0.
h→0

We call the transformation Dfx0 the derivative of f at x0 if it exists. This linear trans-
formation is unique.
Lemma 2.2.7. The derivative is unique.
Proof. Let T : Rn → Rm be a transformation that satisfies
|f (x0 +h)−f (x0 )−T(h)|
lim h = 0.
h→0

Rewriting d(h) := f (x0 + h) − f (x0 ) we have that

|T(h)−Dfx0 (h)| |T(h)−d(h)+d(h)−Dfx0 (h)| |T(h)−d(h)| |d(h)−Dfx0 (h)|
lim |h| = lim |h| 6 lim |h| + lim |h| 6 0.
h→0 h→0 h→0 h→0

Therefore, for any x 6= 0 we have, by linearity, that:

|T(hx)−Dfx0 (hx)| |h| |T(x)−Dfx0 (x)| |T(x)−Dfx0 (x)|
lim |hx| = lim = ,
h→0 h→0 |h| |x| |x|

and thus T = Dfx0 as linear transformations. 

The derivative satisfy generalizations of usual properties that we have seen in one variable
calculus. We refer to [Spi65, Theorem 2-2, 2-3] for statements and proofs of these.
12 2. WHAT IS DIFFERENTIAL TOPOLOGY?

2.3. The Jacobian and partial derivatives. Since Dfx0 : Rn → Rm is a linear trans-
formation we can, up to choosing bases, represent Dfx0 as a matrix. This matrix, in fact,
encodes important properties about the derivative. Suppose that f : Rn → R is a func-
tion, then we can hold all variables constant in the domain except at the i-th spot, i.e., fix
a1 , · · · , ai−1 , ai+1 , · · · an ∈ R and consider the function
f :R→R x 7→ f (a1 , · · · , ai−1 , x, ai+1 · · · an ).
In this way, the i-th partial derivative at x0 = (a1 , · · · , an ) is the limit
f (a1 ,··· ,ai−1 ,ai +h,ai+1 ···an )−f (a1 ,··· ,ai−1 ,ai ,ai+1 ···an )
lim h =: Di (f )(x0 )
h→0

Now, write f : Rn → Rm as f = (f 1 , · · · , f m ) noting that a map into Rm is the same thing as

m-functions f 1 , · · · , f m : Rn → R.
Proposition 2.3.1. If f : Rn → Rm is differentiable at x0 , then Dj f i (x0 ) exists and the
Jacobian matrix is given by:
D1 f 1 (x0 ) D2 f 1 (x0 ) · · · Dn f 1 (x0 )
 
 D1 f 2 (x0 ) ··· · · · Dn f 2 (x0 ) 
 
 ··· ··· ··· ··· 
m m
D1 f (x0 ) ··· · · · Dn f (x0 ).
Proof. Apply the chain rule; see [Spi65, Theorem 2-7].

Remark 2.3.2. If f : R → R, we have reformulated the derivative as a linear transformation
Dfx0 : R → R. The Jacobian matrix corresponding to this linear transformation is exactly the
1 × 1-matrix given whose only entry is the number f 0 (x0 ). In this sense the Jacobian matrix
is a generalization of the maneuver we made above to go from the number f 0 (x0 ) to the linear
transformation Dfx0 .
We remark that the converse to Proposition 2.3.1 fails: there exists a function whose partials
exist but need not be differentiable. The following theorem guarantees existence as soon as we
know that the partials are continuous.
Theorem 2.3.3 (Differentiability theorem). If f : Rn → Rn is a function, then Dfx0 exists
if the limits Dj f i (x0 ) exists for all i, j in an open set containing x0 (call it U) and the function
Dj f i : U → Rn is continuous.
Proof. See [Spi65, Theorem 2-8]. 
Definition 2.3.4. Let U ⊂ Rm be an open subset. A function f : U → Rn is continuously
differentiable at x0 if the hypothesis of Theorem 2.3.3 holds. If it is continuously differentiable
at all points of U, then we say that f is continuously differentiable. As usual if A is any
subset and f : A → Rm is a function, we say that it is continuously differentiable at A if for
all x ∈ A there exists an open U ⊂ Rm and a continuously differentiable function F : U → Rm
such that F|U∩A = f |U∩A . We write
f ∈ C1 (A).
In other words, if A ⊂ Rm a function f : A → Rm is C1 (A) is all its partial derivatives
exists and are continuous. To iterate this definition, we have to consider mixed partials: given
f : A → R, assume that Di f : A → R exists. Then we can ask if its derivative exists, this time
maybe with respect to the j-th coordinate. If it exists we write:
Dj,i f := Dj (Di f ) : A → R.
Lemma 2.3.5 (Symmetry of second derivatives). Assume that Dj,i f and Di,j f are contin-
uous in an open set containing x0 . Then
Dj,i f (x0 ) = Di,j f (x0 ).
 3. SOME RECOLLECTIONS ON TOPOLOGY 13

Proof. See [Spi65, Theorem 2-5].



Definition 2.3.6. Let A ⊂ Rm . A function f : A → R is in Ck (A) if Di f is in Ck−1 (A)

for all i = 1, · · · , m. The set C∞ (A) is then defined to be the set of functions such that
Di f is in Ck (A) for all k and for all i = 1, · · · , m. Functions in this set are called smooth
functions. In other words, f is smooth if all partials exists and are continuous. A function
f = (f 1 , · · · , f n ) : A → Rn is smooth if each f i is smooth for i = 1, · · · , n.

Remark 2.3.7 (Additional structure on C∞ (A)). There are two more important things to
note about smooth functions which are not necessarily used in this class, but are fundamental.
First, let f, g ∈ C∞ (A). Then note that we can define

(f + g)(x) := f (x) + g(x) f · g(x) := f (x) · g(x).

The functions f + g and f · g are both smooth. Hence C∞ (A) forms the structure of a ring
with the constant functions at 1 and 0 as the multiplicative and additive identity respectively.
Furthermore, we have a compatible action of R on C∞ (A): if r ∈ R

(r · f )(x) := rf (x)

is also a smooth function; whence C∞ (A) is a R-algebra. Informally, it is a R-vector space

where we know how to compatibly multiply vectors
Next, we know that the the operation of taking derivatives enjoys some very nice properties;
for a list that you need to know see [Spi65, Theorems 2-2, 2-3]. They come under the monikers
of “Liebniz rule, chain rules etc.” It is reasonable to ask: can we interact with derivatives
formally just knowing its properties; we state a result that uniquely characterizes the derivative
(at least in one variable):

Lemma 2.3.8. Let A ⊂ R. Then the derivative

f 7→ f 0 : C∞ (A) → C∞ (A)

is characterized as the R-linear map satisfying the following properties:

(1) transforms constants to zero 1 7→ 0;
(2) transforms linear maps to constants x 7→ 1;
(3) the Liebniz rule
(f g)0 = f g 0 + gf 0 .

This will be left an exercise to the reader.

In essence, the main functions of interest in differential topology are smooth functions
f : Rn → R. In fact, at this point, we can define a manifold as follows: first, let X, Y be subsets
of Rn . We say that a function f : X → Y is a diffeomorphism if f is a bijection such that f
and f −1 are both smooth functions. Then a smooth manifold of dimension n is a subset
M ⊂ Rk such that any x ∈ M has a neighborhood W ⊂ M such that W is diffeomorphic to
W0 where W0 ⊂ Rn . This is a perfectly good definition except that it is rather unsatisfying:
for example what does k have to do with n and it is not clear if it depends on the inclusion
M ⊂ Rk . We want to get rid of these dependence and so we introduce the abstract concept of
a smooth manifold.

3. Some recollections on topology

We run through the definitions from topology which are relevant for us; the notes for the
other class gives a more extensive treatment.
14 2. WHAT IS DIFFERENTIAL TOPOLOGY?

3.1. Hausdorff, second-countable topological spaces.

Definition 3.1.1. A topology on a set X is a collection T ⊂ P(X) subject to the following
properties:
(1) ∅ and X are in T,
(2) T is closed under arbitrary unions,
(3) T is closed under finite intersections.
The pair (X, T) is called a topological space. Elements in P(X) are called open. If x ∈ X,
then any element U ⊂ P(X) such that x ∈ U is called a open neighborhood of x. We often
write X for a topological space (or, simply, a space) and suppress T.
In this class, the language of topology is useful for interfacing with analysis and making
“soft-analysis” argument as we had already seen. For example, the notion of continuity is easy
to formulate: if X, Y are spaces and f : X → Y is a function, then it is continuous if for any
open set V ⊂ Y, the inverse image f − (V) ⊂ X is an open set.
Often, one wants to specify a topology using only a partial collection of open sets. We have
already seen this in real analysis; we begin by reviewing the idea of open balls in a slightly more
abstract context.
Example 3.1.2 (Metric spaces). In your analysis course, you might have learned the
idea/definition of a metric space — perhaps at least on subsets of Rn . Intuitively, a metric
specifies a way to measure distances between points in a set. We recall that a metric space is
a space X equipped with a function
d:X×X→R
such that
(1) d(x, y) > 0 for any x, y ∈ X
(2) d(x, x) = 0
(3) d(x, y) = d(y, x)
(4) we have the triangle inequality
d(x, z) 6 d(x, y) + d(y, z).
Recall that we have the notion of -balls for  > 0, i.e.,
B(x0 , ) := {y : d(x0 , y) < }.
These are the basic building blocks for arguments in real analysis and are the building blocks
for opens in X. The notion of a basis makes this precise.
Definition 3.1.3. Let X be a set. A collection B ⊂ P(X) is a basis for a topology if:
(1) for each x ∈ X, there is at least one basis element that contains x;
(2) if x is in the intersection of two basis elements B1 ∩ B2 , then there is a basis element
B3 with x ∈ B3 ⊂ B1 ∩ B2 .
The topology generated by B is the topology defined as follows: an set U ⊂ X is open if
for each x ∈ U there exists an element B ∈ B such that x ∈ B ⊂ U. The following is a routine
check and left to the reader.
Lemma 3.1.4. The topology generated by a basis B is indeed a topology. In fact, opens are
exactly those which can be expressed as an arbitrary unions of elements in B.
Example 3.1.5 (The Euclidean topology on Rn ). Continuing the example of metric spaces,
we see that the collection {B(x0 , )}x0 ∈X forms a basis for the topology on X; the second
condition is most easily verified by drawing a picture. By Lemma 3.1.4 we see that a set U ⊂ X
is open if and only if for each x ∈ U there exists a  > 0 such that B(x0 , ) ⊂ U. We will use the
notation (Rn , Euc) to denote Rn with its usual metric topology; here Euc is meant to indicate
the Euclidean topology which is the usual name for this topology.
 3. SOME RECOLLECTIONS ON TOPOLOGY 15

In the other class we will learn many interesting, but perhaps pathological, topological
spaces. For example, one can endow a finite set with a topology whose role seems to be
combinatorial in nature. We want to only consider “geometric” topological spaces in this class
and the axioms that let us do these is as follows.
Definition 3.1.6. A space X is said to be Hausdorff or separated if for any two distinct
points x, y there exists open neighborhoods x ∈ Ux , y ∈ Uy such that Ux ∩ Uy = ∅. It is called
second countable or completely separable if its topology has a countable basis.
Remark 3.1.7. Suppose that X is a Hausdorff space with the property that any x ∈ X
admits an open neighborhood homeomorphic to an open subset of Euclidean space. Then it is
second countable if and only if it is metrizable, that is, admits a metric space. We certainly want
differentiable manifolds to have a notion of distance and hence second countability is desired.
At this point, we have completed Definition 2.0.1 and understood what it means for a space
to have a smooth atlas. However, let us explain a quick warning.
Remark 3.1.8 (Invariance of domain). Technically, Definition 2.0.1 is a little abusive. We
have to specify that our atlas is made up of open’s of Rn for a specified n because there is no
reason for this particular n to be well-defined. What allows us to get away with this abuse is a
result called invariance of domain, proved by Brouwer using algebraic topology.
Theorem 3.1.9 (Brouwer). Let U ⊂ Rn be a open subset and f : U → Rn a continuous,
injective function. Then f (U) ⊂ Rn is open.
Using Brouwer’s theorem, we see that a space cannot have an atlas of different dimensions.
Indeed, if we have a chart which constitutes a homeomorphism f : U → V where U ⊂ Rk and
0
V ⊂ Rk where k > k 0 , then we can look at the composite
0
U → V ⊂ Rk ,→ Rk .
This produces a map which is a continuous injection and thus, by Brouwer’s theorem, an open
0
map. But no open subset of Rk can be exclusively contained in Rk , unless it is empty.
Before we go on to define what a differentiable manifold, let us extract an important piece
of data from a smooth atlas.
−1
Remark 3.1.10 (Atlases and cocycles). The maps ψαβ : ϕ−1 α (Vα ∩ Vβ ) → ϕβ (Vα ∩ Vβ ),
which are maps between open subsets of Eucildean space, are called transition functions and
they satisfy
ψαα = id ψαβ ◦ ψβγ = ψαγ .
These conditions make precise what we mean by “compatibility” and is often called the cocycle
condition.
Furthermore, the above two conditions imply that ψαβ ◦ ψβα = id. Hence ψαβ is a diffeo-
morphism: a smooth bijection with a smooth inverse.
In fact, there is another way to specify a differentiable manifolds by glueing together cocy-
cles; we fix n. Let us take as input a set I with elements denoted by i ∈ I. For each i ∈ I we
are given Xi ⊂ Rn an open and for each Xi we are given an open
Uij ⊂ Xi ∀j ∈ I.
For each pair i, j ∈ I we are given diffeomorphisms
ψij : Uij → Uji
−1
for which 1) Uii = Xi , 2) ψij (Uji ∩ Ujk ) = Uij ∩ Uik and the diagram
ψik
Uij ∩ Uik Uki ∩ Ukj
ψij ψjk

Uji ∩ Ujk .
16 2. WHAT IS DIFFERENTIAL TOPOLOGY?

commutes. Then there exists topological manifold X and a smooth atlas of dimension n,
{(Xi , Vi , ϕi )} such that
(1) ϕi : Xi → Vi is a homeomorphism;
(2) ϕi (Uij ) = Vi ∩ Vj
(3) ψij = ϕ−1 j |Vi ∩Vj ◦ ϕi |Uji .
This will be left as exercise in homework 2.
Now, two smooth atlases on X are said to be compatible if their union is an atlas and an
atlas is maximal if it is one with the property that any atlas compatible with it must be in it.
Definition 3.1.11. An n-dimensional smooth manifold or, simply, an n-manifold is
a second-countable, Hausdorff space X with a maximal atlas.
In practice, one finds a smooth atlas on X and then generate a maximal one by the following
lemma:
Lemma 3.1.12. Every smooth atlas is contained in a unique maximal atlas.
Proof. See [Kup, Lemma 2.2.2].

 CHAPTER 3

Examples of smooth manifolds

We now discuss examples on smooth manifolds and how to get new smooth manifolds out of
old ones. From now on, by a manifold we really mean a smooth one unless otherwise stated. We
want to illustrate several methods for showing that something is a manifold. We will repeatedly
use the following fact without further comment.
Lemma 0.0.1. Let X be a space. If A ⊂ X is a subset, then the subspace topology on A
satisfies:
(1) if X is Hausdorff, then A is;
(2) if X is second countable, then A is.

1. Spheres
Let us now give Sn ⊂ Rn+1 the structure of a smooth manifold of dimension n. Evidently,
it is second countable and Hausdorff by Lemma 0.0.1 since it is a subspace of Rn+1 . Consider
U+ := Sn r {(1, 0, · · · , 0)} U− := Sn r {(−1, 0, · · · , 0)}.
We then define
ϕ−1
+ : U+ → R
n 1
1−x0 (x1 , · · · , xn ),
ϕ−1
− : U− → R
n 1
x0 +1 (x1 , · · · , xn ).
Indeed, ϕ−1 −1
+ and ϕ− are diffeomorphisms which generalizes the (inverse of) the example of S
1

at the very beginning of this class.

It can be a little cumbersome to “eyeball” charts for manifolds, especially if it comes as the
vanishing locus of some equation in n-space; what I mean by this is that Sn is the vanishing
locus of the quadratic equation
q : Rn+1 → R (x0 , · · · , xn ) x20 + · · · x2n − 1,
and we want to say that q is nice enough such that its zero locus must have the structure
of a smooth manifold. Furthermore, there is nothing special about the number 1, the equa-
tion (x0 , · · · , xn ) x20 + · · · x2n − 57 is also a nice space with a smooth structure and it is
diffeomorphic to S1 .
However, one must be careful: not all equations will have a vanishing locus which is smooth.
Clearly if we consider x20 + · · · x2n , then its vanishing locus is just the point (0, · · · , 0); it does
have a smooth structure but there’s something “off” because it is zero dimensional. A better
class of examples is discussed in the following remark
Remark 1.0.1 (Real, punctured elliptic curves). Consider the vanishing locus of the fol-
lowing family of functions
fb (x, y) = y 2 − x3 − b.
For all b 6= 0, note that we have manifolds diffeomorphic to R1 . But when b = 0 we get the
“cusp.” This is a space, homeomorphic to R1 but has no smooth structure, because of the cusp
at (0, 0).
The following question is what we would like to answer next:
17
18 3. EXAMPLES OF SMOOTH MANIFOLDS

Question 1.0.2. Suppose that f : Rn+1 → R is a function, when is f −1 (0) ⊂ Rn+1 is an

n-dimensional manifold.

Here’s our strategy: as Z := f −1 (0) is already second countable and Hausdorff, we just
need to produce an atlas and then take something maximal containing it. In particular, the
following is enough: for each x ∈ Z, find an open subset U ⊂ Z such that U is diffeomorphic
to a subset of Rn . In particular, this is a local check: just choose some U first and then find a
U0 ⊂ U (so, possibly after shrinking U) which is indeed diffeomorphic to a subset of Rn .
In multivariable calculus, we have already learned a method of showing that some function
is diffeomorphic, at least for a map f : Rn → Rn .

Theorem 1.0.3 (Inverse function theorem). Let U ⊂ Rn be an open subset and x ∈ U.

Assume that f : U → Rn is a smooth map whose derivative Dfx : Rn → Rn is invertible. Then
there is an open subset U0 ⊂ U containing x such that f (U0 ) is open and f |U0 : U → f (U) is a
diffeomorphism.

Theorem 1.0.3 is plainly very useful: it converts the task of showing that f is diffeomorphism
into one of linear algebra, at least if we are willing to shrink around x. The inverse function
theorem can be used to study maps into Rm when m 6 n. To state this result, let us introduce
a piece of terminology. Let U ⊂ Rn be an open subset. A map f : U → Rm is a submersion
at x ∈ U if the linear map Dfx : Rn → Rm is surjective.

Theorem 1.0.4 (Submersion theorem). Let U0 ⊂ Rn be an open subset, x ∈ U0 and let

g
→ Rn be an open subset. Assume that g : U0 → Rm is a smooth map and is a submersion
U0 −
and m 6 n. Then there exists an open neighborhood U ⊂ U0 containing x, an open V ⊂ Rm
containing g(x) and diffeomorphisms
ψ : Rn → U ϕ : Rm → V
such that
(1) ψ(0) = x,
(2) ϕ(0) = g(x) and a commutative diagram
ψ
Rn U
π g
ϕ
Rm V.
Here, π is the projection map (x1 , · · · , xn ) 7→ (x1 , · · · , xm ).

Here is how we can use Theorem 1.0.4. Let y ∈ g −1 (g(x)); we want to find a diffeomorphism
between an open subset U1 ⊂ g −1 (g(x)) (in the subspace topology) and an open subset of
Rn−m ; this will then assemble (as y varies) into an atlas for g −1 (g(x)). The commutativity of
the diagram in part (2) of Theorem 1.0.4 then says that we have a diffeomorphism
g −1 (g(x)) ∩ g −1 (V) ∼
= g −1 (g(x)) ∩ U ∼
= π −1 (0) = {0} × Rn−m ,
which is the desired diffeomorphism. This provides an easy way to prove that spheres are
smooth.

Example 1.0.5. We have the function g : Rn+1 → R given by (x0 , · · · , xn ) 7→ x20 + · · · + x2n .
Its derivative is given by the matrix


2x0 2x1 · · · 2xn .
This matrix has full rank if and only if (x0 , · · · , xn ) 6= 0. Therefore Theorem 1.0.4 states that
g −1 (c) is a smooth manifold if and only if c 6= 0.
 3. TORI 19

2. Projective spaces
Sn is just one of other possible ways to generalize S1 . An important class of spaces that are
ubiquitous throughout mathematics are the projective spaces. We first introduce real projec-
tive spaces. The real projective space RPn is the space of “lines in Rn+1 through the origin.”
To make this precise, we consider the following equivalence relation on Rn+1 r {0}:
(x0 , · · · , xn ) ∼ (rx0 , · · · , rxn ) r ∈ R r {0}.
The quotient is denoted by
RPn := Rn+1 r 0/ ∼ .
Points in RPn (in other words, equivalence classes) are denoted by [x0 : x1 : · · · : xn ]. We have
the quotient map q : Rn+1 r {0} → RPn and the inverse image over each point is a copy of
R r {0}. It is given the quotient topology: a subset U ⊂ RPn is open if and only if q −1 (U) is
open. One can check easily that it is second countable and Hausdorff.
The interesting part is to write down an smooth atlas for RPn . Consider the open subsets
Ui = {[x0 : x1 : · · · : xn ] : xi 6= 0};
n
then the maps ϕi : Ui → R defined by
ϕi ([x0 : x1 : · · · : xn ]) = (x1 /xi , · · · , x[
i /xi , · · · , xn /xi )

defines the (inverse) of an atlas. We also have an analog for complex space which is discussed
in the homework.

3. Tori
We now discuss tori. To begin with, we note:
Lemma 3.0.1 (Products of manifolds). Let M and N be manifolds of dimension m and n
respectively. Then M × N admits the structure of an m × n-manifold.
Proof. After noting that second countability and Hausdorfness are stable under products
(with the product topology on M × N, we need only equip M and N with an atlas. But now if
{(Uα , Vα , ϕα )} is an atlas for M and {(U0β , Vβ0 , ϕ0β )} is an atlas for N, then we can produce an
atlas for M × N given by {Uα × U0β , Vα × Vβ0 , ϕα × ϕ0β )}.

The n-torus is the n-manifold given by the n-fold product of S1 ’s.
Tn = S1 × · · · × S1
This is yet another way to specify a manifold — by producing them out of old ones. We
want to discuss a construction of the 2-torus which is somewhat “visual” and is useful in practice.
Consider the “box” I × I where I = [0, 1]; its boundary is defined as ∂(I × I) := {(x, y) : x ∈
{0, 1} or y ∈ {0, 1}}. Consider the equivalence relation given by
(0, y) ∼ (1, y) (x, 0) ∼ (x, 1)
0 0 0 0
and (x, y) ∼ (x , y ) if and only if (x, y) = (x , y ) if the points are not on the boundary. In
other words, we are trying to glue the edge of the square to its opposite edge. We can give this
the quotient topology: let us recall that this means, under the map:
q : I × I → I × I/ ∼
a set U ⊂ I × I/ ∼ is open if and only if q −1 (U) is open. Let us endow the quotient with a
smooth structure; tentatively we call this space T2q to indicate that we don’t quite know that it
is diffeomorphic to I × I and we will write elements in it as [(x, y)] to indicate the equivalence
class of (x, y).
There are several flavors of charts:
(Interior) if (x, y) 6∈ q(∂(I × I)) this is the easiest chart: just take a small ball around (x, y).
20 3. EXAMPLES OF SMOOTH MANIFOLDS

(Edges) For [(x, 0)] such that x 6= 0, 1, define

B ((x, 0))(⊂ R2 ) → T2q
given by (
0 0 [(x0 , y 0 + 1)] y0 < 0
(x , y ) 7→
[(x0 , y 0 )] y 0 > 0.
In words, if we are in the “half-moon” where y 0 < 0, translate that point by 1. The
same goes for [(0, y)] where y 6= 0, 1.
(Corner) The most interesting part is the corner: the class [(0, 0)] which contains (0, 0), (1, 0), (1, 1), (0, 1).
The idea is to split up the ball into 4 “quarter-moons.” Explcitly:
B (0, 0) → T2q
 0

 [(x + 1, y 0 + 1)] y 0 < 0, x0 < 0
[(x0 + 1, y 0 )]

x0 < 0
(x0 , y 0 ) 7→ 0 0


 [(x , y + 1)] y0 < 0
 0 0
[(x , y )] else.
What is left to check is that the transition maps are smooth: this is the case because they are
either identity of some translation, which is clearly a smooth function.
 CHAPTER 4

Morphsims of manifolds, tangent spaces and derivatives

We now begin to discuss the following problem:

Question 0.0.1 (Classification problem). Given an integer n > 0, what are all the possible
manifolds of dimension?
It is clear that the only (connected) manifold of dimension 0 is the point. In dimension 1,
we have R and S1 , but how do we know that they are different. In this case, it is quite easy
to tell them apart purely topologically since one is compact and the other is not. To proceed
further, we need to say what we mean for manifolds to be “the same” — evidently we need them
to be the same topologically (aka homeomorphism) and we need preservation of differentiable
structures.

1. Smooth maps of manifolds

Let M and N be manifolds of dimension m and n respectively, with smooth atlases {(Uα , Vα , ϕα )}α∈A
and {(U0β , Vβ0 , ϕ0β )}β∈B respectively. Suppose that f : M → N is a continuous map. Then we
say that f is smooth or a morphism of manifolds if for all α ∈ A and β ∈ B the map
ϕα f ϕ−1
ϕα : ϕ−1 −1
→ Vβ0 −−→ ϕ−1
(Vβ0 )) −−→ Vα ∩ f −1 (Vβ0 ) − 0
β
α (Vα ∩ f β (Vβ )

is a smooth map between open subsets of Euclidean spaces. A smooth map f : M → N is said
to be a diffeomorphism if it is a continuous bijection with a smooth inverse. In this class,
we want to say that two manifolds are the same whenever there is a diffeomorphism between
them. In the homeworks, you will be asked to show that S1 is diffeomorphic to RP1 and S2 is
diffeomorphic to CP1 .
In the next example, we exhibit ways in which we can prove that something is a smooth
map.
Example 1.0.1 (The diagonal). A smooth map that is always defined for any manifold is
the diagonal: as a continuous map of topological spaces we set
∆:M→M×M x 7→ (x, x).
To “officially” prove that ∆ is a smooth map, one needs to pick an atlas on both domain and tar-
get. Of course the atlas on the domain does induce an atlas on the target: if {(Uα , Vα , ϕα )}α∈A
is an atlas then {(Uα × Uβ , Vα × Vβ , ϕα × ϕβ }α,β∈A gives an atlas. However, it turns out that
we do not need to pick an atlas on the target by the next lemma which asserts that smoothness
is a local property.
To use the next lemma, we take a chart (Uα , Vα , ϕα ) containing x and take the chart
(Uα × Uα , Vα × Vα , ϕα ). We are then left to prove that the diagonal map between subsets of
Euclidean spaces:
Uα → Uα × Uα
is smooth.
The proof of the next lemma is completely analogous to the proof of the lemma in topology
that a function is continuous if and only if it is continuous at each point.
21
22 4. MORPHSIMS OF MANIFOLDS, TANGENT SPACES AND DERIVATIVES

Lemma 1.0.2. Let f : M → N be a continuous map of topological spaces. Then f is smooth

if and only if for all x ∈ M there is a chart (Uα , Vα , ϕα ) around x and a chart (U0β , Vβ0 , ϕβ )
around f (x) such that the composite
ϕα f ϕ−1
→ Vβ0 −−→ ϕ−1
ϕα : ϕ−1 (Vα ∩ f −1 (Vβ0 )) −−→ Vα ∩ f −1 (Vβ0 ) − 0
β
β (Vβ )
is smooth.
Another important, albeit somewhat tautological, class of smooth maps are given by open
inclusions.
Lemma 1.0.3. Let M be an n-manfiold, then any open subset U ⊂ M is also an n-manifold.
The inclusion map i : U → M is smooth.
Proof. Note that if {(Uα , Vα , ϕα )} is an atlas for M, then {(Uα ∩ U, Vα ∩ U, ϕα )} is an
atlas for U. The second assertion follows from the fact that if U, V ⊂ Rn are opens such that
U ⊂ V, then the inclusion map i : U → V is smooth.

The next lemma asserts the standard properties of smooth functions:
Lemma 1.0.4. The following holds:
(1) the identity map is smooth;
(2) smooth maps are stable under composition: if f : M → N and g : N → P is smooth,
then g ◦ f : M → P is as well;
(3) Local nature: if {Ui } is an open cover of M (a collection of open subsets U of M) such
that ∪Ui = M. Then f : M → N is smooth if and only if f |Ui : Ui → N is a smooth
map (where we have endowed Ui with the smooth structure as in Lemma 1.0.3.
Proof. Prove it yourself or [Kup, Lemma 4.1.12]. 
1.1. Digression: Lie groups. Now that we know how to define smooth maps of mani-
folds, we can study a large and very interesting class of examples. We note that S1 is actually a
group: it is most easily described using complex coordinates. Note that S1 = {z : |z| = 1} ⊂ C
and so any element in S1 can be denoted by eiθ . The addition law is then described by addition
of angles eiθ · eiϕ = ei(θ+ϕ) . This leads to the definition of a “group in manifolds” also called a
Lie group:
Definition 1.1.1. A Lie group is a group G such that the maps
G×G→G (x, y) 7→ xy
and
ι:G→G x 7→ x−1
are smooth.
It is easy to see that S1 is a Lie group; in fact it is an abelian Lie group: this just means
that the multiplication is commutative. Furthermore, the n-torus Tn is also one since we know
how to take products of groups and manifolds. One of the most important examples of a Lie
group is
Example 1.1.2 (General linear groups). Let GLn (R) be the set of n×n-invertible matrices.
It is a manifold of dimension n2 once we know that opens of a smooth manifold of dimension
n is also one; Lemma 1.0.3. Now, we have the determinant map det : Mn (R) → R which is
a continuous function and GLn (R), as a set, is given by Mn (R) r det−1 (0) and hence is an
2
open subset of Mn (R). The latter is isomorphic to Rn which does have the structure of a
2 2
smooth n -manifold. Therefore GLn (R) is an n -manifold. To check that the multiplication
and inverses are smooth, we note that they can be described as polynomials in the entries which
are smooth functions.
 1. SMOOTH MAPS OF MANIFOLDS 23

Example 1.1.3 (Special linear groups). SLn (R) ⊂ GLn (R) is the subset of invertible ma-
trices with determinant 1. In other words, it is det−1 (1) where det : GLn (R) → R is the
determinant map. Using the submersion theorem, one can show that SLn (R) is smooth of
dimension n2 − 1.
The next example is very important and illustrates the kind of phenomena, in simplified
form, that we want to capture via the general notion of a derivative.
Example 1.1.4. [Orthogonal groups] Let (−, −) : Rn × Rn → R be the standard inner
product on n-space. Explicitly this is given by
(x = (x1 , · · · , xn ), y = (y1 , · · · , yn )) = x1 y1 + x2 y2 + · · · + xn yn .
In matrix multiplication, we can write this as
 
y1
t


x · y = x1 ··· xn · · · · = (x, y).
yn
The orthogonal group is the subset O(n) ⊂ Mn (R) given by those matrices A such that
(x, y) = (Ax, Ay)
n
for all x, y ∈ R . In other words, O(n) is the set of invertible matrices that preserve the inner
product. One should interpret this as the set of matrices that preserve angles and lengths.
Equivalently, this is the set of all matrices whose inverse is the transpose. The set O(n) itself
admits the structure of a manifold of dimension n(n − 1)/2.
Let’s prove this because it illustrates the submersion theorem quite nicely. Let Sn (R) ⊂
Mn (R) be the subset of symmetric matrices: those such that t A = A. Hopefully it is
clear why we call these symmetric matrices. The first claim is that Sn (R) is a vector space
of dimension (n2 − n)/2 + n = n(n + 1)/2 (hint: the first expression gives a hint for a proof).
Indeed: this follows from the standard properties of the transpose T (A + B) =T A +T B and
n(n+1)
T
(rA) = r(T A). Therefore, it is isomorphic to R 2 as a topological space and hence a
smooth manifold. We have the map
t : GLn (R) → Sn (R) A 7→T AA
and t−1 (In ) = O(n). We note that t is actually a map between Euclidean spaces!
2 n(n+1)
t : Rn → R 2 .
We want to prove that the set t is a submersion at all inverse elements in t−1 (In ) in order to
apply the submersion theorem. Hence fix A ∈ t−1 (In ) and we want to say that
2 n(n+1)
DtA : Rn → R 2

has full rank. The best way to compute that the Jacobian matrix has full rank is to straight
up compute it from first principles. We claim that
DtA (B) = BT A +T AB.
Indeed, we note that the following expression does make sense
T
(A+hB)(A+hB)−T AA T
AA+hT AB+hAT B+(h2 )T BB−T AA
lim ( t(A+hB)=t(A)
h = h = h )
h→0

which is equal to BT A +T AB and is the desired linear transformation solving the requirement
for the derivative as discussed in chapter 1. Now, to apply the submersion theorem, we need to
prove that for each C ∈ Sn (R) there indeed exists a B such that
DtA (B) = C.
Using that AA = In by assumption, we see that 12 CA works!
T
24 4. MORPHSIMS OF MANIFOLDS, TANGENT SPACES AND DERIVATIVES

2. The tangent space and the derivative

We now come to a rather tricky part of the course: defining the tangent space at a point
of a manifold and the notion of the derivative of a smooth map between manifolds. The idea is
actually relatively simple: suppose that that our manifolds are embedded in Euclidean space,
or even (secretly) Euclidean space itself (see the example in Example 1.1.4), f : M → N. Then
for each point x ∈ M we have seen that the derivative is a certain linear map
Dfx : Rm → Rn
which can concretely be thought of as a Jacobian matrix. Note that the domain and the range
are all of linear space even though the derivative is “local in nature” in the sense that it is
defined with only reference to open neighborhoods of x. But this makes sense for manifolds
because around x it is supposed to “look like” Euclidean space. Now the vector spaces Rm and
Rn should not be thought of as an “abstract” Euclidean space: they should have something to
do with x and f (x) (even though they will be abstractly isomorphic to Rm and Rn respectively.
So we should write something like
Dfx : Tx M → Tf (x) N.
Furthermore, Dfx should look and feel like the derivative.
It is easy to be lost in the formalism of tangent spaces, but the desiderata is very clear and
so we will state them here:
(1) to each x ∈ M we attach a R-vector space Tx M;
(2) a chart around X should make an identification between Tx M and Rn where n is the
dimension of M — in particular this identification should rely in a meaningful way to
the smooth structure of M around x;
(3) as x varies across M, the “assignment” x 7→ Tx M should be continuous in some precise
way;
(4) Dfx is a R-linear map;
(5) if f is linear (around x) then Dfx is just f again;
(6) if f is the identity then Dfx is the identity linear transformation;
(7) then Dfx is the identity and we should have a chain rule D(g ◦ f )x = Dgf (x) ◦ Dfx .
For practical purposes I have just outlined what the derivative should be and that is most
of how you will interact with the derivative. But, of course, we should go ahead and define it
properly. We will take the “algebraic approach.”
2.1. Germs of functions. Recall that, in the definition of a smooth manifold, we are
equipping a topological manifold with a “smooth structure.” This is made precise by saying
that the transition functions between charts are smooth — hence R-valued smooth functions
from one chart are all preserved. This is the starting point of one approach to the tangent
space: whatever the tangent space is, maps between manifolds must induce a linear function,
the derivative, between them. So the tangent space is the “domain” of derivatives and so must
be some kind of “ space of functions” since derivatives take in functions. To make sense of this
we define germs.
Definition 2.1.1. Let M, N be a manifolds and x ∈ M be fixed. We define an equivalence
relation on the set
{(f, U) : U is an open neighborhood of x, f : U → N is a smooth function}.
where f ∼ g if and only if there exist a V, an open neighborhood of x such that f |V = g|V .
The equivalence class of f will be denoted by f : (M, x) → N and is called a germ or, more
precisely, a germ at x. We also sometimes want to emphasize that f (x) = y and write f :
(M, x) → (N, f (x)) = y for a function germ. A function germ (at x) is a germ f : (M, x) → R
and is denoted by
OM,x := {(f, U) : U is an open neighborhood of x, f : U → R is a smooth function}/ ∼ .
 2. THE TANGENT SPACE AND THE DERIVATIVE 25

Remark 2.1.2 (Germs and sheaves). There is a more “global” notion of germs: to each
chart (U, V, ϕ) we can associate the R-algebra
C∞ (U) =: C∞ (V),
and if (U0 , V0 , ϕ0 ) is another chart such that V ⊂ V0 then we have a map of R-algebras:
C∞ (V) → C∞ (V0 )
given by
ϕ0 ϕ−1 f
f : U → R 7→ U0 −→ V0 ⊂ V −−→ U −
→ R;
we write the latter function as f |V0 . In these terms, a function germ around a point x ∈ M is,
informally, an equivalence class of a function f ∈ C∞ (V) for V containing x and we identify f
with f |U0 whenever V0 ⊂ V as long as x ∈ V0 . The language of sheaves and stalks, which we
will not cover in this class, makes all of this precise.
The following is easy to verify:
Lemma 2.1.3. Let M, N, P be manifolds.
(1) Composition of germs: L let f : (M, x) → (N, f (x)) and g : (N, g(x)) → (P, (g ◦ f )(x))
be function germs. Then the composite, defined by
g ◦ f := g ◦ f
is well-defined.
(2) R-vector space structure: for f, g ∈ OM,x , the operations
f + g := f + g rf := λf
defines a R-vector space structure on OM,x .
(3) R-algebra structure: with the notation as above, multiplication
f · g := f g
defines a R-algebra structure on OM,x .
Construction 2.1.4 (Induced map on germs). Let f : M → N be a smooth map of
manifolds and x ∈ M, the induced map on germs is the map
f ∗ : ON,f (x) → OM,x g 7→ g ◦ f .
Concretely, f ∗ takes the equivalence class of (U, g) where g : U → R and f (x) ∈ U to the
f| 0
equivalence class of g ◦ f : U0 −−−
U
→ U → R where U0 is a neighborhood of x. It is easy to check
that this is, in fact, a R-algebra map.
Notice that the above elaboration on f ∗ only depends on open neighborhoods around x
and f (x). Hence, more generally, if we have a germ f : (M, x) → (N, f (x)) then we have the
induced map on function germs
f ∗ : ON,f (x) → OM,x g 7→ g ◦ f ,
given by the same formula.
We will not really use the R-algebra structure seriously, though we will use that this is a
map of R-vector spaces. The next proposition, albeit tautological, is key to the theory.
Lemma 2.1.5 (Local nature of germs). Let M be a manifold of dimension n and (Uα , Vα , ϕα )
be a chart around x such that ϕα (0) = x. Then there is an induced isomorphism of R-vector
spaces
ϕ∗
OM,x −−→
α
ORn ,0 .
26 4. MORPHSIMS OF MANIFOLDS, TANGENT SPACES AND DERIVATIVES

Proof. In fact, Construction 2.1.4 satisfies the following: suppose that we have germs
f : (M, x) → (N, f (x)) and (N, f (x)) → (P, g(f (x))) we have precompositions
f∗
ON,f (x) −→ OM,x
such that: 1) f ∗ is an R-linear map, 2) id∗ = id and 3) (g ◦ f )∗ = f ∗ ◦ g ∗ . Hence, we deduce
that (ϕ−1 ∗ ∗
α ) furnishes an inverse to ϕα . 
We will use Lemma 2.1.5 repeatedly to reduce a lot of results about tangent spaces to results
about Rn . We now define the tangent space; we explain why this is a reasonable definition later.
Now, recall that we want a map f : M → N to induce a map Tx M → Tf (x) N of tangent spaces.
As explained in Construction 2.1.4, function germs compose in the other direction:
ON,f (x) → OM,x g : (N, f (x)) → R g ◦ f : (M, x) → R.
So we might guess that Tx M is the dual space of R-linear functionals O∨
M,x := HomR (OM,x , R).
This fixes the problem as we will now have a map
O∨
M,x → ON,f (x) .
∨

However, this is not quite what we want: the space O∨M,x is way too big and is not even finite
dimensional. Remember that we want this to just be a Rn where n = dim(M). The right thing
to do is to only consider “germs up to the first order.” This is captured by the notion of a
derivation.
Definition 2.1.6. A derivation (at x ∈ M) is a R-linear map
X : OM,x → R
satisfying the “Liebniz rule”
X(f g) = X(f )g(x) + f (x)X(g(x)).
The tangent space Tx M is the set {X : OM,x → R : X is a derivation}.
A derivation takes in a germ and spits out a number; hence we should think of as a certain
operation on functions which only depends on its germ. Furthermore, these operations only
really depend on “first order information” of f . This is what the Liebniz rule says but the best
way to see this is via the proof of Lemma 2.1.10.
Example 2.1.7. Let c be the germ of a constant funtion, then we claim that X(c) = 0. By
linearity, we may assume that c = 1, the constant function with value 1. Then:
X(1) = X(1 · 1) = 2X(1)
which means that X(1) = 0.
Example 2.1.8. Consider T0 Rn . Then, for each i = 1, · · · , n we have the derivation
∂
∂xi : ORn ,0 → R f 7→ Di f (0).
These are the most important kinds of derivation and they are linearly independent by evalu-
ating them on the germs of the projections: πi : Rn → R, (x1 , · · · , xn ) 7→ xi .
Example 2.1.9. Let f , g : (M, x) → R be function germs such that f (x) = g(x) = 0. Then
X(f g) = f (x)X(g) + g(x)X(f ) = 0 + 0 = 0. Hence X annihilates the product of functions that
vanish around x. We will use this result later.
The next lemma is a key result which verifies that Tx M is indeed what we want.
Lemma 2.1.10. The tangent space Tx M is a R-linear vector space of dimension dim(M).
More precisely: let (U, V, ϕ) be a chart around x such that ϕ(0) = x and thus induces an isomor-
ϕ∗
phism OM,x −−→ ORn ,0 as in Lemma 2.1.5. Then taking R-linear dual induces an isomorphism
∼
=
T 0 Rn −
→ Tx M
 2. THE TANGENT SPACE AND THE DERIVATIVE 27

and the map

R n → T0 Rn ei 7→ ∂
∂xi

is an isomorphism (here {ei } is the standard basis).

Proof. We prove this in a few steps; for this proof we suppress the bar’s above germs and
declare that all functions should be thought of as their germs at x.
(1) The dual space is a vector space is a vector space. We claim that Tx M is a sub-vector
space of O∨
M,x . To do so we just need to check the “Liebniz rule property” is closed
under addition and scalar multiplication. For addition: on the one hand we have
(X + Y)(f g) = X(f g) + Y(f g) = X(f )g(x) + X(g)f (x) + Y(f )g(x) + Y(g)f (x),
on the other
((X + Y)(f ))g(x) + ((X + Y)(g))f (x) = X(f )g(x) + Y(f )g(x) + X(g)f (x) + Y(g)f (x).
Scalar multiplication is similar and left to the reader.
(2) By Lemma 2.1.5 we have that OM,x ' ORn ,0 for some chart, normalized such that
ϕα (0) = x. Therefore we have an isomorphism of R-vector spaces
∼
=
T0 R n −
→ Tx M.
So we are left to prove the result for T0 Rn . Here we can try to write down a basis for
this vector space. Consider the following linear functionals
∂
∂xi : ORn ,0 → R f 7→ Di f (0)(= ∂f
∂xi (0)).

These linear functionals are derivations by the usual Liebniz rule and are linearly
independent. We prove that they span.
(3) Let us write xi : Rn → R for the i-th projection map and we think about it as the
germ at 0. Let X be derivation, our claim is that
n
X
∂
X= X(xi ) ∂x i
.
i=1

Let’s unpack what this means: this means that for all function germ f , X(f ) is the
sum of terms
∂f
X(xi ) ∂x i
(0) = X(xi )Di f (0).
where xi is the germ of the projection function to the i-variable and Di f (0) is a
number: it is the value of the derivation explained in Example 2.1.8.
(4) To do so, recall that we have Taylor’s theorem with remainder: we can write any
smooth function f locally around 0 (in other words, valid for x ∈ U0 a small enough
neighborhood of 0) as
n
X
∂f
f (x = (x1 , · · · , xn )) = f (0) + ∂xi (0)xi + R2 (x)
i=1

where R2 (x) is the remainder term:

n Z 1
X 2
R2 (x) = xi xj (1 − t) ∂∂xfi ∂x
(tx)
j
dt.
i,j 0

It almost doesn’t matter what R2 (x) is —- the only thing we ever need is to be able
to know that each summand can be written as a product of two functions vanishing
at 0 so as to invoke Example 2.1.9.
28 4. MORPHSIMS OF MANIFOLDS, TANGENT SPACES AND DERIVATIVES

(5) Now, we calculate:

n
X
∂f
X(f ) = X(f (0)) + X( ∂xi (0)xi ) + X(R2 )
i=1
n
X
∂f
= X( ∂xi (0)xi )
i=1
n
X
∂f
= ∂xi (0)X(xi ).
i=1

The first equality comes from plugging in the Taylor expansion and the linearity of X,
the second equality comes from Examples 2.1.7 and 2.1.9 and the last equality comes
from linearity.


Lemma 2.1.10 tells us that, in practice, the tangent space at a point x ∈ M is an n-

∂
dimensional vector space whose bases are given by the linear functions ∂x i
. This is useful for
performing explicit calculations. As promised, we define the tangent space in order to get to
the derivative of a smooth function between smooth manifolds; more generally, the germ of a
smooth function. The definition is relatively easy: first given a germ
f : (M, x) → (N, f (x))
we get an induced map (in the opposite direction) on function germs
f ∗ : ON,f (x) → OM,x
which induces a map on functionals
(f ∗ )∨ := f∗ : O∨
M,x → ON,f (x)

We note that f∗ preserves derivation, i.e., if X is a derivation then f∗ X is a derivation. Indeed:

if g, g 0 : (N, f (x)) → R are function germs then
(f∗ X)(gg 0 ) = X(gg 0 ◦ f ) = X(g ◦ f · g 0 ◦ f ) = g(f (x))X(g 0 ◦ f ) + g 0 (f (x))X(g ◦ f ).
We define the derivative this way.
Construction 2.1.11 (Induced maps on tangent spaces). Let f : M → N be a smooth
map of manifolds and let x ∈ M. Then the derivative at x is given by
Dfx : Tx M → Tf (x) N
given by
X 7→ f∗ X = (g ∈ ON,f (x) 7→ X(g ◦ f )).
We verify the properties of the derivation.
Lemma 2.1.12. Let f : M → N be a smooth map of manifolds and let x ∈ M, then
Dfx : Tx M → Tf (x) N
satisfies the following properties:
(1) the map Dfx is an R-linear map;
(2) if f = id, then Didx = id;
(3) if g : N → P is another smooth map then
D(g ◦ f )x = Dgf (x) ◦ Dfx .
 2. THE TANGENT SPACE AND THE DERIVATIVE 29

(4) Let (U, V, ϕ) and (U0 , V0 , ϕ0 ) be charts around x and f (x) respectively such that ϕ(0) =
x, ϕ0 (0) = f (x), then the composite
Dfx
Tx M Tf (x) N
Dϕ0 ,∼
= (Dϕ00 )−1 ,∼
=

T0 Rm T 0 Rn

coincides with the total derivative.

Proof. Properties (1)-(3) all hold for the composition
f∗ : O ∨
M,x → ON,f (x)
∨

by the formal properties of taking duals of linear maps. Thus, it holds for Dfx because f∗
preserves derivatives. Now we explicitly want to compute the composition
(Dϕ00 )−1 ◦ Dfx ◦ Dϕ0 = (ϕ0−1
0 ◦ f ◦ ϕ0 )∗ .
where the equality follows from (1)-(3). So let us rewrite
g := ϕ00−1 ◦ f ◦ ϕ0 : (Rm , 0) → (Rn , 0)
∂
as g = (g 1 , · · · , g n ) where each g i : Rm → R. Since we have already shown that ∂xj for
j = 1, · · · m forms a basis for T0 Rm we need only calculate
∂
g∗ ( ∂x j
)
∂
in terms of the bases ∂yi for i = i, · · · , n. In fact, to prove that g∗ is indeed the usual total
∂ ∂g i
derivative, we need to know the coefficient corresponding to ∂yi is exactly the number ∂xj (0);
∂ ∂
in the basis ( ∂x j
)j=1,··· ,m for the source and ( ∂y i
)i=1,···n we will have written out the Jacobian
of g.
So take a function germ h : (Rn , 0) → R. Then, for any fixed j ∈ {1, · · · m} we have
∂ ∂(h◦g)
g∗ ( ∂x j
)(h) = ∂xj
n
∂g j
X
∂h
= ∂yi (0) ∂xj (0),
i=1

where the first equality is by definition and the second equality comes from the multivariable
chain rule. Since this is true for all function germ h, we conclude that the coefficient the term
∂
∂yi is as desired. 

Remark 2.1.13 (Tangent space as curves). We remark on another, equivalent way of defin-
ing the tangent space. It is somehow more “visual” and evokes the idea that a tangent vector
should be a point on the manifold and a “vector pointing parallel to it.” Let (−1, 1) := I be
the open interval and we think of it as a smooth manifold. A smooth curve around x ∈ M
(or simply a curve if the context is clear) is a smooth map γ : I → M such that γ(0) = x.
A smooth curve in M can be rather wild, but we are only interested in the first order
information derived from γ. To do so we associate to a curve γ a derivation given by:
Xγ : OM,x → R g 7→ (g ◦ γ)0 (0).
In other words, we look at g ◦ γ : I → R and just take its derivative at 0. By the usual Liebniz
and chain rules, this is a derivation. Set
Γx M := {γ : γ is smooth curve around x}
30 4. MORPHSIMS OF MANIFOLDS, TANGENT SPACES AND DERIVATIVES

and define an equivalence relation by

γ ∼ γ 0  (f ◦ γ)0 (0) = (f ◦ γ)0 (0)∀f ∈ OM,x .
Then one can check that Vx M := Γx M/ ∼ is a R-vector space and that the map
Vx M → T x M γ 7→ Xγ
is an isomorphism of R-vector spaces.
2.2. Vector bundles and tangent bundles. We now want to assemble the tangent
spaces into a global object. The idea of the tangent space is simple once we have granted
ourselves the Whitney embedding theorem or we had started the class with manifolds embedded
in Euclidean space. Let M be a manifold and suppose that it comes as a subset of RN : so we
have M ⊂ RN . Informally speaking, Tx M is the n-dimensional affine plane in RN through x
which is the best linear approximation. So Tx M is the translate of a n-dimensional subspace of
RN . With this definition, we set
TM := {(x, v) : v ∈ Tx M} ⊂ RN ,
and one can show that this subspace of RN is a smooth manifold.
Example 2.2.1. Let X = Sn embedded in Rn+1 . Then the tangent space at x ∈ Sn is given
by
Tx Sn ∼
= {x + v : v ⊥ x} ⊂ Rn+1 .
Since we will not really use the above “embedded” definition of the tangent space, we will
not bother making the idea precise. Instead, we will make the tangent bundle precise. To do
so, we will put in the context of the general theory of vector bundles
Definition 2.2.2. Let X be a space. Then a real vector bundle of rank n is a topological
space E equipped with a surjective continuous map p : E → X subject to the following condition:
(1) the fibre p−1 (x) has the structure of an n-dimensional R-vector space;
(2) for each x ∈ M there exists a neighborhood U of x and a homeomorphism
∼
=
Φ : p−1 (U) −
→ U × Rn
such that: pU ◦ Φ = p where pU : U × Rn → U is the projection;
(3) for each p ∈ U, the restriction of Φ to
Φ : p−1 (x) → {x} × Rn
is a R-linear isomorphism.
If X and E are smooth manifolds and p is a smooth function, then we say that q is a smooth
vector bundle.
Some terminology surrounding Definition 2.2.2 we say that E is the total space of the
vector bundle and p is the projection.
Example 2.2.3. Let E := X × Rn , then the projection onto X map E → X is a vector
bundle. If X is a smooth manifold of dimension m, then E is a smooth manifold of dimension
n + m. This bundle is called the trivial bundle or sometimes called the product bundle.
Example 2.2.4. Consider X = RPn ; a point x = [x0 : · · · : xn ] defines a line, which we
denote by x ⊂ Rn+1 by considering the span of (x0 , · · · , xn ); in this way we think of X as the
space of lines in n + 1-Euclidean space. Consider
Tautn := {(x, v) : v ∈ x} ⊂ X × Rn+1 .
pX
Then Tautn ⊂ X × Rn+1 −−→ X is a smooth vector bundle of rank 1; this is called the tauto-
logical bundle on real projective space.
Now, we construct the most important example of a vector bundle.
 2. THE TANGENT SPACE AND THE DERIVATIVE 31

Construction 2.2.5 (The tangent bundle). Let M be a manifold of dimension n. Begin

by considering the set
TM := tx∈M Tx M.
Elements can be thought of as pairs (x, v) where x ∈ M and v ∈ Tx M. We have the map
p : TM → M that sends (x, v) to x.
(1) For each x ∈ M we take a chart (U, V, ϕ). What we then do is define
n
X
e : p−1 (V) → R2n
ϕ (x, v = v i ∂x
∂
i
) 7→ (ϕ−1 (x), v i ).
i=1
−1
The image is given by ϕ (V) × R = U × Rn which is an open subset of R2n . Note
n

that this is a bijection onto its image by using ϕ.

(2) Given two charts (U, V, ϕ) and (U0 , V0 , ψ 0 ) we have diffeomorphisms
U ∩ U0 × Rn = ϕ(p
e −1 (V) ∩ p−1 (V0 )) → (ϕe0 )(p−1 (V0 ) ∩ p−1 (V)) = U0 ∩ U × Rn
where the map is given by ψe ◦ ϕe−1 .
(3) Now, pick a countable atlas {(Ui , Vi , ϕi )} of M and consider the data of Ui ×Rn ⊂ R2n ,
and for each i, Uij := (Ui ∩ Uj ) × Rn and the maps Uij → Uji as defined in part (2).
One checks that this satisfies the condition of the “cocyles” explained in homework 2.
Hence TM assembles into a manifold of dimension 2n.
(4) Locally the function p is modeled by the projection U × Rn → U and thus it is smooth.
Lastly, we note the following “functoriality” property of the tangent bundle which follows
easily from the construction of the tangent bundle.
Lemma 2.2.6. Let f : M → N be a map of smooth manifolds, then there is a smooth map
Df making the following diagram commute:
Df
TM TN

f
M N
such that, for each x ∈ X the map
Tx M → Tf (x) N
is the map of Lemma 2.1.12.
Proof. Note that the commutativity of the diagram says that Df (x, v) = (f (x), Dfx (v))
so the map Df is constructed this way by defining it on the underlying set and checking that it
glues together to a map Df : TM → TN. The only thing we really need to explain is why Df
is smooth. The point here is that
∂f
Df (x, v = (v1 , · · · , vn )) = (f (x) = (f 1 (x), · · · , f n (x)), ∂x i
(x)vi ).
which is evidently smooth since f was. 
It then follows that the following global version of Lemma 2.1.12 holds:
Corollary 2.2.7. The map
Df : TM → TN
satisfies the following properties:
(1) if f = id, then Df = id;
(2) if g : N → P is another smooth map then
D(g ◦ f )x = Dg ◦ Df
coincides with the total derivative.
32 4. MORPHSIMS OF MANIFOLDS, TANGENT SPACES AND DERIVATIVES

3. Application: globalizing calculus

We now see some of the formalism above in action. To do so, we have the following
terminology which makes things convenient o state.
Definition 3.0.1. Let f : M → N be a smooth map of manifolds. We say that:
(1) x ∈ M is a regular point of f if Dfx is surjective;
(2) y ∈ N is a regular value if any x ∈ f −1 (y) is a regular point;
(3) x is a critical point if Dfx is not surjective and y = f (x) is a critical value in this
case.
For the rest of this section, when we write M we always mean a smooth m-dimensional
manifold and write N for a smooth n-dimensional manifold. We will prove the following three
theorems.
Theorem 3.0.2 (Global inverse function theorem). Let f : M → N be a bijective smooth
map such that for all x ∈ M, the derivative Dfx is bijective. Then f is a diffeomorphism.
Theorem 3.0.3 (Global submersion theorem). Let f : M → N be a smooth map and y ∈ N
a regular value, then f −1 (y) is a smooth m − n-dimensional manifold.
Theorem 3.0.4 (Stack of records theorem). Let f : M → N be a smooth map and assume
that dim(M) = dim(N). Further assume that M is compact. For any y ∈ N which is a regular
value, then:
(1) f −1 (y) is the finite set, f −1 (y) = {x1 , · · · , xn } ⊂ X.
(2) there is an open neighborhood U of y such that f −1 (U) = V1 t · · · t Vn with xi ∈ Vi
and f |Vi : Vi → U is a diffeomorphism.
The three theorems have the same theme: “define globally and prove locally.” This will be
evident in the proofs.
3.1. Proofs of theorems. We begin with Theorem 3.0.2.
Lemma 3.1.1. Let f : M → N be a smooth map and x ∈ M. Assume that Dfx : Tx M →
Tf (x) N is bijective. Then there exists an open neighborhood V of x such that f |V : V → f (V)
is a diffeomorphism.
Proof. 
Proof of Theorem 3.0.2. 
3.1.2. Milnor’s proof of the fundamental theorem of algebra. We now see how the concepts
that we have studied so far can be used synchronously to reprove the fundamental theorem of
algebra.
3.2. Sard’s theorem.
 CHAPTER 5

The Whitney embedding theorem

33
APPENDIX A

35
 Bibliography

[Kup] A. Kupers, Lectures on differential topology, http://www.utsc.utoronto.ca/people/kupers/teaching/

differential-topology-lecture-notes/
[Mil56] J. Milnor, On manifolds homeomorphic to the 7-sphere, Ann. of Math. (2) 64 (1956), pp. 399–405,
https://doi.org/10.2307/1969983
[Spi65] M. Spivak, Calculus on manifolds. A modern approach to classical theorems of advanced calculus, W.
A. Benjamin, Inc., New York-Amsterdam, 1965

37