Professional Documents
Culture Documents
Topology Polytechnique
Topology Polytechnique
Kleber Carrapatoso
[email protected]
Bachelor of Sciences
École Polytechnique
2023 – 2024
Contents
4 Completeness 21
4.1 Cauchy sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.2 Complete metric spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.3 Uniformly continuous and Lipschitz maps . . . . . . . . . . . . . . . . . 23
4.4 Fixed-point theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.5 Baire’s category theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5 Compactness 25
5.1 Open covers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.2 Compact spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.3 Compact metric spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.4 Product of compact spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.5 Local compactness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
6 Connectedness 31
6.1 Connected spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
6.2 Path-connected spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
7 Function spaces 35
7.1 Pointwise and uniform convergences . . . . . . . . . . . . . . . . . . . . 35
7.2 Space of continuous functions on a compact set . . . . . . . . . . . . . . 36
7.3 Equicontinuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
7.4 Arzelà-Ascoli theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
7.5 Stone-Weierstrass theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 39
iii
iv Contents
9 Differentiable maps 55
9.1 Fréchet-differentiable maps . . . . . . . . . . . . . . . . . . . . . . . . . . 55
9.2 Operations on differentiable maps . . . . . . . . . . . . . . . . . . . . . . 57
9.3 Maps with values in a product space . . . . . . . . . . . . . . . . . . . . . 59
9.4 Partial derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
12 Optimization 75
12.1 Extremum points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
12.2 First-order conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
12.3 Optimization with constraints . . . . . . . . . . . . . . . . . . . . . . . . 76
12.4 Second-order conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Remark 1.2. Axiom (i) can actually be deduced from the others, and it was added in
the definition for the sake of clarity. Indeed, with our convention, taking an empty
intersection in (ii) implies that 𝑋 ∈ 𝒪; and taking an empty union in (iii) implies that
∅ ∈ 𝒪.
Remark 1.3. One can verify that (ii) is equivalent to: “𝑋 ∈ 𝒪 and every intersection of
two elements of 𝒪 is an element of 𝒪”.
Elements of 𝑋 are called points. Elements of 𝒪 are called open sets in 𝑋 (with respect
to the topology 𝒪). We often denote by 𝑋 the couple (𝑋 , 𝒪).
Definition 1.5. Let 𝒪1 and 𝒪2 be two topologies on a set 𝑋 . If 𝒪1 ⊆ 𝒪2 one says that
𝒪1 is coarser than 𝒪2 , or that 𝒪2 is finer than 𝒪1 . If moreover 𝒪1 ≠ 𝒪2 , then 𝒪1 is said to
be strictly coarser than 𝒪2 , or 𝒪2 is said to be strictly finer than 𝒪1 .
Example 1.1. The collection {∅, 𝑋 } is a topology on any set 𝑋 , called the trivial topology.
Example 1.2. The powerset 𝔓(𝑋 ) of 𝑋 is a topology on any set 𝑋 , called the discrete
topology. A topological space endowed with the discrete topology is called a discrete
space.
1
2 Chapter 1. Topological and metric spaces
Example 1.3. The standard topology on R is given by the collection of arbitrary unions of
open intervals. We shall always consider this topology on R (unless otherwise specified).
Example 1.4. (Extended real number line) Consider the set R ∶= R ∪ {−∞, +∞}, where
+∞ and −∞ are two sets that do not belong to R. This set is equipped with the topology
given by the collection of unions of sets of the form {−∞} ∪ (−∞, 𝑎), (𝑏, +∞) ∪ {+∞} and
(𝑎, 𝑏), with 𝑎, 𝑏 ∈ R.
Example 1.5. Let (𝑋 , 𝒪) be a topological space and 𝑌 ⊆ 𝑋 a subset. The collection of
subsets {𝑈 ∩ 𝑌 ∶ 𝑈 ∈ 𝒪} is a topology on 𝑌 , called the subspace topology.
Example 1.6. Let 𝑋 be a set, 𝑌 a topological space and 𝑓 ∶ 𝑋 → 𝑌 a map. The collection
of subsets {𝑓 −1 (𝑈 ) ∶ 𝑈 open in 𝑌 } is a topology on 𝑋 , called the preimage topology.
Example 1.7. The intersection of an arbitrary family of topologies on a set is a topology
on this set.
Remark 1.7. If (i) is replaced by the weaker assumption “𝑑(𝑥, 𝑥) = 0”, then 𝑑 is called a
pseudo-metric.
Remark 1.8. If 𝑑 satisfies 𝑑(𝑥, 𝑧) ⩽ sup(𝑑(𝑥, 𝑦), 𝑑(𝑦, 𝑧)) for any 𝑥, 𝑦, 𝑧 ∈ 𝑋 , which is
stronger than the triangle inequality, then we say that 𝑑 is an ultrametric and (𝑋 , 𝑑) an
ultrametric space.
Remark 1.9. For any 𝑥, 𝑦, 𝑧 ∈ 𝑋 we also have |𝑑(𝑥, 𝑧) − 𝑑(𝑦, 𝑧)| ⩽ 𝑑(𝑥, 𝑦).
Let (𝑋 , 𝑑) be a metric space. For any 𝑥 ∈ 𝑋 and 𝑟 > 0, we define the open ball centred
at 𝑥 of radius 𝑟 by
𝐵(𝑥, 𝑟) ∶= {𝑦 ∈ 𝑋 ∣ 𝑑(𝑥, 𝑦) < 𝑟},
the closed ball centred at 𝑥 of radius 𝑟 by
Example 1.8. On any non-empty set 𝑋 , we can define the distance 𝑑(𝑥, 𝑦) = 1 if 𝑥 ≠ 𝑦
and 𝑑(𝑥, 𝑦) = 0 if 𝑥 = 𝑦. This is called the discrete distance.
Example 1.9. The set of real numbers R (resp. complex numbers C) is endowed with the
metric 𝑑(𝑥, 𝑦) = |𝑥 − 𝑦| for any 𝑥, 𝑦 in R (resp. in C) where | ⋅ | denotes the absolute value
(resp. the modulus).
Example 1.10. If (𝑋 , 𝑑) is as metric space and 𝑌 is a subset of 𝑋 , we define the map
𝑑𝑌 ∶ 𝑌 × 𝑌 → R+ as the restriction of 𝑑 to 𝑌 × 𝑌 . Then 𝑑𝑌 is a metric on 𝑌 and hence
(𝑌 , 𝑑𝑌 ) is a metric space, called a (metric) subspace of (𝑋 , 𝑑).
Example 1.11. Any normed vector space (𝐸, ‖ ⋅ ‖) is a metric space with the distance
𝑑(𝑥, 𝑦) ∶= ‖𝑥 − 𝑦‖, for any 𝑥, 𝑦 ∈ 𝐸.
Example 1.12. Let (𝑋 , 𝑑𝑋 ) and (𝑌 , 𝑑𝑌 ) be metric spaces. On the product set 𝑋 × 𝑌 we can
define the following distances: for any (𝑥, 𝑦), (𝑥 ′ , 𝑦 ′ ) ∈ 𝑋 × 𝑌
and
𝑑∞ ((𝑥, 𝑦), (𝑥 ′ , 𝑦 ′ )) ∶= max (𝑑𝑋 (𝑥, 𝑥 ′ ), 𝑑𝑌 (𝑦, 𝑦 ′ )).
𝒪𝑑 ∶= {𝑈 ⊆ 𝑋 ∣ ∀𝑥 ∈ 𝑈 , ∃𝑟 > 0, 𝐵(𝑥, 𝑟) ⊆ 𝑈 },
(ii) equivalent if there exist constants 𝑐1 , 𝑐2 > 0 such that for any 𝑥, 𝑦 ∈ 𝑋 one has
2.1 Neighborhoods
Let 𝑋 be a topological space, 𝐴 ⊆ 𝑋 a subset and 𝑥 ∈ 𝑋 a point.
Definition 2.2. A neighborhood basis or local basis for 𝑥 (resp. 𝐴) is a family 𝒮 of neigh-
borhoods of 𝑥 (resp. 𝐴) such that every neighborhood of 𝑥 (resp. 𝐴) contains an element
of 𝒮 .
Example 2.1. Consider a point 𝑥 in R endowed with its standard topology. Each of the
following families are neighborhood basis for 𝑥:
{(𝑥 − 𝑛1 , 𝑥 + 𝑛1 ) ∶ 𝑛 ∈ N∗ }, {[𝑥 − 𝑛1 , 𝑥 + 𝑛1 ] ∶ 𝑛 ∈ N∗ }.
One easily verifies the following properties:
Proposition 2.4. Let 𝑋 be a topological space. For any 𝑥 ∈ 𝑋 , the neighborhood system
𝒱 (𝑥) satisfies:
5
6 Chapter 2. Neighborhoods, basis, interior and closure
𝒪 = {𝑈 ⊆ 𝑋 ∣ ∀𝑥 ∈ 𝑈 , 𝑈 ∈ 𝒱 (𝑥)}.
𝒱 (𝑥) = {𝑉 ⊆ 𝑋 ∣ ∃𝐵 ∈ 𝒮 (𝑥), 𝐵 ⊆ 𝑉 }.
In view of this, one can also define a topology on a set 𝑋 by prescribing the sets 𝒮 (𝑥) of
neighborhood basis for all points of 𝑋 , provided that, for any 𝑥 ∈ 𝑋 , 𝒮 (𝑥) is non-empty
and verifies:
2.2 Basis 7
2.2 Basis
Let (𝑋 , 𝒪) be a topological space.
Definition 2.6. A family ℬ ⊆ 𝒪 of open sets in 𝑋 is a basis for the topology 𝒪 if every
open set in 𝑋 can be written as a union of elements of ℬ.
Proposition 2.7. A subset ℬ ⊆ 𝒪 is a basis for the topology 𝒪 if and only if for any open
set 𝑈 ∈ 𝒪 and any 𝑥 ∈ 𝑈 , there exists 𝐵 ∈ ℬ such that 𝑥 ∈ 𝐵 ⊆ 𝑈 .
Proof. Suppose that ℬ is a basis for 𝒪, namely that every open set 𝑈 ∈ 𝒪 is a union of
elements of ℬ. Let 𝑈 ∈ 𝒪 and 𝑥 ∈ 𝑈 . Then there is ℬ1 ⊆ ℬ such that 𝑈 = ⋃𝐵∈ℬ1 𝐵,
therefore there is 𝐵 ∈ ℬ1 such that 𝑥 ∈ 𝐵 ⊆ 𝑈 .
Conversely, let ℬ ⊆ 𝒪 satisfy
∀ 𝑈 ∈ 𝒪, ∀ 𝑥 ∈ 𝑈 , ∃ 𝑉 ∈ ℬ, 𝑥 ∈ 𝑉 ⊆ 𝑈 .
Let 𝑈 ∈ 𝒪, then for each 𝑥 ∈ 𝑈 , there is 𝐵𝑥 ∈ ℬ such that 𝑥 ∈ 𝐵𝑥 ⊆ 𝑈 . This implies that
𝑈 = ⋃𝑥∈𝑈 𝐵𝑥 , that is, 𝑈 is a union of elements of ℬ and hence ℬ is a basis for 𝒪.
Example 2.2. In a metric space (𝑋 , 𝑑), the family ℬ = {𝐵(𝑥, 𝑟) ∶ 𝑥 ∈ 𝑋 , 𝑟 > 0} is a basis
for the topology associated to the metric 𝑑.
Then the family 𝒪 of arbitrary unions of elements of ℬ coincides with the topology gener-
ated by ℬ.
Proof. One easily observes that if 𝒪 is a topology then it coincides with the topology
generated by ℬ. Therefore, we only need to show that 𝒪 is a topology.
It is clear that ∅ ∈ 𝒪 and 𝑋 ∈ 𝒪. Let 𝒰 be a family of elements of 𝒪, then each
member of 𝒪 is a union of elements of ℬ, therefore the union ⋃ 𝒰 is also a union of
elements of ℬ. Let 𝑈 , 𝑉 ∈ 𝒪, then there is ℬ1 , ℬ2 ⊆ ℬ such that 𝑈 = ⋃ ℬ1 and
𝑉 = ⋃ ℬ2 . Therefore
𝑈 ∩ 𝑉 = ( ⋃ 𝐴) ∩ ( ⋃ 𝐵) = ⋃ 𝐴 ∩ 𝐵,
𝐴∈ℬ1 𝐵∈ℬ2 𝐴∈ℬ1 , 𝐵∈ℬ2
and the proof will be complete if we show that any intersection of two elements of ℬ is
a union of elements of ℬ. Consider then 𝐴, 𝐵 ∈ ℬ. For each 𝑥 ∈ 𝐴 ∩ 𝐵, there is 𝑊𝑥 ∈ ℬ
such that 𝑥 ∈ 𝑊𝑥 ⊆ 𝐴 ∩ 𝐵. Thus 𝐴 ∩ 𝐵 = ⋃𝑥∈𝐴∩𝐵 𝑊𝑥 , which completes the proof.
It is easy to verify that 𝐴̊ is the union of all open sets contained in 𝐴, therefore also
the largest open set contained in 𝐴. Indeed let 𝐵 be the union of all open sets contained
in 𝐴. Since a union of open sets is open, 𝐵 is also the largest open set contained in 𝐴.
We now check that 𝐴̊ = 𝐵. Let 𝑥 ∈ 𝐴,̊ then there is an open set 𝑈 such that 𝑥 ∈ 𝑈 ⊆ 𝐴,
but 𝑈 ⊆ 𝐵 thus 𝑥 ∈ 𝐵. Let 𝑥 ∈ 𝐵, then there is an open set 𝑈 such that 𝑥 ∈ 𝑈 ⊆ 𝐴, hence
𝑥 ∈ 𝐴.̊
Definition 2.11. One says that 𝑥 is an adherent point (or closure point) of 𝐴 if every
neighborhood of 𝑥 intersects 𝐴. The set of all adherent points of 𝐴 is called the closure
of 𝐴, and is denoted by 𝐴 or cl(𝐴).
It is easy to verify that 𝐴 is the intersection of all closed sets containing 𝐴, there-
fore also the smallest closed set containing 𝐴. The proof of this fact follows by taking
complement sets in the above argument for proving that 𝐴̊ is the union of all open sets
contained in 𝐴.
Definition 2.12. One says that 𝑥 is a limit point (or cluster point) of 𝐴 if every neigh-
borhood of 𝑥 intersects 𝐴 ⧵ {𝑥}.
Definition 2.14. One says that 𝑥 is an isolated point of 𝐴 if 𝑥 ∈ 𝐴 and there exists a
neighborhood of 𝑥 which does not contain any other points of 𝐴.
2.4 Separation and countability 9
One can verify that 𝑋 is T0 if and only if for any two distinct points there is one of
them that is not an adherent point of the singleton formed by the other. Moreover, 𝑋 is
T1 if and only if every singleton is closed. Furthermore a Hausdorff space is regular if
and only if for any 𝑥 ∈ 𝑋 and any neighborhood 𝑈 of 𝑥, there exists a neighborhood 𝑉
of 𝑥 such that 𝑥 ∈ 𝑉 ⊆ 𝑉 ⊆ 𝑈 ; and it is normal if and only if for any closed set 𝐹 in 𝑋 and
any neighborhood 𝑈 of 𝐹 , there exists a neighborhood 𝑉 of 𝐹 such that 𝐹 ⊆ 𝑉 ⊆ 𝑉 ⊆ 𝑈 .
One can also remark that in a first-countable space every point admits a neighbor-
hood basis composed by a decreasing sequence of open sets. Indeed let 𝑥 be a point in a
first-countable space, then there exists a neighborhood basis {𝑊𝑖 }𝑖∈N of 𝑥. For any 𝑛 ∈ N
let 𝑉𝑛 = ⋂0⩽𝑖⩽𝑛 𝑊̊ 𝑖 , therefore the family {𝑉𝑛 }𝑛∈N is a decreasing sequence of open sets
and it is a neighborhood basis for 𝑥.
Proof. Let 𝑋 be a second-countable space, then there is a countable basis ℬ = {𝐵𝑛 }𝑛∈N
for the topology of 𝑋 .
We first show that 𝑋 is separable. For any 𝑛 ∈ N we pick some 𝑥𝑛 ∈ 𝐵𝑛 , and then
define the countable subset 𝐴 = {𝑥𝑛 ∶ 𝑛 ∈ N} of 𝑋 . We now verify that 𝐴 = 𝑋 . Let
𝑈 be a non-empty open set in 𝑋 , then there is some non-empty set 𝐽 ⊆ N such that
𝑈 = ⋃𝑛∈𝐽 𝐵𝑛 . Therefore 𝑈 ∩ 𝐴 = {𝑥𝑛 ∶ 𝑛 ∈ 𝐽 } is non-empty, thus 𝐴 = 𝑋 .
We now show that 𝑋 is first-countable. For any 𝑥 ∈ 𝑋 , define 𝒮 (𝑥) = {𝐵𝑛 ∈ ℬ ∣ 𝑥 ∈
𝐵𝑛 } which is a countable family of neighborhoods of 𝑥. Let 𝑉 be a neighborhood of 𝑥,
then there is a open set 𝑈 in 𝑋 such that 𝑥 ∈ 𝑈 ⊆ 𝑉 . Since ℬ is a basis, 𝑈 is a union
of elements of ℬ, therefore there is 𝑛 ∈ N such that 𝑥 ∈ 𝐵𝑛 ⊆ 𝑈 ⊆ 𝑉 , hence 𝒮 (𝑥) is a
neighborhood basis for 𝑥.
Proof. Let (𝑋 , 𝑑) be a metric space. For any 𝑥 ∈ 𝑋 define 𝒮 (𝑥) = {𝐵(𝑥, 𝑛1 )}𝑛∈N∗ , which
is a countable family of neighborhoods of 𝑥. Let 𝑉 be a neighborhood of 𝑥, then there
exists a open set 𝑈 in 𝑋 such that 𝑥 ∈ 𝑈 ⊆ 𝑉 . Since 𝑈 is open, there is 𝑛 ∈ N∗ large
enough such that 𝐵(𝑥, 𝑛1 ) ⊆ 𝑈 ⊆ 𝑉 , thus 𝒮 (𝑥) is a neighborhood basis for 𝑥.
Proposition 2.21. For metric spaces second-countability and separability are equivalent.
Proof. Let (𝑋 , 𝑑) be a metric space. Thanks to Proposition 2.19 we only need to show that
separable implies second-countable. Assume 𝑋 is separable, then there is a countable
subset 𝐴 ⊆ 𝑋 such that 𝐴 = 𝑋 . Define ℬ = {𝐵(𝑦, 𝑛1 )}𝑦∈𝐴,𝑛∈N∗ , which is a countable
family of open subsets of 𝑋 .
2.4 Separation and countability 11
Let us show that ℬ is a basis for the topology of 𝑋 . Let 𝑈 be a open set in 𝑋 and
𝑥 ∈ 𝑈 . Since 𝑈 is open, there is 𝜀 > 0 such that 𝐵(𝑥, 𝜀) ⊆ 𝑈 , hence there exists 𝑛 ∈ N∗
large enough such that 𝐵(𝑥, 𝑛2 ) ⊆ 𝑈 . Since 𝐴 = 𝑋 and 𝐵(𝑥, 𝑛1 ) is a neighborhood of 𝑥,
there is a point 𝑦 ∈ 𝑋 such that 𝑦 ∈ 𝐴 ∩ 𝐵(𝑥, 𝑛1 ). Therefore 𝐵(𝑦, 𝑛1 ) ⊆ 𝐵(𝑥, 𝑛2 ) ⊆ 𝑈 .
Chapter 3
13
14 Chapter 3. Continuity and limits
Theorem 3.3. Let 𝑓 ∶ 𝑋 → 𝑌 be a map between two topological spaces. The following
are equivalent:
(i) 𝑓 is continuous on 𝑋 ;
(ii) for any open set 𝑈 in 𝑌 , 𝑓 −1 (𝑈 ) is an open set in 𝑋 ;
(iii) for any closed set 𝐹 in 𝑌 , 𝑓 −1 (𝐹 ) is a closed set in 𝑋 ;
(iv) for any 𝐴 ⊆ 𝑋 , one has 𝑓 (𝐴) ⊆ 𝑓 (𝐴).
For any 𝑖 ∈ 𝐽 one has 𝑥 ∈ 𝑓𝑖−1 (𝑈𝑖 ), thus 𝑓𝑖 (𝑥) ∈ 𝑈𝑖 . Since ℬ𝑖 is a basis for the topology 𝒪𝑖 ,
there is 𝐵𝑖 ∈ ℬ𝑖 such that 𝑓𝑖 (𝑥) ∈ 𝐵𝑖 ⊆ 𝑈𝑖 , hence 𝑥 ∈ 𝑓𝑖−1 (𝐵𝑖 ) ⊆ 𝑓 −1 (𝑈𝑖 ). Therefore,
𝑥 ∈ ⋂ 𝑓𝑖−1 (𝐵𝑖 ) ⊆ ⋂ 𝑓𝑖−1 (𝑈𝑖 ) ⊆ 𝑈 ,
𝑖∈𝐽 𝑖∈𝐽
16 Chapter 3. Continuity and limits
𝑥 ∈ ⋂ 𝑓𝑖−1 (𝑈𝑖 ) ⊆ 𝑉 .
𝑖∈𝐽
For any 𝑖 ∈ 𝐽 one has 𝑥 ∈ 𝑓𝑖−1 (𝑈𝑖 ), thus 𝑓𝑖 (𝑥) ∈ 𝑈𝑖 . Since 𝒮𝑖 is a neighborhood basis for
𝑓𝑖 (𝑥) in 𝑌𝑖 and 𝑈𝑖 is a neighborhood of 𝑓𝑖 (𝑥), there is some 𝑉𝑖 ∈ 𝒮𝑖 such that 𝑓𝑖 (𝑥) ∈ 𝑉𝑖 ⊆ 𝑈𝑖 ,
hence 𝑥 ∈ 𝑓𝑖−1 (𝑉𝑖 ) ⊆ 𝑓 −1 (𝑈𝑖 ). Therefore,
Therefore
𝑧 ∈ ⋂ 𝑔 −1 (𝑓𝑖−1 (𝑈𝑖 )) = ⋂(𝑓𝑖 ∘ 𝑔)−1 (𝑈𝑖 ) ⊆ 𝑔 −1 (𝑉 ),
𝑖∈𝐽 𝑖∈𝐽
and the set in the left-hand side is open in 𝑋 since it is a finite intersection of open sets,
hence 𝑔 −1 (𝑉 ) is a neighborhood of 𝑧 in 𝑍 , which completes the proof.
Two particular examples of initial topologies are the subspace topology and the prod-
uct topology that we describe below.
𝒪𝑌 ∶= {𝑈 ∩ 𝑌 ∣ 𝑈 ∈ 𝒪}.
where 𝐽 ⊆ 𝐼 is finite and 𝑈𝑗 ∈ ℬ𝑗 for any 𝑗 ∈ 𝐽 , is a basis for the product topology.
(iii) Let 𝑥 = (𝑥𝑖 )𝑖∈𝐼 ∈ 𝑋 and 𝒱𝑖 be a neighborhood basis for 𝑥𝑖 in 𝑋𝑖 for any 𝑖 ∈ 𝐼 , then
the family of sets of the form
The next result shows that any product of Hausdorff spaces is also Hausdorff.
Proposition 3.12. Let {𝑋𝑖 }𝑖∈𝐼 be a family of Hausdorff topological spaces. Then the product
space 𝑋 = ∏𝑖∈𝐼 𝑋𝑖 is Hausdorff.
Proof. Let 𝑥 = (𝑥𝑖 )𝑖∈𝐼 ≠ (𝑦𝑖 )𝑖∈𝐼 = 𝑦 belong to 𝑋 . There is 𝑘 ∈ 𝐼 such that 𝑥𝑘 ≠ 𝑦𝑘 thus,
since 𝑋𝑘 is Hausdorff, there are disjoint open sets 𝑈𝑘 and 𝑉𝑘 in 𝑋𝑘 such that 𝑥𝑘 ∈ 𝑈𝑘 and
𝑦𝑘 ∈ 𝑉𝑘 . Therefore the sets pr−1𝑘
(𝑈𝑘 ) and pr−1
𝑘
(𝑉𝑘 ) are disjoint open sets in 𝑋 containing
𝑥 and 𝑦, respectively.
The function 𝜑 is called an extraction map and it verifies 𝜑(𝑛) ⩾ 𝑛 for all 𝑛 ∈ N.
Let (𝑥𝑛 )𝑛⩾0 be a sequence in 𝑋 and 𝑥 ∈ 𝑋 a point.
Definition 3.16. One says that (𝑥𝑛 )𝑛⩾0 converges to 𝑥 when 𝑛 goes to +∞ if for any
neighborhood 𝑉 of 𝑥 in 𝑋 , there is 𝑁0 ∈ N such that, for any 𝑛 ⩾ 𝑁0 , one has 𝑥𝑛 ∈ 𝑉 .
When (𝑥𝑛 )𝑛⩾0 converges to 𝑥, we denote it by
lim 𝑥𝑛 = 𝑥 or 𝑥𝑛 −−−−−→ 𝑥.
𝑛→+∞ 𝑛→+∞
We observe that if (𝑋 , 𝑑) is a metric space, then (𝑥𝑛 )𝑛⩾0 converges to 𝑥 if and only if the
real sequence (𝑑(𝑥𝑛 , 𝑥))𝑛⩾0 converges to 0.
Definition 3.17. One says that 𝑥 is a cluster point of the sequence (𝑥𝑛 )𝑛⩾0 if for any
neighborhood 𝑉 of 𝑥 in 𝑋 and any 𝑁 ∈ N, there exists an integer 𝑛 ⩾ 𝑁 such that
𝑥𝑛 ∈ 𝑉 .
It easily follows from the definition that the set of all cluster points of a sequence
(𝑥𝑛 )𝑛⩾0 is given by
⋂ {𝑥𝑝 ∶ 𝑝 ⩾ 𝑛}.
𝑛∈N
Proof. Let (𝑥𝜑(𝑛) )𝑛⩾0 be a subsequence of (𝑥𝑛 )𝑛⩾0 that converges to 𝑥. Let 𝑉 be a neigh-
borhood of 𝑥 and 𝑁0 ∈ N. Since (𝑥𝜑(𝑛) )𝑛⩾0 converges to 𝑥, there is 𝑁1 ∈ N such that for
all integer 𝑛 ⩾ 𝑁1 one has 𝑥𝜑(𝑛) ∈ 𝑉 . Choosing 𝑛 = max(𝑁0 , 𝑁1 ) ∈ N, one has 𝜑(𝑛) ⩾ 𝑁0
and 𝑥𝜑(𝑛) ∈ 𝑉 . This proves that 𝑥 is a cluster point of (𝑥𝑛 )𝑛⩾0 .
Assume now that 𝑋 is first-countable and let 𝑥 be a cluster point of the sequence
(𝑥𝑛 )𝑛⩾0 . Let {𝑉𝑛 }𝑛∈N be a countable neighborhood basis for 𝑥 composed by a decreasing
sequence of opens sets. Since 𝑉0 is an open neighborhood of 𝑥, there is 𝑛0 ∈ N such
that 𝑥𝑛0 ∈ 𝑉0 . By induction one can therefore construct a sequence (𝑥𝑛𝑘 )𝑘⩾0 such that
𝑛𝑘 > 𝑛𝑘−1 and 𝑥𝑛𝑘 ∈ 𝑉𝑘 for all 𝑘 ∈ N∗ , thus (𝑥𝑛𝑘 )𝑘⩾0 is a subsequence of (𝑥𝑛 )𝑛⩾0 .
Let us now verify that (𝑥𝑛𝑘 )𝑘⩾0 converges to 𝑥. Let 𝑉 be a neighborhood of 𝑥, then
there is 𝑁0 ∈ N such that 𝑥 ∈ 𝑉𝑁0 ⊆ 𝑉 . Hence for any 𝑘 ⩾ 𝑁0 one has 𝑛𝑘 ⩾ 𝑘 and thus
𝑥𝑛𝑘 ∈ 𝑉𝑛𝑘 ⊆ 𝑉𝑁0 ⊆ 𝑉 .
As a consequence, in a first-countable space 𝑋 the closure of a subset can be charac-
terized using sequences: If 𝐴 is a subset of 𝑋 , then the closure 𝐴 is the set of all limits
of all convergent sequences of points in 𝐴.
Chapter 4
Completeness
One easily verifies that every convergent sequence is a Cauchy sequence, and that
every Cauchy sequence is bounded.
Proposition 4.2. Let (𝑋 , 𝑑) be a metric space, then the following properties hold:
(i) Every convergent sequence is a Cauchy sequence.
(ii) Every Cauchy sequence is bounded.
Proof. (i) Let (𝑥𝑛 )𝑛⩾0 be a sequence in 𝑋 that converges to 𝑥 ∈ 𝑋 . Let 𝜀 > 0, then there
is 𝑁0 ∈ N such that 𝑑(𝑥𝑛 , 𝑥) < 𝜀/2 for all 𝑛 ⩾ 𝑁0 . Thanks to the triangle inequality, for
all 𝑛, 𝑚 ⩾ 𝑁0 one obtains
(ii) Let (𝑥𝑛 )𝑛⩾0 be a Cauchy sequence in 𝑋 . We want to show that there is some 𝑎 ∈ 𝑋
and 𝑀 > 0 such that 𝑑(𝑥𝑛 , 𝑎) ⩽ 𝑀 for all 𝑛 ∈ N. There is 𝑁0 ∈ N such that 𝑑(𝑥𝑛 , 𝑥𝑁0 ) ⩽ 1
for all 𝑛 ⩾ 𝑁0 . Denote
Proposition 4.3. If a Cauchy sequence in a metric space (𝑋 , 𝑑) has a cluster point, then
it converges to it.
21
22 Chapter 4. Completeness
Proof. Let (𝑥𝑛 )𝑛⩾0 be a Cauchy sequence in 𝑋 that has a cluster point 𝑥 ∈ 𝑋 . From
Proposition 3.19, there is some subsequence (𝑥𝜑(𝑛) )𝑛⩾0 that converges to 𝑥. Let 𝜀 > 0. On
the one hand, since (𝑥𝜑(𝑛) )𝑛⩾0 converges to 𝑥, there is 𝑁0 ∈ N such that 𝑑(𝑥𝜑(𝑛) , 𝑥) < 𝜀/2
for any integer 𝑛 ⩾ 𝑁0 . On the other hand, since (𝑥𝑛 )𝑛⩾0 is a Cauchy sequence, there
is 𝑁1 ∈ N such that 𝑑(𝑥𝑛 , 𝑥𝑚 ) < 𝜀/2 for any integers 𝑛, 𝑚 ⩾ 𝑁1 . Therefore for any
𝑛 ⩾ max(𝑁0 , 𝑁1 ) it follows
𝑛 1/𝑝
𝑑𝑝 ((𝑥1 , … , 𝑥𝑛 ), (𝑦1 , … , 𝑦𝑛 )) = {∑ 𝑑𝑖 (𝑥𝑖 , 𝑦𝑖 )𝑝 } , 1 ⩽ 𝑝 < +∞,
𝑖=1
Proof. Let (𝑋 , 𝑑) be a complete metric space and 𝑌 ⊆ 𝑋 be closed. Let (𝑥𝑛 )𝑛⩾0 be a
Cauchy sequence in 𝑌 . Then (𝑥𝑛 )𝑛⩾0 is a Cauchy sequence in 𝑋 and hence there is
𝑦 ∈ 𝑋 such that (𝑥𝑛 )𝑛⩾0 converges to 𝑦 in 𝑋 . Since 𝑌 is closed, one concludes that 𝑦 ∈ 𝑌
by Proposition 3.18.
Proof. Since 𝑓 is a contraction map, there is 0 ⩽ 𝑘 < 1 such that 𝑑(𝑓 (𝑥), 𝑓 (𝑦)) ⩽ 𝑘𝑑(𝑥, 𝑦)
for any 𝑥, 𝑦 ∈ 𝑋 .
We first prove uniqueness. Assume 𝑥,̄ 𝑦 ̄ ∈ 𝑋 are fixed points of 𝑓 , then 𝑑(𝑥,̄ 𝑦)̄ =
𝑑(𝑓 (𝑥),
̄ 𝑓 (𝑦))
̄ ⩽ 𝑘𝑑(𝑥,̄ 𝑦),̄ which implies 𝑑(𝑥,̄ 𝑦)̄ = 0 and hence 𝑥 ̄ = 𝑦.̄
We shall now construct a convergent sequence in 𝑋 whose limit is a fixed point. Let
𝑥0 ∈ 𝑋 and define 𝑥𝑛 = 𝑓 (𝑥𝑛−1 ) for any 𝑛 ∈ N∗ . Let us verify that (𝑥𝑛 )𝑛⩾0 is a Cauchy
sequence. For any 𝑛 ∈ N∗ one has 𝑑(𝑥𝑛+1 , 𝑥𝑛 ) = 𝑑(𝑓 (𝑥𝑛 ), 𝑓 (𝑥𝑛−1 )) ⩽ 𝑘𝑑(𝑥𝑛 , 𝑥𝑛−1 ), hence
one gets by induction 𝑑(𝑥𝑛+1 , 𝑥𝑛 ) ⩽ 𝑘 𝑛 𝑑(𝑥1 , 𝑥0 ), therefore for any integers 𝑛, 𝑝 ∈ N one
has
𝑑(𝑥𝑛+𝑝 , 𝑥𝑛 ) ⩽ 𝑑(𝑥𝑛+𝑝 , 𝑥𝑛+𝑝−1 ) + 𝑑(𝑥𝑛+𝑝−1 , 𝑥𝑛+𝑝−2 ) + ⋯ + 𝑑(𝑥𝑛+1 , 𝑥𝑛 )
⩽ (𝑘 𝑛+𝑝−1 + 𝑘 𝑛+𝑝−2 + ⋯ + 𝑘 𝑛 ) 𝑑(𝑥1 , 𝑥0 )
𝑘𝑛
⩽ 𝑑(𝑥1 , 𝑥0 ).
1−𝑘
Fron this inequality we deduce that (𝑥𝑛 )𝑛⩾0 is a Cauchy sequence in 𝑋 . Indeed, let 𝜀 > 0,
𝑘 𝑁0
then we choose 𝑁0 ∈ N large enough such that 1−𝑘
𝑑(𝑥1 , 𝑥0 ) < 𝜀, then for all integers
𝑛, 𝑚 ⩾ 𝑁0 one has
𝑘 min(𝑛,𝑚) 𝑘 𝑁0
𝑑(𝑥𝑚 , 𝑥𝑛 ) ⩽ 𝑑(𝑥1 , 𝑥0 ) ⩽ 𝑑(𝑥1 , 𝑥0 ) < 𝜀.
1−𝑘 1−𝑘
Since 𝑋 is complete one deduces that there is 𝑥 ̄ ∈ 𝑋 such that (𝑥𝑛 )𝑛⩾0 converges to 𝑥.̄
Since 𝑓 is continuous at 𝑥 ̄ and 𝑋 is Hausdorff (so that limits are unique), one can pass
to the limit 𝑛 → +∞ in the relation 𝑥𝑛+1 = 𝑓 (𝑥𝑛 ), which yields 𝑥 ̄ = 𝑓 (𝑥).
̄
24 Chapter 4. Completeness
Theorem 4.12 (Baire’s category theorem, first version). Every complete metric space is
a Baire space.
Proof. Let 𝑋 be a complete metric space and {𝑈𝑛 }𝑛∈N be a countable family of dense open
sets in 𝑋 . Denoting 𝑈 ∶= ⋂𝑛∈N 𝑈𝑛 the intersection, we want to show that 𝑈 is dense in
𝑋 , that is, every non-empty open set 𝑊 in 𝑋 intersects 𝑈 .
Let 𝑊 be a non-empty open set in 𝑋 . Since 𝑈0 is dense, 𝑊 intersects 𝑈0 , hence there
is 𝑥0 ∈ 𝑈0 ∩ 𝑊 . Therefore there is 0 < 𝜀0 < 1 such that 𝐵(𝑥0 , 𝜀0 ) ⊆ 𝑈0 ∩ 𝑊 . By induction,
1
for any 𝑛 ∈ N there is 𝑥𝑛 and 0 < 𝜀𝑛 < 𝑛+1 such that 𝐵(𝑥𝑛 , 𝜀𝑛 ) ⊆ 𝐵(𝑥𝑛−1 , 𝜀𝑛−1 ) ∩ 𝑈𝑛 . The
sequence (𝑥𝑛 )𝑛⩾0 is a Cauchy sequence in 𝑋 which is complete, hence it converges to
some 𝑦 ∈ 𝑋 . For any 𝑛 ∈ N, (𝑥𝑘 )𝑘⩾𝑛 is a sequence in 𝐵(𝑥𝑛 , 𝜀𝑛 ) which is closed, and it also
is a subsequence of (𝑥𝑛 )𝑛⩾0 , therefore 𝑦 ∈ 𝐵(𝑥𝑛 , 𝜀𝑛 ) ⊆ 𝑊 ∩ 𝑈𝑛 .
Chapter 5
Compactness
25
26 Chapter 5. Compactness
Proof. (i) We shall prove that 𝑋 ⧵ 𝐴 is open, so let 𝑥 ∈ 𝑋 ⧵ 𝐴. Since 𝑋 is Hausdorff, for
any 𝑦 ∈ 𝐴 there exist an open neighborhood 𝑈𝑦 of 𝑥 in 𝑋 and an open neighborhood 𝑉𝑦
of 𝑦 in 𝑋 such that 𝑈𝑦 ∩ 𝑉𝑦 = ∅. Hence 𝐴 ⊆ ⋃𝑦∈𝐴 𝑉𝑦 and there exist {𝑦1 , … , 𝑦𝑁 } such
that 𝐴 ⊆ ⋃1⩽𝑖⩽𝑁 𝑉𝑦𝑖 . Then the set 𝑈 = ⋂1⩽𝑖⩽𝑁 𝑈𝑦𝑖 is open in 𝑋 and 𝑈 ∩ 𝐴 = ∅, hence
𝑥 ∈ 𝑈 ⊆ 𝑋 ⧵ 𝐴.
(ii) Let 𝐴1 , … , 𝐴𝑛 be compact in 𝑋 . The union 𝐴 ∶= ⋃1⩽𝑘⩽𝑛 𝐴𝑘 is Hausdorff, since it
is a subset of a Hausdorff space. Let {𝑈𝑖 }𝑖∈𝐼 be a cover of 𝐴 by open sets in 𝑋 . For
any 𝑘 ∈ {1, … , 𝑛} we have 𝐴𝑘 ⊆ ⋃𝑖∈𝐼 𝑈𝑖 , hence there exists a finite 𝐽𝑘 ⊆ 𝐼 such that
𝐴𝑘 ⊆ ⋃𝑖∈𝐽𝑘 𝑈𝑖 . Then 𝐽 = ⋃1⩽𝑘⩽𝑛 𝐽𝑘 is a finite set contained in 𝐼 and 𝐴 ⊆ ⋃𝑖∈𝐽 𝑈𝑖 , hence
𝐴 is compact.
Proof. Let 𝐴 be a closed set of a compact topological space 𝑋 . Recall that we already
know that 𝐴 is Hausdorff. Let {𝑈𝑖 }𝑖∈𝐼 be a cover of 𝐴 by open sets in 𝑋 . Let 𝑘 ∉ 𝐼 and
define 𝑈𝑘 = 𝑋 ⧵ 𝐴 which is open since 𝐴 is closed. Therefore {𝑈𝑖 }𝑖∈𝐼 ∪{𝑘} is a open cover
of 𝑋 , and since 𝑋 is compact, there exists a finite set 𝐽 ⊆ 𝐼 ∪ {𝑘} such that ⋃𝑖∈𝐽 𝑈𝑖 is a
finite cover of 𝑋 . This implies that ⋃𝑖∈𝐽 ⧵{𝑘} 𝑈𝑖 is a finite cover of 𝐴.
Proof. Let {𝐾𝑛 }𝑛∈N be a decreasing sequence of non-empty compact sets in 𝑋 . By way
of contradiction, suppose that ⋂𝑛∈N 𝐾𝑛 = ∅. Consider, for any 𝑛 ∈ N, the open set 𝑈𝑛 =
𝑋 ⧵ 𝐾𝑛 in 𝑋 . The collection {𝑈𝑛 }𝑛∈N is a open cover of 𝐾0 , which is compact, therefore
there is a finite set 𝐽 ⊆ N such that {𝑈𝑛 }𝑛∈𝐽 is a open cover of 𝐾0 . Denoting 𝑝 = max 𝐽 ,
it follows that 𝐾0 ⊆ 𝑈𝑝 which in turn implies 𝐾𝑝 ⊆ 𝑈𝑝 , hence a contradiction.
Proof. One has that 𝑓 (𝑋 ) is Hausdorff since it is a subspace of 𝑌 which is Hausdorff. Let
{𝑈𝑖 }𝑖∈𝐼 be an open cover of 𝑓 (𝑋 ) by open sets in 𝑌 . Therefore, since 𝑋 = 𝑓 −1 (𝑓 (𝑋 )) and
𝑓 is continuous, one obtains that {𝑓 −1 (𝑈𝑖 )}𝑖∈𝐼 is an open cover of 𝑋 . Since 𝑋 is compact,
there is a finite 𝐽 ⊆ 𝐼 such that {𝑓 −1 (𝑈𝑖 )}𝑖∈𝐽 is a finite cover of 𝑋 , which implies that
{𝑈𝑖 }𝑖∈𝐽 is a finite sub-cover of 𝑓 (𝑋 ).
Proposition 5.8. In a compact topological space, every sequence has at least one cluster
point.
Proof. Let (𝑥𝑛 )𝑛⩾0 be a sequence in a compact topological space 𝑋 . Recall that the set of
cluster points of (𝑥𝑛 )𝑛⩾0 is given by 𝐹 = ⋂𝑛∈N {𝑥𝑝 ∶ 𝑝 ⩾ 𝑛}. Denote 𝐹𝑛 ∶= {𝑥𝑝 ∶ 𝑝 ⩾ 𝑛}
and consider the family of closed sets {𝐹𝑛 }𝑛∈N . The intersection of any finite family of
{𝐹𝑛 }𝑛∈N is non-empty, indeed if 𝐽 ⊆ N is finite then ⋂𝑛∈𝐽 𝐹𝑛 contains 𝑥max 𝐽 . Therefore,
since 𝑋 is compact, we deduce that 𝐹 = ⋂𝑛∈N 𝐹𝑛 is non-empty.
5.3 Compact metric spaces 27
Proof. (i) ⇒ (ii) It follows from Proposition 5.8 and Proposition 3.19.
(ii) ⇒ (iii) Assume that 𝑋 is sequentially compact. We want to show that 𝑋 is com-
plete and totally bounded. Let (𝑥𝑛 )𝑛⩾0 be a Cauchy sequence in 𝑋 , then it admits a
subsequence that converges to some 𝑥 ∈ 𝑋 , and one deduces that 𝑥 is a cluster point of
(𝑥𝑛 )𝑛⩾0 from Proposition 3.19. Therefore, by Proposition 4.3, one concludes that (𝑥𝑛 )𝑛⩾0
converges to 𝑥.
Let 𝜀 > 0. We argue by contradiction and assume that 𝑋 does not admit any finite
cover by a finite family of open balls of radius 𝜀. We fix 𝑥0 ∈ 𝑋 , then by hypothesis
we can chose some 𝑥1 ∈ 𝑋 ⧵ 𝐵(𝑥0 , 𝜀). By induction, we can then construct a sequence
𝑛−1
(𝑥𝑛 )𝑛⩾0 of elements of 𝑋 such that 𝑥𝑛 ∉ ⋃𝑖=1 𝐵(𝑥𝑖 , 𝜀) for all 𝑛 ⩾ 1. This sequence satisfies
𝑑(𝑥𝑖 , 𝑥𝑗 ) ⩾ 𝜀 for any integers 𝑖 ≠ 𝑗 and hence does not admit any convergent subsequence,
which is a contradiction.
(ii) ⇒ (i) Assume that 𝑋 is sequentially compact. We want to show that 𝑋 is compact.
We already know that 𝑋 is Hausdorff, so we need to prove that 𝑋 is quasi-compact.
Let {𝑈𝑖 }𝑖∈𝐼 be an open cover of 𝑋 . We now show that there is 𝛿 > 0 such that, for any
𝑥 ∈ 𝑋 , there is 𝑖 ∈ 𝐼 such that 𝐵(𝑥, 𝛿) ⊆ 𝑈𝑖 . We argue by contradiction and assume that it
is not the case, hence for any 𝑛 ∈ N∗ there is 𝑥𝑛 ∈ 𝑋 such that 𝐵(𝑥𝑛 , 𝑛1 ) ⊈ 𝑈𝑖 for any 𝑖 ∈ 𝐼 .
The sequence (𝑥𝑛 )𝑛⩾1 has a subsequence that converges to some 𝑥 ∈ 𝑋 , which implies
that 𝑥 is a cluster point of (𝑥𝑛 )𝑛⩾1 . Let 𝑖0 ∈ 𝐼 such that 𝑥 ∈ 𝑈𝑖0 , thus for 𝑛 large enough
one obtains 𝐵(𝑥𝑛 , 𝑛1 ) ⊆ 𝑈𝑖0 , which is a contradiction.
By repeating the above argument used to show that (ii) implies (iii), we show that 𝑋
can be recovered by a finite number of ball of radius 𝛿.
(iii) ⇒ (ii) Assume now that 𝑋 is complete and totally bounded. We wan to prove that 𝑋
is sequentially compact. Let x = (𝑥𝑛 )𝑛⩾0 be a sequence in 𝑋 . For 𝑗 = 1 there exists a finite
family of open ball of radius 1 covering 𝑋 , hence there is a subsequence x1 = (𝑥𝜑1 (𝑛) )𝑛⩾0
of x that is included in some open ball of radius 1. By induction, for any integer 𝑗 ⩾ 2,
one can then find a subsequence x𝑗 = (𝑥𝜑1 ∘⋯∘𝜑𝑗 (𝑛) )𝑛⩾0 of x𝑗−1 (hence also a subsequence
of x) that is included in some open ball of radius 1𝑗 . By a diagonal extraction argument,
one defines the sequence (𝑦𝑛 )𝑛⩾1 with 𝑦𝑛 = 𝑥𝜑1 ∘⋯∘𝜑𝑛 (𝑛) , which is a subsequence of (𝑥𝑛 )𝑛⩾0
and satisfies 𝑑(𝑦𝑛 , 𝑦𝑚 ) ⩽ 𝑛1 for any positive integers 𝑛 ⩽ 𝑚. Therefore (𝑦𝑛 )𝑛⩾1 is a Cauchy
sequence, thus it converges since 𝑋 is complete.
Proof. Assume that every open cover of 𝑋 by elements of 𝒮 has a finite sub-cover. By
way of contradiction, suppose that there is a open cover of 𝑋 with no finite sub-cover.
Consider the collection 𝔓 of open covers of 𝑋 with no finite sub-cover, which is
non-empty, partially ordered by inclusion Let 𝔘 be a chain (totally ordered subset) in
𝔓. We now define 𝒰 = ⋃ 𝔘 which is an upper bound of 𝔘, and we shall verify that
𝒰 has no finite sub-cover. Consider any finite sub-collection 𝑈1 , … , 𝑈𝑛 of 𝒰, then for
any 1 ⩽ 𝑖 ⩽ 𝑛 there is 𝒰𝑖 ∈ 𝔘 such that 𝑈𝑖 ∈ 𝒰𝑖 . Since 𝔘 is totally ordered, there is
some 𝑘 ∈ {1, … , 𝑛} such that 𝒰𝑘 contains all the sets of { 𝑈1 , … , 𝑈𝑛 }, therefore this finite
sub-collection cannot cover 𝑋 . By Zorn’s Lemma (Lemma A.4), one deduces that 𝔓 has
a maximal element ℳ. Since ℳ is a maximal element, for any open 𝑈 ∉ ℳ the cover
ℳ ∪ {𝑈 } has a finite sub-cover, necessarily of the form ℳ ′ ∪ {𝑈 } for some finite subset
ℳ ′ ⊆ ℳ.
We claim that if 𝑈 , 𝑉 are open sets in 𝑋 such that 𝑈 , 𝑉 ∉ ℳ, then 𝑈 ∩ 𝑉 ∉ ℳ. Indeed,
there is finite subsets ℳ ′ ⊆ ℳ and ℳ ″ ⊆ ℳ such that ℳ ′ ∪ {𝑈 } and ℳ ″ ∪ {𝑉 } both
cover 𝑋 . Therefore {𝑈 ∩ 𝑉 } ∪ ℳ ′ ∪ ℳ ″ is a finite cover of 𝑋 , hence 𝑈 ∩ 𝑉 ∉ ℳ since ℳ
has no finite sub-cover.
We also claim that if 𝑈 , 𝑉 are open sets in 𝑋 such that 𝑈 ∉ ℳ and 𝑈 ⊆ 𝑉 , then
𝑉 ∉ ℳ. Indeed, if ℳ ′ ∪ {𝑈 } is a finite cover, with ℳ ′ ⊆ ℳ finite, then also is ℳ ′ ∪ {𝑉 },
thus 𝑉 ∉ ℳ.
We now show that ℳ ∩ 𝒮 covers 𝑋 . Let 𝑥 ∈ 𝑋 , since ℳ is a cover, there is some
open set 𝑈 ∈ ℳ such that 𝑥 ∈ 𝑈 . Since 𝒮 is a sub-basis, there are 𝑆1 , … , 𝑆𝑛 ∈ 𝒮 such that
𝑥 ∈ 𝑆1 ∩ ⋯ ∩ 𝑆𝑛 ⊆ 𝑈 . Thanks to the two claims above, there is some 𝑖 such that 𝑆𝑖 ∈ ℳ,
whence 𝑥 ∈ 𝑆𝑖 ⊆ ℳ ∩ 𝒮 .
Since ℳ ∩ 𝒮 is a cover of 𝑋 , it follows that ℳ ∩ 𝒮 has a finite sub-cover, which
contradicts the fact that ℳ has no finite sub-cover. Therefore the original collection 𝔓
is empty and thus every open cover of 𝑋 has a finite sub-cover.
Theorem 5.11 (Tychonoff’s theorem). Any product of compact topological spaces is com-
pact.
Proof. Let {𝑋𝑖 }𝑖∈𝐼 be a family of compact topological spaces and consider the product
space 𝑋 = ∏𝑖∈𝐼 𝑋𝑖 , equipped with the product topology, and denote by pr𝑗 ∶ 𝑋 → 𝑋𝑗 ,
(𝑥𝑖 )𝑖∈𝐼 ↦ 𝑥𝑗 , the canonical projections. Since every 𝑋𝑖 is Hausdorff, we already know
that 𝑋 is also Hausdorff (Proposition 3.12). Recall that the collection 𝒮 of subsets of the
form pr−1𝑖 (𝑈𝑖 ), where 𝑖 ∈ 𝐼 and 𝑈𝑖 is open in 𝑋𝑖 , is a sub-basis for the product topology
on 𝑋 .
We argue by contradiction and assume that 𝑋 is not compact. Thanks to Lemma 5.10,
there is a cover 𝒰 of 𝑋 by elements of 𝒮 with no finite sub-cover. For each 𝑖 ∈ 𝐼 , define
the collection 𝒮𝑖 of open sets 𝑈 ⊆ 𝑋𝑖 such that pr−1 𝑖 (𝑈 ) ∈ 𝒮 .
We now show that 𝒮𝑖 does not cover 𝑋𝑖 . Indeed, since 𝑋𝑖 is compact, if 𝒮𝑖 covers 𝑋𝑖
then there is a finite collection 𝑈1 , … , 𝑈𝑛 ∈ 𝒮𝑖 that covers 𝑋𝑖 . If follows that
pr−1 −1 −1
𝑖 (𝑈1 ) ∪ ⋯ ∪ pr𝑖 (𝑈𝑛 ) = pr𝑖 (𝑋𝑖 ) = 𝑋 ,
Proposition 5.13. In a locally compact space 𝑋 , every point has a neighbourhood basis
composed by compact sets.
Theorem 5.14 (Baire’s category theorem, second version). Every locally compact topo-
logical space is a Baire space.
Proof. Let 𝑋 be a locally compact topological space and {𝑈𝑛 }𝑛∈N be a countable family
of dense open sets in 𝑋 . Denote 𝑈 ∶= ⋂𝑛∈N 𝑈𝑛 the intersection.
Let 𝑊 be a non-empty open set in 𝑋 . Since 𝑈0 is dense, 𝑊 intersects 𝑈0 , which implies
that there is 𝑥0 ∈ 𝑈0 ∩ 𝑊 . Since 𝑋 is locally compact, there exists a compact set 𝐾0 that
is a neighborhood of 𝑥0 , hence with non-empty interior, such that 𝑥0 ∈ 𝐾0 ⊆ 𝑊 ∩ 𝑈0 .
Since 𝑈1 is dense, one has 𝑈1 ∩ 𝐾̊0 ≠ ∅, which implies that there is 𝑥1 ∈ 𝑈1 ∩ 𝐾̊0 . As
before, there exists a compact set 𝐾1 with non-empty interior such that 𝐾1 ⊆ 𝑈1 ∩ 𝐾0̊ .
By induction, for any 𝑛 ∈ N, there is a compact set 𝐾𝑛 with non-empty interior such
that 𝐾𝑛 ⊆ 𝑈𝑛 ∩ 𝐾̊𝑛−1 . Therefore ⋂𝑛∈N 𝐾𝑛 ⊆ 𝑊 ∩ ⋂𝑛∈N 𝑈𝑛 , and the intersection ⋂𝑛∈N 𝐾𝑛 is
non-empty by Proposition 5.6.
Chapter 6
Connectedness
Definition 6.2. A topological space 𝑋 is locally connected if every point has a neighbor-
hood basis formed by connected sets.
31
32 Chapter 6. Connectedness
Proof. Let 𝑎 ∈ 𝑋 and denote by 𝐴 the set of all points that can be joined to 𝑎 by a path.
We will show that 𝐴 is both open and closed, which concludes the proof since 𝑋 is
connected.
For any 𝑥 ∈ 𝑋 there is a open neighborhood 𝐵𝑥 of 𝑥 such that every point 𝑦 ∈ 𝐵𝑥
can be joined to 𝑥 by a path. Therefore, for any 𝑥 ∈ 𝐴 and 𝑦 ∈ 𝐵𝑥 there is a path joining
𝑎 to 𝑦, this implies 𝑦 ∈ 𝐴 and thus 𝐴 is open.
Let 𝑧 ∈ 𝐴. Let 𝑦 ∈ 𝐴 ∩ 𝐵𝑧 , then there is a path joining 𝑦 to 𝑧 and a path joining 𝑎 to 𝑦,
hence the existence of a path joining 𝑎 to 𝑧. Therefore 𝑧 ∈ 𝐴 and hence 𝐴 is closed.
Chapter 7
Function spaces
35
36 Chapter 7. Function spaces
for any 𝑖 = 1, … , 𝑘, one has 𝑓 (𝑥𝑖 ) ∈ 𝑈𝑖 and hence there is 𝑛𝑖 ∈ 𝑁 such that 𝑓𝑛 (𝑥𝑖 ) ∈ 𝑈𝑖
for any integer 𝑛 ⩾ 𝑛𝑖 . Hence, taking 𝑁 = max1⩽𝑖⩽𝑘 𝑛𝑖 , for any integer 𝑛 ⩾ 𝑁 one has
𝑓𝑛 ∈ pr−1 −1
𝑥1 (𝑈1 ) ∩ ⋯ ∩ pr𝑥𝑘 (𝑈𝑘 ) ⊆ 𝑉 .
One can easily check that 𝑑∞ is a metric on ℱ (𝑋 , 𝑌 ), and that a sequence of maps (𝑓𝑛 )𝑛⩾0
converges uniformly to 𝑓 on 𝑋 if and only if (𝑓𝑛 )𝑛⩾0 converges to 𝑓 with respect the
metric 𝑑∞ . The metric 𝑑∞ is called the metric of uniform convergence or the uniform
metric, and its associated topology is called the topology of uniform convergence.
Proposition 7.1. Let 𝑋 be a set and (𝑌 , 𝑑) a complete metric space. Then (ℱ (𝑋 , 𝑌 ), 𝑑∞ )
is complete.
Proof. Let (𝑓𝑛 )𝑛⩾0 be a Cauchy sequence in (ℱ (𝑋 , 𝑌 ), 𝑑∞ ). Then for any 𝑥 ∈ 𝑋 , (𝑓𝑛 (𝑥))𝑛⩾0
is a Cauchy sequence in (𝑌 , 𝑑), since 𝑑(𝑓𝑛 (𝑥), 𝑓 (𝑥)) ⩽ sup𝑦∈𝑋 (𝑓𝑛 (𝑦), 𝑓 (𝑦)) if 𝑑∞ (𝑓𝑛 , 𝑓 ) < 1.
Therefore, for any 𝑥 ∈ 𝑋 , (𝑓𝑛 (𝑥))𝑛⩾0 converges to some point 𝑓 (𝑥) in 𝑌 . We have hence
constructed a map 𝑓 ∶ 𝑋 → 𝑌 , 𝑥 ↦ 𝑓 (𝑥), and it remains to show that (𝑓𝑛 )𝑛⩾0 converges
to 𝑓 with respect to the metric 𝑑∞ .
Let 𝜀 > 0, then there is 𝑛0 ∈ N such that one has, for any for any integers 𝑛, 𝑚 ⩾ 𝑛0
and any 𝑥 ∈ 𝑋 , 𝑑(𝑓𝑛 (𝑥), 𝑓𝑚 (𝑥)) < 𝜀. Hence, with 𝑥 ∈ 𝑋 fixed and letting 𝑚 → +∞, we
obtain 𝑑(𝑓𝑛 (𝑥), 𝑓 (𝑥)) ⩽ 𝜀 for all 𝑛 ⩾ 𝑛0 . Thus (𝑓𝑛 )𝑛⩾0 converges uniformly to 𝑓 on 𝑋 .
and any 𝑥 ∈ 𝑋 , 𝑑(𝑓𝑛 (𝑥), 𝑓𝑚 (𝑥)) < 𝜀. Hence, with 𝑥 ∈ 𝑋 fixed and letting 𝑚 → +∞, we
obtain 𝑑(𝑓𝑛 (𝑥), 𝑓 (𝑥)) ⩽ 𝜀 for all 𝑛 ⩾ 𝑛0 . Thus (𝑓𝑛 )𝑛⩾0 converges uniformly to 𝑓 on 𝑋 .
We now prove that 𝑓 is continuous on 𝑋 . Let 𝑎 ∈ 𝑋 and 𝜀 > 0. There is 𝑛0 ∈ N such
that sup𝑥∈𝑋 𝑑(𝑓𝑛0 (𝑥), 𝑓 (𝑥)) < 𝜀/3. Since 𝑓𝑛0 is continuous at 𝑎, there is a neighbourhood
𝑉 of 𝑎 in 𝑋 such that for any 𝑥 ∈ 𝑉 one has 𝑑(𝑓𝑛0 (𝑥), 𝑓𝑛0 (𝑎)) < 𝜀/3. Therefore for any
𝑥 ∈ 𝑉 one has
𝑑(𝑓 (𝑥), 𝑓 (𝑎)) ⩽ 𝑑(𝑓 (𝑥), 𝑓𝑛0 (𝑥)) + 𝑑(𝑓𝑛0 (𝑥), 𝑓𝑛0 (𝑎)) + 𝑑(𝑓𝑛0 (𝑎), 𝑓 (𝑎)) < 𝜀,
thus 𝑓 is continuous at 𝑎. We conclude that 𝑓 is continuous on 𝑋 since 𝑎 is arbitrary.
7.3 Equicontinuity
Let 𝑋 be a topological space and (𝑌 , 𝑑) a metric space.
Definition 7.3. Let ℋ be a subset of ℱ (𝑋 , 𝑌 ) and 𝑥 ∈ 𝑋 a point. One says that ℋ is
equicontinuous at 𝑥 if for any 𝜀 > 0, there is a neighbourhood 𝑈 of 𝑥 in 𝑋 such that,
∀𝑦 ∈ 𝑈 , ∀𝑓 ∈ ℋ , 𝑑(𝑓 (𝑦), 𝑓 (𝑥)) < 𝜀.
One says that ℋ is equicontinuous on 𝑋 if ℋ is equicontinuous at every point of 𝑋 .
Example 7.1. A finite family ℋ = {𝑓1 , … , 𝑓𝑝 } of functions that are continuous at a point
𝑥 ∈ 𝑋 is equicontinuous at 𝑥.
Example 7.2. If 𝑋 is a metric space and 𝑘 > 0, the set of all 𝑘-Lipschitz functions from 𝑋
to 𝑌 is equicontinuous on 𝑋 .
Proposition 7.4. Let (𝑓𝑛 )𝑛⩾0 be a sequence of maps from 𝑋 to 𝑌 . Suppose that {𝑓𝑛 ∶ 𝑛 ∈ N}
is equicontinuous at 𝑎 ∈ 𝑋 and that (𝑓𝑛 )𝑛⩾0 converges pointwisely on 𝑋 to some map
𝑓 ∶ 𝑋 → 𝑌 . Then 𝑓 is continuous at 𝑎.
Proof. Let 𝜀 > 0. Since {𝑓𝑛 ∶ 𝑛 ∈ N} is equicontinuous at 𝑎, there is a neighbourhood 𝑈
of 𝑎 such that for all 𝑥 ∈ 𝑈 and any 𝑛 ∈ N one has 𝑑(𝑓𝑛 (𝑥), 𝑓𝑛 (𝑎)) < 𝜀/3. Since (𝑓𝑛 )𝑛⩾0
converges pointwisely to 𝑓 , for any 𝑥 ∈ 𝑈 there is 𝑁0 ∈ N such that for any 𝑛 ⩾ 𝑁0 one
has 𝑑(𝑓𝑛 (𝑥), 𝑓 (𝑥)) < 𝜀/3. Hence, for any 𝑥 ∈ 𝑈 and any 𝑛 ⩾ 𝑁0 , one obtains
𝑑(𝑓 (𝑥), 𝑓 (𝑎)) ⩽ 𝑑(𝑓 (𝑥), 𝑓𝑛 (𝑥)) + 𝑑(𝑓𝑛 (𝑥), 𝑓𝑛 (𝑎)) + 𝑑(𝑓𝑛 (𝑎), 𝑓 (𝑎)) < 𝜀,
thus 𝑓 is continuous at 𝑎.
Proposition 7.5. Let (𝑓𝑛 )𝑛⩾0 be a sequence of maps from 𝑋 to 𝑌 and suppose that 𝐴 is
dense in 𝑋 and that 𝑌 is complete. Assume that {𝑓𝑛 ∶ 𝑛 ∈ N} is equicontinuous on 𝑋 and
that (𝑓𝑛 )𝑛⩾0 converges pointwisely on 𝐴. Then (𝑓𝑛 )𝑛⩾0 converges pointwisely on 𝑋 .
Proof. We show that, for any 𝑥 ∈ 𝑋 , the sequence (𝑓𝑛 (𝑥))𝑛⩾0 is a Cauchy sequence
in 𝑌 . Let 𝑥 ∈ 𝑋 and 𝜀 > 0. Since {𝑓𝑛 ∶ 𝑛 ∈ N} is equicontinuous at 𝑥, there is a open
neighbourhood 𝑈 of 𝑥 such that for all 𝑦 ∈ 𝑈 and any 𝑛 ∈ N one has 𝑑(𝑓𝑛 (𝑥), 𝑓𝑛 (𝑦)) < 𝜀/3.
By density of 𝐴, there exists a point 𝑧 ∈ 𝐴 ∩ 𝑈 and in particular one has 𝑑(𝑓𝑛 (𝑥), 𝑓𝑛 (𝑧)) <
𝜀/3. Thanks to the pointwise convergence, the sequence (𝑓𝑛 (𝑧))𝑛⩾0 is a Cauchy sequence
in 𝑌 , therefore there is 𝑁0 ∈ N such that for any 𝑛, 𝑚 ⩾ 𝑁0 one has 𝑑(𝑓𝑛 (𝑧), 𝑓𝑚 (𝑧)) < 𝜀/3.
One finally gets, for any 𝑛, 𝑚 ⩾ 𝑁0 , that
𝑑(𝑓𝑛 (𝑥), 𝑓𝑚 (𝑥)) ⩽ 𝑑(𝑓𝑛 (𝑥), 𝑓𝑛 (𝑧)) + 𝑑(𝑓𝑛 (𝑧), 𝑓𝑚 (𝑧)) + 𝑑(𝑓𝑚 (𝑧), 𝑓𝑚 (𝑥)) < 𝜀,
hence (𝑓𝑛 (𝑥))𝑛⩾0 is a Cauchy sequence in 𝑌 .
38 Chapter 7. Function spaces
Proposition 7.6. Let (𝑓𝑛 )𝑛⩾0 be a sequence of maps from 𝑋 to 𝑌 and suppose that 𝑋 is
compact. Assume that {𝑓𝑛 ∶ 𝑛 ∈ N} is equicontinuous on 𝑋 and that (𝑓𝑛 )𝑛⩾0 converges
pointwisely on 𝑋 to some map 𝑓 ∶ 𝑋 → 𝑌 . Then (𝑓𝑛 )𝑛⩾0 converges uniformly on 𝑋 to 𝑓 .
𝑑(𝑓𝑛 (𝑥), 𝑓 (𝑥)) ⩽ 𝑑(𝑓𝑛 (𝑥), 𝑓𝑛 (𝑥𝑖 )) + 𝑑(𝑓𝑛 (𝑥𝑖 ), 𝑓 (𝑥𝑖 )) + 𝑑(𝑓 (𝑥𝑖 ), 𝑓 (𝑥)) < 𝜀,
Remark 7.8. If 𝑌 = R or C then property (i) just means that ℋ (𝑥) is bounded in R or
C for any 𝑥 ∈ 𝑋 .
Proof of Theorem 7.7. We start with the easier part of the result and assume that ℋ has
compact closure in 𝒞 (𝑋 , 𝑌 ). For any 𝑥 ∈ 𝑋 , the evaluation map 𝑓 ↦ 𝑓 (𝑥) is continuous
from 𝒞 (𝑋 , 𝑌 ) into 𝑌 since 𝑑(𝑓 (𝑥), 𝑔(𝑥)) ⩽ 𝑑∞ (𝑓 , 𝑔), thus the set ℋ (𝑥) is compact as the
image of a compact under a continuous function. This gives (i).
Let 𝑥 ∈ 𝑋 and 𝜀 > 0. Since ℋ is compact it is totally bounded, hus there is a finite
collection {𝑓1 , … , 𝑓𝑝 } ⊆ ℋ such that, for any 𝑓 ∈ ℋ there is a label 𝑖 ∈ {1, … , 𝑝} such that
𝑑∞ (𝑓 , 𝑓𝑖 ) < 𝜀/3. Since the finite family {𝑓1 , … , 𝑓𝑝 } is equicontinuous at 𝑥, there is 𝑈 a
neighbourhood of 𝑥 such that for all 𝑦 ∈ 𝑈 and any 𝑖 ∈ {1, … , 𝑝} one has 𝑑(𝑓𝑖 (𝑦), 𝑓𝑖 (𝑥)) <
𝜀/3. Therefore, for any 𝑦 ∈ 𝑈 and any 𝑓 ∈ ℋ one obtains, with the choice of label
𝑖 ∈ {1, … , 𝑝} described above,
𝑑(𝑓 (𝑦), 𝑓 (𝑥)) ⩽ 𝑑(𝑓 (𝑦), 𝑓𝑖 (𝑦)) + 𝑑(𝑓𝑖 (𝑦), 𝑓𝑖 (𝑥)) + 𝑑(𝑓𝑖 (𝑥), 𝑓 (𝑥)) < 𝜀.
7.5 Stone-Weierstrass theorem 39
𝑑(𝑓 (𝑥), 𝑔(𝑥)) ⩽ 𝑑(𝑓 (𝑥), 𝑓 (𝑥𝑖 )) + 𝑑(𝑓 (𝑥𝑖 ), 𝑦𝜙(𝑖) ) + 𝑑(𝑦𝜙(𝑖) , 𝑔(𝑥𝑖 )) + 𝑑(𝑔(𝑥𝑖 ), 𝑔(𝑥))
𝜀 𝜀 𝜀 𝜀
< + + + = 𝜀,
4 4 4 4
therefore ℋ𝜙 is contained in a ball of radius 𝜀 for the uniform metric 𝑑∞ , which concludes
the proof.
Lemma 7.10. There exists a sequence of real polynomials (𝑃𝑛 )𝑛⩾0 in 𝑡 that converges uni-
formly to the function 𝑡 ↦ |𝑡| on [−1, 1].
Proof. Define 𝑃0 (𝑡) = 0 for any 𝑡 ∈ [−1, 1] and, for any 𝑛 ∈ N, 𝑃𝑛+1 (𝑡) = 𝑃𝑛 (𝑡) + 12 (𝑡 2 −
𝑃𝑛 (𝑡)2 ) for any 𝑡 ∈ [−1, 1]. By induction we easily obtain that 0 ⩽ 𝑃𝑛 (𝑡) ⩽ 𝑃𝑛+1 (𝑡) ⩽ |𝑡| for
any 𝑡 ∈ [−1, 1]. Therefore, for any 𝑡 ∈ [−1, 1], the real sequence (𝑃𝑛 (𝑡))𝑛⩾0 is increasing
and bounded above, hence it converges to some limit that we denote 𝑃(𝑡). By passing
to the limit 𝑛 → ∞, we obtain that 0 ⩽ 𝑃(𝑡) ⩽ |𝑡| and 𝑃(𝑡) = 𝑃(𝑡) − 12 (𝑡 2 − 𝑃(𝑡)2 ), thus
𝑃(𝑡) = |𝑡|.
40 Chapter 7. Function spaces
This implies that the sequence (𝑃𝑛 )𝑛⩾0 converges pointwisely to 𝑡 ↦ |𝑡| defined on
[−1, 1]. By Dini’s theorem (see Exercise sheet) we deduce that (𝑃𝑛 )𝑛⩾0 converges uni-
formly to 𝑡 ↦ |𝑡| defined on [−1, 1].
Proof of Theorem 7.9. First of all, observe that 𝒜 is also a subalgebra of 𝒞 (𝑋 , R). We
split the proof into three steps.
Step 1. We shall prove that if 𝑓 , 𝑔 ∈ 𝒜 then |𝑓 | ∈ 𝒜 , min(𝑓 , 𝑔) ∈ 𝒜 , and max(𝑓 , 𝑔) ∈ 𝒜 .
Let 𝑓 ∈ 𝒜 . If 𝑓 = 0 then it is clear that 𝑓 ∈ 𝒜 , so we assume that 𝑓 ≠ 0. The sequence
𝑃𝑛 (𝑓 /‖𝑓 ‖∞ ) converges uniformly to |𝑓 |/‖𝑓 ‖∞ thanks to Lemma 7.10. Since 𝑃𝑛 (𝑓 /‖𝑓 ‖∞ ) ∈
𝒜 for any 𝑛 ∈ N, it follows that |𝑓 |/‖𝑓 ‖∞ ∈ 𝒜 and hence |𝑓 | ∈ 𝒜 .
Let 𝑓 , 𝑔 ∈ 𝒜 . Since
1 1
min(𝑓 , 𝑔) = (𝑓 + 𝑔 − |𝑓 − 𝑔|) and min(𝑓 , 𝑔) = (𝑓 + 𝑔 + |𝑓 − 𝑔|),
2 2
𝛽 −𝛼 𝛼𝑔(𝑦) − 𝛽𝑔(𝑥)
ℎ(𝑧) = 𝑔(𝑧) + , ∀𝑧 ∈ 𝑋 ,
𝑔(𝑦) − 𝑔(𝑥) 𝑔(𝑦) − 𝑔(𝑥)
which is non-empty and open in 𝑋 . Since 𝑋 = ⋃𝑥∈𝑋 𝑉𝑥 and 𝑋 is compact, there are
𝑥1 , … , 𝑥𝑚 ∈ 𝑋 such that 𝑋 = 𝑉𝑥1 ∪ ⋯ ∪ 𝑉𝑥𝑚 . Consider the function 𝑔 = max(𝑔𝑥1 , … , 𝑔𝑥𝑚 )
so that 𝑔 ∈ 𝒜 . On the one hand we have 𝑔 < 𝑓 + 𝜀 since 𝑔𝑥𝑖 < 𝑓 + 𝜀 for any 𝑖 ∈ {1, … , 𝑚}.
On the other hand, for any 𝑧 ∈ 𝑋 there is 𝑖 ∈ {1, … , 𝑚} such that 𝑧 ∈ 𝑉𝑥𝑖 , hence 𝑔𝑥 (𝑧) ⩾
𝑔𝑥𝑖 (𝑧) > 𝑓 (𝑧) − 𝜀.
Therefore for any 𝑓 ∈ 𝒞 (𝑋 , R) and 𝜀 > 0 there exists 𝑔 ∈ 𝒜 such that 𝑑∞ (𝑓 , 𝑔) < 𝜀,
which means that 𝒜 is dense in 𝒞 (𝑋 , R), hence 𝒜 is also dense in 𝒞 (𝑋 , R).
It is worth noting that the above result regards the space of continuous real-valued
maps and one can prove that it is false in the case of complex-valued maps. One can
however prove a complex version of the Stone-Weierstrass theorem by adding a further
hypothesis.
7.5 Stone-Weierstrass theorem 41
Through this chapter we shall always consider K-vector spaces where K is the field of
real numbers R or complex numbers C. Moreover, by a linear map we shall always mean
a K-linear map.
Proof. (i) Let 𝑥0 ∈ 𝑋 , we deduce that the map 𝑥 ↦ ‖𝑥‖ is continuous at 𝑥0 thanks to the
inequality
|‖𝑥‖ − ‖𝑥0 ‖| ⩽ ‖𝑥 − 𝑥0 ‖.
43
44 Chapter 8. Banach spaces and continuous linear maps
(ii) Let (𝜆0 , 𝑥0 ) ∈ K × 𝑋 , we deduce that the map (𝜆, 𝑥) ↦ 𝜆𝑥 is continuous at (𝜆0 , 𝑥0 )
since one has
‖𝜆𝑥 − 𝜆0 𝑥0 ‖ ⩽ |𝜆 − 𝜆0 |‖𝑥‖ + |𝜆0 |‖𝑥 − 𝑥0 ‖.
(iii) Let (𝑥0 , 𝑦0 ) ∈ 𝑋 × 𝑋 , we deduce that the map (𝑥, 𝑦) ↦ 𝑥 + 𝑦 is continuous at (𝑥0 , 𝑦0 )
since one has
‖(𝑥 + 𝑦) − (𝑥0 + 𝑦0 )‖ ⩽ ‖𝑥 − 𝑥0 ‖ + ‖𝑦 − 𝑦0 ‖.
Definition 8.5. Let (𝑋 , ‖ ⋅ ‖) be a normed vector space and 𝐴 a subset of 𝑋 . One says
that:
(i) 𝐴 is bounded if sup𝑥∈𝐴 ‖𝑥‖ < ∞;
(ii) 𝐴 is convex if for any 𝑥, 𝑦 ∈ 𝐴 and any 𝑡 ∈ [0, 1] one has 𝑡𝑥 + (1 − 𝑡)𝑦 ∈ 𝐴.
One remarks that 𝐴 is bounded if and only if 𝐴 is contained in some open ball.
Definition 8.6. Two norms ‖ ⋅ ‖1 and ‖ ⋅ ‖2 on a vector space 𝑋 are said to be equivalent
if there are positive constants 𝛼, 𝛽 > 0 such that
Definition 8.7. A Banach space is a normed space that is complete with respect to the
metric associated to its norm.
Example 8.1. On K𝑁 we can define the norms, for 𝑥 = (𝑥1 , … , 𝑥𝑁 ) ∈ K𝑁 ,
𝑁 1/𝑝
‖𝑥‖𝑝 = ( ∑ |𝑥𝑘 𝑝
| ) for 1 ⩽ 𝑝 < ∞, ‖𝑥‖∞ = sup |𝑥𝑘 |.
𝑘=1 1⩽𝑘⩽𝑁
The space (𝒞 ([0, 1]; K), ‖ ⋅ ‖∞ ) is a Banach space, however (𝒞 ([0, 1]; K), ‖ ⋅ ‖1 ) is not.
8.2 Finite-dimensional normed vector spaces 45
‖𝑓 ‖∞ ∶= sup |𝑓 (𝑥)|.
𝑥∈𝑋
which implies that ‖𝑥‖ ⩽ 𝑐1 ‖𝑥‖∞ , where 𝑐1 = ‖𝑒1 ‖ + ⋯ + ‖𝑒𝑛 ‖ > 0. This gives the first
inequality.
Moreover, last inequality implies that the map ‖ ⋅ ‖ ∶ (𝑋 , ‖ ⋅ ‖∞ ) → R is continuous.
Since 𝑆 ∶= {𝑥 ∈ 𝑋 ∣ ‖𝑥‖∞ = 1} is compact, the image of 𝑆 under ‖ ⋅ ‖ is a compact
subset of R and therefore it possess a minimum, that is, there is some 𝑥0 ∈ 𝑆 such that
‖𝑥‖ ⩾ ‖𝑥0 ‖ > 0 for any 𝑥 ∈ 𝑆. By homogeneity of the norm and setting 𝑐2 = ‖𝑥0 ‖, we
finally get that
∀𝑥 ∈ 𝑋 , 𝑐2 ‖𝑥‖∞ ⩽ ‖𝑥‖,
which completes the proof.
Proposition 8.10. If 𝑋 is a finite-dimensional vector space, then the compact sets in 𝑋
are the closed and bounded sets.
Proof. We will show that the compact sets in (R𝑁 , ‖ ⋅ ‖∞ ) are the closed and bounded
sets, and the case of a finite-dimensional vector space follows from it since all norms are
equivalent.
We already know that if 𝐴 ⊆ R𝑁 is compact then it is closed and bounded. Con-
versely, let 𝐴 ⊆ R𝑁 be bounded and closed, then there is 𝑀 > 0 such that 𝐴 ⊆ [−𝑀, 𝑀]𝑁 ,
and the set [−𝑀, 𝑀]𝑁 is compact as the product of compact sets. Therefore 𝐴 is a closed
subset of a compact set, thus 𝐴 is compact.
Theorem 8.11 (Riesz theorem). The closed unit ball 𝐵(0, 1) of a normed vector space 𝑋
is compact if and only if 𝑋 is finite-dimensional.
46 Chapter 8. Banach spaces and continuous linear maps
Proof. If 𝑋 is finite-dimensional, we already know that the closed unit ball 𝐵(0, 1) is
compact since it is bounded and closed.
Conversely, assume that 𝐵(0, 1) is compact. Since {𝐵(𝑥, 1/2)}𝑥∈𝐵(0,1) is a cover of
𝐵(0, 1) by open sets in 𝑋 , there is a finite subset {𝑒1 , … , 𝑒𝑁 } ⊆ 𝐵(0, 1) such that 𝐵(0, 1) ⊆
𝐵(𝑒1 , 12 ) ∪ ⋯ ∪ 𝐵(𝑒𝑁 , 21 ). Define 𝑌 = span(𝑒1 , … , 𝑒𝑁 ), we will show that 𝑌 = 𝑋 which will
complete the proof. Since 𝑌 is closed, it suffices to show that 𝑌 is dense in 𝑋 . Let 𝑥 ∈ 𝑋 ,
then there are 𝑦 ∈ 𝑌 and 𝑎 ∈ Z such that ‖𝑥 − 𝑦‖ ⩽ 2−𝑎 . Then we have 2𝑎 (𝑥 − 𝑦) ∈ 𝐵(0, 1)
and hence, by definition of 𝑌 , there is 𝑗 ∈ {1, … , 𝑁 } such that ‖2𝑎 (𝑥 − 𝑦) − 𝑒𝑗 ‖ ⩽ 1/2.
Therefore one has that 𝑦1 ∶= 𝑦 + 2−𝑎 𝑒𝑗 ∈ 𝑌 and ‖𝑥 − 𝑦1 ‖ ⩽ 2−𝑎−1 . By induction we can
construct a sequence (𝑦𝑛 )𝑛⩾0 of elements of 𝑌 such that ‖𝑥 − 𝑦𝑛 ‖ ⩽ 2−𝑎−𝑛 , thus 𝑥 belongs
to the closure of 𝑌 .
∀𝑥 ∈ 𝑋 , ‖𝑥‖𝑋 ⩽ 𝛿 ⇒ ‖𝑢(𝑥)‖𝑌 ⩽ 1.
‖𝑥‖𝑋 𝛿
Now for any 𝑥 ≠ 0 we write 𝑥 = 𝛿 ‖𝑥‖𝑋
𝑥, hence
‖𝑥‖𝑋 ‖𝑥‖𝑋
‖𝑇 𝑥‖𝑌 = ‖𝑇 ( ‖𝑥‖𝛿 𝑥)‖ ⩽ .
𝛿 𝑋 𝑌 𝛿
(iii) ⇒ (i) Assume that 𝑇 is bounded. Then, for any 𝑥, 𝑦 ∈ 𝑋 , one has
and also
‖𝑇 𝑥‖𝑌 ⩽ ‖𝑇 ‖ℒ (𝑋 ,𝑌 ) ‖𝑥‖𝑋 .
Proof. Let (𝑇𝑛 )𝑛⩾0 be a Cauchy sequence in ℒ (𝑋 , 𝑌 ). For any 𝜀 > 0 there is 𝑁0 ∈ N such
that for any integers 𝑛, 𝑚 ⩾ 𝑁0 we have ‖𝑇𝑛 − 𝑇𝑚 ‖ℒ (𝑋 ,𝑌 ) < 𝜀. Thus, for any 𝑥 ∈ 𝑋 , we
have
‖𝑇𝑛 𝑥 − 𝑇𝑚 𝑥‖𝑌 ⩽ ‖𝑇𝑛 − 𝑇𝑚 ‖ℒ (𝑋 ,𝑌 ) ‖𝑥‖𝑋 < 𝜀 ‖𝑥‖𝑋 . (⋆)
It follows that, for any 𝑥 ∈ 𝑋 , the sequence (𝑇𝑛 𝑥)𝑛⩾0 is a Cauchy sequence in 𝑌 , which is
complete, hence there is 𝑦𝑥 ∈ 𝑌 such that (𝑇 𝑥𝑛 )𝑛⩾0 converges to 𝑦𝑥 in 𝑌 . Define the map
𝑇 ∶ 𝑥 ↦ 𝑦𝑥 from 𝑋 to 𝑌 , which clearly is linear. It remains to show that 𝑇 is continuous
and that (𝑇𝑛 )𝑛⩾0 converges to 𝑇 in ℒ (𝑋 , 𝑌 ).
From (⋆) we obtain that, for any 𝑥 ∈ 𝑋 ,
which implies that 𝑇 is continuous. Finally, passing to the limit 𝑚 → ∞ in (⋆) we obtain,
for any 𝑥 ∈ 𝑋 and 𝑛 ⩾ 𝑁0 ,
Proof. Let 𝑥 ∈ 𝑋 , then there is a sequence (𝑥𝑛 )𝑛∈N in 𝐷 converging to 𝑥 in 𝑋 . For all
𝑛, 𝑚 ∈ N we have
‖𝑇 𝑥𝑛 − 𝑇 𝑥𝑚 ‖𝑌 ⩽ ‖𝑇 ‖ℒ (𝐷,𝑌 ) ‖𝑥𝑛 − 𝑥𝑚 ‖𝑋 ,
thus (𝑇 𝑥𝑛 )𝑛∈N is a Cauchy sequence in 𝑌 which is complete, hence it converges to some
limit.
We define the map 𝑇̂ ∶ 𝑋 → 𝑌 , 𝑥 ↦ lim𝑛→∞ 𝑇 𝑥𝑛 , where (𝑥𝑛 )𝑛∈N is a sequence in
𝐷 converging to 𝑥. We first check that 𝑇̂ is well-defined, that is, it does not depend on
the choice of the sequence converging to 𝑥. Indeed if (𝑥𝑛′ )𝑛∈N is another sequence in 𝐷
converging to 𝑥, then
The map 𝑇̂ is clearly linear. For all 𝑥 ∈ 𝐷, we consider the constant sequence (𝑥)𝑛∈N
so that 𝑇̂𝑥 = lim𝑛→∞ 𝑇 𝑥 = 𝑇 𝑥, that is, 𝑇̂ is an extension of 𝑇 .
We now prove that 𝑇̂ is continuous. Let 𝑥 ∈ 𝑋 and (𝑥𝑛 )𝑛∈N a sequence in 𝐷 converg-
ing to 𝑥, then
8.3.2 Isomorphisms
Definition 8.15. Let 𝑋 and 𝑌 be normed vector spaces.
(i) A map 𝑓 ∶ 𝑋 → 𝑌 is an isometry if it preserves the distance, that is, if ‖𝑓 (𝑥) −
𝑓 (𝑦)‖𝑌 = ‖𝑥 − 𝑦‖𝑋 for all 𝑥, 𝑦 ∈ 𝑋 .
(ii) A map 𝑓 ∶ 𝑋 → 𝑌 is an isomorphism if it is bijective and if both 𝑓 and its inverse
map 𝑓 −1 are linear and continuous. Two normed vector spaces are said to be
isomorphic if there is a isomorphism between them.
(iii) A map 𝑓 ∶ 𝑋 → 𝑌 is an isometric isomorphism if it is an isomorphism that is also
an isometry. Two normed vector spaces are said to be isometrically isomorphic if
there is a isometric isomorphism between them.
Remark 8.16. Remark that if a map 𝑓 ∶ 𝑋 → 𝑌 is linear and bijective, then its inverse
map is also linear.
We shall denote by Isom(𝑋 , 𝑌 ) the set of all isomorphisms from 𝑋 into 𝑌 . We have
the following result concerning the set of isomorphisms and the inversion map, which
is left as an exercise.
or
‖(𝑥1 , 𝑥2 )‖𝑋1 ×𝑋2 = ‖𝑥1 ‖𝑋1 + ‖𝑥2 ‖𝑋2 ,
or still any other equivalent norm.
Proof. The proof is similar to the proof of Theorem 8.12, and hence is left as an exercise.
‖𝑇 (𝑥1 , 𝑥2 )‖𝑌
‖𝑇 ‖ℒ2 (𝑋1 ,𝑋2 ;𝑌 ) ∶= sup
𝑥1 ∈𝑋1 ⧵{0}, 𝑥2 ∈𝑋2 ⧵{0} ‖𝑥1 ‖𝑋1 ‖𝑥2 ‖𝑋2
and thus (ℒ2 (𝑋1 , 𝑋2 ; 𝑌 ), ‖⋅‖ℒ2 (𝑋1 ,𝑋2 ;𝑌 ) ) is a normed vector space. As for continuous linear
maps, one can easily obtain that
Proposition 8.21. The spaces ℒ2 (𝑋1 , 𝑋2 ; 𝑌 ) and ℒ (𝑋1 , ℒ (𝑋2 , 𝑌 )) are isomorphic. More
precisely, the map
ℒ2 (𝑋1 , 𝑋2 ; 𝑌 ) → ℒ (𝑋1 , ℒ (𝑋2 , 𝑌 ))
𝑇 ↦ {𝑥 ↦ (𝑦 ↦ 𝑇 (𝑥, 𝑦))}
is an isometric isomorphism between normed spaces, and its inverse map is given by
Therefore the map 𝑥 ↦ 𝑇𝑥 is continuous with norm less or equal than ‖𝑇 ‖ℒ2 (𝑋1 ,𝑋2 ;𝑌 ) .
Since one has
‖𝑇𝑥 ‖ℒ (𝑋2 ,𝑌 ) ‖𝑇 (𝑥, 𝑦)‖𝑌
sup = sup sup = ‖𝑇 ‖ℒ2 (𝑋1 ,𝑋2 ;𝑌 ) ,
𝑥∈𝑋1 ⧵{0} ‖𝑥‖𝑋1 𝑥∈𝑋1 ⧵{0} 𝑦∈𝑋2 ⧵{0} ‖𝑥‖𝑋1 ‖𝑦‖𝑋2
and thus 𝐵𝑌 (0, 𝑟/2) ⊆ 𝑇 (𝐵𝑋 (0, 1)). Observe that this implies that for any 𝑦 ∈ 𝑌 such that
‖𝑦‖𝑌 < 𝜆 2𝑟 with 𝜆 > 0, we have
∀𝜀 > 0, ∃𝑧 ∈ 𝑋 , ‖𝑧‖𝑋 < 𝜆 and ‖𝑦 − 𝑇 𝑧‖𝑌 < 𝜀.
Now let 𝑦 ∈ 𝑌 with ‖𝑦‖𝑌 < 𝑟/4, then taking 𝜀 = 𝑟/4 we find 𝑧1 ∈ 𝑋 such that
‖𝑧1 ‖𝑋 < 1/2 and ‖𝑦 − 𝑇 𝑧1 ‖𝑌 < 𝑟/4. By applying the same argument to 𝑦 − 𝑇 𝑧1 that
verifies ‖𝑦 − 𝑇 𝑧1 ‖𝑌 < 𝑟/4 and taking 𝜀 = 𝑟/8, we find 𝑧2 ∈ 𝑋 such that ‖𝑧2 ‖𝑋 < 1/4 and
‖𝑦 − 𝑇 𝑧1 − 𝑇 𝑧2 ‖𝑌 < 𝑟/8. By induction we construct a sequence (𝑧𝑛 )𝑛∈N∗ in 𝑋 such that
for any 𝑛 ∈ N∗ we have
1
‖𝑧𝑛 ‖𝑋 < 𝑛 and ‖𝑦 − 𝑇 (𝑧1 + ⋯ + 𝑧𝑛 )‖𝑌 < 𝑟2−𝑛−1 .
2
Therefore the sequence (𝑥𝑛 )𝑛∈N∗ , defined by 𝑥𝑛 = 𝑧1 + ⋯ + 𝑧𝑛 , is a Cauchy sequence
in 𝑋 , which is complete. Hence there is some 𝑥 ∈ 𝑋 such that 𝑥𝑛 → 𝑥, and we have
‖𝑥‖𝑋 < 1 and 𝑦 = 𝑇 𝑥 since 𝑇 is continuous. We have hence proved that for any 𝑦 ∈ 𝑌
with ‖𝑦‖𝑌 < 𝑟/4 there is 𝑥 ∈ 𝑋 with ‖𝑥‖𝑋 < 1 such that 𝑦 = 𝑇 𝑥, which implies that
𝐵𝑌 (0, 𝑟/4) ⊆ 𝑇 (𝐵𝑋 (0, 1)).
We now prove that the above estimate implies that 𝑇 is an open map. Let 𝑉 be open
in 𝑋 . Let 𝑦 ∈ 𝑇 (𝑉 ), then there is 𝑥 ∈ 𝑉 such that 𝑦 = 𝑇 𝑥. Since 𝑉 is open, there is
𝑟 > 0 so that 𝐵𝑋 (𝑥, 𝑟) ⊆ 𝑉 . Hence 𝑥 + 𝐵𝑋 (0, 𝑟) = 𝐵𝑋 (𝑥, 𝑟) ⊆ 𝑉 from which we obtain
𝑦 + 𝑇 (𝐵𝑋 (0, 𝑟)) ⊆ 𝑢(𝑉 ). By hypothesis we have 𝐵𝑌 (0, 𝑟𝑐) ⊆ 𝑇 (𝐵𝑋 (0, 𝑟)), then it follows
𝐵𝑌 (𝑦, 𝑟𝑐) = 𝑦 + 𝐵𝑌 (0, 𝑟𝑐) ⊆ 𝑦 + 𝑇 (𝐵𝑋 (0, 𝑟)) ⊆ 𝑇 (𝑉 ). This proves that 𝑇 (𝑉 ) is open in
𝑌.
52 Chapter 8. Banach spaces and continuous linear maps
Proof. The map 𝑇 is bijective, continuous and open by the Open mapping theorem (The-
orem 8.23), hence 𝑇 is a homeomorphism. Therefore 𝑇 −1 is continuous. Since 𝑇 −1 is
also linear, this implies that 𝑇 is an isomorphism.
Recall that the converse is true, indeed we already know that the graph of a contin-
uous map (not necessarily linear) is closed.
We remark that saying that the graph Γ of 𝑇 is closed in 𝑋 × 𝑌 is equivalent to: For
every sequence (𝑥𝑛 )𝑛⩾0 in 𝑋 that converges to 𝑥 ∈ 𝑋 and is such that (𝑇 𝑥𝑛 )𝑛⩾0 converges
to 𝑦 ∈ 𝑌 , then one has 𝑦 = 𝑇 𝑥.
Proof of Theorem 8.25. Recall that the product space 𝑋 × 𝑌 is a Banach space. Since 𝑇 is
linear, it follows that its graph Γ is a closed vector subspace of 𝑋 × 𝑌 , thus also a Banach
space.
The projection 𝜋 ∶ (𝑥, 𝑦) ↦ 𝑥 is a continuous linear bijective map from Γ onto 𝑋 ,
hence by Corollary 8.24, the map 𝜋 −1 ∶ 𝑥 ↦ (𝑥, 𝑇 𝑥) from 𝑋 onto Γ is also continuous.
The projection 𝜃 ∶ (𝑥, 𝑦) ↦ 𝑦 from 𝑋 ×𝑌 into 𝑌 is also a continuous linear map, therefore
one deduces that 𝑇 = 𝜃 ∘ 𝜋 −1 is continuous.
Definition 8.26. One says that a map 𝑓 ∶ [𝑎, 𝑏] → 𝑌 is a step function if there is a
partition 𝜎 = {𝑡0 , … , 𝑡𝑛+1 } of [𝑎, 𝑏] and 𝑦0 , … , 𝑦𝑛 ∈ 𝑌 such that, for all 𝑘 = 0, … , 𝑛, one has
One can check that this definition does not depend on the choice of partition, that
is, if 𝜎 ′ is another finite partition of [𝑎, 𝑏] such that 𝑓 is constant on the open intervals
of 𝜎 ′ , then the value of (8.1) associated to 𝜎 coincides with the value of (8.1) associated
to 𝜎 ′ .
Denote by 𝑆([𝑎, 𝑏]; 𝑌 ) ⊆ ℬ([𝑎, 𝑏]; 𝑌 ) the set of all step functions from [𝑎, 𝑏] to 𝑌 . The
map
ℐ ∶ 𝑆([𝑎, 𝑏]; 𝑌 ) → 𝑌
𝑏
𝑓 ↦ ∫ 𝑓 (𝑡) d𝑡
𝑎
is linear and continuous. Moreover we easily obtain from the definition the following
properties:
Lemma 8.27. For any step function 𝑓 ∈ 𝑆([𝑎, 𝑏]; 𝑌 ) there holds:
(i) Let 𝑎 < 𝑐 < 𝑏 then
𝑏 𝑐 𝑏
∫ 𝑓 (𝑡) d𝑡 = ∫ 𝑓 (𝑡) d𝑡 + ∫ 𝑓 (𝑡) d𝑡.
𝑎 𝑎 𝑐
(ii) We have
𝑏 𝑏
‖∫ 𝑓 (𝑡) d𝑡‖ ⩽ ∫ ‖𝑓 (𝑡)‖𝑌 d𝑡 ⩽ (𝑏 − 𝑎)‖𝑓 ‖∞ .
𝑎 𝑌 𝑎
We now define the integral of a larger class of functions by extending the continuous
linear map ℐ .
Definition 8.28. One says that a map 𝑓 ∶ [𝑎, 𝑏] → 𝑌 is a regulated function if it belongs
to 𝑆([𝑎, 𝑏]; 𝑌 ), the closure of 𝑆([𝑎, 𝑏]; 𝑌 ) in ℬ([𝑎, 𝑏]; 𝑌 ).
where (𝑓𝑛 )𝑛⩾0 is a sequence of step functions that converges uniformly to 𝑓 , and we
recall that, by the proof of Theorem 8.14, this limit is independent of the chosen sequence.
Thanks to this procedure, properties (i) and (ii) above remain true for 𝑓 ∈ 𝑆([𝑎, 𝑏]; 𝑌 ).
Chapter 9
Differentiable maps
Through this chapter, let (𝑋 , ‖ ⋅ ‖𝑋 ) and (𝑌 , ‖ ⋅ ‖𝑌 ) be Banach spaces over K. When there
is no confusion, we omit the subscripts 𝑋 and 𝑌 of the norms.
Proof. Assume that there are two maps 𝐿1 , 𝐿2 ∈ ℒ (𝑋 , 𝑌 ) and two neighborhoods 𝑉1 , 𝑉2
of 0 in 𝑋 such that 𝑎 + 𝑉1 ⊆ 𝑈 , 𝑎 + 𝑉2 ⊆ 𝑈 and
This implies that, for any 𝜂 > 0, there is 𝛿 > 0 such that for any ℎ ∈ 𝑉1 ∩ 𝑉2 one has
By homogeneity of the norm, this implies that ‖𝐿1 − 𝐿2 ‖ ⩽ 𝜂, and we conclude that
𝐿1 = 𝐿2 since 𝜂 > 0 is arbitrary.
Definition 9.3. The map 𝐿 in Definition 9.1 is called the Fréchet derivative of 𝑓 at the
point 𝑎, and it is denoted by
d𝑓 (𝑎) or 𝑓 ′ (𝑎).
55
56 Chapter 9. Differentiable maps
Remark 9.4. The differentiability of 𝑓 and the value of 𝑓 ′ (𝑎) only depend on the topolo-
gies of 𝑋 and 𝑌 . In particular, they do not change if we replace the norms of 𝑋 and 𝑌 by
equivalent ones.
Definition 9.5. Consider a map 𝑓 ∶ 𝑈 → 𝑌 .
(i) One says that 𝑓 is Fréchet-differentiable on 𝑈 if 𝑓 is Fréchet-differentiable at every
point of 𝑈 . In that case we define the Fréchet derivative of 𝑓 , denoted by d𝑓 or 𝑓 ′ ,
as the map
𝑓 ′ ∶ 𝑈 → ℒ (𝑋 , 𝑌 )
.
𝑎 ↦ 𝑓 ′ (𝑎)
(ii) One says that 𝑓 is continuously differentiable or of class 𝒞 1 on 𝑈 if 𝑓 is Fréchet-
differentiable on 𝑈 and the Fréchet derivative 𝑓 ′ is continuous.
We denote by 𝒞 1 (𝑈 , 𝑌 ) the vector space of all continuously differentiable maps on
𝑈.
Hereafter, when there is no possible misunderstanding, Fréchet differentiable shall be
referred to simply as differentiable, and Fréchet derivative to derivative.
Example 9.1. A constant map 𝑓 ∶ 𝑈 → 𝑌 is differentiable on 𝑈 and for any 𝑎 ∈ 𝑈 one
has 𝑓 ′ (𝑎) = 0.
Example 9.2. A continuous linear map 𝑓 ∶ 𝑈 → 𝑌 is differentiable on 𝑈 and for any
𝑎 ∈ 𝑈 one has 𝑓 ′ (𝑎) = 𝑓 .
Example 9.3. When 𝑋 = R, then 𝑓 ′ (𝑎) ∈ ℒ (R, 𝑌 ) is identified with the vector 𝜆 =
𝑓 ′ (𝑎)(1) ∈ 𝑌 , through the canonical isomorphism ℒ (R, 𝑌 ) ≃ 𝑌 , where
𝑓 (𝑥) − 𝑓 (𝑎)
𝜆 ∶= lim
𝑥→𝑎 𝑥 −𝑎
is the usual derivative for functions defined on R.
and one has limℎ→0 𝜆𝜀(ℎ) = 0. The map 𝑋 ∋ ℎ ↦ 𝜆𝑓 ′ (𝑎)(ℎ) ∈ 𝑌 is clearly linear and
continuous, thus the proof is complete.
(ii) There are 𝑉1 , 𝑉2 neighborhoods of 0 in 𝑋 such that 𝑎 + 𝑉1 ⊆ 𝑈 , 𝑎 + 𝑉2 ⊆ 𝑈 , and
(𝑓 + 𝑔)(𝑎 + ℎ) = 𝑓 (𝑎 + ℎ) + 𝑔(𝑎 + ℎ)
= (𝑓 + 𝑔)(𝑎) + 𝑓 ′ (𝑎)(ℎ) + 𝑔 ′ (𝑎)(ℎ) + ‖ℎ‖(𝜀1 (ℎ) + 𝜀2 (ℎ)),
and one has limℎ→0 (𝜀1 (ℎ) + 𝜀2 (ℎ)) = 0. The map 𝑋 ∋ ℎ ↦ 𝑓 ′ (𝑎)(ℎ) + 𝑔 ′ (𝑎)(ℎ) ∈ 𝑌 is
clearly linear and continuous, thus the proof is complete.
In the next result we show that the composition of differentiable maps is also differ-
entiable.
Proposition 9.8. Let 𝑋 , 𝑌 , 𝑍 be Banach spaces, 𝑈 an open subset of 𝑋 , and 𝑉 an open
subset of 𝑌 . Let 𝑓 ∶ 𝑈 → 𝑌 and 𝑔 ∶ 𝑉 → 𝑍 be two maps. Assume that 𝑓 is differentiable
at 𝑎 ∈ 𝑈 and 𝑔 is differentiable at 𝑏 = 𝑓 (𝑎) ∈ 𝑉 . Then the composition map 𝑔 ∘ 𝑓 ∶ 𝑈 → 𝑍
is differentiable at 𝑎 and
(𝑔 ∘ 𝑓 )′ (𝑎) = 𝑔 ′ (𝑓 (𝑎)) ∘ 𝑓 ′ (𝑎).
Proof. By hypothesis there are 𝛿, 𝛿 ′ > 0 such that
with 𝜀 ∶ 𝐵𝑋 (0, 𝛿) → 𝑌 and 𝜂 ∶ 𝐵𝑌 (0, 𝛿 ′ ) → 𝑍 satisfying limℎ→0 𝜀(ℎ) = 0 and lim𝑘→0 𝜂(𝑘) =
0. We set 𝑘 = 𝑓 (𝑎 + ℎ) − 𝑓 (𝑎) = 𝑓 ′ (𝑎)(ℎ) + ‖ℎ‖𝜀(ℎ) so that
Let 𝛿 > 0 be small enough such that ‖ℎ‖ < 𝛿 implies ‖𝑘‖ < 𝛿 ′ . Hence one obtains
It remains to check that lim𝑘→0 𝜂(𝑘) = 0 and the proof will be complete. Since ‖ℎ‖ ⩽ 2𝑐‖𝑘‖
for ℎ small enough, we obtain ‖𝜂(𝑘)‖ ⩽ 2𝑐 2 ‖𝜀(ℎ)‖ and we conclude by composition of
limits since 𝑘 → 0 implies ℎ → 0.
9.3 Maps with values in a product space 59
𝑝𝑖 ∶ 𝑌 → 𝑌𝑖
(𝑦1 , … , 𝑦𝑚 ) ↦ 𝑦𝑖
𝜄𝑖 ∶ 𝑌 𝑖 → 𝑌
𝑦𝑖 ↦ (0, … , 𝑦𝑖 , … , 0)
where there are 0’s everywhere except at the 𝑖-th coordinate. One easily checks that 𝑝𝑖
and 𝜄𝑖 are continuous linear maps and verify
𝑚
𝑝𝑖 ∘ 𝜄𝑖 = id𝑌𝑖 and ∑ 𝜄𝑖 ∘ 𝑝𝑖 = id𝑌 .
𝑖=1
Remark 9.13. A map that is differentiable with respect to every variable at some point
is not necessarily differentiable at this point.
Proof of Proposition 9.12. Let 𝜄𝑖 denote the canonical injection
𝜄𝑖 ∶ 𝑋𝑖 → 𝑋
𝑥𝑖 ↦ (0, … , 𝑥𝑖 , … , 0)
where there are 0’s everywhere except at the 𝑖-th coordinate, which clearly is a contin-
uous linear map. Remark that
thus 𝜆𝑖′ (𝑥𝑖 ) = 𝜄𝑖 for all 𝑥𝑖 ∈ 𝑋𝑖 . By composition of differentiable maps, we deduce that
𝑓 ∘ 𝜆𝑖 is differentiable at the point 𝑎𝑖 ∈ 𝑋𝑖 and moreover
(𝑓 ∘ 𝜆𝑖 )′ (𝑎𝑖 ) = 𝑓 ′ (𝑎) ∘ 𝜄𝑖 ,
9.4 Partial derivatives 61
63
64 Chapter 10. Mean-value theorem and applications
𝜕𝑖 𝑓 ∶ 𝑈 → ℒ (𝑋𝑖 , 𝑌 )
𝑥 ↦ 𝜕𝑖 𝑓 (𝑥)
is continuous.
Proof. Assume that 𝑓 is of class 𝒞 1 on 𝑈 , then by Proposition 9.12 𝑓 is differentiable with
respect to any variable on 𝑈 . Moreover 𝜕𝑖 𝑓 is continuous by composition of continuous
maps.
Conversely, assume that 𝑓 is differentiable with respect to every variable on 𝑈 and
𝜕𝑖 𝑓 is continuous for any 𝑖 ∈ {1, … , 𝑛}. Let 𝑎 = (𝑎1 , … , 𝑎𝑛 ) ∈ 𝑈 . We want to prove that for
any 𝜀 > 0, there is 𝛿 > 0 such that for any ℎ ∈ 𝐵𝑋 (0, 𝛿) one has
𝑛
‖𝑓 (𝑎1 + ℎ1 , … , 𝑎𝑛 + ℎ𝑛 ) − 𝑓 (𝑎1 , … , 𝑎𝑛 ) − ∑ 𝜕𝑖 𝑓 (𝑎)(ℎ𝑖 )‖ < 𝜀‖ℎ‖.
𝑖=1
We hence write
𝑛
𝑓 (𝑎1 + ℎ1 , … , 𝑎𝑛 + ℎ𝑛 ) − 𝑓 (𝑎1 , … , 𝑎𝑛 ) − ∑ 𝜕𝑖 𝑓 (𝑎)(ℎ𝑖 )
𝑖=1
= 𝑓 (𝑎1 + ℎ1 , … , 𝑎𝑛 + ℎ𝑛 ) − 𝑓 (𝑎1 , 𝑎2 + ℎ2 , … , 𝑎𝑛 + ℎ𝑛 ) − 𝜕1 𝑓 (𝑎)(ℎ1 )
+ 𝑓 (𝑎1 , 𝑎2 + ℎ2 , … , 𝑎𝑛 + ℎ𝑛 ) − 𝑓 (𝑎1 , 𝑎2 , 𝑎3 + ℎ3 , … , 𝑎𝑛 ) − 𝜕2 𝑓 (𝑎)(ℎ2 )
+ ⋯ +
+ 𝑓 (𝑎1 , … , 𝑎𝑛−1 , 𝑎𝑛 + ℎ𝑛 ) − 𝑓 (𝑎1 , … , 𝑎𝑛−1 , 𝑎𝑛 ) − 𝜕𝑛 𝑓 (𝑎)(ℎ𝑛 )
so that it suffices to prove that for any 𝜀 > 0 there is 𝛿 > 0 such that ‖ℎ‖ < 𝛿 implies
𝜀
‖𝑓 (𝑎1 + ℎ1 , … , 𝑎𝑛 + ℎ𝑛 ) − 𝑓 (𝑎1 , 𝑎2 + ℎ2 , … , 𝑎𝑛 + ℎ𝑛 ) − 𝜕1 𝑓 (𝑎)(ℎ1 )‖ < ‖ℎ1 ‖,
𝑛
𝜀
‖𝑓 (𝑎1 , 𝑎2 + ℎ2 , … , 𝑎𝑛 + ℎ𝑛 ) − 𝑓 (𝑎1 , 𝑎2 , 𝑎3 + ℎ3 , … , 𝑎𝑛 ) − 𝜕2 𝑓 (𝑎)(ℎ2 )‖ < ‖ℎ2 ‖,
𝑛 (⋆)
⋮
𝜀
‖𝑓 (𝑎1 , … , 𝑎𝑛−1 , 𝑎𝑛 + ℎ𝑛 ) − 𝑓 (𝑎1 , … , 𝑎𝑛−1 , 𝑎𝑛 ) − 𝜕𝑛 𝑓 (𝑎)(ℎ𝑛 )‖ < ‖ℎ𝑛 ‖.
𝑛
Let then 𝜀 > 0. We will show that there is 𝛿 > 0 such that ‖ℎ‖ < 𝛿 implies the first
inequality in (⋆). The other inequalities can be obtained in a similar fashion, hence at
the end we can take 𝛿 > 0 to be the smallest one so that all the above inequalities hold.
Define the map 𝑔 ∶ 𝑡 ↦ 𝑓 (𝑎1 + 𝑡, 𝑎2 + ℎ2 , … , 𝑎𝑛 + ℎ𝑛 ) − 𝜕1 𝑓 (𝑎)(𝑡) for 𝑡 ∈ 𝑋1 small
enough. Then 𝑔 is differentiable and
Moreover 𝜕1 𝑓 is continuous at 𝑎, hence there is 𝛿 > 0 such that for any ‖ℎ‖ < 𝛿 one has
𝜀
‖𝜕1 𝑓 (𝑎 + ℎ) − 𝜕1 𝑓 (𝑎)‖ <
𝑛
and this implies that, for any ‖𝑡‖ ⩽ ‖ℎ1 ‖, there holds
𝜀
‖𝑔 ′ (𝑡)‖ = ‖𝜕1 𝑓 (𝑎1 + 𝑡, 𝑎2 + ℎ2 , … , 𝑎𝑛 + ℎ𝑛 ) − 𝜕1 𝑓 (𝑎1 , … , 𝑎𝑛 )‖ < .
𝑛
By Theorem 10.2 one deduces that ‖𝑔(ℎ1 ) − 𝑔(0)‖ < 𝑛𝜀 ‖ℎ1 ‖ which is exactly the first
inequality in (⋆).
We have then proved that 𝑓 is differentiable at 𝑎, for any 𝑎 ∈ 𝑈 , and that
𝑛
𝑓 ′ (𝑎)(ℎ 1 , … , ℎ𝑛 ) = ∑ 𝜕𝑖 𝑓 (𝑎)(ℎ𝑖 ),
𝑖=1
Proof. Let 𝑎 ∈ 𝑈 . For any 𝑥 ∈ 𝐵𝑋 (𝑎, 𝑟) and 𝑛, 𝑚 ∈ N there holds, thanks to Theorem 10.2,
‖𝑓𝑛 (𝑥) − 𝑓𝑚 (𝑥) − (𝑓𝑛 (𝑎) − 𝑓𝑚 (𝑎))‖ ⩽ ‖𝑥 − 𝑎‖ sup ‖𝑓𝑛′ (𝑦) − 𝑓𝑚′ (𝑦)‖
𝑦∈𝐵𝑋 (𝑎,𝑟)
(∗)
⩽𝑟 sup ‖𝑓𝑛′ (𝑦) − 𝑓𝑚′ (𝑦)‖.
𝑦∈𝐵𝑋 (𝑎,𝑟)
Therefore, if (𝑓𝑛 (𝑎))𝑛⩾0 converges in 𝑌 it follows that (𝑓𝑛 )𝑛⩾0 is a Cauchy sequence for
the uniform norm on 𝐵𝑋 (𝑎, 𝑟), hence (𝑓𝑛 )𝑛⩾0 converges uniformly on 𝐵𝑋 (𝑎, 𝑟) to some
map 𝑓 which is continuous.
Moreover, let 𝐴 = {𝑥 ∈ 𝑈 ∣ (𝑓𝑛 (𝑥))𝑛⩾0 converges in 𝑌 }, which is non-empty by
hypothesis. One easily obtains that 𝐴 is open and closed, hence 𝐴 = 𝑈 since 𝑈 is
connected, that is, lim𝑛→∞ 𝑓𝑛 (𝑥) = 𝑓 (𝑥) for any 𝑥 ∈ 𝑈 .
66 Chapter 10. Mean-value theorem and applications
Let 𝜀 > 0, then there is 𝑁0 ∈ N such that for any integer 𝑛, 𝑚 ⩾ 𝑁0 one has
𝜀
sup ‖𝑓𝑛′ (𝑦) − 𝑓𝑚′ (𝑦)‖ < ,
𝑦∈𝐵𝑋 (𝑎,𝑟) 3
𝑋 ×𝑋 → 𝑌
(ℎ, 𝑘) ↦ 𝑓 ″ (𝑎)(ℎ)(𝑘).
10.4 High-order derivatives 67
Define the map 𝑔 ∶ 𝑡 ↦ 𝑓 ′ (𝑎+𝑡ℎ+𝑘)−𝑓 ′ (𝑎+𝑡ℎ) from [0, 1] into 𝑌 , then 𝑔 is differentiable
(by composition of differentiable maps) and
‖𝑔(1) − 𝑔(0) − 𝑓 ″ (𝑎)(𝑘, ℎ)‖ ⩽ sup ‖𝑔 ′ (𝑡) − 𝑓 ″ (𝑎)(𝑘, ℎ)‖ ⩽ 2𝜀(‖ℎ‖ + ‖𝑘‖)2 ,
0⩽𝑡⩽1
By homogeneity, this last inequality holds for any (ℎ, 𝑘) ∈ 𝑋 × 𝑋 , and one concludes that
𝑓 ″ (𝑎)(ℎ, 𝑘) = 𝑓 ″ (𝑎)(𝑘, ℎ) since 𝜀 > 0 is arbitrary.
which relates the second-order derivative of 𝑓 at 𝑎 with the partial derivatives 𝜕𝑖 (𝜕𝑗 𝑓 )(𝑎).
Applying twice (the proof of) Theorem 10.3, one obtains:
Proposition 10.8. A map 𝑓 ∶ 𝑈 → 𝑌 is twice differentiable at 𝑎 ∈ 𝑈 if and only if
(i) the partial derivatives 𝜕𝑗 𝑓 exist on 𝑈 for any 𝑗 = 1, … , 𝑛 and they are continuous on
𝑈;
(ii) for any 𝑗 = 1, … , 𝑛, the partial derivatives 𝜕𝑖 (𝜕𝑗 𝑓 ) exist on 𝑈 for any 𝑖 = 1, … , 𝑛 and
they are continuous at 𝑎.
ℒ𝑛 (𝑋 , 𝑌 ) → ℒ𝑛 (𝑋 , … , 𝑋 ; 𝑌 )
𝑢 ↦ {(ℎ1 , … , ℎ𝑛 ) ↦ 𝑢(ℎ1 )(ℎ2 ) ⋯ (ℎ𝑛 )}.
One also obtains an analogue result of Proposition 10.7, which states that if 𝑓 ∶ 𝑈 →
𝑌 is 𝑛-times differentiable at some point 𝑎 ∈ 𝑈 , then its derivative 𝑓 (𝑛) (𝑎) ∈ ℒ𝑛 (𝑋 , 𝑌 ) is
a 𝑛-multilinear symmetric map, that is, for any (ℎ1 , … , ℎ𝑛 ) ∈ 𝑋 𝑛 and any permutation 𝜎
on {1, … , 𝑛}, one has
(ii) Assume that there is 𝑀 > 0 such that ‖𝑣 (𝑛+1) (𝑡)‖ ⩽ 𝑀 for any 𝑡 ∈ [0, 1], then
1 ″ 1 𝑀
‖𝑣(1) − 𝑣(0) − 𝑣 ′ (0) − 𝑣 (0) − ⋯ − 𝑣 (𝑛) (0)‖ ⩽ .
2! 𝑛! (𝑛 + 1)!
1
Proof. (i) Recall that if 𝑓 ∶ 𝐼 → 𝑌 is of class 𝒞 1 then 𝑓 (1) − 𝑓 (0) = ∫0 𝑓 ′ (𝑡) d𝑡. The result
follows applying this to the function
(1−𝑡)𝑛+1
(ii) Consider the function 𝑓 (𝑡) defined above, and define 𝑔(𝑡) = −𝑀 (𝑛+1)!
so that 𝑔 ′ (𝑡) =
(1−𝑡)𝑛
𝑀 𝑛!
. One has
𝑣 ′ (𝑡) = 𝑓 ′ (𝑎 + 𝑡ℎ)(ℎ)
𝑣 ″ (𝑡) = 𝑓 ″ (𝑎 + 𝑡ℎ)(ℎ, ℎ)
⋮
𝑣 (𝑛) (𝑡) = 𝑓 (𝑛) (𝑎 + 𝑡ℎ)(ℎ, … , ℎ).
Therefore, applying Proposition 10.10 we obtain the following two Taylor formulas.
70 Chapter 10. Mean-value theorem and applications
1 ″ 1
‖𝑓 (𝑎 + ℎ) − 𝑓 (𝑎) − 𝑓 ′ (𝑎)(ℎ) − 𝑓 (𝑎)(ℎ, ℎ) − ⋯ − 𝑓 (𝑛) (𝑎)(ℎ, … , ℎ)‖
2! 𝑛!
𝑀
⩽ ‖ℎ‖𝑛+1 .
(𝑛 + 1)!
One can actually obtain a third formula that is similar to that obtained in Theo-
rem 10.12, but under weaker hypothesis.
Using the symmetry of the derivatives 𝑓 (𝑘) (𝑎) and denoting (ℎ)𝑘 = (ℎ, … , ℎ) ∈ 𝐸 𝑘 , one
gets
1 (3) 2 1
𝜑 ′ (ℎ) = 𝑓 ′ (𝑎 + ℎ) − 𝑓 ′ (𝑎) − 𝑓 ″ (𝑎)(ℎ) − 𝑓 (ℎ) − ⋯ − 𝑓 (𝑛) (𝑎)(ℎ)𝑛−1 .
2! (𝑛 − 1)!
Therefore, by the induction hypothesis applied to 𝑓 ′ , one obtains that for any 𝜀 > 0,
there is 𝛿 > 0 such that for any ℎ ∈ 𝐸, ‖ℎ‖ ⩽ 𝛿 implies ‖𝜑 ′ (ℎ)‖ ⩽ 𝜀‖ℎ‖𝑛−1 . Hence, thanks
to Theorem 10.2, on deduces that ‖𝜑(1) − 𝜑(0)‖ ⩽ 𝜀‖ℎ‖𝑛 for any ‖ℎ‖ ⩽ 𝛿, which concludes
the proof since 𝜑(0) = 0.
Chapter 11
11.1 Diffeomorphisms
Let 𝑋 and 𝑌 be Banach spaces.
𝑔 ′ (𝑏) = (𝑓 ′ (𝑔(𝑏)))−1 , ∀𝑏 ∈ 𝑊 .
71
72 Chapter 11. Inverse function and implicit function theorems
Optimization
One says that 𝑓 has an (strict local/local/strict) extremum at 𝑎 if 𝑓 has a (strict lo-
cal/local/strict) minimum or maximum at 𝑎. Moreover, if 𝑉 is a subset of 𝑈 , one says that
𝑓 has an extremum on 𝑉 if the restriction map 𝑓|𝑉 ∶ 𝑉 → R of 𝑓 to 𝑉 has an extremum.
Proof. Assume that 𝑓 has a local minimum at 𝑎 (the proof of the other case being similar).
Then there is some neighborhood 𝑉1 of 0 in 𝑋 such that 𝑓 (𝑎 +ℎ)−𝑓 (𝑎) ⩾ 0 for any ℎ ∈ 𝑉1 .
75
76 Chapter 12. Optimization
hence
𝑓 ′ (𝑎)(ℎ) + ‖ℎ‖𝜀(𝑡ℎ) ⩾ 0,
and taking the limit 𝑡 → 0 one gets 𝑓 ′ (𝑎)(ℎ) ⩾ 0. Applying the same argument with −ℎ
instead of ℎ one obtains 𝑓 ′ (𝑎)(ℎ) ⩽ 0, which concludes the proof.
Proof. By way of contradiction, assume that there is some 𝑏 ∈ 𝑈 ⧵ {𝑎} such that 𝑓 (𝑏) <
𝑓 (𝑎). Since 𝑈 is convex one has 𝑡𝑏 + (1 − 𝑡)𝑎 ∈ 𝑈 for all 𝑡 ∈ [0, 1] and
Define the function 𝑔 ∶ 𝑡 ↦ 𝑓 (𝑡𝑏 + (1 − 𝑡)𝑎) from 𝐽 into R, where 𝐽 is some open interval
containing [0, 1]. It is clear that 𝑔 is differentiable and
𝑔(𝑡) − 𝑔(0)
𝑔 ′ (0) = lim ⩽ 𝑓 (𝑏) − 𝑓 (𝑎) < 0.
𝑡→0, 𝑡>0 𝑡
On the other hand one has 𝑔 ′ (0) = 𝑓 ′ (𝑎)(𝑏 − 𝑎), thus 𝑓 ′ (𝑎)(𝑏 − 𝑎) < 0, which is a
contradiction with 𝑓 ′ (𝑎) = 0.
Φ(𝑥, 𝑡) = 𝜑(𝑥 + 𝑡1 𝑢1 + ⋯ + 𝑡𝑝 𝑢𝑝 ).
In particular, one has 𝜕𝑡 Φ(𝑎, 0) = idR𝑝 ∈ ℒ (R𝑝 , R𝑝 ). Hence by the Implicit Function
Theorem 11.4 there is some open neighborhood 𝑊 of 𝑎 in 𝑋 and some function 𝑔 =
(𝑔1 , … , 𝑔𝑝 ) = 𝑊 → R𝑝 of class 𝒞 1 such that 𝑔(𝑎) = 0 and Φ(𝑥, 𝑔(𝑥)) = 0 for any 𝑥 ∈ 𝑊 ,
that is
𝑥 + 𝑔1 (𝑥)𝑢1 + ⋯ 𝑔𝑝 (𝑥)𝑢𝑝 ∈ 𝑆 = Φ−1 ({0}),
and moreover
𝑔 ′ (𝑎)(ℎ) = −(𝜕𝑡 Φ(𝑎, 0))−1 ∘ 𝜕𝑥 Φ(𝑎, 0)(ℎ),
hence
𝑔𝑖′ (𝑎)(ℎ) = −𝜑𝑖′ (𝑎)(ℎ) ∀𝑖 = 1, … , 𝑝.
Assume now that 𝑓 has a local maximum on 𝑆 at the point 𝑎 (the other case can be
treated in a similar fashion). For 𝑥 close enough to 𝑎 in 𝑊 one has
It follows that
𝑝
𝑓 ′ (𝑎)(ℎ) = ∑(−𝑔𝑖′ (𝑎)(ℎ))𝑓 ′ (𝑎)(𝑢𝑖 )
𝑖=1
which implies
𝑝
𝑓 ′ (𝑎)(ℎ) = ∑(𝑓 ′ (𝑎)(𝑢𝑖 ))𝜑𝑖′ (𝑎)(ℎ),
𝑖=1
which concludes the proof by taking 𝜆𝑖 = 𝑓 ′ (𝑎)(𝑢𝑖 ) for any 𝑖 = 1, … , 𝑝.
Proof. Since 𝑓 ″ is continuous there is 𝑟 > 0 such that ‖𝑓 ″ (𝑎 + ℎ) − 𝑓 ″ (𝑎)‖ ⩽ 𝛼2 for any
ℎ ∈ 𝐵𝑋 (0, 𝑟).
Define the function 𝑔 ∶ 𝑥 ↦ 𝑓 (𝑎 + 𝑥) − 12 𝑓 ″ (𝑎)(𝑥, 𝑥) from 𝐵𝑋 (0, 𝑟) into R. The map
𝑔 is also of class 𝒞 2 and one has
A.1 Sets
Let 𝑋 be a set. We denote by ∅ the empty set, and by 𝔓(𝑋 ) the powerset of 𝑋 , that is,
𝔓(𝑋 ) = {𝐴 ∶ 𝐴 ⊆ 𝑋 } is the set of all subsets of 𝑋 .
We shall adopt the convention that the union of an empty family of subsets of 𝑋 is equal
to the empty set ∅, and that the intersection of an empty family of subsets of 𝑋 is equal
to 𝑋 .
79
80 Chapter A. Reminders: Sets and maps
𝑥 ∈ 𝑋 ⧵ ⋂ 𝐴𝑖 ⟺ 𝑥 ∉ ⋂ 𝐴𝑖 ⟺ ∃𝑖 ∈ 𝐼 , 𝑥 ∉ 𝐴𝑖
𝑖∈𝐼 𝑖∈𝐼
⟺ ∃𝑖 ∈ 𝐼 , 𝑥 ∈ 𝑋 ⧵ 𝐴𝑖 ⟺ 𝑥 ∈ ⋃(𝑋 ⧵ 𝐴𝑖 ).
𝑖∈𝐼
Let 𝑥 ∈ 𝑋 , then
𝑥 ∈ 𝑋 ⧵ ⋃ 𝐴𝑖 ⟺ 𝑥 ∉ ⋃ 𝐴𝑖 ⟺ ∀𝑖 ∈ 𝐼 , 𝑥 ∉ 𝐴𝑖
𝑖∈𝐼 𝑖∈𝐼
⟺ ∀𝑖 ∈ 𝐼 , 𝑥 ∈ 𝑋 ⧵ 𝐴𝑖 ⟺ 𝑥 ∈ ⋂(𝑋 ⧵ 𝐴𝑖 ).
𝑖∈𝐼
Proposition A.2. Let 𝑋 be a set, {𝐴𝑖 }𝑖∈𝐼 be a family of subsets of 𝑋 , and 𝐵 be a subset of
𝑋 . Then
(⋃ 𝐴𝑖 ) ∩ 𝐵 = ⋃(𝐴𝑖 ∩ 𝐵) and (⋂ 𝐴𝑖 ) ∪ 𝐵 = ⋂(𝐴𝑖 ∪ 𝐵).
𝑖∈𝐼 𝑖∈𝐼 𝑖∈𝐼 𝑖∈𝐼
Let 𝑥 ∈ 𝑋 , then
𝑥 ∈ (⋂ 𝐴𝑖 ) ∪ 𝐵 ⟺ 𝑥 ∈ ⋂ 𝐴𝑖 or 𝑥 ∈ 𝐵 ⟺ (∀𝑖 ∈ 𝐼 , 𝑥 ∈ 𝐴𝑖 ) or 𝑥 ∈ 𝐵
𝑖∈𝐼 𝑖∈𝐼
⟺ ∀𝑖 ∈ 𝐼 , 𝑥 ∈ 𝐴𝑖 ∪ 𝐵 ⟺ 𝑥 ∈ ⋂(𝐴𝑖 ∪ 𝐵).
𝑖∈𝐼
(𝑥, 𝑦) ∈ 𝐴 × (𝐶 ∩ 𝐷) ⟺ 𝑥 ∈ 𝐴, 𝑦 ∈ 𝐶 ∩ 𝐷 ⟺ 𝑥 ∈ 𝐴, 𝑦 ∈ 𝐶, 𝑦 ∈ 𝐷
⟺ (𝑥, 𝑦) ∈ 𝐴 × 𝐶, (𝑥, 𝑦) ∈ 𝐴 × 𝐷
⟺ (𝑥, 𝑦) ∈ (𝐴 × 𝐶) ∩ (𝐴 × 𝐷).
(𝑥, 𝑦) ∈ 𝐴 × (𝐶 ∪ 𝐷) ⟺ 𝑥 ∈ 𝐴, 𝑦 ∈ 𝐶 ∪ 𝐷 ⟺ 𝑥 ∈ 𝐴, 𝑦 ∈ 𝐶 or 𝑦 ∈ 𝐷
⟺ (𝑥, 𝑦) ∈ 𝐴 × 𝐶 or (𝑥, 𝑦) ∈ 𝐴 × 𝐷
⟺ (𝑥, 𝑦) ∈ (𝐴 × 𝐶) ∪ (𝐴 × 𝐷).
(𝑥, 𝑦) ∈ 𝐴 × (𝐶 ⧵ 𝐷) ⟺ 𝑥 ∈ 𝐴, 𝑦 ∈ 𝐶 ⧵ 𝐷 ⟺ 𝑥 ∈ 𝐴, 𝑦 ∈ 𝐶, 𝑦 ∉ 𝐷
⟺ (𝑥, 𝑦) ∈ 𝐴 × 𝐶, (𝑥, 𝑦) ∉ 𝐴 × 𝐷
⟺ (𝑥, 𝑦) ∈ (𝐴 × 𝐶) ⧵ (𝐴 × 𝐷).
(𝑥, 𝑦) ∈ (𝐴 ∩ 𝐵) × (𝐶 ∩ 𝐷) ⟺ 𝑥 ∈ 𝐴 ∩ 𝐵, 𝑦 ∈ 𝐶 ∩ 𝐷
⟺ 𝑥 ∈ 𝐴, 𝑥 ∈ 𝐵, 𝑦 ∈ 𝐶, 𝑦 ∈ 𝐷
⟺ (𝑥, 𝑦) ∈ 𝐴 × 𝐶, (𝑥, 𝑦) ∈ 𝐵 × 𝐷
⟺ (𝑥, 𝑦) ∈ (𝐴 × 𝐶) ∩ (𝐵 × 𝐷).
Consider now a family {𝑋𝑖 }𝑖∈𝐼 of sets indexed by 𝐼 , we then define the (Cartesian)
product ∏𝑖∈𝐼 𝑋𝑖 of the sets {𝑋𝑖 }𝑖∈𝐼 by
∏ 𝑋𝑖 = {𝑥 ∶ 𝐼 → ⋃ 𝑋𝑖 | ∀𝑖 ∈ 𝐼 , 𝑥(𝑖) ∈ 𝑋𝑖 } .
𝑖∈𝐼 𝑖∈𝐼
We often denote 𝑥 ∈ ∏𝑖∈𝐼 𝑋𝑖 by 𝑥 = (𝑥𝑖 )𝑖∈𝐼 . Moreover, for each 𝑗 ∈ 𝐼 we define the
canonical projection pr𝑗 ∶ ∏𝑖∈𝐼 𝑋𝑖 → 𝑋𝑗 given by pr𝑗 (𝑥) = 𝑥(𝑗).
Lemma A.4 (Zorn’s Lemma). If 𝑃 is a partially ordered set with the property that every
chain has an upper bound, then 𝑃 contains a maximal element.
Proposition A.5. There is no surjective map from a non-empty set 𝑋 onto 𝔓(𝑋 )
Proof. We only show that the product of two countable sets is countable, the general
case being deduced from this by induction. Let 𝑋 and 𝑌 be countable sets. We know
that there is an injection 𝑋 → N and an injection 𝑌 → N, thus there exists an injection
𝜙 ∶ 𝑋 × 𝑌 → N × N. But N × N is countable, since the map 𝑓 ∶ N × N → N given by
𝑓 (𝑛, 𝑚) = 2𝑛 3𝑚 is injective. Therefore the map 𝑓 ∘ 𝜙 ∶ 𝑋 × 𝑌 → N is injective, thus 𝑋 × 𝑌
is countable.
Proof. Let {𝑋𝑛 }𝑛∈N be a countable family of countable sets and consider their union 𝑋 =
⋃𝑛∈N 𝑋𝑛 . Define, for any 𝑛 ∈ N, the set ℐ𝑛 of all injections from 𝑋𝑛 to N. We remark
that ℐ𝑛 ≠ ∅ since 𝑋𝑛 is countable.
Thanks to the Axiom of Choice (actually we need only a weaker version of it, called
the “Axiom of Countable Choice”) there is a sequence (𝑖𝑛 )𝑛∈N such that 𝑖𝑛 ∈ ℐ𝑛 for all
𝑛 ∈ N. Define the map
𝜙 ∶ 𝑋 → N×N
𝑥 ↦ (𝑛, 𝑖𝑛 (𝑥))
where 𝑛 is the smallest natural number verifying 𝑥 ∈ 𝑋𝑛 . By construction 𝜙 is an injec-
tion, since every 𝑖𝑛 is an injection. But there is an injection 𝑓 ∶ N × N → N since N × N
is countable. Therefore 𝑓 ∘ 𝜙 ∶ 𝑋 → N is a injective map and thus 𝑋 is countable.
A.2 Maps 83
A.2 Maps
A.2.1 Injectivity and surjectivity
Let 𝑋 and 𝑌 be sets and 𝑓 ∶ 𝑋 → 𝑌 a map.
The map 𝑓 is said to be injective (or one-to-one) provided for any 𝑥, 𝑥 ′ ∈ 𝑋 , if 𝑓 (𝑥) =
𝑓 (𝑥 ′ )
then 𝑥 = 𝑥 ′ .
The map 𝑓 is is said to be surjective (or onto) if for any 𝑦 ∈ 𝑌 there is 𝑥 ∈ 𝑋 such that
𝑦 = 𝑓 (𝑥).
The map 𝑓 is is said to be bijective (or to be a one-to-one correspondence) if it is both
injective and surjective.
Proposition A.8. Let 𝑓 ∶ 𝑋 → 𝑌 and 𝑔 ∶ 𝑌 → 𝑍 be two maps between sets, then there
holds:
(i) If 𝑓 and 𝑔 are injective, then 𝑔 ∘ 𝑓 is injective.
(ii) If 𝑓 and 𝑔 are surjective, then 𝑔 ∘ 𝑓 is surjective.
(iii) If 𝑔 ∘ 𝑓 is injective, then 𝑓 is injective.
(iv) If 𝑔 ∘ 𝑓 is surjective, then 𝑔 is surjective.
Proof of Proposition A.8. (i) Let 𝑥, 𝑥 ′ ∈ 𝑋 such that 𝑔 ∘ 𝑓 (𝑥) = 𝑔 ∘ 𝑓 (𝑥 ′ ). This means that
𝑔(𝑓 (𝑥)) = 𝑔(𝑓 (𝑥 ′ )) and hence, since 𝑔 is injective, one obtains 𝑓 (𝑥) = 𝑓 (𝑥 ′ ). Since 𝑓 is
injective one deduces that 𝑥 = 𝑥 ′ .
84 Chapter A. Reminders: Sets and maps
(ii) Let 𝑧 ∈ 𝑍 . Since 𝑔 is surjective there is 𝑦 ∈ 𝑌 such that 𝑔(𝑦) = 𝑧. Since 𝑓 is surjective
there is 𝑥 ∈ 𝑋 such that 𝑦 = 𝑓 (𝑥), therefore 𝑔 ∘ 𝑓 (𝑥) = 𝑔(𝑦) = 𝑧.
(iii) Let 𝑥, 𝑥 ′ ∈ 𝑋 such that 𝑓 (𝑥) = 𝑓 (𝑥 ′ ). Applying the map 𝑔 to last equality, one
obtains that 𝑔 ∘ 𝑓 (𝑥) = 𝑔 ∘ 𝑓 (𝑥 ′ ) and hence, since 𝑔 ∘ 𝑓 is injective, one deduces 𝑥 = 𝑥 ′ .
(iv) Let 𝑧 ∈ 𝑍 . Since 𝑔 ∘ 𝑓 is surjective there is 𝑥 ∈ 𝑋 such that 𝑔 ∘ 𝑓 (𝑥) = 𝑧. Therefore
𝑦 = 𝑓 (𝑥) ∈ 𝑌 and 𝑔(𝑦) = 𝑧.
Proposition A.9. For any subset 𝐴 ⊆ 𝑋 one has 𝐴 ⊆ 𝑓 −1 (𝑓 (𝐴)). Moreover, the map 𝑓 is
injective if and only if 𝐴 = 𝑓 −1 (𝑓 (𝐴)) for any subset 𝐴 ⊆ 𝑋 .
Proof. Let 𝐴 ⊆ 𝑋 be a subset and 𝑥 ∈ 𝐴, then 𝑓 (𝑥) ∈ 𝑓 (𝐴) and hence 𝑥 ∈ 𝑓 −1 (𝑓 (𝐴)).
This proves 𝐴 ⊆ 𝑓 −1 (𝑓 (𝐴)).
We give an example in which the previous inclusion is proper: Consider the map
𝑓 ∶ {0, 1} → {0, 1} defined by 0 ↦ 0 and 1 ↦ 0. Then {0} ⊊ {0, 1} = 𝑓 −1 (𝑓 ({0})).
We now prove the second assertion. Suppose 𝑓 injective and let 𝐴 ⊆ 𝑋 . Let 𝑥 ∈
−1
𝑓 (𝑓 (𝐴)), then 𝑓 (𝑥) ∈ 𝑓 (𝐴). Therefore there is 𝑎 ∈ 𝐴 such that 𝑓 (𝑥) = 𝑓 (𝑎), which
implies that 𝑥 = 𝑎 ∈ 𝐴, whence 𝑓 −1 (𝑓 (𝐴)) ⊆ 𝐴. From the first assertion one deduces
𝐴 = 𝑓 −1 (𝑓 (𝐴)).
Conversely, assume that 𝐴 = 𝑓 −1 (𝑓 (𝐴)) for any subset 𝐴 ⊆ 𝑋 . Let 𝑥, 𝑦 ∈ 𝑋 such that
𝑓 (𝑥) = 𝑓 (𝑦). Then, by hypothesis, {𝑥} = 𝑓 −1 (𝑓 ({𝑥})) = 𝑓 −1 ({𝑓 (𝑥)}) = 𝑓 −1 ({𝑓 (𝑦)}) =
𝑓 −1 (𝑓 ({𝑦})) = {𝑦}, therefore 𝑥 = 𝑦 and thus 𝑓 is injective.
Proposition A.10. For any subset 𝐵 ⊆ 𝑌 one has 𝑓 (𝑓 −1 (𝐵)) ⊆ 𝐵. Moreover, the map 𝑓 is
surjective if and only if 𝑓 (𝑓 −1 (𝐵)) = 𝐵 for any subset 𝐵 ⊆ 𝑌 .
Proof. Let 𝐵 ⊆ 𝑌 and 𝑦 ∈ 𝑓 (𝑓 −1 (𝐵)), then there is 𝑥 ∈ 𝑓 −1 (𝐵) such that 𝑦 = 𝑓 (𝑥). Since
𝑓 (𝑥) ∈ 𝐵 we conclude that 𝑦 ∈ 𝐵. This proves 𝑓 (𝑓 −1 (𝐵)).
We give an example in which the inclusion is proper: Consider the map 𝑓 ∶ {0, 1} →
{0, 1} defined by 0 ↦ 0 and 1 ↦ 0. Then 𝑓 (𝑓 −1 ({0, 1})) = {0} ⊊ {0, 1}.
We now prove the second assertion. Suppose 𝑓 surjective and let 𝐵 ⊆ 𝑌 . Let 𝑦 ∈ 𝐵,
then there is 𝑥 ∈ 𝑋 such that 𝑓 (𝑥) = 𝑦, that is 𝑥 ∈ 𝑓 −1 (𝐵), whence 𝑦 ∈ 𝑓 (𝑓 −1 (𝐵)). From
the first assertion one deduces that 𝑓 (𝑓 −1 (𝐵)) = 𝐵.
Conversely, assume that 𝑓 (𝑓 −1 (𝐵)) = 𝐵 for any subset 𝐵 ⊆ 𝑌 . Let 𝑦 ∈ 𝑌 , then
𝑓 (𝑓 −1 ({𝑦})) = {𝑦} is non-empty, therefore there is 𝑥 ∈ 𝑋 such that 𝑓 (𝑥) = 𝑦, and thus 𝑓
is surjective.
Proof. Let 𝐵 ⊆ 𝑌 and 𝑥 ∈ 𝑋 . Then 𝑥 ∉ 𝑓 −1 (𝐵) if and only if 𝑓 (𝑥) ∉ 𝐵, which means
𝑥 ∈ 𝑓 −1 (𝑌 ⧵ 𝐵).
Consider a non-empty family {𝐴𝑖 }𝑖∈𝐼 of subsets of 𝑋 . We obtain the following prop-
erties concerning the image of unions and intersections.
Proposition A.12. There holds
𝑓 (⋃ 𝐴𝑖 ) = ⋃ 𝑓 (𝐴𝑖 ) and 𝑓 (⋂ 𝐴𝑖 ) ⊆ ⋂ 𝑓 (𝐴𝑖 ).
𝑖∈𝐼 𝑖∈𝐼 𝑖∈𝐼 𝑖∈𝐼
Moreover, if 𝑓 is injective then the above inclusion is an equality.
Proof. Let 𝑦 ∈ 𝑌 , then
𝑦 ∈ 𝑓 (⋃ 𝐴𝑖 ) ⟺ ∃𝑥 ∈ ⋃ 𝐴𝑖 , 𝑦 = 𝑓 (𝑥) ⟺ ∃𝑖 ∈ 𝐼 , ∃𝑥 ∈ 𝐴𝑖 , 𝑦 = 𝑓 (𝑥)
𝑖∈𝐼 𝑖∈𝐼
⟺ ∃𝑖 ∈ 𝐼 , 𝑦 ∈ 𝑓 (𝐴𝑖 ) ⟺ 𝑦 ∈ ⋃ 𝑓 (𝐴𝑖 ).
𝑖∈𝐼
This proves the first assertion.
Let 𝑦 ∈ 𝑓 (⋂𝑖∈𝐼 𝐴𝑖 ), then there is 𝑥 ∈ ⋂𝑖∈𝐼 𝐴𝑖 such that 𝑦 = 𝑓 (𝑥), which means that
𝑥 ∈ 𝐴𝑖 for any 𝑖 ∈ 𝐼 and 𝑦 = 𝑓 (𝑥). This implies that for any 𝑖 ∈ 𝐼 there is a 𝑥 ∈ 𝐴𝑖 such
that 𝑦 = 𝑓 (𝑥), thus 𝑦 ∈ 𝑓 (𝐴𝑖 ) for any 𝑖 ∈ 𝐼 . Therefore 𝑦 ∈ ⋂𝑖∈𝐼 𝑓 (𝐴𝑖 ). This proves the
second assertion.
Let us give an example of a proper inclusion: Take 𝑓 ∶ 𝑥 ↦ 𝑥 2 from R to R, 𝐴1 = {1}
and 𝐴2 = {−1}, then 𝑓 (𝐴1 ∩ 𝐴2 ) = ∅ and 𝑓 (𝐴1 ) ∩ 𝑓 (𝐴2 ) = {1}.
Finally, we assume that 𝑓 is injective and we prove the inclusion ⋂𝑖∈𝐼 𝑓 (𝐴𝑖 ) ⊆ 𝑓 (⋂𝑖∈𝐼 𝐴𝑖 ).
Let 𝑦 ∈ ⋂𝑖∈𝐼 𝑓 (𝐴𝑖 ), then 𝑦 ∈ 𝑓 (𝐴𝑖 ) for any 𝑖 ∈ 𝐼 . This means that, for any 𝑖 ∈ 𝐼 , there is
𝑥𝑖 ∈ 𝐴𝑖 such that 𝑦 = 𝑓 (𝑥𝑖 ). Since 𝑓 is injective we deduce that all the 𝑥𝑖 are the same,
that is, there is 𝑥 ∈ 𝑋 such that 𝑥 = 𝑥𝑖 ∈ 𝐴𝑖 for any 𝑖 ∈ 𝐼 . Therefore we have that
𝑥 ∈ ⋂𝑖∈𝐼 𝐴𝑖 and 𝑦 = 𝑓 (𝑥), hence 𝑦 ∈ 𝑓 (⋂𝑖∈𝐼 𝐴𝑖 ).
Finally, we consider a non-empty family {𝐵𝑗 }𝑗∈𝐽 of subsets of 𝑌 and obtain the fol-
lowing properties concerning the preimage of unions and intersections.
Proposition A.13. There holds
𝑓 −1 (⋃ 𝐵𝑗 ) = ⋃ 𝑓 −1 (𝐵𝑗 ) and 𝑓 −1 (⋂ 𝐵𝑗 ) = ⋂ 𝑓 −1 (𝐵𝑗 ).
𝑗∈𝐽 𝑗∈𝐽 𝑗∈𝐽 𝑗∈𝐽
⟺ ∃𝑗 ∈ 𝐽 , 𝑥 ∈ 𝑓 −1 (𝐵𝑗 ) ⟺ 𝑥 ∈ ⋃ 𝑓 −1 (𝐵𝑗 ).
𝑗∈𝐽
⟺ ∀𝑗 ∈ 𝐽 , 𝑥 ∈ 𝑓 −1 (𝐵𝑗 ) ⟺ 𝑥 ∈ ⋂ 𝑓 −1 (𝐵𝑗 ).
𝑗∈𝐽