Chapter 1

Metric Spaces

1.1 Definition and examples

In your first calculus class, you encountered the notion of a limit of a function at
a point. Recall the definition:

Let f ∶ R → R and let x0 , y0 ∈ R. We say that

lim f (x) = y0
x→x0

if

∀" > 0 ∃ > 0 such that ∀x ∶ 0 < �x − x0 � < � f (x) − y0 � < ".

Let’s examine what is needed for this definition to make sense and how much we
can generalize it. The domain and the range are both R. Can we replace them
by other sets? Clearly, we could take them to be “nice enough” subsets of R. In
addition, it is assumed that we can subtract elements in the domain (x − x0 ) and
elements in the range ( f (x) − y0 ). Vectors, for example, can be subtracted. Can
we replace the domain and the range by any vector space? The answer is positive,
but we need something to replace the absolute value. In vector spaces, the role of
the absolute value (the distance from zero) is played by a norm.
Thus, we have the following generalization:
2 Chapter 1

Let (V, � ⋅ �V ) and (W, � ⋅ �W ) be normed spaces. Let f ∶ V → W and

let x0 ∈ V and y0 ∈ W. We say that

lim f (x) = y0
x→x0

if

∀" > 0 ∃ > 0 such that ∀x ∶ 0 < �x−x0 �V < � f (x)−y0 �W < ".

Expressions such that �x − x0 �V and � f (x) − y0 �W can be interpreted as distances

between elements in two respective vector spaces. In fact, we could reformulate
the above as follows:

Let (V, � ⋅ �V ) and (W, � ⋅ �W ) be normed spaces. Let f ∶ V → W and

let x0 ∈ V and y0 ∈ W. We say that

lim f (x) = y0
x→x0

if ∀" > 0 there exists a > 0, such that the distance between f (x) and
y0 is less than " whenever the distance between x and x0 is greater
than zero and less than .

In this reformulation, we observe that we don’t really need the domain and the
range of f to be vector spaces. All we need is a notion of distance. Sets endowed
with a distance are called metric spaces, and they are the subject of this chapter.

1.1.1 Basic definitions

Definition 1.1 (Metric space) A metric space (*9)/ "(9/) is a pair (X, d),
where X is a non-empty set, and d ∶ X × X → R+ is a distance function, or a
metric (%8*9)/). The metric d assigns to every pair of points x, y ∈ X a non-
negative number, d(x, y), which we call the distance between x and y. A metric d
must satisfy three defining properties:

1. Symmetry: for every x, y ∈ X,

d(x, y) = d(y, x).

 Metric Spaces 3

2. Positivity: for every x, y ∈ X,

d(x, y) ≥ 0,
with equality if and only if x = y.
3. The triangle inequality: for every x, y, z ∈ X,
d(x, z) ≤ d(x, y) + d(y, z).

y
x
z

It follows, by induction, from the triangle inequality that for every finite set of
points (xi )ni=1 in a metric space,
n−1
d(x1 , xn ) ≤ � d(x j , x j+1 ).
j=1

x2 xn−1
x3 xn−2
x1

xn

Likewise, it follows from the triangle inequality that

d(x, y) ≥ d(x, z) − d(y, z).
Since this expression must hold if the switch the roles of x and y, we obtain the
reverse triangle inequality,
d(x, y) ≥ �d(x, z) − d(y, z)�. (1.1)

Comment: Metric spaces were first introduced by Maurice Frechet in 1906, uni-
fying work on function spaces by Cantor, Volterra, Arzelà, Hadamard, Ascoli,
and others. Topological spaces (.**#&-&5&) .*"(9/), of which metric spaces are a
special instance, were first introduced by Felix Haussdorf in 1914; their current
formulation is a slight generalization from 1922 by Kazimierz Kuratowski.
4 Chapter 1

Examples:

(a) Both R and C are metric spaces when endowed with the distance function
d(x, y) = �x−y�. We will often refer to “the metric space R”, without explicit
mention of the metric.
(b) Let F ∶ R → R be any strictly monotonic function (say increasing). Then, R
endowed with the distance function

d(x, y) = �F(x) − F(y)�

is a metric space. Symmetry and positivity are immediate. It only remains

to verify the triangle inequality. For x, y, z ∈ R,

d(x, z) = �F(x) − F(z)�

≤ �F(x) − F(y)� + �F(y) − F(z)�
= d(x, y) + d(y, x).

Note that the strict monotonicity of F is only needed to ensure the positivity
of d.
(c) Rn endowed with the Euclidean distance,

n 1�2
d(x, y) = ��(xi − yi ) � 2
i=1

is a metric space. Unless otherwise specified, Rn will always be assumed to

be endowed with the Euclidean metric.
(d) The surface of earth is a metric space, with d being the (shortest) distance
along great circles.
(e) Let (G, V) be a non-directed graph. For u, v ∈ V, a path (%-*2/) from u to
v is a finite sequence of vertices,

= (x0 , x1 , x2 , . . . , xn1 , xn )

such that (x j , x j+1 ) ∈ E for each j = 0, . . . , n − 1, x0 = u and xn = v. The

length `( ) of that path is the number of edges, n. Define then,

d(u, v) = min{`( ) ∶ is a path from u to v}.

 Metric Spaces 5

We claim that d is a metric on V. Indeed, it is positive because d(u, u) = 0

and d(u, v) ≥ 1 for u ≠ v. It is symmetric, because every path from u to v is
a path from v to u, namely, we have an equality between sets,

{`( ) ∶ is a path from u to v} = {`( ) ∶ is a path from v to u}.

It remains to show that d satisfies the triangle inequality. If 1 and 2 are two
paths, such that 1 ends at the point where 2 starts, then we can concatenate
(9:9:-) them; denote the compound path by 2 ∗ 1 . Then,

`( 2 ∗ 1) = `( 1 ) + `( 2 ).

Let u, v, w ∈ V. Let 1 be a shortest path from u to v and let 2 be a shortest

path from v to w. Then 2 ∗ 1 is a (not necessarily shortest!) path from u to
w, hence

d(u, w) ≤ `( 2 ∗ 1) = `( 1 ) + `( 2 ) = d(u, v) + d(v, w).

(f) Any non-empty set X can be endowed with the discrete metric (%8*9)/%
%$*$"%),
�
�
�0 x = y
d(x, y) = �
�1 x ≠ y.
�
�
This is a particularly boring metric, since it provides no information about
the structure of the space. We will often use it, however, as a clarifying
example (or as a counter-example). Note that the discrete metric coincide
with the graph metric, if we endow X with the structure of a complete graph
(every two points are connected by an edge).

TA material 1.1 Let X be a non-empty set. A function d ∶ X × X → [0, ∞) is

called a pseudo-metric if

1. For every x, y ∈ X, d(x, y) = d(y, x).

2. For every x ∈ X, d(x, x) = 0.
3. For every x, y, z ∈ X, d(x, z) ≤ d(x, y) + d(y, z).

(a) Show that d induces an equivalence relation on X,

x∼y ⇐⇒ d(x, y) = 0.
6 Chapter 1

(b) Show that d̃ ∶ (X� ∼) × (X� ∼) → [0, ∞) given by

d̃([x], [y]) = d(x, y)

is well-defined and is a metric on X� ∼.

. Exercise 1.1 The goal of the following exercise is to rehearse elementary set-
theoretic relations. Let X be a set and let A, B, C ⊂ X. Let f ∶ X → Y and let
E, F ⊂ Y. Prove the following:

(a) (A ∪ B) ∩ C = (A ∪ C) ∪ (B ∪ C).
(b) (A ∩ B)c = Ac ∪ Bc .
(c) (A ∪ B)c = Ac ∩ Bc .
(d) f −1 (E ∩ F) = f −1 (E) ∩ f −1 (F).
(e) f −1 (E ∪ F) = f −1 (E) ∪ f −1 (F).
(f) f −1 (E c ) = ( f −1 (E))c .
(g) f (A ∪ B) = f (A) ∪ f (B).
(h) f (A ∩ B) ⊂ f (A) ∩ f (B) (note that this is not an equality).
(i) A ⊂ f −1 ( f (A)).
(j) E ⊃ f ( f −1 (E)).
(k) Is there any relation between f (A) and f (Ac )?

. Exercise 1.2 Prove that the set of infinite sequences of real numbers, endowed
with the function
1 �xn − yn �
∞
d((xn ), (yn )) = �
n=1 2 1 + �xn − yn �
n

is a metric space.

. Exercise 1.3 For each of the following pairs (X.d), determine whether it is a
metric space:

(a) X is the set of Riemann-integrable functions on [0, 1] and

d( f, g) = �
1
� f (x) − g(x)� dx.
0
 Metric Spaces 7

(b) X is the set of continuous functions on [0, 1] and

d( f, g) = �
1
� f (x) − g(x)� dx.
0

(c) X is the set of continuously-di↵erentiable functions on [0, 1] and

d( f, g) = max � f ′ (x) − g′ (x)�.
0≤x≤1

(d) X is the set of continuously-di↵erentiable functions on [0, 1] and

d( f, g) = max � f (x) − g(x)� + max � f ′ (x) − g′ (x)�.
0≤x≤1 0≤x≤1

Given a metric space (X, d) and a non-empty subset Y ⊂ X, there is a canonical

metric defined on Y:

Proposition 1.2 Let (X, d) be an arbitrary metric space, and let Y ⊂ X. Then the
set Y with the function d restricted to Y × Y is a metric space.

We will call d�Y×Y the metric on Y induced by the metric on X (;*9:&/ %8*9)/).
A subspace of a metric space always refers to a subset endowed with the induced
metric.

Proof : This is really immediate. Since y1 , y2 ∈ Y are also points is X,

d�Y×Y (y1 , y2 ) = d(y1 , y2 ) ≥ 0
with equality if and only if y1 = y2 . By the same argument, d�Y×Y is symmetric and
satisfies the triangle inequality. n

Example: Let X = R2 endowed with the Euclidean metric, and let Y = S 1 be the
unit circle. Then, the induced metric on Y is depicted below:
8 Chapter 1

Note that this metric is di↵erent from the “natural” metric on S 1 which determined
by the shortest distance along the arc. ▲▲▲

Definition 1.3 Let (X, d) be a matric space. A sphere (%9*52) of radius r > 0
centered at a ∈ X is the set of points,

Sr (a) = {x ∈ X ∶ d(x, a) = r} .

(Note that this set may be empty.) An open ball ((&;5 9&$,) (or for short, a ball)
of radius r centered at a is the set

Br (a) = {x ∈ X ∶ d(x, a) < r} .

A closed ball (9&#2 9&$,) of radius r centered at a is defined by

B̂r (a) = {x ∈ X ∶ d(x, a) ≤ r} .

Comment: In Rn these definitions coincide with our intuitive pictures of spheres

and balls. Note, however, that for a space endowed with the discrete metric,
�
�
�{a} r ≤ 1
Br (a) = �
�
� r > 1.
�X

We now show number of elementary properties of balls in metric spaces:

Lemma 1.4 For every x, y ∈ X and every r > 0,

x ∈ Br (y) if and only if y ∈ Br (x).

x y
 Metric Spaces 9

Proof : This is immediate as

x ∈ Br (y) ⇐⇒ d(x, y) < r ⇐⇒ y ∈ Br (x).
n

Lemma 1.5 Let x ∈ X and suppose that y ∈ Br (x) for every r > 0. Then, y = x.

Proof : The assumption is that

d(y, x) < r for all r > 0.
Since the metric is non-negative, it follows that d(y, x) = 0, and by the positivity
of the metric, y = x. n

Lemma 1.6 Let x ∈ X and r > 0. Let y ∈ Br (x) and let s > 0 satisfiy
d(x, y) + s < r.

Then,
B s (y) ⊂ Br (x).

r
s
y x

Proof : Let z ∈ B s (y). Then, by the triangle inequality

d(x, z) ≤ d(x, y) + d(y, z) < d(x, y) + s < r,
hence z ∈ Br (x). Since this holds for every z ∈ B s (y), we conclude that B s (y) ⊂
Br (x). n
10 Chapter 1

Proposition 1.7 Let (X, dX ) be a metric space and let Y ⊂ X be a non-empty

subspace. Denote the induced metric on Y by dY . Likewise, denote by BrX and BrY
the open balls with respect to the metrics dX and dY . Then, for every y ∈ Y,

BrY (y) = Y ∩ BrX (y).

r
Y
y

Proof : We need to show that two sets are equal. Just follows the definitions,

BrY (y) = {z ∈ Y ∶ dY (z, y) < r}

= {z ∈ Y ∶ dX (z, y) < r}
= {z ∈ X ∶ z ∈ Y, dX (z, y) < r}
= {z ∈ X ∶ z ∈ Y} ∩ {z ∈ X ∶ dX (z, y) < r}
= Y ∩ BrX (y).

n
Recall the definition of a bounded set in R from your first calculus course. We
have the following generalization for metric spaces:

Definition 1.8 Let (X, d) be a metric space. A set A ⊂ X is called bounded

(.&2() if it is enclosed in some open ball. That is,

∃a ∈ X and ∃r > 0 such that A ⊂ Br (a).

(Note that a does not need to be an element of A; in particular, the empty set is
always bounded.)
 Metric Spaces 11

The following lemma shows that the center of the bounding ball is immaterial to
the definition of boundedness:

Lemma 1.9 If A ⊂ X is bounded, then

∀b ∈ X ∃⇢ > 0 such that A ⊂ B⇢ (b).

Proof : Suppose that A is bounded. That is,

A ⊂ Br (a)

for some a ∈ X and r > 0. Let b ∈ X. Then, for every x ∈ A,

d(x, b) ≤ d(x, a) + d(a, b) < r + d(a, b).

Since this holds for every x ∈ A,

A ⊂ B⇢ (b)

for ⇢ = r + d(a, b). n

Lemma 1.10 Every open ball Br (a) contains a closed ball centered at a.

Proof : By definition,
B̂r�2 (a) ⊂ Br (a).
n

1.1.2 Normed spaces

Recall the following definition from your first-year linear algebra class:

Definition 1.11 Let X be a vector space over R (we might consider vector spaces
over C as well). A norm (%/9&1) on X is a function � ⋅ � ∶ X → R satisfying

1. Positivity: �x� ≥ 0 with equality if and only if x = 0.

12 Chapter 1

2. Homogeneity: for every x ∈ X and a ∈ R, �ax� = �a��x�.

3. The triangle inequality: for every x, y ∈ X,

�x + y� ≤ �x� + �y�.

A vector space X endowed with a norm is called a normed space (*/9&1 "(9/).

Normed space are an important instance of metric spaces, as the following propo-
sition asserts:

Proposition 1.12 Let (X, � ⋅ �) be a normed space. Then,

d(x, y) = �x − y�

defines a metric on X. That is, every normed space is automatically a metric space
with a canonical metric.

Proof : Immediate. n
Note, however, that normed spaces carry properties that are not pertinent to all
metric spaces. Metric spaces need not be vector spaces. In addition, normed
spaces, as metric spaces, are translational invariant,

d(x + z, y + z) = d(x, y),

and homogeneous,
d( x, y) = � � d(x, y),
where is a scalar.

. Exercise 1.4 Let (X, d) be a metric space. Suppose that X is also a vector
space over R and that

d(x + z, y + z) = d(x, y), and d( x, y) = � � d(x, y).

Show that there exists a norm on X such that d is the corresponding metric.

Examples:
 Metric Spaces 13

(a) Both R and C are normed space, with �x� = �x�, and the resulting metric is
the same as the one above.
(b) For X = Rn and 1 ≤ p < ∞, the function � ⋅ � p ∶ X → R defined by
n 1�p
�x� p = �� �xi � p �
i=1

is a norm, hence defines a metric. We denotes this family of metric spaces

by `np . It is easy to see that this function satisfies the positivity and the homo-
geneity conditions. Let’s check that it satisfies also the triangle inequality
for the particular cases of p = 1 and p = 2. For p = 1, this follows from the
triangle inequality in R,
n n
�x + y�1 = � �xi + yi � ≤ �(�xi � + �yi �) = �x�1 + �y�1 .
i=1 i=1

For p = 2, this follows from the Cauchy-Schwarz inequality,

n
�x + y�22 = � �xi + yi �2
i=1
n n
= �(�xi �2 + �yi �2 ) + 2 � �xi yi �
i=1 i=1
n n 1�2 n 1�2
≤ �(�xi � + �yi � ) + 2 �� �xi � �
2 2 2
�� �yi � �
2
i=1 i=1 i=1
= �x�22 + �y�22
+ 2�x�2 �y�2
= (�x�2 + �y�2 )2 .

(c) The function � ⋅ �∞ ∶ Rn → R defined by

�x�∞ = max �xi �.
1≤i≤n

is a norm. The corresponding metric space is denoted by `∞

n .

(d) The set of infinite sequences (xi ) for which

∞ 1�p
�x� p = �� �xi � � p
i=1

is finite is an example of an infinite-dimensional normed space. We denote

this space by ` p .
 14 Chapter 1

(e) Consider C([a, b]), the set of continuous real-valued functions on [a, b]. It
is a vector space over R, where addition and multiplication by a scalar are
defined pointwise, namely, for f, g ∈ C([a, b]) and ↵, ∈ R,

(↵ f + g)(x) = ↵ f (x) + g(x).

The function � ⋅ �∞ ∶ C([a, b]) → R defined by

� f �∞ = max � f (x)�
a≤x≤b

is a norm, therefore C([a, b]) is a metric space if endowed with the distance
function
d( f, g) = max � f (x) − g(x)�.
a≤x≤b

. Exercise 1.5 Draw the unit balls B1 (0) in R2 for the metrics induced by the
1-, 2-, and ∞-norms.

. Exercise 1.6 Show that the functions � ⋅ � p ∶ Rn → R, defined by

n 1�p
�x� p = �� �xi � � p
1< p<∞
i=1

satisfy the definition of a norm. (Hint: the triangle inequality is proved through
Young’s inequality, followed by Hölder’s inequality, followed by Minkowski’s
inequality. Prove all three inequalities.)

. Exercise 1.7 Let x = (x1 , x2 , . . . , xn ). Prove that the function

'(p) = �x� p

defined for 1 < p < ∞ is monotonically decreasing, and that '(p) → �x�∞ as
p → ∞.
—2h(2019)—

1.2 Point set topology in metric spaces

Throughout this section we will assume a metric space (X, d). We will denote
sequences x1 , x2 , . . . in X by (xn )∞
n=1 , or in short by xn , when the identity of the
running index is evident.
 Metric Spaces 15

1.2.1 Limits and convergence

Definition 1.13 (Limit of a sequence) A sequence xn in a metric space (X, d) is
called convergent (;21,;/) if there exists a point x ∈ X such that

lim d(xn , x) = 0.
n→∞

That is,
∀" > 0 ∃N ∈ N such that ∀n > N d(xn , x) < ".
The point x is called a limit (-&"#) of the sequence xn , and we write

lim xn = x or xn → x.
n→∞

We must first make sure that in a convergent sequence, the notion of the limit is
well-defined:

Proposition 1.14 If a sequence converges then its limit is unique.

Proof : Suppose that a sequence xn converges to x and to y. Then, for every " > 0
there exists a (sufficiently large) n such that

d(xn , x) < " and d(xn , y) < ",

from which we conclude that for every " > 0,

d(x, y) ≤ d(x, xn ) + d(xn , y) < 2",

i.e., d(x, y) = 0, namely x = y. n

The following is an equivalent characterization of a convergent sequence:

Proposition 1.15 A sequence xn in a metric space (X, d) converges to x ∈ X if

and only if for every r > 0 the sequence is eventually in Br (x). That is, there exists
an N, which depends on r, such that for every n > N,

xn ∈ Br (x).
16 Chapter 1

Br (x)

Proof : This follows immediately from the definition. n

Examples:

(a) In every metric space, a constant sequence xn = x0 converges to x0 .

(b) Let (X, � ⋅ �) be a normed space. Then,

lim xn = x if and only if lim �xn − x� = 0.

n→∞ n→∞

(c) In particular, let X = C([a, b]) endowed with the maximum norm. Let
fn , f ∈ C([a, b]). Then,

lim fn = f if and only if lim max � fn (x) − f (x)� = 0.

n→∞ n→∞ a≤x≤b

That is, convergence in C([a, b]) coincides with uniform convergence

(%&&: %$*/" ;&21,;%).

Proposition 1.16 A metric d on X is continuous in the following sense: if xn → x

and yn → y, then
lim d(xn , yn ) = d(x, y).
n→∞

Proof : By the triangle inequality,

d(x, y) − d(x, xn ) − d(yn , y) ≤ d(xn , yn ) ≤ d(xn , x) + d(x, y) + d(y, yn ).

 Metric Spaces 17

xn y
x
yn

Letting n → ∞ and using the definition of the limit, both left- and right-hand sides
tend to d(x, y). n

Corollary 1.17 In a normed space, xn → x implies that

lim �xn � = �x�.
n→∞

(Note that the converse is not true.)

Proof : Taking yn , y = 0 and applying the previous proposition,

lim �xn � = lim d(xn , yn ) = d(x, y) = �x�.

n→∞ n→∞

Proposition 1.18 Let xn be a sequence converging to x. Then, every subsequence

of xn converges to x as well.

Proof : We need to show that if

lim d(xn , x) = 0,
n→∞

then, for every increasing sequence (nk ) of indexes,

lim d(xnk , x) = 0.
k→∞

This follows from the convergence of subsequences in R (which you proved in

your first calculus course). n
The following is a very useful characterization for when sequences do not con-
verge.: a sequence in which the distance between every two elements is bounded
form zero uniformly does not converge.
18 Chapter 1

Proposition 1.19 Let xn be a sequence in a metric space X and suppose that there
exists an " > 0 such that

d(xi , x j ) ≥ " for every i ≠ j.

Then, the sequence xn does not converge.

Proof : Suppose by contradiction that xn → x. Then, there exists an N ∈ N, such

that for every n > N,
d(xn , x) < "�2.
By the triangle inequality,

d(xN+1 , xN+2 ) ≤ d(xN+1 , x) + d(x, xN+2 ) < ",

which is a contradiction. n

Proposition 1.20 Convergence in Rk with respect to the Euclidean norm is equiv-

alent to the convergence of each component: for xn = (xn1 , . . . , xnk ) and x =
(x1 , . . . , xk ),

lim xn = x if and only if lim x j = xj for j = 1, . . . , k.

n→∞ n→∞ n

Proof : Suppose that xn → x. Then, for every j = 1, . . . , k,

k 1�2
�xnj −x �≤
j
�� �xnj − x j� �
2
,
i=1

hence the left-hand side tends to zero for every j as n → ∞.

Conversely, suppose that xnj → x j for j = 1, . . . , k. Then,

k 1�2
lim �� �xnj −x � �
j2
=0
n→∞
i=1

by limit arithmetics. n
 Metric Spaces 19

. Exercise 1.8 Let (X, ⇢) be a metric space and define for every x, y ∈ X,
⇢(x, y)
d(x, y) =
1 + ⇢(x, y)
.

1. Show that d is a metric on X.

2. Show that convergence with respect to this new metric d is equivalent to
convergence with respect to the original metric ⇢. That is, prove that a
sequence xn is convergent in (X, d) if and only if it is convergent in (X, ⇢).

. Exercise 1.9 Let X = {0, 1}N (the space of infinite binary sequences), and
define for every two sequences xn , yn ∈ X,
∞
d1 ((xn ), (yn )) = � 2−n �xn − yn �
�
n=1
�
�2− min(i�xi ≠yi ) (xn ) ≠ (yn )
d2 ((xn ), (yn )) = �
�
.
�
�0 otherwise

1. Show that these are indeed metrics.

2. Show that these two metrics are equivalent.

. Exercise 1.10 Let (Xn , dn ) be an infinite sequence of metric spaces. We define

the product space
∞
X = � Xn = {(x1 , x2 , . . . , ) ∶ x1 ∈ X1 , x2 ∈ X2 , . . . }
n=1

of infinite sequence for which the i-th term is in Xn .

(a) Prove that

∞
dn (xn , yn )
d(x, y) = � 2−n
n=1 1 + dn (xn , yn )
is a metric on X.
(b) Prove that a sequence xk ∈ X converges to x ∈ X if and only if limk→∞ xnk = xn
for every k.
(c) Conclude that on the space of infinite real numbers (i.e., Xn = R for all
n) there exists a metric for which convergence coincides with element-by-
element convergence. Show, however, that there is no norm satisfying this
convergence property (hint: consider the sequence ei = (0, . . . , 0, 1, 0, . . . )).
—4h(2019)—
20 Chapter 1

1.2.2 Open and closed sets

A topological space (*#&-&5&) "(9/) is a non-empty set endowed with a structure
in which some if its subsets are called open (and the notion of openness must
satisfy some axiomatic properties). Metric spaces are an important sub-class of
topological spaces, with the notion of open sets defined below:

Definition 1.21 In a metric space (X, d), a set A ⊂ X is called open ((&;5) if
every point a ∈ A is the center of an open ball contained in A. That is,

∀a ∈ A ∃r > 0 such that Br (a) ⊂ A.

In other words, a set is open if every point in it has a “neighborhood” contained

in that set.

a
A

Comments:

(a) We can replace “open ball” by “closed ball”, since Br (a) ⊂ A implies
B̂r�2 (a) ⊂ A.
(b) By definition, X itself is always an open set.
(c) By convention, the empty set � is considered to be open as well (it satisfies
the conditions in a null sense).

One can define open sets in the following equivalent way:

Definition 1.22 Let (X, d) be a metric space and let A ⊂ X. A point a ∈ A is

called an interior point (;*/*15 %$&81) of A if there exists an r > 0 such that
Br (a) ⊂ A. A set is called open if all its points are interior points.
 Metric Spaces 21

Proposition 1.23 In every metric space (X, d),

1. The intersection of any finite number of open sets is open.
2. The union of any collection (not necessarily countable) of open sets is open.
3. An open ball is an open set.
4. A set is open if and only if it is the union of open balls.

Proof :
1. It suffices to show that the intersection of two open sets is open, since the
extension to finitely-many will follow by induction (but note that induction
does not hold for infinitely-many sets, even if there are countably-many).
Let A, B ⊂ X be open sets of X and set C = A ∩ B. If C = � then we are done,
for the empty set is open. Otherwise, let c ∈ C. Since c ∈ A and A is open,
there exist an r1 > 0, such that
Br1 (c) ⊂ A.
Likewise, since c ∈ B and B is open, there exist an r2 > 0, such that
Br2 (c) ⊂ B.
Set r = min(r1 , r2 ). Then,
Br (c) ⊂ A and Br (c) ⊂ B hence Br (c) ⊂ C,
proving that C is open.

c
A B

2. Let {A↵ ∶ ↵ ∈ I} be a collection of open sets and let

A = � A↵ .
↵∈I

Every point a ∈ A belongs to at least one open set A , therefore there exists
an r > 0 such that Br (a) ⊂ A ⊂ A.
22 Chapter 1

3. Let Br (a) be an open ball. By Lemma 1.5, for every x ∈ Br (a),

B⇢ (x) ⊂ Br (a),
for ⇢ = r − d(a, x), proving that Br (a) is an open set.

a x

4. We have already seen that every open ball is open and that any union of open
sets is open. Thus, every union of open balls in open. It remains to show that
every open set can be represented as a union of open balls. By definition, to
every point x in a open set A corresponds an open ball Br(x) (x) ⊂ A. Then,
A = � {x} ⊂ � Br(x) (x) ⊂ A,
x∈A x∈A

proving that A is a union of open balls.

n
An infinite intersection of open sets is not necessarily open. Take for example
X = R with the standard metric. For every integer n, the open ball B1�n (0) =
(−1�n, 1�n) is open, however
∞ 1 1
� �− , � = {0}
n=1 n n
is not open. There is no r > 0 for which Br (0) ⊂ {0}.

Proposition 1.24 In a discrete space every set is open.

Proof : Let A be a non-empty set in a discrete metric space (X, d). Then,
A = � B1�2 (a).
a∈A

(Alternatively, every singleton is open and every set is a union of singletons.) n

 Metric Spaces 23

Definition 1.25 A set A ⊂ X is called closed (9&#2) if its complement Ac is open.

The following proposition is a direct consequence of the Proposition 1.23, using
de Morgan’s laws,

Proposition 1.26 In a metric space (X, d),

1. Every finite union of closed sets is closed.
2. Every arbitrary intersection of closed sets is closed.

. Exercise 1.11 Show, by example, that it is not generally true that an infinite
union of closed sets is closed.

Comments:
(a) Being closed does not mean not being open. The set [0, 1) in R is neither
open nor closed.
(b) By duality, a set is open if and only if its complement is closed.
(c) In every metric space (X, d), the sets X and � are both open and closed
(one sometimes uses the term clopen to designate the property of being
both open and closed).
(d) A metric space is called connected (9*:8) if X and � are the only sets
that are both open and closed. The reason for this terminology is the fol-
lowing: suppose that a non-trivial set A is both open and closed, then X is
the disjoint union of two open sets, A and Ac , i.e., you can disconnect it
into two separated open sets. Discrete spaces are the extreme example of a
non-connected set.
(e) For every x ∈ X the singleton (0&$*(*) {x} is a closed set, since for every
y ∈ {x}c ,
Bd(x,y) (y) ⊂ {x}c .

(f) By Proposition 1.26, the union of every finite collection of points is closed
as well; it is not generally true, however, that a countable union of points is
closed.
24 Chapter 1

(g) Every closed ball B̂r (a) is closed. Indeed, if x ∈ (B̂r (a))c then Bd(x,a)−r (x) ⊂
(B̂r (a))c .

r
a x

(h) In a set endowed with the discrete metric, every subset is both open and
closed. This follows from the fact that every set is open.

Examples:
(a) Consider X = R2 with the Euclidean metric and let A = S 1 , the unit circle.
Then, A is closed because every point not in S 1 is the center of a ball not in
S 1.
(b) Let X = C([a, b]), let h1 , h2 ∈ X with h1 < h2 and let

A = { f ∈ C([a, b]) ∶ ∀x ∈ [a, b] h1 (x) < f (x) < h2 (x)}.

Let f ∈ A and set

" = min min{� f (x) − h1 (x)�, � f (x) − h2 (x)�}.

x∈[a,b]

Then, every function in

B" ( f ) = {g ∈ C([a, b]) ∶ ∀x ∈ [a, b] f (x) − " < g(x) < f (x) + "}

is an element of A, hence A is an open set.

. Exercise 1.12 Let X be a normed space. For A, B ⊂ X we define

A + B = {a + b ∶ a ∈ A, b ∈ B}.

(a) Prove that if either A or B is open then A + B is open.

(b) Give an example where both A and B are closed however A+ B is not closed.
 Metric Spaces 25

1.2.3 Limits and topology

The following proposition elucidates what is “closed” about a closed set:

Proposition 1.27 A subset C ⊂ X is closed if and only if

xn ∈ C and lim xn = x implies x ∈ C.
n→∞

That is, a closed set is a set that contains all its limit points.

Proof : There are two directions to prove:

1. A closed set contains the limits of all convergent sequences: Let C be

closed (i.e., its complement is open), and let

xn ∈ C and lim xn = x.
n→∞

Suppose by contradiction that x ∈� C. Since C c is open, there exists an r > 0,

such that
Br (x) ⊂ C c ,
and therefore
xn ∈� Br (x) ∀n ∈ N,
in contradiction to x being the limit of xn .

2. A set containing the limits of all converging sequences is closed: Suppose

that C has the property that every convergent sequence in C has its limit in
C. We need to show that C is closed, i.e., that C c is open. Suppose that C c
is not open. Then,

∃x ∈ C c such that ∀r > 0 Br (x) ⊂� C c ,

and equivalently,

∃x ∈� C such that ∀r > 0 Br (x) ∩ C ≠ �.

In particular,

∃x ∈� C such that ∀n ∈ N ∃xn ∈ B1�n (x) ∩ C.

26 Chapter 1

Then, xn is a sequence in C converging to x ∈� C, as

1
lim d(xn , x) ≤ lim = 0.
n→∞ n→∞ n

This contradicts the assumption, hence we conclude that C is closed.

Examples:

(a) Consider X = C([a, b]) endowed with the max-norm and let

A = { f ∈ C([a, b]) ∶ f is di↵erentiable}.

A is not closed because there are sequences of di↵erentiable functions con-

verging (uniformly) to a non-di↵erentiable limit. A is neither open because
for every f ∈ A and every r > 0, there are non-di↵erentiable functions in
Br ( f ).
(b) Let (X, d) be an arbitrary metric space and let xn be a sequence converging
to x. Then
A = {xn ∶ n ∈ N} ∪ {x}
is closed because every converging sequence in A is either eventually con-
stant or converges to x. In either case, the limit is an element of A.

. Exercise 1.13 Prove the above assertion: if xn is a sequence converging to x,

then
A = {xn ∶ n ∈ N} ∪ {x}
is a closed set.

As explained above, sets in metric spaces are not necessarily either open or closed.
Every set, however, has a largest open set contained in it, and a smallest closed set
containing it.

Definition 1.28 Let A ⊂ X be a set in a metric space. Its interior (.*15), A○ , is

the union of all the open subsets of A,

A○ = �{U ∶ U is open and U ⊂ A}

 Metric Spaces 27

(it is the largest open subset of A). Its closure (9&#2), A, is the intersection of all
the closed sets that include A,

A = �{C ∶ C is closed and A ⊂ C}

(it is the smallest closed set containing A).

Comment: Recall the notion of an interior point in a set. Likewise, for A ⊂ X,

we call x ∈ X an exterior point (;*1&7*( %$&81) of A if there exists an r > 0 such
that Br (x) ⊂ Ac . It is easy to verify that the interior of a set is the union of all its
interior points, whereas the closure of a set is the union of all the points that are
not exterior to it.

Proposition 1.29 The interior of a set is open and the closure of a set is closed.

Proof : Any union of open sets is open and any intersection of closed sets is closed.
n
Also,

Proposition 1.30 If U is open and C is closed, then

U○ = U and C = C.

Proof : Immediate from the definitions of U ○ being the maximal open set con-
tained in U and C being the minimal closed set containing C. n

Proposition 1.31 For every set A ⊂ X,

A○ ⊂ A ⊂ A.

Proof : This is immediate, as A○ is a union of sets all contained in A and A is an

intersection of sets all containing A. n
 28 Chapter 1

Example: Let X = R. Then,

Q○ = � and Q = R.
Indeed, Q does not contain any open segment, and since every open set is a union
of open segments, Q does not contain any non-empty open set. The second state-
ment can be justified as follows. Let C be a closed set containing Q. Since every
point in R is a limit of a sequence in Q, and since we have seen that closed sets
contain all their limit points, we conclude that C = R, i.e., R is the only closed set
—6h(2019)— containing Q. ▲▲▲

. Exercise 1.14 Prove that for every A ⊂ X

(A○ )c = Ac and (Ac )○ = (A)c .

The following are immediate consequences of the definitions of the interior and
the closure:

Lemma 1.32 Let (X, d) be a metric space and let A ⊂ X.

1. If U ⊂ A is open, then U ⊂ A○ .
2. If A ⊂ C and C is closed, then A ⊂ C.

Corollary 1.33 For every set A, the open sets contained in A coincide with the
open sets contained in A○ and the closed sets containing A coincide with the closed
sets containing A. As a result, the operations of interior and closure are idempo-
tent. For every set A,

(A○ )○ = A○ and A = A.

Proof : The first part is an immediate corollary of the lemma. For the second part,
A○ = �{U is open ∶ U ⊂ A}
= �{U is open ∶ U ⊂ A○ }
= (A○ )○ ,
 Metric Spaces 29

and
A = �{C is closed ∶ A ⊂ C}
= �{C is closed ∶ A ⊂ C}
= A.

n
We have seen that the property of being closed is related to limits being in that set.
The following proposition relates the closure of a set to limits:

Proposition 1.34 The closure of a set A is the set of points that are limits of
sequences in A,
A = {x ∈ X ∶ ∃xn ∈ A such that xn → x} .

Proof : Define
B = {x ∈ X ∶ ∃xn ∈ A such that xn → x} .
We need to show that A = B, i.e., that A ⊂ B and B ⊂ A.
Let x ∈ B. By definition, there exists a sequence xn ∈ A converging to x. Since
A ⊂ A, xn is a sequence in A. Since A is closed, x ∈ A, implying that B ⊂ A.
In the other direction, we first assert that B is closed. Let (bn ) be a sequence in B
converging to x ∈ X. By the definition of B, there exists for every n ∈ N a sequence
in A converging to bn , and in particular, an an ∈ A, such that
1
d(bn , an ) < .
n
It follows that
d(x, an ) ≤ d(x, bn ) + d(bn , an ) → 0.
By definition, x ∈ B, and by Proposition 1.27, B is closed.
Clearly, every point in A belongs to B (take the trivial sequence). Thus, B is a
closed set containing A, and by Lemma 1.32, A ⊂ B. n

Example: Let X = R2 and consider the set

A = {(cos 2⇡q, sin 2⇡q) ∶ q ∈ Q}.
30 Chapter 1

Then A ⊂ S1 , because every point

(cos 2⇡x, sin 2⇡x) ∈ S1

is a limit of points in A. On the other hand, since S 1 is closed and contains A,

A ⊂ S 1 , from which we deduce that A = S1 . ▲▲▲

Definition 1.35 The boundary (%5:) @A of a set A is defined as

@A = A � A○ .

Proposition 1.36 The boundary of a set is closed.

Proof : The boundary is the intersection of two closed sets, A, and (A○ )c , hence it
is closed. n

Example: In any normed space,

@Br (x) = B̂r (x) � Br (x)

= {y ∶ �y − x� ≤ r} � {y ∶ �y − x� < r}
= Sr (x).

This is not true in a general metric space. For example, in the case of a discrete
space,
@B1 (x) = B1 (x) � B1 (x) = � ≠ S1 (x),
where we used the fact that every set is both open and closed. ▲▲▲

Example: Consider the space X = C([0, 1]) endowed with the maximum norm,
and consider the set

A = { f ∈ C([0, 1]) ∶ f is strictly decreasing}.

We will obtain its interior, its closure and its boundary.

Set
B = { f ∈ C([0, 1]) ∶ f is decreasing}.
 Metric Spaces 31

Let fn ∈ B and fn → f . Then, f must be decreasing as well, otherwise if x < y and

f (x) < f (y), then it must hold that for n sufficiently large, fn (x) < fn (y), which is
a contradiction. Thus, B contains all it limit, and therefore it is closed.
Next, we claim that every element of B is a sequence of elements of A. Given
f ∈ B, set
fn (x) = f (x) − x�n.
Then,
d( fn , f ) = max x�n = 1�n,
0≤x≤1

and for every x < y,

fn (x) = f (x) − x�n > f (y) − y�n = fn (y),

i.e., fn ∈ A. Since B is closed, and since every point in B is a limit of a sequence

in A, it follows that
B ⊂ A.
On the other hand, since A ⊂ B and B is closed, it follows that A ⊂ B, namely,

A = B.

Let now f ∈ A and let r > 0. Since f is continuous, there exists a pair of points
x < y, such that f (y) > f (x) − r�2. Let g ∈ C([0, 1]) be any function satisfying

g(y) − g(x) = r�2 and �g�∞ < r�2.

Then,
f + g ∈ Br ( f ),
however
( f + g)(x) < f (y) + r�2 + g(y) − r�2 = ( f + g)(y),
i.e., f + g ∈� A. Thus, for every f ∈ A and every r > 0, the open ball Br (x)
contains elements not in A, namely, A does not have any non-empty open subset.
We conclude that
A○ = �.
Finally,
@A = A � A○ = B.
▲▲▲
32 Chapter 1

. Exercise 1.15 Let A, B be subsets of a metric space (X, d). Prove that

1. A ∪ B = A ∪ B.
2. A ∩ B ⊂ A ∩ B.
3. If A is open then (@A)○ = �.
4. A is closed if and only if @A ⊂ A.
5. A is open if and only if @A ∩ A = �.

. Exercise 1.16 Let A be a set in a metric space (X, d). Show that its boundary
@A coincides with the set of points x, such that every open ball with center at x
intersects both A and Ac .

Consider now a metric space (X, d). Any Y ⊂ X is a metric space with the induced
metric. Note, however, that a set A ⊂ Y can be open in the metric space (Y, d), but
closed in the metric space (X, d) (e.g., the set Y itself is always open in (Y, d), but
could be closed in (X, d)).

Proposition 1.37 Let (X, dX ) be a metric space and let Y ⊂ X; we denote the
induced metric on Y by dY . A set A ⊂ Y is open in (Y, dY ) if and only if it is the
intersection of Y and a set that is open in (X, dX ). The same applies if we replace
“open” by “closed”.

Proof : By Proposition 1.7, every ball in Y is the intersection of Y and a ball in X.

Suppose that A is open in Y. Then, for every a ∈ A, there exists an r(a) > 0, such
that Br(a)
Y
(a) ⊂ Y. It follows that

A = � Br(a)
Y
(a) = � (Y ∩ Br(a)
X
(a)) = Y ∩ � � Br(a)
X
(a)� ,
a∈A a∈A a∈A

proving that A is the intersection of Y and a set open in X.

Conversely, let
A = Y ∩ B,
where B is open in X. Since B is a union of open balls in X, A is a union of open
balls in Y. n
 Metric Spaces 33

Definition 1.38 Let (X, d) be a metric space. A set A ⊂ X is called dense (4&57)
if
A = X.

Example: The rationals Q are dense in R because Q = R. ▲▲▲

Definition 1.39 A metric space (X, d) is called separable (*-*"952) if it contains

a countable dense set. That is, there exists a countable set {xn ∶ n ∈ N} ⊂ X such
that
X = {xn ∶ n ∈ N}.

Example: R is separable because Q is countable and Q = R. ▲▲▲

. Exercise 1.17 Show that every subspace (Y, d) of a separable metric space
(X, d) is separable.
. Exercise 1.18 (a) Prove that in a normed space,
(B̂r (a))○ = Br (a) and Br (a) = B̂r (a).
(b) Prove that this is not necessarily true in a general metric space.
. Exercise 1.19 A point a in a metric space X is called an accumulation point
(;&9")7% ;$&81) of a set A if in any open ball centered at a there exists a point of
A (excluding a!). Prove that a is an accumulation point of A if and only if there
exists a sequence of distinct points in A that converges to a.
. Exercise 1.20 Let A be a set in a metric space. The set of its accumulation
points is called its derived set (;9'#1% ;7&"8) and is denoted by A∗ . Prove that A∗
is closed.
. Exercise 1.21 Prove that the only sets in R that are both open and closed are
R and �; that is, R is a connected space.
. Exercise 1.22 A point x ∈ A ⊂ X is called isolated ($$&"/) in A if there exists
a ball B̂r (x) such that B̂r (x) ∩ A = {x}. Prove that if A ⊂ R is a non-empty,
countable and closed set, then it contains at least one isolated point.
. Exercise 1.23 Prove that `∞ , the space of infinite sequence with the metric
induced by the norm
�x�∞ = sup �xn �
n
is not separable. Hint: use Cantor’s diagonalization technique.
34 Chapter 1

1.3 Continuous functions on metric spaces

Metric spaces are an abstraction generalizing the real line. Abstractions, such
as metric spaces, groups, rings, etc., never live on their own. They are always
associated with mappings (morphisms, in the language of categories) between
instances of the same abstraction, which preserve certain properties (e.g., groups
and homomorphisms). In the first calculus courses we learned about continuous
mappings (i.e., functions) R → R. This concept will be generalized into continu-
ous mappings between metric spaces.

1.3.1 Definitions and basic properties

Definition 1.40 Let (X, d) and (Y, ⇢) be two metric spaces. A mapping f ∶ X →
Y is said to be continuous (4*79) at a point a ∈ X if for every " > 0 there exists a
> 0 such that

d(x, a) < implies ⇢( f (x), f (a)) < ".

That is, f is continuous at a if for every open ball B" ( f (a)) ⊂ Y there exists an
open ball B (a) ⊂ X whose image under f is contained in B" ( f (a)). A function
is called continuous if it is continuous at all points.

f
f (a) f (B (a))
a

B (a) B" ( f (a))

Examples:
(a) Every constant function is continuous.
(b) Let (X, d) be an arbitrary metric space and let a ∈ X be an arbitrary point.
The function f ∶ x � d(x, a) is a continuous mapping f ∶ (X, d) → (R, � ⋅ �)
because for every " > 0 take = ". Then d(x, z) < implies that

� f (x) − f (z)� = �d(x, a) − d(z, a)� ≤ d(x, z) < ".

 Metric Spaces 35

(c) Every function defined on a discrete space and whose range is any metric
space is continuous. Indeed, for every " > 0 take = 1�2. Then,

f (B (a)) = f ({a}) = { f (a)} ⊂ B" ( f (a)).

Proposition 1.41 (Sequential continuity (;*;9$2 ;&5*79)) Let f ∶ (X, d) →

(Y, ⇢) be a function between two metric spaces. Then f is continuous at a if and
only if for every sequence xn in X

lim xn = a implies lim f (xn ) = f (a).

n→∞ n→∞

(This notion of continuity is attributed to Heine, unlike the "- notion, which is
attributed to Cauchy.)

Proof :

1. Suppose that f is continuous at a and let xn be a sequence in X that con-

verges to a. Since f is continuous at a,

∀" > 0 ∃ > 0 such that x ∈ B (a) implies f (x) ∈ B" ( f (a)).

Since xn → a,

∀ >0 ∃N ∈ N such that ∀n > N xn ∈ B (a).

Putting both assertions together,

∀" > 0 ∃N ∈ N such that ∀n > N f (xn ) ∈ B" ( f (a)),

i.e., f (xn ) → f (a).

2. Suppose that xn → a implies f (xn ) → f (a). Suppose that f were not con-
tinuous at a: then
1
∃" > 0 such that ∀n ∈ N ∃xn ∶ d(xn , a) < and ⇢( f (xn ), f (a)) ≥ ".
n
The sequence xn converges to a, however the sequence f (xn ) does not con-
verge to f (a), contradicting the assumption.
36 Chapter 1

Comment: In the more general context of topological spaces, continuity always

implies sequential continuity, but the converse is not true. —8h(2019)—

Examples:
(a) Let (X, � ⋅ �) be a normed space and let B = B1 (0) be the unit ball centered
at the origin. The indicator function (;1**7/% %*781&5%)
�
�
�1 x∈B
B (x) = �
�
� x ∈� B
�0
is not continuous. We will show it by constructing a converging sequence
xn → x for which f (xn ) →
� f (x).
Let x be a unit vector; by definition it is not in B, i.e., f (x) = 0. Take the
sequence
1
xn = �1 − � x ∈ B.
n
Then,
lim xn = x however lim f (xn ) = 1 ≠ f (x).
n→∞ n→∞

(b) Let X = C([a, b]). For x0 ∈ [a, b] define the function E ∶ X → R by

E( f ) = f (x0 )

(E is called an evaluation map). Then, E is continuous. Indeed, for every

f ∈ X and every sequence X ∋ fn → f ,

lim d( fn , f ) = lim max � fn (x) − f (x)� = 0.

n→∞ n→∞ a≤x≤b

In particular,

lim �E( fn ) − E( f )� = lim � fn (x0 ) − f (x0 )� = 0.

n→∞ n→∞

(c) Let X = C([−1, 1]) and let A ⊂ X be the subset of functions that are di↵er-
entiable on [−1, 1]. Consider the function D ∶ A → R defined by

D( f ) = f ′ (0).
 Metric Spaces 37

Then, D is not continuous (with A is endowed with the subspace metric).

Indeed, take the sequence of functions

1
fn (x) = sin nx.
n
and set f = 0. Then

lim d( fn , f ) = lim max � fn (x) − f (x)� = 0,

n→∞ n→∞ x∈[−1,1]

whereas
�D( fn ) − D( f )� = 1 →
� 0.

(d) Every function f ∶ X → Rn is continuous if and only if each of its component

is a continuous function X → R; this is because a sequence converges in Rn
if and only if each of its components converges.

The following theorem provides alternative characterizations to the continuity of

functions between metric spaces. In fact, these definitions are more general in the
sense that they remain valid in (non-metric) topological spaces (i.e., spaces that
are only endowed with a structure of open sets).

Theorem 1.42 Let f ∶ (X, d) → (Y, ⇢) be a function between two metric spaces.
Then the following three statements are equivalent:

1. f is continuous.
2. The pre-image of every set open in Y is a set open in X ( f pulls back (+:&/
9&(!-) open sets into open sets).
3. The pre-image of every set closed in Y is a set closed in X ( f pulls back
closed sets into closed sets).

Proof :

1. The equivalence between Property 2 and Property 3 follows from the com-
mutativity between the pre-images and set-theoretic operations. Thus, if the
38 Chapter 1

pre-image of every set open in Y is a set open in X and C ⊂ Y is closed in

Y, then
f −1 (C) = ( f −1 (C c ))c
is closed in Y. The other direction is obtained similarly.
2. We next show that Property 1 implies Property 2. Let A ⊂ Y be an open
set, and let x ∈ f −1 (A). Since A is open and f (x) ∈ A, there exists a ball
B" ( f (x)) ⊂ A. Since f is continuous, there exists a ball B (x) whose image
under f is a subset of B" ( f (x)), i.e., B (x) ⊂ f −1 (A), which implies that
f −1 (A) is open.

f −1 (A)
A
f

x
f (x)

B (x) f (B (x)) B" ( f (x))

3. It remains to show that Property 2 implies Property 1. Let x ∈ X be an

arbitrary point; by assumption, for every " > 0 the pre-image of the set
B" ( f (x)) is open in X, and includes the point x, i.e., there exists an open
ball B (x) ⊂ f −1 (B" ( f (x))), or, f (B (x)) ⊂ B" ( f (x)) which implies the
continuity of f at x.
n

Comment: Continuous function not necessarily map open sets into open sets. For
example the function f ∶ R → R defined by f ∶ x � 0 maps the open set (0, 1) into
the closed set {0}. Functions that map open sets into open sets are called open
maps (;&(&;5 ;&8;3%) (and likewise, functions that map closed sets into closed
sets are called closed maps (;&9&#2 ;&8;3%)). Open maps, on the other hand are
not necessarily continuous.

Definition 1.43 (Function composition) Let f ∶ (X1 , d1 ) → (X2 , d2 ) and g ∶

(X2 , d2 ) → (X3 , d3 ) be functions between metric spaces. Their composition g ○ f
is a function (X1 , d1 ) → (X3 , d3 ) defined by
(g ○ f )(x) = g( f (x)).
 Metric Spaces 39

Proposition 1.44 (Composition of continuous mappings) The composition of

continuous mappings is a continuous mapping.

Proof : Let f ∶ (X1 , d1 ) → (X2 , d2 ) and g ∶ (X2 , d2 ) → (X3 , d3 ) be continuous func-

tions between metric spaces. The proof is immediate if we use the “topological
interpretation” and observe that

(g ○ f )−1 = f −1 ○ g−1 .

Then, for every open A ⊂ X3 , g−1 (A) is open in X2 and f −1 (g−1 (A)) is open in X1 ,
proving that g ○ f is continuous.
Alternative proof using Heine’s characterization: Let xn → a in X1 . Then, f (xn ) →
f (a) in X2 and g( f (xn )) → g( f (a)) in X3 . n

Lemma 1.45 Let (X, d) be a metric space and let f ∶ X → R and g ∶ X → R be

continuous. Then, the map h ∶ X → R2 defined by

h(x) = ( f (x), g(x))

is continuous.

Proof : This is an immediate consequence of sequential continuity. Let xn → a.

Then,
�
d(( f (xn ), g(xn )), ( f (a), g(a))) = ( f (xn ) − f (a))2 + (g(xn ) − g(a))2
≤ 2� f (xn ) − f (a)� + 2�g(xn ) − g(a)�
→ 0.

Corollary 1.46 For continuous functions f, g ∶ X → R, the functions f ± g, and

f ⋅ g are continuous. Furthermore, f �g is continuous at all points where g ≠ 0.
40 Chapter 1

Proof : Let h ∶ X → R2 be defined as in Lemma 1.45. Let s1 , s2 , s3 ∶ R2 → R be

defined by
s1 (x, y) = x + y s2 (x, y) = x − y and s3 (x, y) = xy.
The maps s1 , s2 , s2 are continuous, because, for example, as we know from ele-
mentary calculus, if xn → x and yn → y, then xn + yn → x + y. By Proposition 1.44,
the map
s1 ○ h(x) = f (x) + g(x)
is continuous, and the same argument holds for subtraction and multiplication. To
prove the continuity of division, it suffices to show that the mapping x → 1�x is
continuous at all points where x ≠ 0 (which we know from elementary calculus).
n

Proposition 1.47 Let (X, d) and (Y, ⇢) be metric spaces and let f ∶ X → Y be
continuous. Let X′ ⊂ X be non-empty. Then,

f �X′ ∶ X′ → Y

is continuous, where X′ is endowed with the subspace metric.

Proof : Since f is continuous, for every A ⊂ Y,

f −1 (A) = {x ∈ X ∶ f (x) ∈ A}
is open in X. By Proposition 1.37,
f �−1
X′ (A) = {x ∈ X ∶ f (x) ∈ A} = X ∩ {x ∈ X ∶ f (x) ∈ A} = X ∩ f (A)
′ ′ ′ −1

is open in X′ , thus proving that f �−1

X′ is continuous. n
The following proposition generalizes a theorem you proved in first year calculus
for the case of real-valued functions:

Proposition 1.48 Let (X, d) and (Y, ⇢) be metric spaces and let fn ∶ X → Y be a
sequence of continuous functions satisfying

lim sup ⇢( fn (x), f (x)) = 0,

n→∞ x∈X

where f ∶ X → Y. Then, f is also continuous.

 Metric Spaces 41

Proof : Let xk → x in X. By Heine’s characterization, we need to show that

lim ⇢( f (xk ), f (x)) = 0.
k→∞

For every k and every n,

⇢( f (xk ), f (x)) ≤ ⇢( f (xk ), fn (xk )) + ⇢( fn (xk ), fn (x)) + ⇢( fn (x), f (x))
≤ 2 sup ⇢( fn (z), f (z)) + ⇢( fn (xk ), fn (x)).
z∈X

Letting k → ∞, using the fact that all the fn are continuous,

lim sup ⇢( f (xk ), f (x)) ≤ 2 sup ⇢( fn (z), f (z)).
k→∞ z∈X

Since this holds for every n, and since the right-hand side tends to zero as n → ∞
we obtain the desired result. n

. Exercise 1.24 Let A ⊂ X. The indicator function (;1**7/ %*781&5) of the the
�
set A is defined as
�
�1 x ∈ A
A (x) = �
�
�0 x ∈� A.
�
Show that the points of discontinuity of A coincide with @A.

. Exercise 1.25 Let f ∶ R → R be a bounded function.

(a) Prove that f is continuous if and only if its graph
{(x, f (x)) ∶ x ∈ R} ⊂ R2
is closed in R2 .
(b) Show why this is not generally true if f is not bounded.

. Exercise 1.26 Let F, G be sets in a metric space X and let Y be a metric space.
Given two functions f ∶ F → Y and g ∶ G → Y such that f = g on F ∩ G, we define
�
the function
�
� f (x) x ∈ F
( f ∧ g)(x) = �
�
�g(x) x ∈ G.
�
1. Prove that if F, G are closed in X and f, g are continuous, then f ∧ g is
continuous.
2. Prove that this is also correct if F, G are open in X.
3. Show that this is not necessarily the case if F is open and G is closed.
42 Chapter 1

1.3.2 Functions from Rk to Rm

Functions between vector spaces, and in particular, functions between Rk to Rm
for some k, m ∈ N play a prominent role in analysis. Unless otherwise specified,
we will always assume that those spaces are endowed with the Euclidean metric.

Comments:
(a) A function f ∶ Rk → Rm gets for input a k-tuple of real numbers and returns
an m-tuple of real numbers. We can represent it as

f (x1 , . . . , xk ) = ( f1 (x1 , . . . , xk ), . . . , fm (x1 , . . . , xk )).

(b) The graph of a function f ∶ X → Y between two sets is the set

Graph( f ) = {(x, f (x)) ∶ x ∈ X} ⊂ X × Y.

In particular, for X = Rk and Y = Rm ,

Graph( f ) = {(x, f (x)) ∶ x ∈ Rk } ⊂ Rk × Rm = Rk+m .

(c) When m = 1, the function is called a scalar function (;*9-82 %*781&5). Its
graph is a surface in Rm+1 .
(d) When k = 1, the function

f (t) = ( f1 (t), . . . , fm (t)).

is called a path (%-*2/). You could think of the parameter t as “time”.

Example: Consider the path f ∶ [0, 2⇡] → R2 given by

f (t) = (cos t, sin t).

Its image is the unit circle. Note that if the domain is [0, 4⇡] or the whole of R,
the image is still the unit circle. What about its graph? It is the set

(t, cos t, sin t) ⊂ R3 ,

which is a helix (-*-2) whose axis is along the x-axis. ▲▲▲

 Metric Spaces 43

Proposition 1.49 A function f ∶ Rk → Rm is continuous if and only if each of its

m components f j ∶ Rk → R is continuous.

Proof : Denote by ⇡ j the projection, ⇡ j ∶ Rm → R,

⇡ j (y1 , . . . , ym ) = y j , j = 1, . . . , m.
Each ⇡ j is a continuous function, because if A ⊂ R is open, then

j (A) = R × ⋅ ⋅ ⋅ × R ×A × R × . . . R
⇡−1
��� � � � � � � � � �� � � � � � � � � � � ��� � � � � � ��� � � � � � �
j − 1 times m − j times

is open in Rm .
Now suppose that f ∶ Rk → Rm is continuous; then f j = ⇡ j ○ f is continuous as
a composition of continuous functions. Conversely, suppose that all the f j are
continuous, and let xn = (x1n , . . . , xkn ) ∈ Rk converge to x ∈ Rk . Then
lim f j (xn ) = f j (x) for all j = 1, . . . , m,
n→∞

hence
lim ( f1 (xn ), . . . , fm (xn )) = ( f1 (x), . . . , fm (x)).
n→∞
n

Example: Any linear function A ∶ Rk → Rm is continuous because

k
(A(x1 , . . . , xk )) j = � a ji xi ,
i=1

which is clearly continuous. ▲▲▲

1.4 Compactness
1.4.1 Definition and basic properties
In this section, we study an important property that certain metric spaces (or sub-
sets of) possess: compactness. As we will see, compactness, is a certain measure
of “smallness”.
44 Chapter 1

Definition 1.50 (Compactness) A metric space (X, d) is called compact (*)85/&8)

if every sequence xn has a convergent subsequence. A subset K ⊂ X is called
compact if the subspace (K, d) is compact (i.e., if every sequence in K has a sub-
sequence that converges to a limit in K).

Comment: The compactness of a set requires two conditions: (i) that every se-
quence has a converging subsequence, and (ii) that the limit belongs to that set.
The second conditions is superfluous when the set is the entire space.

Comment: Note that the compactness of a set is a property pertinent only to the
set, whereas openness/closedness is property of a set in relation with the entire
space.

. Exercise 1.27 Let (X, d) be a metric space and let Y ⊂ X with the subspace
metric dY . Prove that a set A ⊂ Y is compact as a subset of Y if and only if it is
compact as a subset of X.

. Exercise 1.28 Let (X, d) be a metric space and xn a sequence that converges
to a limit x. Prove that the set {x} ∪ {xn ∶ n ∈ N} is compact.

. Exercise 1.29 Consider the space {0, 1}N (infinite binary sequences) with the
metric
d(x, y) = 2min{i ∶ xi ≠yi } .
(First prove that it is indeed a metric.) Prove that this space is compact. Hint:
show first that convergence in this space amounts to component-by-component
convergence.

Proposition 1.51 A finite union of compact sets is compact.

Proof : Given a sequence xn ∈ �ni=1 Ki = A, with Ki compact, it must have an

infinite subsequence in at least one of the Ki ’s. Since this Ki is compact, there is a
sub-subsequence converging in Ki hence in A. n

Examples:
(a) Every singleton in a metric space is compact.
 Metric Spaces 45

(b) Every metric space having a finite number of elements is compact.

(c) The Bolzano-Weierstraß theorem implies that a closed segment on the line
is compact.

Definition 1.52 A set A in a metric space (X, d) is called ↵-discrete if

∀x ≠ y ∈ A d(x, y) ≥ ↵.

Lemma 1.53 In a metric space (X, d), a set containing an infinite ↵-discrete set
for some ↵ > 0 is not compact.

Proof : By definition, one can construct an infinite sequence xn ∈ A, such that

∀m ≠ n d(xm , xn ) ≥ ↵.

If xn has a subsequence xnk converging to a limit x, then there exist k ≠ ` such that

d(xnk , x) < d(xn` , x) <

↵ ↵
and .
2 2
Hence,
d(xnk , xn` ) ≤ d(xnk , x) + d(x, xn` ) < ↵,
which is a contradiction. n

Comment: The other direction is not true: a set may fail to contain an ↵-discrete
set for every ↵ > 0 and yet not be compact. The segment (0, 1) ⊂ R is an example.

. Exercise 1.30 True of false: if a set A in a metric space contains a subset

B ⊂ A which is not compact, then A is not compact.
—10h(2019)—

Proposition 1.54 A compact set in a metric space is both bounded and closed.
46 Chapter 1

Proof : Closedness: Let K be a compact set in a metric space (X, d). By defini-
tion, every sequence xn ∈ K has a subsequence that converges to a point in K. In
particular, if xn is a sequence converging to a point x ∈ X, then by the uniqueness
of the limit, x ∈ K. It follows from Proposition 1.27 that K is closed.
Boundedness: Suppose, by contradiction, that K was not bounded. We will show
that it contains a 1-discrete sequence, hence it is not compact. Take an arbitrary
point x1 . Since K is not bounded, there exists a point x2 ∈ K � B1 (x1 ). Inductively,
there exists for every n a point

n−1
xn ∈ K � � B1 (x j ),
j=1

forming a 1-discrete sequence in K. n

Comment: One often confuses between “compact” and “bounded and closed”. In
general, the implication is only uni-directional, although we will soon see situa-
tions in which there is indeed a correspondence between the two. An elementary
counter-example is the segment (0, 1), which is closed (as a space) and bounded,
but not compact.

Proposition 1.55 If (X, d) is a compact metric space, then every closed subset of
X is compact.

Proof : Let K be closed in X and let xn ∈ K. Since X is compact, xn has a subse-

quence xnk that converges to some limit x ∈ X. Since K is closed, x ∈ K, hence K
is compact. n

Corollary 1.56 In a compact metric space closedness and compactness are the
same.

. Exercise 1.31 Prove that in a metric space in which all the closed balls are
compact, the compact sets coincide with the sets that are closed and bounded.
 Metric Spaces 47

Definition 1.57 Let (X, d) be a metric space with A, B ⊂ X and a ∈ X. We

define the distance between a point and a set, and the distance between two sets
as follows:
d(a, A) = inf d(a, x) and d(A, B) = inf d(x, y).
x∈A x∈A,y∈B

(Note that we write inf because we can’t guarantee the existence of a minimum.)

Proposition 1.58 Let A, B ⊂ X be nonempty, closed disjoint sets, such that at

least one of them is compact. Then d(A, B) > 0.

Example: A non-example: let X = R2 with A = {(x, 1�x) ∶ x > 0} and B =

{(x, 0) ∶ x ≥ 0}. The sets A, B are non-empty, disjoint and closed (why?). Yet,
d(A, B) = 0. There is no contradiction because neither A nor B is compact. ▲ ▲ ▲

Proof : Suppose that A is compact and B is closed. Assume, by contradiction, that

d(A, B) = 0. Then, there exist sequences xn ∈ A, yn ∈ B, such that
lim d(xn , yn ) = 0.
n→∞

Since A is compact, the sequence xn has a subsequence xnk with limit x ∈ A, Then,
d(x, ynk ) ≤ d(x, xnk ) + d(xnk , ynk ) → 0,
from which follows that ynk → x. Since B is closed, x ∈ B, contradicting the
disjointness of A and B. n

Corollary 1.59 If U is open in (X, d) and K ⊂ U is compact, then U c is closed

and disjoint from K, from which follows that d(K, U c ) > 0. In particular, since
@U ⊂ U c , d(K, @U) > 0.

Uc

U
48 Chapter 1

. Exercise 1.32 For a countable sequence metric spaces (Xn , dn ) define the prod-
uct space
∞
X = � Xn = {(x1 , x2 , . . . , ) ∶ x1 ∈ X1 , x2 ∈ X2 , . . . }
n=1
with the metric ∞
dn (xn , yn )
d(x, y) = � 2−n
1 + dn (xn , yn )
.
n=1
Prove that if each Xn is a compact space, then the product space X is compact.
Hint: think of Cantor’s diagonalization.

1.4.2 Topological characterizations of compactness

In this section we will see equivalent characterizations of compactness (which in
fact are the definitions in the more general context of topological spaces).

Definition 1.60 Let (X, d) be a metric space. A covering (*&2*,) of X is a col-

lection of subsets {A↵ ∶ ↵ ∈ I}, whose union is X. A covering is called a finite
covering if the number of sets A↵ is finite. It is called an open covering (*&2*,
(&;5) if all the A↵ are open sets (in X).

Let {A↵ } be a covering of X. One may ask whether we could still cover X with
a smaller collection of subsets. Consider for example the case where X = R; the
countable collection of subsets
{(n − 1, n + 1) ∶ n ∈ Z}
is an open covering of R. The omission of any of those (n − 1, n + 1) would
fail to cover the point x = n; that is, this open covering does not have a strict sub-
covering, not to speak of a finite sub-covering. As the main theorem below shows,
compactness can be defined alternatively as “every open covering can be reduced
to a finite sub-covering”.

Lemma 1.61 Let (X, d) be a compact metric space, and let F an open covering
of X (F is a collection of open subsets of X). Then, there exists an " > 0 such that
every open ball of radius " is contained in one of the sets in F . That is,

∃" > 0 such that ∀x ∈ X ∃A ∈ F such that B" (x) ⊂ A.

 Metric Spaces 49

Comment: Given an open covering, the supremum over all such " is called the
Lebesgue number (#⌥"- 952/) of the covering.

Proof : We need to prove that

∃n ∈ N ∀x ∈ X ∃A ∈ F such that B1�n (x) ⊂ A.

Suppose that the desired property does not hold. That is,

∀n ∈ N ∃xn ∈ X such that ∀A ∈ F B1�n (xn ) ⊂� A.

Since X is compact, the sequence xn has a subsequence xnk converging to some

x ∈ X. Since F is a covering of X, x ∈ A for some A ∈ F . Moreover, since A is
open,
∃m ∈ N, B1�m (x) ⊂ A.
Since x is a partial limit of xn ,

∃n > 2m, xn ∈ B1�2m (x).

Thus, for every y ∈ B1�n (xn ),

1 1 1
d(y, x) ≤ d(y, xn ) + d(xn , x) < + < ,
n 2m m
i.e., y ∈ B1�m (x), which implies that

B1�n (xn ) ⊂ B1�m (x) ⊂ A ∈ F ,

in contradiction with our assumption. n

Definition 1.62 A metric space (X, d) is called totally-bounded (-*-, .&2() if

for every " > 0, X can be covered by a finite set of open balls of radius ". That is,
n
∀" > 0 ∃x1 , . . . , xn such that � B" (xi ) = X.
i=1

Proposition 1.63 Let (X, d) be a compact metric space. Then, it is totally-

bounded.
50 Chapter 1

Proof : Suppose that the space was not totally bounded. Then, there exists an
" > 0 such that no finite collection of balls of radius " covers X. As a result, we
can construct inductively an "-discrete sequence, proving that X is not compact.
n

Lemma 1.64 Let xn be a sequence in a metric space (X, d). If xn doesn’t have a
converging subsequence, then the set

{xn ∶ n ∈ N}

is closed.

Proof : Denote A = {xn ∶ n ∈ N}. Let x ∈� A. Since x is not a partial limit of the
sequence xn , it is not true that
∀" > 0 ∀n0 ∈ N ∃k > n0 such that d(xk , x) < ".
That is,
∃" > 0 ∃n0 ∈ N such that ∀k > n0 d(xk , x) ≥ ".
Since x ∈� A, there exists an
r = min{d(xn , x) ∶ n ≤ n0 } > 0.
Then,
∀n ≤ n0 d(xn , x) > r and ∀n > n0 d(xn , x) > ",
i.e., Bmin(r,") (x) ⊂ Ac , proving that Ac is open, i.e., that A is closed. n

Theorem 1.65 (Heine-Borel) The following three properties of a metric space

(X, d) are equivalent:

1. X is compact.
2. Every open covering of X has a finite sub-covering.
3. Every collection of closed sets {F↵ }↵∈I has the property that if the intersec-
tion of every finite sub-collection is not empty, then the intersection of the
entire collection is not empty.
 Metric Spaces 51

Proof : We start by proving that Property 2 implies Property 3. Let {F↵ ∶ ↵ ∈ I}

be a collection of closed sets, hence {F↵c ∶ ↵ ∈ I} is a collection of open sets. If
Property (2) holds, then

� F↵c = X �⇒ ∃ finite J ⊂ I such that � F↵c = X.

↵∈I ↵∈J

By duality.
c c
�� F ↵ �
c
=� �⇒ ∃ finite J ⊂ I such that � � F↵ �
c
= �.
↵∈I ↵∈J

Using de Morgan’s laws,

� F↵ = � �⇒ ∃ finite J ⊂ I such that � F↵ = �.

↵∈I ↵∈J

Recalling that A ⇒ B is the same as Bc ⇒ Ac ,

∀ finite J ⊂ I � F↵ ≠ � �⇒ � F↵ ≠ �.
↵∈J ↵∈I

We next show that Property 3 implies Property 1. Suppose that Property 3 holds,
and let xn be a sequence. We need to show that it has a converging subsequence.
Suppose, by contradiction, that it doesn’t, and consider the sets
An = {xk ∶ k ≥ n}.
By Lemma 1.64, the sets An are closed; they are also decreasing, and have have
the property that every finite sub-collection has a non-empty intersection, as for
every finite set J ⊂ N,
� An = Amax(J) .
n∈J
By Property 3, there exists an
∞
x ∈ � An .
n=1
This means that x repeats infinitely many times in the sequence xk , hence x is a
partial limit, contradicting the assumption. —12h(2019)—
It remains to show that Property 1 implies Property 2. Let X be compact and let
U be an open covering. By Lemma 1.61 there exists an " > 0 such that every
open ball of radius " is contained in an element of U , namely,
∀x ∈ X ∃V x ∈ U ∶ B" (x) ⊂ V x .
52 Chapter 1

By Proposition 1.63, X is totally-bounded, i.e., we can cover X with a finite num-

ber of open balls of radius ",
n
∃{xi }ni=1 ∶ � B" (xi ) = X.
i=1

Then,
n
� V xi = X,
i=1

i.e., U has a finite subcover of X n

Comment: The term compactness was introduced by Maurice Fréchet in 1906. It

was primarily introduced in the context of metric spaces in the sense of “sequential
compactness”. The “covering compact” definition has become more prominent
because it allows us to consider general topological spaces, and many of the old
results about metric spaces can be generalized to this setting. This generalization
is particularly useful in the study of function spaces, many of which are not metric
spaces.
One of the main reasons for studying compact spaces is because they are in some
ways very similar to finite sets. In other words, there are many results which are
easy to show for finite sets, the proofs of which carry over with minimal change
to compact spaces.

Example: A non-example: the space X = (0, 1) is not compact, yet for every
" > 0 we can cover (0, 1) with a finite number of segments of size " (this space is
totally-bounded). However, the open covering of (0, 1),

F = {(1�(n + 2), 1�n) ∶ n ∈ N} ,

does not have a finite sub-covering. ▲▲▲

The Heine-Borel theorem refers to compact metric spaces. An analogous theorem
applies for compact sets in metric spaces:

Theorem 1.66 (Heine-Borel for sets) A set K in a metric space X is compact if

and only if for every collection of open sets {A↵ }↵∈I , such that K ⊂ �↵∈I A↵ , there
exists a finite sub-collection {A↵ }↵∈J , such that K ⊂ �↵∈J A↵ .
 Metric Spaces 53

. Exercise 1.33 Prove it.

Recall that a set A ⊂ X is called dense (4&57) in X if A = X (every non-empty
open set in X intersects A), and that a metric space is called separable (*-*"952)
if is contains a dense countable subset.

Example: R is separable because it contains the countable dense set Q. So is any

Rn , which contains the countable dense set Qn . ▲▲▲

Proposition 1.67 A compact metric space is separable.

Proof : Let X be compact. Then, it is totally bounded, hence it can be covered by

a finite number of open balls of size 1,
n1
∃x11 , . . . , xn11 such that X = � B1 (xi1 ).
i=1

Likewise, every k ∈ N,
nk
∃x1k , . . . , xnk k such that X = � B1�k (xik ).
i=1

Consider the countable set,

∞ nk
A = � �{xik }.
k=1 i=1

The set A is dense in X, because for every x ∈ X and every " > 0, take k > 1�".
Since
nk
X = � B1�k (xik ),
i=1

there exists an xik such that d(xik , x) < 1�k < ". n

TA material 1.2 The Hausdor↵ distance (also called Pompeiu-Hausdor↵ dis-

tance) measures how far two subsets of a metric space are from each other. It
turns the set of non-empty compact subsets of a metric space into a metric space
in its own right.
54 Chapter 1

. Exercise 1.34 Let (X, d) be a metric space in which the closed-and-bounded

sets are compact (like Rn ). Prove that for any a ∈ X and F ⊂ X non-empty and
closed, there exists a point x ∈ F such that

d(a, x) = d(a, F).

. Exercise 1.35 Show that the finite product space of compact spaces is com-
pact: Let (X1 , d1 ), . . . , (Xn , dn ) be metric spaces and consider the product space
n
X = � {(x1 , . . . , xn ) ∶ x j ∈ X j ∀ j} ,
i=1

with
d(x, y) = max di (xi , yi ).
j=1,...,n

(1) Prove that (X, d) is a metric space. (2) Prove that if the (Xi , di ) are compact
then (X, d) is compact.

1.4.3 Continuous functions on compact sets

Proposition 1.68 (Continuity and compactness) Continuous functions between

metric spaces map compact sets into compact sets.

Proof : Let f ∶ X → Y be continuous and let K ⊂ X be a compact set. We need to

show that f (K) ⊂ Y is compact. Take a sequence yn in f (K). By the definition of
f (K),
∀yn ∈ f (K) ∃xn ∈ K such that f (xn ) = yn .
Since K is compact, xn has a subsequence xnk converging to a limit x ∈ K,

lim xnk = x ∈ K.
k→∞

Since f is continuous,

lim ynk = lim f (xnk ) = f (x) ∈ f (K),

k→∞ k→∞

hence f (K) is compact. n

 Metric Spaces 55

. Exercise 1.36 Suppose that f ∶ X → Y is a continuous function between two

metric spaces. Is it true that the pre-image of every compact set is compact.

Corollary 1.69 Suppose that X is compact and f ∶ X → Y is continuous. Then,

the image of every closed set is closed.

Proof : Let A ⊂ X be closed. By Corollary 1.55 it is compact. By Proposition 1.68

f (A) is compact, and by once again Corollary 1.55, it is closed. n
The following theorem is a generalization of a well-known theorem in elementary
calculus:

Proposition 1.70 (Extremal values in compact spaces) A continuous function

from a compact metric space (equivalently, from a compact set in a metric space)
into R assumes a minimum and a maximum.

Proof : Let f ∶ X → R with X compact. By the previous proposition, the set

f (X) ⊂ R is compact, hence bounded and closed. Since it is bounded, it has a
finite supremum, and since it is closed this supremum must belong to the set, i.e.,
it is a maximum. n
(Note the advantage of working with abstract concepts: we prove powerful theo-
rems without having to deal with problem-specific details.)

Definition 1.71 (Uniform continuity) A mapping f ∶ (X, d) → (Y, ⇢) is called

uniformly continuous (%&&: %$*/" %5*79) if
∀" > 0 ∃ >0 ∶ ∀{x, y ∶ d(x, y) < } ∶ ⇢( f (x), f (y)) < ".
That is, f is uniformly continuous if it is continuous, and the same can be chosen
for all points. The function (") (with (") being the supremum of all ’s) is called
the continuity modulus (;&5*79% 2&-&$&/) of the function.

Definition 1.72 (Hölder and Lipschitz continuity) A mapping f ∶ (X, d) → (Y, ⇢)

is called Hölder continuous of order ↵ if there exists a real number k > 0 such
that for every a, b ∈ X,
⇢( f (a), f (b)) ≤ k [d(a, b)]↵ .
56 Chapter 1

The mapping is called Lipschitz continuous (6*:5*-) if it is Hölder continuous of

order ↵ = 1.

. Exercise 1.37 Let F ⊂ X be a subset in a metric space. Prove that the function
X → R defined by x � d(x, F) is Lipschitz continuous.

Theorem 1.73 A continuous function on a compact metric space (equivalently, on

a compact set in a metric space) is uniformly continuous.

Proof : Let f ∶ (X, d) → (Y, ⇢) be continuous with X compact. Suppose that f

were not uniformly continuous. Then,

∃" > 0 ∶ ∀ > 0 ∃x, y ∶ d(x, y) < and ⇢( f (x), f (y)) ≥ ".

In particular,

∃" > 0 ∶ ∀n ∈ N ∃xn , yn ∶ d(xn , yn ) < 1�n and ⇢( f (xn ), f (yn )) ≥ ".

Since X is compact, there exists a subsequence xnk , ynk such that xnk → x and
ynk → y. For every n,

d(x, y) ≤ d(x, xnk ) + d(xnk , ynk ) + d(ynk , y),

hence,
d(x, y) ≤ lim inf (d(x, xnk ) + d(xnk , ynk ) + d(ynk , y)) = 0,
k→∞

from which follows that x = y. By the continuity of the metric ⇢,

lim ⇢( f (xnk ), f (ynk )) = ⇢( f (x), f (x)) = 0,

k→∞

which is a contradiction. n

1.4.4 Homeomorphisms and isometries

In this section we study notions of equivalences between metric spaces:
 Metric Spaces 57

Definition 1.74 (Homeomorphism) A mapping f ∶ (X, d) → (Y, ⇢) between met-

ric spaces is called a homeomorphism (.'*59&/&!/&%) if it is one-to-one, onto, and
both f and f −1 are continuous, i.e., A ⊂ X is open if and only if f (A) ⊂ Y is open.
Two metric spaces are called homeomorphic if there exists a homeomorphism
between them.

Comment: It is easy to see that being homeomorphic is an equivalence relation

between metric spaces: it is symmetric, reflexive and transitive. In fact, homeo-
morphism is a topological concept (only requires the definition of open sets), and
is also called topological equivalence.

Examples:
(a) The identity map from (Rn , �⋅� p ) to (Rn , �⋅�q ) is a homeomorphism, because
open balls with respect to the � ⋅ � p metric are open sets (although not balls)
with respect to the � ⋅ �q metric, and vice versa.
(b) Every two open segments (a, b) and (c, d) on the line are homeomorphic,
because the linear mapping x → c + (d − c)(x − a)�(b − a) is a homeomor-
phism.
(c) Every open segment (a, b) is homeomorphic to the entire real line because
it is homeomorphic to the open segment (−⇡�2, ⇡�2) and the latter is home-
omorphic to the real line though the homeomorphism x � tan x.

An homeomorphism is before all an equivalence between sets, i.e., a bijection; it

identifies points in one space with points in another space. The additional property
of being a homeomorphism is summarized by the following proposition:

Proposition 1.75 A homeomorphism f ∶ X → Y preserves the topological prop-

erties of openness, closedness and compactness. That is, A ⊂ X is open (resp.
closed or compact) if and only if f (A) ⊂ Y is open (resp. closed or compact).

Proof : This is an immediate corollary on the fact that the pre-image of an open set
under f or f −1 is open and the image of a compact set under f or f −1 is compact.
n
58 Chapter 1

Proposition 1.76 If f is a continuous, one-to-one mapping from a compact metric

space (X, d) onto an (arbitrary) metric space (Y, ⇢), then f is a homeomorphism.

Proof : Since f is one-to-one and onto, f −1 is well-defined, and it only remains to

prove that it is continuous. By Theorem 1.42, it suffices that the pre-image under
f −1 of a closed set in X is closed in Y, namely, that f maps closed sets into closed
sets. This follows from Corollary 1.69. n
How does one show that two metric spaces are not homeomorphic? In principle,
we should show that there does not exist a homeomorphism between them. We
can’t of course check all mappings. The idea is to show that the two spaces di↵er
in a topological invariant—a property that is preserved by homeomorphisms.

Examples:

(a) R is not homeomorphic to a discrete space because in a discrete space all

compact sets are finite, whereas this not true in R. Thus, the image of [0, 1]
in the discrete space cannot be compact.
(b) (0, 1) is not homeomorphic to [0, 1] because the first is not compact whereas
the second is.
(c) The unit square [0, 1] × [0, 1] is not homeomorphic to [0, 1]. Both spaces
are compact, and it requires more subtle arguments to show it.
(d) Rn is not homeomorphic to Rm for m ≠ n. This is a very deep theorem
(invariance of domain, Brouwer 1912), which is hard to prove and beyond
the scope of this course (requires advanced topology).

A homeomorphism preserves topological properties of sets but it does not neces-

sarily preserve metric properties such that boundedness. Metric equivalence is a
particular case of topological equivalence:

Definition 1.77 Two metric spaces (X, d) and (Y, ⇢) are called isometric (.**9)/&'*!)
if they are homeomorphic, and there exists a homeomorphism f ∶ X → Y which is
distance-preserving: for all x, y ∈ X,

⇢( f (x), f (y)) = d(x, y).

 Metric Spaces 59

Comment: Like homeomorphism, isometry is an equivalence relation between

metric space.

. Exercise 1.38

1. Show that every distance-preserving function from X onto Y is an homeo-

morphism. Give an example to show that the converse is not true.
2. Prove that every isometry from R onto itself is of the form x � ±x + a.

. Exercise 1.39 Show that every distance-preserving function from a compact

space into itself is onto.

. Exercise 1.40 Let f ∶ (0, 1) → R be continuous. Show that f can be extended

into a continuous function on [0, 1] i↵ f is uniformly continuous on (0, 1).

. Exercise 1.41 Prove that the following spaces are homeomorphic (using the
natural metrics):

1. R2 .
2. The unit disc: D2 = {(x, y) ∶ x2 + y2 < 1}.
3. The unit square: Sq2 = {(x, y) ∶ max(�x�, �y�) < 1}.
4. The unit sphere in R3 less one point, e.g.
S 2 � � = {(x, y, z) ∶ x2 + y2 + z2 = 1} � {(0, 0, 1)}.

1.4.5 Compact sets in normed spaces

We have seen that a compact set in a metric space is always closed and bounded.
What about the opposite: is every closed and bounded set compact? The gen-
eral answer is no. In this section, we will examine this question in the particular
context of normed spaces.

Theorem 1.78 (Bolzano-Weirestraß in Rm ) In Rm (endowed with the Euclidean

metric) closed and bounded sets are compact.
60 Chapter 1

Proof : Let K ⊂ Rm be closed and bounded. Since it is bounded, there exists an

M > 0, such that for every x = (x1 , . . . , xm ) ∈ K,

m 1�2
d(x, 0) = �� �x � � i2
< M.
i=1

In particular,
�xi � < M ∀i = 1, . . . , m.

Let now xn = (xn1 , . . . , xnm ) be a sequence in K. Consider the sequence of real

numbers xn1 , formed by the first component. Since this sequence is bounded, xn
has by the Bolzano-Weierstraß theorem a subsequence xnk such that xn1k has a limit
x1 . Consider now the second component xn2k of that subsequence: once again,
this is a bounded sequence in R, hence there exists a sub-subsequence xnk` , such
that xn2k converges to a limit x2 . We proceed component-by-component, until we
`
obtain a subsequence, let’s denote it yn , such that each component yin converges to
a limit xi . It follows that yn → x = (x1 , . . . , xm ). Since K is closed, x ∈ K, hence K
is compact. n

Corollary 1.79 The closed unit ball B̂1 (0) in Rm is compact, and so is the unit
sphere S1 (0).

Definition 1.80 Let V be a vector space and let � ⋅ � and � ⋅ �′ be norms on V. The
norms are called equivalent (;&-&8:) if there exist constants c1 , c2 > 0, such that
for all v ∈ V,
c1 �v� ≤ �v�′ ≤ c2 �v�.
Another way to state it is that the mapping Id ∶ (V, � ⋅ �) → (V, � ⋅ �′ ) is Lipschitz
and so is its inverse (it is bi-Lipschitz).

Proposition 1.81 Let � ⋅ � and � ⋅ �′ by equivalent norms on V. Then the identity

Id ∶ (V, � ⋅ �) → (V, � ⋅ �′ )

is a homeomorphism mapping bounded sets to bounded sets.

 Metric Spaces 61

Proof : By symmetry, it suffices to prove that Id is continuous and maps bounded

sets to bounded sets. Let xn → x with respect to the � cot �-norm. Then,
� (xn ) − Id(x)�′ = �xn − x�′ ≤ c2 �xn − x� → 0,
i.e., Id is continuous. Furthermore, let A ∈ Br (0), i.e., for all x ∈ A,
�x� < r.
Then,
�x�′ ≤ c2 �x� < c2 r,
i.e.,
Id(A) ⊂ Bc′ 2 r (0).
n —14h(2019)—

Theorem 1.82 All the norms on Rm are equivalent.

Proof : Since the equivalence of norms is an equivalence relation, it suffices to

show that all the norms are equivalent to the Euclidean norm. Let � ⋅ � be a norm
on Rm . Every vector x ∈ Rm can be written as
m
x = � xi ei ,
i=1

where ei is the i-th unit vector. Using the triangle inequality and the Cauchy-
Schwarz inequality,
m m 1�2 m 1�2 m 1�2
�x� ≤ � �xi ��ei � ≤ �� xi � �� �ei � �
2 2
≤ �� �ei � �
2
�x�2 .
i=1 i=1 i=1 i=1

In the other direction, write

x
�x� = � � �x�2 ≥ (inf{�y� ∶ �y�2 = 1}) �x�2 .
�x�2
We will be done if we prove that inf{�y� ∶ �y�2 = 1} > 0. Since the function
x � �x� is continuous (every norm is continuous) and since the unit sphere with
respect of the Euclidean norm is compact, the infimum attains a minimum. This
minimum cannot be zero by the positivity of the norm. n
62 Chapter 1

Corollary 1.83 Let � ⋅ � be a norm on Rm . Then, all the closed and bounded sets
are compact.

Proof : Let A ⊂ Rm be closed and bounded. Since the identity map is a homeo-
morphism from (Rm , � ⋅ �) to (Rm , � ⋅ �2 ) mapping bounded sets to bounded sets, it
follows that Id(A) = A is closed and bounded in (Rm , � ⋅ �2 ), hence compact. Fi-
nally, using again the fact that the identity map is a homeomorphism, A is compact
in (Rm , � ⋅ �). n

Proposition 1.84 Let (V, � ⋅ �) be a normed space of finite dimension m. Then, it

is isometric to Rm endowed with some norm.

Proof : Let ai be a basis for V and consider the linear map T ∶ V → Rm , given by
m
T �� xi ai � = (x1 , . . . , xm ).
i=1

This map is invertible and in fact a homeomorphism. For x ∈ Rm define

�x�m = �T −1 (x)�.

Clearly, �x�m ≥ 0 with equality if and only if T −1 (x) = x = 0. For homogeneity,

�↵x�m = �T −1 (↵x)� = �↵ T −1 (x)� = �↵��T −1 (x)� = �↵��x�m .

Finally, for x, y ∈ Rm ,

�x + y�m = �T −1 (x + y)� = �T −1 (x) + T −1 (y)� ≤ �T −1 (x)� + �T −1 (y)� = �x�m + �y�m .

It follows that � ⋅ �m is a norm on Rm . By construction, T is an isometry, since for

every u, v ∈ V,

�u − v� = �T −1 T (u − v)� = �T (u − v)�m = �T u − T v�m .

n
 Metric Spaces 63

Corollary 1.85 Let V be a finite-dimensional vector space and let � ⋅ � and � ⋅ �′ be

norms on V. Then they are equivalent.

Proof : It follows from the fact that (V, � ⋅ �) are (V, � ⋅ �′ ) are isometric to Rm
endowed with two norms and that every two norms on Rm are equivalent. n

Corollary 1.86 Let (V, � ⋅ �) be a finite-dimensional normed space. Then, all the
closed and bounded sets are compact.

Proof : It follows from the fact that (V, � ⋅ �) is isometric to a normed space in
which this property holds. n

Example: If V is an infinite-dimensional vector space, then, not every two norms

on V are equivalent. Take for example the vector space C([0, 1]) and consider the
sequence of elements fn (x) = xn . Then,

lim � fn − 0�1 = lim � �xn − 0� dx = 0,

1

n→∞ n→∞ 0

however
lim � fn − 0�∞ = lim sup �xn − 0� = 1,
n→∞ n→∞ 0≤x≤1

so the sequence converges to zero in one norm and not in the other, implying that
the norms are not equivalent. ▲▲▲

Example: If (V, � ⋅ �) is an infinite-dimensional normed space, then, closed and

bounded sets are not necessarily compact. Take for example the unit ball in
(C[0, 1], � ⋅ �∞ ). For every n ∈ N denote
In = [1�(n + 1), 1�n],
and let fn be a sequence of non-negative continuous functions satisfying � fn �∞ =
1, vanishing outside In and assuming the value 1 is an interior point of In . Then
for every n ≠ n,
� fn − fm �∞ = 1,
i.e., fn is a 1-discrete sequence, hence does not have a converging subsequence.
▲▲▲
64 Chapter 1

1.4.6 Convergence of sequences of functions

This section deals with sequences of functions between metric spaces.

Definition 1.87 (Pointwise and uniform convergence) A sequence fn of func-

tions from (X, d) to (Y, ⇢) is said to converge pointwise (;*;$&81 ;&21,;/) to
a limit function f if for every x ∈ X,

lim fn (x) = f (x),

n→∞

i.e.,
∀x ∈ X ∀" > 0 ∃N ∈ N ∶ ∀n > N ⇢( fn (x), f (x)) < ".
It is said to converge uniformly (%&&: %$*/" ;&21,;/) if N can be chosen inde-
pendently of x; that is,

∀" > 0 ∃N ∈ N ∶ ∀n > N ∀x ∈ X ⇢( fn (x), f (x)) < ".

Comment: Pointwise convergence amounts to

∀x ∈ X lim ⇢( fn (x), f (x)) = 0,
n→∞

whereas uniform convergence amounts to

lim sup ⇢( fn (x), f (x)) = 0.

n→∞ x∈X

Proposition 1.88 (Cauchy criterion for uniform convergence) Let (X, d) be a

metric spaces and let fn ∶ X → R be a sequence of functions (not necessarily
continuous) satisfying the following condition:

∀" > 0 ∃N ∈ N such that ∀m, n > N sup � fn (x) − fm (x)� < ".
x∈X

Then, there exists a function f ∶ X → R, such that fn → f uniformly.

 Metric Spaces 65

Proof : We start by showing that the sequence fn converges pointwise. Fix x.

Then,

∀" > 0 ∃N ∈ N such that ∀m, n > N � fn (x) − fm (x)� < ".

It follows from Cauchy’s criterion for sequences in R that the sequence fn (x) has
a limit, which we denote by f (x).
It remains to prove that the convergence is uniform. Given " > 0, there exists an
N ∈ N such that for every m, n > N and for every x ∈ X,

� fn (x) − fm (x)� < .

"
2
Thus, for every x ∈ X and m, n > N,

� fn (x) − f (x)� ≤ � fn (x) − fm (x)� + � fm (x) − f (x)�,

from which follows, letting m → ∞, that

� fn (x) − f (x)� ≤ < ".

"
2
Thus,
∀" > 0 ∃N ∈ N ∶ ∀n > N ∀x ∈ X � fn (x) − f (x)� < ",
which completes the proof. n

Examples:

(a) The functions between the discrete spaces N and {0, 1} defined by fn ( j) =
j,n converge to the constant function f ∶ j � 0 pointwise but not uniformly.
In fact, one can show that a sequence of functions fn into a discrete space
converges uniformly to f if and only if fn coincides with f eventually.
(b) The sequence of functions fn ∶ [0, 1) → [0, 1) (with the standard metric)
defined by fn (x) = xn converges to f (x) = 0, pointwise, but not uniformly,
because for every n,

sup � fn (x) − f (x)� = sup �xn − 0� = 1.

0≤x<1 0≤x<1
66 Chapter 1

(c) Consider the following sequence of functions fn ∶ R → R,

�
� 0 < x < 1�n
�
�
�
n2 x
fn (x) = �2n − n2 x 1�n < x < 2�n
�
�
�
�
�0 otherwise.

It converges pointwise to f = 0, but not only it does not converges uniformly,

we even have
lim sup � fn (x) − f (x)� = lim n = ∞.
n→∞ x∈R n→∞

. Exercise 1.42 Let (X, d) be a metric space and let Y be an non-empty set
endowed with the discrete metric. Suppose that a sequence of functions fn ∶ X → Y
converges to a function f ∶ X → Y uniformly. Show that there exists an N ∈ N
such that for all n > N, fn ≡ f .

Consider the sequence of functions fn (x) = xn on the closed interval [0, 1]. Al-
though each of the fn is continuous, this sequence converges pointwise to the
�
discontinuous function
�
�0 0 ≤ x < 1
f (x) = �
�
�1 x = 1.
�
The following theorem shows that this can’t happen if the convergence is uniform.

Theorem 1.89 Let fn be a sequence of continuous functions from (X, d) to (Y, ⇢),
converging uniformly to a limit f ; then, the limit is continuous. Moreover, if all
the fn are uniformly continuous on X, then f is uniformly continuous on X.

Proof : Fix a point x ∈ X and let xm be a sequence converging to x. For every pair
of indexes n and m,

⇢( f (xm ), f (x)) ≤ ⇢( f (xm ), fn (xm )) + ⇢( fn (xm ), fn (x)) + ⇢( fn (x), f (x))

≤ 2 sup ⇢( f (y), fn (y)) + ⇢( fn (xm ), fn (x)).
y∈X

Let " > 0 be given. Because the fn converge to f uniformly, we can choose n
sufficiently large such that the first term is less than "�2. Having fixed n, there
exists an N such that for all m > N, the second term is less than "�2 (since fn is
 Metric Spaces 67

continuous). To conclude, for all " > 0 there exists an N such that for all m > N,
⇢( f (xm ), f (x)) < ", i.e., f is continuous.
Assume next that the fn are uniformly continuous. For x, y ∈ X we write

⇢( f (x), f (y)) ≤ ⇢( f (x), fn (x)) + ⇢( fn (x), fn (y)) + ⇢( fn (y), f (y))

≤ sup ⇢( f (z), fn (z)) + ⇢( fn (x), fn (y)).
z∈X

Since fn converges uniformly to f , given " > 0, there exists an n such that the first
term is less than "�2. Since this fn is uniformly continuous, there exists a > 0
such that d(x, y) < implies ⇢( fn (x), fn (y)) < "�2, i.e., implies ⇢( f (x), f (y)) < ".
n
The importance of uniform convergence is exemplified in the following proposi-
tion:

Proposition 1.90 Let (X, d) and (Y, ⇢) be metric spaces and let fn , f ∶ X → Y. If
fn are continuous and converge uniformly to f , then for every sequence xn in X,

xn → x implies fn (xn ) → f (x).

Example: A non-example: the sequence fn (x) = xn defined on [0, 1] converges

pointwise to the function
�
�
�0 0≤x<1
f (x) = �
�
� x = 1.
�1
Consider the sequence xn = 1 − 1�n, which converges to 1. Then,
1 n 1
lim fn (xn ) = �1 − � = ≠ f (1).
n→∞ n e
▲▲▲

Proof : We have

⇢( fn (xn ), f (x)) ≤ ⇢( fn (xn ), f (xn )) + ⇢( f (xn ), f (x))

≤ sup ⇢( fn (y), f (y)) + ⇢( f (xn ), f (x)).
y∈X
68 Chapter 1

The first term on the right hand side converges to zero by the uniform convergence
of fn , whereas the second term converges to zero by the continuity of f (which
results from the previous proposition). n —16h(2019)—

Definition 1.91 Let fn be a sequence of functions from a metric space (X, d) into
R (in fact, we only need the target to be a normed space). The series (9&)) ∑∞
n=1 fn
is said to converge uniformly to S if the sequence of partial sums,

n
S n = � fk
k=1

converges uniformly to S .

Theorem 1.92 (Weierstraß’ M-test) Let fn be a sequence of functions from a

metric space (X, d) into R. Suppose that there exists a sequence of real numbers
Mn such that
sup � fn (x)� ≤ Mn ,
x∈X

∑n=1 Mn converges. Then, the series ∑∞

∞
and that the series n=1 fn converges uni-
formly on X. In particular, if the fn are continuous, so is the series.

Example: The figure below shows the functions

n
sin kx
S n (x) = �
k=1 k
1.1

for n = 2, 20, 200, 2000. By the Weierstraß M-test, the sequence S n converges
uniformly on [0, 2⇡]. Moreover, since each of the partial sums is a uniformly
continuous function, the limit is uniformly continuous.
 Metric Spaces 69

1.5 1.5

1 1

0.5 0.5

0 0

−0.5 −0.5

−1 −1

−1.5 −1.5
0 2 4 6 0 2 4 6
n=2 n=20

1.5 1.5

1 1

0.5 0.5

0 0

−0.5 −0.5

−1 −1

−1.5 −1.5
0 2 4 6 0 2 4 6
n=200 n=2000

▲▲▲

Proof : For every m > n,

m m
sup �S m (x) − S n (x)� = sup � � fk (x)� ≤ sup � � fk (x)�
x∈X x∈X k=n+1 x∈X k=n+1
m ∞
≤ � Mk ≤ � Mk .
k=n+1 k=n+1

Since the right-hand side is the tail of a convergent series, there exists for every
" > 0 an N ∈ N, such that for all m, n > N,

sup �S m (x) − S n (x)� < ",

x∈X

hence S n converges uniformly by Cauchy’s criterion. n

70 Chapter 1

Proposition 1.93 Let fn ∶ [a, b] → R be a sequence of continuous functions con-

verging uniformly on [a, b] to a limit f . Then,
b
fn (x) dx = �
b
lim � f (x) dx.
n→∞ a a

Comment: We restrict ourselves to functions R → R because we have so far only

learned about integrals of real-valued univariate functions.

Proof : The integrals exists because the fn and their limit f are continuous (The-
orem 1.89), hence Riemann integrable. By the linearity of the integral, for every
n,

�� f (x) dx − � fn (x) dx� = �� ( f (x) − fn (x)) dx�

b b b

a a a

≤�
b
� f (x) − fn (x)� dx
a

≤�
b
sup � f (y) − fn (y)� dx
a a≤y≤b

= (b − a) sup � f (y) − fn (y)�.

a≤y≤b

Since fn converges to f uniformly, the right hand side tends to zero. n

Theorem 1.89 provides a criterion for the limit of a sequence of continuous func-
tions to be continuous (respectively, uniformly continuous). What about di↵er-
entiability? Under what conditions is the limit of a sequence of di↵erentiable
functions di↵erentiable?

Theorem 1.94 Let fn ∶ (a, b) → R be a sequence of di↵erentiable functions such

that (i) the sequence fn (x) converges at at least one point c ∈ (a, b), and (ii) the
sequence of derivatives, fn′ converges uniformly on (a, b) to a limit function g.
Then,

1. The sequence fn converges uniformly on (a, b) to a limit f .

2. f is di↵erentiable with f ′ = g.
 Metric Spaces 71

Comments:

(a) We need the first condition, because the non-convergent sequence fn (x) = n
satisfies the second condition with g = 0.
(b) Since we have not assumed the fn to be continuously di↵erentiable, we can’t
assume g to be continuous, and not even integrable. This prevents us from
defining f (x) = ∫a g(y) dy, and then show that fn converges to f uniformly.
x

This will “cost us” in technical complications. Providing a simpler proof

for the case where the fn are continuously di↵erentiable is left as an exercise
(see below).

Proof : We first prove the uniform convergence of fn based on Cauchy’s criterion.

For every m, n, we use the mean-value theorem to obtain

� fm (x) − fn (x)� ≤ �( fm − fn )(x) − ( fm − fn )(c)� + � fm (c) − fn (c)�

= �x − c��( fm′ − fn′ )(⇠mn )� + � fm (c) − fn (c)�,

for some ⇠mn between x and c. Hence,

sup � fm (x) − fn (x)� ≤ (b − a) sup � fm′ (⇠) − fn′ (⇠)� + � fm (c) − fn (c)�.
x ⇠

Since fn′ converges uniformly on (a, b), given " > 0, there exists an N1 ∈ N such
that for all m, n > N1 ,

sup � fm′ (⇠) − fn′ (⇠)� <

"
2(b − a)
.
⇠

Likewise, since fn (c) converges, there exists an N2 ∈ N such that for all m, n > N2 ,

� fm (c) − fn (c)� < .

"
2
Thus, for all m, n > max(N1 , N2 ),

sup � fm (x) − fn (x)� < ",

x

i.e., fn satisfies Cauchy’s criterion for uniform convergence; we denote the limit
by f . Since the fn are di↵erentiable, and in particular continuous, f is continuous.
72 Chapter 1

It remains to show that f ′ = g. Let x be an arbitrary point, and set

�
�
� 1h [ f (x + h) − f (x)] h ≠ 0
'(h) = �
�
� h = 0,
�g(x)
which is defined on some neighborhood I of zero. We will be done if we prove
that ' is continuous at zero.
How can we show that a function is continuous? For example, by showing that it
is the uniform limit of a sequence of continuous functions. Define the sequence
of functions on I,
�
�
� 1 [ fn (x + h) − fn (x)] h ≠ 0
'n (h) = � h ′
�
� fn (x)
� h = 0.
These functions are, by definition, continuous at zero, and we already know that
they converge pointwise to '. To show that they converge uniformly, we show that
they satisfy Cauchy’s criterion: if h ≠ 0 then
1
�'n (h) − 'm (h)� = �( fn − fm )(x + h) − ( fn − fm )(x)�
h
≤ sup � fn′ (⇠) − fm′ (⇠)�,
⇠∈X

whereas at h = 0,
�'n (0) − 'm (0)� = � fn′ (0) − fm′ (0)� ≤ sup � fn′ (⇠) − fm′ (⇠)�.
⇠∈X

Thus, given " > 0 there exists an N ∈ N such that for all m, n > N,
sup �'n (h) − 'm (h)� ≤ sup � fn′ (⇠) − fm′ (⇠)� < ",
h ⇠∈X

proving that 'n satisfies Cauchy’s criterion, hence converges to ' uniformly. This
concludes the proof. n

. Exercise 1.43 Let fn ∶ (a, b) → R be a sequence of continuously di↵erentiable

functions such that (i) the sequence fn (x) converges at at least one point c ∈ (a, b),
and (ii) the sequence of derivatives, fn′ converges uniformly on (a, b) to a limit
function g. Prove that

lim ( fn (x) − fn (c)) = lim �

x
fn′ (y) dy = �
x
g(y) dy.
n→∞ n→∞ c c

Conclude then that fn has a limit, which we denote by f , and that f ′ = g.

 Metric Spaces 73

Proposition 1.93 and Theorem 1.94 can be reformulated in terms of convergent

series:

Corollary 1.95 (1) If the series of continuous real-valued functions, ∑∞

n=1 fn , con-
verges uniformly on an interval [a, b], then
b ∞ ∞
�� fn (x)� dx = � �
b
�a fn (x) dx.
n=1 n=1 a

(2) Let fn be a sequence of di↵erentiable functions on (a, b) such that the series
of derivatives converges uniformly on (a, b), and there exists a point c ∈ (a, b) at
which the series ∑∞n=1 fn converges, then the series converges uniformly on (a, b)
and ′
∞ ∞
�� fn (x)� = � fn′ (x).
n=1 n=1

We have seen that the uniform convergence of continuous functions implies a

continuous limit. What about the other direction? Does the fact that a (pointwise)
converging sequence of continuous functions have a continuous limit imply that
the convergence is uniform? The following theorem show that the answer may be
affirmative under certain additional conditions.

Theorem 1.96 (Dini) Let fn be a sequence of continuous real-valued functions

defined on a compact metric space (X, d), and converging pointwise to a function
f , which is also continuous. If for every x ∈ X the sequence fn (x) is monotone,
then fn converges to f uniformly.

Comment: There is no requirement that fn (x) and fn (y) be monotone in the same
direction.

. Exercise 1.44 Find a non-example to Dini’s theorem where the monotonicity

assumption does not hold.
74 Chapter 1

Proof : Suppose, by contradiction, that the convergence was not uniform. Then, it
is not true that

∀" > 0 ∃k ∈ N ∶ ∀n > k ∀x ∈ X � fn (x) − f (x)� < ",

i.e.,
∃" > 0 ∶ ∀k ∈ N ∃nk > k ∃xk ∈ X ∶ � fnk (xk ) − f (xk )� ≥ ",
Since the sequence fn (x) is monotone for every x, it follows that

� fm (xk ) − f (xk )� ≥ � fnk (xk ) − f (xk )� ≥ " ∀m ≤ nk .

Putting it together,

∃" > 0 ∶ ∀k ∈ N ∃nk > k ∃xk ∈ X ∶ ∀m ≤ nk � fm (xk ) − f (xk )� ≥ ",

We can state it di↵erently,

∃" > 0 ∶ ∀k ∈ N ∃nk > k ∃xk ∈ X ∶ ∀m � fm (xk )− f (xk )� ≥ " for large enough k.

Since X is compact, the sequence xk has a converging subsequence xkl with limit
x. Since fm and f are continuous, the limit l → ∞ may be taken, yielding

∃" > 0 ∶ ∀m ∈ N � fm (x) − f (x)� ≥ ".

This violate the assumption that fm converges to f pointwise. n

Proof : Here is another proof that uses the “covering version” of compactness.
Since the sequence fn converges pointwise and monotonically to f , then there
exists for every " > 0 and every point x a number N(x, "), such that

� fn (x) − f (x)� < ∀n > N(x, ").

"
3
Moreover, since both fn and f are continuous, there exists a (x, ") > 0 such that

� f (y) − f (x)� < � fN(x,") (y) − fN(x,") (x)� < ∀y ∈ B (x,") (x).
" "
and
3 3
The union of all B (x,") (x) is an open covering of X. Since X is compact, there
exists a finite sub-covering,
m
X = �B (xi ,") (xi ).
i=1
 Metric Spaces 75

Let now N = maxN(xi ,") (here we exploit the finiteness). Since every y ∈ X is
contained in one of these balls, say the k-th ball, then for all n > N (here we
exploit the monotonicity),

� fn (y) − f (y)� ≤ � fn (y) − fn (xk )� + � fn (xk ) − f (xk )� + � f (xk ) − f (y)� < ",

which prove uniform convergence. n

Corollary 1.97 Let fn be a sequence of non-negative, real-valued function defined

on a compact metric space (X, d). If the sequence ∑∞ n=1 fn converges pointwise
and its limit is continuous, then it converges uniformly on X.

1.4.7 The Arzela-Ascoli theorem

For a compact metric space K, the space C(K) of real-valued functions over K is
a normed space with respect to

� f � = max � f (x)�,
x∈K

where the maximum exists by Proposition 1.70. Convergence in this space is

uniform convergence, as

d( f, g) = � f − g� = max � f (x) − g(x)�.

x∈K

We have already seen that closed and bounded sets in C(K) are not necessarily
compact. It turns out that if closedness and boundedness are supplemented with
another property, sets in C([0, 1]) are compact.

Definition 1.98 (Equicontinuity) Let K be a compact metric space and let A be

a collection of functions in C(K). This collection is said to be equicontinuous
(%$*(! %$*/" %5*79) if

∀" > 0 ∃ > 0 ∶ ∀ f ∈ A ∀x, y ∶ d(x, y) < � f (x) − f (y)� < ".

That is, all the f ∈ A are uniformly continuous, and the same modulus of continuity
(") applies for all f ∈ A.
76 Chapter 1

Theorem 1.99 (Arzelà-Ascoli) A closed and bounded set A ⊂ C(K) is compact if

and only if it is equicontinuous.

Proof : Suppose that A equicontinuous. Since it is bounded, there exists an M

such that
A ⊂ B M (0),
i.e., for all f ∈ A, � f � < M, or

sup �max � f (x)�� < M.

f ∈A x∈K

Consider a sequence of functions fn ∈ A. We need to show that it has a subse-

quence converging uniformly.
Let xn be a sequence of points which is dense in K; such a sequence exists by
Proposition 1.67 (compact spaces are separable). The sequence of real numbers
fn (x1 ) is bounded, hence there exists a subsequence fn of fn converging at x1 .
(1)

Take then the sequence fn (x2 ); by the same argument, there exists a subsequence
(1)

(2) (1)
fn of fn converging at x2 (and by inheritance also at x1 ). We proceed induc-
(k)
tively, constructing a sequence of subsequences, fn , which for given k converge
at x1 , �, xk .
The following picture is helpful:

f1 f2 � fn �
� �
(1) (1) (1)
f1 f2 fn converges at x1
� �
(2) (2) (2)
f1 f2 fn converges at x1 , x2
� � � �
� �
(k) (k) (k)
f1 f2 fn converges at x1 , . . . , xk .

Every row is a sequence of functions which is (i) a subsequence of the previous

row, and (ii) converges at a finite number of points.
(n) (1)
Consider now the “diagonal” sequence fn . Since it is a subsequence of fn it
converges at x1 . By the same argument, it converges at every point xk , i.e., it
converges pointwise at a dense set of points.
 Metric Spaces 77

(n)
So far, we have not used equicontinuity. We will use it now to show that fn
converges uniformly on K. Let " > 0 be given, and take > 0 according to the
definition of equicontinuity. Consider the collection of open sets

U = {B (xk ) ∶ k ∈ N}.

Since xk is dense in K, U is an open cover of K (every point in K it at a distance

less than from one of the points xk ). Since K is compact, there exists a finite
sub-cover,
K ⊂ {B (xk j ) ∶ j = 1, . . . , p}.

As the sequence fn (xk ) converges for all k, there exists for each j an N j , such
(n)

that for every m, n > N j ,

� fn (xk j ) − fm (xk j )� < ".

(n) (m)

Finally, as there are only finitely many N j , we may set

N = max N j .
p
j=1

Let x ∈ K. Since the finite sets of open balls cover K, there exists a j such that

x ∈ B (xk j ).

For every m, n > N,

� fn (x) − fm (x)� ≤ � fn (x) − fn (xk j )� + � fn (xk j ) − fm (xk j )�

(n) (m) (n) (n) (n) (m)

+ � fm (xk j ) − fm (x)�
(m) (m)

< " + " + ",

where in the first and the third term we used the definition of . It follows that for
every " > 0 there exists an N such that for all m, n > N

� fn − fm �∞ < ".
(n) (m)

(n)
By Cauchy’s criterion, the sequence fn converges uniformly on K (and its limit
is continuous). Finally, since A is closed, the limit is in A, hence A is compact.
We leave the other direction as an exercise. n
78 Chapter 1

. Exercise 1.45 Let K ⊂ X be compact. Show that if a set A ⊂ C(K) is compact,

then it is equicontinuous.

Comment: Note that we used the closedness only in the last step. Without clos-
enedess, we still have that if A ⊂ C(K) is bounded and equicontinuous, then every
sequence in A has a subsequence converging uniformly; the limit is in C(K) but
not necessarily in A.

Comment: A classical application of the Arzela-Ascoli theorem is the local exis-

tence proof of solutions of ordinary di↵erential equations.

Example: Consider a sequence of di↵erentiable functions fn ∶ [a, b] → R, that

are uniformly bounded, � f �∞ ≤ M1 and have a uniformly bounded derivative,
� fn′ �∞ ≤ M2 . Then the sequence has a subsequence that converges uniformly.
This is because the set of di↵erentiable functions that satisfy those conditions is
bounded and equicontinuous, as

� fn (x) − fn (y)� ≤ M2 �x − y�.

▲▲▲

. Exercise 1.46 1. Let f and fn be continuous mappings from a compact

metric space X into a metric space Y. Show that if

xn → x ⇒ fn (xn ) → f (x)

for all x ∈ X and sequences xn ∈ X, then fn converges uniformly to f .

2. Show, by a counter-example, that this is not true if X is not compact.

. Exercise 1.47 Let fn be a sequence of real-valued functions uniformly bounded

(i.e., fn (x) ≤ M) and Riemann integrable on [a, b]. Define for x ∈ [a, b]:

Fn (x) = �
x
fn (y) dy.
a

Show that the sequence (Fn ) has a subsequence that converges uniformly on
[a, b].
 Metric Spaces 79

1.4.8 The cantor set and the Cantor function

TA material 1.3

1.5 Completeness
1.5.1 Definitions and properties
Definition 1.100 (Cauchy sequence) Let (X, d) be a metric space. A sequence
of points xn in X is called a Cauchy sequence (*:&8 ;9$2) if there exists for every
" > 0 an N such that d(xn , xm ) < " for all m, n > N.

Comment: A converging sequence is always a Cauchy sequence; if xn → x, then

there exists for every " > 0 an N such that for every n > N, d(xn , x) < "�2. It
follows that for all m, n > N,

d(xn , xm ) ≤ d(xn , x) + d(xm , x) < ".

On the other hand, a Cauchy sequence does not necessarily converge. Take the
sequence xn = 1�n in (0, 1]. Then,

d(xn , xm ) < 1� min(m, n) → 0,

but the sequence has no limit in this space.

The following proposition should be familiar from your first calculus class:

Proposition 1.101 If a Cauchy sequence has a converging subsequence then the

whole sequence converges.

Proof : Suppose that the Cauchy sequence xn has a subsequence xnk that converges
to x. Let " > 0 be given. By the Cauchy property,

∃N ∶ ∀m, n > N d(xm , xn ) < "�2,

and by the definition of the limit,

∃K ∶ ∀k > K d(xnk , x) < "�2.

 80 Chapter 1

Let k be such that k > K and nk > N. Then, for every n > N,
d(xn , x) ≤ d(xn , xnk ) + d(xnk , x) < "�2 + "�2.
n

Definition 1.102 (Complete metric space) A metric space is called complete (.-:)
if every Cauchy sequence in X converges.
—18h(2019)—

Examples:
(a) R is a complete metric space (elementary calculus course).
√ metric space because the sequence of finite decimal
(b) Q is not a complete
approximations of 2 is a Cauchy sequence but it has no limit in Q.
(c) Completeness should not be confused with closedness. The metric space
(Q, � ⋅ �) is closed (as every metric space).
(d) If (X, d) is a complete metric space and Y ⊂ X, then (Y, d) is a complete
metric space if and only if Y is closed in X.
(e) Every compact metric space is complete (a consequence of Proposition 1.101).
(f) A complete normed space is called a Banach space; a complete inner prod-
uct space is called a Hilbert space.

Proposition 1.103 If K is a compact metric space then C(K) is a complete metric

space.

Proof : We have already seen that C(K) is a normed space, hence a metric space.
Let fn be a Cauchy sequence in C(K), i.e.,
∀" > 0 ∃N ∶ ∀m, n > N sup � fn (x) − fm (x)� < ".
x∈K

This is precisely Cauchy’s criterion for uniform convergence, that is, fn converges
uniformly, and its limit is in C(K). n

Comment: Completeness is not a topological invariant: it is not conserved under

a homeomorphism, as shows the example of f (x) = tan x which is a homeomor-
phism between the incomplete space (−⇡�2, ⇡�2) and the complete space R. In
fact, the Cauchy property of a sequence is not a topological invariant.
 Metric Spaces 81

. Exercise 1.48 Prove that a metric space (X, d) is complete if and only if for
every decreasing sequence of closed balls,

B̂r1 (x1 ) ⊃ B̂r2 (x2 ) ⊃ ⋅ ⋅ ⋅ ⊃ B̂rn (xn ) ⊃ . . . ,

such that rn → 0, we have

∞
� B̂rn (xn ) ≠ �.
n=1

. Exercise 1.49 This exercise shows that the condition rn → 0 in the previous
exercise was necessary. Let X = N, and define
�
�
�0 m=n
d(m, n) = �
�
�1 + m+n m ≠ n.
.
�
1

1. Show that d is a metric on X.

2. Show that X is complete with respect to this metric.
3. Show that every point in X is isolated, i.e., for every n ∈ N there exists an rn
such that Brn (xn ) = {xn }.
4. Show that the sequence of closed balls Bn = B̂1+1�2n (n) is decreasing, i.e.,
Bn+1 ⊂ Bn , and yet their countable intersection is empty.

. Exercise 1.50 Let (X, d) be a metric space and D ⊂ X a dense set. Show that
X is complete if and only if every Cauchy sequence in D converges in X.

. Exercise 1.51 Show that if X and Y are metric spaces, Y is complete, D ⊂ X is

a dense subset of X, and f ; D → Y is a uniformly continuous function, then f has
a unique extension which is a function on the whole X. That is, that there exists
a unique continuous function, f˜ ∶ X → Y, such that f˜�D = f . Show, furthermore,
that the uniform continuity of f is necessary.

. Exercise 1.52 Prove that the following spaces are Banach spaces:

1. The space `∞ of infinite sequences with the norm

�x�∞ = sup �xn �.

n
82 Chapter 1

2. The space `1 of infinite sequences with the norm

∞
�x�1 = � �xn �.
n=1

TA material 1.4 Every non-complete metric space has a canonical completion

(just as Q has a canonical completion yielding R). We formulate this as a theorem:

Theorem 1.104 (Completion of metric spaces) Let (X, d) be a (non-complete)

metric space. Then there exists a complete metric space (X̂, D), such that

(a) There exists a one-to-one map ◆ ∶ X → X̂, which is distance preserving.

(b) The image of ◆ is dense in X̂.

A metric space (X̂, D) satisfying those conditions is called a completion (%/-:%)

of (X, d). Finally, the completion is unique modulo an isomorphism; that is, if
(Y, ⇢) is a completion of (X, d), then (Y, ⇢) and (X̂, D) are isometric.

Proof : The first task is to identify the natural candidate for the completion (X̂, D).
Since the “holes” in (X, d) are detected by Cauchy sequences that do not converge,
we would like to define X̂ to be the “set of all limits of Cauchy sequences in X”.
The problem of course is there there are no such limits... Thus, the only plausible
candidates for the elements of the completion would be the Cauchy sequences
themselves.
So let X̂ be temporarily the set of all Cauchy sequences in X. How would we then
define the distance between two Cauchy sequences (xn ) and (yn )? The natural
choice is
D((xn ), (yn )) = lim d(xn , yn ).
n→∞

We need to verify that the limit exists. Note that

�d(xn , yn ) − d(xm , ym )� ≤ �d(xn , yn ) − d(xn , ym )� + �d(xn , ym ) − d(xm , ym )�

≤ d(yn , ym ) + d(xn , xm ).

Since both (xn ) and (yn ) are Cauchy sequence, the sequence of real numbers
(d(xn , yn )) is a Cauchy sequence, hence converges.
 Metric Spaces 83

The nest step is to show that D is a metric on the space of Cauchy sequences.
Symmetry is obvious. For the triangle inequality, let (xn ), (yn ) and (zn ) be Cauchy
sequences in X, then

D((xn ), (yn )) = lim d(xn , yn )

n→∞
≤ lim d(xn , zn ) + lim d(zn , yn )
n→∞ n→∞
= D((xn ), (zn )) + D((zn ), (yn )).

Remains positivity; D is clearly non-negative and D((xn ), (xn )) = 0. It is however

not strictly positive. Take any two di↵erent sequence (xn ) and (yn ) converging to
the same limit a. Both are Cauchy sequences, however

D((xn ), (yn )) = lim d(xn , yn ) ≤ lim d(xn , a) + lim d(a, yn ) = 0.

n→∞ n→∞ n→∞

It follows that D is a pseudo-metric on the space of Cauchy sequences.

On the first week of this course we saw that every pseudo-metric space can be
turned into a metric space by defining an equivalence relation in X,

(xn ) ∼ (yn ) if and only if D((xn ), (yn )) = 0.

For a sequence (xn ), we denote its equivalence class by [(xn )]. Then, we redefine
X̂ to be the space of all equivalence classes of Cauchy sequences in X, endowed
with the metric
D([(xn )], [(yn )]) = lim d(xn , yn ).
n→∞

We have seen that this definition does not depend on the choice of representatives;
that is, if [(xn )] = [(zn )] and [(yn )] = [(wn )], then

lim d(xn , yn ) = lim d(zn , wn ).

n→∞ n→∞

We finally have a bone fide metric space which we will show to be a completion
of (X, d). For x ∈ X we define

◆(x) = [(x)],

where [(x)] stands for the equivalence class of the constant sequence (x, x, . . . ).
This map is distance preserving as

D([(x)], [(y)]) = lim d(x, y) = d(x, y).

n→∞
84 Chapter 1

In particular, it is one-to-one. It remains to prove that its image is dense in (X̂, D).
Let [(xn )] be an element of X̂, and consider the sequence of elements of X̂, ◆(xk )
(it is a sequence of equivalence classes of constant sequences). Then,

D([(xn )], ◆(xk )) = lim d(xn , xk ).

n→∞

Since (xn ) is a Cauchy sequence,

lim lim d(xn , xk ) = 0,

k→∞ n→∞

proving that the image of ◆ is dense in X̂.

The next step is to show that (X̂, D) is complete. Let [(xnk )] be a Cauchy sequence
(in k) of elements of X̂. That is, for every " > 0 there exists a K such that for every
k1 , k2 > K,
D([(xnk1 )], [(xnk2 )]) = lim d(xnk1 , xnk2 ) < ".
n→∞

Consider the sequence yn = xnn .

It is easy to see that it is a Cauchy sequence in X;
furthermore,
D([(xnk )], [(yn )]) = lim d(xnk , xnn ) ≤ ".
n→∞

This proves that

lim [(xnk )] = [(yn )],
k→∞

i.e., (X̂, D) is complete.

It remains to show that the completion is unique modulo isometries. Let (Y, ⇢) be
a complete metric space such that there exists a distance preserving map | ∶ X → Y
whose image is dense of Y. Consider the map

A ∶ X̂ ∋ Image(◆) → Image( |) ∈ Y

defined by A = | ○ ◆−1 . This is a distance preserving map from a dense set in X to

a dense set in Y. It can be extended to an isometry A ∶ X̂ → Y. Let’s show it with
simpler notations: for ↵ ∈ X̂ there exists a sequence ↵n ∈ Image(◆) converging to
↵. Since A is distance-preserving, A(↵n ) is a Cauchy sequence in Y, and since Y
is complete we may define

A(↵) = lim A(↵n ).

n→∞

A is onto Y as every element ∈ Y is a limit of elements A(↵n ). It follows that

A ∶ X̂ → Y is distance-preserving and onto, hence an isometry. n
 Metric Spaces 85

1.5.2 Contractive mapping theorem

Definition 1.105 Let X be a set and let f ∶ X → X. A point x ∈ X is called a fixed
point (;": ;$&81) of f if f (x) = x.

Definition 1.106 Let (X, d) be a metric space. A mapping f ∶ X → X is called

contractive (;7&&,/ %8;3%) if there exists a real number < 1 such that for all
x, y ∈ X
d( f (x), f (y)) ≤ d(x, y).

Lemma 1.107 A contractive mapping is continuous.

Proof : Let f ∶ X → X be contractive with parameter < 1. Let xn → x. Then,

lim d( f (xn ), f (x)) ≤ lim d(xn , x) = 0.

n→∞ n→∞

Theorem 1.108 (Contractive mapping theorem) Let (X, d) be a complete metric

space. If f ∶ X → X is contractive, then it has a unique fixed point x ∈ X.

Proof : The proof is constructive: let x0 be an arbitrary point in X and consider the
sequence xn defined inductively by

xn+1 = f (xn ).

We first show that xn is a Cauchy sequence. Note that

d(xn+1 , xn ) = d( f (xn ), f (xn−1 )) ≤ d(xn , xn−1 ),

and by induction, d(xn+1 , xn ) ≤ n d(x , x ).

1 0 For n > m,
n−1 m
d(xn , xm ) ≤ � k
d(x1 , x0 ) ≤ d(x1 , x0 ).
k=m 1−

Thus, given " > 0, choose N such that N �(1− ) d(x1 , x0 ) < " and for all n, m > N,

d(xn , xm ) < ".

86 Chapter 1

Since X is complete it follows that xn converges to a limit, which we denote x.

Taking the n → ∞ limit of the relation xn+1 = f (xn ), using the continuity of f , we
obtain that x is a fixed point of f .
Finally, if x, y are two distinct fixed points of f , then

d(x, y) = d( f (x), f (y)) ≤ d(x, y) < d(x, y),

which is a contradiction. n

Example: Let p > 1. What is the value of the following expression:

� � √
x= p+ p+ p + �.
√
By this expression, we mean the limit of the sequence x0 = p, xn+1 = p + xn .
To show that this limit exists, we use the contractive mapping theorem. First, we
√
claim that the mapping
x � p + x ≡ (x)

√ interval K = [0, 2p] into itself. Indeed, the mapping is monotonic, and
maps the
2p � 3p ≤ 2p. Thus, ∶ K → K, where K is a complete metric space.
We then show that this mapping is contractive: for x, y ∈ K,

(x) − (y) = ′
(⇠)(x − y),
′ (⇠)
√
where = 1�2 p + ⇠ < 1�2, and ⇠ is a point between x and y. It follows that
1
� (x) − (y)� ≤ �x − y�,
2

√
i.e., the mapping is contractive. Thus xn converges to the unique fixed point,
satisfying x = p + x, i.e.,
1 1�
x= + 1 + 4p.
2 2
▲▲▲

Example: Suppose we want to solve the nonlinear equation f (x) = 0, where

f ∶ R → R is continuous. Consider the following iterations,

xn+1 = xn − ↵ f (xn ) ≡ (xn ),

 Metric Spaces 87

with x0 chosen arbitrarily, and ↵ is to be determined. Then. for every x, y ∈ R,

� (x) − (y)� = �1 − ↵ f ′ (⇠)��x − y�,
where ⇠ is a point between x and y. If f ′ (x) has the property that 0 < 1 ≤ f ′ (x) ≤
2 < M, then we can choose ↵ = 2�M, so that

�1 − ↵ f ′ (⇠)� ≤ max (�1 − 2 1 �M�, �1 − 2 2 �M�) < 1,

so that the mapping is contractive and the iterations converge. ▲▲▲

1.5.3 Newton’s method for root finding

TA material 1.5 Let f ∶ R → R. In many applications, one is interested is com-
puting roots of f , i.e., points r ∈ R, where f (r) = 0. Unless f is particularly
simple, this cannot be done by analytical means. One resorts then to numerical
methods.
Suppose that f is di↵erentiable, and that x0 is “close” to a root r; furthermore,
assume that f ′ (x0 ) ≠ 0. The tangent to f at the point x0 is the line
P(x) = f (x0 ) + f ′ (x0 )(x − x0 ).
It seems natural to approximate the root of f by the root of P; the latter can be
calculated analytically,
f (x0 )
x = x0 − ′
f (x0 )
.

While nothing guarantees that x is a root of f , this suggested the following ap-
proach; start with some initial guess x0 , and then generate a sequence
f (x)
xn+1 = (xn ) where (x) = x −
f ′ (x)
with the hope that it converges to a root of f . This algorithm is called the Newton-
Raphson scheme.
88 Chapter 1

√
Example: Suppose that you want to evaluate R using just the four operations of
arithmetics. The goal then is to find a root of the function

f (x) = x2 − R.

The Newton-Raphson scheme is

x2 − R x R
xn+1 = (xn ) where (x) = x − = + .
2x 2 2x

Below we show the first eight iterates for R = 2 starting with

√ x0 = 1. The digits in
bold are correct. See how fast this sequence converges to 2.

x0 = 1
x1 = 1.5
x2 = 1.4166666666666666666666666666666666666666666666666666666666675
x3 = 1.4142156862745098039215686274509803921568627450980392156862745
x4 = 1.4142135623746899106262955788901349101165596221157440445849057
x5 = 1.4142135623730950488016896235025302436149819257761974284982890
x6 = 1.4142135623730950488016887242096980785696718753772340015610125
x7 = 1.4142135623730950488016887242096980785696718753769480731766796
▲▲▲

Theorem 1.109 Let f ∶ R → R be continuously twice di↵erentiable. Let r be a root

of f such that f ′ (r) ≠ 0. Then, there exists a > 0 such that the Newton-Raphson
scheme converges to r for every x0 ∈ B̂ (r).

Proof : We will use the contractive mapping theorem, showing that there exists a
. > 0 such that
∶ B̂ (r) → B̂ (r)
is contractive. From this will follows that xn converges to a fixed point r of , i.e.,
f (r)
r=r−
f ′ (r)
,

i.e., r is a root of f .
 Metric Spaces 89

Note that
f (x) f ′′ (x)
′
(x) = −
( f ′ (x))2
.

Since f (r) = 0, f ′ (r) ≠ 0 and f, f ′ , f ′′ are continuous, there exists a > 0, such
that � ′ (x)� < 1�2 for all x ∈ B̂ (r). For such x,

� (x) − r� = � (x) − (r)� ≤ sup � ′ (⇠)� �x − r� ≤ ,

⇠∈B̂ (r) 2

i.e., ∶ B̂ (r) → B̂ (r). Moreover, for x, y ∈ B̂ (r),

1
� (x) − (y)� = sup � ′ (⇠)� �x − y� ≤ �x − y�,
⇠∈B̂ (r) 2

proving that the mapping is contractive.

We can say even more: for every x ∈ B̂ (r) there exists a ⇠ ∈ (x, r) such that

f (⇠) − f (r) f ′′ (⇠)

� (x) − r� = � ′ (⇠)� �x − r� ≤ � �� ′ � �x − r�2 ,
⇠−r ( f (⇠))2

from which follows that

�xn+1 − r� f ′′ (r)
lim = � �.
n→∞ �xn − r�2 f ′ (r)

Thus, eventually, there exists an M such that

�xn+1 − r� ≤ M �xn − r�2

which explains why each error is of the order of the previous error squared (this
is called a quadratic convergence rate). n

1.5.4 Existence of solutions of ODEs

We will now use the contractive map theorem to prove the fundamental theorem
of existence of solutions to ordinary di↵erential equations. In fact, we shall prove
a simplified version of the general theorem in order to avoid some technical com-
plications.
90 Chapter 1

Theorem 1.110 (Picard, Existence and uniqueness of solutions to ODEs) Let

F ∶ R2 → R be continuous, bounded with constant M and Lipschitz continuous in
its second argument with constant L, i.e.,

�F(x, y)� ≤ M and �F(x, y1 ) − F(x, y2 )� ≤ L�y1 − y2 �.

Then, there exists for every (x0 , y0 ) ∈ R2 a > 0 and a unique function y ∈
C([x0 , x0 + ]) satisfying the di↵erential equation,

y′ (x) = F(x, y(x))

subject to the initial condition (%-(;% *!1;) y(x0 ) = y0 .

Proof : Set,
< min(1�M, 1�L),

and consider the set of functions

K = {y ∈ C([x0 , x0 + ]) ∶ �y − y0 �∞ ≤ 1},

which is a closed subset of C([x0 , x0 + ]) (hence complete in the subspace metric).

Consider now the mapping

∶ C([x0 , x0 + ]) → C([x0 , x0 + ]),

defined by
( (y))(x) = y0 + �
x
F(t, y(t)) dt.
x0

We first show that maps K into itself: if y ∈ K, then for every x ∈ [x0 , x0 + ],

�( ( ))(x) − y0 � = ��
x
F(t, y(t)) dt� ≤ �
x
�F(t, y(t))� ds ≤ M < 1,
x0 x0

i.e., (y) ∈ K.
 Metric Spaces 91

We next show that is contractive: let y, z ∈ K. Then, for every x ∈ [x0 , x0 + ],

�( (y))(x) − ( (z))(x)� = �� (F(t, y(t)) − F(t, z(t)) dt�

x

x0

≤�
x
�F(t, y(t)) − F(t, z(t))� dt
x0

≤ L�
x
�y(t) − z(t)� dt
x0
≤ L�y − z�∞ ,
Taking the supremum over all x,
� (y) − (z)�∞ ≤ (L )�y − z�∞ ,
proving that is indeed contractive.
It follows from the contractive mapping theorem that has a unique fixed point,
y ∈ K, i.e., for every x ∈ [x0 , x0 + ],

y(x) = y0 + �
x
F(t, y(t)) dt.
x0

It remains to show that such a y solves the di↵erential equation. Indeed, y is

di↵erentiable, and
y′ (x) = F(x, y(x)),
along with the initial condition y(x0 ) = y0 . n

Comment: Under the conditions of this version of Picard’s theorem, a solution for
the initial value problem can be obtained for all x ∈ R, by continuing the solution
over -intervals. —19h(2019)—

1.5.5 Baire’s category theorem (not taught in 2019)

The subject of this section is the classification of “large sets” in a metric space.
We have already seen the following definition:

Definition 1.111 Let (X, d) be a metric space. A set A ⊂ X is called dense

(%5&57) if every non-empty open set U ⊂ X intersects A. This condition is equiva-
lent to the condition that
A = X.
That is, every point in X is a limit of a sequence in A.
92 Chapter 1

We can say that a dense set is “almost everywhere”, but in many respects, dense
√
sets can be “small”. For example, two dense sets may have an empty intersection
(e.g., Q and 2+Q in R). The rationals and the irrationals are another example of
dense sets in R that do not intersect. It turns out that this could not have happened
if we had added the additional requirement that the sets be open.
First, a lemma:

Lemma 1.112 A metric space (X, d) is complete if and only if every decreasing
sequence of closed balls, B̂rn (xn ) with rn → 0 has a non-empty intersection.

Proof : Suppose first that X is complete and let

B̂r1 (x1 ) ⊃ B̂r2 (x2 ) ⊃ . . . ,

with rn → 0. Taking yn ∈ B̂rn (xn ) we obtain a Cauchy sequence. Since the space
is complete, the sequence converges, yn → y. Fix n; for every k ≥ n, yk ∈ B̂rn (xn );
since the balls are closed, y ∈ B̂rn (xn ) for every n.
Conversely, suppose that the non-empty intersection property holds. Let xn be a
Cauchy sequence. For every n ∈ N let

"n = sup d(xm , xn ).

m≥n

Since the sequence xn is a Cauchy sequence, it follows that "n → 0. Then, we can
take a subsequence nk of natural number, such that
∞
� "nk < ∞.
k=1

Define
rk = � "n j .
j≥k

Then, for every k,

B̂rk+1 (xnk+1 ) ⊂ B̂rk (xnk ).
This is because rk = rk+1 + "nk and d(xnk+1 , xnk ) ≤ "nk . By assumption, there exists
an ∞
x ∈ � B̂rk (xnk ).
n=1
 Metric Spaces 93

Since d(xnk , x) ≤ "nk → 0, x is a partial limit of the sequence xn . We have already

seen on several occasions that if a Cauchy sequence has a partial limit, then it
converges. n

Theorem 1.113 (Baire) Let X be a complete metric space. If Un ⊂ X is a sequence

of open and dense sets, then
∞
Y = � Un
n=1

is dense.

Proof : Let Br (x) be some open ball. We need to show that it intersects Y. The
idea is to construct a decreasing sequence of closed balls. Since U1 is dense, it
intersects Br (x) and the intersection is open. Hence, there exists a close ball,

B̂r1 (x1 ) ⊂ U1 ∩ Br (x).

Similarly, since U2 is dense, it intersects Br1 (x1 ) and the intersection is open.
Hence there exists a closed ball,

B̂r2 (x2 ) ⊂ Br1 (x1 ) ∩ U2 ⊂ U2 ∩ U1 ∩ Br (x).

We proceed inductively, constructing

B̂rn (xn ) ⊂ Brn−1 (xn−1 ) ∩ Un ⊂ Un ∩ ⋅ ⋅ ⋅ ∩ U1 ∩ Br (x).

Moreover, we may always let rn → 0. By the lemma, there exists a point in the
intersection of all those closed balls, which belongs to Y ∩ Br (x). n
It should be noted that a countable intersection of open sets is not necessarily
open. We have the following definition:

Definition 1.114 Let (X, d) be a metric space. A set Y ⊂ X is called a G -set if

it can be represented as a countable intersection of open sets.

Comment:

1. Every open set U is of class G as it is its own countable intersection.

94 Chapter 1

2. Every closed set C is of class G as

∞
C = � Un , Un = � B1�n (x).
n=1 x∈C

(Explain!)

3. There are G -sets that are neither open nor closed, for example

{−1} ∪ (0, 1) ⊂ R.

4. The requirement that the intersection be countable is crucial; every set can
be represented as an intersection of open sets,

A = �{z}c ,
z�∈A

so that this classification would be of no interest.

5. It is not yet clear that there exist sets that are not of class G .

Corollary 1.115 If (X, d) is a complete metric space, then the class of dense G -
sets is closed under countable intersections.

Proof : Let Yn be dense G -sets. Then, for each n there exist open Un,k , such that
∞
Yn = � Un,k .
k=1

Clearly, the Un,k must be dense. By Baire’s theorem,

∞ ∞ ∞
� Yn = � � Un,k ,
n=1 n=1 k=1

is dense, and by definition it is also of class G . n

Thus, the dense G -sets can be viewed as large sets, since even when intersected
countably many times with sets of that class, they remain dense.
 Metric Spaces 95

Corollary 1.116 Q is not of class G in R.

Proof : Since Q is countable, we can write its elements as a sequence (qn ). Then,
∞
Qc = � {qn }c ,
n=1

is both dense and of class G . Q is dense. If it were also of class G its intersection
with Qc must have been dense, but it is actually empty. n
We are finally in measure to characterize “large sets”.

Definition 1.117 Let (X, d) be a complete metric space. A set Y ⊂ X is called

residual (%1/:) if it contains a dense G -set.

Corollary 1.118 Let (X, d) be a complete metric space. Let R be the collection
of all residual sets. Then, R satisfies the following properties:

1. X ∈ R and � ∈� R.

2. If A ∈ R and A ⊂ B then B ∈ R.

3. R is closed under countable intersections.

The notion of “large sets” has a dual notion of “small sets”, which we now review
briefly:

Definition 1.119 Let (X, d) be a metric space.

1. A set Y ⊂ X is called nowhere dense (%-*-$) if its closure has an empty

interior (i.e., its closure does not contain open balls).

2. A set Y ⊂ X is called of class F" if it is a countable union of closed sets.

3. A set Y ⊂ X is called of the first category if it is a countable union of

nowhere dense sets. A set which is not of the first category is called of the
second category.
96 Chapter 1

Proposition 1.120 Y is nowhere dense if and only it its complement contains an

open dense set.

Proof :
Y is nowhere dense ⇐⇒ (Y)○ = �
⇐⇒ ((Y)○ )c = X
⇐⇒ (Y)c = X
⇐⇒ (Y c )○ = X
⇐⇒ (Y c )○ is dense
⇐⇒ Y c contains an open dense set.
n

Proposition 1.121 A is an F" -set if and only if Ac is a G -set.

Proof : Immediate from duality. n

Proposition 1.122 Y is of the first category if and only if Y c is residual.

Proof : Let Y be of first category. By definition,

∞
Y = � An ,
n=1

where An are nowhere dense. Hence,

∞
Y c = � Acn ,
n=1

where Acn contain an open dense set, i.e., Y c is residual.

Conversely, let Y c be residual, i.e., Y c contains a dense G -set,
∞
Y c ⊃ � Bn ,
n=1
 Metric Spaces 97

where the Bn are open and dense. It follows that

∞
Y ⊂ � Bcn ,
n=1

where the Bcn are nowhere dense, hence

∞
Y = � (Bcn ∩ Y)
n=1

is of the desired form. n

Proposition 1.123 Let (X, d) be a complete metric space and let M denote the
collection of sets of the first category. Then,

1. � ∈ M and X ∈� M.

2. If A ∈ M and B ⊂ A then A ∈ M.

3. M is closed under countable union.

1.5.6 Semicontinuity and the continuity of pointwise limits

Suppose that fn ∈ C([a, b]) converges pointwise to a function f . As we know,
the limit is not necessarily continuous. A question is “how bad” can the limit be?
We will that the set of points in which f is continuous is residual in [a, b]. To
prove it, we’ll first study the notion of semicontinuity (%7(/- ;&5*79), which is
interesting in its own.

Definition 1.124 Let (X, d) be a metric space. A function f ∶ X → R is called

lower-semicontinuous if
f −1 ((t, ∞)) is open
for all t ∈ R.

Proposition 1.125 Lower-semicontinuity is equivalent to either of the following

two characterizations:
98 Chapter 1

1. for every a ∈ X and " > 0 there exists a > 0 such that

f (x) > f (a) − " ∀x ∶ d(x, a) < .

2. For every sequence xn → a,

f (a) ≤ lim inf f (xn ).

n→∞

Proof : (i) Suppose that for every t,

{x ∈ X ∶ f (x) > t} is open.

Let a ∈ X. For every " > 0,

f (a) > f (a) − ".
Hence,
a ∈ {x ∈ X ∶ f (x) > f (a) − "}.
Since the latter is open, there exists a open ball B (a) such that for every x ∈ B (a),

f (x) > f (a) − ".

(ii) Suppose that the second condition holds and let xn → a. For every " > 0 and n
large enough,
f (xn ) > f (a) − ".
Letting n → ∞,
lim inf f (xn ) ≥ f (a) − ".
n→∞

Since this holds for every " > 0 we obtain the desired result.
(iii) Suppose that the sequential condition holds and consider the set

A = {x ∶ f (x) > t}.

Suppose that this set was not open. It would imply the existence of a ∈ A such that
for every n there exists an xn ∈ B1�n (a) for which

f (xn ) ≤ t.
 Metric Spaces 99

Then, xn → a and
lim inf f (xn ) ≤ t < f (a),
n→∞

which is a contradiction. n

Example: For X = R and a < b,

�
�
�1 x ∈ (a, b)
f (x) = �
�
�
�0 otherwise

is lower-semicontinuous, as easily verified by the sequential characterization, or

by noting that
�
� � 1<t
�
�
�
f ((0, t)) = �(a, b)
−1
0<t≤1
�
�
�
�
�R t ≤ 0.
▲▲▲

Theorem 1.126 Let (X, d) be a metric space and let f ∶ X → R be lower-

semicontinuous. Then, the set of points in which f is continuous is a dense G -set
(hence residual).

Proof : As a first step, suppose that f is bounded from above. Define

Cn = {a ∈ X ∶ ∃r > 0, � f (x) − f (y)� < 1�n ∀x, y ∈ Br (a)}.

Cn is an open set for all n, because suppose that a ∈ Cn and that for every k there
exists xk ∈ B1�k (a) which is not in Cn . This would imply that for every k there
exists sequences yk` , zk` ∈ B1�` (xk ) such that

� f (yk` ) − f (zk` )� ≥ 1�n.

Now, xk → a and yk` , zk` → xk . Consider the “diagonal” sequences ykk , zkk ; both
converges to a and
� f (ykk ) − f (zkk )� ≥ 1�n,
contradicting the fact that a ∈ Cn .
100 Chapter 1

Then,
∞
C ≡ � Cn = {a ∈ X ∶ ∀n ∃r > 0, � f (x) − f (y)� < 1�n ∀x, y ∈ Br (a)}
n=1
⊂ {a ∈ X ∶ ∀n ∃r > 0, � f (x) − f (a)� < 1�n ∀x ∈ Br (a)},
is the set of continuity points of f . Since Cn is open for every n, it suffices to show
that Cn is dense.
Let Br (a) be an arbitrary open ball; we need to show that it intersects Cn . Let

M = sup f (x),
x∈Br (a)

which is finite by assumption. Since f is lower-semicontinuous, the set

f −1 ((M − 1�n, ∞)) = {x ∈ X ∶ f (x) > M − 1�n}

is open. Moreover, by the definition of M as the supremum of f in Ba (r),

Br (a) ∩ f −1 ((M − 1�n, ∞)) ≠ �,

and the intersection is open. Hence, there exists an open ball

B⇢ (b) ⊂ Br (a) ∩ f −1 ((M − 1�n, ∞)).

For every x, y ∈ B⇢ (b),

M − 1�n < f (x), f (y) ≤ M,

hence b ∈ Cn , proving that Cn is dense.

Remains the case where f is not bounded. For this, we note that that (i) if f is
lower-semicontinuous and g is continuous, the g ○ f is lower-semicontinuous, and
(ii) if f is any function and g is continuous, then every continuity point of f is a
continuity point of g ○ f . Since tan−1 ∶ R → (−⇡�2, ⇡�2) is continuous,

F(x) = tan−1 ( f (x))

is bounded, lower-semicontinuous, and its continuity points coincides with those

of f . Since the continuity points of F are residual, so are the continuity points of
f. n
We now turn to consider pointwise limits of continuous functions. First a lemma:
 Metric Spaces 101

Lemma 1.127 Let (X, d) be a metric space. Let fn ∈ C(X) be a sequence of

continuous functions. Suppose that

f (x) = sup fn (x)

n

is finite for every x. Then, f is lower-semicontinuous.

Proof : For every t ∈ R,

{x ∶ f (x) > t} = {x ∶ sup fn (x) > t}

n
= {x ∶ ∃n, fn (x) > t}
∞
= � {x ∶ fn (x) > t},
n=1

and since fn are continuous, the latter is a union of open sets. n

Comment: The lemma would still hold if fn were only lower-semicontinuous and
the collection of functions was non-countable.

Comment: There exists a dual notion of upper-semicontinuity, which requires

f −1 ((−∞, t)) to be open for every t ∈ R. It can be shown similarly to above that

Proposition 1.128 Upper-semicontinuity is equivalent to either of the following

two characterizations:

1. for every a ∈ X and " > 0 there exists a > 0 such that

f (x) < f (a) + " ∀x ∶ d(x, a) < .

2. For every sequence xn → a,

f (a) ≥ lim sup f (xn ).

n→∞
102 Chapter 1

Theorem 1.129 Let (X, d) be a complete metric space and let fn ∈ C(X) converge
pointwise to f . Then, the continuity points of f form a residual set.

Proof : For every n define

Fn (x) = sup fk (x).
k≥n

By the previous lemma, the Fn ’s are lower-semicontinuous. Moreover, Fn is a

decreasing sequence, converging pointwise to f (the lim sup equals the limit).
Let
En = {continuity points of Fn }.
This is a dense G -set, hence
∞
E = � En
n=1
is also a dense G -set.
Let x ∈ E and let " > 0. Since Fn (x) → f (x), there exists an n such that

Fn (x) < f (x) + .

"
2
Since x ∈ En is a continuity point of Fn , there exists an r such that

Fn (y) < Fn (x) + ∀y ∈ Br (x).

"
2
Hence, for every y ∈ Br (x),

f (y) ≤ Fn (y) < Fn (x) + < f (x) + ".

"
2
In other words, there exists a dense G -set E, such that

∀x ∈ E, ∀" > 0 ∃r > 0, f (y) < f (x) + " ∀y ∈ Br (x).

Now, we can proceed symmetrically, replacing lower-semicontinuity with upper-

semicontinuity. We infer the existence of a dense G -set F, such that

∀x ∈ F, ∀" > 0 ∃r > 0, f (y) > f (x) − " ∀y ∈ Br (x).

The intersection E ∩ F is a dense G -set, containing all the continuity points of f .

n
 Metric Spaces 103

Example: Let f ∶ R → R be di↵erentiable. It derivative f ′ is not necessarily

continuous, however, since
f (x + 1�n) − f (x)
f ′ (x) = lim
n→∞ 1�n

is a pointwise limit of continuous functions, it follows that the continuity points

of f ′ are a residual set. ▲▲▲