Chapter 2

Metric Spaces

A normed space is a vector space endowed with a norm in which the length of a vector
makes sense and a metric space is a set endowed with a metric so that the distance between
two points is meaningful. There is always a metric associated to a norm. Normed spaces
form a sub-class of metric spaces and metric spaces form a sub-class of topological spaces.
In Section 1 the definitions of a normed space and a metric space are given and some
examples are present. In Section 2 limit of sequences and continuity of functions in a
metric space is defined. Next open and closed sets are introduced and used to describe
the convergence of sequences and continuity of functions in Section 3. Relevant notions
such as the boundary points, interior points, closure and interior of a set are collected in
Section 4.

2.1 Definitions and Examples

By far we are familiar with the Euclidean space and the space of continuous functions. You
may have already observed there are certain similarities between these two mathematical
entities. For instance, the completeness theorem for R states that every Cauchy sequence
converges. This property can be extended to Rn without much effort. In fact, it suffices
to observe that a Cauchy sequence in Rn is defined by replacing the absolute value by the
Euclidean norm v
u n 2
uX
|x| = t xj , x = (x1 , · · · , xn ) .
j=1

In other words, a sequence {xk } in Rn is called a Cauchy sequence if for each ε > 0, there
is some k0 such that
v
u n
uX
k m
|x − x | = t (xkj − xm 2
j ) < ε , ∀k, m ≥ k0 .
j=1

1
2 CHAPTER 2. METRIC SPACES

On the other hand, in the space of continuous function C[a, b], a corresponding com-
pleteness theorem states every Cauchy sequence of functions converges uniformly to a
continuous function. Recall that a sequence {fk } is a Cauchy sequence if for every ε > 0,
there is some k0 such that
kfk − fm k∞ < ε , ∀k, m ≥ k0 .
You can see things are the same except the Euclidean norm is now replaced by the sup-
norm.

In order to unify the Euclidean spaces and the space of continuous functions, we
introduce the general definition of a normed space. This is our first step of abstraction.
A norm k · k is a function on a real vector space X to [0, ∞) satisfying the following three
conditions, for all x, y ∈ X and α ∈ R,
N1. kxk ≥ 0 and “=” 0 if and only if x = 0,
N2. kαxk = |α| kxk, and
N3. kx + yk ≤ kxk + kyk. (Triangle inequality)

A normed space is a vector space endowed with a norm. The pair (X, k · k) is called
a normed space. Here are some examples of normed spaces.

Example 2.1. Let R be the set of all real numbers. For x ∈ R, set its Euclidean norm
|x| to be the absolute value of x. It is easily seen that |x| satisfies N1-N3 above and
so it defines a norm. In particular, N3 is the usual triangle inequality. Thus (R, | · |) is
a normed space. From now on whenever we talk about R, it is understood that it is a
normed space endowed with the Euclidean norm.
Example 2.2. More generally, let Rn be the n-dimensional real vector space consisting
of all n-tuples x = (x1 , . . . , xn ), xj ∈ R, j = 1, . . . , n. Introduce the Euclidean norm
q
|x| = x21 + · · · + x2n .
It reduces to the previous example when n = 1. Apparently, N1 and N2 are fulfilled.
Moreover, taking square of both sides of the triangle inequality, N3 follows from the
Cauchy-Schwarz Inequality
X n  n n
2 1/2
X X
1/2
yj2


xj yj  ≤ xj .
 

1 1 1

Example 2.3. In general, for p ∈ [1, ∞), we may define

∞
!1/p
X
kxkp = |xj |p ,
j=1
2.1. DEFINITIONS AND EXAMPLES 3

and for p = ∞, define kxk∞ = max{|x1 |, · · · , |xn |}. It can be shown that this is a norm
on Rn . To verify N3 one needs the Minkowski’s Inequality, see my notes on inequalities.
As the 2-norm comes up much more often than other p-norm, and it justifies to use a
simpler notation | · |.

Example 2.4. Let C[a, b] be the real vector space of all continuous, real-valued functions
on [a, b]. For f ∈ C[a, b], define the sup-norm

kf k∞ ≡ max{|f (x)| : x ∈ [a, b]}.

It is easily checked that it defines a norm on C[a, b]. Parallel to Rn , for each p ∈ [1, ∞),
we define the p-norm by
ˆ b 1/p
p
kf kp = |f | , p ∈ [1, ∞) ,
a

on C[a, b]. Each p-norm defines on C[a, b]. The triangle inequality is the integral form of
Minkowski’s Inequality, see the end of the chapter.

Example 2.5. Let E be a nonempty subset of Rn , let Cb (E) be the family of all bounded,
continuous functions defined in E. It is readily checked that Cb (E) forms a vector space
under the usual addition and scalar multiplication of functions. Moreover, the supnorm is
well defined and makes Cb (E) into a normed space. It reduces to C[a, b] when E = [a, b]
as continuous functions on a closed, bounded interval must be bounded.

In passing we point out same notations such as k · kp have been used to denote norms
on different spaces. As they arise in quite different context, hopefully it would not cause
much confusion.

Example 2.6. Let R[a, b] be the vector space of all Riemann integrable functions on [a, b]
and consider kf k∞ and kf k1 as defined above. It is routine to verify that while kf k∞
defines a metric, kf k1 is not as N2 and N3 are satisfied but not N1. In fact, we know from
the previous chapter that kf k1 = 0 if and only if f vanishes almost everywhere. Hence
R[a, b] does not form a normed space under k·k1 . The space of integrable functions, where
Riemann integrability is replaced by Lebesgue integrability, will be studied in detailed in
Real Analysis.

Observe that whenever (X, k · k) is a normed space and Y is a vector subspace of X,

we can make Y into a normed space by restricting the norm to Y . However, apparently
such hereditary property does not hold when Y is merely a subset but not a subspace
of X. We have come to the second step of abstraction. We will detach the vector space
4 CHAPTER 2. METRIC SPACES

structure from the norm structure by introducing the notion of a metric space as follows.

Let X be a non-empty set. We would like to define a concept of distance which assigns
a positive number to every two points in X, that is, the distance between them. In analysis
the name metric is used instead of distance. (But “d” not “m” is used in notation. I have
no idea why it is so.) A metric on X is a function from X × X to [0, ∞) which satisfies
the following three conditions: ∀x, y, z ∈ X,
M1. d(x, y) ≥ 0 and equality holds if and only if x = y,
M2. d(x, y) = d(y, x), and
M3. d(x, y) ≤ d(x, z) + d(z, y) . (Triangle inequality)
M3 is a key property of a metric. M2 and M3 together imply another form of the triangle
inequality, namely,

|d(x, y) − d(x, z)| ≤ d(y, z) .

There is always a metric associated to a norm in a normed space (X, k · k). Indeed,
letting
d(x, y) = kx − yk,
it is readily checked that d defines a metric on X. This metric is called the metric
induced by the norm. For instance, in Rn we have
dp (x, y) = kx − ykp , p ∈ [1, ∞] .
On C[a, b], we have
dp (f, g) = kf − gkp , p ∈ [1, ∞] .

Given a metric space (X, d), the metric ball centered at x and with radius r is the
set
Br (x) = {y ∈ X : d(y, x) < r} .
As there could be more than one metrics on a set, it is interesting to compare the metric
balls with respect to different metrics. For instance, on Rn there are infinitely many met-
rics given by dp . Denote the metric balls in d2 , d1 and d∞ metrics by Br (x), Br1 (x) and
Br∞ (x) respectively. Then Br (x) is the usual ball of radius r centered at x and Br∞ (x) is
the cube of length r centered at x. I let you draw Br1 (x) as an exercise.

In the following we give two examples of metrics defined on a set without the structure
of a vector space. They are not metric spaces induced by normed spaces.
2.1. DEFINITIONS AND EXAMPLES 5

Example 2.7. Let X be a non-empty set. For x, y ∈ X, define


1, x =6 y,
d(x, y) =
0, x = y.

The metric d is called the discrete metric on X. The metric ball Br (x) consists of x
itself for all r ∈ (0, 1] and is equal to X when r > 1.

Example 2.8. Let H be the collection of all strings of words in n digits. For two strings
of words in H, a = a1 · · · an , b = b1 · · · bn , aj , bj ∈ {0, 1, 2, . . . , 9}. Define

dH (a, b) = the number of digits at which aj is not equal to bj .

By using a simple induction argument one can show that (H, dH ) forms a metric space.
Indeed, the case n = 1 is straightforward. Let us assume it holds for n-strings and
show it for (n + 1)-strings. Let a = a1 · · · an an+1 , b = b1 · · · bn bn+1 , c = c1 · · · cn cn+1 , a0 =
a1 · · · an , b0 = b1 · · · bn , and c0 = c1 · · · cn . As dH (a0 , b0 ) ≤ dH (a, b) always holds for any
a, b, it suffices to consider two cases, namely, (a) an+1 = bn+1 and (b) an+1 6= bn+1 .
In (a), we have dH (a, b) = dH (a0 , b0 ). By induction hypothesis, dH (a, b) = dH (a0 , b0 ) ≤
dH (a0 , c0 ) + dH (c0 , b0 ) ≤ dH (a, c) + dH (c, b), done. In (b), dH (a, b) = dH (a0 , b0 ) + 1. Since
an+1 is not equal to bn+1 , either cn+1 is not equal to an+1 or bn+1 . Assume it is the former.
Then dH (a, c) = dH (a0 , c0 ) + 1, so d(a, b) = d(a0 , b0 ) + 1 ≤ dH (a0 , c0 ) + dH (c0 , b0 ) + 1 ≤
dH (a, c) + dH (c0 , b0 ) ≤ dH (a, c) + dH (c, b), done. (Thanks to a student who suggested this
simplified proof.)

There is a common way to construct metrics based on the following observation.

Let Φ be a bijection from a metric space (X, d) to a set Z. For z1 , z2 ∈ Z, define
ρ(z1 , z2 ) = d(Ψ(z1 ), Ψ(z2 )) where Ψ is the inverse of Φ. Then ρ becomes a metric on Z.

Example 2.9. The Euclidean metric on R is induced from the Euclidean norm. We now
construct another metric on R as follows. Imagine that R is the x-axis on the plane and
consider the unit circle S = {p = (x, y) : x2 + y 2 = 1}. The unit circle is endowed with a
natural metric, that is, the arclength between two points. (I let you define it rigorously.)
For every point p on the circle not equal to the north pole (0, 1), the ray connecting the
north pole and p would hit the x-axis at a unique point x. It is clear that p 7→ x sets
up a one-to-one correspondence between S \ {(0, 1)} and R, the x-axis. For x1 , x2 ∈ R,
we define ρ(x1 , x2 ) to be the arclength of the arc formed by their corresponding points on
the circle. The metric is bounded in the sense that ρ(x1 , x2 ) ≤ 2π for all x1 , x2 .

Let Y be a non-empty subset of (X, d). Then (Y, d|Y ×Y ) is again a metric space. It
is called a metric subspace of (X, d). The notation d|Y ×Y is usually written as d for
simplicity. Every non-empty subset of a metric space forms a metric space under the
restriction of the metric. In the following we usually call a metric subspace a subspace
6 CHAPTER 2. METRIC SPACES

for simplicity. Again a metric subspace of a normed space needs not be a normed space.
It is so only if the subset is also a vector subspace. For example, consider the subsets
E = {(x, y) : 2x + 3y = 0} and F = {(x, y) : xy = 1, x, y > 0} in R2 . The restriction
of the Euclidean metric to these two sets make them metric spaces. The first one is a
one-dimensional vector subspace of R2 , E is a normed space. On the other hand, while
the restriction of the Euclidean metric on F makes it a metric space, it is no longer a
metric induced from any norm. In fact, F is no longer a vector space. Taking (1, 1) and
(2, 1/2) from F , that the point (3, 2.5) = (1, 1) + (2, 1/2) does not satisfy 3 × 2.5 = 1
shows F does not inherit the vector space structure of R2 .

2.2 Limits and Continuity

Convergence of sequences of real numbers and uniform convergence of sequences of func-

tions are the main themes in Mathematical Analysis I and II and sequences of vectors
were considered in Advanced Calculus I and II. With a metric d on a set X, it makes
sense to talk about limits of sequences in a metric space. Indeed, a sequence in (X, d) is
a map ϕ from N to (X, d) and usually we write it in the form {xn } where ϕ(n) = xn . A
sequence {xn } is said to converge to x if limn→∞ d(xn , x) = 0, that’s, for every ε > 0,
there exists n0 such that d(xn , x) < ε, for all n ≥ n0 . When this happens, we write or
limn→∞ xn = x or xn → x in X.
Convergence of sequences in (Rn , d2 ) reduces to the old definition encountered before.
From now on, we implicitly refer to the Euclidean metric whenever convergence of se-
quences in Rn is considered. For sequences of functions in (C[a, b], d∞ ), it is simply the
uniform convergence of sequences of functions in C[a, b]. Likewise, uniform convergence
of sequences in Cb (E) in Example 2.5 is the convergence with respect to d∞ .

As there could be more than one metrics defined on the same set, it is natural to make
a comparison among these metrics. Let d and ρ be two metrics defined on X. We call ρ
is stronger than d, or d is weaker than ρ, if there exists a positive constant C such that
d(x, y) ≤ Cρ(x, y) for all x, y ∈ X. They are equivalent if d is stronger and weaker than
ρ simultaneously, in other words,

d(x, y) ≤ C1 ρ(x, y) ≤ C2 d(x, y), ∀x, y ∈ X,

for some positive C1 and C2 . When ρ is stronger than d, a sequence converging in ρ is

also convergent in d. When d and ρ are equivalent, a sequence is convergent in d if and
only if it is so in ρ.
Take dp and d∞ on Rn as an example. It is elementary to show that for all x, y ∈ Rn ,

dp (x, y) ≤ n1/p d∞ (x, y) ≤ n1/p dp (x, y),

2.2. LIMITS AND CONTINUITY 7

and
d1 (x, y) ≤ nd∞ (x, y) ≤ nd1 (x, y),
hence all dp and d∞ are all equivalent. The convergence of a sequence in one metric
implies its convergence in all others.
It is a basic result in functional analysis that every two induced metrics in a finite
dimensional normed space are equivalent. Consequently, examples of inequivalent induced
metrics can only be found when the underlying space is of infinite dimension.
Let us display two inequivalent metrics on C[a, b]. For this purpose it suffices to
consider d1 and d∞ . On one hand, clearly

d1 (f, g) ≤ (b − a)d∞ (f, g), ∀f, g ∈ C[a, b],

so d∞ is stronger than d1 . But the reverse is not true. Consider the sequence given by
(taking [a, b] = [0, 1]) 
−n3 x + n, x ∈ [0, 1/n2 ],
fn (x) =
0, x ∈ (1/n2 , 1].
We have d1 (fn , 0) → 0 but d∞ (fn , 0) → ∞ as n → ∞. Were d∞ (fn , 0) ≤ Cd1 (fn , 0) true
for some positive constant C, d1 (fn , 0) must tend to ∞ as well. Now it tends to 0, so d1
cannot be stronger than d2 and these two metrics are not equivalent.

A metric space (X, d) is called bounded if d(x, y) ≤ M for all x, y ∈ X. A set E

is bounded if the subspace (E, d) is bounded. It is clear that E is bounded if and only
if E ⊂ BR (x) for some x ∈ X and R > 0. The diameter of a set is defined to be
diam(E) = sup{d(x, y) : x, y ∈ E} ≤ ∞.

It is a little unexpected that two inequivalent metrics may have the same strength of
convergence.

Proposition 2.1. Let (X, d) be a metric space. Define

d(x, y)
ρ(x, y) = .
1 + d(x, y)

Then ρ is a metric on X. Moreover, a sequence converges in d if and only it converges

in ρ.

I leave the proof of this proposition as an exercise. We note that ρ is always bounded,
ρ(x1 , x2 ) < 1 for all x1 , x2 . For instance, on R there is the Euclidean metric and the new
metric
|x − y|
ρ(x, y) = .
1 + |x − y|
8 CHAPTER 2. METRIC SPACES

We have ρ(x, y) ≤ |x − y|, ∀x, y , but there is no constant C to make |x − y| ≤ Cρ(x, y)

holds. These two metrics are not equivalent.

Now we define continuity in a metric space. First of all, recall that there are two
ways to describe it, namely, the behavior of sequences and the ε-δ formulation. Specif-
ically, the function f is continuous at x ∈ E if for every sequence {xn } ⊂ E satisfying
limn→∞ xn = x, limn→∞ f (xn ) = f (x). Equivalently, for every ε > 0, there exists some
δ > 0 such that |f (y) − f (x)| < ε whenever y ∈ E, |y − x| < δ. Both definition can be
formulated on a metric space. We will adapt the sequence approach. Let (X, d) and (Y, ρ)
be two metric spaces and f : (X, d) → (Y, ρ). Let x ∈ X. We call f is continuous at x
if f (xn ) → f (x) in (Y, ρ) whenever xn → x in (X, d). It is continuous on a set E ⊂ X
if it is continuous at every point of E.

First we show the sequence formulation is equivalent to the ε-δ-formulation.

Proposition 2.2. Let f be a mapping from (X, d) to (Y, ρ) and x0 ∈ X. Then f is

continuous at x0 if and only if for every ε > 0, there exists some δ > 0 such that
ρ(f (x), f (x0 )) < ε for all x, d(x, x0 ) < δ.

Proof. ⇐) Let ε be given and δ is chosen accordingly. For any {xn } → x0 , given δ > 0,
there exists some n0 such that d(xn , x0 ) < δ , ∀n ≥ n0 . It follows that ρ(f (xn ), f (x0 )) < ε
for all n ≥ n0 , so f is continuous at x0 .
⇒) Suppose that the implication is not valid. There exist some ε0 > 0 and {xk } ∈ X
satisfying d(f (xk ), f (x0 )) ≥ ε0 and d(xk , x0 ) < 1/k. However, the second condition tells
us that {xk } → x0 , so by the continuity at x0 one should have d(f (xk ), f (x0 )) → 0,
contradiction holds.

Shortly we will face a third formulation, that is, by open/closed sets to describe
continuity in a metric space.
As usual, continuity of functions is closed under compositions of functions.

Proposition 2.3. Let f : (X, d) → (Y, ρ) and g : (Y, ρ) → (Z, m) be given.

(a) If f is continuous at x and g is continuous at f (x), then g ◦ f : (X, d) → (Z, m) is

continuous at x.

(b) If f is continuous in X and g is continuous in Y , then g ◦ f is continuous in X.

Proof. It suffices to prove (a). Let xn → x. Then f (xn ) → f (x) as f is continuous at x.

Then (g ◦ f )(xn ) = g(f (xn )) → g(f (x)) = (g ◦ f )(x) as g is continuous at f (x).
2.3. OPEN AND CLOSED SETS 9

To end this section, we consider the following question: Are there any continuous
functions in a metric space? Of course, constant functions are obviously continuous.
But we want some non-trivial ones. It turns out that many continuous functions can be
constructed from the distance function. (What else?) Indeed, let A be a non-empty set
in (X, d). We define the distance from a point x to A by

ρA (x) = inf {d(y, x) : y ∈ A} .

We claim:
|ρA (x) − ρA (y)| ≤ d(x, y), ∀x, y ∈ X.
It shows that ρA is not only continuous but also “Lipschitz continuous”. To prove the
claim, let ε > 0, we pick z ∈ A such that ρA (y) + ε > d(y, z). The

ρA (x) ≤ d(x, z)
≤ d(x, y) + d(y, z)
≤ d(x, y) + ρA (y) + ε .

Since ε could be any positive number, we conclude that ρA (x) − ρA (y) ≤ d(x, y), and the
claim follows after noting that the roles of x and y are exchangeable.
It is convenient to introduce the notations:

d(x, F ) = inf{d(x, y) : y ∈ F } ,

and
d(E, F ) = inf{d(x, y) : x ∈ E, y ∈ F } .

The collection of continuous functions is large in the sense that they separate points.
Indeed, let x1 and x2 be two distinct points in X. The continuous function ρ{x1 } satisfies
ρ{x1 } (x1 ) = 0 and ρ{x1 } (x2 ) > 0. Further related results can be found in the exercise.

2.3 Open and Closed Sets

The existence of a metric on a set enables us to talk about the convergence of a sequence
and continuity of a map. It turns out that, in order to define continuity, a structure
less stringent then a metric structure is needed. It suffices the set be endowed with a
topological structure. In a word, a metric induces a topological structure on the set but
not every topological structure comes from a metric. In a topological space, continuity
can no longer be defined via the convergence of sequences. Instead one uses the notion
of open and closed sets in the space. As warm up for topology we discuss how to use the
language of open/closed sets to describe the convergence of sequences and the continuity
of functions in this section.
10 CHAPTER 2. METRIC SPACES

First the definition. Let (X, d) be a metric space. A set G ⊂ X is called an open
set if for each x ∈ G, there exists some ρ such that Bρ (x) ⊂ G. The number ρ may vary
depending on x. We also define the empty set φ to be an open set. Roughly speaking, an
open set is a subset in which every point is surrounded by points in the set.

Proposition 2.4. Let (X, d) be a metric space. We have

(a) X and φ are open sets.

S
(b) If α∈A Gα is an open set provided that all Gα , α ∈ A, are open where A is an
arbitrary index set.
TN
(c) If G1 , . . . , GN are open sets, then j=1 Gj is an open set.

Note the union in (b) of this proposition is over an arbitrary collection of sets while
the intersection in (c) is a finite one.

Proof. (a) Obvious.

S
(b) Let x ∈ α∈A Gα . There exists some S x ∈ Gα1 . As Gα1 is open, there
S α1 such that
is some Bρ (x) ⊂ Gα1 . But then Bρ (x) ⊂ α∈A Gα , so α∈A Gα is open.

(c) When N
T
j=1 Gj is empty, it is open by definition. On the other hand, let x ∈
TN
j=1 Gj . For each j, there exists Bρj (x) ⊂ Gj . Let ρ = min {ρ1 , . . . , ρN }. Then Bρ (x) ⊂
TN TN
j=1 Gj , so j=1 Gj is open.

The complement of an open set is called a closed set. Taking the complement of
Proposition 2.4, we have

Proposition 2.5. Let (X, d) be a metric space. We have

(a) X and φ are closed sets.

T
(b) If Fα , α ∈ A, are closed sets, then α∈A Fα is a closed set.
SN
(c) If F1 , . . . , FN are closed sets, then j=1 Fj is a closed set.

Note that X and φ are both open and closed. The terminology of a closed set will
become evident soon.
2.3. OPEN AND CLOSED SETS 11

Example 2.10. Every ball in a metric space is an open set. Let Br (x) be a ball and
y ∈ Br (x). We claim that Bρ (y) ⊂ Br (x) where ρ = r − d(y, x) > 0. For, if z ∈ Bρ (y),

d(z, x) ≤ d(z, y) + d(y, x)

< ρ + d(y, x)
= r,

by the triangle inequality, so z ∈ Br (x) and Bρ (y) ⊂ Br (x) holds. Next, the set E = {y ∈
X : d(y, x) > r} for fixed x and r ≥ 0 is an open set. For, let y ∈ E, d(y, x) > r. We
claim Bρ (y) ⊂ E, ρ = d(y, x) − r > 0,. For, letting z ∈ Bρ (y),

d(z, x) ≥ d(y, x) − d(y, z)

> d(y, x) − ρ
= r,

shows that Bρ (y) ⊂ E, hence E is open. Finally, consider F = {x ∈ X : d(x, z) = r > 0}

where z and r are fixed. Observing that F is the complement of the two open sets Br (z)
and {x ∈ X : d(x, z) > r}, we conclude that F is a closed set.

Example 2.11. In the real line every open interval (a, b), −∞ ≤ a ≤ b ≤ ∞, is an open
set. Other intervals such as [a, b), [a, b], (a, b], a, b ∈ R, are not open. It can be shown
that every open set G in R can be written as a disjoint union of open intervals. Letting
(an , bn ) = (−1/n, 1/n),
∞  
\ 1 1
− , = {0}
n=1
n n

is not open. It shows that Proposition 2.3(c) does not hold when the intersection is over
infinite many sets. On the other hand, [a, b] is a closed set since it is the complement of
the open set R \ (−∞, a) ∪ (a, ∞). Setting a = b, {a} is closed, that is, a single point is al-
ways a closed set. Finally, some sets we encounter often are neither open nor closed. Take
the set of all rational numbers as example, as every open interval containing a rational
number also contains an irrational number, we see that Q is not open. The same reason-
ing shows that the set of all irrational numbers is not open, hence Q is also not a closed set.

Example 2.12. When we studied multiple integrals in Advanced Calculus II, we encoun-
tered many domains or regions as the domain of integration. These domains are bounded
by nice curves in the plane or by nice surfaces in the space. Without counting its bound-
ary points, the interior of these domains form open sets. The exterior of these domains,
again excluding the boundary points, are also open sets. Therefore, the set consisting of
all boundary points is a closed set as it is the complement of the union of two open sets,
namely, the interior and exterior of the domain.
12 CHAPTER 2. METRIC SPACES

Example 2.13. Consider the set E = {f ∈ C[a, b] : f (x) > 0, ∀x ∈ [a, b]} in C[a, b]. We
claim that it is open. For f ∈ E, it is positive everywhere on the closed, bounded interval
[a, b], hence according to extremal value theorem it attains its minimum at some x0 . It
follows that f (x) ≥ m ≡ f (x0 ) > 0. Letting r = m/2, for g ∈ Br (f ), d∞ (g, f ) < r = m/2
implies

g(x) ≥ f (x) − |g(x) − f (x)|

m
> m−
2
m
= > 0,
2
for all x ∈ [a, b], hence g ∈ E which implies Br (f ) ⊂ E, E is open. Likewise, sets like
{f : f (x) > α, ∀x}, {f : f (x) < α, ∀x} where α is a fixed number. On the other hand,
by taking complements of these open sets, we see that the sets {f : f (x) ≥ α, ∀x}, {f :
f (x) ≤ α, ∀x} are closed.

Example 2.14. Consider the extreme case where the space X is endowed with the discrete
metric. We claim that every set is open and closed. Clearly, it suffices to show that every
singleton set {x} is open. But, this is obvious because the ball B1/2 (x) = {x} ⊂ {x}.

We now use open sets to describe the convergence of sequences.

Proposition 2.6. Let (X, d) be a metric space. A sequence {xn } converges to x if and
only if for each open G containing x, there exists n0 such that xn ∈ G for all n ≥ n0 .

Proof. Let G be an open set containing x. According to the definition of an open set,
we can find Bε (x) ⊂ G. It follows that there exists n0 such that d(xn , x) < ε for all
n ≥ n0 , i.e., xn ∈ Bε (x) ⊂ G for all n ≥ n0 . Conversely, taking G = Bε (x), we see that
xn → x.

From this proposition we deduce the following result which explains better the termi-
nology of a closed set.
Proposition 2.7. The set A is a closed set in (X, d) if and only if whenever {xn } ⊂ A
and xn → x as n → ∞ implies that x belongs to A.

Proof. ⇒). Assume on the contrary that x does not belong to A. As X \ A is an open
set, we can find a ball Bε (x) ⊂ X \ A. However, as xn → x, there exists some n0 such
that xn ∈ Bε (x) for all n ≥ n0 , contradicting the fact that xn ∈ A.
T
⇐). If X \ A is not open, say,Twe could find a point x ∈ X \ A such that B1/n (x) A 6= φ
for all n. Pick xn ∈ B1/n (x) A to form a sequence {xn }. Clearly {xn } converges to x.
By assumption, x ∈ A, contradiction holds. Hence X \ A must be open.
2.3. OPEN AND CLOSED SETS 13

Now we use open sets to describe continuity.

Proposition 2.8. Let f : (X, d) → (Y, ρ).

(a) f is continuous at x if and only if for every open set G containing f (x), f −1 (G)
contains Bρ (x) for some ρ > 0.

(b) f is continuous in X if and only if for every open G in Y , f −1 (G) is an open set in
X.

These statements are still valid when “open” is replaced by “closed”.

Proof. We consider (a) and leave (b) as an exercise.

⇒). Suppose there exists some open G such that f −1 (G) does not contain B1/n (x) for all
n ≥ 1. Pick xn ∈ B1/n (x), xn ∈ / f −1 (G). Then xn → x but f (xn ) does not converge to
f (x), contradicting the continuity of f .
⇐). Let {xn } → x in X. Given any open set G containing f (x), we can find Br (x) ⊂
f −1 (G). Thus, there exists n0 such that xn ∈ Br (x) for all n ≥ n0 . It follows that
f (xn ) ∈ G for all n ≥ n0 . By Proposition 2.7, f is continuous at x.

This proposition shows in particular that for a continuous function F : (X, d) → R,

the sets F −1 ((a, b)) are open and F −1 ([a, b]), −∞ ≤ a < b ≤ ∞, are closed. This gives an
effective way to determine whether a set is open or closed. Let us look at the following
examples.
Example 2.15. Consider Example 2.11 again. In R2 , the domains are obtained as the
intersection of several curves. A curve may be described as the zero set of some function.
A typical description of a domain would be like

D = (x, y) ∈ R2 : f (x) < y < g(x), x ∈ (a, b)



where f, g are continuous function from Rn to R satisfying f (x) < g(y). If we let A1 =
{(x, y) : y > f (x)}, A2 = {(x, y) : y < g(x)} and B = {(x, y) : x ∈ (a, b), y ∈ (−∞, ∞)}.
It is not hard to see that A1 and A2 are open. In fact, let F (x, y) = y − f (x) which is
clearly continuous, so A1 = F −1 (0, ∞) is open. Similarly A2 is open. (You may also verify
it using the definition of an open set.) Since it is clear that B is open, D = A1 ∩ A2 ∩ B
is also open.
Some open sets are the regions bounded by closed curves. In this case, a single function
is sufficient to define them. For instance, the region bounded by an ellipse is described
as Ω = {(x, y) ∈ R2 : x2 /a2 + y 2 /b2 < 1}. As F (x, y) = x2 /a2 + y 2 /b2 is obviously a
continuous function, and by Ω = F −1 (−∞, 1), it is an open set.
14 CHAPTER 2. METRIC SPACES

Example 2.16. Let us consider an example in C[a, b]. In Example 2.12 we showed that
the set P = {f : f (x) > 0, ∀x ∈ [a, b]} is open in C[a, b]. We claim that the set
D = {f : f 2 (x) − sin f (x) > 0} is also an open set. For, the map Φ(f ) = f 2 − sin f
defines a map from C[a, b] to itself (check it). Now we realize that D = {f : (Φ ◦ f )(x)) >
0 , ∀x ∈ [a, b]}. From D = Φ−1 P we conclude that D is open too.

2.4 Points in a Metric Space

We describe some further useful notions associated to sets in a metric space.

Let E be a set in (X, d). A point x is called a boundary point of E if G ∩ E and
G \ E are non-empty for every open set G containing x. Of course, it suffices to take G of
the form Bε (x) for all small ε or ε = 1/n, n ≥ 1. We denote the boundary of E by ∂E.
The closure of E, denoted by E, is defined to be E ∪ ∂E. Clearly ∂E = ∂(X \ E). The
boundary of the ball Br (x) in Rn is theSsphere Sr (x) = {y ∈ Rn : d2 (y, x) = r}. Hence,
the closed ball Br (x) is given by Br (x) Sr (x), which is precisely the closure of Br (x).

Example 2.17. Let E = [0, 1) × [0, 1) ⊂ R2 . It is easy to see that ∂E = [0, 1] × {0, 1} ∪
{0, 1} × [0, 1]. Thus some points in ∂E belong to E and some do not. The closure of E,
E, is equal to [0, 1] × [0, 1].

It can be seen from definition that the boundary of the empty set is the empty set.
Also, the boundary of a set is always a closed set. For, let {xn } be a sequence in ∂E
converging to some x. For any ball Br (x), we can find some xn in it, so the ball Bρ (xn ), ρ =
r − d(xn , x) > 0, is contained in Br (x). As xn ∈ ∂E, Bρ (xn ) has non-empty intersection
with E and X \ E, so does Br (x) and x ∈ ∂E too. The following proposition characterizes
the closure of a set as the smallest closed set containing this set.

Proposition 2.9. Let E be a set in (X, d). We have

(a) x ∈ E if and only if Br (x) ∩ E 6= φ for all r > 0.

(b) A ⊂ B implies A ⊂ B.

(c) E is a closed set.

T
(d) E = {C : C is a closed set containing E}.

Proof. (a) If x ∈ E, then x is always contained in Br (x) ∩ E for all r. On the other hand,
if x ∈ ∂E, by definition Br (x) ∩ E 6= φ for all r too. Conversely, it is trivial when x ∈ E.
When x does not belong to E but Br (x) ∩ E 6= φ, then x ∈ ∂E ⊂ E, done.
2.4. POINTS IN A METRIC SPACE 15

(b) Let x ∈ A. By (a) Br (x) ∩ A 6= φ for all r. But as A ⊂ B, Br (x) ∩ B 6= φ for all r.
By (a), x ∈ B.
(c) Let xn ∈ E and xn → x. We need to show x ∈ E. If not true, x neither belong to E
nor ∂E. So there is some Br (x) disjoint from E, contradicting xn → x.
(d) Denote the right hand side by F . It is a closed set. By (c) E is a closed set containing
E, so F ⊂ E already. TOn the other hand, for any closed C satisfying E ⊂ C, (b) implies
E ⊂ C = C , so E ⊂ C = F .

A point x is called an interior point of E if there exists an open set G such that
x ∈ G ⊂ E. It can be shown that all interior points of E form an open set call the
interior of E, denoted by E o . It is not hard to see that E 0 = E \ ∂E. The
interior of
o
a set is related to its closure by the following relation: E = X \ X \ E . Using this
relation, one can show that the interior of a set is the largest open set sitting inside E.
More precisely, G ⊂ E 0 whenever G is an open set in E.

Example 2.18. Consider the set of all rational numbers E in [0, 1]. It has no interior point
since there are irrational numbers in every open interval containing a rational number, so
I o is the empty set. On the other hand, since every open interval contains some rational
numbers, the closure of E, E, is [0, 1]. It shows the interior and closure of a set could be
very different.

Example 2.19. In Example 2.11 we consider domains in R2 bounded by several contin-

uous curves. Let D be such a domain and the curves bounding it be S. It is routine to
verify that ∂D = S, that is, the set of all boundary points of D is precisely the S and the
closure of D, D, is D ∪ S. The interior of D is D.

Example 2.20. For any two sets E and F in the same space, it is not hard to show
E ∪ F = E ∪ F . But (E ∪ F )o may not always equal to E o ∪ F o . As an extreme case,
take E = Q and I in R. We have (Q ∪ I)0 = R0 = R, but Qo ∪ Io = φ ∪ φ = φ. In general,
we only have E o ∪ F 0 ⊂ (E ∪ F )o .

Example 2.21. Let

S = {f ∈ C[0, 1]) : 1 < f (x) ≤ 5, x ∈ [0, 1]}.

We have
S = {f ∈ C[0, 1] : 1 ≤ f (x) ≤ 5, x ∈ [0, 1]} .
For, denote this set by A. As we have A = f −1 ([1, 5] where f is continuous, A is a closed set
containing S. On the other hand, let f ∈ A, the functions fn (x) = max{f (x), 1 + 1/n} ∈
S, n ≥ 1, and fn → f in sup-norm. (Recall the fact that max{f, g} is continuous when
16 CHAPTER 2. METRIC SPACES

f and g are continuous.) Hence every function in A is the limit of some sequence in S,
so A is contained in any closed set containing S. We conclude that A is the smallest
closed set containing S, that is, A = S. On the other hand, the interior of S is given
by S o = {f ∈ C[0, 1] : 1 < f (x) < 5, x ∈ [0, 1]}. Denoting this set by B, from
B = f −1 ((1, 2)) we see that B is open set in S. On the other hand, if g is an interior
point of S, there is some δ > 0 such that Bδ (g) ⊂ S. In other words, 1 < h(x) ≤ 5 for all
h, kh − gk∞ < δ. In particular, g(x) < h(x) − δ ≤ 5 on [0, 1]. So g belongs to B, that is,
B is the largest open set in S.

2.5 Elementary Inequalities for Functions.

We start with the Young’s Inequality covered in MATH2060.

Young’s Inequality. For a, b > 0 and p > 1,

ap b q 1 1
ab ≤ + , + = 1,
p q p q

and equality sign holds if and only if ap = bq .

The number q is called the conjugate of p. Note that q > 1. The proof of this inequality
is left to you. Basically, we use calculus to show the function

ap b q
ϕ(a) = + − ab ,
p q

where b is fixed, has a unique minimum over (0, ∞) at the point a = b1/(1−p) , that is,
ap = b q .

Theorem 2.10. (Hölder’s Inequality.) Let f, g ∈ R[a, b] and p > 1. Then

ˆ b ˆ b 1/p ˆ b 1/q
p q
|f (x)g(x)| dx ≤ |f (x)| dx |g(x)| dx , q is conjugate to p .
a a a

Equality sign in this inequality holds if and only if either (a) f or g vanish almost every-
where, or (b) there is some positive λ such that |g|q = λ|f |p almost everywhere.

Proof. Assume kf kp and kgkq are positive, otherwise the inequality holds trivially.
For ε > 0 to be chosen, by Young’s Inequality,

εp |f (x)|p ε−q |g(x)|q

|f (x)g(x)| = |εf (x)ε−1 g(x)| ≤ + .
p q
2.5. ELEMENTARY INEQUALITIES FOR FUNCTIONS. 17

Integrate this inequality to get

ˆ b ˆ ˆ
εp b p ε−q b
|f (x)g(x)| dx ≤ |f (x)| dx + |g(x)|q dx . (2.1)
a p a q a
We now choose ε so that
ˆ b ˆ b
p p −q
ε |f (x)| dx = ε |g(x)|q dx ,
a a

that is,
ˆ b  ˆ b −1
p+q q p
ε = |g(x)| dx |f (x)| dx .
a a

Using this epsilon to plug in (1), the right hand side becomes
ˆ b ˆ b ˆ b 1/p ˆ b 1/q
εp p ε−q q p q
|f (x)| dx + |g(x)| dx = |f (x)| dx |g(x)| dx . (2.2)
p a q a a a

The Hölder’s Inequality follows.

To characterize the inequality sign in this inequality, observe case (a) is obvious so let
us assume kf kp , kgkq are both positive, so |f (x)| and |g(x)| are positive almost everywhere.
From (2.1) and (2.2) we see that the inequality sign in (1) becomes equality, that is,
ˆ b p
ε |f (x)|p ε−q |g(x)|q

+ − |f (x)g(x)| dx = 0 .
a p q
The integrand is a non-negative function by Young’s Inequality. The vanishing of this
integral implies that the integrand must vanish almost everywhere, that is,
εp |f (x)|p ε−q g(x)q
+ − |f (x)g(x)| = 0 a.e. .
p q
By the equality sign condition in Young’s Inequality, we conclude that

εp |f (x)|p = ε−q g(x)q a.e.,

that is, |g(x)|q = λ|f (x)|p almost everywhere where λ = ε−p−q .

Remarks. (a) We have used the following proposition proved in Chapter 1: For f ∈
R[a, b],
ˆ b
|f | dx = 0 if and only if f = 0 a.e. .
a
We also point out, when f ∈ C[a, b],
ˆ b
|f | dx = 0 if and only if f = 0 everywhere .
a
18 CHAPTER 2. METRIC SPACES

(b) When f and g in Hölder’s Inequality are continuous, almost everywhere in the char-
acterization of equality sign becomes everywhere.
(c) The inequality still holds in the limiting cases. In fact, when g ∈ C[a, b] and p = 1,
we have ˆ b ˆ b
|f (x)g(x)| dx ≤ |f (x)| dxkgk∞ .
a a

When f ∈ C[a, b] and p = ∞,

ˆ b ˆ b
|f (x)g(x)| dx ≤ kf k∞ |g(x)| dx .
a a

But there is no clean characterization of the equality sign.

Theorem 2.11. (Minkowski’s Inequality.) For f, g ∈ R[a, b] and p > 1,

kf + gkp ≤ kf kp + kgkp .

Equality sign in this inequality holds if and only if either (a) f or g vanishes almost
everywhere, or (b) kf kp , kgkp > 0 and there is some positive λ such that g(x) = λf (x)
almost everywhere.

Proof. Using

|f + g|p = |f + g|p−1 |f + g| ≤ |f + g|p−1 |f | + |f + g|p−1 |g| ,

integrate both sides to get

ˆ b ˆ b ˆ b
p p−1
|f + g| dx ≤ |f + g| |f | dx + |f + g|p−1 |g| dx . (2.3)
a a a

Applying the Hölder’s Inequality to the two integrals on the right separately, we have
ˆ b ˆ b 1/q ˆ b 1/p
q p
|f + g||f | dx ≤ |f + g| dx |f | dx ,
a a a

and
ˆ b ˆ b 1/q ˆ b 1/p
q p
|f + g||g| dx ≤ |f + g| dx |g| dx ,
a a a

where q is conjugate to p. Putting this back to (2.3), we obtain the desired inequality
after some simplifications.

The equality case, in principle, could be treated as in the Hölder’s case. It is easy
to get |g(x)|q = λ|f (x)|p almost everywhere, but rather tedious (or need to use Lebsegue
integral) to get f (x)p = λg(x)q . Luckily, this property has no consequence in our later
2.5. ELEMENTARY INEQUALITIES FOR FUNCTIONS. 19

development.

Comments on Chapter 2. A topology on a set X is a collection of sets τ consisting

the empty set and X itself which is closed under arbitrary union and finite intersection.
Each set in τ is called an open set. The pair (X, τ ) is called a topological space. From
Proposition 2.2 we see that the collection of all open sets in a metric space (X, d) forms
a topology on X. This is the topological space induced by the metric. Metric spaces
constitute a large class of topological spaces, but not every topological space comes from
a metric. However, from the discussions in Section 3 we know that continuity can be de-
fined solely in terms of open sets. It follows that continuity can be defined for topological
spaces, and this is crucial for many further developments. In the past, metric spaces were
covered in Introductory Topology. Feeling that the notion of a metric space should be
learned by every math major, we move it here.

Although the metric space is a standard topic, I found it difficult to fix upon a single
reference book. Rudin’s Principles covers some metric spaces, but his attention is mainly
on the Euclidean space. There are many text books or lecture notes available in the
internet with more or less the same content. Here I simply list the old but very readable
book by E.T. Copson, Metric Spaces, as the main reference.