A2: Metric Spaces

Alan Lauder 1

1
Aside from Chapter 1, these are mildly edited versions of notes written by Ben Green and Kevin McGerty
which were in turn based upon notes due to Richard Earl. Many thanks to them all, and also Dmitry
Belyaev, Panos Papazoglou and Melanie Rupflin.
 Contents

Chapter 1. Differentiability in R2 1
1.1. The total derivative 1
1.2. Directional derivatives 2
1.3. Continuous partial derivatives give differentiability 3
1.4. Example, counterexample and some geometric intuition 5
1.5. Continuity and the chain rule 6
1.6. Some topology in R2 8
1.7. The Inverse function theorem 11

Chapter 2. Metric spaces 15

2.1. The real numbers and the axiom of choice 15
2.2. The definition of a metric space 16
2.3. Some examples of metric spaces 17
2.4. Norms 19
2.5. New metric spaces from old ones 21
2.6. Balls and boundedness 22

Chapter 3. Limits and continuity 25

3.1. Basic definitions and properties 25
3.2. Function spaces 26

Chapter 4. Isometries, homeomorphisms and equivalence 29

4.1. Isometries 29
4.2. Homeomorphisms 29
4.3. Equivalent metrics (non-examinable) 30

Chapter 5. Open and closed sets 33

5.1. Basic definitions 33
5.2. Basic properties of open sets 34
5.3. Continuity in terms of open sets 35
5.4. Topological spaces (non-examinable) 35
5.5. Subspaces 36
iii
iv CONTENTS

Chapter 6. Interiors, closures, limit points 39

6.1. Interiors and closures 39
6.2. Limit points 40

Chapter 7. Completeness 43
7.1. Basic definitions and examples 43
7.2. First properties of complete metric spaces 44
7.3. Completeness of function spaces 45
7.4. The contraction mapping theorem 46

Chapter 8. Connectedness and path-connectedness 49

8.1. Connectedness 49
8.2. Path-connectedness 52
8.3. Connectedness and path-connectedness 53

Chapter 9. Sequential compactness 55

9.1. Definitions 55
9.2. Closure and boundedness properties 55
9.3. Continuous functions on sequentially compact spaces 56
9.4. Product spaces 57
9.5. Sequentially compact equals complete and totally bounded 58

Chapter 10. Compactness 61

10.1. Open covers and the definition of compactness 61
10.2. Compactness implies sequential compactness 62
10.3. Sequential compactness implies compactness 62

Chapter 11. Conformal maps 65

11.1. The extended complex plane C∞ 65
11.2. Conformal transformations 70
 CHAPTER 1

Differentiability in R2

We shall use the usual Euclidean notion of size and distance on Rn . So, for example, with
p
x = (x, y) ∈ R2 we have |x| = x2 + y 2 . We are primarily interested in functions on R2 ,
but in many places it is easier just to work in general dimension.
For f : Ω ⊆ Rn → Rm we of course write f (x) → b as x → a to mean

∀ε > 0, ∃δ > 0 such that ∀x ∈ Ω with 0 < |x − a| < δ we have |f (x) − b| < ε.

And when this holds we write limx→a f (x) = b. (So everything as usual, but just with
“underlines” to indicate vectors.)

1.1. The total derivative

We begin with a definition.

Definition 1.1.1. We say that a function f : Ω ⊆ Rn → Rm is differentiable at a ∈ Ω if

there exists a linear map L : Rn → Rm such that
f (a + h) − f (a) − Lh
(1.1) lim = 0.
h→0 |h|
We call L the (total) derivative of f at a and denote it df (a). We say f is differentiable in
Ω if it is differentiable at every point a ∈ Ω.

We divided by |h| rather than h here because of course it does not make sense to divide
by a vector. (But when f : R2 → R2 and you identify R2 with C, then this is fine, but leads
to a much stronger condition and the whole subject of complex analysis.) The matrix for the
linear map L with respect to the usual bases of Rn and Rm is called the Jacobian matrix.
(The matrix just contains the partial derivatives, as in Prelims Multivariable Calculus: see
for example, equation(1.5).)
We shall shortly see that L is unique when it exists. But where did this definition come
from? Let us assume that f is differentiable at a and do some rearranging. Define
f (a + h) − f (a) − Lh
R(h) := ∈ Rm .
|h|
So we then have

(1.2) f (a + h) − f (a) = Lh + R(h)|h|

1
2 1. DIFFERENTIABILITY IN R2
and when h ̸= 0  
f (a + h) − f (a) h
=L + R(h)
|h| |h|
where now we have R(h) → 0 as h → 0. So then letting h → 0 and noting L is independent
of h one has for h ̸= 0 that
 
f (a + h) − f (a) h
(1.3) lim = L lim
h→0 |h| h→0 |h|
provided both of these two limits exists. And indeed one exists if and only if the other does.
Suppose now that n = 1. Then we have h = h and h/|h| = 1 or −1 according to whether
h > 0 or h < 0. Either way our limit on each side exists, and after taking into account a
sign one gets
f (a + h) − f (a)
lim = L.
h→0 h
And hey presto! we see that L is the usual derivative, but now interpreted as a linear map
on R.
With a sense of false confidence we now look at n = 2. And now one sees that
 
h
lim
h→0 |h|
does not exist. Here h/|h| is a vector of unit length in the direction of h ̸= 0. And since we
allowed complete freedom in letting h → 0 we can take a sequence of such h going to zero
for which the corresponding normalised vectors h/|h| are anything you like at all.
So the moral here is that our usual definition of the derivative is fine when n = 1, but
wholly inadequate for n > 1.

1.2. Directional derivatives

But we can salvage something from the ashes. Let n ∈ Rn be a unit vector.

Definition 1.2.1. We say that a function f : Ω ⊆ Rn → Rm is differentiable at a ∈ Ω

in direction n if
f (a + λn) − f (a)
lim
λ→0 λ
exists.

Assume that f is differentiable at a. And in (1.3) restrict to those h = λn for λ ∈ R.

Well now  
h
lim =n
h→0 |h|
does exist. And so we find
f (a + λn) − f (a)
(1.4) lim = Ln.
λ→0 λ
 1.3. CONTINUOUS PARTIAL DERIVATIVES GIVE DIFFERENTIABILITY 3

That is, the derivative in direction n exists and is just given by applying our matrix L to
the unit vector n. (That n was a unit vector was unimportant; we just need it is non-zero,
as one can easily check that only the direction would matter.)
Note in all this I made no mention of the dimension of the codomain Rm . The reason is
that if m > 1 one can write f as an n-tuple of functions (f1 , · · · , fm ) each with codomain R.
And it is easy to check that our definitions apply “component-wise”, i.e., f is differentiable
at a point, or has some directional derivative, if and only if all the functions f1 , · · · , fm do.
This is just because a vector y ∈ Rm tends to zero exactly when all its entries do. So we
may as well assume that m = 1 in what follows.
Here is the million dollar question:

If all of the directional derivatives exist at a does the total derivative exist?

When n > 1, no. We shall see some counterexamples to this shortly. But let us keep our
momentum going and prove a partial converse which is true.

1.3. Continuous partial derivatives give differentiability

We shall restrict for simplicity and notational convenience to n = 2, and also m = 1, and
take Ω = R2 . So we are considering functions f : R2 → R. I will write a = (a, b) and
h = (h, k), and f (a, b) rather than f ((a, b)) etc. For our partial converse I will only need
to consider two directions, given of course by the unit vectors (1, 0) and (0, 1). I will write
these directional derivatives when they exist as fx (a, b) and fy (a, b). (They are also denoted
∂f
∂x
(a, b) etc but this requires more effort to write.) We will of course follow convention and
call these special directional derivatives the partial derivatives.
Assume briefly f : R2 → R is differentiable at a point a = (a, b). We have the linear map
L = df (a) : R2 → R, the derivative. And we see from (1.4) that

df (a) : (1, 0) 7→ fx (a, b), df (a) : (0, 1) 7→ fy (a, b).

So writing df (a) as a 1 × 2 matrix with respect to the usual basis (the Jacobian matrix), we
can see it is just

(1.5) df (a) = (fx (a, b), fy (a, b)).

Note this proves that df (a) is unique, since the partial derivatives are uniquely defined (and
the same argument works for general dimension). And this little calculation also inspires the
next proof.

Theorem 1.3.1. Suppose f : R2 → R has continuous partial derivatives. Then f is

differentiable in R2 with derivative at a = (a, b) given by df (a) = (fx (a, b), fy (a, b)).
4 1. DIFFERENTIABILITY IN R2
Proof. Let (a, b) ∈ R2 . Define

L := (fx (a, b), fy (a, b)).

We need to show that

f (a + h, b + k) − f (a, b) − L(h, k)
(1.6) lim = 0.
(h,k)→0 |(h, k)|
Our proof uses the telescoping sum

f (a + h, b + k) − f (a, b) = (f (a + h, b + k) − f (a + h, b)) + (f (a + h, b) − f (a, b)).

Note in each bracket we move only one of the arguments. And by the Mean Value Theorem
in one variable from Prelims Analysis: there exists θh , θk ∈ (0, 1) such that
f (a + h, b + k) − f (a + h, b) = fy (a + h, b + θk k)k
f (a + h, b) − f (a, b) = fx (a + θh h, b)h.
(Note the function f is differentiable in each argument when one fixes the other, i.e., the
partial derivatives exist. That is why the MVT can be applied.) So

f (a + h, b + k) − f (a, b) = fy (a + h, b + θk k)k + fx (a + θh h, b)h.

Also we have (thinking of (h, k) as a column vector so our matrix for L can act)

L(h, k) = fx (a, b)h + fy (a, b)k.

So the numerator in (1.6) is just

(fx (a + θh h, b) − fx (a, b))h + (fy (a + h, b + θk k) − fy (a, b))k

= (fx (a + θh h, b) − fx (a, b), fy (a + h, b + θk k) − fy (a, b)) · (h, k).

This is the dot product of two vectors and we shall apply the Cauchy-Schwarz inequality to
deduce its absolute value is at most
q √
(fx (a + θh h, b) − fx (a, b))2 + (fy (a + h, b + θk k) − fy (a, b))2 × h2 + k 2 .

So to show (1.6) we must show that

q
(fx (a + θh h, b) − fx (a, b))2 + (fy (a + h, b + θk k) − fy (a, b))2 → 0

as (h, k) → 0. But as (h, k) → 0 we have that the argument (a + θh h, b) of fx tends to (a, b)

and since fx is continuous on R2 we get that fx (a + θh h, b) − fx (a, b) → 0. And likewise
(a + h, b + θk k) → (a, b) and by continuity of fy we see the second term in the sum tends to
zero. This completes the proof.
 1.4. EXAMPLE, COUNTEREXAMPLE AND SOME GEOMETRIC INTUITION 5

Note that the partial derivative fx itself is a function from (some subset of) R2 to R. So
we may consider its partial derivatives with respect to x and y, which I will denote fxx and
fxy . In Prelims calculus you often used that fxy = fyx . On problem sheet 1 you will prove a
sufficient condition for this equality to be true, and see that it is not always so.
The proof of Theorem 1.3.1 in fact works for any f : Ω ⊆ R2 → R with continuous partial
derivatives, provided one assumes Ω is open (Definition 1.6.1); that is, every point a ∈ Ω
lies in a disk contained within Ω. One needs this assumption in the proof simply to ensure
f is defined on the line segments implicit in our use of the telescoping sum and Mean Value
Theorem.

1.4. Example, counterexample and some geometric intuition

Let us focus now on n = 1 or 2 and m = 1. So we have a map f : Ω ⊆ R → R or

f : Ω ⊆ R2 → R which is differentiable at a point a. The reason for doing this is that we
can now visualise the graph of the function

Gf = {(x, f (x)) : x ∈ Ω ⊆ R} or Gf = {(x, y, f (x, y)) : (x, y) ∈ Ω ⊆ R2 }

in R2 or R3 near the point (a, f (a)).

Take n = 1 first. Then we have from (1.2) that

f (a + h) = f (a) + Lh + R(h)|h|

where L = fx (a) is the derivative and R(h) → 0 as h → 0. So we recover that the tangent
line f (a) + Lh is a good approximation to the function f (x) near x = a.
Consider now n = 2. We take as above a = (a, b) and h = (h, k), and recall from (1.5)
our derivative df (a) = L is just the matrix (fx (a, b), fy (a, b)). We see now that

(1.7) f (a + h, b + k) = f (a, b) + fx (a, b)h + fy (a, b)k + R(h, k)|(h, k)|.

And now the tangent plane f (a, b) + fx (a, b)h + fy (a, b)k is a good approximation to the
function f (x, y) near (a, b).
Put slightly differently, (1.7) is the “first-order Taylor expansion” around (a, b), as in the
one-variable case. And as R(h, k) → 0 as |(h, k)| → 0, then the remainder R(h, k)|(h, k)|
goes to zero faster than linearly. So the linear function f (a, b) + fx (a, b)h + fy (a, b)k provides
a good numerical approximation near (a, b).
Recall when you first learned calculus, computing the derivative of a function directly
from the definition was rather hard work. The same is true in this higher dimensional case.
Working directly from Definition 1.1.1 would be an effort. But our description of the matrix
L in terms of partial derivatives, as in (1.5), makes life much easier.
6 1. DIFFERENTIABILITY IN R2
Example 1.4.1. Let
p
f (x, y) = 1 − (x2 + y 2 )
and Ω = {(x, y) : x2 + y 2 ⩽ 1} ⊆ R2 . Then obviously our graph Gf is just the top half
of a sphere of radius one with centre the origin. And if we compute the matrix of partial
derivatives (fx , fy ) it is
!
−x −y
p ,p .
1 − (x2 + y 2 ) 1 − (x2 + y 2 )
These are defined and continuous except on the boundary x2 +y 2 = 1 of Ω. So Theorem 1.3.1
tells us the function is differentiable there, and the derivative at such a point (a, b) is just
the evaluation (fx (a, b), fy (a, b)). And this gives an explicit description of the tangent plane
at any such point (which I won’t write down). (Note that on the boundary the function is
not especially badly behaved, it is just that its derivative becomes “infinite”, in the same
way a function like x1/3 on R does at the origin, and our present set-up is not robust enough
to handle this.)

Example 1.4.2. Let f : R2 → R be defined by

(
xy(x+y)
x2 +y 2
when (x, y) ̸= (0, 0)
f (x, y) :=
0 for (x, y) = (0, 0).
For a unit vector n note that f (λn) = λf (n). (That is, the rational function f (x, y) is
homogeneous of degree 1.) Hence
f ((0, 0) + λn) − f (0, 0)
lim = f (n).
λ→0 λ
Thus all the direction derivatives exist at (0, 0), and are just equal to the value of the function
in that direction. But if f were differentiable at (0, 0) then by (1.4) there would exists a
linear map L : R2 → R such that Ln = f (n). And this is not the case, since f is not linear.
So f is not differentiable at (0, 0).

You can use the 3D calculator at [Link]/3d to see the graph of this function,
or any other.

1.5. Continuity and the chain rule

I am sure you all remember the ε − δ definition of continuity from Prelims Analysis II.
Let me just recall it for you, but in a rather unusual manner.
We shall say that a function f : R → R is “not continuous” at a point a ∈ R if

∃ε > 0 such that ∀δ > 0, ∃x ∈ R with |x − a| < δ but |f (x) − f (a)| ⩾ ε.

 1.5. CONTINUITY AND THE CHAIN RULE 7

This seems a pretty good way of capturing that the function f is not continuous in the
intuitive sense, since our statement suggests it “jumps by ε” at a. And you might also
remember how to negate logical statements, by sweeping through them and switching ∀ and
∃. So let’s do that. We find our function is not “not continuous” at a when

∀ε > 0, ∃δ > 0 such that ∀x ∈ R with |x − a| < δ we have |f (x) − f (a)| < ε.

Let us call such a function continuous at a, since writing not not continuous continually
would be tiresome.
I find this a much more intuitive approach to the definition, since really it is the property
of being “not continuous”, i.e. a function jumping, that we have a feeling for and should
capture in a logical manner. But anyway let us stick with convention:

Definition 1.5.1. A function f : Ω ⊆ Rn → Rm is continuous at a ∈ Ω if

∀ε > 0, ∃δ > 0 such that ∀x ∈ Ω with |x − a| < δ we have |f (x) − f (a)| < ε.

Note this equivalent to f (x) − f (a) → 0 as x → a.

Theorem 1.5.2. Assume f : Ω ⊆ Rn → Rm is differentiable at a ∈ Ω. Then f is

continuous at a.

Proof. Let the derivative be L : Rn → Rm . Then from (1.2) we have

f (a + h) − f (a) = Lh + R(h)|h|

where R(h) → 0 as h → 0. Taking h → 0 on the righthand side gives zero, and hence

f (a + h) − f (a) → 0 as h → 0.

You shall see on problem sheet 1 though, that even if all the directional derivatives exist
at a point, the function is not necessarily continuous at that point.
We also note here a useful result which should not be surprising to you.

Theorem 1.5.3 (Chain Rule). Let f : Rn → Rm , g : Rm → Rp and a ∈ Rm . Assume

that f is differentiable at a and g is differentiable at f (a). Let h := g ◦ f . Then h is
differentiable at a with dh(a) = dg(f (a))df (a).

The proof is omitted and non-examinable. It is quite straightforward, but requires some
simple properties on sizes of matrices (that is to say, matrix norms) which we have not
covered. (See also the end of Section 1.7: matrix norms are also important in the proof of
the inverse function theorem.)
8 1. DIFFERENTIABILITY IN R2
1.6. Some topology in R2

In this section we shall introduce some properties of sets in R2 which you will need for the
complex analysis course. We revisit all of these properties in greater depth and generality
later in the course. (Note I will drop the underlines from vectors in most of this section to
reduce clutter on notation.)

1.6.1. Path-connected and connected sets. Let r > 0 and x ∈ R2 . We call

B(x, r) := {y ∈ R2 ||x − y| < r} and B(x, r) := {y ∈ R2 ||x − y| ⩽ r}

the (respectively) open and closed balls of radius r around x.

Definition 1.6.1. We say that a set U ⊆ R2 is open if for all x ∈ U there exists r > 0
such that B(x, r) ⊆ U . We say that U ⊆ R2 is closed if the complement R2 \U is open.
Given a ∈ R2 , we call any open set U ⊆ R2 containing a an open neighbourhood of a.

Notice that ∅ and R2 itself are open. And they are also both closed: in particular, it is
not the case that a set is closed when it is not open. It can be both, or neither.
The collection of all open sets in R2 gives what is called a topology on R2 . That is,
a collection of subsets containing ∅ and R2 which is closed under finite intersections and
arbitrary unions. The properties of sets we shall introduce below depend only upon this
topology, in the following sense. Suppose we used a different notion of distance on R2 which
defined the same collection of open sets. Then we could use that different notion of distance
in our definitions below, and end up defining the same property of sets.
We first look at the notion of paths and path-connectedness.

Definition 1.6.2. Let U ⊆ R2 and a0 , a1 ∈ U . A path in U between a0 and a1 is a

continuous map γ : [0, 1] → U with γ(0) = a0 and γ(1) = a1 . We say that U is path-
connected if for every a0 , a1 ∈ U there exists a path between a0 and a1 .

One thinks of the value γ(t) for t ∈ [0, 1] as continuously tracing the path between a0
and a1 . The significance of paths in complex analysis is that, after identifying R2 and C,
you will integrate complex valued functions along them. (Though you will usually make the
stronger assumption that the map γ is in fact differentiable.)

Definition 1.6.3. Let U ⊆ R2 and a0 , a1 ∈ U . Let γ0 and γ1 be paths in U connecting

a0 and a1 . We say that the paths γ0 and γ1 are homotopic if there exists a continuous map
Γ : [0, 1] × [0, 1] → U with Γ(s, 0) = a0 and Γ(s, 1) = a1 for all s ∈ [0, 1], and Γ(0, t) = γ0 (t)
and Γ(1, t) = γ1 (t) for all t ∈ [0, 1].
 1.6. SOME TOPOLOGY IN R2 9
Here for each s ∈ [0, 1] we get a path Γ(s, t) : [0, 1] → U , and one thinks of the first
path γ0 (t) = Γ(0, t) being continuously deformed into the second γ1 (t) = Γ(1, t) as s moves
between 0 and 1.

Definition 1.6.4. We shall say that U ⊆ R2 is simply-connected if it is path-connected

and given any points a0 , a1 ∈ U then any two paths in U between a0 and a1 are homotopic.

So “up to homotopy” there is a unique path between any a0 and a1 in a simply-connected

set. We can put this another way. For any point a ∈ U there is the trivial path from a to a
given by γ(t) = a for all t ∈ [0, 1]. Then a path-connected set U will be simply-connected
exactly when for every a ∈ U any path between a and a can be “shrunk to a point”, i.e. is
homotopic to this trivial path. (That this is equivalent to the definition is a little work to
prove.)
The significance of all this for complex analysis is that it will turn out the path integrals
only depend on paths up to homotopy (which makes life much easier).

Example 1.6.5. The open unit ball B(0, 1) is simply-connected. Given any two points
a0 , a1 ∈ B(0, 1) and any two paths connecting them, one can visualise deforming one into the
other. By contrast, the open “doughnut” B(0, 3)\B(0, 1) is not simply-connected. Following
the circle of radius 2 anticlockwise from (2, 0) gives a path from (2, 0) to itself. But it is
intuitively clear that there is no way to shrink this path to a point. (I do not claim either
of these visually properties is especially easy to prove, though you will show the first on
problem sheet 1.)

Here is a useful property of path-connectedness.

Theorem 1.6.6. Let U ⊆ R2 be path-connected and f : U → R2 continuous. Then f (U )

is path-connected.

Proof. Let v0 , v1 ∈ f (U ). Then v0 = f (u0 ) and v1 = f (u1 ) for some u0 , u1 ∈ U . Since U is

path-connected there exists a path γ joining u0 and v0 . Then f ◦ γ gives the required path
between v0 and v1 . (I will leave it to you to prove directly from Definition 1.5.1 that the
composition of two continuous maps is continuous.)

There is a slightly less intuitive related notion in topology, namely that a set is connected.

Definition 1.6.7. Let U ⊆ R2 . We shall say that U is connected if the following is true.
For any open sets A and B in R2 with U ⊆ A ∪ B and (A ∩ B) ∩ U = ∅, then either U ⊆ A
or U ⊆ B.

We shall look closely at this notion later. Path-connected sets are always connected
(Theorem 8.3.1). It turns out that open connected sets in R2 are always path-connected (by
10 1. DIFFERENTIABILITY IN R2
Theorem 8.3.2). So in complex analysis when you read an “open connected set in C” this is
just the same as it being an open path-connected set in R2 .

Example 1.6.8. The space R2 itself is connected, since it is obviously path-connected.

This implies that the only sets in R2 which are both open and closed are ∅ and R2 . For
suppose A was such a set, with A ̸= ∅ or R2 . Then B := R2 \A is also open with B ̸= R2 .
And R2 = A ∪ B with A ∩ B = ∅ but R2 ̸⊆ A or B, contradicting the fact R2 is connected.

And again the property of being connected is preserved under continuous maps (Lemma
8.1.6).

Theorem 1.6.9. Let U ⊂ R2 be connected and f : U → R2 be continuous. Then f (U ) is

connected.

1.6.2. Compactness. We now turn to a quite different but equally important notion.

Definition 1.6.10. Let U ⊆ R2 . We shall say that U is compact if given any collection
{Vi : i ∈ I} of open sets with U ⊆ ∪i∈I Vi , there exists a finite subset J ⊆ I such that
U ⊆ ∪j∈J Vj .

The mantra to remember here is “U is compact exactly when every open covering of
U has a finite subcovering”. This is not at all an intuitive notion, but the essential point
really is that it is a good proxy for finiteness when dealing with infinite sets. That is to say,
compactness allows you to apply to infinite sets certain arguments you would normally only
apply to finite sets. Fortunately in R2 we can relate compactness to more intuitive notions.

Definition 1.6.11. We shall say that U ⊆ R2 is bounded if there exists x ∈ R2 and r > 0
such that U ⊆ B(x, r).

It turns out that in R2 being compact is the same as being closed and bounded (Theorem
10.3.3). It is important to stress though that these two notions are not the same in more
general settings.
You will recall from Prelims that a continuous function f : U → R on a non-empty closed
and bounded set is bounded and achieves that bound. This is a special case of a much more
general theorem in topology, that the image of a compact set under a continuous map is
compact. In particular, we also have the following.

Theorem 1.6.12. Let f : U ⊆ R2 → R be a continuous function where U is a non-empty

compact set. Then f (U ) is bounded and achieves that bound. That is, there exists b ∈ R and
a ∈ U such that |f (x)| ⩽ b for all x ∈ U and furthermore |f (a)| = b.

Proof. We will derive this from some theorems proved later in the course. (So you should
defer reading this for now.) Since U is compact it is sequentially compact (Proposition
 1.7. THE INVERSE FUNCTION THEOREM 11

10.2.1). Note |f | is also continuous, since it is a composition of two continuous functions.

Hence |f (U )| is sequentially compact (Lemma 9.3.1), and so closed and bounded (Lemma
9.2.1). As |f (U )| is non-empty and bounded it has a supremum, and since it is closed this
supremum lies in |f (U )| (Corollary 6.2.5). (Note the supremum is a limit point of |f (U )|.)
That is to say, there exists b ∈ R and a ∈ U such that |f (x)| ⩽ b for all x ∈ U and
furthermore |f (a)| = b.

1.6.3. Notions of distance in R2 . For x = (x1 , x2 ) and y = (y1 , y2 ) ∈ R2 define

p
d2 (x, y) := |x − y| = (x1 − y1 )2 + (x2 − y2 )2 . This is just the usual Euclidean notion of
distance which we have been using throughout. But it is not the only “sensible” notion of
distance on R2 . There are many, some less sensible than others. By “sensible” we mean
primarily that the notion of distance defines a metric, a term to be made precise in the next
chapter.
For example we could take

d∞ (x, y) := max(|x1 − y1 |, |x2 − y2 |)

and then define the open ball of radius r > 0 around a point a ∈ R2 to be

B∞ (x, r) := {y ∈ R2 : d∞ (x, y) < r}.

This ball would not be very good for football (rather square: draw one). But it would be
perfectly good for doing topology as it defines the same open sets as our usual Euclidean
metric. The reason being simply that given any point x and r > 0, then there exists r′ < r
such that B∞ (x, r′ ) ⊆ B(x, r), and vice-versa. In other words, square balls can be fitted
within round ones, and vice-versa.
Here is a notion of distance on R2 which defines a metric, but which is good for neither
football nor topology (in the sense that like all metrics it defines a topology on R2 , but this
one is not of any use in geometry).

Example 1.6.13. Define

(
|x2 − y2 | if x1 = y1
δ(x, y) :=
|x1 − y1 | + |x2 | + |y2 | ̸ y1 .
if x1 =
This metric is effectively the distance between two points in a library of shelves parallel to
the y-axis and with a connecting corridor along the x-axis only.

1.7. The Inverse function theorem

For U, V ⊆ R2 we write C 1 (U, V ) for the set of functions f : U → V with continuous

partial derivatives on U . Note in particular such functions are differentiable on U .
12 1. DIFFERENTIABILITY IN R2
Theorem 1.7.1 (Inverse Function Theorem in R2 ). Let Ω ⊆ R2 be open, f ∈ C 1 (Ω, R2 )
and a ∈ Ω. Suppose that df (a) : R2 → R2 is invertible. Then there exists an open neigh-
bourhood U of a such that f (U ) is open and f : U → f (U ) is a bijection with inverse
f −1 ∈ C 1 (f (U ), U ). Moreover, we have df −1 (f (a)) = df (a)−1 .

The inverse function theorem allows you to “locally invert” functions on R2 , in the same
way that you can “globally invert” (invertible) linear maps on R2 .

Example 1.7.2. We show that the equations u = x3 y + y 2 , v = ln(x + y) have a unique

solution x = x(u, v), y = y(u, v) on some neighbourhood of (6, ln(3)) with x(6, ln(3)) = 1
and y(6, ln(3)) = 2.
Define f (x, y) = (u, v) = (x3 y + y 2 , ln(x + y)) (for say x + y > 0). Then to solve for x, y
in terms of (u, v) it is enough to show f −1 exists, for then we have (x, y) = f −1 (u, v). Now
the Jacobian matrix !
3x2 y x3 + 2y
df (x, y) =
1/(x + y) 1/(x + y)
is continuous in a neighbourhood of (1, 2) and invertible there, and so the Inverse Function
Theorem applies.

The intuition behind the Inverse Function Theorem is very simple. At a point a + h near
a the value f (a + h) is well approximated by the value f (a) + df (a)h. If it were in fact equal
to this, then one could invert the function on a point f (a) + k near f (a) simply by mapping
this to a + df (a)−1 k, provided the derivative is invertible. And the inverse function theorem
tells us this is almost true. That is, there is an inverse near f (a) which is well approximated
by this linear map. Unfortunately the proof itself is rather complicated, and omitted from
the course (and non-examinable).
This complication is mainly notational though, and it is rather straightforward to explain
the main steps in the one variable case. And these carry over almost exactly to higher
dimensions. So let’s look at the existence of the inverse. (This is not the simplest proof in
dimension one, but rather one that can be translated easily to the general case and suggests
a useful numerical algorithm.)
Let f : R → R be continuously differentiable at a ∈ R with f ′ (a) ̸= 0. Without any loss
of generality we may assume a = 0, f (0) = 0 and f ′ (0) = 1 (think about why). Consider
the function

(1.8) ψ(x) = x − f (x).

Note that ψ(0) = ψ ′ (0) = 0 and since ψ ′ (x) is continuous at 0, there exists r > 0 such that
1
(1.9) |x| ⩽ r ⇒ |ψ ′ (x)| ⩽ .
2
 1.7. THE INVERSE FUNCTION THEOREM 13

Hence by the Mean Value Theorem in one variable we see

1 r
(1.10) |x| ⩽ r ⇒ |ψ(x)| ⩽ |x| ⩽ .
2 2
So now let us fix v0 ∈ R with |v0 | ⩽ r/2. Define ϕ(v0 , x) : R → R by

ϕ(v0 , x) := x − f (x) + v0 = ψ(x) + v0 .

Then by (1.10)

|x| ⩽ r ⇒ |ϕ(v0 , x)| = |x − f (x) + v0 | ⩽ |ψ(x)| + |v0 | ⩽ r.

Thus ϕ(v0 , x) maps the closed ball B(0, r) to itself. Also for any x, x′ ∈ B(0, r) using (1.9)
and the Mean Value Theorem again,
1
|ϕ(v0 , x) − ϕ(v0 , x′ )| = |ψ(x) − ψ(x′ )| ⩽ |x − x′ |.
2
So we have a contraction mapping and by Prelims analysis / constructive mathematics, there
exists a unique “fixed point” x∞ ∈ B(0, r) such that ϕ(v0 , x∞ ) = x∞ , that is f (x∞ ) = v0 .
(The choice of notation x∞ rather than x0 will become clearer below: numerically one can
obtain it as the limit of a sequence starting from an arbitrary x0 .)
Note that if |v0 | < r/2 then we have |v0 | ⩽ r′ /2 for some r′ < r. And replacing r by r′ in
the above we have x∞ ∈ B(0, r′ ) ⊂ B(0, r). So in fact we have a map on open sets

g : B(0, r/2) → U := f −1 (B(0, r/2)) ∩ B(0, r), v0 7→ x∞ .

And this gives our inverse g : f (U ) → U with f ◦ g = If (U ) and g ◦ f = IU .

Let’s turn now to an example in the two-dimensional case, and try and understand our
partial proof in this context.
Example 1.7.3. Let f : R2 → R2 be defined by

f (x, y) := (cos(y) + x − 1, sin(x) + y − cos(y) − x + 1).

Then f (0, 0) = (0, 0). The Jacobian matrix L here at the point (x, y) is
!
1 − sin(y)
df (x, y) = .
cos(x) − 1 1 + sin(y)
Note that df (0, 0) = I2 . So we are in a situation analogous to that in our partial proof for
the one-dimensional case.
Inspired by our proof above, let us try to invert the function numerically near the point
(0, 0) = f (0, 0). So pick (v0 , w0 ) ∈ B((0, 0), r/2), and define ϕ(v0 , w0 , x, y) : R2 → R2 by

ϕ(v0 , w0 , x, y) := (x, y) − f (x, y) + (v0 , w0 )

= (x − (cos(y) + x − 1) + v0 , y − (sin(x) + y − cos(y) − x + 1) + w0 ).

14 1. DIFFERENTIABILITY IN R2
On B((0, 0), r) this should define a contraction mapping. And so iteration from any starting
point (x0 , y0 ) ∈ B((0, 0), r) using

(xn+1 , yn+1 ) := ϕ(v0 , w0 , xn , yn )

will converge to a fixed point ϕ(v0 , w0 , x∞ , y∞ ) = (x∞ , y∞ ). And we will have f (x∞ , y∞ ) =
(v0 , w0 ).
But the problem is we have no idea on the value for r! In the one-dimensional case we
chose r small enough so that the derivative of the function ψ(x) = x − f (x) was bounded by
1/2 for |x| ⩽ r. And this is where we used the continuity of the derivative. So it appears in
this case we need to bound the size of the matrix df (x, y) by 1/2 on some disk around (0, 0),
using the continuity of its entries. To do this we need a notion of size on matrices, or linear
maps. And this is really the only ingredient missing if we want to generalise our proof above
to higher dimension. We will not pursue this further, but just end with some calculations.
Experimentally take (v0 , w0 ) = (0.5, 0.5) and start from (x0 , y0 ) = (0, 0). After 5 iterations
we get

(x5 , y5 ) = (0.5898 . . . , 0.4462 . . .) and f (x5 , y5 ) = (0.4919 . . . , 0.5105 . . .).

And more iterations give higher precision. And this works well for (v0 , w0 ) = (0.6, 0.6), and
so on. But somewhere between taking (v0 , w0 ) = (0.79, 0.79) and (0.8, 0.8) things go wrong.
I have not looked closely enough at the example to see why: perhaps you would like to figure
this out.
 CHAPTER 2

Metric spaces

2.1. The real numbers and the axiom of choice

The real numbers. I will assume familiarity with the real numbers R as discussed at some
length in the Prelims course Analysis I. I will not repeat the long list of axioms for the real
numbers here. The most important properties we shall need are

• Any non-empty, bounded subset S ⊆ R has a least upper bound sup(S), which
is a real number c such that x ⩽ c for all x ∈ S, and such that if c′ is any other
number with this property then c′ ⩾ c;
• Similarly, any non-empty, bounded subset S ⊆ R has a greatest lower bound inf(S);
• (Bolzano-Weierstrass) Any bounded sequence of real numbers has a convergent
subsequence;
• Any Cauchy sequence of real numbers converges.

It might be a good idea to remind yourself of the precise meaning of these statements
now, though we shall be going over the last two points in a more general context later in the
course.
The Prelims course Analysis I assumed that the real numbers exist. This is not, by any
means, obvious! We will also assume they exist.
The axiom of choice. The following statement, used for example in the proof of Corollary
6.1.5, seems very uncontroversial: given nonempty subsets S1 , S2 , . . . of some set X, we may
find a sequence (xn )∞
n=1 with xn ∈ Sn for all n. One might have thought that this is the most
trivial induction imaginable: pick x1 ∈ S1 , then pick x2 ∈ S2 , and so on. This does indeed
show that there are x1 , . . . , xN with xn ∈ Sn for n = 1, . . . , N , but it does not show the
infinitary statement about the existence of a sequence. In fact, the existence of a sequence
(xn )∞
n=1 with xn ∈ Sn for all n has the status of a separate axiom of mathematics, called the
axiom of countable choice.
You can learn much more about this and, more particularly, the axiom of choice itself
in the course B1.2: Set Theory. However, the introduction of the Wikipedia page on the
Axiom of Choice is a good read.
15
16 2. METRIC SPACES

2.2. The definition of a metric space

One of the key definitions of Analysis I was that of the continuity of a function. Recall
that if f : R → R is a function, we say that f is continuous at a ∈ R if, for any ε > 0, we
can find a δ > 0 such that if |x − a| < δ then |f (x) − f (a)| < ε.
Now consider what it is about real numbers that we need in order for this definition to
make sense: Really we just need, for any pair of real numbers x1 and x2 , a measure of
the distance between them. Thus we should be able to talk about continuous functions
f : X → X on any set X provided it is equipped with a notion of distance. Even more
generally, provided we have prescribed a notion of distance on two sets X and Y , we should
be able to say what it means for a function f : X → Y to be continuous. In order to make
this precise, we will therefore need to give a mathematically rigorous definition of what a
“notion of distance” on a set X should be. This is the concept of a metric space.

Definition 2.2.1. Let X be a set. Then a distance function on X is a function d : X ×

X → R with the following properties:

(i) (positivity) d(x, y) ⩾ 0 and d(x, y) = 0 if and only if x = y;

(ii) (symmetry) d(x, y) = d(y, x);
(iii) (triangle inequality) if x, y, z ∈ X then we have d(x, z) ⩽ d(x, y) + d(y, z).

The pair (X, d) consisting of a set X together with a distance function d on it is called a
metric space.

Remark. Often we will not be quite so formal, and will refer to X (rather than the pair
(X, d)) as a metric space. However, it is important to note that the same space X can have
many different distances on it, and in fact that different distances on the same space X can
have wildly differing properties.
Occasionally, we will be more formal, for instance when we have two metric spaces (X, dX )
and (Y, dY ) and wish to make it clear which distance we are talking about.
The axioms that a distance function d is required to satisfy are very basic, and one feels
that any “reasonable” notion of distance ought to satisfy these properties. This, coupled
with the fact that using just these axioms one can develop a satisfactory theory of continuity
of functions – as well as many other things – is the point of the definition.
Before moving on, let us record one very simple but useful equivalent form of the triangle
inequality, sometimes (but not by me) known as the reverse triangle inequality.

Lemma 2.2.2. Let x, y, z be points in a metric space. Then we have |d(x, y) − d(x, z)| ⩽
d(y, z).
 2.3. SOME EXAMPLES OF METRIC SPACES 17

Proof. This is two inequalities in one, namely the inequality d(x, y) − d(x, z) ⩽ d(y, z), and
the inequality d(x, z) − d(x, y) ⩽ d(y, z). Both are instances of (in fact, equivalent to) the
triangle inequality.

2.3. Some examples of metric spaces

In this section we look at some examples of metric spaces. A very basic example is that
of the real numbers.

Example 2.3.1. Take X = R and d(x, y) = |x − y|.

Let us generalise this to higher dimensions. In fact, there is no “obvious” generalisation.

Here are several natural ones.

Example 2.3.2. Take X = Rn . Then each of the following functions define metrics on
X.
n
X
d1 (v, w) = |vi − wi |;
i=1
n
X
1/2
d2 (v, w) = (vi − wi )2
i=1

d∞ (v, w) = max |vi − wi |.

i∈{1,2,...,n}

These are called the ℓ1 - (“ell one”), ℓ2 - (or Euclidean) and ℓ∞ -distances respectively. Of
course, the Euclidean distance is the most familiar one.
The proof that each of d1 , d2 , d∞ defines a distance is mostly very routine, with the
exception of proving that d2 , the Euclidean distance, satisfies the triangle inequality. To
establish this, recall that the Euclidean norm ∥v∥2 of a vector v = (v1 , . . . , vn ) ∈ Rn is
n
X
1/2
∥v∥2 := |vi |2 = ⟨v, v⟩1/2 ,
i=1

where the inner product is given by

n
X
⟨v, w⟩ = vi w i .
i=1

Then d2 (v, w) = ∥v − w∥2 , and so the triangle inequality is the statement that

∥u − w∥2 ⩽ ∥u − v∥2 + ∥v − w∥2 .

This follows immediately by taking x = u − v and y = v − w in the following lemma.

Lemma 2.3.3. If x, y ∈ Rn then ∥x + y∥2 ⩽ ∥x∥2 + ∥y∥2 .

18 2. METRIC SPACES

Proof. Since ∥v∥2 ⩾ 0 for all v ∈ Rn the desired inequality is equivalent to

∥x + y∥22 ⩽ ∥x∥22 + 2∥x∥2 ∥y∥2 + ∥y∥22 .

But since ∥x + y∥22 = ⟨x + y, x + y⟩ = ∥x∥22 + 2⟨x, y⟩ + ∥y∥22 , this inequality is immediate

from the Cauchy-Schwarz inequality, that is to say the inequality |⟨x, y⟩| ⩽ ∥x∥2 ∥y∥2 .

The next example is rather a routine and trivial one. However, it behaves very differ-
ently to the Euclidean examples and can often provide counterexamples to over-optimistic
conjectures based on geometric intuition.

Example 2.3.4 (Discrete metric). Let X be an arbitrary set. The discrete metric on a
set X is defined as follows: (
1, if x ̸= y
d(x, y) =
0, if x = y
The axioms for a distance function are easy to check.

Now we turn to some metrics which come up very naturally in diverse areas of mathe-
matics. Our first example is critical in number theory, and also serves to show that metrics
need not conform to one’s most naı̈ve understand of “distance”.

Example 2.3.5 (2-adic metric). Let X = Z, and define d(x, y) to be 2−m , where 2m is the
largest power of two dividing x − y. The triangle inequality holds in the following stronger
form, known as the ultrametric property:

d(x, z) ⩽ max(d(x, y), d(y, z)).

Indeed, this is just a rephrasing of the statement that if 2m divides both x − y and y − z,
then 2m divides x − z.
This metric is very unlike the usual distance. For example, d(999, 1000) = 1, whilst
d(0, 1000) = 81 !
The role of 2 can be replaced by any other prime p, and the metric may also be extended
in a natural way to the rationals Q.

Metrics are also ubiquitous in graph theory:

Example 2.3.6 (path metric). Let G be a graph, that is to say a finite set of vertices V
joined by edges. Suppose that G is connected, that is to say that there is a path joining any
pair of distinct vertices. Define a distance d as follows: d(v, v) = 0, and d(v, w) is the length
of the shortest path from v to w. Then d is a metric on V , as can be easily checked.

They also come up in group theory:

 2.4. NORMS 19

Example 2.3.7 (Word metric). Let G be a group, and suppose that it is generated by
elements a, b and their inverses. Define a distance on G as follows: d(v, w) is the minimal k
such that v = wg1 · · · gk , where gi ∈ {a, b, a−1 , b−1 } for all i.

When G is finite, the word metric is a special case of the path metric – you may wish to
think about why.
There are many metrics with a prominent position in computer science, for instance:

Example 2.3.8 (Hamming distance). Let X = {0, 1}n (the boolean cube), the set of all
strings of n zeroes and ones. Define d(x, y) to be the number of coordinates in which x and
y differ.

Remark. In fact, one can if desired see {0, 1}n as a subset of Rn , and in this case d is
the restriction of one of the metrics already considered in Example 2.3.2 (you may care to
contemplate which one).
It hardly need be said that metrics are ubiquitous in geometry.

Example 2.3.9 (Projective space). Consider the set P(Rn ) of one-dimensional subspaces
of Rn , that is to say lines through the origin). One way to define a distance on this set is to
take, for lines L1 , L2 , the distance between L1 and L2 to be
s
|⟨v, w⟩|2
d(L1 , L2 ) = 1 − ,
∥v∥2 ∥w∥2
where v and w are any non-zero vectors in L1 and L2 respectively. It is easy to see this is
independent of the choice of vectors v and w. The Cauchy-Schwarz inequality ensures that d
is well-defined, and moreover the criterion for equality in that inequality ensures positivity.
The symmetry property is evident, while the triangle inequality is left as an exercise.
It is useful to think of the case when n = 2 here, that is, the case of lines through the
origin in the plane R2 . The distance between two such lines given by the above formula is
then sin(θ) where θ is the angle between the two lines (another exercise).

2.4. Norms

In Example 2.3.2, we looked at three examples of metrics on Rn . They are all, as it turns
out, induced from norms. This is an important notion which we now develop in its general
context.

Definition 2.4.1 (Norms). Let V be any vector space (over the reals). A function
∥ · ∥ : V → [0, ∞) is called a norm if the following are all true:
• ∥x∥ = 0 if and only if x = 0;
• ∥λx∥ = |λ|∥x∥ for all λ ∈ R, x ∈ V ;
20 2. METRIC SPACES

• ∥x + y∥ ⩽ ∥x∥ + ∥y∥ whenever x, y ∈ V .

Given a norm, it is very easy to check that d(x, y) := ∥x−y∥ defines a metric on V . Indeed,
we have already seen that when V = Rn , ∥ · ∥2 is a norm (and so the name “Euclidean norm”
is appropriate) and we defined d2 (x, y) = ∥x − y∥2 .
As we mentioned, the other metrics in Example 2.3.2 also come from norms. Indeed, d1
comes from the ℓ1 -norm n
X
∥x∥1 := |xi |,
i=1
whilst d∞ comes from the ℓ∞ -norm

∥x∥∞ := max |xi |.

i=1,...,n

As the notation suggests, these are special cases of a more general family of norms, the
ℓp -norms n
X
1/p
∥x∥p := |xi |p .
i=1
It is true (but we shall not prove it in this course) that these do indeed define norms for
1 ⩽ p < ∞. Moreover,
lim ∥x∥p = ∥x∥∞ ,
p→∞
∞
which is how the ℓ -norm comes to have its name.
The principle of turning norms into metrics is important enough that we state it as a
lemma in its own right.

Lemma 2.4.2. Let V be a vector space over the reals, and let ∥ · ∥ be a norm on it. Define
d : V × V → [0, ∞) by d(x, y) := ∥x − y∥. Then (V, d) is a metric space.

It is important to note that the converse is very far from true. For instance, the discrete
metric does not arise from a norm. All metrics arising from a norm have the translation
invariance property d(x + z, y + z) = d(x, y), as well as the scalar invariance d(λx, λy) =
|λ|d(x, y), neither of which are properties of arbitrary metrics. Conversely one can show that
a metric with these two additional properties does come from a norm, an exercise we leave
to the reader (Hint: the norm must arise as ∥v∥ = d(v, 0)).
We call a vector space endowed with a norm ∥ · ∥ a normed space. Whenever we talk
about normed spaces it is understood that we are also thinking of them as metric spaces,
with the metric being defined by d(v, w) = ∥v − w∥.
Note that we do not assume that the underlying vector space V is finite-dimensional.
Here are some examples which are not finite-dimensional (whilst we do not prove that they
are not finite-dimensional here, it is not hard to do so and we suggest this as an exercise).
 2.5. NEW METRIC SPACES FROM OLD ONES 21

Example 2.4.3 (ℓp spaces). Let

X
ℓ1 = {(xn )∞
n=1 : |xn | < ∞}
n⩾1
X
ℓ2 = {(xn )∞
n=1 : x2n < ∞}
n⩾1

ℓ∞ = {(xn )∞
n=1 : sup |xn | < ∞}.
n∈N
P
The sets ℓ1 , ℓ2 , ℓ∞ are all real vector spaces, and moreover ∥(xn )∥1 = n⩾1 |xn |, ∥(xn )∥2 =
2 1/2
P

n⩾1 xn , ∥(xn )∥∞ = supn∈N |xn | define norms on ℓ1 , ℓ2 and ℓ∞ respectively. Note that
ℓ2 is in fact an inner product space where
X
⟨(xn ), (yn )⟩ = xn y n ,
n⩾1

(the fact that the right-hand side converges if (xn ) and (yn ) are in ℓ2 follows from the
Cauchy-Schwarz inequality).

The space ℓ2 is known as Hilbert space and it is of great importance in mathematics.

2.5. New metric spaces from old ones

Subspaces. Suppose that (X, d) is a metric space and let Y be a subset of X. Then the
restriction of d to Y × Y gives Y a metric so that (Y, d|Y ×Y ) is a metric space. We call Y
equipped with this metric a subspace.
The word “subspace” is rather overused in mathematics. If X = Rn , so that X is a vector
space, then Y need not be a vector subspace – it is just a subset of X.
Let us give an example of a subspace of a metric space. If X = R, we could take Y = [0, 1],
for instance, or Y = Q (the rationals) or Y = Z (the integers). (It would be perverse to
define the usual metric on Z or on Q by restricting from X = R. Indeed, the metric space
(X, d) with X = Z and d(x, y) := |x − y| is a much more basic object than R.)

Product spaces. If (X, dX ) and (Y, dY ) are metric spaces, then it is natural to try to make
X × Y into a metric space. One method is as follows: if x1 , x2 ∈ X and y1 , y2 ∈ Y then we
set
p
dX×Y ((x1 , y1 ), (x2 , y2 )) = dX (x1 , x2 )2 + dY (y1 , y2 )2
The use of the square mean on the right, rather than the max or the sum, is appealing since
then the product R × R becomes the space R2 with the Euclidean metric. However, either
of those alternative definitions results in a metric which is equivalent, in the sense made
precise in Section 4.3.

Lemma 2.5.1. With notation as above, dX×Y gives a metric on X × Y .

22 2. METRIC SPACES

Proof. Positivity and symmetry are obvious. Less clear is the triangle inequality. We need
to prove that
p p
dX (x1 , x3 )2 + dY (y1 , y3 )2 + dX (x3 , x2 )2 + dY (y3 , y2 )2 ⩾
p
(2.1) dX (x1 , x2 )2 + dY (y1 , y2 )2 .

To make this appear less frightening, write a1 = dX (x2 , x3 ), a2 = dX (x1 , x3 ), a3 = dX (x1 , x2 )

and similarly b1 = dY (y2 , y3 ), b2 = dY (y1 , y3 ) and b3 = dY (y1 , y2 ). Thus we want to show
q q q
(2.2) a22 + b22 + a21 + b21 ⩾ a23 + b23 .

To prove this, note that from the triangle inequality we have a1 + a2 ⩾ a3 , b1 + b2 ⩾ b3 .

Squaring and adding gives

a21 + b21 + a22 + b22 + 2(a1 a2 + b1 b2 ) ⩾ a23 + b23 .

By Cauchy-Schwarz, q q
a1 a2 + b1 b2 ⩽ a1 + b1 a22 + b22 .
2 2

Substituting this into the previous line gives precisely the square of (2.2), and (2.1) follows.

2.6. Balls and boundedness

Definition 2.6.1 (Balls). Let X be a metric space. If a ∈ X and ε > 0 then we define
the open ball of radius ε to be the set

B(a, ε) = {x ∈ X : d(x, a) < ε}.

Similarly we defined the closed ball of radius ε about a to be the set

B(a, ε) = {x ∈ X : d(x, a) ⩽ ε}.

Thus when X = R3 with the Euclidean metric we see that B(0, 1) really is what we
understand geometrically as a ball (minus its boundary, the unit sphere), whilst B(0, 1)
contains the unit sphere and everything inside it.
We caution that this intuitive picture of the closed ball being the open ball “together with
its boundary” is totally misleading in general. For instance, in the discrete metric on a set
X, the open ball B(a, 1) contains only the point a, whereas the closed ball B(a, 1) is the
whole of X.

Definition 2.6.2. Let X be a metric space, and let Y ⊆ X. Then we say that Y is
bounded if Y is contained in some open ball.

Lemma 2.6.3. Let X be a metric space and let Y ⊆ X. Then the following are equivalent.
 2.6. BALLS AND BOUNDEDNESS 23

(i) Y is bounded;
(ii) Y is contained in some closed ball;
(iii) The set {d(y1 , y2 ) : y1 , y2 ∈ Y } is a bounded subset of R.

Proof. That (i) implies (ii) is totally obvious. That (ii) implies (iii) follows immediately
from the triangle inequality. Finally, suppose Y satisfies (iii). Then there is some K such
that d(y1 , y2 ) ⩽ K whenever y1 , y2 ∈ Y . If Y is empty, it is certainly bounded. Otherwise,
let a ∈ Y be an arbitrary point. Then Y is contained in B(a, r) where r = K + 1.
 CHAPTER 3

Limits and continuity

The main purpose of introducing the idea of a metric space is that many notions familiar
over R, such as those of limit and continuous function, can be extended to metric spaces,
and theorems about them proven in that context.

3.1. Basic definitions and properties

Definition 3.1.1 (Limit). Suppose that (xn )∞ n=1 is a sequence of elements of a metric
space (X, d). Let x ∈ X. Then we say that xn → x, or that limn→∞ xn = x, if the following
is true. For every ε > 0, there is an N such that d(xn , x) < ε for all n ⩾ N .

Let us bolster this definition with a couple of easy remarks. First, it is quite possible
and indeed usual for a sequence (xn )∞ n=1 to have no limit. Take, for instance, the sequence
(0, 1, 0, 1, 0, 1, . . . ) in R. Second, if the limit does exist then it is unique. To see this, suppose
that xn → a and xn → b, but that a ̸= b. Let δ := d(a, b). Then, taking ε = δ/2 in the
definition of limit, we see that for n sufficiently large we have d(xn , a), d(xn , b) < δ/2. But
then the triangle inequality yields

δ = d(a, b) ⩽ d(xn , a) + d(xn , b) < δ,

a contradiction.

Definition 3.1.2 (Continuity). Let (X, dX ) and (Y, dY ) be metric spaces. We say a
function f : X → Y is continuous at a ∈ X if for any ε > 0 there is a δ > 0 such that for
any x ∈ X with dX (a, x) < δ we have dY (f (x), f (a)) < ε.
We say f is continuous if it is continuous at every a ∈ X.

Although we will not come across it all that much in this course, it is important to note
that the definition of uniform continuity may be extended to metric spaces as well. As for
real functions, the idea is that “δ should depend only on ε”.

Definition 3.1.3 (Uniform continuity). Let (X, dX ) and (Y, dY ) be metric spaces. We
say a function f : X → Y is uniformly continuous if for any ε > 0 there is a δ > 0 such that
for any x, y ∈ X with dX (x, y) < δ we have dY (f (x), f (y)) < ε.

As for functions on the reals, one may also phrase the definition of continuity in terms of
limits.
25
26 3. LIMITS AND CONTINUITY

Lemma 3.1.4. Let f : X → Y be a function between metric spaces. Then f is continuous

at a if and only if the following is true: for any sequence (xn )∞
n=1 with limn→∞ xn = a, we
have limn→∞ f (xn ) = f (a).

Proof. Suppose first that f is continuous at a. Let ε > 0. Then there is a δ > 0 such that
for all x ∈ X with d(x, a) < δ we have d(f (x), f (a)) < ε. Now let (xn )∞
n=1 be a sequence with
limit a. So, by the definition of limit, there is an N > 0 such that d(a, xn ) < δ for all n ⩾ N .
But then for all n ⩾ N we see that d(f (a), f (xn )) < ε, so indeed limn→∞ f (xn ) = f (a) as
required.
For the other direction we show the contrapositive. Suppose f is not continuous at a.
Then there is an ε > 0 such that for all δ > 0 there is some x ∈ X with d(x, a) < δ and
d(f (x), f (a)) ⩾ ε. Taking δ = 1/n, we see that for each n there is some xn ∈ X with
d(xn , a) < 1/n and d(f (xn ), f (a)) ⩾ ε. Therefore lim xn = a, but lim f (xn ) ̸= f (a).

3.2. Function spaces

A great deal of power comes from considering the set of all functions on a space satisfying
some property, such as continuity, as a metric space in its own right. In this section we
consider some important examples of such spaces.
We begin with the space of bounded real-valued functions on a set X. At this stage we
assume nothing about X.

Definition 3.2.1. If X is any set we define B(X) to be the space of functions f : X → R

for which f (X) = {f (x) : x ∈ X} is bounded. If f ∈ B(X), define ∥f ∥∞ = supx∈X |f (x)|.

Lemma 3.2.2. For any set X, B(X) is a vector space, and ∥ · ∥∞ is a norm.

We leave the proof as an easy exercise.

Now we turn to the space of continuous real-valued functions, C(X). To make sense of
what this means we now need X to be a metric space.

Definition 3.2.3. Let X be a metric space. Then we write C(X) for the space of all
continuous functions f : X → R.

Lemma 3.2.4. The space C(X) is a vector space over R, with pointwise addition and
multiplication by scalars.

Proof. One must check that C(X) is closed under addition and scalar multiplication. We
do the case of addition; scalar multiplication is left as an (easy) exercise.
Suppose that f, g ∈ C(X), and let ε > 0. Let a ∈ X.
Since f is continuous at a, there is some δ1 such that d(x, a) < δ1 implies |f (x) − f (a)| <
ε/2.
 3.2. FUNCTION SPACES 27

Since g is continuous at a, there is some δ2 such that d(x, a) < δ2 implies |g(x) − g(a)| <
ε/2.
Take δ = min(δ1 , δ2 ). Then, if d(x, a) < δ we have

|(f + g)(x) − (f + g)(a)| = |f (x) + g(x) − f (a) − g(a)|

⩽ |f (x) − f (a)| + |g(x) − g(a)|
< ε/2 + ε/2 = ε.

Therefore f + g is continuous at a.

In general, we certainly do not have B(X) ⊆ C(X), and unless X is special we do not
have C(X) ⊆ B(X). We will discuss situations in which this is true later on; you will already
be familiar with a nontrivial example, namely that C([0, 1]) ⊆ B([0, 1]), that is to say all
continuous functions on [0, 1] are bounded.

Definition 3.2.5. Let X be a metric space. Write Cb (X) := C(X) ∩ B(X) for the space
of continuous, bounded functions on X. Since Cb (X) is a subspace of B(X), it inherits the
norm ∥f ∥∞ = supx∈X |f (x)|, and we may define a metric d∞ on Cb (X) in the usual way via
d∞ (f, g) := ∥f − g∥∞

A useful exercise in checking your understanding of these definitions is Example 3.2.6

below. Here, let X = [0, 1]. Then, as we just remarked, C(X) = Cb (X). Instead of
writing C([0, 1]), it is conventional to write C[0, 1] for the vector space of continuous (and
automatically bounded) functions on [0, 1].

Example 3.2.6. Consider the space C[0, 1] together with the metric d∞ induced from the
norm ∥ · ∥∞ . Let (fn )∞
n=1 be a sequence of elements (functions) of this space, and let f be a
further element. Then fn → f in the metric d∞ if, and only if, fn converges to f uniformly.

Proof. This is essentially a tautology, but it takes a little thought to unravel all the
definitions.
 CHAPTER 4

Isometries, homeomorphisms and equivalence

One learns as mathematician that, when one studies a type of structure, one should also
study maps which preserve that structure. In this chapter we will look at various such
notions applicable to metric spaces.

4.1. Isometries

Maps which genuinely preserve the distance function are called isometries.
Definition 4.1.1. Let (X, dX ) and (Y, dY ) be metric spaces. A function f : X → Y
between metric spaces (X, dX ) and (Y, dY ) is said to be an isometry if

(4.1) dY (f (x), f (y)) = dX (x, y) for all x, y ∈ X.

Remarks. An isometry is automatically injective, but not automatically surjective. For
instance, the right-shift map on ℓ2 defined by f ((x1 , x2 , x3 , . . . )) = (0, x1 , x2 , . . . ) is an isom-
etry, but it is not surjective.
If an isometry is surjective as well, we call it a bijective isometry. Some authors use the
word “isometry” to mean “bijective isometry”, but we have refrained from doing this so
that we are consistent with Prelims course M4: Geometry. There, isometries in the case
X = Y = Rn were discussed in considerable detail. It was shown that they all have the
form f (x) = Ax + b for some orthogonal matrix A; in particular, they are automatically
surjective.
For any metric space X the set of all bijective isometries from X to itself is a group under
composition, denoted Isom(X).

4.2. Homeomorphisms

The notion of isometry is rather rigid. A weaker notion is that of a homeomorphism.

Definition 4.2.1. Let f : X → Y be a continuous function between metric spaces X
and Y . We say that f is a homeomorphism if it is continuous, a bijection, and if its inverse
f −1 : Y → X is also continuous.
Remark. Note that it is possible for a map f : X → Y to be both continuous and a
bijection, but for its inverse to fail to be continuous (so in this case f is not a homeomor-
phism). For instance, consider the spaces X = [0, 1)∪[2, 3] and Y = [0, 2]. Then the function
29
30 4. ISOMETRIES, HOMEOMORPHISMS AND EQUIVALENCE

f : X → Y given by (
x, if x ∈ [0, 1)
f (x) =
x − 1, if x ∈ [2, 3]
is a bijection and is clearly continuous. However, its inverse g : Y → X is not continuous at
1 – the one-sided limits of g as x tends to 1 from below and above are 1 and 2 respectively.
The following examples illustrate the extent to which homeomorphisms are less rigid than
isometries.

Example 4.2.2. The closed disk B̄(0, 1) of radius 1 in R2 is homeomorphic to the square
[−1, 1]×[−1, 1]. The easiest way to see this is to inscribe the disk in the square and stretch the
disk radially out to the square. One can write explicit formulas for this in the four quarters
of the disk given by the lines x ± y = 0 to check this does indeed give a homeomorphism.

Example 4.2.3. The open interval (−1, 1) is homeomorphic to R: an explicit homeomor-

phism is given by f (x) = x/(1 − |x|), which has inverse g(x) = x/(1 + |x|). It follows (using
translation and scaling maps) that any open interval is homeomorphic to R. Similarly, the
function h(x) = 1/x shows that (0, 1) and (1, ∞) are homeomorphic, and from this one can
see that the spaces R, (a, b), (−∞, a) and (a, ∞) are all homeomorphic for any a, b ∈ R with
a < b.

Example 4.2.4. A coffee cup (reusable, with a handle) is homeomorphic to a doughnut.

4.3. Equivalent metrics (non-examinable)

One space X can certainly support wildly different metrics. For instance, the 2-adic
metric on Q is very different to the standard Euclidean metric. However, there is a useful
notion of two metrics d1 , d2 on the same space being equivalent.

Definition 4.3.1 (Equivalent metrics). Let X be a set, and let d, d′ be two metrics on
X. Then we say that the metrics d, d′ are equivalent if the identity map ι : (X, d) → (X, d′ )
is a homeomorphism.

An easy exercise in the definitions show that this is equivalent to the following property:
every open ball B(x, ε) with respect to the d-metric contains an open ball B ′ (x, ε′ ) in the
d′ -metric, and vice versa.
If two metrics d, d′ are equivalent then, for example, the notions of limit coincide in the
two metric spaces (X, d) and (X, d′ ). We leave the detailed proof as an exercise.

Proposition 4.3.2. The metrics d1 , d2 , d∞ on Rn are equivalent.

 4.3. EQUIVALENT METRICS (NON-EXAMINABLE) 31
Proof. In fact, we will show that these metrics are strongly equivalent. Two metrics d, d′
on a space X are strongly equivalent if there is a constant C such that

d(x, y) ⩽ Cd′ (x, y) and d′ (x, y) ⩽ Cd(x, y)

for all x ̸= y. We leave it as an easy exercise to show that strongly equivalent metrics are
indeed equivalent (the converse is not true).
The three metrics under consideration all come from norms, and it is enough to find some
constant C such that

(4.2) ∥x∥ ⩽ C∥x∥′

for each pair ∥ · ∥, ∥ · ∥′ of these norms. Four such inequalities are obvious, namely

∥x∥∞ ⩽ ∥x∥1 ⩽ n∥x∥∞

and
∥x∥∞ ⩽ ∥x∥2 ⩽ n1/2 ∥x∥∞ .
The remaining two inequalities follow from these two, or one could use the Cauchy-Schwarz
inequality to get better constants.
 CHAPTER 5

Open and closed sets

5.1. Basic definitions

The definition of an open set, which we give now, is one of the most important in the
course.

Definition 5.1.1 (Open sets). If X is a metric space then we say a subset U ⊆ X is

open (or open in X) if for each y ∈ U there is some δ > 0 such that the open ball B(y, δ) is
contained in U .

To check you have understood the definition, convince yourself of the following facts:
• The interval (0, 1) is open in R, but [0, 1] is not;
• The rationals Q are not open in R;
• If X is a set with the discrete metric, every set is open.
Note carefully that the notion of open set is a relative notion, depending on U being
contained in X. Thus, while [0, 1] is not open in R, it is an open set considered as a
subspace of itself.
The first basic result about open sets is that open balls B(a, ε) are open. Note that this
is not a tautology – at this point “open ball” is just the name we gave to the sets B(a, ε),
and the fact that they are indeed open in the sense of Definition 5.1.1 needs to be proven.

Lemma 5.1.2. Every open ball in an metric space is an open set.

Proof. Let the ball be B(a, ε). Let x ∈ B(a, ε). Then d(x, a) < ε, so there is ε′ > 0
so that d(x, a) < ε − ε′ . We claim that the open ball B(x, ε′ ) is contained in B(a, ε). To
see this, suppose that z ∈ B(x, ε′ ). Then d(z, x) < ε′ and so by the triangle inequality
d(z, a) ⩽ d(z, x) + d(x, a) < ε′ + (ε − ε′ ) = ε.

The complement of an open set is a closed set.

Definition 5.1.3 (Closed sets). If X is a metric space, then a subset F ⊆ X is said to

be a closed subset of X if and only if its complement F c = X\F is an open subset.

It is important to note that the property of being closed is not the property of not being
open! In a metric space, it is possible for a subset to be open, closed, both or neither: In R
33
34 5. OPEN AND CLOSED SETS

the set R is open and closed, the set (0, 1) is open and not closed, the set [0, 1] is closed and
not open while the set (0, 1] is neither.
Just as open balls are open sets, so closed balls are closed sets, and this is also a fact
requiring proof.

Lemma 5.1.4. Every closed ball in a metric space is a closed set. In particular, singleton
sets are closed.

Proof. Let the ball be B̄(a, ε). We will show that the complement B̄(a, ε)c is open. Let
x ∈ B̄(a, ε)c . Then d(x, a) > ε, so there is ε′ > 0 so that d(x, a) > ε + ε′ . We claim that
the open ball B(x, ε′ ) is contained in B̄(a, ε)c . To see this, suppose that z ∈ B(x, ε′ ). Then
d(z, x) < ε′ and so by the triangle inequality d(z, a) ⩾ d(x, a) − d(z, x) > (ε + ε′ ) − ε′ = ε.
The second statement – that singleton sets are closed – follows from the observation that
{a} = B̄(a, 0).

5.2. Basic properties of open sets

Lemma 5.2.1. Let X be metric space. Then we have

(i) The subsets X and ∅ are open.
S
(ii) For any indexing set I and {Ui : i ∈ I} a collection of open sets, the set i∈I Ui is
an open set.
T
(iii) If I is finite and {Ui : i ∈ I} are open sets then i∈I Ui is open.
S
Proof. The first claim is trivial. For the second claim, if x ∈ i∈I Ui then there is some
i ∈ I with x ∈ Ui . Since Ui is open, some open ball B(x, ε) is contained in Ui and hence in
S
i∈I Ui .
T
Finally, for claim (iii), suppose that I is finite and that x ∈ i∈I Ui . For each i ∈ I, we
have x ∈ Ui , and so some ball B(x, εi ) is contained in Ui . Set ε := mini∈I εi ; then ε > 0
(here it is, of course, crucial that I be finite), and B(x, ε) ⊆ B(x, εi ) ⊆ Ui for all i. Therefore
T
B(x, ε) ⊆ i∈I Ui .

Remarks. (i) is in fact a special case of (ii) and (iii), taking I to be the empty set.
It is extremely important to note that, whilst the indexing set I in (ii) can be arbitrary,
the indexing set in (iii) must be finite. In general, an arbitrary intersection of open sets is
not open; for instance, the intervals Ui = (−1/i, 1/i) are all open in R, but their intersection
T∞
i=1 Ui is just the singleton {0}, which is not an open set.
A result equivalent to Lemma 5.2.1 may be formulated in terms of closed sets, simply by
taking complements and applying de Morgan’s laws. We simply state the outcome.
 5.4. TOPOLOGICAL SPACES (NON-EXAMINABLE) 35
Lemma 5.2.2. Let X be a metric space and let {Fi : i ∈ I} be a collection of closed
subsets.
(i) The subsets X and ∅ are closed.
T
(ii) The intersection i∈I Fi is a closed subset.
S
(iii) If I is finite then i∈I Fi is closed.

If X is a metric space, the collection of all open sets in X is called the topology of X.

5.3. Continuity in terms of open sets

An interesting and important fact is that continuity of a function may be expressed in an

ε-δ-free manner using open sets.

Proposition 5.3.1. Let X, Y be metric spaces and let f : X → Y be a map. Then f is

continuous on all of X if and only if for each open subset U of Y , its preimage f −1 (U ) is
open in X.

Proof. Suppose first that f is continuous at every point, and let U ⊆ Y be open; we want to
show that f −1 (U ) is open. Let a ∈ f −1 (U ) be arbitrary. Then f (a) ∈ U , and so, since U is
open, some ball B(f (a), ε) also lies in U . By the definition of continuity, there is some δ > 0
such that if x ∈ B(a, δ) then f (x) ∈ B(f (a), ε), and therefore f −1 (B(f (a), ε)) ⊇ B(a, δ).
Therefore f −1 (U ) contains B(a, δ), which means that f −1 (U ) is open.
Now suppose that f satisfies the open sets preimages property, and let a ∈ X. The ball
B(f (a), ε) is open, and so by assumption the preimage f −1 (B(f (a), ε)) is open. Since a
lies in this set, it follows from the definition of open that there is some δ > 0 such that
B(a, δ) ⊆ f −1 (B(f (a), ε)), whence f (B(a, δ)) ⊆ B(f (a), ε). This is what it means for f to
be continuous at a.

By taking complements, one can show the following version of Proposition 5.3.1 for closed
sets: f : X → Y is continuous if and only if for each closed subset V of Y , its preimage
f −1 (V ) is a closed subset of X.
Finally, it is important to take note of what Proposition 5.3.1 does not say, namely that a
continuous function maps open sets to open sets. This is obvious since, for example, constant
functions are continuous. Less obvious is the fact that it still fails even under the assumption
that f is injective. For instance, the injective map f : [0, 1) → S 1 ⊂ C given by t → e2πit is
continuous. The set [0, 1/2) is open in [0, 1), but its image is not open in S 1 .

5.4. Topological spaces (non-examinable)

In this section we offer a very brief taster of the course A5: Topology by discussing the
notion of a topological space. One may of course observe that Proposition 5.3.1 allows one
36 5. OPEN AND CLOSED SETS

to define the notion of a continuous function without explicitly mentioning the metric X or
concepts equivalent to it such as the notion of an open ball of radius δ. Of course, those
notions are embedded within the definition of the notion of an open set, so this comment is
a little misleading.
In the concept of a topological space, the open sets are to the fore. Thus a topological
space is a set X together with a collection of sets U (which we call the open sets) satisfying
certain properties. The properties we require are precisely those which we proved in Lemma
5.2.1, namely
(i) The subsets X and ∅ are open.
S
(ii) For any indexing set I and {Ui ; i ∈ I} a collection of open sets, the set i∈I Ui is
an open set.
T
(iii) If I is finite and {Ui : i ∈ I} are open sets then i∈I Ui is open in X.
Note that we have not said anything about the “geometry” of the open sets, or anything
about them containing balls - indeed there are no such notions, because X is equipped with
no structure.
Lemma 5.2.1 may then be phrased as follows.

Lemma 5.4.1 (Lemma 5.2.1). Let X be a metric space together with the open sets as
defined in Definition 5.1.1. Then X is a topological space, with the same collection of open
sets.

The concept of a topological space is considerably more general than that of a metric
space, and there are certainly topological spaces which do not have the structure of a metric
space (are not metrizable). However, as a consequence of Proposition 5.3.1 we may still
formulate the notion of a continuous function between two topological spaces, in such a way
that when restricted to metric spaces it coincides with the usual definition.

Definition 5.4.2. Suppose that X and Y are two topological spaces. Then we say that
f : X → Y is continuous if and only if, for every open set U ⊆ Y , the inverse f −1 (U ) is open
in X.

Let us emphasise that in the generality of topological spaces, there is no equivalent form
of this definition in terms of εs and δs.

5.5. Subspaces

If (X, d) is a metric space, then as we noted in Section 2.5, any subset Y ⊆ X is automat-
ically also a metric space since the distance function d : X × X → R⩾0 restricts to a distance
function on Y . We will use the letter d for both metrics, but it is important to distinguish
the balls in Y from the balls in X, because these are quite different objects.
 5.5. SUBSPACES 37

We will write
BY (y, r) = {z ∈ Y : d(z, y) < r}
for the open ball about y of radius r in Y and

BX (y, r) = {x ∈ X : d(x, y) < r}

for the open ball of radius r about y in X.

Note that BY (y, r) = Y ∩ BX (y, r).
Similarly, the notions of a set being open in X and of being open in Y are quite different.
By way of an example, consider X = R2 and Y = R × {0}, that is to say Y is the x-axis.
The the ball BX (0, 1) is simply the open unit disc of radius 1, whilst the ball BY (0, 1) is the
open unit line segment (−1, 1). Note carefully that BY (0, 1), whilst it is open as a subset of
Y , does not look remotely like an open subset of X.
The following lemma clarifies the relationship between open sets in X and open sets in
the subspace Y .

Lemma 5.5.1. Let X be a metric space and suppose that Y ⊆ X. Then a subset U ⊆ Y
is an open subset of Y if and only if there is an open subset V of X such that U = Y ∩ V .
Similarly a subset Z ⊆ Y is a closed subset of Y if and only if there is a closed subset F of
X such that Z = F ∩ Y .

Proof. Suppose first that U = Y ∩ V , where V is open in X. We will show that U is open in
Y . Let y ∈ U . Then, since V is open, there is some ε > 0 such that BX (y, ε) ⊆ V . Therefore

BY (y, ε) = Y ∩ BX (y, ε) ⊆ V ∩ Y = U.

We have shown that some open ball (in Y ) about y is contained in U , and therefore U is
open.
In the other direction, suppose that U is an open subset of Y . Then for each y ∈ U we
S
may pick an open ball BY (y, εy ) contained in U . We have y∈U BY (y, εy ) = U . Now define
S
V = y∈U BX (y, εy ). Then V , being a union of open balls in X, is open. Moreover
[ [ [
Y ∩V =Y ∩ BX (y, εy ) = (Y ∩ BX (y, εy )) = BY (y, εy ) = U.
y∈Y y∈Y y∈Y

The corresponding result for closed sets follows by taking complements – we leave the
detailed verification as an exercise.

The concept of being open in a subspace is a bit confusing when you first meet it, so let
us give an example. Let X = R, and Y = (0, 1] ∪ [2, 3]. Set U = (0, 1]. Then U is not
open as a subset of X – for instance, no open ball BX (1, ε) is contained in U . However, U
38 5. OPEN AND CLOSED SETS

is open as a subset of Y . For example, the ball BY (1, 12 ), which consists of all points of Y
at distance less than 12 from 1, is the set ( 21 , 1], and this is contained in U .
 CHAPTER 6

Interiors, closures, limit points

In this chapter we explore some further concepts in the basic theory of metric spaces.

6.1. Interiors and closures

Definition 6.1.1. Let X be a metric space, and let S ⊆ X. The interior int(S) of S is
defined to be the union of all open subsets of X contained in S. The closure S̄ is defined
to be the intersection of all closed subsets of X containing S. The set S̄\int(S) is known as
the boundary of S and denoted ∂S. A set S ⊆ X is said to be dense if S = X.

It is very important to note that, in Definition 6.1.1, the notion of open and closed is here
being taken in the metric space X, not in the subspace metric on S, which would result in
trivial definitions.
Since an arbitrary union of open sets is open (Lemma 5.2.1), int(S) is itself an open set,
and it is clearly the unique largest open subset of X contained in S. If S is itself open then
evidently S = int(S).
Since an arbitrary intersection of closed sets is closed, S̄ is the unique smallest closed
subset of X containing S. If S is itself closed then evidently S = S.
If x ∈ int(S) we say that x is an interior point of S.

Example 6.1.2. If S = [a, b] is a closed interval in R then its interior is just the open
interval (a, b). If we take S = Q ⊂ R then int(Q) = ∅.

Example 6.1.3. The rationals Q are a dense subset of R, as is the set { 2an : a ∈ Z, n ∈ N}.

Let us give a couple of simple characterisations of the closure of a set.

Lemma 6.1.4. Let X be a metric space, and let S ⊆ X be a subset. Then a ∈ S̄ if and
only if the following is true: every open ball B(a, ε) contains a point of S.

Proof. Suppose a ∈ S̄ and let ε > 0. If B(a, ε) does not meet S, then B(a, ε)c is a closed
set containing S. Therefore B(a, ε)c contains S̄, and hence it contains a, which is obviously
nonsense.
Conversely, let a ∈ X and suppose that every ball B(a, ε) meets S. If a ∈ / S̄ then, since
c c c
S̄ is open, there is a ball B(a, ε) contained in S̄ , and hence in S , contrary to assumption.

39
40 6. INTERIORS, CLOSURES, LIMIT POINTS

Remark. A particular consequence of this is that S ⊆ X is dense if and only if it meets

every open set in X.

Corollary 6.1.5. Let X be a metric space, and let S ⊆ X be a subset. Let a ∈ X.

Then a lies in the closure S̄ if and only if there is a sequence (xn )∞
n=1 of elements of S with
limn→∞ xn = a. In particular, S is closed if and only if the limit of every convergent sequence
(xn )∞
n=1 of elements of S lies in S.

Proof. We use Lemma 6.1.4. Suppose that a ∈ S̄. Then by Lemma 6.1.4 every ball
B(a, 1/n) contains a point of S, so we may pick a sequence (xn )∞
n=1 with xn ∈ B(a, 1/n) ∩ S.
Clearly limn→∞ xn = a.
Conversely, suppose limn→∞ xn = a, where xn ∈ S. If a ∈ / S̄ then by Lemma 6.1.4 there
must be some ball B(a, ε) not meeting S. But if n is large enough then d(xn , a) < ε, and so
xn ∈ S ∩ B(a, ε), contradiction.

We conclude with a cautionary example which has often confused people when they first
meet it.

Example 6.1.6. In general, it need not be the case that B̄(a, ε) is the closure of B(a, ε).
Since we have seen that B̄(a, ε) is closed, it is always true that B(a, ε) ⊆ B̄(a, ε), but the
containment can be proper. Indeed, take any set X with at least two elements equipped
with the discrete metric. Then if x ∈ X we have B(x, 1) = B(x, 1) = {x}, but B̄(x, 1) is the
whole space X.

6.2. Limit points

This section introduces the notion of limit points (also known in some places as cluster
points or accumulation points). The notion is a well-studied one, introduced here for cultural
reference, but we will not come across it in subsequent chapters of the course.

Definition 6.2.1. If X is a metric space and S ⊆ X is any subset, then we say a point
a ∈ X is a limit point of S if any open ball about a contains a point of S other than a itself.

We will write L(S) for the set of limit points of S; I am not sure that there is any
completely standard notation for this. Note that we do not necessarily have S ⊆ L(S), that
is to say it is quite possible for a point a ∈ S not to be a limit point of S. This occurs if
there is some ball B(a, ε) such that B(a, ε) ∩ S = {a}, and in this case we say that a is an
isolated point of S.

Example 6.2.2. Take X = R and S = (0, 1]∪{2}. Then L(S) = [0, 1]. Note in particular
that 0 does not lie in S, but is a limit point; by contrast, 2 does lie in S, but it is not a limit
point, so it is an isolated point.
 6.2. LIMIT POINTS 41

Lemma 6.2.3. Let S be a subset of a metric space X. Then L(S) is a closed subset of X.

Proof. We need to show that the complement L(S)c is open. Suppose a ∈ L(S)c . Then
there is a ball B(a, ε) whose intersection with S is either empty or {a}.
We claim that B(a, ε/2) ⊆ L(S)c . Let b ∈ B(a, ε/2). If b = a, then clearly b ∈ L(S)c . If
b ̸= a, there is some ball about b which is contained in B(a, ε), but does not contain a: the
ball B(b, δ) where δ = min(ε/2, d(a, b)) has this property. This ball meets S in the empty
set, and so b ∈ L(S)c in this case too.

Proposition 6.2.4. Let S be a subset of a metric space X. Let L(S) be its set of limit
points, and S̄ its closure. Then S̄ = S ∪ L(S).

Proof. We first show the containment S ∪ L(S) ⊆ S̄. Obviously S ⊆ S̄, so we need only
show that L(S) ⊆ S̄. Suppose a ∈ S̄ c . Since S̄ c is open, there is some ball B(a, ε) which lies
in S̄ c , and hence also in S c , and therefore a cannot be a limit point of S. This concludes the
proof of this direction.
Now we look at the opposite containment S̄ ⊆ S ∪ L(S). Let a ∈ S̄. We saw in Lemma
6.1.5 that there is a sequence (xn )∞ n=1 of elements of S with xn → a. If xn = a for some n
then we are done, since this implies that a ∈ S. Suppose, then, that xn ̸= a for all n. Let
ε > 0. Then there exists N such that for all n ⩾ N we have xn ∈ B(a, ε) \ {a}, and they all
also lie in S. It follows that a is a limit point of S, as B(a, ε) contains for example xN ̸= a.
So we are done in this case also.

Corollary 6.2.5. Let S be a subset of a metric space X. Then S is closed if and only
if it contains all its limit points.

Proof. We already remarked, in Section 6.1, that S is closed if and only if S = S̄. The
corollary is immediate from this and Proposition 6.2.4.
 CHAPTER 7

Completeness

Students may wish to remind themselves of the Prelims course M2: Analysis I, which
covered some of the topics of this section in the specific case of the real numbers. Much of
the theory in a general metric space is a natural generalisation of what was done there.

7.1. Basic definitions and examples

Definition 7.1.1. Let (xn )∞

n=1 be a sequence in some metric space X. Then we say that
this sequence is
• Bounded if the set {xn : n ⩾ 1} is bounded in the sense of Definition 2.6.2, that is
to say if all the xn lie in some ball B(a, R);
• Cauchy if the xn become arbitrarily close together as n → ∞, in the following sense:
for every ε > 0, there is some N such that d(xn , xm ) < ε whenever n, m ⩾ N .
• Convergent if there is some a ∈ X such that limn→∞ xn = a.

If a sequence σ has any one of these properties, then any subsequence of σ also has the
property. We leave the proof of this as an exercise.
The relation between the above concepts is as follows.

Proposition 7.1.2. A convergent sequence is Cauchy. A Cauchy sequence is bounded.

Neither of the reverse implications holds in general.

Proof. We begin by showing that the reverse implications do not hold, since the examples
we will give serve to illustrate the concepts. Take X = (0, 1]. Then the sequence xn = 1/n
is Cauchy, but not convergent. The sequence in which xn = 1 for n odd and xn = 1/2 for n
even is bounded, but it is not Cauchy since there is no N such that d(xn , xn+1 ) < 1/2 for all
n ⩾ N.
Now we show the two main implications. Suppose that (xn )∞ n=1 is convergent, and that
limn→∞ xn = a. Let ε > 0. By the definition of limit, there is some N such that, if n ⩾ N ,
d(xn , a) < ε/2. Now suppose that m, n ⩾ N . Then

d(xm , xn ) ⩽ d(xn , a) + d(xm , a) < ε/2 + ε/2 = ε,

and so (xn )∞
n=1 is Cauchy.
43
44 7. COMPLETENESS

Now suppose that (xn )∞ n=1 is Cauchy. Taking ε = 1 in the definition, we see that there is
some N such that d(xm , xn ) < 1 whenever m, n ⩾ N . In particular, all points of the sequence
except (possibly) x1 , . . . , xN −1 lie in B(xN , 1). It follows that all points of the sequence lie
in B(xN , R), where R is the largest value of the set {d(xN , x1 ) + 1, . . . , d(xN , xN −1 ) + 1}, and
so (xn )∞
n=1 is bounded.

Now we turn to the key definition of the chapter.

Definition 7.1.3 (Completeness). A metric space is said to be complete if every Cauchy

sequence converges.

One of the main results of the Prelims course was that R is complete, and it is easy to
deduce from this that Rn is complete also (since a sequence in Rn converges if and only if
each of its coordinates converge).
On the other hand, we observed above that (0, 1] is not complete. For much the same
reason, (0, 1) is not complete. Note, however, that (0, 1) is homeomorphic to R, as we
showed earlier. Therefore the notion of completeness is not (necessarily) preserved under
homeomorphisms.
Let V be a normed vector space with norm ∥ · ∥. As previously discussed, we can define
a metric on V by d(v, w) = ∥v − w∥. We say that V is complete if, when endowed with
the structure of a metric space in this way, it is complete. That is, when we talk about
completeness of normed spaces we implicitly assume that the obvious metric has been put
on V , without necessarily mentioning it explicitly.

7.2. First properties of complete metric spaces

In this section we collect a couple of basic properties of complete metric spaces.

Lemma 7.2.1. A subspace of a complete metric space is complete if and only if it is closed.

Proof. Let X be a complete metric space and let Y ⊆ X. Suppose first that Y is closed;
we will show that it is complete. Let (yn )∞
n=1 be a Cauchy sequence in Y . Then it is also a
Cauchy sequence in X. Since X is complete, it converges, say limn→∞ yn = a. By Corollary
6.1.5, a ∈ Y .
In the other direction, suppose that Y is complete. Let (yn )∞
n=1 be a sequence of elements
∞
of Y with limn→∞ yn = a for some a ∈ X. Then (yn )n=1 is certainly a Cauchy sequence in
Y , and so by completeness of Y it converges to an element of Y . So by uniqueness of limits
we have a ∈ Y . Thus Y contains the limits of such sequences, and by Corollary 6.1.5, Y is
closed.
 7.3. COMPLETENESS OF FUNCTION SPACES 45

Definition 7.2.2. Let X be a metric space and Y ⊆ X a non-empty subset. The diameter
of Y , written diam(Y ), is defined to be the supremum of the set {d(x, y) : x, y ∈ Y } when
this set is bounded, and infinity otherwise.

The next lemma is sometimes known as Cantor’s intersection theorem.

Lemma 7.2.3. Let X be a complete metric space and suppose that S1 ⊇ S2 ⊇ . . . form
a nested sequence of non-empty closed sets in X with the property that diam(Sn ) → 0 as
n → ∞. Then ∞
T
n=1 Sn contains a unique point a.

Proof. For each n, pick xn ∈ Sn . We claim that (xn )∞ n=1 is Cauchy. To see this, let ε > 0,
and suppose that N is large enough that diam(SN ) < ε. If n, m ⩾ N then, since the Si are
nested, xn , xm ∈ SN . By the definition of diameter, d(xn , xm ) ⩽ diam(SN ) < ε.
Since X is complete, we have limn→∞ xn = a for some a. For each i, the nesting property
of the sets Si implies that we have xn ∈ Si for all n ⩾ i. Therefore, since Si is closed,
Corollary 6.1.5 tells us that a ∈ Si . Since this is true for all i, we have a ∈ ∞
T
i=1 Si .
T∞
To show that a is unique, suppose that b ∈ i=1 Si . Then d(a, b) ⩽ diam(Si ) for all i.
Since diam(Si ) → 0, we have d(a, b) = 0 and so a = b.

What if we drop the condition diam(Si ) → 0? We certainly could not expect a to be

unique since, for instance, we could take all the Si to be the whole space X. Somewhat
surprisingly at first sight, the intersection ∞
T
i=1 Si may even be empty. For instance, take
Si = [i, ∞) ⊂ R.

7.3. Completeness of function spaces

In this section we show that two natural spaces of functions give rise to complete metric
spaces.
For the first result, recall that if X is a set then B(X) denotes the normed vector space
of bounded functions f : X → R, with norm ∥f ∥∞ = supx∈X |f (x)|.

Theorem 7.3.1. Let X be any set. Then B(X) is complete.

Proof. Let (fn )∞ ∞

n=1 be a Cauchy sequence in B(X). Then for each x the sequence (fn (x))n=1
is a Cauchy sequence of real numbers (convincing yourself of this is a good exercise to check
you have understood the definitions). Since R is complete, each such sequence has a limit,
and we write f (x) for this limit. That is, limn→∞ fn (x) = f (x).
We claim that f is a bounded function. To see this, take ε = 1 in the definition of Cauchy
sequence. This gives an N such that, if n, m ⩾ N , supx |fn (x) − fm (x)| ⩽ 1. In particular,
|fN (x) − fn (x)| ⩽ 1 for all n ⩾ N and for all x ∈ X. Taking the limit as n → ∞, it follows
that |fN (x) − f (x)| ⩽ 1 for all x. Since fN is a bounded function, so is f .
46 7. COMPLETENESS

Finally, we need to show that fn → f in the norm ∥ · ∥∞ (at the moment we have only
shown pointwise convergence). The argument is a simple modification of the preceding one.
Let ε > 0, and let N be such that, if n, m ⩾ N , |fn (x) − fm (x)| ⩽ ε for all x ∈ X. For each
fixed n ⩾ N and x ∈ X, we may let m → ∞, obtaining that |fn (x) − f (x)| ⩽ ε. That is, for
all n ⩾ N we have ∥fn − f ∥∞ ⩽ ε. It follows that fn → f in the ∥ · ∥∞ -norm.

For the second result, recall that if X is a metric space then Cb (X) denotes the normed
vector space of bounded continuous functions f : X → R, again with norm ∥f ∥∞ =
supx∈X |f (x)|.

Theorem 7.3.2. Let X be a metric space. Then Cb (X) is complete.

Proof. We have shown in Theorem 7.3.1 that B(X) is complete, so by Lemma 7.2.1 it is
enough to show that Cb (X) is a closed subset of B(X).
By Corollary 6.1.5, it suffices to show that if (fn )∞
n=1 is a sequence of elements of Cb (X)
converging in the ∥ · ∥∞ -norm to some f ∈ B(X), then f ∈ Cb (X), or in other words f is
continuous.
Let a ∈ X, and let ε > 0. Since fn → f in the ∥ · ∥∞ -norm, there is some n such that
∥fn − f ∥∞ ⩽ ε/3. Since fn is continuous, there is a δ > 0 such that |fn (x) − fn (a)| < ε/3
for all x ∈ B(a, δ). But then for x ∈ B(a, δ) we have
|f (x) − f (a)| ⩽ |f (x) − fn (x)| + |fn (x) − fn (a)| + |fn (a) − f (a)|
< ε/3 + ε/3 + ε/3 = ε.
It follows that f is continuous at a, and since a was arbitrary, f is a continuous function on
X.

Remark. You may have the impression that you have seen something like this argument
before, and indeed that is the case. In Prelims: Analysis II you saw that a uniform limit of
continuous functions on R is continuous, and our task here was essentially the same, but in
the setting of a general metric space.

7.4. The contraction mapping theorem

The final topic of this section is a classic theorem about fixed points of certain maps from
a metric space to itself. We will discuss the result for its own intrinsic interest, but it has
important applications to the solutions of differential equations, as you will see in the course
A1 : Differential Equations.
Let us begin with a couple of definitions.

Definition 7.4.1. Let (X, dX ) and (Y, dY ) be metric spaces and suppose that f : X → Y .
We say that f is a Lipschitz map (or is Lipschitz continuous) if there is a constant K ⩾ 0
 7.4. THE CONTRACTION MAPPING THEOREM 47

such that
dY (f (x), f (y)) ⩽ KdX (x, y).
If Y = X and K ∈ [0, 1) then we say that f is a contraction mapping (or simply a contrac-
tion).

An easy exercise is to check that every Lipschitz map is continuous, and to give an example
of a continuous map between metric spaces which is not Lipschitz.
It is very important to note, in the definition of a contraction, that it says something
stronger than that d(f (x), f (y)) < d(x, y), namely that there is a constant K < 1 such that
d(f (x), f (y)) ⩽ Kd(x, y) for all x, y.

Theorem 7.4.2 (Contraction mapping theorem). Let X be a nonempty complete metric

space and suppose that f : X → X is a contraction. Then f has a unique fixed point, that
is, there is a unique x ∈ X such that f (x) = x.

Proof. We begin by showing that there cannot be two fixed points. Suppose that f (x1 ) = x1
and that f (x2 ) = x2 . Then we have

d(x1 , x2 ) = d(f (x1 ), f (x2 )) ⩽ Kd(x1 , x2 ).

Since d(x1 , x2 ) ⩾ 0 and K < 1, we are forced to conclude that d(x1 , x2 ) = 0 and hence that
x1 = x2 .
Now we show that there is a fixed point. The proof is constructive (and may be used
in practical situations to find fixed points numerically). The idea is as follows. Pick an
arbitrary x0 ∈ X, and form the sequence of iterates x1 := f (x0 ), x2 := f (x1 ), and so on.
We claim that (no matter which x0 we started with) the sequence (xn )∞ n=1 converges to some
limit x, and that f (x) = x.
To show that (xn )∞ n=1 converges, it suffices to show that it is Cauchy, since X is complete.
To do this, first observe that by repeated use of the contraction property and the definition
of the sequence (xn )∞ n=1 we have

d(xn , xn−1 ) ⩽ Kd(xn−1 , xn−2 ) ⩽ K 2 d(xn−2 , xn−3 ) ⩽ . . . ⩽ K n−1 d(x0 , x1 )

(you could prove this formally by induction if you wanted). Therefore if n > m we have

d(xn , xm ) ⩽ d(xn , xn−1 ) + · · · + d(xm+1 , xm )

⩽ (K n−1 + K n−2 + · · · + K m )d(x0 , x1 )
⩽ K m (1 + K + K 2 + . . . )d(x0 , x1 ) = CK m ,

where C = d(x0 , x1 )/(1 − K) (by summing the geometric series).

It follows that if n, m ⩾ N then d(xm , xn ) ⩽ CK N .
48 7. COMPLETENESS

Since K < 1, for any ε > 0 there is some N such that CK N < ε, and therefore (xn )∞
n=1 is
indeed a Cauchy sequence.
Since X is complete, xn → x for some x ∈ X. To complete the proof we must show that
f (x) = x. This is quite straightforward. Indeed, since f is continuous we have

f (x) = lim f (xn ) = lim xn+1 = x.

n→∞ n→∞

This finishes the proof.

Remarks. Over the years, many people have lost a mark in exam questions for forgetting
that X must be non-empty.
Let us conclude by giving a couple of examples to show that the hypotheses of the theorem
are necessary. First, we observe that the weaker condition that d(f (x), f (y)) < d(x, y) for all
x ̸= y is not sufficient. For instance, it may be checked that the function f : [1, ∞) → [1, ∞)
defined by f (x) = x + 1/x has this property, but it obviously has no fixed points.
More obviously, the requirement that X is complete is important. For instance, if we
define f : (0, 1) → (0, 1) by f (x) = x/2 then clearly f is a contraction, but f has no fixed
points in (0, 1).
 CHAPTER 8

Connectedness and path-connectedness

In this section we try to understand what makes a space “connected”. We will consider
two natural approaches to this question, and show that for reasonably nice spaces the two
notions in fact coincide. In particular, the two notions of connectedness coincide for open
subsets of the complex plane.

8.1. Connectedness

The concept of connectedness formulates the intuitive idea of a space which cannot be
split into two “separated” pieces.

Definition 8.1.1. We say that a metric space is disconnected if we can write it as the
disjoint union of two nonempty open sets. We say that a space is connected if it is not
disconnected.

If X is written as a disjoint union of two nonempty open sets U and V then we say that
these sets disconnect X.
If X = [0, 1] ∪ [2, 3] ⊂ R then we have seen that both [0, 1] and [2, 3] are open in X. Since
X is their disjoint union, X is disconnected.
It is a little harder to give a nontrivial example of a connected space. Later on, we will
show that all intervals in R are connected.
The following lemma gives some equivalent ways to formulate the concept of a connected
space.

Lemma 8.1.2. Let X be a metric space. Then the following are equivalent.
(i) X is connected.
(ii) If f : X → {0, 1} is a continuous function then f is constant.
(iii) The only subsets of X which are both open and closed are X and ∅.
(Here the set {0, 1} is viewed as a metric space via its embedding in R, or equivalently with
the discrete metric.)

Proof. (i) ⇒ (ii): Let X be connected, and let f : X → {0, 1} be a continuous function.
The singleton sets {0} and {1} are both open in {0, 1} and so both f −1 (0) and f −1 (1) are
open subsets of X. They are clearly disjoint, and their union is X. Therefore one of them
must be empty, which means that f is constant.
49
50 8. CONNECTEDNESS AND PATH-CONNECTEDNESS

(ii) ⇒ (iii): Suppose that A ⊆ X is both open and closed. Then Ac is open (and closed),
and so the function f : X → {0, 1} defined by f (x) = 1 for x ∈ A and f (x) = 0 for x ∈ Ac
(that is, the characteristic function of A) is continuous. Assuming (ii), it must be constant.
If it takes the constant value 1, then A = X. If it takes the constant value 0, then A = ∅.
(iii) ⇒ (i): Suppose that X = U ∪ V with U, V open and disjoint. Then U c = V is open,
so U is also closed. Thus U is both open and closed, and hence (assuming (iii)) is either X
or ∅. Similarly for V . Hence there is no way to disconnect f (X).

Frequently one has a metric space X and a subset Y of it whose connectedness or otherwise
one wishes to ascertain. To this end, it is useful to record the following lemma.

Lemma 8.1.3. Let X be a metric space, and let Y ⊆ X be a subset, considered as a metric
space with the metric induced from X. Then Y is connected if and only if the following is
true. If U, V are open subsets of X, and U ∩ V ∩ Y = ∅, then whenever Y ⊆ U ∪ V , either
Y ⊆ U or Y ⊆ V .

Proof. The key point here is to recall that the open sets in Y are precisely the sets of the
form U ∩ Y , where U is open in X. This was proven in Lemma 5.5.1. Take a pair U ∩ Y ,
V ∩ Y of such open sets. They disconnect Y if and only if
(i) They are disjoint, thus U ∩ V ∩ Y = ∅;
(ii) They cover Y , which is equivalent to Y ⊆ U ∪ V ;
(iii) Neither is empty.
Thus Y is connected if and only if (i) and (ii) imply that one of U ∩ Y , V ∩ Y is empty or
equivalently that Y ⊆ V or Y ⊆ U .

We now turn to some basic properties of the notion of connectedness. These broadly
conform with one’s intuition about how connected sets should behave, but of course proof is
required in each case.

Lemma 8.1.4 (Sunflower lemma). Let X be a metric space. Let {Ai : i ∈ I} be a collection
T S
of connected subsets of X such that i∈I Ai ̸= ∅. Then i∈I Ai is connected.

Proof. We use the alternative characterisation of connectedness given in Lemma 8.1.2 (ii).
S
Suppose that f : i∈I Ai → {0, 1} is continuous. We must show that f is constant. Pick
T S
x0 ∈ i∈I Ai . Then if x ∈ i∈I Ai there is some i for which x ∈ Ai . But then the restriction
of f to Ai is constant since Ai is connected, so that f (x) = f (x0 ) as x, x0 ∈ Ai . But since x
was arbitrary, it follows that f is constant as required.

Lemma 8.1.5 (Connectedness and closures). Let X be a metric space. If A ⊆ X is

connected then if B is such that A ⊆ B ⊆ Ā, the set B is also connected.
 8.1. CONNECTEDNESS 51

Proof. We use the criterion for a subspace to be connected from Lemma 8.1.3. Suppose
that B ⊆ U ∪ V where U and V are open in X and U ∩ V ∩ B = ∅. Then certainly A ⊆ U ∪ V
and A ∩ U ∩ V = ∅. Hence, since A is connected, either A ⊆ U or A ⊆ V . Without loss
of generality, A ⊆ U , and since A ∩ U ∩ V = ∅ this means that A ⊆ V c . However, V c is
closed and so taking closures we obtain Ā ⊆ V¯c = V c . In particular B ⊆ V c and so, since
B ⊆ U ∪ V , we must have B ⊆ U . We have verified the criterion (Lemma 8.1.3) for a
subspace to be connected.

Lemma 8.1.6 (Connected image of a connected set). Let X be a connected metric space,
and let f : X → Y be continuous. Then f (X) is connected.

Proof. We may as well suppose that f is surjective (otherwise replace Y by f (X)). Suppose
that U and V are disjoint open subsets of Y with U ∪ V = Y . Then f −1 (U ) and f −1 (V ) are
disjoint open subsets of X with f −1 (U ) ∪ f −1 (V ) = X. Since X is connected one of them,
say f −1 (U ), is empty. Therefore U is empty.
It follows that there is no way to disconnect Y .

A simple corollary is that (unlike completeness) the property of connectedness is preserved

under homeomorphisms.
Connected components. A consequence of the Sunflower Lemma is that, for each x ∈ X,
there is a unique maximal connected subset of X containing x, which contains all other such
sets (take the union of all connected subsets of X containing x). This is called the connected
component of X containing x.

Proposition 8.1.7 (Connected components). The connected components of a metric

space partition the space. A space is connected if and only if it has a unique connected
component.

Proof. Let X be the space, and for x ∈ X write Γ(x) for the connected component
containing x. Suppose that Γ(x) and Γ(y) are not disjoint, say a ∈ Γ(x) ∩ Γ(y). We wish
to show that they coincide, which is what it means for them to partition the space. By the
Sunflower Lemma, Γ(x) ∪ Γ(y) is connected. By the definition of connected component, Γ(x)
must contain this set, which of course means that Γ(y) ⊆ Γ(x). Similarly Γ(x) ⊆ Γ(y), and
so Γ(x) = Γ(y).
The second statement is obvious.

As promised we now show that intervals in R are connected.

Theorem 8.1.8. Any interval [x, y] in R is connected.

52 8. CONNECTEDNESS AND PATH-CONNECTEDNESS

Proof. We show that [x, y] is connected using Lemma 8.1.3. Suppose [x, y] ⊆ U ∪ V where
U and V are open subsets of R with [x, y] ∩ U ∩ V = ∅, but that we do not have [x, y] ⊆ U or
[x, y] ⊆ V . Thus [x, y] ∩ U and [x, y] ∩ V are both non-empty. Without loss of generality we
may assume x ∈ [x, y] ∩ U and y ∈ [x, y] ∩ V . (The latter is not so obvious: as [x, y] ∩ V ̸= ∅,
there exists y ′ ∈ [x, y] with y ′ ∈ V . Note y ′ ̸= x ∈ U . We replace [x, y] by the possibly
smaller interval [x, y ′ ]: Then still [x, y ′ ] ⊆ U ∪ V , with [x, y ′ ] ∩ U ∩ V = ∅, and moreover
[x, y ′ ] ̸⊆ U and [x, y ′ ] ̸⊆ V , as x ∈ U and y ′ ∈ V . Note we shall reach a contradiction, which
shows our assumption on the original interval [x, y] was false.) Note that as [x, y]∩U ∩V = ∅
this tells us y ̸∈ U and x ̸∈ V .
Now define S = {z ∈ [x, y] : z ∈ U }. Then S is non-empty and bounded and so c = sup(S)
exists. Note c ∈ [x, y]. Since [x, y] ⊆ U ∪ V , we have either c ∈ U or c ∈ V .
Assume c ∈ U . Since y ̸∈ U we have c ̸= y. So as U is open, there is some interval
[c, c + ε) contained in U and also in [x, y]. This means that [c, c + ε) ⊆ S, which contradicts
the fact that c = sup(S) (for instance, c + ε/2 lies in S and is bigger than c).
If c ∈ V then c ̸= x and so, since V is open, there is some interval (c − ε, c] contained in
V and also in [x, y]. In particular, [c − ε/2, c] is disjoint from S, which contradicts the fact
that c = sup(S) (for instance, c − ε/2 as an upper bound for S, and is smaller than c).
These two contradictions show that we were wrong to assume that neither [x, y] ⊆ U or
[x, y] ⊆ V . Therefore [x, y] is connected.

To finish this section, let us remark that the intermediate value theorem is an almost
immediate consequence of Theorem 8.1.8 and Lemma 8.1.6. Indeed, suppose f : [a, b] → R
is continuous. Then, since [a, b] is connected, f ([a, b]) is connected. Therefore this latter set
is an interval and in particular it contains every c with between f (a) and f (b).

8.2. Path-connectedness

We now turn to a different, but equally intuitive, notion of what it means for a set to be
connected: that one should be able to “continuously move” from any point to another. Here
is the precise definition.

Definition 8.2.1 (Path connectedness). Let X be a metric space. Then we say that
X is path-connected if the following is true: for any a, b ∈ X there is a continuous map
γ : [0, 1] → X with γ(0) = a and γ(1) = b.

A continuous map γ : [0, 1] → X is called a path. To develop the basic theory of path-
connectedness, we introduce a couple of simple operations on paths.
 8.3. CONNECTEDNESS AND PATH-CONNECTEDNESS 53

Given two paths γ1 , γ2 in X such that γ1 (1) = γ2 (0) we can form the concatenation γ1 ⋆ γ2
of the two paths to be the path
(
γ1 (2t), 0 ⩽ t ⩽ 1/2
γ1 ⋆ γ2 (t) =
γ2 (2t − 1), 1/2 ⩽ t ⩽ 1.
We leave it as an easy exercise to show carefully that γ1 ⋆ γ2 is continuous, and hence really
is a path.
If γ : [0, 1] → X is a path, then the opposite path γ − is defined by γ − (t) = γ(1 − t).

Lemma 8.2.2. Let X be a metric space. Define a relation ∼ on X as follows: a ∼ b if and

only if there is a path γ : [0, 1] → X with γ(0) = a and γ(1) = b. Then ∼ is an equivalence
relation.

Proof. To show that a ∼ a, use the path γ which takes the constant value a. To show that
a ∼ b implies b ∼ a, take a path γ from a to b and consider its opposite path γ − . Finally, to
show transitivity, use the join of two paths.

The equivalence classes into which this relation partitions X are called the path-components
of X.

8.3. Connectedness and path-connectedness

In the final part of this chapter, we explore the link between connectedness and path-
connectedness. The key points to be covered are as follows:
• Path-connectedness implies connectedness;
• Connectedness does not imply path-connectedness in general, but it does for open
sets in normed vector spaces.

Theorem 8.3.1. A path-connected metric space is connected.

Proof. Suppose that X is path-connected, and let f : X → {0, 1}. We claim that f
is constant, which is enough to establish connectedness of X by Lemma 8.1.2 (ii). Let
a, b ∈ X. Since X is path-connected, there is a path γ : [0, 1] → X such that γ(0) = a
and γ(1) = b. Consider the composition f ◦ γ. This is a continuous function from [0, 1]
to {0, 1} and hence, since [0, 1] is connected by Theorem 8.1.8, it is constant. Therefore
f (a) = (f ◦ γ)(0) = (f ◦ γ)(1) = f (b). Since a and b were arbitrary, this implies that f is
indeed constant.

Theorem 8.3.2. A connected open subset of a normed space is path-connected.

Proof. Write X for the connected open set. The key observation is that any path-component
of X is open. To see this, suppose that P is a path-component of X, and let a ∈ P . Since
54 8. CONNECTEDNESS AND PATH-CONNECTEDNESS

X is open, there is a ball B(a, ε) contained in X. Let b be a point in this ball. We can now
write down an explicit path γ between a and b, namely γ(t) = (1 − t)a + tb. This is easily
seen to be continuous, and its image is contained in B(a, ε) since

∥γ(t) − a∥ = t∥a − b∥ ⩽ ∥a − b∥ = d(a, b) < ε

for all t. Therefore b lies in the same path-component P .

With this observation in place, the theorem follows easily. Indeed, the path-components
partition X, and so if there was more than one of them we could write X as a disjoint union
of non-empty open sets, contrary to the assumption that X is connected.

We now return to the topic of path-connectedness, and give a famous example.

Theorem 8.3.3. There is a connected subset of R2 which is not path-connected.

Proof. (Non-examinable) There is a classic example, known as the Topologist’s sine-curve.

This is the set A ⊆ R2 given by

{(0, y) : −1 ⩽ y ⩽ 1} ∪ {(x, sin(1/x) : x ∈ (0, 1]}.

Why is A connected? It is quite easy to convince oneself that A = Ē, where E =

{(x, sin(1/x) : x ∈ (0, 1]}. However, E is connected, being the image of the connected
set (0, 1] under a continuous map, and so the connectedness of A is immediate from Lemma
8.1.5.
Why is A not path-connected? It is “intuitively clear” that there is no path γ : [0, 1] → A
with γ(0) = (0, 0) and γ(1) = (1, sin(1)), but we must prove this. Suppose we have such a
path γ. Write ℓ for the vertical line {0} × [−1, 1], thus A = E ∪ ℓ. Since ℓ is closed in A,
γ −1 (ℓ) is closed, and in particular contains its supremum t. Thus γ(t) ∈ ℓ, whilst γ(u) ∈ E
for all u > t.
Let pY : R2 → R be projection onto the y-coordinate, i.e. pY (x, y) = y. Since pY is
continuous, so is the composition pY ◦ γ : [0, 1] → R. Thus there is some δ > 0 such that

(8.1) |pY (γ(u1 )) − pY (γ(u2 ))| ⩽ 1 for all u1 , u2 ∈ [t, t + δ].

Now let pX be projection onto the x-coordinate, i.e. pX (x, y) = x. The composition pX ◦γ
is continuous, and so by the intermediate value theorem and the fact that pX (γ(t + δ)) > 0,
(pX ◦ γ)[t, t + δ] contains some interval [0, c], c > 0.
However, as x ranges over (0, c], sin(1/x) takes all values in [−1, 1] (infinitely often), so
there are u1 , u2 ∈ [t, t + δ] such that pY (γ(u1 )) = 1, pY (γ(u2 )) = −1. This contradicts (8.1).
 CHAPTER 9

Sequential compactness

In this chapter (and in Chapter 9) we will be talking a lot about sequences and sub-
sequences, so let us be clear about what these concepts are. If X is some space, let
σ = (xn )∞ n=1 = (x1 , x2 , . . . ) be a sequence of elements of X. Any sequence of the form
σ = (xnk )∞
′
k=1 , where n1 < n2 < n3 < . . . , is called a subsequence of σ. For instance,
(x1 , x4 , x9 , x16 , . . . ) is a subsequence of (x1 , x2 , x3 , x4 , . . . ).

9.1. Definitions

In this chapter we study metric spaces which satisfy the metric-space analogue of the
Bolzano-Weierstrass property. Recall what the Bolzano-Weierstrass property of R is: any
bounded sequence has a convergent subsequence. More precisely, if (xn )∞ n=1 is a sequence
of elements in some closed bounded interval [a, b], there is a subsequence of the xn which
converges to some c ∈ [a, b].
There is an obvious way to generalise this notion to subsets of metric spaces, and the
resulting notion is called sequential compactness.

Definition 9.1.1 (Sequential compactness). Let X be a metric space. Then X is said to

be sequentially compact if any sequence of elements in X has a convergent subsequence.

Example 9.1.2. The closed interval [0, 1] is sequentially compact, by the Bolzano-Weierstrass
theorem.
The open interval (0, 1) is not sequentially compact. For instance, the sequence xn = 1/n
has no convergent subsequence in this space.
The set of rational numbers in [0, 1] is not sequentially compact – for instance, the sequence
0.1, 0.14, 0.141, 0.1415, . . . consisting of decimal approximations to π − 3 has no convergent
subsequence.
Finally, the real line R is not sequentially compact. For instance, the sequence xn = n
has no convergent subsequence in this space.

9.2. Closure and boundedness properties

In this section we prove a couple of basic lemmas about sequentially compact spaces.

Lemma 9.2.1. A sequentially compact subspace of a metric space is closed and bounded.
55
56 9. SEQUENTIAL COMPACTNESS

Proof. Let X be the space and Y the sequentially compact subspace.

Suppose first that Y is not closed. Then Ȳ \ Y is nonempty. Let a be a point in this set.
By Lemma 6.1.5, there is a sequence (yn )∞ n=1 of elements of Y with limn→∞ yn = a. Then
any subsequence of (yn )∞ n=1 converges to a and hence, by the uniqueness of limits, does not
converge to an element of Y . Therefore Y cannot be sequentially compact.
Suppose next that Y is not bounded. Pick an arbitrary point y0 ∈ Y , and pick a sequence
(yn )∞ ∞
n=1 such that d(y0 , yn ) ⩾ n for all n. Suppose there is a subsequence (ynk )k=1 converging
to b. Then for k sufficiently large we have d(ynk , b) < 1, which implies that

d(y0 , b) ⩾ d(y0 , ynk ) − d(ynk , b) ⩾ nk − 1.

Since nk → ∞ as k → ∞, whilst d(y0 , b) is a fixed finite quantity, this is a contradiction.

The converse is not true – for instance, take X = Y = (0, 1) (noting that Y is closed as
a subset of X).

Lemma 9.2.2. A closed subset of a sequentially compact metric space is sequentially com-
pact.

Proof. Let X be the space and Y the closed subspace. Consider a sequence (yn )∞ n=1 of
elements of Y . It is also a sequence of elements of X and so, by sequential compactness of
X, has a subsequence converging to some a ∈ X. However, Y is closed, so the limit of any
convergent sequence of elements of Y lies in Y . In particular, a ∈ Y .

9.3. Continuous functions on sequentially compact spaces

Sequential compactness has some nice properties with respect to continuous maps.

Lemma 9.3.1. The image of a sequentially compact metric space under a continuous map
is sequentially compact.

Proof. Let X be sequentially compact, and suppose that f : X → Y is continuous. Let

σ = (f (xn ))∞
n=1 be a sequence of elements of f (X). The sequence (xn ) contains a convergent
subsequence (xnk ) say, with xnk → a as k → ∞ for some a ∈ X. But then, since f is
continuous, we have f (xnk ) → f (a), and so σ ′ = (f (xnk ))∞
k=1 is a convergent subsequence of
σ.

As a consequence of Lemma 9.2.1, we see that continuous function f from a sequentially

compact metric space X to R has closed and bounded image, so in particular f is bounded
and attains its bounds.
Another consequence of Lemma 9.3.1 is if X and Y are homeomorphic metric spaces and
if X is sequentially compact, then so is Y .
 9.4. PRODUCT SPACES 57

Proposition 9.3.2. A continuous function from a sequentially compact metric space to

R is uniformly continuous.

Proof. Let X be a sequentially compact metric space, and suppose that f : X → R is

continuous but not uniformly continuous. Then there exists some ε > 0 such that for each
n ∈ N we may find an , bn ∈ X such that d(an , bn ) < 1/n but |f (an ) − f (bn )| ⩾ ε. Since
X is sequentially compact, (an )∞ ∞
n=1 has a subsequence, (ank )k=1 converging to some point ℓ.
Consider the corresponding sequence (bnk )∞ k=1 . Since d(ank , bnk ) ⩽ 1/nk → 0, it follows that
bnk also converges to ℓ as k → ∞.
Relabelling (to avoid double subscripts) we may now assume we have sequences (an )∞ n=1 ,
∞
(bn )n=1 with limn→∞ an = limn→∞ bn = ℓ and |f (an ) − f (bn )| ⩾ ε for all n.
Now f is continuous at ℓ, so there is a δ > 0 such that for all x ∈ X with d(ℓ, x) < δ, we
have |f (ℓ) − f (x)| < ε/2. If n is sufficiently large, we have d(ℓ, an ), d(ℓ, bn ) < δ and hence

ε ⩽ |f (an ) − f (bn )| ⩽ |f (an ) − f (ℓ)| + |f (ℓ) − f (bn )| < ε/2 + ε/2 = ε,

which is a contradiction.
We were therefore wrong to assume that f is not uniformly continuous.

9.4. Product spaces

Recall that if (X, dX ) and (Y, dY ) are metric spaces then their Cartesian product X × Y
can be equipped with a metric dX×Y by setting
p
dX×Y ((x1 , y1 ), (x2 , y2 )) = dX (x1 , x2 )2 + dY (y1 , y2 )2 .
The main result of this section, Proposition 9.4.2 below, is that the product of two se-
quentially compact spaces is compact. Before proving this, we note an important lemma.

Lemma 9.4.1. Let X and Y be metric spaces. A sequence ((xn , yn ))∞

n=1 in X ×Y converges
if and only if (xn )∞ ∞
n=1 converges in X and (yn )n=1 converges in Y .

Proof. The projection maps pX : X × Y → X and pY : X × Y → Y are continuous. In fact

it is easy to see that they are Lipschitz continuous with Lipschitz constant 1. It follows that
if limn→∞ (xn , yn ) = (a, b) then

lim xn = lim pX (xn , yn ) = pX (a, b) = a,

n→∞ n→∞

and similarly limn→∞ yn = b.

Conversely, if xn → a and yn → b then
p
dX×Y ((xn , yn ), (a, b)) = dX (xn , a)2 + dY (yn , b)2 → 0
58 9. SEQUENTIAL COMPACTNESS

as n → ∞ and so (xn , yn ) → (a, b) as n → ∞, as required.

Now we prove that the product of two sequentially compact spaces is compact, with
apologies for using a rather unpleasant triple subscript notation in the argument.

Proposition 9.4.2. The product of two sequentially compact metric spaces is sequentially
compact.

Proof. Let ((xn , yn ))∞

n=1 be a sequence in X ×Y . As X is sequentially compact, the sequence
σ = (xn )n=1 in X has a convergent subsequence σ ′ = (xnk )∞
∞
k=1 , with xnk → a as k → ∞.
∞
Now consider the sequence (ynk )k=1 in Y . Since Y is sequentially compact this in turn has
a convergent subsequence (ynkr )∞ ′′
r=1 , say ynkr → b as r → ∞. Let σ be the corresponding
subsequence of xs, that is to say σ ′′ = (xnkr )∞ ′′ ′
r=1 . Then σ is a subsequence of σ , and so it
converges to a.
By the previous Lemma it follows that (xnkr , ynkr ) → (a, b) as r → ∞, and so we have ex-
hibited a convergent subsequence of ((xn , yn ))∞ n=1 . Therefore X × Y is sequentially compact.

A corollary of this is the following result, which is often called the Bolzano-Weierstrass
theorem (being a generalisation of the version on R).

Corollary 9.4.3 (Bolzano-Weierstrass). Any closed and bounded subset of Rn is se-

quentially compact.

Proof. Let X ⊆ Rn be the set. Since X is bounded, it is contained in some cube [−M, M ]n .
The Bolzano-Weierstrass theorem on R implies that [−M, M ] is sequentially compact, and
therefore by Proposition 9.4.2, [−M, M ]n is sequentially compact. Since X is closed, it is
sequentially compact by Lemma 9.2.2.

9.5. Sequentially compact equals complete and totally bounded

As a warm-up to the main business of this section, we prove the following.

Proposition 9.5.1. A sequentially compact metric space X is complete and bounded.

The converse is not true in general.

Proof. Suppose that X is sequentially compact. We have already shown that X is bounded
in Lemma 9.2.1. Let us now show that X is complete. Suppose that (xn )∞ n=1 is a Cauchy
∞
sequence in X. Since X is sequentially compact, (xn )n=1 has a convergent subsequence
(xnk )∞
k=1 . Suppose that limk→∞ xnk = a. We claim that in fact limn→∞ xn = a.
 9.5. SEQUENTIALLY COMPACT EQUALS COMPLETE AND TOTALLY BOUNDED 59

Let ε > 0. Then, since (xn )∞ n=1 is Cauchy, there is some N such that for all n, m ⩾ N
we have d(xn , xm ) < ε/2. Since limk→∞ xnk = a, we may find a k such that nk ⩾ N and
d(xnk , a) < ε/2. But then if n ⩾ N we have

d(xn , a) ⩽ d(xn , xnk ) + d(xnk , a) < ε/2 + ε/2 = ε,

as required.
To show that the converse is not true in general, consider the following example. Take
Cb (R) to be the normed space of continuous bounded functions on the real line equipped as
usual with the ∥ · ∥∞ -norm and the associated metric. Let X = B̄(0, 1) (that is, functions
having sup norm bounded by 1). Then X is bounded. Moreover, by Lemma 7.2.1 as X is
closed and Cb (R) is complete (Theorem 7.3.2) we have that X is complete. Define a function
ϕ : R → R by (
2t + 1, −1/2 ⩽ t ⩽ 0;
ϕ(t) =
1 − 2t, 0 ⩽ t ⩽ 1/2
and ϕ(t) = 0 for t ∈ / [−1/2, 1/2]. For each n ∈ N set fn (t) = ϕ(t − n) (we might call
this sequence of functions a “moving bump”). All of the functions fn lie in X = B̄(0, 1).
However, if n ̸= m then fn (n) = 1, whilst fm (n) = 0, so ∥fn − fm ∥∞ = 1. Thus the sequence
(fn )∞
n=1 has no Cauchy subsequence, and hence certainly no convergent subsequence.

Remark. The Bolzano-Weierstass theorem (Corollary 9.4.3) implies that the converse is
true for subsets of Rn .
It turns out that there is a stronger notion of boundedness called total boundedness which
– together with completeness - implies sequential compactness and in fact is equivalent to it.

Definition 9.5.2. A metric space is said to be totally bounded if, for any ε > 0, it may
be covered by finitely many open balls of radius ε.

Here is one of the more substantial theorems of the course.

Theorem 9.5.3. A metric space is sequentially compact if and only if it is complete and
totally bounded.

Proof. Suppose first that we have a space X which is sequentially compact. We have
already shown in Proposition 9.5.1 that X is complete. Let us now show that it is totally
bounded. Suppose X is not totally bounded, and let ε be such that there is no way to cover
X by finitely many open balls of radius ε.
Using a greedy algorithm, we select an infinite sequence (xn )∞ n=1 of elements of X which
are separated by at least ε, that is to say d(xi , xj ) ⩾ ε whenever i ̸= j.
60 9. SEQUENTIAL COMPACTNESS

To do this, suppose that x1 , . . . , xn have already been selected. By assumption, the balls
B(xi , ε) do not cover X, and so we may select a point xn+1 ∈ X which does not lie in any
of these balls, and therefore d(xi , xn+1 ) ⩾ ε for i = 1, . . . , n.
It is clear that such a sequence has no convergent subsequence, and so we were wrong to
assume that X is not totally bounded.
We turn now to the more substantial direction of the theorem, which is to show that a
complete and totally bounded metric space X is sequentially compact. Let σ be a sequence
of elements of X. We will use the total boundedness assumption for balls of radii 1, 21 , 14 , . . . .
(m) (m)
Thus, for each nonnegative integer m there is a finite collection of open balls B1 , . . . , Bkm
of radius 2−m which cover X.
(0) (0)
Start with the balls B1 , · · · , Bk0 of radius 1. One of these balls contains infinitely many
elements of the sequence σ. Write B0 for the ball with this property, and let σ (0) be the
infinite subsequence of σ of elements contained in this ball.
(1) (1)
Now look at the balls B1 , . . . , Bk1 of radius 21 . One of these balls contains infinitely
many elements of the new subsequence σ (0) . Write B1 for such a ball, and let σ (1) be the
infinite subsequence of σ (0) of elements contained in it.
Continue in the obvious fashion, producing new subsequences σ (2) , σ (3) , . . . with σ (r) con-
tained in Br and a subsequence of σ (r−1) .
Now consider the sequence σ ∗ obtained by a diagonal argument: the ith element of σ ∗
is taken to be the ith element of σ (i) . Clearly σ ∗ is a subsequence of σ and, if we write
σ ∗ = (xn )∞n=1 , we have xn ∈ Br for all n ⩾ r.
It is now clear that σ ∗ is a Cauchy sequence. Indeed, given ε > 0, let N be such that
2−N < ε/2. If n, m ⩾ N then xn , xm both lie in BN , which is a ball of radius 2−N , and hence
d(xn , xm ) < ε/2 + ε/2 = ε.
Finally, since X is complete the sequence σ ∗ converges. We have shown that σ, which was
an arbitrary sequence in X, has a convergent subsequence, and therefore X is sequentially
compact.

Remark. Observe that the argument in fact shows that any sequence in a totally bounded
metric space has a subsequence which is Cauchy. We only used completeness right at the
end.
 CHAPTER 10

Compactness

10.1. Open covers and the definition of compactness

In this final chapter of the metric spaces part of the course, we come to one of the most
powerful and important notions in all of mathematics: compactness.
Let us start by giving the definition.

Definition 10.1.1. Let X be a metric space and U = {Ui : i ∈ I} a collection of open

S
subsets of X. We say that U is an open cover of X if X = i∈I Ui . If J ⊆ I is a subset such
S
that X = i∈J Ui then we say that {Ui : i ∈ J} is a subcover of U and if |J| < ∞ then we
say that it is a finite subcover.

Definition 10.1.2 (Compactness). A metric space is said to be compact if every open

cover has a finite subcover.
S
Example 10.1.3. The real line R is not compact. For instance, the open cover n∈N (−n, n)
has no finite subcover.

Motivation. It is quite hard to motivate the definition of compactness when one first sees
it. Indeed, von Neumann’s famous quote “... in mathematics you don’t understand things.
You just get used to them” is quite apposite. Nonetheless, a couple of comments are in
order. First of all, it turns out that compactness and sequential compactness are the same
concept in metric spaces. We prove this in Sections 10.2 and 10.3 below (with the second of
these being non-examinable). Second, the notion of compactness looks rather natural in the
context of topological spaces, since it talks about open sets in a very basic way. Whilst the
notion of sequential compactness can also be formulated in topological spaces, it is somehow
less basic and, in this more general situation, not equivalent to compactness.

Subspaces. Sometimes, we will have a metric space X and a subspace Y ⊆ X, and we

wish to talk about whether Y is compact. In this context, by convention an open cover U of
S
Y is a collection {Ui : i ∈ I} of open subsets of X, such that Y ⊆ i∈I Ui . A subcollection
S
{Ui : i ∈ J} is called a subcover if Y ⊆ i∈J Ui .
Then Y is compact if and only if every open cover has a finite subcover. The reason this
notion is the same as the previous one (which was “internal to Y ”, making no reference to
61
62 10. COMPACTNESS

open sets in X) is Lemma 5.5.1, which says that open sets in Y are the same thing as open
sets in X intersected with Y .
It should be said that this abuse of nomenclature of using the phrase “open cover” in two
slightly different ways can be a touch confusing when you first see it.

10.2. Compactness implies sequential compactness

Proposition 10.2.1. A compact metric space is sequentially compact.

We isolate a lemma from the proof.
Lemma 10.2.2. Suppose that X is a compact metric space and that we have a nested
sequence S1 ⊇ S2 ⊇ S3 ⊇ · · · of nonempty, closed subsets of X. Then the intersection
T∞
n=1 Sn is nonempty.

Remark. You might be interested in comparing this with Lemma 7.2.3, where a similar
conclusion was reached assuming that X is complete and that the diameters of Si tend to 0.

Proof. Suppose the intersection is empty. Then the complements Sic (which are open sets)
are an open cover of X. By compactness, there is a finite subcover. In particular, for some n
the sets S1c , . . . , Snc cover X. However, we have S1c ⊆ S2c ⊆ · · · ⊆ Snc , and therefore Snc covers
(is equal to) X. But this is a contradiction, since Sn is nonempty.

Proof. (Proof of Proposition 10.2.1.) Let X be the space in question, and suppose that
(xn )∞ n=1 is a sequence of elements of X. We wish to find a convergent subsequence of this
sequence.
For each natural number n, set An := {xn , xn+1 , xn+2 , . . . }. Obviously, A1 ⊇ A2 ⊇ A3 ⊇
· · · , and so Ā1 ⊇ Ā2 ⊇ Ā3 ⊇ · · · . Applying Lemma 10.2.2, we see that ∞
T
n=1 Ān is nonempty.
Let a be a point in this intersection. We inductively construct a subsequence (xnk )∞ k=1
such that d(xnk , a) < 1/k for all k; it is then clear that this subsequence converges (to a) and
the proof will be complete. Suppose that n1 , . . . , nk have already been constructed. Now a
lies in Ānk +1 , that is to say the closure of the set {xnk +1 , xnk +2 , . . . }. In particular, there is
some element of this sequence at distance less than 1/(k + 1) from a, and we can take this
to be our xnk+1 .

10.3. Sequential compactness implies compactness

The converse of Proposition 10.2.1 is also true.

Proposition 10.3.1. A sequentially compact metric space is compact.
As a consequence of this, Proposition 10.2.1 and Theorem 9.5.3, we have the following
substantial and important theorem.
 10.3. SEQUENTIAL COMPACTNESS IMPLIES COMPACTNESS 63

Theorem 10.3.2. Let X be a metric space. Then the following are equivalent:
(i) X is compact;
(ii) X is sequentially compact;
(iii) X is complete and totally bounded.

Lemma 9.2.1, Corollary 9.4.3, Proposition 10.2.1 and Proposition 10.3.1 also imply the
following important result.

Theorem 10.3.3 (Heine-Borel). A subset of Rn is compact if and only if it is closed and

bounded.

We turn now to the proof of Proposition 10.3.1, which is nonexaminable.

Proof. (Proof of Proposition 10.3.1, non-examinable.) Let X be a sequentially compact met-
ric space. By (the easy direction of) Proposition 9.5.3, X is complete and totally bounded.
(m) (m)
For m = 1, 2, 3, . . . , fix some collection of balls B1 , . . . , Bkm of radius 2−m which cover X.
Suppose we have an open cover of X by sets Ui , i ∈ I, which has no finite subcover. Then
(1)
one of the balls Bj is not covered by finitely many of the Ui ; let us write B1 for this ball.
(2)
Now consider the balls Bj which intersect B1 . One of these is not covered by finitely
many of the Ui (otherwise B1 would be). Write B2 for this ball.
(3)
Now consider the balls Bj which intersect B2 , and so on.
Continuing in this fashion, we obtain a sequence B1 , B2 , . . . of open balls, with Bm having
radius 2−m , Bm ∩ Bm+1 ̸= ∅ for all m, and with none of the Bj covered by finitely many of
the Ui . Let xm be the centre of Bm . Then, since Bm and Bm+1 intersect in some point t, we
have
d(xm , xm+1 ) ⩽ d(xm , t) + d(xm+1 , t) < 2−m + 2−(m+1) < 2 · 2−m .
By the triangle inequality and summing the geometric series, it follows that for any n ⩾ m
we have

d(xm , xn ) ⩽ d(xm , xm+1 ) + · · · + d(xn−1 , xn )

< 2(2−m + 2−(m+1) + · · · ) = 4 · 2−m .

Therefore (xn )∞n=1 is a Cauchy sequence. Since X is complete, we have limn→∞ xn = x for
some x ∈ X. Since the sets Ui cover X, one of them must contain x. Let us suppose U1
contains x. Then, since U1 is open, some ball B(x, ε) is contained in U1 .
Choose n large enough that d(xn , x) < ε/2, and also that 2−n < ε/2. Recalling that Bn
is the ball of radius 2−n centred on xn , it follows that Bn ⊆ B(x, ε). But then Bn ⊆ U1 ,
contrary to the assumption that Bn is not covered by finitely many of the Ui .
We were wrong to assume the existence of an open cover of X with no finite subcover,
and so X is indeed compact.
 CHAPTER 11

Conformal maps

11.1. The extended complex plane C∞

It is a remarkable fact that it is possible to “add the point at infinity” to C in such a way
that the resulting space C∞ has pleasant analytic properties. For instance, one can extend
the function f (z) = 1/z to a continuous bijection on this space, by setting f (0) = ∞ and
f (∞) = 0, and one can make rigorous sense of such statements as ∞ + 1 = ∞. The aim
of this section is to study C∞ , which is known as the extended complex plane. Note that
C∞ is a very basic example of a Riemann surface, one of the main objects of study in the
Geometry of Surfaces course in Part B.

11.1.1. Stereographic projection. Let

S = {(x, y, z) ∈ R3 : x2 + y 2 + z 2 = 1}

be the unit sphere of radius 1 centred at the origin in R3 . View the complex plane C as
the copy of R2 inside R3 given by the plane {(x, y, 0) ∈ R : x, y ∈ R}. Thus z = x + iy
corresponds to the point (x, y, 0). Let N be the “north pole” N = (0, 0, 1) of S.
We can define a bijective map S : C → S \ {N } as follows. To determine S(z), join z to
N by a straight line, and let S(z) be the point where this line meets the sphere S. This map
(or more accurately its inverse) is called stereographic projection.
It is not too hard to give an explicit formula for S(z).

Lemma 11.1.1. Suppose that z = x + iy. Then

x2 + y 2 − 1
 
2x 2y
S(z) = , , .
x2 + y 2 + 1 x2 + y 2 + 1 x2 + y 2 + 1

Proof. The general point on the line joining z and N is t(0, 0, 1) + (1 − t)(x, y, 0). There
is a unique value of t with t ̸= 1 for which this point lies on the sphere, namely t =
(x2 + y 2 − 1)/(x2 + y 2 + 1), as can be easily checked.

We remark that the same formula can be written in the alternative form
1 2


S(z) = 2ℜ(z), 2ℑ(z), |z| − 1 .
1 + |z|2
65
66 11. CONFORMAL MAPS

Figure 1. The stereographic projection map.

As we have seen, C may be identified with S \ {N } by stereographic projection. The set

S \ {N } has a natural metric, namely the one induced from the Euclidean metric on R3 .
This induces a metric d on C, the unique metric on C such that S is an isometry. To spell
it out,
d(z, w) := ∥S(z) − S(w)∥.
Here is a formula for this metric.

Lemma 11.1.2. For any z, w ∈ C we have

2|z − w|
d(z, w) = p p
1 + |z|2 1 + |w|2
Proof. Since ∥S(z)∥ = ∥S(w)∥ = 1 we have ∥S(z)−S(w)∥2 = 2−2⟨S(z), S(w)⟩, where ⟨, ⟩ is
the usual Euclidean inner product on R3 . Using the formulæ (and after a little computation),
2|z − w|2
⟨S(z), S(w)⟩ = 1 − .
(1 + |z|2 )(1 + |w|2 )
Therefore
4|z − w|2
∥S(z) − S(w)∥2 =
(1 + |z|2 )(1 + |w|2 )
as required.

This d is not the same as the usual metric. However, it is very similar to it. For instance,
on any bounded set {z ∈ C : |z| ⩽ K} we have

c1 |z − w| ⩽ d(z, w) ⩽ c2 |z − w|

for some c1 , c2 > 0 depending on K. In fact, we could take c2 = 2 and c1 = K12 for K ⩾ 1
(exercise). That is, d is strongly equivalent to the usual metric on any such set, in the sense
 11.1. THE EXTENDED COMPLEX PLANE C∞ 67

described in Section 4.3 of the metric spaces part of the course. Therefore d is equivalent
(but not strongly equivalent) to the usual metric on all of C. Recall what this means: any
ball B(z, ε) in the usual metric is contained in some ball Bd (z, ε′ ) in the metric d, and vice
versa.
Therefore, as remarked in the metric spaces notes, notions such as limit and continuity
are the same whether we work with the usual metric or with d.

11.1.2. Adding in ∞. Now it is time to add in the point at infinity, which we will call
∞ (note this is just a symbol).
Now (exercise) as |z| → ∞, S(z) → N . Therefore, once we have identified C with S\{N },
it is natural to identify ∞ with N , and hence C∞ = C ∪ {∞} with the whole sphere S. We
extend the map S to a map S : C∞ → S by defining S(∞) = N .
Using, once again, the Euclidean metric on S, we can extend d to a metric on C∞ , the
unique metric for which the map S is an isometry.

Lemma 11.1.3. For any z ∈ C we have

2
d(z, ∞) = p .
1 + |z|2
Proof. By definition, d(z, ∞) = ∥S(z) − S(∞)∥ = ∥S(z) − N ∥, where N is the north pole
on the sphere. We may now proceed in much the same way as before, except the calculation
is easier this time. The details are left as an exercise.

We turn now to a few examples, which show that adding ∞ to C in this way leads to a
space with nice analytic properties.

Example 11.1.4 (Translations). Let a ∈ C. Define f : C∞ → C∞ by f (z) = z + a for

z ∈ C and f (∞) = ∞. Then f is a continuous bijection.

Proof. Clearly f is continuous with respect to the usual metric on C. Therefore, restricted
to C, it is also continuous with respect to d, since d is equivalent to the usual metric.
It remains
q q at ∞. Let ε > 0. Now if δ > 0 and if d(z, ∞) < δ then
to check continuity
4 4
|z| > δ2
− 1 and so |f (z)| > δ2
− 1 − |a|. This tends to ∞ as δ → 0, so by choosing δ
small enough in terms of ε it will follows that d(f (z), ∞) = √ 2 2 < ε.
1+|f (z)|

Example 11.1.5 (Dilations). Let b ∈ C with b ̸= 0. Define f : C∞ → C∞ by f (z) = bz

for z ∈ C and f (∞) = ∞. Then f is a continuous bijection.

Proof. This is very similar to the argument for translations and we leave the details as an
exercise.
68 11. CONFORMAL MAPS

The final example is the most interesting one.

Example 11.1.6 (Inversion). Define f : C∞ → C∞ by f (z) = 1/z for z ∈ C \ {0},

f (0) = ∞ and f (∞) = 0. Then f is a continuous bijection.

Proof. As before, the equivalence of d and the usual metric on C means that f is continuous
except possibly at 0 and ∞.
We prove that f is continuous at 0, leaving the continuity at ∞ as an exercise (similar to
Example 11.1.4).
Let ε > 0 be small. Then there is δ such that √ 2t 2 ⩽ ε for all t ∈ [0, δ]. If |z| < δ, then
1+t

2 2|z|
d(f (z), f (0)) = d(1/z, ∞) = q =p ⩽ ε.
1+ 1 1 + |z|2
|z|2

This indeed shows that f is continuous at 0.

There is a nice way to analyse Example 11.1.6, by considering what f looks like under
the identification of C∞ with the unit sphere S. One can easily check using Lemma 11.1.1
that if S(z) = (s, u, v) ∈ S then S(f (z)) = (s, −u, −v). That is, under the identification
S : C∞ → S, f corresponds to the (obviously continuous) map (s, u, v) 7→ (s, −u, −v), that
is to say rotation by π about the x-axis.

11.1.3. Möbius maps. In this subsection and subsequent ones we look at an important
class of maps from C∞ to itself, the Möbius maps.

! The general linear group GL2 (C) consists of all nonsingular 2 × 2

Definition 11.1.7.
a b
matrices g = with a, b, c, d ∈ C, with the group operation being matrix multiplica-
c d
tion.

Each element g ∈ GL2 (C) gives a Möbius map Ψg : C∞ → C∞ . Roughly, this is given by
the formula
az + b
Ψg (z) := ,
cz + d
but one needs to be careful about ∞, as follows:
• If c ̸= 0 then we define Ψg (−d/c) = ∞ and Ψg (∞) = a/c;
• If c = 0 then we define Ψg (∞) = ∞.
We remark that two elements g, g ′ ∈ GL2 (C) give the same Möbius map if g = λg ′ for some
λ ̸= 0.
An important fact about Möbius maps is that composing them corresponds to multiplying
the relevant matrices. That is to say, we have the following.
 11.1. THE EXTENDED COMPLEX PLANE C∞ 69

Proposition 11.1.8 (Composition of Möbius maps). We have Ψg1 g2 = Ψg1 ◦ Ψg2 . That
is, GL2 (C) acts on C∞ via Möbius maps.

One can prove this by a direct calculation (which we leave as an exercise).

11.1.4. Decomposing Möbius maps. Earlier in this section, we looked at transla-

tions, dilations and inversion from C∞ to C∞ . It turns out that these are all Möbius maps,
and moreover that an arbitrary Möbius map can be built from maps of these types.
That they are all Möbius maps is straightforward. Indeed,
!
1 a
• The translation z 7→ z + a is the Möbius map ΨT (a) , where T (a) = ;
0 1
!
b 0
• The dilation z 7→ bz is the Möbius map ΨD(b) , where D(b) = ;
0 1
!
0 1
• The inversion z 7→ 1/z is the Möbius map ΨJ , where J = .
1 0

Lemma 11.1.9. Every Möbius map can be written as a composition of translations, dila-
tions and inversions.
!
a b
Proof. Let Ψg , g = , be the Möbius map we are interested in. Suppose first that
c d
c ̸= 0. Then, putting aside any worries about ∞, we have the following chain of compositions:
ΨD(c) ΨT (d) J Ψ 1 Ψ bc−ad
D( c ) b − ad
c
ΨT ( a ) az + b
z −−−→ cz −−−→ cz + d −→ −−−−− −→ −−−c→ .
cz + d cz + d cz + d
This certainly suggests (very strongly!) that

Ψg = ΨT ( ac ) ◦ ΨD( bc−ad ) ◦ ΨJ ◦ ΨT (d) ◦ ΨD(c) .

c

A rigorous proof follows from Proposition 11.1.8 and the following identity of matrices (which
is of course an easy check):
!  
a b a bc − ad
=T ·D · J · T (d) · D(c).
c d c c
The case c = 0 is much easier and we leave this as an exercise (in this case only a dilation
and a translation are required).

11.1.5. Basic geometry of Möbius maps. In this final subsection, we look at one
key example of how Möbius maps transform C∞ .

Definition 11.1.10. A circline is either

70 11. CONFORMAL MAPS

• A circle in C (considered as a subset of C∞ ) or

• A line in C (considered as a subset of C∞ ) together with the point {∞}.

Note that lines in C are given by equations of the form |z − a| = |z − b| for a and b distinct
in C (think about why).

Proposition 11.1.11. Möbius maps take circlines to circlines.

Proof. (sketch: non-examinable) By Lemma 11.1.9 it is enough to check this for transla-
tions, dilations and inversions. So it is a straightforward case-by-case analysis. (The only
interesting cases being circles through the origin which map under inversion to lines not
through the origin, and vice-versa.)

11.2. Conformal transformations

Another important feature of the stereographic projection map is that it is conformal,

meaning that it preserves angles. The following definition helps us to formalize what this
means:

Definition 11.2.1. If γ : [−1, 1] → C is a C 1 path which has γ ′ (t) ̸= 0 for all t, then we
say that the line {γ(t) + sγ ′ (t) : s ∈ R} is the tangent line to γ at γ(t), and the vector γ ′ (t)
is a tangent vector at γ(t) ∈ C.

Note that this definition gives us a notion of tangent vectors at points on paths in subsets
of Rn , since the notion of a C 1 path extends readily to such paths (we just require all n
component functions are continuously differentiable). In particular, if γ : [−1, 1] → S ⊂ R3
is a C 1 path, it is easy to check that the tangent vectors at a point γ(t) ∈ S all lie in the
plane perpendicular to γ(t) – simply differentiate the identity f (γ(t)) = 1 where f (x, y, z) =
x2 + y 2 + z 2 using the chain rule.
We can now state what we mean by a conformal map:

Definition 11.2.2. Let U be an open subset of C and suppose that T : U → C (or S)

is continuously differentiable in the real sense (so all its partial derivatives exist and are
continuous). If γ1 , γ2 : [−1, 1] → U are two paths with z0 = γ1 (0) = γ2 (0) then γ1′ (0) and
γ2′ (0) are two tangent vectors at z0 , and we may consider the angle between them (formally
speaking this is the difference of their arguments). By our assumption on T , the compositions
T ◦ γ1 and T ◦ γ2 are C 1 -paths through T (z0 ), thus we obtain a pair of tangent vectors at
T (z0 ). We say that T is conformal at z0 if for every pair of C 1 paths γ1 , γ2 through z0 , the
angle between their tangent vectors at z0 is equal to the angle between the tangent vectors
at T (z0 ) given by the C 1 paths T ◦ γ1 and T ◦ γ2 . We say that T is conformal on U if it is
conformal at every z ∈ U .
 11.2. CONFORMAL TRANSFORMATIONS 71

One of the main reasons we focus on conformal maps here is because holomorphic functions
give us a way of producing many examples of them, as the following result shows.

Proposition 11.2.3. Let f : U → C be a holomorphic map and let z0 ∈ U be such

that f ′ (z0 ) ̸= 0. Then f is conformal at z0 . In particular, if f : U → C has nonvanishing
derivative on all of U , it is conformal on all of U .

Proof. We need to show that f preserves angles at z0 . Let γ1 and γ2 be C 1 -paths with
γ1 (0) = γ2 (0) = z0 . Then we obtain paths η1 , η2 through f (z0 ) where η1 (t) = f (γ1 (t)) and
η2 (t) = f (γ2 (t)). We show that a version of the chain rule applies to these compositions.
For i = 1, 2 we have

f (γi (h)) − f (γ(0)) f (γi (h)) − f (z0 ) γi (h) − z0

ηi′ (0) = lim = lim ·
h→0 h h→0 γi (h) − z0 h
For suitably small h, γi (h) ̸= z0 as γi′ (0) ̸= 0. (If there was arbitrarily small h > 0 with
f (γi (h)) − f (z0 )
γi (h) = z0 , then the limit γi′ (0) could only be zero.) Also limh→0 = f ′ (z0 ).
γi (h) − z0
So if we set f ′ (z0 ) = ρeiθ we have

ηi′ (0) = f ′ (z0 )γi′ (0) = ρeiθ γi′ (0), i = 1, 2.

Hence if ϕ1 and ϕ2 are the arguments of γ1′ (0) and γ2′ (0), then the arguments of η1′ (0) and
η2′ (0) are ϕ1 + θ and ϕ2 + θ respectively. It follows that the difference between the two pairs
of arguments, that is, the angles between the curves at z0 and f (z0 ), are the same.

Example 11.2.4. The function f (z) = z 2 has f ′ (z) nonzero everywhere except the origin.
It follows f is a conformal map from C× to itself. Note that the condition that f ′ (z) is non-
zero is necessary – if we consider the function f (z) = z 2 at z = 0, f ′ (z) = 2z which vanishes
precisely at z = 0, and it is easy to check that at the origin f in fact doubles the angles
between tangent vectors.

Lemma 11.2.5. The sterographic projection map S : C → S is conformal.

Proof. (Non-examinable sketch) Let z0 be a point in C, and suppose that γ1 (t) = z0 + tv1
and γ2 (t) = z0 + tv2 are two paths having tangents v1 and v2 at z0 = γ1 (0) = γ2 (0). Then
the lines L1 and L2 they describe, together with the point N , determine planes H1 and H2 in
R3 , and moreover the image of the lines under stereographic projection is the intersection of
these planes with S. Since the intersection of S with any plane is either empty or a circle, it
follows that the paths γ1 and γ2 get sent to two circles C1 and C2 passing through P = S(z0 )
and N . Now by symmetry, these circles meet at the same angle at N as they do at P . Now
the tangent lines of C1 and C2 at N are just the intersections of H1 and H2 with the plane
72 11. CONFORMAL MAPS

tangent to S at N . But this means the angle between them will be the same as that between
the intersection of H1 and H2 with the complex plane, since it is parallel to the tangent
plane of S at N . Thus the angles between C1 and C2 at P and L1 and L2 at z0 coincide as
required.

Although it follows easily from what we have already done, it is worth highlighting the
following:

Lemma 11.2.6. Möbius transformations are conformal.

Proof. As we have already shown, any holomorphic map is conformal wherever its derivative
is nonzero.
For a Möbius map we have f (z) = az+b
cz+d
and
ad − bc
f ′ (z) = ̸= 0,
(cz + d)2
for all z ̸= −d/c, thus f is conformal at each z ∈ C\{−d/c}.

Since a Möbius map is given by the four entries of a 2 × 2 matrix, up to simultaneous

rescaling, the following result is perhaps not too surprising.

Proposition 11.2.7. If z1 , z2 , z3 and w1 , w2 , w3 are triples of pairwise distinct complex

numbers, then there is a unique Möbius transformation f such that f (zi ) = wi for each
i = 1, 2, 3.

Proof. It is enough to show that, given any triple (z1 , z2 , z3 ) of complex numbers, we can
find a Möbius transformations which takes z1 , z2 , z3 to 0, 1, ∞ respectively. Indeed if f1 is
such a transformation, and f2 takes 0, 1, ∞ to w1 , w2 , w3 respectively, then clearly f2 ◦ f1−1
is a Möbius transformation which takes zi to wi for each i.
Now consider
(z − z1 )(z2 − z3 )
f (z) =
(z − z3 )(z2 − z1 )
It is easy to check that f (z1 ) = 0, f (z2 ) = 1, f (z3 ) = ∞, and clearly f is a Möbius transfor-
mation as required. If any of z1 , z2 or z3 is ∞, then one can find a similar transformation (for
2 −z3
example by letting zi → ∞ in the above formula). Indeed if z1 = ∞ then we set f (z) = zz−z 3
;
z−z1 z−z1
if z2 = ∞, we take f (z) = z−z3 ; and finally if z3 = ∞ take f (z) = z2 −z1 .
To see the f is unique, suppose f1 and f2 both took z1 , z2 , z3 to w1 , w2 , w3 . Then taking
Möbius transformations g, h sending z1 , z2 , z3 and w1 , w2 , w3 to 0, 1, ∞ the transformations
hf1 g −1 and hf2 g −1 both take (0, 1, ∞) to (0, 1, ∞). But suppose T (z) = az+b cz+d
is any Möbius
transformation with T (0) = 0, T (1) = 1 and T (∞) = ∞. Since T fixes ∞ it follows c = 0.
Since T (0) = 0 it follows that b/d = 0 hence b = 0, thus T (z) = a/d · z, and since T (1) = 1
 11.2. CONFORMAL TRANSFORMATIONS 73

it follows a/d = 1 and hence T (z) = z. Thus we see that

hf1 g −1 = hf2 g −1 = id,

and so f1 = f2 as required.

Example 11.2.8. The above lemma shows that we can use Möbius transformations as a
source of conformal maps. For example, suppose we wish to find a conformal transformation
which takes the upper half plane H = {z ∈ C : ℑ(z) > 0} to the unit disk B(0, 1). The
boundary of H is the real line, and we know Möbius transformations take lines to lines or
circles, and in the latter case this means the point ∞ ∈ C∞ is sent to a finite complex
number. Now any circle is uniquely determined by three points lying on it, and we know
Möbius transformations allow us to take any three points to any other three points. Thus
if we take f the Möbius map which sends 0 7→ −i, and 1 7→ 1, ∞ 7→ i the real axis will be
sent to the unit circle. Now we have
iz + 1
f (z) =
z+i
(one can find f in a similar fashion to the proof of Proposition 11.2.7).
So far, we have found a Möbius transformation which takes the real line to the unit circle.
Since C\R has two connected components, the upper and lower half planes, H and iH,
and similarly C\S1 has two connected components, B(0, 1) and C\B̄(0, 1). Since a Möbius
transformation is continuous, it maps connected sets to connected sets, thus to check whether
f (H) = B(0, 1) it is enough to know which component of C\S1 a single point in H is sent
to. But f (i) = 0 ∈ B(0, 1), so we must have f (H) = B(0, 1) as required.