Anda di halaman 1dari 66

Topology 2008/2009

Notes c _ S Carter and J M Speight 2008


0 Preliminaries
0.1 Contact Details
Lecturer: Martin Speight
Oce: School of Mathematics, 9.19i
E-mail: speight@maths.leeds.ac.uk
Phone: 343 5169
Web: http://www.maths.leeds.ac.uk/speight/teaching/math3224
All handouts will be made available on the above webpage.
0.2 Glossary
This course will make heavy use of set theoretic notation. Youve met almost all of it before but
here is a glossary of terms and symbols for future reference.
x A, A x: object x is an element of set A, set A contains object x.
A B, B A: A is a subset of B, that is, x A x B. Note that A A (we will not use
the symbol in this course).
A B = x : x A and x B, the intersection of A and B.
A B = x : x A or x B, the union of A and B.
A B = x : x A and x / B, the complement of B in A.
A B = (a, b) : a A and b B, the Cartesian product of sets A and B (the set of all
ordered pairs).
R the set of real numbers
Z the set of integers
Z
+
the set of positive integers (note 0 / Z
+
)
N the set of non-negative integers, or natural numbers (note 0 N)
Q the set of rational numbers
the empty set (the unique set containing no elements).
f : X Y , f is a function mapping domain X to codomain Y .
For any x X, f(x) is an element of Y called the image of x under f.
f(A) = f(x) : x A, the image of A X under f. Note f(A) Y . In particular, f(X) is
the range of f.
f
1
(B) = x X : f(x) B, the preimage of B Y under f (the collection of objects in
X which get mapped to B). Note f
1
(B) X, and this set is well dened whether or not
function f has an inverse function.
f is injective (one-to-one): f(x
1
) = f(x
2
) x
1
= x
2
.
f is surjective (onto): f(X) = Y , the range is the whole codomain.
f is bijective: f is both injective and surjective.
1
1 Introduction: what is topology?
Given a set X, what does it mean to say that a function
f : X X
is continuous? Or if we have an innite sequence (x
n
) of elements of X, what does it mean to say
(x
n
) converges? If X happens to be R we know how to answer these questions (because we have a
photographic memory of our Real Analysis modules). But what if X is something more exotic, for
example, the set of 1-dimensional vector subspaces of R
6
? A priori, continuity and convergence in
such a set have no meaning.
A topology is the extra structure that we must equip the set X with in order for notions like
continuity and convergence to make sense. Once X has a topology, we can meaningfully think of it
as a space, rather than just a collection of unrelated objects. It is crucial to note that a given set
may allow many dierent topologies, and that the spaces associated with each of these may have
very dierent (and sometimes counter-intuitive) properties. This applies even to the most familiar
sets, such as R. Consider the function
f : R R, f(x) =
_
0 x is rational
1 x is irrational.
Hopefully you recall that this function is not continuous. However, it is possible to equip R with
a topology such that any function R R is continuous, including the pathological example above.
Actually, there is more than one topology with this bizarre characteristic.

What is the point of this? Why should we want to equip sets other than R, C etc. with natural
notions of continuity and convergence? Is this just pure mathematical abstraction for abstractions
sake? Emphatically not! Topology turns out to be extremely useful for many real world applications.
For example, lets say we want to solve the ordinary dierential equation
dy
dx
= x(x +y)
with initial data y(0) = 0. I have no idea how to solve this equation (do you?). In fact, how do
we know that a solution even exists? Towards the end of the course, we will use a deep topological
result called the Contraction Mapping Theorem to prove that a solution to this initial value problem
exists, and is unique. Furthermore, the theorem will give us a method to approximate this solution
to arbitrarily high accuracy. The key idea is to think of the solution y(x) as a point in a suitable
space of functions, equipped with the right topology. This point of view is fundamental to modern
analysis.
Where to begin? Since our goal is to nd an abstract generalization of the idea of continuity,
lets rst recall that a function f : R R is continuous at a if
for all > 0 there exists > 0 such that [x a[ < implies [f(x) f(a)[ < ,
or, to paraphrase, if for all > 0, there exists > 0 such that we can ensure that f(x) lies within
distance of f(a) by demanding that x lies within distance > 0 of a. The obvious way to
generalize this to functions f : X X is to equip X with a notion of distance. A set equipped
with a distance function is called a metric space, so we will start with these. We will then use the
general properties of metric spaces as a guide for how to dene topological spaces in the completely
abstract sense.

A word of reassurance: there is a topology on R called the usual or standard topology, with respect to which
the denitions of continuity and convergence reduce to the familiar Real Analysis denitions you know and love.
2
2 Metric spaces
2.1 Basic denitions
Denition 1 Let X be a set. A metric or distance function on X is a function d : X X R
which has the properties
(i) d(x, y) = 0 if and only if x = y
(ii) d(x, y) = d(y, x) for all x, y X (symmetry)
(iii) d(x, y) +d(y, z) d(x, z) for all x, y, z X (the triangle inequality).
The number d(x, y) is called the distance between x and y. Note that it cant be negative since
putting z = x in (iii) gives
d(x, y) +d(y, x) d(x, x) = 0 by (i)
d(x, y) +d(x, y) = 2d(x, y) 0 by (ii)
Example 2 (a). For any (nonempty) set X, d : X X R such that
d(x, y) =
_
0 x = y
1 x ,= y
is a metric on X. The distance between any pair of distinct elements in X is unity by denition.
(Exercise: show that the triangle inequality holds.) This metric is called the discrete metric.
(b). d : R R R such that d(x, y) = [x y[ is a metric on R. If we visualize R as an innite
straight line, then d(x, y) is what we intuitively think of as the distance between the numbers
x and y.
(c). d : R
n
R
n
R such that
d(x, y) =

_
n

i=1
(x
i
y
i
)
2
is a metric on R
n
. Clearly this function satises the rst two axioms of a metric, but its not
so obvious that is satises the triangle inequality. To show this, we will need
Cauchys inequality:
_
n

i=1
a
i
b
i
_
2

_
n

i=1
a
2
i
__
n

i=1
b
2
i
_
which, for completeness, we will now prove. For all t R,
0
n

i=1
(a
i
+tb
i
)
2
= t
2
n

i=1
b
2
i
+ 2t
n

i=1
a
i
b
i
+
n

i=1
a
2
i
= p(t)
so the quadratic polynomial p(t) has at most one real root. Hence its discriminant is non-
positive:
b
2
4ac = 4
_
n

i=1
a
i
b
i
_
2
4
_
n

i=1
b
2
i
__
n

i=1
a
2
i
_
0,
and Cauchys inequality immediately follows.
3
Returning to the triangle inequality for our example, for all x, y, z R
n
,
(d(x, y) +d(y, z))
2
=
_
_

(x
i
y
i
)
2
+
_

(y
i
z
i
)
2
_
2
=

(x
i
y
i
)
2
+ 2
_

(x
i
y
i
)
2

(y
i
z
i
)
2
+

(y
i
z
i
)
2

(x
i
y
i
)
2
+ 2
_
(

(x
i
y
i
)(y
i
z
i
))
2
+

(y
i
z
i
)
2
(Cauchys inequality)

(x
i
y
i
)
2
+ 2(

(x
i
y
i
)(y
i
z
i
)) +

(y
i
z
i
)
2
(since

2
= [[ )
=

[(x
i
y
i
) + (y
i
z
i
)]
2
=

(x
i
z
i
)
2
= d(x, z)
2
.
Since both d(x, y)+d(y, z) and d(x, z) are non-negative, it follows that d(x, y)+d(y, z) d(x, z).
Again, this metric is geometrically natural because d(x, y) coincides with what we normally
think of as the distance between points x and y in R
n
, namely the length of the straight line
segment joining them.
(d). Other metrics on R
n
are, of course, possible. One example is
d(x, y) = max[x
i
y
i
[ : i = 1, 2, . . . , n.
Again the rst two properties are obvious. To check the triangle inequality, note that for all
i 1, 2, . . . , n,
d(x, y) +d(y, z) [x
i
y
i
[ +[y
i
z
i
[ [(x
i
y
i
) + (y
i
z
i
)[ = [x
i
z
i
[
so
d(x, y) +d(y, z) max[x
i
z
i
[ : i = 1, 2, . . . , n = d(x, z).
In R
2
, for example, the distance between x = (1, 2) and y = (7, 1) is
d(x, y) =
_
(1 7)
2
+ (2 + 1)
2
=

45 using metric (c)


and
d(x, y) = max[1 7[, [2 + 1[ = 6 using metric (d).
Note that for n = 1 the metrics in (b), (c) and (d) coincide.
Denition 3 A metric space is a pair (X, d) where X is a set and d is a metric on X.
Denition 4 Let (X, d) be a metric space, let c X and > 0. Then
B

(c) = x X : d(c, x) <


is called the open ball with centre c and radius .
Example 2 (revisited) (a). X = any set, d(x, y) = 1 if x ,= y, d(x, y) = 0 if x = y.
For any c X,
B

(c) =
_
c if 1
X if > 1.
(b). R, d(x, y) = [x y[.
B
3
(2) = x R : [x 2[ < 3 = (1, 5)
B

(c) = x R : [x c[ < = (c , c +)
4
(c). R
2
, d(x, y) =
_
(x
1
y
1
)
2
+ (x
2
y
2
)
2
.
B

((0, 0)) = (x
1
, x
2
) R
2
:
_
x
2
1
+x
2
2
< = (x
1
, x
2
) R
2
: x
2
1
+x
2
2
<
2

(c) = (x
1
, x
2
) R
2
: (x
1
c
1
)
2
+ (x
2
c
2
)
2
<
2

= disk of radius centred on (c


1
, c
2
)
(excluding boundary circle)
In R
3
, B

((c
1
, c
2
, c
3
)) is the interior of a solid spherical ball of radius centred on (c
1
, c
2
, c
3
),
hence the name ball in general.
(d). R
2
, d(x, y) = max[x
1
y
1
[, [x
2
y
2
[.
B

((c
1
, c
2
)) = (x
1
, x
2
) R
2
: [x
1
c
1
[ < and [x
2
c
2
[ <
= (x
1
, x
2
) R
2
: c
1
< x
1
< c
1
+ and c
2
< x
2
< c
2
+
= a square of side length 2 centred on (c
1
, c
2
).
Denition 5 Let (X, d) be a metric space. A subset U X is an open set if for all x U there
exists > 0 such that B

(x) U.
Example 6 (a). R, d(x, y) = [x y[.
(0, ), R0 are open sets, as are all open intervals (a, b), (a, ), (, b).
Q is not an open set. For example, 0 Q, but for all > 0, the ball B

(0) contains an
irrational number (e.g.

2/n for n Z
+
suciently large). Hence no B

(0) is contained in Q.
[1, 2) is not an open set, because for all > 0, B

(1) 1 /2 / [1, 2).


(b). X, d(x, y) = 1 if x ,= y (and 0 if x = y).
Every U X is open, since for all x U, B1
2
(x) = x U.
(c). (X, d) = any metric space.
X is an open set. So is (careful!).
Theorem 7 In a metric space (X, d), every open ball B

(c) is an open set.


Proof: Let x B

(c). Choose = d(c, x) > 0. We will show that B

(x) B

(c). If y B

(x),
then d(x, y) < . But then, by the triangle inequality,
d(c, y) d(c, x) +d(x, y) < d(c, x) + =
so y B

(c).
2.2 Continuity
Denition 8 Let (X, d) and (Y, d

) be metric spaces. A function f : X Y is continuous at


c X if for all > 0 there exists > 0 such that d(x, c) < d

(f(x), f(c)) < . Further, f is


continuous if it is continuous at every c X.
Notes:
5
If X = Y = R and d(x, y) = d

(x, y) = [x y[ this clearly reduces to the old denition of


continuity for maps R R.
The statement d(x, c) < d

(f(x), f(c)) < can be rewritten in several useful ways:


x B

(c) f(x) B

(f(c))
f(B

(c)) B

(f(c))
f
1
(B

(f(c))) B

(c).
When reading the above, note that B

(c) X, since c X, while B

(f(c)) Y since f(c) Y ,


and recall that these sets (X and Y ) and their metrics (d and d

) may be completely unrelated.


Theorem 9 Let (X, d) and (Y, d

) be metric spaces. Then f : X Y is continuous if and only if


for every open set V in Y , f
1
(V ) is an open set in X.
Proof: () Let V be an open set in Y . If f
1
(V ) = , then f
1
(V ) is open in X and were done.
If not, let c f
1
(V ). We must show there exists > 0 such that B

(c) f
1
(V ). Now f(c) V ,
by denition, and V is open, so there exists > 0 such that B

(f(c)) V . But f is continuous at


c, so there exists > 0 such that f(B

(c)) B

(f(c)) V . Hence B

(c) f
1
(V ) as required.
() Let c X. Given any > 0, B

(f(c)) is an open set in Y , by Theorem 7. Hence, by assumption,


f
1
(B

(f(c))) is an open set in X. Now clearly c f


1
(B

(f(c))), so there must exist > 0 such


that B

(c) f
1
(B

(f(c))), by the denition of openness. Hence f(B

(c)) B

(f(c)), so f is
continuous at c. But this is true for all c X, so f is continuous.
Example 10 (a). X = Y = R, d(x, y) = d

(x, y) = [x y[,
f(x) =
_
0 x Q
1 x / Q
is not continuous. For example V = (1/2, 1/2) is an open set in Y , but f
1
(V ) = Q which
is not open in X.
(b). X, Y = any sets, d = the discrete metric [d(x, y) = 1 if x ,= y, 0 if x = y], d

= any metric.
Then any function f : X Y is continuous: given any open set V Y , f
1
(V ) is some subset
of X, and hence is automatically open since, with this choice of metric on X, all subsets are
open (see Example 6). Returning to the specic example above, if we equip X = R with the
discrete metric then the function f is continuous (whatever metric we choose for Y = R).
2.3 Properties of open sets
Having proved Theorem 9, we can, if we wish, use it to give an alternative denition of continuity:
Denition 8 (*) Let (X, d) and (Y, d

) be metric spaces. Then f : X Y is continuous if for


every open set V in Y , f
1
(V ) is open in X.
An interesting point about this denition is that it uses only the open sets in (X, d), (Y, d

), not
the metrics d, d

directly (as we will see shortly, dierent metrics on the same set can dene exactly
the same collection of open sets, so the distinction between a metric and its collection of open sets is
meaningful). A topology on a set X will be a collection of subsets of X which we simply decree to
be open, by at; that is, there isnt necessarily an underlying metric, we simply pick a collection of
subsets and announce that henceforth these are the open sets in X. Denition 8(*) clearly generalizes
to this (yet more) abstract setting. However, it wont do to allow any collection of subsets to be
6
decreed to be open: we want our denition of continuity to preserve the nice properties provided
by the old one (e.g. it would be nice if all constant functions were automatically continuous). To
ensure this, we will write down a list of axioms that the collection must satisfy. These axioms will
be chosen to mimic important general properties of open sets in metric spaces, which we will shortly
establish. Before that, we need some new notation.
Denition 11 Let X and be sets. For each let A

be a subset of X. is called an
indexing set. The collection of sets A

is called an indexed family of subsets of X, denoted


A

or A

: . We can dene the union and intersection of the sets in an indexed


family:
union:
_

= x X : there exists such that x A

intersection:

= x X : for all , x A

Remark 12 If is a nite set, say = 1, 2, . . . , n, then


_

=
n
_
=1
A

= A
1
A
2
A
n

=
n

=1
A

= A
1
A
2
A
n
.
It follows immediately from the denition, that for every ,
U

, and

.
Example 13 (a). X = R, = Z
+
= 1, 2, 3, . . ., A
n
=
_

1
n
,
1
n
_
_
nZ
+
A
n
= (1, 1),

nZ
+
A
n
= 0.
(b). X = R
2
, = (0, ), A

= (x, y) R
2
: x
2
+y
2
=
2

= R
2
(0, 0),

=
Theorem 14 Let (X, d) be a metric space. Then
(a). and X are open sets in (X, d).
(b). If U and V are open sets in (X, d) then U V is an open set in (X, d).
(c). If U

is an open set in (X, d) for all , then

is an open set in (X, d).


Proof:
(a). is vacuously open since there does not exist x . Given any x X, B
37
(x) X, so X is
open.
(b). Let x U V . Then x U and x V . But U and V are open, so there exist
1
> 0 and

2
> 0 such that B
1
(x) U and B
2
(x) V . Let = min
1
,
2
. Then B

(x) is a subset of
both B
1
(x) and B
2
(x). Hence B

(x) U V , so U V is open.
7
(c). Let x

. Then there exists such that x U

. But U

is open, so there exists


> 0 such that B

(x) U

. Hence

is open.

Remark 15 Its easy to prove by induction that property (b) extends to arbitrary nite collections
of open sets. However, it does not extend to arbitrary indexed families. For example in (R, d) with
d(x, y) = [x y[, U
n
= (
1
n
,
1
n
) is open for all n Z
+
, but

nZ
+
U
n
= 0
which is not open in (R, d).
8
3 Topological spaces
3.1 The basic denition
In a manoeuvre absolutely characteristic of 20th century mathematics, we will now convert the
theorem concerning properties of open sets in metric spaces (Theorem 14) into a set of axioms for
open sets in a general set. The result gives us a denition of a space which makes no use of the
concept of distance at all.
Denition 16 Let X be a set. A topology on X is a set of subsets of X having the properties
that
(a). and X .
(b). If U and V then U V .
(c). If U

for all , then

.
A topological space is a pair (X, ) where X is a set and is a topology on X. The elements of
X are called the points of the space. The elements of are called the open sets of the space, and
are said to be open in (X, ).
Notes:
It follows by induction from (b) that the intersection of any nite collection of open sets is
open.
Once we have specied a topology on a set X, we often refer to X itself as a topological space,
rather than the pair (X, ), the topology now being understood. This slightly slapdash habit
is common throughout mathematics; for example, we often refer to the group of integers Z,
even though strictly speaking the group we mean is (Z, +) where + is the group operation
(addition in this case).
Example 17 (a). X any set, = , X is the trivial topology on X.
(b). X any set, = set of all subsets of X is the discrete topology on X.
(c). If X = a, b, a set containing two objects, then

1
= , a, b,

2
= , a, a, b,

3
= , b, a, b,

4
= , a, b, a, b
are all the topologies on X.
(d). X = a, b, c
= , b, a, b, b, c, a, b, c
is a topology on X, but
= , a, b, b, c, a, b, c
is not. Why not?
9
3.2 Metric topologies
A metric space (X, d) comes automatically equipped with a natural topology derived from the metric
d:
Denition 18 Let X be a set with a metric d. Let be the set of open sets dened by d, as in
Denition 5, that is:
= U X : for all x U there exists > 0 such that B

(x) U.
Then is a topology for X, by Theorem 14. This is called the topology obtained from (or
induced by) d. Any topology on X which is induced by a metric is called a metric topology
on X.
Example 19 (a). X = any set, d = discrete metric [d(x, y) = 1 if x ,= y, d(x, y) = 0 if x = y].
Then every subset of X is an open set (see Example 6), so the topology induced by the discrete
metric is the discrete topology, and we conclude that the discrete topology is a metric topology.
(b). R, d(x, y) = [x y[. The topology induced by this metric is called the usual or standard
topology on R. So in the usual topology, a subset U of R is an element of if
for all x U, there exists > 0 such that (x , x +) U.
(c). R
n
, d(x, y) =
_
i
(x
i
y
i
)
2
. The topology obtained from this metric is called the usual
or standard topology on R
n
.
(d). R
n
, d(x, y) = max[x
i
y
i
[ : i = 1, 2, . . . , n.
Whats the topology induced by this metric? In fact, even though this metric is dierent from
that of part (c), it turns out that the topology it induces is precisely the same.
Denition 20 Two metrics d
1
and d
2
on a set X are equivalent if they induce the same topology
on X (i.e. they dene precisely the same open sets).
Remark 21 Equivalence is, as the name suggests, clearly an equivalence relation on the set of
metrics on X. Writing d
1
d
2
to denote that d
1
and d
2
are equivalent we see that
(a). d d for every metric d (reexivity).
(b). If d
1
d
2
then d
2
d
1
(symmetry).
(c). If d
1
d
2
and d
2
d
3
then d
1
d
3
(transitivity).
Wed now like to show that the metrics of Example 19 (c) and (d) on R
n
are equivalent. To do
this, well use the following handy lemma which gives a convenient sucient (but not necessary)
condition for equivalence of d
1
and d
2
.
Lemma 22 Let d
1
and d
2
be metrics on X. If there exist constants C

C > 0 such that, for all


x, y X,
Cd
1
(x, y) d
2
(x, y) C

d
1
(x, y)
then d
1
and d
2
are equivalent.
Proof: We must show that a set U X is open in the metric space (X, d
1
) if and only if it is open
in the metric space (X, d
2
). Let us denote open balls in these spaces by B
1

(c) and B
2

(c) respectively.
Let U X be open in (X, d
1
). Then for all x U there exists > 0 such that B
1

(x) U. Let

2
= C > 0. Then,
y B
2
2
(x) d
1
(x, y) d
2
(x, y)/C <
2
/C =
y B
1

(x) U.
10
Hence, for all x U there exists
2
> 0 such that B
2
2
(x) U, so U is open in (X, d
2
).
Conversely, let U X be open in (X, d
2
). Then for all x U there exists > 0 such that
B
2

(x) U. Let
1
= /C

> 0. Then,
y B
1
1
(x) d
2
(x, y) C

d
1
(x, y) < C

1
=
y B
2

(x) U.
Hence, for all x U there exists
1
> 0 such that B
1
1
(x) U, so U is open in (X, d
1
).
Example 23 We have (so far) dened two metrics on R
n
:
d
1
(x, y) =

_
n

i=1
(x
i
y
i
)
2
, d
2
(x, y) = max[x
i
y
i
[ : i = 1, 2, . . . , n.
By denition, there exists j 1, 2, . . . , n such that
d
2
(x, y) = [x
j
y
j
[.
Hence
d
2
(x, y)
2
= (x
j
y
j
)
2

i=1
(x
i
y
i
)
2
= d
1
(x, y)
2
,
so d
2
(x, y) d
1
(x, y). Furthermore, by denition,
[x
i
y
i
[ [x
j
y
j
[ = d
2
(x, y)
for all i, so
d
1
(x, y)
2
=
n

i=1
[x
i
y
i
[
2
n[x
j
y
j
[
2
= nd
2
(x, y)
2
,
and hence d
1
(x, y)

nd
2
(x, y). Applying Lemma 22 with C = 1/

n and C

= 1 we see that d
1
and d
2
are equivalent metrics. Both induce the usual topology on R
n
.
Exercise 24 Prove that
d
3
(x, y) =
n

i=1
[x
i
y
i
[
is a metric on R
n
, then use Lemma 22 to show that its equivalent to d
1
and d
2
. Hence this metric,
too, induces the usual topology on R
n
. In the case n = 2, what do its open balls look like?
3.3 Hausdor spaces
Is every topology induced by some metric on X? We will now demonstrate that the answer is a
resounding no!
Denition 25 A topological space X is Hausdor if for all x, y X, x ,= y there exist open sets
U, V in X such that x U, y V , and U V = .
So in a Hausdor space, any pair of distinct points can be housed o into disjoint open sets.
(This pun works best if you imagine H.M. the Queen reciting it.)
Example 26 (a). Any set X with the discrete topology (every subset of X is open) is Hausdor,
since for any x, y X we can take U = x and V = y.
11
(b). X = a, b, = , a, a, b is not Hausdor since a, b is the only open set to which
b belongs, and it also contains a.
Theorem 27 Let X be a topological space whose topology is induced by a metric. Then X is
Hausdor.
Proof: Let x, y X, x ,= y, and let =
1
2
d(x, y) > 0. Let U = B

(x) and V = B

(y). Then U x
and V y, and both U and V are open in X (Theorem 7). Furthermore, U V = , for if there
exists z U V , then d(x, z) < and d(z, y) < , so
d(x, y) d(x, z) +d(z, y) < + = 2 = d(x, y) [triangle inequality]
which is a contradiction.
It immediately follows that if a topological space is not Hausdor then its topology cannot be
induced by a metric (Example 26(b), for example). It does not follow that every Hausdor topology
can be induced by a metric.
3.4 Continuity
Returning to the notion of continuity, we can now execute our plan to convert Theorem 9 into a
denition of continuity for maps between topological spaces.
Denition 28 Let X, Y be topological spaces. A function f : X Y is continuous if for every
open set V Y , f
1
(V ) is open in X.
Example 29 (a). Constant functions are always continuous.
Let X and Y have any topologies and dene f : X Y so that for all x X, f(x) = c Y
(constant). Now let V Y be open. If c V then f
1
(V ) = X which is open in X. On the
other hand, if c / V then f
1
(V ) = which is also open in X. Hence f is continuous.
(b). The identity map is always continuous.
On any topological space X dene the function Id : X X by Id(x) = x (note both domain
and codomain have the same topology). Now let V X be open in the codomain. Then
Id
1
(V ) = V which is open in the domain. Hence Id is continuous.
Its important to note that Id may fail to be continuous if we give the domain X a dierent
topology,
1
say, from the topology on the codomain X,
2
say. This is because (unless
2

1
)
V
2
does not imply that Id
1
(V ) = V
1
.
(c). If X has the discrete topology, every f : X Y is continuous.
Let V Y be open. Then f
1
(V ) is automatically open since all subsets of X are open.
Hence f is continuous.
(d). If Y has the trivial topology, every f : X Y is continuous.
Let V Y be open. Then V = Y or V = . f
1
(Y ) = X, which is open in X, and f
1
() =
which is also open in X. Hence f is continuous.
You may have fond memories of proving, in your rst Real Analysis course, that the composition
of two continuous functions is continuous. Note that this theorem becomes absolutely trivial once
we dene continuity as in Denition 28:
Proposition 30 Let f : X Y and g : Y Z be continuous. Then g f : X Z is continuous.
Proof: Let W Z be open. Then g
1
(W) Y is open, by continuity of g, and so f
1
(g
1
(W)) =
(g f)
1
(W) X is open by continuity of f.
12
3.5 Closed sets
Denition 31 Let X be a topological space. Then C X is closed if XC is open in X.
Example 32 (a). X = any set with the discrete topology. Then given any subset C X, XC
is open in X (since all subsets are open), so C is closed. Hence all subsets are both open and
closed in the discrete topology.
(b). By contrast, if X has the trivial topology, only X and are closed.
(c). X = a, b, = , a, a, b.
The closed sets are a, b, b, .
(d). R with the standard topology.
[0, 1] is closed, since R[0, 1] = (, 0) (1, ) is open. All closed intervals [a, b], [a, ),
(, b] are closed in R. (0, 1] is neither open nor closed.
(e). R
2
with the standard topology.
(x, 0) : x R is closed. (x
1
, x
2
) : x
2
0 is closed.
Closed sets have properties which are dual to those of open sets. These follow quickly from
Denition 16 and some elementary rules of set theory called De Morgans Laws:
Proposition 33 (De Morgans Laws) .
DM1: X(
_

) =

(XA

)
DM2: X(

) =
_

(XA

)
Proof: DM1:
x X(
_

) x X and x /
_

x X and x / A

(x X and x / A

)
x XA

XA

DM2: exercise.
Theorem 34 (Properties of closed sets) Let X be a topological space. Then
(a). X, are closed in X.
(b). If C, D are closed in X then C D is closed in X.
(c). If C

is closed in X for all , then

is closed in X.
13
Proof:
(a) XX = is open, so X is closed. X = X is open, so is closed.
(b) XC, XD are open, so X(C D) = (XC) (XD) by DM1, and this is open by axiom
(b) of a topological space.
(c) XC

is open for all , so X(

) =

(XC

) by DM2, and this is open by


axiom (c) of a topological space.

14
4 New topological spaces from old
4.1 Subspaces
Denition 35 Let (X, ) be a topological space and let A be a subset of X. Let

A
= U A : U .
Then
A
is the subspace topology (or relative topology) for A. (A,
A
) is called a subspace of
(X, ).
Theorem 36
A
is a topology for A.
Proof: We must check that
A
satises the axioms of a topology (Denition 16).
(a) A
A
since A = X A and X .

A
since = A and .
(b) Let U, V
A
. Then there exist U

, V

such that U = U

A and V = V

A. Hence
U V = (U

A) (V

A) = (U

) A
which is in
A
since U

.
(c) Let U


A
for all . Then for each there exists U

such that U

= U

A.
Hence
_

=
_

(U

A) =
_
_

_
A
which is in
A
since

In the subspace topology on A, the open sets are, by denition, those which arise by intersecting
open sets in X with A. We will now show that the closed sets in (A,
A
) are those which arise by
intersecting closed sets in X with A.
Theorem 37 Let A be a subspace of X. A subset B of A is closed in A if and only if there exists
some closed set C in X such that B = C A.
Proof: (): Let B A be closed in A. Then AB is open in A, and hence there exists U X,
open in X, such that
AB = U A.
I claim that
B = (XU) A.
Note that XU is closed in X, so once weve proved the claim, weve established the () part of
the theorem. To show that B = (XU) A, we must establish two things:
(i) B (XU) A and (ii) (XU) A B.
So, assume x B. Then x A (since B A). If x U, then x U A = AB, so x / B, a
contradiction. Hence x XU, and (i) is established. Now assume y (XU) A. Then y / U, so
y / U A = AB. But y A, so y B, and (ii) is established.
(): Let B = C A where C is closed in X. Then
AB = A(C A) = AC = x X : x A and x / C = (XC) A.
But XC is open in X, so AB is open in A, and hence B is closed in A.
15
Example 38 Let X = R with the usual topology.
(a). A = [0, 1) with the the subspace topology.
[0,
1
2
) is open in A since [0,
1
2
) = (1,
1
2
) A, and (1,
1
2
) is open in R.
[
1
2
, 1) is closed in A since [
1
2
, 1) = A[0,
1
2
) (or because [
1
2
, 1) = [
1
2
, 37] A and [
1
2
, 37] is closed
in R, using Theorem 37).
(b). A = [0, 1] (2, 4) with the subspace topology.
[0, 1] is open in A since [0, 1] = (1,
3
2
) A.
(2, 4) is open in A since (2, 4) = (2, 4) A.
[0, 1] is also closed in A since [0, 1] = A(2, 4).
(c). A = Z, the set of integers.
Here the subspace topology is the discrete topology for A since given any B A
B =
_
mB
m =
_
mB
[(m
1
2
, m+
1
2
) A]
which is a union of open sets in A, hence open in A.
(d). A = Q, the set of rationals with the subspace topology.
B = x A :

2 < x <

2 is both open and closed in A since B = (

2,

2) A, and
(

2,

2) is open in R, and B = [

2,

2] A and [

2,

2] is closed in R.
The Hausdor property (see Denition 25) is inherited by subspaces:
Theorem 39 Let (X, ) be a Hausdor topological space and (A,
A
) be a subspace of (X, ). Then
(A,
A
) is Hausdor.
Proof: Let x, y A, x ,= y. Since (X, ) is Hausdor, there exist U, V such that U V = ,
U x, V y. Hence U A, V A
A
, (U A) (V A) = (U V ) A = , x U A and
y V A.
Given any subset A of a set X, there is a natural map : A X called inclusion, (x) = x. If
we equip X with some topology , and A with the resulting subspace topology
A
, it turns out that
is continuous:
Proposition 40 Let (X, ), (Y,

) be topological spaces and (A,


A
) be a subspace of (X, ).
(a) The inclusion map : (A,
A
) (X, ), (x) = x, is continuous.
(b) If f : (X, ) (Y,

) is continuous, so is its restriction f[


A
: (A,
A
) (Y,

).
Proof: (a) Let U X be open. Then
1
(U) = U A which is open in A by denition. Hence is
continuous.
(b) Note that f[
A
= f , which is a composition of two continuous functions (by part (a)), hence
itself continuous by Proposition 30.
4.2 Product spaces
Given two topological spaces (X, ), (Y, ), there is a natural topology on their cartesian product
X Y = (x, y) : x X, y Y .
16
Denition 41 Let (X, ), (Y, ), be topological spaces, and let
= W X Y : (x, y) W U , V such that x U, y V and U V W.
Then is called the product topology on XY , and (XY, ) is called the topological product
of (X, ) and (Y, ).
Theorem 42 is a topology on X Y .
Proof:
(a). X Y since for all (x, y) X Y we can take U = X and V = Y .
vacuously.
(b). Let W
1
, W
2
, and let (x, y) W
1
W
2
. Then there exist U
1
, U
2
and V
1
, V
2
such
that x U
1
, y V
1
, and U
1
V
1
W
1
; and x U
2
, y V
2
, and U
2
V
2
W
2
. Let
U = U
1
U
2
and V = V
1
V
2
. Then x U and y V , and U V U
1
V
1
W
1
and U V U
2
V
2
W
2
, so U V W
1
W
2
. Hence W
1
W
2
.
(c). Let W

for all , and let (x, y)

. Then there exists such that


(x, y) W

. Hence, there exist U and V such that x U, y V and U V W

. Hence

Proposition 43 The open sets of (X Y, ) are precisely those of the form

(U

), where
U

and V

for all .
Proof: Every member of the indexed family U

: is open, hence their union is open.


Thus all sets of the form

(U

) are in . Conversely, let W . Then for each (x, y) W,


there exist U
(x,y)
and V
(x,y)
such that x U
(x,y)
, y V
(x,y)
, and U
(x,y)
V
(x,y)
W. Hence
W =

(x,y)W
(U
(x,y)
V
(x,y)
), so W can be written as

(U

), the indexing set being W


itself.
Example 44 Let (X, ) = (Y, ) = (R, usual topology). Then the product topology is a topology
on R
2
= RR. In fact, it is the usual topology on R
2
again. Recall we showed in Example 23 that
the metric topology on R
2
induced by the metric
d(x, y) = max[x
1
y
1
[, [x
2
y
2
[
is the usual topology. Assume that W R
2
is open with respect to metric d. Then for each
(x
1
, x
2
) R
2
there exists > 0 such that B

((x
1
, x
2
)) W. But for this metric B

((x
1
, x
2
)) =
(x
1
, x
1
+) (x
2
, x
2
+), and both (x
1
, x
1
+) and (x
2
, x
2
+) are open in R with
the usual topology. Hence W , the product topology.
Conversely, assume W . Then for each (x
1
, x
2
) W there exists U
1
, U
2
open in R such that
x
1
U
1
, x
2
U
2
and U
1
U
2
W. Since U
1
, U
2
are open in the standard topology on R, there
exist
1
,
2
> 0 such that (x
1

1
, x
1
+
1
) U
1
and (x
2

2
, x
2
+
2
) U
2
. Let = min
1
,
2
> 0.
Then B

((x
1
, x
2
)) (x
1

1
, x
1
+
1
) (x
2

2
, x
2
+
2
) U
1
U
2
W. Hence W is open in the
standard topology on R
2
.
There are natural maps p
X
: X Y X and p
Y
: X Y Y called projection maps,
p
X
(x, y) = x, p
Y
(x, y) = y.
Proposition 45 The projection maps p
X
: X Y X, p
X
(x, y) = x, and p
Y
: X Y Y ,
p
Y
(x, y) = y are continuous.
17
Proof: Let U X be open. Then p
1
X
(U) = U Y which is open in X Y by Proposition 43. p
Y
is handled similarly.
We can use the projection maps to give a convenient characterization of continuous maps f :
Z X Y .
Proposition 46 Let X, Y, Z be topological spaces. Then f : Z X Y is continuous if and only
if p
X
f : Z X and p
Y
f : Z Y are continuous.
Proof: () follows immediately from Propositions 30 and 45 (p
X
f, p
Y
f are compositions of
continuous maps, hence continuous).
() Assume that p
X
f : Z X and p
Y
f : Z Y are continuous, and let W be open
in X Y . Then W =

(U

) for some indexed open families U

X : and
V

Y : , by Proposition 43. Now


f
1
(
_

(U

)) =
_

f
1
(U

) (check it!)
so it suces to show that f
1
(U V ) is open in Z for each U open in X and V open in Y (since
f
1
(W) is then a union of open sets, hence open). Now
f
1
(U V ) = z Z : f(z) U V
= z Z : (p
X
f)(z) U and (p
Y
f)(z) V
= (p
X
f)
1
(U) (p
Y
f)
1
(V ).
But p
X
f and p
Y
f are continuous by assumption, so both (p
X
f)
1
(U) and (p
Y
f)
1
(U) are
open in Z. Hence their intersection is open in Z.
Note: This generalizes the (hopefully) familiar fact that a function f : R R
2
, f(x) = (f
1
(x), f
2
(x)),
is continuous if and only if its component functions f
1
, f
2
are continuous.
The Hausdor property is preserved by the topological product.
Proposition 47 Let X and Y be topological spaces. Their topological product is Hausdor if and
only if both X and Y are Hausdor.
Proof: Exercise.
4.3 Quotient spaces

Given a topological space (X, ), a set Y (no topology as yet), and a function f : X Y , we can
use f to transfer the topology on X to Y . The idea is to dene

f
= U Y : f
1
(U) .
So
f
is the collection of subsets of Y whose preimages under f are open in (X, ). I claim that
f
is a topology on Y :
(a) f
1
(Y ) = X , so Y
f
. Similarly f
1
() = , so
f
.
(b) Let U, V
f
. Then f
1
(U), f
1
(V ) . Hence f
1
(U V ) = f
1
(U) f
1
(V ) (since
is a topology). Hence U V
f
.
18
(c) Let U


f
for all . Then f
1
(U

) for all . Hence


f
1
(
_

) =
_

f
1
(U

)
since is a topology. Hence


f
.
Note that f : (X, ) (Y,
f
) is automatically continuous.
Exercise: Let f : X Y be a constant function. Show that
f
is the discrete topology on Y ,
whatever the choice of topology on X.
This construction is rather trivial (and doesnt, as far as I know, have a name). But it is a useful trick
for equipping one type of set with a natural topology. Recall that given an equivalence relation
on a set X, we dene the equivalence class of x X to be the set
[x] = y X : y x.
The collection of equivalence classes is itself a set, called the quotient of X by ,
X/ = [x] X : x X.
Note there is a natural map X X/ called projection
p : X X/ , p(x) = [x].
It just maps each x X to its equivalence class. If X has a topology, , we can now use the above
construction to equip the quotient set X/ with a natural topology inherited from , called the
quotient topology:

p
= U X/ : p
1
(U) .
In many examples, this is the most sensible way to give the quotient set a topology.
Example: Lets say we want to equip the set of one-dimensional subspaces of R
n
with a natural
topology. The best way to do this is to let X = R
n
(0, 0, . . . , 0) and dene an equivalence relation
on X by x y if there exists R, ,= 0, such that x = y. For each x X, [x] is the straight line
in R
n
through the origin containing the point x. So X/ can be identied with the set of original
interest. (This set is called RP
n1
, the real projective (n 1)-space.) It inherits a natural topology
from the topology on X, itself the subspace topology given by the inclusion X R
n
, where R
n
has
the usual topology.
One slightly unpleasant thing about the quotient topology is that it doesnt preserve the Hausdor
property in general. That is, even if (X, ) is Hausdor, (X/ ,
p
) may fail to be Hausdor.
Example: On X = R, usual topology, let x y if there exists n Z such that x = 2
n
y. This
is an equivalence relation (check it!). The quotient space R/ is not Hausdor. To see this, let
U R/ be any open set containing [0]. Then p
1
(U) R is open, and contains 0. Hence there
exists > 0 such that (, ) p
1
(U). Note that [0] = 0. Now let x be any real number
other than 0. Then there exists some n Z such that 2
n
x (, ). Hence 2
n
x p
1
(U), so
p(2
n
x) = [2
n
x] = [x] U. Thus any open set containing [0] contains all other points [x] R/ . So
R/ certainly is not Hausdor ([0] cant be housed o from any other point [x]).
19
5 Topological invariants
5.1 Homeomorphisms
Denition 48 Let X, Y be topological spaces. A function f : X Y is a homeomorphism if
(a). f is bijective

(i.e. f is both one-to-one and onto), and


(b). both f and f
1
are continuous.
X is homeomorphic to Y if there exists a homeomorphism f : X Y .
Example 49 (a). X = R with usual topology, Y = (0, ) with subspace topology (from inclusion
(0, ) R).
f : R (0, ), f(x) = e
x
.
f is injective (one-to-one), surjective (onto) and continuous. Its inverse function is
f
1
: (0, ) R, f
1
(y) = log y
which is also continuous. Hence f is a homeomorphism, and R and (0, ) are homeomorphic.
(b). X = a, b, c,
1
= , X, a, b, a, b,
2
= , X, b, c, b, c.
f : (X,
1
) (X,
2
), f(a) = b, f(b) = c, f(c) = a
is a homeomorphism (check it!). Is it the only homeomorphism (X,
1
) (X,
2
)?
(c).
C = (x
1
, x
2
) R
2
: x
2
1
+x
2
2
= 1 circle
S = (x
1
, x
2
) R
2
: max [x
1
[, [x
2
[ = 1 square
_
both with subspace topology from (R
2
, usual).
Dene functions f, g by radial projection:
f : C S, f(x
1
, x
2
) =
_
x
1
max[x
1
[, [x
2
[
,
x
2
max[x
1
[, [x
2
[
_
g : S C, g(y
1
, y
2
) =
_
y
1
_
y
2
1
+y
2
2
,
y
2
_
y
2
1
+y
2
2
_
.
Its not hard to check that g f = Id
C
and f g = Id
S
. It follows that f is injective:
f(x) = f(x

) g(f(x)) = g(f(x

)) Id
C
(x) = Id
C
(x

) x = x

;
and that f is surjective: let y S; then y = Id
S
(y) = f(g(y)), so there exists x C such that
f(x) = y (namely x = g(y)). So f satises property (a) of a homeomorphism, and g = f
1
.
Clearly f and f
1
are continuous. Actually, to prove this is not entirely trivial. One could
dene
F : R
2
(0, 0) R
2
(0, 0), F(x
1
, x
2
) =
_
x
1
max[x
1
[, [x
2
[
,
x
2
max[x
1
[, [x
2
[
_
G : R
2
(0, 0) R
2
(0, 0), G(y
1
, y
2
) =
_
y
1
_
y
2
1
+y
2
2
,
y
2
_
y
2
1
+y
2
2
_
,

It follows that the inverse function f


1
: Y X exists.
20
and note that f = F[
C
and g = G[
S
. Now using standard theorems from Real Analysis,
one can argue that F, G are continuous, and hence their restrictions f, g are continuous by
Proposition 40.
(d). Lets dene f : [0, 2) C by
f(x) = (cos x, sin x).
Is this a homeomorphism, when [0, 2) R is given the subspace topology? It is injective,
surjective and continuous. But f
1
is not continuous. To see this, note that [0, ) is open in
[0, 2), but f([0, )) = (f
1
)
1
([0, )) is not open in C.
If two topological spaces are homeomorphic, they are, as abstract topological spaces, entirely
equivalent. That is, if we use f to identify each x X with f(x) Y , then X and Y are
essentially the same space, but with the points labelled dierently. Note that homeomorphism
(i.e. being homeomorphic) is an equivalence relation on the collection of topological spaces. This
follows from the rst 3 parts of the following proposition.
Proposition 50 .
(a). For any topological space X, Id
X
: X X is a homeomorphism.
(b). If f : X Y is a homeomorphism, so is f
1
: Y X.
(c). If f : X Y and g : Y Z are homeomorphisms, so is g f : X Z.
(d). If f : X Y is a homeomorphism and A is a subspace of X then f[
A
: A f(A) is a
homeomorphism.
(a),(b),(c) are trivial, and are left as an exercise. To prove (d), we will need the following slightly
tricky lemma:
Lemma 51 Let g : X Y be injective and A, B X. Then g(A B) = g(A) g(B).
Proof: If y g(A) g(B) then there exists x A such that g(x) = y and there exists x

B such
that g(x

) = y. But g is injective, so x = x

. Hence, there exists x A B such that g(x) = y, so


y g(A B).
Conversely, if y g(A B) then there exists x A B such that g(x) = y. Hence there exists
x A such that g(x) = y, and there exists x B such that g(x) = y. Hence y g(A) g(B).
Note that injectivity of g is crucial for the inclusion g(A) g(B) g(A B).
Proof of Proposition 50(d): f[
A
(x) = f[
A
(x

) f(x) = f(x

) x = x

by injectivity of
f. Hence f[
A
is injective. Also f[
A
: A f(A) is surjective by denition. So f[
A
is bijective. Now
f
A
is continuous by continuity of f (Proposition 40), so it remains to show that f[
1
A
: f(A) A is
continuous. Let U A be open. Then there exists U

open in X such that U = U

A. Hence
(f[
1
A
)
1
(U) = f[
A
(U) = f(A) = f(U

A) = f(U

) f(A),
by Lemma 51. Now f(U

) is open in Y by continuity of f
1
, so f(U) is open in f(A), by the
denition of subspace topology.
21
Denition 52 A topological invariant is any property of topological spaces which, if held by a
space X, is held by all spaces homeomorphic to X.
Proposition 53 The Hausdor property is a topological invariant. That is, if X is Hausdor and
Y is homeomorphic to X, then Y is Hausdor.
Proof: Let X be a Hausdor topological space, and f : X Y be a homeomorphism. Let
y
1
, y
2
Y , y
1
,= y
2
. Since f is surjective, there exist x
1
, x
2
X such that f(x
1
) = y
1
and
f(x
2
) = y
2
. Clearly x
1
,= x
2
. Since X is Hausdor, there exist open sets U, V in X such that
x
1
U, x
2
V and U V = . But now f(U) and f(V ) are open in Y (by continuity of f
1
),
y
1
f(U), y
2
f(V ) and f(U) f(V ) = f(U V ) = f() = by Lemma 51. Hence Y is Hausdor.

5.2 Connectedness
Denition 54 A topological space X is disconnected if there exist two disjoint, nonempty open
subsets U, V of X such that U V = X. (Disjoint means U V = .) The pair of sets U, V is called
an open partition of X. A topological space X is connected if it is not disconnected.
Example 55 (a). If X has more than one point and is given the discrete topology, it is discon-
nected: given any x X,
x, Xx
is an open partition of X.
(b). X with the trivial topology = , X is connected, as there is clearly no open partition.
Denition 56 A subset A of a topological space (X, ) is a connected subset of X if (A,
A
) is
connected, where
A
is the subspace topology.
Example 57 (a). A = [0, 1] (2, 4) is a disconnected subset of (R, usual topology) since [0, 1],
(2, 4) form an open partition of A (recall [0, 1] and (2, 4) are open subsets of A in the subspace
topology, see Example 38).
(b). A = (x
1
, x
2
) R
2
: x
1
x
2
0(0, 0) is a disconnected subset of (R
2
, usual topology). The
sets
U = A (x
1
, x
2
) R
2
: x
1
> x
2

V = A (x
1
, x
2
) R
2
: x
1
< x
2

are open in the subspace topology, disjoint and U V = A. Hence they form an open partition
of A.
(c). A = Q is a disconnected subset of (R, usual topology), since
(,

2) Q, (

2, ) Q
form an open partition of Q.
22
In general its easy to demonstrate that a disconnected space is diconnected: one simply has to
exhibit an open partition. Its much harder to show that a connected space is connected: one must
prove that no open partition exists, and this may be quite dicult. For example, its intuitively
clear (I hope) that R, with the usual topology, is connected, but its far from trivial to actually prove
it. Our next goal is to prove something slightly more general than this, that a subset of (R, usual) is
connected if and only if it is an interval. First well give a slick denition of interval. This will allow
us to consider all the dierent types of interval (open, closed, half closed, bounded etc.) in one go.
Denition 58 A nonempty subset J of R is an interval if for all x, y J, if x < c < y then c J.
You should check that this denition includes all the types of interval youre familar with:
(a, b), [a, b), (a, b] , [a, b], (, b), (, b], (a, ), [a, ), (, ).
It also includes every singleton set a = [a, a], which you have previously probably not considered
to be an interval.
Theorem 59 A subset J of R is connected

if and only if J is an interval.


Proof: () Let J be connected. Let x, y J such that x < y (if J contains only one element,
then J = [a, a] which is an interval, so were done). For each c (x, y) let U = (, c) J and
V = (c, ) J. U, V are disjoint open sets in J, x U and y V , so both are nonempty. If
c / J then U, V form an open partition of J, a contradiction. Hence c J. Since this holds for all
c (x, y), for all x, y J, J is an interval.
() Let J be an interval and suppose, towards a contradiction, that J is disconnected. Then
there exist open sets U, V in R such that
(U J) (V J) = J, (U J) (V J) =
and both U J, V J are nonempty. Let a U J and b V J. Then a ,= b and, without
loss of generality, we may assume that a < b. Since J is an interval, it contains a, b and all numbers
between, that is, [a, b] J. Let
U

= U [a, b] = U J [a, b], V

= V [a, b] = V J [a, b].


Both U

and V

are open in [a, b], a U

and b V

, so there exist ,

> 0 such that


[a, b] (a , a +) = [a, a +) U

, [a, b] (b

, b +

) = (b

, b] V

.
Now U

is bounded above (by b) so there exists c = sup U

, and c a + > a. Since U

= ,
every x > b

is not in U

, so c b

< b. Hence c (a, b).


If c U

, then since U

is open there exists > 0 such that (c , c + ) U

. But then
c +/2 U

so c is not an upper bound on U

. Hence c / U

.
If c V

, then since V

is open there exists > 0 such that (c , c +) V

. Since c = sup U

,
c is not an upper bound on U

so there exists x U

with c < x c. Then x is in both U

and V

which is impossible (theyre disjoint). Hence c / V

.
Hence c / U

= [a, b], a contradiction.


We want to show that connectedness is a topological invariant (if X is connected then every Y
homeomorphic to X is connected). This follows easily from the following theorem.
Theorem 60 Let f : X Y be continuous and X be connected. Then f(X) is connected.

From now on, R, R


n
will always be assumed to have the usual topology, unless otherwise stated.
23
Proof: Assume, towards a contradiction, that f(X) is disconnected and let U, V be open sets in Y
such that U f(X), V f(X) is an open partition of f(X). Since f is continuous,
U

= f
1
(U) = f
1
(U f(X)) and V

= f
1
(V ) = f
1
(V f(X))
are open in X. Neither Uf(X) nor V f(X) is empty, so U

and V

are also nonempty. Furthermore,


U

= f
1
(U) f
1
(V ) = f
1
(U V ) = X
since f(X) U V , and,
U

= f
1
(U f(X)) f
1
(V f(X)) = f
1
((U f(X)) (V f(X))) = f
1
() = .
Hence U

, V

is an open partition of X, a contradiction (X is connected).


Corollary 61 Let X be connected and Y be homeomorphic to X. Then Y is connected.
Proof: By assumption, there exists a homeomorphism f : X Y . Since f is continuous and X is
connected, f(X) is connected, by Theorem 60. But f is surjective, so f(X) = Y .
Another quick (and very famous) corollary of Theorem 60:
Corollary 62 (The Intermediate Value Theorem) Let f : [a, b] R be continuous and y be
some number between f(a) and f(b). Then there exists c [a, b] such that f(c) = y.
Proof: Since f is continuous and [a, b] is connected (Theorem 59), f([a, b]) is connected (Theorem
60) and hence is an interval (Theorem 59). f(a), f(b) f([a, b]), so any y between f(a) and f(b)
also lies in f([a, b]) (denition of an interval). Hence there exists c [a, b] such that f(c) = y.
In fact Theorem 60 is a very versatile tool, as the next example illustrates.
Example 63 (a). The unit circle is connected.
f : R R
2
such that f(x) = (cos x, sin x) is continuous, and R is connected (Theorem 59).
Hence
f(R) = (x
1
, x
2
) R
2
: x
2
1
+x
2
2
= 1 = C,
the unit circle, is connected.
(b). [a, b] is not homeomorphic to (c, d).
Suppose there does exist a homeomorphism f : [a, b] (c, d). Then
f[
(a,b]
(a, b] (c, d)f(a)
is also a homeomorphism, by Theorem 50(d). But (a, b] is connected, while (c, d)f(a) =
(c, f(a)) (f(a), d) is not, a contradiction.
(c). Let f : [a, b] [c, d] be a homeomorphism. Then f(a) = c and f(b) = d, or f(a) = d and
f(b) = c.
If f(a) (c, d), then we have a homeomorphism
f[
(a,b]
: (a, b] [c, f(a)) (f(a), d]
from a connected to a disconnected set, which is impossible. Hence f(a) = c or f(a) = d.
Similarly f(b) = c or f(b) = d. But f is injective, so f(a) ,= f(b), whence the result follows.
24
So far weve proved that all intervals are connected, and the unit circle in R
2
is connected. What
about higher dimensional Euclidean spaces? Is R
n
connected, n 2? What about the unit n-sphere,
S
n
= (x
1
, x
2
, . . . , x
n+1
) R
n+1
: x
2
1
+x
2
2
+ +x
2
n+1
= 1 ?
Again, its intuitively clear that these are connected spaces. But how do we prove it? The following
rather technical looking theorem will help.
Theorem 64 Let C

: be a family of connected subsets of X such that


_

= X and

,= .
Then X is connected.
Proof: Assume, towards a contradiction, that X is disconnected, and let U, V be an open partition
of X. Let x

. Then x is in precisely one of U, V , and without loss of generality, we


may assume x U. For each , U C

, V C

are open in C

, and (U C

) (V C

) =
(U V ) C

= , since U, V is a partition. Now x U C

, so V C

= , or else U C

, V C

is an open partition of C

(contradicting the assumption that each C

is connected). So V C

=
for all , and hence V X = . Therefore V = , a contradiction (U, V was assumed to be an
open partition of X).
More informally, Theorem 64 says that if you can split a space up into connected pieces all of which
meet in at least one point, the space is connected.
Example 65 (a). R
2
is connected.
R
2
is the union of all straight lines through
(0, 0). Each line is homeomorphic to R,
hence is connected. The intersection of all
such lines is nonempty (they all contain
(0, 0)). Hence by Theorem 64, R
2
is con-
nected. A similar argument works for R
n
,
n 3.
(b). The 2-sphere S
2
is connected.
Recall that
S
2
= (x
1
, x
2
, x
3
) R
3
: x
2
1
+x
2
2
+x
2
3
= 1.
we can split this into a union of all the unit
circles through (0, 0, 1) and (0, 0, 1):
C

= (sin t cos , sin t sin, cos t) R


3
: t R,
R. Each C

is connected by Theorem
60 since it is the image of the (connected)
set R under the continuous map
f

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


The intersection of all these arcs is
nonempty,

R
C

= (0, 0, 1), (0, 0, 1)


so S
2
is connected by Theorem 64. Can
this argument be generalized to deal with
S
n
, n 3? .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
25
(c) R is not homeomorphic to R
2
.
Assume there is a homeomorphism f : R R
2
. Then its restriction to R0 is a homeomor-
phism
f : (, 0) (0, ) R
2
f(0).
But (, 0) (0, ) is disconnected, while R
2
f(0) is connected, a contradiction.
To see that R
2
f(0) is connected, let C be
the unit circle with centre f(0),
C = f(0) + (cos t, sin t) : t R
and for each R dene
R

= f(0) + (t cos , t sin ) : t (0, ).


Both C and R

are the images of continuous


maps from intervals (R and (0, )) to R
2
, and
hence are connected by Theorem 60. They
have non-empty intersection:
C R

= f(0) + (cos , sin )


so C

= C R

is connected by Theo-
rem 64. Now

R
C

= R
2
f(0) and

R
C

= C ,= . Hence R
2
f(0) is con-
nected by Theorem 64.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Theorem 66 X Y is connected if and only if both X and Y are connected.
Proof: () Assume X Y is connected. The projection maps p
X
: X Y X, p
Y
: X Y Y
are surjective and continuous (Proposition 45), so p
X
(XY ) = X and p
Y
(XY ) = Y are connected
by Theorem 60.
() The idea is to split X Y into a union
of intersecting connected pieces. Assume that X
and Y are connected. For each y Y , X is home-
omorphic to X y X Y . For example
f : X X y, f(x) = (x, y)
is a homeomorphism. Similarly, for each x X,
Y is homeomorphic to xY (g(y) = (x, y) is a
homeomorphism). Hence for each (x, y) XY ,
Xy and xY are both connected. Choose
and x y
0
Y , and for each x X let
C
x
= (X y
0
) (x Y ).
.
.
.
.
.
.
.
.
C
x
is connected, by Theorem 64, since it is the union of two connected sets with nonempty in-
tersection, (x, y
0
). Now (x, y) X Y (x, y) C
x
, so

xX
C
x
= X Y , and

xX
C
x
= X y
0
,= , so X Y is connected by Theorem 64.
Note: It follows immediately from Theorem 66 that R
2
is connected. We can show that R
3
is
homeomorphic to R
2
R, whence it follows that R
3
is also connected. An obvious inductive argument
then shows that R
n
is connected for all n Z
+
.
26
Denition 67 A component of a topological space X is a maximal connected subset of X. That
is C X is a component if C is connected and for all subsets D of X with C D and C ,= D,
D is not connected. The space is said to be totally disconnected if all the components of X
are singleton sets (sets with only one element), that is, if all sets with two or more elements are
disconnected.
Example 68 (a). The components of A = [0, 1] (2, 4), as a subspace of R, are [0, 1] and (2, 4).
This space is not totally disconnected.
(b). If X has the discrete topology then x is a component of X for all x X. To see this, let D
be any subset of X containing x and at least one other element. Then x, Dx is an open
partition of D, so D is disconnected. Hence X is totally disconnected.
(c). Consider Q with the subspace topology inherited from R. Again x is a component of Q for
all x Q. Again, let D Q contain x and at least one other number y. Then between the
distinct rational numbers x, y there is an irrational number , and
D (, ), D (, )
is an open partition of D. So D is disconnected. Hence Q is totally disconnected.
Its important to realize that Q does not have the discrete topology here, since not all subsets
of Q are open in the subspace topology. In fact every open subset of Q contains innitely
many elements (so nite sets, for example, are not open in this topology). To see this, note
that a general open set takes the form U Q where U is open in R. For each x U Q there
exists > 0 such that (x , x + ) U, and there are innitely many rational numbers in
this interval.
5.3 Path connectedness
There is a more geometrically intuitive notion of connectedness for topological spaces.
Denition 69 Let X be a topological space and x, y X. A path in X from x to y is a continuous
map : [0, 1] X such that (0) = x and (1) = y. The space X is path connected if for all
x, y X there is a path in X from x to y.
Theorem 70 If X is path connected then X is connected.
Proof: Let X be path connected and assume, towards a contradiction, that X is not connected.
Then there is an open partition U, V of X. Let x U and y V . Since X is path connected, there
is a path : [0, 1] X from x to y. Consider
1
(U),
1
(V ) [0, 1]. Since is continuous and
U, V are open in X, both
1
(U),
1
(V ) are open in [0, 1]. Neither is empty (0
1
(U), and
1
1
(V )). They are disjoint,

1
(U)
1
(V ) =
1
(U V ) =
1
() = ,
and their union is [0, 1],

1
(U)
1
(V ) =
1
(U V ) =
1
(X) = [0, 1].
Hence
1
(U),
1
(V ) forms an open partition of the set [0, 1]. But [0, 1] is connected, by Theorem
59, a contradiction.
WARNING! The converse of this theorem is false! There exist connected topological spaces which
are not path connected.
27
Counterexample 71 Let A = (x, sin(1/x)) R
2
: x > 0 (the graph of the function f(x) =
sin(1/x) for x > 0, and consider the space X = A (0, 0), given the subspace topology from R
2
.
X is connected (note that (0, 0), A is not an open partition of X because (0, 0 is not open).
But x is not path connected: there is no path from (0, 0) to any point in A.
Theorem 72 Let f : X Y be continuous and X be path connected. Then f(X) is path connected.
Proof: Let f(x), f(x

) f(X). Since X is path connected, there is a path : [0, 1] X from x to


x

. Then f : [0, 1] Y is a path in f(X) from f(x) to f(x

): being a composition of continuous


functions, it is continuous, and f (0) = f((0)) = f(x) while f (1) = f((1)) = f(x

).
Corollary 73 Let X be path connected and Y be homeomorphic to X. Then Y is path connected.
Proof: Let f : X Y be a homeomorphism. Then f is continuous, so f(X) Y is path connected.
But f is surjective, so f(X) = Y .
Theorem 74 X Y is path connected if and only if both X and Y are path connected.
Proof: () Assume X Y is path connected. Then the projection maps p
X
: X Y X and
p
Y
: X Y Y are continuous and surjective, so p
X
(X Y ) = X and p
Y
(X Y ) = Y are path
connected by Theorem 72.
() Assume both X and Y are path connected. Then given any pair (x, y), (x

, y

) X Y ,
there exist continuous maps : [0, 1] X and : [0, 1] Y such that (0) = x, (1) = x

,
(0) = y and )1) = y

. Let : [0, 1] Y such that (t) = ((t), (t)). Then is continuous


(Proposition 46), and (0) = ((0), (0)) = (x, y), (1) = ((1), (1)) = (x

, y

).
One can often prove path connectedness of specic spaces without having to use big theorems.
Furthermore, checking path connectedness is a constructive process: given any pair x, y X we
must construct a path between them. Since path connectedness implies connectedness, this often
provides a sneaky way of proving that a given space is connected. Checking connectedness directly is,
by contrast, quite logically subtle (we have to rule out the existence of an open partition). Compare
the following almost trivial argument showing that R
2
is path connected with the chain of reasoning
we used to prove it was connected.
Example 75 R
2
is path connected (and hence connected).
Given any (x
1
, x
2
), (y
1
, y
2
) R
2
, let : [0, 1] R
2
such that
(t) = (x
1
+t(y
1
x
1
), x
2
+t(y
2
x
2
)).
This is continuous and (0) = (x
1
, x
2
), (1) = (y
1
, y
2
), so is a path from (x
1
, x
2
) to (y
1
, y
2
).
Of course, the hard work hasnt really just disappeared. The proof that path connectedness implies
connectedness relies on the fact that [0, 1] is connected (Theorem 59) whose proof was itself hard
work.
We can streamline the process of checking that a space X is path connected using the following
proposition, which says that, rather than checking that every pair of points in X is connected to
28
each other by a path, its enough to show that all points are connected by paths to a single base
point.
Proposition 76 X is path connected if and only if there exists x
0
X such that for all y X
there is a path in X from x
0
to y.
Proof: () is clear. To prove (), choose any x, y X. Then there exist paths from x
0
to x
and from x
0
to y. Let : [0, 1] X such that (t) = (1 t). This is a path from x to x
0
(check
it!). Now dene
: [0, 1] X, (t) =
_
(2t) t [0,
1
2
]
(2t 1) t [
1
2
, 1].
This is continuous by the Glue Lemma (see Problem Set 2), and has (0) = x, (1) = y. Hence X
is path connected.
Example 77 R
n
is path connected.
Given any x R
n
, let : [0, 1] R
n
such that (t) = tx = (tx
1
, tx
2
, . . . , tx
n
). This is a path from
0 = (0, . . . , 0) to x.
29
6 Compactness
Compactness is yet another topological invariant, but its so important we will give it its own section.
Denition 78 let X be a topological space and A be a subset of X. An open cover of A is a
family C = U

: of open sets in X such that


_

A.
The cover is nite if the indexing set is nite. Let be a subset of . Then C

= U

: C
is a subcover of C if it is an open cover, that is,
_

A.
A is a compact subset of X if every open cover of A has a nite subcover. In particular, X is
compact if it is a compact subset of itself.
Note: It is important to realize that A compact means that every open cover of A has a nite
subcover. It does not simply mean that A has a nite open cover. Indeed, every subset of a
topological space X has (at least) one nite open cover, namely C = X (recall X is open by
axiom (a) of a topology).
Example 79 (a). R is not compact. For example C = (n, n) : n Z
+
is an open cover which
has no nite subcover.
(b). (0, 3) is not a compact subset of R. For example C = (
1
n
, 3
1
n
) : n Z
+
is an open cover
which has no nite subcover.
(c). R
2
is not compact. For example C = U
n
: n Z
+
where
U
n
= (x
1
, x
2
) : x
2
1
+x
2
2
< n
2

is an open cover which has no nite subcover.


(d). If A is a nite subset of X, it is compact, whatever topology we put on X. C = U

:
be an open cover of A. Then for each x A there exists
x
such that x U
x
. Hence
C

= U
x
: x A is a subcover of C, and is nite since A is nite.
(e). If we give X the trivial topology = , X, every subset of X is compact, since and X
are the only open covers, and these are both nite.
(f). X with the discrete topology is compact if and only if X is nite. The if part weve seen
already. The only if part follows once we recognize that
C = x : x X
is an open cover which has no subcovers (other than itself). If X is innite, C is innite, so
has no nite subcover.
Proposition 80 A subset A of a topological space (X, ) is compact if and only if (A,
A
) is a
compact topological space, where
A
is the subspace topology.
Proof: () Assume A is a compact subset of X, and let C = U

: be an open cover of
(A,
A
). Then for each there exists V

such that U

= V

A. Then C

= V

:
is an open cover of the set A. But A is a compact subset, so there exists , nite, such that
C

= V

: is a cover of the set A. Hence C

= V

A : = U

: is a
nite subcover of C. Hence (A,
A
) is compact.
() Exercise.
30
Example 81 Its not hard to show that every type of interval other than [a, b], b a, is not compact
as a subset or R.
Interval Open cover with no nite subcover
(a, b] (a +
1
n
, b + 1) : n Z
+

(a, b) (a +
1
n
, b) : n Z
+

[a, b) (a 1, b +
1
n
) : n Z
+

(a, ) (a, n) : n Z
+

[a, ) (a 37, n) : n Z
+

(, b) (n, b) : n Z
+

(, b] (n, b + 37) : n Z
+

(, ) (n, n) : n Z
+

Its rather more dicult to prove that [a, b], a < b, is compact (note [a, a] is nite, hence trivially
compact). This is (a special case of) the celebrated Heine-Borel Theorem.
Theorem 82 (Heine-Borel, R) Every interval [a, b], where a, b R and a < b, is compact.
Proof: Let C be an open cover of [a, b] and dene the set B as follows
B = x [a, b] : there is a nite subcover C

C of [a, x].
Since C covers [a, b] and a [a, b], there is some U C containing a. But U is open, so there exists
> 0 such that (a , a + ) U. Hence, for each x [a, a + ) there exists a nite subcover
C

C of [a, x], namely C

= U. Hence [a, a + ) B, so B is nonempty. Also B [a, b], so it


is bounded above (by b). Hence there exists s = sup B. Clearly a < s b. We aim to show that
(i) s B, and (ii) s = b.
It will follow that b B, and hence that the interval [a, b] itself has a nite subcover in C. This
holds for all open covers C, so the result follows.
(i) Since a < s b, s [a, b] so there exists V C such that s V . But V is open, so
there exists > 0 such that (s , s + ) V . Now s is not an upper bound on B, so there
exists r (s , s) such that r B. Hence there is a nite subcover C

C of [0, r]. But then


C

V C is a nite subcover of [0, s]. Hence s B.


(ii) Assume that s < b. Then there exists t (s, s + ) with s < t < b. But then C

V C
is a nite subcover of [a, t], so t B. But s is an upper bound on B, a contradiction.
Its tempting to think of closed bounded intervals, like [0, 1], as archetypal compact sets, but this
can be misleading. For example [a, b] [c, d] is (if nonempty) another closed bounded interval, hence
compact. But in a general topological space theres no reason why the intersection of two compact
subsets should be compact. Heres a counterexample.
Counterexample 83 Let X be an innite set and let a, b X, a ,= b. Let X have topology
where U if and only if U = X, U = or U Xa, b (U contains neither a nor b). You should
check that is a topology.
Let A = Xb and B = Xa. I claim A is a compact subset of X. For if U

: is a
family of open sets in X with A

, then there exists such that a U

. But then
U

= X (the only open set containing A is X). Hence U

is a nite subcover of A. Similarly B


is a compact subset of X.
Now consider AB = Xa, b. Every subset of AB is open in X, hence the subspace topology
on A B is the discrete topology. Since X is innite, so is A B. Hence A B is not compact (for
example x : x X, x ,= a, x ,= b is an open cover with no nite subcover).
We will now prove that compactness, like connectedness and path connectedness is preserved by
continuous maps.
31
Theorem 84 Let X be compact and f : X Y be continuous. Then f(X) is a compact subset of
Y .
Proof: Let U

: be a family of open sets in Y with f(X)


. Then
X = f
1
_

_
=
_

f
1
(U

).
Since f is continuous, f
1
(U

) is open in X. Hence
f
1
(U

) :
is an open cover of X. But X is compact, so there exists a nite subcover
f
1
(U
i
) : i = 1, 2, . . . , n
of X. Hence
f(X) = f
_
n
i=1
f
1
(U
i
)
_
=
n
_
i=1
f(f
1
(U
i
))
n
_
i=1
U
i
.
So U
i
: i = 1, 2, . . . , n is a nite subcover of the original cover of f(X). Since this holds for all
open covers of f(X), it follows that f(X) is compact.
It follows immediately that compactness is a topological invariant:
Corollary 85 Let X be compact and Y be homeomorphic to X. Then Y is compact.
Proof: By assumption there is a homeomorphism f : X Y . Since f is continuous, f(X) Y is
compact (Theorem 84). But f is surjective, so f(X) = Y .
As with connnectedness, we can use Theorem 84 to show that some spaces are compact by
demonstrating that they are the image of a compact set under a continuous map.
Example 86 The unit circle C R
2
is compact.
Note that f : [0, 2] R
2
where f(x) = (cos x, sin x) is continuous, and [0, 2] is compact. Hence
f([0, 2]) = C is compact.
We can show that X Y is compact if and only if both X and Y are compact. The proof of
the only if part is essentially identical to the corresponding proof for connectedness (and path
connectedness). The proof of the if part is considerably more subtle. This is a special case of a
famous theorem due to Tychono.
Theorem 87 (Tychono) X Y is compact if and only if both X and Y are compact.
Proof: () Assume X Y is compact. The projection functions p
X
: X Y X and p
Y
:
X Y Y are continuous and surjective. Hence X and Y are compact by Theorem 84.
() Let W

: be an open cover of X Y . For each (x, y) X Y there exists


such that (x, y) W

. Now W

is open, so there exist U


(x,y)
open in X, containing x, and V
(x,y)
open in Y , containing y, such that U
(x,y)
V
(x,y)
W

. Consider the open cover


C = U
(x,y)
V
(x,y)
: (x, y) X Y .
We will show that this cover has a nite subcover
C

= U
(x
k
,y
k
)
V
(x
k
,y
k
)
: k = 1, 2, . . . , n.
32
Since for each k there is some
k
such that U
(x
k
,y
k
)
V
(x
k
,y
k
)
W

k
, it will follow that our
original open cover W

: has a nite subcover W

k
: k = 1, 2, . . . , n, which was to be
proved.
For b Y , consider the collection U
(x,b)
: x X. This is an open cover of X, which is
compact, so there is a nite subcover U
(x1,b)
, . . . , U
(xn,b)
. Dene
V
b
=
n

i=1
V
(xi,b)
.
V
b
is open in Y (its a nite intersection of open sets), and b V
b
. Also,
X V
b

n
_
i=1
U
(xi,b)
V
(xi,b)
() (check it!).
Now V
y
: y Y is an open cover of Y , which is compact, so there is a nite subcover V
y1
, . . . , V
ym

of Y . Then
X Y =
m
_
j=1
X V
yj
.
But then for each j = 1, 2, . . . , m, there is some nite collection of points x
ij
X, i = 1, 2, . . . , n
j
,
such that
X V
yj

nj
_
i=1
U
(xij,yj)
V
yj
by (). Hence
X Y
m
_
j=1
nj
_
i=1
U
(xij,yj)
V
(xij,yj)
,
so X Y is covered by a nite subcollection from C, which completes the proof.
Note: by induction, X
1
X
2
X
n
is compact if and only if every X
k
, k = 1, 2, . . . , n is
compact.
Example 88 It follows that the rectangle [a, b] [c, d] R
2
is compact, whence it follows that the
cuboid [a, b] [c, d] [e, f] R
3
is also compact.
There are connexions between the concepts of closedness and compactness, but the interplay is
rather subtle.
Theorem 89 A closed subset of a compact space is compact.
Note: both hypotheses are necessary:
A subset of a compact space need not be compact, e.g. (0, 1) [0, 1].
A closed subset of a space need not be compact, e.g. [0, ) R.
Proof of Theorem 89: Let X be compact and C be closed in X. Let U

: be a family of
open sets in X such that C

. Now XC is open in X, and


X =
_
_

(XC),
so XC U

: is an open cover of X. But X is compact, so there is a nite subcover


XC, U
1
, . . . , U
n
. Since this collection covers X, and C (XC) = , it follows that
C
n
_
i=1
U
i
.
33
Hence C is compact.
Example 90 The 2-sphere S
2
R
3
is compact.
Recall S
2
= (x
1
, x
2
, x
3
) R
3
: x
2
1
+x
2
2
+x
2
3
= 1. This is closed in the cube [2, 2][2, 2][2, 2],
which is compact.
Note: A compact subset of a compact space need not be closed.
Counterexample 91 Let X = a, b with topology = , X, a. Then X and the subset
C = a are trivially compact (they are both nite). But XC = b which is not open, so C is not
closed.
Note that this counterexample is not Hausdor. Theres a good reason for that:
Theorem 92 A compact subset of a Hausdor space is closed.
Proof: Let X be a Hausdor space and C X be compact. We must show that XC is open.
Let x XC. For each c C we can nd open sets U
c
, V
c
in X such that c U
c
, x V
c
and
U
c
V
c
= (X is Hausdor). The family of open sets U
c
: c C covers C, and hence has a nite
subcover U
c1
, U
c2
, . . . , U
cn
(C is compact). Let
W
x
=
n

i=1
V
ci
.
W
x
is open in X (its a nite intersection of open sets), and x W
x
. Also
W
x
C W
x

_
n
_
i=1
U
ci

=
n
_
i=1
W
x
U
ci
=
since y U
ci
implies y / V
ci
W
x
. Hence W
x
XC, so
XC =
_
xX\C
W
x
which, being a union of open sets, is open.
It follows that in a Hausdor space, the intersection of two compact sets is compact. This is false for
general topological spaces (see Counterexample 83). In the following we prove compactness of the
intersection of an arbitrary indexed family of compact subsets (rather than just a pair of subsets).
Corollary 93 Let C

: be a family of compact subsets of a Hausdor space. Then

is a compact subset of X.
Proof: For all , C

is a compact subset of a Hausdor space, and hence is closed in X


(Theorem 92). Hence

is closed (Theorem 34). Let . Then

, which is
compact, so

is compact by Theorem 89.


Recall that, in general, to show that a bijective map f : X Y is a homeomorphism we must check
that both f and f
1
are continuous. The following theorem shows that if X is compact and Y is
Hausdor, then we need only check that f is continuous.
Theorem 94 Let X be a compact space, Y be a Hausdor space and f : X Y be bijective and
continuous. Then f is a homeomorphism.
34
Proof: We must show that f
1
: Y X is continuous. So, let U X be open. Then XU is
closed in X, and hence is compact (by Theorem 89). Since f is continuous, f(XU) is compact
(Theorem 84) and so, being a compact subset of a Hausdor space, is closed (Theorem 92). But
f(XU) = f(X)f(U) = Y f(U), since f is both injective and surjective. Hence (f
1
)
1
(U) =
f(U) is open, so f
1
is continuous.
Corollary 95 The only Hausdor topology on a nite set is the discrete topology.
Proof: Let X be a nite set and be a Hausdor topology on X. Let
0
be the discrete topology
on X, and consider the function
f : (X,
0
) (X, ), f(x) = x.
It is clearly bijective and continuous (every map g : (X,
0
) Y is trivially continuous). Hence, by
Theorem 94, f is a homeomorphism. Now every subset U X is open in (X,
0
), so f(U) = U is
open in (X, ) by continuity of f
1
. Hence
0
, and
0
(trivially), so =
0
.
Recall that a metric space is a pair (X, d) where X is a set and d is a metric (distance function)
on X (see Denition 1). Recall also that every metric space has a natural topology induced by the
metric d (see Denition 18).
Denition 96 Let (X, d) be a metric space, and A be a subset of X. Then A is bounded with
respect to d if there exist x X and > 0 such that A B

(x).
Theorem 97 Every compact subset of a metric space is closed and bounded.
Proof: Let (X, d) be a metric space, and C X be compact (with respect to the topology induced
by d). Then X is Hausdor (Theorem 27), so C is closed (Theorem 92).
Fix x X. Then C

nZ
+ B
n
(x) since if y X, then y B
n
(x) for any n > d(x, y). Hence
B
n
(x) : n Z
+
is an open cover of C. But C is compact, so there exists a nite subcover
B
ni
(x) : i = 1, 2, . . . , k. Hence
C
k
_
i=1
B
ni
(x) = B
N
(x)
where N = maxn
1
, n
2
, . . . , n
k
, so C is bounded.
WARNING! The converse of this Theorem is false!
Counterexample 98 We seek a closed bounded subset of a metric space which is not compact.
Let X be any innite set and d be the discrete metric:
d(x, y) =
_
0 x = y
1 x ,= y.
Let C be any innite subset of X. C is closed in X (all subsets are closed, since the topology induced
by d is the discrete topology). Further,
C B
2
(x) = X
for any x X, so C is bounded. However, C is not compact, since c : c C is an open cover
of C which has no nite subcover.
35
The converse is true in the special case where X is R
n
with the usual topology. Recall this is the
metric topology induced by the standard (or Euclidean) metric
d(x, y) =

_
n

i=1
(x
i
y
i
)
2
, ()
or by any metric equivalent to d. This result is the general version of the Heine-Borel Theorem:
Theorem 99 (Heine-Borel, R
n
) A subset of R
n
, with the usual topology, is compact if and only
if it is closed and bounded with respect to the standard metric (see ()).
Proof: () follows immediately from Theorem 97.
() Let C R
n
be closed and bounded (w.r.t. standard metric d). Then C B

(a) for some


> 0 and a R
n
. But
x B

(a) [(x
1
a
1
)
2
+ + (x
n
a
n
)
2
]
1
2

[x
i
a
i
[ for all i = 1, 2, . . . , n
x
i
[a
i
, a
i
+] for all i = 1, 2, . . . , n.
Hence C [a
1
, a
1
+] [a
2
, a
2
+] [a
n
, a
n
+]. Therefore C is a closed subset of
a compact space (Tychonos Theorem), and hence is compact (Theorem 89).
Theorem 100 Let X be a compact topological space and f : X R be continuous. Then f is
bounded and attains both its supremum and inmum. Hence f attains both a maximum and a
minimum value.
Proof: Since f is continuous and X is compact, f(X) R is compact (Theorem 84), and hence
closed and bounded with respect to d(x, y) = [xy[ (Theorem 99). So sup f(X) and inf f(X) exist.
Let s = sup f(X), and suppose s / f(X). Then s Rf(X), which is open (f(X) is closed), so
there exists > 0 such that (s , s + ) Rf(X). In particular (s , s] f(X) = , which
contradicts s = sup f(X) (for example, s /2 is an upper bound on f(X), and is less than s).
Hence s f(X). That inf f(X) f(X) follows similarly.
Note that a continuous real valued function on a noncompact domain may be unbounded, e.g.
f : (0, 1) R where f(x) = 1/x,
or even if bounded, may attain neither its supremum nor inmum, e.g.
f : (0, 1) R where f(x) = x.
36
7 Closure, interior and boundary
7.1 Closure and derived set
Denition 101 Let A be a subset of a topological space X. The closure of A, denoted A, is the
intersection of all closed subsets of X which contain A.
Example 102 X = a, b, c, = , X, a, a, b. What are the closures of the singleton sets
A = a, B = b, C = c?
Proposition 103 A is the smallest closed subset of X containing A. That is, A is closed and
contains A, and every closed set containing A contains A.
Proof: Let C denote the collection of all closed subsets of X containing A:
C = C X : C is closed and A C.
Then A =

CC
C. A is closed, since it is the intersection of a familiy of closed sets (Theorem 34).
If x A then x C for all C C, so x A. Hence A A, so A C (it is a closed set containing
A). If x A then x C for all C C, so A C for all C C. This completes the proof.
Corollary 104 A is closed if and only if A = A.
Example 105 In R with the usual topology, A = (0, 1) has closure A = [0, 1], and B = (0, 1)(1, 2)
has closure B = [0, 2], by Proposition 103.
We can give another characterization of the set A by using the notion of neighbourhoods.
Denition 106 Let X be a topological space and x X. A neighbourhood of x is any open set
containing x.

Example 107 (a). In X = R, (0, 5) is a neighbourhood of 1, since it is open and contains 1. [1, 5)
is not a neighbourhood of 1.
(b). In X = [0, ) with the subspace topology from R, [0, 37) is a neighbourhood of 0: it contains
0 and is open in X since [0, 37) = (37, 37) X for example.
(c). In X with the discrete topology, any set containing x is a neighbourhood of X, since all sets
are open. E.g. x is a neighbourhood of X.

Many authors dene a neighbourhood of x to be any set which contains an open set containing x, so that a neigh-
bourhood is not necessarily open. They would then call what we are calling a neighbourhood an open neighbourhood.
This distinction will not be important for our purposes, but you should bear it in mind when reading topology texts.
37
(d). In X with the trivial topology, the only neighbourhood of x X is X itself, since X is the
only open set containing x.
Theorem 108 Let A be a subset of topological space X and let x X. Then x A if and only if
every neighbourhood U of x meets A, that is, U A ,= .
Proof: () Let x A and U be neighbourhood of x. If U A = then A XU, which is closed,
so A XU (Proposition 103). But then A U = , a contradiction (x A U).
() Assume that every neighbourhood of x meets A. XA is open and does not meet A, since
A A. Hence it does not contain x. Hence x A.
Example 109 In R with the usual topology, the closure of Q is Q = R. This follows immediately
from Theorem 108. Let x be any real number. Then every open set U containing x contains a
rational number, so U Q ,= . Hence x Q.
Denition 110 A subset A of a topological space X is dense in X if A = X.
Clearly Q is dense in R. What are the dense subsets of X = a, b, c with respect to the topology
of Example 102?
Example 102 (revisited) The set of subsets of X is
, A, B, C, b, c, a, c, a, b, X.
Clearly is not dense and X is dense. The only dense singleton set is A = a. Hence any set
containing A must also be dense (why?). So a, b and a, c are dense. That leaves b, c. But
this is closed, so b, c = b, c (Corollary 104), and hence b, c is not dense. So the list of dense
subsets of X is
a, a, b, a, c, a, b, c.
Theorem 111 Let A be a connected subset of a topological space X and let B be a subset of X such
that A B A. Then B is connected. In particular, A is connected.
Proof: Suppose there is an open partition U B, V B of B, where U, V are open subsets of X.
Neither set in this partition is empty, so there exist x U B and y V B. U is an open set
containing x and x A, so U A ,= by Theorem 108. Similarly V A ,= . U A and V A are
disjoint, since
(U A) (V A) (U B) (V B) = ,
and their union is A, since
(U A) (V A) = (U A B) (V A B) = [(U B) (V B)] A = B A = A.
Hence U A, V A is an open partition of A, a contradiction since A is connected. Hence no open
partition of B exists, that is, B is connected.
So the closure of a connected set is connected.
WARNING! The closure of a compact set need not be compact!
Counterexample 112 On Z dene the following topology: U Z is open if U = or 0 U. Then
0 is a compact subset of Z (its nite!). The only closed set containing 0 is Z (all nonempty
open sets contain 0). Hence 0 = Z. But Z is noncompact. For example 0, n : n Z is an
open cover of Z which has no nite subcover.
38
Closely related to the closure of a set A is the derived set, or set of cluster points of A:
Denition 113 Let A be a subset of a topological space X. A cluster point (or limit point or
accumulation point) of A is any point x X with the property that every neighbourhood of X
contains at least one point in A dierent from x. The set of cluster points of A is called the derived
set of A, denoted A

.
Example 114 (a). Let X = a, b, c with topology = , X, a, a, b and A = b, c. What
is A

?
a A

? No: U = a is a neighbourhood of a and U A = .


b A

? No: V = a, b is a neighbourhood of b and V A = b which does not contain


a point dierent from b.
c A

? Yes: X is the only neighbourhood of c, and X A = b, c does contain a point


dierent from c, namely b.
Hence A

= c. In this case A

A, but this is not so in general.


(b). Let X = R and A = (0, 1). Then A

= [0, 1], so in this case A A

.
(c). Let X = R and A = 1/n : n Z
+
. Then A

= 0, so in this case A A

= .
Proposition 115 Let X be a topological space and A be a subset of X. Then A = A A

.
Proof: If x A

then every neighbourhood of x meets A, so x A by Theorem 108. Hence A

A.
Also A A by denition, so A A

A. Conversely, assume x A. Then every neighbourhood of


x meets A. Hence, if x / A, then neighbourhood of x meets A at some point dierent from x, so
x A

. Hence x A, or x A

. Thus A A A

.
7.2 Interior and boundary
Denition 116 Let A be a subset of a topological space X. The interior of A, denoted A

or
int A, is the union of all open subsets of X which are contained in A. The boundary of A in X is
A = AA

.
Example 117 X = a, b, c, = , X, a, a, b. What are the interiors and boundaries of the
nonempty subsets of X?
39
Proposition 118 A

is the largest open subset of X contained in A. That is, A

is open and
contained in A, and every open set contained A is a subset of A

.
Proof: Let C denote the collection of all open subsets of X contained in A:
C = U X : U is open and U A.
Then A

UC
U. A

is open, since it is the union of a family of open sets. If x A

then there
exists U C such that x U A, so A

A. Hence A

C (it is an open set contained in A). If


x U C then x A

(denition of union), so U A

for all U C. This completes the proof.


Corollary 119 A is open if and only if A = A

.
Example 120 In R with the standard topology, (0, 1]

= (0, 1) and Q

= (since no subset of Q
is open in R). It follows that
(0, 1] = (0, 1](0, 1]

= [0, 1](0, 1) = 0, 1
and
Q =
Example 121 Consider Z with the topology of Counterexample 112 (open sets are and any
set containing 0). Let O, E denote the subsets of odd and even integers respectively. What are
O, O

, O, E , E

, E ?
O =
O

=
O =
E =
E

=
E =
Recall that the closure of a connected set is connected.
WARNING! The interior of a connected set need not be connected.
Counterexample 122 In R
2
, let
A = (x
1
, x
2
) R
2
: [x
1
[ 1 or x
2
= 0.
40
This is connected (prove it!). Its interior is
A

= (x
1
, x
2
) R
2
: [x
1
[ > 1
which is not.
While were here, lets determine the closure and
boundary of A. A = A since A is closed, and so
A = (x
1
, x
2
) R
2
: [x
1
[ = 1
(x
1
, x
2
) R
2
: [x
1
[ < 1 and x
2
= 0.
In fact, we can dene interior and boundary directly using only closure:
Proposition 123 For any subset A of a topological space X,
A

= X(XA) and A = A (XA)


Proof: Exercise.
41
8 Sequences
8.1 Basic denitions
A sequence in a topological space X is a map x : Z
+
X (or perhaps x : N X). Following
historical custom, we will denote the image of n under the mapping x by x
n
rather than x(n), and
denote the map as a whole by (x
n
) rather than x.
A key question concerning sequences in R, C, etc. is whether they converge. Our rst task is to
write down a denition of convergence for sequences which generalizes the familar N denition
of convergence for real-valued sequences.
Reminder 124 A sequence (x
n
) in R converges to x R if for each > 0 there exists some
N() Z
+
such that [x
n
x[ < for all n N().
The trick to generalizing this is to identify an appropriate analogue of (x , x + ) R, a set of
points close to x.
Proposition 125 A real sequence (x
n
) converges to x R if and only if for each neighbourhood A
of x, there exists N Z
+
such that x
n
A for all n N.
Proof: () Let A be a neighbourhood of x. Then there exists > 0 such that (x , x +) A.
Now (x
n
) converges to x, so there exists N Z
+
such that, for all n N, [x x
n
[ < , that is,
x
n
(x , x +) A.
() Choose any > 0. Then A = (x , x +) is a neighbourhood of x, so there exists N Z
+
such that, for all n N, x
n
(x , x +), that is, [x x
n
[ < .
In time-honoured fashion we will now convert this theorem into a denition for convergence in a
general topological space (where, recall, we may have no notion of distance).
Denition 126 A sequence (x
n
) in a topological space X converges to x X if for every neigh-
bourhood A of x there exists N Z
+
such that x
n
A for all n N. We write x
n
x. The point
x is called a limit of (x
n
).
Note: x is a limit, not the limit. The reason will become clear shortly.
Example 127 (a). Eventually constant sequences converge.
Let X be any topological space and (x
n
) be a sequence which is eventually constant, that
is, there exist x X and k Z
+
such that x
n
= x for all n k. Then x
n
x, whatever the
topology on X. To see this, note that given any neighbourhood A of x, we can take N = k,
since then for all n N, x
n
= x A.
(b). In the discrete topology ONLY eventually constant sequences converge.
Let (x
n
) x in X with the discrete topology. Then A = x is a neighbourhood of x. Hence
there exists N Z
+
such that n N implies x
n
A = x, that is, x
n
= x. So (x
n
) is
eventually constant.
(c). In the trivial topology, all sequences converge to everything!
Let (x
n
) be any sequence and x X. The only neighbourhood of x is X, and all terms of (x
n
)
are in X, so we can take N = 1 (for example). Clearly, limits of sequences are not necessarily
unique!
(d). Let X = a, b, c with topology = , X, a, a, b. Dene the sequence (x
n
) by
x
n
=
_
a if n is odd
b if n is even
42
Then x
n
c since the only neighbourhood of c is X, which contains the whole sequence.
Also x
n
b since the neighbourhoods of b are a, b and X, and each of these contains the
whole sequence.
However x
n
a since a is a neighbourhood of a, and there is no N Z
+
such that x
n
= a
for all n N.
The good news is that the limit of a convergent sequence in a Hausdor space is unique.
Theorem 128 Let X be a Hausdor topological space and (x
n
) be a convergent sequence in X.
Then (x
n
) has a unique limit.
Proof: Suppose x
n
x and x
n
y, where x ,= y. Since X is Hausdor, there exist neighbourhoods
U, V of x, y respectively such that U V = . But x
n
x, so there exists N
1
such that n N
1
implies x
n
U, and x
n
y, so there exists N
2
such that n N
2
implies x
n
V . Hence
x
N1+N2
U V = , a contradiction.
It follows that convergent sequences in metric spaces have unique limits (recall metric topologies
are Hausdor). The next result reformulates convergence in metric spaces to make contact with the
familiar N defnition of convergence you know and love from Real Analysis (see Reminder 124
above).
Theorem 129 Let (X, d) be a metric space, equipped with the topology induced by d. Then x
n
x
if and only if d(x
n
, x) 0 in R (in the sense of Reminder 124).
Proof: () Suppose x
n
x. Let > 0. Then B

(x) is a neighbourhood of x, so there exists


N Z
+
such that x
n
B

(x) for all n N, that is, d(x, x


n
) = [d(x, x
n
) 0[ < for all n N.
() Let A be a neighbourhood of x. Then there exists > 0 such that B

(x) A. Now, since


d(x, x
n
) 0, there exists N Z
+
such that d(x, x
n
) < for all n N. Hence x
n
B

(x) for all


n N, which implies that x
n
A for all n N. So x
n
x.
We can use the convergence properties of sequences in a topological space X to give alternative
denitions of the notions of closure, continuity and compactness in X. These sequential denitions
turn out to be weaker than the original denitions in general, but are equivalent in the (important)
special case that X is a metric space.
8.2 Closure in terms of sequences
Denition 130 Let A be a subset of a topological space X. The limit set of A in X is the set
A
lim
= x X : there is a sequence (x
n
) in A such that x
n
x in X,
that is, the set of all limits of convergent sequences taking values in A.
Example 131 Let X = a, b, c, = , a, a, b, X and A = a. What is A
lim
?
In general A
lim
is not closed. However, for a large class of topological spaces (including metric
spaces) it is closed, and coincides with A, the closure of A.
43
Theorem 132 Let A be a subset of a topological space X. Then
(a). A A
lim
A.
(b). If the topology on X is induced by a metric, then A
lim
= A.
Proof:
(a). Let a A. Then the constant sequence x
n
= a converges to a, so a A
lim
. Let x A
lim
.
Then there is a sequence (x
n
) in A such that x
n
x in X. Let U X be any neighbourhood
of x. Since x
n
x, there exists N Z
+
such that x
n
U for all n N. Hence U A ,=
(it contains x
n
: n N). Hence x A (Theorem 108).
(b). Let the topology on X be induced by a metric d. We must establish that A A
lim
. So, let
x A. We seek a sequence (x
n
) in A which converges to x in X. For each n Z
+
, consider
the open ball B1
n
(x). This is an open set containing x, so it must meet A. Choose any point
in B1
n
(x) A and call it x
n
. Clearly this denes a sequence in A. Furthermore, d(x, x
n
) <
1
n
,
so d(x, x
n
) 0 by the Squeeze Rule. Hence x
n
x (Theorem 129).

8.3 Continuity in terms of sequences


Denition 133 Let X and Y be topological spaces and x X. We say that f : X Y is
sequentially continuous at x if given any sequence (x
n
) in X such that x
n
x, the image
sequence f(x
n
) f(x) in Y . f is sequentially continuous if it is sequentially continuous at every
x X.
Continuity implies sequential continuity:
Theorem 134 Let f : X Y be continuous. Then f is sequentially continuous.
Proof: Let (x
n
) be any sequence converging to x and V be any neighbourhood of f(x). Then
U = f
1
(V ) is open in X and contains x. Hence it is a neighbourhood of x. Hence, there exists
N Z
+
such that x
n
U for all n N. But then f(x
n
) f(U) = f(f
1
(V )) V for all n N.
Hence f(x
n
) f(x).
WARNING! The converse is false! Sequential continuity does not imply continuity. Unfortu-
nately, its hard to write down a simple counterexample to illustrate this. If you know what it
means for a set to be countable, you will understand the following counterexample. If not, dont
worry about it. You just need to know that counterexamples exist, and that they are rather bizarre.
Counterexample* Let X = R equipped with the topology where U R is open if it is , R, or
if RU is countable. Let Y = R equipped with the usual topology, and let f : X Y such that
f(x) = x. This function is not continuous. For example (0, ) is open in Y , but f
1
((0, )) = (0, )
is not open in X since (, 0] is not countable. But f is sequentially continuous at each x X.
To see this, let (x
n
) be a sequence converging to x in X and let V Y be any neighbourhood of
f(x) = x in Y . Consider the following subset of X:
C = x
n
: n Z
+
, x
n
/ V x.
C is countable and x / C, so U = XC is a neighbourhood of x in X (its open and contains x).
But x
n
x in X, so there exists N Z
+
such that x
n
U for all n N. But then for all n N,
x
n
/ C so x
n
V . Hence x
n
x in Y . Note that X is not Hausdor, so the bizarre topology in
this counterexample is not a metric topology. Theres a good reason for this.
44
Theorem 135 Let X be a metric space, Y be a topological space, and f : X Y be sequentially
continuous. Then f is continuous.
Proof: We will prove the contrapositive: if f is not continuous, then f is not sequentially continuous.
Let d be the metric on X and asume that f is discontinuous. Then there exists V Y , open,
such that f
1
(V ) X is not open. Hence, there exists x f
1
(V ) such that for all > 0,
B

(x)/f
1
(V ). Hence, for each n Z
+
, there exists some x
n
B1
n
(x) such that x
n
/ f
1
(V ), that
is, f(x
n
) / V . Consider the sequence (x
n
). Clearly x
n
x by Corollary 129 (since d(x, x
n
) <
1
n
).
But f(x
n
) does not converge to f(x), since V is a neighbourhood of f(x) and for all n Z
+
,
f(x
n
) / V .
8.4 Compactness in terms of sequences
Recall that given a sequence (x
n
) we can obtain subsequences from it by omitting some (possibly
innitely many) of the terms.
Denition 136 Given a sequence (x
n
), a subsequence of (x
n
) is any sequence (y
k
) where y
k
=
x
n
k
, (n
k
) being an increasing sequence of positive integers (n
1
< n
2
< n
3
< ).
Example 137 In R the sequence y
k
= 1/2
k
is a subsequence of x
n
= 1/n (we take n
k
= 2
k
). It
is also a subsequence of z
n
= (1)
n
/n (again n
k
= 2
k
). It is not a subsequence of w
n
= 1/n
2
(for
example y
3
= 1/8 and there is no n such that w
n
= 1/8).
Is v
k
= 1/(2[k 3[ + 1) a subsequence of (x
n
)? No! Although every term of (v
k
) is a term of
(x
n
), the terms dont appear in the same order as they do in (x
n
):
v
1
=
1
5
= x
5
, v
2
=
1
3
= x
3
, v
3
= 1 = x
1
, v
4
=
1
2
= x
3
,
and (x
5
, x
3
, x
1
, x
3
, x
5
, x
7
, . . .) is not a subsequence of (x
n
).
We now dene an alternative notion of compactness, called sequential compactness. (Recall: X is
compact if every open cover of X has a nite subcover.)
Denition 138 A topological space is sequentially compact if every sequence in X has a con-
vergent subsequence.
Example 139 (a). Any set X with the trivial topology = , X is sequentially compact (every
sequence converges). Note that X is compact.
(b). A nite set X with any topology is sequentially compact. To see this note that any sequence
(x
n
) can only take a nite number of values, so at least one value, c X, say must occur
innitely often. But then the constant sequence (c, c, c, . . .) is a convergent subsequence of
(x
n
). Note that (X, ) is compact.
(c). Let X = (0, 1) with the subspace topology from R. Then X is not sequentially compact. For
example
x
n
=
_
1
n
n odd
1
1
n
n even
is a sequence in X with no convergent subsequence in X. Note that X is not compact either.
(d). Let X = [a, b], again with the subspace topology fromR. Its a classic result from Real Analysis
(the Bolzano-Weierstra theorem) that every sequence in [a, b] has a convergent subsequence.
For example, the sequence above, thought of as a sequence in [0, 1] has (at least) two convergent
subsequences, x
2k
1 and x
2k+1
0. So the Bolzano-Weierstra theorem implies that [a, b]
is sequentially compact. Note that [a, b] is compact.
45
In fact, we will show that for metric spaces compactness and sequential compactness are equivalent.
Along the way, we will prove a metric space generalization of the Bolzano-Weierstra theorem. Not
surprisingly compact sequentially compact is somewhat easier to prove than the converse. The
key is to consider the properties of cluster points in metric spaces:
Proposition 140 Let X be a metric space, A be a subset of X and x be a cluster point of A. Then
every neighbourhood of x contains innitely many points in A.
Proof: Assume not, so there is some neighbourhood U of x such that U (Ax) is nite
x
1
, x
2
, . . . , x
n
. Then U contains an open set containing x, so there exists > 0 such that B

(x)
U. Let = min, d(x, x
1
), d(x, x
2
), . . . , d(x, x
n
) > 0. Then B

(x) contains no points in A, except,


perhaps x. But B

(x) is a neighbourhood of x, a contradiction (x is a cluster point of A).


Lemma 141 Let X be a metric space (x
n
) be a sequence in X and A = x
n
: n Z
+
. If x is a
cluster point of A then there exists a subsequence (x
n
k
) which converges to x.
Proof: B
1
(x) contains innitely many points of A (Proposition 140). Pick one and call it x
n1
.
Then B1
2
(x) also contains innitely many points of A (Proposition 140). Some of these must appear
after x
n1
in the sequence (x
n
), since only n
1
terms do not. Pick one of these and call it x
n2
. Then
B1
3
(x) also contains innitely many points of A (Proposition 140). Some of these must appear after
x
n2
in the sequence (x
n
), since only n
2
terms do not. Pick one of these and call it x
n3
.
In this way we dene, for each k Z
+
some x
n
k
B1
k
(x) such that n
k
> n
k1
. Hence (x
n
k
) is
a subsequence of (x
n
). Clearly x
n
k
x since d(x, x
n
k
) <
1
k
(Theorem 129).
Theorem 142 Let X be a compact metric space. Then X is sequentially compact.
Proof: We will prove the contrapositive. So assume that X is not sequentially compact, that is,
there is some sequence (x
n
) in X with no convergent subsequence. We will construct an open cover
of X which has no nite subcover.
Let A = x
n
: n Z
+
. Clearly A is innite (if not, some value x
n
= c of the sequence is
repeated innitely many times, so there is a constant, hence convergent, subsequence (c, c, c, . . .)).
By Lemma 141, A has no cluster point. Hence, given any x XA, there exists some open set in
X, U
x
say, such that U
x
A = . Also, given any x A, there exists some open set in X, V
x
say,
such that V
x
A = x. Consider the open cover
C = U
x
: x XA V
x
: x A.
Each x A is contained in V
x
, but in no other set in C, so any subcover of C must contain every
V
x
, x A. But A is innite, so C has no nite subcover. Thus X is noncompact.
Denition 143 Let A be a subset of a metric space (X, d). The induced metric on A is d
A
:
A A R, where d
A
(x, y) = d(x, y). So d
A
is the restriction of d to A. Note that d
A
clearly is a
metric on A (check it!).
Proposition 144 Let (X, d) be a metric space and A be a subset of X. The metric topology induced
by d
A
on A coincides with the subspace topology on A.
Proof: Exercise.
46
Corollary 145 (The Bolzano-Weierstra Theorem) Every bounded sequence in (R
n
, d), where
d is the usual metric, has a convergent subsequence.
Proof: Since the sequence is bounded, it is contained in some ball B

(x), and hence in the closed


ball A = y R
n
: d(x, y) . But A is closed and bounded, hence compact (Heine-Borel), and
is a metric space (Proposition 144), so is sequentially compact (Theorem 142).
So every compact (subset of a) metric space is sequentially compact. We will now prove the converse:
every sequentially compact metric space is compact. The proof rests on two lemmas (lemmata if
youre prissy about Greek plurals).
Lemma 146 (Lebesgue Covering Lemma) Suppose a metric space (X, d) is sequentially com-
pact and C = U

: is an open cover of X. Then there exists > 0 such that for each x X
there is some
x
such that B

(x) U
x
.
Proof: Suppose no such exists. Then for each > 0 there is some x X such that B

(x)/U

for all . In particular, for each n Z


+
there is some x
n
X such that B1
n
(x
n
)/U

for all
(take =
1
n
in the previous statement). Consider the sequence (x
n
). Since X is sequentially
compact, it has a convergent subsequence (x
n
k
). Denote by y X its (unique) limit. Since C covers
X, there is some such that y U

. U

is open, so there exists > 0 such that B

(y) U

.
Now x
n
k
y, so there exists some K Z
+
such that x
n
k
B
2
(y) for all k K. Take any k K
such that
1
n
k


2
. Then
z B 1
n
k
(x
n
k
) d(x
n
k
, z) <
1
n
k


2
d(y, z) d(y, x
n
k
) +d(x
n
k
, z) <

2
+

2
=
z B

(y) U

.
Hence B 1
n
k
(x
n
k
) U

. But this contradicts the denition of x


n
k
.
Lemma 147 Suppose (X, d) is a sequentially compact metric space. Then for each > 0 there is a
nite collection of points x
1
, x
2
, . . . , x
n
such that X =

n
i=1
B

(x
i
).
Proof: Suppose not. Then there exists > 0 such that no nite collection of radius balls covers
X. But then we can nd a sequence (x
n
) in X such that d(x
m
, x
n
) for all m, n Z
+
, m ,= n
(pick any x
1
X, then choose x
2
XB

(x
1
) ,= , x
3
X(B

(x
1
) B

(x
2
)) ,= etc.). Since X is
sequentially compact, there must exist a convergent subsequence x
n
k
, converging to y, say. Hence,
there exists K Z
+
such that k K implies d(x
n
k
, y) <

2
. But then
d(x
nK
, x
nK+1
) d(x
nK
, y) +d(y, x
nK+1
) <

2
+

2
=
which is a contradiction.
Theorem 148 Let X be a sequentially compact metric space. Then X is compact.
Proof: Let C = U

: be an open cover of X. By Lemma 146 there is some > 0 such


that for all x X there is some
x
for which B

(x) U
x
. But then, by Lemma 147, there is
some nite colection of points x
1
, x
2
, . . . , x
n
such that X =

n
i=1
B

(x
i
). Hence

n
i=1
U
x
i
= X, so
C

= U
x
i
: i = 1, 2, . . . , n is a nite subcover from C.
47
9 Complete metric spaces
9.1 Basic denitions
For the rest of the course, we will concentrate (almost) exclusively on metric spaces. An awkward
aspect of the denition of convergence of a sequence is that we must guess a limit (the limit in a
Hausdor space) before we can test convergence. In a metric space, we can dene a property which
is rather convergence-like, and concerns only the terms of the sequence itself, not some putative
limit. It is called the Cauchy property. It will turn out that every convergent sequence has the
Cauchy property. The class of metric spaces for which the converse is true (that is, every sequence
with the Cauchy property is convergent) is the class of complete metric spaces. They are of immense
practical importance.
Denition 149 Let (X, d) be a metric space. A sequence (x
n
) in X is a Cauchy sequence if for
all > 0 there exists N Z
+
such that d(x
n
, x
m
) < for all n, m > N.
Example 150 (x
n
) =
_
1
n
_
is Cauchy in R with d(x, y) = [x y[. Given > 0, we can take any
N Z
+
such that N >
1

. Then for all n, m N


d(x
n
, x
m
) =

1
n

1
m

< max
1
n
,
1
m

1
N
< .
Theorem 151 Let (x
n
) be a Cauchy sequence in (X, d). Then (x
n
) is bounded.
Proof: There exists N Z
+
such that d(x
n
, x
m
) < 1 for all n, m N. Then d(x
N
, x
n
) < 1 for all
n N. Let
= maxd(x
N
, x
1
), d(x
N
, x
2
), . . . , d(x
N
, x
N1
), 1 > 0
Then x
n
B
+1
(x
N
) for all n Z
+
. Hence (x
n
) is bounded.
Of course, a bounded sequence need not be Cauchy, e.g. x
n
= (1)
n
. A nice thing about Cauchy
sequences is that if any subsequence of a Cauchy sequence converges, the Cauchy sequence converges.
Again, this is not true of sequences in general, e.g. x
n
= (1)
n
again.
Theorem 152 Let (x
n
) be a Cauchy sequence in (X, d) and (x
n
k
) be a subsequence converging to
x X. Then (x
n
) converges to x.
Proof: Given any > 0, there exists K Z
+
such that d(x, x
n
k
) <

2
for all k K (since x
n
k
x).
Also, as (x
n
) is Cauchy, there exists N Z
+
such that d(x
n
, x
m
) <

2
for all n, m N. Choose any
k Z
+
such that k K and n
k
N (for example, k = maxK, N will do). Then for all n N,
d(x, x
n
) d(x, x
n
k
) +d(x
n
k
, x
n
) <

2
+

2
= ,
since k K and n
k
, n N. Hence x
n
x.
Theorem 153 Let (x
n
) be a convergent sequence in (X, d). Then (x
n
) is Cauchy.
Proof: Suppose x
n
x X. Then given any > 0 there exists N Z
+
such that d(x, x
n
) <

2
for all n N. Hence, for all n, m N,
d(x
n
, x
m
) d(x
n
, x) +d(x, x
m
) <

2
+

2
+,
so (x
n
) is Cauchy.
WARNING! The converse is false! For example x
n
= 1/n is Cauchy in (0, ) with d(x, y) = [xy[,
but it is not convergent. (Note 0 / (0, )!)
48
Denition 154 A metric space (X, d) is complete if every Cauchy sequence in X is convergent.
Theorem 155 R
n
with the usual metric is complete. (In particular, R with d(x, y) = [x y[ is
complete.)
Proof: Let (x
n
) be a Cauchy sequence in R
n
with the usual metric. Then (x
n
) is bounded by
Theorem 151. Hence (x
n
) has a convergent subsequence by Corollary 145 (Bolzano-Weierstra).
Hence (x
n
) is convergent by Theorem 152.
WARNING! Completeness is not a topological invariant!
Counterexample 156 Let d be the usual metric on R. Then (R, d) is homeomorphic to ((0, ), d
(0,)
),
e.g. f(x) = e
x
is a homeomorphism. But R is complete, while (0, ) is incomplete, e.g. x
n
=
1
n
is
Cauchy but not convergent in (0, ).
Theorem 157 Let (X, d) be a compact metric space. Then (X, d) is complete.
Proof: Let (x
n
) be a Cauchy sequence in (X, d). Since X is a compact metric space, it is sequentially
compact (Theorem 142), so (x
n
) has a convergent subsequence (x
n
k
). But then (x
n
) is convergent
by Theorem 152.
Note: The converse is false! For example R is complete but is noncompact.
Denition 158 A subset A of a metric space (X, d) is complete if (A, d
A
) is complete.
Example 159 Q, d(x, y) = [x y[, is not complete. For example, for each n Z
+
choose x
n

Q (

2
1
n
,

2 +
1
n
). Then, as a sequence in R, (x
n
) is convergent (it converges to

2). Hence
it is Cauchy with respect to d (Theorem 153). Hence it is Cauchy in Q (the metric is the same).
But it cannot converge in Q: it converges to

2 R and R is Hausdor, so it cannot converge to


anything else, and in particular, it cannot converge to anything rational.
Theorem 160 Let (X, d) be a metric space and A be a complete subset of X. Then A is closed in
X.
Proof: Let x A. Then there exists a sequence (x
n
) in A such that x
n
x (Theorem 132).
Since (x
n
) is convergent, it is Cauchy in (X, d) (Theorem 153), and hence in (A, d
A
). But (A, d
A
) is
complete, so (x
n
) converges in A (that is, it converges to some point in A). But metric spaces are
Hausdor, so the limit of (x
n
) is unique. Hence x A. Hence A A. Clearly A A by denition,
so A = A, and hence A is closed.
Theorem 161 Let A be a closed subset of a complete metric space (X, d). Then A is complete.
Proof: Let (x
n
) be a Cauchy sequence in (A, d
A
). Then (x
n
) is Cauchy in (X, d), hence convergent
in (X, d) ((X, d) is complete). Hence (x
n
) x X. Then x A by Theorem 132 so x A since
A = A (A is closed).
9.2 A metric space of functions
Let X and Y be topological spaces and C(X, Y ) be the set of continuous maps X Y . Can we
give C(X, Y ) a natural topology? If X is compact and Y is a metric space, we can. In fact, we
can dene a metric on C(X, Y ), and it turns out that the resulting metric space is complete if Y is
complete.
49
Denition 162 Let X be a compact topological space, (Y, ) be a metric space, and C(X, Y ) be
the set of continuous maps X Y . We dene the distance between f, g C(X, Y ) to be
d(f, g) = sup(f(x), g(x)) : x X.
This is called the sup metric on C(X, Y ).
So d(f, g) is the maximum pointwise deviation of g from f.
Example 163 (a). Let X = [0, 1], Y = R, (x, y) = [x y[. Then f(x) = x
2
and g(x) = 2x 1
are elements of C(X, Y ). What is the distance between f and g?
d(f, g) = sup[f(x) g(x)[ : x [0, 1] = sup[x
2
2x + 1[ : x [0, 1]
= sup(x 1)
2
: x [0, 1] = 1
(b). Note: its important that X is compact, or else d(f, g) may not exist. E.g. take Y, f, g as above,
but X = R. Then
d(f, g) = sup(x 1)
2
: x R
which does not exist.
Before going any further, we should check that the sup metric really is a metric on C(X, Y ).
Proposition 164 (a). d(f, g) exists for all f, g C(X, Y ).
(b). The function d is a metric on C(X, Y ).
Proof:
(a). Since X is compact and f, g are continuous, both f(X) Y and g(X) Y are compact
subsets of the metric space Y (Theorem 84) and hence are bounded (Theorem 97). Hence,
there exist open balls B
1
(y
1
) f(X) and B
2
(y
2
) g(X). Let =
1
+
2
+d(y
1
, y
2
). Then
B

(y
1
) (f(X) g(X)) (check it!), so given any pair of points y, y

f(X) g(X), (y, y

) <
2. In particular, given any x X, (f(x), g(x)) < 2. Hence (f(x), g(x)) : x X R is
bounded above, so d(f, g) exists.
(b). We must show that d satises the 3 axioms of a metric.
First: d(f, g) = 0 (f(x), g(x)) = 0 x X f = g.
Second: d(f, g) = d(g, f) since (f(x), g(x)) = (g(x), f(x)).
Third: d(f, g) +d(g, h) = sup(f(x), g(x)) : x X + sup(g(y), h(y)) : y X
sup(f(x), g(x)) +(g(x)h(x)) : x X
sup(f(x), h(x)) : x X = d(f, h).

So (C(X, Y ), d) is a genuine metric space. What does convergence mean in this metric space? If
f
n
f in C(X, Y ) then for all > 0 there is some N Z
+
such that whenever n N
sup(f
n
(x), f(x)) : x X <
and hence (f
n
(x), f(x)) < for all x X. Hence f
n
(x) f(x) uniformly in x. This implies, but
is stronger than, the condition of pointwise convergence (that is, f
n
(x) f(x) for each x X).
50
Example 165 Let X = [0, 1], Y = R, (x, y) = [x y[ and consider the sequence of functions in
C(X, Y ),
f
n
(x) = nxe
nx
This converges pointwise to the constant function f(x) = 0 (why?). Since uniform convergence
implies pointwise convergence, if (f
n
) converges uniformly, it must converge to f(x) = 0. But
d(f
n
, f) = sup[f
n
(x) 0[ : x [0, 1] [f
n
(
1
n
) 0[ =
1
e
so d(f
n
, f) 0. So f
n
does not converge uniformly.
We next prove that (C(X, Y ), d) is complete if Y is complete.
Theorem 166 Let X be a compact topological space and (Y, ) be a complete metric space. Then
the metric space (C(X, Y ), d), where d is the sup metric, is complete.
Proof: We must show that every Cauchy sequence is convergent. So, let (f
n
) be a Cauchy sequence
in C(X, Y ). Then for all > 0 there exists N Z
+
such that
d(f
n
, f
m
) = sup(f
n
(x), f
m
(x)) : x X < n, m N
(f
n
(x), f
m
(x)) < n, m N, x X.
Hence, for each x X, the sequence (f
n
(x)) is Cauchy in (Y, ). But (Y, ) is complete, so (f
n
(x))
is convergent for each x X. Call its limit (which is unique since Y is Hausdor) f(x). In this way,
we obtain a function f : X Y , f(x) = limf
n
(x). I claim that f
n
f in C(X, Y ). To prove this,
I must show
(A) f C(X, Y ), that is, f is continuous.
(B) d(f, f
n
) 0.
The result then follows from Theorem 129. Both (A) and (B) will follow from
(C) For all > 0, there exists N Z
+
such that (f(x), f
n
(x)) < for all x X and all n N.
Proof of (C): Choose and x > 0. For all x X, f
n
(x) f(x), so there exists N
x
Z
+
such
that (f
n
(x), f(x)) <

2
for all n N
x
. Since (f
n
) is Cauchy in C(X, Y ), there exists N Z
+
such
that
d(f
n
, f
m
) = sup(f
n
(x), f
m
(x)) : x X <

2
n, m N,
and hence,
(f
n
(x), f
m
(x)) <

2
n, m N, x X.
Now for each x X let

N
x
= maxN
x
, N. Then for each x X, we have that for all n N,
(f(x), f
n
(x)) (f(x), f

Nx
(x)) +(f

Nx
(x), f
n
(x)) <

2
+

2
= .
This establishes claim (C).
Proof of (A): Let V Y be open and consider U = f
1
(V ) X. We must show that U is open.
Let x U. Then f(x) V , so there exists > 0 such that B

(f(x)) V . Now for all z X and


n Z
+
,
(f(z), f(x)) (f(z), f
n
(z)) +(f
n
(z), f
n
(x)) +(f
n
(x), f(x)), ()
by the triangle inequality. By (C), there exists N Z
+
such that [f(z) f
n
(z)[ <

3
for all n N
and all z X. Applying this in the case n = N to () (1st and 3rd terms) yields
(f(z), f(x))

3
+(f
N
(z), f
N
(x)) +

3
().
51
But f
N
: X Y is continuous, and B

3
(f
N
(x)) Y is open, so U
x
= f
1
N
(B

3
(f
N
(x))) is open. But
then, for all z U
x
,
(f(z), f(x))

3
+

3
+

3
= ,
by (), so f(z) B

(f(x)) V , and hence U


x
f
1
(V ) = U. Thus, for each x U, there is an
open set U
x
U such that x U
x
. Hence U =

xU
U
x
, which, being a union of open sets, is open
in X.
Proof of (B): Choose and x > 0. Then by (C) there exists N Z
+
such that (f(x), f
n
(x)) <

2
for all x X and all n N. Hence, for all n N,
d(f, f
n
) = sup(f(x), f
n
(x)) : x X

2
< .
Hence, d(f, f
n
) 0.
It follows immediately that C([a, b], R
n
) and C([0, 1], [0, 1] [0, 1]), equipped with their sup metrics
are complete. We will use these metric spaces to prove existence of solutions of ordinary dierential
equations, and to construct a space lling curve, respectively. Before that we need one more big
theorem.
9.3 Contraction mappings
Denition 167 Let (X, d) be a metric space. A mapping : X X is a contraction mapping
if there exists k (0, 1) such that d((x), (y)) kd(x, y) for all x, y X.
Proposition 168 Every contraction mapping is continuous.
Proof: Since X is a metric space, it suces to show that the contraction mapping : X X
is sequentially continuous (Theorem 135). Let (x
n
) be a sequence converging to x in X. Then
d(x, x
n
) 0, and d((x), (x
n
)) kd(x, x
n
), so d((x), (x
n
)) 0 by the Squeeze Rule, and
hence (x
n
) (x).
Example 169 (a). X = [0, 1], d(x, y) = [x y[, : X X such that (x) =
1
8
(x
3
+2x
2
+4) is a
contraction mapping. Check:
d((x), (y)) =

1
8
(x
3
+ 2x
2
+ 4)
1
8
(y
3
+ 2y
2
+ 4)

=
1
8
[(x
3
y
3
) + 2(x
2
y
2
)[

1
8
[x
3
y
3
[ +
1
4
[x
2
y
2
[
=
1
8
[x y[[x
2
+xy +y
2
[ +
1
4
[x y[[x +y[

1
8
[x y[ 3 +
1
4
[x y[ 2 =
7
8
[x y[.
So we can take k =
7
8
(or any number in [
7
8
, 1)).
(b). The mean value theorem is a useful tool in establishing that smooth real valued functions are
contraction mappings. Heres an illustrative example: X = R, usual metric,
: R R, (x) =
1
4
sin3x.
52
Now d((x), (y)) = [(x) (y)[ and we want to show this is k[x y[ for some k (0, 1).
Clearly this holds automatically if x = y, and if x ,= y, it is equivalent to

(x) (y)
x y

k. ()
But this is dierentiable on R, so we can apply the mean value theorem: there is some c
between x and y such that
(x) (y)
x y
=

(c) =
3
4
cos 3c.
Hence () holds with k =
3
4
, so is a contraction mapping.
(c). X = C([0, 1], R) with the sup metric,
: X X, (f) =
_
x
_
x
0
t(t +f(t)) dt
_
.
You should check this really does map points in X to points in X, that is, for any continuous
function f : [0, 1] R, the image (f) is a continuous function [0, 1] R. I claim that is a
contraction mapping. Check:
d((f), (g)) = sup[(f)(x) (g)(x)[ : x [0, 1].
Now [(f)(x) (g)(x)[ =

_
x
0
[(t(t +f(t)) t(t +g(t))] dt

_
x
0
t(f(t) g(t)) dt

_
x
0
t[f(t) g(t)[ dt

_
x
0
t d(f, g) dt
= d(f, g)
_
x
0
t dt =
x
2
2
d(f, g).
Hence d((f), (g)) sup
x
2
2
: x [0, 1]d(f, g) =
1
2
d(f, g).
So is a contraction mapping (we can take k =
1
2
).
Theorem 170 Let (X, d) be a metric space and : X X be a contraction mapping. Then for
each x
1
X the sequence obtained by iterating the map , that is, x
n+1
= (x
n
) for all n 1, is
Cauchy.
Proof: We rst claim that d(x
n
, x
n+1
) k
n1
d(x
1
, x
2
) for all n Z
+
. Check by induction: clearly
the claim holds for n = 1. Assume it holds for n = p. Then d(x
p+1
, x
p+2
) = d((x
p
), (x
p+1
))
kd(x
p
, x
p+1
) k.k
p1
d(x
1
, x
2
) so it holds for n = p + 1. Hence, by induction, the claim holds for
all n Z
+
.
Now if n < m, then by the triangle inequality,
d(x
n
, x
m
)
m1

p=n
d(x
p
, x
p+1
)
m1

p=n
k
p1
d(x
1
, x
2
) = d(x
1
, x
2
)k
n1
mn1

q=0
k
q
= k
n1
1 k
mn
1 k
d(x
1
, x
2
)
k
n1
1 k
d(x
1
, x
2
).
53
Similarly, if n > m,
d(x
n
, x
m
)
k
m1
1 k
d(x
1
, x
2
).
Hence, for all n, m Z
+
,
d(x
n
, x
m
)
d(x
1
, x
2
)
1 k
maxk
n1
, k
m1
. ()
It follows that (x
n
) is Cauchy.
To see this, note that k (0, 1), so the sequence k
n1
0. Hence for all > 0 there exists
N Z
+
such that
k
n1

1 k
d(x
1
, x
2
)

for all n N. But then for all n, m N, d(x


n
, x
m
) < by ().
WARNING! You might be tempted to think that : X X is a contraction mapping is
equivalent to the condition
d((x), (y)) < d(x, y) for all x, y X, x ,= y ()
but you would be wrong! Condition () is fatally weaker than Denition 167. It does not guarantee
that iterative sequences x, (x), ((x)), . . . are Cauchy in X.
Counterexample 171 Let X = R, usual metric, and
: R R, (x) = log(e
x
+ 1).
This satises condition (): by the mean value theorem, for any pair x, y with x ,= y, there is some
c between x and y such that
d((x), (y))
d(x, y)
=

(x) (y)
x y

= [

(c)[ =

e
c
e
c
+ 1

< 1.
But consider the iterative sequence starting at x
1
= 0:
x
2
= log(e
0
+ 1) = log 2, x
3
= log(e
log 2
+ 1) = log 3, . . . , x
n
= log n.
This is unbounded so cannot be Cauchy (Theorem 151). It follows that is not a contraction
mapping (if it were, wed have a counterexample to Theorem 170).
Denition 172 x X is a xed point of : X X if (x) = x.
Example 173 : R R, (x) = x
2
has xed points 0, 1.
We are now ready to state and prove the big theorem of this section:
Theorem 174 (The Contraction Mapping Theorem) Let (X, d) be a complete metric space
and : X X be a contraction mapping. Then has a unique xed point.
Proof: Take any x
1
X and dene the sequence (x
n
) by iterating , that is, x
n
= (x
n1
)
for all n 2. Then (x
n
) is Cauchy (Theorem 170), and hence convergent ((X, d) is assumed
complete). Hence (x
n
) x for some x X. Now is (sequentially) continuous (Proposition 168),
so (x
n
) (x). But (x
n
) = x
n+1
, and clearly x
n+1
x also. Hence, since limits are unique in
metric spaces, we have that (x) = x, that is, x is a xed point of .
Suppose that (y) = y also. Then
d(x, y) = d((x), (y)) kd(x, y)
54
for some k (0, 1) since is a contraction mapping. This is a contradiction unless d(x, y) = 0.
Hence x = y.
Note: This theorem doesnt just establish the existence of a xed point, it gives us a way to nd
it, approximately, to any prescribed accuracy! The point is that the unique xed point is x = limx
n
where x
1
is arbitrary and x
n
= (x
n1
). So, pick any x
1
X. Then for n suciently large x
n
will
be close to x. How close?
d(x
2
, x) = d((x
1
), x) = d((x
1
), (x)) kd(x
1
, x)
d(x
3
, x) = d((x
2
), x) = d((x
2
), (x)) kd(x
2
, x) = k
2
d(x
1
, x)
.
.
.
d(x
n
, x) = d((x
n1
), x) = d((x
n1
), (x)) kd(x
n1
, x) = k
n1
d(x
1
, x).
Of course, we dont know x, so we dont know d(x
1
, x). But we can get round this with the following
cunning ruse:
d(x
1
, x) d(x
1
, x
2
) +d(x
2
, x) d(x
1
, x
2
) +kd(x
1
, x)
(1 k)d(x
1
, x) d(x
1
, x
2
)
d(x
1
, x)
d(x
1
, x
2
)
1 k
.
So we have an upper bound on the error if we approximate x by x
n
, in terms of things we can
compute:
d(x
n
, x)
k
n1
1 k
d(x
1
, x
2
).
If we are happy to accept an approximation to the xed point x lying within distance > 0 of it,
then we can use any x
n
with n chosen suciently large that
k
n1
1 k
d(x
1
, x
2
) < .
Note that such an n always exists, and we can compute it knowing only the rst two terms of the
sequence.
Example 175 (a). Show that x
3
+ 2x
2
8x + 4 = 0 has a unique solution in [0, 1].
Consider the function : [0, 1] [0, 1], (x) =
1
8
(x
3
+2x
2
+4), and note that x [0, 1] satises
the equation above if and only if (x) = x. Now [0, 1] with the usual metric is a complete
metric space (its compact) and we showed in Example 169 that is a contraction mapping
with k =
7
8
. Hence, by the Contraction Mapping Theorem, the equation has a unique solution
in [0, 1]. Furthermore, lets say we want to nd the solution within an error of = 0.001. Then
we can take the sequence x
n
= (x
n1
) starting at x
1
= 0, say. Now
x
2
= (0) =
1
8
(0 + 0 + 4) =
1
2
,
so d(x
1
, x
2
) = [0
1
2
[ =
1
2
. Hence we must choose n large enough so that
(7/8)
n1
1 (7/8)

1
2
< 0.001

_
7
8
_
n1
< 0.00025
(n 1) log
7
8
< log 0.00025
n 1 >
log 0.00025
log(7/8)
n 64.
55
Is this is a sensible method to solve this problem numerically?
(b). Show that 4x = sin3x x = 0.
Clearly x = 0 solves the equation, so we just need to show that no other x does. Note that x
satises the equation if and only if it is a xed point of : R R, (x) =
1
4
sin 3x. But is a
contraction mapping (Example 169), so its xed point is unique. Hence 0 is the only solution.
(c). Show that the initial value problem y(0) = 0 for the ordinary dierential equation
dy
dx
= x(x +y)
has a unique solution valid for all x [0, 1].
We must show that there is one and only function f : [0, 1] R with the properties that
f

(x) = x(x +f(x)) for all x [0, 1], and f(0) = 0.


Note that such a function must be dierentiable, and hence is continuous. So we can think of
it as a point in C([0, 1], R). Furthermore,
f(x) =
_
x
0
f

(t)dt =
_
x
0
t(t +f(t)) dt = (f)(x)
by the fundamental theorem of calculus, where : C([0, 1], R) C([0, 1], R) is the function
we dened in Example 168,
(f) =
_
x
_
x
0
t(t +f(t)) dt
_
.
So y = f(x) solves the initial value problem under consideration if and only if f is a xed
point of . But we showed that is a contraction mapping, and crucially that C([0, 1], R) is
complete (Theorem 166) so has one and only one xed point by the Contraction Mapping
Theorem. Again, we can approximate it by taking any f
1
C([0, 1], R), for example, f
1
(x) = 0,
and generating the sequence f
n
= (f
n1
) for n 2.
56
9.4 Picards method
The argument used in the last part of Example 175 can be generalized to give a a general method
to establish existence and uniqueness of solutions to a large class of ordinary dierential equations.
The case of linear ODEs turns out to be crucially dierent from the case of nonlinear ODEs. We will
start with the former. In fact, for reasons that will become clear shortly, we will consider coupled
systems of rst order linear dierential equations. The main idea is to reinterpret the solution of
a system of ODEs as a xed point of a related mapping on a suitable function space. For systems
of n rst order ODEs, the required function space is C([a, b], R
n
). We know this is complete with
respect to the sup metric
d(f, g) = sup(f(x), g(x)) : x [a, b]
if we give R
n
the Euclidean metric
Euc
(x, y) = [

n
i=1
(x
i
y
i
)
2
]
1
2
by Theorem 166, since (R
n
,
Euc
)
is complete (Theorem 155). For our purposes it is rather more convenient to give R
n
the metric
(x, y) = max[x
i
y
i
[ : i = 1, 2, . . . , n. ()
We showed that and
Euc
are equivalent metrics, that is, they induce the same topology on R
n
.
Recall, however, that completeness is not a topological invariant so completeness of (R
n
,
Euc
) does
not imply completeness of (R
n
, ).
Proposition 176 R
n
equipped with the max metric () is complete.
Proof: Let X = 1, 2, . . . , n equipped with the discrete topology. Note X is trivially compact.
We can interpret any x R
n
as a map x : X R (x(i) = x
i
). Every such map is continuous, so R
n
is identied with C(X, R). If we give R the usual metric,

(a, b) = [a b[ then (R,

) is complete,
so C(X, R) is complete with respect to the sup metric (Theorem 166). But the sup metric is
d(x, y) = sup

(x(i), y(i)) : i X = max[x


i
y
i
[ : i = 1, 2, . . . , n = (x, y).
Hence (R
n
, ) is complete.
Theorem 177 Let M = (M
ij
), A = (A
i
) be an n n and an n 1 matrix of continuous functions
M
ij
: [a, b] R, A
i
: [a, b] R respectively, x

[a, b] and y

R
n
. The vector-valued ODE
dy
dx
= M(x)y(x) +A(x) ()
has a unique solution with y(x

) = y

valid for all x [a, b].


Note: The ODE is really a system of n coupled linear ODEs for real variables y
1
, y
2
, . . . , y
n
,
d
dx
_

_
y
1
y
2
.
.
.
y
n
_

_
=
_

_
M
11
(x) M
12
(x) M
1n
(x)
M
21
(x) M
22
(x) M
2n
(x)
.
.
.
.
.
.
.
.
.
M
n1
(x) M
n2
(x) M
nn
(x)
_

_
_

_
y
1
y
2
.
.
.
y
n
_

_
+
_

_
A
1
(x)
A
2
(x)
.
.
.
A
n
(x)
_

_
.
If we choose the matrix M(x) and the column vector A(x) to have the special form
M(x) =
_

_
0 1 0 0
0 0 1 0
0 0 0 0
.
.
.
.
.
.
.
.
.
.
.
.
a
0
(x) a
1
(x) a
2
(x) a
n1
(x)
_

_
_

_
y
1
y
2
y
3
.
.
.
y
n
_

_
, A(x) =
_

_
0
0
0
.
.
.
c(x)
_

_
57
then any solution of system () has y
2
= y

1
, y
3
= y

2
= y

1
etc., so that () is equivalent to the single
nth order linear ODE problem
d
n
z
dx
n
+a
n1
(x)
d
n1
z
dx
n1
+ +a
1
(x)
dz
dx
+a
0
(x)z = c(x), x [a, b], (z(x

), z

(x

), . . . , z
(n1)
(x

)) = y

for the scalar function z(x) = y


1
(x). So our theorem will cover higher order scalar ODEs as a special
case.
Proof of Theorem 177: First we need an upper bound on M. Let
M

= sup[M
ij
(x)[ : x [a, b], i, j = 1, 2, . . . , n.
M

exists since every M


ij
: [a, b] R is continuous and [a, b] is compact. Choose any N Z
+
suciently large so that T = (b a)/N <
1
2nM
, and split the interval [a, b] into N subintervals of
equal length T,
I
i
= [a +iT, a +iT +T], i = 0, 1, . . . , (N 1).
We will prove that the restriction of () to each interval I
i
has, for any choice of x
0
I
i
and y
0
R
n
,
a unique solution with y(x
0
) = y
0
.
So, for each i = 0, . . . , N1, and each x
0
I
i
, y
0
R
n
, dene the mapping
i,x0,y0
: C(I
i
, R
n
)
C(I
i
, R
n
) by

i,x0,y0
(f) =
_
x y
0
+
_
x
x0
(M(t)f(t) +A(t))dt
_
Note that y : I
i
R
n
solves () on I
i
with initial data y(x
0
) = y
0
if and only if it is a xed
point of
i,x0,y0
. I claim that every such (for short) is a contraction mapping. Check: for each
i = 1, 2, . . . , n
[(f)(x)
i
, (g)(x)
i
[ =

_
x
x0
[M(t)(f(t) g(t)]
i
dt

_
x
x0
[[M(t)(f(t) g(t)]
i
[ dt

_
x
x0

j
M
ij
(t)(f
j
(t) g
j
(t))

dt

_
x
x0
nM

(f, g) dt

= nM

(f, g)[x x
0
[
Hence
((f)(x), (g)(x)) = max[(f)(x)
i
, (g)(x)
i
[ : i = 1, 2, . . . , n nM

(f, g)[x x
0
[
so
d((f), (g)) supnM

(f, g)[x x
0
[ : x I nM

(f, g)T
1
2
(f, g).
Hence, by the Contraction Mapping Theorem, a xed point exists, and is unique.
Now, (at least) one of the intervals, I
j
say, contains x

, so the above proves existence and


uniqueness of the required initial value problem on the interval I
j
(take x
0
= x

and y
0
= y

). Call
this solution f
j
. We can now use the same argument on I
j+1
with x
0
= a+(j +1)T and y
0
= f
j
(x
0
)
to extend the solution (uniquely) to I
j
I
j+1
. Note the extension is continuous by the Glue Rule.
58
x
y
a b I
1
I
2
I
3
I
4
y
y
y
y
(0)
(1)
(2)
(3)
Repeating this process a nite number of times gives us a solution to () on [a +jT, b]. We extend
to the left similarly: on I
j1
, let x
0
= a +jT and y
0
= f
j
(x
0
). The argument above shows existence
and uniqueness of an extension of the solution to I
j1
[a +jT, b]. Repeating this process a nite
number of times gives us a solution to () on [a, b]. If the solution constructed were not unique, its
restriction to (at least) one of the intervals I
i
would not be unique, so that one of the mappings

i,x0,y0
would have more than one xed point, a contradiction.
Corollary 178 Consider the ODE () where M and A are continuous on some interval J. Given
any x

J and y

R
n
there is a unique solution y : J R
n
of () with y(x

) = y

.
Proof: For each x J, the interval I = [x

, x] (if x x

, or I = [x, x

] otherwise) is a subset of J
so M, A are continuous on I, and x

I. Hence, by Theorem 177 there is a unique solution f


x
of
() on I with f
x
(x

) = y

. Let g(x) = f
x
(x). This denes a mapping g : J R
n
satisfying () with
g(x

) = y

. Hence a solution to the initial value problem on J exists. If it were not unique, then its
restriction to some interval [a, b] J would not be unique, contradicting Theorem 177.
So linear ODEs with continuous coecients have global solutions, uniquely determined by their
initial data. The same is not true of nonlinear ODEs.
Counterexample 179 Consider the scalar ODE
dy
dx
= y
2
.
Note that this is nonlinear and that the right hand side is a continuous function of x (its constant
in x!). Lets nd the solution with y(0) = 1. Whilever y(x) ,= 0,
1
y
2
dy
dx
=
d
dx
1
y
= 1

1
y(0)

1
y(x)
= x
y(x) =
1
1 x
.
This solution is unique and is clearly only well dened on (, 1), not the whole of R.
This is a generic feature of nonlinear ODEs. Solutions exist locally, but not necessarily globally:
the solution can grow unbounded in nite time (thinking of x as a time variable). We can still
use Picards Method to prove existence and uniqueness of local solutions of nonlinear ODEs. Once
again, we consider systems of coupled rst order ODEs, since these include the case of higher order
scalar ODEs.
59
Theorem 180 Let F : [a, b] R
n
R
n
be continuously dierentiable and w R
n
. Then there
exists b

(a, b] such that the ODE


dy
dx
= F(x, y) ()
has a unique solution with y(a) = w, valid for all x [a, b

].
Proof: Choose and x > 0, and let Y = [w
1
, w
1
+] [w
2
, w
2
+] [w
n
, w
n
+] =
B

(w). Let G
ij
: [a, b] R
n
R be the partial derivative of F
i
with respect to y
j
,
G
ij
(x, y) =
F
i
y
j

(x,y)
.
Note that F
i
and G
ij
, i, j = 1, 2, . . . , n, are continuous, and hence are bounded on the compact
set [a, b] Y , [F
i
(x, y)[ F

and [G
ij
(x, y)[ G

for all i, j, say. It follows that for each xed x,


F(x, ) : Y R
n
is globally Lipschitz with respect to the max metric . Check: let y, z Y . Then
for each i, there exists, by the Mean Value Theorem, q Y , such that
[F
i
(x, y) F
i
(x, z)[ [F
i
(x, y
1
, y
2
, . . . , y
n
) F
i
(x, z
1
, y
2
, . . . , y
n
)[
+[F
i
(x, z
1
, y
2
, . . . , y
n
) F
i
(x, z
1
, z
2
, . . . , y
n
)[ +
+[F
i
(x, z
1
, z
2
, . . . , z
n1
, y
n
) F
i
(x, z
1
, z
2
, . . . , z
n1
, z
n
)[
= [G
i1
(x, q
1
, y
2
, . . . , y
n
)[[y
1
z
1
[ +[G
i2
(x, z
1
, q
2
, y
3
, . . . , y
n
)[[y
2
z
2
[ +
+[G
in
(x, z
1
, z
2
, . . . , q
n
)[[y
n
z
n
[
G

[[y
1
z
1
[ +[y
2
z
2
[ + +[y
n
z
n
[] nG

(y, z).
Hence (F(x, y), F(x, z)) = max[F
i
(x, y) F
i
(x, z)[ : i = 1, 2, . . . , n nG

(y, z).
Now, choose b

(a, b] such that


b

a <

F

and b

a
1
2nG

.
Let C = C([a, b

], Y ). Note that Y is compact, hence complete, so C is complete with respect to


the sup metric.
Now consider the mapping
: C C([a, b

], R
n
), (f) =
_
x w +
_
x
a
F(t, f(t)) dt
_
.
I claim that (C) C. Clearly (f) is continuous, so we must check that (f)(x) Y for all
x [a, b

]. We will show the stronger condition that (f)(x) Y

= B

(w) for all x [a, b

]. Let
f C, so (f(t), w) for all t [a, b

]. Then for all x [a, b

]
[(f)(x)
i
w
i
[ =

_
x
a
F
i
(t, f(t)) dt


_
x
a
[F
i
(t, f(t))
i
[ dt

_
x
a
F

dt = (x a)F

(b

a)F

< .
Hence for all x [a, b

]
((f)(x), w) = max[(f)(x)
i
w
i
[ : i = 1, 2, . . . , n < ()
so (f)(x) B

(w), and the claim is established.


I claim further that : C C is a contraction mapping. For all f, g C and x [a, b

], and
each i = 1, 2, . . . , n
[(f)(x)
i
(g)(x)
i
[
_
x
a
[F
i
(t, f(t)) F
i
(t, g(t))[ dt
_
x
a
nG

(f(t), g(t)) dt

_
x
a
nG

d(f, g) dt = nG

(x a)d(f, g)
60
Hence
((f)(x), (g)(x)) = max[(f)(x)
i
(g)(x)
i
[ : i = 1, 2, . . . , n nG

(x a)d(f, g)
so
d((f), (g)) = sup((f)(x), (g)(x)) : x [a, b

] nG

(b

a)d(f, g)
1
2
d(f, g).
This establishes the claim.
Hence, by the Contraction Mapping Theorem, has a unique xed point f, and y = f(x) is
clearly a solution of the ODE on [a, b

] satisfying y(a) = w.
Assume that the ODE has another solution g(x), dierent from f(x), but valid on [a, b

] and
with g(a) = w. Since the xed point of is unique, g must lie outside C. But g is manifestly
continuous (its dierentiable) so it must be that g(x) strays outside Y . Since g(a) Y , there must,
by continuity, be a rst x, call it b

, for which g(b

) Y , that is, (g(b

), w) = . Note that f and


g coincide on [a, b

] since they are both xed points of the contraction mapping

: C([a, b

], R
n
) C([a, b

], R
n
),

(f) =
_
x w +
_
x
a
F(t, f(t)) dt
_
.
Hence (f(b

), w) = . But f = (f) and ((f)(b

), w) < by (), a contradiction.


Heres a subtle point. What we showed is that, given an initial value w, we can integrate the
ODE forward in time for a time interval of
x = min/2F

, 1/2nG

say. Why cant we just repeat the argument and continue the solution for another interval of width
x? Why not keep going until we exhaust [a, b], and thereby get a global solution? The answer is
that F

, G

, and hence x, depend on the initial value w. Recall that F

, G

are upper bounds for


[F
i
[, [G
ij
[ on [a, b] Y , and Y is a (closed) ball centred on the initial value of the solution. If we
repeat the argument, w is replaced by the end value of our initial solution, f(b

) so Y is replaced
by Y

= B

(f(b

)) and it may happen that the maximum of [F


i
[ or [G
ij
[ on [a, b] Y

is larger than
F

or G

, in which case x

may be smaller than x. In fact, repeating the process, we may nd


that the xs tend to 0 very fast, so that our solution never reaches x = b. This is what happens in
Counterexample 179 (with [a, b] = [0, 37] say).
61
9.5 A space-lling curve
In this section, I will always denote the interval [0, 1]. What, to a topologist, is a curve in a
topological space? The obvious denition is:
Denition 181 A curve in a topological space X is a continuous map f : I X.
We can think of f(0) and f(1) as the start and end points of the curve. Heres an example of a
curve in R
2
(with the usual topology):
f : I R
2
, f(t) = ([2t 1[, t
2
).
It starts at f(0) = (1, 0) and ends at f(1) = (1, 1). Heres a picture of f(I):
x
1
x
2
f(0)
f(1)
Its intuitively clear that the image f(I) of any curve in R
2
(usual topology) must be one-dimensional.
In particular, its clear that no continuous map f : I I I can be surjective. The image of such
a curve would completely ll the unit square in [0, 1]. Like many things which are intuitively clear,
this turns out to be false!
Theorem 182 There exists a surjective continuous map f : I I I.
Such a map is called, for obvious reasons, a space-lling curve. Our aim in this section is to
prove this theorem. The metric space of interest is
C = C(I, I I), d(f, g) = supd
1
(f(t), g(t)) : t I
where d
1
(x, y) = [x
1
y
1
[ +[x
2
y
2
[. Note that d
1
induces the usual topology on R
2
, so (I I, d
1
)
is compact, hence complete, so (C, d) is complete by Theorem 166.
Our strategy is to dene a certain sequence f
n
C, show that it is Cauchy, and hence converges
to some f C. We will then show that f(I) is dense in I I, and is closed. It follows that
f(I) = I I. The sequence will be dened iteratively.
First, note that given any nite ordered list of points
(x
0
, y
0
), (x
1
, y
1
), . . . , (x
k
, y
k
) I I,
there is a continuous map g : I I I whose image is the union of the k line segments joining
consecutive points in the list. To see this, rst split I into k equal subintervals
I
1
= [0,
1
k
], I
2
= [
1
k
,
2
k
], . . . , I
j
= [
j 1
k
,
j
k
], . . . , I
k
= [
k 1
k
, 1].
Now for each j dene g
j
: I
j
I I by
g
j
(t) = (x
j1
+k(t
j 1
k
)(x
j
x
j1
), y
j1
+k(t
j 1
k
)(y
j
y
j1
)).
62
Note that g
j
is manifestly continuous,
g
j
(
j 1
k
) = (x
j1
, y
j1
) and g
j
(
j
k
) = (x
j
, y
j
),
so g
j
(I
j
) is the straight line segment from (x
j1
, y
j1
) to (x
j
, y
j
). Hence, by the Glue Lemma,
g : I I I, g(t) = g
j
(t) t I
j
is well-dened and continuous, and g(I) is the union of line segments claimed.
Now let f
0
be the continuous function associated with the list
f
0
(0, 0), (
1
2
,
1
2
), (0, 1).
Clearly f
0
(I) is the triangular path shown below left.
f
0
f
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1


0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
To obtain f
1
we replace this triangle by a sequence of 4 triangles joined together, as depicted
above right. So f
1
is the continuous function associated with the list
f
1
= (0, 0), (
1
4
,
1
4
), (0,
1
2
), (
1
4
,
3
4
), (
1
2
,
1
2
), (
3
4
,
3
4
), (1,
1
2
), (
3
4
,
1
4
), (0, 1).
We can think of this as a path consisting of 4 triangles.
Now to obtain f
2
we perform the same operation to each of the 4 triangles in f
1
. The rule by
which a triangle is replaced by 4 triangles is depicted below. Its not hard to write down the ordered
list of points dening f
2
(try it!). Note that f
2
(0) = (0, 0) and f
2
(1) = (1, 0). So f
2
is a path
consisting of 16 triangles. We apply the rule again, replacing each by a set of 4, to obtained f
3
.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1


0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1


0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1


0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1


0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Repeating this construction indenitely we obtain, for each n N a continuous map f
n
: I II
whose image is a path consisting of 4
n
triangles. If we split I I into a grid of subsquares of side
length 2
n
then each constituent triangle is contained in one of these subsquares, and the path visits
every such subsquare at least once. Here are pictures of f
0
to f
7
:
63
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
I claim that (f
n
) is a Cauchy sequence in C. To see this, consider d(f
n
, f
n+1
). In passing from
f
n
to f
n+1
, we replace each triangle in f
n
(I), which lies in a square of side length 2
n
, by a set of
4 triangles, each of which lies in the same square. Hence for all t I,
d
1
(f
n
(t), f
n+1
(t)) =

[f
n
(t)]
1
[f
n+1
(t)]
1

[f
n
(t)]
2
[f
n+1
(t)]
2


1
2
n
+
1
2
n
=
1
2
n1
.
Hence
d(f
n
, f
n+1
) = supd
1
(f
n
(t), f
n+1
(t)) : t I
1
2
n1
.
So for all n, p Z
+
,
d(f
n
, f
n+p
) d(f
n
, f
n+1
) +d(f
n+1
, f
n+2
) + +d(f
n+p1
, f
n+p
)

1
2
n1
+
1
2
n
+ +
1
2
n+p2
=
1
2
n1
_
1 +
1
2
+
1
2
2
+ +
1
2
p1
_
<
1
2
n1

j=0
1
2
j
=
1
2
n1
1
1
1
2
=
1
2
n2
.
It follows immediately that (f
n
) is Cauchy (why?).
64
Since C is complete, (f
n
) is convergent, that is, f
n
f for some (unique) f C. This
f : I I I is continuous by denition. We will now prove that it is surjective, that is, f(I) = I I.
The strategy is to prove that every x I I is in the closure of f(I), and then to prove that f(I)
is closed (so coincides with its closure).
Let x be any point in I I, and be any positive number. I claim that there exists N Z
+
such that for all n N, f
n
(I) B

(x) ,= . Consider again the square I I split into a grid of


subsquares, each of sidelength 2
n
. The point x is in (at least) one of these subsquares. If we choose
n N where 2
N
<

2
, then the subsquare containing x is contained in B

(x). But recall that f


n
visits every subsquare in the grid, so f
n
(I) B

(x) ,= .
Now let U I I be any open set containing x. There exists > 0 such that B

(x) U. Since
d(f
n
, f) 0, there exists N

Z
+
such that d(f
n
, f) <

2
for all n N

. Also, by the previous


paragraph, there exists N

Z
+
such that f
n
(I)B

2
(x) ,= for all n N

. Let N = maxN

, N

.
Then there exists t
0
I such that d
1
(x, f
N
(t
0
)) <

2
, and hence
d
1
(x, f(t
0
)) d
1
(x, f
N
(t
0
)) +d
1
(f
N
(t
0
), f(t
0
)) <

2
+d(f
N
, f) <

2
+

2
= .
so f(t
0
) B

(x) U. Hence every open set containing x meets f(I), so x f(I) (Theorem 108).
This holds for all x I I, so f(I) = I I.
Now I is compact and f is continuous, so f(I) is a compact subset of I I (Theorem 84). But
I I is Hausdor, so every compact subset of I I is closed (Theorem 92). Hence f(I) is closed,
that is f(I) = f(I) = I I. So f is surjective, as was to be proved.
65
Advice on exam preparation
This is a 15 credit level 3 module in pure mathematics. Hence
PROOFS ARE EXAMINABLE!
In fact, with a few exceptions which I will list below, all the theorems, propositions, lemmas,
corollaries in this course are examinable. Most of the proofs are, if you understand the basic
denitions, quite straightforward. Just ask yourself:
(a). What do I know (i.e. what do the hypotheses tell me)?
(b). What am I supposed to deduce?
If you can sort that out, a sensible proof strategy usually suggests itself. Of course, there may
be a clever idea required, and its a good idea to learn (memorize) what the clever idea is. It is
probably not a good idea to try to learn the proofs verbatim without really understanding them. Its
extremely unlikely that you will get down on paper exactly what I wrote in the notes, and youre
quite likely to write complete nonsense.
The scope of the paper will be similar to previous years. I have put slightly more emphasis on
proving standard results than my predecessor. As always, a substantial portion of the marks can be
gained by accurately recalling denitions. Of course, if you dont know the denitions, the theorems,
examples etc. are pretty meaningless and youre very unlikely to pass the exam. So my number one
piece of advice is
LEARN ALL THE DEFINITIONS!
A few of the proofs in the course are too intricate to be examinable. Heres a complete list of
them:
Theorem 37.
The if part of Theorem 59. I could legitimately ask you to prove that every connected
subset of R (usual topology) is an interval, but it would be unreasonable to ask you to prove
the converse, that every interval is connected.
Theorem 82 (the Heine-Borel Theorem for R).
Theorem 87 (Tychonos Theorem).
Lemma 146 (Lebesgues Covering Lemma).
Lemma 147.
Theorem 148.
Theorem 166.
Theorem 177.
Theorem 180.
Theorem 182.
Note that, although Theorems 177 and 180, which we proved using Picards method, are not exam-
inable, Picards method itself certainly is examinable. It would be perfectly reasonable for me to
set a question similar to questions 6 or 7 from Problem Set 5, for example.
66

Anda mungkin juga menyukai