Notes on Topology for Functional Analysis

Eric A. Carlen1
Rutgers University

January 30, 2013

1 Introduction
What is it that one analyzes in functional analysis? Very often the analysis involves functions
defined on a domain in an infinite dimensional vector space with values in the field, R or C, over
which the vectors space is defined. Since the elements of these infinite dimensional vector spaces
are often functions themselves; e.g., the vectors space may consist of square integrable functions on
[0, 1], functions on such vectors spaces are called functionals. More generally, a functional is simply
a scalar valued function defined on a set of functions on which there is some linear structure.
Many problems in mathematics can be treated effectively by relating them to variational prob-
lems; i.e., problems of finding maximizers and minimizers of functions – or functionals. The Spectral
Theorem is a simple but important example. Many other problems can be treated effectively by
some sort of method of successive approximation. The development of both strategies requires
a through exploration of the notions of continuity and compactness, and their interactions with
convexity, completeness, separability, and duality that we shall develop in this course. We begin by
recalling some basic theorems on continuity and compactness.

1.1 Approximation
A basic strategy in functional analysis is approximation. In particular, one often tries to approx-
imate “general” elements of some infinite dimensional vector space by “nice” elements from some
well understood vector space, possibly even finite dimensional.
For example, consider the vector space C([0, 1]) consisting of continuous real valued functions
on [0, 1]. The Weirstrass Approximation Theorem says that for any f ∈ C([0, 1]), and any  > 0,
there is a polynomial p such that

sup |f (t) − p(t)| ≤  .

0≤t≤1

The quantity sup0≤t≤1 |f (t) − p(t)| is a measure of the distance between f and p in C([0, 1]).
That is, the function duniform on C([0, 1]) × C([0, 1]) defined by

duniform (f, g) = sup |f (t) − g(t)|

0≤t≤1

is a metric on C([0, 1]), and equipped with this metric, C([0, 1]) is a metric space.
1
c 2012 by the author. This article may be reproduced, in its entirety, for non-commercial purposes.

1
EAC January 30, 2013 2

According to the Wierstrass Approximation Theorem, given any f ∈ C([0, 1]), for each n ∈ N,
we can find a polynomial pn such that duniform (f, pn ) ≤ 1/n, so that
lim duniform (f, pn ) = 0 ,
n→∞

which we express by writing

f = lim pn ,
n→∞
or by saying that the sequence {pn }n∈N converges (uniformly) to f .
Many of the central questions in real analysis have to do with convergence, and especially with
two closely related concepts, continuity and compactness.
A basic theorem from elementary topology says that if K is a closed and bounded subset of Rn ,
and if f is a continuous real valued function on K, then there exist an x ∈ K such that
f (x) ≥ f (y) for all y ∈ K .
Such an x is called a maximizer of f .
In recalling this statement, we did not specify the topology with respect to which “closed” and
“continuous” are defined, nor the metric with respect to which “bounded” is defined. There is only
one topology on Rn with respect to which the algebraic operations on it are continuous, which is its
usual metric topology, defined in terms of the Euclidean distance. In infinite dimensions, this not
the case, and we shall need a variety of topologies for many problems, as we shall explain below.
Moreover, not all of these topologies are metrizable. That is, we shall deal with topologies that
cannot be induced by any metric.
Still, metric topologies are central to our subject, and are extremely convenient to work with,
largely because questions of continuity and compactness can both be reduced to questions about
sequences in the metric space setting.

1.2 Continuity in metric spaces

A function f from X to Y is continuous if a sufficiently small change in the input results in a small
change in the output. This is most conveniently expresses in terms of the open metric balls in X:
1.1 DEFINITION (Open ball in metric spaces). Let X be a metric space with metric d. Given
a number r > 0, and a point x ∈ X, let Br (x) be defined by
Br (x) = { y ∈ X : d(y, x) < r } .
This set is called the open ball of radius r about x.
Another piece of notation will be convenient: If X and Y are sets, and f : X → Y , we write
f (A) to denote the image of A ⊂ X under f , and we write f −1 (B) to denote the pre-image of
B ⊂ Y . We now define continuity in the metric space setting:

1.2 DEFINITION (Continuous functions from one metric space to another). Let (X, dX ) and
(Y, dY ) be two metric spaces. Let f be a function from X to Y . Then f is continuous at x0 ∈ X
in case for every  > 0, there is a δ > 0 such that
f (Bδ (x0 )) ⊂ B (f (x0 )) . (1.1)
The function f is continuous in case it is continuous at each x0 ∈ X.
EAC January 30, 2013 3

1.3 DEFINITION (Convergent sequences in a metric space). Let (X, dX ) be a metric spaces.
Then a sequence {xn } in X converges to x0 ∈ X if and only if for every  > 0, B (x0 ) contains all
but finitely many terms of the sequence {xn }. We express this by writing limn→∞ xn = x0 .

1.4 THEOREM (Continuity and sequences). Let (X, dX ) and (Y, dY ) be two metric spaces. Let
f be a function from X to Y . Then f is continuous at x0 ∈ X if and only if for every sequence
{xk } in X
lim xk = x0 ⇒ lim f (xk ) = f (x0 ) . (1.2)
k→∞ k→∞

Proof. Suppose that f is continuous, and that limk→∞ xk = x0 . Pick any  > 0. Since f is
continuous, there exists a δ > 0 so that f (Bδ (x0 )) ⊂ B (f (x0 )). Since limk→∞ xk = x0 , all but
finitely many terms of {xk } lie inside Bδ (x0 ). Hence all but finitely many terms of {f (xk )} lie
inside f (Bδ (x0 )) ⊂ B (f (x0 )), and this proves that limk→∞ f (xk ) = f (x0 ).
Next suppose that f is not continuous at x0 . Then there exists some  > 0 so that for each
positive integer k,
f (B1/k (x0 )) 6⊂ B (f (x0 )) .
Define a sequence {xk } by choosing
\
xk ∈ B1/k (x0 ) (f −1 (B (f (x0 ))))c 6= ∅ .

Then {xk } converges to x0 , but {f (xk )} does nor converge to f (x0 ).

There is another characterization of continuity involving the notion of open sets, which we now
define:

1.5 DEFINITION (Open sets in metric spaces). Let X be a metric space with metric d. A subset
U of X is open in case either:
(1) It is the empty set ∅, or else
(2) For each x ∈ U , there is an r > 0, depending on x, such that

Br (x) ⊂ U .

It is possible, and as we shall see, useful to characterize the (global) continuity of functions
between two metric spaces simply in terms of open sets, without explicit reference to the specific
metrics themselves.

1.6 THEOREM (Continuity and open sets). Let X and Y be metric spaces with metrics dX and
dY resepctively. Let f be a function from X to Y . Then f is continuous if and only if for every
open set U in Y , f −1 (U ) is open in X.

Proof. Suppose that f is continuous, and let U be an open set in Y . If f −1 (U ) = ∅, then f −1 (U )

is open by (1). Otherwise, if f −1 (U ) 6= ∅, consider any x ∈ f −1 (U ). Then f (x) ∈ U , and since U
is open, there exists a  > 0 such that B (f (x)) ⊂ U . Then, since f is continuous at x0 , there is a
δ > 0 so that f (Bδ (x)) ⊂ B (f (x)) ⊂ U . This means

Bδ (x) ⊂ f −1 (U ) ,
EAC January 30, 2013 4

and since x is an arbitrary element of f −1 (U ), U is open.

Conversely, suppose that f has the property that whenever U is open in Y , f −1 (U ) is open in
X. Fix any x ∈ X and any  > 0. f −1 (B (f (x))) is open and contains x. Therefore, there is some
δ > 0 such that
Bδ (x) ⊂ f −1 (B (f (x))) .
Since  is arbitrary, f is continuous at x. Since x is arbitrary, f is continuous.

1.3 Topological spaces

Since we can characterize continuous functions in terms of open sets, without explicitly mentioning
a metric at all, we can “strip away” the metric structure, and deal directly with open sets. This
will turn out to be useful.

1.7 DEFINITION (Topological Spaces). Let X be any set, and let O be any collection of sets
in X satisfying:
(1) The empty set ∅ belongs to O, as does X itself.
(2) The union over any arbitray set of sets in O belongs to O.
(3) The intersection over any finite set of sets in O belongs to O.
In this case, O is said to be a topology on X, and the sets belonging to O are called open sets
in X (for the topology in question). A subset A of X is closed in case its complement, Ac is open.
The pair (X, O) is said to be a topological space.

Note that by De Morgans laws, the intersection of any arbitrary set of closed sets in X is itself
closed.
The next definitions introduces some more useful terminology

1.8 DEFINITION (Interior, closure and neighborhoods). Let (X, O) be a topological space, and
A a subset of X.The interior of A, Ao , is the union of all of the open sets contained in A. The
closure of A, A, is the intersection of all of the closed sets containing A. Finally for any x ∈ X,
the set Nx of neighborhoods of x consists of all sets B such that x ∈ B o .

1.9 DEFINITION (Hausdorff, normal). A topological space (X, O) is called Hausdorff if for any
two distinct elements x, y ∈ X, there are disjoint open sets sets U and V with x ∈ U and y ∈ V .
It is called normal if for each x ∈ X, the singleton {x} is closed, and moreover, for any two
disjoint closed sets A and B, there are disjoint open sets U and V with A ⊂ U and B ⊂ V .

It is easy to see, and left as an exercise, that if X is any metric space, and O is the collection
of all open sets in X, as defined above in terms of open balls, O does indeed constitute a topology
on X. Thus by Theorem 1.6, the definition of continuity that we make next is consistent with our
existing notion of continuity in the metric space setting.

1.10 DEFINITION (Continuous functions between topological spaces). Let (X, OX ) and (Y, OY )
be two topological spaces. A function f from X to Y is continuous at x ∈ X in case for every
neighborhood V of f (x), there is a neighborhood U of x such that f (U ) ⊂ V .
A function f from X to Y is continuous whenver U is open in Y , f −1 (U ) is open in X.
EAC January 30, 2013 5

It is easy to see that f : X → Y is continuous if and only if it is continuous at each x ∈ X.

We are almost ready to get back to proving theorems. The next theorem we prove concerns one
of our core issues: approximation. Given a topological space (X, OX ), and a subset A of X, what
does it mean to “approximate” x ∈ X\A by elements of A? We shall take it to mean that every
neighborhood U of x contains points in A:

1.11 DEFINITION (Limit points in a topological space). Let (X, OX ) be a topological space.
If A is any set in X, a point x ∈ X is a limit point of A in case every for every open set U that
contains x,
A ∩ U 6= ∅ .

We must be careful to distinguish this notion of limit point of a set from the limit of a sequence.
In particular: Let {xk } be a sequence of elements of X. Recall that {xk } is convergent to x in
case every neighborhood U of x also contains all but finitely many terms in the sequence {xk }. On
the other hand, x is a limit point of {xk } in case every neighborhood U of x contains at least one
element of {xk }.
In a metric space, there is of course a characterization of limit points in terms of sequences;
x is a limit point of A if and only if there is a sequence {xn }n∈N of elements in A such that
limn→∞ xkn = x.
We are now ready for the theorem that justifies the terminology “closed”:

1.12 THEOREM (Closed sets and limit points). Let (X, O) be any topological space. A subset
A of X is closed if and only if A contains all of its limit points.

Proof. Suppose that A is closed, and x ∈ Ac . Since Ac is open, there is an open set U containing x
that has an empty intersection with A. Thus, x is not a limit point of A. Since x was an arbitrary
point outside A, A must contain all of its limit points.
On the other hand, suppose that A contains all of its limit points. We must show that A is
closed, or, what is the same thing, that Ac is open. Consider any point x ∈ Ac . Since it is not a
limit point of x, there is an open set Ux containing x that has empty intersection with A. For each
x ∈ Ac , chose such a Ux . But then, since for each x ∈ Ac , x ∈ Ux ⊂ Ac ,
[
Ac = Ux .
x∈Ac
[
Thus, by (2) in the definition of topological spaces, Ac = Ux is open.
x∈Ac

We close this subsection with one more definition:

1.13 DEFINITION (Density). Let (X, O) be a toplogical space. Let A ⊂ B ⊂ X. Then A is

dense in B in case the closure of A contains B.

By Theorem 1.12, A is dense in B if and only if every point in B is a limit point in A; i.e, if
every point in B can be approximated arbitrarily well by points in A.
EAC January 30, 2013 6

1.4 Compactness
1.14 DEFINITION (Compact Sets). Let (X, OX ) be a topological space. A subset K is called
compact in X in case for every collection U of open sets that covers K; i.e.,
\
K⊂ U ,
u∈U

there is a finite subset G of U that also covers K:

\
K⊂ U .
u∈G

G is called a finite subcover of A.

Our first example of a theorem involving compactness is the classical result known as Dini’s
Theorem. In proving this, we shall make use of the following fact: If U is any set of open subsets
of X, then by De Morgan’s laws, !c
[ \n
U = Uc ,
U ∈U U ∈U

Thus U is an open cover if and only if { Uc

: U ∈ U } is a set of closed subsets of X with empty
intersection.
Therefore, X is compact if and only if whenever K is a set of closed subsets of X such that
\
K=∅,
K∈K

there is a finite subset {K1 , . . . , Kn } ⊂ K such that

n
\
Kj = ∅ .
j=1

This analysis is often summarized by saying that X is compact if and only if X has the “finite
intersection property”.

1.15 THEOREM (Dini’s Theorem). Let X be a compact space, and let {fn }n∈N be a sequence
of real valued functions on X, and suppose that there is a continuous real valued function f on X
such that for each x ∈ X, the sequence {fn (x)}n∈N is monotone non-decreasing,

lim fn (x) = f (x) .

n→∞

Then
lim fn = f
n→∞
uniformly.

In other words, pointwise convergence, together with compactness and montonicity, imply uni-
form convergence. Also note that replacing each fn by −fn , one converts a monotone non-decreasing
sequence into a montone non-increasing sequence, and so the theorem remains true if one replaces
“monotone non-decreasing” by “monotone non-increasing”.
EAC January 30, 2013 7

Proof. Fix  > 0. Define the sets K` , ` ∈ N, by

K` := { x ∈ X : f (x) − f` (x) ≥  } = { x ∈ X : |f (x) − f` (x)| ≥  } .

Since f and f` are continuous, K` is closed. Also, since for each x, lim`→∞ f` (x) = f (x),
∞
\
K` = ∅.
k=1

Then, by the compactness of X, there is some n ∈ N such that

n
\
K` = ∅.
`=1

Hence, for all ` > n, and all x, |f` (x) − f (x)| < . Since the sequence is monotone non-decreasing,
it then follows that
m≥` ⇒ |fm (x) − f (x)| <  ,
and so d(fm , f ) <  for all m ≥ `.

Next, we turn to one of the main theorems on compactness.

1.16 THEOREM (Compactness, Continuity, and Minima). Let (X, O) be any topological space,
and let K be a compact subset of X. Let f be a functions from X to R that is continuous when R
is equipped with its usual metric topology. Then there exosts and x ∈ K so that

f (x) ≥ f (y) for all y∈K . (1.3)

Proof. Consider the open sets (−n, ∞) in R, since f is continuous,

U = { f −1 ((−n, ∞)) : n ∈ N }

is an open cover of X, and hence K. Since K is compact, there exists a finite subcover. But since
the sets in our cover are nested; i.e., since for n > m > 0, f −1 ((−m, ∞)) ⊂ f −1 ((−n, ∞)), a single
set in our cover suffices; i.e., there is an n with K ⊂ f −1 ((−n, ∞)). In particular, f is bounded
from below on K.
Now let a be the greatest lower bound of the numbers f (y) for y ∈ K. We claim that there
exists an x ∈ K with f (x) = a. If so, then plainly (1.3) is true.
To prove this, let us suppose that there is no such x. Then

U = { f −1 ((a + 1/n, ∞)) : n ∈ N }

is an open cover of K. This means that there is a finite subcover, and again, since the sets in the
open cover are nested, a single one of them, say f −1 ((a + 1/n, ∞)), covers K. But this would mean
that f (y) ≥ a + 1/n for each y in k, which is not possible since a is the greatest lower bound.
EAC January 30, 2013 8

Any point x for which (1.3) is true is called a maximizer of f on X. Likewise, any point x for
which
f (x) ≤ f (y) for all y∈K . (1.4)
is called a minimizer of f on X.
There are several important things to notice from this proof. First, if f is continuous, so is
−f , and a minimizer of −f is a maximizer of f . Hence, the theorem implies the existence of both
minimizers and maximizers for continuous functions on compact sets.
Now suppose we have a real valued function f defined on a set K, and we want to know if f has
a minimizer in K. If we can find a toplogy on K that makes f continuous, and makes K compact,
then we can apply the previous theorem.
However, the demands of continuity and compactness pull in opposite directions when we look
for our topology: The topology has to have sufficiently many open sets in it for f to be continuous,
since we need f −1 (U ) to be open for every open set U in R. On the other hand, the more open
sets we include in our topology, the more open covers we have to consider when showing that every
open cover has a finite subcover.
Very often, one is stuck between a rock and a hard place, and there is no topology that both
makes f continuous, and K compact. Indeed, there are many very nice functions f on R – such
as the identity function, f (x) = x – that simply do not have minimizers or maximizers. While R
is compact under the trivial topology O = {∅, R}, and while the identity function is continuous
under the usual metric topology on R, the fact that the identity function does not have either a
maximizer or a minimizer shows that there is no topology O on R under which R is compact and
the identity function is continuous from (R, O) to R equipped with its usual metric topology.
A situation that is frequently encountered in applications is that a function f on X does have,
say, a minimizer, but not a maximizer. Also in this situation, it is impossible to find a topology for
which f is continuous and X is compact, since them both minima and maxima would exist.
However, if we are only looking for minima, it is worth noticing that in our proof of Theo-
rem 1.16, we did not use the full strength of the continuity hypothesis. The same proof yields the
same conclusion if we assume only the property that f −1 ((t, ∞)) is open for each t in R.
1.17 DEFINITION (Upper and lower semicontinuous function). Let (X, OX ) be a topological
space. A function f from X to R is called lower semicintinuous in case for all t in R, f −1 ((t, ∞))
is open. It is called upper semicintinuous in case for all t in R, f −1 ((−∞, t)) is open.
Thus, we can prove existence of minimizers for f on X by finding a topology that makes f lower
semicontinuous, and K compact. This turns out to be a very useful strategy, as we shall see.
Still, to use either Theorem 1.15 or Theorem 1.16, we need criteria for compactness. How can
we tell if a set X is compact? In metric spaces, we can reduce this to a question about sequences.
We first make three definitions, two for metric spaces, and one for topological spaces.
1.18 DEFINITION (Total boundedness). A set K in a metric space (X, d) is totally bounded in
case for every  > 0, there exists a finite set {x1 , . . . , xn } ⊂ K such that
n
[
K⊂ B (xj ) ;
j=1

i.e., for each  > 0, K has a finite cover by open balls of radius .
EAC January 30, 2013 9

1.19 DEFINITION (Cauchy sequences and completeness). Let (X, d) be a metric space. A
sequence {xk }k∈N is a Cauchy sequence in case for each  > 0, there is some m so that B (xm )
contains all but finitely many terms in the sequence.
A set K ⊂ D is complete in case whenever {xk }k∈N is Cauchy, then {xk }k∈N converges to an
element x0 ∈ K; i.e., limk→∞ xk = x0 .

1.20 DEFINITION (totoal boundedness). A set K in a metric space (X, d) is totally bounded in
case for every  > 0, there exists a finite set {x1 , . . . , xn } ⊂ K such that
n
[
K⊂ B (xj ) ;
j=1

i.e., for each  > 0, K has a finite cover by open balls of radius .

1.21 DEFINITION (Sequential compactness). A set K in a topological space (X, O) is sequen-

tially compact in case every sequence {xn }n∈N in K has a convergent subsequence {xnk }k∈N that
converges to an elements of K.

1.22 THEOREM (Compactness in a Metric space). Let (X, d) be any metric space, and let K be
any subset of X. Then the following are equivalent:
(1) K is sequentially compact.
(2) K is totally bounded and complete.
(3) K is compact.

Proof of Theorem 1.21. We first show that (3) implies (2). Suppose K is compact. For each
 > 0, { B (x) : x ∈ K } is an open cover of K. Since K is compact, there exist a finite set
{x1 , . . . , xn } ⊂ K such that K ⊂ nj=1 B (xj ). Thus, K is totally bounded.
S

Next, let {xk }k∈N be a Cauchy sequence in K. Since {xn } is Cauchy, for each m we can fins a
ball B1/m (zm ) that contains all but finitely many of the terms in the sequence. For each m ∈ N,
define  c
[
Cm :=  B1/m (y) ,
d(y,zm )>2/m

and note that Cm is closed. If z ∈ B1/m (zm ), and d(y, zm ) > 2/m, then
2 1 1
d(z, y) ≥ d(zm , y) − d(zm , z) ≥ − = .
m m m
Thus, z ∈
/ B1/m (y). This is true for any y with d(y, zm ) > 2/m, z ∈ Cm ; i.e, B1/m (zm ) ⊂ Cm .
S
Next, since obviously z ∈ B1/m (z), if d(z, zm ) > 2/m, z ∈ d(y,zm )>2/m B1/m (y). That is,
(B2/m (zm ))c ⊂ Cm
c . Thus, C ⊂ B
m 2/m (zm ). Altogether, we have:

B1/m (zm ) ⊂ Cm ⊂ B2/m (zm ) . (1.5)

n
\
Then, for any finite n, Cm 6= ∅ since each Cm contains all but finitely many terms in {xn }n∈N .
m>1
∞
\
Therefore, by compactness, and the finite intersection property in particular, Cm 6= ∅. Chose
m>1
EAC January 30, 2013 10

x0 in this infinite intersection. By (1.5), limn→∞ xn = x0 , and x0 ∈ K. This shows that every
Cauchy sequence in K converges to a point in K, and thus, K is complete.
We now show that (2) implies (1). Let {xn }n∈N be any sequence in K. For each m ∈ N, choose
a finite 1/m cover of K by open ball of radius 1/m. By the pigeonhole principle at least one of these
contains an infinite subsequence. Using Cantor’s diagonal construction as above, only more simply,
(m)
we construct a subsequence {xm }m∈N such that for each r > 0, all but finitely many terms lie in
(m)
a ball of radius r. Thus, {xm }m∈N is a Cauchy sequence, and since K is complete, it converges to
some x0 ∈ K. This shows that K is sequentially complete.
Finally, we show that (1) implies (3) This part of the proof is more complicated, and we proceed
in four steps.
Step 1: K is bounded: We first show that K is bonded, which means that

sup d(x, y) < ∞ .

x,y∈K

This supremum is called the diameter of K.

To see that the diameter is finite, suppose that it is not. Under this hypothesis, we construct a
sequence {xn }n∈N as follows. First, fix any x ∈ X. Now for each n ∈ N , choose some xn ∈ K\Bn (x).
The set K\Bn (x) is not empty when the diameter of K is infinite.
Then, by hypothesis, there is a subsequence {xnk }k∈N and some y ∈ K such that

lim xnk = y . (1.6)

k→∞

Then by the triangle inequality, we would have

d(x, xnk ) ≤ d(x, y) + d(y, xnk ) .

But this cannot be: By construction, d(x, xnk ) > nk , while d(x, y) is some fixed, finite number,
and for all sufficiently large k, d(y, xnk ) ≤ 1, by (1.6). This contradiction shows that K must be
bounded.
Step 2: K contains a dense sequence: We next show that there is a sequence {xn }n∈N that is dense
in K; i.e., that for every  > 0, and every x ∈ K, there is some n such that d(xn , x) < .
In other words, the sequence {xn }n∈N passes arbitrarily close to every point in K. Here is how
to construct it:
Pick the first term x1 arbitrarily. We then define the rest of the sequence recursively as follows:
Suppose that {x1 , . . . , xk } have been chosen. For each y ∈ K, define

dk (y) := min {d(y, xj )} .

1≤j≤k

This is, by definition, the distance from y to the set {x1 , . . . , xk } ⊂ K, and of course, this is no
greater than the diameter of K, which is finite by the first step.
Therefore, dk , defined by
dk := sup dk (y)
y∈K

is no greater than the diameter of K.

EAC January 30, 2013 11

Armed with this knowledge, we are ready to choose xk+1 : We choose xk+1 to be any element
of K with
1
dk (xk+1 ) ≥ dk .
2
We now claim that limk→∞ dk = 0. It should be clear that {xn }n∈N is dense if and only if this
is the case. So, to complete Step 2, we need to prove that limk→∞ dk = 0.
Towards this end, the first thing to observe is that {dk }k∈N is a monotone decreasing sequence,
bounded below by zero: Indeed, for any sets A ⊂ B ⊂ K, the distance from y to B is no greater
than the distance from y to A. Therefore, we only have to show that some subsequence of {dk }k∈N
converges to zero.
To do this, let {xkn }n∈N be a convergent subsequence of {xk }k∈N , and let y be the limit; i.e.,

lim xkn = y .
k→∞

Then by the triangle inequality d(xkn , xkn+1 ) ≤ d(xkn , y) + d(y, xkn+1 ), and since

lim d(xkn , y) = lim d(y, xkn+1 ) = 0 ,

n→∞ n→∞

we have
lim d(xkn , xkn+1 ) = 0 .
k→∞

But since
xkn ∈ {x1 , . . . , xkn+1 −1 } ,
1
d(xkn , xkn+1 ) ≥ dkn+1 −1 (xkn+1 ) ≥ dkn+1 −1 .
2
Therefore,
lim dkn+1 −1 = 0 ,
n→∞

and then, since the entire sequence is monotone decreasing, limk→∞ dk = 0. Hence, the sequence
we have constructed is dense.
Step 3: Given any open cover of K, there exists a countable subcover. To prove this, consider any
open cover G of K. Consider the set of open balls Br (xk ) where r > 0 is rational, and {xk }k∈N
is the dense sequence that we have constructed in Step 2. This set of balls is countable since a
countable union of countable sets is countable.
The countable subcover is constructed as follows: For each rational r > 0 and each k ∈ N,
choose Ur,k to be some open set in G that contains Br (xk ) if there is such a set, and otherwise,
do not define Ur,k . Let U be the set of open sets defined in this way; clearly U is countable by
construction.
We now claim that U is an open cover of K. Clearly the sets in U are open. To see that they
cover, pick any x ∈ K. Since G is an open cover of K, x ∈ V for some V ∈ G. Then, since V is
open, for some rational r > 0, B2r (x) ⊂ V .
Then, since {xk }k∈N is dense, there is some k with xk ∈ Br (x). But then x ∈ Br (xk ) and

Br (xk ) ⊂ B2r (x) ⊂ V ,

EAC January 30, 2013 12

(where the first containment holds by the triangle inequality). This shows that for the pair (r, k),
there is some V ∈ G containing Br (xk ). Therefore, by construction, Ur,k ∈ U contains Br (xk ), and
hence x ∈ Ur,k . Since x is an arbitrary element of K, U covers K.
Step 4: Some finite subcover of the countable cover is a cover. Order the sets in our countable
cover U into a sequence of open sets {Uk }k∈N that covers K.
Suppose that for each n, it is not the case that
n
[
K⊂ Uk . (1.7)
k=1

Then we can construct a sequence {xn }n∈N by choosing

n
!
[
xn ∈ K\ Uk .
k=1

Let {xnj }j∈N be a convergent subsequence with limj→∞ xnj = y ∈ K. Then, since U is an open
cover of K, there is some Uk with y ∈ Uk . But then all but finitely many terms of the sequence
{xnj }j∈N lie in Uk , and so the whole sequence lies in some finite union of the sets in U. This is a
contradiction, and so (1.7) is true for some n ∈ N.

In the broader setting of topological spaces, there is no relation between compactness and se-
quential compactness. There are topological spaces that are compact, but not sequentially compact,
and there are sequentially compact spaces that are not compact: This theorem does not extend to
the general setting of topological spaces; it is important that the topology be a metric topology.
Likewise, in the general topological setting, it is not true that a function f is continuous if and only
if it takes convergent sequences to convergent sequences.
The notion of compactness as we have defined it in terms of open covers is a 20th century notion.
In the 19th century, mathematicians thought about compactness issues in terms of sequential
compactness.

1.5 The Arzelà-Ascoli Theorem

Let X be a compact topological space. Let C(X, R) denote the set of continuous real valued
functions on X. If f, g ∈ C(X, R), then x 7→ |f (x) − g(x)| also belongs to C(X, R), Therefore, by
Theorem 1.16, there exists an x0 ∈ X such that

|f (x0 ) − g(x0 )| ≥ |f (x) − g(x)| .

Define d(f, g) to be this maximum value of |f (x) − g(x)|. It is obvious that d(f, g) = d(g, f ) and
that d(f, g) = 0 if and only if f = g. Next, for f , g and h in C(X, R),

d(f, g) = |f (x0 ) − g(x0 )| ≤ |f (x0 ) − h(x0 )| + |h(x0 ) − g(x0 )|

≤ sup |f (x) − h(x)| + sup |h(x) − g(x)|
x∈X x∈X
= d(f, h) + d(h, g) .

Thus, d is a metric on C(X, R), known as the uniform metric. For brevity of notation, we shall
write C(X, R) to denote the metric space (C(X, R), d).
EAC January 30, 2013 13

1.23 THEOREM (Completeness of C(X, R)). C(X, R) is a complete metric space.

Proof. Suppose that {fn }n∈N is a Cauchy sequence in (C(X, R), d). For each x ∈ X,

|fn (x) − fm (x)| ≤ d(fn , fm ) ,

and so {fn (x)}n∈N is also a Cauchy sequence. Since R is complete, {fn (x)}n∈N is convergent. Define
f (x) to be the limit of this sequence. (Since R is Hausdorff, the limit is unique, and so f (x) is
well-defined.)
Now, fix  > 0, Since {fn }n∈N is a Cauchy sequence, there exists an N such that

n, m ≥ N ⇒ d(fn , fm ) ≤ .
2
For each x,
|fn (x) − f (x)| ≤ |fn (x) − fm (x)| + |fm (x) − f (x)| .
Now choose m so large that m ≥ N and |fm (x) − f (x)| < /2. Thus, for all all x ∈ X,

n ≥ N ⇒ |fn (x) − f (x)| <  .

That is, since N does not dependent on x, f is the uniform limit of the sequence {fn }.
We now claim that f ∈ C(X, R). For all x, y ∈ X,

|f (x) − f (y)| ≤ |f (x) − fn (x)| + |fn (x) − fn (y)| + |fn (y) − f (y)| .

Then, for n > N/3

2
|f (x) − f (y)| ≤  + |fn (x) − fn (y)| .
3
Since fn is continuous, there is a δ > 0 so that

d(y, x) < δ ⇒ |fn (x) − fn (y)| < .
3
Thus,
d(y, x) < δ ⇒ |f (x) − f (y)| <  .
Thus f is continuous a x, and since x ∈ X is arbitrary, f ∈ C(X, R), and have already seen that
limn→∞ d(fn , f ) = 0. We have shown that C(X, R) is complete.

We shall now characterize the compact sets in C(X, R).

1.24 DEFINITION (Equicontinuous, pointwise bounded). Let F ⊂ C(X, R). Then F is equicon-
tinuous in case for each  > 0 and each x ∈ X, there is a neighborhood U of x such that

y ∈ U ⇒ |f (y) − f (x)| <  for all f ∈ F .

Also, F is pointwise bounded in case for each x ∈ X, {f (x) : f ∈ F } is a bounded subset of R.

1.25 THEOREM (Arzelà-Ascoli). Let X be a compact topological space space, and let F be an
equicontinuous and pointwise bounded subset of C(X, R). Then the closure of F is compact if and
only if F is pointwise bounded and equicontinuous.
EAC January 30, 2013 14

Proof. Suppose that F is pointwise bounded and equicontinuous. Since C(X, R), the closure of F
is complete, and hence it suffices to show that F is totally bounded.
Fix  > 0. Since F is equicontinous, at each x ∈ X, there is an open set Ux such that
|f (y) − f (x)| < /3 for all y ∈ Ux . Since X is compact, and since {Ux : x ∈ X} is an open cover
of X, there exists a finite subcover {Ux1 , . . . , Uxn }.
We now claim that for this set {x1 , . . . , xm },

max{|f (xj ) − g(xj )|} < ⇒ d(f, g) <  .
3
To see this, note that every x belongs to some Uxj , and hence

|f (x) − g(x)| ≤ |f (x) − f (xj )|| + |f (xj ) − g(xj )| + |g(xj ) − g(x)| <  .

Now define a map Φ from F to Rn by

Φ(f ) = (f (x1 ), . . . , f (xn )) .

Now, since F is pointwise bounded, Φ(F) is a bounded, and hence totally bounded subset of
Rn . Conider any open cover of Φ(F) by balls of radius /3. Then the inverse images of these balls
are an open cover of F ball by sets of diameter no more than . Replacing them with the balls of
radius  about any point in them, we have our finite open cover by balls of radius . sine  > 0is
arbitrary, F is totally bounded.
Next, suppose that the closure of F is compact. Then F is totally bounded, so that for each
 > 0, there is a finite set {f1 , . . . , fn } ⊂ F such that
n
[
F⊂ B/4 (fj ) .
j=1

For any x, and choosing, say,  = 4, we have

n
[
{f (x) : f ∈ F } ⊂ (fj (x) − 1, fj (x) + 1) .
j=1

Thus, F is point wise bounded.

Next, for each x ∈ X, define Ux ⊂ X by
n
\
Ux = { y ∈ X : |fj (y) − fj (x)| < /4} .
j=1

Evidently, Ux is a finite intersection of open sets, and therefore open. Also, x ∈ Ux . Thus
[
Ux = X ,
x∈X

and then since X is compact, there is a finite set {x1 , . . . , xm } ⊂ X such that
m
[
Uxk = X .
k=1
EAC January 30, 2013 15

Now, fix any x ∈ X. By the above, x ∈ Uxk for some k ∈ {1, . . . , m}. For any f ∈ F, there is
some j ∈ {1, . . . , n} such that d(f, fj ) < /4. But then for any y ∈ Uxk ,

|f (y) − f (x)| ≤ |f (y) − fj (y)| + |fj (y) − fj (x)| + |fj (x) − f (x)|
 
≤ + |fj (y) − fj (x)| +
4 4
(1.8)

where the first inequality is the triangle inequality, the second uses the fact that d(f, fj ) < /4.
Then, once more by the triangle inequality, and then the definition of Uxk ,

|fj (y) − fj (x)| ≤ |fj (y) − fj (xk )| + |fj (xk ) − fj (x)| ≤ .
2
Altogether, we have |f (x) − f (y)| <  whenever y ∈ Uxk , and f ∈ F, so that Uxk is the required
neighborhood of x. Hence, F is equicontinuous.

1.6 The Stone-Wierstrass Theorem

The Stone-Wierstrass Theorem is an approximation theorem that generalizes the classical Wier-
strass Approximation Theorem that we discussed at the beginning of these notes.
We begin with two definitions. Let X be a compact topological space, and let C(X, R) be the
space of continuous real valued functions on X equipped with the uniform metric. A subset A of
C(X, R) is an algebra in case A is a vector subspace over R of C(X, R) equipped with its usual rules
of addition and scalar multiplication, and if, moreover, for every f, g ∈ A, the pointwise product
f g also belongs to A.
A subset A of C(X, R) is separating in case for pair of distinct points x, y in X, there is an
f ∈ A such that f (x) 6= f (y).
Notice that if X is not Hausdorff, not even C(X, R) the space of all continuous real valued
functions on X is separating. Indeed, if X is not Hausdorff, there exist two distinct points x and
y in X such that every neighborhood U of x contains y. But then for any continuous function
f , f (x) = f (y). Hence throughout this subsection, we shall only be concerned with Hausdorff
topological spaces.
The primary example of a separating algebra to keep in mind is X = [0, 1], with A being the
algebra of all polynomials in the real variable x ∈ [0, 1]. To see that this algebra is separating,
consider the polynomial p(x) = x. Then for x0 6= x1 in X, p(x0 ) 6= p(x1 ).

1.26 THEOREM (Stone-Wierstrass). Let X be a compact topological space, and let A be a subset
of C(X, R) that is a separating algebra. Let B be the uniform closure of A. Then either B = C(X, R),
or else B consisits of all continuous functions on X that vanish at some fixed point x0 . In particular,
if A contains the constant functions, B = C(X, R).

We will prove Theorem 1.26 as a consequence of two lemmas, and shall make use of the partial
order in C(X, R): If f, g ∈ C(X, R), we write f ≤ g in case f (x) ≤ g(x) for all x ∈ X. With this
partial order, C(X, R) is a lattice: Given any f, g ∈ C(X, R) there is a unique function g∧f ∈ C(X, R)
such that g ∧ f ≤ f, g, and such that h ≤ g ∧ f whenever h ≤ f, g. Of course, g ∧ f is defined by

g ∧ f (x) = min{ f (x) , g(x) } ,

EAC January 30, 2013 16

which is continuous.
Likewise, given any f, g ∈ C(X, R) there is a unique function g ∨ f ∈ C(X, R) such that f, g ≤
g ∨ f , and such that g ∨ f ≤ h whenever f, g ≤ h. Of course, g ∧ f is defined by

g ∨ f (x) = max{ f (x) , g(x) } ,

which is continuous.
A subset F of C(X, R) is itself a lattice if whenever f, g ∈ F, then both f ∧ g and f ∨ g belong
to F. Then observing that
1 1
f ∧ g = (f + g − |f − g|) and f ∨ g = (f + g + |f − g|) , (1.9)
2 2
we see that a subset F of C(X, R) is itself a lattice if whenever f ∈ F, then |f | ∈ F.

1.27 LEMMA (Limit point criterion for lattices in C(X, R)). Let X be a compact Hausdorff space.
Let F ⊂ C(X, R) be a lattice.
If f is any element of C(X, R) with the property that for every x, y ∈ X, there exists a function
fx,y ∈ F for which
fx,y (x) = f (x) and fx,y (y) = f (y) . (1.10)
Then f is a limit point of F; i.e., it belongs to the closure of F.

Proof. Fix any f ∈ C(X, R) with the property every x, y ∈ X, there exists a function fx,y ∈ F
such that (1.10) is satisfied. Fix any  > 0. We must show that there exists some g ∈ F with
|g(x) − f (x)| <  for all x ∈ X.
First, for each (x, y) ∈ X × X, make some choice of fx,y , and define the open set Ux,y ⊂ X by

Ux,y = { z : fx,y (z) < f (z) +  } .

Evidently, x, y ∈ Ux,y . Therefore [

X= Ux,y ,
x∈X

and then, since X is compact, there exists a finite set {x1 , . . . , xn } ⊂ X such that
n
[
X= Uxj ,y .
j=1

Now define the function fy by

fy = fx1 ,y ∧ fx2 ,y ∧ · · · ∧ fxn ,y .

Since F is a lattice, fy ∈ F, and

fy ≤ f + 
in the lattice order; i.e., everywhere on X.
Furthermore, since fxj ,y (y) = f (y) for each j, fy (y) = f (y). Therefore, defining the open set
Vy by
Vy := { z ∈ X : f (z) −  < fy (z) } ,
EAC January 30, 2013 17

we have y ∈ Vy , and hence [

X= Vy ,
y∈X

hen, since X is compact, there exists a finite set {y1 , . . . , ym } ⊂ X such that
m
[
X= Vyk .
k=1

Now define g by
g = fy1 ∨ fy2 ∨ . . . , ∨fym .
Then since F is a lattice, g ∈ F, and by construction,

f −≤g ≤f + ,

which means that |f (x) − g(x)| <  for all x ∈ X.

1.28 LEMMA (A closed algebra in C(X, R) is a lattice). Let X be a compact Hausdorff space.
Let B be a closed subset of C(X, R) that is also a subalgebra of C(X, R). Then B is a lattice.

Proof. By the remarks we have made concerning (1.9), it suffices to show that for all f ∈ B, |f | ∈ B.
Since X is compact and f is continuous, f is bounded above and below, and hence there is a finite
positive number c such that |cf | ≤ 1. Then since |cf | = c|f |, we may freely suppose that |f | ≤ 1.
Therefore, fix any f ∈ B with |f | ≤ 1, We shall complete the proof by showing that there exists
a sequence of polynomials {pn }n∈N so that

|f | = lim pn (f 2 ) (1.11)
n→∞

in the uniform topology. Since B is an algebra, pn (f 2 ) ∈ B for each n, and then since B is closed,
|f | ∈ B.
For any number a ∈ [0, 1], we define a sequence {bn }n∈N recursively as follows: We set b1 = 0
and then for al n ∈ N,
a − b2n
bn+1 = bn + .
2
Notice that
a a2
b1 = 0 , b2 = , b3 = a − ,
2 8
and so forth. It is easy to see by induction that for each n, there is a polynomial pn , independent
of the value of a, so that such that bn = pn (a).
√
We claim that a = limn→∞ bn . This will give us a sequence of polynomials {pn }n∈N such that
for each a ∈ [0, 1],
√
a = lim pn (a) ,
n→∞

and therefore, such that

|f (x)| = lim pn (f 2 (x))
n→∞
EAC January 30, 2013 18

for all x in X. Then, since X is compact, Dini’s Theorem implies that (1.11) is true with uniform
convergence.
√
Hence, we need only verify the claim that a = limn→∞ bn . To do this, note that
√ √ √
√ √ √
 
( a − bn )( a + bn ) a + bn
a − bn+1 = a − bn − = ( a − bn ) 1 − .
2 2
√ √
Since a ≤ 1, as long as bn ≤ a, the right hand side is non-negative, and therefore bn+1 ≤ a.
√ √
Since b1 ≤ a, it follows that a is an upper bound for the sequence {bn }n∈N .
Now, knowing that b2n ≤ a for all n, it is clear from the definition that {bn }n∈N is a monotone
non-decreasing sequence. Therefore the limit b = limn→∞ bn exists and satisfies

a − b2
b=b+ .
2
√
This means that b2 = a, and since b ≥ 0, b = a.

Proof of Theorem 1.26: Fix x 6= y in X, and consider the linear transformation from A to R2 given
by
f 7→ (f (x), f (y)) .
The range of this linear transformation is a subspace S of R2 .
Since A separates, there can be at most one point x0 ∈ X for which g(x0 ) = 0 for all g ∈ A.
Let us first assume first that neither x nor y is such a point. Since A is an algebra, and a vector
space in particular, if g is in A so is very multiple of g. By assumption, there is some g ∈ A such
that g(x) 6= 0, and by choosing an appropriate multiple, we may arrange that g(x) = 1.
Thus, S contains a vector of the form (1, a). ( with Since A separates, we can choose g ∈ A so
that a 6= 1.
Now there are two cases to consider. If also a 6= 0, then the two vectors (1, a) and (1, a2 ) are
linearly independent, and (1, a2 ) also belongs to S since A is an algebra (so that g 2 ∈ A). On the
other hand if a = 0 then S contains the vector (1, 0), and, since there is some other g with g(y) = 1,
there is some b ∈ R such that (b, 1) inS. Hence in this case, S contains the two vectors (1, 0) and
(b, 1) which are linearly independent. Either way, S = R2 , and so we have proved that as long as
g(x) 6= 0 and h(y) ne0 for some g, h ∈ A, then S is all of R2 .
This has the consequence that for any f ∈ C(X, R), we can find a function fx,y ∈ A for which

(f (x), f (y)) = (fx,y (x), fx,y (y)) . (1.12)

Now we have two cases once more: Suppose first that there is no point x0 ∈ X with f (x0 ) = 0
for all f ∈ A. Then the above argument applies for all x and y in X and all f ∈ C(X, R), we can
find fx,y ∈ A such that (1.12) is true. Moreover, by Lemma 1.28, B is a lattice. Therefore, by
Lemma 1.27, f is a linit point of B, and since B is closed, f ∈ B. Since f is an arbitrary element
of C(X, R), we see that in this case, B = C(X, R).
The remaining case to consider is that in which there is one point x0 such that g(x0 ) = 0 for
all g ∈ A, and henceB, so that B is certainly contained in the closed subset of C(X, R) consisting
of continuous functions f on X such that f (x0 ) = 0.
EAC January 30, 2013 19

Let f be any such function. The argument made above show that as long as neither x nor y
equals x0 , then there is some g ∈ A, and hence B, for which (1.12) is true. Now suppose that
x = x0 , and y 6= x0 . Then we trivially have
f (x0 ) = g(x0 ) = 0
for all g ∈ B. And since A separates, and is a vector space, we can choose g so that f (y) = g(y).
Therefore, for any f ∈ C(X, R) with f (x0 ) = 0, no matter how x and y are chosen, we can can find
gx,y ∈ B so that (1.12) is true.
Then the argument made above shows that every f ∈ C(X, R) with f (x0 ) = 0 is a limit point
of B, and hence belongs to B. Therefore, in this second case, B is the subset of C(X, R) consisting
of functions f with f (x0 ) = 0.

In our proof of Theorem 1.26, we made use of the fact that our functions f were real valued,
and not complex valued: The real numbers are ordered, while the complex numbers are not, and
the order on the complex number played a crucial role in the proof through our use of Lemma 1.27.
This is not simply an artifact of the proof: If in the statement of the theorem we replace C(X, R)
by, C(X, C), the space of continuous complex valued functions on X, the statement becomes false.
To see this, take X to be the closed unit disc in the complex plane C. Take A to be the algebra
of all complex polynomials in the complex variable z, which clearly separates. Polynomials in z
are analytic, and uniform limits of analytic functions are analytic, and so the closure of A consists
of functions that are analytic in the interior of the the unit disc. Obviously, not every continuous
function of the closed unit disc is analytic in the interior of the disc; f (z) = z ∗ , the complex
conjugate of z, is an example. Hence, the uniform closure of A is not the full set of continuous
complex valued functions on the closed unit disc.
However, under one simple additional condition on the algebra A, one can reduce the complex
valued case to the real case.
A (complex) subalgebra A of the algebra of complex valued function on a compact Hausdorf
space is called a ∗-algebra in case it is closed under complex conjugation. That is, whenver f ∈ A,
then f ∗ ∈ A, where f ∗ is the function defined by f ∗ (x) = (f (x))∗ for all x ∈ X.
In this case, for every f ∈ A, the real and imaginary parts of f both belong to A. It is also
easy to see that when A separates, so does the real algebra consisting of the real and imaginary
parts of functions in A. Applying the Stone-Wierstrass Theorem to this algebra, one can separately
approximate, in the uniform metric, the real and imaginary parts of any continuous complex valued
function on X by functions in A.
In summary, we have:
1.29 THEOREM (Complex Stone-Wierstrass). Let X be a compact topological space, and let A
be a subset of C(X, C) that is a separating ∗-algebra. Let B be the uniform closure of A. Then
either B = C(X, C), or else B consisits of all continuous functions on X that vanish at some fixed
point x0 . In particular, if A contains the constant functions, B = C(X, R).
Here is one important application of Theorem 1.29: Let X be the unit circle in C, with its usual
topology. Let A ⊂ C(x, C) be the set consisting of functions f of the form
n
X
f (z) = aj z n
j=−n
EAC January 30, 2013 20

for some n ∈ N, and some numbers a−n , . . . , an in C. (Each element of X is a complex number
z, and z n denotes the nth power of z.) The elements of A are called complex trigonomentric
polynomials
It is easy to see that A is a ∗-algebra, and that A separates. Hence, by Theorem 1.29, A is
dense in C(X, C). This proves:

1.30 THEOREM (Density of Complex Trigonometric Polynomials). Let X be the unit circle in
C, with its usual topology. Then the set of complex trigonometric polynomials is dense in C(X, C),
with respect to the uniform topology.

2 Tychonoff ’s Theorem
Let X be a set. The Cartesian product of X with itself, X ×X, is the set of all ordered pairs (x1 , x2 )
of elements of X. Of course (x1 , x2 ) is the graph of a unique function f : {1, 2} → X, namely the
one with f (1) = x1 and f (2) = x2 . (One can accomodate Cartesian products of two different sets
Y and Z in this framework by considering X = Y ∪ Z and restricting attention to functions f such
that f (1) ∈ Y and f (2) ∈ Z. No real generality is lost in taking the sets to be the same, and the
notation is much simpler, so that is how we shall proceed.)
More generally, given any set non-empty S, the Cartesian product of X indexed by S, denotes
S
X , is the set of all functions from S to X. For example, X N is the set of all infinite sequences
{xn }n∈N of elements of X.
On any Cartesian product, there is a natural family of functions with values in X, namely the
coordinate functions: For each s ∈ S, define

ϕs : X S → X

by
ϕ(f ) = f (s) .
That is, one simply evaluates the function f ∈ X S at s.
Note that when S = {1, 2}, ϕj ((x1 , x2 )) = xj , which is why the ϕs are called coordinate
functions.
Now suppose that X is a topological space. Is there a nice topology on X S for which all of the
coordinate functions are continuous? Of course, there are plenty of topologies on X S for which all
of the coordinate functions are continuous, but they are not necessarily very nice: For example, if
we equip X S with the power set topology; i.e., the topology consisting of all subsets of X S , then
every function on X S is continuous.
However, with such a topology, as long as X has infinitely many elements, X S will not be
compact. On the other hand, if we take X = [0, 1] and S = {1, 2}, then we can identify X S with
the closed unit square in R2 . If we equip this with its usual metric toplogy, then both coordinate
functions are clearly continuous, and the Cartesian product itself is compact.
Tychonoff’s theorem says that there is a natural topology on any Cartesian product that is
“nice” in the sense that under this topology we have both:
(1) If X is compact, then so is X S , no matter what non-empty set S is.
(2) For each s ∈ S, the coordinate function ϕs is continuous
EAC January 30, 2013 21

To find such a topology O, we must include enough open sets to make each φs continuous,
and should avoid introducing too many more than that. The more open sets we include, the more
open covers there are to consider, and if we produce to many of these, some might well lack finite
subcovers.
The way to proceed turns on the following very simple observation: Suppose that Y is any set,
and A is any set of subsets of Y . Let U and V be two topologies on Y such that A ⊂ U and A ⊂ V.
Then U ∩ V is also a topology – it is very easy to see that is satisfies the three requirements in the
definition, and clearly, it contains A.
More generally, consider the set of all topologies U on Y that contain A. The intersection of all
of the topologies is again a topology that contains A. By construction, it is the smallest topology
on Y that contains A in the sense that any other topology that contains A also contains every set
in this one. One often refers to this topology as the topology of Y generated by A.

2.1 DEFINITION (Product topology). Let (X, O) be a topological space, and S a nonempty
set. The product toplogy on a Cartesian product X S the smallest topology on X S containing all of
the sets of the form
φ−1
s (U ) , s∈S , U ∈O .

By construction, each of the coordinate functions is continuous when X S is equipped with the
product Topology. Moreover:

2.2 THEOREM (Tychonoff’s Theorem). Let X be a compact topological space, and S any non-
empty set. Then X S , equipped with the product topology, is compact.

The special case of this theorem in which X is a compact metric space and S is countable (or
finite) is fairly easy to prove using the theorems presented so far in these notes. This is developed
in the exercises that follow. The general case involves either the theory of “nets” or the theory of
“filters”, and this would be a digression, since we shall not invoke the general case in this book,
nor shall we have any other occasion to use the theory of nest of filters. Furthermore, the proof
of the general case involves the axiom of choice in a much more subtle way than does the spacial
case. This is not a problem, but discussion of these subtleties would take us far afield. (The axiom
of choice enters the subject, even in the special case, in an essential way: It is the axion of choice
whihc assures us that X S is non-empty: one can always choose, for each x ∈ s, some x(s) ∈ X.)
It is well worth knowing the general case nonetheless. It shows that advantage of the 20th
century notion of compactness, as defined above, in terms of open covers, and the 19th century
notion of sequential compactness