Chapter IV

Completeness and Compactness

1. Introduction
In Chapter III we introduced topological spaces a generalization of pseudometric spaces. This
allowed us to see how certain ideas continuity, for example can be extended into a situation
where there is no distance between points. Continuity does not really depend on the
pseudometric . but only on the topology.
Looking at topological spaces also highlighted the particular properties of pseudometric spaces
that were really important for certain purposes. For example, it turned out that first countability is
the crucial ingredient for proving that sequences are sufficient to describe a topology, and that
Hausdorff property, not the metric . , is what matters to prove that limits of sequences are unique.
We also looked at some properties that are equivalent in pseudometric spaces for example,
second countability and separability but are not equivalent in topological spaces.
Most of the earlier definitions and theorems are just basic tools for our work. Now we look at
some deeper properties of pseudometric spaces and some significant theorems related to them.

2. Complete Pseudometric Spaces

Definition 2.1 A sequence B8 in a pseudometric space \ . is called a Cauchy sequence
if a% ! b M N such that if 7 8 R , then .B7 B8 %.
Informally, a sequence B8 is Cauchy if its terms get closer and closer to each other. It should
be intuitively clear that this happens if the sequence converges, and the next theorem confirms
this.
Theorem 2.2 If B8 B in \ ., then B8 is Cauchy.
Proof Let % !. Because B8 B, there is an R such that .B8 B #% for 8 R .
So if 7 8 R , we get .B7 B8 .B7 B .B B8 #% #% %. Therefore B8 is
Cauchy.
A Cauchy sequence does not always converge. For example, look at the space . where . is
the usual metric. Consider a sequence in that converges to # in for example, we could
used ;8 " "% "%" "%"% . Since ;8 is convergent in ;8 is a Cauchy
sequence in (Cauchy depends only on . and the numbers ;8 , not whether we are thinking of
these ;8 's as elements of or ..)
But ;8 has no limit in . (Why? To say that # is part of the answer.)

154

Definition 2.3 A pseudometric space \ . is called complete if every Cauchy sequence in

\ . has a limit in \ .

Example 2.4

Throughout this example, . denotes the usual metric on subsets of .

1) . is complete. (For the moment, we take this as a simple fact from analysis;
however we will prove it soon.) But . and . are not complete completeness is not
hereditary!
2) Let B8 be a Cauchy sequence in .. Then there is some R such that
lB7 B8 l 1 for 7 8 R . But the B8 's are integers, so this means that B7 B8 for
7 8 R . So every Cauchy sequence is eventually constant and therefore converges. . is
complete.
3) Consider with the metric ." 7 8 l 7" 8" l. Then the sequence B8 8 is
Cauchy. If 5 , then ." B8 5 l 8" 5" l 5" ! as 8 , so B8 does not converge to
5 . The space ." is not complete.
However ." has the discrete topology, so the identity function 0 8 8 is a
homeomorphism between ." and the complete space ( . So completeness is not a
topological property.
A homeomorphism takes convergent sequences to convergent sequences and non
convergent sequences to nonconvergent sequences. For example, 8 does not converge in
." and 0 8 does not converge in .. These two homeomorphic metric spaces have
exactly the same convergent sequences but they do not have the same Cauchy sequences.
Changing a metric . to a topologically equivalent metric . w does not change the open sets and
does not change which sequences converge. But the change may create or destroy Cauchy
sequences since this property depends on distance measurements.
Another similar example: we know that any open interval + , is homeomorphic to .
But + , is not complete and is complete (where both space have the metric . ).
4) In light of the comments in 3) we can ask: if \ . is not complete, might there be a
different metric . w . for which \ . w is complete? To be specific: could we find a metric .w
on so . . w but where . w is complete? We could also ask this question about .
Later in this chapter we will see that the answer is no in one case but yes in the other.
5) ." and 8" 8 . are both countable discrete spaces, so they are
homeomorphic. In fact, the function 0 8 8" is an isometry between these spaces: ." 7 8
| 7" 8" l .0 7 0 8 An isometry is a homeomorphism, but since it preserves distances,
an isometry also takes Cauchy sequences to Cauchy sequences and non-Cauchy sequences to
non-Cauchy sequences. Therefore if there is an isometry between two pseudometric spaces, then
one space is complete iff the other space is complete.
For example, the sequence B8 8 is Cauchy is ." and the isometry 0 carries
B8 to the Cauchy sequence 0 B 8 8" in 8" 8 .. It is clear that 8 has no limit
in the domain and 8" has no limit in the range.
." and 8" 8 . look exactly alike not just topologically but
as metric spaces. We can think of 0 as simply renaming the points in a distance-preserving way.
Theorem 2.5 If B is a cluster point of a Cauchy sequence B8 in \ ., then B8 B.

155

Proof Let % !. Pick R so that .B7 B8 #% when 7 8 R Since B8 clusters at B, we

can pick a O R so that BO F #% B. Then for this O and for 8 R
.B8 B .B8 BO .BO B #% #% %
so B8 B.
Corollary 2.6 A Cauchy sequence B8 in \ . converges iff it has a cluster point.
Corollary 2.7 In a metric space \ ., a Cauchy sequence can have at most one cluster point.

Theorem 2.8 A Cauchy sequence B8 in \ . is bounded that is, the set B" B# B8
has finite diameter.
Proof Pick R so that .B7 , B8 " when 7 8 R . Then B8 F" BR for all 8 R
Let < max " .B" BR .B# BR .BR " BR Therefore, for all 7 8 we have
.B7 B8 .B8 BR .BR B7 #<, so diam B" B# B8 #<
Now we will prove that (with its usual metric . ) is complete. The proof depends on the
completeness property (also known as the least upper bound property) of . We start with two
lemmas which might be familiar from analysis. A sequence B8 is called weakly increasing if
B8 B8" for all 8, and weakly decreasing if B8 B8" for all 8. A monotone sequence is one
that is either weakly increasing or weakly decreasing.
Lemma 2.9 If B8 is a bounded monotone sequence in , then B8 converges.
Proof Suppose B8 is weakly increasing. Since B8 is bounded, B8 8 has a least upper
bound in . Let B sup B8 8 . We claim B8 B
Let % !. Since B % B we know that B % is not an upper bound for B8 8 .
Therefore B % BR B for some R Since B8 is weakly increasing and B is an upper bound
for B8 8 it follows that B % BR B8 B for all 8 R . Therefore lB8 Bl %
for all 8 R so B8 B.
If B8 is weakly decreasing, then B8 is a weakly increasing sequence so
B8 + for some +; then B8 +
Lemma 2.10 Every sequence B8 in has a monotone subsequence.
Proof Call B5 a peak point of B8 if B5 B8 for all 8 5 . We consider two cases.
i) If B8 has only finitely many peak points, then there is a last peak point B8! in B8 .
Then B8 is not a peak point for 8 8! Pick an 8" 8! . Since B8" is not a peak point, there is
an 8# 8" with B8" B8# . Since B8# is not a peak point, there is an 8$ 8# with B8# B8$ .
We continue in this way to pick an increasing subsequence B85 of B8 .
ii) If B8 has infinitely many peak points, then list the peak points as a subsequence
B8" B8# (where 8" 8# . Since each of these points is a peak point, we have B85 B85"
for each 5 , so B85 is a weakly decreasing subsequence of B8 .

156

Theorem 2.11 . is complete.

Proof Let B8 be a Cauchy sequence in .. By Theorem 2.8, B8 is bounded and by the
Lemma 2.10, B8 has a monotone subsequence B85 Of course B85 is also bounded. By
Lemma 2.9, B85 converges to some point B . Therefore B is a cluster point of the Cauchy
sequence B8 , so B8 B
Corollary 2.12 8 . is complete.
Proof Exercise (Hint: For 8 #, the sequence B8 C8 in # is Cauchy iff the sequences
B8 and C8 both are Cauchy sequences in . If B8 B and C8 C, then
B8 C8 B C in # .)

3. Subspaces of Complete Spaces

Theorem 3.1 Let . be a metric and suppose E \ ..

1) If \ . is complete and E is closed, then E . is complete.
2) If E . is complete, then E is closed in \ ..
Proof 1) Let +8 be a Cauchy sequence in E. Then +8 is also Cauchy in the complete space
\ ., so +8 B for some B \ But E is closed, so this limit B must be in E, that is,
+8 B E. Therefore E . is complete.
2) If B cl E, we can pick a sequence in E such that +8 B. Since +8 converges,
it is a Cauchy sequence in E. But E . is complete, so we know +8 + for some
+ E \ . Since limits of sequences are unique in metric spaces, we conclude that B +.
Therefore if B cl E, then B E. So E is closed.
Part 1) of the proof is valid when . is a pseudometric, but the proof of part 2)
requires that . be a metric. Can you provide an example of a pseudometric space \ .
for which the part 2) of the theorem is false?

The definition of completeness is stated in terms of the existence of limits for Cauchy sequences.
The following theorem characterizes complete spaces differently: in terms of the existence of
points in certain intersections. This illustrates an important idea: completeness for \ . can be
expressed in different ways, but all equivalent characterizations of completeness are statements
that guarantee the existence of certain points.

Theorem 3.2 (The Cantor Intersection Theorem) The following are equivalent for a metric
space \ .

157

1) \ . is complete
2) Whenever J" J# ... J8 ... is a decreasing sequence of nonempty
closed sets with diamJ8 !, then
8" J8 B for some B \
Proof 1 2 For each 5 , pick a point B5 J5 . Since diamJ5 !, the sequence B5 is
Cauchy so B5 B for some B \
For any 8, the subsequence B85 B8" B8# B. Since the J8 's are
decreasing, this subsequence is inside the closed set J8 , so B J8 . Therefore B
8" J8 .
If B C
J
,
then
.B
C

diam
J

for
every
8
Since
diam(
J

!,
8
8
8" 8

this means that .B C ! Therefore B C, and so 8" J8 B.

2 1 Suppose 1) is false and let B8 be a nonconvergent Cauchy sequence B8 .
Without loss of generality, we may assume that all the B8 's are distinct. (If any value B5 C
were repeated infinitely often, then C would be a cluster point of B8 and we would
have B8 C. So, since each term occurs only finitely often, it is possible to pick a subsequence
from B8 whose terms are all distinct, and this subsequence is also a nonconvergent Cauchy
sequence.) We will construct sets J8 which violate condition 2).
Let J8 B8 B8" B85 the 8>2 tail of the sequence B8 . For each 8,
%
J8 g J8 J8" and
8" J8 g Also, for % !, we can pick R so that .B7 B8 # if
%
7 8 R .
This means that for 8 R , diamJ8 diamJR # % and therefore
diamJ8 !.
We claim that the J8 's are closed which contradicts 2) and completes the proof.
Suppose that some J8! is not closed. Then there is a point B cl J8! J8! and we can find a
sequence of distinct terms in B8! B8! " B8! # J8! that converges to B. This sequence
is a subsequence B85 B of the original sequence B8 , so B is a cluster point of B8 But
this is impossible a nonconvergent Cauchy sequence B8 cannot have a cluster point.
How would Theorem 3.2 be different if . were only a pseudometric? How would the proof
change? Can you locate where the fact that . is a metric is used in each half of the proof ?

Example 3.3
1) In the complete space . consider:
a) J8 8
b) J8 ! 8"
c) J8 8 8 8"

In each case,
8" J8 g. Why are these examples not inconsistent with the Theorem 3.2?

2) Let J8 B # 8" B # 8" . The J8 's satisfy all the

hypotheses in Theorem 3.2), but
8" J8 g. Why is this not inconsistent with Theorem 3.2?
Example 3.4 Here is a proof, using the Cantor Intersection Theorem, that the closed interval
! " is uncountable.

158

If ! " were countable, we could list its elements as a sequence B8 Pick a subinterval
+" ," of ! " with length less than "# and excluding B" . Then pick an interval
+# ,# +" ," of length less than "% and excluding B# . Continuing inductively, choose
"
an interval +8" ,8" +8 ,8 of length less than #8"
and excluding B8" . Then
+" ," +# ,# +8 ,8 and these sets clearly satisfy the conditions in part
2) of the Cantor Intersection Theorem. But
8" +8 ,8 g. This is impossible since
! " is complete.

The following theorem has some interesting consequences. In addition, the proof is a very nice
application of Cantor's Intersection Theorem. The completeness is used to prove the existence of
very many points in \ .

Theorem 3.6 Suppose \ . is a nonempty complete metric space with no isolated points. Then
l\l -
To prove Theorem 3.5, we will use the following lemma.

Lemma 3.5 Suppose \ . is a metric space that E \ If + , int E and + , then we

can find disjoint closed balls J+ and J, (centered at + and , ) with J+ int E and J, int E
Proof Let % .+ ,. Since + , int E we can choose positive radii %+ and %, so that
F%+ + int E and F%, , int E. We may choose both %+ and %, less than #% Let %
w
min%+ %,
Then we can use the closed balls J+ B .B + % w and
J, B .B , % w .
Proof of Theorem 3.6 The idea of the proof is to inductively construct - different descending
sequences of closed sets, each of which satisfies condition 2) in the Cantor Intersection Theorem,
and to do this in a way that the intersection of each sequence gives a different point in \ . It then
follows that l\l - The idea is simple but the notation gets a bit complicated. First we will
give the idea of how the construction is done. The actual details of the induction step are
relegated to the end of the proof.
Stage 1 \ is nonempty and has no isolated points, so there must exist two points
B! B" \ . Pick disjoint closed balls J! and J" , centered at B! and B" , each with
diameter "#
Stage 2 Since B! int J! and B! is not isolated, we can pick distinct points B!! and B!"
(both B! ) in int J! and use Lemma 3.5 to pick disjoint closed balls J!! and J!" (centered at
B!! and B!" ) and both int J! We can then shrink the balls, if necessary, so that each has
diameter #"#
We can repeat the same steps inside J" since B" int J" and B" is not isolated, we can
pick distinct points B"! and B"" (both B" ) in int J" and use the Lemma 3.5 to pick disjoint
closed balls J"! and J"" (centered at B"! and B"" ) and both int J" We can then shrink the
balls, if necessary, so that each has diameter #"# At the end of Stage 2, we have 4 disjoint
closed balls: J!! J!" J"! J""
Stage 3 We now repeat the same construction inside each of the 4 sets J!! J!" J"! J"" .
For example, we can pick distinct points B!!! and B!!" (both B!! ) in int J!! and use the Lemma

159

3.5 pick disjoint closed balls J!!! and J!!" (centered at B!!! and B!!" ) and both int J!! Then
we can shrink the balls, if necessary, so that each has diameter #"$ See the figure below.

After Stage 3, we have 8 disjoint closed balls: J!!! J!!" J!"! We have the beginnings of 8
descending sequences of closed balls at this stage. In each sequence, at the 8th stage, the sets
have diameter #"8 .
J!!!
J!!
J!!"
J!
J!"!
J!"
J!""
and

J"!!
J"!
J"!"
J"
J""!
J""
J"""
We continue inductively (see the details below) in this way at a given stage, each descending
sequence of closed balls splits into two disjoint branches. The sequence splits when we
choose two new nonempty closed balls inside the current one, making sure that their diameters
keep shrinking toward !.

160

In the end, each binary sequence = 8" 8# 85 ! " will correspond to a

descending sequence of nonempty closed sets whose diameters !:
J8" J8# 8# J8" 8# 8$ J8" 8# 85
For example, the binary sequence = ! " " ! ! " ! " corresponds to the descending
sequence of closed sets
J! J!" J!"" J!""! J!""!! J!""!!"
The Cantor Intersection theorem tells us for each such =, there is an B= \ such that
J8" J8# 8# J8" 8# 8$ J8" 8# 85
5" J8" 8# 85 B=

For two different binary sequences, say > 7" 7# 75 8" 8# 85 =, there
is a smallest 5 for which 75 85 . Then B= J8" 8# 85" 85 and B> J8" 8# 85" 75 . Since these
sets are disjoint, we have B= B> . Thus, mapping = B= gives a " " function from ! "
into \ . We conclude that - #i! l! "l l\l
Here are the details of the formal induction step in the proof..
Induction Hypothesis: Suppose we have completed 5 stages that is, for each 3 " 5
and for each 3-tuple 8" 83 ! "3 we have defined points B8" ...83 and closed balls
J8" 83 centered at B8" ...83 , with diamJ8" 83 #"3 and so that
for each 8" 83 , J8" J8" 8# J8" 8# 83
Induction step: We must construct the sets for stage 5 " For each 5 "-tuple
8" 85 85" ! "5" , we need to define a point B8" 85 85" and a closed ball
J8" 85 85" in such a way that the conditions in the induction hypothesis remain true with
5 " replacing 5 .
For any 8" 85 : we have B8" 85 int J8" 85 . Since B8" 85 is not
isolated, we can pick distinct points B8" 85 ! and B8" 85 " (both B8" 85 ) in int J8" ...85
and use the Lemma 3.5 to pick disjoint closed balls J8" ...85 ! and J8" ...85 " (centered at
B8" 85 ! and B8" 85 " ) and both int J8" ...85 . We can then shrink the balls, if necessary,
so that each has diameter #"5
Corollary 3.7 If \ . is a nonempty complete separable metric space with no isolated points,
then l\l - .
Proof Theorem II.5.21 (using separability) tells us that l\l - ; Theorem 3.6 gives us
l\l -

The following corollary gives us another variation on the basic result.

Corollary 3.8 If \ . is an uncountable complete separable metric space, then l\l - .

161

(So we might say that the Continuum Hypothesis holds among complete separable metric
spaces. )
Proof Since \ . is separable, Theorem II.5.21 gives us l\l - . If we knew that \ had no
isolated points, then Theorem 3.6 (or Corollary 3.7) would complete the proof. But \ might have
isolated points, so a little more work is needed.
Call a point B \ a condensation point if every neighborhood of B is uncountable. Let
G be the set of all condensation points in \ . Each point + \ G has a countable open
neighborhood S. Each point of S is also a non-condensation point, so + S \ G .
Therefore \ G is open so G is closed, and therefore G . is a complete metric space.
Since \ . is separable (and therefore second countable), \ G is also second
countable and therefore Lindelof.
Since \ G can be covered by countable open sets, it can be
covered by countably many of them, so \ G is countable: therefore G g (in fact, G must
be uncountable).
Finally, G . has no isolated points in G : if , G were isolated in G , then there would
be an open set S in \ with S G ,. Since S , \ G , S would be
countable which is impossible since , is a condensation point.
Therefore Corollary 3.7 therefore applies to G ., and therefore l\l lGl -
Why was the idea of a condensation point introduced in the example? Will the argument
work if, throughout, we replace condensation point with non-isolated point? If not,
precisely where would the proof break down?
The next corollary answers a question we raised earlier: is there a metric . w on which is
equivalent to the usual metric . but for which . w is complete that is, is completely
metrizable?

Corollary 3.9 is not completely metrizable.

Proof Suppose . w is a metric on equivalent to the usual metric ., so that g. g. w Then
. w is a nonempty metric space and, since . . w , the space . w has no isolated points
(no set ; is in g. g. w isolated point is a topological notion.). If . w were complete,
Theorem 3.6 would imply that is uncountable.

162

Exercises

E1. Prove that in a metric space \ ., the following are equivalent:

a) every Cauchy sequence is eventually constant
b) \ . is complete and g. is the discrete topology
c) for every E \ each Cauchy sequence in E converges to a point in E
(that is, every subspace of \ . is complete).

E2. Suppose that X is a dense subspace of the pseudometric space ] . and that every Cauchy
sequence in X converges to some point in ] . Prove that ] . is complete.

E3. Suppose that \ . is a metric space and that B8 is a Cauchy sequence with only finitely
many distinct terms. Prove that B8 is eventually constant, i.e., that for some 8 ,
B8 B8"

E4. Let p be a fixed prime number. On the set of rationals, the :-adic absolute
value | |: (or :-adic norm) is defined by:
k

For ! B , write B pnm for integers 5 7 8, where p does not divide 7 or 8. (Of
course, 5 may be negative. ) Define |B|: :5 :"5 . . We define |!|: !
For all B C ,
a) l Bl: ! and l Bl: ! iff B !
b) l BCl: l Bl: l Cl:
c) l B Cl: l Bl: l Cl:
d) l B Cl: max l Bl: l Cl: Bl: l Cl:
The :-adic metric .: is defined on by .: B C lB Cl:
Prove or disprove that ( .: ) is complete.

E5. Suppose that \ . is complete, that ] g is a Hausdorff space and that 0 \ ]

is continuous. Suppose that J" J# ... J8 ... is a sequence of closed sets in \ for which

diamJ8 0. Prove that

8" 0 J8 0 8" J8 .

E6. A metric space \ . is called locally complete if every point B has a neighborhood RB

163

(necessarily closed) which is complete.

a) Give an example of a metric spce \ . that is locally complete but not complete.
b) Prove that if H . is a locally complete dense subspace the complete metric space
\ ., then H is open in \ .
Hint: it may be helpful to notice if O is open and D is dense, then cl (S H) = cl(O).
This is true in any topological space \ g .
E7. Suppose E is an uncountable subset of with lEl 7 - (Of course, there is no such
set E if the Continuum Hypothesis is assumed.) Is it possible for E to be closed? Explain.

164

4. The Contraction Mapping Theorem

Definition 4.1 A point B \ is called a fixed point for the function 0 \ \ if 0 B B
We say that a topological space \ has the fixed point property if every continuous 0 \ \
has a fixed point.
It is easy to check that the fixed point property is a topological property.
Example 4.2
1) A function 0 has a fixed point if and only if the graph of 0 intersects the line
C B.
If E , then E is the set of fixed points for some function: for example
0 B B ;E B where ;E is the characteristic function of E. Is every set E the set of
fixed points for some continuous function 0 ? Are there any restrictions on the
cardinality of E?
2) The interval + , has the fixed point property. To see this, suppose that
0 + , + , is continuous. If 0 + + or 0 , ,, then 0 has a fixed point so assume that
0 + + and 0 , ,
Define 1B 0 B B. Then
1+ 0 + + !,
1, 0 , , ! and 1 is continuous. By the Intermediate Value Theorem (from analysis)
there is a - + , where 1- 0 - - !, that is, 0 - -
3) The Brouwer Fixed Point Theorem states that H8 B 8 .B ! " has the
fixed point property. For 8 " the proof of Brouwer's theorem is rather difficult; the usual
proofs use techniques from algebraic topology. For 8 ", H8 is homeomorphic to + ,, so
Brouwer's Theorem genralizes Part 2.
In fact, Brouwer's Theorem can be generalized as the Schauder Fixed Point
Theorem: every nonempty compact convex subset O of a Banach space has the fixed point
property.

Definition 4.3 0 \ . \ . is a contraction mapping (for short, a contraction) if there is

a constant ! " such that .0 B 0 C .B C for all B C \ .
Notice that a contraction 0 is automatically continuous: for % !, choose $ %. Then, for all
B \ .B C $ implies .0 B 0 C $ %. In fact, this choice of $ depends only on %
and not the point B. In this case we say that 0 is uniformly continuous.
The definition of uniform continuity for 0 on \ reads:
a% b$ aB aC .B C $ .0 B 0 C %
Compare this to the weaker requirement for 0 to be continuous on \ :
aB a% b$ aC .B C $ .0 B 0 C %

Example 4.4

165

1) The function 0 ! "% ! "% given by 0 B B# is a contraction because for any

B C ! "% , we have l0 B 0 Cl lB# C# l lB CllB Cl "# lB Cl.
2) But the function 0 given by 0 B B# is not a contraction.
contraction would require some ! " which

To be a

lB# C# l lB Cl holds for all choices of B C, which would imply that

lB Cl for all B C (which is false).
We could also argue that 0 is not a contraction by noticing that 0 is not uniformly
continuous. For example, if we choose % #, then no $ ! will satisfy the condition
aB aC .B C $ .0 B 0 C #. For example, given $ !, we can let B #$ and
#
C #$ $# . Then l0 B 0 Cl # $% #.
Theorem 4.5 (The Contraction Mapping Theorem) If \ . is a nonempty complete metric
space and 0 \ \ is a contraction, then 0 has a unique fixed point. (Nonempty is included
in the hypothesis because the empty function g from the complete space g to g is a contraction
with no fixed point.)
A fixed point provided by the Contraction Mapping Theorem can be useful, as we will soon see
in Theorem 4.9 (Picard's Theorem) below. The proof of the Contraction Mapping Theorem is
also useful because it shows how to make some handy numerical estimates (see the example
following Picard's Theorem.)
Proof Suppose ! " and that .0 B 0 C .B C for all B C \ . Pick any point
B! \ and apply the function 0 repeatedly to define
B" 0 B!
B# 0 B" 0 0 B! 0 # B!

B8 0 B8" 0 8 B!

We claim that this sequence B8 is Cauchy and that its limit in \ is a fixed point for 0 . (Here
you can imagine a control system where the initial input is B! and each output becomes the
new input in a feedback loop. This system approaches a steady state where
input output.)
Suppose % !. We want to show that .B8 B7 % for large enough 7 8 Assume 7 8.
Applying the contraction property 8 times gives:
.B8 B7

.0 8 B! 0 7 B! .0 0 8" B! 0 0 7" B!
.0 8" B! 0 7" B! # .0 8# B! 0 7# B!
8 .B! 0 78 B! 8 .B! B78

By the triangle inequality (and the contraction property, again) we get

166

.B8 B7 8 .B! B78

8 .B! B" .B" B# .B# B$

.B78" B78

8 .B! B" .B! B" # .B! B" 78" .B! B"

5
8 .B! B" " # 78" 8 .B! B"
5!

So

8 .B! B"
"

8 .B! B"
"

! as 8 .

% if 8 some R Then .B8 B7 % if 7 8 R . Therefore B8 is

Cauchy and since \ . is complete B8 B for some B \ .

By continuity, 0 B8 0 B. But also 0 B8 B because 0 B8 B8" . Therefore
0 B B because a sequence in a metric space has at most one limit. (If . were just a
pseudometric, we could conclude only that .B 0 B !.)
If C is also a fixed point for 0 , then .B C .0 B 0 C .B C. Since ! ", this
implies that .B C !, so B C that is, the fixed point B is unique. (If . were just a
pseudometric, 0 could have several fixed points at distance ! from each other.)

Notes about the proof

8 .B! B"
for each 8 . If we
"
8 .B! B"
" . Thus we have a computable

1) The proof gives us that if 7 8, then .B8 B7

fix 8 and let 7 , Then B7 B, so .B8 B
bound on how well B8 approximates the fixed point B.

2) For large 8 we can think of B8 as an approximate fixed point in the sense that
0 doesn't move the point B8 very much. To be more precise,
.B8 0 B8 .B8 B .B 0 B8 .B8 B .0 B 0 B8
.B8 B .B B8 " .B8 B
"

8 .B! B"
"

! as 8 .

Example 4.6 Consider the following functions that map the complete space into itself:
1) 0 B #B ": 0 is (uniformly) continuous, 0 is not a contraction, and 0 has a
unique fixed point: B "
2) 1B B# " is not a contraction and has no fixed point.
3) 2B B sin B has infinitely many fixed points.
We are going to use the Contraction Mapping Theorem to prove a fundamental theorem about the
existence and uniqueness of a solution to a certain kind of initial value problem in differential

167

equations. We will use the Contraction Mapping Theorem applied to the complete space in the
following example.
Example 4.7 Let \ be a topological space and G \ be the set of bounded continuous realvalued functions 0 \ , with the metric 30 1 sup l0 B 1Bl B \. (Since 0 1
are bounded, so is l0 1l and therefore this sup is always a real number. It is easy to check
that 3 is a metric.) The metric 3 is called the metric of uniform convergence because
30 1 % implies that 0 is uniformly close to 1
30 1 %

l0 B 1Bl % at every point B \

3
In the specific case of G 3, the statement that 08 0 is equivalent the statement (in
analysis) that 08 converges to 0 uniformly on .
Theorem 4.8 G \ 3 is complete.
Proof Suppose that 08 is a Cauchy sequence in G \ 3: given % !, there is some R so
that 308 07 % whenever 8 7 R . This implies (see *, above) that for any fixed B \ ,
l08 B 07 Bl % whenever 8 7 R . Therefore 08 B is a Cauchy sequence of real
numbers, so 08 B some <B . We define 0 B <B . Then 0 \ To complete the
3
proof, we claim that 0 G \ and that 08 0
First we argue that 0 is bounded. Pick R so that 308 07 " whenever 8 7 R .
Using 7 R , this gives 308 0R " for 8 R and therefore l08 B 0R Bl " for every
B \ . If we now let 8 , we conclude that
l0 B 0R Bl ", that is
" 0 B 0 R B " for every B.
Since 0R is bounded, there is a constant Q such that
Q 0 R B Q

for every B.

Adding the last two inequalities gives

l0 Bl Q "

for every B, so 0 is bounded.

3
We now claim that 08 0 . (Technically, this is an abuse of notation because 3 is defined
on G \ and we don't yet know that 0 is in G \ ! However, the same definition for 3 makes
sense on any collection of bounded real-valued functions, continuous or not. We are using this
extended definition of 3 here.) Let % ! and pick R so that 308 07 #% whenever
7 8 R . Then l08 B 07 Bl #% for every B whenever 8 7 R . Letting 7 , we
get that l08 B 0 Bl #% for all B and all 8 R . Therefore 308 0 % if 8 R .
Finally, we show that 0 is continuous at every point + \ and % !. For B \ and R we
have

168

l0 B 0 +l l0 B 0R Bl l0R B 0R +l l0R + 0 +l
3
Since 08 0 , we can choose an R large enough that 30R 0 $% This implies that
l0 B 0R Bl $% and l0R + 0 +l $% . Since 0R is continuous at +, we can choose a
neighborhood [ of + so that l0R B 0R +l $% for all B [ So, if B [ then we have
l0 B 0 +l

%
$

%
$

%
$

%, so 0 is continuous at +.

Theorem 4.9 (Picard's Theorem) Let 0 H be continuous, where H is a closed box in #

and B! C! int H. Suppose that there exists a constant Q such that for all B" C" and B" C#
in H
l0 B" C" 0 B" C# l Q lC" C# l (L)
Then
i) there exists an interval M B! + B! + B lB B! l + and
ii) there exists a unique differentiable function 1 M
such that

C! 1B!
1 w B 0 B 1B for B M

In other words, on some interval M centered at B! , there is a unique solution C 1B to the initial
value problem
C w 0 B C
CB C
!
!

Before beginning the proof, we want to make some comments about the hypotheses.
1) Condition (L) says that 0 satisfies a Lipschitz condition in the variable C on H. This
may seem a bit obscure. For our purposes, it's enough to notice that (L) is true if the partial
derivative 0C exists and is continuous on the box H. In that case, we know (from analysis) that
|0C | some constant Q on H. Thinking of 0 B" C as a single-variable function of C and using
the ordinary Mean Value Theorem gives us a point D between C" and C# for which
|0 B" C" 0 B" C# l l0C B" Dl lC" C# l Q lC" C# l
2) As a specific example, suppose 0 B C C B, that H " " " ", and
that B! C! ! !. Since 0C ", the preceding comment tells us that condition (L) is true.
Cw CB
Consider the initial value problem
. Picard's Theorem states that there
C! !
is a unique differentiable function C 1B, defined on some interval + + that satisfies this
system: 1! ! and 1 w B 0 B 1B that is, C w C B. Moreover, the proofs of
Picard's Theorem and the Contraction Mapping Theorem can actually help us to actually find this
solution. Of course, once a solution is actually found, a direct check might show that the solution
is actually valid on an interval larger than the interval M that comes up in the proof.

169

Proof (Picard's Theorem) We begin by transforming the initial value problem into an
equivalent problem that involves an integral equation rather than a differential equation.
Let C 1B be a continuous function defined on an interval M B! + B! +:
Suppose 1 satisfies (*): 0 is continuous, so 1 w B 0 B 1B is continuous. (Why?)
Therefore the Fundamental Theorem of Calculus tells us for all B M ,
1B 1B! B! 1 w > .> B! 0 >, 1> .>, so
B

1B C! B! 0 >, 1> .>

B

(**)

B
Suppose 1 satisfies (**): then 1B! ) C! B!! 0 >, 1> .> C! . Since 0 > 1> is
continuous, the Fundamental Theorem gives that for all B M , 1 w B = 0 B 1B.
Therefore the function 1B satisfies (*).

Therefore a function continuous 1 on M satisfies * iff it satisfies **. (Note that the initial
condition C! 1B! is built into the single equation ) We will find a 1 that satisfies
**.
Now we establish some notation:
We are given a constant Q in the Lipschitz condition.
Pick a constant O so that l0 B Cl O for all B C H. (We will prove later in this
chapter that a continuous function on a closed box in # must be bounded. For now, we
assume that as a fact from analysis.)
Because B! C! int H, we can pick a constant + ! so that
i) B lB B! l + C lC C! l O+ H, and
ii) +Q "
Let M B lB B! l + B! + B! +
We consider G M 3, where 3 is the metric of uniform convergence. By Theorem 4.8, this
space is complete. Let F 1 G M a> M l1> C! l O+.
For an arbitrary 1 G M, 1> might be so large that > 1> H If we look only at
functions in F , we avoid this problem: if 1 F and > M , then > 1> H because of
how we chose +. So for 1 F , we know that 0 > 1> is defined.)
Notice that F g since the constant function 1B C! is certainly in F .
F is closed in G M: if 2 cl F , then there is sequence of functions 18 in F such that
18 2 in the metric 3. Therefore 18 B 2B for all B M , and subtracting C! we get
l18 B C! l l2B C! l.

170

Since l18 B C! l O+ for each B M , then l2B C! l O+ for each B M . Therefore

2 F , so F is closed. Therefore F 3 is a nonempty complete metric space.
For 1 F , define a function 2 X 1 by

2B X (1)B C! B! 0 > 1> .>.

B

Notice that 2 is continuous (in fact, differentiable) and that, for every B M ,

B
B
B
l2Bl |C! | |B! 0 > 1> .> | | C! | B! |0 > 1>| .> | C! | B! O .> lC! l O+,

so 2 is bounded. Therefore X F G M. But in fact, even more is true: X F F , because

B
aB M , |X 1B C! | | 2B C! l lB! 0 > 1> .> | O+
We now claim that X F 3 F 3 is a contraction.
distances: if 1" and 1# F , then
3X 1" , X 1# l

To see this, we simply compute

sup | X 1" B X 1# B| B M
B
sup | B! 0 > 1" > 0 > 1# > .>| B M
B
sup |B! |0 > 1" > 0 > 1# >| .>| B M
B
sup | B! Q l1" > 1# >| .> | B M
( from Condition L )
B
sup | B! Q 31" 1# .> | B M
sup Q 31" 1# B B! l B M
+Q 31" 1#
31" 1# , where +Q 1.

Then the Contraction Mapping Theorem gives us a unique function 1 F for which 1 X 1.
From the definition of X , that simply means:
B
aB M , 1B X 1B C! B! 0 >, 1>.>

which is precisely condition (**).

Example 4.10 The method in the proof of Picard's Theorem and numerical estimates in the proof
of the Contraction Mapping Theorem can be used to get useful information about a specific initial
value problem. To illustrate, we consider
Cw CB
C! !
and find a solution that is valid on some interval containing 0.
We begin by choosing a box H with B! C! ! ! in its interior. Rather arbitrarily, we select
the box H " " " " Since l0 B Cl lC Bl lBl lCl # on H, we can use
O # in the proof.
Because |0C B Cl " throughout H, the Lipschitz condition (L) is satisfied with Q "

171

Following the proof of Picard's Theorem, we now choose a constant + so that

i) B lBl + C lCl O+ #+ H and
ii) +Q + " "
Again rather arbitrarily, we choose + "# , so that M B! + B! + "# "# .
Then F 1 G M l1B !l O+ 1 G M |1Bl " for all B "# "# .
Finally, we choose any function 1! F; to make things as simple as possible, we might as well
choose 1! to be the constant function 1! B ! on "# "# .
According to the proof of the Contraction Mapping Theorem, the sequence of functions
3
18 X 8 1! 1, where 1 is a fixed point for X , and 1 is the solution to our initial value
problem. We calculate:
1" B C! B! 0 > 1! > .> ! ! 0 > ! .> ! > .>
B

#
B
B
1# B X 1" B ! ! 0 > ># .> !

1$ B X 1# B ! 0 >
B

>$
'

># .> !
B

>#
#

> .>

>$
'

>#
#

B$
'

B#
#

> .>

B#
#

B%
#%

B$
'

B#
#

and, in general,
18 B X 18" B

B#
#x

B$
$x

B8"
8"x

The functions 18 converge (uniformly) to the solution 1.

In this particular problem, we are lucky enough to recognize that the functions 18 B are just the
B8
B8
B
partial sums of the series
" B
Therefore
8# 8x
8! 8x " B /
1B " B /B is a solution of our initial value problem, and we know it is valid for all
B M "# "# . (You can check by substitution that the solution is correct and that it
actually is a solution that works for all B .)
Even if we couldn't recognize a neat formula for the limit 1B, we could still make some useful
approximations. From the proof of the Contraction Mapping Theorem, we know that
3(18 1

8 31! 1"
" .

In this example +Q "# , so that 3(18 1

#

"
"
#8"
sup l! B# l B M #8"
#"$
"
distance #8#
of 1B on the interval "# "# .

"
#8# .

"
#8 31! 1"
" "#

Therefore 18 B is uniformly within

Finally, recall that our initial choice of 1! F was arbitrary. Since lsin Bl ", we know that
sin F , and we could just as well have chosen 1! B sin B. Then the functions 18 B
computed as X 1! 1" X 1# 1# would be quite different try computing 1" and 1#
but it would still be true that 18 B 1B " B /B uniformly on M the same limit 1
because the solution 1 is unique.

172

The Contraction Mapping Theorem can be used to prove other results for example, the Implicit
Function Theorem. (You can see details, for example, in Topology, by James Dugundji.)

173

Exercises

E8. + Suppose 0 is differentiable and that there is a constant O 1 such that

|0 w B | O for all B Prove that 0 is a contraction (and therefore has a unique fixed point.)
b) Give an example of a continuous function 0 such that
| 0 B 0 C l lB C l
for all B C but such that 0 has no fixed point.
Note: The function 0 is not a contraction mapping. If we allow " in the definition of
contraction, then the Contraction Mapping Theorem is not be true not even if we
compensate by using instead of in the definition.
E9. Let 0 \ . \ ., where \ . is a nonempty complete metric space. Let 0 5 denote
the k th iterate of 0 that is, 0 composed with itself 5 times.
a) Suppose that b5 for which 0 5 is a contraction. Then, by the Contraction Mapping
Theorem, 0 5 has a unique fixed point :. Prove that : is also the unique fixed point for 0 .
b) Prove that the function cos is not a contraction.
c) Prove that cos5 is a contraction for some 5 .
H3nt: the Mean Value Theorem may be helpful. )
d) Let 5 be such that 1 cos5 is a contraction and let : be the unique fixed point of
1. By a), : is also the unique solution of the equation cos B B. Start with 0 as a first
approximation for : and use the technique in the proof of the Contraction Mapping Theorem to
find an 8 so that | 18 ! : < 0.00001.
e) For this 8, use a computer or calculator to evaluate 18 !. (This solves the equation
cos B B with l Error l 0.00001.)
E10. Consider the differential equation C w B C with the initial condition C! ". Choose a
suitable rectangle H and suitable constants O Q and + as in the proof of Picard's Theorem. Use
the technique in the proof of the contraction mapping theorem to find a solution for the initial
value problem. Identify the interval M in the proof. Is the solution you found actually valid on an
interval larger than M ?

174

5. Completions
The set of rationals (with the usual metric . ) is not complete. However, the rationals are a
dense subspace of the complete space ( .. This is a model for the definition of a completion
for a metric space.

Definition 5.1 \ , . is called a completion of \ . if: . |\ \ ., \ is a dense

subspace of \ , and \ ,. is complete.
Loosely speaking, a completion of \ . adds the additional points needed and no others) to
provide limits for the Cauchy sequences that fail to converge in \ . . In the case of ., we
can think of the additional points as the irrational numbers, and the resulting completion is
.
If a metric space \ . is complete, what would a completion look like? Since \ . is a

complete subspace of \ , . , \ must be closed in \ . But \ must also be dense. So

\ \ that is, a complete metric space is its own completion. (If \ . is a complete

pseudometric space, then \ might not be closed in \ ; but each C \ \ is at distance

! from some B \ so if B8 C ] , then B8 already converges to B \ and adding the
point C was unnecessary.
Of course different spaces may have the same completion for example, . is a completion
for both . and .
Notice that a completion of \ . depends on . , not just the topology g. For example, .
and 8" 8 , . are homeomorphic topological spaces (both have the discrete topology).
But the completion of . itself!) is not homeomorphic to ! 8" 8 ., which is a
completion of 8" 8 .
Recall that if there is an (onto) isometry 0 between two metric spaces \ . and ] =, then
\ . and ] = can be regarded as the same metric space: we can think of 0 as just
assigning new names to the points of \ . If 0 \ . ] = is an isometry from \ into ] ,
then we can identify \ . with 0 \ = an exact metric copy of \ . inside ] = So
if ] = happens to be complete, then 0 \ = is dense in the complete space cl] 0 \ =. If
we identify 0 \ = with \ ., then we can call cl] 0 \ = a completion of \ . even
though \ is not literally a subset of cl] 0 \. Therefore, to find a completion of \ ., it is
sufficient to find an isometry 0 from \ . into some complete space ] . w .
Example 5.2
Consider with the metric . w 8 7 l 8" 7" l. . w is isometric to
( 8" 8 . where . is the usual metric on (0 8 8" is an isometry). These two spaces
are metrically identical.. Therefore ! 8" 8 . contains a dense isometric copy of
. w . The closure of that copy is ! 8" 8 , and we can view
! 8" 8 . as a completion of . w even though is not literally a subspace
of ! 8" 8

175

The next theorem tells us every metric space has a completion and, as we will see later, it is
essentially unique.

Theorem 5.3 Every metric space \ . has a completion.

Proof Our previous comments show that it is sufficient to produce an isometry of \ . into
some complete metric space: the complete space we will use is G \ 3 where, as usual, 3 is
the metric of uniform convergence.
The theorem is trivial if \ g so we assume that we can pick a point : \ . For any
point + \ , define 9+ \ by 9+ B .B + .B : The map 9+ is a difference of
continuous
functions
so
9+
is
continuous.
For
each
B \,

l9+ Bl l.B + .B :l .+ :, so 9+ is bounded. Therefore 9+ G \

Define F \ . G \ 3 by F+ 9+ We complete the proof by showing that
F is an isometry. This just involves computing some distances: for any + , \ ,
39+ 9, sup l9+ B 9, Bl B \
sup l.B + .B : .B , .B :l B \
sup l.B + .B ,l B \
For any B, l.B + .B ,l .+ ,.
And letting B , gives l.B + .B ,l .+ ,. Therefore
39+ 9, sup l.B + .B ,l B \ .+ ,

The completion of \ . given in this proof is clG \ F\,3. The proof is slick but it seems
to lack any intuitive content. For example, if we applied the proof to ., it would not at all
clear that the resulting completion is isometric to . (as we would expect). In fact, there is
another more intuitive way to construct a completion of \ ., but verifying all the details is
much more tedious. We will simply sketch this alternate construction here.
Call two Cauchy sequences B8 and C8 in \ . equivalent if .B8 C8 ! It is
easy to check that is an equivalence relation among Cauchy sequences in \ ..
Clearly, if B8 D \ then any equivalent Cauchy sequence also converges to D . If
two nonconvergent Cauchy sequences equivalent, then they are trying to converge to the
same point but the necessary point is missing in \ .

We denote the equivalence class of a Cauchy sequence B8 by B8 and let \ be the

set of equivalence classes. Define the distance . between two equivalence classes by

. B8 C8 lim .B8 C8 (Why must this limit exist?) It is easy to check that
8

. does not depend on the choice of representative sequences from the equivalence

~
classes, and that . is a metric on \ . One then checks (this is only tricky part) that

\ . is complete: any . -Cauchy sequence of equivalence classes B8 must

converge to an equivalence class in \ . .

For each B \ , the sequence B B B is Cauchy, and so B B B \ .

176


The mapping 0 \ . \ . given by 0 B B B B is an isometry, and it


is easy to check that 0 \ is dense in \ . so that \ . is a completion for
\ .

This method is one of the standard ways to construct the real numbers from the rationals: is
defined as the set of these equivalence classes of Cauchy sequences of rational numbers.
The real numbers can also be constructed as a completion for the rationals by using a
method called Dedekind cuts. However that approach also uses the order structure in
and therefore we cannot mimic it in the general setting of metric spaces.
The next theorem gives us the good news that, in the end, it doesn't really matter how we
construct a completion because the completion of \ . is essentially unique: all
completions are isometric (and, in a special way!). We state the theorem for metric spaces.
(Are there modifications of the statement and proof to handle the case where . is merely a
pseudometric? )
Theorem 5.4 The completion of \ . is unique in the following sense: if ] = and ^ > are
both complete metric spaces containing \ as a dense subspace, then there is an (onto) isometry
0 ] = ^ > such that 0 l\ is the identity map on \ . (In other words, not only are two
completions of \ . isometric, but there is an isometry between them that holds \ fixed. The
isometry merely renames the new points in the outgrowth ] \ )
Proof Suppose C ] Since \ is dense in ] , we can pick a sequence B8 in \ which
converges to C. Since B8 is convergent, it is Cauchy in \ ., and since >l\ \ . , B8 is
also a Cauchy sequence in the complete space ^ >. Therefore B8 some point D ^ . We
define 0 C D . Then 0 ] = ^ > (It is easy to check that 0 is well-defined that is,
we get the same D no matter which of the possibly many sequences B8 we first choose
converging to C.)
In particular, if B \ we can choose B8 to be the constant sequence B8 B; then
0 B B, so that 0 l\ is the identity map on \ .
We need to verify that 0 is an isometry, Suppose C w ] and we choose B8w C w .
We then have that
.B8 Bw8 =B8 B8w =C C w and
.B8 Bw8 >B8 B8w >D D w >0 C 0 Cw , so
=C C w >0 C 0 C w

According to Theorem 5.4, the space . no matter how we construct it is the completion
of the space ..

177

6. Category
In mathematics there are many different ways to compare the size of sets. These different
methods are useful for different purposes. One of the simplest ways is to say that one set is
bigger if it has more points than another set that is, by comparing their cardinal numbers.
In a totally different spirit, we might call one subset of # bigger than another if it has a larger
area. In analysis, there is a generalization of area. A certain collection ` of subsets of #
contains sets that are called measurable, and for each set W ` a nonnegative real number .W
is assigned. .W is called the measure of W and we can think of .W as a kind of
generalized area. A set with a larger measure is bigger.
In this section, we will look at a third completely different and useful topological idea for
comparing the size of certain sets. The most interesting results in this section are be about
complete metric spaces, but the basic definitions make sense in any topological space \ g

Definition 6.1 A subset E of the topological space \ g is called nowhere dense in \

if int\ (cl\ E) g (In some books, a nowhere dense set is called rare.)
A set has empty interior iff its complement is dense. Therefore E is nowhere dense in \ iff
int\ (cl\ E) g iff \ cl\ E is dense.
Intuitively, we can think of an open set S as including some elbow room around each of its
points if B S, then all sufficiently nearby points are also in S. Then we think of a nowhere
dense set as being skinny so skinny that not only does it contain no elbow room around
any of its points, but even its closure contains no elbow room around any of its points.

Example 6.2
1) A closed set J is nowhere dense in \ iff int J g iff \ J is dense in \ . In
particular, if a singleton set : is a closed set in \ , then : is nowhere dense unless : is
isolated in \
2) Suppose E E is nowhere dense iff cl E contains no interval of positive length.
For example, each singleton < is nowhere dense in . In particular, for 8 8} is nowhere
dense in . But notice that 8 is not nowhere dense in , since 8 is isolated in . The property
of nowhere dense is relative to the space in which a set lives.
If F E \ , then int\ cl\ F int\ cl\ E. Therefore if E is nowhere dense in
\ , then F is also nowhere dense in \ . However, the set F might not be nowhere dense in E.
For example, consider "
8" 8 is nowhere dense in .
and are not nowhere dense in (the awkward double negative in English is
one reason why some authors prefer to use the term rare for nowhere dense.

178

4) Since cl\ E cl\ cl\ E, the set on the left side has empty interior iff the set on the right side
has empty interior that is E is nowhere dense in \ iff cl\ E is nowhere dense in \ .
Theorem 6.3 Suppose F E \ g . If F is nowhere dense in E, then F is nowhere dense
in \ .
Proof If not, then there is a point B int\ cl\ F cl\ F. Since int\ cl\ F is an open set
containing B and B cl\ F , then g int\ cl\ F F .
So g int\ cl\ F F int\ cl\ F E cl\ F E clE F .
Since int\ cl\ F E is a nonempty open set in E, we conclude that intE clE F g, which
contradicts the hypothesis that F is nowhere dense in E.
The following technical results are sometimes useful for handling manipulations involving open
sets and dense sets.
Lemma 6.4 In any space \ g
1) If S open in \ and H is dense in \ , then cl S cl S H
2) If S is open and H is dense in \ , then S H is dense in S
3) If S is open and dense, and H is dense in \ , then S H is dense in \ . In particular,
the intersection of two and therefore finitely many dense open sets is dense. (This
result is not true for countable intersections; can you provide an example? )
Proof 1) We need to show that cl S cl S H Suppose B cl S and let Y be an open set
containing B. Then Y S g, so, since H is dense, we must have g Y S H
Y S H. Therefore B cl S H
2) Using part 1), we have clS S H cl\ S H S cl\ S S clS S S
3) If S is dense, then part i) gives clS H cl S \ .

Theorem 6.5 A finite union of nowhere dense sets in \ g is nowhere dense in \ .

Proof We will prove the result for the union of two nowhere dense sets. The general case
follows using a simple induction. If E" and E# are nowhere dense in \ , then
\ cl E" E# \ cl E" cl E# \ cl E" \ cl # .
Since the last two sets are open and dense, Lemma 6.4(3) gives that \ cl E" E# is dense.
Therefore E" E# is nowhere dense.
Theorem 6.5 is false for infinite unions: for each ; , ; is nowhere dense in , but
; ; is not nowhere dense in .
Definition 6.6 Suppose E \ g E is called a first category set in \ if E can be written as
a countable union of sets that are nowhere dense in X. If E is not first category in \ , we say that
E is second category in \ . (In some books, a first category set is called meager.)

179

If we think of a nowhere dense set in \ as skinny, then a first category set is a bit
larger merely thin.
Example 6.7
" If E is nowhere dense in \ , then E is first category in \ .
2) is a first category set in .
3 If F E \ and E is first category in \ , then F is first category in \ . However
F may not be first category in E, as the example ! ! " shows. However, using
Theorem 6.3, we can easily prove that if F is first category in E, then F is also first category in
\.
4) A countable union of first category sets in \ is first category in \ .
5) E ! 8" 8 is second category in E since any subset of E containing "
is not nowhere dense in E. But E is first category (in fact, nowhere dense) in .
6) Cardinality and category are totally independent ways to talk about the size of a set.
a) Consider the right-ray topology g g + + on .
is uncountable, but
8" J8 , where J8 8 Each J8 is nowhere
dense in g , so is first category in g .
b) On the other hand, is countable but (with the usual topology) is second
category in .
7) For 8 ", a straight line is nowhere dense in 8 , so a countable union of straight
lines in 8 is first category in 8 . (Aside: it follows from the Baire Category Theorem, below,
that the union of countably many straight lines cannot equal 8 . But nothing that fancy is
needed: can you give an argument that is based just on countable and uncountable sets to
show why a countable union of lines cannot 8 ? Can you extend your argument to show
why $ is not a countable union of planes?
More generally, if 5 8, a countable union of 5 -dimensional linear subspaces of 8 is
first category in 8 (Can a countable union of 5 -dimensional linear subspaces 8 ? )
It is not always easy to see what the category of a set is. Is a first or second category set in
? For that matter, is first or second category in ? (If you know the answer to either of these
questions, you also know the answer to the other: why? )

As we observed in Lemma 6.4, the intersection of two (and therefore, finitely many) dense open
sets is dense. However, in certain cases, the intersection of a countable collection of dense open
sets is dense. The following theorem discusses this condition and leads us to a definition.
Theorem 6.8 The following are equivalent in any space \ g :
1) If E is first category in \ , then \ E is dense in \
2) If K" , K# ,..., K5 ,... is a sequence of dense open sets, then
5" K5 is dense.
180

Proof (1 2) Let K" , K# ,..., K5 ,... be a sequence of dense open sets. Each \ K5 is closed

and nowhere dense, so

5" \ K5 is first category. By 1), \
5" \ K5
5" K5
is dense.
(2 " Let E be a first category set, say E
8" R5 , where each R5 is nowhere
dense. Then each \ cl R5 K5 is dense so, by 2),
5" \ cl R5 is dense. Since

5" \ cl R5 \
5" cl R5 \
5" R5 \ E, we see that \ E is dense.
Definition 6.9 A space \ g is called a Baire space if one (therefore both) of the conditions in
Theorem 6.8 hold.
Intuitively we can think of a Baire space as one which is thick a thin first category set E
can't fill up \ . In fact \ E is not merely nonempty but actually dense in \ . In fact, this is
often the way a Baire space is used in applications: if you want to prove that there is an element
in a Baire space having a certain property T , you can consider E B \ B does not have
property T . If you can then show that E is first category in \ , it follows that \ E g.

The following theorem tells us a couple of important things.

Theorem 6.10 Suppose S is an open set in a Baire Space \ g
1) If S g then S is second category in \ .
(In particular, a nonempty Baire space \ is second category in itself.)
2) S is a Baire space.
Proof 1) If S were first category in \ , then (by definition of a Baire space) \ S would be
dense in \ ; since \ S is closed, it would follow that \ S \ , and therefore S g
2) Suppose S is open in the Baire space \ , and let E be a first category set in S. We
must show that S E is dense in S. Suppose E
5" R5 , where R5 is nowhere dense in S.
By Theorem 6.3, R5 is also nowhere dense in \ . Therefore E is first category in \ so \ E is
dense in \ . Since S is open, S \ E S E is dense in S by part 2) of the Lemma 6.4.

Example 6.11 By Theorem 6.10.1, a nonempty Baire space is second category in itself. The
converse is false. To see this, let \ ! " #, with its usual topology. If we write
\
5" R5 , then the isolated point 2 must be in some R5 , so R5 is not nowhere dense.
Therefore \ is second category in itself. If \ were Baire, then, by Theorem 6.10, the open
subspace ! " would also be Baire and therefore second category in itself. However this is
false since ! " is a countable union of (nowhere dense) singleton sets.
To make use of properties of Baire spaces, it would be helpful to have a large supply of Baire
spaces. The next theorem provides us what we want.

181

Theorem 6.12 (The Baire Category Theorem) A complete metric space \ . is a Baire
space. (and therefore, by 6.10, a nonempty complete metric space is second category in itself.)
Proof If \ g then \ is Baire, so we assume \ g
Let E
5" R5 , where R5 is nowhere dense. We must show that \ E is dense in
\ . Suppose B! \ The closed balls of the form B .B! B % form a neighborhood base
at B! . Let J! be such a closed ball, centered at B! with radius % ! We will be done if we can
show J! \ E g We do that by using the Cantor Intersection Theorem.
Since int J! g (it contains B! ) and since R" is nowhere dense, we know that
int J! is not a subset of cl R" . Therefore we can choose a point B" in the open set
int J! cl R" . Pick a closed ball J" , centered at B" , so that B" J" int J! cl R" J! .
If necessary, choose J" even smaller so that diamJ" "#
(This is the first step of an inductive construction. We could now move to the induction step, but
actually include step two to be sure the process is clear. )
Since int J" g (it contains B" ) and since R# is nowhere dense, we know that
int J" is not a subset of cl R# . Therefore we can choose a point B# in the open set
int J" cl R# . Pick a closed ball J# , centered at B# , so that B# J# int J" cl R# J" .
If necessary, choose J# even smaller so that diamJ# "$

For the induction step: suppose we have defined closed balls J" J# J5 centered at points
B" B# B5 , with diamJ3 3 " " and so that B3 J3 int J3" cl R3 J3" .
Since int J5 g (it contains B5 ) and since R5" is nowhere dense, we know that int J5 is not a
subset of cl R5" . Therefore we can choose a point B5" in the open set int J5 cl R5" . Pick
a closed ball J5" , centered at B5" , so that B5" J5" int J5 cl R5" J5 . If necessary,
choose J5" even smaller so that diamJ5" 5 " # .

182

By induction, the closed sets J5 are defined for all 5 , diamJ5 ! and
J! J" J5 Since \ . is complete, the Cantor Intersection Theorem says that
there is a point B
5" J5 J! For each 5 ", B J5 int J5" cl R5 , so B cl R5 , so
B R5 Therefore B \
5" R5 \ E, so B J! \ E and we are done.
Example 6.13
1) Now we can see another reason why is not completely metrizable. Suppose . w is a
metric on that is equivalent to the usual metric . and for which . w is complete. By the
Baire Category Theorem, . w must be second category in itself. But this is false since the
space is topologically the same as the usual space of rationals.
2) Suppose \ . is a nonempty complete metric space without isolated points. We
proved in Theorem 3.6 that l\l - . We can now that, in addition, that each point of \ must be
a condensation point.
Otherwise, there would be a non-condensation point B! \ , and there would be some
countable open set S B! B" B# B8 Since each B8 is non-isolated, the singleton sets
B8 are nowhere dense in \ ., so S is first category in \ .. Since \ . is Baire, this
would mean that the closed set \ S is dense which is impossible.
Notice that this example illustrates again that cannot be completely metrizable: if its
topology were produced by a some complete metric, then each point in would have to be a
condensation point.
3) The set is a second category set in : if not, we could write , so that
would be a first category set in which contradicts the Baire Category Theorem. (Is second
category in ?
4) Recall that E \ g is called an J5 set if E can be written as a countable union of
closed sets, and that E is a K$ set if it can be written countable intersection of open sets. The
complement of a K$ set is an J5 set and vice-versa.
is not an J5 set in . To see this, suppose that
8" J8 , where the J8 's are
closed in . Since is second category in , one of these J8 's must be not nowhere dense in .
This J8 must therefore contain an open interval + ,. But then + , J8 , which is
impossible because there are also rational numbers in + ,
Taking complements, we see that is not a K$ -set in .

183

Example 6.14 Suppose 0 ! " is continuous. By the Fundamental Theorem of Calculus,

0 has an antiderivative 0" . Since 0" is differentiable (therefore continuous), it has an
antiderivative 0# . Continuing in this way, let 08 denote an antiderivative of 08" . It is a trivial
observation that:
b5 aB 05 B ! 0 B ! for all B ! "
We will use the Baire Category Theorem to prove a less obvious fact:
aB b5 05 B ! 0 B ! for all B in ! " (*)
In other words, if 0 B is not identically 0 on ! ", then there must exist a point B! ! " such
that every antiderivative 05 B! !

By hypothesis, we can write ! "

5" J5 , where J5 B ! " 05 B !
05" !. Since each 05 is continuous, J5 is closed in ! ". Since ! " is second category in
itself, one of the sets J5! is not nowhere dense and therefore must contain an interval + ,.
Since 05! is identically ! on + ,, 0 is identically ! on + ,
We can repeat the same argument on any closed subinterval N - . ! " letting
1 0 lN and 15 05 lN , we can conclude that N contains an open interval on which 0 is
identically !. In particular, each closed subinterval N - . contains a point B at which
0 B !. Therefore B 0 B ! is dense in ! ". Since 0 is continuous and is ! on a dense
set, 0 must be ! everywhere in ! ". See Theorem II.5.12 and Exercise III.E.16.)

Example 6.15 (The Banach-Mazur Game) The equipment for the game consists of two disjoint
sets E F for which E F ! " Set E belongs to Andy and set F belongs to Beth. Andy
gets the first move and selects a closed interval M" ! ". Beth then chooses a closed interval
M# M" , where M# has length "# Andy then selects a closed interval M$ M# , where M$ has
length "$ . They continue back-and-forth in this way forever. (Of course, to finish in a finite
time they must make their choices faster and faster.) When all is done, they look at

8" M8 B If B E, Andy wins; if B F Beth wins. We claim that if E is first category,

then Beth can always win.
th
Suppose E
choice be N
8" R5 , where E5 is nowhere dense in ! " Let Andy's 5
M#5" . Of course, N R5 is nowhere dense in ! "; but actually N R5 is also nowhere
dense in the interval N :

If intN clN N R5 g, then clN N R5 must contain an nonempty open

interval + , But then
+ , clN N R5 clN R5 cl!" R5 N cl!" R5 ,
which contradicts the fact that R5 is nowhere dense in ! ".
Then the open set int N clN N R5 is nonempty and Beth can make her 5 >2 choice M#5 to be
a closed interval + , int N clN N R5 . This implies that M#5 R5 g, since
M#5 int N clN N R5 int N N R5 N N R5 N R5

Therefore
5" M5
5" R5 g and Beth wins!

184

Example 6.16 The Baire Category Theorem can also be used to prove the existence of a
continuous function on ! " that is nowhere differentiable. The details can be found, for
example, in Willard's General Topology. In outline, one looks at the space G ! " with the
uniform metric 3, and argues that the set R of functions which have a derivative at one or more
points is a (thin) first category set in G ! " But the complete space G ! " 3 is a
(thick) Baire space. Therefore it is second category in itself so G ! " R g. In fact,
G ! " R is dense in G ! " Any function in G ! " R does the trick.

7. Complete Metrizability
Which metrizable spaces \ . are completely metrizable that is, when does there exist a
metric . w on \ with . . w and with \ . w complete? We have already seen that is not
completely metrizable (using the Baire Category Theorem), and that certain familiar spaces like
8" 8 are completely metrizable. In this section, we will give an answer to this question.
Lemma 7.1 If \ . is a pseudometric space and 1 \ continuous and g (a) 0. Then
1
g \ is continuous at a.
Proof The proof is a perfect mimic of the standard proof in advanced calculus of the same
result for \ .

Our first theorem tells us that certain subspaces of complete spaces are completely metrizable.
Theorem 7.2 If \ . is complete and S is open in \ , then there is a metric d w . on S such
that O, d w is complete that is, S is completely metrizable.
If S is not already complete with metric . , it is because there are some nonconvergent Cauchy
sequences in S. However those Cauchy sequences do have limits in \ they converge to points
outside S but in Fr S. The idea of the proof is to create a new metric . w on S equivalent of . but
which blows up near the boundary of S: in other words, the new metric destroys the
Cauchyness of the nonconvergent Cauchy sequences.
Here is a concrete (but slightly simpler) example to illustrate the idea.
Let . be the usual metric on the interval N 1# 1# The sequence B8 1# 8"
is a nonconvergent Cauchy sequence in N The function tan N is a
homeomorphism for which 0 B8
. w B C l tan B tan C l defines a new metric on N , and tan N . w . is an
isometry because . w B C ltan B tan Cl .tan B tan C Since . is
complete, so in N . w The sequence B8 is still nonconvergent in N . w because
. . w but B8 is no longer a Cauchy sequence in N . w
In the proof of Theorem 7.2, we will not have a homeomorphism like tan to use.
Instead, we define a continuous 0 S . and use 0 to create a new metric . w on S
that destroys Cauchy sequences B8 when they approach Fr S

185

Proof of 7.2 Define 0 B

"

.B \ S .

Since S is open, the denominator is never ! for B S

so 0 S is continuous by Lemma 7.1. On S, define . w B C .B C l0 B 0 Cl

It is easy to check that . w is a metric on S .
Since sequences are sufficient to determine the topology in metric spaces, we show that . w .
on S by showing that they produce the same convergent sequences in S. Suppose B8 is a
sequence in S and B S
If . w B8 B !, then since . w . we get . B8 B ! .
.

If . B8 B !, then l0 B8 0 Bl ! (by continuity) so . w B8 B !.

S . w is complete suppose B8 is a . w -Cauchy sequence in S. Since . w . B8 is also

.
. -Cauchy so B8 some B \ . But we claim that this B must be in S
If B \ S then .B8 \ S ! and so 0 B8 In particular, this means
that for every 8 we can find 7 8 for which 0 B7 0 B8 " Then
l0 B7 0 B8 l ", so . w B7 B8 ". This contradicts the assumption that B8 is
. w -Cauchy.
.
.w
Therefore B8 B S. Since . . w on S, B8 B S and so S . w is complete.
The next theorem generalizes this result. In fact, the interesting idea is in Theorem 7.2. The
proof of Theorem 7.3 just uses Theorem 7.2 repeatedly to patch together a more general result.
Theorem 7.3 (Alexandroff) If \ . is complete and E is a K$ in \ , then there is a metric 3 98
A, with 3 . on E and such that E 3 is complete. In other words: a K$ subset of a complete
space is completely metrizable.
w
Proof Let E S3 , where S3 is open in \ . Let .3 be a metric on S3 , equivalent to . on S3 ,

3"
w

such that S3 , .3 is complete. Let .3 B C min .3 B C ". Then .3 .3 . on S, and

w
S3 .3 is complete. (Note that .3 and .3 are identical for distances smaller than one: this is why
they produce the same convergent sequences and the same Cauchy sequences. )
For B C E, define 3B C .3 BC
23

(the series converges since all .3 B C ").

3"

3 is a metric on E:
Exercise
3 . on E:
We show that 3 and . produce the same convergent sequences in E
.
Suppose +8 + E S3 . Let % !.

186

Choose 5 so that

45"

1
24

%
2

. For every 3, . .3 on S3 , so

.3
+8 + S3 . Therefore, for 3 " ... , 5 , we can pick R3 so that 8 R3
.3 +283 + 2%5 . Then if 8 R max {R" ..., R5 , we get
5

3
3+8 ,+ .3 +283 + .3 +283 + 5 2%5 + 2% %, so +8 + E.
3"

35"

3
Conversely, if +8 + E, then for % !, we can pick R so that

8 R 3+8 +

3"

.3 +8 +
23

%
2

." +8 + %, so

."
.
+8 + S" But ." . on S" , so +8 +.
E 3 is complete:
Let +8 be 3-Cauchy in E. Let % ! and let 3 . Choose R3 so that
if 7 8 R3 , then 3+8 +7

%
23 ,

and therefore .3 +8, +7 %.

Therefore +8 is Cauchy in the complete space S3 .3 and there is a point

.3
.
B3 S3 such that +8 B3 S3 . But .3 . on S3 , so +8 B3 S3 \ .
This is true for each 3. But +8 can have only one limit + \ ., so we
conclude that B" B# ... B3 ... + Since + S3 for each 3, we
3
.
have + E, so (+8 + E. But . 3 on E, so (+8 + E. Therefore
E 3 is complete.

Corollary 7.4 is completely metrizable (and therefore is a Baire space, so is second

category in ).

Proof ; ;, so is an J5 set in . Taking complements, we get that

; ; is a K$ set in .
In fact, a sort of converse to Alexandroff's Theorem is also true.
Theorem 7.5 For any metric space \ ., the following are equivalent:

i) \ . is completely metrizable
ii) \ . is homeomorphic to a G$ set in some complete metric space ] . w
iii) If^ . ww is any metric space and 0 \ . 0 \ .ww ^ . ww is a
homeomorphism then 0 \ is a K$ set in ^ . ww
(so we say that \ is an absolute K$ set among metric spaces)

iv) \ is a K$ set in the completion \ , . .
Various parts of Theorem 7.5 are due to Mazurkiewicz (1916) & Alexandroff (1924).

187

The proof of one of the implications in Theorem 7.5 requires a technical result whose proof we
will omit. (See, for example Willard, General Topology)
Theorem 7.6 (Lavrentiev) Suppose \ . and ] . w are complete metric spaces with E \
and F ] . Let 2 be a homeomorphism from E onto F. Then there exist K$ sets E w in \ and
F w in ] with E E w cl E and F F w cl F and there exists a homeomorphism
2 w from E w onto F w such that 2 w lE 2 (Loosely stated: a homeomorphism between subsets of
two complete metric spaces can always be extended to a homeomorphism between K$ sets.)
Assuming Lavrentiev's Theorem, we now prove Theorem 7.5.
Proof i) ii) Suppose \ . is completely metrizable and let . w be a complete metric on \
equivalent to . . Then the identity map 3 \ . \ . w is a homeomorphism and 3\ \
is certainly a K$ set in \ . w
ii) iii) Suppose we have a homeomorphism 1 \ . 1\ E, where E is a K$
ww
set in a complete space ] . w Let 0 \ . ^ . ww the completion ^ , . be a
homeomorphism (into ^ ). We want to prove that F 0 \ must be a K$ set in ^ .
We have that 2 0 1" E F is a homeomorphism. Using Lavrentiev's Theorem, we get an
extension of 2 to a homeomorphism 2 w E w F w where E w , F w are K$ sets with E E w ]

and F F w ^ (see the figure).

Since E is a K$ in ] there are open sets S8 in ] such that E

8" S8

w
w

E w
S

E
K
E
But 2 w E w F w is a
.
Therefore
is
also
a
in
$
8
8" 8
8"
homeomorphism, so 2 w E 2E F is a K$ set in F w . Therefore F
8" Z8 , where Z8 is

open in F w and, in turn, Z8 F w [8 where [8 is open in ^ .

Also, F w is a K$ in ^ , so F w
8" Y8 , where the Y8 's are open in ^ .

188

Putting all this together, F

8" Z8
8" [8 F

w
8" [8 F 8" [8 8" Y8 . The last expression shows that F is a K$ set in

^ . But then F ^ F ^
8" [8 8" Y8 8" [8 ^ 8" Y8 ^
is a K$ set in ^ .

iii) iv) iii) states that any homeomorphic copy of \ in a metric space ^ . ww must

be a K$ in ^ . Letting ^ . ww \ , . , it follows that \ must be a K$ set in the completion

\ , . .
iv) i) This follows from Alexandroff's Theorem 7.3: a K$ set in a complete space is
completely metrizable.

Theorem 7.5 raises the questions: what spaces are absolutely open? what spaces are
absolutely closed? Of course in each case a satisfactory answer might involve some kind of
qualification. For example, among Hausdorff spaces, a space \ is absolutely closed iff ...
Example 7.7 Since is not a K$ in (an earlier consequence of the Baire Category Theorem),
Theorem 7.5 gives us yet another reason why is not completely metrizable.
Example 7.8 Since is completely metrizable, it follows that is a Baire space. If . is the
usual metric on , then . is an example of a Baire metric space that isn't complete.

To finish this section on category, we mention, without proof and just as a curiosity, a
generalization of an earlier result (Blumberg's Theorem: see Example II.5.8).
Theorem 7.9 Suppose \ . is a Baire metric space. For every 0 \ , there exists a dense
subset H of \ such that 0 lH H is continuous,
The original theorem of this type was proved for \ or # \ # (Blumberg, New
properties of all real-valued functions, Transactions of the American Mathematical Society,
24(1922) 113-128).
The more general result stated in Theorem 7.9 is (essentially) due to Bradford and Goffman
(Metric spaces in which Blumberg's Theorem holds, Proceedings of the American Mathematical
Society 11(1960), 667-670)

189

Exercises
E11. In a pseudometric space \ ., every closed set is a K$ set and every open set is an J5
set (see Exercise II.E.15).
a) Find a space \ g which contains a closed set J that is not a K$ set.
b) Recall that the scattered line is the space g where
g Y Z Y is a usual open set in and Z
Prove that the scattered line is not metrizable. (Hint: K$ or J5 sets are relevant.)
E12. a) Prove that if O is open in \ g , then Fr S is nowhere dense.
b) Suppose that D is a discrete subspace of the Hausdorff space \ g and that \ has no
isolated points. Prove that H is nowhere dense in \ .
E13. Let g be the confinite topology. Prove that \ g is a Baire space if and only if \ is
either finite or uncountable.

E14. Suppose . is any metric on which is equivalent to the usual metric on . Prove that the
~ ~
completion . is uncountable.

E15. Suppose that \ . is a nonempty complete metric space and that Y is a family of
continuous functions from \ to with the following property:
aB \ b a constant QB such that l0 Bl QB for all 0 Y
Prove that there exists a nonempty open set Y and a constant Q (independent of B) such that
l0 Bl Q for all B Y and all 0 Y .
This result is called the Uniform Boundedness Principle.
Hint: Let I5 B \ l0 Bl 5 for all 0 Y . Use the Baire Category Theorem.

190

E16. Suppose 0 \ g ] ., where ] . is a metric space where . is a metric. For each

B \, define
=0 B the oscillation of 0 at B inf {diam0 R R is a neighborhood of B
a) Prove that 0 is continuous at + if and only if =0 + !.
b) Prove that for 8 , B \ =0 B 81 is open in \ .
c) Prove that B \ 0 is continuous at B is a K$ -set in \ .
Note: Since is not a K$ -set in , a continuous function 0 cannot have as its set of
points of continuity. On the other hand, is a K$ -set in and, from analysis, you should
probably know an example of a continuous function 0 that is continuous at each :
and discontinuous at each point ; .
d) Prove that there cannot exist a function 0 ! such that for all B and all
C , 0 B0 C B C Hint for d): Suppose 0 exists.
i) First prove that if (B8 ) is a sequence of rationals converging to an irrational,
then 0 B8 ! and, likewise, if C8 is a sequence of irrationals converging
to a rational, then 0 C8 !.
ii) Define a new function 1 by 1 ! and 1C 0 C for C
irrational. Examine the set of points where 1 is continuous. )
E17. There is no continuous function 0 for which the set of points of continuity is .
The reason (see problem E16) ultimately depends on the Baire Category Theorem. Find the error
in the following more elementary proof:
Suppose 0 is continuous and that 0 is continuous at B iff B
"
Then B =0 B !
8" Y8 , where Y8 B =0 B 8
Each Y8 , and Y8 can be written as a countable union of disjoint open intervals. If
+ , and - . are consecutive intervals in Y8 , then , - or else there would be a
rational in , -. Therefore Y8 consists only of the endpoints of some disjoint open
intervals. Therefore Y8 is countable.

It follows that B 0 is not continuous at B

8" Y8
8" Y8 is
countable. Therefore 0 is continuous on more than just the points in .

E18. Suppose that for each irrational :, we construct an equilateral triangle X: (including its
interior) in # with one vertex at : ! and its opposite side above and parallel to the B-axis.
Prove that X: : must contain an open box of the form + , ! 51 for some
+ , and some 5 . In the hypothesis, we could weaken equilateral to read ... ?
(Hint: Consider R5 : X: has height 5" .)

191

E19. Suppose 0 . Prove that if 0 is discontinuous at every irrational : , then there

must exist an interval M + , such that 0 is discontinuous at every point in M
E20. \ g is absolutely open if whenever 0 \ g ] g w is a homeomorphism
(into), then 0 \ must be open in ] . Find all absolutely open spaces.

192

8. Compactness
Compactness is one of the most important topological properties. Formally, it is a strengthening
of the Lindelof
property.
Compact spaces are often particularly simple to work with because of the rule of thumb that
compact spaces often act like finite spaces.
Definition 8.1 A topological space \ g is called compact if every open cover of \ has a
finite subcover.
Example 8.2
1) A finite space is compact.
2) Any set \ with the cofinite topology is compact because any nonempty open set,
alone, covers \ except for perhaps finitely many points.
3) An infinite discrete space is not compact because the cover consisting of all
singleton sets has no finite subcover. In particular, 8" 8 is not compact.
4) The space \ ! 8" 8 is compact: if h is any open cover, and
! Y h, then the set Y covers \ except for perhaps finitely many points.
5) is not compact since h 8 8 8 has no finite subcover.

Definition 8.3 A family Y of sets has the finite intersection property (FIP) if every finite
subfamily of Y has nonempty intersection.
Sometimes it is useful to have a characterization of compact spaces in terms of closed sets, and
we can do this using FIP. As you read the proof, you should see that Theorem 8.4 is hardly any
more than a reformulation, using complements, of the definition of compactness in terms of
closed sets.
Theorem 8.4 \ g is compact iff whenever Y is a family of closed sets with the finite
intersection property, then Y g.
Proof Suppose \ is compact and that Y is a family of closed sets with FIP. Let
h \ J J Y . For any J" J# J8 Y , the FIP tells us that \ \ 83" J8
\ J" \ J8 In other words, no finite subcollection of h covers \ . Therefore
h itself cannot cover \ , that is, \ h \ Y Y h Y g.
The proof of the converse is similar.
Theorem 8.5 For any space \ ,g
a) If \ is compact and J is closed in \ , then J is compact.
b) If \ is a Hausdorff space and O is a compact subset, then O is closed in \ .
Proof a) The proof is like the proof that a closed subspace of a Lindelof
space is Lindelof.

(Theorem III.3.10) Suppose J is a closed set in a compact space \ , and let h Y E

be a cover of J by sets open in J . For each , pick an open set Z in \ such that Y Z J
Since J is closed, the collection i \ J Z E is an open cover of \ .
193

Therefore we can find Z" Z8 so that \ J Z" Z8 covers \ . Then Y" Y8

covers J , so J is compact.
Note: the definition of compactness requires that we look at a cover of J by sets open in J ,
but it should be clear that it is equivalent to look at a cover of J using sets open in \
b) Suppose O is a compact set in a Hausdorff space \ . Let C \ O For each
B O , we can choose disjoint open sets YB and ZB in \ with B YB and C ZB . Since
h YB B O covers O there are finitely many points B" B8 O so that YB" YB8
covers O . Then C Z ZB" ZB8 \ O . Therefore \ O is open, so O is closed.

Notes 1) Reread the proof of part b) assuming O is a finite set. This highlights how
compactness has been used in place of finiteness and illustrates the rule of thumb that
compact spaces often behave like finite spaces.
2) The Hausdorff hypothesis cannot be omitted in part b). For example, let \ be a set
with l\l " and let . be the trivial pseudometric on \ . Then each singleton set B is compact
but not closed.
3) Compactness is a clearly a topological property, so part b) implies that if a compact
space \ is homeomorphic to a subspace O in a Hausdorff space ] , then O is closed in ] . So we
can say that a compact space is absolutely closed among Hausdorff spaces it's closed
wherever you put it. In fact, the converse is also true among Hausdorff spaces a space which
is absolutely closed is compact but we do not have the machinery to prove that now.
Corollary 8.6 A compact metric space \ . is complete.

Proof \ . is a dense subspace of its completion \ , . . But \ . is compact, so \ must

be closed in \ . Therefore \ . \ , . , so \ . is complete.
(Note: this proof doesn't work for pseudometric spaces (why?). But is Corollary 8.6 still true
for pseudometric spaces? Would a proof using the Cantor Intersection Theorem work in that
case?)
You may have seen a different definition of compactness in some other book. In fact there are
several different kinds of compactness and, in general they are not equivalent. But we will see
that they are equivalent in certain spaces for example, in 8
Definition 8.7 A topological space \ g is called
sequentially compact
countably compact
pseudocompact

if every sequence in \ has a convergent subsequence in \

if every countable open cover of \ has a finite subcover
(therefore Lindelof
+ countably compact compact )
if every continuous 0 \ is bounded (You should check
that this is equivalent to saying that every continuous real
-valued function on \ assumes both a maximum and a
minimum value).

We want to look at the relations between these compactness-like properties.

194

Lemma 8.8 If every sequence in \ g has a cluster point, then every infinite set in \ has a
limit point. The converse is true if \ g is a X" -space that is, if all singleton sets B are
closed in \ .
Proof Suppose E is an infinite set in \ . Choose a sequence +8 of distinct terms in E and let
B be a cluster point of +8 Then every neighborhood R of B contains infinitely many +8 's so
R E B g Therefore B is a limit point of E.
Suppose \ is a X" -space in which every infinite set has a limit point. Let B8 be a
sequence in \ . We want to show that B8 has a cluster point. Without loss of generality, we
may assume that the terms of the sequence are distinct (why?), so that E B8 8 is
infinite Let B be a limit point of the set E. We claim B is a cluster point for B8
Suppose Y is an open set containing B and 8 . Let Z Y B" B8 B
Since B" B8 is closed, Z is open and B Z Y Then Z E B g, so
Y E B g Therefore Y contains a term B5 for some 5 8 so B is a cluster point
of B8
Example 8.9 The X" hypothesis in the second part of Lemma 8.8 cannot be dropped. Let
\ 8" 8 :8 8 where the :8 's are distinct points and :8 5" for all 8 5
The idea is to make .:8 B8 ! for each 8 so that :8 is a sort of double for 8" . So we define
.:8 B8 !
" "
"
"
. 8 7 l 8 7 l .:8 :7 if 7 8
.:8 " l " " l if 7 8
7
8
7

It is easy to check that . is a pseudometric on \ . In \ ., every nonempty set E (finite or

infinite) has a limit point if B E, then its double is a limit point of E. But the sequence 8"
has no cluster point in \ .

Theorem 8.10 \ g is countably compact iff every sequence in \ has a cluster point
(Then by Lemma 8.8, a X" -space \ g is countably compact iff every infinite set in \ has a
limit point).
Proof Suppose \ g is countably compact and consider any sequence B8 . Let X8 the
8>2 tail of B8 B5 5 8 We claim that
8" cl X8 g
Assume not. Then every B is in \ cl X8 for some 8, so \ cl X8 8
is an open cover of \ . Since \ is countably compact, there exists an R such
that \ \ cl X" \ cl XR . Taking complements gives
g R
8" cl X8 cl XR , which is impossible.

Let B
8" cl X8 and let R be a neighborhood of B. Then R X8 g for every 8, so R
contains an B5 for arbitrarily large values of 8. Therefore B is a cluster point of B8
Conversely, suppose \ is not countably compact. Then \ has a countable open cover
Y" Y8 with no finite subcover. For each 8, pick a point B8 \ 83" Y3 . Then the
sequence B8 has no cluster point: for any B \ , we know B is in some Y8 . Y8 is a
neighborhood of B and B7 Y8 for 7 8.

195

The next theorem tells us some connections between the different types of compactness.
Theorem 8.11 In any space \ g , the following implications hold:
\ is compact
or

\ is sequentially compact

\ is countably compact

\ is pseudocompact

Proof It is clear from the definitions that a compact space is countably compact.
Suppose \ is sequentially compact. Then each sequence B8 in \ has a subsequence
that converges to some point B \ . Then B is a cluster point of B8 . From Lemma 8.8 we
conclude that \ is countably compact.
Suppose \ is countably compact and that 0 \ is continuous. The sets
Y8 0 " 8 8 B \ 8 0 B 8 form a countable open cover of \ .
Y" YR cover \ for some R . Since Y" Y# YR , this implies that YR \ . So
R 0 B R for all B \ that is, 0 is bounded. Therefore \ is pseudocompact.

So

In general, no other implications hold in among these four types of compactness, but we do not
have the machinery to provide counterexamples now. ( See Corollary 8.5 in Chapter VIII and
Example 6.5 in Chapter X ). We will prove, however, that they are all equivalent in a
pseudometric space \ . Much of the proof is developed in the following sequence of
lemmas some of which have intrinsic interest of their own.
Lemma 8.12 If \ is first countable, then \ is sequentially compact iff \ is countably compact.
(So, in particular, sequential and countable compactness are equivalent in pseudometric spaces.)
Proof We know from Theorem 8.11 that if \ is sequentially compact, then \ is countably
compact. So assume \ is countably compact and that B8 is a sequence in \ . By Theorem
8.10, we know B8 has a cluster point, B. Since \ is first countable, there is a
subsequenceB85 B (see Theorem III.10.6). Therefore \ is sequentially compact.
Definition 8.13 A pseudometric space \ . is called totally bounded if for each % ! \ can
be covered by a finite number of %-balls. If \ g, this is the same as saying that for % !, there
exist B" B# B8 \ such that \ F% B" F% B8 .
More informally, a totally bounded pseudometric space is one that can be kept under full
surveillance using a finite number of policemen with an arbitrary degree of nearsightedness.

Example 8.14
1) Neither nor (with the usual metric) is totally bounded since neither can be covered
by a finite number of "-balls.

196

2) A compact pseudometric space \ . is totally bounded: for any % !, we can

pick a finite subcover from the open cover h F% B B \
Lemma 8.15 If \ . is countably compact, then \ . is totally bounded.
Proof If \ g, \ is totally bounded, so assume \ g If \ . is not totally bounded, then
for some % !, no finite collection of %-balls can cover \ . Choose any point B" \ . Then we
can choose a point B# so that .B" B# % (or else F% B" would cover \ ).
We continue inductively. Suppose we have chosen points B" B# B8 in \ such that
.B3 B4 % for each 3 4 3 4 " 8. Then we can choose a point B8" at distance %
from each of B" B8 because otherwise the %-balls centered at B" B8 would cover \ .
The sequence B8 chosen in this way cannot have a cluster point because, for any B,
F #% B contains at most one B8 . Therefore \ is not countably compact.
Lemma 8.16 A totally bounded pseudometric space \ . is separable.
Proof For each 8, choose a finite number of 8" -balls that cover \ and let H8 be the (finite) set
consisting of the centers of these balls. For any C \ .B C 8" for some B H8 This
means that H
8" H8 is dense. Since H is countable, \ is separable.
Theorem 8.17 In a pseudometric space \ ., the properties of compactness, countable
compactness, sequential compactness and pseudocompactness are equivalent.
Proof We already have the implications from Theorem 8.11.
If \ . is countably compact, then \ . is totally bounded (see Lemma 8.15) and
therefore separable (see Lemma 8.16). But a separable pseudometric space is Lindelof
(see
Theorem III.6.5) and a countably compact Lindelof
space is compact. Therefore compactness and
countable compactness are equivalent in \ .
We observed in Lemma 8.12 that countable compactness and sequential compactness are
equivalent in \ ..
Since a countably compact space is pseudocompact, we complete the proof by showing
that if \ . is not countably compact, then \ . is not pseudocompact. This is the only part
of the proof that takes some work. (A bit of the maneuvering in the proof is necessary because .
is a pseudometric. If . is actually a metric, some minor simplifications are possible. )
If \ . is not countably compact, we can choose a sequence B8 with no cluster point
(see Theorem 8.10). In fact, we can choose this B8 so that all the B8 's are distinct and
.B7 B8 ! if 7 8 why?. We can then find open sets Y8 such that i) B8 Y8 ,
ii) Y8 Y7 g for 7 8, and so that iii) diam Y8 ! as 8
Here is a sketch for finding the Y8 's; the details are left to check as an exercise.
First check that for any + , \ and positive reals <" <# if <" <# .+ , then
F<" + and F<# , are disjoint in fact, they have disjoint closures.
Since B8 has no cluster point we can find, for each 8, a ball F$8w B8 that contains
w
w
no other B7 that is, .B8 B7 $8 for all 7 8. Let $8 $8 #. Then for
7 8 we have B7 F$8 B8 and $8 .B8 B7 Of course, two of these balls

197

might overlap. To get the Y8 's we want, we shrink these balls (choose smaller radii
%8 to replace the $8 ) to eliminate any overlap. We define the %8 's inductively:
Let
Pick

%" $" .B" B#

%# ! so that
%" %# .B" B# and
%# $#

Since
%" $" .B" B$ and
%# $# .B# B$ we can
pick %$ ! so that
%" %$ .B" B$ and
%# %$ .B# B$ and
%$ $$
Continue in this way. At the 8th step, we get a new ball F%8 B8 for which
if 7 8, B7 F%8 B8 (because %8 $8 and
F%8 B8 F%4 B4 g for all 4 8 (because %4 %8 .B4 B8
Finally, when we choose %8 at each step, we can also add the condition that
%8 8" , so that diamF%8 B8 !.
Then we can let Y8 F%8 B8
8.B \ Y

For each 8, define 08 \ by 08 B .B \ Y 8 . Since B8 Y8 and \ Y8 is closed,

8
8
the denominator of 08 B is not !. Therefore 08 is continuous, and 08 B8 8 Define
0 B 08 B.

(For any B \ , this series converges because B is in at most one Y8 and

8"

therefore at most one term 08 B !.) Then 0 is unbounded on \ because 0 B8 08 B8

8. So we are done if we can show that 0 is continuous at each in \
Let + \ . We claim that there is an open set Z+ containing + such that Z+ Y8 g for at most
finitely many 8's.
If .+ B8 ! for some B8 , we can simply let Z+ Y8 .
Suppose .+ B8 ! for all 8. Since + is not a cluster point of B8 , we can choose
% ! so that F% + contains none of the B8 's. For this %, choose R so that if 8 R
then diam Y8 #% . Then F #% + Y8 g for 8 R (if 8 R and D F #% + Y8 ,
then .+ B8 .+ D .D B8 #% #% %). Then let Z+ F #% +
Then 08 lZ+ is identically ! for all but finitely many 8. Therefore 0 lZ+ is really only a finite sum
of continuous functions, so 0 lZ+ Z+ is continuous. Suppose [ is any neighborhood of
0 +. Then there is an open set Y in Z+ containing + for which 0 lZ+ Y [ . But Z+ is open
in \ so Y is also open in \ , and 0 Y 0 lZ+ Y [ . So 0 is continuous at +.
198

We now explore some properties of compact metric spaces.

Theorem 8.18 Suppose E is a compact subset of a metric space \ . Then E is closed and
bounded (that is, E has finite diameter).
Proof A compact subset of the Hausdorff space must be closed (see Theorem 8.5b).
If E g, then E is certainly bounded. If E g, choose a point +! E and let
0 E by 0 + .+ +! . Then 0 is continuous and 0 is bounded (since \ is
pseudocompact): say l0 Bl Q for all B E. But that means that E FQ +! , so E is a
bounded set .
Caution: the converse of Theorem 8.18 is false. For example, suppose . is the discrete unit
metric on an infinite set \ . Every subset of \ . is both closed and bounded, but an infinite
subspace of \ is not compact. However, the converse to Theorem 8.18 is true in 8 , as you may
remember from analysis.
Theorem 8.19 Suppose E 8 . The E is compact iff E is closed and bounded.
Proof Because of Theorem 8.18, we only need to prove that if E is closed and bounded, then E
is compact. First consider the case 8 ". Then E is a closed subspace of some interval
M + , and it is sufficient to show that M is compact. To do this, let B8 be a sequence in M
and choose (see Lemma 2.10) a monotone subsequence B85 . Since B85 is bounded, we know
that B85 some point < M (see Lemma 2.9). Since B8 has a convergent subsequence, M is
sequentially compact and therefore compact.
When 8 ", the proof is similar. We illustrate for 8 2. If E is a closed bounded set in
# , then E is a closed subspace of some closed box of the form M + , - .. Therefore it is
sufficient to prove that M is compact. To do this, choose a sequence B8 C8 in M . Since + , is
compact, the sequence B8 has a subsequence B85 that converges to some point B + , .
Now consider the subsequence B85 C85 in M . Since - . is compact, the sequence C85 has a
subsequence C856 that converges to some point C - .. But in # , ?8 @8 ? @ iff
?8 ? and @8 @ in . So B856 C856 B C M This shows that M is sequentially
compact and therefore compact.
For 8 #, a similar argument clearly works it just involves taking subsequences 8
times.

199

9. Compactness and Completeness

We already know that a compact metric space \ . is complete. Therefore all the big
theorems that we proved for complete spaces are true in a compact metric space \ . for
example Cantor's Intersection Theorem, the Contraction Mapping Theorem and the Baire
Category Theorem (which implies that a nonempty compact metric space is second category in
itself).
Of course, a complete space \ . need not be compact: , for example. We want to see the
exact relationship between compactness and completeness.
We begin with a couple of preliminary results.
Theorem 9.1 If \ . is totally bounded, then \ . is bounded.
Proof For % ", pick points B" B8 so that \ F" B" F" B8 Let
Q max .B3 B4 3 4 " 8. If B C \ , then we have B F" B3
and C F" B4 for some 3 4 , so that
.B C .B B3 .B3 C4 .C4 C " Q " Q #.
Therefore \ . has finite diameter that is \ . is bounded.
Notice that the converse to the theorem is false: total boundedness is stronger than
boundedness. For example, let . w B C min " lB Cl on . Then diam . w " so
( . w is bounded. Because . w and the usual metric . agree for distances smaller that ",
w
F". B B " B ". Since cannot be covered by a finite number of these "-balls,
( . w is not totally bounded.

Theorem 9.2 If \ . is totally bounded and E \ then E . totally bounded: that is, a
subspace of a totally bounded space is totally bounded.
Proof
Let % ! and choose B" B8 so that \ F % B" F #% B8 . Of course these
#
balls also cover E. But to show that E . is totally bounded, we need to show that we can cover
E with a finite number of %-balls centered at points in E
Let N 4 F #% B4 E g and, for each 4 N , pick +4 F #% B4 E
We claim that the balls F%E +4 cover E. In fact, if + E, then + F #% B4 for some 4;
and for this 4, we also have +4 F #% B4 Then .+ +4 .+ B4 .B4 +4 #% #% %, so
+ F%E +4 .
The next theorem gives the exact connection between compactness and completeness. It is
curious because it states that compactness (a topological property) is the sum of two
nontopological properties.

200

Theorem 9.3 \ . is compact iff \ . is complete and totally bounded.

Proof ( ) We have already seen that the compact space \ . is totally bounded and
complete. (For completeness, here is a fresh argument: let B8 be a Cauchy sequence in \ ..
Since \ is countably compact, B8 has a cluster point B in \ . But a Cauchy sequence must
converge to a cluster point. So \ . is complete.)
Let B8 be a sequence in \ .. We will show that B8 has a Cauchy
subsequence which (by completeness) must converge. That means \ . is sequentially
compact and therefore compact. Without loss of generality, we can assume that the terms of the
sequence B8 are distinct (why?).
Since \ . is totally bounded, we can cover \ with a finite number of 1-balls, and one of
them call it F" must contain infinitely many B8 's. Since F" . is totally bounded (see
Theorem 9.2) we can cover F" with a finite number of "# -balls and one of them call it
F# F" must contain infinitely many B8 's.
Continue inductively in this way.
At the induction step, suppose we have 5" -balls
F" F# F5 where each ball contains infinitely many B8 's. Since F5 . is totally
bounded, we can cover it with a finite number of 5 " " -balls, one of which call it
F5" F5 must contain infinitely many B8 's. This inductively defines an infinite descending
sequence of balls F" F# F5 where each F5 contains B8 for infinitely many
values of 8.
B 8" F"
B8# F# , with 8# 8"

B85 F5 , with 85 85" 8"

The subsequence B85 is Cauchy since B86 F5 for 6 5 and diam F5 !

Then we can choose

Example 9.4 With the usual metric . :

i) is complete and not compact (and therefore not totally bounded).
ii) 8" 8 is totally bounded, because it is a subspace of the totally
bounded space ! ". But it is not compact (and therefore not complete).
iii) and 8" 8 are homeomorphic because both are countable discrete
spaces. Total boundedness is not a topological property.
iv) 8" 8 ! is compact and is therefore both complete and totally
bounded.
During the earlier discussion of the Contraction Mapping Theorem, we defined uniform
continuity. For convenience, the definition is repeated here.
Definition 9.5 A function 0 \ . ] . w is uniformly continuous if
a% b$ aB aC \ (.B C $ . w 0 B 0 C %

201

Clearly, if \ B" B8 is finite, then every continuous function 0 \ . ] = is

uniformly continuous. (For % ! and each 3 " 8, we pick the the $3 that works at B3 in the
definition of continuity. Then $ min $" $8 works uniformly across the space \ .)
The next theorem generalizes this observation to compact metric spaces \ . and illustrates
once more rule of thumb that compact spaces act like finite spaces.
Theorem 9.6 If \ . is compact and 0 \ . ] . w is continuous, then 0 is uniformly
continuous.
Proof Suppose 0 is not uniformly continuous. Then for some % !, no $ works. In particular,
for that %, $ 8" doesn't work therefore we can find, for each 8, a pair of points ?8 and @8
with .?8 @8 8" but . w 0 ?8 0 @8 % Since \ is compact, there is a subsequence
?85 B \ . It follows that the corresponding subsequence @85 B also because
.@85 B .@85 ?85 .?85 B 8"5 .B85 B ! Therefore, by continuity, we should
have 0 ?85 0 B and also 0 @85 0 B But this is impossible since
. w 0 ?85 0 @85 % for all 5 .
Uniform continuity is a strong and useful condition. For example, the preceding theorem implies
that a continuous function 0 + , must be uniformly continuous. This is one of the
reasons why in calculus continuous functions on closed intervals are so nice to work with. The
following example is a simple illustration of how uniform continuity might be used in a simple
analysis setting.
Example 9.7 Suppose 0 + , be bounded. Choose any partition of + ,
B! + B" B3" B3 B8 ,
and let Q3 and 73 denote the sup and inf of the set 0 B3" B3 . Let ?3 B B3 B3" The
sums
8
8
73 ?3 B Q3 ?3 B
3"

3"

are called the lower and upper sums for 0 associated with this partition.
It is easy to see that the sup of the lower sums (over all possible partitions) is finite: it is called
,
the lower integral of 0 on + , and denoted + 0 B .B. Similarly, the inf of all the upper sums is

,
finite: it is called the upper integral of 0 on + , and denoted + 0 B .B It is easy to verify that
+, 0 B .B +, 0 B .B If the two are actually equal, we say 0 is (Riemann) integrable on + ,

,
,
,
and we write + 0 B .B + 0 B .B + 0 B .B

Uniform continuity is just what we need to prove that a continuous 0 on + , is integrable. If 0 is

continuous on + ,, then 0 is uniformly continuous. Therefore, for % !, we can choose $ !
such that l0 B 0 Cl , % + whenever lB Cl $ Pick a partition of + , for which each
?3 B $ Then for each 3, Q3 73 , % + Therefore

202

+, 0 B .B +, 0 B .B Q3 ?3 B 73 ?3 B
8

Q3 73 ?3 B % ?3 B
8

3"

3"

%
, + ,

+ %

Since % ! was arbitrary, we conclude that + 0 B .B + 0 B .B.

3"

3"

10. The Cantor Set

The Cantor set is an example of a compact subspace of with a surprising combination of
properties. Informally, we can construct the Cantor set G as follows. Begin with the closed
interval ! " and delete the open middle third. The remainder is the union of 2 disjoint closed
intervals: ! "$ #$ ". At the second stage of the construction, we then delete the open middle
thirds of each interval leaving a remainder which is the union of 2# disjoint closed intervals:
! "* #* "$ #$ (* )* ". We repeat this process of deleting the middle thirds forever.
The Cantor set G is the set of survivors the points that are never discarded.
Clearly there are some survivors: for example, the endpoint of a deleted middle third for
example, "* clearly survives forever and ends up in G .
To make this whole process precise, we name the closed subintervals that remain after each stage.
Each interval contains the # new remaining intervals below it:
! "$
J!
! "*
J!!

#$ "
J#
#* "$
J!#

"
# "
! #(
#(
*
(J!!! ) (J!!# )

(
) "
#* #(
#(
$
(J!#! ) J!##

#$ (*
J#!

)* "
J##

"*
#! (
$# #(
#(
*
J#!! J#!#

#&
#'
*) #(
#(
"
J##! J###

etc.

etc.

At the 5 >2 stage the set remaining is the union #5 closed subintervals each with length
label them J8" 8# 85 where 8" 8# 85 ! #5

"
.
$5

We

Notice that, for example, J! J!# J!#! J!#!! J!#!!# . As we go down the chain
toward the Cantor set, each new ! or # in the subscript indicates whether the next set down is
the remaining left interval or the remaining right subinterval.
For each sequence = 8" 8# 85 ! # , we have
J8" J8" 8# J8" 8# 85

Since ! " is complete, the Cantor Intersection Theorem gives that

5" J8" 8# 85 contains a
single point of ! ", and we call that point B= . The Cantor set G is then defined by

203

G B= = ! # ! ".
If = = w 8"w 8#w 85w ! # , we can look at the first 5 for which 85 8w5 . Then
B= J8" 85" 85 and B=w J8" 85" 8w5 . Since these closed intervals are disjoint, we conclude
B= B=w . Therefore each sequence = ! # corresponds to a different point in G , so
lGl l! # l #i! - . Since G ! ", we conclude lGl - .
Each B ! " can be written using a ternary decimal expansion B $>88 !>" ># >5 three ,

8"

where each >5 ! " # Just as with base 10 decimals, the expansion for a particular B may not
be unique: for example, "$ !"!!!three but also "$ !!#####three It's easy to check that
an B ! " has two different ternary expansions iff B is the endpoint of one of the deleted open
middle thirds in the construction. Therefore it is easy to see that a point is in G iff it has a
ternary expansion involving only 0's and 2's. For example, "$ G and "# G
What are some of the properties of G ?
i) ! " G is the union of the deleted open middle third intervals, so ! " G is
open in ! ". Therefore G is closed in ! ", so G is a compact metric space (and therefore
complete in the usual metric).
ii) Every point in G is a limit point of G that is G has no isolated points. (Note: such
a space is sometimes called dense-in-itself. This is an awkward but well-established term; it is
awkward because it means something different from the obvious fact that every space \ is a
dense subset of itself.)
Suppose B= G , where = 8" 8# 85 85" , 85# ! # . Given % !,
pick 5 so $"5 %. Define 8w5" ( ! or #) so that 8w5" 85" and let
w
=w 8" 8# 85 85"
852 ! #i! .

Then B= and B=w are distinct points in G , but both are in J8" 85 and diam J8" 85
Therefore lB= B=w l %, so B= is not isolated in G .

"
$5

%.

iii) With the usual metric ., G is a nonempty complete metric space with no isolated
points. It follows from Theorem 3.6 that lGl - and, since G ! ", then lGl - However,
we can also see this from the discussion in ii): there are #i! - different ways we could redefine
the infinite tail 85" 856 ) of = and each version produces a different point in G at
distance % from B=
iv) In some ways, G is big. For example, the Cantor set has as many points as ! ".
Also, since G is complete with the usual metric, G is second category in itself. But G is small
in other ways. For one thing, G is nowhere dense in ! " (and therefore in . Since G is
closed, this simply means that G contains no nonempty open interval.
To see this, suppose M + , G . At the 5 th stage of the construction, we must have
M exactly one of the J8" 85 . (This follows immediately from the definition of an interval.)
Therefore , + $"5 for every 5 , so + , and M g

204

v) G is also small in another sense. In the construction of G , the open intervals we

"
"
delete have total length "$ # "* % #(
"$ # ". In some sense, the length of what
$

remains is 0 ! This is made precise in measure theory, where we say that ! " G has measure
1 and that G has measure 0. Measure is yet another way, differing from both cardinality and
category, to measure the size of a set.

The following theorem states that the topological properties of G which we have discussed
actually characterize the Cantor set. The theorem can be used to prove that the generalized
Cantor set resulting from any of these modified constructions is topologically the same as G .
Theorem 10.1 Suppose E is a nonempty compact subset of which is dense-in-itself and
contains no nonempty open interval. Then E is homeomorphic to the Cantor set G .
Proof The proof is omitted. (See Willard, General Topology.)
It is possible to modify the construction of G by changing middle thirds to, say, middle
fifths or even by deleting middle thirds at stage one, middle fifths at stage two, etc. The
resulting generalized Cantor set is homeomorphic to G (using Theorem 10.1) but the total
lengths of the deleted open intervals may be different: that is, the result is can be a topological
Cantor set with positive measure. Measure is not a topological property.

As we noted earlier, G consists of those points B ! " for which we can write
B $>88 !>" ># >5 three , with each >8 ! #

8"

We can obviously rewrite this as B 2$,88 , where each ,8 ! " and therefore we can define

8"

a function 1 G ! " as follows:

for B G 1B 1 2$,88 #,88 .

8"

8"

More informally,:
1! #," #,# #,$ three !," ,# ,$ two

This mapping 1 is not one-to-one . For example,

1 (* 1!#!####three ) !"!""""two
1 )* 1!##!!!!three ) !""!!!two $%

$
%

and

In fact, for + , G , we have 1+ 1, iff the interval + , is one of the deleted middle
thirds. It should be clear that 1 is continuous and that if + , in G , then 1+ 1, that is,
1 is weakly increasing. (Is 1 onto? )

205

The function 1 is not defined on ! " G . But ! " G is, by construction, a union of disjoint
open intervals. We can therefore extend the definition of 1 to a mapping K ! " ! " in the
following simple-minded way:
KB

1B if B G
1+ if B + ,, where + , is a deleted middle third

Since we know 1+ 1,, this amounts to extending the graph of 1 over each deleted middle
third + , by using a horizontal line segment from + 1+ to , 1, The result is the graph
of the continuous function K.
K is sometimes called the Cantor-Lebesgue function. It satisfies:
i) K! ! K" "
ii) K is continuous
iii) K is (weakly) increasing
iv) at any point B in a deleted middle third + ,, K is differentiable
and K w B !
Recall that G has measure !. Therefore we could say K w B ! almost everywhere even
though K rises monotonically from ! to 1!

Note: The technique that we used to extend 1 works in a similar way for any closed set J
and any continuous function 1 J . Since J is open in , we know that we can write
J as the union of a countable collection of disjoint open intervals M8 which have for + ,,
, or + We can extend 0 to a continuous function J simply by extending
the graph of 0 over linearly over each M8
if M8 + ,,

then let the graph J over M8 be the straight line segment joining
+ 0 + >9 , 0 ,
if M8 ,,
then let J have the constant value 0 , 98 I8
if M8 + , then let J have the constant value 0 + on M8
There is a much more general theorem that implies that whenever J is a closed subset of \ .,
then each 0 GJ and be extended to a function J G\. In the particular case \ ,
proving this was easy because is ordered and we completely understand the structure of the
open sets in ..
Another curious property of G , mentioned without proof, is that its difference set
B C B C G " ". Although G has measure !, the difference set in this case has
measure #!

206

Exercises
E21. Suppose \ is compact and that 0 \ ] is continuous and onto. Prove that ] is
compact. (A continuous image of a compact space is compact.)

E22. a) Suppose that \ is compact and ] is Hausdorff. Let 0 \ ] be a continuous

bijection. Prove that 0 is a homeomorphism.
b) Let g be the usual topology on ! " and suppose g" and g# are two other topologies
on ! " such that g" g g# Prove that ! ", g" is not Hausdorff and that ! ", g# is not

compact. (Hint: Consider the identity map 3 ! " ! " )
c) Is part b) true if ! " is replaced by an arbitrary compact Hausdorff space \ g ?

E23. Prove that a nonempty space \ g is pseudocompact iff every continuous 0

achieves both a maximum and a minimum value.
E24. Suppose 0 \ . ] . w Prove that 0 is continuous iff 0 lO is continuous for
every compact set O \ .
E25. Suppose that E and F are nonempty disjoint closed sets in \ . and that E is compact.
Prove that .E F !

E26.

Let \ and ] be topological spaces.

a) Prove that \ is compact iff every open cover by basic open sets has a finite subcover.

b) Suppose \ ] is compact. Prove that if \ ] g, then \ and ] are compact. (By

induction, a similar statement applies to any finite product.)
c) Prove that if \ and ] are compact, then \ ] is compact. (By induction, a similar
statement applies to any finite product.)
(Hint: for any B \ B ] is homeomorphic to ] . Part a) is also relevant.)
d) Point out explicitly why the proof in c) cannot be altered to prove that a product of two
countably compact spaces is countably compact. (An example of a countably compact space \
for which \ \ is not even pseudocompact is given in Chapter X, Example 6.8.)
Note: In fact, an arbitrary product of compact spaces is compact. This is the Tychonoff Product
Theorem which we will prove later.
It is true that the product of a countably compact space and a compact space is countably
compact. You might trying proving this fact. A very similar proof shows that a product of a

207

compact space and a Lindelof

space is Lindelof.
That proof does not generalize, however, to
show that a product of two Lindelof
spaces is Lindelof.
Can you see why?
E27. Prove that if \ . is a compact metric space and l \ l i! , then \ has infinitely many
isolated points.

E28. Suppose \ is any topological space and that ] is compact. Prove that the projection map
1\ \ ] \ is closed. (A projection parallel to a compact factor is closed.)
E29. Let \ be an uncountable set with the discrete topology g . Prove that there does not exist
a totally bounded metric d on \ such that g. g .
E30. a) Give an example of a metric space \ . which is totally bounded and an isometry
from \ . into \ . which is not onto.
Hint: on the circle W " , start with a point : and keep rotating it around the boundary by
increments of some angle " .
b) Prove that if \ . is totally bounded and if 0 is an isometry from \ . into \ .,
then 0 \ must be dense in \ ..
Hint: Given B % ! cover \ with finitely many #% spheres; if the sequence B 0 B 0 0 B,
consists of distinct terms, there must be infinitely many of them in one sphere.
c) Show that a compact metric space cannot be isometric to a proper subspace of itself.
Hint: you might use part b).
d) Prove that if each of two compact metric spaces is isometric to a subspace of the other,
then the two spaces are isometric to each other.
Note: Part d) is an analogue of the Cantor-Schroeder-Bernstein Theorem for compact metric
spaces.
E31. Suppose \ . is a metric space and that \ . w is totally bounded for every metric
. w . . Must \ . be compact?

E32. Let h be an open cover of the compact metric space \ .. Prove that there exists a
constant $ ! such that for all B: F$ B Y for some Y h . (The number $ is called a
Lebesgue number for h .)
E33. Suppose \ . is a metric space with no isolated points. Prove that \ is compact iff
.E F ! for every pair of disjoint closed sets E F \

208

Chapter IV Review
Explain why each statement is true, or provide a counterexample.

1. Let E denote the set of fixed points of a continuous function 0 \ \ where \ is a

Hausdorff space. E is closed in \ .
2. Let W denote the set of all Cauchy sequences in which converge to a point in . Then
lW -.
3. For a metric space \ ., if every continuous function 0 \ assumes a minimum
value, then every infinite set in \ has a limit point.
4. There exists a continuous function 0 for which the Cantor set is the set of fixed
points of 0 .
5. If \ has the cofinite topology, then every subspace is pseudocompact.
6. Let ! " have the subspace topology from the Sorgenfrey line. Then ! " is countably
compact.
7. Let E E is first category in , and let g be the topology on for which is a
subbase. Then g is a Baire space.
8. Every sequence in ! "# has a Cauchy subsequence.
9. If E is a dense subspace of \ . and every Cauchy sequence in E converges to some point
in \ , then \ . is complete.
10. If g is the cofinite topology on , then ( g ) has the fixed point property.
11. Every sequence in ! " has a Cauchy subsequence.
12. Let ^ ! 13 G , where G is the Cantor set. Then ^ is completely metrizable.
13. The Sorgenfrey plane is countably compact.
14. If every sequence in the metric space \ . has a convergent subsequence, then every
continuous real valued function on \ must have a minimum value.
15. Suppose \ . is complete and 0 \ \ . The set G B 0 is continuous at B is
second category in itself.
16. In the space G! ", with the metric 3 of uniform convergence, the subset of all
polynomials is first category.
17. A nonempty open set in \ g cannot be nowhere dense.

209

18. Every nonempty nowhere dense subset of contains an isolated point.

19. There are exactly - nowhere dense subsets of .
20. Suppose F is nonempty subset of with no isolated points. F cannot be completely
metrizable.
21. Suppose \ . is a countable complete metric space. If E \ , then E . may not be
complete, but E is completely metrizable.
22. If E is first category in \ and F E, then F is first category in E.
23. There are - different metrics . on , each equivalent to the usual metric, for which . is
complete.
24. In , with the cofinite topology, every infinite subset is sequentially compact.
25. Suppose 0 ! "# and E : (! ")# :
usual metric, is totally bounded.

`0
`B

and

`0
`C

both exist at :}. Then E, with its

26. A space \ g is a Baire space iff \ is second category in itself.

27. Let I denote the subset of the Cantor set G consisting of the endpoints of the open intervals
deleted from ! " in the construction of G . Suppose 0 ! " is continuous and that
0 I ! !". Then 0 G ! !" .
28. Let . be a metric on the irrationals which is equivalent to the usual metric and such that
. is complete. Then . must be totally bounded.
29. Suppose \ . is a nonempty complete metric space and E is a closed subspace such that
E . is totally bounded. Then every sequence in E has a cluster point.
30. There is a metric d on ! ", equivalent to the usual metric, such that ! " . is not
complete.
31. Let ^ ! "# have the usual topology. Every nonempty closed subset of ^ is second
category in itself.
32. Suppose 0 ! "8 7 and G B ! "8 0 is continuous at B. There is a metric .
equivalent to the usual metric on G for which G . is complete.
33. If the continuum hypothesis (CH) is true, then cannot be written as the union of fewer than
c nowhere dense sets.
34. Suppose E is a complete subspace of ! " Then E is a K$ set in
35. There are 2- different subsets of none of which contains an interval of positive length.
36. Suppose \ . is a nonempty complete metric space and E is a closed subspace such that

210

E . is totally bounded. Then every sequence in E has a cluster point.

37. Let . w be a metric on the irrationals which is equivalent to the usual metric and for which
. w is totally bounded. Then . w cannot be complete.
38. The closure of a discrete subspace of may be uncountable.
39. Suppose 0 ! "8 7 and G B ! "8 0 is continuous at B. Then there is a
metric d on G , equivalent to the usual metric, such that G . is totally bounded.
40. Suppose \ g is compact and 0 \ is a continuous function such that 0 B ! for
all B \ Then there is an % ! such that 0 B % for all B \
41. If every point of \ . is isolated, then \ . is complete.
42. If \ . has no isolated points, then its completion has no isolated points.
43. Suppose that for each point B \ ., there is an open set YB such that YB . is
complete. Then there is a metric d w d on \ such that \ . w is complete.

44. For a metric space \ . with completion (\ . , it can happen that \ \ i! .

45. Let . be the usual metric on . There is a completion ( . of . for which

l l i!
46. A subspace of is bounded iff it is totally bounded.
47. A countable metric space must have at least one isolated point.
48. The intersection of a sequence of dense open subsets in must be dense in .
49. If G is the Cantor set, then G is an J5 set in .
50. A subspace E of a metric space \ . is compact iff E is closed and bounded.
51. Let U + , ! " + , + , be the base for a topology g on ! ". The
space ! " g is compact.
52. j# ., where . is the usual metric on j# , is sequentially compact.
53. A subspace of a pseudocompact space is pseudocompact.
54. j# ., where . is the usual metric on j# , is countably compact.
55. An uncountable closed set in must contain an interval of positive length.
56. Suppose G is the Cantor set and 0 G G is continuous. Then the graph of 0 (a subspace
of G G , with its usual metric) is totally bounded.

211

57. Let G denote the Cantor set ! ", with the metric .B C

BC
"BC

. Then G . is

totally bounded.
58. A discrete subspace of must be closed in
59. A subspace of which is discrete in its relative topology must be countable.
60. If E and F are subspaces of \ g and each is discrete in its subspace topology, then E F
is discrete in the subspace topology.
61. Let cos8 denote the composition of cos with itself 8 times. Then for each B " ", there
exists an 8 (perhaps depending on B) such that cos8 B B
62. Suppose . is any metric on equivalent to the usual metric. Then ! " . is totally bounded.
63. Let W " denote the unit circle in # . W " is homeomorphic to a subspace of the Cantor set G .
64. There is a metric space \ . satisfying the following condition: for every metric d w d,
\ . w is totally bounded.
65. ! "# , with the subspace topology from the Sorgenfrey plane, is compact.

212