CHAPTER 1

Metric Spaces
1. Denition and examples
Metric spaces generalize and clarify the notion of distance in the real line. The
denitions will provide us with a useful tool for more general applications of the
notion of distance:
Definition 1.1. A metric space is given by a set X and a distance function d :
X X R such that
i) (Positivity) For all x, y X
0 d(x, y) .
ii) (Non-degenerated) For all x, y X
0 = d(x, y) x = y .
iii) (Symmetry) For all x, y X
d(x, y) = d(y, x)
iv) (Triangle inequality) For all x, y, z X
d(x, y) d(x, z) + d(z, y) .
Examples:
i) X = R, d(x, y) = [x y[.
ii) X = R
2
= R R, x = (x
1
, x
2
), y = (y
1
, y
2
)
d
1
(x, y) = [x
1
y
1
[ +[x
2
y
2
[ .
iii) X = R
2
, x = (x
1
, x
2
), y = (y
1
, y
2
)
d
2
(x, y) =
_
[x
1
y
1
[
2
+[x
2
y
2
[
2
_1
2
.
iv) Let X = p
1
, p
2
, p
3
and
d(p
1
, p
2
) = d(p
2
, p
1
) = 1 ,
d(p
1
, p
3
) = d(p
3
, p
1
) = 2 ,
1
2 1. METRIC SPACES
d(p
2
, p
3
) = d(p
3
, p
2
) = 3 .
Can you nd a triangle (p
1
, p
2
, p
3
) in the plane with these distances?
v) Let X = p
1
, p
2
, p
3
and
d(p
1
, p
2
) = d(p
2
, p
1
) = 1 ,
d(p
1
, p
3
) = d(p
3
, p
1
) = 2 ,
d(p
2
, p
3
) = d(p
3
, p
2
) = 4 .
Can you nd a triangle (p
1
, p
2
, p
3
) in the plane with these distances?
vi) The French railway metric (Chicago suburb metric) on X = R
2
is dened
as follows: Let x
0
= (0, 0) be the origin, then
d
SNCF
(x, y) =
_

_
d
2
(x, y) if there exists a t R such that x
1
= ty
1
and x
2
= ty
2
d
2
(x, x
0
) + d
2
(x
0
, y) else
.
Exercise: Show that the railroad metric satises the triangle inequality.
It is by no means trivial to show that d
2
satises the triangle inequality. In the
following we write 0 = (0, ..., 0) for the origin in R
n
.
CS Lemma 1.2. Let x, y R
n
, then
[
n

i=1
x
i
y
i
[
_
n

i=1
[x
i
[
2
_1
2
_
n

i=1
[y
i
[
2
_1
2
Lemma 1.3. On R
n
the metric
d
2
(x, y) =
_
n

i=1
[x
i
y
i
[
2
_1
2
satises the triangle inequality.
Proof. Let x, y, z R
n
. Then we deduce from Lemma
CS
1.2
d(x, y)
2
=
n

i=1
[x
i
y
i
[
2
=
n

i=1
[(x
i
z
i
) (y
i
z
i
)[
2
=
n

i=1
[(x
i
z
i
)[
2
2
n

i=1
(x
i
z
i
)(y
i
z
i
) +
n

i=1
[y
i
z
i
[
2
d(x, z)
2
+ 2d(x, y)d(y, z) + d(y, z)
1. DEFINITION AND EXAMPLES 3
= (d(x, z) + d(y, z))
2
.
Hence,
d(x, y) d(x, z) + d(y, z)
and the assertion is proved.
More examples:
(1) Let n be a prime number. On Z we dene
dd
n
(x, y) = n
max{mN: n
m
divides x-y}
.
The n-adic metric satises a stronger triangle inequality
dd
n
(x, y) maxdd
n
(x, z), dd
n
(z, y) .
(2) Let 1 p < . Then
d
p
(x, y) =
_
n

i=1
[x
i
y
i
[
p
_1
p
denes a metric n R
n
.
(3) For p =
d

(x, y) = max
i=1,..,n
[x
i
y
i
[
also denes a metric on R
n
.
Project 1: Let 1 < p, q < such that 1/p+1/q = 1. Show Minkowskis inequality.
Mink (1.1) xy
x
p
p
+
y
q
q
holds for all x, y > 0. Hint: the function f(x) = ln x is convex on (0, ).
Proof of the triangle inequality for d
p
. The triangle inequality for p =
1 is obvious. We will st show
mink2 (1.2) [
n

i=1
x
i
y
i
[
_
n

i=1
[x
i
[
p
_1
p
_
n

i=1
[y
i
[
q
_1
q
whenever
1
p
+
1
q
= 1. Let t > 0. We rst observe that
[
n

i=1
x
i
y
i
[ =
n

i=1
[tx
i
[[t
1
y
i
[
n

i=1
1
p
[tx
i
[
p
+
1
q
[t
1
y
i
[
q
=
t
p
p
n

i=1
[x
i
[
p
+
t
q
q
n

i=1
[y
i
[
q
.
4 1. METRIC SPACES
What is best choice of t? Make
t
p
n

i=1
[x
i
[
p
= t
q
n

i=1
[y
i
[
q
i.e.
t
p+q
=
n

i=1
[y
i
[
q
n

i=1
[x
i
[
p
.
This yields
[
n

i=1
x
i
y
i
[ t
p
n

i=1
[x
i
[
p
=
_
n

i=1
[y
i
[
q
_
p
p+q
_
n

i=1
[x
i
[
p
_
p
p+q
n

i=1
[x
i
[
p
=
_
n

i=1
[y
i
[
q
_1
q
_
n

i=1
[x
i
[
p
_
1
1
q
Now, we proof the triangle inequality. Let x = (x
i
), (y
i
) and z = (z
i
) in R
d
. Then
we apply (
mink2
1.2)
d
p
(x, y)
p
=
d

i=1
[x
i
y
i
[
p

i=1
[x
i
y
i
[
p1
([x
i
z
i
[ +[z
i
y
i
[)

i=1
[x
i
y
i
[
p1
[x
i
z
i
[ +
d

i=1
[x
i
y
i
[
p1
[z
i
y
i
[

_
d

i=1
([x
i
y
i
[
p1
)
q
_
1
q
_
_
_
d

i=1
[z
i
x
i
[
p
_
1
p
+
_
d

i=1
[z
i
y
i
[
p
_
1
p
_
_
.
However, 1 = 1/p + 1/q implies p 1 = p/q and thus q(p 1) = p. Hence we get
d
p
(x, y)
p
d
p
(x, y)
p1
(d
p
(x, z) + d
p
(z, y)) .
If x ,= y we may divide and deduce the assertion.
2. EXCURSION: CONVEX FUNCTIONS 5
2. Excursion: Convex functions
Definition 2.1. Let I be an interval. A function f : I R is called convex if
f(x + (1 )y) f(x) + (1 )f(y)
holds for all x, y I, 0 < < 1.
Lemma 2.2. Let f : [a, b] R be continuous, diernetiable on (a, b) such that f

is
increasing. Then f is convex.
Proof. Let x [a, b]. We will show that
g(z) =
f(y + z) f(y)
z
is monotone increasing on (0, bx). Indeed, by the fundamental theorem and change
of variables we deduce for z
1
< z
2
and =
z
1
z
2
(s = t, ds = dt
g(z
1
) =
z
1
_
0
f

(s)
ds
z
1
=
z
2
_
0
f

(t)
dt
z
1
=
z
2
_
0
f

(t)
dt
z
2

z
2
_
0
f

(t)
dt
z
2
= g(z
2
) .
Now, we x y < x and u = x +(1 )y = y +(x y), z
1
= (x y), z
2
= x y.
Then, we get
f(y + z) f(y)
(x y)

f(x) f(y)
(x y)
.
This implies
f(x + (1 )y) f(y) + (f(x) f(y)) = f(x) + (1 )f(y) .
Proof of
Mink
1.1. Let x, y > 0. Since ln x is convex we have
ln(
1
p
x
p
+
1
q
y
q
)
1
p
(ln x
p
) +
1
q
(ln y
q
) .
This shows by the monotonicity of exp that
1
p
x
p
+
1
q
y
q
e
ln x+ln y
= xy .
Minkowskis inequality is proved.
6 1. METRIC SPACES
3. Continuous functions between metric spaces
Continuous functions preserve properties of metric spaces and allow to describe
deformation of one metric space into another. There are three dierent (but equiv-
alent) ways of dening continuity, the --criterion, the sequence criterion and the
topological criterion. Each of them is interesting in its own right.
Definition 3.1. Let (X, d) and (Y, d

) be metric spaces. A map f : X Y is called

continuous if for every x X and > 0 there exists a > 0 such that
edelt (3.1) d(x, y) < = d

(f(x), f(y)) < .

Let us use the notation
B(x, ) = y : d(x, y) < .
For a subset A X, we also use the notation
f(A) = f(x) : x A .
Similarly, for B Y
f
1
(B) = x X : f(x) B .
Then (
edelt
3.1) means
f(B(x, )) B(f(x), ) .
Or in a very non-formal way
f maps small balls into small balls .
Our aim is to prove a criterion for continuity in terms of so called open sets. This
criterion illustrates simultaneously the role of open sets and its interaction with
continuity and has a genuinely geometric avor.
Definition 3.2. A subset O of a metric space is called open if
x O : > 0 : B(x, ) O .
Examples:
O = (1, 1) , O = R , O = (1, 1) (2, 2)
are open in R, (R
2
, d
2
) respectively.
Remark 3.3. The sets B(x, ), x X, > 0 are open.
Proposition 3.4. Let (X, d), (Y, d

) be metric spaces and f : X Y be a map. f

is contiuous i f
1
(O) is open for all open subsets O Y .
3. CONTINUOUS FUNCTIONS BETWEEN METRIC SPACES 7
Proof. : We assume that f is continuous and O is open. Let x f
1
(O),
i.e. f(x) O. Since O is open, there exists an > 0 such that B(f(x), ) O. By
continuity, there exists a > 0 such that
f(B(x, )) B(f(x), ) O .
Therefore
B(x, ) f
1
(O) .
Since x f
1
(O) was arbitrary, we deduce that f
1
(O) is open.
: Let x X and > 0. Let us show that
B(f(x), )
is a on open subset of (Y, d

). Indeed, let y B(f(x), ) dene

= d

(y, f(x)).
Let z Y such that
d(z, y) <

then
d(f(x), z) d(f(x), y) + d(y, z) < d(f(x), y) + d

(y, f(x)) = .
Thus
B(y, d

(f(x), y)) B(f(x), ) .

By the assumption, we see that f
1
(B(f(x), )) is an open set. Since x f
1
(B(f(x), )),
we can nd a > 0 such that
B(x, ) f
1
(B(f(x), )) .
Hence, for all x with d(x, x) < , we have
d

(f(x), f( x)) < .

The assertion is proved.
Examples:
(1) Let (X, d) be a metric space and x
0
X be a point , then f(x) = d(x, x
0
)
is continuous. Indeed, the triangle inequality implies
d(d(x, x
0
), d(d(y, 0)) = [d(x, x
0
) d(y, x
0
)[ d(x, y)
This easily implies the assertion.
(2) On R
n
with the standard euclidean metric d = d
2
, the function f : R
n
R
n
dened by f(x) = d(x, 0)x is continuous.
8 1. METRIC SPACES
(3) (Exercise) The function f : R
3
R
3
, f(x) = (cos(x
1
), sin(x
2
), cos(x
1
)) is
continuous.
Definition 3.5. Let (X, d), (Y, d

) be a metric space. The space C(X, Y ) is the set

of all continuous functions from X to Y . Let x
0
X be a point. Then
C
b
(X, Y ) = f : X Y : f is continuous and sup
xX
d

(f(x), f(x
0
)) <
is the subset of bounded continuous functions.
Proposition 3.6. Let (X, d), (Y, d

) be metric spaces and x

0
X. Then C
b
(X, Y )
equipped with
d(f, g) = sup
xX
d

(f(x), g(x))
is a metric space.
Problem: Show that d is not well-dened on C(R, R).
Proof: d(f, g) = 0 if and only if f(x) = g(x) for all x X. This means f = g. Let
us show that d is well-dened. Indeed, if f, g C
b
(X, Y ). Then
sup
x
d

(f(x), g(x)) sup

x
d

(f(x), f(x
0
)) + d

(f(x
0
), g(x
0
)) + d

(g(x
0
), g(x))
sup
x
d

(f(x), f(x
0
)) + d

(f(x
0
), g(x
0
)) + sup
x
d(g(x
0
), g(x))
is nite. Let h be a third function and x X. Than
d

(f(x), g(x)) d

(f(x), h(x)) + d(h(x), g(x)) d(f, h) + d(h, g) .

Taking the supremum yields the assertion.
alg Proposition 3.7. Let (X, d) be a metric space. Then C(X, R) is closed under
(pointwise-) sums, products and multiplication with real numbers. (C(X, R) is an
algebra over R).
Remark 3.8. Let X = N and d(x, y) = 1 of x ,= y and d(x, y) = 0 for x = y. (This
is called the discrete metric). Then C(X, R) is an innite dimensional vector space.
Proof of
alg
3.7. Let f, g C(X, R) be continuous and x X. Consider x

X.
Then
fg(x) fg(y) = f(x)g(x) f(y)g(y) = (f(x) f(y))g(x) + f(y)(g(x) g(y))
= (f(x) f(y))g(x) + f(x)(g(x) g(y)) + (f(y) f(x))(g(x) g(y)) .
Let > 0 and = min, 1. We may choose
1
> 0 such that
d(f(x), f(y))(1 +[g(x)[) <

3
3. CONTINUOUS FUNCTIONS BETWEEN METRIC SPACES 9
holds for all d(x, y) <
1
. Similarly, we may choose
2
> 0 such that
d(g(x), g(y))(1 +[f(x)[) <

3
.
Let = min(
1
,
2
) and d(x, y) < . Then we deduce that
d(fg(x), fg(y)) = [fg(x) fg(y)[ <

3
+

3
+

2
9
< .
Thus fg is again continuous. The other assertions are easier.
Corollary 3.9. The polynomials on R are continuous.
Lemma 3.10. Let 1 p and x, y R
n
, then
1
n
1
p
d
p
(x, y) d

(x, y) d
p
(x, y) .
Proof. The last inequality is obvious. For the rst one, we consider x, y R
n
and 1 p < , then by estimating every element in the sum against the maximum
d
p
(x, y)
p
=
n

i=1
[x
i
y
i
[
p
nmax[x
i
y
i
[
p
.
Taking the p-th root, we deduce the assertion.
Corollary 3.11. Let 1 p, q , then the identity map id : (R
n
, d
p
) (R
n
, d
q
)
is continuous.
Proof. We have for all x R
n
and > 0
B
d
p
(x,

n
) B
d
q
(x, ) .
This easily implies the assertion.
Corollary 3.12. The metrics d
p
dene the same open sets on R
n
.
Definition 3.13. Let (X, d) be a metric space. We say that a sequence (x
n
) con-
verges to x if for all > 0 there exists n
0
such that for n > n
0
we have
d(x
n
, x
0
) < .
In this case we write
lim
n
x
n
= x
or more explicitly
d lim
n
x
n
= x .
A sequence (x
n
) is convergent, if there exists x X with lim
n
x
n
= x.
10 1. METRIC SPACES
Examples: d
2
lim
n
1
n
= 0, dd
3
lim
n
3
n
= 0. (What axioms of the natural
numbers are involved?).
Proposition 3.14. Let (X, d), (Y, d

) be metric spaces and f : X Y be a map.

Then f is continuous if for every convergent sequence (x
n
) in X
lim
n
f(x
n
) = f(lim
n
x
n
) .
Proof: : Let x = lim
n
x
n
and > 0, then there exists a > 0 such that
d(y, x) < d

(f(y), f(x)) < .

Let n
0
N be such that
d(x
n
, x) <
for all n > n
0
, then
d

(f(x
n
), f(x)) <
for all n > n
0
. Hence
lim
n
f(x
n
) = f(x) .
Let x X and assume in the contrary that
> 0 > 0y : d(y, x) < and d

(f(x), f(y)) .
Applying these successively for all =
1
k
, we nd a sequence (x
k
) such that
d(x
k
, x) <
1
k
and d

(f(x
k
), f(x))

.
and thus
lim
k
x
k
= x .
By assumption, we have
lim
k
f(x
k
) = f(x) .
Hence, there exists a k
0
such that for all k > k
0
d(f(x
k
), f(x)) < .
a contradiction.
4. COMPLETE METRIC SPACES AND COMPLETION 11
4. Complete metric spaces and completion
Complete metric space are crucial in understanding existence of solutions to many
equations. Complete spaces are also important in understanding spaces of inte-
grable functions. We will review basic properties here and show the existence of a
completion.
We will say that a sequence in a metric space is a Cauchy sequence of for every
> 0 there exists n
0
N such that
d(x
n
, x
m
) <
for all n, m > n
0
.
Definition 4.1. A metric space (X, d) is called complete, if every Cauchy sequence
converges.
model Proposition 4.2. The space (R
2
, d
1
) is complete.
Proof: Let x
n
be a Cauchy sequence in (R
2
, d
1
). Then x
n
= (x
n
(1), x
n
(2)) is a
sequence of pairs.
Claim: The sequences (x
n
(1))
nN
and (x
n
(2))
nN
are Cauchy sequences.
Indeed, let > 0, then there exists an n
0
such that
d
1
(x
n
, x
m
) <
for all n, m > n
0
. In particular, we have
[x
n
(1) x
m
(1)[ [x
n
(1) x
m
(1)[ +[x
n
(2) x
m
(2)[ d
1
(x
n
, x
m
) <
for all n, m > n
0
and
[x
n
(2) x
m
(2)[ [x
n
(1) x
m
(1)[ +[x
n
(2) x
m
(2)[ d
1
(x
n
, x
m
) < .
Therefore, (x
n
(1)) and (x
n
(2)) are Cauchy.
Since R is complete, we can nd x(1) and x(2) such that
lim
n
x
n
(1) = x(1) and lim
n
x
n
(2) = x(2) .
Claim: lim
n
x
n
= (x(1), x(2)).
Indeed, Let > 0 and choose n
1
such that
[x
n
(1) x(1)[ <

2
for all n > n
1
. Choose n
2
such that
[x
n
(2) x(2)[ <

2
12 1. METRIC SPACES
for all n > n
2
. Set n
0
= maxn
1
, n
2
, then for every n > n
0
, we have
d
1
(x
n
, (x(1), x(2)) = [x
n
(1) x(1)[ +[x
n
(2) x(2)[ <
Thus
lim
n
x
n
= x
and the assertion is proved.
Examples:
(1) Let X = R 0 and d(x, y) = [x y[, them (X, d) is not complete. The
sequences (
1
n
) is Cauchy and does not converge.
(2) Let p be a prime number. On the set of integers, we dene
dd
p
(z, w) = p
n
,
where n = maxn : p
n
divides (zw) . This satises the triangle inequal-
ity. The sequence (x
n
) given by x
n
= p + p
2
+ + p
n
is a non convergent
Cauchy sequence.
rncomp Theorem 4.3. Let n N. The space (R
n
, d
2
) is a complete metric space.
Proof. Similar as in Proposition
model
4.2 using the following Lemma .
Lemma 4.4. Let x, y R
n
, then
d
2
(x, y)
n

i=1
[x
i
y
i
[ .
Proof. We proof this by induction on n N. The case n = 1 is obvious.
Assume the assertion is true for n and let x, y R
n+1
. We dene the element
z = (x
1
, ..., x
n
, y
n+1
), then we deduce from the triangle inequality
d
2
(x, y) d
2
(x, z) + d
2
(z, y)
=
_
n+1

i=1
[x
i
z
i
[
2
_1
2
+
_
n+1

i=1
[z
i
y
i
[
2
_1
2
= [x
n+1
y
n+1
[ +
_
n

i=1
[x
i
y
i
[
2
_1
2
.
To apply the induction hypothesis, we dene x = (x
1
, ..., x
n
) and y = (y
1
, .., y
n
).
Then the induction hypothesis yields
_
n

i=1
[x
i
y
i
[
2
_1
2
= d
2
( x, y)
n

i=1
[x
i
y
i
[ .
4. COMPLETE METRIC SPACES AND COMPLETION 13
Hence,
d
2
(x, y) [x
n+1
y
n+1
[ +
_
n

i=1
[x
i
y
i
[
2
_1
2
[x
n
y
n
[ +
n

i=1
[x
i
y
i
[
=
n+1

i=1
[x
i
y
i
[ .
The assertion is proved.
Definition 4.5. A subset C X is called closed if X C is open.
complete1 Proposition 4.6. Let C be closed subset of a complete metric space (X, d), then
(C, d[
CC
) is complete.
Proof. Let (x
n
) C be Cauchy sequence. Since X is complete, there exists
x X such that
x = lim
n
x
n
.
We have to show x C. Assume x / C. Then there exists a > 0 such that
B(x, ) X C. By denition of the limit there exists n
0
such that d(x
n
, x) <
for all n > n
0
. Set n = n
0
+ 1. Then d(x
n
, x) < implies x
n
X C and x
n
C
by denition. This contradiction nished the proof.
comp1 Theorem 4.7. Let (Y, d

) be complete metric space. Let h C(X, Y ) and

C
h
(X, Y ) = f C(X, Y ) : sup
xX
d

(f(x), h(x)) <

Then C
g
(X, Y ) is complete with respect to
d(f, g) = sup
xX
d

(f(x), g(x)) .
Proof. Let (f
n
) C
h
(X, Y ) be Cauchy sequence. This means that for every
> 0 there exists an n
0
such that
eqq (4.1) sup
xX
d

(f
n
(x), f
m
(x)) <

2
.
In particular, for xed x X, f
n
(x) is Cauchy. Therefore f(x) := lim
m
f
m
(x) is a
well-dened element in Y . We x n > n
0
and consider m n
0
such that
d

(f
m
(x), f(x))

3
.
14 1. METRIC SPACES
This implies
d

(f
n
(x), f(x)) d

(f
n
(x), f
m
(x)) + d

(f
m
(x), f(x))
5
6

for all x X. In particular,

eqq2 (4.2) sup
nn
0
sup
xX
d

(f
n
(x), f(x))
5
6
.
Let us show that f is continuous. Let z X and > 0. Choose n
0
according to (
eqq
4.1).
Choose n = n
0
+ 1. Let > 0 such that d(x, y) < implies d

(f
n
(x), f
n
(y)) < .
Then, we have
d

(f(x), f(y)) d

(f(x), f
n
(x)) + d

(f
n
(x), f
n
(y)) + d

(f
n
(y), f(y)) < 3 .
Since > 0 is arbitrary, we see that f is continuous. Moreover, (
eqq2
4.2) implies that
f
n
converges to f. Finally, (
eqq2
4.2) for = 1 implies that
sup
x
d(f(x), h(x)) sup
x
d(f(x), f
n
(x)) + sup
x
d(f
n
(x), h(x)) <
implies that f C
h
(X, Y ).
Definition 4.8. Let (X, d) be a metric space and C X. O X is called sense if
for ever x C and > 0 B(x, ) O ,= .
Definition 4.9. Let O X be a subset. Then

O =
OC,Cclosed
C
is called the closure.
dens Lemma 4.10. O is dense in

O and

O is closed.
Proof. Let x

O. Assume B(x, ) O = . Then C = X B(x, ) contains
O. Thus

O C .
This implies that x /

O, a contradiction. Now, we show that

O is closed. Indeed,
let y /

O. Then there has to be a closed set C such that O C but y / C. This
means y X C which is open. Hence there exists > 0 such that
B(y, ) X C
By denition every element z B(y, ) does not belong to

O. This means B(y, )
X

O.
5. UNIQUE EXTENSION 15
comp2 Theorem 4.11. Let (X, d) be a non-empty metric space. For every x X we dene
f
x
(y) = d(x, y) .
Let x
0
X. The map f : X C
f
x
0
(X, R) satises the following properties.
i) d(f(x), d(f(y)) = d(x, y),
(1) The closure C = f(X) is complete,
(2) f(X) is dense in the closure C = f(X).
Proof. Let x, y X and z X. Then the converse triangle ineqquality
implies
[f
x
(z) f
y
(z)[ = [d(x, z) d(y, z)[ d(x, y) .
Moreover,
[f
x
(z) f
x
(y)[ = [d(x, z) d(x, y)[ d(z, y) .
Therefore f
x
C
f
x
0
(X, R) for every x X and
d(f
x
, f
y
) d(x, y) .
However,
d(f
x
, f
y
) [f
x
(x) f
y
(x)[ = [0 d(y, x)[ = d(y, x) .
This shows i). According to Proposition
complete1
4.6 and Theorem
comp1
4.7, we see that C is
complete. According to Lemma
dens
4.10, we deduce that f(X) is dense in C.
Project: On C([0, 1]) we dene
d
1
(f, g) =
_
[f(s) g(s)[ds .
Show that (C([0, 1]), d
1
) is not complete.
Project: In the literature you can nd another description of the completion of a
metric space. Find it and describe it.
5. Unique extension of densely dened uniformly continuous functions
In this section we will show that the completion C constructed in Theorem
comp2
4.11 is
unique (in some sense). This is based on a simple observation-the unique extension.
This principle is very often used in analysis.
Definition 5.1. Let (X, d), (Y, d

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

called uniform continuous if for every > 0 there exists a > 0 such that
d(x, y) < d

(f(x), f(y)) < .

16 1. METRIC SPACES
u-ext Proposition 5.2. Let O C be a dense set and f : O Y be uniformly contin-
uous function with values in a complete metric space. Then there exists a unique
continuous function

f : O Y such that

f(x) = f(x) for all x O.
Proof. Let x X. Since B(x,
1
n
) O is not empty, we may nd (x
n
) O such
that lim
n
x
n
= x. We try to dene

f(x) = lim
n
f(x
n
) .
Let us show that this is well-dened. So we consider another Cauchy sequence (x

n
)
such that lim
n
x

n
= x. Let > 0. Then there exists > 0 such that
d

(f(x

, y)) <
holds for d(x

, y) < . We may nd n
0
such that
d(x
n
, x) <

2
and
d(x

n
, x) <

2
holds for all n, n

> n
0
. Thus
d

(f(x

n
), f(x
n
)) < .
This argument also shows that (f(x
n
)) is Cauchy and hence

f(x) is well-dened. If
x O, we may choose for (x
n
) the constant sequence x
n
= x and hence

f(x) = f(x).
Now, we want to show that

f is uniformly continuous. Indeed, let > 0, then there
exists > 0 such that d(x

, y

) < implies
d(f(x

), f(y

)) <

2
.
Given x, y C with d(x, y) < , we may nd (x
n
) converging to x and (y
n
) con-
verging to y such that
d(x
n
, x) <
d(x, y)
2
.
Thus for all n N we have
d(x
n
, y
n
) d(x, y) + d(x
n
, x) + d(y
n
, y) < .
This implies
d(f(x), f(y)) = lim
n
d(f(x
n
), f(y
n
)) <

2
.
6. A FAMOUS EXAMPLE 17
This shows that

f is uniformly continuous. If g is another continuous function such
that g(x) = f(x) holds for elements x O, then we may choose a Cauchy sequence
(x
n
) converging to x and get
g(x) = lim
n
g(x
n
) = lim
n
f(x
n
) = f(x) .
Example If f : (0, 1] R is uniformly continuous, then f is bounded (why).
f(x) = 1/x is not uniformly continuous.
Theorem 5.3. The completion of a metric space is unique. More precisely, let C be
the set constructed in Theorem
comp2
4.11. Let C

be a complete metric space and

: X
C

be uniformly continuous with uniformly continuous inverse

1
: (X) X such
that

(X) is dense. Then there is a bijective, bicontinuous map u : C C

such
that u((x)) =

(x).
Proof. The map

1
: (X) C

is uniformly continuous and hence admits

a unique continuous extension u : C C

. Also

1
:

(X) C admits a
unique extension v : C

C. Note that vu : C C is an extension of the map

vu((x)) = (x). Thus there is only one extension, namely the identity. This show
vu = id. Similarly uv = id. Thus v = u
1
and u is bijective and bi-continuous.
Project: Find the completion of (Z, d
3
).
6. A famous example
In this section we want to identify the completion of C([0, 1]) with respect to
d
1
(f, g) =
1
_
0
[f(t) g(t)[dt .
We will also use the function
I(f) =
1
_
0
f(t)dt
dened by the Riemann integral.
Lemma 6.1. I is uniformly continuous.
Proof. It suces to show that
[I(f)[
_
[f(t)[dt .
18 1. METRIC SPACES
(This implies that I is 1-Lipschitz, i.e.
[I(f) I(g)[ d
1
(f, g) .)
We dene f
+
= maxf, 0 and f

= maxf, 0. Then we have

I(f) = I(f
+
) I(f

) I(f
+
) + I(f

) = I([f[) =
_
[f(t)[dt .
Similarly, we may show that
I(f) = I(f
+
) I(f

) I(f
+
) I(f

) = I([f[) =
_
[f(t)[dt .
The assertion follows.
The characteristic function is given by
1
A
(x) =
_
_
_
1 if x A
0 if x / A
.
Lemma 6.2. I(1
[a,b]
) = b a.
Proof. We only consider [a, b] = [0, b]. For 2/n b we dene
f
n
(t) =
_

_
nt if t
1
n
1 if
1
n
t = b
1
n
n(b t)
.
Then we deduce that for m n we have
d
1
(f
n
, f
m
) = 2
1
m
_
0
mt dt +
1
n

1
m

1
n
_
0
ntdt = 2(
1
2m
+
1
n

1
m

1
2n
) =
1
n

1
m
.
Thus (f
n
) is Cauchy and
lim
n
_
f
n
(t)dt = lim
n
b
1
n
= b .
For general a we simply shift.
In the following we denote the length of an interval by
[[a, b][ = b a .
6. A FAMOUS EXAMPLE 19
chy Lemma 6.3. Let f be a continuous positive function on [0, 1], > 0 and > 0.
Then there exists intervals J
1
, ...., J
m
such that
t : f(t) > J
1
J
k
and

k
[J
k
[
1
2
_
f(t)dt .
Proof. Let <

2
. Since f is uniformly continuous, we may nd n N such
that [x y[ 1/n implies
[f(x) f(y)[ < .
We dene x
k
= k/n. Let S = k : f(x
k
) > . Let t [0, 1] such that f(t) > .
We consider k = tn|. Then x
k
t
1
n
and hence
f(x
k
) > f(t) > .
Therefore
t : f(t) >
_
kS
[x
k
, x
k+1
] .
However, f(x
k
) > implies f(t) > 2 for all t [x
k
, x
k+1
]. By the denition
of the lower sums we deduce
1
_
0
f(t)dt

kS
( 2)[J
k
[ .
Since 2 > 0, we deduce the assertion.
Definition 6.4. i) A subset A [0, 1] is said to have measure 0 if for every
> 0 there exists a sequence (J
k
) of intervals such that
A
_
k
J
k
and

k
[J
k
[ < .
ii) A sequence (f
n
) converges almost everywhere (a.e.) to a function f if there
exists a set A of measure 0 such that
lim
n
f
n
(t) = f(t)
for all t A
c
= [0, 1] A.
limit1 Proposition 6.5. Let (f
n
) be a d
1
-Cauchy sequence. Then there exists a subse-
quence (n
k
) and a function f such that f
n
k
converges to f a.e.
20 1. METRIC SPACES
Proof. We may choose (n
k
) such that
d
1
(f
n
k
, f
n
k+1
) 6
k
.
Let us denote g
k
= f
n
k
. We apply Lemma
chy
6.3 to = 2
k
and =
2
k
3
k
2
and nd
intervals J
k
1
, ..., J
k
m(k)
such that

l
[J
k
l
[
6
k
2
k
2
= 2
k
and
t [0, 1] : [g
k
(t) g
k+1
(t)[ > 2
k

_
l
J
k
l
.
We dene
A
k
=
_
nk
_
l
J
n
l
and
A =

k
A
k
.
Then, we see that A A
k
and

nk

l
[J
n
l
[

nk
2
n
= 2
1k
.
Thus shows that A has measure 0. For t / A, we may nd k such that for all n k
we have t /

l
J
n
l
. This implies
[g
n
(t) g
n+1
(t)[ 2
n
for all n k. In particular, (g
k
(t)) is Cauchy for all t A
c
. We may dene
f(t) =
_
_
_
lim
k
g
k
(t) if t / A
0 else
.
Then (g
k
) converges to f almost everywhere.
This leads us to dene the set of possible limits.
L = f : [0, 1] R : (f
n
) C[0, 1], f
n
converges to f a.e.
on L we dene the equivalence relation f g if there exists a set A of measure 0
such that f1
A
c = g1
A
c.
Exercise: Show that is an equivalence relation.
We dene
L = L/ .
6. A FAMOUS EXAMPLE 21
For a function f L we dene the equivalence class [f] = g : g f. In the
following we denote by X the completion of C[0, 1] with respect to the d
1
-metric.
Our main theorem is the following.
inj Theorem 6.6. There is an injective map : X L such that
(x) = [f]
holds whenever (f
n
) is a Cauchy sequence converging to x (with respect to d
1
) and
converging to f. a.e. Moreover, I can be extended to (X).
Problem: Give a description of (X
1
). This is done in the real analysis course
(441=540).
We need some more preparation.
uniqueext Lemma 6.7. Let A =

k
J
k
the union of intervals.
i) Let f C[0, 1], then f1
A
X, f1
A
c X and
d
1
(f1
A
, g1
A
) d
1
(f, g)
and
d
1
(f1
A
c, g1
A
c) d
1
(f, g) .
ii) There is are continuous maps m
A
: X X, m
A
c : X X such that
m
A
(f) = f1
A
and m
A
c(f) = f1
A
c for f C[0, 1].
iii) There is a Lipschitz map add : X X X such that add(f, g) = f + g.
iv) add(m
A
(x), m
A
c(x)) = x for all x X.
v) d
1
(f1
A
c, 0) sup
tA
c [f(t)[.
Proof. We will start with i) for A = [a, b]. We use the functions Let f
n
dened
for [0, b a] and dene g
n
(t) = f
n
(t a). Then we see that for every f C[0, 1]
we have
d
1
(ff
n
, ff
m
) =
b
_
a
[f(t)(f
n
(t) f
m
(t))[dt sup
t
[f(t)[d(f
n
, f
m
)
sup
t
[f(t)[d
1
(f
n
, f
m
) sup
t
[f(t)[ (
1
n

1
m
) .
Thus (ff
n
) is Cauchy. We denote the limit by f1
[a,b]
. (Moreover, ff
n
converges
pointwise to f1
[a,b]
.) Now, we observe that [f
n
(t)[ 1 and hence
d
1
(f1
[a,b]
, g1
[a,b]
) = lim
n
d
1
(ff
n
, gf
n
) = lim
n
1
_
0
[f
n
(t)(f(t) g(t))[dt
22 1. METRIC SPACES

1
_
0
[f(t) g(t)[dt = d
1
(f, g) .
By Proposition
u-ext
5.2 we nd a map u
A
: X X such that u
A
(f) = f1
A
and
d
1
(u
A
(x), u
A
(y)) d
1
(x, y) .
Now, we will prove iii). The metric on X X is given by
d((x, y), (x

, y

)) = d
1
(x, x

) + d
1
(y, y

) .
Now, we consider add : C[0, 1] C[0, 1] C[0, 1] and want to show that add is
uniformly continuous. Indeed, elementary properties of the integral imply
d
1
(add(f, g), add(f

, g

)) =
1
_
0
[(f + g) (f

+ g

)[dt

1
_
0
[f f

[dt +
1
_
0
[g g

[dt = d((f, g), (f

, g

) .
Thus Proposition
u-ext
5.2 implies the assertion iii). In the nest step we prove i) for
A = J
1
J
n
. The key observation here is that we can nd new intervals
J

1
, ...., J

m
such that the J

l
only overlap in one point and
A =
_
l
J

l
.
Therefore, we may dene
u
A
(x) =
m

l=1
1
J

l
x = add(1
J

1
x, add(1
J

2
x, add(1
J

m1
x, 1
J

m
x) ) .
Being a composition of continuous function that is continuous. Moreover, for every
l we may consider the sequence of function f
l
n
constructed for the interval J

l
. The
function
f
n
(t) =
m

l=1
f
l
n
(t)
is positive, continuous and vanishes in the overlapping endpoints. Hence 0 f
n
(t)
1 and the argument from above shows
d
1
(f
n
f, f
n
g) d
1
(f, g) .
This yields
d
1
(f1
A
, g1
A
) = lim
f
d
1
(ff
n
, gf
n
) d
1
(f, g) .
6. A FAMOUS EXAMPLE 23
Then we dene
f1
A
c = f f1
A
= lim
n
add(f, ff
n
) = lim
n
f(1 f
n
) .
Since 0 1 f
n
1 we also prove as above that
d
1
(f1
A
c, g1
A
c) d
1
(f, g) .
Therefore i) is proved for A being a nite union of intervals. Let us show iv) for this
particular case. Indeed,
add(f1
A
, f1
A
c) = lim
n
add(ff
n
, f(1 f
n
)) = lim
n
ff
n
+ f(1 f
n
) = f .
We need an additional estimate for showing the general case:
8-cont (6.1) d
1
(f1
A
, 0) sup
tA
[f(t)[

k
[J
k
[ .
Indeed, inductively we may choose the non-overlapping J

l
in groups J

1
, ..., J

l
1
,
J

l
1
+1
, ..., J
l

2
, ... such
l
k+1

l=l
k1
+1
[J

l
[ [J
k
[ .
Then, we have
d
1
(f1
A
, 0) = lim
n
1
_
0
[f
n
f(t)[dt sup
tA
[f(t)[ lim
n
1
_
0
f
n
(t)dt
= sup
tA
[f(t)[

l
[J

l
[ sup
tA
[f(t)[

k
[J
k
[ .
Now, we consider the general case A =

k
J
k
. We dene A
n
=

kn
J
k
. We
want to show that f1
A
n
is Cauchy. For this we choose non-overlapping intervals
J

l
k1
+1
, ...., J

l
k
J
k
. Then, we deduce from (
8-cont
6.1) that for n m
chy3 (6.2) d
1
(f1
A
n
, f1
A
m
) sup
t
[f(t)[
m

k=n+1
l
k

l=l
k1
+1
[J

l
k
[ sup
t
[f(t)[
m

k=n+1
[J
k
[ .
Thus we may dene f1
A
= lim
n
f1
A
n
. Again, we have
d
1
(f1
A
, g1
A
) = lim
n
d
1
(f1
A
n
, g1
A
n
) d
1
(f, g) .
By the unique extension principle, we nd a Lipschitz map u
A
: X X such that
u
A
(f) = f1
A
. We use the unique extension principle to dene x and the dene
u
A
c(x) = add(x, u
A
(x)). For f, g C[0, 1] we have
d
1
(u
A
c(f), u
A
c(g)) = lim
n
d
1
(f1
A
c
n
, g1
A
c
n
) d
1
(f, g) .
24 1. METRIC SPACES
Thus by unique extension this also holds for x, y X. Finally, we note that for
f C[0, 1]
add(f1
A
, f1
A
c) = add(f1
A
, add(f, f1
A
)) = lim
n
add(f1
A
n
, add(f, f1
A
n
) = f .
Indeed, the equality holds for every n N. Now, we will prove v). We may assume
that
A =
_
k
J
k
such that the J
k
s are non-overlapping. We dene
A
n
=
_
kn
J
k
.
Let > 0 and > 0 such that
[t s[ < [f(t) f(s)[ <

2
.
Now, we consider x A such that
d(x, A
c
) = inf
yB
[x y[
This means that
B(x, ) =
_
k
J
k
B(x, ) .
This implies
lim
n
[B(x, ) A
c
n
[ = 0 .
Moreover, we can nd a maximal family x
1
, ..., x
m
of such points such that d(x, A
c
)
implies d(x, x
j
) < for some j. Then, we may choose n large enough such that
sup
t
[f(t)[
m

j=1
[B(x
j
, ) A
c
n
[

2
.
Now, we dene D =

B(x
j
, ). Since A
c
n
D
c
is again a collection of intervals we
see that
d
1
(f1
A
c
n
, 0) d
1
(f1
A
c
n
D
, 0) + d
1
(f1
A
c
n
D
c, 0)
sup
t
[f(t)[
m

j=1
[B(x
j
,

2
) A
c
n
[ + sup
tA
c
n
D
c
[f(t)[ .
Now, we consider t A
c
n
D
c
. If d(t, B)

2
, then d(t, x
j
)

2
hence t D. Thus
we may assume d(t, B) <

2
. Then we nd s B such that [t s[ < and thus
[f(t)[ sup
sB
[f(s)[ + .
6. A FAMOUS EXAMPLE 25
This implies
d
1
(f1
A
c
n
, 0)

2
+ + sup
sB
[f(s)[ .
for n n
0
. The assertion is proved.
Remark 6.8. For J = [a, b] we have
I(f1
[a,b]
) =
b
_
a
f(t)dtpl.
approx Lemma 6.9. Let (A
m
) be a sequence such that
A
m
=
_
k
J
m
k
and
lim
m

k
[J
m
k
[ = 0 .
Then
lim
m
d
1
(x1
A
m
, 0) = 0
holds for every x X.
Proof. Let (f
n
) be Cauchy sequence converging to x. Let > 0. Then, we
may choose n such that
d
1
(f
n
, x) <

2
.
Since u
A
m
is Lipschitz, we deduce
d
1
(f
n
1
A
m
, u
A
m
(x)) <

2
.
for all m N. According to (
8-cont
6.1) we nd
d
1
(f
n
1
A
m
, 0) sup
t
[f
n
(t)[

k
[J
m
k
[ .
By assumption, we may nd m
0
such that
d
1
(f
n
1
A
m
, 0) <

2
holds for all m m
0
. This implies
d
1
(u
A
m
(x), 0) d
1
(u
A
m
(x), f
n
1
A
m
) + d
1
(f
n
1
A
m
, 0) <
for all m > m
0
.
26 1. METRIC SPACES
Proof of Theorem
inj
6.6. Let (f
n
), (g
n
) be Cauchy such that d
1
lim
n
f
n
= x
and d
1
lim
n
g
n
= y and f
n
converges to f and g
n
converges to g a.e. We consider
h

n
= f
n
g
n
and z = add(x, y). We want to show
d
1
(x, 0) = 0 .
Clearly, h

n
converges to 0 almost everywhere. Passing to subsequence (h
k
) we may
assume that d
1
(h
k
, h
k+1
) 6
n
. According to the proof of Proposition
limit1
6.5, we nd
A
k
=

l
J
k
l
such that

l
[J
l
k
[ 2
1k
such that
[h
n
(t) h
n+1
(t)[ 2
n
for all t / A
k
. By the denition of a.e. we nd B
k
=

J
k
l
such that

l
[

J
l
k
[ 2
1k
and
lim
t
h
n
(t) = 0
holds for all t / B
k
. Thus we dene
C
k
=
_
l
J
l
k

_
l

J
l
k
.
Then

l
[J
l
k
[ +

l
[

J
l
k
[ 2
2k
and for all t / C
k
we have
[h
n
(t)[ = [ lim
m
h
n
(t) h
m
(t)[ limsup
m
[h
n
(t) h
m
(t)[ 2
1n
.
According to Lemma (
approx
6.9) we nd a k
0
such that
d
1
(u
C
k
(z), 0) <
for all k k
0
. For all n k we deduce from Lemma
uniqueext
6.7 that
d
1
(1
C
c
k
f
n
, 0) 2
1n
.
Therefore
d
1
(u
C
c
k
(z), 0) = lim
n
d
1
(u
C
c
k
(f
n
), 0) = 0 .
Since x = add(u
C
c
k
(x), u
C
c
k
(x)), we deduce
d
1
(z, 0) = d
1
(add(u
C
k
(z), u
C
c
k
(z)), add(u
C
k
(0), u
C
c
k
(0)))
6. A FAMOUS EXAMPLE 27
d
1
(u
C
k
(z), 0) + d
1
(u
C
c
k
(z), 0) < .
Finally, we observe that
d
1
(x, y) = d
1
(add(x, y), 0)
holds by unique extension. Thus x = y.
28 1. METRIC SPACES
7. Closed and Compact Sets
Let (X, d) be a metric space. We will say that a subset A X is closed if X A is
open.
closed Proposition 7.1. Let (X, d) be a complete metric space and C X a subset. C
is closed i every Cauchy sequence in C converges to an element in C.
Proof: Let us assume C is closed and that (x
n
) is a Cauchy sequence with elements
in C. Let x = lim
n
x
n
be te limit and assume x / C. Since X C is open
B(x, ) X C
for some > 0. Then there exists an n
0
such that d(x
n
, x) < for n > n
0
. In
particular,
x
n
0
+1
B(x, )
and thus x
n
0
+1
/ C, a contradiction.
Now, we assume that every Cauchy sequence with values in C converges to an
element in C. If X C is not open, then there exists an x / C and no > 0 such
that
B(x, ) X C .
I.e. for every n N, we can nd x
n
C such that
d(x, x
n
) <
1
n
.
Hence, limx
n
= x C but x / C, contradiction.
The most important notion in this class is the notion of compact sets. We will say
that a subset C X is compact if For every collection (O
i
) of open sets such that
C
_
i
O
i
= x X [
iI
x O
i

There exists n N and i

1
, ..., i
n
such that
C O
i
1
O
i
n
.
In other words
Every open cover of C has a nite subcover .
7. CLOSED AND COMPACT SETS 29
Definition 7.2. Let X

O
i
be an open cover. Then we say that (V
j
) is an open
subcover if
X
_
j
V
j
all the V
j
are open and for every j there exists an i such that
V
j
O
i
.
It is impossible to explain the importance of compactness right away. But we can
say that there would be no discipline Analysis without compactness. The most
clarifying idea is contained in the following example.
Proposition 7.3. The set [0, 1] R is compact.
Proof: Let [0, 1]

i
O
i
. For every x [0, 1] there exists an i I such that
x O
i
.
Since O
i
is open, we can nd > 0 such that
x B(x, ) O
i
.
Using the axiom of choice, we ne a function
x
and i
x
such that
x B(x,
x
) O
i
x
.
Let us dene the relation x _ y if x < y and
y x e
x
+ e
y
.
The crucial point here is to dene
S = x [0, 1] [ x
1
, ..., x
n
:
1
2
_ x
1
_ _ x
n
_ x .
We claim a) sup S S and b) sup S = 1.
Ad a): Let y = sup S [0, 1]. Then there exists an x S with
y
y
< x y .
Then obviously x _ y. Since x S, we can nd
1
2
_ x
1
_ _ x
n
_ x _ y .
Thus y S.
Ad b): Assume y = sup S < 1. Let 0 < = min(e
y
, 1 y). Then
y + y =
y
+
y+
.
30 1. METRIC SPACES
By a), we nd
1
2
_ x
1
_ _ x
n
_ y _ y +
and thus y + S. Contradiction to the denition of the supremum.
Assertion a) and b) are proved.
Therefore we conclude 1 S and thus nd x
1
, ..., x
n
such that
1
2
_ x
1
_ _ x
n
_ 1 .
Let x
0
=
1
2
and x
n+1
= 1, then by denition of _, we have
[x
j
, x
j+1
] B(x
j
,
x
j
) B(x
j+1
,
x
j+1
) O
i
x
j
O
i
x
j+1
for j = 0, .., n. Thus, we deduce
[
1
2
, 1]
n
_
j=0
[x
j
, x
j+1
]
n+1
_
j=0
O
i
x
j
.
Doing the same trick with [0,
1
2
], we nd
[0, 1]
m+1
_
j=0
O
i
x

n+1
_
j=0
O
i
x
j
and we have found our nite subcover.
subcom Proposition 7.4. Let B X be closed set and C X be a compact set, then
B C
is compact
Proof: Let B C

O
i
be an open cover. then
C (X B)
_
i
O
i
is an open cover for C, hence we can nd a nite subcover
C (X B) O
i
1
O
i
n
.
Thus
B C O
i
1
O
i
n
is a nite subcover.
7. CLOSED AND COMPACT SETS 31
ccc Lemma 7.5. Let (X, d) be a metric space and D X be a countable dense set in
X, then for every subset C X and every open cover
C
_
i
O
i
we can nd a countable subcover of balls.
Proof: Let us enumerate D as D = d
n
[ n N. Let x C and nd i I and
> 0 such that
x B(x, ) O
i
.
Let k >
2

. By density, we can nd an n N such that

d(x, d
n
) <
1
2k
.
Then
x B(d
n
,
1
2k
) B(x,
1
k
) B(x, ) O
i
.
Let us dene
M = (n, k) [
iI
B(d
n
,
1
2k
) O
i
.
Then M N
2
is countable and hence there exists a map : N M which is
surjective (=onto). Hence for V
m
= B(d

1
(m)
,
1
2
2
(m)
),
1
,
2
the 2 components of
we have
C
_
m
V
m
and (V
m
) is a countable subcover of balls of the original cover (O
i
).
main Theorem 7.6. Let (X, d) be a metric space. Let C X be a subset. Then the
following are equivalent
i) a) Every Cauchy sequence of elements in C converges to a limit in C.
b) For every > 0 there exists points x
1
, ..., x
n
X such that
C B(x
1
, ) B(x, ) .
ii) Every sequence in C has a convergent subsequence.
iii) C is compact.
Proof: i) ii). Let (x
n
) be a sequence. Inductively, we will construct innite
subset A
1
A
2
A
3
and y
1
, y
2
, y
3
,... in X such that

lA
j
: d(x
l
, y
j
) < 2
j1
.
32 1. METRIC SPACES
Put A
0
= N. Let us assume A
1
A
2
A
n
and y
1
, ..., y
n
have been constructed.
We put = 2
n2
and apply condition i)b) to nd z
1
, ..., z
m
such that
C B(z
1
, ) B(z
m
, ) .
We claim that there must be a 1 k m such that
A
n
(k) = l A
n
[ x
n
B(z
k
, )
has innitely many elements. Indeed, we have
A
n
(1) A
n
(m) = A
n
.
If they were all nite, then a nite union of nite sets would have nitely many
elements. However A
n
is innite. Contradiction! Thus, we can nd a k with A
n
(k)
innite and put A
n+1
= A
n
(k) and y
n+1
= z
k
. So the inductive procedure is nished.
Now, we can nd an increasing sequence (n
j
) such that n
j
A
j
and deduce
d(x
n
j
, x
n
j+1
) d(x
n
j
, y
j
) + d(y
j
, x
n
j+1
) <
1
2
2
j
+
1
2
2
j
= 2
j
because n
j
A
j
and n
j+1
A
j+1
A
j
. Thus (x
n
j
) is Cauchy. Indeed, be induction,
we deduce for j < m that
d(x
n
j
, x
n
m
) d(x
n
j
, x
n
j+1
) + d(x
n
j+1
, x
n
j+2
) d(x
n
m1
, x
n
m
)
2
j
m1

k=0
2
k
= 2
1j
.
This easily implies the Cauchy sequence condition. By a) it converges to some
x C. We got our convergent subsequence.
ii) iii): We will rst show ii) i)b). Indeed, let > 0 and assume for all n N,
y
1
, ..., y
n
C we may nd
x(n, y
1
, ..., y
n
) C (B(y
1
, ) B(y
n
, )) .
Then we dene x
1
C and nd x
2
C B(x
1
, ). Then we nd x
3
C B(x
1
, )
B(x
2
, ). Thus inductively we nd x
n
C such that
d(x
n
, x
k
)
for all 1 k n. It is easily seen that (x
n
) has no convergent subsequence. Thus i)b)
is showed (with points in C). For every
k
=
1
k
we nd these points y
k
1
, ...., y
k
m(k)
C
such that
C B(y
k
1
,
1
k
) B(y
m(k)
,
1
k
) .
7. CLOSED AND COMPACT SETS 33
Then, we see that D = y
k
j
: k N, 1 j m(k) is dense in C. Therefore, we
may work with the closure

X =

D and show that C is compact in

X. (It will then
be automatically compact in X). By Lemma
ccc
7.5, we may assume that
C
_
k
O
k
and O
k
s open. If we can nd an n such that
C O
1
O
n
the assertion is proved. Assume that is not the case and choose for every n N
an x
n
C O
1
O
n
. According to the assumption, we have a convergent
subsequence, i.e. lim
k
x
n
k
= x C. Then x O
n
0
for some n
0
and there exists a
> 0 such that
B(x, ) O
n
0
.
By convergence, we nd a k
0
such that d(x, x
n
k
) < for all k > k
0
. In particular,
we nd a k > k
0
such that n
k
> n
0
. Thus
x
n
k
B(x, ) O
n
0
O
1
O
n
k
.
Contradicting the choice of the (x
n
)s. We are done.
iii) i)b) Let > 0 and then
C
_
xC
B(x, ) .
thus a nite subcover yields b).
iii) i)a) Let (x
n
) be a Cauchy sequence. Assume it is not converging to some
element x C. This means
cccc (7.1) x C(x) > 0n
0
n > n
0
d(x
n
, x) > .
Then
C
_
xC
B(x,
(x)
2
) .
Let
C B(y
1
,
(y
1
)
2
) B(y
1
,
(y
1
)
2
)
be a nite subcover (compactness). Then there exists at least one 1 k m such
that
A
k
= n N [ d(x
n
, y
k
) <
(y
k
)
2

34 1. METRIC SPACES
is innite. Fix that k and apply the Cauchy criterion to nd n
0
such that
d(x
n
, x
n
) <
(y
k
)
2
for all n, n

> n
0
. By (
cccc
7.1), we can nd an n > n
0
such that
d(x
n
, y
k
) > (y
k
) .
Since A
k
is innite, we can nd an n

> n
0
in A
k
thus
(y
k
) < d(x
n
, y
k
) d(x
n
, x
n
) + d(x
n
, y
k
)
<
(y
k
)
2
+
(y
k
)
2
= (y
k
) .
A contradiction. Thus the Cauchy sequence has to converge to some point in C.
Corollary 7.7. Every intervall [a, b] R with a < b R is compact
Proof: It is easy to see that X [a, b] is open. Hence, by Proposition
closed
7.1 [a, b] is
complete, i.e. i)a) is satised. Given > 0, we can nd k >
1

. For m > k(b a) we

derive
[a, b]
m
_
j=0
B(a +
j
k
, ) .
Thus the Theorem
main
7.6 applies.
cube Lemma 7.8. Let r > 0 and n N, the set C
r
= [r, r]
n
is compact.
Proof: Let x / C
r
, then there exists an index j 1, .., n such that [x
j
[ > r. Let
= [x
j
[ r and y R
n
such that
max
i=1,..,n
[x
i
y
i
[ < ,
then
[y
j
[ = [y
j
x
j
+ x
j
[ [x
j
[ [y
j
x
j
[ > [x
j
[ = r .
thus y / C
r
. Hence, C
r
is closed and according to Proposition
rncomp
4.3, we deduce that
C
r
is complete.
For n = 1 and > 0, we have seen above that for k >
1

and m >
2r
k
[r, r]
m
_
j=0
B(r +
j
k
, ) .
Therefore
[r, r]
n

_
j
1
,....j
n
=0,...m
B

((r +
j
1
k
, .., r +
j
n
k
), ) .
Thus i)a) and i)b) are satises and the Theorem
main
7.6 implies the assertion (The
separable dense subset is Q
n
.)
7. CLOSED AND COMPACT SETS 35
Theorem 7.9. Let C R
n
be a subset. The following are equivalent
1) C is compact.
2) C is closed and there exists an r such that
C B(0, R) .
(That is C is bounded.)
Proof: 2) 1) Let
C B(0, R) [R, R]
n
be a closed set. Since [R, R]
n
is compact, we deduce from Proposition
subcom
7.4 that C
is compact as well.
1) 2) Let C subset R
n
be a compact set. According to Theorem
main
7.6 i)b), we nd
C B(x
1
, 1) B(x
m
, 1)
thus for r = max
i=1,..,m
(d(x
i
, 0) + 1) we have
C B(0, r) .
Moreover, by Theorem
main
7.6 i)a) and Proposition
closed
7.1, we deduce that C is closed.
We will now discuss one of the most important applications.
Theorem 7.10. Let (X, d) be a compact metric space and f : X R be a continuous
function. The there exists x
0
X such that
f(x
0
) = supf(x) : x X .
Proof. Let us rst assume
A = f(x) : x X
is bounded and s = sup A. For every n N, we know that s
1
n
is no upper bound.
Hence there x
n
X such that
s f(x
n
) > s
1
n
.
Let (n
k
) be such that lim
k
x
n
k
= x X. Then we deduce from continuity that
f(x) = lim
k
f(x
n
k
) lims
1
n
k
= s .
By denition of s we nd f(x) = s. Now, we show that A is bounded. Indeed, if
note we nd x
n
X such that f(x
n
) n. Again we nd a convergent subsequence
36 1. METRIC SPACES
(x
n
k
). Since f(x
n
k
) is convergent it is bounded. We assume (f
n
k
) is bounded above
by m nz. Choosing k m + 1 we get
m f(x
n
k
) n
k
> n
m
m .
This contradiction shows that A is bounded and hence the rst argument applies.