Lecture 14 - Dense sets and separable spaces

We still need to prove “part 2” of the theorem from the last lecture, namely that
“sequentially compact ⇒ compact”.
For this we need some preparation, which will take up the next two lectures.
Definition
A set D is dense in the metric space (X, d) if each point of X\D is a limit point
of D.
(i.e. elements of X \ D can be approximated arbitrarily well by elements of D)
Examples
(1) Both Q and R\Q are dense in (R, | · |).
(2) C ∞ [a, b] is dense in (C[a, b], k · ksup )
(This follows from Weierstrass Approximation Theorem - see later)
Definition
A metric space (X, d) is separable if it has a countable dense subset.
Examples
(1) (R, | · |) is separable, and Rn with k · k2 , k · k1 , k · kp are separable as they
contain the dense countable set Qn .
(2) `∞ is not separable.
For any G ⊂ N let eG ∈ `∞ be the sequence such that eG
n = 1 if n ∈ G and
eG
n = 0 otherwise.
(So e.g. e{2,4,5} = 0, 1, 0, 1, 1, 0, 0, . . .)
0
{eG : G ⊂ N} is an uncountable set in `∞ such that keG − eG k∞ = 1
whenever G 6= G0 . (Uncountability follows by a “diagonal argument”)
Hence there are uncountably many non-intersecting open balls of radius 1/2
(if we center them in eG ). Here disjointness holds because
0
keG − eG k∞ = 1 ≥ 1/2 + 1/2 for G 6= G0 .
Suppose D is a countable set in `∞ .
It cannot have points in all the balls above so some balls will have no points
from D in them.
The centre of any such ball is not a limit point of D.
Lemma
If X is sequentially compact then (X, d) is separable.
Proof:
Let X be sequentially compact.
We claim that for every δ there exists Dδ = {x1 , . . . , xn(δ) } with the property
that for every x ∈ X there is y ∈ Dδ such that d(x, y) < δ. (That is,
n(δ)
X ⊂ ∪i=1 B o (xi , δ).)

To prove this, suppose that δ > 0 is such that no such finite set Dδ exists. Pick
any x1 ∈ X, and then pick x2 , x3 , . . . inductively as follows: having chosen
x1 , . . . , xn ∈ X, choose xn+1 ∈ X so that d(xi , xn+1 ) ≥ δ for i = 1, . . . , n. (The
existence of such an xn+1 is guaranteed by our assumption on δ.)
By this procedure we have constructed a sequence (xn )n∈N such that
d(xn , xm ) ≥ δ for all n 6= m.
Since X is sequentially compact there is a convergent subsequence xni → x ∈ X.
But then (xni ) is Cauchy, contradicting the fact that d(xn , xm ) ≥ δ for n 6= m.
Having proved the existence of Dδ for each δ > 0, set D = ∪n∈N D1/n .
D is countable as a countable union of finite sets.
D is also dense since for every x ∈ X\D and every B o (x, r) we can first pick n so
that 1/n < r and then pick y ∈ D1/n satisfying d(x, y) < 1/n < r. Hence
B o (x, r) contains y ∈ D, y 6= x, so x is a limit point of D.
 Lecture 15 - Bases

Definition
A collection {Vα }α∈A of open sets in a metric space (X, d) is said to be a base
for X if for every open set G ⊂ X and every x ∈ G we have x ∈ Vα ⊂ G for some
α ∈ A.

Example
The collection of all open balls is a base.

So the notion of a “base” generalises the set of open balls. This is important in
the setting of topological spaces, where we don’t have a metric with which to
define balls - see later.

Remark
Every open set can be written as the union of a subcollection of the base.
Proof: G = ∪α∈A0 Vα where A0 = {α(x) : x ∈ G} ⊂ A, where, for x ∈ G, α(x)
denotes any α ∈ A for which x ∈ Vα ⊂ G.
Lemma
Every separable metric space has a countable base.

Example
In (R, | · |), take all open balls centered at rational points with rational radii, i.e.,
all open intervals with rational endpoints. (Extension to (Rn , k · k2 ) is obvious)
(This suggests a strategy for the proof in the general case!)

Proof:
Let D be a countable dense subset of X.
Consider the collection {B o (z, 1/n)}z∈D,n∈N , which is countable.
To show that it is a base, pick any open set G ⊂ X and x ∈ G.
Since G is open, there is B o (x, r) ⊂ G.
Let n be such that 1/n < r.
1 1
Since D is dense in X there is y ∈ D such that d(x, y) < 2n , so x ∈ B o (y, 2n ).
o 1 o
It remains to show that B (y, 2n ) ⊂ G. But since B (x, r) ⊂ G it’s enough to
1
show that B o (y, 2n ) ⊂ B o (x, r).
1
For this, note that for every z ∈ B o (y, 2n ) we have
d(x, z) ≤ d(x, y) + d(y, z) < 2n + 2n = n1 < r.
1 1
Lemma
If X is sequentially compact then X is compact.

Proof:
Let {Gα }α∈A be a cover of X.
We want to show there is a finite subcover. But let’s first try and show there is a
countable subcover.
Since X is sequentially compact, (X, d) is separable and so has a countable base
{Vn }n∈N .
Note that for every x ∈ X there exists Gα and Vn such that x ∈ Vn ⊂ Gα .
Denote these by Vn(x) .
Since {Vn(x) }x∈X is a sub-collection of the base it is countable and covers X.
Denote it by {Vn0 }n∈N .
Observe that by construction for every Vn0 there exists Gαn such that Vn0 ⊂ Gαn .
Thus there is a countable subcover {Gαn }n∈N of X.

Now we need to prove that this countable subcover has a finite subcover!
Proof: (continued)
Suppose there is no finite subcover.
For each n ∈ N let Fn = X \ Gα1 ∪ · · · ∪ Gαn ), which is closed and non-empty
(else {Gα1 , . . . , Gαn } would be a finite subcover). Also Fn+1 ⊂ Fn for all n.
Since {Gαn }n∈N covers X, we have ∩∞ ∞
n=1 Fn = X \ (∪n=1 Gαn ) = X \ X = ∅.
However, in a sequentially compact space, a decreasing sequence of nested
non-empty closed sets has a non-empty intersection. To prove this let (xn ) be any
sequence such that xn ∈ Fn for each n. By sequential compactness it has a
subsequence (xni ) converging to some x ∈ X. By nestedness it follows that
x ∈ Fni for all i, and hence that x ∈ ∩∞ ∞
i=1 Fni = ∩n=1 Fn .
So ∩∞
n=1 Fn is non-empty, which gives a contradiction.
Hence there is a finite subcover, so X is compact.
We’ve shown that “sequentially compact ⇒ compact” for metric spaces.
It now remains to transfer this to subsets of a given (not necessarily compact)
metric space. This is easy using the following trivial result.

Lemma
Let (X, d) be a metric space and let S ⊂ X.
Then the metric space (S, d) is compact (sequentially compact) if and only if S is
compact (sequentially compact) as a subset of (X, d).

Proof: (of forward implication for compactness - the rest is an exercise)

Let {Gα }α∈A be a cover of S in X.
Then the sets G0α = Gα ∩ S form a cover of (S, d).
By the compactness of (S, d) there is a finite subcover {G0α1 , · · · , G0αm } of S in
(S, d) and so {Gα1 , · · · , Gαm } is a finite subcover of S in (X, d).

Proof: (of part 2 of theorem: “sequentially compact ⇒ compact”)

Suppose S is sequentially compact in (X, d).
Then the metric space (S, d) is sequentially compact, implying by the previous
lemma that (S, d) is compact.
Hence S is a compact subset of X.
 Lecture 16 - Totally bounded sets
We always have that “compact ⇒ closed and bounded”.
The converse holds in (Rn , k · k2 ), but not in general.
What can we say in complete metric spaces more general than (Rn , k · k2 )?
Definition
Let ε > 0. A set Sε ⊂ S is an ε-net for a set S if S ⊂ ∪x∈Sε B o (x, ε).
S is totally bounded if, for any ε > 0, it has a finite ε-net.
 Lecture 16 - Totally bounded sets
We always have that “compact ⇒ closed and bounded”.
The converse holds in (Rn , k · k2 ), but not in general.
What can we say in complete metric spaces more general than (Rn , k · k2 )?
Definition
Let ε > 0. A set Sε ⊂ S is an ε-net for a set S if S ⊂ ∪x∈Sε B o (x, ε).
S is totally bounded if, for any ε > 0, it has a finite ε-net.

Remark
Totally bounded ⇒ bounded. To see this, take e.g. ε = 1 and note that
diam S ≤ 2 + diamS1 < ∞.
Totally bounded ⇒ separable. (Exercise)
If there exists ε > 0 such that S contains infinitely many non-intersecting
open balls of radius ε then S is not totally bounded.
(Proof: if Sε is an ε-net for S then each of the balls must contain at least
one point of Sε , to ensure that the ball centres are covered by
∪x∈Sε B o (x, ε), and hence Sε must be infinite.)
Examples
(1) In R and Rn , totally bounded ⇔ bounded.
(2) Finite ⇒ totally bounded.
(3) In a discrete space, totally bounded = finite.
In particular, bounded 6⇒ totally bounded
(4) The unit ball in `∞ is not totally bounded.
Proof 1: Use the fact that `∞ is not separable.
Proof 2: There are infinitely many non-intersecting open balls of radius 1/4
centered around 21 ek , k ∈ N. Disjointness holds because
k 12 em − 12 en k∞ = 1/2 for m 6= n.
(Compare to our proof of non-separability of `∞ , where we constructed an
uncountable collection of non-intersecting balls of a fixed radius.)
Theorem
Suppose (X, d) is complete.
A set S ⊂ X is compact if and only if it is closed and totally bounded.

Proof:
(⇒) Suppose S is compact. We know it is then closed.
To prove that it is totally bounded, let ε > 0 and consider the cover
{B o (x, ε)}x∈S .
By compactness there is a finite subcover {B o (x1 , ε), . . . , B o (xm , ε)} and so
{x1 , . . . , xm } is a finite ε-net.
(⇐) Suppose S is closed and totally bounded, so a finite ε-net Sε exists ∀ε > 0.
Suppose S is not compact, i.e., there is a cover {Gα }α∈A which has no finite
subcover.
We construct a sequence (xn ) as follows.
Let x1 ∈ S1/2 be such that there is no finite subcover of B o (x1 , 1/2) ∩ S.
(If no such x1 exists then, by the finiteness of S1/2 , S has a finite subcover.)
Let x2 ∈ S1/22 be such that B o (x1 , 1/2) ∩ B o (x2 , 1/22 ) 6= ∅ and there is no
finite subcover of B o (x2 , 1/22 ) ∩ S.
···
Let xn ∈ S1/2n be such that B o (xn−1 , 1/2n−1 ) ∩ B o (xn , 1/2n ) 6= ∅ and
there is no finite subcover of B o (xn , 1/2n ) ∩ S.
Proof: (continued)
The sequence (xn ) is Cauchy: for n > m
n−1 n−1 1 n−1 ∞
X X 1  3X 1 3X 1 3
d(xn , xm ) ≤ d(xi , xi+1 ) < + = < = m → 0.
i=m i=m
2i 2i+1 2 i=m 2i 2 i=m 2i 2

Since (X, d) is complete, xn converges to some x ∈ X.

Since all xn ∈ S and S is closed we have x ∈ S.
Let α ∈ A be such that x ∈ Gα . Since Gα is open there is B o (x, r) ⊂ Gα .
Pick n large enough so that d(xn , x) < r/2 and 1/2n < r/2.
Then y ∈ B o (xn , 1/2n ) implies
d(x, y) ≤ d(x, xn ) + d(xn , y) < r/2 + 1/2n < r.
Hence B o (xn , 1/2n ) ⊂ B o (x, r) ⊂ Gα , contradicting B o (xn , 1/2n ) ∩ S not
having a finite subcover.

(Note: This proof is similar to the proof of Heine-Borel)

Exercise: draw some pictures to help you understand the idea of the proof!
Remark
The completeness assumption isn’t needed for the ⇒ implication. But if (X, d) is
not complete then the ⇐ implication may fail. In particular, the next theorem
shows that, in this case, X itself is non-compact, even if X is totally bounded.
(Remember that X is closed in (X, d).)

Theorem
A metric space is compact if and only if it is complete and totally bounded.
Proof:
(⇒) Suppose X is compact. Then it is totally bounded (by the remark above).
To prove that (X, d) is complete, let (xn ) be a Cauchy sequence in X.
Since X is sequentially compact there is a convergent subsequence xni → x.
Then the whole sequence xn → x.
To see this, for any ε > 0, pick N so that for all n, m ≥ N we have
d(xn , xm ) < ε/2.
Then choose i so that ni ≥ N and d(xni , x) < ε/2.
Now for all n ≥ N we have
d(xn , x) ≤ d(xn , xni ) + d(xni , x) < ε/2 + ε/2 = ε.
(⇐) Suppose (X, d) is complete and totally bounded.
X is also closed so it is compact.