Math 4061(Topics in Modern Analysis) c Math Department, HKUST

Chapter 1. Preliminary Facts

§1. Notations and Terminologies

We begin with some basic abbreviations, notations and terminologies. For convenience, we will write
“iff” for “if and only if”. For the common number systems, we will use
Z to denote the set of all integers, N or Z+ to denote the set of all positive integers,
Q to denote the set of all rational numbers, Q+ to denote the set of all positive rational numbers,
R \ Q to denote the set of all irrational numbers,
R to denote the set of all real numbers, R+ to denote the set of all positive real numbers and
C to denote the set of all complex numbers.
Now for sets and mappings, we use the following terminologies.

Definitions. (1) The notation X ⊆ Y or Y ⊇ X means X is a subset of Y and possibly X = Y.

(2) The notation X ⊂ Y or Y ⊃ X means X is a subset of Y and X 6= Y.
(3) For sets A and B, the set A × B = {(a, b) : a ∈ A, b ∈ B} is the Cartesian product of A and B.
(4) A function f : A → B is injective or one-to-one iff for all x, y ∈ A, f(x) = f(y) implies x = y.
(5) A function f : A → B is surjective or onto iff for all y ∈ B, there exists x ∈ A such that f(x) = y.
(6) A function f : A → B is bijective or a one-to-one correspondence iff f is injective and surjective.
(7) For f : A → B and C ⊆ A, the restriction of f to C is the function f|C : C → B defined by f|C (x) = f(x)
for all x ∈ C.

§2. Equivalence Relation

Equivalence relation is an important concept and it appears in almost all branches of mathematics.

Definitions. (1) A (binary) relation on a set S is a subset R of S × S.

(2) An equivalence relation R on a set S is a (binary) relation on S such that
(a) (reflexivity) for all x ∈ S, (x, x) ∈ R
(b) (symmetry) if (x, y) ∈ R, then (y, x) ∈ R
(c) (transitivity) if (x, y) ∈ R and (y, z) ∈ R, then (x, z) ∈ R.
(3) For an equivalence relation R on S and a ∈ S, the set [a] = {x ∈ S : (x, a) ∈ R} is called the
equivalence class of a.
(4) A partition of S is a collection of pairwise disjoint nonempty subsets of S whose union is S.

Theorem. Let R be an equivalence relation on S. Then

(1) for every a ∈ S, a ∈ [a]
(2) [a] = [b] iff (a, b) ∈ R
(3) for every pair [a] and [b], either [a] = [b] or [a] ∩ [b] = ∅.
Thus, the equivalence classes form a partition of S.
Proof. (1) a ∈ [a] iff (a, a) ∈ R, which follows from reflexivity.

1
(2) If [a] = [b], then a ∈ [a] = [b] implies (a, b) ∈ R. Conversely, if (a, b) ∈ R, then for every x ∈ [a],
(x, a) ∈ R. By transitivity, (x, b) ∈ R. So x ∈ [b]. Hence, [a] ⊆ [b]. Next, by symmetry, (b, a) ∈ R. So the
same argument also gives [b] ⊆ [a]. Therefore, [a] = [b].
(3) If [a] ∩ [b] 6= ∅, then there exists x ∈ [a] ∩ [b]. Since x ∈ [a] and x ∈ [b], we get (x, a) ∈ R and (x, b) ∈ R.
By (2), [a] = [x] = [b].

Example. Let R = {(a, b) : a, b ∈ Z and a − b is even}. We can easily check this is an equivalence relation
on Z. There are two equivalence classes, namely the set of all even integers [0] = {a : a ∈ Z, a is even} and
the set of all odd integers [1] = {a : a ∈ Z, a − 1 is even}.

Remarks. (1) The purpose of equivalence relation on a set is to partition the set into subsets which have
some structure for solving certain types of problems. For the last example, we have three levels of structures
for the set of integers. The outer level is the set of all integers as a whole. The intermediate level is the
division of the integers into two parts, namely the even integers and the odd integers. The inner level is the
division of the integers into individual numbers.
For the outer level, we have facts like integer+integer=integer and integer×integer=integer. For the
intermediate level, we have facts like odd+odd=even, even+odd=odd, even+even=even, odd×odd=odd,
even×odd=even, even×even=even. For the inner level, we have facts like 0+1 = 1, 0+2 = 2, 1+2 = 2, . . . .
Some problems can be solved faster by using facts from the different levels.
Consider the following three problems:
(a) Is there any integral solution to the equation x + y = 1/2?
(b) Is there any integral solution to the equation 2x(2y + 1) = 15?
(c) Is there any integral solution to the equation x3 + y + y3 x = 23?
Solutions. (a) No, use integer+integer=integer. (Facts from the outer level solves this problem.)
(b) No, use even×odd=even. (Facts from the intermediate level solves this problem. Facts from the outer
level are not too effective.)
(c) Yes, x = 0, y = 23. (Trying small numbers is faster. Facts from the intermediate level helps, but still
more case-by-case considerations are needed.)

(2) Equivalence relations appear in different parts of mathematics. For example, in linear algebra, a subspace
of a vector space can be used to form a quotient space. In number theory, the modulo concept (mod m)
came from partitioning integers into m subsets (that generalize the even and odd concepts).

§3. Useful Inequalities

The following famous inequalities will be useful in later chapters.

1 1 p αp β q
Theorem. (Young’s Inequality) For p > 1, let + = 1 (i.e. q = ). If α, β ≥ 0, then αβ ≤ + .
p q p−1 p q
Proof. The case α = 0 or β = 0 is clear. So consider α, β > 0. Define f : R → R by f(x) = ex . Now
f 00 (x) = ex > 0 implies f(x) is convex on R. Let x = p ln α, y = q ln β and t = 1/p ∈ (0, 1). By convexity,
αp β q
αβ = f(tx + (1 − t)y) ≤ tf(x) + (1 − t)f(y) = + .
p q

1 1
Theorem. (Hölder’s Inequality) For p > 1, let + = 1. If λ = (λ1 , λ2, . . . , λn), µ = (µ1 , µ2, . . . , µn) ∈ Cn,
p q
then
X
n X
n 1/p X
n 1/q
|λi µi| ≤ |λi |p |µi |q .
i=1 i=1 i=1

2
Proof. If λ or µ is (0, 0, . . ., 0), then both sides are zero. So suppose λ and µ are not (0, 0, . . ., 0). Let
.X
n 1/p .X n 1/q X
n X
n
αj = |λj | |λi|p and βj = |µj | |µi |q , then αpi = 1 and βiq = 1. The desired inequal-
i=1 i=1 i=1 i=1
ity follows from Young’s inequality as
Pn n  p
βiq  1 1
n
X X
i=1 |λi µi | α i
P 1/p P 1/q = α i βi ≤
p
+
q
= + = 1.
p q
n p n
i=1 |λi | i=1 |µi|q i=1 i=1

Theorem. (Minkowski’s Inequality) For p ≥ 1, if x = (x1, x2, . . . , xn), y = (y1 , y2 , . . . , yn) ∈ Cn, then
X
n 1/p X
n 1/p X
n 1/p
|xi + yi |p ≤ |xi|p + |yi|p .
i=1 i=1 i=1

Proof. The case p = 1 follows from the triangle inequality for C. Let p > 1 and ri = |xi + yi |. Then
X
n X
n X
n X
n
ri ≤ |xi| + |yi| and so rip ≤ (|xi| + |yi |)rip−1 = |xi|rip−1 + |yi |rip−1. Applying Hölder’s inequality
i=1 i=1 i=1 i=1
to the two terms on the right and recalling (p − 1)q = p, the right side is at most
X
n 1/p X
n 1/q X
n 1/p X
n 1/q  X
n

1/p Xn

1/pX
n 1/q
(p−1)q (p−1)q
|xi|p ri + |yi|p ri = |xi|p + |yi |p rip .
i=1 i=1 i=1 i=1 i=1 i=1 i=1

X
n

1/q
Cancelling rip , we get the desired inequality.
i=1

Similarly, we can prove the integral versions of these inequalities. Below m denotes the Lebesgue measure
on R.
1 1
Theorem. (Hölder’s Inequality) For p > 1, let + = 1. If |f|p and |g|q are Lebesgue integrable on a
p q
Lebesgue measurable set X, then
Z Z 1/p Z 1/q
|fg| dm ≤ |f|p dm |g|q dm .
X X X

|f(t)| |g(t)|
(To get a proof, just let α(t) =  1/p , β(t) = R 1/p , apply Young’s inequality to
R
X
|f|p dm X
|g|p dm
α(t)β(t) and integrate over X. Imitate the above proof of Hölder’s inequality for numbers.)

Theorem. (Minkowski’s Inequality) For p ≥ 1, if |f|p and |g|p are Lebesgue integrable on a Lebesgue
measurable set X, then
Z 1/p Z 1/p Z 1/p
p p
|f + g| dm ≤ |f| dm + |g|p dm .
X X X

(To get a proof, just let r(t) = |f(t) + g(t)| and replace summation by integration over X to imitate the
above proof of Minkowski’s inequality for numbers.)
Remark. For the case p = 2, the Hölder inequality is also known as the Cauchy-Schwarz inequality.

3
 Chapter 2. Metric Spaces

§1. Definition and Examples of Metric Spaces

We first define metric spaces and then we will provide some examples and remarks.

Definitions. (1) A metric space is a nonempty set M with a function d : M × M → R such that for every
x, y, z ∈ M,
(a) d(x, y) ≥ 0 and equality holds if and only if x = y,
(b) d(x, y) = d(y, x) and
(c) (Triangle Inequality) d(x, z) ≤ d(x, y) + d(y, z).
Such a function d on M × M is called a metric on M. To be precise, we should call (M, d) the metric space.
(2) If M is a metric space with metric d and N ⊆ M, then we can restrict d to N × N to provide a metric
for N. We call this the induced metric from d for N.
(3) For a sequence {xn} in M, we say {xn} converges to x iff lim d(xn, x) = 0 or more precisely, for every
n→∞
ε > 0, there is a positive integer K such that n ≥ K ⇒ d(xn, x) < ε. In this case, our notation is lim xn = x
n→∞
or just xn → x.

Remark. If {xn } converges to x and y, then x = y since d(x, y) ≤ d(x, xn) + d(xn, y) → 0 implies d(x, y) = 0
by the sandwich theorem. So limits of sequences are unique if they exist.

1 if x 6= y
Examples. (1) For every nonempty set M, define d(x, y) = for all x, y ∈ M. It is easy to check
0 if x = y
that d satisfies conditions (a), (b), (c) above. Hence, this function d is a metric on M. This is called the
discrete metric on M. So every nonempty set M has at least one metric.
(2) Some sets can have more than one metric! Let M = Rn or Cn . For every real number p ≥ 1, let
x = (x1 , x2, . . . , xn) and y = (y1 , y2, . . . , yn ) be in M. Define
 1/p
dp (x, y) = |x1 − y1 |p + |x2 − y2 |p + · · · + |xn − yn |p .

It is easy to see that dp satisfies conditions (a) and (b) above. Condition (c) follows from Minkowski’s
inequality. Therefore, dp is a metric.
If 0 < p < 1, then dp is not a metric, since taking x = (1, 0, 0, . . ., 0), y = (0, 1, 0, . . ., 0) and z =
(0, 0, 0, . . ., 0), we get dp (x, z) + dp (z, y) = 1 + 1 < 21/p = dp(x, y). So triangle inequality fails.
(3) Let S be a nonempty set. Let B(S, R) be the set of all bounded functions f : S → R and for every
f, g ∈ B(S, R), define d(f, g) = sup{|f(x) − g(x)| : x ∈ S}. Clearly, d satisfies conditions (a) and (b) for a
metric. Taking supremum of the left side of the inequality

|f(x) − h(x)| ≤ |f(x) − g(x)| + |g(x) − h(x)| ≤ d(f, g) + d(g, h),

we see condition (c) also hold. Therefore, B(S, R) is a metric space. Similarly, we may replace R by C and
use the same metric d to form a metric space B(S, C).
n
X
1/2
Remarks. (1) For Rn or Cn , d2(x, y) = |xi − yi |2 is called the usual metric.
i=1

X
n

1/p
(2) Observe that lim |ai|p = max{|a1|, |a2|, . . ., |an|}. (To see this, let M = max{|a1|, |a2|, . . . , |an|},
p→∞
i=1
X
n

1/p
then M ≤ |ai|p ≤ (nM p )1/p = n1/p M and taking limit as p → ∞, we get the desired equation from
i=1

4
the sandwich theorem.) Thus, we may define d∞ (x, y) = max{|x1 − y1 |, |x2 − y2 |, . . ., |xn − yn |}. We can see
this is also a metric on Rn and Cn by taking the limit of the triangle inequality for dp as p → ∞.

§2. Topological Properties in Metric Spaces

Definitions. Let M be a metric space with metric d(x, y). Let S ⊆ M, x0 ∈ M and r > 0.
(1) The set B(x0 , r) = {x ∈ M : d(x, x0) < r} is the open ball in M with center x0 and radius r > 0.
(2) x0 is an interior point of S iff there exists r > 0 such that B(x0 , r) ⊆ S. (Interior points of S are in S.)
(3) The set S ◦ = {x0 ∈ M : x0 is an interior point of S} is called the interior of S in M. (S ◦ ⊆ S) So ∅◦ = ∅.
(4) S is open in M iff S ◦ = S. So ∅ is open in M.
(5) x0 ∈ M is a cluster point (or accumulation point or limit point) of S iff every B(x0 , r) contains at least
one point x ∈ S and x 6= x0 . (Cluster points of S may or may not be in S.)
(6) The set S 0 = {x0 ∈ M : x0 is a cluster point of S} is called the derived set of S in M. (∅0 = ∅)
(7) The set S = S ∪ S 0 is called the closure of S in M. (S ⊆ S) ∅ = ∅.
(8) S is closed in M iff S = S (or equivalently S 0 ⊆ S). So ∅ is closed.
(9) S is dense in M iff S = M.
(10) S is bounded in M iff S is a subset of some open ball in M.
(11) A point a ∈ S is an isolated point of S iff there exists r > 0 such that B(a, r) ∩ S = {a}.
(12) A point x ∈ M is a boundary point of S iff for all r > 0, B(x, r) contains a point in S and a point not
in S. (Boundary points of S may or may not be in S.)
(13) The set ∂S = {x ∈ M : x is a boundary point of S} is called the boundary of S in M.

Examples. (1) In M
= R, the usual metric is d(x, y) = |x − y|. B(x0 , r) = (x0 − r, x0 + r) is an open interval.
Let S = Q ∩ [0, 1) ∪ (2, 3) ∪ {4}. Then
(i) S ◦ = (2, 3) 6= S (so S is not open in M ).
(ii) S 0 = [0, 1] ∪ [2, 3], S = [0, 1] ∪ [2, 3] ∪ {4} =
6 S (S is not closed in M ), S 6= M (S is not dense in M ).
(iii) S ⊆ B(0, 5) (so S is bounded in M ).
(iv) 4 is the only isolated point of S.
(v) ∂S = [0, 1] ∪ {2, 3, 4}.

(2) In M = R2, for 1 ≤ p ≤ ∞, let Bp (0, 1) denote the open ball in M with the dp metric, center at 0 = (0, 0)
and radius 1. Then
(i) B1 (0, 1) = {(x, y) ∈ R2 : |x| + |y| < 1} is the set of all points inside the square with vertices at
(1, 0), (0, 1), (−1, 0) and (0, −1).
(ii) B2 (0, 1) = {(x, y) ∈ R2 : x2 + y2 < 1} is the set of all points inside the circle center at (0, 0) and radius
1. B3 (0, 1) is oval-shaped and contains B2 (0, 1).
(iii) B∞ (0, 1) = {(x, y) ∈ R2 : |x| < 1 and |y| < 1} is the set of all points inside the square with vertices at
(1, 1), (−1, 1), (−1, −1) and (1, −1).
Next consider 0 = (0, 0) and the closed convex set L = {(1, y) ∈ R2 : y ∈ R}, which is the vertical line
through (1, 0). For 1 ≤ p < ∞, there is a unique closest point (in the dp metric) to 0 on L, namely (1, 0).
For p = ∞, the closest points (in the d∞ metric) to 0 on L are {(1, y) ∈ R2 : −1 ≤ y ≤ 1}. So closest point
on L, although exists, is not unique in this case.

5
(3) For every nonempty set M, take the discrete metric d. For every x ∈ M, B(x, 2) = M and so open balls
can be any shape! In particular, open balls may not be convex!

Below we will prove some basic facts. Let M be a metric space with metric d and let S ⊆ M.

Theorem. S is closed in M iff M \ S is open in M. In particular, ∅ and M are both open and closed in M.
Proof. S is closed in M iff S 0 ⊆ S iff (M \ S) ∩ S 0 = ∅ iff for every x ∈ M \ S, x 6∈ S 0 , i.e. there exists
B(x, r) ⊆ M \ S iff (M \ S)◦ = M \ S iff M \ S is open in M. Next, since ∅◦ = ∅ and ∅0 = ∅, ∅ is open and
closed in M. By the first part, M \ ∅ = M is both open and closed in M.

Ball-in-Ball Lemma. Let x ∈ B(x0 , r0). Then B(x, r) ⊆ B(x0 , r0) ⊆ B(x, R) for 0 < r ≤ r0 − d(x, x0) and
R ≥ r0 + d(x, x0).
Proof. For all y ∈ B(x, r), we have d(y, x0) ≤ d(y, x) + d(x, x0) < r + d(x, x0) ≤ r0. So B(x, r) ⊆ B(x0 , r0).
For z ∈ B(x0 , r0), we have d(z, x) ≤ d(z, x0) + d(x0, x) < r0 + d(x0 , x) ≤ R. So B(x0 , r0) ⊆ B(x, R).
Structure Theorem for Open Sets. S is open in M iff S = ∪ B(xα , rα) for some xα ∈ M and rα > 0.
α

Proof. If S = ∪ B(xα , rα), then every x ∈ S is in some B(xα , rα). By the ball-in-ball lemma, B(x, r) ⊆
α
B(xα , rα) ⊆ S, where r = rα − d(x, xα). So every x ∈ S is an interior point of S, i.e. S is open in M.
Conversely, if S is open in M, then every x ∈ S is an interior point of S, i.e. there is rx > 0 such that
B(x, rx ) ⊆ S. Hence, S = ∪ {x} ⊆ ∪ B(x, rx) ⊆ S. Therefore, S = ∪ B(x, rx ).
x∈S x∈S x∈S

Sequential Closure Theorem. (1) x ∈ S 0 iff there exist distinct x1 , x2, x3, . . . ∈ S such that lim xn = x.
n→∞

(2) x ∈ S iff there exist x1 , x2, x3, . . . ∈ S such that lim xn = x.

n→∞

Proof. (1) If there exist distinct x1, x2, x3, . . . ∈ S such that lim xn = x, then for every r > 0, there is K
n→∞
such that xK , xK+1, xK+2 , . . . ∈ B(x, r) ∩ S. Since xn ’s are distinct, so xK 6= x or xK+1 6= x, which implies
x ∈ S0 .
Conversely, if x ∈ S 0 , then we will get x1, x2, x3, . . . by induction. As cluster point, there is x1 ∈
B(x, 1) ∩ S and x1 6= x. Suppose distinct x1 , . . ., xk ∈ S are found with all xi 6= x. Choose a positive rk+1 <
d(x, x1), . . . , d(x, xk) and rk+1 < 1/(k + 1). By the definition of cluster point, there is xk+1 ∈ B(x, rk+1) ∩ S
and xk+1 6= x. Now d(x, xk+1) < rk+1 < d(x, xi) for i = 1, . . . , k implies x1, . . . , xk, xk+1 are distinct. For
k > 1, we have 0 < d(x, xk) < rk < 1/k, which implies lim xn = x by the sandwich theorem.
n→∞

(2) If there are x1, x2, x3, . . . ∈ S with lim xn = x, then either x ∈ S (so x ∈ S) or x 6∈ S. In the latter case,
n→∞
for every r > 0, there is xK ∈ B(x, r) ∩ S, hence xK 6= x. This leads to x ∈ S 0 . So, in both cases, x ∈ S.
Conversely, if x ∈ S, then either x ∈ S (in which case we take all xn = x) or x ∈ S 0 (in which case we
apply (1)). Both cases lead to the desired conclusion.

Density Theorem. S is dense in M iff for every x ∈ M and r > 0, B(x, r) contains some point of S.
Proof. If for every x ∈ M and r > 0, B(x, r) contains some point of S, then taking r = 1, 1/2, 1/3, . . ., we
get x1, x2, x3, . . . ∈ S satisfying d(xn, x) < 1/n. By the sandwich theorem, lim xn = x. By the sequential
n→∞
closure theorem, S = M.
If S is dense in M, then for every x ∈ M, there is a sequence {xn} in S converges to x by the
sequential closure theorem. For every r > 0, by the definition of limit, there is a positive integer K such
that xK , xK+1, xK+2, . . . ∈ B(x, r). In particular, B(x, r) contains xK ∈ S.

Remark. Note the condition “every B(x, r) contains a point of S” is equivalent to M \ S does not contain
any B(x, r). So we also have S is dense in M iff (M \ S)◦ = ∅.

6
Exercises. Let M be a metric space and S ⊆ M.
(1) Prove that any union of open sets in M is open in M. Prove that finite intersection of open sets in M is
open in M. Prove that any intersection of closed sets in M is closed in M. Prove that finite union of closed
sets in M is closed in M. (The collection of all open sets in M is called the topology of M.)
(2) Prove that S ◦ is the union of all open sets in M contained in S (so that S ◦ is the largest open set in
M contained in S). Prove that S is the intersection of all closed sets in M containing S (so that S is the
smallest closed set in M containing S).
(3) Prove that ∂S = S \ S ◦ is closed in M. (So S ◦ , ∂S form a partition of S.)
(4) Let S ∗ be the set of all isolated points of S. Prove that S 0 = S \ S ∗ is closed in M. (So S 0 , S ∗ form a
partition of S.)

§3. Separability and Countability of Bases

Definitions. Let M be a metric space.

(1) M is separable iff M has a countable dense subset. (For example, R with the usual metric is separable
because Q is a countable dense subset of R.)
(2) A collection S = {Oα : α ∈ A} of open sets in M is a basis in M iff every open set in M is a union of
some of the Oα’s. (For example, {B(x, r) : x ∈ M, r > 0} is a basis in M by the structure theorem for open
sets.) Later we will write {Oα} as a shorthand for {Oα : α ∈ A}.
(3) M is second countable (or M satisfies the second axiom of countability) iff M has a countable basis.

Theorem. S is a basis in M iff for every x in an open set O, there is Oαx ∈ S such that x ∈ Oαx ⊆ O.
Proof. If for every x in an open set O, there is Oαx ∈ S such that x ∈ Oαx ⊆ O, then {x} ⊆ Oαx ⊆ [Link]
union over all x ∈ O, we have O = ∪ {x} ⊆ ∪ Oαx ⊆ ∪ O = O. So O = ∪ Oαx . Therefore, S is a basis.
x∈O x∈O x∈O x∈O

Conversely, if S is a basis, then every open set O is the union of some of the Oα’s in S by the definition
of basis. Hence, for every x ∈ O, we have x ∈ Oαx ⊆ O for some Oαx ∈ S.

Theorem. M is separable iff M is second countable, i.e. M has a countable basis.

Proof. If M has a countable basis S = {On : n = 1, 2, 3, . . .}, then choosing a xn ∈ On , we claim A =
{x1, x2, x3, . . .} is a countable dense subset of M. To see this, let x ∈ M and k = 1, 2, 3, . . .. By the last
theorem, x ∈ B(x, 1/k) implies there is Onk ∈ S such that x ∈ Onk ⊆ B(x, 1/k). Also, xnk ∈ A ∩ Onk ⊆
B(x, 1/k). Then d(xnk , x) < 1/k so that lim xnk = x. By the sequential closure theorem, A = M.
k→∞

Conversely, if M is separable, then M has a countable dense subset A = {x1, x2, x3, . . .}. We claim
S = {B(xn , r) : xn ∈ A, r ∈ Q+ } is a basis in M, which is clearly countable. To prove the claim, let x be in
an open set O. Since x is an interior point of O, there is rx > 0 such that B(x, rx) ⊆ O. Since A is dense,
there is xn ∈ B(x, rx/2). Then d(xn, x) < rx /2 < rx − d(xn, x). By the density of Q, there is r ∈ Q such
that rx /2 < r < rx − d(xn, x). Then B(xn , r) ∈ S and x ∈ B(xn , rx/2) ⊆ B(xn , r) ⊆ B(x, rx) ⊆ O, where
the second inclusion is by the ball-in-ball lemma. Therefore, by the last theorem, the claim is true.

Lindelöf ’s Covering Theorem. Let M be second countable with basis S = {O1, O2, O3, . . .} and A ⊆ M.
∞
If A ⊆ ∪ Uα for some open Uα ’s, then there are countably many Uα ’s, say Uαk , such that A ⊆ ∪ Uαk .
α k=1

Proof. For every x ∈ A, there is Uα,x such that x ∈ Uα,x. Since S is a basis, there exists β(x) ∈ N
such that x ∈ Oβ(x) ⊆ Uα,x . Since S is countable, the set T = {Oβ(x) : x ∈ A} (being a subset of S)
is also countable, say T = {Oβ(x) : x ∈ A} = {On1 , On2 , On3 , . . .}. (If T = {On1 , . . . , Onj } is a finite
set, then for k > j, let Onk = On1 .) Since Onk ∈ T , Onk = Oβ(xk ) ⊆ Uα,xk for some xk ∈ A. Then
∞ ∞ ∞
A = ∪ {x} ⊆ ∪ Oβ(x) = ∪ Onk ⊆ ∪ Uα,xk . Let Uαk = Uα,xk . Then A ⊆ ∪ Uαk .
x∈A x∈A k=1 k=1 k=1

7
§4. Continuity and Equivalent Metrics

Definitions. Let M1 and M2 be metric spaces with metrics d1 and d2 respectively. Let x0 ∈ M1 , y0 ∈ M2
and f : M1 → M2.
(1) We define lim f(x) = y0 iff lim d2(f(x), y0 ) = 0 or more precisely, for every ε > 0, there exists δ > 0
x→x0 x→x0
such that 0 < d1(x, x0) < δ ⇒ d2(f(x), y0 ) < ε.
(2) We define f is continuous at x0 iff lim f(x) = f(x0 ) or more precisely, for every ε > 0, there exists δ > 0
x→x0
such that d1 (x, x0) < δ ⇒ d2 (f(x), f(x0 )) < ε. Now the implication “d1(x, x0) < δ ⇒ d2(f(x),

f(x0 )) < ε”

is
the same as “x ∈ B(x0 , δ) ⇒ f(x) ∈ B(f(x0 ), ε)”, which can be simplified to “f B(x0 , δ) ⊆ B f(x0 ), ε ”.
(3) We define f is continuous iff f is continuous at every x0 ∈ M1 .

Sequential Continuity Theorem. f : M1 → M2 is continuous at x0 iff (*) for every sequence {xn}
converges to x0 in M1 , lim f(xn ) = f lim xn = f(x0 ).
n→∞ n→∞

Proof. If f : M1 → M2 is continuous at x0, then for every ε > 0, there is δ > 0 such that for all x ∈ M1,
d1(x, x0) < δ ⇒ d2(f(x), f(x0 )) < ε. If lim xn = x0 , then there is K such that n ≥ K ⇒ d1(xn , x0) < δ,
n→∞
which implies d2(f(xn ), f(x0 )) < ε. So lim f(xn ) = f(x0 ) and (*) is true.
n→∞
Conversely, let (*) be true. Assume f is discontinuous at x0. Then there is ε > 0 such that for every
δ > 0, there is x ∈ M1 with d1(x, x0) < δ and d2 (f(x), f(x0 )) ≥ ε. Now, setting δ = n1 , there is xn ∈ M1
with d1(xn , x0) < δ = n1 and d2(f(xn ), f(x0 )) ≥ ε. Then xn → x0. By (*), lim f(xn ) = f(x0 ). Then
n→∞
0 = lim d2(f(xn ), f(x0 )) ≥ ε, a contradiction.
n→∞

Topological Continuity Theorem. A function f : M1 → M2 is continuous iff for every open set O in
M2 , f −1 (O) = {x ∈ M1 : f(x) ∈ O} is open in M1 iff for every closed set C in M2 , f −1 (C) is closed in M1.
Proof. Suppose f is continuous. Let O be open in M2 and x0 ∈ f −1 (O). Then f(x0 ) ∈ O = O◦ . So there
exists B(f(x0 ), ε) ⊆ O for some ε > 0. Since f is continuous at x0 , for this ε > 0, there exists δ > 0 such
that f(B(x0 , δ)) ⊆ B(f(x0 ), ε) ⊆ O. Then B(x0 , δ) ⊆ f −1 (O). So f −1 (O) is open in M1 .
Conversely, let x0 ∈ M1 . For every ε > 0, O = B(f(x0 ), ε) is open in M2 and f(x0 ) ∈ O. Then
f −1 (O) is open in M1 and x0 ∈ f −1 (O). So there exists δ > 0 such that B(x0 , δ) ⊆ f −1 (O). This implies
f(B(x0 , δ)) ⊆ O = B(f(x0 ), ε). So f is continuous at x0. Finally, the second iff follows easily from applying
M1 \ f −1 (S) = f −1 (M2 ) \ f −1 (S) = f −1 (M1 \ S) in case S is closed and in case S is open.

Exercise. Prove that a function f : M1 → M2 is continuous at x0 if and only if for every open set O2 in M2
containing f(x0 ), there exists an open set O1 in M1 containing x0 such that f(O1 ) ⊆ O2.

Definitions. (1) Let f : M1 → M2 be bijective. We say f is a homeomorphism between M1 and M2 iff f

and its inverse function are both continuous. In that case, we say M1 and M2 are homeomorphic.
(2) We say metrics d1 and d2 on M are equivalent (or topological equivalent) iff the identity function i :
(M, d1) → (M, d2) defined by i(x) = x is a homeomorphism.
If f : M1 → M2 is a homeomorphism with inverse g : M2 → M1 , then there is a bijection between open
sets in M1 and M2 . Namely let open set O = f −1 (f(O)) in M1 correspond to open set f(O) = g−1 (O) in M2.
Then a basis in M1 corresponds to a basis in M2 . Also, M1 is second countable iff M2 is second countable.
We have xn → x in M1 iff f(xn ) → f(x) in M2 . A closed set C in M1 corresponds to a closed set f(C) in
M2 . A dense set S in M1 corresponds to a dense set f(S) in M2. M1 is separable iff M2 is separable.
The topology of a metric space (M, d) is the set of all its open sets. Let M be a set with two metrics d1
and d2. Now we raise the
Question: When will d1 and d2 have the same topology on M ? I.e. when will the collections of open sets
for d1 and d2 be the same?

8
 Assume (M, d1 ) and (M, d2) have the same collection of open sets. Then the identity function i :
(M, d1) → (M, d2 ) assigning i(x) = x would be continuous by the topological continuity theorem since
i−1 (O) = O for all open sets O in M2 . Similarly, the inverse function of i would also be continuous. That
means i is a homeomorphism and d1, d2 are equivalent metrics on M.
Next we will look at a common sufficient condition for two metrics on the same set to be equivalent.

Theorem. Let M have two metrics d1 and d2. If there are c1 , c2 > 0 such that for every x, y ∈ M, we have
c1 d1(x, y) ≤ d2 (x, y) ≤ c2 d1(x, y), then d1 and d2 are equivalent. The converse is false in general.
Proof. Observe that applying the sandwich theorem to the given inequalities, we have d1(xn, x) → 0 iff
d2(xn , x) = d2(i(xn ), i(x)) → 0. Therefore, the identity function i : (M, d1) → (M, d2) is continuous by the
sequential continuity theorem. Similarly, the inverse function of i is also continuous.

Remarks. Many people defined metric equivalence by the condition c1 d1(x, y) ≤ d2(x, y) ≤ c2 d1(x, y). Below
we will show by an example that there are two metrics d and dS 2 on C which are (topological, but not metric)
equivalent, i.e. they yield the same topology, but there is no c > 0 such that cd(z, w) ≤ dS 2 (z, w) for all
z, w ∈ C.

Examples. (1) In Rn or Cn, all dp (1 ≤ p ≤ ∞) metrics are metric equivalent. To see this, first consider
1 1
1 < p, q < ∞ with + = 1. By Hölder’s inequality,
p q

X
n X
n X
n 1/p X
n 1/q X
n 1/p
|ai| = |ai · 1| ≤ |ai |p 1q = n1/q |ai|p .
i=1 i=1 i=1 i=1 i=1

X
n 1/p n
X
Summing |ai |p ≤ (|a1| + · · ·+ |an |)p from i = 1 to n and taking p-th root, we get |ai|p ≤ n1/p |ai|.
i=1 i=1
So
X
n X
n

1/p X
n
n−1/q |ai| ≤ |ai|p ≤ n1/p |ai |.
i=1 i=1 i=1

X
n X
n
As p → ∞, q → 1 and the inequality becomes n−1 |ai| ≤ max{|ai| : i = 1, 2, . . ., n} ≤ |ai |. Now for
i=1 i=1
x = (x1, . . . , xn), y = (y1 , . . . , yn ) ∈ Cn , let ai = xi − yi and get n−1/q d1 (x, y) ≤ dp (x, y) ≤ n1/p d1(x, y),
i.e. d1 and dp are metric equivalent. Composing inequalities, we can see any two dp metrics are metric
equivalent. By the last theorem, they all yield the same collection of open sets (i.e. topology) on Rn or Cn.
(2) In C = R2 , the usual metric is d(z, w) = |z − w|. Define

2|z − w|
dS 2 (z, w) = p .
(1 + |z|2)(1 + |w|2)

This is called the chordal metric on C. This metric comes from the stereographic projection of the complex
plane. To describe this projection, identify every complex number z = x+iy with (x, y, 0) ∈ R3 . Consider the
unit sphere S 2 = {(x1, x2, x3) ∈ R3 : x21 + x22 + x23 = 1} in R3 and the north pole N = (0, 0, 1). Let Z be the
intersection of S 2 \{N } with the line joining N to z = (x, y, 0). Z is the stereographic projection of z ∈ C into
the unit sphere S 2 . By a calculation (see John Conway’s book, Functions of One Complex Variable, second
edition, pp. 8-9), we can show dS 2 (z, w) is the length of the chord joining the stereographic projections of z
and w on S 2 . From this, we can see dS 2 is a metric on C.
Clearly, dS 2 (z, w) ≤ 2d(z, w). If there is c > 0 such that for all z, w ∈ C, cd(z, w) ≤ dS 2 (z, w), then

2 4
c≤ p =⇒ (1 + |z|2)(1 + |w|2) ≤ .
(1 + |z| )(1 + |w|2)
2 c2

9
However, (1 + |z|2)(1 + |w|2) is unbounded. So no such c can exist. Yet d and dS 2 are still equivalent! This
will be explained below.
Since dS 2 (z, w) ≤ 2d(z, w), for every ε > 0, we can take δ = ε/2 to see that the identity function
i : (C, d) → (C, dS 2 ) is continuous. Next to see i−1 : (C, dS 2 ) → (C, d) is continuous, we
Claim: dS 2 (zn , z) → 0 implies {|zn| : n ∈ N} is bounded.
Proof. Assume the |zn |’s are unbounded. Then there are |znk | → ∞. Then

2|znk − z| 1/|znk | 2|1 − z/znk | 2

dS 2 (znk , z) = p · = p →p 6= 0,
(1 + |znk | )(1 + |z| ) 1/|znk |
2 2 2 2
(1/|znk | + 1)(1 + |z| ) 1 + |z|2

contradicting dS 2 (znk , z) → 0. This finishes the proof of the claim.

Now suppose dS 2 (zn , z) → 0. By the claim, there is k > 0 such that for all n, |zn| ≤ k. Then

2|zn − z| 2d(zn , z)
dS 2 (zn , z) = p ≥ p =⇒ d(zn , z) → 0.
2 2
(1 + |zn| )(1 + |z| ) (1 + k2)(1 + |z|2)

By the sequential continuity theorem, this implies i−1 is continuous. Thus, we conclude d and dS 2 are
equivalent metrics on C, but there does not exists c > 0 such that cd(z, w) ≤ dS 2 (z, w).

Definitions. (1) f : (M1 , d1) → (M2 , d2) is uniformly continuous iff for every ε > 0, there exists a δ > 0
such that for all x, y ∈ M1, d1(x, y) < δ implies d2 (f(x), f(y)) < ε.
(2) We say two metrics d1 and d2 on a set M are uniformly equivalent iff the identity function i : (M, d1) →
(M, d2) and its inverse function are uniformly continuous.

Remarks. (1) The condition c1 d1(x, y) ≤ d2(x, y) ≤ c2 d1(x, y) implies d1 and d2 are uniformly equivalent.


The converse is false. Let d be the usual metric on R. The metric d1(x, y) = d(x, y)/ 1 + d(x, y) on R is
uniformly equivalent to d, but there is no c such that |y| = d(0, y) ≤ cd1(0, y) for all y ∈ R.
(2) Clearly, uniform equivalence implies equivalence. The converse is false. Above we showed d and dS 2 are
equivalent. We may ask if d and dS 2 on C are uniformly equivalent. For this, we need to ask if the inverse
function of i is uniformly continuous. The answer is negative since we can check lim dS 2 (n, n + 1) = 0, but
n→∞
d(n, n + 1) = 1. This means for ε = 1, there is no δ > 0 such that dS 2 (n, n + 1) < δ =⇒ d(n, n + 1) < 1.

§5. Completeness and Contractions

Definitions. (1) A sequence {xn} is a Cauchy sequence iff for all ε > 0, there exists a positive integer K
such that m, n ≥ K ⇒ d(xm , xn) < ε.
(2) A metric space M is complete iff every Cauchy sequence in M converges in M.

Examples. (1) R with the usual metric is complete because the famous Cauchy theorem asserts that a
sequence in R converges in R iff it is a Cauchy sequence. The usual proof of the Cauchy theorem is consisted
of four steps:
Step 1. If {xn} converges, then it is a Cauchy sequence.
Step 2. If {xn} is a Cauchy sequence, then it is bounded.
Step 3. (Bolzano-Weierstrass Theorem) Every bounded sequence has a convergent subsequence.
Step 4. If {xn} is a Cauchy sequence and some subsequence {xnk} converges to x, then {xn} converges to x.
We remark that steps 1, 2 and 4 are true for all metric spaces with the same proofs as in the Cauchy theorem.
(2) Every nonempty set M with the discrete metric is complete. This follows by taking ε = 1/2 to see that
every Cauchy sequence {xn } is eventually constant, i.e. there is K such that xK = xK+1 = xK+2 = · · · .

10
(3) Q with the metric d(x, y) =
√|x−y| is not complete. For example, the sequence 1, 1.4, 1.41, 1.414, 1.4142, . . .
(or more precisely, xn = [10n 2]/10n) is a Cauchy sequence in Q, but it does not converge in Q.
(4) R or C with the dS 2 metric is not complete. For example, xn = n is a Cauchy sequence since for ε > 0,
1
2|m − n| mn 2| n1 − m
1
| 1 1 2 2
dS 2 (m, n) = p · 1 =q ≤2 − < + <ε
(1 + m2 )(1 + n2) mn 1 1
( n2 + 1)( m2 + 1) n m n m

if m, n ≥ K = d4/εe. However, the sequence does not converge. Assume it converges to x. Then

2|n − x| 1/n 2|1 − nx | 2

dS 2 (n, x) = p · = q →√ 6= 0,
(1 + n2)(1 + x2) 1/n 1 2
( n2 + 1)(1 + x ) 1 + x2

which is a contradiction. Although (R, d), (R, dS2 ) are homeomorphic, they have different Cauchy sequences.
Note i−1 : (R, dS 2 ) → (R, d) is continuous, {n} is Cauchy in (R, dS 2 ), but {i−1 (n)} is not Cauchy in (R, d).

Next we will look at some basic facts about completeness.

Lemma. If f : (M1 , d1) → (M2 , d2) is uniform continuous and {xn } is a Cauchy sequence in M1, then
{f(xn )} is a Cauchy sequence in M2 .
Proof. For every ε > 0, since f is uniformly continuous, there exists δ > 0 such that d1(x, y) < δ ⇒
d2(f(x), f(y)) < ε. Since {xn} is a Cauchy sequence, there exists K such that m, n ≥ K ⇒ d1(xm , xn) < δ.
Then d2 (f(xm ), f(xn )) < ε.

Theorem. If f : (M1 , d1) → (M2 , d2) is a uniform homeomorphism (i.e. f is bijective and f, f −1 are
uniformly continuous), then (M1 , d1) is complete iff (M2 , d2) is complete. In particular, if d1 and d2 are
uniformly equivalent metrics on M, then (M, d1) is complete iff (M, d2) is complete.
Proof. Suppose (M2 , d2) is complete. Let {xn} be a Cauchy sequence in (M1 , d1). Since f : (M1 , d1) →
(M2 , d2) is uniformly continuous, by the lemma, {f(xn )} is a Cauchy sequence in (M2 , d2). By the complete-
ness of (M2 , d2), {f(xn )} converges to some y in (M2 , d2). Since f −1 : (M2 , d2) → (M1 , d1) is continuous, by
the sequential continuity theorem, {xn} converges to f −1 (y) in (M1 , d1). The other direction is similar.

Remark. Since (R, d) is complete and (R, dS 2 ) is not complete, by the last theorem, we see that d and dS 2
are not uniformly equivalent.

Theorem. (1) If M is complete and A is closed in M, then A is complete with the induced metric.
(2) If A ⊆ M and A is complete, then A is closed in M (even if M may not be complete).
(3) Let M be complete and both {xn} and {yn } are Cauchy sequences in M. Then lim xn = lim yn iff
n→∞ n→∞
lim d(xn, yn) = 0.
n→∞

Proof. (1) Let {xn} be a Cauchy sequence in A, then it is a Cauchy sequence in M. Since M is complete,
{xn} converges to some x in M. By the sequential closure theorem and the fact A is closed in M, x ∈ A = A.
Therefore, A is complete.
(2) Let x ∈ A. By the sequential closure theorem, there is a sequence {xn} in A converges to x. Then {xn}
is a Cauchy sequence in A. Since A is complete, x ∈ A. Therefore, A = A, i.e. A is closed in M.
(3) Since M is complete and {xn} and {yn } are Cauchy sequences, there exist x, y ∈ M such that d(xn, x) → 0
and d(yn , y) → 0. Using d(x, y) ≤ d(x, xn) + d(xn , yn) + d(yn , y), d(xn, yn) ≤ d(xn, x) + d(x, y) + d(y, yn ) and
the sandwich theorem, we get x = y (i.e. d(x, y) = 0) iff d(xn, yn ) → 0.

Exercise. Prove the Nested Ball Theorem: M is complete iff for every Kn = {x ∈ M : d(xn, x) ≤ rn}
∞
satisfying K1 ⊇ K2 ⊇ K3 ⊇ · · · and rn converges to 0, ∩ Kn is always a single point in M.
n=1

11
 We now come to two powerful theorems. The first one is the Banach Contractive Mapping Theorem.

Definitions. (1) f : (M1 , d1) → (M2 , d2) is a contractive

mapping (or a contraction) iff there exists 0 < k < 1
such that for all x, y ∈ M1, we have d2 f(x), f(y) ≤ kd1 (x, y).
(2) f : M → M has a fixed point iff there exists x ∈ M such that f(x) = x.

Banach Contractive Mapping Theorem. If M is complete and f : M → M is a contraction, then f

has a unique fixed point.
There are many applications of this theorem. Here are just a few: the proof of the inverse function
theorem for Rn as well as the version for infinite dimensional spaces, the existence and uniqueness of of
local solutions to some differential equations (for example, dy/dx = f(x, y), y(a) = y0 ) and math olympiad
problems as well as PhD qualifying examinations (for example, every map of USA placed flat inside USA
has a unique point directly over the same location it represents).

Remarks. In the theorem, the contractive condition cannot be weakened to d(f(x), f(y)) ≤ d(x, y) or
d(f(x), f(y)) < d(x, y) for x 6= y. For example, let M = R with the usual metric and f(x) = x + 1,
then d(f(x), f(y)) = d(x, y), but f has no fixed point. Again, let M = R with the usual metric and
f(x) = ln(1 + ex ), then 0 < |f 0 (c)| = ec /(1 + ec ) < 1. By the mean value theorem, d(f(x), f(y)) < d(x, y)
for x 6= y, but f has no fixed point.
Proof of the Banach Contractive Mapping Theorem. We first prove the uniqueness of fixed point. If
f(x) = x and f(y) = y, then d(x, y) = d(f(x), f(y)) ≤ kd(x, y) and 0 < k < 1 imply d(x, y) = 0, i.e. x = y.
For the existence part, take any x0 ∈ M. If f(x0 ) = x0, then x0 is a fixed point. Otherwise define
xn = f(xn−1 ) for n = 1, 2, 3, . . . and get d(x0, x1) > 0. Observe that d(xn, xn+1) = d(f(xn−1), f(xn )) ≤
kd(xn−1, xn) ≤ · · · ≤ knd(x0, x1). So for m > n ≥ N,

kN
d(xn, xm ) ≤ d(xn, xn+1)+d(xn+1, xn+2)+· · ·+d(xm−1 , xm) ≤ (kn +kn+1 +· · ·)d(x0, x1) ≤ d(x0, x1) < ε
1−k


ln ε(1 − k)/d(x0, x1)
if we let N > . Thus, {xn} is a Cauchy sequence in M. By the completeness of M,
ln k
{xn} converges to some x ∈ M. Since 0 ≤ d(f(xn ), f(x)) ≤ kd(xn, x), by the sandwich theorem, we get
f(x) = lim f(xn ) = lim xn+1 = x.
n→∞ n→∞

Next, we present some terminologies for the second theorem, the famous Baire Category Theorem.

Definitions. Let M be a metric space and S ⊆ M.

(1) S is nowhere dense in M iff (S)◦ = ∅, i.e. S contains no open balls in M.
∞
(2) S is of the first category (or meager) in M iff S = ∪ Sn , where Sn ’s are nowhere dense in M.
n=1

(3) S is of the second category (or non-meager) in M iff S is not of the first category in M.
(4) S is comeager or residual in M iff M \ S is of the first category in M.
(5) M is a Baire space iff the intersection of every sequence of open dense sets in M is dense in M.

Baire Category Theorem. Let M be a complete metric space.

∞
(1) M is a Baire space, i.e. if U1 , U2, U3 , . . . are open dense sets in M, then ∩ Un is dense in M.
n=1

(2) If S is of the first category in M, then M \ S is dense in M (equivalently S ◦ = ∅ by the remark on p. 6).
In case S = M, the contrpositive statement asserts M complete implies M is of the second category in M.

Again, there are many applications of this theorem. Here are just a few: continuous, nowhere differ-
entiable functions on R are dense among continuous functions on R, existence of linear partial differential

12
equations with no local solution anywhere (due to Hans Lewy) and the proofs of major theorems of functional
analysis and topology.

Examples. (1) Let M be a metric space without any isolated point. So for every x ∈ M, {x} 6= B(x, r).
Since {x} = {x} by the sequential closure theorem, we get ({x})◦ = {x}◦ = ∅. Hence, every set with one
element in M is nowhere dense in M and so every countable set in M is of the first category in M.
(2) M = R with the usual metric is complete, hence R is of the second category in R by part (2) of the Baire
category theorem. Both Q and R \ Q are dense in R and both have empty interior. By example (1), Q is
of the first category in R. However, R \ Q is of the second category in R for otherwise R \ Q is of the first
category in R would imply R = (R \ Q) ∪ Q is a countable union of nowhere dense sets in R, contradicting
R is of the second category in R.
(3) Let M be a complete metric space and S ⊆ M. By part (2) of the Baire category theorem, if S contains
some nonempty open ball, then S is of the second category in M. However, the converse is false as R \ Q is
of the second category in R, but the interior of R \ Q is empty. Also, first category sets may not be nowhere
dense. For example, Q is of first category in R, but (Q)◦ = R◦ = R.
(4) Let M = {(x, 0) : x ∈ R} with metric defined by d((a, 0), (b, 0)) = |a − b|. As R is complete, so is M.
Hence M is of the second category in M. However, M is nowhere dense (hence of the first category) in
R2 with the usual metric since M is complete, hence closed in R2 and M = M has empty interior in R2.
Therefore, being first or second category are relative concepts.

Proof of the Baire Category Theorem. (1) Take any B(x0 , r0) in M. Since U1 is dense, there exists
x1 ∈ U1 ∩ B(x0 , r0). Since U1 ∩ B(x0 , r0) is open in M, U1 ∩ B(x0 , r0) ⊇ B(x1 , 2r1) ⊇ B(x1 , r1) for some
r1 < 1 by shrinking r1. Inductively repeating this, we get a sequence B(xn , rn) such that Un ∩B(xn−1, rn−1) ⊇
B(xn , rn) for some rn < 1/n. Note B(xn , rn) contains xn, xn+1, xn+2, . . ..
We claim {xn} is a Cauchy sequence. To see this, for every ε > 0, let K > 2/ε, then B(xK , rK ) contains
xK , xK+1, xK+2, . . . . So

2
m, n ≥ K ⇒ xm , xn ∈ B(xK , rK ) ⇒ d(xn, xm ) ≤ d(xn, xK ) + d(xK , xm) < 2rK < < ε.
K

So {xn} converges to some w ∈ M. Now for n = 1, 2, 3, . . ., we get w =lim xn+j ∈ B(xn , rn) ⊆ Un ∩ B(x0 , r0).
j→∞
∞ ∞
Hence, w ∈ ∩ Un is in B(x0 , r0). Therefore, by the density theorem, ∩ Un is dense in M.
n=1 n=1
∞
(2) Suppose S = ∪ Sn , where Sn ’s are nowhere dense in M. Then Un = M \ Sn is open and dense in M
n=1
∞ ∞ ∞
because (M \ Un ) = (Sn )◦ = ∅. By part (1), ∩ Un = ∩ (M \ Sn ) = M \ ∪ Sn (⊆ M \ S) is dense in M.
◦
n=1 n=1 n=1
Then M \ S is also dense in M.

§6. Compactness

Recall the Bolzano-Weierstrass theorem asserts that every sequence in [a, b] has a subsequence converging
to a limit in [a, b]. This theorem is very important. It is used in the proofs of the Cauchy theorem, the extreme
value theorem, the uniform continuity theorem and many more. This leads us to ask the following
Question: Which sets in a metric space have the property that every sequence in the set has a subsequence
converging to a limit in that set?

Definitions. Let M be a metric space and S ⊆ M.

(1) S is sequentially compact iff every sequence in S has a subsequence converging to a limit in S.
(2) A collection U of subsets of M covers (or is a cover of) S in M iff S is a subset of the union of all sets
in U. If U0 ⊆ U and U0 also covers S in M, then U0 is said to be a subcover of S.

13
(3) A cover U of S is a finite cover iff U is a finite set. A cover U of S is a countable cover iff U is a countable
set. A cover U of S is an open cover iff every set in U is an open set.
(4) S is compact in M iff every open cover of S in M has a finite subcover.
n
(5) S is totally bounded in M iff for every ε > 0, there are x1, x2, · · · , xn ∈ M such that S ⊆ ∪ B(xi , ε).
i=1

Examples. (1) Finite sets are compact (for each x, find an open set in the open cover containing x).
(2) If a sequence x1, x2, x3, . . . converges to a in a metric space M, then K = {a, x1, x2, x3, . . .} is compact
(since in any open cover {Uα } of K, some open Uα0 contains a, hence contains some B(a, εa ), which contains
xN , xN +1 , xN +2 , . . . for some N. The remaining x1, . . . , xN −1 are in finitely many other Uα ’s).
(3) If S is a closed subset of a compact metric space M, then S is compact (since for any open cover of S, we
can add the open set M \ S to the cover, then it will be an open cover of M, hence it has a finite subcover,
which we can delete M \ S from the subcover (if it is in the subcover) to get a finite subcover of S).

Remarks. (1) Total boundedness implies boundedness since B(x1 , 1)∪B(x2, 1)∪· · ·∪B(xn, 1) ⊆ B(x1 , c+1),
where c = max{d(x1, x2), . . . , d(x1, xn)}.
In general, boundedness does not imply total boundedness. For example, let M be an infinite set with the
discrete metric d. Then M ⊆ B(x, 2), hence bounded. However, for ε < 1, B(x1 , ε)∪B(x2, ε)∪· · ·∪B(xn, ε) =
{x1, x2, . . . , xn} never contains M, hence M is not totally bounded.
(2) In Rn , boundedness implies totally boundedness since every B(a, r) is a subset of an open cube with side
h r√n i n
2r, which can be divided into + 1 subcubes and each subcube can be covered by an open ball of
ε
radius ε. So in Rn, a set is bounded iff it is totally bounded.

Theorem. Let S ⊆ M1 ⊆ M2 , where M2 has a metric d and M1 has the induced metric from d.
(1) S is open in M1 iff there is an open set O in M2 such that S = O ∩ M1 .
(2) S is closed in M1 iff there is a closed set C in M2 such that S = C ∩ M1 .
Proof. (1) For i = 1, 2, let Bi (x, r) = {y ∈ Mi : d(x, y) < r}, then B1 (x, r) = B

2 (x, r)∩M1. By the structure
theorem for open sets, S is open in M1 iff S = ∪ B1 (xα, rα) = ∪ B2 (xα , rα) ∩ M1 . Let O = ∪ B2 (xα, rα),
α α α
then O is open in M2 and S = O ∩ M1 .
(2) S is closed in M1 iff M1 \ S is open in M1 iff by (1), M1 \ S = O ∩ M1 for some open set O in M2 iff
S = M1 \ O = (M2 \ O) ∩ M1 , where C = M2 \ O is closed in M2 .

Theorem. Let M be a metric space and S ⊆ M.

(1) S is compact in M iff S is compact in S.
(2) S is totally bounded in M iff S is totally bounded in S.
Proof. (1) By the last theorem, if {OM,α : α ∈ A} is any open cover of S in M, then {OM,α ∩ S : α ∈ A} is
also an open cover of S in S. Conversely, every open cover {OS,α : α ∈ A} of S in S can be extended to an
open cover {OM,α : α ∈ A} of S in M. The converse follows.
(2) The if direction is clear. For the only if direction, for every ε > 0, by total boundedness in M,
N

S ⊆ ∪ B xi , ε/2 for some x1 , x2, . . . , xN in M. We may assume every one of of these N balls intersects
i=1

S, otherwise remove it. Then there exists si ∈ S in each B xi , ε/2 . So d(xi, si ) < ε/2. By the ball-in-ball


N

lemma, since ε > ε/2 + d(xi , si), we get B xi, ε/2 ⊆ B si , ε . Then S ⊆ ∪ B si , ε .
i=1

Remarks. The first theorem above tells us that whether a set is open depends on the metric space containing
it. For example, (0, 1] is open in [−1, 1], but not open in R. So being open is a relative concept.

14
 The second theorem tells us that whether a set is compact or totally bounded does not depend on the
metric space containing it. So being compact or totally bounded is an absolute concept. Hence, often we say
a set is compact or totally bounded without mentioning which metric space contains it.
The main point of part (2) of the last theorem is the centers of the balls can be chosen to be in S.
Metric Compactness Theorem. Let M be a metric space and S ⊆ M. The following are equivalent.
(1) S is compact
(2) S is sequentially compact
(3) S is complete and totally bounded.
∞

Proof. (1)⇒(2) Let {xn} be a sequence in S. Let Kn = {xn, xn+1, xn+2, . . .}. Assume ∩ Kn ∩ S = ∅.
n=1
∞ ∞
Then S ⊆ M \ ∩ Kn = ∪ (M \ Kn ). Since M \ Kn is open in M and S is compact in M, for some
n=1 n=1
m m
m

m ∈ N, we have S ⊆ ∪ (M \ Kn ) = M \ ∩ Kn . This contradicts xm ∈ S ∩ ∩ Kn . So, there exists
n=1 n=1 n=1
∞

x ∈ ∩ Kn ∩ S.
n=1

Next we construct a subsequence {xnk } converging to x ∈ S. Since x ∈ K1, there exists xn1 ∈ B(x, 1).
Since x ∈ Kn1 +1 = {xn1 +1 , xn2+1 , xn3+1 , . . .}, there exist n2 ≥ n1 + 1 > n1 and xn2 ∈ B(x, 1/2). Repeating
inductively, we get a subsequence {xnk } of {xn } such that xnk ∈ B(x, 1/k). Then d(xnk , x) < 1/k → 0, so
xnk → x ∈ S. Therefore, S is sequentially compact.
(2)⇒(3) (Completeness) Let {xn} be a Cauchy sequence in S. Since S is sequentially compact, {xn} has
a subsequence {xnk } converges to some limit x ∈ S. By step 4 of the proof of the Cauchy theorem, {xn}
converges to x ∈ S.
(Total Boundedness) Assume S is not totally bounded. Then there exists ε > 0 such that S cannot be
covered by finitely many open ball of radius ε. Let x1 ∈ S. Then there exist x2 ∈ S \ B(x1 , ε), x3 ∈
S \ (B(x1 , ε) ∪ B(x2 , ε)), · · · , xn ∈ S \ (B(x1 , ε) ∪ · · · ∪ B(xn−1 , ε)), · · · . Then for m 6= n, d(xm , xn) > ε.
Hence, all subsequences of {xn} cannot be Cauchy sequences, in particular cannot converge, a contradiction.
(3)⇒(1) Assume S has an open cover U = {Oα : α ∈ A} without any finite subcover. Let S0 = S. Since
S0 is totally bounded, S0 can be covered by finitely many B(xβ , 1/2)’s. This implies S0 is the union of the
S0 ∩ B(xβ , 1/2)’s. Not all of these S0 ∩ B(xβ , 1/2) can be covered by finitely many Oα’s, otherwise S0 can
be covered by finitely many Oα’s, a contradiction. So say S1 = S0 ∩ B(x1, 1/2) cannot be covered by finitely
many Oα’s.
Inductively, for every n = 1, 2, 3, . . ., there exists Sn = Sn−1 ∩ B(xn , 1/2n) = · · · = S0 ∩ B(x1 , 1/2) ∩
· · · ∩ B(xn , 1/2n), which cannot be covered by finitely many Oα’s. In particular, Sn 6= ∅, say zn ∈ Sn .
We claim {xn} is a Cauchy sequence. First observe that zn+1 ∈ Sn+1 ⊆ B(xn , 1/2n) ∩ B(xn+1 , 1/2n+1)
implies d(xn, xn+1) ≤ d(xn, zn+1) + d(zn+1 , xn+1) < 1/2n + 1/2n+1 = 3/2n+1. For every ε > 0, let K > 3/ε.
Then m > n ≥ K implies
3 3 3 3 3
d(xn, xm ) ≤ d(xn, xn+1) + d(xn+1, xn+2) + · · · + d(xm−1 , xm) < n+1 + n+2 + · · · = n < ≤ < ε.
2 2 2 n K

Since S is complete, {xn } converges to some x ∈ S. Then x ∈ Oα for some α. So there exists B(x, ε) ⊆
Oα. Choose n large so that 1/2n < ε/2 and d(xn, x) < ε/2. If y ∈ B(xn , 1/2n), then y ∈ B(x, ε) since

1 ε
d(y, x) ≤ d(y, xn) + d(xn, x) < n
+ < ε.
2 2
Then Sn ⊆ B(xn , 1/2n) ⊆ B(x, ε) ⊆ Oα . Sn is covered by this Oα, contradiction. So S is compact in M.

Exercise. Prove that a set S is totally bounded in M iff every sequence in S has a Cauchy subsequence.
(Note the if direction followed from (2)⇒(3) of the proof above.)

15
Theorem. If S is a totally bounded metric space, then S is separable.
Nk 1

Proof. For every positive integer k, by total boundedness, S ⊆ ∪ B xk,i, for some xk,1, xk,2, . . . , xk,Nk
i=1 k
∞
in S. Let D = ∪ {xk,1, xk,2, . . . , xk,Nk }, then D is a countable subset of S.
k=1
Now we check D is dense in S. For every w ∈ S, there exists xk,iw such that w ∈ B(xk,iw , 1/k). So
d(xk,i, w) < 1/k → 0, i.e. w = lim xk,i ∈ D. Therefore, D = S.
k→∞

Corollary. If S is compact, then S is closed and separable (as it is complete and totally bounded).

Heine-Borel Theorem. Let S ⊆ Rn . S is compact if and only if S is closed and bounded in Rn.
Proof. In Rn, a set is closed iff it is complete; a set is bounded iff it is totally bounded. The result follows
from the metric compactness theorem.

Compact Image Theorem. If f : M1 → M2 is continuous and M1 is compact, then f(M1 ) is compact.

Proof. Suppose we have f(M1 ) ⊆ ∪ Oα , where the Oα ’s are open in M2 . Taking f −1 on both sides,


α∈A
M1 ⊆ f −1 f(M1 ) ⊆ f −1 ∪ Oα = ∪ f −1 (Oα ). By the topological continuity theorem, every f −1 (Oα )
α∈A α∈A
n
is open in M1 . By the compactness of M1 , there exists α1, . . . , αn such that M1 ⊆ ∪ f −1 (Oαi ). Then
 n  n n
i=1
f(M1 ) ⊆ f ∪ f −1 (Oαi ) = ∪ f(f −1 (Oαi )) ⊆ ∪ Oαi . Therefore, f(M1 ) is compact.
i=1 i=1 i=1

Extreme Value Theorem. If f : M1 → R is continuous and M1 is compact, then sup f(M1 ) and inf f(M1 )
are in f(M1 ), i.e. the range of f attains maximum and minimum values at some points of M1 .
Proof. By the compact image theorem, f(M1 ) is compact in R. By the Heine-Borel theorem, f(M1 ) is closed
and bounded in R. f(M1 ) bounded implies sup f(M1 ) and inf f(M1 ) are in R. Since f(M1 ) is closed in R,
by the supremum limit theorem, there is a sequence {xn} ∈ M1 such that

sup f(M1 ) = lim f(xn ) ∈ f(M1 ) = f(M1 ).

n→∞

Similarly, inf f(M1 ) ∈ f(M1 ).

Uniform Continuity Theorem. If f : (M1 , d1) → (M2 , d2) is continuous and M1 is compact, then f is
uniformly continuous.
Proof. For every ε > 0, since f is continuous at x ∈ M1 , there exists δx > 0 such that d1(x, y) < δx ⇒
d2(f(x), f(y)) < ε/2. Since M1 = ∪ {x} ⊆ ∪ B1 (x, δx/2), by compactness, there are t1 , · · ·, tn such
x∈M1 x∈M1
n
that M1 ⊆ ∪ B1 (ti , δti /2).
i=1

Let δ = min{δt1 /2, . . ., δtn /2}. Let x, y ∈ M1 and d1(x, y) < δ. Since x ∈ M1 , there exists B1 (ti , δti /2)
containing x. By the definitions of δ and δt , we have d1 (x, ti) < δti /2 < δti and

d1(x, y) < δ =⇒ d1(y, ti ) ≤ d1(y, x) + d1 (x, ti) < δ + δti /2 ≤ δti .

ε ε
So d2(f(x), f(y)) ≤ d2(f(x), f(ti )) + d2(f(ti ), f(y)) < 2
+ 2
= ε.

Definition. Let A and B be nonempty subsets of a metric space M. Define the distance between A and B
to be d(A, B) = inf{d(x, y) : x ∈ A, y ∈ B}. Clearly, d(A, B) = d(B, A). In case B = {x}, we may write
d(A, B) = d(A, x) = d(x, A).

Theorem. Let A and B be nonempty subsets of a metric space M.

16
(1) If f : M → [0, ∞) is defined by f(x) = d(x, B), then f satisfies |f(x) − f(y)| ≤ d(x, y) for all x, y ∈ M,
hence f is uniformly continuous.
(2) If A is compact, then there is a ∈ A such that d(A, B) = d(a, B). In case B = {x}, we have d(A, x) =
d(a, x) for some a ∈ A.
(3) If A and B are compact, then there exist a ∈ A and b ∈ B such that d(A, B) = d(a, b).
(4) Let B be compact. Then d(A, B) = 0 iff A ∩ B 6= ∅. In case B = {x}, d(A, x) = 0 iff x ∈ A.
Proof. (1) Observe that for all b ∈ B,

f(x) = d(x, B) ≤ d(x, b) ≤ d(x, y) + d(y, b) =⇒ f(x) ≤ d(x, y) + d(y, B) = d(x, y) + f(y).

So f(x) − f(y) ≤ d(x, y). Similarly, f(y) − f(x) ≤ d(y, x) = d(x, y). Hence, |f(x) − f(y)| ≤ d(x, y). For every
ε > 0, taking δ = ε, we see f is uniformly continuous.
(2) By (1), g : A → [0, ∞) defined by g(x) = d(x, B) is continuous. Since A is compact, by the extreme value
theorem, there exists a ∈ A such that d(A, B) = inf{d(x, B) : x ∈ A} = inf{g(x) : x ∈ A} = g(a) = d(a, B).
(3) By (2), d(A, B) = d(a, B) = d(a, b) for some a ∈ A and b ∈ B.
(4) If there is b ∈ A ∩ B, then by the sequential closure theorem, there is a sequence {an} in A converging
to b. So d(A, B) ≤ lim d(an, b) = 0, which implies d(A, B) = 0.
n→∞

Conversely, if d(A, B) = 0, then since B is compact, by (2) and the infimum limit theorem, there is
b ∈ B and a sequence {an } in A such that 0 = d(A, B) = d(A, b) = inf{d(a, b) : a ∈ A} = lim d(an, b). So
n→∞
b = lim an ∈ A by the sequential closure theorem. Therefore, b ∈ A ∩ B.
n→∞

§7. Connectedness

Definitions. Let S be a metric space.

(1) S is disconnected iff S = O1 ∪ O2 for some disjoint, nonempty open sets O1, O2 in S. (Then O1 = S \ O2
is also closed in S. Similarly, O2 is closed in S.)
(2) S is connected iff S is not disconnected (equivalently, ∅ and S are the only open and closed sets in S).

Example. R with usual metric is connected. (Assume R = O1 ∪O2, where O1 and O2 are disjoint, nonempty
open sets in R. Let x1 ∈ O1 and x2 ∈ O2 , say x1 < x2. Let w = sup{x ∈ [x1, x2] : x ∈ O1 }. If w ∈ O1, then
(w, x2] ⊆ O2 contradicts w is an interior point of O1. If w ∈ O2 , then for all ε > 0, (w − ε, w) ∩ O1 6= ∅
contradicts w is an interior point of O2.)

Connected Union Theorem. Let {Sα : α ∈ A} be a collection of connected subsets in a metric space M.
If ∩ Sα 6= ∅, then S = ∪ Sα is connected.
α∈A α∈A

Proof. Assume S = O1 ∪ O2, where O1 and O2 are disjoint, nonempty open
sets in S. Since there is
x ∈ ∩ Sα , so x ∈ O1 or x ∈ O2. Suppose x ∈ O1. Now ∅ =
6 O2 = O2 ∩ ∪ Sα = ∪ (O2 ∩ Sα ) implies
α∈A α∈A α∈A
O2 ∩ Sα 6= ∅ for some α ∈ A. Also, x ∈ O1 ∩ Sα . Then Sα = S ∩ Sα = (O1 ∩ Sα ) ∪ (O2 ∩ Sα ) with
O1 ∩ Sα , O2 ∩ Sα disjoint, nonempty open in Sα , contradicting Sα is connected. Therefore, S is connected.

Remarks. Let M be a metric space and x ∈ M. Since {x} is connected, so we can form the union of all
connected subsets Sα of M containing x. By the theorem, the union is connected. This set is the “largest”
connected subset of M containing x. We call it the component of x in M.

Connected Image Theorem. Let M1 and M2 be metric spaces. If f : M1 → M2 is continuous and M1 is

connected, then f(M1 ) is connected.

17
Proof. Assume f(M1 ) = O1 ∪ O2, where O1 and O2 are disjoint, nonempty open sets in f(M1 ). Say
f(a) ∈ O1 and f(b) ∈ O2. Then M1 = f −1 (O1 ) ∪ f −1 (O2). Also, f −1 (O1) ∩ f −1 (O2 ) = f −1 (O1 ∩ O2 ) = ∅,
a ∈ f −1 (O1 ), b ∈ f −1 (O2) and f −1 (O1), f −1 (O2 ) are open by the topological continuity theorem. This
contradicts M1 connected. So f(M1 ) is connected.

Connected Sandwich Theorem. Let M be a metric space and S ⊆ M. If S is connected and S ⊆ S0 ⊆ S,

then S0 is connected.
Proof. Assume S0 = O1 ∪O2 for some disjoint nonempty open sets O1, O2 in S0 . Then S = (O1 ∩S)∪(O2 ∩S)
and O1 ∩ S, O2 ∩ S are disjoint open sets in S. Since S is connected, one of O1 ∩ S or O2 ∩ S is empty, say
O1 ∩ S = ∅. Then S ⊆ S0 \ O1. Since S ⊆ S0 ⊆ S, S is dense in S0 . Hence, S0 \ O1 is dense in S0 . Since O1
is open in S0 , so S0 \ O1 is closed in S0 . Then S0 \ O1 = S0 \ O1 = S0 , i.e. O1 = ∅, a contradiction.

Theorem. Let S ⊆ R. S is connected iff S is an interval.

Proof. S = ∅ or R is clear. For other intervals, we have three cases. Let a, b ∈ R with a < b.
b−a π

Case 1: S = (a, b). Define f : R → R by f(x) = Arctan x + + a. We have f is continuous and R is
π 2
connected. By the connected image theorem, f(R) = (a, b) is connected.
Case 2: S = (−∞, a) or (a, ∞). Define f : R → R by f(x) = a − ex or f(x) = a + ex . Since f is continuous,
by the connected image theorem, f(R) = (−∞, a) or (a, ∞) is connected.
Case 3: S is a (not necessarily open) interval. Let I be the interval obtained from S by removing its
endpoints. By cases 1 and 2, I is connected. Since I ⊆ S ⊆ I, by the last theorem, S is connected.
Conversely, if S is connected, to see S is an interval, it suffices to show a, b ∈ S and a < c < b imply
c ∈ S. Assume c 6∈ S. Then S = (S ∩ (−∞, c)) ∪ (S ∩ (c, ∞)) with S ∩ (−∞, c), S ∩ (c, ∞) both open in S,
disjoint and nonempty as a ∈ S ∩ (−∞, c) and b ∈ S ∩ (c, ∞). This contradicts S is connected. Therefore,
c ∈ S and S is an interval.

Remark. If f : M → R is continuous and M connected, then f(M ) is connected, hence an interval. This is
the intermediate value theorem for connected metric spaces.

Definitions. (1) S is path-connected iff for every x, y ∈ S, there is a continuous function f : [0, 1] → S such
that f(0) = x and f(1) = y. Such a function f is called a path from x to y in S.
(2) Let S ⊆ Rn . S is polygonally-connected iff S is path-connected with the paths can be chosen to be
polygonal paths (i.e. the ranges consist of polygonal segments).

Remarks. By definition, polygonally connected implies it is path connected. However, a circle is path
connected, but not polygonally connected. If S is connected, it is possible that S is not path connected (for
example, take S = A ∪ B, where A = {(x, y) : y = sin(1/x), x ∈ (0, 1]} and B = {(0, y) : y ∈ [−1, 1]}).

Theorem. If S is a path connected metric space, then S is connected.

Proof. Assume S = O1 ∪ O2 for some disjoint, nonempty open O1 and O2 in S, say x ∈ O1 and y ∈ O2. Let
f : [0, 1] → S be a path from x = f(0) to y = f(1). Then [0, 1] = f −1 (O1 )∪f −1 (O2), where f −1 (O1), f −1 (O2 )
are disjoint, nonempty as 0 ∈ f −1 (O1 ), 1 ∈ f −1 (O2) and open in [0, 1]. This contradicts [0, 1] is connected.

Theorem. Let O be a nonempty open set in Rn . Then O is connected iff O is polygonally connected.
Proof. By definition, O polygonal connected implies O path connected. By the last theorem, O is connected.
Conversely, let O be connected and x ∈ O. Let O1 be the set of y ∈ O such that there exists a polygonal
path in O from x to y. Note x ∈ O1 . So O1 6= ∅. All we have to show is O1 = O.
First, O1 is open in O. To see this, let y ∈ O1, then since O is open, there is B(y, r) ⊆ O. From y, we
can draw a radius to connected to any point in B(y, r). Since there is a polygonal path from x to y, we see
B(y, r) ⊆ O1 .

18
 Second, O \ O1 is open. If y ∈ O \ O1, then again there is B(y, r) ⊆ O and a radius can connect y to
any point in B(y, r). Since there is no polygonal path from x to y, there cannot be any polygonal path from
x to any point of B(y, r). So B(y, r) ⊆ O \ O1.
Finally, O = O1 ∪ (O \ O1) and O1, O \ O1 are disjoint open in O. Since O is connected and O1 is
nonempty, we must have O \ O1 = ∅. Therefore, O1 = O.

§8. Product Spaces

For a metric space M with metric d, as simple as asking if d : M × M → [0, +∞) is continuous, we need
to consider M × M. So we now discuss the formation of new metric spaces by taking the Cartesian products
of metric spaces and introduce metrics on them as in the Rn example.

Definition. Let (M1 , δ1 ) and (M2 , δ2) be two metric spaces. For 1 ≤ p ≤ ∞, we can define a dp metric on
M1 × M2 = {(x1, x2) : x1 ∈ M1 , x2 ∈ M2 } by assigning to x = (x1 , x2) and y = (y1 , y2 ) the distance

1/p
dp (x, y) = δ1 (x1, y1 )p + δ2 (x2, y2 )p if 1 ≤ p < ∞ .
max{δ1 (x1, y1 ), δ2(x2 , y2)} if p = ∞

Remarks. (1) Just as in Rn , the dp metrics above are uniformly equivalent for 1 ≤ p ≤ ∞. In particular,
the open sets, the Cauchy sequences and the limits (if exist) are the same for all dp metrics.
(2) We prefer to use the d∞ metric because d∞ (x, y) < r iff δ1 (x1 , y1) < r and δ2 (x2, y2 ) < r. Thus, the open
ball in M1 × M2 with the d∞ metric is B(x, r) = B(M1 ,δ1 ) (x1, r) × B(M2 ,δ2 ) (x2 , r).

Example. Let (M, d) be a metric space. We would like to check d : M ×M → [0, ∞) is uniformly

continuous.
Let us use the d∞ metric on M × M. For every ε > 0, let δ = ε/2, then d∞ (x, y), (x0, y0) < δ implies


|d(x, y)−d(x0, y0 )| ≤ |d(x, y)−d(y, x0)|+|d(y, x0)−d(x0, y0 )| ≤ d(x, x0)+d(y, y0 ) ≤ 2d∞ (x, y), (x0, y0 ) < ε.

To study properties of M1 × M2 , we often look at each of the coordinate spaces M1 and M2 .

Definitions. (1) For i = 1, 2, πi : M1 × M2 → Mi with πi(x1 , x2) = xi is called the projection onto Mi .
(2) Let X, Y be metric spaces. We say f : X → Y is open iff for every open set O in X, f(O) is open in Y.

Theorem. For i = 1, 2, πi : M1 × M2 → Mi with πi(x1 , x2) = xi is open and uniformly continuous.

Proof. Let O be open in M1 × M2 . To see π1(O) is open in M1 , let x1 ∈ π1(O). Then there exists x2 ∈ M2
such that x = (x1, x2) ∈ O. Since O is open, using the d∞ metric, there exists

B(M1 ,δ1 ) (x1 , r) × B(M2 ,δ2 ) (x2 , r) = B(x, r) ⊆ O.

Then B(M1 ,δ1 ) (x1, r) = π1 B(x, r) ⊆ π1(O). So x1 ∈ π1 (O)◦ . Therefore, π1 (O) is an open set in M1 and π1
is open. Similarly, π2 is open.
Next, for every ε > 0, take δ = ε. For x = (x1, x2) and y = (y1 , y2 ) ∈ M1 × M2 , d∞ (x, y) < δ = ε implies
max{δ1 (x1 , y1), δ2 (x2, y2 )} < ε, which implies δ1 (π1(x), π1(y)) = δ1 (x1 , y1) < ε and δ2 (π2(x), π2(y)) =
δ2 (x2 , y2) < ε. Therefore, π1 and π2 are uniformly continuous.

The next theorem will tell us many of the properties introduced in the earlier sections hold for a product
space iff it holds for each of the coordinate spaces.

Product Theorem. Let (M1 , δ1) and (M2 , δ2 ) be metric spaces.

(1) For a sequence {(x1,n, x2,n)} in M1 × M2 , we have (x1,n, x2,n) → (x1, x2) in M1 × M2 iff x1,n → x1 in
M1 and x2,n → x2 in M2 .

19
(2) M1 × M2 is separable iff M1 and M2 are separable.
(3) {(x1,n, x2,n)} is a Cauchy sequence in M1 × M2 iff {x1,n} is a Cauchy sequence in M1 and {x2,n} is a
Cauchy sequence in M2 .
(4) M1 × M2 is complete iff M1 and M2 are complete.
(5) M1 × M2 is totally bounded iff M1 and M2 are totally bounded.
(6) M1 × M2 is compact iff M1 and M2 are compact.
(7) M1 × M2 is connected iff M1 and M2 are connected.
Proof. (1) Since δi (xi, yi ) ≤ d∞ (x, y) ≤ δ1 (x1, y1 ) + δ2 (x2, y2 ), by the sandwich theorem,


d∞ (x1,n, x2,n), (x1, x2) → 0 ⇐⇒ δ1 (x1,n, x1) → 0 and δ2 (x2,n, x2) → 0.
(2) Let xn = (x1,n, x2,n) ∈ M1 × M2 . By the sequential closure theorem and (1), {x1, x2, x3, . . .} = M1 × M2
implies {x1,1, x1,2, x1,3, . . .} = M1 , {x2,1, x2,2, x2,3, . . .} = M2 , which imply {(x1,n, x2,m) : n, m = 1, 2, 3, . . .}
= M1 × M2 .
(3) We need to show


∀ ε > 0, ∃ a positive integer K such that m, n ≥ K ⇒ d∞ (x1,m, x2,m), (x1,n, x2,n) < ε
iff
∀ ε > 0, ∃ a positive integer K1 such that m, n ≥ K1 ⇒ δ1 (x1,m , x1,n) < ε
and
∀ ε > 0, ∃ a positive integer K2 such that m, n ≥ K2 ⇒ δ2 (x2,m , x2,n) < ε.
For the if direction, take K = max{K1 , K2}. For the only-if direction, take K1 = K and K2 = K.
(4) This follows from (1), (3) and the definition of completeness.
(5) For every ε > 0, if M1 can be covered by U1 = {B(M1 ,δ1 ) (x1,i, ε) : i = 1, 2, . . ., m} and M2 can be


covered by U2 = {B(M2 ,δ2 ) (x2,j , ε) : j = 1, 2, . . ., n}, then M1 × M2 can be covered by {B (x1,i, x2,j ), ε : i =
1, 2, . . ., m; j = 1, 2, . . ., n}.


Conversely, if M1 × M2 can be covered by U = {B xk , ε : k = 1, 2, . . ., r}, then for i = 1, 2, Mi can be
covered by {B(Mi ,δi ) (πi(xk ), ε) : k = 1, 2, . . ., r}.
(6) This follows from (3), (4) and the metric compactness theorem.
(7) If M1 × M2 is connected, then since the projection πi onto Mi is continuous for i = 1, 2, so by the
connected image theorem, Mi = πi (M1 × M2 ) is connected for i = 1, 2.
Conversely, suppose M1 and M2 are connected. For every x1 ∈ M1 and x2 ∈ M2 , define f : M1 →
M1 × M2 by f(x) = (x, x2) and g : M2 → M1 × M2 by g(x) = (x1 , x). By the sequential continuity
theorem and (1), we see f, g are continuous. Then, by the connected image theorem, M1 × {x2} = f(M1 )
and {x1} × M2 = g(M2 ) are connected. Since their intersection contain (x1, x2), by the connected union
theorem, their union S(x1 ,x2 ) = (M1 × {x2}) ∪ ({x1 } × M2 ) is connected.
For all (y1 , y2 ) ∈ M1 × M2 different from (x1, x2), the set S(x1 ,x2 ) ∪ S(y1 ,y2 ) is connected because their
intersection contains (x1, y2 ) and (x2 , y1). Since S(x1 ,x2) ∪ S(y1 ,y2 ) contains (x1, x2) for all (y1 , y2 ) 6= (x1 , x2),
so [

M1 × M2 = S(x1 ,x2) ∪ S(y1 ,y2 )
(y1 ,y2 )6=(x1 ,x2 )

is connected.

Remarks. Product of finitely many metric spaces can be defined similarly and the results above hold for
them by mathematical induction. In particular, Rn or Cn are complete.

§9. Completion of Metric Spaces

20
 In this section, our goal is to show every metric space M can be “extended densely” to a complete
c in a “unique” way. We begin with a few definitions.
metric space M


Definitions. (1) f : (M1 , d1) → (M2 , d2) is an isometry iff for all x, y ∈ M1 , d2 f(x), f(y) = d1(x, y).
(2) M1 and M2 are isometric iff there exists a bijective isometry f : M1 → M2 .
(3) Mc is a completion of M iff M
c is complete and there is an isometry f : M → M c with f(M ) dense in M.
c


Remarks. (1) Every isometry is injective because f(x) = f(y) implies d1(x, y) = d2 f(x), f(y) = 0,
i.e. x = y. Every isometry is uniformly continuous since we can take δ = ε to check the definition.
(2) Isometric spaces are considered the same metric spaces since there is a one-to-one correspondence between
the elements and the distances between corresponding pairs of elements are preserved.
The next theorem will show a metric space can have at most one completion (up to isometry). For the
proof, we first recall some facts from the section on completeness and contractions. Namely, (1) complete
subsets are closed in any metric space; (2) for a complete metric space, closed subsets are complete; (3) for
a uniformly continuous function f : M → M 0, if {xn} is a Cauchy sequence in M, then {f(xn } is a Cauchy
sequence in M 0. So the image of a Cauchy sequence under an isometry is also a Cauchy sequence. Therefore,
for two isometric spaces, one of them is complete iff the other one is also complete.

Theorem. Let (M, dM ), (N, dN ) be metric spaces, N complete and S ⊆ M. If f : S → N is uniformly

continuous, then there exists a uniformly continuous F : S → N such that for all x ∈ S, F (x) = f(x). If f
is an isometry, then F is an isometry.
Proof. Let x ∈ S. Suppose {xn} and {x∗n} are sequences in S both converges to x. Then both are Cauchy
sequences and dM (xn , x∗n) → dM (x, x) =
0. Since f is uniformly continuous, both {f(xn )} and {f(x∗n )} are
Cauchy sequences and dN f(xn ), f(x∗n ) → 0. Since N is complete, both {f(xn )} and {f(x∗n )} converge to
the same limit, call it F (x), in N. This defines a function F : S → N. For x ∈ S, taking xn = x, we see
F (x) = f(x).
For every
ε > 0, since f is uniformly continuous, there is δ > 0 such that for all x, y ∈ S, dM (x, y) < δ ⇒
dN f(x), f(y) < 0.99ε. For a, b ∈ S, take sequences {xn } in S converging to a and {yn} in S converging to


b. If dM (a, b) < δ, then for large n, dM (xn, yn ) < δ, hence dN f(xn ), f(yn ) < 0.99ε. Taking limit, we get


dN F (a), F (b) ≤ 0.99ε < ε. Therefore, F is uniformly continuous. If f is an isometry, then



dN F (a), F (b) = lim dN f(xn ), f(yn ) = lim dM xn, yn ) = dM (a, b).
n→∞ n→∞

Theorem. If (M1 , d1) and (M2 , d2) are both completions of (M, d), then M1 and M2 are isometric, i.e. com-
pletion is unique (up to isometry).
Proof. Suppose for i = 1, 2, there are isometries fi : (M, d) → (Mi , di) with fi (M ) dense in Mi . Observe
that f = f2 ◦ f1−1 : f1 (M ) → f2 (M ) ⊆ M2 is an isometry, hence is uniformly continuous. By the theorem
above, f can be extended to F : f1 (M ) = M1 → M2 . Since f1 and f2 are isometries, f and F are also
isometries. By the underlined statement in remark (2) above, M1 complete and F isometry imply F (M1 )
is complete, hence closed in M2 . Since f2 (M ) = (f ◦ f1 )(M ) = F (f1(M )) ⊆ F (M1) ⊆ M2 = f2 (M ), we get
F (M1) = F (M1) = M2 . Therefore, F is a bijective isometry and M1 , M2 are isometric.

b up to isometry.
c, d)
Theorem. Every metric space (M, d) has a (unique) completion (M
Proof. Let Cb(M, R) be the set of all bounded continuous functions from M to R. For f, g ∈ Cb(M, R),
b g) = sup{|f(x) − g(x)| : x ∈ M }. (Cb(M, R), d)
define d(f, b is a complete metric space (Exercise).
Fix m ∈ M. For every x ∈ M, define fx : M → R by fx (y) = d(y, x) − d(y, m). This is continuous since
d is continuous on M × M. Observe that since |fx (y)| = |d(y, x) − d(y, m)| ≤ d(x, m), fx is bounded. So we
may define p : M → Cb(M, R) by p(x) = fx . Since for all y ∈ M,

|fx1 (y) − fx2 (y)| = |d(y, x1) − d(y, m) − d(y, x2 ) + d(y, m)| = |d(y, x1) − d(y, x2)| ≤ d(x1, x2)

21
 b x , fx ) = d(x1, x2). Hence, p is an isometry. Now M
and fx1 (x2) − fx2 (x2) = d(x1, x2), so d(f c = p(M ) is a
1 2

closed subset in Cb (M, R), hence it is complete. Therefore, (Mc, d)

b is a completion of (M, d).

Remarks. There is another way to get a completion for M. Let S be the set of all Cauchy sequences in M.
(1) Let R = {({xn}, {x∗n}) : {xn}, {x∗n} ∈ S, lim d(xn, x∗n) = 0}. We can check R is an equivalence relation
n→∞
c be the set of all equivalence classes [{xn}].
on S. Let M
(2) We can check that if ({xn}, {x∗n}), ({yn}, {yn∗ }) ∈ R, then lim d(xn, yn) = lim d(x∗n, yn∗ ). This allows
n→∞ n→∞
b
us to define a metric d([{x c c b
n }], [{yn}]) = lim d(xn , yn ) on M. We can check (M , d) is complete.
n→∞

c defined by f(x) = [{xn}], where all xn = x, is an isometry and f(M ) = M

(3) We can check f : M → M c.

For example, using this procedure to get Q b = R, we have [0.3, 0.33, 0.333, . . .] = [1/3, 1/3, 1/3, . . .] ∈ Q
b
b
correspond to 1/3 ∈ R and [3, 3.1, 3.14, 3.141, 3.1415, . . .] ∈ Q correspond to π ∈ R.

Appendix
Here we present a few special theorems that are useful in certain applications.
Urysohn’s Theorem. Let A and B be disjoint closed sets in M. Then there exists a continuous function
f : M → [0, 1] such that f(A) = {0} and f(B) = {1}.
Proof. Note d(x, A) + d(x, B) = 0 implies d(x, A) = 0 = d(x, B), which implies x ∈ A ∩ B = A ∩ B = ∅, a
d(x, A)
contradiction. So for all x ∈ M, d(x, A) + d(x, B) > 0. Let f(x) = , then f is continuous,
d(x, A) + d(x, B)
0 ≤ f(x) ≤ 1, f(A) = {0} and f(B) = {1}.

Theorem. Let A and B be disjoint closed sets in M. Then there exists disjoint open sets C and D in M
such that A ⊆ C and B ⊆ D.
Proof. Take f as in Urysohn’s theorem, C = f −1 ((−∞, 1/3)) and D = f −1 ((2/3, ∞)). Then C ∩ D =
f −1 ((−∞, 1/3) ∩ (2/3, ∞)) = f −1 (∅) = ∅, C and D are open in M by the topological continuity theorem,

A ⊆ f −1 ({0}) ⊆ f −1 ((−∞, 1/3)) = C and B ⊆ f −1 ({1}) ⊆ f −1 ((2/3, ∞)) = D.

Definition. Let f : M → R be continuous. The support of f is supp f = {x ∈ M : f(x) 6= 0}.

Partition of Unity Theorem. (1) If U = {O1, O2, . . . , On} is a finite open cover of a metric space (H, δ),
then there exist continuous h1 , h2, . . . , hn : H → [0, 1] such that for all x ∈ H, h1(x) + h2 (x) + · · ·+ hn (x) = 1
and for every i = 1, 2, . . ., n, supp hi ⊆ Oi .
(2) Let K be compact in a metric space (M, d) and U = {O1, O2, . . . , On} be an open cover of K. Then there
exist continuous f1 , f2, . . . , fn : M → [0, 1] such that for all x ∈ K, f1 (x) + f2 (x) + · · · + fn (x) = 1 and for
every i = 1, 2, . . ., n, supp fi ⊆ Oi.
Proof. (1) (Due to Michael Mather) Since U covers H, for every x ∈ H, there exist Oi ∈ U and r > 0
such that B(x, r) ⊆ Oi, which implies δ(x, H \ Oi ) ≥ r > 0. So the function p : H → R defined by
Xn n p(x) o
p(x) = δ(x, H \ Oi) is positive-valued. Define gi : H → [0, ∞) by gi(x) = max δ(x, H \ Oi) − ,0 .
i=1
n+1
n 
p(x) 
X n X p(x)
Then gi is continuous, gi(x) ≥ δ(x, H \ Oi) − = > 0 and
i=1 i=1
n+1 n+1

n p(x) o n p(x) o
supp gi = x : δ(x, H \ Oi ) > ⊆ x : δ(x, H \ Oi) ≥ ⊆ {x : δ(x, H \ Oi) > 0} = Oi.
n+1 n+1
Therefore, taking hi(x) = gi(x)/(g1 (x) + g2(x) + · · · + gn(x)), we are done.

22
(2) H = O1 ∪ O2 ∪ · · · ∪ On is open in M. Since M \ H and K are disjoint closed sets, by Urysohn’s theorem,
there is a continuous function f : M → [0, 1] such that f(M \ H) = {0} and f(K) = {1}. By part (1), there
exist continuous h1 , h2, . . . , hn : H → [0, 1] such that for all x ∈ H, h1(x) + h2 (x) + · · · + hn (x) = 1 and
for every i = 1, 2, . . ., n, supp hi ⊆ Oi . The functions fi = fhi : M → [0, 1] are continuous, for all x ∈ K,
f1 (x) + f2 (x) + · · · + fn(x) = h1(x) + h2 (x) + · · · + hn(x) = 1 and supp fi ⊆ supp hi ⊆ Oi .

Remarks. In part (2) of the theorem, if g : M → R is continuous and supp g ⊆ K, then g = gf1 +gf2 +· · ·+
gfn and supp gfi ⊆ Oi for i = 1, 2, . . ., n. This is useful when we need to shift the support of a continuous
function to small sets. For example, see Walter Rudin’s book Principle of Mathematical Analysis, third
edition, pp. 252-253 for the proof of the change of variable formula for multiple integration.

Lemma. Let Y be a nonempty closed set in M. Let f : Y → R be continuous with sup f(Y ) = k > 0 and
inf f(Y ) = −k. Then there exists a continuous g : M → R such that
(a) ∀x ∈ M, |g(x)| ≤ k/3, (b) ∀x ∈ M \ Y, |g(x)| < k/3 and (c) ∀x ∈ Y, |f(x) − g(x)| ≤ 2k/3.

Proof. By the topological continuity theorem, A = f −1 ([−k, −k/3]) and B = f −1 ([k/3, k]) are closed
(and disjoint) in Y. Since Y is closed in M, so A and B are closed in M. The function g : M → R by
k  d(x, A) − d(x, B) 
g(x) = has positive denominator (as in the proof of Urysohn’s theorem), is continuous
3 d(x, A) + d(x, B)
and condition (a) holds. If x ∈ M \ Y, then x 6∈ A and x 6∈ B so that condition (b) holds. For condition (c),
let x ∈ Y. Then
x∈A ⇒ g(x) = −k/3, −k ≤ f(x) ≤ −k/3 ⇒ |f(x) − g(x)| ≤ 2k/3,
x∈B ⇒ g(x) = k/3, k/3 ≤ f(x) ≤ k ⇒ |f(x) − g(x)| ≤ 2k/3,
x ∈ Y \ (A ∪ B) ⇒ |g(x)| < k/3, |f(x)| < k/3 ⇒ |f(x) − g(x)| < 2k/3.

Tietze’s Extension Theorem. Let Y be a nonempty closed set in M. If f : Y → R is continuous, then

there exists continuous g : M → R such that for all y ∈ Y, g(y) = f(y).
Proof. If f is boundeded, then let k = 12 (sup f(Y ) + inf f(Y )), k0 = 12 (sup f(Y ) − inf f(Y )) and f1 (x) =
f(x) − k on Y. We have |f1(x)| ≤ k0 .
Let ki = (2/3)i−1k0 for k = 1, 2, 3, . . .. By mathematical induction, we can get two sequences of functions
fi : Y → [−ki , ki] and gi : M → [−ki/3, ki/3] satisfying conditions (a), (b), (c) of the lemma (with fi , gi, ki
replacing f, g, k). This is done by applying the lemma to get a gi for fi and define fi+1 = fi − gi on Y.
X
∞ X
∞
ki X
∞
Since for all x ∈ M, |gi(x)| ≤ < ∞, by the Weierstrass M -test, g∞ = gi is continuous on
i=1 i=1
3 i=1
∞
X ∞
X ∞
X ki
M. On Y, g∞ = gi = = k0. Define g : M → R by g(x) = g∞ (x)+k.
(fi+1 − fi ) = f1 and |g∞| <
i=1 i=1
3 i=1
Then g is continuous on M, g = f on Y and inf f(Y ) < g(x) < sup f(Y ) for all x ∈ M \ Y.
Finally, if f is unbounded, then we let F (x) = Arctan f(x) on Y. Apply the bounded case above to F
to get a G. Then take g(x) = tan G(x) on M.

Remarks. We cannot go without mentioning Harald Bohr’s proof of Tietze’s extension theorem, where a
formula can be given for such an extension function. We first reduce to the case f is bounded and f ≥ 0.
Define Mx (r) = sup{f(y) : y : Y ∩ B(x, r)} and δ(x) = d(x, Y ). Note Mx (r) is bounded and increasing in r,
hence integrable. For x ∈ M \ Y, Bohr’s formula is
Z 2δ(x)
1
g(x) = Mx (r) dr.
δ(x) δ(x)
For details of Bohr’s proof, see Avner Friedman’s book Foundation of Modern Analysis, pp. 111-112. Harald
Bohr was the brother of the famous Niels Bohr in Quantum Mechanics. Harald Bohr played football for the
Danish team in the 1908 Summer Olympics!

23
 Chapter 3. Normed Vector Spaces

§1. Definition and Examples of Normed Vector Spaces

In advanced courses, vector spaces are also called linear spaces. In analysis, we consider vector spaces
over K = R or C.

Definitions. (1) A normed vector space (or normed linear space or normed space) is a vector space V over
K with a function that assigns to every x ∈ V a norm kxk ∈ R such that for every x, y ∈ V and α ∈ K,
(a) kxk ≥ 0 (with kxk = 0 if and only if x = 0);
(b) kαxk = |α|kxk;
(c) kx + yk ≤ kxk + kyk.

(2) Defining d(x, y) = kx − yk for all x, y ∈ V will make a normed vector space V into a metric space. We
say the metric d is induced by the norm k · k. (Convergence of sequences, continuity of functions, Cauchy
sequences and completeness are defined for normed spaces by using the induced metrics as for metric spaces.)

(3) A normed vector space is called a Banach space iff it is complete as a metric space.

(4) k · k1 and k · k2 on V are equivalent iff the metrics they induced are equivalent.

Remarks. (1) A metric space is a set, possibly no addition, no scalar multiplication and not even a 0
element! A normed vector space is a vector space with addition, scalar multiplication and a 0 element!
(2) Balls in metric spaces may have arbitrary shapes (for example, take an arbitrary subset of the unit disk
in R2 as the metric space with the same metric as in R2 ). However, in normed vector spaces, every ball B
is convex (i.e. for every x, y ∈ B and t ∈ [0, 1], we have tx + (1 − t)y ∈ B).
(3) In normed vector space, Bclosed (x, r) = {y : d(x, y) ≤ r} is equal to B(x, r). (This is because yn ∈ B(x, r)
converges to y implies d(y, x) = lim d(yn , x) ≤ r, hence Bclosed (x, r) is a closed set containing B(x, r). So
n→∞
1 1
B(x, r) ⊆ Bclosed (x, r). Conversely, for every y ∈ Bclosed (x, r), we see yn = x + (1 − )y is in B(x, r)
n n
1
because d(yn , x) = (1 − n1 )ky − xk < r. Also d(yn , y) = kx − yk → 0, which implies y ∈ B(x, r). So
n
Bclosed (x, r) ⊆ B(x, r).)
(4) In a normed vector space V, addition and scalar multiplication are continuous. To see this, observe that
if xn → x, yn → y in V and αn → α in K, then d(xn, x) = kxn − xk → 0, d(yn , y) = kyn − yk → 0 and
|αn − α| → 0 so that |αn| is bounded and

d(xn + yn , x + y) = k(xn + yn ) − (x + y)k ≤ kxn − xk + kyn − yk → 0,

d(αnxn, αx) = kαnxn − αxk ≤ kαnxn − αnxk + kαnx − αxk = |αn|kxn − xk + |αn − α|kxk → 0.

Also, the norm is continuous since xn → x implies kxnk → kxk because kxnk − kxk ≤ kxn − xk → 0.

X
n 1/p
Examples. (1) Kn = Rn or Cn with x = (x1, x2, . . . , xn) assigned k(x1, x2, . . . , xn)kp = |xj |p is a
j=1
Banach space for 1 ≤ p ≤ ∞. (Recall Rn is complete from product space discussion.)
n X
∞ o X
∞ 1/p
(2) For 1 ≤ p < ∞, `p = (x1, x2, x3, . . .) : all xj ∈ K, |xj |p < ∞ with k(x1, x2, x3, . . .)kp= |xj |p
j=1 j=1
is a normed space. Also, `∞ = {(x1 , x2, x3, . . .) : all xj ∈ K, sup |xj | < ∞} with k(x1, x2, x3, . . .)k∞ = sup |xj |

24
is a normed space. The triangle inequality holds by taking limit in Minkowski inequality. They will be shown
to be Banach spaces in the examples of section 4.
X∞ 1/p X ∞ 1/p X∞ 1/p
However, for 0 < p < 1, the inequality |xj + yj |p ≤ |xj |p + |yj |p does not
j=1 j=1 j=1
hold in general. For example, let x = (1, 0, 0, 0, . . .) and y = (0, 1, 0, 0, . . .), then x + y = (1, 1, 0, 0, . . .), but
for 0 < p < 1, (1p + 1p )1/p = 21/p > 2 = 11/p + 11/p .
(3) Let m denote the Lebesgue measure on R and X be a nonempty Lebesgue measurable subset of R.
n Z o
For 1 ≤ p < ∞, consider measurable functions f : X → R and let Lp = f : |f|p dm < ∞ .
Z 1/p X
p
If we define kfkp = |f| dm , then we do not get a norm since kfkp = 0 if and only if f = 0
X
almost everywhere on X (not f = 0 on X).
For p = ∞, considering measurable f : X → R, we define the essential supremum of f by

kfk∞ = ess sup|f| = inf{c ∈ [0, +∞] : |f(x)| ≤ c almost everywhere on X}

and L∞ = {f : kfk∞ < ∞}. Again, kfk∞ = 0 if and only if f = 0 almost everywhere on X.
To make these into normed vector spaces, for 1 ≤ p ≤ ∞, consider

R = {(f, g) : f, g ∈ Lp , f = g almost everywhere on X}.

We easily check this is an equivalence relation on Lp . The collection of equivalence classes Lp = {[f] : f ∈ Lp }
is a vector space with addition and scalar multiplication defined by [f] + [g] = [f + g] and c[f] = [cf].
Next, define k[f]kp = kfkp . This is well-defined since [f] = [g] (⇐⇒ (f, g) ∈ R) implies f = g a.e. on X
and so kfkp = kgkp . In particular, k[f]kp = 0 if and only if [f] = [0]. Also, kα[f]kp = |α|k[f]kp is clear. The
triangle inequality follows from the integral version of Minkowski’s inequality. Therefore, Lp is a normed vec-
tor space. In fact, we will prove later that Lp is a Banach space. It is a common practice to write f ∈ Lp to
mean [f] ∈ Lp (with the understanding that f = g in Lp still means [f] = [g], i.e. f = g almost everywhere).

(4) Let M be a metric space. For every bounded function f : M → K, define the sup-norm of f to be kfkM
or kfk∞ = sup{|f(x)| : x ∈ M }. Then B(M ) = {f : kfk∞ < ∞} is a Banach space because {fn } Cauchy
implies for all x ∈ M, {fn(x)} Cauchy, hence lim fn(x) = f(x) exists. Since there exists K such that for
n→∞
m, n ≥ K, kfn − fm k∞ < ε, so |fn (x) − f(x)| = lim |fn(x) − fm (x)| ≤ ε. Then |f(x)| ≤ |fn (x)| + ε. Hence,
m→∞
kfk∞ ≤ kfK k∞ + ε. Therefore, f ∈ B(M ) and d(fn, f) = kfn − fk∞ → 0.
The collection C(M ) = {f : f is continuous on M } is a vector space, but the functions may be un-
bounded. The collection Cb(M ) or BC(M ) = B(M ) ∩ C(M ) is a normed vector space with sup-norm.
Since convergence in sup-norm is unform convergence, so limit of continuous functions is continuous. Hence
Cb(M ) is a closed subspace of B(M ) and so Cb(M ) is a Banach space. In particular, when M is compact,
C(M ) = Cb(M ) is a Banach space with sup-norm.
(5) For a metric space M, the support of f : M → K is supp f = {x ∈ M : f(x) 6= 0}. Observe that x 6∈ supp f
iff f = 0 on some B(x, r). From this, we get supp (f +g) ⊆ supp f ∪supp g and supp (αf) = supp f for α 6= 0.
These facts imply Cc (M ) = {f ∈ C(M ) : supp f is compact} is a vector space, i.e. for every f, g ∈ Cc (M )
and α ∈ K, we have f + g, αf ∈ Cc (M ). With sup-norm, Cc(M ) is a normed vector subspace of B(M ).
When M is compact, Cc(M ) = C(M ) is a Banach space. However, Cc (M ) is not complete in general.
1
To see this, take M = R. Then f(x) = 6∈ Cc(R) as supp f = R. For every positive integer n, the
1 + x2
function 
 f(x) if |x| ≤ n − 1,
fn (x) = f(n − 1)(n − |x|) if n − 1 < |x| ≤ n,

0 if |x| ≥ n

25
 2
has [−n, n] as support. Hence fn ∈ Cc(M ). Now d(fn , f) = kfn − fk∞ ≤ 2f(n − 1) = →0
1 + (n − 1)2
implies fn → f in B(M ). So {fn } is a Cauchy sequence in B(M ), hence also in Cc (M ) as both use sup-norm
as norm. However, f not in Cc (M ) implies {fn} diverges in Cc(M ). Therefore, Cc(M ) is not complete.
Z 1 1/p
(6) Other than sup-norm, on C([0, 1]), we can use kfkp = |f(x)|p dx with 1 ≤ p < ∞ as norm.
0
This will also make C([0, 1]) into a normed vector space, but it is not complete. We will show this for p = 2.
Let f(x) = x−1/4 for x ∈ (0, 1] and

x−1/4 if 1/n ≤ x ≤ 1
fn (x) =
n1/4 if 0 ≤ x < 1/n.
Z 1
Now fn’s are in C([0, 1]), but f is not in C([0, 1]). We have f(x)2 dx = 2 and
0
Z 1 Z 1/n Z 1/n
3
(f(x) − fn (x))2 dx = (x−1/4 − n1/4)2 dx ≤ (x−1/2 + n1/2) dx = √ .
0 0 0 n

For every ε > 0, let K > 144ε−4, then for n, m ≥ K, by Minkowski’s inequality, we have
s s s
Z 1 Z 1 Z 1
kfn − fm k2 = (fn (x) − fm (x))2 dx ≤ (fn (x) − f(x))2 dx + (f(x) − fm (x))2 dx
0 0 0
√ √ √
3 3 2 3
< + ≤ 1/4 < ε.
n1/4 m1/4 K
This checks {fn } is Cauchy in C([0, 1]). Assume {fn } converges to g ∈ C([0, 1]). By Minkowski’s inequality,
s s s √
Z 1 Z 1 Z 1
2 2 2
3
(f(x) − g(x)) dx ≤ (f(x) − fn (x)) dx + (fn (x) − g(x)) dx ≤ 1/4 + kfn − gk2 → 0,
0 0 0 n

which implies the left hand side is zero. So f = g on (0, 1] as they are continuous on (0, 1]. Then g is
unbounded, contradicting g ∈ C([0, 1]). Therefore, C([0, 1]) with k · k2 is not complete. From measure
theory, we can see the completion of C[0, 1] with k · kp is Lp ([0, 1]) for 1 ≤ p < ∞.

§2. Continuous Linear Transformations

Definition. Let V and W be normed spaces with norm k·kV and k·kW , respectively. A linear transformation
T : V → W is bounded iff there exists c > 0 such that for every x ∈ V, kT xkW ≤ ckxkV .

Theorem. Let V and W be normed spaces and T : V → W be linear. Then the following are equivalent:
(a) T is bounded;
(b) T is uniform continuous;
(c) T is continuous;
(d) T is continuous at 0;
(e) for every bounded set S in V, T (S) is a bounded set in W.
Proof. For (a) implies (b), use δ = ε/c. Next, (b) implies (c) and (c) implies (d) are trivial. For (d) implies
(e), since T is continuous at 0, for ε = 1, there exists δ > 0 such that kwk < δ implies kT wk < ε = 1.
δx
Let c = 2/δ. If S is bounded, say S ⊆ B(0, r), then x ∈ S implies kxk < r. Let w = . Then kwk < δ. So
r
r r
kT xk = kT wk < . Therefore, T (S) is in B(0, r/δ).
δ δ
26
 Finally, for (e) implies (a), letting S = B(0, 1), by
 (e), we have T (B(0, 1)) ⊆ B(0, c) for some c > 0. For
x x 
every nonzero x ∈ V, since ∈ B(0, 1), we get T < c, which implies kT xk ≤ ckxk. Clearly, this
kxk kxk
also holds if x = 0. Therefore, T is bounded.

Corollary. Let k · k1 and k · k2 be two norms on a vector space V. The following are equivalent:
(a) k·k1 is equivalent to k·k2 (i.e. the identity function i : (V, k·k1) → (V, k·k)2) is a homeomorphism);
(b) i : (V, k · k1) → (V, k · k)2 ) is a uniform homeomorphism, i.e. i and i−1 are uniformly continuous
(in this case, we say k · k1, k · k2 are uniformly equivalent);
(c) there exists c1 , c2 > 0 such that for all x ∈ V, c1 kxk1 ≤ kxk2 ≤ c2 kxk1.

Example. Unlike Kn , on C([0, 1]), if 1 ≤ p < q ≤ ∞, then the k · kq norm and the k · kp norm are
not equivalent. (Assume they are equivalent. By (c), there would exist some c > 0 such that for all
f ∈ C([0, 1]), we have kfkq ≤ ckfkp . So kfkq /kfkp is bounded by c. Consider fn (x) = xn, where n ∈ N.
Then kfnk∞ /kfn kp = 1/(np + 1)−1/p ≈ n1/pp1/p → ∞ as n → ∞. We get a contradiction. Similarly, for
1 ≤ p < q < ∞, kfnkq /kfn kp ≈ n1/p−1/q p1/p q−1/q → ∞ as n → ∞. So no such c can exist.)

§3. Finite Dimensional Normed Spaces

In this section, we will establish some useful facts about finite dimensional normed spaces.
Theorem. Let V be a normed space over K with n = dim V < ∞. Then V is uniform homeomorphic to
(Kn , k·k2). Consequently, any two norms on V are equivalent. Also, any finite dimensional normed subspace
is complete (hence closed) in any normed space.
Proof. Let {x1, x2, . . . , xn} be a basis of V. Define g : (Kn, k · k2 ) → (V, k · k) by g((α1, α2, . . . , αn)) =
α1x1 + α2x2 + · · · + αnxn. Thenp g is a bijective linear transformation by the definition of a basis. Also, g is
bounded because letting c = kx1k2 + · · · + kxnk2 and applying Hölder’s inequality, we have
v v
u n u n
X
n X
n
uX uX
kg((α1 , α2, . . . , αn)) = α i xi ≤ |αi|kxik ≤ t |αi| t
2 kxik2 = ck(α1, α2, . . . , αn)k2.
i=1 i=1 i=1 i=1

Next, assume g−1 is not bounded. Then for all j ∈ N, there exists xj ∈ V such that kg−1 (xj )k > jkxj k.
Let uj = xj /kxj k. We have kuj k = 1 and kg−1(uj )k > j. Let βj = g−1 (uj )/kg−1 (uj )k ∈ Kn. Since kβj k = 1,
by compactnesss of unit sphere S in Kn , there exists a subsequence βjm converges to some β ∈ S. Then
kg(βjm )k = 1/kg−1(ujm )k < 1/jm implies g(β) = 0. Now g injective implies β = 0, contradicting β ∈ S.
Hence, g and g−1 are bounded. By the theorem in the last section, g is a uniform homeomorphism.
Let k · k0, k · k1 be two norms on V. From the uniform homeomorphisms g0 : (Kn , k · k2) → (V, k · k0) and
g1 : (Kn , k·k2) → (V, k·k1) as above, we see g1 ◦g0−1 : (V, k·k0) → (V, k·k1) and g0 ◦g1−1 : (V, k·k1) → (V, k·k0)
are identity maps and they are uniformly continuous. Therefore, the two norms are uniformly equivalent.
Finally, since completeness is preserved under uniform homeomorphism and Kn is complete, any finite
dimensional normed subspace is complete (hence closed) in any normed space.

Riesz’ Lemma. Let V be a normed vector space and W be a closed vector subspace of V with W 6= V. Then
for every ε > 0, there exists x ∈ V such that kxk = 1 and d(x, W ) = inf{kx − wk : w ∈ W } > 1 − ε.
Proof. It suffices to consider 0 < ε < 1. Let y ∈ V \ W. Since y 6∈ W = W , we get d(y, W ) > 0. Then
d(y, W ) d(y, W )
d(y, W ) < . By the infimum property, for some y0 ∈ W, we have 0 < d(y, W ) ≤ ky − y0 k < .
1−ε 1−ε
y − y0
In particular, y 6= y0 . Let x = , then kxk = 1. Next observe that T : W → W defined by T (w) = z,
ky − y0 k

27
 z − y0
6 0. In fact, T −1(z) =
where z = y0 + wky − y0 k, is a bijection since ky − y0 k = . Using this, we have
ky − y0 k
n y − y0 o 1 n o
d(x, W ) = inf −w : w ∈W = inf y − y0 − wky − y0 k : w ∈ W
ky − y0 k ky − y0 k
1 d(y, W )
= inf{ky − zk : z ∈ W } = > 1 − ε.
ky − y0 k ky − y0 k

Remark. For infinite dimensional space V, it may not be possible to have kxk = 1 and d(x, W ) = 1 in
general. See Robert Kuller’s book Topics in Modern Analysis, pp. 97-98, example 1.

Theorem. Let V be a normed vector space. We have dim V < ∞ if and only if B(0, 1) is compact in V.
Proof. Suppose m = dim V < ∞. Let {x1, x2, . . ., xm } be a basis of V. Recall the function g : Km → V
defined by g((α1 , α2, . . ., αm )) = α1x1 + α2x2 + · · ·+ αm xm is a linear homeomorphism. Now g−1 (B(0, 1)) is
closed and bounded in Km , hence it is compact by the Heine-Borel theorem. Since g is a homeomorphism,
B(0, 1) is compact in V.
Conversely, suppose B(0, 1) is compact. Let 0 < ε < 1. Since B(0, 1) is totally bounded, there exists
[
n
x1, x2, . . . , xn ∈ V such that B(0, 1) ⊆ B(xj , 1 − ε).
j=1

Let W = span{x1, x2, . . . , xn}, then dim W < ∞. We claim W = V. Assume W 6= V. By Riesz’ lemma,
there exists x ∈ V such that kxk = 1 and d(x, W ) > 1 − ε. For every j = 1, 2, . . ., n, xj ∈ W implies
[
n
d(x, xj ) ≥ d(x, W ) > 1 − ε. Then x ∈ B(0, 1) and x 6∈ B(xj , 1 − ε), a contradiction. Therefore, W = V
j=1
and dim V = dim W < ∞.

Theorem. Let V be a normed vector space. Let W be a finite dimensional vector subspace of V. If x 6∈ W,
then there exists w ∈ W such that d(x, W ) = kx − wk.
Proof. By the infimum limit theorem, there exists a sequence wn ∈ W such that d(x, W ) = lim kx − wn k.
n→∞
Then {x−wn } is a bounded sequence in W. So {wn} is a bounded sequence in W. Since m = dim W < ∞, W
is homeomorphic to Km . By the Bolzano-Weierstrass theorem, there exists a subsequence {wnk } converging
to some w ∈ W. Then d(x, W ) = lim kx − wnk k = kx − wk.
k→∞

§4. Completeness of Normed Spaces

X
∞ X
n
Definitions. In a normed space V, a series xn converges iff the partial sum sequence sn = xj converges
j=1 j=1
∞
X ∞
X
in V. xn converges absolutely iff kxj k converges in R.
j=1 j=1

Theorem. (Completeness Criterion) Let V be a normed space. V is complete if and only if every absolute
X
∞
convergent series xj converges in V.
j=1
n
X ∞
X n
X
Proof. Suppose V is complete. Let sn = xj . If kxj k < ∞, then tn = kxj k is Cauchy in R. Hence,
j=1 j=1 j=1
for every ε > 0, there exists K ∈ N such that n > m > K implies |tn − tm | < ε. Then
X
n X
n
ksn − sm k = xj ≤ kxj k = tn − tm < ε.
j=m+1 j=m+1

28
 X
∞
Hence, {sn } is a Cauchy sequence in V. Since V is complete, xj = lim sn converges in V.
n→∞
j=1

Conversely, let {xk} be a Cauchy sequence in V. To show {xk} converges in V, by step 4 of Cauchy’s
theorem, it is enough to show some subsequence {xkn } converges in V. For every positive integer j, consider
ε = 1/2j > 0. Since {xk } is Cauchy, there exists strictly increasing Kj ∈ N such that m, n ≥ Kj implies
X
∞ X∞
1
kxm −xnk < ε = 1/2j . Let y1 = xK1 and for j > 1, let yj = xKj −xKj−1 . Then kyj k ≤ ky1 k + < ∞.
2j
j=1 j=1
∞
X n
X ∞
X
So yj converges absolutely. Therefore, lim xKn = lim yj = yj exists in V.
n→∞ n→∞
j=1 j=1 j=1

Examples. (1) We will show B(M ) = {f : M → K : kfk∞ = sup{|f(x)| : x ∈ M } < ∞} is complete by

checking the completeness criterion.
X
∞ X
k
Suppose kfnk∞ < ∞, where fn ∈ B(M ). Then we need to show sk = fn converges in B(M ). Now
n=1 n=1
X
∞ X
∞ X
∞
for every x ∈ M, |fn(x)| ≤ kfn k∞ < ∞ and so by the absolute convergence test, fn (x) converges
n=1 n=1 n=1
∞
X ∞
X ∞
X ∞
X ∞
X
in K. Then for every x ∈ M, fn(x) ≤ |fn (x)| ≤ kfn k∞ < ∞. So fn ≤ kfn k∞ < ∞.
∞
n=1 n=1 n=1 n=1 n=1
X
∞
Then f = fn ∈ B(M ). Similarly, summing from n = k + 1 to infinity, we have
n=1

∞
X k
X ∞
X ∞
X
kf − sk k∞ = fn − fn = fn ≤ kfn k∞ → 0 as k → ∞.
∞ ∞
n=1 n=1 n=k+1 n=k+1

X
k
Therefore, sk = fn (x) converges to f in B(M ).
n=1

(2) For 1 ≤ p ≤ ∞ and measurable X ⊆ R, we will show Lp = {[f] : f measurable on X and kfkp < ∞} is
complete.
For p = ∞, completeness is similar to B(M ) (just change B(M ) to L∞ and “every” to “almost every”).
X
∞ X
k X
∞
p
For 1 ≤ p < ∞, suppose kfnkp < ∞, where fn ∈ L . We have gk = |fn| increases to g = |fn |.
n=1 n=1 n=1
Then gkp increases to gp . By the monotone convergence theorem and the triangle inequality in Lp ,
Z Z k
X p X
k p X
∞ p
p
g dm = lim gkp dm = lim kgk kpp = lim |fn | ≤ lim kfnkp = kfn kp < ∞.
X k→∞ X k→∞ k→∞ p k→∞
n=1 n=1 n=1

∞
X ∞
X
This implies gp ∈ L1 , g ∈ Lp and g = |fn| converges a.e. on X. By the absolute convergence test, fn
n=1 n=1
converges a.e. to some f on X. So kfkp ≤ kgkp < ∞, i.e. f ∈ Lp .
k
X
Next we will show sk = fn converges to f in Lp . These follow from observing that
n=1

∞
X p  X
∞ p
|f − sk |p = fn ≤ |fn | ≤ gp ∈ L1 .
n=k+1 n=k+1

29
By the Lebesgue dominated convergence theorem,
Z Z
p p
lim kf − sk kp = lim |f − sk | dm = lim |f − sk |p dm = 0.
k→∞ k→∞ X X k→∞

n X∞ 1/p o
(3) `p = (a1 , a2, a3, . . .) : k(a1, a2, a3, . . .)kp = |an|p < ∞ is complete. This follows by considering
n=1 Z
p p
` as L on N with the counting measure (that is, f : N → K corresponds to (f(1), f(2), f(3), . . .) and f dm
∞
N
X
corresponds to f(n).)
n=1

§5. Arzela-Ascoli’s Theorem

Let M be a compact metric space. Then C(M ) is complete. By the metric compactness theorem, a
subset of C(M ) is compact if and only if it is complete and totally bounded in C(M ). Since complete subsets
of C(M ) are closed and totally bounded subsets are bounded, so compact subsets in C(M ) are closed and
bounded. It is useful to have conditions (along with closed and boundedness) that imply compactness.

Definitions. Let S be a set of functions from a metric space M to K.

(1) Let x0 ∈ M. S is equicontinuous at x0 iff for every ε > 0, there exists δ > 0 such that for every x ∈ M,
d(x, x0) < δ implies for every f ∈ S, |f(x) − f(x0 )| < ε. S is equicontinuous iff it is equicontinuous at every
x0 ∈ M.
(2) S is uniformly equicontinuous iff for every ε > 0, there exists δ > 0 such that for every x, y ∈ M,
d(x, y) < δ implies for every f ∈ S, |f(x) − f(y)| < ε.
(3) S is pointwise bounded iff for every x ∈ M, {|f(x)| : f ∈ S} is bounded by a number Cx in R.
(4) S is uniformly bounded iff there exists a positive number C such that for every f ∈ S and x ∈ M,
|f(x)| ≤ C. (For S ⊆ B(M ) or C(M ), S is uniformly bounded iff S is a bounded subset of B(M ) or C(M ).)

Theorem. Let M be a compact metric space and S ⊆ C(M ). Then S is equicontinuous iff S is uniformly
equicontinuous.
Proof. The if direction is clear. For the only-if direction, for every ε > 0, since S is equicontinuous at
every x ∈ M, there exists δx > 0 such that d(x, y) < δx implies for all f ∈ S, |f(x) − f(y)| < ε/2. Since
n
M = ∪ {x} ⊆ ∪ B(x, δx /2), by compactness, there are t1 , t2, . . . , tn ∈ M such that M ⊆ ∪ B(ti , δti /2).
x∈M x∈M i=1

Let δ = min{δt1 /2, δt2 /2, . . . , δtn /2}. Let x, y ∈ M with d(x, y) < δ. Since x ∈ M, there exists B(ti , δti /2)
containing x. By the definition of δ and δt , we have d(x, ti) < δti /2 < δti and

d(x, y) < δ =⇒ d(y, ti ) ≤ d(y, x) + d(x, ti) < δ + δti /2 ≤ δti .

So for all f ∈ S, |f(x) − f(y)| ≤ |f(x) − f(ti )| + |f(ti ) − f(y)| < ε/2 + ε/2 = ε.

Arzela-Ascoli’s Theorem. Let M be compact and S ⊆ C(M ). The following are equivalent:
(a) S is compact in C(M );
(b) S is closed, (uniformly) bounded and uniformly equicontinuous;
(c) S is closed, pointwise bounded and equicontinuous.

Corollary. Let M be compact and S ⊆ C(M ). If S is pointwise bounded and equicontinuous, then every
sequence in S has a (uniformly) convergence subsequence with the limit in S.

30
Proof. From the sequential closure theorem, we see S pointwise bounded and equicontinuous imply S
pointwise bounded and equicontinuous. So we may apply Arzela-Ascoli’s theorem to S.

Since in analysis, we often study sequences of approximations to solve problems and compactness pro-
vides convergent subsequences, this theorem has many applications. We will mention two applications here.
In ordinary differential equations, the Peano existence theorem provides conditions for existence of local so-
lutions. In complex analysis, the Riemann mapping theorem asserts that for every simply connected region
U, there exists a bijective holomorphic function (in particular, a homeomorphism) from the open unit disk
D onto U. The common proofs of these theorems relied on Arzela-Ascoli’s theorem.

Proof of Arzela-Ascoli’s Theorem. (a)⇒(b) As mentioned above, S compact implies it is closed and
bounded in C(M ). By the previous theorem, it remains to prove that S is equicontinuous at every y ∈ M.
n
[
For every ε > 0, since S is totally bounded, there are f1 , f2 , . . . , fn ∈ C(M ) such that S ⊆ B(fi , ε/3).
i=1
Since every fi (i = 1, 2, . . ., n) is continuous at y, there exists δi > 0 such that for every x ∈ M, d(x, y) < δi
implies |fi (x) − fi (y)| < ε/3. Let δ = min{δ1 , δ2 , . . . , δn} > 0. For every f ∈ S, there exists some fi such that
f ∈ B(fi , ε/3), i.e. kf − fi k∞ < ε/3. Then for every x ∈ M, d(x, y) < δ implies d(x, y) < δi and
ε ε ε ε
|f(x) − f(y)| ≤ |f(x) − fi (x)| + |fi (x) − fi (y)| + |fi (y) − f(y)| < kf − fi k∞ + + kfi − fk∞ < + + = ε.
3 3 3 3

(b)⇒(c) Clear (as bounded implies pointwise bounded and uniform equicontinuous implies equicontinuous).
(c)⇒(a) We will show S is sequentially compact by proving
(i) every sequence {fn } in S has a subsequence converging pointwise on a countable dense subset E of M ;
(ii) the subsequence is a Cauchy sequence in C(M ) (since C(M ) is complete and S is closed, the subsequence
will have a limit in S).
For (i), M compact (hence totally bounded) implies M separable (i.e. M has a countable dense subset
E = {y1 , y2 , y3, . . .}.) Since S is pointwise bounded, for all yj ∈ E, {fn (yj )} is bounded in K. By the Bolzano-
Weierstrass theorem, for j = 1, there exists a subsequence {f1n} of {fn} such that f11(y1 ), f12(y1 ), f13(y1 ), . . .
converges. Similarly, for j = 2, there exists a subsequence {f2n } of {f1n } such that f21(y2 ), f22(y2 ), f23(y2 ), . . .
converges, etc. Consider fnk = fkk . Then fnk converges pointwise on E.
For (ii), by the previous theorem, S is uniformly equicontinuous. So for every ε > 0, there exists δ > 0
such that d(x, y) < δ implies (*) for all f ∈ S, |f(x) − f(y)| < ε/4. Since M is totally bounded, there are
m
[
x1, x2, . . . , xm ∈ M such that M = B(xi , δ/2). Since E is dense, there exists yi ∈ E ∩ B(xi , δ/2). By the
i=1
m
[
ball-in-ball lemma, B(yi , δ) ⊇ B(xi , δ/2) and so (**) M = B(yi , δ).
i=1

Since {fnk } converges at every yi ∈ E, so for i = 1, 2, . . ., m, {fnk (yi )} is a Cauchy sequence in K. Then
for every ε > 0, there exists N ∈ N such that p, q > N implies |fnp (yi ) − fnq (yi )| < ε/4 for i = 1, 2, . . ., m.
By (**), for every x ∈ M, there exists yi such that x ∈ B(yi , δ). So d(x, yi) < δ. By (*), since all fnk ∈ S, we
have |fnk (x) − fnk (yi )| < ε/4. Then p, q > N implies
ε ε ε
|fnp (x) − fnq (x)| ≤ |fnp (x) − fnp (yi )| + |fnp (yi ) − fnq (yi )| + |fnq (yi ) − fnq (x)| < + + < ε.
4 4 4
This implies kfnp − fnq k∞ ≤ 3ε/4 < ε. So {fnk } is a Cauchy sequence in C(M ).

§6. Stone-Weierstrass Theorem

In this section, we will exploit the algebraic structure (instead of the vector space structure) of C(M )
to identify some of its dense subalgebras.

31
Definitions. (1) An algebra A is a vector space with an multiplication from A × A to A assigning every pair
(x, y) ∈ A × A an element xy ∈ A such that
(a) for every x, y, z ∈ A, x(yz) = (xy)z;
(b) for every x, y, z ∈ A, x(y + z) = xy + xz and (x + y)z = xz + yz;
(c) for every c ∈ K and x, y ∈ A, c(xy) = (cx)y = x(cy).
If A is an algebra, B ⊆ A and B is an algebra with the same multiplication as A, then we will say B is
a subalgebra of A.
(2) An algebra A is commutative iff for every x, y ∈ A, xy = yx.
(3) A normed vector space is a normed algebra iff it is an algebra such that for every x, y ∈ A, kxyk ≤ kxkkyk.
(This inequality is motivated by the sup-norm as k cos x sin xk∞ = 1/2 ≤ k cos xk∞k sin xk∞ = 1.)
(4) A Banach algebra is a complete normed algebra.

Remarks. In a normed algebra, multiplication is continuous in the sense that xn → x and yn → y implies
xnyn → xy. This follows from xn → x implies kxnk bounded and

kxnyn − xyk = kxnyn − xny + xn y − xyk ≤ kxnyn − xnyk + kxny − xyk ≤ kxnkkyn − yk + kxn − xkkyk → 0.

Examples. If M is compact, then C(M ) is a Banach algebra. For every f, g ∈ C(M ), fg ∈ C(M ) is given by
(fg)(x) = f(x)g(x). The conditions (a),(b),(c) in the definitions above are easily checked. For the normed
inequality, we observe that for every f, g ∈ C(M ) and x ∈ M, we have |f(x)g(x)| ≤ kfk∞ kgk∞ and so
kfgk∞ ≤ kfk∞ kgk∞ .
If M is a compact subset of R, then the set of all polynomials on M is a normed subalgebra of C(M ).

In case K = R, we will write C(M, R) for C(M ) for emphasis. Similarly, in case K = C, we will write
C(M, C) for C(M ). Recall that the Weierstrass approximation theorem asserts that for every f ∈ C([0, 1], R)
and ε > 0, there exists a polynomial P (x) such that for all x ∈ [0, 1], |f(x) − P (x)| < ε. Equivalently, this as-
serts that the collection of all polynomials on [0, 1] is a dense normed subalgebra of C([0, 1], R). Marshall
Stone generalized this by providing conditions for a subalgebra of C(M ) to be dense.

Stone-Weierstrass Theorem (Real-Valued Case). Let M be a compact metric space and A be a subal-
gebra of C(M, R) such that
(a) for all distinct x, y ∈ M, there exists f ∈ A such that f(x) 6= f(y) (we say A separates points of M ) and
(b) for all x ∈ M, there exists g ∈ A such that g(x) 6= 0 (we say A vanishes at no points of M ).
Then for every ε > 0 and h ∈ C(M, R), there exists p ∈ A such that kh − pk∞ < ε. Equivalently, A is dense
in C(M, R).

Remarks. Let M = [a, b] and A be the set of all polynomials on [a, b]. Since P (x) = x − a + 1 separates
points of [a, b] and vanishes at no points of [a, b], we get the Weierstrass approximation theorem.
We will provide a proof of the Stone-Weierstrass theorem without using the Weierstrass approximation
theorem. To do this, we will introduce a lemma.

Lemma. For every a > 0, there is a sequence of polynomials Pn(x) converging uniformly on [−a, a] to |x|
and Pn(0) = 0.


Proof. By substituting t = x/a, we may assume a = 1. Let P0 (t) = 0 and Pn+1
(t) = Pn(t) + 12 t2 −
Pn2(t)
for n ∈ N. By induction, we get Pn(0) = 0. We have |t| − Pn+1(t) = |t| − Pn(t) 1 − 12 (|t| + Pn(t)) . Using
this, by induction, we have 0 ≤ Pn(t) ≤ Pn+1(t) ≤ |t| ≤ 1. Next,

|t| − Pn (t) Y |t| − Pk+1(t) n−1

n−1 Y |t| + Pk (t)   |t| n
= = 1− ≤ 1− .
|t| |t| − Pk (t) 2 2
k=0 k=0

32
Multiplying both sides by |t| and applying the AM -GM inequality, we have for all t ∈ [−1, 1],
 |t| n 2  n2 |t| + (1 − 12 |t|) + · · · + (1 − 12 |t|) n+1 2  n n+1 2
0 ≤ |t| − Pn(t) ≤ |t| 1 − ≤ = ≤ → 0.
2 n n+1 n n+1 n

Proof of the Stone-Weierstrass Theorem. We will divide the proof into six steps.
Step 1: (A is a subalgebra of C(M, R).)
If f, g ∈ A, then there exist fn , gn ∈ A such that fn → f and gn → g. Then fn ± gn → f ± g ∈ A,
cfn → cf ∈ A and fn gn → fg ∈ A.
Step 2: (f ∈ A implies |f| ∈ A.)
As f(M ) is compact, there exists a > 0 such that f(M ) ⊆ [−a, a]. Let Pn(x) be as in the lemma. Then
Pn ◦ f ∈ A and |f| − Pn ◦ f ∞ → 0 as n → ∞. This implies |f| ∈ A.
Step 3: (f, g ∈ A implies max(f, g), min(f, g) ∈ A.)
This follows from steps 1, 2 and the identities max(f, g) = (f +g+|f −g|)/2, min(f, g) = (f +g−|f −g|)/2. By
mathematical induction, we also have f1 , f2 , . . ., fn ∈ A implies max(f1 , f2 , . . . , fn), min(f1 , f2, . . . , fn) ∈ A.
Step 4: (For all a, b ∈ M and h ∈ C(M, R), there is gab ∈ A such that gab(a) = h(a) and gab (b) = h(b).)
If a = b, then by (b), there exists g ∈ A such that g(a) 6= 0 and we can take gaa (t) = h(a)g(t)/g(a).
If a 6= b, then by (a), there exists j ∈ A such that j(a) 6= j(b) and by (b), there exist k, m ∈ A such that
k(a) 6= 0 and m(b) 6= 0. Then u(t) = j(t)m(t) − j(a)m(t) and v(t) = j(t)k(t) − j(b)k(t) are in A such that
h(a) h(b)
u(a) = v(b) = 0 and u(b) 6= 0 and v(a) 6= 0. Now we can take gab(t) = v(t) + u(t) in A to get
v(a) u(b)
gab(a) = h(a) and gab(b) = h(b).
Step 5: (For every ε > 0, a ∈ M and h ∈ C(M, R), there exists ga ∈ A such that ga (a) = h(a) and for all
t ∈ M, ga(t) > h(t) − ε/2.)
For every b ∈ M and ε > 0, consider gab in step 4. We have (gab − h)(b) = 0 ∈ (−ε/2, +∞). Now Jb =
(gab − h)−1 ((−ε/2, +∞)) is open and contains b. So {Jb : b ∈ M } is an open cover of M. Since M is compact,
n
M = ∪ Jbi . Let ga = max(gab1 , gab2 , . . . , gabn ) ∈ A. Now ga (a) = max(h(a), h(a), . . . , h(a)) = h(a). For
i=1
every t ∈ M, t ∈ Jbi for some i. Then (gabi − h)(t) > −ε/2. So ga(t) ≥ gabi (t) > h(t) − ε/2.
Step 6: (For every ε > 0 and h ∈ C(M, R), there exists p ∈ A such that for all t ∈ M, |h(t) − p(t)| < ε.)
For every a ∈ M, consider ga ∈ A in step 5. We have (ga − h)(a) = 0 ∈ (−∞, ε/2). Now Ia =
(ga − h)−1 (−∞, ε/2) is open and contains a. So {Ia : a ∈ M } is an open cover of M. Since M is com-
m
pact, M = ∪ Iai . Let g = min(ga1 , ga2 , . . . , gam ) ∈ A. For every t ∈ M, t ∈ Iai for some i. Then
i=1
(gai − h)(t) < ε/2. So g(t) ≤ gai (t) < h(t) + ε/2. By step 5, for all j, gaj (t) > h(t) − ε/2. So g(t) =
min(ga1 (t), ga2 (t), . . . , gam (t)) > h(t) − ε/2. Hence, |h(t) − g(t)| < ε/2. Since g ∈ A, there exists p ∈ A such
that kg − pk∞ < ε/2. Then kh − pk∞ < ε.

Next, by adding one more condition, we can also have a complex-valued version of the Stone-Weierstrass
theorem.

Stone-Weierstrass Theorem (Complex-Valued Case). Let M be a compact metric space and A be a

subalgebra of C(M, C) such that
(a) for all distinct x, y ∈ M, there exists f ∈ A such that f(x) 6= f(y) (we say A separates points of M ),
(b) for all x ∈ M, there exists g ∈ A such that g(x) 6= 0 (we say A vanishes at no points of M ) and
(c) f ∈ A implies f ∈ A (we say A is self-adjoint).

33
Then for every ε > 0 and h ∈ C(M, C), there exists p ∈ A such that kh − pk∞ < ε. Equivalently, A is dense
in C(M, C).
Proof. Let A0 = {f ∈ A : f(M ) ⊆ R}. Then A0 is an algebra in C(M, R). For f ∈ A, we have Re f =
(f + f )/2, Im f = (f − f )/(2i) ∈ A0 . For distinct x, y ∈ M, by (a), there exists f ∈ A such that f(x) 6= f(y).
Then either Re f(x) 6= Re f(y) or Im f(x) 6= Im f(y). So A0 separates points of M. For every x ∈ M, by (b),
there exists g ∈ A such that g(x) 6= 0. Then either Re g(x) 6= 0 or Im g(x) 6= 0. So A0 vanishes at no points
of M. Therefore, A0 is dense in C(M, R).
For all ε > 0 and h ∈ C(M, C), since Re h, Im h ∈ C(M, R), there exist u, v ∈ A0 such that ku−Re hk∞ <
ε/2 and kv − Im hk∞ < ε/2. Then u + iv ∈ A and k(u + iv) − hk∞ ≤ ku − Re hk∞ + kv − Im hk∞ < ε.

Corollary 1. If M is a compact metric space, then C(M ) is separable.

Proof. M compact implies M separable, which implies M has a countable basis {On : n = 1, 2, 3, . . .}.
Let f0 (t) = 1 and fn(t) = d(t, M \ On), which are continuous. Note if x, y ∈ M and x 6= y, then there
exists On such that x ∈ On ⊆ B(x, 12 d(x, y)). We have fn (x) = d(x, M \ On) > 0 since x 6∈ M \ On and
fn (y) = 0 since y ∈ M \ On . Let A be the subalgebra generated by fn over K (i.e. A is the set of all finite
linear combinations of finite products of fn ’s with K coefficients). We see A separates points of M using
fn as above. A vanishes at no points of M because f0 ≡ 1 on M. Also, if K = C, then since fi = fi , A is
self-adjoint. By the Stone-Weierstrass theorem, A is dense in C(M ). Let A0 be the subset of A consisted of
all finite linear combinations of finite products of fn ’s with (Q + iQ) ∩ K coefficients, then A0 is countable
and dense in A, hence dense in C(M ). Therefore C(M ) is separable.

Corollary 2. If M is compact in Rn, then the set Q of all polynomials with rational coefficients in n-variables
nXk1 kn
X o
is dense in C(M, R). Here Q = ··· ai1 ...in xi11 · · · xinn : ai1 ...in ∈ Q, k1, . . . , kn = 0, 1, 2, . . . .
i1 =0 in =0

Proof. Let P be the set as defining Q, but with Q replaced by R. We easily check that P is an algebra. P
separates points of M because distinct points in M have at least one distinct coordinate xi . P vanishes at no
points of M because 1 ∈ P. By the Stone-Weierstrass theorem for real-valued case, P is dense in C(M, R).
Then Q is dense in P, hence dense in C(M, R).

§7. Hilbert Space

Definition. An inner product space V is a vector space with a function that assigns to every (x, y) ∈ V × V
an inner product hx, yi ∈ K such that for every x, y, z ∈ V and α ∈ K,
(a) hx, xi ≥ 0 (with hx, xi = 0 if and only if x = 0);
(b) hx, y + zi = hx, yi + hx, zi and hx + y, zi = hx, zi + hy, zi;
(c) hαx, yi = αhx, yi and hx, αyi = αhx, yi;
(d) hx, yi = hy, xi.

p
Remarks. For an inner product space V, we may define kvk = hv, vi for every v ∈ V. We will check this is
a norm on V and so V becomes a normed vector space. Condition (a) of norm follows from condition (a) of
inner product space. Condition (b) of norm follows from condition (c) of inner product space. For condition
(c) of norm, we first observe that taking α = 0 in condition (c) of inner product, we get h0, yi = 0 = hx, 0i.
Next we will need the Cauchy-Schwarz inequality below to prove the triangle inequality for norm.

Theorem. Let V be an inner product space.

(1) (Cauchy-Schwarz Inequality) For all x, y ∈ V, |hx, yi| ≤ kxkkyk. Equality holds iff x, y are linearly
dependent.
(2) (Triangle Inequality) For all x, y ∈ V, kx + yk ≤ kxk + kyk.

34
(3) If kxn − xk → 0 and kyn − yk → 0, then hxn, yn i → hx, yi.
(4) (Parallelogram Law) For all x, y ∈ V, kx + yk2 + kx − yk2 = 2(kxk2 + kyk2 ).
(5) (Polar Identity or Polarization Identity) In case K = C, for all x, y ∈ V,

1X n
3
1
hx, yi = (kx + yk2 + ikx + iyk2 − kx − yk2 − ikx − iyk2 ) = i kx + in yk2 .
4 4 n=0

hx, yi
Proof. (1) The case x = 0 or y = 0 is trivial. For x 6= 0 and y 6= 0, let t = − . By condition
kxkkyk
D x y x y E |hx, yi|2
(a) of inner product, +t , +t ≥ 0, which can be expanded to yield 1 − ≥ 0.
kxk kyk kxk kyk kxk2kyk2
x y
This is equivalent to the Cauchy-Schwarz inequality. Equality holds if and only if +t = 0, which is
kxk kyk
equivalent to x and y linearly dependent.
(2) For all x, y ∈ V,
kx + yk2 = hx + y, x + yi = hx, xi + hx, yi + hy, xi + hy, yi
= kxk2 + 2 Rehx, yi + kyk2
≤ kxk2 + 2kxkkyk + kyk2 = (kxk + kyk)2 .

(3) Since kxn − xk → 0, kxnk is bounded. Therefore,

|hxn, yn i − hx, yi| = |hxn, yni − hxn , yi + hxn, yi − hx, yi|

≤ |hxn, yn − yi| + |hxn − x, yi|
≤ kxnkkyn − yk + kxn − xkkyk → 0 as n → ∞.

(4) and (5) follows from using kx + cyk2 = hx + cy, x + cyi = hx, xi + 2 Re chx, yi + |c|2hy, yi.

Definitions. Let V be an inner product space.

(1) A set {xα} in V is orthogonal iff for every α 6= β, hxα, xβ i = 0. {xα} is orthonormal iff it is orthogonal
and for every α, kxαk = 1.
(2) Let S1 , S2 ⊆ V. S1 is orthogonal to S2 (denoted by S1 ⊥ S2 ) iff for every x ∈ S1 , y ∈ S2 , hx, yi = 0.

Theorem. (Properties of Inner Product Spaces) Let V be an inner product space.

(1) (Pythagoras’ Theorem) If {x1, x2, . . . , xn} is orthogonal in V, then

kx1 + x2 + · · · + xnk2 = kx1k2 + kx2k2 + · · · + kxnk2 .

(2) If {x1 , x2, . . . , xn} is orthogonal in V and all xi 6= 0, then {x1, x2, . . ., xn} is linearly independent.
(3) Let x ∈ V be orthogonal to a dense subset S of V. Then x = 0.
(4) (Best Approximation Inequality) Let S be orthonormal, then for every x ∈ V, x1 , x2, . . ., xn ∈ S and
c1 , c2, . . ., cn ∈ K,
Xn Xn
x− hx, xiixi ≤ x − ci xi ,
i=1 i=1

where equality holds if and only if ci = hx, xii for all i = 1, 2, . . ., n.

35
 X
(5) (Bessel’s Inequality) Let S = {xα} be orthonormal, then for every x ∈ V, |hx, xαi|2 ≤ kxk2, where
α

X nX
n o
|hx, xαi|2 = sup |hx, xii|2 : {x1, x2, . . ., xn} ⊆ S .
α i=1

(Remark. The inequality implies there are only countably many nonzero terms |hx, xαi| because for every
 1
positive integer k, Jk = xα : ≤ |hx, xαi|2 can only have at most kkxk2 elements.)
k
Proof. (1) We have
DX
n X
n E Xn X
n X
n
kx1 + x2 + · · · + xnk2 = xi , xj = hxi, xj i = hxi, xii = kx1k2 + kx2k2 + · · · + kxnk2 .
i=1 j=1 i=1 j=1 i=1

X
n
(2) If αi xi = 0, then for every j = 1, 2, . . ., n, we have
i=1

DX
n E Xn
0= αixi , xj = αi hxi, xj i = αj hxj , xj i =⇒ αj = 0.
i=1 i=1

(3) Since x ∈ V = S, x = lim yn for some yn ∈ S. Then x ⊥ S implies hx, yn i = 0 and so

n→∞

kxk2 = hx, xi = hx, lim yn i = lim hx, yni = 0 =⇒ x = 0.

n→∞ n→∞

(4) Let βi = hx, xii for i = 1, 2, . . ., n. Using |βi − ci|2 = |βi |2 − 2 Re ci βi + |ci|2 , we have
X
n 2 X
n X
n
x− ci xi = kxk2 − 2 Re cihxi , xi + |ci|2 kxik2
i=1 i=1 i=1
n
X n
X
= kxk2 − 2 Re ciβi + |ci |2
i=1 i=1
X
n X
n
= kxk2 − |βi |2 + |βi − ci |2
i=1 i=1
X
n X
n 2
≥ kxk2 − |βi |2 = x − βi x i .
i=1 i=1

(5) Let {x1, x2, . . . , xn} ⊆ S and βi = hx, xii for i = 1, 2, . . ., n. Setting ci = βi in the proof of (4), we have
Xn 2 Xn X n X
0≤ x− βi xi = kxk2 − |βi |2. So kxk2 ≥ sup |βi |2 = |hx, xαi|2.
i=1 i=1 i=1 α

Theorem (Closest Point Property). Let A be a nonempty complete convex subset of an inner product
space V. For every x ∈ V, there exists a unique y ∈ A such that d(x, A) = d(x, y).
Proof. By the infimum limit theorem, there exists yn ∈ A such that lim d(x, yn) = d(x, A) = inf kx − ak.
n→∞ a∈A
Since (yn + ym )/2 ∈ A, by the parallelogram law,
yn + ym 2
kyn − ym k2 = 2(kyn − xk2 + kym − xk2) − 4 −x
2
≤ 2(d(x, yn)2 + d(x, ym )2) − 4d(x, A)2 → 0.

36
This implies {yn} is Cauchy in A. So {yn } converges to some y ∈ A. Then d(x, A) = lim d(x, yn) = d(x, y).
n→∞

Suppose w ∈ A also satisfies d(x, A) = d(x, w). Then (y + w)/2 ∈ A implies kx − (y + w)/2k ≥ d(x, A).
By the parallelogram law,

kw − yk2 = 2(kx − wk2 + kx − yk2 ) − 4kx − (y + w)/2k2 ≤ 2(d(x, A)2 + d(x, A)2) − 4d(x, A)2 = 0

implies w = y.

Definition. Let V be an inner product space and S ⊆ V. The orthogonal complement of S is

S ⊥ = {x : x ∈ V, x ⊥ S}.

Observe that S ⊥ is a vector subspace of V and x ∈ S ∩ S ⊥ implies 0 = hx, xi = kxk2. So S ∩ S ⊥ ⊆ {0}.

Theorem. Let M be a vector subspace of an inner product space V and z ∈ V. Then z ∈ M ⊥ if and only if
for all m ∈ M, kz − mk ≥ kzk (i.e. 0 is the closest point on M to z).
Proof. If z ∈ M ⊥ and m ∈ M, then by the Pythagoras theorem, kz − mk2 = kzk2 + kmk2 ≥ kzk2.
Conversely, assume there exists m ∈ M such that hz, mi =
6 0. Let w = |hz, mi|/hz, mi. For all t > 0,
since twm ∈ M, so kz − twmk2 ≥ kzk2. Expanding and cancelling kzk2, we see −2 Re twhz, mi + t2 kmk2 =
−2t|hz, mi| + t2kmk2 ≥ 0. Cancelling 2t from both sides and letting t → 0+ , we get −|hz, mi| ≥ 0. Then
hz, mi = 0, a contradiction. So z ∈ M ⊥.

Projection Theorem. Let M be a complete vector subspace of an inner product space V. For every x ∈ V,
there exist unique y ∈ M and z ∈ M ⊥ such that x = y + z. Also, we have d(x, M ) = d(x, y) and d(x, M ⊥) =
d(x, z).
(Definitions. The function PM : V → M defined by PM (x) = y is called the orthogonal projection onto M .
Similarly, PM ⊥ : V → M ⊥ defined by PM ⊥ (x) = z is called the orthogonal projection onto M ⊥ .)
Proof. For uniqueness, suppose x = y + z = y0 + z 0, where y, y0 ∈ M and z, z 0 ∈ M ⊥ , then y − y0 =
z 0 − z ∈ M ∩ M ⊥ = {0}, i.e. y = y0 and z = z 0. For existence, by the closest point property, there is
a unique y ∈ M such that d(x, M ) = d(x, y). Let z = x − y. For every m ∈ M, since y + m ∈ M, so
kzk = kx − yk = d(x, M ) ≤ kx − (y + m)k = kz − mk. By the last theorem, z ∈ M ⊥ . For all w ∈ M ⊥ , we
have y ⊥ z − w. By the Pythagoras theorem, kx − wk2 = ky + z − wk2 = kyk2 + kz − wk2 ≥ kyk2 = kx − zk2 ,
which implies d(x, M ⊥) = d(x, z).
p
Definition. A Hilbert space is an inner product space, where the normed space defined by kvk = hv, vi
is complete. (Thus, all facts about metric spaces or normed vector spaces or inner product spaces may be
applied to a Hilbert space.)

Examples. (1) Kn = Rn or Cn with h(x1, x2, . . . , xn), (y1, y2 , . . . , yn)i = x1y1 + x2y2 + · · ·+ xn yn is a Hilbert
space.
(2) `2 with h(x1 , x2, x3, . . .), (y1 , y2, y3 , . . .)i = x1 y1 + x2 y2 + x3 y3 + · · · is a Hilbert space.
Z
(3) For a nonempty measurable subset X of R, L2 with hf, gi = fg dm is a Hilbert space.
X

Definitions. Let E = {xα} be an orthonormal set in an inner product space V .

(1) E is maximal in V iff every orthonormal set containing E equals E.
(2) E is complete in V iff for every x ∈ V, x ⊥ E implies x = 0.
nXn o
(3) E is total (or fundamental) in V iff span E = ci xi : n ∈ N, ci ∈ K, xi ∈ E is dense in V.
i=1

37
Remarks. (1) (Total implies maximal) If E = {xα} is a total orthonormal set in an inner product space V,
then it is maximal. To see this, let F be an orthonormal set containing E. Assume there exists x ∈ F \ E.
Then kxk = 1 and x ⊥ E. So x ⊥ span E and span E dense in V imply x = 0 by property (3) of inner
product space. This contradicts kxk = 1. So F = E, i.e. E is maximal.
(2) Let V be the set of all complex-valued continuous functions on the unit circle T. For f, g ∈ V, define
Z 2π
1
hf, gi = f(eiθ )g(eiθ ) dθ. Then V becomes an inner product space. The set E = {en : n ∈ Z, en (z) =
2π 0
z n } is orthonormal. Now z ∈ E separates points of T, 1 ∈ E vanishes at no point of T and span E
is a self-adjoint algebra. By the Stone-Weierstrass theorem, span E is dense in V (with sup norm). Since
Z 2π Z 2π
1 1
|f(eiθ ) − p(eiθ )|2 dθ ≤ kf − pk2∞ dθ = kf − pk2∞ , span E is dense in V (with norm from inner
2π 0 2π 0
product). Then E is total in V, but still V is not a Hilbert space (see section 1, example (6)).

Lemma. Let V be a Hilbert space and S = {x1, x2, x3, . . .} be orthogonal in V.

∞
X ∞
X
(1) xi converges in V if and only if kxik2 converges in R.
i=1 i=1

X
∞
(2) If S is orthonormal, then for every x ∈ V, hx, xiixi converges to some y ∈ V and y − x ⊥ S.
i=1

Proof. (1) For n > m, by Pythagoras’ theorem,

X
n X
m 2 X
n 2 X
n X
n X
m
xi − xi = xi = kxik2 = kxik2 − kxi k2.
i=1 i=1 i=m+1 i=m+1 i=1 i=1

X
n X
n
Therefore, xi is a Cauchy sequence in V if and only if kxik2 is a Cauchy sequence in R.
i=1 i=1

X
∞ X
∞ X
∞
(2) By Bessel’s inequality, khx, xiixi k2 = |hx, xii|2 ≤ kxk2 < ∞. By (1), hx, xiixi converges to
i=1 i=1 i=1
some y ∈ V. Next, y − x ⊥ S because for every xj ∈ S,
DX
∞ E X∞
hy − x, xj i = hx, xiixi − x, xj = hx, xiihxi , xj i − hx, xj i = hx, xj i − hx, xj i = 0.
i=1 i=1

Theorem. Let E = {xα} be orthonormal in an inner product space V. Among the following, (a) and (b) are
equivalent and also (c), (d), (e) are equivalent. If V is a Hilbert space, then all of them are equivalent.
(a) E is maximal;
(b) E is complete;
(c) E is total;
X
∞
(d) for every x ∈ V, x = hx, xαi ixαi for every ordering xα1 , xα2 , xα3 , . . . of the elements of J = {xα :
i=1
hx, xαi 6= 0};
X
∞
(e) (Parseval’s Equation) for every x ∈ V, kxk2 = |hx, xαi i|2 for every ordering xα1 , xα2 , xα3 , . . . of the
i=1
elements of J = {xα : hx, xαi =
6 0}.
Proof. (a)⇒(b) Let x ⊥ E. Assume x 6= 0. Then E ∪ {x/kxk} is an orthonormal set containing E. By (a),
this implies y = x/kxk ∈ E. Using x ⊥ E, we get 0 = hx, yi = kxk, contradicting x 6= 0.

38
(b)⇒(a) Assume there is an orthonormal set F ⊃ E. Taking x ∈ F \ E, we get x ⊥ E and (kxk = 1 implies)
x 6= 0, a contradiction to (b).
For the equivalence of (c), (d), (e), recall the remark on Bessel’s inequality that J = {xα : hx, xαi =
6 0}
is countable. Let xα1 , xα2 , xα3 , . . . be an arbitrary ordering of the elements of J.
k
X
(c)⇒(d) Let x ∈ V. We will show lim wk = 0, where wk = x − hx, xαi ixαi . By the best approximation
k→∞
i=1
kj
X
inequality, w1 , w2, w3, . . . is decreasing in R. By the density of span E, there exists cαi xαi ∈ span E with
i=1
kj kj
X 1 X
cαi ∈ K satisfying x − cαi xαi < . By the best approximation inequality, wkj = x − hx, xαi ixαi
j
i=1 i=1
kj
X 1
≤ x− cαi xαi < . Then lim wk = inf{wk : k = 1, 2, 3, . . .} = 0.
j k→∞
i=1

(d)⇒(e) From the proof of Bessel’s inequality and (d), for every x ∈ V,
n
X n
X ∞
X
2
kxk2 − |hx, xαi i|2 = x − hx, xαi ixαi →0 =⇒ kxk2 = |hx, xαi i|2 .
i=1 i=1 i=1

(e)⇒(c) From the proof of Bessel’s inequality and (e), for every x ∈ V,

X
n 2 X
n
x− hx, xαi ixαi = kxk2 − |hx, xαi i|2 → 0.
i=1 i=1

n
X
Since hx, xαi ixαi ∈ span E, so span E is dense in V.
i=1

Finally, by remark (1) above, (c) implies (a). If V is a Hilbert space, then it is enough to conclude by
showing
X
∞
(b)⇒(d) By the lemma, hx, xαi ixαi = y for some y ∈ V and y − x ⊥ E. By (b), y − x = 0. So
i=1
∞
X
hx, xαi ixαi = x.
i=1

Definition. Any orthonormal set E = {xα } satisfying any of the condition (a) to (e) in the above theorem
in a Hilbert space V is an orthonormal basis for V.
Caution: The usage of basis here is different from the usage of basis in vector spaces! Here, an orthonormal
basis expands an element into a sum of possibly infinitely many terms, while a basis in a vector space expands
an element into a finite linear combination of vectors in the basis.

Theorem. Every inner product space V has a maximal orthonormal set. In particular, if V is a Hilbert
space, then V has an orthonormal basis.
Proof. Order the collection X of orthonormal sets in V by inclusion. For every chain in X, the union of all
elements of the chain is an orthonormal set. So this union is in X and is an upper bound of the chain. By
Zorn’s lemma, there is a maximal element in X, which is a maximal orthonormal set in V.

Theorem. A Hilbert space V is separable iff it has a countable orthonormal basis. In that case, every
orthonormal basis for such V is countable.

39
Proof. If {xn} is a countable dense set in V, by discarding recursively any xn lying in span{x1, x2, . . ., xn−1},
we get a linearly independent sequence {yn } whose span is dense in V. Applying the Gram-Schmidt process
to {yn}, we get an orthonormal sequence {zn }, whose span is dense in V. Hence, {zn } is total in V. Then
{zn } is a countable orthonormal basis for V.
Conversely, from a countable orthonormal basis {zn }, form the set of all finite linear combinations of
{zn } with coefficients in (Q + iQ) ∩ K. Then it is a countable dense set in V. So V is separable.
Moreover, if {wα : α ∈ A} is another orthonormal basis of V, then for every n = 1, 2, 3, . . ., the set
An = {α ∈ A : hzn , wαi = 6 0} is countable by the remarks to Bessel’s inequality. Since {zn} is complete,
for every wα , there exists zn such that hzn , wαi =
6 0, i.e. α ∈ An. This implies A = A1 ∪ A2 ∪ A3 ∪ · · · is
countable.

Definitions. Let V1 , V2 be Hilbert spaces. A unitary transformation T : V1 → V2 is a bijective linear

transformation such that hT (u), T (v)i = hu, vi for all u, v ∈ V1 . V1 and V2 are isomorphic Hilbert spaces iff
there exists a unitary transformation T : V1 → V2 . (In such case, by switching the names of the corresponding
elements, we may regard V1 and V2 as the “same” space.)

Theorem. Every separable Hilbert space V is isomorphic to `2 .

X
∞
Proof. Let {xn} be a countable orthonormal basis of V. For every x ∈ V, x = hx, xnixn. Define T : V → `2
n=1
by T (x) = (hx, x1i, hx, x2i, hx, x3i, . . .). Clearly, T is a linear transformation. T is bijective because T has the
X∞ X
∞ X∞
inverse T −1 : `2 → V defined by T −1(c1 , c2, c3, . . .) = ci xi ∈ V as 2
kcixi k = |ci |2 < ∞. Finally,
i=1 i=1 i=1
for every u, v ∈ V,
DX
∞ X
∞ E X∞ X
∞ X
∞
hu, vi = hu, xiixi , hv, xj ixj = hu, xiihv, xj ihxi , xj i = hu, xiihv, xi i = hT (u), T (v)i.
i=1 j=1 i=1 j=1 i=1

§8. Dual Spaces

Let X and Y be normed linear spaces. Let L(X, Y ) denote the set of all bounded (equivalently, contin-
uous) linear transformations from X to Y. In the case Y = X, we simply write L(X) for this set and refer
to its elements as bounded linear operators on X. In the case Y = K, we write X ∗ for L(X, K) and refer to
its elements as bounded linear functionals on X. X ∗ is called the dual space or conjugate space of X.

Now we can assign a norm to every T ∈ L(X, Y ) as follow. First k0k = 0. Next, for T 6= 0, since
T is bounded, there exists a nonnegative real number c such that for all x ∈ X, kT xk ≤ ckxk. So x 6= 0,
kT xk/kxk ≤ c. We define kT k to be the least such c, which is sup{kT xk/kxk : x 6= 0}. Substituting
w = x/kxk, it is also sup{kT wk : kwk = 1}, which is often written as sup kT wk. In particular, for all
kwk=1
x ∈ X, kT xk ≤ kT kkxk. We can check this definition of kT k satisfies all the required properties of a norm.
So L(X, Y ) is a normed linear space.

Theorem. If X is a normed linear space and Y is a Banach space, then L(X, Y ) is a Banach space. In
particular, X ∗ is a Banach space since Y = K is complete.

Question: Why are we interested in dual spaces?

There are many reasons. Here are two reasons.
(1) In ODE, often we are interested in knowing if an equation like

dn y dn−1y
+ p(x) + · · · + q(x)y = g(x)
dxn dxn−1
40
(with p(x), . . . , q(x) infinitely differentiable) always has a solution for every infinitely differentiable g(x).
Define
dnf dn−1f
T (f) = + p(x) + · · · + q(x)f
dxn dxn−1
for infinitely differentiable f. To know if the equation above always has a solution is the same as to know if
the linear transformation T is surjective. In linear algebra, for a finite dimensional vector space V, a linear
transformation T : V → V is surjective if and only if T is injective. In general, to check injectivity is much
simpler than to check surjectivity. There is a similar result in functional analysis for infinite dimensional
Banach spaces X and Y. It asserts that for every T ∈ L(X, Y ), there exists T ∗ ∈ L(Y ∗ , X ∗ ) such that T is
surjective if and only if T ∗ is bounded below (in the sense that there exists δ > 0 such that for all v ∈ Y ∗ ,
kT ∗ vk ≥ δkvk). Now checking the bounded below condition involves manipulating inequalities, which is still
easier than checking surjectivity. For those who are interested, this result can be found in Walter Rudin’s
book Functional Analysis, 2nd ed., p. 100.
(2) For normed linear spaces, we would like to see they can be completed to Banach spaces. Dual spaces are
the key tool for producing completions. Here is a theorem for that purpose.

Theorem. For every normed vector space V. Define i : V → V ∗∗ = (V ∗ )∗ by i(v)(f) = f(v) for all f ∈ V ∗ .
(i is often called the canonical embedding of V into V ∗∗ .) Then i is a linear transformation and ki(v)k = kvk
for all v ∈ V , which implies i is injective. Since V ∗∗ is complete, so i(V ) is complete. If we identify v ∈ V
with i(v), then i(V ) is a completion of V.

Next, every inner product space can also be completed to a Hilbert space. In presntation project XV,
there is a proof of the Jordan-von Neumann theorem that a Banach space is a Hilbert space if and only
if it satisifies the parallelogram law. This implies every inner product space V has a completion. To see
this, from the inner product space V, we can take the induced norm to make V a normed space. Taking the
completion Vb of this normed space, we get a Banach space. Since the norm in V satisfies the parallelogram
law, the embedding of V into Vb is dense and ki(v)k = kvk, the norm of Vb will also satisfy the parallelogram
law. Hence the norm of Vb is induced from an inner product. Therefore, Vb is a Hilbert space completion of
V and uniqueness follows from uniqueness of norm space completion.
Following the theorem, we will introduce some related terms.

Definitions. (1) Let V be a vector space. Vtwin is the vector space having the same elements and the same
addition as V, but the scalar multiplication of every c ∈ K with v ∈ Vtwin is assigned to the element cv as
defined in V. For normed spaces, Vtwin is assigned the same norm as V. In case K = R, Vtwin = V.
(2) An isomorphism between vector spaces V and W is a bijective linear transformation T : V → W. V
is isomorphic to W iff there exists an isomorphism between V and W. V is anti-isomorphic to W iff there
exists an isomorphism between V and Wtwin. An isomorphism between normed spaces V and W is a linear
transformation T : V → W that is also a homeomorphism.
(3) An isometric isomorphism between normed spaces V and W is a bijective linear transformation T : V →
W such that for all v ∈ V, kT (v)k = kvk. (It is clearly a homeomorphism.) Two normed spaces V and W are
isometric isomorphic iff there exists an isometric isomorphism between them. In such case, we write V ∼
= W.
(In identifying dual spaces, this is often abused and is written simply as V = W since changing names of
the corresponding elements will give the “same” space).
(4) A Banach space V is reflexive iff i(V ) = V ∗∗ , where i is the canonical embedding in the theorem above.

We begin by identifying the dual spaces of some familiar normed linear spaces.
1 1
Examples. (1) For 1 ≤ p ≤ ∞, (Kn, k · kp )∗ = (Kn , k · kq )twin, where + = 1. There is an isometric iso-
p q
X
n
morphism sending (y1 , y2 , . . ., yn ) ∈ (Kn , k · kq )twin to T(y1 ,y2 ,...,yn ) (x1 , x2, . . ., xn) = xiyi ∈ (Kn, k · kp )∗ .
i=1
In particular, k(y1 , y2, . . . , yn )kq = kT(y1 ,y2 ,...,yn ) k. Then, (Kn , k · kp ) is reflexive.

41
 1 1
(2) For 1 ≤ p < ∞, (`p , k·kp)∗ = (`q , k·kq )twin and (Lp , k·kp)∗ = (Lq , k·kq )twin, where + = 1. There is an
p q
X∞
isometric isomorphism sending (y1 , y2, . . .) ∈ (`q , k · kq )twin to T(y1 ,y2 ,...) (x1, x2, . . .) = xi yi ∈ (`p , k · kp )∗ .
i=1
In particular, k(y1 , y2, . . .)kq = kT(y1 ,y2 ,...) k.
Z
Also, there is an isometric isomorphism sending g ∈ Lqtwin to Tg (f) = fg dm ∈ (Lp )∗ . In particular,
[a,b]
kgkq = kTg k. Therefore, for 1 ≤ p < ∞, (`p , k · kp ) and (Lp , k · kp ) are reflexive.

Next, we will identify the dual of a Hilbert space.

Riesz Representation Theorem for Hilbert Space. Let V be a Hilbert space. Then V ∗ = Vtwin due to
(a) for every y ∈ V, Ty : V → K defined by Ty (x) = hx, yi is in V ∗ and kTy k = kyk,
(b) if T ∈ V ∗ , then there exists a unique w ∈ V such that T = Tw .
(Remarks. This means there is a bijection between V and V ∗ so that w ↔ Tw , kwk = kTw k, w + w0 ↔
Tw+w0 = Tw + Tw0 , but cw ↔ Tcw = cTw . So V ∗ = Vtwin. Then V ∗∗ = V and hence every Hilbert space is
reflexive.)
Proof. (a) Clearly Ty is linear. Since |Ty (x)| = |hx, yi| ≤ kykkxk, Ty is bounded. So Ty ∈ V ∗ . Now
kTy k = sup{kTy (x)k : ∀x ∈ V, kxk ≤ 1} ≤ kyk. Conversely, kyk2 = hy, yi = |Ty (y)| ≤ kTy kkyk, which implies
kyk ≤ kTy k. Therefore, kTy k = kyk.
(b) We check uniqueness of w first. Suppose Tw = Tw0 . Then

0 = Tw (w − w0 ) − Tw0 (w − w0) = hw − w0 , wi − hw − w0 , w0i = hw − w0, w − w0i = kw − w0 k2 ⇒ w = w0.

Next we show existence of w. If T = 0, then take w = 0. Otherwise, let M = {x ∈ V : T x = 0} = ker T. Then

M = T −1 ({0}) is closed, hence complete in V. Also, T 6= 0 implies M 6= V. So there exists u ∈ V \ M. By the
projection theorem, u = y + z for some y ∈ M and z ∈ M ⊥ . Since u ∈ V \ M, z 6= 0. Let w = T (z)z/hz, zi.
We now show T = Tw . For every x ∈ V, let r = T (z)x − T (x)z, where T (x), T (z) ∈ K = R or C. Then
T (r) = T (z)T (x) − T (x)T (z) = 0 implies r ∈ M. Since z ∈ M ⊥ ,

0 = hr, zi = hT (z)x − T (x)z, zi = T (z)hx, zi − T (x)hz, zi.

T (z) D T (z) E
Solving for T (x), we get T (x) = hx, zi = x, z = hx, wi. Therefore, T = Tw .
hz, zi hz, zi

Remark. Riesz identified the dual spaces of many familar normed linear spaces. Hence, there are many Riesz
representation theorems. See Marian Fabian’s bookFunctional Analysis and Infinite-Dimensional Geometry,
pp. 44-49.

Appendix

In this section, we will identify the dual space of C[0, 1]. First we need to introduce the following
terminologies.

Definitions. For every function f : [a, b] → K, if a = b, then define Vab (f) = 0 and if b > a, then define
nX
n o
Vab (f) = sup |f(ti ) − f(ti−1 )| : a = t0 < t1 < · · · < tn = b ,
i=1

which is called the total variation of f on [a, b]. If Vab (f) < ∞, then f is of bounded variation on [a, b]. Finally,
let BV [a, b] denote the set of all functions f : [a, b] → K such that Vab (f) < ∞.

42
Remarks. (1) For continuous f : [a, b] → R, if f(t) denotes the position of a particle moving on the x-axis
at time t ∈ [a, b], then Vab(f) gives the total distance the particle travelled from time t = a to t = b.
(2) For continuous f : [a, b] → C, if f(t) = x(t) + iy(t), then Vab(f) gives the arc length of the path
{(x(t), y(t)) : t ∈ [a, b]} on R2 .

Examples. (1) If f : [a, b] → R is monotone, then Vab (f) = |f(b) − f(a)| < ∞ and so f ∈ BV [a, b].
(2) If f : [a, b] → R is continuous and has a bounded derivative on (a, b), then by the mean value theorem,

X
n X
n X
n
|f(ti ) − f(ti−1 )| = |f 0 (ci )|(ti − ti−1) ≤ kf 0 k∞ (ti − ti−1 ) = kf 0 k∞ (b − a).
i=1 i=1 i=1

Taking supremum, we get Vab(f) ≤ kf 0 k∞ (b − a) and so f ∈ BV [a, b].

(3) There exists continuous function f : [0, 1] → R not in BV [a, b]. For example, we can define f(0) = 0 and
π 1 1 1
for 0 < x ≤ 1, define f(x) = x cos . By considering 0 < < < · · · < < 1, it follows by a simple
2x 2n 2n − 1 2
1 1
computation that V01(f) ≥ 1 + + · · · + for every positive integer n. Therefore, V01 (f) = ∞.
2 n

Basic Facts. (1) Vab(f) = 0 if and only if f is a constant function on [a, b].
(2) If f ∈ BV [a, b], then f is bounded on [a, b] because for a < x ≤ b, |f(x)| ≤ |f(a)| + |f(x) − f(a)| ≤
|f(a)| + |f(a) − f(x)| + |f(x) − f(b)| ≤ |f(a)| + Vab (f) < ∞.
(3) Vab (f + g) ≤ Vab (f) + Vab(g) and Vab (cf) = |c|Vab(f) for c ∈ K. So BV [a, b] is a vector space.
X∞ X
∞  X ∞
b
(4) If fi converges pointwise on [a, b], then Va fi ≤ Vab(fi ). This follows from taking supremum
i=1 i=1 i=1
over all a = t0 < t1 < · · · < tn = b of the estimate
n X
X ∞ ∞
X n X
X ∞ ∞ X
X n ∞
X
fi (tj ) − fi (tj−1) ≤ |fi (tj ) − fi (tj−1)| = |fi (tj ) − fi (tj−1 )| ≤ Vab (fi ).
j=1 i=1 i=1 j=1 i=1 i=1 j=1 i=1

(5) Vab(|f|) ≤ Vab (f). Vab (fg) ≤ kfk∞ Vab(g) + Vab(f)kgk∞ . Hence, we see BV [a, b] is an algebra as f, g ∈
BV [a, b] implies fg ∈ BV [a, b].

Remarks. For all f ∈ BV [a, b], defining kfk = |f(a)| + Vab(f) will turn BV [a, b] into a normed vector space
since kfk = 0 implies f(a) = 0 = Vab(f), which by (1) implies f = 0 everywhere and triangle inequality
follows from

kf + gk = |f(a) + g(a)| + Vab (f + g) ≤ |f(a)| + |g(a)| + Vab (f) + Vab (g) = kfk + kgk.

X
∞ X
∞ X
∞
Using basic fact (4), if fi converges pointwise on [a, b], then fi ≤ kfi k.
i=1 i=1 i=1

For the rest of this section, we will focus on real-valued functions, i.e. K = R, to see BV [a, b] is a
Banach algebra and study C[0, 1]∗. The inequality kfgk ≤ kfkkgk requires a canonical (i.e. God-given)
decomposition of functions in BV [a, b] into a difference of two increasing functions.

Definition. For f : [a, b] → R and x ∈ [a, b], define Tf (x) = Vax (f), which is called the total variation of f.
(Clearly, if f ∈ BV [a, b], then Tf (x) is bounded and increasing.)

Lemma. For f ∈ BV [a, b], Tf + f and Tf − f are increasing functions on [a, b].

43
Proof. Let a ≤ x ≤ y ≤ b and ε > 0. By the supremum property, there exist a = t0 < t1 < · · · < tn = x
X
n


such that |f(ti ) − f(ti−1 )| ≥ Tf (x) − ε. Since |f(y) − f(x)| ± f(y) − f(x) ≥ 0, we have
i=1

X
n 

Tf (y) ± f(y) ≥ |f(ti ) − f(ti−1 )| + |f(y) − f(x)| ± f(y) − f(x) ± f(x) ≥ Tf (x) − ε ± f(x).
i=1

Letting ε tend to 0, we get Tf (y) ± f(y) ≥ Tf (x) ± f(x).

Theorem. If f ∈ BV [a, b], then f can be represented as the difference of two bounded increasing functions,
for example f = Tf − (Tf − f).

Remark. The converse of the theorem is true by example 1, basic facts (2) and (3).

Corollary. Every f ∈ BV [a, b] is Riemann integrable on [a, b], must have left and right limit at every
x ∈ [a, b] and has countably many discontinuities only.

Remarks. (1) In Tom Apostol’s book Mathematical Analysis, 2nd edition, pp. 132-3, it is showed that for
all f ∈ BV [a, b], f is continuous at x ∈ [a, b] if and only if Tf is continuous at x ∈ [a, b]. So every continuous
f in BV [a, b] is the difference of two continuous bounded increasing functions.
1
(2) Let f : [a, b] → R. Define Pax(f) = (Tf (x) + f(x) − f(a)). This measures the positive variation of f.
2
To see the meaning of this, let g = max(g, 0) and g− = − min(g, 0). Then g = max(g, 0) + min(g, 0), |g| =
+

max(g, 0)−min(g, 0) and so |g|+g = 2 max(g, 0) = 2g+ and |g|−g = −2 min(g, 0) = 2g− . Now Tf (x) = Vax (f).
X n
For a = t0 < t1 < · · · < tn = x, we have f(x) − f(a) = (f(ti ) − f(ti−1 )). So
i=1

1 nXn Xn o
Pax(f) = sup |f(ti ) − f(ti−1 )| + (f(ti ) − f(ti−1 )) : a = t0 < t1 < · · · < tn = x
2 i=1 i=1
nX
n o
= sup (f(ti ) − f(ti−1 ))+ : a = t0 < t1 < · · · < tn = x .
i=1

1
Similarly, Nax (f) = (Tf (x) − f(x) + f(a)) measures the negative variation of f. In terms of the positive
2
and negative variations of f, we have Vax (f) = Tf (x) = Pax (f) + Nax (f). Also, f(x) = f(a) + Pax(f) − Nax(f),
which is called the Jordan decomposition of f.

Theorem. For f, g ∈ BV [a, b], we have kfgk ≤ kfkkgk. So BV [a, b] is a normed algebra.
Proof. We have f(x) = f(a) + Pax(f) − Nax (f) and g(x) = g(a) + Pax(g) − Nax (g). So

(fg)(x) = f(a)g(a) + f(a)Pax (g) + g(a)Pax (f) − f(a)Nax (g) − g(a)Nax (f)
+ Pax(f)Pax (g) − Pax(f)Nax (g) − Pax (g)Nax (f) + Nax(f)Nax (g).

By basic fact (3), Vab(fg) is at most the sum of the total variations of the terms on the right. Since the
terms are all constant or monotone, we have

Vab (fg) ≤ 0 + |f(a)|Pab(g) + |g(a)|Pab(f) + |f(a)|Nab(g) + |g(a)|Nab(f)

+ Pab(f)Pab (g) + Pab(f)Nab (g) + Pab(g)Nab (f) + Nab (f)Nab (g).

Adding |f(a)g(a)| to both sides, then factor the right side, we get





kfgk ≤ |f(a)| + Pab(f) + Nab(f) |g(a)| + Pab(g) + Nab (g) = |f(a)| + Vab (f) |g(a)| + Vab(g) = kfkkgk.

44
Next, for the set of all functions of bounded variations, we will introduce a normalization procedure for them.

Definition. We say f ∈ BV [a, b] is normalized iff f(a) = 0 and f(x) = f(x+), where f(x+) = lim f(t) for
t→x+
all x ∈ (a, b). The set of all normalized functions of bounded variation will be denoted by N BV [a, b]. For
f ∈ N BV [a, b], kfk = Vab (f).

Remark. Following page 274 of Robert Kuller’s book Topics in Modern Analysis, let

N = {h ∈ BV [a, b] : ∀s ∈ [a, b), t ∈ (a, b], h(a) = h(s+) = h(t−) = h(b)}.

Let R = {(f, g) : f, g ∈ BV [a, b], f − g ∈ N }. Then R is easily checked to be an equivalence relation. In

every equivalence class [f], there is a unique fb ∈ N BV [a, b] given by

0 if t = a
fb(t) = f(t+) − f(a) if a < t < b .

f(b) − f(a) if t = b

Then 
 f(a) if t = a
f(t) − fb(t) = f(t) − f(t+) + f(a) if a < t < b ∈N

f(a) if t = b

as (1) f(t) − fb(t) ∈ BV [a, b] implies left limit exists at every t ∈ (a, b] and right limit exists at every s ∈ [a, b)
and (2) f is continuous except on a countable set and f(x) − fb(x) = f(a) at those x, where f is continuous.

Theorem. BV [a, b] is a Banach algebra. Similarly, N BV [a, b] is a Banach algebra.

X
∞ X
∞
Proof. Let fn be absolutely convergent in BV [a, b] (i.e. kfn k < ∞). For all t ∈ [a, b], |fn (t)| ≤
n=1 n=1
∞
X ∞
X ∞
X
|fn(a)| + |fn(t) − fn (a)| ≤ |fn(a)| + Vab (fn ) = kfnk. So |fn (t)| ≤ kfn k < ∞. Hence, g(t) = fn (t)
n=1 n=1 n=1
converges absolutely and uniformly on [a, b] (and for N BV [a, b], uniform convergence implies g(t) = g(t+)
for all t ∈ (a, b)). We claim g ∈ BV [a, b]. To see this, observe that by basic fact (4),

X
∞ X
∞
kgk = |g(a)| + Vab (g) ≤ (|fn (a)| + Vab (fn )) = kfnk < ∞.
n=1 n=1

X
N X
∞ X
∞ X
∞
Next, g − fn = fn ≤ kfn k → 0 as N → ∞. So fn converges to g in BV [a, b].
n=1 n=N +1 n=N +1 n=1

Remark. On page 214 of Kuller’s book, Theorem 5.9.5 and its corollary asserted that functions in BV [a, b]
are differentiable almost everywhere and the derivatives of functions in BV [a, b] are Lebesgue integrable.

Next, we will identify the dual space of C[0, 1].

Riesz-Markov Theorem. C[0, 1]∗ is isometric isomorphic to N BV [0, 1].

We will give a sketch of the isometric isomorphism. First we extend the Riemann integral to the
Riemann-Stieljes integral.
Let f be bounded and g be increasing on [0, 1]. Consider P = {0 = t0 < t1 < · · · < tn = 1}, a
partition of [0, 1]. For i = 1, 2, . . ., n, let mi = inf{f(x) : ti−1 ≤ x ≤ ti} and Mi = sup{f(x) : ti−1 ≤ x ≤ ti }.

45
 X
n X
n
Define the lower sum L(f, P, g) = mi (g(ti ) − g(ti−1 )), the upper sum U (f, P, g) = Mi (g(ti ) − g(ti−1 )),
i=1 i=1
the lower Riemann-Stieljes integral
Z 1
(L) f dg = sup{L(f, P, g) : P is a partition of [0, 1]}
0

and the upper Riemann-Stieljes integral

Z 1
(U ) f dg = inf{U (f, P, g) : P is a partition of [0, 1]}.
0
Z 1 Z 1 Z 1
If (L) f dg = (U ) f dg, then we write f dg for this common number and say f is Riemann-Stieljes
0 0 0
integrable on [0, 1] with respect to g, which we denote by f ∈ R(g). As in Riemann integration, if f is continu-
ous on [0, 1] (and g bounded increasing on [0, 1]), then f ∈ R(g). The proof involved a simple modification of
the similar theorem for the Riemann integral (see Walter Rudin’s book Principles of Mathematical Analysis,
3rd ed., pp. 125-6).
For g ∈ N BV [0, 1], we can take the Jordan decomposition of g = g1 − g2 and define
Z 1 Z 1 Z 1
f dg = f dg1 − f dg2.
0 0 0
Z 1
Next, we define Tg : C[0, 1] → R by Tg (f) = f dg. We can check Tg ∈ C[0, 1]∗ and kTg k = kgk.
0

Conversely, every T ∈ C[0, 1]∗ is of the form Tg for some unique g ∈ N BV [0, 1]. To get g, we recall the
Bernstein polynomials Pn of f is
Xn  k n
Pn (x) = f xk (1 − x)n−k .
n k
k=0

In the proof of the Weierstrass approximation theorem, we learned Pn converging uniformly to f on [0, 1].
n 
For k = 0, 1, 2, . . ., n, let αn,k = T xk (1 − x)n−k .
k
n 1 3 2n − 1 o
Consider 0 = t0 < t1 = < t2 = < · · · < tn = < tn+1 = 1 . Define
2n 2n 2n

0 if x ∈ [t0, t1)
gn(x) = αn,0 + αn,1 + · · · + αn,k−1 if x ∈ [tk , tk+1), where k ∈ {1, 2, . . ., n}

αn,0 + αn,1 + · · · + αn,n if x = 1.

Then gn ∈ N BV [0, 1] and for k = 0, 1, 2, . . ., n, gn(tk+1 ) − gn(tk ) = αn,k. Let sn,k = ±1 so to satisfy
X
n   Xn  
n k n−k n k
sn,k αn,k = |αn,k|. On [0, 1], sn,k x (1 − x) ≤ x (1 − x)n−k = (x + (1 − x))n = 1 and
k k
k=0 k=0

n
X n
X
|gn(x)| ≤ V01 (gn ) = |αn,k| = sn,k αn,k
k=0 k=0
n
X n  Xn  
n k 
= sn,k T xk (1 − x)n−k = T sn,k x (1 − x)n−k
k k
k=0 k=0
X
n  
n k
≤ kT k sn,k x (1 − x)n−k ≤ kT k.
k ∞
k=0

46
Next we would like to apply the following theorem.

Helly Selection Theorem. Let gn(t) be a sequence in BV [a, b] with all kgnk∞ and Vab (gn ) bounded by M.
Then some subsequence gni (t) converges pointwise on [a, b] to a function g(t) ∈ BV [a, b] with kg||∞ and
Z b Z b
b
Va (g) bounded by M and for every f(t) ∈ C[a, b], lim f dgni = f dg.
i→∞ a a
For a proof of the theorem, see Robert Kuller’s book Topics in Modern Analysis, pp. 277-281. From
the sequence Pn converging uniformly on [0, 1] to f, we apply the Helly selection theorem to gn to get a
subsequence gni converging to a g ∈ BV [0, 1]. We have
ni
X  k  n  
i
T (f) = lim T (Pni ) = lim f T xk (1 − x)ni −k
i→∞ i→∞ ni k
k=0
Xni k Xni k

= lim f αni,k = lim f gni (tk+1 ) − gni (tk )
i→∞ ni i→∞ ni
k=0 k=0
Z 1 Z 1
= lim f dgni = f dg.
i→∞ 0 0

Z 1 Z 1
Finally, for two equivalent functions g, g0 ∈ BV [0, 1] and f ∈ C[0, 1], we have f dg = f dg0 (see
0 0
pp.273-4, lemma 6.4.1 of Robert Kuller’s book Topics in Modern Analysis). So we may replace g by the
normalized representative in its equivalent class to get the isometric isomorphism.

47
 Presentation Project I – Applications of the Banach Contractive Mapping Theorem

In this project, we will look at some applications of the Banach Contractive Mapping Theorem.

(1) (1979 UC Berkeley PhD preliminary examination) An accurate map of California is spread out flat on
a table in Evans Hall, in Berkeley. Prove that there is exactly one point on the map lying directly over the
point it represents.
(1978 USA Math Olympiad) ABCD and A0 B 0 C 0D0 are square maps of the same region, drawn to
different scales and ABCD is placed totally inside A0B 0 C 0D0 (both flatly opened). Prove that there is only
one point O on ABCD which lies directly over the point O0 of A0B 0 C 0D0 such that O and O0 each represent
the same place of the region.
Solution. For every point X on the large map, define f(X) to be the corresponding point in the small map.
This is a contractive mapping on the large map, which is a closed subset of R2 , hence complete. By the
contractive mapping theorem, there is a unique fixed point.
(2) The contractive mapping theorem is used to prove the inverse function theorem (see Lecture Notes for
Math 3033), which is a key theorem in differential equations, differential geometry and several complex
variables.
(3) The contractive mapping theorem is used to prove the following local existence and uniqueness theorem
for certain first order ordinary differential equations.

Picard’s Theorem. Let f(x, y) be continuous on a closed rectangle G containing (x0 , y0) in its interior.
Suppose there is M > 0 such that |f(x, y) − f(x, y0 )| ≤ M |y − y0 | for every (x, y), (x, y0 ) ∈ G. (For example,
∂f
this is the case if is bounded on G◦ by the mean value theorem.) Then there is an interval [x0 − δ, x0 + δ]
∂y
dy
with δ > 0 where the differential equation = f(x, y) has a unique solution y = h(x) satisfying the initial
dx
condition h(x0 ) = y0 .
Proof. Note if there is a continuous function h on some interval [x0 − δ, x0 + δ] such that
Z x

h(x) = y0 + f t, h(t) dt,
x0

dy

then setting y = h(x), we have = h0 (x) = f x, h(x) = f(x, y) and h(x0) = y0 , which is what we want.
dx
dy

(Conversely, for every solution y = h(x) of = f(x, y) and h(x0 ) = y0 , we have h0 (x) = f x, h(x) and so
Z x dx


h(x) = y0 + f t, g(t) dt.) To get such a function, let K = max{|f(x, y)| : (x, y) ∈ G}. Now choose δ > 0
x0
1
such that δ < M and Sδ = {(x, y) : |x − x0| ≤ δ and |y − y0 | ≤ Kδ} ⊆ G.

Let C(I) be the set of all continuous functions on I = [x0 −δ, x0 +δ]. Define d∞ (f, g) = sup{|f(x)−g(x)| :
x ∈ I}. Then C(I) with metric d∞ is a complete metric space. Let C ∗ be the subset of C(I) consisting of
g ∈ C(I) such that for all x ∈ I, |g(x) − y0 | ≤
Kδ. Then C ∗ is nonempty because the constant function y0
is in C ∗ . For every g ∈ C ∗ , its graph { t, g(t) : t ∈ I} ⊆ Sδ ⊆ G. Also, if a convergent sequence {gn} in
C ∗ has a limit g ∈ C(I), then for all x ∈ I, |g(x) − y0 | = lim |gn(x) − y0 | ≤ Kδ. Hence, g ∈ C ∗. So C ∗ is
n→∞
closed in the complete metric space C(I) with d∞ metric. Hence C ∗ is also complete.

Define T : C ∗ → C ∗ as follow. For every g ∈ C ∗ , define T (g) = g0 , where

Z x

g0 (x) = y0 + f t, g(t) dt
x0

48


for all x ∈ I. Clearly, g0 ∈ C(I). Now for every x ∈ I, since the graph of g is in G, we have f t, g(t) ≤K
for all t ∈ I and Z x Z x


|g0(x) − y0 | = f t, g(t) dt ≤ K dt = K|x − x0| ≤ Kδ.
x0 x0

Next, we check T is a contractive mapping.

Let g, h
∈ C ∗ with T (g) = g0 and T (h) = h0 . Since the
graphs of g and h are in G, we have f t, g(t) − f t, h(t) ≤ M |g(t) − h(t)| ≤ M d∞ (g, h) for all t ∈ I and
Z Z
x

x
|g0(x) − h0 (x)| ≤ f t, g(t) − f t, h(t) dt ≤ M d∞(g, h) dt ≤ M δd∞(g, h).
x0 x0

Taking supremum over x ∈ I on the left side, we get d∞ T (g), T (h) ≤ M δd∞ (g, h) for all g, h ∈ C ∗ . Since
M δ < 1, T is a contractive mapping on C ∗ .

By the contractive mapping theorem, there is a unique fixed point h in C ∗, which is the function we
want.

49
 Presentation Project II – Applications of the Banach Contractive Mapping Theorem

This project is to give a proof of the Inverse Function Theorem using the Contractive Mapping Theorem.
We begin by recalling some facts taught in Math 3033.

Continuity
v Lemma. Let e1 , . . . , en be the usual basis of Rn , T : Rn → Rm be a linear transformation and
u n
uX
kT = t kT (ei )k2. We have kT (v)k ≤ kT kvk for every v = (x1 , . . . , xn) ∈ Rn .
i=1

Proof. By the triangle inequality and the Cauchy-Schwarz inequality,

v v
n n n u n u n
X
X X uX uX
kT (v)k = T xiei = xiT (ei ) ≤ |xi|kT (ei )k ≤ t xi t
2 kT (ei )k2 = kT kvk.
i=1 i=1 i=1 i=1 i=1

n kT (v)k o
Definition. For a linear transformation T : Rn → Rm with matrix M, since the set : v ∈ Rn \ {O}
kvk
is bounded above by kT as in the continuity lemma, we may define the norm of M and the norm of T to be
n kT (v)k o
kM k = kT k = sup : v ∈ Rn \ {O} = sup{kT (u)k : kuk = 1}.
kvk | {z }
u = v/kvk

Then kT (v)k ≤ kT kkvk ⇔ kM vk ≤ kM kkvk for every v ∈ Rn .

Theorem (Mean Value Inequality). If F : Rn → Rm is differentiable at every point of B(p, r) and

kF 0(a)k ≤ k for every a ∈ B(p, r), then kF (x) − F (y)k ≤ kkx − yk for every x, y ∈ B(p, r).


Proof. For x, y ∈ B(p, r) and t ∈ [0, 1], let γ(t) = (1 − t)y + tx and g(t) = F γ(t) . Then



kF (x) − F (y)k = kg(1) − g(0)k ≤ kg0(c)k(1 − 0) = kF 0 γ(c) (x − y)k ≤ kF 0 γ(c) kkx − yk ≤ kkx − yk.

Definition. F : Rn → Rm is C 1 near q ∈ Rn iff all partial derivatives of F are continuous in some B(q, r).

Inverse Function Theorem. If F : Rn → Rn is C 1 near q ∈ Rn and det F 0(q) 6= 0, then there exists

0

an open ball B(q, r) such that F restricted to B(q, r) has a C 1 inverse function F −1 and F −1 F (x) =

−1
F 0(x) for every x ∈ B(q, r).

Proof. First we consider the special case q = O, F (O) = O and F 0(O) = I.

(Step 1) Since F is C 1 near O, all the partial derivatives are continuous near O. Considering the n2 partial
2
derivatives as coordinate functions, we see that the function F 0 : Rn → Rn is continuous near x = O.
1 1
So for ε = > 0, there is a δ > 0 such that kx − Ok < δ implies kF 0(x) − F 0(O)k = kF 0(x) − Ik < .
2 2
n n 0 0 1
Let G : R → R be given by G(x) = F (x) − x. On B(O, δ), kG (x)k = kF (x) − Ik < . By the mean
2
1
value inequality, we get kG(x) − G(y)k ≤ kx − yk for every x, y ∈ B(O, δ). By triangle inequality, for every
2
x, y ∈ B(O, δ),

1
kF (x) − F (y)k = kx − y− G(y) − G(x) k ≥ kx − yk − kG(y) − G(x)k ≥ kx − yk. (∗)
2

(Step 2) Let 0 < r < δ/2. For each t ∈ B(O, r), we will find a s ∈ B(O, δ) such that F (s) = t. This
is the
same as finding a fixed point in B(O, δ) for the function E : Rn → Rn defined by E(s) = s + t − F (s) . Now

50
 1 1
kE 0(s)k = kI − F 0(s)k ≤ for every s ∈ B(O, δ). By the mean value inequality, kE(x) − E(y)k ≤ kx − yk
2 2
1
for every x, y ∈ B(O, δ). Since 2r < δ, so for every s ∈ B(O, 2r), we have kE(s) − E(0)k ≤ ksk. Then
2
1
kE(s)k ≤ ksk + kE(0)k ≤ r + ktk ≤ 2r. So E takes B(O, 2r) to B(O, 2r). By the contractive mapping the-
2
orem, there is a unique s ∈ B(O, 2r) ⊆ B(O, δ) such that E(s) = s, then F (s) = t. Note there is exactly one
such s ∈ B(O, δ) satisfying F (s) = t by (*).

(Step 3) Now for every t ∈ B(O, r), define F −1(t) = s, where s ∈ B(O, δ) is as in step 2. In particular,
F −1(O) = O. By (*), we have (**) kF −1(t) − F −1(t0 )k ≤ 2kt − t0 k, which implies F −1 is continuous. Since
F is differentiable at q = O, F (O) = O and DF (O) = I, so

F (x) − F (O) − DF (O)(x − O) F (x) − x

r(x) = = →O as x → O.
kxk kxk

By (**), as x → O,


kF −1(x) − F −1(O) − I(x − O)k kF −1 F (x) − r(x)kxk − O − xk
=
kxk kxk



kF −1 F (x) − r(x)kxk − F −1
F (x) k 2k(F (x) − r(x)kxk) − F (x)k
= ≤ = 2kr(x)k → 0.
kxk kxk

So F −1 is differentiable at O.


(General Case) Define φ, ψ : Rn → Rn by φ(x) = x − q and ψ(y) = DF (q)−1 y − F (q) . Note that φ and
ψ are C 1 functions and have C 1 inverses given by φ−1(x) = x + q and ψ−1 (y) = DF (q)(y) + F (q). Let
H : Rn → Rn be defined by H(x) = ψ ◦ F ◦ φ−1(x). Then H(O) = ψ(F (φ−1 (O))) = ψ(F (q)) = O and


H 0(O) = ψ0 F (q) F 0(q)(φ−1 )0(0) = (F 0(q))−1 F 0(q)I = I.

By the special case above, H has an inverse near O, which is differentiable at O. Then F has the inverse near q
given by F −1 = φ−1 ◦H −1 ◦ψ and it is differentiable at F (q). If in the above argument,we replace q by another
point q0 near q where F is C 1 , then similarly we will get F −1 is also differentiable at F (q0). Finally, F −1 is


−1
−1
C 1 near F (q) because the chain rule gives (F −1 )0 F (x) = F 0(x) (⇔ (F −1)0 (y) = F 0(F −1(y)) if we
set y = F (x)). So the partial derivatives of F −1 are polynomials of compositions of the partial derivatives
of F with the function F −1, which are continuous near q.

Remarks. (1) In the statement of the inverse function theorem, if we replace C 1 near q ∈ Rn by differentiable
1
at q ∈ Rn , then F may not have any inverse near q. For example, take F (0) = 0 and F (x) = x + x2 sin2 for
x
x 6= 0. Then F is differentiable everywhere and F 0(0) = 1. However, F has no inverse on any open interval
containing 0.
(2) If F is only C 1 at q ∈ Rn , then from the proof above, we will see that F has an inverse function near q,
which is differentiable at F (q). However, the inverse function may not be C 1 at F (q) as the partial derivatives
of the inverse function are not known to exist near F (q).

51
 Presentation Project III – Applications of the Baire Category Theorem

Definition. A set S is perfect in a metric space M iff S is closed in M and every point of S is a limit point
of S (in short, S = S 0 ).
Examples. (1) If a < b, then [a, b] is perfect in R as it is closed in R and every x ∈ [a, b] is the limit of a
sequence in [a, b] \ {x}.
(2) The Cantor set K of all numbers in [0, 1] which have base 3 expansions containing only the digits 0 or 2
is perfect in R. It is closed in R because
 1 2 1 2 7 8 
K = [0, 1] \ ( , ) ∪ ( , ) ∪ ( , ) ∪ · · · .
3 3 9 9 9 9
For every x ∈ K, let xn ∈ K such that xn and x have the same base 3 expansion except for the n-th digit.
2
Then xn 6= x and |xn − x| = n → 0 as n → ∞ so that {xn} converges to x.
3

Theorem. Every nonempty perfect set S (in particular the Cantor set) in a complete metric space M is
uncountable.
Proof. Since M is complete and S is a closed in M, S is also complete. Suppose S is countable, say

[
∞
S = {r1, r2, r3, . . .} = {rn}.
n=1

However, each {rn} is nowhere dense because {rn} = {rn} and being perfect, each rn is the limit of a
sequence in S \ {rn} and hence if there is any nonempty open ball contained in {rn}, it would contain more
than one point, contradiction. So S is of first category in itself, contradicting the Baire category theorem.

Theorem. Let C([0, 1]) be the set of all continuous real-valued functions on [0, 1] with metric d∞(f, g) =
sup{|f(x) − g(x)| : x ∈ [0, 1]}. Then the set of all continuous, nowhere differentiable functions is dense in
C([0, 1]).
Proof. For every m ∈ N, let Em be the set of all f ∈ C([0, 1]) such that there exists a s ∈ [0, 1] so that
|f(x) − f(s)| ≤ m|x − s| for every x ∈ [0, 1].

|f(x) − f(a)|/|x − a| if x ∈ [0, 1] and x 6= a
If f ∈ C([0, 1]) is differentiable at a ∈ [0, 1], then g(x) =
|f 0 (a)| if x = a
∞
[
is continuous on [0, 1]. So g(x) has a maximum value M on [0, 1]. Then f ∈ EdM e . Therefore, Em contains
m=1
every continuous function on [0, 1] that is differentiable in at least one a ∈ [0, 1].
Next, we will show every Em is closed in C([0, 1]). Let f ∈ C([0, 1]) be a cluster point of Em . Then there
is a sequence {fn } in Em converging to f, i.e. lim d∞ (fn , f) = 0. For every fn , there exists a sn ∈ [0, 1]
n→∞
satisfying |fn(x) − fn (sn )| ≤ m|x − sn | for every x ∈ [0, 1]. By the Bolzano-Weierstrass theorem, there is a
subsequence {snk } converging to some s ∈ [0, 1]. Passing to a subsequence if necessary, we may assume sn is
snk and fn is fnk . Now

|fn(sn ) − f(s)| ≤ |fn(sn ) − f(sn )| + |f(sn ) − f(s)| ≤ d∞(fn , f) + |f(sn ) − f(s)| → 0 as n → ∞

by the convergence of {fn } to f and the continuity of f at s. Then

|f(x) − f(s)| = lim |fn (x) − fn (sn )| ≤ lim m|x − sn | = m|x − s|.
n→∞ n→∞

So f ∈ Em . Therefore Em is closed.

52
 Also, every Em has no interior point. To see this, suppose g is an interior point of Em . Then there
is r > 0 such that B(g, r) ⊆ Em . Since [0, 1] is compact, g is uniformly continuous on [0, 1]. So for r > 0,
there exists δ = 1/k > 0 such that we can partition [0, 1] into k closed subintervals Ii = [(i − 1)/k, i/k]
where |g(x) − g(y)| < r/2 for every x, y ∈ Ii . In the rectangle Ii × [g((i − 1)/k) − r/2, g((i − 1)/k) + r/2],
we can draw a polygonal graph joining ((i − 1)/k, g((i − 1)/k)) and (i/k, g(i/k)) which remains within the
rectangle and whose line segments have slopes greater than m in absolute value. Let’s consider the function
f ∈ C([0, 1]) whose graph consisted of the k polygonal paths just described. Then d∞ (g, f) ≤ r/2 < r so
that f ∈ B(g, r), but f ∈
/ Em , a contradiction to B(g, r) ⊆ Em .
∞
[
By the Baire Category theorem, C([0, 1]) \ Em is dense in C([0, 1]). Such functions are continuous
m=1
on [0, 1], but nowhere differentiable.

53
 Presentation Project IV – Application of the Baire Category Theorem

Amazingly there exist functions continuous on R\Q and discontinuous on Q. Here is an example. Define

 0 if x ∈ R \ Q
f(x) = 1 if x = 0
1
n if x ∈ Q and x = m
n in lowest term with n ∈ N.

This function f is discontinuous at every x ∈ Q because by density of irrational numbers, there exists
xn ∈ R \ Q such that xn → x, but then f(xn ) = 0 does not converge to f(x) 6= 0.
Next, f is continuous at every x ∈ R \ Q because for every ε > 0, there is a k ∈ N such that k1 < ε. Since
x ∈ R \ Q, x is at some distance δ from rational numbers of the form m n , where n = 1, 2, · · ·, k − 1. Then
there is no such rational numbers in (x − δ, x + δ) so that |y − x| < δ implies |f(y) − f(x)| = |f(y)| ≤ k1 < ε.

Question: Is there a function continuous on Q and discontinuous on R \ Q?

Definitions. For every f : R → R and bounded nonempty open interval J, define the oscillation of f over
J to be
ω(f, J) = sup{|f(x) − f(y)| : x, y ∈ J}
and for x0 ∈ R, define the oscillation of f at x0 to be

ω(f, x0 ) = inf{ω(f, J) : J bounded nonempty open interval containing x0 }.

Lemma. A function f : R → R is continuous at x0 if and only if ω(f, x0 ) = 0.

Proof. If f is continuous at x0, then for every ε > 0, there is an interval J = (x0 − δ, x0 + δ) such that x ∈ J
implies |f(x) − f(x0 )| < 0.4ε. Then for every x, y ∈ J, |f(x) − f(y)| ≤ |f(x) − f(x0 )| + |f(x0 ) − f(y)| < 0.8ε.
So ω(f, J) < ε. Then 0 ≤ ω(f, x0 ) < ε for every ε. Therefore, ω(f, x0 ) = 0.
Conversely, if ω(f, x0 ) = 0, then by the infimum property, for every ε > 0, there is a bounded nonempty
open interval J containing x0 such that 0 ≤ ω(f, J) < ε. As x0 is an interior point of J, there is Bδ (x0) =
(x0 − δ, x0 + δ) ⊆ J. So |x − x0 | < δ implies x ∈ B(x0 , δ) ⊆ J, which implies |f(x) − f(x0 )| ≤ ω(f, J) < ε.
Therefore, f is continuous at x0.

Theorem. If f : R → R is continuous on a dense set, then the set D of discontinuous points of f is first
category in R. In particular, if f : R → R is continuous on Q, then the set of discontinuous points cannot be
R \ Q.
∞
[
1
Proof. For every m ∈ N, define Em = {x : ω(f, x) ≥ m }. By the lemma, D = Em .
m=1

We will show every Em is closed. Let x0 be a limit point of Em . Then every bounded nonempty open
interval J containing x0 will contain a Br (x0) = (x0 − r, x0 + r), which has an x ∈ Em and x 6= x0. So
1
ω(f, J) ≥ ω(f, x) ≥ m . Then ω(f, x0 ) = inf{ω(f, J) : J bounded nonempty open interval containing x0 } ≥
1
m , i.e. x 0 ∈ Em . Therefore, Em is closed.
Every Em has no interior point for otherwise there will be an open ball B in Em . Since f is continuous on
a dense set, this will imply f is continuous at some point in B ⊆ Em ⊆ D, a contradiction to the definitipon
of D. Therefore, D is of first category.

Remarks. In the proof, we saw that the set of discontinuous points of a function f : R → R is the countable
union of closed sets. Taking complement, the set of continuous points is a countable intersection of open
sets. These sets come up in measure theory. Any set that is a countable union of closed sets is called a Fσ
set and any set that is a countable intersection of open sets is called a Gδ set. In Karl Stromberg’s book,

54
An Introduction to Classical Real Analysis, p. 120, exercise # 3, it shows that a subset S of R is the set of
continuous points for a function f : R → R if and only if S is a Gδ subset of R.

Hans Lewy’s Example. Finally we mention one more application of the Baire category theorem. In 1957,
Hans Lewy published an amazing result that there existed a linear partial differential equation having no
solution anywhere! This is a big contrast with the general knowledge that many linear differential equations
have solutions in at least some open intervals. Hans Lewy’s result via the Baire category theorem can be
found in Fritz John’s book Partial Differential Equation, 4th ed., pp. 235-9.

55
 Presentation Project V – Application of Baire Category Theorem

In this project, we raise two questions about how bad the discontinuities of some functions can be and
resolve them using the Baire category theorem.

Question 1 For a differentiable function f : R → R, can f 0 (x) be discontinuous at every x ∈ R?

Next, we will point out that there exists a sequence of continuous functions fn : R → R such that it is
pointwise convergent on R to some function f : R → R, but f is not continuous on R. For this, let fn : R → R
be defined by fn (x) = 1/(1 + nx2). Then each fn is continuous, yet

1 if x = 0
f(x) = lim fn (x) = .
n→∞ 0 if x 6= 0

Question 2 Can the pointwise limit of a sequence of continuous functions fn : R → R be nowhere continu-
ous? In particular, can the limit of such fn be XQ , the characteristic function of Q?

We will deal with question 2 first. For a function f : R → R, let D(f) = {x ∈ R : f is discontinuous at x}.
In project IV, we have the following definitions and lemma.

Definitions. For every f : R → R and bounded nonempty open interval J, define the oscillation of f over
J to be
ω(f, J) = sup{|f(x) − f(y)| : x, y ∈ J}
and for x0 ∈ R, define the oscillation of f at x0 to be

ω(f, x0 ) = inf{ω(f, J) : J bounded nonempty open interval containing x0 }.

Lemma. A function f : R → R is continuous at x0 if and only if ω(f, x0 ) = 0. In particular,

[
∞
D(f) = {x : ω(f, x) > 0} = {x : ω(f, x) ≥ 1/k}.
k=1
Fact. If f : R → R is any function, then for any r > 0, Sr = {x : ω(f, x) < r} is an open set. (To see that,
take any x0 ∈ Sr , i.e. ω(f, x0 ) < r. By infimum property, there exists a bounded nonempty open interval J
containing x0 such that ω(f, x0 ) ≤ ω(f, J) < r. For any y ∈ J, we have ω(f, y) ≤ ω(f, J) < r, i.e. y ∈ Sr . So
J ⊆ Sr and J is open interval containing x0 . Hence, Sr is open.)

Next, we will present a theorem which asserts that the pointwise limit of a sequence of continuous
functions fn : R → R must be continuous on some dense subset of R. The gives a negative answer to
question 2.

The Baire-Osgood Theorem. Let fn : R → R be continuous for each n. Suppose fn converges pointwise
on R to f. Then D(f) is of the first category in R. In particular, f is continuous on some dense subset of R
by part (2) of the Baire category theorem.

Proof. The lemma and the fact above gives us D(f) is the countable union of closed sets. It remains to
show that for every k, the closed set F = {x : ω(f, x) ≥ 1/k} has empty interior. (This may look like the
theorem in project IV, but it is not, since we do not know where f is continuous.) Assume the contrary that
F contains an interval I with nonempty interior. Since F is closed, F contains I and we may assume I is a
closed interval.
1 \
Set ε = and consider En = {x : |fi (x) − fj (x)| ≤ ε}. Since fn is pointwise convergent on R, for
5k
i,j≥n
∞
[
every x ∈ R, fn (x) is a Cauchy sequence. From the definition of a Cauchy sequence, we have En = R.
n=1

56
 [
∞
Since all fn are continuous, so En is closed. Since I is complete, En ’s are closed and I = (En ∩ I), by
n=1
the Baire category theorem, there exists some n such that En ∩ I contains a nonempty open interval T. We
will show T ⊆ R \ F = {x : ω(f, x) < 5ε}, which will contradict I ⊆ F.
Since T ⊆ En , we have for all x ∈ T and all i, j ≥ n, |fi (x) − fj (x)| ≤ ε. Let i → ∞ and j = n,
we have |f(x) − fn (x)| ≤ ε for all x ∈ T. Since fn is continuous, for each x0 ∈ T, there is an open
interval Ix0 ⊆ T containing x0 such that |fn (x) − fn (x0 )| ≤ ε for all x ∈ Ix0 . By the triangle inequality,
for all x ∈ Ix0 , |f(x) − fn (x0 )| ≤ |f(x) − fn (x)| + |fn (x) − fn (x0)| ≤ 2ε. Finally, for all x, y ∈ Ix0 ,
|f(x) − f(y)| ≤ |f(x) − fn (x0| + |fn(x0 ) − f(y)| ≤ 4ε. Then ω(f, x0 ) ≤ ω(f, Ix0 ) ≤ 4ε. Hence, x0 6∈ F.

Remarks. (a) For question 2, since D(XQ ) = R and R is of the second category in R, by the Baire-Osgood
theorem, XQ cannot be written as the limit of a sequence of continuous functions on R. However, for all
x ∈ R, XQ (x) = lim lim (cos m!πx)2n.
m→∞ n→∞

(b) For question 1, if f : R → R is differentiable, then f 0 (x) = lim fn (x), where fn (x) = (f(x+1/n)−f(x))/n
n→∞
is continuous on R. So by the Baire-Osgood theorem, f 0 (x) is continuous on some dense subset of R.

57
 Presentation Project VI – Application of Baire Category Theorem

Part (a) of the following problem is a simple exercise of using the Baire categoty theorem and appeared
in the PhD Qualifying Exam of many famous universities. Part (b) is a much deeper result and required
more analysis.

Problem. Let f : R → R be a C ∞ -function (i.e. infinitely differentiable) such that for every x ∈ R there is
a nonnegative integer n such that f (n) (x) = 0.
(a) Prove that on some nonempty open interval (a, b), f is a polynomial.
(b) Prove that f is a polynomial on R.

Solution. (b) Call f locally polynomial at x if there exists a polynomial P such that f = P on some open
interval containing x. In that case, call P a local polynomial for x.
First we claim that if f is locally polynomial at each point of some open interval (c, d), then f is a
polynomial on (c, d). To see this, suppose P is the local polynomial for some x ∈ (c, d). Let S be the closed
set of all accumulation points of T = {t ∈ (c, d) : f(t) = P (t)} in (c, d). S is also open in (c, d) because if
y ∈ S and Q is the local polynomial for y, then Q = f = P on a sequence in T with limit y, so Q = P and
some open interval about y will be in T and hence in S. Since (c, d) is connected, so S = (c, d). Now t ∈ S
implies there is a sequence tn ∈ T converging to t. Hence f(t) = lim f(tn ) = lim P (tn) = P (t). Therefore
n→∞ n→∞
f = P on S = (c, d).
Next, the set G = {x ∈ R : f is locally polynomial at x} is open by the definition of locally polynomial.
Let A = R \ G, then A is closed in R. Now A contains no isolated points because if c is isolated in A,
then f would be given by polynomials in the components of R \ A to the left and right of c. Since f is
C ∞ , the polynomials on the components must be the same and f would then be locally polynomial at c, a
contradiction.
Assume A is nonempty. For every nonnegative integer m, let Am = {t ∈ A : f (m) (t) = 0}. Then Am is
∞
closed in A and A = ∪ Am . Next assume the interior of Am is nonempty in A, i.e. there is an open interval
m=0
I such that ∅ 6= A ∩ I ⊆ Am . Since A has no isolated points, every x ∈ A ∩ I is the limit of a sequence of
distinct tn’s in A, which except for finitely many terms, also belong to I, hence these tn ’s will eventually be
in Am . This implies x is an accumulation point of Am . By Rolle’s theorem, f (k) (x) = 0 for k ≥ m. Now on
every component of I \ A, f is given by a polynomial and the polynomial can have degree at most m − 1 (if
the degree is k ≥ m, then f (k) would be a nonzero constant in the component and f (k) would be nonzero at
the endpoints of the component, at least one endpoint of these belongs to A ∩ I, contradiction). So on I,
f (m) ≡ 0. Then I ⊆ G and hence A ∩ I = ∅, a contadiction. So Am is nowhere dense in A. However, as A is
a nonempty complete metric space, this contradicts the Baire category theorem. Hence, A must be empty
and f is locally polynomial at all x ∈ R. Therefore, f is a polynomial on R.
Remarks. The problem is still true if the domain of f is an open interval instead of R. This is because the
Baire category theorem is topological, i.e. all we need to apply the Baire category theorem is the domain
homeomorphic to a complete meteric space.

58
 Presentation Project VII – the Cantor-Hausdorff Completion of Metric Spaces

In this project, we will present the Cantor-Hausdorff completion of metric spaces. We will denote a
sequence x1, x2, x3, . . . by {xn}, where n is a dummy variable. Let (M, d) be a metric space.

Step 1 For any two Cauchy sequences {rn}, {sn} in M, we will check lim d(rn, sn ) exists in R. For every
n→∞
ε > 0, since {rn}, {sn} are Cauchy, there exists K0 ∈ N such that m, n ≥ K0 implies d(rm , rn) < ε/2 and
d(sm , sn) < ε/2. Then

|d(rm , sm ) − d(rn, sn )| ≤ |d(rm, sm ) − d(rn, sm )| + |d(rn, sm ) − d(rn, sn)| ≤ d(rm , rn) + d(sm , sn) < ε.

Step 2 Let S be the set of all Cauchy sequences in M. For each {xn} ∈ S, let

[{xn}] = {yn} ∈ S : lim d(xn, yn) = 0 and c = {[{xn}] : {xn} ∈ S}.
M
n→∞

c we define
So [{xn}] = [{yn}] if and only if lim d(xn, yn) = 0. For [{an}], [{bn}] ∈ M,
n→∞

b
d([{an }], [{bn}]) = lim d(an , bn).
n→∞

We have to check db is well-defined. If [{an}] = [{xn}] and [{bn}] = [{yn}], then

|d(an, bn)−d(xn , yn)| ≤ |d(an, bn)−d(xn , bn)|+|d(xn, bn)−d(xn , yn )| ≤ d(an, xn)+d(bn , yn ) → 0 as n → ∞.

So lim d(an , bn) = lim d(xn, yn).

n→∞ n→∞

Step 3 We check db is a metric on M c. Now d([{a

b b b
n }], [{bn}]) = d([{bn }], [{an}]) ≥ 0. Also d([{an }], [{bn}]) = 0
if and only if [{an}] = [{bn}]. Triangle inequality follows from lim d(an, bn) + lim d(bn, cn) ≥ lim d(an, cn).
n→∞ n→∞ n→∞

Step 4 Define p : M → M c by assigning x ∈ M to p(x) = [x, x, x, . . .] ∈ M.

c This is an isometry because
b
d(p(x), p(y)) = lim d(x, y) = d(x, y).
n→∞

c by showing for every [{xn}] ∈ M

Step 5 We will show p(M ) is dense in M c, lim p(xm ) = [{xn}]. For every
m→∞
ε > 0, since x1 , x2, x3, . . . is Cauchy, there exists K1 ∈ N such that m, n ≥ K1 implies d(xm , xn) < ε/2.
Then for m ≥ K1 , we have d(p(x b m ), [{xn}]) = lim d(xm , xn ) ≤ ε/2 < ε. So p(x1 ), p(x2 ), p(x3 ), . . . converges
n→∞
c, d).
to [{xn}] in (M b

Step 6 We will show (M, c d)b is complete. Let W1, W2 , W3, . . . be a Cauchy sequence in (M b For every
c, d).
c, there is yn ∈ M such that d(p(y
Wn , since p(M ) is dense in M b n ), Wn ) < 1/n. We will check {yn } is a
Cauchy sequence in M. For every ε > 0, there exists K2 ∈ N such that m, n ≥ K2 implies d(Wb m , Wn) < ε/3.
Then for m, n ≥ max{K2 , d3/εe}, we have

b b b b 1 ε 1
d(ym , yn) = d(p(y m ), p(yn )) ≤ d(p(ym ), Wm ) + d(Wm , Wn ) + d(Wn , p(yn )) < + + < ε.
m 3 n
Let W = [{yn}] ∈ M. c We will show W1 , W2, W3 , . . . converges to W in M c. For every ε > 0, since
b
lim d(p(y b
n ), W ) = 0, there is a K3 ∈ N such that n ≥ K3 implies d(p(yn ), W ) < ε/2. If n ≥ max{K3 , d2/εe},
n→∞
then
b n , p(yn)) + d(p(y
b n , W ) ≤ d(W b 1 ε
d(W n ), W ) < + < ε.
n 2
c is complete.
Therefore, M

59
 Finally, we will give an example of using the completion concept to solve a problem.

Problem. Let M be a metric space such that every closed and bounded subset of M is compact. Prove
that M is complete.

Solution. Assume M is not complete. Then there exists a divergent Cauchy sequence {xn} in M. Embed
M in a completion M c. Then {xn} is also a Cauchy sequence in M c. So {xn} converges to some x ∈ M c and
c
x 6∈ M. In M , {x1, x2, x3, . . .} = {x1, x2, x3, . . .} ∪ {x} is closed. So M ∩ {x1, x2, x3, . . .} = {x1, x2, x3, . . .} is
closed in M. It is also bounded in M as it is a Cauchy sequence. By the property of M, we see {x1, x2, x3, . . .}
is compact. By metric compactness theorem, it is complete, contradicting {xn} diverges in M.

60
 Presentation Project VIII – Application of Arzela-Ascoli’s Theorem

From uniform convergence, we know that if a sequence {fn } converges in C([a, b]), i.e. a sequence of
continuous functions {fn } converges uniformly on [a, b], then the limit function f is continuous on [a, b] and
Z b Z b Z b
f(x) dx = lim fn (x) dx = lim fn (x) dx. (∗)
a a n→∞ n→∞ a

Similar to C([a, b]), for a compact metric space K, we can define a metric on the collection C(K) of all
continuous functions f : K → R by defining d∞ (f, g) = sup{|f(x) − g(x)| : x ∈ K}. By the extreme value
theorem, this is the same as max |f(x) − g(x)|. As C([a, b]), the metric space C(K) is also complete. Also,
x∈K
in later courses, we also consider complex-valued functions, i.e. f : K → C. Then we denote that collection
by C(K, C) and give it the same metric. It is also complete.

Definitions. Let F be a collection of real-valued functions on a metric space M.

(a) F is pointwise bounded on M iff for every x ∈ M, there exists Cx ≥ 0 (depending on x) such that for
every f ∈ F, we have |f(x)| ≤ Cx.
(b) F is uniformly bounded on M iff there exists C ≥ 0 such that for every f ∈ F and every x ∈ M, we
have |f(x)| ≤ C.
(c) F is uniform equicontinuous on M iff for every ε > 0, there exists δ > 0 such that for every f ∈ F and
x, x0 ∈ M, d(x, x0) < δ implies |f(x) − f(x0 )| < ε. (Note this implies every function f ∈ F is uniformly
continuous.)

Arzela-Ascoli’s Theorem. Let K be a compact metric space and {fn } is a sequence of functions on K. If
{fn } is pointwise bounded and uniform equicontinuous on K, then {fn } is uniformly bounded on K and has
a subsequence {fnj } converging to some f ∈ C(K).

The theorem is also true for complex-valued functions. Here we will give an important application of
the Arzela-Ascoli theorem on the theory of ordinary differential equation.

Peano’s Theorem. Let f(x, y) be continuous on a closed rectangle G = [x0−r, x0 +r]×[y0 −s, y0 +s]. Suppose
s
c = max{|f(x, y)| : (x, y) ∈ G} > 0 and δ = min(r, ). Then on the interval [x0 − δ, x0 + δ], the differential
c
dy
equation = f(x, y) has at least one solution y = h(x) satisfying the initial condition h(x0) = y0 .
dx
Remarks. In Picard’s theorem, in addition to the continuity of f we also required f to be Lipschitz on the
y-variable (i.e. there is M > 0 such that |f(x, y) − f(x, y0 )| ≤ M |y − y0 | for every (x, y), (x, y0 ) ∈ G). Then
we get existence and uniqueness of solution. In Peano’sp theorem, no Lipschitz condition is required, but we
get only an existence theorem. In fact, for f(x, y) = |y| on [−1, 1] × [−1, 1], there are two solutions to the
dy x2 /4 if x ≥ 0
differential equation = f(x, y) with y(0) = 0, namely y1 (x) = and y2 (x) = 0.
dx −x2 /4 if x < 0

Proof of Peano’s Theorem. (Due to Leonida Tonelli) First we consider solution on [x0, x0 + δ]. As in
the proof of Picard’s theorem, all we need to Zshow is that there exists a continuous function h on the
x

interval K = [x0, x0 + δ] such that h(x) = y0 + f t, h(t) dt for every x ∈ K. Note K is compact by the
x0
Heine-Borel theorem.
Z x− nδ
 δ

For every n = 1, 2, 3, . . ., define hn(x) = y0 for x ∈ x0, x0 + and hn(x) = y0 + f t, hn(t) dt
n x0
 δ 2δ   2δ 3δ   (n − 1)δ 
for x ∈ x0 + , x0 + , then for x ∈ x0 + , x0 + , . . . , then for x ∈ x0 , , x0 + δ . Note
nZ n n n n
x0 +δ

|hn(x) − y0 | ≤ c dt = cδ ≤ s so that f is defined at x, hn(x) for every x ∈ K and n = 1, 2, 3, . . ..
x0

61
 Z x0 +δ
Since |hn(x)| ≤ |y0 | + c dt ≤ |y0| + cδ for every x ∈ K and n = 1, 2, 3, . . ., the sequence {hn}
x0
ε ε
is uniformly bounded on K. For every ε > 0, we have > 0 and for every hn and x, x0 ∈ K, |x − x0 | <
Z x− δ c c
n
0 0
implies |hn(x) − hn(x )| ≤ c dt = c|x − x | < ε. So {hn} is equicontinuous on K. By the Arzela-Ascoli
δ
x0 − n
theorem, {hn} has a subsequence {hnj } converging to some h ∈ C(K).



Using the uniform continuity of f on G, we can see that gj (t) = f t, hnj (t) converges to g(t) = f
t, h(t)
in C(K). The details are as follow. For every ε > 0, there is δ0 > 0 such that d (x, y), (x0, y0 ) < δ0
0 0
implies |f(x, y) − f(x
, y )| < ε; take K0 so that j ≥ K0 implies d∞(hnj , h) < δ0 , then
for every t
∈ K,
d (t, hnj (t)), (t, h(t)) = |hnj (t)−h(t)| ≤ d∞(hnj , h) < δ0 implies |gj (t)−g(t)| = |f t, hnj (t) −f t, h(t) | < ε.
Finally, since Z Z
x
x

hnj (x) = y0 + f t, hnj (t) dt − f t, hnj (t) dt
x0 x− nδ
j
Z x
δ
and f t, hnj (t) dt ≤ c → 0 as j → ∞, so taking limits on both sides and using (*) on the right
x− nδ nj
j Z x


side, we get h(x) = y0 + f t, h(t) dt for every x ∈ K. This gives h(x0 ) = y0 , h0 (x) = f(x, h(x)) for all
x0
x ∈ (x0 , x0 + δ] by the fundamental theorem of calculus and

h(x0 + c) − h(x0 )
h0(x0 +) = lim = f(x0 , h(x0)) = f(x0 , y0 ).
c→0+ c

Finally, the argument for [x0 − δ, x0] is similar, in particular h0(x0 −) = f(x0 , y0). Then combining solutions
of both sides, we see h is differentiable at x0.

62
 Presentation Project IX – Application of Arzela-Ascoli’s Theorem

Below we will look at some problems that require the Arzela-Ascoli theorem for their solutions.

Question 1. Let fn : R → R be differentiable for each n = 1, 2, . . . with |fn0 (x)| ≤ 1 for all n and x. Suppose
lim fn (x) = g(x) for all x. Prove that g : R → R is continuous.
n→∞

Solution. Let N > 0. It suffices to show g is continuous on [−N, N ]. For all n ∈ N and x, y ∈ [−N, N ],
by the mean value theorem, |fn(x) − fn (y)| = |fn0 (θ)(x − y)| ≤ |x − y|. So for every ε > 0, let δ = ε, we
get |x − y| < δ implies for all n ∈ N, |fn(x) − fn(y)| ≤ |x − y| < δ = ε. Hence, {fn } is equicontinuous.
Since the limit of fn (0) exists, we see {fn (0)} is bounded in R, say |fn(0)| ≤ C for all n. Then |fn (x)| ≤
|fn(x) − fn (0)| + |fn (0)| ≤ |x − 0| + C ≤ N + C implies fn (x) is uniformly bounded. Hence, by the Arzela-
Ascoli theorem, some subsequence fnk converges uniformly on [−N, N ] to g(x). By continuity theorem for
uniform convergence, since fnk ’s are continuous, g is continuous on [−N, N ].

Question 2. For n = 1, 2, 3, . . ., let {fn } be a uniformly bounded equicontinuous sequence of real-valued

functions on the compact metric space (X, d). Define gn : X → R by gn(x) = max{f1 (x), f2 (x), . . ., fn (x)}.
Prove that {gn} converges uniformly on X.

Solution. For every ε > 0, by the equicontinuity of {fn }, there is δ > 0 such that d(x, y) < δ implies
|fn(x) − fn (y)| < ε for all n. Fix n and fix x, y with d(x, y) < δ. Let j, k ≤ n be such that gn (x) = fj (x) and
gn(y) = fk (y). Then gn(y) ≥ fj (y). So gn (x) − gn (y) ≤ fj (x) − fj (y) < ε. Similarly, gn (x) ≥ fk (x) and so
gn(y)−gn (x) ≤ fk (y)−fk (x) < ε. Then |gn(x)−gn (y)| < ε. It follows that {gn} is uniformly equicontinuous.
Since {fn } is uniformly bounded, it follows {gn } is also uniformly bounded. By the Arzela-Ascoli
theorem, some subsequence {gnk } converges uniformly on X to a limit function g. Since {gn} is nondecreasing,
for n > nk , we have gnk ≤ gn ≤ g, which implies {gn} converges uniformly on X to g.

Question 3. For n = 1, 2, 3, . . ., let the functions fn : [0, 1] → [0, 1] satisfy |fn(x)−fn (y)| ≤ |x−y| whenever
|x − y| ≥ 1/n. Prove that the sequence {fn } has a uniformly convergent subsequence.

Solution. For each n, let gn be the continuous function that equals fn at k/n for k = 0, 1, 2, . . ., n and the
graph of gn is a line segment on [(k − 1)/n, k/n]. Then gn(x) ∈ [0, 1] implies the gn’s are uniform bounded.
Next, by the assumption on fn , the absolute values of the slopes of gn on these intervals are

gn (k/n) − gn ((k − 1)/n) fn (k/n) − fn ((k − 1)/n)

= ≤ 1.
1/n 1/n
i i+1 i j
If 0 ≤ n
≤x≤y≤ n
≤ 1, then |gn(x) − gn (y)| ≤ |x − y|. Otherwise, if 0 ≤ x ≤ n
≤ n
≤ y ≤ 1, then

i j−1
X k  k + 1 j 
|gn(x) − gn(y)| ≤ gn(x) − gn + gn − gn + gn − gn(y)
n n n n
k=i
j−1
X k k+1
i j
≤ x− + − + − y = |x − y|
n n n n
k=i

For every ε > 0, taking δ = ε, we see {gn} is uniformly equicontinuous. By the Arzela-Ascoli theorem, some
subsequence {gnj } converges uniformly on [0, 1] to a limit function g. Fix x ∈ [0, 1]. For each j, let xj be a
point in [0, 1] of the form k/nj (k = 0, 1, 2, . . ., nj ) such that 1/nj ≤ |x − xj | < 2/nj . For every x ∈ [0, 1],

2 2
|g(x) − fnj (x)| ≤ |g(x) − gnj (x)| + |gnj (x) − gnj (xj )| + |fnj (xj ) − fnj (x)| ≤ kg − gnj k∞ + + →0
nj nj

as j → ∞. Therefore, {fnj } converges uniformly on [0, 1] to g.

63
 Presentation Project X – Application of Arzela-Ascoli’s Theorem

Definitions. Let G be a nonempty domain in C. A closed curve in G is a continuous function γ : [0, 1] → G

such that γ(0) = γ(1). G is simply connected if and only if every closed curve γ in G can be shrunk to a
point p ∈ G without leaving G (more precisely, there exists a continuous function H : [0, 1] × [0, 1] → G such
that for all t ∈ [0, 1], H(0, t) = γ(t) and H(1, t) = p).

Our goal is to prove the following.

Riemann Mapping Theorem. For every simply connected domain G 6= C and every a ∈ G, there exists
a bijective holomorphic (or analytic) function f : G → B(0, 1) such that f(a) = 0 and f 0 (a) > 0.

The following facts are needed and proofs can be found in many textbooks on complex analysis.

Weierstrass’ Theorem. If {fn} is a sequence of holomorphic functions on a nonempty domain U such

that fn converges uniformly on every compact subsets of U, then the limit function f is holomorphic in U.
Hurwitz’ Theorem In Weierstrass’ theorem, if fn is also injective on U, then either f is injective on U
or constant on U.
Square Root Theorem. Let G be a simply connected domain. If h : G → C is holomorphic and h has no
roots on G, then there exists a holomorphic function φ : G → C such that φ2 (z) = h(z) for all z ∈ G.
Remarks. For proofs of Weierstrass’ and Hurwitz’ Theorems, see p. 136 and p. 232 of Gamelin’s book
Complex Analysis respectively. For a proof of the square root theorem, see pp. 274-5 of Rudin’s book
Real and Complex Analysis, 3rd edition.

Next, we will recall Arzela-Ascoli’s theorem and use it to prove a useful fact.

Arzela-Ascoli Theorem. Let K be a compact metric space and {fn } is a sequence of functions on K. If
{fn } is pointwise bounded and equicontinuous at each z ∈ K, then {fn} has a subsequence {fnj } converging
uniformly to some f ∈ C(K).

Normality Theorem. If {fn } is a sequence of uniformly bounded holomorphic functions on a nonempty

domain U, then it has a subsequence {fnk } converging uniformly on every compact subset of U.
Proof. For n = 1, 2, 3, . . ., the set Kn = B(0, n) ∩ {z ∈ G : d(z, C \ G) ≥ n1 } is closed and bounded, hence
◦
compact by the Heine-Borel theorem. We have Kn ⊆ Kn+1 ⊆ U and K1 ∪ K2 ∪ K3 ∪ · · · = U. If K is a
n
compact subset of U, then since {Kn◦ : n = 1, 2, 3, . . .} covers K, we have K ⊆ ∪ Ki◦ ⊆ Kn for some n.
i=1

Let M be a uniform bound for all fn . We will show {fn } is equicontinuous at every a ∈ U. Take
B(a, r) ⊆ U. For every ε > 0, let δ = min{ 2r , 2M
rε
} > 0. Then for |z − a| < δ ≤ r/2 and for all fn , we have
Z  f (w) f (w) 
1 n n
|fn(z) − fn (a)| = − dw
2π |w−a|=r w − z w−a
Z
1 fn (w)(z − a) M |z − a| 2M δ
≤ |dw| ≤ 2πr < ≤ ε.
2π |w−a|=r (w − z)(w − a) 2πr(r/2) r

Let f0,n = fn . By Arzela-Ascoli’s theorem, we can inductively (for j = 1, 2, 3, . . .,) get a subsequence
{fj,n} of {fj−1,n} converging uniformly on Kj . Taking the diagonal subsequence {fnj } = {fj,j }, we see it
converges uniformly on every Kn, hence on every compact subset K of U. This concludes the proof.

Next, we will present the proof of the Riemann mapping theorem. Let F be the set of all injective
holomorphic functions g : G → B(0, 1) with g(a) = 0 and g0 (a) > 0. (The idea is to find a g ∈ F that is

64
surjective, i.e. with g(G) largest possible. Intuitively, the larger g0 (a) is, the faster g(G) should grow. So we
may consider the supremum c of g0 (a) over all g ∈ F. From supremum, we get a sequence fn ∈ F with limit
of fn0 (a) tends to c. Arzela-Ascoli’s theorem is the tool to let us a get a subsequence fnk converging (in the
sense of Weierstrass’ theorem) to a holomorphic function f. It is matter of checking to see f ∈ F and it is
surjective.)

Step 1 (F 6= ∅) Pick b ∈ C \ G. By the square root theorem, there exists a holomorphic function φ : G → C
such that φ2 (z) = z − b for all z ∈ G. If φ(w) = ±φ(z), then w = b + φ2(w) = b + φ2 (z) = z. So φ is
injective. By the open mapping theorem, the range φ(G) is open. Hence, it contains some B(c, r), where we
can choose 0 < r < |c| by shrinking r. From above, w, z ∈ G with φ(w) = −φ(z) ∈ B(c, r) is impossible, so
φ(G) ∩ B(−c, r) = ∅. Then (φ(G) + c) ∩ B(0, r) = ∅ and it follows that g(z) = r/(φ(z) + c) ∈ F.

Z
1 g(z) 1
Step 2 Observe that if d(a, C \ G) = R > r > 0 and g ∈ F, then 0 < g0 (a) ≤ |dz| ≤ .
2π |z−a|=r (z − a)2 r
So c = sup{g0 (a) : g ∈ F} is a positive number.
As a supremum, there exists a sequence {fn } in F such that lim fn0 (a) = c. Since fn (G) ⊆ B(0, 1),
n→∞
{fn } is uniformly bounded by 1. By the normality, Weierstrass’ and Hurwitz’ theorems above, there exists
a subsequence {fnk } such that f(z) = lim fnk (z) is injective holomorphic on G because f(a) = 0 and
k→∞

Z Z
1 f(z) 1 fnk (z)
f 0 (a) = 2
dz = lim dz = lim fn0 k (a) = c > 0, i.e. f ∈ F.
2πi |z−a|=r (z − a) k→∞ 2πi |z−a|=r (z − a)2 k→∞

Step 3 (f(G) = B(0, 1)) Assume there is d ∈ B(0, 1) \ f(G). By the square root theorem, there exists a
f(z) − d h(z) − h(a)
holomorphic function h : G → C such that h2 (z) = . Define g(z) = on G. We have
1 − df(z) 1 − h(a)h(z)
g(a) = 0, g(G) ⊆ B(0, 1) and g is injective, holomorphic, i.e. g ∈ F. By computation, f(a) = 0 implies

f(a) − d (1 − |d|2)f 0 (a) (1 − |d|2)c |h0(a)| c(1 + |d|)

|h(a)|2 = = |d| < 1, |h0(a)| = = p , |g0 (a)| = = p > c,
1 − df(a) 2|h(a)| 2 |d| 1 − |h(a)|2 2 |d|

a contradiction.
z−w
Remarks. (1) In step 3, we implicitly used the fact that if w, z ∈ B(0, 1), then ∈ B(0, 1). One can
1 − wz
see this is true by properties of Möbius maps or by expanding |z − w|2 < |1 − wz|2 to get it equivalent to
(1 − |w|2)(1 − |z|2) > 0. In step 3, since d, f(z) ∈ B(0, 1), these imply h(z) ∈ B(0, 1) and g(z) ∈ B(0, 1).

(2) Using the Schwarz lemma, it can be proved that the f in the Riemann mapping theorem is unique! If
f and g are two such functions, then f ◦ g−1 and g ◦ f −1 maps B(0, 1) to B(0, 1) and maps 0 to 0. Also,
f 0 (a)/g0 (a) = (f ◦ g−1 )0 (0) > 0 and g0 (a)/f 0 (a) = (f ◦ g−1)0 (0) > 0. Since their product is 1, one of them is
at least 1, say f 0 (a)/g0(a) = (f ◦ g−1 )0 (0) ≥ 1. By Schwarz lemma, we must have f ◦ g−1 equals the identity
function so that f = g.

65
 Presentation Project XI – Applications of the Stone-Weierstrass Theorem

Now we present some applications of the Weierstrass Approximation Theorem and the Stone-Weierstrass
theorem. These theorems are used often to prove certain facts for continuous functions by first showing them
to be true for the polynomial case, then completing the continuous case by taking limits as illustrated below.
Z 1
Theorem. Let f be a real-valued continuous function on [0, 1] such that f(x)xn dx = 0 for every n =
0
0, 1, 2, . . .. Then f(x) = 0 for every x ∈ [0, 1].
X
k
Proof. For a polynomial P (x) = anxn , we have
n=0
Z 1 Z 1 X
k X
k Z 1
f(x)P (x) dx = f(x) an xn dx = an f(x)xn dx = 0.
0 0 n=0 n=0 0

By the Weierstrass approximation theorem, there is a sequence of polynomials {Pn} converging to f in

C([0, 1]). Let M = max{|f(x)| : x ∈ [0, 1]}. Then as n → ∞,
d∞(fPn , f 2) = max |f(x)Pn(x) − f 2 (x)| ≤ M max |Pn(x) − f(x)| = M d∞ (Pn, f) → 0,
x∈[0,1] x∈[0,1]

i.e. {fPn } converges to f 2 in C([0, 1]). By the integration theorem for uniform convergence in Math 301,
Z 1 Z 1 Z 1
f 2 (x) dx = lim f(x)Pn (x) dx = lim f(x)Pn (x) dx = 0, which implies f(x) = 0 for every x ∈
0 0 n→∞ n→∞ 0
[0, 1].

Fubini’s Theorem for Continuous Functions. Let K = [a, b] × [c, d] and f ∈ C(K). Then
Z bZ d Z dZ b
f(x, y) dydx = f(x, y) dxdy.
a c c a

Proof. The polynomial case follows because

Z bZ dX m Xn Xm Xn Z b Z d  Z d Z m X
bX n
i j i j
aij x y dydx = aij x dx y dy = aij xi yj dxdy.
a c i=0 j=0 i=0 j=0 a c c a i=0 j=0

For the general case, by the Heine-Borel theorem, K is compact. The collection A of all polynomials of two
X
m X n
variables P (x, y) = aij xi yj on K is easily checked to be an algebra. Any two points on K can be
i=0 j=0
separated by P1(x, y) = x or P2 (x, y) = y. Also A contains P (x, y) = 1, which vanishes at no point of K.
For every f ∈ C(K), by the Stone-Weierstrass theorem, there is a sequence {Pn(x, y)} in A converging to f
in C(K). Then as n → ∞, εn = d∞ (Pn, f) = max |Pn(x, y) − f(x, y)| → 0. So, as n → ∞,
(x,y)∈K
Z bZ d Z bZ d Z b Z d
Pn(x, y) dydx − f(x, y) dydx ≤ |Pn(x, y) − f(x, y)| dydx ≤ εn (b − a)(d − c) → 0,
a c a c a c
Z bZ d Z bZ d
which implies the equation f(x, y) dydx = lim Pn(x, y) dydx. Similarly, we also have the
a c Z n→∞ a c
Z dZ b d Z b
equation f(x, y) dxdy = lim Pn(x, y) dxdy. From the polynomial case, we have
c a n→∞ c a
Z b Z d Z b Z d Z d Z b Z d Z b
f(x, y) dydx = lim Pn(x, y) dydx = lim Pn (x, y) dxdy = f(x, y) dxdy.
a c n→∞ a c n→∞ c a c a

66
 Presentation Project XII – Applications of the Stone-Weierstrass Theorem

Below we will look at some problems that require the Stone-Weierstrass theorem for their solutions.

Question 1. Show that for any continuous function f : [0, 1] → R and ε > 0, there is a function of the
form g(x) = C0 + C1x4 + · · · + Cn x4n for some nonnegative integer n and C0, C1, . . . , Cn ∈ Q such that
|g(x) − f(x)| < ε for all x ∈ [0, 1].

Solution. Observe that h : [0, 1] → [0, 1] defined by h(x) = x4 is a homeomorphism. By the Stone-
Weierstrass theorem (or the Weierstrass approximation theorem) and the density of Q, for every ε > 0, there
is a polynomial p(x) = C0 + C1x + · · · + Cn xn with Ci ∈ Q (i = 0, 1, . . ., n) such that kp − f ◦ h−1k∞ < ε.
Let g = p ◦ h, then kg − fk∞ < ε.
Z 1
Question 2. Let f : [0, 1] → R be a continuous function. Prove that lim (n + 1) xn f(x) dx = f(1).
n→∞ 0

Solution. Let p(x) = a0 + a1 x + · · · + anxn be a polynomial. We have

Z 1 k
X n+1
lim (n + 1) xn p(x) dx = lim aj = p(1). (∗)
n→∞ 0 n→∞
j=0
n + j +1

So the result holds for any polynomial. For the continuous function f, let ε > 0, by the Stone-Weierstrass
theorem (or the Weierstrass approximation theorem), there is a polynomial p such that kf − pk∞ < ε/3. By
Z 1
(*), there exists N ∈ N such that n ≥ N implies (n + 1) xnp(x) dx − p(1) < ε/3. Then
0
Z 1  Z 1  Z 1
(n + 1) x f(x)dx − f(1) ≤ (n + 1) x |f(x) − p(x)|dx + (n + 1) xn p(x)dx − p(1) +|p(1) − f(1)|
n n
0 0 0
≤ ε/3 + ε/3 + ε/3 = ε.

Z 1
Question 3. Let φ1 , φ2, φ3, . . . be nonnegative continuous functions on [0, 1] such that lim xk φn(x) dx
n→∞ 0
Z 1
exists for every k = 0, 1, 2, . . .. Show that lim f(x)φn (x) dx exists for every continuous function f on
n→∞ 0
[0, 1].

Solution. By the Stone-Weierstrass theorem (or the Weierstrass approximation theorem), there exists a
Z 1
sequence of polynomial pn converging uniformly on [0, 1] to f. For k = 0, since the limit of φn exists,
Z 1 0

we have a constant M > 0 such that for all n ∈ N, φn ≤ M . For every ε > 0, there exists j ∈ N such
Z 1 0

that kf − pj k < ε/(3M ). Since pj φn converges, it is a Cauchy sequence. So there exists N ∈ N such that
Z 1 0 Z
1
ε nZ 1 o
for m, n ≥ N implies pj φn − pj φm < . It suffices to show fφn is a Cauchy sequence. For
0 0 3 0
n, m ≥ N, we have
Z 1 Z 1 Z 1 Z 1 Z 1 Z 1 Z 1 Z 1
fφn − fφm ≤ fφn − pj φn + pj φn − pj φm + pj φm − fφm
0 0 0 0 0 0 0 0
Z 1 Z 1
ε ε ε ε
< kf − pj k∞ φn + + kf − pj k∞ φm ≤ M+ + M = ε.
0 3 0 3M 3 3M

67
 Presentation Project XIII – Application of the Stone-Weierstrass Theorem

Let D = {(x1, x2, . . . , xn) ∈ Rn : x21 + x22 + · · · + x2n ≤ 1} be the closed unit ball in Rn . Also, let
T = {(x1 , x2, . . . , xn) ∈ Rn : x21 + x22 + · · · + x2n = 1} be the boundary (unit sphere) of D. We will prove the
following famous

Brouwer Fixed Point Theorem. Any continuous function G : D → D has a fixed point, i.e. there exists
a x ∈ D such that G(x) = x.
Below by a smooth mapping, we mean an infinitely differentiable function. In the proof of the Brouwer
Fixed Point Theorem, some clever computations allowed us to get the smooth case. The Stone-Weierstrass
Theorem is needed to reduce the continuous case to the smooth case.
We first establish two lemmas.

Lemma 1. There does not exist any smooth mapping f : D → T such that f(x) = x for all x ∈ T.
Proof. Assume such f = (f1 , f2, . . . , fn ) : D → T exists. Then for all x ∈ D, f12 (x) + f22 (x) + · · ·+ fn2 (x) = 1.
Taking partial derivative with respect to xi on both sides for i = 1, 2, . . ., n and putting the n equations in
matrix form, we have  ∂f1   
· · · ∂fn  
∂x1 (x) ∂x1 (x) 2f1(x) 0
 . .   .  .
 .. . .. ..   ..  = ..  .

∂f1 ∂fn 0
(x) · · · ∂x (x) 2fn(x)
| ∂xn {z n }| {z }
=J (x) nonzero vector

Observe that the matrix J(x) has a nonzero null space. So det J(x) = 0. Let ω0 = f1 df2 ∧ · · · ∧ dfn on D.
Then dω0 = df1 ∧ df2 ∧ · · · ∧ dfn = det J(x) dx1 ∧ dx2 ∧ · · · ∧ dxn = 0. For x ∈ T, f(x) = x implies fi (x) = xi
for i = 1, 2, . . ., n. By Stokes’ Theorem,
Z Z Z Z
0= dω0 = ω0 = x1 dx2 ∧ · · · ∧ dxn = dx1 ∧ dx2 ∧ · · · ∧ dxn = volume(D) > 0,
D T T D
contradiction.

Lemma 2. Any smooth mapping g : D → D has a fixed point.

Proof. Assume g has no fixed point. For x ∈ D, let f(x) be the point on T at which the ray from g(x) to
x intersects T. Thenp f(x) = x + tu, where u = (x − g(x))/kx − g(x)k and f(x) · f(x) = 1. Solving for t,
we get t = −x · u + 1 − x · x + (x · u)2. Since g(x) 6= x, u is a smooth mapping of x. Since x · x ≤ 1, so
1 − x · x + (x · u)2 ≥ 0. The zero case would require x · x = 1 and 0 = x · u = (x · x − x · g(x))/kx − g(x)k.
Then x · x − x · g(x) = 1 − x · g(x) > 0 by the Cauchy-Schwarz inequality. Hence 1 − x · x + (x · u)2 > 0.
Then t is a smooth mapping of x. So f : D → T is a smooth mapping with f(x) = x for all x ∈ T, which
contradicts lemma 1.

Finally, we will prove the Brouwer Fixed Point Theorem. Let G : D → D be continuous. Assume
G(x) 6= x for all x ∈ D. By the extreme value theorem, the continuous function kG(x) − xk attains a
minimum value m > 0. By the Stone-Weierstrass theorem, for ε = m/2 > 0, there exists a polynomial
function P : Rn → Rn such that for all x ∈ D, we have |G(x) − P (x)| < ε. Let Q(x) = P (x)/(1 + ε). Then
|P (x)| ≤ |G(x)| + ε ≤ 1 + ε.
So Q is a continuous map from D to D. Also,
|G(x) − Q(x)| ≤ |G(x) − P (x)| + |P (x) − Q(x)| < ε + |P (x)|ε/(1 + ε) ≤ 2ε = m.
Then Q : D → D is continuous, but for all x ∈ D, since |G(x) − x| ≥ m and |G(x) − Q(x)| < m, so Q(x) 6= x,
which contradicts lemma 2.
As an application, squeeze a round map of D = U SA to a ball (without tearing off any bit) and put
the ball in D = U SA. Then there is at least one point in the map lying directly over the place it repesents
in the map. There is no uniqueness assertion in the Brouwer Fixed Point Theorem.

68
 Presentation Project XIV – Weierstrass Approximation Theorem

For an infinitely differentiable function, such as cos x, we can compute the values of the function using
its Taylor series most of the time. However, if the function is only continuous, but not differentiable, then
there is no Taylor series. The following theorem asserts that for every continuous function f(x) on a closed
and bounded interval, there is a sequence of polynomials Pn(x) converging uniformly on the interval to f(x).

Weierstrass Approximation Theorem. Let f: [0, 1] → R be continuous. For every ε > 0, there is a
polynomial P (x) such that |P (x) − f(x)| < ε for all x ∈ [0, 1]. Taking ε = 1/n, we get a sequence of
polynomials Pn(x) converging uniformly on [0, 1] to f(x).

The theorem is also true for every closed and bounded interval [a, b] by substituting x = (t − a)/(b − a)
and compact subsets of R. (Note for ε = 10−12, P (x) will agree with f(x) to at least 10 decimal places for
all x ∈ [0, 1]. Thus, we can find f(x) to 10 decimal places by computing P (x).)

To prove the Weierstrass approximation theorem, we will introduce the Bernstein polynomials of a
continuous function.
Xn  
k
n k
Definition. For each n ∈ N, the n-th Bernstein polynomial of f(x) is fn (x) = f x (1 − x)n−k.
n k | {z }
k=0 | {z } polynomial
number
Examples. Find the n-th Bernstein polynomial of f(x) = 1, g(x) = x and h(x) = x2 respectively.
Xn  
n n k n−k
Solution. The binomial theorem asserts that (x + a) = x a . Differentiating x, then multiplying
k
k=0
x
by , and repeating these operations, we get
n

X    
x X k2 n k n−k
n n
k n k n−k
(x + a)n−1x = x a and [(n − 1)(x + a)n−2x + (x + a)n−1 ] = x a .
n k n n2 k
k=0 k=0

n  
X n
X  
n k k n k
Setting a = 1 − x, we get fn (x) = x (1 − x)n−k = 1, gn(x) = x (1 − x)n−k = x and
k n k
k=0 k=0

n
X  
k2 n k (n − 1)x2 + x
hn (x) = 2
x (1 − x)n−k = .
n k n
k=0

As n → ∞, we get fn (x) → f(x), gn (x) → g(x) and hn (x) → h(x).

Formula. We have

n    n   k   k  2  n 
X k 2 n k X
x− x (1 − x)n−k = x2 − 2x + xk (1 − x)n−k
n k n n k
k=0 k=0
(n − 1)x2 + x x − x2
= x2 − 2x2 + = .
n n

Proof of Weierstrass Approximation Theorem. For any ε > 0, since f is uniformly continuous on
ε
[0, 1], there is δ > 0 such that for every x, w ∈ [0, 1], |x − w| < δ ⇒ |f(x) − f(w)| < . By the extreme value
2
M
theorem, |f(x)| ≤ M for some M . Choose integer n > 2 . Let P (x) be the n-th Bernstein polynomial of f.
εδ
69
 (†)
1 1 1
For every x ∈ [0, 1], we have 0 ≤ (x − )2 = x2 − x + , which implies x − x2 ≤ . Next define S =
2 4 4
n k o 1 k 2(•)
k: k = 0, 1, . . ., n and x − < δ . Note k 6∈ S implies 2 (x − ) ≥ 1. Then for every x ∈ [0, 1],
n δ n

X n   Xn  
n k n−k k n k
|f(x) − P (x)| = f(x) x (1 − x) − f( ) x (1 − x)n−k
k n k
k=0 k=0
| {z }
=1 by examples
n     
k  n k
X Xn
k n k
= f(x) − f( ) x (1 − x)n−k ≤ f(x) − f( ) x (1 − x)n−k
n k n k
k=0 k=0
X   X  
k n k k n k
≤ f(x) − f( ) x (1 − x)n−k + f(x) − f( ) x (1 − x)n−k
n k n k
k∈S k6∈S
X ε n X  
n k
≤ xk (1 − x)n−k + 2M x (1 − x)n−k
2 k k
k∈S k6∈S
   
εX n k X 1
n
k 2 n k
≤ x (1 − x)n−k + 2M 2
(x − ) x (1 − x)n−k
2 k δ n k
k=1 k6∈S | {z }
by(•)
2
ε x−x ε M ε ε
≤ + 2M 2
≤ + 2
< + = ε.
2 nδ
| {z } 2 2nδ
| {z } 2 2
by formula by(† )

70
 Presentation Project XV – Which Banach Spaces are Hilbert Spaces?

In this project, we will answer the interesting question“Which Banach spaces are Hilbert spaces?” We
will do this by establishing the following

Jordan-von Neumann Theorem. A normed space is an inner product space if and only if the parallelo-
gram holds. Similarly, a Banach space is a Hilbert space if and only if the parallelogram holds.

The second statement follows easily from the first statement. Also, we know inner product spaces satisfy
the parallelogram law kx + yk2 + kx − yk2 = 2(kxk2 + kyk2 ). So we just need to prove the converse, namely
for a normed space V satisfying the parallelogram, from its norm k · k, we can define an inner product h·, ·i
so that V becomes an inner product space with kxk2 = hx, xi for all x ∈ V.

In the case the scalar field is K = R, let V be a normed space satisfying the parallelogram law. (Note
in inner product space kx + yk2 = kxk2 + 2hx, yi + kyk2.) This motivate us to define

hx, yi = (kx + yk2 − kxk2 − kyk2 )/2.

We need to prove (i) hx + y, zi = hx, zi + hy, zi; (ii) for all r ∈ R, hrx, yi = rhx, yi; (iii) hx, yi = hy, xi and
(iv) hx, xi = kxk2. From the definition of hx, yi, we immediately have

hx, yi = hy, xi, hx, xi = kxk2 and h0, yi = 0. (1)

So (iii) and (iv) are true. Also, if kxn − xk → 0 (i.e. xn → x) in V, then for all y ∈ V, hxn, yi → hx, yi
because first we have k(xn + y) − (x + y)k = kxn − xk → 0, kxnk − kxk ≤ kxn − xk → 0 and so

kxn + yk2 − kx + yk − kxnk2 + kxk2

|hxn, yi − hx, yi| = → 0. (2)
2
By the parallelogram law,

kx + zk2 ky + zk2 kx + y + 2zk2 kx − yk2 kx − yk2 kxk2 kyk2 kx + yk2

+ = + and − − = . (3)
2 2 4 4 4 2 2 4
By the definition of h·, ·i and (3), we have

kx + zk2 kxk2 kzk2 ky + zk2 kyk2 kzk2 kx + y + 2zk2 kx + yk2

hx, zi + hy, zi = − − + − − = − − kzk2
2 2 2 2 2 2 4 4
x+y 2 x+y 2 x + y 
= +z − − kzk2 = 2 ,z .
2 2 2
(4)
Let y = 0 in (4), then hx, zi = 2hx/2, zi. Applying this to the last part of (4), we get

hx, zi + hy, zi = hx + y, zi, (5)

which is (i). Also, let y = −x, then hx, zi + h−x, zi = 0, so (*) h−x, zi = −hx, zi.

For (ii), setting y = nx in (5), we can get by induction that hkx, zi = khx, zi for k = 0, 1, 2, 3 . . .. Then
using (*), we get the cases k = −1, −2, −3, . . . as well. For w ∈ V and m ∈ N, letting x = w/m, we have

k kD w E k Dw E D w E D kw E
hw, zi = m ,z = m ,z = k ,z = ,z . (6)
m m m m m m m
So for all r ∈ Q and w, z ∈ V, we have hrw, zi = rhw, zi.
Finally, for any c ∈ R, there are rn ∈ Q converging to c. We have rnw → cw and so taking limits of both
sides of hrnw, zi = rn hw, zi, by (2), we get hcw, zi = chw, zi, which is (ii). This complete the case K = R.

71
 In the case K = C, if V is an inner product space with the inner product denoted by (·, ·), then
kx + yk2 = kxk2 + 2 Re(x, y) + kyk2 . So Re(x, y) = (kx + yk2 − kxk2 − kyk2 )/2 = hx, yi as before. Let
(x, y) = a + ib, then Im(x, y) = b = Re{−i(a + ib)} = Re{−i(x, y)} = − Re(ix, y) = −hix, yi. If V is a
normed space over C satisfying the parallelogram law, this motivated us to define

(x, y) = Re(x, y) + i Im(x, y) = hx, yi − ihix, yi, (7)

where hx, yi = (kx + yk2 − kxk2 − kyk2 )/2. We have

hix, iyi = (kix + iyk2 − kixk2 − kiyk2 )/2 = (kx + yk2 − kxk2 − kyk2 )/2 = hx, yi (8)

and (ix, y) = hix, yi − ih−x, yi = hix, yi + ihx, yi = i(hx, yi − ihix, yi) = i(x, y). (9)

We need to prove (v) (x + y, z) = (x, z) + (y, z); (vi) for all r ∈ C, (rx, y) = r(x, y); (vii) (x, y) = (y, x) and
(viii) (x, x) = kxk2.
For (v), by (7) and (5),

(x + y, z) = hx + y, zi − ihi(x + y), zi = hx, zi + hy, zi − ihix, zi − ihiy, zi = (x, z) + (y, z).

For (vi), if r ∈ R, then by (7) and (i), (rx, y) = hrx, yi − ihirx, yi = rhx, yi − rihix, yi = r(x, y). If
r = a + ib, then by (7), (v) and (9),

(rx, y) = (ax, y) + (ibx, y) = a(x, y) + b(ix, y) = a(x, y) + ib(x, y) = r(x, y).

For (vii), by (7), (8) and (*),

(x, y) = hx, yi − ihix, yi = hx, yi − ih−x, iyi = hy, xi + ihx, iyi = hy, xi − ihiy, xi = (y, x).

For (viii), by (vii), (x, x) = (x, x) ∈ R and so (x, x) = Re(x, x) = hx, xi = kxk2.

72