You are on page 1of 83

Contents

1 Measure Theory 2
1.1 Riemann Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Lebesgue Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Lebesgue Null Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Cantor Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Translates and Non-Measurability . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.6 Measurable Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.7 Limits of Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.8 Simple Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.9 Lebesgue Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.10 Lp spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

1.11 Complete Normed Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24


1.12 L∞ -space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.13 Lp -space Containment Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

1.14 Improper Riemann Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29


1.15 Functional Analysis on Lp -Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

2 Fourier Analysis 37
2.1 Complex Valued Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.2 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3 Approximating with Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.4 Averaging Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.5 Coefficients of Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.6 Point-wise Convergence of Naive Fourier Series . . . . . . . . . . . . . . . . . 63
2.7 Inner Product and Hilbert Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.8 Hilbertion Fourier Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
2.9 Behavior of Fourier Series Near Jump Discontinuities . . . . . . . . . . . . 81

1
1 Measure Theory

1.1 Riemann Integration


We shall begin by recalling a few basic definitions:
Definition 1.1.1. A norm k · k : X → R is a mapping that satisfies the following properties:
1.) ∀x ∈ X, kxk ≥ 0 (Non-negativity)
2.) kxk = 0 ⇔ x = 0 (Non-degeneracy)
3.) ∀x, y ∈ X, kx + yk ≤ kxk + kyk (Sub-additivity)
4.) kαxk = |α|kxk (Homogeneity)
Definition 1.1.2. A sequence (xn )∞
n=1 ⊂ X is said to be Cauchy if ∀ > 0, ∃N > 0 such that ∀m, n ≥
N , kxn − xm k < 
Definition 1.1.3. A normed vector space X is said to be complete if every Cauchy sequence (xn )∞ n=1
converges to an element x ∈ X. That is, lim kxn − xk = 0. Any normed vector space that is also complete
n→∞
is called a Banach Space

With this small bit of review in place, we are now in a position to introduce the notion of a Riemann
sum. Naively, we recall the notion of a Riemann sum from introductory calculus courses. While the
following is a bit more abstract, the general idea is nonetheless the same. Consider a Banach space X and
real numbers a < b with a partition on [a, b] of the form
P = {a = x0 < x1 < . . . < xn−1 < xn = b}
Definition 1.1.4. A Riemann sum is any sum of the form
n
X
S(f, P ) = f (ti )(xi − xi−1 ) for ti ∈ [xi−1 , xi ]
i=1

Since this is similar to the concept of an integral, we want to be able to define what it means for an
object to be integrable, especially in the context of the Riemann sums above.
Definition 1.1.5. Let a < b be real numbers and P = {a = x0 < x1 < . . . < xn−1 < xn = b} be a
partition on this interval. We say that a function f : [a, b] → X is Riemann Integrable if there is a fixed
element x ∈ X such that ∀ > 0, ∃δ > 0 such that whenever L(P ) = max (xi − xi−1 ) < δ then
i=1,...,n

kS(f, P ) − xk < 
Z b
It is easy to show that any such point of convergence above is unique, and is often denoted as x = f.
a
In general however, this definition is not easy to apply to particular functions. Luckily there is a theorem
to help us avoid using this definition.
Theorem 1.1.6 (Cauchy Criterion for Integrability). If f : [a, b] → X is a function, then f is Riemann
integrable ⇔ ∀ > 0∃δ > 0 such that for any pair of partitions P, Q on [a, b] such that L(P ), L(Q) < δ,
then
kS(f, P ) − S(f, Q)k < 

2
Proof. (⇒) Let  > 0 and assume that f is a Riemann integrable function. By the definition of Riemann
integrability, there exists x ∈ X and δP , δQ for partitions P, Q respectively such that L(P ) < δP , L(Q) < δQ
implies that
 
kS(f, P ) − xk < kS(f, Q) − xk <
2 2
Define δ = min{δP , δQ }, then
kS(f, P ) − S(f, Q)k ≤ kS(f, P ) − xk + kx − S(f, Q)k
 
< + =
2 2
(⇐) Let (δn )∞ 1
n=1 ⊂ R be such that kS(f, P ) − S(f, Q)k < 2n whenenver L(P ), L(Q) < δn and assume that
δ1 ≥ δ2 ≥ . . .. We can do this by assumption. Now for each n let Pn be a partition with L(Pn ) < δn and
let xn = S(f, Pn ) be a fixed Riemann sum. Then for m < n we have
kxn − xm k ≤ kxn − xn−1 k + kxn−1 − xn−2 k + . . . + kxm+1 − xm k
= kS(f, Pn ) − S(f, Pn−1 )k + . . . + kS(f, Pm+1 ) − S(f, Pm )k
1 1 1
< n−1 + n−2 + . . . + m by construction
2 2 2
1
< m−1
2
It is easy to see from here that (xn )∞
n=1 must then be a Cauchy sequence, and be completeness of X there
exists x ∈ X such that lim kxn − xk = 0. We claim that this is the unique x which will satisfy Riemann
n→∞
integrability for f .
1
Indeed, let  > 0 and choose a sufficiently large n such that 2n−1 < 2 . Let P be an arbitrary partition
with L(P ) < δn where δn is the same as that defined above. Then for any Riemann sum S(f, P ) we have
that
kS(f, P ) − xk ≤ kS(f, P ) − xn k + kxn − xk
1 1
< n + n by Cauchy, and hypothesis
2 2
1
= n−1 < 
2

However we very quickly encounter problems with Riemann integration. There are some functions that
are defined everywhere, but simply cannot be integrated.
(
1 if t ∈ Q
Proposition 1.1.7. Define χQ = . Then we claim that χQ is not Riemann integrable on
0 if t ∈/Q
the interval [0, 1].

Proof. Let δ > 0 and P {a = x0 < x1 < . . . < xn−1 < xn = b} be a partition of [0, 1] such that L(P ) < δ.
We note that the rational numbers are dense in [0, 1] whereas the irrational numbers are dense in (0, 1).
Hence consider the following two Riemann sums
n
X n
X
SR (χQ , P ) = χQ (ri )(xi − xi−1 ) SI (χQ , P ) = χQ (yi )(xi − xi−1 )
i=1 i=1

3
Where ri ∈ [xi−1 , xi ] ∩ Q and yi ∈ [xi−1 , xi ] ∩ R\Q . We can expand these as
n
X
SR (χQ , P ) = (xi − xi−1 )
i=1
= (x1 − x0 ) + (x2 − x1 ) + . . . + (xn − xn−1 )
= −x0 + xn = 1
SI (χQ , P ) = 0

Thus for any 0 <  ≤ 1, all δ > 0 result in


|SR (χQ , P ) − SI (χQ , Q)| ≥ 
By the Cauchy Criterion it then follows that χQ is not Riemann integrable.

Since such a relatively simple function could not be integrated, we are immediately motivated to find
some other way to perform an integration. This encourages the following section.

1.2 Lebesgue Integration


Note to self, input stuff here:
We want to know if there’s a simple way to characterize Borel sets. Naively, define G = G0 , and define
Gn+1 = ((Gn )δ )σ . This yields a properly nested set of intervals
G0 ( G1 ( G2 ( . . .
It is a non-trivial matter
S∞ to show that the inclusions are proper, and is beyond the scope of this course.
Naively, define Gω = n=1 Gn . We wish that this would be a characterization of Borel sets, but it turns
out that this simply isn’t true. In order to properly examine the characteristics of Borel sets, we require
some further definitions from set theory.
Definition 1.2.1. A well-ordering on a countable set S is a relation ≤ that satisfies

1. (total ordering) ∀s, t ∈ S, s ≤ t, t ≤ s


2. (anti-reflexive) if s ≤ t, t ≤ s then s = t
3. (transitive) if r ≤ s, s ≤ t then r ≤ t
4. (well-ordering) ∀A ⊆ S\∅ , A has a minimal element
Definition 1.2.2. An ordinal number is the cardinality associated with sets of the natural numbers.
Successor ordinals are defined as those that increase the ordinal number by 1, limit ordinals are multiples
of the first countable ordinal.

Since our above attempt to examine the likeness of Borel sets failed with a naive definition, we consider
the following construction. If α is a successor ordinal, define S
mathcalGα+1 = ((Gα )δ )σ . If on the other hand α is a limit ordinal, define Gα = β<α Gβ .
S
It turns out that B(R) = a∈[0,Ω] Gα where we’ve taken [0, Ω] to be the set of all ordinal numbers.

4
1.3 Lebesgue Null Sets
Definition 1.3.1. A subset N ⊂ R is a (Lebesgue) null set if λ∗ (N ) = 0

Proposition 1.3.2. If N ⊂ R is a null set, then it is Lebesgue measurable

Proof. We begin by observing that, via the non-negativity and monotonicity of the Lebesgue outer measure,
if M ⊆ N then 0 ≤ λ∗ (M ) ≤ λ∗ (N ) = 0. This implies that all subsets of a null set are also a null set. Now
let E ⊆ R be any set. We immediately see that

λ∗ (E) ≤ λ∗ (E ∩ N ) + λ∗ (E\N ) by sub-additivity

However, we note that E ∩ N ⊂ N and so λ∗ (E ∩ N ) = 0. Also by monotonicity, λ∗ (E\N ) ≤ λ∗ (E). This


gives the opposite inequality, and we conclude that N is a measurable set.

[
Corollary 1.3.3. If A1 , A2 , . . . is a countable collection of null subset of R, then Ai is also a null set
i=1

Proof. Using σ-subadditivity, we have


∞ ∞
!
[ X

λ Ai ≤ λ∗ (Ai )
i=1 n=1
=0

Hence the union is also a null se

Corollary 1.3.4. The rationals Q are a null set.

Proof. Recall that the rationals are countable, so we can enumerate


S∞ them as Q = {q1 , q2 , q3 , . . .}. Since
any finite set has measure 0, define Ai = {qi }. Then since Q = i=1 Ai we conclude by the above corollary
that λ∗ (Q) = 0

1.4 Cantor Set


Define the following sets,
C0 = [0, 1]
   
1 2
C1 = 0, ∪ , 1 = C0 \I1,1
3 3
       
1 2 1 2 7 8
C2 = 0, ∪ , ∪ , ∪ , 1 = C0 \I2,1 ∪I2,2
9 9 3 3 9 9
..
.
Cn = Cn−1 \S2n −1 I n>1
k=1 n,k

where we’ve defined In,k to be the open middle third of the k th interval of Cn−1 .

5
T∞
Definition 1.4.1. The Cantor set is defined as C = i=1 Ci

Proposition 1.4.2. The following properties hold for the Cantor set

1. C 6= ∅

2. C is compact and as a result, Lebesgue measurable

3. λ(C) = 0

4. C is uncountable

Proof. 1. Each Cn is a closed set since it is a finite union of closed sets. Furthermore, as a collection,
the Cn form a proper nested arrangement of sets.

C1 ⊃ C2 ⊃ C3 ⊃ . . .

[
Thus by the finite intersection property, since each Cn 6= ∅ we have that C = 6= ∅
n=1

2. Filled in after assignment

3. Filled in after assignment

4. Every element of [0, 1 admits a ternary expansion. That is,



X
t = e1 e2 e3 . . . = ek 3−k , ek ∈ {0, 1, 2}
k=1

However we recall that ternary expansions are not necessarily unique. Notice that
   
1 2
I1,1 = , = a = 0.1e1 e2 . . . a satisfiesP
3 3

where P = {el 6= 2 for some l > 1} and not all el = 0, l > 1. In general we can define
 
In,k = a = 0.e1 e2 . . . en−1 1en+1 . . . a satisfies Q

n−1
( )
X el
where Q = el 6= 2, l > n. not all el = 0, l > n.k = 1 + . Now from before we recall that
2k−1
l=1
the Cantor set is defined as
C = [0, 1]\S∞ S2n −1
In,k
n=1 k=1

Now we can bijectively map the Cantor set onto the set of all infinite binary strings via the mapping
φ : C → {0, 1}N defined by e e e 
1 2 3
φ(0.e1 e2 e3 . . .) = , , ,...
2 2 2
And we know that all infinite binary strings is indeed uncountable. Thus we can conclude that the
Cantor set is uncountable as required.

6
1.5 Translates and Non-Measurability
Definition 1.5.1. If E ⊂ R, x ∈ R, define the translate of E by x as
 
x+E = x+y y ∈E

Proposition 1.5.2. Let E ⊂ R, x ∈ R, then the following holds

1. λ∗ (x + e) = λ∗ (E)

2. E ∈ L(R) ⇒ x + E ∈ L(R)

3. If E ∈ L(R) then λ(x + E) = λ(E)

Proof. 1. We have that (In )∞ ∞


n=1 is an open interval if and only if {x + In }n=1 is an open interval cover
for x + E. Thus we have that for any open interval (a, b) ∈ R that x + (a, b) = (x + a, x + b) and
`(x + a, x + b) = b − a. Using the definition of λ∗ (E) and λ∗ (x + E) the result immediately follows.

2. If A ∈ P(R) we have

λ∗ (A ∩ (x + E)) + λ∗ (A\(x+E) ) = λ∗ (−x + A ∩ (x + E)) + λ∗ (−x + A\(x+E) )


= λ∗ ((−x + A) ∩ E) + λ∗ ((−x + A)\E )
= λ∗ (−x + A) since E was measurable

Thus the result follows.

3. (1) and (2) together immediately imply the desired result.

While this was an easy result to show, it is nonetheless very important to showing the existence of
non-measurable sets. It turns out that the axiom of choice will be essential to the construction, and the
absence of this axiom leads to all sets being measurable.

Theorem 1.5.3. Given a > 0 a real number, ∃E ⊂ (−a, a) such that E ∈


/ L(R). That is, E is not a
measurable set.

Proof. We begin by defining an equivalence relation of (−a, a) by x ∼ y if x − y ∈ Q ∩ (−2a, 2a). This is


indeed an equivalence relation since

1. (Reflexivity) If x ∈ (−a, a), then x − x = 0 ∈ Q, so x ∼ x

2. (Symmetry) If x ∼ y then we know that x − y ∈ Q. But x − y = −(y − x) ∈ Q and so y − x ∈ Q,


thus y ∼ x

3. (Transitivity) Assume x ∼ y, y ∼ z; that is, x − y, y − z ∈ Q. Then x − z = (x + y) − (y − z) ∈ Q by


the fact that Q is closed under addition. Hence x ∼ z.

7
Notationally, denote ∀x ∈ (−a, a) the equivalence class, define as
   
[x] = y ∈ (−a, a) x ∼ y = y ∈ (−a, a) x − y ∈ Q = (x + Q) ∩ (−a, a)

Invoking the axiom of choice, we can choose E ⊂ (−a, a) such that

1. If x, y ∈ E and x 6= y, then x  y

2. |(E ∩ [x]| = 1, ∀[x] ∈ (−a, a)

Let us note that E ∩ (r + E) = ∅, ∀r ∈ Q. Thus let (rk )∞


k=1 be an enumeration of Q ∩ (−2a, 2a). Then
we have

[
(−a, a) ⊂ (rk + e) ⊂ (−3a, 3a) (1)
n=1

The first inclusion, note that if x ∈ (−a, a), E ∩ [x] = {xE }. Then x − xE ∈ Q since x ∼ xE , and since
x, xE ∈ (−2a, 2a) we have that |x − xE | < 2a. The second inclusion follows easily since E ⊂ (−a, a), rk ∈
(−2a, 2a) then ∀x ∈ E, x + rk ∈ (−3a, 3a). Let us prove that E ∈ / L(R).
First, if λ∗ (E) = 0, then E is a null set and hence measurable. Then λ∗ (rk + E) = λ∗ (E) = 0, ∀K.
and hene
X∞ ∞
X

0 < 2a = λ((−a, a)) ≤ λ (rk + E) = 0=0
k=1 k=1

which is a contadiction.
Thus if E were measurable, ∃α > 0 such that λ(E) = α. Of course, as E ⊂ (−a, a), α ≤ λ((−a, a)) =
2a < ∞. But then by (1), we have

6a = λ((−3a, 3a))

!
[
≥λ (rk + E)
k=1
n
X n
X
= λ(rk + E) = λ(E)
k=1 k=1
= nα nα

Hence 6a ≥ nα, ∀n which implies that α = 0. But we’ve already seen that this leads to a contradiction.
Thus E ∈
/ L(R).

1.6 Measurable Functions


 
For notation sake, define f −1 (A) = x ∈ R f (x) ∈ A . This allows us to easily make the following
definition

8
Definition 1.6.1. Let f : R → R, then we say that f is measurable if it satisfies f −1 ((α, ∞)) =

x ∈ R f (x) > α ∈ L(R)

Proposition 1.6.2. 1. If f : R → R is continuous, then f is measurable

2. If A ⊂ R, the indicator function on A is measurable if and only if A is measurable

Proof. 1. Since (α, ∞) is an open function and f is continuous then we know that f −1 ((α, ∞)) is also
an open set. Furthermore, we showed that all open sets are measurable and our result follows.

2. Fix α ∈ R. Then 
∅
 if α ≥ 1
χA ((α, ∞)) = A if 0 ≤ α < 1

if α < 0

R
Since ∅, R ∈ L(R) we find that χA is measurable if and only if A is measurable, as required.

Proposition 1.6.3. Let f : R → R. Then the following are equivalent

1. f is measurable

2. f −1 ((−∞, α]) ∈ L(R), ∀α ∈ R

3. f −1 (((−∞, α)) ∈ L(R), ∀α ∈ R

Proof. (1) ⇒ (2)


 
−1
f ((−∞, a]) = x ∈ R f (x) ≤ α

= R\



x∈R f (x)≥α
 

= R\f −1 ((α,∞))


[ 1
(2) ⇒ (3) Note that (−∞, α) = (−∞, α − ] and hence
n
n=1


−1
[ 1
f ((−∞, α)) = f −1 ((−∞, α − ))
n
n=1

so since f −1 ((−∞, α)) ∈ L(R) we have that f −1 ((−∞, α]) ∈ L(R)


(3) ⇒ (1) As before, for each α ∈ R we have that f −1 ([α, ∞)) ∈ L(R) given condition (3). Proceeding
as we did in the previous case, we obtain (1).

9
Proposition 1.6.4. Let f : R → R then f is measurable if and only if f −1 (A) ∈ L(R), ∀A ∈ B(R)

Proof. (⇐) Trivial (⇒) Assume that f is a measurable function. Consider an open set given by (a, b).
Then we can write (a, b) = (a, ∞) ∩ (−∞, b) and so f −1 ((a, b)) = f −1 ((−∞, b)) ∩ f −1 ((a, ∞)). Since we
assumed that f is measurable, then we can conclude that f −1 ((a, b)) ∈ L(R).

[
Now consider a general open set G. We know that we can write G = (ai , bi ) and so
i=1


[
−1
f (G) = f −1 ((ai , bi )) ∈ L(R)
i=1

since we know from above that f −1 ((ai , bi )) ∈ L(R).


 
−1
Let MF = A ⊂ R f (A) ∈ L(R) . From above, MF contains all open sets. Let us see that MF
is a σ-algebra of sets.

1. ∅, R ∈ MF since f −1 (∅) = ∅ ∈ L(R) and f −1 (R) = R ∈ L(R)

2. A ∈ MF ⇒ R\A ∈ MF since f −1 (R\A ) = R\f −1 (A)


∞ ∞ ∞
!
[ [ [
3. If A1 , A2 , . . . ∈ MF then ∈ MF since f −1 AI = f −1 (Ai )
i=1 i=1 i=1

Thus we have that MF is a σ-algebra which contains all open sets. Thus B(R) ⊂ MF since B(R) is
the smallest σ-algebra containing all open sets.

Proposition 1.6.5. Let f, g : R → R be measurable fucntions, with c ∈ real and φ : R → R a contin-


uous function. Then the following are true, which tells us that the set of measurable functions admits a
continuous functional calculus:

1. cf

2. f + g

3. φ f

4. f g

Proof. 1. Fix α ∈ R and consider the function





 f −1 (( αc , ∞)) if c>0

−1
R if c = 0, α ≤ 0
(cf ) (α, ∞) =


 ∅ if c = 0, α > 0
f −1 ((−∞, α ))

if c<0
c

It is immediately obvious (since ∅, R ∈ L(R) that since f is measurable, this function is measurable.

10
2. Let us begin by enumerating the rational numbers as Q = (qk )∞
k=1 . If α ∈ R then we have
 
−1
(f + g) ((α, ∞)) = x ∈ R f (x) + g(x) ≥ α
 
= x ∈ R f (x) > α − g(x)
∞ 
[   
= x ∈ R f (x) > rk ∩ x ∈ R rk > α − g(x)
k=1
[∞
f −1 ((rk , ∞)) ∩ g −1 ((α − rk , ∞)) ∈ L(R)

=
k=1

since f, g are both measurable functions.

3. Let α ∈ R, then
(φ · f )−1 ((α, ∞)) = f −1 φ−1 ((α, ∞))


but φ continuous implies that φ−1 ((α, ∞)) is open. So this function is measurable.

4. Let us observe that


(f + g)2 − (f − g)2
fg =
4
By (2), each of f + g and f − g are measurable. Furthermore, h(x) = x2 is a continuous function.
Thus (1) and (3) imply that f g ∈ L(R) as required.

Corollary 1.6.6. If f : R → R is measurable, then so too is |f |, f + , f − where we define

|f |(x) = |f (x)|, f + = max{f (x), 0}, f − (x) = min{−f (x), 0}

and we note that |f | = f + − f − .

Proof. —f— is measurable since the absolute value function is continuous. Furthermore,
|f | + f |f | − f
f + (x) = , f − (x) =
2 2
which are measurable by the above proposition.

1.7 Limits of Functions


For the purposes of this section, it will be useful to consider the option of limits of functions tending to
infinity. This motivates the following definitions.

Definition 1.7.1. We define the extended real numbes as R̄ = R ∪ {±∞}. If f : R → R̄ then we say
that f is and “extended real valued function.” We say that f : R → R̄ is measurable if f −1 ((α, ∞]) ∈
L(R), ∀α ∈ R

11
Note that is it easy to show, by methods similar to those above, that

f : R → R̄ measurable ⇔ f −1 (A) ∈ L(R), ∀A ∈ B(R), and f −1 ({±∞}) ∈ L(R)



\
Of particular use may be the identity that f −1 ({∞}) = f −1 ((n, ∞])). And that the set of measurable
n=1
extended-real valued function is denoted by M̄(R).
A pitfall that one must be wary to avoid is that M̄(R) does not necessarily form an algebra. It is
possible to encounter things such as ∞ − ∞ which are not terribly well defined. In the case of continuous
functions, we will adopt the convention that if φ is a continuous function, and lim exists and is possibly
t→±∞
±∞, then φ(±∞) is that limit. We note that under this convention, the absolute value of a measurable
function is always defined.

Proposition 1.7.2. Let (fn )∞


n=1 ⊂ B(R). Then the following functions are all measurable

1. sup fn where (supn∈N fn ) (x) = supn∈N fn (x)


n∈N

2. inf fn which is defined similar to that above


n∈N
!
3. lim sup fn = lim sup fk
n→∞ n→∞ k≥n
 
4. lim inf fn = lim inf fk
n→∞ n→∞ k≥n

Proof. 1.
 −1  
sup fn ([−∞, α]) = x ∈ R sup fn (x) ≤ α
n∈N n∈N

\  
= x ∈ R fn (x) ≤ α
n=1
\∞
= fn−1 ([−∞, α]) ∈ L(R)
n=1

2. This follows from the fact that (inf n∈N fn ) = − (supn∈N (−fn ))

3. Let gn = supk≥n fk , gk ∈ M̄(R) Furthermore, g1 ≥ g2 ≥ g3 ≥ . . ., so

lim sup fn = inf sup fn


n→∞ k≥n n∈N k≥n

= inf gn ∈ gn by 2
n∈N

4. This fact follows from lim inf n→∞ fn = − (lim supn→∞ (−fn ))

12
Corollary 1.7.3. If (fn )∞
n=1 ⊂ M(R) and f (x) = lim fn (x) ∀x ∈ R exists, then f ∈ M̄(R)
n→∞

Proof. If f = lim fn exists in a pointwise fashion, then f = lim sup fn = lim inf fn
n→∞ n→∞ n→∞

1.8 Simple Functions


Definition 1.8.1. Let φ : R → R. We say that φ is a simple function provided that
 
φ(R) = φ(x) x ∈ R is finite

Furthermore, if A ⊂ R, then φ : A → R is said to be simple on A if


 
φ(A) = φ(x) x ∈ A is finite

Note that if φ(A) is finite, then we can write φ(A) = {a1 < a2 < . . . < an } implying that each ai is
distinct and the list is exhaustive. In turn, define the preimage of each element as Ei = φ−1 ({ai }). With
Xn
this in hand, we can write φ = ai χEi . It is also immediately obvious that Ei ∩ Ej = ∅, ∀i 6= j.
i=1

Proposition 1.8.2. Let A be a measurable set and φ : A → R be simple, with phi(A) = {a1 < a2 < . . . < an }.
Then φ is a measurable function if and only if Ei = φ−1 ({ai }) is measurable as a set forall i = 1, . . . , n.

Proof. (⇒) Each {ai } is closed, and hence a Borel set. Thus Ei = φ−1 ({ai }) ∈ meas by an earlier
proposition.
(⇐) We saw earlier that a set being measurable satisfies

Ei ∈ L(R) ⇔ χEi ∈ M(R)

We have that M(R) is closed under scalar multiplication and pointwise addition. The result then follows.

For simplicity of notation, we introduce the following set


 
S(A) = φ : A → R φ is simple and measurable

 
+
S (A) = φ ∈ S(A) φ(x) ≥ 0, ∀x ∈ A

Definition 1.8.3. If φ ∈ S + (A) for measurable A, then with φ(A) = {a1 < a2 < . . . < an } and Ei =
φ−1 ({ai }), then we define the proto-Lebesgue integral
n
X
IA (φ) = ai λ(Ei )
i=1

13
For the purpose of this definition, we accept the convention that αλ(E) = ∞ if α > 0 and λ(E) = ∞
and 0λ(E) = regardless of the value of λ(E)
Proposition 1.8.4. Let A ∈ L(R), φ, ψ ∈ S + (A), c ≥ 0. Then

1. cφ ∈ S + (A) and IA (cφ) = cIA (φ)


2. φ + ψ ∈ S + (A) and IA (φ + ψ) = IA (φ) + IA (ψ)
3. if φ ≤ ψ in a pointwise fashion, then IA (φ) ≤ IA (ψ)

Proof. 1. Trivial
2. Let φ = {a1 < a2 < . . . < an } , ψ = {b1 < b2 < . . . < bm }, and define Ei = inf φ({ai }), Fi = ψ −1 ({bi }),
so that
X n Xm
φ= ai χEi , ψ = bj χFj Ei ∩ Ej = Fi ∩ Fj = ∅
i=1 j=1
  [
Let {γ1 < γ2 < . . . < γl } = ai + bj i = 1, . . . , n; j = 1, . . . m and define Dk = Ei ∩ Fj .
{i,j|ai +bj =γk }
Then we have that
n
X m
X
φ+ψ = ai χEi + bj χFj
i=1 j=1
Xn Xm
= (ai + bj )χEi ∩Fj
i=1 j=1
`
X X
= γk χEi ∩Fj
k=1 {i,j|ai +bj =γk }
`
X
= γk χDk
k=1

Now we have that


`
X
IA (φ + ψ) = γk λ(Dk )
k=1
X` X
= γk λ(Ei ∩ Fj )
k=1 {i,j|ai +bj =γk }
` X
X m
= (ai + bj )λ(Ei ∩ Fj )
k=1 j=1
Xn Xm m
X n
X
= ai λ(Ei ∩ FJ ) + bj λ(Ei ∩ Fj )
i=1 j=1 j=1 i=1
Xn m
X
= ai λ(Ei ) + βi λ(Fi )
i=1 j=1

14
1.9 Lebesgue Integration
Definition 1.9.1. Let f ∈ M̄ (A). Then we say that f is Lebesgue integrable on A if both A f + , A f − <
R R

∞, where f +R = max{f, 0}, fR− = max {−f, 0}. Furthermore, in this case we define the Lebesgue integral
of f on a as A f = A f − A f −
R +

 
For notational simplicity, let L̄(A) = f : A → R̄ f measurable, Lebesgue integrable

Lemma 1.9.2. Let f ∈ M̄ (A), then

 
1. f ∈ L̄ ⇒ λ x ∈ A f (x) = ±∞

 
R
2. A |f | = 0 ⇔ λ x ∈ A f (x) 6= 0

 
Proof. 1. Suppose that f ∈ L̄(A). Let E± = x ∈ A f (x) = ±∞ . Then for any n ∈ N, we have that
nχE+ ≤ f and thus
Z Z
nλ(E+ ) = nχE+ ≤ f+ < ∞ since f ∈ L̄(A)
A A

 + ) = 0. A similar argument holds to show that λ(E− ) = 0. Thus E+ ∪ E− =


Thus we must have λ(E

x ∈ A f (x) = ±∞ , and the union of null sets is a null set, so the result follows.

R
2. (⇒) Suppose that A |f | = 0, for n ∈ N let
 
1
En = x ∈ A |f (x)| ≥
n

We note that |f | = f + − f − , so En = f −1 [ n1 , ∞] is a measurable set. Thus 1


 
n χEn ≤ |f |, and we
can conclude that Z Z
1 1
λ(En ) = χEn ≤ |f | = 0
n A n A

which allows us to conclude that λ(En ) = 0. Let



[    
E= En = x ∈ A |f (x)| > 0 = x ∈ A f (x) 6= 0
n=1

then since a countable union of null sets is null, we have that λ (E) = 0 as required.

15
 
+
(⇐) Suppose that λ x ∈ A f (x) 6= 0 = 0. Let φ ∈ S|f | (A), so φ is simple, measurable, and

φ ≤ |f |. Let φ(A) = {α1 < . . . < αn }, and Ei = φ−1 ({α}i ). So that


n
X
φ= αi χEi Ei ∩ Ej = ∅, i 6= j
i=1

If αi > 0, then αi χEi ≤ φ ≤ |f |, so


   
Ei ⊂ x ∈ A |f (x)| ≥ αi ⊂ x ∈ A f (x) > 0

  n
X
However, x ∈ A f (x) > 0 is a null set, and so λ(Ei ) = 0. Thus we know that IA (φ) = αi λ(Ei ) =
i=1
0, and Z
|f | = sup IA (φ) = 0
A φ∈Sf+ (A)

Definition 1.9.3. Let f, g ∈ M̄ (A). We say that f = g almost everywhere on A if


 
λ x ∈ A f (x) 6= g(x) =0

 
Let L(A) = f : A → R f is measurable, Lebesgue integrable

Corollary
R 1.9.4. If f ∈ L̄(A) then ∃f0 ∈ L(A) such that f = f0 almost everywhere on A. Moreover
A |f − f0 | = 0

(
f (x) if f (x) 6= ±∞
Proof. Define f0 (x) = . By the above Lemma, this function satisfies all the neces-
0 otherwise
sary properties.

Theorem 1.9.5. If f, g ∈ L(A) and c ∈ R, then

R R
1. cf ∈ L(A) and A cf =c Af
R R R
2. f + g ∈ L(A) and A (f + g) = A f + A g
R R
3. |f | ∈ L(A) and A f ≤ A |f |. In fact, f ∈ L(A) ⇔ f ∈ M (A) and |f | ∈ L(A)

Proof. 1. Straightforward

16
2. First note that (f + g)+ ≤ f + + g + , (f + g)− ≤ f − + g − . This imples that
Z Z Z Z
(f + g)± ≤ (f ± + g ± ) = f± + g± < ∞
A A A A

Claim If h, k, φ, ψ ∈ L+ (A) with h − k = φ − ψ then


Z Z Z Z
h− k= φ− ψ
A A A A

Proof. We note that h − k = φ − ψ ⇒ h + ψ = φ + k. Since these are all positive function, we then
have that
Z Z Z
h+ ψ = (h + ψ) by MCT
A A ZA
= (φ + k)
ZA Z
= φ+ k
A A
R R
Hence since A ψ, A k < ∞, we can subtract to get the desired result

Now we have that f + g = (f + g)+ − (f + g)− = (f + − f − ) + (g + − g − ) = (f + + g + ) − (f − + g − ).


Now we can use the above claim to get that
Z Z Z
(f + g) = (f + g) + (f + g)−
+
A ZA AZ
= (f + g ) − (f − + g − )
+ +

ZA Z AZ Z

= +
f + +
g − f − g−
ZA Z A A A

= f+ g
A A

3. Note that |f | = f + − f − . Hence


Z Z Z Z

+
|f | = (f − f ) = +
f − f− < ∞
A A A A

Thus we can conclude |f | ∈ L(A). (⇒) Now


Z Z Z
f = f+ + f−
A
ZA A
Z
≤ +
f + f−
A A
Z Z
= f+ + f − since both are positive
ZA A

= |f |
A

17
+ − +, −
R R
R(⇐) If f ∈ M (A) and |f | ∈ L(A), then we have that f , f ≤ |f |, which implies that Af Af ≤
A |f | ≤ ∞, and so we can conclude that f ∈ L(A)

We note that if E ∈ \P(R) L(R), say E ⊂ (a, b), then f = χ(a, b)\ − χE . We have |f | = χ(a,b) ∈
E
M ((a, b)) but f ∈
/ M ((a, b)).

Lemma 1.9.6 (Fatou’s Lemma). If {f } n ⊂ M̄ + (A) for A ∈ L(R) then


Z Z
lim inf fn ≤ lim inf fn
A n→∞ n→∞ a

Proof. Let gn = inf fk so that g1 ≤ g2 ≤ g3 ≤ . . ., which we know is true since gk+1 = inf fn ≥ inf fn =
k≥n n≥k+1 n≥k
fk , and lim inf gn = lim inf fn by definition of lim inf. Hence, by the monotone convergence theorem,
Z n→∞ Z n→∞
R R
lim inf fn = lim gn . However, since gn ≤ fk whenever k ≥ n, we have that A gn ≤ A fk for k ≥ n.
A n→∞ n→∞ A
Hence Z Z Z
gn ≤ inf fk ≤ lim inf fk
A k≥n A n→∞ A
The above two statements give the desired inequality.

The inequality in the Lemma may be strict. For example, consider A = [0, 1] and define the function
fn = χ(0, 1 ) . It can be shown that
n
Z Z
lim inf fn = 0, lim inf fn = 1
A n→∞ n→∞ A

Definition 1.9.7. Let A ∈ L(R), (fn )∞


n=1 ⊂ M (A), and f : A → R, then we write

lim fn f almost everywhere on A


n→∞
 
if N = x∈A lim fn (x) 6= f (x), or DNE
n→∞

Note that if (fn )∞


n=1 ⊂ M (A) satisfies f = n→∞
lim fn almost everywhere on A, then f is measurable
on A. Furthermore, the MCT and Fatou’s Lemma remain valid if pointwise convergence is replaced by
convergence almost everywhere. Also, the advantage to using almost everywhere convergence is that though
(fn )∞
n=1 ⊂ M (A), we may have f = lim fn ∈ M̄ (A). If however f is Lebesgue integrable, then we replace
n→∞
f by f0 ∈ M (A) where f = f0 almost everyhwere, and f0 = limn→∞ fn almost everywhere.

Theorem 1.9.8 (Lebesgue’s Dominated Convergence Theorem). If (fn )∞ n=1 ⊂ L(A), f : A → R, and
g ∈ L+ (A) = {g ∈ L(A)|g ≥ 0} such that f = lim fn almost everywhere on A and |fn | ≤ g almost
n→∞ Z Z
everywhere on A for all fn , then f is measurable and Lebesgue integrable with f = lim fn
A n→∞ A

Note that we refer to the function g as the integrable majorant for (fn )∞
n=1 .

18
  ∞
[
Proof. Let N = x∈A lim fn (x) 6= f (x), or DNE ∪ {x ∈ A||fx (x)| > g(x)}, then N is clearly a
n→∞
n=1
null-set and so we consider A\N , and all assumptions will hold pointwise on this refined set. Thus for the
rest of this proof we will relabel A 7→ A\N .
R
Now f = lim fN is measurable and |f | = lim |fn | ≤ g and it fllows that A |f | < ∞, so f is Lebesgue
n→∞ R R n→∞
integrable, since A |f | < A g < ∞. Since fn + g ≥ 0 by our assumption, we have that

g + f = lim (fn + g) = lim inf (fn + g)


n→∞ n→∞

which tells us that


Z Z Z
g+ f= lim inf (g + fn )
A A A n→∞ Z
≤ liminfn→∞ (g + fn )
Z AZ

= g + lim inf fn
A n→∞ A

Z Z Z Z
Similarily, it can be shown that since g − fn ≥ 0, then g− f ≤ g − lim sup fn . These two
A A A n→∞ A
inequalities tell us that Z Z Z Z
lim inf fn ≥ f, lim sup fn ≤ f
n→∞ A A n→∞ A A
We then have that Z Z Z
lim sup fn ≤ f ≤ lim inf fn
n→∞ A A n→∞ A
Z Z
which implies that the limit exists and lim fn = f
n→∞ A A

1.10 Lp spaces
Let A ∈ L(R) (usually an open interval or the entire space).
Proposition 1.10.1. For f ∈ L(A), define
Z
kf k1 = f
A

This norm has the following two properties: If f, g ∈ L(A), c ∈ R then

1. kcf k1 = |c|kf k1

2. kf + gk ≤ kf k1 + kgk2

Proof. Let f, g ∈ L(A), c ∈ R, then


R R R
1. kcf k1 = A |cf | = A |c||f | = |c| A |f | = |c|kf k1

19
2.
Z
kf + gk1 = |f + g|
ZA
≤ (|f | + |g|)
ZA Z
= |f | + |g|
A A

We are tempted to call this the one-norm; however we note that we are currently lacking non-degeracy.
It turns out that this is problematic since, as we saw earlier,
Z
kf k1 = |f | = 0 ⇔ f = 0 almost everywhere
A

We can resolve this problem by defining an equivalence relation on L(A).

f ∼ g ⇔ f = g almost everywhere on A

In this regard, now define


L1 (A) = L(A)/∼
Hence we think of L1 (A) as the space of integrable functions on A where we agree that f = g if f = g
almost everywhere on A. Under this equivalence class, it’s easy to check that k · k1 defines a norm on
L1 (A).
Warning: Since {x} is null, for x ∈ A, then value ”f (x)” for f ∈ L1 (A) is meaningless. That is, we
lose the notion of point-wise evaluation.
Next we wish to discuss the topic of convergence in L1 (A). If (fn )∞
n=1 ⊂ L1 (A) and f ∈ L1 (A) such
+
that lim fn = f almost everywhere, and there is g ∈ L1 (A) such that |fn | ≤ g then we can conclude that
n→∞

lim kfn − f k1 = 0
n→∞

Proof. First note that |f | = lim |fn | ≤ g almost everywhere, so


n→∞

|fn − f | ≤ |fn | + |f |
≤ 2g

so we can use Lebesgues dominated convergence theorem to show that


Z Z
n→∞
kfn − f k1 = |fn − f | −−−→ 0=0
A A

since fn → f as n → ∞ almost everywhere

20
We are forced to ask ourselves, if lim kfn −f k1 = 0 does this imply that lim fn = f almost everywhere?
n→∞ n→∞
The answer is no.
Example: Let A = [0, 1], and

f1 = χ[0, 1 ] , f2 = χ[ 1 ,1] , f3 = χ[0, 1 ] , f4 = χ[ 1 , 2 ] , f5 = χ[ 2 ,1] , . . .


2 2 3 3 3 3

Then check that Z Z


lim |fn − 0| = lim fn = 0
n→∞ [0,1] n→∞ [0,1]

lim fn (x) exists for no x ∈ [0, 1]


n→∞

In particular, we see that lim |fn | = 0 does not exist almost everywhere on [0, 1].
n→∞

Definition 1.10.2. Let 1 < p < ∞, then we define the conjugate index to p as the number q which
satisfies p1 + 1q = 1.

Note that if p = 1 then we get that q = ∞, and similarily if p = ∞ then q = 1.

Definition 1.10.3. For f ∈ M (A), define


Z 1
p
p
kf kp = |f |
A

where we note that the integrand could be infinite.

Definition 1.10.4. Define  


p
Lp (A) = f ∈ M (A) |f | ∈ L(A) /∼

where ∼ is the ”almost everywhere” equivalence relation.

Lemma 1.10.5. If 1 < p < ∞, and q is the conjugate index with a, b ∈ [0, ∞) then

ap bq
ab ≤ +
p q

and moreover, equality holds precisely when ap = bq .

Proof. If ab = 0 we are done, thus assume that ab > 0. Let 0 < α < 1, and φ : (0, ∞) → R such that
φ(t) = αt − tα . Then we have that
 
1
φ0 (t) = α − αtα−1 = α 1 −
t1−α

and consider that

φ0 (t) < 0 if 0 < t < 1, φ0 (t) > 0 if t > 1, φ0 (t) = 0 ⇔ t = 1

21
Thus by applications of the mean value theorem we see that

φ(t) = αt − tα
≥ φ(1)
=α−1
α
⇒ t ≥ αt + (1 − α)

ap
with equality holding only when t = 1. Now set t = . Then we have
bq

ap ap

≤α + (1 − α)
bq bq
⇒ apα bq(1−α) ≤ αap + (1 − α)bq
1 1 ap
And we can take α = p so 1 − α = q and we note that equality holds only when bq =1

Proposition 1.10.6 (Holder’s Inequality). If f ∈ Lp (A) for 1 < p < ∞ and g ∈ Lq (A) for q the conjugate
to p, then f g ∈ L1 (A) in the almost everywhere sense, and

kf gk1 ≤ kf kp kgkq

Moreover, equality holds ⇔ kgkq |f | = kf kp |g| almost everywhere.

Proof. If kf kp kgkq = 0 then it is the case that f g = 0 almost everywhere, and so the inequality is trivially
true. Thus assume that kf kp kgkq > 0. Let for almost everywhere x ∈ A, define

|f (x)| |g(x)|
a(x) = , b(x) =
kf kp kgkq

Then by the lemma above


|f (x)g(x)|
= a(x)b(x)
kf kp kgkq
a(x)p b(x)q
≤ +
p q
|f (x)|p |g(x)|q
= +
p kf kpp q kgkqq

Note that since f, g are each measurable almost everywhere, we have that f g is measurable, hence so
too is |f g|. This being the case, we can always make sense of the integral, and so

|f |p |g|q
Z Z
1
|f g| ≤ kgkq + kf kp
kf kp kgkq A A p q
kf kpp kgkqq
= +
p kf kpp q kgkqq
1 1
= + =1
p q

22
Z
This implies that kf gk1 = |f g| ≤ kf kp kgkq
A
To see when equality holds, use the equality case from the Lemma and follow through.

Proposition 1.10.7 (Minkowski’s Inequality). If 1 < p < ∞, f, g ∈ Lp (A), A ∈ L(R) then we can conclude
that f + g ∈ Lp (A) and
kf + gkp ≤ kf kp + kgkp
Moreover, equality holds only if ∃c1 , c2 > 0 such that c1 f = c2 g almost everywhere

Corollary 1.10.8. k·kp is a norm on Lp (A)

Proof of Corollary. Absolute value scalar homogeneity follows from

kcf kp = |c| kf kp

Non-degeneracy follows from


Z
kf kp = 0 ⇔ |f |p = 0 ⇔ |f |p = 0 almost everywhere ⇔ f = 0 almost everywhere
A

From which it immediately follows that f = 0 in Lp (A). Subadditivity will follow for free from Minkowski’s
inequality.

Proof of Minkowski’s Inequality. First, if f, g ∈ Lp (A) we have

|f + g|p ≤ (2 max {|f |, |g|})p


= 2p max {|f |p , |g|p }
≤ 2p (|f |p + |g|p )
Z
⇒ |f + g|p ≤ 2p (|f |p + |g|p )
A
Z Z 
= 2p |f |p + |g|p < ∞
A A

Hence we can conclude that f + g ∈ Lp (A). Now we can move on to proving subadditivity. First observe
the following:

|f + g|p = |f + g||f + g|p−1


≤ |f ||f + g|p−1 + |g||f + g|p−1 (*)

1 1
Notice that |f + g|p−1 ∈ Lq (A) where q is the conjugate index of p. This is true since p + q = 1
1 1 p−1 p
by definition of the conjugate, and so q = 1− p = p and so we can conclude that q = p−1 , and so
(p − 1)q = p and we have that
Z Z
q
|f + g|p−1 = |f + g|(p−1)q
A ZA
= |f + g|p < ∞
A

23
Hence we have
Z Z Z
p p−1
|f + g| ≤ |f ||f + g| + |g||f + g|p−1
A A A
Z  1 Z 1 Z  1 Z 1
p q p q
(Holder0 sInequality) ≤ |f |p |f + g|(p−1)q + |g|p |f + g|(p−1)q (**)
A A A A
p p
= kf kp kf + gkpq + kgkp kf + gkpq
  p
= kf kp + kgkp kf + gkpq

p
Hence, dividing both sides by kf + gkpq we have
p− pq
kf + gkp ≤ kf kp + kgkp
p
kf + gkp ≤ kf kp + kgkp since p − =1
q

We have equality at (*) for the following

equality at (*) ⇔ |f + g| = |f | + |g| almost everywhere ( or |f + g| = 0, trivial)


⇔ f (x)g(x) > 0 almost everywhere ∀x ∈ A

That is, they are ”sticking out in the same direction”. We have equality at (**) for the following

kf + gkqp kf + gkqp
equality at (**) ⇔ |f |p = |f + g|(p−1)q = |g|p
kf kpp kgkpp

Note that we required that kf + gkp 6= 0, but in the case when kf + gkp = 0 is trivial. The previous
two conditions can both hold simultaneously if there are c1 , c2 > 0 such that

c1 f = c2 g almost everywhere

Also one can check that this condition implies that kf + gkp = kf kp + kgkp . Note that we choose to use
two scalar instead of one in order to allow for the case when one of f, g is equal to 0 almost everywhere.

1.11 Complete Normed Spaces


1. If X is a R-vector space with norm k · k, then (X, k · k) is complete if every Cauchy sequence in
this norm convergence to a point in X. That is, if (xn )∞
n=1 such that  > 0 there is n ∈ N so that
∀m, n > n , ⇐ kxn − xm k <  then ∃x ∈ X such that lim kxn − xk = 0
n→∞

2. If (X, k · k) is a normed vector space, then it is complete if and only if for each sequence (xn )∞
n=1 ⊂ X,
X∞ X∞ XN
having that kxn k < ∞ implies that xn = lim xn exists and is in X.
N →∞
n=1 n=1 n=1

24
3. If (xn )∞ ∞
n=1 is a Cauchy sequence in X, then there is a subsequence (xnk )k=1 (where we have chosen
the indices such that they are in strictly increasing order), such that kxnk − xnk+1 k < 21k .

 
Theorem 1.11.1. Let A ∈ L(R), 1 ≤ p < ∞, then Lp (A), k·kp is complete

Proof. Let (fn )∞


n=1 ⊂ Lp (A) be a Cauchy sequence for k·kp . By dropping to a subsequence and relabelling
1
we may assume that kfn+1 − fn kp < n (Thus if we find a limit for the subsequence we obtain one for the
2
entire sequence). We let

X
f = f1 + (fn+1 − fn ) almost everywhere
n=1

We must show that the limit is real-valued almost everywhere, which will be true if |f |p ∈ L1 (A). For each
k ∈ N define
k
X
gk = |f1 | + |fn+1 − fn |
n=1

Then we see that gk ≥ 0 almost everywhere, and g1 ≤ g2 ≤ g3 ≤ . . . with almost everywhere limit


X
g = lim gk = |f1 | + |fn+1 − fn |
k→∞
n=1

By Minkowski’s inequality, we have

k
X
kgk kp ≤ k|f1 |kp + k|fn+1 − fn |kp
n=1
k
X 1
≤ kf1 kp +
2n
n=1
≤ kf1 kp + 1

 p
and thus gkp = kgk kpp ≤ kf1 kp + 1 . Also, g1p ≤ g2p ≤ . . . so by the monotone convergence theorem, we
Z Z
p
have that g ∈ L1 (A) with p
g = lim gkp . Hence g is real valued almost everywhere. Now
A k→∞ A

∞ p
X
p
|f | = f1 + (fn+1 − fN )
n=1
∞ p
X
≤ |f1 | + |fn+1 − fn |
n=1
≤ gp

25
so |f |p ∈ L1 (A) and f is real valued almost everywhere and f ∈ Lp (A). Now for n ∈ N,

!
X
kfn − f kp = fn − f1 + (fn+1 − fn )
n=1 p
= kfn − (f1 + (f2 − f1 ) + (f3 − f2 ) + . . .)kp

X
= (fk+1 − fk )
k=n+1 p

X
≤ kfk+1 − fk kp
k=n+1

X 1 n→∞
≤ −−−→ 0
2k
k=n+1

so lim fn = f
n→∞

1.12 L∞ -space
Definition 1.12.1. Let p= ∞, and take A ∈L(R). Then a function f ∈ M (A) is essentially bounded
on A if ∃c > 0 such that x ∈ A |f (x)| > c is null.

Note that for such a function f , we refer to the essential supremum, and use that notation
 
kf k∞ = ess − supx∈A |f (x)| = inf c ≥ 0 c is essential bound for f

 
1
Furthemore kf k∞ is an essential bound. Indeed, En = x ∈ A |f (x)| > kf k∞ + is null. That is,
n
∞  
1 [
λ(En ) = 0 since kf k∞ + must be an essential bound. Thus if we take E = En = x ∈ A |f (x)| > kf k∞ ,
n
n=1
we see that λ(E) = 0, so kf k∞ is also an essential bound.

Definition 1.12.2. Define L∞ (A) = {f ∈ M (A)|f is essentially bounded} / ∼. . Where we have taken
f ∼ g almost everywhere ⇔ f = g almost everywhere on A.

Proposition 1.12.3. If f, g ∈ L∞ (A), c ∈ R, then

1. kf k∞ = 0 ⇔ f = 0 almost everywhere

2. kcf k∞ = |c| kf k∞

3. kf + gk∞ ≤ kf k∞ + kgk∞

Thus we can conclude that L∞ (A) is a vector space, and k·k∞ is a norm.

26
Proof. 1. Straightforward (Exercise)
2. Straightforward (Exercise)
3. We will see that simultaneously, that kf k∞ + kgk∞ is an essential bound for |f + g|, and hence
kf k∞ + kgk∞ ≥ kf + gk∞ since it will be larger than the infimum of all essential bounds. Consider
   
x ∈ A |f (x) + g(x)| > kf k∞ + kgk∞ ⊆ x ∈ A |f (x)| + |g(x)| > kf k∞ + kgk∞

since |f (x)| + |g(x)| ≥ |f (x) + g(x)|. Then,


   
x ∈ A |f (x)| > kf k∞ ∪ x ∈ A| |g(x)| > kgk∞

Is a superset of the previous set, since logically if |f (x)| + |g(x)| > kf k∞ + kgk∞ then both f (x), g(x)
cannot be larger than their corresponding norm. However, the last set in the nesting is null, and
subsets of null sets are null. Thus we can conclude that
 
x ∈ A |f (x) + g(x)| > kf k∞ + kgk∞

is also a null set.

Note that if λ(A) < ∞ then for f ∈ L∞ (A) we have that kf k∞ = lim kf kp . This is a tough exercise,
p→∞
and the proof is omitted here.
Theorem 1.12.4. (L∞ (A), k·k∞ ) is complete and hence a Banach space

Proof. Let (fn )∞n=1 ⊂ L∞ (A) be an arbitrary Cauchy sequence. By dropping to a subsequence and rela-

1 X
beling if necessary, we may suppose that kfn+1 − fn k∞ < n . Then we let f = f1 + (fn+1 − fn ). We
2
n=1
first establish that f ∈ L∞ (A). Define
   
F = x ∈ A |f1 (x)| > kf1 k∞ , En = x ∈ A |fn+1 (x) − fn (x)| > kfn+1 − fn k∞

We immediately see that F, En are null, and so the union of all of these sets is null as well. Define

[
E=F ∪ En . Thus for x ∈ A\E we have
n=1

X
|f (x)| ≤ f1 (x) + (fn+1 (x) − fn (x))
n=1
X∞
≤ |f1 (x)| + (fn+1 (x) − fn (x))
n=1

X
≤ kf1 k∞ + kfn+1 − fn k∞
n=1
≤ kf1 k∞ + 1

27
Thus kf1 k∞ + 1 is an essential bound for f , and so f is essentially bounded. Clearly on A\E we have that
f ∈ M (A\E ), and so since E is null, we have that f ∈ L∞ (A).
Now lim kf − fn k∞ = 0 is proved in precisely the same manner as in the Lp (A) case.
n→∞

1.13 Lp -space Containment Relations


Let A = [a, b] be a bounded interval, so that λ([a, b]) = b − a.
Proposition 1.13.1. Let 1 ≤ p < r < ∞, then Lr ([a, b]) ⊂ Lp ([a, b]). Moreover we can show that
r−p
kf kp ≤ kf kr (b − a) rp

for f ∈ Lr ([a, b])

r
Proof. Let f ∈ Lr ([a, b]). Then |f |p ∈ L pr ([a, b]), since when we raise this to the p power, we retrieve |f |r
which we know is integrable. Thus by Holders inequality
Z Z
p
|f | = |f |p · 1
[a,b] [a,b]

≤ |f |p k1kq
r
p

r−p
where q is conjugate to pr . That is 1
r + 1
q = 1 which implies that 1
q = r . Then
p

kf kpp ≤ |f |p k1k r
r r−p
p
!p ! r−p
Z r Z r
r r
p
= (|f | ) p 1 r−p

[a,b] [a,b]
p r−p
= (kf krr ) r (b − a) r

r−p
= kf kpr (b − a) r

which is precisely what we wanted.

Remarks:

1. Note that L∞ ([a, b]) ⊂ Lp ([a, b]) for all 1 ≤ p < ∞.

2. Let A = R, [0, ∞) Then do we have, if 1 ≤ p < r < ∞ that

Lr (A) ⊂ Lp (A), Lp (A) ⊂ Lr (A)

Certainly we know that Lr ([0, ∞)) 6⊂ Lp ([0, ∞)), and consider the function
 
1
f (x) = max 1, 1 , x ∈ [0, ∞)
xp

28
Then f ∈ Lr ([0, ∞)) but f ∈
/ Lp ([0, ∞)).
1
Furthermore, we know that LP ([0, 1]) 6⊂ Lr ([0, 1]) and consider the function f (x) = 1 almost
xr
everywhere. Then f ∈ Lp ([0, 1]) but f ∈
/ Lr ([0, 1]). Which implies that

Lp ([0, ∞]) 6⊂ Lr ([0, ∞))

On assignment 3, it is shown that f : [a, b] → R is Riemann integrable implies that f is Lebesgue


integrable. Moreover,
Z Z b
f= f
[a,b],L a,R
 
Also, f : [a, b] is Riemann Integrable implies that λ x ∈ [a, b] f not continuous at x = 0 and that
f is bounded. The converse is also true, but the proof is tricky. To show this, prove that if the sets of
discontinuity
 
1
En = x ∈ [a, b] ∀δ > 0, ∃y, z ∈ (x − δ, x + δ) ∩ [a, b], |f (y) − f (z)| ≥
n

is null, then
( k k
)
X [
inf `(Ii ) E ⊂ Ii , Ii open interval, k ∈ N =0
i=1 i=1

1.14 Improper Riemann Integral

Definition 1.14.1. If f : [0, ∞) → R, then we say that f is improperly Riemman integrable if the
following two conditions hold

1. for any b > 0, f is Riemann integrable on [0, b]


Z b
2. lim f exists
b→∞ 0

Proposition 1.14.2. If f [0, ∞) → R is improper-Riemann integrable as defined above, and f (x) ≥ 0 for
x ∈ [0, ∞), then f is also Lebesgue integrable on [0, ∞) and moreover with Lebesgue integral
Z Z ∞ Z b
f= f = lim f
[0,∞) 0 b→∞ 0

Z Z b
Proof. From assignment 3, we have that f is Lebesgue integrable on each interval [0, b] with f= f.
[0,b] 0
Define fn = χ[0,n] Then f1 ≤ f2 ≤ . . . and lim fn = f in a pointwise sense. Thus be the monotone
n→∞

29
convergence theorem we have
Z Z
f = lim fn
[0,∞) n→∞ [0,∞)
Z
= lim f χ[0,n]
n→∞ [0,∞)
Z
= lim f
n→∞ [0,n]
Z n Z ∞
= lim f= f
n→∞ 0 0

Note that if we do not assume that f ≥ 0, then this may fail. That is, the improper Riemann integral can
exists, but f is not Lebesgue integrable on the whole unbounded interval [0, ∞) (Exercise). Take as a hint,
∞ ∞
X (−1)n X 1
the fact that converges, but does not. Furthermore, there is also an improper Riemann
n n
n=1 n=1
integral for unbounded functions f : (0, 1] → R where f is Riemann integrable on each [a, 1], a > 0. The
same facts regarding Lebesgue integrability holds.

1.15 Functional Analysis on Lp -Spaces


Definition 1.15.1. If X, Y are Banach spaces, a linear operator T : X → Y is bounded provided that
 
|||T ||| = sup kT xky x ∈ X, kxkX ≤ 1 < ∞

Note that if Y = R then T are linear operators. If f : X → R then f are called linear functionals. In
this context we will notate kf k∗ = |||f |||. We denote the space of linear functionals on X by X ∗ , and call
it the dual space. Furthermore,
 
|||T ||| = sup kT xkY kxkX ≤ 1 (*)
 
= sup kT xkY kxkX < 1 (**)

We show this by dual inequalities. Clearly (∗∗) ≤ (∗). To see that (∗) ≤ (∗∗) note that for kxkX ≤ 1 then
for 0 < η < 1 we have kηxkX = η kxkX < 1. and we have that kxkX = sup kηxkX . Thus (*) becomes
0<η<1
   
sup kT xkY kxkX ≤ 1 = sup kT (ηx)kY kxkX ≤ 1, 0 < η < 1

≤ (∗∗)

Proposition 1.15.2. Let X, Y be Banach spaces, and T : X → Y be a linear operator. Then the following
are equivalent

30
1. T is continuous

2. T is bounded; that is |||T ||| < ∞

3. T is Lipschitz, moreover, the Lipschitz constant is |||T |||

Proof. (1) ⇒ (2)


 
If T is continuous on X, then T is continuous at 0X . Since B1 (0Y ) = y ∈ Y ky − 0Y kY < 1 is
 
open in Y and T (0X ) = 0Y , then there is a neighbourhood Bδ (0X ) = x ∈ X kx − 0X kX < δ such
that T (Bδ (0X )) ⊂ B1 (0Y ). If kxkX < 1 then kδxkX = δ kxkX < δ so δx ∈ Bδ (0X ). Thus δ kT xkY =
kT (δx)kY < 1, and thus
 
|||T ||| = sup kT xkY kxkX < 1

1 1
= since kT xkY <
δ δ
<∞

(2) ⇒ (3)
First note that kT xkY ≤ |||T ||| kxkX for x ∈ X. Indeed this
 holds for 0X . If x = 6 0X we have that
1 1 1
x = 1 ≤ 1 and thus kT xkY = T x ≤ |||T |||. So we are done, since for
kxkX X kxkX kxkX Y
x, x0 ∈ X, we have
T x − T x0 Y = T (x − x0 ) Y ≤ |||T ||| x − x0 X
Which shows the Lipschitz property, with a Lipschitz bound of |||T |||. Note that it is routing to check that
|||T ||| is the smallest C ≥ 0 such that T x − T x0 Y ≤ C x − x0 Y .
(3) ⇒ (1)
This is quite trivial, since Lipschitz continuity implies uniform continuity which in turn implies conti-
nuity.

Note that we are commenting on linear functions on the Banach space X, thus if Γ is such an operator,
it is unsurprising to demand that

Γ(x + y) = Γ(x) + Γ(y), Γ(ax) = aΓ(x), ∀x, y ∈ X, α ∈ R

We say that an operator is bounded if it is finite in the operator norm, that is,
 
kΓk∗ = sup |Γ(x)| x ∈ X, kxk ≤ 1 < ∞

31
Theorem 1.15.3. Let 1 < p < ∞, A ∈ L(R), and q the conjugate index to p. Then for g ∈ Lq (A) we have
Z
Γg : Lp (A) → R given by Γg (f ) = fg
A

defines a bounded linear functional on Lp (A), with kΓgk∗ = kgkq

Note that in turns out that every bounded linear functional Γ : Lp (A) → R is of the form described in
the theorem Γ = Γg for some choice of g ∈ Lq (A). The proof of this statement requires the Radon-Nikodym
theorem. In the language of functional analysis we say that Lq (A) is the dual space of Lp (A) and write
this relation as Lp (A)∗ = Lq (A).

Proof. First, if g ∈ Lq (A), f ∈ Lp (A) then


R by Holder’s inequality we see that f g ∈ L1 (A), and hence f g is
integrable, so that statement Γg (f ) = A f g makes sense. It is obvious that Γg is linear since the Lebesgue
integral is linear. Again by Holder’s inequality, we have that for any choice of f ∈ Lp (A),
Z
|Γg (f )| = fg
A
Z
≤ |f g|
A
≤ kf kp kgkq by Holder’s inequality

and hence
 
kΓg k∗ = sup |Γg (f )| f ∈ Lp (A), kf kp ≤ 1
 
≤ sup kf kp kgkq f ∈ Lp (A), kf kp ≤ 1

Thus Γg is bounded with kΓg k∗ ≤ kgkq . Next we want to show that kΓg k∗ ≥ kgkq . Let us imitate the

p
1
 x>0
proof of the equality case in Holder’s inequality. Choose f = C |g| q sgn ◦ g where sgn(x) = 0 x=0

−1 x < 0

which is Borel measurable. As an exercise, show that sgn ◦ f ∈ M(A), if f ∈ M(A). Let us check that
f ∈ Lp (A).

Z
kf kpp = |f |p
ZA
q p
= C|g| p sgn ◦ g
AZ

= Cp |g|q |sgn ◦ g|p


ZA
p
=C |g|q since |sgn ◦ g| = 1
A
= C kgkqq < ∞
p

32
q
1
Hence f ∈ Lp (A) with kf kp = C kgkqp . If we choose C = q for g 6= 0 almost everywhere, then we
kgkqp
obtain kf kp = 1. Note that we can make this assumption on g since if g = 0 almost everywhere, then
Γg (f ) = 0 and hence kΓg k∗ = 0 = kgkq . Thus we have
 
kΓg k∗ = sup |Γg (f )| f ∈ Lp (A), kf kp ≤ 1
 
1 q
≥ Γg  q |g| sgn ◦ g
p 
p
kgkq
Z
1 q
= q |g| (sgn ◦ g)g
p

A kgk p
q
Z
1 q
+1
= q |g| p since (sgn ◦ g)g = |g|
p A
kgkq
 
1 q q 1 1
= q kgkq since + 1 = q + =q
p q p
kgkqp
q− pq
= kgkq
 
q 1 1
= kgkq since q − =q 1− =q =1
p p q

Hence kΓg k∗ ≥ kgkq and the two inequalities give us the desired equality. Thus we are finished.

Theorem Z 1.15.4. Let g ∈ L∞ (A), A ∈ L(R) such that 0 < λ(A) < ∞. Define Γg : L1 (A) → R by
Γg (f ) = f g. Then Γg is a bonded linear functional with kΓg k∗ = kgk∞
A

Proof. Let f ∈ L1 (A) and g ∈ L∞ (A). Then note that |g(x)| ≤ kgk∞ for almost every x ∈ A. This implies
that
|f (x)g(x)| = |f (x)||g(x)| ≤ |f (x)| kgk∞ sometimes called Holder’s inequality

Thus, we can proceed as follows


Z Z
|f g| ≤ |f | kgk∞
A A Z
= kgk∞ |f |
A
= kgk∞ kf k1

SInce we’ve shown that f g ∈ L1 (A), we have that


Z Z
fg ≤ |f g| ≤ kf k1 kgk∞
A A

33
As in the proof above , it folows that kΓg k∗ ≤ kgk∞ . Let us see that 
kΓg k∗ ≥ kgk∞ . Fist,
 if g = 0 almost
everywhere then this is trivial. Thus assume that kgk∞ > 0. Let B = x ∈ A |g(x)| > 0 so that λ(B) >
 
1
0 since g 6= 0 almost everywhere. It is possible that λ(B) = ∞. We let Bn = x ∈ A |g(x)| > kgk∞ − ,
n
so that λ(Bn ) 6= 0

1
fn = sgn ◦ gχBn
λ(Bn )
This then gives us that
Z
1
kf k1 = |sgn ◦ gχBn |
A λ(Bn )
Z
1
= χB
λ(Bn ) A n
=1

which in turn gives us that


 
kΓg k∗ = sup |Γg (f )| f ∈ L1 (A), kf k1 ≤ 1

≥ |Γg (fn )|
Z
1
= (sgn ◦ g)χBn g
A λ(Bn )
Z
1 1
= |g| since |g(x)| ≥ kgk∞ −
λ(Bn ) Bn n
Z  
1 1
≥ kgk∞ −
λ(Bn ) Bn n
1
= kgk∞ −
n
Theorem 1.15.5. Let 1 ≤ p < ∞, A ∈ L(R), φ ∈ L∞ (A). Then define Mφ : Lp (A) → Lp (A) by Mφ f = φf .
Then Mφ is a linear operator with
|||Mφ ||| = kφk∞

Proof. We must first check that φf ∈ Lp (A) for f ∈ LP (A)

|φf |p = |φ|p |f |p
≤ kφkp∞ |f |p
Z Z
⇒ p
|φf | ≤ kφkp∞
A A Z
p
= kφk∞ |f |p < ∞
A
⇒ φf ∈ Lp (A)

34
Also,
Z 1/p
p
kMφ f kp = |φf |
A
 Z 1/p
p p
≤ kφk∞ |f |
A
= kφk∞ kf kp

 
and hence |||Mφ ||| = sup kφf kp f ∈ Lp (A), kf kp ≤ 1
 
≤ sup kφk∞ kf kp f ∈ Lp (A), kf kp ≤ 1 = kφk∞

Let’s see that |||Mφ ||| ≥ kφk∞ . Let  > 0 and define
 
A = x ∈ A |φ(x)| > kφk∞ − 

By abuse of notation, φ is a representative of its equivalence class. Then A ∈ L(A) and λ(A ) > 0 by
definition of kφk∞ . Find A0 ⊂ A such that A0 ∈ L(A) and 0 < λ(A0 < ∞. Let
1
f= χA 0
λ(A0 )1/p 
Then
Z Z p
p 1
|f | = χ 0
0 1/p A
A A λ(A )
Z
1
= χ 0
0 ) A
A λ(A 
Z
1
= χA0
λ(A0 ) A 
=1
Z 1/p
p
This implies that kf kp = |f | = 1, and so in particular, f ∈ Lp (A). Then
A
Z
kMφ f kpp = |φf |p
ZA p
1
= +A φ χA0
λ(A0 )1/p 
Z
1 p
= 0
φχA0
λ(A ) A
Z
1
= |φ|p
λ(A0 ) A0
Z
1
≥ (kφk∞ − )p = (kφk∞ − )p
λ(A0 ) A0

35
h i1/p
Hence kMφ f kp ≥ (kf kφ ∞ − )p = kφk∞ − , and so

|||Mφ ||| ≥ kMφ f k ≥ kφk∞ − 


Since  > 0 was arbitrary, we see that |||Mφ ||| ≥ kφk∞ .

Theorem 1.15.6. Let a < b in R, f ∈ L1 [a, b]. Then


R
1. If we define Γf : L∞ [a, b] → R by Γf (φ) = A f φ then Γf is a bounded linear functional with
kΓf k∞ = kf k1
R
2. Define Γf : C[a, b] → R by Γf (h) = A f h. Then Γf is a bounded linear functional with kΓf k∗ = kf k1

Proof. 1. If φ ∈ L∞ [a, b], we have


Z Z
φf ≤ |φf | ≤ kφk∞ kf k1
[a,b] [a,b]

as seen last class. It is immediate that Γf is linear and kΓf k∗ ≤ kf k1 . Thus we need only check
the reverse inequality. Let φ = sgn ◦ f ∈ L∞ (A). Then kφk∞ = 1 (since sgn returns only ±1 in its
range), and Z Z
kΓf k∗ ≥ |Γf (φ)| = f sgn ◦ f = |f | = kf k1
A A

2. We compute, exactly as above, that if h ∈ C[a, b] ⊂ L∞ [a, b] that


Z Z
|Γf (h)| = fh ≤ |f h|
A A
≤ kf k1 khk∞
so we can conclude that kΓf k∗ ≤ kf k1 . From assignment 4 (Q2, b and c), any measurable simple
function (e.g sgn ◦ f can be approximated pointwise almost everywhere by a sequence of continuous
functions (hn )∞ lim hn = sgn ◦ f almost everywhere, and khn k∞ = ksgn ◦ f k∞ = 1.
n=1 such that n→∞
Then it follows that
|f hn | ≤ |f ||hn | ≤ |f | khk∞ ≤ |f |
so by Lebesgue’s Dominated Convergence Theorem we see that
Z
|Γf (hn )| = f hn
A
Z
n→∞
−−−→ f sgn ◦ f
A
= kf k1
Hence
kΓf k∗ ≥ lim |Γ(hn )| since each khn k∞ ≤ 1
n→∞
= kf k1

36
Theorem 1.15.7. If 1 ≤ p < ∞, a < b in R, then Lp [a, b] is separable.

Proof. Recall that C[a, b] is separable in the uniform norm. We note that Q[x], considered as elements
on C[a, b] is dense in C[a, b]. We have that R[x] considered as functions on [a, b] is dense in C[a, b] by
Stone-Weierstrass Theorem. By careful approximation, Q[x] ¯ k·k∞ ⊃ R[x] (since if q(x) = P an xn ∈ R[x]
n
then approximate each ak by rationals rk and we find that q(x) − ( n rn xn ) = n (an − rn )xn ). Hence
P P
¯ k·k∞ = C[a, b]. Let Q[x] = (qn )∞ since Q[x] is countable. Now let f ∈ Lp [a, b],  > 0. By assignment
Q[x] n=1

4, Q2, there is h ∈ C[a, b] such that kf − hkp < . On the other hand, since Q[x] = (qn )∞ n=1 is uniformly
2

dense in C[a, b], there is n0 ∈ N such that kf − qn0 k∞ < . We have that
2(b − 1)1/p

kf − qn0 kp ≤ kf − hkp + kh − qn0 kp Minkowski


Z !1/p

< + |h − qn0 |p
2 [a,b]
!1/p
p
Z

≤ + p
2 [a,b] 2 (b − a)
 
= + =
2 2

Note that L∞ [a, b] is not separable. Let [a, b] = [0, 1] and define ψn = χ[ 1 , 1 ] . If B ⊂ N define
X n+1 n

φB = ψn in a pointwise fashion. Check that {φB }B∈P(N) satisfy


n∈B

kφB − φB 0 k∞ = 1, if B 6= B 0

and that {φB }B∈P(N) = |P(N)| = c and use this to show there is not countable dense subset.

2 Fourier Analysis
Let us begin by motivating our study of this subjection. Consider the heat equation on a disc: Let D be a
homogeneous disc, and
 for mathematical
 convenience we will describe it as the unit circle in the complex
plane, namely D = z ∈ C |z| ≤ 1 .

[insert picture here]


Furthermore, recall polar coordinate representation, which is given by x + iy = reiθ . The coordinate
2 2 2 y
transformation is given by r = x + y , and θ = arctan x . We will consider Dirichlet boundary
conditions, wherein the temperature of the disc is specified at the boundary r = 1. Characterize this
boundary temperature as T (eiθ ) = f (θ), 0 ≤ θ ≤ 2π. We desire a function u defined on D such that
describes the temperature throughout the interior points of the disc. For simplicity sake, we will make the
assumption that the temperature is affected only by the boundary, and that we’re considering only the

37
steady state solution. Applied mathematics tells us that if u is twice continuously partially differentiable
on the interior of the disc, then u is governed by
∂2 1 ∂u 1 ∂2u
0= 2
+ + 2 2 on the interior
∂r r ∂r r ∂θ
u(1, θ) = f (θ) on the boundary
Candidate solutions include
u0 (r, θ) = a0
un (r, θ) = rn (an cos nθ + bn sin nθ)
einθ + e−inθ einθ − e−inθ
 
n
= r an + bn
2 2i
 
= rn c+ ne
inθ
+ c−
ne
−inθ

The general solution to this boundary value problem is given by



X
u(r, θ) = a0 + rn (an cos nθ + bn sin nθ)
n=1

X
= cn r|n| einθ
n=−∞

X
and the boundary condition then implies that u(1, θ) = einθ f (θ). Thus our general question, which
n=−∞
will motivate this section, is if we can devise a way to test the convergence of this last expression. That is,
we want to find a way to write f (t) as a series.

2.1 Complex Valued Functions


Definition 2.1.1. If A ∈ L(R), a function f : A → C is measurable is <(f ), =(f ) : A → R are both
measurable. Furthermore, we say such f is integrable if each of <(f ), =(f ) are integrable and
Z Z Z
f= <(f ) + i =(f )
A A A

 
Note that if we define MC (A) = f : A → C f measurable , then MC (A) is an algebra of functions
if operations are taken pointwise. Furthermore, the Lebesgue Dominated Convergence Theorem holds in
this case. That is, if fn → f pointwise almost everywhere, and |fn | ≤ g for non-negative real valued
g, then LDCT follows. Also, the Holder and Minkowski inequalities still hold. However, the Monotone
Convergence Theorem and Fatou’s Lemma pertain to non-negative real valued functions, and can only be
applied to such.
From now on, if [a, b] is a compact interval in R we let
 
C[a, b] = f : [a, b] → C f continuous

38
 
Lp [a, b] = f : [a, b] → C f ∈ MC [a, b], |f |p ∈ L1 [a, b]
 
L∞ [a, b] = f : [a, b] → C f ∈ MC [a, b], |f | essentially bounded

Definition 2.1.2. Let θ ∈ R\{0} . We say that f : R → C is almost everywhere θ-periodic if


f (t + θ) = f (t) for almost every t ∈ R.

Notice that if we define en (t) = eint then en , n ∈ N is 2π-periodic. If we define the boundary ∂D = T ,
then for t ∈ R then the map t 7→ eit takes R → T . Thus functions on T will naturally correspond to
2π-periodic functions on R.
Definition 2.1.3. Define the following:
 
C(T) = f : R → C f continuous, 2π-periodic
 
≡ f : [−π, π] → C f continuous, f (−π) = f (π) ⊂ C[−π, π]

( )
Lp (T) = f : R → C f ∈ MC (R), f almost everywhere 2π periodic, |f |p ∈ L1 [−π, π]
[−π,π]
≡ Lp [−π, π]
 
L∞ (T) = f ∈ L∞ (R) f 2π-periodic

Define the norm for f ∈ C[T], Lp (T) for 1 ≤ p < ∞ as

kf k∞ = esssupt∈[−π,π] |f (t)|
 Z π 1/p Z π 1/p
1 p 1 p
= |f | = |f |
2π −π (2π)1/p −π

Clearly Minkowski’s inequality still holds. Furthermore, Holder’s inequality also holds since if f ∈
Lp (T), g ∈ Lq (T) for p, q conjugate indices. Then
Z π
1
kf gk1 = |f g|
2π −π
Z π 1/p Z π 1/q
1 p q
≤ |f | |g|
2π −pi −π
 Z π 1/p  Z π 1/q
1 p 1 q
= |f | |g|
2π −π 2π −π
= kf kp kgkq

In summary we have that


 Z π 
L(T) = f : R → C f ∈ MC (R), f is 2π periodic and |f | < ∞
−π

39
 
L1 (T) = f : R → C f inMC (R), f is a.e 2π periodic and f |[−π,π] ∈ L1 [−π, π]

and we recall that L1 (T) ⊃ Lp (T) ⊃ C(T). We are then faced with the basic question: ”If f ∈ L(T), how
can we represent f as a Fourier series. That is, find coefficients such that

X
f (t) = cn eint
n=−∞

If we allow the naive interchanging of the sum and the integral, without any consideration of convergence,
X∞
suppose we can write f in the following form, f (t) = cn en where we recall that en (t) = eint , and
n=−∞
f ∈ L(T).

!
Z π Z π X
f (t)e−ikt dt = cn eint e−ikt dt
−π −π n=−∞

X Z π
= cn ei(n−k)t dt
n=−∞ −π

Let us focus now on the integrand itself, and break this into its real and imaginary parts
Z π Z π Z π
i(n−k)t
e dt = cos((n − k)t)dt + i sin((n − k)t)dt
−π
(−π −π

2π if n = k
=
0 otherwise

Thus we get that


Z π ∞
X Z π
f (t)e−ikt dt = cn ei(n−k)t dt = 2πck
−π n=−∞ −π

Definition 2.1.4. The k th Fourier coefficient k ∈ Z of f ∈ L(T) is given by


Z π Z π
1 −ikt 1
ck = f (t)e dt = f e−k
2π −π 2π −π

Note that if f = g almost everywhere, then f e−k = ge−k almost everywhere, which implies that
ck (f ) = ck (g). This allows us to conclude that ck (f ) is well-defined for f ∈ L1 (T). This allows us to refine
our above question:
Let f ∈ L1 (T), Lp (T), C(T). Then does the following hold:

X N
X
f= cn (f )en = lim cn (f )en
N →∞
n=−∞ n=−N

in the L1 , Lp norm?
Another important motivational computation is as follows:

40
n
X
If f ∈ L1 (T), we let Sn (f ) = ck (f )ek . Furthemore, ∀t ∈ [−π, π], we may denote Sn (f, t) =
k=−n
n
X
ck (f )eikt , and note that the summand is always a continuous 2π periodic function. Thus for t ∈ [−π, π]
k=−n
we have
n
X
Sn (f, t) = ck (f )eikt
k=−n
n  Z π 
X 1 iks
= f (s)e ds eikt
2π −π
k=−n
Z π n
1 X
= f (s) eik(t−s) ds
2π −π k=−n
Z π n
1 X
= f (s)Dn (t − s)ds where Dn = ek
2π −π k=−n
Z π−t
1
= Dn (−σ)f (σ + t)dσ σ = s − t
2π −π−t
Z π
1
= Dn (−σ)f (σ + t)dσ by assg 5
2π −π
Z π
1
= Dn (s)f (−s)ds, where s = −σ
2π −π
= Dn ∗ f (t) called the convolution of Dn with f

Where we refer to Dn as the Dirichlet kernel of order n.

2.2 Convolution
Let us examine the notion of convolution in a more rigorous, theoretical sense. Let X be a Banach space.
We say that X is homogeneous of T if there is a group action R × X → X where (t, x) 7→ t ∗ x. That is, if
s, t ∈ R, x ∈ X then a continuous group action by T is defined by the following characteristics:

• 0∗x=x

• (s + t) ∗ x = s ∗ (t ∗ x)

• s 7→ s ∗ x : R → X is continuous

• (s + 2π) ∗ x = s ∗ x

Moreover, we have for s ∈ R, x, y ∈ X, α ∈ R then

• s ∗ (x + y) = s ∗ x + s ∗ y

• s ∗ (αx) = α(s ∗ x)

41
• ks ∗ xk = kxk

that is, this group action is linear, and isometric.


Example: Let X = C(T); that is, the continuous 2π periodic functions. The for s ∈ R, f ∈ C(T), define
the group action by
s ∗ f (t) = f (t − s) for t ∈ R
We next move to confirm that this is indeed a continuous group action, causing C(T) to be a homogeneous
Banach space over T.

Proof. Let s, s1 ∈ T, f ∈ C(T), then we have

• 0 ∗ f (t) = f (t − 0) = f (t) ⇒ 0 ∗ f = f

(s + s1 ) ∗ f (t) = f (t − (s + s1 ))
= f ((t − s1 ) − s)
= s ∗ (s1 ∗ f (t))

• If s 7→ s0 in R, we want to show that lim ks ∗ f − s0 ∗ f k∞ = 0. Recall that


s→s0
 
C(T) ≈ f : [−π, π] → C f continuous, f (−π) = f (π) ⊂ C[−π, π]

Thus if f ∈ C[−π, π] then since [−π, π] is compact, we have that f is uniformly continuous. Hence
given  > 0 there is a δ > 0 such that |t − t0 | < δ ⇒ |f (t) − f (t0 )| < . Thus for any t ∈ [−π, π] we
have that
s ∗ f (t) − s0 ∗ f (t) = f (t − s) − f (t − s0 )
But |(t − s) − (t − s0 )| = |s0 − s|, so if |s − s0 | < δ we find that
ks ∗ f − s0 ∗ f k∞ = max |f (t − s) − f (t − s0 )| < 
t∈R

Alternatively, say s 7→ s0 in R. Given  > 0 find δ > 0 such that |s − s0 | < δ ⇒ ks ∗ f − s0 ∗ f kp < .

Then for assignment 4, find h ∈ C(T) such that kf − hkp < . Then from the fact that C(T) is a
3
homogeneous Banach space over T, ∃δ > 0 such that

|s − s0 | < δ ⇒ ks ∗ h − s0 ∗ hk∞ <
3
Thus we have for |s − s0 | < δ then
ks ∗ f − s0 ∗ f kp ≤ ks ∗ f − s ∗ hkp + ks ∗ h − s0 ∗ hkp + ks0 ∗ h − s0 ∗ f kp
 Z π 1/p
1 p
= ks ∗ (f − h)kp + |s ∗ h − s0 ∗ h| + ks0 ∗ (h − f )kp
2π −π
≤ kf − hkp + ks ∗ h − s0 ∗ hk∞ + kh − f kp
  
< + + =
3 3 3

42
• (2π) ∗ f (t) = f (t − 2π) = f (t), by 2π periodicity

Now for linearity and isometry, we have for f, g ∈ C(T), α ∈ C we can easily check that

• s ∗ (f + g) = s ∗ f + s ∗ g, s ∗ (αf ) = α(s ∗ f )

• Let us check that ks ∗ f k∞ = kf k∞ . Note that


Z π Z π Z π−s
1 1 1 1
ks ∗ f kpp = |s ∗ f | =p p
|f (t − s)| = |f (t0 )|dt0
2π 2π −π 2π −π 2π −π−s
Z π
1
= |f (t)|p dt = kf kpp
2π −π

Trivial Example: Let X be a Banach space, for s ∈ R define s ∗ x = x for x ∈ X. This defines a
homogeneous Banach space over T.
Definition 2.2.1 (Convolution). Let h ∈ C(T) ( where we could relax continuity to piecewise continuous,
bounded, 2π-periodic function ). Let X be a homogeneous Banach space over T. Define

Fh,x : R → X by Fh,x (s) = h(s)s ∗ x, x∈X

Then Fh,x is a continuous, 2π periodic function from R → X. Now define the convolution of h with x by
the following: Z π Z π
1 1
h∗x= Fh,x (s)ds = h(s)s ∗ xds
2π −π 2π −π
Note that this is a vector valued Riemann integral.
Now suppose that f ∈ C(T) [or f ∈ Lp (T)]. Then for almost every choice of t ∈ R we can write
Z π Z π
1 1
h ∗ f (t) = h(s)s ∗ f (t)ds = h(s)f (t − s)ds
2π −π 2π −π

Note: Given h ∈ C(T) we can define an operator Φh : X → X by Φh (x) = h ∗ x. Using properties of


the vector-valued Riemann integral, we have that Φh is a linear operator. Moreover, Φh is bounded, since
if x ∈ X, then
Z π Z π
1 1
kΦh (x)k = h(s)s ∗ xds ≤ kh(s)s ∗ xkds
2π −π 2π −π
Z π Z π
1 1
= |h(s)|ks ∗ xkds = |h(s)|kxkds
2π −π 2π −π
Z π
1
= |h(s)|ds · kxk = khk1 kxk
2π −π

Thus |||Φh ||| ≤ khk1 . We will see that if X = L1 (T), C(T) then |||Φh ||| = khk1 . However, if X = L2 (T)
then we can find examples where |||Φh ||| < khk1 .

43
2.3 Approximating with Fourier Series
n
X Z π
Let f ∈ C(T) and define Sn (f, t) = ck (f )eikt , where ck (f ) = f (s)e−iks ds. If we let Dn (t) =
k=−n −π
Pn 1

k=−n eiktthen we have that −π Dn (s)f (t − s)ds. This suggests that properties of Sn (f, t)
Sn (f, t) = 2π
should be in a bijective correspondence with Dn , the Dirichlet kernel.
Theorem 2.3.1 (Properites of the Dirichlet Kernel). The Dirichlet kernel of order n, Dn satisfies the
following relations

1. Dn is real-valued, 2π-periodic and even


Z π
1
2. Dn = 1
2π −π
3. For t ∈ [−π, π], we have that Dn (t) admits the following description
 1
 sin[(n+ 2 )t] , t 6= 0
1
Dn (t) = sin[ 2 t]
2 + 1, t = 0
n

Z π
1
4. lim kDn k1 = lim |Dn (t)|dt = ∞
n→∞ n→∞ 2π −π
n
X
Theorem 2.3.2 (Properites of the Dirichlet Kernel). The Dirichlet kernel of order n, Dn = eikt
k=−n
satisfies the following relations

1. Dn is real-valued, 2π-periodic and even


Z π
1
2. Dn = 1
2π −π
3. For t ∈ [−π, π], we have that Dn (t) admits the following description
 1
 sin[(n+ 2 )t] , t 6= 0
1
Dn (t) = sin[ 2 t]
2 + 1, t = 0
n

Z π
1
4. lim kDn k1 = lim |Dn (t)|dt = ∞
n→∞ n→∞ 2π −π

Proof. 1. 2π periodicity is obvious, since the Dirichlet kernel is the sum of functions that are 2π periodic,
evenness and real-valuedness follow from item 3.

2. Note the following


π π n n Z
1 X π ikt
Z Z
1 1 X
ikt
Dn = e dt = e dt = 1
2π −π 2π −π k=−n 2π −π
k=−n

44
X 1 1
3. Dn (t) = kneikt = e−int + · · · + e−it + 1 + eit + · · · + eint . Now we multiply this sum by e−i 2 t + ei 2 t
t get
 1 1
 h 1 1
i h 1 1
i
Dn (t) e−i 2 t − ei 2 t = e−i(n+ 2 )t + · · · + ei(n− 2 )t − e−i(n− 2 )t + · · · + ei(n+ 2 )t
1 1
= e−i(n+ 2 )t − ei(n+ 2 t)
     
1 1
Dn (t) −2i sin t = −2i sin n + t
2 2

Thus for t ∈ [−π, π]\{0} we have that

−2i sin n + 12 t sin n + 12 t


     
Dn (t) = =
−2i sin 21 t sin 12 t
   

X
Also, we have that Dn (0) = kneik0 = 2n + 1.

4. Recall that | sin θ| ≤


Z |θ| for θ ∈ R: this will be very useful in the following calculation. We have that
π
1
Ln = kDn k1 = |Dn |. By part 3, this is equivalent to
2π −π

1 π
Z
Ln = |Dn |
π 0
1 π sin n + 12 t
Z 
= dt
π 0 sin 12 t
2 π | sin n + 12 t|
Z 
≥ dt
π 0 t
1
2 (n+ 2 )π | sin s|
Z  
1 1
= s 1 ds s = n + 2 t

π 0 1 n +
( 2)
n+ 2
Z nπ
| sin s|
≥ ds
0 s
n Z
2 X jπ | sin s|
= ds
π s
j=1 (j−1)π
n Z jπ
2X 1
≥ | sin s|ds since s ≤ jπ on [(j − 1)π, jπ]
π jπ (j−1)π
j=1
n
2 X 1
=
π2 j
j=1
n→∞
−−−→ ∞
n
2 X 1 n→∞
Hence Ln = kDn k1 ≥ −−−→ ∞.
π2 j
j=1

45
n→∞
We will use the fact that lim Ln = ∞ to show that if f ∈ C(T) then Sn (f, t) −
6 −−→ f (t) in the uniform
n→∞
sense.

Theorem 2.3.3. Let h ∈ C(T). Then

1. If Φh : L1 (T) → L1 (T) is defined by Φh (f ) = h ∗ f , then |||Φh ||| = khk1 .

2. If Ψh : C(T) → C(T) is defined by Ψh (f ) = h ∗ f , then |||Ψh ||| = khk1 .

Proof. Recall that for any convolution operator Φh : X → X for X a homogeneous Banach space over T,
defined as Φh (x) = h ∗ x, then |||Φh ||| ≤ khk1 . Thus it is sufficient to show the reverse inequality.

 
1. Recall that |||Φh ||| = sup kΦh (f )k1 f ∈ L1 (T), kf k1 ≤ 1 . Define fn = nπχ[− 1 , 1 ] on [−π, π] and
n n

extend to a 2π periodic function. Notice that kfn k≤ 1 since

Z π
1
kfn k1 = nφχ[− 1 , 1 ]
2π −π n n
Z π
n n2
= χ[− 1 , 1 ] = =1
2 −π n n 2n

Now we have that

Z π
1
h ∗ fn (t) = h(s)fn (t − s)ds
2π −π
Z π
n
= h(s)χ[− 1 , 1 ] (t − s)ds
2 −π n n
Z t−π
n
=− h(t − σ)χ[− 1 , 1 ] (σ)dσ σ =t−s
2 t+π n n
Z π−t
n
= h(s + t)χ[− 1 , 1 ] (s)ds
2 −π−t n n

n π
Z
= h(s + t)χ[− 1 , 1 ] (s)ds
2 −π n n
Z 1
n n
= h(s + t)ds
2 −1
n

46
Now we have that
Z π
kh − h ∗ fn k1 = |h − h ∗ fn |
−π
Z π Z 1
1 n n
= h(t) = h(s + t)ds dt
2π −π 2 −1
n
Z π Z 1 Z 1
1 n n n n
= [h(t) − h(t + s)] ds dt h(t) = h(t)ds
2π −π 2 − 1 2 −1
n n
Z πZ 1
n n
≤ |h(t) − h(t + s)| dsdt
4π −π − 1
n
Z πZ 1
n n
= |h(t) − (−s) ∗ h(t)|dsdt
4π −π − 1
n
Z πZ 1
n n
≤ sup |h(t) − (−s) ∗ h(t)|dsdt
4π −π − 1 s∈[− 1 , 1 ]
n n n
Z πZ 1
n n
≤ sup kh − (−s) ∗ hk∞ dsdt
4π −π − 1 s∈[− 1 , 1 ]
n n n
n→∞ n→∞
−−−→ 0 since sup kh − (−s) ∗ hk∞ −−−→
[− n1 , n1 ]

Theorem 2.3.4. Let h ∈ C(T). Then

1. If Φh : L1 (T) → L1 (T) is defined by Φh (f ) = h ∗ f , then |||Φh ||| = khk1 .

2. If Ψh : C(T) → C(T) is defined by Ψh (f ) = h ∗ f , then |||Ψh ||| = khk1 .

Proof. Recall that for any convolution operator Φh : X → X for X a homogeneous Banach space over T,
defined as Φh (x) = h ∗ x, then |||Φh ||| ≤ khk1 . Thus it is sufficient to show the reverse inequality.

 
1. Recall that |||Φh ||| = sup kΦh (f )k1 f ∈ L1 (T), kf k1 ≤ 1 . Define fn = nπχ[− 1 , 1 ] on [−π, π] and
n n

extend to a 2π periodic function. Notice that kfn k≤ 1 since


Z π
1
kfn k1 = nφχ[− 1 , 1 ]
2π −π n n
Z π
n
= χ 1 1
2 −π [− n , n ]
n2
= =1
2n

47
Now we have that

Z π
1
h ∗ fn (t) = h(s)fn (t − s)ds
2π −π
Z π
n
= h(s)χ[− 1 , 1 ] (t − s)ds
2 −π n n
Z t−π
n
=− h(t − σ)χ[− 1 , 1 ] (σ)dσ σ =t−s
2 t+π n n
Z π−t
n
= h(s + t)χ[− 1 , 1 ] (s)ds
2 −π−t n n
Z π
n
= h(s + t)χ[− 1 , 1 ] (s)ds
2 −π n n
Z 1
n n
= h(s + t)ds
2 −1
n

Now we have that

Z π
kh − h ∗ fn k1 = |h − h ∗ fn |
−π
Z π Z 1
1 n n
= h(t) = h(s + t)ds dt
2π −π 2 −1
n
Z π Z 1 Z 1
1 n n n n
= [h(t) − h(t + s)] ds dt h(t) = h(t)ds
2π −π 2 − 1 2 −1
n n
Z πZ 1
n n
≤ |h(t) − h(t + s)| dsdt
4π −π − 1
n
Z πZ 1
n n
= |h(t) − (−s) ∗ h(t)|dsdt
4π −π − 1
n
Z πZ 1
n n
≤ sup |h(t) − (−s) ∗ h(t)|dsdt
4π −π − 1 s∈[− 1 , 1 ]
n n n
Z πZ 1
n n
≤ sup kh − (−s) ∗ hk∞ dsdt
4π −π − 1 s∈[− 1 , 1 ]
n n n
n→∞ n→∞
−−−→ 0 since sup kh − (−s) ∗ hk∞ −−−→
[− n1 , n1 ]

Thus we have

n→∞
|khk1 − kh ∗ fn k1 | ≤ kh − h ∗ fn k1 −−−→ 0 ⇒ lim kh ∗ fn k1 = khk1
n→∞

48
this allows us to conclude that
 
|||Φh ||| = sup normΦh (f )1 f ∈ L1 (T), kf k1 ≤ 1
 
= sup kh ∗ fn k1 f ∈ L1 (T), kf k1 ≤ 1

≥ sup kh ∗ fn k1 ≥ lim kh ∗ fn k1 = khk1


n∈N n→∞

2. Define Γ0 : C(T) → C by Γ0 (f ) = f (0). Then we have that Γ0 is linear and


 
kΓ0 k∗ = sup |Γ0 (f )| f ∈ C(T), kf k∞ ≤ 1
 
= sup |f (0)| f ∈ C(T), kf k∞ ≤ 1

≤ sup |f (t)| = kf k∞
t∈[−π,π]

Now we have Ψh (f ) = h ∗ f ∈ C(T) and


Z π
1
h ∗ f (0) = h(s)s ∗ f (0)ds
2π −π
Z π Z π
1 1
= h(s)f (0 − s)ds = h(s)f˜(s)ds where f˜(s) = f (−s)
2π −π 2π −π

Note that f˜ = sup |f (−t)| = sup |f (t)| = kf k∞ and so the operation of taking f 7→ f˜ is
∞ t∈[−π,π] t∈[−π,π]
isometric. Thus we have for f ∈ C(T) that kf k∞ ≤ 1, yielding
kΨh (f )k∞ = kh ∗ f k∞
Z π
1
≥ |h ∗ f (0)| = h(s)f˜(s)ds
2π −π

1

Recall that for f ∈ L1 (T), Γf : C(T) → C given by Γf (g) = 2π −π f g for all g ∈ C(T), then this
satisfies kΓf k∗ = kf k1 . From above, we have that
Z π
1
kΨh (f )k∞ ≥ h(s)f˜(s)ds = Γh (f˜)
2π −π
and thus
 
|||Ψh ||| = sup kΨh (f )k∞ f ∈ C(T), kf k∞ ≤ 1
 
˜
≥ sup Γh (f ) f ∈ C(T), kf k∞ ≤ 1
 
= sup |Γh (f )| f ∈ C(T), kf k∞ ≤ 1 since f 7→ f˜ isometric bijection

= khk1
and thus we can conclude that |||Ψh ||| ≥ khk1 and we are done.

49
2.3.5. Let X be a Banach space. A set F ⊂ X is called a set of first category or a meager
Definition S
set if F = ∞ n=1 Fn , where each set Fn is nowhere dense. We recall that a set Fn is nowhere dense if
0
F̄n = ∅, that is, the interior of the closure is empty. We say that U ⊂ X is of second category or
non-meager if it is not of type first category.
Theorem 2.3.6 (Baire Category Theorem). If X is a Banach space, then any open subset U ⊂ X is non
meager.

That is, Banach spaces contain many non-meager sets. Note that F is nowhere dense ⇔, G = X\F̄ is
S∞
dense. Thus we can conclude that if F1 , F2 , . . . is a sequence of meager sets, then n=1 Fn is also meager.
Theorem 2.3.7 (Banach-Steinhaus Theorem).If X, Y are Banach spaces and (Ti )∞ i=1 is a family of

bounded linear operator Ti : X → Y , then if sup kTi xk i ∈ I < ∞ for any x ∈ U ⊂ X, for non-meager
U , then      
sup |||Ti ||| i ∈ I < ∞ sup kTi k∗ i ∈ I , if Y = C

Corollary 2.3.8. If {T } n is a sequence of bounded operators (functionals if Y = C) such that sup |||Tn ||| =
  n∈N

∞ then x ∈ X sup ||Tn x|| < ∞ is meager in X.


n∈N

Proof. For each n ∈ N define the set


 
Fn = x ∈ X sup kTi xk ≤ n
i∈I
[ 
= x ∈ X kTi xk ≤ n
i∈I

which is closed since this is the intersection of closed sets. By assumption, there is a non-meager set U
[∞
such that x ∈ U ⇒ sup kTi xk < ∞ and thus U ⊆ Fn . Since U is non-meager, then we can write
i∈I n=1

(Fn ∩ U ) we have for at least n0 ∈ N that (Fn0 ¯∩ U )o 6= ∅ and in particular, Fno0 ⊃ (Fn0 ¯∩ U )o .
[
U =
n=1
Thus there is x0 ∈ Fn0 and r > 0 such that
 
Br (x0 ) = x ∈ X kx − x0 k < r ⊂ Fn0

1h r r i
If x ∈ X for kxk ≤ 1 then x0 ± 2r x ∈ Br (x0 ) ⊂ Fn0 , then x = (x0 + ) − (x0 − x) , which yields
r 2 2
 h 
1 r r i
kTi xk = Ti (x0 + x) − (x0 − x)
r 2 2
1 r 1 r
≤ Ti (x0 + x) + Ti (x0 − x)
r 2 r 2
n0
≤2
r

50
 
2n0 n0
Thus |||Ti ||| = sup kTi xk x ∈ X, kxk ≤ 1 ≤ r and hence sup |||Ti ||| ≤ 2 <∞
i∈I r

Theorem 2.3.9. 1. There is a set U ⊂ L1 (T) whose complement is meager, such that sup kSn (f )k1 =
n∈N
∞ for f ∈ U

2. There is a subset U ⊂ C(T) whose complement is meager such that sup kSn (h)k∞ = ∞ for h ∈ U
n∈N

Proof. 1. Recall that


n Z π
X 1
Sn (f ) = j
cj (f )e = Dn ∗ f where e (t) = ej ijt
and cn (f ) = f e−j
2π −π
j=−n
Z π
1 n→∞
And we have that kDn k1 = |Dn | = Ln −−−→ ∞. Also,
2π −π
   
sup kSn (f )k1 f ∈ L1 (T), kf k1 ≤ 1 = sup kDn ∗ f k1 f ∈ L1 (T), kf k1 ≤ 1

= |||ΦDn ||| = kDn k1 = Ln = ∞

By (Corollary to) the Banach-Steinhaus Theorem,


 
F = f ∈ L1 (T) sup kΦDn (f )k1 < ∞
n∈N

then F is meager of L1 (T). The Baire Category theorem implies that L1 (T) is non-meager, and so
U ⊂ L1 (T)\F 6= ∅. That is, in L1 (T), our Fourier series don’t necessarily converge

2. Proved similarily

In light of this fact, there are two ways that we can proceed. We can average the Fourier sums (an idea
due to Fejer), and we can look at specific functions where convergence holds (Dini’s Theorem).

2.4 Averaging Fourier Series


Definition 2.4.1. Let X be a Banach space, and x = (xn )∞
n=1 ⊂ X. Define the n
th Cesaro aver-
∞ 1
age/mean of (xn )n=1 as σn (x) = n (x1 + · · · + xn )
Proposition 2.4.2. If lim xn exists, then lim σn (x) exists and
n→∞ n→∞

lim xn = lim σn (x)


n→∞ n→∞

The proof of this proposition is left as an exercise.

51
Definition 2.4.3. If f ∈ L(T) [ or f ∈ L1 (T) ], we define

n n j
1 X 1 X X
σn (f ) = Sj (f ) = ck (f )ek
n+1 n+1
j=0 j=0 k=−j

Note that we can think of the Cesaro average as


 
n n
1 X 1 X
σn (f ) = Dj ∗ f =  Dj  ∗ f
n+1 n+1
j=0 j=0

n n j
1 X 1 X X k
Let Kn = Dj = e , and call this the Fejer kernel of order n. The philosophy
n+1 n+1
j=0 j=0 k=−j
and motivation of understanding the behaviour of the Fejer kernel is that it will give us some insight into
the convergence of σn (f ).

Theorem 2.4.4 (Properties of the Fejer Kernel). The Fejer kernel of order n, Kn satisfies the following

1. Kn is real-valued, continuous, and 2π-periodic


  2
 1 sin[ 12 (n+1)t]

n+1 sin[ 12 t]
t 6= 0
2. For t ∈ [−π, π], we have Kn (t) =

n + 1 t=0
Z π Z π
1 1
3. kKn k1 = |Kn | = Kn = 1
2π −π 2π −π

4. If 0 < |t| ≤ π, then


π2
Kn (t) ≤
(n + 1)t2

Proof. 1. Since each Dj , j = 0, 1, . . . , n is real valued, continuous, and 2π-periodic, then the same is
n
1 X
true for Kn = Dj
n+1
j=0

2. Let us calculate Kn (t) as follows

n j
" n #
1 X X ikt 1 X
ikt
Kn (t) = e = (n + 1 − |k|)e
n+1 n+1
j=0 k=−j k=−n
1 h −int i
= e + 2e−i(n−1)t + · · · + ne−it + (n + 1) + neit + · · · + eint
n+1

52
Now consider
(n + 1)Kn (t) e−it − 2 + eit

h i
= e−i(n+1)t + 2e−int + 3e−i(n−1)t + · · · + (n + 1)e−it + n + (n − 1)eit + · · · + ei(n−1)t
h i
− 2e−int + 2 · · · 2e−i(n−1)t + · · · + 2ne−it + 2(n + 1) + 2neit + · · · + 2 · 2ei(n−1)t + 2eint
h i
+ e−(n−1)t + · · · + (n − 1)e−it + n + (n + 1)eit + · · · + (n − 2)ei(n−1)t + 3ei(n−1)t + 2eint + ei(n+1)t
= e−i(n+1)t − 2 + ei(n+1)t
If t ∈ [−π, π]\{0} then
h
(n+1) (n+1)
i2
−i 2 t t
1 e−i(n+1)t
−2+ e
ei(n+1)t − e 2

Kn (t) = −it
=
n+1 e − 2 + eit t 2
h t i
e−i 2 − ei 2
2  !2
−i sin 21 (n + 1)t sin 12 (n + 1)t
 
1 1
=  2 =
sin 12 t
 
n+1 −i sin 12 t n+1

Also, we have that


n j n
1 X X ik0 1 X
Kn (0) = e = (2j + 1)
n+1 n+1
j=0 k=−j j=0
 
1 n(n + 1)
= 2 +n+1 =n+1
n+1 2

3. Since Kn ≥ 0 on [−π, π] we have that


Z π Z π
1 1
kKn k1 = |Kn | = Kn
2π −π 2π −π
n Z
1 1 X π
Z π
= Dj but Dj = 2π
2π n + 1 −π −π
j=0

=1
2
4. We have that πθ ≤ sin θ for 0 ≤ θ ≤ π2 . Thus for 0 < t < π we have
1 1 π
1  ≤ t =
sin 2 t π
t
Thus we have that
1  !2
1 sin 2 (n + 1)t
Kn (t) = 1 
n+1 sin 2 t
1 1  π 2
≤  1 2 ≤
(n + 1) sin 2 t n+1 t
π2
=
(n + 1)t2

53
for t > 0. This also holds for t < 0 since Kn is even Kn (t) = Kn (−t)

Definition 2.4.5. A sequence (kn )∞


n=1 of (piece-wise) continuous, real-valued, 2π-periodic functions on R,
is a summability kernel provided that
Z π
1
1. (Limit over average 1) kn = 1
2π −π
Z π
1
2. (L1 -bounded) sup kkn k1 = sup |kn | < ∞
n∈N n∈N 2π −π

Z −δ Z π 
3. (mass concentrated about 0) If 0 < δ ≤ π, then lim |kn | + |kn | =0
n→∞ −π δ

Example: The Fejer kernel {Kn }∞


n=1 is a summability kernel. Notice that (1),(2) are satisfied since
Z π
1 n→∞
Kn = kKn k1 = 1 −−−→ 1 < ∞
2π −π

π2
To see part (3), note that for 0 < δ ≤ π we have that Kn (t) ≤ . By even-symmetry of Kn we get
(n + 1)t2
Z −δ Z π Z π
Kn + =2 Kn (t) dt
−π δ δ
Z π
π2 π2
 
dt 1 1
≤2 = −
n + 1 δ t2 n+1 δ π
n→∞
−−−→ 0

Example: The Dirichlet kernel {Dn }∞


n=1 is not a summability kernel. We notice that condition (2)
n→∞
fails since Ln = kDn k1 −−−→ ∞.
Example: Consider kn = nπχ[− 1 , 1 ] on [−π, π], continued 2π-periodically on R. This is a summability
n n
kernel, and is left as an exercise to prove.
2
Example: Define gn (t) = cn e−nt , t ∈ [−π, π] continued 2π-periodically, is a summability kernel for
appropriate normalization constant cn . Parts (1),(2) come for free, and part (3) comes from calculus
estimates.
(
cn sin(nx)
x x ∈ [−π, π]\{0}
Example: Define fn = , continued 2π-periodically is a summability kernel.
cn n x=0

Theorem 2.4.6 (Abstract Summability Kernel Theorem). Let X be a homogeneous Banach space over
T. If (kn )∞
n=1 is a summability kernel, then

lim kkn ∗ x − xkX = 0


n→∞

54
Proof. Fix x ∈ X\{0} , and define F : R → X by F (s) = s∗x, so we see immediately that F (s) is continuous
Z π
1
and 2π-periodic, and F (0) = 0 ∗ x = x. Then by definition, kn ∗ x = kn (s)s ∗ x. Then we have
2π −π

Z π
1
kkn ∗ x − xkX = kn (s)F (s) ds − F (0)
2π −π
Z π Z πX Z π
1 1 1
= kn (s)F (s) ds − kn (s)F (0) ds since kn (s) ds = 1
2π −π 2π −π X 2π −π
Z π  
1
= kn (s) F (s) − F (0) ds
2π −π X
Z π
1
≤ |kn (s)| kF (s) − F (0)kX ds
2π −π

Now let 0 < δ ≤ π and re-write this last line in terms of three integrals

Z π
1
|kn (s)| kF (s) − F (0)kX ds
2π −π
Z −δ Z δ Z π
1 1 1
= kn (s)| kF (s) − F (0)kX ds + kn (s)| kF (s) − F (0)kX ds + kn (s)| kF (s) − F (0)kX ds
2π −π 2π −δ 2π δ
= († † −) + (†) + († † +)

Let us consider († † −),

Z −δ
kn (s)| kF (s) − F (0)kX ds ≤ kF (s)kX + kF (0)kX ≤ sup 2 kF (t)kX
−π t∈[−π,π]
" #
Z −δ
≤ |kn (s)| sup 2 kF (t)kX ds
−π t∈[−π,π]
Z −δ
= |kn (s)|2 kxkX since kt ∗ xkX = kxkX
−π
Z −δ
n→∞
= 2 kxkX |kn (s)| ds −−−→ 0 since summability kernel
−π

n→∞
Similarily, († † +) −−−→ 0 as well. Thus all that is left to check is (†).

55
Z δ
1
|kn (s)| kF (s) − F (0)kX ds ≤ sup kF (t) − F (0)kX constant in s
2π −δ t∈[−δ,δ]
" #
Z δ
1
≤ |kn (s)| sup kF (s) − F (0)kX ds
2π −δ t∈[−δ,δ]
" # Z δ
1
= sup kF (s) − F (0)kX |kn (s)| ds
t∈[−δ,δ] 2π −δ

Z δ Z π Z π
1 1 1
but notice that |kn (s)| ds ≤ |kn (s)| ds ≤ sup |kn | = C < ∞
2π −δ 2π −π n∈N 2π −π

δ→0
≤ sup kF (s) − F (0)kX C −−−→ 0 since F is continuous
t∈[−δ,δ]

Thus for any δ > 0 we have


Z π
1
lim sup kkn ∗ x − xkX ≤ lim sup |kn (s)| kF (s) − F (0)kX ds
n→∞ n→∞ 2π −π
≤ sup kF (s) − F (0)kC
t∈[−δ,δ]

δ→0
But since sup kF (s) − F (0)kX −−−→ 0, we find that
t∈[−δ,δ]

0 ≤ lim sup kkn ∗ x − xkX ≤ 0 ⇒ lim kkn ∗ x − xk= 0


n→∞ n→∞

Corollary 2.4.7. 1. If f ∈ C(T) then lim kσn (f ) − f k∞ = 0


n→∞

2. If f ∈ Lp (T) for 1 ≤ p < ∞ then lim kσn (f ) − f kp = 0


n→∞

Proof. 1. We see that for f ∈ C(T) that σn (f ) = Kn ∗ f where (Kn )∞


n=1 is the Fejer kernel. By the
abstract summability kernel theorem, we have that

lim kkn ∗ f − f k∞ = 0
n→∞

2. Proof is nearly identical

Let us focus for the moment on L1 (T). We recall that if f = g almost everywhere ( i.e. f = g in L1 (T))
then ck (f ) = ck (g), ∀k ∈ Z, which leads us to the following corollary

56
Corollary 2.4.8. If f, g ∈ L1 (T) and ck (f ) = ck (g) then f = g almost everywhere; that is, f = g in
L1 (T).

Proof. Note that k·k1 − lim σn (f ) = f ; that is lim kσn (f ) − f k1 = 0. Thus if ck (f ) = ck (g) for every
n→∞ n→∞
k ∈ Z, then σn (f ) = σn (g) for each n ∈ N. Hence it is the case that f = k · k1 − lim σn (f ) =
n→∞
k · k1 − lim σn (g) = g
n→∞

In a more classical setting, define


 Z π 
L(T) = f : R → C f is 2π periodic, measurable, |f | < ∞
−π

and note that L1 (T) = L(T)\∼ where f ∼ g if f = g almost everywhere. If f ∈ L(T), x ∈ R we define the
mean-value of f at x by

1
ωf (x) = lim [f (x + s) + f (x − s)] provided the limit exists
s→0 2

If f is real-valued, then ±∞ are acceptable values for this limit.


Example: Let f (t) = √1 for t ∈ [−π, π] extended 2π-periodically to R, then ωf (0) = ∞.
|t|

Theorem 2.4.9 (Fejer’s Theorem). 1. If f ∈ L(T), x ∈ [−π, π] and ωf (x)exists, then

lim σn (f, x) = lim σn (f )(x) = ωf (x)


n→∞ n→∞

2. If [a, b] ⊂ [−π, π] is a closed subset on which f is continuous, then

lim σn (f, x) = f (x) uniformly for x ∈ [a, b]


n→∞

n→∞
where we note that by the limit converging uniformly, we mean that sup |σn (f, x) − f (x)| −−−→ 0
x∈[a,b]

Proof. 1. Recall some properties of the Fejer kernel Kn .

(a) Kn ∗ f (x) = σn (f, x)


(b) Kn is even, non-negative, 2π-periodic
Z π Z π
1 1
(c) Kn = |Kn | = 1
2π −π 2π −π
π2 π2
(d) If 0 < |t| ≤ π, then Kn (t) ≤ (n+1)t2
. In particular, if 0 < δ ≤ π then Kn (t) ≤ (n+1)δ 2
for
t ∈ [−π, −δ] ∪ [δ, π]

1
Now suppose x ∈ [−π, π] and ωf (x) = lim [f (x + s) + f (x − s)] exists and |ωf (x)| < ∞ (the case
2s→0
of ±∞ still holds, and is left as an exercise). Let  > 0 be given, and choose 0 < δ ≤ π such that for

57
1
0 < |s| ≤ δ we have that [f (x + s) − f (x − s)] − ωf (x) < , then we have
2
Z π
1
|σn (f, x) − ωf (x)| = |Kn ∗ f − ωf (x)| = Kn (s)f (x − s) ds − ωf (x)
2π −π
Z π
1
= Kn (s) [f (x − s) − ωf (s)] ds Lebesgue integral if necessary
2π −π
Z −δ
1
≤ Kn (s) |f (x − s) − ωf (x)| ds
2π −π
Z δ
1
+ Kn (s) [f (x − s) − ωf (x)] ds
2π −δ
Z π
1
+ Kn (s)|f (x − s) − ωf (x)|ds
2π δ

π2
Consider († † +). We have Kn (s) ≤ (n+1)δ 2
for s ∈ [δ, π], so

π π
π2
Z Z
1 1
Kn (s)|f (x − s) − ωf (x)|ds ≤ 2
|f (x − s) − ωf (x)| ds
2π δ 2π δ (n + 1)δ
Z π
π
= x ∗ f˜(s) − ωf (x) ds f˜(s − x) = x ∗ f˜(s)
2(n + 1)δ 2 δ
Z π
π ˜ − ωf (x) ds = π2
≤ x ∗ f x ∗ f˜ − ωf (x)
2(n + 1)δ 2 −π (n + 1)δ 2 1
n→∞
−−−→ 0 since x ∗ f˜ − ωf (x) <∞
1

n→∞
Similiarily, († † −) −−−→ 0. Now let’s check (†). Note that

Z δ Z −δ
Kn (s) [f (x − s) − ωf (x)] ds = Kn (−s) [f (x + s) − ωf (x)(−1)] ds s 7→ −s
−δ δ
Z δ
= Kn (s) [f (x + s) − ωf (s)] ds since Kn is even
−δ

Thus we get that


Z δ
1
Kn (s) [f (x − s) − ωf (x)] ds
2π −δ
 Z δ
1 δ
Z 
1 1
= Kn (s) [f (x − s) − ωf (x)] ds + K − n(s) [f (x + s) − ωf (x)] ds
2π 2 −δ 2 −δ
Z δ   
1 1 
= Kn (s) f (x − s) + f (x + s) − ωf (x) ds
2π −δ 2

58
Hence
Z δ  
1 1 
Kn (s) f (x − s) + f (x + s) − ωf (x) ds
2π −δ 2
Z δ
1 1
= Kn (s) [f (x + s) + f (x − s)] − ωf (x) ds
2π −δ 2
Z δ Z π
1 
≤ Kn (s) ds ≤ Kn (s)ds = 
2π −δ 2π −π

So we can conclude that


Z δ
1
lim sup Kn (s) [f (x − s) − ωf (x)] ds ≤ 
n→∞ 2π −δ

Thus we find that

0 ≤ lim sup |σn (f, x) − ωf (x)| ≤ 


n→∞

But since  > 0 was chosen arbitrarily, we find that

lim sup |σn (f, x) − ωf (x)| = 0 ⇒ lim |σn (f, x) − ωf (x)| = 0


n→∞ n→∞

2. Note that if f is continuous on the compact interval [a, b], then f is uniformly continuous on [a, b]
and thus given  > 0 in (1) above, there is a δ > 0 chosen in the uniform sense for all x ∈ (a, b).
Thus, we can proceed in precisely the same manner as above and get the desired result.

n
X
Proposition 2.4.10. Let f ∈ L(T), x ∈ R. If sn (f, x) = cj (f )eijx converges as n → ∞ then
j=−n

lim sn (f, x) = lim σn (f, x)


n→∞ n→∞

Proof. If a sequence of real (complex) numbers converges, then so to does the sequence of Cesaro means,
with the same limit point.

Corollary 2.4.11 (Corollary to Fejer’s Theorem). If f ∈ L(T) such that ωf (x) exists and lim sn (f, x)
n→∞
converges, then lim sn (f, x) = ωf (x)
n→∞

Note: There are improvements to Fejer’s theorem, and we will present some of them as follows
Definition 2.4.12. Let f ∈ L[a, b], and x ∈ (a, b) We say x is a Lebesgue point of f if
Z h  
1 1
lim f (x − s) + f (x + s) − f (x) ds exists
h→0 h 0 2

A fact that will be shown in PMath 454 is that every x ∈ (a, b) is a Lebesgue point for f .

59
Theorem 2.4.13 (Lebesgue-Fejer Theorem). If f ∈ L(T), x ∈ [−π, π] is a Lebesgue point for f , then

lim σn (f, x) = f (x)


n→∞

Hence this statement of convergence holds for almost every x ∈ [−π, π].

The proof for this will be omitted as it is beyond the scope of this course.

2.5 Coefficients of Fourier Series


A natural question that arises in our study is if (cn )∞
n=1 is a sequence of real (complex) numbers, is there
a f ∈ L1 (T) such that Z π
1
cn = cn (f ) = f (t)e−int dt for n ∈ N
2π −π
Lemma 2.5.1. If f ∈ L1 (T) and n ∈ Z then |cn (f )| ≤ kf k1

Proof. Consider the following calculation


Z π Z π
1 1
|cn (f )| = f (t)e−int dt ≤ |f (t)| |e−int |dt
2π −π 2π −π
Z π
1
≤ |f |dt = kf k1 since |e−int | = 1
2π −π

Lemma 2.5.2 (Riemann-Lebesgue Lemma). If f ∈ L1 (T), then

lim cn (f ) = 0
|n|→∞

Proof. From (the Corollary to) the abstract summability kernel theorem, we have that lim kσn (f ) − f k1 =
n→∞
0. Thus given  > 0, there is n0 ∈ N such that kσn (f ) − f k1 <  for n ≥ n0 . P Furthermore, note that σn (f )
is in the span of e−n0 , e−n0 +1 , . . . , en0 −1 , en0 , thus we can write σn0 (f ) = nj=−n
0
b ej for some bj ∈ C.
0 j
If |k| > n0 we have
Z π
1
ck (σn0 (f ) − f ) = (σn0 (f ) − f ) e−k
2π −π
 
Z π n0
1 X
=  bj ej − f  e−k
2π −π
j=−n0
n0 Z π Z π
1 X 1
= bj e j−k
+ f e−k
2π −π 2π −π
j=−n0

= 0 + ck (f ) since j − k 6= 0 and averages to 0

Thus for k > n0 we have from our previous lemma that

|ck (f )| = |ck (σn0 (f ) − f )| ≤ kσn0 (f ) − f k1 < 

60
Thus we can conclude that lim ck (f ) = 0
|k|→∞

Corollary 2.5.3. If f ∈ L1 (T) (but will hold for L(T)), we have


Z π Z π
lim f (t) cos(nt) dt = 0, lim f (t) sin(nt) dt = 0
n→∞ −π n→∞ −π

e + e−int and so
1
 int 
Proof. Recall that cos(nt) = 2
Z π Z π Z π 
1 int −int
f (t) cos(nt) dt = f (t)e dt + f (t)e dt
−π 2 −π −π
n→∞
= π (c−n (f ) + cn (f )) −−−→ 0 by Riemann-Lebesgue

The proof for the sine is obviously similar


 
Let A(Z) = (cn (f ))n∈Z f ∈ L1 (T) . The Riemann-Lebesgue Lemma then impies that

 
A(Z) ⊂ c0 (Z) = (cn )n∈Z lim cn = 0
|n|→∞

(cn )n∈Z + (dn )n∈Z = (cn + dn )n∈Z , α(cn )n∈Z = (αcn )n∈Z
Then c0 (Z) is a Banach space with norm k(cn )n∈Z k∞ = sup |cn |. Another question we can ask ourselves is
n∈Z
if the map f 7→ (cn (f ))n∈Z : L1 (T) → c0 (Z) surjective?
From PMath 453, we have the following theorem
Theorem 2.5.4 (Open Mapping Theorem). If X, Y are Banach spaces, T : X → Y is a bounded linear
operator ( |||T ||| = sup kT xk < ∞), then if T is surjective then T is open. That is T (Br (X)) ⊃ B1 (Y )
kxk ≤ 1
 
for some r > 0 and where Br (x) = x ∈ X kxk < r .

Theorem 2.5.5 (Open Mapping Theorem). If X, Y are Banach spaces, T : X → Y is a bounded linear
operator ( |||T ||| = sup kT xk < ∞), then if T is surjective then T is open. That is T (Br (X)) ⊃ B1 (Y )
kxk ≤ 1
 
for some r > 0 and where Br (x) = x ∈ X kxk < r .

Z b
Note on notation: The majority of times f (t)dt will denote the Lebesgue integral, although if f
a
is Riemann integrable (that is, bounded and piecewise continuous) we may take this to be the Riemann
integral as well. The major exception to this convention is the case of vector-valued integrals, such as
Z π
1
h∗x= h(s)s ∗ xds
2π −π

which is always a Riemann integral.

61
 
Notes on the Open Mapping Theorem: Note that B1 (X) = x ∈ X kx − 0k < 1 is open, and so if
T is linear, bounded, and surjective, then T (B1 (X)) is open and contains 0 = T (0). Hence there is δ > 0
such that
 
T (B1 (X)) ⊃ Bδ (Y ) = y ∈ Y kyk < δ

1
rT (B1 (X)) ⊃ rBδ (Y ) taking r =
δ
T (Br (X)) ⊃ B1 (Y )

Corollary 2.5.6 (Corollary to Open Mapping Theorem).


   
A(Z) = (cn (f ))n∈Z f ∈ L1 (T) ( c0 (Z) = (cn )n∈Z lim cn = 0
|n|→∞

Proof. Recall that (L1 (T), k·k1 ) is a Banach space, and (c0 (Z), k·k∞ ) where

k(cn )n∈Z k∞ = sup |cn | = max |cn |


n∈Z n∈Z

Define T : L1 (T) → c0 (Z) by T (f ) = (cn (f ))n∈Z which is in c0 (Z) by the Riemann-Lebesgue Lemma.
Clearly T is linear, and f ∈ L1 (T), then

kT (f )k∞ = k(cn (f ))n∈Z k∞ = max |cn (f )| ≤ kf k1


n∈Z

via the Lemma prior to the Riemann-Lebesgue Lemma. Thus we have that
 
|||T ||| = sup kT (f )k∞ f ∈ L1 (T), kf k1 ≤ 1 ≤ 1

and so T is bounded. If A(Z) = T (L1 (T) is equal to c0 (Z) then the Open Mapping
Pn Theorem tells us that
there is some r > 0 such that T (Br (L1 (T))) ⊃ B1+ (c0 (Z)). Let Dn = j=−n e be the nth Dirichlet
j
Z π
1 n→∞
kernel. We saw that Ln = kDn k1 = |Dn | −−−→ ∞; however, T (Dn ) = (ck (Dn ))k∈Z and
2π −π

n Z
(
π
1 X π j−k
Z
1 −k 2π j=k
ck (Dn ) = Dn e = e =
2π −π 2π −π 0 otherwise
j=−n

thus T (Dn ) = (. . . , 0, 0, 1, 1, . . . , 1, 1, 0, 0, . . .) over −n, . . . , n so that

kT (Dn )k∞ = k(. . . , 0, 0, 1, 1, . . . , 1, 1, 0, 0, . . .)k∞ = 1

and T (Dn ) ∈ B1+ (c0 (Z)). However, (Dn )∞


n=1 6⊂ Br (L1 (T)) for any fixed r > 0. This contradicts the Open
Mapping Theorem, and thus T cannot be surjective.

62
2.6 Point-wise Convergence of Naive Fourier Series
Note that by naive Fourier series, we mean working without the notion of Cesaro means. Recall that if
f ∈ L(T), t ∈ R (usually t ∈ [−π, π]), then
n Z π
X
ijt 1
Sn (f, t) = cj (f )e = Dn (s)f (t − s)ds Lebesgue Integral
2π −π
j=−n
 1
 sin[(n+ 2 )s] s 6= 0
1
Also, Dn (s) = sin[ 2 s] for t ∈ [−π, π].
n + 1 s=0
Rπ f (t) n→∞
Lemma 2.6.1. If f ∈ L(T) satisfies −π t dt < ∞ then Sn (f, 0) −−−→ 0.

Proof. We have that

sin n + 12 s
Z π

1
Sn (f, 0) = f (0 − s)ds
2π −π sin 12 s
sin n + 12 s
Z π

1
= f (s)ds since Dirichlet kernel is even
2π −π sin 12 s

We recall that sin(x+y) = sin x cos y +sin y cos x, and so sin n + 12 t = sin(nt) cos 12 t +sin 12 t cos(nt),
    

and so
sin n + 12 t
Z π 
1
Sn (f, 0) = f (t)dt
2π −π sin 12 t
cos 12 t
Z π Z π
1 1
= sin(nt)f (t)dt + cos(nt)f (t)dt
2π −π sin 12 t 2π −π
Z π
1 n→∞
Though we note that cos(nt)f (t)dt −−−→ 0 by the Riemann-Lebesgue Lemma. Thus let us estimate
2π −π
the following integral
π cos 12 t π
Z Z
f (t) 1
f (t) dt ≤ dt since | cos t| ≤ 1
−π sin 21 t −π sin 21 t 2
Z π
f (t)
≤ π dt < ∞
−π t

2 π 1 1
Where we’ve used the fact that | sin θ| ≥ π |θ| for |θ| < 2, and so ≤ π |t| for |t| ≤ π. Thus, the
|sin 21 t|
cos 12 t
map t 7→ is in L(T) and so by corollary to the Riemann-Lebesgue Lemma, we have that
sin 12 t
π cos 21 t
Z
lim f (t) sin(nt) dt = 0
n→∞ −π sin 12 t

63
Thus in conclusion, we have that
π cos 12 t π
Z Z
1 1
Sn (f, 0) = 1 f (t) sin(nt) dt + 2π f (t) cos(nt) dt
2π −π sin 2 t −π
n→∞
−−−→ 0 + 0 = 0

Theorem 2.6.2 (Localization Principle). Let f ∈ L(T) and suppose there is an open interval I ⊂ R such
that f (x) = 0 for almost every x ∈ I. Then for x ∈ I,

lim sn (f, x) = 0
n→∞

Proof. For x ∈ I, let g = x∗(fˇ), so that g(s) = fˇ(s−x) = f (x−s) for s ∈ R. There is a small open interval
(−δ, δ) for δ > 0 such that for s ∈ (−δ, δ) implies that x − s ∈ I which entails that g(s) = f (x − s) = 0 for
almost every s ∈ (−δ, δ). Then
Z π Z −δ Z δ Z |
g(t) g(t) g(t) g(t)
dt = dt + dt + pi dt
−π t −π t −δ t δ t
1 −δ 1 π
Z Z
≤ |g(t)| dt + 0 + |g(t)| dt
δ −π δ δ
2 π
Z
≤ |g(t)| dt < ∞ by assumption that g ∈ L(T)
δ −π

By the previous lemma, we then have that lim sn (g, 0) = 0. Now we have
n→∞
Z π Z π
1 1
sn (f, x) = Dn (s)f (x − s) ds = DN (s)fˇ(s − x) ds
2π −π 2π −π
Z π Z π
1 ˇ 1
= Dn (s)x ∗ f (s) ds = Dn (s)g(s) ds
2π −π 2π −π
Z π
1
= Dn (s)g(−s) ds via inversion invariance
2π −π
Z π
1 n→∞
= Dn (s)g(0 − s) ds = sn (g, 0) −−−→ 0
2π −π
n→∞
Hence we can conclude that sn (f, x) −−−→ 0

Corollary 2.6.3. If f, g ∈ L(T) and I ⊂ R is an open interval for which f (x) = g(x) for almost every
x ∈ I, then
lim sn (f, x) exists ⇔ lim sn (g, x) exists
n→∞ n→∞

and the limits are the same for x ∈ I.

Proof. Let h = f − g, and apply the localization principle.

64
Theorem 2.6.4 (Dini’s Theorem). If f ∈ L(T) and f is differentiable at x(∈ [−π, π]), then

lim sn (f, x) = f (x)


n→∞

Proof. Let f 0 (x) denote the derivative of f at x, which exists by assumption. Given  > 0, there is a δ > 0
such that the following holds

f (x − s) − f (x)
|s| < δ ⇒ − f 0 (x) < 
−s

Thus if |s| < δ we have that

f (x − s) − f (x) f (x − s) − f (x)
− |f 0 (x)| <  ⇒ <  + |f 0 (x)|
−s s

and so the mapping s 7→ f (x−s)−f


s
(x)
is bounded (by  + |f 0 (x)|) on (−δ, δ), and let g(s) = x ∗ fˇ(s) − f (x)
and note that f (x) is constant in this expression for fixed x. Then g ∈ L(T), and moreover we have
Z π Z −δ Z δ Z π
g(t) g(t) g(t) g(t)
dt = dt + dt + dt
−π t −π t −δ t δ t
Z π Zδ
2
≤ |g(t)| dt + Cdt < ∞ similar to localization proof
δ −π −δ

Thus by the lemma, we have that lim sn (g, 0) = 0. However, as in the proof above,
n→∞

sn (g, 0) = sn (x ∗ fˇ − f (x), 0) = sn (x ∗ fˇ, 0) − sn (f (x), 0)


Z π
1
= sn (x ∗ fˇ0) − f (x) since c0 (f (x)) = f (x)dt = f (x)
2π −π
= sn (f, x) − f (x)

and so we conclude that lim sn (f, x) = f (x).


n→∞

2.7 Inner Product and Hilbert Spaces


Definition 2.7.1. Let X be a complex vector space. An inner product < ·, · >: X × X → C is a map
which satisfies for f, g ∈ X and α ∈ C

1. (Non-negativity) < f, f >≥ 0

2. (Non-degeneracy) < f, f >= 0 ⇔ f = 0

3. (Additivity) < f + g, h >=< f, h > + < g, h >

4. (Scalar Homogeneity) < αf, g >= α < f, g >

5. (Skew Symmetry) < f, g >= < g, f >

65
Note that by axiom (3) and (5), we can conclude < f, g + h >=< f, g > + < f, h >. Furthermore, (4)
and (5) tells us that < f, αh >= ᾱ < f, h >.

We call the pair (X, < ·,√· >) an inner product space . Furthermore, we can define the induced
norm for f ∈ X by kf k = < f, f >. Note that while we have called it a norm, we have yet to prove as
such. Before we can prove that this is a norm, we must introduce a few results.
Theorem 2.7.2 (Cauchy-Schwartz Inequality). If X is an inner product space and f, g ∈ X, then

|< f, g >| ≤ kf kkgk

Proof. Fix t ∈ R so that t̄ = t. We have that

0 ≤< tf + g, tf + g > =< tf, tf > + < tf, g > + < g, tf > + < g, g > via non-negativity and additivity
= t kf k + t < f, g > +t< g, f > + kgk = t kf k + 2t<(< f, g >) + kgk2
2 2 2 2 2
√ p
≤ t2 kf k2 + 2t| < f, g > | + kgk2 since x ≤ x2 ≤ x2 + y 2 = |x + iy|

We have then that p(t) is a quadratic function from R → R satisfying p(t) ≥ 0. Thus the discriminant
4| < f, g > |2 − 4kf kkgk ≤ 0 which implies the desired inequality.

Note that we have |hf, gi| = kf kkgk ⇔ f = tg for some t ≥ 0. If |hf, gi| = kf kkgk then the discriminant

(2|hf, gi|)2 − 4kf k2 kgk2

|hf, gi|
is precisely zero, so for t0 = − we have
kf k2

0 ≤ ht0 f + g, t0 f + gi ≤ 0 ⇒ ht0 f + g, t0 f + gi = 0

which implies that t0 f + g = 0 thus −t0 f = g. Note that by assumption this implies that

kf kkgk kf k
f =g⇒g= f
kf k2 kgk

Conversely, if f = tg for t ≥ 0, then

|hf, gi| = |htg, gi| = t|hg, gi| = t kgk2


= ktk g kgk = kf kkgk
p
Corollary 2.7.3. If (X, h·, ·i) is an inner product space, then kf k = hf, f i defines a norm on X
p
Proof. If f ∈ X, then hf, f i ≥ 0 ⇒ hf, f i ≥ 0 and hf, f i = 0 ⇔ f = 0 so kf k = 0 ⇔ f = 0, thus
non-degeneracy and non-negativity are satisfied.
If f ∈ X and α ∈ C then
p p p
kαf k = hαf, αf i = αᾱhf, f i = |α|2 hf, f i
p
= |α| hf, f i = |α| kf k

66
If f, g ∈ X we have that

kf + gk2 = hf + g, f + gi = kf k2 + 2<hf, gi + kgk2


≤ kf k2 + 2|hf, gi| + kgk2
≤ kf k2 + 2 kf k kgk + kgk2
= (kf k + kgk)2

Thus we can conclude that kf + gk ≤ kf k + kgk

Definition 2.7.4. A Hilbert space is a complete inner product space (H, h·, ·i) such that H is complete
under the induced norm

Examples:

1. Consider Cn for x = (x1 , . . . , xn ), y = (y1 , . . . , yn ) and define the associate inner product as
n
X
hx, yi = xi ȳi
i=1

Since Cn =R
˜ 2n , we can conclude that Cn is complete and hence a Hilbert space.

2. Let A ⊂ L(R) with λ(A) > 0. Then on L2 (A) define


Z
hf, gi = f ḡ
A

Note that by Holder’s inequality, since ḡ ∈ L2 (A) if g ∈ L2 (A) then f ḡ ∈ L2 (A), so that the inner
product is well defined. The norm is defined as
Z  1 Z 1
2 2
f f¯ =
p 2
kf k = hf, f i = |f | = kf k2
A A

Since we’ve seen that L2 (A) is complete in k·k2 , then (L2 (A), h·, ·i) is a Hilbert space.

3. Consider the set


 Z π 
2
L2 (T) = f : R → C f ∈ MC (R), 2πperiodic, |f | < ∞
−π

and the inner product given by


Z f
1
hf, gi = ḡ
2π −f

This is a Hilbert space

4. Recall the definition of the following set


 
C(T) = f : R → C f continuous and 2π periodic

67
Then define the inner product as
Z π
1
hf, gi = f ḡ can consider Riemann integral
2π −π

Now note that (C(T), k·k2 ) is not complete since this space is dense in L(T). Thus this is an inner
product space, but not a Hilbert space.

5. Define the following set



( )
X
2
`2 (Z) = (cn )n∈Z cn ∈ C, |cn | < ∞
n=−∞

the set of Z-indexed square summable sequences. Define the following inner product

X
hx, yi = xn ȳn
n=−∞

However, we should verify that this sum converges. We recall that if we can show that this series is
N
X
absolutely convergent, then it is convergent. Note that the finite sum |xn ȳn | can be considered
n=−N
as a sum in C2N +1 and so we can apply Cauchy-Schwartz, thus we get the following calculation

X N
X
|xn ȳn | = lim |xn ȳn |
N →∞
n=−∞ n=−N

N
! 12 N
! 12
X X
≤ lim |xn |2 |yn |2 <∞
N →∞
n=−N n=−N

N
!∞
X PN
Now since |xn ȳn | is increasing and bounded above, then limN →∞ n=−N |xn ȳn | exists.
n=−N N =1
Now we must establish that `2 (Z) is a vector space, under

(xn )n∈Z + (yn )n∈Z = (xn + yn )n∈Z , α(xn )n∈Z = (αxn )n∈Z

Furthermore, |xn yn | = |xn ||yn | = |xn ||ȳn | = |xn ȳn | and so



X N
X
= lim |xn + yn |2
N →∞
n=−∞ n=−N
N
X
|xn |2 + 2|xn yn | + |yn |2

≤ lim
N →∞
n=−N
N N N
!
X X X
= lim |xn |2 + 2 |xn yn | + |yn |2
N →∞
n=−N n=−N n=−N
<∞

68
Cleary (αxn )n∈Z ∈ `2 (Z) if α ∈ C and (xn )n∈Z ∈ `2 (Z). It turns out that (`2 (Z), k·k2 ) is complete
and so a Hilbert space, where


!1
X 2

k(xn )n∈Z k2 = |xn |2


n=−∞

The proof of this follows from the Plancherel Theorem.

Definition 2.7.5. Let (X, h·, ·i) be an inner-product space. A family of vectors {fi }i∈I( ⊂ X is called
1 i=j
orthogonal if hfi , fj i = 0, i 6= j. Moreover, if {fi }i∈I is called orthonormal if hfi , fj i = .
0 i 6= j

Proposition 2.7.6. If {f1 , . . . , fn } ⊂ X is orthogonal, then

kf1 + f2 + . . . + fn k2 = kf1 k2 + . . . + kfn k2

Proof. Note the following calculation

n 2 n n
X X X
fk =h fk , f` i
k=1 k=1 `=1
n X
X n n
X
= hfk , f` i = hfk , fk i
k=1 `=1 k=1
n
X
= kfk k2
k=1

which is precisely the result we desired.

Lemma 2.7.7 (Linear Approximation Lemma). Let {e1 , e2 , . . . , en } be an orthonormal set in an inner-
product space (X, h·, ·i), and let En = span {e1 , . . . , en }. If f ∈ X then

n
( )
X
dist(f, En ) = inf f− αi ei αi ∈ C
i=1
n n
!1
X X 2
2 2
= f− hf, ei iei = kf k − |hf, ei i|
i=1 i=1

69
Pn
Proof. Arbitrary elements of En are of the form g = i=1 αi ei . Let us estimate
kf − gk2 = hf − g, f − gi = f 2 − 2< (hf, gi) + kgk2
" n
# n 2
2
X X
= kf k − 2< hf, αi e i i + αi ei
i=1 i=1
" n # n
X X
= kf k2 − 2< ᾱi hf, ei i + |αi |2 by Pythagoras
i=1 i=1
n
X n
X
≥ kf k2 − 2 ᾱi hf, ei i + |αi |2 (†)
i=−n i=−n
n
!1 n
!1 n
X 2 X 2 X
2 2 2
≥ kf k − 2 |ᾱi | |hf, ei i| + |αi |2 Cauchy-Schwarz (††)
i=−n i=−n i=−n
n n n
!1 n
!1 n
X X X 2 X 2 X
2 2 2 2 2
= kf k − |hf, ei i| + |hf, ei i| − 2 |ᾱi | |hf, ei i| + |αi |2
i=−n i=−n i=−n i=−n i=−n

n

n
!1 n
! 1 2
X X 2 X 2

= kf k2 − |hf, ei i|2 +  |hf, ei i|2 − |αi |2 


i=−n i=−n i=−n
n
X
≥ kf k2 − |hf, ei i|2 († † †)
i=−n

Thus we can conclude that


 
 n 2 
X
dist(f, En )2 = inf f− αi ei αi ∈ C
 
i=−n
n
X
≥ kf k2 − |hf, ei i|2 via the above calculation
i=−n

The inequality yields equality when the following holds:


n
X
(†) ᾱi hf, ei i ∈ R
i=−n

(††) αi = hf, ei i
Pn 2
Pn 2
(† † †) i=−n |αi | = i=−n |hf, ei i| which follows from (††)

Thus we find that the inequalities become equalities exactly when αi = hf, ei i for all i = 1, . . . , n, and so
n 2
X
2
dist(f, En ) = f − hf, ei i
i=−n
Xn
= kf k2 − |hf, ei i|2
i=−n

70
Theorem 2.7.8 (Orthonormal Basis Theorem). Let X be an infinite dimensional inner product space,
and {ei }∞
i=1 is an orthonormal set in X. Then the following are equivalent

n ∞
( )
span {ei }∞
X [
1. i=1 = αi ei n ∈ N, αi ∈ C = span {e1 , . . . , en } is dense in X with respect to
p i=−n n=1
kf k = hf, f i
2. For every f ∈ X we have that

X
kf k2 = |hf, ei i|2 Bessel’s Inequality
i=−∞

3. For every f ∈ X

X
f= hf, ei iei
i=−∞
n
X
That is, lim f− hf, ei iei = 0
n→∞
i=−n

4. For every f, g ∈ X we have



X
hf, gi = hf, ei ihei , gi Parseval’s Identity
i=−∞

Note that in (3) we will often refer to the numbers hf, ei i as (abstract) Fourier coefficients.

Proof. (1) ⇔ (3) Let En = span {e1 , . . . , en }. Notice that En ⊂ En+1 for each n and thus

!
span {ei }∞ ∞
S [
i=1 = n=1 En is dense in X ⇔ for every f ∈ X, dist f, En = 0
n=1
⇔ ∀f ∈ X, lim dist(f, En ) = 0
n→∞
n
!
2
X
2
⇔ lim kf k − |hf, ei i| =0 Lin.Approx.Lemma
n→∞
i=−n

Which is the desired result


(2) ⇔ (3) By the linear approximation lemma, we have that
n 2 n
X X
f− hf, ei iei = kf k2 − |hf, ei i|2
i=−n i=−n

for all f ∈ X. Thus


n 2 n
X X
lim f− hf, ei iei =0 ⇔ lim |hf, ei i|2 = kf k2
n→∞ n→∞
i=−n i=−n

71
(3) ⇒ (4) Let us fix g ∈ X. Note that Γg : X → C, Γg (f ) = hf, gi is a bounded linear function. Indeed
it is linear, and moreover
   
kΓg k∗ = sup |Γg (f )| f ∈ X, kf k ≤ 1 = sup |hf, gi| f ∈ X, kf k ≤ 1

≤ kgk since |hf, gi| ≤ kf k kgk ≤ kgk

In particular Γg : X → C is continuous, and so


n
!
X
Γg (f ) = Γg lim hf, ei iei by assumption of 3
n→∞
i=−n
n
X n
X
= lim hf, ei iΓg (ei ) = lim hf, ei ihei , gi
n→∞ n→∞
i=−n i=−n

(4) ⇒ (2) We have for f ∈ X



X
hf, f i = hf, ei ihei , gi
i=−∞
X∞ ∞
X
= hf, ei ihf, ei i = |hf, ei i|2
i=−∞ i=−∞

Note: There is also a Bessel inequality, which is as follows: Let (ek )∞


k=1 be an orthonormal sequence in
an inner product space X. Then for f ∈ X we have

X
2
hf, f i = kf k ≥ |hf, ek i|2
n=1

The difference between this and (2) of the above theorem is that we have used less assumptions, and hence
do not have equality.

Proof. If En = span {e1 , . . . , en } then


n
X
0 ≤ dist(f, En )2 = f 2 − |hf, ek i|2
k=−n


X n
X
Hence |hf, ek i|2 = sup |hf, ek i|2 ≤ f 2
k=−∞ n∈N k=−n

∞ N
( )
X X X
Example: Consider `2 (Z) = x = {xn }n∈Z |xn |2 = |xn |2 = lim |xn |2 < ∞ . With
N →∞
n∈Z n=−∞ n=−N

X
the inner product defined as hx, yi = xn ȳn (a sum which always converges absolutely), consider for
n=−∞

72
each n ∈ Z the element en = (. . . , 0, 0, 1, 0, 0, . . .) where the 1 occurs in the nth position. Now for m, n ∈ Z,
we have
(
1, n = m
hen , em i =
0 otherwise

Thus {en }n∈Z is an orthonormal set. Now if x = {xn }n∈Z we have


X
hx, en i = xk en,k = xn 1̄ = xn sinceen,k 6= 0 ⇔ n = k
k=−∞

We thus have

n 2
X
x− hx, ek iek = kx − (. . . , 0, x−n , x−n+1 , . . . , xn , 0, 0 . . .)k2
k=−n

= k(. . . , x−n−2 , x−n−1 , 0, . . . , 0, xn+1 , xn+2 , . . .)k2


−n−1
X ∞
X
2
= |xk | + |xk |2
k=−∞ k=n+1
X∞ Xn
= |xk |2 − |xk |2
k=−∞ k=−n
∞ ∞
n→∞
X X
−−−→ |xk |2 − |xk |2 = 0
k=−∞ k=−∞


X
Thus we can conclude that x = hx, ek iek and so by the orthonormal basis theorem, {ek }k∈Z is dense
k=−∞
in `2 (Z).

2.8 Hilbertion Fourier Analysis


 

Recall that L2 (T) = f : R → C f ∈ MC (R), a.e 2π-periodic, −π |f |2 < ∞ . and has a norm and inner
product defined as
 Z π 1 Z π
1 2
2 1
kf k = |f | , hf, gi = f ḡ
2π −π 2π −π

Theorem 2.8.1. ek k∈Z where ek (t) = eikt is an orthonormal basis (i.e. an orthonormal sequence such


that its span is dense) in L2 (T).

73
Proof. Let us begin by showing that ek

k∈Z
is an orthonormal set. If k, ` ∈ Z, then
Z π
k ` 1
he , e i = eikt ei`t dt
2π −π
Z π Z π
1 ikt −i`t 1
= e e dt = ei(k−`)t dt
2π −π 2π −π
Z π" Z ( #
1
= cos(k − `)t + i k − `)t dt
2π −π −(
(
1, k = `
=
0, otherwise

We next want to show that the span of this set is dense in L2 (T). In fact, there are two useful proofs that
we will present.

1. (Abstract Summability Kernel Theorem Method) We saw as a Corollary to the aforementioned


theorem that
lim kf − σn (f )k2 = 0
n→∞
n j
1 X X n on
We notice that σn (f ) = ck (f )ek ∈ span ek and hence if f ∈ L2 (T) then
n+1 k=−n
j=0 k=−j
n o  n on 
dist(f, span ek = lim dist f, span ek
k∈Z n→∞ k=−n

≤ lim kf − σn (f )k2 = 0
n→∞

and hence dist(f, span ek ) = 0 so f ∈ span ek


 
k∈Z k∈Z
 
Pn
ek ek

2. (Stone-Weierstrass Theorem) Let Trig(T) = span k∈Z
= k=−n αk n ∈ Z, αk ∈ C and

refer to this as the set of trigonometric polynomials, since we have point-wise multiplication ek e` =
ek+` . We note the following:
(a) Trig(T) ⊂ C(T) is an algebra of functions
(b) Trig(T) separates points in [−π, π). This is, if t 6= s in [−π, π), then ∃k ∈ Z such that eikt 6= eiks
(c) Trig(T) contains the constant function 1
(d) Trig(T) is conjugation closed. That is, nk=−n αk ek = nk=−n ᾱk e−k ∈ Trig(T)
P P

and hence by the Stone-Weierstrass theorem, we have


k·k
Trig(T) ∞ = C(T)

Thus if f ∈ L2 (T) and  > 0, find h ∈ C(T) such that kh − f k2 < and find p ∈ Trig(T) such that
2

kp − hk∞ < . Then
2
 
kf − pk2 ≤ kf − hk2 + kh − pk2 ≤ kf − hk2 + kh − pk∞ < + = 
2 2

74
Corollary 2.8.2. If f ∈ L2 (T), then lim ksn (f ) − f k2 = 0.
n→∞

Proof. If k ∈ Z, then
Z π Z π
1 ikt 1
ck (f ) = f (t)e dt = f ek = hf, ek i
2π −π 2π −π

Since ek

k∈Z
is an orthonormal basis for L2 (T) we have

n
X
kf − sn (f )k2 = f − ck (f )ek
k=−n 2
n
n→∞
X
= f− hf, ek iek −−−→ 0 by Orthonormal Basis Theorem
k=−n 2

Warning: If the dimension of the space is infinite, the an orthonormal basis is not necessarily a linear
(Hamel) basis. We saw that if ek (t) = eikt for k ∈ Z then ek k∈Z is an orthonormal basis for L2 (T).
Recall that L2 (T) ⊂ L1 (T) by Holder’s inequality.

Theorem 2.8.3 (Riesz-Fischer Theorem). Let f ∈ L1 (T), then


X
f ∈ L2 (T) ⇔ |ck (f )|2 < ∞
k=−∞

That is, if (ck (f ))k∈Z ∈ `2 (Z).

Proof. (⇒) By Bessel’s equality, we have


X
|ck (f )|2 = kf k22 < ∞
k=−∞


X n
X n
X
(⇐) Suppose that |ck (f )|2 = lim |ck (f )|2 < ∞. Then define fn = ck (f )ek and note
n∈N
k=−∞ k=−n k=−n

75
that fn = sn (f ). If m < n we have

n m 2
X X
kfn − fm k22 = k
ck (f )e − ck (f )ek

k=−n k=−m 2
−m−1 n 2
X X
k k
= ck (f )e + ck (f )e
k=−n k=m+1 2
−m−1
X n
X
= |ck (f )|2 + |ck (f )|2
k=−n k=m+1
−m−1
X ∞
X
≤ |ck (f )|2 + |ck (f )|2
k=−∞ k=m+1

m→∞
X
−−−−→ 0 by assumption that |ck (f )|2 < ∞
k=−∞

Hence (fn )∞
n=1 is Cauchy is L2 (T) and since L2 (T) is complete, it must have a limit in L2 (T), say
F = lim fn . We have for k ∈ Z
n→∞

n
X
ck (F ) = hF, ek i = lim hfn , ek i = lim h cj (f )ej , ek i
n→∞ n→∞
j=−n
n
X
= lim hej , ek i = ck (f )
n→∞
j=−n

By a corollary to the Abstract Summability Kernel theorem, we have that if ck (f ) = ck (F ) for all k ∈ Z,
then f = F almost everywhere. In particular, f = F ∈ L2 (T), which is precisely what we wished to
show.

Theorem 2.8.4 (Plancherel Theorem). Let U : L2 (T) → `2 (Z) be given by U f = (ck (f ))k∈Z . Then


!1 1
2  Z π
X
2 1 2
2
kU f k2 = |ck (f )| = kf k2 = |f |
2π −π
k=−∞

that is, U is an isometry. Furthermore, U is invertible. That is, U is surjective and the inverse map
U −1 : `2 (T) → L2 (T) is a bounded linear operator. These two properties together imply that U is unitary.

Proof. This proof is immediate from Riesz-Fischer and the orthonormal basis theorem.

Corollary 2.8.5. `2 (Z) is complete and hence a Hilbert space

Proof. If c(n) = {(ck,n )k∈Z }∞


n=1 ⊂ `2 (T) is Cauchy, then


X
fn = U −1 {(ck,n )k∈Z } = ck,n ek ∈ L2 (T)
k=−∞

76
satisfies
kfn − fm k2 = kU fn − U fm k2 = c(n) − c(m) 2

the later of which is Cauchy in `2 (Z), so (fn )∞


n=1 is Cauchy in L2 (T), but L2 (T) is complete and so
f = lim fn exists and {(ck,n )k∈Z }∞
n=1 = lim (c ∞ ∞
k,n )k=−∞ so the limit (c(n) )n=1 exists in `2 (Z)
n→∞ n→∞

Spaces of (almost everywhere equivalent classes) of functions:

A(T) ⊂ C(T) ⊂ L2 (T) ⊂ L1 (T)


l l l l
`1 (Z) ⊂ C ∗ (Z) ⊂ `2 (Z) ⊂ A(Z) ⊂ c0 (Z)

But the spaces C ∗ (Z), A(Z) are ”mystery spaces” in that we don’t have an intrinsic description of these
spaces.
Spaces of Z-sequences :
Via assignment 6, we will see that D(T) ⊂ A(T) where D(T) is the set of continuous, piecewise bounded,
differentiable functions.
Consider the function F ∈ L(T) defined by F (t) = 21 − 2π t
for t ∈ [0, 2π). Note that if we want to
rewrite this in terms of our more fundamental interval [−π, π), we can write this as

(
1 t
2 − 2π , t ∈ [0, π)
F (t) = 1 t
2 − 2π , t ∈ (−π, 0)

Notice that

1h i
ωF (0) = lim F (0 + s) + lim F (0 − s)
2 s→0 s→0
1h i
= lim F (s) + lim F (−s)
2 s→0 s→0
 
1 − +
 1 1 1
= F (0 ) + F (0 ) = −
2 2 2 2
=0

Also note that F is continuous except at points 2πn for n ∈ Z.

Pn sin kt
Proposition 2.8.6. For n ∈ N, we have that sn (F, t) = k=−n kπ

Proof. Note that F is almost everywhere odd; that is, F (−t) = −F (t) for almost every t ∈ R. Thus for

77
any k ∈ Z we have
Z π
1
ck (F ) = F (t)e−ikt dt
2π −π
Z π Z π 
1
= F (t) cos(kt)dt = i F (t) sin(kt)dt
2π −π −π
Z π
i
=− F (t) sin(kt)dt by the fact that F (t) cos(kt) is odd
2π −π
Z 2π
i
=− F (t) sin(kt) dt
2π 0
Z 2π  
i 1 t
=− − sin(kt) dt
2π 0 2 2π
Z 2π
i
= t sin(kt) dt
(2π)2 0
" #

1 2π
Z
i t
= − cos(kt) + cos(kt)dt integration by parts
(2π)2 k 0 k 0
 
i 2π 1
= 2
− cos(2πk) + 0 =
(2π) k 2πik

We note that here we’ve used the fact that


Z π Z π Z π Z π
F (t) cos(kt)dt = F (−t) cos(−kt)dt = − F (t) cos(kt)dt ⇒2 F (t) cos(kt) = 0
−π −π −π −π

Thus we have
−1
X n
X
sn (F, t) = ck (f )eikt + ck (f )eikt
k=−n k=−n
Xn h i
= c−k (f )e−ikt + ck (f )eikt
k=−n
n h
X i 1
= −ck (f )e−ikt + ck (f )eikt since c−k (f ) = − = −ck (f )
2πik
k=−n
n n
X X sin(kt)
= ck (f )2i sin(kt) =

k=−n k=−n

Definition 2.8.7. For f ∈ L(T) we say that f is boundedly piecewise differentiable if sup f 0 (t) < ∞
Lemma 2.8.8. 1. lim sn (F, 0) = 0 = ωF (0)
n→∞

2. lim sn (F, t) = F (t) for t 6= 0


n→∞

Moreover, for any compact interval I ⊂ R\{2πn}n∈Z we have


lim sn (F, T ) = F (t) uniformly for t ∈ I
n→∞

78
n
X sin(0) n→∞
Proof. 1. sn (F, 0) = = 0 −−−→ 0

k=−n

2. We may suppose, by 2π periodicity that I = [a, b] ⊂ (0, 2π). Find δ > 0 such that 0 < δ < a and
b < 2π − δ so that [a, b] ⊂ (δ, 2π − δ). Define Fδ ∈ C(T) by

F (t) = 1 − t , t ∈ (δ, 2π − δ)
2 2π
Fδ (t) = for t ∈ [−δ, 2π − δ]
 δ 1 −t δ , t ∈ [−δ, δ]
( 2 2π )

Then Fδ ∈ C(T) and is boundedly piecewise differentiable on [−π, π). Thus Fδ ∈ D(T), and so
lim sn (Fδ , t) = Fδ (t) uniformly on R (Assignment 6). Thus if t ∈ [a, b] ⊂ (δ, 2π − δ) then by the
n→∞
Localization Principle
lim sn (F, t) = lim sn (Fδ , t) = Fδ (t) = F (t)
n→∞ n→∞

which holds uniformly for t ∈ [a, b].

Theorem 2.8.9. Suppose that f ∈ L(T) such that f is boundedly piecewise differentiable (we do not
assume that f is continuous). Then for t ∈ R we have that

f (t+ ) = lim f (s), f (t− ) = lim f (s)


s→t+ s→t−

exist for each t ∈ R and


1 +
f (t ) + f (t− )

lim sn (f, t) = ωf (t) =
n→∞ 2
In particular, if f is continuous at t, then ωf (t) = f (t). Moreover, if I ⊂ R is a compact interval which
excludes any point of discontinuity of f , then

lim sn (f, t) = f (t) uniformly for t ∈ I


n→∞

Proof. Clearly if f is continuous at t, then f (t+ ), f (t− ) exists and are equal to f (t). Let t1 < . . . < tm
denote the set of points of discontinuity in [−π, π). For convenience, let t0 = tm − 2π, tm+1 = t + 1 + 2π.
Let us see for j = 1, . . . , m that f (t− + −
j ), f (tj ) exists. Recall that f (tj ) = lim f (s). Choose any sequence
s→t−
tj−1 < s1 < . . . < tj with lim sn = t. We have for m < n in N
n→∞

|f (sn ) − f (sm )| = f 0 (s∗ )(sn − sm ) , s∗ ∈ (sn , sm ) by mean value theorem


0 ∗ 0
= |f (s )||sn − sm | ≤ sup |f (s)||sn − sm |
s∈[−π,π]

= M < ∞ since f is boundedly piecewise differentiable


 ∞
Since (sn )∞
n=1 is Cauchy, we find that f (sn ) is also Cauchy and hence converges. Since we can do
n=1
this for an arbitrary sequence sn → t− − +
j we find that f (tj ) = lim f (s) exists. Similarily f (tj ) exists.
s→t−
j

Let I ⊂ R be a compact interval which does not contain any points of discontinuity. By 2π-periodicity,
we may assume that I ⊂ [−π, π) 6= ∅. Write the point of discontinuity of f in [−π, π) by t1 < . . . < tm ,

79
and for convenience let t0 = tm − 2π, tm+1 = t1 + 2π. We can write I = [a, b], tj < a < a0 < b0 < b < tj+1 ,
for some j = 0, 1, . . . , m. Define h ∈ C(T) by the following
(
f (t), t ∈ [a, b]
h(t) = for t ∈ [a, a + 2π]
M t + B, t ∈ (b, a + 2π)

where M, B are chosen such that M b + B = f (b), M (a + 2π) + B = f (a). It is clear that h is continuous,
and piecewise boundedly differentiable; that is, h ∈ D(T) and from assignment 6, we have that

lim sn (f, t) = h(t) uniformly on R


n→∞

Thus by localization principle for t ∈ [a0 , b0 ] ⊂ (a, b) we have

lim sn (f, t) = lim sn (h, t) = h(t) = f (t)


n→∞ n→∞

uniformly for t ∈ [a0 , b0 ].


Now we must consider the discontinuities t1 , . . . , tm . Define at tj the gap in f as

γ = γf (tj ) = f (t+
j ) − f (tj )
(
t 1
− 2π ,2 t ∈ [0, π)
and recall that we defined F (t) = 1 t
has the property that ωF (0) = 0, γF (0) = 1,
− 2 − 2π , t ∈ [−π, 0)
and F is continuous except at 2nπ. Furthermore,
(
ωF (0) = 0, t = 0
lim sn (F, t) = t ∈ [−π, π)
n→∞ F (t, t 6= 0

and convergence is uniform on compact intervals I 63 2nπ. Define


(
f (t) = γF (t − tj ), t 6= tj
g(t) =
ωf (tj ), t = tj

Then

g(t−
j ) = lim g(s) − lim [f (s) − γF (s − t)]
s→t− s→t−
h i  1
= f (t−
j )

− γF (0 ) = f (t−
j ) − f (t+
j ) − f (t−
j ) −
2
1h + i
= f (tj ) − f (t−
j ) = ωf (t) = g(tj )
2
Similarily, g(t+ 1
j ) = ω(tj ) = g(tj ) and hence g is continuous at tj . Let 0 < δ < 2 min {tj − tj−1 , tj+1 − tj },
and on [tj − δ, tj + δ] , the same localization principle and assignment 6 (Q2) technique used above allows
us to conclude that
lim sn (g, t) = g(t) uniformly for t ∈ [tj − δ, tj + δ]
n→∞

We have g = f − γ(tj ∗ F ) almost everywhere, so f = g + γ(tj ∗ F ) almost everywhere and thus for
t ∈ [tj − δ, tj + δ]
lim sn (f, t) = lim [sn (g, t) + γsn (tj ∗ F, t)]
n→∞ n→∞

80
where we recall that
Z π
1
sn (tj ∗ F, t) = Dn (s)tj ∗ F (t − s) ds
2π −π
Z π Z π
1 1
= Dn (s)F ((t − s) − tj ) ds = Dn (s)F ((t − tj ) − s) ds
2π −π 2π −π
= sn (F, t − tj )
Hence for t ∈ [tj − δ, tj + δ] we get
lim sn (f, t) = lim [sn (g, t) + γsn (F, t − tj )]
n→∞ n→∞
(
g(t) + γF (t − tj ), t 6= tj
=
ωf (tj ) + γωF (0), t = tj
(
f (t), t 6= tj
=
ωf (tj ), t = tj

Corollary 2.8.10. If f ∈ L(T) and there is t0 ∈ R and δ > 0 such that f is boundedly piecewise
differentiable and continuous on (t0 − δ, t0 + δ), then for any compact interval I ⊂ (t0 − δ, t0 + δ), the
Fourier series converge uniformly, that is
lim sn (f, t) = f (t) uniformly for t ∈ I
n→∞

Proof. Embedded in the proof above

2.9 Behavior of Fourier Series Near Jump Discontinuities


(
1 t
− 2π
2t ∈ [0, π)
,
Recall that F (t) = 1 t
for t ∈ [−π, π). Which will be useful in the following result
− 2 − 2π
t ∈ [−π, 0)
,
 π  1 Z π sin x
Lemma 2.9.1 (Gibbs Lemma). lim sn F, = dx. Thus we have that
n→∞ n π 0 x
h  π  π i 1 Z π sin x 1
lim sn F, −F = dx −
n→∞ n n π x 2
| 0 {z }
Gs ≈0.0898

Pn sin kt
Proof. Recall that sn (F, t) = k=−n πk , thus
n
 π X sin kπ
n
sn F, =
n kπ
k=−n
n
1 X sin kπ
n π
= kπ
π n
n
k=−n
Z π
n→∞ 1 sin x
−−−→ dx
π 0 x

81
n j
1 X X sin kt
σn (F, t) =
n+1 kπ
j=0 k=−j
n  
X k sin kt
= 1−
n+1 kπ
k=−n
 π  1 π sin x sin x 
Z 
lim σn F, = − dx
n→∞ n π 0 x π
 π π 
n→∞
σn F, −F −−−→ Gσ = −0.11
n n

82
Index
Lp - space, 21 function, 9

conjugate index, 21 norm, 2


Fejer kernel, 52
inner product operator
definition, 65 bounded, 30
space, 66 ordinal, 4
p-norm, 21 orthogonal vectors, 69

Banach space real


definition, 2 extended real numbers, 11
bounded Riemann
essentially, 26 integrability
definition, 2
Cantor, 6 integral
Category improper, 29
Baire’s Theorem, 50 sum, 2
First, 50
Second, 50 summability kernel, 54
Cauchy
translate, 7
Criterion, 2
sequence well-order, 4
definition, 2
Cesaro average, 51
Complete space
definition, 2
convolution, 43

Fourier
coefficient, 40
function
simple, 13
functional calculus, 10

Hilbert space, 67

L space
L∞ , 26
Lebesgue
point, 59
Integral
integrable, 15
Proto-, 13

measurable
complex function, 38

83

You might also like