See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/344486946

Introduction to Fourier Series

Article · October 2020

CITATIONS READS
2 1,879

1 author:

Zhuo Chen
Imperial College London
1 PUBLICATION 2 CITATIONS

SEE PROFILE

All content following this page was uploaded by Zhuo Chen on 06 October 2020.

The user has requested enhancement of the downloaded file.

 Introduction to Fourier Series
Zhuo Chen, Xueying Wang, Yunfei Wang, and Tianchen Zhao
Department of Mathematics, Imperial College London

Abstract

Fourier series has always been a heated topic in mathematics and physics.
This article summarizes contents in Fourier Analysis: an Introduction,
from the discovery that some functions can be written as the sum of sin
and cos, into the smoothness condition of its convergence and provides an
insight into its application in different fields such as number theory and
physics, which works as the guide for the development of Fourier analysis.

Contents

1 Introduction 3

2 The Discovery of Fourier Series 4

2.1 Wave equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.1.1 Derivation of wave equation . . . . . . . . . . . . . . . . . 5

2.1.2 Solutions of wave equation . . . . . . . . . . . . . . . . . 6

2.2 The Heat Equation . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.2.1 Derivation of heat equation . . . . . . . . . . . . . . . . . 10

2.2.2 Solution of the steady-state heat equation in unit discs . . 10

2.3 Schrödinger equation . . . . . . . . . . . . . . . . . . . . . . . . . 13

3 Properties of Fourier Series 14

3.1 Fourier Trigonometric Series . . . . . . . . . . . . . . . . . . . . . 15

1
 3.2 Fourier Exponential Series . . . . . . . . . . . . . . . . . . . . . . 16

3.3 Restrictions on f . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3.3.1 Riemann integrability . . . . . . . . . . . . . . . . . . . . 16

3.3.2 Functions on the circle . . . . . . . . . . . . . . . . . . . . 17

3.4 Examples of Fourier Series . . . . . . . . . . . . . . . . . . . . . . 18

3.5 Uniqueness of Fourier series . . . . . . . . . . . . . . . . . . . . . 22

3.6 Convergence of Fourier series . . . . . . . . . . . . . . . . . . . . 24

3.7 Convolutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3.8 Good kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.9 The Dirichlet kernel . . . . . . . . . . . . . . . . . . . . . . . . . 31

3.10 Cesàro summability . . . . . . . . . . . . . . . . . . . . . . . . . 32

3.11 Abel summability . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

3.12 Gibbs phenomenon . . . . . . . . . . . . . . . . . . . . . . . . . . 40

3.13 Dirichlet’s problem in the unit disc . . . . . . . . . . . . . . . . . 40

4 Convergence of Fourier Series 42

4.1 Mean-square convergence . . . . . . . . . . . . . . . . . . . . . . 42

4.1.1 Vector space and inner product . . . . . . . . . . . . . . . 42

4.1.2 Proof of Mean-Squared Convergence . . . . . . . . . . . . 47

4.2 Pointwise Convergence . . . . . . . . . . . . . . . . . . . . . . . . 52

5 Application of Fourier Series 53

5.1 The Isoperimetric problem . . . . . . . . . . . . . . . . . . . . . . 53

5.1.1 The isoperimetric inequality . . . . . . . . . . . . . . . . . 53

5.1.2 Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

5.1.3 Length of the curve . . . . . . . . . . . . . . . . . . . . . 54

2
 5.1.4 Green’s theorem . . . . . . . . . . . . . . . . . . . . . . . 54

5.1.5 Area bounded by the curve . . . . . . . . . . . . . . . . . 55

5.1.6 Isoperimetric inequality . . . . . . . . . . . . . . . . . . . 56

5.2 Weyl’s equidistribution theorem . . . . . . . . . . . . . . . . . . . 58

5.3 A continuous but nowhere differentiable function . . . . . . . . . 65

5.4 The heat equation on the circle . . . . . . . . . . . . . . . . . . . 69

5.4.1 Useful tool: separation of variables . . . . . . . . . . . . . 71

5.4.2 Heat kernel . . . . . . . . . . . . . . . . . . . . . . . . . . 73

1 Introduction

Although the idea was motivated only to solve the heat equation, Fourier series
was then proven to be useful in many fields such as differential equations, algebra
and electrical engineering. Based on Fourier Analysis: an Introduction, we will
introduce the motivation, properties, convergence and applications of Fourier
series.

Not only to the heat equation, Fourier series applies to wave equation, Schrödinger
equation and so on. In fact, most linear differential equations with solutions in
trigonometric functions or exponentials have solutions in the form of Fourier
series, which we will explore in section 2.

Also, Fourier series of a given function doesn’t always converge to it. In sections
3 and 4, we will explore the smoothness conditions and we will discuss the
pointwise and uniform convergence.

Next, Fourier series discomposes a periodic function into a linear combination

or infinite sum of sin and cos functions, which is an orthogonal set of functions
of unit norm and analogous to orthonormal basis in linear algebra, that is,
Fourier series is a generalized orthonormal basis for periodic functions. We will
introduce this concept in section 4.

Finally, we will give an insight into different applications of Fourier series in

many fields such as geometry, number theory, analysis and physics. It meanwhile
provides guidance applications of Fourier analysis.

3
Acknowledgement

We would like to express our thanks to Doctor Disheng Xu for his help in reading
and guidance for writing the report. We also want to thank Imperial College
for this wonderful experience in UROP.

2 The Discovery of Fourier Series

This section introduces how mathematicians approached the idea of Fourier se-
ries. Along with the development of physics, mathematicians found many linear
differential equations having solutions in the form of trigonometric functions,
which can be sum up to a series.

We will go through three important differential equations in physics in details,

including the derivation, the relationship with Fourier series, the solutions and
some examples.

2.1 Wave equation

Waves are considered as derivation from equilibrium that oscillates periodically.

We now introduce two important types of wave, standing waves and travelling
waves: the first with constant peak amplitude and the other with changing peak
amplitude.

Travelling waves consist of an initial profile F (x), the profile at t = 0, and are
travelling at a rate of c as t changes. It is characterized as the following.

u(x, t) = F (x − ct)

Standing waves consist of an initial profile φ(x) and an amplifying factor ψ(t).
The displacement u(x, t) is therefore their product.

u(x, t) = φ(x)ψ(t)

Note that x and t are completely separable, an important characteristic of stand-

ing waves in solving the wave equations.

4
2.1.1 Derivation of wave equation

Wave equation concerns with the displacement x, the height u(x, t) and time t
of a string hanging with both ends fixed. We can assume without the loss of
generality that it is fixed at x = 0 and x = π, which gives us the first initial
conditions. The heights at the two points are zero.

u(0, t) = u(π, t) = 0 (2.1)

Moreover, the string is steady at the beginning, that is, the initial velocity is
zero.

u(x, 0) = 0 (2.2)

We now start the derivation. We will first look at the discrete case, that is, we
consider several N particles with little masses connected by sticks. This is a
microcosm of the string. These particles have horizontal coordinates xi equally
distributed in x = 0 and x = π. y = u(x, t) denotes the displacement in given x
π
and t. The distance between each neighbor particle is h = N : x0 , x1 , ...xN and
xn = nh, yn = u(xn , t).

Our final result is based on Newton’s Second Law F = ma. The total force F is
the sum of the forces from the left and right, which can be obtained by Hooke’s
Law. For the nth particle,

 
τ
FL = (yn+1 − yn )
h
 
τ
FR = (yn−1 − yn )
h

where τ is a constant. Let ρ be the density of the string.

Assume the density of the sticks with respect to the horizontal axis is ρ. The
right hand side is then

ρhyn00 (t)

Equalling both sides, we have

 
τ
(yn+1 − 2yn + yn−1 ) = ρhyn00 (t)
h

5
  
yn+1 − 2yn + yn−1
τ = ρyn00 (t)
h2

q
τ
Let h → 0 and we have the following equation with c = ρ.

∂2u ∂2u
τ = ρ
∂x2 ∂t2
1 ∂2u ∂2u
2 2
=
c ∂t ∂x2

Make substitution x = X(just for better format) and t = bT with b = 1c . The

equation is also arranged into the standard form of the wave equation.

∂2u ∂2u
= (2.3)
∂T 2 ∂X 2

2.1.2 Solutions of wave equation

Now we are able to apply the definition of two kinds of waves into the problems as
each of them represents a solution. The one using the travelling waves is more
comprehensive, while the one with standing waves introduce us into Fourier
series.

Note that the forms of travelling waves F (x + t) or F (x − t) solve the equation

when F is twice differentiable. In fact, all solutions can be written as the
sum of F (x + t) and G(x − t) for twice differentiable F and G. We show
this by substitution m = x + t, n = x − t so u can be written in m and n,
u = u(x, t) = u(m, n). With chain rule,

∂u 1 ∂u 1 ∂u
= +
∂m 2 ∂x 2 ∂t

∂2u 1 ∂ ∂u 1 ∂ ∂u 1 ∂2u 1 ∂2u

= + = − =0
∂m∂n 2 ∂n ∂x 2 ∂n ∂t 4 ∂x2 4 ∂t2

Therefore, all solutions can be written as travelling waves. Let’s then consider
the initial conditions. We have u(x, 0) = f (x) as f (x) is the initial profile.
u(x, 0) = u(π, t) = 0 as the string is fixed at x = 0 and x = π. Next, ∂u
∂t (x, 0) =
g(x), in which g(x) represents the initial velocity and g(0) = g(π) = 0, which
leads to the following equations (note that these should only apply to interval
[0, π], but we can extend the function to be odd and periodic of 2π).

6
 
F (x) + G(x) = f (x)
F 0 (x) − G0 (x) = g(x)

Differentiate the first equation and we solve for F and G by elimination.

 Z x 
1
F (x) = f (x) + g(u) du + C1
2 0
 Z x 
1
G(x) = f (x) − g(u) du + C2
2 0

Note that F (x) + G(x) = f (x) implies C1 + C2 = 0. Substitute in u(x, t) =

F (x + t) + G(x − t) and we get d’Alembert’s formula.

Z x+t
1  1
u(x, t) = f (x + t) + f (x − t) + g(u) du (2.4)
2 2 x−t

We now turn to the second method. When one guesses for solutions of a dif-
ferential equation, it is common to use separation of variables. In this case, we
can assume the solution u(x, t) can be written as φ(x)ψ(t), which coincides our
definition of standing waves. Substitute and we get

φ(x)ψ 00 (t) = φ00 (x)ψ(t)

ψ 00 (t) φ00 (x)
=
ψ(t) φ(x)

Note that the left-hand side only depends on x and the other side only depends
on t. The equality holds only when both are a constant λ.

ψ 00 (t) φ00 (x)

= =λ
ψ(t) φ(x)

ψ 00 (t) − λψ(t) = 0

(2.5)
φ00 (x) − λφ(x) = 0

If λ > 0, u(x, t) doesn’t oscillate so we assume λ < 0 and let λ = −m2 .

ψ(t) = A cos mt + B sin mt

φ(x) = Ā cos mx + B̄ sin mx

7
As the string is fixed at x = 0 and x = π, φ(0) = φ(π) = 0 ⇒ Ā = 0. Then
B = 0 is trivial so assume B 6= 0 ⇒ m ∈ Z. Next, m = 0 is trivial and m ≤ −1
can be ignored by symmetry, so assume m ≥ 1 and arrange the above results as

um (x, t) = (Am cos mt + Bm sin mt) sin mx

Recall the definition of an equation being linear: if u and v are two solutions,
so is αu + βv with α and β two constants. Due to the linearity of derivatives,
the wave equation is linear as well, so we can superimpose the above results as
a sum.

∞
X
u(x, t) = (Am cos mt + Bm sin mt) sin mx (2.6)
n=1

Moreover, given the initial profile f (x), we have

∞
X
f (x) = u(x, 0) = Am sin mx
m=1

The convergence of the infinite sum is to be proven in sections 3 and 4. Although

unsure about the convergence, this problem leads to the study of Fourier series.

Also, to solve for Am , we integrate

Z π ∞ Z π
X π
f (x) sin nx dx = Am sin mx sin nx dx = An ·
0 m=1 0 2

Z π
2
An = f (x) sin nx dx (2.7)
π 0

Example 2.1 (the plucked string). Let’s consider an example of plucked string
as in the diagram and with c = 1. The string has initial velocity 0 and is plucked
at x = π/4, y = 1, given by the following formula.


4
 π x,

 0 ≤ x ≤ π/4




f (x) = − π4 x + 2h, π/4 ≤ x ≤ π/2 (2.8)






0, π/2 ≤ x ≤ π

8
 h
O
π/4 π/2 π

Figure 1: the plucked string

We are given the initial position f (x) and initial velocity g(x) = 0 to find Am
and Bm in (2.6). First, we have

∞
X
f (x) = u(x, 0) = An sin nx
n=1
π
8 2 sin nπ/4 − sin nπ/2
Z
2
An = f (x) sin nx dx =
π 0 π2 n2

Differentiate (2.6) with respect to t to get the velocity

∞
∂u X
(x, t) = (−mAm sin t + mBm cos mt) sin mx
∂t n=1

∞
∂u X
(x, 0) = mBm sin mx = 0
∂t n=1

⇒ Bm = 0
∞
X 8 2 sin nπ/4 − sin nπ/2
u(x, t) = Am cos mt sin mx, An =
n=1
π2 n2

2.2 The Heat Equation

The heat equation deals with the temperature u(x, y, t) on a plane, which also
leads to Fourier series. First, we will derive the time-dependent heat equation
and then solve it when t = 0, which is known as the steady-state heat equation.

9
2.2.1 Derivation of heat equation

Similar to wave equation, we begin with a sufficiently small piece of square S

centered at (x0 , y0 ). The amount of heat at t is given by
ZZ
H(t) = σ u(x, y, t)dxdy
S

in which σ is a positive constant, the specific heat of the material. We differen-

tiate both sides to find the heat flow.
ZZ
∂H ∂u
=σ dxdy
∂t S ∂t

For sufficiently small h, the above can be approximated by

∂u
σh2 (x0 , y0 , t)
∂t
By Newton’s Law of Cooling, the heat flows through S can also be written as

∂H ∂u ∂u
= κh (x0 + h/2, y0 , t) − (x0 − h/2, y0 , t)+
∂t ∂x ∂x

∂u ∂u
(x0 , y0 + h/2, t) − (x0 , y0 − h/2, t)
∂y ∂y
where κ > 0 is the conductivity of the material. Let h → 0 and we have
!
∂H 2 ∂2u ∂2u
= κh + 2
∂t ∂x2 ∂y

Therefore, we have the following equation and arrange into the time-dependent
heat equation. !
∂H 2 ∂u 2 ∂2u ∂2u
= σh = κh + 2
∂t ∂t ∂x2 ∂y

σ ∂u ∂2u ∂2u
= + 2 (2.9)
κ ∂t ∂x2 ∂y

2.2.2 Solution of the steady-state heat equation in unit discs

Steady state means the heat is in equilibrium and no heat flow, which gives the
steady-state heat equation.
∂2u ∂2u
+ 2 =0 (2.10)
∂x2 ∂y

10
Let’s consider the equation in a unit disc. This is known as the Dirichlet prob-
lem. Although we can solve the equation in rectangular coordinates, it is too
complicated to consider the boundary conditions. Therefore, we first change to
polar coordinates. Using chain rules, we have the following results.

∂u ∂u ∂x ∂u ∂y ∂u ∂u
= + = −r sin θ + r cos θ
∂θ ∂x ∂θ ∂y ∂θ ∂x ∂y
∂u ∂u ∂x ∂u ∂y ∂u ∂u
= + = cos θ + sin θ (2.11)
∂r ∂x ∂r ∂y ∂r ∂x ∂y
∂2u ∂2u ∂2u
2
= cos2 θ 2 + sin2 θ 2 (2.12)
∂r ∂x ∂y
∂2u ∂u ∂u 2
2
2 ∂ u 2
2
2 ∂ u
= −r cos θ − r sin θ + r sin θ + r cos θ (2.13)
∂θ2 ∂x ∂y ∂x2 ∂y 2

Combining (2.11), (2.12) and (2.13), we get

∂2u ∂2u ∂ 2 u 1 ∂u 1 ∂2u

2
+ 2 = 2
+ + 2 2 =0
∂x ∂y ∂r r ∂r r ∂θ

∂2u ∂u ∂2u
r2 2
+r =− 2
∂r ∂r ∂θ

Again, we apply the separation of variables and let u(r, θ) = F (r)G(θ).

F 00 (r) + rF 0 (r) G00 (θ)

r2 =−
F (r) G(θ)

The left and right hand sides are each function of r and θ, so the equality holds
only when they are both equal to λ, a constant.

G00 (θ) + λG(θ) = 0


(2.14)
r F (r) + rF 0 (r) − λF (r) = 0
2 00

If λ ≤ 0, it doesn’t oscillate so assume λ = m2 > 0 and m ∈ Z. We solve for

G(θ).
G(θ) = Aeimθ + Be−imθ

To solve for F , we substitute r = ex and by chain rule,

dF dF
= e−x
dr dr

11
 d2 F dF d2 F
2
= −e−2x + e−2x 2
dr dx dx

Thus,

d2 F
− λF = 0
dx2

dx
We further substitute u = dF and the equation is now easy to solve.

d2 F 1 du 1 dF du 1 du
2
=− 2 =− 2 =− 3
dx u dx u dx dF u dF
d2 F
= λF
dx2

Two simple solutions for F are rm and r−m when m 6= 0, 1 and log r when
m = 0. Note that F (r) is unbounded at the origin if the exponential is negative
which disobeys our intuition. We assume positive exponential. Following the
same reason, we reject m = log r.

F (r) = r|m|

We yield solutions in the forms of um (r, θ) = r|m| eimθ As the heat equation is
linear, we superimpose these solutions as an infinite sum. Again, we leave the
convergence to be proven in the next section.

∞
X imθ
u(r, θ) = am r|m|e (2.15)
m=−∞

Take r = 1 and we again encounter a problem in Fourier series.

∞
X
u(1, θ) = am eimθ
m=−∞

Exactly the same mechanism works for any harmonic functions as all of them
satisfy Laplace’s equation (2.8). Therefore, we will be seeing this solution in
Fourier series also in many other fields such as electromagnetism, fluid dynamics
and complex analysis.

12
2.3 Schrödinger equation

Schrödinger equation describes the wave function Ψ(x, y, z, t) of a particle in

space, which depends on its position and time. It follows

~2 2 ∂Ψ
∇ Ψ + V Ψ = i~ (2.16)
2m ∂t

We solve it using separation of variables. Write the wave function as the follow-
ing and substitute into (2.16)

Ψ(x, y, z) = ψ(x, y, z)T (t)

~2 dT
T ∇2 ψ + V T Ψ = i~ψ
2m dt
~2 1 2 1 dT
∇ ψ + V = i~
2m ψ T dt

Again, the left-hand side is completely dependent on x, y, z while the right-hand

side is a function of t. The equality holds only if they are both constant, so we
have the space equation and time equation.


~2 1 2
2m ψ ∇ ψ +V =E


(2.17)
i~ T1 dT


dt =E

Let’s consider the problem in V = 0 and in one dimension, that is, ψ = ψ(x).
The solutions are then given by

T = e−iEt/~
2mE
ψ = A sin kx + B cos kx, k 2 =
~2
Ψn (x, t) = (An sin nx + Bn cos nx)e−iEn t/~
∞
X
Ψ(x, t) = (An sin nx + Bn cos nx)e−iEn t/~
n=1

If we substitute t = 0, this is again a problem about Fourier series. Take

”particle in a box” problem for example.

13
Example 2.2 (Particle in a box). We need to solve Schrödinger equation with
V = 0 on (0, π) and Ψ vanishes at x = 0 and x = π. Substitutions give the
following results.

∞
X
Ψ(x, t) = An (sin nx)e−iEn t/~ (2.18)
n=1

Moreover, given Ψ(x, t) = f (x), we can find An exactly the same as in (2.7).

Recall that in geometry, we discompose a vector into basis vectors and solve
for the coefficients by taking their inner products. Similarly, the essence of
Fourier series is to discompose a function into exponential or trigonometric
functions(which forms basis in Hilbert Space) and the coefficients can be cal-
culated by taking the integral of their products(which is the definition of inner
product for functions). We will explore the convergence of the series and make
these intuitions into rigorous definitions in the rest of the article.

3 Properties of Fourier Series

Fourier analysis is the study of how general functions can be decomposed into
trigonometric or exponential functions with definite frequencies.There are two
types of Fourier expansions:

• Fourier series: If a (reasonably well-behaved) function is periodic, then

it can be written as a discrete sum of trigonometric or exponential func-
tions with specific frequencies.

• Fourier transform: A general function that is not necessarily periodic

(but that is still reasonably well-behaved) can be written as a continuous
integral of trigonometric or exponential functions with a continuum of
possible frequencies.

In the following sections, we will introduce basic Fourier series and Fourier
transform and investigate some detailed questions about Fourier analysis.

The first question is that, what does ”well-behaved” in the above text mean for
a function? In other words, what are the restrictions on the functions that can
be decomposed into Fourier series?

Next, we will investigate the uniqueness of Fourier series. Are two functions
with the same Fourier coefficient necessarily equal?

14
After that, we will take a closer look at the partial sum of Fourier series and
try to solve the question arisen in the last section: in what sense does SN (f )
converge to f as N → ∞?

Finally, to better apply Fourier series to solve mathematics and physics prob-
lems, we will introduce convolution, good kernels and Cesàro and Abel summa-
bility.

3.1 Fourier Trigonometric Series

Definition 3.1. A trigonometric series is a series of the form

∞
A0 X
+ (An cos nx + Bn sin nx)
2 n=1

By Euler’s Theorem, where eiθ = cos θ+i sin θ, we can also write a trigonometric
series in exponential form
∞
2πinx
X
cn e L
n=−∞

If a trigonometric series involves only finitely many non-zero terms, that is,
cn = 0 for all large |n|, it is called a trigonometric polynomial; its degree
is the largest value of |n| for which cn 6= 0.

Definition 3.2. Consider a function f (x) that is periodic on the interval [0, L],
then the Fourier series of F can be expressed as
∞
X 2πnx 2πnx
a0 + (an cos ( ) + bn sin ( )) (3.1)
n=1
L L

where Z L
1
a0 = f (x)dx,
L 0
Z L
2 2πnx
an = f (x) cos ( )dx,
L 0 L
Z L
2 2πnx
bn = f (x) sin ( )dx.
L 0 L

15
3.2 Fourier Exponential Series

By Euler’s theorem, given a function f (x) which is periodic on [0, L], its Fourier
series can also be written as
∞
2πinx
X
cn e L

n=−∞

where the Fourier coefficients cn are given by

Z L
1 −2πinx
cn = f (x)e L dx, n ∈ Z. (3.2)
L 0

We also use fˆ(n) to represent the Fourier coefficients of f .

Remark. Even though we are given the Fourier series of f , we cannot say for
sure that they are necessarily equal, so we will temporarily use ∼ to represent
the relations between them, written as
∞
2πinx
X
f (x) ∼ cn e L .
n=−∞

Definition 3.3. The N th partial sum of the Fourier series of f , for N a

positive integer, is given by

N
2πinx
X
SN (f )(x) = fˆ(n)e L .
−N

3.3 Restrictions on f

3.3.1 Riemann integrability

With the definition of Fourier coefficients, it seems necessary to place some

integrability conditions on f . In the following sections we will assume that all
functions are at least Riemann integrable.

Definition 3.4. A real-valued f defined on [0, L] is Riemann integrable (or

simply integrable ) if it is bounded, and if for every  > 0, there is a subdivision
0 = x0 < · · · < xN −1 < xN = L of the interval [0, L], so that if U and L are,
respectively, the upper and lower sums of f for this subdivision, namely

16
 XN
U =[ sup f (x)](xj − xj−1 )
j=1 xj−1 ≤x≤xj

and

XN
L=[ inf f (x)](xj − xj−1 ),
xj−1 ≤x≤xj
j=1

then we have U − L < .

Remark. From the definition of Riemann integrability, we can see that the func-
tion is not necessarily continuous. It can be everywhere continuous, piecewise
continuous (meaning that it has finitely many discontinuities), or even with
infinitely many discontinuities as long as it is Riemann integrable.

Finally, we say that a complex-valued function is integrable if its real and imagi-
nary parts are integrable. Also, the sum and product of two integrable functions
are integrable.

3.3.2 Functions on the circle

We emphasize that we will not generally restrict our attention to real-valued

functions. We will almost always allow functions that take values in the complex
numbers C. Furthermore, we sometimes think of our functions as being defined
on the circle rather than an interval.

There is a natural connection between 2π-periodic functions on R like the

exponentials einθ , functions on an interval of length 2π, and functions on
the unit circle.

Concisely, 2π-periodic functions on R and functions on an interval of length 2π

that takes on the same value at its end-points, are two equivalent descriptions
of the same mathematical objects, namely, functions on the circle.

We will show this relation in a simple graph:

17
 restrict it to any interval of length 2π
2π periodic functions functions on an 2π-interval
extend it to a periodic function on R

functions on the circle

Figure 2: relations

3.4 Examples of Fourier Series

Example 3.1 (Sawtooth function). Find the Fourier series for the periodic
function. It is defined by f (x) = Ax for −L L
2 < x < 2 and it has period L.

The answer to this question could be quite long but is a great example to show
the detailed process of finding both Fourier trigonometric and exponential series.

Since f (x) is an odd function, f (x) cos 2πnx

L is also odd. So an = 0 for n ∈ Z
and the Fourier series of sawtooth function is a sine series.

By the definition of bn , we get

Z L  
2 2 2πnx
bn = Ax sin dx
L −L
2
L

Integrating by parts gives the general result,

Z Z
1
1
x sin(rx)dx = x − cos(rx) − − cos(rx)dx
r r
x 1
= cos(rx) + 2 sin(rx)
r r

Let r ≡ 2πn/L yields

 L
2A L
2πnx
L
2 2πnx
2
bn = −x cos + sin
L 2πn L 2πn L −L
2

AL
=− cos(πn)
πn
AL
= (−1)n+1 .
πn

18
Therefore, the Fourier trigonometric series of f is
∞
AL X 1 2πnx

f (x) ∼ (−1)n+1 sin
π n=1 n L

We can also obtain the Fourier series using exponentials. By definition of Fourier
coefficients, we get

Z L
1 2
cn = Axe−i2πnx/L dx (3.3)
L −L
2

xe−rx can be found by integration by parts.

R
The integral of the general form
The result is

Z
x 1
xe−rx = − e−rx − 2 e−rx .
r r

This is valid unless r = 0, in which case the integral equals x = x2 /2. For the
R

specific integral in Eq.(3), we have r = i2πn/L, so we obtain (for n 6= 0)

 
A xL −i2πnx/L L L
2 −i2πnx/L L
cn = − e 2
−L + e 2
−L .
L i2πn 2 i2πnx 2

The second of these terms yields zero because te limits produce equal terms.
The first term yields

A (L/2)L −iπn
+ eiπn .


cn = − · e
L i2πn

The sum of the exponentials is simply 2(−1)n . So we have (getting the i out of
the denominator)

iAL
cn = (−1)n forn 6= 0.
2πn

L
If n = 0, then the integral yields c0 = (A/L)(x2 /2)|−2 L = 0. Basically, the area
2
under the curve is zero since Ax is an odd function. Putting everything together
gives the Fourier exponential series,

19
 X iAL i2πnx/L
f (x) ∼ (−1)n e .
2πn
n6=0

The sum runs over all the integers (positive or negative) with the exception of
0.
Example 3.2. Define f (θ) = (π − θ)2 /4 for 0 ≤ θ ≤ 2π. Then successive
integration by parts similar to that performed in the previous example yield

∞
π 2 X cos nθ
f (θ) ∼ + . (3.4)
12 n=1 n2

Actually this solves the Basel problem, which asks for the precise summation
of the reciprocals P
of the squares of the natural numbers, i.e. the precise sum of
∞
the infinite series n=1 .

Let θ = 0. we get

π2
f (0) =
4
∞
π2 X 1
= + ,
12 n=1 n2

which gives us the solution

∞
X 1 π2
2
= .
n=1
n 6

Example 3.3 (Dirichlet kernel).

Definition 3.5. The trigonometric polynomial defined for x ∈ [−π, π] by
N
X
DN (x) = einx
−N

is called the N t h Dirichlet kernel.

Dirichlet kernel is of the fundamental importance of the whole theory. Its Fourier
coefficients an have the property that an = 1 if |n| ≤ N and an = 0 otherwise.
Theorem 3.1. Dirichlet kernel can be expressed as

20
 sin (N + 21 )x


DN (x) = .
sin(x/2)

Proof. This can be seen by summing the geometric progressions

N
X −1
X
ωn and ωn
n=0 n=−N

with ω = eix . These sums are, respectively, equal to

1 − ω N +1 ω −N − 1
and .
1−ω 1−ω

By Euler’s theorem, their sum is then

sin (N + 12 )x


ω −N − ω N +1 ω −N −1/2 − ω N +1/2
= = .
1−ω ω −1/2 − ω 1/2 sin(x/2)

Example 3.4 (Poisson kernel).

Definition 3.6. The function, Pr (θ), called the Poisson kernel, is defined for
θ ∈ [−π, π] and 0 ≤ r < 1 by the absolutely and uniformly convergent series

∞
X
Pr (θ) = r|n| einθ .
n=−∞

Remark.When calculating the Fourier coefficients of Pr (θ) we can interchange

the order of integration and summation since the sum converges uniformly in θ
for each fixed r.
Theorem 3.2.
1 − r2
Pr (θ) = .
1 − 2r cos θ + r2

Proof. Similar to the previous proof, we first split Pr (θ) into two parts,
∞
X ∞
X
Pr (θ) = ωn + ω̄ n with ω = reiθ ,
n=0 n=1

where both series converge absolutely. The first sum (an infinite geometric
progression) equals 1/(1 − ω), and likewise, the second is ω̄/(1 − ω̄).

21
Together, they combine to give
1 − ω̄ + (1 − ω)ω̄ 1 − |ω|2 1 − r2
= = .
(1 − ω)(1 − ω̄) |1 − ω|2 1 − 2r cos θ + r2

3.5 Uniqueness of Fourier series

If we were to assume that the Fourier series of functions f converge to f in an

appropriate sense, then we could infer that a function is uniquely determined
by its Fourier coefficients. This would lead to the following statement:

If f and g have the same Fourier coefficients, then f and g are necessarily equal.
By taking the difference f − g, this proposition can be reformulated as : if
fˆ(n) = 0 for all n ∈ Z, then f = 0.
Theorem 3.3. Suppose that f is an integrable function on the circle with
fˆ(n) = 0 for all n ∈ Z. Then f (θ0 ) = 0 whenever f is continuous at the
point θ0 .

Proof. We first suppose that f is real-valued, and argue by contradiction.

Assume, without loss of generality, that f is defined on [−π, π], that θ0 = 0,
and f (0) > 0.

Our contradiction aims to prove that when assuming fˆ(n) = 0, f (θ0 ) 6= 0 gives
a contradiction. Therefore when fˆ(n) = 0, there should be f (θ0 ) = 0 whenever
f is continuous at the point θ0 .

The idea now is to construct

R a family of trigonometric polynomials {pk } that
”peak” at 0, and so that pk (θ)f (θ) → ∞ as k → ∞. This will be our desired
contradiction since these integrals are equal to zero by assumption.

Here is how we construct it:

• First we put restrictions on f . Choose 0 < δ ≤ π/2, so that f (θ) > f (0)/2
whenever |θ| < δ.
• Then we define p(θ). Let
p(θ) =  + cos θ
where  > 0 is chosen so small that |p(θ)| < 1 − /2 whenever δ ≤ |θ| ≤ π.
• Choose η > 0 with η < δ so that

p(θ) ≥ 1 + for |θ| < η.
2

22
 • Finally, Let pk (θ) = [p(θ)]k .

Now we use a graph to demonstrate the parameters more clearly.

p(θ) ≥ 1 + /2 |p(θ)| < 1 − /2

O η δ π π x
2

f (θ) > f (0)/2

Figure 3: graph showing the relation of η, δ and π

Since f (θ) is Riemann integrable, it is bounded. We select B so that |f (θ)| ≤ B

for all θ.

Since Z π
1
fˆ(n) = f (θ)e−inθ dθ = 0,
2π −π

and since pk (θ) is defined as a real trigonometric polynomial, it can be expressed

in the form
N
X N
X
pk (θ) = cn einθ = c¯n e−inθ .
−N −N

Therefore, we have
Z π Z π N
X
pk (θ)f (θ) dθ = f (θ) c¯n e−inθ dθ
−π −π −N
N
X Z π (3.5)
= c¯n f (θ)e−inθ dθ
−N −π

= 0.

However, we have the estimate

Z  k

f (θ)pk (θ) dθ ≤ 2πB 1 − . (3.6)
δ≤|θ| 2

Also, our choice of δ guarantees that p(θ) and f (θ) are non-negative whenever
|θ| < δ, thus:

23
 Z
f (θ)pk (θ) dθ ≥ 0, (3.7)
η≤|θ|≤δ

and Z  k
f (0) 
f (θ)pk (θ) dθ ≥ 2η 1+ . (3.8)
|θ|<η 2 2

When k → ∞, the integral in (8) goes to infinity, and since the

R second integral is
non-negative and the first one is bounded, we can say that f (θ)pk (θ) dθ → ∞.
This contradicts with (5). So the proof is concluded when f is real-valued.

If interested in the general case, refer to Chapter 2 of [1].

Corollary 3.4. If f is continuous on the circle and fˆ(n) = 0 for all n ∈ Z,

then f = 0.

This corollary can be derived from Theorem 3.3 and it indicates the uniqueness
of Fourier series.

3.6 Convergence of Fourier series

Theorem 3.5. Suppose that f is a continuous function on the circle and that
the Fourier series of f is absolutely convergent. Then, the Fourier series
converges uniformly to f , that is,

lim SN (f )(θ) = f (θ) uniformly in θ.

N →∞

P∞ P∞
Proof. Let g(x) = n=−∞ fˆ(n)einx . Since n=−∞ |fˆ(n)| < ∞, g(x) is defined
everywhere.

By triangle-inequality, we get that

X ∞
X
SN (f )(x) − g(x) = fˆ(n)einx − fˆ(n)einx
|n|≤N n=−∞
X
= fˆ(n)einx
|n|>N
X
≤ fˆ(n) .
|n|>N

24
Since the P
Fourier series of f is bounded, for any  > 0, we can find large N
such that |n|>N |fˆ(n)| < . Hence, SN (f )(x) converges uniformly to g(x) as
N → ∞. Note that for large N , the Fourier coefficients of g(x) are

Z 2π
1
ĝ(n) = g(x)e−inx dx
2π 0
Z 2π
1 X
fˆ(n)einx e−inx dx


=
2π 0 n≤N

= fˆ(n).

In other words, f and g have identical Fourier coefficients. Now we define a

new function h = f − g. By distributivity of the integral, it is easy to see that
ĥ(n) = 0. By the uniqueness of Fourier series, we get that h = 0 or f = g.
Combining this result with the uniform convergence of SN (f )(x) to g(x), we
get the desired result.

We want to generalize the initial condition where f is absolutely convergent. It

turns out that the smoothness of f is directly related to the decay of Fourier
coefficients. To further interpret the smoothness conditions, we will first intro-
duce two definitions:
Definition 3.7. The function f is said to be of differentiability class C k if
0 00
the derivatives f , f , . . . , f (k) exists and are continuous, or we can say that f
is k times continuously differentiable.

Definition 3.8. The Hölder condition of f of order α, with α > 1/2 indicates,

sup |f (θ + t) − f (θ)| ≤ A|t|α for all t.

θ

We say f belonging to class C k or satisfying Hölder condition are two possible

ways to describe the smoothness of a function. In general, the smoother the
function, the faster f decay.

To state the results concisely, we introduce the standard O-notation. Instead

of defining it, it would be more straightforward to present an example:

We say fˆ(n) = O(1/|n|2 ) as |n| → ∞ when the left hand side is bounded by
a constant multiple of the right hand side, i.e. there exists C > 0 such that
|fˆ(n)| ≤ C/|n|2 . Particularly, f (x) = O(1) means that f is bounded.

25
Corollary 3.6. Suppose that f is a twice continuously differentiable function
on the circle. Then

fˆ(n) = O(1/|n|2 ) as |n| → ∞,

so that the Fourier series of f converges absolutely and uniformly to f .

Proof. We first assume that n 6= 0. The twice differentiability indicates that we

can obtain the Fourier coefficient of f by integrating by parts twice. We omit
the process and present the result directly:

−1 2π 00
Z
ˆ
2π f (n) = 2 f (θ)e−inθ dθ
n 0

Therefore,
Z 2π Z 2π
00 00
2π|n|2 |fˆ(n)| ≤ f einθ dθ ≤ |f (θ|dθ ≤ C,
0 0

where C is a constant independent of n.

00
1/n2 converges, we obtain
P
Take C = 2πB where B is a bound for f . Since
the desired result.

The case of n = 0 is pretty obvious and is left for the readers.

From this proof, we also established an important identity:

fˆ0 (n) = infˆ(n), for alln ∈ Z.

We can also investigate the further smoothness conditions on f .

Corollary 3.7. If f is a periodic function of period 2π which belongs to the
class C k , then
fˆ(n) = O(1/|n|k ) as n → ∞.

The proof is quite similar to the one of 3.6.

There are also stronger versions of 3.6. Generally, the Fourier series of f con-
verges absolutely if f satisfies Hölder condition with α > 1/2.

26
3.7 Convolutions

Definition 3.9. Given two 2π-periodic functions f and g on R, we define

convolution f ∗ g on [−π, π]by

Z π

1
f ∗ g (x) = f (y)g(x − y) dy. (3.9)
2π −π

Corollary 3.8. We can also write the convolution of f and g as

Z π
1
(f ∗ g)(x) = f (x − y)g(y) dy.
2π −π

Proof. We can prove this by proving that if f is continuous and 2π-periodic,

then Z π Z π
f (y)dy = f (x − y)dy for any x ∈ R.
−π −π

Let y = −z, we get

Z π Z −π Z π
I= f (y)dy = − f (−z)dz = f (−z)dz
−π π −π

Now let z = y − x and we get

Z π
I= f (x − y)dy.
−π

The convolution plays a role as ”weighted averages”. It is similar to or replaces

pointwise product f (x)g(x). We can even express the partial sum of the Fourier
series of f in terms of convolutions.
N
X
SN (f )(x) = fˆ(n)einx
−N
N  Z π 
X 1 −iny
= f (y)e dy einx
2π −π
−N
Z π N
X 
1 in(x−y)
= f (y) e dy
2π −π −N

= (f ∗ DN )(x),

where DN is the N th Dirichlet kernel.

27
Theorem 3.9. Suppose that f , g, and h are 2π-periodic integrable functions.
Then:

f ∗ (g + h) = (f ∗ g) + (f ∗ h).
(cf ) ∗ g = c(f ∗ g) = f ∗ (cg) for any c ∈ C.
f ∗ g = g ∗ f.
(f ∗ g) ∗ h = f ∗ (g ∗ h).
f ∗ g is continuous.
∗ g(n) = fˆ(n)ĝ(n).
f[

Proof. The first two properties, as the linearity of convolutions, can be proved
by the linearity of the integral.

The third and fourth properties describe commutativity and associativity of

convolutions respectively. We can prove them by interchanging integral signs
and appropriate change of variables.

The fifth and the sixth properties are extremely important. The fifth one shows
that f ∗ g is more ”regular” than f and g, meaning that without restrictions
on the continuity of f and g, only assuming that they are Riemann integrable,
f ∗ g is continuous.

We will first separate the proof into two parts. First we assume that f and g
are continuous, we may write

Z π
1
(f ∗ g)(x1 ) − (f ∗ g)(x2 ) = f (y)[g(x1 − y) − g(x2 − y)] dy.
2π −π

Since g is continuous and periodic, it must be continuous on all of R. The

continuity of g indicates that given  > 0, there exists δ > 0 so that |g(s)−g(t)| <
 whenever |s − t| < δ. Also, |x1 − x2 | < δ implies that |(x1 − y) − (x2 − y)| < δ,
thus
Z π
1
|(f ∗ g)(x1 ) − f ∗ g)(x2 )| ≤ |f (y)||g(x1 − y) − g(x2 − y)| dy
2π −π

≤ |f (y)| dy
2π

≤ 2πB,
2π
where B is the bound of f . Hence we conclude that f ∗ g is continuous when f
and g are continuous.

We should also prove that f ∗ g is continuous if f and g are merely integrable.

This part of proof is omitted here.

28
Now it comes to the sixth property, which is essential for interpreting the con-
volution. It shows relation between convolution and pointwise product. The
proof is quite simple
Z π
1
∗ g(n) =
f[ (f ∗ g)(x)e−inx dx
2π −π
Z π Z π 
1 1
= f (y)g(x − y) dy e−inx dx
2π −π 2π −π
Z π  Z π 
1 1
= f (y)e−iny g(x − y)e−in(x−y) dx dy
2π −π 2π −π
Z π  Z π 
1 1
= f (y)e−iny g(x)e−inx dx dy
2π −π 2π −π
ˆ
= f (n)ĝ(n).

3.8 Good kernels

We say a family of kernels {KN (x)}∞

N =1 is a family of good kernels if and only
if the following are true:

Property (a): The area under the kernels is always fixed, specifically, 2π. We
can rewrite it as
Z π
1
KN (x)dx = 1, for all natural numbers N
2π −π

This property gives us an intuition that good kernels act like weight distribution
function. Moreover, the weighted average of good kernels is always 1, indicating
that good kernels assign unit mass to the whole circle. The “unit” doesn’t mean
1, but a fixed value, in this case 2π.

Property (b): The integral of the absolute value of good kernels is bounded,
Z π
1
|KN (x)|dx ≤ M, for some M ∈ R.
2π −π

A trivial result is that if the good kernel is positive for all x ∈ [−π, π], this prop-
erty is actually a consequence of the previous one. In this case, the infimum of
M is 1.

29
Property (c): As N increases, the tail of good kernels gets slimmer. Mathe-
matically, it can be expressed as followed:
Z
0 0
∀ > 0, ∀δ > 0, ∃N ∈ N, such that N ≥ N =⇒ |KN (x)|dx < .
δ<|x|≤π

The weight is more concentrated near the origin and to be more extreme, the
good kernel will look alike a continuous straight line segment on the x-axis,
except for the discontinuity at the origin, of which the value will be infinity.
Usually, we call this function as Dirac Delta function.
Theorem 3.10. For any function f on the circle, given a family of good kernels
{KN (x)}∞
N =1 , we have

lim (KN ∗ f )(x) = f (x)

N →∞

at the point of continuity x of f . If f is continuous everywhere, then the con-

vergence is uniform.

Proof. Assume that f is continuous at x, then we have

∀ > 0, ∃δ > 0 s.t. |y| < δ =⇒ |f (x − y) − f (x)| < .

Also, f is integrable on the circle, then ∃B ∈ R, such that |f (x)| ≤ B. Hence for
any two points on the good kernels, the difference between their values cannot
be greater than 2B. Also, say the integral of the absolute value of good kernels
is bounded by M .

For the same , we choose a suitable N 0 ∈ N such that

Z
|KN (x)|dx < , for all N ≥ N 0
δ<|x|≤π

Now let’s take a look at the difference between the convolution and the function

30
itself.
Z π
1
(KN ∗ f )(x) − f (x) = f (x − y)KN (y)dy − f (x)
2π −π
Z π
1

= f (x − y) − f (x) KN (y)dy, by the first property of good kernels
2π −π
Z
1

= f (x − y) − f (x) KN (y)dy
2π δ≤|y|≤π
Z
1

+ f (x − y) − f (x) KN (y)dy
2π |y|<δ
Z
1
≤ f (x − y) − f (x) KN (y) dy
2π δ≤|y|≤π
Z
1
+ f (x − y) − f (x) KN (y) dy
2π |y|<δ
Z Z
1 1
≤ 2B KN (y) dy +  KN (y) dy
2π δ≤|y|≤π 2π |y|<δ
2B 
< + M, for all N ≥ N 0 for some N 0
2π
 2π 
2B M
= +  −→ 0
2π 2π

Hence (f ∗ Kn )(x) → f (x). Moreover, if f is continuous everywhere, we can

choose δ independent of x, then it is uniform convergence by the definition.

Here is another way to understand this theorem. Since the weight of good
kernels concentrates on the origin. Therefore, for the following integral
Z π
1
(f ∗ Kn )(x) = f (x − y)Kn (y)dy
2π −π

it is valid over smaller and smaller neighbourhood near y = 0, so f (x − y)

becomes closerR π to f (x) as n increases. Hence the convolution approximately
1
equals to 2π −π
f (x)dy, which equals to f (x).

3.9 The Dirichlet kernel

Proposition 3.11. The Dirichlet kernel is not a good kernel.

If the Dirichlet kernel was a good kernel, then the theorem could apply to it and
as a consequence, the Fourier series, which is the convolution of the Dirichlet
kernel and a given function would converge to itself at points of continuity.
However, this is demonstrated to be false.

31
Another reason can be the Dirichlet kernel fails to satisfy the second property,
i.e. the Dirichlet kernel is not uniformly bounded.
Z π
4
|Dn (x)|dx = 2 lnN + O(1) −→ ∞
−π π

Proposition 3.12. The Dirichlet kernel satisfies the property (a) of good ker-
nels.

PN
Proof. Nevertheless, using the exponential form of DN = n=−N einx , the
integral of itself is 2π, satisfying property (a).
Z π Z π N
1 1 X
DN (x)dx = einx dx
2π −π 2π −π n=−N

N Z π
1 X
= einx dx
2π −π
n=−N
N  π Z π
1 X 1 inx 1
= e + 1dx
2π in −π 2π −π
n=−N,n6=0
1
=0+ [π − (−π)]
2π
=1

3.10 Cesàro summability

Given a sequence {an }∞ n=0 and its N

th
partial sum SN , we define the Cesàro
∞
sum of {an }n=0 , or in other words, the Cesàro mean of SN , σN as followed
S0 + S1 + ... + SN −1
σN := .
N

If σN → σ ∈ R, we say the sequence {an }∞

n=0 is Cesàro summable to σ.

This can be applied to diverging sequences, especially ones that are not divergent
to +∞ or −∞, which gives a more precise average value of sequences.
Example 3.5. One of the most classic examples is Grandi’s series.

ck = (−1)k and sn = 1 − 1 + 1 − 1 + ... + (−1)n

This sum doesn’t converge if considering it merely as a series, but it is Cesàro

summable and the Cesàro sum equals 21 .

32
In terms of the Dirichlet kernel, we define its Cesàro mean of the Dirichlet
kernels, what we call, the Fejér kernel, as followed,

D0 (x) + D1 (x) + ... + DN −1 (x)

FN (x) :=
N
We know that the Dirichlet kernel is not a good kernel from previous section,
but the Fejér kernel is and we shall prove it later.
2
1 sin (N x/2)
Theorem 3.13. FN (x) = N sin2 (x/2) and the Fejér kernel is a good kernel.

Proof.
 
N −1
1 X
FN (x) = Dk (x)
N
k=0
 
N −1
1  X sin(k + 21 )x 
=
N
k=0
sin 21 x
N −1  
1 1 X 1
= sin k + x
N sin 21 x 2
k=0
N −1  
1 X 1 1
= 2 1 2 sin (k + )x sin x
2N sin 2 x 2 2
k=0
N −1
1 X
= cos kx − cos(k + 1)x
2N sin2 21 x k=0
1
= (1 − cos N x)
2N sin2 21 x
 
2 Nx
1 sin 2
=
N sin2 ( x2 )
(3.10)

Start from proving the Fejér kernel satisfies the three properties.

33
Property (a):
Z π Z π N −1
1 1 1 X
FN (x) dx = Dk (x) dx
2π −π 2π
−π N k=0
N −1 Z
1 X π
= Dk (x) dx
2N π
k=0 −π
N −1 Z π !
1 X 1
= Dk (x) dx
N 2π −π
k=0
N −1
1 X
= 1=1
N
k=0
(3.11)

sin2 (N x/2)
Property (b): Since FN (x) = N1 sin2 (x/2) , trivially, the numerator and the de-
nominator are always non-negative, we therefore can deduce that property (b)
is satisfied.

Property (c): As the function sin2 (x) is not only bounded above but also con-
tinuous, given an arbitrary δ > 0, ∀|x| ∈ [δ, π] we have 0 < cδ ≤ sin2 ( x2 ).
Then
1 sin2 N x/2
Z
Z
1 1
2 dx ≤ dx → 0
δ≤|x|≤π N sin (x/2) δ≤|x|≤π N cδ
as N → ∞.

Therefore, the Fejér kernel is a good kernel.

A trivial result from the last theorem is about the convergence of the convolution
of the Fejér kernel and functions. Details will be discussed in the following
theorem.
Theorem 3.14. For any integrable 2π-periodic function f , the Fourier series
of f is Cesàro summable to f at every point of continuity, in other words,
the convolution of Fejér kernel and f is convergent to f . Moreover, if f is
continuous everywhere, the convergence is uniform.

Proof. Since FN (x) is a good kernel, then

(FN ∗ f )(x) → f ⇐⇒ σN (f )(x) → f.
By the definition of Cesàro mean of the Fourier series, it implies
N −1
1 X
Sk (f )(x) → f.
N
k=0

34
Hence, the Fourier series of f is Cesàro summable to f .
Corollary 3.15. If the nth Fourier coefficients of an integrable function f are
0 for all natural numbers n, then f = 0 at all points of continuity.

Proof.
N
1 X ˆ
SN (f )(x) = f (n)einx
2π
n=−N
N
1 X
= 0
2π
n=−N

=0

By Fejér’s theorem,
σN (f )(x) = 0 → f
for all points of continuity of f . If f is continuous, f = 0 everywhere.

Understanding this corollary can be started from another view: Since f is Rie-
mann integrable, it is continuous almost everywhere. S∞ That is, for a fixed
 > 0, the set of discontinuity points is covered by n−1 Un , where Un is an
open interval and the total length of Un is less than .
Corollary 3.16. If f is continuous on the circle,f can be uniformly approxi-
mated by trigonometric polynomials.

PN
Proof. SN (f )(x) = 1
2π n=−N fˆ(n)einx is a trigonometric polynomial, then so
is σN (f )(x).

Weierstrass approximation theorem, which states that every continuous func-

tion on a compact interval can be approximated by a polynomial, is the result
of this corollary, as exponential functions can be approximated uniformly by
polynomials on a compact interval.

3.11 Abel summability

P∞
Given a series n=0 an , we define the Abel mean of the series as
∞
X
A(r) := rn an , r ∈ [0, 1)
n=0

35
If limr→1 A(r) converges to some constant s ∈ R, we say the series is Abel
summable and the limit of the Abel mean is s.

Note that the following implications about series are true:

Convergent =⇒ Cesàro summable =⇒ Abel summable,

and the arrows cannot be reversed.

In functions, we define the Abel means of the function f (θ) ∼

P∞terms of inθ
n=−∞ an e as
∞
X
Ar (f )(θ) := r|n| an einθ .
n=−∞

Notice that we take the absolute value of n is because we want it to fit in the
definition of Abel summability of series which requires the power of r to be non-
negative. Hence, we don’t simply write ck = ak eikθ , but ck = ak eikθ + a−k einθ .
P∞
Theorem 3.17. Let n=0 an be a series and the Abel mean is A(r), converging
for every r ∈ [0, 1). If the series is Abel summable to s, and an = o( n1 ), then
we have
X∞
an = s.
n=0

Proof. First, define a sequence of {rN }∞

N =1 as

1
rN := 1 − .
N
The little o-notation indicates that limn→∞ nan = 0, so

∀ > 0, ∃M ∈ N such that ∀n ≥ M, we have |nan | < .

Our aim is to show that for arbitrary  > 0, there exists N ∈ N and N ≥ M ,
PN P∞ n
such that limr→1− n=0 an − n=0 an r < . This is equivalent to show that

36
 PN P∞
limN →∞ n=0 an − n=0 an (rN )n = 0.

N
X ∞
X N
X ∞
X
an − an (rN )n = an (1 − (rN )n ) − an (rN )n
n=0 n=0 n=0 n=N +1

M
X N
X ∞
X
= an (1 − (rN )n ) + an (1 − (rN )n ) − an (rN )n
n=0 n=M +1 n=N +1

M
X N
X ∞
X
≤ an (1 − (rN )n ) + an (1 − (rN )n ) + an (rN )n
n=0 n=M +1 n=N +1

= S1 + S2 + S3

If N → ∞, trivially we have S1 converges to 0 as there are only finite number

of terms.
For S2 , we have the following:

N
X
S2 = (1 − rN ) an (1 + rN + ... + (rN )n−1 )
n=M +1

N
X
< (1 − rN ) an n
n=M +1

< (1 − rN )N 
1
= (1 − 1 + )N 
N
=

37
 ∞
X
S3 = an (rN )n
n=N +1

∞
(rN )n
X  
= an n
n
n=N +1

∞
X (rN )n
< 
n
n=N +1

∞
 X
< (rN )n
N +1
n=N +1

∞
 X
< (rN )n
N + 1 n=0

 1
= <
N + 1 1 − rN

by the definition of rN = 1 − N1 .
Therefore, we are able to deduce that for arbitrary  < 0, by choosing suitable
PN P∞ n
N , we have n=0 an − n=0 an (rN ) < . Then the proof is finished.

In terms of the Dirichlet kernel, we again, define the Abel mean as

∞
X
Pr (θ) := r|n| einθ ,
n=−∞

which is called the Poisson kernel.

Theorem 3.18. If 0 ≤ r < 1, then

1 − r2
Pr (θ) =
1 − 2r cos(θ) + r2

and the Poisson kernel is a good kernel as r → 1.

N.B. In this case, the family of kernels is indexed by a continuous parameter

0 ≤ r < 1 rather than the discrete n considered previously. In the definition
of good kernels, we simply replace n by r and take the limit in property (c)
appropriately, for example r → 1 in this case.

38
Proof.
−1 
X  X∞  
Pr (θ) = r|n| einθ + r|n| einθ + 1
n=∞ n=1
X∞   ∞ 
X 
= rn e−inθ + rn einθ + 1
n=1 n=1
X∞  n ∞ 
X n
= r e−iθ + r eiθ + 1
n=1 n=1
r e−iθ r eiθ
= + +1
1 − r e−iθ 1 − r eiθ
1 − r2 1
= , since cos θ = (eiθ + e−iθ )
1 − 2r cos θ + r2 2
We then prove it meets the three requirements of good kernels.

Property (a):
Z π Z π ∞
1 1 X
P r(θ)dθ = r|n| einθ dθ
2π −π 2π −π n=−∞
∞ Z π !
1 X |n|
= r einθ dθ
2π n=−∞ −π
∞ Z π
1 X   1
= r|n| 0 + 1 e0 dθ
2π 2π −π
n=∞,n6=0

=1

Property (b): The expression of P r(θ) and the condition that 0 < r < 1 tell us
P r(θ) ≥ 0, which implies property (b) is satisfied.

Property (c): Since we are considering the case when r is close to 1, so we can
assume that 12 < r < 1, then

1 − 2r cos θ + r2 > cδ > 0

for δ ≤ |θ| ≤ π. Hence,
1 − r2
P r(θ) <
cδ
As r → 1, P r(θ) → 0. Therefore,
Z
P r(θ) → 0
δ≤|θ|≤π

under the above premise.

39
Theorem 3.19. The Fourier series of an integrable function on the circle is
Abel summable to f at every point of continuity. If the function f is continuous
on the circle, then the Fourier series is Abel summable uniformly to f .

P∞
Proof. The Abel mean of SN (f )(x) is Ar(f )(x) = n=−∞ an r|n| einθ . Since
P r(x) is a good kernel, then

(P r ∗ f )(x) → f ⇐⇒ Ar(f )(x)

It is equivalent to say that SN (f )(x) is Abel summable to f . The following

statement is trivial if we look back to the proof of Fejér’s theorem.

3.12 Gibbs phenomenon

We have already talked a lot about convergence of the Fourier series at the point
of continuity, but what happens at the point of discontinuity? We will briefly
introduce this phenomenon.

Suppose we have a discontinuous function, e.g. top-hat function. If we use

truncated Fourier series Sn (f )(x) to approximate at the discontinuity point, we
have the so-called Gibbs phenomenon.

There are some “damped oscillations” of the truncated Fourier series near the
discontinuity points. As n increases, the oscillating intervals are getting nar-
rower, but the overshoot remains the same size about 9%. Equivalently, from
the other side of the jump, the truncated Fourier series undershoot about 9%.
Hence in total, the partial Fourier series is 18% larger than the discontinuity of
the function. Moreover, as n increases, the truncated Fourier series at the jump
converges to the midpoint of the jump.

Informally, the Fourier series has the big “jump” at the discontinuity because
most of the amplitudes of sine and cosine overlap at that point in a perfect
way. But since the truncated Fourier series has only finite terms of sine and
cosine, the amplitudes cannot cancel out perfectly near the jump, leading to a
damped oscillation. As more terms are added, there will be more cancellations,
hence the truncated Fourier series can approximate the function better, which
explains why the damping intervals become narrower.

3.13 Dirichlet’s problem in the unit disc

We now will check if the solution to the steady-heat equation we obtained before
is true and unique.

40
Theorem 3.20. Let f be an integrable function on the unit circle (the heat
distribution function on the unit circle). Then the heat distribution function u
defined in the unit disc
u(r, θ) = (f ∗ Pr )(θ)
has the following properties:

• u has two continuous derivatives in the unit disc and satisfies ∆u = 0.

• If θ is any point of continuity of f , we have
lim u(r, θ) = f (θ)
r→1

If f is continuous everywhere, then the limit is uniform.

• If f is continuous, u(r, θ) is the unique solution to the steady-state heat
equation in the disc which satisfies the first two properties.

Proof. Property 1: We should first notice that in the unit disc problem, we
defined r ∈ [0, 1), with 1 excluded. However, when we prove the sequence of
derivatives of functions converges to the derivatives of limit of function sequence,
we need the function variables to be defined in a compact set. P∞As a result, we
should introduce ρ s.t. 0 < ρ < 1. Then ∀r ∈ [0, ρ], u(r, θ) = n=−∞ an r|n| einθ
can be differentiated once, twice and so on term by term.

Since x = r cos θ and y = r sin θ, we have

∂2u ∂2u
∆u := + 2
∂x2 ∂y
2
∂ u 1 ∂u 1 ∂2u
= + +
∂r2 r ∂r r2 ∂θ2
∞ ∞ ∞
X X 1 X 1
= |n|(|n| − 1)an r|n|−2 eineθ + |n|an r|n|−1 eineθ + 2
(−n2 )an r|n| eineθ
n=−∞ n=−∞
r n=−∞
r
∞
X  
= an einθ |n|(|n| − 1)r|n|−2 + |n|r|n|−2 + n2 r|n|−2
n=−∞

=0

Property 2: In section 2, we name f to be the heat distribution on the unit cir-

cle, so we have u(1, θ) = f (θ). Rigorously, by the fact that u(r, θ) = (f ∗ Pr )(θ)
is Abel summable to f and Pr (θ) is a good kernel, we obtain limr→1 u(r, θ) = f.

Property 3: Suppose we have two different solutions, called u1 and u2 . Then we

take the difference u1 − u2 . By the first property, we can deduce that
∆(u1 − u2 ) = ∆u1 − ∆u2 = 0 − 0 = 0.

41
Both of the solutions equal to f (θ) at the disc boundary, where u1 − u2 = 0. So
now you probably have a picture in your mind: the boundary has value 0 and
the derivative over the disc is also 0. By intuition, you may guess the whole
disc has zero-value of u. We finish the proof if we use this lemma:

For any point p in the unit disc, any subdiscs at p contained in the unit disc, we
call them Dp and the average value of u in Dp is u(p). The proof of this lemma
is out of the scope of this section so we won’t look much into it.

Then the final result we get is that the whole disc will have u = 0, which is
a contradiction. Therefore, we cannot have two solutions unless they are equal.

4 Convergence of Fourier Series

In this section, we will focus on the convergence of the Fourier series. The
convergence of the Fourier series are divided into two parts: The mean-square
convergence on the entire interval and the pointwise convergence in separate
intervals.

4.1 Mean-square convergence

In this section, we will discover the overall behavior of a function f on the entire
interval [0, 2π]. We will first lay out the definitions and notations that will be
used throughout this section.

4.1.1 Vector space and inner product

Definition 4.1. A vector space consists of a set V (elements of V are called

vectors), a field F (elements of F are called scalars), and two operations:

• An operation called vector addition that takes two vectors v, w ∈ V ,

and produces a third vector, written v + w ∈ V
• An operation called scalar multiplication that takes a scalar c F and a
vector v ∈ V , and produces a new vector, written cv ∈ V .

And satisfy following properties (called Axioms):

42
 • Associative law: For all vectors u, v, w ∈ V, u + (v + w) = (u + v) + w
• Additive identity: The set V contains an additive identity element, de-
noted by 0, such that for any vector v ∈ V , 0 + v = v and v + 0 = v.
• Existence of negatives: For every u ∈ V , there is a vector in V , written
−u and called the negative of u, which has the property that u+(−u) = 0.
• Associativity of multiplication: (ab)u = a(bu) for any a, b ∈ F and u ∈ V .
• Distributivity: (a + b)u = au + bu and a(u + v) = au + av for all a, b ∈ F
and u, v ∈ V
• Unitarity: 1u = u for all u ∈ V
Definition 4.2. An inner product space is a vector space V over the field
F together with an inner product, i.e., with a map (. , .) : V × V → F that
satisfies the following three properties for all vectors x, y, z ∈ V and all scalars
αβ ∈ F :

• Conjugate Symmetry: (x, y) = (y, x)

• Linearity: (αx + βy, z) = α(x, z) + β(y, z)
• Positive Definite: (x, x) ≥ 0

Given an inner product (·, ·) we may define the norm of X by kXk = (X, X)1/2

The presence of an inner product on a vector space allows one to define the
geometric notion of “orthogonality.” Let V be a vector space (over R or C)
with inner product (·, ·) and associated norm k.k Two elements X and Y are
orthogonal if (X, Y ) = 0, and we write X ⊥ Y . Three theorems can be derived
from this notion of orthogonality:
Theorem 4.1. The Pythagorean theorem: if X and Y are orthogonal, then
2 2 2
kX + Y k = kXk +kY k (4.1)

Proof.
2
kX + Y k = (X + Y, X + Y ) = (X, X) + (X, Y ) + (Y, X) + (Y, Y ) (4.2)
2 2
= (X, X) + (Y, Y ) = kXk +kY k (4.3)

Theorem 4.2. The Cauchy-Schwarz inequality: for any X, Y ∈ V , we have

(X, Y ) ≤ kXkkY k (4.4)

43
Proof. If kY k =0, we want to show that for all X, (X,Y )=0. For all real t we
have:
2 2
0 ≤ kX + tY k = kXk + 2tRe(X, Y ) (4.5)
So if Re(X, Y ) 6= 0 contradicts the inequality if we pick some big enough t.
Then we have Re(X, Y ) = 0. Similarly, we have Im(X, Y ) = 0.We finished the
case for kY k =0.

If kY k =
6 0,we may set c = (X, Y )/(Y, Y ); then X − cY is orthogonal to Y , and
therefore also to cY . If we write X = X − cY + cY and apply the Pythagorean
theorem, we get
2 2 2 2
kXk = kX − cY k +kcY k ≥ |c|2 kY k (4.6)

Then we get the result.

Theorem 4.3. The triangle inequality: for any X, Y ∈ V , we have

kX + Y k ≤ kXk +kY k (4.7)

Proof. According to the proof of i, we have

2
kX + Y k = (X, X) + (X, Y ) + (Y, X) + (Y, Y ) = kXk + (X, Y ) + (Y, X) +kY k
(4.8)
And by the Cauchy-Schwartz inequality,

|(X, Y ) + (Y, X)| ≤ 2kXkkY k (4.9)

So we have
2 2 2
kX + Y k ≤ kXk + 2kXkkY k +kY k = (kXk +kY k)2 (4.10)

44
Theorem 4.4. The vector space l2 (Z) over C is the set of all (two-sided) infinite
sequences of complex numbers

(..., a−n , ..., a−1 , a0 , a1 ..., an , ...) (4.11)

such that X
|an |2 < ∞ (4.12)
n∈Z

Proof. We will first prove that l2 (Z) is a vector space. That is, if A and B are
two elements in l2 (Z), then A + B is an element in l2 (Z). We assume

A = (..., a−1 , a0 , a1 , ...) (4.13)

B = (..., b−1 , b0 , b1 , ...) (4.14)

Then for each integer N > 0, we define the truncated element

AN = (..., 0, 0, a−n , ..., a−1 , a0 , a1 , ...an , 0, 0, ...) (4.15)

Which we have an = 0 when |n| > N Then according to the triangle inequality,
we have
kAN + BN k ≤ kAN k +kBN k ≤ kAk +kBk (4.16)
According to the definition of the norm, we have
X 2
|an + bn |2 = kAN + bN k ≤ (kAk +kBk)2 (4.17)
|n|≤N

When N → ∞, X
|an + bn |2 < ∞ (4.18)
n∈Z

So we finished the proof

Definition 4.3. An inner product space with these two properties is called a
Hilbert space:

• The inner product is strictly positive-definite, that is, kXk = 0 implies

X = 0.

• The vector space is complete, which by definition means that every

Cauchy sequence in the norm converges to a limit in the vector space.

We see that Rd and Cd are examples of finite-dimensional Hilbert spaces, while

l2 (Z) is an example of an infinite-dimensional Hilbert space. If either of the
conditions above fails, it is called a pre-Hilbert space.

45
Lemma 4.5. : The vector space l2 (Z) is complete

Proof. : Suppose Ak = {ak , n}n∈Z with k = 1, 2, ... is a Cauchy sequence. For

each n, Ak = {ak , n}k=1 is a Cauchy sequence of complex numbers since

|ak,n − ak0 ,n | ≤ Ak − A0k l2

(4.19)

Therefore it converges to a limit bn . Given  > 0, there is a N ∈ N such that

for all k, k 0 ≥ N ,
M
X 2 2
|ak,n − ak0 ,n |2 ≤ Ak − A0k l2
< (4.20)
n=1
2

for all M. PM
For each k > N and M , we let k 0 → ∞ to get n=1 |ak,n − bn |2 < 2 . Since
M is arbitrary, one has shown that kAk − Bkl2 <  as k ≥ N , where B =
(..., b−1 , b0 , b1 , ...). Also, b ∈ l2 (Z). Since  is arbitrary, Ak → B in l2 (Z) as
k→∞

We now give an important example of a pre-Hilbert space where both conditions

(i) and (ii) fail.

Proposition 4.6. Let R denote the set of complex-valued Riemann integrable

functions on [0, 2π] (or equivalently, integrable functions on the circle). This is
a vector space over C. Addition is defined pointwise by

(f + g)(θ) = f (θ) + g(θ) (4.21)

Naturally, multiplication by a scalar λ ∈ C is given by

(λf )(θ) = λ · (θ) (4.22)

An inner product is defined on this vector space by

Z 2π
1 ¯
(f, g) = f (θ)g(θ)dθ (4.23)
2π 0

This set of complex-valued Riemann integrable functions is a pre-Hilbert space

46
Proof. It is easy to find that the analogue of the Cauchy-Schwarz and triangle
inequalities hold in this example.

We now check for the conditions of Hilbert space.

For condition (i), it fails since kf k = 0 implies only that f vanishes at its points
of continuity.

For condition (ii), we can use the following corollary to check

Corollary 4.7. : Let
(
0, for θ = 0
f (θ) = (4.24)
log(1/θ), for 0 < θ ≤ π

and define a sequence of functions in R by

(
0, for θ = 0
fn (θ) = (4.25)
f (θ), for 1/n < θ ≤ π

Prove that {fn }∞

n=1 is a Cauchy sequence in R. However, f does not belong to
R.

Proof. By L’Hospital’s Rule, We have limθ→0 θ(log(θ))2 → 0 and limθ→0 θ(log(θ)) →

0 if 0 < a < b and b → 0 And we have
Z
(log(θ))2 dθ = θ(log(θ))2 − 2θ(log(θ)) + 2θ + C (4.26)

which C is a constant
Rb
Therefore, we have a ((log(θ))2 dθ) → 0 Thus, for all  > 0, there exist a N > 0,
for n > m > N ,
Z 1/m Z 1/m
1 1
fn (θ) − fm (θ) = ( (log(1/θ))2 dθ)1/2 = ( (logθ)2 dθ)1/2 < 
2π 1/n 2π 1/n
(4.27)
showing that {fn }∞n=1 is a Cauchy sequence in R and lim n→∞ nf = f . However,f ∈
/
R, since it is not bounded.

So this set of complex-valued Riemann integrable functions is a pre-Hilbert

space.

4.1.2 Proof of Mean-Squared Convergence

Theorem 4.8. (Mean-Squared Convergence)

47
Consider the space R of integrable functions on the circle with inner product
Z 2π
1
(f, g) = f (θ)g(θ)dθ (4.28)
2π 0

And norm kf k defined by

Z 2π
2 1
kf k = (f, f ) = |f (θ)|2 dθ (4.29)
2π 0

With this notation, prove that f − SN (f ) → 0 as N tends to infinity.

Notation-wise, we use an to denote the Fourier coeffcients of f , where f is an

integrable function on the circle. And for each integer n, we have en = einθ
Lemma 4.9. : The family {en }n∈Z is orthonormal

Proof. When n = m,
Z 2π
1
(en , em ) = einθ eimθ dθ (4.30)
2π 0
Z 2π
1
= ei(n−m)θ dθ (4.31)
2π 0
Z 2π
1
= 1dθ (4.32)
2π 0
=1 (4.33)

When n 6= m,
Z 2π
1
(en , em ) = einθ eimθ dθ (4.34)
2π 0
Z 2π
1
= ei(n−m)θ dθ (4.35)
2π 0
i(n−m)θ
1 e
= |2π (4.36)
2π i(n − m) 0
1 ei(n−m)2π − e0
= (4.37)
2π i(n − m)
1 1+0−1
= (4.38)
2π i(n − m)
=0 (4.39)

Therefore, the family {en }n∈Z is orthonormal

48
Besides for the family {en }n∈Z , we can find the Fourier coefficient can be rep-
resented by an inner product:

Z 2π
1
an = f (θ)einθ dθ = (f, en ) (4.40)
2π 0

Also, we should note that

X
SN (f ) = an en (4.41)
|n|≤N

Lemma 4.10. : For any complex number bn , we have

X X
(f − an en ) ⊥ bn en (4.42)
|n|≤N |n|≤N

P
Proof. We will first show (f − |n|≤N an en ) ⊥ en
X X
(f − an en , en ) = (f, en ) − ( an en , en ) (4.43)
|n|≤N |n|≤N

= an − an (4.44)
=0 (4.45)

P P
So for any complex number bn , we have (f − |n|≤N an en ) ⊥ |n|≤N bn en

P P
Therefore if we choose an as bn , we have (f − |n|≤N an en ) ⊥ |n|≤N an en .
Then by the Pythagoras theorem, we have
2 2
2
X X
kf k = f − an en + an en (4.46)
|n|≤N |n|≤N

which can be rewritten as

2 2 X
kf k = (f − SN (f )) + |an |2 (4.47)
|n|≤N

Next, we will introduce the Best Approximation Lemma.

Lemma 4.11. (Best Approximation Lemma): If f is integrable on the circle
with Fourier coefficients an , then for any complex numbers cn ,
X
f − SN (f ) ≤ f − cn e n (4.48)
|n|≤N

Moreover, equality holds precisely when cn = an for all |n| ≤ N .

49
Proof. Let bn =an -cn , so we have
X X X
f− cn en = f − an en + bn e n (4.49)
|n|≤N |n|≤N |n|≤N

By previous lemma, we know that

X X
(f − an en ) ⊥ (f − bn e n ) (4.50)
|n|≤N |n|≤N

Hence, by Pythagoras theorem, we have,

2 2
X 2 X
f − (f − cn en ) = f − SN (f ) + f− bn e n (4.51)
|n|≤N |n|≤N

When cn = an , bn = 0,

X
f − (f − cn en ) = f − SN (f ) (4.52)
|n|≤N

When cn 6= an ,
2
Z 2π
X 1
bn en = |bn en |2 dθ > 0 (4.53)
2π 0
|n|≤N

Therefore, X X
(f − an en ) ⊥ (f − bn e n ) (4.54)
|n|≤N |n|≤N

Then we come to the final step of the proof of Mean-Squared Convergence.

Theorem 4.12. (Mean Square Convergence): Suppose f is integrable on the
circle. Then
Z 2π
1
|f (θ) − SN (f )(θ)|2 dθ → 0 as N → ∞ (4.55)
2π 0

Proof. Apply Lemma from section 3 and choose a continuous function g on the
circle such that
sup |g(θ)| ≤ sup |f (θ)| = B (4.56)
θ∈[0,2π] θ∈[0,2π]

50
and

Z 2π
|f (θ) − g(θ)|dθ < 2 (4.57)
0

Since supx∈[−π,π] |f (x)| = B

Z 2π
2 1
kf − gk = |f (θ) − g(θ)|2 dθ (4.58)
2π 0
Z 2π
1
= |f (θ) − g(θ)||f (θ) − g(θ)|dθ (4.59)
2π 0
Z 2π
B
≤ |f (θ) − g(θ)|dθ (4.60)
π 0
≤ C2 (4.61)

Where C is a constant

Moreover, we can approximate g by a trigonometric polynomial P of degree M

such that for all θ
|g(θ) − P (θ)| <  (4.62)

After taking square and integrating, we have

kg − P k <  (4.63)

Then we have kf − P k < C 0 

By the best approximation lemma we finished our proof.

From the result

2 2 X
kf k = (f − SN (f )) + |an |2 (4.64)
|n|≤N

We have the Parserval’s Identity:

∞
X 2
|an |2 = kf k (4.65)
n=−∞

51
4.2 Pointwise Convergence

Theorem 4.13. Let f be an integrable function on the circle which is differen-

tiable at a point θ0 . Then SN (f )(θ0 ) → f (θ0 ) as N tends to infinity.

f (θ0 −t)−f (θ0 )

Proof. Define F (t) = t if t 6= 0 and |t| < π, F (t) = −f 0 (θ0 ) if t = 0

For a small δ the function F is integrable on [−π, −δ] ∪ [δ, π], therefore the
function F is integrable on all of [−π, π].

R SN (f )(θ0 ) = (f ∗ DN )(θ0 ), where DN is a Dirchlet

In section 3, we learned that
1
kernel. Then we have 2π DN = 1. Therefore,
Z π
1
SN (f )(θ0 ) − f (θ0 ) = f (θ0 − t)DN (t)dt − f (θ0 ) (4.66)
2π −π
Z π
1
= (f (θ0 − t) − f (θ0 ))DN (t)dt (4.67)
2π −π
Z π
1
= F (t)tDN (t)dt (4.68)
2π −π

Also,
t
tDN (t) = sin((N + 1/2)t) (4.69)
sin(t/2)

t
And the quotient sin(t/2) is continuous in the interval [−π, π], so

sin((N + 1/2)t) = sin(N t) cos(t/2) + cos(N t) sin(t/2) (4.70)

Then we apply the Riemann-Lebesgue Lemma to the functions F (t)t cos(t/2)/ sin(t/2)
and F (t)t:

Z π
1
F (t)t cos(t/2)/ sin(t/2)) sin(N t)dt → 0 (4.71)
2π −π

and

Z π
1
F (t)t cos(N t)dt → 0 (4.72)
2π −π

52
Therefore,
Z π
1
SN (f )(θ0 ) − f (θ0 ) = F (t)tDN (t)dt (4.73)
2π −π
Z π
1
= F (t)(t cos(N t) + t cos(t/2)/ sin(t/2) sin(N t))dt
2π −π
(4.74)
Z π
1
= F (t)t cos(t/2)/ sin(t/2)) sin(N t)dt (4.75)
2π −π
Z π
1
+ F (t)t cos(N t)dt (4.76)
2π −π
→0 (4.77)

Then we finished the proof of this theorem.

5 Application of Fourier Series

5.1 The Isoperimetric problem

5.1.1 The isoperimetric inequality

In this section, we will investigate among all simple closed curves of length `,
in the plane R2 , which one encloses the largest area. The intuition tells us the
answer should be a circle, because if the region enclosed by the curve is concave,
then we can reflect the concave part to be convex and we can convince ourselves
that the flatter the curve is in some portion, the less efficient it is in an enclosing
area. We are going to prove it is the case rigorously later.

5.1.2 Curves

In this section, we only focus on curves that are closed and simple. This means
the curves don’t intersect themselves and the endpoints are joined. For example,
a square and a circle are both closed and simple.

For any point on the simple closed curve Γ, we define its orientation by a
parametrized expression, γ(s) : [a, b] → R2 .

γ(s) = (x(s), y(s))

53
x(s) and y(s) denote the x and y axis positions respectively. For simplicity, we
limit γ to be in class C 1 , that is, γ is continuously differentiable.

Remark: If we substitute s by another variable t, we can rewrite the function

η(t) : [c, d] → R2 as
η(t) = γ(s(t)).
In real life, the variable s is the distance travelled by the walker and we can
imagine t as the time spent in travelling distance s. Undoubtedly, the orientation
of the walker can be expressed by a function of t and s0 (t) represents the speed.
If s0 (t) > 0, then η and γ induce the same direction on the curve Γ.

5.1.3 Length of the curve

In Cartesian plane, given a function f : [a, b] → R, y = f (x), we know the

length is
Z b  21
`= (dx)2 + (dy)2 .
a
Here x and y are functions of s, then by the chain rule, the length in parametrized
form is
Z b Z b  2  2 ! 21
0 dx dy
`= |γ (s)|ds = + ds.
a a ds ds

Specifically, if |γ 0 (s)| = 1, i.e. the walker travels in a constant speed, we say γ is

a parametrization by arc-length. In this case, the length of the curve is simply
b − a.

5.1.4 Green’s theorem

Theorem 5.1 (Green’s theorem). Given a closed simple curve C in the vector
field f~(x, y) = P (x, y)~i+Q(x,y)~j and the region R bounded by C, where the
position vector along the curve C is defined as ~r(t) = x(t)~i + y(t)~j, then closed
curve line integral of C in the vector field is
I ZZ
~ ∂Q ∂P
f · d~r = − dxdy
C ∂x ∂y
R

Proof. Suppose we have a vector field with ~i-component only, that is f~ =

P (x, y)~i. Though, in the position vector, x and y are functions of t, we still
can write y as a function of x and break the curve into C1 and C2 , representing

54
g(x) and f (x). WLOG, say C1 and C2 are the upper and the lower half of the
curve, and the direction along C is anticlockwise. Then,
I Z b Z a
f~ · d~r = f~(x, g(x))d~r + f~(x, f (x))d~r
C a b
Z b Z b
= f~(x, g(x))d~r − f~(x, f (x))d~r
a a
Z b h i
=−

~
d~r
f~ x, f (x) x, g(x)
a
Z b  
f (x)
=− f~(x, y) g(x) · d~r
a
Z b  
f (x)
=− P (x, y) g(x) dx
a
!
Z b Z f (x)
∂P
=− dy dx
a g(x) ∂y
ZZ
∂P
=− dydx
∂y
R

Now suppose the vector field only has ~j-component f~ = Q(x, y)~j, and repeat.

Finally, for an arbitrary vector field f~, we can write it as

f~(x, y) = P (x, y)~i + Q(x, y)~j,
then I ZZ ZZ
∂P ∂Q
f~ · d~r = − dydx + dydx.
C ∂y ∂x
R R

5.1.5 Area bounded by the curve

RR RR
We know the area of the region R is A = 1dA = 1 dydx. In order to use
R R
Green’s theorem, we need
∂Q ∂P
− = 1.
∂x ∂y
Easily, we can choose Q := x2 and P := − y2 , then
y x
f~(x, y) = − ~i + ~j,
2 2
and the closed line integral in this vector field is
I I     1I
~ y~ x ~ ~ ~
f · d~r = − i + j · dxi + dy j = xdy − ydx
C C 2 2 2 C

55
To avoid negative values of the area, we put absolute value sign over the integral.
I
1
A= xdy − ydx
2 C

5.1.6 Isoperimetric inequality

Theorem 5.2. Suppose that Γ is a simple closed curve in R2 of length `, and

let A denote the area of the region enclosed by this curve. Then

`2
A≤
4π
and the equality holds if and only if Γ is a circle.

Proof. Let ` = 2π and the aim is now to show the curve enclosed the largest
area is the unit circle. General cases can be done by rescaling.

Again, we write the parametrized expression for the curve Γ as

γ(s) = (x(s), y(s)),

and we assume that γ is a parametrization by arc-length.

Since the area bounded by the curve Γ is given as

I
1
A= xdy − ydx ,
2 Γ
which is equivalent as
Z 2π Z 2π Z 2π
1 1 1
A= (x(s)y 0 (s) − y(s)x0 (s))ds = x(s)y 0 (s)ds − y(s)x0 (s)ds (∗).
2 0 2 0 2 0

We notice that γ is 2π-periodic, so are x(s) and y(s). Then we write

∞
X ∞
X
x(s) ∼ an eins and y(s) ∼ bn eins .
n=−∞ n=−∞

Since we have shown that the Fourier series can be differentiated term by term,
hence their derivatives can be written as
∞
X ∞
X
x0 (s) ∼ inan eins and y 0 (s) ∼ inbn eins ,
n=−∞ n=−∞

and
x0 (s) = x0 (s) and y 0 (s) = y 0 (s)

56
as x(s) and y(s) are real functions from which we deduce that an eins +a−n e−ins
is real, and hence an = a−n
∞
X ∞
X
x0 (s) ∼ (inan )eins and y 0 (s) ∼ (inbn )eins
n=−∞ n=−∞

Recall the Parseval’s theorem, then A becomes

∞
X
A=π n(an bn − bn an ) .
n=−∞

Since γ is a parameterization by arc-length, i.e. |γ 0 (s)| = 1, this is equivalent to

say that |x0 (s) + y 0 (s)| = 1. By Parseval’s identity, we have
∞
X
|n|2 (|an |2 + |bn |2 ) = 1.
n=−∞

Apparently, |n| < |n|2 if n 6= −1, 0 and 1, hence it is without difficulty to deduce
that
∞
X
A≤π n2 (an bn − bn an ) . (∗)
n=−∞

By triangle inequality, we have

∞
X ∞
X ∞
X
n2 (an bn − bn an ) ≤ n2 (an an + bn bn ) = n2 (|an |2 + |bn |2 ) = 1, (∗∗)
n=−∞ n=−∞ n=−∞

P∞
recalling Parseval’s identity, which implies that n=−∞ n2 (|an |2 + |bn |2 ) =
|x0 (s)2 + y 0 (s)2 | as n2 |an |2 is the nth-Fourier coefficient of x0 (s) and n2 |bn |2 is
that of y 0 (s).

To show the curve enclosed the greatest area is the unit circle starts from as-
suming the area A is π. In this case, (*) and (**)must be equality, so there
must be finite terms. Then only n = −1, 0 and 1 are valid, so write

x(s) = a−1 e−is + a0 + a1 eis and y(s) = b−1 e−is + a0 + a1 eis ,

and
1
X 1
X
n(an bn − bn an ) = |n|2 (|an |2 + |bn |2 ) = 1,
n=−1 n=−1

while the second equality holds by the Parseval’s identity.

57
So (|a−1 |2 + |b−1 |2 ) + (|a1 |2 + |b1 |2 ) = 1, and we deduce that |a1 | = |a−1 | = 12
and |b1 | = |b−1 | = 12 . Since we don’t know if they are real values, then we can
write them as
1 1
a1 = eiα and b1 = eiβ .
2 2
P1
The relationship between α and β is connected by | sin (α − β)| = 1 as n=−1 n(an bn − bn an ) =
| − a−1 b1 + b1 a1 + a1 b1 − b1 a1 | = 2|a1 b1 − a1 b1 | = 1, then eα−β − e−α+β = 1,
hence |sin(α − β)| = 1. So α − β = kπ + π2 and k is any integer.

Finally, we can write

x(s) = a0 + cos(α + s) and y(s) = b0 ± sin(α + s).
This is exactly the parametric form of a circle. Therefore, we have shown that
Γ is a circle.

5.2 Weyl’s equidistribution theorem

We now look at the application of Fourier series into number theory, starting
with the following definitions.
Definition 5.1. For x ∈ R, its integer part [x] is defined as the greatest
integer less than x. Its frictional part hxi is then defined as x − [x]. For
a, b ∈ R, they are equivalent module over Z if and only if their frictional
parts are equal, denoted by
a ≡ b mod Z

This definition is inspired by the congruence of integers. Recall that when

dealing with integer, division with remainders says for arbitrary a and b in Z,
there uniquely exists r such that
a≡r mod b with 0 ≤ r < b
Similarly, it is not hard to prove for arbitrary x ∈ R, there uniquely exists
r = hxi ∈ [0, 1) such that
a ≡ r mod Z

To introduce Weyl’s equidistribution theorem, let’s first look at the following

example.
Example 5.1. Consider the set {cos k|k ∈ Z \ {0}} bounded above by 1, there-
fore having a supremum. Without the constraints on n, the supremum is ob-
viously 1. By intuition, we expect the supremum is still 1 even with the con-
straints, but how can we rationalize and prove it? By observation, the closer k

58
is to 2nπ for any n, the closer cos k is to 1. So the problem is now if 2nπ can
be as close as to an integer, that is, having a frictional part that is as small as
possible. Equivalently, we want to prove that

∀ > 0, ∃k ∈ Z, n ∈ N such that |ξn − k| < , where ξn is 2nπ

The integer part of a real might not be the closest integer to it, but in fact, the
above statement can be changed to be stronger

∀ > 0, ∃n ∈ N such that hξn i < , where ξn is 2nπ

The proof of this theorem requires Weyl’s equidistribution theorem, so we will

come back to it later. This example leads us to consider the distribution of hnγi
in [0, 1), which is said to be equidistributed.

Definition 5.2. A sequence of reals ξ1 , ξ2 , ξ3 , ... is equidistributed in [0, 1) if

∀(a, b) ⊂ [0, 1),

| {ξ1 , ξ2 , ..., ξN } ∩ (a, b)|

lim =b−a
N →∞ N

Remark. The left hand side refers to the proportion of the terms in the interval
(a, b) among ξ1 , ξ2 , ξ3 , ..., ξN . The right hand side is the proportion of the
length of the interval (a, b) in [0, 1). Informally, ξn is the microcosm of [0, 1)
such that its cover rate of the interval (a, b) is the same as that of [0, 1).

We can now test our definition on hnγi for different γ.

γ = 1/2 : 0.5, 0, 0.5, 0, 0.5, ...
γ = 1/3 : 0.3333, 0.6666, 0, 0.3333, ...
γ=√ 1/7 : 0.142857, 0.285714, 0.428571, 0.571428, 0.714285, 0.857142, 0, 0.142857, ...
γ = 2 : 0.4142, 0.8284, 0.2426, 0.6568, 0.0710, 0.4852, ...
γ = π : 0.1415, 0.2831, 0.4247, 0.5663, 0.7079, ...
γ = e : 0.7182, 0.4365, 0.1548, 0.8731, 0.5914, ...

It is not hard to discover that if γ is rational, the sequence repeats itself and
obviously cannot be equidistributed, but what about γ is irrational? This is
what Weyl’s equidistribution theorem discusses.
Theorem 5.3. hnγi is equidistributed in [0, 1) if γ is irrational.

Proof. Before we start the proof, we need a few lemmas. The definition of
equidistribution involves the countability of a set, which is hard to use in our

59
proof. Our first job is to translate it into analytical language. For fixed (a, b) ⊂
[0, 1), we define its characteristic function χ(a,b) : [0, 1) → R as

1, x ∈ (a, b)


χ(a,b) (x) =

0, x ∈ [0, 1) − (a, b)


This function works as indicative variable that tells whether the input is in
(a, b). We extend it to be periodic 1. Now we can translate the countability
into
N
X
| {ξ1 , ξ2 , ..., ξN } ∩ (a, b)| = χ(a,b) (nγ)
n=1
Moreover, the RHS is translated into
Z 1
χ(a,b) (x) dx
0

The theorem is now translated into

N Z 1
1 X
χ(a,b) (nγ) → χ(a,b) (x) dx as N → ∞
N n=1 0

We will first prove a weaker version of this statement.

Lemma 5.4. For continuous f of period 1 and irrational γ,
N Z 1
1 X
f (nγ) → f (x) dx as N → ∞ (5.1)
N n=1 0

Proof. Now we are free to apply Fourier series in this question. We will first
prove that the statement holds for all exponentials, therefore all sums of expo-
nentials that are Fourier series of the continuous function. Initially, we compute
the LHS when f (x) = e2πikx for k ∈ N ∪ {0}. The case when k = 0 is obvious
so we assume k 6= 0.
N
1 X 2πikγ n e2πikγ 1 − e2πikN γ
(e ) =
N n=1 N 1 − e2πikγ

The numerator is bounded and 1 − e2πikγ 6= 0 as γ is irrational. The limit is

0 as N tends to infinity. Note that the integral on the right is 0 as well. (5.1)
holds for all exponentials. Moreover, (5.1) is linear so it applies to all sums
of exponentials including the partial sums P (x) of Fourier series of continuous
function f . Therefore, ∃N ∈ N such that
N Z 1
1 X
P (nγ) − P (x) dx < /3 (5.2)
N n=1 0

60
From section 3, we know continuous functions f (x) can be approximated by
trigonometric polynomials, that is, there exists trigonometric polynomials P (x)
such that |f (x) − P (x)| < /3, so we have

N
1 X
f (nγ) − P (nγ) < /3 (5.3)
N n=1

Z 1
f (nγ) − P (nγ) dx < /3 (5.4)
0

Combining (5.2), (5.3), (5.4) with Cauchy’s inequality, we have

N Z 1 N
1 X 1 X
f (nγ) − f (x) dx ≤ f (nγ) − P (nγ) +
N n=1 0 N n=1
Z 1
f (nγ) − P (nγ) dx+
0 (5.5)
N Z 1
1 X
P (nγ) − P (x) dx
N n=1 0

<

Then the lemma is proven.

To prove that the statement also applies to χ(a,b) , we need to bound it with
two continuous functions. As in the figure. They agree with χ(a,b) (x) but for
intervals of length 2. So,

f+
f−

a+ a a− b− b b+

Figure 4: approximate f

61
 Z 1 Z 1
b − a − 2 ≤ f− (x) dx and f+ (x) dx ≤ b − a + 2 (5.6)
0 0

Note that f− (x) ≤ χ(a,b) ≤ f+ (x). We have

N N N
1 X − 1 X 1 X +
f (nγ) ≤ χ(a,b) (nγ) ≤ f (nγ)
N n=1 N n=1 N n=1 

Apply Lemma 5.2 to f− (x) and f+ (x).

N Z 1
1 X −
f (nγ) → f− (x) dx as N → ∞
N n=1  0

N Z 1
1 X +
f (nγ) → f+ (x) dx as N → ∞
N n=1 0

Therefore,

Z 1 N Z 1
1 X
f− (x) dx ≤ lim χ(a,b) (nγ) ≤ f+ (x) dx
0 N →∞ N 0
n=1

Combining with (5.6), we get

b − a − 2 ≤ lim inf SN and lim sup SN ≤ b − a + 2

N →∞ N →∞

Since this is true for all , the limit exists and tends to a − b, proving Theorem
5.1.

We can generalize Theorem 5.1 that says (5.1) holds for χ(a,b) (x). In fact, (5.1)
for all integrable functions on [0, 1).
Corollary 5.5. (5.1) holds for merely integrable f of period 1 and irrational γ,

Proof. The proof is quite similar to that of Theorem 5.1, so we will just outline
it. Define functions fU (x) = supxn <x<xn+1 f (x) for x ∈ (xn , xn +1) and fL (x) =
inf xn <x<xn+1 f (x) for x ∈ (xn , xn + 1). We now have f bounded by fU and fL .

Z 1 Z 1 Z 1
fL (x) dx ≤ f (x) dx ≤ fU (x) dx
0 0 0

62
Moreover, recall the definition of integrability and we have

Z 1 Z 1
fU (x) dx − fL (x) dx < 
0 0

Note that fU and fL are sums of characteristic functions. By the linearity of

(5.1) and Theorem 5.1, (5.1) holds for fU and fL as well. Therefore, the limit
PN
limN →∞ N1 n=1 χ(a,b) (nγ) is bounded between two functions that can both be
R1
as close as to 0 f (x) dx, proving the corollary.

We now come back to the example in the beginning, where we are to prove that

∀ > 0, ∃n ∈ N such that hξn i < , where ξn is 2nπ

We can assume there is 0 that doesn’t satisfy the above statement and hξn i ≥ 
for all n. Also, ξn satisfies the following for all (a, b) ⊂ [0, 1).

| {ξ1 , ξ2 , ..., ξN } ∩ (a, b)|

lim =b−a
N →∞ N

Take (a, b) = (0, ) and the left hand side is 0. However, the right hand side is
obviously not, a contradiction.

After having some experience in equidistribution, we now ask is there an equiv-

alent condition for equidistribution? We understand that ξn being equidis-
tributed is essentially satisfying the following equality with f (x) = χ(a,b) (x) for
any (a, b) ⊂ [0, 1).
N Z 1
1 X
f (ξn ) → f (x) dx as N → ∞ (5.7)
N n=1 0

χ(a,b) is periodic and if the above equation holds for exponentials, we can deduce
that it also holds for χ(a,b) and prove equidistribution. However, can we go
reversely? This question leads to the following theorem.
Theorem 5.6. A sequence of reals ξ1 , ξ2 , ξ3 , ... in [0,1) is equidistributed if and
only if for all integers k 6= 0,
N
1 X 2πinkξn
e →0 as N → ∞ (5.8)
N n=1

63
 P P

Figure 5: reflection of the light

Since we have discussed one direction of the theorem, we will focus on the other
one when ξn is equidistributed and we want to deduce (5.8).

First, we know all characteristic functions in [0, 1) satisfies (5.7), therefore so

does the sums of all characteristic functions. Similar to the proof of Corollary
5.3, for f (x) = e2πinkx , we define fU and fL , two functions both satisfying (5.7)
and bound f (x) in a region that can be made to be as small as possible, proving
that f (x) satisfies (5.7) as well.

In fact, equidistribution is not merely a concept in number theory. Its idea

pervades in many fields of mathematics such as geometry.
Example 5.2. Consider a beam of light reflecting in a square. Note that during
the reflection, the incidence angle equals the reflection angle, as in the figure.

Let the slope of the ray be γ. Consider the grids formed by the reflections. In
fact, if γ is rational, the path repeats. If γ is irrational, the path is equidis-
tributed in the rectangle.

Figure 6: equidistribution of lines

64
Remark. Imagine a group of parallel lines that ”equally fill” the plane. Each of
them can be uniquely represented by the intercept with an axis. We can general-
ize that the equidistribution of these lines is the equidistribution of coordinates
of these intercepts on the axis. In this case, the reflection of rays produce two
group of parallel lines(labelled blue and red in figure 6) and actually both of
them are equidistributed given irrational slope. As in the figure, each red line
has an intercept with the axis and if these red lines ”equally fill” the rectangle,
their intercepts will be equidistributed along the axis.

5.3 A continuous but nowhere differentiable function

It is not difficult to find a obvious example of a continuous function that is not

differentiable at one single point, like the function f (x) = |x|. But is there a
continuous function that can be differentiable nowhere?

In this section, we will prove the following theorem:

Theorem 5.7. If 0 < α < 1, then the function
∞
X n
fα (x) = f (x) = 2−nα e−i2 x
(5.9)
n=0

is continuous but nowhere differentiable.

The proof of theorem can be split into two parts. First start with continuity
Lemma 5.8. If 0 < α < 1, then the function
∞
X n
fα (x) = f (x) = 2−nα e−i2 x
(5.10)
n=0

is continuous.

Proof. We will first check the absolute convergence of this series.

n
Define an = 2−nα e−i2 x . Using the Ratio test,
(
|an+1 | 2−(n+1)α e−i2 n+1)x
lim = lim (5.11)
n→∞ |an | n→∞ 2−nα e−i2n x
= lim 2−α e−i2x (5.12)
n→∞
<1 (5.13)

Therefore, the series fα (x) is absolute convergence thus it is a continuous func-

tion.

65
Then we come to the vital part of this theorem, it is nowhere differentiable.

To prove this, we need to introduce three methods:

• Partial Sums: SN (g) = g ∗ DN

• Cesaro summability: σN (g) = g ∗ FN
• Delayed Means: ∆N (g) = 2σ2N (g) − σN (g)

Hence, ∆N (g) = g ∗ [2F2N − FN ]

an einx ,then
P
We suppose g(x) ∼
X
SN (g)(x) = an einx ,
n≤N

S0 (g)(x) + S1 (g)(x) + · + SN −1 (g)(x)

σN (g)(x) =
N
N −1 X
1 X
= ak eikx
N n=0
|k|≤n
X |n|
= (1 − )an eiex ,
N
|n|≤N

and

∆N (g) = 2σ2N (g) − σN (g)

X |n| X |n|
=2 (1 − )an eiex − (1 − )an eiex
2N N
|n|≤2N |n|≤N
X X |n|
= an eiex + 2(1 − )an eiex .
2N
|n|≤N N ≤|n|≤2N

These methods can be visualized as in figures below.

According to the figures,we find out the delayed means are essentially equal to
the partial sums.

So we pick N 0 which is the largest integer of the form 2k with N 0 ≤ N . Then

we have
SN (f ) = ∆N 0 (f ) (5.14)

Then we will introduce a lemma related with this observation

66
 1

−N 0 N

an einx
P
Figure 7: SN (g)(x) = |n|≤N

−N 0 N
P |n| iex
Figure 8: σN (g)(x) = |n|≤N (1 − N )an e

Lemma 5.9. If 2N = 2n , then

n
∆2N (f ) − ∆N (f ) = 2−nα ei2 x
(5.15)

Proof. We have
∆2N (f ) = S2N (f )
∆N (f ) = SN (f )

Therefore,

∆2N (f ) − ∆N (f ) = S2N (f ) − SN (f )
n
= 2−nα ei2 x

Then, we will use the proof by contradiction to finish the proof of the theorem.
We assume f 0 (x0 ) exists for some x0
Lemma 5.10. Let g be any continuous function that is differentiable at x0 .
Then, the Cesaro means satisfy σN (g)0 (x0 ) = O(log N ), therefore

67
 1

−2N −N 0 N 2N

|n|
an eiex + iex
P P
Figure 9: ∆N (g) = |n|≤N N ≤|n|≤2N 2(1 − 2N )an e

∆N (g)0 (x0 ) = O(log N )

Proof. The Cesaro means

Z π Z π
0 0
σN (g) (x0 ) = FN (x0 − t)g(t)dt = FN0 (t)g(x0 − t)dt
−π −π

And FN is periodic, we have

Z π
FN0 (t)dt = 0
−π

And by g is differentiable, we have

Z π Z π
0
|σN (g) (x0 )| = | FN0 (t)[g(x0 − t) − g(x0 )]dt| ≤ C ||FN0 (t)||t|dt
−π −π

FN is a trigonometric polynomial of degree N whose coefficients are bounded by

1. Therefore, FN0 is a trigonometric polynomial of degree N whose coefficients
are no bigger than N . Hence

|F (t)0 | ≤ (2N + 1)N ≤ C1 N 2

Also,

1 sin2 (N t/2) 0
|FN (t)0 | = | |
N sin2 (t/2)
sin(N t/2) cos(N t/2) 1 cos(t/2) sin2 (N t/2)
= 2 −
sin (t/2) N sin3 (t/2)
C2
≤ 2
|t|

68
By using these two inequalities above, we have
Z Z
0 0
|σN (g) (x0 )| ≤ C |FN (t)||t|dt + C |FN0 (t)||t|dt
|t|≥1/n |t|≤1/n
Z Z
1
≤ CC1 dt + CC1 N dt
|t|≥1/n |t| |t|≤1/n

= O(log N ) + O(1)
= O(log N )

Now we know
∆2N (f )0 (x0 ) − ∆N (f 0 )(x0 ) = O(log N )

But according to the previous lemma, we have

|∆2N (f )0 (x0 ) − ∆N (f 0 )(x0 )| = 2n(1−α) ≥ cN 1−α

Then we get the contradiction and we finish our proof.

P∞ n
Remark : Actually, the real part and the imaginary part of fα (x) = n=0 2−nα e−i2 x

are both nowhere differentiable.

5.4 The heat equation on the circle

The investigation of heat flow is a vital part of the application of Fourier analysis.
This brings us a special partial differential equation, the heat equation.
Definition 5.3. Consider an infinite metal plate which we model as the plane
R2 , and suppose we are given an initial heat distribution at time t = 0. Let the
temperature at the point (x, y) at time t be denoted by u(x, y, t).

Consider a small square centered at (x0 , y0 ) with sides parallel to the axis and
of side length h. Using skills of calculus and the mean value theorem, we find
the time-dependent heat equation:

σ ∂u ∂2u ∂2u
= + 2 (5.16)
κ ∂t ∂x2 ∂y

where σ > 0 is a constant called the specific heat of the material and κ > 0 is
the conductivity of the material.

69
When the whole system reaches thermal equilibrium and ∂u
∂t = 0, the time-
dependent heat equation reduces to the steady-state heat equation

∂2u ∂2u
+ 2 =0 (5.17)
∂x2 ∂y

We can use Laplacian operator to simplify the equation

u=0

If we rewrite the Laplacian in polar coordinates and apply the chain rule, we
get
∂ 2 u 1 ∂u 1 ∂2u
u= 2
+ + 2 2
∂r r ∂r r ∂θ

Besides temperature on a plate, the investigation proceeds to describe the tem-

perature at points on a ring. Consider the ring as a thin circular wire where
there is no radial temperature dependence, i.e. u(r, θ) = u(θ) so that ∂u
∂r = 0,
in this case, the polar Laplacian reduces to

∂ 2 u 1 ∂u 1 ∂2u ∂2u
u= + + = (5.18)
∂r2 r ∂r r2 ∂θ2 ∂(rθ)2

Model the ring by the unit circle. A point on the circle is described by its angle
∂2u ∂2u
θ = 2πx, where the variable x lies between 0 and 1. Then we get ∂(rθ) 2 = ∂x2 .

If u(x, t) denotes the temperature at time t of a point described by the angle θ,

then (5.16) shows that u satisfies the differential equation

∂u ∂2u
=c 2
∂t ∂x
where the constant c is a positive physical constant which depends on the ma-
terial of the ring.

Another notation writes ut = cuxx . Also, in the following calculation, we will

assume that c = 1.

To solve this partial differential function,the boundary conditions and initial

conditions are needed. First we impose the initial condition assuming that f is
our initial data
u(x, 0) = f (x)
Also, by the property of unit circle, we obtain the periodic boundary conditions

u(L, t) = u(−L, t),

70
 ∂ ∂
(L, t) = (−L, t)
∂x ∂x

To solve this partial differential equation, we will need a useful technique, sep-
aration of variables.

5.4.1 Useful tool: separation of variables

Here’s how we use separation of variables to solve the heat equation.

Firstly, separate variables and look for special solutions of the form

u(x, t) = A(x)B(t).

Plugging this into the two sides of the heat equation gives
∂
(x, t) = A(x)B 0 (t)
∂t
∂2
(x, t) = A00 (x)B(t)
∂x2

From the initial heat equation, we get

A(x)B 0 (t) = A00 (x)B(t)

Therefore
B 0 (t) A00 (x)
=
B(t) A(x)

The left side of this equation is a function only of the variable t, while the
right side is a function only of the variable x. The equality exists only if these
functions are constant functions. Therefore, we deduce that there must exist a
constant λ such that
B 0 (t) A00 (x)
= =λ
B(t) A(x)

In other words, we now have two ODEs to solve

B 0 (t) = λB(t),

A00 (x) = λA(x)

To tackle these, one need to have a revision of principles of solving ODEs.

Theorem 5.11. Any function f for which f 0 (t) = cf (t) where c is a constant
is of the form f (t) = aect .

71
Theorem 5.12. Let k be a square root of the constant c. If f is a solution of
the equation f 00 (x) = cf (x), then there exists constants α and β such that

f (x) = αekx + βe−kx .

The key to the proofs of these theorems is to define a new function based on
the equation to prove, say φ(x), and use this to investigate the relation between
f (x) and ekx .

Now with the results of the two theorems,

B(t) = aeλt

and √ √
A(x) = αe λx
+ βe− λx
.

From boundary conditions, we know that A(x) = A(−x), so

√ √ √ √
αe λx
+ βe− λx
= αe− λx
+ βe λx

√ √
(α − β)(e λx
− e− λx
)=0
Since x is a variable in the interval [0, 1], we get α = β.

Also, since the point lies on the unit circle, A must be periodic of period 1,
which means
A(x) = A(x + 1)
√ √ √ √ √ √
e λx
+ e− λx
=e λx
e λ
+ e− λx − λ
e
√ √
λ(x+1) − λx
e =e

This is not possible just in R since the value of x varies in the interval [0, 1].
√
As a result, we can deduct that λ ∈ C and assume that λ = −k 2 where k is a
constant.

Then the equation reduces to

eik(x+1) = e−ikx

eik(2x+1) = 1.
So the only possibility is k = 2πn and λ = −4π 2 n2 where n ∈ Z.

By superposing these solutions, we are led to

∞
2
n2 t 2πinx
X
u(x, t) = an e−4π e (5.19)
n=−∞

72
Then a question arises: do we have u(x, t) → f (x) as t tends to 0? For a better
understanding of our above solution and to investigate this problem, we will
define the heat kernel in the following section.

5.4.2 Heat kernel

Definition 5.4. The heat kernel for the circle, written as Ht , is given by
∞
2
n2 t 2πinx
X
Ht (x) = e−4π e ,
−∞

so that we can rewrite u(x, t) as

u(x, t) = (f ∗ Ht )(x).

Corollary 5.13. If u(x, t) = (f ∗ Ht )(x) where Ht is the heat kernel, and f is

Riemann integrable, then
Z 1
|u(x, t) − f (x)|2 dx → 0 as t → 0.
0

Proof. Since f is a Riemann integrable function on [0, 1], by definition of the

norm of f , we get
Z 1
|u(x, t) − f (x)|2 dx = ||u(x, t) − f (x)||2
0

2 2
Using the property of convolution, û = fˆĤt = fˆ(x)e−4π n t , combined with
Parseval’s identity, we get
1
2
n2 t 2
X
||u(x, t) − f (x)||2 = |fˆ(ω)|2 |1 − e−4π |
n=0

2
n2 t
When t → 0, (1 − e−4π ) → 0, meaning that ||u(x, t) − f (x)||2 → 0 and we
are done!

The heat kernel can actually be considered as an analogue of the Poisson kernel.
A change of variables in (5) leads to the solution
X 2
u(θ, τ ) = an e−n τ einθ = (f ∗ hr )(θ)

73
 of the equation

ur = uθθ with 0 ≤ θ ≤ 2π and τ > 0

P∞ 2
with boundary condition u(θ, 0) = f (θ) ∼ an einθ . Here hr (θ) = n=−∞ e−n τ einθ .
P

This version of the heat kernel on [0,

P2π] is the analogue of the Poisson kernel,
∞ −|n|τ inθ
which can be written as Pr (θ) = n=−∞ e e with r = e−τ (and so
0 < r < 1 corresponds to τ > 0).

References
[1] Elias M. Stein and Rami Shakarchi. Fourier Analysis, an introduction.
Princeton University Press, New Jersey 08540.

74

View publication stats