INTRODUCTION TO FOURIER ANALYSIS

SAGAR TIKOO

Abstract. In this paper we introduce basic concepts of Fourier analysis, de-

velop basic theory of it, and provide the solution of Basel problem as its ap-
plication. We also include a proof of the uniqueness of trigonometric series.

Contents
1. Introduction 1
2. Convergence 2
3. Uniqueness 6
4. Parseval’s Identity and the Basel Problem 9
Acknowledgements 11
References 11

1. Introduction
The development and study of Fourier series is motivated by the desire to model
general functions as sums of simple trigonometric functions. The ability to model
functions in a simple manner is tremendously valuable to science and technology,
as it allows for the understanding and mathematical manipulation of the original
function in terms of the simpler decomposition.
Jean-Baptiste Joseph Fourier, after whom the field is named, worked extensively
with trigonometric series and used the Fourier series in order to solve the heat
equation in physics. The ability of Fourier analysis to model complicated func-
tions has led to its application in partial differential equations, quantum physics,
thermodynamics, signal processing, and many other fields.
We begin our study of Fourier analysis with some definitions:
Definition 1.1. If f is an integrable function defined on the interval [a, b] of length
L, the n-th Fourier coefficient of f is defined by
Z b
1
fˆ(n) = f (x)e−2πinx/L dx
L a

Definition 1.2. The Fourier series of f is formally defined by

∞
X
F (f (x)) = fˆ(n)e2πinx/L
n=−∞

Date: August 28, 2015.

1
2 SAGAR TIKOO

Definition 1.3. The N -th Fourier partial sum of f is defined by

N
X
SN (f (x)) = fˆ(n)e2πinx/L
n=−N

Due to Euler’s formula (eix = cos x + i sin x), this definition of Fourier series
expresses a function f in terms of a summation of sine and cosine functions. This
fact leads to our first intuition about Fourier series: like the trigonometric functions
it is composed of, the Fourier series of a function is periodic in nature. As we will
see later, a function need only satisfy a mild condition in order for its Fourier series
to converge to the function itself.

2. Convergence
In order to justify the use of Fourier series to model functions and explore the
various application of Fourier analysis, we must first investigate whether the Fourier
series is, indeed, a good approximation of the original function. Until now, we have
only stated the definition of Fourier series, and demonstrated that it is a summation
of sine and cosine waves dependent on the original function, but have not proven
that the series converges to its original function.
Before embarking on the subtle task of proving convergence, we must first under-
stand what the formula provided in the definition conceptually entails. The idea of
convolutions is central to understanding Fourier series, and will also serve a pivotal
role in more advanced Fourier analysis.
Definition 2.1. Given two functions f and g integrable on [a, b], their convolution
f ∗ g on [a, b] is defined as
Z b
(f ∗ g)(x) = f (y)g(x − y)dy
a

Conceptually, convoluting two functions gives the area of their overlap as one is
shifted, or translated, over the other. The operation is, loosely speaking, a way of
determining the “blend” or “weighted average” of the two functions.
Some of the essential properties of convolutions are stated here:
Proposition 2.2. Let c be any complex number and let f , g, and h be functions
integrable on the interval of convolution. Then we have:
(1) f ∗ g = g ∗ f
(2) f ∗ (g ∗ h) = (f ∗ g) ∗ h
(3) f ∗ (g + h) = (f ∗ g) + (f ∗ h)
(4) c(f ∗ g) = (cf ) ∗ g
(5) (f ∗ g) is continuous
The first few properties are easily checked using the definition of convolution;
the operation preserves commutativity, associativity, distributivity, and scalar as-
sociativity. Property (5) illuminates a key aspect of the convolution: it is more
“regular” than the original functions. Whereas f and g need only be integrable,
their convolution f ∗ g is continuous.
 INTRODUCTION TO FOURIER ANALYSIS 3

Given the concept of convolution, we can rearrange terms in the Fourier partial
sum to more deeply study Fourier series. We will be using 2π-periodic functions
defined on the interval [−π, π] henceforth when discussing the properties of Fourier
series, for the sake of simplicity and standardization. Since L is set to 2π, the
exponent in the Fourier coefficient (2πinx/L) reduces to (inx). We have:

N
X
SN (f )(x) = fˆ(n)einx
n=−N
N Z π
X 1
f (y)e−iny dy einx


=
2π −π
n=−N
Z π N
1 X
ein(x−y) dy


= f (y)
2π −π n=−N

After some rearranging of terms and exchanging the integral and the summation,
the Fourier partial sum reveals itself as a convolution of f and the trigonometric
PN
sum n=−N einx . This trigonometric sum, critically important in studying Fourier
series, is called the Dirichlet kernel.
Definition 2.3. The N -th Dirichlet kernel is defined to be
N
X
DN (x) = einx
n=−N

Therefore, the Fourier partial sum can be represented as a convolution of its

original function and the Dirichlet kernel:

SN (f )(x) = (f ∗ Dn )(x)
The Dirichlet kernel has certain essential properties that will be necessary in the
investigation of the convergence of Fourier series.
Proposition 2.4.
sin((N + 21 )x)
DN (x) =
sin( x2 )
PN
Proof. Let r = eix . Recall that the sum of a geometric series n=−N rn is given
n
by 1−r
1−r .

N −1
X X 1 − rN +1 r−N − 1
DN = rn + rn = +
n=0
1−r 1−r
n=−N
−N −1/2
r − rN +1/2 e−(N +1/2)ix − e(N +1/2)ix
= =
r−1/2
−r 1/2 e−ix/2 − eix/2
−2i sin((N + 1/2)x) sin((N + 1/2)x)
= =
−2i sin(x/2) sin(x/2)

4 SAGAR TIKOO

Given that many other kernel

P∞ functions used in convolutions exist (e.g. the
|n| inx
Poisson kernel Pr (x) = n=∞ r e ), it is useful to define certain sets of
properties that cause the kernel to be “good.”
Definition 2.5. A family of functions {Kn (x)}∞n=1 on the circle is a family of good
kernels if:
Z π
1
(a) For all n ∈ N, Kn (x)dx = 1.
2π −π
Z π
(b) There exists M > 0 such that for all n ∈ N, |Kn (x)|dx ≤ M.
−π
Z
(c) For all δ > 0, |Kn (x)|dx → 0 as n → ∞
δ≤|x|≤π

Functions on the circle and 2π-periodic functions on R are naturally connected.

Since any point on the circle can be represented as eiθ , we may define a 2π-periodic
function f by its counterpart F on the circle: f (θ) = F (eiθ ). The importance of
good kernels arises from their use in convolutions with functions.
Proposition 2.6. Let {Kn }∞ n=1 be a family of good kernels, and f be an integrable
function on the circle. Then, whenever f is continuous at x,
lim (f ∗ Kn )(x) = f (x)
n→∞

If the Dirichlet kernel was a good kernel, the problem of Fourier series con-
vergence would be easily solved with the proof of this sole proposition. However,
the Dirichlet kernel does not satisfy all of the requirements in Definition 2.5. The
definition of the Dirichlet kernel as a sum of exponentials quickly shows us that
Property (a) holds.
However, the “goodness” of DN falls apart at Property (b). While the mean
value of DN is 1, the integral of its absolute value is very large, due to the fact that
DN (x) oscillates rapidly between positive and negative values. More precisely,
Z π
|DN (x)|dx ≥ c log n, as N → ∞
−π

All of this is to demonstrate that the problem of Fourier convergence is more

subtle than expected at first glance, due to the Dirichlet kernel’s extreme behavior
that prevents it from being a good kernel.
Convergence of Fourier series can be understood in several ways (e.g. mean-
square convergence, Abel summability, etc.). Our proof will focus on the pointwise
convergence of a Fourier series to its original function. The Riemann-Lebesgue
Lemma is central to the proof of pointwise convergence.
Lemma 2.7. Riemann-Lebesgue Lemma: If f is an arbitrary Riemann-integrable
function on compact interval I = [a, b], then:
Z
lim f (x)e−inx dx = 0
n→∞ I

Proof. Suppose first that f is a piecewise constant function on a compact interval

I = [a, b]. This means that I can be subdivided into finitely many subintervals
 INTRODUCTION TO FOURIER ANALYSIS 5

Ik = [ak , bk ] such that f is constant on each subinterval. Then, f can be written

as a summation of the constant functions:
N
X
For all x ∈ (ak , bk ), f (x) = ck gk (x)
k=1

where gk (x) = 1 in Ik and gk (x) = 0 outside of Ik .

We then get:
N Z N N
e−inx
Z X X Z X
−inx −inx −inx
f (x)e dx = ck gk (x)e dx = ck e dx = ck ||Ik ||
I I Ik −in
k=1 k=1 k=1

This is a sum of finitely many terms, all of which approach 0 as n approaches

infinity.
Now, let f be an arbitrary Riemann-integrable function on compact I. By
the definition of Riemann-integrability, the integral of f can be approximated by
rectangles (i.e. the integral of a piecewise constant function). In other words, for
any  > 0 there exists a piecewise constant function g such that:
Z
|f (x) − g(x)|dx < 
I
Let g be a function as described above.
Z Z Z
f (x)e−inx dx = (f (x) − g(x))e−inx dx + g(x)e−inx dx
I I I
Z Z
≤ |f (x) − g(x)||e−inx |dx + g(x)e−inx dx
I I
Z Z
≤ |f (x) − g(x)|dx + g(x)e−inx dx <  + 0
I I
as n → ∞. 
Proposition 2.8. If f is bounded on [a, b], c ∈ [a, b], and for any δ > 0, f is
integrable on [a, c − δ] ∪ [c + δ, b], then f is integrable on [a, b].
Proof. Since f is bounded, there exists a number M such that |f | ≤ M . Let  > 0.
Choose δ > 0 such that 4δM ≤ /3. Since f is integrable on [a, c − δ] ∪ [c +
δ, b], we can choose P1 and P2 to be partitions of [a, c − δ] ∪ [c + δ, b] such that
U (Pi , f ) − L(Pi , f ) < /3, i = 1, 2. Let P = P1 ∪ (c − δ, 0] ∪ (0, c + δ) ∪ P2 . Then
U (P, f ) − L(P, f ) < /3 + 2δM + 2δM + /3 ≤ . Thus, f is integrable on the whole
interval [a, b]. 
With the powerful Riemann-Lebesgue lemma, the dirichlet kernel, and Proposi-
tion 2.8, the pointwise convergence of Fourier series is readily investigated.
Theorem 2.9. Let f be an integrable function on [−π, π] differentiable at a point
x0 . Then SN (f )(x0 ) → f (x0 ) as N → ∞.
Proof. Let x0 ∈ [−π, π] be such that f is differentiable at x0 . Let F (t) be the
function defined by:
(
f (x0 −t)−f (t)
t t 6= 0, t ∈ (−π, π)
F (t) = 0
−f (x0 ) t=0
6 SAGAR TIKOO

Due to the differentiability at x0 and integrability of f , we know that F is bounded

on [−π, π] and integrable on [−π, −δ] ∪ [δ, π] for every δ ∈ (0, π). By Proposition
2.8, F is integrable on [−π, π].
We know from our analysis of the Dirichlet kernel that SN (f )(x0 ) = (f ∗DN )(x0 ),
sin((N + 21 )x) 1
Rπ
DN (x) = sin( x , and 2π −π
Dn dx = 1. Then,
2)
Z π
1
SN f (x0 ) − f (x0 ) = f (x0 − t)DN (t)dt − f (x0 )
2π −π
Z π
1
= [f (x0 − t) − f (x0 )]DN (t)dt
2π −π
Z π
1
= F (t)tDN (t)dt
2π −π
Z π
1 t
= F (t) sin((N + 1/2)t)dt
2π −π sin(t/2)
Z π
1 t
= F (t) [sin(N t) cos(t/2) + cos(N t) sin(t/2)]dt
2π −π sin(t/2)
Z π Z π
1 t cos(t/2) 1
= F (t) sin(N t)dt + F (t)t cos(N t)dt
2π −π sin(t/2) 2π −π
Let Q(t) = F (t)t cos(t/2)
sin(t/2) and let R(t) = F (t)t. We know that Q and R are
t
continuous since F (t), t, cos(t/2), and sin(t/2) are all continuous in the interval
[−π, π]. Then, we can apply the Riemman-Lebesgue lemma to Q and R to finish
the proof:
Z π Z π
1 1
SN f (x0 ) − f (x0 ) = Q(t) sin(N t)dt + R(t) cos(N t)dt → 0 as N → ∞
2π −π 2π −π


3. Uniqueness
In previous sections we used Fourier series to express periodic functions as
trigonometric series. A natural question is whether this expression is unique. In
this section, we are going to prove the Cantor-Riemann theorem which answers the
above question affirmatively.
Several concepts and lemmas must be addressed before the uniqueness proof may
be initiated. The first and most essential of these is the Cantor-Lebesgue Theorem.
Theorem 3.1. Cantor-Lebesgue Theorem Let {an }∞ ∞
n=1 , {bn }n=1 be two se-
quences of real numbers. If
lim an cos x + bn sin x = 0
n→∞
for every x ∈ R, then an , bn → 0 as n → ∞.
p
Proof. Let pn = a2n + b2n and choose θn such that pn sin θn = bn , pn cos θn = an .
Then an cos(nx) + bn sin(nx) = pn cos(nx − θn ) → 0 as n → ∞. If we can prove
that pn → 0 then, since sin and cos are bounded, an , bn must approach 0.

Assume pn 6→ 0. Then there exists a subsequence {nk } of {n} and a number

δ such that for any positive integer k, pnk ≥ δ > 0. We can further assume that
 INTRODUCTION TO FOURIER ANALYSIS 7

nk+1
nk ≥ 3. Then we attempt to put a lower bound on the cos(nx − θ) term for all
n at at least one point. We define I1 to be
(θn1 − π3 ) (θn1 + π3 )
 
,
n1 n1
Then we have: cos(n1 x − θn1 ) ≥ 12 , for all x ∈ I1 . It is known that |I1 | = 3n 2π
1
and nn12 ≥ 3. As x ranges over I1 , (n2 x − θn2 ) ranges over (n2 I1 − θn2 ). We have:
2π
|n2 I1 − θn2 | = n2 |I1 | = n2 ( 3n 1
) ≥ 2π. Just as we defined I1 to be the interval for
which cos(n1 x − θn1 ) ≥ 1/2, there exists as interval I2 for which cos(n2 x − θn2 ) ≥
1/2 for all x ∈ I2 . Since the range of (n2 I1 − θn2 ) exceeds 2π, there exists I2 ⊂ I1
such that:
cos(n2 x − θn2 ) ≥ 1/2 for all x ∈ I2
2π
|I2 | =
3n2
Proceeding by induction, we can construct intervals Ik , k ∈ Z>0 such that I1 ⊃
I2 ⊃ I3 ⊃ . . . for which cos(nk x − θnk ) ≥ 1/2 for T∞all k, for all x ∈ Ik . By
Cantor’s closed interval theorem, there exists ξ ∈ k=1 Ik such that cos(nk ξ −
θnk ) ≥ 21 for all k. Therefore pnk cos(nk ξ − θnk ) ≥ δ/2 for all k. This contradicts
the convergence of pn cos(nx − θn ) to 0, since it will always be greater than δ/2 at
x = ξ. 

One of the first limitations encountered in attempting to prove the uniqueness of

Fourier series, is the need to analyze function’s second derivatives without imposing
the restriction of differentiability at each point. For this reason, we use a generalized
version of the second derivative rather than differentiating at x.
Definition 3.2. The Schwarz second derivative of a function F is defined by:
F (x + h) − 2F (x) + F (x − h)
DF (x) = lim
h→0 h2
Remark 3.3. We can see that the Schwartz second derivative coincides with f 00 (x)
by using L0 H ôpital0 s Rule for f with continuous f 00 (x):
f (x + h) − 2f (x) + f (x − h) f 0 (x + h) − f 0 (x − h)
lim = lim
h→0 h2 h→0 2h
f 00 (x + h) + f 00 (x − h)
= lim = f 00 (x)
h→0 2
Lemma 3.4. A continuous function F on R with zero Schwarz second derivative
everywhere is a linear function.
Proof. Let A(x) = x2 . We prove that F + A is convex for every  > 0. Suppose
it is not, then there exist a < t < b and a linear function L(x) = µx + λ such
that G(x) = F (x) + A(x) − L(x) satisfies G(a) = G(b) = 0 and G(t) > 0. Choose
s ∈ (a, b) where G attains its maximum in [a, b]. Then DG(s) ≤ 0. Since DL(s) = 0
and DA(s) = 2 > 0, we see that DF (s) < 0, a contradiction.
As  tends to 0, we see that F itself is convex. Similarly, we can prove that F is
concave. Therefore, F is linear. 

Theorem 3.5. Cantor-Riemann Uniqueness Theorem

8 SAGAR TIKOO

Let {cn }n∈Z be complex numbers such that

N
X
lim cn einx = 0
N →∞
n=−N
for all x ∈ R, then cn = 0 for all n ∈ Z.
P∞
Proof. Since n=−∞ cn einx converges everywhere, cn einx + c−n e−inx → 0 for all x
as n → ∞. By the Cantor-Lebesgue Theorem, |cn | + |c−n | → 0 as n → ∞. Let F
be the formal integral:
N
x2 X cn inx
Z Z X
F = cn einx = c0 + e
2 (in)2
n=−N n6=0

Using the Weierstrass M-test on the following expression

cn einx c−n e−inx sup(|cn | + |c−n |)
2
+ 2
≤
(in) (−in) n2
2
we deduce that F (x) − c0 x2 is a continuous function and F is the uniform limit of
its partial sums. Applying the Schwarz second derivative, first to eix and then to
F , we find:
h 2
ei(x+h) − 2eix + ei(x−h)

ix sin 2
D(eix ) = = −e h
h2 2
  nh 2 
F (x + h) − 2F (x) + F (x − h) inx sin 2
X
DF (x) = lim = lim c0 + cn e nh
h→0 h2 h→0
2
n6=0
 
sin nh P cn inx
Since limh→0 nh
2
= 1 and c0 + n6=0 (in) 2e = 0,
2

  nh 2 
inx sin 2
X cn
DF (x) = lim c0 + e nh
=0
h→0 (in)2 2
n6=0

Since DF (x) = 0, the Schwarz lemma tells us that F is linear, so for some α, β:
x2 X cn inx
F (x) = c0 + e = αx + β
2 (in)2
n6=0
X cn x2
2
einx = −c0 + αx + β
(in) 2
n6=0
P cn inx
Since n6=0 (in)2 e is bounded in x, we have c0 = 0 = α. We now know that
sN (x) converges uniformly to 0, where sN (x) is defined by:
X cn inx
sN (x) = −β + e
(in)2
0<|n|≤N

However, for each n 6= 0 we can also express cn in terms of sN :

(in)2 2π
Z
cn = sN (x)e−inx dx
2π 0
Since sN (x) converges uniformly to 0, we have
(in)2 2π
Z
cn = lim sN (x)e−inx dx = 0
N →∞ 2π 0
 INTRODUCTION TO FOURIER ANALYSIS 9


Riemann’s idea of examining the double integral of the trigonometric series and
Cantor’s insight that the static value cn could be analyzed and determined using a
sequence of functions sN were seminal advances in mathematical understanding of
trigonometric series and beyond.

4. Parseval’s Identity and the Basel Problem

The Basel problem, relevant to both mathematical analysis and number theory,
was first introduced
P∞ in the 17th century when Pietro Mengoli asked for the exact
sum of the series n=1 n12 . The problem puzzled mathematicians for many years
2
before Euler finally offered a solution ( π6 ) in 1735, but it took him even longer to
develop a rigorous proof. Using Fourier Analysis, however, the problem’s treatment
becomes simple.
The solution to the Basel problem will require a property of Fourier series called
Parseval’s Identity, which will first require examination of Fourier series convergence
on the L2 norm.
Definition 4.1. The Hilbert inner product of two functions f and g on the
interval [−π, π] (where g ∗ denotes the complex conjugate of g such that if g(c) =
a + bi then g ∗ (c) = a − bi) is defined by:
Z π
1
hf, gi = f (x)g(x)∗ dx
2π −π
Definition 4.2. The L2 norm of a function f is defined by:
p
||f || = hf, f ∗ i
Definition 4.3. Two functions f, g on [−π, π] are orthogonal (denoted by f ⊥ g)
if hf, gi = 0.
Proposition 4.4. Pythagoras Theorem If f and g are orthogonal, then
||f + g||2 = ||f ||2 + ||g||2
.
Lemma 4.5. Best L2 Approximation Amongst all the functions of the form
PN
PN = n=1 bn en , where eN = einx , the n-th Fourier partial sum SN f best approx-
imates f in the L2 norm. More precisely:
||f − PN || ≥ ||f − SN f ||
Proof. Let EN f = f − SN f . Notice that the family en is orthonormal and that
fˆ = hf, ei. Consequently:
hEN f, en i = 0 for all n ≤ N
h(SN f − PN ), EN f i = 0
We can write:
f − PN = f − SN f + SN f − PN = EN f + (SN f − PN )
The Pythagoras theorem then gives us:
||f − SN f || = ||EN f || ≤ ||f − PN ||

10 SAGAR TIKOO

Theorem 4.6. Mean-square convergence If f is a square-integrable function,

then ||f − SN f || → 0 as N → ∞.
Proof. Fix  > 0. We begin by approximating f using a continuous function g that
satisfies:
sup |g(x)| ≤ sup |f (x)| = B on [−π, π]
Z π
|f (x) − g(x)|dx < 2
−π
We then find:
Z π
2 1
||f − g|| = |f (x) − g(x)|2 dx
2π −π
Z π
1
= |f (x) − g(x)||f (x) − g(x)|dx
2π −π
Z π
2B
≤ |f (x) − g(x)|dx
2π −π
≤ C2
Since g is continuous, we may approximate it by a trigonometric polynomial P :
||g − P || < 
Then,
||f − P || < C 0 
We conclude with the Best Approximation Lemma:
||f − SN f || ≤ ||f − P || < C 0 


Lemma 4.7. For an integrable function f on [−π, π], we have

hf − SN f, SN f i = 0
Proof. The lemma follows from the statement in our proof of the Best Approxima-
tion Lemma that hEN f, en i = 0. 

Theorem 4.8. Parseval’s Identity For a periodic square-integrable function f

on the circle with convergent Fourier series, Fourier coefficient fˆ, and complex
conjugate f ∗ :
∞ Z π
X
ˆ 2 1
|f (n)| = (f (x))2 dx
n=−∞
2π −π

Proof. By the above lemma and Pythagoras theorem,

||f ||2 = ||f − SN f ||2 + ||SN f ||2
for every N ≥ 0. By theorem 4.4, we see that ||f ||2 = limN →∞ ||SN f ||2 . Since
{einx }n∈Z is an orthonormal set, we have
N
X
2
||SN f || = |fˆ(n)|2
n=−N
 INTRODUCTION TO FOURIER ANALYSIS 11

Let N → ∞, we get Parseval’s identity

∞
X
||f ||2 = |fˆ(n)|2
n=−∞

Theorem 4.9. Solution to Basel Problem
∞
X 1 π2
=
n=1
n2 6

Proof. Let f (x) = x. Since f is continuous, differentiable, integrable, and real, we

apply Parseval’s identity to get:
∞ Z π
X
ˆ 2 1
|f (n)| = x2 dx
n=−∞
2π −π

Taking the Fourier coefficient of f (x) = x and integrating by parts:

Z π π
1 ixe−inx e−inx

ˆ 1 −inx
f (n) = xe dx = +
2π −π 2π n n2 −π
 π
1 x 1
= (i cos(nx) + sin(nx)) + 2 (cos(nx) − i sin(nx))
2π n n −π
cos(nπ)i (−1)n
= = i
n n
Then we find
∞ ∞ Z π
X 1 1 X ˆ 2 1 2 π2
= |f (n)| = x dx =
n=1
n2 2 n=−∞ 4π −π 6

Acknowledgements. It is a pleasure to thank my mentors, Zhiyuan Ding for his
indispensable guidance and suggestions in the development of this paper and Yun
Cheng for helpful discussions, as well as Peter May for organizing the Research
Experience for Undergraduates.

References
[1] Elias M. Stein, and Rami Shakarchi. Fourier Analysis: An Introduction. Princeton University
Press, 2003.
[2] Ash J M. Uniqueness of representation by trigonometric series[J]. American Mathematical
Monthly, 1989, 96(10): 873-885.
[3] Cody Dianopoulos, Joran Layne, and Alex Yokokawa. Basel Problem Proof.
http://www.academia.edu/3669174/Basel Problem Proof.