AN INTRODUCTION TO FOURIER SERIES AND

TRANSFORMS

JUSTIN A. TARQUINO

Abstract. This paper offers a brief introduction to the theory, calculation,

and application of Fourier series and transforms. First, we define the trigono-
metric and exponential representations of the Fourier series, coupled with some
examples of its use. We then define the Fourier transform, followed by an il-
lustrative example of its function and distinctness from the Fourier Series.
Thirdly, we establish the definition and properties of the Dirac Delta Func-
tion, which we proceed to use in taking the Fourier transform of constant
functions. Lastly, we briefly discuss the application of Fourier analysis to sig-
nal processing through the advent of the magnitude Fourier transform. This
paper is designed for readers willing to accept Fourier theory at face value,
merely motivated to know its formulae and methods of use. This paper seeks
to satisfy such motivation by providing a straightforward and relatively con-
cise walkthrough of the core formulae and real-world application of Fourier
analysis.

Contents
1. Introduction 1
2. Fourier Series 2
3. Fourier Transforms 7
4. The Dirac Delta Distribution 12
5. Application to Signal Processing 17
Acknowledgments 20
References 21

1. Introduction
Named after its founder, the great French mathematician Joseph Fourier (1768-
1830), Fourier analysis allows for the decomposition of periodic and aperiodic func-
tions alike into a series of sinusoidal and exponential functions. This is done by
taking advantage of the orthogonality between the sine and cosine functions, which
act as a basis in the space of functions. The pillars of Fourier analysis are Fourier
Series and Fourier Transforms. The first deals with periodic functions, and the
second deals with aperiodic functions. Fourier series and transforms have powerful
real-world applications in signal processing, seismology, econometrics, and physics,
to name a few. Fourier analysis is embedded in the technology we find so essen-
tial to our modern lifestyle, for example, in the storage and transmission of digital
images.
1
2 JUSTIN A. TARQUINO

2. Fourier Series
Theorem 2.1 (Trigonometric Fourier Representation). Let a function f : R → R
be periodic for 0 ≤ x ≤ L, L being the length of the period. This means that for all
x ∈ R, f (x) = f (x + L). Fourier’s Theorem states that f (x) can be written as the
series

∞     
X 2πnx 2πnx
(2.2) f (x) = a0 + an cos + bn sin
n=1
L L
where

Z L
1
(2.3) a0 = f (x)dx,
L 0
Z L  
2 2πnx
(2.4) an = f (x) cos dx,
L 0 L
Z L  
2 2πnx
(2.5) bn = f (x) sin dx.
L 0 L
a0 , an , and bn are called Fourier coefficients.
Remark 2.6. The bounds of integration for (2.3), (2.4), and (2.5) are from 0 to
L. However, Fourier coefficients can be determined just the same as long as you
integrate over a period length. For example, you can integrate from − L2 to L2 , − L4
to 3L
4 , −7L to 6L, and in general, for p ∈ R, from p to L + p. One may want to use
alternate bounds of integration in order to avoid discontinuities or if it produces a
cleaner way of evaluating an antiderivative.
The definition just outlined can be daunting. Therefore, working through a
couple of examples will be useful for understanding how it works and that it works.
Example 2.7. (Sawtooth Function) Consider a sawtooth function where f (x) =
Ax for some A ∈ R and for − L2 < x < L2 with a period L. Such a function looks
like

Figure 1. Odd sawtooth function f (x) = x with a period of 2

 AN INTRODUCTION TO FOURIER SERIES AND TRANSFORMS 3

This function is odd because we have defined it such that it is symmetric across
the origin. Therefore, by integrating from −L L
2 to 2 , it is clear that a0 will integrate
to 0 since we only want to integrate over the original function. Furthermore, all of
our an terms will also integrate to 0 since an odd function multiplied by an even
function remains an odd function. However, an odd function multiplied by an odd
function is an even function. Therefore, we only need to consider bn since it is the
only term that does not integrate to 0. Plugging in our function Ax, we get
Z L2  
2 2πnx
bn = Ax sin dx.
L − L2 L
By integrating by parts for a fixed constant r ∈ R, we find that
  Z
−1 −1
Z
x sin(rx)dx = x cos(rx) − cos(rx)dx.
r r
−x 1
= cos(rx) + 2 sin(rx)
r r
Letting r = 2πn
L , we have
L
L2
     
2A Lx 2πnx 2πnx 2
bn = − cos + 2 2 sin dx
L 2πn L 4π n L −L
2

L2 L2 L2 L2
 
2A
= − cos(πnx) + 2 2 sin(πnx) − cos(−πnx) 2 2 sin(πnx) .
L 4πn 4π n 4πn 4π n
Since sin(πn) = 0 for any integer n, we’re left with
AL AL
bn = − cos(πn) − cos(−πn).
2πn 2πn
Since cos(x) = cos(−x) because it is an even function,
AL
bn = − cos(πn)
πn
AL
=⇒ bn = (−1)n+1 .
πn
Therefore, our Fourier Series for f (x) will look like,
∞  
n+1 AL 2πnx
X
f (x) = (−1) sin
n=1
πn L
∞
AL X (−1)n+1
 
2πnx
(2.8) =⇒ f (x) = sin .
π n=1 n L
Writing out the first few terms of this series, we get
       
AL 2πx 1 4πx 1 6πx
(2.9) f (x) = sin − sin + sin − ··· .
π L 2 L 3 L
4 JUSTIN A. TARQUINO

Plugging in L = 2 and A = 1 for this series, we get

 
2 1 1
(2.10) f (x) = sin (πx) − sin (2πx) + sin (3πx) − · · · ,
π 2 3
which graphically looks like the following:

(a) The first term of (2.10) (b) The first five terms of (2.10)

(c) The first ten terms of (2.10) (d) The first twenty terms of (2.10)

Figure 2. Increasing terms of our series (2.10) approaches f (x) =

x when L = 2 and A = 1

As the number of terms increases, our series looks more and more like the func-
tion it is intended to describe. This is exactly the purpose of Fourier analysis.

We can restate Fourier’s Theorem by reinterpreting the sine and cosine terms as
components of complex exponential functions. This results in a series of exponential
functions consisting of a real part and an imaginary part.
Theorem 2.11 (Exponential Fourier Representation). Let a function f : R → R
be periodic for 0 ≤ x ≤ L, L being the length of the period. This means that for all
x ∈ R, f (x) = f (x + L). Fourier’s Theorem states that f (x) can be written as the
following series:

∞
X
(2.12) f (x) = Cn e2iπnx/L
n=−∞

where the Cn coefficients are given by

Z L
1
(2.13) Cn = f (x)e−2iπnx/L dx.
L 0

Proof. We will start by examining Euler’s identity for sine and cosine, which states
that for z ∈ C,
 AN INTRODUCTION TO FOURIER SERIES AND TRANSFORMS 5

eiz + e−iz
(2.14) cos z = ,
2

eiz − e−iz
(2.15) sin z = .
2i
2πnx
Plugging this in to (2.2) by letting z = L , we have
∞
X eiz + e−iz eiz − e−iz
f (x) = a0 + an + bn
n=1
2 2i

∞
X an eiz + an e−iz − ibn eiz + ibn e−iz
= a0 +
n=1
2
∞    
X an − ibn an + ibn
= a0 + eiz + e−iz .
n=1
2 2
Now, we can take both terms in the parentheses and assign them to the same
variable by letting
an + ibn an − ibn
(2.16) C−n = , Cn = .
2 2
Thus, we can rewrite our series as
∞
X
f (x) = a0 + Cn eiz + C−n e−iz .
n=1
Plugging back in our definition for z, we have
∞
X
f (x) = a0 + Cn e2πinx/L + C−n e−2πinx/L .
n=1
We can conglomerate our exponential terms and coefficients by accounting for
only Cn while allowing the sum to operate along all integers. Doing so multiplies
2π − inx/L by −1, and thus both exponential terms go to e2πinx/L . Lastly, we can
let our a0 = C0 to account for n = 0. Therefore, we have
∞
X
f (x) = Cn e2πinx/L .
n=−∞

an − ibn
Now, we can calculate Cn based on how we defined it. In (2.16), Cn = .
2
By plugging in (2.4) and (2.5), we have that

" Z  #
1 2 L 2i L
  Z 
2πnx 2πnx
Cn = f (x) cos dx − f (x) sin dx .
2 L 0 L L 0 L

Recalling (2.14) and (2.15), we then have

6 JUSTIN A. TARQUINO

1 L + e−2πinx/L
Z  2πinx/L 
e
Cn = f (x) dx
L 0 2
i L − e−2πinx/L
Z  2πinx/L 
e
− f (x) dx
L 0 2i
Combining terms and canceling out the i’s in the second term, we get

L
e2πinx/L + e−2πinx/L − e2πinx/L + e−2πinx/L
Z  
1
Cn = f (x) dx
L 0 2

1 L
Z
=⇒ Cn = f (x)e−2πinx/L dx
L 0
□
Remark 2.17. We can now switch between exponential and trigonometric repre-
sentations of the Fourier Series relatively easily. This process of switching is even
simpler when one knows an and bn for the trigonometric series and Cn for the
exponential series. The following formulae allow us to relate all three:

an + ibn
(2.18) Cn = ,
2

(2.19) an = Cn + C−n ,

(2.20) bn = i(Cn − C−n ).

Let’s work through an example, though, this time using the exponential repre-
sentation of a Fourier series.
Example 2.21. We will repeat the sawtooth function example. Plugging in Ax to
(2.13),

Z L
1 2
(2.22) Cn = Axe−2πnx/L .
L −L
2

Integrating by parts, we know that for a constant r,

xe−rx e−rx
Z
xe−rx dx = − − 2 .
r r
This is valid except for the case in which r = 0. Therefore, by letting r = i2πn/L,
Z L2
we need to consider the special case in which n = 0. If n = 0, (2.22) will be Ax
−L
2
which just integrates to 0 since its an odd function. Therefore, considering all n
except for 0, we have
L !
−A Lxe−2πinx/L L2 e−i2πnx/L 2
Cn = + .
L i2πn 4π 2 n2 −L 2
 AN INTRODUCTION TO FOURIER SERIES AND TRANSFORMS 7

The second term evaluates to 0 because it’s limits are equal, and we are left with

L2 e−πin L2 eπin
 
A
Cn = − +
L i4πn i4πn
A L2 −πin
=−· (e + eπin ).
L i4πn
The exponential terms sum to 2(−1)n . So, factoring i out, we get

iAL
Cn = (−1)n for n ̸= 0.
2πn
Therefore, our series can be expressed as
X iAL −2πnx/L
f (x) = (−1)n e .
2πn
n̸=0
We can double-check that this answer agrees with our previous result by writing
this exponential in terms of sines and cosines:
    
n iAL 2πn 2πn
X
f (x) = (−1) cos + i sin .
2πn L L
n̸=0

Since (−1)n /n is an odd function of n and the cosine terms are even functions
of n, those terms sum to 0. However, since the sine terms are odd in sine, we only
consider the positive terms and then double the result. Lastly, i2 (−1)n = (−1)n+1 ,
leaving us with

∞
AL X (−1)n+1
 
2πnx
(2.23) f (x) = sin ,
π n=1 n L
which agrees with (2.8).

3. Fourier Transforms
The Fourier series we have been working with applies to periodic functions.
However, an appropriate next question is asking if we can do the same for a function
that is not periodic: does there exist a Fourier series for an aperiodic function? At
first, this question may sound like it necessitates an entirely different approach,
but in reality, we can extend Fourier series over periodic functions to aperiodic
functions by initially examining a function over a period length, − L2 ≤ x ≤ L2 , but
then extending L → ∞. This move intuitively makes sense because we can think of
an aperiodic function as a function that repeats only at infinity, which essentially
means it doesn’t repeat at all.
Theorem 3.1 (Exponential Fourier Transform Representation). Let a function
f : R → R be aperiodic. We can express this function as

Z ∞
(3.1) f (x) = fˆ(k)eikx dk
−∞

where
8 JUSTIN A. TARQUINO

Z ∞
1
(3.2) fˆ(k) = f (x)e−ikx dx.
2π −∞

Proof. We know by Theorem 2.11 that a periodic function f (x) with period L can
be expressed as
∞
X
f (x) = Cn e2πinx/L .
−∞
Let kn = 2πn/L. Taking the derivative of this, we have that dkn = 2π dn/L.
However, recall that we are still operating over a discrete sum of integers, and thus
d
dn = 1 because that is the amount we are changing from one n term to another.
Therefore, we have that dkn = 2π/L. If we take the limit of L → ∞, we have that
limL→∞ dkn = limL→∞ 2π/L = 0 Now, since dn is 1, it is equivalently true to write
our function as
X∞
f (x) = Cn eikn x dn
−∞
∞  
X
ikn x L
=⇒ f (x) = Cn e dkn .
−∞
2π
Here is the critical step. We want to take the limit of L as it approaches infinity.
By doing this, we get the following:

lim dkn = lim 2π/L → 0,

L→∞ L→∞

∞
"  #
X
ikn x L
lim f (x) = lim Cn e dkn .
L→∞ L→∞
−∞
2π
As L approaches infinity, dkn gets really small, eventually transforming our
discrete sum into a continuous sum. Thus we can take the integral:
Z ∞  
L
f (x) = Cn eikn x dkn .
−∞ 2π
Lastly, because we are no longer operating on a discrete sum, we can drop the
index n and let kn = k and for notation’s sake, we can let fˆ(k) = Cn 2π
L
. Therefore,
plugging back in for k, we end up with
Z ∞  
L
f (x) = fˆ(k) eikx dkn
−∞ 2π
where

L 1 L
Z
L
fˆ(k) = Cn = · f (x)e−2πinx/L dx
2π 2π L 0
Z L
1
=⇒ fˆ(k) = f (x)e−ikx dk.
2π 0
□
Let’s move through some examples to further understand how this works and
that it works.
 AN INTRODUCTION TO FOURIER SERIES AND TRANSFORMS 9

Example 3.3. First, consider a periodic, odd, step function. For n ∈ N, let

(
A if (n−1)L
2 < x < nL
2
f (x) = (1−n)L
−A if − nL2 < x < 2

The graph for this looks like

Figure 3. Odd periodic function where A = 1 and L = 2

This function is periodic, and so by taking the trigonometric representation of

the Fourier Series for this function, we need to only consider the bn terms since this
is an odd function, and therefore the a0 and an equal 0. We thus have

Z L   Z L  
2 2 2πnx 2 2 2πnx
bn = f (x) sin = ·2 A sin
L −L
2
L L 0 L

  L !
4A L 2πnx 2
2A
=− · cos = (1 − cos(πn)).
L 2πn L 0 πn

Because for all even n, cos(πn) = 1, we need to only consider summing odd n
terms. Thus we have

∞  
2A X 1 2πnx
(3.4) f (x) = sin
π n L
odd n

       
2A 2πx 1 6πx 1 10πx
(3.5) = sin + sin + sin + ··· .
π L 3 L 5 L
In exponential terms, this series looks like

∞
−iA X 1   πnx 
(3.6) f (x) = 1 − cos e2iπnx/L .
π n=−∞ n L
10 JUSTIN A. TARQUINO

Plugging in 1 = A and L = 2 for (3.6), our calculation will yield expected results:

(a) The first term of (3.6) (b) The first five terms of (3.6)

(c) The first ten terms of (3.6) (d) The first twenty terms of (3.6)

Now, equipped with Fourier transforms, let us now consider a step function with
just one step and then zero everywhere else, i.e., consider a constant A and a period
L in which

A
 if 0 < x < L2
(3.7) f (x) = −A if − L2 < x < 0
if x < − L2 or 1 < L

0

2

Figure 5. Odd aperiodic function where A = 1

Here,
Z ∞
ˆ 1
f (k) = f (x)e−ikx dk
2π −∞
Z 0 Z L2
1 1
=⇒ fˆ(k) = −Ae−ikx dk + Ae−ikx dk
2π − L2 2π 0
 AN INTRODUCTION TO FOURIER SERIES AND TRANSFORMS 11

! L !
0
−A e−ikx A e−ikx 2
=⇒ fˆ(k) = +
2π −ik −L 2π −ik 0
2

−A  
=⇒ fˆ(k) = −(1 − e−ikL/2 ) + (eikL/2 − 1)
2πik
  
ˆ −Ai kl
=⇒ f (k) = 1 − cos .
πk 2
So for f (x) we get,
Z ∞
f (x) = fˆ(k)e−ikx dk
−∞
∞   
−Ai
Z
kL
=⇒ f (x) = 1 − cos e−ikx dk
−∞ πk 2
∞   
−iA
Z
1 kL
(3.8) =⇒ f (x) = 1 − cos e−ikx dk.
π −∞ k 2
When we compare (3.8) with (3.6), they are highly similar. The key difference
is that in (3.6), we are taking a discrete sum over the integers n whereas in (3.8)
we are integrating over a continuous variable k.
A lurking concern may be how we can know that (3.8) accurately describes (3.8).
To do this, we can approximate the integral by taking a discrete sum over a very
small dk. The smaller dk is, the better the approximation we will have.

(a) Zoomed in view (b) Zoomed out view

Figure 6. Approximation of (3.8) using dk = .1, A = 1, L = 2,

using the first 200 terms

Our approximation is highly revealing. We can say that for a small dk, our
approximation will look like Figure 5, though only when zoomed in to the origin.
When zoomed out enough, however, we can see that the behavior at the origin
repeats regularly. The smaller we take our dk to be, the better our approximation
for (3.7) will be as the step will repeat at further and further distances from the
origin. This is equivalent to letting the period approach infinity, which is exactly
what we did originally.
12 JUSTIN A. TARQUINO

Theorem 3.9. Fourier transforms can be expressed in terms of trigonometric func-

tions as the following:
Z ∞ Z ∞
(3.10) f (x) = Ce (k) cos(kx)dk + i Co (k) sin(kx)dk
−∞ −∞

where fˆ(k) is split into its even component,

∞
−i
Z
(3.11) Ce (k) = fe (x) cos(kx)dx,
2π −∞
and its odd component:
Z ∞
1
(3.12) Co (k) = fo (x) sin(kx)dx.
2π −∞
The even and odd components of the original function can be expressed respec-
tively as
f (x) + f (−x) f (x) − f (−x)
fe (x) = and fo (x) =
2 2
such that fe (x) + fo (x) = f (x).

4. The Dirac Delta Distribution

We have thus determined methods for determining trigonometric and exponential
Fourier representations of periodic and aperiodic functions alike. However, we have
yet to consider functions that are merely non-zero constants. A non-zero constant
function f is intriguing because f (x) is the same everywhere, and thus the period
would be so small as to be 0 since the function is always repeating. In order to
tackle such functions, we must first consider the Dirac delta function.
Definition 4.1 (Dirac’s Delta “Function”). The Dirac delta function, δ(x), is zero
everywhere except at the origin and has area 1.
Remark 4.2. This “function” is a function in name alone as it is actually a distri-
bution. Distributions, rather than functions themselves, are generalizations of the
concept of a function, describing where functions lie in the differentiable function
space. It is analogous to the statistical distribution of probabilities in the prob-
ability space. This definition may appear outlandish as it seems impossible for a
function that is non-zero except for at a single point to have a non-zero area. How-
ever, this function does make sense when the limits of certain functions are taken.
We will see what this means later.
Let’s begin by considering some properties of the delta function. First, consider
letting f : R → R be an arbitrary function. Then,
Z ∞
(4.3) δ(x)f (x)dx = f (0).
−∞
This is because the delta function is zero everywhere at the origin. Thus, the
function integrates to 0 on all points that are not at the origin, and thus we need
only consider f (0). Furthermore, f (0) is a constant and thus can be factored out.
Lastly, we have defined the area under δ(x) to be exactly one, and thus we have
 AN INTRODUCTION TO FOURIER SERIES AND TRANSFORMS 13

Z ∞ Z ∞ Z ∞
δ(x)f (x)dx = δ(x)f (0)dx = f (0) δ(x)dx = f (0).
−∞ −∞ −∞
We can further generalize this property by considering the following:
Z ∞
δ(x − x0 )f (x)dx = f (x0 ).
−∞
This fact follows from similar logic as laid out in (4.3 ). However, we result in f (x0 )
because δ(x − x0 ) is only non-zero at x0 , and thus it is the only point that does
not integrate to 0. With these properties in mind, we can thus conclude with the
following.
Definition 4.4. The Dirac delta function is the function that satisfies the following
for an arbitrary function f :
Z ∞
(4.5) f (x)δ(x)dx = f (0).
−∞

Let us now examine an example of what the delta function actually looks like.
There is only one true delta distribution (zero everywhere except at the origin with
an area of 1), though it can be written down in various ways. We will use Fourier
analysis to write the delta distribution down explicitly. We will first present proof
for one and then merely list examples for others.
Example 4.6. Let us take the Fourier transform of the exponential function f (x) =
Ae−b|x| where both A and b are constants in R (the shape of this function can be
found in Figure 7).
Plugging into our formula for Theorem 3.2, we get
Z ∞
1
fˆ(k) = Ae−b|x| e−ikx dx
2π −∞
Z 0 Z ∞
1 1
= Aebx e−ikx dx + Ae−bx e−ikx dx
2π −∞ 2π 0
Z 0 Z ∞ 
A 1
= e(b−ik)x dx + e(−b−ik)x dx
2π −∞ 2π 0
!
(b−ik)x 0 −∞
A e e(−b−ik)x
= +
2π b − ik −∞ (−b − ik) 0
 
A 1
= + 1−b − ik
2π b − ik
 
A b + ik + b − ik
=
2π b2 + k 2
Ab
(4.7) fˆ(k) = .
π(b2 + k 2 )
Equation (4.7) has a special name, the Lorentzian function, formalized as
c1
(4.8) f (k) =
c2 + c3 k
where c1 , c2 and c3 are constants in R.
14 JUSTIN A. TARQUINO

With this calculation of (4.7), we can now compute the Fourier transform for
f (x) = 1. To do so, we will use a “tapering” method, in which we take a function
that is equal to 1 or very nearly equal to 1 for very large x, though eventually
tapers off to 0. By “tapering” infinitely, we will therefore arrive at a function in
which is identical to f (x) = 1. We will first taper with the function given above.
Let f (x) = Ae−bix . If we fix A = 1 and let b → 0, our function would approach
f (x) = 1 because e0·ix = 1. We can see this graphically by letting b = 0.01:

Figure 7. Exponential decay function with A = 1 for all lines,

b = 1 for the black line, b = 0.1 for the red line, and b = 0.01 for
the green line

As b gets smaller and smaller, i.e., as our exponential decay function approaches
f (x) = 1, our fˆ(k) will look more and more like a delta function because a taller
and taller spike will appear. We can see this graphically:
Letting b approach 0, the Lorentizan Fourier transform starts to look like a delta
function. We can see this by taking the integral of
Z ∞
b dk
2 + k2 )
.
−∞ π(b
Z ∞  
b dk −1 k
Using the fact that 2 2
= tan , we have
−∞ (b + k ) b
Z ∞   ∞ !
bdk 1 −1 k 1
tan−1 (∞) − tan−1 (−∞)


2 2
= tan =
−∞ π(b + k ) π b −∞ π
 
1 π −π
= − = 1.
π 2 2
Since we arrived at this result regardless of the choice of b, and because this
function becomes sharply peaked at 0 when b → 0, we have found an equivalent
representation of the delta function, which we can formalize as

b
(4.9) δ(k) = lim .
b→0π(b2 + k 2 )
Other representations of the delta function can be found by taking the Fourier
transform of other “tapered” functions. Some prominent examples are the square
wave function and the Gaussian function. The Gaussian function gains its signif-
icance by representing the probability density function of a normal distribution,
 AN INTRODUCTION TO FOURIER SERIES AND TRANSFORMS 15

(a) Fourier transform when b = 1 (b) Fourier transform when b = 0.1

(c) Fourier transform when b = 0.01

Figure 8. Fourier transform for f (x) = Ae−b|x| for A = 1 for

different values of b

taking on a “bell” shape. It is special because it accurately describes the proba-

bility distribution of a wide range of phenomena, such as rolling a pair of dice or
height. The Gaussian function can be generalized as f : R → R where
(x−b)2
(4.10) f (x) = ae− 2c2

where a describes the height of the peak, b describes the x-coordinate of the center
of the peak, and c describes the standard deviation.
Example 4.11. Let’s consider the Gaussian function, which has a peak height
of 1 and is centered at the origin. Such a function can be expressed as f (x) =
2
e−dx where d is a constant in R. We have simplified the exponent such that now,
q
1
2d = c, where c is the standard deviation. The Fourier transform of this Gaussian
function is

2
e−k /4d
(4.12) fˆ(k) = √ .
2 πd
Its corresponding delta function representation would be

2
e−k /4d
(4.13) δ(k) = lim √ .
d→0 2 πd
16 JUSTIN A. TARQUINO

(a) Gaussian and its transform for d = 1 (b) Gaussian and its transform d = .01

(c) Gaussian and its transform for d = .001

2
Figure 9. The Gaussian function, f (x) = e−dx shown in red and
its transform shown in green for different values of d

Remark 4.14. The result from Example 4.11 is particularly notable as the Fourier
transform of a Gaussian function is also a Gaussian function. It is one of the few
examples of functions that are their own transforms.
Example 4.15. The Fourier transform of a square wave, i.e., for constants a, A ∈
R,

(
A if − a < x < a
(4.16) f (x) =
0 if x < −a or x < a
If we let A = 1, our Fourier transform will be

sin(ka)
(4.17) fˆ(k) = .
πk
This Fourier transform attains another representation of the delta function by let-
ting a → ∞ (which correspondingly allows the function to approach f (x) = 1):

sin(ka)
(4.18) δ(k) = lim .
a→∞πk
In general, we have seen that the Fourier transform for f (x) = 1 is a delta function.
Therefore, plugging in 1 to (3.2), we are left with
Z ∞
1
(4.19) δ(k) = e−ikx dx.
2π −∞
For curious readers, the calculations for (4.12),(4.13), (4.17), and (4.18) are
derived in Section 3.5.3 of Morin’s Fourier Analysis[1].
 AN INTRODUCTION TO FOURIER SERIES AND TRANSFORMS 17

(a) Square wave and transform for a = 1 (b) Square wave and transform for a = 10

(c) Square wave and transform for a = 100

Figure 10. The square wave function shown in red and its trans-
form shown in blue for different values of a

5. Application to Signal Processing

Fourier transforms and series are hugely applicable in signal processing. A signal
is a function that expresses the behavior and condition of a physical measure within
a system. Signals are measured by the intensity of the behavior over time, which
can be highly informative, though even more information about the system can be
learned. By taking the Fourier transform, we can also determine the underlying
fundamental frequencies a system exhibits. As we have demonstrated, the Fourier
transform does this by decomposing a signal into sine and cosine components, which
can be reinterpreted as frequency components. In other words, the original signal
displays how a system changes over time, and its Fourier transform displays the
same change over frequency.
Let’s further break down how this works in terms of sound. Sound is produced by
a vibrating object which causes surrounding air molecules to undergo compression
or rarefaction by bouncing off of each other at various rates. The air molecules
remain relatively stable in space though sound energy is transferred through space.
This ultimately translates to oscillations in air pressure, and therefore, sound is
measured as air pressure over time. The shape it takes on is called its waveform.
The waveform of a sound appears, on a macroscopic scale, aperiodic, though it is,
in fact, periodic on a microscopic scale. This phenomenon can be seen in Figure
11.
Specifically in regards to sound, “signal tells us when certain notes are played
in time, but hides the information about frequencies. In contrast, the Fourier
transform of music displays which notes (frequencies) are played but hides the
information about when the notes are played.”[2] This speaks to the utility of the
Fourier transform as it extracts underlying phenomena mostly imperceptible to
humans.
18 JUSTIN A. TARQUINO

Figure 11. The waveform of a recording of the first five measures

of Beethoven’s Fifth with seconds 7.3 to 7.8 magnified

Graphing the frequencies of a signal is most often done digitally through what is
called a Discrete Fourier Transform (DFT). As opposed to analog systems, which
take continuous readings of systems, digital technology can only store and process
a finite, albeit large, number of data. Therefore, a signal can only be digitized by
converting it into what is called a discrete-time signal. A discrete-time signal is a
signal expressed over a set of finite points, where points in time equidistant from
each other are each associated with the amplitude of the sound at that point in
time. The sampling period, T , is the time between samplings, and the sampling
rate is the inverse of T , measured in Hertz. The following graphs display the signal
as amplitude over time.

Figure 12. The digitizing of an arbitrary signal with a sampling

rate of 32Hz

Once a signal has been digitized, a Fourier transform can be taken rather readily.
Firstly, we also want to digitize both a sine wave and a cosine wave.

Figure 13. The digitizing of a cosine wave with frequency 2Hz

 AN INTRODUCTION TO FOURIER SERIES AND TRANSFORMS 19

We then want to multiply the signals together.

Figure 14. The digitizing of a cosine wave multiplied by our ar-

bitrary signal

Our last step deviates from the Fourier transform as we must now take the Rie-
mann sum of our new signal, as opposed to integrating. The greater our sampling
rate is, the more closely our DFT will align with the true Fourier transform.

Figure 15. The digitizing of a cosine wave multiplied by our ar-

bitrary signal

Once this sum is calculated, we then want to repeat the same process with a sine
wave of the same frequency. Using both sums, we can calculate the magnitude of a
given frequency by summing the squares of both sums and then taking their square
root. This entire process is called a magnitude Fourier Transform, whose results
are graphed in Figure 16(B). The algorithm for our DFT is a function g : Z → R,
with a given sampling of T ,

∞
X
(5.1) g(k) = T f (nT )e(−iknT ) .
n=−∞

To calculate the magnitude of a frequency using our method of sinusoidal func-

tions, we recall Euler’s Formula to get

∞
X ∞
X
(5.2) g(k) = T f (nT ) cos(−knT ) + i T f (nT ) sin(−knT ).
n=−∞ n=−∞

This will yield a complex number. We can√calculate the magnitude of any com-
plex number of the form z = a + bi as |z| = a2 + b2 . We are therefore motivated
to define our magnitude function as |g(k)| : Z → R where

v !2 !2
∞ ∞
u
u X X
(5.3) |g(k)| = t T f (nT ) cos(−knT ) + T f (nT ) sin(−knT ) .
n=−∞ n=−∞
20 JUSTIN A. TARQUINO

By repeating this process for a range of frequencies, we are able to graph against
magnitude, which displays the underlying frequency components of a given wave-
form.

(a) Amplitude over time(seconds) (b) Magnitude over frequency(Hz)

Figure 16. Waveform and magnitude Fourier transform of a tone

C4(”the middle C”) played on the piano

The above example of Figure 16 indicates that the tone C4 possesses a funda-
mental frequency of 262Hz, which is supported by the spikes around 262Hz, 524Hz,
and 786Hz, all of which are multiples of 262Hz.
The Dirac delta function explains the physical properties of impulses on the
spectrum of sound. Impulse sounds make sharp sounds for a short amount of time,
such as a drum beat or a clap of the hand. Therefore, by taking its magnitude
Fourier transform, we would expect to see a constant magnitude over a range of
frequencies. This is indeed the case, as can be seen in Figure 17, where we take the
magnitude Fourier transform of an impulse sound.

(a) Amplitude over time(seconds) (b) Magnitude over frequency(Hz)

Figure 17. Waveform and magnitude Fourier transform for a

clapping sound

We can now see that the energy of an impulse sound is distributed relatively
evenly across the frequency spectrum.

Acknowledgments
It is more than a pleasure to thank my mentor, Micah Gay, for all her help
and kindness in guiding my understanding and focus. Of course, I would like
to thank professors Daniil Rudenko and László Babai for leading and organizing
the apprentice lectures. I would also like to thank Professor Peter May for his
organization of the entire Math REU. Lastly, I would like to thank Otto Reed for
helping me feel like I wasn’t falling behind.
 AN INTRODUCTION TO FOURIER SERIES AND TRANSFORMS 21

References
[1] David Morin. Fourier Analysis. scholar.harvard.edu/files/david-morin/files/waves fourier.pdf
[2] Meinard Müller. Fundamentals of Music Processing: Audio, Analysis, Algorithms, Applica-
tions. https://doi.org/10.1007/978-3-319-21945-5
[3] Marı́a Cristina Pereyra and Lesley A. Ward. Harmonic Analysis: From Fourier to Wavelets.
http://gibbs.if.usp.br/ marchett/fismat2/harmonic-analysis-fourier-wavelet pereyra-ward.pdf