MATH.APP.

240 Fourier methods

Merja Laaksonen and Petteri Laakkonen, TUT 2020

M
F
 MATH.APP.240 F OURIER METHODS / M e / 2020 2

Forewords

This text contains the lecture notes of the course MATH.APP.240 Fourier methods. Our aim is
to give an introduction to the mathematics involved with Fourier methods, and the text should
give the reader skills required for understanding engineering texts on the topic. Because of this
emphasis, since we only assume very basic knowledge on mathematics, and time restrictions of
the course, we have (with our mathematician’s hearts bleeding) chosen to omit some mathemat-
ical details. The text contains the most important aspects and examples. An interested reader
can turn to other text books for a more comprehensive treatment. There are nice selection of
books available from the low level to hard core texts. A low level text with a similar emphasis
as this one can be found in Advanced Modern Engineering Mathematics by Glyn James. A nice
but mathematically bit more challenging treatment can be found in Fourier Analysis: An intro-
duction by Elias M. Stein and Rami Shakarchi. A word of caution is in place before proceeding.
The definitions of the Fourier transformation and related concepts may vary from book to book.
We have just adopted one of the multiple alternatives.
The definition of the Fourier series and the Fourier transform involve an integral, so knowing
the elementary integration methods and formulas makes things much easier. When dealing with
Fourier series, trigonometric functions and the related identities are heavily used, so the reader
might want to do some recap on them. The Fourier transform turns a real valued function into
complex valued one, so complex numbers are also naturally involved and make certain consid-
erations easier. The required results related to complex numbers and trigonometric functions
can be found in Appendix B. When dealing with the Fast Fourier Transform (FFT) things are
most conveniently represented by matrices and vectors. To sum up: The reader of this material
is assumed to be familiar with basic concepts of analysis, linear algebra, and complex numbers.
No specialized knowledge is needed.
If possible, use a (symbolic) software. Using computer is only in a minor role since we con-
centrate in understanding certain aspects related to Fourier transforms. However, if you are
applying Fourier methods in a real application using computers is compulsory. On this course,
we use Matlab (with symbolic toolbox). While you should be able to do the basic calculations
on paper, drawing and evaluating the Fourier series and transforms using a computer makes the
theory more approachable and you can verify your calculations.
Contents
1 Introduction 5
1.1 Basic concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.1.1 Even and odd functions . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.1.2 Heaviside step function . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.1.3 Dirac delta function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.1.4 Scaling- and shifting operations . . . . . . . . . . . . . . . . . . . . . 16

2 Fourier series 19
2.1 Fourier theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2 Exponential Fourier series . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3 Some Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4 The coefficients of the Fourier series . . . . . . . . . . . . . . . . . . . . . . . 27
2.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.5.1 Finding values of infinite sums using Fourier series . . . . . . . . . . . 33
2.6 Linearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.6.1 Fourier series of scaled and shifted functions . . . . . . . . . . . . . . 36
2.7 Convergence of Fourier series . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.7.1 Rate of convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.7.2 Gibbs phenomenon . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.7.3 *Uniform convergence . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.8 Periodic extensions of functions over finite intervals . . . . . . . . . . . . . . . 44
2.9 *Term-by-term differentiation and integration of Fourier series . . . . . . . . . 47
2.10 Yet another form of Fourier series . . . . . . . . . . . . . . . . . . . . . . . . 50
2.11 Discrete frequency spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.12 Parseval’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

3 Discrete Fourier transform (DFT) 57

3.1 Definition of Discrete Fourier transform . . . . . . . . . . . . . . . . . . . . . 57
3.2 Properties of discrete Fourier transforms . . . . . . . . . . . . . . . . . . . . . 62
3.3 Fast Fourier transform (FFT) . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

4 Fourier transform 66
4.1 Definition of Fourier transform . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.2 Sinc function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.3 Spectrum of a non periodic function . . . . . . . . . . . . . . . . . . . . . . . 71
4.4 Properties of Fourier transform . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.5 Fourier transforms involving impulse functions . . . . . . . . . . . . . . . . . 84
4.6 Parseval’s identity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
CONTENTS MATH.APP.240 F OURIER METHODS / M e / 2020 4

5 Discrete-time Fourier transform (DTFT) 90

6 Sampling frequency 92

7 A summarizing example 95

A Tabular. Selected Fourier transforms. 100

B Selected results on complex numbers and trigonometry 101

B.1 Some trigonometric identities . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
1 Introduction MATH.APP.240 F OURIER METHODS / M e / 2020 5

1 Introduction

The reader is probably familiar with the idea of representing a function as a combination of other
functions, e.g. we can write infinitely differentiable function sin(x) as a Maclaurin (Taylor)
series ∞
n
1 3 1 5 X −1

x2n+1 .


sin x = x − x + x − + · · · =
3! 5! n=0
2n + 1 !
The function and its series representation on the right hand side correspond to each other pro-
vided that the series converges, i.e. x belongs to the interval of convergence. Thus, on the
interval of convergence we have represented the function as a combination of exponents

1, x, x2 , x3 , . . . , xn , . . .

and their coefficients an . The Taylor series expansion plays a crucial role in many applica-
tions (e.g. approximations, numerical integration, etc.) as well as in theoretical considerations
(e.g. we can show that term-by-term integration/differentiation yields of the series yields the
integral/derivative of the function).
It is clear that the Taylor series is not a unique way of representing a function as a combination
of other functions. Alternative representations might give us further insights into the function’s
behavior or properties. The need for such representations can stem from a specific engineering
application. For example, representing a sound wave as a linear combination of sine functions
would enable us to design filters for unwanted frequencies. Such filters are used for example in
CD-players.
So, we may now ask if we can represent any periodic function with period T = 2π ω
as a com-
bination of simple periodic function such as sine functions. The answer is affirmative. If f is a
periodic function with period T = 2πω
then we can find constants An and φn such that
∞
X

A0 + An sin nωx + φn .
n=1

This representation is called Fourier series expansion after the French mathematician and physi-
cist Joseph Fourier (1768-1830) ) who developed the ideas dating back to ancient Babylonian
astronomers into a significant mathematical tool. Fourier applied his results to study heat con-
duction in a metal plate and was the first to provide a general solution to the heat equation.
At first, there was some criticism due to lack of mathematical preciseness of Fourier’s results,
but as soon as proper mathematical tools were developed the mathematical background was
confirmed. Applications of the theory soon followed and nowadays they involve applications
related to vibration processes and signal processing.
Since we can describe a periodic function as a Fourier series, we can actually take any func-
tion on a bounded interval and give it a Fourier series representation on that specific interval.
One may ask what happens if we extend the interval to infinity and change the Fourier series
accordingly. Roughly speaking, this kind of limiting process leads to a continuous distribution
of frequency components instead of a discrete one appearing in the Fourier series. This leads to
the Fourier transform Z ∞
F f ω = f (u)e−jωu du.


−∞

The Fourier series can be seen as a special case of this more general transform.
1.1 Basic concepts MATH.APP.240 F OURIER METHODS / M e / 2020 6

The data available when dealing with real world applications is discretized. This means that
we do not usually have a function that we work on, but a sequence of samples. In order to
deal with such an information we can use discrete Fourier transform. With modern computers,
this enables a wide range of applications in signal processing. However, the computational
cost of evaluating the transform or approximating it with a desired accuracy can be too steep.
In the famous paper by Cooley and Tukey, an algorithm reducing this cost significantly was
proposed. The idea of the algorithm itself was not new since a similar method for reducing the
computational cost was used earlier. However, it made the explosion of digital signal processing
possible and is referred as one of the most important algorithm of our life time.

1.1 Basic concepts

A function f is periodic if there exists a constant T , such that for all real numbers t we have



f t+T =f t .
The constant T is called the period of the function. If T is a period, then mT where m ∈ Z is
also a period. However, T is not automatically a period if mT with fixed m ∈ Z is.
The shortest of all periods is called the fundamental period and is often the most interesting
one, provided that it exists. However, when applying Fourier methods it is sufficient to find any
period. We make a standing assumption that the period T is positive, and if a negative value is
required we denote −T .

T
2T

Figure 1. A periodic function.

According to the definition, a periodic function should attain a value at any real number. How-
ever, we often deal with functions that are discontinuous or that have no specified value at some
isolated point. Such a situation is not a crucial restriction to the theory we are about to develop.
At such a point, we can set a desired value or just ignore the behavior there. This is justified,
since the Fourier series or transform will not be affected by the value at one point (this is related
to the fact that the value of the integral will not change if we change the value of the function
at one point) and by doing so we can avoid mathematical clumsiness that is not important for
applications or the theory.
♠Example 1.1. Sine and cosine are periodic functions with fundamental period 2π. Let us find
the fundamental period T of



f : f t = sin ωt + ϕ , ω > 0.



By the definition, we have f t = f t + T for every real number t, i.e.




sin ωt + ϕ = sin ω(t + T ) + ϕ = sin ωt + ϕ + ωT .
1.1 Basic concepts MATH.APP.240 F OURIER METHODS / M e / 2020 7

Because sine has the fundamental period 2π, we must have

ωT = n2π, n ∈ Z

or alternatively
2π
.
T =n
ω
Since zero is not an eligible period, the smallest possible period is
2π
T = .
ω
♠

♠Example 1.2. Find a period of

f : f t = 3 cos 2t + 2 cos 3t .



Since T = 2π is a period

of both of the functions cos 2t and sin(3t , it is a period of their
linear combination f t . The following calculation verifies this:




f t + T = 3 cos 2(t + T ) + 2 cos 3(t + T )



= 3 cos 2t + 4π + 2 cos 3t + 6π



= 3 cos 2t + 2 cos 3t = f (t).

Does the function has a shorter period. Plot the graph of the function and deduce why not. ♠

♠Example 1.3. Show that

√


f : f t = cos 2t + cos 3t

is not periodic.

√
Both cosine functions are periodic having fundamental periods T1 = 2π/ 2 and T2 = 2π/3.
Assuming that the sum function has a period T > 0, then for every real number t

√ √


f t + T = cos 2t + 2T + cos 3t + 3T = f (t).

In particular for t = 0,

√


f T = cos 2T + cos 3T = f (0) = 2.

Since cosine has maximum value √ 1, both of the cosine functions in the sum must attain their
maximum simultaneously, i.e. 2T = n2π, n ∈ N and 3T = k2π, k ∈ N. This implies that
k
n
= √32 ∈
/ Q. This is a contradiction and f is not periodic. ♠

Frequency tells us the number of repetitions of periods per unit of t (usually the unit of t is
second), so
1 1
frequency = = .
period T
1.1 Basic concepts MATH.APP.240 F OURIER METHODS / M e / 2020 8

1
For example, the period of sine is 2π, so it has the frequency 2π ≈ 0.159 (not necessarily an
integer). This means that if t increases by one, then sin(t) has repeated approximately one
sixteenth 16% of its period. The common unit 1/s of the frequency is called hertz (Hz). If we
allow negative periods, then the frequency can be negative as well. In such a case, the absolute
value of the frequency gives the number of repetitions of the period.
The term circular frequency is often referred in engineering literature. It is defined by

2π
circular frequency = 2π · frequency = .
T
The circular frequency of sine is 1. The unit of the circular frequency is rad/s. Sometimes we
omit the word "circular" and just talk about frequency, assuming that it is clear from the context
which kind of frequency we are talking about.

♠Example 1.4. The function f : f t = sin ωt + ϕ has frequency
1 1 ω
= 2π =
T ω
2π

and circular frequency

2π 2π
= 2π = ω.
T ω
♠

Definition 1.1. One-sided limits of a function f of real variable t are those limits where
t approaches a specified point either from left/below or from right/above.
The left-hand limit of a function f at t = a is the limit of f at a as t approaches a from
below, i.e. t < a. It is denoted by

f a− = lim− f t = lim+ f a − h .




t→a h→0

Similarly, the right-hand limit at t = a is the limit of f at t = a as t approaches a from

above, i.e. t > a. It is denoted by

f a+ = lim+ f t = lim+ f a + h .




t→a h→0

♠Example 1.5. Consider the function f :


t  1, if t > 0,


 , if t 6= 0,

f t = |t| = 0, if t = 0,
0, if t = 0
 
−1, if t < 0


is discontinuous at t = 0. It has finite one-sided limits

f 0+ = 1 ja f 0− = −1.

This function is called the signum function (or sign function ) and it extracts the sign of a real
number. The signum function is often denoted by sgn(·). ♠
1.1 Basic concepts MATH.APP.240 F OURIER METHODS / M e / 2020 9

 
Definition 1.2.
 A function f , defined on the interval a, b , is piecewise continuous on
the interval a, b , if f (a+ ), f (b− ) exist as finite values and f (c+ ) = f (c− ) = f (c) if
c ∈ (a, b) except at a finite number of points. The points where the left- or right-hand
limit is not equal to the value of the function are called points of discontinuities.

Remark. The definition allows discontinuity at the endpoints of the interval. The one sided
limits at the endpoints must exist when we approach the point inside the interval [a, b]. However,
f (a− ) and f (b+ ) are not required to exist.
Remark. The definition excludes those functions that are unbounded at the discontinuity, e.g.
f (t) = 1/t, or functions that don’t have one-sided limit at a point due to oscillation, e.g.
sin(1/t).

Definition 1.3. If the nth derivative f (n) of f exists and is continuous then we say that
the function is n times continuously differentiable.

 
♠Example 1.6. The function of Figure 2 is defined on the interval a, b and has one point of
discontinuity at t = c. The function has one sided limits at t = c (the loops of the red lines).
The function is not continuous, but it is piecewise continuous. Because of the discontinuity, the
function cannot be differentiable at t = c.

b t t
a c a c b

Figure 2. Discontinuous and non- Figure 3. Continuous and non-

differentiable. differentiable.
 

of Figure 3 is defined on the interval a, b and is continuous. The one sided limits
The function


f t −f c
of t−c
at t = c are not equal (compare to the slopes of the red lines in Figure 2), so the
function cannot be differentiable there. ♠

Theorem
  1.1. Let f be a periodic function with period
 T that
 is piecewise continuous on
0, T . Then f is differentiable over the interval d, d + T and
Z d+T

Z T

f t dt = f t dt.
d 0

Proof. The existence of the integral is clear, since we can divide the interval at points of dis-
continuity into subintervals where the function is continuous and then do piecewise integration.
Remember that discontinuity at end point of the interval of integration does not affect the value
1.1 Basic concepts MATH.APP.240 F OURIER METHODS / M e / 2020 10

of the integral and every continuous functious function is integrable over a finite interval.

Let us split the integral into three integrals

Z d+T Z 0 Z T Z d+T





f t dt = f t dt + f t dt + f t dt.
d d 0 T

Substitute x = t − T , to the rightmost integral to observe that

Z d+T Z 0 Z T Z d





f t dt = f t dt + f t dt + f x + T dx
d d 0 0

Z T

Z d

Z d

= f t dt − f t dt + f x dx
0 0 0

Z T

= f t dt.
0

The above theorem shows that we can choose the interval of integration freely provided that the
length of the interval is the length of the period of a periodic function. Choosing the interval of
integration wisely can make the calculations easier (in particular, with even and odd functions).

1.1.1 Even and odd functions

Definition 1.4. A function f is even if for every t in its domain

f −t =f t .

A function g is odd if for every t in its domain

g − t = −g t .

Remark. The domain of an odd or even function needs to be symmetric with respect to zero.
The graph of an even function is symmetric with respect to the line t = 0 (Figure 4). The graph
of an odd function is symmetric with respect to the origin (Figure 5)

t t0 t
t0 t0 t0

Figure 4. Even function. Figure 5. Odd function.

1.1 Basic concepts MATH.APP.240 F OURIER METHODS / M e / 2020 11

♠Example 1.7. Examples of even functions are

cos t , t2 , t4 + t2 + 1.

Examples of odd functions are

sin t , t, t3 + t.




Most of the functions are neither even nor odd, for example f : f t = t3 + 1. ♠

Remark. Even and odd functions have some nice properties. It can be shown for example that
the product of two even functions is an even function (the first line in the table below). The
following rules for odd and even functions apply. :

Even · Even = Even,

Odd · Odd = Even,

Odd · Even = Odd,

Even + Even = Even,

Odd + Odd = Odd.

♠Example 1.8. We verify the the rule on the third line in the above table as an example. If f is
even and g is odd, then their product is








f − t g − t = f t · − g t = −f t g t ,

showing that the product is odd by the definition of odd function. ♠

In addition to the above product and sum rules, differentiation changes the parity of a function:

f t = f − t ⇒ f 0 t = f 0 − t · − 1 = −f 0 − t .

Integrals of even and odd function over a symmetric interval [−a, a] where a > 0 have the
following useful properties:
For an even function Even(·) we have
Z a Z 0 Z a




Even t dt = Even t dt + Even t dt
−a −a 0

Z 0


Z a

= Even − y − dy + Even t dt
a 0
Z a

Z a

= Even y dy + Even t dt
0 0
Z a

=2 Even t dt
0
1.1 Basic concepts MATH.APP.240 F OURIER METHODS / M e / 2020 12

and for an odd function Odd(·) we have

Z a Z 0 Z a




Odd t dt = Odd t dt + Odd t dt
−a −a 0

Z 0


Z a

= Odd − y − dy + Odd t dt
a 0
Z a

Z a

= −Odd y dy + Odd t dt
0 0

= 0.

♠Example 1.9. Cosine is an even function, sine is an odd function, and the interval [−1, 1] is
symmetric, so

1
1 1


sin 2t
Z Z






cos 2t dt = 2 cos 2t dt = 2 = sin 2 − sin 0 = sin 2
−1 0 0 2
and Z
1

sin 2t dt = 0. ♠
−1

1.1.2 Heaviside step function

Certain discontinuous functions appear frequently in engineering applications. English electri-

cal engineer Oliver Heaviside (1850-1925) defined a function that can be used to force dis-
continuities to a function. The function is called the Heaviside step function or the unit step
function. In what follows, we denote this function by H (in Maple its name is "Heaviside" and
in Matlab "heaviside" ).

Heaviside function obtains H(t)

1
only two values
(

0, t < 0
H t = t
1, t > 0.

Figure 6. Heaviside step function has a step

discontinuity at t = 0.


Remark. In the above definition we did not define the value of H at 0. We can define H 0
to be zero, one, or something else (for example heaviside(0) is 1/2 in Matlab). From theoretical
point of view, the value at zero does not matter, but we do not want undefined function values
when using computers. However, the crucial feature of the function is the step discontinuity at
t = 0.
1.1 Basic concepts MATH.APP.240 F OURIER METHODS / M e / 2020 13

The step discontinuity moves to an-

H(t a)
other location by defining 1
(

0, t < a
H t−a =
1, t > a. t
a b

Figure 7. The Heaviside step function with

jump discontinuity at t = a.

By combining Heaviside step func-

tions, we can obtain other interest- H(t a) H(t b)
1
ing functions. For example,
(


1, a < t < b
H t−a −H t−b = t
0, otherwise. a b

Figure 8. H t − a − H t − b
(compare with the previous figure ).

The practical value of Heaviside step function is in that it allows us to write down certain
functions in one line without defining them piecewise. For example,
(
0, t<0
f (t) = H(t)e−t = −t
e , t > 0.

Rectangular pulse function

( 1

1, |t| < T
rectT t =
0, |t| > T.

Clearly t
T T




rectT t =H t+T −H t−T .


Figure 9. The rectangular pulse function rectT t .
1.1 Basic concepts MATH.APP.240 F OURIER METHODS / M e / 2020 14

1.1.3 Dirac delta function

In some applications, it is useful to model

A things that happen instantaneously as an ide-
alization of phenomena that happens in a very
t short time window. A mathematical equivalent
a of such a phenomena is an impulse function,
that is localized at a certain point. Often this
localized impulse is described as an upright
Figure 10. An impulse function. arrow, see Figure 10.

A sound theoretical theory for impulse functions can be developed, but we only present the
underlying ideas and some of the related properties that we need. The impulse function is not
a function in the sense we are used to. In particular, it is not the function that obtains value
A at t = a and is zero otherwise. It is the generalized function, that is defined by using the
following pulse function

 A/T

 0, t < a − T /2


ϕ t = A/T, a − T /2 ≤ t < a + T /2 a t


t ≥ a + T /2.

0, a–T / 2 a+T / 2

Figure 11. A pulse function.

The height of the pulse is A/T and its duration is T time units, so the area between the graph
of the pulse and the t axis is A, i.e.
Z ∞ Z a+T /2

A
ϕ t dt = dt = A.
−∞ a−T /2 T

Next we let the pulse duration T approach zero, but we require that the area between the graph
and the t axis remains constant A. As the non-zero part of the graph gets narrower the height
A/T must increase. Mathematically
Z ∞
A
lim = ∞ and lim ϕ(t) dt = A.
T →0+ T T →0+ −∞

The limit process leads us to the impulse function having value A (in the generalized sense) at
t = a.

The impulse function having value one is called the unit impulse function or Dirac delta func-
tion and is denoted by δ(t − a). Paul Dirac (1902-1984) was a British theoretical physicist and
developer of quantum physics. The Dirac delta function have many physical applications. For
example, it is useful in modeling point masses and point charges (or any quantity that is local-
ized at one point either in time or in space).
1.1 Basic concepts MATH.APP.240 F OURIER METHODS / M e / 2020 15

The Dirac delta function has the following properties:

1. δ t − a = 0 whenever t 6= a,

Z ∞

2. δ t − a dt = 1,
−∞

1

3. δ at = δ t .
|a|

Assuming that f is continuous at t = a, then

Z ∞




4. f t δ t − a dt = f a ,
−∞

which is called the sifting property since the impulse function sifts through the function
and pulls out the value f (a).

♠Example 1.10. According to the sifting property

Z π

 π π  1
cos t δ t − dt = cos = .
0 3 3 2
♠

The impulse function

with value A localized at a can be presented using the Dirac delta func-
tion: Aδ t − a .

Remark. We can think that

f t δ t−a =f a δ t−a ,

where f is assumed to be continuous at t = a. This is justified because

1. f t δ t − a = 0 = f a δ t − a if t 6= a,

Z ∞



Z ∞


2. f t δ t − a dt = f a = f a δ t − a dt.
−∞ −∞

It is important that we understand that the impulse functions are not functions in the traditional
sense. However, if used with care, they have practical implications that cannot be expressed
otherwise. We cannot express impulse functions geometrically and we just have to accept it as
the result of the following limiting process.

According to the definition of the Dirac delta function

Z t (

0, t < 0
δ u du =
−∞ 1, t > 0.
1.1 Basic concepts MATH.APP.240 F OURIER METHODS / M e / 2020 16

We cannot define the value of the integral precisely at t = 0. We need to know if the location
of the impulse belongs to the interval of integration or not which we need to agree on.
On the other hand,
H 0 t = 0, t 6= 0.

Although, the functions δ and H are not continuous and δ is not even a function in the common
sense, the above results hint that we can think that H 0 (t) = δ(t). We now state the following
results without proof:
Z t

d


δ u du = H t and H t =δ t .
−∞ dt

Proving the results require theory that is out of the scope of this text. The results are proved in
James’ book Advanced Modern Engineering Mathematics. At this point we just note that the
problems stem from the fact that the function H is not differentiable t = 0 because of disconti-
nuity and that δ is not a function in the traditional sense.

♠Example 1.11. Consider

f t = H t 3e−2t .

The one-sided limits at t = 0 are f 0− = 0 ja f 0+ = 3. The generalized derivative is by the

usual product rule of differentiation (we omit the detailed justifications)

f 0 t = −H t 6e−2t + δ t 3e−2t = −H t 6e−2t + 3δ t .

The term 3δ

−2t t "corresponds to the jump of f from zero to three at the origin". The term
−H t 6e where t 6= 0, is the usual piecewise defined derivative. ♠

1.1.4 Scaling- and shifting operations

We end this chapter by recalling some simple operations on functions/signals. We use the sine
function as an example, but the principle generalizes easily to any periodic function.
Consider the sine function f (t) = sin(t) with amplitude 1 and period 2π:

t
1

2π



Figure 12. The sine function y = f t = sin t .

Amplitude scaling can be expressed mathematically by multiplying f (t) by a scaling factor

A > 0, i.e. y(t) = Af (t):
1.1 Basic concepts MATH.APP.240 F OURIER METHODS / M e / 2020 17

2
2π


Figure 13. Doubling the amplitude (stretching in the y-direction): y(t) = 2f t .

Time shifting of the function/signal by S > 0 units to the left is expressed by adding S to t, i.e.
y(t) = f (t + S):

2
2π


Figure 14. Shifting the function two units to the left: y = f t + 2 .

Why to the left? At t = −2 we obtain the value of the original function at t = 0. It is easy to
deduce that adding a negative number shifts the function to the right.
Time scaling of the function/signal can be expressed by multiplying t by a scaling factor B, i.e.
y(t) = f (Bt):

2π / 3
t

2π


Figure 15. Time scaling the function (shrinking in the t-direction) to one third: y = f 3t .

We can combine the above operations:

1
2π / 3
t

2π


Figure 16. Shifting the above time scaled signal one unit to the right: y = f 3t − 3 .
1.1 Basic concepts MATH.APP.240 F OURIER METHODS / M e / 2020 18

In the last case where we have combined shifting and time scaling operations, there is a risk of
confusion. We have




f 3t − 3 = f 3 t − 1 = f 3u ,
where u = t − 1 or alternatively t = u + 1, i.e. the y-axis at u = 0 in (u, y)-coordinates locates
at t = 1 in the (t, y)-coordinates. This means that the scaled function is sifted by one unit to the
right. In particular, the shift is not by three units!
2 Fourier series MATH.APP.240 F OURIER METHODS / M e / 2020 19

2 Fourier series

2.1 Fourier theorem

We mentioned in the introduction that the Fourier series is a decomposition of a periodic func-
tion f with period T = 2π ω
into a (typically infinite) sum of oscillating terms. Fourier theorem
states that every reasonably continuous periodic function can be written in the form
∞
X

A0 + An sin nωt + φn ,
n=1

the terms An and φn are constants for every index n = 0, 1, 2, 3, . . .. The term A1 sin ωt+
where
φ1 is called first harmonic. Its (circular) frequency is
2π
ω= .
T
In general,

An sin nωt + φn
is called nth harmonic having the frequency nω. In addition, we call the constant term A0
zeroth harmonic. The constant An is the amplitude and φn the phase angle of nth harmonic
measuring the lag/shift of pure sine wave with the same frequency. According to the sum
formula for sine we have






An sin nωt + φn = An cos φn sin nωt + An sin φn cos nωt .

By denoting (

an = An sin φn


bn = An cos φn .
and setting A0 = a20 (the reason for this will be explained later) we can represent the series
decomposition in the form
∞  
a0 X


+ an cos nωt + bn sin nωt .
2 n=1

We can find the original sum decomposition when an and bn are given since the amplitude of
the nth harmonic is p
An = a2n + b2n
and the phase angle  
an
φn = arctan
bn
where one should be careful when choosing the quadrant.
Now we need to answer two mathematically relevant questions. First, when does the series exist,
or in other words when does the series converge? Secondly, how can we find the coefficients
an and bn ? We postpone answering the second questions to Section 2.4 (we want to present the
exponential form of the Fourier series and some lemmas before that). We discus the convergence
of Fourier series briefly in this section. More detailed discussion is postponed to Section 2.7.
Before proceeding to the next section we want to illustrate possible gains using Fourier series
by an example.
2.1 Fourier theorem MATH.APP.240 F OURIER METHODS / M e / 2020 20

♠Example 2.1. The location x of a mass m of a mass-spring system with spring constant k can
be modeled as
mx00 t + cx0 t + kx t = f t .

In the differential equation c is a positive constant and f represents an external force. Provided
that c2 − 4mk < 0, the homogeneous differential equation (f(t)=0) has solution


x(t) = eαt C1 cos βt + C2 sin βt ,

where √
c 4mk − c2
α=− and β= .
2m 2m
Assume that the external force f (t) is periodic and has the Fourier series
∞  
a0 X


f (t) = + an cos nωt + bn sin nωt .
2 n=1

By the principle of super position, we can solve the differential equation for each harmonic
component fn (t) = an cos(nωt) + bn sin(nωt) separately, i.e. we solve

mx00 t + cx0 t + kx t = fn t .

A particular solution of the original differential equation is then the sum of these separate so-
lutions which are harmonic oscillators. As we remember from the earlier courses, the general
solution is then the general solution of the homogeneous differential equation added to the par-
ticular solution which consists of harmonic oscillators. The harmonic oscillators with large
amplitudes dominate the solution. For harmonic components fn (t) on certain frequency range
the amplitude of the solution is increases due to resonance.
In order to make the above discussion more formal, consider the system with c = 1, m = 1,
k = 6, and with two separate forcing functions f0 (s) = 2 and f1 (s) = 2 sin(3t). We
can think that both of the forcing functions have the same amplitude. However, as the time
passes
√ the solution with f0 approaches the value 1/3 while the solution with f1 approaches
2/3 cos(3t − 5π/4). We observe that the amplitude in the latter case is larger. In the non-
realistic case where we do not have damping, i.e. c = 0 the resonance can actually increase the
amplitude of the solution unboundedly. ♠

As in the case of Taylor series, we can formally write down the Fourier series of a periodic
function. However, the function and its series representation correspond to each other only if
the series converges. By definition, the series converges if and only if the sequence of its partial
sums converge. The partial sum

N  

a0 X


SN t = + an cos nωt + bn sin nωt
2 n=1

of Fourier series is a function (variable t), so we need to study if the sequence of functions
converges to a limit function S, i.e. if



lim SN t = S t .
N →∞
2.1 Fourier theorem MATH.APP.240 F OURIER METHODS / M e / 2020 21

A natural question that arises is that what kind of function f has a convergent Fourier series. If
the Fourier series converges there might not be a one to one correspondence between f and its
Fourier series, so we want to know what is the limit function S and how does it correspond to
f.

The functions of practical applications are typically well behaving in the sense that the above
theoretical questions are irrelevant, i.e. the Fourier series converges to the function of interest.
The following definition gives a class that contain most of the practically interesting functions
having convergent Fourier series. The conditions posed are only sufficient for convergence, so
there are functions outside the defined class that have convergent Fourier series.

Definition

2.1. A periodic function satisfies the Dirichlet conditions if
a
it is piecewise continuous,
b every bounded interval can be partitioned into finitely many subintervals on each
of which the function is monotone, i.e. increasing or decreasing.

Remark. For most functions the condition b can be replace by

b’ it has a finite number of local extrema over any bounded interval.

The condition b implies b’
, but not vise

versa. For example, the piecewise constant func-
tions satisfy the condition b , but not b’ .
Note that the assumption of piecewise continuity contains assumption of boundedness. The
class of functions satisfying the Dirichlet conditions exclude mathematically problematic func-
tions but is large enough to contain most of the practically interesting ones. The conditions are
named after German mathematician Johann Peter Gustav Lejeune Dirichlet (1805-1859 ) who
studied number theory, theory of Fourier series, and other topics in mathematical analysis. He
showed proved the following theorem (known as Dirichlet’s theorem (there are other theorem’s
that have the same name))

Theorem 2.1. If T-periodic function f satisfies the Dirichlet conditions, then it has the
(convergent) Fourier series
∞  

a0 X
ˆ



f t = + an cos nωt + bn sin nωt ,
2 n=1

2π
where ω = T
and
Z d+T Z d+T
2

2


an = f t cos nωt dt and bn = f t sin nωt dt.
T d T d

The above series is called the trigonometric form of Fourier series . The coefficients an and
bn have the above form whenever the Fourier series exists, not just for functions satisfying
Dirichlet conditions. We postpone showing this later to Section 2.4, since the proof is easy if
we have the exponential form of Fourier series at our disposal.
2.1 Fourier theorem MATH.APP.240 F OURIER METHODS / M e / 2020 22

The limit function fˆ is not always f . Because of this we make a distinction between the limit
function fˆ and the given periodic function f . The functions f and fˆ are equal almost every-
where:

Fixing the value t = t0 , we can show that the series converges (pointwise) to the average of
left and right hand limits of f . In particular, if the function is continuous, then the limit is f (t0 ).
We state this result as a theorem:

Theorem 2.2. If T -periodic function f satisfies the Dirichlet conditions and has the
Fourier series ∞  

a0 X
fˆ t =



+ an cos nωt + bn sin nωt ,
2 n=1
2π
where ω = T
, then



 f t0 , if f is continuous at t = t0


ˆ
f t0 = f t−
+ f t+

0 0


 2 , if f is discontinuous at t = t0 .

In practice, we only use a finite number of terms, i.e. we approximate the function by the partial
sum of Fourier series. In order to understand the precision of this approximation we need to
understand something about the convergence. We return to this in Section 2.7.

There are convergent series that formally look like Fourier series, but are not a Fourier series of
any function. Such a series is given in the following example.

♠Example 2.2. The series

∞
X sin(kt)
S(t) =
k=2
ln(k)
has the form of Fourier series. The series is pointwise convergent and defines a function, but
there exist no function having such a Fourier series. By using the Parseval’s theorem (Section
2.12), we can show that (this is not difficult but we omit the details) if
∞  
a0 X


+ an cos nωt + bn sin nωt
2 n=1

is a Fourier series, then the series

∞
X bk
k=1
k
1
converges. In the given series bk = ln(k)
. According to the integral test,
∞
X 1
k=2
k ln(k)

diverges, so S(t) cannot be a Fourier series. ♠

2.2 Exponential Fourier series MATH.APP.240 F OURIER METHODS / M e / 2020 23

2.2 Exponential Fourier series

The exponential form of Fourier series is a compact way of writing the Fourier series using
complex exponential function. It makes many otherwise tedious task easier, e.g. see Section
2.4. The exponential form follows by representing sine and cosine functions using complex
exponential (see Appendix B):
∞  

a0 X
ˆ



f t = + an cos nωt + bn sin nωt
2 n=1

∞ 
ejnωt + e−jnωt ejnωt − e−jnωt

a0 X
= + an + bn
2 n=1
2 2j

∞    
a0 X an − jbn jnωt an + jbn −jnωt 1
= + e + e note that = −j .
2 n=1
2 2 j

Denote
a0
c0 =
2
an − jbn
cn = , n = 1, 2, 3, . . .
2
an + jbn
c−n = = c∗n , n = 1, 2, 3, . . .
2

where c∗n is the complex conjugate of cn (the sign of the imaginary part is changed). We have
found the exponential form
∞ 
X  ∞
X
fˆ t = c0 + jnωt −jnωt
cn ejnωt .


cn e + c−n e =
n=1 n=−∞

♠Example 2.3. It is relatively easy to find the trigonometric form of the Fourier series starting
from the exponential form:
∞ ∞   X ∞ ∞
X −j jnt X −j jnt 2 

 X 2

e = 2Re e = Re − j cos nt + sin nt = sin nt
n=−∞
n n=1
n n=1
n n=1
n
n6=0

or
∞ ∞   X ∞ ∞
X −j jnt X −j jnt −j −jnt −j jnt −jnt

X 2
e = e + e = e −e = sin(nt)
n=−∞
n n=1
n −n n=1
n | {z
} n=1 n
n6=0 =2j sin nt

−j
or alternatively, by observing that ω = 1 and cn = n
, we see that an = 0 and bn = n2 . ♠
2.3 Some Integrals MATH.APP.240 F OURIER METHODS / M e / 2020 24

2.3 Some Integrals

Next we present some integrals that help us when considering Fourier series. We can express
many of the formulas in a compact form by using Kronecker delta ( Leopold Kronecker, (1823
- 1891) ).

Definition 2.2. Kronecker delta δk−j is a symbol (function), having the value 0 or 1 de-
pending on the integers k and j
(
1, if k = j,
δk−j =
0, if k 6= j.

The notation δk,j is often used instead of δk−j . Then δk,j = 1 is k = j and 0 otherwise. We
adopt the above definition, since it allows us to write the Kronecker delta using one integer. In
particular, δk = δk−0 = δk,0 .
♠Example 2.4. We have
∞
X 1
0 1 0 0 1 0 0 1 1
k
δk + δk−3 = −1 + 0 + 1 + 2 + 3 + 4 + · · · + k + · · · = 0 + 3 .
k=−1
2 2 2 2 2 2 2 2 2 2

Both Kronecker deltas picks up exactly one term from the sum. ♠

In the following formulas we assume the period to be

2π
T = .
ω
We assume that m and n are positive integers unless otherwise stated. Proving the formulas
is a simple task, so you should not try to memorize them, but try to understand how we use
periodicity and parity of the functions.

Z d+T

1. cos nωt dt = T δn .
d

Z d+T

2. sin nωt dt = 0.
d

Z d+T

1
3. cos nωt cos mωt dt = T δn−m .
d 2

Z d+T

1
4. sin nωt sin mωt dt = T δn−m .
d 2
2.3 Some Integrals MATH.APP.240 F OURIER METHODS / M e / 2020 25

Z d+T


5. sin nωt cos mωt dt = 0.
d

Let us prove the formulas.

1. We have two cases.

a) If n = 0, then
Z d+T Z d+T T


cos nωt dt = dt = t = T.
d d 0
b) If n 6= 0, then

T
d+T




sin nωt sin n2π sin 0
Z


cos nωt dt = = − = 0.
d 0 nω nω nω

2. Since sine is odd, we have

Z d+T

Z T /2

sin nωt dt = sin nωt dt = 0.
d −T /2 | {z }
odd

3. Assume that n 6= 0 and m 6= 0. We have two cases.

a) If n = m, then
Z d+T T
1 T
Z Z  
2





cos nωt cos mωt dt = cos nωt dt = 1 + cos 2nωt dt
d 0 2 0
T

!
1 sin 2nωt 1
= t+ = T.
2 0 2nω 2
b) If n 6= m, then
Z d+T Z T  


1


cos nωt cos mωt dt = cos (n − m)ωt + cos (n + m)ωt dt
d 0 2
T


!
1 sin (n − m)ωt sin (n + m)ωt
=
+
= 0.
2 0 n−m ω n+m ω

4. Assume that n 6= 0 and m 6= 0. This is similar to item 3. We have two cases.

a) If n = m, then
Z d+T Z T
1
sin2 nωt dt = · · · = T.




sin nωt sin mωt dt =
d 0 2
b) If n 6= m, then
Z d+T Z T  


1


sin nωt sin mωt dt = cos (n − m)ωt − cos (n + m)ωt dt = · · · = 0.
d 0 2
2.3 Some Integrals MATH.APP.240 F OURIER METHODS / M e / 2020 26

5. By using the parity properties of sine, cosine and their product we have
Z d+T


Z T /2


sin nωt cos mωt dt = sin nωt cos mωt dt = 0.
d −T /2 | {z }
odd

Recall the integration by parts formula

Z b b Z b
0
f t g 0 t dt.







f t g t dt = f t g t −
a a a

♠Example 2.5. Let a 6= 0 be a constant. Then

Z



sin at sin at sin at cos at
Z


t cos at dt = t − dt = t + + C.
a a a a2

Note that it is important to choose the functions f 0 and g in a sensible way. There is no simple
rule for this, and doing the right choice is learned by practice. Wrong choice usually makes
the integral more difficult. Here we reduced the order of the polynomial function by choosing
g = t, so that the differentiation reduced the order to zero. If you are not sure if the calculations
were correct, you can check the result by differentiation:


!



d sin at cos at sin at cos at a sin at a

t + 2
+C = +t − 2
= t cos at .
dt a a a a a

In what follows we need to know how to integrate complex valued functions of real variable.
We present the required results next. You can learn more about integration of complex function
on the course MATH.APP.440 Complex functions.
Integration of complex valued functions of real variable can be reduced to integration of two
real integrals. Every complex valued function z of real variable t can be written in the form




z t = u t + jv t , t ∈ R,

where u and v are real functions. Then we define the integral of z with respect to real variable
t as Z Z Z




z t dt = u t dt + j v t dt.
2π
In particular, if m ∈ Z and T = ω
Z d+T Z d+T Z d+T
jmωt



e dt = cos mωt dt + j sin mωt dt = T δm + 0 = T δm .
d d d

Integration (as well as differentiation) can be carried out if we consider j as any constant (which
it is for complex functions), i.e.
2.4 The coefficients of the Fourier series MATH.APP.240 F OURIER METHODS / M e / 2020 27


d+T

t = T, if m = 0,

d+T
Z 

jmωt d
e dt = d+T  
d 1 ejmωd
ejmωt ejm2π − 1 = 0,



 jmω
= jmω
if m 6= 0,
d

or in a simplified form

Z d+T
ejmωt dt = T δm , m ∈ Z. (1)
d

2.4 The coefficients of the Fourier series

Assume that ∞
X
fˆ(t) = ck ejkωt ,
k=−∞

where the function f is T -periodic and ω = 2π

T
. We multiply the formula by e−jnωt , to obtain
∞
X ∞
X
fˆ(t)e−jnωt = ck ejkωt e−jnωt = ck ej(k−n)ωt .
k=−∞ k=−∞

Integration with respect to t over one period yields

Z d+T ∞ Z d+T ∞
−jnωt
X (1) X
f (t)e dt = ck ej(k−n)ωt dt = ck T δk−n = T cn .
d k=−∞ d k=−∞

So Z d+T
1
cn = f (t)e−jnωt dt.
T d

This shows that

∞ 
X  ∞
X
ˆ jnωt −jnωt
cn ejnωt ,


f t = c0 + cn e + c−n e =
n=1 n=−∞

where Z d+T
1
f t e−jnωt dt,


cn = n = 0, ±1, ±2, ±3, . . .
T d

an − jbn
By finding the real and imaginary parts of the coefficient cn = , n ≥ 1, we obtain
2
d+T d+T
2 d+T
Z Z Z
2 2
f t e−jnωt dt =






2cn = f t cos nωt dt − j f t sin nωt dt
T d T d T d
2.5 Examples MATH.APP.240 F OURIER METHODS / M e / 2020 28

and

∞ 

a0 X

ˆ


f t = + an cos nωt + bn sin nωt ,
2 n=1

where Z d+T
2


an = f t cos nωt dt, n = 0, 1, 2, 3, . . .
T d
Z d+T
2


bn = f t sin nωt dt, n = 1, 2, 3, . . .
T d

Both forms are equally important and you can choose the one that is most suitable for your task.
You should handle both of them and know how you can find the exponential by starting from
the trigonometric form and vice versa.

2.5 Examples

♠Example 2.6. Let us find the Fourier series of 2π-periodic function f , defined by

f t = t2
+ t,
if t < π,

f t + 2π = f t .

t
−5π −3π −π π 3π 5π

 
Figure 17. Graph of f on the interval − 5π, 5π .

2π
The period is T = 2π and the frequency ω = T
= 1. We substitute the values to the formulas
to obtain the coefficients:

π π
2π 3 2π 2
Z  
2 2

1 1 3 1 2
a0 = t + t dt = t + t = = .
2π −π π −π 3 2 3π 3

Finding the remaining coefficient is a bit more laborious, since we need to integrate by parts.
We have (n ≥ 1)
2.5 Examples MATH.APP.240 F OURIER METHODS / M e / 2020 29

Z π
1
t2 + t cos nt dt



an =
π −π
Z π  Z π
1 2



= t cos nt dt +
t cos nt dt
π −π | −π |
{z } {z }
even odd
π 2
2 π 2
Z 

2 t
2t
2

= t cos nt dt = sin nt + 2 cos nt − 3 sin nt
π 0 π 0 n n n
4
n
= − 1 ,
n2
and
1 π 2
Z



bn = t + t sin nt dt
π −π
Z π Z π 
1 2



= t sin nt dt + t sin nt dt
π −π | {z } −π | {z }
odd even
Z π π 
2
2 t
1

= t sin nt dt = − cos nt + 2 sin nt
π 0 π 0 n n
2
n
= − −1 .
n
The Fourier series is
∞

π2 X
 
ˆ 4
n
2
n

f t = + − 1 cos nt − − 1 sin nt .
3 n=1
n2 n

 
In Figure 18, the partial sum S10 of the above series is plotted on the interval − 15, 15 . In
Figure 19 we have added the graph of the original function.

f(t)
S10(t)

t t
−5π −3π −π π 3π 5π −5π −3π −π π 3π 5π

Figure 18. Partial sum S10 . Figure 19. The function and the partial sum S10 .

According to Figure 19, the partial sum approximates the function relatively well at the points
where

the function is continuous. The function f has jump discontinuity at t = t0 = 2k +
1 π, k = 0, ±1, ±2, . . . The one-sided limits are

f t− 2
f t+ 2



0 = π + π ≈ 13.01, 0 = π − π ≈ 6.73.

The average value π 2 is approximately 9.87. Let us examine the behavior or the partial sums
 discontinuity at t = π. We plot the function and several partial sums on the
near the jump
interval 2, 4 . In Figures 20-23 we have marked the point (π, a), a ≈ 9.87 by a cross. We
observe that the partial sums approach the point as the number of terms increases.
2.5 Examples MATH.APP.240 F OURIER METHODS / M e / 2020 30

f(t) f(t)
S2(t) S5(t)
a a

3 t 3 t
2 π 4 2 π 4

Figure 20. Partial sum S2 . Figure 21. Partial sum S5 .

f(t) f(t)
S20(t) S100(t)
a a

3 t 3 t
2 π 4 2 π 4

Figure 22. Partial sum S20 . Figure 23. Partial sum S100 .

We can observe rapid oscillation near the jump discontinuity in Figure 23. This is somewhat un-
expected since the number of terms is rather large and consequently the approximation should
be very good. This is a general phenomenon that can be observed at any jump discontinuity. It
is called Gibbs phenomenon and is studied more closely in Section 2.7.2. ♠

♠Example 2.7. Let us find the Fourier series of 2π-periodic function f , that is defined by

f t = t2
,



if t < π,
f t + 2π = f t .

f(t)

t
−5π −3π −π π 3π 5π

 
Figure 24. Graph of f on the interval − 5π, 5π .

The function f is even and periodic. The graph has sharp edges at


t = t0 = 2k + 1 π, k = 0, ±1, ±2, . . .
2.5 Examples MATH.APP.240 F OURIER METHODS / M e / 2020 31

2π
The period is T = 2π and the frequency ω = T
= 1.

We find the coefficients

Z π π
2 1 1 3 2π 3 2π 2
a0 = t2 dt = t = = ,
2π −π π −π 3 3π 3

and for n ≥ 1 Z π Z π
1 2 2 4
n
t2 cos nt dt = · · · = 2 − 1



an = t cos nt dt =
π −π π 0 n
and Z π
1
t2 sin nt dt = 0.


bn =
π −π | {z }
odd

The Fourier series is ∞

π2 X 4
n
ˆ


f t = + − 1 cos nt .
3 n=1
n2
 
The graph of the partial sum S5 is plotted over the  − 15, 15 in Figure 25. Figures
 interval
26-28 illustrate the convergence over the period − 1, 4 . We observe that the partial sums
converge to the original function rapidly. Convergence is relatively rapid even near the sharp
edges, and the rate of convergence is greater than in the previous example.

f(t)
S1(t)

f(t)
S5(t)

t t
−15 15 −1 2 π 4

Figure 25. Partial sum S5 . Figure 26. Partial sum S1 .

f(t) f(t)
S3(t) S10(t)

t t
−1 2 π 4 −1 2 π 4

Figure 27. Partial sum S3 . Figure 28. Partial sum S10 .

♠
2.5 Examples MATH.APP.240 F OURIER METHODS / M e / 2020 32

♠Example 2.8. Consider the function of Example 2.6. Let us find the coefficients of the
trigonometric Fourier series by using the exponential form. According to the formula
1 d+T
Z
f t e−jnωt dt.


cn =
T d

The constant term was found as earlier and a0 = 2c0 . For n ≥ 1

Z π
1
t2 + t e−jnt dt


cn =
2π −π
!
π
1 t2 + t −jnt 2t + 1 −jnt 2 −jnt
= e −
2 e +
3 e
2π −π −jn − jn − jn

n 
−1

2π 4π
= j + 2
2π n n
2
n 1
n
= 2
−1 +j −1 ,
n n

which yields
4
n 2
n
an = − 1 and bn = − −1 .
n2 n
♠

♠Example 2.9. It is obvious that the Fourier series of sine or cosine function should contain
exactly one term which is the function itself. Let us verify this for cosine functions. Consider


f (t) = cos mωt ,
2π
where m ∈ N0 , ω = T
and T is the period.

The coefficient a0 Zis

2 T

a0 = cos mωt dt
T 0

T
 2 1

sin mωt , if m 6= 0


T
0 mω

= T
 2


T
 t, if m = 0
0

(
0, if m 6= 0
=
2, if m = 0.

Z n ≥ 1 are
The coefficients when
2 T


an = cos mωt cos nωt dt
T 0

1 T
Z  



= cos (m − n)ωt + cos (m + n)ωt dt
T 0
2.5 Examples MATH.APP.240 F OURIER METHODS / M e / 2020 33


T  
 1 1
1

sin (m − n)ωt + sin (m + n)ωt , if m 6= n


T (m − n)ω (m + n)ω

0
= T  
 1 1



T t+ sin 2mωt , if m = n
2mω

0

(
0, if m 6= n
=
1, if m = n,

and Z T /2
2


bn = cos mωt sin nωt dt
T −T /2 | {z }
odd
= 0.

Thus, the Fourier series is

∞  

a0 X
ˆ



f t = + an cos nωt + bn sin nωt
2 n=1


= cos mωt .

Show the result for sine functions as an example. ♠

√

♠Example √
2.10. Continuing the previous example. If f (t) = cos 2t , we choose m = 3
and ω = 32 . Then T = 2πω
=√ 6π
2
. There is only one non-zero coefficient a3 = 1 in the Fourier
series, and √ !
2
fˆ t = 1 · cos 3 ·



t =f t .
3

√ the index m, as long as we choose the frequency accordingly,

It does not matter how we choose
i.e. we need to have mω = 2. Then the index of the non-zero coefficient just changes corre-
spondingly. ♠

2.5.1 Finding values of infinite sums using Fourier series

The Fourier series is known to converge pointwise. If we can evaluate the value of the function
and find its Fourier series, then we can deduce the values of the related infinite sums at a point.

♠Example 2.11. In Example 2.7 we found the Fourier series

∞
n
1 2 π2 X − 1

t − = 2
cos nt , |t| ≤ π.
4 12 n=1 n
2.6 Linearity MATH.APP.240 F OURIER METHODS / M e / 2020 34

The function is continuous, so we
have pointwise convergence everywhere. If we choose the
value t ∈ [−π, π] so that cos nt can be simplified, we can find values for certain series. For
example,
∞
X 1 1 2 π2 π2

= π − = substitution t = π ,
n=1
n2 4 12 6
∞
n
X −1 π2

− = substitution t = 0 .
n=1
n2 12
♠

2.6 Linearity

Theorem 2.3. Let f be a T -periodic function that is a linear combination of two T -

periodic functions g and h having Fourier series, i.e.




f t = ag t + bh t ,

where a and b are constants. Then

fˆ t = aĝ t + bĥ t .

Proof. Denote
∞
X
cn ejnωt


ĝ t =
n=−∞

and ∞
X
γn ejnωt .


ĥ t =
n=−∞

We need to find the coefficients of f .

Z d+T Z d+T
1 1 

f t e−nωt dt = ag t + bh t e−nωt dt



Cn =
T d T d
Z d+T Z d+T
1 1
g t e−nωt dt + b h t e−nωt dt



= a
T d T d

= acn + bγn .

Thus,
∞
X ∞
X ∞
X
ˆ jnωt jnωt
γn ejnωt = aĝ t + bĥ t .




f t = Cn e =a cn e +b
n=−∞ n=−∞ n=−∞
2.6 Linearity MATH.APP.240 F OURIER METHODS / M e / 2020 35

♠Example 2.12. Square wave f is the function defined by

f t = signum

t
, if t < π,
f t + 2π = f t .
In the definition we have not defined the values at multiples of π where the function has jump
discontinuities. If needed, we can set an appropriate value.

1
π π t

 
Figure 29. Graph of square wave f over the interval − 3π, 3π .

The Fourier series of square wave is

∞

4X 1
ˆ


f t = sin (2n − 1)t .
π n=1 2n − 1

Let us scale the square wave according to Figure 30:

1

g t = 1 + signum t , if t < π,
2


g t + 2π = g t .

1
π π t

 
Figure 30. Graph of modified square wave g over the interval − 3π, 3π .

By linearity, the Fourier series of g is

∞

1 ˆ

 1 2 X 1

ĝ t = 1+f t = + sin (2n − 1)t .
2 2 π n=1 2n − 1

♠
♠Example 2.13. Let us verify that the function

1
1

f t =2+ cos 2t + sin 4t
10 100
is its own Fourier series. Since the frequency or period is not given readily let us find them first.
Since the constant function 2, cos(2t), and sin(4t) are π-periodic we can observe that

1
1


f t+π =2+ cos 2t + 2π + sin 4t + 4π = f t
10 100
2.6 Linearity MATH.APP.240 F OURIER METHODS / M e / 2020 36

and can choose period T = π (this happens to be the shortest period). The frequency is ω = 2
and

4 1
1

f t = + cos ωt + sin 2ωt ,
2 10 100
1 1
which shows that a0 = 4, a1 = 10 , and b2 = 100 . The rest of the coefficients an and bn are
zeros. We can also use the linearity together with observation of Example 2.9. ♠

2.6.1 Fourier series of scaled and shifted functions

It is trivial how the amplitude scaling affects the Fourier series. Next we study how the time
scaling affects the Fourier series. Consider a T -periodic function f and denote its frequency by
ω. Assume that it has the Fourier series
∞ ∞
ˆ a0 X X
cn ejnωt .


f (t) = + an cos(nωt) + bn sin(nωt) =
2 n=1 n=−∞

Let us define the function g as

g(t) = f (αt), α 6= 0.
The period of the function g is T 0 = Tα and the frequency ω 0 = αω. The properties of the
function remain the same apart from the period and the frequency, so g has Fourier series as
well. Denote the exponential Fourier series of g by
∞
0
X
c0n ejnω t .
n=−∞

Its coefficients are

Z T0 Z T /α
1 −jnω 0 t α
c0n = 0 g(t)e dt = f (αt)e−jnαωt dt
T 0 T 0

Let us substitute x = αt. Then dx = α dt, the lower limit is still 0, and the upper limit is T .
Thus,
α T 1 T
Z Z
0 −jnωx 1
cn = f (x)e dx = f (x)e−jnωx dx = cn .
T 0 α T 0
This shows that the coefficient do not change, and the time scaling just changes the frequency
components of Fourier series, i.e.
∞ ∞
0
X X
ĝ(t) = fˆ(αt) = jnωαt
cn e = cn ejnω t .
n=−∞ n=−∞

Let us now consider the time shift. Consider the function

g(t) = f (t + s),
i.e. we perform time shift to the right or to the left. Obviously, the period and the frequency do
not change in time shift, i.e. T 0 = T and ω 0 = ω. Let us find the exponential Fourier series.
∞
0
X
c0n ejnω t .
n=−∞
2.7 Convergence of Fourier series MATH.APP.240 F OURIER METHODS / M e / 2020 37

The coefficients are

Z T0 Z T
1 −jnω 0 t 1
c0n = 0 g(t)e dt = f (t + s)e−jnωt dt
T 0 T 0

We substitute x = t + s. Then dx = dt, the lower limit is s and the upper limit is T + s. Thus,

1 T +s
Z Z T +s
0 −jnω(x−s) jnωs 1
cn = f (x)e dx = e f (x)e−jnωx dx = ejnωs cn ,
T s T s
which implies
∞
X
ĝ(t) = ejnωs cn ejnωt .
n=−∞

By replacing t in the Fourier series of f by t + s gives the Fourier series of g if we separate the
term ejnωs from the frequency component. Then
∞
X ∞
X
fˆ(t + s) = cn ejnω(t+s) = ejnωs cn ejnωt .
n=−∞ n=−∞

Finding the trigonometric form is not as straightforward. Write

∞  
ˆ a0 X


f (t + s) = + an cos nω(t + s) + bn sin nω(t + s) .
2 n=1

By using the sine and cosine sum rules or observing that

a0n b0
c0n = −j n
2 2
and comparing this to the coefficient
 
cos(nωs)an + sin(nωs)bn cos(nωs)bn − sin(nωs)an
c0n jnωs
=e cn = −j
2 2

we find the trigonometric form

∞
a0 X 


ĝ(t) = + cos(nωs)an + sin(nωs)bn cos(nωt)+ cos(nωs)bn − sin(nωs)an sin nωt) .
2 n=1 | {z } | {z }
=a0n =b0n

This shows that the time shift changes the coefficient of the Fourier series.

2.7 Convergence of Fourier series

If a function f satisfies the Dirichlet conditions, then the coefficients of the Fourier series
∞  
a0 X


+ an cos nωt + bn sin nωt
2 n=1
2.7 Convergence of Fourier series MATH.APP.240 F OURIER METHODS / M e / 2020 38

have the limits Z T

2
lim an = lim f (t) cos(nωt) dt = 0
n→∞ n→∞ T 0
and Z T
2
lim bn = lim f (t) sin(nωt) dt = 0.
n→∞ n→∞ T 0

The precise proof is out of the scope of this text. Geometrical justification is due to the fact that
the rabid oscillation sin(nωt) (or cos(nωt)) as n increases causes the areas below and above the
graph of the function f (t) sin(nωt) (f (t) cos(nωt))to be approximately the same, see Figure 31.
t2cos( 40t )
t2
t2

t
−π π

 
Figure 31. Graphs of g : g(t) = t2 cos(40t) and g : g(t) = ±t2 over the interval − π, π .

This show that the terms of the series have limit

 



lim an cos nωt + bn sin nωt = 0,
n→∞

which shows that the convergence is possible. However, this is not enough to guarantee conver-
gence. We omit the actual proof of convergence. You can find many proofs for convergence in
the literature.
Note that convergence can be understood in many different sense, for example as pointwise,
uniform convergence(see Section 2.7.3), or in the sense of the function norm
Z T  p1
p
kf kp = |f (t)| dt .
0

Then we say that Sn → S as n → ∞ if kSn − Skp → 0 as n → ∞. The vector space of all

functions for which the above integral converges for fixed p ≥ 1 is denoted by Lp (0, T ). These
spaces have nice properties and are important for example in functional analysis and mathemat-
ical systems theory. This theory is out of the scope of this text. These spaces are considered for
example on the course MATH.MA.810 Introduction to Functional Analysis.

2.7.1 Rate of convergence

Above we have explained geometrically why the terms an and bn of the Fourier series converge
to zero. The accuracy of the approximation by a truncated Fourier series, i.e. by a partial sum,
is related to the rate of convergence. A rule of the thumb is that the smoother the function, the
2.7 Convergence of Fourier series MATH.APP.240 F OURIER METHODS / M e / 2020 39

faster the rate of convergence. This is formalized by the following theorem.

Theorem 2.4. Let f be a T -periodic function having the Fourier series

∞  
ˆ a0 X


f (t) = + an cos nωt + bn sin nωt .
2 n=1

(a) If the function f is piecewise continuous, but not continuous, then |an | or |bn | ap-
proach zero by the rate 1/n.

(b) If f is continuous, but not continuously differentiable, then |an | or |bn | approach zero
by the rate 1/n2 .

(c) If f is k ≥ 1 times continuously differentiable, but the k + 1 derivative is discontinu-

ous, then |an | or |bn | approach zero by the rate 1/nk+2 .

♠Example 2.14. The piecewise continuous function of Example 2.6 having discontinuities has
Fourier coefficients
an = 4(−1)n /n2 and bn = −2(−1)n /n
which is in accordance with the item (a) of Theorem 2.4. On the other hand, the continuous but
not continuously differentiable function of Example 2.7 has Fourier coefficients

an = 4(−1)n /n2 and bn = 0

which is in accordance with the item (b). ♠

2.7.2 Gibbs phenomenon

Recall Example 2.6. We observed rapid oscillation near the discontinuity. More precisely, here
we observe that the approximation by a partial sum obtains larger values than the actual values
on the left hand side of the jump discontinuity. This error is called overshoot. Similarly, on the
right hand side of the jump discontinuity the approximation tends to be too small, i.e. there is
undershoot. This approximation error does not converge vanish if we add more terms to the
partial sum. Merely, the spike near the jump discontinuity gets sharper and sharper. The figures
below illustrate this in the case of Example 2.6. Here a denotes the average of the one-sided
limits.
2.7 Convergence of Fourier series MATH.APP.240 F OURIER METHODS / M e / 2020 40

f(t) f(t)
S100(t) S1000(t)
a a

t t
3 π 3.5 3 π 3.5

Figure 32. Approximation by the partial Figure 33. Approximation by the partial
sum S100 . sum S1000 .

The phenomenon observed is general, i.e. whenever the function has a jump discontinuity and
we approximate it by a partial series there exist undershoot and overshoot on opposite sides
of the
jump. This phenomenon is called Gibbs phenomenon Josiah Willard Gibbs (1839-
1903) . Gibbs was American scientist whose worked on physics, chemistry, and mathematics.

The total magnitude of overshoot and undershoot due to Gibbs phenomenon is about 18% larger
than the true difference of the one-sided limits. That means that we have approximately 9%
overshoot on one side of the jump discontinuity and 9% undershoot on the other. This is a
mathematically well established fact that stems from the difficulty of approximating a discon-
tinuous function by continuous ones. Proving this is out of the scope of this text. However, we
present the related concept of uniform discontinuity in the next section which gives us a tool to
consider the phenomenon in mathematical terms.

On this course we use approximation 9% for the overshoot/undershoot (you should memorize
this). At each point of discontinuity we take the absolute value of the jump, i.e. the difference of
right- and left-hand limits, and the approximation is 9% of that value. We add this value to the
larger one-sided limit and reduce from the smaller one. We have illustrated this in the following
figure, where the green lines illustrate the 9% overshoot/undershoot at the jump discontinuity
located at t = t0 . The left-hand limit is a and the right-hand limit is b. We denote absolute value
of their difference by c = |a − b|.

f(t)
a+(9/100)c
a

Janan pituus c=a–b

b
b–(9/100)c
t
t0

Figure 34. Overshoot with jump of c units.

2.7 Convergence of Fourier series MATH.APP.240 F OURIER METHODS / M e / 2020 41

♠Example 2.15.
In Example 2.6 the one-sided limits at the discontinuity t0 = π are f π + =
π 2 − π and f π − = π 2 + π, so the absolute value of the jump is 2π. If we approximate the
function by a partial sum of Fourier series near t = π then the maximum (in this case on the
left-hand side of the discontinuity) obtained is about
9
π2 + π + 2π ≈ 13.577.
100
9
The minimum near T able = π is approximately 100
2π ≈ 0.565 less than the left-hand limit
π 2 − π. ♠

The number

π
sin t
Z
dt ≈ 1.851937052
0 t


is called Wilbraham-Gibbs constant. In Maple it is denoted by Si π . The exact magnitude
of the overshoot/undershoot is


1 π sin t
Z
1
dt − ≈ 0.0894898722.
π 0 t 2
To be precise this means the following. Approximate the function by the partial sum
N

a0 X 


SN t = + an cos nωt + bn sin nωt ,
2 n=1

and denote the difference between one-sided limits at jump discontinuity t = t0 by

−
f t+



0 − f t0 = d.

We denote the location of the first local extrema of SN (t) − f (t) on the left-hand side of the
jump discontinuity by tN and on the right-hand side by t0N . These are the points at which the
maximum overshoot/undershoot are attained. It can be shown (skipping the proof again) that
tN → t0 and t0N → t0 as N → ∞, and
−

f t+



0 + f t0
lim SN t0 = ,
N →∞ 2
Z π
!
1 sin t 1
lim SN (tN ) = f t+


0 + dt − · d,
N →∞ π 0 t 2

!
π
sin t
Z
1 1
lim SN (t0N ) = f t−


0 − dt − · d.
N →∞ π 0 t 2

In order to smooth out the overshoot/undershoot due to Gibbs phenomenon we can modify the
Fourier series approximation slightly. One such modification is σ-approximation, where we
wight every nth harmonic by a coefficient depending on n. Then the new approximation is
N

a0 X 


g:g t = + σn an cos nωt + bn sin nωt ,
2 n=1
2.7 Convergence of Fourier series MATH.APP.240 F OURIER METHODS / M e / 2020 42

where
nπ


sin N
σn = nπ , n = 1, 2, . . . , N.
N
Naturally, by using such an approximation we give up on some convergence properties. How-
ever, the overshoot/undershoot of g at a discontinuity has smaller magnitude when compared to
that of SN . For further information search the Internet for "sigma approximation".

2.7.3 *Uniform convergence

Let us consider the convergence of a sequence fn of function over the interval I. We say that
fn converges pointwise to the function f , if for every point x in I we have



lim fn x = f x .
n→∞

In this expression, we fix the point x ∈ I first and then let n approach infinity. However, the
convergence to the limit function needs not happen "smoothly".

♠Example 2.16. As an example, consider the function sequence

n  
fn x = nx 1 − x , x ∈ 0, 1 , n = 1, 2, 3, . . . , s

i.e. the sequence is

2
3
n
x 1 − x , 2x 1 − x , 3x 1 − x , . . . , nx 1 − x , . . .
 
Fixing x = a ∈ 0, 1 = I, we obtain the sequence


2
3
n
a 1 − a , 2a 1 − a , 3a 1 − a , . . . , na 1 − a , . . .

numbers. For a equal to zero or one every member of the sequence are zeros. If a ∈
of real
0, 1 , then the limit of the sequence is

n na l’H a
lim na 1 − a = lim
−n =
∞
lim
−n
= 0.
n→∞ n→∞ 1−a ” ∞ ” n→∞ − 1−a ln 1 − a


This shows that the sequence fn converges pointwise to the zero function f : f x = 0. Let
us study the convergence more closely by sketching the graphs of the functions fn for several
n. This is done in Figure 35 for f1 , f3 , f5 , f7 , f9 and in Figure 36 for f50 , f100 , f150 . According
to the figures, fn has a hump that gets sharper and sharper and moves closer to y-axis as n
increases. The convergence is not "smooth" in the sense that the maximum error between the
functions fn and the limit f does not seem to approach zero.
2.7 Convergence of Fourier series MATH.APP.240 F OURIER METHODS / M e / 2020 43

0.3 0.3
f1
f3
f5
f7
f9
t t
0.2 0.6 1 0.2 0.6 1

Figure 35. Graphs of elements of fn . Figure 36. Graphs of elements of fn .

Let us now show that the maximum value of fn does not tend to zero as n → ∞. The maximum
is obtained at a zero of the derivative:

n
n−1
n−1
fn0 x = n 1 − x − n2 x 1 − x



=n 1−x 1 − x − nx .
1
The zeros are 1 and n+1
. The maximum is located at the latter zero, so the maximum value is
   n  n+1
1 n 1 1 1
fn = 1− = 1− → ≈ 0.3678794412, as n → ∞.
n+1 n+1 n+1 n+1 e

A stronger convergence type is uniform convergence. The idea of uniform convergence is

that the maximal difference between the elements of the function sequence and its limit tends
to zero everywhere or, in other words, uniformly. Mathematically, fn converges uniformly to
the limit f over the interval I if



lim sup fn x − f x = 0,
n→∞ x∈I

where sup stands for the smallest upper bound (supremum ) over the interval I. The sequence
x∈I
of the above example does not converge uniformly since the above condition does not hold.

Theorem 2.2 shows that the Fourier series of a periodic function f satisfying the Dirichlet condi-
tions converges pointwise to f wherever f is continuous. Mathematical interpretation of Gibbs
phenomenon is that the convergence of the Fourier series is not uniform. However, for certain
classes of function we have uniform convergence, e.g. continuous differentiability is a sufficient
condition for uniform convergence of the Fourier series:

Theorem 2.5. Let f be a T -periodic continuously differentiable function. The Fourier

series of f converges to f uniformly, i.e.
k  
a0 X



lim sup + an cos nωt + bn sin nωt − f t = 0.
k→∞ t∈R 2
n=1
2.8 Periodic extensions of functions over finite intervals MATH.APP.240 F OURIER METHODS / M e / 2020 44

2.8 Periodic extensions of functions over finite intervals

Assume that we have a function f that is defined over a finite interval 0, T or that it is
a
general (not necessarily periodic) function whose values are of interest over the interval 0, T .
Is it still possible to find a Fourier series of the function over this specific period? The answer
lies in the following two observations. First, the Fourier series of a periodic function with pe-
riod T can be determined just by considering the function over the finite interval (0, T ) since
the Fourier coefficients are determined by integrals over (0, T ). Secondly, the function and its
Fourier series correspond to each other outside the interval (0, T ) because of periodicity. So, if
we are not interested of what happens outside the period (0, T ) and we can extend the function
f into a periodic function having values corresponding to f over (0, T ), then we can find the
desired Fourier series for f . There are multiple ways to do the extension. In what follows, we
present three ways doing this.

1. Periodic extension: We interpret that the

values of f over (0, T ) define one period of
a periodic function g. Then g is obtained by
copying this period infinitely many times. The
t
new function has period T , i.e. g(t + T ) = g(t),
and f (t) = g(t) over (0, T ). In Figure 37 T
the original function is plotted in red and its
values are copied to other periods without Figure 37. Periodic extension.
modifications.

2. Even periodic extension: Alternatively we

can define the function
(



f t , t ∈ 0, T ,
g t =

t
f −t , t ∈ − T, 0 , T T


2T
g t + 2T = g t .
Figure 38. Even periodic extension.

This function has period 2T and it corresponds to f over (0, T ). The graph of g is illustrated in
Figure 38. The graph of g over (−T, 0) is the mirror

image of that of the original function (in
red). Thus, g is an even function over − T, T , and the values over this period are copied to
form a periodic function. In the corresponding Fourier series, bn = 0, i.e. the series is of the
form ∞

a0 X
ˆ


f t = + an cos nωt .
2 n=1


Because the terms of the series only contain cosine terms and a constant ( cos 0 = 1 ), we call
this the cosine series expansion.
2.8 Periodic extensions of functions over finite intervals MATH.APP.240 F OURIER METHODS / M e / 2020 45

3. Odd periodic extension: We can define an

odd function g using f by setting
(

2T

f t , t ∈ 0, T , t
g t =

T T
−f − t , t ∈ − T, 0 ,



g t + 2T = g t .

Figure 39. Odd periodic extension.

As illustrated in Figure 39, we have first reflected the function f with respect to the origin, and
then copied the resulting odd function over the period (−T, T ) to other periods. The resulting
function is periodic with period 2T . For all odd function an = 0, so the corresponding Fourier
series is of the form ∞

X
ˆ


f t = bn sin nωt
n=1
and it is called the sine series expansion.

The three different extensions all result to a different Fourier series which all are representatives
of the function f over the interval (0, T ). It is important to note that outside (0, T ) they all con-
verge to a different function. The properties of the different Fourier series are illustrated by the
following example.

♠Example

2.17. Consider the line segment defined by the function y = t over the interval
0, π . Let us find the Fourier series corresponding to the periodic, even period, and odd peri-
odic extensions.

1. The Fourier series of the periodic extension


g t = t, 0 < t < π,
is ∞


π X sin 2nt
ĝ t = − .
2 n=1 n
g(t)
S5(t)
t t
−2π −π π 2π −2π −π π 2π

Figure 40. Periodic extension g. Figure 41. Function g and the partial sum
S5 .

2. The Fourier series of the even periodic extension




h t = |t|, −π < t < π, h t + 2π = h t ,
is ∞


π 2 X (−1)n − 1

ĥ t = + 2
cos nt (cosine series expansion).
2 π n=1 n
2.8 Periodic extensions of functions over finite intervals MATH.APP.240 F OURIER METHODS / M e / 2020 46

h(t)
S5(t)
t t
−2π −π π 2π −2π −π π 2π

Figure 42. Even periodic extension h. Figure 43. Function h and the partial sum
S5 .

3. The Fourier series of the odd periodic extension




r t = t, −π < t < π, r t + 2π = r t ,

is ∞
n

X −1

r̂ t = −2 sin nt (sine series expansion).
n=1
n

r(t)
S5(t)
t t
−2π −π π 2π −2π −π π 2π

Figure 44. Odd periodic extension r. Figure 45. Function r and the partial sum
S5 .

Can you explain why r̂ is sine series expansion but ĝ is not?

 
Let us focus on what happens over the interval 0, π of interest.

g(t)
S5(t)
h(t) r(t)
S3(t) S5(t)

t t
π π

Figure 46. The partial sum ĥ : S3 . Figure 47. The partial sums ĝ : S5 and
r̂ : S5 .

Figure 46 shows that the truncated Fourier series related to h with only four terms approximates
the straight line very well. There is only small error near the endpoints of the line segment. On
the other hand, Figure 47 shows that the partial sum approximations with more terms related to
periodic and odd periodic extensions are much poorer. The partial sums are
2.9 *Term-by-term differentiation and integration of Fourier series MATH.APP.240 F OURIER

METHODS / M e / 2020 47

π
1
1
1
1

ĝ : S5 = − sin 2t − sin 4t − sin 6t − sin 8t − sin 10t ,
2 2 3 4 5
π 4
4

ĥ : S3 = − cos t − cos 3t ,
2 π 9π


2
2
2

r̂ : S5 = 2 sin t − sin 2t + sin 3t − sin 4t + sin 5t .
3 4 5

It seems that the cosine series expansion out performs the other two when it comes to the accu-
racy of the approximation. We can explain this theoretically. Since the even periodic extension
is continuous, Theorem 2.4 guarantees that the Fourier coefficients have convergence rate 1/n2
which outperforms the convergence rate 1/n ralated to the Fourier series of discontinuous pe-
riodic and odd periodic extensions. This means that the dominance of the first terms in cosine
series expansion is stronger than in the other two. In addition, there is no Gibbs phenomenon
related to the continuous functions. ♠

2.9 *Term-by-term differentiation and integration of Fourier series

Recall that the power series defining the function f t can be differentiated and integrated
term-by-term, and the corresponding series are again power series whose limits converge to
the derivative and integral of f inside the interval of convergence (with possible exceptions at
the end points). Unfortunately, we cannot proceed this straightforwardly with the Fourier series.

In what follows, the smoothness of a function is to be understood in the sense of the times the
function is continuously differentiable. The higher the least order of continuous derivative the
smoother the function is. Integration tends to smooth a function. In particular, if f has only
jump discontinuities, its intergal function
Z t



I t = f t dt
a

is continuous. Differentiation has on opposite effect, and the derivative of a continuous func-
tion is not necessarily continuous. On the other hand, smoothness is related to the rate of
convergence of the Fourier coefficients towards zero by Theorem 2.4, and consequently to that
convergence of the Fourier series itself. That is why term-by-term integration of a Fourier series
produces another Fourier series more often than term-by-term differentiation. We formalize this
in the following two theorems. Their proofs are omitted.

Theorem 2.6. The term-by-term integrated Fourier series of a periodic function f satis-
fying the Dirichlet conditions converges to an integral function of f .
2.9 *Term-by-term differentiation and integration of Fourier series MATH.APP.240 F OURIER

METHODS / M e / 2020 48

Remark. Every integrable function has infinite number of integral functions that differ by a
constant. The above theorem states that we can find such a constant that the term-by-term in-
tegrated Fourier series is a representative of the corresponding integral function. However, one
should note that the term-by-term integrated Fourier series needs not be a Fourier series.

♠Example 2.18. Consider the periodic extension g of Example 2.17:

g t = t,
if 0
≤ t < π,
g t+π =g t

Its Fourier series is ∞

π X sin 2nt
ĝ t = − .
2 n=1 n
If we integrate the series term-by-term we obtain the function series
∞
! ∞

π X sin 2nt cos 2nt
Z
π X
− dt = t + + C.
2 n=1 n 2 n=1
2n2

Assuming that the series converges to the integral function of g we obtain

∞

1 2 π X cos 2nt
t = t+ + C.
2 2 n=1
2n2
 
We can deduce the constant of integration by fixing an appropriate value t ∈ 0, π . Substituting
t = 0 yields
∞
X 1
0= + C,
n=1
2n2
and consequently
∞
1X 1 1 π2 π2
C=− = − · = − (see Example 2.11).
2 n=1 n2 2 6 12

The term a20 t = π2 t does not appear in any Fourier series, so the series is not a Fourier series. If
we move this term to the left-hand side of the equation we obtain the series
2 ∞

π X cos 2nt
fˆ t = − +


12 n=1 2n2

which is the Fourier series of the periodic function f

1 π
f t = t2 − t, if 0 < t < π,
2 2



f t+π =f t .

(Verify this by finding the coefficients an and bn of the Fourier series of f by using the defini-
tion.) ♠
2.9 *Term-by-term differentiation and integration of Fourier series MATH.APP.240 F OURIER

METHODS / M e / 2020 49

Theorem 2.7. Consider a periodic function f satisfying the Dirichlet conditions. The
derivative f 0 , whenever it exists, may be found by term-by-term differentiation of the
Fourier series of f if and only if the function f is continuous everywhere and the function
f 0 has a Fourier series expansion.

Remark. If f 0 satisfied the Dirichlet conditions, it has the Fourier series expansion.

♠Example 2.19. Consider the even periodic extension h of Example 2.17. The function


h t = |t|,

if −
π ≤ t < π,
h t + 2π = h t
is continuous everywhere and it is differentiable everywhere except at the locations where its
graph has sharp edges. The derivative is
(

0

1, if 2nπ < t < 2n + 1 π,
h t =

−1, if 2n − 1 π < t < 2nπ, n ∈ Z.

This is the square wave, and it satisfied the Dirichlet conditions. Consequently, it has the Fourier
series expansion.

The Fourier series of h is

∞

π X −4

ĥ t = +
2 cos (2n − 1)t
2 n=1 π 2n − 1
and term-by-term differentiation yields
∞
∞

X 4 2n − 1

4X sin (2n − 1)t

2 sin 2n − 1 t =
,
n=1 π 2n − 1 π n=1
2n − 1
which is the Fourier series of square wave given in Example 2.12. ♠

♠Example 2.20. Consider the odd periodic extension r

r t = t,
if −
π ≤ t < π,
r t + 2π = r t
of Example 2.17. It is not continuous, and its derivative is
r0 t = 1, t 6= 2k − 1 π.





The Fourier series of the derivative is rb0 t = 1. Obviously, we cannot find it by term-by-term
differentiation: The Fourier series of r is
∞
n

X −1

r̂ t = −2 sin nt
n=1
n
and term-by-term differentiation yields
∞
n ∞
X −1 n
X
n

−2 cos nt = −2 − 1 cos nt ,
n=1
n n=1
which does not converge. (plot graphs of partial sums to see what happens!) ♠
2.10 Yet another form of Fourier series MATH.APP.240 F OURIER METHODS / M e / 2020 50

2.10 Yet another form of Fourier series

We have considered the trigonometric and exponential Fourier series that are equivalent forms
of Fourier series. We also mentioned the form where the harmonic component functions are
represented by sine functions with varying amplitude and phase angle. Let us find yet another
form for Fourier series where the harmonic components are given in terms of cosines. In order
to do this, consider the coefficients of the exponential Fourier series
1p 2
cn = an + b2n = c−n
2
and


arg cn = − arg c−n .

 
Denote the phase angle arg cn by θn . We choose θn from the interval − π, π . When the
phase angle is ±π we can choose the value that is suitable for us. The coefficients can be written
in the form
cn = cn ejθn and c−n = cn e−jθn
so we find the following new form for the Fourier series:
∞ 
X 
fˆ t = c0 + jnωt −jnωt


cn e + c−n e
n=1

∞
X 


j nωt+θn −j nωt+θn
= c0 + cn e +e
n=1

∞
X
= c0 + cn 2 cos (nωt + θn ) .
n=1

The form where the harmonic functions are written as sine functions can be obtained from the
above representation of the Fourier series by observing the π/2-phase angle between the cosine
and sine functions: π

  π
cos α = sin − α = − sin α − .
2 2
In what follows we only consider the form with cosine functions.

Let us write down the three different forms of Fourier series:

2.10 Yet another form of Fourier series MATH.APP.240 F OURIER METHODS / M e / 2020 51

∞  

a0 X
fˆ t =



+ an cos nωt + bn sin nωt
2 n=1

∞
X
= cn ejnωt
n=−∞

∞
X
= c0 + 2 cn cos (nωt + θn ) ,
n=1

where
p

2cn = an − bn j, 2c−n = an + bn j, 2 cn = a2n + b2n , θn = arg cn .

♠Example 2.21. Let us find the three forms of Fourier series of the function s, defined by

(

t, 0 < t < π π
s t = t
0, −π < t < 0 −π


−2π π 3π
s t + 2π = s t

Figure
 48. The graph of the function s over the
interval − 2π, 4π .

Instead of using the definition we observe that

t + |t| r t +h t
s t = = , −π < t < π,
2 2
where r and h are the periodic extensions of Example 2.17. We already know their Fourier
series so we can use linearity to find the Fourier series


∞
∞
n

r̂ t + ĥ t π 1 X 2 (−1)n − 1
1X 2 −1

ŝ t = = + 2
cos nt + − sin nt ,
2 4 2 n=1 nπ 2 n=1 n

so  
∞
n
n

π X  −1 −1
−1

ŝ t = + 
2π
cos nt − sin nt .
4 n=1 | n{z } n
| {z }

=an =bn

The exponential Fourier series is

∞
n
n !

π X −1 −1 −1
ŝ t = + +j ejnt .
4 n=−∞ 2n2 π 2n
n6=0 | {z }
=cn

The third form is ∞

π X
ŝ t = + 2 cn cos (nt + θn ) ,
4 n=1
2.11 Discrete frequency spectra MATH.APP.240 F OURIER METHODS / M e / 2020 52

where θn = arg(cn ). Writing this down explicitly does not give us a nice form with further in-
formation. However, finding the amplitudes |cn | and phase angles θn using a computer software
and representing them appropriately gives us a nice way to represent the crucial information
provided by the Fourier series. This is the topic of the next section. ♠

2.11 Discrete frequency spectra

The Fourier series gives a decomposition of a periodic function into an infinite number of har-
monic or frequency components. The frequency components have frequencies that are multi-
ples of the frequency
2π
ω0 =
T
of the original function. We assume ω0 to be positive. However, in the mathematical model
we have negative frequencies −nω0 which correspond to negative indices. The
amplitudes of
the frequency components are An = 2 cn and the phase angle is θn = arg cn . This means
that the information about the function is stored in the amplitudes and phase angles. Plotting
the amplitude and phase angle against the frequency gives us amplitude and phase spectrum,
respectively. These plots give us a graphical illustration of the distribution of the function into
different frequency components. In explicit, amplitude spectrum is the sequence
. . . , c−3 , c−2 , c−1 , c0 , c1 , c2 , c3 , . . .
and the phase spectrum is the sequence
. . . , −θ3 , −θ2 , −θ1 , θ0 , θ1 , θ2 , θ3 , . . . .

Spectrum is often illustrated by the following kind of plots where every vertical line segment
represent a frequency component:

| cn |

ω
−7ω0 −5ω0 −3ω0 −ω0 ω0 3ω0 5ω0 7ω0

Figure 49. Amplitude spectrum.

arg(cn) π

−7ω0 −5ω0 −3ω0 −ω0 ω

ω0 3ω0 5ω0 7ω0

−π
2.12 Parseval’s theorem MATH.APP.240 F OURIER METHODS / M e / 2020 53

Figure 50. Phase spectrum.

The amplitude spectrum is symmetric with respect to zero frequency and the phase spectrum is
symmetric with respect to the origin. Thus, the crucial information can be extracted from the
positive frequency components. This is natural, since in real applications negative frequencies
do not make sense. They just happen to be mathematically convenient tool in modelling. The
amplitude of the nth frequency component An = 2 cn is split evenly into components with the
frequencies ±nω0 in the mathematical model.

♠Example 2.22. Let the frequency of the original function be ω0 = 1. The frequency and
amplitude spectra of the signal are presented in the following figures. We assume that cn =
0, n = ±4, ±5, . . .

| cn |
arg(cn)
3
2
1.5
1
0.1 ω −3ω0 ω0 ω
−3ω0 −ω0 ω0 3ω0 −ω0 3ω0

−π

Figure 51. Amplitude spectrum. Figure 52. Phase spectrum.

The figure contains the crucial information about the coefficients cn . For example,
1
c0 = 1 · ejπ = −1, c1 = 3 · e−j2 , c2 = 0, c3 = · ej1.5
10
and so on. Thus,
1
fˆ t = −1 + 2 · 3 cos (ω0 t − 2) + 0 · cos (2ω0 t + 0) + 2 ·


cos (3ω0 t + 1.5)
10
1
= −1 + 6 cos (t − 2) + cos (3t + 1.5) .
5
♠

2.12 Parseval’s theorem

The average power P associated with the periodic signal f with period T is defined as
1 d+T 2

Z
f t dt
T d
2.12 Parseval’s theorem MATH.APP.240 F OURIER METHODS / M e / 2020 54

which is the mean square value of the function f over one period. Sometimes we need the mean
value of the product of two functions. These can be represented using Fourier series.

Theorem 2.8 (The multiplication theorem). Let f and g be T -periodic functions that have the
Fourier series
∞ ∞
ˆ
X X 2π
cn ejnωt γn ejnωt



f t = and ĝ t = where ω = .
n=−∞ n=−∞
T

Then ∞ ∞
Z d+T
1 X X
cn γn∗ = c∗n γn .



f t g t dt =
T d n=−∞ n=−∞

Proof. The functions g are equal ĝ except at isolated points of discontinuities. Thus, their
integrals over a specified interval are the same. Assuming that the term-by-term integration is
allowed we have
1 d+T 1 d+T
Z Z





f t g t dt = f t ĝ t dt
T d T d
∞
!
1 d+T
Z

X
= f t γn ejnωt dt
T d n=−∞

∞ 
1 d+T
X Z 

jnωt
= f t e dt γn
n=−∞
T d

∞
X ∞
X
= c−n γn = c∗n γn .
n=−∞ n=−∞

If we substitute n = −k, we observe that the series is the same as the series with terms cn γn∗ .
This is natural since we can change the order of f and g in the integral.

Theorem 2.9 (Parseval’s theorem). Consider a T -periodic function f having the Fourier
series ∞
ˆ
X 2π
cn ejnωt , where ω =


f t = .
n=−∞
T
We have that ∞
Z d+T
1
2
X 2
f t dt = cn .
T d n=−∞

This formula is called Parseval’s equation or Parseval’s identity.

Proof. The result is a direct consequence of the multiplication theorem.

2.12 Parseval’s theorem MATH.APP.240 F OURIER METHODS / M e / 2020 55

The coefficients of the exponential Fourier series are

an − b n j an + b n j a0
cn = , c−n = , c0 = ,
2 2 2
so the terms in the series appearing Parseval’s equation are

2 2 an − b n j an + b n j a2 + b2n
cn = c−n = cn c−n = · = n ,
2 2 4
so we can write it in the form
∞ ∞
1 d+T 2
a2 X a2n + b2n
Z
2
X
f t dt = cn = 0+ .
T d n=−∞
4 n=1
2

♠Example 2.23. Let us consider the 2π-periodic function signal s of Example 2.21 defined
by (

t, 0 < t < π
s t =
0, −π < t < 0 .



s t + 2π = s t
Its exponential Fourier series is
∞
n
n !

π X −1 −1 −1
ŝ t = + +j ejnt .
4 n=−∞ 2n2 π 2n
n6=0 | {z }
=cn

2
The power spectrum of s is the plot of the terms cn against frequencies.

| cn |2
0.6

0.3

ω
−10 −5 −1 1 5 10

Figure 53. Power spectrum of s (having frequency 1) over the interval [−10, 10].

2
The terms cn approach zero as n increases (in other words as the frequency nω of the har-
monic components increases). If we sum the three largest values we have
1
2
X
cn ≈ 1.319492642.
n=−1

The average power of s is

Z π Z π π 3
1 2 1 1 t π2
t2 dt =


s t dt = = ≈ 1.644934068.
2π −π 2π 0 2π 0 3 6
2.12 Parseval’s theorem MATH.APP.240 F OURIER METHODS / M e / 2020 56

We observe that the average power associated with the first harmonic component and the con-
stant term is approximately
1.319492642
· 100 ≈ 80.216
1.644934068
percent of the total power of the signal. If we add three first harmonic components and the
constant term the associated average power is
3
2
X
cn
n=−3 1.502549956
· 100 ≈ · 100 ≈ 91.344
1.644934068 1.644934068
percent of the total power. ♠
3 Discrete Fourier transform (DFT) MATH.APP.240 F OURIER METHODS / M e / 2020 57

3 Discrete Fourier transform (DFT)

3.1 Definition of Discrete Fourier transform

In many applications the exact function is not known, only a sequence of samples of the function
at certain time instants. A typical situation is a measured data, for example in a physical experi-
ment, where the values of the quantity of interest is measured at subsequent time instants. Then
we can use the discrete Fourier transform which is an estimation of the Fourier series based on
the known values of the underlying function. Next we define the discrete Fourier transform.

A periodic function f satisfying the Dirichlet conditions has the exponential Fourier series
∞ 
X  ∞
X
fˆ t = c0 + cn ejnωt + c−n e−jnωt = cn ejnωt ,


n=1 n=−∞

where
1 d+T
Z
f t e−jnωt dt, n = 0, ±1, ±2, ±3, . . .


cn =
T d
The coefficients cn , n = 0, ±1, ±2, . . . form a sequence that is called the finite Fourier trans-
form. Finiteness refers to the finite period T . The general Fourier transform that is defined in
the next chapter does not require the function to be periodic, or in other words the function can
have an infinite period.

When applying the Fourier transform we might not know whether the function is periodic or
not. Then we the approach introduced in Section 2.8, we have general function over the interval
with measured data and outside the interval the function is assumed to be periodic.

T
 
Consider the period 0, T and divide it into N equal size subintervals of length N
. The lower
bounds of the subintervals are


T 2T N −1 T
0, , ,..., .
N N N
The values of the function f (the

samples)

are given at these time instants. We assume that the
function is T -periodic, so f 0 = f T . The coefficients of the exponential Fourier series are

1 T
Z
f t e−jnωt dt.


cn =
T 0
We approximate this integral by the Riemann sum, where the values of the function are evalu-
ated at the left boundary t = tk on each subinterval. The lengths of the subintervals are equal
T
∆t = N . Thus, the Riemann sum approximation of cn is
N −1 N −1 
X
−jnωtk
X kT −jnω kT T
f tk e ∆t = f e N ,
k=0 k=0
N N

which we denote by dn , i.e.

N −1   N −1  
1 X kT −jnω kT T 1 X kT −jnω kT
cn ≈ d n = f e N = f e N .
T k=0 N N N k=0 N
3.1 Definition of Discrete Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 58

We simplify the notation by denoting the samples by

 
kT
gk = f , k = 0, 1, 2, . . . , N − 1
N
2π
and by replacing ω by T
. Then
N −1
1 X 2π
dn = gk e−jnk N , n = 0, 1, 2, . . . , N − 1.
N k=0

We denote the sum in the above formula by

N −1
2π
X
Gn = gk e−jnk N , n = 0, 1, 2, . . . , N − 1,
k=0

so dn = Gn /N .

The sequence
  N −1
g0 , g1 , g2 , . . . , gN −1 = gk k=0
that is based on the samples of the function is associated with the sequence
  N −1
G0 , G1 , G2 , . . . , GN −1 = Gn n=0 ,

which we call the discrete Fourier transform (DFT). If we divide the elements of the G-
sequence by the number N of elements, we obtain an approximation of the original Fourier
coefficients cn .

Remark. A set and a sequence are two different things. Here we use the same notation for
them. The meaning of the notation should be clear from the context. Whenever we use the
notation for the sequence, it is important to keep all the repetitions and the order of elements.

Remark. The term  2πn k

2π
e−jnk N = e−j N
has interesting properties. We recall the definition of complex root and solve the equation

z N = 1.

The solutions are  k

j 2π
z= e N , k = 0, 1, . . . , N − 1.
The roots locate symmetrically on the unit circle. The sum of the roots is a geometric sum
which we can calculate (ratio q 6= 1 ):
N −1  N −1
X
j 2π
k X 1 − qN
e N = qk = = 0.
k=0
|{z}
k=0
1−q
=q

Similarly, (the only difference is the power n) the elements

 2πn k
e−j N , k = 0, 1, 2, . . . , N − 1
3.1 Definition of Discrete Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 59

of the G-sequence locate symmetrically on the unit circle (e−j2πn = ej2π = 1).
 N −1
These observations help us to construct the inverse transform. Given the sequence Gn n=0
 N −1
we want to find the sequence gk k=0 . To this end, we consider the sum

N −1 N −1  N −1 
jnk 2π −jnm 2π 2π
X X X
Gn e N = gm e N ejnk N
n=0 n=0 m=0

N −1 N −1 
−jn 2π
X X
(m−k)
= gm e N

m=0 n=0

N −1 N −1  n
X X (m−k)
−j2π
= gm e N .
m=0 n=0

The indices m and k satisfy the inequality

0 ≤ m − k ≤ N − 1,
(m−k)
which implies that N
is an integer only if m − k = 0. If m = k, then

N −1  n N −1
−j 2π
X X
(m−k)
gm e N = gk 1 = gk N .
n=0 n=0

If m 6= k, then

N −1 
X 2π
n 1 − qN 1−1
gm e|−j N{z(m−k)} = gm = = 0.
n=0
1−q 1−q
=q6=1

Thus,
N −1 N −1 N −1  n
jnk 2π 2π
X X X
Gn e N = gm e−j N (m−k) = 0 + 0 + · · · + N gk + 0 + · · · + 0 = N gk ,
n=0 m=0 n=0

which implies that

N −1
1 X 2π
Gn ejnk N = gk ,
N n=0


which defined the sequence which we call the inverse discrete Fourier transform IDFT .

To sum up,

N −1
2π
X
gk e−jnk N , n = 0, 1, 2, . . . , N − 1


Gn = DFT
k=0

and
3.1 Definition of Discrete Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 60

N −1
1 X 2π
Gn ejnk N , k = 0, 1, 2, . . . , N − 1.


gk = IDFT
N n=0

There are multiple different definitions for the discrete Fourier transform and its inverse. The
definition adopted here coincides with those of MATLAB and the book by James. The alterna-
tive definitions are similar, e.g. the difference might be the division by the element N the other
way around or the symmetric normalization
N −1
1 X 2π
Gn = √ gk e−jnk N , n = 0, 1, 2, . . . , N − 1
N k=0
and
N −1
1 X 2π
gk = √ Gn ejnk N , k = 0, 1, 2, . . . , N − 1.
N n=0
It is recommended to check the definition of the transform when using a new software or read-
ing new source material. The most important thing is that the discrete Fourier transform and its
inverse work as a pair. The definition adopted usually depends on the application.

♠Example 3.1. Consider

f t = 1 + 2t, 0 ≤ t < 4.
We assume that the function is periodic with period T = 4. We take N = 4 samples at equal
intervals 1. We obtain the sequence
 3  
gk k=0 = g0 , g1 , g2 , g3 = 1, 3, 5, 7 .
Next we find the discrete Fourier transform term-by-term

3  k X3
−j 2π·0
X
G0 = gk e 4 = gk = 1 + 3 + 5 + 7 = (1 + 5) + (3 + 7) = 16
k=0 k=0
3  k X3
X
−j 2π·1

k
G1 = gk e 4 = gk −j = 1−3j −5+7j = (1−5)−j(3−7) = −4+4j
k=0 k=0
3  k X3
X
−j 2π·2

k
G2 = gk e 4 = gk − 1 = 1 − 3 + 5 − 7 = (1 + 5) − (3 + 7) = −4
k=0 k=0
3  k X3
X
−j 2π·3

k
G3 = gk e 4 = gk j = 1 + 3j − 5 − 7j = (1 − 5) + j(3 − 7) = −4 − 4j.
k=0 k=0

Thus, the discrete Fourier transform is

 
 3 1 1 1 1 
dn n=0 = G0 , G1 , G2 , G3 = 4, −1 + j, −1, −1 − j .
4 4 4 4
Let us compare the terms of the DFT to the coefficients of the exponential Fourier series ck
which they should approximate. The Fourier series is
 
∞ sin n π t ∞ ∞
8 X 2 X 4j jn π t X −4j −jn π t
fˆ t = 5 −


=5+ e 2 + e 2 ,
π n=1 n n=1
πn n=1
πn
3.1 Definition of Discrete Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 61

which defines the sequence

 
 3  4j 2j 4j
cn n=0 = c0 , c1 , c2 , c3 = 5, , , .
π π 3π

Obviously, the elements dn do not approximate cn very well. Let us take N = 12 samples with
reduced sampling interval. We use a computer software to find the transform (doing this by
hand is too hard). The absolute values are

n 0 1 2 3 4 5 6 7 8 9 10 11

dn 4.67 1.29 0.667 0.470 0.384 0.343 0.333 0.343 0.384 0.470 0.667 1.29

cn 5. 1.27 0.636 0.423 0.318 0.254 0.212 0.182 0.159 0.141 0.127 0.116

and the phase angles

n 0 1 2 3 4 5 6 7 8 9 10 11


arg dn 0 1.83 2.09 2.36 2.62 2.88 3.14 −2.88 −2.62 −2.36 −2.09 −1.83


arg cn 0 1.57 1.57 1.57 1.57 1.57 1.57 1.57 1.57 1.57 1.57 1.57

The values are not the same, which is natural since dn only approximate cn . We observe some
similarities and differences. The sequence is decreasing cn , but the sequence dn decreases
at first and then start to grow from the sixth element on in a symmetric manner. Theoretic ex-
planation for this is given in the next section. There is a big jump from negative phase angles to
positive ones. However,
 this
 is not as severe as it first appears since the phase angle is chosen
from the interval − π, π and the jump is because the angle increases over π. When consider-
ing the situation on the unit circle the difference is not that big. The approximate and the actual
amplitude and phase spectra are.

| dn | | cn |
5
4.67

ω ω

Figure 54. dn , n = 0, ±1, ±2, . . . , ±11. Figure 55. cn , n = 0, ±1, ±2, . . . , ±11.
3.2 Properties of discrete Fourier transforms MATH.APP.240 F OURIER METHODS / M e / 2020 62

π
arg(dn) arg(cn)

π/2

ω ω

−π/2

−π

Figure 56. arg dn , n = 0, ±1, ±2, . . . , ±11. Figure 57. arg cn , n = 0, ±1, ±2, . . . , ±11.

3.2 Properties of discrete Fourier transforms

Periodicity

Let us now consider the sequences related to the discrete Fourier transform with indices 0, ±1, ±2, . . .
The function f was assumed to be periodic, so
   
(k + N )T kT
gk+N = f =f + T = gk .
N N
The discrete Fourier transform turns out to be periodic as well:
N −1 N −1
−j(n+N )k 2π 2π
X X
Gn+N = gk e N = gk e−jnk N |e−jk2π
{z } = Gn , n ∈ Z.
k=0 k=0 =1

Symmetry

The symmetry observed in Example 3.1 is a general property, i.e. if 1 ≤ n ≤ N − 1

N −1 N −1 N −1
−jk(N −n) 2π jnk 2π −jnN 2π 2π
X X X
GN −n = gk e N = gk e N e N = gk ejnk N = G∗n ,
k=0 k=0 k=0
so


GN −n = Gn ∧ arg GN −n = − arg Gn .
By the periodicity,
G−n = GN −n = G∗n .

The function f is not necessarily even, so the sequence gk does not have this symmetry
property in general. To sum up,
gk+N = gk

Gn+N = Gn (periodicity)

GN −n = G∗n , n, k ∈ Z. (symmetry)
3.3 Fast Fourier transform (FFT) MATH.APP.240 F OURIER METHODS / M e / 2020 63

Dividing the elements of the discrete Fourier transform by N does not affect the periodicity or
symmetry, which implies that thesequence dn has these properties as well. Because  of this
only the first half of the sequence dn approximates the corresponding elements of cn well.
Consequently, the remaining part should be forgot. We return to this in Chapter 6.

The periodicity and symmetry must be taken into account if the G-sequence is modified, e.g.
by filtering out some of the terms. For example, if G3 ← 0 then we must set GN −3 ← 0 as well.
This is because, otherwise the modified sequence would not be a discrete Fourier transform of
any function. Consequently, the inverted sequence would make no sense.

3.3 Fast Fourier transform (FFT)

In various application we must compromise between accuracy and computational efficiency. In

particular, large problems require fast computation time. Finding the discrete Fourier transform
with N terms using the definition requires approximately a constant times N 2 computations.
The Fast Fourier Transformation is an algorithm that reduces the computational time of discrete
Fourier transform to a constant times N · log(N ) computations. It is now a standard tool used
by mathematical software which was proposed in 1965 by J.W Cooley and J.W. Tukey. The idea
was already known by then, for example Johann Carl Friedrich Gauss (1777 - 1855) applied
similar idea in number theory, but Cooley and Tukey were the first to apply it to compute the
discrete Fourier transform. The FFT algorithm computes the time consuming sums
N −1
2π
X
gk e−jnk N , n = 0, 1, 2, . . . , N − 1
k=0

avoiding unnecessary repetitions. There are different implementations of the algorithm, but the
underlying idea remains the same. The idea is illustrated by the following example. The real
potential of the FFT algorithm was revealed as the computational efficiency of the computers
increased, and it is considered to be one of the most important numerical algorithms of our
lifetime.

♠Example 3.2. Let us illustrate how to reduce the number of computations when N = 8 = 23 .
We assume that the sequence
 N −1  7
gk k=0 = gk k=0
is given. By the definition, the discrete Fourier transform is
7
X
k 2π
Gn = gk q n , n = 0, 1, 2, 3, 4, 5, 6, 7, where q = e−j 8 ,
k=0

or in the matrix form

   0
q0 q0 q0 q0 q0 q0 q0
 
G0 q g0
G1  q 0 q1 q2 q3 q4 q5 q6 q7  g1 
   0  
G2  q
   0 q2 q4 q6 q8 q 10 q 12 q 14 
 g2 
 
G3  q
 = 0 q3 q6 q9 q 12 q 15 q 18 21   
q  g3 
G4  q q4 q8 q 12 q 16 q 20 q 24 q 28  g4  .
   0   
G5  q
   0 q 5 q 10 q 15 q 20 q 25 q 30 q 35 
 g5 
 
G6  q q 6 q 12 q 18 q 24 q 30 q 36 q 42   
g6
G7 q0 q 7 q 14 q 21 q 28 q 35 q 42 q 49
g7
3.3 Fast Fourier transform (FFT) MATH.APP.240 F OURIER METHODS / M e / 2020 64

We only need to compute the elements of the matrix once, so we assume that it is given. We
observe that the computation of the above matrix product directly involves 56 summations. We
do not consider the number of multiplications here.

Let us now to the evaluations, but avoiding unnecessary repetitions. We represent the indices in
the (binary) form

n = 4n2 + 2n1 + n0 , k = 4k2 + 2k1 + k0 , kj , nj ∈ {0, 1}.

The product of the indices is

nk = 4n0 k2 + 2n0 k1 + n0 k0 + 8n1 k2 + 4n1 k1 + 2n1 k0 + 16n2 k2 + 8n2 k1 + 4n2 k0

= 8 2n2 k2 + n2 k1 + n1 k2 + 4 n0 k2 + n1 k1 + n2 k0 + 2 n0 k1 + n1 k0 + n0 k0 .

We apply the facts q 4 = −1 and q 8 = 1. According to the formula

7 1 1 1
!!
2 n0 k1 +n1 k0
X X X X
gk q nk = g4k2 +2k1 +k0 (−1)n0 k2 +n1 k1 +n2 k0 q q n0 k0


Gn = ,
k=0 k0 =0 k1 =0 k2 =0
1 1 1
!!
2 n0 k1
X X X
n2 k0 (n0 +2n1 )k0 n1 k1 n0 k2


= (−1) q (−1) q (−1) g4k2 +2k1 +k0 .
k0 =0 k1 =0 k2 =0

In the following tabular, the sums are grouped appropriately.

n Gn 





0 g0 + g4 + g2 + g6 + g1 + g5 + g3 + g7





4 g0 + g4 + g2 + g6 − g1 + g5 + g3 + g7





2 g0 + g4 − g2 + g6 + q2 g1 + g5 − g3 + g7





6 g0 + g4 − g2 + g6 − q 2 g1 + g5 − g3 + g7




2
1 g0 − g4 + g2 − g6 q 2 + q g1 − g5 + g3 − g7 q





5 g0 − g4 + g2 − g6 q 2 − q g1 − g5 + g3 − g7 q 2





3 g0 − g4 − g2 − g6 q 2 + q 3 g1 − g5 − g3 − g7 q 2




2
7 g0 − g4 − g2 − g6 q 2 3
− q g1 − g5 − g3 − g7 q

The important observation is that we do not need to do the calculations term
by term if we

store
certain sums. For example, the sum g0 +g4 appears in four term and g0 +g4 + g2 +g6 in two
terms. By evaluating the repeated sums or subtractions only once we can reduce the number of
summations to 24. The following diagram illustrates how the computations are done.
3.3 Fast Fourier transform (FFT) MATH.APP.240 F OURIER METHODS / M e / 2020 65

b0 b0 a0 a0 G0
g0 + + +

b4 b2 a2 a1 G4
g4 -1 + -1 + -1 +

b2 b4 a4 a2 G2
g2 + + +

b6 b6 a6 a3 G6
g6 -1 + q2 -1 + -1 +

b1 b1 a1 a4 G1
g1 + + +

b5 b3 a a5 G5
g5 -1 + -1 + q2 3 -1 +

b3 b5 a5 a6 G3
g3 + + q +

b7 b7 a7 a7 G7
g7 -1 + q2 -1 + q3 -1 +

Figure 58. FFT.

4 Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 66

4 Fourier transform

4.1 Definition of Fourier transform

In the preceding chapters, we assumed that the function is periodic. Next we present the (con-
tinuous) Fourier transform that
 Tdoes  not require periodicity. The idea of the construction is to
T
start from the time window − 2 , 2 and consider the periodic extension of the given function
(Section 2.8) and construct its Fourier series (we already know how to do this). Then we let the
time T approach to infinity, i.e. we let the period to be infinite, or, in other words, the we drop
the requirement that the function is periodic.

Let us consider a general function f that is not necessarily periodic or continuous. We denote
the T -periodic extension of  

T T
f t , t∈ − ,
2 2
by g (see Figures 59 and 60).

f(t)

t
T T
2 Τ 2

Figure 59. Graph of f and the time window − T2 , T2 .

 

g(t)

t
T T
2 Τ 2

Figure 60. Periodic extension g of f .

The periodic extension g has the exponential Fourier series

∞
X
cn ejnω0 t ,


ĝ t =
n=−∞

where Z T
1 2
f u e−jnω0 u du,


cn = n = 0, ±1, ±2, ±3, . . .
T − T2

The frequency of the periodic extension is

2π
ω0 = ,
T
4.1 Definition of Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 67

so the difference of the frequencies of two subsequent terms is ∆ω = ω0 . If we let T to increase

unboundedly, the frequency decreases to zero and the the lines of the frequency spectrum (see
the figures of frequency spectrum) approach each other. On the limit T = ∞ we obtain a
continuous distribution of frequency elements ω, i.e.

lim ω0 = lim ∆ω = 0.
T →∞ T →∞

The Fourier series ĝ and the periodic extension g are the same almost everywhere, so




lim ĝ t ' lim g t = f t .
T →∞ T →∞

The series
∞  Z T 
X 1 2
−jnω0 u
du ejnω0 t


ĝ t = f u e
n=−∞
T T
−2

∞ Z T 
X 2
−jnω0 u 2π 1
= f u e du ejnω0 t
n=−∞ − T2 T 2π

∞ Z T 
1 X jnω0 t 2
−jnω0 u
= e f u e du ∆ω.
2π n=−∞ − T2

We denote Z T
2
f u e−jωu du.



GT ω =
− T2

Then ∞

1 X jnω0 t

ĝ t = e GT nω0 ∆ω.
2π n=−∞
We let T → ∞ when ∆ω → 0. Under some assumptions (f needs to be smooth enough . . .)
we obtain
∞ Z ∞ Z ∞ 

1 X jnω0 t
1 jωt

−jωu
lim ĝ t = lim e GT nω0 ∆ω = e f u e du dω.
∆ω→0 ∆ω→0 2π 2π −∞ −∞
n=−∞

Thus, Z ∞ Z ∞ 

1 jωt

−jωu
f t ' e f u e du dω.
2π −∞ −∞
The double integral on the right-hand side is called the Fourier integral of f . The Fourier inte-
gral and the original function f are the same almost everywhere just like in the case of Fourier
series (that is why we denoted ' ). We have equality wherever f is continuous. At jump dis-
continuities the value of the integral is the average of the one-sided limits.

Remark. The value Z a

lim f x dx
a→∞ −a

is called Cauchy principal value (CPV) . In the above considerations, the time window − T2 , T2
 

was symmetric and the time window was extended to whole real axis by letting the parameter
4.1 Definition of Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 68

T approach infinity. Thus, we considered the CPV. In general, the CPV can exist even if the
improper integral Z ∞


I= f x dx
−∞
does not since we cannot let the lower and upper bounds to approach infinity simultaneously,
i.e. we actually define
Z ∞ Z c Z b




I= f x dx = lim f x dx + lim f x dx.
−∞ a→−∞ a b→∞ c

If the improper integral I converges, then its value is the Cauchy principal value, but existence


of Cauchy principal value does not imply convergence of I (consider the function f x = x).

Theorem 4.1 (Dirichlet conditions for Fourier integrals). If the function f is such that

• it has a finite number of local extrema and discontinuities on any finite interval,

• the improper integral Z ∞

f t dt
−∞


converges i.e. it has a finite value ,

then the Fourier integral converges to f at every point where f is continuous. The Fourier
integral converges to the average of one-sided limits at a point of discontinuity.

Now we can define the Fourier transform of f :

Z ∞
F f t =F f f u e−jωu du.






ω =F ω =
−∞

The Fourier integral gives the inverse Fourier transform

Z ∞
1
F −1
=F −1
F ω ejωt dω.





F ω F t =f t =
2π −∞

We interpret the associated integrals as Cauchy principal values. Remember that the values of
the functions appearing in the formulas of the Fourier transform and its inverse are those of f
only when the functions are continuous.

Remark. In the book by James, the notation F jω was adopted instead of F ω that is used
here. Motivation behind this is that the Fourier transform can be seen as a restriction of the
Laplace transforms having the complex variable

s = σ + jω, σ, ω ∈ R.

Remark. Evaluation of the Fourier transform can be often carried out as two real integrals if
we consider j as an constant or an symbol since
Z ∞ Z ∞ Z ∞


−jωu




F ω = f u e du = f u cos ωu du − j f u sin ωu du.
−∞ −∞ −∞
4.1 Definition of Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 69

Evaluating the inverse transform using the definition is a bit more complicated and requires
more knowledge on complex integrals. However, the exercise and exam problems of this course
are chosen so that the mathematical skills learned on basic courses are enough. This means
that we usually use the properties and inverse transforms that we already know to deduce which
function has the given Fourier transform.

♠Example 4.1. Let us find the Fourier transform of the function

f : f t = H t e−at , a > 0.

The area between the graph of the function and t-axis

is
Z ∞ Z b
e−at dt


t f t dt = lim
−∞ b→∞
0 
1 −ab 1 0 1
= lim − e + e = .
b→∞ a a a
Figure 61. The graph of f .
It is finite. The function f has a finite number of points of discontinuities and local extrema, so
it satisfies Dirichlet conditions. Thus, the Fourier transform exists and is
Z ∞ Z ∞

−jωu
e−au e−jωu du


F ω = f u e du =
−∞ 0
b
−1 −u


a+jω
= lim e
b→∞
0 a + jω

−1  −ba −jbω  1
= lim e e − e0 = .
b→∞ a + jω a + jω
♠

The assumption that the improper integral

Z ∞

f t dt
−∞

converges is a strong one, and there are several common functions
that do not satisfy it e.g.,
polynomials, trigonometric functions, exponential function . In practice, the signals usually are
finite, i.e., they do not start from the negative infinity or continue forever. Then we actually have
only a proper integral and the the above integral converges.

♠Example 4.2. Let us find Fourier transform of

(

1, if 0 ≤ t ≤ 2
f :f t =
0, otherwise.

Then
2
2


−1 −jωu −e−jω2 + e0 ejω − e−jω 2 sin ω
Z
ω6=0
e−jωu du = = e−jω = e−jω


F ω = e =
0 0 jω jω jω ω
4.2 Sinc function MATH.APP.240 F OURIER METHODS / M e / 2020 70

and


Z 2
F 0 = 1 du = 2.
0
♠

4.2 Sinc function

A particular impulse is given by sinc function that is defined by



sin t


sin x 
, t 6= 0
sinc t = lim = t
x→t x 
1, t = 0.

Figure 62. The graph of f : f t = sinc 30t .

Using MacLaurin series of sine we obtain

∞
k ∞
k
sin x X −1 x2k+1 X − 1 1 1 1
=
=
x2k = 1 − x2 + x4 − x6 + − · · · .
x k=0
2k + 1 ! x k=0
2k + 1 ! 3! 5! 7!
This reveals that sinc is continuous at t = 0 (the limit in the definition is not actually needed).

Using the above definition, the Fourier transform of Example 4.2 can be written in the simple
form
F (ω) = 2e−jω sinc ω .

Alternative definitions of sinc exist in the literature, so we must always check which definition
is used in which text. The following function appears often in applications
Z t
∞
k
sin x X −1

t2k+1 .


Si t = dx =

0 x k=0
2k + 1 ! 2k + 1
The above definition does not have a closed form, i.e., it cannot be defined as a combination
of elementary functions. However, the corresponding series is convergent and well defined.
Applying results of complex integrals (for example residue theorem that is presented in Complex
analysis), we can show that
Z ∞

sin x π
dx = .
0 x 2
The name Si is used in mathematical software and is due to the words The Sine Integral.
4.3 Spectrum of a non periodic function MATH.APP.240 F OURIER METHODS / M e / 2020 71

♠Example 4.3. Let us find the Fourier transform of the rectangular pulse
(




 A, |t| < T
f t = A rectT t = A H t + T − H t − T =
0, |t| > T.

t
T T

 
Figure 63. Rectangular pulse of height A over the interval − T, T .

∞ T T
−1 −u
Z Z

ω6=0 jω
f u e−jωu du = −jωu



F ω = Ae du = A e
−∞ −T −T jω



T jω −T jω


e −e ejT ω − e−jT ω sin T ω
=A = 2AT = 2AT
jω 2j · ωT Tω

Z ∞

Z T
F 0 = f u du = A du = 2AT.
−∞ −T

Thus,


F ω = 2AT sinc T ω .

4.3 Spectrum of a non periodic function

In Examples 4.1 and 4.3, we evaluated the Fourier transforms

n o
1 a − jω
F H t e−at ω = F ω =



= 2
a + jω a + ω2
and n
o

F A rectT t



ω = G ω = 2AT sinc T ω .


The first transform

obtains imaginary values Im 6= 0 while the values of the latter are real
Im = 0 . We can always plot the graph of the real function, so we have a nice graphical
representation for the Fourier transform of the square pulse. Every complex number can be
presented in the polar form, in particular we can write
4.3 Spectrum of a non periodic function MATH.APP.240 F OURIER METHODS / M e / 2020 72

j arg F (ω)



F ω = F ω e ,



where F ω and arg F (ω) are real valued functions of a real variable ω. These functions are
the amplitude and phase spectra. Unlike the amplitude and phase spectra of a periodic function
obtained from its Fourier series, the phase and amplitude spectra are continuous. This makes
sense if we consider the way in which we derived the Fourier transform as a limit of Fourier
series. Let us plot the spectra of the function in Example 4.1 with a = 3, i.e.,
n o
1
F H t e−3t ω = F ω = = r ω ejθ(ω) ,




3 + jω
where

1
ω 
r ω =√ and θ ω = − arctan .
32 + ω 2 3

1
3 θ(ω) π
2
r(ω)
ω
ω
π
2

Figure 64. Amplitude spectrum of f . Figure 65. Phase spectrum of f .

In real life, the signals f are real functions in the time domain. The Fourier transform
Z ∞ Z ∞ Z ∞


−jωu




F ω = f u e du = f u cos ωu du − j f u sin ωu du
−∞ −∞ −∞

of such signal satisfies the equation

Z ∞ Z ∞ Z ∞


−j(−ω)u




∗
F −ω = f u e du = f u cos ωu du+j f u sin ωu du = F ω .
−∞ −∞ −∞

Thus,


F −ω = F ω ,
i.e., the absolute value of the transform is an even function. If in addition f is an even function
in the time domain, then    
Im F (−ω) = Im F (ω) = 0,
which implies that


F −ω =F ω
which shows that the transform F is real and even. What if f is real and odd?

As ω increases the amplitude spectra of f decreases to zero (real and imaginary parts converge
to zero). Roughly speaking, this is due to the rapid oscillation of sine and cosine terms at large
frequencies (the areas bounded by the the graph that stay above and below the t-axis are roughly
the same). That is why a large portion of the information about f is contained in the narrow
spectrum band around the origin.
4.4 Properties of Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 73

4.4 Properties of Fourier transform

We next study properties of the Fourier transform. They are useful properties in system design
and analysis. We denote
n
o
n
o

F f t ω =F ω , F g t



ω = G ω , t, ω ∈ R, ja a, b ∈ C.

1. Linearity

n
o

F af t + bg t




ω = aF ω + bG ω

The linearity is a direct consequence of the linearity of integrals, i.e.,

n Z ∞

o

 −jωu
F af t + bg t



ω = af u + bg u e du
−∞
Z ∞ Z ∞
f u e−jωu du + b g u e−jωu du



=a
−∞ −∞



= aF ω + bG ω .

2. Scaling property

 
n
o
1 ω
F f at ω = F , a 6= 0
|a| a

Let us prove the scaling property for a continuous function f . The result easily extends to a
piecewise continuous signals. Let us substitute y = au to the integral
Z ∞
f au e−jωu du.


−∞

Depending on the sign of a 6= 0 we have two situations.

• If a > 0, then the upper and lower limits of the integral do not change.

• If a < 0, the upper and lower limits are swapped due to substitution. Swapping the limits
again so that we have lower bound −∞ changes the sign of the integral.

Thus,
Z ∞ Z ∞ Z ∞  

−jωu

y
−jω a 1 1
−j ω y 1 ω
f au e du = signum a f y e dy = f y e a dy = F .
−∞ −∞ a |a| −∞ |a| a
4.4 Properties of Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 74

In particular, if a = −1, then we observe that

n
o

F f −t


ω =F −ω .

♠Example 4.4. Let us find the Fourier transform of the function

g : g t = 5H t e−at + 2H − t eat , a > 0.

5 We denote

f t = H t e−at ,



2
t
which yields

1
Figure 66. Graph of g. F ω = .
a + jω
By using the linearity and the scaling property we obtain

5 2 7a − 3jω
G ω = 5F (ω) + 2F (−ω) = + = 2 .
a + jω a − jω a + ω2
♠

3. Differentiation with respect to time

We assume that the function f is continuous and that the assumptions of Theorem 4.1 hold for
f and its derivative f 0 . In addition, we assume that

lim f (t) = 0.
|t|→∞

Then

n
o

F f0 t


ω = jωF ω

Let us justify the property.

n
o

Z ∞
F f t0
f 0 u e−jωu du


ω =
−∞

∞ Z ∞
f u e−jωu − − jω f u e−jωu du




=
−∞ −∞

= I + jωF ω ,
4.4 Properties of Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 75

where
N  
CPV
f u e−jωu = lim f N e−jωN − f − N ejωN .




I = lim
N →∞ N →∞
−N


Since limN →∞ f ± N = 0, we have I = 0. Thus,
n
o

F f0 t


ω = jωF ω .

Assume that the assumptions of Theorem
4.1 hold for f and its derivatives f (k) , k = 1, 2, . . . , n,
and that f and f (k) , k = 1, 2, . . . , n − 1 are continuous and satisfy lim|t|→∞ f (k) (t) = 0. Ap-
plying the above result repeatedly yields

dn
 

n
F




n
f t ω = jω F ω .
dt

♠Example 4.5. Consider

g t = H t e−at + H − t eat , a > 0.

1
| {z
} | {z
}
=f t =f −t
t

Then

1 1 2a
G ω = + = 2 . Figure 67. Graph of g.
a + jω a − jω a + ω2




The function g is continuous if we set g 0 = 1 (f t + f − t does not have a defined value
t = 0 otherwise). Let us differentiate g with respect to time
g 0 t = −aH t e−at + aH − t eat , t 6= 0.

Its Fourier transform is

n
o
−a a 2ajω
F g0 t




ω = −aF ω + aF − ω = + = 2 2
= jωG ω .
a + jω a − jω a +ω
We obtain the result directly by using the differentiation property. Which way is simpler? ♠

♠Example 4.6. Let us find the Fourier transform of f : f t = H(t)te−at , a > 0.

It is continuous and satisfies Dirichlet conditions. Let us use the differentiation property to the
equation  
f 0 (t) = H(t) e−at − ate−at = H(t)e−at − aH(t)te−at
to obtain
1
jωF (ω) = − aF (ω).
a + jω
We can solve the transformation from the equation:
1
F (ω) =
(a + jω)2
♠
4.4 Properties of Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 76

4. Shifting properties

n
o

F f t−a ω = e−jaω F ω



time shift .
n
o

F e f t
jbt



ω = F ω−b frequency shift .

An alternative name for the time shifting property is time delay property. The frequency
shifting property has the alternative name modulation property. The physical interpretations
justify the names. We derive the formulas in the following way
n Z ∞

o

F f t−a f u − a e−jωu du


ω = ( substitution: u = y + a)
−∞
Z ∞
f y e−jω(y+a) dy


=
−∞
Z ∞
f y e−jωy e−jωa dy


=
−∞
Z ∞
−jωa
f y e−jωy dy = e−jωa F ω



=e
−∞

and
Z ∞
F ejbt f t ejbu f u e−jωu du




ω =
−∞
Z ∞
f u e−j(ω−b)u du = F ω − b .



=
−∞

♠Example 4.7. In Example 4.3, we computed the Fourier transform

F ω = 2AT sinc T ω

of the rectangular pulse

(




 A, |t| < T
f t = A rectT t = A H t + T − H t − T =
0, |t| > T.

t
T T

Figure 68. Rectangular pulse with height A over the interval − T, T .
4.4 Properties of Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 77

We use this to find the Fourier transform of the rectangular pulse




g t = π H t − H t − 2π .

t
2π

Figure 69. Rectangular pulse with height π over the interval 0, 2π .

We observe that we obtain g by shifting π rectπ t π time units to the right. Thus,



g t = π rectπ t − π

and according to the time shifting property

G ω = e−jωπ F ω = e−jωπ 2π 2 sinc πω .

In telecommunications, when transmitting a signal f over a large distance we want to reduce

the amount of energy required, or when having a multiple different signals in the same medium
we want to distinguish them in some way. For this we need a carrier wave.

♠Example 4.8. Let us assume that we are given f and its Fourier transform F . We form a new
signal g by multiplying f by cosine function with frequency α, i.e.,




g t = f t cos αt .

According to the linearity and phase shifting properties,

e + e−jαt
 jαt 
1 
1  1
1
G ω =F f t = F ejαt f t + F e−jαt f t = F ω−α + F ω+α .





2 2 2 2 2

F(ω)

G(ω)

ω ω
−α α

Figure 70. An arbitrary F and the corresponding G.

4.4 Properties of Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 78

We observe that that multiplication by cosine moves the graph of the transform to the neigh-
borhood of t = α and multiplies its height and width by one half. The information about f
is maintained as long as we know the carrier wave. Note that the negative frequencies do not
contain any further information. They are required because of our mathematical construction.

Exercise. If we multiply g by cosine again to built the function

k t = g t cos αt = f t cos2 αt ,








how does its Fourier transformation K ω look like? ♠

Exercise. Find the Fourier transformation of

g t = f t sin αt ,
and substitute it to the formula below.

n
o

F f 1
ω − α + 21 F ω + α ,




t cos αt ω = 2
F
n o
F f




t sin αt ω =

5. Convolution property

The convolution of continuous functions f and g is defined to be

Z ∞ Z ∞






(f ∗ g) t = f t − x g x dx = f x g t − x dx.
−∞ −∞

We observe that convolution in the time domain corresponds to the multiplication in the fre-
quency domain, and wise versa:

n
o

F (f ∗ g) t



ω = F ω G ω ,
n

o
1
F f t g t


ω = (F ∗ G) ω .
2π

Let us show the first formula:

∞ ∞
Z Z
−juω
g(u)e−juω F ω du




F ω G ω =F ω g(u)e du =
−∞ −∞
Z ∞ n
o
Z ∞ Z ∞
g(u)f t − u e−jtω du dt


= g(u)F f t − u du =
−∞ −∞ −∞
!
Z ∞ Z ∞ Z ∞ 

−jtω
(f ∗ g) t e−jtω dt


= g(u)f t − u du e dt =
−∞ −∞ −∞

n
o
= F (f ∗ g) t .
4.4 Properties of Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 79






♠Example 4.9. Consider f t = t H t −H t−1 and g t = rect1 t . Sketch the figures!
Their convolution is Z ∞





k t = (f ∗ g) t = f t − x g x dx.
−∞

We must split the integral into different intervals.

• If t < −1 or t > 2, then

Z ∞


k t = 0 dx = 0.
−∞

• If −1 < t < 0, then

t
(t + 1)2
Z



k t = − x + t dx = .
−1 2

• If 0 < t < 1, then

Z t


1
k t = − x + t dx = .
t−1 2

• If 1 < t < 2, then

1
1 − (t − 1)2
Z



k t = − x + t dx = .
t−1 2

♠Example 4.10. Consider a phenomenon that is modeled by the second order linear differential
equation
y 00 t + 5y 0 t + 6y t = f t ,







where the external force is f : f (t) = cos t . We consider only positive time instants, i.e.,
t > 0.
By using standard tools, we can find the solution of the differential equation



sin t + cos t
y(t) = C1 e−2t + C2 e−3t + .
10
Let us transform the differential equation to the frequency domain:

2




jω Y ω + jω5Y ω + 6Y ω = F ω

which simplifies to  
− ω 2 + jω5 + 6 Y ω = F ω ,

i.e.,

1

Y ω = F ω
−ω 2 + jω5 + 6
The function

−1 j2 1
G ω = 2
=

=


ω − 5jω − 6 ω − 2j ω − 3j 2 + jω 3 + jω
4.4 Properties of Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 80

is called the transfer function of the systems. The partial fraction decomposition of G is
1 1 −1
G(ω) =

= + ,
2 + jω 3 + jω 2 + jω 3 + jω

which shows that

g(t) = H(t)e−2t − H(t)e−3t .




Inverse transforming Y ω = G ω F ω back to the time domain yields
Z ∞
y(t) = g(t − x)f (x) dx.
−∞

Let us calculate the convolution

Z ∞ 
−2(t−x) −3(t−x)


y(t) = H(t − x)e − H(t − x)e cos x dx
−∞
Z ∞ Z ∞
−2(t−x)
H(t − x)e−3(t−x) cos x dx



= H(t − x)e cos x dx −
−∞ −∞

Z t Z t
−2t 2x −3t
e3x cos x dx



=e e cos x dx − e
−∞ −∞

t t
e2x 
 e3x 

= e−2t 2 cos x + sin x − e−3t



3 cos x + sin x
−∞ 5 −∞ 10

1




= 4 cos t + 2 sin t − 3 cos t − sin t
10
√
1
1
2

= cos t + sin t = cos t − π/4 .
10 10 10
This method gives a particular solution that satisfied the differential equation. It is not the gen-
eral solution. The value of the technique is not in solving the differential equation for which
there are easier ways. Engineers use the frequency domain representations in systems design,
because the calculations are relatively simple in the frequency domain. One can find convolu-
tions easily using Maple
√
or Matlab. However, we do not actually need to find it. It is sufficient
2 −jπ/4
to compute G(1) = 10 e . See Example 4.19. ♠

6. Symmetry

n 
o
n
o

F F f ω =F F t


t ω = 2πf − ω .

In some references the symmetry property is called

the duality property. Here we first trans-
form f to the frequency domain

obtaining F ω . We then consider the Fourier transform as a
"time domain function F t " and apply the transformation once more. It is really easy to get
4.4 Properties of Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 81

confused here!

Let us prove the formula. By the inverse Fourier transformation formula

Z ∞
1
ejωt F ω dω



f t =
2π −∞

or equivalently Z ∞
e−jωt F ω dω



2πf − t =
−∞

which shows that

Z ∞ n
o

e−juω F u du = F F t



2πf − ω = ω .
−∞

Again we must remember that the equality should be considered in the sense that the functions
are the same only whenever f is continuous. At the points of discontinuity we can have differ-
ent values.

Sometimes we use arrows pointing both ways to denote the transformation pairs. However,
such arrows are not commutative





f t ↔ F ω and F t ↔ 2πf − ω .

Remark. Stick to the roles of the variables; use t in the time domain and ω in the frequency do-
main. If you make an exception, write it down clearly. Small and large letters are not sufficient
to distinguish the function and its Fourier transform.

♠Example 4.11. Consider the time domain sinc-pulse

f t = sinc 3t .

We want to transform it to the frequency domain.

Remark. Finding the Fourier transform using only knowledge from elementary courses and
the definition is very hard. We can find the transform if we know that
Z ∞
sin(x) π
dx = .
0 x 2
However, it is easier to use the symmetry property. We recall that the sinc function is the Fourier
transform of a rectangular pulse:

F rectT t




ω = 2T sinc T ω .

According to the symmetry,

F 2T sinc T t




ω = 2πrectT − ω .

The square pulse function is even, so

F 2T sinc T t




ω = 2πrectT ω .
4.4 Properties of Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 82

Thus,

1 
F sinc 3t ω = F 2 · 3 · sinc 3t




ω
6
1

= · 2πrect3 ω
6
π


= H ω+3 −H ω−3 .
3

Just to compare,
1


π

rect3 t ↔ sinc 3ω and sinc 3t ↔ rect3 ω .
6 3

1
1
6
t ω
3 3 −π π

Figure 71. Time domain signal 16 rect3 t .

Figure

72. Frequency domain signal
sinc 3ω .
π
1 3

t ω
−π π 3 3

Figure 73. Time domain signal sinc 3t . Figure 74. Frequency domain signal
π


3
rect3 ω .
♠

♠Example 4.12. Let us consider the rectangular pulse

f t =H t −H t−6 .

Its Fourier transform is

F ω = e−j3ω 6sinc 3ω .

By the symmetry,
n
o


F e−j3t 6 sinc 3t



ω = 2πf − ω = 2π H ω + 6 − H ω .
4.4 Properties of Fourier transform MATH.APP.240 F OURIER METHODS / M e / 2020 83

t ω
6 −π π

Figure 75. Time domain signal H t −H t−6 . Figure
76. Frequency domain signal
e−j3ω 6sinc 3ω .

2π
6

t ω
−π π 6

Figure 77. Time domain signal e−j3t 6sinc 3t .



Figure 78. Frequency domain signal 2π H ω + 6 −


H ω .
♠

♠Example 4.13. Let us consider the signal




f t =t H t −H t−6 .

Its Fourier transform is

( −j3ω ω+e−j3ω −ej3ω
e−j3ω 6je



ω2
, if ω 6= 0 hard work
F ω =
18, if ω = 0

We consider this as a time domain function and use the symmetry property to obtain


F F t





ω = 2πf − ω = −2πω H ω + 6 − H ω .

18

t ω
6 −π π

Figure 79. Time domain signal f t . Figure 80. Frequency domain signal F ω .
4.5 Fourier transforms involving impulse functions MATH.APP.240 F OURIER METHODS / M e / 2020 84

12π
18

t ω
−π π 6

Figure 81. Time domain signal F t . Figure 82. Frequency domain signal 2πf − ω .
♠

4.5 Fourier transforms involving impulse functions

Let us consider Fourier transforms involving the Dirac delta function. The Dirichlet conditions
are not automatically satisfied. However, we can proceed by using the definition (care must be
taken).

♠Example 4.14. Let us find the Fourier transform of

f :f t =δ t−a .
According to the definition
Z ∞

Z ∞

f t dt = δ t − a dt = 1.
−∞ −∞

The integral is finite. The function f has only a finite number of discontinuities and local
extrema, so it satisfies the Dirichlet conditions. Thus,
Z ∞
δ u − a e−jωu du = e−jωa .



F ω =
−∞

♠Example 4.15. Let us find the Fourier transform of

f : f t = A.
The function does not satisfy the Dirichlet conditions. However, we can proceed by using the
symmetry property. If


g t = Aδ t ,
then its transform is
G(ω) = Ae−jω·0 = A.
Thus, n
o

F ω =F G t




ω = 2πg(−ω) = 2πAδ − ω = 2Aπδ ω .
♠
4.5 Fourier transforms involving impulse functions MATH.APP.240 F OURIER METHODS / M e / 2020 85

♠Example 4.16. Let us find the Fourier transforms of sine and cosine functions. They do not
satisfy the Dirichlet conditions. However, we can use the symmetry property. First we observe
that
δ t − a ↔ e−jωa and e−jta ↔ 2πδ − ω − a = 2πδ ω + a .

Thus,
n
o
n ejat + e−jat o

F cos at ω =F ω
2
1 n o
1 n o

= F ejat ω + F e−jat ω
2 2



=π δ ω−a +δ ω+a

and
n
o
n ejat − e−jat o

F sin at ω =F ω
2j

1 n jat o
1 n o

= F e ω − F e−jat ω
2j 2j



πδ ω − a πδ ω + a
= −
j j



= jπ δ ω + a − δ ω − a .

♠Example 4.17. Results of complex analysis (omitted) give the inverse transform

1
  Z ∞  2
, if t > 0,
1 1 1 jtu 
F −1


t = e du = 0, if t = 0,
jω 2π −∞ ju 
 1
− 2 , if t < 0.

It follows that

2
signum t ↔ , ω 6= 0.
jω

In addition, H(t) = 21 signum t + 12 , so we obtain the transformation pair

1 2 1
1

H(t) ↔ + 2πδ ω = + πδ ω .
2 jω 2 jω
♠
4.5 Fourier transforms involving impulse functions MATH.APP.240 F OURIER METHODS / M e / 2020 86

♠Example 4.18. The function

f t = H t e−at , a > 0

has a jump discontinuity at the origin (Figure 61 ), so we cannot use the differentiation property
(Property 3). By using the generalized derivative we get

f 0 t = −aH t e−at + δ t e−at .

Assuming that the property 3 is applicable

jωF ω = −aF ω + 1

or equivalently

1
F ω = ,
a + jω
which is the correct result. ♠

♠Example 4.19. Recall Example 4.10 where we found the following transfer function of a
linear system

1
G ω =

.
2 + jω 3 + jω
If we use Dirac delta function to obtain





cos at ↔ π δ ω + a + δ ω − a = F ω ,

we can transform



Y ω =G ω F ω
to the time domain
Z ∞ Z ∞

1

jωt 1


 jωt
y t = G ω F ω e dω = G ω π δ ω + a + δ ω − a e dω
2π −∞ 2π −∞

1

G − a e−jat + G a ejat .


=
2
We observe that

1
∗
G −a =

=G a .
2 − ja 3 − ja
Denote
G a = rejϕ ,

where



 

r= G a = G −a and ϕ = arg G a = − arg G − a .

Thus,
1
jat
−jat  ej(at+ϕ) + e−j(at+ϕ)

G a e +G −a e =r = r cos at + ϕ .
2 2
We used convolution to find a particular solution of a non-homogeneous differential equation in
Example 4.10. We observe that we obtain the same result by substituting a = 1. The example
4.6 Parseval’s identity MATH.APP.240 F OURIER METHODS / M e / 2020 87

shows why the transfer function is interesting. Its spectrum G ω determines the amplitude
of the cosine wave of the result.

Remark. The Fourier transform of r cos at + ϕ is

 

jϕ ω −jϕ jϕ



πre a δ ω + a + δ ω − a = πr e δ ω + a + e δ ω − a ,



that does not look anything like G ω F ω . We lose the information about the transfer function
due to the properties of Dirac delta function which only saves the local data at one point and is
zero elsewhere.
If F is a combination of Dirac delta functions, we have a simple way to determine the corre-
sponding y(t):








Y ω =G ω F ω =π G −a δ ω+a +G a δ ω−a


jϕ ω


= πre a δ ω+a +δ ω−a .

Using the shifting property we obtain

y(t) = r cos a(t + ϕ/a) = r cos at + ϕ .

4.6 Parseval’s identity

We showed that the Fourier series of a T -periodic real function f obeys the Parseval’s identity
Z d+T ∞
1 2

X 2
f t dt = cn ,
T d n=−∞

where cn is the coefficient of the term ejnωt in the Fourier series. The corresponding result for
the Fourier transform of a possibly non-periodic function f is

Z ∞ Z ∞

2 1
2

f u du = F ω dω Parseval’s identity .
−∞ 2π −∞

The proof for the identity is quite straightforward. First we use the convolution property to
obtain
Z ∞ Z ∞
n
o
1 1
du = F f t
2

−jωu 2




f u e ω = (F ∗ F ) ω = F ω − x F x dx.
−∞ 2π 2π −∞

Substituting ω = 0 shows that

Z ∞ Z ∞

2 1


f u du = F − x F x dx.
−∞ 2π −∞
4.6 Parseval’s identity MATH.APP.240 F OURIER METHODS / M e / 2020 88

∗
Since the transform F satisfies F − x = F x , we have
Z ∞ Z ∞
2

1
2
f u du = F ω dω.
−∞ 2π −∞

Earlier we observed that the average power of periodic functions was related to the Parseval’s
identity for Fourier series. The concept of average power does not make sense for non-periodic
signals. However, we use the concept of (total) energy , which is the number on both sides of
the Parseval’s identity, i.e.,

Z ∞ Z ∞
2 1

2

E= f u du = F ω dω Energy .
−∞ 2π −∞

The function
1
2
2
F ω or F ω ,
2π


with the frequency as its variable is called the spectral energy density of the signal f and its
graph is called the energy spectrum.

♠Example

4.20. In Example 4.11 we found that the Fourier transform of the signal f : f t =
sinc 3t is n
o
π
F sinc 3t


ω = rect3 ω .
3
The energy of the signal is
Z ∞ Z ∞

2 1 π
2
E= sinc 3t dt = rect3 ω dω.
−∞ 2π −∞ 3

We evaluate the easier integral, which in this case is

3
π2 6π 2
Z
1 π
E= dω = = .
2π −3 9 18π 3
♠

♠Example

14.21. The
frequency domain representation (i.e., the Fourier transform) of the signal
f : f t = 6 rect3 t is

F (ω) = sinc 3ω .
Its energy is distributed in the frequency domain according to the following figure.
4.6 Parseval’s identity MATH.APP.240 F OURIER METHODS / M e / 2020 89

The energy is
Z ∞  2
1

E= rect3 t dt
ω −∞ 6
Z ∞
2π 2π 1
2 1
3 3 = sinc 3ω dω = .
2π −∞ 6
Figure 83. Energy spectrum.
2π
The amount of energy on the frequency band |ω| ≤ 3
is
2π 2π
sin2 3ω
Z Z

1 3
2 2 3
sinc 3ω dω = dω
2π − 2π
3
2π 0 (3ω)2
2π
Z 2π
!
1 3 sin2 3ω 3 3 sin 6ω
= − + dω
9π 0 ω 0 ω

2π
!
4π
sin x
Z
1 3


= −3 sin 3ω sinc 3ω + 3 dx
9π 0 0 x

1

= − (0 · 0 − 0 · 1) + Si 4π
3π


Si 4π
= ≈ 0.1583232233.
3π

It is

6 Si 4π · 100
% ≈ 94.99393400%
3π
of total energy. ♠
5 Discrete-time Fourier transform (DTFT) MATH.APP.240 F OURIER METHODS / M e / 2020 90

5 Discrete-time Fourier transform (DTFT)

In Section 3, we approximated the Fourier series by the discrete Fourier transform based on the
samples of the given function. We can take a similar

approach with Fourier transform as well.
We assume that we only know a sequence f kh of samples of (possibly non-periodic) f
that are equally spaced in time (sampling period h). We should estimate the Fourier transform
F of f by using it.

In theory, the sequence covers the whole time axis, i.e., k runs through all integers. We assume
that the function has a Fourier transform (the assumption is satisfied, for example, if the func-
tion satisfies the Dirichlet conditions).

The Fourier transform of the function f is the integral

Z ∞
f t e−jωt dt.



F ω =
−∞

Consider the associated Riemann sum

∞
X
f tk e−jωtk ∆tk ,


k=−∞

where the time interval is split into intervals of equal length h, and the value of the function is
evaluated at the lower bound of each interval. The Riemann sum can be written in the form
∞
X
f kh e−jωkh h.


k=−∞

The function

∞
X
f kh e−jωkh




F̂ ω = h DTFT
k=−∞

is called the discrete-time Fourier transform (DTFT). Here we have constructed a continuous
function of variable ω using a discrete signal (or a sequence of samples).

♠Example 5.1. V Let us choose h = 1 and

(

ak , k ≥ 0,
x k =
0, k < 0, where |a| < 1, k ∈ Z.

Then
∞ ∞ ∞
X
−jωk
X
k −jωk
X
k 1
ae−jω , if ae−jω = a < 1.



X̂ ω = x k e = a e = =
k=−∞ k=0 k=0
1 − ae−jω

The resulting function of ω is continuous. ♠

5 Discrete-time Fourier transform (DTFT) MATH.APP.240 F OURIER METHODS / M e / 2020 91

Remark. In the literature, the sum

∞
X
x k e−jωk



X̂ ω =
k=−∞


is often called the discrete-time Fourier transform. Here the value x k of the signal is just a
term of a sequence and the index k expresses its location in the sequence, but not necessarily
the time instant t which is kh. We have used time scaling x(t) = f (ht), and the relation for the
transforms is

1 ω 
X̂ ω = F̂ .
h h


at time instant t = 0 and end it at t = b. Then we only observe

In practice, we start sampling
values on the interval 0, b . We assume that the values are zero elsewhere. We take N samples
on equally spaced time instants. The sampling interval is denoted h = Nb , and we get the
sequence
n oN −1 n



o
gk = f 0 , f h , f 2h , . . . , f (N − 1)h .
k=0
The discrete-time Fourier transform is (the rest of the terms have zero coefficients)
N
X −1 N
X −1

−jωkh
gk e−jωkh .


F̂ ω = h f kh e =h
k=0 k=0



The sum F̂ ω approximates the Fourier transform F ω of f . We will study the relation to the
discrete Fourier transform next, which shows that there are certain limitations for the accuracy
of the approximation. The accuracy is studied more closely in the next chapter.

We observe that
  N −1 N −1
2π 2π
gk e−j (ω+ h )kh = h
X X
gk e−jωkh e−j2πk = F̂ ω .


F̂ ω + =h
h k=0 k=0

This means that F̂ is periodic with period

2π
,
h
unlike the Fourier transform F it approximates. If we split the period into N subintervals and
take equally spaces samples of F̂ , we observe that
  N −1 N −1
2π X 2π
−j (n hN )
X 2π
F̂n = F̂ n =h gk e kh
=h gk e−jkn N = hGn ,
hN k=0 k=0

where Gn are the terms of the discrete Fourier transform, which we can evaluate efficiently
using FFT. To summarize, we have the relation
n oN −1 n oN −1
F̂n = hGn .
n=0 n=0

The symmetry and periodicity properties of the sequence implies that only approximately one
half of the sequence contains useful information. These problems are discussed in more details
in the next chapter.
6 Sampling frequency MATH.APP.240 F OURIER METHODS / M e / 2020 92

6 Sampling frequency

 N −1
Both discrete methods based on samples require evaluation of the sequence Gn n=0 that has
periodicity and symmetry properties. Because of the properties, half of the elements of the
sequence are just mirror images of those obtained earlier and do not contain any useful infor-
mation.



If we split the interval 0, b into N subintervals of length h = b/N , the frequency is on the
interval
2π
2π

− N −1 ≤ω ≤ N −1 ,
b b
so  
2π 2π 2π
− <ω< = the periof of the sequence G .
h h h
1
We call the reciprocal h
of the length of the sampling interval the sampling frequency.

If we want the discrete-time Fourier transform (or discrete Fourier transform) to estimate the
Fourier transform (series) accurately on the frequency band

−ωm < ω < ωm

we need the highest circular frequency to be greater than 2ωm , i.e., we have the condition
2π 1
2ωm < ⇔ ωm < · 2π.
h 2h
The limit frequency
1
fNyq =
unit Hz
2h
is called the Nyquist frequency (Harry Nyquist, (1889 - 1976) ). In terms of the circular fre-
quency ωNyq = 2πfNyq = πh with unit rad/s.

The above limit works usually relatively well, but not always as illustrated by the next exam-
ple. I.e., choosing h = π/ωm is not generally enough if we want good approximation up to
frequency ωm .



♠Example 6.1. We take samples on the interval 0, 16 . Let us choose h = 0.5 and N = 32.
The Nyquist frequency is 1 Hz, or as circular frequency ωNyq = 2π. We use FFT to obtain the
amplitude spectrum.

ω
6 12
6 Sampling frequency MATH.APP.240 F OURIER METHODS / M e / 2020 93

(31·2π)
Figure 84. Amplitude spectrum on the interval 0 ≤ ω ≤ 0.5·32
≈ 12.17.

The green line in the figure is the midpoint after which the spectra is just a mirror image of the
first half. According to the results, the greatest amplitudes are at frequencies ω ≈ 1.96 where
|c5 | ≈ 1.95 and arg(c5 ) ≈ −1.29, and ω ≈ 3.14 where |c8 | ≈ 0.95 and arg(c8 ) ≈ 0.07. By
using these frequency components we obtain the estimate

fˆ t ≈ 3.9 cos 1.96t − 1.29 + 1.9 cos 3.14t + 0.07 .

The original f would have been

f t = 4 sin 2t + 2 cos 3πt ,

which was not available when we found our estimate.

f (t)
f^(t)

t
16

Figure 85. Values of f and an inaccurate approximation.

Let
 us
proceed as above, but with larger sampling frequency. We take samples on the interval
0, 16 , and choose h = 0.25 and N = 64. Nyquist frequency is 2 Hz or as an circular fre-
quency ωNyq = 4π. We obtain the amplitude spectrum using FFT.

ω
12 24

(63·2π)
Figure 86. Amplitude spectrum over the interval 0 ≤ ω ≤ 0.25·64
≈ 24.74.

According to the obtained results the largest amplitudes are at the frequencies ω ≈ 1.96 where
|c5 | ≈ 1.95 and arg(c5 ) ≈ −1.28, and ω ≈ 9.425 where |c24 | ≈ 0.98 and arg(c24 ) ≈ 0.008 .
By using these two amplitudes we obtain the approximation

fˆ t ≈ 3.9 cos 1.96t − 1.28 + 1.96 cos 9.425t + 0.008 .

6 Sampling frequency MATH.APP.240 F OURIER METHODS / M e / 2020 94

f (t)
f^(t)

t
16

Figure 87. Values of f and an improved estimation.

If we take samples of cos 3πt at t = k 12 we obtain the sequence

  31   31   31

1 π 1
cos 3π k = cos 2πk − k = cos πk ,
2 k=0 2 k=0 2 k=0

which is exactly the same sequence that is obtained by using frequency π. This phenomenon is
called folding. The number of samples is such that at least two different harmonic components
can be fitted into the sequence of samples.

Let us take a closer look at this phenomenon. Consider the function f (t) = cos at + b let us
take N samples with sampling interval h. If ω is such a frequency that


cos ω · kh + c = f (kh), ∀k = 0, 1, 2, . . . , N − 1,

then
 
2πn   
cos ω± kh + c = cos ω · kh ± 2πnk + c , n ∈ Z,
h
 
= cos ω · kh + c = f (kh), ∀k = 0, 1, 2, . . . , N − 1.

In addition    
cos − ω · kh − c = cos ω · kh + c .

Thus, we can take a to be any number

2πn 2πn
ω± ∧ −ω ± , n ∈ Z.
h h
This choice determines the sign of c.

In order to avoid folding, we should have all frequency components of the function inside
the sampling interval, i.e., the amplitudes outside the interval are zero. However, we only
know the measured values so we cannot be sure. Every engineer should live with this kind of
uncertainty.
7 A summarizing example MATH.APP.240 F OURIER METHODS / M e / 2020 95

7 A summarizing example

We have encountered four different situations. The original function was either periodic or non-
periodic with continuous or discrete time variable. The related Fourier transforms had either
discrete or continuous variable.

Let us illustrate the different situations

by using a simple signal function that is suitable for
all situations not generally possible .

♠Example 7.1. Consider the signal

f t =H t −H t−π .

• We assume that the signal has period T = 3π, i.e.,

(

1, 0 < t < π
f t =
0, π < t < 3π



f t + 3π = f t .

We do not define the value at the points of discontinuity.

t
π 3π
T=3π

Figure 88. Graph of f .

The Fourier series of the signal is

∞ !
sin n 2π 2π



−
  

1 X 2 1 cos n 2
fˆ t = + 3
cos n t + 3
sin n t ,
3 n=1 πn 3 πn 3

or in more useful form

∞  
ˆ

X 2
f t = c0 + An cos n t + θn ,
n=1
3

where
r
1 p 1 2π


c0 = , An = a2n + b2n = 2 − 2 cos n , ja θn = arg an − jbn .
3 πn 3
The coefficients An show which frequency components are the strongest ones. Approxi-
mately

A1 ≈ 0.55, A2 ≈ 0.28, A3 = 0, A4 ≈ 0.14, A5 ≈ 0.11, A6 = 0, A7 ≈ 0.08, A8 ≈ 0.07, . . .

7 A summarizing example MATH.APP.240 F OURIER METHODS / M e / 2020 96

The amplitude spectrum is given in the figure below (we have dropped out the left hand
side, since it is the mirror image of the right hand side). Note that the frequency of the
signal is ω0 = 2/3

ω
2 4 6 8 10 12 14

Figure 89. Amplitude spectrum at frequencies ω = nω0 , n = 0, 1, 2, . . . , 21.

Let us use 5 first frequency components in our approximation.

t
π 3π

Figure 90. Graphs of f and the partial sum.

Here the time variable is continuous and the (angular) frequencies 2n/3 are discrete.
There is a number, e.g., amplitude, related to every frequency (Figure 89 ).

• Let us consider the situation where we assume that the function is non-periodic
(

1, 0 < t < π,
f t =
0, elsewhere.

t
π

Figure 91. Graph of f .

The function has the Fourier transform

π
 πω 
F ω = e−j 2 ω πsinc


.
2
Its values are complex numbers, but its absolute values and phase angles are real valued,

2
so we can plot them. Let us plot the graph of the square of the amplitude F ω .
7 A summarizing example MATH.APP.240 F OURIER METHODS / M e / 2020 97

|F(ω)|2

ω
4 2 2 4 6

Figure 92. Graph of the square of the amplitude.

The above figure shows the energy distribution with respect to frequencies. By using
appropriate mathematical software, we can calculate that approximately 95 percent of the
total energy is contained in the frequency band from zero to four, i.e.,
R4
2
−4
F ω dω
95 ≈ R ∞
2 · 100.
−∞
F ω dω
In this case, the continuous time variable transformed to a continuous frequency variable.

• Let us next consider a discrete time variable, which corresponds to having sampled data.
Then cannot use the actual rule of the function, since it is considered to be unknown. We
only have a sequence of samples at our disposal.

Let us take N = 32 samples over the time interval of length T = 3π. The samples are
equally spaced, so the sampling time is h = T /N = 3π/32 ≈ 0.2945. We have made
this choice in order to compare the results with the previous ones. Other choices of h and
N can be made and the length of the sampling interval would be T = N h in that case.

We obtain the sequence

n o n
o31 n o
gk = f kh = 1, . . . , 1, 0, . . . , 0 .
k=0 | {z } | {z }
11kpl 21kpl

Let us find an approximation for the Fourier series or the strongest frequency components
by using the sequence. The Fourier tranform is obtained by using FFT, and is
31
X
2πnk
Gn = f kh e−j 32 , n = 0, 1, . . . , 31.
k=0

If we only want to locate the strongest harmonic components, it is sufficient to evaluate

the absolute values Gn and choose the greatest ones. If we need to approximate the co-
efficients cn of the Fourier serieswe need to divide the absolute values by N . This gives
us the approximations
 
cn ≈ 0.344, 0.280, 0.133, 0.01, 0.075, 0.051, 0.011, 0.047, 0.031, 0.012, 0.037,
0.022, 0.013, 0.033, 0.018, 0.015, 0.031, 0.015, 0.018, 0.033, 0.013, 0.022,
0.037, 0.012, 0.031, 0.047, 0.011, 0.051, 0.075, 0.01, 0.133, 0.280 .

Do you observe anything peculiar about the values written in blue?

7 A summarizing example MATH.APP.240 F OURIER METHODS / M e / 2020 98

The real coefficients approach zero in a diminishing wave pattern. Let us choose the


frequency components from zero to five. The related phase angles are θn = arg Gn .
This gives the approximation
5  
ˆ

X 2π
f t ≈ c0 + 2 cn cos n t + θn ,
n=1
T

i.e.,

fˆ t ≈ 0.344 + 0.562 cos 32 t − 0.982 + 0.266 cos 34 t − 1.96 +

0.021 cos 36 t + 0.196 + 0.151 cos 83 t − 0.785 + 0.102 cos 10

3
t − 1.767 .

S5
1
Arvio

Figure 93. Partial sum and its approximation.

Exercise. Study how does the approximation changes if the harmonic components do not
have the frequencies n 32 that are the same as in the Fourier series. Choose some h and
use T = N h, which gives discrete frequencies n 2π T
. Draw a figure and compare to the
original function.

In this case we had a discrete time variable, and using the sample we obtained discrete


frequencies. The related transformation is called discrete Fourier transform DFT .

• Let us consider the same discrete time variable as above, i.e., we consider the sequence
n o n
o31 n o
gk = f kh = 1, . . . , 1, 0, . . . , 0 .
k=0 | {z } | {z }
11kpl 21kpl

By using

it, we obtain the following approximation of the original Fourier transformation
F ω :
X31
f kh e−jωkh .



F̂ ω = h
k=0

|F(ω)|
^
|F(ω)|

ω
21.33 21.33
7 A summarizing example MATH.APP.240 F OURIER METHODS / M e / 2020 99

Figure 94. Absolute values of F and F̂ .

2π

period h ≈ 21.33. We observe that starting from
The approximation F̂ is periodic with
ω = 10 the green line in the figure the approximated values do not correspond to the
values of F .

This transformation from discrete time samples

to a function of continuous frequency is
called discrete-time Fourier transform DTFT .

Remark. If we only need samples of F at certain frequencies, we can use the sequence
31
X
2πnk
Gn = f kh e−j 32 , n = 0, 1, . . . , 31,
k=0

which we obtained by using FFT with a small modification. If we multiply the terms by
the interval length h, we obtain the sequence
n o
hGn

which is the same as the sequence

31
X
f kh e−jωkh



F̂ ω = h
k=0
at frequencies
2π 2πN 2
ω=n =n = n , n = 0, 1, 2, . . . , 31.
Nh N 3π 3
That is indeed the case since
  31 31
2 X
−jn 2 k 3π X
2π
F̂ n =h f kh e 3 32 =h f kh e−j 32 kn = hGn .
3 k=0 k=0

|hGn|
^
|F(ω)|

ω
21.33


Figure 95. Illustration over one period. We have joined the terms hGn by a
straight line.

In the latter examples we sampled the function over the period of length 3π and took 32
samples. The time between subsequent samples was h = 3π/32. The frequency of F̂ in
the frequency domain is 2πh
, which means that the values of F̂ after the frequency πh are
just mirror images of the earlier values and do not approximate F . ♠
A Tabular. Selected Fourier transforms. MATH.APP.240 F OURIER METHODS / M e / 2020 100

A Tabular. Selected Fourier transforms.

In the tabular a > 0.

f t F ω

1
H t e−at


a + jω

1
H t te−at



2
a + jω

2a
e−a|t|
a2 + ω2

A rectT t 2AT sinc T ω

sinc t πrect1 ω

1 2πδ ω

δ t 1

2
signum t
jω

1

H t + πδ ω
jω

1

−jπsignum ω
t




cos at π δ ω+a +δ ω−a




sin at πj δ ω + a − δ ω − a
B Selected results on complex numbers and trigonometry MATH.APP.240 F OURIER METHODS / M e / 2020 101

B Selected results on complex numbers and trigonometry

The reader should have basic knowledge on trigonometry and complex numbers. We present
some important results here.

A complex number is of the form

z = a + bj, a, b ∈ R,

where a is the real part Re(z), b is the imaginary part Im(z), and j is the imaginary unit. Other
common notation for the imaginary unit is i, but we have chosen j since it is more common in
engineering literature. The imaginary unit satisfies

j 2 = −1.

We can illustrate complex numbers in the complex plane, and in some situations we can identify
it with the xy-plane.

y Im
(x, y) z
r r
θ x θ Re
−θ

z*

Figure 96. Real plane. Figure 97. Complex plane.

In the xy-plane we can write the points x, y on the origin centered circle with radius r in the
polar coordinates (

x = r cos θ


y = r sin θ .

In the complex plane the elements z on the origin centered circle with radius r can be written
in the form


z = a + bj = r cos θ + jr sin θ .

The angle θ is called the phase angle (argument) of z and denoted by

arg z = θ.

The complex conjugate of z is

z ∗ = a − bj = r cos − θ + jr sin − θ = r cos θ − jr sin θ .

The absolute value |z| of the complex number is the distance of z from the origin, i.e.,
√ √
|z| = a + bj = a2 + b2 = zz ∗ = r.
B Selected results on complex numbers and trigonometry MATH.APP.240 F OURIER METHODS / M e / 2020 102

Euler’s formula
ejθ = cos θ + j sin θ

is an important formula that can be proved using series. It implies that

z = z ej arg(z) = rejθ .

The above exponential form (and complex analysis in general) is a useful tool for an engi-
neer since the calculations using it can be much simpler than manipulation of trigonometric
functions. Most (but not all) of the related formulas of real functions generalize to complex
functions, e.g.,

k
ejθ ejϕ = ej(θ+ϕ) , ejθ = ejkθ , k ∈ Z.

By the Euler’s formula, 

 ejθ
 = cos θ + j sin θ

 e−jθ = cos θ
− j sin θ
,



so 
 ejθ + e−jθ = 2 cos θ




 ejθ − e−jθ = 2j sin θ
.



Thus we can write sine and cosine in the form

  

1 jθ −jθ
cos θ = e +e


2



 
1

 jθ −jθ


 sin θ = e −e .


2j

The values of sine and cosine at angles related to the following triangles should be known.

AA

π A

6 A
π
A
√ 4 2
A 2
2 1

A

√ A

3 A

A
π
π πA
4
3 3 A
1 1 1

In addition, the values of the trigonometric functions at multiples of pi should be known, e.g.,
(

1, if n is even
n

cos nπ = = −1 and sin nπ = 0.
−1, if n is odd
B.1 Some trigonometric identities MATH.APP.240 F OURIER METHODS / M e / 2020 103

B.1 Some trigonometric identities

1. sin2 x + cos2 x = 1,

2. sin − x = − sin x , 3. cos − x = cos x ,





4. sin x + π = − sin x , 5. cos x + π = − cos x ,





6. sin π − x = sin x , 7. cos π − x = − cos x ,
   
π
π

8. cos − x = sin x , 9. sin − x = cos x ,
2 2

10. cos x − y = cos x cos y + sin x sin y ,

11. cos x + y = cos x cos y − sin x sin y ,

12. sin x + y = sin x cos y + cos x sin y ,

13. sin x − y = sin x cos y − cos x sin y ,

14. sin 2x = 2 sin x cos x ,

15. cos 2x = cos2 x − sin2 x = 2 cos2 x − 1 = 1 − 2 sin2 x ,

   


x+y x−y
16. sin x + sin y = 2 sin cos ,
2 2
   


x−y x+y
17. sin x − sin y = 2 sin cos ,
2 2
   


x+y x−y
18. cos x + cos y = 2 cos cos ,
2 2
   


x+y x−y
19. cos x − cos y = −2 sin sin ,
2 2
 

1




20. sin x cos y = sin x − y + sin x + y ,
2
 

1




21. cos x cos y = cos x − y + cos x + y ,
2
 


1


22. sin x sin y = cos x − y − cos x + y .
2