Indian Institute of Technology Bombay

Dept of Electrical Engineering

Handout 8 EE 603 Digital Signal Processing and Applications

Lecture Notes 2 August 26, 2016

Fourier Series
Main reference material: Stein and Shakarchi, Fourier Analysis, 2003.
Most of you might have heard about Fourier expansion. This is a term so dear to signal
processing, a panacea for many problems there. A natural question (often forgotten) here is
“why Fourier analysis?”. Let us trace the roots and motivation behind Fourier expansion,
in particular Fourier Series and Fourier Transform.

0.1 Why Fourier?

A good example can answer this question logically. We don’t have to search far, remember
our coffee machine example from the last chapter (Example 6). Let us assume that the
exchange-rates are fixed and the machine accepts any currency as input. Let us also not
worry about conversion to numbers and available denominations, think that Indian currency
(rupee) is a real currency (all real values are possible and available).

Coffee
Input
Converter Machine Coffee
$,|,e, ¥ | α ml/|

An automatic currency converter can make a dump device like a coffee machine appear
intelligent. In signal processing parlance, we are searching for a converter which transforms
input signals to some currency signals. The currency signals have the property that they
can be manipulated quite easily by any given system, in particular linear systems. Following
Fourier’s trail will lead to such a signal representation.

Specifically, we like to write,

x(t) = ∑ αk φk (t), (1)

k∈Z

where x(t) is the input signal and φk (t) are the currency signals. Furthermore, we
insist that the currency signals obey

φk (t) ∗ h(t) = βk φk (t), ∀k ∈ Z, (2)

for any given linear system h(t).

In other words, convolution of h(t) and φk (t) amounts to simply scaling φk (t) by some
value βk (typically a complex number). Notice that βk depends on the system (or function)
h(t), while the scalars αk , k ∈ Z in (1) depend on the signal x(t).
 In a nut-shell, Fourier Analysis is an essential part of the Signal Processing toolbox,
which helps us solve even seemingly complex problems, by transforming them to simple
parallel problems with an additive structure, from which the global answers can be syn-
thesised. Let us now study an example, which illustrates the origin and fundamentals of
Fourier Series representation1 .

1 Analysing the Vibrating String

Let us start with the equation of the string that we did in class. A string of length L and
placed along the line segment from origin to the point (L, 0), with both ends fixed. It is
plucked, or placed at an initial position. Let u be the horizontal distance and s(u, t) be its
displacement in the vertical direction. We know that s(0, t) = s(L, t) = 0, ∀t. We showed
that the dynamics of a vibrating string can be approximated by

δ2s τ δ2s
= . (3)
δt2 ρ δu2

By taking c2 = τ /ρ,

1 δ2s δ2s
= . (4)
c2 δt2 δu2
By suitable scaling of the variables, we can equivalently solve

δ2s δ2s
= . (5)
δt2 δu2
Furthermore, let us take L = π, a suitable scaling can generalize the results to all L ∈ R+ .
There are many ways to solve such partial differential equations. The initial conditions
hold the key in resolving the coefficients of the general solution obtained by various means.
The fundamental constraint imposed here is by the initial position of the string. We
assume that the initial position is f (u), 0 ≤ u ≤ L for some meaningful function f (⋅). It so
happens that we need one more constraint or initial condition to completely characterize
the solution. Let us assume that the initial velocity of the string is given by the function
g(u). Another convenient simplification that we can do is to assume that the length of the
string is π. A linear transformation of the horizontal distance can then adapt the solution
to any value of L. Putting it all together,

Solve,

δ2s δ2s
= .
δt2 δu2
subjected to:

0≤u≤π
t≥0
s(u, 0) = f (u)
s′ (u, 0) = g(u)
1
Refer: Kreyszig: ‘Engg Maths’ or Stein and Schakarchi, ‘Fourier Analysis’
J. Fourier2 in the early 1800s advocated the solution of the above problem using the
method of separation. Notice that the left side of the p.d.e has derivative with re-
spect to (wrt) t and the right side wrt x. It seems natural to look for solutions to (5) which
factor into two functions, i.e.,

s(x, t) = w(t)φ(u). (6)

Substituting this in (5),

w′′ (t) φ′′ (u)

= . (7)
w(t) φ(u)
Since the LHS is a function of t and RHS that of u, equality in above implies that both
sides equal some real constant, say λ.

w′′ (t) − λw(t) = 0

φ′′ (u) − λφ(u) = 0 (8)

Exercise 1. Show that equation (7) implies the above equations.

Notice that λ > 0 cannot be a solution to the wave equation, as the resulting w is not
a wave at all.
Exercise 2. Show that λ > 0 is not a solution to our wave equation.
Thus assume that λ = −m2 for some m. It is easy to verify that Âm cos(mt), Âm ∈ R is
a solution to the time equation, and so is B̂m sin(mt). So the general solutions take the
form,

φ(u) = Ãm cos(mu) + B̃m sin(mu) (9)

m
˜ sin(mt).
w(t) = Ø cos(mt) + B̃ m (10)

Substituting the initial condition that φ(0) = 0 will imply

Ãm = 0. (11)

Further φ(π) = 0 implies that m is an integer. We can then re-scale the solution as,

s(u, t) = (Am cos(mt) + Bm sin(mt)) sin(mu). (12)

We know that if there are two solutions to the PDE, the sum of these solutions will also
satisfy the PDE. In particular, our p.d.e may be satisfied for several values of m. We can
then write,

s(u, t) = ∑ (Am cos(mt) + Bm sin(mt)) sin(mu). (13)

m∈Z+

How can we check that the above s(u, t) covers all the possible solutions. One way to verify
is to see whether u(x, t) satisfies the initial condition. If all the solutions are included in
(12), then we will have,

f (u) = ∑ Am sin(mu) , 0 ≤ u ≤ π. (14)

m≥1

2
his interest was more in heat equations
Notice that the initial position of the string is arbitrary (depends on the way one holds it).
So, if an arbitrary function (zero at both ends) can be expressed as a sum of sine waves as
in (14), then we have some confidence in our solution to the wave equation.
Many didn’t believe in the validity of this representation, and it was Fourier’s die-hard
belief which made the difference, and he applied this technique to some very important
problems of his time.
The representation in (14) is valid in the range [0, π], but we can extend this to [−π, π]
by making f (u) an odd function. i.e. f (−u) = −f (u). Since sine is an odd function, this
extension is immediate. On the other hand, if the function is even in the interval [−π, π],
we expect the representation in (14) to take the form,
feven (x) = ∑ Âm cos(mx). (15)
m≥0

Any function can be written as a sum of an odd function and even function.
1 1
f (u) = (f (u) + f (−u)) + (f (u) − f (−u))
2 2
= feven (u) + fodd (u)
= ∑ Âm cos(mu) + ∑ Am sin(mu)
m≥0 m≥1
1 1
= Â0 + ∑ (Âm − jAm )ejmu + ∑ (Âm + jAm )e−jmu (16)
m≥1 2 m≥1 2

Exercise 3. For an arbitrary function f (⋅), verify that 12 (f (u)−f (−u)) is an odd function.
Defining,
⎧
⎪ 1 (Âm − jAm ) , m ≥ 1
⎪
am = ⎨ 12 (17)
⎩ 2 (Â−m + jA−m ) , m ≤ −1,
⎪
⎪
we can rewrite,
∞
f (u) = ∑ am ejmx . (18)
m=−∞

Is this true for an arbitrary function?. “Oui”, said Fourier. We will come back to the
veracity of this statement shortly. Before that, we will show how to find the coefficients am
if the above formula is true.
π π
∫−π f (u)e
−jnu
du = ∫ ∑ am ej(m−n)u du = 2πan + ∑ ∫ ej(m−n)u du = 2πan . (19)
−π m≠n

Thus
1 π
an = f (u) exp(−jnu)du. (20)
2π ∫−π
The coefficient an is called the nth Fourier coefficient. Often we will replace an by fˆ(n)
to explicitly show its dependency on the function f . In signal processing scenarios, we
generally deal with the interval [− T2 , + T2 ] as opposed to [−π, π]. Let us stretch our function
f (u) to [− T2 , + T2 ] and write,
1 π T
am = ∫ f (u )e−jmu du
2π −π 2π
1 + T2 2π
−j 2π mv
= ∫ f (v)e T dv . (21)
2π − T2 T
 Fourier Series Coefficients
T
1 +2 2π
am = ∫ T f (u)e−j T mu du. (22)
T −2

Fourier Series Expansion

2π
f (u) = ∑ am ej T mu
. (23)
m∈Z

In spite of Fourier being adamant that his representation holds true for arbitrary func-
tions, it took another 150years for the issue to get settled, mostly in Fourier’s favor. We
will adopt a particularly elegant formalism as a fact, but this is one of those things which
we will not prove rigorously in this course.

2 Periodicity in Fourier Expansion

Consider a function f (u) for which the Fourier Series expansion holds good in the interval
− T2 ≤ u ≤ + T2 . What will happen if we simply consider u outside [− T2 , + T2 ]. At u′ = u + kT ,
2π 2π
mu′ mu j 2π
∑ am ej T = ∑ am ej T T
kT
e
m∈Z m∈Z
2π
= ∑ am ej T mu
(24)
m∈Z

Thus the FS expansion defines a function f (u) such that,

f (u + kT ) = f (u). (25)

In other words, f (u) is T − periodic. In general, text books and other material say Fourier
Series is for periodic functions. The meaning is that the FS expansion gives a periodic
function, even if the coefficients are evaluated for a function residing in [− T2 , + T2 ].

3 Uniqueness of FS Expansion
Is the FS expansion unique. The following theorem positively answers it except for some
number of possible discontinuities in the signal.

Theorem 1 (Stein and Shakarchi 2003). Suppose f is a T −periodic function which is

integrable (locally), with FS coefficients an = 0, ∀n ∈ Z. Then f (t) = 0 whenever f is
continuous at t.

Proof: The proof uses contradiction arguments. We will start with the assumption that
an = 0, ∀n. If the assertion of the theorem does not follow, i.e. f (t) ≠ 0 at some instant
t where f (⋅) is continuous, then we will show that an ≠ 0 for some n, thus violating
our original assumption. In other words, the assumption and the assertion has to co-
exist. These arguments also show a glimpse of the beauty of proofs in functional analysis.
Consider the function
2π
p(t) = ∑ αk exp(−j kt), (26)
k∈K T
where K is some finite set of indices. Clearly, since am = 0, ∀m, we get
T
1 2 l
∫ f (t) [p(t)] = 0, ∀l ∈ N. (27)
T −2T

1
By choosing α0 =  > 0, α1 = α−1 = 2 and αk = 0 for ∣k∣ ≥ 2, we get

2π
p(t) =  + cos( t). (28)
T
Denote
l
pl (t) ∶= (p(t)) . (29)

Take any point in t ∈ [− T2 , + T2 ], w.l.o.g assume t = 0. Otherwise we can shift our arguments
to any t. Also let f (t) be continuous at t = 0.
Suppose, contrary to the assertion of theorem, f (0) ≠ 0. Then assuming f (0) > 0 entails
no loss of generality since we can have the same arguments for f (t) or −f (t). We will now
show that an > 0 for some n ∈ Z, thus achieving the desired violation of the assumption.
By the definition of continuity, there exists 0 < δ ≤ T2 such that

1
f (t) ≥ f (0) , t ≤ ∣δ∣. (30)
2
Fix the above parameter δ. We can now choose the parameter  in (28) such that the
p(∣δ∣) = 1. Since p(0) = 1 +  and p(δ) = 1, we can choose a parameter µ ∈ (0, δ) such that
p(t) ≥ 1 + 2 whenever ∣t∣ ≤ µ. See the illustration to get the meaning of these parameters.

f (t)

f (0)
2

t
− T2 −δ +δ + T2
p(t)

1.0

t
− T2 −δ −µ +µ
+δ + T2
−1.0

T
2
∫− T f (t)pl (t)dt = ∫∣t∣≤µ f (t)pl (t)dt + ∫µ<∣t∣<δ f (t)pl (t)dt + ∫∣t∣≥δ f (t)pl (t)dt
2

= I1 + I2 + I3 (31)

We enumerate each of these integrals, and show that their sum diverges as l → ∞.
 1. ∣I3 ∣ < ∞
Since ∣p(t)∣ ≤ 1 when ∣t∣ ≥ δ (see Figure),

∣I3 ∣ ≤ ∫ ∣f (t)pl (t)∣dt (32)

∣t∣≥δ

 l
≤ (1 − ) ∫ ∣f (t)∣dt (33)
2 ∣t∣≥δ

→ 0 as l ↑ ∞, (34)
b
where we used the assumption that f (⋅) is locally integrable (∫a ∣f (t)∣dt ≤ ∞). Since
∣I3 ∣ ≤ ∞, there exist some c1 such that I3 ≥ c1 > −∞.

2. I2 ≥ 0
f (0)
Since f (t) ≥ 2 > 0 and pk (t) > 0 when ∣t∣ ≤ δ.

3. I1 → ∞

 l
∫∣t∣≤µ f (t)p l (t)dt ≥ (1 + ) f (t)dt (35)
2 ∫∣t∣≤µ
 l f (0)
≥ (1 + ) (36)
2 2
→ ∞ , as l ↑ ∞. (37)

Collecting the results together,

I1 + I2 + I3 → ∞.
Comparing with (27), I1 + I2 + I3 = 0 by assumption, which is invalid if we assume f (t) ≠ 0.
Hence the required contradiction is achieved, and the theorem is proved.

4 Examples
It seems that there has been too much talk and it is time to get wet. Our first example is
to find the Fourier Series of a train of rectangular signals. Imagine a goods train whistling
past. Let us find its frequency components.

Example 1. A rectangular signal, we will call it rectτ (t) is a symmetric waveform which
is defined as,

⎧
⎪ 1.0 if − τ2 ≤ t ≤
⎪ τ
2
x(t) = ⎨ (38)
⎩ 0 otherwise
⎪
⎪

− T2 − τ2 0 + τ2 + T2 t
Consider T −periodic repetitions of rectτ (t). i.e. f (t) = ∑n rectτ (t − nT ). Let us note down
a few points about f (t).

1. f (t) is periodic: we expect it to have a Fourier Series representation.

2. f (t) is T −periodic: only multiples of T1 are the frequencies present in Fourier Series
(we know this from the vibration of a string).

3. The FS coefficients are

T
1 2 2π
am = ∫ T f (t) exp(−j mt) dt. (39)
T −2 T

For the rectangle train in consideration, using 22

τ
1 +2 2π
am = ∫ τ exp(−j mt) dt
T −2 T
τ
1 exp(−j 2π mt) + 2
= − ∣ j 2π Tm ∣
T T − τ2
1 1 2π τ 2π τ
= 2π (exp(j m ) − exp(−j m ))
TjTm T 2 T 2
1 1 2π τ
= π sin( m )
T Tm T 2
τ
τ sin πm T
=
T πm Tτ
τ τ
= sinc(m ). (40)
T T
The second to last equality was obtained by multiplying both numerator and denominator
by τ . The last equality uses the definition of cardinal sine function.

sin(πx)
sinc(x) = . (41)
πx
Example 2. Consider a periodic train of impulses, denoted as ∆T (t).

∆T (t) = ∑ δ(t − nT ) (42)

n

The Fourier series coefficients are given by

T
1 2 2π
am = ∫ T δ(t) exp(−j mt) dt (43)
T −2 T
1
= . (44)
T
The FS coefficients corresponds to a periodic sequence of weighted Kronecker delta, at
multiples of the interval T1 .
5 A Fourier Currency on Signals
Our search initially was a currency for the signals, with a countable number of designations
and easy manipulations. Now we do have a representation of signals, particularly for
periodic ones, and for those have support in [− T2 , T2 ]. Remember that the denominations
are in integral multiples of the fundamental frequency, which is physically evident from
the vibrations of the string. This is good news. Since a signal can be represented by its
frequency components, let us study the effect of the system on a signal. Specifically, we
need to analyze the effect of each input frequency, represented as exp(jωt), ω = 2πf , when
passed through the system h(t). Then by using linearity of the system, we can add the
effect of each frequency to find the system output.

6 Fourier Transform
Consider a system h(t). We assume that

∫t ∣h(t)∣dt < ∞. (45)

This means that the system is integrable and we keep on with this definition for the rest of
the section. We have already learnt that an input x(t) to an LTI system h(t) will generate
an output,

y(t) = h(t) ∗ x(t). (46)

By using the formula for convolution and since x(t) = exp(j2πf t),

y(t) = ∫ h(τ )ej2πf (t−τ ) dτ (47)

= ej2πf t ∫ h(τ )e−j2πf τ dτ. (48)

τ

Notice that for a given input frequency f in Hz, the output is a constant multiple of the
input. That is surprising, but looks elegant. It also says that when we pass a frequency
through an LTI system, no new frequencies can appear at the output. Thus the input
frequency, unless filtered out, will come out of the LTI system. The effect of the system on
each frequency is given by,

H(f ) = ∫ h(τ )e−j2πf τ dτ. (49)

τ

In other words, with the input x(t) = exp(j2πf t),

y(t) = H(f )ejωt . (50)

We will call H(f ) as the Fourier Transform of h(τ ). This may ring a bell (maybe alarm)
to many. Equation (50) bears no similarity to the Fourier Series that we have learnt, at
least in the first inspection. Why should we call this Fourier Transform, since we already
felicitated Fourier by naming a series after him. On the other hand, equation (49) has a
striking resemblance to the Fourier Series expansion in (22). It looks like replacing f in
(50) by 2πm/T will make it similar to (22), but not precisely the same.
Exercise 4. Write down 3 visible differences between the formulas for Fourier Series and
Transform.
 The issue of nomenclature is one thing, but being partial is unbearable. Why do we
define just the Fourier Transform of the system h(t), and not for the signals. So in similar
vein, if x(t) is integrable, we define its Fourier Transform as

X(f ) = ∫ x(t) exp(−j2πf t)dt. (51)

t

This probably needs no explanation, we have learnt that a system and signal can be in-
terchanged in the convolution formula. However, the Fourier Transform of a system had a
physical meaning, i.e. the multiplicative gain when we pass a sinusoid through the system.
In case of signals, one such strategy is to visualize x(t) as a system for a moment and
imagine that we are passing various sinusoids through the system, and measuring the effect
at the output. X(f ) is precisely the multiplicative coefficient, possibly complex, for the
frequency f .
To make matters more clear, imagine our system being an impulse. Any sensible input
will come out unruffled through the system. We input a sinusoid of amplitude α, a sinusoid
of α comes out. So what is the Fourier Transform of an impulse (or Dirac delta). Though
the guess is correct, please verify by doing this exercise.

Exercise 5. Evaluate the Fourier Transform of a Dirac delta from its definition.

In retrospect, we defined the impulse that way to have the same response at all fre-
quencies.
Let us interchange our input and the system. A unit impulse, having all frequencies
equally weighted is passed through a system x(t). Assume x(t) to be continuous and
integrable. If x(t) is not periodic, the frequency contents at the output for a small frequency
interval df around f is X(f )df . On the other hand, if x(t) is periodic with period T , then
Fourier Series tells us that only those frequencies which are multiples of T1 will come out
of the system. From this point of view, it seems plausible that the Fourier Series and
Transform are intimately connected. Let us try the following procedure.

1. Compute the Fourier Transform X(f ) of a signal x(t) which is limited in [−L, L].

2. Repeat the signal in time with period T = 2L, to obtain xp (t).

3. The Fourier Series coefficients an of xp (t) are obtained as,

1 n
an = X( ). (52)
T T

The scaling by T1 in the last step is to avoid unwanted amplification. There is something
missing, we did not prove whether the algorithm for converting FT to FS is correct. The
validity of the claim follows from a celebrated theorem, which we will discuss in a later
section.
Let us see whether this approach works for some known cases.

Example 3. Recall the rectτ (t) function that we saw in Example 1.

⎧
⎪ 1.0 if − τ2 ≤ t ≤
⎪ τ
2
x(t) = ⎨ (53)
⎩ 0 otherwise
⎪
⎪

What is the Fourier Transform X(f ) of this signal. Let us note some points.
 1. Since the function seems to be well behaved and defined in a bounded time interval,
we expect it to have a Fourier Transform.
2. The lack of periodicity points that the frequency components can be continuous,
unlike the Fourier Series expansion.
3. The Fourier Transform can be obtained by the formula,

X(f ) = ∫ x(t) exp(−j2πf t) dt. (54)

t

In fact,
τ
2
X(f ) = ∫ τ exp(−j2πf t) dt (55)
2
1 τ τ
= exp(j2πf ) − exp(−j2πf )
j2πf 2 2
sin(πf τ )
=τ
πf τ
= τ sinc(f τ ). (56)
Once we go through these steps, it is quite clear that the Fourier Series is nothing but the
Fourier Transform sampled at multiples of T1 and scaled by T1 .
Example 4. Constant Function: We have seen that the Fourier Transform of an im-
pulse is a constant function. What about the Fourier Transform of a constant function. Do
we expect it to be an impulse.
Consider the rectangular function rectτ (t) in the previous example. If we take the width
τ to infinity, we can approximate a constant function. Intuitively, we expect the Fourier
Transform to be
X(f ) = lim τ sinc(f τ ). (57)
τ →∞

f (x) = 8sinc8f

f (x) = 2sinc2f

f (x) = sincf

Figure 1: τ sinc(f τ ) approaching an Impulse

For the strict minds, this is not an impulse for any value of τ , however arbitrarily large.
The function sinc(⋅) is not even integrable. Nevertheless, as τ ↑ ∞, an impulse like shape
emerges, as evident from Figure 1.
7 Energy of Signals
Upto this we didn’t worry about how much energy a signal has. This question so often
relates to parameters like battery life, transmission range etc in practical communica-
tion/signal processing situations. The rate of energy spent/received will give the power.
We found that an impulse or Dirac delta is a powerful concept. Is it full of energy too?.
The energy of a signal is given by

E(f ) = ∫ ∣f (t)∣2 dt. (58)

t

Exercise 6. What is the energy of a Dirac delta measure?. Do you find the answer sur-
prising?. Can you imagine a physical interpretation.

One concept that we should keep in mind when transforming to frequency is that,
Fourier Transform preserves the energy of the signals.

2 2
∫t ∣x(t)∣ dt = ∫f ∣X(f )∣ df. (59)

This has a straight forward proof, which we will see as we go on.

8 LTI Output: Convolution-Multiplication Theorem

If signals and systems have Fourier Transform, why not do it for the output too. After all,
the output is just another signal.

Y (f ) = ∫ y(t) exp(−j2πf t)dt (60)

t

Since y(t) = x(t) ∗ h(t),

Y (f ) = ∫ ∫ h(τ )x(t − τ )dτ exp(−j2πf t)dt (61)

t τ

= ∫ h(τ ) ∫ x(t − τ ) exp(−j2πf (t − τ + τ ))dtdτ (62)

τ t

= ∫ h(τ ) exp(−j2πf τ ) ∫ x(t − τ ) exp(−j2πf (t − τ ))dtdτ (63)

τ t

= ∫ h(τ ) exp(−j2πf τ )dτ ∫ x(t − τ ) exp(−j2πf (t − τ ))dt (64)

τ t
= H(f )X(f ) (65)

In writing this, we assumed that x(⋅) and h(⋅) are integrable functions 3 .

Exercise 7. Justify the steps above

Equation (65) is so important to us, that it warrants a statement as a separate theorem.

Convolution-Multiplication Theorem
Convolution of two time domain signals corresponds to multiplication of their respec-
tive Fourier Transforms in the frequency domain.

3
The interchange of integrals is justified by Fubini’s theorem in analysis