Frequency-Domain Analysis: the Discrete Fourier

Series and the Fourier Transform

John Chiverton

School of Information Technology

Mae Fah Luang University

1st Semester 2009/ 2552

Outline

Overview
Lecture Contents

Introduction to Frequency-Domain Analysis

Discrete Fourier Series

Spectra of Periodic Digital Signals
Magnitude and Phase of Line Spectra
Other Types of Signals

The Fourier Transform

Aperiodic Digital Sequences
Lecture Contents

Overview
Lecture Contents

Introduction to Frequency-Domain Analysis

Discrete Fourier Series

Spectra of Periodic Digital Signals
Magnitude and Phase of Line Spectra
Other Types of Signals

The Fourier Transform

Aperiodic Digital Sequences
Outline

Overview
Lecture Contents

Introduction to Frequency-Domain Analysis

Discrete Fourier Series

Spectra of Periodic Digital Signals
Magnitude and Phase of Line Spectra
Other Types of Signals

The Fourier Transform

Aperiodic Digital Sequences
Frequency-Domain Analysis

I Point 1:
I Sinusoidal and exponential signals occur everywhere. Most
signals can be decomposed into component frequencies;
I Response of a LTI processor to each frequency component can
only alter the amplitude and phase, not the frequency;
I The output can then be found with the Principle of
Superposition.
I Point 2:
I If an input signal has a frequency spectrum and the LTI
processor has a frequency response then the output signal
spectrum is found by multiplication;
I Simpler than time-domain convolution.
I Point 3:
I Frequency response is a typical requirement for DSP algorithms
I Such as low-pass, bandstop or bandpass filters.
Continuous-time Fourier Analysis

I A signal can be decomposed into sinusoidal components;

I Even functions are composed of only cosine functions;
I Odd functions are composed of only sine functions;
I A finite number of frequency components can be used to approximate a
signal;
I Frequency components of a periodic signal are harmonically related with
discrete spectral lines, line spectrum, described by Fourier series;
I Fourier series can be expressed in exponential form;
I Aperiodic (non-repetitive) signals are decomposed into non-harmonically
related sinusoids. Resulting spectrum is continuous, described by Fourier
Transform;
I The inverse Fourier Transform reverses the process;
I The frequency response of an LTI system specifies how each sinusoidal or
exponential component is modified in amplitude and phase;
I Multiplication of the frequency response along with the input signal
spectrum gives the output signal spectrum.
Discrete-time Fourier Analysis

I Parallel set of Fourier techniques applicable to digital signals

I Two representations:
I A discrete-time Fourier Series, applicable to periodic digital
signals
I A discrete-time Fourier Transform, applicable to aperiodic
digital signals and LTI processors
Outline

Overview
Lecture Contents

Introduction to Frequency-Domain Analysis

Discrete Fourier Series

Spectra of Periodic Digital Signals
Magnitude and Phase of Line Spectra
Other Types of Signals

The Fourier Transform

Aperiodic Digital Sequences
Spectra of Periodic Digital Signals

I Periodic digital signal x[n] can be represented by Fourier

Series
I Line spectrum coefficients can be found using the analysis
equation:
N −1  
1 X −j2πkn
a[k] = x[n] exp
N N
n=0

where a[k] is the k spectral component or harmonic and N is

the number of sample values in each period of the signal.
I Regeneration of the original signal x[n] can be found using
the synthesis equation:
N −1  
X j2πkn
x[n] = a[k] exp .
N
k=0
Finding Line Spectra with a computer

I Remember Euler’s identity:

real part imaginary part

  z  }| { z }|
 {
−j2πkn 2πkn 2πkn
exp = cos −j sin
N N N
| {z }| {z }
Re(·) Im(·)
    
−j2πkn 2πkn
Re exp = cos
N N
    
−j2πkn 2πkn
Im exp = −j sin
N N
I The real Re(a[k]) and imaginary Im(a[k]) components can be
calculated individually
Finding Line Spectra with a computer

N −1 „ « N −1 „ „ « „ ««
1 X −j2πkn 1 X 2πkn 2πkn
a[k] = x[n] exp = x[n] cos − j sin
N n=0 N N n=0 N N

Two loops:
I A loop over n
I A loop over k

Pseudo-code:
1. Given x[n] and N :
2. Create a[k] size N and set all elements to zero a[k] = 0
3. For k=0 to N -1 % outer loop
3.1 For n=0 to N -1 % inner loop
3.1.1 Re(a[k]) = a[k] + x[n] × cos(2πkn/N ) % real part
3.1.2 Im(a[k]) = a[k] − x[n] × sin(2πkn/N ) % imaginary part
3.2 End For
3.3 Let Re(a[k]) = Re(a[k])/N and Let Im(a[k]) = Im(a[k])/N
4. End For
Finding Line Spectra Simple Example - manually
Let N = 2 and x[0] = 7, x[1] = 1. Find a[k]. To start find the real part of a[k]...
I To start Let n = 0 and k = 0, then

Re(a[0]) = Re(a[0]) + x[0] × cos(2πkn/2) = 0 + 7 × cos(2π0 × 0/2) = 7

I Increment n, so n = 1, then

Re(a[0]) = Re(a[0]) + x[1] × cos(2πkn/2) = 7 + 1 × cos(2π0 × 1/2) = 8

I Divide by N : Re(a[0]) = Re(a[0])/N = 4

I If we increment n any more then it will be larger than N − 1 so
Let n = 0 and increment k so that k = 1, then

Re(a[1]) = Re(a[1]) + x[0] × cos(2πkn/2) = 0 + 7 × cos(2π1 × 0/2) = 7

I Increment n, so n = 1, then

Re(a[1]) = Re(a[1]) + x[1] × cos(2πkn/2) = 7 + 1 × cos(2π1 × 1/2) = 6

I Divide by N : Re(a[0]) = Re(a[0])/N = 3

So the real part of a[k] is given by Re(a[0]) = 4 and Re(a[1]) = 3.

Finding Line Spectra Simple Example - manually
Finding the imaginary part of a[k]...
I To start Let n = 0 and k = 0, then

Im(a[0]) = Im(a[0]) − x[0] × sin(2πkn/2) = 0 + 7 × sin(2π0 × 0/2) = 0

I Increment n, so n = 1, then

Im(a[0]) = Im(a[0]) − x[1] × sin(2πkn/2) = 0 + 1 × sin(2π0 × 1/2) = 0

I Divide by N : Im(a[0]) = Im(a[0])/N = 0

I If we increment n any more then it will be larger than N − 1 so Let n = 0 and
increment k so that k = 1, then

Im(a[1]) = Im(a[1]) − x[0] × sin(2πkn/2) = 0 + 7 × sin(2π1 × 0/2) = 0

I Increment n, so n = 1, then

Im(a[1]) = Im(a[1]) − x[1] × sin(2πkn/2) = 0 + 1 × sin(2π1 × 1/2) = 0

I Divide by N : Im(a[0]) = Im(a[0])/N = 0

So the imaginary part of a[k] is given by Im(a[0]) = 0 and Im(a[1]) = 0.

Finding Line Spectra Another Example - but by computer
+2
0 1
B +1 C
−2
B C
B C
This time let N = 7 so that x = B +3 C = (+2 + 1 − 2 + 3 − 1 − 1 + 1)T or
B C
−1
B C
B C
@ −1 A
+1
x[0] = +2, x[1] = +1, ..., x[6] = +1. A plot of this periodic signal is given by
Finding Line Spectra Another Example - but by computer
+2
0 1
B +1 C
−2
B C
B C
This time let N = 7 so that x = B +3 C = (+2 + 1 − 2 + 3 − 1 − 1 + 1)T or
B C
−1
B C
B C
@ −1 A
+1
x[0] = +2, x[1] = +1, ..., x[6] = +1. A plot of this periodic signal is given by
Finding Line Spectra Another Example - but by computer

The line spectra in this case are found by a computer program using the algorithm
described earlier.

Notice (for the real coefficients) the mirror image, where a[1] = a[6] and a[2] = a[5]
etc.
Finding Line Spectra Another Example - but by computer

The line spectra in this case are found by a computer program using the algorithm
described earlier.

Notice (for the imaginary coefficients) a similar effect, except for a change of sign, e.g.
a[1] = −a[6].
Finding Line Spectra Another Example - but by computer

The line spectra in this case are found by a computer program using the algorithm
described earlier.

The line spectra are also periodic. The line spectra have the same periodicity as the
original signal.
Finding Line Spectra Another Example - but by computer

The line spectra in this case are found by a computer program using the algorithm
described earlier.

The line spectra are also periodic. The line spectra have the same periodicity as the
original signal.
Magnitude and Phase of Line Spectra

I The real and imaginary components are often quoted in terms of

magnitude and phase, or more commonly magnitude only.
I The magnitude can be calculated with:
p
Mag(a[k]) = Re(a[k])2 + Im(a[k])2

I and the phase with:

„ «
Im(a[k])
φ(a[k]) = tan−1 .
Re(a[k])

I The magnitude indicates the relative strength of the signal at different

frequencies;
I The phase indicates the phase angle of the signal at different frequencies;
I The strength of the signal at different frequencies is often the more
important information for a description of a system or a signal.
Magnitude and Phase of Line Spectra Example

Original Signal
x[n] = sin(2πn/64)
The real and imaginary components
contain separate information about
the frequency content of the signal.
The sine function is not present in
the real part.

Real Component Imaginary Component

Magnitude and Phase of Line Spectra Example

Original Signal

x[n] = sin(2πn/64)
The magnitude provides a
convenient overview of the
frequency content of the signal.

Magnitude Phase
Magnitude and Phase of Line Spectra Example

Original Signal

x[n] = sin(2πn/64) + cos(2πn/32)

The cosine part of the signal is

present in the real part.

Real Component Imaginary Component

Magnitude and Phase of Line Spectra Example

Original Signal

x[n] = sin(2πn/64) + cos(2πn/32)

The cosine frequency is not present

in the phase as it has zero phase.
The sine part is present as it has
±π/2 phase.

Magnitude Phase
Magnitude and Phase of Line Spectra Example

Original Signal

x[n] = sin(2πn/64) + cos(2πn/32)

+0.1 sin(2πn/4)

Real Component Imaginary Component

Magnitude and Phase of Line Spectra Example

Original Signal

x[n] = sin(2πn/64) + cos(2πn/32)

+0.1 sin(2πn/4)

Magnitude Phase
Magnitude and Phase of Line Spectra Example

Original Signal

x[n] = sin(2πn/64) + cos(2πn/32)

+0.1 sin(2πn/4) + 0.5 cos(2πn/16)

Real Component Imaginary Component

Magnitude and Phase of Line Spectra Example

Original Signal

x[n] = sin(2πn/64) + cos(2πn/32)

+0.1 sin(2πn/4) + 0.5 cos(2πn/16)

Magnitude Phase
Magnitude and Phase of Signals with Discontinuities

Original Signal x[n] = sin(2πn/26) I This signal is not periodic as

the sine function is analyzed
over ∼ 2 25 periods. The end
of the signal does not join up
with the beginning, resulting
in a discontinuity.
I The discontinuity has many
frequency components with
different phases.

Magnitude Phase
Magnitude and Phase of Impulse Function

Original Signal x[n] = δ[0]

I The line spectra of a (periodic) impulse
function is composed of all frequencies
I This illustrates the usefulness of an
impulse function in characterizing a
system’s frequency response
I Zero phase because the function is even,
i.e. x[n] = x[−n], frequency response
composed cosine functions only.

Magnitude Phase
Magnitude and Phase of Impulse Function

Original Signal x[n] = δ[1]

I The line spectra of a (periodic) impulse
function is composed of all frequencies
I This illustrates the usefulness of an
impulse function in characterizing a
system’s frequency response
I Phase components present when odd
function, i.e. x[n] = −x[−n], composed
of sine functions only.

Magnitude Phase
Useful Properties of Discrete Fourier Series

Parseval’s theorem
Equates the total power of a signal in the time and frequency domains:
N −1 N −1
1 X X
(x[n])2 = (Mag(a[k]))2
N n=0
k=0

Example
Impulse function, δ[0] = 1
N −1
1 X 1
(x[n])2 =
N n=0 N

and
N −1 „ «2
X 1 1
(Mag(a[k]))2 = N × =
N N
k=0

which are equal.

Other Example Useful Properties of Discrete Fourier Series

x[n] ↔ a[k] symbolizes a[k] is the discrete Fourier Series of x[n].

I Linearity:
If x1 [n] ↔ a1 [k] and x2 [n] ↔ a2 [k] then

w1 x1 [n] + w2 x2 [n] ↔ w1 a1 [k] + w2 a2 [k]

I Time-shifting (invariance):
If x[n] ↔ a[k] then

x[n − n0 ] ↔ a[k] exp(−j2πkn0 /N ),

i.e. The shift is just a phase shift and does not affect the
magnitude.
Outline

Overview
Lecture Contents

Introduction to Frequency-Domain Analysis

Discrete Fourier Series

Spectra of Periodic Digital Signals
Magnitude and Phase of Line Spectra
Other Types of Signals

The Fourier Transform

Aperiodic Digital Sequences
Aperiodic Digital Sequences
I Most signals do not endlessly repeat (i.e. not periodic);
I Most signals are therefore known as aperiodic.
I Different analysis and synthesis equations are necessary for
aperiodic sequences, known as the Fourier Transform for
aperiodic digital sequences
∞
X
X(Ω) = F(x[n]) = x[n] exp(−jΩn)
n=−∞

I and the inverse Fourier Transform for aperiodic digital

sequences
Z
−1 1
x[n] = F (X(Ω)) = X(Ω) exp(jΩn)dΩ.
2π 2π
I Note: X(Ω) is a continuous function. It is also periodic which
is a result of the ambiguities in discretely sampled signals.
Fourier Transform for Aperiodic Digital Sequences

Comparing the Fourier Transform:

∞
X
X(Ω) = x[n] exp(−jΩn), (1)
n=−∞

with the Fourier Series analysis equations:

N −1  
1 X −j2πkn
a[k] = x[n] exp . (2)
N N
n=0

We can see that Ω = 2πkN . n has also been taken to ±∞ and

because of this the Fourier Transform is no longer divided by N
(otherwise X(Ω) would be zero) so that X(Ω) can in some way be
equated with N a[k].
Fourier Transform Boxcar Example

The Fourier Transform of the impulse function δ[0]:

x[n] = δ[0]
X∞
∴ X(Ω) = δ[0] exp(−jΩn)
n=−∞
= exp(−jΩ × 0)
= 1.

In other words, the Fourier Transform of an impulse function

consists of all frequencies. Similar to the Fourier Series
representation of a periodic impulse function, calculated earlier.
Fourier Transform Example

0.2 if −2 ≤ n ≤ 2,
If x[n] = X(Ω) =
0 otherwise.
0.2 × (2 cos(Ω2) + 2 cos(Ω1) + 1).

F
⇒

Then

∞
X 2
X
X(Ω) = x[n] exp(−jΩn) = 0.2 exp(−jΩn)
n=−∞ n=−2

=0.2 × (exp(jΩ2) + exp(jΩ1) + exp(−jΩ0) + exp(−jΩ1) + exp(−jΩ2))

=0.2 × (cos(Ω2) + j sin(Ω2) + cos(Ω1) + j sin(Ω1) + 1
+ cos(Ω1) − j sin(Ω1) + cos(Ω2) − j sin(Ω2))
=0.2 × (2 cos(Ω2) + 2 cos(Ω1) + 1).
Periodicity of Fourier Transform

Also note that the Fourier Transform of an aperiodic signal is

periodic.

The periodicity is every 2π periods, a result of the sampling in the

digitisation process.
Frequency Response of LTI Systems

An LTI system has an input x[n] and an output y[n]:

Linear Time
Input, x[n] Output, y[n]
Invariant System

Recall (see lecture 2) that an LTI system has an impulse response, h[n]:

h[n], Linear Time

Input, x[n] Output, y[n]
Invariant System

which describes the response of the system when an impulse function is given
as the input. The impulse response is useful as it can be used to calculate the
output signal for a given input signal:

y[n] = x[n] ∗ h[n]

where ∗ is convolution NOT multiplication.

Frequency Response of LTI Systems

An LTI system can also be described in the frequency domain:

H(Ω ), Linear Time

Input, X(Ω ) Output, Y( Ω)
Invariant System

where
I The input frequency domain signal is X(Ω) = F(x[n]),
I The output frequency domain signal is Y (Ω) = F(y[n])
I The LTI system is described by H(Ω) = F(h[n]) which is known as the
frequency response of the system and is the Fourier Transform of the
impulse response.
In the frequency domain, the output can be calculated more easily:

Y (Ω) = X(Ω) × H(Ω),

where multiplication IS used here. In other words, convolution is performed by

multiplication in the frequency domain.
Frequency Response of LTI Systems

The frequency domain convolution (multiplication) equation:

Y (Ω) = X(Ω) × H(Ω),

can be re-arranged so that:

Y (Ω)
H(Ω) =
X(Ω)
so if we want to find the frequency response of a system then we can find it via
this equation or via the Fourier transform of the time domain representation
h[n].
Frequency Response of LTI Systems

Recall the general form of LTI difference equations (see lecture 2):
N
X M
X
a[m]y[n − m] = b[m]x[n − m].
m=0 m=0

Using the linearity and time-shifting properties of Fourier transforms we can

convert it to an expression using frequency domain terms:
N
X M
X
a[m] exp(−jkΩ)Y (Ω) = b[m] exp(−jkΩ)X(Ω).
m=0 m=0

Therefore the frequency response of a system can also be described by

M
P
b[m] exp(−jmΩ)
m=0
H(Ω) = N
.
P
a[m] exp(−jmΩ)
m=0

This equation can be used to directly find the frequency response of a system
even if only the coefficients a[m] and b[m] are known.
Frequency Response Example

Q. A moving average filter has y[n] = 13 (x[n] + x[n − 1] + x[n − 2]) . Find the
frequency response of this filter.
A. We can find the frequency response by using the coefficients:
I There is only 1 output coefficient, a[0] = 1.
I There are 3 input coefficients, b[0] = b[1] = b[2] = 1
3
I Therefore M
1 P
3
exp(−jmΩ)
m=0 1
H(Ω) = = (1 + exp(−jΩ) + exp(−j2Ω))
exp(−j0Ω) 3
1
= (1 + cos(Ω) − j sin(Ω) + cos(2Ω) − j sin(2Ω))
3
1
= (1 + cos(Ω) + cos(2Ω) − j(sin(Ω) + sin(2Ω)))
3
Frequency Response Example cont’d.

I Magnitude:
r
1
Mag(H(Ω)) = ((1 + cos(Ω) + cos(2Ω))2 + (sin(Ω) + sin(2Ω))2 )
3

I Phase: „ «
(sin(Ω) + sin(2Ω))
φ(H(Ω)) = tan−1 −
(1 + cos(Ω) + cos(2Ω))
The magnitude can be simplified using:

I 2 sin(Ω) sin(2Ω) = cos(Ω) − cos(3Ω) I sin2 (2Ω) = 1−cos(4Ω)

2
I 2 cos(Ω) cos(2Ω) = cos(Ω) + cos(3Ω) 1+cos(2Ω)
I cos2 (Ω) = 2
I sin2 (Ω) = 1−cos(2Ω)
2 1+cos(4Ω)
I cos2 (2Ω) = 2
Resulting in:
q
I Mag(H(Ω)) = 1
3
(3 + 2(2 cos(Ω) + cos(2Ω)))
Frequency Response Example cont’d.

Moving Average Filter (k=3) Frequency Response

Magnitude Phase

Mag(H(Ω)) = φ(H(Ω))
“ = ”
(sin(Ω)+sin(2Ω))
tan−1 − (1+cos(Ω)+cos(2Ω))
q
1
3
(3 + 2(2 cos(Ω) + cos(2Ω)))
Lecture Summary

This lecture has covered...

I Introduction to frequency domain analysis
I Discrete Fourier Series
I Spectra of Periodic Digital Signals
I Magnitude and Phase of Line Spectra
I The Fourier Transform for aperiodic digital sequences