Version:

TEX’d:

11 THE DISCRETE Oct. 23, 2013

FOURIER
TRANSFORM AND
THE FFT

PREVIEW

Classical numerical analysis techniques depend largely on polynomial

approximation of functions for differentiation, integration, interpolation, and
solution of differential equations. Fourier techniques, as presented in
Chapter 8, use sinusoids and exponential functions to describe a function.
Moreover, the Fourier techniques lead to the possibility of understanding
physical phenomena in terms of the frequency components associated with
a system or signal.
In Chapter 8, the Fourier series and Fourier transform were applied to
signals and systems described by functions that are continuous functions of
time. For computer analysis, which is in fact the modern approach to
analyzing signals and systems, a signal or system is described by samples of
the continuous function associated with the signal or system. Fourier

507
 techniques are ideally suited to studying the effects of sampling such
continuous functions.
This chapter introduces the Discrete Fourier transform (DFT) and an
important algorithm to compute the DFT, called the Fast Fourier transform
(FFT). Our emphasis is on the practical use of the FFT and the errors that
can arise from sampling a signal and then computing its DFT. The final
Section 11.6 explains the FFT algorithm.

11.1 FREQUENCY ANALYSIS OF SIGNALS

One of the most important uses of Fourier analysis, although not the only
one by any means, is to analyze the frequency components of a signal
derived from measurements of a physical variable of interest. These sig-
nals can be functions that represent quantities changing with time, such
as voltage, force, or temperature. Light from a star can be analyzed to
determine the spectrum of the light from the star. Similarly, the Fourier
transform of radar echoes or audio (speech) signals is used to analyze the
characteristics of the source of the signals. The study of the frequency
content of such signals is called spectral analysis, and the Fourier trans-
form is the primary method to compute the spectrum analytically.
Before attempting to apply Fourier analysis by computer to physical
signals, it is necessary to thoroughly understand the frequency properties
of such signals. This is one area where the motto, “The purpose of com-
puting is insight, not numbers,” mentioned in the preface, must be taken
very seriously.1
First, suppose the physical signal of interest is a continuous func-
tion of time or another continuous variable. For computer analysis it is
necessary to represent the signal by a finite number of values of the sig-
nal, usually called samples of the signal. Second, the analytical form of
the Fourier methods, as presented in Chapter 8, assumes that the sig-
nal being analyzed is infinitely long in time. Obviously, any computer
representation of the signal must be finite in length. Thus, representing
a continuous signal for computer analysis leads to a finite-length vector
of sampled points with the possibility of various errors in representation.
Later in the chapter, we shall see that the spacing of the samples of a
time-dependent signal and the length of time of observation of the sig-
nal will determine the information that can be derived from frequency
analysis of the signal.
1
The motto is from Richard Hamming’s book, Numerical Methods for Scientists and Engineers.

508 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT

 As with any physical problem, some knowledge of the characteristics
of the system being analyzed must be introduced into the Fourier analysis
of the signals created by the system. For example, in human speech
processing, the frequencies of interest generally range from 0 hertz (dc)
to about 5000 hertz. White light, by contrast, has a frequency range
from about 0.4 × 1015 hertz to 0.7 × 1015 hertz. The range of frequencies
computed as the maximum frequency in the signal minus the minimum
frequency is called the bandwidth of the signal. Thus, a typical speech
signal would have a bandwidth of 5000 hertz. The bandwidth for white
light is about 0.3 × 1015 hertz.

EXAMPLE 11.1 Frequency Spectrum

Comparing the frequency spectrum of single pulses gives important insight
into the effect of changes in the characteristics of the pulse on the spectrum.
Consider the pulse 
A, |t| < T/2,
P (t) = (11.1)
0, otherwise.
According to the analysis in Chapter 8, the frequency spectrum of the pulse is
determined by the Fourier transform of the pulse as
sin(2πf T/2)
F [P (t)] = F (f ) = AT . (11.2)
(2πf T/2)

The effect on the spectrum F (f ) of changing the pulse width is shown

clearly in Figure 11.1. The first zero of the 2-second pulse occurs at f = 1/2
hertz.
Two pulses and spectrum Figure 9.1
1.5

1
Amplitude

0.5

0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
t seconds

1.5

1
F(f)

0.5

−0.5
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
f hertz

FIGURE 11.1 Rectangular pulses and their spectra

From Equation 11.2, the zeros of the frequency spectrum occur at the points

11.1 Frequency Analysis of Signals 509

 where 2πf T/2 = nπ and n is an integer. Solving for f leads to the zero points
as f = n/T , in which T is the total width of the pulse.
As the pulse width is decreased, the first zero crossing of the frequency
axis moves up in frequency. The spectrum of the narrower 1-second pulse has
its first zero at f = 1 hertz. If we assume that the frequencies of interest are
primarily contained in the region defined from 0 hertz to the first zero, the
bandwidth of the pulse for analysis could be considered to be approximately
1/T hertz. However, studying the analytical solution of Equation 11.2 shows
that the actual bandwidth is infinite.
The magnitude of the frequency spectrum for a pulse of width T seconds
is
sin(2πf T/2)
|F (f )| = AT ,
(2πf T/2)
which has zeros at f = n/T and maxima where

d sin(2πf T/2)
AT = 0.
df (2πf T/2)

Differentiating leads to the equation

  i
1 T 2πT/2 T
h  i h
AT − 2 sin 2πf + cos 2πf = 0.
f 2 f 2

After dividing by the cosine term, the equation for the maxima is
T T
h  i h  i
tan 2πf = 2πf ,
2 2
which becomes x = tan x by substituting x = 2πf T/2. The solutions are tab-
ulated in the Handbook of Mathematical Functions by Abramowitz and Stegun
listed in the Annotated Bibliography for this chapter. For a pulse with ampli-
tude A = 1 and width T = 1 second, Table 11.1 shows the first four frequencies
at which there are relative maxima of the magnitude of the spectrum. The
values at these points show that the magnitudes of the frequency components
diminish rapidly with increasing frequency. For practical purposes, the spec-
trum can be assumed to be zero above some frequency.

TABLE 11.1 Maxima of F (f )

f hertz |F (f )|
0.0000 1.0000
1.4303 0.2172
2.4590 0.1284
3.4709 0.0913

The ideal pulses just discussed cannot exist in nature since every physical
pulse has a finite rise time. The rise time is usually defined as the time for a

510 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT

 signal to change from 10% to 90% of its maximum value. A triangular pulse,
as shown in Figure 11.2, will be used to demonstrate the effect of the rise time
on its frequency spectrum.
Triangular Pulse.The triangular pulse of width T is defined by the equation
  
 A 1 − 2 |t| , |t| <
T
,
PT (t) = T 2 (11.3)

0, otherwise.

This pulse has the Fourier transform spectrum

AT sin2 (2πf T/4)

F (f ) = , (11.4)
2 (2πf T/4)2
with zeros at f = 2n/T . Comparing the spectra in Figure 11.1 and Figure 11.2
shows the frequency spectrum to the first zero for the triangular pulse is wider
than that of a square pulse of the same width. However, the higher-frequency
components in the spectrum of the triangular pulse go to zero more rapidly
than those of a pulse with a zero rise time.
Triangular pulse and spectrum Figure 9.2
1.5

1
Amplitude

0.5

0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
t seconds

1.5

1
F(f)

0.5

−0.5
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
f hertz

FIGURE 11.2 Triangular pulse and its spectrum

For the triangular pulse, the time taken for the signal to change from 0
to A is T/2 seconds, and the slope of the rising edge of the triangular pulse is
dPT /dt = 2A/T . Rewriting Equation 11.4 as
 sin2 (2πf T/4)
2A

F (f ) =
T (πf )2
shows that the magnitude of the frequency components is proportional to the
slope of the rising edge of the signal. In the case of typical pulses used in
modern electronic equipment, the signal could change from 0 to 5 volts in 10

11.1 Frequency Analysis of Signals 511

 nanoseconds (10−8 seconds) or less. Thus, significant frequencies greater than
1/10−8 hertz, or 100 megahertz, could be present in the spectrum of the pulse.

11.2 DISCRETE AND FAST FOURIER TRANSFORMS

In this section, we treat the approximation of the exponential Fourier

series and the integral Fourier transform of real-valued signals by sums
of finite lengths and then present an algorithm for efficient computation.
The discrete sum is called the discrete Fourier transform, or DFT, and the
algorithm is the fast Fourier transform, or FFT. Table 11.2 summarizes
the relationship between the time function and various Fourier techniques.
In the table, f(t) and f(ti ) are functions of time. The parameter ω = 2πf
radians per second is angular frequency, where f is frequency in hertz.

TABLE 11.2 Table of Fourier techniques

Name Characteristics Typical use

Fourier series f (t) continuous Analysis of periodic functions
F (ωi ) discrete and signals
Fourier transform f (t) continuous Frequency analysis of signals
F (iω) continuous and systems
Discrete Fourier f (ti ) discrete Computation of other transforms
transform (DFT) F (ωi ) discrete Analysis of sampled signals
Fast Fourier f (ti ) discrete Algorithm to compute the DFT
transform (FFT) F (ωi ) discrete

Both the DFT and FFT deal with discrete functions in time and
frequency. Thus, the DFT and the FFT transform a discrete sequence of
values fn , n = 0, . . . , N − 1, into another discrete sequence of values

Fk , k = 0, . . . , N − 1.

In this section, we discuss the DFT and FFT in order to interpret the
meaning of the Fk for practical signal processing.

512 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT

DISCRETE Consider the portion of a continuous signal f(t) shown in Figure 11.3.
FOURIER To determine the Fourier transform of f(t) by computer analysis requires
TRANSFORM that the signal be sampled at a finite number of points. Typically, signals
(DFT) are sampled at equally spaced points in time. In Figure 11.3, assume the
signal is sampled at intervals ∆t = Ts seconds to create a set of N points.
The length in time of the sampled signal is (N − 1)Ts seconds. Thus, the
original continuous signal is sampled and truncated for computer process-
ing.

f(t) f(nT s)

t
0
T
∆t = = Ts
N
T = NT s

FIGURE 11.3 Approximation of a signal by sampling

Definition of DFT and IDFT Assume that a function f(t) is defined

at a set of N points, f(nTs ) for n = 0, . . . , N − 1 values, as shown in
Figure 11.3. The DFT yields the frequency spectrum at N points by the
formula
  N−1
k X
Fk = F = f(nTs )e−i2πnk/N (11.5)
N Ts n=0

for k = 0, . . . , N − 1. Thus, N sample points of the signal in time lead

to N frequency components in the discrete spectrum spaced at intervals
fs = 1/(N Ts ). The Inverse DFT (IDFT) is defined as
N−1  
1 X k
fn = f(nTs ) = F ei2πnk/N (11.6)
N N Ts
k=0

for n = 0, . . . , N − 1. The IDFT is used to re-create the signal from its

spectrum.
Real f(t) For a real function of time sampled at an even number of
points N , the transform produces symmetry about the point N/2 be-
cause the real part of the transform is even and the imaginary part of
the transform is odd. Table 11.3 summarizes the properties of the DFT
applied to a real function of time. The DFT is assumed to have the form

F (f) = Fr (f) + iFi (f), (11.7)

11.2 Discrete and Fast Fourier Transforms 513

 where Fr (f) is the real part of the transform and Fi (f) is the imaginary
part. A real and even signal has a transform that is a real and even
function. Similarly, a real and odd signal produces an imaginary and odd
function for the transform, as described in Table 11.3.

TABLE 11.3 Properties of the DFT of f (t)

f (t) Fr (f ) and Fi (f )

Real Fr (f ) even; Fi(f ) odd

Real and even F (f ) = Fr (f ); even
Real and odd F (f ) = Fi (f ); odd

In this chapter, we consider only functions f(t) that are real functions
of time.
Frequency Range The DFT frequencies described by Equation 11.5
appear to range from f = 0 to f = (N − 1)fs . However, this is not
correct because of the symmetry of the transform results. Summarizing
the relationship between sampling in time and the frequency components
in the spectrum leads to the following conclusions:
1. The sample spacing in time Ts determines the highest positive
frequency in the DFT as
1
Fmax = ;
2Ts

2. The frequency spacing of the DFT components is

1 1
fs = = .
T N Ts

We stress these conclusions in discussing the DFT since it is neces-

sary to understand that the frequency components computed by DFT
analysis of a signal are determined by the choice of sample spacing and
number of samples, not necessarily by the characteristics of the signal
being analyzed. An important part of the analysis is to define the sam-
pling parameters correctly so that the DFT represents the spectrum of
the signal accurately. We shall reinforce these results in the chapter by
numerous examples, discussions, and problems at the end of the chapter.
Figure 11.4 presents a comparison of the Fourier series, Fourier trans-
form, and the DFT. The spectrum of a periodic signal computed as a
Fourier series consists of a sequence of frequency components spaced at
ω = 2πfs = 2π/T hertz apart, where T is the period of the signal in

514 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT

 seconds. The Fourier transform is used to determine the spectrum of an
infinite length signal which is continuous in time. The spectrum of the
Fourier transform and Fourier series is defined on the frequency interval
(−∞, ∞).

–T T t 0 2π ω
–– –– ––
2 2 T
Periodic function Fourier Series
Discrete Spectrum
(a) Fourier series

f(t) F(ω)

Continuous f(t) t ω
Fourier Transform
Continuous Spectrum
(b) Fourier transform

N t N ω
Discrete f(ti) Discrete Periodic
Spectrum (DFT)
(c) Discrete Fourier transform
FIGURE 11.4 Comparison of Fourier techniques

We find that use of the discrete Fourier transform implies periodicity

in both the time and frequency domain. This property of the DFT is
proven in several of the texts listed in the Annotated Bibliography at the
end of this chapter. In Problem 11.1, you are asked to show that the DFT
as defined in Equation 11.5 is periodic. A similar result holds for the time
function computed from the IDFT in Equation 11.6.
Thus, the discrete signal and its transform have the periodic proper-
ties

F (kfs ) = F ([k + N ]fs ) = F (kfs + Fp ),

f(nTs ) = f([n + N ]Ts ) = f(nTs + T ),

where Fp = 1/Ts is the period of the spectrum computed by the DFT.

Thus, it is easily shown by direct substitution in the DFT that the result
is periodic since
F (N fs ) = F0 .

11.2 Discrete and Fast Fourier Transforms 515

 If the discrete spectrum from the DFT is replicated as a periodic function
of period N fs defined over the entire real line, the result is called the pe-
riodic extension of the DFT. If the IDFT is used to compute the sampled
function of time from its spectrum, the reconstructed signal has a periodic
extension with period T seconds. This periodicity in time and frequency
is characteristic of the DFT regardless of the true characteristics of the
real signal being analyzed or its spectrum. It can be shown that the DFT
and the IDFT give the exact results, neglecting numerical errors, for any
discrete-time signal that is periodic with period N .

APPROXIMATION OF FOURIER TRANSFORMS

The DFT can be used to approximate the continuous Fourier transform.
As defined in Chapter 8, the continuous Fourier transform is
Z ∞
F [f(t)] = F (f) = f(t)e−i2πft dt. (11.8)
−∞

The frequency f in hertz is used as the parameter in this integral. The

function F (iω), where ω = 2πf is the frequency in radians per second,
could be calculated as well. For the transform of a physical signal, we
assume that f(t) = 0 for t < 0. Such signals are called causal . If the
signal samples need to be shifted in time to meet this restriction, only the
phase of the Fourier transform changes according to the shifting theorem
presented in Chapter 8.
Using the sampled f(t) with t = nTs and replacing f by the dis-
crete frequencies fs = k/(N Ts ) leads to the approximation of the Fourier
transform as
  N−1
k X
F = Ts f(nTs )e−i2πnk/N (11.9)
N Ts n=0

for k = 0, . . . , N − 1. The factor ∆t = Ts replaced dt in the integral and

is used as a multiplier of the DFT defined by Equation 11.5 in order to
approximate the continuous Fourier transform. Problem 11.2 presents
another derivation of the DFT approximation to the Fourier transform.
Various computer algorithms, called collectively fast Fourier trans-
forms have been developed to compute the DFT. Few of them take the
sampling time Ts in seconds into account since the typical FFT result,
F0 , F1, . . . , Fk , . . . FN−1 , is simply a function of the index k. Thus, the
spectrum must be interpreted and scaled properly if the spectral compo-
nents are to be displayed versus frequency in hertz.

FAST FOURIER There are many algorithms that are generally called fast Fourier transform
TRANSFORM algorithms. In fact, developing techniques to improve the efficiency of
(FFT) calculation for the Fourier transform has been an active research area for
many years. The algorithm is discussed in more detail in Section 11.6.

516 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT

 The FFT algorithms take advantage of the symmetry in the ex-
ponential functions exp(−i2πnk/N ) to reduce the number of computa-
tions while computing the DFT. For example, a direct calculation of the
DFT requires N 2 multiplications. The basic FFT requires approximately
N log2 N multiplications. If N = 4096 points, the FFT reduces the num-
ber of multiplications from more than 16 million to less than 50,000.
Figure 11.5 shows the important parameters for the DFT and the
common FFT algorithms when applied to physical signals. In the figure,
the magnitude of the spectrum computed by the FFT is plotted as
p
|F (f)| = [Fr (f)]2 + [Fi (f)]2

using the notation of Equation 11.7. In the upper plot, the DFT com-
ponents are plotted for the index k = 0 to k = N − 1 which emphasizes
the symmetry about the index value k = N/2. The k-th positive digital
frequency has the value F = k/N . The digital frequency, described in
Section 10.9, corresponding to Fmax is F = 1/2.
|FFT|

Positive Negative
Frequencies Frequencies

0 Fmax Frequency 0
012 ... N/2 Harmonic N

Negative Positive
frequencies frequencies

–fs 0 fs Fmax
0 Harmonic N/2
FIGURE 11.5 Discrete Fourier transform spectrum

The output of most FFT algorithms is folded in frequency, as shown

in the upper plot of Figure 11.5 and the DFT spectrum is symmetrical
around the frequency Fmax . This is called the folding frequency when the
symmetry of the spectrum is being discussed.
Folding Frequency The theory of the folding shown in Figure 11.5
is that N points in time produce N points in the frequency domain.
However, for a real signal, there are N complex numbers in the transform

11.2 Discrete and Fast Fourier Transforms 517

 with N real parts and N imaginary parts. The real part is even and the
imaginary part is odd around the folding point as indicated in Table 11.3.
Including the component at f = 0, there are really only N/2 + 1 unique
points on the positive frequency axis.
Symmetry about the origin It is often convenient to plot the DFT
spectrum showing the symmetry about the origin as in the lower plot of
Figure 11.5. In terms of the components, the DFT of a real sequence
possesses conjugate symmetry about the origin.
The procedure for plotting is to shift the upper half of the spectral
components from k > N/2 to k = N − 1 to negative values k − N so that
the origin represents the zero of frequency. The results for k > N/2 are
thus negative frequency results. The component at N/2 corresponds to
the maximum positive frequency.
DFT summary Table 11.4 summarizes the DFT or FFT parameters
when a real signal is sampled every Ts seconds for (N − 1)Ts seconds.

TABLE 11.4 DFT parameters

Parameter Notation
Time domain:
Sample interval Ts (s)
Sample size N points
Length (N − 1)Ts (s)
Period (from IDFT) T = N Ts (s)
Frequency domain:
1 1
Frequency Spacing fs = = (Hz)
T N Ts
Spectrum size N components
N
Maximum frequency fs = Fmax (Hz)
2
1
Frequency period F p = N fs = (Hz)
Ts

The two important parameters in the time domain are the sampling
interval Ts and the number of sample points N since other parameters can
be derived in terms of these. Be careful to not confuse the period of the
reconstructed signal from the IDFT, T = N Ts , with the total sampling
time for the actual time signal. Also, the maximum frequency in the
spectrum is (N/2)fs = 1/(2Ts ) hertz, but the spectrum is periodic with
period N fs = Fp hertz.
DFT frequency in hertz In the frequency domain, the spectral lines
of the DFT are spaced a distance fs hertz or ωs = 2πfs radians/second

518 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT

 apart. The frequency resolution is thus fs = 1/(N Ts ) hertz. As pre-
viously discussed, the spectral values for k > N/2 are simply negative
frequency results. For example, F [(N − 1)fs ] in the FFT output corre-
sponds to the frequency component F (−fs ).
The practical consequences of the characteristics of the DFT and
FFT are presented later in the section discussing practical signal analysis.
First, we present computer examples to demonstrate the use of MATLAB
commands to compute the DFT.

11.3 MATLAB FOURIER COMMANDS

MATLAB contains a number of commands to compute, manipulate and

plot the DFT of a function. These are listed in Table 11.5. Except for the
command fourier, the commands are used for numerical computation.
The symbolic command fourier is part of the Symbolic Math Toolbox . If
a function can be defined symbolically, fourier computes the Fourier inte-
gral transform. The Signal Processing Toolbox has additional commands
for more advanced signal processing.

TABLE 11.5 MATLAB commands for frequency analysis

Command Result
abs Magnitude of FFT
angle Phase angle of FFT in radians (−π to π)
fft FFT
fftshift Moves zero frequency to center of spectrum
ifft Inverse FFT
nextpow2 Returns next power of 2
unwrap Unwraps phase angle beyond −π to π
stem Plots discrete sequence data
fourier Symbolic Fourier transform

The MATLAB commands fft and ifft compute the DFT and inverse
DFT, respectively, without regard for the actual sample spacing. The
input to the functions is a given number of samples of the time function
for fft or frequency function for ifft. Thus, the discrete functions that
result from MATLAB Fourier computations functions must be scaled if
an answer in physical units, such as hertz, is desired.
The result of fft is generally an array of complex numbers. The com-
mands abs and angle can be used to compute the magnitude and angle

11.3 MATLAB Fourier Commands 519

 of the complex values, respectively. For plotting a two-sided transform,
the command fftshift will rearrange the output of fft to move the zero
component to the center of the spectrum.

INDEXING AND MATLAB vectors are indexed from 1 to N , where N is the number of ele-
TRANSPOSE ments. However, the standard definition of the DFT, as in Equation 11.5,
is indexed from 0 to N −1. Care must be taken in indexing when analyzing
and plotting the results of MATLAB FFT calculations.
If the MATLAB transpose operator (’) is used on a complex vector,
the operation gives the complex conjugate transpose; that is, the sign of
the imaginary part is changed as part of the transpose operation. This is
not generally desired if the transpose of the DFT results are to be used
since the sign of the phase angle would be changed. The operator (.’)
transposes a complex array but does not conjugate it.

EXAMPLE 11.2 Function to Compute DFT

The accompanying MATLAB script presents the function clfftf, which
computes the approximate Fourier transform and the magnitude and phase of
the two-sided spectrum of the function ft, defined at N sample points. The
variable Ts defines the sample interval in seconds. Notice that the MATLAB
command fft is multiplied by the sampling interval to approximate the actual
Fourier transform.

MATLAB Script
Example 11.2
function [FT,FTmag,FTang] = clfftf(ft,N,Ts)
% CALL: [FT,FTmag,FTang] = clfftf(ft,N,Ts) Compute the DFT
% approximation of the Fourier Transform
% Inputs:
% ft Sampled function of time f(nTs)
% N Number of sample points
% Ts Sample interval in seconds
% Outputs:
% FT Approximate Fourier transform using DFT
% FTmag Magnitude of spectrum
% FTang Phase in degrees
% Determine the two-sided spectrum
FT1=Ts*(fft(ft,N)); % Scale to approximate FT
FT=fftshift(FT1); % Shift 0 to center
%
% Compute the magnitude and phase for the frequency values
% in hertz fs=1/(N*Ts); fmax=1/(2*Ts)
%
FTmag=abs(FT); % Magnitude
FTang=(180/pi)*angle(FT); % Phase in degrees

520 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT

 This function is used to compute the Fourier transform in other exam-
ples in this chapter. The file can also be modified to provide the spectrum
in radians per second or to plot the results.
You will notice in this function and other scripts in this chapter that
calculation of the FFT requires only a call to the MATLAB command
fft. Most of the other executable statements scale or shift the result for
plotting in physical units of frequency.

DFT OF THE The decaying exponential function is a good test function to use for ex-
EXPONENTIAL ploring the accuracy and problems of the DFT applied to a continuous
FUNCTION function.

EXAMPLE 11.3 FFT Computation of Spectrum

Consider the function

e−t, t ≥ 0,
f (t) =
0, t < 0,

with the Fourier amplitude spectrum

1
F (ω) = p ,
1 + (2πf )2

as computed in Chapter 8. The phase of the spectrum is

2πf
 
θ(f ) = − tan−1 .
1
The accompanying MATLAB script is used to compute the DFT of the
function. The magnitude and phase of the result are plotted in Figure 11.6.
After the number of points is input and the various parameters are computed,
the FFT of the sampled signal is computed using the function clfftf described
in Example 11.2. Even though the computed spectrum is discrete, the function
plot, which interpolates through the points, is used when a large number of
points are to be plotted. For a small number of plotted points, the MATLAB
plotting command stem could be used.
A comparison of the DFT and exact Fourier transform plotted in the figure
shows a close agreement in the amplitude spectrum. Experimenting with the
number of points or the length of the period (T) in this example will show that
taking a shorter interval or fewer sampling points leads to decreased accuracy.
MATLAB Script
Example 11.3
% EX11_3.M Compute and plot the DFT of f(t)=exp(-t)
% Creates f(t) sampled each Ts seconds for T seconds
% Input: N -number of points input
% T -Period of signal
% t0 =0 -start of time points
% Calls clfftf to compute DFT
N=input(’Number of points N= ’) % Sample N points
T=input(’Period of signal T= ’)

11.3 MATLAB Fourier Commands 521

 Ts=T/N; % Sampling interval
% Form the vector of time points and f(n*Ts)
t0=0; % Start of signal
ts=(t0:Ts:Ts*(N-1)); % Compute N points
ft=exp(-ts);
% Determine the spectrum
[Fft,Ffmag,Ffang]=clfftf(ft,N,Ts);
% Compute the frequency values in hertz fs=1/(N*Ts); fmax=1/(2*Ts)
%
fs=1/(N*Ts); % Frequency spacing
f=fs*linspace(-N/2,N/2-1,N); % N points in frequency
% Plot Fexact and DFT result
w=2*pi*f;
Fexact=1./(sqrt(1+w.^2)); % Magnitude
Thetaex=-(180/pi)*atan(w); % Angle in degrees
clf
subplot(2,1,1),plot(f,Fexact(1:N),’--’,f,Ffmag(1:N));
title([’FT and DFT of exp(-t), N=’,num2str(N), ’ T= ’,num2str(T),’ sec’])
xlabel(’Frequency in hertz’)
ylabel(’FT and DFT’)
legend(’FT’,’DFT’)
subplot(2,1,2),plot(f,Thetaex(1:N),’--’,f,Ffang(1:N));
xlabel(’Frequency in hertz’)
ylabel(’Phase FT and DFT’)
legend(’FT’,’DFT’)

FT and DFT of exp(−t), N=64 T= 8 sec

1.4
FT
1.2 DFT
1
FT and DFT

0.8

0.6

0.4

0.2

0
−4 −3 −2 −1 0 1 2 3 4
Frequency in hertz

100
FT
DFT
Phase FT and DFT

50

−50

−100
−4 −3 −2 −1 0 1 2 3 4
Frequency in hertz
−t
FIGURE 11.6 Magnitude and phase of spectrum of e

522 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT

 DFT Phase Errors One potential problem with the DFT is that the
signal being analyzed appears as a periodic function with period N Ts ,
where Ts is the sampling interval. Considering the decaying exponential
in Example 11.3, the function is actually zero for all t < 0. However,
applying the DFT to the time function can be viewed as taking the Fourier
transform of a periodic waveform of exponential pulses.
The effect of periodicity of the analyzed signal was not obvious in com-
puting the magnitude of the transform in Example 11.3. However, com-
paring the computed phase with the correct value (− arctan 2πf) shows
that the computed phase angle can be badly in error at the higher fre-
quencies. The true phase changes from π/2 at negative frequencies to
−π/2 at positive frequencies.
Caution. The function angle computes the arctangent of the ratio of
the real to the imaginary part of a complex number. However, the result
is very sensitive to the magnitude and sign of the elements. Values close
to the minimum representable by the computer may cause unpredictable
variations in the phase angle of the DFT when the command angle is
used. These minimum values should be zero but are not due to roundoff
errors. In some cases, it is best to plot the real and imaginary part of the
DFT and set to zero any values that are below a predefined amount. A
value such as 10×eps could be used as the minimum allowed value, where
eps is the smallest MATLAB number for a particular computer.

11.4 PRACTICAL SIGNAL ANALYSIS

Assume that a physical signal is to be analyzed using the DFT to deter-

mine the spectral components. Figure 11.7 presents a simplified diagram
of the input stage of a data acquisition system.

Bandlimited
analog signal

f(t) Anti-Aliasing
Filter ADC Processor

nTs Sampled
|F( f )| data signal DFT
1
Ts < ––
2B
0 B Hz
fs > 2B

FIGURE 11.7 Data acquisition system

11.4 Practical Signal Analysis 523

 The signal will be assumed to be a continuous-time signal, which is
often referred to as an analog signal. A system that converts the analog
signal to a sequence of time samples suitable for computer processing is
called a data acquisition system. The first component is a filter that elim-
inates unwanted frequency components of a signal. In this case, the filter
limits the frequencies in the signal to the range 0 to B. A signal filtered
in this way is said to be bandlimited to B hertz. In a data acquisition
system, the filter is sometimes called a presampling filter . Such a filter is
also called an anti-aliasing filter because it is intended to prevent an error
called aliasing, as explained later after the sampling theorem is presented.
The analog-to-digital converter (ADC) samples the filtered signal each
Ts seconds to create a sampled-data or digital signal, which is the discrete-
time representation of the analog signal after sampling. Typically, the
samples are stored in a computer’s memory for processing.

SAMPLING Two of the most important questions in the specification of a data acqui-
sition system such as that shown in Figure 11.7 are the following:

1. How often should the analog signal be sampled?

2. How long should the signal be sampled?

If the highest frequency of interest in the signal is B hertz and the

frequency spectrum of the signal is limited to B hertz by the anti-aliasing
filter, the sampling theorem answers the question in Part 1. The theo-
rem is the cornerstone of practical and theoretical studies in electronic
communication.
Sampling Theorem Suppose the highest frequency contained in an
analog signal f(t) is fmax = B hertz. Then, if f(t) is sampled periodically
at a rate of
fsample = 1/Ts > 2B (11.10)
samples per second, the signal can be exactly recovered from the sample
values.2
Considering the sampling interval Ts , the sampling theorem states
that the analog signal must be sampled so that more than two samples
per cycle of the highest frequency in the signal are taken. Since a cycle
of the sinusoid at B hertz is 1/B seconds in length, the sampling interval
must be less than 1/(2B) seconds, as indicated in the sampling theorem.
The frequency 2B is called the Nyquist frequency, and the corresponding
sampling rate is the Nyquist rate. Thus, an analog signal must be sampled
at a rate greater than the Nyquist rate if errors due to sampling at too
low a rate are to be avoided.
2
The interpolation function to reconstruct the analog signal is discussed in most textbooks that cover
digital signal processing. Several of these books are listed in the Annotated Bibliography at the end of this
chapter.

524 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT

 For example, if the signal contains frequencies up to B = 1000 hertz,
the Nyquist frequency is 2000 hertz and the sampling theorem requires
sampling at a rate higher than 2000 samples per second. In terms of the
parameters defined earlier for the DFT, the sampling interval correspond-
ing to the Nyquist rate must be
1 1
Ts < = seconds,
2B 2000
or Ts < 0.5 milliseconds, between samples. Another way to view this is
to realize that a 1000-hertz sinusoid repeats every millisecond. Sampling
more than two samples per cycle requires a sampling interval of less than
1/(2 × 103 ) seconds. The length of time the signal is sampled would be
(N − 1)Ts seconds if N samples are taken.
Sampling and the DFT Suppose that an analog signal is sampled at
a rate
1
fsample = > 2B samples/second.
Ts
The highest possible frequency in the DFT spectrum would be Fmax =
1/(2Ts ) hertz. If the signal is bandlimited to B hertz and sampled prop-
erly, the component at the DFT maximum frequency should be zero since
Fmax > B hertz.
Although the sampling theorem gives an elegant solution to the sam-
pling problem, a number of practical considerations intervene to compli-
cate the selection of a sampling rate. First, the reader should review the
examples in this chapter and in Chapter 8 and notice that none of the
ideal example signals have a bandlimited spectrum. In practice, this is
not a problem because any physical signal is in fact bandlimited. The
physical system that created the signal cannot oscillate above some fi-
nite frequency, and the energy content of the signal is negligible beyond
a certain frequency. However, the highest frequency may not be known,
so an anti-aliasing filter can be used in a data acquisition system to limit
the maximum frequency to a known value, say, B hertz. An ideal filter
would leave frequency components of the signal below B hertz unaltered
and eliminate all components above B hertz. Sampling the signal at a
rate greater than 2B samples per second would thus satisfy the sampling
theorem.
Because the physical filter in a data acquisition system is not ideal,
frequencies above the desired bandlimited frequency B can be present in
the signal even after filtering. To compensate for the nonideal character-
istics of the anti-aliasing filter, the filtered signal is often sampled at a
much higher sampling rate than that required by the theorem. In prac-
tice, a signal may be sampled at 3 to 10 or more times the rate dictated
by the minimum ideal rate of 2B samples per second. This oversampling
reduces sampling errors but increases the number of samples that must
be stored and processed.

11.4 Practical Signal Analysis 525

 Another problem cannot be completely overcome. For the sampling
theorem to be correct, the signal must theoretically be sampled in time
over the interval (−∞, ∞). This is often stated as “a bandlimited signal
cannot be limited in time.” The ideal pulses in previous examples are
time limited and thus cannot have bandlimited spectra. However, the
desired frequency resolution in the spectrum of a signal can be used to
define the length of time necessary to sample the signal. Since the DFT
resolution is
∆f = fs = 1/(N Ts ) hertz,
where Ts is the sampling interval for the analog signal, if N Ts = 2 seconds,
it is possible to resolve frequencies as low as f = 1/2 = 0.5 hertz.
If the signal is not sampled at a high-enough rate to satisfy the sam-
pling theorem, the errors cannot be undone once the sampling occurs.
Also, if the signal is not sampled long enough, the resolution in frequency
cannot be improved.

EXAMPLE 11.4 Sampling Example

This example defines the relationship between sampling interval, frequency
resolution, and number of samples for the DFT. In terms of previous notation,
Ts is the sampling interval in seconds, fs is the frequency resolution, and N is
the number of sample points in time and in frequency.
Consider an analog signal with frequencies of interest up to 1200 hertz. The
desired frequency resolution is 0.5 hertz. Thus, the signal should be filtered so
that B = 1200 hertz. This filtering removes frequencies in the signal above 1200
hertz and noise above B hertz. The noise consists of unwanted signals added
to the desired signal that are the result of environmental effects as the signal is
transmitted to the data acquisition system.
By the sampling theorem, the sampling interval in time must be
1 1
Ts < = seconds,
2B 2400
so that at least 2400 samples per second are needed. For a resolution of 0.5
hertz, T = 1/0.5 = 2 seconds. The total number of points required is thus
T 2
N= = = 4800.
Ts (2400)−1

If N is to be a power of 2 for the FFT algorithm, 213 = 8192 samples would be

taken. The sampling rate could be increased to 4096 samples per second, which
is sampling at a rate corresponding to about 3.4 times the highest frequency of
interest.

Spectral Analysis The measurement and study of the frequency con-

tent of a signal is called spectral analysis. Many systems designed to mea-
sure the Fourier spectrum of an arbitrary signal f(t) sample the signal
and perform Fourier analysis using the DFT. If the sampled data points
are stored in computer memory, the discrete signal can be analyzed using
the DFT.

526 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT

 EXAMPLE 11.5 MATLAB Spectral Analysis
The time signal shown in Figure 11.8 was sampled with the parameters

N = 128 and Ts = 1/128.

The DFT analysis yields the spectrum of Figure 11.8, with a maximum fre-
quency of 64 hertz and a resolution of 2 hertz. From the DFT analysis, the
spectrum appears to have a strong component at 20 hertz. In fact, the MAT-
LAB function that generated the signal was
>> N=128;
>> Ts=1/128;
>> t=0:Ts:Ts*(N-1);
>> ft=sin(2*pi*20*t);
>> ft=ft+randn(size(t));
>> save clex115.mat

where the MATLAB function randn generates random values (noise) superim-
posed on the sine wave. The MATLAB script analyzes the signal and plots the
result.
Signal
4

2
f(t)

−2

−4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time in seconds

0.8

0.6
Spectrum

0.4

0.2

0
−80 −60 −40 −20 0 20 40 60 80
Frequency in hertz

FIGURE 11.8 Noisy signal and spectrum

MATLAB Script
Example 11.5
% EX11_5.M Compute the spectrum of a signal saved
% in file CLEX95.MAT. The data are:
% N samples of the |FFT| are plotted
% Ts sampling interval in seconds
% t time points

11.4 Practical Signal Analysis 527

 % ft function f(t)
% Calls clfftf to compute DFT
load clex95.mat % Load N,Ts,t,ft
fs=1/(N*Ts); % Frequency spacing
fhertz=fs*linspace(-N/2,N/2-1,N); % Create frequency axis
[FT,FTmag,FTarg]=clfftf(ft,N,Ts); % Compute DFT
% Plot f(t) and DFT
subplot(2,1,1), plot(t,ft) % Time function
title(’Signal’)
xlabel(’Time in seconds’), ylabel(’f(t)’)
subplot(2,1,2),plot(fhertz, FTmag) % Spectrum
xlabel(’Frequency in hertz’)
ylabel(’Spectrum’)

11.5 PRACTICAL SIGNAL SAMPLING AND DFT ERRORS

One way to approach an error analysis of the DFT is to compare the

transform to the exponential (complex) Fourier series studied in Chap-
ter 8. The Fourier series represents a periodic analog signal that exists
for −∞ < t < ∞. The resulting line spectrum can cover the range
−∞ < f < ∞ with a resolution of 1/T hertz, where T is the period of
the analog signal in seconds. In contrast, the DFT analyzes the analog
signal using a finite number of N samples in the range 0 ≤ nT /N < T .
The useful frequency range is limited to 0 ≤ k/T < N/2T and is periodic
with a period of N frequency points. The limits imposed by the finite
nature of the DFT lead to several possible errors that must be considered
when the DFT is used to determine the spectrum of an analog signal.

ALIASING If a signal is sampled at a rate that is equal to or less than the Nyquist
ERROR rate defined by the sampling theorem, aliasing errors can occur. The
name implies that one signal can be “impersonated” by another signal.
The impersonation caused by too-coarse sampling (undersampling) de-
scribes high-frequency components of the true spectrum appearing as
low-frequency components in the DFT spectrum.
The effect can be seen in several practical situations. Spoked wagon
wheels on a stagecoach going forward occasionally appear to rotate back-
ward when western films are shown on TV because the image of the
rotating wheel is being sampled. The continuous motion is represented
by a series of pictures (samples) taken at a time interval that is too slow
to capture the true motion of the wheel. The effect can also be seen when
a stroboscope is used to capture the motion of rotating equipment. If the

528 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT

 light from the stroboscope is flashed at times that are slightly less than
the period of rotation, the equipment appears to be rotating a slower rate
than the true rate. If the stroboscope flashes exactly at the period of
rotation, the rotating portion of the machine appears to be stationary.

EXAMPLE 11.6 Aliasing

A dramatic example of aliasing in signal processing is shown by the function

f (t) = cos(2πf t),

which should be sampled at a rate greater than 2f samples per second. Suppose
we sample at a rate corresponding to one-half the Nyquist frequency, or only f
samples per second. The sampling interval is thus Ts = 1/f , and the times of
sampling are
1 2 N −1
0, , , . . . , seconds
f f f
if N points are taken. Notice that f (nTs ) = cos(2πf nTs ) = 1 for every point
0 ≤ n < N . Now consider the constant function f (t) = 1 sampled at the same
rate. Both the cosine and the constant yield the time series

f (nTs ) = [1, 1, 1 . . . , 1],

so we conclude that the cosine sampled at one-half of the Nyquist rate imper-
sonates a constant or dc value. The DFT of the cosine signal would indicate a
spectral value at f = 0 but not at the frequency of the cosine, since the maxi-
mum frequency in the DFT spectrum is f /2. The only solution to eliminate the
aliasing error is to sample at a rate that satisfies the sampling theorem. In prac-
tice, if the high-frequency components of the signal diminish rapidly, the effect
of aliasing may not be significant. The anti-aliasing filter in Figure 11.7 ensures
that the signal will be bandlimited in a practical data acquisition system.

LEAKAGE The IDFT produces an N -sample periodic time-series from the spectrum
ERROR created by the DFT. An error called leakage occurs when the actual signal
being sampled is not perfectly periodic over an N -sample interval. The
term implies that energy “leaks” from one frequency to another in the
DFT spectrum. Hence, the signal appears to have frequency components
that may not be present in the spectrum of the analog signal.
Truncating the time signal at t = (N − 1)Ts has the effect of intro-
ducing a discontinuity in the signal unless it is zero at the end point or
the signal is periodic with period T = N Ts . Increasing the DFT period
of the signal by increasing N tends to reduce the leakage effect unless the
analog signal is actually periodic. Problem 11.9 allows you to investigate
the effect.

11.5 Practical Signal Sampling and DFT Errors 529

PICKET FENCE The fact that only discrete frequencies k/T appear in the DFT spectrum
EFFECT leads to the conclusion that actual signal components at other frequencies
will not be present in the spectrum. This is sometimes colorfully referred
to as the picket fence effect since the scene behind this fence can be
viewed only between the pickets. The pickets in this case correspond to
the frequencies that cannot be resolved in the spectrum. Increasing the
overall sampling time T while holding the sampling interval constant leads
to greater resolution in the spectrum. This is equivalent to increasing the
number of samples N .

SUMMARY OF Table 11.6 summarizes the major DFT errors that may arise from sam-
ERRORS pling at too low a rate or not sampling long enough. For the parameters
listed in the table, T = N Ts , so there are really only two parameters
to manipulate. Several problems at the end of this chapter are intended
to emphasize the effects on the DFT spectrum of changing the sampling
interval or the number of samples.

TABLE 11.6 DFT errors

Condition Cure
Aliasing Increase maximum frequency
(decrease Ts )
Leakage Increase frequency resolution
(increase N )
Picket fence Increase frequency resolution
(increase N )

11.6 ANALYSIS OF DFT FOR COMPUTATION (OPTIONAL)

Using and interpreting the FFT results previously presented in this chap-
ter does not require a complete understanding of the algorithm itself since
the FFT is only one particular method of computing the discrete Fourier
transform (DFT). In this section, we only describe particular cases of the
FFT algorithm for simplicity. For example, we restrict the value of N
to an even number, and furthermore the discussion often assumes that
N = 2m for convenience in understanding and computation. References
in the Annotated Bibliography for this chapter treat other cases with a
more general FFT algorithm.

530 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT

 Consider the discrete Fourier transform of Equation 11.5 in the form
N−1
X
Fk = fn e−i2πnk/N k = 0, 1, . . ., N − 1, (11.11)
n=0

where nTs is replaced with n and k/N Ts is replaced by k. This describes

the components in terms of their index values. After the DFT is cal-
culated, the values can be scaled to reflect the frequency resolution and
range.
To explore the characteristics of the exponential factor in Equation 11.11,
we first define
WN = e−i2π/N . (11.12)
This is the complex number
2π 2π
WN = e−i2π/N = cos − i sin
N N
with the property that |WN | = 1 so that every one of these complex
numbers lies on the unit circle in the complex plane. In Equation 11.11,
the exponential factors have the form WNnk with the values

WNnk = (e−i2π/N )nk = cos 2πnk/N − i sin 2πnk/N. (11.13)

The term WN0 = 1. In cases where the argument is a multiple of π, the

value is ±1. If the argument is a multiple of π/2, the value is ±i.
Since WN = z is a complex number, we can write the terms as the
solutions of an equation of the form

z N = a,

which is the formula for the N th root of a. The solution uses the theorem
of De Moivre
(cos θ + i sin θ)N = cos N θ + i sin N θ, (11.14)
which follows from the formula for multiplication of complex numbers
written in polar form as presented in Chapter 2. Writing a in polar
coordinates as a = r(cos θ + i sin θ), the roots of a are

√
    
1/N N
θ 2π θ 2π
a = r cos +p + i sin +p , (11.15)
N N N N
√
where p takes the values 0, 1, . . ., n − 1. The term N r is the positive real
N th root of r, and for the specific case |a| = 1, this value is 1.
Consider the values p/N for integers p = 0, 1, 2, 3 with N = 4. Writing
W4p = e−θ , where θ = 2πp/4, these values correspond to the angles

θ = 0 (2π), θ = π/2, θ = π and θ = 3π/2

11.6 Analysis of DFT for Computation (Optional) 531

 in the complex plane, respectively. The exponential W4p then takes the
values
W40 = e0 = 1
W41 = e−iπ/2 = −i
(11.16)
W42 = e−iπ = −1
W43 = e−i3π/2 = i.
Notice that when N = 4, these results correspond to the four roots of 1
in the equation z 4 = 1, where z is a complex number. The square roots
of unity are ±1, and the square roots of these roots are 1, −1, i, −i as is
easily verified. The sum of the roots (1 − i − 1 + i) is zero.
The repeated values in the relationship of Equation 11.13 can be
written
WNnk = WNnk mod N , (11.17)
where nk mod N is the remainder upon division of nk by N . Thus, the
term W44 , if formed, would be the same as W40 . Similarly, W45 = W41 . The
results thus far lead to observations about the periodicity and symmetry
of the terms in the DFT.
Periodicity and Symmetry in the DFT Considering the exponen-
tial factor in the DFT of Equation 11.11, the periodic extension of the
sequence of frequency components F (k) is periodic with period N . This
is shown by writing
N−1
X
Fk+N = fn e−i2πn(k+N)/N
n=0
N−1
X
= fn e−i2πnk/N e−i2πn
n=0
N−1
X
= fn e−i2πnk/N = Fk (11.18)
n=0

because n is an integer.
When f(t) is real, we expect that the real part of Fk is even and
the imaginary part is odd, using k = N/2 as the origin. Remember from
the previous discussions that the frequencies for k > N/2 are actually
negative frequencies so that

F̄k = FN−k

when f(t) is real and N is an even number, where F̄k is the complex conju-
gate of the term Fk . We say that the terms in the DFT exhibit conjugate
symmetry about the point N/2 in this case. Thus, computation can be
simplified if the DFT is to be computed directly from the definition.

532 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT

 EXAMPLE 11.7 DFT Example
Consider the sequence f (n) = {1, 1, 0, 0, 0, 0, 0, 0}, for which the DFT re-
duces to the sum
1
X
F (k) = f (n)e−i2πnk/8 = 1 + exp(−iπk/4) k = 0, 1, . . . , 7. (11.19)
n=0

The terms are thus

F0 = 1+1 = 2
F1 = 1 + exp(−iπ/4) = 1.7071 − 0.7071i
F2 = 1 + exp(−iπ/2) = 1−i
F3 = 1 + exp(−iπ3/4) = 0.2929 − 0.7071i
F4 = 1 + exp(−iπ) = 0
F5 = 1 + exp(−iπ5/4) = 0.2929 + 0.7071i
F6 = 1 + exp(−iπ3/2) = 1+i
F7 = 1 + exp(−iπ7/4) = 1.7071 + 0.7071i.

The argument of each exponential term is a multiple of −π/4.

√ The numerical
value can be computed from Equation 11.13, using ±1/ 2 ≈ ±0.7071 as an
approximation.
Notice that there is conjugate symmetry about N/2 = 4; that is, since
FN −k = F̄k ,
F5 = F̄3 , F6 = F̄2 , F7 = F̄1 .

WHAT IF? Compare the result of Example 11.7 with the results of
the MATLAB script

>>ft=[1,1,0,0,0,0,0,0]
>> Fw=fft(ft)

What conclusions do you draw about the MATLAB fft command

with respect to calculating the DFT?

Four-Point DFT It is convenient to write Equation 11.11 for the

Fourier component Fk in matrix form to explore the characteristics of
the DFT. We will consider the expansion of Equation 11.11 with N = 4
to show the pattern. Expanding Equation 11.11 leads to the result

F0 = f0 W40 + f1 W40 + f2 W40 + f3 W40

F1 = f0 W40 + f1 W41 + f2 W42 + f3 W43
F2 = f0 W40 + f1 W42 + f2 W44 + f3 W46
F3 = f0 W40 + f1 W43 + f2 W46 + f3 W49 . (11.20)

11.6 Analysis of DFT for Computation (Optional) 533

 In Equation 11.20, N = 4 = 22 , so the exponential terms can be
written as W4nk = W4nk mod 4 according to Equation 11.17. Thus, we
compute
W40 = W44 = 1
W41 = W49 = −i
(11.21)
W42 = W46 = −1
W43 = i.

using the result of Equation 11.16. In Equation 11.20, the terms W45 =
W41 , W47 = W43 , and W49 = (W43 )3 = W41 .
Rewriting Equation 11.20 in matrix terms, considering the results
defined in Equations 11.21, yields
    
F0 1 1 1 1 f0
 F1   1 −i −1 i 
  f1  .
 
 F2  =  1 −1 (11.22)
  
1 −1   f2 
F3 1 i −1 −i f3

N -Point DFT Using the approach used for the four-point problem,
the DFT for N points can be written
   1 1 1 ··· 1
 
F0 f0
1 2 N−1
 F1   1 WN WN · · · WN 
f1 
   2 4

2(N−1)   
 F2  
= 1 WN WN · · · WN   f2 

.
.. .. .. .. .. .. ..

    
.   . . . . . .
  
 
FN−1 N−1 2(N−1) (N−1)2 f
1 WN WN · · · WN N−1
(11.23)
The Fourier equation (11.23), which can be written as

F = WNnk f ,
 
(11.24)

performs the analysis of the signal represented by f . The inverse relation-

ship represents the synthesis of a signal from its Fourier components. As
you are asked to show in Problem 11.10, the N × N Fourier matrix in
Equation 11.23 has orthogonal columns. The inverse matrix is formed by
taking the conjugate transpose and dividing by N .
Notice that N 2 multiplications and N (N −1) additions are performed
in the direct matrix computation of Equation 11.23, and many of these
involve complex values. One complex multiplication requires four real
multiplications and three real additions. The number of direct calcula-
tions can be reduced if we let N = 2m , where m is an integer, since
there are many repeated values among the real and imaginary parts in
Equation 11.13.

534 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT

 EXAMPLE 11.8 Four-point DFT
Let f (n) = [1, 2, 3, 4] so that the four-point DFT from Equation 11.22 is

F0 1 1 1 1 1 10
      
 F1   1 −i −1 i   2   −2 + 2i 
 F = = . (11.25)
2 1 −1 1 −1   3   −2
F3 1 i −1 −i 4 −2 − 2i

Formulation of the DFT matrix from direct application of the DFT defi-
nition has a number of computational drawbacks. Obviously, it would not be
very efficient to store the N × N matrix if N is a large number. In any case, the
direct computation method is not considered computationally efficient, and it
is desirable to find methods of speeding up the computer calculation by reduc-
ing the number of operations. When comparing computational methods, only
the multiplications are usually counted since they are by far the most time-
consuming compared to the time for addition, subtraction, and retrieval and
storage of data values.

FFT As we have seen, the direct calculation of the N -point DFT requires
ALGORITHMS N 2 complex multiplications. A number of researchers have developed
algorithms called fast Fourier transform algorithms that more efficiently
compute the DFT. These algorithms essentially reduce the problem of
calculating an N -point DFT to that of calculating many DFTs for smaller
values of N . Although there are numerous algorithms, the algorithms
generally perform an efficient DFT by taking advantage of the following:

1. The symmetry and periodicity of WNnk = exp(−i2πnk/N ).

2. Careful choice of N or the factors of N .

3. Separate calculation on samples of even index and odd index.

4. Performing calculations in place.

Radix-2 FFT The choice N = 2m for the number of points, where m

is an integer, leads to the radix -2 FFT. We will present an intuitive de-
velopment (without proof) of the fast Fourier transform (FFT) algorithm
for the particular case that the number of points is a power of two.3 The
key to the efficiency of the FFT algorithm we discuss is the factorization
of an N -point transform into two N/2-point transforms. Even with the
computation necessary to assemble the two transforms into the N -point
transform, considerable savings in computation time results, particularly
when N is a large number. In fact, N 2 /2 multiplications are required,
which is half the number needed in the original problem. The reduction
3
For a more rigorous approach and for a description of other algorithms, the textbook by Brigham listed
in the Annotated Bibliography for this chapter is highly recommended.

11.6 Analysis of DFT for Computation (Optional) 535

 procedure can be repeated until two-point transforms are computed. In
Example 11.9, we will show a particular case of the FFT known as a
radix-2 transform for N = 4 points.4

EXAMPLE 11.9 Two-and Four-Point FFT

Consider the case of the two-point DFT
1
X 1
X
Fk = fn e−i2πnk/2 = fn W2nk k = 0, 1.
n=0 n=0

Because W20k = 1 and W21k = e−iπk = (−1)k , the two-point result is

F0 = f0 + f1
F1 = f0 − f1 .

In matrix form, the two-point transform is

    
F0 1 1 f0
= . (11.26)
F1 1 −1 f1

To factor the DFT for the FFT algorithm, the DFT is written in terms
of the even and odd indices of f (n). We assume that N = 2m and show the
case for N = 4 written in terms of two, two-point transforms. This algorithm
is described as a decomposition-in-time or decimation-in-time, radix-2 FFT.
The kth term in the four-point transform is
3
X
Fk = f (n)e−i2πnk/4
n=0

= f0 W40k + f1 W41k + f2 W42k + f3 W43k , (11.27)

where k = 0, 1, 2, 3. Using the results in Equation 11.21 to simplify W4nk leads
to the expression

Fk = [f0 + f2 (−1)k ] + [f1 + f3 (−1)k ]W41k (11.28)

for k = 0, 1, 2, 3. The idea is to write this expression in terms of two-point
transforms. Define the even and odd points as

fen = f2n , n = 0, 1
fon = f2n+1 , n = 0, 1

so that
Fk = [fe0 + fe1 (−1)k ] + [fo0 + fo1 (−1)k ]W41k , (11.29)
where the factors in brackets are two-point DFTs. If we define

Feq = fe0 + fe1(−1)q q = 0, 1

q
Foq = fo0 + fo1 (−1) q = 0, 1, (11.30)
4
An illuminating article that covers the history and derivation of the FFT is “A Guided Tour of the Fast
Fourier Transform,” by G.D. Bergland in the IEEE Spectrum, Vol. 6, July 1969, pp41–52.

536 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT

 you are asked to show in Problem 11.13 that the four-point FFT can be written
as

F0 = Fe0 + Fo0
F1 = Fe1 − Fo1 (i)
F2 = Fe0 − Fo0
F3 = Fe1 + Fo1 (i). (11.31)

In the final expression, the values W41k = (−i)k for each k were used.
The conclusion is that the four-point DFT can be written as follows

[Four-point DFT of f (n)] = [Two-point DFT of fe (n)]

+ W41k [Two-point DFT of fo(n)].

The result in Example 11.9 can be generalized for radix-2 FFT al-
gorithms to show that an N -sample DFT can be written as the sum of
two N/2-sample DFTs formed from the even- and odd-indexed samples
of the original sequence. When N = 2m , this process of decomposition
or decimation can be continued until the N -point FFT can be computed
by properly combining N/2 two-point transforms. The reader should see
the references in the Annotated Bibliography for this chapter for more
information about this important algorithm.

11.7 REINFORCEMENT EXERCISES AND EXPLORATION

PROBLEMS

In these problems, do the computations by hand unless otherwise indicated and then check those
that yield numerical or symbolic results with MATLAB.

REINFORCEMENT EXERCISES
P11.1. Periodicity of the DFT Show that the DFT defined in Equation 11.5 is periodic.
Thus, for an arbitrary integer r,
r+N r
   
F =F .
N Ts N Ts
P11.2. Approximation to Fourier transform Consider the Fourier transform of a function
Z ∞
F (ω) = f (t)e−iωt dt
−∞

applied to a function sampled at intervals t = nTs . Assume that the function is defined only for
t > 0 and further that f (nTs ) is almost zero for n ≥ N when N is a sufficiently large number. By

11.7 Reinforcement Exercises and Exploration Problems 537

integrating over one sampling interval and summing the results, show that the discrete
approximation to the Fourier transform is
N −1
X 2π
F (kΩ) ≈ Ts f (nTs )e−inkΩTs Ω= ,
N Ts
n=0

which agrees with the definition of the DFT in this chapter.

Hint: Assume that the sampling interval is sufficiently small that exponentials not involving k
or n can be expanded in a Taylor series with only two terms.

P11.3. DFT conditions Determine the functions f (t) that have the properties that meet the
criteria for exact DFT representation:

a. f (t) is periodic and the spectrum is bandlimited.

b. f (t) can be sampled at a rate that yields more than two samples per cycle of the
highest frequency in F (f ).
c. f (t) can be sampled exactly over one period or an integral number of periods.

Hint: Consider the answer in terms of the Fourier series expansion of the function.

P11.4. Direct calculation of DFT A good way to understand the DFT is to calculate the
DFT for a simple function by hand. Determine the DFT for the discrete function

1, 0 ≤ n ≤ 4,
fn =
0, 5 ≤ n ≤ 9.

Plot the amplitude and phase of the transform and describe the symmetries in the transform.
P11.5. Aliasing Given that the sampling rate is 10 samples per second for the following signals,
are they aliased?

a. f (t) = cos(2πt)
b. f (t) = cos(4πt)
c. f (t) = cos(10πt)
d. f (t) = cos(12πt)
e. f (t) = cos(14πt)

P11.6. MATLAB discrete signal Create a 256-point function with a 25% duty cycle of ones
(64 points) and compute the Fourier spectrum and plot it.

a. Try using the MATLAB command kron to create the discrete pulse.
b. Compute the maximum error when compared with the Fourier transform of a pulse.

538 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT

P11.7. MATLAB FFT Compare the Fourier spectrum for the signal
πt
(  
sin , for 0 ≤ t ≤ 4,
f (t) = 4
0, for t < 0 or t > 4,
sampled every Ts seconds at N points in the following cases:
a. N = 16 and Ts = 1 second;
b. N = 64 and Ts = 1 second;
c. N = 128 and Ts = 0.5 second.
P11.8. MATLAB FFT resolution Numerical techniques usually require some trial and error.
For example, the proper number of sample points N and the sampling period of the FFT are not
generally known for an arbitrary signal. Assume that a signal is given analytically and various
values of sampling times and sampling period are to be tried.
One method is to choose a reasonable value for the number of samples N and compute the
spectrum of the signal. Then, let N 1 = 2 ∗ N and recompute. This is repeated until the spectrum
of two subsequent calculations are very close. Start with 64 samples.
Assume the function
sin(0.6πt) + 0.5 sin(0.64πt),
sampled every second for 512 points, is to be analyzed.
a. Compute the magnitude of the FFT for 64, 128, 256, and 512 points and plot the
results.
b. Analyze each plot and compare the frequency resolution for each.
c. For the final plot, if there appear to be several distinct frequencies, pick off the peaks
and print the frequencies that correspond.

Hint: Use MATLAB command ginput if you have a mouse.

P11.9. MATLAB FFT leakage Compare the FFT results for the functions:
f (t) = sin(2π20t) and g(t) = sin(2π19t)
by sampling 64 points with a time resolution of 1/128 second.
a. Plot the magnitude of each FFT on a subplot to compare the results.
b. Explain the extra frequency components in each result.
P11.10. DFT matrix Show that the inverse of the N × N Fourier matrix of Equation 11.23 is
the conjugate transpose divided by N . Therefore, the algorithm to compute the DFT can be used,
with minor modifications, to compute the inverse DFT.
P11.11. DFT 8-point matrix Work out the 8-point DFT matrix from Equation 11.23.
P11.12. Four-point FFT matrix Show that the four-point FFT as in Example 11.9 can be
written as the product of matrices as
F0 1 0 1 0 1 1 0 0 1 0 0 0 f0
      
 F1   0 1 0 −i   1 −1 0 0  0 0 1 0   f1 
 F = 1 0 −1 0  0 0 1 1  0 1 0 0   f2 
.
2
F3 0 1 0 i 0 0 1 −1 0 0 0 1 f3
Explain the meaning of each matrix.

11.7 Reinforcement Exercises and Exploration Problems 539

P11.13. FFT factorization Write out the terms in Equation 11.29 and show that the
four-point FFT can be written in terms of 2 two-point FFTs.
P11.14. FFT four-point factorization Write out the terms in Equation 11.31 and show that
the result is the same as Equation 11.22.
P11.15. Four point FFT Using the decomposition described in Example 11.9, compute the
DFT of the function
fn = [1, 2, 3, 4].

EXPLORATION PROBLEMS
P11.16. MATLAB FFT phase Given the function

e−t , t ≥ 0,
f (t) =
0, t < 0,

as treated in Example 11.3, compute the phase of the function and compare the results with the
actual value. Use the MATLAB command fftshift to plot the two-sided spectrum from −Fmax to
Fmax, where the frequency range will be determined by the sampling interval of f (t) and the
number of points taken.

Hint: Try to approach the Fourier transform result (at least near the origin) by adding zeros to
increase the period of the time function as much as possible. This method is called zero padding to
lengthen the sequence of sampled points. However, zero padding does not add any information
about the signal being analyzed.

P11.17. MATLAB discrete spectrum plot Write an M-function to plot a discrete spectrum
versus frequency in hertz or radians per second. The function call should be:

function plotdscf(f,F,xunit)
% CALL: plotdscf(f,F,xunit); plot a discrete spectrum [f F]
% Inputs to function are
% f - frequencies
% F - spectral values
% xunit - units of frequency (Hz or rad/s)
P11.18. FFT Investigation (Optional) Consider the unit step function, U (t) = 1 for f > 0,
defined in Chapter 6. Its ordinary Fourier transform does not exist (because the signal would have
infinite energy) but a generalized transform can be defined. The transform is written
1
F [U (t)] = πδ(ω) + ,
iω
where δ(ω) is the unit impulse function. Attempt to compute the DFT of the unit step function
and explain the result. Use the inverse FFT (ifft) on the result and reconstruct the time function.
Compare the results with the Laplace transform of the unit step function.
P11.19. MATLAB DFT  nk matrix Write a MATLAB script with N as input to compute the
N -point DFT matrix WN .
P11.20. MATLAB Script Write a MATLAB script that computes the DFT directly from the
definition with the inputs

540 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT

 a. N , the number of sample points;
b. f , the vector of sample points.

Test the routine on the sequence f = {1 1 0 0 0 0 0 0}.

P11.21. MATLAB FFT Consider the Fourier transform of the function

1, 0 ≤ t ≤ 2s,
P (t) = (11.32)
0, otherwise.

Compare the FFT result with |F (ω)| and find the maximum error for various values of N , the
number of samples; N = 16, 128, 1024. For each value of N , plot |F (ω)| and the FFT result on the
same graph.
Remember you are comparing the Fourier transform of a time signal with the FFT results.
P11.22. MATLAB Script comparing DFT and FFT Write a MATLAB script to compare
the time of the DFT routine from the previous problem and the MATLAB FFT routine. The
result should be a table of the form
Comparison of DFT and FFT times
N DFT FFT FFT/DFT log2 N/N
8 Time1 etc.
64
128
1024
In the table, N is the number of sample points and the times are in seconds (or fractions of a
second). Test the functions:

a. f = {1 1 0 0 0 0 0 0 . . .}.
b. A random sequence of N numbers.

P11.23. MATLAB Script comparing Fourier transform and FFT Compute the Fourier
transform of the function
f (t) = e−t cos 100t.
Then, plot the following on the same figure:

a. Plot the function from 0 ≤ t ≤ 10.

b. Plot the Fourier transform amplitude and phase versus radians/second over the range
[0, 314];
c. Plot the amplitude and phase of the FFT result.
d. Comment on the error in the amplitude and phase as determined by the FFT.

11.7 Reinforcement Exercises and Exploration Problems 541

11.8 ANNOTATED BIBLIOGRAPHY

1. Abramowitz, Milton, and Irene A. Stegun, Handbook of Mathematical

Functions, Dover Publications, Inc., New York, 1972. This handbook contains
many formulas, graphs and tables of mathematical functions. In particular, the
solutions to the equation x = tan x are tabulated.
2. Bracewell, Ronald N., The Fourier Transform and Its Applications, 2nd
edition, revised, McGraw-Hill Book Company, New York, 1986. An excellent
book that covers many applications of the Fourier transform as well as a number
of other transforms.
3. Brigham, E. Oran, The Fast Fourier Transform and its Applications,
Prentice Hall, Englewood Cliffs, NJ, 1988. A very readable treatment of Fourier
techniques with emphasis on the FFT. There is also an extensive bibliography.
4. Burrus, C. Sidney, J. H. McClellan, A. V. Oppenheim, T. W. Parks,
R. W. Schafer, and H. W. Schuessler, Computer-Based Exercises for Signal
Processing Using MATLAB, Prentice Hall, Englewood Cliffs, NJ, 1994. The
book contains a collection of computer exercises about signal processing using
MATLAB.
5. Cooley, James W., P. Lewis, and P. Welch, “Application of the Fast
Fourier Transform to Computation on Fourier Integrals, Fourier Series, and
Convolution Integrals,” AU-15, No. 2, June 1967. This paper discusses the use
of the FFT and the errors involved in various approximations.
6. Hamming, Richard W., Numerical Methods for Scientists and Engineers,
McGraw-Hall Book Company, New York, 1973. An interesting treatment of
numerical techniques with a number of personal insights from the author. The
book is particularly notable for dividing numerical analysis into classical analysis
(polynomials) and modern analysis (Fourier methods).

11.9 ANSWERS

11.1 Periodicity of the DFT For an arbitrary integer r, the DFT component at index
k = r + N is
N −1
r+N
  X
F = f (nTs )e−i2πn(r+N )/N .
N Ts
n=0

Considering the exponential factor, we find

e−i2πn(r+N )/N = e−i2πnr/N e−i2πn = e−i2πnr/N

since
e−i2πn = cos 2πn − i sin 2πn = 1

542 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT

for n an integer. Thus,
r+N r
   
F =F ,
N Ts N Ts
which shows that the DFT defined in Equation 11.5 is periodic with a period N .
11.3 DFT conditions The only functions f (t) for which all the ideal DFT conditions hold
are those functions with finite Fourier series.
11.5 Aliasing Given that the sampling rate is 10 samples per second, the sampling theorem
requires the maximum frequency in the signal to be less than fmax = 10/2 = 5 hertz to avoid
aliasing.
a. cos(2πt), f = 1 Hz, not aliased;
b. cos(4πt), f = 2 Hz, not aliased;
c. cos(10πt), f = 5 Hz, aliased;
d. cos(12πt), f = 6 Hz, aliased;
e. cos(14πt), f = 7 Hz, aliased.
11.9 FFT leakage The 20-hertz sine wave is sampled over a number of periods with a DFT
spectrum resolution of 2 hertz. If a sufficient number of sampling points are taken, the spectrum
appears as a “spike” at 20 hertz (with a component at negative frequencies). The 19-hertz signal is
truncated by sampling but not over an integral number of periods. The energy at 19 hertz “leaks”
into adjacent frequencies, and the spectrum near 19 hertz appears broadened. Another way to view
this is to recognize that there is no component of this DFT spectrum at 19 hertz. Energy at 19
hertz will appear at all the frequencies in the spectrum, with the largest values at 18 and 20 hertz.
11.11 DFT 8-point matrix For N = 8, the terms of exp(i2πm/N ), m = 0, . . . 7 are as follows:

W80 = 1
√
W81 = (1 − i)/ 2
W82 = −i
√
W83 = −(1 + i)/ 2
W84 = −1
√
W85 = −(1 − i)/ 2
W86 = i
√
W87 = (1 + i)/ 2.

Since W8nk = W8nk mod 8 , relationships such as W80 = W88 = W816 = W824 , etc., reduce the matrix to
the form  
1 1 1 1 1 1 1 1
 1 W81 W82 W83 W84 W85 W86 W87 
 1 W82 W84 W86 W80 W82 W84 W86
 

 nk    1 W83 W86 W81 W84 W87 W82 W85

W8 = .

4 0 4 0 4
 1 W8 W8 W8 W8 W8 W80 W84 
 1 W5 W2 W7 W4 W1 W86 W83 
 8 8 8 8 8 
 1 W6 W4 W2 W0 W6 W84 W82 
6 8 8 8 8
1 W87 W86 W85 W84 W83 W82 W81

11.9 Answers 543

This can be further simplified by using the relationships

W84 = −W80 W85 = −W81 W86 = −W82 W87 = −W83 .

11.13 FFT factorization By direct expansion of Equation11.29, the four-point transform is

F0 = fe0 + fe1 + [fo0 + fo1 ] = Fe0 + Fo0

F1 = fe0 − fe1 + [fo0 − fo1 ](−i) = Fe1 − Fo0 (i)
F2 = fe0 + fe1 + [fo0 + fo1 (−1)] = Fe0 − Fo0
F3 = fe0 − fe1 + [fo0 + fo1 ](i) = Fe1 + Fo1 (i).

11.15 Four point FFT Verify your answer with the result of Example 11.8.

544 Chapter 11 THE DISCRETE FOURIER TRANSFORM AND THE FFT