FOURIER TRANSFORMS

An Introduction for Engineers

 FOURIER TRANSFORMS
An Introduction for Engineers

by

Robert M. Gray
Joseph W. Goodman

Information Systems Laboratory

Department of Electrical Engineering
Stanford University

....
"
SPRINGER SCIENCE+BUSINESS MEDIA, LLC
ISBN 978-1-4613-6001-8 ISBN 978-1-4615-2359-8 (eBook)
DOI 10.1007/978-1-4615-2359-8

Library of Congress Cataloging-in-Publication Data

A C.I.P. Catalogue record for this book is available

from the Library of Congress.

Copyright ~ 1995 by Springer Science+Business Media New York

Second Printing 2001.
Originally published by Kluwer Academic Publishers in 1995
Softcover reprint ofthe hardcover Ist edition 1995
Ali rights reserved. No part of this publication may be reproduced, stored in
a retrieval system or transmitted in any form or by any means, mechanical,
photo-copying, recording, or otherwise, without the prior written permission of
the publisher, Springer Science+Business Media, LLC.

Printed on acid-free paper.

This printing is a digital duplication of the original editioll.

 to Ron Bracewell
whose teaching and research on Fourier transforms
and their applications have been an inspiration
to us and to generations of students
Contents

Preface xi

1 Signals and Systems 1

1.1 Waveforms and Sequences 1
1.2 Basic Signal Examples 5
1.3 Random Signals. . . 18
1.4 Systems . . . . . . . 19
1.5 Linear Combinations 20
1.6 Shifts . . . . . . . . 24
1.7 Two-Dimensional Signals 28
1.8 Sampling, Windowing, and Extending 35
1.9 Probability Functions 47
1.10 Problems . . . . . . . 48

2 The Fourier Transform 53

2.1 Basic Definitions . . . . . . . . . 53
2.2 Simple Examples . . . . . . . . . 59
2.3 Cousins of the Fourier Transform 70
2.4 Multidimensional Transforms .. 74
2.5 * The DFT Approximation to the CTFT 79
2.6 The Fast Fourier Transform 81
2.7 * Existence Conditions . 92
2.8 Problems . . . . . . . . . . 107

3 Fourier Inversion 115

3.1 Inverting the DFT . . . . . . . . . . . 115
3.2 Discrete Time Fourier Series . . . . . . 121
3.3 Inverting the Infinite Duration DTFT 122
3.4 Inverting the CTFT . . . . . . 126
3.5 Continuous Time Fourier Series . . . . 137
viii CONTENTS

3.6 Duality . . . . . . . . 140

3.7 Summary . . . . . . . 144
3.8 * Orthonormal Bases . 145
3.9 * Discrete Time Wavelet Transforms 147
3.10 * Two-Dimensional Inversion 152
3.11 Problems . . . . . . . . . . . . . . . 153

4 Basic Properties 161

4.1 Linearity.. 161
4.2 Shifts . . . . 162
4.3 Modulation 164
4.4 Parseval's Theorem . 166
4.5 The Sampling Theorem 170
4.6 The DTFT of a Sampled Signal . 173
4.7 * Pulse Amplitude Modulation (PAM) 180
4.8 The Stretch Theorem 181
4.9 * Downsampling . . . . . . . . . . . . 182
4.10 * Upsampling . . . . . . . . . . . . . . 186
4.11 The Derivative and Difference Theorems. 187
4.12 Moment Generating . . . . . 190
4.13 Bandwidth and Pulse Width 197
4.14 Symmetry Properties. 199
4.15 Problems . . . . . . . . . . . 203

5 Generalized Transforms and Functions 217

5.1 Limiting Transforms . . . . . . . . 217
5.2 Periodic Signals and Fourier Series .. . 219
5.3 Generalized Functions . . . . . . . . . . 227
5.4 Fourier Transforms of Generalized Functions 233
5.5 * Derivatives of Delta Functions 235
5.6 * The Generalized Function 6(g(t» 237
5.7 Impulse Trains 238
5.8 Problems . . . . . . . . . . 245

6 Convolution and Correlation 251

6.1 Linear Systems and Convolution 251
6.2 Convolution......... 257
6.3 Examples of Convolution . . . . 261
6.4 The Convolution Theorem . . . . 267
6.5 Fourier Analysis of Linear Systems 271
6.6 The Integral Theorem 274
6.7 Sampling Revisited . . . . . . . . . 275
CONTENTS ix

6.8 Correlation . . . . . . . . . . . . . . . . 279

6.9 Parseval's Theorem Revisited . . . . . . 285
6.10 * Bandwidth and Pulsewidth Revisited. 285
6.11 * The Central Limit Theorem. 289
6.12 Problems . . . . . . . . . . . . . . 293

7 Two Dimensional Fourier Analysis 309

7.1 Properties of 2-D Fourier 'Transforms. 310
7.2 Two Dimensional Linear Systems 312
7.3 Reconstruction from Projections .. . 317
7.4 The Inversion Problem . . . . . . . . . 320
7.5 Examples of the Projection-Slice Theorem 321
7.6 Reconstruction . . . . . . . . . . . . . 324
7.7 *
Two-Dimensional Sampling Theory. 327
7.8 Problems . . . . . . . . . . . . . . . . 332

8 Memoryless Nonlinearities 333

8.1 Memoryless Nonlinearities 334
8.2 Sinusoidal Inputs . . . 335
8.3 Phase Modulation .. 337
8.4 Uniform Quantization 338
8.5 Problems . . . . . . . 344

A Fourier Transform Tables 347

Bibliography 353

Index 356
Preface

The Fourier transform is one of the most important mathematical tools

in a wide variety of fields in science and engineering. In the abstract it
can be viewed as the transformation of a signal in one domain (typically
time or space) into another domain, the frequency domain. Applications
of Fourier transforms, often called Fourier analysis or harmonic analysis,
provide useful decompositions of signals into fundamental or "primitive"
components, provide shortcuts to the computation of complicated sums
and integrals, and often reveal hidden structure in data. Fourier analysis
lies at the base of many theories of science and plays a fundamental role in
practical engineering design.
The origins of Fourier analysis in science can be found in Ptolemy's
decomposing celestial orbits into cycles and epicycles and Pythagorus' de-
composing music into consonances. Its modern history began with the
eighteenth century work of Bernoulli, Euler, and Gauss on what later came
to be known as Fourier series. J. Fourier in his 1822 Theorie analytique de la
Chaleur [16] (still available as a Dover reprint) was the first to claim that
arbitrary periodic functions could be expanded in a trigonometric (later
called a Fourier) series, a claim that was eventually shown to be incorrect,
although not too far from the truth. It is an amusing historical sidelight
that this work won a prize from the French Academy, in spite of serious
concerns expressed by the judges (Laplace, Lagrange, and Legendre) re-
garding Fourier's lack of rigor. Fourier was apparently a better engineer
than mathematician. (Unhappily for France, he subsequently proved to
be an even worse politician than mathematician.) Dirichlet later made
rigorous the basic results for Fourier series and gave precise conditions un-
der which they applied. The rigorous theoretical development of general
Fourier transforms did not follow until about one hundred years later with
the development of the Lebesgue integral.
The current extent of the influence of Fourier analysis is indicated by a
partial list of scientists and engineers who use it:
xii PREFACE

• Circuit designers, from audio to microwave, characterize circuits in

terms of their frequency response.
• Systems engineers use Fourier techniques in signal processing and
communications algorithms for applications such as speech and im-
age processing and coding (or compression), and for estimation and
system identification. In addition to its widespread use for the analy-
sis of linear systems, it also plays a fundamental role in the analysis of
nonlinear systems, especially memoryless nonlinearities such as quan-
tizers, hard limiters, and rectifiers.
• Audio engineers use Fourier techniques, partially because the ear
seems to be sensitive to frequency domain behavior.
• Statisticians and probabilists characterize and compute probability
distributions using Fourier transforms (called characteristic functions
or operational transforms). Fourier transforms of covariance func-
tions are used to characterize and estimate the properties of random
processes.
• Error control code designers use Fourier techniques to characterize
cyclic codes for error correction and detection.
• Radio astronomers use the Fourier transform to form images from
interferometric data gathered from antenna arrays.
• Antenna designers evaluate beam patterns of periodic arrays using z-
transforms, a form of Fourier transform, and evaluate beam patterns
for more general arrays using Fourier transforms.
• Spectroscopists use the Fourier transform to obtain high resolution
spectra in the infrared from interferograms (Fourier spectroscopy).
• Crystallographers find crystal structure using Fourier transforms of
X-ray diffraction patterns.
• Lens designers specify camera performance in terms of spatial fre-
quency response.
• Psychologists use the Fourier transform to study perception.
• Biomedical engineers use Fourier transforms for medical imaging, as
with magnetic resonance imaging (MRI) wherein data collected in the
frequency domain is inverse Fourier transformed to obtain images.
• Mathematicians and engineers use Fourier transforms in the solution
of differential, integral, and other equations.
PREFACE xiii

This book is devoted to a development of the basic definitions, proper-

ties, and applications of Fourier analysis. The emphasis is on techniques
important for applications to linear systems, but other applications are oc-
casionally described as well. The book is intended for engineers, especially
for electrical engineers, but it attempts to provide a careful treatment of
the fundamental mathematical ideas wherever possible. The assumed pre-
requisite is familiarity with complex variables and basic calculus, especially
sums and Riemann integration. Some familiarity with linear algebra is also
assumed when vector and matrix ideas are used. Since knowledge of real
analysis and Lebesgue integration is not assumed, many of the mathemat-
ical details are not within the scope of this book. Proofs are provided
in simple cases when they can be accomplished within the assumed back-
ground, but for more general cases we content ourselves with traditional
engineering heuristic arguments. These arguments can always be made rig-
orous, however, and such details can be found in the cited mathematical
texts.
This book is intended to serve both as a reference text and as a teach-
ing text for a one quarter or one semester course on the fundamentals of
Fourier analysis for a variety of types of signals, including discrete time (or
parameter), continuous time (or parameter), finite duration, and infinite
duration. By "finite duration" we mean a signal with a finite domain of
definition; that is, the signal is only defined for a finite range of its indepen-
dent variable. The principal types of infinite duration signals considered
are absolutely summable (or integrable), finite energy, impulsive, and pe-
riodic. All of these signal types commonly arise in applications, although
sometimes only as idealizations of physical signals. Many of the basic ideas
are the same for each type, but the details often differ significantly. The
intent of this book is to highlight the common aspects in these cases and
thereby build intuition from the simple examples, which will be useful in
the more complicated examples where careful proofs are not included.
Traditional treatments tend to focus on infinite duration signals, either
beginning with the older notion of a Fourier series of a periodic function and
then developing the Fourier integral transform as the limit of a Fourier series
as the period approaches infinity, or beginning with the integral transform
and defining the Fourier series as a special case of a suitably generalized
transform using generalized functions (Dirac delta functions). Most texts
emphasize the continuous time case, with the notable exception of treat-
ments in the digital signal processing literature. Finite duration signals are
usually considered late in the game when the discrete Fourier transform
(DFT) is introduced prior to discussing the fast Fourier transform (FFT).
We here take a less common approach of introducing all of the basic types
of Fourier transform at the beginning: discrete and continuous time (or
xiv PREFACE

parameter) and finite and infinite duration. The DFT is emphasized early
because it is the easiest to work with and its properties are the easiest
to demonstrate without cumbersome mathematical details, the so-called
"delta-epsilontics" of real analysis. Its importance is enhanced by the fact
that virtually all digital computer implementations of the Fourier trans-
form eventually reduce to a DFT. Furthermore, a slight modification of
the DFT provides Fourier series for infinite duration periodic signals and
thereby generalized Fourier transforms for such signals.

This approach has several advantages. Treating the basic signal types
in parallel emphasizes the common aspects of these signal types and avoids
repetitive proofs of similar properties. It allows the basic properties to
be proved in the simplest possible context. The general results are then
believable as simply the appropriate extensions of the simple ones even
though the detailed proofs are omitted. This approach should provide more
insight than the common engineering approach of quoting a result such as
the basic Fourier integral inversion formula without proof. Lastly, this
approach emphasizes the interrelations among the various signal types, for
example, the production of discrete time signals by sampling continuous
time signals or the production of a finite duration signal by windowing
an infinite duration signal. These connections help in understanding the
corresponding different types of Fourier transforms.

No approach is without its drawbacks, however, and an obvious prob-

lem here is that students interested primarily in one particular signal type
such as the classical infinite duration continuous time case or the finite
duration discrete time case dominating digital signal processing texts may
find the variety of signal types annoying and regret the loss of additional
detailed results peculiar to their favorite signal type. Although a sequential
treatment of the signal types would solve this problem, we found the cited
advantages of the parallel approach more persuasive. An additional fac-
tor in favor of the parallel approach is historical. The Stanford University
course (EE261) in which these notes were developed has served as a general
survey and review for a diverse group of students from many departments.
Most students taking this course had had bits and pieces of Fourier theory
scattered throughout their undergraduate courses. Few had had any form
of overview relating the apparently different theories and taking advantage
of the common ideas. Such a parallel treatment allows readers to build
on their intuition developed for a particular signal class in order to under-
stand a collection of different applications. A sequential treatment would
not have accomplished this as effectively.
PREFACE xv

Synopsis
The topics covered in this book are:

1. Signals and Systems. This chapter develops the basic definitions and
examples of signals, the mathematical objects on which Fourier trans-
forms operate, the inputs to the Fourier transform. Included are
continuous time and discrete time signals, two-dimensional signals,
infinite and finite duration signals, time-limited signals, and periodic
signals. Combinations of signals to produce new signals and systems
which operate on an input signal to produce an output signal are
defined and basic examples considered.

2. The Fourier Transform. The basic definitions of Fourier transforms

are introduced and exemplified by simple examples. The analytical
and numerical evaluation of transforms is considered. Several trans-
forms closely related to the Fourier transform are described, including
the cosine and sine transforms, the Hartley transform, the Laplace
transform, and z-transforms.

3. Fourier Inversion. Basic results on the recovery of signals from their

transform are developed and used to describe signals by Fourier inte-
grals and Fourier series.

4. Basic Properties. This chapter is the heart of the book, developing the
basic properties of Fourier transforms that make the transform useful
in applications and theory. Included are linearity, shifts, modulation,
Parseval's theorem, sampling, the Poisson summation formula, alias-
ing, pulse amplitude modulation, stretching, downsampling and up-
sampling, differentiating and differencing, moment generating, band-
width and pulse width, and symmetry properties.

5. Generalized Transforms and Functions. Here the Fourier transform

is extended to general signals for which the strict original definitions
do not work. Included are Dirac delta functions, periodic signals, and
impulse trains.

6. Convolution. Fourier methods are applied to the analysis of linear

time invariant systems. Topics include impulse responses, superpo-
sition and convolution, the convolution theorem, transfer functions,
integration, sampling revisited, correlation, the correlation theorem,
pulsewidth and bandwidth in terms of autocorrelation, the uncer-
tainty relation, and the central limit theorem.
xvi PREFACE

7. Two-Dimensional Fourier Transforms. Methods of Fourier analysis

particularly useful for two-dimensional signals such as images are
considered. Topics include two-dimensional linear systems and re-
construction of 2D signals from projections.

8. Memoryless Nonlinearities. Fourier methods are shown to be useful

in certain nonlinear problems. The so-called "transform method"
of nonlinear analysis is described and applied to phase modulation,
frequency modulation, uniform quantization, and transform coding.

The final two topics are relatively advanced and may be tackled in any
order as time and interest permit. The construction of Fourier transform
tables is treated in the Appendix and can be referred to as appropriate
throughout the course.

Instructional Use
This book is intended as an introduction and survey of Fourier analysis for
engineering students and practitioners. It is a mezzanine level course in the
sense that it is aimed at senior engineering students or beginning Master's
level students. The basic core of the course consists of the unstarred sections
of Chapters 1 through 6. The starred sections contain additional details
and proofs that can be skipped or left for background reading without
classroom presentation in a one quarter course. This core plus one of the
topics from the final chapters constitutes a one quarter course. The entire
book, including many of the starred sections, can be covered in a semester.
Many of the figures were generated using Matlab ™ on both unix ™
and Apple Macintosh ™ systems and the public domain NIH Image pro-
gram (written by Wayne Rasband at the U.S. National Institutes of Health
and available from the Internet by anonymous ftp from zippy.nimh.nih.gov
or on floppy disk from NITS, 5285 Port Royal Rd., Springfield, VA 22161,
part number PB93-504568) on Apple Macintosh ™ systems.
The problems in each chapter are intended to test both fundamentals
and the mechanics of the algebra and calculus necessary to find transforms.
Many of these are old exam problems and hence often cover material from
previous chapters as well as the current chapter. Whenever yes/no answers
are called for, the answer should be justified, e.g., by a proof for a positive
answer or a counterexample for a negative one.
PREFACE xvii

Recommended Texts
Fourier analysis has been the subject of numerous texts and monographs,
ranging from books of tables for practical use to advanced mathematical
treatises. A few are mentioned here for reference. Some of the classic texts
still make good reading.
The two most popular texts for engineers are The Fourier Transform
and its Applications by R. Bracewell [6] and The Fourier Integral and its
Applications by A. Papoulis [24]. Both books are aimed at engineers and
emphasize the infinite duration, continuous time Fourier transform. Cir-
cuits, Signals, and Systems by W. McC. Siebert [30] is an excellent (and
enormous) treatment of all forms of Fourier analysis applied to basic circuit
and linear system theory. It is full of detailed examples and emphasizes ap-
plications. A detailed treatment of the fast Fourier transform may be found
in O. Brigham's The Fast Fourier Transform and its Applications [9]. Treat-
ments of two-dimensional Fourier transforms can be found in Goodman [18J
and Bracewell [8] as well as in books on image processing or digital image
processing. For example, Gonzales and Wintz [17] contains a variety of
applications of Fourier techniques to image enhancement, restoration, edge
detection, and filtering.
Mathematical treatments include Wiener's classic text The Fourier In-
tegral and Certain of its Applications [36], Carslaw's An Introduction to the
Theory of Fourier's Series and Integrals [10], Bochner's classic Lectures on
Fourier Integrals [3], Walker's Fourier Analysis [33], and Titchmarsh's In-
troduction to the Theory of Fourier Integrals [32]. An advanced and modern
(and inexpensive) mathematical treatment can also be found in An Intro-
duction to Harmonic Analysis by Y. Katznelson [21]. An elementary and
entertaining introduction to Fourier analysis applied to music may be found
in The Science of Musical Sound, by John R. Pierce [25].
Discrete time Fourier transforms are treated in depth in several books
devoted to digital signal processing such as the popular text by Oppenheim
and Schafer [23].

Some Notation
We will deal with a variety of functions of real variables. Let n denote
the real line. Given a subset r of the real line n, a real-valued function g
of a real variable t with domain of definition r is an assignment of a real
number g(t) to every point in T Thus denoting a function 9 is shorthand
for the more careful and complete notation {g(t)j t E r} which specifies
the name of the function (g) and the collection of values of its argument
xviii PREFACE

for which it is defined. The most common cases of interest for a domain of
definition are intervals of the various forms defined below.
• T = n, the entire real line.
• T = (a, b) = {r : a < r < b}, an open interval consisting of the points
between a and b but not a and b themselves. The real line itself is
often written in this form as n = (-00,00).
• T = [a, b] = {r : a ~ r ~ b}, a closed interval consisting of the points
between a and b together with the endpoints a and b (a and b both
finite).
• T = [a, b) = {r : a ~ r < b}, a half open (or half closed) interval con-
sisting of the points between a and b together with the lower endpoint
a (a finite).

III T = (a, b] = {r : a < r ~ b}, a half open (or half closed) interval
consisting of the points between a and b together with the upper
endpoint b (b finite).
• T = Z = {- .. , -1,0, 1, ... }, the collection of integers.
• T = ZN
N-l.
= {O, 1, ... , N - I}, the collection of integers from °
through

Complex numbers can be expressed as

z = x + iy,
where x = ~(z) is the real part of z, y = ~(z) is the imaginary part of z,
and

(often denoted by j in the engineering literature). Complex numbers can

also be represented in polar coordinates or magnitude/phase form as

The magnitude or modulus A is given by

If we restrict the phase angle () to be within [-~,~) (or, equivalently, in

[0,7r), then () is given by the principal value of the inverse tangent (or
arctangent): () = arctan(~) or () = tan-l(~) in radians. The quantity ()/27r
is the phase in units of cycles. Figure 0.1 illustrates these quantities. Here,
PREFACE xix

Imaginary

z
y

Real
x

Figure 0.1: Complex Numbers

the complex number is denoted by the vector z, with magnitude A and

phase 8. It has a real component x and imaginary component y.
The complex conjugate z* is defined as x - iy. The real and imaginary
parts of z are easily found by

x = !R(z) = ~(z + z*), y = ~(z) = ~(z - z*)

Sinusoids and complex exponentials play a basic role throughout this

book. They are related by Euler's formulas: For all real ()

ei9 = cos(} + isin(}

eilJ + e- ilJ
cos(} = 2
eilJ _ e- ilJ
sin 8 = 2i
A complex-valued function g = {get); t E T} is an assignment of
complex-valued numbers get} to every t E T.
A signal {get}; t E R} is said to be even if

g( -t) = get}; t E R.
xx PREFACE

It is said to be odd if
g( -t) = -g(t); tEn.
A slight variation of this definition is common: strictly speaking, an odd
signal must satisfy g(O) = 0 since -g(O) = g(O). This condition is some-
times dropped so that the definition becomes g( -t) = -g(t) for all t ;f O.
For example, the usual definition of the sign function meets the strict def-
inition, but the alternative definition (which is +1 for t ~ 0 and -1 for
t < 0) does not. The alternative definition is, however, an odd function if
one ignores the behavior at t = O. A signal is Hermitian if
g(-t) = g*(t); tEn.
For example, a complex exponential g(t) = ei21r/ot is Hermitian. A signal
is anti-Hermitian if
g( -t) = -g*(t); tEn.
As examples, sin t and te-'>'Itl are odd functions of tEn, while cos t
and e-'>'Itl are even.
We shall have occasion to deal with modular arithmetic. Given a posi-
tive real number T > 0, any real number a can be written uniquely in the
form a = kT + r where the "remainder" term is in the interval [0, T). This
formula defines a mod T = r, that is, a mod T is what is left of a when the
largest possible number of integer multiples of T is subtracted from a. This
is often stated as "a modulo T." The definition can be summarized by
a mod T = r if a = kT + r where r E [0, T) and k E Z. (0.1)
More generally we can define modular arithmetic on any interval [a, b),
b> a by
a mod [a, b) = r if a = kT + r where r E [a, b) and k E Z. (0.2)
Thus the important special case a mod T is an abbreviation for a mod
[0, T). By "modular arithmetic" is meant doing addition and subtraction
within an interval. For example, (0.5 + 0.9) mod 1 = 1.14 mod 1 = .14 mod
1 and (-0.3) mod 1 = 0.7.

Acknowledgements
We gratefully acknowledge our debt to the many students who suffered
through early versions of this book and who considerably improved it by
their corrections, comments, suggestions, and questions. We also acknowl-
ege the Industrial Affiliates Program of the Information Systems Labora-
tory, Stanford University, whose continued generous support provided the
computer facilities used to write and design this book.
Chapter 1

Signals and Systems

1.1 Waveforms and Sequences

The basic entity on which all Fourier transforms operate is called a signal.
The model for a signal will be quite general, although we will focus almost
entirely on a few special types. Intuitively, the definition of a signal should
include things like sinusoids of the form sin t or, more precisely, {sin tj t E
(-00, oo)}, as well as more general waveforms {g(t)j t E T}, where T could
be the real line or the positive real line [0,00) or perhaps some smaller
interval such as [O,T) or [-T/2,T/2). In each of these cases the signal is
simply a function of an independent variable (or parameter), here called t
for "time." In general, however, the independent variable could correspond
to other physical quantities such as "space."
T provides the allowable values of the time index or parameter and is
called the domain of definition or, simply, domain of the signal. It is also
called the index set of the signal since one can think of the signal as an
indexed set of values, one value for each choice of the index t. This set can
be infinite duration, i.e., infinite in extent, as in the case of the real line
R = (-00,00) , or finite duration, i.e., finite in extent as in the case [0, T).
There is potential confusion in this use of the phrase "finite duration" since

°
it could mean either a finite extent domain of definition or a signal with
an infinite extent index set with the property that the signal is except
on a finite region. We adopt the first meaning, however, and hence "finite
duration" is simply a short substitute for the more precise but clumsy

°
"finite extent domain of definition." The infinite duration signal with the
property that it is except for a finite region will be called a time-limited
signal. For example, the signal = {sin tj t E [O,211")} has finite duration,
2 CHAPTER 1. SIGNALS AND SYSTEMS

while the signal {h(t)j t E (-oo,oo)} defined by

h(t)
°
= {sint t E [0,211')
tEn, t ¢ [0,211')

is infinite duration but time-limited. Clearly the properties of these two

signals will strongly resemble each other, but there will be subtle and im-
portant differences in their analysis.
When the set T is continuous as in these examples, we say that the
signal is a continuous time signal, or continuous parameter signal, or simply
a waveform.
The shorter notation get) is often used for the signal {g(t)j t E T},
but this can cause confusion as g(t) could either represent the value of the
signal at a specific time t or the entire waveform g(t) for all t E T. To
lessen this confusion it is common to denote the entire signal by simply the
name of the function gj that is,

9 = {g(t)j t E T}

when the index set T is clear from context. It is also fairly common practice
to use boldface to denote the entire signalj that is, g = {g(t)j t E T}.
Signals can also be sequences, such as sampledsinusoids {sin(nT)j n E
Z}, where Z is the set of all integers { ... , -2, -1,0, 1, 2, ... }, a geometric
progression {rnj n = 0, 1,2, ... }, or a sequence of binary data {unj n E Z},
where all of the Un are either 1 or 0. Analogous to the waveform case we
can denote such a sequence as {g(t)j t E T} with the index set T now
being a set of integers. It is more common, however, to use subscripts
rather than functional notation and to use indexes like k, I, n, m instead
of t for the index for sequences. Thus a sequence will often be denoted
by {gnj nET}. We still use the generic notation 9 for a signal of this
type. The only difference between the first and second types is the nature
of the index set T. When T is a discrete set such as the integers or the
nonnegative integers, the signal 9 is called a discrete time signal, discrete
parameter signal, sequence, or time series. As in the waveform case, T may

°
have infinite duration (e.g., all integers) or finite duration (e.g., the integers
from to N - 1).
In the above examples the index set T is one-dimensional, that is, con-
sists of some collection of real numbers. Some signals are best modeled as
having multidimensional index sets. For example, a two-dimensional square
sampled image intensity raster could be written as {gn,kj n = 1, ... , Kj k =
1, ... ,K}, where each gn,k represents the intensity (a nonnegative number)
of a single picture element or pixel in the image, the pixel located in the
nth column and kth row of the square image. Note that in this case the
1.1. WAVEFORMS AND SEQUENCES 3

dummy arguments correspond to space rather than time. There is nothing

magic about the name of a dummy variable, however, and we could still
use t here if we wished.
For example, a typical magnetic resonance (MR) image consists of a

°
256 x 256 square array of pixel intensities, each represented by an integer
from (black, no light) to 29 - 1 = 511 (white, fully illuminated). As we
shall wish to display such images on screens which support only 8 and not
9 bits in order to generate examples, however, we shall consider MR images

°
to consist of 256 x 256 square array of pixel intensities, each represented by
an integer from to 28 -1 = 255. We note that, in fact, the raw data used
to generate MR images constitute (approximately) the Fourier transform of
the MR image. Thus when MR images are rendered for display, the basic"
operation is an inverse Fourier transform.
A square continuous image raster might be represented by a wave-
form depending on two arguments {g(x, y)j x E [0, al, y E [0, a]}. A
three-dimensional sequence of image rasters could be expressed by a sig-
nal {gn,k,lj n = 0,1,2, ... , k = =
1,2, ... ,K, 1 1,2, ... ,K}, where now
n is the time index (any nonnegative integer) and k and 1 are the spatial
indices. Here the index set T is three-dimensional and includes both time
and space.
In all of these different signal types, the signal has the general form
9 = {g(t)j t E T},

where T is the domain of definition or index set of the signal, and where g(t)
denotes the value of the signal at "time" or parameter or dummy variable
t. In general, T can be finite, infinite, continuous, discrete, or even vector
valued. Similarly g(t) can take on vector values, that is, values in Euclidean
space. We shall, however, usually focus on signals that are real or complex
valued, that is, signals for which g(t) is either a real number or a complex
number for all t E T. As mentioned before, when T is discrete we will often
write gt or gn or something similar instead of g(t).
In summary, a signal is just a function whose domain of definition is T
and whose range is the space of real or complex numbers. The nature of T
determines whether the signal is continuous time or discrete time and finite
duration or infinite duration. The signal is real-valued or complex-valued
depending on the possible values of g(t).
Although T appears to be quite general, we will need to impose some
structure on it to get useful results and we will focus on a few special cases
that are the most important examples for engineering applications. The
most common index sets for the four basic types of signals are listed in
Table 1.1. The subscripts of the domains inherit their meaning from their
place in the table, that is,
4 CHAPTER 1. SIGNALS AND SYSTEMS

DTFD = discrete time, finite duration

CTFD = continuous time, finite duration
DTID = discrete time, infinite duration
CTID = continuous time, infinite duration.

The superscripts will be explained shortly.

Duration Time
Discrete Continuous

(1) _ ~ (1) [
Finite TDTFD - ZN - {O, 1,···,N -I} TCTFD = O,T)

(2) .D. (2)

Infinite TDTID = Z = {"" -1,0, I,"'} TCTJD = 'R

Table 1.1: Common Index Sets for Basic Signal Types

It should be pointed out that the index sets of Table 1.1 are not the
only possibilities for the given signal types; they are simply the most com-
mon. The two finite duration examples are said to be one-sided since only
nonnegative indices are considered. The two infinite duration examples are
two-sided in that negative and nonnegative indices are considered. Com-
mon alternatives are to use two-sided sets for finite duration and one-sided
sets for infinite duration as in Table 1.2. Superscripts are used to dis-
tinguish between one- and two-sided time domains when convenient. They
will be dropped when the choice is clear from context. The modifications in

Duration Time
Discrete Continuous

(2) _ {
Finite TDTFD - -N,···,-I,O,I,···,N } 7,(2) _ [ T T)
CTFD - -2' 2

Infinite (1)
TDTID = {
0,1,'" }
Ti:~ID = [0,00)

Table 1.2: Alternative Index Sets for Basic Signal Types

transform theory for these alternative choices are usually straightforward.

1.2. BASIC SIGNAL EXAMPLES 5

We shall emphasize the choices of Table 1.1, but we shall often encounter
examples from Table 1.2. The careful reader may have noticed the use
of half open intervals in the definitions of finite duration continuous time
signals and be puzzled as to why the apparently simpler open or closed
intervals were not used. This was done for later convenience when we con-
struct periodic signals by repeating finite duration signals. In this case one
endpoint of the domain is not included in the domain as it will be provided
by another domain that will be concatenated.

1.2 Basic Signal Examples

One of the most important signals in Fourier analysis is the sinusoid.
Figure 1.1 shows several periods of a continuous time sinusoid {get) =
sin(21l"t); t E R}. A period of a signal g is aTE T such that g(t+T) = get)
for all t satisfying t + T E T. When we refer to the period rather than a
period of a signal, we mean the smallest period T > O. The period of the
sinusoidal g under consideration is 1. Since figures obviously cannot depict
infinite duration signals for the entire duration, it is not clear from the
figure alone if it is a piece of an infinite duration signal or the entire finite
duration signal. This detail must always be provided.

Figure 1.1: Continuous Time Sine Signal: Period 1

6 CHAPTER 1. SIGNALS AND SYSTEMS

A sinusoid can also be considered as a discrete time signal by sampling

the continuous time version. For example, consider the discrete time signal
g defined by

g(n) = sin(~;n)j n E Z. (1.1)

A portion of this signal is plotted in Figure 1.2. Note that the period has

0.5

~
0
~
0)

-0.5

-\

-20 20

Figure 1.2: Discrete Time Sine Signal: Period 10

changed because of the change in the definition of the variable correspond-

ing to "time." Not only has time been made discrete, it has also been
scaled.
We can consider yet another signal type using the same sinusoid by
both sampling and truncating. For example, form the one-sided discrete
time signal defined by

g(n) = sin(2; n)j n = {O, 1,··· ,31} (1.2)

as shown in Figure 1.3. Here the figure shows the entire signal, which is not
possible for infinite duration signals. For convenience we have chosen the
number of time values to be a power of 2j this will lead to simplifications
when we consider numerical evaluation of Fourier transforms. Also for
convenience we have chosen the signal to contain an integral number of
1.2. BASIC SIGNAL EXAMPLES 7

.. ····0···

O.S .

-0.5 .

-I

Figure 1.3: One-Sided Finite Duration Discrete Time Sine Signal: Period 8

periods. This simplifies some of the manipulations we will perform on this

signal.
The two previous figures are, of course, different in a trivial way. It
is worth understanding, however, that they depict distinct signal types
although they originate with the same function. For historical reasons,
the most common finite duration signals are one-sided. This often makes
indexing and numerical techniques a little simpler.
Another basic signal is the continuous time exponential

g(t) = ae-Atj t E [0,00), (1.3)

where A > 0 and a is some real constant. The signal is depicted in Figure 1.4
for the case a = 1 and A = .9. This signal is commonly considered as a
two-sided signal {g(t)j t E 'R} by defining

ae-At t >0
9 ()
t ={ o -.
otherwise
(1.4)

The discrete time analog of the exponential signal is the geometric sig-
nal. For example, consider the signal given by the finite sequence

9 = {rnjn = 0, 1, ... ,N -1} = {1,r, ... ,r N - 1 }. (1.5)

8 CHAPTER 1. SIGNALS AND SYSTEMS

Figure 1.4: Continuous Time Exponential Signal

This signal has discrete time and finite duration and is called the finite
duration geometric signal because it is a finite length piece of a geometric
progression. It is sometimes also called a discrete time exponential. The
signal is plotted in Figure 1.5 for the case of r = .9 and N = 32.
Given any real T > 0, the box function DT(t) is defined for any real t
by
DT(t) = {1 It I ~ T. .
o otherWise (1.6)

A variation of the box function is the rectangle function considered by

Bracewell [6):
I It I < !
n(t) = { ~ It I = I
o otherwise
Note that n(t/2T) = DT(t) except at the edges t = T. The reason for this
difference in edge values is that, as we shall see later, signals with discon-
tinuities require special care when doing Fourier inversion. We shall see
that the rectangle function is chosen so as to ensure that it is reproduced
exactly after a sequence of a Fourier transform and an inverse Fourier trans-
form. Both box functions, when considered as continuous time signals, will
give the same transform, but Fourier inversion in the continuous time case
yields Bracewell's definition. On the other hand, the extra simplicity of
1.2. BASIC SIGNAL EXAMPLES 9

1.2r---~--~---~--~--~-----'--'

0.8

0.6
,--.
~
Ol 0.4

0.2 0

•
0
0
0
0
o 0 0 o 0
o 0 0 0
0

-0.2
0 10 15 20 25 30

Figure 1.5: Finite Duration Geometric Signal

DT(t) will prove useful, especially when used to define a discrete time sig-
nal, since then discontinuities do not pose problems. The notation rect(t)
is also used for n(t).
As an example, a portion of the two-sided infinite duration discrete time
box signal {D5(n); n E Z} is depicted in Figure 1.6 and a finite duration
one-sided box signal {D5(n); n E {O, 1"", 15} is depicted in Figure 1.7.
The corresponding continuous time signal {D5 (t); tEn} is depicted in
Figure 1.8.
The Kronecker delta function Ot is defined for any real t by
t=O
t;f:O'
This should not be confused with the Dirac delta or unit impulse which
will be introduced later when generalized functions are considered. The
Kronecker delta is primarily useful as a discrete time signal, as exemplified
in Figure 1.9. The Kronecker delta is sometimes referred to as a "unit
sample" or as the "discrete time impulse" or "discrete impulse" in the lit-
erature, but the latter terms should be used with care as "impulse" is most
associated with the Dirac delta and the two deltas have several radically
different properties. The Kronecker and Dirac delta functions will play sim-
ilar roles in discrete time and continuous time systems, respectively, but the
10 CHAPTER 1. SIGNALS AND SYSTEMS

1.2,,----~--~~--r_--~--~--___r_"1

0.6
.--,
~
0) 0.4 .;

0.2 .

30-----"_2.,-0---'-:-----'------'-10-----"20,-------,3',-10
-0.2 L _-:'c

Figure 1.6: Discrete Time Box Signal

1.2.-----r---..,.---r--~---.---r--~--,

0.8

0.6
....
.--,
'-"
0) 0.4

0.2

0 ... :, . ~ .0 ..
-0.2
0 2 4 6 10 12 14

Figure 1.7: One-sided Discrete Time Box Signal

1.2. BASIC SIGNAL EXAMPLES 11

1.2~--~--~r----.----"-----r------r-o

0.8 .C

-.:::,
0)
0.6

0.4

0.2

-0.2
-30 -20 -10 0 10 20 30

Figure 1.8: Continuous Time Box Signal

Kronecker delta function is an ordinary signal, while a Dirac delta is not;

the Dirac delta is an example of a generalized function and it will require
special treatment.
Note that
iSn = Oo(n); n E Z.

The sinc function sinc t is defined for real t by

. sin 1l't
slUct = - -
1l't
and is illustrated in Figure 1.10. The sinc function will be seen to be of
fundamental importance to sampling theory and to the theory of Dirac
delta functions and continuous time Fourier transform inversion.
The unit step function is defined for all real t by
t_{l t~O (1.7)
0 otherwise
U-l ( ) -

The notation indicates that the unit step function is one of a class of special
functions udt) related to each other by integration and differentiation. See,
e.g., Siebert [30]. The continuous time step function and the discrete time
step function are depicted in Figures 1.11 and 1.12.
12 CHAPTER 1. SIGNALS AND SYSTEMS

1.2,...------,-----,----,...------,-----,------,

............................ ........................, ........................., ........................................

0.8

0.6 ...................................................1.........................1 ..........................1. . . . . . . . . . . . .1 . . . . . . .

---....
'-'
Q') 0.4

0.2 ..........................1 ........................ 1. . . . . . . . . . . . .1. ..........................1. . . . . . . . . . . . .1. . . . . . .

-O·~3L.0---_~20:----....J_\':-0---0':-----:':\0,----~20:-----:'30

Figure 1.9: Discrete Time Kronecker Delta

1.2,...--...,---,----,-----,:---,...--...,---,----,-----,:-----,

0.8

0.6

---....
'-'
Q')
0.4

0.2

o.

-0.2

-\0 -8 -6 -4 -2 0 2 4 6 \0

Figure 1.10: Sinc Function

1.2. BASIC SIGNAL EXAMPLES 13

1.2r------.-----,-----.-----.----...-----,

0.8

0.6
....
.........
........
Q) 0.4

0.2

-0.2
-30 -20 -\0

Figure 1.11: Continuous Time Step Function

1.2r------.-----.-----.-,---...,:~----.-,----,

I I
..............,............................ ! ........................ ~GGo·o·o.eGG-o.!·oeeo.o.o.oeG;·o.o«).&o.o.o
;
0.8

0.6
....
.........
........
Q) 0.4 ................ ., ·······················1 ......................... !.... . ....................... -f ...

0.2 ....................... .

: :
o .o-o.o.oeeo.~.o.o.&o.o.o-G.!o.o.oeGoo.o.o+.. ........... ~..... .

~.2L---~---~---~---~---~--~
-30 -20 -\0 0 \0 20 30

Figure 1.12: Discrete Time Step Function

14 CHAPTER 1. SIGNALS AND SYSTEMS

As with the box functions, we consider two slightly different forms of

the unit step function. Following Bracewell [6], define the Heaviside unit
step function H(t) for all real t by

H(t) = {! t>O
t
t<O
= O. (1.8)

The Heaviside step function is used for the same purpose as the rectangle
function; the definition of its value at discontinuities as the midpoint be-
tween the values above and below the discontinuity will be useful when
forming Fourier transform pairs. Both step functions share a common
Fourier transform, but the inverse continuous time Fourier transform yields
the Heaviside step function.
The signum or sign function also has two common forms: The most
common (especially for continuous time) is
+1 if t > 0
sgn(t) ={ 0 if t = 0 (1.9)
-1 if t < o.
The most popular alternative is to replace the 0 at the origin by +1. The
principal difference is that the first definition has three possible values while
the second has only two. The second is useful, for example, when modeling
the action of a hard limiter (or binary quantizer) which has two possible
outputs depending on whether the input is smaller than a given threshold or
not. Rather than add further clutter to the list of names of special functions,
we simply point out that both definitions are used. Unless otherwise stated,
the first definition will be used. We will explicitly point out when the second
definition is being used.
The continuous time and discrete time sgn(t) signals are illustrated in
Figures 1.13-1.14.
Another common signal is the triangle or wedge /\(t) defined for all real
t by
/\(t) = { ~ - It I if It I < 1 (1.10)
otherwise.
The continuous time triangle signal is depicted in Figure 1.15. In order to
simplify the definition of the discrete time triangle signal, we introduce first
the time scaled triangle function /\T(t) defined for all real t and any T > 0:

(1.11)

Thus /\(t) = /\1 (t). The discrete time triangle is defined as /\T(n); n E Z
for any positive integer T. Figure 1.16 shows the discrete time triangle
signal /\5 (n).
1.2. BASIC SIGNAL EXAMPLES 15

0.5

,.-..,
0
~
0:>

·0.5

-I

-30 -20 -10 0 10 20 30

Figure 1.13: Continuous Time Sign Function

0.5

,.-..,
0
~
0:>

-0.5

20 30

Figure 1.14: Discrete Time Sign Function

16 CHAPTER 1. SIGNALS AND SYSTEMS

1.2,..---,----.,---,------,--,...--...,---.,---,----,.----,

0.8

0.6

----
.....
'"5; 0.4

0.2

-0.2
·5

Figure 1.15: Continuous Time Triangle Signal

1.2,.-----,----,...----,----,...----,-----,

0.8

0.6

0.2

o e&o.o.o..,eo.~.o OOo-&o.o.oe~o
0 0 0'0& ....... ; ............. O-e00.0.~-()eoo.o.OiJOo-~.O 0-090-0-0,0,

.0·~3LO---.-:'-20O----'.1':"0---0~---:':1O:----~20O----:'30

Figure 1.16: Discrete Time Triangle Signal

1.2. BASIC SIGNAL EXAMPLES 17

As a final special function that will be encountered on occasion, the nth

order ordinary Bessel function of the first kind is defined for real t and any

J"
integer n by
In(t) = 2.. eitsinq,-inq, d</>. (1.12)
271" _"
Bessel functions arise as solutions to a variety of applied mathematical
problems, especially in nonlinear systems such as frequency modulation
(FM) and quantization, as will be seen in Chapter 8. Figure 1.17 shows a
plot of the In(t) for various indices n.

Figure 1.17: Bessel Functions J n: Solid line n 0, dotted line n = 1,

dashed line n = 2, dash-dot line n = 3.

The basic properties and origins of Bessel functions may be found in

standard texts on the subject such as Watson [35) or Bowman [5]. For
example, Bessel functions can be expressed in a series form as

(1.13)

Table 1.3 summarizes several examples of signals and their index sets
along with their signal type. w, >. > 0, and x are fixed real parameters, and
m is a fixed integer parameter.
18 CHAPTER 1. SIGNALS AND SYSTEMS

Signal Time Domain Duration Comments

{sin1Tt; t E R} continuous infinite sinusoid, periodic
{sin1Tt; t E [a, I]} continuous finite sinusoid
{e iwt ; t E R} continuous infinite complex exponential
ie-At; t E [O,oo)} continuous infinite real exponential
{I; t E R} continuous infinite constant (dc)
{e-lI't2; t E R} continuous infinite Gaussian
{ei ll't 2; t E R} continuous infinite chirp
{H(t); t E R} continuous infinite Heaviside step function
{U-l (t); t E R} continuous infinite unit step function
{sgn(t); t E R} continuous infinite signum (sign) function
{2Bsinc(2Bt); t E R} continuous infinite sinc function
{Or(t); t E R} continuous infinite box, time-limited
{t; t E [-1, 1]} continuous finite ramp
{I\(t); t E R} continuous infinite triangle, time-limited
{t0 1 (t); t E R} continuous infinite time-limited ramp
{Jm(21Tt); t E R} continuous infinite Bessel function
{In(x); n E Z} discrete infinite Bessel function
{rn; n E {a,l, ... }} discrete infinite geometric progression
{rn; n E ZN} discrete finite
{e iwn ; n E Z} discrete infinite
{u_l(n); n E Z} discrete infinite unit step
ibn; n E Z} discrete infinite Kronecker delta

Table 1.3: Common Signal Examples

1.3 Random Signals

All of the signals seen so far are well defined functions and knowing the sig-
nal means knowing exactly its form and structure. Obviously in real life one
will meet signals that are not so well understood and are not constructible
in any useful and simple form from the above building blocks. One way
of modeling unknown signals is to consider them as having been produced
by a random process. Although probability and random processes are not
prerequisite for this book and we do not intend to teach such a general
subject here, it is fair to use important examples of such processes as a
means of producing interesting signals for illustration and analysis. We
shall confine attention to discrete time when considering randomly gener-
ated signals to avoid the mathematical and notational complexity involved
in dealing with continuous time random processes. Figure 1.18 shows an
1.4. SYSTEMS 19

example of a discrete time signal produced by a Gaussian random process

produced by Matlab's random number generator. The values are produced
as a sequence of independent, zero mean, unit variance Gaussian random
variables. (Don't be concerned if these terms are unfamiliar to you; the
point is the signal is produced randomly and has no nice description.) This

4,-----,------,-----,------,-----,-----"

···················+·······················f·········· ················;························k············ ............., ...................oc •••• ,••• ~

-:;:; 0 ·······a .. ·~·········;··························;······ ................... +.........................+.........................;.......................... !! ••• ~

0;
-I ..•.....
. ' •
.
-2 ,
......................... ;, ......................... ........................ .

-3

40·L-----~S----~IOL-----~IS----~WL---~~~----~30~

Figure 1.18: Sequence Produced by an Independent, Identically Distributed

Gaussian Random Process.

is an example of what is often called discrete time white Gaussian noise

and it is a common and often good model of random phenomena in signal
processing and statistical applications.

1.4 Systems
A common focus of many of the application areas mentioned in the in-
troduction is the action of systems on signals to produce new signals. A
system is simply a mapping which takes one signal, often called the input,
and produces a new signal, often called the output. A particularly trivial
system is the identity system which simply passes the input signal through
to the output without change (an ideal "wire"). Another trivial system is
one which sets the output signal equal to 0 regardless of the input (an ideal
"ground"). More complicated systems can perform a variety of linear or
20 CHAPTER 1. SIGNALS AND SYSTEMS

nonlinear operations on an input signal to produce an output signal. Of

particular interest to us will be how systems affect Fourier transforms of
signals.
Mathematically, a system (also called a filter) is a mapping C of an input
signal v = {v(t)j t E 7i} into an output signal w = {w(t)j t E To} = C(v).
When we wish to emphasize the output at a particular time, we write

w(t) = Ct(v).
Note that the output of a system at a particular time can, in principle,
depend on the entire past and future of the input signal (if indeed t cor-
responds to "time"). While this may seem unphysical, it is a useful ab-
straction for introducing properties of systems in their most general form.
We shall later explore several physically motivated constraints on system
structure.
The ideal wire mentioned previously is modeled as a system by C(v) = v.
An ideal ground is defined simply by C( v) = 0, where here 0 denotes a signal
that is 0 for all time.
In many applications the input and output signals are of the same type,
that is, 7i and To are the samej but they need not always be. Several
examples of systems with different input and output signal types will be
encountered in section 1.8. As the case of identical signal types for input
and output is the most common, it is usually safe to assume that this is the
case unless explicitly stated otherwise (or implied by the use of different
symbols for the input and output time domains of definition).
The key thing to remember when dealing with systems is that they map
an entire input signal v into a complete output signal w.
A particularly simple type of system is a memoryless system. A mem-
oryless system is one which maps an input signal v = {v(t); t E T} into an
output signal w = {w(t)j t E T} via a mapping of the form

w(t) = at(v(t»j t E T

so that the output at time t depends only on the current input and not on
any past or future inputs (or outputs).

1.5 Linear Combinations

By scaling, shifting, summing, or combining in any other way the pre-
viously treated signals a wide variety of new signals can be constructed.
Scaling a signal by a constant is the simplest such operation. Given a sig-
nal g = {g(t)j t E T} and a complex number a, define the new signal ag by
1.5. LINEAR COMBINATIONS 21

{ag( t); t E T}; that is, the new signal formed by multiplying all values of
the original signal by a. This production of a new signal from an old one
provides another simple example of a system, where here the system £ is
defined by £t(g) = ag(t); t E T.
Similarly, given two signals 9 and h and two complex numbers a and b,
define a linear combination of signals ag+bh as the signal {ag(t) + bh(t); t E
T}. We have effectively defined an algebra on the space of signals. This
linear combination can also be considered as a system if we extend the
definition to include multiple inputs; that is, here we have a system £ with
two input signals and an output signal defined by £t(g, h) = ag(t) + bh(t);
t E T.
As a first step toward what will become Fourier analysis, consider the
specific example of the signal shown in Figure 1.19 obtained by adding two
sines together as follows:

sines1l"n) + sin(~n)
g(n) = 2 .

The resulting signal is clearly not a sine, but it is equally clearly quite well

0.5

o 0 o 0

o
o 0 o 0

-0.5

-\

o 10 15 20 25 30

Figure 1.19: Sum of Sinusoids.

behaved and periodic. One might guess that given such a signal one should
be able to decompose it into its sinusoidal components. Furthermore, one
22 CHAPTER 1. SIGNALS AND SYSTEMS

might suspect that by using general linear combinations of sinusoids one

should be able to approximate most reasonable discrete time signals. The
general form of these two goals constitute a large part of Fourier analysis.
Suppose that instead of combining two sinusoids, however, we combine
a finite duration one-sided sinusoid and a random signal. If the signals of
Figures 1.3 and 1.18 are summed and divided by two, the result is that
of Figure 1.20. Unlike the previous case, it is not clear if one can recover

2.-----~----~----~------~----~----_ro

1.5

0.5 ... 0 0

....
----
-......-
0
Ol
-0.5

-1

-1.5

-2
0 10 15

Figure 1.20: Sum of Random Signal and Sine.

the original sinusoid from the sum_ In fact, several classical problems in
detection theory involve such sums of sinusoids and noise. For example:
given a signal that is known to be either noise alone or a sinusoid plus noise,
how does one intelligently decide if the sinusoid is present or not? Given
a signal that is known to be noise plus a sinusoid, how does one estimate
the amplitude or period or the phase of the sinusoid? Fourier methods are
crucial to the solutions of such problems. Although such applications are
beyond the scope of this book, we will later suggest how they are approached
by simply computing and looking at some transforms.

Linear Systems
A system is linear if linear combinations of input signals yield the corre-
sponding linear combination of outputs; that is, if given input signals v(1)
1.5. LINEAR COMBINATIONS 23

and v(2) and complex numbers a and b, then

Linearity is sometimes referred to as additivity or as superposition. A

system is nonlinear if it is not linear. As a special case linearity implies
that a mapping is homogeneous in the sense that

L(av) = aL(v)
for any complex constant a.
Common examples of linear systems include systems that produce an
output by adding (in discrete time) or integrating (in continuous time)
the input signal times a weighting function. Since integrals and sums are
linear operations, using them to define systems result in linear systems. For

1:
example, the systems with output w defined in terms of the input v by

w(t) = v(T)ht(T) dT

in the infinite duration continuous time case or the analogous

L
00

Wn = Vkhn,k
k=-oo

in the discrete time case yield linear systems. In both cases h t ( T) is a

weighting which depends on the output time t and is summed or integrated
over the input times T. We shall see in Chapter 6 that these weighted
integrals and sums are sufficiently general to describe all linear systems.
A good way to get a feel for linear systems is to consider some nonlinear
systems. The following systems are easily seen to be nonlinear:

Lt (v) = v 2 (t)
Lt (v) a + bv(t)
Lt (v) sgn( v(t))
Lt (v) sin( v( t))
Lt( v) e- i21fv (t) .

(Note that all of the above systems are also memoryless.) Thus a square
law device, a hard limiter (or binary quantizer), a sinusoidal mapping, and
a phase-modulator are all nonlinear systems.
24 CHAPTER 1. SIGNALS AND SYSTEMS

1.6 Shifts
The notion of a shift is fundamental to the association of the argument of a
signal, the independent variable t called "time," with the physical notion of
time. Shifting a signal means starting the signal sooner or later, but leaving
its basic shape unchanged. Alternatively, shifting a signal means redefining
the time origin. If the independent variable corresponds to space instead of
time, shifting the signal corresponds to moving the signal in space without
changing its shape. In order to define a shift, we need to confine interest
to certain index sets T. Suppose that we have a signal g = {get); t E T}
and suppose also that T has the property that if t E T and T E T, then
also t - T E T. This is obviously the case when T = n or T = Z, but it
is not immediately true in the finite duration case (which we shall remedy
shortly). We can define a new signal
g(T) = {g(Tl(t); t E T} = {get - T); t E T}
as the original signal shifted or delayed by T. Since by assumption t - T E T
for all t E T, the values get - T) are well-defined. The shifted signal can be
thought of as a signal that starts T seconds after get) does and then mimics
it. The property that the difference (or sum) of any two members of T is
also a member of T is equivalent to saying that T is a group in mathematical
terms. For the two-sided infinite cases the shift has the normal meaning.
For example, a continuous time wedge signal with width 2T is depicted in
Figure 1.21 and the same signal shifted by 2T is shown in Figure1.22.

get)

t
-2T -T 0 T 2T

Figure 1.21: The Wedge or Triangle Signal I\T(t)

The operation of shifting or delaying a signal by T can be viewed as a

system: the original signal put into the system produces an output that is
a delayed version of the original signal. This is a mathematical model for
an idealized delay line with delay T, which can be expressed as a system
1.6. SHIFTS 25

g(t)

t
-2T -T 0 T 2T

Figure 1.22: Shifted Triangle Signall\T(t - 2T)

C defined by Ct(v) = v(t - r). It should be obvious that this system is a

linear system.
The idea of a shift extends in a natural fashion to finite duration signals
by taking advantage of the relation between a finite duration signal and
an infinite duration signal formed by replicating the finite duration signal
forever: its periodic extension. The continuous time case is considered first.
To define carefully a periodic extension of a finite duration signal, we use
the modulo operation of (0.1).
Define the periodic extension g = {g(t); t E 'R} of a finite duration
signal 9 = {g(t); t E [0,Tn by

g(t) = g(t mod T); t E 'R. (1.14)

In the discrete time case, the natural analog is used. The periodic extension
g = {gn; n E Z} of a finite duration signal 9 = {gn; n E ZN} is defined by
g(t) = g(n mod N); t E Z. (1.15)

Periodic extensions for more complicated finite duration time domains

can be similarly defined using the general definition of modulo of (0.2).
Roughly speaking, given a finite duration signal 9 = {g(t); t E [0, T)}
(discrete or continuous time) then we will define the shifted signal g(T) =
{g(T)(t); t E T} as one period of the shifted periodic extension of g; that is,
if g(t) = g(t mod T), then
g(T)(t) = g(t - r) = g((t - r) mod T); t E [0, T).
This is called a cyclic shift and can be thought of as wrapping the original
finite duration signal around so that what is shifted out one end is shifted
back into the other. The finite duration cyclic shift is depicted in Figure 1.23
26 CHAPTER 1. SIGNALS AND SYSTEMS

• I t
T T
2"

(a) g(t) = I\T/S(t - f)

g(t)

11~6. . . . . .
o T
2"
T
t

(b) g(t) = I\T/S(t - f)

II
g(t)

o
I

T
2"

(c) g(t) = I\T/S(t - T) = I\T/S(t)

Figure 1.23: Shifted Finite Duration Triangle Signal
1.6. SHIFTS 27

for the triangle signal. The choice of index set [0, T) = {t : 0 :::; t < T}
does not include the endpoint T because the periodic extension starts its
replication of the signal at T, that is, the signal at time T is the same as
the signal at time O.
An alternative and equivalent definition of a cyclic shift is to simply
redefine our "time arithmetic" t - l' to mean difference modulo T (thereby
again making T a group) and hence we are defining the shift of {get); t E
[0, T)} to be the signal {g«t-1') mod T); t E [0, Tn.Since (t -1') mod T E
[0, T), the shifted signal is well defined.

Time-Invariant Systems
We have seen that linear systems handle linear combinations of signals in
a particularly simple way. This fact will make Fourier methods particulary
amenable to the analysis of linear systems. In an analogous manner, some
systems handle shifts of inputs in a particularly simple way and this will
result in further simplifying the application of Fourier methods.
A system C is said to be time invariant or shift invariant or stationary
if shifting the input results in a corresponding shift of the output. To be
precise, a system is time invariant if for any input signal v and any shift r,
the shifted input signal veT) = {vet - r); t E 1t} yields the shifted output
signal
(1.16)
In other words, if wet) = Ct ( {v(t); t E 1t}) is the output at time t when
v is the input, then wet - r) is the output at time t when the shifted signal
v T is the input.
One can think of a time-invariant system as one which behaves in the
same way at any time. If you apply a signal to the system next week at
this time the effect will be the same as if you apply a signal to the system
now except that the results will occur a week later.
Examples of time-invariant systems include the ideal wire, the ideal
ground, a simple scaling, and an ideal delay. A memoryless system defined
by wet) = at(v(t)) is time-invariant if at does not depend on t, in which
case we drop the subscript.
As an example of a system that is linear but not time invariant, consider
the infinite duration continuous time system defined by

wet) = Ct(v) = v(t) cos 27rJot.

This is a double sideband suppressed carrier (DSB-SC or, simply, DSB)
modulation system. It is easily seen to be linear by direct substitution. It
is not time invariant since, for example, if vet) = I, all t, then shifting the
28 CHAPTER 1. SIGNALS AND SYSTEMS

input by 7r /2 does not shift the output by 7r /2. Alternatively, the system
is time-varying because it always produces an output of 0 when 27r Jot is
an odd multiple of 7r /2. Thus the action of the system at such times is
different from that at other times.
Another example of a time varying system is given by the infinite du-
ration continuous time system

Lt(V) = vet) n (t).

n
This system can be viewed as one which closes a switch and passes the
input during the interval [- ~, but leaves the switch open (producing a
zero output) otherwise. This system is easily seen to be linear by direct
substitution, but it is clearly not time invariant. For example, shifting an
input of vet) = net) by 1 time unit produces an output of 0, not a shifted
square pulse. Another way of thinking about a time-invariant system is
that its action is independent of the definition of the time origin t = O.
A more subtle example of a time varying system is given by the con-
tinuous time "stretch" system, a system which compresses or expands the
time scale of the input signal. Consider the system which maps a signal
{v(t); t E R} into a stretched signal defined by {v(at); t E R}, i.e., we have
a system mapping L that maps an input signal {vet); t E R} into an output
signal {wet); t E R} where wet) = v(at). Assume for simplicity that a > a
so that no time reversal is involved.
Shift the input signal to form a new input signal {v1'(t);t E R} defined
by v1'(t) = vet - 1'). If this signal is put into the system, the output signal,
say wo(t), is defined by wo(t) = v1'(at) = v(at - 1').
On the other hand, if the unshifted v is put into the system to get
wet) = v(at), and then the output signal is delayed by 1', then w1'(t) =
wet - 1') = v(a(t - 1')) = v(at - a1'), since now w directly plugs t - l' into
the functional form defining w.
Since wo(t) and wet -1') are not equal, the system is not time invariant.
The above shows that it makes a difference in which order the stretch and
shift are done.

1. 7 Two-Dimensional Signals
Recall that a two-dimensional or 2D signal is taken to mean a signal of the
form {g(x,y); x E Tx, y E Ty}, that is, a signal with a two-dimensional
domain of definition. Two dimensional signal processing is growing in im-
portance. Application areas include image processing, seismology, radio as-
tronomy, and computerized tomography. In addition, signals depending on
two independent variables are important in applied probability and random
1.7. TWO-DIMENSIONAL SIGNALS 29

processes for representing two dimensional probability density functions and

probability mass functions which provide probabilistic descriptions of two
dimensional random vectors.
A particularly simple class of 2D signals is constructed by taking prod-
ucts of one-dimensional (or 1D) signals. For example, suppose that we have
two 1D signals {gx(x); x E Tx} and {gy(y); y E Ty}. For simplicity we
usually consider both 1D signals to have the same domain, i.e., Tx = Ty.
We now define a two-dimensional domain of definition as the cartesian prod-
uct space T = Tx x Ty consisting of all (x, y) such that x E Tx and y E Ty.
An example of a 2D function defined on this domain is

g(x,y) = gx(x)gy(Y)j x E Tx, y E Ty,

or, equivalently,

g(x, y) = gx{x)gy(y)j (x, y) E Tx x Ty.

A 2D signal of this form is said to be separable in rectangular coordinates.

A continuous parameter example of such a separable signal is the 2D
box signal
(1.17)
where 'R2 = 'R x 'R is the real plane. For display we often focus on a discrete
parameter version of the basic 2D signals. Consider the discrete time box
function of Figure 1.6. If glen) = g2(n) = D5(n), n E 7i = {-32,··· ,31},
then
g(n, k) = Ds(n)Ds(k)
is a 2D box signal. A question now arises as to how to display such a 2D
signal to help the reader visualize it. There are two methods, both of which
we will use. The first method depicts the signal in three dimensions, that is,
as a surface above the 2D domain. This is accomplished in Matlab by using
the mesh command and is illustrated in Figure 1.24. The mesh figure plots
the signal values at the collection of times in the time domain and collects
the points by straight lines to create a "wire figure" of the signal. Here
the signal appears as a skyscraper on a flat field. We choose a fairly small
domain of definition, 64 x 64 = 4096 index pairs so that the Matlab mesh
graphics can clearly show the shape and the individual pixels. The 64 x 64
points in the grid at which the signal is defined are called pixels or pels as
an abbreviation for "picture elements." The real images we will consider
as examples will have a larger domain of definition of 256 x 256 = 65,536
pixels.
An alternative method, which will be more useful when dealing with
nonnegative signals such as image intensity rasters, is to plot the intensities
30 CHAPTER 1. SIGNALS AND SYSTEMS

Figure 1.24: 2D Box: Three Dimensional Representation

in two dimensions. This we do using the public domain NIH Image program
as in Figure 1.25. Here the light intensity at each pixel is proportional to the
signal value, i.e., the larger the signal value, the whiter the pixel appears.
Image rescales the pixel values of a signal to run from the smallest value to
the largest value and hence the image appears as a light square in a dark
background.
Both mesh and image representations provide depictions of the same
2D signal.
The above 2D signal was easy to describe because it could be written as
a product, in two "separable" pieces. It is a product of separate signals in
each of the two rectangular coordinates. Another way to construct simple
signals that separate into product terms is to use polar coordinates. To
convert rectangular coordinates (x, y) into polar coordinates (r,8) so that

x = rcos8, y = rsin8. (1.18)

The radius r is given by

r = r(x,y) = ';x 2 + y2. (1.19)

If we restrict the phase angle 8 to [-rr/2,rr/2}, then 8 is given by the

1.7. TWO-DIMENSIONAL SIGNALS 31

Figure 1.25: 2D Box: Image Representation

principal value of the inverse tangent:

8(x, y) = tan- 1 'x#... (1.20)

Consider for example the one-dimensional signals 9 R (r) = sinc r for all
positive real r and ge (8) = 1 for all () E [-11', 11'). Form the 2D signal from
these two 1D signals by
g(x,y) = 9R (r)ge ((})
for all real x and y. Once again the signal is separable, but this time in
polar coordinates.
A simple and common special case of separable signals in polar coordi-
nates is obtained by setting
gs((}) = 1; for all (}
so that
g(x, y) = gR(r). (1.21)
32 CHAPTER 1. SIGNALS AND SYSTEMS

A 2D signal having this form is said to be circularly symmetric.

Perhaps the simplest example of a circularly symmetric signal separable
in polar coordinates is the 2D disc

(1.22)

This signal is a disc of radius r and height 1. For purposes of illustration,

we focus on a discrete parameter analog and consider the 2D signal having
a 64 x 64 pixel domain of definition defined by

g(n,k) = D lO (v'n2 + k 2 ); k,n E {-32, .. ·,31}.

The mesh representation of this signal is presented in Figure 1.26 and it is

seen to be approximately a circle or disk above a flat plane. The circle is
not exact because the domain of definition is discrete and not continuous. If
more pixels had been used, the approximation to the continuous parameter
2D disk would be better. At this lower resolution the individual pixels are
seen as small squares. These blocky artifacts are not visible in the higher
resolution images such as the 256 x 256 images to be considered shortly.
The image representation is seen in Figure 1.27, where it is seen as a white
circle on a black background.

Figure 1.26: Disk: Three Dimensional Representation

1.7. TWO-DIMENSIONAL SIGNALS 33

Figure 1.27: Disk: Image Representation

Another polar separable circularly symmetric signal is

g(x, y) = sinc( VX2 + y2); (x, y) E 1(,2. (1.23)

For illustration we again emphasize a discrete parameter analog, a 2D signal
having a 64 x 64 pixel domain of definition defined by

g(n, k) = sinc( Vn 2 + k 2 ); k, n E {-32, .. ·, 31}.

The mesh representation of this signal is presented in Figure 1.28. The

plotting program makes the discrete signal appear continuous. The signal
does not have an obvious image representation because it has negative
values and hence does not correspond to image intensities at all points.
We can make it into an image by taking the absolute value, resulting in
the mesh representation of Figure 1.29 and the image representation of
Figure 1.30. The discrete nature of the signal can be seen in the blockiness
of the image representation.
As alternatives to the previous artificial images representing examples
of 2D signals, we include two real world images as examples. The first is a
34 CHAPTER 1. SIGNALS AND SYSTEMS

Figure 1.28: 2D Sine: Three Dimensional Representation

Figure 1.29: Magnitude 2D Sine: Three Dimensional Representation

1.B. SAMPLING, WINDOWING, AND EXTENDING 35

Figure 1.30: 2D Sine: Image Representation

256 x 256 section of a digitized version of the Mona Lisa taken from the NIH
collection of image examples and depicted in Figure 1.31. The second image
is a magnetic resonance (MR) brain scan image, which we shall refer to as
"Eve." This image is 256 x 256 pixels and is 8-bit gray scale as previously
discussed. The printed version is, however, half-toned.

1.8 Sampling, Windowing, and Extending

The various types of signals are clearly related to each other. We con-
structed some of the examples of discrete time signals by "sampling" con-
tinuous time signals, that is, by only defining the signal at a discrete col-
lection of regularly spaced time values. This suggests that more generally
signals of a given type might naturally produce signals of another type.
Another reason for considering the relations among different types of sig-
nals is that often a physical signal can be well modeled by more than one
of the given types and it is useful to know which, if any, of the models is
best. For example, suppose that one observes a continuous time sine wave
36 CHAPTER 1. SIGNALS AND SYSTEMS

Figure 1.31: Mona Lisa

produced by a oscillator for T seconds. This could be modelled by a finite

duration waveform {sin(wt); t E [0, Tn, but it might be useful to consider

°
it as a piece of an infinite duration sine wave {sin(wt); tEn} or even as a
time-limited waveform that lasts forever and assumes the value for t not
in [0, T). Which model is more "correct"? None; the appropriate choice
for a particular problem depends on convenience and the goal of the anal-
ysis. If one only cares about system behavior during [0, T), then the finite
duration model is simpler and leads to finite limits of integrals and sums.
If, however, the signal is to be used in a system whose behavior outside
this time range is important, then the infinite (or at least larger) duration
model may be better. Knowing only the output during time [0, T) may
force one to guess the behavior for the rest of time, and that this can be
done in more than one way. If we know the oscillator behaves identically
for a long time, then repeating the sinusoid is a good idea. If we do not
know what mechanism produced the sinusoid, however, it may make more
sense to set unknown values to zero or something else. The only general
1.8. SAMPLING, WINDOWING, AND EXTENDING 37

Figure 1.32: Magnetic Resonance Image

rule is to use the simplest useful model.

All of the signal conversions considered in this section are examples of
systems where a signal of one type is converted into a signal of a different
type.

Continuous Time Infinite Duration to Discrete Time In-

finite Duration
Given an infinite duration continuous time signal, say {g(t); t E 'R.}, we
can form a discrete time infinite duration signal {gn; n E Z} by sampling:
given a fixed positive real number T (called the sampling period) define

gn = genT); n E Z; (1.24)

that is, gn is formed as the sequence of successive values of the waveform

g(t) each spaced T units apart. Note that the new signal is different from
38 CHAPTER 1. SIGNALS AND SYSTEMS

the original and the original mayor may not be reconstructible from its
sampled version. In other words, the sampling operation is not necessarily
invertible. One of the astonishing results in Fourier analysis (which we
will prove in a subsequent chapter) is the Whittaker-Shannon-Kotelnikov
sampling theorem which states that under certain conditions having to do
with the shape of the Fourier transform of 9 and the sampling period T,
the original waveform can (in theory) be perfectly reconstructed from its
samples. This result is fundamental to sampled-data systems and digital
signal processing of continuous waveforms. The sampling idea also can be
used if the duration of the original signal is finite.
Figure 1.1 shows a continuous time sinusoid having a frequency of one
Hz; that is, g(t) = sin(211't). Figure 1.33 shows the resulting discrete time
signal formed by sampling the continuous time signal using a sampling
period of T = .1, that is, sin(211't) for t = n/l0 and integer n. Note that
the sampled waveform looks different in shape, but it is still periodic and
resembles the sinusoid. Figure 1.2 shows the resulting discrete time signal
9n = sin(211'n/1O), where we have effectively scaled the time axis so that
there is one time unit between each sample.

O.S

........ 0
~
<:::I)

.(l.S

-1

-2 0 0.5 I.S 2

Figure 1.33: Sampled Sinusoid: Sampling Period .1

If instead we take a sampling period of 1/311' ~ .1061, the resulting signal

sin(211't) for t = n/311' is shown in Figure 1.34. Rescaling the time axis so
1.B. SAMPLING, WINDOWING, AND EXTENDING 39

as to have one time unit between consecutive samples yields the discrete
time signal gn = sin(2n/3) shown in Figure 1.35. This discrete time signal
is not periodic in n (for example, it never returns to the value 0 ;:: sin 0)
and it bears less resemblance to the original continuous time signal. This

o.S .~............~...... ..

° ,
o .................................... !....................;.................... ~ .......................................................,.... .
,0
,
-0.5
,0
.................. ......................................... y ....................... ..................... ··.... o· ...... 0·
·o" .... ·•· ...... ·........ · .. '.·~

.1 ............................0 ........................................°...,o: ............................ ,...... o" ............................. ~..... ~. 0 ...•

·2 -I.S ·1 ·o.S o o.s 1.5 2

Figure 1.34: Sampled Sinusoid: Sampling Period 1/31r

simple example shows that discrete time signals obtained from continuous
time signals can be quite different in appearance and behavior even if the
original signal being sampled is fixed.

Discrete Time Infinite Duration to Continuous Time In-

finite Duration
Given a discrete time signal, we can construct a continuous time signal by
using the discrete time signal to "modulate" (Le., modify) shifted versions
of a waveform. For example, suppose that gn is a discrete time signal and
{pet); t E 'R} is a continuous time signal such as an ideal pulse

p
(t) _
-
{I 0
if 0 ::; t < T
otherwise (1.25)

as depicted in Figure 1.36 for T ;:: .1. This pulse is an example of a time-
limited signal, an infinite duration signal that is nonzero only on an interval
40 CHAPTER 1. SIGNALS AND SYSTEMS

0.5

~ 0
~

-0.5

-1

-20 -15 10 15 20

Figure 1.35: Discrete Time Sinusoid: Sampling Period 1/37r

1.2,-------,---,-------,---,-------.---;-------,------,

0.8

0.6
/"'0,

~
~ 0.4

0.2

-0.2
-2 -1.5 -1 -0.5 0 O.S 1.5 2

Figure 1.36: Ideal Rectangular Pulse

1.8. SAMPLING, WINDOWING, AND EXTENDING 41

of finite length. Note that except for the point t = T

T
pet) = °T/2(t - 2")' (1.26)

a shifted box function. If we had defined pet) to be 1 at t = T then (1.26)

would hold everywhere, but that would cause trouble with the application
we consider next. Alternatively,
(1.27)
For a fixed T, pet - T) is a delayed version of pet), which here means
it is 1 for T S t < 2T and 0 otherwise. Similarly, for any integer n
pet - nT) is a delayed version of pet) that is 1 for nT S t < (n + l)T and
o otherwise. For each n the signal {pet - nT); t E 'R.} is thus a pulse that
is 1 for nT S t < (n + l)T and 0 otherwise and hence the pulses in this
collection have no overlap with each other. Form a continuous time signal
{get); t E 'R.} as

L
00

get) = gnp(t - nT). (1.28)

n=-oo
This is an example of a pulse amplitude modulation (PAM) system where
the "pulses" p(t-nT) are modulated by the values of gn' For the flat pulse
under consideration, during the time interval [nT, (n + l)T) the output of
the PAM signal is simply gn'
If the sequence gn is the sampled sinusoid of Figure 1.2, then the re-
sulting PAM signal is as depicted in Figure 1.37. If the waveforms pet) are
the ideal pulses of (1.25) and if the gn are the samples genT) of a wave-
form, then the overall operation is called a sample-and-hold system. Here
we have chosen the pulse width to match the sampling period used origi-
nally to sample the sinusoid so that the reconstructed signal resembles the
original signal. The pulse width and the original sampling period need not
be the same, however. For example, doubling the pulse width yields the
stretched out sample-and-hold waveform of Figure 1.38.
Note that there is no guarantee that in a sample-and-hold system get)
and get) will be the same; in fact they will almost always be different
waveforms as above. (If the pulse pet) is 0 for t < 0 and t ~ T, then at
least the waveforms get) and get) will agree at the sample times nT.)
An alternative means of generating the continuous time signal would be
to interpolate between successive values of the discrete time signal.

Infinite Duration to Finite Duration

A finite duration signal can be obtained from an infinite duration signal
by simply truncating a sequence g = {gn; n E Z} to form a finite duration
42 CHAPTER 1. SIGNALS AND SYSTEMS

0.5

....-- 0
~
<::ll

-0.5

-1

-2 -1.5 -1 -0.5 o 0.5 1.5

Figure 1.37: Pulse Amplitude Modulated (PAM) Sequence

0.5

....-- 0
~
<::ll

-0.5

-1

-2 -1.5 -1 -0.5 0 0.5 1.5

Figure 1.38: Pulse Amplitude Modulated Sequence: Doubled Pulse Width

1.B. SAMPLING, WINDOWING, AND EXTENDING 43

sequence B = {gn; n E {O, I, ... ,N - I}} . More generally, the truncation

can also include windowing. Define a window function Wn of length N so
that Wn = 0 if n is not in some finite subset W, say ZN = {O, 1, ... , N -I},
of Z. The truncated and windowed

9 = {Bn; n = 0, 1, ... , N - I} = {wngn ; n = 0, 1, ... , N - I}. (1.29)

A common choice of Wn is to set Wn = 1 for all n E ZN and otherwise.
This is called the boxcar window function. Again observe that something
°
has been lost in forming the new signal from the old and that the original
infinite duration signal is in general not perfectly recoverable from the finite
duration signal.
The same idea can be used to construct finite duration continuous time
signals from infinite duration continuous time signals. For example, given

°
a continuous time signal {g(t); t E R}, we can define a continuous time
window function w(t) with w(t) = for t not in [0, T) and then define the
windowed and truncated signal 9 = {g(t) = g(t)w(t); t E [0, T)}. Once
again, the constant window is called a "boxcar" window. If the continuous
I Hz sine wave (a portion of which is shown in Figure 1.1) is truncated
to the time interval [-.5, .5], then the resulting finite duration signal is as
shown in Figure 1.39.

O.S

-:;;- 0 ..
.......
~

-0.5

-I

-O.S -0.4 -0.3 -0.2 -0.1 0 0.1 0.4 0.5

Figure 1.39: Finite Duration Signal by Truncation

44 CHAPTER 1. SIGNALS AND SYSTEMS

Finite Duration to Infinite Duration

Zero Filling
An infinite duration signal can be constructed from a finite duration signal
by "filling in" the missing signal values according to some rule. Suppose
=
that 9 {gn; n = 0,1, ... ,N - I} is a finite duration signal. We can define
an infinite duration signal Y = {Yn; n E Z} by "zero filling" as follows:

_ {gn if n E ZN
gn = 0 otherwise (1.30)

The infinite duration signal has simply taken the finite duration signal and
inserted zeros for all other times. Observe that if the finite duration signal
was originally obtained by windowing an infinite duration signal, then it is
likely that the infinite duration signal constructed as above from the finite
duration signal will differ from the original infinite duration signal. The
one notable exception will be if the original infinite duration signal was in
fact 0 outside of the window ZN, in which case the original signal will be
perfectly reconstructed. Extending a finite duration signal by zero filling
always produces a time-limited signal.
In a similar fashion, a continuous time finite duration signal can be
extended by zero filling. The signal {get); t E [0, T)} can be extended to
the infinite duration signal

9 = {g(t) t E [0, T)
o otherwise
As an example, extending the finite duration sinusoid of Figure 1.39 by
zero filling produces an infinite duration signal which has one period of
a sinusoid and is zero elsewhere, as illustrated in Figure 1.40. Another
example is given by noting that the two-sided continuous time ideal pulse
can be viewed as a one-sided box function extended by zero filling.

Periodic Extension
Another approach to constructing an infinite duration signal from a finite
duration signal is to replicate the finite duration signal rather than just
insert zeros. For example, given a discrete time finite duration signal 9 =
{gn; n = 0,1"", N - I} we can form an infinite duration signal 9 =
{9n; n E Z} by defining
9n = gnmodN, (1.31 )
where the mod operation was defined in (0.1). Note that the infinite dura-
tion signal 9 has the property that 9n+N = 9n for all integers n; that is, it is
1.B. SAMPLING, WINDOWING, AND EXTENDING 45

0.5

01----,-------,

-0.5

-1

-2 -1.5 -1 -0.5 o 0.5 1.5 2

Figure 1.40: Signal Extension by Zero Filling

periodic with period N. We refer to 9 as the periodic extension of the finite

duration signal. Observe that if the finite duration signal were obtained via
a boxcar window function of length N, then the periodic extension of the
finite duration signal will not equal the original signal unless the original
signal was itself periodic with period N.
Given a continuous time finite duration signal 9 = {g(t); t E [0, T)} we
can similarly form a periodic extension g(t) as in the discrete time case:

g(t) = g(t mod T); t E R. (1.32)

As an example, consider the continuous time finite duration ramp 9 =

{g(t); t E [0,1)} defined by
g(t) = t (1.33)
and depicted in Figure 1.41. The periodic extension of this having period
1 is the sawtooth signal t mod (1) depicted in Figure 1.42.
A convenient form for the periodic extension that relates the zero filled
and periodic extension is given by

L
00

g(t) = g(t - nT), (1.34)

n=-oo
46 CHAPTER 1. SIGNALS AND SYSTEMS

1.2 ..--~-~-~-~-~-..,.--..,.--..,.--..,.----,

0.8

0.6

----
~
0) 0.4

0.2

-0.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Figure 1.41: Continuous Time Ramp

-0.2 L . . . . _ L - - _ - ' - - _ - ' - - _ - ' - - _ - ' - - _ - ' - - _ - ' - - _ - ' - - _ - ' - - - - '
-5 -4 -3 -2 -I 0 2 4

Figure 1.42: Sawtooth Signal

1.9. PROBABILITY FUNCTIONS 47

where jj is the zero filled extension.

We shall see that the theory of finite duration signals is intimately con-
nected with that of infinite duration periodic signals and both periodic and
zero filled extensions will prove useful for different applications.
Although each transformation of one signal type can be considered to be
a system, more complex systems may have many such transformations. For
example, a sampled data system may start with continuous time signals and
then sample to form discrete time signals for use in digital signal processing.
Digital audio is a popular example. A digital-to-analog converter, such as
is used in a compact disc player, begins with discrete time signals (possibly
stored in a sequence of memory locations) and produces a continuous time
signal.

1.9 Pro bability Functions

It is important to keep in mind that a signal is just a function of a dummy
variable or parameter. The idea of a waveform or data sequence is perhaps
the most common example in engineering, but transforms are important for
a variety of other examples as well. We close this section with two examples
of important special cases that do not immediately fit in the category of
waveforms and sequences, but which can be profitably modeled as signals.
A discrete parameter signal {Pn; nET} such that Pn :::: 0 for all nand

(1.35)

is called a probability mass function or pm! for short. (Here T is called the
alphabet.) A continuous parameter signal {p( x); x E T} with the properties
that p(x) :::: 0 all x and
r
iXET
p(x)dx =1 (1.36)

is called a probability density function or pdf. These quantities playa fun-

damentally important role in probability theory. Here we simply note that
they are a source of examples for demonstrating and practicing Fourier
analysis.
Probability functions are simply nonnegative signals with integrals or
sums normalized to 1. Nonnegative signals having bounded sums or inte-
grals will also arise in other applications, e.g., power spectral densities.
Some of the more common named probability functions are listed below.
The parameters p, A, A, and a are positive real numbers, with 0 < p < 1.
The parameters m and b > a are real numbers. n is a positive integer.
48 CHAPTER 1. SIGNALS AND SYSTEMS

The Bernoulli pmf: p(k) = pk(1 - p)l-kj k E r = {a, I}.

The binomialpmf: p(k) = ( ~ )pk(l_ p)n-k j k E T= {0,1,2,···,n}.
The geometric pmf: p(k) = p(1 - p)k-lj k E = p, 2", .}. r
The Poisson pmf: p(k) = >..ke->'/k!j k E 7= {a, 1,2", .}.
The uniform pdf: 7 = R. f(r) = *Df(r - m)j r E 7 = R.
The exponential pdf: f(r) = >"e->.ru_1(t)j r E 7 = R.
The Laplacian pdf: f(r) = ~e->'Irl; r E 7 = n.
The Gaussian pdf: f(r) = (2rru 2 )-1/2 e-(r-m)2/2<T 2 j r E 7 = R.
Probability functions can also be two-dimensional, e.g., p = {Pn,kj n E
ZN, k E ZN}. Because of the requirement of nonnegativity, a 2D probabil-
ity function is mathematically equivalent to a 2D intensity image that has
been normalized.

1.10 Problems
1.1. Which of the following signals are periodic and, if so, what is the
period?
(a) {sin(2rrft)j t E (-00, oo)}
(b) {sin(2rrfn); n E { ... , -1,0,1, ... }} with f a rational number.
(c) Same as the previous example except that f is an irrational
number.
(d) n=~=l sin(2rr fatn); t E (-00, oo)}
(e) {sin(2rrfot) +sin(2rr Itt); t E (-00, co)} with fa and It relatively
primej that is, their only common divisor is 1.

1.2. Is the sum of two continuous time periodic signals having different
periods itself periodic? Is your conclusion the same for discrete time
signals? Here and hereafter questions requiring a yes/no answer also
require a justification of the answer, e.g., a proof for a positive answer
or a counter example for a negative answer.

1.3. Suppose that 9 = {g(t)j t E 'R.} is an arbitrary signal. Prove that the
signal
L
00

h(t) = g(t - nT)

n=-oo
is periodic with period T. Sketch h for the case where

g(t) = { 1 -
° ¥l It I ~
else
f
1.10. PROBLEMS 49

and
It I ~ T
else.
1.4. Prove the basic geometric progression formulas:

L
N-l N
rn = __
l-r _
(1.37)
l-r
n=O

If Irl < 1,

Evaluate these sums for the case r = !.

1.5. Evaluate the following sums using the geometric progression formulas
and Euler's relations:

(a) 2::=0 cosnw

(b) 2::=0 sin nw
( C) ""N
L.m=O r
n cos nw
(d) ""N
L....n=O r n'
sm nw
1.6. Evaluate the sum

Hint: What happens if you differentiate the geometric progression

formula of Eq. 1.37 with respect to r? Use a similar trick to relate
the integrals JoT te- 1t dt and JOT e- 1t dt. Verify your answer by the
usual method of integration by parts.
1.7. The following input/output relations describe infinite duration dis-
crete time systems with input v and output w. Are the systems
linear? time invariant? (Justify your answers!)

(a) Wn = Vn - Vn-l'

(b) Wn = sgn(vn ).
(c) Wn = r Vn , Irl < 1.
(d) Wn = 2:~=-oo Vk·
(e) Wn = aV n + b, a, b real constants.
50 CHAPTER 1. SIGNALS AND SYSTEMS

1.8. The following systems describe infinite duration continuous time sys-
tems with input v(t) and output w(t). Are the systems linear? time
invariant? (Justify your answers!)

(a) w(t) = v(t) cos(27rfot + ¢), where fo and ¢ are constants.

(b) w(t) = A cos(27rfot+ ~v(t» where ~ is a constant (this is phase
modulation).
(c) w(t) = f~oo v(r) dr.
(d) w(t) = f~oo v(r)e- i2 11"tT dr.

(e) w(t) = -itv(t).

(f) w(t) = av(t) + b (a and b are constants).
(g) w(t) = t 3 v(t).
(h) w(t) = v(t - 7r).
(i) w(t) = v( -t).
(j) w(t) = n(v(t».

1.9. The operations of sampling, windowing, and extending can all be

viewed as systems. Are these systems linear? (Explain for each oper-
ation.)

1.10. Define the infinite duration continuous time signal g(t) = e- t u_l(t)
for all t E 'R.

(a) Sketch the discrete time signal {gn = g(n)j n E Z} formed by

sampling.
(b) Let p be the pulse defined by (1.25) with T = 1 and sketch the
PAM signal fI defined by

g(t) =L 9np(t - n).

nE2

(c) Evaluate the time-average magnitude error between 9 and the

i:
approximation fI defined by

Ig(t) - fI(t)1 dt.

1.10. PROBLEMS 51

1.11. Define a continuous time finite duration waveform 9 = {g(t)j t E

[-1,1) by
g(t) = {+1 t E [0,1) .
-1 tE[-I,O)
(a) Express 9 as a sum of box functions. Is the expression valid for
all t?
(b) Sketch the zero filled extension {Uj t E R} of g.
(c) Sketch a periodic extension {Uj t E R} having period 2 of g.
(d) Write an expression for 9 in terms of 9 using the modular nota-
tion.
(e) Write an expression for jj in terms of U as an infinite sum.
(f) Sketch a periodic extension {g'j t E R} having period 4 of g.
1.12. Prove that the Poisson pmf sums to 1. Hint: You may find the Taylor
series expansion of an exponential useful.
1.13. Prove that the Gaussian pdf is indeed a pdf, i.e., that it is nonnegative
and integrates to 1. (Do the integral, do not just quote a table and
change variables.)
1.14. Express the uniform pdf in terms of the box function and in terms of
unit step functions.
1.15. Define the circle function
x 2 + y2 <_ 1
circ(x, y) = { 0I otherwise
(1.38)

for all real x, y. Provide a labeled sketch of the mesh and image forms
of the signal

g(x,y) = circ(~, ~O)j x E R,y E R.

1.16. Define a 2-dimensional signal 9

g(x,y) = {01 Ixl + Iyl ~ 1

else
for all real x and y. Provide a labeled sketch of the mesh and image
forms of the signal g. Repeat for the signal h defined by
x-2 y+l
h(x,y) = g(-2-' -2-)' x E R,y E R.
Chapter 2

The Fourier Transform

2.1 Basic Definitions

The definition of a Fourier transform will differ depending on the signal
type. The definitions all have a common form, however, and all can be
thought of as a means of mapping a signal 9 = {g(t); t E T}, which de-
pends on a parameter t in some time index set T, into another signal
G = {GU); I E S}, which depends on a new parameter or independent
variable I, which we shall call frequency. As does t, the independent vari-
able I takes values in a domain of definition or index set, denoted S. We
will eventually show that there are natural frequency domains of definitions
for use with each of the basic signal types, as summarized in Table 2.l.
There are two basic forms of the Fourier transform and several special
cases. Unless specifically stated otherwise, we assume one-dimensional in-
dex sets; that is, we assume that T is a subset of the real line R. The
basic form of the transform depends on whether the index set T is con-
tinuous (e.g., the real line itself) or discrete (e.g., the integers). Given a
signal 9 = {g(t); t E T}, its Fourier transform, Fourier spectrum, or fre-
quency spectrum is defined by G = {GU); IE S}, where S is a set of real
numbers, and
E g(n)e-i21rln Tdiscrete
GU) = :;:I(g) =
{ nET
(2.1)
J g(t)e-i21rltdt T continuous
tET
The frequency parameter I is required to be real, although later we shall
briefly consider generalizations which permit it to be complex (the z trans-
form of discrete time signals and the Laplace transform of continuous time
54 CHAPTER 2. THE FOURIER TRANSFORM

signals). For reasons that will be made clear in the next section, we do not
always need to consider the Fourier transform of a signal to be defined for
all real I; each signal type will have a corresponding domain of definition
for I. There is nothing magic about the sign of the exponential, but the
choice of the negative sign is the most common for the Fourier transform.
The inverse Fourier transform will later be seen to take similar form except
that the sign of the exponent will be reversed.
We sometimes refer to the original signal 9 as a time domain signal and
the second signal G a frequency domain signal or spectrum. We will denote
the general mapping by F; that is,

G = F(g). (2.2)
When we wish to emphasize the value of the transform for a particular
frequency I we write

Another popular notation is

{g(t); t E T} :::> {G(f); IE S} (2.3)

or, more simply,
9 :::>G.
When the index sets are clear from context, (2.3) is commonly written
g(t) :::> G(f)
to denote that G(f), considered as a function of the independent variable
I, is the Fourier transform of g(t), which is considered as a function of the
independent variable t.
The Fourier transform is a mapping or transformation of a signal, in
the time or space domain, into its spectrum or transform, in the frequency
domain. If we do not distinguish between the independent variables as being
associated with time and frequency, then the Fourier transform operation
can be considered to be a system.
In this section we consider the details of the definition of the Fourier
transform for the signal types introduced. The remainder of the book is
devoted to developing the properties and applications of such transforms.
Some of the basic questions to be explored include:
• How do we compute transforms either analytically or numerically?
• Does the transform exist; i.e. does the sum or integral defining the
transform exist? Under what conditions is existence of the transform
guaranteed?
2.1. BASIC DEFINITIONS 55

• Have we lost information by taking the transform; that is, can the
original signal be recovered from its spectrum? Is the Fourier trans-
form invertible?
• What are the basic properties of the mapping, e.g., linearity and
symmetry?
• What happens to the spectrum if we do something to the original
signal such as scale it, shift it, filter it, scale its argument, or modulate
it? By filtering we include, for example, integrating or differentiating
continuous time signals and summing or differencing discrete time
signals.
• Suppose that we are given two signals and their transforms. If we
combine the signals to form a new signal, e.g., using addition, mul-
tiplication, or convolution, how does the transform of the new signal
relate to those of the old signals?
• What happens to the spectrum if we change the signal type, e.g.,
sample a continuous signal or reconstruct a continuous signal from a
discrete one?
Before specializing the basic definitions to the most important cases,
it is useful to make several observations regarding the definitions and the
quantities involved. The basic definitions require that the sum or integral
exists, e.g., the limits defining the Riemann integrals converge. If the sum
or integral exists, we say the Fourier transform exists. To distinguish the
two cases of discrete and continuous T we often speak of the sum form as a
discrete time (or parameter) Fourier transform or DTFT, and the integral
form as the continuous time (or parameter) Fourier transform or CTFT
or integral Fourier transform. Note that even if the original signal is real,
its transform is in general complex valued because of the multiplicative
complex exponential e- i21r It.
The dimensions of the frequency variable f are inverse to those of t.
Thus if t has seconds as units, f has cycles per second or hertz as units.
Often frequency is measured as w = 27T f with radians per second as units. If
t has meters as units, f has cycles/meter as units. If t uses the dimensionless
spatial units of distance/wavelength, then f has cycles as units. The symbol
w/21r is also commonly used as the frequency variable, the units of w being
radians per second (or radians per meter, etc.).
One fundamental difference between the discrete and continuous time
cases follows from the fact that the exponential e-i21rln is a periodic func-
tion in f with period one for every fixed integer nj that is,
e- i21r (f+1)n = e-i21rlne-i27fn = e-i27f/n, all fEn.
56 CHAPTER 2. THE FOURIER TRANSFORM

This means that if we consider a DTFT G(f) to be defined for all real f,
it is a periodic function with period 1 (since sums of periodic functions of
a common period are also periodic). Thus
G(f + 1) = G(f) (2.4)
for the DTFT discrete time signals. G(f) does not exhibit this behavior in
the CTFT casej that is, e- i27r It is not periodic in f with a fixed period for
all values of t E T when T is continuous. The periodicity of the spectrum
in the discrete time case means that we can restrict consideration of the
spectrum to only a single period of its argument when performing our
analysis.
In addition to distinguishing the Fourier transforms of discrete time
signals and continuous time signals, the transforms will exhibit different
behavior depending on whether or not the index set T is finite or infinite,
that is whether or not the signal has finite duration or infinite duration. For
example, if 9 has a finite duration index set T = ZN = {O, 1, ... , N - I},
then
= E g(n)e-i27r/n.
N-l
GU) (2.5)
n=O
To define a Fourier transform completely we need to specify the domain
of definition of the frequency variable f. While the transforms appear to
be defined for all real f, in many cases only a subset of real frequencies will
be needed in order to recover the original signal and have a useful theory.
We have already seen, for example, that all the information in the spec-
trum of a discrete time signal can be found in a single period and hence if
T = Z, we could take the frequency domain to be S = [0,1) or [-1/2,1/2),
for example, since knowing GU) for f E [0,1) gives us G(f) for all real
f by taking the periodic extension of G (f) of period 1. We introduce the
appropriate frequency domains at this point so as to complete the defi-
nitions of the Fourier transforms and to permit a more detailed solution
of the examples. The reasons for these choices, however, will not become
clear until the next chapter. The four basic types of Fourier transform are
presented together with their most common choice of frequency domain of
definition in Table 2.1. When evaluating Fourier transforms it will often
be convenient first to find the functional form for arbitrary real f and then
to specialize to the appropriate set of frequencies for the given signal type.
This is particularly true when we may be considering differing signal types
having a common functional form.
Common alternatives are to use a two-sided finite duration DTFT

E
N
G(f) = gne-i27r/nj
n=-N
2.1. BASIC DEFINITIONS 57

Duration Time
Discrete Continuous

Finite Finite Duration DTFT (DFT) Finite Duration CTFT

N-l T
GU) = E gne-i21rln; G(f) = Jg(t)e-i21rlt dt;
n=O 0
IE {a,l/N,· ",(N - l)/N} IE {k/T; k E Z}

Infinite Infinite Duration DTFT Infinite Duration CTFT

00
J g(t)e- i21r It dt;
00
GU) = n=-oo
E gne-i21rln; GU) =
-00

IE [-1/2,1/2) IE (-00,00)

Table 2.1: Fourier Transform Types

N 1 1 N
f E {-2N + l' 2N + l' a, 2N + 1' , 2N + I}'

and a two-sided finite duration CTFT

G(I) = Jf
g(t)e-i21rlt dt; f E {k/T; k E Z}.
,...
-T

It is also common to replace the frequency domain for the infinite duration
DTFT by [a, 1). There is some arbitrariness in these choices, but as we shall
see the key point is to use a frequency domain which suffices to invert the
transform. The reader is likely to encounter an alternative notation for the
DTFT. Many books that treat the DTFT as a variation on the z transform
(which we will consider later) write G(e i21f /) instead of the simpler G(f).
A discrete time finite duration Fourier transform or finite duration
DTFT defined for the frequencies {a, l/N, ... , (N - l)/N} is also called a
discrete Fourier transform or DFT because of the discrete nature of both
the time domain and frequency domain. It is common to express the trans-
form as G(k) instead of G(k/N) in order to simplify the notation, but one
should keep in mind that the frequency is the normalized k/N and not the
integer k.
The DFT can be expressed in the form of vectors and matrices: Given a
signal 9 = {gn; n = 0,1, ... ,N -I}, suppose that we consider it as a column
58 CHAPTER 2. THE FOURIER TRANSFORM

vector g = (Yo, 91, ... , YN _d t , where the superscript denotes the transpose
of the vector (which makes the row vector written in line with the text a
column vector). We will occasionally use boldface notation for vectors when
we wish to emphasize that we are considering them as column vectors and
we are doing elementary linear algebra using vectors and matrices. Similarly
let G denote the DFT vector (G(O),G(l/N),···,G«N -l)/N))t. Lastly,
define the N x N square matrix W by
-2" k-
W={e- tiV Jik=O,1,···,N-1ij=O,1,···,N-l}. (2.6)

Then the DFT can be written simply as

G=Wg=
1 1 1 1
1 e-i~ e-i:jf e-i7f(N -1)
1 e-ij,J e-i~ e-i"; (N-l)
x

e- i7f (N-l)(N-l)

(2.7)

demonstrating the fact that the DFT is expressible as a matrix multiplica-

tion and hence is clearly a linear operation.
When considering infinite duration DTFTs with infinite sums and CTFTs
with integrals, the Fourier transforms may not be well-defined if the limits
defining the infinite sums or Riemann integrals do not exist. There is no
such problem, however, with the DFT provided the signal can take on only
finite values. Before considering some of the technicalities that can arise
and providing some sufficient conditions for the existence of Fourier trans-
forms, we consider examples of evaluation of the various types of Fourier
transforms. In the examples we adopt a common viewpoint: First try to
evaluate the sums or integrals. If the integrals or sums can be successfully
evaluated, then the Fourier transform exists. If not, then it may be nec-
essary to generalize the definition of the Fourier transform, e.g., by taking
suitable limits or by generalizing the definition of an integral or sum. We
begin with simple examples wherein the calculus of evaluating the sums
or integrals is straightforward. We later consider numerical techniques for
finding Fourier transforms and return to the more general issue of when the
definitions make sense.
2.2. SIMPLE EXAMPLES 59

During much of the book we will attempt to avoid actually doing integra-
tion or summation to find transforms, especially when the calculus strongly
resembles something already done. Instead the properties of transforms will
be combined with an accumulated collection of simple transforms to obtain
new, more complicated transforms. The simple examples to be treated can
be considered as a "bootstrap" for this approach; a modicum of calculus
now will enable us to take many shortcuts later.

2.2 Simple Examples

The simplest possible example is trivial: the all zero signal 9 = {g(t); t E
T} defined by g(t) = 0 for all t E r. In this case obviously G(f) = 0 for
all real f, regardless of the choice of the frequency domain.
The simplest nontrivial discrete time signal is the Kronecker delta or
unit pulse (or unit sample). If we consider the infinite duration signal
9 = {8 n ; n E Z}, then it is trivial calculus to conclude that for any real f

=E
00

GU) c5 n e- i27f / n =1
n=-oo

and hence we have the Fourier transform

(2.8)

where we have restricted the frequency domain to Sg~lD' If instead we

consider the discrete time finite duration signal {8 n i n = O,I, .. ·,N -I},
then the transform relation becomes
k
{c5 n ; n = 0, 1, ... , N - I} ::J {I; f = N for k = 0, 1, ... ,N - I}. (2.9)

In both cases the conclusion is the same, the transform of a Kronecker

delta is a constant for all suitable values of frequency. The difference in
the Fourier transforms of the two signals representing two types is only in
the choice of the frequency domain. As mentioned before, this difference in
the definition of the frequency domain will be explained when the inversion
of the Fourier transform is developed. Essentially, the frequency domain
consists of only those frequencies necessary for the reconstruction of the
original signal, and that will differ depending on the signal type.
A simple variation on a Kronecker delta at the origin is a Kronecker
delta at time I, that is, the shifted Kronecker delta function

c5n -1
=l
= { 0I neIse. (2.10)
60 CHAPTER 2. THE FOURIER TRANSFORM

This example is only slightly less trivial and yields

L
00

G(f) = On_ke-i21fln = e-i21Tlk,

n=-(X)

a complex exponential. This is a special case of a general result that will

be seen in Chapter 4: shifting a signal causes the Fourier transform to be
multiplied by a complex exponential with frequency proportional to the
time shift. We therefore have the new transform

{On-I; n E Z} :J {e-i21TII; J E [-~, ~)}. (2.11)

In a similar manner we can show that for the DFT and the cyclic shift that
for any l E ZN

{On-Ii n=0,1,"',N-1}:J{e-i2111Ii J= ~ fork=0,1, .. ·,N-1},

(2.12)
where as usual for this domain the shift is the cyclic shift.
The delta function examples provide an interpretation of a Fourier trans-
form for discrete time signals. A single shifted delta function has as a
Fourier transform a single complex exponential with a frequency propor-
tional to the shift. Any discrete time signal can be expressed as a linear
combination of scaled and shifted delta functions since we can write

gn =L aIOn-l, (2.13)
I

where al= gl. Since summations are linear, taking the Fourier transform
of the signal 9 thus amounts to taking a Fourier transform of a sum of
scaled and shifted delta functions, which yields a sum of scaled complex
exponentials. Each sample of the input signal yields a single scaled complex
exponential in the Fourier domain, so the entire signal yields a weighted
sum of complex exponentials.
Consider next the infinite duration continuous time signal

g(t) = {e- At t 2: 0
o otherwise
which can be written more compactly as
g(t) = e-Atu_l (t); tEn,

where). > O. From elementary calculus we have that

G(f) = roo e-(A+i21ff) t dt = ~ (2.14)
io ). + z27rJ
2.2. SIMPLE EXAMPLES 61

for f E 'R.. Thus the transform relation is

(2.15)

The magnitude and phase of this transform are depicted in Figure 2.1. (The
units of phase are radians here.) In this example the Fourier transform

2r----,----,----,----,----,----,----,----,

1.5 -.~.":'.~..,:-.: ..,. ..... ---......... -:.:.;:.:.~~.~~

,,
\

0.5

s
e"
0

-0.5 ,
:\

-1 ..... ,'
. .'
. :, ..\ ...
\

-1.5
"' " ---
~~---L----~---L----~--~----~--~----~
-4 -3 -2 -1 o 2 3 4
f

Figure 2.1: Fourier Transform of Exponential Signal: solid line = magni-

tude, dashed line = phase

exhibits strong symmetries:

• The magnitude of the transform is symmetric about the origin, i.e.,

IG( - 1)1 = IGU)lj all f E S. (2.16)

• The phase of the transform is antisymmetric about the origin, i.e.,

LG( - 1) = -LG(f)j all f E S. (2.17)

In Chapter 4 symmetry properties such as these will be considered in detail.

Next consider an analogous discrete time signal, the finite duration ge-
ometric signal of (1.5). For any f we have, using the geometric progression
summation formula, that the Fourier transform (here the finite duration
62 CHAPTER 2. THE FOURIER TRANSFORM

DTFT) is given by

(2.18)

There is a potential problem with the above evaluation if the possibility

re- i21r ! = 1 arises, which it can if r = 1 and f is an integer. In this case the
solution is to evaluate the definition directly rather than attempt to apply
the geometric progression formula:
N-I
G(f) =L In =N if re- i21r ! = 1. (2.19)
n=O

Alternatively, L'H6pital's rule of calculus can be used to find the result from
the geometric progression. Applying the appropriate frequency domain of
definition SDTFD we have found the Fourier transform (here the DFT):

{rn;n = 0, 1, ... ,N - I} = {1,r, ... ,r N - I } :) (2.20)

ll. k 1- r N
{G k =G(N)= l_re-i21rk/N; k=O,l,···,N-l},

where the answer holds provided we exclude the possibility of r = 1. The

subscripted notation for the DFT is often convenient and in common use, it
replaces the frequency by the index of the frequency component and saves
a little writing. The magnitude and phase of the DFT are plotted as o's
and *'s, respectively, in Figure 2.2. Observe the even symmetry of the
magnitude and the odd symmetry of the phase about the central frequency
.5.
It is common in the digital signal processing literature to interchange
the left and right hand parts of the figure. This simply takes advantage of
the periodicity of the DFT in f to replace the original frequencies [.5,1) by
[-.5,0) and hence replace the frequency domain [0,1) by [-.5,.5}. Most
software packages for computing Fourier transforms have a command to
perform this frequency shift. In Matlab it is fftshift, in Image it is Block-
shift. The exchange produces the alternative picture of Figure 2.3. The
advantage of this form is that low frequencies are grouped together and
the symmetry of the figure is more evident. In particular, all frequencies
"near" 0 or dc are adjacent.
For this simple signal for which a closed form formula for the general
definition of G (k / N) could be found, the previous figure could have been
generated simply by plugging in the 32 possible values of frequency into the
formula and plotting the magnitude and phase. If we had not had such a
2.2. SIMPLE EXAMPLES 63

10

~ 4

---
-Ie!
'-"
C!) 2

-2
0 0.1

kiN

Figure 2.2: DFT of Finite Duration Geometric Signal: o=magnitude,

*=phase

10

6 ..............

4
: 0 0
I :

0 0

. ...
2 ,......................... 1............. -

...
0
0 0
• v

..... -
0
; " o 0
o 0
o 0 ~ ~ o 0 0 o 0 0

o ;....; ...

, ,
.... , 0 or' 0

-2
-o.s -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 O.S

kiN

Figure 2.3: DFT of Finite Duration Geometric Signal with shifted frequency
domain: o=magnitude, *=phase
64 CHAPTER 2. THE FOURIER TRANSFORM

useful formula, however, we would have in principle had to do a brute force

calculation of the definition

(2.21)

for k = 0,1"",31. To produce the plot we would have had to perform

roughly 32 multiply/adds for each frequency and hence 32 2 = 1024 multiply
adds to perform the entire transform. This would be the case, for example,
if we wished to compute the Fourier transform of the randomly generated
signal of Figure 1.18 or the sum of the sine and random noise of Figure 1.20.
This would be too much computation if done with a pocket calculator,
but it would be easy enough if programmed for a computer. It is easy
to see, however, that even computers can get strained if we attempt such
brute force computation for very large numbers of points, e.g., if N were
several thousand (which is not unreasonable for many applications). In fact
Figure 2.2 was produced using a program called a "fast Fourier transform"
or FFT which uses the structure of the definition of the DFT to take many
shortcuts in order to compute the values of (2.21) for all values of k of
interest. We shall later describe in some detail how the FFT computes the
DFT of an N-point signal in roughly N logN multiply/adds instead of the
N 2 multiply/adds needed to calculate the DFT by brute force. As examples
of the DFT for less structured signals, the FFT is used to compute the DFT
of the signals of Figure 1.18 and Figure 1.20 and the results are displayed in
Figures 2.4 and 2.5, respectively. Note how the DFT of the purely random
signal itself appears to be quite random. In fact, a standard exercise of
probability theory could be used to find the probability density functions of
the magnitude and phase of the DFT of a sequence of independent Gaussian
variables. The sine plus noise, however, exhibits important structure that
is not evident in the signal itself: There are clearly two components in the
magnitude of the DFT that stand out above the largely random behavior.
To see what these components are, consider the DFT of the sinusoidal term
alone, that is, the DFT of the signal depicted in Figure 1.3. This is shown
in Figure 2.6. The DFT of the sinusoid alone shows the same two peaks
in the magnitude at frequencies of 1/8 and 7/8 along with some smaller
frequency components. The point is that these two peaks are still visible
in the DFT of the signal plus noise, although the sinusoidal nature of the
time signal is less visible. This example provides a simple demonstration of
the ability of a Fourier transform to extract structure from a signal where
such structure is not apparent in the time domain. The two peaks in the
signal plus noise at 1/8 and 7/8 suggest that the noisy signal contains a
discrete time sinusoid of period 8.
2.2. SIMPLE EXAMPLES 65

10

-
4

~
-'<! 2 ...
......-
\!)
0 .. 0 ...

-2
.
-4
0 0.5 0.6 0.7 0.8 0.9

kiN

Figure 2.4: DFT of a Random Signal: o=magnitude, *=phase

10

-
4

~ o 0

-'<! 2 ..
......-
\!)
0

-2

-4
0

kiN

Figure 2.5: DFT of the Sum of a Random Signal and Sinusoid:

o=magnitude, *=phase
66 CHAPTER 2. THE FOURIER TRANSFORM

14

12

10

2 .• .... ".: .
. : -
o 000 ~ o:v: ~ o ... o.. o .. ~ti.~ Ii!. 8 & I{O'O O~.: I) o;U:

-2 ... II' •

_4L-__ ~ __ ~ __L -_ _L -_ _L -_ _L -_ _L -_ _L -_ _ L-~

o 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

kiN

Figure 2.6: DFT of Finite Duration One-Sided Sinuosoid: o=magnitude,

*=phase

Returning to the original example of a geometric signal, if r = 1 and

hence the signal is a constant value of 1, then from (2.18) and (2.19) we
have that
k
{I; n = 0, 1, ... , N - I} ::) {Gk = G( N) = NJ k ; k = 0, 1,···, N - I};
(2.22)
that is, the DFT of a unit constant is proportional to the Kronecker delta.
In fact this is the dual of the previous result that the Kronecker delta time
signal has a constant Fourier transform. We shall later see how all Fourier
transform results provide dual results by reversing the roles of time and
frequency.
If instead of the finite duration signal g, we wished to find the Fourier
transform of the infinite duration signal ii formed by zero filling, i.e., iin = rn
for n = 0,1,·", N - 1 and gn = 0 otherwise, then, assuming Irl i- 1, the
Fourier transform relation would instead be

(2.23)

because of the different frequency domains (here Sl:;f, I D)'

2.2. SIMPLE EXAMPLES 67

Next consider the DTFT of the infinite duration signal

gn -
_{rn0 n = 0, 1, ...
n < 0, (2.24)

where now we require that /r/ < 1. We can write this signal more compactly
using the unit step function as

(2.25)

Observe that the functional form of the time dependence is the same here
as in the previous finite duration example; the difference is that now the
functional form is valid for all integer times rather than just for a finite set.
The DTFT is
=L
00

GU) r n e- i2 1<f n ,
n=O
which is given by the geometric progression formula as
1
G(f) =1- re-''2 1< f'
(2.26)

Note that the spectrum is again complex valued. Restricting frequencies to

SDTID we have the Fourier transform relation

(2.27)

This transform is the discrete time version of (2.15). Note the transforms
do not clearly resemble each other as much as the original signals do.
As another example of a DTFT consider the two-sided discrete time box
function

_ 0 ~ {1 if Inl ::; N
gn - N(n) - 0 n = ±(N + 1), ±(N + 2),·,. (2.28)

Application of the geometric progression formula and some algebra and

trigonometric identities then yield

L
N
G(f) e- i2 1<f n
n=-N
cos(21rfN) - cos(27rf(N + 1»
= 1 - cos(27r f)
sin(27rf(N + !»
= sin( 7r f)
(2.29)
68 CHAPTER 2. THE FOURIER TRANSFORM

Applying the appropriate frequency domain of definition yields the trans-

form
. sin(27rf(N + !)). 1 1
{DN(n), n E Z} :::> { sin(7rf) , f E [-2' 2)}· (2.30)

This spectrum for the case of Figure 1.6 is thus purely real and (N = 5)
is plotted in Figure 2.7. Note the resemblance to the sinc function. In

S 4

c.J
2

-2

-4
-0.5 -0.4 -OJ -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5

Figure 2.7: Transform of {D5(n); n E Z}

fact, this function is often referred to as a "discrete time sinc" function

because of the similarity of form (and because, as we shall see, the analogous
continuous time Fourier transform of a box is indeed a sinc). The Fourier
transform for this example is purely real.
One can also use the box signal to describe a finite duration con-
stant signal {gn = 1; n = -N, ... , N}. The functional form of (2.29)
then simplifies for the frequencies {O, 1/(2N + 1),···, 2N/(2N + I)} to
G(kjN) = (2N + I)J(k). This provides a two-sided version of (2.22).
As an example of the finite duration CTFT, recall the continuous time
ramp function of Figure 1.41. The solution to the integral defining the
CTFT for all real frequencies is given by
if f =0 (2.31)
if f f.: 0
2.2. SIMPLE EXAMPLES 69

°
Note that the calculus would yield G(J) if we considered g to be an infinite
duration signal for which g(t) = for t not in the interval [0,1). Restricting
the result to the frequency domain of definition for finite duration signals
yields the transform

l k=O
{g(t)j t E (0, I)} :::> {G(k) = { ~
21rk
k Z k -I-
E,-r
oJ· (2.32)

If we consider instead the zero filled extension 9 defined by g(t) = t for

t E [0,1) and g(t) = 0 for t =/: [0,1), then the transform becomes

if f = 0
{g(t)j t E R} J {{ t-;2~1 e-;2~/-l if fER, f =/: 0.
}
(2.33)
~ + (21rj) 2 .

As another simple example of a CTFT, consider the continuous time

analog of the box of the previous DTFT example. This example can be
considered finite or infinite duration.
Recall the definition of the rectangle function n(t):

1 if ItI < 1
n(t) ={ ~ if t ±~
= (2.34)
o otherwise

For the signal g(t) = n(t) we have that

G(J) = J
00

n(t)e-i21rltdt = J!
e-i21rltdt
-00 -l
1
e-i21rltI2 _sin(7rf)~.
= -t
'2 7r f _1
- 7r f - smc(J), (2.35)
2

where we have recalled the definition of the sinc function. If we consider

the infinite duration case, the Fourier transform is

{n(t)j t E R} J {sinc(f)j fER}. (2.36)

The sinc function was shown in Figure 1.10. Although that was sinc(t) and
this is sinc(J), obviously the name of the independent variable has nothing
to do with the shape of the function.
Riemann integrals are not affected by changing the integrand by a finite
amount at a finite number of values of its argument. As a result, the Fourier
transforms of the two box functions D 1/ 2 (t) and n(t), which differ in value
70 CHAPTER 2. THE FOURIER TRANSFORM

only at the two endpoints, are the same; hence, we cannot count on inverting
the Fourier transform in an unambiguous fashion. This fact is important
and merits emphasis:

Different signals may have the same Fourier transform and hence
the Fourier transform may not have a unique inverse. As in the
box example, however, it turns out that two signals with the same
Fourier transform must be the same for most values of their argu-
ment ("almost everywhere," to be precise).

More generally, given T > 0 the Fourier transform of DT(t) (and of

n(t/2T)) is given by

G(f) f00

DT(t)e-i21Tftdt = fT
e-i21Tftdt
-00 -T

-i21TftIT . (2 IT)
~2I = sm; = 2Tsinc(2Tf), (2.37)
-z 7l" -T 7l"

and hence
{DT(t); t E R} :) {2Tsinc(2Tf); IE R}. (2.38)
This is another example of a real-valued spectrum. Note the different forms
of the spectra of the discrete time box function of (2.29) and the continuous
time transform of (2.37). We also remark in passing that the spectrum of
(2.37) has the interesting property that its samples at frequencies of the
form k/2T are 0:
k
G(2T) =0; k=±1,±2,···, (2.39)
as can be seen in Figure 1.10. Thus the zeros of the sinc function are
uniformly spaced.
We have completed the evaluation, both analytically and numerically,
of a variety of Fourier transforms. Many of the transforms just developed
will be seen repeatedly throughout the book. The reader is again referred
to the Appendix where these and other Fourier transform relations are
summarized.

2.3 Cousins of the Fourier Transform

The Cosine and Sine Transforms
Observe that using Euler's formulas we can rewrite the CTFT as
2.3. COUSINS OF THE FOURIER TRANSFORM 71

= 1 g(t)[cos( -271Jt) + i sine -27T /t)]dt

1
tET

= g(t)COS(27T't)dt-il g(t)sin(27T/t)dt
tET tET
= C,(g) - is, (g), (2.40)

where c, (g) is called the cosine trons/orm of g( t) and S, (g) is called the sine
trons/orm. (There often may be a factor of 2 included in the definitions.)
Observe that if the signal is real-valued, then the Fourier transform can be
found by evaluating two real integrals. The cosine transform is particularly
important in image processing where a variation of the two-dimensional
discrete time cosine transform is called the discrete cosine transform or
DCT in analogy to the DFT. Its properties and computation are studied
in detail in Rao and Yip [27].

The Hartley Transform

If instead of combining the sine and cosine transforms as above to form the
Fourier transform, one changes -i to 1 and defines

1/.1 (g) = f g(t)[COS(27T/t) + sin(271Jt)] dt

itEr
= C,(g) + S,(g) (2.41)

then the resulting transform is called the Hartley trons/orm and many of its
properties and applications strongly resemble those of the Fourier transform
since the Hartley transform can be considered as a simple variation on the
Fourier transform.
An alternative way to express the Hartley transform is by defining the
cas function
cas(x) ~ cos(x) + sin (x) (2.42)

1
and then write
1/.,(g) = get) Cas(27T/t) dt. (2.43)
tET

The Hartley transform has the advantage that it is real-valued for a

real-valued signal, but it lacks some of the important theoretical properties
that we shall find for Fourier transforms. In particular, it does not prove
a decomposition of compex signals into a linear combination of eigenfunc-
tions of linear time invariant systems. A thorough treatment of Hartley
transforms can be found in R. Bracewell's The Hartley Transform [7].
72 CHAPTER 2. THE FOURIER TRANSFORM

z and Laplace Transforms

The Fourier transform is an exponential transform; that is, it is found
by integrating or summing a signal multiplied by a complex exponential
involving a frequency variable. So far only real values of frequency have
been permitted, but an apparently more general class of transforms can be
obtained if the frequency variable is allowed to be complex. The resulting
generalizations provide transforms that will exist for some signals which
do not have Fourier transforms and the generalized transforms are useful
in some applications, especially those emphasizing one-sided signals and
systems with initial conditions. The added generality comes at a cost,
however; the generalized transforms do not exist for all values of their
arguments (z for a z-transform and s for a Laplace transform) and as a
result they can be much harder to invert than Fourier transforms and their
properties can be more complicated to describe. Most problems solvable by
the general transforms can also be solved using ordinary Fourier transforms.
For completeness, however, we describe the two most important examples
for engineering applications, the z transform for discrete time signals and
the Laplace transform for continuous time signals, so that their connection
to the Fourier transform can be seen.
Given a discrete time signal 9 = {gn; nET}, its z transform is defined
by
Gz(z) = L gnz-n, (2.44)
nET

provided that the sum exists, i.e., that the sum converges to something fi-
nite. The region of z in the complex plane where the sum converges is called
the region of convergence or ROC. When the sum is taken over a two-sided
index set T such as the set Z of all integers or {-N, ... , -1,0,1, ... ,N},
the transform is said to be two-sided or bilateral. If it is taken over the
set of nonnegative integers or a set of the form 0,1"", N, it is said to be
one-sided or unilateral. Both unilateral and bilateral transforms have their
uses and their properties tend to be similar, but there are occasionally dif-
ferences in details. We shall focus on the bilateral transform as it is usually
the simplest. If G(f) is the Fourier transform of gn, then formally we have
that

that is, the Fourier transform is just the z transform evaluated at e i27f f. For
this reason texts treating the z-transform as the primary transform often
use the z-transform notation G(e i27ff ) to denote the DTFT, but we prefer
the simpler Fourier notation of G(f). If we restrict f to be real, then the
Fourier transform is just the z-transform evaluated on the unit circle in the
2.3. COUSINS OF THE FOURIER TRANSFORM 73

z plane. If, however, we permit I to be complex, then the two are equally
general and are simply notational variants of one another.
Why let I be complex or, equivalently, let z be an v.rbitrary complex
number? Provided Izl ::f:. 0, we can write z in magnitude-phase notation as
z = r- 1 ei9 . Then the z transform becomes
Gz(z) = 2: (9nrn)e-inO.
nET

This can be interpreted as a Fourier transform of the new signal 9nrn and
this transform might exist even if that of the original signal 9n does not since
the rn can serve as a damping factor. Put another way, a Fourier transform
was said to exist if G(f) m-,de sense for all I, but the z transform does
not need to exist for all z to be useful, only within its ROC. In fact, the
existence of the Fourier transform of an infinite duration signal is equivalent
to the ROC of its z-transform containing the unit circle {z: Izi = I}, which
is the region of all z = e- i21f ! for real I.
As an example, consider the signal 9n = u_l(n) for n E Z. Then the
ordinary Fourier transform of 9 does not exist for all I, e.g., it blows up for
I = O. Choosing Irl < 1, however, yields a modified signal 9nrn which, as
we have seen, has a transform. In summary, the z transform will exist for
some region of possible values of z even though the Fourier transform may
not. The two theories, however, are obviously intimately related.
The Laplace transform plays the same role for continuous time wave-
forms. The Laplace transform of a continuous time signal 9 is defined by

Gds) = 1
tET
g(t)e- st dt. (2.45)

As for the z transform, for the case where T is the real line, one can define a
bilateral and a unilateral transform, the latter being the bilateral transform
of g(t)U_l(t). As before, we focus on the bilateral case.
If we replace i21r I in the Fourier transform by s we get the Laplace
transform, although the Laplace transform is more general since s can be
complex instead of purely imaginary. Letting the I in a Fourier transform
take on complex values is equivalent in generality. Once again, the Laplace
transform can exist more generally since, with proper choice of s, the orig-
inal signal is modified in an exponentially decreasing fashion before taking
a Fourier transform.
The primary advantage of Laplace and z-transforms over Fourier trans-
forms in engineering applications is that the one-sided transforms provide a
natural means of incorporating initial conditions into linear systems analy-
sis. Even in such applications, however, two-sided infinite duration Fourier
74 CHAPTER 2. THE FOURIER TRANSFORM

transforms can be used if the initial conditions are incorporated using delta
functions.
It is natural to inquire why these two variations of the Fourier transform
with the same general goal are accomplished in somewhat different ways.
In fact, one could define a Laplace transform for discrete time signals by
replacing the integral by a sum and one could define a z transform for con-
tinuous time signals by replacing the sum by an integral. In fact, the latter
transform is called a Mellin transform and it is used in the mathematical
theory of Dirichlet series. The differences of notation and approach for
what are clearly closely related transforms is attributable to history; they
arose in different fields and were developed independently.

2.4 Multidimensional Transforms

The basic definition of a Fourier transform extends to two-dimensional and
higher dimensional transforms. Once again the transform is different de-
pending on whether the index set T is discrete or continuous. As usual a
signal is defined as a function 9 = {g(t); t E T}. In the multidimensional
case T has dimension K. The case K = 2 is the most common because of
its importance in image processing, but three-dimensional transforms are
useful in X-ray diffraction problems and higher dimensions are commonly
used in probabilistic systems analysis when characterizing multidimensional
probability distributions. To ease the distinction of the multidimensional
case from the one-dimensional case, in this section we use boldface for the
index to emphasize that it is a column vector, i.e., t = (tl, t2, ... , tK)t and
the signal is now 9 = {g(t); t E T}. When transforming a signal with a
K -dimensional parameter we also require that the frequency variable be a
K-dimensional parameter, e.g., f= (h,h, ... ,fK)t.
The Fourier transforms of the signal 9 are defined exactly as in the one-
dimensional case with two exceptions: the one-dimensional product ft in
the exponential e- i2rr It is replaced by an inner product or dot product or
scalar product
K
t .f = ttf = L tklk
k=l

and the one dimensional integral or sum over t is replaced by the K-

dimensional integral or sum over t. The inner product is also denoted
by < t, f >. Thus in the continuous parameter case we have that

G(f) = h g(t)e-i2rrt.f dt (2.46)

2.4. MULTIDIMENSIONAL TRANSFORMS 75

which in the two-dimensional case becomes

G(fx'/y) = JJ g(x,y)e- i27r (z/x+Y/Y)dxdy; ix,fy En. (2.47)

The limits of integration depend on the index set I; they can be finite or
infinite. Likewise the frequency domain is chosen according to the nature
of T-
In the discrete parameter case the transform is the same with integrals
replaced by sums:
G(f) = L g(t)e- i27rt-f (2.48)
tE7"

which in the two dimensional case becomes

G(fx,fy) = LLg(x,y)e- i27r (z!x+y!y). (2.49)

Z y

Multidimensional Fourier transforms can be expressed as a sequence

of ordinary Fourier transforms by capitalizing on the property that the
exponential of a sum is a product of exponentials. This decomposition
can be useful when evaluating Fourier transforms and when deriving their
properties. For example, consider the discrete time 2D Fourier transform
and observe that

G(fx, fy) = L Lg(x, y)e- i27r (Z!x+Y!v)

Z Y
= L e- i27rz !x L g(x, y)e-i27rY/Y
x Y

= Le- i27rX !xG z (fy) (2.50)

x

where
Gx(fy) ~ Lg(x,y)e- i27rY !v (2.51)
Y

is the ordinary one-dimensional Fourier transform with respect to y of the

signal {g(x, y); y E 'Ty}. Thus one can find the 2D transform of a signal
by first performing a collection of 1D transforms G x for every value of x
and then transforming G x with respect to x. This sequential computation
has an interesting interpretation in the special case of a finite-duration 2D
transform of a signal such as {g(k,n)j k E ZN, n E ZN}. Here the signal
can be thought of as a matrix describing a 2D sampled image raster. The
76 CHAPTER 2. THE FOURIER TRANSFORM

20 Fourier transform can then be written as

G(fx,fy) = ~e-i2'11"fXk (~>-i211'fYng(k,n»);

k
jx,fy E {N; k E ZN},

where the term in parentheses can be thought of as the 10 transform of

the kth column of the matrix g. Thus the 20 transform is found by first
transforming the columns of the matrix g to form a new matrix. Taking
the 10 transforms of the rows of this matrix then gives the final transform.
Continuous parameter 20 Fourier transforms can be similarly found as a
sequence of 10 transforms.

Separable 2D Signals
The evaluation of Fourier transforms of 2D signals is much simplified in the
special case of 20 signals that are separable in rectangular coordinates; i.e.,
if
g(x,y) = gx(x)gy(y),
then the computation of 20 Fourier transforms becomes particularly simple.
For example, in the continuous time case

G(fx,jy) = 1:1: gx(x)gy(y)e- i21r (fxX+!YY)dxdy

= ([: gx(x)e- i2'11"/XX dX) (1: gy(y)e- i2 11'/Yll dY)

= Gx(fx)Gy(fy),
the product of two I-D transforms. In effect, separability in the space do-
main implies a corresponding separability in the frequency domain. As a
simple example of a 2D Fourier transform, consider the continuous param-
eter 20 box function of (1.17): g(x,y) = 0T(X)OT(Y) for all real x and y.
The separability of the signal makes its evaluation easy:

G(fx,fy) = / / g(x,y)e- i21r (x/x+1I!Y)dxdy

= (J DT(x)e- i2 11'X!x dX) ( / OT(y)e- i2 11'Yfy dY)

= .1'!x({DT(X)j x E R}).1'fy({DT(X); x E 'R.})

= (2Tsinc(2Tfx»(2Tsinc{2Tfy», (2.52)
2.4. MULTIDIMENSIONAL TRANSFORMS 77

where we have used (2.38) for the individual ID transforms.

A similar simplification results for signals that are separable in polar
coordinates. Consider the 2D disc of (1.22) with a radius of 1: g(x, y) =
0 1 (J x 2 + y2) for all real x, y. In other words, the signal is circularly
symmetric and in polar coordinates we have that g(r, 0) = gR(r). To form
the Fourier transform we convert into polar coordinates
x = r cosO
y = r sinO
dx dy = r dr dO.
We also transform the frequency variables into polar coordinates as
fx = pcos¢
fy = psin¢

Substituting into the Fourier transform:

G(p,¢) = JJ 211'
dO
00

drrgR{r)e-i 211'[rpcos8cosHrpsin8sin4>j
o 0

= J J
00 211'
dr rgR(r) dOe-i 211'rpcos(8-4».
o 0
To simplify this integral we use an identity for the zero-order Bessel
function of the first kind:
Jo{a) = ~ (211' e-iacos(8-4» dO. (2.53)
27r 10
Thus
J
00

G(p, ¢) = 27r rgR(r)Jo(27rrp) dr = G(p).

o
Thus the Fourier transform of a circularly symmetric function is itself
circularly symmetric. The above formula is called the zero-order Hankel
transform or the Fourier-Bessel transform of the signal.
For our special case of gR{r) = 01{r):

G(p) = 27r 1 00
r01 (r)Jo(27rrp) dr

= 27r 11 rJo(27rrp) dr
78 CHAPTER 2. THE FOURIER TRANSFORM

Make the change of variables r' = 27rrp so that r = r' /27rp and dr =

1
dr' /27rp. Then
21rP
G(p) = - 122 r'Jo(r')dr'.
7rp 0
It is a property of Bessel functions that

1 al
(Jo(() d( = xJ1 (x),
where J1 (x) is a first-order Bessel function. Thus

G(p) = J 1(27rp) = 7r [2 J1 (27r P)] .

p 27rp
The bracketed term is sometimes referred to as the Bessinc function and it
is related to the jinc function by a change of variables as G(p) = 4jinc2p.

The Discrete Cosine Transform (DCT)

A variation on the 2D discrete Fourier transform is the discrete cosine
transform or DCT that forms the heart of the international JPEG (Joint
Photographic Experts Group) standard for still picture compression [26].
(The "Joint" emphasizes the fact that the standard was developed by coop-
erative effort of two international standards organizations, ISO and ITU-T,
formerly CCITT.) Consider a finite-domain discrete parameter 2D signal
9 = {g(k,j)j k = 0,1, ... , N -1, j = 0,1, ... , N -I}. Typically the g(k,j)
represent intensity values at picture elements (pixels) in a scanned and
sampled image. Often these values are chosen from a set of 28 = 256 pos-
sible levels of gray. The DCT G = {G(l, m)j 1 = 0,1, ... , N - 1, m =
0,1, ... , N - I} is defined by

G(l , m ) = !C(l)C( ) ~ ~1 (k.) (2k + 1)17r (2j + l)m7r

4 m L...J L...J 9 ,J cos 2N cos 2N '
k=O j=O
(2.54)
where
C(n) = {~
if n = 0
1 otherwise
As with the Hartley transform, the prime advantage of the DCT over the
ordinary DFT is that it is purely real. An apparent additional advantage
is that an image of N 2 samples yields a DCT also having N 2 real values
while an FFT yields twice that number of real valuesj i.e., N 2 magnitudes
and N 2 phases. It will turn out, however, that not all 2N2 values of the
FFT are necessary for reconstruction because of symmetry properties.
2.5. * THE DFT APPROXIMATION TO THE CTFT 79

2.5 * The DFT Approximation to the CTFT

We have now surveyed a variety of basic signal examples with Fourier trans-
forms that can be evaluated using ordinary calculus and algebra. Analytical
tools are primarily useful when the signals are elementary functions or sim-
ple combinations of linear functions. Numerical techniques will be needed
for random signals, physically measured signals, unstructured signals, or
signals which are simply not amenable to closed form integration or sum-
mation.
The discrete Fourier transform (OFT) plays a role in Fourier analysis
that is much broader than the class of discrete time, finite duration signals
for which it is defined. When other signal types such as infinite duration
continuous time signals are being considered, one may still wish to com-
pute Fourier transforms numerically. If the Fourier transforms are to be
computed or manipulated by a digital computer or digital signal processing
(OSP) system, then the signal must be sampled to form a discrete time
signal and only a finite number of samples can be used. Thus the original
continuous time infinite duration signal is approximated by a finite dura-
tion discrete time signal and hence the numerical evaluation of a Fourier
transform is in fact a OFT, regardless of the original signal type! For this
reason it is of interest to explore the complexity of evaluating a OFT and
to find low complexity algorithms for finding OFTs.
Suppose that g = {g(t); t E 'R} is an infinite duration continuous time
signal and we wish to compute approximately its Fourier transform. Since
a digital computer can only use a finite number of samples and computes
integrals by forming Riemann sums, a natural approximation can be found
by taking a large number N = 2M + 1 of samples
{g(-MT),··· ,g(O),··· ,g(MT)}
with a small sampling period T so that the total time windowed NT is

I:
large. We then form a Riemann sum approximation to the integral as

G(f) = g(t)e-i21rft dt

L
M
~ g(nT)e-i21rfnTT; (2.55)
n=-M

that is, we make the Riemann approximation that dt ~ T for T small

enough. This has the general form of a OFT. We can put it into the exact
form of the one-sided OFT emphasized earlier if we define the discrete time
signal
On = genT - MT); n = 0,··· ,N-1.
80 CHAPTER 2. THE FOURIER TRANSFORM

(By construction 90 = g( -MT) and 9N-l = g(MT).) Then the Riemann

sum approximation yields

L
M
G(f) ~ g(nT)e-i21r/nTT
n=-M

L g(nT - MT)e- i21r /(n-M)T

N-l
= T
n=O

L
N-l
= ei21r/MTT 9ne-i21r/nT
n=O
= ei21r/MTTG(fT) , (2.56)

where G is the DFT of g. The usual frequency domain of definition for

the DFT is the set Sg~FD = {a, liN,"', (N - l)IN} which means that
(2.56) provides an approximation to G(f) for values of I for which IT E
{a, liN,,,,, MIN, (M + l)IN,"', (N - l)IN = 2MIN}, that is, for I E
{a, liNT"", (N -l)INT}. It is useful, however, to approximate G(f) not
just for a collection of positive frequencies, but for a collection of positive
and negative frequencies that increasingly fills the real line as T becomes
small and NT becomes large. Recalling that the DFT is periodic in I with
period 1, we can add or subtract 1 to any frequency without affecting the
value of the DFT. Hence we can also take the frequency domain to be
(2) M +1 M N - 1 11M
SDTFD = {~-1 = - N""'(~) -1 = - N'O, N"",N}'
(2.57)
This can be viewed as performing the frequency shift operation on the DFT.
Combining these facts we have the following approximation for the in-
finite duration CTFT based on the DFT:

G(~)
NT
~ ei21rkM/NTG(~)
N
= ei21rkM/NTGk' k = -M ...
" "
°... , M ,
(2.58)
which provides an approximation for a large discrete set of frequencies that
becomes increasingly dense in the real line as T shrinks and NT grows.
The approximation is given in terms of the DFT scaled by a complex ex-
ponential; i.e., a phase term with unit magnitude, and by the sampling
period.
The above argument makes the point that the DFT is useful more gener-
ally than in its obvious environment of discrete time finite duration signals.
It can be used to numerically evaluate the Fourier transform of continuous
time signals by approximating the integrals by Riemann sums.
2.6. THE FAST FOURIER TRANSFORM 81

2.6 The Fast Fourier Transform

We consider first the problem of computing the DFT in the most straight-
forward and least sophisticated manner, paying particular attention to the
number of computational operations required. We then take advantage of
the structure of a DFT to significantly reduce the number of computations
required.
Recall that given a discrete time finite duration signal 9 = {gn; n =
0,1, ... , N - I}, the DFT is defined by

(2.59)

Since 9 is in general complex-valued, computation of a single spectral co-

efficient G(mjN) requires N complex multiplications and N - 1 complex
additions. Note that a complex multiply has the form

(a + ib)(e + id) = (ae - bd) + i(be + ad),

which consists of four real multiplies and two real adds, and a complex
addition has the form

(a + ib) + (e + id) = (a + e) + i(b + d),

which consists of two real adds. As an aside, we mention that it is also
possible to multiply two complex numbers with three real multiplies and
five real additions. To see this, form A = (a + b)(e - d), which takes two
additions and one multiply to give A = ae + be - ad - bd, B = be (one
multiply), and C = bd (one mUltiply). Then the real part of the product is
found as A - B + C = ae - bd (two additions), and the imaginary part is
given by B + C = be + ad (one addition). The total number of operations
is then three multiplies and five additions, as claimed.
This method may be useful when the time required for a real multipli-
cation is greater than three times the time required for a real addition.
Complexity can be measured by counting the number of operations,
but the reader should keep in mind that the true cost or complexity of an
algorithm will depend on the hardware used to implement the arithmetic
operations. Depending on the hardware, a multiplication may be vastly
more complicated than an addition or it may be comparable in complexity.
As a compromise, we will measure complexity by the number of complex-
multiply-and-adds, where each such operation requires four real multiplies
and four real adds. The complexity of straightforward computation of a sin-
gle DFT coefficient is then approximately N complex-multiply-and-adds.
82 CHAPTER 2. THE FOURIER TRANSFORM

Computation of all N DFT coefficients G(O), ... , G«N - l)jN) then re-
quires a total of N 2 complex-multiply-and-adds, or 4N2 real multiplies and
4N 2 real adds.
If k represents the time required for one complex-multiply-and-add, then
the computation time Td required for this "direct" method of computing
a DFT is Td = kN2. An approach requiring computation proportional to
N log2 N instead of N2 was popularized by Cooley and Tukey in 1965 and
dubbed the fast Fourier transform or FFT [13]. The basic idea of the algo-
rithm had in fact been developed by Gauss and considered earlier by other
authors, but Cooley and Tukey are responsible for introducing the algo-
rithm into common use. The reduction from N2 to N log2 N is significant
if N is large. These numbers do not quite translate into proportional com-
putation time because they do not include the non-arithmetic operations
of shuffling data to and from memory.

The Basic Principle of the FFT Algorithm

There are several forms of the FFT algorithm. We shall focus on one, known
as "decimation in time," a name which we now explain. A discrete time
signal 0 = {On; nET} can be used to form a new discrete time signal by
downsampling or subsampling in much the same way that a continuous time
signal could produce a discrete time signal by sampling. Downsampling
means that we form a new process by taking a collection of regularly spaced
samples from the original process. For example, if M is an integer, then
downsampling 9 by M produces a signal g(M) = {g~M); n E T(M)} defined
by
g~M) = gnM (2.60)

for all n such that nM E T. Thus g~M) is formed by taking every Mth
sample of gn' This new signal is called a downsampled version of the original
signal. Downsampling is also called decimation after the Roman army
practice of decimating legions with poor performance (by killing every tenth
soldier). We will try to avoid this latter nomenclature as it leads to silly
statements like "decimating a signal by a factor of 3" which is about as
sensible as saying "halving a loaf into three parts." Furthermore, it is an
incorrect use of the term since the decimated legion referred to the survivors
and hence to the 90% of the soldiers who remained, not to the every tenth
soldier who was killed. Unfortunately, however, the use of the term is so
common that we will need to use it on occasion to relate our discussion to
the existing literature.
We change notation somewhat in this section in order to facilitate the
introduction of several new sequences that arise and in order to avoid re-
2.6. THE FAST FOURIER TRANSFORM 83

peating redundant parameters. Specifically, let g(n)j n = 0,1, ... , N - 1,

denote the signal and G(m)j m = 0,1, ... , N - 1 denote the DFT coeffi-
cients. Thus we have replaced the subscripts by the parenthetical index.
Note that G(m) is not the Fourier transform of the sequence in the strict
sense because m is not a frequency variable, G(m) is the value of the spec-
trum at the frequency miN. In the notation used through most of the
book, G(m) would be G m or G(mIN), but for the current manipulations
it is not worth the bother of constantly writing miN.
The problem now is to compute
N-l N-l
G(m) =L g(n)e- i1fnm =L g(n)w nm , (2.61)
n=O n=O
where W = e-i~ , and m = 0,1, ... ,N - 1. We assume that N is a power
of two - the algorithm is most efficient in this case. Our current example
is for N = 8.
Step 1: Downsample g( n) using a sampling period of 2j that is, construct
two new signals go ( n) and gl (n) by
N
gO(n)
A
g(2n)j n = 0, 1, ... , "2 - 1 (2.62)
N
gl(n)
A
g(2n + l)j n = 0,1, ... , "2 - 1. (2.63)

For an 8-point g(n), the downsampled signals go(n) and gl(n) have four
points as in Figure 2.8.
The direct Fourier transforms of the downsampled signals go and gl,
say Go and G 1 , can be computed as
.if-I .if-I
Go(m) = L go(n)e-i~mn =L g(2n)w2mn (2.64)
n=O n=O
.if-I .if-I
G 1 {m) = L gl{n)e-i~mn =L g{2n+ 1)w 2mn , (2.65)
n=O n=O
where m = 0,1, ... , !f - 1. As we have often done before, we now observe
that the Fourier sums above can be evaluated for all integers m and that the
sums are periodic in m, the period here being N 12. Rather than formally
define the periodic extensions of Go and G 1 and cluttering the notation
further (in the past we put tildes over the function being extended), here
we just consider the above sums to define Go and G 1 for all m and keep in
mind that the functions are periodic.
84 CHAPTER 2. THE FOURIER TRANSFORM

go(n)

g(n) n

n gl(n)
I I
~ I
I
I I

I I I
0 1 2
3
.. n

Figure 2.8: Downsampled Signals

Now relate Go and G l to G: For m = 0,1, ... ,N-1

L
N-l
G(m) = g(n)e-i~mn
n=O
.If-l .If-l
= L g(2n)e-i~m(2n) + L g(2n + 1)e- i1fm (2n+l)
n=O n=O
+ Gt{m)e-'"h
Go(m) N m

= Go(m) + G l (m)W m . (2.66)

Note that this equation makes sense because we extended the definition of
Go(m) and Gl(m) from Z/f-l to ZN.
We have thus developed the scheme for computing G(m) given the
smaller sample DFTs Go and G l as depicted in Figure 2.9 for the case
of N = 8. In the figure, arrows are labeled by their gain, with unlabeled
arrows having a gain of 1. When two arrow heads merge, the signal at that
point is the sum of the signals entering through the arrows. The connection
pattern in the figure is called a butterfly pattern.
The total number of complex-multiply-and-adds in this case is now

2(N)2 + N
2 ~

T wo N'7-'DFT
=4 s
Combining Step
2.6. THE FAST FOURIER TRANSFORM 85

g(l) t---~--it--i!--it-~ G( 4)

g(3) t--~-f--+----'\r--~ G( 5)

DFT
N' =4
g(5) t---JI....'-.'--f---~.----~ G(6)

g(7) t----~-----~ G(7)

Figure 2.9: FFT Step 1

86 CHAPTER 2. THE FOURIER TRANSFORM

If N is large, then this is approximately N2/2 and the decomposition of the

DFT into two smaller DFTs plus a combination step has roughly halved
the computation.
Step 2: We reduce the computation even further by using the same
trick to compute the DFTs Go and G1 , that is, by further downsampling
the downsampled signals and repeating the procedure described so far. This
can be continued until we are transforming individual input signal symbols.
We now define

goo(n) = go(2n) = g(4n)

g01 (n) = go(2n + 1) = g(4n + 2)
glO(n) g1(2n) = g(4n + 1)
gl1 (n) g1(2n + 1) = g(4n + 3)
if -
for n = 0,1, ... , 1. In the N = 8 example, each of these signals has
only two samples. It can be shown in a straightforward manner that

Go(m) GOO(m) + G 01 (m)W 2m ; m = 0, 1, ... ,~ - 1 (2.67)

GlO(m) + G 11 (m)W 2m ; m = 0,1, ... , ~ - 1. (2.68)

This implies that we can expand on the left side of the previous flow graph
to get the flow graph shown in Figure 2.10. Recall that any unlabeled
branch has a gain of 1. We preserve the branch labels of W O to highlight
the structure of the algorithm.
The number of computations now required is

4(N )2 + N + N
4 ~ ~
~
This Step Combination Previous Step Combination
4 Nil = 2 DFTs

If N is large, the required computation is on the order of N 2 /4, again

cutting the computation by a factor of two.
Step 3: We continue in this manner, further downs amp ling the sequence
in order to compute DFTs. In our example with N = 8, this third step is
the final step since downsampling a signal with two samples yields a signal
with only a single sample:

gooo(O) = g(O) g100(0) = g(l)

gaOl (0) = g( 4) g101 (0) = g(5)
go 10 (0) = g(2) 9110 (0) = g(3)

gall (0) = g(6) 9111(0) =g(7). (2.69)

2.6. THE FAST FOURIER TRANSFORM 87

r--------,Goo(O) Go(O)
g(O) G(O)

DFT
Nil =2

g(4) 1----<t---if--+:i~-<t-----\----+___+7 G (1)

g(2)

DFT
Nil =2

g(6) I---...;..;....--~-~--i!--+--If--_ G(3)

g(l) I----<t----+:i~-~---*--+-*_~ G( 4)

DFT
Nil =2

g(5) I----<t---if--+:i~-~---f--+-T_~ G( 5)

g(3)

DFT
N"=2

g(7) I---+~--+--"""':"~----~ G(7)

Figure 2.10: FFT Step 2

88 CHAPTER 2. THE FOURIER TRANSFORM

The DFT of a single sample signal is trivial to find, i.e.,

GOOO(O) = gooo(O)W O= g(O) GlOO(O) = glOO(O)WO= g(l)

G001(0) = gOOl(O)W O= g(4) GlOl (0) = glOl (O)W O = g(5)
GOlO(O) = gOlO(O)WO= g(2) GllO(O) = gllO(O)W O= g(3)
GOll(O) = gOll(O)WO= g(6) Glll(O} = glll(O}W O= g(7).
(2.70)

In other words, the DFTs of the one-point signals are given by the signals
themselves. We can now work backwards to find Goo, G01 , G lO , and G ll
as follows:
GOO(m) = g(O) + g(4}w4mj m = 0,1
G Ol (m) = g(2) + g(6}w4mj m = 0,1
GlO(m} = g(l} + g(5}W 4m j m = 0, 1
G u (m) = g(3) + g(7}W 4m j m = 0,1.

The complete flow chart is shown in Fig. 2.11.

Either by counting arrows and junctions or studying the operations
required in each step, it is seen that each stage involves N multiplications
(not counting N of the trivial multiplications by I) and N additions to
combine the signal. We assumed that N = 2R is a power of two and
there are R = log2 N stages. Thus the overall complexity (as measured by
multiply / adds) is

IN log2 N complex-multiply-and-adds·1
This is the generally accepted computational complexity of the FFT.
There are, however, further tricks and variations that can yield lower com-
plexity. For example, of the N log2 N multiplies, if log2 N can be elimi-
nated by noting that W 4 = -Wo, W5 = _Wl, W 6 = _W2, and W 7 =
- W3, allowing half the multiplies to be replaced by inverters as shown in
Fig. 2.12.
The astute reader will have observed a connection between the binary
vector subscripts of the final single sample signals (and the correspond-
ing trivial DFTs) with the corresponding sample of the original signal in
Eq. 2.69. If the index of the original signal is represented in binary and
then reversed, one gets the subscript of the single sample final downsampled
sequence. For example, writing the argument 3 of g(3} in binary yields 011
which yields 110 when the bits are reversed so that g(3} = g110(0}.
The reduction of computation from roughly N2 to N log2 N may not
seem all that significant at first glance. A simple example shows that it is
2.6. THE FAST FOURIER TRANSFORM 89

Goo (0) Go(O)

g(O) G(O)

g(4) G(l)

g(2) G(2)

g(6) G(3)

g(l) G(4)

g(5) G(5)

g(3) G(6)

g(7) G(7)

Figure 2.11: FFT Algorithm Flow Chart: N=8

90 CHAPTER 2. THE FOURIER TRANSFORM

Figure 2.12: Inversion

indeed an enormous help in ordinary tasks. Consider the examples of the

Mona Lisa and MR images of Figures 1.31 and 1.32. These images consist
of 256 x 256 square arrays of pixels, for a total of N = 216 = 65536 pixels.
A brute force evaluation of the DFT by evaluating the double sums of
exponentials would take on the order of N 2 operations and the FFT would
take on the order of N log2 N operations. Hence the brute force method
requires

times as many computations. In our example with N = 216 , this means the
brute force approach will take roughly 216 /2 4 = 212 = 4096 times as long.
On a Macintosh IIci the Matlab FFT of one of these images took about 39
seconds. A brute force evaluation would take more than 45 hours! (In fact
it can take much longer because of the optimized code of an FFT and the
brute force evaluation of the powers of the complex exponentials required
before the multiply and adds.)
As a more extreme example, a typical digitized x-ray image has 2048 x
2048 pixels, yielding a computational complexity of 22 x 222 with the FFT
in comparison to 244 for the brute force method!

FFT Examples
We have already considered ID examples of the FFT when we computed
the DFT of the random signal and the sinusoid plus the random signal
in Figures 2.4-2.5. The advantages of the FFT become more clear when
computing 2D DFTs, e.g., of image signals. Before doing so, however, we
point out an immediate problem. Since images are usually represented as
a nonnegative signal (the intensity at each pixel is a nonnegative number),
they tend to have an enormous DC value. In other words, the value of
the Fourier transform for (Ix, fy) = (0,0) is just the average of the entire
image, which is often a large positive number. This large DC value dwarfs
2.6. THE FAST FOURIER TRANSFORM 91

the values at other frequencies and can make the resulting DFT look like
a spike at the origin with 0 everywhere else. For this reason it is common
to weight plots of the spectrum so as to deemphasize the low frequencies
and enhance the higher frequencies. The most common such weighting is
logarithmic: instead of plotting the actual magnitude spectrum IG (fx, fy) I,
it is common to instead plot

Glog(fX,fy) = log(l + IG(fx,fy)l). (2.71)

The term 1 in this expression is added to assure that when IGI has value
zero, so will Glog (fx, fy). Although other weightings are possible (the 1
and the magnitude spectrum can be multiplied by constants or one can use
a power of the magnitude spectrum rather than the log), this form seems
the most popular. Figures 2.13-2.24 show the Fourier transforms of several
of the 2D signal examples. Both mesh and image plots are shown for the
log weighted and unweighted versions of the simple box and disk functions.
For the Mona Lisa and MR images the mesh figures are not shown as the
number of pixels is so high as to render the mesh figures too black.

Figure 2.13: FFT of 2D Box of Figure 1.25: mesh

An optical illusion obscures the circular shape of the inner contour

curves. This example provides a warning that visual interpretation can
be misleading.
92 CHAPTER 2. THE FOURIER TRANSFORM

Figure 2.14: FFT of 2D Box of Figure 1.25: image

In these artificial images, the de component of the DFT is not so strong

as to completely wipe out the effects of the nonzero frequencies. One can
see in both the mesh and image plots the falling off of the magnitude
spectrum. The log weighting, however, clearly enhances the relative values
of the nonzero frequencies.
The unweighted FFT of natural images typically shows only a single
bright spot at the origin. In order to display the full scale, all nonzero
frequency components become so small as to be invisible. With the loga-
rithmic weighting the nonzero frequencies become visible and the structure
of the spectrum becomes more apparent.

2.7 * Existence Conditions

Recall that the Fourier transform of a finite duration discrete time signal
always exists provided only that the values of the signal are finite. In this
case the Fourier transform is simply a weighted sum of a finite number of
terms and there are no limits to concern us. In all of the other cases, the
2.7. * EXISTENCE CONDITIONS 93

Figure 2.15: Log Weighted FFT of 2D Box of Figure 1.25: mesh

existence of the Fourier transform is not as obvious. The Fourier transform

in those cases is defined either as a sum over an infinite set of indices, which
is a limit of simple sums, or as an integral, which is itself a limit of sums.
Whenever a quantity is defined as a limit, it is possible that the limit might
not exist and hence the corresponding Fourier transform might not be well-
defined. Existence conditions are conditions on a signal which force it to be
"nice" enough for the Fourier transform to exist for all (or perhaps at least
some) relevant frequency values. There is no single theorem providing easy-
to-use necessary and sufficient conditions for a transform to exist. Instead
we must be content with a few of the most important results providing
sufficient conditions for a Fourier transform to exist.
To make matters even worse, there are many ways to define the existence
of a limit and these different notions of limits can give different definitions
of limiting sums and integrals and hence of the Fourier transform. Delving
into these mathematical details is beyond the intended scope of this text,
but we can at least provide some simple conditions under which we are
guaranteed to have no problems and the Fourier transform exists in the
usual sense. Unfortunately these conditions exclude some interesting classes
of signals, so we also briefly describe a more general class for which the
Fourier transforms exist if we use a more general notion of integration
94 CHAPTER 2. THE FOURIER TRANSFORM

Figure 2.16: Log Weighted FFT of 2D Box of Figure 1.25: image

and infinite summation. Some of the details are described in the starred
subsections. For those who do not wish to (or are not asked to) read these
sections, the key points are summarized below.

• Discrete Time Signals.

- A sufficient condition for a discrete time infinite duration signal

{g(n)j n E 7} to have a Fourier transform is that it be absolutely
summable in the sense

L Ig(n)1 < 00.

nET

If this sum is finite and equals, say, M, then the spectrum G(f)
exists for all f E S DT ID and

IG(f)1 < M, all f. (2.72)

- If a discrete time infinite duration signal {g(n)j n E 7} has

2.7. * EXISTENCE CONDITIONS 95

Figure 2.17: FFT of 2D Disk of Figure 1.27: mesh

finite energy in the sense that

L
00

£g =Ig(nW < 00, (2.73)

n=-oo
then the Fourier transform need not exist for all frequencies. For
example, the signal g(n) = lin for n > 0 does not have a Fourier
transform at O. Finite energy is, however, a sufficient condition
for the existence of a Fourier transform in a mean square sense
or limit in the mean sense or L2 sense: There exists a function
GU) with the property that

lim
N-400
12
!

-2
1
IG(f) - L
N

n=-N
g(n)e-i21f/nI2 df = O. (2.74)

(The function is unique in that any two functions satisfying the

formula must be the same except for a set of f of zero measure,
e.g., they can differ on a finite collection of points of f.) In other
words, the truncated Fourier transforms
N
GN(f) = L g(n)e- i21f / n
n=-N
96 CHAPTER 2. THE FOURIER TRANSFORM

Figure 2.18: FFT of 2D Disk of Figure 1.27: image

might not converge in the ordinary sense as N -t 00, but they

do converge in the sense that there is a frequency function G(f)
for which the error energy

10 2 = [: IG(f) - GN(fW df
2

between G(f) and the approximations GN(f) converges to 0 as

N -t 00.

• Continuous Time Signals.

- A sufficient condition for a continuous time signal {g(t)j t E T}
(finite or infinite duration) to have a Fourier transform is that
it be absolutely integrable in the sense

r
itET
Ig(t)1 dt < 00.

If this integral is finite and equals, say, M, then the spectrum

G(f) exists for all f E SCTFD for finite duration or f E SCTID
2.7. * EXISTENCE CONDITIONS 97

Figure 2.19: Log Weighted FFT of 2D Disk of Figure 1.27: mesh

for infinite duration and

IGU)I < M, all f. (2.75)

- If a continuous time signal {get); t E T} has finite energy in the

sense that

(2.76)

then the Fourier transform need not exist for all frequencies. For
example, the signal get) = lit for 0 < t < 1 and 0 otherwise does
not have a Fourier transform at f = O. Finite energy is, however,
a sufficient condition for the existence of a Fourier transform in a
mean square sense or limit in the mean or L2 sense analogous to
the discrete time case. We will not treat such transforms in much
detail because of the complicated analysis required. We simply
point out that the mathematical machinery exists to generalize
most results of this book from the absolutely integrable case to
the finite energy case.
98 CHAPTER 2. THE FOURIER TRANSFORM

Figure 2.20: Log Weighted FFT of 2D Disk of Figure 1.27: image

* Discrete Time
The most common infinite duration DTFT has r equal to the set of all
integers and hence

L
00

GU) = g(n)e- i21r / n . (2.77)

n=-CX)

This infinite sum is in fact a limit of finite sums and the limit may blow
up or not converge for some values of f. Thus the DTFT mayor may not
exist, depending on the signal. To be precise, an infinite sum is defined as

N
L L
00

an = lim an
n=-oo N-+oo,K-+oo n=-K

if the double limit exists, i.e., converges. Mathematically, the double limit
exists and equals, say, a if for any f > 0 there exist numbers M and L such
2.7. * EXISTENCE CONDITIONS 99

Figure 2.21: FFT of Mona Lisa of Figure 1.31

that if N ~ M and K ~ L, then

K
I L an - al :5 f.
n=-N

Note that this means that if we fix either N or K large enough (bigger than
M or L above, respectively) and let the other go to 00, then the sum cannot
be more than € from its limit. For example, the sum L~=-N 1 = K + N + 1
does not have a limit, since if we fix N the sum blows up as K -+ 00.
A Fourier transform is often said to exist if the limiting sum exists in
the more general Cauchy sense or Cauchy principal value sense, that is, if
the limit
00 N

~ an
"" = N-+oo
lim ~ an
""
n=-oo n=-N

exists. We will not dwell on the differences between these limits; we simply
point out that care must be taken in interpreting and evaluating infinite
sums. The infinite sum L~=-oo n does exist in the Cauchy principal value
100 CHAPTER 2. THE FOURIER TRANSFORM

Figure 2.22: Log Weighted FFT of Mona Lisa of Figure 1.31

sense (it is 0). Similarly, the sum

1
L
00

k=-oo,k;to
k

does not exist in the strict sense, but it does exist in the Cauchy sense.
Several of the generalizations of Fourier transforms encountered here and
in the literature are obtained by using a weaker or more general definition
for an infinite sum.
A sufficient condition for the existence of the sum (and hence of the
transform) can be shown (using real analysis) to be

L
00

Ig(n)1 < OOj (2.78)

n=-oo

that is, if the signal is absolutely summable then the Fourier transform
2.7. * EXISTENCE CONDITIONS 101

Figure 2.23: FFT of Magnetic Resonance Image of Figure 1.32

exists in the usual sense of convergence of sums. For example, if

L
00

Ig(n)1 = M < 00,

n=-oo
then application of the inequality

(2.79)

implies that
00

IGU)I = L
n=-oo
gne-i27r/nl

00

< L
n=-oo
Igne-i27r/nl

00

= L
n=-oo
Ignlle-i27r/nl
102 CHAPTER 2. THE FOURIER TRANSFORM

Figure 2.24: Log Weighted FFT of Magnetic Resonance Image of Fig-

ure 1.32

L
00

= Ignl=M.
n=-oo

It should be kept in mind that absolute summability is a sufficient but

not necessary condition. We will encounter sequences which violate ab-
solute summability yet have an ordinary Fourier transform. We will also
encounter sequences which violate the condition and do not have an ordi-
nary Fourier transform, but which do have a Fourier transform in a suitably
generalized sense. As examples of the absolute summability condition, note
that the signal {rkj k = 0,1, ... } is absolutely summable if Irl < 1. It is
not absolutely summable if Irl ~ 1. The sequence {k-1j k = 1,2, ... } is
not absolutely summable, but the sequence {k- 2 j k = 1,2, ... } is.
We shall encounter several means of generalizing the Fourier transform
so that a suitably defined transform holds for a larger class of signals. This
is accomplished by generalizing the idea of an infinite sum (or integral). An
important example of such a generalization is the following. Recall that a
discrete time signal is said to be square summable or to have finite energy
2.7. * EXISTENCE CONDITIONS 103

if it satisfies (2.73). For example, the sequence {g(n) = 1/nj n = 1,2, ... }
has finite energy but is not absolutely summable. If a signal has finite
energy, then it can be proved that the Fourier transform G(f) exists in the
following sense:

lim
N.-+oo
12
1

1
-2
IG(f) - L
n=-N
N
g(n)e- i2 71'/nI2 df = O. (2.80)

When this limit exists we say that the Fourier transform G(f) exists in the
sense of the limit in the mean. This is sometimes expressed as
N
G(f) = N.-+oo
l.i.m. ' " g(n)e- i2 71'/n
~
n=-N

where "l.i.m." stands for "limit in the mean." Even when this sense is
intended, we often write the familiar and simpler form

L
00

G(f) = g(n)e- i2 71'/n,

n=-oo

but if finite energy signals are being considered, the infinite sum should be
interpreted as an abbreviation of (2.80). Note in particular that when a
Fourier transform of this type is being considered, we cannot say anything
about the ordinary convergence of the sum E:=-K g(n)e- i2 71'/n for any
particular frequency f as K and N go to 00, we can only know that an
integral of the form (2.80) converges.
It is not expected that this definition will be natural at first glance,
but the key point is that one can extend Fourier analysis to finite energy
infinite duration signals, but that the definitions are somewhat different.
We also note that discrete time finite energy signals are sometimes called
[2 sequences in the mathematical literature.
In the discrete time case, the property of finite energy is indeed more
general than that of absolutely summablej that is, absolute summability
implies finite energy but not vice versa. For example, if

Llgnl ~ M < 00,

n
then

n n k

n nf.k n
104 CHAPTER 2. THE FOURIER TRANSFORM

* Continuous Time
The most common finite duration continuous time Fourier transform con-
siders a signal of the form g = {get); t E [0, Tn and has the form

GU) = loT g(t)e-i2rrftdt. (2.81)

Unfortunately, unlike the finite duration DTFT case we cannot guarantee

that this integral always exists. It can be shown that the integral exists for
all f if the signal itself is absolutely integrable in the sense that

loT Ig(t)ldt < 00. (2.82)

As in the DTFT case, this condition is sufficient but not necessary. An

example of a signal violating absolute integrability is {t- 1 ; t E (0, In. As
in the discrete time case, we can extend the definition of a Fourier transform
to include signals with finite energy. We postpone that discussion until the
infinite duration case.
The most common infinite duration CTFT considers a signal of the form
g = {get); t E R} and the transform is given by

(2.83)

The usual definition of an improper (infinite limit) Riemann integral is

1 00
-00
g(t)e-i2rrftdt = lim lim
5-t00 T-too
1T
-5
g(t)e-i2rrftdt, (2.84)

if the limits exist. For such a double limit to exist, one must get the
same answer when taking Sand T to their limits separately in any manner
whatsoever.
We formalize the statement of the basic existence theorem so as to ease
comparison with later existence theorems. No proof is given (it is standard
integration theory).

Theorem 2.1 A sufficient condition for the existence of the transform in

the strict sense (that is, an ordinary improper Riemann integral) is that the

I:
signal be absolutely integrable:

Ig(t)ldt < 00. (2.85)

2.7. * EXISTENCE CONDITIONS 105

As in the finite duration case, this is sufficient but not necessary for
the existence of the CTFT. Signals violating the condition include {t(l +
t 2)-1; t E 'R} and {sinet; t E 'R}. Observe that if a signal is absolutely
integrable, then its transform is bounded in magnitude:

IGU)I == 11: g(t)e- i21T / t dtl :::; 1: Ig(t)e- i21T / t l dt

= 1: Ig(t)lle- i21T / t l dt = 1: Ig(t)1 dt < 00.

Note the use of the basic integration inequality

I J g(t) dtl :::; JIg(t)1 dt

in the previous inequality chain.

As a first generalization consider the case where the improper integral
exists in the Cauchy principal value sense; that is,

['X) g(t)e- i21T / t dt =

1-00
lim
T~001_T
rT
g(t)e- i21T / t dt. (2.86)

This is more general than the usual notion. (That is, if the integral exists as
an improper Riemann integral, then it also exists in the Cauchy principal
value sense. The converse, however, is not always true.) The following
theorem gives sufficient conditions for the Fourier transform to exist in this
sense. A proof may be found in Papoulis [24].

Theorem 2.2 If g(t) has the form

g(t) = f(t) sin(wot + <Po) (2.87)

for some constants Wo and <Po, where If(t)1 is monotonically decreasing as

ItI ~ 00, and where g(t)/t is absolutely integrable for It I > A > 0, i.e.,
roo Ig(t) Idt < 00 (2.88)
1A t
then the Fourier transform GU) of g(t) exists in the Cauchy principal value
sense.

This general condition need not hold for all interesting signals. For
example, if g(t) is equal to 1 for all t, the condition of (2.88) is violated. The
most important example of a signal meeting the conditions of this theorem
106 CHAPTER 2. THE FOURIER TRANSFORM

is g(t) = sinc(t). This signal is not absolutely integrable, but it meets the
conditions of the theorem with ¢o = 0, Wo = 1l', and f(t) = 1/(1l't).
To prepare for a final existence theorem (for the present, at least), we
say that the Fourier transform of a signal 9 = {g(t); t E T} exists in a
limit-in-the-mean sense if the following conditions are satisfied:
• The signal can be approximated arbitrarily closely by a sequence of
signals gN = {gN(t); t E T} in the sense that the error energy goes
to zero as N -t 00:

lim
N-too iT
r Ig(t) - gN(tW dt = 0, (2.89)

i.e., 9 is the limit in the mean of gN;

• all of the signals gN have Fourier transforms in the usual sense, i.e.,

GN(f) = 19N(t)e-i21f'tdt (2.90)

is well-defined for all f in the frequency domain of definition; and

• GN has a limit in the mean, i.e., if T = [0, T) and S = {kiT; k E Z},
then there is a function G(f) for which

lim "IG(f) - G N (f)1 2

N-tooL..J
= 0; (2.91)
IES

if T = Rand S = R, then there is a function G(f) for which

lim
N-too isr IG(f) - GN(fW df = 0 (2.92)

(The convergence is called h for the sum in the first case and L2 for
the integral in the second case.)
If these conditions are met then G is a Fourier transform of g. (We say "a
Fourier transform" not "the Fourier transform" since it need not be unique.
For example, changing a G in a finite way at a finite collection of points
yields another G with the desired properties since Riemann integrals are
not affected by changing the integrand at a finite number of points.) This
is usually expressed formally by writing

but in the current situation this formula is an abbreviation for the more
exact definition above.
2.8. PROBLEMS 107

Theorem 2.3 A sufficient condition for a continuous time signal 9 to pos-

sess a Fourier transform in the sense of limit-in-the-mean is that it have
finite energy (or be square integrable):

(2.93)

Finite energy neither implies nor is implied by absolutely integrability.

For example, the signal sinJ;t)
is square integrable but not absolutely inte-
grable, while the signal {[v'i(l + t))-l j t E (O,oo)} is absolutely integrable
but not square integrable. Square integrable functions are often called L2
functions in the mathematical literature (like the notation 12 for square
summable functions). Once again, finite energy signals have a Fourier the-
ory, it is just different (and somewhat more complicated) than that used
for absolutely integrable functions.

2.8 Problems
2.1. Let 9 be the signal in problem 1.11 Find the Fourier transforms of 9
and its zero-filled extension g.
2.2. Find the OTFT of the following infinite duration signals (T = Z)
signals:
(a) gn = rlnl, where Irl < 1. What happens if r = I?
(b) Un = On-k, the shifted Kronecker delta function, where k is a
fixed integer.
(c) gn = Ef=-N On-k'
(d) Un = a for Inl : : ; N °
and Un = otherwise.
n a °
(e) Un = r - for n ~ a and Un = otherwise. Assume that Irl < 1.
2.3. Find the OFT of the following sequences:

(a) {Un} = {l,l,l,l,l,l,l,l}

(b) {Un} = {l,O,O,O,O,O,O,O}
(c) {gn} = {O, -1,0,0,0,0,0, I}
(d) {Un} = {I, -1, 1, -1, 1, -1, 1, -I}
2.4. Find the OFT of the following 8-point sequences:
108 CHAPTER 2. THE FOURIER TRANSFORM

(b) {gn} = {cos(~n)}

(c) {gn} = {sin(1T4n)}

2.5. Find the DFTs of the following signals:

(a) g1 = {2,2,2,2,2,2,2,2}
(b) g2 = {e i1Tn / 2; n = 0,1,2,3,4,5,6, 7}
(c) g3 = {e i1T (n-2 l /2 i n = 0,1,2,3,4,5,6, 7}
(d) g4 ={0,0,0,0,0,0,1,0}

2.6. Find the DFTs of the following sequences:

(a) {gn} = {0,0,0,1,0,0,0,0}

(b) {gn} = {cos(71'n); n = 0, 1, ... , 7}
(c) {gn} = {cos(1T2n); n=0,1, ... ,7}
(d)
{ 1 -~, n = 0,1,2,3
gn= ~-1, n=4,5,6,7.

2.7. Suppose that 9 is an infinite duration continuous time signal that is

nonzero only for the interval [0, T). Find an approximation to the
Fourier transform G(f) of 9 in terms of the DFT G of the sampled
signal {ani n E ZN}' For what values of frequency f does the ap-
proximation hold?

2.8. Find the CTFT of the following signals using the following special
signals (7 is the real line in all cases): The rectangle function

1 if -~ < t < ~
if It I = ~
net) = { ~
° otherwise

and the unit step function

H(t) ={
1
o~ t
t >
=
°
°
otherwise

(a) get) = e->.t H(t)

(b) get) = n(~ - b) where a > 0.
2.8. PROBLEMS 109

(c)

g(t)={ ~(1-~) -a ~ t ~ a
otherwise

where c is a fixed constant.

(d) The raised cosine pulse g(t) = n(t)(l + cos(2wt)); t E R.
(e) The Gaussian pulse g(t) = e-lI"t 2 ; t E R. (Hint: Use the tech-
nique of "completing the square.")

2.9. Find the CTFT of the following signals.

(a) g(t) = rltl; t E R, where 0 < r < 1.

(b) g(t) = sgn(t); t E [-1,1).
(c) g(t) = sin(2wJot); t E [-w,w].
(d) The pulse defined by p(t) = A if t E [0, r] and 0 otherwise, t E R.

2.10. By direct integration, find the Fourier transform of the infinite dura-
tion continuous time signal g(t) = t A (t).

2.11. Find the Fourier transform ofthe signals {e->'It l; t E R} and {sgn(t)e->'Itl;
t E R}.

2.12. Find the CTFT of the following signals:

(a) g(t) = At for t E [-T/2, T/2] and 0 for t E R but t (j. [-T/2, T/2].
(b) {I sintl; It I < w}.
(c)
for 0 < t < T/2
for t= 0
for - T /2 < t < 0 .
for It I ~ T/2

2.13. An important property of the Fourier transform is linearity. If you

have two signals 9 and h with Fourier transforms G and H, respec-
tively, and if a and b are two constants (i.e., complex numbers), then
the Fourier transform of ag + bh is aG + bH.

(a) Prove this result for the special cases of the DFT (finite duration
discrete time Fourier transform) and the infinite duration CT
Fourier transform.
110 CHAPTER 2. THE FOURIER TRANSFORM

(b) Suppose that 9 is the continuous time one-sided exponential

g(t) = e- At u_l(t) for all real t and that h is the signal defined
by h(t) = g( -t). Sketch the signals 9 + hand 9 - h and find the
Fourier transforms of each.

2.14. An important property of the Fourier transform is linearity. If you

have two signals 9 and h with Fourier transforms G and H, respec-
tively, and if a and b are two constants (i.e., complex numbers), then
the Fourier transform of ag + bh is aG + bH.

(a) Prove this result for the special cases of the infinite duration
DTFT.
(b) Use this result to evaluate the Fourier transform of the signal
9 = {gnj n E Z} defined by

n=O,l, ...
otherwise
and verify your answer by finding the Fourier transform directly.

2.15. Suppose that 9 is the finite duration discrete time signal {onj n =
0,···, N - I}. Find F(F(g», that is, the Fourier transform of the
Fourier transform of g. Repeat for the signal h defined by h n = 1 for
n = k (k a fixed integer in {O, ... , N - I}) and hn = 0 otherwise.

2.16. Suppose you know the Fourier transform of a real-valued signal. How
can you find the Hartley transform? (Hint: Combine the transform
and its complex conjugate to find the sine and cosine transforms.)
Can you go the other way, that is, construct the Fourier transform
from the Hartley transform?

2.17. Suppose that 9 = {g(h, v)j h E [0, H], v E [0, V)} (h represents hor-
izontal and v represents vertical) represents the intensity of a sin-
gle frame of a video signal. Suppose further that 9 is entirely black
(g(h, v) = 0) except for a centered white rectangle (g(h, v) = 1) of
width aH and height a V (a < 1). Find the two-dimensional Fourier
transform F(g).

2.18. Suppose that 9 = {g(h, v)j h E [0, H), v E [0, V)} is a two dimen-
sional signal (a continuous parameter image raster). The independent
variables h and v stand for "horizontal" and "vertical", respectively.
The signal g( h, v) can take on three values: 0 for black, 1/2 for grey,
and 1 for white. Consider the specific signal of Figure 2.25.
2.8. PROBLEMS 111

v g=O
4vl5
3vIs
2Vl5 0=1
VIS

Figure 2.25: 20 "Stripes"

(a) Write a simple expression for g(h, v) in terms of the box function

DT(X) = {I Ixi ~ T .
o otherwise
(You can choose 9 to have any convenient values on the bound-
aries as this makes no difference to the Fourier transform.)
(b) Find the 2-D Fourier transform G(/h, 11J} of g.

2.19. Find the Fourier transform of the binary image below.

0000
1001
0000
1111
2.20. Supposethatg = {g(k,n}; k,n E ZN} is an image and G = {G(~, N}; k,n E
ZN} is its Fourier transform. Let W denote the complex exponential
matrix defined in (2.6). Is it true that

(2.94)
Note that gtwt = (Wg}t.
2.21. Consider an N = 4 FFT.

(a) Draw a flow graph for the FFT.

112 CHAPTER 2. THE FOURIER TRANSFORM

(b) Draw a modified flow graph using inverters to eliminate half the
multiplies.

2.22. (a) Express the DFT of the 9-point sequence {gO.gl, ... , gs} in terms
of the DFTs of the 3-point sequences

ga(n) = {go,g3,g6}
gb(n) = {gl,g4,gr}
gc(n) = {g2,g5,gS}.

(b) Draw a legible flow graph for the "base 3" method for computing
the FFT, as suggested above, for N = 9.
2.23. Suppose you have an infinite duration discrete time signal 9 = {gn; n E
Z} and that its Fourier transform is G = {G(f); f E [-1/2,1/2)}.
Consider the new signals

g2n+l; n E Z
Vn = { gn/2 if n is an even number
o otherwise
with Fourier transforms H, W, and V, respectively. hand ware ex-
amples of downsampling or subsampling. v is called upsampling. Note
that downsampling and upsampling are not generally inverse opera-
tions, i.e., downsampling followed by upsampling need not recover the
original signal.
(a) Find an expression for V in terms of G.
(b) Find an expression for G in terms of Wand H. (This is a
variation on the fundamental property underlying the FFT.)
(c) Suppose now that r = {rn; n E Z} is another signal and that p
is the upsampled version of r, i.e., Pn = r n /2 for even nand 0
otherwise. We now form a signal x defined by

xn = Vn + Pn+l; n E Z.

This signal can be considered as a time-division multiplexed (TDM)

version of 9 and r since alternate samples come from one or the
other. Find the Fourier transform X in terms of G and R.
(d) Evaluate the above transforms and verify the formulas for the
case gn = 2- n u_l(n), rn = 4- n u_l(n).
2.8. PROBLEMS 113

2.24. Consider the signal 9 = {gn = nrnu_l(n); n E Z}, where Irl < 1.
(a) Is this signal absolutely summable?
(b) Find a simple upper bound to IG(f)1 that holds for all f.
(c) Find the Fourier transform of the signal g.
(d) Consider the signal h = {h n ; n E Z} defined by h n = g2n' (h is
a downsampled version of g.) Find the DTFT of h.
(e) Find a simple upper bound to IH(f)1 that holds for all f.
(f) Consider the signal W = {w n ; n E Z} defined by W2n = h n and
W2n+l = 0 for all integer n. W is called an upsampled version of
h. Find the DTFT of w.
(g) Find the DTFT of the signal g - w.
Chapter 3

Fourier Inversion

Having defined the Fourier transform and examined several examples, the
next issue is that of invertibility: if G = F(g), can 9 be recovered from the
spectrum G? More specifically, is there an inverse Fourier transform F- 1
with the property that
F- 1 (F(g)) = g? (3.1)
When this is the case, we shall call 9 and G a Fourier transform pair and
write
9 f-t G, (3.2)
where the double arrow notation emphasizes that the signal and its Fourier
transform together form a Fourier transform pair. We have already seen
that Fourier transforms are not always invertible in the strict sense, since
changing a continuous signal at a finite number of points does not change
the value of the Riemann integral giving the Fourier transform. For exam-
ple, {Ol/2(t); t E 'R} and {n(t); t E 'R} have the same transform. In this
chapter we shall see that except for annoying details like this, the Fourier
transform can usually be inverted.
The first case considered is the DFT, the finite duration discrete time
Fourier transform. This case is considered first since an affirmative answer
is easily proved by a constructive demonstration. The remaining cases are
handled with decreasing rigor, but the key ideas are accurate.

3.1 Inverting the DFT

Given a finite duration discrete time signal 9 = {gn; n E {a, 1, ... , N -I}}
the DFT is defined for all real f by GU) = L,~:;} gne-i2rr/n. Although
this formula is well defined for all real f, we earlier restricted the frequency
116 CHAPTER 3. FOURIER INVERSION

domain of definition of G(f) for a discrete time finite duration signal to

the subset of real frequencies Sg~FD = {O, liN, 2/N, ... , (N - l)IN}. We
now proceed to justify this selection by showing that this set of frequencies
suffices to recover the original signal. We point out that this set is not
unique; there are other choices that work equally well. It is, however,
simple and the most common.
The present goal is this: suppose that we know a signal g has a spectrum
G = {G(kIN);k = O,l, ... ,N -I} or, equivalently, G = {G(f); f =
0, liN, ... , (N - l)IN}, but we do not know the signal itself. How do we
reconstruct g from G?
Consider for n = 0,1, ... , N - 1 the sequence
N-l

Yn = ~
N "
~
G(~)
N e i27r~n . (3.3)
k=O

Here Yn can be considered as N- 1 times the DFT of the signal G with -i

replacing i in the exponential. Since

(3.4)

we have that
L L g,e-i27r~lei27r~n
N-l N-l
Yn =~ (3.5)
k=O 1=0
and hence exchanging the order of summation, which is always valid for
finite sums, yields
N-l
L
N-l
Yn = g, ~ L ei27r~(n-I). (3.6)
1=0 k=O

To evaluate this sum, consider

If m = 0 (or is any multiple of N) the sum is 1. If m is not a multiple of

N, then the geometric progression formula implies that

1 N-l. 1 1 _ i27rNN
_ " et27r~m = _ e m = O.
N ~ N l-e a7r N
k=O
3.1. INVERTING THE DFT 117

Readers may recognize this as a variation on the fact that the sum of all
roots of unity of a particular order is O. Alternatively, adding up N equally
spaced points on a circle gives their center of gravity, which is just the origin
of the circle. Observe that the m = 0 result is consistent with the m =P 0
result if we apply L'Hopital's rule to the latter.
Recalling the definition of the Kronecker delta function 8m :

8 _{1a
m -
ifm = a
otherwise' (3.7)

then
1
L
N-l . k

N e t21r N"m = 8mmodN; for any integer m. (3.8)

k=O
Since both nand l in (3.6) are integers between a and N -1, the sum inside
(3.6) becomes
N-l
~
N "~ e i21r j;(n-l) = 8n-l,. n , l E ZN (3.9)
k=O
and hence
N-l

Yn = L
1=0
g18n - 1 = gn; (3.10)

that is, the original sequence has been perfectly recovered.

Thus given a spectrum G = {Gk = G(k/N); k = 0,1, ... ,N - 1} it is
reasonable to define the inverse discrete Fourier transform (or IDFT) by

1 N-l k . k

9n = -N~
" G(_)e
N
t21r N"n.
' n = 0"
1 '", N - 1, (3.11)
k=O

and we have shown that

9 = F- 1 (G) = F- 1 (F(g)). (3.12)

In summary we have shown that the following are a Fourier transform pair:

N-l
L gn e- i21r j;n; kE ZN (3.13)
n=O
1 N-l k . k
gn = -N~" G(_)e'
N
21r N"n. n E Z
' N·
(3.14)
k=O

Several observations on this result are worth making.

118 CHAPTER 3. FOURIER INVERSION

• Recall from the matrix form of the OFT of (2.7) that G = Wg,
·2" k·
where W = {e-a-w J; k = 0, I,···,N -I; j = 0, I,···,N -I}. From
elementary linear algebra this implies that

This is readily identified as the vector/matrix form of (3.14) if we

make the identification

W- 1 = N-l{e+i~kj; k=O,I, .. ·,N-I; j=O,I,· .. ,N-I}

= N-1W*, (3.15)

where the * attached to a matrix or vector indicates the conjugate

transpose, that is, the matrix formed by taking the complex conjugate
and then transposing. The matrix inversion of (3.15) is just the matrix
formulation of (3.9).
• Instead of the frequency domain of definition {O, I/N, ... , (N -I)/N},
we could use any shift of this domain by an integer K, that is, any
frequency domain of the form {K,K +I/N,···,K +(N -I)/N}. With
this choice the Fourier transform of {Un; n = 0, ... , N - I} becomes
{G(k/N); k = KN,KN+l, .. · ,KN+N -I} and the IDFT becomes

because of the periodicity of complex exponentials; that is, because

for each l
e'·2 "II" KX±I n = e'·2 "ll"J'iTn.
I

• Different choices for the time domain of definition lead to similar

alterations in the inversion formula. For example, if the signal is two-
sided: {Un; n = -N,· .. , -1,0, I,· .. , N}, then the OFT is commonly
taken as
k
{G(2N + 1); k = -N, -N + 1,"', -1,0, I,N -I,N}; (3.16)

that is, the frequency domain of definition is

(2) NIl N
SDTFD = {- 2N + I' ... , -
2N + 1,0, 2N + I' , 2N + I}'

As an exercise you should prove the corresponding inversion formula:

3.1. INVERTING THE DFT 119

G(f) = g n e -i21f/n., /ES(2)

DTFD (3.17)
n=-N

I Nk
gn = " G( )ei21f~n. (3.18)
2N + 1 L.,; 2N + 1 '
k=-N
n E {-N, .. ·,O, .. ·,N}.

• The key property yielding the inversion is the orthogonality of the

discrete time exponentials. A collection of finite duration discrete
time complex valued signals {¢hm ) j n E ZN} for some collection of
integers m is said to be orthogonal if

L
N-1
¢kn)¢~)· = C n 8n - 1 (3.19)
k=O

for C n =I- OJ i.e., two signals are orthogonal if the sum of the coordinate
products of one signal with the complex conjugate of the other is 0
for different signals, and nonzero for two equal signals. If C n = 1 for
all appropriate n, the signals are said to be orthonormal. Eq. (3.9)
implies that the exponential family {e i21f *mj k = 0,1, ... , N - I} for
m = 0, 1, ... , N - 1 are orthogonal and the scaled signals

VN
1 {i21f*m.
e , 0 1, ... , N - I}
k =,

are orthonormal since for I, n E Z N

N-1
2.
N "L.,; e-
i21f * l e i21f-f,n =i:
UI-n· (3.20)
k=O

The matrix form of this relation also crops up. Let W be the expo-
nential matrix defined in (2.6) and define

(3.21)

Then (3.20) becomes in matrix form

U- 1 = U*j (3.22)
that is, the inverse of the matrix U is its complex conjugate trans-
pose. A matrix with this property is said to be unitary. Many of
120 CHAPTER 3. FOURIER INVERSION

the properties of Fourier transforms for finite duration sequences (or,

equivalently, finite dimensional vectors) generalize to unitary trans-
formations of the form G = Hg for a unitary matrix H .
• The factor of l/N in the IDFT could have easily been placed in the
definition of the Fourier transform instead of in the inverse. Alterna-
tively we could have divided it up into symmetric factors of l/VN in
each transform. These variations are used in the literature and also
referred to as the discrete Fourier transform.
• The Fourier inversion formula provides a representation or a decompo-
sition into complex exponentials for an arbitrary discrete time, finite
duration signal. Eq. (3.14) describes a signal as a weighted sum of
exponentials, where the weighting is given by the spectrum. As an
example, consider the signal 9 = {gn = rn; n E ZN}. The DFT is
given by (2.20) and hence (3.14) becomes

(3.23)

a formula that might appear somewhat mysterious out of context.

The formula decomposes rn into a weighted sum of complex expo-
nentials of the form
N-l
gn =L Ck ei27r ';;n; n E ZN. (3.24)
k='O

This can be viewed as one way to represent the signal as a simple

linear combination of primitive components. It is an example of a
Fourier series representation of a signal, here a discrete time finite
duration signal. This representation will be extremely important for
finite duration signals and for periodic signals and hence we formalize
the idea in the next section.
As another example of a Fourier series, (3.9) immediately provides a
Fourier series of the Kronecker delta defined on a finite domain; that
is, the signal {6 n ; n = 0,1, .. " N - I} has a Fourier series

1 N-l . k
6n = -N "'"'
~
et27rwn.
, n E Z N· (3.25)
k=O

This again has the general form of (3.24), this time with Ck = 1/N
for all k E ZN.
3.2. DISCRETE TIME FOURIER SERIES 121

3.2 Discrete Time Fourier Series

We can rewrite (3.14) as
N-l
gn = L Ckei27r~n; n E ZN (3.26)
k=O

where
N-l

Ck = N1 G( Nk ) = N1 "'" N; k E ZN.
~ gle -i27rl-"- (3.27)
1=0

This decomposition of a signal into a linear combination of complex ex-

ponentials will be important in the study of linear systems where complex
exponentials play a fundamental role. An exponential sum of this form
is called a Fourier series representation of the discrete time finite dura-
tion signal g. This representation has an interesting and useful extension.
Although the left-hand side of (3.26) is defined only for n E ZN, the right-
hand side gives a valid function of n for all integers n. Furthermore, the
right-hand side is clearly periodic in n with period N and hence consists
of periodic replications of the finite duration signal g. In other words, we
have a convenient representation for the periodic extension y of g:

N-l 1 k . k
g- = "'" _G(_)e'27rNn. n E Z (3.28)
n ~ N N ' .
k=O

Conversely, if we are given a discrete time periodic signal Yn, then we

can find a Fourier series in the above form by using the Fourier transform
of the truncated signal consisting of only a single period of the periodic
signal; that is,

=L
N-l
Yn ckei27r~n; n E Z (3.29)
k=O
N-l
Ck -- ~
N "'"
~ g-I e- i27rl -/!r., k E Z N· (3.30)
1=0
This provides a Fourier representation for discrete time infinite duration
periodic signals. This is a fact of some note because the signal, an infinite
duration discrete time signal, violates the existence conditions for ordinary
Fourier transforms (unless it is trivial, i.e., gn = 0 for all n). If gn is ever
nonzero and it is periodic, then it will not be absolutely summable nor
will it have finite energy. Hence such periodic signals do not have Fourier
transforms in the strict sense. We will later see in Chapter 5 that they have
122 CHAPTER 3. FOURIER INVERSION

a generalized Fourier transform using Dirac delta functions (generalized

functions), but for the moment we have a perfectly satisfactory Fourier
representation without resort to generalizations.

3.3 Inverting the Infinite Duration DTFT

We next turn to the general DTFT with infinite duration. The general idea
will be the same as for the OFT, but the details will be different. Suppose
now that we have a signal 9 = {gn; n E Z}, where as usual Z is the set of
all integers. Assume that the Fourier transform

L
00
G(f) = gke-i21rfk (3.31)
k=-oo

exists. Since G(f) is a periodic function of f with period 1 (because e-i21rfk

is periodic in f with period 1), we can consider S = [-1/2,1/2) (or any
other interval of unit length such as [0, 1)) to be the domain of definition
of the spectrum and we can consider the spectrum to be G = {G(f); f E
[-1/2, 1/2)}. We have that
1

Yn = (2 G(f)ei27rfn df
L!
= [! ( f: gke-i21rkf) ei27rfn df
! k=-oo

L
00 1
= gk (2 ei21rf (n-k) df. (3.32)
k=-oo L!
To complete the evaluation observe for m = 0 that
j-!! ei21rmf df = j!-! df = 1

and for any nonzero integer m that

j-!
1
2 ei21rmf df =0

and hence
[21
1
ei21rmf df = 6m ; mE Z. (3.33)
2
3.3. INVERTING THE INFINITE DURATION DTFT 123

Thus inserting (3.33) into (3.32) we have that

Yn L
00

k=-oo
gk 1 1
2
,

-2
e i27rf (n-k) df

(3.34)

recovering the signal from its transform ..

Observe the close resemblance of (3.33) to (3.8). As in the case of the
DFT, the key property of exponentials leading to the inversion formula is
the orthogonality of the signals {e i21Tkf ; f E [-1/2, 1/2)} for k E Z in the
,
12
sense that
e i21Tkf e-i2rrlf df = 15k-I; k, IE Z, (3.35)
-~
the continuous time finite duration analog of (3.20). It must be admitted
that we have been somewhat cavalier in assuming that the sum and the in-
tegral could be exchanged above, but under certain conditions the exchange
can be mathematically justified, e.g., if the signal is absolutely summable.
We have shown that

gn = /2,
-!
G(f)ei2rrfn df; n E Z (3.36)

and hence that the right hand side above indeed gives the inverse Fourier
transform. For example, application of the discrete time infinite duration
inversion formula to the signal rnU_l (n); n E Z for Irl < 1 yields

(3.37)

Unlike (3.14) where a discrete time finite duration signal was represented
by a weighted sum of complex exponentials (a Fourier series), here a dis-
crete time infinite duration signal is represented as a weighted integral of
complex exponentials, where the weighting is the spectrum. Instead of a
Fourier series, in this case we have a Fourier integral representation of a sig-
nal. Intuitively, a finite duration signal can be perfectly represented by only
a finite combination of sinuosoids. An infinite duration signal, however, re-
quires a continuum of frequencies in general. As an example, (3.33) provides
the infinite duration analog to the Fourier series representation (3.25) of a
124 CHAPTER 3. FOURIER INVERSION

finite duration discrete time Kronecker delta: the infinite duration discrete
time signal {8 n ; n E Z} has the Fourier integral representation

(3.38)

To summarize the infinite duration discrete time Fourier transform and

inversion formula, the following form a Fourier transform pair:

00
1 -)
1
G(f) = L
n=-oo
9n e-''2 71' / n., j E [--2' 2 (3.39)

gn = [ : G(f)e i2 71'/n dj; n E Z. (3.40)

2

With this definition the Fourier inversion of (3.12) extends to the more
general case of the two-sided DTFT. Note, however, the key difference be-
tween these two cases: in the case of the DFT the spectrum was discrete in
that only a finite number of frequencies were needed and the inverse trans-
form, like the transform itself, was a sum. This resulted in a Fourier series
representation for the original signal. In the infinite duration DTFT case,
however, only time is discrete, the frequencies take values in a continuous
interval and the inverse transform is an integral, resulting in a Fourier inte-
gral representation of the signal. We can still interpret the representation
of the original signal as a weighted average of exponentials, but the average
is now an integral instead of a sum.
We have not really answered the question of how generally the result of
(3.39)-(3.40) is valid. We have given without proof a sufficient condition
under which it holds: if the original sequence is absolutely summable then
(3.39)-(3.40) are valid.
One might ask at this point what happens if one begins with a spectrum
{G (f); j E S} and defines the sequence via the inverse Fourier transform.
Under what conditions on G(f) will the formulas of (3.39)-(3.40) still hold;
that is, when will one be able to recover gn from G(f) using the given
formulas? This question may now appear academic, but it will shortly gain
in importance. Unfortunately we cannot give an easy answer, but we will
describe some fairly general analogous conditions when we treat the infinite
duration CTFT.
Analogous to the remarks following the DFT inversion formula, we could
also consider different frequency domains of definition, in particular any
unit length interval such as [0,1) would work, and we could consider dif-
ferent time domains of definition, such as the nonnegative integers which
3.3. INVERTING THE INFINITE DURATION DTFT 125

yield one-sided signals of the form {gn; n = 0,1" .. }. These alternatives

will not be considered in detail. Instead we consider another variation on
the choice of a frequency domain of definition and the resulting inversion
formula. This alternative will prove very useful in our subsequent consid-
eration of continuous time Fourier transforms.

Frequency Scaling
Instead of having a frequency with values in an interval of unit length, we

°
can scale the frequency by an arbitrary positive constant and adjust the
formulas accordingly. For example, we could fix 10 > and define a Fourier
transform as

and hence, changing variables of integration,

We now restate the Fourier transform pair relation with the scaled fre-
quency value and change the name of the signal and the spectrum in an
attempt to minimize confusion. Replace the sequence gn by h n and the
transform G 10 (I) by H (I). We have now proved that for a given sequence
h, the following are a Fourier transform pair (provided the technical as-
sumptions used in the proof hold):

L
00

H(I) = h ne- i27r fon; 1 E [0,10) (3.41)

n=-oo

(3.42)

The idea is that we can scale the frequency parameter and change its
range and the Fourier transform pair relation still holds with minor modifi-
cations. The most direct application of this form of transform is in sampled
data systems when one begins with a continuous time signal {g(t); t E 'R}
and forms a discrete time signal {g(nT); n E Z}. In this situation the
scaled frequency form of the Fourier transform is often used with a scaling
10 = lIT. More immediately, however, the scaled representation will prove
useful in the next case treated.
126 CHAPTER 3. FOURIER INVERSION

3.4 Inverting the CTFT

The inversion of discrete time Fourier transforms is both straightforward
and easy. In both cases the inversion formula follows from the orthogonality
of complex exponentials. For the finite duration case it is their orthogo-
nality using a sum over a set of integers. For the infinite duration case
it is their orthogonality using an integral over a unit interval. Although
the basic orthogonality idea extends to continuous time, the proofs become
significantly harder and, in fact, the details are beyond the prerequisites
assumed for this course. They are typically treated in a course on integra-
tion or real analysis. Hence we shall first summarize the inversion formulas
and conditions for their validity without proof, but we emphasize that they
are essentially just an extension of the discrete time formulas to continuous
time, that the orthogonality of exponentials is still the key idea, and that
the formulas are clear analogs to those for discrete time. In optional sub-
sections we present some "almost proofs" that, it is hoped, provide some
additional insight into the formulas.
Just stating the inversion theorems requires some additional ideas and
notation, which we now develop.

Piecewise Smooth Signals

Let 9 = {g(t)j t E T} be a continuous time signal. 9 is said to have a jump
discontinuity at a point t if the following two limits, called the upper limit
and lower limit, respectively, exist:

g(t+) ~ lim g(t + /ED (3.43)

<-+0

g(C) ~ lim g(t -

<-+0
lei). (3.44)

but are not equal. If the limits exist and equal 9 at t, g(t) = g(t+) = g(t-),
then g(t) is continuous at t.
A real valued signal {g(t)j t E T} is said to be piecewise continuous on
an interval (a, b) c T if it has only a finite number of jump discontinuities
in (a, b) and if the lower limit exists at b and the upper limit exists at a.
A real valued signal {g(t); t E T} is said to be piecewise smooth on an
interval (a,b) if its derivative dg(t)fdt is piecewise continuous on (a,b).
A real valued signal {g(t)j t E T} is said to be piecewise continuous
(piecewise smooth) if it is piecewise continuous (piecewise smooth) for all
intervals (a, b) c T. Piecewise smooth signals are a class of "nice" signals
for which an extension of Fourier inversion works.
One further detail is required before we can formally treat the inversion
of finite duration continuous time signals. Suppose that T = [0, T). What
3.4. INVERTING THE CTFT 127

if the discontinuity occurs at the origin where only the upper limit g(t+)
makes sense? (We do not need to worry about the point T because we have
purposefully excluded it from the time domain of definition [0, T).) We
somewhat arbitrarily redefine the lower limit of a signal defined on [0, T)
°
at by
(3.45)
that is, the limit of g(t) as t approaches T. This definition can be interpreted
as providing the ordinary lower limit for the periodic extension g(t) =
g(t mod T) of the finite duration signal. Alternatively, it can be considered
as satisfying the ordinary definition (3.44) if we interpret addition and
subtraction of time modulo T; that is, t - r means (t - r) mod T. This will
later be seen to be a reasonable interpretation when we consider shifts for
finite duration signals.
We can now state the inversion theorems for continuous time signals.
The reader should concentrate on their similarities to the discrete time
analogs for the moment, the chief difference being the special treatment
given to jump discontinuities in the signal.

Finite Duration Signals

Theorem 3.1 The Finite Duration Fourier Integral Theorem
Suppose that 9 = {g(t); t E [0, T)} is a finite duration continuous time
signal such that
• 9 is absolutely integrable; that is,
T
10 Ig(t)1 dt < 00, (3.46)

• 9 is piecewise smooth.
Define the Fourier transform by

G(f) = foT g(t)e-i27rft dt, f E {kiT; k E Z} = SgfFD' (3.47)

Then
L:
OO
G(-if)e i27r.!!.t
- T T
. (3.48)
n=-oo
If 9 is continuous at t, then

g(t) = f
n=-oo
G~) e i27r ¥<t. (3.49)
128 CHAPTER 3. FOURIER INVERSION

As it is somewhat awkward to always write the midpoints for possible

jump discontinuities, (3.48) is often abbreviated to

g(t) '" f 7)
n=-(X)
G ei21T ¥<t. (3.50)

This formula means that the two sides are equal at points t of continuity
of g(t), but that the more complicated formula (3.48) holds if g(t) has a
jump discontinuity at t. With this notation we can easily summarize the
theorem as stating that under suitable conditions, the following is a Fourier
transform pair:

G(f) = loT g(t)e- i21T / t dt, f E {kiT; k E Z}, (3.51)

g(t) '" f 7)
n=-(X)
G ei27r ¥<t; t E [0, T). (3.52)

As an example of the inversion formula for a finite duration continuous

time signal, consider the signal it; t E [0, I)}. From the inversion formula
and (2.32) we have that

t = !2 + "L.J _i_ei27Tkt. t E (0 1)
21rk ' ,. (3.53)
k~O.kEZ

Note that the spectrum {G(k); k E Z} is not absolutely summable, yet

it has an inverse transform!
The theorem remains true if the time domain of definition is replaced
by a two-sided index set of the same length, i.e., [- T /2, T /2). Thus if
9 = {g(t); t E [-T/2,T/2)} is absolutely integrable and piecewise smooth,
then we have the Fourier transform pair

G(f) = l:T

"'2
g(t)e- i21T / t dt, f E {kiT; k E Z}, (3.54)

g( t) '" ~ G(lf) i27r;"t. t [ T T)

L.J --y;-e , E -2'2 . (3.55)
n=-(X)

As an example of inversion of a two-sided signal, consider the two-sided

continuous time finite duration rectangle function {n(t); t E [-T/2, T/2)},
3.4. INVERTING THE CTFT 129

T > 1, which has discontinuities at ±1/2. From (2.35) the Fourier trans-
form for this signal is found by restricting the frequency domain to the
integer multiples of kiT, that is,

{net); t E [-T/2,T/2)}:J {sinc(;); k E z}.

Using the continuous time finite duration inversion formula then yields

~
n (t ) = L..J T1 smc
. ( k) i27rt.!L
T e T; [T T )
tE - 2' 2 . (3.56)
k=-oo

Since the rectangle function has values at discontinuities equal to the mid-
points of the upper and lower limits, the Fourier inversion works. Although
01/2(t) shares the same Fourier transform, (3.56) does not hold with 0 1 / 2
replacing net) because the right-hand side does not agree with the 01/2
signal at the discontinuities. There are two common ways to handle this
difficulty. The first is to modify all interesting signals with discontinuities
so that their values at the discontinuities are the midpoints. The second
is to simply realize that the Fourier inversion formula for continuous time
signals can only be trusted to hold at times where the signal is continuous.
This latter approach is accomplished by recalling the notation
00
""
0 1/ 2 (t ) '" L..J T1 smc
. ( k) i27rt
T e 'i'; k
tE
[T T)
-2'2' (3.57)
k=-oo

to denote the fact that the right hand side gives the left hand side only at
points of continuity. The right hand side gives the midpoints of the left
hand side's upper and lower limits at points of discontinuity.

Infinite Duration Signals

Next suppose that g = {get); t E 'T} is an infinite duration continuous
time signal. We will focus on the most important special case of an infinite
duration signal with 'T = n.

Theorem 3.2 The Fourier Integral Theorem

Given a continuous time infinite duration signal g = {get); tEn},
suppose that g is absolutely integrable and piecewise smooth. Define the

1:
Fourier transform

G(f) = g(t)e- i27rt / dt; / E n. (3.58)

130 CHAPTER 3. FOURIER INVERSION

Then

1 -00
00 G(f) e i2wlt dlf = {g(t)
g(t+)+g(C)
2
if t is a point of continuity
otherwise.
(3.59)

Note the strong resemblance of the inverse transform to that of the

discrete time infinite duration case. In both cases the inversion formula
is an integral of the Fourier transform times an exponential having the
opposite sign than that used to define the transform. The only difference is
that here the integral is over all f E (-00,00) instead of over [-1/2,1/2).
Note also the resemblance of the inverse CTFT to the CTFT itself-only
the sign of the exponential has been changed.
Again we summarize the theorem by saying simply that subject to suit-
able conditions, the following is a Fourier transform pair:

= I:
I:
G(f) g(t)e- i21f / t dt; fER, (3.60)

g(t) G(f)e i21f / t ; t E Rd!. (3.61)

As an example of the inversion formula for an infinite duration contin-

uous time signal, we have using Eq. 2.15 that

e-tH(t) = J
oo

-00
1
ei21f/t

+
2
7n
'f df; t E R. (3.62)

This is a Fourier integral representation of the given signal.

The following subsections discuss these inversion formulas in some depth.

* Inverting the Finite Duration CTFT

We begin with an educated guess on how to recover a finite duration contin-
uous time signal 9 = {g(t); t E [0, Tn
from its Fourier transform. Observe
that we now have a finite duration continuous time signal and take an in-
tegral transform thereof. This is analogous to the earlier case where we
had a finite duration continuous frequency spectrum and took an inverse
Fourier transform that was an integral transform. Since there is nothing
magic about the interpretation oftime and frequency (mathematically they
are just independent variables), we can view this situation as a dual of the
previous case and guess that our transform will have discrete parameters
(previously time, now frequency) and that the Fourier transform pair will
look like (3.41)-(3.42) except that the roles of time and frequency will be
reversed. One slight additional change is that the signs of the exponen-
tials have also been reversed since we are switching the Fourier transform
3.4. INVERTING THE CTFT 131

with its inverse. Putting this together yields the guess that the appropriate
Fourier transform pair is given by

= loT g(t)e- i21r -f t dtj k E Z (3.63)

get) = f
n=-oo
G;) ei21r~tj t E [0, T). (3.64)

Note that as in the DFT inversion formula, we again have the original
signal represented as a Fourier series. It is not expected that the above pair
should be obvious given the corresponding result for the DTFT, only that
it should be a plausible guess. We now prove that it is in fact equivalent
to the DTFT result. To do this we make the substitutions of Table 3.1 in
the scaled-frequency DTFT Fourier transform pair.

Eq.3.42 Eq.3.47

f (frequency) t (time)
H(f) ( spectrum) get) (signal)
fo T
h n (signal) ~G{-j.) (spectrum)

Table 3.1: Duality of the Infinite Duration DTFT and the Finite Duration
CTFT

With these definitions (3.41) becomes

get) = f:
n=-oo
G(;j.) e- i21r "t.
Changing the summation dummy variable sign yields

Similarly (3.42) becomes

G{;j.) = loT g~) ei21r i<n dt.

132 CHAPTER 3. FOURIER INVERSION

Thus evaluating the formula at -n yields

G(!!:.)
T
= r g(t)e-i27rlTt dt,
io
T

which shows that (3.63)-(3.64) follow from Eqs. (3.41)-(3.42).

As with the DTFT we have swept some critical details under the rug.
The Fourier transform pair formula does not hold for all possible finite du-
ration continuous time signals g(t)j they must be "nice" enough for the
arguments used in the proof (the exchange of sums and integrals, for ex-
ample) to hold. Sufficient conditions for (3.63)-(3.64) to hold were first
developed by Dirichlet and were named after him. They strongly resemble
those that will be encountered in the next section on the infinite duration
CTFT. While the proof of these conditions is beyond the assumed prereq-
uisites of this course, we can describe sufficient conditions for the theorem
to hold and sketch some of the ideas in the proof.
Roughly speaking, the sufficient conditions for the inversion formula
(3.64) to hold are threefold:
1. g(t) is absolutely integrable,

2. g(t) cannot "wiggle" too much in a way made precise below, and

3. g(t) is a continuous function of t.

The continuity condition can be dropped at the expense of complicating
the inversion formula.
The condition that we have adopted to constrain the wiggliness of 9 is
that it be piecewise smooth. Other, similar, conditions are also seen in the
literature. Perhaps the most common is that of bounded variation, that
9 can have at most a finite number of maxima and minima in any finite
length interval.
In order to sketch a partial proof of the theorem that will strongly
resemble the continuous time infinite duration, we consider the behavior of
the partial sums

gN(t) = t
n=-N
G~) e i27rlTt
as N ~ 00. We have that

T1 inr 2::
T N. n
g(r)( e,27r(t-r)'1') dr. (3.65)
o n=-N
3.4. INVERTING THE CTFT 133

The internal sum can be evaluated using (2.29) as

(3.66)

so that
()-iT (
gN t - o gT
)sin(2IT U¥)(N + ~)) d
T' ( (t T))
sm IT T
T. (3.67)

What happens as N -+ oo? The term multiplying the signal inside the
integral can be expressed in terms of the Dirichlet kernel defined by

( ) _ sin(ITt(N + ~)) [_~ ~) (3.68)

DN t - sin (ITt ) , t E 2' 2 '

which should be familiar if we change the t to f as the DTFT of the box

function {ON(n); n E {-N, ... , N}}. The Dirichlet kernel is symmetric
about the origin and has unit integral. To see this latter fact, observe that
the Fourier inversion for the DTFT of the box function evaluated at n = 0
yields
(3.69)

(This is an example of a moment property as will be considered in Chapter

4.) We can now express (3.67) as

gN(t)
r g(T)f1 DN ('i')
= 10
T t-
dT.
T
(3.70)

We point out in passing that the right hand integral is an example of

a convolution integral, an operation that we will consider in some detail
in Chapter 6. DN(t) is plotted for three values of N in Figure 3.1. As
N increases, the signal becomes more and more concentrated around the
origin. Its height at the origin grows without bound and its total area
of 1 is in an increasingly narrow region around the origin. The signal
T-lDNe-~/); t E [-i, i) will behave similarly, except that its peak will
be at t = T. Multiplying T- 1 DN(t-:/) by geT) and integrating with N large
will result in very little contribution being made to the integral except for
T ~ t. If get) is continuous at t, the fact that the integral of T- 1 D N (t l .r)
equals 1 implies that the integral of (3.70) will be approximately get). If
get) has a jump discontinuity, then the approximation will be replaced by
1/2 the upper limit plus 1/2 the lower limit, yielding the midpoint of the
upper and lower limits.
134 CHAPTER 3. FOURIER INVERSION

10,---,----,---,----,---,---,----,---,---,,---,

o ............... ~\:.:
: :,

.'. .... -.-

-2

-4~--~--~--~--~----~--~--~--~--~--~
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5

Figure 3.1: The Dirichlet Kernel DN(t): N=1 (solid line), 2 (dashed line),
4 (dash-dot line)

Making this line of argument precise, that is, actually proving that
gN(t) -t g(t) as N -t 00 would require the Riemann-Lebesgue Lemma, one
of the fundamental results of Fourier series. Alternatively, it is proving that
the limit of the Dirichlet kernels behave like a generalized function - the
Dirac delta to be considered in Chapter 5. We here will content ourselves
with the above intuitive argument, which is reinforced by the similarity of
the result to the discrete time inversion formulas, which were proved in
some detail.
The inversion formula can be extended from absolutely integrable sig-
nals to finite energy signals by considering the infinite sum to be a limit
in the mean. In particular, it can be shown that if the signal 9 has finite
energy, then it is true that

(3.71)

Note that this form of inversion says nothing about pointwise convergence
of the sum to g(t) for a particular value of t and is not affected by the
values of g(t) at simple isolated discontinuities.
3.4. INVERTING THE CTFT 135

* Inverting the Infinite Duration CTFT

Next consider a continuous time signal g = {g(t); E R} with Fourier

I:
transform G = F(g) defined by

G(f) = g(t)e-i27rtf dt (3.72)

for all real f.

We now sketch a proof of the Fourier integral theorem in a way re-
sembling that for the finite duration continuous time inversion formula.
Consider the function

(3.73)

formed by truncating the region of integration in the inverse Fourier trans-

form in a symmetric fashion. We will argue that under the assumed condi-
tions,
r (t) _ g(t+) + g(C) .
a~~ga - 2 '

I:
that is, that the integral

G(f)ei27rft df

exists (in a Cauchy principal value sense) and has the desired value. As in
the DFT and DTFT case, we begin by inserting the formula for G(f) into
the (truncated) inversion formula:

ga(t) = faa ( [ : g(x)e-i27rfx dX) ei27rft df. (3.74)

If the function is suitably well behaved, we can change the order of inte-
gration:
ga(t) = [ : dxg(x) ([: ei27rf (t-x) df ) (3.75)

The integral in parentheses is easily evaluated as

ei27r f(t-x) sin(27ra( t - x))
i27r(t - x) I~a 7r(t-x)
2asinc2a(t - x)
sin 2a(t - x)7r
= (3.76)
(t - x)7r
= Fa(t - x), (3.77)
136 CHAPTER 3. FOURIER INVERSION

where
. sin2at1l'
Fa(t) = 2asmc2at =- -
t1l'
- (3.78)
is called the Fourier integral kernel and is simply the CTFT of the box
function {DT(t)j t E 'R.} with t substituted for f. Like the Dirichlet kernel

ga(t) i:
it is symmetric about the origin and has unit integral. Thus

= g(x)Fa{t - x) dx.

Once again we observe that the right hand integral is an example of a

(3.79)

convolution integral. Figure 3.2 shows the Fourier integral kernel for several
values of a. As with the Dirichlet kernel encountered in the finite duration
continuous time case, the sinc functions become increasingly concentrated
around their center as a becomes large. Thus essentially the same argument

Figure 3.2: The Fourier Kernel Fa(t): a=l (solid line), 2 (dashed line), 4
(dash-dot line)

yields the inversion formula as in the finite duration case. The Riemann-
Lebesgue Lemma implies that the limit of this function as a -+ 00 is as
stated in the theorem under the given conditions.
We also consider an alternative intuitive "proof" of the Fourier integral
theorem for the special case where the signal is continuous. The proof is a
3.5. CONTINUOUS TIME FOURIER SERIES 137

simple limiting argument based on the finite duration inversion of (3.54)-

(3.55). Unfortunately the proof cannot easily be made rigorous and hence
is presented only as an interpretation.
Suppose that in (3.54)-(3.55) we let T ~ 00 in order to extend the
Fourier transform pair to infinite duration continuous time signals. In order
to do this define a frequency variable f = niT. With these definitions (3.54)
becomes in the limit

G(f) = lim
T~oo
12-t
T
g(t)e-i2trft dt = 1
00

-00
g(t)e-i2trft dt,

the usual CTFT, and (3.55) can be considered as a Riemann sum (liT
becomes df) which in the limit as T ~ 00 approximates the integral
get) = lim gT(t)
T~oo

= lim ~ G(~) ei2tr~t

T~oo ~ T
k=-oo

= [: G(f)ei2trft df, (3.80)

which is the claimed inverse transform. This "proof" at least points out
that the form of the infinite duration CTFT is consistent with that for the
finite duration CTFT. Thus the Fourier integral transform relations can be
viewed as a limiting form of the Fourier series relations for continuous time
signals. This is not, however, the way that such results are properly proved.
This is not a rigorous proof because the Riemann integrals are improper
(have infinite limits) and there is no guarantee that the Riemann sums will
converge to the integrals. Furthermore, the improper Riemann integrals in
the argument have only been considered in a Cauchy principal value sense.

3.5 Continuous Time Fourier Series

As in the discrete time finite duration case, the Fourier inversion formula
provides a representation for continuous time finite duration signals as a
weighted sum of complex exponentials, a Fourier series. For simplicity we
now assume that get) is absolutely integrable and piecewise smooth. The
Fourier transform pair of (3.63)-(3.64) can also be written in the form

(3.81)

(3.82)
138 CHAPTER 3. FOURIER INVERSION

As usual, the'" notation means equality if g(t) is continuous at t and the

midpoint of the upper and lower limits otherwise. This is a very common
form for a Fourier series of a finite duration continuous time signal anal-
ogous to the discrete time Fourier series of (3.26)-(3.27). Note that all
we have done is replace G{f)/T by Ck (which are also called Fourier co-
efficients). In some applications the exponential is expanded using Euler's
formulas to form the trigonometric series

(3.83)

The ak and bk can be determined from the Ck and vice-versa. The details
are left as an exercise.
As yet another form of this relation, consider the case where T =
[-T/2,T/2} and replace Ck by dk/T:

g(t) f:
k=-oo
~ ei2 11'tk; t E [-T/2, T/2) (3.84)

dk = 1:T

-~
g(t)e- i2 11't k dt; k E Z. (3.85)

As a final observation, as with the inversion of the finite duration dis-

crete time signal, the inversion formula for a finite duration continuous
time signal is perfectly well defined for all real t and not just for the origi-
nal t E [0, T). Also, as in the DFT case, the formula is periodic, now having
period T. Thus we automatically have a Fourier series representation for
the periodic extension 9 of g:

L
00

g{t) '" Ck ei2 11'+k; tEn, (3.86)

k=-oo

(3.87)

Alternatively, given a continuous time periodic signal 9 we can find a Fourier

series as above using the Fourier transform of the truncated finite dura-
tion signal consisting of only a single period. As an example, consider
the periodic signal {g(t) = t mod 1; tEn}. This has the Fourier series
representation
L
00

g(t) '" Ck ei2 11'+k

k=-oo
3.5. CONTINUOUS TIME FOURIER SERIES 139

with T = 1 and Co = 1/2 and Ck = i/2'1rk for k i- O. Thus at all points of

continuity of t mod 1 (all points except those where t is an integer) we have

t mod 1 = ~ + " _i_ei27rtk. tER t d Z (3.88)

2 L..J 2'1rk ' ,l'"
k#O,kEZ

From the continuous time finite duration inversion formula this series will
have a value of 1/2 at the discontinuities which occur at integer t.
This series can be put in a somewhat simpler form by observing that
the sum can be rewritten as

f _i_(ei27rtk - e-i27rtk)
k=l 2'1rk
=- f
k=l
sin(2'1rkt)
'Irk

yielding the Fourier series

1 1 ~ sin(2'1rkt)
t mo d rv 2- L..J 'Irk . (3.89)
k=l

A simple application of Fourier series yields a famous equality involving

Bessel functions that crops up regularly in the study of nonlinear systems.
Consider the signal
g( t/J) = ei % sin tP

for real t/J. The quantity z can be considered as a parameter. This signal
is periodic in t/J with period 2'1r and hence can be expanded into a Fourier
series
L L
00 00

e i % sin tP = cke i27r f,;k = CkeitPk

k=-oo k=-oo

where
Ck = -1 127r , '.1. '.I'k
el%Sln "'e- ' ''' dt/J = Jk(Z), (3.90)
2'1r 0

the kth order Bessel function of (1.12). Thus we have the expansion

L
00

ei%sin tP = Jk(z)e itPk , (3.91)

k=-oo

a result called the Jacobi-Anger expansion.

140 CHAPTER 3. FOURIER INVERSION

* Continuous time Fourier Series and the DFT

We previously saw that the OFT could be used to approximate the Fourier
transform of a continuous time signal. We here carry this discussion a bit
further to point out that the OFT can be used to approximate the Fourier
series for a continuous time function.
Suppose that {g(t)j t E [0, T)} is a continuous time finite duration
signal with Fourier transform G and hence that we have a Fourier series
representation
~
g(t) = T1 ~
k
G(f)e'
'2 lLt
7rT •
k=-oo

We can approximate the spectrum by the OFT of a sampled waveform

{al = g( iT / N) j i = 0, ... ,N - I} using Riemann sum approximations as

where G is the OFT of the sequence {alj i E ZN}, Approximating the

remaining spectral terms by 0, we have the approximation

(3.92)

which should be increasingly accurate as N becomes large. This is the

way in which a digital computer approximately computes the Fourier series
of a continuous time signal. (It must also digitize the samples, but that
consideration is beyond the scope of this discussion.)

3.6 Duality
When inferring the inversion formula for the finite duration continuous time
Fourier transform from that for the infinite duration discrete time Fourier
transform, we often mentioned the duality of the two cases: interchang-
ing the role of time and frequency turned one Fourier transform pair into
another. We now consider this idea more carefully.
3.6. DUALITY 141

Knowing one transform often easily gives the result of a seemingly dif-
ferent transform. In fact, we have already taken advantage of this duality
in proving the finite duration continuous time Fourier inversion formula by
viewing it as a dual result to the infinite duration discrete time inversion
result.
Suppose, for example, you know that the Fourier transform of an infinite
duration continuous time signal g = {get); t E 'R} is G(f), e.g., we found
that the transform of get) = e- t H(t) is G(f) = (1 + i27r1)-1. Now suppose
that you are asked to find the Fourier transform of the continuous time
signal ret) = (1 + i27rt)-I. This is easily done by noting that ret) = G(t);
that is, r has the same functional dependence on its argument that G
has. We know, however, that the inverse Fourier transform of G(f) is get)
and that the inverse Fourier transform in the infinite duration continuous
time case is identical to a Fourier transform except for the sign of the
exponent. Putting these facts together we know that the Fourier transform
of ret) = G(t) will be g( - I). The details might add some insight: Given
get) and G(f), then the inversion formula says that

get) = J00

G(f)ei21rlt df
-00

= J00

-00
G(a)ei21rQt da,

where the name of the dummy variable is changed to help minimize confu-
sion when interchanging the roles of f and t. Now if ret) = G(t), its Fourier
transform is found using the previous formula to be

J J
00 00

R(f) = r(t)e-i21rtl dt = G(t)e-i21rtl dt

-00 -00

= J00

-00
G(a)e i21rQ (-f)da=g(_I).

To summarize:

If {get); t E 'R} is an infinite duration continuous time signal

with Fourier transform

:r,( {get); t E 'R}) = G(f); f E 'R,

142 CHAPTER 3. FOURIER INVERSION

then the Fourier transform of the infinite duration continuous

time signal {G(t)j tEn} is

Ff({G(t)j tEn})=g(-f)j fEn.

Thus for example, since

then also
(3.93)

Similarly,
Ff({sinc(t)j t E 'R}) = n(f); f E 'R. (3.94)
Almost the same thing happens in the DFT case because both trans-
forms have the same formj that is, they are both sums. Thus with suitable
scaling, every DFT transform pair has its dual pair formed by properly
reversing the role of time and frequency as above.
The finite duration continuous time and infinite duration discrete time
are a little more complicated because the transforms and inverses are not
of the same form: one is a sum and the other an integral. Note, however,
that every transform pair for a finite duration continuous time signal has as
its dual a transform pair for an infinite duration discrete time signal given
appropriate scaling. Suppose, for example, that we have an infinite duration
discrete time signal {gn; n E Z} with Fourier transform {G(f); f E [0, I)}

11
and hence
gn = G(f)ei21rfn df. (3.95)

As a specific case, suppose that gn = r nU_l(n); n E Z and hence G(f) =

(l-r- i21rf )-1; f E [0,1). Suppose that {x(t)j t E [0, T)} is a finite duration
continuous time signal defined by

t
x(t) = G(f)jt E [O,T)

(where now the scaling is needed to permit a time domain more general
than [0,1)). In the specific example the waveform becomes
3.6. DUALITY 143

What is the Fourier transform XU) of x? Since the transform of a finite

duration continuous time signal has a discrete frequency domain, we need
to find

X(~) = loT x(t)e-i21Tft dt

= loT x(a)e- i21Tfa da

= {T G( ~ )e-i21Tfa da
10 T
= T 10 1 G({3)e- i2 11"k{3 d{3,

where we have made the variable substitution {3 = a/To Combining this

equation with (3.95) we have the duality result

which also has the flavor of the infinite duration continuous time result
except that the additional scaling is required. We can summarize this
duality result as follows.

If {gnj n E Z} is an infinite duration discrete time signal with

Fourier transform

F,( {gnj n E Z}) = GU)j f E [0,1),

then the Fourier transform of the finite duration continuous time

signal {G(t/T)j t E [0, T)} is

Fk/T({G(t/T)j t E [O,T)}) = Tg-kj k E Z.

In our example this becomes

k = 0, -1, -2,···
otherwise.

These simple tricks allow every transform pair we develop to play a

double role.
144 CHAPTER 3. FOURIER INVERSION

3.7 Summary
The Fourier transform pairs for the four cases considered are summarized in
Table 3.2. The'" notation in the continuous time inversion formulas can be
changed to = if the signals are continuous at t. Note in particular the ranges
of the time and frequency variables for the four cases: both are discrete in
the finite duration discrete time case (the DFT) and both are continuous in
the infinite duration continuous time case. In the remaining two cases one
parameter is continuous and the other discrete and these two cases can be
viewed as duals in the sense that the roles of time and frequency have been
reversed. These results assume that the transforms and inverses exist and
that the functions of continuous parameters are piecewise continuous. Note
also that the finite duration inversion formulas both involve a normalization
by the length of the duration, a normalization not required in the infinite
duration formulas. In many treatments, this normalization constant is di-
vided between the Fourier and inverse Fourier transforms to make them
more symmetric, e.g., both transform and inverse transform incorporate a
normalization factor of l/vN (discrete time) or l/VT (continuous time).
Such changes of the definitions by a constant do not affect any of the the-
ory, but one should be consistent. If one uses a stretched frequency scale
(instead of [0, 1) or [-1/2,1/2» for the infinite duration discrete time case,
then one needs a scaling of 1/8 in the inversion formula, where 8 is the
length of the frequency domain.
It is also informative to consider a similar table describing the nature
of the frequency domains for the various signal types. This is done in
Table 3.3. The focus is on the two attributes of the frequency domain S. As
with the time domain r, we have seen that the frequency can be discrete
or continuous. The table points out that also like the time domain, the
frequency domain can be "finite duration" or "infinite duration" in the sense
of being defined for a time interval of finite or infinite length. We dub these
cases finite bandwidth and infinite bandwidth and observe that continuous
time signals yield infinite bandwidth frequency domains and discrete time
signals yield finite bandwidth frequency domains. These observations add
to the duality of the time and frequency domains and between signals and
spectra: as there are four basic signal types, there are also four basic spectra
types. Lastly observe that discrete time finite duration signals yield discrete
frequency finite bandwidth spectra and continuous time infinite duration
signals yield continuous frequency infinite bandwidth spectra. Thus in these
cases the behavior of the time and frequency domains are the same. On the
other hand, continuous time finite duration signals yield discrete frequency
infinite bandwidth spectra and discrete time infinite duration signals yield
continuous frequency finite bandwidth spectra. In these cases the time and
3.B. * ORTHONORMAL BASES 145

frequency domain behaviors are reversed.

Duration Time
Discrete Continuous

N-l T
Finite G(-M= L: gne-Z-2 1I"Nk n ; k E ZN G(~) = J g(t)e-i21r~t dt; k EZ
n=O 0
N-l
= -k L: -k
00
gn
k=O
G( )ei211"~n; n E Zj.f g( t) '" L: G<t)ei211"~t; t E [O,T)
k=-oo

G(f) = J g(t)e- i2 11"tf dt; fEn

00 00

Infinite G(f) = L: gne-i211"fn; f E [0, S]

n=-oo -00
S
= J ~G(f)ei211"fn df; J G(f)e i2 11"ft df;
00
gn nEZ get) '" tEn
0 -00

Table 3.2: Fourier Transform Pairs

As with the terminology "finite duration," there is potential confusion

over the name "finite bandwidth" which could describe either a finite length
domain of definition for the spectrum (as above) or a spectrum with infinite
length domain of definition which just happens to be zero outside a finite
length interval, e.g., a spectrum of an infinite duration signal which is zero
outside of [-W, W]. We will refer to the latter case as a "band-limited"
spectrum and the corresponding signal as a band-limited signal. This is
analogous to our use of the term "time-limited" for infinite duration signals
which are zero outside of some finite interval.

Duration Time
Discrete Continuous
Finite Discrete Frequency Discrete Frequency
Finite Bandwidth Infinite Bandwidth
Infinite Continuous Frequency Continuous Frequency
Finite Bandwidth Infinite Bandwidth

Table 3.3: Frequency Domain

3.8 * Orthonormal Bases

Many basic properties of the Fourier transform can be generalized by replac-
ing complex exponentials by more general functions with similar properties.
146 CHAPTER 3. FOURIER INVERSION

We demonstrate this for the special case of discrete time finite duration sig-
nals. The case for continuous time finite duration signals is considered in
the exercises and similar ideas extend to infinite duration signals. In the
next section we consider an important special case, the discrete wavelet
transform.
We here confine interest to signals of the form 9 = {g(n); n = 0,1, ... , N-
I}, where N = 2£ for some integer L. Let ON denote the collection of all
such signals, that is, the space of all real valued discrete time signals of
duration N.
A collection of signals 'lfJk = {'l/Jk(n)j n = 0,1, ... , N -I}; k = 0,1, ... , K-
1, is said to form an orthonormal basis for the space ON if (1) the signals
are orthonormal in the sense that
N-l
L 'lfJk(n)'l/Jt(n) = 8k- l , (3.96)
n-O

and (2), the set of signals is complete in the sense that any signal 9 E 0
can be expressed in the form
K-l
g(n) = L ak'lfJk(n). (3.97)
k=O

It follows from (3.9) that the discrete time exponentials

e i27r -/!;n
'lfJk = { -IN ; n E ZN}, k = 0, ... , N - 1, (3.98)

are orthonormal and the DFT inversion formula (3.14) guarantees that any
signal 9 E ON can be written as a weighted sum of the 'lfJk and hence the
complex exponentials indeed form an orthonormal basis. Thus the idea
of an orthonormal basis can be viewed as a generalization of the Fourier
transform and its inverse.
The smallest integer K for which there exists an orthonormal basis for a
space is called the dimension of the space. In the case of ON the dimension
is N. While we will not actually prove this, it should be believable since
the Fourier example proves the dimension is not more than N.
The general case mimics the Fourier example in another way. If we wish
to compute the linear weights ak in the expansion, observe that

N-l N-l N-l

L'lfJt"(n)g(n) L 'lfJt"(n) L ak'lfJk(n)
n=O n=O k=O
3.9. * DISCRETE TIME WAVELET TRANSFORMS 147

N-l N-l

= L ak L 1/Ji(n)1/Jk(n)
k=O n=O
N-l

= L aktSl-k = al
k=O
using the orthonormality. Thus the "coefficients" ak are calculated in gen-
eral in the same way as in the Fourier special case: multiply the signal by
1/Ji and sum over time to get al. This is sometimes abbreviated using the
inner product or scalar product notation

N-l
< g,1/Jk >~ L g(n)1/J,i;(n) (3.99)
n=O

and we can write the general expansion as

K-l
g(n) = L < g,1/Jk > 1/Jk(n). (3.100)
k=O
The inner product is often denoted by (g,1/Jk) as well as < g, 1/Jk >.
The complex exponentials are not the only possible basis. A simpler
basis is given by the shifted Kronecker delta functions 1/Jk(n) = tSn - k .
Whenever we have an orthonormal basis we can think of the resulting
coefficients {ak; k = 0, ... , N - I} as forming the transform of the signal
with respect to the basis, and (3.100) provides the inverse transform. Eq.
(3.100) is sometimes called a generalized Fourier series.

3.9 * Discrete Time Wavelet Transforms

In this section we consider an alternative transform that has gained wide
popularity in a variety of fields during recent years - the wavelet trans-
form. For simplicity we consider only the special case of discrete time finite
duration signals and we only provide a superficial treatment showing some
of the similarities and differences with Fourier transforms. The interested
reader can pursue the subject in detail in Daubechies [14], Rioul and Vet-
terli [29], Strang [31], or the references cited in these surveys.
A wavelet transform is a form of orthonormal transform with specific
properties. In fact, it need not even be orthonormal, but we will only
consider this important special case. A key idea is to construct all of the
basis functions from a single continuous time signal called a wavelet. Let
148 CHAPTER 3. FOURIER INVERSION

t/J(t)j t E R be a continuous time signal and consider the collection of

N = 2L discrete time signals defined in terms of t/J as follows:
t/Jo,o(n) = 2-1- j n E ZN
t/Jm,k(n) = T~t/J(2-mn - k)j n E ZNj (3.101)
m = 1, ... , Lj k = 0, 1, ... , 2L - m - 1.

Let K denote the set of all possible indices (m, k) specified in the above
collection.
With the exception of the (0,0) signal, these signals are all formed
by dilations and shifts of the basic function t/J. m is called the dilation
parameter and k is called the shift parameter. We say that t/J(t) is a wavelet
or mother wavelet if the collection {t/Jm,kj (m, k) E K} form an orthonormal
basis for gN. (This is not the usual definition, but it will suit our purposes.)
If this is the case, then we can expand any signal g E gN in a series of the
form
g(n) = L
am,kt/Jm,k(n), (3.102)
(m,k}EK:

where am,k =< g, tPm,k >.

The first question is whether or not the definition makes sense, i.e., if
functions tP having this property exist. This is easily demonstrated by a
basic example. Consider the function

1 0 < t < 1/2

t/J(t)= { -1 1/2~t<1
°
(3.103)
otherwise
depicted in Figure 3.3
In this case tP is called the Haar wavelet, but the function and its basic
properties go back to the early part of this century, long before it and
its properties were unified with many previously disparate techniques and
applications under the general topic of wavelets. For the special case where
L = 3 and N = 2L = 8, we have the following functions:
1
tP1,0 = .;2 (1, -1,0,0,0,0,0,0)
1
1/;1,1 = .;2(0,0,1, -1,0,0,0,0)
1
1/;1,2 = .;2(0,0,0,0,1, -1,0,0)
1
tP1,3 = .;2(0,0,0,0,0,0,1,-1)
3.9. * DISCRETE TIME WAVELET TRANSFORMS 149

g(t)

°
1
:2 1
-1

Figure 3.3: The Haar Wavelet

1
'l/J2,O = J4(I, 1, -1, -1,0,0,0,0)
1
'l/J2,1 = J4(0, 0, 0, 0,1,1, -1, -1)
1
'l/J3,o = y'8(1, 1, 1, 1, -1, -1, -1, -1)
1
'l/Jo,o = y'8(1,1,1,1,1,1,1,1).

These signals are easily seen to be orthogonal. The fact that they form a
basis is less obvious, but it follows from the fact that the space is known
to have dimension N and hence N orthonormal signals must form a basis.
It also follows from the fact that we can write every shifted delta function
{8 n -k; n = 0,1, ... , N -I}, k = 0,1, ... , N - 1, as a linear combination of
the 'l/Jm,k and hence since the shifted deltas form a basis, so do the 'l/Jm,k.
For example,

To find relations for the shifted delta functions 8j = {8 n - j; n E ZN} for

150 CHAPTER 3. FOURIER INVERSION

j E ZN, where as usual the shift is modulo N, verify the following formula:

= L 'I/1m,k(j)'I/1m,k(n); j,n E ZN. (3.105)

(m,k)eJC
Why is the wavelet transform special? There are many reasons, among
which is the fact that it provides a convenient multiresolution reconstruction
of the original signal. A Fourier series representation gives an increasingly
accurate approximation to the original signal as one adds up the terms
in the sum. A wavelet reconstruction does this in a particularly useful
way, providing progressively better approximations. The lowest resolution
is achieved by including only the 'I/1m,k with large m in the reconstruction.
As the lower m are added, the reconstructions becomes increasingly good.
The wavelet transform is also of interest because many of the signals used
in the decomposition have short duration in our example and in many other
examples. This is in obvious contrast to the complex exponentials, which
are non zero over almost the entire time domain. This means that errors
in wavelet coefficients will usually have effects that are highly localized and
not spread out over the full domain.
Wavelets also have the big advantage of simple and fast algorithms
for their computation. In the simple example considered, the coefficients
for a signal 9 = (90,91,· .. ,99) are found as follows. First compute the
coefficients a1,k =< 9, 'I/11,k > as

a1,0 = < 9, '1/11,0 >= 90 - 91

a1,1 = < 9, '1/11,1 >= 92 - 93
a1,2 < 9, '1/11,2 >= 94 - g5
a1,3 = < g, '1/11,3 >= g6 - g7·

These coefficients can be found by taking successive differences gn - gn-1 in

the signal and then looking at every other difference, that is, subsampling
the differences by 2. Also compute another set of auxiliary coefficients
which will be used to find ao,o, but in the mean time will have other uses.
Define

b1,0 = go + gl
b1,1 = g2 + g3

b1.2 = g4 + g5
3.9. * DISCRETE TIME WAVELET TRANSFORMS 151

bl ,3 = 96 + 97·

Thus the b1 ,k sequence replaces the differences used for a1,k by sums. Note
for later that summing up the b1 ,k will give ao,o.
Next we wish to compute the coefficients a2,k =< 9, 'I/J2,k > as

a2,0 = < 9, 'I/J2,0 >= 90 + 91 - 92 - 93

a2,1 = < 9,'l/J2,1 >= 94 + 95 - 96 - 97,

but we can use the computations already done to assist this. These coeffi-
cients can be found by forming differences (as we used to find the al,k on
the auxiliary bl,k; i.e., form

a2,0 = bl,o - bl,l

a2,1 = b1,2 - b1,3.

We can also form the auxiliary sequence as before by replacing these dif-
ferences by sums.

b2,0 = bl,o + bl,l

b2,1 = bl ,2 + bl ,3.
To finish, we now note that the coefficient a3,0 can be evaluated as

a3,0 = < 9,'l/J3,0 >

= 90 + 91 + 92 + 93 - 94 - 95 - 96 - 97
= b2,0 - b2,1

and ao,o as

ao,o = < 9,'l/Jo,0 >

= 90 + 91 + 92 + 93 + 94 + 95 + 96 + 97
= b2,0 + b2,1.
Thus by an iteration involving separate pairwise sums and differences, the
coefficients are built up in a sequence of simple combinations. Each of
these operations can be viewed as a linear filtering, as will be considered
in Chapter 6 and the overall operation can be constructed using a cascade
of filters and subsampling operations, a form of subband filterin9. The
interested reader can pursue the subject of subband filtering in Woods [37).
152 CHAPTER 3. FOURIER INVERSION

3.10 * Two-Dimensional Inversion

The basic ideas of Fourier inversion can be extended to two-dimensional
transforms in the obvious way, e.g., for the continuous parameter infinite

i: i:
duration case we have the Fourier transform pair

G(fx,fy) = g{x,y)e- i21r (fxz+fyy) dxdyj Ix E n,fy En,

(3.106)

g{x,y) = i: i: G(fx'/y)e+ i21r (fxz+fYY) dlx dfyj x E n, yEn,

(3.107)
provided that the signal is sufficiently well behaved.
Analogous to the one-dimensional case, the Fourier transform can be
thought of as decomposing the signal g(x, y) into elementary functions of
the form e+ i21r (fxx+fyy) , that is, two dimensional complex exponentials.
To be more concrete, consider the real part of the elementary functionsj
that is, consider the signals cos(211'(fxx + fyy)). For fixed y, this signal is
periodic in x with period 1/ Ix. For fixed x it is periodic in y with period
1/ fy. The signal will have 0 phase (mod 211') when the argument satisfies
Ixx + fyy = kj k E Z.

These formulas define straight lines in the X - Y plane.

2D signals that are separable in rectangular coordinates are easily in-
verted; the inverse is simply the product of the two one dimensional inverses
for each coordinate. Signals that are separable in polar coordinates take
a bit more effort, but it can be shown that we can invert the zero-order
Hankel transform or the Fourier-Bessel transform

J
00

G{p, ¢) = 211' rgR(r)Jo(211'rp) dr = G{p)

o

by

gR(r) = 211' 10 00
pG(p)JO(211'Tp) dp

at all points of continuity of gR. Thus the transform and inverse transform
operations are identical in the circularly symmetric case.
3.11. PROBLEMS 153

3.11 Problems
3.1. Given a signal 9 = {g(t); t E T}, what is .1'(.1'(g))? What is .1'(.1'(.1'(g)))?

3.2. Prove the two-sided discrete time finite duration Fourier inversion
formula (3.18).

3.3. Recall that the DCT of a signal {g(k,j); k = O,I, ... ,N -1; j =
0,1, ... , N - I} is defined by

G(l , m ) = !C(l)C( ) ~ ~1 (k') (2k + I)l1l" (2j + I)m1l"

N m L..J L..J 9 , J cos 2N cos 2N '
k=O j=O
(3.108)
where
C(n)
1
= {~
if n =
otherwise.
°
Show that the inverse DCT is given by

9 (k , J
.) =~ ~ ~ C(l)C( m
N L..J L..J
)G(l )
, m cos
(2k + 1)111"
2N cos
(2j + I)m1l"
2N .
1=0 m=O
(3.109)
Warning: This problem takes some hacking, but it is an important
result for engineering practice.

3.4. Find a Fourier series for the two-sided discrete time signal {r- 1nl ; n =
-N,· .. ,0, N} for the cases Irl < 1, Irl = 1, and r > 1. Compare the
result with the Fourier series of the one-sided discrete time geometric
signal of (3.23). Write the Fourier series for the periodic extensions
of both signals (period N for the one-sided signal and period 2N + 1
for the two sided signal) and sketch the two periodic signals.

3.5. Find a Fourier series for the discrete time infinite duration signal
defined by gn = n mod 10, n E Z. Is the Fourier series accurate
for all integer n? Compare the result to the Fourier series for the
continuous time ramp function g(t) = t mod 10, tEn.

3.6. Define the "roundoff" function q(x) which maps real numbers x into
the nearest integer, that is, q(x) = n if n -1/2 < x ~ n + 1/2. Define
the roundoff error by f(X) = q(x) - x. Find a Fourier series in x for
f(X).

3.7. What signal has Fourier transform e- 1fl for all real f?
154 CHAPTER 3. FOURIER INVERSION

i:
3.8. Suppose that G is the infinite duration CT Fourier transform of a
signal g. Then the definition at f = 0 gives the formula

g(t) dt = G(O). (3.110)

(a) Find an analogous result for the finite duration CT Fourier trans-
form.
(b) Repeat for the finite duration DT Fourier transform.
(c) Suppose now we have an infinite duration time signal h defined
by h(t) = G(t) for all real t. In words, we are now looking at
the function G as a time signal instead of a spectrum. What is
the Fourier transform H of h (in terms of g)?
(d) Use the previous part to find the dual of Eq. (3.110), that is,
relate J~oo GU) df to 9 in a simple way for an infinite duration
CT signal g.
(e) Use the previous part to evaluate the integral

L oo
sinc(t) dt.

(To appreciate this shortcut you might try to evaluate this inte-
gral by straightforward calculus.)
3.9. What is the Fourier transform of the finite duration, continuous time
signal
( ) _ sin(21Tt(~». [_~ ~)?
9t - sin(1Tt) ,t E 2' 2 .
Find a Fourier series representation for the periodic extension (having
period 1) of this signal.
3.10. What infinite duration discrete time signal {gn; n E Z} has Fourier
transform {n(4f); f E [-~, ~)}?
3.11. Define the finite duration continuous time signal 9 = {g(t); t E
[-T/2, T/2)} by g(t) = A for -T/4 ~ t ~ T/4 and g(t) = 0 for
It I > T /4. Find the Fourier transform GU). Is the inverse Fourier
transform of GU) equal to g(t)?
Now let g(t) be the infinite duration continuous time signal formed by
zero-filling g(t). Again find the Fourier transform and inverse Fourier
transform. (Note that the Fourier transform has the same functional
form in both cases, but the frequency domain of definition is different.
The inverse Fourier transforms, however, are quite different.)
3.11. PROBLEMS 155

3.12. What is the Fourier transform of the continuous time finite duration
signal 9 = t; t E [- ~, ~ )? Find an exponential Fourier series repre-
sentation for g. Find a trigonometric Fourier series representation for
g.
3.13. Find a trigonometric Fourier series representation for the infinite du-
ration continuous time periodic signal g(t) = (t - 1/2)modl - 1/2.
(First sketch the waveform.) For what values of t is the Fourier series
not accurate. Repeat for the signal g(t - 1/2).
3.14. Prove the Fourier transform pair relationship of (3.84)-(3.85) by di-
rect substitution.
3.15. Show how the an and bn are related to g(t) in Eq. 3.83. Express the
an and bn in terms of the Cn of the exponential Fourier series.
3.16. Orthogonal Expansions
A collection of signals {4>i(t); t E [0, T]}, i = -N,···, 0,1,2, ... , N
are said to be orthonormal on [0, T] if

loT 4>i(t)4>;(t)dt = 8i- j , (3.111)

that is, the integral is 1 if the functions are the same and ° otherwise.

(a) Suppose that you are told that a real-valued signal 9 is given by
N
g(t) = 2: bk4>k(t).
k=-N

How do you find the bk from g(t) and the 4>k(t)? Evaluate the
energy
£g = loT Ig(tWdt
in terms of the bi. (This is an example of Parseval's theorem.)
(b) Are the functions A sin(27rkt/T); k = 1,2, ... , N orthonormal
on [O,T]?
(c) Suppose that we have an orthonormal set of functions {4>k(t); t E
[0, T]}, k E Z and that we have an arbitrary signal g(t), t E
[0, T]. We want to construct an approximation p(t) to g(t) of
the form
N
p(t) = 2: cn4>n(t).
n=-N
156 CHAPTER 3. FOURIER INVERSION

Define the error signal resulting from this approximation by

e(t) = get) - pet).

We wish to determine the best possible choice of the coefficients
en. Toward this end define the coefficients

and the approximation

N
q(t) = L an<pn(t).
n=-N

Consider the mean squared error

By adding and subtracting q(t) inside of the brackets argue that

r le(tWdt
T1 10
T
=
1
T 10
rT Ig(t) - q(t)1 2 dt
1 rT
+ T 10 Iq(t) - p(tW dt

-2~~ [IoT[g(t) - q(t)][q(t) - pet)]· dt]

Show that the rightmost integral is 0 and use this to conclude

that
~ loT le(t)1 2 dt ~ ~ loT Ig(t) - q(tWdt,
that is, the an are the optimum coefficients in the sense of min-
imizing the mean squared error.

3.17. What signal has Fourier transform AU); fER?

3.18. Suppose that you have a finite duration, discrete time signal {gn; n =
0,1, ... , N - 1} and you have found its DFT {Gk = G(k/N); k =
0,1, ... , N - 1}. Just to be difficult, a local curmudgeon demands a
formula for the Fourier transform for arbitrary real f, that is,
N-l
GU) =L gn e- i27r /n, fER.
n=O
 3.11. PROBLEMS 157

Find an expression for G(f) in terms of the Gk which is valid for

arbitrary real f.
The point of this problem is that knowing the DFT for only the finite
collection of frequencies is enough to compute the formula giving the
transform for all real f. This is a trivial form of "sampling theorem"
since G(f) is recovered from its "samples" G(k/N).
3.19. Suppose that you are told that a signal get); t E 'R is time-limited
to [-~/2, ~/21 and that its spectrum G(f) has the property that
G(k/~) = 0 for all nonzero integers k and G(O) 1:- O. What is get)?
Suppose that you only know that 9 is time-limited to [-~, ~l, but it
still has the property that G(k/~) = 0 for all nonzero integers k. Give
examples of at least two nontrivial signals that meet this condition.

3.20. Suppose that 9 = {e- t ; t E [0, T)} is a finite duration continuous

time signal. (T> 0 is a fixed parameter.)

(a) Find the Fourier transform G of g.

(b) Find a Fourier series representation for g.
(c) Find the Fourier transform GT of a zero filled extension gT of g,
i.e., gT(t) = get) for t E [0, T) and gT(t) = 0 for all other real t.
Sketch gT.
(d) Find a Fourier series for the periodic extension g of g, i.e., get) =
get) for t E [0, T) and 9 is periodic with period T. Sketch g.
(e) Now suppose that T becomes very large. Let G 1 denote the
CTFT of the signal gl = {e- t u_l(t);t E 'R}. Describe the
differences and similarities between G 1 and GT . Is it true that
G 1 (f) = limT~oo GT(f) in some sense? (You are not required
to come up with a mathematical proof, but you are expected to
look at the limit and comment on why it might or might not
converge to the guessed limit.)
3.21. Consider the matrix notation for the DFT and its inverse: G = W g
and g = W- 1 G. Define the energy in the signal and its transform by
N-l
c9 = IIgll2 = g*g = L /g(n)/2
n=O

and
N-l
ca = IIGII2 = G*G = L /G(n/NW,
n=O
158 CHAPTER 3. FOURIER INVERSION

respectively. Show that

This is Parseval's equality for the DFT.

Hint: Use the matrix form of the transform and recall the relation of
W-l and W*.
3.22. This problem treats a special case of an important application of
Fourier (and other transforms) called transform coding. Suppose that
a signal g = (go,··· ,gN-d, represented as a column vector g =
(gO,··· ,gN_l)t (we use vector notation here for convenience), is to
be reconstructed by the following sequence of operations. First we
take its Fourier transform G = Wg, where the matrix W is given
by (2.6). This transform is then approximated by another vector G.
For example, G might be formed by quantizing or digitizing G in
order to store it in a digital medium or transmit it over a digital
channel. We then use G to reconstruct an approximation to g by
inverse transforming to form g = W-1G. Suppose that we define
the mean squared error or MSE of the overall reconstruction using
the notation of Problem 3.21 as
N-l
€~ = N-111g - 911 2 = N- 1 L Ign _ g~12. (3.112)
n=O
We can similarly define the mean squared error in the Fourier domain
by
N-l
€2; = N-111G - oW = N- 1 L IG(~) - O(~)12. (3.113)
n=O

Show that the MSE in the original time domain is proportional to that
in the frequency domain and find the constant of proportionality.
Hint: See Problem 3.2l.
Now suppose that we wish to "compress" the image by throwing away
some of the transform coefficients. In other words, instead of keeping
all N floating point numbers describing the G( N); n = 0, 1, ... ,N -1,
we only keep M < N of these coefficients and assume all the remain-
ing N - M coefficients are 0 for purposes of reconstruction. Assuming
a fixed number of bytes, say m, for representing each floating point
number on a digital computer, we have reduced the storage require-
ment for the signal from Nm bytes to Mm bytes, achieving a com-
pression ratio of N : M. Obviously this comes at a cost as the setting
3.11. PROBLEMS 159

of some coefficients to zero causes error in the reconstruction and one

has to know in advance which of the coefficients are to be kept. If
the reconstruction error is small, it will not be perceivable or will not
damage the intended use of the image.
Suppose that one is allowed to optimally choose for a particular image
which of the M < N are the best coefficients to keep and what is the
resulting mean squared error? Note that in this case one would need
to specify which coefficients are nonzero as well as the values of the M
nonzero coefficients. Thus in this case the "compression ratio" N : M
alone does not really indicate the information needed to reconstruct
an approximate image.
The two dimensional version of transform coding is the most popular
technique for compressing image data and it is an integral part of sev-
eral international standards, including the international ISOjITU-T
JPEG (Joint Photographic Expert Group) standard. For a discussion
of the standard and further details on how the floating point numbers
are quantized and coded, see, e.g., [34, 26).
You might wonder why one would wish to do the approximating in
the frequency domain rather than in the original time domain. There
are many reasons, two of which can be easily described. First, the
Fourier transform of most interesting data sources tends to concen-
trate the energy in the lower frequencies, which means often many of
the higher frequency coefficients can be reproduced very coarsely or
simply thrown away without much damage. This effectively reduces
the dimensionality of the problem and provides some compression.
Second, both the eye and ear seem to be sensitive to signal behav-
ior in the Fourier domain, which means that doing a good job ap-
proximating the original signal in the Fourier domain will result in a
reproduction that looks or sounds good.
3.23. Consider two finite duration signals: h = {h(t) = 6 - t; t E [0,6)}
and 9 = {g(t) = cos (7rt/2), t E [0,4)}. Let h = {h(t); tEn} be the
zero-filled extension of h, Le.,

°
h(t) = { 6 - t t E [O,~) .
otherwIse

Let g be the periodic extension of g, and let k be the periodic extension

of h. Define r = g + k.

(a) Find the Fourier transform of h.

(b) Find the Fourier series coefficients Ch (k) for the signal h.
160 CHAPTER 3. FOURIER INVERSION

(c) Find the Fourier series coefficients cg (k) for the signal g.
(d) What is the period of r(t)?
(e) Express the Fourier series coefficients Cr (k) for the periodic signal
r(t) in terms of the coefficients ch(k) and cg(k) that you found
in parts (b) and (c).

3.24. Suppose that U is a unitary matrix which defines a transform G =

Wg. Suppose that V is yet another unitary matrix which also can be
used to define a transform. Is the matrix product VU also unitary?
If so, then how do you invert the transform UV g?
Chapter 4

Basic Properties

In this chapter the fundamental properties of Fourier transforms are de-

rived. These properties are useful in manipulating, evaluating, verifying,
and applying Fourier transforms.

4.1 Linearity
Recall that a linear combination of two signals 9 and h is a signal of the
form ag + bh = {ag(t) + bh(t)jt E T}. The most important elementary
property of Fourier transforms is given by the following theorem.
Theorem 4.1 The Fourier transform is linear; that is, given two signals
9 and h and two complex numbers a and b, then

F(ag + bh) = aF(g) + bF(h). (4.1)

The theorem follows immediately from the fact that the Fourier trans-
form is defined by a sum or integral and that sums and integrals have the
linearity property. We have already seen that the DFT is linear by express-
ing it in matrix form. Linearity can also easily be proved directly from the
definitions in this case:

L (ag
N-l N-l N-l
n + bhn ) e- i2tr !n = a L gn e- i2tr !n +bL h n e- i2tr !n
n=O n=O n=O
aG(f) + bH(f)
as claimed.
The linearity property is also sometimes called the superposition prop-
erty.
162 CHAPTER 4. BASIC PROPERTIES

Recall that the Fourier transform operation can be considered as an

example of a system mapping an input signal (the original time domain
signal) into an output signal (the Fourier transform or spectrum of the
time domain signal, a frequency domain signal). Viewed in this way, the
theorem simply states that the system described by the Fourier transform
operation is a linear system.
The idea of linearity for more general systems is fundamental to many
applications and so we here point out the general definition and interpret
the linearity property of Fourier transforms as a special case.
The basic linearity property implies a similar result for any finite col-
lection of signals; that is,
K K
F(L ang(n» = L anF(g(n». (4.2)
n=l n=l
This result is quite useful in computing Fourier transforms: if a signal is
the sum of a finite number of signals for which the transforms are known,
then the transform of the sum is the sum of the transforms.
An often useful property which does not immediately follow from linear-
ity, but which can be proved under certain conditions, is called countable
linearity or extended linearity wherein a property like the above one holds
for an infinite summation as well as a finite summation; that is,
00 00

F(L ang(n» =L anF(g(n». (4.3)

n=l n=l
We usually assume that the Fourier transform has the countable linearity
property, but be forewarned that this is true only when the signals g(n) and
the sequence of weights an are sufficiently well behaved.
As an example, suppose that

gn = ar nu_l(n) + bpnu_l(n); nEZ,

where Irl < 1 and Ipi < 1. Then from linearity the spectrum is immediately
a b
GU) = 1 _ re-i21ff + 1 _ pe- i21ff ; f E [0,1).

4.2 Shifts
In the previous section the effect on transforms of linear combinations of
signals was considered. Next the effect on transforms of shifting or delaying
a signal is treated.
4.2. SHIFTS 163

Theorem 4.2 The Shift Theorem

Given a signal 9 = {g(t)j t E T} with Fourier transform G = F(g),
suppose that for rET, gT = {g(t - r)j rET} is the shifted signal (the
shift is a cyclic shift in the finite duration case). Then

(4.4)

Thus a delay in the time domain produces a frequency dependent expo-

nential multiplier in the frequency domain. To interpret this result, consider
two sinusoids (or equivalently complex exponentials), one having a low fre-
quency and the other a high frequency. Suppose both sinusoids are shifted
by r. The sinusoid with the long period may be shifted by some small
fraction of its period, while simultaneously the sinusoid with the short pe-
riod is shifted by a large number of periods. Some further thought shows
that each sinusoidal component is shifted by a fraction of its period that
increases linearly with its frequency. Hence a fixed time shift corresponds
to a phase shift linearly proportional to frequency (as well as to r). For a
complex exponential, this linear phase shift is equivalent to multiplication
by another complex exponential with exponent -i21r fr.
From the systems viewpoint, the theorem tells us how to find the trans-
form of the system output in terms of the transform of the system input.
Many of the results of this chapter will be of this general form.
The proof is similar in all cases, so we only prove two of them. First
consider the DFT case of a signal {gnj n E {a, 1, ... , N - I}}. Here

N-l

= ~
~ g(n-T)modNe -i2"-/frn
n=O
N-l
i2 i2,,T-/fr
= ~ 9
~ (n-T)modN e- "-/fr(n-T)e-
n=O
N-l
~ 9 e-i2,,-/fr«n-T)modN)e-i27rT-/fr
= ~ (n-T)modN
n=O
N-l
'2 .L '2 .L ·2.L
= e-'''T N ~
~gne-'''Nn=e-'''TNF;.(g)j kEZN,
n=O

proving the result. Note that we have used the fact that e- i2";'n =
e- i2 ";'{nmodN) which holds since for n = KN + I with 0 < < N - 1
e-i21r -/fr(KN+I) = e-i21rkKe-i2,,;'1 = e-i2,,;'I.
164 CHAPTER 4. BASIC PROPERTIES

i:
Next consider the infinite duration CTFT. Here we simply change vari-
ables a = t - T to find

i: i:
Ff(gr) = g(t - T)e-i2trft dt

g(a)e- i2trf (a+r) da = e-i2trfr g(a)e-i2trfa da

= e-i2trfr Ff(9)·

The finite duration CTFT and the infinite duration DTFT follow by similar
methods.
As an example, consider the infinite duration continuous time pulse
p(t) = A for t E [0, T) and 0 otherwise. This pulse can be considered as a
scaled and shifted box function
T
p(t) = ADT / 2 (t - 2")
and hence using linearity and the shift theorem

(4.5)

4.3 Modulation
The modulation theorem treats the modulation of a signal by a complex
exponential. It will be seen to be a dual result to the shift theorem; that is,
it can be viewed as the shift theorem with the roles of time and frequency
interchanged.
Suppose that 9 = {g(t); t E T} is a signal with Fourier transform
G = {G(f); f E S}. Consider the new signal ge(t) = g(t)e i2trfot ; t E T
where fo E S is a fixed frequency (sometimes called the carrier frequency).
The signal ge(t) is said to be formed by modulating the complex expo-
nential ei2trfot, called the carrier, by the original signal g(t). In general,
modulating is the methodical alteration of one waveform, here the com-
plex exponential, by another waveform, called the signal. When the signal
and the carrier are simply multiplied together, the modulation is called
amplitude modulation or AM. In general AM includes multiplication by a
complex exponential or by sinusoids as in gc(t) = g(t) cos(27r fot); t E T
and gs(t) = g(t) sin(27r fot); t E T. Often AM is used in a strict sense to
mean signals of the form ga(t) = A[l + mg(t)] cos(27r fot) which contains
a separate carrier term Acos(27rfot). ga(t) is referred to as double side-
band (DSB) or double sideband amplitude modulation (DSB-AM), while
the simpler forms of gc or gs are called double sideband suppressed carrier
4.3. MODULATION 165

(DSB-SC). The parameter m is called the modulation index and sets the
relative strengths of the signal and the carrier. Typically it is required that
m and 9 are chosen so that Img(t)1 < 1 for all t.
Amplitude modulation without the carrier term, i.e., gc or gs and not
ga, are called linear modulation because the modulation is accomplished
by a linear operation, albeit a time varying one.
The operation of modulation can be considered as a system in a math-
ematical sense: the original signal put into the system produces at the
output a modulated version of the input signal. It is perhaps less obvious
in this case than in the ideal delay case that the resulting system is linear
(for the type of modulation considered - other forms of modulation can
result in nonlinear systems).
Theorem 4.3 The Modulation Theorem.
Given a signal {g(t); t E T} with spectrum {G(f); IE S}, then
{g(t)ei21r/ot; t E T} ::) {G(f - 10); IE S}
1 1
{g(t) cos(21r lot); t E T} ::) {2 G (f - 10) + 2G(f + 10); I E S}
i i
{g(t) sin(21r lot); t E T} ::) {2G(f + 10) - 2G(f - 10); IE S}

where the difference I - 10 and sum I + 10 in the frequency domain are

treated like the shift in the time domain, that is, ordinary algebra in the
case 01 S = n or {kiT; k E Z} and modulo the frequency domain in the
case 01 [0, 1) or {kiN; k E ZN}.
Thus modulating a signal by an exponential shifts the spectrum in the
frequency domain. Modulation by a cosine with spectrum G(f) causes
replicas of G (f) to be placed at plus and minus the carrier frequency. These
replicas are sometimes called sidebands.
Prool: As usual we consider only two cases, leaving the others for an

I:
exercise. In the infinite duration continuous time case,

I:
Ge(f) = (g(t)ei21r/ot)e-i21r/t dt

= g(t)e- i21r (f-/o)t dt

= G(f - 10).
The results for cosine and sine modulation then follow via Euler's relations.
For the DFT we have that with a frequency 10 = ko/N

L (gnei21r!J9.n)e-i21r~n
N-l

n=O
166 CHAPTER 4. BASIC PROPERTIES

n=O

= G(k ~kO).
Thus, for example, the Fourier transform of n(t) cos(7I"t) is given by

F/({n(t)cos(7I"t)j t E R}) = ~(sinc(f-1/2)+sinc(f+1/2))j fER. (4.6)

4.4 Parseval's Theorem

In this section we develop a result popularly known as Parseval's theo-
rem, although Parseval only considered the special case of continuous time
Fourier series. The result might be better described as Rayleigh's theorem
after Lord Rayleigh, who proved the result for the continuous time Fourier
transform. We will follow common use, however, and collect the four sim-
ilar results for the four signal types under the common name of Parseval's
theorem.

I:
The energy of a continuous time infinite duration signal is defined by

c9 = Ig(t)12 dt
and it has the interpretation of being the energy dissipated in a one ohm
resistor if 9 is considered to be a voltage. It can also be viewed as a measure
of the size of a signal. In a similar manner we can define the energy of the

I:
appropriate Fourier transform as

ca = IG(fW df.
These two energies are easily related by substituting the definition of the

i:
transform, changing the order of integration, and using the inversion for-
mula:

ca = G(f)G* (f) df

i: (i:
G(f) g(t)e- i27r / t dt) * df

i: (i:
I:
g*(t) G(f)e i27r / t df ) dt

= g*(t)g(t) dt
4.4. PARSEVAL'S THEOREM 167

proving that the energies in the two domains are the same for the continuous
time infinite duration case.
The corresponding result for the DFT can be proved by the analogous
string of equalities for discrete time finite duration or by matrix manipula-
tion as in Problem 3.21. In that case the result can be expressed in terms
of the energies defined by

L Ig(n)12
N-l
cg =
n=O

and
N-l
Co =L IG(n/NW
n=O
as
1
c = NCo.
g

The following theorem summarizes the general result and its specializa-
tion to the various signal types.

Theorem 4.4 Parseval's Theorem

c
Given a signal 9 with Fourier transform G, then g = C Co, where the
energy is defined as the integral or sum of the signal squared over its domain
of definition and C is a constant that equals 1 for infinite duration signals
and the inverse duration of the signal for finite duration signals. To be
specific:

I: I:
1. If the signals are infinite duration continuous time signals, then

Cg = Ig(tW dt = IG(fW df = Ca.

2. If the signals are finite duration continuous time signals, then

[T 1 1
L IG(fW = Tca .
00

cg = io Ig(tW dt =T
o n=-oo

3. If the signals are infinite duration discrete time signals, then

168 CHAPTER 4. BASIC PROPERTIES

4. If the signals are finite duration discrete time signals, then

~ /2 1 ~l n 2 1
tg = L..J /gn = N L..J /G(N)/ = Nta.
n=O n=O

As mentioned earlier, the first relation is better described as Rayleigh's

theorem and the second relation is Parseval's theorem as originally devel-
oped. The second relation can be stated in its classical form as follows. If
a continuous time signal 9 is periodic with period T, then the the signal
has a Fourier series described by (3.81)-(3.82), where Ck = G(~)/T. Thus
the finite duration result becomes

(4.7)

The results extend in a straightforward way to integrals and sums of

products of signals. Consider for example two continuous time infinite

i:
duration signals 9 and h with Fourier transforms G and H and consider
the integral
< g, h >= g(t)h*(t) dt,

where we have used the inner product notation as an abbreviation. The

inner product is also called the scalar product and is often denoted by (g, h).

<G,H> = i:
Exactly as in the earlier case where 9 = h, we have that

G(f)H*(f) df

= L: (L: g(t)e-i21rftdt) H*(f)df

= L: (L: get) H*(J)e-i21rft df ) dt

= L: (L: df) * dt
i:
get) H(J)ei21rft

= g(t)h*(t) dt

= <g,h> .
In a similar fashion we can define inner products for discrete time finite
duration signals in the natural way as
N-l
< g, h >= L g(n)h*(n)
n=O
 4.4. PARSEVAL'S THEOREM 169

and
N-l
< G,H >= L G(~)H*(~)
n=O
and derive by a similar argument that for the DFT case

< g,h >= N- 1 < G,H >.

Repeating these arguments for the various signal types yields the following
general form of Parseval's theorem, which also goes by other names such
as Rayleigh's theorem, Plancherel's theorem, and the power theorem.

Theorem 4.5 Parseval's Theorem: General Form

Given two signals 9 and h with Fourier transforms G and H, respec-
tively, then < g, h > = C < G, H >, where C is defined as in the previous
theorem. In particular,

i: i:
1. If the signals are infinite duration continuous time signals, then

< g, h >= g(t)h*(t) dt = G(f)H*(f) df =< G, H > .

2. If the signals are finite duration continuous time signals, then

r T
< g,h >= 10 g(t)h*(t)dt = T
1 ~ n n
~ G(T)H*(T)
1
= T < G,H >.
o n=-oo

3. If the signals are infinite duration discrete time signals, then

< g,h >= L

00

n=-oo
gnh~ = 1 -!
!
G(f)H*(f)dj =< G,H >.

4. If the signals are finite duration discrete time signals, then

N-l 1 N-l 1
< g,h >= L gnh~ = N L G(~)H*(~) = N < G,H >.
n=O n=O

We shall later see that Parseval's theorem is itself just a special case of
the convolution theorem, but we do not defer its statement as it is a handy
result to have without waiting for the additional ideas required for the more
general result.
170 CHAPTER 4. BASIC PROPERTIES

I:
Parseval's theorem is extremely useful for evaluating integrals. For ex-
ample, the integral
sinc 2 (t) dt
is difficult to evaluate using straightforward calculus. Since sinc(t) +-+ n(J),
where the double arrow was defined in (3.2) as denoting that the signal and
spectrum are a Fourier transform pair, Parseval's Theorem can be applied

1: 1: 1:
to yield that

sinc 2 (t) dt = n2(J) df = n(J) df = l.

1: 1:
As an example of the general theorem observe that

=
I:
sinc3 (t) dt sinc(t)sinc 2 (t) dt

n(J) 1\ (f) df

21'i
1

(1 - 1) df = ~.

4.5 The Sampling Theorem

We now turn to a powerful but simple application of Fourier series. We
use the ideas of Fourier series representations of finite duration or periodic
functions to obtain a representation for infinite duration continuous time
band-limited signals. This provides a surprising connection between con-
tinuous time and discrete time signals formed by sampling the continuous
time signals. This topic will be further explored later in the book, but the
fundamental result is derived here.
Suppose that 9 is an infinite duration continuous time signal with a
spectrum G having the property that G(f) = 0 for If I 2: W; that is, the
signal is band-limited. For the development we can choose any W for which
this is true, but a particularly important example will be the smallest W
for which the assumption is true, a value which we shall call W min and to
which we shall refer as the bandwidth of the signal. For the development
that follows, any W 2: W min will do.
Consider the truncated finite bandwidth spectrum G = {G(J); f E
( - W, W)}. As with a finite duration signal, we can write a Fourier series
for G(f) in this region:
00

(4.8)
n=-oo
4.5. THE SAMPLING THEOREM 171

where

Note that the only thing unusual in this derivation is the interchange
of signs in the exponentials and the fact that we have formed a Fourier
series for a finite bandwidth spectrum instead of a Fourier series for a finite
duration signal. It is this interchange of roles for time and frequency that
suggests the corresponding changes in the signs of the exponentials.
Since G(J) is assumed to be zero outside of (-W, W), we can rewrite
the coefficients as

en _1_ {w G(J)ei27f-rtvn df
2W Lw

= _1 ('XJ G(J)ei27r-rtvn df
2W i-co
1 n
= 2W g (2W); n E Z.

Thus we have the following formula expressing the spectrum of a band-

limited signal in terms of the samples of the signal itself:

f E (-W,W) (4.10)
otherwise
This formula yields an interesting observation. Suppose that we define
the discrete time signal, = {Tn; n E Z} by the samples of g, i.e.,

,n = genTs); nEZ, (4.11)

where Ts = 1/2W ~ 1/2Wm in is the sampling period, the time between

consecutive samples. Then the DTFT of the sampled signal is ,n
L
co
r(f) = ,ne-i27fnl
n==-oo

(4.12)

Comparing (4.10) and (4.12) yields

G(f) = { ~8r(fTs); f E [- 2h ' 2~.l (4.13)

else
172 CHAPTER 4. BASIC PROPERTIES

We will see in the next section how to generalize this relation between the
transform of a continuous time signal to the DTFT of a sampled version
of the same signal to the case where Ts > 1/2Wmin, i.e., the signal is not
bandlimited or it is bandlimited but the sampling period is too large for
the above analysis to hold.
Knowing the spectrum in terms of the samples of the signal means that
we can take an inverse Fourier transform and find the original signal in
terms of its samples! In other words, knowing g(n/2W) for all n E Z
determines the entire continuous time signal. Taking the inverse Fourier
transform we have that

Summarizing the above, provided W is chosen so that W 2:: Wmin and

hence so that GU) = 0 when If I 2:: W, then the resulting bandlimited
signal can be written in terms of Ts = 1/2W as

(4.14)

Alternatively, we can express the formula in terms of the sampling frequency

fs = l/Ts, the number of samples per second, as

:E
00

g(t) = g(; )sinc (fat - n] , (4.15)

n=-oo a

provided only that fs ~ 2Wmin , a number known as the Nyquist frequency

or Nyquist rate.
This is not a Fourier series because sinc functions and not exponentials
appear, but it can be shown to be an orthogonal expansion since the sinc
functions can be shown to be orthogonal on R. Because of its fundamental
importance in applications of Fourier analysis, we summarize this result
formally.
4.6. THE DTFT OF A SAMPLED SIGNAL 173

Theorem 4.6 The Whittaker-Shannon-Kotelnikov Sampling Theorem

Suppose that g = {get); t E 'R.} is a continuous time infinite duration
signal that is band limited to (-W, W),. that is, its Fourier transform G (f)
has the property that

G(f) = 0; f'l. (-W, W).

If the signal is sampled with sampling period Ts and sampling frequency
fs = l/Ts and if fs ~ 2W, then the original signal can be perfectly recovered
from its samples as

L
00

get) = genTs) sinc (fat - n]. (4.16)

n=-oo

The sampling theorem states that provided a bandlimited signal is sam-

pled fast enough (at least twice its maximum frequency in Hz), then the
signal can be recovered perfectly from its samples from the sampling ex-
pansion.
We shall later see a more engineering motivated derivation of the same
result using the ideal sampling train and the properties of convolution. The
previous derivation points out, however, that the result is easily demon-
strated directly using Fourier series.
The basic idea can also be applied to obtain the dual result for time
limited signals. Suppose that a signal g has the property that get} = 0 for
It I ~ T. Then it can be shown that
00

G(f) = L
n=-oo
G(2~)sinc [2T(f - 2~)] . (4.17)

This is sometimes called sampling in the frequency domain.

4.6 The DTFT of a Sampled Signal

The idea of sampling raises an interesting question: We have defined a
continuous time Fourier transform for the original signal, and we could
apply a discrete time Fourier transform to the sampled signal. How do these
two Fourier transforms relate to each other? The answer to this question
will provide a means of resolving another question that was implicit in the
previous section. What happens if you sample a signal too slowly, i.e.,
below the Nyquist rate?
Suppose that {get); t E 'R.} and its Fourier transform {G(f); f E 'R.}
are both absolutely integrable and continuous functions. Fix a positive
174 CHAPTER 4. BASIC PROPERTIES

number Ts and define a discrete time signal 'Y = {Tn; n E Z} by

'Yn = g(nTs); n E Z;

that is, 'Y is the sampled version of g. Unlike the previous section, no
assumptions are made to the effect that 9 is bandlimited so there is no
guarantee that the sampling theorem holds or that 9 can be reconstructed
from 'Y. The immediate question is the following. How does the DTFT r
of the sampled signal, defined by

r(f) = L
00.
'Yke-t27rfkj f E
1 1
[-2' 2) (4.18)
k=-oo

compare to the CTFT G of the original signal? We establish this important

result in two ways, each of which is of interest in its own right. The first
is the more direct and the shorter. The second provides an application of
Fourier series and yields an important side result, the Poisson summation
formula.
First Approach
Since the unknown r(f) is the Fourier transform of the discrete time
signal 'Y = (Tn = g(nT); n E Z}, 'Y must be the inverse DTFT of r:

[21
1

'Yn = ei27rfnr(f) df. (4.19)

2

We also have that

'Yn =
= I:
g(nT)
G(f)ei27rfnT. df

= "
~
00
r .
h.
~

k - 1/ 2 )
G(f)ei27rfnT'df (4.20)
k=-oo ~

where we have broken up the integral into an infinite sum of integrals over
disjoint intervals of length liT•. Each of these integrals becomes with a
change of variables f' = fT. + k

r ~
T.
}l.k-l/2)
G(f)ei27r fnT. df
T.
4.6. THE DTFT OF A SAMPLED SIGNAL 175

so that interchanging the sum and integral in (4.20) yields the formula

"In = [~
~
ei2 11"/n [ f
k=-oo
G(f ~ k)] df. (4.22)

Comparison with (4.19) identifies the term in brackets as the DTFT of "I;
that is,
r(f) = f
k=-oo
G(f ~ k). (4.23)

Second Approach
We establish this relation in an indirect manner, but one which yields an
interesting and well known side result. Let t denote the periodic extension
of r with period 1; that is,

E
00

t(f) = "Ik e- i2 11"/k; f E 'R. (4.24)

k=-oo

In much of the literature the same notation is used for r and t with the
domain of definition left to context, but we will distinguish them.
Consider the following function of frequency formed by adding an infi-
nite number of scaled and shifted versions of G:
1 f-k
a(f) =~
S
E
00

k=-oo
G( ~); fEn.
S
(4.25)

This function is a continuous parameter function (f is a continuous argu-

ment) and it is periodic in f with period 1 since adding any integer to f
simply results in a reindexing of the infinite sum and does not change the
value of the sum. Hence we can expand a(f) in a Fourier series in f, e.g.,

L
00

a(f) = Ck e- i2 11"Ik, (4.26)

k=-oo

where
Ck = 11/2 a(f)e i2 11"/k df.
-1/2
Before evaluating these coefficients, note the similarity of (4.24) and (4.26).
In fact we will demonstrate that Ck = "Ik = g(kTs), thereby showing that
t(f) = a(f) and hence r(f) = a(f) for a E [-1/2,1/2) since the two
continuous functions have the same Fourier series, which will provide the
desired formula relating rand G.
176 CHAPTER 4. BASIC PROPERTIES

To evaluate the coefficients Ck we use the usual trick of substitution

followed by a change of order of integration and summation.

L e G(I - n)ei27rkf df.

00 1
~ (4.27)
Ts n=-oo
1-1
2
Ts

Changing variables I' = (f - n) ITs in the rightmost integrals we have

1 r...L--",-
L _T: G(f')e i2trk (T.!'+n) Tsdl'
00

Ck Ts 1-2~'
n=-oo

f:
2T" Ts

= r2~'
J_--L_.D-
-;, G(f')ei2trkTs!' dj'
I:
n=-oo

g(kTs).
2Ts Ts

G(f')e i2trkT,!, dj'

(4.28)

The infinite sum of integrals over a disjoint collection of intervals together

constituting the entire real line resulted in a single integral in the penulti-
mate step above.
We have thus proved that the DTFT of the discrete time signal 'Y formed
by sampling the continuous time signal 9 with sampling period Ts is

(4.29)

where G is the CTFT of the continuous time signal g. Before interpreting

this result, however, we point out a side result that is important in its
own right. The above development proved the following result, which is
a version of the Poisson summation formula. We state the result here for
later reference.

Theorem 4.7 The Poisson Summation Formula

Given an infinite duration signal {g( t) ; tEn} with absolutely integrable
continuous Fourier transform G(f), then for any T > 0

~ f:
k=-oo
G( f ; k) = f:
n=-oo
g(nT)e- i2trnf j lEn. (4.30)
4.6. THE DTFT OF A SAMPLED SIGNAL 177

In particular, i/ / = 0, then
G(!.)
L ;- L
00 00

= g(nT). (4.31)
k=-oo n=-oo

Returning to the consideration of the DTFT of a sampled process, we

formally state (4.29) in context as a theorem and then interpret it in light
of the sampling theorem.
Theorem 4.8 Given an infinite duration signal {g(t)j t E R} with abso-
lutely integrable continuous time Fourier transform G(f), then the discrete
time Fourier transform 0/ the signal hn = g(nTs)j n E Z} is given by
/-k
r(f) =T
1
L
00

s k=-oo
G(--r)j / E
S
1 1
[-2' 2)' (4.32)

Thus the DTFT of a sampled signal is the sum of frequency scaled

shifted replicas of the spectrum of the original continuous time signal.
As an example, consider the simple Fourier transform of a simple con-
tinuous time signal in Figure 4.1. The DTFT of the sampled signal will

G(f)

f
-W 0 W

Figure 4.1: Original CTFT G(f)

be the sum of shifted replicas of G(f /Ts), the original spectrum with the
argument "stretched." We depict this basic waveform by simply relabeling
the time axis as in Figure 4.2.
If the sampling period Ts is chosen so that 1/Ts ~ 2W or 1/2 ~ WTs ,
then the DTFT of the sampled signal is given by (4.29) and the individual
terms in the sum do not overlap, yielding the picture of Figure 4.3 with
separate "islands" for each term in the sum. Only one term, the k = 0
term, will be nonzero in the frequency region [-1/2,1/2]. In this case,
r(f) = G(f /Ts)/Tsj f E (-1/2,1/2) (4.33)
178 CHAPTER 4. BASIC PROPERTIES

Figure 4.2: Stretched Original CTFT G(f /Ts)

r(f)

f
-1 1 1
2

Figure 4.3: DTFT of Sampled Signal: l/Ts > 2W

and the DTFT and CTFT are simply frequency and amplitude scaled ver-
sions of each other and the continuous time signal 9 can be recovered from
the discrete time signal 'Y by inverting

G(f) = {Tosr(Tsf) f E [-2~.' 2~.J , (4.34)

else

which is the same as (4.13) and provides another proof of the sampling
theorem! If 9 is not bandlimited, however, the separate scaled images of G
in the sum giving r(f) will overlap as depicted in Figure 4.4 so that taking
the sum to form the spectrum will yield a distorted version in (-1/2,1/2),
as shown in Figure 4.5. This prevents recovery of G and hence of 9 in
general. This phenomenon is known as aliasing.
4.6. THE DTFT OF A SAMPLED SIGNAL 179

t(f)

f
-2 ~ 1 2

Figure 4.4: DTFT of Sampled Signal is sum of overlapping islands when

11Ts < 2W

t(f)

f
1 1 1
2T. T. 2T.

Figure 4.5: DTFT of Sampled Signal: 11Ts < 2W

The Poisson Summation

The Poisson summation formula of Theorem 4.7 is primarily important
here because of its basic role in deriving the DTFT of a sampled signal.
It also provides some interesting equalities for sums in the two domains
and is thereby useful in evaluating some apparently complicated infinite
summations. As an example, consider the Gaussian signal g( t) = e-lI't 2 j t E
n which has as Fourier transform G(f) = e-1I'/2 j fEn. Then for any
T>O
(4.35)
180 CHAPTER 4. BASIC PROPERTIES

This equality is surprising and not trivial to prove.

The Poisson summation formula is also seen in the dual form with the
sum of replicas in the time domain instead of the frequency domain. The
following version can be proved in a similar manner to the previous or by
using duality.

Theorem 4.9 The Poisson Summation Formula: Second Form

Given an infinite duration, absolutely integrable, and continuous signal
{g(t)j t E 'R-} with Fourier transform G(J), then for any T > 0

f:
n=-oo
g(t - nT) = f:
k=-oo
G~) ei21r!J.t. (4.36)

Setting t = 0 yields the special case of (4.31).

4.7 * Pulse Amplitude Modulation (PAM)

In the previous section the DTFT of a sampled continuous waveform was
derived. We now reverse the process and see what happens when a discrete
time signal "I with a known DTFT r is used to construct a continuous time
signal by modulating a pulse sequence. We begin without any assumptions
regarding the origin of the discrete time signal. Later the case of sampling
will be revisited and the effect of bandlimiting again considered.
Given the discrete time signal "I, define a continuous time signal r =
{r(t)j t E 'R-} by forming a PAM signal

L
00

r(t) = "Inp(t - nTs)j t E 'R, (4.37)

n=-oo

where the signal (pulse) p = {p(t)j t E 'R} has a Fourier transform P(f).
In the ideal sampling expansion, the pulses would be sine functions and the
transforms box functions. A more realistic pulse might be

p(t) = {1/Ts t E [0, ~s) (4.38)

o otherwlse
which would correspond to the overall system being an idealized sample
and hold circuit. What is the transform of r and how does it relate to those
of rand P? We have that
4.8. THE STRETCH THEOREM 181

L
00

= rnP(f)e-i2rrnIT., (4.39)
n=-oo

where the last step used the shift theorem for continuous time Fourier
transforms. Pulling P(f) out of the sum we are left with

L
00

R(f) = P(f) rne-i2rrn/T.

n=-oo
(4.40)
where f' is the periodic extension of f.
Now return to the case where r is formed by sampling a continuous
time signal 9 with transform G. From (4.29) we then have that the Fourier
transform of the reconstructed signal will be

R(f) = P(f); f
s k=-oo
G(f - ; ).
S
(4.41)

An important aspect of this formula is that if P(f) is itself bandlimited,

then so is R(f). For example, if we choose the sinc pulse of the idealized
sampling expansion, p(t) = sinc(t/Ts), then P(f) is a box function which
is zero for If I 2: 1/2Ts and hence R(f) is similarly bandlimited. If 9 is
bandlimited to the same band, e.g., if 9 is bandlimited to [-W, W] and
the sampling frequency fs = 1/Ts exceeds the Nyquist frequency 2W, then
(4.41) simplifies to R(f) = G(J) and we have once again proved the sam-
pling theorem since the PAM signal constructed by modulating sincs with
the samples has the same transform as the original signal.
Suppose on the other hand that one uses the rectangular pulse of (4.38),
then
P(f) = sinc( ITs )e- i2rr IT./4
2
which is not bandlimited, and hence neither is R(J).

4.8 The Stretch Theorem

We have seen several specific examples of scaling a time or frequency domain
variable. We now develop the general cases. Unfortunately, these results
182 CHAPTER 4. BASIC PROPERTIES

differ in important ways for the different signal types and hence we are
forced to consider the cases separately.

Theorem 4.10 Continuous Time Stretch Theorem

Given a continuous time, infinite duration signal g = {get); t E 'R} with
Fourier transform G = {GU); / E 'R} and a real-valued nonzero constant
a, then
1 /
{g(at); t E R} H {~G(;;); / E R}. (4.42)

Proof: We here consider only the case a > O. The case of negative a is
left as an exersise. If a is strictly positive, just change variables 7 = at to
obtain

As an example of the theorem, the Fourier transform of get) = n(2t) is

immediately seen to be !sinc( i).
The stretch theorem can be applied to a finite duration continuous time
signal, but the stretching changes the domain of definition of the signal, e.g.,
from [0, T) to [0, T / a) if a > o. (For this discussion we confine attention
to the case of positive a.) If 9 = {get); t E [0, T)} has a Fourier transform
{G(~); k E Z}, then the stretched signal ga = {g(at);t E [O,T/a)} will
have a spectrum G a = {rhG(**); k E Z} = {rhG(~); k E Z}, that is,
the multiplier inside G is cancelled by a corresponding stretch in the length
of the time domain. This apparent oddity is in fact consistent with the
infinite duration stretch theorem, as can be seen in two ways. First, we
have seen that the spectrum of any finite duration signal of duration D has
the form H = {H(k/D); k E Z}. Thus since the duration of Ga is T/a,
we have that Ga = {Ga(k/(T/a)) = Ga(ak/T); k E Z}. Comparing this
with the previous formula for G a shows that G a (*) = rhG(~). Setting
/ = k/(T/a) this is
1 /
Ga(f) = ~G(;;),
exactly as in the infinite duration case.

4.9 * Downsampling
Stretching the time domain variable has a much different behavior in dis-
crete time. Suppose that {gn; n E Z} is a discrete time signal. If analogous
to the continuous time case we try to form a new signal {gan; n E Z} we run
4.9. * DOWNSAMPLING 183

into an immediate problem: if an is not an integer, then gan is not defined

and the new signal does not make any sense. In order to have gan be a
reasonable signal we must do one of two things. Either we must restrict
an to be an integer for all integers n so that gan is well-defined or we must
extend the definition of our original signal to cover a wider index set; that
is, define the values of 9an for all of those an which are not integers. We
focus on the first approach and later return to the second.
If an is required to be an integer for all integers n, then a must itself be
an integer (or else an would not be an integer for the case n = 1). Hence let's
rename a as M and consider the new discrete time infinite duration process
g[MJ = {gMn; n E Z}. Observe that the new process is formed by taking
every Mth sample from the original process. Hence stretching in discrete
time is called downsampling. Unlike the continuous time case, we have not
just stretched the original signal, we have thrown away several of its values
and hence have likely caused more distortion than a mere compression of
the time scale. Alternatively, stretching a continuous time signal by a could
be undone by restretching using l/a; nothing was irretrievably lost. On the
other hand, stretching in the discrete time case has destroyed data.
The questions are now: Is there an analogous stretching theorem for
the discrete time case and are there conditions under which one can recover
the original signal from the downsampled signal? We shall see that these
questions are interrelated and are resolved by a discrete time version of the
sampling theorem.
As previously, let 9 denote the original signal and let g[MJ denote the
downsampled signal. Let G and G[M] denote the corresponding Fourier
transforms. We have that

L
00

G[MJ(f) = gnMe- i2 71'/n

n=-oo

which in general bears no simple relationship to the spectrum

L
00

G(f) = gne- i2 71'/n.

n=-oo

Unlike the continuous time case we cannot simply change variables.

The original signal 9 can be obtained by inverting G as

[21
1

gn = G(f)e i2 71'/n dJ
2

and hence we can find the downsampled signal by simply plugging in the
184 CHAPTER 4. BASIC PROPERTIES

appropriate values of time:

1
gnM = {2 G(f)ei21f/nM df.
J-!
Here we can change variables (1 = f M to write

(4.43)

This formula can be compared with the inverse Fourier transform represen-
tation for the downsampled process:

gnM = I: 1

2
G[Ml(f)e i21f / n df. (4.44)

It is tempting to identify G[Ml(f) with G(f /M)/M, but this inference can-
not in general be made because of the differing regions of integration. Sup-
pose, however, that G(f /M) = 0 for If I ~ 1/2; that is, that
1
G(f) = 0 for Ifl ~ 2M'

This is a requirement that the discrete time signal be bandlimited analogous

to the definition for continuous time signals. If we make this requirement,
then (4.43) becomes

gnM - -12"-! 1 L
G(MM) ei21f/n dlf (4.45)

and this formula combined with (4.44) and the uniqueness of Fourier trans-
forms proves the following discrete time stretch theorem.

Theorem 4.11 Discrete Time Stretch Theorem (Downsampling Theorem)

Given a discrete time, infinite duration signal 9 = {gn; n E Z} with
Fourier transform G = {G(f); f E [-~, ~)}. If there is an integer L such
that G(f) = 0 for If I ~ 1/2L, then for any positive integer M :5 L
1 f 1 1
{gnM; n E Z} f-t {M G( M); f E [-'2' '2)}. (4.46)

The theorem is illustrated for M = 2 in Figure 4.6.

The bandlimited condition and the implied result easily provide a dis-
crete time sampling theorem.
4.9. * DOWNSAMPLING 185

G(f)

I
1 1
4' 2'

(a) DTFT of {gni n E Z}:

G(f) = 0 for III > 2lt
M=2

G[Ml(f)

G(O)

I
1 1
4' 2'

(b) DTFT of {g2ni n E Z}

Figure 4.6: DTFT of Downsampled Signal

186 CHAPTER 4. BASIC PROPERTIES

Theorem 4.12 The Discrete Time Sampling Theorem

Suppose that g = {gn; n E Z} is a discrete time, infinite duration signal
with Fourier tronsform G = {GU); f E [-!, !)}. Then if there is an
integer L such that GU) = 0 for ifi ~ 1/2L, then for any M ~ L

L
00

gn::: gkMSinc(~ - k) (4.47)

k=-oo

and hence the original signal can be recovered from its downsampled version.
Proof: We have, using the bandlimited property and making a change
of variables, that

L 121
00 1
= gkM e i2tr !(j;r-k) df
k=-oo -~

L
00

= gkMsinc<;-k).
k=-oo

The theorem can also be proved in a manner much like the contin-
uous time sampling theorem. One begins by expanding the bandlimited
transform GU) on the region [-1/2M,1/2M) into a Fourier series, the
coefficients of which are in terms of the samples. An inverse DTFT then
yields the theorem.

4.10 * Upsampling
As previously mentioned, another method of stretching the time variable in
discrete time is to permit a scaling constant that is not an integer, but to
extend the definition of the signal to the new time indices. The most com-
mon example of this is to take the scale constant to be of the form a = 11M,
4.11. THE DERIVATIVE AND DIFFERENCE THEOREMS 187

where M is an integer. Here we begin with a signal 9 = {gnj n E Z} and

form a new signal, say h = {hnj n E Z} where h n = gn/M for those values
of n for which n/M is an integer and where we have to define h n for other
values of n. One common, simple definition is to make h n = 0 for all n that
are not multiples of M. This has the effect of producing a signal that has
every Mth sample equal to a sample from gn with Os in between. This is
the reverse process of downsampling and is called upsampling. An alterna-
tive (and perhaps more sensible) approach would be to interpolate between
the values. We consider zero filling, however, because of its simplicity. In-
terpolation could be accomplished by subsequent smoothing using a linear
filter.
If we upsample and fill with zeros, then observe that the Fourier trans-
form of his

L
00

H(f) = hne-i21f/n
n=-oo

L
00

= hnMe-i21f/nM
n=-oo
00. 11
= '"' 9 e-,21f/nM.
L...J n ,
f E [-- -)
2'2·
n=-oo

The latter formula resembles G(f), the Fourier transform of g, with a scale
change of the frequency range. Strictly speaking, G(f) is only defined for
f E [-t, t), but the right-hand side of the above equation has frequencies
f M which vary from [- ~ , ~). Thus H (f) looks like M periodic replicas
of a compressed G(f). In particular, if a(f) is the periodic extension of
G(f), then
- 1 1
H(f) = G(Mf)j f E [-2' 2)·

The case for M = 2 is depicted in Figure 4.7.

4.11 The Derivative and Difference Theorems

Suppose that 9 = {g(t)j t E 'R} is an infinite duration continuous time
signal that is everywhere differentiable and suppose that g' = {g'(t)j t E 'R}
is the signal defined by
g'(t) = dg(t).
dt
188 CHAPTER 4. BASIC PROPERTIES

G(f)

G(O)

f
1 1
4" 2

(a) DTFT of {gn; n E Z}

H(f)

f
1 1
4" 2

(b) DTFT of {h n = gn/2; n E Z}

Figure 4.7: DTFT of Upsampled Signal

Let G(f) denote the Fourier transform of g. Then if the signal is nice
enough for the Fourier inversion formula to hold and for us to be able to
interchange the order of differentiation and integration:

g'(t)
4.11. THE DERIVATIVE AND DIFFERENCE THEOREMS 189

= J00

G(f).!!..ei2trlt df
dt
-00

J
00

= G(f)(i211-j)e i2trlt df.

-00

From the inversion formula we can identify i27r fG(f) as the Fourier trans-
form of g', yielding the following result.

Theorem 4.13 The Derivative Theorem

Given a continuous time infinite duration everywhere differentiable sig-
nal 9 with Fourier transform G(f), then

F/(g') = i27rfG(f); fER. (4.48)

Similarly, if 9 is n times differentiable and if g(n) is the signal defined by

g (n) (t) = £.9J.!l.
dtn , then

(4.49)

As an example of the derivative theorem, consider the infinite duration

continuous time Gaussian pulse 9 = {e-trt 2; t E R}. The theorem implies
that the CTFT of the signal g'(t) = _27rte- trt2 is i27rfe-tr /2 .
The analogous discrete time result follows by replacing the derivative
with a difference, sometimes called a discrete time derivative. If 9 =
{gn; n E Z} is an infinite duration discrete time signal, define the difference
signal g' by
g'n = gn - gn-l·
In this case we have
00 00

n=-oo n=-oo

L L
00 00

= gne-i2trln - gne-i2trl(n+l)
n=-oo n=-oo

L
00

= (1 - e- i2tr I) gn e- i2tr In.

n=-oo

This yields the discrete time equivalent of the derivative theorem.

190 CHAPTER 4. BASIC PROPERTIES

Theorem 4.14 The Difference Theorem

Given a discrete time infinite duration signal 9 with Fourier transform
G(f), then
F/(g') = (1 - e- i21r /)G(f)j / E [-~, ~). (4.50)

Similarly, i/ g(k) is the kth order difference 0/ the signal g, then

(4.51)

Note that the nth order difference of a signal can be defined iteratively
by
gn(k) = g(k-l)
n
_ g(k-l)
n-l ,
e.g.,
(2) _ (1) (1)
gn - gn - gn-l - gn -
- 2gn-l + gn-2,
and so on.
The derivative and difference theorems are similar, but the dependence
on / in the extra term is different. In particular, the multiplier in the
discrete time case is periodic in / with period 1 (as it should be).
Both results give further transform pairs considering the dual results.
In the continuous time case we have that the Fourier transform of the signal
tg(t) is
i
F/({tg(t)j t E 'R}) = 27rG'(f)j f E 'R.

This can be proved directly by differentiating the formula for G(f) with
respect to / which results in a multiplier of -i27rt. The discrete time
equivalent is left as an exercise.

4.12 Moment Generating

This section collects a group of results often called moment theorems be-
cause they show how moments or averages in one domain can be evaluated
in the other domain. These results are useful as a tool for evaluating in-
tegrals of a particular kind and for studying the relation between "pulse
width" and "bandwidth" of a signal. The results are also useful simply
because they provide practice in manipulating Fourier sums and integrals.
Because of the importance of Fourier transforms in finding moments,
a variation of the Fourier transform used in probability theory is often
called the moment generating function. Such variations may differ from the
ordinary Fourier transform in the sign in the exponential and in the nature
4.12. MOMENT GENERATING 191

of the independent variable, but the basic ideas of evaluating moments from
transforms remain the same.
We focus on moments in the time domain, but duality can be invoked
to find similar formulas for frequency domain moments. We will consider
both continuous and discrete time, but we focus on the infinite duration
case for simplicity. It is often convenient to normalize moments, as we shall
do later.
We begin with the general definition of moments (unnormalized) and
we will then consider several special cases. Given a signal g, define for any
nonnegative integer n the nth order moment

M(n) ={
'J tng(t) dt
-00
continuous time
9
L
00
kngk discrete time
k=-oo

If 9 is a probability density function or probability mass function (a nonneg-

ative signal that integrates or sums to one, respectively) then the nth order
moment is the usual statistical nth order moment, the mean or average of
the nth power of the independent variable weighted by the signal.
The simplest example of a moment is the Oth order moment or area of
the signal:

M(O) ={
'J g(t) dt
-00
continuous time
9 00
L gk discrete time
k=-oo

This is easily recognized in both cases as being G(O), where G(f) is the
Fourier transform of g. Thus for both continuous and discrete time,

(4.52)

The dual of this result is similarly found to be

'J G(f) df continuous time

M G(O) -
-
{ -~
2' = g(O). (4.53)
J G(f) df discrete time
-!
Eq. (4.52) is called the area property.
Note the difference of appearance of the two results: both continuous
time and discrete time signals have continuous spectra and hence both
192 CHAPTER 4. BASIC PROPERTIES

spectral moments are integrals. Similar results could be obtained for the
finite duration counterparts, but that is left as an exercise.
As an example of the use of the area property consider the continuous
time signal g(t) = Jo(211't), a zeroth order ordinary Bessel function. From
the transform tables G(f) = ~. Thus from the area property
1f..jl-j2

1-0000 Jo(211't) dt = G(O) = -.

1
11'
Next consider the first moment. We treat the continuous time case;
the discrete time case follows in exactly the same way with sums replacing
integrals. Observe that if we differentiate the spectrum with respect to f
we have

G'(f) = ~ ([00 g(t)e-i21fftdt) = (-i211') roo tg(t)e-i21fftdt.

# 1- 00 1- 00

(This is the dual of the derivative theorem.) Thus choosing f = 0 we have

that
1-00
00 tg(t) dt = 2i G'(O).
11'
(4.54)

The above formula of course holds only if the derivative and integral can

i:
be interchanged. This is the case if the absolute moment exists,

Itg(t)1 dt < 00.

In general (continuous or discrete time) we have the following:

M(l)
9
= ~G'(O).
211'
(4.55)

This result is called the first moment property.

The importance of this result is that it is often easier to differentiate
than integrate. Thus if one knows the spectrum, it is easier to take a single
derivative than to integrate to find a moment.
As an example of the first moment property, consider the continuous
time signal g(t) = n(t - 3) and find J~oo tg(t) dt. We have that G(f) =
sinc(f)e- i61f f and hence

dG (f)
df
= -611'i sinc(f)e -6i1f f + e-6i". f ~
df
sinc(f)
.
4.12. MOMENT GENERATING 193

°
At f = 0, dSi:(f(f) = (any even function has zero slope at origin if the
derivative is well-defined) and therefore G' (0) = -i611" and hence

1-00
00 iG' (0) 611"
tg(t)dt = - - = - = 3.
211" 211"

n times and evaluate the result at °

This same procedure works for any moment: Differentiate the spectrum
f = to obtain the nth moment times
some constant. We now summarize these results:

Theorem 4.15 Moment Theorem

Suppose that 9 is an infinite duration signal which satisfies the following

i:
condition:

Itng(t)1 dt < ,
00· continuous time (4.56)
00

L Ikngkl
k=-oo
< , discrete time;
00· (4.57)

then
(4.58)

where

As an example of a proof, in the infinite duration continuous time case

we have that

= ~
dfn
J 00

g(t)e-i21rft dt
-00

J
00

(_211"it)ng(t)e-i21rft dt
-00

J
00

= (-211"i)n tng(t)e-i21rftdt.
-00

°
Setting f = then yields the theorem. Not setting f = 0, however, leaves
one with another result of interest, which we state formally.
194 CHAPTER 4. BASIC PROPERTIES

Theorem 4.16 Given a continuous time infinite duration signal get) with
spectrum G(f),

(4.59)

Similar results can be proved for other signal types.

In the special case n = 2 the second moment is called the moment of
inertia and the moment theorem reduces in the continuous time case to

called the second moment property. This result has an interesting impli-
cation. Suppose that the signal is nonnegative and hence can be thought
of as a density (say of mass or probability) so that the moment of inertia
can be thought of as a measure of the spread of the signal. In other words,
if the second moment is small the signal is clumped around the origin. If
it is large, there is significant "mass" away from the origin. The above
second moment property implies that a low moment of inertia corresponds
to a spectrum with a low negative second derivative which means a small
curvature or relatively flat behavior around the origin. Correspondingly,
a large moment of inertia means that the spectrum has a large curvature
at the origin and hence is very "peaky". Thus a signal with a peak at the
origin produces a spectrum that is very flat at the origin and a signal that
is very flat produces a spectrum that is very steep. This apparent tradeoff
between steepness in one domain and flatness in another will be explored
in more depth later.
As an example that provides a simple evaluation of an important inte-
2
gral, consider the continuous time signal g(t) = e- 1ft which has spectrum
G(f) = e- 1f / 2 • We have that

G'(f) = _27r!e- 1f / 2

which implies from the moment theorem that

1 -00
00 2
e- 1ft dt = G(O) = 1,

1 00

-00
2
te- 1ft dt
.

= -G'
Z

27r
(0) = 0,
4.12. MOMENT GENERATING 195

(since the integrand is odd) and

f oo
-00
t2e-trt2 dt = __I_G II (O) = ~.
411"2 211"
As a second example and an object lesson in caution when dealing with
moments, consider the continuous time signal g(t) = 2/(1 + (211"t)2) and its

I:
transform G(f) = e- 1fl . The Oth moment is

g(t) dt = G(O) = 1.

If one tries to find the first moment, however, one cannot use the moment
theorem because G' is not continuous at O! In particular, G' (0+) = -1 and
G'(O-) = +1. The problem here is that the integrand in

f oo

-00
tg(t) dt =
foo
-00
1
+
2t
(2 )2 dt
1I"t

falls as lit for large t, which is not integrable. In other words, the first
moment blows up. The integral does exist in a Cauchy sense (it is 0). In
fact, this signal corresponds to the so-called Cauchy distribution in proba-
bility theory. Note that it violates the sufficient condition for the moment
theorem to hold, i.e., Itg(t)1 is not integrable.
As a final example pointing out a more serious peril, consider the signal
sinc(t). Its spectrum is n(f) which is infinitely differentiable at the origin
and the derivatives are all o. Thus one would suspect that the moments
are all 0 (except for the area). This is easily seen to not be the case for the
second moment, however, by direct integration. Integration by parts shows
that in fact

f -T
T t2 sinc(t) dt = sin T - Tcos T

does not converge as T -t 00 and hence the second moment does not exist.
The problem is that the conditions for validity of the theorem are violated
since the second absolute moment does not exist. Thus existence of the
derivatives is not sufficient to ensure that the formula makes sense. In
order to apply the formula, one needs to at least argue or demonstrate by
other means that the desired moments exist.

* Normalized Moments
Normalizing moments allows us to bring out more clearly some of their
basic properties. If the signal is nonnegative, the normalized signal can
196 CHAPTER 4. BASIC PROPERTIES

be thought of as a probability or mass density function in the continuous

time case and a probability mass function or point mass function in the
discrete time case. The normalized moments are obtained by replacing the
weighting by the signal 9 in the sum or integral by a signal having unit area;
that is, the weighting is g(t)/ I~oo g(a) da in the continuous time case and
gn/ L:~-oo gk in the discrete time case. Clearly this normalization makes
sense only when the signal has nonzero area. Normalizing in this way is
equivalent to dividing the nth moment by the Oth order moment (the area
of the signal). To be specific, define the normalized moments < t n >g by
M(n)
< tn - 9
>g-~. (4.60)
Mg
For both continuous and discrete time,

The normalized first moment is called the centroid or mean of the signal.
In the continuous time case this is given by

where the moment theorem has been used. The second moment is called
the mean squared abscissa and it is given in the continuous time case by

In addition to normalizing moments, they are often centralized in the

sense of removing the area or mean before taking the power. The principal
example is the variance of a signal, defined by

The variance or its square root, the standard deviation, is often used as
a measure of the spread or width of a signal. If the signal is unimodal,
then the "hump" in the signal will be wide (narrow) if the variance is
large (small). This interpretation must be made with care, however, as the
variance may not be a good measure of the physical width of a signal. It
can be negative, for example, if the signal is not required to be nonnegative.
When computing the variance, it is usually easier to use the fact that

(4.61)
4.13. BANDWIDTH AND PULSE WIDTH 197

which can be proved by expanding and manipulating the definition.

We will see later in this chapter that any real even signal will have a

°
real and even spectrum. Since an even function has 0 derivative at time 0,
this means that < t >g= for any real even signal. If the centroid is 0,
then the variance and the mean squared abscissa are equal.

4.13 Bandwidth and Pulse Width

We have several times referred to the notions of the width of a signal or
spectrum. While the width of a signal such as n(t) or I\(t) is obvious,
a meaningful definition of width for an arbitrary non-time-limited signal
(or an arbitrary non-band-limited spectrum) is not obvious. Intuitively,
the width of a signal or spectrum should measure the amount of time or
frequency required to contain most of the signal or spectrum. When con-
sidering the second order moment property, we observed that there is a
tradeoff between width in the two domains: A narrow (broad) signal cor-
responds to a broad (narrow) spectrum. This observation can be further
quantified in the special case of the rectangle function. From the continuous
time stretch (or similarity) theorem, the transform of n(t/T) is Tsinc(T I).
The box has a width of T by any reasonable definition. The width of the
sinc function is at least indicated by the difference between the first two
zeros, which occur at -liT and liT. Thus if we increase (decrease) the
width of the time signal, the spectrum decreases (increases) in proportion.
The object of this section is to obtain a general result along these lines
which provides a useful quantitative notion of time width and band-width
and which generalizes the observation that the two widths are inversely
proportional: wide (narrow) signals yield narrow (wide) spectra.
We have already encountered in Section 4.5 applications where the no-
tion of width is important: band-limited spectra are required for the sam-
pling theorem and time-limited signals can be expanded in a Fourier series
over the region where they are nonzero. Both of these results required
absolute time or band limitation, but a weaker notion of time-limited or
band-limited should allow these results to at least provide useful approxima-
tions. Other applications abound. If a narrow signal has a wide spectrum,
then a quantitative measure of the signal width is necessary to determine
the bandwidth of a communications channel necessary to pass the signal
without distortion. An important attribute of an antenna is its beamwidth,
a measure of its angular region of maximum sensitivity. Resolution in radar
is determined by pulse width.
As one might guess, there is not a single definition of width. We shall
consider a few of the simplest and most commonly encountered definitions
198 CHAPTER 4. BASIC PROPERTIES

along with some properties. We will introduce the definitions for signals,
but they have obvious counterparts for spectra. The entire section con-
centrates on the case of continuous time infinite duration signals so that
indices in both time and frequency are continuous. The notion of width for
discrete time or discrete frequency is of much less interest.

Equivalent Width
The simplest notion of the width of a signal is its equivalent width defined as
the width of a rectangle signal with the same area and the same maximum
height as the given signal. The area of the rectangle n(t/T) is T. Given
a signal 9 with maximum height gmax and area J~oo g(t) dt = G(O) (using
the moment property), then we define the equivalent width Wg so that the
rectangular pulse gmax n (t/Wg) has the same area as g; that is, gmaxWg =
G(O). Thus
Wg = G(O).
gmax
In the special but important case where g(t) attains its maximum at the
origin, this becomes
G(O)
Wg = g(O).
An obvious drawback to the above definition arises when a signal has
zero area and hence the width is zero. For example, the signal n(t -1/2) -
n(t + 1/2) is assigned a zero width when its actual width should be 2.
Another shortcoming is that the definition makes no sense for an idealized
pulse like the impulse or Dirac delta function.
The equivalent width is usually easy to find. The equivalent width of
n(t) is 1, as are the equivalent widths of I\(t) and sinc(t). These signals
have very different physical widths, but their equal areas result in equal
equivalent widths.
When used in the time domain, the equivalent width is often called the
equivalent pulse width. We shall also use the term equivalent time width.
The same idea can be used in the frequency domain to define equivalent
bandwidth:
Wa = J~oo G(f) df = g(O) .
Gmax G max
In the common special case where G(f) attains its maximum value at the
origin, this becomes
g(O)
Wa = G(O)·
4.14. SYMMETRY PROPERTIES 199

Note that even if GU) is complex, its area will be real if get) is real. Note
also that if the signal and spectra achieve their maxima at the origin, then

and hence the width in one domain is indeed inversely proportional to the
width in the other domain.

Magnitude Equivalent Width

In the communications systems literature, equivalent bandwidth is usually
defined not in terms of the area under the spectrum, but in terms of the
area under the magnitude spectrum. To avoid confusion, we will call this
the magnitude equivalent bandwidth and denote it by

Beq = J~oo IG(f)1 df .

Gmax
We again assume that Gmax = G(O), which is reasonable for some baseband
pulses. Defining the pulse width as before we still have a simple relation
between pulse and bandwidth, except that now it is an inequality instead
of an equality. In particular, since for all t

f
00

Ig(t)1 = I G(f)ei21flt dfl

-00

f
00

< IGU)ei21fltl df
-00

f
00

= IGU)I df
-00

= BeqGmax,
then choosing t so that get) = gmax we have that

B eq > gmax _ 1
--.
- G max Wg

4.14 Symmetry Properties

The remaining simple properties have to do with the symmetry properties of
Fourier transforms. These properties are useful for checking the correctness
200 CHAPTER 4. BASIC PROPERTIES

of Fourier transforms of signals since symmetry properties of the signals

imply corresponding symmetries in the transform. They are also helpful
for occasionally suggesting shortcuts to computation.
We here focus on the infinite duration CTFT; similar properties are
easily seen for the infinite duration DTFT.
The Fourier transforms of even and odd signals have special symmetries.
The following result shows that all signals can be decomposed into even and
odd parts and the signal is a linear combination of those parts.
Theorem 4.17 Any signal g(t) can be uniquely decomposed into an even
part and an odd part; that is,
g(t) = ge(t) + go(t)
where ge(t) is even and go(t) is odd.
To get such a representation just define
1
ge(t) = "2(g(t) + g( -t)) (4.62)
1
go(t) = "2(g(t) - g(-t)) (4.63)

By construction the two functions are even and odd and their sum is g(t).
The representation is unique, since if it were not, there would be another
even function e(t) and odd function o(t) with g(t) = e(t) + o(t). But this
would mean that
e(t) + o(t) = ge(t) + go(t)
and hence
e(t) - ge(t) = go(t) - o(t).
Since the left-hand side is even and the right-hand side is odd, this is
only possible if both sides are everywhere 0; that is, if e(t) = ge(t) and
o(t) = go(t).

Remarks
1. The choices of ge(t) and go(t) depend on the time origin, e.g., cost is
even while cos( t - 7r /2) is odd.
2. I~CXJ go(t) dt = 0, at least in the Cauchy principal value sense. It
is true in the general improper integral sense if go(t) is absolutely
integrable. (This problem is pointed out by the function
t>O
sgn{t) ={ ~-1 t=O (4.64)
t<O
4.14. SYMMETRY PROPERTIES 201

which has a 0 integral in the Cauchy principal value sense but which
is not integrable in the usual improper Riemann integral sense. The
function get) = lit is similarly unpleasant.)
3. If el (t) and e2 (t) are even functions and 01 (t) and 02 (t) are odd func-
tions, then el(t) ± e2(t) is even, 01(t) ± 02(t) is odd, el(t)e2(t) is
even, 01 (t)02(t) is even, and edt)02(t) is odd. The proof is left as an
exercise.
4. All of the ideas and results for even and odd signals can be applied
to infinite duration two-sided discrete time signals.
We can now consider the Fourier transforms of even and odd functions.
Again recall that get) = ge(t) + go(t), where the even and odd parts in

1:
general can be complex. Then

1:
G(f) = g(t)e-i21Tftdt

1:
= (ge(t) + go(t)) (cos(27rft) - isin(27rft)) dt

1 : ge(t) cos(27r ft)dt - i ge(t) sin(27r ft) dt +

[ : go(t) cos(27r ft) dt - i [ : go(t) sin(27r ft) dt.

Since the second and third terms in the final expression are the integrals of
odd functions, they are zero and hence

G(f) = 1:ge(t)COS(27rft)dt-i[:9o(t)Sin(27rft)dt
= Ge(f) + Go(f), (4.65)
where Ge(f) is the cosine transform of the even part of get) and Go(f) is -i
times the sine transform of the odd part. (Recall that the cosine and sine
transforms may have normalization constants for convenience.) Note that
if get) is an even (odd) function of t, then G(f) is an even (odd) function
of f.
As an interesting special case, suppose that get) is a real-valued sig-

i:
nal and hence that ge(t) and go(t) are also both real. Then the real and
imaginary parts of the spectrum are immediately identifiable as

i:
lR(G(f)) = ge(t)cos(27rft)dt (4.66)

SS(G(f)) = - go(t) sin(27rft) dt. (4.67)

202 CHAPTER 4. BASIC PROPERTIES

Observe that the real part of G(f) is even in f and the imaginary part of
G(f) is odd in f. Observe also that if g(t) is real and even (odd) in t, then
G(f) is real (imaginary) and even (odd) in f.
If g(t) is real valued we further have that

G( - J) = ~(G( - J)) + m(G( - J))

= ~(G(f)) - i~(G(f))
= G*(f),

which implies that the Fourier transform of a real-valued signal is Hermi-

tian.
By a similar analysis we can show that if the signal g(t) is purely imag-

I:
inary and hence ge(t) and go(t) are imaginary, then

~(G(f)) = -i go(t) sin(27rtJ) dt

I:
is odd in f, and

<:J(G(f)) = -i ge(t) cos(27rtJ) dt

is even in f, and
G(-J) = -G*(f)j (4.68)
that is, the spectrum is anti-Hermitian.
We can summarize the symmetry properties for a general complex signal
g(t) as follows:

g(t)= ge(t) + go(t) = eR(t) + ieI(t) + OR(t) + iOI(t) (4.69)

G(f) = Ge(f) + Go(f) = ER(f) + iEI(f) + OR(f) + iOl(f) (4.70)

where

g(t) ~ G(f)
ge(t) ~ Ge(f)
go(t) ~ Go (f)
eR(t) ~ ER(f) (4.71)
el(t) ~ El(f)
OR(t) ~ iOI(f)
01(t) ~ -iOR(f)
4.15. PROBLEMS 203

These formulas can be useful in computing Fourier transforms of compli-

cated signals in terms of simpler parts. They also provide a quick check on
the symmetry properties of Fourier transforms. For example, the transform
of a real and even signal must be real and even and the transform of an
odd and real signal must be odd and imaginary.

4.15 Problems
4.1. What is the DFT of the signal

gn = sin(27r3i) - sin(27ri); n = 0,1, ... , 7?

4.2. Prove the shift theorem for infinite duration discrete time signals and
finite duration continuous time signals.
4.3. Let 9 = {gn; n E Z}, where Z is the set of all integers, be defined by

n = 0, 1,2,···
otherwise
Find the Fourier transform G of g.
4.4. Prove that the Fourier transform of the infinite duration continuous
time signal {g(at - b); tEn} is 'farG(f /a)e-i21rfb/a, where G is the
Fourier transform of g.
4.5. Find the Fourier transform of the following continous time infinite
duration signal: g(t) = e- 1t - 3I j tEn. Repeat for the discrete time
case (now t E Z).
4.6. Suppose that g(t) and G(f) are infinite duration continuous time
Fourier transform pairs. What is the transform of cos2(27rfot)g(t)?
4.7. What is the Fourier transform of {sinc(t)cos(27rt); tEn}?
4.8. State and prove the following properties for the two-dimensional finite
duration discrete time Fourier transform (the two-dimensional DFT):
Linearity, the shift theorem, and the modulation theorem.
4.9. Given a discrete time, infinite duration signal gn = rn for n ~ 0
and gn = 0 for n < 0 with Irl < 1, suppose that we form a new
signal h n which is equal to gn whenever n is a multiple of 10 and is
o otherwise: hn = gn for n = ... , -20, -10,0,10,20, .... What is the
Fourier transform H(f) of hn ?
204 CHAPTER 4. BASIC PROPERTIES

It was pointed out that the true meaning of "decimation" is not to

produce the sequence h n (Le., the dead centurions in the Roman use
of the word), but to produce the remaining or live centurions. Thus
the correct meaning of "decimating" the sequence Un is to produce
the sequence In for which
f = {gn if n is not a multiple of 10
n 0 if n is a multiple of 10
Find the Fourier transform F(f) of In.
4.10. Suppose you have two discrete time signals {h n ; n = 0, 1, ... ,N -I}
and {gn; n = 0,1, ... , N - I} with OFTs H(k/N) and G(k/N), k =
0,1, ... ,N -1, respectively. Form a new signal {w n ; n = 0, 1, ... ,2N-
I} by "multiplexing" these two signals to form ho, go, h1' gl, ... ,
hN-logN-1; that is,
if n is zero or even
if n is odd

What is the OFT W(l/2N); l = 0,1, ... , 2N - 1 in terms of Hand

G?
4.11. Prove Parseval's equality for energy for the continuous time finite
duration case.
4.12. Prove Parseval's equality for energy for the discrete time infinite du-
ration case.

i:
4.13. Evaluate the integral

sinc[2B(t - 2~)] sinc[2B(t - ;)] dt (4.72)

for all integers n, m. What does this say about the signals sinc(2B(t-
2'1 )) for integer n?
4.14. This problem considers a simplified model of "oversampling" tech-
niques used in CD player audio reconstruction.
A continuous time infinite duration signal 9 = {g(t); t E 'R} is band-
limited to ±22 kHz. It is sampled at 10 = 44 kHz to form a discrete
time signal U= {Un; n E Z}, where Un = g(n/lo).
A new discrete time signal, h = {h n ; n E Z} is formed by repeating
each value of U four times; that is,

h4n = h4n+1 = h4n+2 = h4n+3 = Un; n E Z.

 4.15. PROBLEMS 205

Note that this step can be expressed in terms of the upsampling op-
eration: If we define the signal T = {Tn; n E Z} by

_ {9n/4 if n = 0, ±4, ±8, ...

Tn - 0 if n is not a multiple of 4

then

Lastly, a continuous time signal h = {h(t); tEn} is formed from h

using a sampling expansion:

h(t) = r: h
00

n=-oo
n sinc(176000t - n).

(a) Find H(f); fEn in terms of G(f); fEn.

(b) If g(t) = cos20001rt + cos40001rt, what frequencies have nonzero
values of H(f)?
(c) If h(t) is then filtered to ±22 kHz by an ideal low pass filter (i.e.,
its Fourier transform is multiplied by a unit magnitude box filter
having this bandwidth), is the result an exact reproduction of
g(t)? Why or why not?
4.15. Suppose that a discrete time signal 9 = {gn; n E Z} is not band
limited so that it cannot be recovered from its downsampled version
h = {g2n; n E Z}. Suppose that we also form a second downsampled
signal T = {g2n+1; n E Z} which contains all of the samples in 9
that are not in h. Find a formula for G in terms of Hand R. Thus
if we form all possible distinct downsampled versions of a signal and
compute their transforms, the transforms can be combined to find the
transform of the original signal. The construction strongly resembles
that used for the FFT.

4.16. Develop an analog of the discrete time sampling theorem for the DFT,
that is, for finite duration discrete time signals.
4.17. The DFT of the sequence {go,gl, ... ,gN-d is {Go,Gl, ... ,GN-d.
What is the DFT of the sequence {gO, O,gl, 0, ... , gN-l, O}?
4.18. If 9 = {gn; n = 0,1, ... , N - I} has Fourier transform
N-l
G(f) = L gne-i2rrfn,
n=O
206 CHAPTER 4. BASIC PROPERTIES

find an expression for

N-l
Lngn
n=O
in terms of GU). (Hint: Stick with the Fourier transform here, not
the DFT, i.e., do not restrict the range of f at the start. This problem
makes the point that sometimes it is useful to consider a more general
frequency domain than that required for inversion.)
4.19. Suppose you have an infinite-duration discrete-time signal x = {Xk j k E
Z} with Fourier transform X = {XU)j f E [-1/2,1, 2)}. A second
discrete-time signal Y = {Ykj k E Z} is defined in terms of x by

Yk = Xk + Xk-l + Xk-2 + Xk-3 + Xk-4 + Xk-5 + Xk-6 + Xk-7·

(This is an example of what we will later call a discrete-time lin-
ear filter, but you do not need any linear filtering theory to do this
problem.)
(a) Find the Fourier transform Y of Y for the case where
Xk = e-ku_l(k)j k EZ

where U-l (k) is the unit step function defined as 1 for k ~ 0 and
o otherwise.
(b) Form the truncated signal f) = {Ykj k E Z16} defined by Yn = Yn
for n E Z16. Find the Fourier transform Y of Yfor the case where
x is defined by
Xk = sin(21r~:) k E Z.

4.20. Suppose that 9 = {g(n)j n E Z} is a discrete time Signal with Fourier

transform G = {G(f)j f E [0, I)}, that is not necessarily bandlimited.
For a fixed integer M define the subsampled signal 'Y = b(n) =
g(nM)j n E Z}. Find the DTFT r of 'Y.
Hint: Analogous to the continuous time case, consider the function
M-l
a(f) =L G(f ~m),
m=O

where the shift inside G is cyclic on the frequency domain, e.g., using
[0,1) as the frequency domain,
4.15. PROBLEMS 207

Expand 0: in a Fourier series on [0,1) and compare it to the definition

of 'Y.

4.21. Suppose g = {g(t); t E R} where

g(t) = sinc2(~).
(a) What is G(f)? (Give an explicit expression.) Make a sketch
labeling points of interest in frequency and amplitude.
(b) The signal g is now sampled at the Nyquist rate and a discrete
time sequence h n is formed with the samples: h n = g(nTs).
What is the transform of the sampled sequence? (Give an ex-
plicit expression for H(f) for the given g.) Make a labeled sketch.
(c) h n is then upsampled by 5 to form

_ {hn/5 if n is an integer multiple of 5

Vn -
o otherwise
Give an explicit expression and a labeled sketch of the transform
V (f) of the signal V n ·
(d) Suppose we want to obtain the sequence Wn which is formed by
sampling the continuous time signal g with a sampling period
Ts = 1.2. How can this be done given the sequence Vn alone?
Make a block diagram of your procedure.
(e) Now the system you've designed in part (d) is altered so that
h n is upsampled by 2 instead of 5. Can the resulting sequence
at the output be used to reconstruct the signal g? Why or why
not?

4.22. Consider the 9 point sequence g = {12 0120121}. Define the Fourier
transform (DFT) of g to be G = {G(k/9); k = 0,1, ... ,8}.
Let 9 be the periodic extension of g. Define two new sequences:
h n = r n g"nu_l(n), Irl < 1; and Vn = h 3n + 1 .

(a) What is G(1/3)?

(b) What is V(f)?
(c) What is H(f)?

4.23. Sketch and find the Fourier transform of the infinite duration dis-
crete time signal gn = 2D4(n) - D2(n). What is the inverse Fourier
transform of the resulting spectrum?
208 CHAPTER 4. BASIC PROPERTIES

4.24. Given a continuous time infinite duration signal {pet); t E R} with

Fourier transform P(f), find the Fourier transform of the signal g( t) =
peT - t), where T is fixed. Specialize your result for the cases pet) =
= =
net) and pet) /\(t) and T 1 (e.g., what is the transform of n(l -
t)).
4.25. What is the Fourier transform of the signal 9 = {gn = (1/2)n +
(1/3)n-l; n = 0,1, ... }.
4.26. Suppose that you have a discrete time signal 9 = {g(n); n E Z} with
a Fourier transform G that is bandlimited:
1 1
G(f) = 0; for "2 2: III 2: S·
(a) Find the Fourier transform H ofthe signal h defined by
1l"n
hen) = g(n) cos(""4); n E Z.

(b) Derive the Fourier transform of the upsampled signal gl/3 defined
by
( ) _ { g( i) if n is an integer multiple of 3
gl/3 n - 0 otherwise
(c) Find the Fourier transform of the signal w defined by
Wn = gl/3(n) - g(n).
Compare this signal and its Fourier transform with h and its
Fourier transform.
4.27. An infinite duration continuous time signal 9 = {get); t E R} is band-
limited to (- W, W), i.e., its Fourier transform G satisfies
G(f) = 0 for III 2: w.
We showed that we can expand G in a Fourier series on [- W, W) for
this case. Use this expansion to find an expression for

[ : IG(fWdl

in terms of the samples genT).

4.28. A continuous time signal 9 = {get); t E R} has a spectrum G =
{G(f); fER} defined by
~ I E [-~,~]
G(f) = { 1 IE [-l,-t) or I E (t,l]
o otherwise
4.15. PROBLEMS 209

(a) Find g.
(b) Write a sampling expansion for g using a sampling period of T
and state for which T the formula is valid.
For the remainder of this problem assume that T meets
this condition.
(c) Find the DTFT r for the sampled sequence 'Y = bn = g(nT}j nE
Z}.
(d) Evaluate the sum

L
00

g(nT}.
n=-oo

(e) Prove or disprove

L g2(nT} = 1-
00

n=-oo
[00

-00
l(t} dt

Keep in mind that in Parts (c}-(e) it is assumed that T meets the

condition of (b) necessary to ensure the validity of the sampling the-
orem.
4.29. Suppose that a bandlimited continuous time signal y(t} is sampled
faster than the Nyquist rate to produce a discrete time signal x(n} =
y(nT}. Relate the energy of the discrete time signal, I: Ix(n)j2, to
that of the continuous time process, f ly(t}12 dt.
4.30. An infinite duration continuous time signal g = {g(t)j t E R} is band-
limited to (-1/2,1/2), Le., its Fourier transform G satisfies
1
GU} = 0 for If I ~ 2'
A discrete time signal U = {Unj n E Z} is then defined by sampling
as Un = g(n}. A new continous time signal h = {h(t}j t E R} is then
defined by
L
00

h(t} = Unp(t - n},

n=-oo

where p = {p(t}j t E R} is a continous time signal (not necessarily a

pulse or time limited) with Fourier transform P.
Find an expression for the Fourier transform H of h in terms of G
and P. Specialize your answer to the cases where p(t} = sinc(t}
and p(t} = 01/2(t}. Both of these results are forms of PAM, but you
210 CHAPTER 4. BASIC PROPERTIES

should find that h is trivially related to 9 in one case and has a simple
relation in the other. Note that neither of these p are really physical
pulses, but both can be approximated by physical pulses.
4.31. What is the Fourier transform of 9 = {te- rrt2 ; t E R}?
4.32. Define the signal 9 = {g(t); t E R} by
I + e- 1tl It I < ~
g(t) = { e- 1tl
It I ~ ~.
(a) Find the Fourier transform G of g.
Write a Fourier integral formula for 9 in terms of G. Does the
formula hold for all t?
(b) What signal h has Fourier transform {H(f); fER} given by
H(f) = 2G(f) cos(87rf)?
Provide a labeled sketch of h.
(c) Define the truncated finite duration signal 9 = {g(t); t E [-1/2, 1/2)},
where g(t) = g(t) for t E [-1/2,1/2). Find the Fourier trans-
form 6 of 9 and write a Fourier series representation for g. Does
the Fourier series give [} for all t E [-1/2, 1/2)?

4.33. (a) Define the discrete time, finite duration signal 9 = {go, gl, g2, g3, g4, gs}
by
9 = {+1, -1, +1, -1, +1, -I}
and define the signal h by
h= {+1,-1,+1,+1,-1,+1}.
Find the DFTs G and H of 9 and h, respectively. Compare and
contrast G and H. (Remark on any similar or distinct proper-
ties. )
(b) Define the continuous time finite duration pulse p = {p(t); t E
[0,6)} by

p
(t)
otherwise
Find the Fourier transform P of p.
°
= {10::; t < 1 .

(c) Let g, h, and p be as above. Define the continuous time finite

duration signals 9 = {g(t); t E [0,6)} and h = {h(t); t E [0,6)}
by
5
[}(t) =L gnp(t - n); t E [0,6)
n=O
4.15. PROBLEMS 211

5
h(t) =L hnP{t - n); t E [0,6).
n=O

Find the Fourier transforms G and iI of 9 and h. How do these

transforms differ from each other? How do they differ from the
results of the previous part?
Hint: First try to find the required transforms in terms of P
without plugging in the details from (b).
4.34. Consider a finite duration real-valued continuous time signal g
{get); t E [0, I)} with Fourier transform G = {G(f); fEZ}.

(a) Find an expression for the energy

in terms of G.
(b) Suppose now that

G(f) = r- I/I ; fEZ, (4.73)

°
where r > is a real parameter. What is g? (Your final answer
should be a closed form, not an infinite sum.) What did you
have to assume about r to get this answer?
(c) Given g as in part (b) of this problem, evaluate

11 g(t) dt.

(d) Given g as in part (b) of this problem, evaluate

L
00

G(k).
k=-oo

(e) Given g as in part (b) of this problem, evaluate the energy

4.35. State and prove the moment theorem for finite duration discrete time
signals.
212 CHAPTER 4. BASIC PROPERTIES

4.36. State and prove the moment theorem for finite duration continuous
time signals.
4.37. For the function get) = A(t) cos(1I"t)j tEn, find

(a) The Fourier transform of g.

(b) The equivalent width of g.

4.38. For the function g(t) = e-1t1j t E 'R, find

(a) the area of g.
(b) The first moment of g.
(c) The second moment of g.
(d) The equivalent width of g.

4.39. Suppose that

get) = ge(t) + go(t)
J(t) = Je(t) + Jo(t)
where the subscripts e and 0 denote even and odd parts, respectively.

(a) Find expressions for

and
I: g(t)g( -t) dt

in terms of integrals involving ge and go.

(b) If h(t) = g(t)J(t), find expressions for ho and he in terms of the
even and odd parts of f and g.

4.40. Find the odd and even parts of the following signals (T = 'R):

(a) eit
(b) e- it H(t) (Where H(t) is the Heaviside step function.)
(c) It Isin(t - 11"/4)
(d) ei 1l" sin(t)

4.41. Find the even and odd parts of the continuous time infinite duration
signals
4.15. PROBLEMS 213

(a) get) = Itl! cos(t -~)

(b) get) = ei1r (t-to)2

4.42. Suppose that we decompose an infinite duration continuous time sig-

nal get) into its odd and even parts: get) = ge(t) + go(t) and suppose
that get) f+ G(f) form a Fourier transform pair. What is the Fourier
transform of G(t) in terms of ge and go?

4.43. Is it true that the magnitude spectrum IG(f)1 of a real signal must
be even?

4.44. Match the signals in the first list with their corresponding DFT's in
the second. The DFT's are rounded to one decimal place.

Hint: Very little computation is required here!

A (1,2,3,4,3,4,3,2)
B (0,2,3,4,0, -4, -3, -2)
C (i,3i,2i,4i,3i,4i,2i,3i)
D (1,2,2,3,3,4,4,3)
E (0,3,2, 5i, 0, -5i, -2, -3)
1 (22, -3.4 + 3.4i, -2, -.6 - .6i, -2, -.6 + .6i, -2, -3.4 - 3.4i)
2 (22i, -3.4i, 0, -.6i, -.6i, -.6i, 0, -3.4i)
3 (22, -4.8, -2, .8, -2, .8, -2, -4.8)
4 (0,7.1- 8.2i, -10 - 6i, 7.1- .2i, 0, -7.1 + .2i, 10 + 6i, -7.1 + 8.2i)
5 (0, -14.5i, 4i, -2.5i, 0, 2.5i, -4i, 14.5i)

4.45. Match the signals in the first list with their corresponding DFT's in
the second. The DFT's are rounded to one decimal place.

Hint: Very little computation is required here: no programmable

calculators allowed on this problem!
214 CHAPTER 4. BASIC PROPERTIES

A (-1,-1,-3,4,1,4,-3,-1)
B (0,1, -3, 4, 0, -4,3, -1)
C (j, -j, -4, 2,j, -2,4, -j)
D (6,-4,j,-2,6,-2,-j,-4)
E (0, j, 4j, -5j, 0, -5j, 4j, j)
F (-2,3,j, -2, 1, -2, -j,3)
G (2 + j, 0, 0, 0, -2 - j, 0, 0, 0)
1 (0, 8.5j, -8j, -8.5j, 16j, -8.5j, -8j, 8.5j)
2 (0, -l.lj, 6j, -13.1j, 0, 13.1j, -6j, 1.1j)
3 (0, -9.1,6,5.1, -12, 5.1, 6, -9.1)
4 (0,4+ 2j, 0,4 + 2j,0,4 + 2j,0,4 + 2j)
5 (0, 3.8j, 6j, -9.4j, 4j, 12.2j, -2j, -6.6j)
6 (1,6.1, -1, -12.1, -3, -8.1, -1,2.1)
7 (0, -0.8, 12,0.8,24,4.8,12, -4.8)
4.46. Table 4.1 has two lists of functions. For each of the functions on the
left, show which functions are Fourier transform pairs by means of an
arrow drawn between that function and a single function on the right
(as illustrated in the top case).
4.47. What can you say about the Fourier transform of a signal that is
(a) real and even?
(b) real and odd?
(c) imaginary and even?
(d) complex and even?
(e) even?
(f) odd?
4.15. PROBLEMS 215

{g(t); t E R} {G(f); fER}

{gn; n E Z} {G(f); f E [0, I)}
{g( -t); t E R} {G*(f); fER}
{g*(t); t E R} {G(-f); fER}
{g-n; n E Z} {G* ( - f); fER}
{g~; n E Z} {G*(f); f E [0, I)}
{G(t); t E R} {G( - f mod 1); f E [0, I)}
{G(-t); t E R} {G*( - f mod 1); f E [0, I)}
{G*( -t mod 1); t E [0, I)} {g(f); fER}
{G(-t mod 1); t E [0, I)} {g(-f); fER}
{g*(f); fER}
{g*(-f); fER}
{gk; k E Z}
{g-k; k E Z}
{gk; k E Z}
{g:'k; k E Z}

Table 4.1: Symmetry Properties

Chapter 5

Generalized Transforms
and Functions

In Chapter 2 the basic definitions of Fourier transforms were introduced

along with sufficient conditions for the transforms to exist. In this chap-
ter we extend the Fourier transform definitions to include some important
signals violating the sufficient conditions seen thus far. We shall see that
in some cases the extension is straightforward and simply replaces the ba-
sic definitions by a limiting form. In other cases the definitions cannot be
patched up so easily and we need to introduce more general ideas of func-
tions or signals in order to construct useful Fourier transform pairs. The
emphasis in this chapter is on continuous time, since that is where most of
the difficulties arise.

5.1 Limiting Transforms

In this section we treat by example a class of processes for which the Fourier
transform defined previously does not exist, yet a simple but natural trick
allows us to define a meaningful Fourier transform that behaves in the de-
sired way. The trick is to express the signal as a limit of better behaved
signals which have Fourier transforms in the original sense. If the trans-
forms of these signals converge, then the limiting transform is a reasonable
definition of the transform of the original signal. A primary example of this
technique is the infinite duration CTFT of the signum signal
t>O
g(t) = sgn(t) = { ~ t=O (5.1)
-1 t<O
218 CHAPTER 5. GENERALIZED TRANSFORMS AND FUNCTIONS

This signal clearly violates the absolute integrability criterion since the
integral of its absolute magnitude is infinite. Can a meaningful transform be
defined? One approach is to consider a sequence of better behaved signals
that converge to g(t). If the corresponding sequence of Fourier transforms
also converges to something, then that something is a candidate for the
Fourier transform of g(t) (in a generalized sense). One candidate sequence
is
t>O
t =0 j k = 0, 1, .... (5.2)
t<O
The signal sequence is depicted in Figure 5.1 for k = 1,10,100 along with
the step function. These signals are absolutely integrable and piecewise

0.8

0.6 , '.",
..... ...... ....
\
,
0.4 . . \ ··'>"<k=lO
0.2
\ "
\k=t·'~<:..~_
---
\'
~ \
0 . . ..... ,:-:- -----;,...-------.:....------
~ ----------:--...T...~-------:---'.'<
~
t;;
·0.2 -- .. .. ..... \
............... ,: I
-0.4 .. ~.....
"
..1..
,
.'.,... ',..
-0.6 \.
, ,
....... \, ..
'.
\
-1f----~--~----'~·

-30 ·20 -10 10 20 30

Figure 5.1: Function Sequence: The solid line is the sgn signal.

smooth and
lim gk(t) = g(t). (5.3)
k .... oo

The CTFT of gk(t) is

(5.4)
 5.2. PERIODIC SIGNALS AND FOURIER SERIES 219

(do the integral for practice). We could then define the Fourier transform
of sgn(t) to be

(5.5)

This approach has some obvious shortcomings: We may be able to find

a gk(t) sequence that converges to g(t), yet the Fourier transform sequence
Gk(f) may not converge for that sequence. To make matters worse, there
may not be any such sequence of gk (t) which converges to g(t) and yields a
convergent sequence of Fourier transforms. Hence this trick may be useful
for some specific signals, but it is limited in its utility.
The trick works in a similar fashion for DTFT evaluation. As an exercise
try finding the Fourier transform of the discrete time analog to the signum
function:
n>O
g(n) = sgn(n) = { ~-1 n=O (5.6)
n<O
(Just use the same sequence of limiting functions used in the continuous
time case, but now use them only for discrete time.) You should be able to
show that G(O) = 0 and for f =I 0
1 - e i21f f -i sin(2n})
G(f) = 1 _ cos 27r f - 1 = 1 _ cos(27r f)" (5.7)

In the discrete-time case it is occasionally useful to modify the definition

of the signum function slightly to
In> 0
sgno(n) = { -1 0, n:< (5.8)

where the subscript 0 is intended to emphasize that this signum is different

at the origin. Linearity and the previous result imply that the transform
of the modified signum is G(O) = 1 and for f =I 0
1 - ei21f f
1- cos27rJ"
What continuous time infinite duration signal has Fourier transform
G(f) = isgn(f)?

5.2 Periodic Signals and Fourier Series

In this section we modify one of the Fourier transforms already introduced
to obtain the most important generalization of Fourier transforms to a class
220 CHAPTER 5. GENERALIZED TRANSFORMS AND FUNCTIONS

of signals for which the basic definitions fail. In particular, we show that
the transforms of finite duration signals can be used to define useful trans-
forms for periodic infinite duration signals, signals which are not absolutely
summable or integrable and hence signals for which the usual definitions
cannot be used. We begin with the simpler case of discrete time signals.
Recall that an infinite duration discrete time signal 9 = {gn; n E Z} is
periodic with period N if
gn+N = gn (5.9)
for all integers n. Recall also that if we truncate such a signal to produce
a finite duration discrete time signal

and then we replicate 9 to form the periodic extension g = {gnmodN; n E

Z}, then g = g. In other words, the periodic extension of one period of a
periodic function is exactly the original periodic function. Intuitively this
just says that we should be able to describe the transform of an infinite
duration periodic function in terms of the transform of the finite duration
signal consisting of one period of the infinite duration signal; that is, the
Fourier transform of the periodic signal should be "the same" as the DFT
of one period of the signal. We explore how this idea can be made precise.
Given a DTID periodic signal 9 that is not identically 0, then the signal
is clearly not absolutely summable and hence violates the sufficient condi-
tions for the existence of the DTFT. Recall from Chapter 3 however, that
if we truncate 9 to form 9 and then form the DFT
N-l N-l
a(f) =L gn e- i27r !n =L gn e- i27r !n, (5.10)
n=O n=O

then a Fourier series representation for the infinite duration signal is given
from (3.28) as

(5.11)

Thus we have a Fourier-type decomposition of the original infinite dura-

tion signal into a weighted combination of exponentials, but unlike the
basic DTFT formula, the signal is given by a sum of weighted exponentials
rather than by an integral of weighted exponentials (e.g., the inverse trans-
form is a sum rather than an integral). Unlike the DTFT, we have only a
discrete spectrum since only a finite number of frequencies are required in
this representation.
 5.2. PERIODIC SIGNALS AND FOURIER SERIES 221

Exactly the same idea works for a continuous time infinite duration
periodic signal g. Suppose that get) has period T, that is, g(t+T) = get) for
all t E 'R. Then the finite duration CTFT of the signal 9 = {get); t E [0, Tn
is given by
a(f) = loT g(t)e- i27r !tdt (5.12)

and a Fourier series representation is then given by

get) = f:
n=-oo
a;) e i27r y;t. (5.13)

Again this formula resembles an inverse Fourier transform for g, but it is

a sum and not an integral, that is, it does not have the usual form of an
inverse CTFT.
For reasons of consistency and uniformity of notation, it would be nice
to have the basic transform and its inverse for periodic infinite duration
signals more closely resemble those of ordinary infinite duration signals.
To see how this might be accomplished, we focus for the moment on a
fundamental example, the complex-valued discrete time periodic signal e
defined by

where 10 = miN for some integer m. This signal has period N (and
no smaller period if miN is in lowest terms). For the moment the fre-
quency 10 will be considered as a fixed parameter. From the linearity of
the Fourier transform, knowing the DTFT of any single exponential signal
would then imply the DTFT for any discrete time periodic signal because
of the representation of (5.11) of any such signal as a finite sum of weighted
exponentials! Observe that the discrete time exponential is unchanged if we
replace m by m + M N for any integer M. In other words, all that matters
is m mod N.
The DFT of one period of the signal is

E(n = {Io 1 = 10 = ~
1=j"l=O, ... ,N-1,lim
The ordinary DTFT of the signal defined by

L
00

E(f) = en e- i27r !n
n=-oo

does not exist (because the limiting sum does not converge). Observe also
that en is clearly neither absolutely summable nor does it have finite energy
222 CHAPTER 5. GENERALIZED TRANSFORMS AND FUNCTIONS

since

n=-oo n=-oo

Suppose for the moment that the transform did exist and was equal to some
function of I which we call for the moment E(f). What properties should
E(f) have? Ideally we should be able to use the DTFT inversion formula
on E(f) to recover the original signal en, that is, we would like E(f) to
solve the integral equation

1
1

2 E(f)ei27ffndl == ei27ffon, nEZ, (5.14)

-!
in which case we would know that en was indeed the inverse DTFT of E(f).
It will simplify things a bit if we consider the frequency domain to be
[0, 1) instead of [- ~, ~) for the current discussion and hence we seek an
E(f) that solves

(5.15)

What E(f) will do this; that is, what E(f) is such that integrating
E(f) times an exponential ei27rln will exactly produce the value of the
exponential in the integrand with I == 10 for all n? The answer is that
no ordinary function E(f) will accomplish this, but by using the idea of
generalized functions we will be able to make rigorous something like (5.15).
So for the moment we continue the fantasy of supposing that there is a
function E(f) for which (5.15) holds and we look at the implications of the
formula. This will eventually lead up to a precise definition.
Before continuing it is convenient to introduce a special notation for
E(f), even though it has not yet been precisely defined. Intuitively we
would like something which will have a unit area, i.e.,

11 E(f) dl == 1,

which is concentrated in an infinitesimally small region around 10. If E(f)

has these properties, then the integrand will be 0 for I not near 10 and
E(f)e i27f In ~ E(f)e i27f fon for I ~ 10 and hence

11 E(f)e i27f In dl ~ e i27f Ion 11 E(f) dl

~ ei27f lon,
5.2. PERIODIC SIGNALS AND FOURIER SERIES 223

as desired. Again we emphasize that no ordinary function behaves like

this, but the idea of a very narrow, very tall pulse with unit area is useful
for intuition (provided it is not taken too literally) and hence we give this
hypothetical object the name of a unit impulse function or Dirac delta
function at 10, denote it by 8(/ - 10), and depict it graphically as in Figure
5.2.

8(/ - 10)

I
10 1

Figure 5.2: Graphical representation of a Dirac delta

Equation (5.15) is an example of what is called a sifting property. If

E(/) = 8(/ - 10) is multiplied by a complex exponential and integrated, the
resulting integration yields exactly the value of the complex exponential at
the fixed frequency 10 so that the combination of multiplication by 8(/ -
10) and integration exactly sifts out or samples one value of the complex
exponential. We next show that if 8(/ - 10) sifts exponentials at 10, it must
also sift more general signals.
Suppose that X(/); IE [0,1) is a frequency domain signal. Analogous
to a continuous time finite duration signal, we can expand X (/) as a Fourier
series
X(/) = L
xne-i21fnf
nEZ

11
where
Xn = X (/)e i21f In df.

We have changed the signs in the exponentials because we have reversed the
usual roles of time and frequency, i.e., we are writing a Fourier series for a
frequency domain signal rather than a time domain signal. We assume for
simplicity that X(/) is continuous at fo (that is, X(/o + e:) and X(/o - e:)
go to X(/o) as e: -+ 0) so that the Fourier series actually holds with equality
224 CHAPTER 5. GENERALIZED TRANSFORMS AND FUNCTIONS

at 10. Consider the integral

11 8(f - 10)X(f) dl = (I 8(f _ 10)(2: Xne-i27rn/) dl

10 nEZ
= L: Xn r 8(f - 10)e-i27rnl d!
1

nEZ 10
= L: Xne-i27rnlo
nEZ
= X(fo), (5.16)

where we have used the property (5.15) that 8(/ - 10) sifts complex expo-
nentials at 10. The point is we have shown that if 6(/ - 10) sifts complex
exponentials, it also sifts all other continuous frequency domain signals
(assuming they are well behaved enough to have Fourier series).
We now summarize our hand-waving development to this point: If the
signal {e i27r Ion; n E Z} has a Fourier transform 6(/ - 10); 1 E [0,1), then
this Fourier transform should satisfy the sifting property, i.e., for any suit-
ably well behaved continuous frequency domain signal G(f)

11 6(f - 10)G(/) dl = G(fo), (5.17)

for 10 = miN E [0,1).

When suitable theoretical machinery is introduced, the sifting property
will in fact provide a rigorous definition of {8(/ - 10); 1 E [0, I)}, the
Dirac delta function at 10' There are two equivalent ways of making the
sifting property rigorous. The first is operational: ordinary integrals are
used in a limiting statement to obtain a result that looks like the sifting
property. The second is essentially a redefinition of the integral itself using
the concept of a generalized lunction or distribution. We shall see and use
both methods.
The important fact to keep in mind when dealing with Dirac delta
functions is that they only really make sense inside of an integral; one
cannot hope that they will behave in a reasonable way when alone.
As in the modulation theorem, the difference between frequencies 1 - 10
in (5.17), is handled like shifts in the time domain; that is, it is taken modulo
the frequency domain of definition S so that 1 - 10 E S.
The Dirac delta sifting property can be extended to non-continuous
functions in the same way that Fourier series were so extended. If G(f)
has a jump discontinuity at 10, then a reasonable requirement for a sifting
5.2. PERIODIC SIGNALS AND FOURIER SERIES 225

property is
(5.18)

the midpoint of the upper and lower limits of G(I) at 10. This is the general
form of the sifting property. Again observe that no ordinary function has
this property.
Before making the Dirac delta rigorous, we return to the original ques-
tion of finding a generalized DTFT for periodic signals and show how the
sifting property provides a solution.
If we set E(I) = 8(1 - ~), then (5.15) holds. Thus we could consider
the DTFT of an exponential to be a Dirac delta function; that is, we would
have the Fourier transform pair for k E Z N

(5.19)

where the frequency difference I - kiN is here taken modulo 1, that is,
8(1 - kiN) = 8((1 - kiN) mod 1).
If (5.19) were true, then (5.11) and linearity would imply that the DTFT
G of a periodic discrete time signal 9 would be given by

G(f) = 'I:
k=O
G;;) a(f - ~); f E [0,1). (5.20)

Thus the spectrum of a periodic discrete function with period N can be

considered to be a sequence of N Dirac deltas with areas weighted by the
Fourier series coefficients for the periodic function, as depicted in Figure 5.3.
The values G(kIN) which label the arrows are areas, not magnitudes.
It is important to differentiate between the two forms of Fourier trans-
forms being used here: for a periodic signal g, G is the DFT of one period
of 9 while G is the DTFT of the infinite duration signal g. Each of these
can be inverted according to the usual rule for inverting a DFT or DTFT,
thereby giving a weighted exponential form for g. Using the DFT inversion
we have that
1 N-l, k . k
9n = -N~ ' " G(_)e t2 11'Nn. all n E Z (5.21)
N ' .
k=O
Using the DTFT inversion formula we have that

gn = 11 G(f)e i2 11'fn df

= r1'I:G;)a(f_ ~)ei211'fndf
Jo k=O
226 CHAPTER 5. GENERALIZED TRANSFORMS AND FUNCTIONS

G(f)

C(O)

f
o 1
N
2
N
N-l
IV
1

Figure 5.3: Graphical representation of DTFT of a periodic signal

which is the same as the previous equation because of the sifting proper-
ties of Dirac delta functions. Thus we can represent gn either as a sum of
weighted exponentials as in (5.21) (usually referred to as the Fourier series
representation) or by an integral of weighted exponentials (the Fourier in-
tegral representation). Both forms are Fourier transforms, however. Which
form is best? The Fourier series representation is probably the simplest to
use when it suffices, but if one wants to consider both absolutely summable
and periodic infinite duration signals together, then the integral represen-
tation using delta functions allows both signal types to be handled using
the same notation.
To summarize the discussion thus far: given a discrete time periodic
function 9 with period N, the following can be considered to be a Fourier
transform pair:

N-1C(.!.) k
G(f) = L ; o(f - N); f E [0,1) (5.22)
k=O

1: 1

2
G(f)ei2rrfn df; nEZ, (5.23)

where
,k N-l . k
G( -) - ' " 9n e-t2rrwn., k -- 0 , ... , N - 1 .
N -L..J (5.24)
n=O
The same scenario works for continuous time periodic functions assum-
ing the supposed properties of the Dirac delta function if we also assume
that the Fourier transform is linear in a countable sense; that is, the trans-
form of a countable sum of signals is the sum of the corresponding trans-
5.3. GENERALIZED FUNCTIONS 227

forms. In this case the fundamental Fourier transform pair is that of the
complex exponential:

{ei2711ot;t E 'R.} t+ {O(J - /0);/ E 'R.} (5.25)

which allows us to write for a general continuous time periodic function 9

with period T the Fourier transform pair

f: GC;) ~);
i:
G(J) = 0(J - / E 'R. (5.26)
k=-oo

get) G(J)e i21f / t d/; t E 'R. (5.27)

where
• -)
G( k
T
= iT
0
g(t)e- t°21f'1'k t dt; k E Z. (5.28)

Thus as in the discrete time case we can represent a continuous time

periodic signal either as a sum (Fourier series) of weighted exponentials
(5.13) or as an integral (5.27) (Fourier integral). Both representations have
their uses. We next turn to the chore of making the idea of a Dirac delta
more precise so as to justify the above development.

5.3 Generalized Functions

In this section we add rigor to the intuitive notion of a Dirac delta func-
tion using the theory of generalized functions. Recall that an (ordinary)
function is just a mapping of one space into another. In our case we have
been considering complex-valued functions defined on the real line, that
is, functions of the form 9 : 'R. ~ C which are mappings assigning a value
g(x) E C to every x E 'R.. We here use x as a dummy variable since it could
represent 2either time or frequency. Ordinary functions include things like
sin x, e- X ,U-1 (x), etc. A generalized function or distribution is a mapping
of functions into complex numbers according to certain properties. Given a
=
function 9 {g(x);x E 'R.}, a distribution assigns a complex number D(g)
to 9 in a way satisfying the following rules:

1. (Linearity)
Given two functions gl and g2 and complex constants a1 and a2, then
228 CHAPTER 5. GENERALIZED TRANSFORMS AND FUNCTIONS

2. (Continuity)
If limn-too gn(X) = g(x) for all x, then also
lim D(gn) = D(g)
n-too

A generalized function is called a linear functional or linear operator

in functional analysis. The key idea to keep in mind is that a generalized
function assigns a complex number to every function.
A common example of a generalized function is the following: Suppose

i:
that h(x) is a fixed ordinary function and define the generalized function
D by
Dh(g) = g(x)h(x)dx,

that is, D h (g) assigns the value to 9 equal to the integral of the product of 9
with the fixed function h. The properties of integration then guarantee that
D meets the required conditions to be a generalized function. Note that
this generalized function has nothing strange about it; the above integral
is an ordinary integral.
As a second example of a generalized function, consider the operator
Dii defined as follows:

if g(t) is continuous
at t = 0 (5.29)
otherwise

where g(O+) and g(O-) are the upper and lower limits of gat 0, respectively.
We assume that g(t) is sufficiently well behaved to ensure the existence of
these limits, e.g., g(t) is piecewise smooth.

i:
We write this generalized function symbolically as

Do(g) = 8(x)g(x)dx,

but this is only a symbol or alternative notation for DIi, it is NOT an

ordinary integral. The defined distribution is both linear and continuous.
This is in fact the rigorous definition of a Dirac delta function. It is not
really a function in the ordinary sense; it is a generalized function which
assigns a real number to every function. Alternatively, it is not really 8(x)
that we have defined, it is Dii or an integral containing 8(x) that has been
defined. Because of the linearity and continuity of a generalized function,
we can often (but not always) deal with the above integral as if it were an
ordinary integral.
5.3. GENERALIZED FUNCTIONS 229

i:
The shifted Dirac delta D6cois interpreted in a similar fashion. That is,
when we write the integral

D6 co (g) = <5(x - xo)g(x) dx

what we really mean is that this is a generalized function defined by

D () = g(xci) + g(xi))
6. 0 9 2'

The shifted Dirac delta can be related to the unshifted Dirac delta with a
shifted argument. Define the shifted signal gxo (x) by
gxo(x) = g(x + xo).
Then it is easy to see that
D6(gxo) = D6. 0 (g).
This relationship becomes more familiar if we use the integral notation for

i: i:
the generalized function:

<5(x)g(x + xo) dx = <5(x - xo)g(x) dx. (5.30)

In this form it looks like an ordinary change of variables formula. If instead

i: i:
of a generalized function <5 we had an ordinary function h, then

h(x)g(x + xo) dx = h(x - xo)g(x) dx

would follow immediately by changing the variable x - Xo to x. A subtle

but important point has been made here: although <5 is not an ordinary
function and the integrals in (5.30) are not ordinary integrals, under the
integral sign it behaves like ordinary functions, at least in the special case
of the simple change of variables by a shift. We shall see many more ways
in which generalized functions inside integral signs behave as one would
expect.
As a third example of a generalized function we consider an example
which is fairly close to the usual physical description of a Dirac delta func-
tion. Suppose that we have a sequence of functions hn(x); n = 1,2, ...
with the following properties:
1.

1 -00
00 hn(x)dx = 1; n = 0,1,2, ...
230 CHAPTER 5. GENERALIZED TRANSFORMS AND FUNCTIONS

2. For any function 9

. Joo g(O) if g(t) is continuous at t = 0

hm
n-too -00
hn(x)g(x) dx ={ 9(0+)+9(0-)
2 otherwise
(5.31)
Define D6 by
D6(g) = nl~~ [ : hn(t)g(t) dt,
if the limit exists, where now the integrals are ordinary integrals. This
provides an alternative description of the generalized function D 6 •
The limiting function description of the Dirac delta function is conve-
nient for also describing the shifted delta function. If h n is such a sequence,
then an ordinary change of variables yields

lim Joo hn(x - xo)g(x) dx

n-too -00
= J~~ [: hn(y)g(y + xo) dy
g(xri) + g(xi))
= 2
(5.32)

and hence the shifted sequence satisfies the properties of a shifted Dirac
delta. In fact, we can also prove (5.30) by defining the generalized function
D6 zo in terms of the limiting behavior of integrals of the shifted functions
hn (x - xo) in the above sense. This provides a very useful general approach
to proving properties for the Dirac delta: find a sequence h n describing the
generalized function, prove the property for the members of the sequence
using ordinary signals and integrals, then take the limit to get the implied
property for the generalized function.
It is important to note that we are NOT making the claim that the
delta function is itself the limit of the hn' i.e., that "<5(t) = limn-too hn(t)."
In fact, most sequences offunctions hn(t) satisfying the required conditions
will not have a finite limit at t :::: O! In spite of this warning, it is some-
times useful to think of an impulse as a limit of functions satisfying these
conditions. We next consider a few such sequences .

• hn(x) :::: nD..1...(x). (See Figure 5.4.) Alternatively, one can use n n
2n
(nx). This is the simplest such sequence, a rectangle with vanishing
width and exploding height while preserving constant area. If g(t) is
continuous at t :::: 0, then the mean value theorem of calculus implies
that
n Jin
_..1...
2n
g(t) dt r:::J n g(O) .
n
5.3. GENERALIZED FUNCTIONS 231

9,----------r----------,---------,----------,
8

_ .. _ - .1 .. :-
7

I-
n=8
3 --I--j"
I I

2 .. -- --,--"----- ---
I I

n=2
n=1
~1~--------.~O~.5----~--~O~--------~O.~5--------~

1 (t)
Figure 5.4: Impulse via Box Functions: nO 2n

Note that hn(x) has 0 as a limit as n -+ 00 for all x except the point
x = O. Although this sequence is simple, it has the disadvantage that
its derivative does not exist at the edges .

• The sequence hn(x) = nt\(nx) also has the required properties, where
t\ is the triangle function of (1.10). (See Figure 5.5.)
It is also not differentiable at its edges and at the origin. It is piecewise
smooth. Again hn(x) has a limit (0) except for the point x = o.

• The Dirichlet kernel hn(x) = Dn(x) = Sin[:i:(~;:!J. (See Figure 3.1.)

Since this sequence is periodic with period 1, it is only useful for
defining a Dirac delta when the domain of definition of x has width
1, as is the case here. It can be modified for any finite width interval
of length T by scaling ~DN(T) as in (3.70). This sequence was
used in the "proof" of the convergence of the finite duration CTFT
and hence also in the proof of the Fourier series representation of a
periodic function. Thus a rigorous proof that Dn(x) indeed has the
properties required to define the Dirac delta also thereby provides a
rigorous proof of those Fourier inversion formulas.
232 CHAPTER 5. GENERALIZED TRANSFORMS AND FUNCTIONS

9r----------r----------.-------~_r--------_.

8 ··r
;1

n
7 ;!

6 i:'
; I

5 iii.
I I
. ./ I
I I
I
n = 8
3 ................. ,. ...:...! ................ .

I
2 'j' ...;.,~ .. ~ .
...( : ... ~... n =~
... "".. 1.. I .....

·0.5 o 0.5
t

Figure 5.5: Impulse via Triangle Functions: n /\ (nt)

• hn(x) = nsinc(nx) = sin(mrx)/1rx. (See Figure 3.2.) In this example

the limit lim n --+ oo hn(x) does not exist for any x, yet the sequence is
quite suitable for defining the Dirac delta. This is in fact the sequence
of functions that arose in the "proof" of the Fourier integral theorem.
That proof can now be viewed as accurate given the fact (which we do
not prove) that this sequence indeed satisfies the required conditions
of (5.29) to define the Dirac delta generalized function. Observe that
these functions are everywhere differentiable.
As an aid to picturing the functions n sinc( nx), observe that

1I· msinx
--=
1
:1:--+0 X

and
lim sin (ax) = a.
:1:--+0 X
Thus
. sin(n1rx)
11m =n.
:1:--+0 1rX
These function sequences are useful for proving properties of Dirac delta
functions. For example, consider the meaning of o(ax), a delta function
5.4. FOURIER TRANSFORMS OF GENERALIZED FUNCTIONS 233

with an argument scaled by a > O. Using the simplest limiting sequence

I:
we can argue that this should behave under the integral sign like

c5(ax)g(x)dx = 1
lim
00
nO---.L (x)g(x)dx

.1
n-+oo -00
2ria

hm
00
nO y y dy
(-)g(-)-
1
a a a

1
n--too -00 2n<i"
00 y dy
= lim nOJ..(y)g(-)-
n--too -00 2n a a

by a simple change of variables. This, however, is just

g(O)
a
= 1 ~c5(x)g(x)
-00
00

a
dx.

Identifying these two generalized functions (remember the above formulas

hold for all continuous g(x)) implies that
1
c5(ax) = -c5(x).
a
A similar argument for negative a using the fact that now there is a sign
change when the integration variable is changed gives the general result
1
c5(ax) = ~c5(x). (5.33)

A further generalization of this result is

1 b
c5(ax + b) = ~c5(x + a) (5.34)

for any a "I o. The proof is left as an exercise. (See Problem 6.) Other
properties of Dirac c5 functions are also developed in the exercises. One
such property that is often useful is
g(t)c5(t - to) = g(to)c5(t - to) (5.35)

if get) is continuous at to. (See Problem 10.)

5.4 Fourier Transforms of Generalized Func-

tions
With the definition of a generalized function in hand, we can now generalize
the Fourier transform to include integral transforms of signals involving
234 CHAPTER 5. GENERALIZED TRANSFORMS AND FUNCTIONS

Dirac delta functions by interpreting the integral as a generalized function.

In this sense the formulas of the previous section are made precise. Thus,

1:
for example, in the continuous time case we have that

o(t - to)e-i27rftdt

= e-i27rftOj fEn (5.36)

using the sifting property. In particular, setting to = 0 yields

F({o(t)jt En}) = Ijf E nj (5.37)

that is, the Fourier transforII]. of a Dirac delta at the time origin is a con-
stant. This result is the continuous time analog to the result that the DTFT
of a Kronecker delta is a constant.
We have already argued that the Fourier transform of a complex expo-
nential {e i27r fot j tEn} should have the sifting property, that is, behave like
the generalized function defining a Dirac delta. Thus we can also define

F({ei27rtfOjt En}) = {o(f - fo)j f E n}j (5.38)

that is, the Fourier transform of a complex exponential is a delta function.

Alternatively, if we consider a formal integral in a Cauchy sense we should
have that
fL
lim (2 e- i27rt (f-fo) dt
N-+oo Lr;
lim N sinc(N(f - fo)),
N-+oo

which "converges" to o(f - fo) in the sense that the function N sinc[N(f-
fo) 1 inside an integral has the sifting property defining the Dirac delta.
(The formula makes no sense as an ordinary limit because the final limit
does not exist.)
Again the special case of fo = 0 yields the relation

Ff({Ij tEn}) = o(f)j f E nj (5.39)

that is, the Fourier transform of a continuous time infinite duration signal
that is a constant (a dc) is a delta function in frequency.
Note that these two above results are duals: Transforming a delta in
one domain yields a complex exponential in the other domain.
In the discrete time case the intuition is slightly different and there is not
the nice duality because there is no such thing as a Dirac delta in discrete
5.5. * DERIVATIVES OF DELTA FUNCTIONS 235

time. The analogous idea is a Kronecker delta which is a perfectly well

behaved quite ordinary function (whose transform happens to be a complex
exponential). In the infinite duration discrete time case, one can, however,
have Dirac deltas in the frequency domain and their inverse transform is a
discrete time complex exponential since from the sifting property

(5.40)

We have seen that the intuitive ideas of Dirac deltas can be made pre-
cise using generalized functions and that this provides a means of defining
DTFTs and CTFTs for infinite duration periodic signals, even though these
signals violate the sufficient conditions for the existence of ordinary DTFTs
and CTFTs. Generalized functions also provide a means of carefully prov-
ing conjectured properties of Dirac delta functions either by limiting ar-
guments or from the properties of distributions. We will occasionally have
need to derive such properties. One should always be careful about treating
delta functions as ordinary functions.
The previously derived properties of Fourier transforms extend to gen-
eralized transforms involving delta functions. For example, the differen-
tiation theorem gives consistent results for some signals which are not
strictly speaking differentiable. Consider, for example, the box function
{DT(t); t E 'R}. This function is not differentiable at -T and +T in the
usual sense, but one can define the derivative as a generalized function as
d
dt DT(t) = 6(t + T) - 6(t - T)

since if one integrates the generalized function on the right, one gets DT(t).
Now the transform of DT(t) is 2Tsinc(2T f) and hence the differentiation
It
theorem implies that the transform of DT( t) should be i27r f2T sinc(2T f) =
2i sin(2T7r f), which is easily seen from the sifting property and Euler's re-
lations to be the transform of 6(t + T) - 6(t - T).

5.5 * Derivatives of Delta Functions

The Dirac delta is effectively defined by its behavior inside an integral and
hence integrals of Dirac deltas make sense. Since integration and differen-
tiation are inverse operations on ordinary functions, one might ask if one
can define a derivative of a delta function. The answer is yes if we con-
sider the new function as another generalized function. As a first approach,
suppose that we consider the sequence of triangle functions of (1.10) that
236 CHAPTER 5. GENERALIZED TRANSFORMS AND FUNCTIONS

"converges" to the Dirac delta in the sense that (5.31) holds. This sequence
of functions has a derivative defined by

d {_n 2 x E (0, lin)

h~(x)=dxnl\(nx)= +n 2 xE(-l/n,O). (5.41)
o Ixl> lin
The derivative is not defined at the origin and at the points ±l/n (because
the left and right derivatives differ), but we are only interested in the be-
havior of this function under an integral sign and its differential behavior
at these points does not matter. Suppose now that g(x) is a well behaved
function. In particular, suppose that it has a continuous derivative at the
origin. For large n we will then have that

J h' n(x)g(x) dx = n2jO g(x) dx - n 2 (lin g(x) dx

-lin io
1 1
~ n g(-2n)-n g (2n)'

where we have approximated each integral by the value of the function at

the middle of the interval of integration times the length of the integral.
As n ~ 00 this tends to the negative of the derivative, dg(x)ldx evaluated
at x = 0; that is, setting ~x = lin we have that

1 -1 . g(tl.X)_g(-tl.X)
lim [ng( - ) - ng(-)]
n-+oo 2n 2n = hm
tl.x-+O
2
~x
2

= dg(x) I ~ '(0)
dx x=O - 9 .

J
Thus
lim h'n(x)g(x) dx = -g'(O). (5.42)
n-+oo
The intuition here is as follows: If the functions hn "converge" to a Dirac
delta, then their derivatives should "converge" to the derivative of a Dirac
delta, say c5'(x), which should behave under the integral sign as above; that
is, for any g(x) that is differentiable at the origin,

J c5'(x)g(x) dx = -g'(O). (5.43)

The generalized function c5' (x) is called a doublet.

Analogous to the Dirac delta, we can consider the shifted doublet to

J
obtain the formula
c5'(t - x)g(x) dx = -g'(t) (5.44)
5.6. * THE GENERALIZED FUNCTION 8(G(T)) 237

if g(x) is differentiable at t. This type of integral is called a convolution

and the result often described as showing that the convolution of a doublet
with a signal produces the derivative of the signal at the location of the
doublet.

i:
The Fourier transform of the doublet is easily found to be

FJ({8'(t); tEn}) = 8'(t)e-i27rJtdt

= - :te-i27rJtlt=o = i27rf. (5.45)

In a similar manner higher order derivatives of the Dirac delta can be

defined as generalized functions.

5.6 * The Generalized Function c5(g(t))

Generalized functions provide a way to give meaning to a Dirac delta with
an ordinary function as an argument, that is, a generalized function of the
form 8(g(t)). Suppose for the moment that get) is a well-behaved function
with a single zero at t = to; that is, g(to) = 0 and get) :I 0 for t :I to. If we
think of the Dirac delta as a very tall and narrow pulse, then 8(g(t)) will
have a similar form and will be zero except for t = to, where get) = 0 and
hence 8(g(t)) has ajump. Hence a natural guess is that 8(g(t)) = A8(t-to),
where the area A has to be determined. To make this precise we need to

L: L:
demonstrate that for any well-behaved function r(t),

r(t)8(g(t)) dt = r(t)A8(t - to) dt = Ar(to). (5.46)

We will sketch how this result is proved and find the constant A in the
process.
Let hn(t) = nOt.. (t) denote the box sequence of functions converging
to the Dirac delta under the integral sign and consider the limit defining

i: L:
the new generalized function:

r(t)8(g(t)) dt = nl!.~ r(t)hn(g(t)) dt. (5.47)

The function hn(g(t)) is given by

hn(g(t)) = { nO - 21n < get) < 2~ •

otherwise
238 CHAPTER 5. GENERALIZED TRANSFORMS AND FUNCTIONS

To see how this behaves in the limit as n -+ 00, we suppose that t is very
near to and expand get) in a Taylor series around to as

get) ~ g(to) + (t - to)g'(to) + higher order terms ~ (t - to)g'(t o).

With this approximation we have approximately for large n that

hn(g(t)) = {n -2~ <y-to)g'(to) < 2~.

o otherwIse
Assume for the moment that g'(to) is positive and not zero and we can
rewrite the above equation as

hn(g(t)) = {n to - 2n?~(to) < t < to + 2n/(to) . (5.48)

o otherwIse
Thus hn(g(t)) is a box function of height n and width l/(ng'(to)) and cen-
ter at t = to. By an argument analogous to that used to demonstrate that
the sequence h n provides a limiting definition of the Dirac delta, multiply-
ing ret) by this box function, integrating, and taking the limit will yield
r(to)/g'(to), giving (5.46) with A = l/g'(to).
If g'(to) is negative, then g'(to) in (5.48) is replaced by Il(to)1 and the
remainder of the argument proceeds with this substitution. If 9 has a finite
collection of zeros, then a similar argument for each zero yields the following
result.
If g( t) is a well- behaved function with zeros at t 1 , t 2, ... , t M, then

(5.49)

:t
that is,

1 00

-00
r(t)t5(g(t)) <it =
k=l
r,((tt))·
9 k
(5.50)

5.7 Impulse Trains

Having defined impulse functions precisely using the idea of generalized
functions and having used these ideas to describe the generalized Fourier
transform pair consisting of delayed delta functions and shifted exponen-
tials, we now develop an alternative Fourier representation of continuous
time delta functions by applying Fourier series ideas to a single delta func-
tion considered as a finite duration signal. This provides a rather surprising
5.7. IMPULSE TRAINS 239

formula for the delta function which can be proved both formally, pretend-
ing that the delta function is an ordinary signal, and carefully, using the
ideas of generalized functions. This representation leads naturally to peri-
odic impulse trains as infinite duration signals, the so-called ideal sampling
function (or sampling waveform or impulse train). As some of the results
may be counter-intuitive at first glance, it is helpful to first develop them
in the simple context of a finite duration signal. The generalizations then
come immediately using the periodic extension via Fourier series.
Suppose that we consider fJ = {fJ(t); -T/2 ~ t < T/2} as a finite dura-
tion signal which equals the Dirac delta function during T = [- T /2, T /2).
Assume for the moment that we can treat this as an ordinary finite du-
ration continuous time signal and derive its Fourier series; that is, we can
represent fJ(t) on T by a series of the form

where the coefficients can be computed as

en = -T1 jf fJ(t)e-·.27T +n dt = -T1

-T
:r
using the defining property of delta functions. This yields the surprising
representation for a Dirac delta:

T T
L
00 ei27T+n
fJ(t) = -T-; t E [-2' 2)· (5.51)
n=-oo

This suggests that an infinite sum of exponentials as above equals a Dirac

delta, regardless of the value of T! This is not, however, a proof, because
we cannot treat a delta function as an ordinary signal in this way. To
verify (5.51) we must demonstrate that the right hand side is a generalized
function with the necessary properties to be a Dirac delta. In particular,
we must show that for any well behaved continuous function g(t)

(5.52)

This still leaves a problem of rigor, however, as the infinite sum inside the
right-hand integral may not exist; in fact it cannot exist if it is to equal
a delta function. In order to make sense of the generalized function we
240 CHAPTER 5. GENERALIZED TRANSFORMS AND FUNCTIONS

wish to call E~-oo e i2 11'+n IT, we resort to the limiting definition of delta
functions. Define the sequence of functions hk(t) by

(5.53)

We note in passing the strong resemblance of this formula to the Fourier

transform of a discrete time rectangle, but we do not take advantage of this
similarity. Instead observe that for all k these functions have integral
f
jt
I.

/ hk(t) dt = ei2~+n dt
-f -f n=-k

= t j ei2~+n
n=-k_I.
I.

dt

2
k
2: c5n = 1,
n=-k

where c5 n is a Kronecker c5 and we have used the fact that the integral of a
complex exponential over one period is 0 unless the exponent is O. Thus the
hk satisfy the first condition for defining an impulse as a limit. There is no
problem with the interchange of integral and sum because the summation
is finite. Next observe that if g(t) is continuous at the origin,
f I.

E
2 k e i2 11' ion
lim /
k--.oo
hk(t)g(t) dt = lim
k--.oo
/ g(t) - T - dt
n=-k
-f -'2
T

E /2 g(t)-T- dt
1:
k e i2 11'+n
= lim
k--.oo
n=-k_1:
2

G(-Tf)
= lim
k
~
k--.oo T
n=-k
G(Tf)
E
00
= ----;y-'
n=-oo

Does the last infinite sum above exist as implied? If the original signal
g(t) is well behaved in the sense that it has a Fourier transform G, then
5.7. IMPULSE TRAINS 241

from the inversion formula for the finite duration continuous time Fourier
transform we can write
G(lf)
L -r'
00

g(O) = (5.54)
n=-oo

that is, the claimed infinite sum exists and equals g(O). But this implies
that the sequence h n also satisfies the second condition required of the
limiting definition of a delta function:

J
f
lim
k-+oo
hk(t)g(t) dt = g(O)
-f
for continuous functions g. This shows that indeed the (5.51) is valid in a
distribution sense.
We omit the details of what happens when the function g(t) is not
continuous at t = 0, but the proof can be completed by the stout of heart.
Eq. (5.51) provides a representation for a Dirac delta defined as a finite
duration signal. As always with a Fourier series, however, we can consider
the series to be defined for all time since it is is periodic with period T.
Thus the Fourier series provides immediately the periodic extension of the
original finite duration signal. In our case this consists of periodic replicas
of a delta function. This argument leads to the formula for an impulse train

IJIr(t) = L
00

n=-oo
i5(t - nT) = L
00

n=-oo
T;
a~+n

-00 < t < 00. (5.55)

Thus the Fourier series for a Dirac delta considered as a finite duration
signal of duration T gives an infinite impulse train when considered as an
infinite duration signal. The infinite impulse train is sometimes referred to
as the "bed of nails" or "comb" function because of its symbolic appearance.
It is also referred to as the ideal sampling function since using the properties
of delta functions (especially (5.35)), we have that

L
00

g(t)lJIr(t) = g(nT)i5(t - nT); (5.56)

n=-oo

that is, multiplying the sampling function by an arbitrary signal yields a

sequence of pulses with area proportional to the samples of the original sig-
nal. This representation will be useful when we later interpret the sampling
theorem.
242 CHAPTER 5. GENERALIZED TRANSFORMS AND FUNCTIONS

The same manipulations can be done in the frequency domain to obtain

iI!s(f)= f:
n==-oo
8(f-nS)= f:
n==-oo
~ei27rfn.
The summation index can be negated without changing the sum to yield

iI!s(f) = f:
n==-oo
~e-i27rin.

Recall that the Fourier transform of a periodic continuous time signal 9 has
the form
G(f)= ~ GU})t5(f_~)
L..J T T
n==-oo

where

Thus we can now write

(5.57)

that is, the spectrum of a periodic signal g(t) is the sampled Fourier trans-
form of the finite duration signal {g(t); t E [0, Tn
consisting of a single
period of the periodic waveform, when that transform is defined for all real
f·
We have seen a Fourier series representation for the sampling function.
An alternative Fourier representation for such a periodic signal is a Fourier
transform. If we proceed formally this is found to be

.1', ({ WT(t); tEn}) ~ I (j:::oo

-00
o(t - nT») .-;'0" dt
= L
00

n==-oo_oo
J
00

8(t - nT)e- i27r !t dt

L
00
= e-i27r/nT
n==-oo

This exponential sum, however, is almost identical to the Fourier series that
we have already seen for a sampling function, the only difference is that we
5.7. IMPULSE TRAINS 243

have replaced the time variable tiT by the frequency variable fT. Thus we
can conclude that

L
00

.rf({WT(t); t E R}) e- i2 n:f nT

n=-oo
1 00

= T- ""
LJ 8(f - 'T!.!:.)
n=-oo

= (5.58)

that is, the Fourier transform of the sampling function in time is another
sampling function in frequency!
The sampling function provides an example in which we can use the
properties of delta functions to demonstrate that generalized Fourier trans-
forms can possess the same basic properties as ordinary Fourier transforms,
as one would hope. The stretch theorem provides an example. Consider
the periodic functions formed by replicating the finite duration signals. Let
9 = {g(t); t E R} be the periodic extension of some signal 9 and ga the
periodic extension of ga = {g(at);t E [O,Tla)}, with a > 0. Then 9 has
period T, while 9(at) has period T I a. The Fourier transform of 9(t) is

1 k k
L
_ 00
G(f) = -G(-)8(f --)
k=-oo T T T

and the Fourier transform of g(at) is, from the finite-duration argument,

~ _1_!G(!:..)8(f _ ka)
k~Tlaa T T
= f
k=_ooT
~G(!:..)8(f _
T
ka).
T

According to the infinite duration continuous time stretch theorem, how-

ever, this transform should be

1- f 11 k f k
-G(-)
a a
= -- ""
aT LJ
00
G(-)8(- - -),
TaT
k=-oo

which from the stretch theorem for the Dirac delta function is

f:
k=-oo
~G(~)8(f - ~),
which agrees with the previous result.
244 CHAPTER 5. GENERALIZED TRANSFORMS AND FUNCTIONS

An important special case of the sampling function with a special name

is the case with T = 1. Here the signal is called the shah function and is
defined as 00

III(t) = ~ 8(t - n). (5.59)

n=-oo
Bracewell [6] describes the history of the cyrillic letter ill for the histor-
ically minded reader. Using the stretch formula for delta functions, (5.34),
observe that
tOOt
L L
00
ill(f) = t5(f - n) = T t5(t - nT)j (5.60)
n=-oo n=-oo

that is, the sampling function is given by

f: t5(~
n=-oo
- n) = T-1ill(t/T). (5.61)

Observe from the earlier results for sampling functions that ill is its own
Fourier transformj that is,
F( {ill(t)j tEn}) = {ill(f)j fEn}. (5.62)

Impulse Pairs
We close this section with two related generalized Fourier transforms. First
consider the continuous time signal (actually, generalized function) defined
by
II (t) = t5 (t + ~) + t5 (t - ~) j tEn, (5.63)
2
which is called an even impulse pair. Taking the generalized Fourier trans-
form we have that

(5.64)

Similarly we can define the odd impulse pair

8(t + 1) - t5(t - 1)
II (t) = 2 2 2 ; tEn, (5.65)

and find that

(5.66)
5.B. PROBLEMS 245

5.8 Problems
5.1. Use the limiting transform method to find the Fourier transform of
sgn(t + 3).
5.2. Suppose that we attempt to find the Fourier transform of sgn(t) using
the sequence h n (t) defined by

1 O<t:::;n
hn(t)= { 0-1 -n:::;t<O
otherwise
instead of the exponential sequence actually used. Does this approach
yield the same result?

5.3. What is the (generalized) Fourier transform of the signal 9 = {g(t); t E

R} defined by g(t) = l/t for t i= 0 and g(O) = O?

5.4. We have seen that a discrete time infinite duration periodic signal gn
with period N can be expressed as a Fourier series

L
00

gn = bkei27rfrn.
k=-oo

Suppose that we consider a more general but similar form of infinite

duration discrete time signal. Consider the signal h n given by

L
00

hn = bkei27rAkn
k=-oo

not have the form *

the difference being that we are now allowing frequencies which may
of those in the Fourier series. A signal of this
form is sometimes referred to as a generalized Fourier series or a Bohr-
Fourier series after Harald Bohr who developed their theory [4J. Is a
signal of this form periodic?
Hint: Consider a simple such signal h n = e i27r An for which 27r.x is an
irrational number.
Given such a sequence h n how would you recover the values of bn ?
Hint: Consider the sums
246 CHAPTER 5. GENERALIZED TRANSFORMS AND FUNCTIONS

and assume that any limit interchanges are valid. This form of gener-
alized Fourier analysis is useful for a class of signals known as almost
periodic signals and has been used extensively in studying quantiza-
tion noise.
5.5. Find the DTFT of the signal gn = n mod N; n E Z. Find the DTFT
of the periodic extension of the signal 9 given by

gn -
_{Io n=O,1, ... ,N/2-1
n = N j 2, ... , N - 1

(N is even). Find the CTFT of the analogous continuous time signal,

the periodic extension of g(t) = 1 for 0 ::; t < Tj2 and g(t) = 0 for
T /2 ::; t < T. What are the duals to these transform pairs?
5.6. Prove (5.34), i.e., that if a =/: 0,
1 b
c5(at + b) = ~c5(t + ~).
5.7. What, if anything, does c5 2 (t) mean?
5.8. What is the 2D transform of the signal g(x, y) = {sin(27rax) sin(21l'by);
x E R,y E R}?

5.9. Evaluate
1 +00
-00 c5( -2x
X
+ 3) t\ ("3) dx.
5.10. Given an ordinary function h that is continuous at t = 0 and a Dirac
delta function (a distribution) c5(x), show that the product h(x)c5(x)
can also be considered to be a distribution DhO with the property

DhO(g) = h(O)g(O)

1:
if 9 is continuous at the origin. This is symbolically written as

h(x)c5(x)g(x)dx = h(O)g(O).

This property is usually abbreviated as h(x)c5(x) = h(O)c5(x). Extend

the property to prove (5.35), i.e., that h(x)c5(x-xo) = h(xo)6(x-xo)
if h is continuous at x = Xo.
5.11. Find the generalized Fourier transform U(f) of the unit step function
U_l(t); t E R.
5.8. PROBLEMS 247

5.12. Are t8'(t) and -oCt) equal? (That is, are the corresponding distribu-
tions identical?)

5.13. What is the Fourier transform of

L
N
ae-bdt-Tk)u_1 (t),
k=O

where bk > 0 for all k, the 'Tk and a are fixed, and where U-1 (t) is the
unit step function?

5.14. Consider the impulse-like generalized function 01 defined by the se-

quence
2n 1 2n 1
hn(x) = 3 A (n(x - ;)) + 3 A (n(x + ;)).

What is the value of

L: g(X)01 (x) dx

for g(x) continuous at x = 0 and for g(x) discontinuous at x = O?

5.15. Determine the Fourier series coefficients and the Fourier transforms
of the following periodic continuous time signals:
00
(a) 2: A(t -
n=-(X)
2n).
00
(b) E A(t -
n=-oo
4n).
00
(c) E A(~(t -
n=-oo
nT)).

E
00 2
(d) e- 7r (t-n). (Hint: Use the Poisson summation formula.)
n=-oo
(e) I cos(1l"t)I.

5.16. Consider a continuous time finite duration signal g = {get); -1/2 ::;
t < 1/2} defined by

-I if !.4 < t < !.2

if t = ±!. or t = -!.
get) = ~{ . 1 4
If -- < t < -
1 2

-1
t 4 1
if- 2 <t<-4:

(a) What is the Fourier transform GU) of get)?

248 CHAPTER 5. GENERALIZED TRANSFORMS AND FUNCTIONS

(b) Find a Fourier series for g(t). Does the series equal g(t) exactly
for all -1/2 :$ t < 1/2?
(c) Suppose that {h n ; n E Z} is a discrete time signal with Fourier
transform H(f) = g(f), where 9 is as above. What is h n ?
(d) Let g(t) denote the periodic extension of g(t) having period 1.
Sketch g(t) and write a Fourier series for it.
(e) Sketch the shifted signal {g(t - 1/4); t E 'R} and find a Fourier
series for this signal.
(f) What is the Fourier transform of g(t)?

5.17. Derive the frequency domain sampling formula (4.17).

5.18. An infinite duration discrete time signal gn is defined to be

gn={(_~)n ifn=O,I,2, ....

o otherwise
An infinite duration continuous time signal p = {p(t); t E 'R} has
Fourier transform P.
(a) What is the Fourier transform G(f) of g?
(b) Define the infinite duration continuous time signal f(t) by

= L gnp(t -
00

f(t) nT).
n=O

Sketch the signal f(t) for a simple p and find its Fourier trans-
form F(f) in terms of the given information. What happens if
p is allowed to be a Dirac delta?
(c) For an ordinary signal p (not a Dirac delta), find the energy

5.19. Suppose that 9 = {gn, n E Z} is a discrete time signal. New signals

y and ware defined for all integers n by

_ {gn if n is odd
Yn - 0 if n is even
wn = y(2n + 1).
Find the Fourier transform W(f) of w in terms of G(f).
5.8. PROBLEMS 249

Note: This is a challenging problem since you may not assume that
Gis bandlimited here. You can avoid the use of generalized functions
if you see the trick, but straightforward analysis will lead you to an
impulse train in the frequency domain.
5.20. Define a pulse train (pulses, not impulses!)
1
L
00

pet) = 21' DT(t - nT),

n=-oo

where T, 0 < l' < Tj2 are parameters.

(a) Find a Fourier series for pet).
(b) Find the Fourier transform of p( t) .
(c) Suppose that 9 = {get); t E R} is a band limited signal with
Fourier transform G satisfying G(f) = 0 for If I ;::: w. Suppose
that T < 2iv.Find the Fourier transform R(f) of the signal
ret) = (1 + mg(t))p(t), where m is a fixed constant. Sketch R(f)
for an example G(f).
(d) Is the system mapping 9 into r linear? Time-invariant?
(e) A linear time invariant (LTI) filter hBPF{t) is defined by its
Fourier transform
H
BPF
(f) = {I0 If - fol ~ W or If
otherwise,
+ fol ~ W

where fo > W. (In practice usually fo » W.) A filter of this

form is called an ideal band pass filter centered at fo. Find the
impulse response hBPF{t).
(f) Define fo = MjT for some large integer M. Show that if ret)
is the input to the ideal bandpass filter hBPF, then the output
yet) can be written as an amplitude modulation (AM) signal
yet) = A[1 + mg(t)) cos(27rfot) ,
where you provide A in terms of the given parameters.

What happens in the limit as l' -+ O?

Note: This problem is an old fashioned way of generating AM

using a switch. Multiplication by pet) can be be accomplished
by switching the waveform off (multiplying it by 0) and on (mul-
tiplying it by 1). This is called a chopper modulator.
250 CHAPTER 5. GENERALIZED TRANSFORMS AND FUNCTIONS

5.21. List all the signals you know that are their own Fourier transform
or generalized Fourier transform. What unusual properties do these
signals have? (For example, what do the various properties derived
for Fourier transforms of signals imply in this case.)
Chapter 6

Convolution and
Correlation

We have thus far considered Fourier transforms of single signals and of lin-
ear combinations of signals. In this chapter we consider another means of
combining signals: convolution integrals and sums. This leads naturally
to the related topics of correlation and products of signals. As with the
transforms themselves, the details of the various definitions may differ de-
pending on the signal type, but the definitions and the Fourier transform
properties will have the same basic form.
We begin with an introduction to the convolution operation in the con-
text of perhaps its most well known and important application: linear
time-invariant systems.

6.1 Linear Systems and Convolution

Recall from Chapter 1 that a system is a mapping £ of an input signal
v = {v(t); t E 7i} into an output signal w = {w(t); t E To} = £(v). In
this section we will consider the common special case where the input and
output signals are of the same type; that is, 7i = To.
Recall also from Chapter 1 that a system is linear if given input sig-
nals v(1) and V(2) and complex numbers a and b, then £(av(1) + bV(2)) =
a£(v(1)) + b£(V(2)).
As with the linearity property of Fourier transforms, one can iterate the
definition of linearity to argue that for any finite collection of input signals
252 CHAPTER 6. CONVOLUTION AND CORRELATION

v(n); n = 1, ... ,N and any complex constants an; n = 1, ... ,N

N N
L:(2: anv(n)) = L anL:(v(n)). (6.1)
n=l n=l
This does not imply the corresponding result for infinite sums of signals and
hence we often assume (without always explicitly stating it) that a linear
system also has the extended linearity or countable additivity property:
given an infinite collection of signals v(n); n E Z and a complex-valued
sequence an; n E Z, then
00 00

L:(L anv(n)) = L anL:(v(n)). (6.2)

n=l n=l
There is a further extension of extended linearity of which we shall also
have need. It is best thought of as a limiting form of countable additivity:
Suppose that we have a family of signals v(r); r E R, where as before R
is the set of real numbers; that is, we now have an uncountable infinity or
continuum of input signals, and a weighting function a r ; r E R. Then

L:(h arv(r) dr) = h arL:(v(r)) dr, (6.3)

where the weighted linear combination of signals In

arv(r) dr is written
more completely as Un arv(r) (t) dr; t E R} in the infinite duration con-
tinuous time case. This complicated notation just means that one can
pull integrals as well as sums from inside to the outside of a linear system
operator.
As we have stated before, integrals and sums are linear operations. For
this reason systems defined by integrals and sums of input functions are

i:
linear systems. For example, the systems with output u defined in terms
of the input v by
u(t) = v(r)ht(r) dr
in the infinite duration continuous time case or the analogous

L
00

Un = Vkhn,k
k=-oo
in the discrete time case yield linear systems. In both cases ht (r) is a
weighting which depends on the output time t and is summed or integrated
over the input times r. We shall see that these weighted integrals and sums
are sufficiently general to describe all linear systems. A special case will
yield the convolution operation that forms the focus of this chapter. First,
however, some additional ideas are required.
6.1. LINEAR SYSTEMS AND CONVOLUTION 253

The 8-Response
Suppose that we have a system e with input and output signals of the
same type, e.g., they are both continuous time signals or both discrete
time signals and the domains of definition are the same, say Ti = To = r.
Suppose that the input signal is a delta function at time rj that is, if the
system operates on discrete time signals, then the input signal v = {vnj n E
Z} is a Kronecker delta delayed by r, Vn = On-T' and if the system operates
on continuous time signals, then the input signal v = {v(t)j tEn} is a
Dirac delta delayed by r, vet) = oCt - r). In both cases we can call the
input signal otT) to denote a delta delayed by r. The output signal for
this special case, {h(t, r)j t E 7}, is called the delta response or o-response.
For continuous time systems it is commonly called the impulse response
and for discrete time systems it is often called a unit sample response.
The name impulse response is also used for the discrete time case, but
we avoid that use here as the word "impulse" or "unit impulse" is more
commonly associated with the Dirac delta, a generalized function, than
with the Kronecker delta, an ordinary function. While the two types of 0
functions play analogous roles in discrete and continuous times, the Dirac
delta or unit impulse is a far more complicated object mathematically than
is the Kronecker delta or unit sample.
In discrete time

(6.4)

and in continuous time

h(t, r) = e t ( {o(p - r)j p E 7}). (6.5)

Observe that if the system is time invariant, and if {wet) = h(t, O)j t E
T} is the response to a 0 at time r = 0, then shifting the 0 must yield a
response {w(t-r) = h(t,r)j t E T}. Rewriting this as h(t,r) = w(t-r) for
all t and r emphasizes the fact that the o-response of a time invariant system
depends on its arguments only through their difference. Alternatively, if a
system is time invariant, then for all allowable t, r and a
h(t - a,r - a) = h(t,r). (6.6)

Both views imply that if a system is time invariant, then there is some
function of a single dummy variable, say h(t), such that

h(t, r) = h(t - r). (6.7)

In fact, (6.6)-(6.7) imply that h(t - r) = h(t, r) = h(t - r, 0) and hence
h(t) = h(t,O). It is common practice, however, to use the same symbol for
254 CHAPTER 6. CONVOLUTION AND CORRELATION

h and it and to just write

h(t, 7") = h(t - 7") (6.8)

if the system is time invariant. This is actually an abuse of notation as

we are equating a function of two independent variables to a function of
a single independent variable, but it is quite common. A function with
the property of depending on two variables only through their difference is
called a Toeplitz function [19].

Superposition
The 8-response plays a fundamental role in describing linear systems. To
see why, consider the case of an infinite duration discrete time signal v as
input to a linear system C. Recall that
00
Vn = L Vk 8n-k
k=-oo

and assume that the system satisfies the extended linearity property. Then
the output of the system is given by

Un = Cn(v)
= Cn({Vk; k E Z})
00

= Ln({ L v/8k-/; k E Z})

/=-00
00
Cn( L v/{8k-/; k E Z})
/=-00
00

= L v/Cn({8k-/; k E Z})
/=-00
00
L v/hn,/. (6.9)
/=-00

The output at time n of an infinite duration discrete time linear system

can be written as a weighted summation of the input values, where the
weighting is given by the 8-response. Thus the output can always be found
from the input using the 8-response and calculus. The sum is called the
superposition sum.
6.1. LINEAR SYSTEMS AND CONVOLUTION 255

A similar argument holds for the infinite duration continuous time case,

I:
where now the integral form of extended linearity is needed. In the contin-
uous time case
v(t) = v(r)o(t - r) dr

and the output of the system is given by

u(t) L:.t(v)
L:. t ({v (r); r E R} )

L:.t({l: v(r)o(r - r) dr; r E R})

=
I:
L:.t(l: v(r){o(r - r); r E R} dr)

1:
v(r)L:.t({o(r - r); r E R}) dr

v(r)h(t, r) dr. (6.10)

The integral above is called the superposition integral.

Suppose that the system has finite duration discrete time signals as in-
put and output signals. The only difference from the discrete time case
considered above is that now the shift is taken as a cyclic shift. For sim-
plicity we omit the modular notation when writing delays, but all delays
below are taken mod N, e.g., On-k means O(n-k)modN when considering
signals of the form {v n ; n E Z N }. As before
N-I

Vn =L VkOn-k.
k=O

The output of the system is given by

un L:.n({vk;kEZN})
N-I

L:.n({L VlOk-l; k E ZN})

1=0
N-I

L:.n(L VdOk-l; k E ZN})

1=0
N-I

L vlL:.n({Ok-l; k E ZN})
1=0
256 CHAPTER 6. CONVOLUTION AND CORRELATION

N-l
= L vlh n ,/. (6.11)
/=0

Note that here ordinary linearity suffices; that is, we need not assume
extended linearity.
A similar form can be derived for the finite duration continuous time
case.
We have seen that if a system is time invariant, then it must have a
£5-response of the form h(t, r) = h(t - r). Conversely, if a system has a
£5-response of this form, then it follows from the superposition integral or
sum that the system is also time invariant. Thus we can determine whether
or not a system is time invariant by examination of its £5-response.

LTI Systems
Provided that the input and output signals to a linear system are of the
same type, the system always satisfies either the superposition integral
formula or the superposition sum formula expressing the output of the
system as a weighted average (sum or integral) of the inputs. We now
consider in more detail the simplifications that result when the system is
also time invariant.
Suppose that a system .c is both linear and time invariant, a special
case which we refer to as an LTl system or LTl filter. Note that this is
well-defined for all input and output signal types. Suppose further that
the input and output time domains are the same so that the superposition
integral or summation formula holds. Since in this case h(t, r) = h(t - r),
the superposition summation and integral reduce to simpler forms. For
example, in the infinite duration discrete time case we have that

L L
00 00

Un = vlhn,l = vlhn-l. (6.12)

1=-00 1=-00

This operation on the signals v and h to form the signal u is called the
convolution sum and is denoted by

u = v * h. (6.13)

i: i:
Similarly, in the infinite duration continuous time case we have that

u(t) = v(r)h(t, r) dr = v(r)h(t - r) dr. (6.14)

6.2. CONVOLUTION 257

This operation combining a signal v with a signal h is called a convolution

integral. It is the continuous time analog of the convolution sum and it is
also denoted by
u = v * h. (6.15)
Similar forms hold for the finite duration examples with finite limits of
summation or integration and a cyclic shift. (The details are given in the
next section.)
We have now proved the following result.
Theorem 6.1 If C is a linear time invariant (LTI) system for which the
input and output signals have the same type and if h is the 6-response of
the system, then
(6.16)
that is, the output is given by the convolution of the input with the 6-
response.
It should be emphasized that the result holds for both discrete and
continuous time and for both finite and infinite duration signals. This
result is the basis for the application of Fourier analysis to linear systems.
Having introduced the convolution operation, we next turn to develop-
ing its properties with respect to Fourier transforms.

6.2 Convolution
First suppose that v = {v(t);t E 'R.} and h = {h(t); t E 'R.} are two infinite
duration continuous time signals. We formally define the convolution (or

i:
convolution integraQ of these two signals by the signal 9 = {g(t); t E 'R.}
given by
g(t) = v«()h(t - () d(j t E 'R., (6.17)

the integral of the product of one signal with the time reversed and shifted
version of the other signal. We abbreviate this operation on signals by

g = v * h.
We also use the asterisk notation as g(t) = v * h(t) when we wish to em-
phasize the value of the output signal at a specific time. The notation
g(t) = v(t) * h(t) is also common, but beware of the potential confu-
sion of dummy variables: the convolution operation depends on the entire
history of the two signals; that is, one is convolving {v( t); t E T} with
{h(t); t E T}, not just the specific output values v(t) with h(t).
258 CHAPTER 6. CONVOLUTION AND CORRELATION

We have already seen the primary example of convolution: if a signal v

is put into a linear time-invariant system described by an impulse response
h = {h(t); tEn}, then the output signal is the convolution v*h. Another
common application of convolution is in probability theory: if v and h
are two probability density functions describing two independent random
variables X and Y, then the probability density function of the sum X + Y
is given by the convolution v * h. The same is true for probability mass
functions. Convolution also cropped up in the development of inversion in
(3.79), where a signal was convolved with a sinc function.
In the infinite duration discrete time case the definition of convolution
is the same except that the integral becomes a sum: Given discrete time
signals v = {v n ; n E Z} and h = {h n ; n E Z}, then the convolution or
convolution sum g = {gn; n E Z} of the two signals is defined by

L
00

gn = Vkhn-k; n E Z. (6.18)
k=-oo

As in the continuous time case we use the shorthand g = v * h.

In the finite duration case we use the same trick that we used when
extending the notion of the shift to finite duration signals. In fact, this
trick is necessary because convolution is defined as a sum or integral of
the product of a signal and a time reversed and shifted version of another
signal. The basic idea is the same for both continuous and discrete time:
Take the convolution of the periodic extensions of the finite duration signals
over one period. In the discrete time case, given two signals v = {vn; n =
0,1, ... , N - I} and h = {h n ; n = 0,1, ... , N - I}, define the convolution
g = {gn; n = 0, ... , N - I} = v * h by
N-l N-l
gn = L Vkh(n-k)modN = 2:= Vkhn-k (6.19)
k=O k=O

where hn is the periodic extension of h n . This form of convolution is called

a cyclic convolution or circular convolution. Writing out the sums we have

go voho + v1hN-l + V2 h N-2 + ... + VN-1h 1

gl VOhl + vlho + V2 h N-l + ... + VN- 1h2

The h sequence is cyclically rotated to produce successive values of gn.

6.2. CONVOLUTION 259

The discrete time finite duration convolution is far more important in

applications than the continuous time finite duration convolution because of
its use in digital signal processing. For completeness, however, we observe
that the same idea works for continuous time: given two finite duration
continuous time signals v = {v(t); t E [0, Tn
and h = {h(t); t E [0, Tn,
define the convolution

9 = {g(t);t E [0, Tn = v * h
by

g(t) = foT v(()h((t - () mod T) d( = foT v(()h(t - () d( (6.20)

where h(t) is the periodic extension of h(t).

* Signal Algebra
Suppose that we now consider the space of all signals of the form g =
{g(t); t E 'R}. While we will emphasize the infinite duration continuous
time case in this section, the same results and conclusions hold in all cases
for which we have defined a convolution operation. We have defined two
operations on such signals: addition, denoted by +, and convolution, de-
noted by *. This resembles the constructions of arithmetic, algebra, and
group theory where we have a collection of elements (such as numbers,
polynomials, functions) and a pair of operations. A natural question is
whether or not the operations currently under consideration have useful
algebraic properties such as the commutative law, the distributive law, and
the associative law. The following result answers this question affirmatively.

Theorem 6.2 The convolution and addition operators on signals satisfy

the following properties.

1. Commutative Law
9 * h = h * g. (6.21)

2. Distributive Law

(6.22)

3. Associative Law

(6.23)
260 CHAPTER 6. CONVOLUTION AND CORRELATION

Consider the first property in the case of infinite duration continuous

i:
time signals: the signal f * h is defined by

f(()h(t - () de·

i:
Changing variables by defining TJ =t - ( this becomes

L- OO
f(t - TJ)h(TJ) ( -dTJ) = f(t - TJ)h(TJ) dTJ,

which is just h * f, as claimed. The result follows for the other signal types
similarly. The Distributive Law follows from the linearity of integration.
The proof of the Associative Law is left as an exercise.
In order to have an algebra of signals with the convolution and sum
operations, we also need an identity signal; that is, a signal such that if
convolved with any other signal yields the other signal. (The signal that
is identically 0 for all time is the additive identity.) This role is filled
by the Kronecker delta function in discrete time and by the Dirac delta
function in continuous time since if we define the signal 8 by {8(t); t E T}
for continuous time or {8 n ; nET} for discrete time, then 0 * 9 = g. For
example, in the discrete time case

To summarize the identity properties:

4. Identity
0+ 9 = g, (6.24)
where 0 is the 0 signal (it is 0 for all t),

0* 9 = g. (6.25)

A detail not yet treated which is needed for our demonstration that
the space of signals (including generalized functions) is an algebra is the
fact that we can convolve generalized functions with each other; that is, the
convolution of two 0 functions is well-defined. In fact, if 0 is to play the role
of the convolution identity, we should have that 8 *8 = 8. This is immediate
for the Kronecker 8 in discrete time. To verify it in the continuous time
case suppose that hn(t) is a sequence of pulses yielding the Dirac delta in

i: i:
the sense of the limiting definition of a distribution:

o(t)g(t) dt = J~ hn(t)g(t) dt = g(O)

6.3. EXAMPLES OF CONVOLUTION 261

for get) continuous at t = O. We define 15 * 15 by

J J
00 00

(15 * 15) (t)g(t) dt = lim

n--+oo
hn * hn(t)g(t) dt
-00 -00

J J
00 00

= lim
n--+oo
d(hn() dt hn(t - ()g(t).
-00 -00

The rightmost integral approaches g(O in the limit and hence the overall
integral approaches g(O). Thus the convolution of two Dirac delta functions
is another Dirac delta function.
The final requirement for demonstrating that our signal space indeed
forms an algebra is the demonstration of an inverse for addition and for
convolution. The additive inverse is obvious - the negative of a signal
is its additive inverse (g + (-g) = 0) - but the inverse with respect to
convolution is not so obvious. What is needed is a means of finding for
a given suitably well-behaved signal 9 another signal, say g-l, with the
property that 9 * g-1 = 15. This is the signal space analog of the ordinary
multiplicative inverse a(l/a) = 1. This property we postpone until we have
proved the convolution theorem in a later section.

6.3 Examples of Convolution

Consider first the continuous time convolution of two signals J(t) and h(t)
pictured in Figure 6.1 and defined by
0 if t < OJ
J(t) ={ 1- t if 0 ::; t ::; Ij
o if t > 1
t-l
h(t) = n(-2-)'
Note that in drawing the box function the end points can be determined
easily as follows: The right edge of the box occurs when the argument
(t-l)/2 = 1/2 and hence when t = 2. The left edge occurs when (t-l)/2 =

I:
-1/2 and hence when t = O. To perform the convolution and find the signal

get) = J(r)h(t - r) dr
262 CHAPTER 6. CONVOLUTION AND CORRELATION

J(t)

t
012
h(t)

1 +----,
t
012

Figure 6.1: Example Waveforms

we need to integrate the product J(r)h(t - r) over all r. Alternatively, we

could integrate J(t - r)h(r), but it is simpler to time-reverse the simplest
of the signals, which in this case is the box.
The product of the signal J(r) with the reversed and delayed signal
h( t - r) varies in shape depending on the value of t, which in turn affects the
functional form of the convolution. We consider each possibility separately.
1. t < O. Here the waveform J(r)h(t - r) = 0 for all r and hence its
integral get) is also 0 in this region. (See Figure 6.2.)
2. 0 :s; t < 1. Here J(r)h(t - r) is nonzero between 0 and t and its
integral is the area under the product of the waveforms in as depicted
in Fig. 6.3:

get)
rt (1 - r) dr = (r - '2 )I~ = t -
= 10
r2 t2
2'

3. 1:S; t < 2. With reference to Figure 6.4, get) is the area of the product
of the waveforms in the region where both are nonzero, which is now

= 10r (1 -
1 1
get) r) dr = 2'
6.3. EXAMPLES OF CONVOLUTION 263

f(t)h(t - r)

r
o 1 2
t- 2 t
Figure 6.2: t <0

f(r)h(t - r)

r
o 1 2
t- 2 t
Figure 6.3: 0 S t <1

4. 2 S t < 3. As illustrated in Figure 6.5

g(t) = 11

t-2
(1 - r) dr =- -
1
2
(t - 2)
1
+ -(t - 2)2.
2

5. 3 S t. With reference to Figure 6.6, once again there is no overlap

and g(t) = O.

We can now summarize the convolution integral:

t - 1. t 2 if 0 S t S 1;
g(t) = { ~ 2 2 if 1 S t S 2;
l!.2 - 3t + L2 if 2 -< t <
- 3·,
o otherwise.
264 CHAPTER 6. CONVOLUTION AND CORRELATION

f(r)h(t - r)

r
o 1 2
t-2
Figure 6.4: 1 :::; t < 2

f(r)h(t - r)

r
012
t- 2 t
Figure 6.5: 2 :::; t < 3

f(r)h(t - r)

r
012
t-2
Figure 6.6: 3:::; t
6.3. EXAMPLES OF CONVOLUTION 265

The final waveform is depicted in Figure 6.7. Exercise: Prove that the

f * h(t)

t
o 1 2 3

Figure 6.7: g(t) = f * h(t)

convolution of n(t) with itself is the triangle function

A(t) = { 1 - It I if It I < 1; (6.26)

o otherwise .
As a discrete time convolution example consider the signals

f n_- {pn if n = 0,1, ... ;

0 otherwise,

where p:j:. 1, and

n E ZN+1;
otherwise.
To evaluate the convolution

observe that hn-k is 1 if n-k = 0,1, ... , N and hence if k = n, n-l, .. . ,n-
N. Thus the sum becomes
n
gn = L fk.
k=n-N

Again evaluation is eased by separately considering the possible cases:

1. n < 0
In this case 9n = 0 since the summand is 0 for the indexes being
summed over.
266 CHAPTER 6. CONVOLUTION AND CORRELATION

2.0:::; n < N

n 1 pn+l
gn = 2:l = ~ _p
k=O
Note that the above formula is not valid if p = 1.
3. N$ n

n
gn = 2:
k=n-N
pk

n
= pn-N
2:
k=n-N
pk-(n-N)

N
pn-N2:rri
j=O
n-N 1 - pN+1
= p 1 -p . (6.27)

What happens when p = 1?

As a final example, we show that the continuous time infinite duration
impulse train or ill function can be combined with convolution to provide a
convenient means of replicating a signal (or spectrum), e.g., of forming the
periodic extension of a finite duration signal (or finite bandwidth spectrum).
For simplicity we assume throughout that the signal is continuous. In this

i:
case recall that convolving an impulse with the signal simply produces the
original signal:
g(r)8(t - r) dr = g(t).

(This is just the sifting property.) Consider next convolving an infinite

duration continuous time signal 9 with the ideal sampling function WT(t) =
E~=-oo c5(t - nT) to form 9 * WT. This yields

9 * WT(t) J00

-00
g(r) L
00

n=-oo
c5(t - nT - r) dr

= 2:
00

n=-oo_oo
J
00

g(r)c5(t - nT - r) dt
 6.4. THE CONVOLUTION THEOREM 267

2:
00

= g(t - nT). (6.28)

n=-oo

Thus the signal is the sum of an infinite number of replicas of 9 shifted by

multiples of T. If g(t) = 0 for t ¥- [-T /2, T /2), then 9 * "lJiT will be an
infinite sequence of periodic copies of the original signal. The same idea
works in the frequency domain. The result can also be stated in terms of
the ill function by including the extra scaling factor of T.

6.4 The Convolution Theorem

One of the most important applications of the Fourier transform is for the
computation of convolutions of all kinds. It might appear that the convo-
lution integral or sum is not too complicated in the first place and hence a
shortcut requiring the computation of Fourier transforms may in fact not
be a shortcut. This may in fact be the case when convolving some simple
signals, but when convolving more complicated signals whose transforms
are known (perhaps in a handy table) and, more importantly, when doing
multiple convolutions involving several signals, the Fourier transform can
provide significantly simpler evaluations. Because of the importance of the
result we state it formally as a theorem and discuss its proof for several
signal types. The basic idea is simple to state before we get formal: the
Fourier transform of the convolution of two signals is the product of the
Fourier transforms of the two signals.

Theorem 6.3 The Convolution Theorem

Given two signals 9 = {g(t)j t E T} and h = {h(t)j t E T} with Fourier
transforms G = {G(f)j f E S} = F(g) and H = {H(f)j f E S} = F(h),
respectively, then
Ff(9 * h) = G(f)H(f). (6.29)

Proof: First consider the case of infinite duration continuous time signals.
In this case

FI(9' h) -Z ,-"r/' (Z g«()h(' - () d() <II

~ l (l ,-"<I'h(' - ()d')
g«() <1(,
268 CHAPTER 6. CONVOLUTION AND CORRELATION

Changing variables with t' = t - ( yields

J
00

= g(O (e- i27r /(H(f)) d(

-00

!
00

H(f) g(Oe- i21T /( d( = H(f)G(f),

-00

which proves the claim.

In the case of infinite duration discrete time signals the proof is the same
except that the integrals over continuous variables are replaced by sums
over discrete variables. For completeness we note this yields the sequence
of equalities

F/(g * h) = nJ;;oo e- i27r / n C~oo g(k)h(n - k))

= kf:oo g(k) c~oo e- i21T / nh(n - k))
kf:oo g(k) C,~oo e-i21T/(nl+klh(n'))
L
00

= g(k) (e- i21T / kH(f))

k=-oo

L
00

H(f) g(k)e- i21T / k = H(f)G(f).

k=-oo
In the case of the DFT the manipulations are similar with the sum over
infinite limits replaced by the cyclic sum. As usual let h be the periodic
extension of h defined by h(n) = h(n mod N). Then for any l E ZN

Ee-i21T~n (~1 g(k)k(n - k))

= ~ g(k) (~ e-i21Tilnh(n - k))

6.4. THE CONVOLUTION THEOREM 269

= ~ g(k) (:~: e- i2"j,(n'H) h(n'»)

= ~ g(k)e- 2"j,k (:~: e- i2"j,n' h(n'»)
= ~ g(k)e- 2"j,k (f: e- i2 "j.,n' h(n'»)

L g(k)e- 2"kk H( ~)
N-l
=
k=O
1 1
= H(N)G(N)'
The above proofs make an important point: in all cases the proofs look
almost the samej the only differences are minor. We used the functional
notation g(k) throughout instead of using gk for the discrete time case to
emphasize the similarity. We omit the proof for the case of finite duration
continuous time signals since the modifications required to the above proofs
should be clear.
We state without proof the dual result to the convolution theorem:
Theorem 6.4 The Dual Convolution Theorem
Given two signals 9 and h with spectra G and H, then
.1'( {g(t)h(t)j t E T}) = cG * Hj (6.30)
where c = 1 for infinite duration signals, liN for the DFT of duration N,
and liT for the CTFT of duration T. In words, multiplication in the time
domain corresponds to convolution in the frequency domain.
The extra factor comes in from the Fourier inversion formula. For example,
for the DFT case the Fourier transform of {gnhn; n = 0, ... , N - I} at
frequency kiN is
N-l N-l N-l
L gnhne- i2"nt, = L gn[N- 1 L H(~)ei2"mNle-i2"nt,
n=O n=O m=O
N-l N-l
= N- 1 L H(~) L gne-i2"n~
m=O n=O

L
N-l
= N- 1 H(m)G(k -m mod 1)
m=O N N

= N-IH*G(~).
270 CHAPTER 6. CONVOLUTION AND CORRELATION

The convolution theorem provides a shortcut to computing Fourier

transforms. For example, suppose that we wish to find the transform of
A(t). A straightforward exercise shows that A(t) is the convolution of n(t)
with itself. Since the Fourier transform of n(t) is sinc(f), we have easily
that
(6.31)
As a second and less obvious example, what is the Fourier transform
of the convolution of sinc(t) with itself? Duality implies that the Fourier
transform of sinc(t) is n(f), and hence the convolution theorem implies
that the transform of sinc * sinc{t) is just the product of n(f) with itself,
which is n(f) (except at the endpoints, which do not affect the Fourier
transform). Thus

Ff({sinc(t);t E 'R} * {sinc(t);t E 'R}) = n(f), (6.32)

and hence, after inverse transforming,

sinc * sinc(t) = sinc t. (6.33)

* Convolution Inverses Revisited

The convolution theorem also permits us to complete the discussion of signal
space algebra by showing how to define an inverse signal with respect to
convolution. Suppose that we have a signal g. We wish to find a signal (if it
exists) g-1 with the property that 9 *g-1 = 8. (Note that even in ordinary
algebra not every element of the space has an inverse, e.g., in the real
numbers 0 has no inverse.) To accomplish this take the Fourier transform
of both sides of the equality. The transform of a 8 function (Kronecker in
discrete time and Dirac in continuous time) is just a constant 1. If G- 1 is
the transform of the signal g-l, then we have that G(f)G-l (f) = 1 or
1
G- 1 (f) = G(f);
that is, the inverse of 9 is just the inverse Fourier transform of one over the
Fourier transform of g.
Clearly this inverse will not exist in general, e.g., if G(f) has zeros
then 1/G(f) is not well-behaved. For example, a band-limited function
cannot have an inverse. It can have an inverse in an approximate sense,
however. Suppose that g(t) is an infinite duration continuous time band-
limited signal with the property that G(f) > 0 if and only if If I < W; that
is, the spectrum is bandlimited and has no zeros inside of its bandwidth.
6.5. FOURIER ANALYSIS OF LINEAR SYSTEMS 271

In this case G(f) = G(f) n (f /2W) and we can define a "pseudo-inverse"

signal as the signal, say h(t), having spectrum
1 f
H(f) = G(f) n (2W)'
The signal h(t) is now well-defined and now the convolution 9 * h has spec-
trum n(f /2W) and hence
9 * h(t) = 2W sinc(2Wt); tEn.
As W --t 00, this convolution looks more and more like an impulse and the
pseudo-inverse looks more like an inverse.
Note that the Dirac delta has an inverse: itself, since d * d = d. This is
analogous to the number 1 being its own multiplicative inverse in the real
number system.
As a simple example of an exact inverse, consider the two infinite dura-
tion discrete time signals

gn = {Op
n n>O
th- . ; fn = dn - pdn-l =
{I n=O
-p n = 1
o erWlse 0 otherwise
Then the convolution is

9 * f(n) = f
k=-oo
gkfn-k = { ~
1 x pn - p X pn-l =0
~: ~
otherwise
and hence
g * f(n) = <5n.
In this case
G(f) = 1. j F(f) = 1- pe- i21T !.
1 - pe- t21T !

6.5 Fourier Analysis of Linear Systems

Since LTI systems are described by convolutions, the convolution theorem
immediately provides a means of relating the input and output spectra of
such a system.
Theorem 6.5 Suppose that C is an LTI system with <5-response h. Let
H = F(h) be the Fourier transform of the <5-response. If the input signal
v has Fourier transform V, then the output signal w is the inverse Fourier
transform of W = {W(f)j f E S}, where
W(f) = V(f)H(f). (6.34)
272 CHAPTER 6. CONVOLUTION AND CORRELATION

The Fourier transform of the o-response is called the system junction or

transfer function of the LTI system. Note that the transfer function is not
in general defined if the system is not time invariant. It is well-defined if
the system is time invariant but nonlinear, but in that case superposition
does not hold and hence the transfer function is of little use. The power
of the above result lies in the fact that it is much easier to multiply than
convolve. This is particularly useful if a signal v is the input to a cascade of
N linear systems with o-responses ht(t)j 1= 1, ... , N. The output signal u
is then the difficult N-fold convolution hI * h2 * ... * hN * v, but the overall
transfer function is found easily by
N
H(f) = II H, (f).
1=1

Eigenfunctions of LTI Systems

The second principal application of Fourier analysis to linear systems is a
simple observation of the fact that delta functions (Kronecker in discrete
time and Dirac in continuous time) are Fourier transform pairs with expo-
nentials. This fact has an interesting implication to Fourier analysis itself
in that it demonstrates why exponentials are so important. To state the
result first requires a definition.
Given a system C, an eigenfunction of C is an input signal e with the
property that
C(e) = Ae
for some complex constant Aj that is, an eigenfunction is a signal which is
passed through a system without any change of shape, the sole effect of the
system being to multiply the signal by a complex constant (not depending
on time). For those old (or cultured) enough to remember the Firesign
Theater, an eigenfunction can be viewed as an example of what they called
"Fudd's third law": What goes in must come out. The complex constant A
is called an eigenvalue of C.

Theorem 6.6 Let C be an LTI system with common input and output
signal types. Let T be the time domain of definition and S the frequency
domain of definition. Then for any fo E S, the signal

e = {e i2 11"Jotj t E T}

is an eigenfunction of c. The corresponding eigenvalue is H(fo), where H

is the transfer function of the system.
 6.5. FOURIER ANALYSIS OF LINEAR SYSTEMS 273

Proof: Given an input signal v and t5-response h, the output signal is

w = v * h. Transforming we have that W(f) = V(f)H(f); 1 E S. First
suppose the system is an infinite duration discrete time system, in which
case S = [-~, ~). The input signal Vn =
ei21rlon has Fourier transform
t5(f - 10), whence
W(f) = H(f)t5(f - 10) = H(fo)t5(f - 10)
which implies that

as claimed.
If the system is instead an infinite duration continuous time system,
in which case S is the real line, then v(t) = ei21rlot has Fourier transform
V(f) = o(f - 10) and the same equality chain as above proves the result.
If the system is finite duration and discrete time, then S = {~; k E Z N}
and Vn = ei21r/on (with 10 = -k for some l E ZN) has Fourier transform
t5 / -f.i for 1 E S. As previously (except that now only 1 E {~; k E ZN}
are possible)
W(f) = H(f) t5/-/o = H(fo) t5/-/o
which implies that

proving the result.

The finite duration continuous time case is left as an exercise.
The theorem indicates the reason why the complex exponentials are
the most important functions in the study of LTI systems: they are the
eigenfunctions of such systems. In general sines and cosines are not eigen-
functions. For this reason the decomposition of an arbitrary signal into a
linear combination (by integration or summation) of complex exponentials
is the most basic such decomposition; it breaks down an arbitrary signal
into pieces that pass through a linear system without changing shape. Thus
we can decompose an input signal into a linear combination of complex ex-
ponentials, infer the resulting output for each exponential (it is the same
exponential times the transfer function at the same frequency), and then
recombine these elementary outputs in a linear fashion (by summation or
integration) to find the complete output to the original signal.
As a corollary to the theorem, observe that if an infinite duration contin-
uous time LTI system has a purely real impulse response h, then an input
of cos(27r/ot) produces an output of IH(fo)lcos(27r/ot + LH(fo)), where
H(fo) = IH(fo)leiLH(fo). (This follows by expanding the cosine as the sum
of exponentials using Euler's relations, applying the eigenfunction property
274 CHAPTER 6. CONVOLUTION AND CORRELATION

to each, and then using the fact that if h is real, then H(- 1) = H*(f) to
write the resulting sum as ?R(H(fo)ei27r/ot).) Thus if H(f) also is purely
real, then the cosine is in fact an eigenfunction. In general, however, trans-
fer functions are not purely real and cosines are not eigenfunctions.
Another corollary to these results is that in an LTI system, only fre-
quencies appearing at the input can appear at the output. This is not the
case if the system is either nonlinear or time varying. For example, if the
system output w(t) for an input v(t) is given by w(t) = v 2(t) (a memoryless
square law device), then if v(t) = cos(21f'fot) we have that

w(t) = cos 2(21f'fot) = 2"1 + 2"1 cos(41f'fot)j

that is, new frequencies not appearing in the input do appear in the output.
In this case the frequency has been doubled. We have seen that a DSB-SC
modulation system is linear but time varying and that it too produces new
frequencies at the output. In fact, any modulation system which performs
frequency translation must either involve a nonlinear mapping or a time-
varying mapping.

6.6 The Integral Theorem

The convolution theorem provides a form of converse result to the derivative
theorem by relating the Fourier transform of the integral of a signal to the
Fourier transform of the original signal.
Suppose that 9 = {g(t)j t E R} is an infinite duration continuous time
signal and form the integral ¢(t) = J~oo g(() d(. Assuming that the integral
is well-defined (the signal g(t) is not too pathological), the Fourier transform

I:
of ¢(t) can be easily found from the convolution theorem by using the fact
that
¢(t) = H * g(t) = U-l(t - ()g(() d(, (6.35)

where U-l (t) is the unit step function

It ~ 0
U-l(t)= { 0 t<O'

Since the Fourier transform of the step function is ~ 8(f) - 2;1 (1 - 81), the
Fourier transform of ¢(t) is given by the convolution theorem as
1 G(f)
<P(f) = 2"G(O)8(f) + i21f'f (1 - 81)'
We have now proved the following result:
6.7. SAMPLING REVISITED 275

Theorem 6.7 The Integral Theorem

Given an infinite duration continuous time integrable signal g = {g(t); tE
R}, then

F/({[too g(() d(; t E R}) = ~G(0)6(f) + ~~J (1- 6/). (6.36)

The first term can be thought of as half the transform of the DC com-
ponent of get) represented by its area J~oo get) dt. The second term shows
that integration in the time domain corresponds to division by I in the
frequency domain (except where I = 0).
The discrete time analog to the integration theorem is the Fourier trans-
form of the sum (or discrete time integral) of a discrete time signal. Given
{gn; n E Z}, what is the Fourier transform of

The answer is left as an exercise.

The integration and the summation properties have dual results formed
by interchanging the roles of time and frequency. What are they?

6.7 Sampling Revisited

As another application of convolution and the convolution theorem we pro-
vide an alternative development of the sampling theorem. We do this in an
apparently more general fashion in order to better consider what happens
when the sampling frequency is not chosen in exactly the same fashion as
in the previous derivation. Assume now that 9 is a continuous time infinite
duration band-limited signal with W chosen so that G(f) = 0 for III ~ W
as before. Fix a sampling period T (which we do not assume to equal
1/2W). Multiply the signal g by the scaled impulse train

t
L
00

III(r) =T 6(t - nT)

n==-oo

to form the idealized sampled waveform

t
L
00

9(t) = III(r)g(t) = T g(nT)8(t - nT).

n=-oo
276 CHAPTER 6. CONVOLUTION AND CORRELATION

Using the convolution theorem, the fact that the ill function is its own
transform, and the scaling formula for delta functions yields

au) = Ff(9(t)) = Ff(ill(f)g(t))

Till(T 1) * G(f) (6.37)

(T n'foo 6(Tf - n)) * G(f)

C~oo 6(f - f)) * G(f)

L f); fER.
00

G (f - (6.38)
n=-oo

For example, given an input signal spectrum G(f) having the shape
depicted in Figure 6.8, then a
is the sum of an infinite sequence of shifted

G(f)

f
-W 0 W

Figure 6.8: Input Spectrum

copies of the original spectrum G as indicated in Figure 6.9.

H we choose T so that liT> 2W, the Nyquist frequency or Nyquist
a
rate, then the copies of the spectra do not overlap and is seen to consist of
an infinite sequence of replicas of G centered at frequencies niT for n E Z.
Since these spectral "islands" are disjoint, we can recover the original signal
g by passing the sampled signal 9 through an ideal low-pass filter, that is,
an operation which multiplies the input spectrum by the transfer function

(6.39)

where B is chosen so that W <B < ~ - W, selecting only the central

6.7. SAMPLING REVISITED 277

au)

f
1 1
2T T

Figure 6.9: Replicated Spectra

island. Thus the output of this operation has spectrum

, f
GU) n (2B) = GU),
the spectrum of the original signal! This action is depicted in Figure 6.10
where the dashed box is the filter magnitude which selects only the central
island.

au)

f
1 1
2T T

Figure 6.10: Signal Recovery

Again invoking the convolution theorem, taking inverse Fourier trans-

forms then yields

g(t) = g(t) * 2B sinc(2Bt)

278 CHAPTER 6. CONVOLUTION AND CORRELATION

(T n'foo g(nT)r5(t - nT») * 2B sinc(2Bt)

L
00

= 2BT g(nT) sinc[2B(t - nT)],

n=-oo

which reduces to the previously derived sampling expansion (4.14) when

T = 1/2B and B = W.
The above derivation points out what happens if we sample too slowly
in the sense that liT < 2W. In this case the repeated copies of G in
a overlap as shown in Figure 6.11, and there is no clear central island to

au)

f
1 1 2
2T T T

Figure 6.11: Overlapping Replicas

remove by low pass filtering. The figure shows the repeated copies and the
final spectrum is the sum of these copies, indicated by the curve forming the
"roof" of the overlapping islands. The low pass spectrum will be corrupted
by portions of other islands and this will cause the resulting signal to differ
from g. This distortion is called aliasing and some invariably occurs in any
physical system since no physical signal can be perfectly band-limited.
The final comment above merits some elaboration. The basic argument
is that all physical systems are time-limited assuming that the universe has
a finite lifetime. A signal cannot be both time-limited and band-limited
since, if it were, we could write for sufficiently large T and W that

g(t) = g(t) n (2~); GU) = GU) n (2~)

and hence, applying the convolution theorem to the relation on the left,

GU) = Ff(9(t» = Ff(9(t) n (2~» = GU) * 2Tsinc(2Tf).

 6.8. CORRELATION 279

This yields a contradiction, however, since the convolution with a sinc func-
tion expands the bandwidth of G(f), while multiplication by n(f /2W) in
general limits the extent of the spectrum.
As a final note, the sampling theorem states that a signal can be com-
pletely reconstructed from its samples provided l/T > 2W and that it is
not possible for l/T < 2W because the resulting aliasing by overlapping
spectral islands results in a distorted spectrum at low frequencies. In order
to avoid such aliasing when sampling too slowly, the original signal can be
first passed through a sharp low pass filter, that is, have its spectrum multi-
plied by n(f /2W), and then this new signal will be recreated perfectly from
its sampled version. The original low pass filtering introduces distortion,
but it results in a signal that can be sampled without further distortion.

6.8 Correlation
Correlation is an operation on signals that strongly resembles convolution
and which will be seen to have very similar properties. Its applications,
however, are somewhat different. The principal use of correlation functions
is in signal detection and estimation problems and in communications the-
ory where they provide a measure of how similar a signal is to a delay of
itself or to another signal. It also is crucial in defining bandwidth of sig-
nals and filters (as we shall see) and in describing the frequency domain
behavior of the energy of a signal.
Suppose, as earlier, that we have two signals 9 = {g(t)j t E f} and
h = {h(t)j t E f}. The cross correlation function rgh(r) of 9 and h is
defined for the various signal types as follows:

00 00
f g*(t - r)h(t) dt =f g*(t)h(t + r) dt CTIDj
-00 -00
00 00
L: g*(n - r)h(n) = L: g*(n)h(n + r) DTIDj
n=-oo n=-oo
T T _
f g* (t - r)h(t) dt = f g*(t)h(t + r) dt CTFDj
o 0
N-l N-l
L: g*(n - r)h(n) = L: g*(n)k(n + r) DTFD,
n=O n=O
where as usual g denotes the periodic extension and where the finite dura-
tion correlation is a cyclic correlation (as was convolution). Analogous to
the asterisk notation for convolution we abbreviate the correlation opera-
tion by a star: r gh = 9 * h. The argument of the correlation function (r
above) is often called the lag.
280 CHAPTER 6. CONVOLUTION AND CORRELATION

The autocorrelation function r 9 (7) of a signal g is the correlation of g

with itself:
rg(7) == rgg(7) = 9 * g(7).
Note that the fact that one of the signals is conjugated is implied by, but
not specifically stated in the shorthand notation. Both crosscorrelation and
autocorrelation functions are referred to as simply correlation functions.
The convolution and correlation operations on signals are similar in
that both take the product of two signals and then integrate (if continuous
time) or sum (if discrete time). The operations differ in that correlation
involves the complex conjugate of one signal (and convolution does not)
and convolution reverses one signal in time (and correlation does not). In
fact, we can express an crosscorrelation as a convolution as follows: given
signals hand g, define g_ to be the signal 9 reversed in time; that is, if
g = {g(t); t E T}, then g_ = {g( -t); t E T}. Then

(6.40)

A major difference between the convolution and correlation operations is

that convolution is commutative, i.e., g*h = h*g, but in general rgh i= r hg .
In fact, in the continuous time infinite duration case

J
00

r hg (7) h*(t)g(t + 7) dt
-00

J
00

h*((-7)g(Od(

(l
-00

= g·(Oh((-7)d()·

= r;h( -7). (6.41)

The same result holds for discrete time and finite duration signals.
A function r( 7) with the property that r( -7) = r* (7) is said to be
Hermitian and hence we have proved that autocorrelation functions are
Hermitian. If g is a real function, this implies that the autocorrelation
function is even.
All of the definitions for correlation functions were blithely written as-
suming that the various integrals or sums exist. As usual, this is trivial in
the discrete time finite duration case. It can be shown that the other defi-
nitions all make sense (the integral or the limiting integral or sum exists) if
6.8. CORRELATION 281

the signals have finite energy. Recall that the energy of a signal 9 is defined
by
f Ig(t)12 dt continuous time;
£ _ { tET
9 - 2: Ig(n)12 discrete time.
nET
Note that the autocorrelation of a signal evaluated at 0 lag is exactly this
energy; that is,
(6.42)
It is often convenient to normalize correlation functions by the signal
energies. Towards this end we define the correlation coefficient

() rgh(r)
'Ygh T = J£g£h (6.43)

and the corresponding normalized autocorrelation function

(6.44)

In order to derive one of the principal properties of correlation functions,

an upper bound on the magnitude of the correlation in terms of the ener-
gies of the signal, we require one of the most useful inequalities of applied
mathematics: the Cauchy-Schwarz inequality. Because of its importance
we pause from the development to state and prove the inequality in its
general form.

Theorem 6.S The Cauchy-Schwartz Inequality

If {g(t); t E T} and {h(t); t E 7} are two complex-valued continuous
time functions on 7, then

Ir
itET
g(t)h(t) dtl2 ~
itET
r Ig(tW dt r
itET
Ih(tW dt (6.45)

with equality if g(t) = Kh*(t) for some complex K; that is, 9 is a complex
constant times the complex conjugate of h.
Similarly, if {gn; nET} and {h n ; n E 7} are two complex-valued
discrete time signals on 7, then

~ L Ignl 2 L Ihnl2 (6.46)

nET nET
282 CHAPTER 6. CONVOLUTION AND CORRELATION

Proof: We prove the result only for the continuous time (integral) case. The
corresponding result for discrete time follows in exactly the same manner
by replacing the integrals by sums. Most proofs in the literature use a
calculus of variations argument which is needlessly complicated. The proof
below is much simpler. It is based on a simple trick and the fact that the
answer is known and we need only prove it. (The calculus of variations is
mainly useful when you do not know the answer first and need to find it.)
Define as usual the energy of a signal by

Cg = r
itET
Ig(tW dt.

If either signal has infinite energy, then the inequality is trivially true.
Hence we can assume that both signals have finite energy. Observe that
obviously

( lg(t)1 _ Ih(t)I)2 > O.

vr; v'0. -
Expanding the square we have that

or
Ig(t)12 + Ih(tW > 2 Ig(t)IIh(t)l.
cg Ch - JC9Ch
Integrating over t then yields

1+1 ~ ~
VCgCh
r
itET
Ig(t)h(t)1 dt.

The right hand side above can be bounded from below using the fact that
for any complex valued function x(t)

r
itEr
Ix(t)1 dt r x(t) dtl
~ IitEr (6.47)

and hence setting x(t) = g(t)h(t)

JCgCh ~ IlET g(t)h(t) dtl '
which proves the Cauchy-Schwartz inequality. If get) = Kh*(t), then both
sides of the inequality equallKlc; and hence it is an equality. The integral
6.8. CORRELATION 283

inequality is due to Schwartz and the sum inequality is due to Cauchy.

The name Cauchy-Schwartz inequality was originally given to the general
inequality for inner products of which these are both special cases, but the
name is often used for the special cases as well.
We now can apply the Cauchy-Schwarz inequality to derive the promised
bound: we claim that
(6.48)
and hence in particular that

(6.49)

Thus the autocorrelation function achieves its maximum at zero lag. We

prove this for the continuous time case; the discrete time case follows simi-
larly. Just replace get) in the Cauchy-Schwartz inequality by g*(t) and h(t)
by h(t + 7) and we have that

(6.50)

since the energy of g* is also the energy of get) and the energy in h(t + 7)
is also the energy in h(t).
We summarize for latter use the principal properties of autocorrelation
functions:

1. fg(O) = £g, ,g(O) =1

2. fg(-7) = f;(7), ,g(-7) = ,;(7) (Hermitian)
3. /fg(7)1 :S /fg(O)I, l,g(7)1 :S 1.
4. The autocorrelation is shift invariant in the sense that if h is a shifted
version of g (h(t) = get - to) for some to), then fg(7) = fh(7).

The notion of an autocorrelation function can be extended to signals

i:
with infinite energy but finite average power by a suitable normalization.
For example, if
Ig(tW dt = 00, (6.51)

but
lim 21T
T ..... oo
jT Ig(t)1 2dt <
-T
00, (6.52)

then we can define the generalized autocorrelation function

fg(7) = lim 21T

T ..... oo
jT g*(t - 7)g(t) dt.
-T
(6.53)
284 CHAPTER 6. CONVOLUTION AND CORRELATION

As one might expect, the similarity between convolution and correlation

will lead to similar properties when the Fourier transform is taken.
Theorem 6.9 The Correlation Theorem

F,(y * h) = G*(f)H(f)j (6.54)

that is, the Fourier transform offgh(r) is G*(f)H(f). Thus also

F,(y*y) = IG(fWj (6.55)

that is, the Fourier transform of the autocorrelation is the squared magni-
tude of the Fourier transform of the signal.
The proof parallels very closely the proof of the convolution theorem
and hence is not presented. Alternatively, the result follows immediately
from the representation of (6.40) of a correlation as a convolution combined
with the fact that if y_ is the signal y reversed in time, then in the infinite
duration continuous time case

J ,t
00

;::"(g:") = g*( _t)e- i21f dt

-00

- Jg*«()ei21f'~
-00

= d(
00

J
00

y*«()e i21f f( d(
-00

Thus
F,(g*h) = F,(g:"*h)
= G*(f)H(f).
An implication of the correlation theorem is that the Fourier transform
of the autocorrelation is real and nonnegative.
Note that all phase information in the spectrum G(f) is lost in IG(f)12.
This implies that many functions (differing from one another only in phase)
have the same autocorrelation function. Thus the mapping of y(t) into
f 9 (r) is many to one. In general, without further a priori information or
restrictions, a unique y(t) cannot be found from fg(r).
6.9. PARSEVAL'S THEOREM REVISITED 285

6.9 Parseval's Theorem Revisited

The correlation theorem stated that given two signals 9 and h with trans-
forms G(f) and H(f), then the Fourier transform of their crosscorrelation
rgh = 9 * h is given by G*(f)H(f). For example, in the infinite duration

I: I:
continuous time case this is

rgh(r) = g*(t)h(t + r) dt = G*(f)H(f)ei2trIT df·

Application of the inversion formula then implies that r gh (r) must be the
inverse transform of G*(f)H(f). Applying this result to the special case
where T = 0 immediately yields the general form of Parseval's theorem of
Theorem 4.5. The general form is

rgh(O) = F01({G*(f)H(f)j f E S}). (6.56)

6.10 * Bandwidth and Pulsewidth Revisited

The properties of autocorrelation functions can be used to provide addi-
tional definitions of bandwidth and pulsewidth of signals and to relate the
two quantities.

Autocorrelation Width
Because of the shortcomings of the definition of equivalent width, it is
desirable to find a better definition not having these problems. Toward
this end we introduce the autocorrelation width of a signal defined (easily)
as the equivalent width of the autocorrelation of the signal. Since the
autocorrelation function has its maximum at the origin, the autocorrelation
width of 9 is defined by

From the correlation theorem, the Fourier transform of r(t) is IG(f)12 and
hence from the zeroth moment property

where the area property was used to express the denominator in terms of
the spectrum. The right hand side is just one over the equivalent width
286 CHAPTER 6. CONVOLUTION AND CORRELATION

of the magnitude squared spectrum, IG (1) 12 , and hence as with equivalent

width,
Wry WIGI2 = 1,
the corresponding widths in the time and frequency domain are inversely
proportional.
The autocorrelation width of a signal does not give a nonsensical result
if the signal has zero area, but it can still give nonsensical results. Consider,
for example, the autocorrelation width of the signal sinc(t) cos(101I"t), which
can be thought of as an amplitude modulated sinc function. The spectrum
of this signal can be found from the modulation theorem to be

Thus IG(O)12 = 0 and hence the autocorrelation width of this signal is O.

The autocorrelation width does provide a meaningful result for the spe-
cial case of a Dirac delta. The autocorrelation function of a delta function
is another delta function (as with the convolution of two delta functions)
and hence Wry = 1/00 = 0, which is what one would expect for a Dirac
delta.

Mean-Squared Width
Yet another definition of width is the standard deviation of the instanta-
neous power or the mean squared width, which we denote f:l.tg. It is defined
as the square root of the variance
2
0"Igl 2
= < t2 >lgl 2 - <t >lgl2

J~oot2Ig(tWdt _ (J~ootlg(t)12dt)2
J~oo Ig(t)12 dt J~oo Ig(t)12 dt

Unlike the previous moments we have considered, the magnitude square of

the signal is always nonnegative and hence the weighting of the integrand
is nonnegative. This width is independent of the origin (not affected by
translation) and it is strictly positive for all physical functions. Its primary
difficulty is that it can be infinite for a signal with a physically finite area.
Consider, for example, the signal g(t) = sinc(t). We have that

roo
Loo t 2 sinc2 (t) dt =
1
11"2
roo
Loo sin 2 (1I"t) dt = 00.
6.10. * BANDWIDTH AND PULSEWIDTH REVISITED 287

The corresponding frequency domain width is the standard deviation of

the energy spectrum, the square root of the variance

These two quantities have a famous relation to each other called the
uncertainty relation which is given in the following theorem.
Theorem 6.10 The Uncertainty Relation
For any continuous time infinite duration signal g(t) with spectrum
GU), the timewidth-bandwidth product tltgtlJo (called the uncertainty prod-
uct) satisfies the following inequality:
1
tltgtlJo ~ 411"' (6.57)

The bound is achieved with equality by the Gaussian signal.

The uncertainty relation can be interpreted as saying that one cannot
have too narrow a width in both domains simultaneously. This is not as
simple a relation as the corresponding time-bandwidth products of the other
width definitions, but at least it provides a lower bound to the width in one
domain in terms of the inverse of the width in the other.

I: I:
Proof of the Uncertainty Relation: For convenience we assume that

Ig(t)12 dt = IGUW df =1

I: tlg(tW dt = I: JIGUW dJ = O.

Define
I: t 2Ig(t)12 dt

I:
(tJ.t)2 =

(tlf? = J 21GUW df.

The goal is then to prove that

(tJ.t)(tJ.f) ~ 4~'
1. If D..t = 00, the relationship is obviously true, so we assume that
tJ.t < 00. If this is true, then limHoo tlg(t)12 = 0 since otherwise the
integral would blow up.
288 CHAPTER 6. CONVOLUTION AND CORRELATION

2. Applying the Cauchy-Schwarz inequality yields

with equality if tg(t) = Kd~~t).

3. To evaluate the right-hand side of the previous equation, first note
that f~oo t 2lg(tWdt = (Llt)2. Next observe that since d~~t) -+ i21rfG(f),
Rayleigh's theorem implies that

Thus

4. Now consider the left-hand side. For any complex number z we have
that

i:
whence

1 (tg(t))(;tg*(t)) dtl ~~ Ii: (tg*(t) :tg(t) + tg(t) :tg*(t)) dtl·

Furthermore, we have that

* dg dg* d * d 1 12
9 dt + g"dt = dt gg = dt 9 .

Thus we have

5. Integrating the expression inside the magnitude of the left-hand side

of the previous equation by parts yields

But f~oo Igl2dt = 1 and from point 1, tlgl 2 = 0 at -00 and 00.
6.11. * THE CENTRAL LIMIT THEOREM 289

6. Finally, we have
~(_1)2 ~ 47T2(At)2(Af)2
or
AlAf ~ 4~'
When does AtAf actually equal 1/47T? In the first inequality (step
(2», equality is achieved if and only if tg(t) = k/tg(t) for some complex
constant k. In the second inequality (step (4» equality is achieved if and
only if Z = J~oo tg(t)/tg(t)* dt is real valued. Thus if tg(t) = k/tg(t) for
some real constant k, then both conditions are satisfied since then Z =
J~oo t 2Ig(t)12 dt/k. Thus the lower bound will hold if

d
k dtg(t) - tg(t) = 0

for some real-valued k. The solution to this differential equation is

= ce2i".
,2
g(t)

In addition we require that limltl-too tlg(t)1 = 0 or 2k = -l/Ct (for Ct > 0)

and J~oo Ig(t)12 dt = 1 or c = {f2Ct/7T. Thus the lower bound is achieved
by a signal having the form

g () = f#
t
2Ct _<>t2
-e
7T
for Ct > O. It can be shown that a signal achieving the lower bound neces-
sarily has this form.

6.11 * The Central Limit Theorem

The convolution of n(t) with itself yields A(t). Let us denote this self
convolution by n(t)*2. If in turn we convolve A(t) with net), we get a new
signal that is more spread out and smoother. As this is the convolution of
three copies of the signal we denote the result by n(t)*3. We can continue
convolving further copies of net) with itself in this way. The question
now is, if we convolve n(t) with itself n times to produce n(t)*n, does
the result tend to a limit as n gets large? More generally, what happens
when we convolve an arbitrary signal with itself n times? Does it converge
to a limit and is the limit different for different signals? The answer is
called the central limit theorem and states that if the signal is suitably
290 CHAPTER 6. CONVOLUTION AND CORRELATION

normalized in time scale and magnitude and satisfies certain constraints

to be discussed, then the result of convolving it with itself many times
will be a Gaussian signal, regardless of the original shape. The moment
properties play a key role in proving this result. The key applications of
approximating the result of a multiple convolution of a signal with itself are
in systems theory and probability theory where the convolutions correspond
respectively to cascading many identical linear filters and summing many
identically distributed independent random variables.

Theorem 6.11 The Central Limit Theorem

Given an infinite duration continuous time signal get) with Fourier
transform C(f), suppose that C(f) has the property that for small f
C(f) ~ a - Cf2 (6.58)

for positive a and c.

Then
(6.59)

The proof of the theorem will take the remainder of the subsection.
First, however, some comments are in order.
The small f approximation may seem somewhat arbitrary, but it can
hold under fairly general conditions. For example, suppose that C(f) has
a Taylor series at f = 0:

= L bkf k ,
00

C(f)
k==O

where
C(k) (0)
bk = k!
(the derivatives of all orders exist and are finite). Suppose further that
1. bI = C' (0) = 0 (This is assumed for convenience.)
2. bo = C(O) > O. Define a = boo
3. b2 = C"(0)/2 < O. Define c = -b 2 .
These assumptions imply (6.58), an equation which is commonly written
as
C(f) =a - cf2 + o(f2),
where o(j2) means a term that goes to zero faster than j2. In fact this is
all we need and we could have used this as the assumption.
 6.11. * THE CENTRAL LIMIT THEOREM 291

An additional implication of assuming that G(f) has a Taylor series is

that the entire spectrum can be computed from the moments by using the
moment properties; that is, since

then

G(f) =~ (-211"i)k < tk > fk =~ (-211"if)k < tk >

~ k! 9 ~ k! g'
k=O k=O

This formula provides one of the reasons that moments are of interest-they
can be used to compute (or approximate) the original signal or spectrum if
it is nice enough to have a Taylor series.
The condition that b1 = G/(O) = 0 is satisfied by any real even G(f).
From the first moment property this implies also that < t >g= O.

Proof of the Central Limit Theorem: From the stretch theorem, the signal
y'ng( y'nt) has Fourier transform G(f I y'n). From the convolution theorem,
(y'ng( y'nt)) *n has Fourier transform Gn (f I y'n). Thus

= an y'n
P
f +0(-)
-1 ( a-c(-)2
n
)n
=
c PIP
( 1- - - + -0(-)
)n
a nan

where the o(P In) notation means a term that goes to zero with increasing
n faster than Pin. From elementary real analysis, as n --+ 00 the right-
most term goes to e- cl2 fa. This result is equivalent to the fact that for
small f,
In(! - f) ~ -f.
Thus for large n

Gn(fly'n) ~ (1- ::.P)n ~ e-~.

an a n

The .Inverse .Founer transform of e- !!E..

a = e - 11" (v""
rI,)2. /Jflf _ ( ~t)2
IS V ¥'e 11" V C .
We now assume that we can interchange limits and transforms to infer that
the inverse transform of a limit of the transforms is the limit of the inverse
292 CHAPTER 6. CONVOLUTION AND CORRELATION

transforms. Applying this to (vng(:,:nt) rn should give us the limit of the

(vng(vnt»)*n
an
h· h . Id
, W lC Yle s

· h,mg(y'nt»*n
11m = /fa
-e -7r(·lllt)2
Yo ,
n-+oo an c

which completes the proof of the theorem.

The central limit theorem can be used to provide an approximation for
an n-fold self convolution g(t)*n by suitable scaling. We have seen that

and hence
c /2
Gnu) ~ ane-"n .
Inverting this result we have that

As an example consider the signal

g(t)= p.
nU·)
7r 1 - t

Does g(t)*n converge to a Gaussian signal? If so, what is its approximate

form? If not, why not? From the Fourier transform tables, the transform of
g(t) is Jo(27r I). Using the series representation Eq. 1.13 for Bessel functions,
it can be verified that GU) indeed has finite derivatives of all orders and
that
= =
G(O) a 1 > 0,
G'(O) = 0,
and
G"(O)
c= - - - =7r 2 >
2
o.
From Eq. 1.13, for small f J o(27r f) ~ 1 - 7r 2 P and hence a = 1 and c = 7r 2 .
Since all of the conditions are met, the CLT holds and we have for large
n that
6.12. PROBLEMS 293

As another example, consider the same questions asked about get) =

JO(21Tt). Here the transform is

G(f) = n(f) .
1T~
Here G(O) > 0 as required, but Gil (0) > 0 violates the sufficient condition
and hence the CLT cannot be applied. What is happening here is that for
small 1
G(f) = .!:.(l - 1 2)-! ~ .!:. + ~ 12.
1T 1T 21T
Thus a = 1/1T and c = -1/1T. A positive c is required, however, for the
derivation to hold.

6.12 Problems
6.1. A certain discrete time system takes an input signal x = {xn; n E Z}
and forms an output signal Y = {Yn; n E Z} in such a way that the
input and output are related by the formula

Yn = Xn - aYn-l, n E Z;

where lal < 1.

(a) If the signals x and Y have Fourier transforms X and Y, respec-
tively, find Y(f)/ X(f).
(b) Find the a-response of the system, that is, the response to a
Kronecker delta a = {On; n E Z}.
(c) Express the output Y as a convolution of the input x and a signal
h (which you must describe).
(d) Is the system linear? Time invariant?

6.2. The formula (3.79) is a convolution. Use the convolution theorem to

relate the Fourier transform G a of 9a to G, the Fourier transform of
g. Describe what happens to both G a and 9a as a -t 00.

6.3. Suppose that an infinite duration, continuous time linear time invari-
ant system has transfer function H(f) given by

H(f) = -isgn(f) = { ee ~" if f > 0

2 if 1 < 0
294 CHAPTER 6. CONVOLUTION AND CORRELATION

This filter causes a 90 0 phase shift for negative frequencies and a -90 0
phase shift for positive frequencies. Let h(t) denote the corresponding
impulse response. Let v(t) be a real valued signal with spectrum V(f)
which is band limited to [-W, W]. Define the signal v = v * h, the
convolution of the input signal and the impulse response.

(a) Show that

v(t) =! r
v(,x) d,x,
7r J>'ER,>4t t - ,x
where the integral is interpreted in the Cauchy sense. The neg-
ative of v(t) is called the Hilbert transform of the signal v(t).
(b) Suppose that the spectrum of v(t) has the shape given in Fig-
ure 6.8. For fc > > W sketch the spectrum of the waveform
vc(t) = v(t) cos(27rJct) + v(t) sin(27rJct).
This is called single-sideband amplitude modulation or SSB for
short. How does the bandwidth of the above signal compare
with that of ordinary double sideband-suppressed carrier (DSB)
modulation y(t) = v(t) cos(27r Jct)?
6.4. Using naked brute force, prove that convolution satisfies the associa-
tivity property:

6.5. Prove the dual convolution theorem (6.30).

6.6. Evaluate the convolution

6.7. Evaluate the following convolutions using Fourier transforms.

(a) Jo(27rt) * sinc(2t)

(b) sin(7rt) * I\(t).
(c) sinc2 (2t) * sinc(2t)
(d) cos(7rt) * e- 7rt2

6.8. The error function erf(t) is defined by

erf(t) = Jrr lot e-(2 d(.

Find the Fourier transform of erf(t).

6.12. PROBLEMS 295

6.9. Find and sketch the following convolutions and cross-correlations. All
signals are infinite duration continuous time signals.

(a) t 2U_l (t)* etU_l (t)

(b) net + 1/2) * n(3t + 1)
(c) n(t-1)*n(3t+3)
(d) net) * net) * net)
6.10. Find and sketch the following convolutions:

(a) {tU_l(t); t E R} * {etu_l(t);t E R}

(b) {net - 1); t E R} * {n(2t + 4); t E R}
(c) {anU_l(n); n E Z} * {U_l(n); n E Z}

6.11. Find and sketch the following convolutions and cross-correlations:

(a) net - 1) * n(2t + 3)

(b) net - 1) * n(2t + 3)
(c) n(t+1)*n(2t-3)

6.12. Find the convolution A(t) * A(t).

6.13. Given a signal 9 = {get); t E 'R}, denote by gT the shifted signal
defined by gT(t) = get - r). Show that

that is, shifting does not affect autocorrelation.

6.14. Define the discrete time signals (7 = Z)

_ {I, n=0,1,2,3,4;
gn - 0, otherwise

and
h n = {(~)n, n=O,l, ... ;
0, otherwise.

(a) Find the autocorrelation function r 9 (n) of 9 and provide a la-

beled sketch.
(b) Sketch the signals Xn = gn+2 and Yn = h 2n . Find the Fourier
transforms of these two signals.
296 CHAPTER 6. CONVOLUTION AND CORRELATION

6.15. Find the autocorrelation function and the Fourier transform of the
signal {e i27rk / 8 ; k = 0, 1, ... , 7}. Find the circular convolution of this
signal with itself and the Fourier transform of the resulting signal.
6.16. Find and sketch the auto correlations and the crosscorrelation of the
discrete time signals

n = 0,1, ... ;
otherwise

(where Irl < 1) and

hn = {e- 0,
in , n = -N,-N
otherwise.
+ 1, ... , -1,0, 1, .. . ,N;

Verify in this example that

6.17. Evaluate the continuous time convolution t5(t - 1) * t5(t - 2) * n(t - 1).
6.18. Find a continuous time signal g(t) whose autocorrelation function is

rg(t) = A(2(t - 4)) + 2 A (2t) + A(2(t + 4)).

Is the answer unique?
6.19. Suppose that h = f * g, where the signals are infinite duration con-
tinuous time. Prove that the area under h equals the product of the
areas under f and g; that is,

1: h(t) dt = (1: f(t) dt) (1: g(t) dt) .

6.20. Evaluate the continuous time convolutions t5(2t + 1) * n(t/3) and c5(t-
0.5) * c5(t + 1) * sinc(t).
6.21. Two finite duration discrete time complex-valued signals 9 = {gn; n =
0,1, ... , N -I} and h = {h n ; n = 0,1, ... , N -I} are to be convolved
(using cyclic or circular convolution) to find a new signal y = 9 * h.
This can be done in two ways:
• by brute force evaluation of the convolution sum, or
• by first taking the DFT of each signal, then forming the product,
and then taking the IDFT to obtain 9 * h.
6.12. PROBLEMS 297

Assuming that an FFT algorithm is used to perform both the DFT

and IDFT, estimate the number of real multiplies required for each
approach. Do not count the multiplication by l/N in the inverse
DFT (the IFFT). (In practice, N is a power of 2 and this can be
accomplished by bit shifting.)
Specialize your answer to the case N = 1024.

6.22. The ambiguity function of a signal {g(t); t E 'R} is defined by

X(T, I) = [: g*(()g(( + T)e- i2tr !( d(.

(This function is important in radar signal analysis.)

(a) What is the maximum value of IX(T, 1)1 over all T and f. When
is it achieved?
(b) Evaluate X(T, I) for g(t) = n(t).

6.23. Define two discrete time, infinite duration signals x = {xn; n E Z}

and Y = {Yn; n E Z} by

where u_I(n) is the unit step function (1 if n ;::: 0 and 0 otherwise)

and Ipi < 1,
Yn = c5n - On-I,
where on is the Kronecker delta (1 if n = 0 and 0 otherwise).
Find the Fourier transforms of the following signals:

(a) x and Y
(b) {X n -5; n E Z}
(c) {X5-n; n E Z}
(d) {xlnlj n E Z}
(e) {xn cos(27Tn/9)j n E Z}
(f) {xnYn; n E Z}
(g) x * Y (the convolution of x and y)
(h) x +Y
(i) {x~; n E Z}
298 CHAPTER 6. CONVOLUTION AND CORRELATION

6.24. Suppose that you are told two continuous time, infinite duration sig-
nals x = {x(t); t E R} and h = {h(t); t E R} are convolved to
form a new signal y = x * h = {y(t); t E R}. You are also told that
x( t) = e-tu_l (t) and

1
y(t) = Ol/2(t - 2") - e-tu_l(t) + e-(t-l)U_l(t -1).

(a) Find the Fourier transform Y(f) of y.

(b) Find and sketch the signal h. (This is an example of deconvolu-
tion.)

6.25. Find the circular convolution of the sequence {I, 1,0,1,0,0,0, O} with
itself.

°
6.26. Let g(t) be a bandlimited signal with spectrum G(f) for If I ~ w.
We wish to sample g(t) at the slowest possible rate that will allow
recovery of
[ : g(t) dt.

We do not wish to reconstruct g(t), only to know its area. What

is the minimum sampling rate that will allow recovery of the above
area? How do we express the above integral in terms of the samples
g(t)?
6.27. Evaluate the following integrals using Fourier transforms:

(a)

(b)

(c)
[ : (sinc(t +~) + sinc(t - ~))2 dt.
6.28. Evaluate the following integral using Fourier transforms (there is an
easy way).
6.12. PROBLEMS 299

6.29. Suppose that an infinite duration continuous time signal g(t} has a
zero spectrum outside the interval (- ~, ~). It is asserted that the
magnitude of such a function can never exceed the square root of its
energy. For what nontrivial signal g(t} is equality achieved? That is,
for what function is the magnitude of g(t) actually equal to ..;t; for
some t? Prove your answer. (Hint: G(J) = n(J)G(J).)

6.30. A signal g(t} is known to have a spectrum that is non-zero only in

the range 4B < III < 6B. This signal could, of course, be recovered
from samples taken at a rate at least as large as 12B samples per
second. It is claimed that g(t} can be recovered from samples taken
at a rate much slower than 12B samples per second using a single
linear time-invariant interpolation filter. What is the slowest rate for
which this is true? Describe the interpolation filter.

6.31. A bandlimited signal of maximum frequency W is to be sampled using

a sampling signal ill(t/T} where l/T < 2W. Specify the bandwidth
of the widest rectangular low-pass filter to be used prior to sampling
which will eliminate aliasing. If no filter is used prior to sampling,
specify the widest rectangular low-pass filter that can be used after
sampling to provide an output free from aliasing.

6.32. A continuous time signal g(t} is bandlimited and has 0 spectrum G(J)
for all I with III ~ W.

(a) Instead of ideal sampling with impulses, the signal is sampled

using a train of very narrow pulses, each with pulse shape p(t)
and spectrum P(J). The resulting sampled pulse train is given
by

L
00

g(t) = g(t)T p(t - nT).

n=-oo

This is sometimes called natural sampling. What is the effect

of the nonideal pulse shape on the procedure for recovering g(t)
from g(t}? State any assumptions you make relating T, W, and
p(t}.
(b) Next suppose that the signal is first multiplied by an ideal im-
pulse train to produce a sampled waveform

L
00

y(t} =T g(nT)8(t - nT),

n=-oo
300 CHAPTER 6. CONVOLUTION AND CORRELATION

but y(t) then is passed through a filter whose output spectrum

U(f) is given by Y(f)P(f) so that the final sampled signal is

L
00

u(t) =T g(nT)p(t - nT).

n=-oo

How can the original signal be recovered from this signal?

6.33. What is the convolution of the infinite duration continuous time sig-
nals
1
1 + [211"(t - 1))2
and
1 ?
1 + [211"(t + 1))2 .
6.34. Name two different nontrivial signals or generalized signals 9 that
have the property that 9 * 9 = g, Le., the signal convolved with itself
is itself. (A trivial signal is zero everywhere.) What can you say
about the spectrum of such a signal?

6.35. Suppose that 9 is an arbitrary continuous time infinite duration signal

and a new signal y is formed by convolving 9 with a box function 01/2.
What can be said about Y (k) for integer k?

6.36. You are given a linear, time-invariant system with transfer function
H(f) = e- 1r / 2 , fEn.

(a) Suppose the input to the system is x(t) = e- 1rt2 , tEn. Write
the output v(t) as a convolution integral of the input with an-
other function h(t), which you must specify.
(b) Find the output v(t) with the input as given in part (a).
(c) A signal g(t) = 3 cos 311"t, tEn, is input to the system. Find the
output.
(d) The output to a particular input signal y(t) is e- 1rt2 , t E R.
What is the input y(t)?
(e) Now suppose that the signal x(t) of part (a) is instead put into
a filter with impulse response 01/2(t) to form an output w(t).

I:
Evaluate the integrals

w(t)dt
6.12. PROBLEMS 301

and
[ : tw(t)dt.

6.37. Suppose that we approximate the impulse train interpretation of the

sampling theorem by substituting narrow tall square pulses for the
Dirac deltas as follows. Define the pulse p(t) to be 1/ ~ for It I :$ ~/2
and then define the pulse train

L
00

s(t) = p(t - nT).

n=-oo

Assume that 9 = {g(t); t E 'R.} is a bandlimited signal with a Fourier

transform G such that G(f) is nonzero only for If I :$ W. Define the
signal y(t) = g(t)s(t). (If the p(t) were indeed impulses, then this
would be the idealized impulsive PAM signal.)
Find the Fourier transform Y (f) of y and show that if T is chosen
large enough (and specify what that means), then 9 can be recovered
from y. Show how to accomplish this recovery.
6.38. An applications problem (courtesy of Rick Wesel):
Modem Training Tones. You may have noticed that when a modem
first begins transmission it makes a tonal sound before the white-noise
"shhhhh" sound of normal data transmission. In this problem you will
use a simplified model of modem training to determine (quite accu-
rately) the frequencies and relative power of these "training tones"
for the V.29 modem.
The V.29 modem sends a training signal that alternates between
two phase-amplitude points of a 1700 Hz sinusoid (the carrier). The
switching between the two points occurs at the symbol rate of 2400
Hz.
For our calculation, one of these points will have zero amplitude, and
the other point will be a unit amplitude cosine. Make the (obviously
untrue) assumption that the training signal is bi-infinite to simplify
your calculations. When the cosine is switched back on, it picks up
where it would have been if it had been left on, not where it was
when it was switched off. Note that the training signal can then
be modeled as s(t) = cos(27rl700t)p(t), where p(t) is a square wave
oscillating between 0 and 1 and holding each value for the duration
of a single (ASCII) symbol.

(a) How many tones are there in this signal?

302 CHAPTER 6. CONVOLUTION AND CORRELATION

(b) What are the frequencies of the three most powerful tones?
(c) What are the frequencies of the next two most powerful tones?
(d) What is the power difference in dB between the most powerful
of the three terms and the next two most powerful?

6.39. Let p = {p(t); t E [0,2)} be the signal defined by

(t) = { +1 0$ t <1
P -1 1$ t < 2'
(a) Find the Fourier transform P of p.
(b) Find a Fourier series representation for p.
(c) Find a Fourier series representation for p, the periodic extension
of p with period 2. Provide a labeled sketch of p.
(d) Find the Fourier transform P of p.
(e) Suppose that p is put into a filter with an impulse response h
whose Fourier transform H is defined by

H (f) = {Io otherwise

1f ± ~ I. l $

to produce an output y = {y(t); t E 'R.}. (This is called an ideal

bandpass filter.) Find a simple expression for y(t).

6.40. Suppose that a discrete time system with input x = {xn; n E Z} and
output y = {Yni n E Z} is defined by the difference equations
M
Yn = aXn - L akYn-k; nEZ,
k=1

where a > 0 and the ak are all real. This system is linear and time
invariant and hence has a Kronecker delta response h with DTFT
H. Assume that h is a causal filter, that is, that h n = 0 for n < O.
(Physically, the difference equations cannot produce an output before
time 0 if the input is 0 for all negative time.) Assume also that the
ak are such that h is absolutely summable.

(a) Find an expression for H(f) in terms of a and

M
A(f) =1+L ake- i2 11"/k
k=1
6.12. PROBLEMS 303

(A filter of this form is called an autoregressive filter.)

Observe that if ak is extended to all integer k by defining ao =1

and ak = 0 if k is not in {O, 1, ... , M}, then

L
00

AU)= ake-i21r/k,
k=-oo

and AU) can be considered as the DTFT of the time limited

signal a = {an; n E Z}.
(b) Is h real?
(c) Define the signal w = h * a, i.e.,
M
Wn = L hn-k a k , (6.60)
k=O

where ao = 1. Find a simple expression for Wn .

(d) What is ho in terms of a and the ak?
(e) Suppose that a Kronecker delta is the input to h. Show that the
autocorrelation function of the output signal satisfies

L:
00

rh(l - k) = hn-khn-l; 1, k E Z. (6.61)

n=-oo

(Recall that the autocorrelation function of a signal is given by

rh(k) = En h~hn+k')
(f) Multiply both sides of (6.60) by hn-l and sum over n to obtain
a simple expression for
M
/31 = L: rh(k -l)ak.
k=O

Specialize this expression to the two cases 1 = 0 and 1 > O.

(g) Now suppose that we observe a discrete time signal Sn from an
unknown source. We measure its autocorrelation r s (k) for M + 1
values k = 0,1, ... , M. We can form a model ofthis signal called
the correlation matching autoregressive model by asking for what
autoregressive filter h will it be true that

rh(k) = rs(k); k = 0,1, ... , M? (6.62)

304 CHAPTER 6. CONVOLUTION AND CORRELATION

This filter h will have the property that when a Kronecker delta
is input, the autocorrelation of the output will equal ("match")
the measured autocorrelation for the first M + 1 values.
Use the previous parts of this problem to provide a set of equa-
tions in terms of a, the ak, and r 8 that will satisfy (6.62).

Note: What you have done above is to derive the Yule-Walker or

Normal equations that are at the basis of LPC (linear predictive cod-
ing) of speech [20, 1, 22]. Intuitively, forcing the autocorrelation of
the model to match the measured short-term autocorrelation of the
genuine speech will result in the model producing a sound similar to
the original speech.
These equations can be rapidly solved numerically using an algorithm
called the Levinson-Durbin algorithm. These methods are also useful
in a variety of other signal processing problems not related to speech.
6.41. Consider the signals 9 = {g(t) = ).e-,Xtu_1(t); t E R} and h(t)
{D 1 / 2 (t); t E R}. ). is a positive real parameter.
Evaluate the following moments:

(a) J~oo tkg(t) dt for k = 0,1.

(b) J~oo tkh(t) dt for k = 0, l.
(c) J~oo P H(f) dl, where H is the Fourier transform of h.
(d) Define y as 9 * h. Find an expression for y(t) for all t.
(e) Evaluate
I: y(t) dt,

where y is as in the previous part.

6.42. Suppose that 9 = {gn; n E Z} is a discrete time signal with Fourier

transform G = {GU); IE [-1/2, 1/2)} defined by

G(f)
°
= {1 iii::; 1(6
otherWIse
.

(a) Suppose that a system with input 9 has output y defined by

Yn = gn cos(27r~n); n E Z. Find Y(f) and provide a labeled
sketch.
(b) Is the system described in the previous part linear? Time-
invariant?
6.12. PROBLEMS 305

(c) Suppose that a system with input 9 has output w defined by

Wn = g3n; n E Z. Find W(f) and provide a labeled sketch.
(d) Is the system described in the previous part linear? Time-
invariant?
(e) Suppose that a system with input 9 has output z defined by
Zn = g;; n E Z. Find Z(f) and provide a labeled sketch.

(f) Is the system described in the previous part linear? Time-

invariant?
(g) What is the Kronecker delta response to each of the above three
systems?
6.43. A continuous time infinite duration signal 9 is bandlimited to [-.5, .5],
i.e., its Fourier transform G satisfies G(f) = 0 for f ~ .5. The
signal is passed through a filter with impulse response h, which is
also bandlimited to [-.5, .5]. The output of the system, y, is sampled
at a rate of 1 sample per second.
(a) Using the sampling theorem, express get) and h(t) in terms of
their respective samples if they are both sampled at 1 sample
per second.
(b) Express the sampled values of the filter output yet) in terms of
the sampled values of 9 and h. (Be as rigorous as possible.)
Your solution should look familiar.
(c) Let {GU); f E [-1/2, 1/2)} denote the Fourier transform of the
discrete time signal {g(n); n E Z}. Express G in terms of G.
6.44. Suppose that a continuous time signal get) is sampled to form a dis-
crete time signal 'Yn = genTs). Consider also the idealized sampled
waveform get) = g(t)ill(t/Ts), that is, the continuous time waveform
with the samples imbedded on impulses. Let ref)denote the periodic
extension of the DTFT of 'Y and let G denote the CTFT of g. Is it
true that
(6.63)
6.45. Given a finite duration signal, show that the generalized autocorre-
lation function of its periodic extension is just the periodic extension
of the autocorrelation function of the finite duration signal.
6.46. Calculate the following integrals.
(a)
j ! [Sin(211"~(N
_1
2
+ 1/2))]
sm(1I"t)
2 dt
306 CHAPTER 6. CONVOLUTION AND CORRELATION

(b)
ro 2
Loo 1 + (27rt)2 sinc(t) dt
(c)
[00 2
J- oo [sinc(t) + e- 1tl sgn(t)] dt

(d)

for real t.
(e)
j! 1
a.
e-·2", ei21r4 / dlf ,
-! 1 +-a-
where lal > 1.
6.47. Does the signal sinc3 (t) satisfy the conditions for the central limit
theorem?
6.48. For the function g(t) = t\(t) cos(7rt)j tEn, find
(a) The Fourier transform of g.
(b) The equivalent width of g.
(c) The autocorrelation width of g.

6.49. For the function g(t) = sinc2 (t) cos(7rt)j tEn, find

(a) The Fourier transform of g.

(b) The equivalent width of g.
(c) The autocorrelation width of g.

6.50. Find the equivalent width and the autocorrelation width of g(t) =
n(t) + o(t - 2).

6.51. Prove (4.61)

6.52. For the function g(t) = Jo(27rt),
(a) Find the equivalent width and the autocorrelation width.
(b) Does g(t) obey the central limit theorem? If yes, what is an
approximate expression for g(t)*n for large n? If no, why not?
6.12. PROBLEMS 307

6.53. Find the Fourier transform of the get) = J 1 (27rt)/(2t).

(a) Find the equivalent width and the autocorrelation width.
(b) Does get) obey the central limit theorem? If yes, what is an
approximate expression for g(t)*n for large n? If no, why not?
6.54. Given the signal get) = 2e- 21t1 j t E 1<-,

(a) Find its area.

(b) Find its second moment.
(c) Find its equivalent width.
(d) Find its autocorrelation width.
(e) Does the central limit theorem apply to this signal? If so, find
an approximation for g(t)*20, the convolution of g(t) with itself
20 times.

6.55. Given a band-limited signal get) with spectrum G(J) = G(J) n (ztv),
define the semi-inverse signal get) as the inverse Fourier transform of
(l/G(J» n (iv). Evaluate 0';,0';,
and 0';*0
in terms of G(J) and its
derivatives.
Chapter 7

Two Dimensional Fourier

Analysis

The basic definitions for two-dimensional (2D) Fourier transforms were in-
troduced in Chapter 2 as Fourier transforms of signals with two arguments.
As in the one-dimensional case, a variety of two-dimensional signal types
are possible. In this chapter we focus on a particular signal type, and
demonstrate the natural extensions of many of the one-dimensional results
to two-dimensional Fourier transforms. We here consider briefly several ex-
tensions of previously described results to the two-dimensional case. More
extensive treatments of two-dimensional Fourier analysis to images may be
found in Goodman [18) and Bracewell [8).
For this chapter we consider two dimensional signals having continuous
"time" and infinite duration. The word "time" is in quotes because in most
two-dimensional applications the parameters correspond to space rather
than time. Thus a signal will have the form

9 = {g(x,y); x E n,y En}.

As we only consider this case in this chapter, we can safely abbreviate the
full notation to just g(x, y) when appropriate. Repeating the definition for
the two dimensional Fourier transform using this notation, we have

G(fx'/y) = [: [: g(x, y)e- i27r (fxx+fYY) dx dy (7.1)

310 CHAPTER 7. TWO DIMENSIONAL FOURIER ANALYSIS

7.1 Properties of 2-D Fourier Transforms

In general the properties of 2-D Fourier transforms are similar to corre-
sponding properties of the 1-D transform. In this section, several of these
properties are collected. They are stated without proof as the proofs are
straightforward extensions of the corresponding 1-D results.

1. Linearity
If gk(X, y) :J Gk(fx, fy) for k = 1 ... K, then
K K
LO!k9k(X,Y):J LO!kGk(fx,fy). (7.2)
k=l k=l

Thus, just as in the 1-D case, if we can decompose a complicated

function into a sum of simpler functions whose Fourier transforms we
know, we can easily find the Fourier transform of the more compli-
cated function by adding known transforms of the simpler functions.
2. Symmetry
If g(-x,-y) = g*(x,y), then G(fx,/y) is entirely real-valued. If
g(-x,-y) = -g*(x,y), then G(fx,fy) is entirely imaginary.

3. Inversion (Fourier Integral Theorem)

I: I:
If g(x,y):J G(fx,fy), then

g(x, y) = G(fx, fy )e i21f (fx Z +/YY) dfx dfy (7.3)

at points of continuity of g(x, y) if

(a)
I: I: Ig(x,y)ldxdy < 00
and
(b) g(x,y) is smooth, e.g., it is piecewise continuous or it has no
infinite discontinuities and only a finite number of maxima and
minima in any finite area.
At a point of discontinuity, we recover the angular average of g(x, y) at
that point as indicated in Fig. 7.1. This is best expressed in polar co-
ordinates, where the angular average can be written limr-to fo21f g(r, 0) dO
and r represents radius measured from the point of discontinuity.
7.1. PROPERTIES OF 2-D FOURIER TRANSFORMS 311

Figure 7.1: Values recovered at two dimensional discontinuities. The func-

tion is one in the shaded area and zero outside that area.

4. Stretch (Similarity)
If g(x,y) :::> G(fx,/y), then

( 1 /x /y (7.4)
9 ax,by):::> labIG(~'T)·

In the 2-D case, a stretch or a contraction of a function can be con-

sidered to be a magnification or a demagnification of that function.
5. Shift
If g(x,y) :::> G(fx,fy), then

g(x - a,y - b):::> G(fx,/y)e-i27f(fxa+IYb). (7.5)

Thus a shift in the (x, y) domain results in the introduction of a linear

phase function in the (fx, fy) domain.
312 CHAPTER 7. TWO DIMENSIONAL FOURIER ANALYSIS

6. Parseval's Theorem

£: £: 91(X,Y)9~(X,y)dxdy =
I: I: G1{/x,fy)G;{/x,fy)dfx dfy. (7.6)

When the functions 91 and 92 are both equal to the same function 9,
Parseval's theorem reduces to a statement that the energy of 9 can
be calculated as the volume lying under either 191 2or IGI2.

7. Convolution

:Flx,fy {I: i: 9«(,1])h(x - (,y -1])d(d1]} = G{/x,fy)H{/x,fy).

(7.7)
Two-dimensional convolution serves the same basic purpose as one-
dimensional convolution: it represents a general linear invariant fil-
tering of the input signal to produce an output signal. The operation
is depicted in Fig. 7.2.

7.2 Two Dimensional Linear Systems

Consider a 2-D system that maps an input {v(x,y)j x,y E 'R} into an
output {w(x,y)j x,y E'R}. If the system is linear, we can characterize
it by an impulse response h(x, Yj (, 1]), the response at output coordinates
(x,y) to an input 6-function at coordinates «(,1]).
As an aside, the unit impulse in two dimensions is represented by the
symbol c5(x,y). By analogy with the 1-D case, its critical property is the

i: i:
sifting property,

c5(x - (, y -1])p«(, 1]) d( d1] = p(x, y)

valid at all points of continuity of p{x, y). In some cases it may be convenient
to consider the 2-D unit impulse to be separable in the (x, y) coordinate
system, leading to the representation

6(x, y) = 6(x)6(y).
7.2. TWO DIMENSIONAL LINEAR SYSTEMS 313

y y

h(x,y)

x x

h(-x,-y)

Figure 7.2: 2-D convolution

However, other forms of the 2-D unit impulse are also possible. For example,
a circularly symmetric form could be defined through a limiting sequence
of circ functions, i.e.
8(x, y) = n-too
lim ~circ( v(nx)2 + (ny)2)
7r

where the factor 1/ 7r assures unit volume. Of course the above equation
should not be taken literally, for the limit does not exist at x = y = 0, but
equality of the left and right will hold if the limit is applied to integrals of
the right hand side.
In two dimensions, the unit impulse response of a linear system is often

i: i:
called the point-spread function. By direct analogy with the 1-D case:

w(x, y) = hex, y; (, 7})v((, 7}) d( d"1, (7.8)

a 2-D superposition integral. If the impulse response simply shifts with

movements of the impulse and does not change its shape, that is, if it is
314 CHAPTER 7. TWO DIMENSIONAL FOURIER ANALYSIS

space invariant, then

h(x, y; (,1]) = h(x - (, Y -1]) (7.9)

and
w(x, y) = i: i: h(x - (, Y -1])v«(, 1]) d( d1], (7.10)

a 2-D convolution integral. A space invariant 2-D filter is a shift invariant

system where the shift is in space instead of in time.
As in the I-D case, convolution is greatly simplified if we represent it as
its frequency domain equivalent. By the convolution theorem,

W(fx,fy) = H(fx,/y)V(fx,fy), (7.11)

where
H(fx,fy) = :F'x,/y(h)
is the 2-D transfer function.
Example
As an example of a linear 2-D system, consider the imaging system
of Fig. 7.3. The system maps an object into an out-of-focus image (the
distances z of the object plane and the image plane from the imaging lens
do not satisfy the lens law). As a consequence, the image of the object is
blurred. Our goal is to represent this blur operation as a linear, invariant
system.

Lens
Y

Y
~ Zo --~~I---- Zl - - -......
~I

Object Plane Image Plane

Figure 7.3: Defocused optical system

According to the laws of geometrical optics, for a point source object

and a perfect imaging lens, the optical rays diverging from the object will
 7.2. TWO DIMENSIONAL LINEAR SYSTEMS 315

be focused by the lens to a perfect geometrical focus in the correct image

plane, i.e. in the plane where the lens law
1 1 1
-+-=-
Zo Zi f
is satisfied, where Zo and Zi are the distances of the object and the image
planes from the lens, respectively, and f is the focal length of the lens.
In the case of current interest, the image plane is slightly too far from
the lens, with the result that the rays pass through perfect focus in front of
the image plane, and diverge from that focus to form an illuminated region
that has the same shape as the aperture of the lens used for imaging. For
a circular lens, the system has the impulse response pictured in Fig. 7.4.
The impulse response is uniformly bright (or of uniform intensity) over a
c5(x,y)

x
x

Object Image

Figure 7.4: System impulse response

circle of radius €, where the value of € depends on geometrical factors and,

in particular, the degree of defocus (a larger defocus implies a larger f).
We represent the imaging system acting on the point source to produce a
circular defocused image through the equation

C{c5(x,y)} = kcirc(!:) = h(r).

€

Image inversion can be shown to destroy space invariance according to

the strict definition (you prove it). The same is true for image magnification
or demagnification. However, the system can be made space invariant by
redefining the image coordinate axes, inverting their directions and normal-
izing their scale to assure a magnification of unity. In this current case, the
object and image planes are assumed to be equal distances from the lens,
so the magnification of the system happens to be unity. (The magnification
316 CHAPTER 7. TWO DIMENSIONAL FOURIER ANALYSIS

can be shown to be the ratio of the distance of the image plane from the
lens to the distance of the object plane from the lens.) We assume that
proper normalizations of the object coordinate system have been carried
out to make the system space invariant.
To find the transfer function H(p) = Fp{h(r)} recall that

. J 1 (27rp)
c1rc(r) ~ . (7.12)
p

The scaling theorem for the zero-order Hankel transform can be stated

(7.13)

and therefore
(7.14)

Thus
(7.15)

where k' is a constant.

It is convenient to normalize the transfer function to have value unity
at the origin:
11.( ) = H(p) = 2lt (2m:p) . (7.16)
p H(O) 27rt:p
Note the following properties of this transfer function:

1. The first zero occurs at Po = 0.610/f. The larger f (Le. the larger the
defocus), the smaller Po and hence the smaller the bandwidth of the
system.

2. A negative transfer function at a particular frequency can be thought

of as a 180 0 phase reversal of sines or cosines at that frequency, or
equivalently a contrast reversal of a spatial frequency component.

3. If 1I.(p) = 11I.(p)le i q,(p), then an input of the form a + b cos(27r{fxx +

Jyy» yields an output of a+bI1l.(p)1 cos(27r {fx X + Jyy)+¢(p», where
¢(p) is the phase of the transfer function at radial frequency p. The
physical significance of this result is two-fold. First the contrast of the
cosinusoidal component of image intensity (defined as the difference
between the maximum and minimum intensities normalized by the
sum of the maximum and minimum intensities) is reduced. Equiva-
lently, the contrast can be expressed as the ratio of the peak intensity
7.3. RECONSTRUCTION FROM PROJECTIONS 317

to the average intensity. Thus

C = peak i~tensit~ = ~ -t ~11l(p)l.

average mtenslty a a
Thus the magnitude of the normalized transfer function represents the
amount by which the contrast of the sinusoidal component of intensity
is reduced. Note that at any zero of the magnitude of the transfer
function, the sinusoidal part of the intensity distribution completely
disappears, leaving only the constant background or average intensity.
Second, the phase of the cosinusoidal component of image intensity
is changed in proportion to the phase of the transfer function, which
means that the cosinusoidal intensity is shifted by some fraction <l>i:)
of its two-dimensional period.

7.3 Reconstruction from Projections

The field of radiology has undergone dramatic changes since the invention
and commercialization of computerized tomography as a measurement and
diagnostic technique. Using such techniques it is now possible to obtain
images of portions of the body that were not accessible by previously used
techniques, and to differentiate different types and constituents of tissue
with sensitivity and discrimination not previously possible. In this section
the principles of tomographic measurement and image reconstruction will
be discussed, not only because the subject is intrinsically important, but
also because the problem provides instructive connections between 1-D and
2-D Fourier analysis.
Figure 7.5 depicts a simplified model of data collection techniques used
in computerized tomography. For the purpose of detailed analysis of the
technique, consider an (x, y) coordinate system attached to the patient to
be scanned. The (x', y') coordinate system is attached to the scanning
geometry and rotates with the rotation of the measurement apparatus. For
future use observe that the relations between the two coordinate systems
are

x' = x cos (J + y sin (J (7.17)

y' = - x sin (J + Y cos (J (7.18)
x = x' cos () - y' sin () (7.19)
y = x' sin () + y' cos () (7.20)

The data is gathered with a combination of linear scanning and rotation.

To begin, the angle () is fixed, and the X-ray source and detector are scanned
318 CHAPTER 7. TWO DIMENSIONAL FOURIER ANALYSIS

y
X'

()

, x

"'d
A
o
I Detector

Top View

~X-Ray
U Detector
I
Side View

Figure 7.5: Computerized tomography data collection

linearly in the chosen direction in synchronism. The result of such a scan

is a measurement of the total attenuation observed by the X-ray source-
detector pair at each position along the linear scan. After each such scan,
the angle () is incremented and the scanning operation is repeated, until the
angle has been incremented through 7r radians or 180 degrees.
Let [0 represent the constant X-ray intensity emitted by the source.
Let [(x') represent the X-ray intensity falling on the detector when it is at
location x' along the scan. Let g(x, y) be the X-ray attenuation coefficient
of the head or body in a transverse slice (constant z). Pathology will be
indicated by unusual structure of g(x,y).
The measured X-ray intensity at scanner coordinates x' and (J will be

16 (X ') = .IOe -foo

T
-00
g(x'cos6-y'sin6,x'sin6+y'cos6)dy'
• (7.21)

The integral is required because absorption occurs along the entire y' path
7.3. RECONSTRUCTION FROM PROJECTIONS 319

of the X-ray beam.

After detection, the ratio

-In 19(X') (7.22)

10
is formed, yielding the measured projection of the X-ray attenuation coef-

I:
ficient

plJ(X') = g(x' cos e - y' sin e, x' sin e + y' cos e) dy'. (7.23)

To illustrate the concept of a projection, consider the projection of the

function circ(r) of Fig. 7.6 onto the x'-axis. Since circ(r) is circularly sym-
metric, projection onto the x'-axis is independent of e. The shaded plane

y'

Figure 7.6: Projection of circ(r)

cutting through the cylinder has area P9(X') for the specific x'. To find this
area we need the height of the rectangle, which is unity, and its width. To
find the width refer to Fig. 7.7. The result is:

po(x') = height x width = 1 x (2\1'1 - (x')2).

While the tomographic apparatus measures a series of projections of the

X-ray attenuation coefficient onto a series of rotated axes, it is not those
projections that are of ultimate interest, but rather the two-dimensional
320 CHAPTER 7. TWO DIMENSIONAL FOURIER ANALYSIS

Figure 7.7: Width of rectangle

distribution of X-ray absorption that gave rise to those projections. Thus

the mathematical challenge is to find a way of calculating the "image"
g(x, y) of the attenuation coefficient from the projection data that has been
collected. This problem is known as the inversion problem, and a solution
is of course critical to utility of the tomographic method.

7.4 The Inversion Problem

Given a set of measured projections PIi(X') (for all () E [0,11")) through
g(x,y), how do we recover g(x,y)? Alternatively, how do we reconstruct
the two dimensional signal g(x, y) from its projections? The answer is given
by the Projection-Slice theorem .(also known as the Central Slice theorem).
 7.5. EXAMPLES OF THE PROJECTION-SLICE THEOREM 321

Theorem 7.1 The 1-D Fourier transform Pe(f) of a projection Pe(x') is

identical to the 2-D Fourier transform G(fx, fy) of g(x, y) evaluated along
a slice through the origin at angle +0 to the f x axis:

Pe(f) = G(scosO,ssinO). (7.24)

Proof:

J
00

Pe(f) pe(x')e- i27rX ' f dx'

-00

This integral is to be carried out over the entire (x', y') plane. Equivalently,

I: I:
we can integrate over the entire (x,y) plane

Pe(f) = g(x,y)e-i 27r(xCose+ysine)8 dxdy.

Thus Pe(f) = G(scosO,ssinO). 0

Figure 7.8 illustrates the projection-slice theorem. The projection through
an attenuation function onto the axis at angle 0 to the x axis is shown, as
is also the slice in the frequency domain at angle 0 to the fx axis.
The projection slice theorem also works in reverse. A slice of g(x, y)
at angle 0 to the x axis in the space domain has a 1-D transform that is
the projection of G(fx'!y) onto an axis at angle 0 to the Ix axis in the
frequency domain.
In conclusion, the above analysis shows that if we can determine the
projections pe(x') for all 0 and x', then we can determine the spectrum
G(fx,fy) for all (fx,fy). With an inverse 2-D transform we can then
recover the image g(x, y) of interest.

7.5 Examples of the Projection-Slice Theo-

rem
Example 1. As a first example, consider a circularly symmetric function
g(x,y) = gR(r) with known radial profile gR(r). We wish to determine
gR(r) from a single projection pe(x') known to be
pe(x') = 2sinc(2x').
322 CHAPTER 7. TWO DIMENSIONAL FOURIER ANALYSIS

g(x,y)

pe(x')

Area A Value
x

(b)
Ge(f)

Figure 7.8: Illustration of the projection-slice theorem. (a) Calculation of

the projection at one point for the attenuation function g(x, y)j (b) The
complete projection through g(x,y) at angle OJ (c) the Fourier transform of
the proection, which is a central slice through the 2-D transform of g(x, y).

Find the radial profile gR(r).

Solution: We outline the various steps in the solution as follows:
1. Pe(f) = .1'f{2 sinc(2x')} = n(f /2).
2. From the projection slice theorem, Pe(f) is a slice through a circularly
symmetric 2-D spectrum. Clearly, then,
G(p) = circp.

3. The inverse zero-order Hankel transform can then be taken to yield

the original function,
gR(r) = J 1 (27rr).
r
Note that a single projection at any angle suffices for recovery of any
circularly symmetric function.

Example 2. Show that any function g(x,y) = gx(x)gy(y) separable in

rectangular coordinates can be reconstructed from two projections.
7.5. EXAMPLES OF THE PROJECTION-SLICE THEOREM 323

Solution:
1. Any function that is separable in the (x, y) domain has a Fourier
transform that is separable in the (fx, fy) domain. Therefore
g(x,y) = gx(x)gy(y) ::> Gx(fx)Gy(fy).

i: i:
2. First find the projection through g(x, y) for () = 0:
Po(x) = g(x,y)dy = gx(x) gy(y)dy.

In addition, Fourier transform the resulting projection using frequency

variable fx
Po(x) ::> Gx(fx)Gy(O).
Thus the I-D transform of this first projection yields information re-
garding the fx dependence of the spectrum, up to an as yet unknown
multiplier Gy(O).

i: i:
3. Next find the projection through g(x, y) at angle () = 7r /2. We obtain

Pf(x') = g(-y',x')dy' = gx(-y')dy'gy(x').

A I-D transform of this projection using frequency variable fy yields

Thus the I-D transform of this projection with respect to variable x'
yields information about the fy dependence of the spectrum, up to
an unknown multiplier Gx(O).
4. Now the undefined multipliers must be found. Note that the value
of the I-D spectrum of either projection at zero frequency yields the
product of these two constants,

Po(O) = Pf (0) = Gx(O)Gy(O).

5. Divide out the unknown multipliers by taking the ratio

Po(fx)Pi(fy) = Gx(fx)Gy(O)Gx(O)Gy(fy) = G (I )G (I )
Po (0) Gx(O)Gy(O) x x y y.

6. Finally, inverse transform to obtain

Gx(fx)Gy(fy) c gx(x)gy(y) = g(x,y).
324 CHAPTER 7. TWO DIMENSIONAL FOURIER ANALYSIS

Example 3. Find an expression for any projection through the function

Solution:
It can be shown (see, e.g., Bracewell, p. 249)

1
7raJo(27rar) :J "20(p - a) = G(p).
Therefore
1 1
Pe(f) = G(f) = "20(8 - a) + '20(8 + a)
for any () and hence
PI! (x') = cos 27rax' .

7.6 Reconstruction by Convolution and Back-

projection
We can summarize the straightforward approach to reconstructing a signal
based on the projection-slice theorem as follows:

1. Collect projections at a discrete set of closely spaced angles.

2. Fourier transform the projections, obtaining samples of G(fx, Jy) on

the grid shown in Fig. 7.9.
3. Interpolate these samples to a rectangular grid.
4. Apply the inverse 2-D Fourier transform.

The straightforward approach has two serious problems:

1. Interpolation is computationally expensive.

2. We cannot begin the final inverse FFT until all data has been gathered
and transformed.
This leads to an alternative approach, developed by Bracewell in the
1950s, called convolution and backprojection. The goal is to reconstruct
g(x,y) from the projections Pe(x' ) taken at all () between 0 and 7r. We will
neglect the fact that () is sampled discretely.

i:i:
The starting point is the definition of the inverse 2-D Fourier transform,

g(x,y) = G(fx,fy)e i27r (fxx+/YY)d/x d/y.

7.6. RECONSTRUCTION 325

fx

Figure 7.9: Sampling grid for G(lx, fy) in Reconstruction

Convert from (Ix, fy) coordinates to (s, B), where s scans 'R and Bruns
from 0 to 11':

fx = scosB
fy = ssinB
dfx dfy = lsi ds dB.
Then by straightforward substitution,

g(x, y) =
o
1 100 dslslG(s cos B, s sin B)e
211"

-00
dB i2 11"S(z cos 9+ysin e);

g(x, y) = 1 de 100 dslsIPe(s)e

hence
211" i2 71"B(z cos 9+ysin 8).
o -00
This result can be rewritten in the following form:

10r11" dBfe(xcosB+ysinB)
2
g(x,y) = (7.25)

where
(7.26)
326 CHAPTER 7. TWO DIMENSIONAL FOURIER ANALYSIS

The function le(x') is called the filtered backprojection of the projection

pe(x'). To explain the name and interpret the result, suppose that we obtain
a projection pe(x') at angle 0; then the backprojection of pe(x') is the 2-D
function pe(x cos 0+ y sin 0). The backprojection can be obtained by taking
the 1-D function Pe(x') and "spreading" it uniformly across the 2-D plane,
with the spreading operation being titled at angle 0 with respect to the x'
axis, as shown in Figure 7.10 for the specific case of a backprojection angle
that is 22.5 degrees counterclockwise from the y axis ..

Figure 7.10: Illustration of a backprojection. The function being backpro-

jected is a 2D cosine oriented at angle 22.5° to the x axis.

We can not reconstruct g(x, y) simply by adding up backprojections;

that is,

g(x,y) "I r
Jo
21T
dOpo(xcosO+ysin(}).

However we can reconstruct g(x,y) if we filter each projection first, before

backprojecting. The filter required is one dimensional and has a transfer
function lsi- This transfer function compensates for the fact that simple
(unfiltered) backprojection gives too heavy a weighting near the origin in
the (fx,fy) plane.
 7.7. * TWO-DIMENSIONAL SAMPLING THEORY 327

Because the 2-D object spectrum is approximately bandlimited, so too

are the Fourier transforms of the projections. Therefore the transfer func-
tion H(I) = lsi need not rise indefinitely. A transfer function

H(I) = lsi n (2~J

will do, where Ie is chosen just large enough so that all significant Fourier
components of the spectrum will be included. Recognizing that the result-
ing transfer function is the difference between a rectangle function of height
and width Ie and a triangle function of height and width Ie, the impulse
response is found to be

The procedure used in practice would be

1. Collect a projection at angle e.
2. While collecting it, convolve with h(x) in real time.
3. As soon as convolution is finished, backproject. Accumulate.
4. Repeat for all 0 ~ e ~ 7r.
A procedure based on backprojection and convolution was used in early
computerized tomography X-ray scanners. However, modern X-ray tomog-
raphy machines typically have more complex data gathering apparatus than
described here, often gathering several different measurements simultane-
ously to increase the scanning speed. More complex reconstruction algo-
rithms have been developed for use with such machines.

7.7 * Two-Dimensional Sampling Theory

Just as a bandlimited 1-D function of a single independent variable can be
reconstructed from samples taken at a uniform rate that exceeds a criti-
cal lower bound, so too a 2-D function that is bandlimited in the (Ix, fy)
plane can be perfectly reconstructed from samples that are uniformly dense
provided that density exceeds a certain lower bound. However, the theory
of 2-D sampling is richer than the theory of 1-D sampling, due to the ex-
tra degree of freedom afforded by the second dimension. Two dimensional
sampling theory provides interesting examples of the similarities and differ-
ences between 1-D theory and 2-D theory, and therefore is briefly discussed
here.
328 CHAPTER 7. TWO DIMENSIONAL FOURIER ANALYSIS

To embark on a discussion of 2-D sampling, we first need a clear def-

inition of a bandlimited function in two dimensions. A function g(x, y) is
called bandlimited if its 2-D Fourier spectrum G(fx, fy) is identically zero
for all radial frequencies P greater than some finite limit Pc. There is consid-
erable flexibility as to the exact shape of the region on which the spectrum
is non-zero, as indicated in Figure 7.11. The only requirement is that it
vanish eventually at large enough radius in the frequency domain.

II =Frequency-domain area occupied by spectrum

o =Minimum circular region within which the spectrum is non-zero
Figure 7.11: Various types of bandlimiting in two dimensions.

The geometrical pattern in which the samples are taken (Le. the sam-
pling grid) is also quite flexible. Most common is the use of a rectangular
sampling function defined by

L L
00 00

s(x,y) =ill(;)ill(~) =XY 8(x-nX)8(y-mY) (7.27)

n=-oom=-oo

and illustrated in Figure 7.12. However, other sampling grids are also
possible. For example, the sampling function

s(x,y) =ill ( V2Ll

x+y)
ill (x-y)
V2L2 (7.28)

can be thought of as resulting from a rotation of the (x, y) axes with respect
to the delta functions by 45 degrees, thus preserving the volume under the
delta functions but changing their locations. The arguments of the two ill
functions are simultaneously zero at the new locations

(7.29)

where nand m run over the integers. The above locations are the new
locations of the delta functions in this modified sampling grid.
7.7. * TWO-DIMENSIONAL SAMPLING THEORY 329

• • • •
• • • •
x

• • • •
• • • •
Figure 7.12: Rectangular sampling grid. 8 functions are located at each
dot.

When a sampling function is multiplied by a signal g(x,y), the product

operation in the time domain results in a convolution operation in the
frequency domain. Thus the Fourier transform of the sampled function
becomes
G(jx, fy) = S(jx, fy) * G(jx, fy).
For the sampling function defined on the original rectangular grid, the
spectrum of the sampled data becomes
00 00

= '~
" " 8(jx - X)
'~ y) * G(jx,fy)
n 8(jy - m
n=-oom=-oo

L L
00 00

C(lx-;,fy-;). (7.30)
n=-oom=-oo

As in the 1-D case, the effect of the sampling operation is to replicate the
signal spectrum an infinite number of times in the frequency domain, as
shown in Figure 7.13.
In order for the zero-order (n = 0, m = 0) term in the frequency domain
to be recoverable, it is necessary that there be no overlap of the various
spectral terms. This will be assured when the sampling intervals X and
Yare chosen sufficiently small. For the original rectangular sampling grid,
if all frequency components of g(x,y) lie within a rectangle of dimensions
2Bx x 2By in the Ix and Iy directions, respectively, centered on the origin
in the frequency domain, then sampling intervals X = 2~x and Y = 2~Y or
330 CHAPTER 7. TWO DIMENSIONAL FOURIER ANALYSIS

Iy

Ix
(a)

Ix

(b)

Figure 7.13: Replication in the frequency domain caused by sampling in

the space domain. (a) Original spectrum, (b) Spectrum after sampling

smaller will suffice. Under such conditions, the spectral islands will separate
and the n = 0, m = 0 island will be recoverable by proper filtering.
To recover g(x, y), the sampled data g(x, y) is passed through a 2-D lin-
ear invariant filter which passes the (n = 0, m = 0) spectral island without
change, and completely eliminates all other islands. For this sampling grid,
a suitable linear filter is one with transfer function

H(fx,fy) = n (/;x) n (/;y) .

Thus the identities

G(fx,fy) = n (/;
x
) n (2;) f f
y n=-oo m=-oo
G(fx -nBx,fy -mBy)

and the space domain equivalent

g(x,y) = g(x,y) * 4BxBy sinc(2Bxx)sinc(2Byy) (7.31)
 7.7. * TWO-DIMENSIONAL SAMPLING THEORY 331

f f
n=-oo m=-oo
9 (2; , 2; ) sine
X Y
(x - 2; X
) sine (y - 2; Y
)

hold provided X ::; 2~x' Y ::; 2~Y' This equation is the 2-D equivalent of
(4.15) which was valid in one dimension. Note that at least 4Bx By samples
per unit area are required for reconstruction of the original function.
This discussion would not be complete without some mention of the
wide variety of other possible sampling theorems that can be derived in the
2-D case. For example, suppose that the reconstruction or interpolation
filter were chosen to be circularly symmetric, rather than separable in the
(x, y) coordinate system. A filter with transfer function

H(fx,fy) = H(p) = circ(p/pc)

could be applied to the sampled data, with the cutoff frequency Pc chosen to
equal the diagonal width of the rectangle used above, i.e. Pc = B'i + B~. J
This filter, like the one considered earlier, will completely isolate the zero-
order spectral island. Noting that

and assuming that the intervals between samples satisfy X = Y = -2 1 , it is

Pc
a simple matter to show that the original function g(x,y) can be recovered
with another interpolation function, i.e.

00
'"' '"'
00
n m
( )[J1(27rPc\!(X - ..!!-)2 + (y -
2
m2)]
9 (x,y ) = ~ ~ 9 -,- 4"7r Pc Pc
2
n=-oo m=-oo Pc Pc 27rPcV(x - ..!!-)2
Pc
+ (y - m
Pc
)
(7.32)
where the factor 7r / 4 is the ratio of the unit rectangular cell in the 2-
D frequency domain replication grid (4p~) to the frequency domain area
covered by the transfer function of the interpolation filter (7r p~). Thus it is
clear that even for a fixed choice of sampling grid there are many possible
forms for the sampling theorem.
If the sampling grid is allowed to change, then additional forms of the
sampling theorem can be found. For example, an hexagonal sampling grid
plays an important role in the sampling of functions with circularly band-
limited spectra. Such a grid provides the densest possible packing of circular
regions in the frequency domain.
We will not pursue the subject of sampling further here. The purpose
of the discussion has been to point out the extra richness of the theory that
occurs when the dimensionality of the signals is raised from 1 to 2.
332 CHAPTER 7. TWO DIMENSIONAL FOURIER ANALYSIS

7.8 Problems
7.1. Find the following convolutions.
(a) g(r) = Jl(;1rr) * ~JlS1rr).
(b) g(x,y) = n(x) n (y) * n(x) n (y).
7.2. Evaluate the integral

7.3. Find the 2-D Fourier transforms of

(a) 8(r - ro) (ro a constant).

(b) e _1rr 2 (be clever).
(c) n(r - 1).
(d) cos(21Tax) cos(21Tby).
(e) cos (21T(ax + by)).
7.4. (a) Find an expression for any projection through the 2-D function

g(r) = 1TaJo(21Tar).
(b) A projection through a 2-D circularly symmetric function is
found to be
p(x) = e- 1raz •
2

Specify the two dimensional function through which the projec-

tion was taken.
7.5. The projection through a certain circularly symmetric function is
known to be p(x') = cos 21TX' sinc (x'). Specify the original circularly
symmetric function.
7.6. Find an expression for any projection through the circularly symmet-
ric function gR(r) = (1 + r2)-!.
Chapter 8

Memoryless Nonlinearities

We have seen that Fourier analysis is a powerful tool for describing and
analyzing linear systems. A particularly important application has been
that of sampling continuous time signals to produce discrete time signals
and the quantifying of the conditions under which no information is lost
by sampling. The purpose of this chapter is twofold. First, we demon-
strate that Fourier analysis can also be a useful tool for analyzing simple
nonlinear systems. The techniques used in this chapter are a relatively mi-
nor variation of techniques already seen, but this particular application of
Fourier theory is often overlooked in the engineering literature. Although a
standard component of courses is devoted to nonlinear systems, the relative
scarcity of such courses and the lack of examples in engineering transform
texts has led to a common belief of near mythological nature that Fourier
methods are useful only in linear systems. Using an idea originally due to
Rice [28) and popularized by Davenport and Root [15) as the "transform
method," we show how the behavior of memoryless nonlinear systems can
be studied by applying the Fourier transform to the nonlinearity rather
than to the signals themselves.
The second goal of the chapter is to consider the second step in the
conversion of continuous signals to digital signals: quantization. A uniform
quantizer provides an excellent and important example of a memoryless
nonlinearity and it plays a fundamental role at the interface of analog and
digital signal processing. Just as sampling "discretizes" time, quantization
converts a continuous amplitude signal into a discrete amplitude signal.
The combination of sampling and quantization produces a signal that is
discrete in both time and amplitude, that is, a digital signal. A benefit of
focusing on this example is that we can use the tools of Fourier analysis
to consider the accuracy of popular models of and approximations for the
334 CHAPTER 8. MEMORYLESS NONLINEARITIES

behavior of quantization error, the distortion that occurs in a signal when

it is mapped from an analog amplitude into a discrete approximation.

8.1 Memoryless Nonlinearities

A memory less system is one which maps an input signal v = {v(t)j t E T}
into an output signal w = {w(t)j t E T} via a mapping of the form

w(t) = O:t(v(t))j t E T
so that the output at time t depends only on the current input and not
on any past or future inputs (or outputs). We will emphasize real-valued
nonlinearities, i.e., O:t(v) E n for all v E n. When O:t does not depend on
t, the memoryless nonlinearity is said to be time invariant and we drop the
subscript.
Let A c n denote the range of possible values for the input signal v(t),
that is, v(t) E A for all tEn. Depending on the system, A could be the
entire real line n or only some finite length interval of the real line, e.g.,
[- V, V]. We assume that A is a continuous set since our principal interest
will be an O:t that quantizes the input, i.e., that maps a continuous input
into a discrete output.
The function O:t maps A into the real linej that is, 0: can be thought
of as being a signal O:t = {o:(v)j v E A}. This simple observation is the
fundamental idea needed. Since O:t is a signal, it has a Fourier transform

At(f) = L O:t(x)e- i27rx / dx.

Keep in mind that here t can be considered as a fixed parameter. If the

nonlinearity is time-invariant, the t disappears.
Assuming that the usual technical conditions are met, the Fourier trans-
form can be inverted to recover 0:, at least at its points of continuity. First
suppose that A = n, the case where the range has "infinite duration"

I:
(infinite length is better terminology here). In this case we have that

O:t(x) = At(f)ei27rX/ dfj x E n. (8.1)

If, on the other hand, A has finite length, say L (we don't use T since the
parameter no longer corresponds to time), then the inversion is a Fourier
series instead of a Fourier integral:

~ At(!?c
L ) ·2 n
O:t(X) = ~ -L-e' 7rTXj x EA. (8.2)
n=-oo
8.2. SINUSOIDAL INPUTS 335

This representation is also useful if A = n, but the nonlinearity at is a

periodic function of its argument with period L.
Thus we can represent the memory less nonlinearity as a weighted linear
combination of exponentials. Why is this of interest? Observe that we
can now write the output of a memory less nonlinearity as follows: Since
w(t) = at(v(t)),
if A =n
if A has finite width L or at (u)
is periodic in u with period L.
(8.3)
Thus we have what resembles a Fourier integral or Fourier series repre-
sentation of the output in terms of the input, even though the system is
nonlinear! It is, of course, not an ordinary Fourier integral or series because
of the appearance of the input signal in the exponents. In fact, one can
view the complex exponential terms containing the input as a form of phase
modulation or PM of the input. Nonetheless, this expansion of the output
can be useful in analyzing the behavior of the system, as shall be seen. This
method of analyzing is sometimes called the transform method because of
its reliance on the Fourier transform.
This is about as far as we can go in general without narrowing the devel-
opment down to a particular memoryless nonlinearity or input signal. First,
however, we observe a basic fact about combining sampling and memoryless
operations. Suppose that our system first takes a continuous time signal
u(t), samples it to form a discrete time signal Un = u(nTs), and then passes
the sampled signal through a discrete time memoryless time-invariant non-
linearity a to form the final signal Wn = a(u(nTs )). Now suppose that
instead the order of the two operations is reversed and we first pass the
continuous time signal u(t) through a continuous time nonlinearity a (the
same functional form as befote) to form a(u(t)) and then we sample to
form the final output Wn = a(u(nTs)). Regardless of the order, the out-
puts are the same. For this reason we often have our choice of considering
continuous or discrete time when analyzing nonlinearities.

8.2 Sinusoidal Inputs

The transform method does not provide simple general results for Fourier
representations of output signals of memory less nonlinear systems in terms
of the Fourier representation of the input. Such generality and simplicity is
usually only possible for linear systems. A general result can be obtained
for memoryless nonlinearities operating on sinusoidal inputs. This result
is not as important as the corresponding result for linear systems because
336 CHAPTER 8. MEMORYLESS NONLINEARITIES

sinusoids are not the fundamental building blocks for nonlinear systems
that they are for linear systems. Nonetheless it is an important class of
"test signals" that are commonly used to describe the behavior of nonlinear
systems.
Suppose now that u(t) = asin(211"/ot) for t E T and suppose that the
Fourier series representation of the memory less nonlinearity is given. Then

wet) = f: AtlY)ei21Ttasin(21T/ot). (8.4)

n=-oo

The exponential term is itself now periodic in t and can be further expanded
in a Fourier series, which is exactly the Jacobi-Anger formula of (3.91):

L
00

ei21Ttasin(21Tfot) = Jk(211"~a)ei21Tfotk, (8.5)

k=-oo

where Jk is the kth order Bessel function of (1.12). Incorporating (8.5) into
(8.4) yields

wet)
00

L
n=-oo
AtlV f: Jk(211"~a)ei21Tfotk
k=-oo

f:
k=-oo
ei21Tfotk f: Jk(211"~a).
n=-oo
Atl£) (8.6)

If the nonlinearity is time-invariant and hence At does not depend on t,

then this result has the general form of a Fourier series

L
00

wet) = bkei21Tfotk (8.7)

k=-oo

with coefficients
(8.8)

An interesting technicality arises in (8.7). If T = n, then (8.7) is indeed a

Fourier series for a continuous time periodic function having period 1/ fo.
In the discrete time case of T = Z, however, we can only say that (8.7)
has the form of a Fourier series because the sum is not actually a Fourier
series because wen) is not a periodic function in general. It will be periodic
only if 211" 10 is a rational number. If 10 is irrational en is not periodic since,
for example, sin(21l"lon) is only 0 if n = o. Another difference with the
8.3. PHASE MODULATION 337

ordinary Fourier series is that the frequencies do not have the form n/N for
an integer period N (since there is no period). A series of this form behaves
in many ways like a Fourier series and is an example of a generalized Fourier
series or a Bohr-Fourier series. The signal en turns out to be an example of
an almost periodic function. See, e.g., Bohr for a thorough treatment [4].
In both the continuous and discrete time case, (8.7) immediately gives
the Fourier transform of the output signal as

L
00

W(f) = bk 8(f - kfo). (8.9)

k=-oo

Thus the memoryless nonlinearity produces an infinite collection of har-

monics of the original input signal, that is, an infinite sum of sinusoids with
frequencies equal to integer multiples of the fundamental frequency fo of
the original signal. Recall that a linear time invariant system could not
have any output frequencies that were not also present at the input, but
that a linear time varying system (such as amplitude modulation) could
(and does) produce such frequencies.
We have not been rigorous in the above derivation in that we have not
proved that the exchange of order of summation in (8.6) is valid. This would
have to be done for a specific nonlinearity in order to prove the result.

8.3 Phase Modulation

The simplest example of the previous derivation is that of phase modula-
tion. In Chapter 4 we considered amplitude modulation of an input signal
where the output of a system was formed by multiplying the input signal
by a sinusoid. In particular, an input signal 9 = {get); t E 'R} resulted
in an output signal ge(t) = ag(t)ei21r/ot. This complex signal provided
a useful means of handling the transform manipulations. In real life the
modulated waveform would be formed by taking the real part to form the
signal gc(t) = ~ge(t) = ag(t) cos(27rfot). In contrast to this linear system,
phase modulation is formed by modulating the carrier phase rather than
its amplitude:
gp(t) = acos(27rfc t + Ag(t)) = ~(aei(21r/ct+ag(t))).
This system is nonlinear, but memoryless. If the real part is ignored for
convenience, then the nonlinearity is complex. It is also a time varying
nonlinearity. To apply the results of the previous section, observe that
gp(t) is the real part of the signal yet) = ae i(21r IcHag(t)) = ae i21r Icteiag(t).
This is a memoryless nonlinearity with
Clt(U) = aei21r/cteiau
338 CHAPTER 8. MEMORYLESS NONLINEARITIES

which has the form of (8.2) with ~ = 27r / L

At(L) = {ei27r/ct n= 1
L 0 otherwise,
so that only one term in the Fourier series is needed. If the input signal is
taken as a sinusoid g(t) = asin(27rJot) we then have from (8.6) that

L
00

w(t) = ei27r/ct Jk(~a)ei27r/otk

k=-oo

L
00

= Jk(~a)ei27r(fc+klo)t
k=-oo

and hence the spectrum is given by

L
00

W(f) = Jk(~a)d(f - (fe + kJo)), (8.10)

k=-oo

a weighted sum of Dirac delta functions including the carrier frequency

and all sidebands formed by adding or subtracting integral multiples of the
input signal frequency to the carrier. This is a complicated spectrum, even
though the input signal is simple! A single frequency component at the
input results in an infinite collection of components at different frequencies
at the output. This is typical behavior for a nonlinear system.

8.4 Uniform Quantization

A uniform quantizer can be thought of as a form of rounding off. A specified
input range, say [-b, b] is divided up into M bins of equal size ~ so that

~ = 2b.
M
The number R = log2 M is called the rate or bit rate of the quantizer and is
measured in bits per sample. We usually assume for convenience that M is
a power of 2 and hence R is an integer. If the input falls within a bin, then
the corresponding quantizer output is the midpoint (or Euclidean centroid)
of the bin. If the input is outside of the bins, its representative value is the
midpoint of the closest bin. Thus a uniform quantizer is represented as a
"staircase" function. Fig. 8.1 provides an example of a uniform quantizer
M = 8 and hence R = 3 and b = 4~. The operation of the quantizer can
also be summarized as in Table 8.1.
8.4. UNIFORM QUANTIZATION 339

q(v)

4~

-4~ -3~ -2~

-2~

-3~

-4~

-5~

Figure 8.1: Uniform Quantizer

Since quantization introduces error, it is of interest to study the behavior

of the error. Note that unlike the sampling case, information has genuinely
been lost by quantization and there is in general no way to recover it. Hence
one must include the distortion generated by quantization into any noise
analysis of the system and any practical system must have enough bits (a
large enough M) to ensure acceptable quality. Typical digitized images and
pulse coded modulated speech have eight bits per sample. Computerized
tomography images have 12 bits per sample. High fidelity digital audio
typically has 14 bits.
Given an input v and its quantized value q(v), define the quantizer error
t(v) = q(v) - v. (8.11)

If the quantizer is applied to a discrete time signal {v n }, then the corre-

340 CHAPTER 8. MEMORYLESS NONLINEARITIES

If input is then output is

[3A,00} 7f:!
2

[2A,3A} 5f:!
2

[A,2A} 3f:!
2

[O,A} ~
2'
~
[-A,O} -2'

(-2A, -A) -3~

(-3A, -2A) -5~

(-00, -3.1.) -7~

Table 8.1: Uniform Quantizer, M = 8

sponding error sequence is

(8.12)

We write it in this form so that the quantizer output can be written as its
input plus an error term, viz.

(8.13)

This yields the so-called "additive noise model" of quantization error as

depicted in Fig. 8.2.
This is not really a "model" at all since it is simply an alternative
way of writing the equation defining the quantizer error. The modeling
comes in when one makes statistical assumptions about the error behaving
like random noise. Here, however, we will not consider such assumptions
but will focus on exactly derived properties. The quantizer error signal is
the output of primary interest here as it quantifies how well the quantizer
approximates the original signal. Thus in this section the memory less non-
linearity of primary interest is not the quantizer itself, but the quantizer
error function.
8.4. UNIFORM QUANTIZATION 341

vn --+-~ + l - - - - I - - -

uantizer

Figure 8.2: Additive noise model of a quantizer

We further simplify matters by focusing on normalized signals. In par-

ticular, define the normalized error

e(v) = €~) = q~) - ~.

The error function e(v) is plotted as a function of v in Fig. 8.3 for the case
of M = 8.
e(v)

v
-4~ -3~ -2~ -~ -1 ~ 2~ 3~ 4

Figure 8.3: Normalized Quantizer Error Function

Several important facts can be inferred from the picture. First, the error
function satisfies
le(v)1 S ~ if Ivl S 4~.
In other words, the normalized error cannot get too big provided the input
does not lie outside of the M quantizer bins. For general M the condition
becomes
le(v)1 S ~ if Ivl S ~ ~ = b,
342 CHAPTER 8. MEMORYLESS NONLINEARITIES

that is, provided the input does not exceed the nominal range b. When an
input falls outside this range [-b, b] we say that the quantizer is overloaded
or saturated. When a quantizer is not overloaded, the normalized quanti-
zation error magnitude cannot exceed 1/2. In this case the quantization
error is often called granular noise. If the quantizer is overloaded, the error
is called overload noise. We here consider only granular noise and assume
that the input range is indeed [-b, b] and hence e(v) E [-1/2,1/2]. If this
is not true in a real system, it is often forced to be true by clipping or lim-
iting the input signal to lie within the range of the quantizer. Sometimes
it is useful to model the quantizer as having an infinite number of levels, in
which case the no-overload region is the entire real line.
Next observe that e( u) is a periodic function of u for u E [-b, b] and that
its period is ~. In other words, the error is periodic within the no-overload
region. In fact it can be shown by direct substitution and some algebra
that
1 u
e(u) = 2" - ~ mod 1; u E [-b,b]. (8.14)

This periodic signal can then be expanded in a Fourier series as

00 1 i"u 00. U
e(u) = I: i2rrle~ = I:sm(2rrl ~). (8.15)
1=-00 1=1
1,,0

This series gives the error function except at the points of discontinuity.
This is the Fourier series of the time-invariant memoryless nonlinearity
a(u) = e(u) of interest in this section.
Now suppose that the input to the quantizer is a sampled signal Un =
u(nTs) as previously. We now have that the resulting normalized error
sequence en = e( un) is given by

(8.16)

This application of Fourier series to the error function in a uniform

quantizer was first developed by Clavier, Panter and Grieg in 1947 [11, 12]
and it was also suggested by Rice (as reported by Bennett in 1948 [2]).
Suppose next that the input is a sinusoid Un = asin(2rr Jon), where a :5 b
is chosen so as not to overload the quantizer. Making the identifications
L = ~ and
A( L) = {i2;n
n::j; 0
L O n =0
8.4. UNIFORM QUANTIZATION 343

we immediately have from (8.7-8.8) that the normalized quantizer error

signal is given by
L
00

en = bkei27r/onk (8.17)
k=-oo

and that its Fourier transform is therefore

L
00

E(f) = bkb(f - kfo), (8.18)

k=-oo

where here the frequency shift is modulo the frequency domain of definition,
and where the coefficients are

(8.19)

The formula for bk can be simplified since

and hence
k odd
(8.20)
k even
which yields

e
n
= ~
L
ein27r/o(2m-l) (~_I_J
L i27r1 2m-l
(27r1!!:.))
. ~ (8.21)
m=-oo 1#0

This formula has the form

L
00

en = eiAmnCm, (8.22)
m=-oo

where
Am = fo(2m - 1)
(taken mod 1 so that it is in [0,1)) and
A
=L
1
em i2 I J2m - 1 (27rl ~).
1#0 7r

As previously discussed, this is actually a Bohr-Fourier or generalized Fourier

series because en is not periodic.
344 CHAPTER 8. MEMORYLESS NONLINEARITIES

We thus have a (generalized) Fourier series for the error signal when
a sinusoid is input to a sampler and a uniform quantizer. The Fourier
transform can be defined in the same way as it was for periodic signals: it
is a sum of impulses at frequencies Am having area em. Note that although
a single frequency sinusoid is put into the system, an infinity of harmonics
is produced-a behavior not possible with a time invariant linear system.
As one might guess, this technique is not useful for all memoryless non-
linearities. It only works if the input/output mapping in fact has a Fourier
transform or a Fourier series. This can fail in quite ordinary cases, for ex-
ample a square-law device o:(x) = x 2 acting on an unbounded input cannot
be handled using ordinary Fourier transforms. Laplace transforms can play
a useful role for such cases. See, for example, Davenport and Root [15].

8.5 Problems
1. Frequency modulation (FM) of an input signal g = {get); tEn} is
defined by the signal

YFM(t) = cos (27rfct + ~FM lot g(r) dr) ; t > O.

Suppose that the input signal g is a sinusoid. Find the Fourier trans-
form ofYFM.
2. Prove Equation (8.14).
3. A hard limiter is a memoryless nonlinear device with input/output
relation
+a u(t) > 0
wet) = asgn(u(t)) = { 0 u(t) = o.
-a u(t) < 0
Find an exponential series for the output of a hard limiter when the
input is a sampled sinusoid.
4. Suppose that a uniform quantizer has an odd number of levels with
the middle level at the origin. Derive the series representation of the
output when the input is a sampled sinusoid.
5. An alternative form of a hard limiter which arises in imaging takes
the form
wet) = {a u(t) E [0, E] .
o otherwise
Find an exponential series for the output when the input is a sampled
sinusoid.
8.5. PROBLEMS 345

6. Suppose that an image 9 is transformed using a unitary transform

T into G = TG, which is quantized to form G, which is inverse
transformed to form g. Prove that the mean squared error in the two
domains is the same, that is, that

L L /gn,k - gn,k /2 = L L /Gn,k - Gn,k /2.

n k n k
Appendix A

Fourier Transform Tables

We here collect several of the Fourier transform pairs developed in the book,
including both ordinary and generalized forms. This provides a handy
summary and reference and makes explicit several results implicit in the
book. We also use the elementary properties of Fourier transforms to extend
some of the results.
We begin in Tables A.I and A.2 with several of the basic transforms
derived for the continuous time infinite duration case. Note that both the
Dirac delta o(x) and the Kronecker delta Ox appear in the tables. The
Kronecker delta is useful if the argument x is continuous or discrete for
representations of the form hex) = I(x)ox + g(x)(1- ox) which means that
h(O) = 1(0) and hex) = g(x) when x =f. O.
The transforms in Table A.2 are all obtained from transforms in Ta-
ble A.I by the duality property, that is, by reversing the roles of time and
frequency.
Several of the previous signals are time-limited (Le., are infinite duration
signals which are nonzero only in a finite interval) and hence have corre-
sponding finite duratiun signals. The Fourier transforms are the same for
any fixed real frequency I, but we have seen that the appropriate frequency
domain S is no longer the real line but only a discrete subset. Table A.3
provides some examples.
Table A.4 collects several discrete time infinite duration transforms.
Remember that for these results a difference or sum in the frequency domain
is interpreted modulo that domain.
Table A.5 collects some of the more common closed form DFTs.
Table A.6 collects several two-dimensional transforms.
 348 APPENDIX A. FOURIER TRANSFORM TABLES

9 :F(g)

{n(t)j t E R} {sinc(f)j fER}

{DT(t)j t E R} {2Tsinc(2Tf)j fER}
{e- t u_l(t)j t E R} { l+i1ri/ j fER}
{e-1t1j t E R} {1+(i1rf)2 j fER}
{tn (t - ~)jt E R} {2Ir
Uf + (ie-
~ + e-(21r/)2 1) (1 -
i2 ,,' i2 ", r)
Uf j
f ER}
{t\(t)j t E R} {sinc2 (f)j fER}
{e- 1rt2 jt E R} {e- 1r / 2 j fER}
{e+ i1rt2 j t E R} {f-:te- i21r / 2 j IE R}

{sgn(t)j t E R} t;}jl E R}
{U_l(t)jt E R} Ho(f) - 2;/(1- 8/)jl E R}
{oCt - to)j t E R} {e-i21rftoj fER}

{8(at + b) j t E R} {rare-i21r/!jl E R}
{O(t)jtER} {lj fER}
{o'(t)jt E J?} {i211}j fER}
{'l1 T (t) = L:~=-oo o(t - nT)j t E R} {,J.'l1 1/ T (f)j fER}
{m(t)j t E R} {m(f)j IE R}
H(8(t + !) + 8(t - !))j t E R} {COS(7r f)j I E R}
H(8(t + ~) - 8(t - !))j t E R} {i sin(7r f)j fER}
{sech(7rt)j t E R} {sech(7rf)jf E R}

{Jo(27rt)j t E R} { 1rv':-f2 fER, III < 1

o otherwise

Table A.l: Continuous Time, Infinite Duration

 349

g F(g)

{sinc(t)j t E 'R} {n(f)j ! E 'R}

{sinc2 (t)j t E 'R} {A(f)j! E 'R}

7rt' E 'R}
{l·t {-isgn(f)j! E 'R}
{e- i27r !otj t E 'R} {8(f + !o)j! E 'R}
{ljt E 'R} {&(f)j! E 'R}
{COS(1Tt)jt E 'R} {II(f)j! E 'R}
{sin(1Tt)jt E 'R} {iII(f)j! E 'R}
H«S(t) + 2~t (1 - «St)j t E 'R} {U-l(f)j! E 'R}

Table A.2: Continuous Time, Infinite Duration (Duals)

350 APPENDIX A. FOURIER TRANSFORM TABLES

g F(g)

{n(t)j t E [-f, f)}, T ~ I {sinc(f)jf E {~j k E Z}}

{A(t)j t E [-t, t)}, T ~ 2 {sinc 2(f)j f E {~j k E Z}}

{6(t)jt E [-t, t)} {Ij f E {~j k E Z}}

{II(t)j t E [-t, t)}, T > I {cos(7rf)j f E {~j k E Z}}

{II(t)j t E [-t, t)}, T > I {isin(7rf)j f E {~j k E Z}}

{tj t E [0, I)} {%J k + 2;k(l- Jk)j k E Z}

Table A.3: Continuous Time, Finite Duration

 351

9 F(g)

{r nU_I(n); n E Z} (Irl < 1) { 1

I-re-. •• [ 1 I)}
2,,/, J E -2' 2

{ON(n); n E Z} {Sin(27r/(N+!)). f E [_1 1)}

sin(7rJ)' 2' 2

l_e;2,,/ . 1 1
{sgn(n); n E Z} f
{ I-cos 27r / ' E [- 2"' 2")}
I
{8 n- no ;n E Z} {e-i27r/no; f E [-~,~)}

{8 n ; n E Z} {I; f E [-~,~)}

{e- i27r *n; n E Z} {8(f + :kr)j f E [-~,~)}

Table A.4: Discrete Time, Infinite Duration

9 F(g)

{rn; n E ZN} f l-reI-r;2"k/N;

N k E ZN }
{I; n E ZN} {N8k; k E ZN}
{8 n ; n E ZN} {I; k E ZN}
{8 n- k ; n E ZN} {e- i27r *; k E ZN}
Table A.5: Discrete Time, Finite Duration (DFT)
352 APPENDIX A. FOURIER TRANSFORM TABLES

Function 'fransform

{exp[-1I'(X2 + y2)]; X,y E R} {exp [-11' (11 + If)] j lx, fy E R}

{n(x) n (y); x, y E R} sinc(fx) sinc(fy); fx,fy E R}
{A(x) A(y); x, y E R} {sinc 2(fx) sinc2(fy)j lx, fy E R}
{t5(x,y); X,y E R} {I; fx,fy E R}
{exp[i1l'(x+y)]; x,yER} {t5(fx -1/2,fy -1/2)j fx,fy E R}
{sgn(x) sgn(y); x, y E R} L1r;x i1r;Y; fx'/y E R}
{comb(x) comb(y); x,y E R} comb(fx)comb(fy)j fx,fy E R}
{exp[i1l'(x 2 +y2)]; X,Y E R} iexp [-i1l' (11 + If)] j Ix,fy E R}
{exp[-(Ixl + lyD]; X,Y E R} {!+(2;lx)2 !+(2;fy)2j Ix,fy E R}

circ( y'x 2 + y2j Y E R} Jl(21rV'i'f+if). f I E R}

X,
v'/i+!-: ' x, y

Table A.6: Two-dimensional Fourier transform pairs.

Bibliography

[1] B. S. Atal and S. L. Hanauer. Speech analysis and synthesis by linear

prediction of the speech wave. J. Acoust. Soc. Am., 50:637-655, 1971.

[2] W. R. Bennett. Spectra of quantized signals. Bell Systems Technical

Journal, 27:446-472, July 1948.

[3] S. Bochner. Lectures on Fourier Integrals. Princeton University Press,

1959.

[4] H. Bohr. Almost Periodic Functions. Chelsea, New York, 1947. Trans-
lation by Harvey Cohn.

[5) F. Bowman. Introduction to Bessel Functions. Dover, New York, 1958.

[6) R. Bracewell. The Fourier Transform and Its Applications. McGraw-

Hill, New York, 1965.

[7] R. Bracewell. The Hartley Transform. Oxford Press, New York, 1987.

[8) R.N. Bracewell. Two-Dimensional Imaging. Prentice Hall, Englewood

Cliffs, New Jersey, 1995.

[9) E. O. Brigham. The fast Fourier transform. Prentice-Hall, Englewood

Cliffs, New Jersey, 1974.

[10) H. S. Carslaw. An Introduction to the Theory of Fourier's Series and

Integrals. Dover, New York, 1950.

[11] A. G. Clavier, P. F. Panter, and D. D. Grieg. Distortion in a pulse

count modulation system. AlEE Transactions, 66:989-1005, 1947.

[12] A. G. Clavier, P. F. Panter, and D. D. Grieg. PCM distortion analysis.

Electrical Engineering, pages 1110-1122, November 1947.
354 BIBLIOGRAPHY

[13) J. W. Cooley and J. W. Tukey. An algorithm for the machine cal-

culation of complex Fourier series. Math. Computation, 19:297-301,
1965.
[14) 1. Daubechies. Ten Lectures on Wavelets. Society for Industrial and
Applied Mathematics, Philadelphia, 1992.
[15] W. B. Davenport and W. L Root. An Introduction to the Theory of
Random Signals and Noise. McGraw-Hill, New York, 1958.
[16) J. B. J. Fourier. Theorie analytiqe de la chaleur. 1922. Available as a
Dover Reprint.
[17) R. C. Gonzalez and P. Wintz. Digital Image Processing. Addison-
Wesley, Reading, Mass., second edition, 1987.
[18) J.W. Goodman. Introduction to Fourier Optics. McGraw-Hill, New
York,1988.
[19) U. Grenander and G. Szego. Toeplitz Forms and Their Applications.
University of California Press, Berkeley and Los Angeles, 1958.
[20) F. Itakura and S. Saito. A statistical method for estimation of speech
spectral density and formant frequencies. Electron. Commun. Japan,
53-A:36-43, 1970.
[21] 1. Katznelson. An Introduction to Harmonic Analysis. Dover, New
York, 1976.
[22) J. D. Markel and A. H. Gray, Jr. Linear Prediction of Speech. Springer-
Verlag, New York, 1976.
[23) A. V. Oppenheim and R. W. Schafer. Digital Signal Processing. Pren-
tice Hall, Englewood Cliffs,New Jersey, 1975.
[24) A. Papoulis. The Fourier Integral and Its Applications. McGraw-Hill,
New York, 1962.
[25) J. R. Pierce. The Science of Musical Sound. Scientific American Books,
New York, 1983.
[26) M. Rabbani and P. W. Jones. Digital Image Compression Techniques,
volume TT7 of Tutorial Texts in Optical Engineering. SPIE Optical
Engineering Press, Bellingham, Washington, 1991.
[27) D.R. Rao and P. Yip. Discrete Cosine Transform. Academic Press,
San Diego, California, 1990.
BIBLIOGRAPHY 355

[28] S. O. Rice. Mathematical analysis of random noise. In N. Wax and

N. Wax, editors, Selected Papers on Noise and Stochastic Processes,
pages 133~294. Dover, New York, NY, 1954. Reprinted from Bell Sys-
tems Technical Journal,Vol. 23:282~332 (1944) and Vol. 24: 46~ 156
(1945).
[29] O. Rioul and M. Vetterli. Wavelets and signal processing. IEEE Signal
Processing Magazine, 8(4):14~38, October 1991.

[30] W. McC. Siebert. Circuits,Signals,and Systems. M.I.T. Press, Cam-

bridge, 1986.
[31] Gilbert Strang. Wavelet transforms versus fourier transforms. Bulletin
(New Series) of the AMS, 28(2):288~305, April 1993.

[32] E. C. Titchmarsh. Introduction to the Theory of Fourier Integrals.

Oxford University Press, Oxford, 1937.
[33] James S. Walker. Fourier Analysis. Oxford University Press, New
York, 1988.
[34] G.K. Wallace. The JPEG still picture compression standard. Commu-
nications of the ACM, 34(4):30~44, April 1991.
[35] G. N. Watson. A Treatise on the Theory of Bessel Functions. Cam-
bridge University Press, Cambridge, 1980. Second Edition.
[36] N. Wiener. The Fourier Integral and Certain of Its Applications. Cam-
bridge University Press, New York, 1933.
[37] John W. Woods, editor. Subband Image Coding. Kluwer Academic
Publishers, Boston, 1991.
Index

«5-response, 253 infinite, 144

z transform, 53, 72 Bernoulli, 48
2D Fourier transform, 309 Bessel function, 17,77, 139,336
Bessinc function, 78
DTFT (discrete time finite dura- bilateral, 72
tion Fourier transform), Binomial, 48
57 bit rate, 338
Bohr-Fourier series, 245, 337
absolutely integrable, 96, 104
box function, 8
absolutely summable, 94, 100
boxcar, 43
additivity
boxcar window, 43
countable, 252
butterfly pattern, 84
aliasing, 178, 278
almost periodic function, 337
carrier frequency, 164
almost periodic signal, 246
Cauchy-Schwartz inequality, 281
alphabet, 47
central limit theorem, 290
AM,164
ambiguity function, 297 central slice theorem, 320
amplitude modulation, 164 centroid, 196, 338
anti-Hermitian, 202 circular convolution, 258
area, 191 circularly symmetric, 32
area property, 191 closed interval, xviii
arithmetic Commutative Law, 259
modular, xx complete, 146
Associative Law, 259 compression ratio, 158
autocorrelation function, 280 continuous, 126
autocorrelation width, 285 convolution, 257, 258, 312
circular, 258
backprojection, 324 cyclic, 258
band-limited, 170 convolution and backprojection, 324
bandlimited, 184 convolution identity, 260
bandwidth, 170, 190,279 convolution integral, 133, 136,257
equivalent, 198 convolution sum, 256, 258
finite, 144 convolution, two-dimensional, 312
INDEX 357

correlation coefficient, 281 finite duration, 57

Correlation Theorem, 284 dual, 66
cosine transform, 71 dual convolution theorem, 269, 294
countable additivity, 252 duality, 140
countable linearity, 162
crosscorrelation, 279 eigenfunction, 272
CTFT, 55 energy, 166
cyclic convolution, 258 equivalent bandwidth, 198
cyclic shift, 25 equivalent pulse width, 198
equivalent time width, 198
DCT, 71, 78 equivalent width, 198
decimation, 82 error
deconvolution, 298 mean squared, 156
delay, 24 quantizer, 339
delta even, xix
Kronecker, 9,117 expansions
Dirac, 9, 223 orthogonal, 155
delta function, 9 exponential, 48
Dirac, 260 discrete time, 8
Kronecker, 260 extended linearity, 162, 252
Derivative Theorem, 187
DFT,57 fast Fourier transform, 82
inverse, 117 FFT, 64, 82
Difference theorem, 187 filter, 20
dilations, 148 filtered backprojection, 326
dimension, 146 finite bandwidth, 144
Dirac delta, 9 finite duration, 1, 56
Dirac delta funtion, 223 finite energy, 95, 97
Dirichlet kernel, 133, 231 first moment property, 192
discrete Fourier transform FM, 17,344
inverse, 117 Fourier analysis, xi
distribution, 224 Fourier coefficients, 138
Distributive Law, 259 Fourier integral, 123
domain, 1 Fourier integral kernel, 136
domain of definition, 1, 3 Fourier series, 120, 121
dot product, 74 continuous time, 138
doublet, 236 generalized, 147
doubly exponential, 48 Fourier transform, 53
downsampling, 82, 183 2D,309
DSB-AM, 164 continuous time, 55
DSB-SC, 165 discrete, 57
DTFT,55 discrete time, 55
358 INDEX

fast, 64 inner product, 74, 147, 168

finite duration integral theorem, 274, 275
discrete time (see DFT), 57 interpolation, 299, 331
inverse, 115 inverse Fourier transform, 115
two-dimensional, 309 inversion, 310
Fourier transform pair, 115 inversion problem, 320
Fourier-Bessel transform, 77, 152
frequency, 53 Jacobi-Anger expansion, 139
frequency domain signal, 54 Jacobi-Anger formula, 336
frequency modulation, 17, 344 jinc function, 78
frequency scaling, 177 JPEG,159
function, xvii jump discontinuity, 126
generalized, 11
box, 8 Kronecker delta, 9, 117

Gaussian pdf, 48 lag, 279

generalized function, 11, 224 Laplace transform, 72
geometric, 48 Laplacian, 48
granular noise, 342 limit in the mean, 95, 97, 103
group, 24 linear, 22, 251
linear functional, 228
Haar wavelet, 148 linear modulation, 165
Hankel transform, 77, 152 linear operator, 228
hard limiter, 344 linearity, 109, 310
harmonic analysis, xi countable, 162
Hartley transform, 71 extended, 162, 252
Heaviside, 14 lower limit, 126
Heaviside step function, 14 LTI,256
Hermitian, xx, 280
Hilbert transform, 294 mean, 196
homogeneous, 23 mean square, 95, 97
mean squared abscissa, 196
ideal sampling function, 241 mean squared error, 156, 158
IDFT (inverse discrete Fourier trans- mean squared width, 286
form), 117 Mellin transform, 74
imaginary part, xviii memory less nonlinearity, 334
impulse pair, 244 memory less system, 20, 334
impulse response, 253, 258 mesh,29
impulse train, 238 modular arithmetic, xx
index set, 1, 3 modulating, 164
multidimensional, 2 Modulation, 164
infinite bandwidth, 144 modulation
infinite duration, 1, 56 amplitude, 164
INDEX 359

frequency, 344 piecewise continuous, 126

linear, 165 piecewise smooth, 126
phase, 337 pixel, 2,29
modulation index, 165 PM,335
moment generating function, 190 pmf,47
moment of inertia, 194 Bernoulli, 48
moment theorem, 193 doubly exponential, 48
moment theorems, 190 Poisson, 48
mother wavelet, 148 Binomial, 48
MSE,158 geometric, 48
multiresolution, 150 Poisson, 48
Poisson summation, 179
natural sampling, 299 Poisson summation theorem, 176
nonlinear, 23 probability density function, 47
nonlinearity probability mass function, 47
memoryless, 334 progression
Nyquist frequency, 172 geometric, 2, 61
Nyquist rate, 172 projection, 317, 319
projection-slice theorem, 320, 321
odd,xx
pulse amplitude modulation, 41
one-sided, 4, 72
pulse width, 190
open interval, xviii
pulse width, equivalent, 198
orthogonal, 119
orthogonal expansions, 155 quantization, 17
orthogonality, 119, 123 uniform, 338
orthonormal, 119, 146, 155 quantizer error, 339
overload, 342
raster, 2
PAM, 41,180,209 rate, 338
Parseval's equality, 158 Rayleigh's Theorem, 167
Parseval's Theorem, 166, 167, 169 Rayleigh's theorem, 166
Parseval's theorem, 312 real part, xviii
pdf,47 region of convergence (ROC), 72
Laplacian, 48 ROC, 72
exponential, 48
Gaussian, 48 sample-and-hold, 41
uniform, 48 sampling, 37,170
pel, 29 sampling frequency, 172
period,5 sampling fmiction
periodic extension, 25,44, 45 ideal, 239
periodic signal, 45 sampling period, 37, 171
phase, xviii sampling theorem, 173
phase modulation, 335, 337 saturation, 342
360 INDEX

scalar product, 74, 147 system, 20

second moment property, 194 memoryless, 20, 334
separable signal, 29 linear, 22
sequence, 2 system function, 272
series systems, 19
Bohr-Fourier, 245, 337
shah function, 244 Taylor series, 238
shift, 24, 311 time invariant, 27, 334
cyclic, 25 time series, 2
shift invariant, 27 time width
sidebands, 165 equivalent, 198
sifting property, 223, 225 time-limited, 1
signal, 1, 3 Toeplitz function, 254
continuous time, 2 transfer function, 272
discrete time, 2 transform
finite duration, 1, 2 z,53
frequency domain, 54 cosine, 71
geometric, 8 discrete cosine, 78
infinite duration, 2 Fourier, 53
periodic, 45 Fourier-Bessel, 77
separable, 29 Hankel, 77
time-limited, 1 Mellin, 74
even, xix multidimensional, 74
infinite duration, 1 sine, 71
odd,xx Fourier-Bessel, 152
signum, 217 Hankel, 152
similarity, 311 Hartley, 71
sinc, 11, 18 Hilbert, 294
sine transform, 71 transform coding, 158
single-sideband modulation, 294 transform method, 335
space invariant, 314 triangle, 14
spectrum, 53, 54 trigonometric series representation,
88B,294 138
stationary, 27 two-dimensional convolution, 312
stretch, 311 two-dimensional Fourier transform,
stretch theorem, 181 309
subband filtering, 151 two-sided, 4, 72
subsampling, 82
superposition integral, 255 uncertainty product, 287
superposition property, 161 uncertainty relation, 287
superposition summation, 254 uniform, 48
symmetry, 310 uniform quantization, 338
INDEX 361

unilateral, 72
unit impulse function, 223
unit sample, 9
unit sample response, 253
unit step, 11
unitary, 119
upper limit, 126
upsampling, 187

variance, 196

waveform, 2
wavelet, 147
Haar, 148
wedge, 14
Whittaker-Shannon-Kotelnikov sam-
pling theorem, 173
width
autocorrelation, 285
mean squared, 286
window, 43
windowing, 43

zero filling, 44