TEAM LinG

INTRODUCTION TO
DIGITAL SIGNAL
PROCESSING AND
FILTER DESIGN
INTRODUCTION TO
DIGITAL SIGNAL
PROCESSING AND
FILTER DESIGN

B. A. Shenoi

A JOHN WILEY & SONS, INC., PUBLICATION

Copyright © 2006 by John Wiley & Sons, Inc. All rights reserved.

Published by John Wiley & Sons, Inc., Hoboken, New Jersey.

Published simultaneously in Canada.

No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any
form or by any means, electronic, mechanical, photocopying, recording, scanning, or otherwise,
except as permitted under Section 107 or 108 of the 1976 United States Copyright Act, without
either the prior written permission of the Publisher, or authorization through payment of the
appropriate per-copy fee to the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers,
MA 01923, (978) 750-8400, fax (978) 750-4470, or on the web at www.copyright.com. Requests
to the Publisher for permission should be addressed to the Permissions Department, John Wiley &
Sons, Inc., 111 River Street, Hoboken, NJ 07030, (201) 748-6011, fax (201) 748-6008, or online at
http://www.wiley.com/go/permission.

Limit of Liability/Disclaimer of Warranty: While the publisher and author have used their best
efforts in preparing this book, they make no representations or warranties with respect to the
accuracy or completeness of the contents of this book and specifically disclaim any implied
warranties of merchantability or fitness for a particular purpose. No warranty may be created or
extended by sales representatives or written sales materials. The advice and strategies contained
herein may not be suitable for your situation. You should consult with a professional where
appropriate. Neither the publisher nor author shall be liable for any loss of profit or any other
commercial damages, including but not limited to special, incidental, consequential, or other
damages.

For general information on our other products and services or for technical support, please contact
our Customer Care Department within the United States at (800) 762-2974, outside the United
States at (317) 572-3993 or fax (317) 572-4002.

Wiley also publishes its books in a variety of electronic formats. Some content that appears in print
may not be available in electronic formats. For more information about Wiley products, visit our
web site at www.wiley.com.

Library of Congress Cataloging-in-Publication Data:

ISBN-13 978-0-471-46482-2 (cloth)

ISBN-10 0-471- 46482-1 (cloth)

Printed in the United States of America.

10 9 8 7 6 5 4 3 2 1
 CONTENTS

Preface xi

1 Introduction 1
1.1 Introduction 1
1.2 Applications of DSP 1
1.3 Discrete-Time Signals 3
1.3.1 Modeling and Properties of Discrete-Time Signals 8
1.3.2 Unit Pulse Function 9
1.3.3 Constant Sequence 10
1.3.4 Unit Step Function 10
1.3.5 Real Exponential Function 12
1.3.6 Complex Exponential Function 12
1.3.7 Properties of cos(ω0 n) 14
1.4 History of Filter Design 19
1.5 Analog and Digital Signal Processing 23
1.5.1 Operation of a Mobile Phone Network 25
1.6 Summary 28
Problems 29
References 30

2 Time-Domain Analysis and z Transform 32

2.1 A Linear, Time-Invariant System 32
2.1.1 Models of the Discrete-Time System 33
2.1.2 Recursive Algorithm 36
2.1.3 Convolution Sum 38
2.2 z Transform Theory 41
2.2.1 Definition 41
2.2.2 Zero Input and Zero State Response 49

v
vi CONTENTS

2.2.3 Linearity of the System 50

2.2.4 Time-Invariant System 50
2.3 Using z Transform to Solve Difference Equations 51
2.3.1 More Applications of z Transform 56
2.3.2 Natural Response and Forced Response 58
2.4 Solving Difference Equations Using the Classical Method 59
2.4.1 Transient Response and Steady-State Response 63
2.5 z Transform Method Revisited 64
2.6 Convolution Revisited 65
2.7 A Model from Other Models 70
2.7.1 Review of Model Generation 72
2.8 Stability 77
2.8.1 Jury–Marden Test 78
2.9 Solution Using MATLAB Functions 81
2.10 Summary 93
Problems 94
References 110

3 Frequency-Domain Analysis 112

3.1 Introduction 112
3.2 Theory of Sampling 113
3.2.1 Sampling of Bandpass Signals 120
3.3 DTFT and IDTFT 122
3.3.1 Time-Domain Analysis of Noncausal Inputs 125
3.3.2 Time-Shifting Property 127
3.3.3 Frequency-Shifting Property 127
3.3.4 Time Reversal Property 128
3.4 DTFT of Unit Step Sequence 138
3.4.1 Differentiation Property 139
3.4.2 Multiplication Property 142
3.4.3 Conjugation Property 145
3.4.4 Symmetry Property 145
3.5 Use of MATLAB to Compute DTFT 147
3.6 DTFS and DFT 154
3.6.1 Introduction 154
 CONTENTS vii

3.6.2 Discrete-Time Fourier Series 156

3.6.3 Discrete Fourier Transform 159
3.6.4 Reconstruction of DTFT from DFT 160
3.6.5 Properties of DTFS and DFT 161
3.7 Fast Fourier Transform 170
3.8 Use of MATLAB to Compute DFT and IDFT 172
3.9 Summary 177
Problems 178
References 185

4 Infinite Impulse Response Filters 186

4.1 Introduction 186
4.2 Magnitude Approximation of Analog Filters 189
4.2.1 Maximally Flat and Butterworth Approximation 191
4.2.2 Design Theory of Butterworth Lowpass Filters 194
4.2.3 Chebyshev I Approximation 202
4.2.4 Properties of Chebyshev Polynomials 202
4.2.5 Design Theory of Chebyshev I Lowpass Filters 204
4.2.6 Chebyshev II Approximation 208
4.2.7 Design of Chebyshev II Lowpass Filters 210
4.2.8 Elliptic Function Approximation 212
4.3 Analog Frequency Transformations 212
4.3.1 Highpass Filter 212
4.3.2 Bandpass Filter 213
4.3.3 Bandstop Filter 216
4.4 Digital Filters 219
4.5 Impulse-Invariant Transformation 219
4.6 Bilinear Transformation 221
4.7 Digital Spectral Transformation 226
4.8 Allpass Filters 230
4.9 IIR Filter Design Using MATLAB 231
4.10 Yule–Walker Approximation 238
4.11 Summary 240
Problems 240
References 247
viii CONTENTS

5 Finite Impulse Response Filters 249

5.1 Introduction 249
5.1.1 Notations 250
5.2 Linear Phase Fir Filters 251
5.2.1 Properties of Linear Phase FIR Filters 256
5.3 Fourier Series Method Modified by Windows 261
5.3.1 Gibbs Phenomenon 263
5.3.2 Use of Window Functions 266
5.3.3 FIR Filter Design Procedures 268
5.4 Design of Windowed FIR Filters Using MATLAB 273
5.4.1 Estimation of Filter Order 273
5.4.2 Design of the FIR Filter 275
5.5 Equiripple Linear Phase FIR Filters 280
5.6 Design of Equiripple FIR Filters Using MATLAB 285
5.6.1 Use of MATLAB Program to Design Equiripple
FIR Filters 285
5.7 Frequency Sampling Method 289
5.8 Summary 292
Problems 294
References 301

6 Filter Realizations 303

6.1 Introduction 303
6.2 FIR Filter Realizations 305
6.2.1 Lattice Structure for FIR Filters 309
6.2.2 Linear Phase FIR Filter Realizations 310
6.3 IIR Filter Realizations 312
6.4 Allpass Filters in Parallel 320
6.4.1 Design Procedure 325
6.4.2 Lattice–Ladder Realization 326
6.5 Realization of FIR and IIR Filters Using MATLAB 327
6.5.1 MATLAB Program Used to Find Allpass
Filters in Parallel 334
6.6 Summary 346
 CONTENTS ix

Problems 347
References 353

7 Quantized Filter Analysis 354

7.1 Introduction 354
7.2 Filter Design–Analysis Tool 355
7.3 Quantized Filter Analysis 360
7.4 Binary Numbers and Arithmetic 360
7.5 Quantization Analysis of IIR Filters 367
7.6 Quantization Analysis of FIR Filters 375
7.7 Summary 379
Problems 379
References 379

8 Hardware Design Using DSP Chips 381

8.1 Introduction 381
8.2 Simulink and Real-Time Workshop 381
8.3 Design Preliminaries 383
8.4 Code Generation 385
8.5 Code Composer Studio 386
8.6 Simulator and Emulator 388
8.6.1 Embedded Target with Real-Time Workshop 389
8.7 Conclusion 389
References 390

9 MATLAB Primer 391

9.1 Introduction 391
9.1.1 Vectors, Arrays, and Matrices 392
9.1.2 Matrix Operations 393
9.1.3 Scalar Operations 398
9.1.4 Drawing Plots 400
9.1.5 MATLAB Functions 400
9.1.6 Numerical Format 401
x CONTENTS

9.1.7 Control Flow 402

9.1.8 Edit Window and M-file 403
9.2 Signal Processing Toolbox 405
9.2.1 List of Functions in Signal Processing Toolbox 406
References 414

Index 415
 PREFACE

This preface is addressed to instructors as well as students at the junior–senior

level for the following reasons. I have been teaching courses on digital signal
processing, including its applications and digital filter design, at the undergraduate
and the graduate levels for more than 25 years. One common complaint I have
heard from undergraduate students in recent years is that there are not enough
numerical problems worked out in the chapters of the book prescribed for the
course. But some of the very well known textbooks on digital signal processing
have more problems than do a few of the books published in earlier years.
However, these books are written for students in the senior and graduate levels,
and hence the junior-level students find that there is too much of mathematical
theory in these books. They also have concerns about the advanced level of
problems found at the end of chapters. I have not found a textbook on digital
signal processing that meets these complaints and concerns from junior-level
students. So here is a book that I have written to meet the junior students’ needs
and written with a student-oriented approach, based on many years of teaching
courses at the junior level.
Network Analysis is an undergraduate textbook authored by my Ph.D. thesis
advisor Professor M. E. Van Valkenburg (published by Prentice-Hall in 1964),
which became a world-famous classic, not because it contained an abundance of
all topics in network analysis discussed with the rigor and beauty of mathematical
theory, but because it helped the students understand the basic ideas in their sim-
plest form when they took the first course on network analysis. I have been highly
influenced by that book, while writing this textbook for the first course on digital
signal processing that the students take. But I also have had to remember that the
generation of undergraduate students is different; the curriculum and the topic of
digital signal processing is also different. This textbook does not contain many of
the topics that are found in the senior–graduate-level textbooks mentioned above.
One of its main features is that it uses a very large number of numerical problems
as well as problems using functions from MATLAB® (MATLAB is a registered
trademark of The MathWorks, Inc.) and Signal Processing Toolbox, worked out
in every chapter, in order to highlight the fundamental concepts. These prob-
lems are solved as examples after the theory is discussed or are worked out first
and the theory is then presented. Either way, the thrust of the approach is that
the students should understand the basic ideas, using the worked, out problems
as an instrument to achieve that goal. In some cases, the presentation is more
informal than in other cases. The students will find statements beginning with
“Note that. . .,” “Remember. . .,” or “It is pointed out,” and so on; they are meant
xi
xii PREFACE

to emphasize the important concepts and the results stated in those sentences.
Many of the important results are mentioned more than once or summarized in
order to emphasize their significance.
The other attractive feature of this book is that all the problems given at the
end of the chapters are problems that can be solved by using only the material
discussed in the chapters, so that students would feel confident that they have an
understanding of the material covered in the course when they succeed in solving
the problems. Because of such considerations mentioned above, the author claims
that the book is written with a student-oriented approach. Yet, the students should
know that the ability to understand the solution to the problems is important but
understanding the theory behind them is far more important.
The following paragraphs are addressed to the instructors teaching a junior-
level course on digital signal processing. The first seven chapters cover well-
defined topics: (1) an introduction, (2) time-domain analysis and z-transform,
(3) frequency-domain analysis, (4) infinite impulse response filters, (5) finite
impulse response filters, (6) realization of structures, and (7) quantization filter
analysis. Chapter 8 discusses hardware design, and Chapter 9 covers MATLAB.
The book treats the mainstream topics in digital signal processing with a well-
defined focus on the fundamental concepts.
Most of the senior–graduate-level textbooks treat the theory of finite wordlength
in great detail, but the students get no help in analyzing the effect of finite word-
length on the frequency response of a filter or designing a filter that meets a set
of frequency response specifications with a given wordlength and quantization
format. In Chapter 7, we discuss the use of a MATLAB tool known as the “FDA
Tool” to thoroughly investigate the effect of finite wordlength and different formats
of quantization. This is another attractive feature of the textbook, and the material
included in this chapter is not found in any other textbook published so far.
When the students have taken a course on digital signal processing, and join an
industry that designs digital signal processing (DSP) systems using commercially
available DSP chips, they have very little guidance on what they need to learn.
It is with that concern that additional material in Chapter 8 has been added,
leading them to the material that they have to learn in order to succeed in their
professional development. It is very brief but important material presented to
guide them in the right direction. The textbooks that are written on DSP hardly
provide any guidance on this matter, although there are quite a few books on
the hardware implementation of digital systems using commercially available
DSP chips. Only a few schools offer laboratory-oriented courses on the design
and testing of digital systems using such chips. Even the minimal amount of
information in Chapter 8 is not found in any other textbook that contains “digital
signal processing” in its title. However, Chapter 8 is not an exhaustive treatment
of hardware implementation but only as an introduction to what the students have
to learn when they begin a career in the industry.
Chapter 1 is devoted to discrete-time signals. It describes some applications
of digital signal processing and defines and, suggests several ways of describing
discrete-time signals. Examples of a few discrete-time signals and some basic
 PREFACE xiii

operations applied with them is followed by their properties. In particular,

the properties of complex exponential and sinusoidal discrete-time signals are
described. A brief history of analog and digital filter design is given. Then the
advantages of digital signal processing over continuous-time (analog) signal pro-
cessing is discussed in this chapter.
Chapter 2 is devoted to discrete-time systems. Several ways of modeling them
and four methods for obtaining the response of discrete-time systems when
excited by discrete-time signals are discussed in detail. The four methods are
(1) recursive algorithm, (2) convolution sum, (3) classical method, and (4) z-
transform method to find the total response in the time domain. The use of
z-transform theory to find the zero state response, zero input response, natural
and forced responses, and transient and steady-state responses is discussed in
great detail and illustrated with many numerical examples as well as the appli-
cation of MATLAB functions. Properties of discrete-time systems, unit pulse
response and transfer functions, stability theory, and the Jury–Marden test are
treated in this chapter. The amount of material on the time-domain analysis of
discrete-time systems is a lot more than that included in many other textbooks.
Chapter 3 concentrates on frequency-domain analysis. Derivation of sam-
pling theorem is followed by the derivation of the discrete-time Fourier trans-
form (DTFT) along with its importance in filter design. Several properties of
DTFT and examples of deriving the DTFT of typical discrete-time signals are
included with many numerical examples worked out to explain them. A large
number of problems solved by MATLAB functions are also added. This chapter
devoted to frequency-domain analysis is very different from those found in other
textbooks in many respects.
The design of infinite impulse response (IIR) filters is the main topic of
Chapter 4. The theory of approximation of analog filter functions, design of
analog filters that approximate specified frequency response, the use of impulse-
invariant transformation, and bilinear transformation are discussed in this chapter.
Plenty of numerical examples are worked out, and the use of MATLAB functions
to design many more filters are included, to provide a hands-on experience to
the students.
Chapter 5 is concerned with the theory and design of finite impulse response
(FIR) filters. Properties of FIR filters with linear phase, and design of such filters
by the Fourier series method modified by window functions, is a major part of
this chapter. The design of equiripple FIR filters using the Remez exchange algo-
rithm is also discussed in this chapter. Many numerical examples and MATLAB
functions are used in this chapter to illustrate the design procedures.
After learning several methods for designing IIR and FIR filters from Chapters
4 and 5, the students need to obtain as many realization structures as possible,
to enable them to investigate the effects of finite wordlength on the frequency
response of these structures and to select the best structure. In Chapter 6, we
describe methods for deriving several structures for realizing FIR filters and IIR
filters. The structures for FIR filters describe the direct, cascade, and polyphase
forms and the lattice structure along with their transpose forms. The structures for
xiv PREFACE

IIR filters include direct-form and cascade and parallel structures, lattice–ladder
structures with autoregressive (AR), moving-average (MA), and allpass struc-
tures as special cases, and lattice-coupled allpass structures. Again, this chapter
contains a large number of examples worked out numerically and using the func-
tions from MATLAB and Signal Processing Toolbox; the material is more than
what is found in many other textbooks.
The effect of finite wordlength on the frequency response of filters realized
by the many structures discussed in Chapter 6 is treated in Chapter 7, and the
treatment is significantly different from that found in all other textbooks. There
is no theoretical analysis of finite wordlength effect in this chapter, because it
is beyond the scope of a junior-level course. I have chosen to illustrate the use
of a MATLAB tool called the “FDA Tool” for investigating these effects on the
different structures, different transfer functions, and different formats for quan-
tizing the values of filter coefficients. The additional choices such as truncation,
rounding, saturation, and scaling to find the optimum filter structure, besides the
alternative choices for the many structures, transfer functions, and so on, makes
this a more powerful tool than the theoretical results. Students would find expe-
rience in using this tool far more useful than the theory in practical hardware
implementation.
Chapters 1–7 cover the core topics of digital signal processing. Chapter 8,
on hardware implementation of digital filters, briefly describes the simulation
of digital filters on Simulink®, and the generation of C code from Simulink
using Real-Time Workshop® (Simulink and Real-Time Workshop are registered
trademarks of The MathWorks, Inc.), generating assembly language code from the
C code, linking the separate sections of the assembly language code to generate an
executable object code under the Code Composer Studio from Texas Instruments
is outlined. Information on DSP Development Starter kits and simulator and
emulator boards is also included. Chapter 9, on MATLAB and Signal Processing
Toolbox, concludes the book.
The author suggests that the first three chapters, which discuss the basics of
digital signal processing, can be taught at the junior level in one quarter. The pre-
requisite for taking this course is a junior-level course on linear, continuous-time
signals and systems that covers Laplace transform, Fourier transform, and Fourier
series in particular. Chapters 4–7, which discuss the design and implementation
of digital filters, can be taught in the next quarter or in the senior year as an
elective course depending on the curriculum of the department. Instructors must
use discretion in choosing the worked-out problems for discussion in the class,
noting that the real purpose of these problems is to help the students understand
the theory. There are a few topics that are either too advanced for a junior-level
course or take too much of class time. Examples of such topics are the derivation
of the objective function that is minimized by the Remez exchange algorithm, the
formulas for deriving the lattice–ladder realization, and the derivation of the fast
Fourier transform algorithm. It is my experience that students are interested only
in the use of MATLAB functions that implement these algorithms, and hence I
have deleted a theoretical exposition of the last two topics and also a description
 PREFACE xv

of the optimization technique in the Remez exchange algorithm. However, I have

included many examples using the MATLAB functions to explain the subject
matter.
Solutions to the problems given at the end of chapters can be obtained by the in-
structors from the Website http://www.wiley.com/WileyCDA/WileyTitle/
productCd-0471464821.html. They have to access the solutions by clicking
“Download the software solutions manual link” displayed on the Webpage. The
author plans to add more problems and their solutions, posting them on the Website
frequently after the book is published.
As mentioned at the beginning of this preface, the book is written from my
own experience in teaching a junior-level course on digital signal processing.
I wish to thank Dr. M. D. Srinath, Southern Methodist University, Dallas, for
making a thorough review and constructive suggestions to improve the material
of this book. I also wish to thank my colleague Dr. A. K. Shaw, Wright State
University, Dayton. And I am most grateful to my wife Suman, who has spent
hundreds of lonely hours while I was writing this book. Without her patience
and support, I would not have even started on this project, let alone complete it.
So I dedicate this book to her and also to our family.

B. A. Shenoi

May 2005
 CHAPTER 1

Introduction

1.1 INTRODUCTION

We are living in an age of information technology. Most of this technology is

based on the theory of digital signal processing (DSP) and implementation of
the theory by devices embedded in what are known as digital signal processors
(DSPs). Of course, the theory of digital signal processing and its applications
is supported by other disciplines such as computer science and engineering, and
advances in technologies such as the design and manufacturing of very large
scale integration (VLSI) chips. The number of devices, systems, and applications
of digital signal processing currently affecting our lives is very large and there
is no end to the list of new devices, systems, and applications expected to be
introduced into the market in the coming years. Hence it is difficult to forecast
the future of digital signal processing and the impact of information technology.
Some of the current applications are described below.

1.2 APPLICATIONS OF DSP

Digital signal processing is used in several areas, including the following:

1. Telecommunications. Wireless or mobile phones are rapidly replacing

wired (landline) telephones, both of which are connected to a large-scale telecom-
munications network. They are used for voice communication as well as data
communications. So also are the computers connected to a different network
that is used for data and information processing. Computers are used to gen-
erate, transmit, and receive an enormous amount of information through the
Internet and will be used more extensively over the same network, in the com-
ing years for voice communications also. This technology is known as voice
over Internet protocol (VoIP) or Internet telephony. At present we can transmit
and receive a limited amount of text, graphics, pictures, and video images from

Introduction to Digital Signal Processing and Filter Design, by B. A. Shenoi

Copyright © 2006 John Wiley & Sons, Inc.

1
2 INTRODUCTION

mobile phones, besides voice, music, and other audio signals—all of which are
classified as multimedia—because of limited hardware in the mobile phones and
not the software that has already been developed. However, the computers can
be used to carry out the same functions more efficiently with greater memory and
large bandwidth. We see a seamless integration of wireless telephones and com-
puters already developing in the market at present. The new technologies being
used in the abovementioned applications are known by such terms as CDMA,
TDMA,1 spread spectrum, echo cancellation, channel coding, adaptive equaliza-
tion, ADPCM coding, and data encryption and decryption, some of which are
used in the software to be introduced in the third-generation (G3) mobile phones.
2. Speech Processing. The quality of speech transmission in real time over
telecommunications networks from wired (landline) telephones or wireless (cel-
lular) telephones is very high. Speech recognition, speech synthesis, speaker
verification, speech enhancement, text-to-speech translation, and speech-to-text
dictation are some of the other applications of speech processing.
3. Consumer Electronics. We have already mentioned cellular or mobile
phones. Then we have HDTV, digital cameras, digital phones, answering
machines, fax and modems, music synthesizers, recording and mixing of music
signals to produce CD and DVDs. Surround-sound entertainment systems includ-
ing CD and DVD players, laser printers, copying machines, and scanners are
found in many homes. But the TV set, PC, telephones, CD-DVD players, and
scanners are present in our homes as separate systems. However, the TV set can
be used to read email and access the Internet just like the PC; the PC can be
used to tune and view TV channels, and record and play music as well as data
on CD-DVD in addition to their use to make telephone calls on VoIP. This trend
toward the development of fewer systems with multiple applications is expected
to accelerate in the near future.
4. Biomedical Systems. The variety of machines used in hospitals and biomed-
ical applications is staggering. Included are X-ray machines, MRI, PET scanning,
bone scanning, CT scanning, ultrasound imaging, fetal monitoring, patient moni-
toring, and ECG and EEC mapping. Another example of advanced digital signal
processing is found in hearing aids and cardiac pacemakers.
5. Image Processing. Image enhancement, image restoration, image under-
standing, computer vision, radar and sonar processing, geophysical and seismic
data processing, remote sensing, and weather monitoring are some of the applica-
tions of image processing. Reconstruction of two-dimensional (2D) images from
several pictures taken at different angles and three-dimensional (3D) images from
several contiguous slices has been used in many applications.
6. Military Electronics. The applications of digital signal processing in mili-
tary and defense electronics systems use very advanced techniques. Some of the
applications are GPS and navigation, radar and sonar image processing, detection

1
Code- and time-division multiple access. In the following sections we will mention several technical
terms and well-known acronyms without any explanation or definition. A few of them will be
described in detail in the remaining part of this book.
 DISCRETE-TIME SIGNALS 3

and tracking of targets, missile guidance, secure communications, jamming and

countermeasures, remote control of surveillance aircraft, and electronic warfare.
7. Aerospace and Automotive Electronics. Applications include control of air-
craft and automotive engines, monitoring and control of flying performance of
aircraft, navigation and communications, vibration analysis and antiskid control
of cars, control of brakes in aircrafts, control of suspension, and riding comfort
of cars.
8. Industrial Applications. Numerical control, robotics, control of engines and
motors, manufacturing automation, security access, and videoconferencing are a
few of the industrial applications.

Obviously there is some overlap among these applications in different devices

and systems. It is also true that a few basic operations are common in all the
applications and systems, and these basic operations will be discussed in the
following chapters. The list of applications given above is not exhaustive. A few
applications are described in further detail in [1]. Needless to say, the number of
new applications and improvements to the existing applications will continue to
grow at a very rapid rate in the near future.

1.3 DISCRETE-TIME SIGNALS

A signal defines the variation of some physical quantity as a function of one

or more independent variables, and this variation contains information that is of
interest to us. For example, a continuous-time signal that is periodic contains the
values of its fundamental frequency and the harmonics contained in it, as well
as the amplitudes and phase angles of the individual harmonics. The purpose of
signal processing is to modify the given signal such that the quality of information
is improved in some well-defined meaning. For example, in mixing consoles for
recording music, the frequency responses of different filters are adjusted so that
the overall quality of the audio signal (music) offers as high fidelity as possible.
Note that the contents of a telephone directory or the encyclopedia downloaded
from an Internet site contains a lot of useful information but the contents do
not constitute a signal according to the definition above. It is the functional
relationship between the function and the independent variable that allows us to
derive methods for modeling the signals and find the output of the systems when
they are excited by the input signals. This also leads us to develop methods for
designing these systems such that the information contained in the input signals
is improved.
We define a continuous-time signal as a function of an independent variable
that is continuous. A one-dimensional continuous-time signal f (t) is expressed
as a function of time that varies continuously from −∞ to ∞. But it may be
a function of other variables such as temperature, pressure, or elevation; yet we
will denote them as continuous-time signals, in which time is continuous but the
signal may have discontinuities at some values of time. The signal may be a
4 INTRODUCTION

x1(t) x2(t)

0 t 0 t

(a) (b)

Figure 1.1 Two samples of continuous-time signals.

real- or complex-valued function of time. We can also define a continuous-time

signal as a mapping of the set of all values of time to a set of corresponding
values of the functions that are subject to certain properties. Since the function is
well defined for all values of time in −∞ to ∞, it is differentiable at all values
of the independent variable t (except perhaps at a finite number of values). Two
examples of continuous-time functions are shown in Figure 1.1.
A discrete-time signal is a function that is defined only at discrete instants of
time and undefined at all other values of time. Although a discrete-time function
may be defined at arbitrary values of time in the interval −∞ to ∞, we will
consider only a function defined at equal intervals of time and defined at t = nT ,
where T is a fixed interval in seconds known as the sampling period and n
is an integer variable defined over −∞ to ∞. If we choose to sample f (t) at
equal intervals of T seconds, we generate f (nT ) = f (t)|t=nT as a sequence of
numbers. Since T is fixed, f (nT ) is a function of only the integer variable n and
hence can be considered as a function of n or expressed as f (n). The continuous-
time function f (t) and the discrete-time function f (n) are plotted in Figure 1.2.
In this book, we will denote a discrete-time (DT) function as a DT sequence,
DT signal, or a DT series. So a DT function is a mapping of a set of all integers
to a set of values of the functions that may be real-valued or complex-valued.
Values of both f (t) and f (n) are assumed to be continuous, taking any value
in a continuous range; hence can √ have a value even with an infinite number of
digits, for example, f (3) = 0.4 2 in Figure 1.2.
A zero-order hold (ZOH) circuit is used to sample a continuous signal f (t)
with a sampling period T and hold the sampled values for one period before the
next sampling takes place. The DT signal so generated by the ZOH is shown in
Figure 1.3, in which the value of the sample value during each period of sam-
pling is a constant; the sample can assume any continuous value. The signals of
this type are known as sampled-data signals, and they are used extensively in
sampled-data control systems and switched-capacitor filters. However, the dura-
tion of time over which the samples are held constant may be a very small
fraction of the sampling period in these systems. When the value of a sample
 DISCRETE-TIME SIGNALS 5

7/8

6/8

5/8

4/8

3/8

2/8

1/8

0.0
−4 −3 −2 −1 0 1 2 3 4 5 6 7 8 n
−1/8

−2/8

−3/8

Figure 1.2 The continuous-time function f (t) and the discrete-time function f (n).

−3 −2 −1 0 1 2 3 4 5 6 n

Figure 1.3 Sampled data signal.

6 INTRODUCTION

is held constant during a period T (or a fraction of T ) by the ZOH circuit as

its output, that signal can be converted to a value by a quantizer circuit, with
finite levels of value as determined by the binary form of representation. Such a
process is called binary coding or quantization. A This process is discussed in
full detail in Chapter 7. The precision with which the values are represented is
determined by the number of bits (binary digits) used to represent each value.
If, for example, we select 3 bits, to express their values using a method known
as “signed magnitude fixed-point binary number representation” and one more
bit to denote positive or negative values, we have the finite number of values,
represented in binary form and in their equivalent decimal form. Note that a
4-bit binary form can represent values between − 78 and 78 at 15 distinct levels
as shown in Table 1.1. So a value of f (n) at the output of the ZOH, which lies
between these distinct levels, is rounded or truncated by the quantizer according
to some rules and the output of the quantizer when coded to its equivalent binary
representation, is called the digital signal. Although there is a difference between
the discrete-time signal and digital signal, in the next few chapters we assume
that the signals are discrete-time signals and in Chapter 7, we consider the effect
of quantizing the signals to their binary form, on the frequency response of the

TABLE 1.1 4 Bit Binary Numbers

and their Decimal Equivalents

Binary Form Decimal Value

0 111 7
8 = 0.875
0 110 6
8 = 0.750
0 101 5
8 = 0.625
0 100 4
8 = 0.500
0 011 3
8 = 0.375
0 010 2
8 = 0.250
0 001 1
8 = 0.125
0 000 0.0 = 0.000
1 000 −0.0 = −0.000
1 001 − 18 = −0.125
1 010 − 28 = −0.250
1 011 − 38 = −0.375
1 100 − 48 = −0.500
1 101 − 58 = −0.625
1 110 − 68 = −0.750
1 111 − 78 = −0.875
 DISCRETE-TIME SIGNALS 7

filters. However, we use the terms digital filter and discrete-time system inter-
changeably in this book. Continuous-time signals and systems are also called
analog signals and analog systems, respectively. A system that contains both the
ZOH circuit and the quantizer is called an analog-to digital converter (ADC),
which will be discussed in more detail in Chapter 7.
Consider an analog signal as shown by the solid line in Figure 1.2. When it
is sampled, let us assume that the discrete-time sequence has values as listed
in the second column of Table 1.2. They are expressed in only six significant
decimal digits and their values, when truncated to four digits, are shown in the
third column. When these values are quantized by the quantizer with four binary
digits (bits), the decimal values are truncated to the values at the finite discrete
levels. In decimal number notation, the values are listed in the fourth column,
and in binary number notation, they are listed in the fifth column of Table 1.2.
The binary values of f (n) listed in the third column of Table 1.2 are plotted in
Figure 1.4.
A continuous-time signal f (t) or a discrete-time signal f (n) expresses the
variation of a physical quantity as a function of one variable. A black-and-white
photograph can be considered as a two-dimensional signal f (m, r), when the
intensity of the dots making up the picture is measured along the horizontal axis
(x axis; abscissa) and the vertical axis (y axis; ordinate) of the picture plane
and are expressed as a function of two integer variables m and r, respectively.
We can consider the signal f (m, r) as the discretized form of a two-dimensional
signal f (x, y), where x and y are the continuous spatial variables for the hor-
izontal and vertical coordinates of the picture and T1 and T2 are the sampling

TABLE 1.2 Numbers in Decimal and Binary Forms

Values of f (n)
Decimal Truncated to Quantized Binary
n Values of f (n) Four Digits Values of f (n) Number Form

−4 −0.054307 −0.0543 0.000 1 000

−3 −0.253287 −0.2532 −0.250 1 010
−2 −0.236654 −0.2366 −0.125 1 001
−1 −0.125101 −0.1251 −0.125 1 001
0 0.522312 0.5223 0.000 0 000
1 0.246210 0.2462 0.125 0 001
2 0.387508 0.3875 0.375 0 011
3 0.554090 0.5540 0.500 0 100
4 0.521112 0.5211 0.500 0 100
5 0.275432 0.2754 0.250 0 010
6 0.194501 0.1945 0.125 0 001
7 0.168887 0.1687 0.125 0 001
8 0.217588 0.2175 0.125 0 001
8 INTRODUCTION

7/8

6/8

5/8

4/8

3/8

2/8

1/8
−4 −3 −2 −1
0 1 2 3 4 5 6 7 8 n

Figure 1.4 Binary values in Table 1.2, after truncation of f (n) to 4 bits.

periods (measured in meters) along the x and y axes, respectively. In other words,
f (x, y)|x=mT1 ,y=rT2 = f (m, r).
A black-and-white video signal f (x, y, t) is a 3D function of two spatial
coordinates x and y and one temporal coordinate t. When it is discretized, we
have a 3D discrete signal f (m, p, n). When a color video signal is to be modeled,
it is expressed by a vector of three 3D signals, each representing one of the
three primary colors—red, green, and blue—or their equivalent forms of two
luminance and one chrominance. So this is an example of multivariable function
or a multichannel signal:
⎡ ⎤
fr (m, p, n)
F(m, r, n) = ⎣ fg (m, p, n) ⎦ (1.1)
fb (m, p, n)

1.3.1 Modeling and Properties of Discrete-Time Signals

There are several ways of describing the functional relationship between the
integer variable n and the value of the discrete-time signal f (n): (1) to plot the
values of f (n) versus n as shown in Figure 1.2, (2) to tabulate their values as
shown in Table 1.2, and (3) to define the sequence by expressing the sample
values as elements of a set, when the sequence has a finite number of samples.
For example, in a sequence x1 (n) as shown below, the arrow indicates the
value of the sample when n = 0:
 
x1 (n) = 2 3 1.5 0.5 −1 4 (1.2)
↑
 DISCRETE-TIME SIGNALS 9

We denote the DT sequence by x(n) and also the value of a sample of the
sequence at a particular value of n by x(n). If a sequence has zero values for
n < 0, then it is called a causal sequence. It is misleading to state that the
causal function is a sequence defined for n ≥ 0, because, strictly speaking, a DT
sequence has to be defined for all values of n. Hence it is understood that a causal
sequence has zero-valued samples for −∞ < n < 0. Similarly, when a function
is defined for N1 ≤ n ≤ N2 , it is understood that the function has zero values for
−∞ < n < N1 and N2 < n < ∞. So the sequence x1 (n) in Equation (1.2) has
zero values for 2 < n < ∞ and for −∞ < n < −3. The discrete-time sequence
x2 (n) given below is a causal sequence. In this form for representing x2 (n), it is
implied that x2 (n) = 0 for −∞ < n < 0 and also for 4 < n < ∞:
 
x2 (n) = 1 −2 0.4 0.3 0.4 0 0 0 (1.3)
↑

The length of a finite sequence is often defined by other authors as the number
of samples, which becomes a little ambiguous in the case of a sequence like x2 (n)
given above. The function x2 (n) is the same as x3 (n) given below:
 
x3 (n) = 1 −2 0.4 0.3 0.4 0 0 0 0 0 0 (1.4)
↑

But does it have more samples? So the length of the sequence x3 (n) would be
different from the length of x2 (n) according to the definition above. When a
sequence such as x4 (n) given below is considered, the definition again gives an
ambiguous answer:
 
x4 (n) = 0 0 0.4 0.3 0.4 (1.5)
↑

The definition for the length of a DT sequence would be refined when we

define the degree (or order) of a polynomial in z−1 to express the z transform of
a DT sequence, in the next chapter.
To model the discrete-time signals mathematically, instead of listing their
values as shown above or plotting as shown in Figure 1.2, we introduce some
basic DT functions as follows.

1.3.2 Unit Pulse Function

The unit pulse function δ(n) is defined by

1 n=0
δ(n) = (1.6)
0 n= 0

and it is plotted in Figure 1.5a. It is often called the unit sample function and also
the unit impulse function. But note that the function δ(n) has a finite numerical
10 INTRODUCTION

δ(n) δ(n − 3) δ(n + 3)

−2 −1 0 1 2 3 n −1 0 1 2 3 n −3 −2 −1 0 1 2 3 n

(a) (b) (c)

Figure 1.5 Unit pulse functions δ(n), δ(n − 3), and δ(n + 3).

value of one at n = 0 and zero at all other values of integer n, whereas the unit
impulse function δ(t) is defined entirely in a different way.
When the unit pulse function is delayed by k samples, it is described by

1 n=k
δ(n − k) = (1.7)
0 n = k

and it is plotted in Figure 1.5b for k = 3. When δ(n) is advanced by k = 3, we

get δ(n + k), and it is plotted in Figure 1.5c.

1.3.3 Constant Sequence

This sequence x(n) has a constant value for all n and is therefore defined by
x(n) = K; −∞ < n < ∞.

1.3.4 Unit Step Function

The unit step function u(n) is defined by

1 n≥0
u(n) = (1.8)
0 n<0

and it is plotted in Figure 1.6a.

When the unit step function is delayed by k samples, where k is a positive
integer, we have

1 n≥k
u(n − k) = (1.9)
0 n<k
 DISCRETE-TIME SIGNALS 11

u(n) u(n − 2)

0 1 2 3 4 5 n 0 1 2 3 4 5 6 7 n
(a) (b)

u(−n) u(−n + 2)

−6 −5 −4 −3 −2 −1 0 1 2 3 n −4 −3 −2 −1 0 1 2 3 n
(c) (d )

u(−n − 2)

−6 −5 −4 −3 −2 −1 0 1 2 3 4 n

(e)

Figure 1.6 Unit step functions.

The sequence u(n + k) is obtained when u(n) is advanced by k samples. It is

defined by

1 n ≥ −k
u(n + k) = (1.10)
0 n < −k

We also define the function u(−n), obtained from the time reversal of u(n), as a
sequence that is zero for n > 0. The sequences u(−n + k) and u(−n − k), where
k is a positive integer, are obtained when u(−n) is delayed by k samples and
advanced by k samples, respectively. In other words, u(−n + k) is obtained by
12 INTRODUCTION

delaying u(−n) when k is positive and obtained by advancing u(−n) when k is a

negative integer. Note that the effect on u(−n − k) is opposite that on u(n − k),
when k is assumed to take positive and negative values. These functions are
shown in Figure 1.6, where k = 2. In a strict sense, all of these functions are
defined implicitly for −∞ < n < ∞.

1.3.5 Real Exponential Function

The real exponential function is defined by

x(n) = a n ; −∞ < n < ∞ (1.11)

where a is real constant. If a is a complex constant, it becomes the complex

exponential sequence. The real exponential sequence or the complex exponential
sequence may also be defined by a more general relationship of the form
 n
a k≤n<∞
x(n) = (1.12)
bn −∞ < n < k

A special discrete-time sequence that we often use is the function defined for
n ≥ 0:

x(n) = a n u(n) (1.13)

An example of x1 (n) = (0.8)n u(n) is plotted in Figure 1.7a. The function x2 (n) =
x1 (n − 3) = (0.8)(n−3) u(n − 3) is obtained when x1 (n) is delayed by three sam-
ples. It is plotted in Figure 1.7b. But the function x3 (n) = (0.8)n u(n − 3) is
obtained by chopping off the first three samples of x1 (n) = (0.8)n u(n), and as
shown in Figure 1.7c, it is different from x2 (n).

1.3.6 Complex Exponential Function

The complex exponential sequence is a function that is complex-valued as a
function of n. The most general form of such a function is given by

x(n) = Aα n , −∞ < n < ∞ (1.14)

where both A and α are complex numbers. If we let A = |A| ej φ and α =

e(σ0 +j ω0 ) , where σ0 , ω0 , and φ are real numbers, the sequence can be expanded
to the form

x(n) = |A| ej φ e(σ0 +j ω0 )n

= |A| eσ0 n ej (ω0 n+φ)
= |A| eσ0 n cos(ω0 n + φ) + j |A| eσ0 n sin(ω0 n + φ) (1.15)
= xre (n) + j xim (n)
 DISCRETE-TIME SIGNALS 13

Plot of x1(n)
1

0.8
Value of x1(n)

0.6

0.4

0.2

0
0 2 4 6 8 10 12
Value of n
(a)

Plot of x2(n)
1

0.8
Value of x2(n)

0.6

0.4

0.2

0
0 2 4 6 8 10 12
Value of n
(b)

Plot of x3(n)
1

0.8
Value of x3(n)

0.6

0.4

0.2

0
0 2 4 6 8 10 12
Value of n
(c)

Figure 1.7 Plots of x1 (n), x2 (n), and x3 (n).

14 INTRODUCTION

When σ0 = 0, the real and imaginary parts of this complex exponential

sequence are |A| cos(ω0 n + φ) and |A| sin(ω0 n + φ), respectively, and are real
sinusoidal sequences with an amplitude equal to |A|. When σ0 > 0, the two
sequences increase as n → ∞ and decrease when σ0 < 0 as n → ∞. When
ω0 = φ = 0, the sequence reduces to the real exponential sequence |A| eσ0 n .

1.3.7 Properties of cos(ω0 n)

When A = 1, and σ0 = φ = 0, we get x(n) = ej ω0 n = cos(ω0 n) + j sin(ω0 n).
This function has some interesting properties, when compared with the

continuous-time function ej ω0 t and they are described below.

First we point out that ω0 in x(n) = ej ω0 n is a frequency normalized by fs =
1/T , where fs is the sampling frequency in hertz and T is the sampling period
in seconds, specifically, ω0 = 2πf0
/fs = ω0
T , where ω0
= 2πf0
is the actual real
frequency in radians per second and f0
is the actual frequency in hertz. Therefore
the unit of the normalized frequency ω0 is radians. It is common practice in
the literature on discrete-time systems to choose ω as the normalized frequency
variable, and we follow that notation in the following chapters; here we denote
ω0 as a constant in radians. We will discuss this normalized frequency again in
a later chapter.

Property 1.1 In the complex exponential function x(n) = ej ω0 n , two frequen-

cies separated by an integer multiple of 2π are indistinguishable from each
other. In other words, it is easily seen that ej ω0 n = ej (ω0 n+2πr) . The real part and
the imaginary part of the function x(n) = ej ω0 n , which are sinusoidal functions,
also exhibit this property. As an example, we have plotted x1 (n) = cos(0.3πn)
and x2 (n) = cos(0.3π + 4π)n in Figure 1.8. In contrast, we know that two
continuous-time functions x1 (t) = ej ω1 t and x2 (t) = ej ω2 t or their real and imag-
inary parts are different if ω1 and ω2 are different. They are different even if they
are separated by integer multiples of 2π. From the property ej ω0 n = ej (ω0 n+2πr)
above, we arrive at another important result, namely, that the output of a discrete-
time system has the same value when these two functions are excited by the
complex exponential functions ej ω0 n or e j (ω0 n+2πr) . We will show in Chapter 3
that this is true for all frequencies separated by integer multiples of 2π, and
therefore the frequency response of a DT system is periodic in ω.

Property 1.2 Another important property of the sequence ej ω0 n is that it is

periodic in n. A discrete-time function x(n) is defined to be periodic if there
exists an integer N such that x(n + rN ) = x(n), where r is any arbitrary integer
and N is the period of the periodic sequence. To find the value for N such that
ej ω0 n is periodic, we equate ej ω0 n to ej ω0 (n+rN ) . Therefore ej ω0 n = ej ω0 n ej ω0 rN ,
which condition is satisfied when ej ω0 rN = 1, that is, when ω0 N = 2πK, where
 DISCRETE-TIME SIGNALS 15

Plot of x(n) = cos(0.3pn) Plot of y(n) = cos(0.3p + 4p)n

1 1
0.8 0.8
0.6 0.6
0.4 0.4
Value of x(n)

Value of y(n)
0.2 0.2
0 0
−0.2 −0.2
−0.4 −0.4
−0.6 −0.6
−0.8 −0.8
−1 −1
0 10 20 30 40 0 10 20 30 40
Value of n Value of n
(a) (b)

Figure 1.8 Plots of cos(0.3πn) and cos(0.3π + 4π)n.

K is any arbitrary integer. This condition is satisfied by the following equation:

ω0 K
= (1.16)
2π N
In words, this means that the ratio of the given normalized frequency ω0 and 2π
must be a rational number. The period of the sequence N is given by
2πK
N= (1.17)
ω0
When this condition is satisfied by the smallest integer K, the corresponding
value of N gives the fundamental period of the periodic sequence, and integer
multiples of this frequency are the harmonic frequencies.

Example 1.1

Consider a sequence x(n) = cos(0.3πn). In this case ω0 = 0.3π and ω0 /2π =

0.3π/2π = 20 3
. Therefore the sequence is periodic and its period N is 20 samples.
This periodicity is noticed in Figure 1.8a and also in Figure 1.8b.
Consider another sequence x(n) = cos(0.5n), in which case ω0 = 0.5. There-
fore ω0 /2π = 0.5/2π = 1/4π, which is not a rational number. Hence this is not
a periodic sequence.
When the given sequence is the sum of several complex exponential functions,
each of which is periodic with different periods, it is still periodic. We consider
an example to illustrate the method to find the fundamental period in this case.
16 INTRODUCTION

Plot of x3(n)
1

0.8

0.6

0.4
Value of x3(n)

0.2

−0.2

−0.4

−0.6

−0.8

−1
0 5 10 15 20 25 30 35 40
Value of n

Figure 1.9 Plot of x3 (n).

Suppose x3 (n) = cos(0.2πn) + cos(0.5πn) + cos(0.6πn). Its fundamental

period N must satisfy the condition

2πK1 2πK2 2πK3

N= = = (1.18)
0.2π 0.5π 0.6π
10K3
= 10K1 = 4K2 = (1.19)
3

where K1 , K2 , and K3 and N are integers. The value of N that satisfies this
condition is 20 when K1 = 2, K2 = 5, and K3 = 6. So N = 20 is the fundamental
period of x3 (n). The sequence x3 (n) plotted in Figure 1.9 for 0 ≤ n ≤ 40 shows
that it is periodic with a period of 20 samples.

Property 1.3 We have already observed that the frequencies at ω0 and at ω0 +

2π are the same, and hence the frequency of oscillation are the same. But con-
sider the frequency of oscillation as ω0 changes between 0 and 2π. It is found
that the frequency of oscillation of the sinusoidal sequence cos(ω0 n) increases
as ω0 increases from 0 to π and the frequency of oscillation decreases as ω0
increases from π to 2π. Therefore the highest frequency of oscillation of a
discrete-time sequence cos(ω0 n) occurs when ω0 = ±π. When the normalized
frequency ω0 = 2πf0
/fs attains the value of π, the value of f0
= fs /2. So the
highest frequency of oscillation occurs when it is equal to half the sampling
 DISCRETE-TIME SIGNALS 17

Plot of y1 = cos(0.1pn) Plot of y2 = cos(0.3pn)

1 1

0.5 0.5

Value of y2(n)
Value of y1(n)

0 0

−0.5 −0.5

−1 −1
0 20 40 60 0 20 40 60
Value of n Value of n

Plot of y3 = cos(pn) Plot of y4 = cos(1.8pn)

1 1

0.5 0.5
Value of y3(n)

Value of y4(n)

0 0

−0.5 −0.5

−1 −1
0 20 40 60 0 20 40 60
Value of n Value of n

Figure 1.10 Plot of cos(ω0 n) for different values of ω0 between 0 and 2π.

frequency. In Figure 1.10 we have plotted the DT sequences as ω0 attains a few

values between 0 and 2π, to illustrate this property. We will elaborate on this
property in later chapters of the book.
Since frequencies separated by 2π are the same, as ω0 increases from 2π to
3π, the frequency of oscillation increases in the same manner as the frequency
of oscillation when it increases from 0 to π. As an example, we see that the
frequency of v0 (n) = cos(0.1πn) is the same as that of v1 (n) = cos(2.1πn). It
is interesting to note that v2 (n) = cos(1.9πn) also has the same frequency of
oscillation as v1 (n) because

v2 (n) = cos(1.9πn) = cos(2π − 0.1πn) (1.20)

= cos(2πn) cos(0.1πn) + sin(2πn) sin(0.1πn) (1.21)
= cos(0.1πn)
= v0 (n)
18 INTRODUCTION

v1 (n) = cos(2.1πn) = cos(2πn + 0.1πn) (1.22)

= cos(2πn) cos(0.1πn) − sin(2πn) sin(0.1πn) (1.23)
= cos(0.1n)
= v0 (n)

We have plotted the sequences v1 (n) and v2 (n) in Figure 1.11, to verify this
property.

Remember that in Chapter 3, we will use the term “folding” to describe new
implications of this property. We will also show in Chapter 3 that a large class
of discrete-time signals can be expressed as the weighted sum of exponential
sequences of the form ej ω0 n , and such a model leads us to derive some powerful
analytical techniques of digital signal processing.
We have described several ways of characterizing the DT sequences in this
chapter. Using the unit sample function and the unit step function, we can express
the DT sequences in other ways as shown below. 
For example, δ(n) = u(n) − u(n − 1) and u(n) = m=n m=−∞ δ(m). A mathe-
matical way of modeling a sequence
 
x(n) = 2 3 1.5 0.5 −1 4 (1.24)
↑

Plot of v1(n) = cos(2.1pn) Plot of v2(n) = cos(1.9pn)

1 1

0.8 0.8

0.6 0.6

0.4 0.4
Value of v1(n)

Value of v2(n)

0.2 0.2

0 0

−0.2 −0.2

−0.4 −0.4

−0.6 −0.6

−0.8 −0.8

−1 −1
0 10 20 30 40 0 10 20 30 40
Value of n Value of n

Figure 1.11 Plots of cos(2.1πn) and cos(1.9πn).

 HISTORY OF FILTER DESIGN 19

is the weighted sum of shifted unit sample functions, as given by

x(n) = 2δ(n + 3) + 3δ(n + 2) + 1.5δ(n + 1) + 0.5δ(n) − δ(n − 1) + 4δ(n − 2)

(1.25)
If the sequence is given in an analytic form x(n) = a n u(n), it can also be
expressed as the weighted sum of impulse functions:
∞
∞

x(n) = x(m)δ(n − m) = a m δ(n − m) (1.26)
m=0 m=0

In the next chapter, we will introduce a transform known as the z transform,

which will be used to model the DT sequences in additional forms. We will
show that this model given by (1.26) is very useful in deriving the z transform
and in analyzing the performance of discrete-time systems.

1.4 HISTORY OF FILTER DESIGN

Filtering is the most common form of signal processing used in all the appli-
cations mentioned in Section 1.2, to remove the frequencies in certain parts
and to improve the magnitude, phase, or group delay in some other part(s) of
the spectrum of a signal. The vast literature on filters consists of two parts:
(1) the theory of approximation to derive the transfer function of the filter such
that the magnitude, phase, or group delay approximates the given frequency
response specifications and (2) procedures to design the filters using the hardware
components. Originally filters were designed using inductors, capacitors, and
transformers and were terminated by resistors representing the load and the inter-
nal resistance of the source. These were called the LC (inductance × capacitance)
filters that admirably met the filtering requirements in the telephone networks for
many decades of the nineteenth and twentieth centuries. When the vacuum tubes
and bipolar junction transistors were developed, the design procedure had to
be changed in order to integrate the models for these active devices into the
filter circuits, but the mathematical theory of filter approximation was being
advanced independently of these devices. In the second half of the twentieth
century, operational amplifiers using bipolar transistors were introduced and fil-
ters were designed without inductors to realize the transfer functions. The design
procedure was much simpler, and device technology also was improved to fabri-
cate resistors in the form of thick-film and later thin-film depositions on ceramic
substrates instead of using printed circuit boards. These filters did not use induc-
tors and transformers and were known as active-RC (resistance × capacitance)
filters. In the second half of the century, switched-capacitor filters were devel-
oped, and they are the most common type of filters being used at present for
audio applications. These filters contained only capacitors and operational ampli-
fiers using complementary metal oxide semiconductor (CMOS) transistors. They
used no resistors and inductors, and the whole circuit was fabricated by the
20 INTRODUCTION

very large scale integration (VLSI) technology. The analog signals were con-
verted to sampled data signals by these filters and the signal processing was
treated as analog signal processing. But later, the signals were transitioned as
discrete-time signals, and the theory of discrete-time systems is currently used to
analyze and design these filters. Examples of an LC filter, an active-RC filter,
and a switched-capacitor filter that realize a third-order lowpass filter function
are shown in Figures 1.12–1.14.
The evolution of digital signal processing has a different history. At the begin-
ning, the development of discrete-time system theory was motivated by a search
for numerical techniques to perform integration and interpolation and to solve
differential equations. When computers became available, the solution of phys-
ical systems modeled by differential equations was implemented by the digital

R2

+
L2
Rs

+ c1 c2 v2 RL
_ v1

Figure 1.12 A lowpass analog LC filter.

+
−
+
−

+
−
+
+
−

v2
v1 +
−

Figure 1.13 An active-RC lowpass analog filter.

 HISTORY OF FILTER DESIGN 21

f1 f1 f1 f1

f2 f2 f2 f2

+
v1 −
+
−

f2 f1

f1 f2

f1 f1 f2 f1

−
+ −
+ +
f2 f2 f1 f2 v2
−

Figure 1.14 A switched-capacitor lowpass (analog) filter.

computers. As the digital computers became more powerful in their computa-

tional power, they were heavily used by the oil industry for geologic signal
processing and by the telecommunications industry for speech processing. The
theory of digital filters matured, and with the advent of more powerful computers
built on integrated circuit technology, the theory and applications of digital signal
processing has explosively advanced in the last few decades. The two revolution-
ary results that have formed the foundations of digital signal processing are the
Shannon’s sampling theorem and the Cooley–Tukey algorithm for fast Fourier
transform technique. Both of them will be discussed in great detail in the follow-
ing chapters. The Shannon’s sampling theorem proved that if a continuous-time
signal is bandlimited (i.e., if its Fourier transform is zero for frequencies above
a maximum frequency fm ) and it is sampled at a rate that is more than twice
the maximum frequency fm in the signal, then no information contained in the
analog signal is lost in the sense that the continuous-time signal can be exactly
reconstructed from the samples of the discrete-time signal. In practical applica-
tions, most of the analog signals are first fed to an analog lowpass filter—known
22 INTRODUCTION

as the preconditioning filter or antialiasing filter —such that the output of the
lowpass filter attenuates the frequencies considerably beyond a well-chosen fre-
quency so that it can be considered a bandlimited signal. It is this signal that
is sampled and converted to a discrete-time signal and coded to a digital signal
by the analog-to-digital converter (ADC) that was briefly discussed earlier in
this chapter. We consider the discrete-time signal as the input to the digital filter
designed in such a way that it improves the information contained in the original
analog signal or its equivalent discrete-time signal generated by sampling it. A
typical example of a digital lowpass filter is shown in Figure 1.15.
The output of the digital filter is next fed to a digital-to-analog converter
(DAC) as shown in Figure 1.17 that also uses a lowpass analog filter that smooths
the sampled-data signal from the DAC and is known as the “smoothing filter.”
Thus we obtain an analog signal yd (t) at the output of the smoothing filter as
shown. It is obvious that compared to the analog filter shown in Figure 1.16, the
circuit shown in Figure 1.17 requires considerably more hardware or involves a
lot more signal processing in order to filter out the undesirable frequencies from
the analog signal x(t) and deliver an output signal yd (t). It is appropriate to
compare these two circuit configurations and determine whether it is possible to
get the output yd (t) that is the same or nearly the same as the output y(t) shown
in Figure 1.16; if so, what are the advantages of digital signal processing instead
of analog signal processing, even though digital signal processing requires more
circuits compared to analog signal processing?

x(n) Σ Σ

z−1
z−1

z−1 Σ y(n)

Figure 1.15 A lowpass third-order digital filter.

Analog Analog Analog

Signal
Input x(t) Processor Output y(t)

Figure 1.16 Example of an analog signal processing system.

 ANALOG AND DIGITAL SIGNAL PROCESSING 23

Analog Preconditioning Sample Digital Analog Analog

Analog and ADC Signal DAC Low Pass
Input Low Pass Filter Hold Processor Filter Output
x(t) y(t)

Figure 1.17 Example of a digital signal processing system.

1.5 ANALOG AND DIGITAL SIGNAL PROCESSING

The basic elements in digital filters are the multipliers, adders, and delay ele-
ments, and they carry out multiplication, addition, and shifting operations on
numbers according to an algorithm determined by the transfer function of the
filters or their equivalent models. (These models will be discussed in Chapter 3
and also in Chapter 7.) They provide more flexibility and versatility compared
to analog filters. The coefficients of the transfer function and the sample values
of the input signal can be stored in the memory of the digital filter hardware or
on the computer (PC, workstation, or the mainframe computer), and by changing
the coefficients, we can change the transfer function of the filter, while chang-
ing the sample values of the input, we can find the response of the filter due
to any number of input signals. This flexibility is not easily available in ana-
log filters.
The digital filters are easily programmed to do time-shared filtering under time-
division multiplexing scheme, whereas the analog signals cannot be interleaved
between timeslots. Digital filters can be designed to serve as time-varying filters
also by changing the sampling frequency and by changing the coefficients as a
function of time, namely, by changing the algorithm accordingly.
The digital filters have the advantage of high precision and reliability. Very
high precision can be obtained by increasing the number of bits to represent
the coefficients of the filter transfer function and the values of the input signal.
Again we can increase the dynamic range of the signals and transfer function
coefficients by choosing floating-point representation of binary numbers. The
values of the inductors, capacitors, and the parameters of the operational amplifier
parameters and CMOS transistors, and so on used in the analog filters cannot
achieve such high precision. Even if the analog elements can be obtained with
high accuracy, they are subject to great drift in their value due to manufacturing
tolerance, temperature, humidity, and other parameters—depending on the type
of device technology used—over long periods of service, and hence their filter
response degrades slowly and eventually fails to meet the specifications. In the
case of digital filters, such effects are nonexistent because the wordlength of
the transfer coefficients as well as the product of addition and multiplication
within the filter do not change with respect to time or any of the environmental
conditions that plague the analog circuits. Consequently, the reliability of digital
filters is much higher than that of analog filters, and this means that they are more
economical in application. Of course, catastrophic failures due to unforeseen
factors are equally possible in both cases. If we are using computers to analyze,
24 INTRODUCTION

design, and simulate these filters, we can assume even double-precision format
for the numbers that represent filter coefficients and signal samples. We point
out that we can carry out the simulation, analysis, and design of any number
of filters and under many conditions, for example, Monte Carlo analysis, worst-
case analysis, or iterative optimization to test the design before we build the
hardware and test it again and again. Of course, we can do the same in the case
of analog filters or continuous-time systems also (e.g., analog control systems)
using such software as MATLAB and Simulink.2 During the manufacture of
analog filters, we may have to tune each of them to correct for manufacturing
tolerances, but there is no such need to test the accuracy of the wordlength in
digital filters.
Data on digital filters can be stored on magnetic tapes, compact disks(CDs),
digital videodisks (DVDs), and optical disks for an indefinite length of time.
They can be retrieved without any degradation or loss of data; a good example
is the music recorded on CDs. In contrast, analog signals deteriorate slowly
as time passes and cannot be retrieved easily without any loss. There is no
easy way of storing the transfer function coefficients that defines the analog
system and feeding the input signals stored on these storage devices to the ana-
log system.
By using digital filters, we can realize many transfer functions that cannot be
realized by analog filters. For example, in addition to those already mentioned
above, we can realize the following characteristics from digital filters:

1. Transition bands much smaller than what can be achieved from analog fil-
ters; an example would be a lowpass filter with a bandwidth of 5000 Hz
and a passband ripple of 0.5 dB, and 100 dB attenuation above 5010 Hz. In
spectrum analyzers and synthesizers, vocoders (voice recorders), and simi-
lar devices, extremely low tolerances on the magnitude and phase responses
over adjacent passbands are required, and digital filters can be designed to
meet these specifications.
2. Finite duration impulse response and filters with linear phase. Neither of
these characteristics can be achieved by analog filters. Digital filters with
these characteristics are used extensively in many applications.
3. Bandwidth of the order 5 Hz or even a fraction thereof that are commonly
required to process biomedical or seismic signals.
4. Programmable filters, multirate filters, multidimensional filters, and adap-
tive filters. Programmable filters are used to adjust the frequency-selective
properties of the filters. Multirate filters are used in the processing of
many complex signals with different rates of fluctuation, whereas two-
dimensional digital filters are the filters used in image processing. Adaptive
filters are used invariably when the transmission medium between the trans-
mitter and receiver changes—either as the transmission line is switched to

2
MATLAB and Simulink are registered trademarks of The MathWorks, Inc. Natick, MA. The soft-
ware is available from The MathWorks, Inc.
 ANALOG AND DIGITAL SIGNAL PROCESSING 25

different receivers or as it changes continuously between the transmitter

and the receiver. For example, when a telephone conversation is switched
from one point to another and the cable or the microwave link changes, or
when the mobile phone moves as the talker moves over a wide territory,
adaptive filters are absolutely necessary to compensate for the distortion of
the signal as it passes through the transmission link.
5. We have chosen filters only as an example to compare digital and analog
signal processors. There are many other types of digital signal processing
that are feasible and are being used, and these are not possible or very effi-
cient in analog filters. For example, error detection in transmitted signals
and correction to reduce the error rate is an advanced technique used in
many applications. Another example is our ability to compress the data by
a significant factor and receive the input signal at lower cost and very good
quality. To point out the power of digital signal processing theory and the
digital signal processors available, let us again consider the mobile phone.
Bateman and Patterson-Stephans state that “Within the phone, a single DSP
device may be performing real-time speech compression, video compres-
sion, echo cancellation, noise cancellation, voice recognition, waveform
coding, modulation/demodulation, interleaving, multipath equalization, soft
decision decoding, convolution, automatic frequency-, power- and gain-
control” [3], and all of them done in a triband phone with TDMA, CDMA,
and analog signal processing! The mobile phone is just an example to illus-
trate the large number of digital signal processing techniques that are built
into any of the applications described above. But an application such as
the mobile phone implements other functions also, and their features are
briefly described given below.

1.5.1 Operation of a Mobile Phone Network

Consider a geographic area in which a part of the mobile phone network operates.
It is divided into cells as indicated in Figure 1.18. The cells are not really equal
in area but could be as small as 300 m where the telephone traffic is high and as
large as 35 km in rural areas. The size and shape of each cell is determined by
the radiation pattern and the power output of the antenna (and is not hexagonal
of equal shape and size) serving the mobile phones. A base station controller
(BSC), usually installed on a tower, serves as many as 124 base transceiver
stations (BTSs). These stations communicate with all the cell phones that are
known to be located within the area covered by the cell. The BTSs operate on
different frequencies, also called “channels,” to transmit to and receive signals
from the cell phones. Global System for Mobile Communication (GSM) is one
of the most widely used mobile cellular phone network systems in the world,
and in that system, the frequencies for transmitting from the mobile phone and
receiving by the BTS lie in the band 890–915 MHz, and the frequencies for
transmitting from the BTS and receiving by the mobile phone lie in the band
935–960 MHz. But in order to utilize the frequency spectrum efficiently, cells
26 INTRODUCTION

2
3
1
4
2 2 2
3 3 3
1 1 1
4 4 4
2 2 2
3 3 3
1 1 1
4 4 4

Figure 1.18 Cell repeat pattern in a mobile network coverage area.

using the same frequency are kept sufficiently apart so that there is no cochannel
interference; also the frequencies used in a BTS are separated by 200 kHz. The
base transceiver stations located on towers over a coverage area are connected by
fixed digital lines to a mobile switching center (MSC), and the mobile switching
center is connected to the public switched telephone network (PSTN) as shown
in Figure 1.19 as well as the Internet, to which other MSCs are also connected.
When a phone initiates a call to send voice, text, an instant message, or other
media, it registers with the network via the BTS closest to its location and the
BTS tracks its location and passes this information to the mobile switching center
(MSC) over fixed digital lines, which updates this information continuously as
received from the BTS. Each mobile phone has a home location register (HLR)
and a visitor location register (VLR) assigned to it. The HLR contains information
such as the identity of the user and phone number assigned to the user in the
user’s home network, the services to which the user has subscribed, whereas the
VLR contains information about the mobile phone when it is used outside the
home network. So when a mobile phone initiates a call, it sends the information
to the BTS about its identity and so on from the VLR or the HLR depending on
the location of the phone at the time the call originates. The mobile switching
center checks the data from its HLR or VLR to authenticate the call and gives
permission for the phone to access the network. As the caller moves within the
cell, the BTS monitors the strength of the signal between the phone and the
receiver, and if this falls below a certain level, it may transfer control of the
phone to the BTS in the next cell, which may offer a stronger signal. If no such
cell is nearby, the caller is cut off (i.e., will not be able to receive or to send
 ANALOG AND DIGITAL SIGNAL PROCESSING 27

Base transceiver Cell

station (BTS) phone

Base station
controller (BSC)

Base transceiver Cell

station (BTS) phone

Mobile
PSTN switching
center (MSC) Base
stations

Visitor Home
location location Base transceiver Cell
register register station (BTS) phone
(VLR) (HLR)
Base station
controller (BSC)

Base transceiver Cell

station (BTS) phone

Figure 1.19 Block diagram of a mobile cell phone network.

a call). As the caller moves from one cell to another cell, the BTS serving it
will transfer control to the BTS in the cell that it has moved to. This is the
main feature that makes mobile telephony possible. All of these operations are
carried out by the computers serving the mobile cellular phone network, and that
technology is known as computer networking technology. It is different from the
theory of digital signal processing. This textbook offers an introduction to the
fundamental theory of digital signal processing, which is used in such techniques
as speech compression, multipath equalization, and echo cancellation, mentioned
in the previous section.
There are some disadvantages and limitations in digital signal processing in
comparison with analog signal processing. By looking at the two circuit config-
urations in Figures 1.16 and 1.17, it is obvious that the digital signal processor
is a more complex system, because of the additional components of analog low-
pass filters (ADC and DAC) on either side of the digital signal processor besides
the additional control and programming circuits, which are not shown in the
figures. Another disadvantage is that the digital signal processor can process sig-
nals within a range of frequencies that is limited mainly by the highest sampling
28 INTRODUCTION

TABLE 1.3 ADCs Currently Available

Maximum
Sampling Rate Resolution Frequency in
(samples per second) (bits) Input Signal Power

96,000 24 48 kHz 90 mW
96,000 18 48 kHz 60 mW
96,000 16 48 kHz 40 mW
65,000,000 14 500 MHz 0.6 W
400,000,000 8 1 GHz 3W

frequency of the ADC and DAC that are available. As the frequency is increased,
the wordlength of these devices decreases and therefore the accuracy and dynamic
range of the input and output data decrease.
For example, data on a few ADCs currently available are given in Table 1.3 [3].
Hence digital signal processing is restricted to approximately one megahertz,
and analog signal processors are necessary for processing signals above that
frequency, for example, processing of radar signals. In such applications, analog
signal processing is a more attractive and viable choice, and currently a lot of
research is being directed toward what is known as mixed-signal processing. Note
in Table 1.3, that as the resolution (wordlength) for a given signal decreases, the
power consumption also decreases, but that is just the power consumed by the
ADCs; the power increases as the sampling frequency increases, even when
the resolution is decreased. The digital signal processor itself consumes a lot
more power, and hence additional power management circuits are added to the
whole system. In contrast, the analog signal processors consume less power.
The LC filters consume almost negligible power and can operate at frequencies
in the megahertz range. The active-RC filters and switched-capacitor filters are
restricted to the audiofrequency range, but they consume more power than do
the LC filters. It is expected that mixed-signal processing carried out on a single
system or a single chip will boost the maximum frequency of the signal that can
be processed, by a significant range, beyond what is possible with a strictly digital
signal processing system. Therefore we will see more and more applications of
DSP with increasing frequencies because the advantages of DSP outweigh the
disadvantages in analog signal processing.

1.6 SUMMARY

In this introductory chapter, we defined the discrete-time signal and gave a few
examples of these signals, along with some simple operations that can be applied
with them. In particular, we pointed out the difference between a sinusoidal
signal, which is a continuous-time signal, and a discrete-time signal. We dis-
cussed the basic procedure followed to sample and quantize an analog signal
 PROBLEMS 29

and compared the advantages and disadvantages of digital signal processing with
those of directly processing the analog signal through an analog system, tak-
ing a filter as an example. In doing so, we introduced many terms or acronyms
that we have not explained. Some of them will be explained in great detail in
the following chapter. In Chapter 2 we will discuss several ways of modeling
a discrete-time system and the methods used to find the response in the time
domain, when excitation is effected by the discrete-time signals.

PROBLEMS

1.1 Given two discrete-time signals x1 (n) = {0.9 0.5 0.8 1.0 1.5 2.0
↑
0.2} and x2 (n) = {1.0 0.3 0.6 0.4}, sketch each of the following:
↑
(a) y1 (n) = x1 (n) + 3x2 (n)
(b) y2 (n) = x1 (n) − x2 (n − 5)
(c) y3 (n) = x1 (n)x2 (n)
(d) y4 (n) = x1 (−n + 4)
(e) y5 (n) = y4 (n)x2 (n)
(f) y6 (n) = x2 (−n − 3)
(g) y7 (n) = y4 (n)y6 (n)
1.2 Sketch each of the following, where x1 (n) and x2 (n) are the DT sequences
given in Problem 1.1:
(a) v1 (n) = x1 (n)x2 (4 − n)

(b) v2 (n) = ∞ k=−∞ x1 (k)x2 (n − k)
∞
(c) v3 (n) = k=−∞ x2 (k)x1 (n − k)

(d) v4 (n) = n=10 2
n=0 x2 (n)
(e) v5 (n) = x1 (2n)
1.3 Repeat Problem 1.1 with x1 (n) = {1.0 0.8 0.2 − 0.2 − 0.5
↑
− 0.7} and x2 (n) = {0.5 0.2 0.1 0.2 0.6}.
↑

1.4 Repeat Problem 1.2 with x1 (n) and x2 (n) as given in Problem 1.3.
1.5 Find the even and odd parts of x1 (n) and x2 (n) given in Problem 1.1.
Even part of x1 (n) is defined as [x1 (n) + x1 (−n)]/2, and the odd part as
[x1 (n) − x1 (−n)]/2
1.6 Repeat Problem 1.5 with x1 (n) and x2 (n) as given in Problem 1.3.

1.7 Find S1 (k) = n=k n=0 n and sketch it, for k = 0, 1, . . . 5.
 
1.8 Find and sketch S2 (k) = kn=−∞ x1 (n) and S3 (k) = kn=−∞ x2 (n), where
x1 (n) and x2 (n) are as given in Problem 1.1.
30 INTRODUCTION

1.9 Repeat Problem 1.8 with x1 (n) and x2 (n) as given in Problem 1.3.
1.10 Express the sequences x1 (n) and x2 (n) as a summation of weighted and
shifted unit step functions, where x1 (n) and x2 (n) are as given in Prob-
lem 1.1.
1.11 Repeat Problem 1.10 with the sequences x1 (n) and x2 (n) given in Prob-
lem 1.3.
1.12 Given x(n) = [0.5ej (π/6) ]n [u(n) − u(n − 4)], calculate the values of
|x(n)| and sketch them. What are the real and imaginary parts of x(n)?

1.13 Express the real and imaginary parts of x(n) = 4k=0 [0.5ej (π/6) ]k
δ(n − k).

1.14 What are the real and imaginary parts of q(n) = ∞ n=0 (0.3 − j 0.4) ?
n

1.15 What is the fundamental period of p1 (n) = ej 0.3πn ?

1.16 Find the fundamental period of p2 (n) = ej 0.4πn + 2ej 0.6πn .
1.17 Find the fundamental period of p3 (n) = cos(0.5πn) + 4 cos(2.5πn).
1.18 What is the fundamental period of p4 (n) = cos(0.2πn) + cos(0.7πn) +
cos(πn)?
1.19 Find the fundamental period of p5 (n) = cos(0.5n) + 3 cos(0.2n).
1.20 Find the fundamental period of p6 (n) = p1 (n)p2 (n), where p1 (n) and
p2 (n) are as given above.
1.21 Find the fundamental period of p7 (n) = cos(1.2πn) + 4 sin(0.2πn).
1.22 What is the fundamental period of p8 (n) = cos(0.1πn)p3 (n)?
1.23 The sinusoidal sequence h(n) = A cos(ω0 n + φ) has the following values
over one period; find the values of A, ω0 and φ:

h(n) = {2.00 0.00 − 2.00 − 0.00}

REFERENCES

1. S. K. Mitra and J. F. Kaiser, Handbook for Digital Signal Processing, Wiley-

Interscience, 1993, Chapter 1.
2. J. G. Proakis and D. G. Manolakis, Digital Signal Processing, Prentice-Hall, 1996.
3. A. Bateman and I. Patterson-Stephans, The DSP Handbook, Algorithms, Applications
and Design Techniques, Prentice-Hall, 2000.
4. S. K. Mitra, Digital Signal Processing, A Computer-Based Approach, McGraw-Hill,
1998.
5. A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing, Prentice-
Hall, 1989.
 REFERENCES 31

6. R. D. Strum and D. E. Kirk, Discrete Systems and Digital Signal Processing,

Addison-Wesley, 1989.
7. S. S. Soliman and M. D. Srinath, Continuous and Discrete Signals and Systems,
Prentice-Hall, 1990.
8. L. R. Rabiner and B. Gold, Theory and Application of Digital Signal Processing,
Prentice-Hall, 1975.
9. E. C. Ifeachor and B. W. Jervis, Digital Signal Processing, Prentice-Hall, 2002.
10. V. K. Ingle and J. G. Proakis, Digital Signal Processing Using MATLAB(R) V.4, PWS
Publishing, 1997.
 CHAPTER 2

Time-Domain Analysis
and z Transform

2.1 A LINEAR, TIME-INVARIANT SYSTEM

The purpose of analysis of a discrete-time system is to find the output in either

the time or frequency domain of the system due to a discrete-time input signal. In
Chapter 1, we defined the discrete-time signal as a function of the integer variable
n, which represents discrete time, space, or some other physical variable. Given
any integer value in −∞ < n < ∞, we can find the value of the signal according
to some well-defined relationship. This can be described as a mapping of the set
of integers to a set of values of the discrete-time signal. Description of this
relationship varied according to the different ways of modeling the signal. In this
chapter, we define the discrete-time system as a mapping of the set of discrete-
time signals considered as the input to the system, to another set of discrete-time
signals identified as the output of the system. This mapping can also be defined
by an analytic expression, formula, algorithm, or rule, in the sense that if we
are given an input to the system, we can find the output signal. The mapping
can therefore be described by several models for the system. The mapping or
the input–output relationship may be linear, nonlinear, time-invariant, or time-
varying. The system defined by this relationship is said to be linear if it satisfies
the following conditions.
Assume that the output is y(n) due to an input x(n) according to this rela-
tionship. If an input Kx(n) produces an output Ky(n), the system satisfies the
condition of homogeneity, where K is any arbitrary constant. Let K1 y1 (n) and
K2 y2 (n) be the outputs due to the inputs K1 x1 (n) and K2 x2 (n), respectively,
where K1 and K2 are arbitrary constants. If the output is K1 y1 (n) + K2 y2 (n)
when the input is K1 y1 (n) + K2 y2 (n), then the system satisfies the superposition
property. A system that satisfies both homogeneity and superposition is defined as
a linear system. If the output is y(n − M) when the input is delayed by M sam-
ples, that is, when the input is x(n − M), the system is said to be time-invariant
or shift-invariant. If the output is determined by the weighted sum of only the
previous values of the output and the weighted sum of the current and previous

Introduction to Digital Signal Processing and Filter Design, by B. A. Shenoi

Copyright © 2006 John Wiley & Sons, Inc.

32
 A LINEAR, TIME-INVARIANT SYSTEM 33

values of the input, then the system is defined as a causal system. This means
that the output does not depend on the future values of the input. We will discuss
these concepts again in more detail in later sections of this chapter. In this book,
we consider only discrete-time systems that are linear and time-invariant (LTI)
systems.
Another way of defining a system in general is that it is an interconnection
of components or subsystems, where we know the input–output relationship of
these components, and that it is the way they are interconnected that determines
the input–output relationship of the whole system. The model for the DT system
can therefore be described by a circuit diagram showing the interconnection of
its components, which are the delay elements, multipliers, and adders, which are
introduced below. In the following sections we will use both of these definitions
to model discrete-time systems. Then, in the remainder of this chapter, we will
discuss several ways of analyzing the discrete-time systems in the time domain,
and in Chapter 3 we will discuss frequency-domain analysis.

2.1.1 Models of the Discrete-Time System

First let us consider a discrete-time system as an interconnection of only three
basic components: the delay elements, multipliers, and adders. The input–output
relationships for these components and their symbols are shown in Figure 2.1.
The fourth component is the modulator, which multiplies two or more signals
and hence performs a nonlinear operation.
A simple discrete-time system is shown in Figure 2.2, where input signal
x(n) = {x(0), x(1), x(2), x(3)} is shown to the left of v0 (n) = x(n). The signal

Delay Element X1(n) z−1 Y1(n) = X1(n − 1)

Multiplier X2(n) K Y2(n) = KX2(n)

X3(n)
Adder Σ Y3(n) = X3(n) + X4(n)
X4(n)

Modulator X5(n) Σ Y5(n) = X5(n)m6(n)

m6(n)

Figure 2.1 The basic components used in a discrete-time system.

34 TIME-DOMAIN ANALYSIS AND z TRANSFORM

X(n)
x(0)
x(3)
x(1)
x(2)

b(0)
V0(n)
n
0 1 2 3 4 5 6 7

x(0)
x(3) z−1
x(1)
x(2)

b(1)
V1(n)
n
0 1 2 3 4
x(0)
z−1
x(3)
x(1) Σ y(n)
x(2)

b(2)
V2(n)
0 1 2 3 4 5 6 7 n

x(0) z−1
x(3)
x(1)
x(2)
b(3)
V3(n)
0 1 2 3 4 5 6 n
7

Figure 2.2 Operations in a typical discrete-time system.

v1 (n) shown on the left is the signal x(n) delayed by T seconds or one sam-
ple, so, v1 (n) = x(n − 1). Similarly, v(2) and v(3) are the signals obtained
from x(n) when it is delayed by 2T and 3T seconds: v2 (n) = x(n − 2) and
v3 (n) = x(n − 3). When we say that the signal x(n) is delayed by T , 2T , or 3T
seconds, we mean that the samples of the sequence are present T , 2T , or 3T
seconds later, as shown by the plots of the signals to the left of v1 (n), v2 (n),
and v3 (n). But at any given time t = nT , the samples in v1 (n), v2 (n), and v3 (n)
are the samples of the input signal that occur T , 2T , and 3T seconds previous to
t = nT . For example, at t = 3T , the value of the sample in x(n) is x(3), and the
values present in v1 (n), v2 (n) and v3 (n) are x(2), x(1), and x(0), respectively.
A good understanding of the operation of the discrete-time system as illustrated
above is essential in analyzing, testing, and debugging the operation of the sys-
tem when available software is used for the design, simulation, and hardware
implementation of the system.
It is easily seen that the output signal in Figure 2.2 is

y(n) = b(0)v(0) + b(1)v(1) + b(2)v(2) + b(3)v(3)

= b(0)x(n) + b(1)x(n − 1) + b(2)x(n − 2) + b(3)x(n − 3)
 A LINEAR, TIME-INVARIANT SYSTEM 35

0.8

−0.1
y1(z) y2(z) y3(z)
X(z) z−1 Σ Σ Σ

z−1
z−1
0.5

0.3 z−1y2(z)
−0.4 0.6
z−1

z−2y1(z)
−0.2

Figure 2.3 Schematic circuit for a discrete-time system.

where b(0), b(1), b(2), b(3) are the gain constants of the multipliers. It is also
easy to see from the last expression that the output signal is the weighted sum of
the current value and the previous three values of the input signal. So this gives
us an input–output relationship for the system shown in Figure 2.2.
Now we consider another example of a discrete-time system, shown in
Figure 2.3. Note that a fundamental rule is to express the output of the adders
and generate as many equations as the number of adders found in this circuit
diagram for the discrete-time system. (This step is similar to writing the node
equations for an analog electric circuit.) Denoting the outputs of the three adders
as y1 (n), y2 (n), and y3 (n), we get

y1 (n) = 0.3y1 (n − 1) − 0.2y1 (n − 2) − 0.1x(n − 1)

y2 (n) = y1 (n) + 0.5y1 (n − 1) − 0.4y2 (n − 1)
y3 (n) = y2 (n) + 0.6y2 (n − 1) + 0.8y1 (n) (2.1)

These three equations give us a mathematical model derived from the model
shown in Figure 2.3 that is schematic in nature. We can also derive (draw
the circuit realization) the model shown in Figure 2.3 from the model given in
Equations (2.1). We will soon describe a method to obtain a single input–output
relationship between the input x(n) and the output y(n) = y3 (n), after eliminat-
ing the internal variables y1 (n) and y2 (n); that relationship constitutes the third
model for the system. The general form of such an input–output relationship is

N
M
y(n) = − a(k)y(n − k) + b(k)x(n − k) (2.2)
k=1 k=0
36 TIME-DOMAIN ANALYSIS AND z TRANSFORM

or in another equivalent form

N
M
a(k)y(n − k) = b(k)x(n − k); a(0) = 1 (2.3)
k=0 k=0

Equation (2.2) shows that the output y(n) is determined by the weighted sum
of the previous N values of the output and the weighted sum of the current and
previous M + 1 values of the input. Very often the coefficient a(0) as shown in
(2.3) is normalized to unity.
Soon we will introduce the z transform to represent the discrete-time signals
in the set of equations above, thereby generating more models for the system, and
from these models in the z domain, we will derive the transfer function H (z−1 )
and the unit sample response or the unit impulse response h(n) of the system.
From any one of these models in the z domain, we can derive the other models in
the z domain and also the preceding models given in the time domain. It is very
important to know how to obtain any one model from any other given model
so that the proper tools can be used efficiently, for analysis of the discrete-time
system. In this chapter we will elaborate on the different models of a discrete-
time system and then discuss many tools or techniques for finding the response of
discrete-time systems when they are excited by different kinds of input signals.

2.1.2 Recursive Algorithm

Let us consider an example of Equation (2.2) as y(n) = y(n − 1) − 0.25y(n −
2) + x(n), where the input sequence x(n) = δ(n), and the two initial conditions
are y(−1) = 1.0 and y(−2) = 0.4.
We compute y(0), y(1), y(2), . . . in a recursive manner as follows: y(0) =
y(−1) − 0.25y(−2) + x(0). Since x(n) = δ(n), we substitute x(0) = 1 and get
y(0) = 1.0 − 0.25(0.4) + 1 = 1.9. Next y(1) = y(0) − 0.25y(−1) + x(1). We
know y(0) = 1.9 from the step shown above, and also that x(1) = 0. So
we get y(1) = 1.9 − 0.25(1.0) + 0 = 1.65. Next, for n = 2, when we compute
y(2) = y(1) − 0.25y(0) + x(2). Substituting the known values from above, we
get y(2) = 1.65 − 0.25(1.9) + 0 = 1.175.
Next, when n = 3, we obtain

y(3) = y(2) − 0.25y(1) + x(3)

= 1.175 − 0.25(1.65) + 0 = 0.760

We can continue to calculate the values of the output y(n) for n = 4, 5, 6, 7, . . . .

This is known as the recursive algorithm, which we use to calculate the output
when we are given an equation of the form (2.2); it can be used when there is
any other input. For a system modeled by an equation of the form (2.2), the
output is infinite in length in general. As a special case, when the input is the
unit impulse function δ(n), and the initial conditions are assumed to be zero, the
 A LINEAR, TIME-INVARIANT SYSTEM 37

resulting output is called the unit impulse response h(n) (or more appropriately
the unit sample response) and is infinite in length.
Consider a system in which the multiplier constants a(k) = 0 for k =
1, 2, 3, . . . , N . Then Equation (2.2) reduces to the form

M
y(n) = b(k)x(n − k) (2.4)
k=0
= b(0)x(n) + b(1)x(n − 1) + b(2)x(n − 2) + · · · + b(M)x(n − M)

Let us find the unit impulse response of this system, using the recursive algo-
rithm, as before:

y(0) = b(0)(1) + 0 + 0 + 0 + · · · = b(0)

y(1) = b(0)x(1) + b(1)x(0) + b(2)x(−1) + 0 + 0 + · · · = b(1)
y(2) = b(0)x(2) + b(1)x(1) + b(2)x(0) + 0 + 0 + 0 + · · · = b(2)

Continuing this procedure recursively, we would get

y(3) = b(3)
y(4) = b(4)
·
·
·
y(M) = b(M)

This example leads to the following two observations: (1) the samples of the unit
impulse response are the same as the coefficients b(n), and (2) therefore the unit
impulse response h(n) of the system is finite in length.
So we have shown without proof but by way of example that the unit impulse
response of the system modeled by an equation of the form (2.2) is infinite in
length, and hence such a system is known as an infinite impulse response (IIR)
filter, whereas the system modeled by an equation of the form (2.4), which has
an unit impulse response that is finite in length, is known as the finite impulse
response (FIR) filter. We will have a lot more to say about these two types of
filters later in the book. Equation (2.3) is the ordinary, linear, time-invariant,
difference equation of N th order, which, if necessary, can be rewritten in the
recursive difference equation form (2.2). The equation can be solved in the time
domain, by the following four methods:

1. The recursive algorithm as explained above

2. The convolution sum, to get the zero state response, as explained in the
next section
38 TIME-DOMAIN ANALYSIS AND z TRANSFORM

3. The classical method of solving a difference equation

4. The analytical solution using the z transform.

We should point out that methods 1–3 require that the DT system be modeled by a
single-input, single-output equation. If we are given a large number of difference
equations describing the DT system, then methods 1–3 are not suitable for finding
the output response in the time domain. Method 4, using the z transform, is the
only powerful and general method to solve such a problem, and hence it will be
treated in greater detail and illustrated by several examples in this chapter. Given
a model in the z-transform domain, we will show how to derive the recursive
algorithm and the unit impulse response h(n) so that the convolution sum can be
applied. So the z-transform method is used most often for time-domain analysis,
and the frequency-domain analysis is closely related to this method, as will be
discussed in the next chapter.

2.1.3 Convolution Sum

In the discussion above, we have assumed that the unit impulse response of a
discrete-time system when it is excited by a unit impulse function δ(n), exists
(or is known), and we denote it as h(n). Instead of using the recursive algo-
rithm to find the response due to any input, let us represent the input sig-
nal x(n) not by its values in a sequence {x(0), x(1), x(2), x(3), . . .} but as
the values of impulse function at the corresponding instants of time. In other
words, we consider the sequence of impulse functions x(0)δ(n), x(1)δ(n − 1),
x(2)δ(n − 2), . . . as the input—and not the sequence of values {x(0), x(1),
x(2), x(3), . . .}. The difference between the values of the samples as a sequence
of numbers and the sequence of impulse functions described above should be
clearly understood. The first operation is simple sampling operation, whereas the
second is known as impulse sampling, which is a mathematical way to repre-
sent the same data, and we represent the second sequence in a compact form:
x(n) = ∞ k=0 x(k)δ(n − k). The mathematical way of representing impulse sam-
pling is a powerful tool that is used to analyze the performance of discrete-time
systems, and the values of the impulse functions at the output are obtained by
analytical methods. These values are identified as the numerical values of the
output signal.
Since h(n) is the response due to the input δ(n), we have x(0)h(n) as the
response due to x(0)δ(n) because we have assumed that the system is linear.
Assuming that the system is time-invariant as well as linear, we get the output
due to an input x(1)δ(n − 1) to be x(1)h(n − 1). In general, the output due to
an input x(k)δ(n − k) is given by x(k)h(n − k). Adding the responses due to all
the impulses in x(n) = ∞ k=0 x(k)δ(n − k), we get the total output as the sum

∞

y(n) = x(k)h(n − k) (2.5)
k=0
 A LINEAR, TIME-INVARIANT SYSTEM 39

This is known as the convolution sum, denoted by a compact notation y(n) =

x(n) ∗ h(n). The summation formula can be used to find the response due to any
input signal. So if we know the unit impulse response h(n) of the system, we
can find the output y(n) due to any input x(n)—therefore it is another model
for the discrete-time system. In contrast to the recursive algorithm, however,
note that the convolution sum cannot be used to find the response due to given
initial conditions. When and if the input signal is defined for −∞ < n < ∞ or
−M ≤ n < ∞, obviously the lower index of summation is changed to −∞. In
this case the convolution sum formula takes the general form
∞

y(n) = x(k)h(n − k) (2.6)
k=−∞

For example, even though we know that h(n) = 0 for −∞ < n < 0, if the input
sequencex(n) is defined for −M < n < ∞, then we have to use the formula
y(n) = ∞  − k). If x(n) = 0 for −∞ < n < 0, then we have to use
k=−∞ x(k)h(n
the formula y(n) = ∞ k=0 x(k)h(n − k).
To understand the procedure for implementing the summation formula, we
choose a graphical method in the following example. Remember that the recur-
sive algorithm cannot be used if the DT system is described by more than one
difference equation, and the convolution sum requires that we have the unit pulse
response of the system. We will find that these limitations are not present when
we use the z-transform method for analyzing the DT system performance in the
time domain.

Example 2.1

Given an h(n) and x(n), we change the independent variable from n to k and
plot h(k) and x(k) as shown in Figure 2.4a,b. Note that the input sequence is
defined for −2 ≤ k ≤ 5 but h(k) is a causal sequence defined for 0 ≤ k ≤ 4. Next
we do a time reversal and plot h(−k) in Figure 2.4c. When n ≥ 0, we obtain
h(n − k) by delaying (or shifting to the right) h(−k) by n samples; when n < 0,
the sequence h(−k) is advanced (or shifted to the left). For every value of n,
we have h(n − k) and x(k) and we multiply the samples of h(n − k) and x(k)
at each value of k and add the products.
For our example, we show the summation of the product when n = −2 in
Figure 2.4d, and show the summation of the product when n = 3 in Figure 2.4e.
The output y(−2) has only one nonzero product = x(−2)h(0). But the output
sample y(3) is equal to x(0)h(3) + x(1)h(2) + x(2)h(1) + x(3)h(0).
But note that when n > 9, and n < −2, the sequences h(n − k) and x(k) do
not have overlapping samples, and therefore y(n) = 0 for n > 9 and n < −2.

Example 2.2

As another example, let us assume that the input sequence x(n) and also the
unit impulse response h(n) are given for 0 ≤ n < ∞. Then output y(n) given
40 TIME-DOMAIN ANALYSIS AND z TRANSFORM

X(k) h(k)

−2 −1 0 1 2 3 4 5 k 0 1 2 3 4 k

(a) (b)
h(−2 − k)
h(− k)

−6 −5−4 −3−2 −1 0 1 2 k
−4 −3 −2 −1 0 k
(d) (c)
h(3 − k)

−1 0 1 2 3 k

(e)
h(10 − k)

0 1 2 3 4 5 6 7 8 9 10 k

(f )

Figure 2.4 Convolution sum explained.

by (2.5) can be computed for each value of n as shown below:

y(0) = x(0)h(0)
y(1) = x(0)h(1) + x(1)h(0)
y(2) = x(0)h(2) + x(1)h(1) + x(2)h(0)
y(3) = x(0)h(3) + x(1)h(2) + x(2)h(1) + x(3)h(0)
y(4) = x(0)h(4) + x(1)h(3) + x(2)h(2) + x(3)h(1) + x(4)h(0)
·
·
·
y(n) = x(0)h(n)+x(1)h(n − 1)+x(2)h(n − 2)+x(3)h(n − 3)+· · ·+x(n)h(0)
· (2.7)
·
·
 z TRANSFORM THEORY 41

It is interesting to note the following pattern. In the expressions for each

value of the output y(n) above, we have x(0), x(1), x(2) . . . and h(n), h(n − 1),
h(n − 2) . . . multiplied term by term in order and the products are added, while
the indices of the two samples in each product always add to n.
Convolution is a fundamental operation carried out by digital signal processors
in hardware and in the processing of digital signals by software. The design of
digital signal processors and the software to implement the convolution sum
have been developed to provide us with very efficient and powerful tools. We
will discuss this subject again in Section 2.5, after we learn the theory and
application of z transforms.

2.2 z TRANSFORM THEORY

2.2.1 Definition
In many textbooks, the z transform of a sequence x(n) is simply defined as
∞

Z[x(n)] = X(z) = x(n)z−n (2.8)
n=−∞

and the inverse z transform defined as

1
Z−1 [X(z)] = x(n) = X(z)zn−1 dz (2.9)
2πj C

Equation (2.8) represents the (double-sided or) bilateral z transform of a sequence

x(n) defined for −∞ < n < ∞. The inverse z transform given in (2.9) is obtained
by an integration in the complex z plane, and this integration in the z plane is
beyond the scope of this book.
We prefer to consider signals that are of interest in digital signal processing
and hence consider a sequence obtained by sampling a continuous-time signal
x(t) with a constant sampling period T (where T is the sampling period), and
generate a sequence of numbers x(nT ). Remember that according to the sifting
theorem, we have x(t)δ(t) = x(0)δ(t). We use this result to carry out a proce-
dure called impulse sampling by multiplying x(t) with an impulse train p(t) =
 ∞
n=0 δ(t − nT ). Consequently we consider a sequence of delayed impulse func-
tions weighted by the strength equal to the numerical values of the signal instead
of a sequence of numbers. By doing so, we express the discrete sequence as a
function of the continuous variable t, which allows us to treat signal processing
mathematically. The product is denoted as
∞

∗
x (t) = x(t)δ(t − nT )
n=0
∞

= x(nT )δ(t − nT ) (2.10)
n=0
42 TIME-DOMAIN ANALYSIS AND z TRANSFORM

This expression has a Laplace transform denoted as

∞

∗
X (s) = x(nT )e−snT (2.11)
n=0

Now we use a frequency transformation esT = z, (where z is a complex variable),

and substituting it in expression (2.11), we get
∞


X∗ (s)esT =z = x(nT )z−n
n=0

Since T is a constant, we consider the samples x(nT ) as a function of n and

obtain the z transform of x(n) as
∞


∗ 
X (s) esT =z = x(nT )z−n
n=0
∞

X(z) = x(n)z−n (2.12)
n=0

Although the first definition of a discrete sequence given in (2.8) is devoid of

any signal concepts, soon concepts such as frequency response and time-domain
response are used in the analysis of discrete-time systems and signal processing.
Our derivation of the z transform starts with a continuous-time signal that is sam-
pled by impulse sampling and introduces the transformation esT = z to arrive at
the same definition. In Chapter 3, we will study the implication of this transfor-
mation in more detail and get a fundamental understanding of the relationship
between the frequency responses of the continuous-time systems and those of
the discrete-time systems. Note that we consider in this book only the unilateral
z transform as defined by (2.12), so we set the lower index in the infinite sum
as n = 0.

Example 2.3

Let us derive the z transform of a few familiar discrete-time sequences. Consider

the unit pulse

1 n=0
δ(n) =
0 n = 0

There is only term in the z transform of δ(n), which is one when n = 0. Hence
Z [δ(n)] = 1.
 z TRANSFORM THEORY 43

Example 2.4

Consider the unit sample sequence u(n)


1 for n ≥ 0
u(n) = (2.13)
0 for n < 0

From the definition of the z transform, we get

Z[u(n)] = 1 + z−1 + z−2 + z−3 + · · · (2.14)

∞

= z−n (2.15)
n=0

This is an infinite series that converges to a closed-form expression (2.16), only

when z−1  < 1, or |z| > 1. This represents the region outside the unit circle in
the z plane and it is called the region of convergence (ROC). This means that
the closed-form expression exists only for values of z that lie in this region:
∞
1 z
z−n = = (2.16)
1 − z−1 z−1
n=0

It is obvious that the region of convergence for the z transform of δ(n) is the
entire z plane.

Example 2.5

Let x(n) = α n u(n), where α is assumed to be a complex number in general.

From the definition for the z transform, we obtain
∞
∞
 −1 n
X(z) = α n z−n = αz (2.17)
n=0 n=0
 
This power series converges to (2.18), when αz−1  < 1, that is, when |z| > |α|.
This shows that the region of convergence for the power series is outside the
circle of radius R = α. It is important to know the region of convergence in
which the closed-form expression for the z transform of a sequence of infinite
length is valid.1
1 z
X(z) = = (2.18)
1 − αz−1 z−α

1
It can be shown that the z transform for the anticausal sequence f (n) = −α n u(−n − 1) is F (z) =
−1 n −n
n=−∞ α z , which also converges to z/(z − α), which is the same as X(z) in (2.18), but its ROC
is |z| < α. So the inverse z transform of a function is not unique; only when we know its ROC does
the inverse z transform become unambiguous.
44 TIME-DOMAIN ANALYSIS AND z TRANSFORM

Example 2.6

Let us consider another example, x(n) = ej θn u(n), which is a complex-valued

sequence. Its z transform is
∞
1 z
X(z) = ej θn z−n = −1
= (2.19)
1−e z
j θ z − ej θ
n=0

and its region of convergence is the region outside the unit circle in the z plane:
|z| > 1.

Example 2.7

Given a sequence x(n) = r n cos(θ n)u(n), where 0 < r ≤ 1, to derive its z trans-
form, we express it as follows:
 
ej θn + e−j θn
x(n) = r n
u(n)
2
 n j θn 
r e r n e−j θn
= + u(n)
2 2
 j θ n  −j θ n
re re
= u(n) + u(n) (2.20)
2 2
Now one can use the previous results and obtain the z transform of x(n) =
r n cos(θ n)u(n) as
z(z − r cos(θ ))
X(z) = (2.21)
z2 − (2r cos(θ ))z + r 2

and its region of convergence is given by |z| > r. Of course, if the sequence
given is x(n) = e−an cos(ω0 n)u(n), we simply substitute e−a for r in (2.21),
to get the z transform of x(n). It is useful to have a list of z transforms for
discrete-time sequences that are commonly utilized; they are listed in Table 2.1.
It is also useful to know the properties of z transforms that can be used to
generate and add more z transforms to Table 2.1, as illustrated by the following
example.

Property 2.1: Differentiation If X(z) is the z transform of x(n)u(n),

−z[dX(z)]/dz is the z transform of nx(n)u(n). We denote this property by

dX(z)
nx(n)u(n) ⇐⇒ −z (2.22)
dz
 z TRANSFORM THEORY 45

TABLE 2.1 List of z-Transform Pairs

x(n), for n ≥ 0 X(z)

1 δ(n) 1
2 δ(n − m) z−m
z
3 u(n)
z−1
az
4 au(n)
z−1
z
5 an
z−a
az
6 na n
(z − a)2
z(z + 1)
7 n2
(z − 1)3
z(z2 + 4z + 1)
8 n3
(z − 1)4
az(z + a)
9 n2 a n
(z − a)3
n(n − 1) n−2 z
10 a
2! (z − a)3
n(n − 1)(n − 2) · · · · · · · (n − m + 2) n−m+1 z
11 a
(m − 1)! (z − a)m
z
12 r n ej θn
z − rej θ
z(z − r cos(θ ))
13 r n cos(θn)
z2 − (2r cos(θ ))z + r 2
rz sin(θ )
14 r n sin(θn)
z2 − (2r cos(θ ))z + r 2
z(z − e−α cos(θ ))
15 e−αn cos(θn)
z2 − (2e−α cos(θ ))z + e−2α

∞
Proof : X(z) = n=0 x(n)z−n . Differentiating both sides with respect to z,
we get
∞ ∞
dX(z)  
= x(n) −nz−n−1 = −z−1 nx(n)z−n
dz
n=0 n=0
∞

dX(z)
−z = nx(n)z−n = Z[nx(n)u(n)]
dz
n=0
46 TIME-DOMAIN ANALYSIS AND z TRANSFORM

Now consider the z transform given by (2.18) and also listed in Table 2.1:
z
x(n) = a n u(n) ⇐⇒ = X(z) (2.23)
z−a
Using this differentiation property recursively, we can show that
az
na n u(n) ⇐⇒ (2.24)
(z − a)2
and
az(z + a)
n2 a n u(n) ⇐⇒ (2.25)
(z − a)3

From these results, we can find the z transform of 1

2 (n + 1)(n + 2)a n u(n) =
2 (n + 3n + 2)a u(n) as follows:
1 2 n

1 z3
(n + 1)(n + 2)a n u(n) ⇐⇒ (2.26)
2 (z − a)3
The transform pair given by (2.26) is an addition to Table 2.1. Indeed, we can
find the z transforms of n3 a n u(n), n4 a n u(n), . . . , using (2.22) and then find the
z transforms of
1
(n + 1)(n + 2)(n + 3)a n u(n)
3!
1
(n + 1)(n + 2)(n + 3)(n + 4)a n u(n) (2.27)
4!
·
·
·

which can be added to Table 2.1.

Properties of z transform are useful for deriving the z transform of new
sequences. Also they are essential for solving the linear difference equations and
finding the response of discrete-time systems when the input function and initial
conditions are given. Instead of deriving all the properties one after another, as is
done in many textbooks, we derive one or two at a time and immediately show
their applications.

Property 2.2: Delay Let the z transform of x(n)u(n) be X(z) = ∞ n=0 x(n)z
−n
=
−1 −2 −3
x(0) + x(1)z + x(2)z + x(3)z + · · · .

Then the z transform of x(n − 1)u(n − 1) is z−1 X(z) + x(−1):

x(n − 1)u(n − 1) ⇐⇒ z−1 X(z) + x(−1) (2.28)

 z TRANSFORM THEORY 47

Proof : The z transform of x(n − 1)u(n − 1) is obtained by shifting to the

right or delaying x(n)u(n) by one sample, and if there is a sample x(−1) at
n = −1, it will be shifted to the position n = 0. The z transform of this delayed
sequence is therefore given by

x(−1) + x(0)z−1 + x(1)z−2 + x(2)z−3 + · · ·

 
= x(−1) + z−1 x(0) + x(1)z−1 + x(2)z−2 + · · ·
= x(−1) + z−1 X(z)

By repeated application of this property, we derive

x(n − 2)u(n − 2) ⇐⇒ z−2 X(z) + z−1 x(−1) + x(−2) (2.29)

x(n − 3)u(n − 3) ⇐⇒ z−3 X(z) + z−2 x(−1) + z−1 x(−2) + x(−3) (2.30)

and

x(n − m)u(n − m) ⇐⇒ z−m X(z−1 ) + z−m+1 x(−1) + z−m+2 x(−2)

+ · · · + x(−m)

or

m−1
−m
x(n − m)u(n − m) ⇐⇒ z X(z) + x(n − m)z−n (2.31)
n=0

If the initial conditions are zero, we have the simpler relationship

x(n − m)u(n − m) ⇐⇒ z−m X(z) (2.32)

Example 2.8

Let us consider an example of solving a first-order linear difference equation

using the results obtained above. We have

y(n) − 0.5y(n − 1) = 5x(n − 1) (2.33)

where
x(n) = (0.2)n u(n)
y(−1) = 2

Let Z[y(n)] = Y (z). From (2.28), we have Z[y(n − 1)] = z−1 Y (z) + y(−1)
and Z[x(n − 1)] = z−1 X(z) + x(−1) where X(z) = z/(z − 0.2) and x(−1) = 0,
since x(n) is zero for −∞ < n < 0. Substituting these results, we get
   
Y (z) − 0.5 z−1 Y (z) + y(−1) = 5 z−1 X(z) + x(−1)
 
Y (z) − 0.5 z−1 Y (z) + y(−1) = 5z−1 X(z)
48 TIME-DOMAIN ANALYSIS AND z TRANSFORM

 
Y (z) 1 − 0.5z−1 = 0.5y(−1) + 5z−1 X(z) (2.34)
0.5y(−1) 5z−1
Y (z) = + X(z)
(1 − 0.5z−1 ) (1 − 0.5z−1 )
0.5y(−1)z 5
Y (z) = + X(z)
(z − 0.5) (z − 0.5)
Substituting y(−1) = 2 and X(z) = z/(z − 0.2) in this last expression, we get

z 5z
Y (z) = + (2.35)
(z − 0.5) (z − 0.5)(z − 0.2)
= Y0i (z) + Y0s (z) (2.36)

where Y0i (z) is the z transform of the zero input response and Y0s (z) is the
z transform of the zero state response as explained below.
Now we have to find the inverse z transform of the two terms on the right
side of (2.35). The inverse transform of the first term Y0i (z) = z/(z − 0.5) is
easily found as y0i (n) = (0.5)n u(n). Instead of finding the inverse z transform
of the second term by using the complex integral given in (2.9), we resort to
the same approach as used in solving differential equations by means of Laplace
transform, namely, by decomposing Y0s (z) into its partial fraction form to obtain
the inverse z transform of each term. We have already derived the z transform of
Ra n u(n) as Rz/(z − a), and it is easy to write the inverse z transform of terms
like Rk z/(z − ak ). Hence we should expand the second term in the form
R1 z R2 z
Y0s (z) = + (2.37)
z − 0.5 z − 0.2
by a slight modification to the partial fraction expansion procedure that we are
familiar with. Dividing Y0s (z) by z, we get

Y0s (z) 5 R1 R2
= = +
z (z − 0.5)(z − 0.2) z − 0.5 z − 0.2
Now we can easily find the residues R1 and R2 using the normal procedure and
get
 
Y0s (z)  5 
R1 = (z − 0.5) =  = 16.666
z z=0.5 (z − 0.2) z=0.5
 
Y0s (z)  5 
R2 = (z − 0.2) =  = −16.666
z z=0.2 (z − 0.5) z=0.2

Therefore
Y0s (z) 5 16.666 16.666
= = −
z (z − 0.5)(z − 0.2) z − 0.5 z − 0.2
 z TRANSFORM THEORY 49

Multiplying both sides by z, we get

16.666z 16.666z
Y0s (z) = − (2.38)
z − 0.5 z − 0.2
Now we obtain the inverse z transform y0s (n) = 16.666[(0.5)n − (0.2)n ]u(n).
The total output satisfying the given difference equation is therefore given as
 
y(n) = y0i (n) + y0s (n) = (0.5)n + 16.666[(0.5)n − (0.2)n ] u(n)
 
= 17.6666(0.5)n − 16.666(0.2)n u(n)

Thus the modified partial fraction procedure to find the inverse z transform of
any function F (z) is to divide the function F (z) by z, expand F (z)/z into its
 form, and then multiply each of the terms by z to get
normal partial fraction
F (z) in the form  k=1 Rk z/(z − ak ). From this form, the inverse z transform
f (n) is obtained as k=1 Rk (ak )n u(n).
However, there is an alternative method, to expand a transfer function expressed
in the form, when it has only simple poles

N (z−1 )
H (z−1 ) =  −1
k=1 (1 − ak z )

to its partial fraction form

Rk
H (z−1 ) = (2.39)
(1 − ak z−1 )
k=1

where

Rk = H (z−1 )(1 − ak z−1 )z=a
k

Then the inverse z transform is the sum of the inverse z transform of all the terms
in (2.39): h(n) = K n
k=1 Rk (ak ) u(n). We prefer the first method because we are
already familiar with the partial fraction expansion of H (s) and know how to
find the residues when it has multiple poles in the s plane. This method will be
illustrated by several examples that are worked out in the following pages.

2.2.2 Zero Input and Zero State Response

In Section 2.2.1, the total output y(n) was obtained as the sum of two outputs
y0i (n) = (0.5)n u(n) and y0s (n) = 16.666[(0.5)n − (0.2)n ]u(n).
If the input function x(n) is zero, then X(z) = 0, and Y (s) in (2.34) will
contain only the term Y0i (z) = 0.5y(−1)z/(z − 0.5) = z/(z − 0.5); therefore the
response y(n) = (0.5)n u(n) when the input is zero. The response of a system
described by a linear difference equation, when the input to the system is assumed
50 TIME-DOMAIN ANALYSIS AND z TRANSFORM

to be zero is called the zero input response and is determined only by the initial
conditions given. The initial conditions specified with the difference equation are
better known as initial states. (But the term state has a specific definition in the
theory of linear discrete-time systems, and the terminology of initial states is con-
sistent with this definition.) When the initial state y(−1) in the problem presented
above is assumed to be zero, the z transform of the total response Y (z) contains
only the term Y0s (z) = 5/(z − 0.5)X(z) = 5/[(z − 0.5)(z − 0.2)], which gives a
response y0s (n) = 16.666[(0.5)n − (0.2)n ]u(n). This is the response y(n) when
the initial condition or the initial state is zero and hence is called the zero state
response. The zero state response is the response due to input only, and the zero
input response is due to the initial states (initial conditions) only. We repeat it
in order to avoid the common confusion that occurs among students! The zero
input response is computed by neglecting the input function and computing the
response due to initial states only, and the zero state response is computed by
neglecting the initial states (if they are given) and computing the response due to
input function only. Students are advised to know the exact definition and mean-
ing of the zero input response and zero state response, without any confusion
between these two terms.

2.2.3 Linearity of the System

If the input x(n) to the discrete-time system described by (2.33) is multiplied
by a constant, say, K = 10, the total response of the system y(n) is given as
y0i (n) + 10y0s (n), which is not 10 times the total response y0i (n) + y0s (n). This
may give rise to the incorrect inference that the system described by the difference
equation (2.33) above, is not linear. The correct way to test whether a system
is linear is to apply the test on the zero state response only or to the zero input
response only as explained below.
Let the zero state response of a system defined by a difference equation be
y1 (n) when the input to the system is x1 (n) and the zero state response be y2 (n)
when the input is x2 (n), where the inputs are arbitrary. Here we emphasize that
the definition should be applied to the zero state response only or to the zero input
only. So the definition of a linear system given in Section 2.1 is repeated below,
emphasizing that the definition should be applied to the zero input response or
zero state response only.
Given a system x(n) ⇒ y(n), if Kx(n) ⇒ Ky(n) and K1 x1 (n) + K2 x2 (n) ⇒
K1 y1 (n) + K2 y2 (n), then the system is linear, provided y(n) is the zero state
response due to an input signal x(n) or the zero input response due to initial
states. Now it should be easy to verify that the system described by (2.33) is a
linear system.

2.2.4 Time-Invariant System

Let a discrete-time system be defined by a linear difference equation of the
general form (2.3), which defines the input–output relationship of the system.
 USING z TRANSFORM TO SOLVE DIFFERENCE EQUATIONS 51

Let us denote the solution to this equation as the output y(n) when an input x(n)
is applied. Such a system is said to be time-invariant if the output is y(n − N )
when the input is x(n − N ), which means that if the input sequence is delayed
by N samples, the output also is delayed by N samples. For this reason, a time-
invariant discrete-time system is also called a shift-invariant system. Again from
the preceding discussion about linearity of a system, it should be obvious the
output y(n) and y(n − N ) must be chosen as the zero state response only or
the zero input response only, when the abovementioned test for a system to be
time-invariant is applied.

2.3 USING z TRANSFORM TO SOLVE DIFFERENCE EQUATIONS

We will consider a few more examples to show how to solve a linear shift-
invariant difference equation, using the z transform in this section, and later we
show how to solve a single-input, single-output difference equation using the
classical method. Students should be familiar with the procedure for decompos-
ing a proper, rational function of a complex variable in its partial fraction form,
when the function has simple poles, multiple poles, and pairs of complex conju-
gate poles. A “rational” function in a complex variable is the ratio between two
polynomials with real coefficients, and a “proper” function is one in which the
degree of the numerator polynomial is less than that of the denominator polyno-
mial. It can be shown that the degree of the numerator in the transfer function
H (s) of a continuous-time system is at most equal to that of its denominator. In
contrast, it is relevant to point out that the transfer function of a discrete-time
system when expressed in terms of the variable z−1 need not be a proper func-
tion. For example, let us consider the following example of an improper function
of the complex variable z−1 :

z−4 − 0.8z−3 − 2.2z−2 − 0.4z−1

H (z−1 ) = (2.40)
z−2 − z−1 + 2.0
In this equation, the coefficients of the two polynomials are arranged in descend-
ing powers of z−1 , and when we carry out a long division of the numerator by
the denominator, until the remainder is a polynomial of a degree lower than that
in the denominator, we get the quotient (z−2 + 0.2z−1 − 4.0) and a remainder
(−4.8z−1 + 8.0):

−4.8z−1 + 8.0
H (z−1 ) = z−2 + 0.2z−1 − 4.0 + (2.41)
z−2 − z−1 + 2.0
= z−2 + 0.2z−1 − 4.0 + H1 (z−1 ) (2.42)

Since the inverse z transform of z−m is δ(n − m), we get the inverse z trans-
form of the first three terms as δ(n − 2) + 0.2δ(n − 1) − 4.0δ(n), and we add it
to the inverse z transform of the H1 (z−1 ), which will be derived below.
52 TIME-DOMAIN ANALYSIS AND z TRANSFORM

Example 2.9: Complex Conjugate Poles

Let us choose the second term on the right side of (2.41) as an example of a
transfer function with complex poles:

−4.8z−1 + 8.0
H1 (z−1 ) =
z−2 − z−1 + 2.0
Multiplying the numerator and denominator by z2 , and factorizing the denomina-
tor, we find that H1 (z−1 ) has a complex conjugate pair of poles at 0.25 ± j 0.6614:

8(z2 − 0.6z)
H1 (z) =
2z2 − z + 1
8(z2 − 0.6z)
=
2(z2 − 0.5z + 0.5)
(z2 − 0.6z)
=4
(z − 0.25 − j 0.6614)(z − 0.25 + j 06614)
Let us expand H1 (z)/z into its modified partial fraction form:
H1 (z) 2 + j 1.0583 2 − j 1.0583
= + .
z z − 0.25 − j 0.6614 z − 0.25 + j 0.6614
It is preferable to express the residues and the poles in their exponential form
and then multiply by z to get
 
2.2627ej 0.4867 z 2.2627e−j 0.4867 z
H1 (z) = +
z − 0.7071ej 1.209 z − 0.7071e−j 1.209
The inverse z transform of H1 (z) is given by

h1 (n) = 2.2627ej 0.4867 (0.7071ej 1.209 )n
 
+ 2.2627e−j 0.4867 (0.7071e−j 1.209 )n u(n)

= 2.2627(0.7071)n ej 1.209n ej 0.4867

+ 2.2627(0.7071)n e−j 1.209n e−j 0.4867 u(n)
 
= 2.2627(0.7071)n ej (1.209n+0.4867) + e−j (1.209n+0.4867) u(n)
= 2.2627(0.7071)n {2 cos(1.209n + 0.4867)} u(n)
= 4.5254(0.7071)n {cos(1.209n + 0.4867)} u(n).

Adding these terms, we get the inverse z transform of

 
−1 −2 −1 2.2627ej 0.4867 z 2.2627e−j 0.4867 z
H (z ) = z + 0.2z − 4.0 + +
z − 0.7071ej 1.209 z − 0.7071e−j 1.209
 USING z TRANSFORM TO SOLVE DIFFERENCE EQUATIONS 53

h(n) = δ(n − 2) + 0.2δ(n − 1) − 4.0δ(n) + 2.2627ej 0.4867 (0.7071ej 1.209 )n

+ 2.2627ej −0.4867 (0.7071e−j 1.209 )n u(n)
= δ(n − 2) + 0.2δ(n − 1) − 4.0δ(n) + 2.2627(0.7071)n
× {ej (1.209n+0.4867) + e−j (1.209n+0.4867) }u(n)
= δ(n − 2) + 0.2δ(n − 1) − 4.0δ(n) + 2.2627(0.7071)n
× {2 cos(1.209n + 0.4867)}u(n)
= δ(n − 2) + 0.2δ(n − 1) − 4.0δ(n)
+ 4.5254(0.7071)n {cos(1.209n + 0.4867)} u(n)

Note that the angles in this solution are expressed in radians.

Example 2.10

Let us consider a discrete-time system described by the linear shift-invariant

difference equation of second order given below

y(n) = 0.3y(n − 1) − 0.02y(n − 2) + x(n) − 0.1x(n − 1)

where

x(n) = (−0.2)n u(n)

y(−1) = 1.0
y(−2) = 0.6

Using the z transform for each term in this difference equation, we get

Y (z) = 0.3[z−1 Y (z) + y(−1)] − 0.02[z−2 Y (z) + z−1 y(−1) + y(−2)]

+ X(z) − 0.1[z−1 X(z) + x(−1)]

We know X(z) = z/(z + 0.2) and x(−1) = 0. Substituting these and the given
initial states, we get

Y (z)[1 − 0.3z−1 + 0.02z−2 ] = [0.3y(−1) − 0.02z−1 y(−1) − 0.02y(−2)]

+ X(z)[1 − 0.1z−1 ]
[0.3y(−1) − 0.02z−1 y(−1) − 0.02y(−2)]
Y (z) =
[1 − 0.3z−1 + 0.02z−2 ]
X(z)[1 − 0.1z−1 ]
+
[1 − 0.3z−1 + 0.02z−2 ]
54 TIME-DOMAIN ANALYSIS AND z TRANSFORM

When the input x(n) is zero, X(z) = 0; hence the second term on the right side
is zero, leaving only the first term due to initial conditions given. It is the z
transform of the zero input response y0i (n).
The inverse z transform of this first term on the right side

[0.3y(−1) − 0.02z−1 y(−1) − 0.02y(−2)]

Y0i (z) =
[1 − 0.3z−1 + 0.02z−2 ]
gives the response when the input is zero, and so it is the zero input response
y0i (n). The inverse z transform of

X(z)[1 − 0.1z−1 ]
= Y0s (z)
[1 − 0.3z−1 + 0.02z−2 ]
gives the response when the initial conditions (also called the initial states) are
zero, and hence it is the zero state response y0s (n).
Substituting the values of the initial states and for X(z), we obtain
[0.288 − 0.02z−1 ] [0.288z2 − 0.02z] z[0.288z − 0.02]
Y0i (z) = −1 −2
= =
[1 − 0.3z + 0.02z ] z − 0.3z + 0.02
2 (z − 0.1)(z − 0.2)

and
 
X(z)[1 − 0.1z−1 ] z [1 − 0.1z−1 ]
Y0s (z) = =
[1 − 0.3z−1 + 0.02z−2 ] z + 0.2 [1 − 0.3z−1 + 0.02z−2 ]
z[z2 − 0.1z] z2 (z − 0.1)
= =
(z + 0.2)(z2 − 0.3z + 0.02) (z + 0.2)(z − 0.1)(z − 0.2)
We notice that there is a pole and a zero at z = 0.1 in the second term on the
right, which cancel each other, and Y0s (z) reduces to z2 /[(z + 0.2)(z − 0.2)]. We
divide Y0i (z) by z, expand it into its normal partial fraction form
Y0i (z) [0.288z − 0.02] 0.376 0.088
= = −
z (z − 0.1)(z − 0.2) (z − 0.2) (z − 0.1)
and multiply by z to get
0.376z 0.088z
Y0i (z) = −
(z − 0.2) (z − 0.1)
Similarly, we expand Y0s (z)/z = z/[(z + 0.2)(z − 0.2)] in the form −0.5/(z +
0.2) + 0.5/(z − 0.2) and get

z2 −0.5z 0.5z
Y0s (z) = = +
(z + 0.2)(z − 0.2) (z + 0.2) (z − 0.2)
Therefore, the zero input response is y0i (n) = [0.376(0.2)n − 0.088(0.1)n ]u(n)
and the zero state response is y0s (n) = 0.5[−(−0.2)n + (0.2)n ]u(n).
 USING z TRANSFORM TO SOLVE DIFFERENCE EQUATIONS 55

Example 2.11: Multiple Poles

Here we discuss the case of a function that has multiple poles and expand it into
its partial fraction form. Let
G(z) N (z)
=
z (z − z0 ) (z − z1 )(z − z2 )(z − z3 ) · · · (z − zm )
r

Its normal partial fraction is in the form

G(z) C0 C1 Cr−1
= + + ··· +
z (z − z0 )r (z − z0 )r−1 (z − z0 )
k1 k2 km
+ + + ··· +
(z − z1 ) (z − z2 ) (z − zm )
The residues k1 , k2 , . . . , km for the simple poles at z1 , z2 , . . . are obtained by
the normal method of multiplying G(z)/z by (z − zi ), i = 1, 2, 3, . . . , m and
evaluating the product at z = zi . The residue C0 is also found by the same
method:
 
r G(z) 

C0 = (z − z0 ) 
z z=z0

The coefficient C1 is found from

 
d 
r G(z) 
(z − z0 ) 
dz z z=z0

and the coefficient C2 is found from

 
1 d2 
r G(z) 
2
(z − z0 ) 
2 dz z z=z0

The general formula for finding the coefficients Cj , j = 1, 2, 3, . . . , (r − 1) is

 
1 dj 
r G(z) 
Cj = j
(z − z0 )  (2.43)
j ! dz z z=z0

After obtaining the residues and the coefficients, we multiply the expansion by z:
C0 z C1 z Cr−1 z
G(z) = + + ··· +
(z − z0 )r (z − z0 )r−1 (z − z0 )
k1 z k2 z km z
+ + + ··· +
(z − z1 ) (z − z2 ) (z − zm )
Then we find the inverse z transform of each term to get g(n), using the z
transform pairs given in Table 2.1. To illustrate this method, we consider the
56 TIME-DOMAIN ANALYSIS AND z TRANSFORM

function G(z), which has a simple pole at z = 1 and a triple pole at z = 2:

z(2z2 − 11z + 12)

G(z) = (2.44)
(z − 1)(z − 2)3
G(z) (2z2 − 11z + 12) C0 C1 C2 k
= = + + +
z (z − 1)(z − 2) 3 (z − 2) 3 (z − 2)2 (z − 2) (z − 1)

(2z − 11z + 12) 
2
k=  = −3
(z − 2)3 z=1

(2z2 − 11z + 12) 
C0 =  = −2
(z − 1) z=2
   2 
d (2z2 − 11z + 12)  2z − 4z − 1 
C1 =  =  = −1
dz (z − 1) z=2 (z − 1)2 z=2
   
1 d 2 (2z2 − 11z + 12)  1 d 2z2 − 4z − 1 
C2 =  =  =3
2 dz2 (z − 1) z=2 2 dz (z − 1)2 z=2

Therefore we have
−2z −z 3z −3z
G(z) = + + + (2.45)
(z − 2)3 (z − 2)2 (z − 2) (z − 1)

Now note that the inverse z transform of az/(z − a)2 is easily obtained from
Table 2.1, as na n u(n). We now have to reduce the term −z/(z − 2)2 to −( 12 )2z/
(z − 2)2 so that its inverse z transform is correctly written as −( 12 )n2n u(n). From
the transform pair 6 in Table 2.1, we get the inverse z transform of z/(z − a)3
as n(n − 1)/2!a n−2 u(n).
Therefore the inverse z transform of −2z/(z − 2)3 is obtained as

−2n(n − 1) n−2 −n(n − 1) n

(2) u(n) = (2) u(n)
2! 4
Finally, we get the inverse z transform of G(z) as
 
−n(n − 1) n n n
(2) − (2) − 3(2) − 3 u(n)
n
4 2

2.3.1 More Applications of z Transform

In this section, we consider the circuit shown in Figure 2.5 and model it by
equations in the z domain, instead of the equivalent model given by equations
(2.1) in the time domain. This example is chosen to illustrate the analysis of a
discrete-time system that has a large number of adders and hence gives rise to
a large number of difference equations in the z domain. Writing the z transform
 USING z TRANSFORM TO SOLVE DIFFERENCE EQUATIONS 57

0.8

−0.1 y2(z) y3(z)

y1(z)
X(z) z−1 Σ Σ Σ

z−1
z−1
0.5

0.3 −1
−0.4 z y2(z) 0.6
z−1

z−2y1(z)
−0.2

Figure 2.5 A discrete-time system.

for each of the equations in (2.1), we get

Y1 (z) = 0.3[z−1 Y1 (z) + y1 (−1)] − 0.2[z−2 Y1 (z) + z−1 y1 (−1) + y1 (−2)]

− 0.1z−1 X(z)
Y2 (z) = Y1 (z) + 0.5[z−1 Y1 (z) + y1 (−1)] − 0.4[z−1 Y2 (z) + y2 (−1)]
Y3 (z) = Y2 (z) + 0.6[z−1 Y2 (z) + y2 (−1)] + 0.8Y1 (z) (2.46)

Note that these are linear algebraic equations—three equations in three unknown
functions Y1 (z), Y2 (z), and Y3 (z), where the initial states and X(z) are known.
After rearranging these equations as follows

Y1 (z)[1 − 0.3z−1 + 0.2z−2 ] = 0.3y1 (−1) − 0.2z−1 y1 (−1)

+ y1 (−2) − 0.1z−1 X(z)
Y1 (z)[−1 − 0.5z−1 ] + Y2 (z)[1 + 0.4z−2 ] = 0.5y1 (−1) − 0.4y2 (−1)
0.8Y1 (z) + Y2 (z)[1 + 0.6z−1 ] − Y3 (z) = −0.6y2 (−1) (2.47)

we express them in a matrix form as

⎡ ⎤⎡ ⎤
(1 − 0.3z−1 + 0.2z−2 ) 0 0 Y1 (z)
⎣ (−1 − 0.5z−1 ) (1 + 0.4z−2 ) 0 ⎦ ⎣ Y2 (z) ⎦
0.8 −1
(1 + 0.6z ) −1 Y3 (z)
⎡ ⎤
0.3y1 (−1) − 0.2z−1 y1 (−1) + y1 (−2) − 0.1z−1 X(z)
= ⎣ 0.5y1 (−1) − 0.4y2 (−1) ⎦ (2.48)
−0.6y2 (−1)
58 TIME-DOMAIN ANALYSIS AND z TRANSFORM

By use of matrix algebra, we can now find any one or all three unknown functions
Y1 (z), Y2 (z), and Y3 (z), when the input X(z) is zero—their inverse z transforms
yield zero input responses. We can find them when all the initial states are
zero—their inverse z transform will yield zero state responses. Of course we
can find the total responses y1 (n), y2 (n), and y3 (n), under the given initial states
and the input function x(n). This outlines a powerful algebraic method for the
analysis of discrete-time systems described by any large number of equations in
either the discrete-time domain or the z-transform domain. We use this method
to find the zero input response and the zero state response and their sum, which
is the total response denoted as y1 (n), y2 (n), and y3 (n).

2.3.2 Natural Response and Forced Response

It is to be pointed out that the total response can also be expressed as the sum
of the natural response and forced response of the system. First let us make
it clear that the natural response is not the same as the zero input response
of the system. The natural response is defined as the component of the total
response, which consists of all terms displaying the natural frequencies of the
system. Natural frequencies are also known as the “characteristic roots” of the
system, eigenvalues of the system determinant, and poles of the transfer
function.
A few methods are used to find the natural frequencies of a system. Suppose
that the system is described by its single input–output relationship as a0 y(n) +
a1 y(n − 1) + a2 y(n − 2) + · · · + aN y(n − N ) = b0 x(n) + b1 x(n − 1) + · · · +
bM x(n − M).
If we assume the solution to the homogeneous equation to be of the form
yc (n) = A(c)n , and substitute it as well as its delayed sequences, we get the
following characteristic equation:

A(c)n [a0 + a1 (c)−1 + a2 (c)−2 + · · · + aN (c)−N ]

= A(c)n−N [a0 (c)N + a1 (c)N −1 + · · · + aN −1 (c) + aN ] = 0.

Let the N roots of the characteristic polynomial

[a0 (c)N + a1 (c)N −1 + · · · + aN −1 (c) + aN ]

be denoted by (c1 ), (c2 ), . . . , (cN ), which are the natural frequencies. Assuming
that all the roots are distinct and separate, the natural response assumes the form

yc (n) = A1 (c1 )n + A2 (c2 )n + · · · + AN (cN )n

which in classical literature is known as the complementary function or comple-

mentary solution.
 SOLVING DIFFERENCE EQUATIONS USING THE CLASSICAL METHOD 59

If, however, the characteristic polynomial has a repeated root (cr ) with a mul-
tiplicity of R, then the R terms in yc (n) corresponding to this natural frequency
(cr ) are assumed to be of the form

[B0 + B1 n + B2 n2 + · · · + BR nR ](cr )n

Suppose that the system is described by a set of linear difference equations such
as (2.47). When we solve for Y1 (z), Y2 (z), or Y3 (z), we get the determinant of
the system matrix shown in (2.48) as the denominator in Y1 (z), Y2 (z), and Y3 (z).
The roots of this system determinant are the poles of the z transform Y1 (z),
Y2 (z), Y3 (z) and appear in the partial fraction expansion for these functions. The
inverse z transform of each term in the partial fraction expansion will exhibit
the corresponding natural frequency. All terms containing the natural frequencies
make up the natural response of the system. The important observation to be made
is that terms with these natural frequencies appear in the zero input response as
well as the zero state response; hence the amplitudes Aj of the term with the
natural frequency (cj ) have to be computed as the sum of terms with the natural
frequency (cj ) found in both the zero input and zero state response. It follows,
therefore, that the natural response is not the same as the zero input response
and the forced response is not the same as the zero state response. Computation
of these different components in the total response must be carried out by using
the correct definition of these terms.

2.4 SOLVING DIFFERENCE EQUATIONS USING

THE CLASSICAL METHOD

Now that we have described the method for finding the complementary func-
tion for a system described by an nth linear ordinary difference equation, we
discuss the computation of the particular function or particular solution, due to
the specified input function. Note that this classical method can be used when
there is only one such equation, and it is not very easy when there are many
equations describing the given discrete-time system. Also, when the order of the
characteristic polynomial or the system determinant is more than 3, finding the
zeros of the characteristic polynomial or the system determinant analytically is
not possible. We have to use numerical techniques to find these zeros, which are
the natural frequencies of the system. If and when we have found the natural
frequencies, the natural response can be identified as the function yc (n) given in
the preceding section. Next we have to choose the form of the particular function
that depends on the form of the input or the forcing function. Hence it is the
forced response, and the sum of the natural response (complementary function)
and the forced response (particular function) is the total response. The form of
the particular function is chosen as listed in Table 2.2.
We substitute the particular function in the nonhomogeneous difference
equation, and by comparing the coefficients on both sides of the resulting
60 TIME-DOMAIN ANALYSIS AND z TRANSFORM

TABLE 2.2 Form of Input Function and Forced Response

Input or Forcing Function Particular Function or Forced Response

1 A(α)n , α = c1 (i = 1, 2, . . .) B(α)n
2 A(α)n , α = ci [B0 + B1 n] (α)n
3 A cos(ω0 n + θ ) B cos(ω0 n + φ)
 m i
n  m i
n
4 i=0 Ai n α i=0 Bi n α

equation, we compute the coefficient B or the coefficients Bi . Next we apply

the given initial conditions on the sum of the complementary function and the
particular function, in which there are n unknown constants of the complemen-
tary function. When we obtain these constants that satisfy the initial conditions,
and substitute them, the solution for the total output is complete. Example 2.12
illustrates the classical method of solving a difference equation.

Example 2.12

Solve the linear difference equation given below, using the classical method:

y(n) − 0.5y(n − 1) + 0.06y(n − 2) = 2(0.1)n (2.49)

y(−1) = 1 and y(−2) = 0 (2.50)

The characteristic polynomial is z2 − 0.5z + 0.06 = (z − 0.3)(z − 0.2), which

has the characteristic roots at z1 = 0.3 and z2 = 0.2. Since these are simple
zeros, the complementary function yc (n) = A1 (0.3)n + A2 (0.2)n . Since the input
x(n) is given as 2(0.1)n , we choose from Table 2.2, the particular function yp to
be of the form yp (n) = B(0.1)n . Thus we substitute yp (n − 1) = B(0.1)n−1 and
yp (n − 2) = B(0.1)n−2 and get the following:

B(0.1)n − 0.5B(0.1)n−1 + 0.06B(0.1)n−2 = 2(0.1)n

B(0.1)n − 0.5(10)B(0.1)n + 0.06(100)B(0.1)n = 2(0.1)n
[1.0 − 5 + 6]B(0.1)n = 2(0.1)n

Therefore B = 1 and the particular function yp (n) = (0.1)n . So the total solu-
tion is

y(n) = yc (n) + yp (n)

= A1 (0.3)n + A2 (0.2)n + (0.1)n

When we apply the initial conditions on this total response, we get

y(−1) = 3.3333A1 + 5A2 + 10 = 1

y(−2) = 11.111A1 + 25A2 + 100 = 0
 SOLVING DIFFERENCE EQUATIONS USING THE CLASSICAL METHOD 61

Solving these two equations, we get A1 = 9.903 and A2 = −8.4. So the total
response is

y(n) = 9.903(0.3)n − 8.4(0.2)n + (0.1)n

Example 2.13

Let us reconsider Example 2.10. The zero input response and the zero state
response in this example were found to be

y0i (n) = [0.376(0.2)n − 0.088(0.1)n ]u(n)

y0s (n) = 0.5[−(−0.2)n + (0.2)n ]u(n) = [0.5(0.2)n − 0.5(−0.2)n ]u(n)

The characteristic polynomial for the system given in Example 2.10 is easily
seen as [1 − 0.3z−1 + 0.02z−2 ]. After multiplying it by z2 , we find the natu-
ral frequencies as the zeros of [z2 − 0.3z + 0.02] = [(z − 0.2)(z − 0.1)] to be
(c1 ) = (0.2) and (c2 ) = (0.1). Note that the zero input response y0i (n) has a
term 0.376(0.2)n u(n), which has the natural frequency equal to (0.2) and the
term −0.088(0.1)n u(n) with the natural frequency of (0.1), while the zero state
response y0s (n) also contains the term 0.5(0.2)n u(n) with the natural frequency
of (0.2). We also noticed that the pole of Y0s (z) at z = 0.1 was canceled by a zero
at z = 0.1. Therefore there is no term in the zero state response y0s (n) with the
natural frequency of (0.1). So the term containing the natural frequency of (0.2) is
the sum 0.5(0.2)n u(n) + 0.376(0.2)n u(n) = 0.876(0.2)n u(n), whereas the other
term with the natural frequency of (0.1) is −0.088(0.1)n u(n). Consequently, the
natural response of the system is 0.876(0.2)n u(n) − 0.088(0.1)n u(n).
The remaining term −0.5(−0.2)n u(n) is the forced response with the fre-
quency (−0.2), which is found in the forcing function or the input function
x(n) = (−0.2)n u(n). Thus the total response of the system is now expressed as
the sum of its natural response 0.876(0.2)n u(n) − 0.088(0.1)n u(n) and forced
response −0.5(−0.2)n u(n). We repeat that in the zero state response, there are
terms with natural frequencies of the system, besides terms with input frequen-
cies; hence it is erroneous to state that the zero input response is equal to the
natural response or that the zero state response is the forced response.

Example 2.14

As another example, let us analyze the discrete-time system model shown in

Figure 2.6. Assuming the initial states are zero, we get the equations for the
outputs of the two adders as

y1 (n) = x(n − 1) − 0.2y1 (n − 1) − 0.4y1 (n − 2)

y2 (n) = 2y1 (n − 1) − 0.1y2 (n − 1)

When a discrete-time system is described by several linear difference equations

like the equations above, it is difficult to derive a single-input, single-output
62 TIME-DOMAIN ANALYSIS AND z TRANSFORM

2.0
X(z) Y1(z) Y2(z)
z−1 ∑ z−1 ∑

z−1
−0.2 z−1

−0.1
z−1

−0.4

Figure 2.6 Example of a discrete-time system.

equation, and hence solving for output by using the recursive algorithm or
the classical method is not possible. However, we can transform the difference
equations to their equivalent z-transform equations. They become linear, alge-
braic equations that can be solved to find the z transform of the output using
matrix algebra. The inverse z transform of the output function gives us the final
solution in the time domain. So it is the z-transform method that is the more pow-
erful method for time-domain analysis. To illustrate this method, let us transform
the two equations above in the time domain to get the following:

Y1 (z) = z−1 X(z) − 0.2z−1 Y1 (z) − 0.4z−2 Y1 (z)

Y2 (z) = 2z−1 Y1 (z) − 0.1z−1 Y2 (z)

Rearranging these equations in the form

Y1 (z)[1 + 0.2z−1 + 0.4z−2 ] + Y2 (z)[0] = z−1 X(z)

Y1 (z)[−2z−1 ] + Y2 (z)[1 + 0.1z−1 ] = 0

and expressing them in a matrix equation as

    −1 
(1 + 0.2z−1 + 0.4z−2 ) 0 Y1 (z) z X(z)
=
−2z−1 (1 + 0.1z−1 ) Y2 (z) 0

from Cramer’s rule, we find the z transform of the output y2 (n):

 
 (1 + 0.2z−1 + 0.4z−2 ) z−1 X(z) 
 
 −2z−1 0 
Y2 (z) =  
 (1 + 0.2z + 0.4z ) 0
−1 −2 
 
 −2z−1 (1 + 0.1z−1 ) 
 SOLVING DIFFERENCE EQUATIONS USING THE CLASSICAL METHOD 63

 
2z−2
= X(z)
1 + 0.3z−1 + 0.42z−2 + 0.04z−3
 
2z
= 3 X(z)
z + 0.3z2 + 0.42z + 0.04

When we substitute the z transform of the given input above and find the inverse
z transform, we get the output y2 (n).
In this example, the natural frequencies of the system are computed as the
zeros of the system determinant
 
 (1 + 0.2z−1 + 0.4z−2 ) 0 
 
 −2z−1 −1 
(1 + 0.1z )

which is 1 + 0.21z−1 + 0.42z−2 + 0.04z−3 = z−3 [z3 + 0.3z2 + 0.42z + 0.04]. It

is obvious that the zeros of this determinant are the same as the poles of the
transfer function
 
Y2 (z) 2z
H (z) = = 3
X(z) z + 0.3z2 + 0.42z + 0.04

As long as these poles of H (z) are not canceled by its zeros, that is, if there
are no common factors between its numerator and the denominator, its inverse z
transform will display all three natural frequencies. If some poles of the transfer
function are canceled by its zeros, and it is therefore given in its reduced form, we
may not be able to identify all the natural frequencies of the system. Therefore
the only way to find all the natural frequencies of the system is to look for
the zeros of the system determinant or the characteristic polynomial. We see
that in this example, the system response does contain three terms in its natural
response, corresponding to the three natural frequencies of the system. But if
and when there is a cancellation of its poles by some zeros, the natural response
components corresponding to the canceled poles will not be present in the zero
state response h(n). So we repeat that in some cases, the poles of the transfer
function may not display all the natural frequencies of the system.
Note that the inverse z transform of Y2 (z) is computed from Y2 (z) = H (z)X(z)
when the initial states are zero. Therefore the response y2 (n) is just the zero state
response of the system, for the given input x(n).

2.4.1 Transient Response and Steady-State Response

The total response can also be expressed as the sum of its transient response and
steady-state response. But there is again a misconception that the natural response
of a system is the same as the transient response, and hence an explanation is
given below to clarify this misconception.
The transient response is the component of the total response, which approaches
zero as n → ∞, whereas the steady-state response is the part that is left as the
64 TIME-DOMAIN ANALYSIS AND z TRANSFORM

nonzero component. All terms with their frequencies that lie within the unit circle
of the z plane approach zero as n → ∞, and terms with simple poles that lie on
the unit circle contribute to the steady-state response.
For example, let us consider a function
◦ ◦
0.5z z 0.4z 0.5ej 40 z 0.5e−j 40 z
Y (z) = + + + +
z − 1 (z − 0.2)2 (z + 0.4) (z − ej 50◦ ) (z − e−j 50◦ )

The response y(n) is obtained as

 ◦ ◦ 
y(n) = 0.5 + 5n(0.2)n + 0.4(−0.4)n + cos(50 n + 40 ) u(n)

In this example, Y (z) has a double pole at z = 0.2 and a simple pole at z = −0.4,
and the terms [5n(0.2)n + 0.4(−0.4)n ] u(n) corresponding to these frequencies
inside the unit circle constitute the transient response in y(n) since these terms
approach zero as n → ∞. The other terms in Y (z) have a pole at z = 1 and
◦
another one at z = ±ej 50 . These are frequencies that lie on the unit circle,
and their inverse z transform is 0.5 + cos(50◦ n + 40◦ ) u(n), which remains
bounded and is nonzero as n → ∞. It is the steady-state component in y(n), and
obviously the sum of the transient response and steady-state response is the total
◦
response y(n) of the system. The frequencies at z = ±ej 50 may be the natural
frequencies of the system or may be the frequencies of the forcing function; this
also applies for the other frequencies that show up as the poles of Y (z). The
natural response and forced response are therefore not necessarily the same as
the transient response and the steady-state response. Only by using the different
definitions of these terms should one determine the different components that add
up to the total response. In summary, we have shown how to express the total
response as the sum of two terms in the following three different ways:

• The zero state response and the zero input response

• The natural response and the forced response
• The transient response and the steady-state response

2.5 z TRANSFORM METHOD REVISITED

The transfer function H (z) of a system is defined as the ratio of the z transform
of the output and the z transform of the input, under the condition that all initial
states are zero and there are no other independent sources within the system. For
the system described in Figure 2.6, the ratio
Y2 (z) 2z
= 3 = H (z)
X(z) z + 0.3z2 + 0.02z + 0.8
is the transfer function. So we can also use the relationship Y2 (z) = H (z)X(z).
That means that when X(z) = 1 and when the initial states are zero, we have
 CONVOLUTION REVISITED 65

Y2 (z) = H (z), or the output response of a system when it is excited by a unit

pulse function δ(n), under zero initial states, is given by the inverse z transform
of H (z). Thus the unit pulse response denoted by h(n) is given by Z−1 [H (z)].
So h(n) is the response of the system due to an excitation δ(n) only. However,
from the general relationship Y2 (z) = H (z)X(z), we observe that if we know the
transfer function H (z) or if we know the unit pulse response h(n) of the system,
we can find the response due to any other input x(n). Therefore H (z) or the
unit impulse response h(n) constitutes another model for the system. If we have
derived or have been given H (z) or h(n), next we find the z transform X(z)
of the given input, and multiply H (z) and X(z) to get Y (z) = H (z)X(z) as the
z transform of the output. Then we find the inverse z transform of Y (z) to get
the output y(n). For these operations, which are algebraic in nature, finding the
output y(n) as the inverse z transform of H (z)X(z) is an efficient method for
finding the system output. It is this z-transform method that is used extensively
in system analysis, but it depends on the satisfaction of two conditions: (1) we
can find the z transform of the input sequence and (2) we know or can find the
transfer function of the system under investigation. Students should be aware that
in practice, either one or both of these conditions may not be satisfied and other
methods of analysis or design of systems are called for. For example, finding a
closed-form expression for a discrete-time signal obtained by sampling a speech
is not easy. Finding the transfer function of physical systems may not be as easy
and straightforward as the one shown in (2.46). In this book, we assume that
these conditions are always satisfied.

2.6 CONVOLUTION REVISITED

In a previous section on convolution, we had shown that the output y(n) of a

linear, shift-invariant, discrete-time system isobtained by convolution of x(n)
and h(n), specifically, y(n) = x(n) ∗ h(n) = ∞ k=0 x(k)h(n − k).
Since Y (z) = H (z)X(z) = X(z)H (z), we now conclude that convolution sum
operation is commutative:

x(n) ∗ h(n) = h(n) ∗ x(n)

∞ 
Therefore y(n) = k=0 x(k)h(n − k) = ∞ k=0 h(k)x(n − k).
Another way of proving this result is as follows. Let
∞

y(n) = x(k)h(n − k).
k=0

Then
∞ ∞
∞ 

−n
Y (z) = y(n)z = x(k)h(n − k) z−n
n=0 n=0 k=0
66 TIME-DOMAIN ANALYSIS AND z TRANSFORM

Interchanging the order of summation, we obtain

∞
∞

Y (z) = x(k) h(n − k)z−n
k=0 n=0

Let us make a substitution m = n − k, and now we have

∞
∞

Y (z) = x(k) h(m)z−(m+k)
k=0 m=−k
∞
∞

= x(k)z−k h(m)z−m
k=0 m=−k

But h(m) = 0 for −k ≤ m ≤ −1, so that

∞
∞

Y (z) = x(k)z−k h(m)z−m
k=0 m=0

= X(z)H (z) = H (z)X(z)

So we have proved that

1. x(n) ∗ h(n) = h(n) ∗ x(n), which means that the convolution sum is com-
mutative. It is now easy to prove that this satisfies the following additional
properties, by using the algebraic relationships for the z transforms of the
discrete-time sequences.
2. KX1 (z)X2 (z) = X1 (z)KX2 (z). Hence convolution sum operation is linear:
Kx1 (n) ∗ x2 (n) = x1 (n) ∗ Kx2 (n).
3. [X1 (z)X2 (z)]X3 (z) = X1 (z)[X2 (z)X3 (z)]. Hence convolution sum opera-
tion is associative:

[x1 (n) ∗ x2 (n)] ∗ x3 (n) = x1 (n) ∗ [x2 (n) ∗ x3 (n)]

4. X1 (z)[X2 (z)+X3 (z)] = X1 (z)X2 (z)+X1 (z)X3 (z). Convolution sum opera-
tion is distributive:

x1 (n) ∗ [x2 (n) + x3 (n)] = x1 (n) ∗ x2 (n) + x1 (n) ∗ x3 (n)

It is interesting to make a new interpretation of the convolution sum operation

as explained below. Let the z transforms X(z), H (z), and Y (z) be expressed in
their power series expansion:

X(z) = x0 + x1 z−1 + x2 z−2 + x3 z−3 + · · ·

H (z) = h0 + h1 z−1 + h2 z−2 + h3 z−3 + · · ·
Y (z) = y0 + y1 z−1 + y2 z−2 + y3 z−3 + y4 z−4 + y5 z−4 + y6 z−6 + · · ·
 CONVOLUTION REVISITED 67

The coefficients x0 , x1 , x2 , . . . and h0 , h1 , h2 , h3 , . . . are the known samples of

the input x(n) and the unit sample response h(n). Either one or both sequences
may be finite or infinite in length. If we multiply the polynomial or the power
series for X(z) and H (z), and group all the terms for the coefficients of z−n , in
the polynomial or the power series, we get
X(z)H (z) = (x0 + x1 z−1 + x2 z−2 + x3 z−3 + · · ·)
× (h0 + h1 z−1 + h2 z−2 + h3 z−3 + · · ·)
= (x0 h0 ) + (x0 h1 + x1 h0 )z−1 + (x0 h2 + x1 h1 + x2 h0 )z−2
+ (x0 h3 + x1 h2 + x2 h1 + x3 h0 )z−3 + · · ·
By comparing the coefficients of z−n in Y (z) and those in this expression, we
notice that
y0 = (x0 h0 )
y1 = (x0 h1 + x1 h0 )
y2 = (x0 h2 + x1 h1 + x2 h0 )
y3 = (x0 h3 + x1 h2 + x2 h1 + x3 h0 )
·
· (2.51)
·
yn = (x0 hn + x1 hn−1 + x2 hn−2 + x3 hn−3 + · · · + xn h0 )
·
·
·
These are the same results
 as given in (2.7), which are obtained by expanding the
convolution sum yn = ∞ k=−∞ x(k)h(n − k). We can multiply the polynomial
or the power series as H (z)X(z) and identify the coefficients of the resulting
polynomial as yn = ∞ k=−∞ h(k)x(n − k). [We can also find the coefficients of
H (z)X(z) by computing the convolution of the coefficients of H (z) and X(z).]
Then we would get the following expressions for the coefficients, which are
the same as those given in (2.51):
y0 = (h0 x0 )
y1 = (h0 x1 + h1 x0 )
y2 = (h0 x2 + h1 x1 + h2 x0 )
y3 = (h0 x3 + h1 x2 + h2 x1 + h3 x0 )
·
· (2.52)
·
68 TIME-DOMAIN ANALYSIS AND z TRANSFORM

yn = (h0 xn + h1 xn−1 + h2 xn−2 + h3 xn−3 + · · · + hn x0 )

·
·
·

One can store the coefficients h(n) and x(n) for a system being investigated,
on a personal computer or workstation, do the time reversal off line, delay the
time-reversed sequence, and multiply the terms and add the products as explained
in Figure 2.4. Computer software has been developed to perform the convolution
of two sequences in a very rapid and efficient manner—even when the sequences
are very long.2 But a real hardware that contains the electronic devices such as
the delay element, multiplier, and the adder cannot reverse a sequence in real
time, but it operates on the incoming samples of the input as follows. When the
sample x0 enters the system at t = 0, it launches the sequence x0 h(nT ), which
appears at the output; when the next sample x1 enters the system at t = T , it
launches the sequence x1 h(nT − T ), which appears at the output, and when the
next sample x2 enters the system, the sequence at the output is x2 h(nT − 2T ),
and so on. At any time t = mT , the value of the output sample is

y(mT ) = x0 h(mT ) + x1 h(mT − T ) + x2 h(mT − 2T ) + x3 h(mT − 3T ) + · · ·

This is the physical process being implemented by the real hardware; an example
of this process was described in Figure 2.2. However, a real hardware can be
programmed to store the input data x(n) and h(n) in its memory registers and to
implement the convolution sum.
It is important to remember that convolution can be used to find the output,
even when the input sequence does not have a z transform, that is, when we
cannot use the z-transform approach. This makes convolution a very fundamen-
tal operation for signal processing and is one of the most powerful algorithms
implemented by the electronic hardware as it does not know what z transform is!

Example 2.15
2
Suppose that the input sequence is x(n) = (0.1)n u(n) and the unit impulse
response h(n) = {0.2 0.4 0.6 0.8 1.0}.
The z transform X(z) for the infinite sequence x(n) does not have a closed-
form expression, whereas it is easy to write the z transform H (z) = 0.2 +
0.4z−1 + 0.6z−2 + 0.8z−3 + z−4 . Therefore we cannot find X(z)H (z) = Y (z) as
a rational function and invert to get y(n). However, the polynomial H (z) can be

multiplied by the power series X(z) = ∞ n2 −n
n=0 (0.1) z to get y(n), according

2
Two methods used to improve the efficiency of computation are known as the overlap-add and
overlap-save methods. Students interested in knowing more details of these methods may refer to
other books.
 CONVOLUTION REVISITED 69

to either one of the algorithms x(n) ∗ h(n) or h(n) ∗ x(n). For example

y(0) = 0.2
y(1) = 0.4 + 0.1(0.2)
y(2) = 0.6 + (0.1)(0.4) + (0.1)4 (0.2)
y(3) = 0.8 + (0.1)(0.6) + (0.1)4 (0.4) + (0.1)9 (0.2)
·
·
·

Recollect that we have obtained two different equations for finding the output
due to a given input. They are the convolution sum (2.6) and the linear difference
equation (2.2), which are repeated below.
∞

y(n) = x(k)h(n − k) (2.53)
k=0

N
M
y(n) = − a(k)y(n − k) + b(k)x(n − k) (2.54)
k=1 k=0

In Equation (2.53), the product of the input sequence and the current and previ-
ous values of the unit impulse response are added, whereas in Equation (2.54)
the previous values of the output and present and past values of the input are
multiplied by the fixed coefficients and added. The transfer function H (z) for
the first case is given by H (z) = ∞ −n
n=0 h(n)z , and for the second case, we use
the z transform for both sides to get

N
M
Y (z) = − a(k)z−k Y (z) + b(k)z−k X(z)
k=1 k=0
 

N
M
Y (z) 1 + a(k)z−k = b(k)z−k X(z)
k=1 k=0
M −k
Y (z) k=0 b(k)z
H (z) = =  
X(z) 1+ N −k
k=1 a(k)z

So we can derive the transfer function H (z) from the linear difference equation
(2.54), which defines the input–output relationship.
We can also obtain the linear difference equation defining the input–output
relationship, from the transfer function H (z), simply by reversing
 the steps as
−k
follows. Given the transfer function H (z), we get Y (z)[1 + N k=1 a(k)z ] =
70 TIME-DOMAIN ANALYSIS AND z TRANSFORM

M
b(k)z−k X(z). Finding the inverse z transform for each term, we arrive at
k=0
the input–output relationship for the system, as shown by the following example.

Example 2.16

Let us assume that we are given a transfer function

0.2 + 0.1z
H (z) =
0.8 + 0.6z + 0.2z2 + z3
We rewrite it as a transfer function in inverse powers of z, by dividing both the
numerator and denominator by z3 to get

0.1z−2 + 0.2z−3 Y (z−1 )

H (z−1 ) = −1 −2 −3
=
1 + 0.2z + 0.6z + 0.8z X(z−1 )
Therefore
   
Y (z−1 ) 1 + 0.2z−1 + 0.6z−2 + 0.8z−3 = X(z−1 ) 0.1z−2 + 0.2z−3

By expressing the inverse z transform of each term, we get the linear difference
equation or the input–output relationship

y(n) + 0.2y(n − 1) + 0.6y(n − 2) + 0.8y(n − 3) = 0.1x(n − 2) + 0.2x(n − 3)

Since the transfer function has been defined and derived by setting the initial
conditions to zero , one may assert that from the transfer function we cannot
find the response due to initial conditions, but this is not true. In the preceding
example, after we have derived the input–output relationship from the given
transfer function, we write the corresponding z-transform equation including the
terms containing the initial conditions, in the form

Y (z) + 0.2[z−1 Y (z) + y(−1)] + 0.6[z−2 Y (z) + z−1 y(−1) + y(−2)]

+ 0.8[z−3 Y (z) + z−2 y(−1) + z−1 y(−2) + y(−3)]
= 0.1z−2 X(z) + 0.2z−3 X(z)

We substitute the initial conditions y(−1), y(−2) and y(−3), in these equations
and obtain the zero input response as well as the zero state response of the
system. Therefore the transfer function H (z) constitutes a complete model of the
discrete-time system.

2.7 A MODEL FROM OTHER MODELS

In this section, we review the important concepts and techniques that we have
discussed so far. For this purpose, we select one more example below.
 A MODEL FROM OTHER MODELS 71

Example 2.17

The circuit for a discrete-time system is shown in Figure 2.7. The equations that
describe it are.

y1 (n) = −x(n) + y3 (n − 2)
y2 (n) = d2 y1 (n) + x(n − 1) − y3 (n − 1) (2.55)
y3 (n) = x(n − 2) + d1 y2 (n)

Let us try to eliminate the internal variables y1 (n) and y2 (n) and get a difference
equation relating the output y3 (n) and x(n):
 
y3 (n) = x(n − 2) + d1 d2 y1 (n) + x(n − 1) − y3 (n − 1)
 
= x(n − 2) + d1 d2 {−x(n) + y3 (n − 2)} + x(n − 1) − y3 (n − 1)

The difference equation given below is the input–output relationship obtained

by substituting the expression for y2 (n) and y1 (n) successively in the expression
for y3 (n):

y3 (n) + d1 y3 (n − 1) − d1 d2 y3 (n − 2) = x(n − 2) + d1 x(n − 1) − d1 d2 x(n)

(2.56)
But remember that in general, it may not be so easy to obtain the single-input,
single-output relationship from the many equations written in the time domain,
by successive elimination. It is always easier to z-transform Equations (2.55) or
write them directly from the circuit diagram, use matrix algebra to obtain the
transfer function H (z) and then obtain the difference equation as shown below:

Y1 (z) = −X(z) + z−2 Y3 (z)

x(n − 1) x(n − 2)
x(n) z−1 z−1 Σ y3(n)

−1

y1(n) d2 d1
y2(n)
Σ Σ

−1

z−1 z−1

Figure 2.7 Structure of a discrete-time system.

72 TIME-DOMAIN ANALYSIS AND z TRANSFORM

Y2 (z) = d2 Y1 (z) + z−1 X(z) − z−1 Y3 (z) (2.57)

Y3 (z) = z−2 X(z) + d1 Y2 (z)

Solving these algebraic equations, by using matrix algebra, we get

Y3 (z) z−2 + d1 z−1 − d1 d2

= H (z) = (2.58)
X(z) 1 + d1 z−1 − d1 d2 z−2

Now we can derive the difference equation relating the input and output, in the
form of (2.56).
Let us choose d1 = 0.5 and d2 = −0.5. Then the preceding transfer function
reduces to

z−2 + 0.5z−1 + 0.25

H (z) =
1 + 0.5z−1 + 0.25z−2
1 + 0.5z + 0.25z2
= (2.59)
z2 + 0.5z + 0.25
The unit impulse response h(n) is computed as the inverse z transform of H (z):

H (z) k1 k1∗
=4+ +
z (z + 0.25 − j 0.433) (z + 0.25 + j 0.433)
◦ ◦
4 (1.9843ej 160.9 ) (1.9843e−j 160.9 )
= + ◦ + ◦
z (z − 0.5ej 120 ) (z − 0.5e−j 120 )

Therefore we have
◦ ◦
(1.9843ej 160.9 )z (1.9843e−j 160.9 )z
H (z) = 4 + ◦ + ◦
(z − 0.5ej 120 ) (z − 0.5e−j 120 )
 ◦ ◦ ◦ ◦ 
h(n) = 4δ(n) + 1.9843 e160.9 (0.5ej 120 )n + e−160.9 (0.5e−j 120 )n
 ◦ ◦ 
= 4δ(n) + 3.9686(0.5)n cos(120 n + 160.9 ) u(n) (2.60)

2.7.1 Review of Model Generation

Using the example above, we review the different models derived for the system
and show how we can obtain one model from the other models. Recollect the
two definitions for a model of the discrete-time system given in Section 2.1:

1. The circuit shown in Figure 2.7 is model 1, in a block diagram represen-

tation.
2. The equations in the discrete-time domain is model 2—an example is
Equation (2.55).
 A MODEL FROM OTHER MODELS 73

3. The single difference equation for the input–output relationship is model

3—an example is Equation (2.56).
4. The equations in the z domain is model 4—an example is Equation (2.57).
5. The z transform for model 3 or the transfer function H (z) is model 5—an
example is Equation (2.59).
6. The unit impulse response h(n) is model 6—an example is Equation (2.60).

The first model is a circuit diagram, whereas the remaining ones are mathematical
models describing the discrete-time system.
In the example worked out above, we have shown how to derive model 2
from model 1, model 3 from model 2, to model 6 from model 5. It is easy to see
that we can get model 5 from model 6 and model 3 from model 5. But when we
get a model 2 or 4 from model 3 or 4, the result is not unique. We will show
that getting a circuit model from model 5 is not unique, either. Yet the flexibility
to generate one model from many of the other models makes the analysis of
discrete-time systems very versatile and requires that we learn how to choose the
most appropriate model to find the output of a system for a given input, with
initial states also given. Even when the transfer function of a system is derived
under zero initial states, we can get model 3 and then can include the previous
values of the output as the initial states and obtain the total output.
Property 2.3: Time Reversal If X(z) is the z transform of a causal sequence
x(n), n ≥ 0, then the z transform of the sequence x(−n) is X(z−1 ). The sequence
x(−n) is obtained by reversing the sequence of time, which can be done only by
storing the samples of x(n) and generating the sequence x(−n) by reversing the
order of the sequence. If a discrete-time sequence or data x(n) is recorded on an
audiocassette or a magnetic tape, it has to be played in reverse to generate x(−n).
The sequence x(−n) and its z transform X(z−1 ) are extensively used in the
simulation and analysis of digital signal processing for the purpose of designing
digital signal processors, although the sequence x(−n) cannot be generated in
real time, by actual
 electronic−nsignal processors.
0 ∞
Let X(z) = ∞ n=0 x(n)z . Then, n=−∞ x(−n)z =
n
n=0 x(n)z =
n
−1
X(z ).

Example 2.18
∞
1 z
Z[(0.5)n u(n)] = (0.5)n z−n = −1
= = X(z)
1 − 0.5z z − 0.5
n=0

Then

0
z−1 1
Z[(0.5)−n u(−n)] = (0.5)−n zn = X(z−1 ) = −1
=
n=−∞
z − 0.5 1 − 0.5z

If
0.1 + 0.25z−1 + 0.6z−2
F (z) = (2.61)
1.0 + 0.4z−1 + 0.5z−2 + 0.3z−3 + 0.08z−4
74 TIME-DOMAIN ANALYSIS AND z TRANSFORM

is the z transform of f (n)u(n), multiplying both the numerator and denominator

polynomial by z4 , we can express F (z) in the form

0.1z4 + 0.25z3 + 0.6z2

F (z) = (2.62)
z4 + 0.4z3 + 0.5z2 + 0.3z + 0.08

The z transform of f (−n)u(−n) is obtained by replacing z by z−1 in (2.62) for

F (z), and therefore the z transform of f (−n)u(−n) is

0.1z−4 + 0.25z−3 + 0.6z−2

F (z−1 ) = (2.63)
z−4 + 0.4z−3 + 0.5z−2 + 0.3z−1 + 0.08
0.1 + 0.25z + 0.6z2
= (2.64)
1.0 + 0.4z + 0.5z2 + 0.3z3 + 0.08z4
which is different from either (2.61) or (2.62) for F (z) given above.

Property 2.4: Initial Value If the z transform of a sequence x(n) is known as

X(z), the value of its sample x(0) is called the initial value and is easily found
from

x(0) = lim X(z)

z→∞

We prove this result by expressing X(z) = x(0) + x(1)z−1 + x(2)z−2 + · · · and

noting that all terms except the first term x(0) go to zero as z → ∞. If, how-
ever, we are interested in finding a few samples x(0), x(1), x(2), . . . , x(k) at the
beginning, and not just the initial value x(0), then we use long division to get a
few samples of the quotient as illustrated below.
Consider, for example, the transfer function

0.1 + 0.21z−1 − 0.134z−2 − 0.0514z−3 + · · ·


1.0 + 0.4z−1 + 0.5z−2 0.1 + 0.25z−1
0.1 + 0.04z−1 + 0.05z−2
0.21z−1 − 0.05z−2
0.21z−1 + 0.084z−2 + 0.105z−3
−0.134z−2 − 0.105z−3
−0.134z−2 − 0.0536z−3 − 0.067z−4
−0.0514z−3 + 0.067z−4
..
.
(2.65)
 A MODEL FROM OTHER MODELS 75

From the coefficients in the quotient, we see that x(0) = 0.1, x(1) = 0.21, x(2) =
−0.134, x(3) = −0.0514 and by continuing this procedure, we can get x(4),
x(5), . . ..
The method of finding the few coefficients of the inverse z transform of a
transfer function X(z) can be shown to have a recursive formula [1,6] which is
given as follows. Let the transfer function X(z) be expressed in the form
M −n
n=0 bn z
N = x0 + x1 z−1 + x2 z−2 + x3 z−3 + · · · (2.66)
−n
n=0 an z

The samples of the inverse z transform are given by the recursive formula
 
1 n
xn = bn − x(n − i)ai , n = 1, 2, . . . . (2.67)
a0
i=1

where x0 = b0 /a0 .

Property 2.5: Final Value To find the value of x(n), as n → ∞, we use

lim x(N ) = lim (z − 1)X(z)

N →∞ z→1

when and if (z − 1)X(z) has all its poles inside the unit circle.

Proof : Consider X(z) = x(0) + x(1)z−1 + x(2)z−2 + · · ·. If we shift the

sequence x(0), x(1), x(2), . . . to the left (i.e., advance it) by one sample, we
have the sequence of values x(1) at n = 0, x(2) at n = 1, x(3) at n = 2, and so
on. This sequence is represented as x(n + 1), and the unilateral z transform of
this sequence is

Z[x(n + 1)] = x(1) + x(2)z−1 + x(3)z−2 + x(4)z−3 + · · ·

 
= z x(1)z−1 + x(2)z−2 + x(3)z−3 + x(4)z−4 + · · ·

= z −x(0) + x(0) + x(1)z−1 + x(2)z−2 + x(3)z−3

+ x(4)z−4 + · · ·
= z [−x(0) + X(z)]

Let us express

N
N
−N
Z[x(n)] = lim x(n)z and Z[x(n + 1)] = lim x(n + 1)z−N
N→∞ N →∞
n=0 n=0
76 TIME-DOMAIN ANALYSIS AND z TRANSFORM

Then

Z[x(n + 1)] − Z[x(n)]

 N 

= lim {x(n + 1) − x(n)} z−N = z [−x(0) + X(z)] − X(z)
N→∞
n=0

= [(z − 1)X(z) − zx(0)]

Letting z → 1, we get
 

N
lim {x(n + 1) − x(n)} = lim [(z − 1)X(z) − zx(0)]
N→∞ z→1
n=0

= lim [{x(N + 1) − x(0)}]

N →∞

= lim [(z − 1)X(z) − x(0)]

z→1

= x(∞) − x(0)

where we have assumed that limN →∞ [{x(N + 1)}] = x(∞) has a finite or zero
value. This condition is satisfied when (z − 1)X(z) has all its poles inside the
unit circle. Under this condition, we have proved that

x(∞) = lim [(z − 1)X(z)]

z→1

Property 2.6: Multiplication by r n If we have Z[x(n)u(n)] = X(z), then

Z[r n x(n)u(n)] = X(z/r).

Proof :
∞
∞
 z −n z
n n −n
Z[r x(n)u(n)] = r x(n)z = x(n) =X
r r
n=0 n=0

We have already used this property in deriving a few z transforms shown in

Table 2.1. As another example, let Z[a n u(n)] = z/(z − a). Then
z
z
Z[r n a n u(n)] =  z r =
−a z − ra
r
A few other properties have already been discussed in this chapter, and
Table 2.3 summarizes the properties of z transforms.
 STABILITY 77

TABLE 2.3 Properties of z Transforms

Operation x(n)u(n) X(z)

Addition x1 (n) + x2 (n) X1 (z) + X2 (z)

Scalar multiplication Kx(n) KX(z)
Delay x(n − 1)u(n − 1) z−1 X(z) + x(−1)
x(n − 2)u(n − 2) z−2 X(z) + z−1 x(−1) + x(−2)
x(n − 3)u(n − 3) z−3 X(z) + z−2 x(−1) +
z−1 x(−2) + x(−3)

x(n − m)u(n − m) z−m X(z) + m−1n=0 x(n − m)z
−n

Time reversal x(−n)u(−n) X(z−1 )

dX(z)
Multiplication by n nx(n) −z
dz
Multiplication by r n
r n x(n) X(r −1 z)
Time convolution x1 (n) ∗ x2 (n) X1 (z)X2 (z)
 z
Modulation x1 (n)x2 (n) (1/2πj ) C X1 (z)X2 u−1 du
u
Initial value x(0) limz→∞ X(z)
Final value limN →∞ x(N) limz→1 (z − 1)X(z), when poles
of (z − 1)X(z) are inside the
unit circle

2.8 STABILITY

It is essential that every system designed by an engineer be extremely stable in

practical use. Hence we must always analyze the stability of the system under
various operating conditions and environments. The basic requirement is that
when it is disturbed by a small input, the response of the system will eventually
attain a zero or a constant value or at most be bounded within a finite limit.
There are definitions for various kinds of stability, but the definition used most
often is that the output asymptotically approaches a constant or bounded value
when a bounded input is applied. This is known as the bounded input–bounded
output (BIBO) stability condition. It 
satisfies the condition when the unit impulse
response h(n) satisfies the condition ∞ n=0 |h(n)| < M < ∞. To prove this result,
let us assume that H (z) = z/z − γi , where γi is the pole of H (z). The unit
impulse response is γin for n ≥ 0:
∞
∞

|h(n)| = |γi |n
n=0 n=0
1
= when |γi | < 1
1 − |γi |
78 TIME-DOMAIN ANALYSIS AND z TRANSFORM


When H (z) = N (z)/ K i=1 (z − γi ), where γi are the poles of H (z) such that
|γi | < 1 for i = 1, 2, 3, . . . , K, we get

∞

K
1
|h(n)| ≤ <∞
1 − |γi |
n=0 i=1

Next consider an input x(n) that is bounded in magnitude:

|x(n)| < B for all n

From the convolution property, we get

 ∞   ∞ 
   
   
|y(n)| =  h(k)x(n − k) ≤  |h(k)| |x(n − k)|
   
k=−∞ k=−∞

Since |x(n)| < B for all values of n and k, we get

∞

|y(n)| ≤ B |h(k)| for all n
k=0

Therefore, we conclude that if the impulse response is absolutely summable, that

is, if
∞

|h(n)| < ∞
n=0

then the output y(n) is bounded in magnitude when the input x(n) is bounded
and the system is BIBO-stable.
There are a few tests that we can use to determine whether the poles of a
transfer function

b0 + b1 z−1 + b2 z−2 + · · · + bM z−M

H (z) =
a0 + a1 z−1 + a2 z−2 + a3 z−3 + a4 z−4 + · · · + aN z−N

b0 zN + b1 zN −1 + · · · + bM zN −M
=
a0 zN + a1 zN −1 + a2 zN −2 + · · · + aN

are inside the unit circle in the z plane.

2.8.1 Jury–Marden Test

To determine whether the poles are inside the unit circle in the z plane, we
choose the Jury–Marden test [1,4] because it has some similarity with the
 STABILITY 79

TABLE 2.4 The Jury–Marden Array of Coefficients

Row Coefficients

1 a0 a1 a2 . . . aN −1 aN
2 aN aN −1 aN −2 . . . a1 a0
3 c0 c1 c2 . . . cN −1
4 cN −1 cN −2 . . . c1 c0
5 d0 d1 d2 . . . dN −2
6 dN −2 dN −3 . . . d0
7
8
..
.
2N − 3 r0 r1 r2

Routh–Hurwitz test that the students have learned from an earlier course, and it
is easier than the other tests that are available.3
We consider the coefficients of the denominator arranged in descending pow-
ers of z, specifically D(z) = a0 zN + a1 zN −1 + a2 zN −2 + · · · + aN where a0 > 0.
The first row of the Jury–Marden array lists the coefficients a0 , a1 , a2 , . . . , aN
(see Table 2.4), and the second row lists these coefficients in the reverse order,
aN , aN −1 , aN −2 , . . . , a2 , a1 , a0 So we start with the two rows with elements
chosen directly from the given polynomial as follows:
a0 a1 a2 . . . aN −1 aN
aN aN −1 aN −2 . . . a1 a0
The elements of the third row are computed as second-order determinants
according to the following rule:
 
 a0 aN −i 
ci =   for i = 0, 1, 2, . . . , (N − 1)
aN ai 

For example
 
a aN 
c0 =  0
aN a0 
 
a aN −1 
c1 =  0 
aN a1
 
a aN −2 
c2 =  0 
aN a2

3
However, in Chapter 6 we describe the use of a MATLAB function tf2latc, which is based on
the Schur–Cohn test.
80 TIME-DOMAIN ANALYSIS AND z TRANSFORM

Note that the entries in the first column of the determinants do not change as i
changes in computing ci . The coefficients of the fourth row are the coefficients
of the third row in reverse order, as shown in the array below. The elements of
the fifth row are computed by
 
 c0 cN −1−i 

di = 
cN −1 ci  for i = 0, 1, 2, . . . , (N − 2)

For example
 
c cN −1 
d0 =  0 
cN −1 c0
 
c cN −2 
d1 =  0 
cN −1 c1
 
c cN −3 
d2 =  0 
cN −1 c2

and the elements of the sixth row are those of the fifth row in reverse order.
Note that the number of elements in these rows are one less than those in the
two rows above. As we continue this procedure, the number of elements in each
successive pair of rows decreases by one, until we construct (2N − 3) rows and
end up with the last row having three elements. Let us denote them as r0 , r1 , r2.
The Jury–Marden test states that the denominator polynomial D(z) = a0 zn +
a1 zn−1 + · · · + aN has roots inside the unit circle in the z plane if and only if the
following three conditions are satisfied. Note here that we need to express the
denominator polynomial in positive powers of z, because we have to evaluate it
at z = ±1 in the first two criteria shown below:

1. D(1) = D(z)|z=1 > 0

2. (−1)N D(−1) > 0
3. a0 > |aN |

Also
|c0 | > |cN −1 |
|d0 | > |dN −2 |
..
.
|r0 | > |r2 |

Example 2.19

Let us consider the denominator polynomial D(z) = 5z5 + 4z4 + 3z3 + z2 +

z + 1. We construct the Jury–Marden array following the method described
above:
 SOLUTION USING MATLAB FUNCTIONS 81

Row Coefficients

1 5 4 3 1 1 1
2 1 1 1 3 4 5
3 24 19 14 2 1
4 1 2 14 19 24
5 575 454 322 29
6 29 322 454 575
7 329784 251712 171984

Check whether the three test criteria are satisfied:

D(1) = D(z)|z=1 = 15(−1)5 R(−1) = 3
a0 = 5; a5 = 1; a0 > |a5 |
c0 = 24; c4 = 1; |c0 | > |c4 |
d0 = 575; d3 = 29; |d0 | > |d3 |
r0 = 329784; r2 = 171984; |r0 | > |r2 |
All criteria are satisfied, and therefore the D(z) above has its five zeros inside
the unit circle.

Example 2.20
Now consider another example: D(z) = 3z4 + 5z3 + 3z2 + 2z + 1. The Jury–
Marden array is constructed as shown below:

jury–marden array
Row 1 3 5 3 2 1
2 1 2 3 5 3
3 8 13 6 1
4 1 6 13 8
5 63 98 35

Although we have calculated all the entries in the array, we find that the second
criterion is not satisfied because (−1)4 D(−1) = 0. We conclude that there is at
least one zero of D(z) that is not inside the unit circle. Indeed, it is found that
there is one zero at z = −1.000. It is a good idea to check at the beginning,
whether the first two criteria are satisfied, because if one or both of these two
criteria (which are easy to check) fail, there is no need to compute the entries in
the rows after the first two rows of the Jury–Marden array.

2.9 SOLUTION USING MATLAB FUNCTIONS

In the previous sections, we have described many models for the discrete-time
system and discussed three methods of finding the output of the system when
the input sequence is given, along with initial conditions in some cases.
82 TIME-DOMAIN ANALYSIS AND z TRANSFORM

The three methods are

1. Recursive algorithm
2. Convolution sum
3. z-Transform method

In this section, we illustrate the use of MATLAB functions to implement some of

the algorithms discussed in the previous sections. At this point it is strongly sug-
gested that the students review the MATLAB primer in Chapter 9 to refresh their
understanding of MATLAB, although it is only an introduction to the software.
First let us consider the case of a system described by the linear shift-invariant
difference equation

y(n) = 0.4y(n − 1) + 0.05y(n − 2) + x(n)

where the initial states are given as y(−1) = 2 and y(−2) = 1.0. We learned
how to find the output of this system for any given input, by using the recursive
algorithm. Assuming x(n) = δ(n) and the initial two states in this example to
be zero, we found the unit impulse response h(n). Knowing the unit impulse
response, we can find the response when any input is given, by using the convo-
lution algorithm. It was pointed out that convolution algorithm can be used to find
only the zero state response since it uses h(n), whereas the recursive algorithm
computes the total response due to the given input and the initial states.
Now we use the z transform to convert the difference equation above to get

Y (z)[1 − 0.4z−1 − 0.05z−2 ] = 0.4y(−1) + 0.05[z−1 y(−1) + y(−2)] + X(z)

= [0.8 + 0.1z−1 + 0.05] + X(z)

Therefore

0.85 + 0.1z−1 X(z)

Y (z) = + (2.68)
[1 − 0.4z−1 − 0.05z−2 ] [1 − 0.4z−1 − 0.05z−2 ]
= Y0i (z) + Y0s (z) (2.69)
= Y0i (z) + H (z)X(z)

We obtain the transfer function H (z) = 1/[1 − 0.4z−1 − 0.05z−2 ] from the given
linear difference equation describing the discrete-time system.
But when we decide to use MATLAB functions, note that if the given input is a
finite-length sequence x(n), we can easily find the coefficients of the polynomial
in the descending powers of z as the entries in the row vector that will be
required for defining the polynomial X(z). But if the input x(n) is infinite in
length, MATLAB  cannot find a closed-form expression for the infinite power
series X(z) = ∞ n=0 x(n)z −n
; we have to find the numerator and denominator
coefficients of X(z).
 SOLUTION USING MATLAB FUNCTIONS 83

Example 2.21

As an example, let us assume that x(n) = [(−0.2)n + 0.5(0.3)n ]u(n). We have

to derive its z transform as

z 0.5z
X(z) = + (2.70)
z + 0.2 z − 0.3
1.5z2 − 0.2z
= (2.71)
z2 − 0.1z − 0.06
1.5 − 0.2z−1
= (2.72)
1 − 0.1z−1 − 0.06z−2

Using (2.70), we have

0.85 + 0.1z−1
Y (z) =
[1 − 0.4z−1 − 0.05z−2 ]
1.5 − 0.2z−1
+ (2.73)
[1 − 0.4z−1 − 0.05z−2 ][1 − 0.1z−1 − 0.06z−2 ]

We illustrate the use of MATLAB function conv to find the product of two
polynomials in the denominator of Y0i (z)

den2=conv(d1,d2)

where the entries for the row vectors d1 and d2 are the coefficients in ascending
powers of z−1 for the two polynomials [1 − 0.4z−1 − 0.05z−2 ] and [1 − 0.1z−1 −
0.06z−2 ].
So we use the following MATLAB statements to find the coefficients of their
product by convolution:

d1=[1 -0.4 -0.05];

d2=[1 -0.1 -0.06];
den2=conv(d1,d2).

MATLAB gives us the vector den2 = [1.00 -0.50 -0.07 0.029 0.003].

Example 2.22

We introduce three MATLAB functions residuez, impz, and filter, which

are very useful in time domain analysis of discrete-time systems:

[r,p,k]=residuez(num,den)
84 TIME-DOMAIN ANALYSIS AND z TRANSFORM

This function gives us the partial fraction expansion of the z transform

B(z−1 ) b0 + b1 z−1 + b2 z−2 + · · · + bM z−M

Y (z−1 ) = = (2.74)
A(z−1 ) 1 + a1 z−1 + a2 z−2 + a3 z−3 + · · · + bN z−N

in the form

r(1) r(2) r(N )

k1 + k2 z−1 + k2 z−2 + · · · + + + ··· +
1 − p(1)z−1 1 − p(2)z−1 1 − p(N )z−1
(2.75)
which can be expressed in a more familiar form:

r(1)z r(2)z r(N )z

k1 + k2 z−1 + k2 z−2 + · · · + + + ··· +
z − p(1) z − p(2) z − p(N )

The vector num = [b0 b1 b2 . . . bM ] and the vector den = [1 a1 a2 a3

. . . aN ] list the coefficients of the numerator and denominator polynomial in
ascending powers of z−1 . In the output, the vectors r and p list the residues and
corresponding poles, while k is the vector of gain constants which are present
when M > N . If there is a pole p(j ) of multiplicity m, then the partial fraction
expansion will show terms in the form

r(j ) r(j + 1) r(j + 2) r(j + m − 1)

 + 2 +  3 + · · · · +  m
1 − p(j )z−1 1 − p(j )z −1 1 − p(j )z −1 1 − p(j )z−1
(2.76)
After we have obtained the partial fraction expansion, we can express the inverse
z transform for each of its terms to get y(n) for all n ≥ 0. This is one method
for finding the response of the discrete-time system.
Instead of this procedure, we can use the MATLAB function impz to obtain
the response y(n), but this procedure yields the value of the response y(n) for
only a finite number of samples n = 0, 1, 2, 3, . . . , K:

[y,T]=impz(num,den,K)

In this function, the column vector T = [0 1 2 3 . . . K]

and the column
vector y gives us the K samples of the inverse z transform of Y (z−1 ) = y(n).
We can then plot the samples in y(n) using the function stem(T,y). If we use
only the command impz(num,den,K), without the output arguments, we will
get the plot of y(n) immediately.
The third MATLAB function that we use is filter, which gives us the
output y(n) of the system with a transfer function H (z−1 ) when its input is a
finite sequence x(n):

y =filter(num,den,x)
 SOLUTION USING MATLAB FUNCTIONS 85

So we enter the samples of the input in the row vector x, besides the vectors for
the coefficients of num and den of H (z−1 ). When the vector x is simply 1, the
output vector y is obviously the unit sample response h(n). This function even
allows us to find the output when initial states are given, if we use

[y, F]=filter(num, den, x, I0 )

where I0 is the vector listing the initial conditions and F is the final value. It is
important to know that although the transfer function H (z−1 ) is the z transform of
the zero state response, the function filter implements the recursive algorithm
based on the transfer function and can find the total response when initial states
are also given. So this function is a more useful function in signal processing
applications.

Example 2.23

Let us consider the z transform of the zero input function found in (2.73):

0.85 + 0.1z−1
Y0i (z−1 ) = (2.77)
[1 − 0.4z−1 − 0.05z−2 ]

To find the partial fraction expansion, we use the following MATLAB script:

num=[0.85 0.1];
den=[1 -0.4 -0.05] ;
[r,p,k]=residuez(num,den)

and we get

r= 0.8750
− 0.0250
p= 0.5000
− 0.1000
k=[]

So the partial fraction expansion of Y0i (z−1 ) = 0.8750z/(z − 0.5) − 0.025z/

(z + 0.1). Therefore the zero input response y0i (n) = [0.8750(0.5)n − 0.025
(−0.1)n ] u(n).

Example 2.24

To find the 20 samples of the zero input response y0i (n) directly from (2.77), we
use the function impz in the following script:
86 TIME-DOMAIN ANALYSIS AND z TRANSFORM

num=[0.85 0.1];
den=[1 -0.4 -0.05];
[y,T]=impz (num,den,20)

and we get the samples of output y0i (n) as

y = 0.8500
0.4400
0.2185
0.1094
0.0547
0.0273
0.0137
0.0068
0.0034
0.0017
0.0009
0.0004
0.0002
0.0001
0.0001
0.0000
0.0000
0.0000
0.0000
0.0000

Example 2.25

As the second example, we consider the z transform of the zero state response
Y0s (z−1 ) in (2.73) and use the following MATLAB program to find the partial
fraction expansion:

1.5 − 0.2z−1
Y0s (z−1 ) =
[1 − 0.4z−1 − 0.05z−2 ][1 − 0.1z−1 − 0.06z−2 ]
 SOLUTION USING MATLAB FUNCTIONS 87

d1=[1 -0.4 -0.05];

d2=[1 -0.1 -0.06];
den2=conv(d1,d2).
num=[1.5 -0.2];
[r,p.k]=residuez(num,den2);

The output from the program is

r= 1.6369
− 0.5625
0.5714
− 0.1458
p= 0.5000
0.3000
− 0.2000
− 0.1000
k=[]

Therefore the partial fraction expansion is given by

1.6369z 0.5625z 0.5714z 0.1458z

Y0s (z−1 ) = − + − (2.78)
z − 0.5 z − 0.3 z + 0.2 z + 0.1

and the zero state response

 
y0s (n) = 1.6369(0.5)n − 0.5625(0.3)n + 0.5714(−0.2)n − 0.1458(−0.1)n u(n)

Example 2.26

To find the zero state response by using the function filter, we choose an input
of finite length, say, 10 samples of

x(n) = [(−0.2)n + 0.5(0.3)n ] for n = 0, 1, . . . , 9

we use the following script:

n=(0:9);
x=[(-0.2).^n+0.5*(0.3).^n];
y=filter(b,a,x)
88 TIME-DOMAIN ANALYSIS AND z TRANSFORM

The output y0s (n) is

y = columns 1–7:
1.5000 0.5500 0.3800 0.1850 0.0987 0.0496 0.0252
columns 8–10:
0.0127 0.0064 0.0032

We compute 10 samples of the response obtained from the partial fraction

expansion method:
 
y0s (n) = 1.6369(0.5)n − 0.5625(0.3)n + 0.5714(−0.2)n − 0.1458(−0.1)n

for n = 0, 1, 2, . . . 9, using the following program and find that the result agrees
with that obtained by the function filter:

n=(0:9);
y=[1.6369*(0.5).^n-0.5625*(0.3).^n+0.5714*(-0.2). ^n-0.1458*(
-0.1).^n]

The output is

y = columns 1–7:
1.5000 0.5500 0.3800 0.1850 0.0986 0.0496 0.0252
columns 8–10:
0.0127 0.0064 0.0032

Example 2.27

Now let us verify whether the result from the function impz also agrees with the
results above. We use the script

d1=[1 -0.4 -0.05];

d2=[1 -0.1 -0.06];
den2=conv(d1,d2).
num=[1.5 -0.2];
[y,T]=impz(num,den2)

We get the following result, which also agrees with the results from the pre-
ceding two methods:

y = 1.5000
0.5500
0.3800
 SOLUTION USING MATLAB FUNCTIONS 89

0.1850
0.0986
0.0496
0.0252
0.0127
0.0064
0.0032
0.0016
0.0008
0.0004
0.0002

Example 2.28

To find the unit impulse response h(n) using the function filter, we identify
the transfer function H (z−1 ) in (2.73) as 1/[1 − 0.4z−1 − 0.05z−2 ].
From the MATLAB program

b=[1];
a=[1 -0.4 -0.05];
[r,p,k]=residuez(b,a),

we get

r= 0.8333
0.1667
p= 0.5000
− 0.1000
k=[]

From this data output, we express the transfer function

0.8333z 0.1667z
H (z−1 ) = −
z − 0.5 z + 0.1

and the unit impulse response of the system is

 
h(n) = 0.8333(0.5)n − 0.1667(−0.1)n u(n)
90 TIME-DOMAIN ANALYSIS AND z TRANSFORM

To find the unit impulse response using the function impz, we use

b=[1];
a=[1 -0.4 -0.05];
[y,T]=impz(b,a,20)

and get

y = 1.0000
0.4000
0.2100
0.1040
0.0521
0.0260
0.0130
0.0065
0.0033
0.0016
0.0008
0.0004
0.0002
0.0001
0.0001
0.0000
0.0000
0.0000
0.0000
0.0000

Example 2.29

To get the same result, using the function filter, we use x =[1 zeros(1,
19)] which creates a vector [1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]:

b=[1 0 0];
a=[1 -0.4 -0.05];
 SOLUTION USING MATLAB FUNCTIONS 91

x=[1 zeros(1,19];
y=filter(b,a,x)

The output is

y = columns 1–7:
1.0000 0.4000 0.2100 0.1040 0.0521 0.0260 0.0130
columns 8–14:
0.0065 0.0033 0.0016 0.0008 0.0004 0.0002 0.0001
columns 15–20:
0.0001 0.0000 0.0000 0.0000 0.0000 0.0000

Example 2.30

Now we consider the use of the function residuez when the transfer function
has multiple poles. Let us choose G(z) from (2.44) and (2.45) and also reduce
it to a rational function in ascending powers of z−1 as shown in (2.80):

z(2z2 − 11z + 12)

G(z) =
(z − 1)(z − 2)3
−2z −z 3z −3z
G(z) = + + + (2.79)
(z − 2) 3 (z − 2)2 (z − 2) (z − 1)
2z3 − 11z2 + 12z
=
z4 − 7z3 + 18z2 − 20z + 8
2z−1 − 11z−2 + 12z−3
= (2.80)
1 − 7z−1 + 18z−2 − 20z−3 + 8z−4
The program used to obtain the partial fraction expansion is

b=[0 2 -11 12];

a=[1 -7 18 -20 8];
[r,p,k]=residuez(b,a)

and the following is the output data we get:

r= 3.0000 + 0.0000i
0.5000 − 0.0000i
− 0.5000
− 3.0000
92 TIME-DOMAIN ANALYSIS AND z TRANSFORM

p= 2.0000 + 0.0000i
2.0000 − 0.0000i
2.0000
1.0000
k=[]

From these data, we construct the partial fraction expansion as

3 0.5 0.5 3
G1 (z) = −1
+ −1
− −1
−
(1 − 2z ) (1 − 2z ) 2 (1 − 2z ) 3 (1 − z−1 )
which can be reduced to the equivalent expression

3z 0.5z2 0.5z3 3z
G(z) = + − −
(z − 2) (z − 2) 2 (z − 2) 3 (z − 1)
which differs from the partial fraction expansion shown in (2.45) or (2.79). But
let us expand  
0.5z2 z 0.5z
= +
(z − 2)2 (z − 2)2 (z − 2)

and
0.5z3 −2z 2z 0.5z
− = − −
(z − 2)3 (z − 2)3 (z − 2)2 (z − 2)

Substituting these expressions in the preceding form for G(z), we get

−2z z 3z 3z
G(z) = − + −
(z − 2)3 (z − 2)2 (z − 2) (z − 1)
which is exactly the same as the form obtained in (2.79).

Example 2.31

We can use a MATLAB function deconv(b,a) to find a few values in the inverse
z transform of a transfer function, and it is based on the recursive formula given
by (2.65). Let us select the transfer function (2.67) to illustrate this function.
%MATLAB program to find a few samples of the inverse z transform

b = [0.1 0.25 0];

a = [1 0.4 0.5];
n = 5;
b= [b zeros(1, n-1)];
[x,r] = deconv(b,a)
 SUMMARY 93

where
x = 0.1000 0.2100 − 0.1340 − 0.0514
r =0 0 0 0 0.0876 0.0257

From these output data, we get

0.1 + 0.25z−1
X(z) = = 0.1 + 0.21z−1 − 0.134z−2 − 0.0514z−3
1 + 0.4z−1 + 0.5z−2
0.086z−4 − 0.0257z−5
+ .
1 + 0.4z−1 + 0.5z−2

Therefore we get x0 = 0.1, x1 = 0.21, x2 = −0.134, x3 = −0.0514, which agrees

with the result obtained from long division, by hand calculation. Note that the
vector b has to be augmented by (n − 1) zeros in the above program above, as
pointed out by Ifeachor and Jervis [6].
Students may find it useful to know the following additional MATLAB func-
tions in their analysis of discrete-time systems, in addition to those used in the
examples above presented. Given a vector of zeros, the coefficients of the poly-
nomial having these zeros is obtained by the function poly. A complex number
entered as a zero must be accompanied by its conjugate so that the coefficients
become real. Given the coefficients of the polynomial in a row vector, its zeros
are found from the function roots. The poles and zeros of a rational function
F (z) are plotted in the z plane by the function zplane. Two other functions
that may be interesting to the students are tfdata and tf. Typing the com-
mands help poly, help roots, help zplane, help tfdata, and help tf
will display the details for using these commands. A list of all MATLAB func-
tions available in the Signal Processing Toolbox is displayed when the command
help signal is typed in the command window and is given in the MATLAB
primer in Chapter 9. Typing Type functionname displays the MATLAB code
as well as the help manual for the function where functionname is the name of
the function. Using the help command, students become familiar with and pro-
ficient in the use of MATLAB functions that are available for conducting many
tasks in the analysis and design of discrete-time systems. It is only by trying as
many functions in MATLAB and the Signal Processing Toolbox as possible that
one becomes familiar with and proficient in their use, and the books by Ingle and
Proakis [9] and Mitra [10] are highly recommended for this purpose, in addition
to the functions we have included in this textbook.

2.10 SUMMARY

In this chapter, we have described several ways of modeling linear shift-invariant

discrete-time systems, highlighting that we should learn how to obtain the one
94 TIME-DOMAIN ANALYSIS AND z TRANSFORM

model from other models that is appropriate for solving a given problem in the
time-domain analysis of the system. The recursive algorithm and the convolution
sum were described first; then the theory and application of z transform was
discussed in detail, for finding the response of the system in the time domain.
In this process, many properties of the z transform of discrete-time signals were
introduced. Some fundamental concepts and applications that we discussed in
this chapter are (1) using a recursive algorithm to find the output in the time
domain, due to a given input and initial conditions; (2) finding the output (zero
input response, zero state response, natural response, forced response, transient
response, steady-state response, etc.) of a discrete-time system from a linear
difference equation (or set of equations), using the z transform; (3) finding the
transfer function and the unit impulse response of the system; and (4) finding
the output due to any input by means of convolution sum. We also showed the
method for obtaining the single input–output relation from the transfer function
and then solving for the zero input and zero state response by introducing the
initial conditions of the output into the linear difference equation.
The concept of stability and a procedure for testing the stability of a discrete-
time system was discussed in detail and followed by a description of many
MATLAB functions that facilitate the time-domain analysis of such systems. In
the next chapter, we consider the analysis of these systems in the frequency
domain, which forms the foundation for the design of digital filters.

PROBLEMS

2.1 Given a linear difference equation as shown below, find the output y(n)
for 0 ≤ n ≤ 5, using the recursive algorithm

y(n) = 0.3y(n − 1) + y(n − 2) + x(n)

where y(−1) = 1.0, y(−2) = 0 and x(n) = (0.1)n u(n).

2.2 An LTI-DT system is described by the following equation

y(n) = 0.3y(n − 1) + y(n − 2) − 0.2y(n − 3) + x(n)

where y(−1) = 1, y(−2) = 0, y(−3) = 2, and x(n) = (0.5)n u(n). Find

the output samples y(n) for 0 ≤ n ≤ 5, using the recursive algorithm.
2.3 An LTI-DT system is described by the recursive equation

y(n) = −0.5y(n − 1) + 0.06y(n − 2) + x(n)

where y(−1) = 0; y(−2) = 0 and x(n) = cos(0.5πn)u(n). Find the out-

put y(n) for 0 ≤ n ≤ 5, using the recursive algorithm.
 PROBLEMS 95

2.4 An LTI-DT system is described by the difference equation

y(n) + 0.5y(n − 1) + 0.06y(n − 2) = 2x(n) − x(n − 1)

where y(−1) = 1.5, y(−2) = −1.0, and x(n) = (0.2)n u(n). Find the out-
put sample y(4) using the recursive algorithm.
2.5 What are the (a) zero state response, (b) zero input response, (c) natu-
ral response, (d) forced response, (e) transient response, (f) steady-state
response, and (g) unit impulse response of the system described in Prob-
lem 2.4?
2.6 Given an input sequence x(−3) = 0.5, x(−2) = 0.1, x(−1) = 0.9, x(0) =
1.0, x(1) = 0.4, x(2) = −0.6, and h(n) = (0.8)n u(n), find the output y(n)
for −5 ≤ n ≤ 5, using the convolution sum.
2.7 Find the samples of the output y(n) for 0 ≤ n ≤ 4, using the convolution
sum y(n) = x(n) ∗ h(n), where x(n) = {1.0 0.5 − 0.2 0.4 0.4}
↑
and h(n) = (0.8)n u(n).
2.8 Given an input sequence x(n) = {−0.5 0.2 0.0 0.2 − 0.5} and
↑
the unit impulse response h(n) = {0.1 −0.1 0.1 − 0.1}, find the
↑
output using the convolution sum, for 0 ≤ n ≤ 6.
2.9 Given an input x(n) = (0.5)n u(n) and h(n) = (0.8)n u(n), find the output
y(n) for 0 ≤ n ≤ 4, using the convolution sum formula and verify that
answer by using the z transforms X(z) and H (z).
2.10 When x(n) = {1.0 0.5 − 0.2 0.4 0.4}, and h(n) = (0.8)n u(n),
↑
find the output y(n) for 0 ≤ n ≤ 6, using the convolution formula.
2.11 Find the output y(n) using the convolution sum formula, y(n) = v(n) ∗
x(n), where v(n) = (−1)n u(n) and x(n) = (−1)n u(n).
2.12 Find the output sample y(3), using the convolution sum formula for
y(n) = x(n) ∗ h(n),where x(n) = e0.5n u(n) and h(n) = e−0.5n u(n).
2.13 Find the output y(5), using the convolution sum, when an LTI-DT system
defined by h(n) = (0.5)n u(n) is excited by an input x(n) = (0.2)n ; 2 ≤
n ≤ ∞.
2.14 Given h(n) = (−1)n u(n) and x(n) = {0.1 0.2 0.3 0.4 0.5 0.6},
↑
find the value of y(n) = x(n) ∗ h(n) at n = 3, from the convolution sum.
2.15 An LTI, discrete-time system is defined by its h(n) = (0.8)n u(n). Find
the output y(n) for n = 1, 2, 3, 4, when the input is given by x(n) =
{1.0 0.5 − 0.5 0.2 0.2 0.4 0.6 0.8}, using the convolution
↑
sum.
96 TIME-DOMAIN ANALYSIS AND z TRANSFORM

2.16 (a) Plot the output y(n) for −3 ≤ n ≤ 3, when x(n) = {1.0 0.5 0.0
↑
0.5 1.0} is convolved with h(n) = (−1)n u(n).
(b) Plot the output y(n) for −4 ≤ n ≤ 4, when x(n) = (−1)n u(−n + 3)
is convolved with h(n) = (−1)n u(n − 2).
2.17 The input sequence is x(n) = {1.0 − 0.5 1.0 −0.5 1.0 − 0.5
↑
1.0 − 0.5} and the unit pulse response h(n) = {0.1 0.2 0.3}. Find
↑
the output sample y(1) and y(4), using the convolution sum formula.
2.18 Show that the z transform of x(n) = (n + 1)a n u(n) is X(z) =
z2 /(z − a)2
2.19 Find the z transform of the following sequences:
(a) x1 (n) = (0.1)n−3 u(n)
(b) x2 (n) = (0.1)n u(n − 3)
(c) x3 (n) = e−j πn cos(0.5πn)u(n)
2.20 Find the z transform of the following two functions:
(a) x1 (n) = n(0.5)n−2 u(n)
(b) x2 (n) = (0.5)n u(n − 2)
2.21 Find the z transform of the following two functions:
(a) x1 (n) = −na n u(−n − 1)
(b) x2 (n) = (−1)n cos( π3 n)u(n)
2.22 Find the z transform of the following functions:
(a) x1 (n) = (−1)n 2−n u(n)
(b) x2 (n) = na n sin(ω0 n)u(n)
(c) x3 (n) = (n2 + n)a n−1 u(n − 1)
(d) x4 (n) = (0.5)n [u(n) − u(n − 5)]
2.23 Show that
⎧
⎨N when z = 1
−1 −2 −(N −1)
X(z) = 1 + z +z +··· +z = 1 − z−N
⎩ when z = 1
1 − z−1
2.24 Find the z transform of an input x(n) = (−1)n [u(n − 4) − u(n −
8)]. When an LTI, discrete-time system, defined by its h(n) = {1.0
↑
0.8 0.6 0.4}, is excited by this x(n), what is the output y(n); n ≥ 0?
2.25 An LTI discrete-time system has an unit pulse response h(n) =
(0.1)n u(n). What is its output y(n) when it is excited by an input
x(n) = (n + 1)(0.5)n u(n)?
2.26 Find the inverse z transform of H (z) = 0.3z + 1.0/[(z + 0.5)(z +
0.2)2 (z + 0.3)].
 PROBLEMS 97

2.27 Find the inverse z transform of H (z) = 0.6z/[(z + 0.1)(z − 0.5)3 ].

2.28 Find f (n) from F (z) = (z + 0.3)/z(z + 0.5)2 .
2.29 Find the inverse z transform of X(z) = z(z + 0.5)/(z2 + 0.6z + 0.5).
2.30 Find the inverse z transform of X(z) = (z + 0.2)/[(z + 0.5)(z − 1)
(z − 0.1)].
2.31 Find the inverse z transform of Y (z−1 ) = (z−1 + 0.4z−2 )/(1 + z−1 +
0.25z−2 ).
2.32 Find the inverse z transforms of the following two transfer functions:

z + 0.6
H1 (z) =
+ 0.8z + 0.5)(z − 0.4)
(z2
(z + 0.4)(z + 1)
H2 (z) =
(z − 0.5)2

2.33 Find the inverse z transform of H (z) = z/[(z + 0.5)2 (z2 + 0.25)].
2.34 Find the inverse z transform of H (z) = [0.1z(z + 1)]/[(z − 1)(z2 − z +
0.9)].
2.35 Find the inverse z transform of F (z) = (z + 0.5)/z(z2 + 0.2z + 0.02).
2.36 Find the inverse z transform of the following two functions:

1 + 0.1z−1 + 0.8z−2
G1 (z) =
(1 + z−1 )
0.2z2 + z + 1.0
G2 (z) =
(z + 0.2)(z + 0.1)

2.37 Find the inverse z transform of X(z) = (3.0 + 1.35z−1 + 0.28z−2 +

0.03z−3 )/(1.0 + 0.5z−1 + 0.06z−2 ).
2.38 Show that the inverse z transform of H (z) = 1/[1 − 2r(cos θ )z−1 +
r 2 z−2 ] is given by

r n sin(n + 1)θ
h(n) = u(n)
sin θ

2.39 Show that the inverse z transform of H (z) = z/(z − a)3 is given by

n(n − 1)a n−2

h(n) = u(n − 2)
2
98 TIME-DOMAIN ANALYSIS AND z TRANSFORM

2.40 Given an LTI-DT system described by

y(n) + 0.25y(n − 2) = x(n − 6)

where y(−1) = 1, y(−2) = 0, and x(n) = u(n), find the (a) zero state
response, (b) zero input response, (c) natural response, (d) forced
response, (e) transient response, and (f) steady-state response of the
system.
2.41 Given an LTI discrete-time system described by

y(n) + 0.2y(n − 1) + 0.2y(n − 2) = 0.5x(n − 1)

where y(−1) = 1, y(−2) = −2, and x(n) = (−0.3)n u(n), find the
(a) zero state response, (b) zero input response, (c) natural response,
(d) forced response, (e) transient response, and (f) steady-state response
of the system above. What is the unit pulse response h(n) of this system?
2.42 An LTI discrete-time system is described by its difference equation
y(n) − 0.09y(n − 2) = u(n), where y(−1) = 1 and y(−2) = 0. Find its
(a) zero state response, (b) zero input response, (c) natural response,
(d) forced response, (e) transient response, and (f) steady-state response,
and (g) the unit pulse response.
2.43 Given an LTI discrete-time system described by

y(n) = −0.2y(n − 1) + 0.3y(n − 2) + 0.1y(n − 3) + 0.5x(n) + x(n − 1)

where y(−1) = 1, y(−2) = 1, y(−3) = 2, and x(n) = (0.5)n u(n), find

its transfer function H (z).
2.44 Given an LTI discrete-time system described by the difference equation

y(n) + 0.6y(n − 1) + 0.25y(n − 2) = x(n) + 0.04x(n − 2)

where y(−1) = 0, y(−2) = 0.4, and x(n) = (−1)n u(n), find the (a) nat-
ural response, (b) forced response, (c) transient response, and (d) steady-
state response of the system.
2.45 Given an LTI-DT system defined by the difference equation

y(n) − 0.5y(n − 1) + 0.06y(n − 2) = u(n)

and y(−1) = y(−2) = 0, find its (a) natural response, (b) forced
response, (c) transient response, and (d) steady-state response, when it is
excited by x(n) = u(n). What is its unit impulse response h(n)?
 PROBLEMS 99

2.46 Find the total response y(n) of the LTI-DT system defined by the fol-
lowing difference equation

y(n) + 0.25y(n − 1) + y(n − 2) = (0.5)n u(n)

where y(−1) = 1 and y(−2) = −1

2.47 Find the total response y(n) for the LTI system given by y(n) + 1.4y(n −
1) + 0.44y(n − 2) = 0.5δ(n − 2), where y(−1) = 1 and y(−2) = 0.5
are the initial states.
2.48 Repeat Problem 2.47 for the system described by the difference equation

y(n) + 0.5y(n − 1) + 0.04y(n − 2) = x(n)

where y(−1) = 0, y(−2) = 0, and x(n) = {1.0 0.5 − 1.0}.

↑

2.49 Solve the following difference equation for y(n), n ≥ 0

y(n) + 0.6y(n − 1) − 0.4y(n − 2) = 2x(n − 2)

where y(−1) = 1, y(−2) = 0.5, and x(n) = (0.1)n u(n).

2.50 Given an LTI, discrete-time system described by the difference equation

y(n) + 0.4y(n − 1) + 0.04y(n − 2) = x(n) − 0.5x(n − 1)

where y(−1) = 2, y(−2) = 2, and x(n) = (e−0.1n )u(n), find its unit pulse
response h(n).
2.51 The difference equation describing an LTI discrete-time system is given
below. Solve for y(n)

y(n) + 0.4y(n − 1) + 0.03y(n − 2) = x(n − 2)

where y(−1) = 1, y(−2) = 1, and x(n) = (0.5)n u(n).

2.52 Find the total response y(n) of the discrete-time system described by the
following difference equation

y(n) − 0.3y(n − 1) + 0.02y(n − 2) = x(n) − 0.1x(n − 1)

where y(−1) = 0, y(−2) = 0, and x(n) = (−0.2)n u(n).

2.53 Repeat Problem 2.52, assuming that the system is described by the dif-
ference equation

y(n) − 0.04y(n − 2) = x(n − 1)

where y(−1) = −0.2, y(−2) = 1.0, and x(n) = (0.2)n u(n).

100 TIME-DOMAIN ANALYSIS AND z TRANSFORM

2.54 An LTI discrete-time system is described by the following difference

equation

y(n) + 0.25y(n − 2) = x(n − 1)

where y(−1) = 0, y(−2) = 2, and x(n) = (0.5)n u(n). Calculate the

(a) zero state response, (b) zero input response, (c) natural response,
(d) forced response, (e) transient response, and (f) steady-state response
of the system.
2.55 Given an LTI discrete-time system described by the difference equation

y(n) − 0.5y(n − 1) = x(n) + 0.5x(n − 1)

where y(−1) = 2 and x(n) = (0.5)n u(n), find y(n) and also the unit
impulse response h(n).
2.56 Given the transfer function H (z) = z/[(z − 1)2 (z + 1)] of a digital filter,
compute and plot the values of h(n) for n = 0, 1, 2, 3, 4, 5. What is the
value of limn→∞ h(n)?
2.57 Given the input X(z−1 ) = 1.0 + 0.1z−1 + 0.2z−2 and the transfer func-
tion H (z) = z/[(z − 0.2)(z + 0.3)], find the output y(n).
2.58 If the z transform of y(n) = x(n) ∗ h(n) is X(z)H (z), what is the con-
volution sum formula for x(−n) ∗ h(n)? What is the z transform of
x(−n) ∗ h(n)?
2.59 Given an LTI discrete-time system described by the difference equation

y(n) = 4 cos(0.4)y(n − 1) − 4y(n − 2) + x(n)

find h(n) and the zero state response when x(n) = u(n).
2.60 Derive the transfer function H (z) of the LTI discrete-time system
described by the circuit shown in Figure 2.8.
2.61 Derive the transfer function H (z) of the LTI-DT system described by the
circuit given in Figure 2.9. Obtain the difference equation relating the
input x(n) to the output y(n).
2.62 Derive the single input–single output relationship as a difference equation
for the LTI-DT system shown in Figure 2.10.
2.63 Obtain the transfer function H (z) = Y3 (z)/X(z) as the ratio of polyno-
mials, for the discrete-time system shown in Figure 2.11.
2.64 Write the equations in the z domain to describe the LTI-DT system shown
in Figure 2.12. and find the z transform Y2 (z).
 PROBLEMS 101

2.0
z−1

X(z)
z−1 Σ Σ Y(z)
1.5
z−1

−0.2

z−1

−0.5

Figure 2.8 Problem 2.60.

X(n)
Σ Σ z−1 Σ Y(n)

z−1
1.0
z−1
−1 0.5 0.8
z

−0.5

Figure 2.9 Problem 2.61.

Σ
1.0
−1
z

Σ z−1 Σ Y(n)
X(n) −0.5
0.5

z−1

0.3

Figure 2.10 Problem 2.62.

102 TIME-DOMAIN ANALYSIS AND z TRANSFORM

2.0
X1(z) z−1 Σ z−1 Σ Y2(z)

z−1
z−1

−0.1

−0.2 z−1

−0.4

Figure 2.11 Problem 2.63.

2.0
X1(z) z−1 Σ z−1 Σ Y2(z)

z−1
z−1

−0.1

−0.2 z−1

−0.4

Figure 2.12 Problem 2.64.

1.5
X(n) z−1 Σ z−1 z−1 Y(n)

0.4

−0.3

Figure 2.13 Problem 2.65.

2.65 Derive the transfer function H (z) for the circuit shown in Figure 2.13
and find its unit impulse response h(n).
2.66 Write the equations in the z domain to describe the LTI-DT system given
in Figure 2.14 and derive the transfer function H (z) = Y3 (z)/X(z), as a
ratio of two polynomials.
 PROBLEMS 103

0.5 −0.5

1.5
X(z) Σ Σ z−1 z−1 Σ Y3(z)
1.5

z−1
−0.6

−0.1 1.2 0.3

Figure 2.14 Problem 2.66.

0.4
z−1

0.5
X(z) Σ Σ Σ Y(z)

z−1 z−1

−0.1 0.3 −0.2 0.2

z−1 z−1

−0.4 0.1

Figure 2.15 Problem 2.67.

X(n) Σ

z−1
1.0

−0.1
z−1 Σ Y(n)

0.02 0.4

Figure 2.16 Problem 2.68.

2.67 Repeat Problem 2.66 for the circuit given in Figure 2.15.
2.68 Find the unit pulse response of the LTI-DT system shown in Figure 2.16.
2.69 Find the unit pulse response h(n) of the discrete-time system shown in
Figure 2.17.
104 TIME-DOMAIN ANALYSIS AND z TRANSFORM

X(z) z−1 z−1

−1.0 V2(z)
Σ
V1(z)
Σ

−1.5 0.5

z−1 z−1

−0.5
−0.2 0.4

Σ
V3(z)
Y(z)

Figure 2.17 Problem 2.69.

1.0

−0.1 0.5
X(z) z−1 ∑

∑ Y(z)
z−1

0.3 0.2
z−1

−0.02

Figure 2.18 Problem 2.70.

2.70 Find the transfer function H (z) of the discrete-time system given in
Figure 2.18.
2.71 Derive the transfer function of the digital filter shown in Figure 2.19 and
find the samples h(0), h(1), and h(2).
2.72 Derive the transfer function H (z) for the digital filter shown in Figure 2.20
and find its unit impulse response h(n).
2.73 Find the unit sample response h(n) of the discrete-time system shown in
Figure 2.21.
 PROBLEMS 105

Σ z−1

−0.1 −0.2

X(z) Σ Y(z)

−1.0
0.2 0.1

Σ z−1

Figure 2.19 Problem 2.71.

Σ z−1 Σ

−0.1 −0.5
z−1

X(z) z−1

0.1
0.5 0.2

Σ z−1 Y(z)

Figure 2.20 Problem 2.72.

Σ z−1 z−1

−0.4

−0.03 Σ Y(n)

X(n)

Σ z−1 z−1

−0.5

Figure 2.21 Problem 2.73.

106 TIME-DOMAIN ANALYSIS AND z TRANSFORM

1.5
z−1

X(z) Σ Σ Y(z)
1.8

z−1

0.75 −2.0

Σ z−1 z−1
−1.0

z−1

1.5

Figure 2.22 Problem 2.74.

2.74 Derive the transfer function H (z) = Y (z)/X(z) for the LTI-DT system
shown in Figure 2.22.
 −1
2.75 A moving-average filter is defined by y(n) = 1/N N k=0 y(n − k). Find
the transfer function of the filter when N = 10.

2.76 In the partial fraction expansion of H (z) = N (z)/ K k=1 (z − zk ) =
K
R
k=1 k /(z − zk ), which has simple poles at z = zk , show that the

residues Rk can be found from the formula Rk = N (zk )/D (zk ), where

D (z) = dD(z)/dz.
2.77 The transfer function H (z) is expanded into its partial fraction form as
shown below:
z
H (z) =
(z − 0.1)(z − 0.2)(1 − 0.3z−1 )(1 − 0.5z−1 )
K1 z K2 z R3 R4
= + + +
(z − 0.1) (z − 0.2) (1 − 0.3z−1 ) (1 − 0.5z−1 )

Find the values of K1 , K2 , R3 , R4 .

2.78 (a) If the transfer function H (z−1 ) is expanded into the form as shown
below, find the values of R1 and R2 :
 PROBLEMS 107

1
H (z−1 ) =
(1 − 0.5z−1 )(1
− 0.1z−1 )
R1 R2
= +
(1 − 0.5z−1 ) (1 − 0.1z−1 )

(b) If we are given a general transfer function of the form

N (z−1 )
N
Rn
H (z−1 ) = N =
n=1 (1 − an z−1 ) (1 − an z−1 )
n=1

what is the general method for finding the residues Rn ? What is the
unit impulse response h(n) of this system?
2.79 (a) In the expression given below, find the values of K1 and K2 and find
h(n):

1 K1 K2
H (z−1 ) = = −1 + −1
(z−1 −1
− 0.5)(z − 0.1) (z − 0.5) (z − 0.1)

(b) If we are given a transfer function of the form

N (z−1 )
N
Kn
−1
H (z ) = N =
n=1 (z−1 − an ) (z−1 − an )
n=1

what is the general expression for finding Kn ?

2.80 Given the three difference equations describing an LTI discrete-time sys-
tem, find Y2 (z)
y1 (n) = 0.1x(n) + 0.2y1 (n − 1) + 0.3y2 (n − 2)
y2 (n) = 0.2y1 (n) + 0.4y3 (n)
y3 (n) = y1 (n − 1) + y3 (n − 1)
and x(n) = δ(n).
2.81 Derive the linear difference equation for the input–output relationship
for the discrete-time system described by its transfer function

z + 0.1
H (z) =
z2 + 0.5z + 0.4
2.82 Derive the linear difference equation for the input–output relationship
for the system with its transfer function H (z)

z(z + 0.4)
H (z) =
z3 + 0.2z2 − 0.4z + 0.05
108 TIME-DOMAIN ANALYSIS AND z TRANSFORM

2.83 Given the transfer function for an LTI-DT system as

z−1
H (z−1 ) =
(1 + 0.3z−1 + 0.02z−2 )

find the zero state response when the input is a unit step function. What
is the zero input response that satisfies the initial conditions y(−1) = 2
and y(−2) = 4?
2.84 Use the Jury–Marden test to determine whether the discrete-time system
defined by the following transfer function is stable:

z + 0.5
H (z) =
z3 + z2 + 2z + 5

2.85 Determine whether the polynomial D(z) = 2 + 2z−1 + 1.7z−2 +

0.6z−3 + 0.1z−4 has all four zeros inside the unit circle |z| = 1, using
the Jury–Marden test.
2.86 Determine whether all the zeros of the polynomial

Q(z) = 2 + 5z + 8z2 + 7z3 + 2z4

are inside the unit circle |z| = 1, using the Jury–Marden test.
2.87 Determine whether the three zeros of the polynomial

P (z) = z3 + 2z2 + 4z + 6

are inside the unit circle |z| = 1, using the Jury–Marden test.
2.88 Apply the Jury–Marden test to determine whether the polynomial has its
zeros inside the unit circle in the z plane:

R(z) = 1 − 1.5z−1 − 0.5z−2 + z−3 .

MATLAB Problems
2.89 Find the roots of the following two polynomials:

N1 (z) = 1 − 0.7z−1 +)0.1725z−2 + 0.1745z−3 − 0.4425z−4

D1 (z) = 1 + 0.8z−1 0.8775z−2 + 0.4333z−3 − 0.1808z−4 − 0.6639z−5

2.90 Plot the poles and zeros of the transfer function H1 (z) = N1 (z)/D1 (z),
where N1 (z) and D1 (Z) are the polynomials given above.
 PROBLEMS 109

2.91 Find the polynomials that have the zeros given below and also the product
of the two polynomials N2 (z)D2 (z):

Zeros of N2 (z) : −0.2 0.3 + j 0.4 0.3 − j 0.4 0.5

Zeros of D2 (z) : 0.4 0.4 0.2 + j 0.2 0.0 − j 0.2

2.92 Plot the poles and zeros of H2 (z) = N2 (z)/D2 (z) in the z plane.
2.93 Find the values of R1 , R2 , and R3 in the expansion of the transfer function
G(z), using the MATLAB function residuez:
(1 + 0.6z) R1 z R2 z R3 z
G(z) = = + +
(z − 0.8)(z + 0.5)2 (z − 0.8) (z + 0.5)2 (z + 0.5)
2.94 Find the values of K1 , K2 , K3 , K4 , K5 in the expansion of the following
transfer functions, using the MATLAB function residuez:
(z − 0.3)
H1 (z) =
(z − 0.2)3 (z + 0.4)(z + 0.5)
K1 z K2 z K3 z K4 z K5 z
= + + + +
(z − 0.2)3 (z − 0.2)2 (z − 0.2) (z + 0.4) (z + 0.5)
(z)2
H2 (z) =
(z + 0.5)2 (z + 0.1)2 (z − 0.2)
K1 z K2 z K3 z K4 z K5 z
= + + + +
(z + 0.5)2 (z + 0.5) (z + 0.1) 2 (z + 0.1) (z − 0.2)
2.95 Plot the magnitude, phase, and group delay of the transfer function
H1 (z−1 ) given below:

0.20 − 0.45z−1
H1 (z−1 ) =
1 − 1.3z−1 + 0.75z−2
2.1 + 1.45z−1 1.8 − 0.60z−1
+ +
1 − 1.07z−1 + 0.30z−2 1 − z−1 + 0.25z−2

2.96 Given H2 (z) = (1 − z−1 )/(1 − 0.9z−1 ), plot the magnitude of H3 (z) =
H2 (zej 1.5 )H2 (ze−j 1.5 ) and the magnitude of H4 (z) = H2 (zej 1.5 ) +
H2 (ze−j 1.5 ).
2.97 Find the partial fraction expansion of the following two transfer functions
and evaluate their unit pulse response for 0 ≤ n ≤ 10:
z(z − 0.5)
H1 (z) =
(z − 0.8)(z − +0.6)
(z − 0.6)
H2 (z) =
(z + 0.6)(z2 + 0.8z + 0.9)
110 TIME-DOMAIN ANALYSIS AND z TRANSFORM

2.98 Repeat Problem 2.97 with H3 (z) as follows:

(z − 0.5)
H3 (z) =
(z + 0.4)(z + 0.2)2

2.99 Find the output y1 (n), y2 (n), and y3 (n) for 0 ≤ n ≤ 15 of the LTI-DT
systems defined by the preceding transfer functions H1 (z), H2 (z), and
H3 (z), respectively, assuming that they are excited by an input sequence
x(n) = {0.5 0.2 − 0.3 0.1}.
↑

Write your code using the MATLAB function filter, and submit it with
the computer output.
2.100 An LTI-DT system is described by the following difference equation

y(n) + 3y(n − 1) + 2y(n − 2) + y(n − 3) = x(n) + 3x(n − 2)

where y(−1) = 1, y(−2) = 2, y(−3) = 1, and x(n) = (0.5)n u(n). Find

the total response y(n) for 0 ≤ n ≤ 20 and plot y(n).
2.101 Find the pulse responses x1 (n) of X1 (z), h2 (n) of H2 (z), and y3 (n) of
Y3 (z) = X1 (z)H2 (z). Convolve the first 9 samples of x1 (n) with the 9
samples of h2 (n) and compare the result with the first 9 samples of y3 (n):
z
X1 (z) =
(z + 0.2)(z + 0.5)
(z + 0.2)
H2 (z) =
z(z + 0.1)

REFERENCES

1. E. I. Jury, Theory and Applications of the z-Transform Method, Wiley, 1964.

2. S. K. Mitra, Digital Signal Processing—A Computer-Based Approach, McGraw-Hill,
2001.
3. B. P. Lathi, Signal Processing and Linear Systems, Berkeley Cambridge Press, 1998.
4. A. Antoniou, Digital Filters, Analysis, Design and Applications, McGraw-Hill, 1993.
5. A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing, Prentice-
Hall, 1989.
6. E. C. Ifeachor and B. W. Jervis, Digital Signal Processing, A Practical Approach,
Prentice-Hall, 2002.
7. S. K. Mitra and J. F. Kaiser, eds., Handbook for Digital Signal Processing, Wiley-
Interscience, 1993.
8. B. A. Shenoi, Magnitude and Delay Approximation of 1-D and 2-D Digital Filters,
Springer-Verlag, 1999.
 REFERENCES 111

9. V. K. Ingle and J. G. Proakis, Digital Signal Processing Using MATLAB (R) V.4, PWS
Publishing, 1997.
10. S. K. Mitra, Digital Signal Processing Laboratory Using MATLAB, McGraw-Hill,
1999.
11. J. G. Proakis and D. G. Manolakis, Digital Signal Processing, Prentice-Hall, 1996.
 CHAPTER 3

Frequency-Domain Analysis

3.1 INTRODUCTION

In the previous chapter, we derived the definition for the z transform of a discrete-
time signal by impulse-sampling a continuous-time signal xa (t) with a sampling
period T and using the transformation z = esT . The signal xa (t) has another
equivalent representation in the form of its Fourier transform X(j ω). It contains
the same amount of information as xa (t) because we can obtain xa (t) from X(j ω)
as the inverse Fourier transform of X(j ω). When the signal xa (t) is sampled
∞ a sampling period T , to generate the discrete-time signal represented by
with
k=0 xa (kT )δ(nT − kT ), the following questions need to be answered:

Is there an equivalent representation for the discrete-time signal in the fre-

quency domain?
Does it contain the same amount of information as that found in xa (t)? If so,
how do we reconstruct xa (t) from its sample values xa (nT )?
Does the Fourier transform represent the frequency response of the system
when the unit impulse response h(t) of the continuous-time system is sam-
pled? Can we choose any value for the sampling period, or is there a limit
that is determined by the input signal or any other considerations?

We address these questions in this chapter, arrive at the definition for the discrete-
time Fourier transform (DTFT) of the discrete-time system, and describe its prop-
erties and applications. In the second half of the chapter, we discuss another trans-
form known as the discrete-time Fourier series (DTFS) for periodic, discrete-time
signals. There is a third transform called discrete Fourier transform (DFT), which
is simply a part of the DTFS, and we discuss its properties as well as its applica-
tions in signal processing. The use of MATLAB to solve many of the problems
or to implement the algorithms will be discussed at the end of the chapter.

Introduction to Digital Signal Processing and Filter Design, by B. A. Shenoi

Copyright © 2006 John Wiley & Sons, Inc.

112
 THEORY OF SAMPLING 113

3.2 THEORY OF SAMPLING

Let us first choose a continuous-time (analog) function xa (t) that can be repre-
sented by its Fourier transform Xa (j
)1

∞
Xa (j
) = xa (t)e−j
t dt (3.1)
−∞

whereas the inverse Fourier transform of Xa (j

) is given by2

∞
1
xa (t) = Xa (j
)ej
t d
(3.2)
2π −∞
Now we generate a discrete-time sequence x(nT ) by sampling xa (t) with a
sampling period T . So we have x(nT ) = xa (t)|t=nT , and substituting t = nT in
(3.2), we can write

∞
1
xa (nT ) = x(nT ) = Xa (j
)ej
nT d
(3.3)
2π −∞
The z transform of this discrete-time sequence is3
∞

X(z) = x(nT )z−n (3.4)
n=−∞

and evaluating it on the unit circle in the z plane; thus, when z = ej ωT , we get
∞

X(ej ωT ) = x(nT )e−j ωnT (3.5)
n=−∞

Next we consider h(nT ) as the unit impulse response of a linear, time-

invariant, discrete-time system and the input x(nT ) to the system as ej ωnT . Then
the output y(nT ) is obtained by convolution as follows:
∞

y(nT ) = ej ω(nT −kT ) h(kT )
k=−∞
∞
∞

−j ωkT
=e j ωnT
e h(kT ) = e j ωnT
h(kT )e−j ωkT (3.6)
k=−∞ k=−∞

1 The material in this section is adapted from a section with the same heading, in the author’s book
Magnitude and Delay Approximation of 1-D and 2-D Digital Filters [1], with permission from the
publisher, Springer-Verlag.
2
We have chosen
(measured in radians per second) to denote the frequency variable of an analog
function in this section and will choose the same symbol to represent the frequency to represent the
frequency response of a lowpass, normalized, prototype analog filter in Chapter 5.
3
Here we have used the bilateral z transform of the DT sequence, since we have assumed that it
is defined for −∞ < n < ∞ in general. But the theory of bilateral z transform is not discussed in
this book.
114 FREQUENCY-DOMAIN ANALYSIS

Note that the signal ej ωnT is assumed to have values for −∞ < n < ∞ in gen-
h(kT ) is a causal
eral, whereas sequence: h(kT ) = 0 for −∞ < k < 0. Hence the
summation ∞ k=−∞ h(kT )e −j ωkT
in (3.6) can be replaced by ∞
k=0 h(kT )e
−j ωkT
.
j ωT
It is denoted as H (e ) and is a complex-valued function of ω, having a magni-
tude response H (ej ωT ) and phase response θ (ej ωT ). Thus we have the following
result
 
y(nT ) = ej ωnT H (ej ωT ) ej θ(e )
j ωT
(3.7)

which shows that when the input is a complex exponential function ej ωnT ,
the magnitude of the output y(nT ) is H (ej ωT ) and the phase of the output
y(nT ) is (ωnT + θ ). If we choose a sinusoidal input x(nT ) = Re(Aej ωnT ) =
A cos(ωnT ), then the  output y(nT ) is also a sinusoidal function given by
y(nT ) = A H (ej ωT ) cos(ωnt
 + θ ). Therefore we multiply the amplitude of the
sinusoidal input by H (ej ωT ) and increase the phase by θ (ej ωT ) to get the ampli-
tude and phase of the sinusoidal output. For the reason stated above, H (ej ωT )
is called thefrequency response of the discrete-time system. We use a similar
expression ∞ k=−∞ x(kT )e
−j ωkT
= X(ej ωT ) for the frequency response of any
input signal x(kT ) and call it the discrete-time Fourier transform (DTFT) of
x(kT ).
To find a relationship between the Fourier transform Xa (j
) of the continuous-
time function xa (t) and the Fourier transform X(ej ωT ) of the discrete-time
sequence, we start with the observation that the DTFT X(ej ωT ) is a periodic func-
tion of ω with a period ωs = 2π/T , namely, X(ej ωT +j rωs T ) = X(ej ωT +j r2π ) =
X(ej ωT ), where r is any integer. It can therefore be expressed in a Fourier series
form
∞

X(ej ωT ) = Cn e−j ωnT (3.8)
n=−∞

where the coefficients Cn are given by

π/T
T
Cn = X(ej ωT )ej ωT dω (3.9)
2π −(π/T )

By comparing (3.5) with (3.8), we conclude that x(nT ) are the Fourier series
coefficients of the periodic function X(ej ωT ), and these coefficients are evaluated
from

π/T
T
Cn = x(nT ) = X(ej ωT )ej ωnT dω (3.10)
2π −(π/T )

Therefore
∞

X(ej ωT ) = x(nT )e−j ωnT (3.11)
n=−∞
 THEORY OF SAMPLING 115

Let us express (3.3), which involves integration from

= −∞ to
= ∞ as
the sum of integrals over successive intervals each equal to one period
2π/T = ωs :
∞
(2r+1)π
1 T
x(nT ) = Xa (j
)ej
nT d
(3.12)
2π r=−∞ (2r−1)π
T

However, each term in this summation can be reduced to an integral over the
range −(π/T ) to π/T by a change of variable from
to
+ 2πr/T , to get
∞
 
T 1 π/T 2πr
x(nT ) = Xa j
+ j ej
nT ej 2πrn d
(3.13)
2π r=−∞ T −(π/T ) T

Note that ej 2πrn = 1 for all integer values of r and n. By changing the order of
summation and integration, this equation can be reduced to

π/T  ∞  
T 1 2πr
x(nT ) = Xa j
+ j ej
nT d
(3.14)
2π −(π/T ) T r=−∞ T

Without loss of generality, we change the frequency variable

to ω, thereby
getting

π/T  ∞  
T 1 2πr
x(nT ) = Xa j ω + j ej ωnT dω (3.15)
2π −(π/T ) T r=−∞ T

Comparing (3.10) with (3.15), we get the desired relationship:

 ∞  
1 2πr
X(e ) =
j ωT
Xa j ω + j (3.16)
T r=−∞ T

This shows that the discrete-time Fourier transform (DTFT) of the sequence
x(nT ) generated by sampling the continuous-time signal xa (t) with a sampling
period T is obtained by a periodic duplication of the Fourier transform Xa (j ω)
of xa (t) with a period 2π/T = ωs and scaled by T . To illustrate this result,
a typical analog signal xa (t) and the magnitude of its Fourier transform are
sketched in Figure 3.1. In Figure 3.2a the discrete-time sequence generated by
sampling xa (t) is shown, and in Figure 3.2b,
 the magnitude of a few terms of
(3.16) as well as the magnitude X(ej ωT ) are shown.
Ideally the Fourier transform of xa (t) approaches zero only as the frequency
approaches ∞. Hence it is seen that, in general, when Xa (j ω)/T is duplicated
and added as shown in Figure 3.2b, there is an overlap of the frequency responses
at all frequencies. The frequency responses of the individual terms in (3.16) add
up, giving the actual response as shown by the curve for X(ej ω ). [We have
116 FREQUENCY-DOMAIN ANALYSIS

(a)

⎮Xa( jw)⎮

0
(b)

Figure 3.1 An analog signal xa (t) and the magnitude of its Fourier transform X(j ω).

xa(nT)

0 n

(a)

⏐X(e jw)⏐

X(jw)
X( j(w − ws))
T
T
X( j(w − 2ws))
T

0 ws 2ws
(b)

Figure 3.2 The discrete-time signal xa (nT ) obtained from the analog signal xa (t) and
the discrete-time Fourier transform H (ej ω ).
 THEORY OF SAMPLING 117

disregarded the effect of phase in adding the duplicates of X(j ω).] Because of
this overlapping effect, more commonly known as “aliasing,” there is no way of
retrieving X(j ω) from X(ej ω ) by any linear operation; in other words, we have
lost the information contained in the analog function xa (t) when we sample it.
Aliasing of the Fourier transform can be avoided if and only if (1) the function
xa (t) is assumed to be bandlimited—that is, if it is a function such that its
Fourier transform Xa (j ω) ≡ 0 for |ω| > ωb ; and (2) the sampling period T is
chosen such that ωs = 2π/T > 2ωb . When the analog signal xb (t) is bandlimited
as shown in Figure 3.3b and is sampled at a frequency ωs ≥ 2ωb , the resulting
discrete-time signal xb (nT ) and its Fourier transform X(ej ω ) are as shown in
Figure 3.4a,b, respectively.
If this bandlimited signal xb (nT ) is passed through an ideal lowpass filter
with a bandwidth of ωs /2, the output will be a signal with a Fourier transform
equal to X(ej ωT )Hlp (j ω) = Xb (j ω)/T . The unit impulse response of the ideal
lowpass filter with a bandwidth ωb obtained as the inverse Fourier transform of
Hlp (j ω) is given by

∞
1
hlp (t) = Hlp (j ω)ej ωt dω
2π −∞

ωs
1 2
= T ej ωt dω (3.17)
2π −ωs
2
   
ωs t πt
sin sin
2 T
=   =   (3.18)
ωs t πt
2 T

xb(t )

0 t

(a)

⏐Xb( jw)⏐

−wb wb
(b)

Figure 3.3 A bandlimited analog signal and the magnitude of its Fourier transform.
118 FREQUENCY-DOMAIN ANALYSIS

xb(nT )

0 n

(a)

Ideal analog lowpass filter

wt wt
− −wb wb wt
2 2

(b)

Figure 3.4 The discrete-time signal obtained from the bandlimited signal and the mag-
nitude of its Fourier transform.

The output signal will be the result of convolving the discrete input sequence
xb (nT ) with the unit impulse response hlp (t) of the ideal analog lowpass fil-
ter. But we have not defined the convolution between a continuous-time signal
and samples of discrete-time sequence. Actually it is the superposition of the
responses due to the delayed impulse responses hlp (t − nT ), weighted by the
samples xb (nT ), which gives the output xb (t). Using this argument, Shannon [2]
derived the formula for reconstructing the continuous-time function xb (t), from
only the samples x(n) = xb (nT )—under the condition that xb (t) be bandlimited
up to a maximum frequency ωb and be sampled with a period T < π/ωb . This
formula (3.19) is commonly called the reconstruction formula, and the statement
that the function xb (t) can be reconstructed from its samples xb (nT ) under the
abovementioned conditions is known as Shannon’s sampling theorem:
π 
∞ sin (t − nT )
xb (t) = xb (nT )  π T  (3.19)
n=−∞ (t − nT )
T
The reconstruction process is indicated in Figure 3.5a. An explanation of the
reconstruction is also given in Figure
 3.5b, where
 it is seen that the delayed
impulse response sin Tπ (t − nT ) / Tπ (t − nT ) has a value of xb (nT ) at t = nT
and contributes zero value at all other sampling instants t = nT so that the
reconstructed analog signal interpolates exactly between these sample values of
the discrete samples.
 THEORY OF SAMPLING 119

Ideal Lowpass analog filter hf(t )

xb (nT ) xb(t ) = xb(nT )∗hf(t )

Hf( jw)
Xb(e jωT) Xb(jw) = Xb(e jwT)Hf(jw)

(a)

x(0) Reconstructed signal

x(T)

−T 0 T 2T 3T 4T
(b)

Figure 3.5 Reconstruction of the bandlimited signal from its samples, using an ideal
lowpass analog filter.

This revolutionary theorem implies that the samples xb (nT ) contain all the
information that is contained in the original analog signal xb (t), if it is bandlim-
ited and if it has been sampled with a period T < π/ωb . It lays the necessary
foundation for all the research and developments in digital signal processing
that is instrumental in the extraordinary progress in the information technology
that we are witnessing.4 In practice, any given signal can be rendered almost
bandlimited by passing it through an analog lowpass filter of fairly high order.
Indeed, it is common practice to pass an analog signal through an analog lowpass
filter before it is sampled. Such filters used to precondition the analog signals
are called as antialiasing filters. As an example, it is known that the maximum
frequency contained in human speech is about 3400 Hz, and hence the sampling
frequency is chosen as 8 kHz. Before the human speech is sampled and input to
telephone circuits, it is passed through a filter that provides an attenuation of at
least 30 dB at 4000 Hz. It is obvious that if there is a frequency above 4000 Hz
in the speech signal, for example, at 4100 Hz, when it is sampled at 4000 Hz,
due to aliasing of the spectrum of the sampled signal, there will be a frequency
at 4100 Hz as well as 3900 Hz. Because of this phenomenon, we can say that
the frequency of 4100 Hz is folded into 3900 Hz, and 4000 Hz is hence called
the “folding frequency.” In general, half the sampling frequency is known as the
folding frequency (expressed in radians per second or in hertz).

4
This author feels that Shannon deserved an award (such as the Nobel prize) for his seminal contri-
butions to sampling theory and information theory.
120 FREQUENCY-DOMAIN ANALYSIS

There is some ambiguity in the published literature regarding the definition of

what is called the Nyquist frequency. Most of the books define half the sampling
frequency as the Nyquist frequency and 2fb as the Nyquist rate, which is the
minimum sampling rate required to avoid aliasing. Because of this definition for
the Nyquist rate, some authors erroneously define fb as the Nyquist frequency.
In our example, when the signal is sampled at 8 kHz, we have 4 kHz as the
Nyquist frequency (or the folding frequency) and 6.8 kHz as the Nyquist rate.
If we sample the analog signal at 20 kHz, the Nyquist frequency is 10 kHz, but
the Nyquist rate is still 6.8 kHz. We will define half the sampling frequency as
the Nyquist frequency throughout this book. Some authors define the Nyquist
frequency as the bandwidth of the corresponding analog signal, whereas some
authors define 2fb as the bandwidth.

3.2.1 Sampling of Bandpass Signals

Suppose that we have an analog signal that is a bandpass signal (i.e., it has a
Fourier transform that is zero outside the frequency range ω1 ≤ ω ≤ ω2 ); the
bandwidth of this signal is B = ω2 − ω1 , and the maximum frequency of this
signal is ω2 . So it is bandlimited, and according to Shannon’s sampling theorem,
one might consider a sampling frequency greater than 2ω2 ; however, it is not
necessary to choose a sampling frequency ωs ≥ 2ω2 in order to ensure that we
can reconstruct this signal from its sampled values. It has been shown [3] that
when ω2 is a multiple of B, we can recover the analog bandpass signal from its
samples obtained with only a sampling frequency ωs ≥ 2B. For example, when
the bandpass signal has a Fourier transform between ω1 = 4500 and ω2 = 5000,
we don’t have to choose ωs > 10,000. We can choose ωs > 1000, since ω2 =
10B in this example.

Example 3.1

Consider a continuous-time signal xa (t) = e−0.2t u(t) that has the Fourier
transform
" X(j ω) = 1/(j ω + 0.2). The magnitude |X(j ω)| = |1/(j ω + 0.2)| =
1/(ω + 0.04), and when we choose a frequency of 200π, we see that the
2

magnitude is approximately 0.4(10−3 ). Although the function xa (t) = e−0.2t u(t)

is not bandlimited, we can assume that it is almost bandlimited with bandwidth
of 200π and choose a sampling frequency of 400π rad/s or 200 Hz. So the sam-
pling period T = 2001
= 0.005 second and ωs = 2π/T = 400π rad/s. To verify
that (3.11) and (3.16) both give the same result, let us evaluate the DTFT at
ω = 0.5 rad/s. According to (3.11), the DTFT of x(nT ) is
∞
∞

e−0.2(nT ) e−j ωnT = e−0.001n e−j ωn(0.005)
n=0 n=0
1
= X(ej ωT ) = (3.20)
1 − e−0.001 e−j (0.005ω)
 THEORY OF SAMPLING 121

and its magnitude at ω = 0.5 is

 
 1 
 
 1 − e−0.001 e−j (0.0025)  = 371.5765

According to (3.16), the DTFT for this example becomes

∞

1 1
(3.21)
0.005 0.2 + j (ω + k400π)
k=−∞

and at ω = 0.5, we can neglect the duplicates at j k400π and give the magnitude
of the frequency response as
 
1  1 
 = 371.3907
0.005 0.2 + j 0.5 


The two magnitudes at ω = 0.5 are nearly equal; the small difference is
attributable to the slight aliasing in the frequency response. See Figure 3.6, which
illustrates the equivalence of the two equations. But (3.16) is not useful when
a sequence of arbitrary values (finite or infinite in length) is given because it
is difficult to guess the continuous-time signal of which they are the sampled
values; even if we do know the continuous-time signal, the choice of a sampling
frequency to avoid aliasing may not be practical, for example, when the signal is
a highpass signal. Hence we refer to (3.11) whenever we use the acronym DTFT
in our discussion.

X(jw)

0 200 p

X(e jw)

0 200 p fs = 400p w

Figure 3.6 Equivalence of the two definitions for the Fourier transform of a discrete-time
signal.
122 FREQUENCY-DOMAIN ANALYSIS

3.3 DTFT AND IDTFT

The expressions for the DTFT X(ej ω ) and the IDTFT x(n) are
∞

jω
X(e ) = x(n)e−j ωn (3.22)
n=0

π
1
x(n) = X(ej ω )ej ωn dω (3.23)
2π −π

The DTFT and its inverse (IDTFT) are extensively used for the analysis and
design of discrete-time systems and in applications of digital signal processing
such as speech processing, speech synthesis, and image processing. Remember
that the terms frequency response of a discrete-time signal and the discrete-time
Fourier transform (DTFT) are synonymous and will be used interchangeably.
This is also known as the frequency spectrum; its magnitude response and
phase response are generally known as the magnitude spectrum and phase spec-
trum, respectively. We will also use the terms discrete-time signal, discrete-time
sequence, discrete-time function, and discrete-time series synonymously.
We will represent the frequency response of the digital filter either by
H (ej ωT ) or more often by H (ej ω ) for convenience. Whenever it is expressed
as H (ej ω )—which is very common practice in the published literature—the
frequency variable ω is to be understood as the normalized frequency ωT =
ω/fs . We may also represent the normalized frequency ωT by θ (radians). In
Figure 3.7a, we have shown the magnitude response of an ideal lowpass filter,
demonstrating that it transmits all frequencies from 0 to ωc and rejects frequencies
higher than ωc . The frequency response H (ej ω ) is periodic, and its magnitude is
an even function. In Figure 3.7b suppose we have shown the magnitude response
of the lowpass filter only over the frequency range [0 π]. We draw its magni-
tude for negative values of ω since it is an even function and extend it by repeated
duplication with a period of 2π, thereby obtaining the magnitude response for all
values of ω over the range (−∞, ∞). Therefore, if the frequency specifications
are given over the range [0 π], we know the specifications for all values of the
normalized frequency ω, and the specifications for digital filters are commonly
given for only this range of frequencies. Note that we have plotted the magni-
tude response as a function of the normalized frequency ω. Therefore the range
[0 π] corresponds to the actual frequency range [0 ωs /2] and the normalized
frequency π corresponds to the Nyquist frequency (and 2π corresponds to the
sampling frequency).
Sometimes the frequency ω is even normalized by πfs so that the Nyquist
frequency has a value of 1, for example, in MATLAB functions. In Figures 3.7c,d,
we have shown the magnitude response of an ideal highpass filter. In Figure 3.8
we show the magnitude responses of an ideal bandpass and bandstop filter.
It is convenient to do the analysis and design of discrete-time systems on
the basis of the normalized frequency. When the frequency response of a fil-
ter, for example, shows a magnitude of 0.5 (i.e., −6 dB) at the normalized
 DTFT AND IDTFT 123

⎮Hlp(e jw)⎮

−2p −p −wc 0 wc p 2p 2w
(a)

⎮Hlp(e jw)⎮

−p −wc 0 wc p
(b)

⎮Hlp(e jw)⎮

−2p −p −wc 0 wc p 2p w
(c)

⎮Hlp(e jw)⎮

0 wc p w
(d)

Figure 3.7 Magnitude responses of ideal lowpass and highpass filters.

frequency 0.3π, the actual frequency can be easily computed as 30% of the
Nyquist frequency, and when the sampling period T or the sampling frequency
ωs (or fs = 1/T ) is given, we know that 0.3π represents (0.3)(ωs /2) rad/s or
(0.3)(fs /2) Hz. By looking at the plot, one should therefore be able to determine
what frequency scaling has been chosen for the plot. And when the actual sam-
pling period is known, we know how to restore the scaling and find the value
of the actual frequency in radians per second or in hertz. So we will choose the
normalized frequency in the following sections, without ambiguity.
The magnitude response of the ideal filters shown in Figures 3.7 and 3.8 cannot
be realized by any transfer function of a digital filter. The term “designing a digital
filter” has different meanings depending on the context.
 One
 meaning is to find a
transfer function H (z) such that its magnitude H (ej ω ) approximates the ideal
magnitude response as closely as possible. Different approximation
  criteria have
been proposed to define how closely the magnitude H (ej ω ) approximates the
ideal magnitude. In Figure 3.9a, we show the approximation of the ideal lowpass
filter meeting the elliptic function criteria. It shows an error in the passband as
124 FREQUENCY-DOMAIN ANALYSIS

⎮Hbp(e jw)⎮

0 wc1 wc2 π w
(a)

⎮Hbs(e jw)⎮

0 wc1 wc2 π w
(b)

Figure 3.8 Magnitude responses of ideal bandpass and bandstop filters.

Magnitude response of elliptic lowpass filter

1

0.8 Passband
Magnitude

0.6

0.4

0.2 Transition band

Stopband
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency w/pi
(a)

Magnitude response of Butterworth highpass filter

1.4
1.2
1
Magnitude

0.8
0.6 Passband
0.4
0.2
Stopband
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency w/pi
(b)

Figure 3.9 Approximation of ideal lowpass and highpass filter magnitude response.
 DTFT AND IDTFT 125

Magnitude response of Chebyshev I bandpass filter

1
Passband
0.8
Magnitude

0.6

0.4

0.2
Stopband Stopband
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency w/pi
(a)

Magnitude response of Chebyshev II bandstop filter

1.4
1.2
1
Magnitude

Passband Passband
0.8
0.6
0.4
0.2 Stopband
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency w/pi
(b)

Figure 3.10 Approximation of ideal bandpass and bandstop digital filters.

well as in the stopband, which is equiripple in nature, whereas in Figure 3.9b,

the magnitude of a highpass filter is approximated by a Butterworth type of
approximation, which shows that the magnitude in the passband is “nearly flat”
and decreases monotonically as the frequency decreases from the passband.
Figure 3.10a illustrates a Chebyshev type I approximation of an ideal band-
pass filter, which has an equiripple error in the passband and a monotonically
decreasing response in the stopband, whereas in Figure 3.10b, we have shown a
Chebyshev type II approximation of an ideal bandstop filter; thus, the error in
the stopband is equiripple in nature and is monotonic in the passband. The exact
definition of these criteria and the design of filters meeting these criteria will be
discussed in the next two chapters.

3.3.1 Time-Domain Analysis of Noncausal Inputs

Let the DTFT of the input signal x(n) and the unit impulse response h(n) of
a discrete-time system be X(ej ω ) and H (ej ω ), respectively.
 The output y(n)
is obtained by the convolution sum x(n) ∗ h(n) = y(n) = ∞ k=−∞ h(k)x(n − k),
which shows that the convolution sum is applicable even when the input signal
126 FREQUENCY-DOMAIN ANALYSIS

is defined for −∞ < n < 0 or −∞ < n < ∞. In this case, the unilateral z trans-
form of x(n) cannot be used. Therefore we cannot find the output y(n) as the
inverse z transform of X(z)H (z). However, we can find the DTFT of the input
sequence even when it is defined for −∞ < n < ∞, and then multiply it by the
DTFT of h(n) to get the DTFT of the output as Y (ej ω ) = X(ej ω )H (ej ω ). Its
IDTFT yields the output y(n). This is one advantage of using the discrete-time
Fourier transform theory. So for time-domain analysis, we see that the DTFT-
IDTFT pair offers an advantage over the z-transform method, when the input
signal is defined for −∞ < n < 0 or −∞ < n < ∞. An example is given later
to illustrate this advantage over the z-transform theory in such cases.
The relationship Y (ej ω ) = X(ej ω )H (ej ω ) offers a greater advantage as it is
the basis for the design of all digital filters. When we want to eliminate certain
frequencies or a range of frequencies in the input signal, we design a filter such
that the magnitude of H (ej ω ) is very small at these frequencies or over the
range of frequencies that would therefore form the stopband. The magnitude of
the frequency response H (ej ω ) at all other frequencies is maintained at a high
level, and these frequencies constitute the passband. The magnitude and phase
responses of the filter are chosen so that the magnitude and phase responses of
the output of the filter will have an improved quality of information. We will
discuss the design of digital filters in great detail in Chapters 4 and 5. We give
only a simple example of its application in the next section.

Example 3.2

Suppose that the input signal has a lowpass magnitude response with a bandwidth
of 0.7π as shown in Figure 3.11 and we want to filter out all frequencies outside
the range between ω1 = 0.3π and ω2 = 0.4π. Note that the sampling frequency
of both signals is set at 2π. If we pass the input signal through a bandpass
filter with a passband between ω1 = 0.3π and ω2 = 0.4π, then the frequency
response of the output is given by a bandpass response with a passband between
ω1 = 0.3π and ω2 = 0.4π, with all the other frequencies having been filtered
out. It is interesting to observe that the maximum frequency in the output is
0.4π; therefore, we can reconstruct y(t) from the samples y(n) and then sample
at a lower sampling frequency of 0.8π, instead of the original frequency of 2π.
If the sampling frequency in this example is 10,000 Hz, then the Nyquist fre-
quency is 5000 Hz, and therefore the input signal has a bandwidth of 3500 Hz,
corresponding to the normalized bandwidth of 0.7π, whereas the bandpass filter
has a passband between 1500 and 2000 Hz. The output of the bandpass filter has
a passband between 1500 and 2000 Hz. Since the maximum frequency in the
output signal is 2000 Hz, one might think of reconstructing the continuous-time
signal using a sampling frequency of 4000 Hz. But this is a bandpass signal
with a bandwidth of 500 Hz, and 2000 Hz is 8 times the bandwidth; according
to the sampling theorem for bandpass signals, we can reconstruct the output sig-
nal y(t) using a sampling frequency of twice the bandwidth, namely, 1000 Hz
instead of 4000 Hz. The theory and the procedure for reconstructing the analog
 DTFT AND IDTFT 127

|Xlp(e jw)| |Hbp(e jw)|

5

0 0.7π π 0 0.3π 0.4π w π

|X(e jw)|
15

0 0.3π 0.4π w π

Figure 3.11 A lowpass signal processed by a bandpass filter.

bandpass signal from its samples is beyond the scope of this book and will not
be treated further.

3.3.2 Time-Shifting Property

If x(n) has a DTFT X(ej ω ), then x(n − k) has a DTFT equal to e−j ωk X(ej ω ),
where k is an integer. This is known as the time-shifting
∞ property and it
is easily
∞proved as follows: DTFT of x(n − k) = n=−∞ x(n − k)e−j ωn =
e −j ωk −j ωn =e −j ωk j ω
n=−∞ x(n)e X(e ). So we denote this property by

x(n − k) ⇔ e−j ωk X(ej ω )

3.3.3 Frequency-Shifting Property

If x(n) ⇔ X(ej ω ), then

ej ω0 n x(n) ⇔ X(ej (ω−ω0 ) )

This is known as the frequency-shifting property, and it is easily proved as

follows:
∞
∞

x(n)ej ω0 n e−j ωn = x(n)e−j (ω−ω0 )n = X(ej (ω−ω0 ) )
n=−∞ n=−∞
128 FREQUENCY-DOMAIN ANALYSIS

3.3.4 Time Reversal Property


Let us consider x(n) = a n u(n). Its DTFT X(ej ω ) = ∞ n −j ωn
n=0 a e . Next, to find
theDTFT of x(−n), if we replace n by −n, we would write the DTFT of x(−n)
as −∞ −n j ωn
n=0 a e (but that is wrong), as illustrated by the following example:
∞

X(ej ω ) = a n e−j ωn = 1 + ae−j ω + a 2 e−j 2ω + a 3 e−j 3ω + · · ·
n=0

But the correct expression for the DTFT of x(−n) is of the form 1 + aej ω +
a 2 ej 2ω + a 3 ej 3ω + · · · . 
So the compact form for this series is 0n=−∞ a −n e−j ωn . With this clarifica-
tion, we now prove the property that if x(n) ⇔ X(ej ω ) then

x(−n) ⇔ X(e−j ω ) (3.24)


Proof : DTFT of x(−n) = ∞ x(−n)e−j ωn . We substitute (−n) = m, and
∞ ∞
n=−∞ 
we get n=−∞ x(−n)e
−j ωn
= m=−∞ x(m)ej ωm = ∞ m=−∞ x(m)e
−j (−ω)m
=
−j ω
X(e ).

Example 3.3

Consider x(n) = δ(n). Then, from the definition for DTFT, we see that δ(n) ⇔
X(ej ω ) = 1 for all ω.
From the time-shifting property, we get

δ(n − k) ⇔ e−j ωk (3.25)

The Fourier transform e−j ωk has a magnitude of one at all frequencies but a
linear phase as a function of ω that yields a constant group delay of k samples. If
we extend this result byconsidering an infinite sequence of unit impulses,
∞ which
can be represented by ∞ k=−∞ δ(n − k), its DTFT would yield k=−∞ e
−j ωk
.
But this does not converge to any form of expression. Hence we resort to a
different approach, as described below, and derive the result (3.28).

Example 3.4

We consider x(n) = δ(n + k) + δ(n − k). Its DTFT is given by X(ej ω ) = ej ωk +

e−j ωk = 2 cos(ωk). In this example, note that the DTFT is a function of the
continuous variable ω whereas k is a fixed number. It is a periodic function of
ω with a period of 2π, because 2 cos((ω + 2rπ)k) = 2 cos(ωk), where r is an
integer. In other words, the inverse DTFT of X(ej ω ) = 2 cos(ωk) is a pair of
impulse functions at n = k and n = −k, and this is given by

cos(ωk) ⇔ 12 [δ(n + k) + δ(n − k)] (3.26)

 DTFT AND IDTFT 129

Example 3.5

Now we consider  the infinite sequence of x(n) = 1 for all n. We represent it in

form x(n) = ∞
the  k=−∞ δ(n − k). We prove below that its DTFT is given as
2π ∞ k=−∞ δ(ω − 2πk), which is a periodic train of impulses in the frequency
domain, with a strength equal to 2π and a period equal to 2π (which is the
normalized sampling frequency).
 We prove this result given by (3.27), by showing
that the inverse DTFT of 2π ∞ k=−∞ δ(ω − 2πk) is equal to one for all n.

∞

2π δ(ω − 2πk) ⇔ 1 (for all n) (3.27)
k=−∞
∞
Proof : The inverse DTFT of 2π k=−∞ δ(ω − 2πk) is evaluated as

 ∞

1 π
2π δ(ω − 2πk) ej ωn dω
2π −π k=−∞

 ∞

π
= δ(ω − 2πk) ej ωn dω
−π k=−∞

From the sifting property we get

 ∞   ∞ 

δ(ω − 2πk) e j ωn
= δ(ω − 2πk) ej 2πkn
k=−∞ k=−∞
 ∞


= δ(ω − 2πk)
k=−∞

where we have used ej 2πkn = 1 for all n. When we integrate the sequence of
impulses from −π to π, we have only the impulse at ω = 0.
Therefore

π ∞

δ(ω − 2πk) ej ωn dω
−π k=−∞

π ∞

= δ(ω)ej ωn dω = 1 (for all n)
−π k=−∞

Thus we have derived the important result

∞
∞

δ(n − k) ⇔ 2π δ(ω − 2πk) (3.28)
k=−∞ k=−∞
130 FREQUENCY-DOMAIN ANALYSIS

To point out some duality in the results we have obtained above, let us repeat
them:

When x(n) = 1 at n = 0 and 0 at n = 0, that is, when we have δ(n), its DTFT
X(ej ω ) = 1 for all ω.

When x(n) = 1 for all  n, specifically, when we have ∞ k=−∞ δ(n − k), its
DTFT X(ej ω ) = 2π ∞ k=−∞ δ(ω − 2πk).

Using the frequency-shifting property, we get the following results:

∞

ej ω0 n ⇔ 2π δ(ω − ω0 − 2πk) (3.29)
k=−∞

From these results, we can obtain the DTFT for the following sinusoidal
sequences:

∞
1 j ω0 n
cos(ω0 n) = [e + e−j ω0 n ] ⇔ π δ(ω − ω0 − 2πk) + δ(ω + ω0 − 2πk)
2 k=−∞
∞
1 j ω0 n −j ω0 n π
sin(ω0 n) = [e −e ]⇔ δ(ω − ω0 − 2πk) − δ(ω + ω0 − 2πk)
2j j k=−∞
(3.30)
Now compare the results in (3.28) and (3.30), which are put together in (3.31)
and (3.32) in order to show the dualities in the properties of the two transform
pairs. Note in particular that cos(ωk) is a discrete-time Fourier transform and a
function of ω, where k is a fixed integer, whereas cos(ω0 n) is a discrete-time
sequence where ω0 is fixed and is a function of n:
1
2 [δ(n + k) + δ(n − k)] ⇐⇒ cos(ωk) (3.31)
∞

cos(ω0 n) ⇐⇒ π δ(ω − ω0 − 2πk) + δ(ω + ω0 − 2πk)
k=−∞
(3.32)
Let us show the duality of the other functions derived in (3.26) and (3.29)

δ(n) ⇐⇒ 1 for all ω

whereas
∞

x(n) = 1 for all n ⇐⇒ 2π δ(ω − 2πk)
k=−∞
 DTFT AND IDTFT 131

Using the time- and frequency-shifting properties on these functions, we

derived the following Fourier transform pairs as well:

δ(n − k) ⇐⇒ e−j ωk
∞

e j ω0 n
⇐⇒ 2π δ(ω − ω0 − 2πk)
k=−∞

Example 3.6

This is another example chosen to highlight the difference between the

expressions for the discrete-time sequence x(n) and its DTFT X(ej ω ). So if
we are given F (ej ω ) = 10 cos(5ω) + 5 cos(2ω) = 5ej 5ω + 5e−j 5ω + 2.5ej 2ω +
2.5e−j 2ω , its IDTFT is obtained by using the result δ(n − k) ⇔ e−j ωk , and we
get f (n) = 5δ(n + 5) + 2.5δ(n + 2) + 2.5δ(n − 2) + 5δ(n − 5), which is plot-
ted in Figure 3.12a. Obviously it is a finite sequence with four impulse functions
and therefore is not periodic.

f(n)
5

2.5

−5 −4 −3 −2 −1 0 1 2 3 4 5 n
(a)

|G(e jw)|

10

−0.5π −0.2π 0 0.2π 0.5π w

(b)

Figure 3.12 A sequence of impulses in the discrete-time domain and a sequence of

impulses in the discrete-frequency domain.
132 FREQUENCY-DOMAIN ANALYSIS

If we are given a function g(n) = 10 cos(0.5πn) + 5 cos(0.2πn), the first thing

we have to recognize is that it is a discrete-time function and is a periodic function
in its variable n. So we find its DTFT, using (3.30), as
∞
∞

G(ej ω ) = 10π δ(ω − 0.5π − 2πk) + 10π δ(ω + 0.5π − 2πk)
k=−∞ k=−∞
∞
∞

+ 5π δ(ω − 0.2π − 2πk) + 5π δ(ω + 0.2π − 2πk)
k=−∞ k=−∞

This DTFT is shown in Figure 3.12b. We notice that it represents an infinite

number of impulses in the frequency domain that form a periodic function in the
frequency variable ω. Because it has discrete components, the impulse functions
are also called the spectral components of g(n).
We have chosen a DTFT F (ej ω ) and derived its IDTFT f (n), which is a
sequence of impulse functions in the time domain as shown in Figure 3.12a;
then we chose a discrete-time function g(n) and derived its DTFT G(ej ω ),
which is a sequence of impulse functions in the frequency domain as shown
in Figure 3.12b.

Example 3.7

Consider the simple example of a discrete-time sinsusoidal signal x(n) =

4 cos(0.4πn). It is periodic when the frequency (0.4N ) is an integer or a ratio of
integers. We choose N = 5 as the period of this function, so x(n) = x(n + 5K) =
4 cos[0.4π(n + 5K)], where K is any integer.
j 0.4πn + e−j 0.4πn ], and therefore its DTFT X(ej ω ) =
∞rewrite x(n) = 2[e
We 
2π k=−∞ δ(ω − 0.4π − 2πk) + 2π ∞ k=−∞ δ(ω + 0.4π − 2πk). It consists of
impulse functions of magnitude equal to 2π, at ω = ±(0.4π + 2πK) in the
frequency domain and with a period of 2π.
Given f (n) = 2δ(n + 4) + 2δ(n − 4), its DTFT ∞ is F (ej ω ) = 4 cos(4ω), and
 = 4 cos(0.4πn), its DTFT is X(e ) = 2π k=−∞ δ(ω − 0.4π − 2πk) +
jω
if x(n)
2π ∞ k=−∞ δ(ω + 0.4π − 2πk).
Examples 3.6 and 3.7 have been chosen in particular to distinguish the differ-
ences between the two Fourier transform pairs.

Example 3.8

Let us consider the DTFT of some more sequences. For example, the DTFT of
x1 (n) = a n u(n) is derived below:
∞
∞
 n
X1 (ej ω ) = a n e−j ωn = ae−j ω
n=0 n=−∞
 DTFT AND IDTFT 133

This
 −j ωinfinite
 series converges to 1/(1 − ae−j ω ) = ej ω /(ej ω − a) when
ae  < 1, that is, when |a| < 1. So, the DTFT of (0.4)n u(n) is 1/(1 − 0.4e−j ω )
and the DTFT of (−0.4)n u(n) is 1/(1 + 0.4e−j ω ). Note that both of them are causal
sequences.
If we are given a sequence x13 (n) = a |n| , where |a| < 1, we split the sequence
as a causal sequence x1 (n) from 0 to ∞, and a noncausal sequence x3 (n)
from −∞ to −1. In other words, we can express x1 (n) = a n u(n) and x3 (n) =
a −n u(−n − 1). We derive the DTFT of x13 (n) as
∞
−1

X13 (ej ω ) = a n e−j ωn + a −n e−j ωn X1 (ej ω ) + X3 (ej ω )
n=0 n=−∞

Substituting m = −n in the second summation for X3 (ej ω ), we get

∞
∞

 n  m
X13 (ej ω ) = ae−j ω + aej ω
n=0 m=1

∞
∞

 n  m
= ae−j ω −1+ aej ω
n=0 m=0

1 1
= −j
−1+ for |a| < 1
1 − ae ω 1 − aej ω
1 aej ω
= +
1 − ae−j ω 1 − aej ω
1 − a2
= for |a| < 1
1 − 2a cos ω + a 2
Hence we have shown that

1 − a2
a |n| ⇔ for |a| < 1
1 − 2a cos ω + a 2

These results are valid when |a| < 1. From the result a n u(n) ⇔ 1/(1 − ae−j ω ),
by application of the time-reversal property, we also find that x4 (n) = x1 (−n) =
a −n u(−n) ⇔ 1/(1 − aej ω ) for |a| < 1 whereas we have already determined that
x3 (n) = a −n u(−n − 1) ⇔ aej ω /(1 − aej ω ). Note that x3 (n) is obtained from
x4 (n) by deleting the sample of x4 (n) at n = 0, specifically, x4 (n) − 1 = x3 (n).
We used this result in deriving X3 (ej ω ) above. The sequence x13 (n) is plotted in
Figure 3.13, while the plots of x1 (n), x3 (n) are shown in Figures 3.14 and 3.15,
respectively.
134 FREQUENCY-DOMAIN ANALYSIS

x13(n)

−8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8
n

Figure 3.13 The discrete-time sequence x13 (n).

x1(n)

1.0

0.8

0.64
0.512
0.41
0.328
0.262
0.21
0.168
0.134

0 1 2 3 4 5 6 7 8 n

Figure 3.14 The discrete-time sequence x1 (n).

x3(n)

0.8

0.64
0.512
0.41
0.328
0.262
0.168

−7 −6 −5 −4 −3 −2 −1 0 n

Figure 3.15 The discrete-time sequence x3 (n).

 DTFT AND IDTFT 135

x5(n)

0.5

0.25
0.125
0.0315 0.0625

−7 −6 −5 −4 −3 −2 −1 0 1 2 n

Figure 3.16 The discrete-time sequence x5 (n).

Now let us consider the case of x5 (n) = α n u[−(n + 1)], where |α| > 1. A plot
of this sequence is shown in Figure 3.16 for α = 2. Its DTFT is derived below:
∞

X5 (ej ω ) = α n u[−(n + 1)]e−j ωn
n=−∞
−∞
−∞ 


−j ω n 1 j ω −n
= αe = e
α
n=−1 n=−1

By a change of variable n = −m, we get

∞ 
 ∞ 

1 jω m 1 jω m
X5 (e ) =
jω
e = −1 + e
α α
m=1 m=0
1
= −1 + 1
1− αe
jω

1
=
αe−j ω −1
So we have the transform pair
1 ej ω
x5 (n) = α n u[−(n + 1)] ⇔ = when |α| > 1 (3.33)
αe−j ω − 1 α − ej ω
It is important to exercise caution in determining the differences in this pair
(3.33), which is valid for |α| > 1, and the earlier pairs, which are valid for |a| < 1.
All of them are given below (again, the differences between the different DTFT-
IDTFT pairs and the corresponding plots should be studied carefully and clearly
understood):

1 ej ω
x1 (n) = a n u(n) ⇔ = when |a| < 1 (3.34)
1 − ae−j ω ej ω − a
136 FREQUENCY-DOMAIN ANALYSIS

1 e−j ω
x4 (n) = a −n u(−n) ⇔ = when |a| < 1 (3.35)
1 − aej ω e−j ω − a
aej ω
x3 (n) = a −n u(−n − 1) ⇔ when |a| < 1 (3.36)
1 − aej ω
1 − a2
x13 (n) = x1 (n) + x3 (n) ⇔ when |a| < 1 (3.37)
1 − 2a cos ω + a 2
For the sequence x5 (n) = α n u[−(n + 1)], note that the transform pair is given
by (3.38), which is valid when |α| > 1:

1 ej ω
x5 (n) = α n u[−(n + 1)] ⇔ = when |α| > 1 (3.38)
αe−j ω −1 α − ej ω

Example 3.9

A few examples are given below to help explain these differences. From the
results given above, we see that

1. If the DTFT X1 (ej ω ) = 1/(1 − 0.8e−j ω ), its IDTFT is x1 (n) = (0.8)n u(n).
2. The IDTFT of X3 (ej ω ) = 0.8ej ω /(1 − 0.8ej ω ) is given by x3 (n) =
(0.8)−n [u(−n − 1)].
3. The IDTFT of X4 (ej ω ) = 1/(1 − 0.8ej ω ) is x4 (n) = (0.8)−n u(−n). But
4. The IDTFT of X5 (ej ω ) = ej ω /(2 − ej ω ) is x5 (n) = (2)n u(−n − 1).

Note the differences in the examples above, particularly the DTFT-IDTFT pair
for x5 (n).
The magnitude and phase responses of X1 (ej ω ), X3 (ej ω ), and X13 (ej ω ) are
shown in Figures 3.17, 3.18, and 3.19, respectively. The magnitude responses of
X1 (ej ω ), X4 (ej ω ), and X3 (ej ω ) given below appear the same except for a scale

5 4.5

4 4

3 3.5

2 3

1 2.5

0 2
−2 −1 0 1 2 −2 −1 0 1 2

Figure 3.17 The magnitude and phase responses of x1 (n).

 DTFT AND IDTFT 137

5 18
16
4
14
12
3
10
2 8
6
1
4
0 2
−2 −1 0 1 2 −2 −1 0 1 2

Figure 3.18 The magnitude and phase responses of x3 (n).

10

0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2

Figure 3.19 The magnitude response of x13 (n).

factor in X3 (ej ω ). The phase response of X3 (ej ω ) exceeds that of X1 (ej ω ) by

ω radians as seen in Figures 3.17 and 3.18. The frequency response (DTFT)
X13 (ej ω ) shown in Figure 3.19 is a real function and therefore has zero phase.
We obtain
  1.0
X1 (ej ω ) = " (3.39)
[1 − 0.8 cos(ω)]2 + [0.8 sin(ω)]2
1.0
=√ (3.40)
[1 + 0.64 − 1.6 cos(ω)
 
a sin (ω)
Ang[X1 (ej ω )] = Ang[X3 (ej ω )] = − tan−1 (3.41)
1 − a cos(ω)
138 FREQUENCY-DOMAIN ANALYSIS

  0.8
X3 (ej ω ) = " (3.42)
[1 − 0.8 cos(ω)]2 + [0.8 sin(ω)]2
0.8
=√ (3.43)
[1 + 0.64 − 1.6 cos(ω)
0.36
X13 (ej ω ) = (3.44)
1 − 1.6 cos ω + 0.64

3.4 DTFT OF UNIT STEP SEQUENCE

Note that a n u(n) ⇔ 1/(1 − ae−j ω ) = ej ω /(ej ω − a) is valid only when |a| < 1.
When a = 1, we get the unit step sequence u(n), but the DTFT 1/(1 − e−j ω )
has an infinite number of poles at ω = 0, ±k2π, where k is an integer. In order
to avoid these singularities in 1/(1 − e−j ω ) = ej ω /(ej ω − 1), the DTFT of the
unit step sequence u(n) is derived in a different way as described below.
We express the unit step function as the sum of two functions
u(n) = u1 (n) + u2 (n)

where

u1 (n) = 1
2 for − ∞ < n < ∞

and
#
1
for n ≥ 0
u2 (n) = 2
− 12 for n < 0

Therefore we express δ(n) = u2 (n) − u2 (n − 1). Using δ(n) ⇔ 1 and u2 (n) −

u2 (n − 1) ⇔ U2 (ej ω ) − e−j ω U2 (ej ω ) = U2 (ej ω )[1 − e−j ω ], and equating the
two results, we get

1 = U2 (ej ω )[1 − e−j ω ]

Therefore
1
U2 (ej ω ) =
[1 − e−j ω ]

We know that the DTFT of u1 (n) = π ∞ k=−∞ δ(ω − 2πk) = U1 (e ). Adding
jω

these two results, we have the final result

∞
1
u(n) ⇔ π δ(ω − 2πk) + (3.45)
(1 − e−j ω )
k=−∞

This gives us the DTFT of the unit step function u(n), which is unique.
 DTFT OF UNIT STEP SEQUENCE 139

Applying the time-shifting property, frequency-shifting property, and time

reversal property on u(n), we can derive the DTFT of a few more discrete-time
functions. For example
 ∞

1
−j ωk
u(n − k) ⇔ e π δ(ω − 2πk) + (3.46)
(1 − e−j ω )
k=−∞
∞
1
ej ω0 n u(n) ⇔ π δ(ω − ω0 − 2πk) + (3.47)
(1 − e−j (ω−ω0 ) )
k=−∞
 ∞
1 1
cos(ω0 n)u(n) ⇔ π δ(ω − ω0 − 2πk) + −j
2 (1 − e (ω−ω0 ) )
k=−∞
∞

1
+π δ(ω + ω0 − 2πk) + (3.48)
(1 − e−j (ω+ω0 ) )
k=−∞

It is worth comparing the DTFT of ej ω0 n u(n) given above with the DTFT of
e−an u(n), where |a| < 1:

1
e−an u(n) ⇔ (3.49)
1 − e−a e−j ω

3.4.1 Differentiation Property


To prove that nx(n) ⇔ j [dX(ej ω )]/dω, we start with X(ej ω ) = ∞ n=−∞
−j ωn
x(n)e
∞ and differentiate both sides to get [dX(e jω
)]/dω =
−j ωn
n=−∞ x(n)(−j n)e  and multiplying both sides by j , we get
j [dX(ej ω )]/dω = ∞ n=−∞ nx(n)e
−j ωn
. The proof is similar to that used
in Chapter 2 to prove that the z transform of nx(n)u(n) is −z[dX(z)]/dz.
Given x(n) = a n u(n) ⇔ 1/(1 − ae−j ω ) = X(ej ω ), we can derive the follow-
ing, using the differentiation property:
 
dX(ej ω ) −j ae−j ω ae−j ω
j =j =
dω (1 − ae−j ω )2 (1 − ae−j ω )2
ae−j ω
na n u(n) ⇔ (3.50)
(1 − ae−j ω )2

Since the DTFT of a n u(n) is 1/(1 − ae−j ω ), we add this DTFT to that of na n u(n)
and get

1
(n + 1)a n u(n) ⇔ (3.51)
(1 − ae−j ω )2
140 FREQUENCY-DOMAIN ANALYSIS

Example 3.10
Consider a rectangular pulse

1 |n| ≤ N
xr (n) =
0 |n| > N
which is plotted in Figure 3.20. It is also known as a rectangular window (of
length 2N + 1) and will be used in Chapter 5 when we discuss the design of
FIR filters. Its DTFT is derived as follows:

N
Xr (ej ω ) = e−j ωn
n=−N

To simplify this summation, we use the identity5

N
r N +1 − r −N
rn = ; r = 1 (3.52)
r −1
n=−N

= 2N + 1; r=1 (3.53)
and get
jNω
e−j (N +1)ω − e
Xr (ej ω ) =
e−j ω − 1

e−j 0.5ω e−j (N +0.5)ω − ej (N +0.5)ω
=
e−j 0.5ω (e−j 0.5ω − ej 0.5ω )
⎧
⎨ sin[(N + 0.5)ω] ω = 0
= sin[0.5ω]
⎩
2N + 1 ω=0
which is shown in Figure 3.21.

Xr(n)

−5 0 1 2 3 4 5 n

Figure 3.20 A rectangular pulse function.

5
Proof:
N
n=−N (r
n+1
− r n) = (r −N +1 + r −N +2 + · · · + r −1 + 1 + r + r 2 + · · · + r N + r N +1 ) − (r −N +
−N +1 −N +2
N
r +r + · · · + 1 + r + r 2 + · · · + r N ) = r N +1 − r −N . Therefore n=−N (r
n+1
− rn) =
  N +1 −N
(r − 1) Nn=−N
n N
r and n=−N r = (r n
− r )/r − 1; r = 1 = 2N + 1; r = 1.
 DTFT OF UNIT STEP SEQUENCE 141

Magnitude of the DTFT of a rectangular pulse

12

10

6
Magnitude

−2

−4
−8 −6 −4 −2 0 2 4 6 8
Normalized frequency

Figure 3.21 The DTFT of a rectangular pulse function.

Using the time-shifting property, we can find the DTFT of the sequence
xr2 (n) = x1 (n − N ) as
sin[(N + 0.5)ω]
Xr2 (ej ω ) = e−j N ω (3.54)
sin[0.5ω]

1 0 ≤ n ≤ 2N
where xr2 (n) =
0 otherwise

Example 3.11

Let us find the IDTFT of a rectangular spectrum H (ej ω ) which is shown as the
magnitude of an ideal lowpass filter in Figure 3.7a with a cutoff frequency of ωc .

π
1
h(n) = H (ej ω )ej ωn dω
2π −π

ωc
1
= ej ωn dω
2π −ωc
 j ωn ωc
1 e 
= 
2π j n −ωc
1 ωc
= sin(ωc n) = sinc(ωc n) (3.55)
πn π
142 FREQUENCY-DOMAIN ANALYSIS

Inverse DTFT of the ideal lowpass filter

0.25

0.2

0.15
Values of the samples

0.1

0.05

−0.05

−0.1
−21 −16 −11 −6 0 4 9 14 19
Value of the index n

Figure 3.22 The inverse DTFT of an ideal lowpass filter.

This is a line spectrum which is shown in Figure 3.22. It is interesting to com-

pare the general shape of the rectangular pulse function xr (n) and its frequency
response Xr (ej ω ) with the frequency response H (ej ω ) of the lowpass filter and
its inverse DTFT h(n) which are derived above. However, it should be noted that
they are not duals of each other, because Xr (ej ω ) is not exactly a sinc function
of ω.

3.4.2 Multiplication Property

When two discrete -time sequences are multiplied, for example, x(n)h(n) = y(n),
the DTFT of y(n) is the convolution of X(ej ω ) and H (ej ω ) that is carried
out in the frequency domain as an integral over one full period. Choosing the
period [−π π] in the convolution integral, symbolically denoted by X(ej ω )*
H (ej ω ) = Y (ej ω ), we have the property

π
1
x(n)h(n) ⇔ X(ej ς )H (ej (ω−ς ) ) dς (3.56)
2π −π

Remember that we will use (3.55) and (3.56) in the design of FIR filters discussed
in Chapter 5.

Example 3.12

The properties and the DTFT-IDTFT pairs discussed here are often used in
frequency-domain analysis of discrete-time systems, including the design of
 DTFT OF UNIT STEP SEQUENCE 143

filters. However, as mentioned earlier, they can also be used in time-domain

analysis, particularly when the z-transform technique cannot be used. To illustrate
this point, we provide the following examples.
Let the unit impulse response of a discrete-time system be given as h(n) =
(0.2)n u(n) and the input sequence be given as x(n) = (0.5)−n u(−n). Therefore

1 ej ω
H (ej ω ) = =
1 − 0.2e−j ω ej ω − 0.2
1 e−j ω
X(ej ω ) = =
1 − 0.5ej ω e−j ω − 0.5
  
ej ω e−j ω
Y (ej ω ) = H (ej ω )X(ej ω ) =
ej ω − 0.2 e−j ω − 0.5

Now let

k1 e j ω k2 e−j ω
Y (ej ω ) = +
ej ω − 0.2 e−j ω − 0.5

so that we can easily obtain the inverse DTFT of each term. Note the difference
in the two terms.
Then we compute k1 from the following method, which is slightly different
from the partial fraction method we have used earlier:
 
e−j ω k2 (ej ω − 0.2) −j 2ω
Y (ej ω )e−j ω (ej ω − 0.2) = = k1 + e
e−j ω − 0.5 (e−j ω − 0.5)

Evaluating both terms at ej ω = 0.2 yields

 
e−j ω  5
k1 =  = = 1.111
e −j ω 
− 0.5 ej ω =0.2 4.5

Similarly, the constant k2 is evaluated as follows (again, this is slightly different

from the partial fraction expansion method we have used earlier):
 
k1 ej 2ω (e−j ω − 0.5) ej ω
Y (ej ω )ej ω (e−j ω − 0.5) = + k2 =
ej ω − 0.2 ej ω − 0.2

Evaluating the two terms at ej ω = 2.0 yields k2 = 1.8

2
= 1.111. Therefore we
have Y (e ) = 1.111e /(e − 0.2) + 1.111e
j ω j ω j ω −j ω −j
/(e ω − 0.5), and the output
is y(n) = 1.111(0.2)n u(n) + 1.111(0.5)−n u(−n).
Now we describe a preferred method of finding the inverse DTFT of
  
ej ω e−j ω
Y (e ) = H (e )X(e ) =
jω jω jω
ej ω − 0.2 e−j ω − 0.5
144 FREQUENCY-DOMAIN ANALYSIS

which we express in the form of

1
(1 − 0.2e−j ω )(1 − 0.5ej ω )

where each term in the denominator is of the general form (1 − ae−j ω ) or

(1 − aej ω ). The partial fraction expansion is now chosen to be in the form of
K1 K2
Y (ej ω ) = +
(1 − 0.2e−j ω ) (1 − 0.5ej ω )
The residue K1 is determined by evaluating

 1 
Y (ej ω )(1 − 0.2e−j ω )ej ω =0.2 =  = 1.1111
(1 − 0.5e ) ej ω =0.2
j ω


 1 
Similarly K2 = Y (ej ω )(1 − 0.5ej ω )ej ω =2 =  = 1.111
(1 − 0.2e −j ω ) ej ω=2

Example 3.13

Let x(n) = ej (0.3πn) and h(n) = (0.2)n u(n).As in Example 3.12, we can find the
DTFT of x(n) = ej (0.3πn) as X(ej ω ) = 2π δ(ω − 0.3π − 2πk) and the DTFT
of h(n) as
   
1 ej ω
H (ej ω ) = =
1 − 0.2e−j ω ej ω − 0.2

Thus

Y (ej ω ) = X(ej ω )H (ej ω ) = 2π δ(ω − 0.3π − 2πk)H (ej ω )
∞

= 2π δ(ω − 0.3π − 2πk)H (ej 0.3π )
k=−∞
 
ej 0.3π
= 2π δ(ω − 0.3π − 2πk)
(ej 0.3π − 0.2)
 
= 1.1146e−j (0.1813) 2π δ(ω − 0.3π − 2πk)
 
Therefore y(n) = 1.1146e−j (0.1813) ej (0.3πn) = 1.1146ej (0.3πn−0.1813) .
As an alternative method, we recollect that from the convolution of ej ωn and
h(n), we obtained y(n) = ej ωn H (ej ω ). In this example H (ej ω ) = [ej ω /(ej ω −
0.2)] and ω = 0.3π. Therefore y(n) = ej 0.3πn H (ej 0.3π ):
 
ej 0.3π
y(n) = e j 0.3πn
= 1.1146ej (0.3πn−0.1813)
(ej 0.3π − 0.2)
 DTFT OF UNIT STEP SEQUENCE 145

By this method, we can also find that when the input is Re{x(n)} = cos(0.3πn),
the output is given by y(n) = Re{1.1146ej (0.3πn−0.1813) } = 1.1146 cos(0.3πn −
0.1813).

3.4.3 Conjugation Property

Here we assume the samples x(n) to be complex-valued and in the general form
x(n) = (aej θ )n and the complex conjugate to be x ∗ (n) = (ae−j θ )n .
Let us find their DTFT:
∞

x(n) ⇔ X1 (ej ω ) = (aej θ )n e−j ωn
n=−∞
∞

= (a)n ej θ n e−j ωn
n=−∞
∞

x ∗ (n) ⇔ X2 (ej ω ) = (ae−j θ )n e−j ωn
n=−∞
∞

= (a)n e−j θ n e−j ωn
n=−∞
∞
Now we replace ω by −ω in X2 (ej ω ) to get X2 (e−j ω ) = n −j θ n
n=−∞ (a) e ej ωn ,
and its conjugate is obtained as X2∗ (e−j ω ):
∞

X2∗ (e−j ω ) = (a)n ej θ n e−j ωn = X1 (ej ω )
n=−∞

In words, this result X2∗ (e−j ω ) = X1 (ej ω ) means that we find the DTFT of the
complex conjugate sequence (ae−j θ )n and replace ω by −ω and then find the
complex conjugate of the result, which is the same as the DTFT of the sequence
(aej θ )n .
However, when x(n) is real, we know that x(n) = x ∗ (n), in which case
X1 (ej ω ) = X2 (ej ω ). So we have the following result:

X1 (ej ω ) = X1∗ (e−j ω ) when x(n) is real (3.57)

3.4.4 Symmetry Property

From the result in (3.57), the following symmetry properties can be derived when
sequence x(n) is real:

Re{X(ej ω )} = Re{X(e−j ω )} (3.58)

−j ω
Im{X(e )} = −Im{X(e
jω
)} (3.59)
146 FREQUENCY-DOMAIN ANALYSIS

   
X(ej ω ) = X(e−j ω ) (3.60)
Ang X(ej ω ) = −Ang X(e−j ω ) (3.61)

The even and odd parts of the sequence x(n) are defined by xe (n) = [x(n) +
x(−n)]/2 and x0 (n) = [x(n) − x(−n)]/2 respectively. When x(n) is real, and
we use the time reversal property, we get xe (n) ⇔ [{X(ej ω )} + {X(e−j ω )}]/2] =
Re{X(ej ω )}:

xe (n) ⇔ Re{X(ej ω )} (3.62)

Similarly, x0 (n) ⇔ [{X(ej ω )} − {X(e−j ω )}]/2 = j Im{X(ej ω )}:

x0 (n) ⇔ j Im{X(ej ω )} (3.63)

 
From (3.60) and (3.61), we see that the magnitude response X(ej ω ) is an even
function of ω and the phase response is an odd function of ω.
The basic properties of the discrete-time Fourier transform and examples of
some common DTFT-IDTFT pairs that have been described above are listed in
Tables 3.1 and 3.2.

TABLE 3.1 Properties of Discrete-Time Fourier Transform

Time Domain Frequency Domain

Property x(n), x1 (n), x2 (n) X(ej ω ), X1 (ej ω ), X2 (ej ω )

Linearity ax1 (n) + bx2 (n) aX1 (ej ω ) + bX2 (ej ω )

Convolution x1 (n) ∗ x2 (n) X1 (ej ω )X2 (ej ω )
Time shifting x(n − k) e−j ωk X(ej ω )
j ω0 n
Frequency shifting e x(n) X(ej (ω−ω0 ) )
Time reversal x(−n) X(e−j ω )
π
Multiplication x1 (n)x2 (n) (1/2π) −π X1 (ej ζ )X2 (ej (ω−ζ ) )dζ
Differentiation nx(n) j [dX(ej ω )]/dω
∗
Conjugation x (n) X ∗ (e−j ω )
Even part of x(n) xe (n) = 12 [x(n) + x(−n)] Re{X(ej ω )}

Odd part of x(n) x0 (n) = 12 [x(n) − x(−n)] j Im{X(ej ω )}

Symmetry Re{X(ej ω )} = Re{X(e−j ω )}
Im{X(ej ω )} = −Im{X(e−j ω )}
   
X(ej ω ) = X(e−j ω )
Ang X(ej ω ) = −Ang X(e−j ω )
 USE OF MATLAB TO COMPUTE DTFT 147

TABLE 3.2 Common IDTFT-DTFT Pairs

Signal (IDTFT):x(n) Discrete-Time Fourier Transform (DTFT):X(ej ω )

δ(n) 1
δ(n − k) e−j
ωk

1 (for all n) 2π ∞ δ(ω − 2πk)

k=−∞
e j ω0 n 2π ∞ δ(ω − ω0 − 2πk)
∞k=−∞
cos(ω0 n) π k=−∞ δ(ω − ω0 − 2πk) + δ(ω + ω0 − 2πk)
π ∞
sin(ω0 n) j k=−∞ δ(ω − ω0 − 2πk) − δ(ω + ω0 − 2πk)
e−an u(n); |a| < 1 1/(1 − e−a e−j ω )
n
a u(n); |a| < 1 1/(1 − ae−j ω )
na n u(n); |a| < 1 ae−j ω /(1 − ae−j ω )2
(n + 1)a n u(n); |a| < 1 1/(1 − ae−j ω )2

1 |n| ≤ N
x(n) = sin(N + 0.5)ω/ sin(0.5ω)
0 |n| > N

1 |ω| ≤ ωc
(1/πn) sin(ωc n)
0 |ω| > ωc

u(n) 1/(1 − e−j ω ) + π ∞k=−∞ δ(ω − 2πk)

3.5 USE OF MATLAB TO COMPUTE DTFT

If a function
is a finite−ksequence, such as the unit impulse response of an FIR filter
H (z−1 ) = N k=0 bk z , then the difference equation for that filter is given by

M
y(n) = bk x(n − k) (3.64)
k=0

We find its frequency response or its Fourier transform (DTFT) to be

M
Y (e ) = H (e )X(e ) =
jω jω jω
bk X(ej ω )e−j ωk
k=0

where the frequency response for this FIR filter is given by

N
H (ej ω ) = bk e−j ωk (3.65)
k=0

It is also the DTFT of the signal h(k) = [b0 , b1 , b2 , . . . , bN ].

148 FREQUENCY-DOMAIN ANALYSIS

Suppose that we are given the transfer function of an IIR filter:

M −k
−1 k=0 bk z
H (z ) = N
1+ −k
k=1 ak z

There are two equivalent approaches to find the frequency response of this filter
as described below.
We find the inverse z transform h(n) of H (z−1 ), which gives an infinite number
of samples of its unit impulse response,
 and now−j we can evaluate its frequency
response or its DTFT as H (e−j ω ) = ∞ h(n)e ωr
. The other approach uses
N r=0 M
the difference equation y(n) + k=1 ak y(n − k) = k=0 bk x(n − k) and finds
the DTFT of both sides as given by
 

N
M
−j ωk
Y (e ) 1 +
jω
ak e = X(ej ω ) bk e−j ωk
k=1 k=0

so that
M −j ωk
−j ω k=0 bk e
H (e )= N (3.66)
1+ −j ωk
k=1 ak e

In short, we can state that H (e−j ω ) = H (z−1 )z=ej ω , provided both exist.
To compute and plot the magnitude, phase, and/or the group delay of the FIR
or IIR filter transfer functions H (z−1 ), we use the MATLAB functions freqz,
abs, angle, unwrap, grpdelay very extensively in signal processing and
filter design. These functions are found in the Signal Processing Toolbox of
MATLAB.
When the sequence of coefficients bk and ak are known, they are entered as
the values in the vectors for the numerator and denominator. The function freqz
is used with several variations for the input variables as described below:

[h,w] = freqz(num,den,w)
[h,w] = freqz(num, den, f, Fs)
[h,w] = freqz(num,den,K,Fs)
[h,w] = freqz(num,den,K,’whole’)
[h,f] = freqz(num,den,K,’whole’,Fs)}

The vectors num and den are the row vectors of the numerator and denominator
coefficients ak and bk , respectively. The function freqz computes the values
of the frequency response as a column vector h at the discrete values of the
frequency w. The set of default frequencies w lie between 0 and π, and the set f
is the vector of values for the frequencies we can arbitrarily distinguish between
0 and Fs/2, where Fs is the sampling frequency in hertz. We can choose a value
for K as the number of frequency points within the default range; preferably K
 USE OF MATLAB TO COMPUTE DTFT 149

is chosen as a power of 2 such as 256, 512, or 1024 to get faster computation.

When K is not specified, the default value of 512 is chosen by the program. When
we include the argument ’whole’ in the function freqz, the frequency range of
w and f changes to [0 2π] and [0 Fs ], respectively.
After we have computed the values of the frequency response at the dis-
crete frequencies, we can compute their magnitude using H = abs(h) or the
magnitude in decibels using HdB=20*log10(H). The phase response (in radi-
ans) of the filter is obtained from the function ph=angle(h), and then we can
use Ph=unwrap(ph) to unwrap the phase response so that the phase angle lies
between 0 and 2π. The group delay of the filter is computed by the MATLAB
function grpdelay (and not from the function freqz) as follows:

[gd,w]=grpdelay(num,den,K,)
[gd,w]=grpdelay(num,den,K,’whole’)

Note that we can change the name for the variables num,den,h,H,
HdB,f,FT,K,ph,Ph,gd in the statements above to other variables as we like.
After we have computed H,HdB, ph,Ph,gd, we plot them using the plotting
function with different choices of variables, as illustrated in the examples given
below. When we plot H, Hdb, ph, Ph, or grpdelay, we normally plot them
as a function of the normalized frequency on a linear scale, between 0 and π.
But the function semilog(......) plots them as a function of log10(w) and
therefore a plot of semilog(HdB) becomes the familiar Bode plot of the digital
filter; alternatively, we define ww = log10(w) as the new frequency variable
and plot the magnitudes using plot(ww, H) or plot(ww, HdB).
The MATLAB function freqz(num,den) without any other arguments com-
putes and plots the magnitude in decibels as well as the phase response as a
function of frequency in the current figure window.

Example 3.14

%Program to compute and plot the magnitude and phase responses

% of a filter
b=[0.0532 0.3725 1.1176 1.8626 1.8626 1.1176 0.3725 0.0532];
a=[1.0000 1.5473 2.1992 1.2240 0.8269 0.0347 0.0587 -0.0790];
[h,w]=freqz(b,a,256);
H=abs(h);
HdB=20*log10(H);
ph=angle(h);
Ph=unwrap(ph);
[gd,w]=grpdelay(b,a,256);
subplot(1,2,1)
plot(w,H);grid
title(’Magnitude of the frequency response’)
ylabel(’Magnitude’)
150 FREQUENCY-DOMAIN ANALYSIS

xlabel(’Normalized frequency’)
subplot(1,2,2)
plot(w,HdB);grid
title(’Magnitude in dB of the frequency response’)
ylabel(’Magnitude in dB’)
xlabel(’Normalized frequency’)
figure(2)
subplot(1,2,1)
plot(w,ph);grid
title(’Phase response of the filter’)
ylabel(’Phase angle in radians’)
xlabel(’Normalized frequency’)
subplot(1,2,2)
plot(w,Ph);grid
title(’Unwrapped phase response filter’)
ylabel(’Phase angle in radians’)
xlabel(’Normalized frequency’)
%end

The magnitude response and phase response of this IIR filter are plotted in
Figures 3.23 and 3.24, respectively.

Magnitude of the Magnitude in dB of the

frequency response frequency response
1 0

0.9

0.8 −50

0.7
−100
Magnitude in dB

0.6
Magnitude

0.5 −150

0.4

−200
0.3

0.2
−250
0.1

0 −300
0 1 2 3 4 0 1 2 3 4
Normalized frequency Normalized frequency

Figure 3.23 Magnitude response of an IIR filter.

 USE OF MATLAB TO COMPUTE DTFT 151

Phase response of the filter Unwrapped phase response filter

4 0

3
−2

−4
Phase angle in radians

Phase angle in radians

1

0 −6

−1
−8

−2

−10
−3

−4 −12
0 1 2 3 4 0 1 2 3 4
Normalized frequency Normalized frequency

Figure 3.24 Phase response of an IIR filter.

Example 3.15

Now we consider an FIR filter with a finite number of coefficients as

[1.0 0.8 0.6 0.4 0.6 0.8 1.0]. The MATLAB script used to compute its
frequency response and plot the magnitude and phase for the whole range of 0
to 2π is shown below:

%Program to compute and plot the frequency response or DTFT

% of an FIR filter
b=[1.0 0.8 0.6 0.4 0.6 0.8 1.0];
[h,w]=freqz(b,1,256,’whole’);
H=abs(h);
ph=angle(h);
plot(w,H);grid
title(’Magnitude of the FIR filter’)
ylabel(’Magnitude’)
xlabel(’Normalized frequency from 0 to 2*pi’)
figure
plot(w,ph);grid
152 FREQUENCY-DOMAIN ANALYSIS

Magnitude of the FIR filter

6

4
Magnitude

0
0 1 2 3 4 5 6 7
Normalized frequency from 0 to 2p

Figure 3.25 Magnitude response of an FIR filter.

title(’Phase response of the FIR filter’)

ylabel(’Phase angle in radians’)
xlabel(’Normalized frequency’)
%end

The magnitude and phase responses for the normalized frequency range from
0 to 2π are shown in Figures 3.25 and 3.26, respectively. The phase response is
found to be linear as a function of the frequency in this example. We will work
out many more examples of computing and plotting the DTFT or the frequency
response of filters in Chapter 4, using the MATLAB functions listed above.

Example 3.16

In this example, we choose the sampling frequency Fs = 200 Hz, and the Nyquist
interval is divided into 100 equal parts as seen in the statement f = [0:99] in the
MATLAB program given below. The sample values of the signal are entered by
us, when prompted by the program. In the example, we entered [0.4 0.6 0.8]
as the input signal. The magnitude and phase are plotted in Figure 3.27 as a
function of the frequency from 0 to 100 Hz. But the group delay is plotted as a
function of the normalized frequency from 0 to π radians.

% Program to compute and plot the magnitude and phase and

% group delay of an FIR filter
 USE OF MATLAB TO COMPUTE DTFT 153

Phase response of the FIR filter

2.5

1.5
Phase angle in radians

0.5

−0.5

−1

−1.5

−2

−2.5
0 1 2 3 4 5 6 7
Normalized frequency from 0 to 2p

Figure 3.26 Phase response of an FIR filter.

Magnitude response of the filter Magnitude response in dB

1.8 6
1.6 4
1.4 2
Magnitude in dB
Magnitude

1.2 0
1 −2
0.8 −4
0.6 −6
0.4 −8
0.2 −10
0 20 40 60 80 100 0 20 40 60 80 100
Frequency in Hz Frequency in Hz

Phase response Group delay response

4 4.5
3
Group Delay in samples

4
Phase angle in radians

2 3.5
1
3
0
2.5
−1
−2 2

−3 1.5
−4 1
0 20 40 60 80 100 0 1 2 3 4
Frequency in Hz Normalized frequency

Figure 3.27 Magnitude, phase, and group delay responses of an FIR filter.
154 FREQUENCY-DOMAIN ANALYSIS

b=input(’Type the values of the signal’);

a=1;
f=0:99;
Fs=200;
[H,w]=freqz(b,a,f,Fs);
mag=abs(H);
DB=20*log10(mag);
phase=angle(H);
[grp,w]=grpdelay(b,a,256);
subplot(2,2,1)
plot(f,mag);grid
title(’Magnitude response of the filter’)
ylabel(’Magnitude’)
xlabel(’Frequency in Hz’)
subplot(2,2,2)
plot(f,DB);grid
title(’Magnitude response in dB’)
ylabel(’Magnitude in dB’)
xlabel(’Frequency in Hz’)
subplot(2,2,3)
plot(f,phase);grid
title(’Phase response’)
ylabel(’Phase angle in radians’)
xlabel(’Frequency in Hz’)
subplot(2,2,4)
plot(w,grp);grid
title(’Group Delay response’))
ylabel(’Group Delay in samples’))
xlabel(’Normalized frequency’)
%end

3.6 DTFS AND DFT

3.6.1 Introduction
We discussed the DTFT-IDTFT pair for a discrete-time function given by
∞

X(ej ω ) = x(n)e−j ωn (3.67)
n=−∞

and

π
1
x(n) = X(ej ω )ej ωn dω (3.68)
2π −π
 DTFS AND DFT 155

The theory for deriving the pair and their properties and applications are very
elegant, but from practical point of view, we see some limitations in computing
the DTFT and IDTFT. For example, the input signal is usually aperiodic and may
be finite in length, but the unit impulse response of an IIR filter is also aperiodic
but infinite in length; however, the values of its samples become almost negligible
in many practical applications as n becomes large but finite. So in (3.67), it is
reasonable to assume that the number of terms is finite, but X(ej ω ) is a function
of the continuous variable ω. We have given some examples of analytically
deriving closed-form expressions for this function and plotting it as a function
of the variable ω. We showed how we can do it by using MATLAB functions.
Let us consider one more example of a discrete-time function x(n) and its
DTFT X(ej ω ). Figure 3.28a shows a nonperiodic discrete-time function x(n)

x(n)

01 2 3 4 5 6 7 8 X

(a)
|X(e jw)|

0 π 2π 3π 4π ω
(b)

∠X(e jw)
π

π 2π 3π 4π ω

−π
(c)

Figure 3.28 A nonperiodic signal; (b) its magnitude response; (c) its phase response.
156 FREQUENCY-DOMAIN ANALYSIS

 
that is of finite length. Figure 3.28b,c show the magnitude response X(ej ω )
and phase response X(ej ω ) of the DTFT X(ej ω ).
The function X(ej ω ) in (3.68) is a function of the continuous variable ω,
and the integration is not very suitable for computation by a digital computer.
Of course, we can discretize the frequency variable and find discrete values
for X(ej ωk ) where ωk are discrete values of the frequency. In contrast to the
case of a continuous-time signal with a frequency response X(j ω), we notice
that we need to compute the DTFT at only a finite number of values since
X(ej ω ) is periodic, and therefore we need to compute it over one period only. In
(3.68), x(n) can be computed approximately, if the integration is substituted by
a summation and such a summation will be finite because the values of X(ej ωk )
have to be chosen only over the interval [−π, π]. [We may also note that the
reconstruction formula used to obtain x(t) from its samples x(n) is not suitable
for digital computer, either.] These limitations are mitigated by a theory based on
the model for a discrete-time signal that is periodic, and in the next section, we
describe the discrete-time Fourier series (DTFS) representation for such discrete-
time periodic signals. This theory exploits the property of the DTFT that it is
periodic, and hence we need to use only a finite frequency range of one period
that is sufficient to find its inverse.

3.6.2 Discrete-Time Fourier Series

We consider a discrete-time aperiodic signal x(n) that is finite in length equal to
N samples and generate a periodic sequence with a period N so that it satisfies the
condition xp (n + KN ) = xp (n), where K is any integer. The complex Fourier
series representation of this periodic signal contains the sum of the discrete-time
complex exponentials ej kω0 n , where ω0 = 2π/N is its fundamental frequency and
ω0 k is its kth harmonic. The Fourier series isa weighted sum of the fundamental
and higher harmonics in the form xp (n) = k=0 Xp (k)ej ω0 kn , where Xp (k) is
the coefficient of the kth harmonic in the Fourier series. These coefficients are
complex-valued
  in general and hence have a magnitude and a phase: Xp (k) =
Xp (k) Xp (k)
In Chapter 1, we pointed out an important property of the discrete-time expo-
nentials (in contrast to the continuous-time exponentials ej ω0 kt ) that the exponen-
tials with frequencies that are separated by integer multiples of 2π are the same

ej (ω0 ±2π)k = ej ω0 k ej ±2πk = ej ω0 k (3.69)

and therefore the coefficient of the complex Fourier series Xp (k) = Xp (k ± N ).

Consequently there are only N independent harmonics (in contrast to the infinite
number of harmonics in the case of the continuous-time periodic function), and
we have used a subscript to denote that these coefficients Xp (k) are periodic.
 DTFS AND DFT 157

Therefore, in its complex Fourier series form, we have

N −1
xp (n) = Xp (k)ej (2π/N )kn (3.70)
k=0

To find these coefficients, let us multiply both sides by e−j mω0 k and sum over n
from n = 0 to (N − 1):

N−1
N −1 N
−1
−j mω0 k
xp (n)e = Xp (k)ej (2π/N )kn e−j mω0 k (3.71)
n=0 n=0 k=0

By interchanging the order of summation on the right side, we get

−1
N −1 

N
j (2π/N )k(n−m)
Xp (k) e (3.72)
k=0 n=0

 −1 j (2π/N )k(n−m)
It is next shown that [ N n=0 e ] is equal to N when n = m and zero
 −1 j 0
for all values of n = m. When n = m, the summation reduces to [ N n=0 e ] =
N , and when n = m, we apply (3.52) and find that the summation yields zero.
Hence there is only one nonzero term Xp (k)N in (3.72). The final result is

N −1
1
Xp (k) = xp (n)e−j nω0 k (3.73)
N
n=0

Now we notice that

N −1
1 1
N−1
xp (n)e−j nω0 k = x(n)e−j (2π/N )nk
N N
n=0 n=0
 
1 
= Xp (ej ω )ω =(2π/N )k = Xp (k) (3.74)
N k

In other words, when the DTFT of the finite length sequence x(n) is evaluated
at the discrete frequency ωk = (2π/N )k, (which is the kth sample when the
frequency range [0, 2π] is divided into N equally spaced points) and dividing
by N , we get the value of the Fourier series coefficient Xp (k).
The expression in (3.70) is known as the discrete-time Fourier series
(DTFS) representation for the discrete-time, periodic function xp (n) and (3.73),
which gives the complex-valued coefficients of the DTFS is the inverse DTFS
(IDTFS). Because both xp (n) and Xp (k) are periodic, with period N , we observe
158 FREQUENCY-DOMAIN ANALYSIS

that the two expressions above are valid for −∞ < n < ∞ and −∞ < k < ∞,
respectively. Note that some authors abbreviate DTFS to DFS.
To simplify the notation, let us denote e−j (2π/N )n by WN so that (3.70) and
(3.73) are rewritten in compact form for the DTFS-IDTFS pair as

N −1
xp (n) = Xp (k)W −kn , −∞ < n < ∞ (3.75)
k=0
N −1
1
Xp (k) = xp (n)W kn , −∞ < k < ∞ (3.76)
N
n=0

Figure 3.29a shows an example of the periodic discrete-time function xp (n)

constructed from
 Figure
 3.28a while Figures 3.29b,c show the samples of the
magnitude N Xp (k) and phase Xp (k) of the DTFT at the discrete frequencies
ωk = k(2π/N ), k = 0, 1, 2, . . . , (N − 1).
By comparing (3.73) and (3.67), we notice the following. Remember that
X(ej ω ) given by (3.67) is the DTFT of a nonperiodic sequence that may be of
finite or infinite length. But when we assume that the signal is of finite length N ,

Xp(n)

−20 −10 0 1 2 3 4 5 6 7 8 9 1011121314 20 n

(a)

|Xp(k )|

DFT |X(k )|
0 10 20 k
(b)

Xp(k )
p

10 20 k

(c)

Figure 3.29 A periodic signal; (b) its magnitude response; (c) its phase response.
 DTFS AND DFT 159

and have xp (n) = x(n) for n = 0, 1, 2, . . . , (N − 1), we have its DTFT as

N −1
1
jω
X(e ) = x(n)e−j ωn
N
k=0

In (3.73) we evaluate the DTFT at N discrete frequencies ωk and divide these

samples by N to get Xp (k). From (3.70), we see that the coefficients Xp (k) of the
DTFS for xp (n) are also the coefficients of the Fourier series for the discrete-time
periodic sequence xp (n), and they are also referred to as the spectral components
of xp (n). In other words, the Fourier series coefficients of a periodic signal xp (n)
is 1/N times the DTFT of x(n) evaluated at the discrete frequencies ωk . We
repeat that Xp (k) = (1/N )X(ej ωk ), where X(ej ω ) is the DTFT of a sequence
x(n). This result is based on the assumption that x(n) is a sequence of finite
length N .

3.6.3 Discrete Fourier Transform

We will discuss some properties of the DTFS-IDTFS pairs later in this section.
Here we use the argument that if (3.75) is valid for all values of n (i.e., for
−∞ < n < ∞), it is valid for 0 ≤ n ≤ N − 1 or for any sequence of length
N ; similarly, (3.76) is valid for 0 ≤ k ≤ N − 1. Thus we denote Xp (k); 0 ≤
n ≤ N − 1 by the notation X(k), whereas we already have represented x(n) =
xp (n); 0 ≤ n ≤ N − 1:

N −1
N −1
x(n) = X(k)ej (2π/N )kn = X(k)W −kn , 0≤n≤N −1 (3.77)
k=0 k=0
N −1
1 1
N−1
−j (2π/N )kn
X(k) = x(n)e = x(n)W kn , 0≤k ≤N −1 (3.78)
N N
n=0 n=0

Note that we have not derived these properties from any new theory but only
defined them as a part of the infinite sequences for the DTFS and IDTFS derived
above.
Also note that whereas (3.75) is termed the discrete-time Fourier series (DTFS)
representation of xp (n), in which Xp (k) are the coefficients of the Fourier
series [and (3.76) is the IDTFS], it is (3.78) that is known as the discrete-time
Fourier transform (DFT) [and (3.77) is known as the inverse DFT (IDFT)]! In
Sections 3.6.1 and 3.6.2, note that we have used different notations to distinguish
the DTFT-IDTFT pair from the DFT-IDFT pair.
In most of the textbooks, and in MATLAB, the DFT-IDFT are simply defined
as given below, without any reference to the theory for deriving the DTFS, from
which the DFT are selected. Also note that the scale factor (1/N ) has been moved
from (3.78) to (3.77) in defining the DFT-IDFT pair as shown in (3.79) and (3.80)
160 FREQUENCY-DOMAIN ANALYSIS

(we will use these two equations for the DFT-IDFT pair in the remaining pages):
N −1 N −1
1 1
x(n) = X(k)ej (2π/N )kn = X(k)W −kn , 0≤n≤N −1 (3.79)
N N
k=0 k=0

N −1
N −1
X(k) = x(n)e−j (2π/N )kn = x(n)W kn , 0≤k ≤N −1 (3.80)
n=0 n=0

In Figure 3.29b, we have shown the DFT as a subset of the Fourier series
coefficients Xp (k), for k = 0, 1, 2, . . . , (N − 1). But we can choose any other
N consecutive samples as the DFT of x(n) [e.g., −[(N − 1)/2] ≤ n ≤ [(N −
1)/2]], so we will use the notation N  = n modulo N to denote that n ranges
over one period of N samples.
Given a nonperiodic discrete-time function x(n), we constructed a mathe-
matical artifact xp (n) and derived the Fourier series representation for it and
also derived its inverse to get xp (n). Then we defined the DFT and IDFT as
argued above so that we could determine the frequency response of the non-
periodic function as samples of the DTFT X(ej ω ) at N equally spaced points
ωk = (2π/N )k. We know x(n) is nonperiodic, but since X(ej ω ) is periodic with
a period 2π, X(ej ωk ) = Xp (k) is periodic with a period N , so one can choose
the range n = N  in Equations (3.79) and (3.80).
The two equations for the DFT and IDFT give us a numerical algorithm
to obtain the frequency response at least at the N discrete frequencies, and
by choosing a large value for N , we get a fairly good idea of the frequency
response for x(n).6 Indeed, we show below that from the samples of X(k), we
can reconstruct the DTFT of x(n) = X(ej ω ), which is a function of the contin-
uous variable ω. This is the counterpart of Shannon’s reconstruction formula to
obtain x(t) from its samples x(n), provided x(n) is bandlimited and the sam-
pling period Ts < (π/ωb ). There are similar conditions to be satisfied in deriving
the formula in the frequency domain, to reconstruct X(ej ω ) from its samples
X(ej ωk ) = X(k).

3.6.4 Reconstruction of DTFT from DFT

First we consider the DTFT of x(n) and substitute x(n) by the formula (3.79)
for finding the IDFT of X(k) as explained below:
−1 −1
 N −1 

N
N
1
X(ej ω ) = x(n)e−j ωn = X(k)ej (2πkn/N) e−j ωn
N
n=0 n=0 k=0

N −1
N −1
1
= X(k) ej (2πkn/N) e−j ωn (3.81)
N
k=0 n=0

6
Later we will discuss what is known as the “picket fence effect,” because of which we may not get
a fairly good idea of the frequency response.
 DTFS AND DFT 161

N −1
Now we use (3.52) in the summation n=0 ej (2πkn/N) e−j ωn and reduce it as
follows:

N −1
1 − e−j (ωN −2πk )
ej (2πkn/N) e−j ωn =
1 − e−j [ω−(2πk/N)]
n=0
 
ωN − 2πk
sin
e−j [(ωN−2πk)/2] 2
= −j [(ωN−2πk)/2N ·  
e ] ωN − 2πk
sin
2N
 
ωN − 2πk
sin
2
=   e−j [ω−(2πk/N)][(N −1)/2] (3.82)
ωN − 2πk
sin
2N

Substituting the last expression in (3.81), we obtain the final result to reconstruct
the DTFT X(ej ω ), from only the finite number of the DFT samples X(k), as
given below:
 
ωN − 2πk
N −1 sin
1 2
X(ej ω ) = X(k)   e−j [ω−(2πk/N)][(N −1)/2] (3.83)
N ωN − 2πk
k=0 sin
2N

If x(n) has M samples and we sample X(ej ω ) at N points in the range [0, 2π],
where N > M, then the N -point IDFT will yield N samples in the discrete-
time domain. It can be shown that this result will give rise to aliasing of the
N -point sequences (in the time domain). So we pad the given function x(n) with
(M − N ) zeros to make it a discrete-time function of length N ; otherwise, we
have to choose N ≤ M. In that case the sampling interval satisfies the condition
(2π/N ) ≥ (2π/M), which is dual to the condition that the sampling period T ≤
(1/2B) to be satisfied for Shannon’s reconstruction formula in the time domain.
To satisfy this condition for reconstruction in the frequency domain, we make
M = N by padding with zeros x(n) to avoid aliasing in the discrete-time domain.

3.6.5 Properties of DTFS and DFT

We have already listed some of the properties of DTFT in Table 3.1. Since the
samples X(k) of DFT are a subset of the Fourier series coefficients Xp (k), there
is some similarity between the properties of Xp (k) and X(k) and also those of
the DTFT X(ej ω ) of x(n). Tables 3.3 and 3.4 list the properties of Xp (k) and
X(k), respectively, which also show some similarities.
Remember again that the DFT X(k) is defined over one period usually for
k = 0, 1, 2, . . . , (N − 1) and it is considered to be zero outside this range. So
162 FREQUENCY-DOMAIN ANALYSIS

TABLE 3.3 Properties of DTFs Coefficients

Periodic Signal DTFS Coefficients Xp (k)

Property x(n) and y(n) with Period N Xp (k) and Yk (k)

Linearity ax(n) + by(n) aXp (k) + bYp (k)

Periodic x(n)  y(n) Xp (k)Yp (k)
convolution
Time shifting x(n − k) e−j (2π/N )km Xp (k) = WNkm Xp (k)
Frequency shifting ej (2π/N )mn x(n) = W −mn x(n) Xp (k − m)
Time reversal x(−n) Xp (−k)
1 N −1
Multiplication x(n)y(n) N m=0 Xp (m)Yp (k − m)

Conjugation x ∗ (n) Xp∗ (−k)

   
Even part of x(n) xe (n) = 1
2x(n) + x ∗ (−n) Re Xp (k)
   
Odd part of x(n) x0 (n) = 12 x(n) − x ∗ (−n) j Im Xp (k)
Symmetry x(n) is a real sequence Xp (k) = Xp∗ (−k)
Xp (k) = Xp∗ (−k)
ReXp (k) = ReXp (−k)
ImXp (k) = −ImXp (−k)
   
Xp (k) = Xp (−k)
Xp (k) = −Xp (−k)

TABLE 3.4 Properties of DFT Coefficients

Signal DFT Coefficients

Property x(n) and y(n) of Length N X(k) and Y (k) of Length N

Linearity ax(n) + by(n) aX(k) + bY (k)

N −1
Convolution m=0 x(m)y((n − m))N X(k)Y (k)
Time shifting x(n − m)N e−j (2π/N )km X(k) = WNkm X(k)
Frequency shifting ej (2π/N )mn x(n) = W −mn x(n) X((k − m))N
1 N −1
Multiplication x(n)y(n) N m=0 X(m)Y ((k − m))N

Conjugation x ∗ (n) X ∗ ((−k))N

 
Even part of x(n) xe (n) = 1
2x(n) + x ∗ ((−n))N ReX(k)
 
Odd part of x(n) x0 (n) = 12 x(n) − x ∗ ((−n))N j ImX(k)
X(k) = X ∗ ((−k))N
ReX(k) = ReX((−k))N
ImX(k) = −ImX((−k))N
|X(k)| = |X((−k))N |
X(k) = −Xp ((−k))N
 DTFS AND DFT 163

in Table 3.4, we have used a notation such as X((−k))N , which means that we
choose DTFS coefficients Xp (k), use time reversal to get Xp (−k), and then
select any N samples that form a period. The double bracket with a subscript
N calls for three operations: choosing the DTFS coefficients Xp (k), carrying
out the operation indicated by the index within the brackets, and then selecting
n = 0, 1, 2, . . . , (N − 1) or n modulo N . This again confirms the statement that
all operations are carried out by the DTFS and then one period of the result
is chosen as the DFT of x(n). This is very significant when we carry out the
periodic convolution of the DTFS of x(n) and f (n) and select the values of this
convolution for n = 0, 1, 2, . . . , (N − 1). We illustrate this by Example 3.17.

Example 3.17

Let x(n) = [1.0 1.0 0.6 0.6] and f (n) = [1.0 0.6 0.4]. We can easily
find the output y1 (n) by using
 either one of two methods: (1) the convolution
−1
sum y1 (n) = x(n) ∗ f (n) = N m=0 x(m)f (n − m) or (2) one of the two trans-
forms, namely, the z transform or the discrete-time Fourier transform (DTFT) of
x(n) and f (n), and find the
 inverse z transform of [X(z)F (z)] or the inverse
DTFT of X(ej ω )F (ej ω ) . Indeed, we can give a proof to show that the z

x(m)
f(−m)

0 1 2 3 m −3 −2 −1 0 1 2 3 m
(a) (b)

f(1−m) f(2−m)

−2 −1 0 1 2 3 m −1 0 1 2 3 m
(c) (d)
f(4−m)
f(3−m)

0 1 2 3 m 0 1 2 3 4 m
(e) (f)

Figure 3.30 Linear convolution of x(n) and f (n).

164 FREQUENCY-DOMAIN ANALYSIS

f(5−m)
f(6−m)

0 1 2 3 4 5 m 0 1 2 3 4 5 6 m
(g) y1(n) (h)

0 1 2 3 4 5 n
(i)

Figure 3.30 (Continued )

transform of the convolution sum y1 (n) = x(n) ∗ f (n) is [X(z)F (z)]. There-
fore, the results from both these methods agree with each other and we get
the output y1 (n) = [1.0 1.6 1.6 1.36 0.6 0.24], which is a sequence of
length 6. It is identified as the result of linear convolution. The three sequences
are shown in Figure 3.30, including the graphical procedure for carrying out the
linear convolution.
Now we ask the following question. What do we get if we compute the DTFS
of the periodic sequences xp (n) and fp (n) generated by extending x(n) and
f
 (n), multiply Xp (k), and Fp (k), and find the inverse DTFS of their product 
Xp (k)Fp (k) to get yp (n). From Table 3.3, we notice that Xp (k)Fp (k) is
N −1
the DTFS of the periodic convolution x(n)  f (n) = m=0 x(m)f (n − m)N =
yp (n)—a result than can be proved analytically. We will provide a numerical
example below to verify this property. However, does this result of periodic
convolution match with the result of applying the familiar linear convolution?
We show in the example chosen below that yp (n) is not the periodic extension
of y1 (n), that is, the result of the periodic convolution and the linear convolution
do not match in the example chosen.

Example 3.18

 −1of x(n),−jwhen
The DTFS coefficients we choose N = 4, are computed from the
formula Xp (k) = Nn=0 x(n)e (2π/N )kn

3
Xp (0) = x(n)e−j (2π/N )(0.n) = x(0)e−j (2π/4)(0) + x(1)e−j (2π/4)(0)
n=0

+ x(2)e−j (2π/4)(0) + x(3)e−j (2π/4)(0) = 3.2

 DTFS AND DFT 165

3
Xp (1) = x(n)e−j (2π/N )(1.n) = x(0)e−j (2π/4)(0) + x(1)e−j (2π/4)(1)
n=0

+ x(2)e−j (2π/4)(2) + x(3)e−j (2π/4)(3) = 0.4 − j 0.4

3
Xp (2) = x(n)e−j (2π/N )(2.n) = x(0)e−j (2π/4)(0) + x(1)e−j (2π/4)(2)
n=0

+ x(2)e−j (2π/4)(4) + x(3)e−j (2π/4)(6) = 0.0 + j 0.0

3
Xp (3) = x(n)e−j (2π/N )(3.n) = x(0)e−j (2π/4)(0) + x(1)e−j (2π/4)(3)
n=0

+ x(2)e−j (2π/4)(6) + x(3)e−j (2π/4)(9) = 0.4 + j 0.4

Similarly, the DTFT of f (n) are computed as

3
Fp (0) = f (n)e−j (2π/N )(0.n) = f (0)e−j (2π/4)(0) + f (1)e−j (2π/4)(0)
n=0

+ f (2)e−j (2π/4)(0) + f (3)e−j (2π/4)(0) = 2.0

3
Fp (1) = f (n)e−j (2π/N )(1.n) = f (0)e−j (2π/4)(0) + f (1)e−j (2π/4)(1)
n=0

+ f (2)e−j (2π/4)(2) + f (3)e−j (2π/4)(3) = 0.6 − j 0.6

3
Fp (2) = f (n)e−j (2π/N )(2.n) = f (0)e−j (2π/4)(0) + f (1)e−j (2π/4)(2)
n=0

+ f (2)e−j (2π/4)(4) + f (3)e−j (2π/4)(6) = 0.8 + j 0.0

3
Fp (3) = f (n)e−j (2π/N )(3.n) = f (0)e−j (2π/4)(0) + f (1)e−j (2π/4)(3)
n=0

+ f (2)e−j (2π/4)(6) + f (3)e−j (2π/4)(9) = 0.6 + j 0.6

The term-by-term product of these vectors gives the DTFS of the output
Yp (k) = Xp (k)Fp (k) as
Yp (k) = [6.4 0 − j 0.48 0 + j 0 0.0 + j 0.48]
and its inverse DTFS is computed from the formula

1
3
yp (n) = Yp (k)ej (2π/N )kn
N
k=0

and we get yp (n) = [1.6 1.84 1.6 1.36].

166 FREQUENCY-DOMAIN ANALYSIS

Using the formula for the periodic convolution x(n)  f (n) =

N−1
m=0 x(m)f (n − m)N directly, as illustrated in Figure 3.31, we get the
same result for yp (n). This verifies the property that the DTFS of the periodic
convolution x(n)  f (n) is Xp (k)Fp (k).

xp(m)

−10 −9 −8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8 9 10 11
(a)

fp(m)

−10 −9 −8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8 9 10 11 12 13
(b)
fp(−m)

−10 −9 −8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8 9 10 11 12 13
(c)
fp(1−m)

−10 −9 −8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8 9 10 11 12 13
(d)
fp(2−m)

−10 −9 −8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8 9 10 11 12
(e)

Figure 3.31 Circular (periodic) convolution of xp (n) and fp (n) with N = 4.

 DTFS AND DFT 167

fp(3−m)

−10 −9 −8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8 9 10 11 12
(f)

f(4−m)

−11 −10 −9 −8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8 9 10 11 12 m
(g)

fp(5−m)

−10 −9 −8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8 9 10 11 12 m
(h)

y(n)

−12 −11 −10 −9 −8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 n

(i )

Figure 3.31 (Continued )

But we notice that this does not match the result of the linear convolution
y1 (n) = x(n) ∗ f (n) = [1.0 1.6 1.6 1.36 0.6 0.24]. It is obvious that the
length of y1 (n) is 6, whereas xp (n), fp (n), Xp (k), Fp (k), and yp (n) are all of
length 4, and for that reason alone, we do not expect the two results to match.
If we look carefully at Figure 3.31 and Figure 3.30, we see another reason why
they do not match. In Figures 3.31, f (4 − m), f (5 − m), f (6 − m), f (7 − m)
are found to be the same as the sequence f (0), f (1 − m), f (2 − m), f (3 − m),
168 FREQUENCY-DOMAIN ANALYSIS

respectively; therefore yp (n) is periodic with a period N = 4, whereas y1 (n) is

nonperiodic and is of length 6. When computing yp (n), we see that products of
the additional terms from the adjacent period are added, whereas when computing
y1 (n), no such overlap takes place. So in order to avoid this overlap, we choose
N = 8 and compute Xp (k), Fp (k) and yp (n) as follows.

Example 3.19

7
Xp (0) = x(n)e−j (2π/N )(0·n)
n=0

= x(0)e−j (2π/8)(0) + x(1)e−j (2π/8)(0) + x(2)e−j (2π/8)(0)

+ x(3)e−j (2π/8)(0) + · · · + x(7)e−j (2π/8)(0) = 3.2

7
Xp (1) = x(n)e−j (2π/N )(1·n)
n=0

= x(0)e−j (2π/4)(0) + x(1)e−j (2π/8)(1) + x(2)e−j (2π/8)(2)

+ x(3)e−j (2π/8)(3) + · · · + x(7)e−j (2π/8)(7) = 1.2828 − j 1.7314

7
Xp (2) = x(n)e−j (2π/N )(2·n)
n=0

= x(0)e−j (2π/8)(0) + x(1)e−j (2π/8)(2) + x(2)e−j (2π/8)(4)

+ x(3)e−j (2π/8)(6) + · · · + x(7)e−j (2π/8)(14) = 0.4 − j 0.4

7
Xp (3) = x(n)e−j (2π/N )(3·n)
n=0

= x(0)e−j (2π/8)(0) + x(1)e−j (2π/8)(3) + x(2)e−j (2π/8)(6)

+ x(3)e−j (2π/4)(9) + · · · + x(7)e−j (2π/8)(21) = 0.7172 − j 0.5314

7
Xp (4) = x(n)e−j (2π/N )(4·n)
n=0

= x(0)e−j (2π/8)(0) + x(1)e−j (2π/8)(4) + x(2)e−j (2π/8)(8)

+ x(3)e−j (2π/4)(12) + · · · + x(7)e−j (2π/8)(28) = 0.0 + j 0.0

7
Xp (5) = x(n)e−j (2π/N )(5·n)
n=0

= x(0)e−j (2π/8)(0) + x(1)e−j (2π/8)(5) + x(2)e−j (2π/8)(10)

+ x(3)e−j (2π/4)(15) + · · · + x(7)e−j (2π/8)(35) = 0.7172 + j 0.5314
 DTFS AND DFT 169

7
Xp (6) = x(n)e−j (2π/N )(6·n)
n=0

= x(0)e−j (2π/8)(0) + x(1)e−j (2π/8)(6) + x(2)e−j (2π/8)(12)

+ x(3)e−j (2π/4)(18) + · · · + x(7)e−j (2π/8)(56) = 0.4 + j 0.4

7
Xp (7) = x(n)e−j (2π/N )(7·n)
n=0

= x(0)e−j (2π/8)(0) + x(1)e−j (2π/8)(7) + x(2)e−j (2π/8)(14)

+ x(3)e−j (2π/4)(21) + · · · + x(7)e−j (2π/8)(49) = 1.2828 + j 1.7314.

Similarly, we compute Fp (k) with N = 8 and get the vector

Fp (k) = [2.0 1.4243 − j 0.8423 0.6 − j 0.6 0.5757 − j 0.0243

0.8 + j 0.0 0.5757 + j 0.0243 0.6 + j 0.6 1.4243 + j 0.8423].

The term-by-term multiplication of Xp (k) and Fp (k) yields

Yp (k) = [6.4 0.4 − j 3.5233 0.0 − j 0.48 0.4 − j 0.3233 0.0 + j 0.0
0.4+j 0.3233 0.0 + j 0.48 0.4 + j 3.5233].

and the inverse DTFS of Yp (k) given by yp (n) = 1/N 7k=0 Yp (k)ej (2π/N )kn is
computed to obtain yp (n) = [1.0 1.6 1.6 1.36 0.6 0.24 0.0 0.0].
As shown in Figures 3.32 and 3.33, we get the same result from the periodic
convolution. Moreover, we see that this result matches the result y1 (n) obtained
by linear convolution!
In general, if the length of x(n) is l1 and that of f (n) is l2 , we know that the
length of y1 (n) from linear convolution will be l1 + l2 − 1. So what we need to do
to match the result of linear convolution and periodic convolution of two signals
is to choose N to be equal to or greater than l1 + l2 − 1. With such a choice, we
can use the DTFS coefficients Xp (k) and Fp (k), each of length N ≥ l1 + l2 − 1,
and then compute the N inverse DTFS coefficients of Xp (k)Fp (k). Because the
formulas for computing the N coefficients of their DFT [i.e., X(k) and F (k)] are
the same as for computing their DTFS coefficients and consequently the DFT
(and inverse DFT) coefficients are a subset of the coefficients of DTFS (and
IDTFS), we conclude that if we are given, say, x(n) of length l1 as the input
signal and h(n) of length l2 as the unit impulse response of a linear discrete-time
system, then we can pad each of the signals with an appropriate number of zeros
to make both of them to be of length N ≥ l1 + l2 − 1, and find the inverse DFT
of X(k)H (k) to get the N samples of the output y(n) of the linear discrete-
time system. Conversely, if we are given any signal, we can easily obtain the N
170 FREQUENCY-DOMAIN ANALYSIS

xp(1−m)

01234567 m
fp(−m)

012345678 m
fp(1−m)

0 7 m
fp(2−m)

0 12 7 m
fp(3−m)

01234567 m

Figure 3.32 Circular convolution of xp (n) and fp (n) with N = 8.

coefficients of its DFT, which indicates the frequency response of the signal at
N discrete frequencies equally spaced between 0 and 2π.

3.7 FAST FOURIER TRANSFORM

Note that when we computed each of the eight samples of the DFT in the previous
example, there was multiplication of the complex number ej (2π/N )kn = W −kn ,
k = 0, 1, 2, . . . , (N − 1), with the eight real-valued samples of the signal and
the product were added. So the total number of multiplications is 82 = 64 and
the number of additions are 72 = 49 in computing the eight samples of the
DFT. The same number of multiplications and additions are required to find the
inverse DFT; in this case, samples of both X(k) and W −kn are complex-valued.
In general, direct computation of the DFT and IDFT using (3.79) and (3.80)
requires N 2 multiplications and (N − 1)2 additions; so they become very large
 FAST FOURIER TRANSFORM 171

fp(4−m)

0 1 2 3 4 5 6 7

fp(5−m)

−8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8

fp(6−m)

−8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8

y(n)

−8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8 9 10 11 12

Figure 3.33 Circular convolution of xp (n) and fp (n) with N = 8 (continued from
Fig. 3.34).

numbers when N is chosen very large, in order to increase the resolution of the
frequency response X(k) of a given signal or to find the unit impulse response
of a filter as the IDFT of the given frequency response of a filter.
Fast Fourier transform (FFT) is a numerical algorithm that has been developed
to improve the computational efficiency by an enormous amount and is the most
popular method used in spectral analysis in digital signal processing, specifically,
to find the DFT of the signal and also the inverse DFT of the frequency response
to get the discrete-time signal. This is only a computational algorithm and not
172 FREQUENCY-DOMAIN ANALYSIS

another transform. In this FFT algorithm, when the value for the radix N is
chosen as 2R , where R is an integer, the number of complex multiplications is
of the order (N/2) log2 (N ) and the number of complex additions is of the order
N log2 N . As an illustration of this efficiency, let us choose N = 256; in this
case the number of complex multiplications is 65,536 in the direct computation,
whereas the number of complex multiplications is 1024 in the FFT algorithm,
which is an improvement by a factor of 64. As N increases to higher values, the
improvement factor increases very significantly, for example, when N = 1024,
we realize an improvement by a factor of 204.
Algorithms based on a radix N = 4 have been developed to further improve
the computational efficiency. Also when the length of a signals to be convolved
is large (e.g., N = 1024), some novel modifications to the FFT algorithm have
also been proposed. They are called the overlap-add method and the overlap-save
method. Basically in these methods, the signals are decomposed as a sequence
of contiguous segments of shorter length, their convolution is carried out in the
basic form, and then the responses are carefully added to get the same result
as that obtained by the direct FFT method applied on their original form. The
MATLAB function y = fftfilt(b,x) and y = fftfilt(b,x,N) implements
the convolution between the input signal x and the unit impulse response b of
the FIR filter, using the overlap-add method, the default value for the radix, is N
512; but it can be changed to any other value by including it as an argument in
the second command.

3.8 USE OF MATLAB TO COMPUTE DFT AND IDFT

Example 3.20

Let us consider the same example for the signal x(n) that was chosen in the
previous example; that is, let x(n) = [1.0 1.0 0.6 0.6]. First we compute
its DTFT and plot it in Figure 3.34. Then we compute a 10-point DFT of the
same signal using the function fft found in the Signal Processing Toolbox of
MATLAB. The function is described by the following simple command:

[X,w] = fft(x,N)

In this function, X is the output vector of the complex-valued DFT of the given
signal x(n) and N is the value for the radix, which is chosen as 10 in this example.
The absolute value of the DFT is computed, and the magnitude |X(k)|, k =
0, 1, 2, . . . , 9 is superimposed on the same plot. It is seen that the values of DFT
match the value of DTFT at the discrete frequencies ej (2π/10)k , k = 0, 1, 2, . . . , 9,
as we expect. But we have chosen this example with N = 10 particularly to
illustrate what is known as the “picket fence effect”. Note that the frequency
response in Figure 3.34 has a local minimum value at the normalized frequency
of 2.5 and 7.5. But if we plot the DFT values alone, we will miss the fact that the
frequency response of the signal has a minimum value at these frequencies. This
 USE OF MATLAB TO COMPUTE DFT AND IDFT 173

DTFT and DFT values of a finite length signal

3.5

3
Magnitude of DTFT and DFT

2.5

1.5

0.5

0
0 1 2 3 4 5 6 7 8 9 10
Normalized frequency 5∗w/p and k

Figure 3.34 DTFT/DFT plot.

phenomenon becomes serious if the signal has a single sinusoidal frequency, in

which case the frequency response will have an impulse function and will appear
as a spike at the corresponding normalized frequency. Because we have chosen
a low value for N , and therefore the fundamental frequency 2π/N is relatively
large, it is likely that the presence of this single frequency will be completely
missed. This is known as the “picket fence effect,” as if we “see” the continuous
frequency response through a picket fence and miss some part of the frequency
response that lies behind the pickets. Hence it is important to select a fairly
high value for the radix N so that we have less chance of missing the single
frequencies but at the same time not too high a value, which would increase the
computation time.
Students are strongly recommended to have a thorough understanding of the
theory while they use MATLAB in signal processing. In this example, when we
compute the DFT using the MATLAB function, its output is actually displayed
for k = 1, 2, 3, . . . , N , since the MATLAB function computes the fft of x(n)
according to the algorithm

N −1
X(k + 1) = x(n + 1)WNkn (3.84)
n=0

and the inverse DFT is computed for n = 1, 2, 3, . . . , N from

N −1
1
x(n + 1) = X(k + 1)W −kn (3.85)
N
k=0
174 FREQUENCY-DOMAIN ANALYSIS

whereas from theory we know that both indices run from 0 to (N − 1) and the
frequency response is normally displayed by MATLAB for the frequency range
of [0 π].
In superimposing the values of DFT on the plot for the DTFT, this fact about
the MATLAB function fft is important. It serves as an example where a thor-
ough understanding of theory is necessary for using MATLAB in digital signal
processing.

Example 3.21

We now consider another example showing the use of the MATLAB function
fft(x,N) and comparing the values of its DFT with the frequency response
(DTFT) of a discrete-time signal. We pick a signal x(n) = sin[0.1(πn)] for 0 ≤
n ≤ 10, which is plotted in Figure 3.35. We find its frequency response using the
function [h,w]=freqz(x,1,’whole’) and plot it for the full period of 2π in
Figure 3.36, in order to compare it with the values of its DFT X(k), which always
gives N samples for 0 ≤ k ≤ (N − 1), which corresponds to the full frequency
range of [0 2π]. The DFT X(k) are computed from the MATLAB function
X=fft(x,64). The  absolute
 values of X(k) are plotted in Figure 3.37 showing
that they match X(ej ω ).

Example 3.22

In this example, we consider the same signal x(n) = [1 1 0.6 0.6] and
h(n) = [1 0.6 0.4], which we considered in Example 3.17 and find the output

Samples of the signal x(n)

1

0.9

0.8
Values of the signal samples

0.7

0.6

0.5

0.4

0.3

0.2

0.1

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

Figure 3.35 A discrete-time signal.

 USE OF MATLAB TO COMPUTE DFT AND IDFT 175

Magnitude of the Frequency Response

7

5
Magnitude

0
0 1 2 3 4 5 6 7
Normalized frequency in radians

Figure 3.36 Magnitude response of the DTFT of a signal.

Magnitude of DFT of x(n)

7

5
Magnitude

0
0 10 20 30 40 50 60 70
Index for k

Figure 3.37 Magnitude of the DFT of a signal.

y(n) of the discrete-time system using the FFT technique. The MATLAB program
we use and the final output from the program are given below:

x = [ 1 1 0.6 0.6];
h = [ 1 0.6 0.4];
176 FREQUENCY-DOMAIN ANALYSIS

X=fft(x,8);
H=fft(h,8);
Y=X.*H;
y=ifft(Y,8)

The output is

y = 1.0000 1.6000 1.6000 1.3600 0.6000 0.2400 0 0

In this program, the MATLAB function y = ifft(Y,8) computes the 8-point

IFFT of the product X(k)H (k) and the last line lists the output y(n), which is,
not surprisingly, the same as the result of linear convolution x(n) ∗ h(n).

Example 3.23

Let us consider the DFT samples of a lowpass filter response as given below
and use the MATLAB function x = ifft(X,8) and plot the output x(n). The
MATLAB program and the output x(n) are given below and the plot of x(n) is
shown in Figure 3.38. Again note that the time index n in this figure runs from
1 to 8, instead of from 0 to 7 in this 8-point IDFT:

X=[1 exp(-j*7*pi/8) 0 0 0 0 0 exp(j*7*pi/8)];

x=ifft(X,8);
stem(x)

Values of the IDFT

0.5

0.4

0.3
Values of the IDFT x(n)

0.2

0.1

−0.1

−0.2
1 2 3 4 5 6 7 8
Index n + 1

Figure 3.38 IDFT of a lowpass filter.

 SUMMARY 177

title(’Values of the IDFT’)

ylabel(’Values of the IDFT’)
xlabel(’Index n’)

The output is

−0.1060 0.0293 0.2207 0.3560 0.3560 0.2207 0.0293 −0.1060

3.9 SUMMARY

We have discussed several topics in this chapter. First we showed that if a signal
is bandlimited and we sample it at a frequency larger than twice the maximum
frequency in the signal, we can use digital signal processing of the signal instead
of analog signal processing, because the digital signal has all the information that
is contained in the analog signal. Shannon’s sampling theorem and the formula
for reconstructing the analog signal from the samples was explained. When such
an analog signal is sampled, the frequency response of the discrete-time signal
is a continuous function of the digital frequency but is periodic with a period
equal to the sampling frequency. Next we discussed several properties of the
frequency response of the discrete-time signal (DTFT), illustrating them with
numerical examples; they are summarized below:

DTFT of a nonperiodic signal x(n):

∞

X(e ) =
jω
x(n)e−j ωn (3.86)
n=0

IDTFT of X(ej ω ):

π
1
x(n) = X(ej ω )ej ωn dω (3.87)
2π −π

Then we considered a discrete-time signal that is periodic with a period N . Its

Fourier series representation was expressed in terms of its Fourier series coeffi-
cients, and the formula for finding the values of the periodic discrete-time signal
from the Fourier series coefficients was presented. Properties of the discrete-time
Fourier series (DTFS) and the inverse DTFS were discussed and illustrated with
examples. They are summarized as follows:

DTFS of a periodic signal with period N :

N −1
N −1
xp (n) = X(k)ej (2π/N )kn = X(k)W −kn , −∞ ≤ n ≤ ∞ (3.88)
k=0 k=0
178 FREQUENCY-DOMAIN ANALYSIS

IDTFS of X(k) with period N :

N −1 N −1
1 1
Xp (k) = x(n)e−j (2π/N )kn = x(n)W kn , −∞ ≤ k ≤ ∞
N N
n=0 n=0
(3.89)

The discrete Fourier transform (DFT) and its inverse (IDFT) are a subset of
the DTFS and IDTFS coefficients, derived from the periodic DTFS and IDTFS
coefficients. They can be considered as nonperiodic sequences. A few examples
were worked out to show that the values of the DTFT, when evaluated at the
discrete frequencies, are the same as the DFT coefficients:

DFT of x(n) with length N :

N−1
N −1
X(k) = x(n)e−j (2π/N )kn = x(n)W kn , 0 ≤ k ≤ (N − 1)
n=0 n=0
(3.90)
IDFT of X(k) with length N :
N −1 N −1
1 1
x(n) = X(k)ej (2π/N )kn = X(k)W −kn , 0 ≤ n ≤ (N − 1)
N N
k=0 k=0
(3.91)

The FFT algorithm for computing the DFT-IDFT coefficients offers very
significant computational efficiency and hence is used extensively in signal pro-
cessing, filter analysis, and design. It provides a unified computational approach to
find the frequency response from the time domain and vice versa. More examples
are added to show that the use of FFT and IFFT functions from MATLAB pro-
vides a common framework for getting the frequency response of a discrete-time
system from the discrete-time signal and finding the discrete-time signal from
the frequency response. Remember that the terms discrete-time (digital) signal,
sequence, or function have been used interchangeably in this book; we have also
used the terms discrete-time Fourier transform (DTFT), frequency response, and
spectrum synonymously in this chapter.

PROBLEMS

3.1 A signal f (t) = e−0.1t u(t) is sampled to generate a DT signal f (n) at such
a high sampling rate that we can assume that there is no aliasing. Find a
closed-form expression for the frequency response of the sequence f (n).
3.2 Find the Fourier transform X(j ω) of the signal x(t) = te−0.1t u(t) and
choose a frequency at which the attenuation is more than 60 dB. Assuming
 PROBLEMS 179

|F(jw)|

Cos(2w)

0 w

Figure 3.39 Problem 3.4

that the signal is bandlimited by that frequency, what is the minimum sam-
pling frequency one can choose to sample x(t) without losing too much
information?
3.3 A continuous-time function f (t) with a bandwidth of 200 Hz is sam-
pled at 1000 Hz, and the sampled values are given by f (nT ) = {1.0
↑
0.4 0.1 0.001}. Find the value of f (t) at t = 0.005.
3.4 A bandlimited analog signal f (t) has a Fourier transform F (j ω) as shown
in Figure 3.39. What is the maximum sampling period T that can be used
to avoid aliasing in the frequency response F (ej ω ) of the sampled sequence
f (n)? Find the Fourier series coefficients for F (ej ω ).
3.5 Find the DTFT of x(n) = {−1 1 0 1 −1} and compute its value
↑
at j ω = j 0.4π. If the 10-point DFT Xk (j ωk ) of this x(n) is computed,
what is the value of the index k at which the DFT is equal to X(ej 0.4π )?
3.6 Find the DTFT of a finite sequence {1.0 0.0 −1.0} and evaluate it at
↑
ω = 0.5π. Calculate the value of the DTFT at ω = 0.5π, using the DFT
for this sequence to verify this result.
3.7 Find the DTFT of the following two functions:
(a) x1 (n) = 10(0.5)n cos(0.2πn + π3 )u(n)
(b) x2 (n) = n(0.2)n u(n)
3.8 Find the DTFT of x1 (n) = (0.5)n u(n) and x2 (n) = (0.5)n ; −5 ≤ n ≤ 5.
3.9 Find the DTFT of the following sequences:
x1 (n) = u(n) − u(n − 6)
x2 (n) = (0.5)n u(n + 3)
x3 (n) = (0.5)n+3 u(n)
x4 (n) = (0.5)−n+2 u(−n + 2)
x5 (n) = (0.3)n−2 u(−n + 2)
180 FREQUENCY-DOMAIN ANALYSIS

3.10 Find the DTFT of the following two functions:

(a) x1 (n) = x(−n − 2) where x(n) = e−0.5n u(n)
(b) x2 (n) = 5−n u(n).
3.11 Given the DTFT of x1 (n) = {1 0 −1 0}x as X1 (ej ω ), express the
DTFT of x2 (n) = {1 0 −1 0 1 0 −1 0} in terms of X1 (ej ω ).
Express the DTFT of x3 (n) = {1 0 −1 1 0 −1 0 0} in terms
of X1 (ej ω ).
3.12 An LTI-DT system is described by the difference equation

y(n) − 0.5y(n − 1) = x(n) − bx(n − 1)

Determine the value of b (other than 0.5) such that the square of the mag-
nitude of its transfer function H (ej ω ) is a constant equal to b2 for all
frequencies.
3.13 A comb filter is defined by its transfer function H (z) = (1 − z−N )/N .
Determine the frequency response of the filter in a closed-form expression
for N = 10.
3.14 Show that the magnitude response of an IIR filter with

2 + 0.2z−1 − 0.2z−3 − 2z−4

H (z) =
(1 − z−4 )

is a real function of ω and an even function of ω.

3.15 A discrete-time signal with a lowpass frequency response that is a con-
stant equal to 5 and has a bandwidth equal to 0.4π is the input to an ideal
bandpass filter with a passband between ωp1 = 0.3π and ωp2 = 0.6π and
a magnitude of 4. What is the bandwidth of the output signal?
3.16 A DT signal x(n) = 4 cos(0.4πn) + 6 cos(0.8πn) + 10 cos(0.9πn) is the
input to an allpass filter with a constant magnitude of 5 for all frequencies.
What is the output y(n) of the filter?
3.17 Given x(n) = a n u(n) and h(n) = bn u(n), where 0 < a < 1, 0 < b < 1,
a = b show that y(n) = x(n) ∗ h(n) ≡ a n+1 − bn+1 /a − b.
3.18 Find the DTFT of x(n) = n2 (0.1)n u(n).
3.19 Prove that

N −1 
N k = 0, ±N, ±2N, . . .
ej (2π/N )kn =
0 otherwise
n=0
 PROBLEMS 181

3.20 Show that the frequency response of a sequence


1 −N ≤ n ≤ N
x(n) =
0 otherwise

is given by X(ej ω ) = 1 + 2 Nn=1 cos(ωn) = sin(N + 0.5)ω/ sin(0.5ω).
π 
3.21 Show that −π X(e )dω = 2πx(0) where X(ej ω ) = ∞
jω
n=0 x(n)e
−j nω
.
3.22 Given H (ej ω ) and X(ej ω ) as shown below, find the output y(n) of the
discrete-time system

ej ω
H (ej ω ) =
(1 − 0.6e−j ω )
X(ej ω ) = 2e−j ω − 5e−j 5ω + e−j 6ω

3.23 Given H (ej ω ) and X(ej ω ) as shown below, find the output y(n)

1
H (ej ω ) =
ej ω + 0.3
ej ω
X(ej ω ) =
(1 + 0.5e−j ω )(1 − 0.5ej ω )

3.24 Find the IDTFT of the function given below:

1 − e−j 2ω
Y (ej ω ) =
(1 + 0.2ej ω )(1 − 0.4e−j ω )(ej ω + 0.5)

3.25 Find the IDTFT of the function given below:

1
Y (ej ω ) =
(ej ω + 0.1)(1 − e−j ω )(1 + ej ω )

3.26 If the input of an LTI-DT system is x(n) = (0.2)n u(−n) and its unit pulse
response h(n) is (0.4)n u(n), what is its output y(n)?
3.27 Given an input x(n) = (0.2)−n u(−n) + (0.5)n u(n) and the unit impulse
response of an LTI-DT system as (0.4)n u(n), find its output y(n).
3.28 Given a sequence x1 (n) = (0.3)−n u(−n) and another sequence x2 (n) =
(0.6)n u(−n), find their convolution sum x1 (n) ∗ x2 (n), using their DTFT.
3.29 Find the convolution y(n) = x1 (n) ∗ x2 (n) where x1 (n) = 0.5−n u(−n) and
x2 (n) = (0.2)−n u(−n).
3.30 Find the DTFT of xe (n) and x0 (n) where x(n) = (0.4)n u(n), and xe (n) =
[x(n) + x(−n)]/2 is the even part of x(n) and x0 (n) = [x(n) − x(−n)]/2
is the odd part of x(n).
182 FREQUENCY-DOMAIN ANALYSIS

3.31 Find the IDTFT of

F (ej w ) = ej ω /[(1 − 0.4e−j ω )(1 + 0.4e−j ω )(1 + 0.4ej ω )].

3.32 Find the IDTFT of X(ej ω ) = 3ej 3ω /(ej ω − 0.4).

3.33 Find the IDTFT of
(a) X(ej ω ) = cos2
 (4ω)
1 for ωc1 ≤ ω ≤ ωc2
(b) X(ej ω ) =
0 otherwise
3.34 Find the IDTFT of X(ej ω ) = 1 + 4 cos(ω) + 3 cos(5ω).
3.35 Find the IDTFT of the two functions given below:
(a) H1 (ej ω ) = 1 + 2 cos(ω) + 4 cos(2ω)
(b) H2 (ej ω ) = 1 + 2 cos(ω) + 4 cos2 (2ω)
3.36 Find the IDTFT of H1 (ej ω ) = 1 + 4 cos(ω) + 3 cos(5ω) cos(ω).
3.37 Find the IDTFT of the following two functions:
(a) Y1 (ej ω ) = j sin(ω)[4 + 4 cos(ω) + 2 cos2 (ω)]
(b) Y2 (ej ω ) = j e−j (ω/2) [4 + 2 cos(ω) + 4 cos2 (ω)] sin(ω/2)
3.38 Find the IDTFT of H2 (ej ω ) shown in Figure 3.40. Find the Fourier series
coefficients of the periodic function H2 (ej ω ).
3.39 An FIR filter is defined by H (z−1 ) = 1 + 0.5z−1 + 0.4z−2 + 0.4z−3 . Find
the magnitude of its frequency response at ω = 0.8π, using the DFT
formula.
3.40 Compute the DFT X8 (3) and X16 (6) of the sequence x(n) = {1.0 0.3 0.2
↑
0.5}.
3.41 A discrete-time sequence x(n) = {1 1 1 1 1 1} is sampled at
↑
√
2400 Hz, and the magnitude of its DTFT at 600 Hz is known to be 2.

|Hz(e jw)|

1.0

0.05
0.1p 0.2p 0.6p p
w

Figure 3.40 Problem 3.38.

 PROBLEMS 183

What is the magnitude at 6600 Hz? Compute the sample X8 (2) of its
8-point DFT.
3.42 Given the 6-point DFT of f (n), as given below compute the value of f (3):
F (0) = 10.0; F (1) = −3.5 − j 2.6; F (2) = −2.5 − j 0.866
F (3) = −2.0; F (4) = −2.5 + j 0.866; F (5) = −3.5 + j 2.6
3.43 Compute the 6-point IDFT of X(k) given below:
X(k) = {3 + j 0 −1 + j 0 −0 + j 1.732 5 + j 0 0 − j 1.732
− 1 − j 0}
3.44 If the N -point DFT of a real sequence x(n) is XN (k), prove that the
∗
DFT of x((−n))N is XN (k), using the property x((−n))N = X(N − n).
Show that the DFT of the even part xe (n) = [x(n) + x(−n)]/2 is given by
ReX(k) and the DFT of the odd part xo (n) = [x(n) − x(−n)]/2 is given
by j ImX(k).
3.45 Find the even part and odd part of the following functions:
x1 (n) = {1 −1 2 0 1 1}
x2 (n) = {1 2 1 −1 0 −2 0 1}
x3 (n) = {1 1 −1 3}
x4 (n) = {0 1 2 −1 1 0}
3.46 Determine which of the following functions have real-valued DFT and
which have imaginary-valued DFT:
x1 (n) = {1 0.5 1 0 0 1 0.5}
x2 (n) = {1 0.5 −1 1 0 1 −1 0.5}
x3 (n) = {0 0.5 −1 1 0 −1 1 − 0.5}
x4 (n) = {1 2 0 0 1 0 0 −2}
3.47 Compute the 4-point DFT and 8-point DFT of x(n) = {1 0.5 −1.5}.
Plot their magnitudes and compare their values.
3.48 Calculate the 5-point DFT of the x(n) = {1 0.5 − 1.5} above.
3.49 Calculate the 6-point DFT of x(n) = {1 1 0.5 0 −0.5}.
3.50 Given the following samples of the 8-point DFT
X(1) = 1.7071 − j 1.5858
X(3) = 0.2929 + j 4.4142
X(6) = −0 + j 2
find the values of X(2), X(5), and X(7).
3.51 Given the values of X(4), X(13), X(17), X(65), X(81), and X(90) of an
128-point DFT function, what are the values of X(124), X(63), X(115),
X(38), X(111), and X(47)?
184 FREQUENCY-DOMAIN ANALYSIS

MATLAB Problems
3.52 Compute the 16-point and 32-point DFTs of the 4-point sequence x(n) =
{1 0.5 0 −0.5}. Plot their magnitudes and compare them.
3.53 Compute the 24-point DFT of the sequence in Problem 3.52, plot the
magnitude of this DFT. Now compute 24-point IDFT of this DFT and
compare it with x(n) given above.
3.54 Plot the magnitude of the following transfer functions:

0.5 + 1.2z−1
X1 (z) =
1 + 0.2z−1 + 0.4z−2 + z−3 − z−4 + 0.06z−5
z−3 − 0.8z−5 + z−1 − 6
X2 (z) =
1 + z−1 + 0.8z−2 − 0.4z−3 − 0.3z−4 + z−5 + 0.05z−6
(1 − 0.3z)(1 + 0.2z + z2 )
X3 (z) =
(z2 + 0.2z + 1.0)(z2 − 0.1z + 0.05)(z − 0.3)
z z + 0.5 0.8
X4 (z) = − +
z + 0.4 (z + 0.1)2 z

3.55 Plot the magnitude and phase responses of the following functions:

0.2ej ω + 0.9ej 2ω
H1 (ej ω ) =
1 − 0.6ej ω + 0.6ej 2ω − 0.5ej 3ω + ej 4ω
1 + 0.4e−j ω
H2 (ej ω ) =
1 + 0.5e−j ω − 0.4e−j 2ω + e−j 3ω + 0.3e−j 4ω + 0.1e−j 5ω
H3 (ej ω ) = H1 (ej ω )H2 (ej ω )

3.56 Evaluate the magnitude response of the transfer function H (z) at ω =

0.365π and at ω = 0.635π:

0.25 + z−1
H (z) =
1− 0.8z−1+ 0.4z−2 − 0.05z−3
3.57 From the real sequence x(n) = {1 −1 2 0.5 0 −1 2 1}, show
that the DFT of [ xe (n)] = ReX(k) where the even part xe (n) = [x(n) +
x((−n))N ]/2.
3.58 From the real sequence in Problem 3.57, obtain its odd part and show that
its DFT = j ImX(k).
 REFERENCES 185

REFERENCES

1. B. A. Shenoi, Magnitude and Delay Approximation of 1-D and 2-D Digital Filters,
Springer-Verlag, 1999.
2. C. E. Shannon, Communication in the presence of noise, Proc. IRE 37, 10–12
(Jan. 1949).
3. J. G. Proakis and D. G. Manolakis, Digital Signal Processing-Principles, Algorithms,
and Applications, Prentice-Hall, 1966.
4. B. P. Lathi, Signal Processing and Linear Systems, Berkeley Cambridge Press, 1998.
5. A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing, Prentice-
Hall, 1989.
6. V. K. Ingle and J. G. Proakis, Digital Signal Processing Using MATLAB (R) V.4, PWS
Publishing, 1997.
7. S. K. Mitra, Digital Signal Processing—A Computer-Based Approach, McGraw-Hill,
2001.
8. S. K. Mitra and J. F. Kaiser, eds., Handbook for Digital Signal Processing, Wiley-
Interscience, 1993.
9. A. Antoniou, Digital Filters, Analysis, Design and Applications, McGraw-Hill, 1993.
 CHAPTER 4

Infinite Impulse Response Filters

4.1 INTRODUCTION

In Chapter 2, we discussed the analysis of discrete-time systems to obtain their

output due to a given input sequence in the time domain, using recursive algo-
rithm, convolution, and the z-transform technique. In Chapter 3, we introduced
the concept of their response in the frequency domain, by deriving the DTFT or
the frequency response of the system. These two chapters and Chapter 1 were
devoted to the analysis of DT systems. Now we discuss the synthesis of these
systems, when their transfer functions or their equivalent models are given. If we
are given the input–output sequence, it is easy to find the transfer function H (z)
as the ratio of the z transform of the output to the z transform of the input. If,
however, the frequency response of the system is specified, in the form of a plot,
such as when the passband and stopband frequencies along with the magnitude
and phase over these bands, and the tolerances allowed for these specifications,
are specified, finding the transfer function from such specifications is based on
approximation theory. There are many well-known methods for finding the trans-
fer functions that approximate the specifications given in the frequency domain.
In this chapter, we will discuss a few methods for the design of IIR filters that
approximate the magnitude response specifications for lowpass, highpass, band-
pass, and bandstop filters. Usually the specifications for a digital filter are given
in terms of normalized frequencies. Also, in many applications, the specifications
for an analog filter are realized by a digital filter in the combination of an ADC in
the front end with a DAC at the receiving end, and these specifications will be in
the analog domain. The magnitude response of ideal, classical analog filters are
shown in Figure 4.1. Several examples of IIR filter design are also included in
this chapter, to illustrate the design of these filters and also filters with arbitrary
magnitude response, by use of MATLAB functions. The design of FIR filters that
approximate the specifications in the frequency domain is discussed in the next
chapter.

Introduction to Digital Signal Processing and Filter Design, by B. A. Shenoi

Copyright © 2006 John Wiley & Sons, Inc.

186
 INTRODUCTION 187

Magnitude

Magnitude
wc wc
Frequency Frequency
(a) (b)
Magnitude

Magnitude

w1 w2 w1 w2
Frequency Frequency
(c) (d )

Figure 4.1 Magnitude responses of analog filters: (a) lowpass filter; (b) highpass filter;
(c) bandpass filter; (d) bandstop filter.

Let us select any one of the following methods to specify the IIR filters. The
recursive algorithm is given by

N
M
y(n) = − a(k)y(n − k) + b(k)x(n − k) (4.1)
k=1 k=0

and its equivalent form is a linear difference equation:

N
M
a(k)y(n − k) = b(k)x(n − k); a(0) = 1 (4.2)
k=0 k=0

The transfer function of the IIR filter is given by

M −k
k=0 b(k)z
H (z) = N ; a(0) = 1 (4.3)
−k
k=0 a(k)z
188 INFINITE IMPULSE RESPONSE FILTERS

Let us consider a few properties of the transfer function when it is evaluated on

the unit circle z = ej ω , where ω is the normalized frequency in radians:
M M
jω k=0 b(k) cos(kω) −j k=0 b(k) sin(kω)
H (e ) = N M (4.4)
k=0 a(k) cos(kω) −j k=0 a(k) sin(kω)
 
= H (ej ω ) ej θ(ω)

In this equation, H (ej ω ) is the frequency  response, or the discrete-time Fourier

transform (DTFT) of the filter, H (e j ω ) is the magnitude response, and θ (ej ω )
 j ω  j α(ω)
) = X(e ) e
jω
is the phase response. If  X(e is the frequency response of
the input signal, where X(e ) is its magnitude and α(j ω) is its phase response,
j ω

 response Y (e ) is given by Y (e ) = X(e )H (e ) =

jω jω jω jω
then jthe  frequency
X(e ω ) H (ej ω ) ej {α(ω)+θ(j ω)} . Therefore the magnitude of the output signal
 
is multiplied by the magnitude H (ej ω ) and its phase is increased by the phase
θ (ej ω ) of the filter:
⎧  2  2 ⎫1/2
⎪ M ⎪
  ⎨ k=0 b(k) cos(kω) + M
k=0 b(k) sin(kω)
⎬
H (ej ω ) =  2  2 (4.5)
⎩ N a(k) cos(kω) + M a(k) sin(kω) ⎪
⎪ ⎭
k=0 k=0
M M
−1 k=0 b(k) sin(kω) a(k) sin(kω)
θ (j ω) = − tan M + tan Nk=0
1
(4.6)
k=0 b(k) cos(kω) k=0 a(k) cos(kω)

The magnitude squared function is

     
H (ej ω )2 = H (ej ω )H (e−j ω ) = H (ej ω )H ∗ (ej ω ) (4.7)

where H ∗ (ej ω ) = H (e−j ω ) is the complex conjugate of H (ej ω ). It can be shown

that the magnitude response is an even function of ω while the phase response
is an odd function of ω.  
Very often it is convenient to compute and plot the log magnitude of H (ej ω )
 2
as 10 log H (ej ω ) measured in decibels. Also we note that H (ej ω )/H (e−j ω ) =
ej 2θ(ω) . The group delay τ (j ω) is defined as τ (j ω) = −[dθ (j ω)]/dω and is
computed from

1 du 1 dv
τ (ω) = − (4.8)
1 + u2 d ω 1 + v 2 d ω

where
M
k=0 b(k) sin(kω)
u = M (4.9)
k=0 b(k) cos(kω)
 MAGNITUDE APPROXIMATION OF ANALOG FILTERS 189

and N
a(k) sin(kω)
v = Nk=0 (4.10)
k=0 a(k) cos(kω)
Designing an IIR filter usually means that we find a transfer function H (z)
in the form of (4.3) such that its magnitude response (or the phase response, the
group delay, or both the magnitude and group delay) approximates the specified
magnitude response in terms of a certain criterion. For example, we may want
to amplify the input signal by a constant without any delay or with a constant
amount of delay. But it is easy to see that the magnitude response of a filter or
the delay is not a constant in general and that they can be approximated only by
the transfer function of the filter. In the design of digital filters (and also in the
design of analog filters), three approximation criteria are commonly used: (1) the
Butterworth approximation, (2) the minimax (equiripple or Chebyshev) approxi-
mation, and (3) the least-pth approximation or the least-squares approximation.
We will discuss them in this chapter in the same order as listed here. Designing a
digital filter also means that we obtain a circuit realization or the algorithm that
describes its performance in the time domain. This is discussed in Chapter 6. It
also means the design of the filter is implemented by different types of hardware,
and this is discussed in Chapters 7 and 8.
Two analytical methods are commonly used for the design of IIR digital fil-
ters, and they depend significantly on the approximation theory for the design
of continuous-time filters, which are also called analog filters. Therefore, it is
essential that we review the theory of magnitude approximation for analog filters
before discussing the design of IIR digital filters.

4.2 MAGNITUDE APPROXIMATION OF ANALOG FILTERS

The transfer function of an analog filter H (s) is a rational function of the complex
frequency variable s, with real coefficients and is of the form1
c0 + c1 s + c2 s 2 + · · · + c m s m
H (s) = , m≤n (4.11)
d0 + d1 s + d2 s 2 + · · · + dn s n
The frequency response or the Fourier transform of the filter is obtained as a
function of the frequency ω,2 by evaluating H (s) as a function of j ω
c0 + j c1 ω − c2 ω2 − j c3 ω4 + c4 ω4 + · · · + (j )m cm ωm
H (j ω) = (4.12)
d0 + j d1 ω − d2 ω2 − j d3 ω3 + d4 ω4 + · · · + (j )n cn ωn
= |H (j ω)| ej φ(ω) (4.13)
1
Much of the material contained in Sections 4.2–4.10 has been adapted from the author’s book
Magnitude and Delay Approximation of 1-D and 2-D Digital Filters and is included with permission
from its publisher, Springer-Verlag.
2
In Sections 4.2–4.8, discussing the theory of analog filters, we use ω and
to denote the angular
frequency in radians per second. The notation ω should not be considered as the normalized digital
frequency used in H (ej ω ).
190 INFINITE IMPULSE RESPONSE FILTERS

where H (j ω) is the frequency response, |H (j ω)| is the magnitude response, and

θ (j ω) is the phase response. We also find the magnitude squared and the phase
response from the following:

|H (j ω)|2 = H (j ω)H (−j ω) = H (j ω)H ∗ (j ω) (4.14)

H (j ω)
= ej 2θ(ω) (4.15)
H (−j ω)

The magnitude response of an analog filter is an even function of ω, whereas

the phase response is an odd function. Although these properties of H (j ω) are
similar to those of H (ej ω ), there are some differences. For example, the frequency
variable ω in H (j ω) is (are) in radians per second, whereas ω in H (ej ω ) is
the normalized frequency in radians. The magnitude response |H (j ω)| (and the
phase response) is (are) aperiodic in ω over  the doubly
 infinite interval −∞ <
ω < ∞, whereas the magnitude response H (ej ω ) (and the phase response) is
(are) periodic with a period of 2π on the normalized frequency scale.

Example 4.1

Let us take a simple example of a transfer function of an analog function as

s+1
H (s) = (4.16)
s 2 + 2s + 2

The first step is to multiply H (s) with H (−s) and evaluate the product at
s = j ω:

{H (s)H (−s)}|s=j ω = |H (j ω)|2 (4.17)

 
(s + 1)(−s + 1) 
|H (j ω)| = {H (s)H (−s)}|s=j ω =
2  (4.18)
(s + s + 2)(s − s + 2) s=j ω
2 2

ω2 + 1
= (4.19)
ω4 + 1

From this example, we see that to find the transfer function H (s) in (4.16) from
the magnitude squared function in (4.19), we reverse the steps followed above in
deriving the function (4.19) from the H (s). In other words, we substitute j ω =
s (or ω2 = −s 2 ) in the given magnitude squared function to get H (s)H (−s)
and factorize its numerator and denominator. For every pole at sk (and zero)
in H (s), there is a pole at −sk (and zero) in H (−s). So for every pole in
the left half of the s plane, there is a pole in the right half of the s plane,
and it follows that a pair of complex conjugate poles in the left half of the s
plane appear with a pair of complex conjugate poles in the right half-plane also,
thereby displaying a quadrantal symmetry. Therefore, when we have factorized
 MAGNITUDE APPROXIMATION OF ANALOG FILTERS 191

the product H (s)H (−s), we pick all its poles that lie in the left half of the
s-plane and identify them as the poles of H (s), leaving their mirror images in
the right half of the s-plane as the poles of H (−s). This assures us that the
transfer function is a stable function. Similarly, we choose the zeros in the left
half-plane as the zeros of H (s), but we are free to choose the zeros in the
right half-plane as the zeros of H (s) without affecting the magnitude. It does
change the phase response of H (s), giving a non–minimum phase response.
Consider a simple example: F1 (s) = (s + 1) and F2 (s) = (s − 1). Then F22 (s) =
(s + 1)[(s − 1)/(s + 1)] has the same magnitude as the function F2 (s) since
the magnitude of (s − 1)/(s + 1) is equal to |(j ω − 1)/(j ω + 1)| = 1 for all
frequencies. But the phase of F22 (j ω) has increased by the phase response of
the allpass function (s − 1)/(s + 1). Hence F22 (s) is a non–minimum phase
function. In general any function that has all its zeros inside the unit circle in the
z plane is defined as a minimum phase function. If it has atleast one zero outside
the unit circle, it becomes a non–minimum phase function.

4.2.1 Maximally Flat and Butterworth Approximation

Let us choose the magnitude response of an ideal lowpass filter as shown in
Figure 4.1a. This ideal lowpass filter passes all frequencies of the input continuous-
time signal in the interval |ω| ≤ ωc with equal gain and completely filters out all
the frequencies outside this interval. In the bandpass filter response shown in
Figure 4.1c, the frequencies between ω1 and ω2 and between −ω1 and −ω2 only
are transmitted and all other frequencies are completely filtered out.
In Figure 4.1, for the ideal lowpass filter, the magnitude response in the
interval 0 ≤ ω ≤ ωc is shown as a constant value normalized to one and is
zero over the interval ωc ≤ ω < ∞. Since the magnitude response is an even
function, we know the magnitude response for the interval −∞ < ω < 0. For

Ideal Magnitude

1.0

1 − dp

Transition Band

Passband
Stopband

ds

wp ws w

Figure 4.2 Magnitude response of an ideal lowpass analog filter showing the tolerances.
192 INFINITE IMPULSE RESPONSE FILTERS

the lowpass filter, the frequency interval 0 ≤ ω ≤ ωc is called the passband,

and the interval ωc ≤ ω < ∞ is called the stopband. Since a transfer function
H (s) of the form (4.11) cannot provide such an ideal magnitude characteristic, it
is common practice to prescribe tolerances within which these specifications
have to be met by |H (j ω)|. For example, the tolerance of δp on the ideal
magnitude of one in the passband and a tolerance of δs on the magnitude of
zero in the stopband are shown in Figure 4.2. A tolerance between the pass-
band and the stopband is also provided by a transition band shown in this
figure. This is typical of the magnitude response specifications for an ideal fil-
ter.
Since the magnitude squared function |H (j ω)| = H (j ω)H (−j ω) is an even
function in ω, its numerator and denominator contain only even-degree terms;
that is, it is of the form
C0 + C2 ω2 + C4 ω4 + · · · + C2m ω2m
|H (j ω)|2 = (4.20)
1 + D2 ω2 + D4 ω4 + · · · + D2n ω2n
In order that it approximates the magnitude of the ideal lowpass filter, let us
impose the following conditions

1. The magnitude at ω = 0 is normalized to one.

2. The magnitude monotonically decreases from this value to zero as ω → ∞.
3. The maximum number of its derivatives evaluated at ω = 0 are zero.

Condition 1 is satisfied when C0 = 1, and condition 2 is satisfied when the coeffi-

cients C2 = C4 = · · · = C2m = 0. Condition 3 is satisfied when the denominator
is 1 + D2n ω2n , in addition to condition 2 being satisfied. The magnitude response
that satisfies conditions 2 and 3 is known as the Butterworth response, whereas
the response that satisfies only condition 3 is known as the maximally flat mag-
nitude response, which may not be monotonically decreasing. The magnitude
squared function satisfying the three conditions is therefore of the form
1
|H (j ω)|2 = (4.21)
1 + D2n ω2n
We scale the frequency ω by ωp and define the normalized analog frequency

= ω/ωp so that the passband of this filter is

p = 1. Now the magnitude of
the lowpass filter satisfies the three conditions listed above and also the condition
that its passband be normalized to
p = 1. Such a filter is called a prototype
lowpass Butterworth filter having a transfer function H (p) = H (s/p), which
has its magnitude squared function given by
1
|H (j
)|2 = (4.22)
1 + D2n
2n
The following specifications are normally given for a lowpass Butterworth filter:
(1) a magnitude of H0 at ω = 0, (2) the bandwidth ωp , (3) the magnitude at the
 MAGNITUDE APPROXIMATION OF ANALOG FILTERS 193

bandwidth ωp , (4) a stopband frequency ωs , and (5) the magnitude of the filter
at ωs . The transfer function of the analog filter with practical specifications like
these will be denoted by H (p) in the following discussion, and the prototype
lowpass filter will be denoted by H (s).
Before we proceed with the analytical design procedure, we normalize the
magnitude of the filter by H0 for convenience and scale the frequencies ωp and
ωs by ωp so that the bandwidth of the prototype filter and its stopband fre-
quency become
p = 1 and
s = ωs /ωp , respectively. The specifications about
the magnitude at
p and
s are satisfied by the proper choice of D2n and n
in the function (4.22) as explained below.
√ If, for example, the magnitude at the
passband frequency is required to be 1/ 2, which means that the log magnitude
required is −3 dB, then we choose D2n = 1. If the magnitude at the passband
frequency
=
p = 1 is required to be 1 − δp , then we choose D2n , normally
denoted by  2 , such that

1 1
|H (j 1)|2 = = = (1 − δp )2 (4.23)
1 + D2n 1 + 2

If the magnitude at the bandwidth

=
p = 1 is given as −Ap decibels, the
value of  2 is computed by

1
10 log = −Ap
1 + 2
10 log(1 +  2 ) = Ap
log(1 +  2 ) = 0.1Ap
(1 +  2 ) = 100.1Ap
"
From the last equation, we get the formula  2 = 100.1Ap − 1 and  = 100.1Ap − 1.
Let us consider the common case of a Butterworth filter with a log magnitude
of −3 dB at the bandwidth of
p to develop the design procedure for a Butter-
worth lowpass filter. In this case, we use the function for the prototype filter, in
the form
1
|H (j
)|2 = (4.24)
1 +
2n

This satisfies the following properties:

1. The magnitude squared of the filter response at

= 0 is one.
2. The magnitude squared at
= 1 is 12 for all integer values of n; so the log
magnitude is −3 dB.
3. The magnitude decreases monotonically to zero as
→ ∞; the asymptotic
rate is −40n dB/decade.
194 INFINITE IMPULSE RESPONSE FILTERS

1.2

0.8
Magnitude

0.6

0.4

0.2
n=2
n=6
0
0 0.5 1 1.5 2 2.5 3 3.5 4
Frequency in rad/sec

Figure 4.3 Magnitude responses of Butterworth lowpass filters.

The magnitude response of Butterworth lowpass filters is shown for n =

2, 3, . . . , 6 in Figure 4.3. Instead of showing the log magnitude of these filters,
we show their attenuation in decibels in Figure 4.4. Attenuation or loss measured
in decibels is defined as

−10 log |H (j
)|2 = 10 log(1 +
2n )

The attenuation over the passband only is shown in Figure 4.4a, and the maximum
attenuation in the passband is 3 dB for all n; the attenuation characteristic of the
filters over 1 ≤
≤ 10 for n = 1, 2, . . . , 10 is shown in Figure 4.4b.

4.2.2 Design Theory of Butterworth Lowpass Filters

Let us consider the design of a Butterworth lowpass filter for which (1) the
frequency ωp at which the magnitude is 3 dB below the maximum value at ω = 0,
and (2) the magnitude at another frequency ωs in the stopband are specified.
When we normalize the gain constant to unity and normalize the frequency by
the scale factor ωp , we get the cutoff frequency of the normalized prototype
filter
p = 1 and the stopband frequency
s = ωs /ωp . After we have found the
transfer function H (p) of this normalized prototype lowpass filter, we restore
the frequency scale and the magnitude scale to get the transfer function H (s)
approximating the prescribed magnitude specification of the lowpass filter.
The analytical procedure used to derive H (p) from the magnitude squared
function of the prototype lowpass filter is carried out simply by reversing the
 MAGNITUDE APPROXIMATION OF ANALOG FILTERS 195

7.0

6.0
Passband attenuation a, dB

5.0

4.0

3.0

2.0
n=1
2
6 3
1.0 5
4 7
8
9
10
0
0 0.2 0.4 0.6 0.8 1.0
ω
(a)

140

10 9
120 8
7

6
Stopband attenuation a, dB

100
5
80
4

60 3

40 2

20 n=1

0
0 2.0 4.0 6.0 8.0 10
ω
(b)

Figure 4.4 Attenuation characteristics of Butterworth lowpass filters in (a) passband;

(b) stopband.
196 INFINITE IMPULSE RESPONSE FILTERS

steps used to derive the magnitude squared function from H (p) as illustrated by
Example 4.1 earlier. First we substitute
= p/j or equivalently
2 = −p 2 in
(4.24):

1  1
 = = H (p)H (−p) (4.25)
2n 
1 +

2 =−p2 1 + (−1)n p 2n

The denominator has 2n zeros obtained by solving the equation

1 + (−)n p 2n = 0 (4.26)

or the equation

1 = ej 2kπ n odd
p2n = (4.27)
−1 = ej (2k+1)π n even

This gives us the 2n poles of H (p)H (−p), which are

pk = ej (2kπ/2n)π k = 1, 2, . . . , (2n) when n is odd (4.28)

and

pk = ej [(2k−1)/2n]π k = 1, 2, . . . , (2n) when n is even (4.29)

or in general

pk = ej [(2k+n−1)/2n]π k = 1, 2, . . . , (2n) (4.30)

We notice that in both cases, the poles have a magnitude of one and the angle
between any two adjacent poles as we go around the unit circle is equal to π/n.
There are n poles in the left half of the p plane and n poles in the right half of
the p plane, as illustrated for the cases of n = 2 and n = 3 in Figure 4.5. For
every pole of H (p) at p = pa that lies in the left half-plane, there is a pole of
H (−p) at p = −pa that lies in the right half-plane. Because of this property,
we identify n poles that are in the left half of the p plane as the poles of H (p)
so that it is a stable transfer function; the poles that are in the right half-plane
are assigned as the poles of H (−p). The n poles that are in the left half of the
p plane are given by
   
2k + n − 1
pk = exp j π k = 1, 2, 3, . . . , n (4.31)
2n
When we have found these n poles, we construct the denominator polynomial
D(p) of the prototype filter H (p) = D(p)
1
from

(
n
D(p) = (p − pk ) (4.32)
k=1
 MAGNITUDE APPROXIMATION OF ANALOG FILTERS 197

q1
π
q2 3
π
2 q3

n=2 n=3

Figure 4.5 Pole locations of Butterworth lowpass filters of orders n = 2 and n = 3.

The only unknown parameter at this stage of design is the order n of the filter
function H (p), which is required in (4.31). This is calculated using the specifi-
cation that at the stopband frequency
s , the log magnitude is required to be no
more than −As dB or the minimum attenuation in the stopband to be As dB.

10 log |H (j
s )|2 = −10 log(1 +
2n
s ) ≤ −As (4.33)

from which we derive the formula for calculating n as follows:

log(100.1As − 1)
n≥ (4.34)
2 log
s

Since we require that n be an integer, we choose the actual value of n = n

that is the next-higher integer value or the ceiling of n obtained from the right
side of (4.34). When we choose n = n, the attenuation in the stopband is more
than the specified value of As . We use this integer value for n in (4.31), to cal-
culate the poles and then construct the denominator polynomial D(p) of order
n. By multiplying (p − pk ) with (p − pk∗ ) where pk and pk∗ are complex con-
jugate pairs, the polynomial is reduced to the normal form with real coefficients
only. These polynomials, known as Butterworth polynomials, have many special
properties. In the polynomial form, if we represent them as

D(p) = 1 + d1 p + d2 p 2 + · · · + dn p n (4.35)

their coefficients can be computed recursively from (d0 = 1)

 
cos (k − 1) π2
dk =   dk−1 k = 1, 2, 3, . . . , n (4.36)
sin kπ
2n

But there is no need to do so, since they can be computed from (4.32). They are
also listed in many books for n up to 10 in polynomial form and in some books
in a factored form also [3,2]. We list a few of them in Table 4.1.
198 INFINITE IMPULSE RESPONSE FILTERS

TABLE 4.1
n Butterworth Polynomial D(p) in Polynomial and Factored Form

1 p+1
√
2 p 2 + 2p + 1
3 p 3 + 2p 2 + 2p + 1 = (p + 1)(p 2 + p + 1)
4 p 4 + 2.61326p 3 + 3.41421p 2 + 2.61326p + 1
= (p 2 + 0.76537p + 1)(p 2 + 1.84776p + 1)
5 p + 3.23607p 4 + 5.23607p 3 + 5.23607p 2 + 3.23607p + 1
5

= (p + 1)(p 2 + 0.618034p + 1)(p 2 + 1.931804p + 1)

6 p + 3.8637p 5 + 7.4641p 4 + 9.1416p 3 + 7.4641p 2 + 3.8637p + 1
6

= (p 2 + 0.5176p + 1)(p 2 + 1.4142p + 1)(p 2 + 1.9318p + 1)

In the case of lowpass filters, usually the magnitude is specified at ω = 0;

hence it is also the magnitude at
= 0. Therefore the specified magnitude is
equated to the value of the transfer function H (p) evaluated at p = j 0. This is
equal to H (j 0) = H0 /D(j 0) = H0 . So we restore the magnitude scale by mul-
tiplying the normalized prototype filter function by H0 . To restore the frequency
scale by ωp , we put p = s/ωp in H0 /D(p) and simplify the expression to get
transfer function H (s) for the specified lowpass filter. This completes the design
procedure, which will be illustrated in Example 4.2.

Example 4.2

Design a lowpass Butterworth filter with a maximum gain of 5 dB and a cutoff

frequency of 1000 rad/s at which the gain is at least 2 dB and a stopband fre-
quency of 5000 rad/s at which the magnitude is required to be less than −25 dB.
The maximum gain of 5 dB is the magnitude of the filter function at ω = 0.
The edge of the passband is the cutoff frequency ωp = 1000, and the frequency
range 0 ≤ ω ≤ ωp is called the bandwidth. So we see that the magnitude of
2 dB at this frequency is 3 dB below the maximum value in the passband. We
say that the filter has a 3 dB bandwidth equal to 1000 rad/s. The frequency scale
factor is chosen as 1000 so that the passband of the prototype filter is
p = 1.
The stopband frequency ωs is specified as 5000 rad/s and is therefore scaled to

s = 5. The magnitude is normalized so that the normalized prototype lowpass

filter function H (p)3 has a magnitude of one (i.e., 0 dB) at
= 0. It is this filter
that has a magnitude squared function
1
|H (j
)|2 = (4.37)
1 +
2n

3
Note that we have chosen p = +j
as the notation for the complex frequency variable of the
transfer function H (p) for the lowpass prototype filter and the notation s = σ + j ω for the variable
of the transfer function H (s) for the specified filter.
 MAGNITUDE APPROXIMATION OF ANALOG FILTERS 199

0 dB 0 dB
−0.5
−3
Magnitude in dB

Magnitude in dB
−30 −30

w w

(a) (b)

Figure 4.6 Magnitude response specifications of prototype filters: (a) Butterworth filter;
(b) Chebyshev (equiripple) filter.

It is always necessary to reduce the given specifications to the specifications of

this normalized prototype filter to which only the expressions derived above are
applicable. The magnitude response of the normalized prototype filter (not to
scale) for this example is shown in Figure 4.6a.
For this example, note that the maximum attenuation in the passband is Ap =
3 dB and the minimum attenuation in the stopband is As = 30 dB. From (4.34)
we calculate the value of n = 2.1457 and choose
√ n = 2.1457 = 3. From (4.31),
we√ get the three poles as p 1 = −0.5 + j 0.75, p2 = −1.0 and p3 = −0.5 −
j 0.75. Therefore the third-order denominator polynomial D(p) is obtained
from (4.32) or from Table 4.1:
√ "
D(p) = (p + 0.5 − j 0.75)(p + 1)(p + 0.5 + 0.75)
= (p 2 + p + 1)(p + 1) = p3 + 2p 2 + 2p + 1 (4.38)

Hence the transfer function of the normalized prototype filter of third order is

1
H (p) = (4.39)
p 3 + 2p 2 + 2p + 1

To restore the magnitude scale, we multiply this function by H0 . Now the filter
function is
H0
H (p) = (4.40)
p3 + 2p 2 + 2p + 1

which has a magnitude of H0 at p = j 0. From the requirement 20 log(H0 ) =

5 dB, we calculate the value of H0 = 1.7783. To restore the frequency scale, we
200 INFINITE IMPULSE RESPONSE FILTERS

substitute p = s/1000 in (4.40) and simplify to get H (s) as shown below:

1.7783
H (p)|p=s/1000 =   
s 3 s 2  s 
+2 +2 +1
1000 1000 1000
(1.7783)109
= 3 (4.41)
s + (2 × 103 )s 2 + (2 × 106 )s + 109
= H (s) (4.42)

The magnitude of H (p) plotted on the normalized frequency scale

shown in
Figure 4.7 is marked as “Example (2).” It is found that the attenuation at the
stopband edge
s = 5 is about 42 dB, which is more than the specified 30 dB.
It must be remembered that in (4.37)
p = 1 is the bandwidth of the prototype
filter, and at this frequency, |H (j
)|2 has a value of 12 or a magnitude of −3 dB.
Hence formulas (4.31) and (4.34) cannot be used if the maximum attenuation Ap
in the passband is different from 3 dB. In this case, we modify the function to
the form (4.43), which is the general case:
1
|H (j
)|2 = (4.43)
1 +  2
2n
Now the attenuation at
= 1 is given by 10 log(1 +  2 ) = Ap , from which we
get  2 = (100.1Ap − 1). We may also note that  2 = 1 in the previous case when
Ap = 3. When Ap is other than 3 dB, the formulas for calculating n and pk are
 
log (100.1As − 1)/(100.1Ap − 1)
n≥ (4.44)
2 log
s

−1 Example (3)
Magnitude in dB

−2

−3
Example (4)

−4
Example (2)

−5
10°
Frequency in radians/sec-linear scale

Figure 4.7 Magnitude responses of the prototype filters in Examples 4.1–4.3.

 MAGNITUDE APPROXIMATION OF ANALOG FILTERS 201

and
   
2k + n − 1
pk =  −(1/n) exp j π k = 1, 2, 3, . . . , n (4.45)
2n

Comparing (4.45) with (4.31), it is obvious that the poles have been scaled by a
factor  −(1/n) . So the maximum attenuation at
p = 1 is the specified value of
Ap ; also the frequency at which the attenuation is 3 dB is equal to  −(1/n) .

Example 4.3

Design a lowpass Butterworth filter with a maximum magnitude of 5 dB, pass-

band of 1000 rad/s, maximum attenuation in the passband Ap = 0.5 dB, and
minimum attenuation As = 30 dB at the stopband frequency of 5000 rad/s.
First we scale the frequency by ωp = 1000 so that the normalized passband
frequency
p = 1 and the stopband frequency ωs is mapped to
s = 5. Also
the magnitude is scaled by 5 dB. The magnitude response for the normalized
prototype filter H (p) is similar to that shown in Figure 4.6a, except that now
Ap = 0.5 dB. Then we calculate  2 = (100.1Ap − 1) = 0.1220 and therefore  =
0.3493. From (4.44), the value of n = 2.7993; it is rounded to n = 3. Next we
compute the three poles from (4.45) as p1 = −0.71 + j 1.2297, p2 = −1.4199,
and p3 = −0.71 − j 1.2297. The transfer function of the filter with these poles is

H0
H (p) =
(p + 1.4199)(p + 0.71 − j 1.2297)(p + 0.71 + j 1.2297)
H0
= (4.46)
(p + 1.4199)(p 2 + 1.42p + 2.0163)

Since the maximum value has been normalized to 0 dB, which occurs at
= 0,
we equate the magnitude of H (p) evaluated at p = j 0 to one. Therefore H0 =
(1.4199)(2.0163) = 2.8629.
√ To raise the magnitude level to 5 dB, we have to
multiply this constant by 100.5 = 1.7783. Of course, we can compute the same
value for H0 in one step, from the specification 20 log |H (j 0)| = 20 log H (0) −
20 log(1.4199)(2.0163) = 5. The frequency scale is restored by putting p =
s/1000 in (4.46) to get (4.47) as the transfer function of the filter that meets
the given specifications:

(2.8629)(1.7783)
H (s) =
[s/1000 + 1.4199][(s/1000)2 + 1.42(s/1000) + 2.0163]
5.09 × 109
= (4.47)
[s + 1419.9][s 2 + 1420s + 2.0163 × 106 ]

The plot is marked as “Example (3)” in Figure 4.7. It is the magnitude response
of the prototype filter given by (4.46). It has a magnitude of −0.5 dB at
= 1
and approximately −33 dB at
= 5, which exceeds the specified value.
202 INFINITE IMPULSE RESPONSE FILTERS

4.2.3 Chebyshev I Approximation

The Chebyshev I approximation for an ideal lowpass filter shows a magnitude
that has the same values for the maxima and for the minima in the passband and
decreases monotonically as the frequency increases above the cutoff frequency.
It has equal-valued ripples in the passband between the maximum and minimum
values as shown in Figure 4.6b. Hence it is known as the minimax approximation
and also as the equiripple approximation. To approximate the ideal magnitude
response of the lowpass filter in the equiripple sense, the magnitude squared
function of its prototype is chosen to be

H02
|H (j
)|2 = (4.48)
1 +  2 Cn2 (
)

where Cn (
) is the Chebyshev polynomial of degree n. It is defined by

Cn (
) = cos(n cos−1
) |
| ≤ 1 (4.49)

The polynomial Cn (
) approximates a value of zero over the closed interval

∈ [−1, 1] in the equiripple sense as shown by examples for n = 2, 3, 4, 5

in Figure 4.8a. These polynomials are

C0 (
) = 1
C1 (
) =

C2 (
) = 2
2 − 1
C3 (
) = 4
3 − 3

C4 (
) = 8
4 − 8
2 + 1
C5 (
) = 16
5 − 20
3 + 5
(4.50)

4.2.4 Properties of Chebyshev Polynomials

Some of the properties of Chebyshev polynomials that are useful for our discus-
sion are described below. Let cos φ =
. Then Cn (n cos−1
) = cos(nφ), and
therefore we use the identity

cos(k + 1) = cos(kφ) cos(φ) − sin(kφ) sin(φ)

= 2 cos(kφ) cos(φ) − cos((k − 1)φ) (4.51)

from which we obtain a recursive formula to generate Chebyshev polynomials

of any order, as

C0 (
) = 1
Ck+1 (
) = 2
Ck (
) − Ck−1 (
) (4.52)
 MAGNITUDE APPROXIMATION OF ANALOG FILTERS 203

C2 (Ω) C3 (Ω)
1 1

1 1

C4 (Ω) C5 (Ω)
1 1

1 1

(a)

10log(l + e2C42 (Ω))

Ap

Ωp = 1
(b)

Figure 4.8 Chebyshev polynomials and Chebyshev filter: (a) magnitude of Chebyshev
polynomials; (b) attenuation of a Chebyshev I filter.

To see that Cn (
) = cos(n cos−1
) is indeed a polynomial of order n, consider
it in the following form:
 
cos(nφ) = Re ej nφ
 n  " n
= Re cos(φ) + j sin(φ = Re φ + j (1 − φ 2
 " n
= Re φ + φ 2 − 1 (4.53)
 " n
Expanding φ + φ 2 − 1 by the binomial theorem and choosing the real part,
we get the polynomial for
n(n − 1) n−2 2
cos(nφ) = φ n + φ (φ − 1)
2!
n(n − 1)(n − 2)(n − 3) n−4 2
+ φ (φ − 1)2 + · · · (4.54)
4!
204 INFINITE IMPULSE RESPONSE FILTERS

Recall that since n is a positive integer, the expansion expressed above has a
finite number of terms, and hence we conclude that it is a polynomial (of degree
n). We also note from (4.50) that

0 n odd
Cn2 (0) = (4.55)
1 n even

But

1 n odd
Cn2 (1) = (4.56)
1 n even

So we derive the following properties:


1 n odd
|H (0)| =
2
1 (4.57)
1+ 2
n even
1
|H (1)|2 = n odd or even (4.58)
1 + 2

The attenuation characteristics of the Chebyshev filter of order n = 4 is shown in

Figure 4.8b as an example. The magnitude |H (j
)| plotted as “Example(4)” in
Figure 4.7 has an equiripple response in the passband, with a maximum value of
0 dB and a minimum value of 10 log[1/(1 +  2 )] decibels. However, the mag-
nitude of Chebyshev I lowpass filters is 10 log[1/(1 +  2 )] at
= 1 for any
order n. The magnitude of the ripple can be measured as either |H (0)| − |H (1)|
or |H (0)|2 − |H (1)|2 = 1 − [1/(1 +  2 )] = [ 2 /(1 +  2 )] ≈  2 . We can always
calculate  2 = (100.1Ap − 1).
Another property of Chebyshev I filters is that the total number of maxima and
minima in the closed interval [−1 1] is n + 1. The square of the magnitude
response of Chebyshev lowpass filters is shown in Figure 4.9a to indicate some
properties of the Chebyshev lowpass filters just described.

4.2.5 Design Theory of Chebyshev I Lowpass Filters

Typically the specifications for a lowpass Chebyshev filter specify the maximum
and minimum values of the magnitude in the passband; the cutoff frequency ωp ,
which is the highest frequency of the passband; a frequency ωs in the stopband;
and the magnitude at the frequency ωs . As in the case of the Butterworth filter, we
normalize the magnitude and the frequency and reduce the given specifications
to those of the normalized prototype lowpass filter and follow similar steps to
find the poles of H (p).
Since
can take real values greater than one in general, let us assume φ to be
a complex variable: φ = ϕ1 + j ϕ2 . From 1 +  2 Cn2 (
) = 0, we get  2 Cn2 (
) =
 MAGNITUDE APPROXIMATION OF ANALOG FILTERS 205

2 2
⏐H(j Ω)⏐ ⏐H( j Ω)⏐

1 1

n = 3 (odd) n = 4 (even)

Ωp Ωs Ωp Ωs
(a)

2 2
⏐H(j Ω)⏐ ⏐H( j Ω)⏐

1 1

n = 5 (odd)
n = 4 (even)

Ω p Ωs Ω p Ωs
(b)

Figure 4.9 Magnitude response of Chebyshev filters: (a) Chebyshev I filters;

(b) Chebyshev II filters.

−1 = j 2 ; we derive
j
Cn (
) = ±

= cos(nφ) = cos(n(ϕ1 + j ϕ2 ))
= cos(nϕ1 ) cosh(nϕ2 ) − j sin(nϕ1 ) sin(nϕ2 ) (4.59)

Equating the real and imaginary parts, we get

cos(nϕ1 ) cosh(nϕ2 ) = 0 (4.60)

and
1
sin(nϕ1 ) sin(nϕ2 ) = ∓ (4.61)

From (4.60) we get
(2k − 1)π
ϕ1 = (4.62)
2n
206 INFINITE IMPULSE RESPONSE FILTERS

Substituting this in (4.61), we obtain sinh(nϕ2 ) = ±(1/), from which we get

 
1 1
ϕ2 = sinh−1 (4.63)
n 

Now
= cos(φ) = cos(ϕ1 + j ϕ2 ) = cos(ϕ1 ) cosh(ϕ2 ) − j sin(ϕ1 ) sinh(ϕ2 ).
Therefore

j
= sin(ϕ1 ) sinh(ϕ2 ) + j cos(ϕ1 ) cosh(ϕ2 ) (4.64)

These are the roots in the p plane that satisfy the condition 1 +  2 Cn2 (
) = 0.
Hence the 2n poles of H (p)H (−p) are given by
   
(2k − 1)π (2k − 1)π
pk = sinh(ϕ2 ) sin + j cosh(ϕ2 ) cos
2n 2n

for

k = 1, 2, . . . , (2n) (4.65)

The 2n poles of H (p)H (−p) given by (4.65) can be shown to lie on an elliptic
contour in the p plane with a major semiaxis equal to cosh(ϕ2 ) alongthe j

axis and a minor semiaxis equal to sinh(ϕ2 ) along  axis, where p = +j

.
We find that the frequency
3 at which the attenuation of the prototype filter is
3 dB is given by
  
1 1

3 = cosh cosh−1 (4.66)
n 

The poles in the left half of the p plane only are given by
   
(2k − 1)π (2k − 1)π
pk = − sinh(ϕ2 ) sin + j cosh(ϕ2 ) cos
2n 2n
= − sinh(ϕ2 ) sin(θk ) + j cosh(ϕ2 ) cos(θk ) k = 1, 2, 3, . . . , n (4.67)

where ϕ2 is obtained from (4.63). In (4.67), note that θk are the angles measured
from the imaginary axis of the p plane and the poles lie in the left half of the
p plane.
The formula for finding the order n is derived from the requirement that
10 log[1 +  2 Cn2 (
s )] ≥ As . It is
) 
cosh−1 (100.1As − 1)/(100.1Ap − 1)
n≥ (4.68)
cosh−1
s
and the value of n is chosen for calculating the poles using (4.67). Given ωp ,
Ap , ωs , and As as the specifications for a Chebyshev lowpass filter H (s), its
 MAGNITUDE APPROXIMATION OF ANALOG FILTERS 207

maximum value in the passband is normalized to one, and its frequencies are
scaled by ωp , to get the values of
p = 1 and
s = ωs /ωp for the prototype
filter at which the attenuations are Ap and As , respectively. The design procedure
to find H (s) starts with the magnitude squared function (4.48) and proceeds as
follows:
"
1. Calculate  = (100.1Ap − 1).
2. Calculate n from (4.68) and choose n = n.
3. Calculate ϕ2 from (4.63).
4. Calculate the poles pk (k = 1, 2, . . . , n) from (4.67).
 
5. Compute H (p) = H0 /[ nk=1 (p − pk )] = H0 /[ nk=0 dk p k ].
6. Find the value of H0 by equating
⎧
H0 ⎨* 1 n odd
H (0) = = 1
d0 ⎩ n even
1 + 2

7. Restore the magnitude scale.

8. Restore the frequency scale by substituting p = s/ωp in H (p) and simplify
to get H (s).

A simple example is worked out below to illustrate this design procedure.

Example 4.4

Let us choose the specifications of a lowpass Chebyshev filter with a maxi-

mum gain of 5 dB, a bandwidth of 2500 rad/s, and a stopband frequency of
12,500 rad/s; Ap = 0.5 dB, and As = 30 dB. For the prototype filter, the maxi-
mum value in the passband is one (0 dB), and we have
p = 1,
s = 5. So
"
1.  = (100.05 − 1 = 0.34931.
) 
2. n ≥ {cosh−1 (103 − 1)/(100.05 − 1) }/[cosh−1 (5)] = 2.2676; choose
n = 3.
 1
3. ϕ2 = 13 sinh−1 0.34931 = 0.591378.
4. pk = −0.313228 ± j 1.02192 and −0.626456.
5. H (p) = H0 /[(p + 0.31228 − j 1.02192)(p + 0.31228 + j 1.02192)(p +
0.626456)] = H0 /[(p 2 + 0.626456p + 1.142447)(p + 0.626456)].
6. H (0) = H0 /[(1.142447)(0.626456)] = 1 (since n = 3 is odd). Hence H0 =
0.715693.
7. The transfer function with a direct-current (DC) gain of 0 dB is H (p) =
0.715693/[(p2 + 0.626456p + 1.142447)(p + 0.626456)]. The magnitude
scale is restored by multiplying H (p) by 1.7783, so that the DC gain is
raised to 5 dB.
208 INFINITE IMPULSE RESPONSE FILTERS

8. The transfer function of the filter is

(0.715693)(1.7783)
H (p) = (4.69)
(p 2 + 0.626456p + 1.142447)(p + 0.626456)

When we substitute p = s/2500 in this H (p) and simplify the expression, we get

19.886 × 1012
H (s) = (4.70)
(s 2 + 1566s + 714 × 106 )(s + 1566)

The magnitude response of the prototype filter in (4.70) is marked as “Example(4)”

in Figure 4.7. The three magnitude responses are plotted in the same figure so
that the response of the three filters can be compared. The attenuation of the
Chebyshev filter at
s = 5 is found to be 47 dB. The abovementioned class of
filters with equiripple passband response and monotonic response in the stopband
are sometimes called Chebyshev I filters, to distinguish them from the following
class of filters, known as Chebyshev II filters.

4.2.6 Chebyshev II Approximation

The Chebyshev II filters have a magnitude response that is maximally flat at ω =
0; it decreases monotonically as the frequency increases and has an equiripple
response in the stopband. Typical magnitudes of Chebyshev II filters are shown
in Figure 4.9b. This class of filters are also called Inverse Chebyshev filters. The
transfer function of Chebyshev II filters are derived by applying the following
two transformations: (1) a frequency transformation
= 1/ω in |H (j
)|2 of
the lowpass normalized prototype filter gives the magnitude squared function of
the highpass filter |H (1/j
)|2 , with an equiripple passband in |
| > 1 and a
monotonically decreasing response in the stopband 0 < |
| < 1; (2) when it is
subtracted from one, we get the magnitude squared function (4.72) of the inverse
Chebyshev lowpass filter:
 
 1 2 1
H 
 j
 = 1 +  2 C 2 (1/
) (4.71)
n

1  2 Cn2 (1/
) 1
1− = =  (4.72)
1+  2 Cn2 (1/
) 1 +  2 Cn2 (1/
) 1 +  2 C 21(1/
)
n

The magnitude squared function |H (j

)|2 of a lowpass Chebyshev I filter
and |H ( j1
)|2 and 1 − |H ( j1
)|2 are shown in Figure 4.10.
We make two important observations in Figure 4.10. The normalized cutoff
frequency
= 1 becomes the lowest frequency in the stopband of the inverse
Chebyshev filter at which the magnitude is  2 /(1 +  2 ). Hence the frequencies ωp
and ωs specified for the inverse Chebyshev filter must be scaled by ωs and not by
ωp to obtain the prototype of the inverse Chebyshev filter. We also observe that
 MAGNITUDE APPROXIMATION OF ANALOG FILTERS 209

|H( j Ω)|2
1

| H( 1 ) |2
jΩ

1 |2
1− | H( )
jΩ

Figure 4.10 Transformation of Chebyshev I–Chebyshev II filter response.

when n is odd, the number of finite zeros in the stopband is (n − 1)/2 = m. When
n is an odd integer, the term sec θk , which is involved in the design procedure
described below, attains a value of ∞ when k = (n + 1)/2. So one of the zeros
is shifted to j ∞; the remaining finite zeros appear in conjugate pairs on the
imaginary axis, and hence the numerator of the Chebyshev II filter is expressed
as shown in step 6 in Section 4.2.7. Note that the value of i calculated in step 1
is different from the value calculated in the design of Chebyshev I filters and
therefore the values of ϕi used in steps 3 and 4 are different from ϕ2 used in
the design of Chebyshev I filters. Hence it would be misleading to state that the
poles of the Chebyshev II filters are obtained as “the reciprocals of the poles of
the Chebyshev I filters.”
210 INFINITE IMPULSE RESPONSE FILTERS

4.2.7 Design of Chebyshev II Lowpass Filters

Given ωp , Ap , ωs , As and the maximum value in the passband, we scale the
frequencies ωp and ωs by ωs and deduce the specifications for the normalized
prototype lowpass inverse Chebyshev filter. Equation (4.72) is the magnitude
squared function of this inverse Chebyshev filter, and we follow the design pro-
cedure as outlined below:
"
1. Calculate i = 1/ (100.1As − 1).
2. Calculate
) 
cosh−1 (100.1As − 1)/(100.1Ap − 1)
n≥
cosh−1
s
and choose n = n.
3. Calculate ϕi from ϕi = (1/n) sinh−1 (1/i ).
4. Compute the poles in the left-half plane pk :
1
pk = k = 1, 2, 3, . . . , n
− sinh(ϕi ) sin(θk ) + j cosh(ϕi ) cos(θk )
5. The zeros of the transfer function H (p) are calculated as zk = ±j
0k =
j sec θk for k = 1, 2, . . . , m = n/2 and the numerator N (p) of H (p) as
 m
k=1 (p +
ok )
2

6. Compute

H0 m (p +
20k )
H (p) = nk=1
k=1 (p − pk )
 
and calculate H0 = nk=1 (pk )/ m 2
k=1 (
0k ) .
7. Restore the magnitude scale.
8. Restore the frequency scale by putting p = s/ωs in H (p) to get H (s) for
the inverse Chebyshev filter.

Example 4.5

Design the lowpass inverse Chebyshev filter with a maximum gain of 0 dB

in the passband, ωp = 1000, Ap = 0.5 dB, ωs = 2000, and As = 40 dB. We
normalize the frequencies by ωs and get the lowest frequency of the stopband
at
= 1, while ωp = 1000 maps to
p = 0.5. We will have to denormalize the
frequency by substituting p = s/2000 when the transfer function H (p) of the
inverse Chebyshev filter, obtained by the steps given above, is completed. The
design procedure gives
√
1. i = ( 104 − 1)−1 = 1
99.995 .
2. n = 5.
 MAGNITUDE APPROXIMATION OF ANALOG FILTERS 211

3. ϕi = 15 sinh−1 (99.995) = 1.05965847.

4. Poles in the left half-plane are pk = (−0.155955926 ± j 0.6108703175),
(−0.524799485 ± j 0.485389011), and (−0.7877702666).
5. Zeros are z1 = ±j 1.0515 and z2 = ±j 1.7013.
6. The transfer function of the inverse Chebyshev filter H (p) is given by

H0 (p 2 + 1.05152 )(p 2 + 1.70132 )

(4.73)
(p 2 + 0.3118311852p + 0.3974722176)
(p 2 + 1.04959897p + 0.5110169847)(p + 0.787702666)

7. Calculate H0 = 0.049995.
8. Hence we simplify H (p) to the final form:

0.049995(p4 + 4.04p 2 + 3.2002)

(4.74)
p 5 + 2.1491328p 4 + 2.30818905p 3 + 1.54997p 2
+ 0.65725515p + 0.15999426

The magnitude response of (4.73) is plotted in Figure 4.11. It is seen that the
prototype filter meets the desired specifications. Now we only have to denormal-
ize the frequency by 2000, so that the passband of the specified filter changes
from 0.5 to 1000 rad/s, and it meets the specifications given in Example 4.5.

20

−20
Magnitude in dB

−40

−60

−80

−100
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Frequency in radians/sec-linear scale

Figure 4.11 Magnitude response of Chebyshev II lowpass filter.

212 INFINITE IMPULSE RESPONSE FILTERS

4.2.8 Elliptic Function Approximation

There is another type of filter known as the elliptic function filter or the Cauer
filter. They exhibit an equiripple response in the passband and also in the stop-
band. The order of the elliptic filter that is required to achieve the given spec-
ifications is lower than the order of the Chebyshev filter, and the order of the
Chebyshev filter is lower than that of the Butterworth filter. Therefore the elliptic
filters form an important class, but the theory and design procedure are complex
and beyond the scope of this book. However, in Section 4.11 we will describe
the use of MATLAB functions to design these filters.

4.3 ANALOG FREQUENCY TRANSFORMATIONS

Once we have learned the methods of approximating the magnitude response

of the ideal lowpass prototype filter, the design of filters that approximate the
ideal magnitude response of highpass, bandpass, and bandstop filters is easily
carried out. This is done by using well-known analog frequency transformations
p = g(s) that map the magnitude response of the lowpass filter H (j
) to that of
the specified highpass, bandpass, or bandstop filters H (j ω). The parameters of
the transformation are determined by the cutoff frequency (frequencies) and the
stopband frequency (frequencies) specified for the highpass, bandpass, or band-
stop filter so that frequencies in their passband(s) are mapped to the passband of
the normalized, prototype filter, and the frequencies in the stopband(s) of the high-
pass, bandpass, or bandstop filters are mapped to the stopband frequency of the
prototype filter. After the normalized prototype lowpass filter H (p) is designed
according to the methods discussed in the preceding sections,the frequency trans-
formation p = g(s) is applied to H (p) to calculate the transfer function H (s) of
the specified filter. With this general outline, let us consider the design of each
filter in some more detail.

4.3.1 Highpass Filter

It is easy to describe the design of a highpass filter by choosing an example.
Suppose that a highpass filter with an equiripple passband ωp ≤ |ω| < ∞ is
specified, along with a stopband frequency ωs . The magnitude at ωp , which is
the cutoff frequency of the passband, and also the magnitude at ωs (or Ap and As )
are given. The lowpass–highpass (LP–HP) frequency transformation p = g(s)
to be used in designing the highpass (HP) filters is
ωp
p= (4.75)
s
It is seen that when s = j ωp , the value of p = −j and when s = −j ωp , the
value of p = j . It can also be shown that under this transformation, all frequen-
cies in the passband of the highpass filter map into the passband frequencies
 ANALOG FREQUENCY TRANSFORMATIONS 213

−1 ≤
≤ 1 of the lowpass prototype filter. We calculate the frequency
s to
which the specified stopband frequency ωs maps, by putting s = j ωs in (4.75).
The stopband frequency is found to be
s = ωp /ωs . So the specified magnitude
response of the highpass filter is transformed into that of the lowpass prototype
equiripple filter. We design the prototype lowpass filter to meet these specifica-
tions and then substitute p = ωp /s in H (p) to get the transfer function H (s) of
the specified highpass filter.

Example 4.6

The cutoff frequency of a Chebyshev highpass filter is ωp = 2500, which is the

lowest frequency in the passband, and the maximum attenuation in the passband
Ap = 0.5 dB. The maximum gain in the passband is 5 dB. At the stopband
frequency ωs = 500, the minimum attenuation required is 30 dB. Design the
highpass filter H (s).
When we apply the LP–HP transformation p = 2500/s, the cutoff frequency
ωp = 2500 maps to
p = 1 and the stopband frequency ωs maps to
s = 5.
In the lowpass prototype filter, we have
p = 1,
s = 5, Ap = 0.5 dB, As =
30 dB, and the maximum value of 5 dB in the passband. This filter has been
designed in Example 4.3 and has a transfer function given by (4.70), which is
repeated below:

(0.715693)(1.7783)
H (p) =
(p 2 + 0.626456p + 1.142447)(p + 0.626456)

Next we substitute p = 2500/s in this transfer function, and when simplified,

the transfer function of the specified highpass Chebyshev filter becomes

(0.715693)(1.7783) 
H (s) = 
(p + 0.626456p + 1.142447)(p + 0.626456) p=2500/s
2

1.7783s 3
= (4.76)
[s 2 + 1370.9s + 5.4707 × 106 ][s + 3990]

The magnitude response of (4.76) is plotted in Figure 4.12 and is found to exceed
the specifications of the given highpass filter. The design of a highpass filter with
a maximally flat passband response or with an equiripple response in both the
passband and the stopband is carried out in a similar manner.

4.3.2 Bandpass Filter

The normal specifications of a bandpass filter H (s) as shown in Figure 4.13
are the cutoff frequencies ω1 and ω2 , the maximum value of the magnitude in
the passband between the cutoff frequencies, the maximum attenuation in this
passband or the minimum magnitude at the cutoff frequencies ω1 and ω2 , and a
frequency ωs (= ω3 or ω4 ) in the stopband at which the minimum attenuation or
214 INFINITE IMPULSE RESPONSE FILTERS

10

−10
Magnitude in dB

−20

−30

−40

−50

−60 −1
10 100 101
Frequency in rad/sec-log scale

Figure 4.12 Magnitude response of a highpass filter.

1.0
1 − d1

d2

w3 w1 w2 w4

Figure 4.13 Typical specifications of a bandpass filter.

the maximum magnitude are specified. The type of passband response required
may be a Butterworth or Chebyshev response.
The lowpass–bandpass (LP–BP) frequency transformation p = g(s) that is
used for the design of a specified bandpass filter is
+ ,
1 s 2 + ω02
p= (4.77)
B s
 ANALOG FREQUENCY TRANSFORMATIONS 215

√
where B = ω2 − ω1 is the bandwidth of the filter and ω0 = ω1 ω2 is the geo-
metric mean frequency of the bandpass filter.
A frequency s = j ωk in the bandpass filter is mapped to a frequency p = j
k
under this transformation, which is obtained by
+ ,
j ω02 − ωk2
j
k = (4.78)
B ωk
 
j ω0 ωk ω0
= − (4.79)
B ω0 ωk

Therefore the frequencies ω1 and ω2 map to

= ∓1, and the frequencies −ω1
and −ω2 map to
= ±1. Similarly, the positive value of the stopband frequency

s to which the frequency ωs maps is calculated from

 + ,
 1 ω2 − ω2 
 s 

s =  0
 (4.80)
B ωs 

The magnitude or the attenuation at the frequencies

= 1 and
s for the
prototype filter are the same as those at the corresponding frequencies of the
bandpass filter. From the specification of the lowpass prototype filter, we obtain
its transfer function H (p), following the appropriate design procedure discussed
earlier. Then we substitute (4.77) in H (p) to get the transfer function H (s) of
the bandpass filter specified.

Example 4.7

The specifications of a Chebyshev I bandpass filter are ω1 = 104 , ω2 = 105 ,

ωs = 2 × 105 , Ap = 0.8 dB, and As = 30 dB, and the maximum magnitude in
the passband = 10 dB. We use the following procedure to design the filter:
√ √
1. B = ω2 − ω1 = 9 × 104 and ω0 = ω2 ω1 = 109 = 31.62 × 103 .
2. The LP–BP transformation is p = 1/9 × 104 [(s 2 + 109 )/s].
3. Let s = j ωs = j 2 × 105 . From the preceding transformation, we get
s =
2.1667.
4. The lowpass Chebyshev prototype filter has a magnitude response as
shown in Fig. 4.6b.
" √
5. Calculate  = 100.1Ap − 1 = 100.08 − 1 = 0.4497.
6. Calculate n from (4.68). Choose n = 3.5 = 4.
7. Calculate ϕ2 from (4.63). We get ϕ2 = 0.3848.
8. Calculate the poles from (4.67): pk = −0.15093 ± j.9931 and −0.36438 ±
j.41137.
216 INFINITE IMPULSE RESPONSE FILTERS

Magnitude response of the bandpass filter

10

−10

−20
Magnitude in dB

−30

−40

−50

−60

−70

−80

−90
103 104 105 106
Frequency in rad/sec-log scale

Figure 4.14 Magnitude response of the bandpass filter in Example 4.6.

9. The transfer function

of the lowpass prototype Chebyshev filter is derived
from H (p) = H0 /[ 4k=1 (p − pk )] where H0 is fixed to match the gain
of 10 dB at
= 0:
0.8788
H (p) = (4.81)
(p 2 + 0.3018p + 1.009)(p 2 + 0.7287p + 0.302)

10. Now substitute p = (s 2 + 109 )/(9 × 104 s) in H (p) and simplify to

get H (s)

5.7658 × 1019 s 4
H (s) =
D(s)

where
 4 
(s + 2.7162 × 104 s 3 + 101.729 × 108 s 2 + 2.7162 × 1013 s + 1018 )
D(s) =
×(s 4 + 6.5583 × 104 s 3 + 44.462 × 108 s 2 + 6.5583 × 1013 s + 1018 )
(4.82)

To verify the design, we have plotted the magnitude response of the bandpass
filter in Figure 4.14.

4.3.3 Bandstop Filter

The normal specification of a bandstop (bandreject) filter is shown in Figure 4.15.
The passband of this filter is given by 0 ≤ ω ≤ ω1 and ω2 ≤ ω ≤ ∞, whereas
 ANALOG FREQUENCY TRANSFORMATIONS 217

1.0
1 − d1

d1
w1 w3 w4 w2

Figure 4.15 Typical magnitude specifications for a bandstop filter.

the frequencies ω3 and ω4 define frequencies in the stopband. Usually these

frequencies and the corresponding magnitudes are given as the specifications.
Note that in Figure 4.13, showing the magnitude response of a bandpass filter,
the stopband is equiripple and the passband between ω1 and ω2 has a monotonic
response; this it is a Chebyshev II type of filter. But in the bandstop filter response
shown in Figure 4.15, the passband is equiripple and the stopband is maximally
flat, so it is a Chebyshev I type of bandstop filter. It is important to remember
that the parameter B in the LP-BS transformation is chosen as ω2 − ω1 , and not
√
the bandwidth ω4 − ω3 of the stopband! The mean frequency ω0 = ω2 ω1 , and
the LP–BS frequency transformation p = g(s) is given by
+ ,
s
p=B 2 (4.83)
s + ω02

This transformation transforms the entire passband of the bandstop filter to the
passband |
| ≤ 1 of the prototype lowpass filter. So we have to find the frequency

s to which the stopband frequency ωs is transformed under the transformation.

It is found from
 
 Bω 
 s 

s =  2  (4.84)
 ωs − ω02 

Thus we have reduced the specification of a bandstop filter to that of a proto-

type lowpass filter. It is designed by the procedures discussed earlier. When the
transfer function H (p) of the prototype filter is completed, the transformation
p = B[s/(s 2 + ω02 )] is used to transform H (p) into H (s). The design of the
bandstop filter is illustrated by the following example.
218 INFINITE IMPULSE RESPONSE FILTERS

Example 4.8
Suppose that we are given the specification of a bandstop filter as shown in
Figure 4.15. In this example, we are given ω1 = 1500, ω2 = 2000, ωs = ω4 =
1800, Ap = 0.2 dB, and As = 55 dB. The passband is required to have a max-
imally flat response. With these specifications, we design the bandstop filter
following procedure given below:
√
1. B = 2000 − 1500 = 500 and ω0 = (2000)(1500) = 1732.1.
2. The LP–BS frequency transformation is p = 500[s/(s 2 + 3 × 106 )].
3. Let s = j ωs = j 1800. Then we get
s = 3.74.
4. Following
√ the design procedure used in Example 4.2, we get  =
100.02 − 1 = 0.21709, and from (4.44), we get n = 5.946 and choose
n = 6.
5. The six poles are calculated from (4.45) as pk = −0.33385 ± j 1.2459,
−0.9121 ± j 0.9121, and −1.246 ± j 0.3329.
 of the lowpass prototype filter H (p) is constructed
6. The transfer function
from H (p) = H0 /[ 4k=1 (p − pk )] as
(1.664)3
(p 2 + 0.6677p + 1.664)(p 2 + 1.824p + 1.664)(p 2 + 2.492p + 1.664)
(4.85)
7. Next we have to substitute p = 500[s/(s 2 + 3 × 106 )] in this H (p) and
simplify the expression to get the transfer function H (s) of the speci-
fied bandstop filter. This completes the design of the bandstop filter. The
magnitude response is found to exceed the given specifications.
The sections above briefly summarize the theory of approximating the piece-
wise constant magnitude of analog filters. This theory will be required for approx-
imating the magnitude of digital filters, which will be treated in the following
sections. The analog frequency transformations p = g(s) applied to the lowpass
prototype to generate the other types of filters are listed in Table 4.2.

TABLE 4.2 Frequency Transformations to Design HP, BP, and BS Filters

Type of Transformation
Transformation p = g(s) Parameters Used

LP–LP p = s/ωp ωp = bandwidth-specified LP

filter
LP–HP p = ωp /s ωp = cutoff frequency of the
specified HP filter
LP–BP p = (1/B)[(s 2 + ω02 )/s] B = ω2 − ω1 , where B is
bandwidth of the specified
√
BP filter: ω0 = ω1 ω2 ;
LP–BS p = B[s/(s 2 + ω02 )] B = ω2 − ω 1
 IMPULSE-INVARIANT TRANSFORMATION 219

4.4 DIGITAL FILTERS

In contrast to analog filters, digital filters are described by two types of transfer
functions: transfer functions of finite impulse response filters and those of infinite
impulse response filters. The methods for designing FIR filters will be treated
in the next chapter. Now that we have reviewed the methods for approximating
the magnitude of analog filters, it is necessary to understand the relationship
between the frequency-domain description of analog and digital filters, in order
to understand the frequency transformation that is used to transform the analog
frequency response specifications to those of the digital filters.

4.5 IMPULSE-INVARIANT TRANSFORMATION

The procedures used for designing IIR filters employ different transformations
of the form s = f (z) to transform H (s) into H (z). The transformation s = f (z)
must satisfy the requirement that the digital filter transfer function H (z) be stable,
when it is obtained from the analog filter transfer functions H (s). The transfer
functions for the analog filters obtained by the methods described above are stable
functions; that is, their poles are in the left half of the s plane. When H (s) and
f (z) are stable in the s and z domains, respectively, the poles of H (s) in the left
half of the s plane map to the poles inside the unit circle in the z plane; therefore
H (z) also is a stable transfer function. We also would like to have frequencies
from −∞ to ∞ on the j ω axis of the s plane mapped into frequencies on the
boundary of the unit circle—without encountering any discontinuities.
We have already introduced the transformation z = esT , in Chapter 2, when
we derived the z transform of a discrete-time signal x(nT ) generated from the
analog signal x(t).
We plot the magnitude response of the analog filter as a function of ω. Under
the impulse-invariant transformation, s = j ω maps to z = ej ωT . Although the
magnitude of the digital filter H (ej ωT ) is a function of the variable ej ωT , we
cannot plot it as a function of ej ωT . We can plot the magnitude response of
the digital filter only as a function of ωT . (Again, we point out that the nor-
malized digital frequency ωT is commonly denoted in the DSP literature by
the symbol ω.) When s = j ω increases values from −j ∞ along the imaginary
axis to +j ∞, the variable ej ωT increases counterclockwise from e−j π to ej π
(passing through z = 1) along the boundary of the unit circle in the z plane and
repeats itself since ej ωT = ej (ωT +2rπ) , where r is an integer. The strips in the
left half of the s plane bounded by ±j [(2r − 1)π/T ] and ±j [(2r + 1)π/T ]
on the j ω axis are mapped to the inside and the boundary of the unit cir-
cle in the zplane as shown in Figure 4.16. Therefore the frequency response
X∗ (j ω) = ∞ n=0 x(nT )e
−j ωT is periodic and will avoid aliasing only if the ana-

log signal x(t) is bandlimited.

Consider the transfer function H (s) of an analog filter. Since the poles of an
analog filter function such as the filters discussed in this chapter are simple, its
220 INFINITE IMPULSE RESPONSE FILTERS

1n 3p
T
p
T

0
p
−
T
3p
−
T

z-plane s-plane

Figure 4.16 Mapping of s plane to z plane under the transformation z = esT .

partial fraction expansion is of the form

K
Rk
H (s) = (4.86)
s + sk
k=1

The unit impulse response hk (t) of a typical term Rk /(s + sk ) is Rk e−sk t . When
it is sampled with a sampling period T , and the z transform is evaluated, it
becomes
∞
1 z
Rk e−sk nT z−n = Rk = Rk (4.87)
1− e−sk T z−1 z − e−sk T
n=0

Hence H (z) derived from H (s) under the transformation z = esT is given by

K
Rk z
H (z) = (4.88)
z − e−sk T
k=1

Because the unit pulse response h(nT ) of the digital filter matches the unit
impulse response h(t) at the instants of sampling t = nT , the transformation z =
esT is called the impulse-invariant transformation. But the frequency response of
H (z) will not match the frequency response of H (s) unless h(t) is bandlimited.
If the magnitude response of the analog filter H (j ω) is very small for frequencies
larger than some frequency ωb , and h(t) is sampled at a frequency greater than
2ωb , the frequency response of the digital filter H (z) obtained from the impulse-
invariant transformation may give rise to a small amount of aliasing that may
or may not be acceptable in practical design applications. However, this method
 BILINEAR TRANSFORMATION 221

is not applicable for the design of highpass, bandstop, and allpass filters since
their frequency responses are not bandlimited at all. If the impulse-invariant
transformation is applied to a minimum phase analog filter H (s), the resulting
digital filter may or may not be a minimum phase filter. For these reasons, the
impulse-invariant transformation is not used very often in practical applications.

4.6 BILINEAR TRANSFORMATION

The bilinear transformation is the one that is the most often used for designing
IIR filters. It is defined as
 
2 z−1
s= (4.89)
T z+1
To find how frequencies on the unit circle in the z plane map to those in the s
plane, let us substitute z = ej ωT in (4.89). Note that ω is the angular frequency
in radians per second and ωT is the normalized frequency in the z plane. Instead
of using ω as the notation for the normalized frequency of the digital filter, we
may denote θ as the normalized frequency to avoid any confusion in this section:
     
2 ej ωT − 1 2 ej (ωT /2) − e−j (ωT /2) 2 ωT
s= = = j tan
T ej ωT + 1 T ej (ωT /2) + e−j (ωT /2) T 2
 
ωT
= j 2fs tan = jλ
2
This transformation maps the poles inside the unit circle in the z plane to the
inside of the left half of the s plane and vice versa. It also maps the frequencies
on the unit circle in the z plane to frequencies on the entire imaginary axis of
the s plane, where s = σ + j λ. So this transformation satisfies both conditions
that we required for the mapping s = f (z) mentioned in the previous section or
its inverse relationship z = b(s). This mapping is shown in Figure 4.17 and may
be compared with the mapping shown in Figure 4.16.
To understand the mapping in some more detail, let us consider the frequency
response of an IIR filter over the interval (0, (ωs /2)), where ωs /2 = π/T is the
 ωT 
Nyquist frequency. As an example, we choose a frequency response H (e ) =
H (e ) of a Butterworth bandpass digital filter as shown in Figure 4.18a.
j θ

In Figure 4.18, we have also shown the curve depicting the relationship between
ωT and λ = 2fs tan (ωT /2). The value of λ corresponding to any value of ωT = θ
can be calculated from λ = 2fs tan (θ/2) as illustrated by mapping a few frequen-
cies such as ω1 T , ω2 T in Figure 4.18. The magnitude of the frequency response
of the digital filter at any normalized frequency ωk T is the magnitude of H (s) at
the corresponding frequency s = j λk , where λk = 2fs tan (ωk T /2).
The plot in Figure 4.17 shows that the magnitude response of the digital filter
over the Nyquist interval (0, π) maps over the entire range (0, ∞) of λ. So there
is a nonlinear mapping whereby the frequencies in the ω domain are warped
222 INFINITE IMPULSE RESPONSE FILTERS

jl

Image of s = j l

s
1

Image of left
half s-plane

s-plane z-plane

Figure 4.17 Mapping of s plane to z plane under bilinear transformation.

l = 2ƒs tan wT
l

2
Ω

l3l4
l1l2

|H( jΩ)| |H( jl)| wT

(c) (b)
|H(jlwT)|

w1T w2T w3T w4T

wT
(a)

Figure 4.18 Mapping of the digital filter response under bilinear transformation and
analog BP ⇒ LP transformation.

when mapped to the λ domain. Similarly, the frequencies in the interval (0, −π)
are mapped to the entire interval (0, −∞) of λ. From the periodic nature of
the function tan(.), we also see that the periodic replicates of the digital filter
frequency response in the ω domain map to the same frequency response in the
λ domain and the transfer function H (s) obtained under the bilinear transform
behaves like that of an analog filter. But it is to be pointed out that we use only
 BILINEAR TRANSFORMATION 223

the mathematical theory of analog filter approximation to solve the problem of

finding such a function H (s), and we are not designing an analog filter. In other
words, the bilinear transformation helps us reduce the mathematical problem of
approximating the frequency response of a digital filter in the variable ω, to
the problem of approximating another function in the variable λ. Because of its
similarity to the analog frequency, the approximation problem is solved by using
the mathematical theory of approximation for the frequency response of analog
filters. So, if the frequency response |H (j λ| is a lowpass frequency response, the
frequency s is linearly scaled by λc to obtain the frequency response |H (j
)| of
a lowpass prototype filter. If it is a highpass, bandpass, or bandstop response, then
the appropriate analog frequency transformation p = g(s) listed in Table 4.2 is
used to convert the specification |H (j λ| to that of an analog prototype lowpass
filter. We obtain the transfer function H (p) of the prototype filter, in which
the complex frequency variable p is shown in bold in order to differentiate it
from H (p), and its magnitude is denoted by |H (j
)|. The theory of analog filter
approximation is used to find H (p) such that its magnitude |H (j
)| approximates
the magnitude response of the lowpass prototype filter. It is important to note
that the unit impulse response of the filter H (p) when sampled with a sampling
period T does not match the unit pulse response of the digital filter H (z), because
the bilinear transformation is not impulse- invariant.
Once we have designed the lowpass prototype filter function H (p), we apply
the appropriate analog frequency transformation p = g(s) to H (p) to get the
function H (s). Then we substitute s = 2fs [(z − 1)/(z + 1)] in H (s) to get H (z)
as the transfer function of the digital filter.

Example 4.9

The specified magnitude response of a maximally flat bandpass digital filter has a
maximum value of 1.0 in its passband, which lies between the cutoff frequencies
θ1 = 0.4π and θ2 = 0.5π. The magnitude at these cutoff frequencies is specified
to be no less than 0.93, and at the frequency θ3 = 0.7π in the stopband, the
magnitude is specified to be no more than 0.004. Design the IIR digital filter that
approximates these specifications, using the bilinear transformation.
It is obvious from these specifications that the frequencies are normalized
frequencies. So θ1 = 0.4π and θ2 = 0.5π are the normalized cutoff frequencies
and θ3 = 0.7π is the frequency in the stopband. The specified magnitude response
is plotted in Figure 4.19a. The two cutoff frequencies ω1 , ω2 and the stopband
frequency ω3 map to and λ1 , λ2 , λ3 as follows. In this example, we have chosen
to scale the frequencies in the s plane by fs ; thus, the values for λ1 , λ2 , and λ3
given below are obtained by the bilinear transform s = 2[(z − 1)/(z + 1)]:

λ1 = 2 tan(0.2π) = 1.453 rad/s

λ2 = 2 tan(0.25π) = 2.00 rad/s
λ3 = 2 tan(0.35π) = 3.95 rad/s

The frequency response of the “analog” filter H (s) is plotted in Figure 4.19b.
224 INFINITE IMPULSE RESPONSE FILTERS

q
λ = 2fs tan
2

λ3
λ2
Ω2
Ω1

λ1
|H(jl)|
1.00
0.93

.004
|H(j Ω)|
1.00
0.93

.004

|H(jθ)|
(c) (b)

1.00
0.93

0.004

q1 q2 q3
(a)

Figure 4.19 Mapping of filter responses under bilinear transformation.

√ the bandwidth (BW) = λ2 − λ1 = 0.547 and the center fre-

Now we find
quency λ0 = λ2 λ1 = 1.705 for the prewarped bandpass filter function |H (j λ)|.
Next we define the lowpass–bandpass frequency transformation
+ ,  2 
1 s 2 + λ20 1 s + 1.7052
p= = .
(B) s 0.547 s

To find the frequency

3 , to which the frequency λ3 = 3.95 maps, we sub-
stitute s = j 3.95 in the preceding transformation and get p = j 5.876 = j
2 ,
whereas the cutoff frequencies map to the normalized frequency
p = 1. Hence
the magnitude response of the lowpass Butterworth prototype filter function is as
shown in Figure 4.19c. Using the same notations as before, we get Ap = 0.63 dB,
As = 48 dB,  = 0.395, and n = 4 for this prototype lowpass Butterworth filter.
The four poles in the left half of the p plane are calculated as the poles of the
lowpass prototype Butterworth filters:

p1, p4 = −0.4827 ± j 1.1654

p2 , p3 = −1.1654 ± j 0.4827
 BILINEAR TRANSFORMATION 225

10

−10

−20
Magnitude in dB

−30

−40

−50

−60

−70

−80
10−1 100 101
Frequency in rad./sec

Figure 4.20 Magnitude response of the analog prototype lowpass filter in Example 4.8.

The transfer function of the prototype lowpass filter is therefore given by

2.5317
H (p) = (4.90)
p4 + 3.296p3 + 5.4325p2 + 5.24475p + 2.5317

The magnitude response of this lowpass filter is plotted in Figure 4.20.

Next we substitute p = (1/0.547)[(s 2 + 1.7052 )/s], in (4.90) and after sim-
plifying, the resulting transfer function is

0.2267s 4
H (s) =
D(s)

where D(s) is given by

 8 
(s + 1.8030s 7 + 13.2535s 6 + 16.5824s 5 + 60.3813s 4
(4.91)
+ 48.205s 3 + 112.0006s 2 + 44.2926s + 71.4135)

Now we apply the bilinear transformation s = 2[(z − 1)/(z + 1)] we chose in

this example, on this H (s), and simplify the transfer function H (z) of the digital
filter to
3.6272z8 − 14.5088z6 + 21.7632z4 − 14.5088z2 + 3.6272
H (z) =  
(3825z8 − 4221z7 + 13127z6 − 9857z5 + 15753z4 − 7615z3
+ 7849z2 − 1934z + 1354)
(4.92)
226 INFINITE IMPULSE RESPONSE FILTERS

−10

−20
Magnitude in dB

−30

−40

−50

−60

−70

−80
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7
Normalized frequency

Figure 4.21 Magnitude response of a digital bandpass filter.

A plot of the magnitude response of this function is shown in Figure 4.21. It is

found that the given specifications are met by this transfer function of the digital
filter.
The design of lowpass, highpass, and bandstop filters use similar procedures.
In contrast to the impulse-invariant transformation, we see that the bilinear trans-
formation can be used for designing highpass and bandstop filters as well. Indeed,
the use of bilinear transformation is the most popular method used for the design
of IIR digital filter functions that approximate the magnitude-only specifications.

4.7 DIGITAL SPECTRAL TRANSFORMATION

In the design procedure described above, we used the bilinear transformation to

convert the magnitude specification of an IIR digital filter to that of H (j λ) by
prewarping the frequencies on the λ axis. Then we either scaled the frequencies
on the λ axis or applied the analog frequency transformations p = g(s) to reduce
the frequency response to that of a lowpass, analog prototype filter function. There
is an alternative method for designing IIR digital filters. It replaces the analog
frequency transformation by a frequency transformation in the digital domain.
Constantinides [1] derived a set of digital spectral transformations (DSTs) that
convert the magnitude of a lowpass digital filter with an arbitrary bandwidth,
say, θp , to that of digital highpass, bandpass, and bandstop filters or digital
lowpass filters with a different passband. These transformations are similar to
the analog frequency transformations, and the parameters of the transformation
are determined by the cutoff frequencies of these filters just as in the case of the
analog frequency transformations. Let us denote the cutoff frequency of the new
 DIGITAL SPECTRAL TRANSFORMATION 227

digital lowpass filter or the highpass filter by θp

; and let us denote the upper
and lower cutoff frequencies of the bandpass and bandstop filters by θu and θl ,
respectively—all of them are less than π radians on the normalized frequency
basis. Whereas the lowpass, prototype analog filter always has a passband of
1 rad/s, the lowpass digital filter has a passband that is chosen arbitrarily as θp ;
yet, we will call it the lowpass digital, “prototype” filter, with a transfer function
H (z−1 ). The digital spectral transformations applied on this digital filter are of
the form z−1 = g(z−1 ). They map points inside the unit circle in the z plane to
points inside the unit circle in the z plane, and map the boundary of the unit
circle in the z plane to the boundary of the unit circle in the z plane. Using these
necessary conditions, Constantinides derived the DSTs for the LP–LP,LP–HP,
LP–BP, and LP–BS transformations and they are listed in Table 4.3, where
θ = ωT are normalized frequencies in radians.

Example 4.10

We choose the same specifications as in Example 4.9 and illustrate the procedure
to design the IIR filter using the digital spectral transformation from Table 4.3.
Let us choose the passband of the lowpass prototype digital filter to be θp =
0.5π. The values for the cutoff frequencies specified for the bandpass filter are
θl = 0.4π, θu = 0.5π. So we calculate
 
0.5π + 0.4π
cos
2
α=   = 0.158
0.5π − 0.4π
cos
2
   
0.5π − 0.4π 0.5π
K = cot tan = 6.314
2 2
z−2 − 0.273z−1 + 0.727
z−1 = −
0.727z−2 − 0.273z−1 + 1
Now we have to find the frequency θs in the lowpass digital “prototype” filter
to which the prescribed stopband frequency θs
= 0.7π of the bandpass filter maps,
by substituting z = ej 0.7π in the digital spectral transformation given above. The
value is found to be θs = 2.8 rad = 0.8913π rad. Therefore the specification
for the lowpass prototype digital filter to be designed is given as shown in
Figure 4.22b.
Using the mapping of λ = 2 tan( θ2 ) versus θ , we map this lowpass frequency
response to the lowpass filter response |H (j λ)| as shown in Figure 4.22(c). We
calculate λp = 2 tan(π/4) = 1.998 and λs = 2 tan 2.8 2 = 11.6 as the edge of the
passband and the edge of the stopband of this filter, respectively. So we scale its
frequency by 1.998 to get the frequency response of the lowpass prototype filter
H (j
) in order to get a normalized bandwidth
p = 1. The stopband frequency

s is scaled down to 5.8, which is slightly different from the value obtained in
228 INFINITE IMPULSE RESPONSE FILTERS

TABLE 4.3 Digital Spectral Transformations

Type of
Transformation Transformation Parameters Used

z−1 − a
LP–LP z−1 = θp = passband of prototype
1 − az−1
filter
θp
= passband of new LP filter
 
θp − θp

sin
2
a=  
θp + θp

sin
  2
z−1 + a
LP–HP z−1 = − θp
= cutoff frequency of the
1 + az−1
HP filter
 
θp + θp

cos
2
a=  
θp θp

cos
⎛ ⎞ 2
z −2 − 2αK z−1 + (K − 1)
⎜ (K + 1) (K + 1) ⎟
LP–BP z−1 = −⎜
⎝ (K − 1) −2
⎟
⎠ θl = lower cutoff frequency of
2αK −1
z − z +1 BP filter
(K + 1) (K + 1) θu = upper cutoff frequency of
BP filter
 
θu + θl
cos
2
α=  
θu − θl
cos
2
   
θu − θl θp
K = cot tan
2 2
2α 1−K
z−2 − z−1 +
−1 (K + 1) 1+K
LP–BS z = θl = lower cutoff frequency of
1 − K −2 2α
z − z−1 + 1 BS filter
1+K (K + 1)
θu = upper cutoff frequency of
BS filter
 
θu + θl
cos
2
α=  
θu − θl
cos
2
   
θu − θl θp
K = tan tan
2 2
 DIGITAL SPECTRAL TRANSFORMATION 229

ls
Ωs

lp
Ωp
1.00
0.93

0.004

1.00
0.93

0.004
(d) (c)
1.00
0.93

0.004

qp qs
(b)

1.00
0.93

0.004

ql qu q′s
(a)

Figure 4.22 Mapping of magnitude responses under digital spectral transformation.

Example 4.8 because of numerical inaccuracies. But the order of the lowpass
prototype analog filter is required to be the same, and hence the transfer function
is the same as in Example 4.8. The transfer function is repeated below: Note, we
use H (p) to denote the lowpass filter in this example
2.5317
H (p) = (4.93)
p4 + 3.2962p3 + 5.4325p2 + 5.2447p + 2.5317
Next we restore the frequency scale by substituting p = s/1.998 in H (p) to get
the transfer function H (s) as
40.5072
H (s) = (4.94)
s 4 + 6.5924s 3 + 21.73s 2 + 41.9576s + 40.5072
230 INFINITE IMPULSE RESPONSE FILTERS

and then apply the bilinear transformation s = 2[(z − 1)/(z + 1)] on this H (s)
to get the transfer function of the lowpass prototype digital filter H (z) as

(40.5072z4 + 162.0288z3 + 243.0432z2 + 162.0288z + 40.5072)

H (z) =
(280.0816z4 + 160.3808z3 + 165.2032z2 + 35.6768z + 6.7728)
(4.95)
The final step is to apply the digital spectral transformation (4.96) derived earlier,
to H (z) in (4.95):

z−2 − 0.273z−1 + 0.727

z−1 = − (4.96)
0.727z−2 − 0.273z−1 + 1
The final result is the transfer function H (z) of the required IIR filter, which
is found to be the same as (4.92) obtained in Example 4.8. The magnitude is
therefore found to be the same as in Figure 4.21 and is not plotted again. When
compared with the method of Example 4.8, the method described above does not
offer any advantages; indeed, it requires more computations, particularly in the
final step. However, if we already have the transfer function of a lowpass IIR
filter, we can use the digital spectral transformation to obtain the transfer function
of a HP, BP, or BS filter if their magnitude responses are of the same type (i.e.,
Butterworth, Chebyshev, or elliptic type) and the tolerances in the passband as
well as the stopband are the same as those for the lowpass filter.

4.8 ALLPASS FILTERS

Allpass filters have a magnitude response that is exactly equal to a constant at

all frequencies, and hence there is no need for any approximation.
Such filters are of the form
a(N ) + a(N − 1)z−1 + · · · + a(2)z−(N −2) + a(1)z−(N −1) + z−N
Hap (z−1 ) = ±
1 + a(1)z−1 + a(2)z−2 + · · · + a(N − 1)z−(N −1) + a(N )z−N
(4.97)
This is the ratio of two polynomials expressed in descending powers of z. Note
that the coefficients of the numerator polynomial are in an order reverse that
of the coefficients of the denominator polynomial. We can express (4.97) in an
alternate form:
⎡ ⎤
a(N )zN + a(N − 1)zN −1 + · · · + a(2)z2
⎢ + a(1)z + 1 ⎥
Hap (z−1 ) = ±z−N ⎢ ⎥
⎣ 1 + a(1)z−1 + a(2)z−2 + · · · + a(N − 1)z−(N −1) ⎦ (4.98)
+ a(N )z−N
⎡ ⎤
1 + a(1)z + a(2)z2 + · · · + a(N − 1)z(N −1)
⎢ + a(N )zN ⎥
= ±z−N ⎢
⎣
⎥ (4.99)
⎦
1 + a(1)z−1 + a(2)z−2 + · · · + a(N − 1)z−(N −1)
+ a(N )z−N
 IIR FILTER DESIGN USING MATLAB 231

If the denominator of (4.99) is denoted as D(z−1 ), its numerator is z−N D(z),

which is the mirror image polynomial of D(z−1 ). Therefore, the allpass filter has
a transfer function expressed in compact form as

z−N D(z)
Hap (z−1 ) = (4.100)
D(z−1 )

When the allpass filter has all its poles inside the unit circle in the z plane, it is
a stable function and its zeros are outside the unit circle as a result of the mirror
image symmetry. Therefore a stable, allpass filter function is non–minimum
function.
From (4.99), it is easy to see that the magnitude response of Hap (ej ω ) is equal
to one at all frequencies and is independent of all the coefficients:
 
   + · · · + a(N )ej Nω 
Hap (ej ω ) =  1 + a(1)e + a(2)e
jω j 2ω
 1 + a(1)e−j ω + a(2)e−j 2ω + · · · + a(N )e−j Nω  = 1 (4.101)

But the phase response (and the group delay) is dependent on the coefficients of
the allpass filter. We know that the phase response—as defined by (4.6)—of an
IIR filter designed to approximate a specified magnitude response is a nonlinear
function of ω and therefore its group delay defined by (4.8) is far from a constant
value. When an allpass filter is cascaded with such a filter, the resulting filter
has a jfrequency response H1 (ej ω )Hap (ej ω ) = H1 (ej ω )Hap (ej ω ) ej [θ(ω)+φ(ω)] =
 j [θ(ω)+φ(ω)]
H1 (e ) e
ω
. So the magnitude response does not change when the IIR
filter is cascaded with an allpass filter, but its phase response θ (ω) changes by the
addition of the phase response φ(ω) contributed by the allpass filter. The allpass
filters Hap (z) are therefore very useful for modifying the phase response (and
the group delay) of filters without changing the magnitude of a given IIR filter
H1 (z), when they are cascaded with H1 (z). However, the method used to find the
coefficients of the allpass filter Hap (z) such that the group delay of H1 (z)Ha (z)
is a very close approximation to a constant in the passband of the filter H1 (z)
poses a highly nonlinear problem, and only computer-aided optimization has
been utilized to solve this problem. When the allpass filters have been designed
to compensate for the group delay of the IIR filters that have been designed to
approximate a specified magnitude only, such that the cascade connection of the
two filters has a group delay that approximates a constant value, the allpass filters
are known as delay equalizers.

4.9 IIR FILTER DESIGN USING MATLAB

The design of IIR digital filters with Butterworth, Chebyshev I, Chebyshev II,
and elliptic filter responses, using MATLAB functions, are based on the theories
of bilinear transformation and analog filters. So they are commonly used to
approximate the piecewise constant magnitude characteristic of ideal LP, HP, BP,
and BS filters. The MATLAB function yulewalk is used to design IIR filters
232 INFINITE IMPULSE RESPONSE FILTERS

with arbitrary magnitude specifications. We will describe all of these functions

in the following sections.
The design of IIR filters based on the bilinear transformation consists of two
steps: (1) estimation of the filter order and (2) computation of the coefficients of
the numerator and denominator of H (z−1 ). The design of these filters starts with
the following specifications:

1. The passband frequency Wp of the lowpass or highpass filter. It is a two-

element vector [Wp1 Wp2] for the design of a bandpass or bandstop filter,
where Wp2>Wp1.
2. The stopband frequency Ws of the lowpass or highpass filter, where Ws>Wp
for the lowpass filter and Ws<Wp for the highpass filter. Ws is a two-
element vector [Ws1 Ws2] for the bandpass or bandstop filter. We have
Ws2>Wp2>Wp1>Ws1 for the bandpass filter and Wp2>Ws2>Ws1>Wp1 for the
bandstop filter. All of these frequencies are specified within the interval
[0 1], where 1 represents the Nyquist frequency.
3. The maximum attenuation Rp (in decibels) in the passband.
4. The minimum attenuation Rs (in decibels) in the stopband.

The four functions to estimate the order of the Butterworth, Chebyshev I, Cheby-
shev II, and elliptic filters are given respectively as

1. [N,Wn] = buttord(Wp,Ws,Rp,Rs)
2. [N,Wn] = cheb1ord(Wp,Ws,Rp,Rs)
3. [N,Wn] = cheb2ord(Wp,Ws,Rp,Rs)
4. [N,Wn] = ellipord(Wp,Ws,Rp,Rs)

where N is the order of the LP and HP filters (2N is the order of the BP and BS
filters) and Wn is the frequency scaling factor. These two variables are then used
in the four MATLAB functions to get the vectors b = [b(1) b(2) b(3) . . .
b(N+1)] and a = [a(1) a(2) a(3) . . . a(N+1)], for the coefficients of the
numerator and denominator of H (z−1 ) in descending powers of z. The constant
coefficient a(1) is equal to unity:

b(1) + b(2)z−1 + b(3)z−2 + · · · + b(N + 1)z−N

H (z−1 ) = (4.102)
1 + a(2)z−1 + a(3)z−2 + · · · + a(N + 1)z−N
Note that in the function [N,Wn] = buttord(Wp,Ws,Rp,Rs), the value of Rp is
restricted to 3 dB because the analog prototype lowpass filter chosen to design the
Butterworth filter obtained under the bilinear transformation uses a 3 dB band-
width. There are several MATLAB functions available in the Signal Processing
Toolbox for the design of analog filters— buttap, cheb1ap, cheb2ap, and
ellipap —and the functions for implementing the analog frequency transforma-
tions discussed in this chapter are lp2bp, lp2hp, lp2bp and lp2bs. However,
we focus on the four MATLAB functions that are available for designing IIR
 IIR FILTER DESIGN USING MATLAB 233

digital filters. They are described below, after we have obtained the order N of
the IIR filter:

1. [b,a] = butter(N,Wn)
2. [b,a] = cheby1(N,Rp,Wn)
3. [b,a] = cheby2(N,Rs,Wn)
4. [b,a] = ellip(N,Rp,Rs,Wn)

After we have obtained the coefficients of the transfer function, we use the
function freqz(b,a,N0) to get the magnitude response, phase response, and
group delay response, which can then be plotted. N0 is the number of dis-
crete frequencies in the interval [0 π] which is chosen by the user. For the
design of a high pass filter and a bandstop filter, we have to include a string
’high’ and ’stop’ as the last argument in the filter functions, for example,
[b,a] = butter(N,Wn,’high’) for designing a Butterworth highpass filter
and [b,a]=cheby2(N,Rs,Wn,’stop’) for designing a Chebyshev II stopband
filter. In these functions, the value of N and Wn are those obtained in the first step,
as the output variables from the functions for estimating the order of the filter.
We illustrate the use of these MATLAB functions by a few examples.

Example 4.11

%MATLAB script to design a Elliptic Lowpass filter

% with the specifications:.Wp = 0.4,Ws = 0.5, Rp = 0.5,
% Rs = 60
[N,Wn]=ellipord(0.4, 0.5, 0.5,60);
[b,a]=ellip(N,0.5,60,Wn);
[h,w]=freqz(b,a,256);
H=abs(h);
HdB=20*log10(H);
plot(w/pi,H);grid
title(’Magnitude response of a Elliptic Lowpass filter’)
ylabel(’Magnitude’)
xlabel(’Normalized frequency’)
figure
plot(w/pi,HdB);grid
title(’Magnitude response of a Elliptic Lowpass filter’)
ylabel(’Magnitude in~dB’)
xlabel(’Normalized frequency’)
%end

The order of this filter is found to be 7, and its magnitude is plotted in

Figures 4.23 and 4.24. Figure 4.23 shows the equiripple in the passband; Figure
4.24 shows the equiripple magnitude (in decibels) in the stopband, and the min-
imum attenuation is seen to be 60 dB.
234 INFINITE IMPULSE RESPONSE FILTERS

Magnitude response of a elliptic lowpass filter

1.4

1.2

1
Magnitude

0.8

0.6

0.4

0.2

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency

Figure 4.23 Magnitude response of an elliptic lowpass filter.

Magnitude response of an Elliptic Lowpass Filter

20

−20
Magnitude in dB

−40

−60

−80

−100

−120
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency

Figure 4.24 Magnitude response (in decibels) of an elliptic lowpass filter.

Example 4.12

%MATLAB Script to design a Chebyshev I (Equiripple) Bandpass

% filter with Ws1 = 0.25, Wp1 = 0.3, Wp2 = 0.4 ,Ws2 = 0.45,
% Rp = 0.5, Rs = 50
[N,Wn]=chebyord([0.3 0.4],[0.25 0.45],0.5,50);
 IIR FILTER DESIGN USING MATLAB 235

Magnitude resopnse of a Chebyshev I bandpass filter

1

0.9

0.8

0.7
Magnitude

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency

Figure 4.25 Magnitude response of a Chebyshev I (equiripple) bandpass filter.

[b,a]=cheby1(N,0.5,Wn)
[h,w]=freqz(b,a,256);
H=abs(h);
HdB=20*log10(H);
plot(w/pi,H);grid
title(’Magnitude response of a Chebyshev I Bandpass filter’)
ylabel(’Magnitude’)
xlabel(’Normalized frequency’)
%end

The order of this filter is found to be 12, and its magnitude response is shown
in Figure 4.25.

Example 4.13

%MATLAB script to design a Butterworth Bandstop filter

% with Wp1 = 0.18, Ws1 = 0.2, Ws2 = 0.205, Wp2 = 0.24,
% Rp = 0.5 and Rs = 50
Wp=[0.18 0.24];
Ws=[0.2 0.205];
[N,Wn]=buttord(Wp,Ws,0.5,50);
236 INFINITE IMPULSE RESPONSE FILTERS

[b,a]=butter(N,Wn,’stop’);
[h,w]=freqz(b,a,256);
H=abs(h);
plot(w/pi,H);grid
title(’Magnitude response of a Butterworth Bandstop filter’)
ylabel(’Magnitude’)
xlabel(’Normalized frequency’)
%end

The order of this filter is 8, and its magnitude response, shown in Figure 4.26,
acts like a notch filter. It can be used to filter out a single frequency at which
the attenuation is more than 65 dB. Since this frequency is ω = 0.2, it is 20%
of the Nyquist frequency or 10% of the sampling frequency. So if the sampling
frequency is chosen as 600 Hz, we can use this filter to filter out the undesirable
hum at 60 Hz due to power supply in an audio equipment.
The coefficients of the digital filter are copied below from the output of the
MATLAB script shown above:

b = numerator coefficients (columns1–9):

0.9168 −5.9000 17.9049 −32.9698 40.1175 −32.9698 17.9049
−5.9000 0.9168

Magnitude response of a Butterworth bandstop filter

1.4

1.2

1
Magnitude

0.8

0.6

0.4

0.2

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency

Figure 4.26 Magnitude response of a Butterworth bandstop filter.

 IIR FILTER DESIGN USING MATLAB 237

a = denominator coefficients (columns1–9):

1.0000 −6.2955 18.6910 −33.6739 40.0927 −32.2433 17.1366
−5.5267 0.8406

Example 4.14

%MATLAB script to design a Chebyshev II highpass filter

%with Wp=0.5,Ws=0.4,Rp=0.5 and Rs=60
[N,Wn]=cheb2ord(0.5,0.4,0.5,60)
[b,a]=cheby2(N,60,Wn,’high’);
[h,w]=freqz(b,a,256);
H=abs(h);
HdB=20*log10(H);
plot(w/pi,H);grid
title(’Magnitude response of a Chebyshev II Highpass Filter’)
ylabel(’Magnitude’)
xlabel(’Normalized frequency’)
%end

The magnitude response of this filter is of order 11, is shown in Figures 4.27
and 4.28.

Magnitude response of a Chebyshev II highpass filter

1

0.9

0.8

0.7

0.6
Magnitude

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency

Figure 4.27 Magnitude response of a Chebyshev II highpass filter.

238 INFINITE IMPULSE RESPONSE FILTERS

Magnitude response of a Chebyshev II highpass filter

0

−20

−40
Magnitude in dB

−60

−80

−100

−120
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency

Figure 4.28 Magnitude response (in decibels) of a Chebyshev II highpass filter.

4.10 YULE–WALKER APPROXIMATION

Now we introduce another function called yulewalk to find an IIR filter that
approximates an arbitrary magnitude response. The method minimizes the error
between the desired magnitude represented by a vector D and the magnitude of
the IIR filter H (ej ω ) in the least-squares sense.
In addition to the maximally flat approximation and the minimax (Chebyshev
or equiripple) approximation we have discussed so far, there is the least- squares
approximation, which is used extensively in the design of filters as well as other
systems. The error that is minimized in a more general case is known as the
least-pth approximation. It is defined by

 p
J3 (ω) = W (ej ω ) H (ej ω ) − D(ej ω ) dω
ω∈R

and when p = 2, it is known as the least-squares approximation. In the error

function shown above, D(ej ω ) is the desired frequency response and H (ej ω )
is the response of the filter designed, whereas W (ej ω ) is a weighting function
chosen by the designer. It has been found that as p approaches ∞, the error
is minimized in the minimax sense, and in practice, choosing p = 4, 5, 6 gives
a good approximation to D(ej ω ) in the least-pth sense [14]. It is best to avoid
sharp transitions in the specifications for the desired magnitude for the IIR filter
when we use the MATLAB function yulewalk. The function has the form

[num,den] = yulewalk(N,F,D)
 YULE–WALKER APPROXIMATION 239

where F is a vector of discrete frequencies in the range between 0 and 1.0, where
1.0 represents the Nyquist frequency; the vector F must include 0 and 1.0. The
vector D contains the desired magnitudes at the frequencies in the vector F; hence
the two vectors have the same length. N is the order of the filter. The coefficients
of the numerator and denominator are output data in the vectors num and den as
shown in (4.102).

Example 4.15

%MATLAB script to design a IIR filter using the function

% yulewalk.
F=[0 0.3 0.7 0.8 1.0];
D=[0 0.8 0.6 0.3 0.5];
[num,den]=yulewalk(10,F,D);
[h,w]=freqz(num,den,256);
H=abs(h);
plot(w/pi,H);grid
title(’Magnitude of an IIR filter by Yulewalker
approximation’)
ylabel(’Magnitude’)
xlabel(’Normalized frequency’)

The magnitude of the IIR filter of order 10 obtained in this example is shown
in Figure 4.29. We can increase or decrease the order of the filter and choose the
design that satisfies the requirements for the application under consideration.

Magnitude of an IIR filter by Yule−Walker approximation

0.8 +

0.7

0.6 +

0.5 +
Magnitude

0.4

0.3 +

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency

Figure 4.29 Magnitude response of an IIR filter, using Yule–Walker approximation.

240 INFINITE IMPULSE RESPONSE FILTERS

4.11 SUMMARY

In this chapter, three major topics have been discussed. First, the theory and
design procedure for approximating the piecewise constant magnitude of ideal
analog filters was discussed, followed by the theory and design procedure for
the design of the IIR filters. These are lowpass, highpass, bandpass, or bandstop
filters that approximate the desired piecewise constant magnitudes in either the
maximally flat sense or the equiripple sense. It is to be pointed out that the
constant group delay of analog filters does not transform to a constant group delay
of the IIR filter obtained by the bilinear transformation. Separate procedures for
designing IIR filters that approximate a constant group delay have been described
in [10].
Next we described the MATLAB functions that are used for designing these
IIR filters as well elliptic function filters. Finally we described the use of the
MATLAB function yulewalk that approximates an arbitrary magnitude response
in the least-squares sense. Design of IIR filters that approximate given fre-
quency specifications with additional approximation criteria are described in
Chapter 7.

PROBLEMS

4.1 Find the function |H (j ω)|2 from the transfer functions given below:
s+3
H1 (s) =
s 2 + 2s + 2
s2 + s + 1
H2 (s) =
s(s 2 + 4s + 20)
4.2 Find the transfer function H (s) from the functions given below:

(ω2 + 9)
|H1 (j ω)|2 = (4.103)
(ω2 + 4)(ω2 + 1)
(ω2 + 4)
|H2 (j ω)|2 = (4.104)
(ω2 + 16)(ω4 + 1)

4.3 An analog signal x(t) = e−2t u(t) is sampled to generate the discrete-
time sequence x(nT ) = e−2nT u(n). Find the z transform X(z) of the DT
sequence for T = 0.1, 0.05, 0.01 s.
4.4 An analog signal x(t) = 10 cos(2t)u(t) is sampled to generate the discrete-
time sequence x(nT ) = 10 cos(2nT )u(n). Find the z transform X(z) of
the DT sequence for T = 0.1, 0.01 s.
4.5 Derive transfer function H1 (z) obtained when the impulse-invariant trans-
formation is applied and H2 (z) when the bilinear transformation s =
 PROBLEMS 241

2[(z − 1)/(z + 1)] is applied to the transfer function H (s) = 1/(s 2 +

√
2s + 1). Simplify the transfer functions as the ratio of polynomials in z.
4.6 Find the z transform obtained by the impulse-invariant transformation of
H (s) given below, assuming T = 0.1 s:
1
H1 (s) =
s(s + 1)(s + 3)
(s + 1)
H2 (s) =
(s + 2)(s 2 + 2s + 5)
s(s 2 + 4s + 10)
H3 (s) =
(s + 5)(s 2 + 6s + 15)
4.7 By application of the impulse-invariant transformation of H (s), the follow-
ing transfer functions have been obtained. Find H (s), assuming T = 0.1 s:
2z
H1 (z) =
(z − e−0.2 )(z − e−0.1 )
z − e−0.6
H2 (z) =
(z − e−0.5 )(z − e−0.4 )
z
H3 (z) =
(z − 0.9)(z − 0.3)
z
H4 (z) =
(z − 0.4)(z − 0.8)
4.8 The following transfer functions of a digital filter are obtained by applying
the bilinear transformation on analog transfer functions H (s). Derive the
transfer functions H (s), assuming T = 0.1 s:
z+1
H1 (z) =
+z+6
z2
z+4
H2 (z) = 2
z + 6z + 8
2z + 5
H3 (z) = 2
z + 2z + 2
4.9 Find the magnitude (in decibels) at the frequency ωs = 1000 rad/s of
a Fourth-order Chebyshev I highpass, analog filter with the magnitude
response shown in Figure 4.30.
4.10 A Chebyshev I bandstop analog filter is to be designed to approximate
the following specifications: ωp1 = 104 and ωp2 = 7 × 104 are the cutoff
frequencies of the passband. Passband ripple = 0.5 dB, ωs1 = 2 × 104 is a
frequency in the stopband and the minimum attenuation in the stopband =
30 dB. What is the stopband frequency
s and the order of the lowpass
prototype filter H (p)?
242 INFINITE IMPULSE RESPONSE FILTERS

|H(jω)|dB

0 dB
−0.4 dB

0 1000 3000 w

Figure 4.30 Problem 4.9.

0 dB
−1.5 dB

−55 dB
w
4 × 104 105 4 × 105

Figure 4.31 Problem 4.11.

4.11 What is the order of an analog bandpass Chebyshev I filter that has a
magnitude response as shown in Figure 4.31?
4.12 Determine the sampling period T such that a frequency s = j 15 of the
analog filter maps to the normalized frequency ω = 0.3π of the digital
filter.
4.13 A digital Butterworth lowpass filter is designed by applying the bilinear
transformation on the transfer function of an analog Butterworth low-
pass filter that has an attenuation of 45 dB at 1200 rad/s. What is the
 PROBLEMS 243

frequency in rad/s at which the digital filter has an attenuation of 45 dB,

if its sampling frequency is 1000 Hz?
4.14 A digital Chebyshev bandpass filter is designed by applying the bilinear
transformation on the transfer function of an analog Chebyshev bandpass
filter that has a bandwidth of 200 rad/s between 1400 and 1600 rad/s.
What is the bandwidth of the digital filter, if its sampling frequency is
1500 Hz?
4.15 A Butterworth bandpass IIR digital filter designed by using the bilinear
transformation has a magnitude response as shown in Figure 4.32. Find
the order of the filter.
4.16 An IIR lowpass Butterworth filter is to be designed, using the bilinear
transformation, with a cutoff frequency ωc = 0.3π at which the maximum
attenuation is 0.5 dB. The minimum attenuation at the stopband cutoff
frequency ωs = 0.8π is 40 dB. Find the transfer function H (p) of the
lowpass prototype filter, when the bilinear transformation is used.
4.17 Find the magnitude of a lowpass Butterworth IIR filter at the frequency
(0.8π). The order of the filter is 10, the sampling frequency is 103 Hz,
and the 3 dB bandwidth of the filter is (0.2π).
4.18 What are the frequencies ωs2 and ωs1 at which the 10th-order Butterworth
analog bandpass filter has a magnitude of 0.045? The filter has a magnitude
response as shown in Figure 4.33.
4.19 What is the order of a Chebyshev bandpass IIR filter that has a magnitude
response such as that shown in Figure 4.34.

|H(e jω)|

1.00
0.96

0.004

2000 3000 4000

Figure 4.32 Problem 4.15.

244 INFINITE IMPULSE RESPONSE FILTERS

1.00
0.98

0.045
ws1 3000 5000 ws2 w

Figure 4.33 Problem 4.18.

|H(e j 2pω)|

1.00
0.95

0.002
200 300 400 1000 f

Figure 4.34 Problem 4.19.

4.20 Design a Chebyshev analog highpass filter that approximates the specifi-
cations as shown in Figure 4.35.

4.21 Design a Butterworth bandpass IIR filter that approximates the specifica-
tions given in Figure 4.36. Show all calculations step by step. Plot the
magnitude using MATLAB.

4.22 A Butterworth bandpass IIR filter of order 10 meets the following spec-
ifications: ωp1 = 0.5π, ωp2 = 0.65π, ωs2 = 0.8π, Ap = 0.5 dB. What is
the attenuation at ωs2 ?

4.23 A Chebyshev I bandstop digital filter meets the satisfy the following spec-
ifications: ωp1 = 0.1π, ωp2 = 0.8π, ωs2 = 0.4π, αp = 0.8, αs = 55. Find
the transfer function H (p) of the lowpass analog prototype filter.
 PROBLEMS 245

|H(jw)|

1.00
0.93

0.004

300 800 w

Figure 4.35 Problem 4.20.

|H(e jwT)|
1.00
0.93

0.004

0.4p 0.5p 0.7p p (wT)

Figure 4.36 Problem 4.21.

4.24 A Butterworth highpass digital filter has a passband cutoff frequency of

1500 Hz, a stopband cutoff frequency of 650 Hz, a passband attenuation
of 2.0 dB, and a stopband attenuation of 45 dB. The sampling period
is 0.2 × 10−3 s. Find the transfer function H (p) of the lowpass analog
prototype filter.
4.25 Design a bandpass Chebyshev I IIR filter that approximates the specifica-
tions as shown in Figure 4.37. Show all calculations step by step, assuming
a sampling frequency fs = 1000 Hz. Plot the magnitude using MATLAB.
4.26 Obtain the transfer function of the Butterworth lowpass IIR filter with ωp =
0.4π, ωs = 0.9π, Rp = 0.5, Rs = 20. Using the digital spectral transforma-
tion LP–HP, find the transfer function of the highpass filter with a cutoff
frequency of 0.75π.
246 INFINITE IMPULSE RESPONSE FILTERS

1.00
0.92

0.003

0.25π 0.35π 0.6π π

Figure 4.37 Problem 4.25.

X(z) Σ Σ Y(z)

a1 a2
Σ Σ

Figure 4.38 Problem 4.27a.

X(z) Σ Σ Y(z)

−1.0
0.5

0.06
Σ Σ

Figure 4.39 Problem 4.27b.

4.27 Derive the transfer function of the two circuits shown in Figures 4.38 and
4.39 and verify that they are allpass filters.
4.28 The transfer function of an analog allpass filter H (s) = (s 2 − as + b)/
(s 2 + as + b) has a magnitude response equal to one at all frequencies.
 REFERENCES 247

Show that the IIR filter obtained by the application of the bilinear trans-
formation on H (s) is also an allpass digital filter.

MATLAB Problems
4.29 Design a Butterworth bandstop filter with Wp1 = 0.2, Ws1 = 0.35, Ws2 =
0.55, Wp2 = 0.7.0, Rp = 0.25, and Rs = 45. Plot the magnitude and the
group delay response.
4.30 Design a Chebyshev I bandpass filter to meet the following specifications:
Ws1 = 0.4, Wp1 = 0.45, Wp2 = 0.55, Ws2 = 0.6, Rp = 0.3, Rs = 50. Plot
the magnitude (in decibels) and the group delay to verify that the given
specifications have been met.
4.31 Design a Chebyshev II highpass filter with Ws = 0.1, Wp = 0.3, Rp = 0.8,
Rs = 60 dB. Plot the magnitude (in decibels) and the group delay of the
filter to verify that the design meets the specifications.
4.32 Design an elliptic lowpass filter with Wp = 0.2, Ws = 0.35, Rp = 0.8,
Rs = 40. Plot the magnitude (in decibels) and the group delay of the
filter.
4.33 Design an elliptic lowpass filter with Wp1 = 0.3, Ws = 0.4, Rp = 0.5,
Rs = 55. Plot the magnitude (in decibels) and the group delay of the
filter. Plot a magnified plot of the response in the stopband to verify that
the specifications have been met.
4.34 Design a Butterworth bandpass filter with Ws1 = 0.3, Wp1 = 0.5, Wp2 =
0.55, Ws2 = 0.8, Rp = 0.5, and Rs = 50. Plot its magnitude and phase
response.
4.35 Design an IIR filter with the following specifications: F = [0 0.2 0.4
0.5 1.0], D = [1.0 0.5 0.7 0.9 1.0], using yulewalk function.
Plot the magnitude of the filter.
4.36 Design an IIR filter with the following specifications, using the MATLAB
function yulewalk: F = [0.0 0.3 0.5 0.7 0.9 1.0]; D = [0.2
0.4 0.5 0.3 0.6 1.0]. Plot the magnitude of the filter.
4.37 Design an IIR filter that approximates the magnitude response with the
specifications F = [0.0 0.2 0.4 0.6 0.8 1.0]; D = [1.0 0.18
0.35 0.35 0.18 1.0] using the MATLAB function yulewalk. Plot
the magnitude and group delay response of the filter.

REFERENCES

1. A. C. Constantinides, Spectral transformations for digital filters, Proc. IEE 117,

1585–1590 (Aug. 1970).
2. A. I. Zverev, Handbook of Filter Synthesis, Wiley, 1967.
248 INFINITE IMPULSE RESPONSE FILTERS

3. R. Schaumann and M. E. Van Valkenburg, Design of Analog Filters, Oxford Univ.

Press, 2001.
4. S. K. Mitra, Digital Signal Processing—A Computer-Based Approach, McGraw-Hill,
2001.
5. J. G. Proakis and D. G. Manolakis, Digital Signal Processing, Prentice-Hall, 1996.
6. A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing, Prentice-Hall,
1989.
7. A. Antoniou, Digital Filters, Analysis, Design and Applications, McGraw-Hill, 1993.
8. E. C. Ifeachor and B. W. Jervis, Digital Signal Processing, A Practical Approach,
Prentice-Hall, 2002.
9. L. R. Rabiner and B. Gold, Theory and Application of Digital Signal Processing,
Prentice-Hall, 1975.
10. B. A. Shenoi, Magnitude and Delay Approximation of 1-D and 2-D Digital Filters,
Springer-Verlag, 1999.
11. V. K. Ingle and J. G. Proakis, Digital Signal Processing Using MATLAB (R) V.4, PWS
Publishing, 1997.
12. R. W. Hamming, Digital Filters, Prentice-Hall, 1989.
13. L. R. Rabiner and B. Gold, Theory and Application of Digital Signal Processing,
Prentice-Hall, 1975.
14. G. C. Temes and D. Y. F. Zai, Least pth approximation, IEEE Trans. Circuit Theory
CT-16, 235–237 (1969).
15. MATLAB(R) and Signal Processing Tool Box, software from The MathWorks, Natick,
MA.
 CHAPTER 5

Finite Impulse Response Filters

5.1 INTRODUCTION

From the previous two chapters, we have become familiar with the magnitude
response of ideal lowpass, highpass, bandpass, and bandstop filters, which was
approximated by IIR filters. In the previous chapter, we also discussed the theory
and a few prominent procedures for designing the IIR filters.
The general form of the difference equation for a linear, time-invariant,
discrete-time system (LTIDT system) is

N
M
y(n) = − a(k)y(n − k) + b(k)x(n − k) (5.1)
k=1 k=0

The transfer function for such a system is given by

b0 + b(1)z−1 + b(2)z−2 + · · · + b(M)z−M

H (z−1 ) = (5.2)
1 + a(1)z−1 + a(2)z−2 + a(3)z−3 + · · · + a(N )z−N

The transfer function of an FIR filter, in particular, is given by

H (z−1 ) = b0 + b(1)z−1 + b(2)z−2 + · · · + b(M)z−M (5.3)

and the difference equation describing this FIR filter is given by

M
y(n) = b(k)x(n − k) (5.4)
k=0
= b(0)x(n) + b(1)x(n − 1) + · · · + b(M)x(n − M) (5.5)

In this chapter, the properties of the FIR filters and their design will be dis-
cussed. When the input function x(n) is the unit sample function δ(n), the

Introduction to Digital Signal Processing and Filter Design, by B. A. Shenoi

Copyright © 2006 John Wiley & Sons, Inc.

249
250 FINITE IMPULSE RESPONSE FILTERS

output y(n) can be obtained by applying the recursive algorithm on (5.4). We

get the output y(n) due to the unit sample input δ(n) to be exactly the values
b(0), b(1), b(2), b(3), . . . , b(M). The output due to the unit sample function δ(n)
is the unit sample response or the unit impulse response denoted by h(n). So
the samples of the unit impulse response h(n) = b(n), which means that the unit
impulse response h(n) of the discrete-time system described by the difference
equation (5.4) is finite in length. That is why the system is called the finite impulse
response filter or the FIR filter. It has also been known by other names such as
the transversal filter, nonrecursive filter, moving-average filter, and tapped delay
filter. Since h(n) = b(n) in the case of an FIR filter, we can represent (5.3) in
the following form:

M
H (z−1 ) = h(k)z−k = h(0) + h(1)z−1 + h(2)z−2 + · · · + h(M)z−(M) (5.6)
k=0

The FIR filters have a few advantages over the IIR filters as defined by (5.1):

1. We can easily design the FIR filter to meet the required magnitude response
in such a way that it achieves a constant group delay. Group delay is defined
as τ = −(dθ/dω), where θ is the phase response of the filter. The phase
response of a filter with a constant group delay is therefore a linear function
of frequency. It transmits all frequencies with the same amount of delay,
which means that there will not be any phase distortion and the input signal
will be delayed by a constant when it is transmitted to the output. A filter
with a constant group delay is highly desirable in the transmission of digital
signals.
2. The samples of its unit impulse response are the same as the coefficients
of the transfer function as seen from (5.5) and (5.6). There is no need to
calculate h(n) from H (z−1 ), such as during every stage of the iterative opti-
mization procedure or for designing the structures (circuits) from H (z−1 ).
3. The FIR filters are always stable and are free from limit cycles that arise
as a result of finite wordlength representation of multiplier constants and
signal values.
4. The effect of finite wordlength on the specified frequency response or the
time-domain response or the output noise is smaller than that for IIR filters.
5. Although the unit impulse response h(n) of an IIR filter is an infinitely
long sequence, it is reasonable to assume in most practical cases that the
value of the samples becomes almost negligible after a finite number; thus,
choosing a sequence of finite length for the discrete-time signal allows us
to use powerful numerical methods for processing signals of finite length.

5.1.1 Notations
It is to be remembered that in this chapterwe choose the order of the FIR
filter or degree of the polynomial H (z−1 ) = N
n=0 h(n)z
−n
as N , and the length
 LINEAR PHASE FIR FILTERS 251

of the filter equal to the number of coefficients in (5.6) is N + 1. If we are

given H (z−1 ) = 0.3z−4 + 0.1z−5 + 0.5z−6 , its order is 6, although only three
terms are present and the correct number of coefficients equal to the length of
the filter is 7, because h(0) = h(1) = h(2) = h(3) = 0. It becomes necessary to
point out the notation
N −1 used −n
in this chapter, because in some textbooks, we may
−1
find H (z ) = n=0 h(n)z representing the transfer function of an FIR filter,
in which case the length of the filter is denoted by N and the degree or order
of the polynomial is (N − 1). (Therefore students have to be careful in using
the formulas found in a chapter on FIR filters, in different books; but with some
caution, they can replace N that appears in this chapter by (N − 1) so that the
formulas match those found in these books.)
The notation often used in MATLAB, is H (z−1 ) = h(1) + h(2)z−1 +
h(3)z−2 + · · · + h(N + 1)z−N , which is a polynomial of degree N , and has
(N + 1) coefficients. In more compact form, it is given by

N
H (z−1 ) = h(n + 1)z−n (5.7)
n=0

The notation and meaning of angular frequency used in the literature on discrete-
time systems and digital signal processing also have to be clearly understood by
the students. One is familiar with a sinusoidal signal x(t) = A sin(wt) in which
w = 2πf is the angular frequency in radians per second, f is the frequency in
hertz, and its reciprocal is the period Tp in seconds. So we have w = 2π/Tp
radians per second. Now if we sample this signal with a uniform sampling
period, we need to differentiate the period Tp from the sampling period denoted
by Ts . Therefore, the sampled sequence is given by x(nTs ) = A sin(wnTs ) =
A sin(2πnTs /Tp ) = A sin(2πf/fs ) = A sin(w/fs ). The frequency w (in radians
per second) normalized by fs is almost always denoted by ω and is called the
normalized frequency (measured in radians). The frequency w is the analog fre-
quency variable, and the frequency ω is the normalized digital frequency. On
this basis, the sampling frequency ωs = 2π radians. Sometimes, w is normalized
by πfs or 2πfs so that the corresponding sampling frequency becomes 2 or 1
radian(s). Note that almost always, the sampling period is denoted simply by T
in the literature on digital signal processing when there is no ambiguity and the
normalized frequency is denoted by ω = wT . The difference between the angular
frequency in radians per second and the normalized frequency usually used in
DSP literature has been pointed out in several instances in this book.

5.2 LINEAR PHASE FIR FILTERS

Now we consider the special types of FIR filters in which the coefficients h(n)
of the transfer function H (z−1 ) = Nn=0 h(n)z
−n
are assumed to be symmetric
or antisymmetric. Since the order of the polynomial in each of these two types
252 FINITE IMPULSE RESPONSE FILTERS

can be either odd or even, we have four types of filters with different properties,
which we describe below.

Type I. The coefficients are symmetric [i.e., h(n) = h(N − n)], and the order
N is even.

Example 5.1

Let us consider a simple example:

H (z−1 ) = h(0) + h(1)z−1 + h(2)z−2 + h(3)z−1 + h(4)z−4

+ h(5)z−5 + h(6)z−6 .

As shown in Figure 5.1a, for this type I filter, with N = 6, we see that h(0) =
h(6), h(1) = h(5), h(2) = h(4). Using these equivalences in the above, we get

H (z−1 ) = h(0)[1 + z−6 ] + h(1)[z−1 + z−5 ] + h(2)[z−2 + z−4 ] + h(3)z−3

(5.8)
This can also be represented in the form
 
H (z−1 ) = z−3 h(0)[z3 + z−3 ] + h(1)[z2 + z−2 ] + h(2)[z + z−1 ] + h(3)
(5.9)

h(n) h(n)

0 1 2 3 4 5 6
0 1 2 3 4 5 6 7

Center of symmetry
Type I N = 6 Type II N = 7
(a) (b)

h(n) h(n)

0 1 2 3 4 5 6 7
0 1 2 3 4 5 6

Center of antisymmetry
Type III N = 6 Type IV N = 7
(c) (d )

Figure 5.1 Unit impulse responses of the four types of linear phase FIR filters.
 LINEAR PHASE FIR FILTERS 253

Let us evaluate its frequency response (DTFT):

H (e−j ω ) = e−j 3ω {2h(0) cos(3ω) + 2h(1) cos(2ω) + 2h(2) cos(ω) + h(3)}

= ej θ(ω) {HR (ω)}

The expression HR (ω) in this equation is a real-valued function, but it can be

positive or negative at any particular frequency, so when it changes from a
positive value to a negative value, the phase angle changes by π radians (180◦ ).
The phase angle θ (ω) = −3ω is a linear function of ω, and the group delay τ is
equal to three samples. Note that on the normalized frequency basis, the group
delay is three samples but actual group delay is 3T seconds, where T is the
sampling period.
In the general case, we can express H (ej ω ) in a few other forms, for example

N
H (ej ω ) = h(n)e−j nω
n=0

= h(0) + h(1)e−j ω + h(2)e−j 2ω + · · · + h(N − 1)e−j (N ω)

     
−j [(N/2)ω] Nω N
=e 2h(0) cos + 2h(1) cos −1 ω
2 2
    
N N
+ 2h(2) cos − 2 ω + ··· + h (5.10)
2 2

We put it in a more compact form:

⎧ ⎫
⎨ N  N/2 
N
 ⎬
H (ej ω ) = e−j [(N/2)ω] h +2 h − n cos(nω) = ej θ(ω) {HR (ω)}
⎩ 2 2 ⎭
n=1
(5.11)
The total group delay is a constant = N/2 in the general case, for a type I FIR
filter.

Type II. The coefficients are symmetric [i.e., h(n) = h(N − n)], and the order
N is odd.

Example 5.2

Here we consider an example in which the coefficients are symmetric but N = 7,

as shown in Figure 5.1b. For this example, we have

H (z−1 ) = h(0) + h(1)z−1 + h(2)z−2 + h(3)z−1 + h(4)z−4 + h(5)z−5

+ h(6)z−6 + h(7)z−7
254 FINITE IMPULSE RESPONSE FILTERS

and because of symmetry

h(0) = h(7), h(1) = h(6), h(2) = h(5), h(3) = h(4).

Therefore

H (z−1 ) = h(0)[1 + z−7 ] + h(1)[z−1 + z−6 ] + h(2)[z−2 + z−5 ]

+ h(3)[z−3 + z−4 ]

The frequency response is given by

H (e−j ω ) = e−j 3.5ω {2h(0) cos(3.5ω) + 2h(1) cos(2.5ω)

+ 2h(2) cos(1.5ω) + 2h(3) cos(0.5ω)}
= ej θ(ω) {HR (ω)}

The phase angle θ (ω) = −3.5ω, and the group delay is τ = 3.5 samples.
In the general case of type II filter, we obtain

N
H (e−j ω ) = h(n)e−j nω = ej θ(ω) {HR (ω)}
n=0
⎧ ⎫
⎨(N
+1)/2     ⎬
−j ( N2 ω) N +1 1
=e 2h − n cos n − ω) (5.12)
⎩ 2 2 ⎭
n=1

which shows a linear phase θ (ω) = −[(N/2)ω] and a constant group delay =
N/2 samples.

Type III. The coefficients are antisymmetric [i.e., h(n) = −h(N − n)], and the
order N is even.

Example 5.3

We consider an example of type III FIR filter of order N = 6 and as shown in

Figure 5.1c, we have h(0) = −h(6), h(1) = −h(5), h(2) = −h(4) and we must
have h(3) = 0 to maintain antisymmetry for these samples:

H (z−1 ) = h(0)[1 − z−6 ] + h(1)[z−1 − z−5 ] + h(2)[z−2 − z−4 ] (5.13)

 
= z−3 h(0)[z3 − z−3 ] + h(1)[z2 − z−2 ] + h(2)[z − z−1 ] (5.14)
 LINEAR PHASE FIR FILTERS 255

Now if we put z = ej ω , and ej ω − e−j ω = 2j sin(ω) = 2ej (π/2) sin(ω), we arrive

at the frequency response for this filter as

H (e−j ω ) = e−j 3ω {h(0)2j sin(3ω) + h(1)2j sin(2ω) + h(2)2j sin(ω)} (5.15)

= e−j 3ω ej (π/2) {2h(0) sin(3ω) + 2h(1) sin(2ω) + 2h(2) sin(ω)} (5.16)
= e−j [3ω−(π/2)] HR (ω) (5.17)

Note that the phase angle for this filter is θ (ω) = −3ω + π/2, which is still a
linear function of ω. The group delay is τ = 3 samples for this filter.
In the general case, it can be shown that
⎧ ⎫
⎨ N/2   ⎬
N
H (e−j ω ) = e−j [(Nω−π)/2] 2 h − n sin(nω) (5.18)
⎩ 2 ⎭
n=1

and it has a linear phase θ (ω) = −[(N ω − π)/2] and a group delay τ = N/2
samples.

Type IV. The coefficients are antisymmetric [i.e., h(n) = −h(N − n)], and the
order N is odd.

Example 5.4

We consider an example of type IV filter with N = 7 as shown in Figure 5.1d, in

which h(0) = −h(7), h(1) = −h(6), h(2) = −h(5), h(3) = −h(4). Its transfer
function is given by

H (z−1 ) = h(0)[1 − z−7 ] + h(1)[z−1 − z−6 ] + h(2)[z−2 − z−5 ]

+ h(3)[z−3 − z−4 ] (5.19)

The frequency response can be derived as

H (e−j ω ) = e−j 3.5ω {h(0)[ej 3.5ω − e−j 3.5ω ] + h(1)[ej 2.5ω − e−j 2.5ω ]
+ h(2)[ej 1.5ω − e−j 1.5ω ] + h(3)[ej 0.5ω − e−j 0.5ω ]}
= e−j 3.5ω {h(0)2j sin(3.5ω) + h(1)2j sin(2.5ω) + h(2)2j sin(1.5ω)
+ h(3)2j sin(0.5ω)}
= e −j [3.5ω−(π/2)] {2h(0) sin(3.5ω) + 2h(1) sin(2.5ω) + 2h(2) sin(1.5ω)
+ 2h(3) sin(0.5ω)} (5.20)
256 FINITE IMPULSE RESPONSE FILTERS

This type IV filter with N = 7 has a linear phase θ (ω) = −3.5ω + π/2 and a
constant group delay τ = 3.5 samples.
The transfer function of the type IV linear phase filter in general is given by
⎧ ⎫
+1)/2 
⎨ (N    ⎬
N +1 1
H (e−j ω ) = e−j [(Nω−π)/2] 2 h − n sin n− ω
⎩ 2 2 ⎭
n=1
(5.21)
The frequency responses of the four types of FIR filters are summarized below:
⎧ ⎫
⎨ N  N/2 
N
 ⎬
H (ej ω ) = e−j [(N/2)ω] h +2 h − n cos(nω)
⎩ 2 2 ⎭
n=1

for type I
⎧ ⎫
⎨ (N +1)/2     ⎬
N + 1 1
H (e−j ω ) = e−j [(N/2)ω] 2 h − n cos n− ω
⎩ 2 2 ⎭
n=1

for type II
⎧ ⎫
⎨ N/2   ⎬
N
H (e−j ω ) = e−j [(Nω−π)/2] 2 h − n sin(nω)
⎩ 2 ⎭
n=1

for type III

⎧ ⎫
⎨(N +1)/2     ⎬
N +1 1
H (e−j ω ) = e−j [(Nω−π)/2] 2 h − n sin n− ω
⎩ 2 2 ⎭
n=1

for type IV (5.22)

5.2.1 Properties of Linear Phase FIR Filters

The four types of FIR filters discussed above have shown us that FIR filters
with symmetric or antisymmetric coefficients provide linear phase (or equiva-
lently constant group delay); these coefficients are samples of the unit impulse
response. It has been shown above that an FIR filter with symmetric or anti-
symmetric coefficients has a linear phase and therefore a constant group delay.
The reverse statement, that an FIR filter with a constant group delay must have
symmetric or antisymmetric coefficients, has also been proved theoretically [4].
These properties are very useful in the design of FIR filters and their applica-
tions. To see some additional properties of these four types of filters, we have
evaluated the magnitude response of typical FIR filters with linear phase. They
are shown in Figure 5.2.
The following observations about these typical magnitude responses will be
useful in making proper choices in the early stage of their design, as will be
 LINEAR PHASE FIR FILTERS 257

Magnitude response of Magnitude response of

type I FIR filter type II FIR filter
4 2.5

2
3
Magnitude

Magnitude
1.5
2
1
1
0.5

0 0
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Normalized frequency Normalized frequency

Magnitude response of Magnitude response of

type III FIR filter type IV FIR filter
3.5 3
3 2.5
2.5
2
Magnitude
Magnitude

2
1.5
1.5
1
1
0.5 0.5

0 0
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Normalized frequency Normalized frequency

Figure 5.2 Magnitude responses of the four types of linear phase FIR filters.

explained later. For example, type I filters have a nonzero magnitude at ω = 0 and
also a nonzero value at the normalized frequency ω/π = 1 (which corresponds to
the Nyquist frequency), whereas type II filters have nonzero magnitude at ω = 0
but a zero value at the Nyquist frequency. So it is obvious that these filters are
not suitable for designing bandpass and highpass filters, whereas both of them
are suitable for lowpass filters. The type III filters have zero magnitude at ω = 0
and also at ω/π = 1, so they are suitable for designing bandpass filters but not
lowpass and bandstop filters. Type IV filters have zero magnitude at ω = 0 and
a nonzero magnitude at ω/π = 1. They are not suitable for designing lowpass
and bandstop filters but are candidates for bandpass and highpass filters.
In Figure 5.3a, the phase response of a type I filter is plotted showing the
linear relationship. When the transfer function has a zero on the unit circle in
the z plane, its phase response displays a jump discontinuity of π radians at the
corresponding frequency, and the plot uses a jump discontinuity of 2π whenever
the phase response exceeds ±π so that the total phase response remains within
the principal range of ±π. If there are no jump discontinuities of π radians,
that is, if there are no zeros on the unit circle, the phase response becomes a
258 FINITE IMPULSE RESPONSE FILTERS

Phase response of type I FIR filter Phase response unwrapped

4 2

3 0

2 −2
Phase angle in radians

Phase angle in radians

1 −4

0 6

−1 −8

−2 −10

−3 −12

−4 −14
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Normalized frequency Normalized frequency
(a) (b)

Figure 5.3 Linear phase responses of type I FIR filter.

continuous function of ω when it is unwrapped. The result of unwrapping the

phase (Fig. 5.3a) is to remove the jump discontinuities in the phase response
such that the phase response lies within ±π (Fig. 5.3b). If the order N of the
FIR filter is even, its group delay is an integer multiple of samples equal to N/2
samples. If the order N is odd, then the group delay is equal to (an integer plus
half) a sample. We will use all of these properties before we start the design of
FIR filters with linear phase.
The linear phase FIR filters have some interesting properties in the z plane
also. As seen in the examples, their transfer functions always contain pairs of
terms such as [zn ± z−n] . Denoting the transfer function of the FIR filters with
symmetric coefficients by H (z), we write

N
N
−n
H (z) = h(n)z = h(N − n)z−n (5.23)
n=0 n=0
N
By making a change of variable m = (N − n), we reduce the series n=0 h(N −
n)z−n to

N
N
h(m)z−N +m = z−N h(m)zm = z−N H (z−1 ) (5.24)
m=0 m=0

so we have the following result:

H (z) = z−N H (z−1 ) (5.25)
 LINEAR PHASE FIR FILTERS 259

Similarly, the FIR filters with antisymmetric coefficients satisfy the property

H (z) = −z−N H (z−1 ) (5.26)

A polynomial H (z) satisfying (5.25) is called a mirror image polynomial , and

the polynomial that satisfies (5.26) is called an anti–mirror image polynomial .
We see that a polynomial H (z) that has symmetric coefficients is a mirror image
polynomial and one with antisymmetric coefficients is an anti–mirror image
polynomial. The reverse statement is also true and can be proved, namely, that
a mirror image polynomial has symmetric coefficients and an anti–mirror image
polynomial has antisymmetric coefficients.
From (5.25) and (5.26), it is easy to note that in a mirror image polynomial
as well as an anti–mirror image polynomial, if z = z1 is a zero of H (z), then
1/z is also a zero of H (z). If the zero z1 is a complex number r1 ej φ ; |r| < 1,
then z1∗ = r1 e−j φ is also a zero. Their reciprocals (1/r1 )e−j φ and (1/r1 )ej φ are
also zeros of H (z), which lie outside the unit circle |z| = 1. Therefore complex
zeros of mirror image polynomials and anti–mirror image polynomials appear
with quadrantal symmetry as shown in Figure 5.4. If there is a zero on the unit
circle (e.g., at z0 = ej φ ), its reciprocal z−1 = e−j φ is already located on the unit
circle, as the complex conjugate of z0 , and therefore zeros on the unit circle do
not have quadrantal symmetry. Obviously a zero on the real axis at zr = r inside
the unit circle will be paired with one outside the unit circle on the real axis at
zr−1 = 1/r.

0.5
Imaginary Part

6
0

−0.5

−1

−1.5 −1 −0.5 0 0.5 1 1.5

Real Part

Figure 5.4 Zero and pole locations of a mirror image polynomial.

260 FINITE IMPULSE RESPONSE FILTERS

Example 5.5

We consider the example of a type I FIR filter with H (z−1 ) = 0.4 + 0.6z−1 +
0.8z−2 + 0.2z−3 + 0.8z−4 + 0.6z−5 + 0.4z−6 , to illustrate these properties. When
it is expressed in the form H (z) = z−6 [0.4 + 0.6z + 0.8z2 + 0.2z3 + 0.8z4 +
0.6z5 + 0.4z6 ] and factorized, we get

(z − z1) (z − z1∗ )(z − z−1 )(z − z−1∗ )(z − z2 )(z − z2∗ )

H (z) = (5.27)
z6
◦ ◦ ◦
where z1 = 0.69ej 128.6 , z1∗ = 0.69e−j 128.6 , z1−1 = 1.45e−j 128.6 , z1−1∗ =
◦ ◦ ◦
1.45ej 128.6 , z2 = ej 54.12 , and z2−1∗ = e−j 54.12 . They are plotted in Figure 5.4
along with the six poles of H (z) at z = 0. The two zeros at z2 and z2∗ are on the
unit circle, and the other four zeros form a quadrantal symmetry in this plot. The
magnitude of this type I filter is illustrated in Figure 5.2a, which shows that the
magnitude has zero value at the two frequencies corresponding to the two zeros
at z2 and z2∗ that are on the unit circle. In Figure 5.3, the phase response also
shows discontinuities at these two frequencies.
Some additional properties of the four types of FIR filters are listed below:

1. Type I FIR filters have either an even number of zeros or no zeros at z = 1

and z = −1.
2. Type II FIR filters have an even number of zeros or no zeros at z = 1 and
an odd number of zeros at z = −1.
3. Type III FIR filters have an odd number of zeros at z = 1 and z = −1.
4. Type IV FIR filters have an odd number of zeros at z = 1 and either an
even or odd number of zeros at z = −1.

These properties confirm the properties of the magnitude response of the filters
as illustrated by Figure 5.2. A zero at z = 1 corresponds to ω = 0, and a zero
at z = −1 corresponds to ω = π. As an example, we note that the type III FIR
filter has zero magnitude at ω = 0 and ω = 1, whereas we stated above that the
transfer function of the type III FIR filter has an odd number of zeros both at
z = 1 and z = −1.
Another important result that will be used in the Fourier series method for
designing FIR filters is given below. This is true of all FIR as well as IIR
filters and not just linear phase FIR filters. The Fourier transform (DTFT) of any
discrete-time sequence x(n) is

n=∞
X(ej ω ) = x(n)e−j nω (5.28)
n=−∞
 FOURIER SERIES METHOD MODIFIED BY WINDOWS 261

Since H (ej ω ) is a periodic function with a period of 2π, it has a Fourier series
representation in the form

n=∞
X(e ) =
jω
c(n)e−j nω (5.29)
n=−∞

where

π
1
c(n) = X(ej ω )ej nω dω (5.30)
2π −π

Comparing (5.28) and (5.29), we see that x(n) = c(n) for −∞ < n < ∞.
n=∞ we consider
When the frequency response of the LTI-DT system H (ej ω ) =
−j nω , where h(n) = 0 for n < 0, we will find that c(n) = 0 for
n=0 h(n)e
n < 0. So we note that the Fourier series coefficients c(n) evaluated from (5.30)
are the same as the coefficients h(n) of the IIR or FIR filter. Evaluating the
coefficients c(n) = h(n) by the integral in the Equation (5.30) is easy when we
choose H (ej ω ) to be a constant in the subinterval within the interval of inte-
gration [−π, π] with zero phase or when H (ej ω ) is piecewise constant over
different disjoint passbands and stopbands, within [−π, π]. This result facili-
tates the design of FIR filters that approximate the magnitude response of ideal
lowpass, highpass, bandpass, and bandstop filters.1 The Fourier series method
based on the abovementioned properties of FIR filters for designing them is
discussed next.

5.3 FOURIER SERIES METHOD MODIFIED BY WINDOWS

The magnitude responses of four ideal classical types of digital filters are shown in
Figure 5.5. Let us consider the magnitude response of the ideal, desired, lowpass
digital filter to be HLP (ej ω ), in which the cutoff frequency is given as ωc . It has
a constant magnitude of one and zero phase over the frequency |ω| < ωc . From
(5.30), we get

π
ωc
1 j ω j nω 1
cLP (n) = HLP (e )e dω = ej nω dω
2π −π 2π −ωc
 j nω ωc
1 e  ej ωc − e−j ωc
=  =
2π j n −ωc 2j (πn)
sin(ωc n)
= = −∞ < n < ∞ (5.31)
πn
1
Two other types of frequency response for which the Fourier series coefficients have been derived
are those for the Hilbert transformer and the differentiator. Students interested in them may refer to
other textbooks.
262 FINITE IMPULSE RESPONSE FILTERS

HLP (e jw) HHP (e jw)

1 1

w w
−p −wc 0 wc p −p −wc 0 wc p
(a) (b)

HBP (e jw) HBS(e jw)

1 1

w w
−p −wc2 −wc1 wc1 wc2 p −p −wc2 −wc1 wc1 wc2 p
(c) (d)

Figure 5.5 Magnitude responses of four ideal filters. (Reprinted from Ref. 9, with per-
mission from John Wiley & Sons, Inc.)

Another form for the Fourier series coefficients is

sin(ωc n)  ωc 
cLP (n) = = sinc(ωc n); −∞ < n < ∞ (5.32)
πn π
Note that sinc(ωc n) = 1 when n = 0, so we find another way of listing the
coefficients as
⎧ ωc
⎪
⎨ π; n=0
cLP (n) = (5.33)
⎪
⎩ sin(ωc n) ; |n| > 0
πn
The Fourier series coefficients for the ideal HP, BP, and BS filter responses
shown in Figures 5.5b–d can be similarly derived as follows:
⎧ ωc
⎪
⎨1− π ; n=0
cHP (n) = (5.34)
⎪
⎩ − sin(ωc n) ; |n| > 0
πn
⎧
⎪ ω c2 − ωc1
⎨ ; n=0
cBP (n) = π (5.35)
⎪
⎩ 1 [sin(ωc2 n) − sin(ωc1 n)]; |n| > 0
πn
⎧
⎪ (ω − ωc1 )
⎨ 1 − c2 ; n=0
cBS (n) = π (5.36)
⎪
⎩ 1 [sin(ωc1 n) − sin(ωc2 n)]; |n| > 0
πn
 FOURIER SERIES METHOD MODIFIED BY WINDOWS 263

Continuing with the design of the lowpass filter, we choose the finite series
n=M −j nω
n=−M cLP (n)e = HM (ej ω ), which contains (2M
 + 1) coefficients from −M
to M, as an approximation to the infinite series n=∞ −j nω . In other
n=−∞ cLP (n)e
words, we approximate the ideal frequency response that exactly matches the given
HLP (ej ω ) containing the infinite number of coefficients by HM (ej ω ), which con-
tains a finite number of coefficients. As M increases, the finite series of HM (ej ω )
approximates the ideal response HLP (ej ω ) in the least mean-squares sense; that is,
the error defined as

π
1  
J (c, ω) = HM (ej ω ) − HLP (ej ω )2 dω (5.37)
2π −π

π  n=M   2

1  sin(ωc n) −j nω 
=  e − HLP (ej ω ) dω
2π −π  πn 
n=−M

attains a minimum at all frequencies, except at points of discontinuity.

We can make the error shown above as small as we like by choosing M
as large as we wish. As M increases, the number of ripples in the passband
(and the stopband) increases while the width between the frequencies at which
the maximum error occurs in the passband (0 ≤ ω ≤ ωc ) and in the stopband
(ωc ≤ ω ≤ π) decreases. In other words, as M increases, the maximum deviation
from the ideal value decreases except near the point of discontinuity, where the
error remains the same, however large the value of M we choose! The maximum
error or the overshoot from the ideal passband value or the stopband value is 11%
of the difference between the ideal passband value that is normalized to 1 and
the stopband value as shown in Figure 5.6. The magnitude response HM (ej ω )
is plotted for two different values of M in Figure 5.6, where Hid (ω) is the ideal
magnitude response of the lowpass filter as shown in Figure 5.5a.

5.3.1 Gibbs Phenomenon

These are some of the features of what is known as the “Gibbs phenomenon,”
which was mathematically derived by Gibbs. We explain it qualitatively as fol-
lows. The finite sequence c(n); −M ≤ n ≤ M can be considered as the result
of multiplying the infinite sequence c(n); −∞ ≤ n ≤ ∞ by a finite window
function:

1; −M ≤ n ≤ M
wR (n) = (5.38)
0; |n| ≥ M

n=M
sin{(2M + 1)ω/2}
(ej ω ) = e−j nω = (5.39)
sin(ω/2)
n=−M

So we have the product hw (n) = c(n) · wR (n), which is of finite length as shown
in Figure 5.7(c). Therefore the frequency response of the product of these two
264 FINITE IMPULSE RESPONSE FILTERS

1.1
1.0
0.9
0.8 Hid(w)
0.7
0.6
H(w)

0.5
0.4
0.3
0.2
0.1
0.0
−0.1
0 0.2π 0.4π 0.6π 0.8π π

Figure 5.6 Frequency response of a lowpass filter, showing Gibbs overshoot. (Reprinted
from Ref. 9, with permission from John Wiley & Sons, Inc.)

g(n)

(a) n
−M 0 M

wR (n)

(b) n
−M 0 M

hw (n)

(c) n
−M 0 M

(d ) n
0 M 2M

Figure 5.7 Coefficients of the FIR filter modified by a rectangular window function.
 FOURIER SERIES METHOD MODIFIED BY WINDOWS 265

functions is obtained from the convolution of (ej ω ) with the frequency response
HLP (ej ω ) of the ideal, desired, frequency response.

π
1
HM (e ) =
jω
HLP (ej ϕ )(ej (ω−ϕ) ) dϕ (5.40)
2π −π

The mainlobe of (ej ω ), centered at ω = 0, has a width defined by the first

zero crossings on either sides of ω = 0, which occur when [(2M + 1) ω2 ] = ±π,
that is, when ω = 2π(2M + 1) so that the width of the mainlobe is 4π/(2M + 1).
As M increases, the width of the mainlobe and the sidelobes decreases, giving
rise to more sidelobes or ripples in the same frequency band. At the same time,
the peak amplitudes of the mainlobe and the sidelobes increase such that the
area under each lobe remains constant. These features of (ej ω ) directly reflect
on the behavior of HM (ej ω ) when it is convolved with HLP (ej ϕ ). The effect
of convolution between HLP (ej ω ) and (ej ω ) is illustrated by looking at the
overlapping interval over which the product HLP (ej ϕ )(ej (ω−ϕ) ) is integrated,
for four different values of ω, in Figure 5.8. It is obvious that if the width of the
mainlobe is extremely narrow, the resulting HM (ej ω ) will have a sharp drop at
ω = ωc . If the number of sidelobes or their peak values in (ej ω ) increases, so
also will the number of ripples and the maximum error in HM (ej ω ).

Hid(q)
y(w − q) y(w − q)
w=p wc < w < p

−wc wc q −wc wc q
p −p p

y(w − q) y(w − q)
w = wc 0 < w < wc

q q
−wc wc p −p −wc wc p

H(w)

−wc wc p w

Figure 5.8 Convolution of the frequency response of a rectangular window with an ideal
filter. (Reprinted from Ref. 9, with permission from John Wiley & Sons, Inc.)
266 FINITE IMPULSE RESPONSE FILTERS

5.3.2 Use of Window Functions

In order to reduce the effects of the Gibbs phenomenon, some researchers have
proposed the use of tapered windows [11,12]; many others have proposed other
types of window functions. Only a few of the more popular window functions
are given below. Note that the number of coefficients generated by the window
functions given below is 2M + 1 = N + 1:

Bartlett window:2
|n|
w(n) = 1 − ; −M ≤ n ≤ M
M +1

Hann window:
  
1 2πn
w(n) = 1 + cos ; −M ≤ n ≤ M
2 2M + 1

Hamming window:
 
2πn
w(n) = 0.54 + 0.46 cos ; −M ≤ n ≤ M
2M + 1

Blackman window:
   
2πn 4πn
w(n) = 0.42 + 0.5 cos + 0.08 cos ; −M ≤ n ≤ M
2M + 1 2M + 1

The frequency responses of the window functions listed above have different
mainlobe widths ωM and different peak magnitudes of their sidelobes. In the
plot of HM (ej ω ) shown in Figure 5.9, it is seen that the difference between the
two frequencies at which the peak error in HM (ej ω ) occurs is denoted as ωM .
When the frequency response of the window functions is convolved with the
frequency response of the desired lowpass filter, the transition bandwidth of the
filter is determined by the width of the mainlobe of the window chosen and hence
is different for filters modified by the different window functions. The relative
sidelobe level Asl is defined as the difference in decibels between the magni-
tudes of the mainlobe of the window function chosen and the largest sidelobe.
It determines the maximum attenuation As = −20 log10 (δ) in the stopband of
the filter.
In Figure 5.9 we have also shown the transition bandwidth ω and the center
frequency ωc = (ωp + ωc )/2, where ωp and ωs are respectively the cutoff fre-
quencies of the passband and the stopband. The value of the ripple δ does not
depend on the length (2M + 1) of the filter or the cutoff frequency ωc of the

2
In many textbooks, the Bartlett window is also called a triangular window, but in MATLAB, the
Bartlett window is different from the triangular window.
 FOURIER SERIES METHOD MODIFIED BY WINDOWS 267

H(w)

1+d
Hid(w)
1−d

0.5 Δw

d
p w
−d wp wc
ws
ΔwM
y(q − we)

wc q

Figure 5.9 Frequency response of an ideal filter and final design. (Reprinted from Ref. 9,
with permission from John Wiley & Sons, Inc.)

filter. The width of the mainlobe ωM , the transition bandwidth ω, and the
relative sidelobe attenuation Asl for the few chosen window functions are listed in
Table 5.1. The last column lists the minimum attenuation As = −20 log10 δs real-
ized by the lowpass filters, using the corresponding window functions. It should
be pointed out that the numbers in Table 5.1 have been obtained by simulating
the performance of type I FIR filters with ωc = 0.4π and M = 128 [1], and they
would change if other types of filters and other values for ωc and M are chosen.
From Table 5.1, we see that as As increases, with fixed value for M, the transition
bandwidth ω also increases. Since we like to have a large value for As and a
small value for ω, we have to make a tradeoff between them. The choice of the
window function and the value for M are the only two freedoms that we have for
controlling the transition bandwidth ω, but the minimum stopband attenuation
As depends only on the window function we choose, and not the value of M.
Two window functions that provide control over both δs (hence As ) and the
width of the transition bandwidth ω are the Dolph–Chebyshev window [6] and
the Kaiser window functions [7], which have the additional parameters r and β,
respectively. The Kaiser window is defined by
5 " 6
I0 β 1 − (n/M)2
w(n) = ; −M ≤ n ≤ M (5.41)
I0 {β}
268 FINITE IMPULSE RESPONSE FILTERS

TABLE 5.1 Some Properties of Commonly Used Windows

Type of Window ωM ω Asl (dB) As (dB)

Rectangular 4π/(2M + 1) 0.92π/M 13 20.9

Bartlett 4π/(M + 1) —a 26.5 —a
Hann 8π/(2M + 1) 3.11π/M 31.5 43.9
Hamming 8π/(2M + 1) 3.32π/M 42.7 54.5
Blackman 12π/(2M + 1) 5.56π/M 58.1 75.3
a The frequency response of the Bartlett window decreases monotonically and therefore does not have
sidelobes. So the transition bandwidth and sidelobe attenuation cannot be found for this window.

where I0 {·} is the modified zero-order Bessel function. It is a power series of

the form
∞ 
2
(x/2)k
I0 {x} = 1 + (5.42)
k!
k=1

We compute the values of the Kaiser window function in three steps as follows:

• The parameter β required to achieve the desired attenuation αs =

−20 log10 (δs ) in the stopband is calculated from the following empirical
formula derived by Kaiser (the ripple in the passband is nearly the same
as δs ):
⎧
⎨ 0.1102(αs − 8.7) for αs > 50
β = 0.5842(αs − 21)0.4 + 0.07886(αs − 21) for 21 ≤ αs ≤ 50
⎩
0 for αs < 21
(5.43)
• Next the order of the filter N (=2M) is estimated from another empirical
formula derived by Kaiser:

(αs − 8)
N= (5.44)
2.285(ω)

where ω = ωs − ωp is the transition bandwidth as shown in Figure 5.9.

• The third step is to compute I0 {x}. In practice, adding a finite number of
terms, say, 20 terms of the infinite series, gives a sufficiently
" accurate value
for I0 {x}. The parameter x in the numerator represents β 1 − (n/M)2 in the
numerator of (5.41), so the value of x takes different values as n changes.

5.3.3 FIR Filter Design Procedures

The steps discussed in the design procedure for linear phase FIR filters are sum-
marized as follows:
 FOURIER SERIES METHOD MODIFIED BY WINDOWS 269

1. Depending on the nature of the magnitude response, we choose a value

for M and use (5.31), (5.34), (5.35), or (5.36) to compute the values of
the coefficients CLP (n), CHP (n), CBP (n), or CBS (n) for −M ≤ n ≤ M.
Then we choose a window function (Bartlett, Hamming, Hann, Kaiser,
or other window) and compute its values w(n) for −M ≤ n ≤ M. In the
case of Kaiser’s window, we find the value of M = N/2 from (5.44),
whereas he has derived a few other empirical formulas to estimate the
value of N , for designing FIR filters using other window functions.3 Note
that we have to choose the lowest even integer N greater than the value
calculated from (5.44), since Kaiser’s window is used for the design of
type I filters only.
2. Then we multiply the coefficients c(n) and w(n) to get the values of
hw (n). The filter with these finite numbers of coefficients has a frequency
response given by Hw (ej ω ) = hw (−M)ej ωM + hw (−M + 1)ej (−M+1)ω +
· · · + hw (1)ej ω + hw (0) + hw (1)e−j ω + · · · + hw (M)e−j Mω .
3. The next step is to multiply Hw (ej ω ) by e−j Mω , which is equivalent to
delaying the coefficients by M samples to get h(n) [i.e., hw (n − M) =
h(n)]. By delaying the product of c(n) and w(n) by M samples, we have
obtained a causal filter of finite length (N + 1) with coefficients h(n) for
0 ≤ n ≤ N.

The procedure becomes a little better understood by considering Figure 5.7

(where a rectangular window has been used). Since Hw (ej ω ) is a real function
of ω, its magnitude does not change
N when we multiply it by e−j Mω . Now we
−1 −n
have an FIR filter H (z ) = n=0 h(n)z  , which  is causal and is of length
(N + 1) and has the same magnitude as Hw (ej ω ). Its phase response is −Mω
with an additional angle of π radians when Hw (ej ω ) attains a negative real
value. Its group delay is a constant equal to M samples. This completes the
general procedure for designing an FIR filter that approximates the ideal magni-
tude response of a lowpass FIR filter; similar procedures are used for designing
highpass, bandpass, and bandstop filters. Let us illustrate this procedure by two
simple examples.

3
The formulas given by Kaiser may not give a robust estimate of the order for all cases of FIR filters.
A more reliable estimate is given by an empirical formula [10] shown below, and that formula is
used in the MATLAB function remezord:
 
(ωs − ωp ) 2
D∞ (δp , δs ) − F (δp , δs )
2π
N∼ =  
(ωs − ωp )
2π
  
where D∞ (δp , δs ) (when δp ≥ δs ) = a1 (log10 δp ) + a2 (log10 δp ) + a3 log10 δs − a4 (log10 δp )2 +
2
  
a5 (log10 δp ) + a6 , and F (δp , δs ) = b1 + b2 log10 δp − log10 δs , with a1 = 0.005309, a2 =
0.07114, a3 = −0.4761, a4 = 0.00266, a5 = 0.5941, a6 = 0.4278, b1 = 11.01217, b2 = 0.51244.
When δp < δs , they are interchanged in the expression for D∞ (δp , δs ) above.
270 FINITE IMPULSE RESPONSE FILTERS

Example 5.6

Design a bandpass filter that approximates the ideal magnitude response given in
Figure 5.5(c), in which ωc2 = 0.6π and ωc1 = 0.2π. Let us select a Hamming
window of length N = 11 and plot the magnitude response of the filter.
The coefficients cBP (n) of the Fourier series for the magnitude response given
are computed from formula (5.35) given below:
⎧
⎪ (ω − ωc1 )
⎨ c2 ; n=0
cBP (n) = sin(ωπ n) sin(ω n)
⎪
⎩ c2 c1
− ; |n| ≥ 0
πn πn
But since the Hamming window function has a length of 11, we need to compute
the coefficients cBP (n) also, from n = −5 to n = 5 only. So also we calculate
the 11 coefficients of the Hamming window, using the formula
 
2πn
wH (n) = 0.54 + 0.46 cos ; −5 ≤ n ≤ 5
N
Their products hw (n) = cBP (n)wH (n) are computed next. The 11 coefficients
cBP (n), wH (n) and hw (n) for −5 ≤ n ≤ 5 are listed below. Next the coefficients
hw (n) are delayed by five samples to get the coefficients of the FIR filter function
[i.e., h(n) = hw (n − 5)], and these are also listed for 0 ≤ n ≤ 10 below. The
plot of the four sequences and the magnitude response of the FIR are shown in
Figures 5.10 and 5.11, respectively.

cBP (n) = 0.00 0.0289 −0.1633 −0.2449 0.1156 0.400 0.1156

−0.2449 −0.1633 0.0289 0.000
wH (n) = 0.08 0.1679 0.0379 0.6821 0.9121 1.0 0.9121
0.6821 0.0379 0.1679 0.0800
hw (n) = 0.00 0.0049 −0.0650 −0.1671 0.1055 0.4000 0.1055
−0.1671 −0.0650 0.0049 0.000

h(n) 0.0000 0.0049 −0.0650 −0.1671 0.1055 0.4000

0.1055 −0.1671 −0.0650 0.0049 0.0000

Example 5.7

Design a lowpass FIR filter of length 11, with a cutoff frequency ωc = 0.3π.
Using a Hamming window, find  the value−nof the samples h(3) and h(9) of the
FIR filter given by H (z−1 ) = 10 n=0 h(n)z .
Since the length of the FIR filter is given as 11, its order is N = 10. The
coefficients hw (n) have to be known for −5 ≤ n ≤ 5 and delayed by five samples.
 FOURIER SERIES METHOD MODIFIED BY WINDOWS 271

Founer series coefficients of the BP filter Coefficients of the Hamming window

0.4 1
0.3
0.8
0.2
0.1 0.6
0
0.4
−0.1
0.2
−0.2
−0.3 0
−5 0 5 −5 0 5
n n

Coefficients of the noncausal BP filter Samples of the FIR BP filter

0.5 0.5
Value of the samples
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
−0.1 −0.1
−0.2 −0.2
−5 0 5 0 2 4 6 8 10
n n

Figure 5.10 Coefficients of the filter obtained during the design procedure.

Magnitude response of the BP filter

1.4

1.2

1
Magnitude

0.8

0.6

0.4

0.2

0
0 0.5 1 1.5 2 2.5 3 3.5
Normalized frequency

Figure 5.11 Frequency response of a bandpass filter.

272 FINITE IMPULSE RESPONSE FILTERS

1.5

1.0 1.0

0.5
0.2
0 wc1 wc2 wc3 wc4 p 0 w p wc p
(a) (b)

Figure 5.12 (a) Ideal magnitude response of a multilevel FIR filter; (b) Magnitude
response of a lowpass filter with a spline function of zero order.

Since only h(3) and h(9) are asked for, by looking at Figure 5.10, we notice that
these samples are the same as hw (−2) and hw (4) because when they are shifted
by five samples, they become h(3) and h(9). So we have to calculate only
cLP (−2), cLP (4) and the values w(−2), w(4) of the Hamming window. Then
hw (−2) = cLP (−2)w(−2) and hw (4) = cLP (4)w(4).
If the frequency response of an FIR filter has multilevel magnitude levels, it
is easy to extend the method as illustrated by Figure 5.12a. We design a lowpass
filter with a cutoff frequency ωc1 and a maximum magnitude of 0.8, another
lowpass filter with a cutoff frequency ωc2 , and a maximum magnitude of 0.2
in the passband; we design a highpass filter with a cutoff frequency ωc3 and a
maximum value of 0.5 and another bandpass filter with cutoff frequencies ωc3 and
ωc4 and a maximum magnitude of 1.0. If all of these filters are designed to have
zero phase or the same phase response, then the sum of the four filters described
above will approximate the magnitude levels over the different passbands. Each
of the four filters should be designed to have very low sidelobes so that they
don’t spill over too much into the passbands of the adjacent filter.
Even when we design an FIR filter having a constant magnitude over one
passband or one stopband, using the methods described above will produce a
transition band between the ideal passband and the stopband. Instead of mitigating
the Gibbs overshoot at points of discontinuity by using tapered windows, we
can make a modification to the ideal piecewise, constant magnitude response to
remove the discontinuities. We choose a spline function of order p ≥ 0 between
the passband and the stopband [5]. The spline function of zero order is a straight
line joining the edge of the passband and the stopband as shown in Figure 5.12b
The Fourier series coefficients for the lowpass frequency response in this are
given by
⎧
⎪ ωc
⎨ π;
⎪ n=0
hLP (n) =     (5.45)
⎪ 2 sin(ωn/2) sin(ωc n)
⎪
⎩ ; |n| > 0
(ωn) πn

where ω = ωs − ωp and ωc = [(ωs + ωp )/2].

 DESIGN OF WINDOWED FIR FILTERS USING MATLAB 273

A smoother transition is achieved when we choose a higher-order spline func-

tion (e.g., p = 2,3,4). In that case, the Fourier series coefficients are given by
⎧ ωc
⎪
⎨ π; n=0
hLP (n) =  sin(ω)n/2p p  sin(ω n)  (5.46)
⎪
⎩ c
; |n| > 0
(ω)n/2p πn

Design procedure using this formula seems easier than the Fourier series method
using window functions—since we do not have to compute the coefficients of
window functions and multiply these coefficients by those of the ideal frequency
response. But it is applicable for the design of lowpass filters only. However,
extensive simulation of this design procedure shows that as the bandwidth ω
is decreased and as p is increased, the magnitude response of the filter exhibits
ripples in the passband as well as the stopband, and it is not much better than
the response we can obtain from the windowed FIR filters.

5.4 DESIGN OF WINDOWED FIR FILTERS USING MATLAB

5.4.1 Estimation of Filter Order

In the discussion of the Fourier series method, it was pointed out that we had a
choice only between the windows (Bartlett, Hamming, Hann, etc.) and the order
N = 2M of the filter. There is no guideline for choosing the type of window
or the value for N ; in other words, they are chosen arbitrarily on a trial-and-
error basis until the specifications are satisfied. But in the case of Kaiser and
Dolph–Chebyshev windows, we have an empirical formula to estimate the order
N that achieves a desired stopband attenuation αs . However, it was pointed
out earlier that some authors have derived empirical formulas for estimating the
order N even when windows like those mentioned above are chosen. We use
the MATLAB M-file kaiserord (available in the MATLAB Signal Processing
Toolbox) to estimate the order N of the filter using the Kaiser window. We will,
however, use the M-file remezord to estimate the order of the filter using the
other windows. After the order of the filter has been obtained, the next step in
the design procedure is to find the values of the unit impulse response h(n) of
the filter that are the same as the coefficients of the FIR filter transfer function.
The MATLAB M-file fir1 is used for designing filters with piecewise constant
magnitudes discussed above, and fir2 is the M-file used for arbitrary magnitude
specifications. In the following examples, note that N is the order of the FIR filter
that therefore has N + 1 coefficients but N is the number of coefficients in such
MATLAB functions such as hamming used for computing the window functions!

Example 5.8

If the magnitude response specified in the passband of an FIR filter lies between
1 + δp and 1 − δp , then the maximum attenuation αp ( in decibels) in the passband
274 FINITE IMPULSE RESPONSE FILTERS

is 20 log 10(1 + δp ) − 20 log 10(1 − δp ) = 20 log[(1 + δp )/(1 − δp )]. Solving for

δp in this case, we get

100.05αp − 1
δp =
100.05αp + 1

If the passband magnitude lies between 1 and (1 − δp ), then the maximum atten-
uation in the passband αp = −20 log(1 − δp ), in which case δp is given by
(1 − 10−0.05αp ). If the magnitude in the stopband lies below δs , the minimum
attenuation in the stopband is given by αs = −20 log(δs ), from which we obtain
δs = 10−0.05αs . These relations are used to find the value of δp and δs if the
attenuations αp and αs in the passband and stopband are specified in decibels.
In the MATLAB function [N, fpoints, magpoints,wt] = remezord
(edgepoints, bandmag, dev, Fs), the input vector edgepoints lists the
edges of the disjoint bands between 0 and the Nyquist frequency but does not
include the frequency at ω = 0 and the Nyquist frequency, as the default value of
the Nyquist frequency is 1.0 (and therefore the sampling frequency Fs=2). The
vector bandmag lists the magnitudes over each of the passbands and stopbands.
If there is transition band between the passband and stopband, it is considered
as a “don’t care” region. Since the first edge at 0 and the last one at the Nyquist
frequency are not included in the vector edgepoints, the length of the vector
edgepoints is two times that of bandmag minus 2. For example, let us choose
a bandpass filter with a stopband [0 0.1], a transition band [0.1 0.12], a pass-
band [0.12 0.3], a transition band [0.3 0.32], and a stopband [0.32 1.0].
The input vector edgepoints and the output vector fpoints are the same,
when Fs=2, namely, [0.1 0.12 0.3 0.32]. The input vector bandmag is of
length 3, and the values may be chosen, for example, as [0 1 0] for the band-
pass filter. The vector dev lists the values for the maximum deviations δp and
δs in the passbands and stopbands, calculated from the specifications for αp and
αs as explained above. The output vector fpoints is the same as edgepoints
when Fs has the default value of 2; if the edgepoints and the sampling fre-
quency Fs are actual frequencies in hertz, then the output vector fpoints gives
their values normalized by the actual Nyquist frequency Fs/2. But it must be
pointed out that the output vector magpoints lists the magnitudes at both ends
of the passbands and stopbands. In the above example, the vector magpoints
is [0 0 1 1 0 0]. The output of this function is used as the input data to
fir1 and (also the function remez, discussed later) to obtain the unit impulse
response coefficients of the FIR filter.
Let us consider a lowpass filter with a passband over [0 0.3] and a magnitude
1.0 and a stopband over [0.4 1.0] with a magnitude 0.0. In this case, there is
transition band between 0.3 and 0.4 over which the magnitude is not specified,
and therefore it is a “don’t care” region. The vector edgepoints is [0.3 0.4],
and the vector bandmag is [1.0 0.0]. For the previous example of a bandpass
filter, we have already mentioned that edgepoints is [0.1 0.12 0.3 0.32]
and the vector bandmag is [0 1 0] . Let us select δp = δs = 0.01 for both
 DESIGN OF WINDOWED FIR FILTERS USING MATLAB 275

filters, namely, a log magnitude of αp = −20 log(1 − δp ) = 0.087 dB for the

ripple in the passband and a gain of −40 dB in the stopband.
The function remezord for estimating the order of the lowpass filter is as
follows:

[N, fpoints, magpoints, wt] = remezord([0.3 0.4], [1.0 0.0],

[0.01 0.01], 2)

and it yields a value N = 39, with the same vector fpoints as bandmag and the
vector wt=[1.0 1.0]. If we choose a sampling frequency of 2000 Hz, we use
remezord([0.3 0.4], [1.0 0.0], [0.01 0.01], 2000), and the output
would be N = 39, fpoints=[0.3 0.4], magpoints=[1 1 0 0] and the vec-
tor wt = [1.0 1.0]. The elements in the vector wt will be unequal if δp = δs
[i.e., wt = [(δs /δp ) 1]]. These output values are used as input in fir1 (or
remez) for the design of the lowpass filter.
The function for estimating the order of the bandpass filter is

[N, fpoints, magpoints, wt]=remezord([0.1 0.12 0.3 0.32],

[0.0 1.0 0.0], [0.01 0.01 0.01], 2).

which gives N = 195, with the same vector fpoints as the input vector edge-
points, magpoints= [0 0 1 1 0 0], and wt = [1 1 1] as in the input.
The MATLAB function kaiserord given below is used to estimate the order
N of the FIR filter using the Kaiser window. The input parameters for this
function are the same as for remezord, but the outputs are the approximate order
N of the Kaiser window required to meet the input specifications, the normalized
frequencies at the bandedges, the parameter beta, and the filtertype:

[N, Wc, Beta, ftype]=kaiserord(edgepoints, bandmag, dev, Fs)

For the lowpass filter specified above, we use

[N, Wc, Beta, ftype] =kaiserord([0.3 0.4], [1.0 0.0],

[0.01 0.01], 2)

and we get N=45, Wc=0.35, Beta=3.3953.

For the bandpass filter, with the same input parameters as those used in reme-
zord, we get the output parameters as N=224, Wc=[0.11 0.31], Beta=3.3953,
and ftype=DC-0. When ftype=DC-0, this means that the first band is a stopband
and when it is DC-1, it indicates that the first band is a passband.

5.4.2 Design of the FIR Filter

After we have found the order of the filter and any other parameters as the output
of the functions remezord and kaiserord, we use the function fir1, which
takes various forms as described below:
276 FINITE IMPULSE RESPONSE FILTERS

b=fir1(N,Wc)
b=fir1(N,Wc,’ftype’)
b=fir1(N,Wc,’ftype’,window)

These forms give the N + 1 samples of the unit impulse response of the lin-
ear phase FIR filter or the coefficients of its transfer function. Wc is the cutoff
frequency of the lowpass filter, and ftype is omitted; it is the cutoff of the high-
pass filter when ’ftype’ is typed as ’high’. But Wc is a two-element vector
Wc=[W1 W2], which lists the two cutoff frequencies, ωc1 and ωc2 (ωc2 ≥ ωc1 ),
of the bandpass filter. (Use help fir1 to get details when there are multiple
passbands.) The term ’ftype’ need not be typed. When ’ftype’ is typed as
’stop’, the vector Wc represents the cutoff frequencies of the stopband filter.
If the filter is a lowpass, it becomes a type I filter when N as obtained from
remezord is even and it is type II filter when N is odd. Note that the frequency
response of type II filters has a zero magnitude at the Nyquist frequency, that is,
their transfer function has a zero at z = −1 and therefore is a polynomial of odd
order. The highpass and bandstop filters that do not have a zero magnitude at the
Nyquist frequency cannot be realized as type II filters. When designing a highpass
or bandstop filter, N must be an even integer, and the function fir1 automatically
increases the value of N by 1 to make it an even number if the output from
remezord is an odd integer. Since the program assumes real values for the
magnitude and zero value for the phase, we do not get types III and IV filters
from this type of frequency specification. The window by default is the Hamming
window in fir1, but we can choose the rectangular (boxcar), Bartlett, triangular,
Hamming, Hanning, Kaiser, and Dolph–Chebyshev (chebwin) windows in the
function fir1. After getting the coefficients of the FIR filter, we can find the
magnitude (phase and group delay also) of the filter to verify that it meets the
specifications; otherwise we may have to increase the value of N , or change the
values in the vector dev.

Example 5.9

Now that we have obtained all the input data needed to design a LP filter with
N = 39 and ωc = 0.3 and a BP filter with N = 195 and ωc = [0.12 0.3], we
design the FIR filters with the Hamming window and the Kaiser window. So we
have four cases, discussed below.
The M-files for designing the four filters are given below, and the resulting
magnitude responses are shown in Figures 5.13–5.17.
It must be noted from Figure 5.9 that the magnitude of the filter designed by
the Fourier series method is 0.5 at ωc , whatever the window function used to
minimize the Gibbs overshoot. The order N = 39 for the lowpass filter, obtained
from the function remezord, is only an estimate that is very conservative because
it results in the magnitude response of the filter that does not meet the passband
error δp = 0.01 specified and used in that function. So we have to change the
value for the cutoff frequency ωc and the order N of the filter by trial and error
 DESIGN OF WINDOWED FIR FILTERS USING MATLAB 277

FIR LP filter response with N = 85 and Hamming window

20

−20
Magnitude in dB

−40

−60

−80

−100

−120
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency

Figure 5.13 Magnitude response of a FIR lowpass filter using Hamming window.

Magnified plot of the LP filter response

1.01

0.99
Magnitude

0.98

0.97

0.96

0.95

0.94
0 0.05 0.1 0.15 0.2 0.25 0.3
Normalized frequency

Figure 5.14 Magnified frequency response of a FIR lowpass filter in the passband.

until the specifications are met. For the lowpass FIR filter, we have had to choose
ωc = 0.35 and N = 65 so that at the frequency ω = 0.3, the error δp ≤ 0.01 and
at ω = 0.4, the error δs ≤ 0.01 (equal to 40 dB). The magnitude response of this
final design is shown in Figures 5.13 and 5.14. Similar changes in the design of
the other filters designed by the Fourier series method are necessary.
278 FINITE IMPULSE RESPONSE FILTERS

FIR BP filter design using Hamming window

0

−20
Magnitude of the filter in dB

−40

−60

−80

−100

−120
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency W/π

Figure 5.15 Magnitude response of a FIR bandpass filter using Hamming window.

Magnitude of LP FIR filter N = 45 with Kaiser window

20

0
Magnitude response

−20

−40

−60

−80

−100
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency W/p

Figure 5.16 Magnitude response of a FIR lowpass filter using Kaiser window.

%Case 1:Design of the LP filter with N=39 and Hamming window

b1=fir1(39, 0.3);
[h1,w]=freqz(b1, 1, 256);
H1db=20*log10(abs(h));
plot(w/pi, H1db);grid
 DESIGN OF WINDOWED FIR FILTERS USING MATLAB 279

Magnitude of BP filter N = 224 with Kaiser window

20

−20
Magnitude response

−40

−60

−80

−100

−120
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency W/p

Figure 5.17 Magnitude response of a FIR bandpass filter using Kaiser window.

title(’Magnitude response of FIR LP filter with N=39 and

Hamming window’);
ylabel(’Magnitude in dB’)
xlabel(’Normalized frequency’)

%Case 2:Design of the BP filter with N=195 and Haming window

b2=fir1(195, [0.12 0.31]);
[h2,w]=freqz(b2, 1, 256);
figure
H2db=20*log10(abs(h));
title(’FIR BP filter with N=195 and Hamming window’)
ylabel(’Magnitude in dB’)
xlabel(’Normalized frequency’)

%Case 3:Design of the LP Filter using the Kaiser window:

%The length of Kaiser window must be one higher than the
% order(N=45) of the FIR filter
%obtained from the function Kaiserord
b3=fir1(45, 0.35, kaiser(46, 3.3953));
[h3,w]=freqz(b3,1,256);
H3db=20*log10(abs(h3));
plot(w/pi,H3db);
title(’Magnitude of LP with N=45 and Kaiser window’)
ylabel(’Magnitude in dB’)
xlabel(’Normalized frequency’).
280 FINITE IMPULSE RESPONSE FILTERS

%Case 4:Design of the BP Filter using the Kaiser window

b4=fir1(224, [0.11 0.31], kaiser(225, 3.3953));
[h4,w]=freqz(b4, 1, 256);
H4=abs(h4);
H4db=20*log10(H4);
figure
plot(w/pi,H4db);grid
title(’Magnitude of BP filter N = 224 with Kaiser window’)
ylabel(’Magnitude in dB’);
xlabel(’Normalized frequency ’)

5.5 EQUIRIPPLE LINEAR PHASE FIR FILTERS

The frequency response in the passband of FIR filters designed by using the mod-
ified Fourier series method as described above has a monotonically decreasing
response and a maximum error from the desired ideal response in the passband, at
the cutoff frequency ωc . Now we discuss another important method that “spreads
out” the error over the passband in an equiripple fashion, such that the maximum
error is the same at several points and can be made very small. This method min-
imizes the maximum error in the passband and is called as the minimax design
or the equiripple design. An example of the equiripple or Chebyshev response
of a lowpass filter is shown in Figure 5.18.

Magnitude response of an equiripple FIR filter

1.4

1.2

1
Magnitude

0.8

0.6

0.4

0.2

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency

Figure 5.18 Magnitude response of an equiripple lowpass filter.

 EQUIRIPPLE LINEAR PHASE FIR FILTERS 281

For types 1–4 FIR filters, the frequency response were shown in (5.22) to be
of the following form:
⎧ ⎫
⎨ N  N/2 
N
 ⎬
H (ej ω ) = e−j [(N/2)ω] h +2 h − n cos(nω)
⎩ 2 2 ⎭
n=1

for type I (5.47)

⎧ ⎫
(N +1)/2 
⎨    ⎬
N +1 1
H (e−j ω ) = e−j [(N/2)ω] 2 h − n cos n− ω
⎩ 2 2 ⎭
n=1

for type II (5.48)

⎧ ⎫
⎨ N/2   ⎬
N
H (e−j ω ) = e−j [(Nω−π)/2] 2 h − n sin(nω)
⎩ 2 ⎭
n=1

for type III (5.49)

⎧ ⎫
(N +1)/2 
⎨    ⎬
N + 1 1
H (e−j ω ) = e−j [(Nω−π)/2] 2 h − n sin n− ω
⎩ 2 2 ⎭
n=1

for type IV (5.50)

In general, Equations (5.47)–(5.50) are of the form H (ej ω ) = e−j (N ω/2) ejβ ×
HR (ω), where4 β is either 0 or π/2 depending on the type of filter, and HR (ω)
is a real function of ω, which can have positive or negative values. It is easy to
see that HR (ω) for type I filters can be reduced to the form (2M = N )

M
HR (ω) = 7
a [k] cos(kω) (5.51)
k=0

where
7
a [0] = h[M],7
a [k] = 2h[M − k], 1≤k≤M (5.52)

Consider HR (ω) for the type II filter shown in (5.48) and given below:
(N +1)/2
    
N +1 1
HR (ω) = 2 h − n cos n− ω
2 2
n=1

This can be reduced to the form

(2M+1)/2   
1
HR (ω) = b[k] cos k− ω (5.53)
2
k=1

4
This is not the same parameter β that is used in Kaiser’s window.
282 FINITE IMPULSE RESPONSE FILTERS

where b[k] = 2h{[(2M + 1)/2] − k}, 1 ≤ k ≤ [(2M + 1)/2]. This can be fur-
ther reduced to the form
 ω  (2M−1)/2

HR (ω) = cos 7
b[k] cos(kω) (5.54)
2
k=0

where
1 7
b[1] = b[1] + 27
b[0]
2
1 7 2M − 1
b[k] = b[k] + 7
b(k − 1) , 2≤k≤ (5.55)
2 2
   
2M + 1 1 7 2M − 1
b = b
2 2 2

Let us consider the function HR (ω) for a type III filter. Equation (5.49) can be
reduced to the form

(2M+1)/2
HR (ω) = c[k] sin(kω) (5.56)
k=1

where c[k] = 2h[M − k], 1 ≤ k ≤ M. This can be reduced to the form

M−1
HR (ω) = sin(ω) 7
c[k] cos(kω) (5.57)
k=0

where

c[1] = 7c[0] − 127
c[1]
c[k] = 1
2 c[k − 1] − 7
(7 c[k]) , 2≤k ≤M −1 (5.58)
c[M] = 12 7
c[M − 1]

Finally we express HR (ω) for a type IV filter as

(2M+1)/2

HR (ω) = d[k] sin k − 12 ω (5.59)
k=1

where d[k] = 2h{[(2M + 1)/2] − k}, 1 ≤ k ≤ (2M + 1)/2. Equation (5.59) can
be reduced to the form
 ω  (2M−1)/2

HR (ω) = sin 7 cos(kω)
d[k] (5.60)
2
k=0
 EQUIRIPPLE LINEAR PHASE FIR FILTERS 283

where 
7 − 1 d[1]
d[1] = d[0] 7
2

7 − 1] − d[k]
d[k] = 12 d[k 7 , 2≤k≤ 2M−1
2
   
2M + 1 2M − 1
d = d7
2 2

Note that we can express the coefficients 7 a [.], 7

b[.], 7 7 in terms of a[.],
c[.], d[.]
b[.], c[.], d[.] since they are linearly related. We express Equations (5.51), (5.54),
(5.57), and (5.60) in a common form, (5.61), in order to develop a common
algorithm that obtains a minimax approximation for all four types of filters

HR (ω) = Q(ω)P (ω) (5.61)

where ⎧
⎪
⎪ 1 for type I
⎨ cos  ω
⎪
for type II
Q(ω) = 2 (5.62)
⎪ sin(ω)
⎪ for type III
⎪
⎩ 
sin ω2 for type IV

and

K
P (ω) = α[k] cos(kω) (5.63)
k=0

where ⎧
⎪
⎪ 7
a [k] for type I
⎪
⎨7b[k] for type II
α[k] = (5.64)
⎪
⎪ 7
c[k] for type III
⎪
⎩7
d[k] for type IV

and ⎧
⎪ M for type I
⎪
⎪
⎪
⎪ 2M − 1
⎨ for type II
K= 2 (5.65)
⎪
⎪ M −1 for type III
⎪
⎪
⎩ 2M − 1
⎪
for type IV
2

We define a weighted error function

   
J (ω) = W (ej ω ) HR (ω) − Hd (ej ω ) = W (ej ω ) Q(ω)P (ω) − Hd (ej ω )
 
Hd (ej ω )
= W (e )Q(ω) P (ω) −
jω
(5.66)
Q(ej ω )
284 FINITE IMPULSE RESPONSE FILTERS

Using the notations W (ej ω )Q(ω) = W 7 (ej ω ) and Hd (ej ω )/Q(ej ω ) = H

7d (ej ω ),
this equation is now rewritten in another form as
 
J (ω) = W 7 (ej ω ) P (ω) − H 7d (ej ω ) (5.67)

In these equations, Hd (ej ω ) is the desired frequency response normally specified

over subintervals of 0 ≤ ω ≤ π and W (ej ω ) is a weighting function chosen by the
designer to emphasize relative magnitude of the error over different subintervals.
In usual filter design applications, the desired magnitude is given as

1 ± δp in the passband(s)
Hd (ej ω ) = (5.68)
0 + δs in the stopband(s)

The weighting function W (ej ω ) can be chosen as

⎧
⎨1 in the passband(s)
W (ej ω ) = δp
⎩ in the stopband(s)
δs
or as
⎧
⎨ δs in the passband(s)
W (ej ω ) = δp
⎩
1 in the stopband(s)

The coefficients α[k] in (5.67) are the unknown variables that have to be found
such that the maximum absolute value of the error |J (ω)| over the subintervals
of 0 ≤ ω ≤ π is minimized. It has been shown [13] that when this minimum
value is achieved, the frequency response exhibits an equiripple behavior:
min max   
W 7d (ej ω ) 
7 (ej ω ) P (ω) − H (5.69)
over {α[k]} over {S}

where {S} is used to denote the union of the disjoint frequency bands in 0 ≤
ω ≤ π.
Once these coefficients are determined, the coefficients h(n) can be obtained
from the inverse relationships between α[k] and a[k], b[k], c[k], and d[k] depend-
ing on the type of filter and then using the relationship between h[n] and these
coefficients.
Parks and McClellan [2] were the original authors who solved the preceding
problem of minimizing the maximum absolute value of the error function J (ω),
using the theory of Chebyshev approximation, and developed an algorithm to
implement it by using a scheme called the Remez exchange algorithm. They also
published a computer program (in FORTRAN) for designing equiripple, linear
phase FIR filters. Although major improvements have been made by others to this
algorithm and to the software [13], it is still referred to as the Parks–McClellan
algorithm or the Remez exchange algorithm. We will work out a few examples of
designing such filters using the MATLAB function remez in the following section.
 DESIGN OF EQUIRIPPLE FIR FILTERS USING MATLAB 285

5.6 DESIGN OF EQUIRIPPLE FIR FILTERS USING MATLAB

The first step is to estimate the order of the FIR filter, using the function reme-
zord which was explained earlier. The next step is to find the coefficients of the
FIR filter using the function remez, which has several options:

b=remez(N, fpoints, magpoints)

b=remez(N, fpoints, magpoints, wt)
b=remez(N, fpoints, magpoints, ’ftype’)
b=remez(N, fpoints, magpoints, wt, ’ftype’)

The vector fpoints lists the edges of the passbands and stopbands, starting
from ω = 0 and ending with ω = 1 (which is the normalized Nyquist frequency).
In contrast to the function remezord, this vector fpoints includes 0 and 1.0
as the first and last entries. The edges between the passbands and the adjacent
stopband must have a separation of at least 0.1; otherwise the program automat-
ically creates a transition band of 0.1 between them, and these transition bands
are considered as “don’t care” regions. The vector magpoints lists the mag-
nitudes in the frequency response at each edge of the passband and stopband.
The weighting function can be prescribed for each frequency band as explained
above. The function remez chooses type I filters for even-order N and type II
filters for odd order as the default choice. The flags ’hilbert’ and ’differ-
entiator’ are used for the option ftype for designing the Hilbert transformer
and the differentiator, respectively. The other input variables are the same as
the outputs obtained from remezord, and hence we can use the two functions
remezord and remez together in one M-file to design an equiripple FIR filter
with linear phase as listed and described below.

5.6.1 Use of MATLAB Program to Design Equiripple FIR Filters

Example 5.10

%Program to obtain the unit impulse response coefficients

% of an equiripple, FIR filter with a linear phase
edgepoints= input(’Type in the edge frequencies of
each band =’);
%Type in normalized edge frequencies between 0 and 1,
% excluding 0 and 1, when Fs=2
%or actual frequencies in Hz and choose FT also in Hz
Fs= input(’Sampling frequency in Hz =’)
%The bandmag is the magnitude in each band from 0 to 1
bandmag =input(’Type in the magnitude values for each
passband and stopband =’);
% There must be one value in dev and in wt for each band
dev=input(’Desired ripple in each band =’);
wt=input(’Type in values for the relative weights in each
286 FINITE IMPULSE RESPONSE FILTERS

band=’);
[N, fpoints, magpoints, wt]=remezord(edgepoints, bandmag,
dev, Fs);
disp(’Order of the FIR filter is’);disp(N);
b=remez(N, fpoints, magpoints, wt);
[h, w]=freqz(b, 1, 256);
H=abs(h);
Hdb=20*log10(H);
plot(w/pi,Hdb);grid
title(’Magnitude response of the equiripple, linear phase
FIR filter’)
ylabel(’Magnitude in dB’)
xlabel(’Normalized frequency’)

This program can be used to design lowpass, highpass, bandpass, and bandstop
filters. If the filter does not meet the given specification, one should increase
the order of the filter by 1, 2, or 3 until the specifications in the passbands and
stopbands are met. But when the cutoff frequencies are very close to 0 or 1, and
when we are designing highpass and stopband filters, the value of N estimated
by remezord may not be acceptable, and we may have to choose it arbitrarily
to meet the given specifications. If one is interested in getting an enlarged view
of the magnitude response over a frequency range such as the passband, then
the following lines may be added to the program listed above. In the function
axis, choose wc1 = 0 and wc2 = passband edge frequency for the lowpass filter
and wc1 and wc2 as the lower and upper cutoff frequencies of the passband of a
bandpass filter or the stopband of a bandstop filter:

%Add the following lines to get an enlarged view of the

% magnitude response
figure
wc1=input(’Type in the lower cutoff frequency =’)
wc2=input(’Type in the upper cutoff frequency=’)
plot(w/pi,Hdb);Grid
axis([wc1 wc2 -0.5 0.5]);
title(’Enlarged plot of the magnitude response’)
ylabel(’Magnitude in dB’)
xlabel(’Normalized frequency’)

Example 5.11

Let us work a few examples using the program displayed above for designing
lowpass and bandpass equiripple filters with the same specifications as the earlier
ones. We type in the following input data for designing an equiripple lowpass
filter:

edgepoints:[0.3 0.4]
 DESIGN OF EQUIRIPPLE FIR FILTERS USING MATLAB 287

bandmag :[1 0]
dev :[0.01 0.01]
wt :[1 1]

The frequency response of the lowpass filter is shown in Figure 5.19, and a
magnified view of the response in the passband is shown in Figure 5.20. It is
seen that the deviation in the passband is within 0.087 dB, which is equivalent
to δp = 0.01 in the passband, but the magnitude in the stopband is not equal to
or less than −40 dB, corresponding to δs = 0.01.
Therefore we increase the value of N from 39 to 41 and show the resulting
filter response displayed in Figure 5.21, which does meet the stopband magnitude
required.
Next we design an equiripple bandpass filter meeting the same specifications
as those that used the Hamming window and the Kaiser window. The input
parameter values are the following:

edgepoints :[0.1 0.12 0.3 0.32]

bandmag :[0 1 0]
dev :[0.01 0.01 0.01]
wt :[1 1 1]

The magnitude response of the equiripple filter is shown in Figure 5.22.

Magnitude response of an equiripple, linear phase FIR filter

10

−10

−20
Magnitude in dB

N = 30
−30

−40

−50

−60

−70

−80
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency

Figure 5.19 Magnitude response of an equiripple FIR lowpass filter with N = 39.
288 FINITE IMPULSE RESPONSE FILTERS

Expanded magnitude response in the passband

0.1

0.08

0.06

0.04
Magnitude in dB

0.02

−0.02

−0.04

−0.06

−0.08

−0.1
0 0.05 0.1 0.15 0.2 0.25 0.3
Normalized frequency

Figure 5.20 Magnified plot of the passband response of an equiripple FIR lowpass filter.

Magnitude response of an equiripple, Lowpass, FIR filter

10

−10

−20
Magnitude in dB

N = 41
−30

−40

−50

−60

−70

−80
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency

Figure 5.21 Magnitude response of an equiripple FIR lowpass filter with N = 41.

When we design an equiripple bandstop filter with the same vector edgepoints
as in the preceding bandstop filter, we get a response that does not meet the
desired specifications even after the order of the FIR filter is increased from 195
to 205. Also, the design of a highpass filter using the Remez algorithm is not
always successful. In Section 5.7, an alternative approach to solve such problems
is suggested.
 FREQUENCY SAMPLING METHOD 289

Magnitude response of an equiripple, linear phase FIR filter

10

−10
N = 195
−20
Magnitude in dB

−30

−40

−50

−60

−70

−80
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency

Figure 5.22 Magnitude response of an equiripple FIR bandpass filter.

It should also be remembered that the Remez algorithm is restricted to filters

of order greater than 3. It is interesting to note that the bandpass filter of order
195, using the Fourier series and the Hamming window achieves a higher stop-
band attenuation than does a bandpass filter of the same length, using the Remez
algorithm.

5.7 FREQUENCY SAMPLING METHOD

In the methods considered above for the design of linear phase FIR filters, the
magnitude response was specified as constant over disjoint bands, and the transi-
tion bands were “don’t care” regions. In this section, we discuss briefly the MAT-
LAB function fir2 that designs a linear phase filter with multistage magnitudes

b=fir2(N, F, M)
b=fir2(N, F, M, window)
b=fir2(N, F, M, window, npt)

and so on. As input parameters for this function, N is the order of the filter and
F is the vector of frequencies between 0 and 1 at which the magnitudes are
specified. In the vector F, we include the end frequencies 0 and 1 and list the
magnitudes at these frequencies in the vector M, so the lengths of F and M are the
same. The argument npt is the number of gridpoints equally spaced between 0
and 1; the default value is 512. Frequencies at the edge of adjacent bands can be
included and will appear twice in the vector F indicating a jump discontinuity.
290 FINITE IMPULSE RESPONSE FILTERS

The output of this function is the N + 1 coefficients of the unit impulse response
of the FIR filter with linear phase.

Example 5.12

%Program to design an FIR fiter using fir2

F =[0 0.25 0.25 0.5 0.65 0.75 0.9 1.0];
M = [0 1 0.5 0.5 0 0 0.3 0.3];
b=fir2(128, F, M);
[h,w]=freqz(b, 1, 256);
H=abs(h);
plot(w/pi,H);Grid;
title(’’Multistage Magnitude Response of an FIR linear
phase filter’)
ylabel(’Magnitude’)
xlabel(’Normalized frequency’)
%end

The magnitude response of this filter is shown in Figure 5.23.

The function fir2 utilizes the relationship between the DFT and IDFT of an
FIR filter. Note that we prescribed the magnitude of the filter response over the
frequency range 0–1.0 and the function assumes that the samples of the DFT
are symmetric about the Nyquist frequency and extrapolates the samples up to
the sampling frequency before finding the inverse DFT. We also prescribe zero
value for the phase, and therefore the function is restricted to types I and II FIR

Magnitude of a multistage linear phase FIR filter

0.9

0.8

0.7

0.6
Magnitude

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency

Figure 5.23 Magnitude response of a multistage FIR filter.

 FREQUENCY SAMPLING METHOD 291

filters only. To design FIR filters with linear phase in general, let us first review
the following results for the DTFT of a digital filter
 
H (ej ω ) = H ∗ (e−j ω ) = H (ej ω ) e−j (Mω−β) 

where M = (N − 1)/2 and β = 0 for types I and II filters and M = (N − 1)/2

and β = ±(π/2) for types III and IV filters.
Next, be aware that in the notation commonly used for the DFT-IDFT pair, N
is the length of the FIR filter or the number of unit impulse response coefficients
h(n) and is also equal to the number of DFT samples H (k) as shown in the
following relationships:

N −1
H (k) = H (ej (2π/N )k ) = h(n)e−j (2π/N )kn (5.70)
n=0
 
= H (ej (2π/N )k ) e−j [M(2π/N )k−β] for k = 0, 1, 2, . . . , (N − 1)

1
N−1
h(n) = H (k)ej (2π/N )nk for n = 0 . . . (N − 1) (5.71)
N
k=0

For an FIR filters with linear phase, h(n) = ±h(N − 1 − n), n =

0, 1, 2, . . . , (N − 1), where the plus sign is used for types I and II linear phase
filters and the minus sign is used for types III and IV linear phase filters.
The DFT H (k) must satisfy the following conditions:

H (0) for k = 0
H (k) =
H (N − k) for k = 1, 2, . . . , (N − 1)
⎧    8 9
⎪
⎪ N −1 2πk N −1
⎪
⎪ − for k = 0, 1, . . . ,
⎪
⎪ 2 N 2
⎪
⎪
⎪
⎨  and types I and II filters
  8 9
H (k) = N −1 2π(N − k) N −1
⎪
⎪ + for k =
⎪
⎪ 2 N 2
⎪
⎪
⎪
⎪ +1, . . . (N − 1)
⎪
⎩ and types I and II filters
⎧    
⎪
⎪ π N −1 2πk
⎪
⎪ ± − for k = 0, 1, . . . ,
⎪
⎪ 2 2 N 8 9
⎪
⎪ N −1
⎪
⎪
⎪
⎪
⎨ 2
= and types 8 III and9IV filters
⎪  π   N − 1   2π(N − k) 
⎪ N −1
⎪
⎪
⎪− ±
⎪ + for k =
⎪
⎪ 2 2 N 2
⎪
⎪
⎪
⎪ +1, . . . , (N − 1) and
⎩ types III and IV filters
292 FINITE IMPULSE RESPONSE FILTERS

These relationships provide us with a general method for finding the unit
sample response of an FIR filter with linear  phase, from the values of the
DFT samples that are the values of H (ej ω ) e−j (Mω−β) at the discrete fre-
quencies ωk = (2π/N )k. Therefore we can prescribe both magnitude and phase
over the entire frequency range, including the transition bands of the filter. The
method followed to find the unit sample response coefficients h(n) is completely
numerical, and we know that the efficient FFT techniques are used to compute
the DFT and IDFT samples. We have already used the MATLAB functions fft
and ifft that compute the DFT and IDFT, respectively, in Chapter 3, and the
function fir2 simply implements IDFT of H (k) specified by the input vectors
F and M in the function fir2.

5.8 SUMMARY

In this chapter, we discussed the design theory of FIR filters with linear phase
and a magnitude response that approximate the magnitudes of ideal LP, HP, BP,
and BS filters as well as some filters that have magnitude specifications that are
smooth but not necessarily piecewise constant. We also described a few very
efficient and well-known MATLAB functions that obtain very good results in
designing these filters. But there are cases when these functions (and a few others
not included in this chapter) may not work satisfactorily. Students are encouraged
to work extensively these MATLAB functions with a variety of specifications and
input arguments and build their experience and insight about the relative mer-
its and advantages of the various methods and the MATLAB functions. It was
pointed out that the function remez does not work very efficiently in designing
highpass and bandstop filters. We suggest below an alternative approach to solve
this problem. However, the three transformations given below [1] are more gen-
eral and useful in transforming the magnitude response of a type I filter into that
of a wide variety of other magnituderesponse characteristics.
−n
Consider a type I filter H (z) = 2M n=0 h(n)z , h(2M − n) = h(n) that has
a passband frequency ωp and stopband frequency ωs . Its zero phase frequency

response is HR (ω) = h(M) + M n=1 2h(M − n) cos(nω). We can obtain three
new classes of type I filters that have the following transfer functions:
⎧ −M
⎨z − H (z) transformation A
G(z) = (−1)M H (−z) transformation B
⎩ −M
z − (−1)M H (−z) transformation C

The corresponding frequency response of these filters is given by

⎧
⎨1 − H (ω) transformation A
G(ω) = H (π − ω) transformation B
⎩
1 − H (π − ω) transformation C
 SUMMARY 293

H(w)
1 + dp

1 − dp

ds
− ds
0 wp ws p w
(a)

G(w) = 1 – H(w)
1 + ds
1 − ds

dp

− dp
0 w p ws p w
(b)

G(w) = H(p − w)
1 + dp

1 − dp

ds
− ds
0 p − ws p − wp p w
(c)

G(w) = 1-H(p − w)
1 + ds
1 − ds

dp

− dp
0 p − ws p − wp p w
(d)

Figure 5.24 Transformation of a lowpass filter response to other types of filter responses.

The responses of these filters are plotted in Figure 5.24. We notice that trans-
formation A transforms a type I, lowpass filter with a passband (cutoff) frequency
ωp and a stopband frequency ωs , into a highpass filter with the cutoff frequency
of ωs and a stopband frequency ωp , whereas transformation B transforms the
294 FINITE IMPULSE RESPONSE FILTERS

lowpass filter response into a highpass filter with a cutoff frequency (π − ωp )

and a stopband frequency (π − ωs ).
To design an equiripple highpass FIR filter, with passband and stopband cutoff
frequencies ωp and ωs , respectively, we first design a lowpass type I equirip-
ple FIR filter with a passband frequency ωp = π − ωp and a stopband frequency
 −n
ωs = π − ωs . Let its transfer function be H (z) = 2M n=0 h(n)z . Then, using the
transformation G(z) = M
(−1) H (−z), we get the transfer function of the high-
pass filter as G(z) = 2M n=0 (−1) M+n
h(n) z−n . Transformation C transforms the
type I lowpass filter into another lowpass filter with a cutoff frequency (π − ωs )
and a stopband frequency (π − ωp ). These transformations provide us with addi-
tional tools to design filters in the event that any one method fails to give the
right results.

PROBLEMS

5.1 Derive the function for the frequency response in the form ej θ {HR (ω)}
for the FIR filters H (z−1 ) given below. Identify the type of filters also:

H1 (z−1 ) = 1 + 0.6z−1 + 0.2z−2 + 0.4z−3 + 0.2z−4 + 0.6z−5 + z−6

H2 (z−1 ) = 1.2 + 0.1z−1 + 0.9z−2 + z−3 − z−4 − 0.9z−5
− 0.1z−6 − 1.2z−7
H3 (z−1 ) = 1 + z−1 + z−2 − z−3 − z−4 − z−5
H4 (z−1 ) = 1 + z−3 + z−6
H5 (z−1 ) = 0.4z−1 − 0.6z−2 + 1.4z−3 + 0.6z−4 − 0.4z−5

5.2 Given that a polynomial√P1 (z−1 ) = 1 − 2.5z−1 + 5.25z−2 − 2.5z−3 + z−4

has a zero at z = 1 + j 3, find the other zeros of P1 (z−1 ).
5.3 The polynomial P2 (z−1 ) = 1.0 − 1.5z−1 + 2.75z−2 + 2.75z−3 − 1.5z−4 +
z−5 has zero at z = 0.25 − j 0.433. Find the other zeros of P2 (z−1 ).
5.4 The polynomial P3 (z−1 ) = 1 + 3.25z−2 − 3.25z−4 − z−6 has zero at z =
0.5ej 1.5π . Find the other zeros of P3 (z−1 ).
5.5 Find the polynomial N (z−1 ) that is the mirror image of the polynomial
D(z−1 ) = 1 + 0.2z−1 − 0.5z−2 + 0.04z−3 . Show that the magnitude of
H (e−j ω ) = N (e−j ω )/D(e−j ω ) is a constant for all frequencies.
5.6 Find the polynomial N (z−1 ) that is the anti–mirror image of the polynomial
D(z−1 ) = 1 + 0.32z−1 + 0.6z−2 − 0.4z−3 + 0.1z−4 . Show that the magni-
tude of H (e−j ω ) = N (e−j ω )/D(e−j ω ) is a constant for all frequencies.
 PROBLEMS 295

5.7 The passband of a lowpass FIR filter lies between 1.04 and 0.96, and its
stopband lies below 0.0016. Find the value of the passband attenuation αp
and the stopband attenuation αs .
5.8 The passband of a lowpass FIR filter lies between 1.15 and 0.9, and its
stopband lies below 0.0025. Find the value of the passband attenuation αp
and the stopband attenuation αs .
5.9 The passband of a lowpass FIR filter lies between 1 + δp and 1 − δp , and
its stopband lies below δs . If the passband attenuation is 0.15 dB and the
stopband attenuation is 45 dB, what are the values of δp and δs ?
5.10 The passband of a lowpass FIR filter lies between 1 + δp and 1 − δp , and
its stopband lies below δs . If the passband attenuation is 0.85 dB and the
stopband attenuation is 85 dB, what are the values of δp and δs ?
5.11 Design a lowpass FIR filter of length of 15, with ωc = 0.6π, using the
Fourier series method; truncate it with a Hann window, and delay the
samples by seven samples to get the transfer function of the causal filter.
5.12 In designing an FIR BP filter with ωc2 = 0.5π and ωc1 = 0.1π, using
the Fourier series method and a rectangular window of length 9, what
are the values of h(3) and h(9) in the transfer function of the causal FIR
filter?

5.13 In designing a bandpass FIR filter H (z−1 ) = 10 −n
n=0 h(n)z , using the
Fourier series method and a Bartlett window in order to approximate the
magnitude response of the filter with ωc2 = 5π/6 and ωc1 = π/2, what
are the values of the samples h(3) and h(7)?
5.14 An FIR bandpass filter has cutoff frequencies at 0.25π and 0.5π. Find the
coefficients h(3) and h(6) of its transfer function H (z−1 ) = 10 −n
n=0 h(n)z ,
assuming that it is designed using the Fourier series method and a Black-
man window.
5.15 Design an FIR filter of length 9, to get a highpass response, with ωc =
0.4π, using a Hamming window.
5.16 The coefficients of the Fourier series for the frequency response of the
differentiator are given by
⎧
⎨0 for n = 0
c(n) = cos(πn)
⎩ for |n| > 0
n
Using a Hann window of length 9, find the value of the coefficient h(6)
of the causal FIR filter that approximates the magnitude response of the
differentiator.
296 FINITE IMPULSE RESPONSE FILTERS

X1(n)

0.8

0 1 2 3 4 5
−5 −4 −3 −2 −1 n

−0.8

(a)

X2(n)

0.8

−5 −4 −3 −2 −1
0 1 2 3 4 5 n

−0.8

(b)

Figure 5.25 Problem 5.18.

5.17 The coefficients of the Fourier series for the frequency response of a dig-
ital filter are c(n) = (0.5)|n| ; −∞ < n < ∞. A window function w(n) =
(−1)n , for −7 ≤ n ≤ 7, is applied to this sequence, and the product is
delayed by seven samples to get a causal sequence h(n). What is the
value of the fourth and eighth samples of h(n)?
5.18 Let x1 (n) be a window of length 11 shown in Figure 5.25a and y1 (n) =
x1 (n) ∗ x1 (n). Plot the function y1 (n) and derive its frequency response
Y1 (ej ω ).
 PROBLEMS 297

X(ejw)

1.0

−0.5p −0.3p
0 0.3p 0.5p p w

(a)

X(ejw)

1.0

0.2

−0.4p −0.2p 0 0.2p 0.4p p w

(b)

Figure 5.26 Problem 5.23.

5.19 Find the convolution sum y2 (n) = x2 (n) ∗ x2 (n) where x2 (n) is as shown
in Figure 5.25b. Plot y2 (n) and derive its frequency response (DTFT)
Y2 (ej ω ).
5.20 What is the frequency response of the filter attained by cascading the two
filters described by Figure 5.27a,b?
5.21 Plot the spectrum of the product y(n) = x1 (n)x2 (n) where x1 (n) = 10 cos
(0.5πn) and x2 (n) = cos(0.25πn).
5.22 If the signal y(n) given in Problem 5.21 is the input to the filter shown
in Figure 5.27b what is the output signal?
5.23 Derive the expressions for the Fourier series coefficients (for −∞ <
n < ∞) for the DTFT of an LTI DT system as shown in Figure 5.26a,b.
5.24 Derive the expressions for the Fourier series coefficients for −∞ < n <
∞ for the frequency response of the LTI-DT system as shown in
Figure 5.27a,b, respectively.
298 FINITE IMPULSE RESPONSE FILTERS

x(ejw)

−0.5p 10 0.5p p w

(a)

x(ejw)

−p −0.5p 0 0.5p p w

(b)

Figure 5.27 Problem 5.25.

5.25 Design an HP, FIR filter of length 21 and ωc = 0.4π, using a Hann
window. Plot its magnitude response using the MATLAB function
fft.

5.26 Derive the expressions for the Fourier series coefficients (for −∞ <
n < ∞) for the DTFT of an LTI-DT system as shown in Figure 5.28a,b.

X(ejw)

1.0 1.0

0.5p p
−0.5p 0.5p p −0.5p 0 w

−1.0 −1.0

(a) (b)

Figure 5.28 Problem 5.26.

 PROBLEMS 299

X(ejw)
1.0

X(ejw)
1.0

0.05
200 300 500 f −0.5p −0.4p 0.4p 0.5p p
w

(a) (b)

Figure 5.29 Problem 5.27.

5.27 Find the Fourier series coefficients for the frequency response of the low-
pass digital filter as shown in Figure 5.29a, in which the Nyquist frequency
is 500 Hz.
5.28 Find the Fourier series coefficients for −5 < n < 5 for the frequency
response of the lowpass filter shown in Figure 5.29b.
5.29 Find the coefficients of the unit impulse response for 0 ≤ n ≤ 64, using
the MATLAB function fir2 after sampling the frequency response of
the lowpass filter shown in Figure 5.29b. Compare the result with that
obtained in Problem 5.28.

MATLAB Problems
5.30 Design a lowpass FIR filter of length 21, with ωp = 0.2π and ωs = 0.5π,
using the spline function of order p = 2, 4 for the transition band. Plot
the magnitude response of these filters on the same plot. Compare their
characteristics.
5.31 Design a lowpass FIR filter of length 41 with ωp = 0.3π and ωs = 0.5π,
using the spline function of order p = 2, 4 for the transition band. Show
the magnitude responses of these filters on the same plot. Compare their
characteristics.
5.32 Design a lowpass FIR filter of length 41 with ωp = 0.4π and ωs = 0.5π,
using the spline function of order p = 2, 4 for the transition band. Give
the magnitude responses of these filters on the same plot. Compare their
characteristics.
5.33 Design a lowpass FIR filter with a passband cutoff frequency ωc =
0.25π and a magnitude of 2 dB, a stopband frequency ωs = 0.4π,
300 FINITE IMPULSE RESPONSE FILTERS

and a magnitude of 0.02. Use a Hamming window and a Hann win-

dow. Plot the magnitude response to verify that the specifications are
met.
5.34 Design a highpass FIR filter with a passband cutoff frequency ωc = 0.45π
and a magnitude of 2 dB, a stopband frequency ωs = 0.2π, and a magni-
tude of 0.02. Use a Hamming window and a Kaiser window of length 50.
Plot the magnitude response to verify that the specifications are met.
5.35 Design a bandpass FIR filter with the passband between ωc1 = 0.4π and
ωc2 = 0.6π with a magnitude of 2.5. Use a Kaiser window. Plot the mag-
nitude response of the filter.
5.36 Design a bandpass FIR filter with a passband between fc1 = 5 kHz and
fc2 = 6.5 kHz with a magnitude of 2.5, and two stopbands with frequen-
cies at fs1 = 4.5 kHz and fs2 = 7 kHz with a magnitude below 0.025. The
sampling frequency is 20 kHz. Use a Kaiser window. Plot the magnitude
response of the filter to verify that the specifications are met.
5.37 Design a bandstop FIR filter with a stopband between ωs1 = 0.35π and
ωs2 = 0.65π with a magnitude of 0.05, and two passband frequencies at
ωp1 = 0.2π and ωp2 = 0.8π with magnitudes above 0.95. Use a Kaiser
window. Plot the magnitude response of the filter.
5.38 Design an equiripple, lowpass FIR filter with a passband cutoff frequency
ωc = 0.25π and δp = 0.05, a stopband frequency ωs = 0.4π, and a mag-
nitude below 0.01. Plot the magnitude and verify that the specifications
are met.
5.39 Using the results obtained in Problem 5.38 and the frequency transforma-
tion B described in the main text at the end of this chapter, design an
equiripple highpass FIR filter. Plot its magnitude response.
5.40 Using the results for the lowpass FIR filter obtained above, and the fre-
quency transformation C described at the end of this chapter, design an
equiripple lowpass FIR filter. Plot its magnitude response.
5.41 Design an equiripple highpass FIR filter with a passband cut-off frequency
ωc = 0.45π and a αp = 0.15 dB, a stopband frequency ωs = 0.2π, and a
magnitude below 0.025.
5.42 Design an equiripple bandpass FIR filter with the passband between ωc1 =
0.4π and ωc2 = 0.6π with a magnitude between 1.02 and 0.8. Its stop-
bands have stopband frequencies at ωs1 = 0.2π and ωs2 = 0.8π with a
magnitude below 0.005. Plot its magnitude to verify that the specifications
are met.
 REFERENCES 301

5.43 Design a bandpass FIR filter with a passband between fc1 = 6 kHz and
fc2 = 7 kHz with αp = 0.2 dB, and two stopbands with stopband fre-
quencies at fs1 = 4 kHz and fs2 = 9 kHz with αs = 35 dB. The sampling
frequency is 20 kHz. Plot the magnitude response to verify that the spec-
ifications are met.

REFERENCES

1. T. Saramaki, Finite impulse response filter design, in Handbook for Digital Signal
Processing, S. K. Mitra and J. F. Kaiser, eds., Wiley-Interscience, New York, 1993,
Chapter 4, pp. 155–278.
2. T. W. Parks and J. H. McClellan, Chebyshev approximation of nonrecursive digital
filters with linear phase, IEEE Trans. Circuit Theory CT-19, 189–194 (1972).
3. G. C. Temes and D. Y. F. Zai, Least pth approximation, IEEE Trans. Circuit Theory
CT-16, 235–237 (1969).
4. L. R. Rabiner and B. Gold, Theory and Application of Digital Signal Processing,
Prentice-Hall, 1975.
5. T. W. Parks and C. S. Burrus, Digital Filter Design, Wiley-Interscience, New York,
1987.
6. H. D. Helms, Nonrecursive digital filters: Design method for achieving specifications
on frequency response, IEEE Trans. Audio Electroacoust. AU-16, 336–342 (Sept.
1968).
7. J. F. Kaiser, Nonrecursive digital filter design using the I0 -sinh window function,
Proc. 1974 IEEE Int. Symp. Circuits and Systems, April 1974, pp. 20–23.
8. S. K. Mitra, Digital Signal Processing—A Computer-Based Approach, McGraw-Hill,
New York, 2001.
9. S. K. Mitra and J. F. Kaiser, eds., Handbook for Digital Signal Processing, Wiley-
Interscience, 1993.
10. O. Herrmann, L. R. Rabiner and D. S. K. Chan, “Practical design rules for optimum
finite impulse response lowpass digital filters”, Bell System Tech. Journal, vol. 52,
pp. 769–799, 1973.
11. F. J. Harris, “On the use of windows for harmonic analysis with discrete Fourier
transform”, Proc. IEEE, 66, pp. 51–83, 1978.
12. J. K. Gautam, A. Kumar, and R. Saxena, “Windows: A tool in signal processing”,
IETE Technical Review, vol. 12, pp. 217–226, 1995.
13. Andreas Antoniou, Digital Filters: Analysis, Design and Applications, McGraw Hill,
1993.
14. Andreas Antoniou, “New improved method for the design of weighted-
Chebyshev, non-recursive, digital filters”, IEEE Trans. on Circuits and Systems,
CAS-30, pp. 740–750, 1983.
15. R. W. Hamming, Digital Filters, Prentice-Hall, 1977.
302 FINITE IMPULSE RESPONSE FILTERS

16. J. H. McClellan and T. W. Parks, “A unified approach to the design of optimum

FIR linear phase digital filters” IEEE Trans. on Circuit Theory, CT-20, pp. 697–701,
1973.
17. A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing, Prentice-Hall,
1989.
18. P. P. Vaidyanathan, “On maximally-flat linear phase FIR filters” IEEE Trans. on
Circuits and Systems, CAS-31, pp. 830–832, 1989.
 CHAPTER 6

Filter Realizations

6.1 INTRODUCTION

Once we have obtained the transfer function of an FIR or IIR filter that approxi-
mates the desired specifications in the frequency domain or the time domain, our
next step is to investigate as many filter structures as possible, before we decide
on the optimal or suboptimal algorithm for actual implementation or applica-
tion. A given transfer function can be realized by several structures or what we
will call “circuits,” and they are all equivalent in the sense that they realize the
same transfer function under the assumption that the coefficients of the transfer
function have infinite precision. But in reality, the algorithms for implementing
the transfer function in hardware depend on the filter structure chosen to realize
the transfer function. We must also remember that the real hardware has a finite
number of bits representing the coefficients of the filter as well as the values
of the input signal at the input. The internal signals at the input of multipliers
and the signals at the output of the multipliers and adders also are represented
by a finite number of bits. The effect of rounding or truncation in the addition
and multiplications of signal values depends on, for example, the type of rep-
resentation of binary numbers, whether they are in fixed form or floating form,
or whether they are in sign magnitude or two-complementary form. The effects
of all these finite values for the number of bits used in hardware implemen-
tation is commonly called “finite wordlength effects,” which we will study in
Chapter 7.
In this chapter we develop several methods for realizing the FIR and IIR filters
by different structures. The analysis or simulation of any transfer function can
be easily done on a general-purpose computer, personal computer, or worksta-
tion with a high number of bits for the wordlength. We can also investigate the
performance of noncausal systems or unstable systems on personal computers.
Simulation of the performance of an actual microprocessor or a digital signal
processor (DSP chip) by connecting it to the PC, a development kit that contains
the microprocessor or the DSP chip, is far preferable to designing and building

Introduction to Digital Signal Processing and Filter Design, by B. A. Shenoi

Copyright © 2006 John Wiley & Sons, Inc.

303
304 FILTER REALIZATIONS

the digital filter hardware with different finite wordlength and testing its perfor-
mance. Of course, extensive analysis (simulation) of a given filter function under
other design criteria such as stability, modularity, pipeline architecture, and noise
immunity is also carried out on a personal computer or workstation using very
powerful software that is available today.
It is true that a real hardware can be programmed to implement a large number
of algorithms, by storing the data that represent the input signals and coefficients
of the filter in a memory. But remember that it can implement an algorithm only
in the time domain, whereas programming it to find the frequency response is
only a simulation. Three algorithms in the time domain that we have discussed
in earlier chapters are the recursive algorithm, convolution sum, and the FFT
algorithm. It is the difference equations describing these algorithms that have to
be implemented by real digital hardware.
Consider the general example of an IIR filter function:
M
n=0 b(n)z−n
H (z) = N (6.1)
1+ n=1 a(n)z−n

The corresponding linear difference equation that implements it directly is

N
M
a(k)y(n − k) = b(k)x(n − k); a(0) = 1 (6.2)
k=0 k=0

It can then be rewritten in the form of a recursive algorithm as follows:

y(n) = −a(1)y(n − 1) + a(2)y(n − 2) + a(3)y(n − 3) + · · · + a(N )y(n − N )

+ b(0)x(n) + b(1)x(n − 1) + · · · + b(M)y(n − M) (6.3)

This recursive algorithm can be easily programmed on a general-purpose

microprocessor, a computer, or a full-function DSP chip. The filter function we
have obtained can be configured on these devices and convolution between its
unit impulse response h(n) and the input signal is the actual process used by the
hardware to produce the output. The convolution sum is given by
∞

y(n) = h(k)x(n − k) (6.4)
k=0

In the following pages, it will be shown that this transfer function (6.1) can
be realized by several structures. We must remember that the algorithms used
to implement them in the time domain will vary for the different structures. All
the equivalent structures realize the same transfer function only under infinite
precision of the coefficients; otherwise their performance depends on the number
of bits used to represent the coefficients, as well as the input signal and the form
for representing the binary numbers. The same statement can be made for the
 FIR FILTER REALIZATIONS 305

realization of an FIR filter function treated in the next section. The purpose of
realizing different structures and studying the effects of quantization is to find the
best possible structure that has the minimum quantization effect on the output of
the system.

6.2 FIR FILTER REALIZATIONS

Example 6.1: Direct Form

 −n
Given the transfer function of an FIR filter as H (z) = M n=0 h(n)z , let us
consider its equivalent algorithm for the output, for example, when M = 4:

y(n) = h(0)x(n) + h(1)x(n − 1) + h(2)x(n − 2)

+ h(3)x(n − 3) + h(4)x(n − 4) (6.5)

We have already discussed one structure employed to implement this algorithm

in Chapter 5, and because the coefficients of the multipliers in it are directly
available as the coefficients h(n) in H (z), it is called the direct form I structure
and is shown in Figure 6.1.
Whenever we have a structure to implement an FIR or an IIR filter, an equiv-
alent structure can be obtained as its transpose by the following operations:

1. Interchanging the input and the output nodes

2. Replacing adders by pickoff nodes and vice versa
3. Reversing all paths

Using these operations, we get the transpose of the structure of Figure 6.1
as Figure 6.2. This is known as direct form II structure; remember that this
(direct form II) structure will be called direct form I transposed structure in the
next chapter.

X(z)
z−1 z−1 z−1 z−1

h(0) h(1) h(2) h(3) h(4)

Σ Σ Σ Σ Y(z)

Figure 6.1 Direct form I of an FIR filter.

306 FILTER REALIZATIONS

z−1 Σ z−1 Σ z−1 Σ z−1 Σ

Y(z)

h(4) h(3) h(2) h(1) h(0)

X(z)

Figure 6.2 Direct form II of an FIR filter.

Example 6.2: Cascade Form

If we have an FIR filter of high order, it may be realized as a cascade of FIR

filters of lower order, preferably as second-order filters when the order is even or
the cascade of second-order filters and one first-order
filter when−nthe order is odd.
We factorize the given FIR filter function H (z) = M n=0 h(n)z in the form
⎧  
⎪ M/2  −1 −2
⎪
⎪ h(0) 1 + h(1m)z + h(2m)z
⎪
⎨
m=1

 when M is even 
H (z) = (M−1)/2  (6.6)
⎪ −1
1 + h(1m)z−1 + h(2m)z−2
⎪ h(0) (1 + h(10)z ) m=1
⎪
⎪
⎩
when M is odd

A cascade realization of this equation when M = 5 is shown in Figure 6.3, and

its transpose is shown in Figure 6.4.

X(z) h(0)
Σ Σ Σ Y(z)

z−1 z−1 z−1

h(10) h(11) h(12)

Σ Σ

z−1 z−1
h(21) h(22)

Figure 6.3 Cascade connection of an FIR filter.

 FIR FILTER REALIZATIONS 307

X(z) h(0) Y(z)

Σ Σ Σ

z−1 z−1
−1
z
h(12) h(11)
Σ Σ
h(10)

z−1 z−1
h(22) h(21)

Figure 6.4 Transpose of the cascade connection shown in Figure 6.3.

Example 6.3: Polyphase Form

This realization is based on the polyphase decomposition of the FIR transfer

function and is illustrated by choosing the following example:

H1 (z) = h(0) + h(1)z−1 + h(2)z−2 + h(3)z−3 + h(4)z−4

+ h(5)z−5 + h(6)z−6 + h(7)z−7 + h(8)z−8 (6.7)

This can be expressed as the sum of two subfunctions, shown below:

 
H1 (z) = h(0) + h(2)z−2 + h(4)z−4 + h(8)z−8
 
+ h(1)z−1 + h(3)z−3 + h(5)z−5 + h(7)z−7
 
= h(0) + h(2)z−2 + h(4)z−4 + h(8)z−8
 
+ z−1 h(1) + h(3)z−2 + h(5)z−4 + h(7)z−6 (6.8)
 
Let us denote A0 (z) = h(0) + h(2)z−2 + h(4)z−4 + h(8)z−8 and
 
A1 (z) = h(1)z−1 + h(3)z−3 + h(5)z−5 + h(7)z−7
 
= z−1 h(1) + h(3)z−2 + h(5)z−4 + h(7)z−6

Since the polynomials in the square brackets contain only even-degree terms,
we denote A0 (z) = A0 (z2 ) and A1 (z) = z−1 A1 (z2 ). Hence we express H1 (z) =
A0 (z2 ) + z−1 A1 (z2 ). A block diagram showing this realization is presented in
Figure 6.5(a), where the two functions A0 (z2 ) and A1 (z2 ) are subfilters connected
in parallel.
308 FILTER REALIZATIONS

Y(z) X(z) Y(z)

X(z) A0(z2) Σ B0(z3) Σ

z−1 z−1

A1(z2) B1(z3) Σ

z−1

B2(z3)

(a) (b)

Figure 6.5 Polyphase structures of FIR filters.

X(z) h(0)

h(2)

h(4)

h(6)

h(8) Y(z)
z−2 Σ z−2 Σ z−2 Σ z−2 Σ

z−1
h(7)

h(5)

h(3)

h(1)

Figure 6.6 Polyphase realization of an FIR filter H1 (z).

 FIR FILTER REALIZATIONS 309

These filters can be realized in either the direct form I or direct form II as
described earlier and illustrated in Figures 6.1 and 6.2, respectively. But there
would be 8 unit delays in building A0 (z) and 7 unit delays in z−1 A1 (z), which
adds up to 15 unit delay elements. We prefer to realize a circuit that would
require a minimum number of unit delays that is equal to the order of the filter.
A realization that contains the minimum number of delays is defined as a canonic
realization. To reduce the total number of delays to 8, we cause the two subfilters
to share the unit delays in order to get a canonic realization. Such a circuit
realization is shown in Figure 6.6.

Example 6.4

Consider the same example and decompose (6.7) as the sum of three terms:
   
H (z) = h(0) + h(3)z−3 + h(6)z−6 + z−1 h(1) + h(4)z−3 + h(7)z−6
 
+ z−2 h(2) + h(5)z−3 + h(8)z−6
= B0 (z3 ) + z−1 B1 (z3 ) + z−2 B2 (z3 ) (6.9)

A block diagram for implementing this decomposition is shown in Figure 6.5b. A

canonic realization of (6.9) is shown in Figure 6.7 and its transpose in Figure 6.8,
each of which uses 8 unit delay elements. The FIR filter of order 8 chosen
in (6.7) can be decomposed as the sum of four subfilters in the form H (z) =
C0 (z4 ) + C1 (z4 ) + C2 (z4 ) + C3 (z4 ) and can be realized by a canonic circuit.
In general, an FIR function of order N (i.e. h(n) = 0 for n > N ) can be
decomposed in the polyphase form with M subfilters connected in parallel as
follows

M−1
H (z) = z−m Em (zM ) (6.10)
m=0

where
+1)/M
(N
Em (z) = h(Mn + m)z−n (6.11)
n=0

6.2.1 Lattice Structure for FIR Filters

FIR filters can be realized in structures called lattice structures shown later in
the chapter in Figure 6.17a and its transpose, in Figure 6.17b—for an example
of a third-order filter. We will describe the design of these structures using a
MATLAB function in Section 6.5.
310 FILTER REALIZATIONS

h(0)

X(z) h(3)

h(6)

h(1)
z−1

h(7) Y(z)
Σ z−3 Σ z−3 Σ

z−1 h(4)

h(8)

h(5)

h(2)

Figure 6.7 Polyphase structure of an FIR filter function H2 (z).

6.2.2 Linear Phase FIR Filter Realizations

When the FIR filter has a linear phase, its coefficients are symmetric or antisym-
metric, and hence the number of multipliers is reduced by almost half. For the
symmetric FIR filter of order N , the samples of the unit impulse response that
are the same as the multiplier coefficients satisfy the condition h(n) = h(N − n)
and it known as a type I FIR filter. The FIR filter with antisymmetric coefficients
satisfies the condition h(n) = −h(N − n) and is known as the type II FIR filter.

Example 6.5

Let us consider the example of a type I FIR filter, of order N = 6:

H3 (z) = h(0) + h(1)z−1 + h(2)z−2 + h(3)z−3 + h(4)z−4 + h(5)z−5 + h(6)z−6

(6.12)
= h(0) + h(1)z−1 + h(2)z−2 + h(3)z−3 + h(2)z−4 + h(1)z−5 + h(0)z−6
= h(0)(1 + z−6 ) + h(1)(z−1 + z−5 ) + h(2)(z−2 + z−4 ) + h(3)z−3 (6.13)

By sharing the multipliers, we get the realization shown in Figure 6.9, which
uses only four multipliers. It is still a canonic realization that uses six delay
elements.
 FIR FILTER REALIZATIONS 311

h(2)

h(5)
Σ

h(8)

z−1
X(z) h(7)
z−3 z−3

h(4)
Σ Σ

h(1)

z−1
h(6)

h(3)
Σ Σ Y(z)

h(0)

Figure 6.8 Transpose of the polyphase structure shown in Figure 6.7.

X(z)
z−1 z−1 z−1

Σ Σ Σ

z−1 z−1 z−1

h(0) h(1) h(2) h(3)

Σ Σ Σ Y(z)

Figure 6.9 Direct-form structure of type I linear phase FIR filter function H3 (z).
312 FILTER REALIZATIONS

X(z)
z−1 z−1 z−1

Σ Σ Σ Σ z−1

z−1 z−1 z−1

h(0) h(1) h(2) h(3)

Σ Σ Σ Y(z)

Figure 6.10 Direct-form structure of type II linear phase FIR filter function H4 (z).

Example 6.6

If we consider a type II FIR filter of order 7, its transfer function is given by

H4 (z) = h(0) + h(1)z−1 + h(2)z−2 + h(3)z−3 + h(3)z−4 + h(2)z−5

+ h(1)z−6 + h(0)z−7 (6.14)
= h(0)(1 + z−7 ) + h(1)(z−1 + z−6 ) + h(2)(z−2 + z−5 )
+ h(3)(z−3 + z−4 ) (6.15)

This is realized by the canonic circuit shown in Figure 6.10, thereby reducing
the total number of multipliers from 7 to 4. Similar cost saving is achieved in
the realization of FIR filters with antisymmetric coefficients.

6.3 IIR FILTER REALIZATIONS

In constructing several equivalent structures of an FIR filter, we used the direct-

form decomposition of the filter transfer function as the product of second-order
sections connected in cascade, polyphase decomposition, and transpose for each
structure obtained by them. We used the symmetry of the coefficients of linear
phase FIR filters to reduce the number of delay elements. We can also generate
their transpose forms. Using similar strategies, in this section we present several
structures for the IIR filters.
 IIR FILTER REALIZATIONS 313

Example 6.7: Direct Forms

The transfer function (6.1) of an IIR filter is the ratio of a numerator polynomial
to a denominator polynomial. First we decompose it as the product of an all-pole
function H1 (z) and a polynomial H2 (z)
M
n=0 b(n)z−n
H (z) = N (6.16)
1+ n=1 a(n)z−n
  
1
M
= H1 (z)H2 (z) = N b(n)z−n (6.17)
1+ −n
n=1 a(n)z n=0

and construct a cascade connection of an FIR filter H2 (z) and the all-pole IIR
filter H1 (z). Again we select an example to illustrate the method. Let H2 (z) =
b0 + b(1)z−1 + b(2)z−2 + b(3)z−3 and

1
H1 (z) =
1+ a(1)z−1 + a(2)z−2 + a(3)z−3

The realization of H1 (z) in direct form I is shown in Figure 6.11 as the filter
connected in cascade with the realization of the FIR filter H2 (z) also in direct
form I structure. The structure for the IIR filter is also called a direct form
I because the gain constants of the multipliers are directly available from the
coefficients of the transfer function.
We note that H1 (z) = V (z)/X(z) and H2 (z) = Y (z)/V (z). We also note that
the signals at the output of the three delay elements of the filter for H1 (z) are

V(z) b(0)
X(z) Σ Σ Y(z)

z−1 z−1
−a(1) b(1)
Σ v(n−1) v(n−1) Σ

z−1 z−1
−a(2) b(2)
Σ v(n−2) v(n−2) Σ

z−1 z−1

−a(3) b(3)
v(n−3) v(n−3)

Figure 6.11 Direct form I of an IIR filter.

314 FILTER REALIZATIONS

b(0)
X(z) Σ Σ Y(z)

z−1
−a(1) b(1)
Σ Σ

z−1
−a(2) b(2)
Σ Σ

z−1

−a(3) b(3)

Figure 6.12 Direct form II structure of an IIR filter.

the same as those at the output of the three delay elements of filter H2 (z). Hence
we let the two circuits share one set of three delay elements, thereby reducing
the number of delay elements. The result of merging the two circuits is shown
in Figure 6.12 and is identified as the direct form II realization of the IIR filter.
Its transpose is shown in Figure 6.13. Both of them use the minimum number
of delay elements equal to the order of the IIR filter and hence are canonic
realizations.
The two filters realizing H1 (z) and H2 (z) can be cascaded in the reverse order
[i.e., H (z) = H2 (z)H1 (z)], and when their transpose is obtained, we see that
the three delay elements of H2 (z) can be shared with H1 (z), and thus another
realization identified as direct form I as well as its transpose can be obtained.

Example 6.8: Cascade Form

The filter function (6.16) can be decomposed as the product of transfer functions
in the form
N1 (z)N2 (z) · · · NK (z)
H (z) = (6.18)
D1 (z)D2 (z)D3 (z) · · · DK (z)
     
N1 (z) N2 (z) N3 (z) NK (z)
= ··· (6.19)
D1 (z) D2 (z) D3 (z) DK (z)
= H1 (z)H2 (z)H3 (z) · · · HK (z) (6.20)
 IIR FILTER REALIZATIONS 315

b(0)
X(z) Σ Y(z)

z−1

b(1) −a(1)
Σ

z−1

b(2) −a(2)
Σ

z−1

b(3) −a(3)
Σ

Figure 6.13 Direct form II transposed structure of an IIR filter.

where K = N/2 when N is even and the polynomials D1 (z), D2 (z), D3 (z), and
so on are second-order polynomials, with complex zeros appearing in conjugate
pairs in any such polynomial. When N is odd, K = (N − 1)/2, and one of the
denominator polynomials is a first-order polynomial. The numerator polynomials
N1 (z), N2 (z), . . . may be first-order or second-order polynomials or a constant:
   
1 + β11 z−1 ( 1 + β1k z−1 + β2k z−2
H (z) = H0 (6.21)
1 + α11 z−1 k 1 + α1k z−1 + α2k z−2

Each of the transfer functions H1 (z), H2 (z), . . . , HK (z) is realized by the direct
form I or direct form II or their transpose structures and then connected in
cascade. They can also be cascaded in many other sequential order, for example,
H (z) = H1 (z)H3 (z)H5 (z) . . . or H (z) = H2 (z)H1 (z)H4 (z)H3 (z) . . . .
There are more choices in the realization of H (z) in the cascade connection in
addition to those indicated above. We can pair the numerators N1 (z), N2 (z), . . .
and denominators D1 (z), D2 (z), D3 (z), . . . in many different combinations; in
other words, we can pair the poles and zeros of the polynomials in different
316 FILTER REALIZATIONS

ways. For example, we can define

     
N1 (z) N2 (z) N3 (z) NK (z)
H (z) = ···
D2 (z) D3 (z) D4 (z) D1 (z)
     
N2 (z) N3 (z) N4 (z) Nk (z)
= ···
D1 (z) D4 (z) D3 (z) D2 (z)
   
NK (z) N2 (z) N4 (z)
= ···
D1 (z) D3 (z) D2 (z)

and cascade them in many different orders.

So the number of realizations that can be obtained from a nominal IIR transfer
function is very large, in general. Besides the difference in the algorithms for
each of these realizations and the consequent effects of finite wordlength when the
coefficients of the filter and the signal samples are quantized to a finite number,
we have to consider the effect on the overall magnitude of the output sequence
and the need for scaling the magnitude of the output sequence at each stage of
the cascade connection and so on.

Example 6.9

Consider a simple example of an IIR filter as follows:

z(0.16z − 0.18)
H (z) = (6.22)
(z − 0.2)(z + 0.1)(z + 0.4)(z2 + z + 0.5)

A few alternate forms of expressing this are given below:

   
1 (0.16z − 0.18) z
H (z) = (6.23)
(z2 − 0.1z − 0.02) (z2 + z + 0.5) (z + 0.4)
   
z 1 (0.16z − 0.18)
=
(z + 0.4) (z2 − 0.1z − 0.02) (z2 + z + 0.5)
   
(0.16z − 0.18) z 1
=
(z2 − 0.1z − 0.02) (z2 + z + 0.5) (z + 0.4)
   
z 1 (0.16z − 0.18)
= (6.24)
(z2 + z + 0.5) (z + 0.4) (z2 − 0.1z − 0.02)

Let us choose the last expression, (6.24), and rewrite it in inverse powers of z,
as given by
   
−1 z−1 z−1 (0.16z−1 − 0.18z−2 )
H (z ) =
1 + z−1 + 0.5z−2 (1 + 0.4z−1 ) (1 − 0.1z−1 − 0.02z−2 )
(6.25)
 IIR FILTER REALIZATIONS 317

X(z)
Σ Σ

z−1 z−1
−1
Σ Σ
−0.4
z−1
z−1
0.1 0.16
−0.5 Σ
Y(z)
z−1 Σ

0.02 −0.18

Figure 6.14 Cascade connection of an IIR filter.

One of the realizations used to implement this transfer function is shown in

Figure 6.14.
Instead of combining the factors (z − 0.2) and (z + 0.1) and getting (z2 −
0.1z − 0.02), in the denominator of (6.22), we can combine (z − 0.2) and (z +
0.4) or (z + 0.1) and (z + 0.4) to generate new second-order polynomials and
select many pole–zero pairs and order of second-order sections connected in
cascade, adding to the many possible realizations of (6.22) in the cascade form.
The cascade connection of second-order sections, each realized in direct form II,
has been a popular choice for a long time and was investigated in great detail,
until other structures became known for their better performance with respect to
finite wordlength effects and practical applications.

Example 6.10: Parallel Form

The IIR transfer function can also be expanded as the sum of second-order
structures. It is decomposed into its partial fraction form, combining the terms
with complex conjugate poles together such that we have an expansion with real
coefficients only. We will choose the same example as (6.22) to illustrate this
form of realization.
One form of the partial fraction expansion of (6.22) is

5.31165z 1.111z 5.21368z 1.1314z2 − 0.15947z

H (z) = − − + (6.26)
z + 0.1 z − 0.2 z + 0.4 z2 + z + 0.5

By combining (z + 0.1), (z − 0.2), (z + 0.4) in different pairs to get the cor-

responding denominator polynomials, we get the following expressions for the
318 FILTER REALIZATIONS

transfer function given above:

0.9797z2 + 1.6033z 1.111z 1.1314z2 − 0.15947z

H (z) = − + (6.27)
z2 + 0.5z + 0.04 z − 0.2 z2 + z + 0.5
−6.3248z2 + 0.5983z 5.31165 1.1314z2 − 0.15947z
= + + (6.28)
z2 + 0.2z − 0.08 z + 0.1 z2 + z + 0.5
4.200z2 − 1.1173z 5.2134z 1.1314z2 − 0.15947z
= − + (6.29)
z2 − 0.1z − 0.02 z + 0.4 z2 + z + 0.5
The three terms in these expressions are rewritten in inverse powers of z, and
any one of the IIR realizations (direct form or their transpose) is used to obtain
the circuit for each of them, and they are connected in parallel. Let us select the
last expression:

4.200z2 − 1.1173z 5.2134z 1.1314z2 − 0.15947z

H (z) = − +
z2 − 0.1z − 0.02 z + 0.4 z2 + z + 0.5
4.200 − 1.117z−1 5.2134 1.1314 − 0.1594z−1
= − + (6.30)
1 − 0.1z−1 − 0.02z−2 1 + 0.4z−1 1 + z−1 + 0.5z−2
Figure 6.15 shows a realization of the filter given by (6.30) in the parallel
form of connection, and by using the transpose of each sections, a new circuit
can be derived.
Another form of expanding the transfer function is the normal form of partial
fraction expansion, indicated by

R1 R2 R3 R4 z + R5
H (z) = + + +
z + 0.1 z − 0.2 z + 0.4 z2 + z + 0.5
R1 z−1 R2 z−1 R3 z−1 R4 z−1 + R5 z−2
= + + + (6.31)
1 + 0.1z−1 1 − 0.2z−1 1 + 0.4z−1 1 + z−1 + 0.5z−2
which gives rise to additional structures.
So, the transfer function given by (6.22) was decomposed in the form of (6.25)
and realized by the cascade structure shown in Figure 6.14; it was decomposed in
the form of (6.30) and realized by the parallel connection in the structure shown
in Figure 6.15.
The algorithm used to implement the structure in Figure 6.14 is of the form

y1 (n) = x(n) − y1 (n − 1) − 0.5y1 (n − 2)

y2 (n) = y1 (n) − 0.4y2 (n − 1)
y3 (n) = y2 (n) + 0.1y3 (n − 1) + 0.02y3 (n − 2)
y(n) = 0.16y3 (n − 1) − 0.18y3 (n − 2)
 IIR FILTER REALIZATIONS 319

4.2
Σ

Σ
z−1

Σ
0.1 −1.117
z−1

0.02
5.2134 Y(z)
X(z)
Σ Σ

z−1
−0.4

1.1314
Σ

z−1 Σ
−1.0
Σ
−0.1594
−1
z

−0.5

Figure 6.15 Parallel connection of the IIR filter function H (z).

whereas the algorithm employed to implement the structure shown in Figure 6.15
has the form

7
y1 (n) = x(n) + 0.17
y1 (n − 1) + 0.027
y1 (n − 2)
7
y2 (n) = 4.27
y1 (n) − 1.1177
y1 (n − 1)
7
y3 (n) = x(n) − 0.47
y3 (n − 1)
y4 (n) = x(n) − 7
7 y4 (n − 1) − 0.57
y4 (n − 2)
7
y5 (n) = 1.13147
y4 (n) − 0.15947
y4 (n − 1)
y(n) = 7 y3 (n) + 7
y2 (n) + 5.21347 y5 (n)

Remember that under ideal conditions both algorithms give the same output
for a given input signal and the two structures realize the same transfer function
(6.22). But when the two algorithms have to be programmed and implemented
by hardware devices, the results would be very different and the accuracy of
320 FILTER REALIZATIONS

the resulting output, the speed of the execution, and the throughput, and other
factors would depend not only on the finite wordlength but also on so many
other factors, including the architecture of the DSP chip, program instructions
per cycle, and dynamic range of the input signal. We will discuss these factors
in a later chapter.

6.4 ALLPASS FILTERS IN PARALLEL

Next in importance is the structure shown in Figure 6.16. The transfer function
G(z) = Y (z)/X(z) is given by 12 [A1 (z) + A2 (z)], where A1 (z) and A2 (z) are
the allpass filters connected in parallel. But in this figure, there is another trans-
fer function, H (z) = V (z)/X(z), which is given by H (z) = 12 [A1 (z) − A2 (z)].
The structure shown in Figure 6.16 is also called the lattice structure or lattice-
coupled allpass structure by some authors. A typical allpass filter function is of
the form
N (z) an + an−1 z−1 + an−2 z−1 + · · · + a1 z−n+1 + a0 z−n
A(z) = =± (6.32)
D(z) a0 + a1 z−1 + a2 z−2 + · · · + an−1 z−n+1 + an z−n

which shows that the order of the coefficients in the numerator is the reverse of
that in the denominator, when both the numerator and denominator polynomial
are expressed in descending powers of z. Equation (6.32) can be expressed in
another form as

z−n (a0 + a1 z + a2 z2 + · · · + an−1 zn−1 + an zn )

A(z) =
a0 + a1 z−1 + a2 z−2 + · · · + an−1 z−n+1 + an z−n
z−n D(z−1 )
= (6.33)
D(z)

The zeros of the numerator polynomial D(z−1 ) are the reciprocals of the zeros
of the denominator D(z), and therefore the numerator polynomial D(z−1 ) is the
mirror image polynomial of D(z).
When the allpass filter has all its poles inside the unit circle in the z plane, it is
a stable function and its zeros are outside the unit circle as a result of the mirror

A1(z) Σ 1/2 Y(z)

X(z)

A1(z) −1 Σ 1/2 V(z)

Figure 6.16 Two allpass filters in parallel (lattice-coupled allpass structure).

 ALLPASS FILTERS IN PARALLEL 321

image symmetry. Therefore a stable, allpass filter function is a non–minimum

phase function.
Letting a0 = 1, we get the magnitude response of this filter to be a constant at
all frequencies, because the numerator is the complex conjugate of the denom-
inator. Since this filter transmits all frequencies with the same gain, it is called
an allpass filter:
 
   + · · · + an ej nω 
A(ej ω ) =  1 + a1 e + a2 e
jω j 2ω
 1 + a e−j ω + a e−j 2ω + · · · + a e−j nω  = 1 (6.34)
1 2 n

But the phase response (and the group delay) is dependent on the coeffi-
cients of the allpass filter. We know that the phase response filter designed
to approximate a specified magnitude response is a nonlinear function of ω,
and therefore its group delay is far from a constant value. When an allpass fil-
ter is cascaded with  such a filter,  the resulting filter  has a frequency response
H1 (ej ω )A(ej ω ) = H1 (ej ω )A(ej ω ) ej [θ(ω)+φ(ω)] = H1 (ej ω ) ej [θ(ω)+φ(ω)] . So the
magnitude response does not change when the IIR filter is cascaded with an all-
pass filter, but its phase response θ (ω) changes by the addition of the phase
response φ(ω) contributed by the allpass filter. The allpass filters A(z) are there-
fore very useful for modifying the phase response (and the group delay) of filters
without changing the magnitude of a given IIR filter H1 (z), when they are cas-
caded with H1 (z). However, the method used to find the coefficients of the allpass
filter A(z) such that the group delay of H1 (z)A(z) is a very close approximation
to a constant in the passband of the filter H1 (z) poses a highly nonlinear problem,
and only computer-aided optimization has been utilized to solve this problem.
Normally IIR filters are designed from specifications for its magnitude only, and
its group delay is far from a linear function of frequency. There are many appli-
cations that call for a constant group delay or a linear phase response, and in
such cases, the filters are cascaded with an allpass filter that does not affect its
magnitude—except by a constant—but is designed such that it compensates for
the phase distortion of the IIR filter. Allpass filters designed for this purpose are
cascaded with the IIR filters and are known as delay equalizers.
An important property of allpass filters is that if the coefficients change in
wordlength, the magnitude response of an allpass filter at all frequencies does
not change. Recall that a second-order allpass filter was analyzed in Chapter 2,
and that if the transfer function of an allpass filter is of a higher order, it can
be realized by cascading second-order filters and possibly one first-order allpass
filter. We illustrate a few more structures that realize first-order allpass trans-
fer functions as well as second-order allpass functions later in the chapter, in
Figures 6.23 and 6.24.
Let us assume that the two allpass filters shown in Figure 6.16 are of order
(N − r) and r, respectively, and are given by

z−(N −r) D1 (z−1 )

A1 (z) = (6.35)
D1 (z)
322 FILTER REALIZATIONS

and

z−r D2 (z−1 )
A2 (z) = (6.36)
D2 (z)

Substituting them in G(z) = 12 [A1 (z) + A2 (z)] and H (z) = 12 [A1 (z) − A2 (z)],
we get
 
1 z−(N −r) D1 (z−1 )D2 (z) + z−r D2 (z−1 )D1 (z)
G(z) = (6.37)
2 D1 (z)D2 (z)

and
 
1 z−(N −r) D1 (z−1 )D2 (z) − z−r D2 (z−1 )D1 (z)
H (z) = (6.38)
2 D1 (z)D2 (z)

If we express them in the form as

N
P (z) n=0 pn z−n
G(z) = = (6.39)
D(z) D(z)

and
N −n
Q(z) n=0 qn z
H (z) = = (6.40)
D(z) D(z)

then it can be shown that the following conditions are satisfied by (6.37) and
(6.38).

Property 6.1 P (z−1 ) = zN P (z). Hence pn = pN −n , that is, the coefficients of

P (z) are symmetric.

Property 6.2 Q(z−1 ) = −zN Q(z). Hence qn = −qN −n , that is, the coefficients
of Q(z) are antisymmetric.

Property 6.3 P (z)P (z−1 ) + Q(z)Q(z−1 ) = D(z)D(z−1 ). Hence

   
G(ej ω )2 + H (ej ω )2 = 1 (6.41)

in other words, G(z) and H (z) are said to form a power complementary pair.
In the next chapter the structure for realizing G(ej ω ) will be termed a lattice-
coupled allpass filter and because of the property stated here, the structure for
realizing H (ej ω ) will be called a lattice-coupled allpass power complementary
filter.
 ALLPASS FILTERS IN PARALLEL 323

Property 6.4
   j θ (ω)   
G(ej ω ) = 1 e 1 + ej θ2 (ω)  = 1 1 + ej (θ1 (ω)−θ2 (ω)  ≤ 1 (6.42)
2 2

where A1 (ej ω ) = ej θ1 (ω) and A2 (ej ω ) = ej θ2 (ω) .

In the following analysis, we will assume that the four conditions described
above are satisfied by G(z) and H (z) and derive the results that they can be
obtained in the form G(z) = 12 [A1 (z) + A2 (z)] and H (z) = 12 [A1 (z) − A2 (z)].
Consider Property 6.3: P (z)P (z−1 ) + Q(z)Q(z−1 ) = D(z)D(z−1 ). Using
Properties 6.1 and 6.2, we get

P (z)zN P (z) − zN Q(z)Q(z) = D(z)D(z−1 ) (6.43)

P 2 (z) − Q2 (z) = D(z)z−N D(z−1 ) (6.44)
[P (z) + Q(z)] [P (z) − Q(z)] = z−N D(z)D(z−1 ) (6.45)
 
From Properties 6.1
 and 6.2, we have P (z−1 ) + Q(z−1 ) = zN [P (z) − Q(z)]
and therefore z−N P (z−1 ) + Q(z−1 ) = [P (z) − Q(z)], and we get

[P (z) + Q(z)] [P (z) − Q(z)] = P 2 (z) − Q2 (z) (6.46)

From (6.44), we have

P 2 (z) − Q2 (z) = D(z)z−N D(z−1 ) (6.47)

 
[P (z) + Q(z)] z−N P (z−1 ) + Q(z−1 ) = D(z)z−N D(z−1 ) (6.48)

Therefore

[P (z) + Q(z)] [P (z) − Q(z)] = z−N D(z)D(z−1 ) (6.49)

This shows that the zeros of [P (z) − Q(z)] are reciprocals of the zeros of
[P (z) + Q(z)].
It has been found that the Butterworth, Chebyshev, and elliptic lowpass filters
of odd order satisfy the four properties described above. We know from Chapter 4
that their transfer function G(z) obtained from the bilinear transformation of
the analog lowpass prototype filters has no poles on the unit circle. In other
words, the zeros of D(z) are within the unit circle, and therefore the zeros of
D(z−1 ) are outside the unit circle, because they are the reciprocals of the zeros
of D(z). From (6.49) we see that the zeros of [P (z) + Q(z)] and [P (z) − Q(z)]
cannot lie on the unit circle. Let us assume that [P (z) + Q(z)] has r zeros
324 FILTER REALIZATIONS

zk (k = 1, 2, 3, . . . , r) that are inside the unit circle and (N − r) zeros zj (j =

r + 1, r + 2, . . . , N ) that are outside the unit circle. Therefore [P (z) − Q(z)]
has r zeros zk−1 (k = 1, 2, 3, . . . , r) outside the unit circle and (N − r) zeros zj−1
(j = r + 1, r + 2, . . . , N ) inside the unit circle. From (6.49), we can therefore
assume that D(z) that has all its zeros inside the unit circle is of the form
⎡ ⎤
(r (
N
D(z) = ⎣ (1 − z−1 zk ) (1 − z−1 zj−1 )⎦ (6.50)
k=1 j =r+1

and

[P (z) + Q(z)] [P (z) − Q(z)] = z−N D(z)D(z−1 )

⎡ ⎤
(
r (
N
= ⎣ (1 − z−1 zk ) (1 − z−1 zj−1 )⎦
k=1 j =r+1
⎡ ⎤
(
r (
N
×⎣ (z−1 − zk ) (z−1 − zj−1 )⎦ (6.51)
k=1 j =r+1

Thus we identify

(
r (
N
P (z) + Q(z) = α (1 − z−1 zk ) (z−1 − zj−1 ) (6.52)
k=1 j =r+1

1 (
r (
N
P (z) − Q(z) = (z −1
− zk ) (1 − z−1 zj−1 ) (6.53)
α
k=1 j =r+1

Then
+ ,
P (z) + Q(z) (
N
z−1 − zj−1
G(z) + H (z) = =α = αA1 (z) (6.54)
D(z)
j =r+1
1 − z−1 zj−1
r  
P (z) − Q(z) 1 ( z−1 − zk 1
G(z) − H (z) = = = A2 (z) (6.55)
D(z) α 1 − z−1 zk α
k=1

But from the power complementary property, we must have α 2 = 1. Therefore,

α = 1, so that

G(z) = 1
2
[A1 (z) + A2 (z)] (6.56)
H (z) = 1
2
[A1 (z) − A2 (z)] (6.57)

So we have proved that when G(z) is a Butterworth, Chebyshev, or elliptic

lowpass filter of odd order (which satisfy the four properties listed above), we
 ALLPASS FILTERS IN PARALLEL 325

can decompose G(z) as the sum of two allpass functions, A1 (z)/2 and A2 (z)/2.
Once we have derived the two allpass functions, we easily obtain H (z) as the
difference of A1 (z)/2 and A2 (z)/2 and realize it by the structure of Figure 6.16.
Because of the complementary power property, we see that H (z) realizes a
highpass filter.

6.4.1 Design Procedure

The procedure to find the poles and zeros of the two allpass filters from the
given Butterworth, Chebyshev, or elliptic lowpass filters of odd order is described
below. We have already pointed out that the transfer function G(z) = P (z)/D(z)
for these filters obtained from the corresponding analog prototype via the bilinear
transformation have all their poles inside the unit circle of the z plane. Their
magnitude response G(ej ω ) has a maximum value at ω = 0, which can be easily
obtained as the value of G(z) at z = 1. In order to satisfy Property 6.4, we have
to divide
 G(z) by G(ej ) or multiply G(z) by a scaling factor k = D(1)/P (1) so

that kG(ej ω ) ≤ 1 as the first step in the design procedure. Let us assume that
G(z) has already been scaled by k in our further discussion.
From (6.47), we write

Q2 (z) = P 2 (z) − D(z)z−N D(z−1 ) (6.58)

We know the numerator polynomial P (z) and the denominator polynomial D(z)
of the filter function G(z), and hence we can compute the  right side of the
−n
Equation 6.58. Let us denote Q2 (z) = Q(z)Q(z) as R(z) = 2N n=0 rn z . The
coefficients of R(z) = Q(z)Q(z) are computed by convolution of the coefficients
of Q(z) with the coefficients of Q(z):

n
rn = qn ∗ qn = qk qn−k (6.59)
k=0

These coefficients can be computed recursively by the following algorithm:

√
q0 = r0 (6.60)
r1
q1 = (6.61)
2q0

rn − n−1
k=1 qk qn−k
qn = , 2≤n≤N (6.62)
2q0
But we need to compute qn for only n = 0, 1, 2, . . . , (N − 1)/2, since these
coefficients are antisymmetric and qn = 0 when n = (N + 1)/2.
When we  have computed the coefficients qn and constructed the polyno-
−n
mial Q(z) = N n=0 qn z , we get P (z) + Q(z), and now we factorize it to
find its N zeros. We identify the zeros inside the unit circle as the r poles
zk (k = 1, 2, 3, . . . , r) of A2 (z). By reversing the coefficients of the polynomial
326 FILTER REALIZATIONS

having these zeros, we get the numerator of A2 (z), which has the zeros at zk−1 .
We identify the zeros of P (z) + Q(z) that are outside the unit circle as the
(N − r) zeros zj (j = r + 1, r + 2, . . . , N ) of A1 (z). By reversing the order of
the coefficients of the numerator polynomial having these zeros, we obtain the
denominator polynomial of A1 (z). It has (N − r) zeros at zj−1 as shown in (6.54).
This completes the design procedure used to obtain A1 (z) and A2 (z) from G(z).
An example is worked out in Section 6.5.

6.4.2 Lattice–Ladder Realization

Another well-known realization of an IIR transfer function of the form (6.1) is
shown in Figure 6.19a, and is known as the lattice–ladder realization. It is also
called the autoregressive moving-average (ARMA) model in the literature on
speech processing, adaptive filters, and signal processing in general. When the
numerator of (6.1) is a constant, we have an all-pole model also known as the
autoregressive (AR) model, and the structure to realize a third-order AR model
is shown in Figure 6.19b, whereas when the denominator is a constant, we get
an FIR model called the moving-average (MA) model. The structure shown in
Figure 6.17a is the model for a third-order FIR filter function or the MA model,
and the structure in Figure 6.17b is its transpose. We do not present the theoreti-
cal analysis of lattice structures for these models as it is beyond the scope of this
book but explain the use of a MATLAB function tf2latc in the next section to
derive the structures. This function implements the theoretical procedure, and
for the ARMA model, it gives the N lattice parameters ki , i = 1, 2, . . . , N
and also the values of the N + 1 ladder coefficients vi , i = 0, 1, . . . , N . But

X(z) Y(z)
Σ Σ Σ

K2 K3
K1
K1 K2 K3

z−1 Σ z−1 Σ z−1 Σ V(z)

(a)

X(z)
Σ Σ Σ Y(z)

K3 K2 K1

K3 K2 K1

U(z)
Σ z−1 Σ z−1 Σ z−1

(b)

Figure 6.17 (a) Lattice structure for an FIR filter; (b) transpose of the lattice structure
for the FIR filter in (a).
 REALIZATION OF FIR AND IIR FILTERS USING MATLAB 327

in Figure 6.19a, the value of the ladder coefficient v5 happens to be zero for the
numerical example, and therefore the multiplier v5 is zero. The lattice parameters
are also known as the reflection coefficients, and it has been shown that the poles
of the IIR filter function are inside the unit circle of the z plane if |ki | ≤ 1. So
this method is used to test whether an IIR filter is stable.

6.5 REALIZATION OF FIR AND IIR FILTERS USING MATLAB

Many of the computations involved in the realization of FIR and IIR filters as
presented in this chapter can be carried out by MATLAB functions. For example,
an FIR filter realization in the cascaded structure can be obtained by finding the
roots of the transfer function and then finding the second-order polynomials with
complex conjugate pair of the roots or a pair of two real zeros.
−n
To find the roots of a polynomial H (z) = N n=0 b(n)z , we use the MATLAB
function R = roots(b) where the vector b = [b(0), b(1), b(2), · · ·
b(N)] and R is the vector of the N roots. Choosing a pair of complex conju-
gate roots or a pair of real roots, we construct the second-order polynomials
using the MATLAB function Pk =poly(Rk ), where Rk is the list of two roots and
Pk is the vector of the coefficients of the second-order polynomial. Of course, if
H (z) is an odd-order polynomial, one first-order polynomial with a single real
root will be left as a term in the decomposition of H (z).

Example 6.11

Using the MATLAB commands

b = [1.965 -3.202 4.435 -3.14 1.591 -0.3667];

R= roots(b)

we get the roots

0.2682 + 0.8986i
0.2682 - 0.8986i
0.3383 + 0.6284i
0.3383 - 0.6284i
0.4166

Then we continue

R1=[0.2682+0.8986*i 0.2682-0.8986*i];
P1=poly(R1)
P1=
1.0000 -0.5364 0.8794
R2=[0.3383+0.6284*i 0.3383-0.6284*i];
P2=poly(R2)
328 FILTER REALIZATIONS

P2 =
1.0000 -0.6766 0.5093

Hence, if H (z) = 1.965 − 3.202z−1 + 4.435z−2 − 3.14z−3 + 1.591z−4 −

−5
0.3667z , we construct a polynomial in positive powers of z and the coefficient
of the highest-degree term normalized to unity:

N (z) = [1.965z5 − 3.202z4 + 4.435z3 − 3.14z2 + 1.591z − 0.3667]

= 1.965[z5 − 1.6295z4 + 2.257z3 − 1.598z2 + 0.8096z − 0.1866]

From the output data for the coefficients of P1 and P2 displayed above, we
construct the polynomial

N (z) = 1.965(z2 − 0.5364z + 0.8794)(z2 − 0.6766z + 0.5093)(z − 0.4166)

Then we get H (z) in the form

H (z) = 1.965(1 − 0.5364z−1 + 0.8794z−2 )(1 − 0.6766z−1 + 0.5093z−2 )

× (1 − 0.4166z−1 )

Example 6.12

Consider the same FIR filter as given in Example 6.11. We use the simple
MATLAB function k = tf2latc(b) to get the vector output k listing the reflec-
tion coefficients ki , i = 1, 2, 3, 4, 5, where b is the vector of the coefficients given
in Example 6.11.
The vector output k for the lattice coefficients is

-0.3597
0.9325
-0.5745
0.5238
-0.1866

and the structure of the lattice realization for the FIR filter or the MA model is
shown in Figure 6.18, where the lattice coefficients are as listed above.

X(z)
Σ Σ Σ Σ Σ Y(z)
K2 K3 K4 K5
K1
K1 K2 K3 K4 K5

z−1 Σ z−1 Σ z−1 Σ z−1 Σ z−1 Σ V(z)

Figure 6.18 Lattice structure for a fifth-order FIR filter.

 REALIZATION OF FIR AND IIR FILTERS USING MATLAB 329

Example 6.13
To get a cascade realization of an IIR filter, one could factorize both the numerator
and the denominator as the product of second-order polynomials (and possibly
one first-order polynomial) as illustrated in Example 6.9. Another approach is to
use the MATLAB functions tf2zp and zp2sos as explained below.
First we use the function [z,p,k] = tf2zp(num,den) to get the output
vector [z,p,k], which lists the zeros, poles, and the gain constant for the IIR
filter. Then the function sos = zp2sos(z,p,k) gives the coefficients of the
second-order polynomials of each section in a matrix of order L × 6 in the
following format:
⎡ ⎤
n01 n11 n21 d01 d11 d21
⎢ n02 n12 n22 d02 d12 d22 ⎥
⎢· · · · · · ⎥
⎣ ⎦
· · · · · ·
n0L n1L n2L d0L d1L d2L
The six elements in each row define the transfer function of each second-order
section Hi (z) used in the product form as indicated below:
(L (L  
n0i + n1i z−1 + n2i z−2
H (z) = Hi (z) =
d0i + d1i z−1 + d2i z−2
i=1 i=1

These two MATLAB functions can be used to factorize an FIR function also.
Instead of the algorithm described above, we let the polynomial H (z) of the FIR
filter, as the denominator polynomial of an IIR filter and the numerator, be unity.
To illustrate this, let us consider the previous example and run the two functions
in the following MATLAB script:

num=1;
den=b
[z,p,k] = tf2zp(num,den);
sos = zp2sos(z,p,k)
sos =
0 0.5089 0 1.0000 -0.4166 0
0 0 1.0000 1.0000 -0.6766 0.5094
0 0 1.0000 1.0000 -0.5363 0.8794

Therefore the product form is given by

  
0.5089z−1 z−2
1 − 0.4166z−1 1 − 0.6766z−1 + 0.5094z−2
 
z−2
×
1 − 0.5363z−1 + 0.8794z−2
The terms in the denominator agree with those obtained in Example 6.11, which
used the functions roots and poly.
330 FILTER REALIZATIONS

Let us illustrate the decomposition of an IIR filter as the product of second-

order functions; consider the transfer function

0.5 + 0.2z−1 + 0.3z−2 + 0.1z−4

H (z) =
1.965 − 3.202z−1 + 4.435z−2 − 3.14z−3 + 1.591z−4 − 0.3667z−5
(6.63)
The MATLAB program used to obtain the factorized form to realize the cas-
cade structure for an IIR filter is

num=[0.5 0.2 0.3 0.0 0.1]

den=[1.965 -3.202 4.435 -3.14 1.591 -0.3667];
[z,p,k]=tf2zp(num,den);
sos=zp2sos(z,p,k)
sos =
0 0.2545 0 1.0000 -0.4166 0
1.0000 0.8204 0.6247 1.0000 -0.6766 0.5094
1.0000 -0.4204 0.3201 1.0000 -0.5363 0.8794

Using the entries in this sos matrix, we write the factorized form of H (z) as
follows:
  
0.2545z−1 1 + 0.8204z−1 + 0.6247z−2
1 − 0.4166z−1 1 − 0.6766z−1 + 0.5094z−2
 
1 − 0.4204z−1 + 0.3201z−2
× (6.64)
1 − 0.5363z−1 + 0.8794z−2

Note that the numerator in this expression seems to be a fifth-order polynomial

in inverse powers of z whereas the numerator of the transfer function (6.63)
is a fourth-order polynomial. But the factorization of a polynomial is carried
out when  it is expressed in positive powers of z since the polynomials are of
the form (z − zi ), where zi are the zeros. So when the preceding factorized
form is converted to the ratio of polynomials in positive powers of z, we get a
fourth-order numerator polynomial and the fifth-order denominator:
   
0.2545 z2 + 0.8204z + 0.6247 z2 − 0.4204z + 0.3201
z − 0.4166 z2 − 0.6766z + 0.5094 z2 − 0.5363z + 0.8794
0.5z4 + 0.2z3 + 0.3z2 + 0.1
=
1.965z5 − 3.202z4 + 4.435z3 − 3.14z2 + 1.591z − 0.3667

This agrees with the result of expressing H (z) as the ratio of a fourth-order
numerator polynomial and a fifth-order denominator polynomial in positive pow-
ers of z. So care is to be taken to express the transfer function in positive powers
of z and then check the results after constructing the factorized form, because the
function zp2sos works only if the zeros are inside the unit circle of the z plane.
 REALIZATION OF FIR AND IIR FILTERS USING MATLAB 331

But the factorized form of H (z) constructed from the sos matrix leads us cor-
rectly to the next step of drawing the realization structures for each section, for
example, by the direct form, and connecting them in cascade. Such a realization
is similar to that shown in Figure 6.14.

Example 6.14

In the previous chapter, we used the function [r,p,k]= residuez(num,den)

to find the zeros, poles, and the gain constant of an IIR filter function H (z). Now
we can select the two residues in a vector [rk(1) rk(2)] and the corresponding
poles [pk(1) pk(2)] in a complex conjugate pair or two real poles at a time
to construct the numerator and denominator of each section in the vector form
[bk,ak]. Then we express the IIR function H (z) as the sum of such second-order
functions—with one first-order section if H (z) has an odd number of poles. We
consider the same function H (z) given by (6.63) to realize the parallel structure.

b =
0.5000 0.2000 0.3000 0 0.1000
a =
1.9650 -3.2020 4.4350 -3.1400 1.5910 -0.3667
[r,p,k]=residuez(b,a)
r =
-0.1632 - 0.1760i
-0.1632 + 0.1760i
0.1516 - 0.0551i
0.1516 + 0.0551i
0.2777
p =
0.2682 + 0.8986i
0.2682 - 0.8986i
0.3383 + 0.6284i
0.3383 - 0.6284i
0.4166
k =
[]
r1 =
[-0.1632 + 0.1760i -0.1632-0.1760i]
p1 =
0.2682 - 0.8986i 0.2682+0.8986i
r2 =
[0.1516 - 0.0551i 0.1516 + 0.0551i]
p2 =
0.3383 + 0.6284i 0.3383 - 0.6284i
[b1,a1]=residuez(r1,p1,0)
b1 =
-0.3264 0.4038 0
332 FILTER REALIZATIONS

a1 =
1.0000 -0.5364 0.8794
[b2,a2]=residuez(r2,p2,0)
b2 =
0.3032 -0.0333 0
a2 =
1.0000 -0.6766 0.5093

The residue and the third pole are 0.2777 and 0.4166. So we construct the transfer
function H (z) as the sum of three terms

−0.3264 + 0.4038z−1 0.3032 − 0.0333z−1

H (z) = +
1 − 0.5364z−1 + 0.8794z−2 1 − 0.6766z−1 + 0.5093z−2
0.2777
+ . (6.65)
1 − 0.4166z−1
The structure found in realizing the parallel connection is similar to that shown
in Figure 6.15.

Example 6.15

We used the MATLAB function k = tf2latc(num) to obtain the lattice real-

ization of an (MA) FIR filter in Example 6.12. It is a special case of the more
general function [k,v] = tf2latc(num,den) used to obtain the lattice–ladder
(ARMA) realization of an IIR filter, where k is the vector of the lattice coeffi-
cients (reflection coefficients) and the vector v gives the ladder coefficients—an
example of which is shown in Figure 6.19a. Note that the lattice coefficients
are in reverse order (i.e., they are shown from right to left) of that shown in
Figure 6.17a; also note that the number of lattice coefficients in Figure 6.18(a) is
N but the number of the ladder coefficients is N + 1. Another special case is the
function k = tf2latc(1,den), which gives the lattice coefficients of an all-pole
(AR) IIR filter. A structure for a third-order AR model is given in Figure 6.19b.
Note the difference in the structure for this AR model and the structures for the
FIR filter that is defined as the MA model shown in Figure 6.17.
Let us consider the transfer function H (z) given in (6.63) to obtain its
lattice–ladder realization. We use the following MATLAB function [k,v] =
tf2latc(num,den) and get the vector outputs k and v immediately. The struc-
ture for this realization is shown in Figure 6.19a, where the values of the lattice
and ladder coefficients are as follows:

b = [ 0.5000 0.2000 0.3000 0 0.1000];

a=[ 1.9650 -3.2020 4.4350 -3.1400 1.5910 -0.3667];
[k,v] = tf2latc(b,a)
k =
-0.3597
0.9325
 REALIZATION OF FIR AND IIR FILTERS USING MATLAB 333

X(z)
Σ Σ Σ Σ Σ

−K5 K5 −K4 K4 −K3 K3 −K2 K2 −K1 K1

Σ z−1 Σ z−1 Σ z−1 Σ z−1 Σ z−1

v4 v3 v2 v1 v0

Σ Σ Σ
Y(z)
(a)

X(z)
Σ Σ Σ Y(z)

−K3 K3 −K2 K2 −K1 K1

V(z)
Σ z−1 Σ z−1 Σ z−1
(b)

Figure 6.19 (a) Lattice–ladder structure for an IIR filter (ARMA model); (b) lattice
structure for an all-pole IIR filter Y (z)/X(z) (AR model) and an allpass filter V (z)/X(z).

-0.5745
0.5238
-0.1866
v =
0.3831
0.3164
0.2856
0.1532
0.1000
0

Example 6.16

In order to illustrate the derivation of a lattice structure for an all-pole (AR model)
filter, we select a transfer function
1
H (z−1 ) = (6.66)
1− 0.2051z−1 − 0.0504z−2 + 0.0154z−3
and use the MATLAB function [k] = tf2latc(1,den) to get the vector output
for the lattice coefficients as shown below:
k = −0.2145
−0.0473
0.0154
334 FILTER REALIZATIONS

The lattice structure for this filter is shown in Figure 6.19b, where H (z−1 ) =
Y (z)/X(z).
Suppose that we select an allpass transfer function Hap (z−1 )

0.0154 − 0.0504z−1 − 0.0205z−2 + z−3

Hap (z−1 ) = (6.67)
1 − 0.2051z−1 − 0.0504z−2 + 0.0154z−3
and use the function [k,v]= tf2latc(num,den) we get the following outputs
for the vectors k and v:

k = −0.2145
−0.0473
0.0154
v = 0.0000
0.0000
0.0000
1.0000

Although this allpass transfer function has a numerator and a denominator and
hence is not an AR model, the lattice structure for realizing it is the same as the
lattice structure for an AR model in Figure 6.19b, but the output is V (z) and not
Y (z). Hence the allpass transfer function realized is Hap (z−1 ) = V (z)/X(z).
When we compare Figure 6.17 for the lattice structure for the third-order
FIR (MA) filter and Figure 6.19b for the lattice structure for the third-order
IIR all-pole (AR) or the allpass (AP) filter, carefully note the direction of the
multipliers and their signs, which are different. Also note that the output terminals
are different for the all-pole filter and allpass filters in Figure 6.19b.

6.5.1 MATLAB Program Used to Find Allpass Filters in Parallel

A MATLAB program lp2apx.m, developed by the author, to obtain the decom-
position of a lowpass filter as the sum of two allpass filters is given below.
The input data consist of the cutoff frequency Wp of the passband, the stopband
frequency Ws , the maximum attenuation Ap in the passband, and the minimum
attenuation As in the stopband, which are entered according to the choice of
the Butterworth, Chebyshev I, Chebyshev II, and elliptic lowpass filters. If the
order of the filter derived from the functions buttord, cheb1ord, cheb2ord
or ellipord is found to be an even integer, it is automatically increased by
one to make it an odd-order integer. The program immediately makes avail-
able the magnitude response of the specified lowpass filter G(z), the response
of the lowpass filter computed from the two allpass filters A1 (z) and A2 (z), and
also the highpass filter H (z). The coefficients of the numerator and denomina-
tor polynomials of A1 (z) and A2 (z) are obtained by typing A1N, A1D, A2N, and
 REALIZATION OF FIR AND IIR FILTERS USING MATLAB 335

A2D, respectively. So also the coefficients of the numerator and denominator of

G(z) can be obtained by typing b and a, respectively. If we type A1, A2, G,
or H, the program displays the transfer function as the ratio of two polynomi-
als. The program may not work successfully if the specifications call for a very
narrow transition band or a very high order for the filter, because of numerical
inaccuracy.
%This Matlab Program lp2apx.m obtains the allpass filters
% A1 (z) and A2 (z) from the lowpass (Butterworth, Chebyshev I
% and Elliptic) filters of odd order.
clear all
Wp=input(’Enter the passband cutoff frequency between 0
and 1’);
Ws=input(’Enter the stopband cutoff frequency Wp<Ws<1’);
Ap=input(’Enter the max.attenuation in the passband’);
As=input(’Enter the min.attenuation in the stopband’);
disp(’As ftype,type in (1) for Butterworth, (2) for
Chebyshev I, (3) Chebyshev II and (4) for Elliptic
filters’);
ftype=input(’’);
if ftype==1
disp(’Butterworth Lowpass Filter’);

[N,Wn]=buttord(Wp,Ws,Ap,As);
M=mod(N,2);
if M==0
N=N+1
end
[b,a]=butter(N,Wn);
end
if ftype==2
disp(’Chebyshev I Lowpass Filter’)
[N,Wn]=cheb1ord(Wp,Ws,Ap,As);
M=mod(N,2);
if M==0
N=N+1
end
[b,a]=cheby1(N,Ap,Wn);
end
if ftype==3
disp(’Chebyshev II Lowpass filter’)
[N,Wn]=cheb2ord(Wp,Ws,Ap,As);
M=mod(N,2);
if M==0
N=N+1
end
336 FILTER REALIZATIONS

[b,a]=cheby2(N,As,Wn);
end
if ftype==4
disp(’Elliptic Lowpass Filter’)
[N,Wn]=ellipord(Wp,Ws,Ap,As);
M=mod(N,2);
if M==0
N=N+1
end
[b,a]=ellip(N,Ap,As,Wn);
end
[h0,w]=freqz(b,a,256);
H0=abs(h0);
plot(w/pi,H0);grid
axis([0.0 1.0 0.0 1.0])
title(’MAGNITUDE OF SPECIFIED LP FILTER’)
ylabel(’Magnitude’)
xlabel(’Normalized frequency’)
% TO FIND Q(z)
k=sum(a)/sum(b);
b=b*k;
fliped a= fliplr(a);
%R(z)= Q2(z)=P2(z)-z^-N D(z^-1)D(z)
R=conv(b,b)-conv(a,fliped a);
% Calculate Q
Q(1)=R(1)^(0.5);
Q(2)=R(2)/(2*Q(1));
for n=2:N
term=0;
for k=1:n-1
term=Q(k+1)*Q(n-k+1)+term;
end
Q(n+1)=(R(n+1)-term)/(2*Q(1));
end
%Zeros of P+Q is calculated
j=1;
k=0;
P plus Q=b+Q;
zeros=roots(P plus Q);
for i=1:N
if abs(zeros(i))<1
zero in(j)=zeros(i);
j=j+1;
else
k=k+1;
 REALIZATION OF FIR AND IIR FILTERS USING MATLAB 337

zero out(k)=zeros(i);
end
end
A1N=poly(zero out);%Numerator of A 1(z)
A1D=fliplr(A1N); %Denominator of A 1(z)
A1=tf(A1N,A1D,1);
A2D=poly(zero in);%Denominator of A 2(z)
A2N=fliplr(A2D);%Numerator of A 2(z)
A2=tf(A2N,A2D,1);
G=0.5*(A1+A2); % LOWPASS FILTER FROM THE TWO ALLPASS FILTERS
[numlp,denlp]=tfdata(G,’v’);
[h1,w]=freqz(numlp,denlp,256);
H1=abs(h1);
figure
plot(w/pi,H1);grid
axis([0.0 1.0 0.0 1.0])
title(’MAGNITUDE OF LP FILTER FROM THE TWO ALLPASS FILTERS’)
ylabel(’Magnitude’)
xlabel(’Normalized frequency’)
H=0.5*(A1-A2); % HIGHPASS FILTER FROM THE TWO ALLPASS FILTERS
[numhp,denhp]=tfdata(H,’v’);
[h2,w]=freqz(numhp,denhp,256);
H2=abs(h2);
figure
plot(w/pi,H2);grid
axis([0.0 1.0 0.0 1.0])
title(’MAGNITUDE OF HP FILTER FROM THE TWO ALL PASS FILTERS’)
ylabel(’Magnitude’)
xlabel(’Normalized frequency’)
%END

Example 6.17

We illustrate the use of this program by taking the example of an elliptic lowpass
filter with the specifications Wp = 0.4, Ws = 0.6, Ap = 0.3, and As = 35, which
have been chosen only to highlight the passband and stopband responses. A
complete session for running this example is given below, including the three
magnitude response plots mentioned above:

Enter the passband cutoff frequency between 0 and 1 0.4

Enter the stopband cutoff frequency Wp<Ws<1 0.6
Enter the max.attenuation in the passband 0.3
Enter the min.attenuation in the stopband 35
As ftype, type in (1) for Butterworth, (2) for Chebyshev I,
(3) for Chebyshev II (4) for Elliptic filters
338 FILTER REALIZATIONS

4
Elliptic Lowpass Filter
N =
5
A1N
A1N =
1.0000 -1.3289 1.9650
A1D
A1D =
1.9650 -1.3289 1.0000
A2N
A2N =
-0.3667 1.1036 -0.9532 1.0000
A2D
A2D =
1.0000 -0.9532 1.1036 -0.3667
A1

Transfer function:
z^2 - 1.329 z + 1.965
-----------------------
1.965 z^2 - 1.329 z + 1

Sampling time: 1

A2

Transfer function:
-0.3667 z^3 + 1.104 z^2 - 0.9532 z + 1
--------------------------------------
z^3 - 0.9532 z^2 + 1.104 z - 0.3667

Sampling time: 1

Transfer function:
0.1397 z^5 + 0.1869 z^4 + 0.3145 z^3 + 0.3145 z^2 + 0.1869 z
+ 0.1397
-------------------------------------------------------------
1.965 z^5 - 3.202 z^4 + 4.435 z^3 - 3.14 z^2 + 1.591 z
- 0.3667

Sampling time: 1
 REALIZATION OF FIR AND IIR FILTERS USING MATLAB 339

Transfer function:
0.8603 z^5 - 2.469 z^4 + 4.021 z^3 - 4.021 z^2 + 2.469 z
- 0.8603
----------------------------------------------------------
1.965 z^5 - 3.202 z^4 + 4.435 z^3 - 3.14 z^2 + 1.591 z
- 0.3667

Sampling time: 1

We rewrite the transfer function G(z−1 ) in the following form for reference in
the next chapter:
 
0.1397 (1 + 1.337z−1 + 2.251z−2 + 2.251z−3 + 1.337z−4 + z−5 )
1.965 (1 − 1.629z−1 + 2.256z−2 − 1.597z−3 + 0.8096z−4 − 0.1866z−5 )
(6.68)
The magnitude response of the lowpass elliptic filter G(z), the magnitude
response of the parallel connection G(z) = 12 [A1 (z) + A2 (z)], and that of the
highpass filter H (z) = 12 [A1 (z) − A2 (z)] are shown in Figures 6.20, 6.21, and
6.22, respectively.
The two allpass filter functions (6.69) and (6.71) obtained in the example
above are expressed in the form of (6.70) and 6.72, respectively. The function
A1 (z) can be realized in the direct form, and A2 (z) can be realized in many
of the structures that we have already discussed, for example, the direct form,

Magnitude of specified LP filter

1

0.9

0.8

0.7

0.6
Magnitude

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency

Figure 6.20 Magnitude response of the elliptic lowpass filter G(z).

340 FILTER REALIZATIONS

Magnitude of LP filter from the two all pass filters

1

0.9

0.8

0.7

0.6
Magnitude

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency

Figure 6.21 Magnitude response of two allpass filters in parallel.

Magnitude of HP filter from the two all pass filters

1

0.9

0.8

0.7

0.6
Magnitude

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency

Figure 6.22 Magnitude response of a highpass filter from the two allpass filters in
parallel.
 REALIZATION OF FIR AND IIR FILTERS USING MATLAB 341

X(z) Y(z) X(z) −1

∑ z−1 ∑ ∑ z−1 ∑

−1

d1 d1
z−1
∑ Y(z)

(a) (b)

X(z) Y(z) ∑ z−1 ∑

z−1 ∑

z−1 −1
d1 d1
−1
Y(z)
∑ ∑

(c) (d)

Figure 6.23 First-order allpass structures.

parallel form, or lattice–ladder form. But the class of allpass functions of first
and second orders can be realized by many structures that employ the fewest
multipliers [1]. A few examples of first-order and second-order allpass filters
are shown in Figures 6.23 and 6.24, respectively. Their transfer functions are
respectively given by

d1 + z−1
AI (z) =
1 + d1 z−1
d1 d2 + d1 z−1 + z−2
AII (z) =
1 + d1 z−1 + d1 d2 z−2
We choose the simpler structure of second-order allpass filter from Figure 6.24a
for A1 and A2 (z), which requires fewer delay elements than do the remaining four
second-order structures. When these two allpass filters are connected in parallel
(as shown in Fig. 6.16), we get the structure shown in Figure 6.25 for the transfer
function G(z) of the fifth-order elliptic lowpass filter chosen in this example:

z2 − 1.329z + 1.965
A1 (z) = (6.69)
1.965z2 − 1.329z + 1
1 − 1.329z−1 + 1.965z−2
=
1.965 − 1.329z−1 + z−2
0.5089 − 0.6763z−1 + z−2
= (6.70)
1 − 0.6763z−1 + 0.5089z−2
342 FILTER REALIZATIONS

X(z) Y(z)
X(z) Y(z) z−1 z−1 Σ
Σ z−1 z−1 Σ

−1
−1
d2 d1
d1 d2 Σ Σ
Σ Σ
−1

z−1 z−1
(a) (b)

−1
X(z) Y(z)
z−1 Σ z−1 Σ
X(z) Y(z)
z−1 Σ z−1 Σ
−1 −d1
Σ d1
Σ
Σ Σ
Σ Σ

z−1
z−1 d2
−1
Σ
d2 Σ
(c) (d)

Figure 6.24 Second-order allpass structures.

1/2 Y(z)
Σ z−1 z−1 Σ Σ

−1

Σ Σ
−0.6763 −0.7525

X(z) 1/2 V(z)

Σ z−1 z−1 Σ Σ z−1 Σ Σ
−1
−1

Σ Σ −0.4165
z−1
−0.5367 −1.640

Figure 6.25 A fifth-order elliptic lowpass IIR filter realized as the parallel connection
of two allpass structures.
 REALIZATION OF FIR AND IIR FILTERS USING MATLAB 343

−0.3667z3 + 1.104z2 − 0.9532z + 1

A2 (z) = (6.71)
z3 − 0.9532z2 + 1.104z − 0.3667
−0.3667 + 1.104z−1 − 0.9532z−2 + z−3
= (6.72)
1 − 0.9532z−1 + 1.104z−2 − 0.3667z−3

We express the third-order allpass function A2 (z) in product form as shown

in (6.73), using a second-order allpass structure from Figure 6.24a connected
in cascade with a first-order allpass structure from Figure 6.23a to realize the
third-order allpass filter A2 (z). It is then connected in parallel with the second-
order allpass filter A1 (z). The structure for the lowpass elliptic filter G(z) =
2 [A1 (z) + A2 (z)] is shown in Figure 6.25. We obtain
1

  
0.8805 − 0.5368z−1 + z−2 −0.4165 + z−1
A2 (z) = (6.73)
1 − 0.5367z−1 + 0.8805z−2 1 − 0.4165z−1

Instead of designing the allpass functions found in Figures 6.24 and 6.25, we
can design them in the form of lattice allpass structures as described earlier.
The lattice coefficients for the second-order filter A1 (z) and the third-order filter
A2 (z) are found by using the MATLAB function [k,v] = tf2latc(num,den),
and after obtaining the lattice structures for them, they are connected in parallel
as shown in Figure 6.27.

%Design of the second order filter A1 (z)

A1num=[0.5089 -0.6763 1];
A1den=[1 -0.6763 0.5089];
[K1,V1]=tf2latc(A1num,A1den)

%Lattice and ladder coefficients for A1 (z)

K1 =
-0.4482
0.5089
V1 =
0
0
1

%Design of third order filter A2 (z)

A2num=[-0.3667 1.104 -0.9532 1];
A2den=fliplr(A2num)
[K2,V2]=tf2latc(A2num,A2den)
%Lattice and ladder coefficients for A2 (z)

K2 =
-0.3385
344 FILTER REALIZATIONS

0.8717
-0.3667

V2 =
0
0
0
1

The circuit realizing the third-order transfer function A2 (z) in the form of a
lattice–ladder structure is shown in Figure 6.26, where the values of V0 = V1 =
V2 = 0 as shown by the vector V2 above.
The circuit realizing the fifth-order lowpass elliptic filter as the parallel connec-
tion of two allpass filters A1 (z) and A2 (z), each realized by the lattice structures,
is shown in Figure 6.27.
Now let us compare the different circuits that we have designed to realize
a lowpass fifth-order, IIR filter. All of these circuits have been designed to
meet the following same specifications—Wp = 0.4, Ws = 0.6, Ap = 0.3, and
As = 35—and have been realized by a cascade connection, a parallel con-
nection, and a lattice–ladder connection as shown in Figures 6.14, 6.15, and
6.19, respectively. They use more than the minimum number of five multipliers,
whereas the lattice-coupled allpass filter shown in Figure 6.25 uses five multipli-
ers—disregarding the multipliers with a gain of −1 or 12 because they represent

X(z)
Σ Σ Σ

−K1
−K3 −K2

K3 K2 K1

Σ z−1 Σ z−1 Σ z−1

V0
V3 V2 V1

Σ Σ Σ Y(z)

Figure 6.26 A third-order allpass filter A2 (z) realized as a lattice–ladder structure.

 REALIZATION OF FIR AND IIR FILTERS USING MATLAB 345

Σ Σ

0.5089 −0.5089 −0.4482 0.4482

Σ z−1 Σ z−1
0.5
X(z)

Y(z)
Σ

Σ Σ Σ

0.8717 −0.3385
−0.3667 0.3667 −0.8717 0.3385

Σ z−1 Σ z−1 Σ z−1

Figure 6.27 Two lattice–ladder allpass structures connected in parallel to realize a fifth-
order lowpass elliptic filter.

minor operations on binary numbers. The direct-form IIR filter for a fifth-order
filter would also require more than five multipliers, whereas the filter shown in
Figure 6.27, which has lattice–ladder, coupled allpass filters, requires 10 mul-
tipliers. Therefore we conclude that the parallel connection of allpass filters as
shown in Figure 6.25 requires a minimum number of delay elements and thus
offers an advantage over the other structures.
The realization of IIR filters as a parallel connection of allpass filters has
another advantage, as explained below. It was pointed out that the magnitude
response of allpass filters does not change when the multiplier constants are
quantized to finite wordlength. The other advantage is that there are many struc-
tures for realizing allpass filters that contain a minimum number of multipliers
(and delay elements). In the method of realizing the lowpass filter by a connec-
tion of two allpass filters in parallel, we used Property (6.4) in Equation (6.42)
which is reproduced below:
     
G(ej ω ) = 1 ej θ1 (ω) + ej θ2 (ω)  = 1 1 + ej (θ1 (ω)−θ2 (ω))  ≤ 1
2 2

This shows that the lowpass filter containing the two allpass structures in parallel
has a magnitude response equal to or less than unity. The magnitude response
in Figure 6.28 illustrates this property in the passband and attains the maximum
value at three frequencies in the passband, which are marked by arrows. As
346 FILTER REALIZATIONS

Magnitude response
1.01

0.99

0.98

0.97
Magnitude

0.96

0.95

0.94

0.93

0.92

0.91

1 2 3 4 5 6 7 8 9 10
Frequency (kHz)

Figure 6.28 Magnified magnitude response of an elliptic lowpass filter.

long as the allpass filters maintain a constant magnitude at all frequencies and
remain stable, as their multiplier constants change in wordlength, the magnitude
response of the lowpass filter cannot exceed this constant at these three frequen-
cies, where the derivative of the magnitude response is zero. Hence for small
changes in wordlength (e.g., by 1 or 2 bits), the change in magnitude response
at these frequencies is almost zero. At other frequencies in the passband, the
change in magnitude is also expected to be small, if not zero. Simulation of
their performance with small changes in wordlength has verified that the change
in their magnitude response is significantly smaller than that displayed by the
other structures. This shows that the structure of allpass filters in parallel has
many advantages compared to the other structures that have been proposed for
realizing IIR filters. In the next chapter, where the effect of finite wordlength is
studied in greater detail, the structure for allpass filters in parallel will be called
lattice-coupled allpass structure. But these structures can be used to design only
lowpass filters (of odd order) whereas the lattice and lattice–ladder structures
can realize any transfer function in general.

6.6 SUMMARY

When we have obtained the transfer functions of FIR and IIR filters that
approximate a given set of specifications—as explained in the previous two
chapters—our next step is to choose the best structures that would meet some
 PROBLEMS 347

important criteria before the algorithm in the time domain can be programmed
or a filter can be designed and built in hardware. It is obvious that the algorithm
for implementing a filter will depend on the particular structure being considered
to realize it. Under ideal assumptions that the magnitude of the input signals and
the values of the multiplier constants are available with infinite precision, any
one of the several alternative structures will realize the transfer function. But
when they are expressed with a finite number of bits, their actual performance
may be quite different, particularly when they are represented by a fixed-point
binary representation. So it is necessary to investigate in great detail their perfor-
mance in the time domain and the frequency domain and compare them. Some
of the performance criteria used for comparison are the effective degradation in
the frequency response, the stability and potential limit cycles, complexity of
the algorithm flow control, number of multiplications, and additions per sample
output. Extensive simulation on a computer is essential to address these issues,
before we chose a few structures for further investigation.
In this chapter, we discussed several structures to realize the FIR and IIR filters
and commented on the effects of finite wordlength. More detailed discussion of
this criterion and other issues will be included in the next chapter.

PROBLEMS

6.1 Draw the direct form and the cascade form of the FIR filter with the
following transfer function:

H (z−1 ) = 4(1 + 0.6z−1 − 0.5z−2 )(1 − 0.25z−1 + 0.9z−2 )

6.2 Find the polyphase structure for the FIR filter in Problem 6.1 and its
transpose.
6.3 Determine the transpose of the direct-form structure realizing the FIR filter

H (z−1 ) = 1 + z−1 − 0.5z−2 + 0.02z−3 + 0.003z−5

6.4 Determine the polyphase structure for the FIR filter given in Problem 6.3.
6.5 Find the polyphase structure for the FIR filter

H (z−1 ) = 1 + 0.5z−1 + 0.4z−2 − 0.6z−3 + z−4 + 1.2z−5 + 0.2z−6

6.6 Obtain the transfer functions H1 (z) = Y (z)/X(z) and H2 (z) = G(z)/X(z)
of the lattice circuit shown in Figure 6.29.
6.7 Draw the direct form and transpose of the circuit shown in Figure 6.29.
6.8 (a) Derive the transfer function H1 (z) = Y (z)/X(z) of the lattice structure
shown in Figure 6.30.
348 FILTER REALIZATIONS

X(z) Σ Σ Y(z)

K1 K2
K1 K2

z−1 Σ z−1 Σ G(z)

Figure 6.29 Problem 6.6.

X(z) Σ Σ Y(z)

K2 K1

−K2 −K1

G2(z) Σ z−1 Σ z−1

Figure 6.30 Problem 6.8.

(b) Derive the transfer function G(z) = G2 (z)/Y (z) and show that the
transfer function H2 (z) = G2 (z)/X(z) is an allpass function.
(c) If the transfer function for the lattice structure shown in Figure 6.30
is H1 (z) = 1/(1 + 1.38z−1 + 1.3z−2 ), what are the values of K1 and
K2 ?
6.9 Draw the transpose of the lattice structure shown in Figure 6.29.
6.10 Plot the unit pulse response of the filter shown in Figure 6.31a,b.
6.11 Derive the transfer function H (z) = Y (z)/X(z) for the structure shown in
Figure 6.32.
6.12 Draw the transpose of the structure shown in Figure 6.32.
6.13 (a) Draw the circuit in a parallel structure, to realize the following transfer
function H (z−1 ) and find its inverse z transform h(n):

(1 + 0.2z−1 )z−2
H (z−1 ) =
(1 − 0.6z−1 + 0.25z−2 )(1 + 0.4z−1 )

(b) Derive the inverse z transform of

(1 + 0.2z)z2
H (z) =
(1 − 0.6z + 0.25z2 )(1 + 0.4z)
 PROBLEMS 349

X(n) z−1 z−1 z−1 z−1 Σ y(n)

1.0
2.0

−1.0

4.0
(a)

X(n)
z−1 Σ z−1 Σ z−1 Σ z−1 Σ y(n)

2.0 −1.0 4.0

(b)

Figure 6.31 Problem 6.10.

X(z) z−1 z−1 z−1

z−1

Σ z−1 z−1 z−1

Σ Σ Σ

1.0 2.0 −1.0 4.0

Y(z)

Figure 6.32 Problem 6.12.

350 FILTER REALIZATIONS

6.14 Find the z transform X(z) of [(0.8)n − (0.4)n ] u(n). What is the inverse
z transform of X(−z)?
6.15 Draw the digital filter circuit in both cascade and parallel forms to realize
the following transfer function:

0.44z2 + 0.36z + 0.02

H (z) =
(z2 + 0.8z + 0.5)(z − 0.4)

6.16 Draw the direct form I, direct form II, cascade and parallel structures for
the transfer function

z−1
H (z−1 ) =
(1 + 0.2z−1 )(1 + 0.6z−1 + 0.2z−2 )

6.17 Draw the transpose of the structures obtained in Problem 6.16.

6.18 Given a transfer function

1 + 0.1z−1
H (z−1 ) =
(1 + 0.3z−1 )(1 + 0.5z−1 )

obtain the cascade and parallel structures to realize it. Draw their transpose
structures also.
6.19 Draw the direct form II structure for the structure shown in Figure 6.33.
Find the unit pulse response of this structure for r = 0.6 and θ = π/5.

r cos q

r sin q
Σ z−1 Y(n)

−r sin q r sin q

X(n)

Σ z−1
−r cos q

r cos q

Figure 6.33 Problem 6.19.

 PROBLEMS 351

6.20 Obtain as many structures as you can to realize the following transfer
function:
(z + 0.2)
H (z) =
(z + 0.1)(z + 0.4)(z2 + 0.5z + 0.06)

6.21 Determine the cascade and parallel structures for the transfer function

(1 + 0.3z−1 )
H (z) = π π
(1 − 0.3z−1 )(1 − 0.5ej 3 z−1 )(1 − 0.5e−j 3 z−1 )

MATLAB Problems
6.22 Find the direct-form and cascade structures realizing the FIR filter function
H (z−1 )

H (z−1 ) = 1 + 0.2z−1 + 0.3z−2 − 0.4z−3 − 0.5z−5 + 0.6z−6 + z−7

6.23 Find the direct-form and cascade structures realizing the FIR filters

H1 (z−1 ) = 1 − 0.5z−1 + 0.3z−2 + 0.1z−3 + 0.02z−4 − 0.05z−5

and

H2 (z−1 ) = 1.0 + 0.8z−1 + 0.8z−2 + 0.08z−3 + 0.01z−4

6.24 Determine the lattice structures to realize the FIR filters in Problem 6.23.
6.25 Find the direct form I and the cascade structures to realize the following
IIR filters:
1 − 0.25z−1 z−1 1 + 0.4z−1
H1 (z) = + +
1 + 0.9z−1 1 + 0.5z−1 1 + 0.2z−1 + 0.08z−2
1 + 0.1z−1 + z−2 − 0.2z−3 4z−1
H2 (z−1 ) = +
1 + z−1 + 0.24z−2 (1 − 0.8z−1 )(1 − 0.4z−1 )

6.26 Find the structure in the parallel and cascade connections to realize the
following filters:
  
−1 z−1 z−1 z−1 z−1
H1 (z ) = + +
1 − 0.5z−1 1 + 0.5z−1 1 − 0.2z−1 1 + 0.2z−1
 
1 − 0.25z−1 z−1
H2 (z−1 ) = +
1 + 0.9z−1 1 + 0.5z−1
 
1 + z−1 2z−2
× +
1 + 0.4z−1 1 + 0.6z−1 + 0.6z−2
352 FILTER REALIZATIONS

6.27 Draw the structures in a cascade connection and in a parallel connection

to realize the following IIR filter:

1.0 + 0.0141z−1 + 0.0284z−2

H (z−1 ) =
1 − 2.9061z−1 + 4.2077z−2 − 3.45412z−3
+1.6046z−4 − 0.3365z−5
6.28 Draw the lattice structures to realize the following all-pole IIR filters. Are
they stable?
1
H1 (z−1 ) =
1.0 + 0.3z−1 − 0.04z−2 − 0.13z−3 − 0.02z−4
1
H2 (z−1 ) =
1.0 + 2.3z−1 + 0.16z−2 − 0.25z−3 − 0.26z−4
1
H3 (z−1 ) =
1.0 + 0.6z−1 − 6.01z−2 − 3.75z−3 − 1.5z−4
6.29 Draw the lattice structure for the following allpass filters:
0.01 − 0.75z−1 + z−2
A1 (z−1 ) =
1 − 0.75z−1 + 0.01z−2
−0.12 − 0.06z−1 + 0.1z−2 + z−3
A2 (z−1 ) =
1 + 0.1z−1 − 0.06z−2 − 0.12z−3
6.30 Find the lattice–ladder structure for the following IIR filter:

1.2 + z−1
H (z−1 ) =
1.0 + 1.1z−1 + 0.5z−2 + 0.1z−3
6.31 Find the lattice–ladder structure for the following IIR filter:

0.01 − 0.75z−1
H (z−1 ) =
1 − 0.75z−1 + 0.01z−2
6.32 Determine the lattice–ladder structure for the following IIR filter:

−0.12 − 0.06z−1 + 0.1z−2 + z−3

H (z−1 ) =
1 + 0.1z−1 − 0.06z−2 − 0.12z−3
6.33 Find the lattice–ladder structure for the following IIR filters:

0.9 + 0.7z−1 − 0.6z−2

H1 (z−1 ) =
1.0 + z−1 + 0.47z−2 + 0.098z−3 + 0.006z−4
1.0 + 0.5z−1 + 0.5z−2
H2 (z−1 ) =
1.0 + z−1 + 0.47z−2 + 0.098z−3 + 0.006z−4
1.0 + z−1 + z−2
H3 (z−1 ) =
1.0 + z−1 + 0.52z−2 + 0.12z−3 + 0.016z−4
 REFERENCES 353

6.34 Design a lowpass Butterworth filter with Wp = 0.3, Ws = 0.6, Ap = 0.25,

As = 55 in an lattice-coupled allpass structure. Give the magnitude response
of the Butterworth filter and the lattice-coupled allpass filter on the same plot.
6.35 Design a Chebyshev lowpass IIR filter with the specifications from Prob-
lem 6.34 in a lattice-coupled allpass structure. Give the magnitude response
of the Chebyshev filter and the lattice-coupled allpass filter structure for
the same filter on the same plot.
6.36 Design an elliptic lowpass IIR filter Wp = 0.25, Ws = 0.5, Ap = 0.20,
As = 60, in a lattice-coupled allpass structure. Give the magnitude
response of the elliptic filter and the lattice-coupled allpass filter for the
same filter on the same plot.

REFERENCES

1. S. K. Mitra and K. Hirano, Digital allpass networks, IEEE Trans. Circuits Syst. CAS-21,
688–700 (1974).
2. J. G. Proakis and D. G. Manolakis, Digital Signal Processing, Prentice-Hall, 1996.
3. S. K. Mitra, Digital Signal Processing—A Computer Based Approach. McGraw-Hill,
2001.
4. S. K. Mitra and J. F. Kaiser, eds., Handbook for Digital Signal Processing, Wiley-
Interscience, 1993.
5. B. A. Shenoi, Magnitude and Delay Approximation of 1-D and 2-D Digital Filters,
Springer-Verlag, 1999.
 CHAPTER 7

Quantized Filter Analysis

7.1 INTRODUCTION

The analysis and design of discrete-time systems, digital filters, and their realiza-
tions, computation of DFT-IDFT, and so on discussed in the previous chapters
of this book were carried out by using mostly the functions in the Signal Pro-
cessing Toolbox working in the MATLAB environment, and the computations
were carried out with double precision. This means that all the data representing
the values of the input signal, coefficients of the filters, or the values of the unit
impulse response, and so forth were represented with 64 bits; therefore, these
numbers have a range approximately between 10−308 and 10308 and a precision
of ∼2−52 = 2.22 × 10−6 . Obviously this range is so large and the precision with
which the numbers are expressed is so small that the numbers can be assumed to
have almost “infinite precision.” Once these digital filters and DFT-IDFT have
been obtained by the procedures described so far, they can be further analyzed
by mainframe computers, workstations, and PCs under “infinite precision.” But
when the algorithms describing the digital filters and FFT computations have
to be implemented as hardware in the form of special-purpose microprocessors
or application-specific integrated circuits (ASICs) or the digital signal processor
(DSP) chip, many practical considerations and constraints come into play. The
registers used in these hardware systems, to store the numbers have finite length,
and the memory capacity required for processing the data is determined by the
number of bits—also called the wordlength —chosen for storing the data. More
memory means more power consumption and hence the need to minimize the
wordlength. In microprocessors and DSP chips and even in workstations and PCs,
we would like to use registers with as few bits as possible and yet obtain high
computational speed, low power, and low cost. But such portable devices such as
cell phones and personal digital assistants (PDAs) have a limited amount of mem-
ory, containing batteries with low voltage and short duration of power supply.
These constraints become more severe in other devices such as digital hearing
aids and biomedical probes embedded in capsules to be swallowed. So there is a

Introduction to Digital Signal Processing and Filter Design, by B. A. Shenoi

Copyright © 2006 John Wiley & Sons, Inc.

354
 FILTER DESIGN–ANALYSIS TOOL 355

great demand for designing digital filters and systems in which they are embed-
ded, with the lowest possible number of bits to represent the data or to store the
data in their registers. When the filters are built with registers of finite length and
the analog-to-digital converters (ADCs) are designed to operate at increasingly
high sampling rates, thereby reducing the number of bits with which the samples
of the input signal are represented, the frequency response of the filters and the
results of DFT-IDFT computations via the FFT are expected to differ from those
designed with “infinite precision.” This process of representing the data with a
finite number of bits is known as quantization, which occurs at several points
in the structure chosen to realize the filter or the steps in the FFT computation
of the DFT-IDFT. As pointed out in the previous chapter, a vast number of
structures are available to realize a given transfer function, when we assume infi-
nite precision. But when we design the hardware with registers of finite length to
implement their corresponding difference equation, the effect of finite wordlength
is highly dependent on the structure. Therefore we find it necessary to analyze
this effect for a large number of structures. This analysis is further compounded
by the fact that quantization can be carried out in several ways and the arithmetic
operations of addition and multiplication of numbers with finite precision yield
results that are influenced by the way that these numbers are quantized.
In this chapter, we discuss a new MATLAB toolbox called FDA Tool avail-
able1 for analyzing and designing the filters with a finite number of bits for the
wordlength. The different form of representing binary numbers and the results of
adding and multiplying such numbers will be explained in a later section of this
chapter. The third factor that influences the deviation of filter performance from
the ideal case is the choice of FIR or IIR filter. The type of approximation chosen
for obtaining the desired frequency response is another factor that also influences
the effect of finite wordlength. We discuss the effects of all these factors in this
chapter, illustrating their influence by means of a design example.

7.2 FILTER DESIGN–ANALYSIS TOOL

An enormous amount of research has been carried out to address these problems,
but analyzing the effects of quantization on the performance of digital filters
and systems is not well illustrated by specific examples. Although there is no
analytical method available at present to design or analyze a filter with finite
precision, some useful insight can be obtained from the research work, which
serves as a guideline in making preliminary decisions on the choice of suitable
structures and quantization forms. Any student interested in this research work
should read the material on finite wordlength effects found in other textbooks
[1,2,4]. In this chapter, we discuss the software for filter design and analysis
that has been developed by The MathWorks to address the abovementioned

1
MATLAB and its Signal Processing Toolbox are found in computer systems of many schools and
universities but the FDA Tool may not be available in all of them.
356 QUANTIZED FILTER ANALYSIS

problem2 . This FDA Tool finite design–analysis (FDA) tool, found in the Filter
Design Toolbox, works in conjunction with the Signal Processing (SP) Toolbox.
Unlike the SP Toolbox, the FDA Tool has been developed by making extensive use
of the object-oriented programming capability of MATLAB, and the syntax for the
functions available in the FDA Tool is different from the syntax for the functions
we find in MATLAB and the SP Toolbox. When we log on to MATLAB and type
fdatool, we get two screens on display. On one screen, we type the fdatool
functions as command lines to design and analyze quantized filters, whereas the
other screen is a graphical user interface (GUI) to serve the same purpose. The
GUI window shown in Figure 7.1a displays a dialog box with an immense array
of design options as explained below.
First we design a filter with double precision on the GUI window using the
FDA Tool or on the command window using the Signal Processing Toolbox and
then import it into the GUI window. In the dialog box for the FDA Tool, we can
choose the following options under the Filter Type panel:
1. Lowpass
2. Highpass
3. Bandpass
4. Bandstop
5. Differentiator. By clicking the arrow on the tab for this feature, we get
the following additional options.
6. Hilbert transformer
7. Multiband
8. Arbitrary magnitude
9. Raised cosine
10. Arbitrary group delay
11. Half-band lowpass
12. Half-band highpass
13. Nyquist
Below the Filter Type panel is the panel for the design method. When the
button for IIR filter is clicked, the dropdown list gives us the following options
specifying the type of frequency response:
• Butterworth
• Chebyshev I
• Chebyshev II
• Elliptic
• Least-pth norm
• Constrained least-pth norm
2
The author acknowledges that the material on the FDA Tool described in this chapter is based on
the Help Manual for Filter Design Toolbox found in MATLAB version 6.5.
 FILTER DESIGN–ANALYSIS TOOL 357

(a)

(b)

Figure 7.1 Screen capture of fdatool window: (a) window for filter design;
(b) window for quantization analysis.

the following options are available for the FIR filter:

• Equiripple
• Least squares
• Window
• Maximally flat
• Least-pth norm
• Constrained equiripple
358 QUANTIZED FILTER ANALYSIS

To the right of the panel for design method is the one for filter order. We can
either specify the order of the filter or let the program compute the minimum order
(by use of SP Tool functions Chebord, Buttord, etc.). Remember to choose an
odd order for the lowpass filter when it is to be designed as a parallel connection
of two allpass filters, if an even number is given as the minimum order. Below
this panel is the panel for other options, which are available depending on the
abovementioned inputs. For example, if we choose a FIR filter with the window
option, this panel displays an option for the windows that we can choose. By
clicking the button for the windows, we get a dropdown list of more than 10
windows. To the right of this panel are two panels that we use to specify the
frequency specifications, that is, to specify the sampling frequency, cutoff fre-
quencies for the passband and stopband, the magnitude in the passband(s) and
stopband(s), and so on depending on the type of filter and the design method
chosen. These can be expressed in hertz, kilohertz, megahertz, gigahertz, or nor-
malized frequency. The magnitude can be expressed in decibels, with magnitude
squared or actual magnitude as displayed when we click Analysis in the main
menu bar and then click the option Frequency Specifications in the drop-
down list. The frequency specifications are displayed in the Analysis panel,
which is above the panel for frequency specifications, when we start with the
filter design.
The options available under any of these categories are dependent on the
other options chosen. All the FDA Tool functions, which are also the functions
of the SP Tool, are called overloaded functions. After all the design options are
chosen, we click the Design Filter button at the bottom of the dialog box. The
program designs the filter and displays the magnitude response of the filter in the
Analysis area. But it is only a default choice, and by clicking the appropriate
icons shown above this area, the Analysis area displays one of the following
features:

• Magnitude response
• Phase response
• Magnitude and phase response
• Group delay response
• Impulse response
• Step response
• Pole–zero plot
• Filter coefficients

This information can also be displayed by clicking the Analysis button in the
main menu bar, and choosing the information we wish to display in the Anal-
ysis area. We can also choose some additional information, for example, by
clicking the Analysis Parameters. At the bottom of this dropdown list is the
option Full View Analysis. When this is chosen, whatever is displayed in the
Analysis area is shown in a new panel of larger dimensions with features that
 FILTER DESIGN–ANALYSIS TOOL 359

are available in a figure displayed under the SP Tool. For example, by clicking the
Edit button and then selecting either Figure Properties, Axis Properties,
or Current Object Properties, the Property Editor becomes active and
properties of these three objects can be modified.
Finally, we look at the first panel titled Current Filter Information.
This lists the structure, order, and number of sections of the filter that we have
designed. Below this information, it indicates whether the filter is stable and
points out whether the source is the designed filter (i.e., reference filter designed
with double precision) or the quantized filter with a finite wordlength. The default
structure for the IIR reference filter is a cascade connection of second-order
sections, and for the FIR filter, it is the direct form. When we have completed
the design of the reference filter with double precision, we verify whether it
meets the desired specification, and if we wish, we can convert the structure of
the reference filter to any one of the other types listed below. We click the Edit
button on the main menu and then the Convert Structure button. A dropdown
list shows the structures to which we can convert from the default structure or
the one that we have already converted.
For IIR filters, the structures are

1. Direct form I
2. Direct form II
3. Direct form I transposed
4. Direct form II transposed
5. Lattice ARMA
6. Lattice-coupled allpass
7. Lattice-coupled allpass—power complementary
8. State space

Items 6 and 7 in this list refer to structures of the two allpass networks in
parallel as described in Chapter 6, with transfer functions G(z) = 12 [A1 (z) +
A2 (z)] and H (z) = 12 [A1 (z) − A2 (z)], respectively. The allpass filters A1 (z) and
A2 (z) are realized in the form of lattice allpass structures like the one shown
in Figure 6.19b. The MA and AR structures are considered special cases of the
lattice ARMA structure, which are also discussed in Chapter 6.
For FIR filters, the options for the structures are

• Direct-form FIR
• Direct-form FIR transposed
• Direct-form symmetric FIR

When we have converted to a new structure, the information that can be

displayed in the Analysis area, like the coefficients of the filter, changes. We also
like to point out that any one of the lowpass, highpass, bandpass, and bandstop
filters that we have designed can be converted to any other type, by clicking
360 QUANTIZED FILTER ANALYSIS

the first icon on the left-hand bar in the dialog box and adding the frequency
specifications for the new filter.

7.3 QUANTIZED FILTER ANALYSIS

When we have finished the analysis of the reference filter, we can move to
construct the quantized filter as an object, by clicking the last icon on the bar
above the Analysis area and the second icon on the left-hand bar, which sets
the quantization parameters. The panel below the Analysis area now changes
as shown in Figure 7.1b. We can construct three objects inside the FDA Tool:
qfilt, qfft, and quantizer. Each of them has several properties, and these
properties have values, which may be strings or numerical values. Currently
we use the objects qfilt and quantizer to analyze the performance of the
reference filter when it is quantized. When we click the Turn Quantization
On button and the Set Quantization Parameters icon, we can choose the
quantization parameters for the coefficients of the filter. Quantization of the filter
coefficients alone are sufficient for finding the finite wordlength effect on the
magnitude response, phase response, and group delay response of the quantized
filter, which for comparison with the response of the reference filter displayed
in the Analysis area. Quantization of the other data listed below are necessary
when we have to filter an input signal:

• The input signal

• The output signal
• The multiplicand: the value of the signal that is multiplied by the multiplier.
• The product of the multiplicand and the multiplier constant
• The output signal

The object quantizer is used to convert each of these data, and this object has
four properties: Mode, Round Mode, Overflow mode, and Format. In order to
understand the values of these properties, it is necessary to review and understand
the binary representation of numbers and the different results of adding them and
multiplying them. These will be discussed next.

7.4 BINARY NUMBERS AND ARITHMETIC

Numbers representing the values of the signal, the coefficients of both the filter
and the difference equation or the recursive algorithm and other properties cor-
responding to the structure for the filter are represented in binary form. They are
based on the radix of 2 and therefore consist of only two binary digits, 0 and 1,
which are more commonly known as bits, just as the decimal numbers based on a
radix of 10 have 10 decimal numbers from 0 to 9. Placement of the bits in a string
determines the binary number as illustrated by the example x2 = 1001 1010,
 BINARY NUMBERS AND ARITHMETIC 361

which is equivalent to x10 = 1 × 20 + 1 × 23 + 2−1 + 2−3 = 9.625. In this dis-

cussion of binary number representation, we have used the symbol  to separate
the integer part and the fractional part and the subscripts 2 and 10 to denote the
binary number and the decimal number. Another example given by

x2 = b2 b1 b0 b−1 b−2 b−3 b−4 (7.1)

has a decimal value computed as

−1 −2 −3
x10 = b22 + b11 + b00 + b−1 + b−2 + b−3 + b−4
4
(7.2)

where the bits b2 , b1 , b0 , b−1 , b−2 , b−3 , b−4 are either 1 or 0. In general, when
x2 is represented as

x2 = bI −1 bI −2 · · · b1 b0 b−1 b−2 · · · b−F (7.3)

the decimal number has a value given by

I −1

x10 = bi 2i (7.4)
i=−F

In the binary representation (7.3), the integer part contains I bits and the bit bI −1
at the leftmost position is called the most significant bit (MSB); the fractional
part contains F bits, and the bit b−F at the rightmost position is called the least
significant bit (LSB). This can only represent the magnitude of positive numbers
and is known as the unsigned fixed-point binary number. In order to represent
positive as well as negative numbers, one more bit called the sign bit is added to
the left of the MSB. The sign bit, represented by the symbol s in (7.5), assigns
a negative sign when this bit is 1 and a positive sign when it is 0. So it becomes
a signed magnitude fixed-point binary number. Therefore a signed magnitude
number x2 = 11001 1010 is x10 = −9.625. In general, the signed magnitude
fixed-point number is given by

I −1

x10 = (−1) s
bi 2i (7.5)
i=−F

and the total number of bits is called the wordlength w = 1 + I + F . When

two signed magnitude numbers with widely different values for the integer part
and/or the fractional part have to be added, it is not easy to program the adders
in the digital hardware to implement this operation. So it is common practice
to choose I = 0, keeping the sign bit and the bits for the fractional part only
so that F = w − 1 in the signed magnitude fixed-point representation. But when
two numbers larger than 0.5 in decimal value are added, their sum is larger
than 1, and this cannot be represented by the format shown above, where I = 0.
362 QUANTIZED FILTER ANALYSIS

So two other form of representing the numbers are more commonly used: the
one’s-complement and two’s-complement forms (also termed one-complementary
and two-complementary forms) for representing the signed magnitude fixed-point
numbers. In the one’s-complement form, the bits of the fractional part are replaced
by their complement, that is, the ones are replaced by zeros and vice versa. By
adding a one as the least significant bit to the one’s-complement form, we get
the two’s-complement form of binary representation; the sign bit is retained in
both forms. But it must be observed that when the binary number is positive, the
signed magnitude form, one’s-complement form, and two’s-complement form are
the same.

Example 7.1

Given: x2 = 0 1100 is the 5-bit, signed magnitude fixed-point number equal to

x10 = +2−1 + 2−2 = 0.75 and v2 = 1 1100 is equal to v10 = −0.75. The one’s
complement of v2 = 1 1100 is 1 0011, whereas the two’s complement of v2 is
1 0011 + 0001 = 1 0100.
The values that can be represented by the signed magnitude fixed-point repre-
sentation range from −2w−F −1 to 2w−F −1 − 2−F . In order to increase the range
of numbers that can be represented, two more formats are available: the floating-
point and block floating-point representations. The floating-point representation
of a binary number is of the form

X10 = (1)s M(2E ) (7.6)

where M is the mantissa, which is usually represented by a signed magnitude,

fixed-point binary number, and E is a positive- or negative-valued integer with
E bits and is called the exponent. To get both positive and negative exponents,
the bias is provided by an integer, usually the bias is chosen as e7 − 1 = 127
when the exponent E is 8 bits or e10 − 1 = 1023 when E is 11 bits. Without
the bias, an 8-bit integer number varies from 0 to 255, but with a bias of 127,
the exponent varies from −127 to 127. Also the magnitude of the fractional part
F is limited to 0 ≤ M < 1. In order to increase the range of the mantissa, one
more bit is added to the most significant bit of F so that it is represented as
(1.F ). Now it is assumed to be normalized, but this bit is not counted in the total
wordlength.
The IEEE 754-1985 standard for representing floating-point numbers is the
most common standard used in DSP processors. It uses a single-precision format
with 32 bits and a double-precision format with 64 bits.
The single-precision floating point number is given by

X10 = (−1)s (1.F )2E−127 (7.7)

According to this standard, the (32-bit) single-precision, floating-point number

uses one sign bit, 8 bits for the exponent, and 23 bits for the fractional part
 BINARY NUMBERS AND ARITHMETIC 363

b8 b7 b0 b−1 b−23
s E (8 bits) F (23 bits)

(a)

b11 b10 b0 b−1 b−52

s E (11 bits) F (52 bits)

(b)

Figure 7.2 IEEE format of bits for the 32- and 64-bit floating-point numbers.

F (and one bit to normalize it). A representation of this format is shown in

Figure 7.2a. But this formula is implemented according to the following rules in
order to satisfy conditions other than the first one listed below:

1. When 0 < E < 255, then X10 = (−1)s (1 F )2E−127 .

2. When E = 0 and M = 0, then X10 = (−1)s (0 F )(2−126 ).
3. When E = 255 and M = 0, then X10 is not a number and is denoted as
N aN .
4. When E = 255 and M = 0, then X10 = (−1)s ∞.
5. When E = 0 and M = 0, then X10 = (−1)s (0).

Here, (1 F ) is the normalized mantissa with one integer bit and 23 fractional bits,
whereas (0 F ) is only the fractional part with 23 bits. Most of the commercial
DSP chips use this 32-bit, single-precision, floating-point binary representation,
although 64-bit processors are becoming available. Note that there is no provision
for storing the binary point ( ) in these chips; their registers simply store the bits
and implement the rules listed above. The binary point is used only as a notation
for our discussion of the binary number representation and is not counted in the
total number of bits.
The IEEE 754-1985 standard for the (64-bit), double-precision, floating-point
number is expressed by

X10 = (−1)s (1.F )2E−1023 (7.8)

It uses one sign bit, 11 bits for the exponent E, and 52 bits for F (one bit is
added to normalize it but is not counted). The representation for this format is
shown in Figure 7.2b.

Example 7.2

Consider the 16-bit floating-point number with 8 bits for the unbiased exponent
and 4 bits for the denormalized fractional part, namely, E = 8 and F = 4. The
364 QUANTIZED FILTER ANALYSIS

binary number is represented as

X2 = 0100000010 0110

Then the exponent E2 = 100000010; therefore E10 = 130, the denormalized

mantissa F2 = 0110, which gives F10 = 0.375. Therefore the normalized man-
tissa M = 1.375. Finally X10 = −(1.375)2130−127 = +(1.375)23 = 11.
Consider another example:

Y2 = 100000111 0110

Then E2 = 00000111, E10 = 7, F2 = 0110, F10 = 0.375, and finally Y10 =

−(1.375)27−127 = (−1.375)2−120 .
The dynamic range of floating-point numbers that are supported by the FDA
Toolbox as well as the SP Toolbox are listed in Table 7.1.
When the reference filter is quantized, we notice that the default value for the
quantized data as shown in Figure 7.1b is the 16-bit signed magnitude, fixed-
point binary number, and this is shown as format [16 15], which means that 15
bits are used for the fractional part and one bit is used as the sign bit. But note
that the quantizer for the product and sum has the default format [32 30].
In the same panel showing the quantization of the different data, there are two
other columns listed as Round Mode and Overflow Mode. When we click the
button for the Round Mode, we get the following options in the dropdown list:

1. Round: round
2. Floor: floor
3. Ceiling: ceil
4. Fix: fix
5. Convergent: convergent

“Rounding” is the operation of choosing the value to the nearest quantized

number. Negative values that lie halfway between two quantization levels are

TABLE 7.1 Dynamic Range of Floating Point Numbers Found in FDA Tool

Type of
Floating-Point Normalized Normalized Exponent
Data Minimum Value Maximum Value Bias Precision

Single precision 2−126 (2 − 2−23 )2127 127 2−23

≈1.18(10−38 ) ≈(3.4)1038 ≈10−7
Double precision 2−1022 (2 − 2−52 )21023 1023 2−52
≈2(10−308 ) ≈(1.7)10308 ≈10−16
Custom precision 21−bias (2 − 2−F )2bias 2E−1 − 1 2−F
 BINARY NUMBERS AND ARITHMETIC 365

rounded toward negative infinity, and positive numbers that lie halfway between
two quantization levels are rounded toward positive infinity. If the number lies
exactly halfway between two levels, it is rounded toward positive infinity. The
operation called ’floor’ is commonly known as truncation since it discards all
the bits beyond the b bits, and this results in a number that is nearest to negative
infinity. These two are the most commonly used operations in binary arithmetic.
They are illustrated in Figure 7.3, where the dotted line indicates the actual value
of x and the solid line shows the quantized value xQ with b bits.
The ceiling operation rounds the value to the nearest quantization level
toward positive infinity, and the fix operation rounds to the nearest level toward
zero. The convergent operation is the same as rounding except that in the case
when the number is exactly halfway, it is rounded down if the penultimate bit is
zero and rounded up if it is one.
Suppose that two positive numbers or two negative numbers in the fixed-point
format with b bits are added together. It is possible that the result could exceed

X X
XQ XQ

Z−b Z−b

Z−b Z−b
2

(a) ROUNDING (b) TRUNCATION (Two's complement)

Z−b< (XQ − X) ≤ Z−b −Z−b< (XQ − X) ≤ O
−
2 2

X
XQ

Z−b
Z−b

(c) TRUNCATION (one's complement and

signed magnitude)
−Z−b < (XQ − X) ≤ 0, X > 0
O ≤ (XQ − X) < Z−b, X < 0

Figure 7.3 Process of rounding and truncation of binary numbers.

366 QUANTIZED FILTER ANALYSIS

the lower or upper limits of the range within which numbers with b bits lie. For
a signed magnitude, fixed-point number with wordlength w and fraction length
f , the numbers range from −2w−f −1 to 2w−f −1 − 2−f , whereas the range for
floating-point numbers is as given in Table 7.1. When the sum or difference of
two fixed-point numbers or the product of two floating-point numbers exceeds
its normal range of values, there is an overflow or underflow of numbers. The
overflow mode in the FDA panel for the quantized filter gives two choices: to
use saturate or to wrap. Choosing the saturate mode sets values that fall
outside the normal range to a value within the maximum or minimum value in
the range; that is, values greater than the maximum value are set to the maximum
value, and values less than the minimum value are set to the minimum value in
the range. This is the default choice for the overflow mode.
There is a third choice: to scale all the data. This choice is made by clicking
the Optimization button. Then from the dialog box that is displayed, we can
use additional steps to adjust the quantization parameters, scale the coefficients
without changing the overall gain of the filter response, and so on. The coefficients
are scaled appropriately such that there is no overflow or underflow of the data
at the output of every section in the realization.
Before we investigate the effects of finite wordlength and the many realization
structures, by using all the options in the dialog box in the FDA Tool, it is useful
to know some of the insight gleaned from the vast amount of research on this
complex subject. It has been found that in general, the IIR filters in the cascade
connection of second-order sections, each of them realized in direct form II, are
less sensitive to quantization than are those realized in the single section of direct
form I and direct form II. The lattice ARMA structure and the special case of
the AR structure are less sensitive to quantization than is the default structure
described above. The lattice-coupled allpass structure, also known as “two allpass
structures in parallel,” is less sensitive than the lattice ARMA structure. We will
determine whether realizing the two allpass filters A1 (z) and A2 (z) by lattice
allpass structures has any advantages of further reduction in the quantization
effects. If the specified frequency response can be realized by an FIR filter,
then the direct-form or the lattice MA structure realizing it may be preferable to
the structures described above, because the software development and hardware
design of the FIR filter is simpler, is always stable, has linear phase, and is free
from limit cycles.
We first design the reference filter that meets the desired specifications; then
we try different structures for the quantized filter with different levels and types of
quantization. Comparing the frequency response, phase response, and group delay
response of the reference filter with those of the quantized filter, we find out which
structure has the lowest deviation from the frequency response, phase response,
and so on of the reference filter, with the lowest finite wordlength. The FDA Tool
offers us powerful assistance in trying a large number of options available for the
type of filter, design method, frequency specification, quantization of the several
coefficients, and other variables, and comparing the results for the reference filter
 QUANTIZATION ANALYSIS OF IIR FILTERS 367

and the quantized filter, it allows us to make a suboptimal choice of the filter.
This is illustrated by the following example.

7.5 QUANTIZATION ANALYSIS OF IIR FILTERS

Let us select the same fifth-order IIR lowpass elliptic filter that was considered
in Example 6.17. Its transfer function G(z) is given by
 
0.1397
G(z) =
1.965
(1 + 1.337z−1 + 2.251z−2 + 2.251z−3 + 1.337z−4 + z−5 )
×
(1 − 1.629z−1 + 2.256z−2 − 1.597z−3 + 0.8096z−4 − 0.1866z−5 )
(7.9)

The frequency specifications for the filter are given as ωp = 0.4, ωs = 0.6,
Ap = 0.3 dB, and As = 35 dB. The transfer function G(z) was decomposed as
the sum of two allpass filters A1 (z) and A2 (z) such that G(z) = 12 [A1 (z) + A2 (z)],
where
0.5089 − 0.6763z−1 + z−2
A1 (z) = (7.10)
1 − 0.6763z−1 + 0.5089z−2
and
  
0.8805 − 0.5368z−1 + z−2 −0.4165 + z−1
A2 (z) = (7.11)
1 − 0.5367z−1 + 0.8805z−2 1 − 0.4165z−1
Recollect the following lattice coefficients used to realize the lattice structures
for A1 (z) and A2 (z) computed in Chapter 6:
For A1 (z):
 
−0.4482
K1 =
0.5089
⎡ ⎤
0
V1 = ⎣ 0 ⎦
1

For A2 (z):
 
−0.2855
K2 =
0.8805
⎡ ⎤
0
V2 = ⎣ 0 ⎦
1
368 QUANTIZED FILTER ANALYSIS

k3 = [−0.4165]
 
0
v3 =
1

Now we log on to the FDA Toolbox by typing fdatool in the MATLAB com-
mand window and enter the following specifications to design the reference filter
under infinite precision. This is a lowpass, IIR, elliptic filter with sampling
frequency = 48,000 Hz, Fpass = 9600 Hz, Fstop = 14,400 Hz, which correspond
to the normalized sampling frequency = 2, Fpass = 0.4, Fstop = 0.6, respectively.
The maximum passband attenuation is set as Ap = 0.3 dB and the minimum stop-
band attenuation, as As = 35 dB. When we design this filter, we find that the
minimum order of the filter is given as 4, and therefore we increase it to 5 as
the order of the filter so that we can realize the allpass networks in parallel and
compare it with the frequency response of other types of filters. With this selec-
tion, the frequency response and phase response displayed in the Analysis area
are as shown in Figure 7.4. The coefficients of the numerator and denominator
of the IIR reference filter are given below.
Numerator coefficients (normalized to render the constant coefficient of the
numerator as one) of this reference filter are

1.000000000000
1.337660698390
2.251235030190
2.251235030190

Magnitude and phase responses

1 0
Filter #1: Magnitude
Filter #1: Phase

0.8 −80
Phase (degrees)

0.6 −160
Magnitude

0.4 −240

Magnitude and Phase Response of Lowpass, Elliptic, IIR

0.2 Reference Filter of Order 5 −320

0 −400
0 5 10 15 20
Frequency (kHz)

Figure 7.4 Magnitude response of an IIR elliptic lowpass (reference filter) filter.
 QUANTIZATION ANALYSIS OF IIR FILTERS 369

1.337660698390
1.000000000000

The denominator coefficients are

1.000000000000000
−1.629530257267632
2.257141351394922
−1.598167067780082
0.809623494277134
−0.186626971448986

As expected these results match the coefficients in Equation (7.9) within an accu-
racy of four digits, because both of the filters were designed by the same Signal
Processing Toolbox function ellip.
Next, we turn on the quantization, and click the Set Quantization Param-
eters button. The quantization parameters are all set to the default values similar
to those shown in Figure 7.1. We change the format for the fixed-point coeffi-
cients of the filter from [16 15] to [9 8] without changing the format for any
of the other data—although most of the DSP chips currently available use 16 or
32 bits. The magnitude response of the cascade connection of two second-order
sections and one first-order section in direct form II when we quantize the filter
coefficients to 9 bit wordlength is shown in Figure 7.5, along with the magnitude

Magnitude and phase responses

1.4 0

1.12 −80
Phase (degrees)

−160
Magnitude

0.84

0.56 −240
Direct From II, Second Order Sections in cascade
Format [9 8] for Filter Coefficients

0.28
Filter #1: Reference magnitude −320
Filter #1: Quantized magnitude
Filter #1: Reference phase
Filter #1: Quantized phase

0 −400
0 5 10 15 20
Frequency (kHz)

Figure 7.5 Magnitude response of reference filter and quantized filter with format
[9 8] in cascade connection of second-order sections.
370 QUANTIZED FILTER ANALYSIS

Magnitude response in dB
2.5
Filter #1: Reference
Filter #1: Quantized
2

1.5

1
Magnitude (dB)

0.5

−0.5

−1 Magnified plot of Magnitude Response in dB in the passband of

Direct From II, Second Order sections in cascade
Format [9 8] for Filter Coefficients
−1.5

−2

1 2 3 4 5 6 7 8 9
Frequency (kHz)

Figure 7.6 Magnified plot of the magnitudes (in decibels) of the two filters in Figure 7.5.

response of the reference filter. The Figure 7.6 shows a magnified plot of the
magnitude in decibels in the passband, which gives the response of the quantized
filter, which is very close to the response of the reference filter. But most of
the DSP chips available on the market have a wordlength that is a power of 2
(wordlengths of 8, 16, 32, etc.). So we try a quantization of 8 bits, and the mag-
nitude response of this filter is shown in Figure 7.7. But we see that the deviation
of the magnitude from that of the reference filter is pronounced near the edge
of the passband. Although we prefer to choose a wordlength of 8 rather than 9,
this deviation is considered excessive, so we must choose other structures. As an
alternative structure, we convert the direct form II structure to the ARMA struc-
ture with the same wordlength of 8 bits, and the resulting magnitude response is
shown in Figure 7.8. A magnified plot of this response in its passband is shown
in Figure 7.9. It does not produce a significant improvement over the response
shown in Figure 7.7 for the quantized filter in the default structure of direct form
II, with the same wordlength of 8 bits.
So we decide to convert the lattice ARMA structure to the lattice-coupled
allpass structure; each allpass structure is realized by lattice allpass structures and
starts with a 9-bit fixed-point quantization for the filter coefficients, and we get the
result shown in Figure 7.10. Hardly any difference is seen between the reference
filter and the quantized filter with the format [9 8], the same as the direct form
 QUANTIZATION ANALYSIS OF IIR FILTERS 371

Magnitude response
1.4
Filter #1: Reference
Filter #1: Quantized
1.2

1
Magnitude

0.8

0.6

0.4
Magnified Response of Direct Form II, Second Order Sections
in Cascade
0.2 Format [8 7] for Filter Coefficients

0
0 5 10 15 20
Frequency (kHz)

Figure 7.7 Magnitude responses of reference filter and quantized filter with format
[8 7] in cascade connection of second-order sections.

Magnitude response
1.4
Filter #1: Reference
Filter #1: Quantized
1.2

1
Magnitude

0.8

0.6
Lattice ARMA Lowpass: Elliptic, IIR Filter
0.4 Format [8 7] for Filter Coefficients

0.2

0
0 5 10 15 20
Frequency (kHz)

Figure 7.8 Magnitude responses of reference filter and quantized filter with format
[8 7] and lattice ARMA structure.
372 QUANTIZED FILTER ANALYSIS

Magnitude response in dB

Filter #1: Reference

1 Filter #1: Quantized

0.5
Magnitude (dB)

−0.5

−1
Magnified Magnitude Response in dB in the passband of Lattice
ARMA Lowpass, IIR: Elliptic Filter
Format [8 7] for Filter Coefficients
−1.5
1 2 3 4 5 6 7 8 9
Frequency (kHz)

Figure 7.9 Magnified plot of the magnitude responses (in decibels) of reference filter
and quantized filter with format [8 7] in lattice ARMA structure.

Magnitude response
1.4
Filter #1: Reference
1.2 Filter #1: Quantized

1
Magnitude

0.8

0.6

0.4

0.2
Lattice Coupled Allpass Structure
Format [9 8] for filter Coefficients
0
0 5 10 15 20
Frequency (kHz)

Figure 7.10 Magnitude response of a lattice-coupled allpass structure, quantized with a

format [9 8].
 QUANTIZATION ANALYSIS OF IIR FILTERS 373

Magnitude response
1.2

Filter #1: Reference

1 Filter #1: Quantized

0.8
Magnitude

0.6

0.4

Lattice Coupled Allpass Structure

0.2 Format [7 6] for Filter Coefficients

0
0 5 10 15 20
Frequency (kHz)

Figure 7.11 Magnitude responses of reference filter and quantized filter with format
[7 6], in lattice-coupled allpass structure.

II structure with 9 bits. Next we try the 7 bit wordlength for this structure and
the magnitude response shown in Figure 7.11. Again, we prefer to choose an
8 bit wordlength for this structure. The magnitude and phase responses of the
filter with 8 bits are shown in Figure 7.12. A magnified plot of the magnitude
in decibels in the passband of this 8-bit filter is shown in Figure 7.13. It shows
that the maximum attenuation for the reference filter is 0.3 dB as specified, and
the deviation from the specified passband magnitude, for the quantized filter is
about 0.1 dB. This amount of deviation is less than that exhibited by the lattice
ARMA filter in Figure 7.9 Therefore this lattice-coupled allpass structure for the
IIR filter is chosen as a compromise.
The lattice coefficients of the second-order allpass filter A1 (z) and those for the
third-order allpass filter A2 (z) realizing the reference filter are printed out and
shown in the right column of Figure 7.14. The lattice coefficients for the two
allpass filters displayed in Figure 7.14 match those given in the vectors K1, V1,
K2, V2, k3 and v3 given at the beginning of this section, within an accuracy
of four digits. In the left column are shown the corresponding coefficients of
the quantized filter with a 8 bit wordlength in the fixed-point, signed magnitude
format [8 7].
374 QUANTIZED FILTER ANALYSIS

Magnitude and phase responses

1.2 0

0.96 −80

Phase (degrees)
0.72 −160
Magnitude

0.48 Filter #1: Reference magnitude −240

Filter #1: Quantized magnitude
Filter #1: Reference phase
Filter #1: Quantized phase
0.24 Magnitude and Phase Response of Lattice Coupled Allpass Structure −320
of IIR Lowpass Elliptic Filter
Format [8 7] for Filter Coefficients

0 −400
0 5 10 15 20
Frequency (kHz)

Figure 7.12 Magnitude and phase responses of reference filter and the quantized filter
with format [8 7] in lattice-coupled allpass structure.

Magnitude response in dB

0.5 Filter #1: Reference

Filter #1: Quantized

0
Magnitude (dB)

−0.5

−1
Magnified Magnitude Response in dB in the passband
of the Lattice Coupled Allpass, IIR, Lowpass, Elliptic Filter
Format [8 7] for Filter Coefficients
−1.5

1 2 3 4 5 6 7 8 9
Frequency (kHz)

Figure 7.13 Magnified plot of the magnitude responses (in decibels) of reference filter
and quantized filter with format [8 7] in lattice-coupled allpass structure.
 QUANTIZATION ANALYSIS OF FIR FILTERS 375

Figure 7.14 Coefficients of reference filter and quantized filter with format [8 7] in a
lattice-coupled allpass structure.

7.6 QUANTIZATION ANALYSIS OF FIR FILTERS

Next we decide to investigate whether the alternative of designing a quantized

FIR filter would give us a better result. Choosing the same frequency-domain
specifications as for the IIR lowpass elliptic filter, we design an FIR lowpass filter
with an equiripple passband and stopband. The reference filter with infinite
precision uses the remez algorithm and yields a linear phase FIR (type I)
filter of order 16. The magnitude response of this filter is shown in Figure 7.15.
When we select the sign magnitude, fixed-point 7 bit wordlength and the 8 bit
wordlength, the results are as shown in Figures 7.16 and 7.17, respectively.
It is apparent that there is not a significant difference between the two filters
with 7 and 8 bit wordlength. In Figure 7.18 we plot a magnified magnitude
in decibels in the passband of the FIR filter with 8 bits for the wordlength. The
maximum deviation from the specified passband ripple of 0.3 dB is ∼0.1 dB. The
coefficients of the reference filter and the quantized filter are listed in Figure 7.19.
It is noted that several coefficients of the quantized filter have an underflow as
indicated by the digit 0 in the first column and have been rounded to zero.
Finally we compare the quantization effects on the IIR with the effect on the
FIR filter by comparing the magnitude responses shown in Figures 7.18 and 7.13.
It is easy to notice that the FIR filter has a lower sensitivity to quantization than
does the IIR filter that we chose above. The IIR filter in the form of the lattice-
coupled allpass structure and the FIR filter in the direct form have a wordlength
376 QUANTIZED FILTER ANALYSIS

Magnitude response
1.4
Filter #1
1.2

1
Magnitude

0.8

0.6

0.4
Magnitude Response of Lowpass, Equiripple, Reference Filter
Order 16
0.2

0
0 5 10 15 20
Frequency (kHz)

Figure 7.15 Magnitude response of a lowpass equiripple FIR (reference filter) filter.

Magnitude and phase responses

1.4 0
Magnitude and Phase Response of Lowpass, Equiripple
FIR Filter Format [7 6] for Filter Coefficients

1.12 −180
Phase (degrees)

0.84 −360
Magnitude

0.56 −540

0.28 Filter #1: Reference magnitude −720

Filter #1: Quantized magnitude
Filter #1: Reference phase
Filter #1: Quantized phase
0 −900
0 5 10 15 20
Frequency (kHz)

Figure 7.16 Magnitude and phase responses of reference FIR filter and quantized filter
with format [7 6] for filter coefficients.
 QUANTIZATION ANALYSIS OF FIR FILTERS 377

Magnitude and phase responses

1.4 0
Magnitude and Phase Response of Lowpass, Equiripple
FIR Filter Format [8 7] for Filter Coefficients

1.12 −180

Phase (degrees)
0.84 −360
Magnitude

0.56 −540

0.28 Filter #1: Reference magnitude −720

Filter #1: Quantized magnitude
Filter #1: Reference phase
Filter #1: Quantized phase
0 −900
0 5 10 15 20
Frequency (kHz)

Figure 7.17 Magnitude and phase responses of FIR reference filter and quantized filter
with format [8 7] for filter coefficients.

Magnitude response in dB

Filter #1: Reference

1.5 Filter #1: Quantized

0.5
Magnitude (dB)

−0.5
Magnified Magnitude in dB in the passband of the Lowpass, Equiripple
FIR Filter
−1
Format [8 7] for Filter Coefficients

−1.5

1 2 3 4 5 6 7 8 9
Frequency (kHz)

Figure 7.18 Magnified magnitude responses (in decibels) of reference FIR filter and
quantized filter with format [8 7] for the filter coefficients.
378 QUANTIZED FILTER ANALYSIS

Figure 7.19 Data for reference FIR filter and quantized FIR filter with 8 bits of word-
length.

of 8 bits. The number of multipliers required in FIR direct form is only 9 because
of the symmetry in its coefficients, whereas the lattice-coupled allpass network
requires 10 multipliers, which is not a significant difference. However, we know
that the phase response of the FIR filter is linear, which is a great advantage over
the IIR filter. Hardware implementation of the FIR filter is simpler than that of
the IIR filter. Unlike the IIR filter, the FIR filter does not exhibit limit cycles
and is always stable. This leads to investigate the 8-bit FIR filter further as a
candidate for generating the code to program a DSP chip of our choice.
It must be pointed out that the specifications we selected for the digital filter
may or may not meet typical application requirements. Also, we would like
to point out that while we argued that a 8 bit wordlength may be preferable
over a 9 bit wordlength, currently most of the digital signal processors (DSPs)
are 16-bit or 32-bit devices. The design process using the fdatool is meant
to illustrate only the different choices and decisions that an engineer may face
before arriving at a particular digital filter that will be considered for further
investigation as described below.
Now we assume that we have designed the digital filter and we have tested
its performance using the fdatool, when the coefficients of the filter and the
 REFERENCES 379

input samples are represented by a finite number of bits. We have also considered
the effect of rounding or truncating the results of adding signals or multiplying
the signal value and the coefficient of the filter and ascertained that there is no
possibility of limit cycles or unstable operation in the filter. Very often a digital
filter is used as a prominent part of a digital system like a cell phone, which has
other components besides power supply, keyboard, or other I/O interfaces. So
we have to simulate the performance of the whole system with all components
connected together in the form of a block diagram.

7.7 SUMMARY

In this chapter we described the use of the MATLAB tool, called the fdatool,
to design digital filters with finite wordlength for the coefficients in fixed-point
and floating-point representations, and investigated several different types of fil-
ter structures and different types of magnitude response specifications. Once we
narrowed down the choice of the filter that meets the frequency response spec-
ifications, we have to simulate the performance of the filter using Simulink, to
check that the filter works satisfactorily under different types of input signals
that will be applied in practice. In Chapter 8 we discuss this and other practical
considerations that are necessary for hardware design of the filter or the whole
digital system in which the filter is embedded.

PROBLEMS

The problems given in Chapter 4 can be assigned either as homework or as

a term project only if the FDA Toolbox is available with MATLAB and Signal
Processing Toolbox in the computer facilities of the school, college, or university.
The students may be asked to investigate the quantization effect of the filters
specified in Chapter 4 and arrive at the suboptimal choice, using these tools.

REFERENCES

1. S. K. Mitra, Digital Signal Processing—A Computer Aided Approach, McGraw-Hill,

1998.
2. E. C. Ifeachor and B. W. Jervis, Digital Signal Processing, A Practical Approach,
Prentice-Hall, 2002.
3. A. Bateman and I. Paterson-Stephens, The DSP Handbook, Algorithms, Applications
and Design Techniques, Prentice-Hall, 2002.
4. S. K. Mitra and J. F. Kaiser, eds., Handbook of Digital Signal Processing, Wiley-
Interscience, 1993.
5. S. M. Kuo and B. H. Lee, Real-Time Digital Signal Processing: Implementations,
Applications and Experiments with TMS320C55X, Wiley, 2001.
380 QUANTIZED FILTER ANALYSIS

6. A. Singh and S. Srinivasan, Digital Signal Processing, Thompson Books/Cole Pub-

lishing, 2004.
7. N. Kehtarnavaz and B. Simsek, C6x-Based Digital Signal Processing, Prentice-Hall,
2000.
8. P. Lapsley, J. Bier, A. Shoham, and E. Lee, DSP Processor Fundamentals, IEEE
Press, 1994.
9. S. M. Kuo and W.-S. Gan, Digital Signal Processors: Architectures, Implementations
and Applications, Pearson Prentice-Hall, 2004.
10. R. Chassaing, Applications Using C and the TMS320C6XDSK, Wiley, 2002.
11. R. Chassaing and D. W. Horning, Digital Signal Processing with the TMS320C25,
John Wiley, 1990.
12. The Mathworks, Inc., Learning Simulink, User’s Guide, 1994.
 CHAPTER 8

Hardware Design Using DSP Chips

8.1 INTRODUCTION

In Chapter 7, we used the fdatool to illustrate the analysis and design of a

digital filter in which the coefficients of the filter and the input samples are
represented by a finite number of bits. We also found the effect of rounding or
truncating the results of adding signals or multiplying the signal value and the
coefficient of the filter and ascertained that there is no possibility of limit cycles
or unstable operation in the filter. In the example chosen we decided that an
FIR filter would meet the frequency response specifications of a lowpass elliptic
filter, with a wordlength of 8 bits. Very often, however, a digital filter is used
as a prominent part of a digital system such as a cell phone, which has other
components such as power supply, keyboard, or other I/O interfaces. So we have
to simulate the performance of the whole system with all components connected
together in the form of a block diagram.

8.2 SIMULINK AND REAL-TIME WORKSHOP

Simulink is the software that is available as a companion toolbox to MATLAB

and is used to model and simulate the performance of dynamic systems, under
varying conditions. Just as MATLAB works with a number of toolboxes, Simulink
has access to a library of many additional tools called blocksets, such as the
DSP blockset, fixed-point blockset, communications blockset, and control system
blockset, as shown on the left side of Figure 8.1.
The Simulink browser library includes blocksets for simulation of aeronautical
and mechanical systems, too, namely, are aerospace blockset and simMe-
chanics.1 Each of these blocksets contains a large number of blocks that are used
to define specific transfer functions or algorithms and a variety of input signals.
1
Depending on the version of Simulink, this may or may not contain some of the blocksets mentioned
in this chapter.

Introduction to Digital Signal Processing and Filter Design, by B. A. Shenoi

Copyright © 2006 John Wiley & Sons, Inc.

381
382 HARDWARE DESIGN USING DSP CHIPS

Figure 8.1 Screen capture of the Simulink browser and block diagram of a model.

The GUI interface is used to drag and drop these blocks from the blockset and
connect them to describe a block diagram representation of the dynamic system,
which may be a continuous-time system or a discrete-time system. A mechanical
system model [6] is shown in Figure 8.1. Simulink is based on object-oriented
programming, and the blocks are represented as objects with appropriate prop-
erties, usually specified in a dialog box. Indeed, the fdatool that we used in
Chapter 7 can be launched from SIMULINK as an object or from the MATLAB
command window because both of them are integrated together to operate in a
seamless fashion. Simulink itself can be launched either by typing simulink in
the MATLAB command window or by clicking the Simulink icon in its toolbar.
For the simulation of a digital filter, we choose the DSP blockset, which
contains the following blocks in a tree structure:
DSP Blockset

→DSP Sinks
→DSP Sources
→Estimation
→Filtering →Adaptive Filters
→Math Functions →Filter Design →Analog Filter Design
→Platform Specific I/O →Multirate filters →Digital Filter Design
→Quantizers →Digital Filter
→Signal Management →Filter Realization
Wizard
→Signal Operations →Overlap-Add FFT filter
→Statistics →Overlap-Save FFT filter
→Transforms
 DESIGN PRELIMINARIES 383

When we open Simulink window, and click File→New→Model in sequence,

we get a window for the new model. Then we drag the block shown above
as Digital Filter Design and drop it in the window for the new model.
When we click on this object in the new window, it opens the same window
as the one for the fdatool shown in Figure 7.1. After we have imported the
parameters of the digital filter that we designed in an earlier session, or after we
have completed the design of the quantized filter as explained in Chapter 7, we
use the Filter Realization Wizard shown above under the DSP Blockset
and get the realization structure for the filter. This serves as the model for the
filter to which we can now connect different types of sources and observe the
output on the scope connected to the filter, as the sink. Very often, we are
required to design a whole system, in which case a digital filter is the only
major block in the system, but there are other subsystems integrated with it.
So it may be necessary to use the blocks for the adaptive filters or multirate
filters or the blocks from the Communication blockset and Controls blockset,
besides the DSP Blockset, and so on. After building the block diagram model
for the total digital signal processing system, and using Simulink to carry out
extensive simulation of the model under varying conditions, we check to ensure
that it meets the specifications satisfactorily; if not, we may have to modify the
design of the filter or tune the parameters. For example, we may simulate the
total system with a finite number of bits in floating-point or fixed-point format,
using the Fixed Point blockset to represent all data. We may have to change
the design completely and simulate the new system.

8.3 DESIGN PRELIMINARIES

All the design and simulation of digital filters and digital systems done by MAT-
LAB and Simulink is based on numerical computation of scientific theory. When
this work is completed, we have to decide on one of the following choices:

1. Design a VLSI chip, using software such as VHDL, to meet our particular
design specifications
2. Select a DSP chip from manufacturers such as Texas Instruments, Analog
Devices, Lucent, or Motorola and program it to work as a digital system
3. Choose a general-purpose microprocessor and program it to work as a
digital signal processor system.
4. Design the system using the field-programmable gate arrays (FPGAs).

In all cases, several design considerations have to be explored as thoroughly as

possible before we embark on the next step in hardware design.
If we decide to select a DSP chip from one of the abovementioned manufac-
tures, we have to consider the bandwidth of the signal(s) that the digital filter or
the the digital system will be processing, based on which sampling frequency of
the ADC is selected. However, the sampling frequency of the ADC may not be
384 HARDWARE DESIGN USING DSP CHIPS

the same as the clock frequency of the CPU in the chip or the rate at which data
will be transferred from and to the memory by the CPU (central processing unit).
This in turn determines the rating in mips (millions of instructions per second).
Depending on the amount of data or memory space required by the processor, the
amount of power is determined. Other considerations are the I/O (input/output)
interfaces, additional devices such as the power supply circuit, and the micro-
controller, add-on memory, and peripheral devices. Finally the most important is
the the cost per chip. We also need to consider the reliability of the software and
technical support provided by the manufacturer; credibility and sustainability of
the manufacturer also become important if the market for the digital filter or the
system is expected to last for many years.
The selection of the DSP chip is facilitated by an evaluation of the chips avail-
able from the major manufacturers listed above and their detailed specifications.
For example, the DSP Selection Guide, which can be downloaded from the TI
(Texas Instruments) Website www.dspvillage.ti.com, is an immense source
of information on all the chips available from them.
The DSP chips provided by TI are divided into three categories. The fam-
ily of TMS3206000 DSP platform are designed for systems with very high
performance, ranging within 1200–5760 mips for fixed-point operation and
600–1350 mflops (million floating-point operations per second) for floating-
point operation. The fixed-point DSPs are designated by TMS320C62x and
TMS320C64x, and the floating-point DSPs belong to the TMS320C67x family.
The fixed-point TMS32062x DSPs are optimized for multichannel, multifunc-
tion applications such as wireless base stations, remote-access servers, digital
subscriber loop (DSL) systems, central office switches, call processing, speech
recognition, image processing, biometric equipment, industrial scanners, pre-
cision instruments, and multichannel telephone systems. They use 16 bits for
multiplication and 32 bits for instructions in single-precision format as well as
double-precision format. The fixed-point TMS320C64x DSPs offer the high-
est level of performance at clock rates of up to 720 MHz and 5760 mips, and
they are best suited for applications in digital communications and video and
image processing, wireless LAN (local area networking), network cameras, base
station transceivers, DSL, and pooled modems, and so on. The floating-point
TMS320C67x DSPs operate at 225 MHz and are used in similar applications.
The TMS320C5000 DSP family is used in consumer digital equipments,
namely, products used in the Internet and in consumer electronics. Therefore
these chips are optimized for power consumption as low as 0.05 mW/mips and
speeds of ≤300 MHz and 600 mips; the TMS320C54x DSPs are well known as
the industry leader in portable devices such as cell phones(2G, 2.5G, and 3G), dig-
ital audio (MP3) players, digital cameras, personal digital assistants (PDAs), GPS
receivers, and electronic books. The TMS320C55x DSPs also deliver the highest
power efficiency and are software-compatible with the TMS320C54x DSPs.
The TMS320C2000 DSPs are designed for applications in digital con-
trol industry, including industrial drives, servocontrol, factory automation,
office equipment, controllers for pumps, fans, HVAC (heating–ventilation–air
 CODE GENERATION 385

conditioning), and other home appliances. The TMS320C28x DSPs offer 32-bit,
fixed-point processing and 150 mips operation, whereas the TMS320C24x DSPs
offer a maximum of 40 mips operation.
More detailed information and specifications for the DSPs and other devices
such as ADCs, and codecs (coders/decoders) supplied by TI can be found in the
DSP Selection Guide. The amount of information on the software and hardware
development tools, application notes, and other resource material that is freely
available in this Website is enormous and indispensable. We must remember that
DSP chips produced by other manufacturers such as Analog Devices may be
better suited for specific applications, and they, too, provide a lot of information
about their chips and the applications.

8.4 CODE GENERATION

The next task is to generate a code in machine language that the DSP we have
selected understands and that implements the algorithm for the digital system
we have designed. First we have to convert the algorithm for the system under
development to a code in C/C++ language. This can be done manually by one
who is experienced in C language programming. Or we simulate the performance
of the whole system modeled in Simulink, and use a blockset available in it,
known as the Real-Time Workshop [7] to generate the ANSI Standard C code
for the model.2 The C code can be run on PCs, DSPs, and microcontrollers in real
time and non–real time in a variety of target environments. We connect a rapid
prototyping target, for example, the xPC Target, to the physical system but use
the Simulink model as the interface to the physical target. With this setup, we test
and evaluate the performance of the physical target. When the simulation is found
to work satisfactorily, the Real-Time Workshop is used to create and download
an executable code to the target system. Now we can monitor the performance of
the target system and tune its parameters, if necessary. The Real-Time Workshop
is useful for validating the basic concept and overall performance of the whole
system that responds to a program in C code.
An extension of Real-Time Workshop called the Real-Time Workshop Embed-
ded Coder is used to generate optimized C code for embedded discrete-time
systems.
Note that the C code is portable in the sense that it is independent of any man-
ufacturer’s DSP chip. But the manufacturers may provide their own software to
generate the C code also, optimized for their particular DSP chip. However, pro-
gramming a code in machine language is different for DSP chips from different
manufacturers, and the different manufacturers provide the tools necessary to
obtain the machine code from the C code for their DSP chips.

2
Depending on the version of MATLAB/Simulink package installed on the computer in the college
or university, software such as FDA Tool, Real-Time Workshop and others mentioned in this chapter
may or may not be available.
386 HARDWARE DESIGN USING DSP CHIPS

8.5 CODE COMPOSER STUDIO

Texas Instruments calls its integrated development tool the Code Composer Stu-
dio (IDE). The major steps to be carried out are outlined in Figure 8.2. Basically,
these steps denote the C compiler, assembler, linker, debugger, simulator, and
emulator functions. It must be pointed out that the other manufacturers also
design DSP chips for various applications meeting different specifications; their
own software bundle follows steps similar to those mentioned above for the Code
Composer Studio (CCS) from Texas Instruments (TI).
First the Code Composer Studio compiles the C/C++ code to an assembly
language code in either mnemonic form or algebraic form, for the particular

C/C++ Source
files

C/C++ Compiler
C/C++ and
Assembly files Assembly Macro library
Source Files files

Assembler
COFF Object
files

Object library RTS library

files files
Linker

Executable COFF
Object Module

DSP Debug DSP Emulation

Environment Hardware

Simulator DSP Application

DSP/BIOS Support DSP/BIOS II Kernel

Host Emulation JTAG Link Target Hardware

Support

Figure 8.2 Software development flow for generating the object code from the C code.
 CODE COMPOSER STUDIO 387

DSP platform that we have chosen. If we choose the TMS320C55x DSPs to

illustrate the software development cycle, then the command used to invoke the
C compiler is of the form

c155[-options] [filenames] [-z[link options] [object files]]

The [filenames] list the C program files, and other assembly language files, and
even object files with their default extensions .c, .asm, and .obj, respectively.
The C language is not very efficient in carrying out a few specific operations,
such as fixed-point data processing that are used in DSP applications. For this
reason, assembly language files are added to the C language program files in order
to improve the efficiency of the program in carrying out time-critical sections
of the assembly language code delivered by the assembler. We can choose from
many options in [-options] and in [link options] to control the way that
the compiler shell processes the files listed in [filenames] and the way that
the linker processes the object files. For more details, students should refer to the
TI simulator user’s guide [25].
The next step is translation of the assembly language code by the assembler
to the object code in binary form (or in machine language) specific to the DSP
platform. The CCS command to invoke the assembler is of the form

asm55 [input file [object file] [list file] [-options]]

Since there might be several C program files that implement the original algo-
rithm in small sections, the assembler produces the output file in several sections.
It may also collect assembly source files from an external library, which imple-
ment processes that are used again and again at several stages of the software and
load them into the list of [filenames]. For example, Texas Instruments pro-
vides a large number of highly optimized functions in three libraries, namely, the
DSP library (DSPLib), the image processing library (IMAGELib), and the chip
support library (CSLib). Then there are assembly files that are long programs
and therefore are shortened to a macro so that they can be invoked by a single
or a few lines of instructions. All of these external files are added to the list
of assembly language files and converted to binary form, under a single format
known as the common-object file format (COFF). The object file produces
the object file in COFF format; the list file shows the binary object code as
well as the assembly source code and where the program and the variables are
allocated in the memory space. But they are allocated in temporary locations, not
in absolute locations. Therefore these relocatable object files can be archived into
a library of reusable files that may be used elsewhere. There are many options
in the assembler, and their use is described in Ref. 25.
The linker utility is invoked to combine all the object files generated by the
assembler to one single linked object code, and this is done by assigning absolute
addresses in the physical memory of the target DSP chip as specified by a memory
map. The memory map is created by a linker command xfile, which lists the
various sections of the assembly code and specifies the location of the starting
388 HARDWARE DESIGN USING DSP CHIPS

address and length of memory space in RAM and ROM (random access and
read-only memory), and where the individual sections are to be located in the
RAM and ROM, as well as the various options. Then the linker command is
invoked as follows:

lnk55 command file.cmd

The linker can call additional object files from an external library and also the
runtime support (RTS) library files that are necessary during the debugging proce-
dure. It also has many options that can be used to control the linker output, which
is an executable COFF object module that has .out as its extension. Detailed
information on the linker can be found in Ref. 17. Remember that information
on compiler, assembler, and linker commands may be different for other DSP
platforms, and information on these commands may be found in TI references
appropriate for the DSP platform chosen.

8.6 SIMULATOR AND EMULATOR

After we have created the executable COFF object module, we have to test and
debug it by using software simulation and/or by hardware emulation. For low-cost
simulation, we use the development starter kits, for example, the TMS3205402
DSP starter kit for the TMS320C54x DSP, and for more detailed evaluation and
debugging, we use an evaluation board such as the TMS320C5409. Finally, we
have the emulator boards such as the XDS510 JTAG emulator, which are used
to run the object code under real-time conditions.
The executable object code is downloaded to the DSP on the DSK board.
The simulator program installed on the PC that is connected to the DSK board
accepts the object code as its input and under the user’s control, simulates the
same actions that would be taken by the DSP device as it executes the object
code. The user can execute the object code one line at a time, by inserting
breakpoints at a particular line of the object program, halt the operation of the
program; view the contents of the data memory, program memory, auxiliary
registers, stacks, and so on; display the contents of the registers, for example,
the input and output of a filtering operation; and change the contents of any
register if so desired. One can also observe or monitor the registers controlling
the I/O hardware, serial ports, and other components. If minor changes are made,
the Code Composer Studio reassembles and links the files quickly to accelerate
the debugging process; otherwise the entire program has to be reassembled and
linked before debugging can proceed. When the monitoring and fixing the bug
at all breakpoints is over, execution of the program is resumed manually. By
inserting probe points, Code Composer Studio enables us to read the data from
a file or written to a file on the host PC, halting the execution of the program
momentarily, and then resume it. It should be obvious that simulation on a DSK
is a slow process and does not check the performance of the peripheral devices
that would be connected to the digital system.
 CONCLUSION 389

In order to test the performance of the object code on the DSP in real time,
we connect an emulator board to the PC by a parallel printer cable, and the
XDS 510 Emulator conforms to the JTAG scan-based interface standard. The
peripheral devices are also connected to the emulator board. A DSP/BIOS II
plug-in is included in the Code Composer Studio to run the emulation of the
software. It also contains the RTDX (real-time data exchange) module that allows
transfer of data between the target DSP and the host PC in real time. The Code
Composer Studio enables us to test and debug the performance of the software
under real-time conditions, at full sampling rate. Without disrupting the execution
of the software, the emulator controls its execution of the breakpoints, single-
step execution, and monitoring of the memory and registers, and checks the
performance of the whole system, including the peripheral devices. When the
emulation of the whole system is found to operate correctly, the software is
approved for production and marketing.
This is a very brief outline of the hardware design process, carried out after
the design of the digital system is completed by use of MATLAB and Simulink.
Students are advised to refer to the extensive literature available from TI and
other manufacturers, in order to become proficient in the use of all software
tools available from them. For example, Analog Devices offers a development
software called Visual DSP++, which includes a C++ compiler, assembler, linker,
user interface, and debugging utilities for their ADSP-21xx DSP chips.

8.6.1 Embedded Target with Real-Time Workshop

Simulink has been expanded to generate and simulate bit-true, timing-accurate
code for directly designing DSP and FPGA targets and produce tests at system
level. This software tool considerably reduces the design effort outlined above,
as it facilitates the design of digital filters and systems obtained by the Signal
Processing Toolbox and FDA Toolbox and generates executable machine code
for hardware design.

8.7 CONCLUSION

The material presented above is only a very brief outline of the design proce-
dure that is necessary to generate the assembly language code from the C code,
generate the object code using the assembler, and link the various sections of
the object code to obtain the executable object code in machine language. Then
this code is debugged by using an evaluation board, simulator, and emulator; all
of these steps are carried out by using an integrated, seamless software such as
the Code Composer Studio that was used to illustrate the steps. Like any design
process, this is an iterative procedure that may require that we go back to earlier
steps to improve or optimize the design, until we are completely satisfied with
the performance of the whole system in real-time conditions. Then the software
development is complete and is ready for use in the DSP chips chosen for the
specific application.
390 HARDWARE DESIGN USING DSP CHIPS

REFERENCES

1. S. M. Kuo and B. H. Lee, Real-Time Digital Signal Processing; Implementations,

Applications and Experiments with TMS320C55X, Wiley, 2001.
2. S. M. Kuo and W.-S. Gan, Digital Signal Processors: Architectures, Implementations
and Applications, Pearson Prentice-Hall, 2004.
3. R. Chassaing, Applications Using C and the TMS320C6XDSK, Wiley, 2002.
4. R. Chassaing and D. W. Horning, Digital Signal Processing with the TMS320C25,
Wiley, 1990.
5. R. Chassaing, Digital Signal Processing and Applications with the C6713 and C416
DSK, Wiley, 2004.
6. The MathWorks, Inc., Learning Simulink, User’s Guide, 1994.
7. The MathWorks, Inc., Real-Time Workshop for Use with Simulink, User’s Guide.
8. P. Embree, C Algorithms for Real-Time DSP, Prentice-Hall, 1995.
9. The MathWorks, Inc., Embedded Target for the TI TMS320C6000T M DSP Platform,
for Use with Real-Time Workshop(R) , User’s Guide.
10. Texas Instruments, TMS320C6411 Fixed-Point Digital Signal Processor (SPRS196).
11. Texas Instruments, TMS320C6000 CPU and Instruction Set Reference Guide
(SPRU189).
12. Texas Instruments, Manual Update Sheet for TMS 320C6000 CPU and Instruction
Set Reference Guide (SPRZ168).
13. Texas Instruments, TMS320C64x Technical Overview (SPRU395).
14. Texas Instruments, Code Composer Studio Tutorial (SPRU301).
15. Texas Instruments, Code Composer Studio User’s Guide (SPRU328).
16. Texas Instruments, TMS320C6000 Programmer’s Guide (SPRU198).
17. Texas Instruments, TMS320C6000 Assembly Language Tools User’s Guide
(SPRU186).
18. Texas Instruments, TMS320C6000 Optimizing C Compiler User’s Guide (SPRU187).
19. Texas Instruments, TMS320C6000 C Source Debugger User’s Guide (SPRU188).
20. Texas Instruments, TMS320C6000 DSP/BIOS User’s Guide (SPRU303).
21. Texas Instruments, TMS320C6000 DSP/BIOS Application Programming Interface
(API) Reference Guide (SPRU403).
22. Texas Instruments, TMS320 DSP Algorithm Standard Rules and Guidelines
(SPRU352).
23. Texas Instruments, TMS320 DSP Algorithm Standard Developer’s Guide (SPRU424).
24. Texas Instruments, TMS320C6000 Simulator User’s Guide (SPRU546).
25. Texas Instruments, TMS320C55x Optimizing C Compiler User’s Guide (SPRU281).
26. Texas Instruments, TMS320C55x Assembly Language Tools User’s Guide (SPRU380).
27. Analog Devices, ADSP-21xxx Family: Assembler Tools and Simulator Model, 1995.
28. Analog Devices, ADSP-2106x SHARC User’s Manual, 1997.
29. Motorola Inc., Motorola DSP Assembler Reference Manual, 1994.
30. Motorola Inc., DSP56xxx Digital Signal Processor: User’s Manual, 1993.
 CHAPTER 9

MATLAB Primer

9.1 INTRODUCTION

MATLAB is a very powerful and well-known software package1 that is used

in science and engineering disciplines, for numerical computation, data analysis,
and graphical visualization. It is available in almost all platforms such as personal
computers, and workstations running under several operating systems. When you
begin a session by typing the command matlab, the first window displayed on
the monitor is the command window with the prompt >>, which is waiting for
your command.2 Use the command exit to end the session.
MATLAB contains a large collection of built-in functions and commands that
are used in an interactive mode, when you are in the command window. As soon
as the name of a function or a command is typed at the prompt in the command
window, with the proper syntax, the answer is displayed immediately. But there
are two other windows, the edit window and graphics window, which will be dis-
cussed later. The software package is designed to use additional sets of functions
that are more applicable in particular disciplines such as control systems, digital
signal processing, communications engineering, and image processing. There are
more than 20 sets known as “toolboxes” (e.g., control toolbox, digital signal pro-
cessing toolbox, communication toolbox, image processing toolbox). All of them
run under MATLAB and implement the functions on the basis of matrix manipu-
lation of numerical data, and that is why the software is called MATLAB (matrix
laboratory). Simulink is another toolbox that is used to simulate the performance
of the systems, when the systems are built by connecting individual blocks rep-
resenting different subsystems and the output is obtained when the systems are

1
The software is available from The MathWorks, Inc., 3 Apple Hill Drive, Natick, MA
01760-2098, phone 508-647-7000, fax 508-647-7001, email [email protected], Website
http://www.mathworks.com.
2
If you are logging on a workstation connected to a computer network, you may have to set the
proper environment by typing setenv DISPLAY network number: or some other com-
mand before launching MATLAB.

Introduction to Digital Signal Processing and Filter Design, by B. A. Shenoi

Copyright © 2006 John Wiley & Sons, Inc.

391
392 MATLAB PRIMER

subjected to different kinds of input signals. MATLAB allows us to construct

new functions using the enormous number of built-in functions, commands, and
operations in MATLAB and in the many toolboxes, without having to know how
to compile, link, load, and create executable code, because MATLAB uses its own
language to carry out all these steps, which are invisible to the user. It carries out
the steps and gives the answer very fast! In most versions of MATLAB, there is
a “symbol” toolbox, which performs does symbolic operations such as differen-
tiation, integration, matrix inversion, and solution of differential equations when
the operations are expressed in their symbolic form. In more recent versions of
this software, new toolboxes such as Filter Design Toolbox and DSP Blockset,
which are based on the object-oriented programming features of MATLAB, have
been added. These have been treated in some chapters of this book.

9.1.1 Vectors, Arrays, and Matrices

Vectors and scalars are special cases of a matrix—all of which are represented
as arrays in MATLAB. A scalar is an array of 1 × 1 dimension, whereas a row
vector is an array of 1 × n dimension, and a column vector is an array of n × 1
dimension. When elements of an array are typed in a row within square brackets,
with a space between the elements, MATLAB displays it as a row vector. For
example, when you type

>>A = [1 2 0 3 1 5]

it displays

A = 1 2 0 3 1 5

If the array is typed without assigning a name for the array, >>[1 2 0 3 1 5],
MATLAB responds with

ans = 1 2 0 3 1 5

When you type elements with a semicolon between them, the elements are dis-
played in a column vector, for example

>>B=[1 2 0; 3 1 5; 0 4 -2] displays the 3 -by-3 matrix

B= 1 2 0
3 1 5
0 4 -2

If a semicolumn is entered at the end of an array or a command, then the

array and the output of the command is not displayed on the command window,
but the command as well as the output variables are saved in a buffer known
as the workspace. The workspace saves the variables, data, contents of arrays
and matrices, and other elements as well as a record of the commands typed
by the user. It is recommended that at the beginning of the session, you change
 INTRODUCTION 393

the directory to the disk drive a: if you have one in your computer so that the
contents of the workspace are saved in the floppy disk in that drive. Instead of
using the semicolumn between the elements, you can type the element on the
next line or by leaving one space, type three dots at the end of the line and
continue on the next line as shown below; this is useful when the array is very
long and extends beyond the end of the line:

>>C=[1 2 0
3 1 5
0 4 -2]
or
>>C=[ 1 2 0; 3 1 5; ...
0 4 -2]
displays the answer
C = 1 2 0
3 1 5
0 4 -2

9.1.2 Matrix Operations

It is now obvious that a column vector can be created by typing the elements
with a semicolumn separating them or creating a row vector and transposing
it. In MATLAB, the transpose of a matrix or a vector is carried out by the
operator, that is, the command x’ gives the transpose of the vector or matrix x.
Since the vectors and matrices listed and described above have been saved in the
workspace, if we type >>A’, we get

ans
1
2
0
3
1
5

and >>D=C’ yields

D =
1 3 0
2 1 4
0 5 -2

If we type C(:), we get a column vector with the columns of C arranged in

a vertical vector:

ans =
1
394 MATLAB PRIMER

3
0
2
1
4
0
5
-2

When a scalar, vector, or matrix is multiplied (or divided) by a scalar c, every

element of the scalar, vector or matrix is multiplied (or divided) by c. When
c is added to matrix, it is added to every element of the vector or matrix. For
example, x = 5; F=x*C gives the output

F= 5 10 0
15 5 25
0 20 -10
FF=x+C gives the output as
FF= 6 7 5
8 6 10
5 9 3

Addition and subtraction of two matrices (and vectors) is carried out by

MATLAB, according to the rules of matrix algebra, when they have the same
dimension. Multiplication of a vector or matrix by a vector or matrix is carried
out according to the rules of algebra when they are compatible or commensu-
rate for multiplication. The matrix operations and their corresponding notations
available in MATLAB are given below:

Addition +
Subtraction -
Multiplication *
Power or exponent ^
Transpose ’
Left division \
Right division /

Note that the command x=M\b gives us the solution to the equation M*x=b,
where M is a square matrix that is assumed to be nonsingular. In matrix algebra,
the solution is given by x = M −1 b. The left division is a more commonly used
operation in application of matrix algebra. (The command for the right division
x=b/M gives the solution to the equation x*M=b, assuming that x and M are
compatible for multiplication and the solution in matrix algebra is given by
x=bM−1 .)
 INTRODUCTION 395

When we use the same variables used above but define them with new values,
they become the current values for the variables, so we define and use them
below as examples of the operations described above:
>>A=[1 2 1;0 1 1;2 1 1];
>>B=[2 1 0;1 1 1;-1 2 1];
>>C=A+B
C=
3 3 1
1 2 2
1 3 2
>>D=A*B
D=
3 5 3
0 3 2
4 5 2
>>M=A;
>>b=[2;4;4];
>>x=M\b
x = 0.0000 -2.0000 6.0000

Whereas the addition and subtraction of matrices are carried out by the addition
and subtraction term by term from the corresponding positions of the elements,
we know that the multiplication and “division” of matrices follow different rules.
MATLAB gives the correct answer in all the preceding operations. It has another
type of operation that is carried out when we use a dot before the sign for the
mathematical operation between the two matrices. The multiplication (.*), divi-
sion (./), and exponentiation (.^) of the terms in the corresponding positions
of the two compatible matrices are the three array operations.
Instead of multiplying the two matrices as D=A*B, now we type a dot before
the sign for multiplication. For example, the answer to >>D=A.*B is
D=
2 2 0
0 1 1
-2 2 1

It is easy to see the result of the command >>A^2 = A ∗ A as

ans =
3 5 4
2 2 2
4 6 4

Let us define a matrix

 
1 2
X =
3 4
396 MATLAB PRIMER

Now we compute U=X.^2 and V=2.^X and get the following outputs:

>>U=X.^2
 
12 22
U =
32 42
>>V=2.^X:
 
21 22
V =
23 24
A matrix can be expanded by adding new matrices and column or row vectors
as illustrated by the following examples:

>>F=[A B]
F=
1 2 1 2 1 0
0 1 1 1 1 1
2 1 1 -1 2 1
>>b=[5 4 2];
>>G=[A;B;b]
G =
1 2 1
0 1 1
2 1 1
2 1 0
1 1 1
-1 2 1
5 4 2

The division operator ./ can be used to divide a scalar by each of the matrix
element as shown below, provided there are no zeros in the matrix:

>>W = 12./X produces the result

W= 12 6
4 3
>> WW= [6 2; 2 3];

W./WW divides the elements of W by the elements of WW term by term:

ans = 2 3
2 1

The element in the (i,j) position of a matrix G is identified by typ-

ing >>G(7,2), and we get ans = 4, and we can change its value by typing
>>G(7,2)=6, so that now we have
 INTRODUCTION 397

G =
1 2 1
0 1 1
2 1 1
2 1 0
1 1 1
-1 2 1
5 6 2
The colon sign : can be used to extract a submatrix from a matrix as shown by
the following examples:
⎡ ⎤
2 5 6
⎢ ⎥
>> Q =⎣ 3 2 4⎦
−3 1 8
>>Q(:,2) gives a submatrix with elements in all rows and the second column only:
ans =
5
2
1
The command Q(3,:) gives the elements in all columns and the third row only:
ans =
-3 1 8
The command Q(1:2,2:3) gives the elements in the rows from 1 to 2 and in
the columns from 2 to 3:
ans =
5 6
2 4
There are many other operations that can be applied on a matrix, such as A,
as listed below:

MATRIX OPERATIONS

rot90(A) Rotates the matrix array by 90◦

fliplr(A) Flips the columns left to right
flipud(A) Flips the rows up to down
triu(A) Gives the upper triangular part of the matrix
tril(A) Gives the lower triangular part of the matrix

There are a few special matrices; we will list only three that are often found
useful in manipulating matrices:
ones(m,n), which gives a matrix with the number one in all its m rows and
n columns
398 MATLAB PRIMER

zeros(m,n), which gives a matrix with zeros in all its m rows and n columns
eye(m), which gives the “identity matrix” of order m × m.

We note that the inverse of a matrix A is obtained from the function inv(A),
the determinant of a matrix A is obtained from the function det(A) and the rank
from rank(A).
Since this is only a primer on MATLAB, it does not contain all the information
on its functions. You should refer to the user’s guide that accompanies every
software program mentioned above or any other books on MATLAB [1–3]. When
you have logged on to MATLAB or any of the subdirectories for the toolboxes,
there is an online help readily available. You type help functionname, where
functionname is the name of the function on which detailed information is
desired, and immediately that information is displayed on the command window.
So there is no need to memorize the syntax and various features of the function
and so on. The best way to learn the use of MATLAB and the toolboxes is to
try the functions on the computer, using the help utility if necessary.

9.1.3 Scalar Operations

If t = (0.1π) radians per second, the MATLAB function sin(t) gives the answer
as 0.3090. To compute and plot v = sin(0.1π t), in the time interval [0 2π]
we have to choose discrete values for the continuous variable t and compute v
at these values. To do so, we create an array t=[0.0:0.1:2.0]; this gives the
sequence of 21 values within t = 0.0–2.0 in increments of 0.1. Now if we type
the function v=sin(pi*t), the result is a sequence of 21 values. The command
stem(v) immediately plots these values for v in a “graphics window” as shown
in Figure 9.1.
We have used the command figure to create a new window and the command
plot(t,v) to get the plot as a continuous plot joining the discrete values of v
(see Fig. 9.2). If we did not use the command figure, the second figure would
replace the first one on the graphics window. Typing the command grid on the
command window results in the plots having grid lines shown in these figures.
Next we use the commands for adding a title and the labels for the y and x
coordinates with the commands title, ylabel, and xlabel.
So the commands entered on the command window to get the two figures are as
follows. Since MATLAB chooses the scales for the x and y coordinates, depend-
ing on the range of their values, we may have to change the literal arguments in
the ylabel and xlabel:

t = [0.0:0.1:2.0];
v=sin(pi*t);
stem(v);grid
title(’Values of sin(pi*t)’)
ylabel(’Values of sin(pi*t)’)
xlabel(’Values of 10t’)
figure
 INTRODUCTION 399

Plot of sin (pi∗t)

1

0.8

0.6

0.4
Values of sin (pi∗t)

0.2

−0.2

−0.4

−0.6

−0.8

−1
0 5 10 15 20 25
Values of 10t

Figure 9.1 Plot of sin(pi*nT).

Plot of sin (pi∗t)

1

0.8

0.6

0.4
Values of sin (pi∗t)

0.2

−0.2

−0.4

−0.6

−0.8

−1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Values of 10t

Figure 9.2 Plot of sin(pi*t).

plot(t,v);grid
title(’Plot of sin(pi*t)’)
ylabel(’Value of sin(pi*t)’)
xlabel(’Value of t’)
400 MATLAB PRIMER

9.1.4 Drawing Plots

Additional arguments can be added in the command plot(t,v) to specify the
color of the curve; for example, plot(t,v,’g’) will show the curve in green
color. The arguments for other colors are as follows:

y yellow
m magenta
c cyan
r red
b blue
w white
k black

The next argument that can be added is a marker used to draw the curve. For
example, plot(t,v,’g’,’+’) will plot the curve with the + sign instead of
the line curve, which is the default marker. Other markers that are available are

o circle
. point
* star
- solid line
: dotted line
-- dashed line
-.- dash–dot–dash

One can plot several curves in the same figure; for example, we can plot both
v and y versus t by the command plot(t,v,’g’,’-’t,y,’r’,’*’). Another
way of plotting more than one variable in the same figure is to use the command
hold on after plotting the first variable and then typing the command for plotting
the second variable:

plot(t,v,’g’);
hold on
plot(t,y,’r’)

The use of the MATLAB commands subplot grid, and axis have been
described and used earlier in the book. The commands gtext and ginput are
also very useful in plotting. There is a tool called fvtool (filter visualization
tool) in the more recent versions of the Signal Processing Toolbox, which offers
several other features in plotting the response of digital filters. You may type help
gtext, help ginput, or help fvtool to get more information about them.

9.1.5 MATLAB Functions

The other functions, in addition to sin, that are available in MATLAB are given
below:
 INTRODUCTION 401

TRIGONOMETRIC FUNCTIONS
sin sine cos cosine
tan tangent asin arcsine
acos arccosine atan arctangent
atan2 four-quadrant arctangent sinh hyperbolic sine
cosh hyperbolic cosine tanh hyperbolic tangent
asinh hyperbolic arcsine acosh hyperbolic arccosine
atanh hyperbolic arctangent

MATHEMATICAL FUNCTIONS
abs absolute value or magnitude
angle phase angle of a complex number
sqrt square root
real real part of a complex number
imag imaginary part
conj complex conjugate
round round toward nearest integer
fix round toward zero
floor round toward −∞
ceil round toward ∞
sign signum function
rem remainder
exp exponential base 2
log natural logarithm
log10 log base 10

9.1.6 Numerical Format

We can specify the format in which MATLAB displays numbers. If the num-
ber is an integer, then by default, it is displayed as an integer. If it is a real
number, it is displayed with approximately four digits to the right of the dec-
imal point (e.g., 12.0945), and this is the default format format short. If
the number has many more significant digits, we specify other formats, using
the scientific notation. For example, let the number be 12.094567832155321.
If we type the MATLAB command format long, this number will be dis-
played with 16 digits as 12.09456783215532. If we declare the format short
e, the number will be displayed with five digits and an exponent in the form
1.2094e+01, whereas the command format long e selects 16 digits and an
exponent: 1.209456783215532e+01.
Remember that these formats are used for display on the monitor, the result
of commands, functions, and operations. But the numerical computations that
implement the functions and scripts are done by MATLAB with a higher degree
402 MATLAB PRIMER

of precision if and when it is necessary, for example, when we use the functions
and scripts in the Signal Processing Toolbox.

9.1.7 Control Flow

Three functions are used to control the flow of command execution that depend
on decisionmaking statements. Such functions are found in other programming
languages and also in MATLAB. They are for loops, If-elseif-end loops
and while loops, which will be illustrated below.
The statement
>> for n=1:10
x(n )=n^2+4*n
end

produces a vector containing 10 values of x(n) = n2 + 4n, for n = 1, 2, 3, . . . , 10.

One can define an array such as n=3:-0.5:-1.0, in which the increment
is −0.5 and the result is an array n = [3.0 2.5 2.0 1.5 1.0 0.5 0.0
-0.5 -1.0]. The default value for the increment is 1.
To define a two-dimensional array, and a function H(i,j)=0.1^i+0.2^j, we
use the statements
for i =1:20;
for j =1:20;
X(i,j) = 0.1^i+0.2^j
end
end

An example of the use of the if statement is

>>n=-10:10
if n<0
x(n)=0;
elseif 0≤n≤5;
x(n)=(0.8).^n;
else
x(n)=0
end

Note that an error message will be shown if we use the statement

x(n) = (0.8)^n without the dot before the exponent.
The while loop is executed step by step as long as the relation holds true.
An example of this is
>>n=1
while n<8
x(n)=0.5^n;
n=n+1
end
 INTRODUCTION 403

If the values of x(n) when n takes the maximum value of 7 is desired, we

type x after the end statement and get the result

0.5000 0.2500 0.1250 0.0625 0.0313 0.0156 0.0078

But this problem is solved more easily by the following two statements to get
the same values for x(n), n = 1, 2, 3, . . . , 7, but we have to insert a dot before
the exponent (^) since n is a row vector of seven elements. It is very helpful to
find the order of a matrix or a vector A by using the statement S= size(A) to
know when to use the dot for the term-by-term operation—particularly when we
get an error message about the dimensions of the matrices:

>>n=1:7;
x(n)=(0.5).^n

9.1.8 Edit Window and M-file

So far we have introduced a few of the common functions and operations of
MATLAB that are used in the command window and the graphics window.
When we are in the command window, we are in an interactive mode, where each
command is executed immediately after it is typed and the answer is displayed. If
we wish to make a change in any one of the previous statements or the input data,
we have to trace it back one line at a time using the ↑ key and edit the line, then
use the ↓ key to get to the line where we had stopped. If that statement is very
many lines before the current line or if we want to make major changes in the
program, or if we wish to find the output of the program for different values for
the input parameters, this is not a convenient procedure. So we create a program
or a script by clicking File-Open-New-M-file and use a text editor that is
built-in MATLAB or any other text editor and save it in the current directory
as a file with a name and an .m extension. We can write this script using the
MATLAB functions and operations; we can even use other functions or functions
that we have written. Such a file is called an M-file, and after it is saved, we click
the command window and type just the name of the M-file without the extension.
The entire script is executed if there are no bugs in it and the results displayed. If
there are any error messages, we go back to M-file in the edit window and make
corrections, save it, get back to the command window, and then type the name
of the M-file to run it again. Either we enter the values of the input variable(s)
in the M-file or add the following command in the M-file:

input(’Type in the input parameters for xyz’)

When the script is to be executed, the program displays the statement Type in
the input parameters for xyz and waits for the input from the keyboard.
We may have requests for input for several parameters, and when the data for
all the parameters are entered by us from the keyboard, the program is executed.
404 MATLAB PRIMER

This is helpful when we wish to find the response (output) of the program with
different values for the input parameters.
Similarly, when we add the statement

disp(’Values of the parameter xx’)

disp(xx)

the program displays the values for the parameter after the script has been
executed, which may not be otherwise displayed as the output from the pro-
gram.
When any statement is preceded by the % character, the statement is not exe-
cuted by the program; it is used only as a comment for information or explanation
of what the program does. It is a good practice to add a few lines with this %
character at the beginning of any script that we write and include the name of
the file also.

Example 9.1

We click File, Open, New from the menu bar when we are in the command
window and then choose M-file. An edit window appears next. Now we give an
example of a M-file that we write using the built-in text editor:

>>clear %clears all the variables in the Workspace

%This program we call Ration.m computes the value of a
% rational function f(x) with a numerator 2x+0.5 and a
% denominator x^2+0.1x+0.05 and plots it over the interval
% 0 ≤ x ≤ 1.0.
>>x=0.0:0.01:1.0;
>>num=2*x+0.5;
>>den=x.^2+0.1*x+0.05; %Note the dot in the first term on
% the right side expression
>>val=num./den;
>>plot(x,val);grid
title(’Plot of the function f(x)’)
ylabel(’Value of f(x)’)
xlabel(’Value of x’)

This file is saved with a name Ration.m on the current drive, and then we get
back to the command window, in which we type >>Ration. All the statements
of the M-file Ration.m are executed immediately, and the plot is shown in the
graphics window (see Fig. 9.3). If there are any error messages, we launch the
file in the edit window, then edit and correct the statements where necessary.
This example is a simple one, but we have many examples of M-files as well as
files used in an interactive mode discussed earlier in the book.
 SIGNAL PROCESSING TOOLBOX 405

Plot of the function f(x)

11

10

8 f(x) = 2x + 0.5/x2 + 0.1x + 0.05

Value of f(x)

2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Value of x

Figure 9.3 Plot of (2x + 0.5)/(x 2 + 0.1x + 0.05).

9.2 SIGNAL PROCESSING TOOLBOX

The Signal Processing Toolbox is a collection of about 160 functions that are
extensively used for the analysis, design, and realization of discrete-time sys-
tems and tasks or operations such as modeling, detection, filtering, prediction,
and spectral analysis in digital signal processing. They run under MATLAB,
which has about 330 functions and operations. By typing help function in
the command window, where function is the name of these functions, detailed
information about them is displayed. By typing help signal, we get a com-
plete list of all the functions in the Signal Processing Toolbox, when this has
been installed as a subdirectory of the MATLAB directory. If we know the name
of the function that does the numerical processing but not the syntax and other
details, we can type help function. But when we have to carry out numerical
processing but don’t know the name of the MATLAB function, we may have to
go through the list of all MATLAB functions and choose the appropriate one for
the purpose. The list of all MATLAB functions in the Signal Processing Toolbox
is given in Section 9.2.1, and students are encouraged to use the help utility and
become familiar with as many of the functions as possible. That should improve
their efficiency in calling up the appropriate function immediately when the need
arises while they write and edit the script. Note that we can use any of the thou-
sands of functions found in all other toolboxes and in the simulation software
called Simulink that runs under MATLAB, which makes this software extremely
powerful and versatile.
406 MATLAB PRIMER

9.2.1 List of Functions in Signal Processing Toolbox

>>help signal
Signal Processing Toolbox
Version 6.0 (R13) 20-Jun-2002

Filter analysis.

abs - Magnitude.
angle - Phase angle.
filternorm - Compute the 2-norm or inf-norm of a digital
filter.
freqs - Laplace transform frequency response.
freqspace - Frequency spacing for frequency response.
freqz - Z-transform frequency response.
fvtool - Filter Visualization Tool.
grpdelay - Group delay.
impz - Discrete impulse response.
phasez - Digital filter phase response.
phasedelay - Phase delay of a digital filter.
unwrap - Unwrap phase.
zerophase - Zero-phase response of a real filter.
zplane - Discrete pole-zero plot.

Filter implementation.

conv - Convolution.
conv2 - 2-D convolution.
convmtx - Convolution matrix.
deconv - Deconvolution.
fftfilt - Overlap-add filter implementation.
filter - Filter implementation.
filter2 - Two-dimensional digital filtering.
filtfilt - Zero-phase version of filter.
filtic - Determine filter initial conditions.
latcfilt - Lattice filter implementation.
medfilt1 - 1-Dimensional median filtering.
sgolayfilt - Savitzky-Golay filter implementation.
sosfilt - Second-order sections (biquad) filter
implementation.
upfirdn - Up sample, FIR filter, down sample.

Discrete-time filter object.

dfilt - Construct a discrete-time, filter object.

(Type ’’doc dfilt’’ for more information)
 SIGNAL PROCESSING TOOLBOX 407

FIR filter design.

cremez - Complex and nonlinear phase equiripple FIR filter

design.
fir1 - Window based FIR filter design - low, high, band,
stop, multi.
fir2 - FIR arbitrary shape filter design using the frequency
sampling method.
fircls - Constrained Least Squares filter design - arbitrary
response.
fircls1 - Constrained Least Squares FIR filter design - low
and highpass.
firgauss - FIR Gaussian digital filter design.
firls - Optimal least-squares FIR filter design.
firrcos - Raised cosine FIR filter design.
intfilt - Interpolation FIR filter design.
kaiserord - Kaiser window design based filter order
estimation.
remez - Optimal Chebyshev-norm FIR filter design.
remezord - Remez design based filter order estimation.
sgolay - Savitzky-Golay FIR smoothing filter design.

IIR digital filter design.

butter - Butterworth filter design.

cheby1 - Chebyshev Type I filter design (passband ripple).
cheby2 - Chebyshev Type II filter design (stopband ripple).
ellip - Elliptic filter design.
maxflat - Generalized Butterworth lowpass filter design.
yulewalk - Yule-Walker filter design.
IIR filter order estimation.
buttord - Butterworth filter order estimation.
cheb1ord - Chebyshev Type I filter order estimation.
cheb2ord - Chebyshev Type II filter order estimation.
ellipord - Elliptic filter order estimation.

Analog lowpass filter prototypes.

besselap - Bessel filter prototype.

buttap - Butterworth filter prototype.
cheb1ap - Chebyshev Type I filter prototype (passband
ripple).
cheb2ap - Chebyshev Type II filter prototype (stopband
ripple).
408 MATLAB PRIMER

ellipap - Elliptic filter prototype.

Analog filter design.
besself - Bessel analog filter design.
butter - Butterworth filter design.
cheby1 - Chebyshev Type I filter design.
cheby2 - Chebyshev Type II filter design.
ellip - Elliptic filter design.

Analog filter transformation.

lp2bp - Lowpass to bandpass analog filter transformation.

lp2bs - Lowpass to bandstop analog filter transformation
lp2hp - Lowpass to highpass analog filter transformation.
lp2lp - Lowpass to lowpass analog filter transformation.

Filter discretization.

bilinear - Bilinear transformation with optional

prewarping.
impinvar - Impulse invariance analog to digital conversion.

Linear system transformations.

latc2tf - Lattice or lattice ladder to transfer function

conversion.
polystab - Polynomial stabilization.
polyscale - Scale roots of polynomial.
residuez - Z-transform partial fraction expansion.
sos2ss - Second-order sections to state-space conversion.
sos2tf - Second-order sections to transfer function
conversion.
sos2zp - Second-order sections to zero-pole conversion.
ss2sos - State-space to second-order sections conversion.
ss2tf - State-space to transfer function conversion.
ss2zp - State-space to zero-pole conversion.
tf2latc - Transfer function to lattice or lattice ladder
conversion.
tf2sos - Transfer Function to second-order sections
conversion.
tf2ss - Transfer function to state-space conversion.
tf2zpk - Discrete-time transfer function to zero-pole
conversion.
zp2sos - Zero-pole to second-order sections conversion.
zp2ss - Zero-pole to state-space conversion.
zp2tf - Zero-pole to transfer function conversion.
 SIGNAL PROCESSING TOOLBOX 409

Windows.

bartlett - Bartlett window.

barthannwin - Modified Bartlett-Hanning window.
blackman - Blackman window.
blackmanharris - Minimum 4-term Blackman-Harris window.
bohmanwin - Bohman window.
chebwin - Chebyshev window.
flattopwin - Flat Top window.
gausswin - Gaussian window.
hamming - Hamming window.
hann - Hann window.
kaiser - Kaiser window.
nuttallwin - Nuttall defined minimum 4-term Blackman-Harris
window.
parzenwin - Parzen (de la Valle-Poussin) window.
rectwin - Rectangular window.
triang - Triangular window.
tukeywin - Tukey window.
wvtool - Window Visualization Tool.
window - Window function gateway.

Window object.

sigwin - Construct a window object.

(Type ’’doc sigwin’’ for more information)

Transforms.

bitrevorder - Permute input into bit-reversed order.

czt - Chirp-z transform.
dct - Discrete cosine transform.
dftmtx - Discrete Fourier transform matrix.
digitrevorder - Permute input into digit-reversed order.
fft - Fast Fourier transform.
fft2 - 2-D fast Fourier transform.
fftshift - Swap vector halves.
goertzel - Second-order Goertzel algorithm.
hilbert - Discrete-time analytic signal via Hilbert
transform.
idct - Inverse discrete cosine transform.
ifft - Inverse fast Fourier transform.
ifft2 - Inverse 2-D fast Fourier transform.
410 MATLAB PRIMER

Cepstral analysis.

cceps - Complex cepstrum.

icceps - Inverse Complex cepstrum.
rceps - Real cepstrum and minimum phase reconstruction.

Statistical signal processing and spectral analysis.

cohere - Coherence function estimate.

corrcoef - Correlation coefficients.
corrmtx - Autocorrelation matrix.
cov - Covariance matrix.
csd - Cross Spectral Density.
pburg - Power Spectral Density estimate via Burg’s method.
pcov - Power Spectral Density estimate via the Covariance
method.
peig - Power Spectral Density estimate via the Eigenvector
method.
periodogram - Power Spectral Density estimate via the
periodogram method.
pmcov - Power Spectral Density estimate via the Modified
Covariance method.
pmtm - Power Spectral Density estimate via the Thomson
multitaper method.
pmusic - Power Spectral Density estimate via the MUSIC
method.
psdplot - Plot Power Spectral Density data.
pwelch - Power Spectral Density estimate via Welch’s
method.
pyulear - Power Spectral Density estimate via the
Yule-Walker AR Method.
rooteig - Sinusoid frequency and power estimation via the
eigenvector algorithm.
rootmusic - Sinusoid frequency and power estimation via
the MUSIC algorithm.
tfe - Transfer function estimate.
xcorr - Cross-correlation function.
xcorr2 - 2-D cross-correlation.
xcov - Covariance function.

Parametric modeling.

arburg - AR parametric modeling via Burg’s method.

arcov - AR parametric modeling via covariance
method.
 SIGNAL PROCESSING TOOLBOX 411

armcov - AR parametric modeling via modified covariance

method.
aryule - AR parametric modeling via the Yule-Walker method.
ident - See the System Identification Toolbox.
invfreqs - Analog filter fit to frequency response.
invfreqz - Discrete filter fit to frequency response.
prony - Prony’s discrete filter fit to time response.
stmcb - Steiglitz-McBride iteration for ARMA modeling.

Linear Prediction.

ac2rc - Autocorrelation sequence to reflection coefficients

conversion.
ac2poly - Autocorrelation sequence to prediction polynomial
conversion.
is2rc - Inverse sine parameters to reflection coefficients
conversion.
lar2rc - Log area ratios to reflection coefficients
conversion.
levinson - Levinson-Durbin recursion.
lpc - Linear Predictive Coefficients using autocorrelation
method.
lsf2poly - Line spectral frequencies to prediction
polynomial conversion.
poly2ac - Prediction polynomial to autocorrelation sequence
conversion.
poly2lsf - Prediction polynomial to line spectral
frequencies conversion.
poly2rc - Prediction polynomial to reflection coefficients
conversion.
rc2ac - Reflection coefficients to autocorrelation sequence
conversion.
rc2is - Reflection coefficients to inverse sine parameters
conversion.
rc2lar - Reflection coefficients to log area ratios
conversion.
rc2poly - Reflection coefficients to prediction polynomial
conversion.
rlevinson - Reverse Levinson-Durbin recursion.
schurrc - Schur algorithm.

Multirate signal processing.

decimate - Resample data at a lower sample rate.

downsample - Downsample input signal.
412 MATLAB PRIMER

interp - Resample data at a higher sample rate.

interp1 - General 1-D interpolation. (MATLAB Toolbox)
resample - Resample sequence with new sampling rate.
spline - Cubic spline interpolation.
upfirdn - Up sample, FIR filter, down sample.
upsample - Upsample input signal.

Waveform generation.

chirp - Swept-frequency cosine generator.

diric - Dirichlet (periodic sinc) function.
gauspuls - Gaussian RF pulse generator.
gmonopuls - Gaussian monopulse generator.
pulstran - Pulse train generator.
rectpuls - Sampled aperiodic rectangle generator.
sawtooth - Sawtooth function.
sinc - Sinc or sin(pi*x)/(pi*x) function
square - Square wave function.
tripuls - Sampled aperiodic triangle generator.
vco - Voltage controlled oscillator.

Specialized operations.

buffer - Buffer a signal vector into a matrix of data

frames.
cell2sos - Convert cell array to second-order-section
matrix.
cplxpair - Order vector into complex conjugate
pairs.
demod - Demodulation for communications simulation.
dpss - Discrete prolate spheroidal sequences
(Slepian sequences).
dpssclear - Remove discrete prolate spheroidal
sequences from database.
dpssdir - Discrete prolate spheroidal sequence
database directory.
dpssload - Load discrete prolate spheroidal sequences
from database.
dpsssave - Save discrete prolate spheroidal sequences
in database.
eqtflength - Equalize the length of a discrete-time
transfer function.
modulate - Modulation for communications simulation.
seqperiod - Find minimum-length repeating sequence in a
vector.
 SIGNAL PROCESSING TOOLBOX 413

sos2cell - Convert second-order-section matrix to cell

array.
specgram - Spectrogram, for speech signals.
stem - Plot discrete data sequence.
strips - Strip plot.
udecode - Uniform decoding of the input.
uencode - Uniform quantization and encoding of the input
into N-bits.

Graphical User Interfaces

fdatool - Filter Design and Analysis Tool.

fvtool - Filter Visualization Tool.
sptool - Signal Processing Tool.
wintool - Window Design and Analysis Tool.
wvtool - Window Visualization Tool.

See also SIGDEMOS, AUDIO, and, in the Filter Design Toolbox, FILTERDESIGN.
If we type help functionname, we get information about the syntax and
use of the function and so on, but if we type type functionname, we get
the program listing also. An example of this given below; one can modify any
function, save it with a different name and run it:

>> type kaiser

function w = kaiser(n est,beta)
%KAISER Kaiser window.
% W = KAISER(N,BETA) returns the BETA-valued N-point Kaiser
% window.
%
% See also BARTLETT, BARTHANNWIN, BLACKMAN, BLACKMANHARRIS,
% BOHMANWIN,
% CHEBWIN, GAUSSWIN, HAMMING, HANN, NUTTALLWIN, RECTWIN,
% TRIANG,
% TUKEYWIN, WINDOW.
% Author(s): L. Shure, 3-4-87
% Copyright 1988-2002 The MathWorks, Inc.
% $Revision: 1.15 $ $Date: 2002/03/28 17:28:33 $
error(nargchk(2,2,nargin));
[nn,w,trivialwin] = check order(n est);
if trivialwin, return, end;
nw = round(nn);
bes = abs(besseli(0,beta));
odd = rem(nw,2);
exind = (nw-1)^2;
n = fix((nw+1)/2);
xi = (0:n-1) + .5*(1-odd);
414 MATLAB PRIMER

xi = 4*xi.^2;
w = besseli(0,beta*sqrt(1-xi/xind))/bes;
w = abs([w(n:-1:odd+1) w])’;
% [EOF] kaiser.m

REFERENCES

1. D. Hanselman and B. Littlefield, Mastering MATLAB 5, Prentice Hall, 1996.

2. The MathWorks, Inc., MATLAB User’s Guide, 1993.
3. D. M. Etter and D. C. Kuncicky, Introduction to MATLAB 6, Prentice-Hall, 2002.
4. W. J. Palm III, Introduction to MATLAB 6 for Engineers, McGraw-Hill, 2001.
5. J. N. Little and L. Shure, Signal Processing Toolbox for Use with MATLAB, User’s
Guide, The MathWorks, Inc., 1994.
6. S. K. Mitra, Digital Signal Processing Laboratory Using MATLAB (R) , McGraw-Hill,
1999.
 INDEX

Active-RC (resistance x capacitance) filters, transformations

19–20, 28 bandpass filter, 213–216
Adaptive equalization, 2 bandstop filter, 216–218
Adaptive filters, 24 bilinear transformations, 223
Adders, 33, 35, 61, 68 highpass filter, 212–213
ADPCM coding, 2 Analog lowpass filters, 21, 323
Aerospace electronics, 3 Analog signal processing, 22–25, 177
Aliasing, 117, 220 Analog signals, discrete-time system, 7
Allpass filters Analog systems, 7
characteristics of, 230–231 Analog-to-digital converter (ADC), 7, 22,
in parallel, realizations of 27–28, 355, 383
design procedure, 325–326 ANSI Standard C code, 385
lattice-ladder realization, 326–327 Antialiasing filters, 22, 119
properties of, 320–325, 334 Anti-mirror image polynomial, 259
quantized filter analysis, 370, 372–375 Antisymmetric coefficient, linear phase FIR
realization using MATLAB, 339–346 filters, 254–256, 259
All-pole (AR) filters, realizations, 333–334 Application-specific integrated circuits
Analog bandpass signal, 120, 126–127 (ASICs), 354
Analog Devices, 389 Associative convolution sum, 66
Analog filters, magnitude approximation Attenuation, 194–195, 213, 232–233
bandpass filter,187 Automotive electronics, 3
bandstop filter, 187 Autoregressive moving-average (ARMA)
Butterworth lowpass filters, design theory of, filter realization, 326, 333, 359
194–201 quantized filter analysis, 366, 370–372
Butterworth response, 192–194
Chebyshev I approximation, 202 Bandpass (BP) filters
Chebyshev I lowpass filters, design theory analog, 187
of, 204–208 digital, 213–216
Chebyshev II approximation, 208–209 discrete-time Fourier transform
Chebyshev II lowpass filters, design of, (DTFT), 126
210–211 equiripple FIR filter design, 286, 389
Chebyshev polynomials, properties of, linear phase FIR, 261–263, 270–271
202–204 windowed FIR, 275
elliptic function approximation, 212 Bandpass signals, sampling, 120–121
highpass filter, 187 Bandstop (BS) filter
lowpass filter, 187 analog, 187
maximally flat magnitude response, 191–192 digital, 216–218
overview of, 189–191 frequency-domain analysis, 122, 124, 126
Analog frequency linear phase FIR, 261–262
finite impulse response (FIR) filters, 251 Bandwidth, digital signal processing, 24

Introduction to Digital Signal Processing and Filter Design, by B. A. Shenoi

Copyright © 2006 John Wiley & Sons, Inc.

415
416 INDEX

Bartlett window, finite impulse response (FIR) Circuit model, discrete-time system, 71–73
filters, 266, 268–269 Closed-form expression, 65, 155
Base station controller (BSC), 25 Code-division multiple access (CDMA)
Base transceiver stations (BTSs), 25–27 technology, 2, 25
Bessel function, 268 Common-object file format (COFF), 387–388
Bilinear transformations, infinite impulse Complementary function/complementary
response (IIR) filters, 221–226 solution, 58
Binary coding, 6 Complementary metal oxide semiconductor
Binary numbers, in quantized filter analysis, (CMOS) transistors, 19, 23
360–367 Complex conjugate poles, 51–54
Binomial theorem, 203 Complex conjugate response, discrete-time
Biomedical systems, 2, 354 Fourier transform, 145
Blackman window, finite impulse response Computed tomography (CT) scanning, 2
(FIR) filters, 266, 268 Computer networking technology, 27
Bode plot, 149 Conjugation property, discrete-time Fourier
Bone scanning, 2 transform, 145
Bounded-input bounded output (BIBO) Consumer electronics, 2
stability, 77–78 Continuous-time filters, see Analog filters
Butterworth bandpass digital filters, 221, 236 Continuous-time function, 113
Butterworth lowpass filters Continuous-time signal, 3–4, 21, 28, 41–42
design theory of, 194–201 Continuous-time systems, 24
filter realization Convolution
generally, 323–324 allpass filters, 325
using MATLAB, 334–337 defined, 25
Butterworth magnitude response, 192–194 discrete-time Fourier series (DTFS),
Butterworth polynomials, 197–198 164–169
linear phase finite impulse response (FIR)
C/C++ language, 385–386 filters, 265
Canonic realization, FIR filters, 309–310 Convolution sum
Cardiac pacemakers, 2 discrete-time Fourier transform
Cascade realization (DTFT), 125
finite impulse response (FIR) filters, filter realizations, 304
306–307 time-domain analysis, 38–41, 82, 94
infinite impulse response (IIR) filters, z-transform theory, 65–70
313–317, 329–331, 366 Cooley–Tukey algorithm, 21
Cauer filter, 212 Cos(ω0 n), properties of, 14–19
Causal sequence, 9, 133 CPU (central processing unit), 384
Causal system, 33 Cramer’s rule, 62
Cell phones, 354 Cutoff frequency
Cell repeat pattern, mobile network system, 26 finite impulse response (FIR) filters, 266,
Channel coding, 2 293–294
Characteristic roots, 58 frequency-domain analysis, 141
Chebyshev (I/II) approximation, 189, 202, infinite impulse response (IIR) filters, 213,
208–209, 284 226–227
Chebyshev (I) bandpass filter, 125, 215, 235 linear phase FIR filters, 272
Chebyshev (I/II) highpass filters, 213, 237–238
Chebyshev (I/II) lowpass filters Data encryption, 2
characterized, 323–324 Decryption, 2
design of, 210–211 Delay, see also Group delay
design theory of, 204–208 defined, 33
realization using MATLAB, 334–337 equalizers, 231, 321
Chebyshev polynomials, properties of, hardware containing, 68
202–204 z-transform theory, 46–49
Circuit boards, filter design and, 19 Demodulation, 25
 INDEX 417

Difference equations, time-domain analysis time-domain analysis of noncausal inputs,

classical method, 59–64 125–127
z-transform theory, 51–56 time reversal property, 128–138, 139
Differentiation property, discrete-time Fourier time-shifting property, 127, 139
transform, 139–142 of unit step sequence, 138–147
Differentiation, z-transform theory, 44–46 Discrete-time function, 122
Digital computers, filter design and, 20–21 Discrete-time sequence, 122
Digital filter, see specific types of filters Discrete-time series synonymously, 122
characteristics of, generally, 6–7, 219 Discrete-time signals
designing, 123, 126 characterized, 3–8
Digital signal, defined, 6 complex exponential function, 12–14
Digital signal processing constant sequence, 10
applications of, 1–3 cos(ω0 n), properties of, 14–19
system, defined, 23 defined, 4, 32
Digital signal processors (DSPs) frequency-domain analysis, 122
defined, 1 modeling and properties of, 8–9
design of, 41, 73 problems, 29–32
filter realizations, 303–304, 320 real exponential function, 12
hardware design, 384–385, 387 unit pulse function, 9–10
quantized filter analysis, 354, 362, 378 unit step function, 10–12
Digital spectral transformations (DSTs), Discrete-time sinusoidal signal, 132
226–230 Discrete-time system
Digital subscriber loop (DSL) systems, 384 defined, 7, 32
Digital-to-analog converter (DAC), 22, 27–28, models, 33–36
186 performance analysis, 38–39
Direct form II structure structure of, 71
finite impulse response (FIR) filters, Distortion, finite impulse response (FIR)
305–306 filters, 250
infinite impulse response (IIR) filters, Distributive convolution sum, 66
313–314, 345, 366 Dolph–Chebyshev window
Discrete Fourier transform (DFT) finite impulse response (FIR) filters, 267
characterized, 159–160 windowed FIR filters, 276
constructed from discrete-time Fourier series
(DTFS), 160–161 ECG mapping, 2
defined, 112 Echo cancellation, 2, 25, 27
finite impulse response (FIR) filters, EEC mapping, 2
290–292 Eigenvalues, 58
MATLAB computations, 172–177 Elliptic function approximation, 212
properties of, 161–170 Elliptic lowpass filters
Discrete-time Fourier series (DTFS) characterized, 234, 323–324
characterized, 156–159 realization using MATLAB, 334–346
defined, 112 Emulator boards, 388
properties of, 161–170, 177 Equiripple, generally
reconstruction from discrete Fourier analog frequency transformations, 217
transform, 160–161 approximation, 189, 202, 212
Discrete-time Fourier transform (DTFT) bandpass filter, 235
characteristics of, 122–125 design, 280
defined, 112, 114 finite impulse response (FIR) filters
frequency response, 177, 253 design using MATLAB, 285–289
frequency shifting property, 127, 130 linear phase, 280–285
linear phase finite impulse response (FIR) lowpass filter, windowed finite impulse
filters, 260–261 response (FIR) filters, 280
MATLAB computations, 147–154 Error detection, 25
properties of, 146–147 Exponent, quantized filter analysis, 362
418 INDEX

Exponential functions, discrete-time system Floating-point numbers, quantized filter

complex, 12, 14–15 analysis, 362–364, 384
real, 12 Folding
defined, 18
Fast Fourier transform (FFT) frequency, 119
computation of, 170–172, 178 Forced response, 58–61, 64, 94
filter realizations, 304 Fourier series method, modified by windows
finite impulse response (FIR) filters, 292 FIR filter design procedures, 268–273
technique, 21 Gibbs phenomenon, 263–265
FDA Tool finite design-analysis (FDA) tool, overview of, 261–263
356–358, 360, 366, 379, 389 window functions, 266–268
Fetal monitoring, 2 Fourth-order polynomial, 330
Field-programmable gate arrays (FPGAs), Frequency control, automatic, 25
383, 389 Frequency-domain analysis
Filter approximation, 19 DTFS and DFT, 154–170, 177–178
Filter design, see specific types of filters DTFT and IDTFT, 122–138, 154–170,
history of, 19–23 177–178
z-transform theory and, 94 fast Fourier transform (FFT), 170–172, 178
Filter Design Toolbox, 356 MATLAB computations, 147–154, 172–177,
Filter realizations 184
allpass filters in parallel, 320–327, 334–346 overview of, 112, 177–178
finite impulse response (FIR) filters, problems, 177–184
305–312, 327–329, 334 sampling theory, 113–122, 177
infinite impulse response (IIR) filters, unit step sequence, DTFT of, 138–147
312–320, 327, 329–334 Frequency sampling, finite impulse response
using MATLAB, 327–346 (FIR) filters, 289–292
overview of, 303–305 Frequency shifting property, discrete-time
problems, 347–353 Fourier transform (DTFT), 127, 130–131
Final value, z-transform theory, 75–76 functionname, MATLAB, 93
Finite impulse response (FIR) filters
defined, 37, 250 Gain control, automatic, 25
equiripple, design using MATLAB, Geophysical data processing, 2
285–289 Gibbs phenomenon, 263–266
equiripple linear phase, 280–285 Global positioning system (GPS), 2
Fourier series method modified by windows, Global System for Mobile Communication
261–273 (GSM), 25
frequency-domain analysis, 140, 142, 147, Graphical user interface (GUI), 356, 382
151, 153 Group delay
frequency sampling method, 289–292 analog filters, 188–189
lattice structures, 309–310 linear phase finite impulse response (FIR)
linear phase, 251–261, 311–312 filters, 250, 253, 255–256, 258, 269
overview of, 249–251, 292–294 response, frequency-domain analysis, 153
problems, 294–301
quantized filter analysis, 375–379 Hann window, finite impulse response (FIR)
realizations filters, 266, 268–269
cascade form, 306–307 Hamming window, finite impulse response
direct form, 305–306 (FIR) filters
lattice structures, 309–310 equiripple design, 287
polyphase form, 307–309 implications of, 266, 268
windowed design using MATLAB, linear phase filters, 270–271
273–280 windowed filters, 276
FIR filters, see Finite impulse response (FIR) Hardware
filters digital filter, 304
First-order polynomials, 315, 327 using DSP chips
 INDEX 419

Code Composer Studio, 386–388 Internet telephony, 1

code generation, 385 Inverse Chebyshev filters, 208
design preliminaries, 383–385 Inverse discrete Fourier transform (IDFT)
emulator, 388–389 characterized, 159–160,170–171
simulator, 388–389 finite impulse response (FIR) filters, 290–291
Simulink, 381–383 MATLAB computations, 172–177
Harmonics, discrete-time Fourier series, Inverse discrete-time Fourier series (IDTFS),
156 characterized, 157–159, 165, 169, 177
Hearing aids, digital, 2, 354 Inverse discrete-time Fourier transform
Hertz per second, 123, 148 (IDTFT)
Highpass (HP) filters, linear phase FIR, differentiation property, 141–142
261–262 multiplication property, 142–143
Highpass filter symmetry property, 146
analog, 187 time-domain analysis of noncausal inputs,
digital, 212–213 126
finite impulse response (FIR), 293–294 time reversal property, 129, 131–132, 136,
frequency-domain analysis, 122, 124 141
linear phase finite impulse response (FIR), Inverse Fourier transform, 117
261–262 Inverse z transform
realization using MATLAB, 339–340 difference equations, 51, 54, 56, 62
Hilbert transformer, 285 frequency-domain analysis, 163
Home location register (HLR), 26 models, 72, 75
z transform theory, 41, 49, 92–93
Ideal bandpass filter, frequency-domain Iterative optimization, 24, 250
analysis, 122–125
IEEE 754–1985 standard, 362–363 Jump discontinuity, linear phase FIR filters,
IIR filters, see Infinite impulse response (IIR) 257–258, 289
filters Jury–Marden test, 78–81
Image processing, 2
Impulse-invariant transformation, 219–221 Kaiser window, finite impulse response (FIR)
Impulse response, 78 filters
Impulse sampling, 38, 41 characteristics of, 267–269
Industrial applications, 3 equiripple design, 287
Infinite impulse response (IIR) filters equiripple linear phase, 281n
allpass filters, 230–231 windowed, 279
bilinear transformation, 221–226
characteristics of, 37, 186–189 Laplace transform, 42
design using MATLAB, 231–238, 240 Lattice-coupled allpass filter, in filter
digital spectral transformation, 226–230 realization
frequency-domain analysis, 148, 150–151, characteristics of, 322, 346
155 power complementary filter, 322
impulse-invariant transformation, 219–221 quantized filter analysis, 370, 372–375
problems, 240–247 structures, 320
quantized analysis, 367–375 Lattice-ladder realization, 326, 332–333
realizations, 304, 312–320 Lattice-ladder structure, filter realization,
Yule–Walker approximation, 238–239 344–345
Initial states, time-domain analysis, 50, 53, 57, Lattice structure, filter realization
63 allpass, 320
Initial value, z-transform theory, 74–75 finite impulse response (FIR) filters,
Input sample response, 67 309–310, 332–334
Input-output relationship, time-domain analysis infinite impulse response (IIR) filters,
implications of, 33, 35, 50, 58, 73, 94 332–334
z-transform relationship, 69–71 LC (inductance × capacitance) filters,
Interleaving, 25 19–20, 28
420 INDEX

Least mean-squares, 263 Mathematical functions, MATLAB, 401

Least significant bit (LSB), 361 MathWorks, 355
Least-squares approximation, 238 MATLAB
Linear, time-invariant, discrete-time system allpass filters in parallel determination,
(LTIDT), defined, 249 334–346
Linear and time-invariant (LTI) systems, 33 arrays, 392–393
Linear convolution sum, 66 control flow, 402–403
Linearity of system, 50 defined, 24
Linear phase finite impulse response (FIR) discrete Fourier transform and inverse
filters discrete Fourier transform computation,
design procedures, 268–273 172–177
overview, 251–256 discrete-time Fourier transform computation,
properties of, 256–261 147–154
realizations, 311–312 drawing plots, 400
Linear system, defined, 32 edit window, 403
Local area network (LAN), 384 equiripple finite impulse response (FIR) filter
Low-pass filters design using, 285–289
analog, 191–192 FDA Tool, 355, 357, 378–379
characteristics of, 27, 117 filter realization
elliptic infinite impulse response (IIR), applications, 327–346
367–375 problems, 351–353
equiripple finite impulse response (FIR) finite impulse response (FIR) filter
filter design, 285–286, 288 realizations, 327–329, 331–332
quantized analysis, 375–376 frequency-domain problems, 184
finite impulse response (FIR), 293–294 functions, 400–401
frequency-domain analysis, 117, 122–124, hardware design, 381–389
141–142, 176 infinite impulse response (IIR) filter
linear phase finite impulse response (FIR), design, 231–238, 240
261, 264, 270, 272 realizations, 327, 329–332
windowed finite impulse response (FIR), matrices, 392
275, 277 matrix operations, 393–398
M-window, 403–405
Magnetic resonance imaging (MRI), 2 numerical format, 401
Magnitude response problems using, 108–110, 184, 247,
allpass filters in parallel, 339–340, 345, 299–301
372–374 scalar operations, 398–399
discrete Fourier transform (DFT), 175 Signal Processing Toolbox, 405–414
discrete-time Fourier transform, 136–137 vectors, 392
elliptic lowpass filter, 339, 346 windowed finite impulse response (FIR) filter
equiripple finite impulse response (FIR) filter design using, 273–280
design, 287–288 z-transform theory, 81–93
finite impulse response (FIR) filters, 270 Matrix algebra, 57–58, 71
frequency-domain analysis, 126, 150, 153 Maximally flat magnitude response, 191–192
infinite impulse response (IIR) elliptic Megahertz, 28
lowpass filter, 368–372 Military electronics, 2–3
linear phase finite impulse response (FIR) Minimax approximation, 202
filters, 256–257, 262–263 Minimax design, 280
lowpass elliptic filter, 339, 346 Mirror image polynomial, 259
lowpass equiripple finite impulse response Mobile phone(s), see Cell phones
(FIR) filter, 376–377 digital signal processing theory, 25
windowed finite impulse response (FIR) network, 2, 25–28
filter, 277–278 Mobile switching center (MSC), 26
Magnitude spectrum, 122, 124–125 Modulation, 25
Mantissa, quantized filter analysis, 362 Monte Carlo analysis, 24
 INDEX 421

Moving average (MA), filter realization, 326, Preconditioning filter, 22

332, 334, 359 Programmable filters, 24
Multidimensional filters, 24 Public switched telephone network (PSTN), 26
Multipath equalization, 25, 27
Multiple poles, 55–56 Quantization, defined, 6
Multiplication property, discrete-time Fourier Quantized filter analysis
transform, 142–145 binary numbers and arithmetic, 360–367
Multiplier, 33, 68 filter design-analysis tool, 355–360
Multirate filters, 24 finite impulse response (FIR) filters, 375–379
infinite impulse response (IIR) filters,
Natural response, 58–59, 61, 63–64, 94 367–375
Noise cancellation, 25 problems, 379
Nonrecursive filter, see Finite impulse response software, 354–355
(FIR) filters
Normalized digital frequency, 251, 257 Radar processing, 2
Nyquist frequency, 120, 122–123, 257, 276 Radians per second, 123, 188, 190
Random access memory (RAM), 388
One-complementary form, quantized filter Read-only memory (ROM), 388
analysis, 362 Real-time data exchange (RTDX), 389
Operational amplifiers, 19, 23 Real-Time Workshop
Oscillation, discrete-time system, 16–17 Embedded Coder, 385
Output noise, finite impulse response (FIR) Simulink, 381–389
filters, 250 Reconstruction formula, 118
Overflow mode, 366 Rectangular pulse function, discrete-time
Overlap-add method, 68n, 172 Fourier transform, 140
Overlap-save method, 68n, 172 Rectangular window, finite impulse response
Overloaded functions, 358 (FIR) filters, 264–265, 268
Recursive algorithm, 36–38, 82, 85, 92, 94,
Parallel forms, infinite impulse response (IIR) 304, 325
filter realization, 317–320 Reflection coefficients, 327, 332
Parks-McClellan algorithm, 284 Region of convergence (ROC), 43–44
Passband filter, 192 remez exchange algorithm, 284, 288–289
Passband filter, finite impulse response (FIR): Remote sensing, 2
characteristics of, 293–294 r n , z-transform theory, 76–77
equiripple design, 285 Rounding, in quantized filter analysis,
linear, 261 364–365
windowed, 275 Routh–Hurwitz test, 79
Patient monitoring, 2 Runtime support (RTS), 388
Personal digital assistants (PDAs), 354
Phase angle, linear phase finite impulse Sampled-data signals, 4–5
response (FIR) filters, 253–255 Sampling frequency, 27–28
Phase response Sampling period, 4
frequency-domain analysis, 136–137, 153 Sampling theory, frequency-domain analysis
finite impulse response (FIR) filter, 376–377 bandpass signals, 120–121
infinite impulse response (IIR) filter, continuous-time function, 113, 118
368–369 discrete-time Fourier transform (DTFT),
lattice-coupled allpass filter, 374 114–116
Picket fence effect, 172–173 inverse Fourier transform, 117
Polyphase form, finite impulse response (FIR) Shannon’s sampling theorem, 118–120
filter realization, 307–311 unit impulse response, 113–114, 117–118
Positron emission tomography (PET) Scalar, defined, 392
scanning, 2 Second-order polynomials, 315, 327
Power control, automatic, 25 Seismic data processing, 2
Power series, 67 Shannon’s reconstruction formula, 160–161
422 INDEX

Shannon’s sampling theorem, 21, 118–120, Three-dimensional (3D) images, 2

177 Time-division multiple access (TDMA)
Shift-invariant system, 32, 51, 53 technology, 2, 25
Signal Processing (SP) Toolbox, 356, Time-domain analysis
358–359, 364, 369, 389, 405–414 convolution, 65–70
Sign bit, quantized filter analysis, 361–362 defined, 38
Signed magnitude fixed-point binary number difference equations, 52–64
representation, 6 linear, time-invariant system, 32–41
Simulink, 24, 405 MATLAB functions, 81–93,
Smoothing filter, 22 108–110
Soft decision decoding, 25 models, 70–77
Sonar processing, 2 problems, 94–110
Speaker verification, 2 stability, 77–81
Spectral components, 132, 159 Time-invariant system
Spectrum analyzers, 24 convolution sum, 38–41
Speech compression, real-time, 25, 27 discrete-time system models, 33–36
Speech enhancement/speech processing/speech overview of, 32–33
recognition/speech synthesis, 2 recursive algorithm, 36–38
Speech-to-text dictation, 2 z-transform theory, 41–59, 64–65
Spline function, linear phase finite impulse Time reversal
response (FIR) filters 272–273 property, discrete-time Fourier transform
Spread spectrum, 2 (DTFT), 128–129
Stability z-transform theory, 73
bounded-input bounded output (BIBO), Time-shifting property, discrete-time Fourier
77–78 transform (DTFT), 127–128, 131,
Jury–Marden test, 78–81 139, 141
z-transform theory, 77–81 Transfer function, 251, 258, 303
Steady-state response, 63–64 Transient response, 63–64
Stopband filter Transition bands, 24
equiripple first impulse response (FIR) filter Transversal filter, see Finite impulse response
design, 285 (FIR) filters
linear phase finite impulse response (FIR), Triangular window, finite impulse response
263 (FIR) filters
realization using MATLAB, 337 implications of, 266n
Stopband frequency, finite impulse response windowed filters, 276
(FIR) filters, 293–294 Trigonometric functions, MATLAB,
Storage, digital filters, 24 401
Switched-capacitor filters, 4, 28 Truncation, quantized filter analysis,
Symmetric coefficients, linear phase finite 365
impulse response (FIR) filters, 252–254, Two-complementary form, quantized filter
256, 258–259 analysis, 362
Symmetry property, discrete-time Fourier Two-dimensional (2D) images, 2
transform, 145–147 Type I finite impulse response (FIR) filters
Synthesizers, 24 characteristics of, 252–253, 257–258, 260,
267, 269, 281, 283, 290
Tapped delay filter, see Finite impulse response filter realizations, 310–311
(FIR) filters Type II finite impulse response (FIR) filters
Telecommunications applications characteristics of, 253–254, 257, 260, 276,
mobile phone network, 2, 25–28 281, 283, 290
overview of, 1–2 filter realizations, 310–311
Texas Instruments, 383–384, 386–388 Type III finite impulse response (FIR) filters,
Text-to-speech translation, 2 254, 257, 260, 281–283, 291
Third-generation (G3) mobile phones, 2 Type IV finite impulse response (FIR) filters,
Third-order lowpass filter, 20, 22 255–257, 260, 281, 283, 291
 INDEX 423

Ultrasound imaging, 2 Weighting function, equiripple linear phase

Unit impulse function, discrete-time system, 9 finite impulse response (FIR) filter,
Unit impulse response 283–284
discrete-time Fourier transform, 143 Windowed finite impulse response (FIR) filters,
infinite impulse response (IIR) filters, 220 design using MATLAB
linear phase finite impulse response (FIR) filter order estimation, 273–275
filters, 252 FIR filter design, 275–280
time-domain analysis, 37–38, 72–73, 89–90 Windows functions, finite impulse response
Unit pulse function, discrete-time system, (FIR) filters, 266–268
9–10 Wordlength, in quantized filter analysis, 354,
Unit sample function, discrete-time system, 9 361, 366, 370, 375
Unit sample response, time-domain analysis, Workspace, defined, 392
37, 67 Worst-case analysis, 24
Unit step function, discrete-time system,
10–12 X-rays, 2
Unit step sequence, discrete-time Fourier XDS510 JTAG Emulator, 388–389
transform xPC Target, 385
conjugation property, 145
differentiation property, 139–142 Yule–Walker approximation, 238–239
multiplication property, 142–145
overview of, 138–139
symmetry property, 145–147 Zero input response, 49–50, 54, 58–59, 61,
Unsigned fixed-point binary number, 361 85, 94
Zero-order (ZOH) circuit, 4, 6–7
Zero state response, 49–50, 54, 58–59, 61, 63,
Vectors 87, 94
finite impulse response (FIR) filter: z transform
equiripple design, 285 defined, 19
realizations, 328, 331, 333 frequency-domain analysis, 163–164
infinite impulse response (IIR) filter theory
design, 239 applications, generally, 56–58
realizations, 328, 331, 333 characterized, 41–49, 82, 93–94, 126
Very large-scale integration (VLSI) technology, convolution, 65–70
20, 383 linearity of system, 50
VHDL, 383 methodology, 64–65
Video compression, 25 models, 70–77
Videoconferencing, 3 problems, 94–110
Visitor location register (VLR), 26 properties of z transforms, 77
Visual DSP++, 389 solution using MATLAB functions,
Vocoders, 24 81–93
Voice over Internet protocol (VoIP), 1–2 solving difference equations, 51–64
Voice recognition, 25 stability, 77–81
time-invariant system, 50–51
Waveform coding, 25 zero input and zero state response,
Weather monitoring, 2 49–50, 94