INTRODUCTION TO

DIGITAL SIGNAL PROCESSING

"This page is Intentionally Left Blank"

INTRODUCTION
TO DIGITAL
SIGNAL PROCESSING

RomanKuc
Department of Electrical Engineering
Yate University

SSP BS Publications
4-4-309, Giriraj Lane, Sultan Bazar,
Hyderabad - 500 095 - AP.
Phone: 040-23445677,23445688
e-mail: ci.: . [email protected]
www.bspublications.net

Reprint 2008
All rights reserved.

No part of this book or parts thereof may be

reproduced, stored in a retrieval system or
transmitted in any language or by any means, ,
electronic, mechanical, photocopying, recording ,
or otherwise without the prior written
permission of the publishers.
I

Copyright 1982, by Author

"This Reprint of Introduction to Digital Signal
Processing, originally published by McGraw Hill, Inc.
1982, is published by agreement with the author".

Published by :

SSP BS
=
__

Publications

4-4-309, Giriraj Lane, Sultan Bazar,

Hyderabad - 500 095 - A. P.
Phone: 040-23445677,23445688
Fax No: 040-23445611

e-mail: [email protected]
www.bspublications.net

Printed at
Adithya Art Printers

Ilyderabad

ISBN

81-7800-168-3

PREFACE

This book is intended to be used in the first course covering digital signal
processing and filter design, typically offered at the senior or first-year
graduate level in electrical engineering. The course is also appropriate for a
graduate course in departments other than electrical engineering, such as
geophysics and mechanical engineering, in which the analysis of discrete"time
data is performed. It is assumed that the student has had a course covering
Fourier series and LaPlace transforms on the level of the first linear circuits or
control systems course. This text also includes projects that require students to
write computer programs to accomplish signal processing projects. The student
should bO'familiar with some programming language 'such as FORTRAN,
BASIC, PASCAL, or C.
This book approaches digital signal processing and filter design in a novel
way, by presenting the relevant theory and then having the student apply it by
implementing signal processing routines on a computer. This mixture of theory
and application has worked successfully for the past six years' in teaching this
course at Yale University. With this approach, the students receive a deeper
and intuitive understanding of the theory, its applications and its limitations.
The text can accommodate a wide variety of courses. Currently, the
course on digital filters at the undergraduate level are taught primarily as a
theory course, with homework problems and exams to determine the course
grade. This book can be used directly in this type of course. However, the
course becomes much more interesting and enjoyable when students apply the
theory to write programs to perform signal processing tasks. Each chapter
describes programming techniques that implement the theory and includes
projects that illustrate different applicatio'!l:>. Digital filter routines are written
from the very beginning of the course, immediately after the time-domain
description of discrete-time systems is presented in Chapter 2. The programs
can be written on any computer from personal to mainframe, requiring only a
standard terminal for displaying the results in graphical form and a printer for
v

vi

PREFACE

generating the hard copy. The programs to be written by the student are not
difficult. Each routine performs only one primitive task in the signal processing
procedure. For example, the discrete Fourier transform subprogram can be
accomplished in only nine lines of code. But once the student has given
sufficient thought to the problem to write these nine lines, the magic associated
with this procedure disappears, being replaced by understanding. To minimize
the programming effort of the student, the routines are used as building blocks
in later projects. Constructing more sophisticated programs from these simpler
building blocks aids the student in conceptualizing projects later in the course.
By the end of the course, the student has a package of signal processing
programs that can be used for a variety of applications.
The graphic outputs for the projects are displayed on the terminal and on
a printer, being generated by programs provided in the Appendix. These plots,
although having crude resolution, are adequate for illustrating the ideas and
also have the advantage of not requiring special plotting hardware. Students
having higher-resolution terminal graphic capabilities have readily written their
own graphic routines using the programs in the Appendix as guides.
This text differs from others in the area of digital filters and signal
processing in the important respect that processing signals and designing filters
on a digital computer, using simple programs composed by the students, are
an integral part of the course. For this reason, more realistic problems can be
assigned and discussed than would be in a course without computers, for whichthese problems would be mathematically untenable or at least tedious. The
organization of the material presented in the text allows the student to start
writing meaningful filtering programs from the beginning of the course.
Projects included in the text illustrate both digital signal processing and digital
filter design concepts.
The book is organized in the following manner. Chapter 1 defines the
basic components of a digital filter and provides motivation by describing
several common applications of digital filters and signal processing. The
material in Chapters 2 through 5 covers the fundamentals of time, frequency,
and z-plane analysis. If some of this material has been introduced in previous
courses, the respective chapters can be treated as a review. Several common
structures that are employed to implement digital filters are described in
Chapter 6.
The synthesis of digital filters to meet a desired specification is presented
in Chapters 7 through 9. To compare the different filter design procedures, the
projects are structured to allow each student to be assigned a personal
specification to be satisfied. In Chapter 7, the digital filter design is accomplished by interactively and intuitively placing poles and zeros in the z plane.
This procedure serves three purposes. First, it develops an understanding of
the interaction of the singularities in generating a filter magnitude response or
system spectrum. Second, it provides an appreciation for the conventional
filter design procedures presented in Chapters 8 and 9. Finally, this intuitive
procedure indicates how to correct the locations of the poles and zeros

PREFACE

vii

generated by conventional design procedures. The computed positions of these

singularities are slightlY, but significantly, displaced from the desired positions
by the round-off errors that are always present in any computation. These
round-off effects become obvious in the filter design projects. The conventional
procedures for designing infinite-impulse response (IIR) digital filters are
described in Chapter 8, and those for finite-impulse response (FIR) digital
filters in Chapter 9.
Two advanced topics are covered in Chapters 10 and 11. Chapter 10
describes the effects of finite precision encountered in practical hardware
implementation of digital filters. Some elementary results from probability
theory are presented to describe noise effects. Chapter 11 provides an
overview of inverse filters. This topic was chosen because it includes and
integrates many of the concepts introduced in the previous chapters to
illustrate an important practical application.
In addition to the Problems, a set of Computer Projects is included, the
programs for which should be written by the student. Instructions for program
development are included as an integral part of the text and Project
description. The ability to manipulate sequences by a computer program is
essential for appreciating the concepts involved for signal processing, which is
important when the studen~ IS faced with novel problems in his or her own
application.
Projects in the analysis part of the book illustrate the practical applications of digital interpolation procedures and the simulation of a frequency-shift
keying (FSK) digital communication channel. In the filter synthesis part of the
book, the Projects include designing digital filters to satisfy a practical
specification. In the chapters describing finite-precision effects and the inverse
filter, the Projects illustrate programming techniques for simulating physical
hardware and systems.
One main benefit of the Projects is to relate the analytic results to those
produced by the computer. At times, the analysis is simple and should then be
used to predict and verify the computer results. At other times, the analytic
procedure is very complicated mathematically, in which case the computer
results should be obtained first to indicate what type of results should be
expected. In this latter case, the computer results indicate the direction in
which the analysis should proceed. For example, the computer results may
indicate that a complex-valued operation in a particular analytic step results in
a real-valued sequence. This knowledge may serve to drive the analytic steps.
In practice, it is often found that computer simulations demonstrate an effect
that can be proven analytically.
An insider's knowledge of the signal processing algorithms is attained by
having the students write their own programs. This knowledge then allow~ the
student comfortably to make the intellectual move to more sophisticated signal
processing and filter design programs and packages, such as those distributed
by the IEEE or SIG or offered by many vendors.
This text has the flexibility of being used in a wide variety of digital signal

viii

PREFACE

processing courses at the undergraduate and graduate levels. For a conventional undergraduate theory course in digital filters, the first nine chapters of
this book supply the necessary topics in discrete-time signals and systems. In
some universities, such a theory course is followed by a digital filter design
laboratory. The Projects in this book can be used as the exercises in such a
course. At the graduate level, a faster pace allows topics from all chapters to
be covered. Courses stressing the signal processing aspects should emphasize
the analytic methods and the discrete Fourier transform and de-emphasize IIR
filter design and quantization effects, while courses stressing filter implementation should stress the latter and reduce the emphasis on the former.
To allow adequate space for this treatment of this material, the analytic
techniques have been restricted to those that are absolutely necessary for
understanding the filter synthesis and signal processing procedures. Topics that
are normally covered in a senior/first-year graduate course on digital signal
processing, but that are not considered in this text, include the solution of the
difference equations with the one-sided z-transform, the numerical methods
for optimizing digital filter designs, and the area of multidimensional signal
processing. These topics are better treated in a second course in which the
students have the familiarity of the material contained in this book.
The author is indebted to the students at Yale University who have taken
this course and provided many useful comments regarding the topics and the
exercises in this book. Among these are Ata Arjomand, Billur Barshan, Leo
Costello. An;lOld Englander, Kannan Parthasarathy, Rob Roy, and David
Starer. The author would like to thank the following reviewers for their
comments and-suggestions, J. I. Aunon, Purdue University; Edward Delp,
Purdue University; Simon Haykin; McMaster University; M. Kaveh, University of Minnesota; and Paul Prucnal, Columbia University.
The author hopes that the instructors teaching digital signal processing as
a combination of theory and computer exercises, as described in this text, have
an enjoyable learning experience. This learning comes about from the
questions posed by the students, who inevitably find some novel approach or
interpretation to filter design on the computer. This has certainly been the case
every time the author has taught this course.
Roman Kuc

CONTENTS

Preface
1 Introduction
1-1
1-2
1-3
1-4
1-5
1-6
1-7
1-8

What is a Digital Filter?

Anatomy of a Digital Filter
Frequency Domain Description of Signals and Systems
Some Typical Applications of Digital Filters
Replacing Analog Filters with Digital Filters
Overview of the Course
Description of the Computer Projects
Summary

v
1
1
3
5
7
12
13
17
18

2 Discrete-time Description of Signals and

Systems
2-1
2-2
2-3
2-4
2-5
2-6
2-7
2-8
2-9
2-10
2-11

Introduction
Discrete-Time Sequences
Superposition Principle for Linear Systems
Unit-sample Response Sequence
Time-Invariant Systems
Stability Criterion for Discrete-time Systems
Causality Criterion for Discrete-time systems
Linear Constant-Coefficient Difference Equations
Suggested Programming Style
Writing a Digital Filter Program
Summary

3 Fourier Transform of Discrete-time Signals

3-1
3-2
3-3

Introduction
Definition of the Fourier Transform
Important Properties of the Fourier Transform

20
20
20
26
27
30
38
40

41
44

45
55
65
65
65
69
ix

CONTENTS

3-4
3-5
3-6
3-7
3-8
3-9
3-10
3-11
3-12

4
4-1
4-2
4-3
4-4
4-5
4-6
4-7
4-8
4-9
4-10
4-11
4-12
4-13
4-14

Properties of the Fourier Transform for Real-valued Sequences

Program to Evaluate the Fourier Transform by Computer
Use of the Fourier Transform in Signal Processing
Fourier Transform of Special Sequences
The Inverse Fourier Transform
Fourier Transform of the Product of Two Discrete-time
Sequences
Sampling a Continuous Function to Generate a Sequence
Reconstruction of Continuous-time Signals from Discrete-time
Sequences
Summary

The Discrete Fourier Transform

Introduction
The Definition of the Discrete Fourier Transform (DFT)
Computing the Discrete Fourier Transform from the Discretetime Sequence
Properties of the DFT
Circular Convolution
Performing a Linear Convolution with the DFT
Computations for Evaluating the DFT
Programming the Discrete Fourier Transform
Increasing the Computational Speed of the DFT
Intuitive Explanation for the Decimation-in-time FFT
Algorithm
Analytic Derivation of the Decimation-in-time FFT Algorithm
Some General Observations about the FFT
Other Fast Realizations of the DFT
Summary

5 The z-transform
5-1
5-2
5-3
5-4
5-5
5-6
5-7
5-8
5-9
5-10
5-11
5-12
5-13
5-14

Introduction
The Definition of the z-transform
Properties of the z-transform
The System Function of a Digital Filter
Combining Filter Sections to Form More Complex Filters
Digital Filter Implementation from the System Function
The Complex z-plane
The Region of Convergence in the z-plane
Determining the Filter Coefficients from the Singularity
Locations
Geometric Evaluation of the z-transform in the z-plane
. Relationship Between the Fourier Transform and the
z-transform
The z-transform of Symmetric Sequences
The Inverse z-transform
Summary

72
80

82
84
92
94
96
105
110
118
118
119
121
128
130
132
133
135
136
142
146
148
150
150
157
157
158
160
162
164
167
169
i74
178
183
184
188
196
203

CONTENTS

6 Digital Filter Structures

6-1
6-2
6-3
6-4
6-5
6-6
6-7
6-8
6-9

Introduction
System Describing Equations
Filter Catagories
The Direct Form I and II Structures
Cascade Combination of Second-order Sections
Parallel Combination of Second-order Sections
Linear-phase FIR Filter Structures
Frequency-sampling Structure for the FIR Filter
Summary

7 From Analysis to Synthesis

7-1
7-2
7-3
7-4
7-5
7-6
7-7
7-8
7-9
7-10
7-11

Introduction
Interrelationships of Analytic Methods
Practical Magnitude Response Specifications
Log-magnitude Response Curves
Programs to Compute the Log-magnitude Response of Poles
and Zeros
Phase Response Considerations
Steps in Performing a Filter Design
Choice of Filter Type
Interactive Filter Design by Intuitive Pole/Zero
Placement
Writing an Interactive Design Program
Summary

8 Infinite Impulse Response Filter Design

Techniques
8-1
8-2
8-3
8-4
8-5
8-6
8-7

Introduction
Analog Filter System Function and Frequency Response
Analog Lowpass Filter Design Techniques
Methods to Convert Analog Filters into Digital Filters
Frequency Transformations for Converting Lowpass Filters
into other Types
All-pass Filters for Phase Response Compensation
Summary

9 Finite Impulse Response Filter Design

Techniques
9-1
9-2
9-3
9-4
9-5

Introduction
Designing FIR Filters with the Windowing Method
The DFT Method for Approximating the Desired Unit-Sample
Response
Designing FIR Filters by Combining the DFT and Window
Methods
Designing FIR Filters with the Frequency-sampling Method

xi

208
208
209
209
210
214
218
220
221
227
230
230
231
236
238
239
243
246
246
247
256
258

262
262
263
263
280
295
310
312

318
318
319
341
348

361

xii

CONTENTS

9-6
9-7

Use of the FFT Algorithm to Perform FIR Filtering

Summary

10 Finite-precision Effects
10-1
10-2
10-3
10-4
10-5
10-6
10-7
10-8

Introduction
Finite-precision Number Representation within the Computer
Quantization Error in Analog-to-digital Conversion
Expressing Coefficients in Finite Precision
Performing Arithmetic in Finite-Precision Number Systems
Effects of Finite-Precision Arithmetic on Digital Filters
Programs to Simulate Quantization Effects
Summary

11 Inverse Filtering
11-1
11-2
11-3
11-4
11-5
11-6
11-7
11-8
11-9

Introduction
Applications of Inverse Filters
Minimum Phase Systems
Formulating the Problem for Applying an Inverse Filter
FIR Filter Approximations to the Inverse Filter
Discrete Hilbert Transform Relationship for Minimum-Phase
Systems
Designing Inverse Filters with the Discrete Hilbert Transform
The Effects of Noise on Inverse Filters
Summary

379
384
390
390
391
395
399
404
409

411
416
422
422
423
426
429
432
433
436
441
448

Appendix A

455

Appendix B

469

Index

471

CHAPTER

1
INTRODUCTION

In this chapter, we present the concept of a digital filter and define the adder,
multiplier and delay elements that are its components. Some typical applications of digital filters for signal processing, system simulation and spectral
analysis are described. An overview of the topics covered in the book
concludes the chapter.

1.1 WHAT IS A DIGITAL FILTER?

Most broadly defined, a digital filter is a numerical procedure, or algorithm,
that transforms a given sequence of numbers into a second sequence that has
some more desirable properties, such as less noise or distortion. As shown in
Fig. 1.1, the entire given, or input, sequence will be denoted in this book by
{x(n}}, and the entire output sequence by (y(n)}, where n is an index.
Typically, the set of index values consists of consecutive integers, which in
some cases may take values from minus infinity to plus infinity.
The desired features in the output sequence depend on the application.
For example, if the input sequence is generated by a sensing device, such as a
microphone, the digital filter may attempt to produce an output sequence
having less background noise or/interference. In radar applications, digital
filters are used to improve the5detection of airplanes. In speech processing,

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

1""____
. .I----o...,uIPul{y(n))sequence
DIGITAL

Input seQueo-nce
__
_
{*)} 0

FILTER

FIGURE I-I
A digital filter transforms the input
sequence {x(n)} into the output
sequence (y(n)}.

digital filters have been employed to reduce the redundancy in the speech
signal so as to allow more efficient transmission, and for speech recognition.
Input sequences can be generated in several ways, as shown in Fig. 1.2.
One common method is to sample a continuous-time signal at a set of
equally-spaced time intervals. If the continuous-time signal is denoted by
x(t), then the values of the discrete-time sequence are denoted as
x(nT.) = x(t)lt=nT,

(1.1)

where T. is the sampling period. In image processing applications, scanning a

line of an image produces an intensity function of position across the image
plane, denoted by i(p). This function can be sampled to produce a discreteposition sequence with values given by
i(nD.) = i(P )lp=nD.

(1.2)

where D. is the sampling distance interval.

One common application of digital filters is to simulate a physical system.
The input sequence is then usually a sequence of numbers generated by the
computer and represents the excitation signal {e(n)}. In this case, {e(n)}
represents instantaneous values of some physical parameter, such as VOltage,
torque, temperature, etc. This sequence is then processed to determine the

CONTINUOUS-TIME SIGNAL

DISCRETE-TIME SEQUENCE
{x(nT.)}

fIII,

IMAGE

III I JlI ,.

DISCRETE-TIME SEQUENCE
{i(nD.)}

-+#-iIR-IllII--IJII--t-Scan line

....IUI.....rr.......

FIGURE 1-%
Deriving one-dimensional discrete-time sequences from analog signals.

nD.

nT.

INTRODUCTION

response of the system, represented by a sequence {r(n)}. This application is

discussed in more detail later in this chapter.
To allow these different sequence types to be treated with the same
notation, we will generalize the element index to be an integer, denoted by n.
That is, we will write x(n), rather than x(nT.), and i(n), rather that i(nDs).
In addition to sampling, the input sequence could have come in a
discrete-time form initially, such as daily sales receipts, {m(n)}, or weekly
measurements of personal weight, {w(n)}. In these cases, we may be
interested in short-term fluctuations or long-term trends.
The above data sequences are indexed in terms of temporal or spatial
increments. To simplify the discussion of the various types of input data
sequences that can exist, we will refer to all of them as discrete-time sequences.
This is done with the understanding that the particular sequence under
discussion may, in practice, be related to a physical parameter other than time.
Although the theory presented in this text is general, the implementation
of a digital filter depends on the application and the user's environment. In
education and research, a digital filter is typically implemented as a program
on a general-purpose computer. The types of computers that can be used vary
widely, from personal computers, through the larger minicomputers, to the
large time-shared mainframes. In commercial instrumentation and industrial
applications, the digital filter program is commonly implemented with a
mierocomputer that may also be used for control and monitoring purposes as
well. For high-speed or large-volume applications, such as use in automobiles
for controlling the engine operation, the digital filter may consist of specialpurpose integrated circuit chips. These circuits perform the computation and
storage functions required for the digital filter operation. These functions,
which are common to all digital filter applications, are described in the next
section.

1.2 ANATOMY OF A DIGITAL FILTER

A digital filter consists of the interconnection of three simple elements: adders,
multipliers and delays. These are shown in Fig. 1.3. The adder and multiplier
are conceptually simple components that are readily implemented in the
arithmetic logic unit of the computer. Delays are components that allow access
to future and past values in the sequence. Open-headed arrows indicate
direction of information flow, while the larger c1o~ed-headed arrows indicate
multipliers. This convention will be useful for drawing complicated digital
filters in the later chapters.
Delays come in two basic "flavors": positive and negative. A positive
delay, or simply delay, is implemented by a memory register that stores the
current value of a sequence for one sample interval, thus making it available
for future calculations. A positive delay is conventionally indicated by a box
denoted by Z-l. The significance of this notation will become clear when we
analyze digital filters with the z-transform. A negative delay, or advance, is

INTRODUCTION TO PlGrrAL SIGNAL PROCESSING

ADDER

---;-cp"
+ "'--_0

xl(n) ...

y(n)

=xl(n) + x2(n)

x2(n)-

MULTIPLIER

POSITIVE AND
NEGATIVE DELAYS

x(n)

K"'---__

----iV-

_0

y(n) = ax(n)

POSITIVE DELAY - "DELAY"

--.".B. .---o
8

x(n) ....

y(n) = x(n-I)

NEGATIVE DELAY - "ADVANCE"

x(n) 0

y(n) = x(n+ I)

nGURE 1-3
Elements that are interconnected to implement a digital filter.

used to look ahead to the next value in the sequence, and is indicated by a box

denoted by z. Advances are typically used for applications, such as image

processing, in which the entire data sequence to be processed is available at
the start of processing, so that the advance serves to access the next data
sample in the sequence. The availability of the advance will simplify the
analysis of digital filters. However, the advance cannot always be implemented
in some appliCations. For example, when the data sequence is obtained by
sampling a function of time, each new sample is usually processed as it is
acquired. In this case, advances are not allowed, since we cannot gain access to
future data values.
A digital filter, design involves selecting and interconnecting a finite
number of these elements and determining the multiplier coefficient values.
Examples of interconnecting these elements to form two simple digital filters
are considered in the following examples.
EumpIe 1.1. Digital filter used as a tbree-sample averager. Consider the
relationship between the values of the output sequence at time n, denoted by
y(n), and the values of the input sequence, x(n -1), x(n) and x(n + 1), given by

y(n) = ix(n

+ 1) + lx(n) + ix(n -1)

As shown in Fig. 1.4, the current output y(n) is equal to the average of the next,
current, and previous input values. The advance serves to access the next value of
the sequence, while the delay stores the previous value.

Any past or future input data values can be accessed by including the
appropriate number of delay or advance elements. Data values can be
combined, however, only through an ad~er.

INTRODUcnON

x(n)--+------..;-c

y(n) = ih [x(n+ I) + x(n)+ x(n-I)]

FIGURE 1....
Nonrecursive digital filter that acts as a three-sample averager.

The term order will be used to denote the minimum number of delays
and advances that are necessary to implement a digital filter. For example, the
3-sample averager is a second-order filter. When the filter output is a function
of only the input sequence, as in the 3-sample averager, it is called a
nonrecursive filter. When the output is also a function of the previous output
values, it is called a recursive filter. Example 1.2 illustrates a recursive filter.
Example 1.2. First-order recursive filter. Consider the relationship between the
input and output sequence values given by
y(n) = ay(n -1)

+ x(n)

where a is a constant. As shown in Fig. 1.5, the current value of the output is
equal to the sum of the input and a times the value of the previo,lls output. Since
only one delay is necessary, this is a first-order filter.

Example 1.2 shows how feedback is implemented in a digital filter. Whereas

both positive and negative delays can be applied to the input sequence, only
positive delays can be applied to the output sequence, since we cannot gain
access to output values that have not yet been computed.

1.3 FREQUENCY DOMAIN

DESCRIPI10N OF SIGNALS AND
SYSTEMS
To discuss some typical applications of digital filters in conventional terms, we
need to define the frequency domain description of signals and systems. Let us

"')If-::-(--a--$~Z_I--
'----l,\r--<<J
FIGURE 15
A first-order recursive digital filter.

y(n) =ay(n-I) + x(n)

(;

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

(x(n)} =

{COS

(wn)}

".

"

nGURE 1-6
The cosine sequence {cos(wn)} for w = :tr/l2.

. DISCRETE-TIME SEQUENCES

MAGNITIJDE SPECrRA

~--~-----:!:~w

(X2(n))

... 1'

I II I I I I

JI

1...

)n
oL _ _ _

---L~

_ _~.;. W

I
nGURE 17
Discrete-time sequences and their magnitude spectra. The sequence {x)(n)} = {x\(n)} + {x2(n)}.

7
consider the discrete-time sinusoidal sequence {x(n)} shown in Fig. 1.6, and
given by
INTRODUcrlON

{x(n)}

= {cos(wn)}

(1.3)

where w is the frequency of the sequence, having units of radians. More

interesting signals can be considered to be composed of a sum of such
sinusoidal sequences having different frequencies. One example is the Fourier
series expansion of a periodic waveform. As described in detail in Chapter 3,
the magnitude spectrum of a discrete-time sequence, denoted by IX(eIW)1
indicates the magnitude of the various sinusoidal components. Examples of
three sequences and their magnitude spectra are shown in Fig. 1.7. The
sequence {xl(n)} contains mostly low-frequency components, {x2(n)} contains
mostly high-frequency components, and {x3(n)} = {xl(n)} + {x2(n)} contains
both.
One common application of digital filters is to separate signals on the
basis of their spectral content. For example, in Fig. 1.8, we start with the
sequence {x3(n)}. To retrieve the low-frequency signal {xl(n)}, we pass
{x3(n)} through a filter whose magnitude response, denoted by IHLP(eiW)I,
passes the low-frequency components of the input, while eliminating the
high-frequency ones. Such a filter is called a lowpass filter. To retrieve the
high-frequency signal {x2(n)}, we pass {x3(n)} through a highpass filter, whose
response is shown as IHHP(eiw)l. Two other important filters are the bandpass
filter, which passes the input components that lie only within a desired band of
frequencies, and the bandreject filter, which eliminates the input components
that lie within a specified band of frequencies.

1.4 SOME TYPICAL APPLICATIONS OF

DIGITAL FILTERS
Studying the field of digital filtering provides a fundamental understanding of
processing of discrete-time data. This is essential for manipulating data by
digital computer and for applying the signal processing procedures to such
applications as robotics, intelligent sensors and seismic signal processing.
Digital filter teChniques are also important for modeling linear systems, such as
speech production models and transmission channels. Often, such modeling
provides insights into the underlying principles of the problem, from which
creative solutions can stem. We now consider some typical applications.
SPEECH PROCESSING. Digital filters are commonly used for speech recognition systems. Let us consider the sequences {xl(n)} and {x2(n)}, shown in Fig.

1.9, which represent samples taken from two speech signals. The first is the lal
sound, as in "father", and the second is the IiI sound, as in "see". These two
periodic signals are most easily differentiated by computing the magnitude
spectrum of one period. We note that the two spectra, IX1(eiw)1 and IX2(e IW)I,
consist of peaks that correspond to the resonances of the vocal tract shape that

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

{X3(n) }

.,1"1,, III

...

1...

IHuo(ei"') I

{X3(n)}~---J

LOWPASS

L{yu.(n)}

I . . ._FI_I:_:rE_R_-,1 =

{x.(n)}

'I

{X3(n)}~ ~IGHPASS L{y~n)}

FILTER

r--

{x~n)}

I..

F:GURE 1-8
Recovering sequences by filtering. The sequence {Xt(n)} is recovered from {x3(n)} by lowpass
filtering. and {x2(n)} is recovered by bighpass filtering.

was used to generate the sound. In speech recognition systems, various speeCh
sounds can be differentiated on the basis of the frequency locations of these
peaks.
COMPUTER SIMULATION OF A PHYSICAL SYSTEM. When the physical
system, such as a robot, is very complicated or expensive to build, often it is
first simulated on a digital computer in its digital filter form. This simulation is
employed to determine the performance and the sensitivities of the analog

INTRODUcnON

(X.(II)}

-~~O------------------------------~nro

Pvr\A,,,.

(X,(II)}

~~O------------------------------~nro

1'

FIGURE
Speech signals and their spectra. The sequence {x1(n)} consists of 512 samples of the speech sound
lal and {x2(n)} of IiI. The sample values are connected by straight lines. The corresponding
spectra IX1(ejO1 and IX2(ei"')I, given in logarithmic units of dB, exhibit peaks that can be used to
differentiate these sounds.

10

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

system before it is actually built. The frequency-dependent transformations,

such as those produced by the inertia and compliance of a robot arm, can be
mimicked by a digital filter.
The advantages of performing a digital filter simulation include the
following. First, researchers can obtain data without performing expensive and
time-consuming experiments. Second, the simulation provides a flexible means
of varying the input and system parameters and observing the effect of these
changes. A third advantage of simulations is that any desired signal in the
model can be observed. In the physical system, it may be very costly or even
impossible to observe the value of certain parameters that are important in the
operation of the system. The following cases illustrate these advantages.
Simulating a computer communication system. One common digital data
communication system that allows computers to transfer data uses the
frequency-shift keying (FSK) scheme. In the FSK system, the binary digits, 1
or 0, are transformed into sections of a sinusoidal waveform. As shown in Fig.
1.10, the frequency of a particular section is determined by the binary value to
be transmitted. A simulation program including a digital filter to model the
transmitter, detector and the transmission channel may attempt to answer the
following questions: What frequencies should be used? How long should the
duration of each segment be? What detector is suitably reliable but reasonably
expensive or fast? To investigate the complicated effects of interference,
random noise can be added to the transmitted signal and the resulting
characteristics of a given detector can be examined. All these procedures can
be performed before any physical equipment is manufactured or installed.
Such a simulation program is developed by the student in the projects at
the end of the chapters. In one project, the transmitter is simulated by a
program that generates the waveform that corresponds to a given binary
sequence. Other projects consider different digital filters that operate on the
received sinusoidal sequence to retrieve the value of the binary digit
transmitted.
Simulating a robot arm. Another advantage of simulation programs is that the
various parameters acting in a physical system can each be examined in detail.
For example, consider the simulation of a robot arm, shown in Fig. 1.11. The
actual arm is constructed of some material that has mass and elasticity. The
input to such an arm is shown as a rotational displacement that is modeled as a
step function at one joint of the arm. The desired output may be the location
of the manipulator at the end of the arm. The time response of the output may
be important for ensuring that the arm construction satisfies the specification
for quickness or overshoot.
One simple digital filter model of the arm is a second-order filter whose
coefficient values are related to the mass and elasticity of the material. The
physical input signal is converted into an input sequence, as shown in the
figure, and the output sequence of the filter indicates the response to be

INTRODUCTION

(a)

11

--"0"---,

r--"O"

"t"----i>I1IE-E

x(t)

~----~-----+--~--~-+--~----~-----,time

(b)

ftGURE 1-10
Frequency-shift keying (FSK) signals. a) Example of continuous-time signal from actual system. b)
Simuated sequence produced by computer.

{e(D)} = {

8(D)~

{e(n)}
___ ...r_.:::_..... Ili n)}
I

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

= (y(n)}

:111 III II III I 1':,

o

{lin)}

~tlill [ rill I "":,

o

!!

ftGURE 1-11
Simulating a robot a'lll by applying an excitation sequence and computmg the response sequenct'

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

expected if the model is accurate. The value of this simulation model is that
the coefficient values of the digital filter can be adjusted to produce the desired
response. Then the final values can be employed to determine the physical
dimensions of the robot arm.

1.S REPLACING ANALOG FILTERS

WITII DIGITAL FILTERS
One of the first applications of digital filters was to replace analog filters,
composed of resistors R, capacitors C and inductors L. This replacement is
usually performed to overcome some of the inherent limitations of analog
components, including the fluctuation of the component values with temperature and age. Another disadvantage of analog components is their large
physical size, especially in the case of large-valued inductors and capacitors.
A further disadvantage of analog filters is the difficulty encountered in
changing the component values iq an existing filter. In many cases, the
equivalent digital filter is implemented as a sequence of instructions in a
computer, which can easily be changed with an editor, while soldering may be
required to change a physical component in the analog filter. When the digital
filter is implemented in hardware, the multiplier coefficients are often stored in
a read-only memory (ROM), which can readily be replaced if a filter
modification is desired. These coefficients can also be stored in read/write
memory and be changed automatically as the nature of the input signal
sequence changes. This added flexibility of digital filters has created an area of
research in so-called adaptive filters that is currently very active.
Replacing an analog filter with a digital filter in a given application
involves several considerations. For illustration, let us examine the situation in
Fig. 1.12, which shows a continuous-time input signal x(t), a simple analog
RC lowpass filter and the output signal y(t). To replace this analog filter with
the equivalent digital filter, analog-to-digital conversion (ADC) must be
performed to produce a discrete-time input sequence {x(n)} from x(t). The
input sequence {x(n)} is then processed with a digital filter, implemented with
special-purpose hardware or as a computer program, to produce the output
sequence {y(n)}. Finally, a digital-to-analog conversion (DAC) must be
performed to reconstruct the continuous-time signal y(t). For simple analog
filters, the added expense involved in the ADC, DAC and the hardware to
implement the digital filter does not usually make this substitution economically feasible. But with the proliferation of microcomputers and the reduced
cost of analog-to-digital converters, performing a filter operation with digital
circuits is becoming common.
In some applications, the analog filters themselves may be ~tly. In
geophysical and biomedical applications, signals with very low-frequency
components, typically less that one Hertz, require very bulky and expensive
inductors and capacitors. In these cases, digital filters are commonly employed
as replacements. In some applications, such as those involving very high

INTRODUCTION

13

ANALOG FILTER
x(t)

-"'I)

DIGITAL FILTER

ftGURE lU
Replacing an analog filter with a digital filter requires the additional operations of an
analog-ta-digital conversion (ADC) and a digital-to-analog conversion (DAC).

frequencies, analog filters still have the advantage. But with decreasing cost
and increasing speed of digital circuits, more and more analog filters are being
replaced with their digital counterparts.

1.6 OVERVIEW OF THE COURSE

The digital filter design and signal processing concepts described in this book
can be divided into five categories: analysis, synthesis, the fast Fourier
transform, finite-accuracy effects, and inverse filters.
ANALYSIS. We analyze digital signal processing algorithms by determining
their time and frequency domain characteristics. This allows us to attain an
intuitive understanding of their capabilities and limitations. The time-domain
performance of digital filters is described in terms of the filter's unit-sample

14

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

response sequence, denoted by {h(n)}. This sequence is analogous to the

impulse response of analog filters. A convolutional equation allows us to
determine the output sequence {y(n)} from the input sequence {x(n)} and the
unit-sample response {h(n)}. The unit-sample response sequence also permits
us to determine whether a filter is stable. Linear difference equations provide
an alternate time-domain description that is be useful for implementing digital
filter structures.
Filter specifications are commonly dpressed in the frequency domain.
Hence, we examine the relevant aspects of the Fourier domain that define the
frequency domain properties of signals and filters. The Fourier transform of
the unit-sample response {h(n)} is the transfer function H(e iw ) of the filter and
it describes the gain of the filter at different frequencies. The Fourier transform
of a data sequence {x(n)} is called the spectrum X(e iw ) and it defines the
frequency content of the signal. We explore the problems encountered when
generating a discrete-time sequence by sampling a continuous-time signal and
when reconstructing the continuous-time signal from the discrete-time sequence. The discrete Fourier transform (DFT) allows us to evaluate the
Fourier transform by computer. The consequences of employing' the DFT for
signal processing and spectral analysis applications are investigated.
An alternate and more general analytic technique than the Fourier
transform is the z-transform. The system function H(z) is defined as the
z-transform of the unit-sample response {h(n)} and is used for the analysis and
synthesis of digital filters. The complex z plane is the domain of interest for the
z-transform. The representation of a digital filter as a collection of poles and
zeros in the z plane will provide a useful interpretation of the filter frequency
response.
SYNTHESIS. After having a firm and intuitive understanding of the time and

frequency behavior of digital filters, the goal is then to design and implement a
digital filter to perform a desired task. The synthesis procedure involves
combining the digital filter elements and specifying the values of the multiplier
coefficients to satisfy a desired specification.
Most often, the transformation that is to be performed by the digital filter
is specified in terms of the magnitude response IH(eiw)l. The magnitude
response for an ideallowpass filter, for which the frequency components in the
input sequence from zero to some cutoff frequency Wc are passed through the
filter unaffected, while all th~ others are completely eliminated from the output
sequence, is shown in Fig. 1.13(a). This ideal cannot be achieved in practice,
but must be approximated. A practical specification that can be satisfied by a
digital filter is shown in Fig. 1.13(b). This desired specification allows slight
deviations from the ideal response: E in the passband, () in the stopband, and a
nonzero frequency interval for the transition between the passband and
stopband. A successful digital filter design produces a magnitude response that
falls within the allowed regions.
We will consider two different digital filter types to satisfy a given

INTRODUCTION

I<

PASS BAND

0
(b)

=-jE

STOPBAND

w,

;.

15

IHdc>"cd(eJ"')1

Successful design

1-

I
I
I
I
I
I

-----------1-----------~.~~--~~--~
OL-------~--------~~~=--=~~

I<

PASS BAND

~IE
wp

TRANSITION
RANGE

STOP BAND

~I W

Ws

FlGURE 1-13
Magnitude response curves for a digital low pass filter. (a) Ideal low pass rlter response. (b)
Practical low pass filter specification allows derivations from the ideal response. A successful
digital filter has a magnitude response that falls within the allowed regions.

specification: infinite impulse response (IIR) filters and finite impulse response
(FIR) filters. IIR digital filters are usually designed by extending classical
analog filter design procedures. An I1R filter design is typically accomplished
in three steps. First, an analog lowpass filter is designed to meet the desired
passband specification. The most commonly employed are the Butterworth,
Chebyshev and elliptic analog lowpass filter design procedures. Second, an
analog-to-digital filter transformation is employed to obtain a digital lowpass
filter. The impulse-invariance method and the bilinear z-transform method are
described for transferring these analog designs into their digital counterparts.
In the final step, a frequency transformation is employed to obtain a highpass,
bandpass or bandreject filter.
/
The design procedures for FIR filters have been develope~SpeciallY for

16

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

digital filters and do not have analog filter counterparts. Two FIR filter
structures will be investigated. The first involves implementing the desired
filter unit-sample response in a tapped-delay line structure. Since the desired
response often has an infinite number of nonzero elements, a finite-duration
window must be applied to eliminate all but a finite number of these elements.
We investigate the effect of this truncation on the resulting filter's transfer
function. The second structure is the frequency-sampling structure, which
employs a comb-filter followed by a bank of resonators and implements the
filter by specifying the magnitude response at a finite number of frequency
points. A comparison of the two design procedures is performed to determine
under which conditions one is more efficient than the other.
FAST FOURIER TRANSFORM (FFf). In addition to digital filters, we also
explore the use of the fast Fourier transform (FFT) algorithm for processing
data and performing spectral analysis. The FFT is an elegant and efficient
method for computing the discrete Fourier transform that allows the filtering
operation to be performed in the frequency domain. We present both an
intuitive description of the FFT, as well as a more rigorous mathematical one,
to indicate how this efficiency is achieved. The utility of employing the FFT for
filtering applications is described and compared with the methods of direct
digital filter implementation.
FINITE-ACCURACY EFFECTS. An important concern for implementing practical digital filters involves finite-accuracy effects caused by the limited
precision to which the sequence and coefficients values are represented within
the computer. Similar errors are also introduced in the numerical round-off
routinely performed in tedious hand calculations. In the analysis and synthesis
discussions above, we dealt with the ideal, or infinite-precision, case. To
process data on a computer, the data values must first be quantized to a finite
number of bits, usually equal to or smaller than the word size on the computer.
The effects of this quantization error will be investigated. The coefficients used
in the digital filter multipliers must also be quantized. The slight change in
these coefficient values modifies the frequency response of the filter. A third
source of error occurs when several numbers are added or multiplied and the
result excee'ds the limits of the computer. In this case, the result must be scaled
and truncated, introducing still another source of error that can produce
unwanted oscillations at the output of some IIR filters. We consider analytic
methods of characterizing these three common quantization problems.
INVERSE FILTERS. In the final chapter, we discuss the interesting and
important topic in signal processing called inverse filtering. This topic integrates many of the digital signal processing concepts presented in the previous
chapters to solve the problem of removing distortions from observed signals
that are introduced by measuring instruments or by physical transmission
media. It is hoped that this topic whets the student's appetite for pursuing
further study in this most interesting area of research and development.

JNTRODUcnON

17

1.7 DESCRIPTION OF THE COMPUfER

PROJECfS
Along with the standard problem sets at the end of each chapter, a set of
projects to be performed on a computer is also included. The pedagogical
value of these projects cannot be overstressed, since they are indicative of the
problems encountered in practice. The computer is a powerful tool for signal
processing and the projects given in this book illustrate its flexibility. Most of
the projects in Chapters 2 through 6 can be analytically determined with little
effort. The analytic results are meant to verify the computer results and to
instill some faith in the algorithms. As the analysis becomes more difficult, the
computer results can help guide the analytic procedure. The design projects in
Chapters 7 through 11 are almost impossible to do analytically, as are most
practical filter designs. By that stage in the course, the intellectual insights are
achieved almost exclusively from the computer results. Using the computer in
this fashion is the basis for computer simulations, which are often performed to
explore the performance of complicated systems before they are implemented.
The subroutines for the projects are meant to be written by the st\1dent,
in FORTRAN, the current universal language for signal processing, or in the
increasingly popular C or PASCAL (BASIC is slightly less desirable, because
of its slow speed, but acceptable). To minimize the amount of programming, a
structured approach is used in which the routines written for the initial projects
are used again in subsequent projects. Each routine performs a simple, and
hence easily verifiable, operation. Constructing more sophisticated programs
from these simpler building blocks. aids the student in conceptualizing the
digital filter operation a!,ld for trouble-shooting the more complicated projects
later in the course. By the end of the course, the student should have an.
intuitive understanding of digital signal processing that is invaluable for
applying it to novel situations. As an added benefit, students will also have a
flexible library of signal processing programs that they have written themselves, and hence understand intuitively.
Graphic outputs produced by the projects are displayed on the terminal
and on a printer, being generated by programs provided in Appendix A. These
plots, although having crude resolution, are adequate for illustrating the ideas.
They also have the advantage of not requiring special plotting hardware. The
plotting subprograms can also easily be modified to display the results on other
types of digital plotters and display terminals.
The experience gained through this course should whet the appetite for
the study of advanced topics in filter design and signal processing. References
to material describing advanced topics are provided at the end of each chapter.
The student will also be well prepared to apply digital filtering techniques to
practical problems.
After this course, the student can easily become facile with the various
programs that are available for automatically designing optimal digital filters.
Examples are the digital filter program packages offered by the IEEE, or in
SIG, the software package from the Lawrence Livermore National Laboratory

18

lNTRODUcnON TO DIGITAL SIGNAL PROCESSING

at the University of California. An important prerequisite for employing these

programs successfully is the intuitive understanding of the tradeoffs and
subtlety involved in the methods for digital signal processing. This is the
material stressed in this book.

1.8 SUMMARY
In this chapter, we have provided the motivation for digital filtering, defined
the components of a digital filter, and indicated some common applications.
An overview of the course material has described the inter-relationship of the
topics that describe the theory and application of digital filters to signal
processing.

FURmER READING
IEEE Acoustics, Speech and Signal Processing Magazine. This bimonthly magazine of the
Acoustics, Speech and Signal Processing Society of the Institute of Electronics and Electrical
Engineers (IEEE) provides tutorial articles on various topics covering digital signal processing
and digital filters.
IEEE Transactions on Acoustics, Speech and Signal Prl?cessing. This bimonthly journal publishes
articles describing the most recent results in the development and implementation of digital
filters and digital signal processing techniques.
Rabiner, L..R. and C. M. Rader (eds.): Digital Signal Processing, IEEE Press, 1972; and Digital
Signal Processing ll, IEEE Press, 1975. [These books contain reprints from various journals
that provide an organized and historical treatment of digital filter theory and applications.]

REFERENCES TO TOPICS
Pascal
Cooper, Doug, and Michael Clancey: Oh! Pascal!, W. W. Norton, 1982.
Jensen, Kathleen, and Niklaus Wirth: Pascal User Manual and Report, Springer-Verlag, Berlin,
1985.
Mallozzi, John, and Nicholas De Lillo: Computability with Pascal, Prentice-Hall, Englewood
Cliffs, NJ, 1984.

C language
Harbison, Samuel, and Guy Steele Jr.: C, A Reference Manual, Prentice-Hall, Englewood Cliffs,
NJ,1984.

FORTRAN
Didday, Richard, and Rex Page: FORTRAN for Humans, West, 1981.
McCracken, Daniel: FORTRAN, Wiley, New York, 1967.
BASIC

Maratek, Samual: BASIC, Academic Press, New York, 1975.

INTROOycnON

19

Comptder (OIIUIIlUIicatioas
Schwartz, Mischa: Telecommunication Networks, Addison-Wesley, Reading, Mass. 1987.
Tanenbaum, Andrew: Computer Networks, Prentice-Hall, Englewood Cliffs, NJ, 1981.
Lathi, B. P.: Modem Digital and Analog Communication Systems, Holt, New York, 1983.

DIPDI sIpaI processIq propam packages

Programs for Digital Signal Processing, IEEE Press, New York, 1979.
Lager, Darrel L., and Stephen G. Azevedo: SIG -A General-Purpose Signal Processing Program.
Proceedings of the IEEE, Vol. 75, No.9, pp. 1322-1332, 1987.

RoIIot

_aIadoa

IEEE Transactions on Pattern Analysis and Machine Intelligence.

Speedt proceaiq

IEEE Transactions on Acoustics, Speech and Signal Processing.

MecIkaI sIpaI procesIIbIc

IEEE Transactions on Biomedicfll Engineering.

CHAPTER

2
DISCRETE-TIME
DESCRIPTION
OF SIGNALS
AND SYSTEMS

,r

2.1 INfRODUcnON

. In this chapter we investigate procedures to describe the time-domain behavior

of discrete-time systems, of which digital filters are a special class. After
establishing the discrete-time signal notation, we define several sequences that
are important for the llnalysis of digital filters. The superposition principle for
linear systems is then employed to derive useful relationships governing digital
filters. The unit-sample response is presented as a time-domain descriptor of a
discrete-time system. The linear convolution operation relates the output
sequence of the filter to the input sequence and the unit-sample response.
Linear constant-coefficient difference equations are employed for the analysis
and synthesis of digital filters. A procedure to write a digital filter program
directly from these time-domain equations is described.

2.2

DISCRETE~TIME

SEQUENCES

In this book, we will be dealing with discrete-time sequences, which will be

denoted either by {x(n)}, or by x(n); for -Nl $;n $;N2, where Nl and N2 may
possibly be infinite. As mentioned in Chapter 1, the term discrete-time can also
refer to sequences of other physical parameters, such as discrete-location for

DISCRETE-TIME DESCRIPTION OF SIGNALS AND SYSTEMS

21

image processing applications, discrete-distance for sonar ranging systems, or

discrete-position for robotic arm sensors_ The particular value of the sequence
at time k will be denoted simply by x(k), with no additional qualifiers on
the index. In this book, we are primarily concerned with sequences that are
produced by sampling a function of a continuous-valued variable at equally
spaced intervals. These, in fact, are the most common types of sequences in
digital filter applications. We now consider several important sequences that
are used throughout this book for analyzing digital filters.
UNIT-8AMPLE SEQUENCE. The unit-sample sequence contains only one
nonzero valued element and is defined by

d n = { 1 for n = 0

()

0 otherwise

(2.1)

This sequence is shown in Fig. 2.1 and will play an important role as the input
sequence to a digital filter. Since the unit-sample sequence contains only one
nonzero element, the form of the output sequence then directly indicates the
effect produced by the digital filter.
The delayed unit-sample sequence, denoted by {d(n - k)}, has its
nonzero element at sample time k:
d(n - k)

= {~

for n = k
otherwise

(2.2)

The unit-sample sequence is also an important analytic tool because of its

sifting property, by which a particular element can be extracted from the entire
sequence. To demonstrate this property, we note that the value of {x(n)} at
time k can be written as
DO

x(k)= ~ d(n-k)x(n)

(2.3)

n=-OCI

as shown in Fig. 2.2. This follows because d(n - k) is nonzero only when its
argument is zero, or when n = k. When multiplying the two sequences together
term by term, only the single element x(k) remains nonzero and the sum
evaluates to x(k). Hence, the unit-sample sequence can be used to select, or
sift, a single. element from the entire sequence. This sifting property will be
important below in the derivation of the convolution relationship between the
digital filter input and output sequences. A more convenient, but equivalent,
form of Eq. (2.3) that is used below employs the time-reversed version of the
{d(n)}

... r. ..

-2-10 123 ...

FIGURE 2-1
Unit-sample sequence.

22

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

(x(n)}

... 111111!]11
I
...

k ... -I 0 1 2

:>

(d(n-k)}

.c.
k

:>

n
(u(n)}

... ...IIIIIII
o

FIGURE 2-2
Sifting property of the unit-sample sequence.

FIGURE 2-3
Unit-step sequence.

unit-sample sequence having its nonzero element at n

= k:

00

x(k)=

d(k-n)x(n)

(2.4)

n=-CIO

UNITSTEP SEQUENCE. The unit-step sequence is defined as

u(n) = {~

forn~O

otherwise

(2_5)

and is shown in Fig. 2.3. The unit-step sequence is used for defining the
starting point of a sequence in analytic expressions. For example, the sequence
an
for n ~O
(2.6)
x(n) = {
otherwise

0,'

can be written simply as

for aU n

(2.7)

SINUSOIDAL SEQUENCES. Sinusoidal sequences play an important role in

the frequency-domain analysis of digital filters. Given some radian frequency

w, we can form the discrete-time sinusoidal sequences {cos(wn)} and
{sin(wn)} by evaluating these functions at the corresponding value of the
argument. Examples are shown in Fig. 2.4. Strictly speaking, the units of ware
radians per sampling interoal. By convention, the sampling interval is taken as
a dimensionless quantity, allowing OJ to be specified in radians. In the next
chapter, we will consider the sampling of continuous-time signals and relate OJ
to continuous-time frequency values.

DISCRETETIME DESCRImON OF SIGNALS AND SYSTEMS

(cos(wn)}

"'1.

1II

w=.~

. iiI.
-I 0 1

(sin(wn)}

1I

23

....
21 II

(o)=~

...
tIllllt
10

III1

-2- l o

1 2

FIGURE l-4
Two discrete-time sinusoidal sequences.

Discrete-time sinusoids exhibit an interesting property that is important

for evaluating the Fourier transform of a discrete-time sequence: The set of all
distinct values that a discrete-time sinusoidal sequence can have occurs for
values of co lying in the range -:re:s; co <:re.
To illustrate this property, let us consider two frequency values, Wo is in
the interval [-:re, :re), and WI outside this range. For example, let
COl

= Wo + 2m:re

(2.8)

where m is a nonzero integer. The values of the cosine sequence at frequency

COl are given by
{cos(co1n)} = {coscoo + 2m:re)n)}

= {cos(won)}

(2.9)

This last result follows because the cosine function is periodic with period 2:re
and the argument is won plus an integral number ( =tnn) of 2:re. In Fig. 2.5,
this effect is shown for COo = 0.1:tr and COl = 2.1:re. This result is also true for the
sine sequence.
COMPLEX EXPONENTIAL SEQUENCE. Before we consider the complex
exponential sequence, we will establish the notation for complex numbers and
the complex plane used throughout this book. Let c be a complex number.
Then it can be written as
c =a + jb

(2.10)

where a is the real part, b is the imaginary part, and j = (_1)112. The
representation of c on the complex plane is shown in Fig. 2.6. An alternate
form of a complex number is the complex vector notation:

c = lei ei Arg[c)
where

(2.11)

lei is the magnitude and Arg[c] is the argument or phase, which is

24

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

{cos (0. lIT) }

...
{cos (2.IIT)}

FIGURE 25
Two analog signals having different frequencies can proouce the same values for a discrete-time
sinusoidal seque\lce.

defined by the angle with respect to the real axis in the complex plane. The
following geometric relationships can be derived from Fig. 2.6:
and

Icl 2 = a2 + b2

(2.12)

Arg[c] = arctan[bla].

(2.13)

The complex exponential sequence is important for a compact frequency

analysis of digital filters, and is given by
{e("+i",)n} = {e<7n ei"'n}
(2.14)
When both a and ro are real numbers, each element of the sequence is
expressed in complex vector notation with magnitude e"" and phase ron.
A special case of this sequence occurs when~ a = O. The exponential
sequence can then be related to the sinusoidal sequences by the geometric
Imaginary
c=a+jb

-------~---'L-.----I:--~

Real

FIGURE 2-6
The complex plane.

DISCRETE.TIME DESCRIPTION OF SIGNALS AND SYSTEMS

2S

Imaginary
c=elNf'I

------4<--"---'--_ Real

Imaginary

-----~-+---_:_.."..__:_~

Real

Imaginary
2)sin(wn)

-e-"'"

eJ"'"

____--l..-______

Real

F1GURE 27
Interpretation of the Euler identities in the
complex plane.

relationships shown in Fig. 2.7, which are known as the Euler identities. These
are given by
ei",n = cos( ron) + j sin( ron)
(2.15)
ei",n + e-i",n
(2.16)
cos(ron) =
2
sin(ron) =

ei",n - e-i",n
2j

(2.17)

The property of sinusoidal sequences, that all values can be represented

with ro in the range [-Jr, Jr), is also true for complex exponentials. To
illustrate this, consider rol = roo + 2mJr, where m is any integer. Then
(2.18)

26

1NTR0DUcnON TO DIGITAL SIGNAL PROCESSING

since ej2mn = 1, when m is any integer. This identity can also be demonstrated
in the complex plane. Increasing the phase of a complex number by a is
equivalent to rotating the complex vector by angle a, counter-clockwise if
a> 0 and clockwise if a < O. If a = 2m:r, where m is any integer, then the
complex vector makes m complete revolutions about the origin and returns to
its original orientation.

2.3 SUPERPOSmON PRINCIPLE FOR

LINEAR SYSTEMS
The digital filters considered in this book are a special class in the general
category of linear systems. Any discrete-time system can be thought of as
transforming an input sequence {x(n)} into an output sequence {y(n)}
through a transformation denoted by G { . }. However, two characteristics
define a linear system: (1) for a given input sequence, a change in the
amplitude scale of the input sequence results in the same amplitude scale
change in the output sequence, and (2) if two input sequences are added and
the sum is applied to the system, the resulting output is the same as the sum of
the responses to the individual inputs. These ideas can be formalized in the
statement of the superposition principle.
SUPERPOSmON PRINCIPLE. Given a system producing a transformation
G{ .}, two arbitrary input sequences {xI(n)} and {x2(n)} that result in
{YI(n)} = G{xI(n)} and {Y2(n)} = G{x2(n)}, and two constants a and (3; then

the system is defined to be linear when

G{axl(n) + (3x2(n)}

= aG{xI(n)} + (3G{x2(n)}
= a{YI(n)}

+ (3{Y2(n)}

(2.19)

The following example illustrates the applicatiOn of the superposition

principle.
Example 2.1. Linearity of the 3-sample averager. The transformation produced
by the 3-sample averager is given by

y(n) = Ux(n + 1) +x(n) +x(n -1)] = G{x(n)}

To test the linearity, we apply the scaled sum of two arbitrary sequences, or
{axt(n) + (3x2(n)}, to the averager. The output is then equal to

G{axt(n) + (3x2(n)} = Uaxt(n + 1) + (3x2(n + 1) + axt(n) + (3x2(n)

+ axt(n -1) + f3x2(n -1)]

="3a [xt(n + 1) + xt(n) + xt(n -1)]

+~[X2(n + 1) +x2(n) +x2(n -1)]

= a{Yt(n)}

+ (3{y2(n)}

DISCRETETlME DESCRIPTION OF SIGNALS AND SYSTEMS

27

This last equation is simply the sum of the two outputs that would have been
observed if only {crxt(n)} or {,8x2(n)} had been applied. Hence, the 3-sample
averager is a linear system.

Example 2.2. A nonlinear system. Let us consider a square-law device, in

which the output is the square of the input:
{y(n)} = G{x(n)} = {x 2(n)}
By applying two scaled inputs into this system, we have at the output:
G{crxt(n) + ,8x2(n)} = [crxt(n) + ,8x2(nW
= ~xi(n)

+ ,82x~(n) + 2a,8Xt(n)x2(n)

which is not equal to ~xi(n) + ,82x~(n), the output required for the system to be
linear. Hence, the square-law device is a nonlinear system.

2.4 UNIT-SAMPLE RESPONSE

SEQUENCE
One important consequence of the superposition principle is that it allows us to
derive a convenient relationship between the input and output of a linear
system in the time domain. Applying the sifting property-of the unit-sample
sequence given by Eq. (2.4), we can write
{y(n)}

= G{x(n)} = GL~oo d(n = G{ ..

k)X(k)}

+ x( -1)d(n + 1) + x(O)d(n) +x(l)d(n -1) + ... } (2.20)

Since the system operates through its transformation G{ . } on sequences,

we can view the output of a system as the superposition of responses to
delayed and scaled unit-sample sequences, as shown in Fig. 2.8. The
unit-sample sequence having its nonzero element at n = k, d(n - k), is scaled
by x(k). Applying the superposition principle, we can 'write the output as
{y(n)}

= ... + x( -1) G{d(n + I)} + x(O) G{d(n)} + x(l) G{d(n -I)} + ...
00

= 2:

x(k) G{d(n - k)}

(2.21)

k=-oo

G{d(n - k)} is the output sequence produced by the system when the input is
a unit-sample sequence having fts nonzero element at n = k. This output is
called the unit-sample response, and is sufficiently important to merit its own
symbol:
G{d(n - k)}

= {h(n, k)}.

(2.22)

In words, {h(n, k)} is the response of a linear system whose input is a

unit-sample sequence with nonzero element at position n = k. The following
examples illustrate the concept of the unit-sample response.

28

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

(..)}
If:
{.t(II)}

---1

~(y(.)}

LINEAR
SYSTEM

= {d(lI)}

I..
0

{Y(II)}

>
~

II

= {h(lI)}

.. IIIt ...
-I 0 I

II

Then:
{Y(II)}

{.t(II)}

~I

0 I

II

-I

.t(

)(d(lI+ I)}

-I
L~-I:

II

o
---.........l-Tl"""'
r-r......'""'...........;..~I .t(-l){h(lI+ I)}
+

.t(O){d(lI)}

+
.t(I){d(lI+ I)}

I.t!O~
0

II

T , >
o

.t(O){h(lI)}

II

... r.
oI

II

Ji'lGURE U
The input sequence to a \inear system can be viewed as the superposition in time of a set of scaled
unit-sample sequences. The output sequence is then the superposition of the scaled unit-sample
responses.

Eumple 2.3. Uait-sample response of the 3-sampIe .veneer. Let

y(n) =lx(n + 1) + lx(n) + lx(n -1)

To find the unit-sample response {h(n, O}}, we set {x(n}} = {d(n}} and solve for
the values of y(n) for each n:
For n = -2,
y( -2) = id( -1) + Id( -2) + ld( -3) = O.
Forn S -2,
y(n) = 0, since d(n) = 0 for n <0.

DISCRETETlME DESCRIPTION OF SIGNALS AND SYSI'EMS

29

{h(n,O) }

.:111 ..

3 21 0 1 2 3

Forn=-I,
For n =0,
For n = 1,
For n=2,

FIGURE 2-9
Unitsample response
sample averager.

y(-I) = id(O) + !d(-I) + id(-2) =

of nonrecursive three-

i.

y(O) = !d(I) + i~(O) + id( -1) '" }.

y(I) = id(2) + id(I) + id(O) = ...
y(2) = id(3) + id(2) + id l" . O.

For n>2, y(n) =0, since the nonzero value of {d(n)} has moved out of the
memory of this filter. Since {d(n)} is the input, the output is defined to be
{h(n, O)}. Hence
h(n, 0) =

{~,

for-I""n",,1
otherwlse

This sequence is shown in Fig. 2.9.

The unit-sample response is computed by starting at a point in time, n,

for which all the previous inputs and the contents of the delays and advances
are all equal to zero. This state of the filter is known as the zero initial
condition.
EumpIe 2.4.

Unit-5lUllple

respoIIIe

of. first-order recursive filter. Let

y(n) = {~(n -1) +x(n)

for n 2:0
otherwise.

To find {h(n, O)}, we let {x(n)} = {d(n)} and apply the zero initial condition.
For n<O, y(n) =0, because d(n) is zero for n <0 and y(-I) =0.
For n = 0, y(O) = ay( -1) + d(O) = 1.
For n = 1, y(I) = ay(O) + d(I) = a.
For n = 2, y(2) = ay(I) + d(2) = a2
For n =k,.y(k) =ay(k -1) + d(k) = a".
Hence,
h(n, 0) =a"u(,n)

for all n

Some typical sequences are shown for different values of a in Fig. 2.10.

This concept of a unit-sample response can also be applied to timevarying systems, in which the multiplier coefficients vary with n, as illustrated
in the next example.

30

INTRODUCfION TO DIGITAL SIGNAL PROCESSING

h(n,O)

II

for all n

.'.
0<.<1

..1111 J
o

= a"u(n)

t , " ";"

)0

-I <a<O

a<-l

FIGURE 2-10
Four possible unit-sample response sequences that can be produced by first-order recursive
filter.

Example 2.5. Unit-sample response of a time-varying system. Let

Y()
n =

{ o~IY(n-l)+x(n)
n+

forn""O

otherwise.

Applying the zero initial condition and using the same approach as in the
previous example, we have
h(O, 0) = (1)(0) + 1 = 1

h(I, 0) = 0>(1) + 0 = ~
h(2, 0) =
h(3, 0)

<no> = ~

= mm = -i4

If the unit-sample is applied at time n = 1, i.e. {x(nn = {d(n

{y(n}} = {h(n, In, where
h(O, 1) = (1)(0) + 0 = 0

-In, then

h(l, 1) = G)(O) + 1 = 1

h(2,I)=m(I)+O=i
h(3,I)=mm=fi

In

On.

Note that {h(n,

is a different-valued sequence from {h(n,
This example
illustrates that, for a time-varying filter (whose coefficient values vary with time
n), the values of the unit-sample response depend upon the time of application of
the unit-sample sequence.

DISCRETE-TIME DESCRIPTION OF SIGNALS AND SYSTEMS

31

The general time-domain relationship between the system output value at

time n and the input values can be obtained by combing Eqs_ (2-21) and
(2.22), to give

y(n) =

(2.23)

x(k) h(n, k)

k=-oo

This form is called a convoLutionaL reLationship. This equation can be simplified

when it is applied to time-invariant systems, which are discussed in the next
section. The references describe adaptive filters whose coefficient values
change to meet various conditions.

2.5 TIME-INVARIANT SYSTEMS

We will further restrict our attention in this book to the important class of
Linear time-invariant (LTI) systems, or those whose transformations do not
change with time. In simplest terms, for an LTI system, a shift of no samples in
the input produces onLy a corresponding shift in the output. That is, if
{yen)} = G{x(n)}, then

{yen - no)}

= G{x(n -

no)}

(2.24)

for all no

One characteristic of an LTI system is that the form of the unit-sample

response does not change with the time of application of the unit-sample
sequence. Then, rather than being a function of two variables, nand k, the
unit-sample response becomes a function of only n - k, or the time as
measured from the application of the unit-sample. Analytically, we then write

G{d(n - k)}

= {hen, k)} = {hen -

(2.25)

k)}

This equation implies that a shift in the time that the unit-sample sequence is
applied to the system results only in a corresponding time shift in its
unit-sample response, as shown in Fig. 2.11. A time-invariant digital filter can
easily be recognized, since its coefficient values do not vary with time.
{d(n))

.'I

0 1 2

.. 1.

o 1

.. IJ
o1

Tt
2

(h(n.k)} = {h(n-k)}

(h(n))

, , ";'
n

.... IITt, ..;

o

F1GURE 211
Shift-invariance properly of the unit-sample response for a time-invariant system.

32

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

Applying Eq. (2.25), the convolution relationship of a LTI system can be

written as
00

y(n)=

x(k)h(n-k)

(2.26)

= {x(n)} * {hen)}

(2.27)

k=-cc

or, more concisely,

{yen)}

where * will denote the convolution operation. In evaluating a convolution

expression, it is often helpful to note that convolution is a commutative
operation, i.e.
{yen)} = {x(n)}

To demonstrate this, we let m

yen) =

L
m=cc

* {hen)} = {hen)} * {x(n)}

=n -

(2.28)

kin Eq. (2.26) to get

00

x(n - m) hem)

= L

h(m)x(n - m)

(2.29)

m=-oo

The desired result is obtained when m is replaced by k. The limits of the first
sum in Eq. (2.29) appear in reverse order, but can be reordered without
consequence because the same terms appear in the evaluation of the sum.
The convolution operation is also distributive, i.e.
{hen)}

* {xt(n) +x2(n)} = {hen)} * {xt(n)} + {hen)} * {x2(n)}

(2.30)

This result can be shown directly by applying the superposition principle.

COMPUTING A DISCRETE-TIME CONVOLUTION. Since the convolution

operation is one of the most common and important in digital signal

processing, we present a step-by-step procedure for the evaluation of (2.26).
Step 1. Choose an initial value of n, the starting time for evaluating the
output sequence. If {x(n)} starts at n = nx and {hen)} starts at n = nh, then
n = nx + nh is a good choice. The values of x(n) for n < nx and of h(n) for
n < nh are then assumed to be zero.
Step 2. Both sequences are expressed in terms of the index k. Note that
x(k) is identical to x(n), while hen - k) is a time-reversed version of hen). If
hen) starts at n = nh' thea the last point of hen - k) occurs at k = n - nh.
Step 3. The two sequences are multiplied element by element and the
products are accumulated over all values of k. This sum of products then
becomes the value of the output for index n, or y(n).
Step 4. The index n is incremented and steps 2, 3 and 4 are repeated until
the sum of products computed in Step 3 is zero for all the remaining values
of n.

DISCRETE-TIME DESCRIPTION OF SIGNALS AND SYSTEMS

33

These four steps are illustrated in the following examples.

Example 2.6. Convolution of two finite-duration sequences. Let

x(n) =

{~

for -1 sn s 1
otherwise

h(n) =

{~

for -1 Sn S 1
otherwise

and

The evaluation of {y(n)} = {h(n)}

given above is shown in Fig. 2.12.

* {x(n)}

by Eq. (2.26) using the four steps

Step 1. The evaluation will start at n = -2.

Step 2 Both sequences are shown indexed by k, with {h(n - k)} being the
time-reversed version of {h(n)}. Since {h(n)} starts at n = -1, the last point of
{h(n -k)} occurs at k =n + 1. When n = -2, k =-1.
Step 3. Computing the product of the pair of sequence values for each k,
only the k = -1 terms yield a nonzero product. This result then becomes the
value of y(-2).
Step 4. The value of n is incremented and the process above is repeated
until n = 3. The relative p'ositions of the two sequences are shown for the values
of n that produce nonzero values for y(n).
Example 2.7. Convolution of a 6nite-duration seqnence with an infiniteduration sequence. Let
forOSn s2
otherwise

x(n)={n+1,
0,
and

h(n) = a"u(n)

for all n

We will perform the convolutio.n in two ways. First, using Eq. (2.29), we have
~

y(n)=

2:

2:

aku(k)x(n-k)

for n -k =0,
for n -k = 1,
for n -k =2,
otherwise

ork=n
ork=n-1
ork =n-2

h(k)x(n-k)=

k=-ao

From Fig. 2.13, note that

z(n

-k)-{~

Hence, the only nonzero terms in the sum are

y(n) = a"u(n) + 2a"-lu(n -1) + 3a"- 2u(n - 2)

The second way of evaluating the convolution is by using Eq. (2.26). The
evaluation, of course, produces the same result, but is simpler because of the

34

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

{x(n)}

{h(n)}

..'I II. .

-3 -2 -1 0 1 2 3

.. 'I II...

-3 -2 -1

n=-1

n=-2

{x(k)}

{~~~~

(x(k)}

1II. ...

I 11. ... :

-4 -3 -2 -1

y(-2)=1
n=O

0 1 2 3

0 1 2 3 4 k

y(-I)=2

n=l

III
...

y(1)=2

y(0)=3

n=2

III .
III ..

(x(4)}

(h(2-k)}

-4 -3 -2 -1 0

(y(n)}.

I'l'I r

-4 -3 -2 -1 0

2 3 4 k

l'

y(2) ... 1

ftGURE loU
1bc CODvolution of two fiDite-duration sequences.

I' .

a.

1 2 3 4

DISCRETETIME DESCRIYnON OF SIGNALS AND SYSTEMS

3S

thIn)}
{x(n))

. rl I..
012

.... I

IIII I I

o1

II

{h(k)l
{x(n-k)}

.IJ

. . IIIIIII [

t"\N-C-

I I I

a:

s:

+I:

I ...

for n = 0

{x(n-k)l

. .1. .1. . . 1. ....._._.

'_00_ _. _

-2-1012

-\ 0 \ 2

for n = 1

for n = 2 ....

II I .

-10 1 2

{y(n)l

..... 1111 J1111 ,.;.

012

FIGURE 2-13
The convolution of a finite-duration sequence with an infinite-duration sequence.

finite duration of {x(n}}:

~

y(n) = .t ~
x(k) h(n - k) = l-O
~ (k + l)a"-tu(n - k)
__ oo

=a"u(n) + 2a"-lu(n -1) + 3a"-2u (n -

2)

Let us define a finite-duration sequence as one that contains only a finite

number of nonzero elements. When both {h(n)} and {x(n)} are finite-duration
sequences, then the convolution produces a finite-duration sequence, as
illustrated by Example 2.6. It is left as an exercise to show that if the input
sequence {x(n)} contains Nx elements and the unit-sample response sequence
{h(n)} contains NH elements, then the output sequence {y(n)} contains
Ny =N x + NH - 1 elements.

36

INfRODUcnON TO DIGITAL SIGNAL PROCESSING

A geometric series is a series in which consecutive elements differ by a

constant ratio. Such a series can be written in the form
x(n)

= r",

(2.31)

where r is a constant and Nl and N2 are any two numbers. In evaluating the
sum of products in the convolution, we commonly encounter sequences that
form such a series. The sum can then be easily computed by employing the
following two geometric sum formulas. To derive these formulas, let us
consider the following Taylor series expansion: Given a complex number c,
where Icl < 1, then

--=1+c+c2+c3 + ...
l-c

= ,,~
L c"

(2.32)

Applying this result in the reverse direction yields the infinite geometric sum
formula: Given a complex number c, where Icl < 1, then

L cit =1z

(2.33)
1- c
We can also compute the sum of a finite number of elements in a
geometric series. Let us consider the following sum
It-O

1+c+~+ +~-l=

N-l

,,0

C"

(2.34)

The sum of this finite-duration sequence can be expressed as the difference

between the sums of two infinite-duration sequences as

N-l
...
...
LC"=LC"-LC"
11==0

n-O

...

n-N

...

=Lc"-cNLc"
It-O

(2.35)

,,=0

Applying the infinite geometric sum formula to the two summations, we obtain
the desired finite geometric sum formula: Given a complex number c, then

N-l
l-cN
LC"=-It.O
1-c

(2.36)

While the infinite geometric sum formula requires that the magnitude of c
be strictly less than unity, the finite geometric sum is valid for any value of c.
EumpIe 2.8. Convolution of two lnfiIIite-dantion seqnences. Let
h(n) =a"u(n)
x(n) = b"u(n)

for all n,
for all n

DISCRETE-TIME DESCRIPTION OF SIGNALS AND SYSTEMS

h(n)=a"u(n) for all n

. .] rl~

.I(n)=/J'tu(n) for all n

11

to . . . .

.... IIIII

)0

-I 0 1 2

bb

-I 0 1 2 3

h(k)=t!u(k) for all k

. . ..'I rri

(.I(n-k)}
n=-1

It

iII: ..
JlIII1 r..

It

... 1JIII

-I 0

n=O

y(-I)=O
)0

It

y(O) = 1
)0

IlIlllH.

n=l

y(l)=a+b
)0

oI

n=2

)0

It

... I IIIII~H
1__ .

y(2)=a2+ab+b2
)0

012

It

y(n)

.I111I1I

)0

-10123

FIGURE Z-14
The convolution of two infinite-duration sequences.

as shown in Fig. 2.14. Then the output sequence values are equal to

y(n) =

L-

" akb"-k
aku(k)b"-ku(n - k) = L

k---

= b"

k-O

"

L (alb)"

k-O

The finite geometric sum formula can be applied to yield

r'

. ( )=b"I-(alb
n
t-(alb)

37

38

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

2.6 STABILITY CRITERION FOR

DISCRETETIME SYSTEMS
For practical applications, we ace interested in designing only stable digit8I
filters, or those whose outputs do not become infinite. In this section, we
define a stability criterion and provide a necessary and sufficient condition for a
filter to meet this condition. In later chapters, we will describe alternate means
of determining whether or not a digital filter is stable.
Let us define a bounded sequence to be one in which the absolute value
of every element is less than some finite number M, i.e. {x(n}} is bounded if
Ix(n)1 <M for all n, where M is some finite number. Then, a system is said to
be stable when every possible bounded input sequence produces a bounded
output sequence. Since the output values of a digital filter should not exceed
the number range of the computer, it is appropriate to consider this bounded
output criterion.
STABll.1TY CRITERION. An LTI system is stable if, and only if, the stability
factor, denoted by S, and defined by

...

S=

L Ih(k)1
k=-CIO

(2.37)

is finite, i.e. S < 00.

The stability of a digital filter is expressed in terms of the absolute values
of its unit-sample response. We now show that this criterion is both sufficient
and necessary. To prove that it is sufficient, we show that if S is finite, then the
system is stable. Let {x(n}} be a bounded input sequence. We must show that
the output is bounded when S is finite. Taking the absolute value of the
output, expressed by the convolution equation, we have
ly(n)1 =

L~", h(k) x(n -

k)1

(2.38)

Recall that the magnitude of the sum of terms is less than or equal to the sum
of the magnitudes. Hence,
ly(n)1 ~

'"

Ih(k)x(n - k)1

(2.39)

k=-oo

Finally, since all the input values are bounded, say by M, we have for all n:

ly(n)1 ~ M

'"

Ih(k)1

k=-oo

or
ly(n)I~MS

(2.40)

Hence, since both M and S are finite, the output is also bounded.
To show that it is necessary for S to be finite for the filter to be stable, we
present a counter-example. We need to finq a bounded input sequence that

DISCRETETIME DESCRIPfION OF SIGNALS AND SYSTEMS

39

produces an unbounded output. For our input sequence, let us choose

x(n)

= {~( -n )/lh( -n)1

for all n for which h( -n)

otherwise

'* 0

(2.41)

This sequence consists of plus and minus ones and is obviously bounded. Let
us now consider the output value for n = 0, or y(O). From the convolution
equation, we have

...

y(O) =

2:

...

h(k)x( -k) =

k--~

h2(k)/lh(k)1

k=-~

...

2:

(2.42)

Ih(k)I=S

k==-oc

Hence, for y(O) to be bounded, it is necessary for S to be finite. We have thus

proved the necessary and sufficient conditions. The following examples
illustrate the application of the stability criterion.
Eumple %.9. StabiHty of the Ceaeralized J-sample avenger. Let us generalize
the 3-sample averager by letting the coefficients be arbitrary constants b -1> b o and
b l . The output is then equal to
y(n)

= b_Ix(n + 1) + boX(n) + blx(n -1)

To test whether this filter is stable, we compute the value of S:

. --

s= ~

Ih(n)1 = Ib_11 + Ibol + Ibll

Hence, S is finite for finite values of b_ .. b o and b l .

Example 2.9 can be easily extended to show that any filter with a
unit-sample response having a finite number of nonzero elements is always
stable.
Eumple 1.10. Stability of first-onler reamive &her. Let
y(n)

=ay(n -1) +x(n)

Recall that h(n) =a"u(n), for all n. Checking for stability, we must calculate

s = ,,_-OD
~ Ih(n)1 = ~ lal"
11-0

It is obvious that S is unbounded for lal2: 1, since then each term in the series is
2:1. For lal < 1, we can aIfply the infinite geometric sum formula, to find

1
S=--

1-a

forlal<1

Since S is finite for lal < 1, the system is stable for lal < 1.

40

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

Example Z.ll. Stability of seoond-order filter. Let

yen) = 2r cos(wo) yen - 1) -,z yen - 2) + x(n) - r cos(wo) x(n -1)

To test the stability of this filter, we must first determine its unit-sample response
and then compute the value of S. To determine {hen)}, we assume zero initial
conditions, or yen) = 0, for n <0 and let the input sequence be {den)}. We then
get the following output values.

yeO) = 1
y(l)

=2r cos(wo) -

y(2) =

r cos(wo)

=r cos(wo)

2r'- cos ( wo) - ,z = ,z(2 co~( wo) - 1) = r2 cos(2wo)

2

y(n) = r" cos(nwo)

Hence,

h(n) = ", cos(nwo) u(n)

Checking for stability, we have
-

00

~ Ih(II)1 = ~ I'" cos(lIwo)1 s ~o Irl" = l-Irl

for Irl < 1

Hence, the filter is stable for Irl < 1. The inequality of the sums in the
above stability calCUlation comes about because the magnitude of the cosine
function is less than or equal to unity.

2.7 CAUSALn:Y CRITERION FOR

DISCRETE-11ME SYSTEMS
A system is said to be causal when the response of the system does not precede
the excitation; All physical systems are causal in that they do not react until a
stimulus is applied. An LTI system is causal if, and only if,
hen) =0

for n <0

(2.43)

:his is true because, by definition, {hen)} is the response of a system to a

unit-sample sequence, whose nonzero element occurs at n = O. For a causal
system, the output cannot have nonzero values for n < 0, the time before the
nonzero-valued element is applied.
Causal filters are typically employed in applications in which the data
samples are processed as they are received. Qearly, in this case we cannot
have access to future sample values. In this book, we will not restrict ourselves
to causal digital filters. When filters are used to process data for which the
entire data record is available at the start of processing, there is no restriction
against looking ahead to future values. One example of such data is that
obtained from images. As described in the previous chapter, this look-ahead
feature is implemented by the advance, or negative delay, component of the
digital filter. For noncausal filters, {hen)} has nonzero elements for n <0, as
illustrated with the 3-sample averager in Example 2.3. Since the unit-sample

DISCRETE-TIME DESCRIPTION OF SIGNALS AND SYSTEMS

41

response of the first-order recursive filter in Example 2-4 obeys the condition
given in Eq. (2.43), it is a causal filter.

2.8 LINEAR CONSTANT-COEFFICIENT

DIFFERENCE EQUATIONS
In this section, we formalize the notation that we have been using above for
describing the time-domain behavior of digital filters. In general, the output of
a finite-order LTI system at time n can be expressed as a linear combination of
the inputs and outputs in the following form:
M

Np

y(n) = ~ a/cy(n-k)+ ~ b/cx(n-k)

k=l

(2.44)

k=-NF

In words, the current output y(n) is equal to the sum of past outputs, from
y(n -1) to y(n - M), which are scaled by the delay-dependent feedback
coefficients ak, plus the sum of future, present and past inputs, which are
scaled by delay-dependent feedforward coefficients bk The output value at
time n is determined by the input values from N F samples in the future
x(n +NF), through the current value x(n), to the sample N p time units in the
past x(n - Np ). For a causal filter NF sO. When NF> 0, then the future values
of the input determine the current output value, and the filter is noncausal.
EumpIe 2.12. Difference equation of a simple filter. We can combine the
equations of the nonrecursive 3-sample averager and the first-order recursive
filter, to get
y(n) = ay(n -1) + lx(n + 1) + lx(n) + lx(n -1)
This filter has both a recursive and a nonrecursive part and is also noncausal.

The difference equation is useful not only for analyzing digital filters but
can be employed for implementing them. A digital filter can be implemented
from the general difference equation by performing the following four steps, as
shown in Fig. 2.15.

Step 1. Two points are drawn, one corresponding to the current input
x(n) and the other corresponding to the current output y(n).
Step 2. To gain access to the past values of the input and output
sequ~nces, delay elements are connected to the output and input points.
Future values of the input sequence are obtained by connecting advances to
the input point.
Step 3. Multipliers are connected to the outputs of the delay elements to
produce the required products. Multipliers with gains of unity need not be
included, but replaced by a simple wire.
Step 4. Finally, the digital filter is accomplished by connecting the
outputs of the multipliers to two adders. The first generates the sum of

42

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

STEP I
x(n)_

y(n)'

STEP 2

..+N,)--+

x(n+2)~

,..---.. y(n)
x(n)---~

I
I
I

~
--..T~y(n-M}
STEP 3

y(n)
x(n)--+----i

ftGURE 2-15
Steps to produce a digital filter structure from the general difference equation. This filter structure

will be implemented in the computer programs.

DISCRETE-TIME DESCRIPTION OF SIGNALS AND SYSTEMS

43

STEP 4

X(II)

_-.------1 >-..--,~

}-----_--oY(II)

products of the inputs and the feedforward coefficients, and the other that of
the outputs and the feedback coefficients.
The following example illustrates the application of these four steps.

_tal

Example 1.13. lmpIemeatatioD of the seeoad-order

filter fro. the
dlrereace eqaatioa. The filter implementation procedure from the difference
equation given in Example 2.11 is shown in Fig. 2.16.

X(II)_-.......----~

}------.---y(II)

FIGURE 2-16
Digital filter structure that implements the difference equation given by
y(n) = 2r cos(ClJo) yen -1)

_,2 yen -

2) + x (n) - , cos(wo)x(n -1)

44

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

x(n) o---+----i

.>--30(

J..------__-- y(n)

ftGURE 1-17
General digital filter structure implemented in the computer programs.

The general digital filter structure shown in Fig. 2.17 is convenient for
describing the logic behind the computer programs described below. However,
it is not the most efficient structure in terms of the number of delays that are
used. In a later chapter, we will consider other filter structures that have a
smaller number of delays but for which the corresponding computer programs
are more complicated. Conceptually, however, these more memory-efficient
programs perform the same types of operations as the simple programs
described below.

2.9 SUGGESTED PROGRAMMING

STYLE
Befqre describing the programs, we provide the motivation for the particular
programming style suggested below. The goal of the programs is to yield the
greatest intuitive understanding of the digital filter operation with the least
programming effort. Techniques to improve the efficiency of these programs
are suggested throughout the text. The logical steps in the operation of a
digital filter are best illustrated by writing simple subprograms that perform the
primitive tasks that are most commonly employed in signal processing. These
simple subprograms are then combined to form the digital filter program.
To provide a FORTRAN example of the recommended programming
style, one common task is to initialize an array to zero. Even though this can
be done quite easily in a simple loop in the main proram, the programming

DlSCRETETIME DESCRIPTION OF SIGNALS AND SYSTEMS

4S

style suggested above would use the following subroutine:

SUBROUTINE NULL(X,NX)
DIMENSION X(NX)
DO 1 I=1,NX
X(I)=O.
RETURN
END

An array B(N), for 1:s N:s NB, can be zeroed in the main program with the
following statement:

CALL NULL(B,NB)
Two additional subroutines, described below, compute the sum of
products of two arrays and implement a shift register. Use of such simple
subroutines will simplify the main program by making it more readable, which
is important for debugging and trouble shooting. This programming style also
allows the student to reason in terms of these primitive tasks.
In trying to explain the programs in a manner that is not specific to a
particular programming language, a problem immediately arises with aI'tay
indexing. The various computer languages used for signal processing can be
divided into two classes: those that allow arrays to be indexed starting with
zero, such as BASIC and C, and those that do not, such as FORTRAN and
PASCAL. In the first set, x(O) is the first element of an array, in the second.
the first element is x(l). Since all languages can start indexing from 1, we will
use this convention in the programs below. For computer languages that allow
the array index to start from zero, a slight modification in the indexing of the
programs should be made. The advantage of starting the array with x(O) is that
it then conforms to the notation used for sequen~es in the analytic sections of
this book, i.e. x(n), for 0 :s n :s N - 1.

2.10 WRITING A DIGITAL FILTER

PROGRAM
In this section, we describe the cOJ1lputer programs to implement a digital filter
from a difference equation. These programs will be used in the following
chapters for analyzing and implementing digital filters to meet some desired
specification. To implement a digital filter as a flexible program, it is most
convenient to write separate subprograms for the nonrecursive part, to be
called NONREC, and the recursive part, to be called REC. The general filter
can then be implemented by first using NONREC followed by REC, as
described below.
First, it is useful to establish the notation for the arrays to be used in the
programs. We are typically given an input array of real values, denoted by

46

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

X(N), for 1 s; N s; NX, where NX is a known value. This array represents the
input data to be processed. In this book, the most common input will be the
unit-sample sequence. We are usually asked to design a digital filter that is
specified by an array of feedback coefficients ACOEF(N), for 1 s; N s; NA, to
be used in REC and an array of feedforward coefficients BCOEF(N), for
1 s; N s; NB, to be used in NONREC.
Since we are implementing nonrecursive and recursive filters separately,
it is helpful to define two output arrays. The output of a nonrecursive filter is
denoted by YN (N) and will have a duration NY = NX + NB - 1. This duration
follows from the convolution of two finite-duration sequences. The output
array of the recursive filter is denoted by YR(N) and has a duration NY. For
the recursive filter, however, the value of NY must be specified. Because of
the feedback loop, the output of an ideal recursive filter contains an infinite
number of non-zero elements. However, for a stable filter, the output values
eventually approach zero when the input has a finite duration. Hence, there
will be only a finite number of output elements whose amplitudes are
significantly greater than zero. For all the filters in this book, NY will be less
than or equal to 128.
Since most programming languages require that the size of all arrays be
specified at the beginning of the program, a length of 128 floating-point, or
Teal, elements for all arrays will be adequate for most of the projects in this
book. In the descriptions of the variables below, Teal is used to specify a
floating-point variable.
We first consider the implementation of a causal filter, for which NF=O
in Eq. (2.44), and then describe the simple modification that can be added to
implement a noncausal filter.
PROGRAMMING A CAUSAL FILTER. The general causal filter is implemented by combining two causal components: a strictly nonTecuTsive causal
part with a strictly TecuTsive part.
The general difference equation (2.44) of a causal nonTecuTsive filter, for
which NF = 0 and ak = 0 for all k reduces to
Np

yen) =

bkx(n - k)

(2.45)

k=O

The feedforward coefficients bk for 0 s; k s; Nr are stored in BCOEF(N) for

1 s; N s; NB (NB = Np + 1). The value of bo is stored in BCOEF(l), b l in
BCOEF(2), etc. Notice that Eq. (2.45) has the form of the following
convolution
Np

y(n) =

L h(k)x(n -k)

(2.46)

k=O

By computing the unit-sample response of the nonrecursive filter, we find that

it has a finite duration and is equal to the feedforward coefficient sequence, or
h(k) = bk

for - NF s; k

s; Nr

(2.47)

DISCRETETIME DESCRIPTION OF SIGNALS AND SYSTEMS

47

Hence, a nonrecursive filter can be implemented directly from its unit-sample

response. For the causal filter considered here, Np = O.
The difference equation for a strictly recursive filter, whic~ employs the
past outputs and only the current input value, for which bk = 0 for k =1= 0,
reduces to
y(n) =

L aky(n-k)+x(n)

(2.48)

k-l

The feedback coefficients ak for 1 s k s M will be stored in ACOEF(N) for

1 s N s NA, where NA = M. The value of a 1 is stored in ACOEF(1), az in
ACOEF(2), etc.
In both Eqs. (2.4S) and (2.48), a sum-of-products calculation employing
an array of coefficients and an array of data is required. We implement this
primitive task as a separate subprogram, called SUMP (sum of products), that
can be used in both NONREC and REC. To simplify the programming logic,
we will use two auxiliary buffer arrays. The first, denoted XBUF(N), for
1 s N s NB, stores the current and past NB values of {X(N)}. Only NB values
are needed since that is the number of coefficients in BCOEF(N). The second
buffer array, denoted YBUF(N), for 1 s N:5 NA, stores the past NA values of
{Y(N)}. A second subprogram, called SHFTRG (shift register), acts as a shift
register to maintain the proper time sequence of values in these arrays.

SUMP Subprogram. The SUMP subroutine computes the sum of products of

two arrays, a coefficient array, denoted COEF(N), for 1sN:5NC, and a
buffer array BUF(N), of the same size, to produce the output value SUM. The
call to the subprogram is then

SUMP(COEF,BUF,NC,SUM)
where
COEF
BUF
NC
SUM

(real input array) is an array of coefficients;

(real input array) is an array of data values;
(integer input) is the duration of the COEF array;
(real output) is the sum of products equal to
NC

SUM = L COEF(I)

* BUF(I)

(2.49)

1-1

as shown in Fig. 2.18. When SUMP is used in NONREC, BUF is equal to

XBUF, COEF corresponds to BCOEF and NC is equal to NB (= Np + 1).
When used in REC, BUF is equal to YBUF, COEF corresponds to ACOEF
and NC is equal to NA ( = M).

SIIFI'RG Subroutine. To update the buffer arrays, we implement a shift

register routine, called SHFTRG, that shifts the contents of the buffer array by
one element at each iteration of the filter operation. The shift is performed

48

INTRODUCi10N TO DIGITAL SIGNAL PROCESSING

COEFFICIENT ARRAY

BUFFER ARRAY

COEF(I)

BUF(I)

COEF(2)

BUF(2)

COEF (3)

BUF(3)

COEF(NC)

BUF(NC)

nGURE 218
Operation performed by the SUMP
routine.

such that the oldest element in the buffer is lost and the current value of the
time sequence is put into the first element of the buffer. That is, at time L,
X(L) is put into XBUF(1) and only the current and past NB -1 values are
retained in the buffer.
In general, given a value VAL and buffer array BUF(N), for 1::5 N ::5
NBUF, we want to write the subroutine called
SHFTRG(VAL,BUF,NBUF)
where
VAL
(real input) is the value to be inserted into the buffer;
BUF
(real input and output array) is the buffer array;
NBUF
(integer input) is the duration of the buffer array.
This subroutine shifts BUF(K) into BUF(K + 1), for 1::5 K::5 NBUF, and puts
VAL into BUF(l), as shown in Fig. 2.19. When used in NONREC, the buffer
BUF is equal to XBUF, NBUF is equal to NB, and at time L, VAL is equal to
X(L). When used in REC, BUF is equal to YBUF, NBUF is equal to NA and
at time L, VAL is equal to YR(L). We now describe how to use these
subroutines to implement the nonrecursive filter subprogram NONREC and
the recursive filter filter subprogram REC.
NONREC Subprogram. For the nonrecursive filter, the input array X(N), for
1::5 N::5 NX and the feedforward coefficient array BCOEF(N), for 1::5 N::5 NB
BEFORE SHFfRG
CALL

AFfER SHFfRG
CALL

VAL
BUF(I)

BUF(I)

BUF(2)

BUF(2)

BUF(NC-I)

BUF(NC-I)

BUF(NC)

BUF(NC)

nGURE 2-19
Operation performed by the SHFrRG
routine.

DISCRETE-TIME DESCRIPTION OF SIGNALS AND SYSTEMS

49

(NB = Np + 1), are specified and the output array YN(N), for 1:5 N:5 NX +
NB - 1, is to be generated_ The nonrecursive filter subroutine is denoted by

NONREC(X,NX,BCOEF,NB,YN)
where
X
NX
BCOEF
NB
YN

(real input array) is the input array to the nonrecursive filter;

(integer input) is the duration of the input array;
(real input array) is the feedforward coefficient array;
(integer input) is the duration of the coefficient array;
(real output array) is the output array of duration NY = NX +
NB-L
The procedure to be performed within NONREC is indicated in the logic
diagram shown in Fig_ 2_20. The NONREC routine is accomplished using the
three basic steps described below.
Step 1. The buffer array XBUF(N), for 1:5 N:5 NB, is defined and
initialized to zero.
Step 2. In a loop with index I, for I equal from one to NX, VAL is set
equal to the Ith element of the input, or X(I). VAL is then entered into the
shift register XBUF. The sum of products of XBUF and BCOEF are
computed to produce SUM. The Ith value of the output sequence for the
nonrecursive filter, YN(I), is then set equal to SUM.
Step 3. When I> NX, all the elements of X(N) have been used. VAL is
then set to zero and then entered into the buffer XBUF. The sum of products
is computed and the result SUM becomes the value of the output YN(I). This
loop is executed until no more elements of the input sequence X(N) remain in
the buffer, or until I > NX + NB - 1.

After this routine is completed, the nonrecursive filter operation has been
accomplished and the output values will reside in the YN array. The
combination of the SHFTRG and SUMP routines results in a convolution
operation.
REC Subroutine. For the recursive filter, the input array X(N), for 1:5 N:5

NX, and the feedback coefficient array ACOEF(N), for 1:5 N:5 NA (NA =
M), are specified and the output array YR(N) , for 1:5 N:5 NY, is to be
generated. Unlike the nonrecursive filter, the duration of a recursive filter can
be infinite because of the feedback. Hence, the value of NY must be
sufficiently large so that the important features of the output are observed. In
the case of determining the unit-sample response, the value of NY must be
chosen large enough such that the values of h(n) for n > NY are approximately
zero. In producing a plot of the sequence, it is sufficient that the omitted
magnitudes be smaller than the resolution of the plotting device. Typical
values for NY used "in the projects will be 55 and will not exceed 128.

50

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

FIGURE 2-20
Logic diagram for the NONREC filter routine. The values for the input array X(N), the size of the
array NX, the feed-forward coefficient array BCOEF(N), and the size of the coefficient array NB
are given; the output array of the nonrecursive filter YN(N) is to be computed.

DISCRETE-TIME DESCRImON OF SIGNALS AND SYSTEMS

51

The recursive filter is implemented by the subroutine

REC(X,NX,ACOEF,NA,YR,NY)
where
X
(real input array) is the input array to the recursive filter;
NX
(integer input) is the duration of the input array;
ACOEF
(real input array) is the feedback coefficient array;
NA
(integer input) is the duration of the coefficient array;
YR
(real output array) is the output array;
NY
(integer input) is the duration of the output array.
The logic diagram for the REC routine is shown in Fig. 2.21. Within REC, the
following steps are performed:
Step 1. The buffer array YBUF(N), for 1:5 N:5 NA, is defined and
initialized to zero.
Step 2. In a loop with index I, for 1:5 1:5 NX, the output at time I,
YR(I), is equal to the sum of the input X(I) and SUM, the product of the past
outputs stored in YBUF and ACOEF. For 1= 1, SUM is equal to zero. VAL
is then set equal to the output YR(I) and is sequenced into YBUF.
Step 3. When I > NX, all the elements of X(N) have been used. At
that point the output YR(I) is then determined from only the past output
values. The recursive filter operation is accomplished after the desired number
of output points NY have been computed.
GENERAL FlLTER IMPLEMENfATION. The general filter can be imple-

mented by calling the NONREC followed by the REC routines, as shown in

Fig. 2.22. The input sequence X(N), for 1:5 N:5 NX, to NONREC produces
the output sequence YN(N) , for 1:5 N :5 NX + NB - 1. This sequence then
becomes the input sequence to the REC routine. The REC routin~ will then
produce the output sequence YR(N) , for 1:5 N :5 NY. Usually NY will be
greater than NX + NB - 1. The array YR is then the output of the general
filter.
If the REC routine is called first and then followed by NONREC, care
must be taken in choosing the value of NY, the duration of the output
sequence YR produced by REC. This sequence then becomes the input to
NONREC. The output of NONREC, YN, then contains NY + NB - 1
elements, a duration that must not exceed the dimension of YN.
NONCAUSAL FILTER IMPLEMENfAnON. By definition, a filter is noncausal

when future values of the input sequence are used in the computation of tb,e
current output value. Noncausal filters can be identified easily from the
difference equation by Np being greater than zero. Since arrays with negative
indices are not generally allowed in programming languages, the simplest way
to implement a noacausal filter is to change the difference equation in such a

52

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

~{

nGURE 2-21
Logic diagram for the REC filter
routine. The values for the input
array X(N), the size of the array NX,
the feedback coefficient array
ACOEF(N), and the size of the
coefficient array NA are given; the
output array of the recursive filter
YR(N) is to be computed.

way as to produce a causal filter and then shift the final output array into the
future by N p index units, as described below. Using this procedure, the causal

filter can then be implemented in the manner described above, and only an
additional shift routine is needed, to obtain the noncausal result.
The noncausal property of the filter is due to the NF feedforward
coefficients having negative index values in the nonrecursive part of the filter.
Hence, a noncausal filter can be made causal by shifting the coefficient

DISCRETETIME DESCRIPTION OF SIGNALS AND SYSTEMS

YN(N)
lo;;No;;NX+NB)

NONREC
X(N)
for lo;;No;;N

:x"

53

B.E.C
YR(N)

ACOEF(N) - 1o;;No;;NY
lo;;No;;M

BCOEF(N)
lo;;No;;NB

FIGURE 2-U
Combining NONREC and REC to form the general filter having both nonrecursive and recursive
parts.

sequence {b k } to produce {ba, where

as shown in Fig. 2.23. This new set of coefficients then generates the causal
output, denoted by {y'(n)}, whose elements are equal to
M

y'(n)=

Np+Np

L aky'(n-k)+ L
k-l

b;'x(n-k)

(2.50)

k~O

NON CAUSAL FILTER

CAUSAL FILTER

Noncausal
y'(n)

part

zen) c---~---l

~~-~

(y(n)}

,... rII"r

(ytn)} =(y(n.NF)}

, .. 1 I lour+ ~n

o
Ji1GURE 2-23
Transforming a noncausal filter into a causal filter.

NF

Np

54

INTRODUCfION TO DIGITAL SIGNAL PROCESSING

CAUSAL FILTER OUTPUT

{/(n))

oorIIIlhr. ..
0

~7
I

NON CAUSAL (y(n)}

FILTER OUTPUT

1 2 NF NF+Np

Jr1 rT

_ .......~T';-L-L-L-!-'L.JWW'L..o...........o--_ _ _ _ _.".:>

- NF

Np

FIGURE 2-24
The noncausal filter output is
obtained by shifting the output
of the causal filter to the left.

This causal tiltf!r can be implemented by using the routines described above.
The values of {y'(n)} are exactly the same as those of {y(n)}, but only shifted
by NF index values. To obtain the noncausal output sequence {y(n)}, a time
shift to the future is required, i.e.
y(n) = y'(n + NF ),

for -NF ::; n::; NY - N F

as shown in Fig. 2.24. This time shift is performed by the subprogram

SHIFT(Y,NY,NF)
where
y

(real input and output array) is the array to be shifted;

(integer input) in the duration of the array;
(integer input) is the number of elements that the input array is to
NF
be shifted to produce the output.
After the shift is performed, the first NF samples of the input array are lost
and the last NF samples of the output array are set to zero, as shown in Fig.
2.25.
NY

AFTERSHlFT
CALL

BEFORE SHIFT
CALL
Y(l)

Y(l)

Y(2)

Y(2)

Y(NF )

Y(Ny-Nr-l)

Y(NF + 1)
Y(NF +2)

Y(Ny-NF)
Y(Ny-NF+l)

Y(Ny-l)

Y(Ny -1)

Y(Ny)

Y(Ny)
0

FIGURE l-2S
Operation performed by the SHIFf
routine.

DISCRETE-TIME DESCRIPTION OF SIGNALS AND SYSTEMS

55

In programming this shift, we still have a problem, because the shifted

array may require negative index values_ To avoid this problem and still
observe the noncausal components, if any, in the filter unit-sample response,
the input to the filter can be the delayed unit-sample sequence, whose nonzero
component occurs at N = NF Using this as the input sequence, the noncausal
components of the filter can then be observed in the output sequence YN(l)
through YN(NF -1).

2.11 SUMMARY
This chapter has presented the time-domain analysis techniques for describing
linear time-invariant digital filters and discrete-time systems, from basic
definitions to programming techniques. A particularly useful sequence that
describes the behavior of a digital filter is its unit-sample response, since from
it we can determine whether a filter is stable and whether it is causal. Further,
for any input sequence to a digital filter, we can determine the output sequence
by performing convolution with the unit-sample response.
An alternate time-domain description of a digital filter is the linear
difference equation. Yhe unit-sample response can be determined from the
difference equation by simply making the input equal to the unit-sample
sequence and computing the output sequence. Of practical importance, it was
found to be a rather straightforward task to implement a digital filter structure
from the difference equation. If the unit-sample response has a finite duration,
a nonrecursive filter can be implemented by identifying the feedforward
coefficients with the corresponding elements of the unit-sample response.
The steps for implementing a digital filter as a computer program were
presented. By writing simple routines to perform a sum-of-products computation and a shift-register operation, the implementation of nonrecursive and
recursive filters becomes conceptually easy to grasp. The general filter
structure containing both nonrecursive and recursive parts is implemente.d by
employing both routines. Noncausal filters can be implemented by employing
the causal filter routines and applying a subprogram that shifts the output
sequence "into the future".

PROBLEMS
2.1. Express the following in complex vector notation:
(a) 1 + e- j8
(b) 1- a e- j8
(c) 1 + 2e- j8 + e-j28
(d) 1 + ja
1 + jb

1 + a eJ8
(e) 1 + b eJ8
(f) (a

+ jb )112

56

INfRODUCfION TO DIGITAL SIGNAL PROCESSING

2.2. Express the following in real plus imaginary notation:

(a) 2cos(0)e- Jw
(b) _I_
1+ ja
1
(c) 1 + a ej8

(d) 1 + e

J; + e J28

(e) (20 + 16 cos 0)112 exp[-j arctan( -sin 0 )]

2+cos~

2.3. Use the Euler identities to derive the formula for the cosine of the sum of two
angles:
R) cos( a + P) + cos( 11'- P)
cos( 11'+ P =
2
2.4. Demonstrate that the first-order recursive filter is linear. (Hint: Express G{x(n)}
as an infinite sum of scaled terms containing current and past values of x(n).
2.5. Evaluate the following summations and, give ,the result in complex vector
notation:
for Irl < 1
n=O

N-)

(b)

L e-

jwn

n=O
N-)

(c)

L cos(wn)

n=O
N-I

(d)

L an e-

jwn

2.6. This problem considers the case of determining the unit-sample response of a
filter having nonzero initial conditions. This condition can occur when the delay
element contents are not initialized to zero.
(a) The 3-sample averager. Determine {h(n,O)} when the content of the delay
element is equal to 2 and that of the advance is equal to 1 at n = O.
(b) The first-order recursive filter. Determine {h(n,O)} when the content of the
delay element is equal to 3 at n = O.
2.7. Given two finite-duration causal sequences, {x(n)} containing Nx elements and
{h(n)} containing NH elements, show that {y(n)} = {x(n)} * {h(n)} produces a
causal, finite-duration sequence containing Ny = Nx + NH -1 elements.
2.S. Given the following two sequences, determine an analytic expression for
{y(n)} = {x(n)} * {h(n)}. Let
x(n) = {~

forOsnsN-1
otherwise

{~

for OSnsN-1
otherwise

and
h(n) =

DISCRETE-TIME DESCRIPTION OF SIGNALS AND SYSTEMS

2.9.
2.10.

2.11.

2.U.

2.13.

2.14.

2.15.

2.16.
2.17.
2.18.

2.19.

57

Hint: First try N = 2 and then N = 3, to obtain the general analytic form of the
sequence.
Prove that any realizable filter having a finite number of elements in its
unit-sample response is always stable.
For the following difference equations, determine and sketch the unit-sample
response sequence and implement the digital filter structure:
(a) y(n)=x(n)-x(n-N)
(b) y(n)=ay(n-1)+x(n)+x(n-1)
Let a = 0.9.
For the digital filter structures shown in Fig. n.ll, determine the difference
equation, compute and sketch the unit-sample response.
The steady-state condition of a discrete-time system is defined as that in which the
output value does not change with time, i.e. y(n + 1) = y(n) for n very large.
Determine this steady-state value when the input sequence is a constant-valued
step sequence, x(n) = Ku(n), for the following system:

y(n) = ay(n -1) +x(n) + x(n -1)

Let a = 0.9.
Implement the digital filter that generates cos( won), for n ~ O. (Hint: see
Example 2.11).
Demonstrate that the sequence {x(n)} = {sin(wn)} produces identical values
when w = Wo and w = Wo + 2mJr, where m is an integer. Sketch the results for a
particular value of wo, with m = 0 and 1.
Demonstrate that the convolution operation is commutative by performing the
convolution given in Example 2.8 in the reverse order to obtain the same result.
Restructure the logic diagram of the nonrecursive filter such that only one loop is
required.
Restructure the logic diagram of the recursive filter such that only one loop is
required.
For each of the following transformations, determine whether the system is
stable, causal, linear and shift-invariant:
(a) Gt{x(n)} = ax(n - no) + bx(n - nt), for constants a, b, no and nl
(b) G2 {x(n)} =x~n)x(n - no)
(c) G3 {x(n)} =x(2n)
That a system is unstable does not imply that the output at time n, or y(n),
approaches infinity for every input sequence. Consider the system whose
unit-sample response is given by

h(n)={~

for n ~O
otherwise

(a) Show that this system is unstable.

(b) Let the input sequence be given by
x(n) = eJwn

for all n

Using the convolution equation, determine the values of w for which y(n) is
finite, and the values of w for which y(n) is infinite.

58

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

(a)

x(n) - ___---~

~---___oy(n)

(b)

x(n)c

(c)

x(n)

y(n)

y(n)

(d)

y(n)

x(n)

(e)

x(n)

y(n)

OGURE Pl-ll
Digital filter structures.

2.20. Find the output sequence produced for the system having the unit-sample
response given by

h(n)=H

forn=O
forn=4
otherwise

DISCRETETlME DESCRIPTION OF SIGNALS AND SYSTEMS

S9

when the input sequence is given by

forn<:!O
otherwise
2.21. A common input to an LTI system used in control situations is the unit-step
sequence, {u(n)}. The response of the system is the unit-step response, denoted
by {g(n)}. Determine the unit-step response of the first-order recursive system,
whose unit-sample response is given by h(n) = a"u(n).
x(n) =

{~

COMPUTER PROJECTS
Preface
Most of the projects contain the following parts:
Object, that is a concise statement of the purpose of the project.
Analytic results, which describe the required analysis to be performed for
later comparison with the computer results. In some cases, references are
made to a problem at the end of the chapter.
Computer results, which describe the programs to be written and the
computations to be performed.

Definitions for the following terms will be helpful for performing the projects.

AnalyticaUy determine means to apply the appropriate mathematical

formulas and evaluate the expression. All of the summations in the analytic
expressions should be evaluated and complex quantities should be expressed in
complex vector notation. The analysis is usually not difficult when requested. If
the student finds that the procedure is becoming very complicated, it means
that either a mistake has been made, or that a simpler approach can be taken.
In this case, the programming should be performed first to provide the answer
towards which the analysis should lead.
Sketch means to provide a graphical interpretation of the analytic results.
If the results were not analytically determined, the graphical interpretation is
to indicate a qualitative understanding of the operation being discussed. In all
cases, the time- or frequency-domain scales and landmarks are to be indicated
on the sketches. In many cases, only relative amplitudes are necessary.
Compute means that the procedure is to be performed numerically by the
computer and compared with the analytic results.
Plot means that the computed data are to be presented in graphical form
and compared with the sketches obtained from analysis. The graphs produced
by the computer should be labeled by hand to indicate the correct time scale
(n) or frequency scale (co). The amplitudes and abscissa locations of features

60

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

of interest in the plots, such as peaks, valleys or discontinuities, should be

noted and related to the sketches.
The project report should be submitted with the material for each project
grouped together. The analysis should be presented first, followed by the
corresponding computer plot for easy comparison. Lists of numerical printouts
should not be included. The programs used in the project should be appended
to the end of the report, to facilitate evaluation of the programming style and
for easy reference in the future.
2.1. Sinusoidal sequences
Object. Demonstrate that the values of sinusoidal sequence are unique for
frequencies only in the range -1r:S W < 1r.
Analytic results. Demonstrate that {x(n)} = {sin(wn)} produces identical values
when w = Wo and w = Wo + 2m1r, where m is any integer and - n :s Wo < n. Sketch
the results for a particular value of wo, with m = 0 and 1.

Computer results. Using the program TESTO.FOR in Appendix A, enter the two
values of frequency used above to produce the sinusoidal sequences. The program
asks for the frequency in Hertz per sample interval F, where F = w /2n. The range
of F that produces unique values is then -0.5:s F < 0.5.
Use the PLOT.FOR subroutine given in Appendix A to plot the data on the
terminal and on the printer. Since 55 lines fit on one piece of paper, limit the
number of points plotted to 55 whenever possible.
2.2. Exponential sequences
Object. Generate an exponential sequence and to determine the value number
range of the computer.

(1) Real-valued exponential

Analytic results. Determine and sketch the exponential sequence
x(n) = r"

for 0 :s n :s 8.

for some positive value of 0.9 < r < 1. Repeat for some negative value of
-l<r< -0.9.
Computer results
(a) Write a program to accept a real-valued constant r and compute 55 elements
of the exponential sequence. Plot the sequences for the two values of r used
above. Compare your results with analytically determined sequences and
sketches.
(b) Determine an approximate range for the largest floating-point number that
can be expressed on your computer by accepting increasingly larger values of r
until an overflow eITor is encountered in computing the sequence. Determine
the smallest magnitude number that is still nonzero by accepting decreasingly
smaller values of r, until either an underflow eITor occurs, or the sequence
values become zero.

DISCRETETIME DESCRIPTION OF SIGNALS AND SYSTEMS

61

(2) CompLex-vaLued exponential

Generate and plot 55 elements of the real and imaginary parts for the complex
exponential sequence given by
x(n) = (r e JWo)"

for Wo = :rIB and for the two values of r used above.

2.3. Determination of the effective duration of discrete-time sequence
Object. Determine the effective duration of an infinite duration unit-sample
response sequence.
For the infinite-duration exponential sequences, the magnitude of the
elements typically decreases as rn. Since the plotting routines have a resolution of
approximately 1 percent, we can define an effective duration of the sequence, as
the number of sequence values that are greater than 1 percent of the maximum
value of the sequence. Beyond this effective duration, all the elements can be
considered as being zero-valued.
Analytic results. For the real-valued exponential sequence x(n) = rnu(n),
determine an analytic expression for the effective duration of the sequence as a
function of r.
Computer results. Verify this result by plotting several representative sequences.
2.4. Verification of the digital filter programs
Object. Write and verify the subroutines to implement nonrecursive and recursive
filters.
Write the subroutines NONREC and REC as described in the text. To
implement the noncausal feature, write the subroutine SHIFT. To verify the filter
operation, let the input array be a delayed unit-sample sequence, whose nonzero
element occurs at N = 6, i.e., X(6) = 1, X(N) = 0, for 1 =s N =s NX, N 6. With
this definition, the sixth element of the output sequence will correspond to n =
and the noncausal components will occur in Y(l) through Y(5).

For each of the filters below, perform the following:

(i) draw the filter structure;
(ii) determine an analytic expression for the unit-sample response sequence and
sketch the sequence;
(iii) compute and plot the unit-sample response;
(iv) compare the analytic and computed results, indicating the agreement or
explaining any differences.
Computer results
(1) Causal nonrecursive fiLters. Implement the following difference equations as
causal digital filters. Let NX = 20.
(a) y(n) =x(n) + 2x(n -1) + 3x(n - 2)
(b) y(n)=x(n)-x(n-lO)
(2) NoncausaL nonrecursive fiLter. Implement the 3-sample averager given by
y(n) = Ux(n + 1) +x(n) +x(n -1)].
(3) Causal recursive filters. Implement the following difference equations. Let
NY=55.

62

INfRODUcnON TO DIGITAL SIGNAL PROCESSING

(a) y(n) = aly(n - 1) + x(n)

for a l = -0.95
(b) y(n) = aly(n - 1) - azY(n - 2) + x(n)
for a l
Let ,

= 0.95 and Wo = :n: /8.

= 2, cos( wo) and a2 = ,2.

Hint:

,n

h(n) = -.-(-) sin[(n + l)woJ

for n 2:0

SID Wo

2.5. AppHcation of a noncausal filter: a Hnear interpolator

Object. Implement a digital interpolator as a noncausal filter.
Consider the following sequence of values
0.95(n-6)

x(n)

={

for 6 :S n :S 36 and n even

otherwise

In practice, this sequence could have been obtained by placing observed data

values into the even-indexed elements of an array. The idea is to determine the
values of {x(n)} that are zeros by employing an interpolation sequence.
To perform the interpolation, let us consider the noncausal filter whose
unit-sample response is given by
h(-I) =0.5
h(O) = 1
h(l) = 0.5

Analytic results. If {y(n)} = {x(n)}

* {h(n)},

y(n) =x(n)

prove that

for n even

and

Y()
n =

x(n - 1) + x(n

+ 1)

for n odd.

Computer results. Implement the causal filter to obtain the intermediate array
{y'(n)}, then shift {y'(n)} to obtain the desired array {y(n)}. Compare the plots
of {y(n)} with {y'(n)} to verify the shift and with {x(n)} to verify the
interpolation.
Notice the strange values at the ends of the interpolated sequence. These are
artifacts produced by assuming the values of the sequence are zero outside the
specified range.
2.6. Cascading filters
Object. To implement more complicated digital filters by cascading simpler filter
sections.
Write a program to cascade NONREC and REC. To verify the filter
operation, let the input array be the delayed unit-sample sequence given by
X(N) =

{~

for N= 6,
otherwise.

For each of the filters below, perform the following:

(i) draw the filter structur.e and determine the difference equation;

DISCRETE-TIME DESCRIPTION OF SIGNALS AND SYSTEMS

63

(ii) analytically determine and sketch the unit-sample response;

(iii) compute and plot the unit-sample response;
(iv) compare the analytic and computed results_
(1) Cascade of nonrecursive filters_ Cascade
yn(n) = x(n) + 2x(n -1) + 3x(n - 2)

and

y(n) = Uyn(n

+ 1) + yn(n) + yn(n -1)]

Determine the number of nonzero elements in the unit-sample response sequence_

(2) Cascade of nonrecursive and recursive filters_ Cascade
yn(n) =x(n) - b l x(n -1)

and

for b l = r cos( wo)

yr(n) = al yr(n -1) - a2yr(n - 2) + yn(n)

Let r

= 0.95,

Wo

for a l = 2rcos(wo) and a2 =

r.

= ;r/8 and NY = 55.

(3) Cascade of recursive filters. Cascade

and
for a = 0.8 and b
2.8.)

yr(n) = ayr(n -1) +x(n)

yr'(n) = b yr'(n - 1) + yr(n)

= -0.9 and NY = 55.

(Compare this result with that of Example

1.7. DiptaI tone generators for frequency-shift keying (FSK) system.

Object. Simulate a frequency-shift keying (FSK) system by implementing two
digital tone generators at different frequencies and an appropriate switching
scheme.

In the FSK system, each bit value (lor 0) will be transformed into 32
elements of a sinusoid at a different frequency: a zero will be transmitted as
cos(won), for OSn s31, and a one will be transmitted as cos(wln), for
OSn s31.
The two frequencies will be personalized to each student: let K be the sum
modulo-8 of the last four digits of your social security number, then

wo=2;rK/32

and

WI

2;r(K + 8)/32

(if K is equal to zero, make K = 1)_

The 32 sinusoidal values are then inserted into the sequence {x(n)} to be
transmitted, the first bit sequence going into x(O) to x(31), the second valu.e into
x(32) to x(63), etc.
Analytic results. Draw a second-order recursive digital filter that acts like a digital
oscillator and generates cos(wn), for n ~ 0, for each of the two frequencies above.
Then. 32 values from the unit-sample response can then be inserted into the
transmitted sequence at the appropriate time from the desired oscillator.

Computer results
(a) Generate and plot the two 32 element sequences that correspond to the two
binary values. On the computer plot, sketch the continuous-time sinusoid from
which the samples were derived_

64

INfRODUcnON TO DIGITAL SIGNAL PROCESSING

(b) To produce an FSK sequence for future projects, write the subroutine

FSK(X)
where X (real 0lltput array) is an array of duration 96 that contains the
discrete-time sequence that results for the binary pattern 010.
Plot the output sequence. Sketch directly on the plot the continuous-time
sinusoidal function that corresponds to the samples.

FURTHER READING
Gold, Bernard, and Charles M. Rader: Digital Processing of Sig1Ulls, McGraw-Hill, 1969. [This
classic is one of the earliest books in the field.]
Oppenheim, Alan V., and Ronald W. Schafer: Digital Signal Processing, Prentice-Hall,
Englewood Cliffs, NJ, 1975. [This book was probably the most popular book in digital
signal processing, before the availability of the present text, and is suitable for advanced
study of many of the topics covered here.]
Rabiner, Lawrence R., and Bernard Gold: Theory and Application of Digital Signal Processing,
Prentice-Hall, Englewood Cliffs, NJ, 1975. [This book provides many practical aids for
digital filter design and implementation. J

REFERENCE TO TOPIC
TlDle-varying filters

Cowan, C. F. N. and P. M. Grant: Adaptive Filters, Prentice-Hall, Englewood Cliffs, NJ, 1985.
Goodwin, C. G. and K. S. Sin: Adaptive Filtering, Prediction and Control, Prentice-Hall,
Englewood Cliffs, ~J, 1984.
Haykin, Simon: Adaptive Filter Theory, Prentice-Hall, Englewood Cliffs, NJ, 1986.
Honig, M. L. and D. G. Messerschmitt: Adaptive Filters, Kluwer Academic, Boston, 1984.

CHAPTER

3
FOURIER
TRANSFORM OF
DISCRETE-TIME
SIGNALS

3.1 INTRODUcnON
In many signal processing applications, the distinguishing features of signals
and filters are most easily interpreted in the frequency domain. This chapter
describes the frequency-domain properties of discrete-time signals and the
frequency-domain behavior of discrete-time systems. The main analytic tool is
the Fourier transform of a discrete-time sequence, which can be applied either
to the unit-sample response of a filter to define the filter transfer function, or to
a data sequence to define the signal spectrum. The important properties of the
Fourier transform applied to discrete-time sequences are itemized and discussed. The magnitude and phase responses of a digital filter are examined. In
practice, discrete-time sequences are commonly obtained by sampling
continuous-time signals. The relationships between the spectra of continuoustime signals and the corresponding discrete-time sequences are investigated.
Interpolation functions to convert discrete-time sequences into continuoustime functions are also desc:ribed.

3.2 DEFINITION OF mE FOURIER

TRANSFORM
The frequency-domain analysis of signals and systems usually employs sinusoidal sequences, given by {sin(On)} and {cos(On)}, or mqre concisely {eieon },
65

66

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

where w is the radian frequency. Sinusoidal sequences have the important

property that when a sinusoid of a particular frequency is applied at the input
'Of a linear time-invariant (LTI) system, then the output is also a sinusoid at the
same frequency, although it may have a different amplitude and phase. The
same statement cannot be made about other functions, such as a square-wave,
for example.
To demonstrate this form of invariance, let the.input to an LTI system be
the complex sinusoid {x(n)} = {eiaJo"}. Then, as shown in the previous chapter,
the output is equal to the convolution of the input and the unit-sample
response of the system:

y(n)=

..
..
L

h(k)x(n-k)

k=-CCI

k=--CIC

= eiaJo"

h(k) eiaJo(n-k)

h(k) e-iwok

(3.1)

k=-CID

We can interpret this last result by using complex-vector notation. Let us

define a complex-valued coefficient H(eiaJo ) that is equal to the evaluation of
the above sum, or

..

H(ei wO) =

h(k) e- iwok = IH(eiWo)1 exp{j Arg[H(eiaJo)]}

(3.2)

k--oc

Including this coefficient in Eq. (3.1), we get

y(n) = H( ei aJO) eia>o"
= IH(eiwo)1 exp{j(won

+ arg[H(eiwo)])}

(3.3)

Hence, the output sequence is a complex exponential at the same frequency

wo, but has been multiplied by a complex-valued time-invariant coefficient
H( ei aJO). This coefficient has introduced both gain and phase factors, as shown
in Fig. 3.l.
{x(nl} = {sin (won)}

wo=rr/4

{y(n)} = {iH(e l .'" llsin (won+Arg [H(e l "")])}

IH(ei"")1

Ar
012_

00

LINEAR
TIME-INVARIANT
SYS1EM

FIGURE 3-1
Magnitude and phase effects produced by a linear time-invariant system.

FOURIER TRANSFORM OF DISCRETE-TIME SIGNALS

67

This result is true for any frequency wand hence this coefficient can be
generalized to be a function of frequency. We will denote this frequency
function by H(e jW ), where
00

H(e jW ) =

h(n) e- jw"

(3.4)

This equation defines the Fourier transform of the discrete-time sequence

{h(n)}.
The Fourier transform of a sequence is said to exist if it can be expressed
in a valid functional form. Since the computation involves summing a possibly
infinite number of terms, the Fourier transform exists only for sequences that
are absolutely summable. That is, given a sequence {h(n)}, H(e jW ) exists only
when
00

(3.5)

Ih(n)l<oo

n=-CIC

When the sequence defines the unit-sample response of a digital filter,

then its Fourier transform is called the transfer function. It defines the
frequency transmission characteristics of the filter. When the sequence is a
discrete-time signal {x(n)}, then its Fourier transform is given by
00

X(e jW)

= L

x(n) e- jw"

(3.6)

n-=-oo

Then X(e jW) defines the frequency content of {x(n)}, and will be called the
signal spectrum.
Eumple 3.1. Transfer function of the 3-sample avenger. Recall that the
unit-sample response for the noncausal 3-sample averager is given by

h(n)={~ol

for-lsnsl
otherwise

Then its transfer function is given by

~

H(ei ,") =

h(n) e- J.... =

le- i ,""

n--i

= 1(ei "

+ 1 + e- i ..)

This last result can be simplified by applying the Euler identity

cia

+ e -ja = 2 COS Q'

to get

Example 3.2. Transfer function of the first-order recursive filter. Recall that the
unit-sample response is given by
h(n) = a"u(n)

for all n

68

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

Because we have an infinite number of elements in {h(n)}, we must check to

determine whether it is absolutely summable:
~

Ih(n)1 =

,. _ _ 00

n-O

lain

If lal ~ 1, the sum becomes infinite and the Fourier transform does not exist. For
lal < 1, we can apply the infinite geometric sum formula

L lal"=-"=0
1-lal

for lal < 1

from which we note that the sum is finite for la I < 1. Then
~

H(eiOJ) =

L a" e-iOJn = L (a e-iOJ)"

Since la e-iOJI < 1 when lal < 1, we can apply the infinite geometric sum formula
to obtain
.
1
H(eIOJ) =
.
for lal < 1.
l-ae 1
01

The necessary condition to apply the infinite geometric sum formula in

evaluating the Fourier transform is the same for the sequence to be absolutely
summable. Hence, it is not necessary to determine whether the sequence is
absolutely summable as a separate step.
Example 3.3. Transfer
Example 2.11, we have

fuadiOD

of the secoIld-order recursive filter. From

h(n) = r" cos(woll) u(n)

Then
H(eiOJ) =

for all

11

L- r" cos(woll) e-J'"

Applying the Euler identity to the cosine, we have

i
=! [i (r e -i(OJ-"'o" + i (r e -i("'+",ol)"]
2 ,,_0
,,-0

H(eio=![i r"ei ..O"e- i.... +

r"e-iO>O"-i ....]
2 .. -0
,,-0

For Irl < 1, we can apply the infinite geometric sum formula to get

H(ei"') = U(1 -

r. e-J(W-GJol)-l + (1- r e-i(OJ+wolt1J

Combining over a common denominator, we have

.
1
2 - r(eJ....o + e-j.,O) e- i'"
H(el ....) =.
.
.
-2
...
21- r(el....O + e- IOJO) e- IOJ + r e-l~
1- r cos(I)o) e- i.,

FOURIER TRANSFORM OF DISCRETETIME SIGNALS

69

In the examples above, the sequences were written in the form of the
unit-sample response of a system {h(n)}. It is to be understood that the same
results would have been obtained for a signal sequence {x(n)}, having the
same element values. In this case, we would be calculating the spectrum,
rather than the transfer function.

3.3 IMPORTANT PROPERTIES OF THE

FOURIER TRANSFORM
We now present some important properties of the Fourier transform. These
are summarized in Table 3.1.
LINEARITY. To demonstrate that the Fourier transform is a linear transformation, we apply the superposition principle to the definition. Let
{y(n)}

= {axl(n)} + {bx2(n)}

(3.7)

Then its Fourier transform is equal to

00

Y(ei"')

= L

y(n) e- i "'"

n=-oo
00

L [axl(n) + bX2(n)] e- i "'"

n=-cc
00

=a L

00

xl(n) e- i "'" + b

n=-oo

L x2(n) e-i "'"

n=-QO

=aX1(ei ",) + bX2(e j ",)

(3.8)

Since X 1 (e J"') is the Fourier transform of {xt(n)} and X 2(ei"') is that of

{x2(n)}, we have the desired result.
PERIODICITY. The Fourier transform is a periodic function of the
continuous-valued variable w, with period 27[. This can be demonstrated by
increasing the frequency in the definition of the Fourier transform by 2k7[,
where k is any integer. We then have

H(ei ("'+2k1r

L
00

h(n) e- i (",+2kn)"

n=-ac

= L

h(n) e- i"'" e-i2knn

(3.9)

n=-tIC

For integer values of k and n, the case here, e-i2knn = 1. Hence, we have
H(ei (",+2kn = H(ei "').
This periodicity can also be noted because the variable w appears in the
function only as e i"', which is itself periodic with period 27[. If the argument of
a function is periodic, it follows that the function itself is also periodic.

70

INfRODUcnON TO DIGITAL SIGNAL PROCESSING

TABLE 3.1

Properties of the Fourier transform

1. Definition:

H(~") =

h(n) e-

(eXists when .. ~~ Ih(n)1 < 00)

I'"

2. Linearity: If

3. Periodicity:
H(~a = H(ei (OJ+2kn

for any integer k

4. Magnitude and phase functions: If

H(~OJ)

= HR(eiOJ) + jHI(eia

then
IH(~OJ)12 = Hi(~OJ) + HUe i "')
Arg[H(~OJ)

= arctan[HI(~")/ HR(eia]

5. Fourier transform of delayed sequence: For any value of no. if

(y(n)}

= {x(n -

no)}

Y(ei"') = X(eiOJ ) e-i"'''o

then

6. Fourier transform of the convolution of two sequences: If

(y(n)}

= {h(n)} * {x(n)}

then

Y(eJOJ ) = H(~"')X(ei"')

7. Fourier transform of the product of two sequences: If

{y(n)}

= {h(n)x(n)}

Y(eiOJ) =H(eiQ X(~OJ).

then

(periodic convolution)
8. For the real-valued sequence {h(n)}:
~

HR(ei"') =

h(n) cos(wn)

H1(eiOJ) = -

h(n) sin(wn)

11_-Il10

(a)
(b)
(c)
(d)
(e)
(f)
(g)

Complex-conjugate symmetry:
H(ei ") = H(ei "') = H(e- iOJ)
Real component is even function:
HR(~") = HR(e- iOJ )
Imaginary component is odd function: HI(el W ) = -H1(e- iOJ)
Magnitude spectrum is even function:
IH(~CV)I = IH(e-i"')1
Phase spectrum is odd function:
Arg[H(eiw)] = -Arg[H(e- 1W)]
If {h(n)} = {h( -n)} (even sequence), then H(~a is purely real.
If {h(n)} = {-h(-n)} (odd sequence), then H(~W) is purely imaginary.

REAL
AND
IMAGINARY
COMPONENTS
OF
THE
FOURIER
TRANSFORM. Since the Fourier transform is a complex function, it is

helpful to define alternate forms that can be plotted on a pair of twodimensional graphs. Let us consider the general case of a complex sequence
{h(n)}, which can be written as
h(n) = hR(n)

+ jhl(n)

(3.10)

FOURIER TRANSFORM OF DlSCRETETIME SIGNALS

71

where hR(n) is the real component and hJ(n) is the imaginary component. By
applying Euler's identity to the definition of the Fourier transform, we have
~

H(e i"') = ~ h(n) e-i",n = ~ [hR(n) + jhJ(n)][cos(wn) - j sin(wn)]

n=-oo

n=-~

(3.11)

This function can be expressed explicitly in terms of its real and imaginary
parts as
(3.12)

where
~

HR(ei"') = ~ [hR(n) cos(wn) + hJ(n) sin(wn)]

(3.13)

n=-oo

and
~

HJ(ei"') = ~ [hJ(n) cos(wn) - hR(n) sin(wn)]

(3.14)

n=-oo

MAGNITUDE AND
PHASE
COMPONENTS
OF THE
FOURIER
TRANSFORM. Being a complex quantity, we can also express H(eij as a

complex vector:
(3.15)
If H(ei"') denotes the transfer function of a system, it is conventional to call
IH(ei"')1 the magnitude response, and Arg[H(ei"')] the phase response. For
signal sequences, H(ei"') is the spectrum, IH(ei"')1 is called the magnitude
spectrum, and Arg[H(ei"')] is the phase spectrum. The magnitude response is
equal to
(3.16)

where H*(eiOJ) is the complex-conjugate of H(ei"'). An equivalent expression

for the magnitude is given by
IH(ei"')1 = (Hi(ei"') + Ht(ei "'ll2
(3.17)
The phase response is equal to
(3.18)

The phase is an indeterminate quantity when IH(ei"')1 =0, which occurs at

values of w for which HR(ei "') = HJ(ei"') = O.
Before giving some examples of computing the magnitude and phase
responses, we show that we can usually simplify the calculation of the Fourier
transform for real-valued sequences by considering the special properties of
their transforms.

72

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

3.4 PROPERTIES OF THE FOURIER

TRANSFORM FOR REAL-VALUED
SEQUENCES
If {hen)} is a sequence of real numbers, then its Fourier transform can be
written as
00

H(e jOJ )=

00

00

L h(n)cos(wn)-j L
n=-oo

h(n)e- jOJn =

n=-tD

h(n)sin(wn)

n=-oo

(3.19)
The Fourier transform of a real sequence has the following useful symmetry
properties.
SYMMETRY OF THE REAL AND IMAGINARY PARTS. Let us consider the
effect of reversing the sign of the frequency variable in the argument of the
transform. We can then write
H(e-io = HR(e-io + jH1(e-iOJ)
(3.20)
Evaluating H(e-iO, we find
00

H(e-iO

= L

hen) e jwn =

n=-oo

00

00

L hen) cos(wn) + j L
n=-oo

hen) sin(wn)

(3.21)

n=-OC

Comparing Eqs. (3.19) and (3.21), we note that

00

HR(eiO =

hen) cos(um) = HR(e- jOJ )

(3.22)

indicating that the real part is an even function of w. Comparing the imaginary
parts, we find
00

HI(eifD) = -

L hen) sin(wn) = -H1(e-ico)

n--co

(3.23)

indicating that the imaginary part is an odd function of (J). Also, since the
complex conjugate is obtained by changing the sign of the j terms, we find
H(e- jCO ) = H*(ejO.
(3.24)
Since H(ei CO) is a periodic function of w with period 2Jr, an important
consequence of these results is that, for real-valued discrete-time sequences, all
the information in the Fourier transform is contained in the frequency range
O:s; w :s; Jr. The unspecified portion of the Fourier transform in the range
-Jr :s; w < 0 can be determined from these symmetry properties.
SYMMETRIES OF THE MAGNITUDE AND PHASE FUNCDONS. The magnitude response is also an even function of frequency, since
IH(ejOJ)1 = (H(e jOJ )H*(e jOJ
= (H*(e- jCO )H(e- jCO = IH(e-iCO)1 (3.25)

ll2

ll2

FOURIER TRANSFORM OF DISCRETETIME SIGNALS

73

The phase response can be written

Arg[H(e- iro )] = arctan[HI(e-1W)/HR(e-1W)]

= -arctan[HI(eiw)/HR(eiro)]
= -Arg[H(eiw )]

(3.26)

indicating that the phase response is an odd function of frequency. To illustrate

the above properties let us consider a few examples.
Example 3.4.
sequence. Let

Magnitude and phase spectra of the delayed unit-sample

d(n-k)={~

for n =k
otherwise

Since the unit-sample sequence is commonly used as the input to a filter, it is

considered to be a signal sequence and it is proper the call its Fourier transform a
spectrum. Let us denote the spectrum of {d(n - k)} by Dk(eIW ). Then
~

Dk(e IW ) =

2:

d(n - k) e- iwn = e- iwk

The real and imaginary parts are most easily determined by applying the Euler
identities:
and
The magnitude spectrum is equal to
IDk(eiw)1 = [cos2 (wk) + sin2 (Wk)]112 = 1
or, alternately,
IDk(eIW)1

= (e-Jwk)(eJcat)112 = 1

The phase spectrum is equal to

Arg[Dk(eiw )] = arctan[ -sin(wk)/cos(wk)] = arctan[ -tan(wk)] = -wk

The magnitude and phase spectra could also have been obtained by inspection by
noting that Dk(eJW) is already in the complex vector form. The magnitude and
phase functions are shown in Fig. 3.2 for different values of k.

In Fig. 3.2, for k = 0, the magnitude spectrum is constant with frequency

and the phase spectrum is identically zero. It is for this reason that {d(n)} is
desirable as a filter input, since the spectrum of the output sequence can then
be attributed to the filter transfer function. For the unit-sample sequence
delayed by k samples, the magnitude response is still constant with frequency,
but the phase response is linear with frequency, equal to -kw.
It is the accepted convention to plot the principal value of the phase lying
in the range [-lC, n]. Even though the phase values of +lC and -lC are
equivalent, we will allow both to occur to accommodate the plotting
conventions described below. Note from Fig. 3.2, that when the phase exceeds

74
(a)

INTRODucnON TO DIGITAL SIGNAL PROCESSING

(b)

k=O

(d")} . .

r. .

(c)

k>0

:}_.J.I_l_)

__{_d(_n-_k....

..n

-2,-1012

o1

k<O

r.

(<I(.-.))

k -10

IDu(ei"') I

-21T

-IT

If

If

I) w
2Ii

Arg[Do(ei"') J
!

-21T

-IT

t
0

-+

If

FIGURE 3-2
Magnitude and phase spectra of the unit-sample sequence. (a) Standard unit-sample sequence
(k = 0); (b) Delayed sequence (k > 0); (c) Delayed sequence (k < 0).

this range, a jump of 2.1t' is needed to bring the phase back into this range.
This principal value range is also convenient, since it is the value returned by
the commonly available ATAN2(Y, X) computer function, where Y is the
imaginary component and X is the real component.
Example 3.5. Mapitllde and pbase response of the 3-samp1e avenger. From
Example 3.1, we have

H(ei "') = !(1 + 2 cos(co

which is a real-valued function of co. The magnitude response is equal to

IH(ei"')1 = 11 +2~(CO)1
and the phase response is equal to

Arg[H(eJ"')] =

{O
1r

when H(eIOl) > 0

when H( el"') < 0

orfor - 2.1t' /3 < co < 2.1t' /3

or for -.1t'S co < -2.1t'/3
and 2.1t'/3 < co < 1r.

The magnitude and phase responses are shown in Fig. 3.3. The appropriate sign
of 1r is chosen to make the Arg[H(eio] an odd function of frequency.

As a filter, the magnitude response indicates that the 3-sample averager is

a lowpass filter. If the same discrete-time sequence were a signal, then the

FOURIER TRANSFORM OF DISCRETE-TIME SIGNALS

75

{h(n)}

...

.~"1.

-2 -10 1 2

nGURE 3-3
Magnitude and phase responses of the 3-sample averager.

magnitude spectrum indicates that it would have more energy at low

frequencies than at high frequencies.
The previous example illustrates that precautions must be taken when
determining the phase response of a filter having a real-valued transfer
function, because negative real values produce an additional phase of 1r
radians. For example, let us consider the following linear-phase form of the
transfer function that will occur repeatedly throughout the book:
H(eiQJ ) = B(eiQJ ) e- ikQJ
(3.27)
where B(eio is a real-valued function of w, called the amplitude function, that
can take on both positive and negative values. When expressing this transfer
function in the complex vector notation,
(3.28)
then the phase function must include the linear phase term and also
accommodate for the sign changes in B(ejW ). Since -1 can be expressed as
ej,., phase jumps of 1r will occur at frequencies where B(e jW ) changes sign.
This jump must be added to the phase component that is calculated from the
arctangent function. If B (eiw ) ~ 0 for all w, then no jumps will occur and the
phase function is given by - kw.
Eumple 3.6. Linear-phase Fourier transform. Let
h(O) =!

h(l) = 1

h(2) = t

h(n) = 0 otherwise

Then the Fourier transform is equal to

H(ei "') =

! + e-J'" + ~e-i2'"

76

INfRODUcnON TO DIGITAL SIGNAL PROCESSING

which can be expressed as

H(eiID) = e- iID + !e-iID(eieo + e- ieo)

= (1

+ cos(we-ieo

This last form is the linear-phase form given above. In this case, the amplitude
function is never negative and the phase response is simply - w.

PHASE JUMPS. From the previous examples, we note that there are two
occasions for which the phase function experiences discontinuities, or jumps.

1. A jump of 2Jr occurs to maintain the phase function within the principal
value range of [-n, n).
2. A jump of n occurs when B(eiev ) undergoes a change in sign.
The sign of the phase jump is chosen such that the resulting phase function is
odd and, after the jump, lies in the range [-n, n].
Example 3.7. Magnitude and phase response of the first-ordel' reausive
filter. From Example 3.2, we have H(ei"') = (1- a e- i 1 The magnitude
response is most easily obtained by calculating the squared magnitude function,
or

"'r

IH(eieoW = H(eieo)H*(e i",)

= (1- a e-i ",)-1(1- a ei ",)-l
= (1- 2a cos(w) + a 2)-1

To find the real and imaginary parts, we make the denominator real by
mUltiplying by its complex conjugate:

1
1 - a ci"'
- 1 - a e i"'1 - a ei "

H( i"') _

1- a ei "'
1- a cos(w) - ja sinew)
1 - 2a cos( w ) + a 2

Since the denominator is real, we can now separate the numerator terms to
determine the real and imaginary parts. The real part is then

/{, (ci") =
R

1- a cos(w)
1-2acos(w)+a2

The real part is an even function of frequency. The imaginary part is equal to

H1(ci"') =

-a sinew)
1- 2a cos(w) + a2

FOURIER TRANSFORM OF DISCRETETlME SIGNALS

77

a= -0.9 IH(e'''')1
a= -.0.5
I
I

---"

".

.-

/~a=-0.9

_-'---_ _ _--:.::::c:..._ _ _ _.L-.;;. w

-IT

Arg[H(e'''')]

Arg[H(e'''')]

n/2

n/2

......'"","" \

a=O.5

a=-0.5
-n/2

ftGURE 3-4
Magnitude and phase responses of first-order recursive filter for different values of Q.

and is an odd function of frequency. The phase response is then equal to

Arg[H(ei"')] = arctan[HJ(ei"')/HR(e'''')]
= arctan[ - a sine co) ]

l-acos(co)
These responses are shown in Fig. 3.4 for different values of a. When 0 < a < 1, a
lowpass filter results, and when -1 < a < 0, a highpass filter results.
Example 3.S. Magnitude and phase response of the second-order recursive
filter. From Example 33, we have
H(e i"')=

1-rcos(co~)e-I'"
1- 2r cos(coo) e- l '" + r2 e-,2O)

For simplicity of notation, let us write H(el "') in the following form
1 + a e-iO)
N(e i"')
H(el "')=
=
-l+ae- i"'+pe-,2O) D(eiw)
where N(eiw ) is the numerator and D(ei"') is the denominator. The magnitude
response is then
(N(eiO N*(e iO)lI2
.
IH(el"')1 = D(ei aJ ) D*(eIW)
_
- [1 + ~ +

[1 + 2a cos(co) + a2f12
p2 + 2a(1 + P) cps(co) + 2P coS(2CO)j112

Note from this rather complicated form that co appears only as an argument of

78

IN1"RODUcnON TO DIGrFAL SIGNAL PROCESSING

jH(eJro)j

Arg[H(elro)]
11

FIGURE 3-5
Magnitude and phase responses of second-order
recursive filter for different values of r.

the cosine function, which is an even function of w. This is reasonable, since

jH(eIW)j is itself an even function of w.
The real and imaginary parts are obtained by computing
'w

H(e

N(eiw ) D*(el "')

= D(eiw ) D*(e iW )

1 + aex + a e- iw + (af3 + ex) eiw + f3 ei2co

1 + a2 + f32 + 2ex(1 + f3) cos( w) + 2f3 cos(2w)

From this expression we find

HR(eij = 1 + aex + (a + af3 + ex) cos(w) + f3 cos(2w)

1 + a2 + f32 + 2ex(1 + f3) cos( w) + 2f3 cos(2w)
and

HI

ei )
( W

(af3

+ ex - a) sin(w) + f3 sin(2w)

= 1 + a2 + f32 + 2ex(1 + f3) cos( w) + 2f3 cos(2w)

The phase response is then given by

Ar [H(eiW)] = arctan[
(af3 + ex-a)sin(w) + f3 sin(2w)
]
g
1 + aex + (a + af3 + ex) cos(w) + f3 cos(2w)
The magnitude and phase response curves are shown in Fig. 3.5 for
various values of r.

Wo

= ,,/4 and

The magnitude response shows a sharp peak close to the frequency

which is called the resonant frequency. This peak will be
employed later to implement the passband features of a digital filter.
00

= n/4 = 000,

Example 3.9.

Magnitude and pbase response of causal3-samp1e avenger. Let

h(n)={lo

forOsns2
otherwise

FOURIER TRANSFORM OF DISCRETETIME SIGNALS

79

Then

We shall evaluate this sum using three methods: the mathematically tedious
approach, the enlightened direct approach and by applying the finite geometric
sum formula.
Mathematically tedious method. Evaluating the sum, we get
H(eiw ) = ill + e- iw + e- i2w]
= ill + cos(w) - j sin(w) + cos(2w) - j sin(2w)]

= ![l + cos(w) + cos(2w)] -1j[sin(w) + sin(2w)]

= HR(e'W) + jHI(eiw )

IH(e1W)12 = Hi(eiw) + H;(e1W)

+ cos(2w + 4 oos(w)]
= Ml + 4ooS2(W) + 400s(w)] = Ml + 2cos(wW

= ~[1 + 2(1

Taking the square root, we get

IH(e1W)1 = 111 +2cos(w)1
This result is shown in Fig. 3.6.
Direct method. To allow an Euler identity substitution, we factor out e- 1W :
H(eiw) = !e-iw(eiw + 1 + e- iw )
=

W + 2cos(we- iw = B(eiw)e-iw

From which we get

IH(e1W)1 = ill + 200s(w)1

h(n)

~ III

-.....l~O~1-72-!j-4~-

....
~ n

Arg(H(eJO>)]

:~

-1T

FlGURE 3-6
Magnitude and phase responses of causal 3-sample averager.

80

INTRODUCfION TO DIGITAL SIGNAL PROCESSING

and

Arg[H(eJW )] =

{-W
-Wlr

for -2lr/3<w<2lr/3,
for -lr:S W < -2lr/3

or when B(eJW ) >0

or when B(eJW ) < O.

and 2lr/3 < w:S lr,

The phase is undefined at frequency points where IH(eJW)1 = o.
Geometric sum method. Applying the finite geometric sum formula with
r = e- Jw, we get

The above form can be simplified by substituting the following identity:

1 - e-'" = e- j<>i2( e J<>i2 - e- Ja/2)
= 2j e- J<>i2 sin(a-/2)

After this substitution, we get

e - j3 0.>'2(e j3 0.>'2 _ e - J3 0.>'2)

H(e JW ) = .J

3 e- 0.>'2(e J0.>'2 - e J0.>'2)

1

= 3

e -jw sin(3w/2)
sin(w/2)

B( j"') -jw
e e

This last form is the amplitude with linear-phase form described above in Eq.
(3.27). From that discussion, we find
IH(eJW)1 = Isin(3w/2)1
13 sin(w/2)1
and

Arg[H(eIW )] =

{-W

-w lr

when B(eIW ) > 0

when B(e jW ) <0

The magnitude response, an even function of w, is expressed as the ratio of

two odd functions of w. The equivalence of this form with that obtained with the
previous two methods is left as an exercise.

3.5 PROGRAM TO EVALUATE THE

FOURIER TRANSFORM BY COMPUTER
To provide computer verification of the analysis above, we will describe a
simple program to evaluate the Fourier transform of a real-valued sequence at
N p points over the frequency range 0:::; W < 2n. In the next chapter, we will
present a more efficient but mathematically more complicated algorithm for
performing this procedure.
Although not the most efficient method, the evaluation of the real and
imaginary parts at a frequency point Wk can easily be performed directly from
the definition. For example, let h(n), for 0:::; n:::; NH - 1, be a sequence of

FOURIER TRANSFORM OF DlSCRETETIME SIGNALS

81

real-valued elements. Then, the real part of the Fourier transform evaluated at
frequency Wk> denoted by HR(e iWk ), is given by
MI-l

HR(eJWk )

2:

(3.29)

hen) cos(wkn)

n=O

(3.30)

To obtain a discrete-frequency approximation to the continuous-frequency

function H(e iw ), we can evaluate the real and imaginary parts at a set of N p
equally-spaced points over 0:::; W < 2n, i.e. at Wk = 2kn 1Np , for 0:::; k :::;
Np-l.
When processing arrays whose indexing starts at one, rather than zero,
the above equations for the real and imaginary parts must be modified. If
H(K), for 1:::; K :::;NH, is the array of discrete-time values, with H(1) equal to
h(O), then the array containing the real part, denot"Ai by HR(K), is given by
HR(K) =

NH

2:

H(N) * cos(2n(K - 1)(N - 1)1 N p )

for1:::;K:::;Np

(3.31)

N=l

The frequency of the Kth cosine sequence is equal to 2n(K -1)INp , for
1:::; K:::; N p The array containing the imaginary part, HI(K), is
HI(K) = -

NH

2:

H(N) *sin(2n(K --1)(N -l)INp )

for 1 :::;K:::;Np

(3.32)

N=l

Using the programming style described in the pre lious chapter, the
computation of the Fourier transform should be performed with the subprogram, denoted by

FT(H,NH,HR,HI,NP)
where
H

(real input array) contains the elements of the discrete-time

sequence;
NH
(integer input) is the number of elements in H;
HR
(real output array) contains the real part of the Fourier transfonn;
HI
(real output array) contains the imaginary part of the Fourier
transform;
NP
(integer input) is the number of frequency points over O:s W < 2n.
For displaying the Fourier transform, the magnitude and phase responses
are conventionally used. The magnitude and phase responses should be
computed using the subprogram denoted by

MAGPHS(HR,HI,HMAG,HPHS,NP)

82

INlRODUcnON TO DIGITAL SIGNAL PROCESSING

where
HR
HI
HMAG
HPHS
NP

(input real array) contains the real part of the Fourier

transform;
(input real array) contains the imaginary part of the Fourier
transform;
(output real array) contains the magnitude values~
(output real array) contains the phase values; and
(input integer) is the number of points over O:s OJ < 2Jr.

The array of magnitude values can be computed as

HMAG(K) = SQRT(HR(K) * HR(K) + HI(K) * HI(K)]

for 1 :s K:s NP
(3.33)

The array of phase values can be computed as

HPHS(K) = ATAN2(HI(K), HR(K

for 1 :s K :s NP

(3.34)

When both HI(K) and HR(K) are zero, the arctangent function should not be
used, but rather a value of zero should be assigned for the pha~e value.
Some programming languages only have the ATAN(Y IX) function,
which generates the phase in the first and fourth quadrants, i.e. in the range
[-n/2, Jr/2]. It is left as an exercise for the student to determine the principal
value of the phase using the ATAN function and the sign of the real part.
Because of the symmetry properties for real-valued sequences, we need
to evaluate the plot the Fourier transform only over the range 0 :s OJ :S Jr. Since
both the frequency points OJ = 0 and OJ = Jr are included, the number of points
to be computed is equal to one more than half the points covering one entire
period, or (Np /2) + 1.

3.6 USE OF THE FOURIER

TRANSFORM IN SIGNAL PROCESSING
The popularity of the Fourier transform for digital signal processing is due to
the fact that the complicated convolutional operation in the time domain i~
converted to a much simpler multiplicative operation in the frequency domain.
To demonstrate this, let {yen)} be the output of a LTI system. Then, its
Fourier transform is given by
Y(eio

yen) e-i "'"

(3.35)

Applying the convolutional definition of the output, we have

Y(eiO

= ,,~'" L~

h(k)x(n - k)] e- i"'"

(3.36)

FOURIER TRANSFORM OF DISCRETE-TIME SIGNALS

MUltiplying this last equation by unity, in the form eJwk e- Jwk , we get
'"
.,
Y(eiw ) =
[h(k) e-iwkx(n - k) e-Jw(n-k)]

2: 2:

83

(3.37)

n=-oo k=-oo

Since h(k) e- iwk is not a function of n, it can be factored out of the summation
with respect to n, giving
Y(eiw )

= k~'" [h(k) e- iwk n~'" x(n -

k) e-iW(n-k)]

(3.38)

In the sum over n, the term (n - k) represents a shift of {x(n)} by k samples.

Further, note that the value of the exponent follows the index of {x(n)}. Since
the summation is over the entire sequence, this shift is inconsequential in
evaluating the sum. By replacing (n - k) with the dummy variable m, we note
that the second summation becomes the definition of X(e iw ), the Fourier
transform of {x(n)}. Since X(e iaJ ) is not a function of k, we can factor it out of
the summation over k. The remaining terms in the sum then define the Fourier
transform of {hen)}. Hence, we have arrived at the desired result
(3.39)

This last result indicates that the input spectrum X(e iW ) is changed through a
multiplicative operation by the filter transfer function H(e iW ) to produce the
output spectrum Y(ei aJ ). Applying the complex vector notation to the output
spectrum, we obtain
Y(eiaJ) = IY(eiaJ)1 exp{j Arg{Y(ei"')]}
= IH(ei"')1 exp{j Arg[H(ei"')]} IX(eiaJ)1 exp{j Arg[X(ei"')]}

(3.40)

=IH(ei"')IIX(ei"')1 exp{j(Arg[H(eiaJ)] + Arg[X(ei"')])}

Equating the magnitudes and the phases, we find
(3.41)

and
(3.42)
In words, the output magnitude spectrum is equal to the product of the input
magnitude spectrum and the filter magnitude response. The output phase
spectrum is equal to the sum of the input phase spectrum and the filter phase
response.
Example 3.10. Sipal processing with the Fourier transform. Let the input
sequence be given by
x(n) =

{~

for -1 Sn S 1

otherwise

84 INTRODUCTION TO DIGITAL SIGNAL PROCESSING

and the filter unit-sample response by

a"

h(n) = { 0

for n 2:0
otherwise

From Examples 3.1 and 3.2, we have

X(e iw ) = i{1 + 2cos(w
and

The output spectrum is then equal to

Y(eiw) =H(eiw)X(eiW ) =

~(; ::o;~~

The output squared-magnitude spectrum is equal to

IY(eiw)12 = (1 + 2 cOS(~)(1 + 2 COS(W)
3(1 - a e IW) 3(1 - a elW )

_
(1+2cos(wW
- 9(1- 2a cos(w) + a2 )
The output phase spectrum can be computed in two ways. First, it can be
obtained from the real and imaginary components of Y(ei W ) , determined from
iW

Y(eiw ) = (1 + 2 COS(~)(1- a e )
3(1- a e- IW ) 1- a elw
_ (1- a cos w)(1 + 2 cos(w - ja sin(w)(1 + 2cos(w
3(1- 2a cos(w) + a2 )
Then
j ..
_
(-aSin(w)(1+2COS(W)_
(-asin(W)
Arg[Y(e )]-arctan (1-acos(w(1+2cos(w -arctan 1-acos(w)

The phase can also be determined by adding the phase response of the filter to
the phase spectrum of the input:
Arg[Y(e i "')] = Arg[H(eIW)] + Arg[X(e i "')]

-a sin(w) )

= 0 + arctan ( 1- a cos(w)

3.7 FOURIER TRANSFORM OF SPECIAL

SEQUENCES
In this section, we consider the Fourier transform of several special sequences
that simplify the computations. These results will be helpful for the analysis of
digital filters later.

FOURIER TRANSFORM OF DlSCRETETIME SIGNALS

8S

FOURIER TRANSFORM OF ADELA YED SEQUENCE. Let us examine the

effect on the Fourier transform when a sequence is delayed by k sample points,
as shown in Fig. 3.7. Let us denote the delayed sequence by {hk{n)} =
{h(n - k)}. Then its Fourier transform, denoted by Hk{e jCJJ ), is equal to
Hk(e jCJJ )

2:'"

hk{n) e- jOJn

n=-oo

'"
2:
h{n n=-OCI

k) e- jWft

(3.43)

(h(n)}

Original
sequence

... llil I
o

Delayed
sequence

"..

1 2

(hJn)}

........
II!II.;.
o
1 k-l k k+l

Arg[H(e JW )]

IH(el"')1

1T

-It
.-

CI)

I.-

I
I

It

CI)

-.-

-------1

I
I

F1GURE

~7

Comparison of the magnitude and phase spectra for original and delayed sequences.

CI)

86

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

Multiplying the right side by e- iwk e1Wk, we have

Hk(e jW ) =

00

2:

hen - k) e-1W(n-k) e- jwk

n=-IXI

=e- jwk

00

hen - k) e-iw(n-k)

(3.44)

n=-oo

By substituting m = n - k in the above summation, we recognize the definition

of H(e iw ). Hence,
Hk(e iw ) = H(eiw) e- iwk
(3.45)
By expressing this result in complex vector form, we note
IHk(eiw)1

= IH(eiW)1

(3.46)

and
(3.47)

In words, when we delay a sequence by k samples, where k is any positive or

negative integer, the magnitude of the Fourier transform is unchanged. The
phase is altered by adding a linear-with-frequency phase term whose slope is
equal to -k. The utility of this result will become evident in computing the
Fourier transform of symmetric sequences in the next section.
FOURIER TRANSFORM OF SEQUENCES WITH SYMMETRIES. In the design

of digital filters, we commonly encounter sequences that have a symmetry

about some time point, denoted by Ns . We consider sequences that have even
symmetry and those that have odd symmetry separately.
Even-symmetric sequences are those whose elements about the point Ns
have equal value, or
heNs + m) = heNs - m)

for all m

(3.48)

The two cases, depending on whether Ns is an integer, are shown in Fig. 3.8.
The special case when Ns = 0 defines an even sequence, for which
hen) = h( -n)

for all n

(3.49)

Antisymmetric sequences are those whose elements about the point Ns are
of equal magnitude, but of opposite sign, or
heNs + n) = -heNs - n)

for all n

(3.50)

The two cases, depending on whether Ns is an integer, are shown in Fig. 3.8.
When Ns is an integer, heNs) must equal zero for {hen)} to be an
antisymmetric sequence. The special case when Ns = 0 defines an odd
sequence, for which
hen) = -h(-n)

for all n

(3.51)

FOURIER TRANSFORM OF DISCRETE-TIME SIGNALS

ANTISYMMETRIC SEQUENCES

EVEN-SYMMETRIC SEQUENCES

11 INs TIT

NsINlEGER

Ns NONINlEGER

87

~J-LT-",---o.>

-.-J'LT1-...1-1-1,1c'-1

Ns

nGURE 3-8
Symmetric and antisymmetric sequences shown for integer and noninteger values of symmetry
point N s -

Fourier transform of even-symmetric sequences. Let us first consider the

Fourier transform of an even sequence, followed by that of a general
even-symmetric sequence. Starting with the definition of the Fourier transform, we have
H(eiw ) =

..

h(n) e- iWrl

n=-oo
-1

..

h(n) e- iwn

n=-oo

+ h(O) + L h(n) e- jWrl

n=1

Letting m = -n in the first sum, we obtain

H(e IOJ )

..

=L

h( -m) el<mt

m=1

h(n) e- iwn

n=1

Noting that h(n) = h( -n), we have

H(e iOJ )

..

+ h(O) + L

..

= h(O) + L

h(n)(e- iwn

+ elOJn )

n=1

= h(O)

+2 L

h(n) cos(wn)

(3.54)

n=1

Hence, the Fourier transform of an even sequence is a real-valued function of

frequency. The magnitude function is then
.
{H(eIW )
when H(eiw ) ~O
(3.55)
IH(eIOJ)1 = -H(eiw )
when H(eiOJ ) <0

88

INTRODUcnON TO DlOITAL SIGNAL PROCESSING

Since its imaginary part is identically zero, the phase function computed from
the arctangent function is zero. The phase response then only accommodates
the sign change of H(e iw ):

Arg[H(ei W )] = {O
Jr,

for
for

iw

(J)
(J)

over which H(e ) > 0

over which H(e iW ) < 0

(3.56)

Recall that the phase is not defined when H(e iW) = O.

Example 3.11. Fourier transform of an even sequence. Let
h(n) = 1

for -2:5 n :52,

=0 otherwise

Then its Fourier transform is given by

n--2

= 1 + 2 cos( w)

+ 2 cos(2w)

Any even-symmetric sequence can be obtained from an even sequence by

applying a suitable delay. The Fourier transform of. a delayed sequence was
shown above to have the same magnitude function as the original sequence,
but also to contain an additionallinear-with-frequency component in the phase
function. To illustrate this in a more general fashion, let us consider the even
sequence {hE(n)}, which is defined for -MS;ns;M, as shown in Fig. 3.9. If
{hE(n)} is the unit-sample response of a digital filter, then the filter is
noncausal because of the nonzero elements for n < o. If a causal filter having
the same magnitude response is desired, we need only add sufficient delay so
that the first nonzero element occurs at n = O. As shown in Fig. 3.9, we can
obtain a causal symmetric sequence {hs(n)} by delaying {hE(n)} by M
samples, or
(3.57)
for OS;n s;2M

{~(II)}
. .I

I
I III_
IIII

-M

I . .

-lOI2 M

{1Is(n)} = {h.!(n-M)}

..I

I II

I I I I II I .

-1 01

2M

FIGURE 3-9
n

Oclaying an even sequence to produce a causal symmetric sequen<:e.

FOURIER TRANSFORM OF DISCRETE-TIME SIGNALS

89

Then, if Hs(eJeo)is the Fourier transform of {hs(n)} and HE(eJW ) is that of

{hE(e jW )}, we have
(3.58)
and
(3.59)
Example 3.U. Fourier transform of an even-symmetric sequence. Let
h(n) = 1

for 0:5n:54

This is just the delayed version of the sequence given in Example 3.11. We can
get the Fourier transform by simply adding the linear phase function - 2w to the
result of Example 3.11 to include the delay.
We verify this result by directly computing the Fourier transform of the
above sequence, which is given by
4

H(e jW ) =

L e-

...

-0

Applying the finite geometric sum formula, we get

1- e- Jsw

H(e JUJ ) = - -JUJ

l_e- -

e - jS on ejSwi2 _ e - jSwi2
e ion

eiwi2 _

e iwi2

= sin(5w/2) e- J2UJ

sin(w/2)
The phase is a linear function of frequency with slope equal to -2. It is left as an
exercise for the student to show that the two amplitude expressions are
equivalent.

Recall that an even sequence has zero phase, except for the phase jumps
of .;rr caused by the sign reversal of the amplitude of the Fourier transform. If
these jumps are ignored, we are left with a phase that is linear with frequency,
which will be called a linear-phase function. This linear-phase property will be
important for filter design in the following chapters.

FOURIER TRANSFORM OF ANflSYMMETRIC SEQUENCES. We first consider the Fourier transform of an odd sequence, followed by that for the
general antisymmetric sequence. For an odd sequence, h(O) = 0, and the
Fourier transform can be written as
00

H(e iw ) = ~ h(n)[e-iwn - eieM ]

n=l
00

=-

j2 ~ h(n) sin(wn)
n=l

(3.60)

90

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

Hence, the Fourier transform of an odd sequence of n is an imaginary function

of frequency. Since only sine functions appear in the sum, H(e iw ) is an odd
function of 0, i.e. H(ei W ) = -H(e- 1CO ). Since the real part is identically zero,
and j = ei J<i2, we can write the function in terms of its imaginary part as

H(e iCO ) = jHI(eiCO ) = HI(eico) ei J<i2

The magnitude response is then equal to

IH(eiCO)1

= HI(e iCO )

for HI(eico ) > 0

=-H1(eico )

for HI(eico ) < 0

(3.61)

and the phase is equal to

for
for

Ar [H(eico) = {lC/2
g
lC/2 lC

0
0

over which HI(ei CO) > 0

over which HI(eico) < 0

(3.62)

Example 3.13. Maguitude and phase response of aD odd sequence. Let

-I
h(n) =

01

forn=-2
for n =2
othe~

The Fourier transform is equal to

H(e 1tu) = _e j2tu + e- j2tu = -2j sin(2w) = 2sin(2w) e-j"fl
The imaginary part is then equal to
H,( e itu) = -2 sin(2w)
The magnitude spectrum is given by
IH(e'"')1 = 12 sin(2w)1
and the phase spectrum is
Arg[H(el."')] = {lC/2

-lC/2

for H,(ei tu ) > 0,

for H,( eitu) < 0,

or for lC/2 < w < lC

or for 0 < w < lC/2

The magnitude and phase functions are shown in Fig. 3.10.

For odd sequences, a jump of lC degrees will always occur at 0 = O. This

happens because HI(eico) is an odd function of 0, which means that a sign
reversal in the amplitude function must occur at 0 = O.
An antisymmetric sequence can be obtained from an odd sequence by
applying a delay. Let us consider the odd sequence {ho(n)}, which is defined
for -M::::: n::::: M. We can obtain a causal antisymmetric sequence {hA(n)} by
delaying {ho(n)} by M samples:
hA(n) = ho(n - M)

for O:::::n :::::2M

(3.63)

Then, we have
(3.64)

fOURIER TRANSFORM OF DISCRETE-TIME SIGNALS

91

IH(eI'")1

'VV\
o

1tfl

1t

-,.

Arg[H(eJ")
1t.
1tfll-

I
1t

-lTfl
-IT'-

FIGURE 3-10
Magnitude and phase spectra for an odd sequence.

and
(3.65)

Ignoring the phase jumps due to sign reversals in the amplitude function,
including the one at (J) = 0, we are also left with a linear-phase function for
antisymmetric sequences. This result is true even when M is not an integer,
although a strictly odd sequence does not exist in this case. This is illustrated in
the next example.
Example 3.14. Fourier transform of aD aDtisymmetric sequence. In this example, we will cOnsider the case when Ns is not an integer. As shown in Fig. 3.11,
IH(e)")1
{h(n)}

-2 -1

~,

-IT

1t

FIGURE 3-11
Magnitude and phase spectra
for an antisymmetric sequence.

92

INTRODUCfION TO DIGITAL SIGNAL PROCESSING

let h(O) = -!, h(1) = 1, h(2) = -1, h(3) =! and h(n) = 0, otherwise. Then, the
Fourier transform is given by
H(ei"') = _! + e-J'" - e-J2", + !e- i3 ",
= -!(1- e- i3 ",) + (e- j'" - e- j2 ",)
= _e-i3WI2H(ei30i2 _ e- i30i2) + ei0i2 _ e- i0i2]
= -je- i3",I2(sin(3ro/2) + 2 sin(ro/2

Substituting -j = e- iJ1f2 , we finally get

H(e iW ) = (sin(3rol2) + 2 sin(ro/2e- i(3wu)/2

When Ns is not an integer, the phase spectrum still has a linear slope equal to
-Nsro. In terms of Eq. (3.65), M = ~.

3.8 THE INVERSE FOURIER

TRANSFORM
The inverse Fourier transform allows us to determine the unit-sample response
of a filter for which we know only the transfer function. Similarly, it also helps
us determine the output discrete-time sequence of a filter from the output
spectrum.
The Fourier transform H(e iw) is a periodic function of the continuousvalued variable (i), with period 2.1l'. Hence, it can be expanded as a Fourier
series. We show that the coefficients of the series terms are equal to the values
of the discrete-time sequence {h(n)}. We first state the Fourier series

argument:
Let s(t) be a periodic function of the continuous-valued variable t with
period T. Then s(t) can be expressed in the following Fourier series
expansion:
co

s(t) =

an e- iot

(3.66)

where On = 2.1l'n/T is the nth harmonic frequency, and the coefficients

are given by

an

=-1 fTI2
T

-TI2

s(t) eiO.,t dt

(3.67)

In exactly the same fashion, H(eiw) can be expressed as the following Fourier
series expansion
co

H(eiw )
where 8n

= 2.1l'n/2.1l' = n,

L en e- iBw
n=-co

(3.68)

can be thought of as a harmonic frequency,.and

(3.69)

FOURIER TRANSFORM OF DISCRETE-TIME SIGNALS

93

Note the similarity of the definition of the Fourier transform in Eq. (3.4) and
the Fourier series expansion of H(eiw ) given in Eq. (3.68). This similarity
allows us to define the inverse Fourier transform as
(3.70)
This integral solution for the inverse Fourier transform is useful for analytic
purposes, as shown later, but it is usually very difficult to evaluate for typical
functional forms of H(eiw ). An alternate, and occasionally more helpful,
method of determining the values of h(n) follows directly from the definition
of the Fourier transform. equal to
00

H(eiw ) =

2:

h(n) e-IW>l

n=-oo

If H(eJW) can be expressed as a series of complex exponentials, then h(n) is

simply the coefficient of e- jOft This is illustrated in the following example.
EDIIlple 3.ts. Inyerse Fourier transform of the 3-sample ayenger. Let
H(e'''') =

W + 2cos(w

Employing the Euler identity for the cosine we obtain

H(ei"') = i(1 + eiw + e- i"').
Noting that h(n) is the coefficient of e- i... , we have by inspection
h(O)=h(I)=h(-I)=t
h(n) = 0
forn<-landn>1
Alternately, we could perform the integration. The value of h(n) is then
h(n) = - 1
2.n=

ilf 3(1 + e'''' + e-'''') e' .... dw

1

_If

J.. ("

[ei'" + eiw(II+I) + e1w(,,-I)] dw

6.n-L.
_J.. (el -

- 6.nsin(n.n-)
3n.n-

e- i.... e i1i(n+l) - e-i 1i(" +I) e i1i(n-l) - e- i1i(n-I)

jn
+
j(n + I)
+
j(n -1)

= --+

sin[(n + 1).n-J sin[(n -1).n-J

+ ---='-'---~
3(n + 1).n3(n -1).n-

Recall that sin(n.n-) =0 for integer values of n. When the denominator is also
zero, an indeterminate form results. Applying L'Hospital's argument, we find
that

lim sin(E)/E = 1
.-0

94

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

Applying this approach to each of the three terms above, we obtain

hen) =

{t

for n = -1, 0,1

otherwise

H H(e iw ) is a rational function of e-i ""', then a series expansion can be

obtained either by performing a division or by computing the Taylor series

expansion, as illustrated in Example 3.16.
Example 3.16. Inverse Fourier transform for the first-order recursive filter. Let
H(e jOJ) = (1- a e-JOJ)-l
Computing the Taylor series expansion, we have
!:f(ejOJ) = 1 + a e- jOJ + a 2 e- j2 Q) + a 3 e- j3OJ + ... + a k e- Jwk + ...
Equating the coefficient of e- jOlk with h(k), we have

a"

h(n)= { 0,

for n 2:0
otherwise

Example 3.17 illustrates the application of frequency domain analysis to

signal processing, in which the output sequence of a filter is determined from
the output spectrum.
Example 3.17. DetermiDing the output sequence from
spectrum. The output spectrum of Example 3.10 is given by
Y(ei OJ)

the

output

=! ejQ) + 1 + ~-j..,
3 1-ae-JOJ

Expanding each term in the parenthesis in a Taylor series expansion, we get

Y(ejO/) = l[ejO/ + a + a 2 e-jO/ + a 3 e- j2 0/ + a 4 e- j3w + .. .
+ 1 + a e-jQ) + a 2 e- j2 0/ + a 3 e- j3 Q) + .. .
+

e-JO/+a e-j2OJ+a2e-J30/+ ... ]

Combining all the terms having the same complex exponential factor and
equating h(k) with the coefficient of e -jOlk, the output sequence is found to be
yen) = l[an+1u(n + 1) + a"u(n) + an-1u(n - 1)]

3.9 FOURIER TRANSFORM OF THE

PRODUCT OF TWO DISCRETE-TIME
SEQUENCES
In the design of nonrecursive filters, we commonly encounter the product of
two sequences {w(n)} and {h(n)}. Let the sequence of term-by-term products

FOURIER TRANSFORM OF DISCRETETIME SIGNALS

95

be denoted by {v(n)}, where

v(n) = w(n) h(n)

for all n

(3.71)

w(n) h(n) e- icm

(3.72)

Then its Fourier transform is given by

V(ei lil) =

"=-00
To determine a more useful form for V(eiW ), we first express h(n) as an
inverse Fourier transform given by
(3.73)
We thus get
(3.74)
Interchanging the order of summation and integration, we have

(i

V(eiO=~

(w(n)e-i(W-cr)n)H(eiJda
2Jf )-_ 11_-"

(3.75)

The term in the parentheses is similar to the definition of the Fourier transform
of {w(n)}, but at a frequency argument of co - a, or W(ei("'-cr. Including this
Fourier transform into the integral, we get
V(eio =

~ (- H(ei J W(ei(o>-cr da
2Jf 1-_

= H(eiO

W(ei"')

(3.76)

This form is that of a convolution operation involving two functions of the

con~uous-valued variable co. Because both these functions are periodic, the
integral is evaluated over one period. We denoted this operation as a circular
convolution. Hence, we have shown that multiplication in the time domain
corresponds to convolution in the .frequency domain. We will use ibis result in
the next section to determine the Fourier transform of sampled continuoustime data.
Eu.pIe 3.18. Fourier traDsfolDl of the prodac:t of two sequaaces. The most
common application of this result occurs in truncating infinite-duration sequences,
as illustrated here.
Let us consider the infinite duration sequence given by
h(n) -

Sin!:;~4)

for all II,

The corresponding H(eI"') is equal to the ideallowpass filter characteristic shown

Fig. 3.12. To truncate this sequence to -3 s n s 3, we multiply h(n) by the

in

96

INTRODUcnON TO DIGrrAL SIGNAL PROCESSING

... 011. 'II

OJ

rlrlhl. 11' .'''....

1~]lllllL. .
(~" I ,,")hi"))

II r

.. .I-3 I 0

-3

HC'")

It. ..
3

FIGURE 3-12
Multiplication of two discrete-time sequences with the corresponding spectral convolution
operation.

following sequence, called a window function,

w(n) =

{~

for -3::s n ::s 3

otherwise

The windowed sequence is then

v(n)

= w(n) h(n) =sin(:rn/4)/(:rn/4)

for -3::s n::s 3

V(ei "'),

we must first compute W(e;") and

and is shown in Fig. 3.12. To find
perform the convolution. The Fourier transform of {wen)} is equal to
3

W(d"') =

2:

e- j .... = ePO>

n--3

. 1 - e- j7 .,
=eI3 ..
.
1 - e I'"

2: e-;'"
n-O

sin(3w/2)
sin (w /2)

The result of the convolution is shown in Fig. 3.12. The main effect of the
convGlution is that the sharp transitions in H(ej .,) have been made smoother. We
will explore this multiplication with a window sequence in more detail in the
design of finite-impulse response filters in a later chapter.

3.10 SAMPLING A CONfINUOUS

FUNcrION TO GENERATE A
SEQUENCE
In practice, discrete-time sequences are often obtained by sampling
continuous-time signals. Typically, this occurs when signals are to be processed

FOURIER TRANSFORM OF DISCRETETIME SIGNALS

97

by computer; situations include speech analysis and extracting information

from sensors for bioengineering and robotic applications. We will apply these
results to sampling the Fourier transform in the next chapter. An application
of sampling continuous-time signals will also occur in a later chapter in the
design of digital filters. In this section we relate the spectrum of the
continuous-time signal to that of the discrete-time sequence that is obtained by
sampling.
We proceed by first stating the relationship between the continuous-time
and discrete-time spectra and then provide the rather involved mathematical
proof. Let us define the continuous-time, or analog, signal by CA(t). Its
spectrum is then given by CAUQ), where Q is the continuous-time frequency,
having units radians per second. For continuous-time signals, the Fourier
transform is defined by
(3.77)

which is similar to the definition of the Fourier transform of discrete-time

signals, but with the sum replaced by an integral. When CA(t) is sampled every
T. seconds, we obtain the discrete-time sequence {cA(nT.)}, whose values are
equal to
(3.78)

T. is commonly called the sampling period and has units seconds per sample
interval. The discrete-time Fourier transform for such sampled data sequences
is defined by
(3.79)
n=-oo

where w has units radians per sample interval. This discrete-time spectrum is
related to the continuous-time spectrum CA(jQ) by
C(e

iw

= ~ k~OO CAOw/T. + j2kn/T.)

(3.80)

In words, the spectrum of the discrete-time sequence, derived by sampling a

continuous-time function, is equal to the sum of an infinite set of shifted
continuous-time spectra. A graphical interpretation of this identity is shown in
Fig. 3.13.
Proof. To show this relationship, we take the following approach. First, we
treat sampling as a multiplicative operation involving the continuous-time
signal CA(t) and a train of Dirac delta functions occurring every T. seconds.

This multiplication produces the intermediate function of continuous-time

cs(t). The continuous-time spectrum of cs(t), denoted by CsOQ), is shown to
have the same analytic form as Cs(eiW ). Finally, we determine CsOQ) in terms
of the original analog spectrum CA(jQ).

98

INfRODUcnON TO DIGITAL SIGNAL PROCESSING

f\
-:bE
Ts

-It

Ts

7t

2r.

Ts

Ts

co,n

FIGURE 3-13
Spectrum of a discrete-time sequence obtained by sampling a continuous-time signal.

We first define the Dirac delta function and present its properties that are
important for this application. By definition, the Dirac delta function, denoted
by c5(t), has infinite height and zero width, but its area, or weight, is equal to
unity. To illustrate how we can approach a Dirac delta function, consider the
rectangular function r.(t), whose width is E and whose height is lIE, so its area
is equal to one, as shown in Fig. 3.14. We can define c5(t) as
c5(t) = lim rE(t)

(3.81)

....0

The following three properties of the Dirac delta function are useful
here.
Property 1. The product of a function and a delta function. In multiplying any continuous-time function f(t), the weight of the delta function
becomes equal to the value of the function at the instant of the delta function,
or
f(t) c5(t - to)

= f(t o) c5(t -

to)

(3.82)

Property 2. The integral of the product (The sifting property). The

integral of the product of any function f(t) and a delta function is a constant,
equal to the value of the function at the instant of the delta function. To show
this, we integrate the product to give
fJ(t) c5(t - to) dt = f./(t o) c5(t - to) dt

,.(1)

II

-'h

'h

o
1"==0

FIGURE 3-14
Limiting proCedure to obtain a Dirac delta function.

(3.83)

FOURIER TRANSFORM OF DISCRETETIME SIGNALS

99

by Property 1. Since f(t o) is not a function of t, we can take it out of the

integral to give
f(t o) L"'", o(t - to) dt = f(t o).

This last equality follows because the area of o(t) is unity. The Dirac delta
function is used to select, or sift, the value of the function at t = to
Property 3. Convolving a function with a delta function (The shifting
property). When the function f(t) is convolved with o(t - to), the origin of
f(t) is shift to t = to, or
o(t - to) * f(t)

= f(t -

to)

To show this, we write out the continuous-time convolution equation,

(3.84)

['" 0(.\ - to)f(t -.\) cIA

and define the variable =.\ - t. We then have

[",f(-) 0(<< + t - to) d = f(t - to)

(3.85)

This last result follows by Property 2.

An infinite train of Dirac delta functions occurring every
denoted by or;(t) and can be expressed as
or,(t) =

'"

L o(t n=-oo

nT.)

T. seconds is
(3.86)

We can now use these properties of the Dirac delta function to determine
the relationship between the spectra. The sampling process is modeled by the
multiplication operation given by
for all

= cA(t)

L'"

o(t - nT.)

for all t

(3.87)

n=-CXl

as shown in Fig. 3.15. Since cs(t) is a continuous-time function, its spectrum

Cs(jQ) is given by
. Cs(jQ) = ['" cs(t) e-

[J

cA(t)

jQt

dt

n~'" o(t -

(3.88)

nT.)] e-

jQt

dt

(3.89)

Changing the order of integration and summation, and applying the sifting

100

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

FIGURE 3-15
Sampling a continuous-time signal can be treated as a multiplicative process with a train of Dirac
impulses.

property of the Dirac delta function, we have

n~.., [f.., 6(t -

Cs(jQ) =

nT.) CA(t) e-

jDt

dt]

...

2:

cA(nT.) e-jnCT,

(3.90)

We have arrived at our first intermediate result. By

(3.79), we observe

compari~g

Eq. (3.90) with

(3.91)

That is, the form of C(eiQ) can be obtained by determining CsOQ) and
replacing Q by (J) I T.. The division of (J) by T. scales the values of (J) to be the
same as Q. For sampled-data applications, it is conventional to express w in
the range [-niTs, nlT.), to have the same numerical values as Q. The units of
wiT. are radians per second, just as for the continuous-time frequency.
To complete the proof, we must determine Cs(jQ) in terms of the
continuous-time spectrum CA(jQ). Since 6r.(t) is a periodic function, it can be
expressed in the following Fourier series expansion given by
..,

6 T ,(t) =
where Q n

= 23m1T., and

i
T.

2: an e-jc
n=-oo

T
,12

an = -1

(3.92)

6 T,(t) ejQ.r dt

(3.93)

-T,12

Over the interval - T./2 :$ t < T./2, the impulse train 6r.(t) is represented by
only the single Dirac impulse situated at t = 0, or 6(t). Making this substitution
and applying the sifting property, we get

i
T.

1
an =-

T
,12

-T,12

1 j o=1
6(t) dQ.rdt=-e
.

T.

T.

for all n

(3.94)

FOURIER TRANSFORM OF DISCRETE-TIME SIGNALS

101

Substituting this Fourier series expansion into the sampling process, we get

cs(t) = CA(t) . ar,(t)

1
= CA(t) . -

=-

oo
L

L
00

T. n=-oo

Tan=-cc

e- IQ t

CA(t) e-IQ t

(3.95) .

The continuous-time spectrum of cs(t) is then equal to

CsGQ) = roo Cs(t) e-

jQt

dt
(3.%)

Interchanging the order of summation and integration, we obtain

Cs(jQ) =

~ n~oo [f"oo CA(t) e-j(Q+Q)t dt ]

(3.97)

Evaluating the kth term of the sum, denoted by CS,k(jQ), we find

CS,k(jQ) =

L:

CA(t) e-j(Q+Qk)t dt

= CAOQ + jQk)

(3.98)

This result is simply the original analog spectrum that has been shifted from
0=0 to 0 = -Ok, as shown in Fig. 3.16. The total spectrum is then the sum
of these shifted versions and is given by
(3.99)

(= -2ITk)
T,

FIGURE 3-16
Shifted version of the analog spectrum.

102

INTRODucnON TO DIGITAL SIGNAL PROCESSING

Cs(jQ) is a periodic function of Q, with period 21f/T.. More concisely, Cs(jQ)

is called the periodic extension of CAOQ).
.
This last identity is the second desired intermediate result. Combining
this result with that given by (3.91), we arrive at the final desired identity given
in (3.80).
(Q.E.D.)
NYQUIST CRITERION. When sampling a continuous-time signal cA(t) to
produce the sequence {cA(nT.)}, we want to ensure that all the information in
the original signal is retained in the samples. There will be no information loss
if we can exactly recover the continuous-time signal from the samples. To
determine the condition under which there is no information loss, let us
consider CA(t) to have a bandlimited spectrum, or one for which
(3.100)
as shown in Fig. 3.17(a). As demonstrated above, when CA(t) is sampled with
sampling period T., then the spectrum of the sampled signal CsOQ) is the
periodic extension of CA(jQ) with period 21f/T., as shown in Fig. 3.17(b). The

(b)

T, < 7T/OM

C (j11)

Ideallowpass filter

_"' ' _~. .J.:- ~-,- r- I<: ~ZS:_- _-" "*-_-_- -:-! -~-L-i-LL_~'_'-+~
- 211

IS

-11
--

Ts

11"

11"

-11

Ts

211

Ts

FIGURE 317
Relationship between continuous-time spectrum of a signal and the spectrum of the discrete-time
sequence obtained by sampling the signal with sampling period T.. (a) Original spectrum of
continuous-time signal; (b) spectrum of sampled sequence when 11M < nIT.; (c) spectrum of
sampled sequence when Q M > nIT.. The latter case illustrates the aliasing error.

FOURIER TRANSFORM OF DISCRETE-TIME SIGNALS

-m

-1[

Ts

IS

1[

21[

IS

IS

CAoy,

-1[

1[

IS

IS

103

~(jn)

U
-

IS

- n
Ts

FIGURE 3-18
Two candidates for the continuoustime spectrum when aliasing occurs_

form of Cs(jQ) in the frequency range [-.nIT., .nIT] is identical to CA(jQ) if

or

(3.101)

In this case, there is no overlap in the adjacent spectral components and we

can then extract CA(jQ) exactly from Cs(jQ) by employing the ideal lowpass
filter magnitude response shown in Fig. 3.17(b).
If T. is chosen to be greater than .n/QM , spectral overlap occurs in the
periodic extension, given by Eq. (3.99), and the form of Cs(jQ) in the range
-.nIT. $. Q < .nIT. is then no longer similar to CAOQ) as shown in Fig. 3.17(c).
This overlap, caused by sampling at too Iowa rate, produces an irretrievable
error in the spectral values, called aliasing. The true spectral shape is
irretrievable since many different CAOQ) functions can produce the same
Cs(jQ). Two possible candidates are shown in Fig. 3.18.
A useful sampling measure is the sampling rate, Is = liT.. Recalling that
Q = 2.n1, and defining 1M as the highest frequency component in the signal,
then no spectral overlap will occur if
(3.102)
To prevent aliasing error, more than two samples are required per period of
the highest frequency sinusoidal component present in the signal. The smallest
sampling rate before aliasing occurs for a particular continuous-time signal is
called the Nyquist rate. The effect of aliasing is most easily demonstrated by
considering a sinusoidal signal in Example 3.19.
Example 3.19. Demonstration of aliasing. Let cA(t) = cos(Qot), for all t, with
Qo = lOOn. This function is sampled to produce the sequence
cs(n 7;) = cos(nQo7;)

for all n

104

INTRODUCl10N TO DIGITAL SIGNAL PROCESSING

(b)

(a)

CA(j!l)

o 'lo

It

T,

k' J
\

-It

T,

~21T

/ 'lo

2!..
T,

C,(eJOl)

C,(el"')

-21T

I
0

2.. - 'loT 0 'loT IT

~~\----~--~~;IT

-+'loT
T,

I T~IT I

21T

--'loT - +'loT
T,
T,

--------

FIGURE 3-19
Demonstration of aliasing caused by choosing too large. (a) Unaliased version, when < 11:/0.0 ;
(b) aliasing caused when > 11:/0.0 In this latter case a low-frequency cosine function is apparent
when the sequence is interpolated.

r.

r.

r.

The continuous-time spectrum of CA(t) is equal to two Dirac delta functions

as shown in Fig. 3.19. Sampling CA(t) causes the spectrum to be extended

periodically:
Cs(jO) =

2;',,~~ [6(0 -

0 0 - 2nK/T.) + 6(0 + 0 0 - 2nKIT.)]

To demonstrate aliasing we note the position of the impulses situated

within -KIT. ~ 0 < KIT., as we vary the sampling period T.. The sampling period
dictated by the sampling rate is T. < KIOo(=10-2s).
For T. < KIOo, say T. = 10- 3 s, the delta functions falling within -KIT. ~
0< KIT., or ....:l000K ~ 0 < 1000K, are only those for the n = 0 term above.
These occur at 0 = Oo. In the discrete-time domain, these impulses will
produce the desired sequence {cos(nOoT.)}.
For T.> KIOo, say T. = 1.1 X 10- 2 s, the impulse in the range 0< 0 < KIT.,
or 0 < 0 < 9OK, is at
0= (2KIT.) - 0 0 = 180K -lOOK = 80K rad S-1

from the term 6(0 + 0 0 - 2KIT.), or the k = 1 term. The impulse located in the
range -KIT. < 0 ~ 0 is at
0= 0 0 - 1K/T. = lOOK -180n = -BOn rad S-1

FOURIER TRANSFORM OF DISCRETE-TIME SIGNALS

105

from b(g - go + 2nllT.), or the k = -1 term_ In this case, the sampled sequence
appears to be derived from a continuous-time function at a lower frequency
2n/T. - go = BOn rad S-I_ This is the aliasing phenomenon_

This effect of aliasing errors is occasionally observed in films of moving

cars in which the wheels appear to be turning in the direction opposite to that
expected from the motion of the car.
ANfI-ALIASING FILTER. In practice, the frequency range contammg sig-

nificant power in the original continuous-time signal may be larger than that
containing the desired information. This commonly occurs when a lowfrequency signal is corrupted by high-frequency noise or in modulated signals,
such as broadcast radio. If sampled at the rate dictated by the Nyquist criterion
for the desired analog signal, unwanted high frequency signals would cause
aliasing errors to occur.
To prevent aliasing errors caused by these undesired high-frequency
signals, an analog lowpass filter, called an anti-aliasing filter, must be used.
This filter is applied to the analog signal prior to sampling and reduces the
power in the analog signal for the frequency range beyond Q = niT.. In
practice, the spectral magnitude level for Q ~ JrIT. should be less than 1%
(-40 dB) of the important spectral features in the desired signal spectrum to
prevent significant aliasing.
For example, let the observed analog signal contain power in an
arbitrarily large frequency range, but have relevant information only in the
frequency range -Q R $ Q $ QR. We wish to determine the appropriate
sampling period T.. The Nyquist criterion tells us that T. must be less than
Jr/QR. If sampled at this rate, the higher frequency components in the analog
signal will be aliased into the relevant discrete-time signal. The anti-aliasing
filter must satisfy two conditions:
Condition 1. The analog signal components with frequencies less than
must be negligibly attenuated by the filter.
Condition 2. To prevent aliasing in the relevant analog frequency range,
the components for IQI ~2JrIT. - QR must be attenuated. After sampling, the
components in this frequency range fall into the range -Q R $ Q $ Q R when
the periodic extension is formed.
QR

We consider how these conditions can be met in Chapter 8, where the design
of analog lowpass filters is discussed.

3.11 RECONSTRUcnON OF
CONTINUOUSTIME SIGNALS FROM
DISCRETETIME SEQUENCES
In many practical applications, discrete-time sequences are required to be
transformed into continuous-time signals. One example is the digital audio

106

lNTRODUcnON TO DIGITAL SIGNAL PROCESSING

CS<jO)

D
27t

-T,

/:\ D
I

_~-OM

T,

OM

7t

27t

T,

T,

H,(jO)
I 1""1- - - . ; . ; . : . . . . . ; . . . . . - - - ,
1
I

L
0

I
I
I

7t

o
FIGURE 3-20
The spectrum of an analog signal can be obtained by multiplying the spectrum of the sample values
with the transfer function of an ideal analog lowpass filter. In practice, H,(jQ) must be
approximated.

compact disk system, in which the audio signal is stored as digital samples on
the disk and must ultimately be converted into the analog signal that drive the
speakers. Another is the digital controller in a robot system, that uses the
digital computer to process the sensor data, but requires an analog signal to
control the actuator motors. In these cases, we are faced with the converse
problem from that of sampling CA(t) to obtain {cA(nT.)}. The relevant
question now becomes how to recover the analog signal CA(t) from the
sampled sequence {csCnT.)}.
We begin by considering the unaliased spectrum of Cs(jQ), shown in Fig.
3.20. CA(jQ) then has the same form as Cs(jQ) over -rclT.:5 Q < rclT.. Let us
consider an analog lowpass filter with the following ideal frequency response
for -rclT.:5Q<rclT.
otherwise

(3.103)

We could retrieve CA(jQ) from Cs(jQ) by filtering with an ideallowpass filter,

or
for all Q

(3.104)

To interpret this result in the time domain, let us examine the corresponding
convolution equation given by

CA(t) =

L"'", cs(c)hICt -

c) dc

(3.105)

The function csCt) is the C<?ntinuous-time sampled signal composed of a series

<

FOURIER TRANSFORM OF DISCRETETIME SIGNALS

107

Approximation to C,(t)

"/

...

FIGURE 3-21
Approximating the sequence of weighted Dirac
delta functions with a sequence of narrow pulses
having different amplitudes.

_ _~--l~~_--':'-_u....._. n T,

of weighted Dirac delta functions. In practice, this continuous-time function

could be approximated by narrow pulses whose amplitudes are proportional to
the corresponding sample values cs(nT.), as shown in Fig. 3.21- The function
hI(t) is the continuous-time impulse response of the ideallowpass filter, found
by the inverse Fourier transform given by

=----j2m

~ Sinc(tlT.)

1 sinemIT.)
lttlT.

= T.

(3.106)

The Sinc(x) function is shown in Fig. 3.22. The function is equal to zero for
integral values of the sampling time, i.e. t = nT., where n is a non-zero-valued
integer. For t = 0, we find
hI(O) = lim.!. sin(mlT.)
1-+0

T.

:Jrt I T.

= .!.lim (mIT.) T. 1-+0

(1/3!)(lttlT.)3 + ...
Jrt I T.

lIT.

(3.107)

Because cs(t) is composed of a series of weighted Dirac delta functions,

the convolution produces a sum of shifted and scaled versions of hI(t), as
demonstrated by Property 3 of the Dirac delta function. As shown in Fig. 3.23,

~.
-3T-2

-T

2T 3T

4T

FIGURE 3-22
The Sinc(tIT) function, used
for interpolation, is the impulse response of the ideal
analog lowpass filter.

188

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

ftGURE 3-23
In the time domain, the interpolation
process is treated as the convolution of
weighted and delayed Sinc(t/T) functions.
The values of the sequence {cs(nT.)} are
shown as dots. The continuous-time
function is shown in heavy solid line.

the original analog signal can be reconstructed from the samples by

oo

CA(t) =

L cs(nT.) hI(t n=-co

=.!.

T.n=-oo

nT.)

cs(nT.) sin(.n-[t - nT.]IT.)

.n-[t-nT.]IT.

(3.108)

The action of hI(t) in this convolution is to fill in, or interpolate, the

values of the continuous-time function between the sample points. Such
functions are called interpolation functions. Because of the infinite limits of the
Sinc(x) function, it is awkward to apply practically. In practice, the ideal
interpolating filter must be approximated by a causal lowpass analog filter.
PRACI1CAL INfERPOLATION FILTERS. The problem with the ideal interpolation function hI(t) is that it extends in time from minus infinity to plus
infinity. Hence, it incorporates the entire sequence, all the future as well as all
the past samples, into the computation of the interpolated function at any time
point. Especially in control situations, we cannot usually wait to observe the
entire sequence before an interpolation is performed. In this section we
consider some of the properties that practical interpolators should have and
provide examples of two commonly employed interpolators.
Let us consider the impulse response of a practical continuous-time
interpolating filter to be given by </>(t). In practice </>(t) should have the
following three properties:
(1)

(2)

(3)

</>(0) = 1
</>(nT.) = 0

for n *0

foo I</>(t)1 dt <

(3.109)

00

The first two properties permit the interpolated function to equal the sample
values at the sampling instants. Property (3) guarantees that the interpolated
function remains finite between the sampling instants. Example 3.20 considers
two commonly employed interpolation functions that have these properties.

FOURIER TRANSFORM OF DlSCRETETIME SIGNALS

109

1;

~(t)

-T,

nGURE 3-24

1;

Two commonly used interpolation functions.

Eumple 3.20. Zero-order hold and linear iDterpolaton. The impulse response
of the zero-order hold interpolator 4>o(t) is given by

4>o(t) =

{~

forO~t<T.

otherwise

That of the linear interpolator 4>1(t) is given by

for -T~t~T
otherwise
These interpolating functions are shown in Fig. 3.24. Notice that both these
functions have the three desirable properties.
Examples of interpolated functions are shown in Fig. 3.25. The zero-order
hold is very easy to implement by quickly charging and discharging a capacitor at
the sampling instants, but the interpolated function CAO(t) exhibits discontinuities

T.

nGURE 3-25
Results of applying the two interpolation functions to the same
sequence.

110

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

at the sampling times. These discontinuities may be filtered by the mechanical

inertia of the syst!,!m that is driven with the signal.
The linear interpolator produces a smoother function CA1 (t). But this
interpolation is noncausal, since it must look ahead to determine the slope value.
To make it causal, a delay can be introduced so that the interpolated function lags
behind the samples by one sampling period. In control systems, this delay is
crucial and may restrict the use of the linear interpolator.

3.12 SUMMARY
In this chapter, we presented the frequency-domain procedures for analyzing
digital filters. The Fourier transform of the unit-sample response provided the
frequency filtering characteristics of the digital filter, while the Fourier
transform of a data sequence indicated its frequency content. The Fourier
transform was shown to be a periodic function of w, with period 2Jr, and for
real-valued sequences to have a magnitude function that is an even function of
frequency and a phase function that is odd. This allows the entire function to
be determined from the values for O:s; w :s; Jr. Even or odd real-valued
sequences were shown to have purely real or imaginary Fourier transforms,
while sequences with symmetries have linear phase functions. The operations
of convolution and multiplication were shown to be converted into the other
when the Fourier transform was computed. The inverse Fourier transform,was
defined through a Fourier series argument, but found to be complicated to
apply directly. An alternate method involving a Taylor expansion was found to
be simpler, but did not provide an analytic form for the time series. The
relationship between continuous-time and discrete-time frequency functions
was discussed and so were procedures to convert a continuous-time function
into a discrete-time sequence. This chapter ended with a short discussion on
reconstructing the analog signal from the sampled sequence by using an
interpolating filter. In the next chapter, we will investigate the signal
processing implications of evaluating the Fourier transform at a finite number
of points by computer.

FURmER READING
Antoniou, Andreas: Digital Filters: Analysis and Design, McGraw-Hill, New York, 1979.
Bracewell, R. N.: The Fourier Transform and Its Applications, 2d ed., McGraw-Hill, New York,
1978.
Childers, D. G. (ed.): Modern Spectral Analysis, IEEE Press, New York, 1978. [This is a book of
reprints which presents a historical perspective of the evolution of spectral analysis.]
Gold, Bernard and Charles M. Rader: Digital Processing of Signals, McGraw-Hill, New York,
1969.
Hamming, R. W.: Digital Filters, 2d ed., Prentice-Hall, Englewood Cliffs, NJ, 1983.
Jong, M. T.: Methods of Discrete Signal and System Analysis, McGraw-Hill, New York, 1982.
Oetkin, G., T. W. Parks, and H. W. Schussler: "New results in the design of digital
interpolators", IEEE Trans. ASSP ASSp23: 301-309 (1975).
Oppenheim, Alan, V., and Ronald W. Schafer: Digital Signal Processing, Prentice-Hall,
Englewood Cliffs, NJ, 1975.

FOURIER TRANSFORM OF DISCRETETIME SIGNALS

111

Papoulis, A.: The Fourier Integral and its Applications, McGraw-Hill, New York, 1962.
Papoulis, A.: Signal Analysis, McGraw-Hili, New York, 1977. [This book contains the
continuous-time results referred to in this chapter.)
Peled, A. and B. Liu: Digital Signal Processing: Theory, Design and Implementation. Wiley, New
York,1976.
PhiIlips, C. L. and'H. T, Nagle, Jr.: Digital Control Systems Analysis and Design, Prentice-Hall,
Englewood Cliffs, NJ, 1984.
Rabiner, Lawrence R. and Bernard Gold: Theory and Application of Digital Signal Processing,
Prentice-Hall, Englewood Cliffs, NJ, 1975.
Schwartz, Mischa; Information, Transmission, Modulation and Noise, McGraw-Hill, New York,
1980. [This book contains the continuous-time results referred to in this chapter.)
Schwartz, Mischa, and Leonard Shaw: Signal Processing: Discrete Spectral Analysis, Detection and
Estimation, McGraw-Hill, New York, 1975.

PROBLEMS
3.1. Given the following unit-sample response sequences compute the Fourier
transform and sketch the magnitude and phase functions.
(a) h( -1) = 1, h(l) = 1, h(n) =0, otherwise.
(b) h(O) = 1, h(l) = -1, hen) = 0, otherwise.
(c) h( -1) = 1, h(O) = -2, h(l) = 1, hen) = 0, otherwise.
3.2. Determine whether the following sequences have Fourier transforms and
compute the Fourier transform when it exists.
(a) hen) = 1, for n ~O, = 0, otherwise.
(b) h(n) = 1, 0:5 n :5 N, = 0, otherwise.
(c) hen) = 1, for all n.
(d) h(n)=1, for -N:5n:5N, =0, otherwise.
3.3. An important theorem of signal processing is Parseva/'s theorem, that can be
stated as
~
1
"~~ x(n)x*(n)=2n -n X(e1W)X*(e1W)dw

fn

Prove this theorem by substituting the definition of X(e 1W ) into the integral and
evaluating.
3.4. Show that a causal filter that has more than two nonzero elements in its
unit-sample response cannot have a phase response that is zero for all
frequencies.
3.S. Consider the sequence {g(n)} that is the time-reversed version of {hen)}, or
g(n)=h(-n)

for all n

(a) Let the H(e1W ) be the transfer function of the filter whose unit-sample
response is {hen)}. Compute the transfer function G(eiW ) in terms of H(e 1W ).
(b) Consider taking the output of the filter described by H(eiw ) and letting it be
the input to a second filter having the transfer function G(e iw). The total
transfer function is then G(e iW) H(e~'). What are the magnitude and phase
responses of this cascade combination?
3.6. For the following difference equations, determine the transfer function of the
filter and sketch the hlagnitude and phase responses.
(a) yen) = x(n) - x(n - N). Let N = 4.
(b) yen) = 2r cos(O) yen - 1) + r2y(n - 2) + x(n). Let 0 = n/4 and r = 0.9.

ill

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

3.7. For the following magnitude and phase functions, determine and sketch the
unit-sample response, the digital filter implementation and the difference
equation.
(a) IH(ei"')1 = [20 + 16 cos(2w WI2,
j

Arg[H(e ,")] = arctan (2

:S:~~;2 )

(b) IH(ej'")1 = [17 - 8 coS(2W)]II2,

4 sin(2w) )
Arg[H(el."')] = arctan ( --~~
1-4cos(2w)
(c) IH(ej"')1 = [1.81-1.8cos(w)]1I2,
. ]
( 0.9sin(w) )
Arg[H (e1"') = -w + arctan 1 _ 0.9 cos(2w)
3.8. Show that the two different analytic forms for the magnitude response in Example
3.9 are equivalent.

3.9. Some computer languages provide only the ATAN(Y/X) function, which
produces the radian angles in the range - : /2 :5 w :5 : /2. Describe a procedure
to determine the principal value of the arctangent over the range -::5 w < :,
from the result of the ATAN(Y/X) function and the values of X and Y.
3.10. If we wanted to generate the sequence
x(n) = cos(won/lO),

for all n

by sampling the continuous-time signal

X(I) = COS(Ool)

for all 1

where Wo = 0 0 1;. What is the value of the sampling period 1;?

3.11. Exploring tbe aliasing phenomenon. The follQwing continuous-time Fourier
transform pairs for the sinusoids of infinite duration are given by:
if

U(I) = COS(Ool)

then

U(jO) = (:12)[c5(0 - 0 0 ) + c5(Q'+ 0 0 )]

if

V(I) = sin(Ool)

then

V(jO) = (j:12)[ c5(0 - 0 0 )

and
-

c5(0 + 0 0 )]

Determine and sketch the real and imaginary parts. of the spectra Us(jO) and
Vs(jO) for the continuous-time sampled functions for the following three sampling
periods:
(i) T. < : 100
(ii) 1;>>>:100
(iii) 1; = :100
3.U. SampUng. bandlimited signal. Let CA(I) have a spectrum CA(jO) that is nonzero
only over the frequency range 0 1 :5101:5 O 2 , as shown in Fig. P3.12. It can be
shown that cA(I) can be sampled at a rate that has to be greater than twice the
bandwidth, O 2 - 0 1 , rather than twice the highest frequency O2 , and yet have no
information loss. This can be demonstrated by doing the following steps.
(a) Let 1; = 2:1(0 2 - 0 1), Determine and sketch the periodically extended
spectrum of the time series.

FOURIER TRANSFORM OF DISCRETETIME SIGNALS

f'1

---.. .uo,....----'o,....---:!!--~U-~---.;.~n
-fi2 -fi]

fi]

fi2

113

FIGURE P3-U
Bandpass spectrum of continuoustime signal.

(b) Sketch the magnitude response of the bandpass filter that retrieves the
desired signal spectrum.
(c) The impulse response response of this filter acts as the interpolation function.
Assume the magnitude response of part (b) and a zero phase response to find
the impulse response. Sketch this continuous-time function. Consider
whether this waveform is reasonable by considering its values at the sampling
instants.

COMPUTER PROJECTS
3.1. High frequency sequences
Objed. Observe the shape of sinusoidal sequences for frequencies in the range
:re/2 < (J) <:re, since they can often appear to be confusing.
Computer results. Use TESTO to obtain plots of either the sine or cosine sequence
for the following values of frequency: (J) = :re/2, 3:re/4, 7:re/8.
Aoalytic results. Determine and sketch the corresponding continuous-time sinusoid that connects the dots of the computer plot.
3.2. Computer evaluation of the Fourier transform
Objed. Write and verify the program that evaluates the Fourier transform of a
real-valued sequence at a specified number of frequency points in the range
Os (J) < 21&.
Analytic results. Determine and sketch the magnitude and phase spectra of the
sequences given below and compare them with the computer results. In all cases,
evaluate the sum in the Fourier transform.
Computer results. Write the subprogram

FT(H.NH.HR.HI.NFT)
where H is the input array containing the real-valued discrete-time sequence of
length NH, HR and HI are the output arrays of size NFT containing the real and
imaginary values of the Fourier transform, with the Kth element corresponding to
the frequency point (J)K = 21&(K - 1)/NFT, for 1 S K S NFT. For this project, let
NFT = 128, which will allow comparison with the fast Fourier transform in the
next chapter.
A clever way of obtaining the value of :re in the program is by using the
arctangent function: PI = 4. *ATAN(1.).
To compute the magnitude and phase spectral sequences, write

MAGPHS(HR. HI. HMAG. HPHS.NFT)

114

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

which converts the input real HR and imaginary HI arrays of the Ff of size NFf
into the output magnitude HMAG array and phase HPHS array. Use the
FORTRAN function ATAN2(XI, XR), where XI is the imaginary component
and XR is the real component, to calculate the principal value of the phase, i.e.,
in the range [-:re, :re). Undefined values of the phase, when HR(K) = HI(K) = 0,
should be set to zero.
To plot the magnitude and phase sequences, use the PLOTMR and
PLOTPR subroutines given in Appendix A. These plot the 65 elements of
HMAG and HPHS that correspond to the range 0!5 W !5 :re.
Compute the spectra for the following sequences.
(a) h(O) = t h(l) =~, h(2) = 1, h(3) = t h(4) = land h(n) = 0, otherwise.
(b) The unit-sample response of first-order recursive filter for the following
values of a: 0.8, 0.95, -0.8, and -0.95 (compute first 128 elements).
3.3. Becoming familiar with the phase spectrum
Object. Illustrate the jumps in the phase response curves.
Analytic results. Determine and sketch the phase spectra of the sequences given
below and compare them with the computer results. In all cases, evaluate the sum
in the Fourier transform. Indicate the phase jump location and type on "the
sketches and computer plots.
Computer project. Use the programs Ff and MAGPHS written in Project 3.2 to
compute the phase and plot with the subroutine PLOTPR given in Appendix A.
Use NFf = 128.
(a) Jumps of 2:rc in the phase spectrum. Compute the phase spectrum of the
delayed unit-sample sequence {d(n - k)} for k = 2. Repeat for k = 3 and
k = 10. Sketch the linear phase functions on the plots.
(b) Jumps of :re in the phase spectrum. Compute the phase spectrum of the
following sequence: h(O) = h(l) = h(2) = 1, and h(n) = 0, otherwise. /
(c) Both jumps of :re and 2:re. Compute the phase spectrum for the following
sequence: h(n) = 1 for 1 !5 n !5 5, and h(n) = 0, otherwise. Indicate the causes
'Of the jumps.
3.4. Demonstration of aliasing
Object. Demonstrate the aliasing phenomenon both analytically and numerically.
Analytic results. Do Problem 3.11 and compare the results with those produced
by the computer.

Computer results. For this project use the TESTO.FOR program in Appendix A
to generate both the sine and cosine sequences. Rather than varying the sampling
period with respect to the frequency of the sinusoid, we will keep the sampling
period constant, equal to 1, and vary the frequency. Demonstrate the aliasing
effect by plotting the time-domain sequences which result when the frequency 0 0
is varied to correspond to the three cases in Problem 3.11.
3.S. Frequency domain interpretation of interpolating sequences
Object. Explore the frequency-domain behavior of two discrete-time interpolation
functions.
"
Consider the sequence
v(n) = cos( won)

for 0!5 n s;NV -1

FOURIER TRANSFORM OF DISCRETE-TIME SIGNALS

115

V(e JW )

---O~--~W~o--------------------~n--+W

X(e JW )

FIGURE Proj 3-5

o
with

Cl.Io=

IT

Spectra for Project 3.5_

n/16 and NY = 128 and x(n) such that

x(n) =v(n)
x(n) =0

for n even, 0:5 n :5 NY - 1

for n odd, O:5n:5NY - L

The idea is to determine the values of {x(n)} that were zeroed, by employing
interpolation functions as was done in ProJect 2.5_
The spectra of the two sequences, denoted by V(ei"') and X(eiOJ ) are shown
in Fig_ Proj. 3.5. They are related by
~

X(eIOJ ) =

2:

V(ei(OJ+kn-

k=-oo

In words, X(eiOJ) is the periodic extension of V(eiOJ), but with period n.

AaaIytk results. We will compare tlte performance of two discrete-time
interpolators:
Zero-order interpolator, where

for0:5n:51
otherwise
First-order interpolator, where

0.5
1.0
ht(n) =

~.5

forn=-1
forn=O
forn=1
otherwise

116

INTRODucnON TO DIGITAL SIGNAL PROCESSING

(a) Determine the magnitude responses IHo(eiw)1 and IH1(e jW )I. Sketch both
responses on the same graph for 0:5 W :5 n.
(b) Let {yo(n)} and {Yl(n)} be the outputs of the zero-order and first-order
interpolators, respectively, when {x(n)} is the input. Compute the magnitude
spectra IYo(eiw)1 and 1y'(e'W)I. Sketch both spectra on the same graph.
Computer results. The interpolation cap be implemented as a nonrecursive filter:

SUBROUTINE NONREC (X, NX, BCOEF, NB, Y)

where X is the input array of size NX, BCOEF is the coefficient array of size NB,
and Y is the output array of size NX + NCOEF - 1. For the nonrecursive filter
recall that the coefficient sequence is equal to the unit-sample response. Since
NX = 128, dimension the Y array sufficiently large to contain the output
sequence.
(a) Generate {v(n)} and {x(n)} for wo=n/16. Plot the first 55 samples.
Implement the two interpolators and plot 55 samples of the output sequences
{yo(n)} and {Yl(n)}. The first-order interpolator corresponds to a noncausal
filter. Repeat the above for Wo = n / 4.
(b) Compute and plot the magnitude spectra of {v(n)} and {x(n)}. Use
NFf= 128. Do the same for Yo(n) and Yl(n). for O:5n :5127.
(c) Explain why there is a difference in the interpolator performance for the two
values of woo
3.6. Detection fiben for the FSK system
Object. Implement and investigate the filters that differentiate between the two
sequences generated by the FSK subroutine written in project in Chapter 2.
In the FSK subroutine, 32 samples of two different frequencies were used to code
the binary values 0 and 1 to be transmitted over communication lines. In this
project, the samples of the two sinusoids will serve a second purpose, other than
the code. They will correspond to the unit-sample response of the two receiver
filters, but in time-reversed order. This is, if {ho(n)} is the unit-sample response of
the "zero-frequency (Ko) filter". then
ho(n) = cos[2nKo(31 - n )/32]

for O:5n :531

The unit-sample response of the "one-frequency (K 1 ) filter" is given by

for O:5n :531
When the unit-sample response is matched in this manner to the desired input
signal, the filter is called a matched filter.
Analytic re:mIts
(a) Sketch the block diagram of the interpolation filter as a nonrecursive
structure, i.e. containing only feed-forward coefficients.
(b) Determine the magnitude response of each filter. (Hint: use the finite
geometric sum formula.)

Computer results
(a) Compute and plot the magnitude spectra of the two 32-sample sequences (use
NFf = 128). Use PLOTMR to plot only the first 65 points. Indicate the
location of the two FSK frequency values on each magnitude spectrum plot.

FOURIER TRANSFORM OF DISCRETETIME SIGNALS

117

(b) Implement the two filters using the NONREC subprogram. Compute and
plot the magnitude response of the two filters (use NFl' = 128). Compare
these to part a.
(c) Generate and plot the four discrete-time sequences that are obtained by
passing each of the 32-bit sequences into each of the two filters.
(d) Generate and plot the discrete-time output sequence for each filter when the
input is the FSK sequence corresponding to bit pattern 010. Based on these
output patterns, describe a simple procedure to determine which binary value
was transmitted.

CHAPTER

.4
THE DISCRETE
FOURIER
TRANSFORM

4.1 INTRODUCl10N
The discrete Fourier transform (DFr) sequence, denoted by {H(k)}, allows us
to evaluate the Fourier transform H(e i "') on a digital computer. This complexvalued sequence is obtained by sampling one period of the Fourier transform
at a finite number of frequency points. The DFf is important for two reasons.
First, it allows us to determine the frequency content of a signal, that is, to
perform spectral analysis. The second application of the DFr is to perform
filtering operations in the frequency domain. After defining the DFr and the
inverse discrete Fourier transform (IDFf), we examine some of its properties
that are relevant for signal processing. The DFr can be applied directly to
causal, finite-duration sequences. Difficulties arise, however, when we try to
apply it to noncausal or infinite-duration sequences. These difficulties are
examined and solutions are suggested. The. fast Fourier transform (FFr)
algorithm is presented as an efficient computational procedure for evaluating
the DFr. Its efficiency is the main reason much current signal processing is
performed in the frequency domain. The FFr algorithm is described in both an
intuitive and analytical form using the decimation-in-time approach.

us

119

THE DISCRETE FOURIER TRANSFORM

4.2 THE DEFINITION OF THE

DISCRETE FOURIER TRANSFORM (DFT)
The DFf is the finite-duration discrete-frequency sequence that is obtained by
sampling one period of the Fourier transform. This sampling is conventionally
performed at N equally spaced points over the period extending over
Os w <2K, or at
forOsksN-1

(4.1)

If {h(n)} is a discrete-time sequence with the Fourier transform H(eiw), then

the DFf, denoted by {H(k)}, is defined as

forOsksN-1

(4.2)

The DFf sequence starts at k = 0, corresponding to w = 0, but does not

include k = N, corresponding to w = 2K.
Let us consider the consequences of sampling H(eirD). In the previous
chapter, it was shown that the Fourier transform H(eij is periodic in w, with
period 2K. Its 'Fourier series expansion was then computed and the coefficients
were found to be equal to the discrete-time sequence {h(n)}. It was also
shown that when a function of continuous-valued time is sampled with
sampling period To, then the spectrum of the resulting discrete-time sequence
becomes a periodic function of frequency with period 2K/T.. A similar result
can be shown to be true here. As shown in Fig. 4.1, when H(ei"'). is sampled
with sampling period w. = 2K/N, the corresponding discrete-time sequence,
denoted by {Ji(n)}, becomes periodic in time with period 2K/w. = N. This

H(el"')

h(n)

/~.'

......... 111 .........

-1 01

21r(J)

tbysamPling

H(k)

h{n)

.111 ..... 111 ..... 111

...
N
-N

III 1 , , I III.

-1

0 1 2

N-IN

FIGURE 4-1
A periodic extension of a time sequence results when a spectrum is sampled. Case for N
samples of spectrum over 0 S w < 2.n-. Bracket indicates Npoint DFr sequence.

=8

120

INfRODUcnON TO DIGITAL SIGNAL PROCESSING

periodic discrete-time sequence can be expressed in terms of {hen)} as

co

hen) =

hen + mN)

(4.3)

m=-oo

The sequence {hen)} is called the periodic extension of {hen)}. The number of
sample points in one period of the spectrum N is also equal to the period of
{hen)}.
The importance of this result is apparent when we employ the DFf for
signal processing and filtering in the frequency domain, for it is {hen)} that is
the corresponding sequence that is being used in the time domain. This chapter
describes the implications and pitfalls that accompany this result.
We are free to choose the value of N, the number of samples of H(eio
over O::s: co < 2.1t'. Since N is also the period of {hen)}, we must be careful not
to choose N too small. We now illustrate the considerations involved in
choosing the value of N. Let {hen)} be a causal finite-duration sequence
containing M samples. Infinite-duration and noncausal sequences are discussed
later. As shown in Fig. 4.2, if M ::s: N, then

hen) =

{~(n)

forO::s:n::S:M-l
for M::S:n::s:N-l

(4.4)

Two points are noteworthy about this result when M ::s: N. The first is that
the finite-duration sequence h (n), for O::s: n ::s: M - 1, can be recovered
uniquely from the first period of {hen)}. The second point is that the excess
number of points in one period of {h(n)}, or those for kM::s: n ::s:
keN -1) for any integer k, are equal to zero.

{h(n)} (M = 5)

..... rIIr
-10-1-2345

000

{h(n)} (N=5)

-1 0 1 2 3 4 5
{h(n)} (N=8)

.. Irllt ... tlllt tII ~

~3

{ii(n)}

4 567,8

(N = 4)

I I I I

1 1 11I 1I I I I 1 I 1
n

FIGURE 4-2
Relationship of duration of sequence M and number of sample
points in spectrum N. When N <
M, a time-aliasing elIect occurs.
Underbar indicates sequence produced by inverse OFf.

TIffi DISCRETE FOURIER TRANSFORM

121

If M > N, however, the duration of {h(n)} is longer than the period of

{h(n)}. An overlap then occurs when the periodic extension is formed, as

shown in Fig. 4.2. This overlap results in an error when we try to retrieve
{h(n)} from the first period of {h(n)}. In this case, time-aliasing is said to
occur. To prevent time-aliasing, we must sample H(ejOO) over 0:5 OJ < 23f at
least as many times as there are elements in {h(n)}, or N? M times. Clearly,
this poses a difficulty when trying to apply this result to infinite-duration
sequences. These problems will be considered below.

4.3 COMPUTING mE DISCRETE

FOURIER TRANSFORM FROM mE
DISCRETE-TIME SEQUENCE
We now derive the relationship of the discrete Fourier transform to the
discrete-time sequence. First, we note that since {h(n)} is a periodic sequence,
it can be expressed as the following Fourier series expansion
h(n)

=..!.
N

H(k) ej'br1cnIN

(4.5)

k=-oo

where {ii(k)} are a set of complex-valued coefficients to be determined, and

the lIN factor has been accepted as the convention. Because the discrete-time
complex exponential sequence {ei'brknIN} is itself periodic in k, with period N,
there are only a finite number of distinct sequences, equal to ei'brknlN, for
0:5 k s N - 1. To show this, let 0:5 k :5 N - 1 and consider a frequency
outside this range, say ei'br(k+N)nIN. Then
ej'br(k+N)IIIN = ei'br1cnIN ei27rn = ej 'br1cnIN
(4.6)
which is identical to the sequence ei'brknIN. Hence, only a finite set of
frequencies need to be included in the above summation, and {h(n)} can be
expressed in terms of the following N frequency components
1 N-l
h(n) = H(k) ej'brnklN
N k=O

2:

(4.7)

To determine the values of H(k), for 0 s k:5 N -1, we multiply both sides of
Eq. (4.7) by e-i'brnrIN and sum over OSn:5N -1, to get

2: h(n)~-i27rnrIN=_1 2: 2: H(k)ei27rnkIN

N-l

n~

N-l [N-l

Nn~

k~

e-j'brnrIN

(4.8)

Changing the order of summations and noting that R(k) is not a function of n,
we get
(4.9)

The sum in the square brackets can be evaluated by applying the finite

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

geometric sum formula, to obtain

N-l

2:

ncO

1 _ ei2Jr(k-r)

ei2Jrn(k-r)/N = -~.:---.,-;---,-=
J2Jr
1- e

(4.10)

(k-r)/N

Since (k - r) takes on only integer values, the numerator term is zero for all
values of k and r. However, for k = r + mN, where m is an integer, the
denominator is also zero. In this case, each term in the sum is equal to 1, and
the sum evaluates to N, or

2:

N-l

ei2Jrn(k-r)lN =

{N
0

n-O

for k=r+mN
otherwise

(4.11)

Since k lies in the range from zero to N -1 in Eq. (4.9), we can apply this
result for k = r, or for m = O. This results in the desired solution for R(r) given
by
N-l

R(r) =

2: hen) e-

j2JrnrIN

for OSrsN-l

(4.12)

n=O

Replacing r with k gives the final desired solution for H(k).

We now relate this result to the Fourier transform of the original
sequence {hen)}. Let {hen)} be a causal, finite-duration sequence containing
M samples. Then its Fourier transform is given by
M-l

H(ej01)

= 2:

,,-0

hen) e- j01n

(4.13)

We can increase the duration of {h(n)} from M to N samples by appending the

appropriate number of zeros. Since zero-valued elements contribute nothing to
the sum, the Fourier transform is equal to
N-l

H(ej01) =

2:

hen) e- j01n

(4.14)

,,=0

Sampling H(ej01) at N equally spaced points over 0 s co < 2.7r produces the
sequence
N-l

H(k) = H(ei01)I O1=2JrklN =

2:

h(n) e- j2:rrknIN

Os k s N - 1

(4.15)

n=O

Since the number of sample points N is equal to the number of elements in the
original discrete-time sequence, the first N samples of {hen)} are equal to
{hen)}, or hen) = hen), for OSn sN i-l. Only the first N values of {hen)}
are needed in the sum in Eq. (4.12). Comparing (4.15) with (4.12), we note
that H(k) = H(k), for 0 s k s N -1. Applying this r.csult to (4.7), we get
1 N-l
h(n) = - 2: H(k) e j2:rr"kIN
for OSnsN-1
(4.16)
N k=O

11iE DISCRETE FOURIER TRANSFORM

123

Summarizing these results, we arrive at the desired Off relationships:

N-l

H(k)

=L

h(n) e-i2mtkIN

for OsksN-1

(4.17)

n=O

which is called an N-point DFT. The inverse discrete Fourier transform, or

IDFT, by which we compute the discrete time sequence from the Off
sequence, is given by
1 N-l
h(n)=H(k)ei2nnkIN

for OSnsN-1

(4.18)

k=O

Equations (4.17) and (4.18) form the basis for the computer algorithms that
evaluate the Off, and are directly applicable to causal, finite-duration
sequences. The number N of frequency points in one period of the Off is
equal to the number of elements in the discrete-time sequence. To obtain more
points in the Off sequence, we can always increase the duration of {h(n)} by
adding additional zero-valued elements. The procedure is called padding with
zeros. These zero-valued elements contribute nothing to the sum in Eq. (4.17),
but act to decrease the frequency spacing ei2nkIN. We will encounter this
padding operation in the computation of the convolution below.
Example 4.1. DFf of the caU$81 3-sample averager. Let

h(n)={~
To compute the N-point DFf, where
for Osk sN -1, is given by
N-l

H(k)

=L

for OSn s2
otherwise
N~3,

the kth complex coefficient H(k),

2

h(n) e-i2nnkIN = L

l e-i2nnkIN

+ e-i2KkIN + e-j4Jrk/~
= l e-i2KkIN(ei2nkIN + 1 + e-i2nkIN)
= l(1

e-J2nklN

1 + 2cos(21tk/N)
3

Comparing this result with Example 3.9, we see that the magnitude and phase
values of {H(k)} are samples of H(ei "') at N points equally spaced over
Os w < 21t. The DFf sequences for different values of N are shown in Fig. 4.3.

The Off is specified over the range 0 s (J) < 23r, rather than -3r s (J) <
the range used for the Fourier transform in the previous chapter. This is the
accepted convention for all the Off algorithms. It is convenient for interpreting the results of computer processing, since the de value, i.e. (J) = 0, then
corresponds to the first element of the Off sequence.
3r,

124

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

{h(n)l

~III. "1.111. .
01234

l........--....

IH(k)1

......

,. ,.

",

.-

/
I

(h(n)}

III
012

;.

IH(e"') I

f\cL.

11

211' w

Arg[H(eJ"')]

:~""'~"~

IN=81

{ h(n)}

.. 111. .... 111..;.

I

IH(k)1

012345678
I

-11

III I Ii
012

3 4

5 6 7

;.

Arg[H(k)]

{121
I

5
1

I6

-11

FIGURE 4-3
Relationship of Fourier transform and discrete Fourier transform for a causal sequence. Cases for
N = 4 and N.,. 8 are shown.

1lIE DISCRETE FOURIER TRANSFORM

(h(n)

125

(M=5)

... 11111 .

-3-2-10 1 23

{ii(n)l (N=5)

IIIIIIIIII!1!!I1I!!1111!! ~

,012345,

{ii(n)l (N=8)

... !I!!L .. !IIII ... II!II.~

{ii(n)l

(N =4)

,0 I 2 3 4 5 6 7,8

IIdlldIIIIIId"rlIlI11,

,0 1 2 3 ,

FIGURE 4-4
Relationship of duration of noncasual sequence M and number of sample points in spectrum N.
The discrete-time sequence is conventionally taken as the first N samples of the inverse OFf
When N < M, a time-aliasing effect occurs.

NONCAUSAL SEQUENCES. Let {hNc{n)} be a finite-duration noncausal

sequence containing M samples. The period extension is given by
h(n) =

hNc{n

+ mN)

(4.19)

".=_00

as shown in Fig. 4.4. However, the conventional definition of the IDFf given
by Eq. (4.18) indicates that the first N samples of the periodic extension are
obtained, or h(n), for 0 ~ n ~ N - 1, when the IDFf is evaluated. The
noncausal part of {hNc{n)} will then appear at the end of the reSUlting
sequence, being the noncausal part of the second period of the periodic
extension, as shown in Fig. 4.4 for N = 8.
Eumple 4.2. 11ae OFf of DODcausal3-S8Dlple avenger. Let

h(n)={~.

for -l:Sn:S 1
otherwise

In the definition of the OFT pair, the time sequence used in the computation is
equal to h(n), for O:s n :s N - 1. Hence, as shown in Fig. 4.5, the sequence that
must be used for evaluating an N-point OFT is

h(n)= {~

for n = 0, 1, N - 1
otherwise

126

INTRODUcrION TO DIGITAl: SIGNAL PROCESSING

{/lIn):

IN=41

ILIILIII

IH(k)1

,,

,,

2 3 ,4

10

{h(n)}

.II 1.
-1 0 1

IH(e'W)1

AcRH{~:[

FIGURE 4-S
Relationship of Fourier transform and discrete Fourier transform for a noncausal sequence.
Cases for N = 4 and N = 8 are shown.

Then
H(k) = j(l + e-ibrklN + e- i2(N-I)",kIN)

Since adding 2Jr to the exponent of the complex exponential does not change its
value, we find
e-1br(N-I)1rkIN

= e- j (2(N-I),..kIN+2J1cNINJ = efb<klN

THE DISCRETE FOURIER TRANSFORM

127

and
H(k) =

hI + e- i2

"kIN

= I

+ eJ2"kIN)

+ 2 cos(2Jlk / N)
3

These are samples of the Fourier transform of the 3-sample averager, as shown in
Fig. 4.5.

We would have obtained the same result if we allowed the DFT sum to
range over negative index values, to include the h( -1) term. But, since the
standard computer algorithms that evaluate the DFT are defined in tenus of
non-negative index values, these equivalent tenus from the next period of
{h(n)} are included in the sum. This convention is often a source of confusion
when interpreting the results of the DFT and IDFT computations.
INFINITE-DURATION SEQUENCES. For infinite-duration sequences, there is
always some amount of time-aliasing which results when the DFT relationships
are applied. In practice, the value of N is usually chosen to be large enough to
make this error negligible for a particular application. We will consider the
effects of choosing the proper value of N in the chapter on IIR filter design.
Example 4.3. DFf of an infinite-dw'lltion sequence. Let us consider the
familiar infinite-duration unit-sample response of the first-order recursive filter,
given by
for all n

g(n) = a"u(n)

As shown in Example 3.2, the Fourier transform is equal to

H(eiOl) = [1- a e-JOIr 1

Sampling this function produces the DFf sequence given by

H(k) = H(e JOI )I"_2"klN= [1- a e- J2 "kINj-!

for 0 s k s N-l

Applying the IDFf formula to determine the corresponding discrete-time

sequence, we have
for OSn sN-l
Expanding the denominator in a Taylor series, we get

N-l
L

1
h(n)=N k=O

eJz"kniN

[ L~

L [N-l
Le

= -1 ~ a'
N ,-0

a'e- j2"k"N

forOsnsN-l

,-0

I2"k(n-,)IN

for OSn sN-l

k-O

From (4.11), the sum in the brackets is zero, except when r = n

+ mN, where m is

128

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

{h(rz)}

.111111

J t

-I 0 1 2 3 4 5

(~ ~) I[1111111 [1111111111

-I I 0 1 2 3 I 4 5

FIGURE 4-6
Time-aliasing effect for an infinite-duration sequence. Underbar indicates sequence
produced by inverse DFT when N = 4.

an integer. The sum then evaluates to N. Hence, we have

~

h(n) =

L:
,=0

for

O~n ~N-l

r=n+mN

For r to be greater than zero, m must also be greater than or equal to O. Summing
over the infinite number of values of m, we have
for

O~n ~N-l

This result is not equal to {g(n)}, although it approaches {g(n)} as N becomes

infinite, as shown in Fig. 4.6. Since the ,mit-sample response of a stable filter
approaches zero with time, a sufficiently l~rge value of N can be found to make
the aliasing effects negligible.
Alternatively, the periodic extension is given by
~

h(n) =

L:

an+mNu(n

+ mN)

for

O~n ~N-l

for

O~n ~N-l

This is the same as the result above.

4.4 PROPERTIES OF mE DFf

Since the DFT sequence consists of samples of the Fourier transform, the
linearity, periodicity and symmetry properties of the Fourier transform are also
true for the DFT. These properties are listed in Table 4.1 and discussed below.
Because of the periodicity and symmetry of the Fourier transform for
real-valued sequences, we need to plot only the values of H(k), for
0::5 k::5 (N /2) + 1, which corresponds to H(e iw ), for 0::5 W ::5 TC.

THE DISCRETE FOURIER TRANSFORM

129

TABLE 4.1

Properties of the discrete Fourier transform

1. Definition: For a finite-duration sequence hen), for O:s n:S N - I,
N-j

2:

H(k) =

hen) e- J2 J<knIN

for O:s k :S N - I

(DFT)

for O:Sn:sN-I

(IDFT)

"=0

and
I
h(n)=-

N-j

2:

H(k)e J2 J<knIN

k-O

2. Periodic extension:
00

2:

hen) =

hen

+ mN)

3. Linearity: If {x(n)} = {hj(n)} + {h2(n)}, then X(k) = Hj(k) + Hik).

4. Periodicity: H(k) = H(k + N).
S. Magnitude and phase functions: If H(k) = HR(k) + jHI(k). then
IH(k)12 = H~(k)

+ Ht(k)

Arg[H(k)] = arctan[HI(k)/ HR(k)]

6. Fourier transform of delayed sequence:

if

{yen)} = {x(n - no)}

then

7. Fourier transform of the convolution of two sequences:

if

{yen)} = {hen)}

* {x(n)}

then Y(k) = H(k)X(k)

A linear convolution of {hen)} and {x(n)} results when Ny-point DFT sequences are computed
for {hen)} and {x(n)}, where Ny is the duration of {yen)}. Otherwise, a circular convolution of
{hen)} and {x(n)} results.
8. Fourier transform of the product of two sequences:

if

{yen)} = {h(n)x(n)}

then

Y(k) = H(k)@X(k)

(circular convolution)
9. For the real-valued sequence hen), for O:Sn:sN -I:
N-j

HR(k) =

2:

hen) cos(2:rckn/N),

for O:sk:sN-l

n=O

N-j

HI(k) = -

2: hen) sin(2:rckn/N)

for 0 :S k

:S N

- 1

n=O

(a)
(b)
(c)
(d)
(e)
(f)
(g)

Complex-conjugate symmetry:
H(k) = H(N - k)
Real component is even function:
HR(k) = HR(N - k)
Imaginary component odd function:
HI(k) = -HI(N - k)
Magnitude function is even function:
IH(k)1 = IH(N - k)1
Phase function is odd function:
Arg[H(k)] = -Arg[H(N - k)]
If {hen)} = {h(-n)} (even sequence), then H(k) is purely real.
If {hen)} = {-h( -n)} (odd sequence), then H(k) is purely imaginary.

130

INfRODUcnON TO DIGITAL SIGNAL PROCESSING

4.5 CIRCULAR CONVOLUTION

Given hen), for O:sn:sN -1, and {x(n)}, their convolution is equal to
N-l

y(n) =

2: h(k)x(n -k)

for all n

(4.20)

k=O

To differentiate it from the operation performed with the DFf, described

below, we denote Eq. (4.20) as being the linear convolution. In the previous
chapter, the Fourier transform was found to be useful for converting a
convolutional operation into a multiplicative one. The values of {yen)} are
then computed from the inverse Fourier transform Y(eico) = H(eico) X(eia.
Here, we would like to determine {yen)} from the inverse DFf of
{Y(k)} , where Y(k)=H(k)X(k), for O:sk:sN-l. But caution must be
exercised in applying the DFf relations because the sequences in the DFf
relations are periodic with period N, being the periodic extension of the
original sequences. The convolution that results is then called a circular
convolution, defined by
N-l

yen) =

2:

h(k)X(n - k)

forO:Sn:sN-l

(4.21)

k-O

or, more concisely,

{yen)} = {hen)} {i(n)}

(4.22)

where denotes the circular convolution operation. The steps involved in

computing this convolution are identical to those for the linear convolution,
except that the sum is taken over only one period. The resulting {y(n)} is also
periodic, with period N.
The main consequence of the periodic nature of the sequences in the
convolution is that the shift produced by the index (n - k) actually represents a
rotation. The periodicity makes the sequence values appear to wrap aroundfrom the beginning of the sequence to the end. The circular convolution is
illustrated in Example 4.4.
Eumple 4.4. CimdIIr CODvolution with the DFf relations. Let
I
x(n) = 00.5
{

forn=O
for n = 1
otherwise

0.5

forn=O
forn=l
otherwise

and
h(n)= { ~

We want to determine the linear convolution of the two sequences

{y(n)}

= {x(n)} * {h(n)}

THE DISCRETE FOURIER TRANSFORM

{x(n)}

I'r:
:" r.
o1

-1
{h(n)}
(a)

-1 0 1 2

(y(n)}

r~

O~5 I .

r:

... -1 0 1 2 ...

{iz{n) }

ddddd
Idddd,
~~I IIIIIIII
~2

{h2 (n)}
(b)

~2

{Y2(n)}

131

...

~2

(c)

FIGURE 4-7
(a) Linear convolution of two finite-duration sequences. (b) Circular convolution result when
2-point DFf sequences are used for the sequences shown in (a). The minimum size of the DFf in
this example is 3. (c) Circular convolution result when 3-point DFf sequences are used for the
sequences shown in (a).

132

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

LiDear convolution. By performing the linear convolution with the finiteduration sequences in the time domain, we get the desired {y(n)} as shown in
Fig.4.7(a).
Circular convolution. Since both {x(n)} and {h(n)} have durations of two
samples, it may be reasonable (but wrong!) to take two 2-point DFfs. Let us
consider the consequences of doing this in both the frequency and time domains.
Frequency domain. Performing the 2-point DFfs we get

X(O) = 1.5 and X(l) = 0.5

and

H(O) = 1.5 and H(I) = -0.5

Multiplying {X(k)} by {H(k)} to compute {Y(k)} and evaluating the product for
k =0,1, we get Y(O) = 2.25 and Y(I) = -0.25.
Applying the inverse DFf, we get the sequence y(O) = 1 and y(1) = 1.25,
which is not the same as the desired linear convolution result. This sequence is
shown in Fig. 4.7(b).
Time domain. Since we computed 2-point DFfs, the corresponding
periodic convolution was between periodic sequences {.i2(n)} and {h2(n)} having
a period of two samples, which were obtained from {x(n)} and {h(n)} by
periodic extension. Performing the periodic convolution, we get
:MO) =.i2(0) h 2(0) + .i2(1) h 2( -1) = 1
.M1) = 2(0) h2(1) + 2(1) h2 (0) = 1.25
.M2) = 2(0) h2(2) + 2(1) h2(1) = 1
The sequence tMn)} is periodic with period 2 and the first period has the same
values as those produced by the 2-point IDFf above.

4.6 PERFORMING A LINEAR

CONVOLUTION WITH THE DFf
To perform a linear convolution with the DFT, it is necessary to evaluate the
DFT at a sufficiently large number of points so that time-aliasing effects do not
occur. To demonstrate this, let {x(n)} and {hen)} be finite-duration sequences
of length Nx and NH , respectively, and {yen)} = {x(n)} * {hen)}. After
performing the linear convolution of {x(n)} and {hen)}, the sequence {yen)}
has duration Ny = N x + NH - 1.
To determine the linear convolution result in the frequency domain, the
DFT of {y(n)} must be evaluated at Ny points to retain the information that is
present in {Y(eiQ)}, i.e. {Y(k)} must contain Ny samples of Y(ei"') over
Os (0 <21r. Then, to compute the product Y(k) = X(k) H(k), Osk sNy -1,
both {X(k)} and {H(k)} must also be evaluated at Ny points. Example 4.5
illustrates this point.
EumpIe 4.5. Performing a linear convolution with the DFf. Let
x(n) =

1
00.5

forn=O
for n = 1
otherwise

TIlE DISCRETE FOURIER TRANSFORM

133

and
0.5
h(n)= { ~

for n =0
forn=1
otherwise

Frequeacy domain. The linear convolution will produce a 3-sample sequence.

To avoid time-aliasing we convert the 2-sample input sequences into 3-sample
sequences by padding with zeros. The 3-point DFr sequences are given by
X(O) = 1.5

X(I) = 1 + 0.5e- j2Jr13

and

X(2) = 1 + 0.5 e- j4,,13

H(I) = 0.5 + e- I2Jr13

and

H(2) = 0.5 + e-)4413

and
H(O) = 1.5

Computing the product

Y(k)

= H(k) X(k)

for 0:sk:s2

we get
Y(O) =2.25
Y(I) = 0.5 + 1.25 e- j2><13 + 0.5 e- j4,.13
Y(2) = 0.5 + 1.25 e- iU13 + 0.5 e- jlln13

Computing the IDFr, we get

y(O) = 0.5

y(l) = 1.25

and

y(2) = 0.5

These are the same values as the linear convolution result obtained in the
previous eJfample.
Time dolUiD. Since we have calculated 3-point DFrs, the result should be
the same as the periodic convolution between {i3(n)} and {h3(n)}. The values of
the result for different values of n are equal to
Y3(0) = i3(0) h3(0)

+ i3(1) h3( -1) + i3(2) h3( -2) = 0.5

Y3(1) = i3(0) h3(1) + i3(1) h3(0) + i3(2) h3( -1) = 1.25

Y3(2) = i3(0) h3(2) + i3(1) h3(1) + i3(2) h3(0) = 0.5

The sequence {Y3(n)} is periodic with period 3 and the first period has the same
values as those produced by the 3-point IDFr, as shown in Fig. 4.7(c).

In Chapter 9, we will indicate how this procedure can be extended to

evaluate a linear convolution when one of the sequences has an infinite
duration. This situation commonly occurs when processing long data records,
such as those of digitized music.

4.7 COMPUTATIONS FOR EVALUATING

THE DFf
In this section, we describe the computations required for the direct evaluation
of the DFf of a complex-valued sequence. In the previous chapter, we

134

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

presented a simple relationship for evaluating the Fourier transform of a

real-valued sequence at a collection of frequency points. By considering
complex-valued sequences here, we find that both the transform and its inverse
can be evaluated with the same algorithm. Even though the time sequence may
be real-valued, the Off sequence is complex-valued. Hence, it is necessary to
consider complex-valued sequences.
Let us consider the complex-valued sequence given by
(4.23)
Including this complex form in the definition of the Off, we obtain
N-l

H(k) = L hen) e- j 2n:nkIN

n=O
N-l

=L

[hR(n) cos(2.n-knIN) + hl(n) sin(2.nknIN)

N-l

L [hR(n) sin(2.7rknIN) -

-j

hI(n) cos(2.n-knIN)

(4.24)

n=O

The DFf sequence can also be expressed explicitly in terms of its real and
imaginary parts as

H(k)=HR(k)+jHI(k)

forOsksN-I

(4.25)

The IOFf relationship is then given by

1 N-l
hen) =H(k) ei2n:nkIN
N k-O

1 N-l

for O:sn:sN-I

=N ~o [HR(k) cos(2.n-knIN) - HI(k) sin(2.n-knIN)

. N-l

+L
N

[HR(k) sin(2.n-knlN) + HI(k) cos(2.n-knIN)]

(4.26)

k=O

Except for the difference in the signs of the exponents and the II N
scaling factor in the inverse, the same operations are performed in both
transforms. Both the Off and IOFf computations require 4N real-valued
multiplications per point, and 4N2 such operations are necessary to compute
the N-point transform. To obtain a simple, but still reasonably accurate,
estimate of the operation count, we ignore the observation that the first real
and imaginary terms are multiplications by 1 and 0 respectively, and hence, not
true multiplications. We then find that for large values of N, the number 4N~,
is a good approximation for the required number of real-valued multiplications
to compute an N -point discrete Fourier transform or its inverse. For
complex-valued sequences, the symmetries of the real and imaginary components, which are valid for real-valued sequences, no longer hold.
In the next section we describe the implementation of a program to

TIlE DISCRETE FOURIER TRANSFORM

135

directly evaluate the Off from (4.24) and the 10FT from (4.26). Because of
the similarity in the operations for computing the Off and IOFf, if an
algorithm were found to perform the evaluation in an efficient manner, the
benefits would be applicable to both the transform and its inverse. Later in this
chapter, we will consider an alternate algorithm to the direct evaluation, called
the fast Fourier transform (FFT). This algorithm requires N to be a power of 2
and takes into account the periodicity of the Off to reduce the number of
multiplications from 4N1 to 4N log2 N.

4.8 PROGRAMMING THE DISCRETE

FOURIER TRANSFORM
The program to evaluate the Fourier transform of a finite-duration sequence
h(n), for O:s; n:S; NH - 1, at a set of discrete-frequency points, was described in
Chapter 3. This same program can also be used to compute the discrete
Fourier transform after making two modifications. First, the number of
frequency points in the Off, denoted by NOFf, should be equal to the
number of elements in the discrete-time sequence NH . To be consistent with
the conventional algorithms and the popular fast Fourier transform (FFf), we
take NOFf to be a power of 2, or
NOFT =2NLOG2
where NLOG2 is an integer. The reason for this is explained when we explore
the mathematically involved FFf algorithm later.
The second modification involves reducing the number of arrays used in
the computation to conserve memory requirements. The standard Off
algorithms conventionally employ the same arrays for both the input and
output, as described below.
Incorporating these two modifications, the subprogram is given by
DFT (NLOG2, HR, HI, NTYPE)

where
NLOG2
HR
HI
NTYPE

(integer input) is the log to the base 2 of the size of the arrays
(real input and output array) is the array of real values
(real input and output array) is the array of imaginary values
(integer input) is the operation to be performed: NTYPE = 1
causes the Off to be computed, NTYPE = -1 produces the
10FT
To compute the Off, the HR and HI arrays contain the real and
imaginary parts of the discrete-time sequence, padded with zeros if necessary.
If the sequence is real, all the elements of HI should be set to zero. After the
routine has been accomplished, HR and HI contain the real and imaginary
parts of the Off sequence. To compute the 10FT, the HR and m arrays
contain the real and imaginary parts of the OFT before the call. After
execution, HR and HI contain the real and imaginary parts of the discrete-time

136

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

sequence. If the OFT sequence corresponds to that for a real-valued sequence,

the elements in the HI array evaluate to values that are approximately equal to
zero, allowing for round-off errors in the computations.
To allow the same arrays to be used in the calling argument, it is helpful
to use temporary arrays, TR and TI, of size NOFT, for the computation of the
real and imaginary parts of the transform within the subprogram. The FFr
algorithm computes the OFT without needing this temporary storage, as
described later.
As in the previous program, a problem arises with the indexing of the
arrays, which usually starts at one, rather than zero. For the Kth element of
the OFT sequence, H(k), for 1 s K s NDFf, the corresponding frequency
value, denoted by OMK, is then equal to
OMK = 2.1l(K - 1)/NDFT

for 1 sKsNOFf

(4.27)

The evaluation is straightforward, being the direct implementation of (4.24)

for the DFf and (4.26) for the IOFf. The real part of the Kth element of the
Off computation then becomes
NDFT

TR(K) =

[HR(M) cos((M - I)OMK) + HI(M) sin((M -1)OMK))

M=l

for 1 sKsNOFT (4.28)

A similar equation can be written for the array TI containing the imaginary
part. For evaluating the DFT, the HR and HI arrays contain the real and
imaginary discrete-time sequence values. After the TR and TI arrays are
computed, these values are transferred to HR and HI, respectively, thus
completing the transform. For evaluating the 10FT, the HR and HI arrays
contain the real and imaginary parts of the OFT sequence.
Evaluating the OFT and 10FT with the above algorithm is a simple
procedure. In many cases, this Off routine is sufficient for occasional use for
determining the spectrum of a discrete-time sequence. However, if spectral
analysis is to be employed routinely as part of a repetitive signal processing
task, this OFT routine is too time-consuming, especially when the number of
points is large, say NOFT = 1024. Recall that the number of mUltiplications
alone is 4NDFf2. In the next section, we investigate procedures to increase the
speed of evaluating the required 4NDFT2 computations. In the following
section, we describe the revolutionary fast Fourier transform, which reduces
the multiplication count to 4NOFT log2 NDFf.

4.9 INCREASING THE

COMPUTATIONAL SPEED OF THE DFf
If the OFT subroutine is to be used 6ften for spectral analysis and filtering
operations, the N 2 operations required to evaluate the N-point OFT become

tediously time-consuming. In the next section, we describe the fast Fourier

TIlE DISCRETE FOURIER TRANSFORM

137

transform (FFf) algorithm that produces exactly the same results in N logz N
operations. For certain applications, however, the FFf may not be the best
way to proceed. For example, som~ applications require only a small number
of frequency points, say at the M frequencies (01' (Oz, , and (OM, where
M N, rather than at the full set of N frequency values (Ok = 2nklN, for
O::s; k ::5 N - 1. A second example is when the M frequency points are
, not
harmonically related, but rather an arbitrary collection of interesting frequencies. In this case the FFf will, in general, not evaluate the spectruin at
these exact frequency values, but only at the harmonic frequencies that
approximate these.
In this section, we discuss programming and implementation procedures
to improve the speed of the N Z computations that are required for the N -point
DFf. These programming techniques save computation time by storing the
sine and cosine coefficients in a table and accessing the discrete-time array
values efficiently. The discussion below illustrates the general principle of
computer processing of a memory / speed tradeoff, that is, a given computation
can be done more quickly if more memory is allowed.
To illustrate the various points below, let us consider computing the OFf
of a real-valued sequence x(n), for 0::5 n ::5 N - 1. The ideas discussed below
can be extended to complex-valued sequences in a straightforward way. The
kth value of the OFf sequence is given by
N-l

X(k)

= L x(n) e-j2JrnkIN
N-l

= L x(n) cos(2nkn/N) n=O

N-l

L x(n) sin(2nkn/N)

for 0::5 k ::5 N - 1

n=O

\
~

(4.29)

COMPUTATION OF THE TRIGONOMETRIC COEFFICIENT VALVES. This

OFf can be programmed directly, as was done in the OFf routine in the
previous section. However, note that for each value of the time:sequence, the
sine and cosine coefficients must be evaluated. In a computer, this evaluation is
usually performed by computing a series expansion of., the trigonometric
identity to the accuracy of the number representatiol,l{ a time-consuming
operation. For example, the Taylor expansion of the cqsine is equal to

cos(a) = 1 - ~/2! + a 4 /4! - a 6 /6!

+...

One way to reduce the computation ti~Je is to note that successive values
of the complex-exponential sequence differ by a constant factor e-jZ1rkIN, or
e- j21rk (n+l)/N

= e-j21rknIN e-j21rkIN

(4.30)

Hence, for each frequency point k, successive values of the sine and cosine
coefficients can be determined by evaluating the constant factor
e-j21rkIN

= cos(2nk/N) -

j sin(2nk/N)

(4.31)

138

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

The two coefficients can then be calculated recursiyely as

cos[2.nk(n + l)/N] = cos[2.nkn/N] cos[2.nk/N] - sin[2.nkn/N) sin[2.nk/N]

(4.32)
and

sin[2.nk(n + l)/N) = cos[2.nkn/ N) sin[2.nk/N) + sin[2.nkn/N) cos[2.nk/N)

The recursion starts with cos(O) = 1 and sin(O) = O.
The undesirable feature of this procedure is that the constant factor in
Eq. (4.31) is computed to the accuracy of the machine. The resulting error in
the coefficients, even though small for small values of n, accumulates as n
increases. This effect is most noticeable for large N-point DFf computations.
A method that does not have this error, at the expense of additional memory
requirements, is to generate and store the coefficient values in an array, as
described next.
STORING DFf COEFFICIENTS IN A TABLE. If the DFf is to be performed

repetitively, as in spectral analysis of a set of data records, the computation

speed can be increased by computing the necessary coefficients once and
storing these values in a table consisting of an array of cosine coefficients,
crAB, and an array of sine coefficients, STAB. These arrays are then used in
the DFf computations.
The memory required for storing crAB and STAB can be reduced by
noting two points. First, because the frequencies in the DFf computations are
harmonically related, or Wk = 2.nk/N, all the needed coefficient values exist in
the set of N coefficients for the fundamental frequency, or W = WI. These are
shown in Fig. 4.8 for N = 32. Let us consider a pointer variable, NPNTR, that
points to the desired coefficient value in the array. To obtain the coefficients
for the spectral value at WI, the pointer is incremented by 1 each time a new
coefficient is needed. For the spectral value at wz, the desired coefficient set is
equal to every second value in the table, starting from the first value. The
pointer is then incremented by 2. The coefficients for the kth spectral value are
then obtained by incrementing the pointer by k. When the end of the table is
encountered, or when NPNTR > N, the pointer is brought back into the
appropriate place in the table by subtracting N, or NPNTR = NPNTR - N.
As shown in Fig. 4.8, the sine and cosine coefficients take on the same
values, with the elements of crAB only shifted with respect to STAB. Hence,
both sets of coefficients can be obtained from the CTAB table. But now there
are two pointers, NCPNTR that points to the cosine values and NSPNTR that
points to the sine values. These pointers are initialized to the values given by
NCPNTR = 1 and NSPNTR = (3N/4) + 1, to produce the correct sequence of
coefficient values. When either pointer value exceeds N, the size of the array,
it is set back to the correct position by substracting N, as before.

TIlE DISCRETE FOURIER TRANSFORM

139

CTAB VALUES

NPNTR
(NCPNTR)

3N
4

+1

STAB VALUES

t
NPNTR

FIGURE 4-8
Cosine (crAB) and sine (STAB)
coefficent values for evaluating
N-point DFf. Case for N = 32 is
shown.
Pointers
NPNTR,
NCPNTR and NSPNTR indicate
the coefficient values that are
needed for a particular computation step.

ARRAY INDEX COMPUTATION OVERHEAD. In computing the DFT re-

lationship, the discrete-time sequence values are processed sequentially. In

FORTRAN, the array index is stored and accessed from memory every time it
is needed. This index overhead also contributes to slowing the computation. In
some programming languages, like C, for example, there is an array
auto-increment feature that eliminates the need for an array index variable.
With this feature, the next value of the array can be accessed automatically.
This feature can be used to increase the computation speed.
GOERTZEL ALGORITHM FOR COMPUTING mE DFf. Let us consider a
second-order filter whose unit-sample response is the cosine sequence given by

h(n)=cos(2nkn/N)

forn~O

(4.33)

This filter is not stable by our definition, but we will use it for only a finite
number of time points, to insure that the output does not become infinite.
Such a filter can be obtained from Example 2.13 by letting r = 1. The filter

140

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

structure is shown in Fig. 4.9(a). If the input to this filter is the real-valued
sequence x(n), for O:s n :s N - 1, the output value at time n = N can be
determined by the convolution equation and is given by
N-l

yeN)

x(m) heN - m)

m=O
N-l

x(m) cos[2nk(N - m)/N]

(4.34)

m=O

(a)
x(n) <>-_._--------3>(

} - - - - - - - - . . . . . - - y ( N ) = XR(k)

x(n)

}--------~---y'(N)

(b)

=-XI(k)

(c)

x(n)

' - - - - - - - I >------oy'(N) = XI(k)

FIGURE 4-9
Filter structures for evaluating the DFT of a sequence x(n), for O:s n :s N - 1, with the Goertzel
algorithm. a) yeN) equals the real part of X(k). b) y'(N) equals the negative imaginary part of
X(k). c) Real and Imagmary parts computed simultaneously at n = N.

TIlE DISCRETE FOURIER TRANSFORM

141

Since
cos[2Jlk(N - m)/N]

= cos[-2Jlkm/N + 2Jlk] = cos[2JTkm/N]

we can write
N-J

y(N) =

2:

x(m) cos(2JTkm/N)

(4.35)

m=O

Comparing this equation with the definition of the OFT, we find that by
putting the finite-duration sequence x(n), for O:Sn:sN -1, through the filter,
the output at time n = N is equal to the real part of the OFT element X(k).
This result is true when the filter has a zero initial condition. In a similar
manner, consider a second filter whose unit-sample response is given by
h'(n)=sin(2JTkn/N)

forn~O

(4.36)

This filter is shown in Fig. 4.9(b) and discussed in Example 5.4. If {x(n)} is
applied to this filter, the output at time N is given by
N-J

y'(N) =

2:

x(m) sin(2JTkm/N)

(4.37)

m=O

Comparing this equation with the definition of the OFT, we find that by
putting the finite-duration sequence x(n), for O:s n :s N - 1, through the filter,
the output at time n = N is equal to the negative imaginary part of the OFT
element, or -Xj(k).
These two filters are combined in Fig. 4.9(c). By changing the recursive
and nonrecursive parts and including the sequence {w(n)} that is common to
both filters, we find that the kth OFT element is equal to
X(k)

= w(N) -

cos(2JTk/N) w(N -1) - j sin(2JTk/N) w(N -1)

(4.38)

Ooing a multiplication count on this filter, we find a multiplication by

2 cos(2Jlk/ N) for each of N time points. The multiplication by -1 is not
counted as a multiplication. At time n = N, the two additional multiplications
given in Eq. (4.38) are performed, for a total of N + 2 for each element in the
OFT sequence. This is approximately one-half of that required for the direct
evaluation of the OFT for a real-valued sequence. Computing the OFT with
this procedure is called the Goertzel algorithm. With this approach the entire
set of coefficient values is not needed, since these values are computed in the
filter as the unit-sample response. The same problem exists of expressing the
recursive coefficient 2 cos(2JTk / N) sufficiently accurately, as described above in
the recursive calculation of the coefficient value set, given in (4.32).
To compute the entire OFT sequence, a separate filter is included for
each spectral value. One desirable feature of this algorithm is that, if
computation time is to be minimized, these filters can be implemented and
operate on the input sequence in parallel. Hence, the entire OFT sequence
could be computed shortly after observing the Nth sample of the input
sequence. For N frequency points, however, the number of multiplications is

142

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

still proportional to N 2 We next describe the fast Fourier transform, an

algorithm that produces the same result with a multiplication count that is
proportional to N iog2 N.

4.10 INTUITIVE EXPLANATION FOR

THE DECIMATION-IN-TIME FFT
ALGORITHM
We describe the FFf algorithm that exploits the symmetry properties of the
discrete-time complex exponential to reduce the multiplication count significantly for the computation of the DFf. For evaluating an N-point DFf, the
FFf algorithm described here achieves its efficiency when N is a power of 2, or
N = 2NLOG2 , for some integer NLOG2. This restriction presents no practical.
problem, since the length of {hen)} can always be increased to the next power
of 2 by padding the sequence with an appropriate number of zeros.
To provide some understanding for the steps in the FIT algorithm, let us
consider the input discrete-time sequence hen), for O::Sn::sN -1, to be a
real-valued sequence. The same basic ideas are also applicable to complexvalued sequences and to the inverse Fourier transform computation.
For real-valued discrete-time sequences, the definition of the DFf
reduces to
N-l

H(k)

2:

n=O

N-l

hen) cos(2nnk/N) - j

2:

n=O

hen) sin(2nnk/N)

for 0 ::S k

::S

N- 1

(4.39)
To compute this N-point DFf sequence requires 2N2 real multiplications. The
set of values for the sine and cosine coefficients required for evaluating each
H(k) term are shown in Fig. 4.10 for N = 8. The coefficients for computing the
terms k and (N /2) + k, for O::s k ::S (N /2) - 1, exhibit a curious similarity. The
coefficients for even-valued n are the same, while those for odd-valued indices
have the same magnitude, but are of opposite sign. To exploit this similarity,
we break up the original sequence {h(n)} into two subsequences: {hE(n)},
containing the elements indexed by the even-values of n; and {ho(n)},
containing the elements indexed by the odd-values. These sets of coefficients
are shown in Fig. 4.1l.
Two features of the even-indexed coefficient set are noteworthy. First, the
set of coefficients for frequency points k and (N /2) + k, for O::s k::s (N /2) - 1,
are identical. Hence, their contribution to the terms H(k) and H(N /2 + k) will
be the same. Only one of these terms need be computed, the other being
obtained for free. The second feature is that this set of coefficient values is
exactly that required for computing the N/2-point DFf of {hE(n)}. To
compute this N /2-point DFf, we need to perform 2(N /2)2 = N 2 /2 real
multiplications. A similar result, except for a sign inversion in the coefficients,
occurs for the odd-indexed sequence. To compute the N /2-point transforms
for the even and odd-indexed sequences requires N 2 multiplications, which is
smaller than the original 2N2 Hence, splitting, or decimating, the sequence

TIIE DISCRETE FOURIER TRANSFORM

143

cosme coefficients:

tIIIIIIL
2

k=O

sine coefficients:

I r
k=!

I t

:=21
I

1
k=3

I
J

1
1

1
r
1

I
k=4

.n

1 2 3 4 5 6

1
k=5

k=6

r r

r
1

k=7

1
n
7

1
t

I r J

ftGURE ....10
The set of cosine and sine coefficient values for evaluating the kth element of an 8-point DFf
sequence.

into even a.nd odd-indexed subsequences leads to a reduction in the number of

computations to evaluate the DFf. We indicate below how these two
N/2-point DFf transforms are combined to form the N-point DFf of the
original sequence.
To reduce further the number of multiplications, we now consider

144

INTRODUcnON TO DIGITAl SIGNAL PROCESSING

EVEN-INDEX

ODD-INDEX

t0 I2 I4 1
t0 I2 I L tI I3 I5
6
4

k=O

2 4 6

k=l

k=O

k=4

..

IJ

t
k=5

k~ k~

1 1 1 In

I I

k=! I I
t t

k=)

11 11

k=2

k=3

I 3 5

5 7

k~ k~ A
."1 r.

7k=4

n
7

I T

k=6

~
ks-r

11 1 I

rr

FiGURE 4-11
The set of cosine and sine coefficient values for evaluating the kth element of an 8-point DFf
sequence broken up into the even-valued indexed and odd-valued indexed sub-sequences.

{hE(n)} to be the original sequence to be transformed into an N/2-point DFT.

We again decimate {hE(n)} into its even and odd-indexed subsequences. The
corresponding set of coefficients is shown in Fig. 4.12. Again, note the same
two features of this set as was apparent above: the set of coefficients for
frequency points k and (N /2) + k, for 0 ~ k ~ (N /4) -1 are identical, and
these are the coefficient values necessary to compute an N /4-point DFT. At
this stage in the decimation, four subsequences of length N /4 are obtained,
each requiring 2(N /4f = N 2 /8 multiplications to compute their N /4-point
DFT, or a total of N 2 /2. Hence, another reduction by 2 is obtained in the
multiplication count.
The decimation operations can continue until only pairs of the original
sequence occur. Each of these require a 2-point DFT computation to be

11fE DISCRETE FOURIER TRANSFORM

ODD-INDEX

EVENINDEX

III
2
.11

k=l

k=2

k=3

t1

III
3

.11

k=O

145

~
1

.-~

FIGURE 4-12
The set of cosine and sine coefficient values for
evaluating the kth element of a 4-point DFI'

sequence broken up into the even-valued indexed and odd-valued indexed sub-sequenccs.

pedormed. At this level, however, all the coefficients are either 0 or 1, and
do not require a multiplication. This can be observed from Eq. (4.39) by
letting N = 2 and setting k = 0 or 1. Hence, by merely adding and subtracting
elements at this lowest possible level of decimation, we can compute 2-point
DFfs. Since the original sequence had N = 2NLOG2 elements, it could be
decimated into subsequences NLOG2 - 1 times, and there are 2NLOG2- 1
2-point DFfs. This procedure is called the decimation-in-time FFI'
algorithm, since the discrete-time samples are continuously split into two
groups. There is also a decimation-in-frequency FFr algorithm. It will not be
covered here, but can be found in the references.
Having described the decimation operation and indicating why there is a
net reduction in the number of computations, we now describe the analysis to
indicate the procedure of combining the N /2-point DFfs to form an N-point
DFf and determine the operation count.

146

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

4.11 ANALYnC DERIVATION OF mE

DECIMATION-IN-TIME FFf ALGORITHM
In this section, we provide the analysis behind the intuition presented in the\
previous section. Let hen), for 0s;ns;N-1, where N=2 NLOG2 , for some
integer NLOG2, be a complex-valued sequence. We split up the OFT defining
equation into two summations consisting of the even and odd-indexed
sequences given by
N-l
H(k) = L hen) e-j21rnkIN
for O::s k s;N-1
,,~O

L hen) e-j2nnkIN

hen) e-j2#nkIN +

"even

II

(4.40)

odd

Letting n = 2r in the first sum, and n = 2r + 1 in the second, we have

(N~-l

H(k) =

h(2r) e-j41frlcIN +

(N~-l

r-O

(N~-l

h(2r + 1) e- j2n (2r+l)kIN

r-O

h(2r) e- j2nrkl(NI2) + e-j2nkIN

(N~-l

~
h(2r + 1) e-j2nrkl(NI2)
(4.41)
,-0
Both sums appear in the form of an N/2-point OFT. Let us define {hE(n)} as
the even-valued index sequence and {ho(n)} as the odd-valued and {HE(k)}
and {Ho(k)} as being, respectively, the OFT of {hE(n)} and {ho(n)}. Since
{HE(k)} and {Ho(k)} are the OFTs of N/2-point sequences, they are periodic
in k with period N /2. These two will be combined to form the N-point OFT
{H(k)} by
~

,~O

H(k) = HE(k) + e-j2nk'NHo(k)

=HE(k -

N /2) + e-j2nk'NHo(k - N /2)

for O::s k s; (N/2)-1

for N/2::s k::s N -1 (4.42)

Since 4N real-valued multiplications are required to compute the N-point

OFT of a complex valued sequence, to calculate {HE(k)} or {Ho(k)} requires
4(N/2)2 = ~ real multiplications, or 2N2 for both. To combine {HE(k)} and
{Ho(k)} to form {H(k)} requires four real multiplications per point for
O::s k s; (N /2) - 1, or 2N for the combination procedure. The points for
N /2 ::s k s; N - 1 contain the same products, but with different signs. Hence, to
compute an N-point OFT by decimating the sequence once requires 2(N2 + N)
multiplications, which is less than 4N2 for N> 2.
If decimating the. sequence once reduces the multiplication count,
additional decimations should reduce it even further. To show this, the evenand odd-indexed sequences can now be considered as original sequences and
also decimated into their own even and odd parts. The Off of the
even-numbered elements of the original sequence can then be written as
HE(k) = HE,E(k) + e-j2nkl(NI2)HE,o(k) for O::s k::s (N /4)-1
=HE;E(k - N/4) + e-j2nk'(NI2)HE,o(k - N/4)

for N/4::sk::s (N/2)-1

(4.43)

TIlE DISCRETE FOURIER TRANSFORM

;B---o

h(O)~

147

HE.E(O)

h(Nj2) r.C.----~

"'>----:Ol

HE E(1)

FIGURE 4-13
The butterfly pattern (G = 1).

where {HE,E(k)} is the DFf of the even-indexed sub-sequence of {hE(n)} and

{HE,O(k)} is that of the odd-indexed sub-sequence of {hE(n)}. A similar
equation can be written for {Ho(k)}.
Since N = 2NLOG2 , the original sequence can be decimated into even- and
odd-indexed sequences NLOG2 - 1 times, until 2NLOG2 - 1 2-element subsequences are obtained. The first even-indexed two-point sequence is composed of h(O) and h(N/2). This 2-point sequence has the 2-point DFf given by
HE,E(k) = h(O) + h(N12) e- jnk

for k = 0, 1

(4.44)

Evaluating this expression, we find

HE,E(O) = h(O) + h(N /2)

and

HE,E(l) =h(O) - h(N /2)

This last operation is shown graphically in Fig. 4.13, and is called a butterfly

pattern. The steps for combining the DFfs of the decimated subsequences are
shown as a sequence of butterflies in Fig. 4.14 for N = 8.
h(O)

H(O)

h(4)

H(l)

h(2)

H(2)

11(6)

H(3)

hell

H(4)

h(5)

H(5)

h(3)

H(6)

h(7)

H(7)

-1

FIGURE 4-14
Decomposition of an N-point OFf into butterfly patterns. Case for N", 8 is shown.

148

INJ'RODucnON TO DIGITAL SIGNAL PROCESSING

TABLE 4.2

Comparison of real mnltiplication counts for N-point

DFT (N=zi')
NLOG2

5
6
7
8
9
10

32
64
128
2S6
512
1024

DFT(W)

4096
16384

6S536
262144
1048576
4194304

J!'FI' (4 N IoIz N)
640
1536
3584
8192
18432
40960

The computations are performed column by column, by operating on al

the pairs first, and then proceeding to the right to perform the computation:
on the results of the first set of butterflies. Notice that as the computatio(
along aparticular vertical set of values is completed, these values are no longel
needed in future computations. Hence, the memory locations that were used
to store these intermediate results are free to be used for storing new values,
and ultimately the values of the DFf itself. This in-place computation is
memory-efficient and a second advantage of the FFr for computers with
limited memory.
To obtain an approximate multiplication count, note that there are N
complex multiplications, or 4N real multiplications, required after each
decimation and that there are 1082 N - 1 splits. Hence, 4N(1082 N - 1)
real-valued multiplications are required. What is important in the multiplication count is not the exact number, but how it is related to the number of
points N. The evaluation of the DFf using the straightforward evaluation
requires on the order of JIl complex multiplications, while evaluating it with
the FFr algorithm requires N 1082 N. These two numbers are compared in
Table 4.2. Notice that reductions from 6-fold for a 32-point DFf, to over a
hundred-fold for a 1024-point DFf can be achieved. Since this reduction is
aiso directly related to computation time, the overwhelming popularity of the
FFr is understandable.

4.12 SOME GENERAL OBSERVATIONS

ON 11IE FFf
In this-section, we discuss some observations on the implementation of the
FFr. These include the unusual ordering of the input sequence, the memory
requirements and the computation of the coefficients.
U-ORDERING TIlE INPUT SEQUENCE. In the decimation-in-time algorithm
described in the previous section, the input sequence is continuously divided
into subsequences having odd- and even-valued. indices, until subsequences

THE DISCRETE FOURIER TRANSFORM

149

TABLE 4.3

Comparison of row location and index nomber

of input sequence
SeqaeDte IIldell

Rowuaber

DedmIII
0
1
2
3
4
5
/6
7

Biliary

DeclmaI

(000)
(001)
(010)
(011)
(100)
(101)
(110)
(111)

0
4
2
6
1
5
3
7

Decimal

(000)
. (100)
(OlO)
(110)
(100)
(101)
(011)
(111)

of length 2 are obtained. The index values on the left side of the
computation graph shown in Fig. 4.14 indicate an unusual ordering of the
input sequence. This ordering has an interesting interpretation, that can be
extended to arrays of any duration, when the index value is expressed as a
binary number. In Table 4.3, we list the number of the row in the computation
diagram and the index of the input sequence that occupies that row. Comparing
the binary representation of the row and index values, we note that the index
value appears as the reversed-bit order of the row number. This reversed-bit
transformation of the index values is valid for any size N-point DFf, where N
is a power of 2. This is commonly observed in decimation-in-time FFf
subroutines, in which the first step involves placing the input sequence values
into the appropriate row location. An algorithm to pedorm this bit reversal by
using the decimal index values is left as a problem for the student.
MEMORY REQUIREMENI'S. In the direct computation of the DFf using

X(k) =

N-l

2: x(n)

e-j2nnkIN

for O:sk:sN-1

(4.45)

n=O

all the values of x(n) are needed for each value of X(k). If the computation is
pedormed by sequentially computing the X(k) values, as in the DFf routine
described in Section 4.8, these spectral values must be stored in a temporary
buffer array. After the computations, the contents of this buffer are transferred
to the arrays passed in the argument of the subroutine. From the computation
diagram for the FFf shown in Fig. 4.14, the data flow is consistently directed
from left to right. This indicates that after the values of a particular column are
used, they are no longer needed in the computations. Hence, those memory
locations can be used for storing the results of future computations. If the
computations are pedormed column by column, the results can be stored at
the same memory locations, and no significant buffer memory is required. This

150

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

memory-efficient procedure is called in-place computation and is increasingly

important as the size of the OFT is increased.
COEFFICIENT VALVES. As discussed previously for the implementation of

the OFT, a memory/speed tradeoff also exists for the FFT. As noted from the
computation diagram, there are only a finite number of coefficients needed to
evaluate the FFf. These coefficient values can be computed as they are
needed, directly from the function or recursively, as described previously, or
they can be pre-computed and stored in a look-up table. In this case the table
look-up procedure is more complicated than for the OFT, since the input
sequence values are in reversed-bit order. Many FFT decimation-in-time
subroutines exist in standard scientific subroutine packages. As a general rule,
the length of the program, in lines of code, and its memory requirements are
inversely related to the execution speed. That is, the longer programs operate
more quickly and the simple programs require more time. But any FFT
subroutine is much faster than the direct method of evaluating the OFT.

4.13 OTHER FAST REALIZATIONS OF

THE DFT
Although the most popular, the FFT is not the only fast realization to evaluate
the DFT. Other procedures offer desirable features of better accuracy or
simpler implementation. These procedures fall into the category of number
theoretic transforms. Since they tend to be esoteric, they are not suitable for
discussion in a beginning text such as this one. These procedures are discussed
in the reference by Elliot and Rao.

4.14 SUMMARY
In this chapter we have defined the discrete Fourier transform (DFT) as the
sequence obtained by sampling one period of the Fourier transform over
O!5 0 < 2.1t' at N frequency points, where N is an integer. The OFT was
shown to be the Fourier series expansion of the periodic extension of a
discrete-time sequence, obtained by repeating the original sequence with a
period of N samples. After defining the DFT and its inverse, the IDFT, we
examined some of its properties relevant to signal processing. The OFT
relations could be used without modification to achieve the desired results
when applied to causal, finite-duration sequences. For noncausal sequences,
the inverse OFT produces the causal part, followed by the noncausal part. For
infinite-duration sequences, an inevitable time-aliasing error occurs, whose
effect decreases as N is increased. When attempting to perform the convolution operation in the frequency domain with the OFT, the equivalent
time-domain operation is the circular convolution. The desired linear convolution result can be obtained by computing the OFT at a sufficiently large

TIlE DISCRETE FOURIER TRANSFORM

151

number of frequency points. The fast Fourier transform (FFf) algorithm was
presented as an efficient computational procedure for evaluating the DFf. The
FFf algorithm was described in both an intuitive and analytical form using the
decimation-in-time approach. Some general remarks were made regarding the
implementation of the FFf algorithm.

FURTHER READING
Bergland, G. D.: "A guided tour of the fast Fourier transform," IEEE Spectrwn, 6: 41-52 (1969)
Blahut, R. E.: Fast Algorithms for Digital Signal Processing, Addison-Wesley, Reading, MA,
1984.
Braoewell, R. N.: The Fourier Transform and Its Applications, McGraw-Hill, New York, 1978.
Burrus, C. S., and T. W. Parks: DFT/FFT and Convolution Algorithms, Wiley-Interscienoe, New
York, 1985. [This book provides a description of the various algorithms to implement a
discrete Fourier transform. Program listings are given in FORTRAN and assembly
language for the Texas Instruments TMS32010 signal prooessing chip.]
Childers, D. G., (ed.): Modem Spectral Analysis, IEEE Press, New York, 1978. (This is a book of
reprints which presents a historical perspective of the evolution of spectral analysis.]
Cooley, J. W., and J. W. Tukey.: "An algorithm for the machine computation of complex Fourier
series," Math. Comp., 19: 297-301 (1965).
Elliott, D. F., and K. R. Rao.: Fast Transformations: Algorithms, Analysis and Applications,
Academic Press, New York, 1982.
Oppenheim, Alan V., and Ronald W. Schafer: Digital Signal Processing, Prentioe-Hall,
Englewood Cliffs, NJ, 1975.
Peled, A., and B. liu: Digital Sgnal processing: Theory, Design and Implementation. Wiley, New
York,1976.
Rabiner, Lawrenoe R., and Bernard Gold: Theory and Applicalion of Digital Signal Processing,
Prentioe-Hall, Englewood Cliffs, NJ, 1975.
Schwartz, Mischa and Leonard Shaw: Signal Processing: Discrete Spectral Analysis, Detection and
Estimation, McGraw-Hill, New York, 1975.

PROBLEMS
4.1. Evaluate and compare the 8-point DFf for the following sequences.
l
for -3sn s3
(a) x(n) = { 0

otherwise
forOSns6
(b) x(n) = {~
otherwise
4.2. V~rify the properties of the DFf relationship given in Table 4.1.
4.3. When discussing the Fourier transform, it was shown that an even discrete-time
sequence, or one for which h(n) =h( -n), has a Fourier transform that is
real-valued. Since the DFf sequence as produced by the computer algorithm can
have only positive index values, indicate the symmetry property of a discrete-time
sequence whose DFf sequence is real.
4.4. Let us consider the finite-duration sequence {x(n)}. It is common to process
sequences that are defined over index values that do not correspond to the range
Os n s N - 1 required for the DFf computations. In these cases, the {h(n)}
must be periodically extended to form {h(n)} , and the sequence h(n), for
Os n s N - 1 is processed to produce the desired results. Determine and sketch

152

INJ'RODUcnON TO DIGITAL SIGNAL PROCESSING

the discrete-time sequence to be processed to find the 8-point OFf, when the
following sequences are given.
(a) x(n) = {OI

for n =-1
otherwise
3- ln l
for-3:S:n:s:3
(b) x(n)= { 0
otherwise
for -3:s: n :s: 3
(c) x(n) = 0
otherwise
4.5. Consider the sequence {x(n)} and its N-point OFf {X(k)} where N is an even
number.
(a) Show that if (z(n)} is an even sequence, then X(N /2) = O.
(b) Show that if {x(n)} is an odd sequence, then X(O) = o.
4.6. Compute the N-point OFf of each of the following sequences. Let N = 8 and
M=5.
for OSn:s:N-1
(a) x(n) = {~
otherwise
for O:s:n:s: M -1,
(b) x(n) = {~
otherwise
for n =0
(c) x(n) =
forn=N-1
forn=O
(d) x(n) =
forn=M,
4.7. The OFf sequence {H(k)} represents samples of the Fourier transform H(el"') at
equally-spaced frequency intervals, or

, {n

{_!
{_!

H(k) = H(e/"')I .._2a*IN

for Osk:s:N-1

To obtain a denser sampling, the finite-duration time sequence h(n), containing

M samples, where M < N, can be increased to any finite duration N by padding
with N - M zeros. The OFf relations can then be applied to obtain N samples for
any desired N. For causal sequences, these zeros are appended to the end of the
given sequence. For noncausal sequences, a more careful padding is necessary, as
illustrated in this problem. For the sequences given below, sketch the 8-point
sequence that is used for computing the 8-point OFf.
for 0:S:n:s:3
n
(a) hen) = { 0
otherwise
Inl
for -IsnsI
(b) hen) = { 0
otherwise

4.8. Perform the linear convolution of the two sequences below in the time domain.
Compare with the result obtained by using the OFf relations.
h(n) =

{~

for O:s: n :s: 3

otherwise

and

z(n) =

{~nl

for -I:s:n:s:I
otherwise

TIlE DISCRETE FOURIER TRANSFORM

153

4.9. Show the symmetry properties of the DFf sequence by computing the 8-point
DFf of the real-valued sequence given by
for -1 Sn sl
otherwise
4.10. Computing the DFf of two real-valued sequences with one execution of the FFT
algorithm. The FFf algorithm can operate on a complex-valued sequence. When
transforming a real-valued sequence, the imaginary part is set to zero. We can
transform two real sequences' simultaneously by letting the real part of the
complex-valued input be set equal to one real sequence, and the imaginary part
equal to the second real sequence. The complex-valued output of the algorithm
must then be manipulated to achieve the desired result, consisting of the two
DFf sequences of the two real-valued input sequences.
Let xt(n), for OsnsN-l be a real-valued sequence having the DFf
sequence whose real and imaginary parts are given by Rt(k) and It(k), for
OsksN-1, or
N-t

Rt(k) =

L xt(n) cos(2:rckn/N)

forOsksN-l

"-0

and
N-t

It(k) = -

L xt(n) sin(2:rckn/N)

for OsksN-1

Let x2(n), for 0 s n s N - 1 be a real-valued sequence having the DFf sequence

given by R2(k) and 12(k), for 0 s k s N-1.
Since the discrete Fourier transform is a linear transformation, it obeys the
superposition principle. Define the discrete-time sequence {y(n)}, is
y(n)=x\(n)+jx2(n)

forOsnsN-l

(a) Find the real and imaginary parts of the DFf of {y(n)}, denoted by R,.(k)
and Iy(k) in terms of Rt(k), R 2 (k), I\(k) and 12(k).
(b) From the above relationships, find the inverse relationship to obtain R\(k),
R2(k), I\(k) and I2(k), to produce the desired DFf sequences.
4.1l. Determine the reversed-bit order of the input sequence to compute a 16-point
DFf with the FFf algorithm.
4.12. Develop the reversed-bit reordering algorithm based on the decimal number of
the sequence index. Demonstrate your algorithm for 8-point DFf and 16-point
DFf calculations.

COMPUTER PROJECfS
4.1. 'Program to evaluate the DFf
Object. Write a subroutine that evaluates the
rete Fourier transform and
compare the results with the Fast Fourier tran.. m (FFf) algorithm given in
Appendix A. It is not necessary to be familiar wi ne FFf algorithm to perform
this project. We consider the FFf to be just like any other packaged program and
simply want to verify that it works correctly.

154

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

Write the subroutine

DFT (NLOG2, HR, HI, NTYPE)

as described in the text. The call to the FFf has the identical arguments as above,
given by
FFT ( NLOG2, HR, HI, NTYPE).

For the problems below, let NLOG2 = 7 (NDFT = 128).

Analytic results. Determine and sketch the magnitude and phase of the Fourier
transform for the sequences below and compare with the computer results.
Computer results
(1) Test the DFT subroutine by computing the DFT of the sequences below and
plotting the entire (NDFT) magnitude array and the phase array for the following
sequences. Use the PLOT subroutine.
Compare these plots with those obtained with the subroutines PLOTMR for
the magnitude response and PLOTPR for the phase response given in Appendix
A. These plotting routines show only (NDFT/2) +1 points, or 65 points when
NDFT= 128.
(a) Causal sequence. Let h(n) = 1, for 0:5 n :5 3, and =0, otherwise ..
(b) Noncausal sequence. Let h( -1) = 0.5, h(O) = 1 and h(l) = 0.5.
(2) Compute the DFT sequences of the two discrete-time sequences given !lbove
with the FFf algorithm. Compare the PLOTMR and PLOTPR plots with those
produced above.
(3) Consider the sequence
h(n)=1

forO:5n:5NDFT-l.

Use the DFT subroutine with NDFT = 64,128,256 and 512 to determine the time
required for each array size. Do not plot or print your results. Use the TIME
function on the computer. Your dimension statements should allow for 512element arrays.
Construct a table of the programming times and compare with those
required by the FFT subroutine to compute the same transforms.
Remark. The results produced by DFT and FFT should be identical, but the
speed of the FFf is much greater. Hence, we will use only the FFf subroutine for
the remainder of the projects in the course.
4.1. DFf of &nite-duration sinusoidal sequences
Object. Investigate the effect on the DFT when varying the period of a
discrete-time sinusoidal sequence and the length of the sequence.
Computer results. For the problems below, let NLOG2 = 7 (NDFT = 128) and plot
the first 65 points of the magnitude sequence corresponding to 0:5 W :5 n with
PLOTMR. Generate the discrete-time sequence
x(n) = {COOS( won)

for 0:5 n :5 K -1
for K:5n:5NDFT-1,

for the following conditions:

(a) Wo on a sampling point of the DFT: let Wo = n/4, and K = 64, 32 and 16.
(b) Wo not on a sampling point of DFT: let Wo = 15n /64 and K = 64, 32 and 16.

mE DISCRETE FOURIER TRANSFORM

15S

4.3. Performing convolutions with the OFf

ObJ~. Perform a linear convolution with the DFf and the observe the effects of
a circular convolution.
We want to compute the linear convolution given by

{y(n)}

= {h(n)} * {x(n)}

where

={~

for OSn s5
otherwise

h(n)={~

for OSn s5
otherwise

x(n)

and

Analytic results. Determine {y(n)} in closed analytic form and sketch.

Computer results
(1) Performing the 11netU" convolution with the OFf. List the steps in performing
the linear convolution by using the smallest possible number of points NDFf in
the DFf, that is still a power of 2. Plot the following by using PLOT:
(a) the original time sequence {x(n)} (plot NDFf points),
(b) the magnitude spectrum of (a),
(c) the magnitude spectrum of the product of the two DFf sequences,
(d) the inverse DFf sequence and compare with {y(n)}.
(2) Performing the dmdar convolution with the OFf. Demonstrate the timealiasing phenomenon produced by choosing NDFf too small. Produce the plots
listed in part (1). Explain the results on the plots.

4.4. laterpolating the OFf by padding the time-domain sequence with zeros
Object. Compute the DFf at a denser set of points in the frequency domain by
padding the time-domain sequence with zeros.
Computer results Let
x(n) = r" cos(01on)

forOSns63

with r = 0.9 and 010 = 9fr/64. Compute the 64-point DFf and plot the 33 points of
the magnitude spectrum corresponding to 0 s 01 S fr. Use the PLOT subroutine.
Generate the 128-point sequence {y(n)} by padding {x(n)} with zeros:
y(n)=

{~(n)

for OSn s63

for 64 s n s 127

Compute the 128-point DFf and plot the 65 points of the magnitude ~pectrum
corresponding to 0 s 01 S fr. Compare the two plots.
4.5. OFf of the product of two sequences
Object. Verify that the product of two time sequences results in the convolution
of their Fourier transforms.
Consider the following sequence: y(n) = w(n) x(n), for all n, where
x(n) =

{COSo(frn/8)

for OSn S 127

otherwise

156

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

and
wen)

~ {~

for 0 ~ n ~ 15 and 113 ~ n ~ 127

otherwise

Analytic results. Determine and sketch the discrete Fourier transforms {Y(k)},
{X(k)} and {W(k)}, and compare with the computer results. (Hint: {X(k)} has
only two nonzero points: use the finite geometric sum formula for {W(k)}, and a
convolution with an isolated point produces only a shift and a scale change.)

Computer results. Compute and plot the magnitudes of the 128-point DFfs for
the three sequences, using PLOTMR.
4.6. IDterpoiatiag time-domain sequences by padding the DFf with zeros
Object. lIIustrate how a time-domain sequence can be interpolated by modifying
the frequency-domain function.
Computer results. Let x(n) = cos(Jrn/4), for O~n ~31.
(1) Plot this sequence. All the plots in this project should be performed with
PLOT.
(2) Compute the 32-point OFf of {x(n)} and plot the real and imaginary
sequences Xa(k) and XI(k), for 0 ~ k ~ 31.
(3) Pad these sequences with 32 zeros to produce Ya(k) and Yi(k). To retain the
symmetry properties of the real and imaginary sequences, the zero-padding is
performed as follows:

Ya(k)=

Xa(k)
0
{
Xa(k-32)

for
for
for

0~k~15
16~k~47
48~k~63

and
for O~ks15
for 16sk s47
for48~k~63

Plot {Ya(k)} and {Yi(k)}.

(4) Compute the 64-point IDFf of {Y(k)}. Plot the real and imaginary components and compare with {x(n)}.
(5) Determine the equivalent time-domain interpolation function by letting
{x(n)} = {den)} and repeat the steps above. Describe the characteristics of the
corresponding interpolation filter.
4.7. Erron when empIoyiDg the DFf for infinite-duration sequences
Object. Observe the errors that are produced when applying the OFf to
infinite-duration sequences.
Let hen) = an for n ~ 0 with a = 0.95. Then

R(ei "') = 1/(1- a e-i "')

Computer results. Generate the 16-point OFf sequence {R(k)} by evaluating
R(ei"') at the appropriate frequency points. Compute the 16-point IDFf of {R(k)}
and compare with hen). Repeat for N = 32. Compare these results with those of
Example 4.3.

CHAPTER

5
THE
z-TRANSFORM

5.1 INTRODUcnON
The z -transform is the primary mathematical tool for the analysis and synthesis
of digital filters. Procedures for determining the z -transform of a discrete-time
sequence are described. The system function of a digital filter is defined as the
z-transform of its unit-sample response. The complex z plane is employed to
display the configuration of the poles and zeros of the system function. The
circle centered at the origin of the z plane and having radius equal to unity is
called the unit circle. The region of convergence (ROC) in the z plane
determines the values of z for which the z-transform is defined. The form of
the ROC is also related to the properties of the discrete-time sequence.
The. location of the system function singularities in the z plane is
exploited in several ways. The stability of the discrete-time system can be
determined by inspection, since stable causal systems have all their poles inside
the unit circle. Stable noncausal systems are allowed to have poles at infinity.
The Fourier transform is shown to be the evaluation of the z-transform along
the unit circle in the z plane. The discrete Fourier transform can be
determined by evaluating the z-transform at a set of equally-spaced points on
the unit circle. Finally, the digital filter coefficients can be determined from the
locations of the poles and zeros. The chapter concludes by considering three
157

158

INTRODUCilON TO DIGITAL SIGNAL PROCESSING

methods for performing the inverse operation of determining the discrete-time

sequence from its z-transform and the region of convergence.

S.2 THE DEFINITION OF THE

z-TRANSFORM
The z-transform of a sequence {h(n)}, denoted by H(z), is defined as
co

H(z) ==

h(n)z-n

(5.1)

n=-oo

where z is a complex-valued variable. The z-transform of a sequence is said to

exist for all values of z for which the sum is finite. When the sequence is the
unit-sample response of a digital filter, as implied by the sequence {h(n)} in
Eq. (5.1), then its z-transform is called the system/unction. The z-transform of
a signal sequence {x(n)}, denoted by X(z), is similarly given by
00

X(z)= Lx(n)z-n

(5.2)

n=-OO

Like the Fourier transform, the z-transform will be shown to convert a

convolution in the time domain into a multiplication in the z domain. Its
primary use, however, is for the analysis and synthesis of digital filters in the z
domain. We illustrate the computation of the z-transform in the following
examples.
Euuaple 5.1. 1he z-tnmsform of the 3-sample avenger. Let

for -lsns1
otherwise
Applying the definition of the z-transform, we have
z 1 Z-I
H(z}="3+"3+"T.
Example 5.2. 1he z-transform of the first-order recursive filter. Let

a"

h(n}= { 0

forn~O

otherwise

The z-transform is given by

H(z} =

00

11-0

.. -0

2: a"z-" = L (az- I)"

Applying the infinite geometric sum formula, we have

1

H(z} = 1- az-I

lHE zTRANSFORM

159

It is common to encounter sequences whose elements are expressed in

terms of sin( ron) and cos( ron). To apply the geometric sum formulae, the
sequences should be transformed by applying the Euler identitiec, as illustrated
in the following examples.
Example 5.3. 11ae ztraDsform of a IeCOnd-order recursive filter. Let

h(n) = {~oos(won),
The z -transform is given by

for n ~O
otherwise

H(z) = L r" oos(won)z-"

,,-0

Applying the Euler identity and the infinite geometric sum formula, we get

H(z) =

..

(ei"'O" + e-j.,O")

,,-0

L r"

z-"

=-1 [ L"(r.e''''oZ-I)" + L.. (r e-I"'Oz-I)" ]

2 ,,-0

,,-0

_![

1
+
1
]
- 2 1- r &"'OZ-I 1- r e-j"'oz-I
j
1[
2-r(&"'o+e- .,0)z-1
]
=2 (1- r &",oz-I)(l- r e-i"'Oz-l)
1- r oos(WO)Z-1

Example 5.4. The ztransrorm of another second-order system. Let

h(n) = {~" sin(won)

The z-transform is given by

for n ~O
otherwise

..

H(z) =

L r" sin(won)z-"
,,-0

Applying the Euler identity and breaking up the sum, we have

1
1]
=2j1 [
1-r&"'Oz-I-1-re iOloZ- 1
r sin( WO)Z-I

160

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

5.3 PROPERTIES OF THE

z-TRANSFORM
In this section, we describe the properties of the z-transform that are relevant
to describing discrete-time signals and systems. These are listed in Table 5.1
and discussed below.
LINEARITY OF THE z-TRANSFORM. To demonstrate that the z-transform is

a linear transformation, we apply the superposition principle to the definition

of the z-transform. Consider the sequenCe {x(n)} = {ahl(n) + bh 2 (n)}, where
a and b are constants. The z-transform'of {x(n)} is equal to
X(z)

...
= L

x(n)z-n

co

= L

[ahl(n) + bh 2(n)]z-n

n=-oo
...

co

=aL h 1(n)z-n+bL h 2(n)z-n

n=-o:>

n~-o:>

=aH1(z) + bH2(z)

(5.3)

The linearity of the z-transform allows us to implement the digital filter

in the z domain using the same four step procedure as that presented in the
time domain in Chapter 2.
THE z;-TRANSFORM OF ADELAYED SEQUENCE. Since one of the main
operations of a digital filter involves delaying the elements in a sequence, let us
TABLE 5.1

Properties of the z-transform

1. Definition:

...

L co

H(z) = n __ h(n)z-n

z.

Region of convergence (ROC): values of z for which

11 __
00

3. Linearity:

4. z-Transform of delayed sequence:

if

(y(n)}

=lx(n -

no)}

then

Y(z) = z-NOX(z)

5. z-Transform of the convolution of two sequences:

if

(y(n)}

= {h(n)} * {x(n)}

then

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

THE ,TRANSFORM

161

consider the z-transform of a delayed sequence. Let {hk(n)} = {hen - k)}

represent a sequence that is delayed by k samples and its z-transform by
Hk(z). From the definition of the z-transform, we have

= L""

Hk(z)

hk(n)z-n

n=-oo

=n -

Changing the index to m

hen - k)z-n

(5.4)

n=-tXl

k, we have

00

Hk(z)

""

=L

00

h(m)z-(m+k) = Z-k

h(m)z-m

= z-kH(z)

(5.5)

m=-oo

m=-oo

In words, delaying the sequence by k samples changes the z-transform of the

sequence only by an additional multiplication by Z-k. It is for this reason that
the unit-delay component in the digital filter is denoted by a box containing a
z -1 .
TIlE z-TRANSFORM OF TIlE CONVOLUDON OF TWO SEQUENCES. As
introduced in Chapter 2, the convolution operation that relates the input and
output values of a linear time-invariant system is given by
00

y(n)=

h(k)x(n-k) ..

(5.6)

k=-""

Applying the definition of the z-transform to {yen)}, we have

Y(z)

= L""

y(n)z-n

n=-oo

(5.7)
From this last equation, we want to determine Y(z) in terms of H(z) and
X(z). The required infinite sums are present, but we need a Z-k for the h(k)
terms and an additional Zk for the x(n - k) terms. Multiplying Eq. (5.7) by
ZkZ-k and changing the order of the summations, we get
Y(z) =

k~"" [h(k)z-k n~"" x(n -

k)z-(n-k) ]

(5.8)

In the sum over n, the variable k represents a delay of k samples in {x(n)}.

Also, the same delay occurs in the exponent of z. Since the sum is taken over
the entire range of {x(n)}, this delay of k samples is inconsequential in
evaluating the sum. By substituting m = n - k, this sum is found to be equal to
X(z). Since X(z) is not a function of k, it can be factored out of the sum over
k, leaving the definition of H(z). Hence, we have the desired result:
Y(z)

= H(z) X(z)

(5.9)

162

INTRODUCI10N TO DIGITAL SIGNAL PROCESSING

5.4 THE SYSTEM FUNCfION OF A

DIGITAL FILTER
The previous result provides a simple description of the behavior of a linear
system in the z domain. The system acts through the z-transform of its
unit-sample response, or its system function H(z). The z-transform of the
output Y(z) is commonly called the response transform, while that of the input
X(z) is the excitation transform. In words, the response transform is equal to
the product of the excitation transform and the system function. We illustrate
the application of this equation in the following example.
Example 50S. An application of the system function. Let us determine the filter
that converts the input sequence

= {~n cos(won)

for n ~O
otherwise

y(n) = {~n sin(won)

for n ~O
otherwise

x(n)

into

From Example 5.3, we have

X(z)

From Example 5.4, we have

Y(z)

1- 2r cos(WO)Z-l + r2z- 2

The required system function is then given by

H(z) = Y(z)/X(z)
r sin( WO)Z-l

1- r cos(WO)Z-l

It is left as an exercise for the student to determine the unit-sample response of

this filter.

The system function was defined above as the z-transform of the

unit-sample response. We can also determine the system function in terms of
the coefficients in the difference equation. In Chapter 2, the finite-order
difference equation describing a linear time-invariant system was given by
(5.10)

TIlE :TRANSFORM

163

Taking the z-transform of the output sequence, we have

00

Y(z}

= 2:

y(n}z-n

n=-aCI

=n~oo [~1 ak y(n -

k) +

k~NF bk x(n -

k) ]z-n

Changing the order of summation, we get

Y(z} =

~1 akL~oo y(n -

k}z-n] +

k~NF bk[n~oo x(n -

k)z-n ] (5.12)

Applying the relation for the z-transform of a delayed sequence, we have

Np

Y(z}

= 2:

k=l

ak Z-k Y(z}

2:

k=-NF

bkz- k X(z}

(5.13)

Since Y(z) and X(z) are not functions of k, they can be factored out to give
M

Y(z) [ 1- ~1 ak z -

k]

Np

=X(Z}k=~NF bkz-

(5.14)

Applying the definition for the system function H(z), we have our desired
result, or
(5.15)

The numerator allows both positive and negative exponent values of z,

corresponding to future and past values of the input. The denominator,
however, does not contain positive exponent values, indicating that only
current and past values of the output are accessible. To implement a digital
filter from the system function, it must be expressed in an analytic form, such
that the denominator contains terms in z with exponent values that are either
zero or negative.
Example 5.6. System function of the 3-sample averager. For the 3-sample
averager, the difference equation is
y(n) = Ux(n

+ 1) +x(n) + x(n -1)]

Taking the z-transform, we have

Y(z) = UzX(z) + X(z)

+ z-IX(z)]

The system function is then

H(z) = Y(z)/X(z) = z/3 + l

+ z-I/3

Example 5.7. System function of the first-order recursive filter. For the

164

INI'RODUcnON TO DIGITAL SIGNAL PROCESSING

first-order recursive filter, the difference equation is

yen) = ay(n -1) +x(n)

Taking the z-transform, we have

Y(z) = az- l Y(z) + X(z)

(1- az- l ) Y(z) = X(z)

or

The system function is then

1

H(z)=-~l

l-az-

Example 5.8 System function of a 5eCOnd-order recursive filter. Consider thl

second-order recursive filter whose difference equation is given by
yen) = al yen -1) + a2y(n - 2) + box(n) + blx(n -1)

Taking the z-transform, we have

Y(z) = alz- l Y(z) + a2z- 2Y(z) + boX(z) + b1z- l X(z)
(1- alz- l - a2z- 2)y(z) = (b o + b1z-l)X(z)

The system function is then

5.5 COMBINING FILTER SECI10NS TO

FORM MORE COMPLEX FILTERS
It is not uncommon for a digital filter to be implemented by combining simple]
component digital filters to form a larger composite filter. There are tw(
common ways of combining additional filters, either in cascade (also calle(
series) or in parallel. In this section, we will consider the system function of the
composite filter in terms of the system functions of its components.

CASCADE CONNECflON. Let us first consider the series combination of twc

filters shown in Fig. S.l(a) Writing the system function explicitly for eact
CASCADE COMBINAnON

PARALLEL COMBINATION

V(z)

X(z)~

'X(z)~

H,(z)H 2 (z)

Y(z)

~Y(z)

X(z)

X(z)O---+,(z)+H 2 (ziJ----oy(z)

FIGURE 5-1
Cascade and parallel combinations of linear systems.

Y(z)

THE z-TRANSFORM

165

component filter, we have

V(z)

= H 1(z) X(z)

Y(z)

= H 2(z) V(z)

(5.16)

Combining these two equations, the output Y(z) can be expressed in the terms
of X(z) as
Y(z)

=H 2(z) H 1(z) X(z)

(5.17)

Since the multiplication operation is commutative, we can also write

Y(z)

= H 1(z) H 2(z) X(z)

(5.18)

This result can be generalized to the cascade connection of any finite number

N of component filters. If H,,(z) is the system function of the nth cascade

section, then the response transform is given by
Y(z)

=H 1(z)H2(z) ... HN(Z)X(Z)

= n H,,(z) X(z)
N

(5.19)

,,=1

Thus the system function of the total filter composed of cascaded components,
denoted by Hc(z), is simply the product of the system functions of each of the
individual series components, or
Hc(z) =

n H,,(z)
N

(5.20)

,,=1

Caseade combiDation 01 diptaI &Iten. Let us cascade two

3-sample averagers to obtain an enhanced lowpass filtering effect. The total
system function is then given by

Eumple 5.9.

H(z) = (z/3 + 1+ z-I/3)2 = (Z2 + 2z + 3 + 2z- 1 + z-2)/9

Notice that in combining the two second-order filters in series, a fourth-order

filter is obtained. In general, the number of delays required in the total filter is
equal to the sum of the delays in the individual components.

It is useful to apply the previous result to the general filter structure,

which can be viewed as the cascade combination of two simpler filters. The
response transform can be expressed as
Y(z) = H(z)X(z) = H 1(z) H 2(z) X(z)

where

and

(5.21)

166

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

H1(z) is the nonrecursive part of the general filter, while H2(z) is the recursive
part. This result will be useful later for deriving an alternate filter structure
that reduces the number of delays required for implementing a digital filter.
Example S.10. Two ways of bnplementing a second-order recnrsive filter. From
Example 5.3, let us consider the following system function:

1- r cos(a>O)Z-l
(z) = 1- 2r cos(a>o) Z 1 + r2z

which we can write as

where

and
H2(z) = (1- 2r cos(a>O)Z-l + ~Z-2)-1
The desired filter can be implemented by either H(z) = H1(z) H 2(z) or H(z) =
H 2(z) H1(z), as shown in Fig. 5.2.

} - - - -__-o y(n)

x(n) 0 - . . , . . - - - - - ; . (

Ha(z)
~

_ _~A~____~

HI(z)
____~A~______~

x(n)

"GURE 5-2
Two digital filter structures that implement the same system function.

y(n)

TIfE zTRANSFORM

167

PARALLEL CONNECI10N. When filters are connected in parallel, each

component has the same input, and the output is the sum of the component
filter outputs. Let us consider N component filters that are connected in
parallel. N = 2 is shown in Fig. 5.1(b). The kth filter system function is
denoted by Hk(z), for 1 ~ k ~ N, and Yk(z) is the output of the kth filter,
equal to

Yiz)

= Hk(z) X(z)

(5.22)

The output response for each of the parallel sections is added to produce the
composite filter output response, given by
,
N

Y(z)

=L

k-l

Yk(z)

=L

k-l

Hk(z) X(z)

=X(z) L Hk(z)
k-l

(5.23)

From this last equation, it is apparent that the system function of the total filter
composed of parallel components, denoted by Hp(z), is given by
Hp(z) =

L Hk(z)
k=l

(5.24)

EUlllpIe 5.11. PanIIeI eoaaedioD of digital filten. Let us combine two

first-order recursive filters, having coefficients a and b, respectively, in parallel.
The overall system function is equal to

1
1-bz- 1+1-az- 1
2-(a+b)z-1
1
1
(z) =1- az- + 1- bz- =(1- az- 1)(1- bz 1) 1- (a + b)Z-1 + abz- 2
1

When combining the two first-order recursive filters, the overall system is
second-order and has a nonrecursive component, if a :/= -b.

5.6 DIGITAL FILTER

IMPLEMENTATION FROM THE
SYSTEM FUNCI10N
Since the z-transform is a linear transformation, the filter implementation
procedure is similar to that in the time domain. The most convenient form for
filter synthesis is the z-transform of the general difference equation given by
Y(z)

k=l

DkZ-k Y(z)

+ }:

bkz- k X(z)

(5.25)

k=-~

This form is convenient because the response transform is expressed in terms

of the past response transforms as well as the future, current and past
excitation transforms. The delay operator acts as a multiplier with coefficient
Z-l. As in the time domain, a digital filter structure can be determined in four
steps, as shown in Fig. 5.3.
Step 1. Two points are drawn, one corresponding to the excitation
transform X(z) and the other to the response transform Y(z).

168

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

STEP!
X(z)_

--oY(z)

STEP 2

.~1!'~X(z)

L:Jo
NF

zX(z)

X(Z)

X(z)

STEP 3

D_NF

~b~NFX(Z)
I

X(z)o--.......- - - t

----11.---<1

boX(z)

Y(z)

STEP 4

I
X(z)o-O

btz-tX(z)

-~:B b

~~---<o Y(z)
L-y-J

f QtZ-l,y(z)
t-I

FIGURE 5-3
Steps in implementing a digital filter from the system function H(z).

TIlE zTRANSFORM

X(z) _ _

169

t-----_-y(z)

---~

FIGURE 54
Implementation of a second-order filter from the system function.

Step 2. Delayed versions of the excitation and response are obtained by

connecting delays, denoted by Z-1, to the respective points. Advances,
denoted by z, are connected only to the excitation transform.
Step 3. Multipliers are connected to the outputs of the delays and
advances to obtain the desired products.
Step 4. The products are combined through adders to obtain the desired
sums.
EumpIe 5.12. Digital filter structure of 5eCOnd-order recursive filter. Let
y(n) = 2rcos(01o)y(n -1) -r2y(n -2) +x(n) -roos(01o)x(n -1)

Taking z -transforms we get

Y(z) = 2r cos(010)Z-1 Y(z) - r 2z- 2 Y(z) + X(z) - r oos(010)Z-1 X(z)

Applying the four steps above, we arrive at the filter structure shown in Fig. 5.4.

5.7 TIlE COMPLEX z PLANE

The z-transform H(z) is a function of the complex-valued variable z. The
complex z plane is shown in Fig. 5.5 with the horizontal axis defining the real
values and the vertical axis the imaginary values. In this section, we consideJ
H(z) in terms of its singularities in the z plane. The configuration of th(

1m

GJ
z=re Jf

----~-~---Re

FIGURE 5-5
The complex z plane.

170

INTRODUCI10N TO DIGITAL SIGNAL PROCESSING

singularities determines the type of the digital filter, either recursive or

nonrecursive, and can be used to interpret its frequency behavior.
To define the singularities of a z-transform, let us consider the system
function of the general digital filter given by
(5.26)
The roots of the numerator, or the values of z for which H(z) = 0, define the
locations of the zeros in the z plane. The roots of the denominator, or the
values of z for which H(z) becomes infinite, define the locations of the poles.
The locations of the poles and zeros can be shown explicitly by factoring the
above sum-of-products form into the following product-of-terms form:
Nr;;~

H(z) = Az~ 11 (1 - CkZ-1)

k=l

/ M

II (1 -

dkz- 1)

(5.27)

k=l

where A is a real-valued gain constant; Ch for 1:s; k :s; Np + NF , is the location

(possibly complex) of the kth zero and dk , for 1 :s; k :s; M, is the location of the
kth pole.
Note the following important points about Eq. (5.27).

1. Taken individually, each numerator term (1- CkZ-1) generates a zero at

z = Ck and also a pole at z = o.

2. Each denominator term generates a pole at z = dk and a zero at z = o.

3. H NF > 0, the filter is noncausal, having NF elements in the unit-sample
response that have negative index values. The term z~ generates NF poles
at z = 00 and NF zeros at z = o.
4. There is a pole/zero cancellation that occurs at z = 0, in which a numerator
pole will typically cancel a denominator zero. When the singularities at
z = 00 are also counted, the net result is that the number of poles will equal
the number of zeros.
5. For the digital filter to have real-valued coefficients, the complex-valued
singularities must occur in complex-conjugate pairs ..
6. The filter type can easily be recognized from the pole/zero configuration.
The system function of a nonrecursive filter can be expressed as a
numerator polynomial. Hence, all the singularities in the z plane, except at
z = 0 and z = 00, will be zeros, whereas recursive filters will hav~ at least
one pole in this region.

These points are summarized in Table 5.2 and illustrated in the following
examples.
Example 5.13. Pole/zero pattern of the 3-sample avenger. From Example 5.1,
the system function is given by

H(z) = z/3 + 1+ z-1/3

TIlE zTRANSFORM

171

TABLE 5.2

Location of singularities in the

z plane

Consider the z-transform given in the fol\owing form:

H(z)=AzNF~NF
11 (l-ckz- I )
k_1

1Mn

(l-dkz- I )

k-I

The locations of the singularities are given by

zeros at z = Ck' for 1 :s k :s N p + N F , and N p + NF poles at z = 0
poles at z = dk , for 1 :s k:s M, and M zeros at z = 0
and
NF poles at z = 00 and ~ zeros at z =0, due to the non-causal part of {h(n)}.
(Pole/zero cancellation will usually occur at z = 0.)

Converted into the product-of-terms form, we get

H(z) = z(1 + Z-I + z-2)/3 = z(1- c1z- I )(1- C2Z-I)/3

where
-1 + j\!3

A = 1/3,

CI=

and

C2=

-1 -

jV3

=e

-j2J<13

The first term, z, generates a pole at z = 00 and a zero at z =O. The two terms
containing Z-1 generate zeros at z = CI and z = C2 and two poles at z = O. One of
these poles cancels the zero at z = O. The pole/zero pattern is shown in Fig. 5.6.
The number of poles is equal to the number of zeros. Notice that the gain term A
does not affect the locations of the singularities.

Eumple 5.14. Pole/zero pattern of the causal3-sample avenger. Let

H(z) = 1+ z-I/3 + z-2/3 = (1- c1 z- I)(1- C2Z- I)/3
where CI and C2 are the same as in Example 5.13. There are two poles at z = 0
and none at z =00. The pole/zero pattern is shown in Fig. 5.7. For a causal filter
there are no poles at z = 00.

Eumple 5.lS. Pole/zero pattern of the first-order recursive filter. From

1m

eJl13_-1+i,j3
- -2-

l'
C?------liI
I
I

X at z=oo

FIGURE 5-6
Pole/zero pattern for
H(z) = Hz + 1 + Z-I).

172

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

1m

eI2.'3

__
0 _

'--2_ _ Re

FIGURE 5-7
Pole/zero pattern for
H(z) = 1(1 + Z-1 + Z-2).

Example 5.2, the system function is given by

r =z/(z -a)

H(z) = [l-az- 1
In this case, there is a pole at z

=a and a zero at z =0, as shown in Fig. 5.8.

The system function in the previous example is expressed with positive

)()wers of z in the denominator. This latter form tends to be less confusing for
ietennining the locations of the poles and zeros. However, it should not be
!Sed for implementing a filter structure, since only zero and negative
~xponents of z are allowed in the denominator of the system function.
EnmpIe 5.16. Pole/zero pattena of IeCOIld-onIer reauslve filter. From
Example 5.3, the system function is given by

Hz) =
(

1- 'cos(I"IIO)Z-1
1 - 2r cos(1"110) Z-1 +

rz 2

_
1-, cos(l"IIo) Z-1
-1-,(el:-O+e-J:-o)z-l+ rz

_
1-,cos(l"IIo)z-1
- (1-, el:-Oz 1)(1_, e-Jo>oz 1)
The numerator generates a zero at z = r cos( 1"110) and a pole at z =0, while the
denominator generates two zeros at z = 0 and poles at z = r eJaoo, as shown in
Fig. 5.9.

1m

---~:r--""*"- Re

FIGURE 5-8
Pole/zero pattern for H(z) = 1/(1 -

QZ-l).

TIlE zTRANSFORM

173

1m

----------~~--~------Re

rcos(wo)
iI

FIGURE 5-9
Pole/zero pattern for H(z) =
(l-rcos(wo)z-I)/(l- 2rcos(WO)Z-1 + rZ z-z).

;/(re-''''

EumpIe 5.17. MlIIdpIe-order siapIulties in the z: plane. Let

H(z) = Z2 + 2z + 3 + 2z- 1 + z-z

This can be factored to give

H(z) = (z + 1 + Z-I)Z = z2(1 - clz- I )Z(l- czz-I)Z
where
CI

-1 + jV3
-j2J<13
2
=e

and

'The first term, zZ, generates two poles at z = 00 and two zeros at z = o. The two
terms containing Z-I each generate two zeros at z = CI and z = Cz and two poles at
z =0. The net pole/zero pattern is shown in Fig. 5.10.

Rumple 5.18. Zeros IyiDI on drde in the

z: plane. In designing filters, we

commonly encounter the system function having the form

H(z) = 1-z-N

The zeros of H(z) occur at values of z for which Z-N = 1. Since H(z) is an Nth
order polynomial, there are N roots. To find these roots, let
where m is an integer

'Then the Nth root of ZN is equal to

z = ,liN ei(8+2mx)lN

1m
X

z at z=oo

---t--Re

FIGURE 5-10
Pole/zero pattern for H(z) =
zZ+ 2z + 3 +2z- 1 + z-z.

114

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

rm

-1

- -.....-Re

FIGURE 511
Pole/zero pattern for H(z) = 1 - z-N. Case is showr:
for N=8.

-j

The roots have distinct values for m = 0, 1, 2, ... , N - 1. In this example,

Hence, the zeros occur at z = ei2nmlN, for m = 0, 1, 2, ... , N -1. These zeros lie
on a circle having unit radius. They start at z = 1 and are separated by angles
equal to 21C/N. The case for N = 8 is shown in Fig. 5.11.

5.8 THE REGION OF CONVERGENCE

IN mE z PLANE
In the definition of the z-transform, Eq. (5.1), the limits of the summation
index can take on both positive and negative infinite values. The concept of a
z-transform is only useful for values of z for which the sum is finite. The values
of z for which H(z) is finite lie within a region, called the region of
convergence (ROC), in the z plane. Given a sequence {hen)}, the ROC is
defined as the set of z values for which
(5.28)
1l=-CO

Since H(z) becomes infinite at the locations of the poles, the poles
cannot lie in the ROC. In fact, it is the location of the poles that determine the
boundaries of the ROC, as illustrated in the following examples. The ROC is
important for determining whether the filter is stable, and is related to the type
of the time sequence, causal or noncausal, finite-duration or infinite-duration.
These relationships are listed in Table 5.3 and discussed below.
Eumple 5.19. ReKion of convergence for 3-sample averager. Let
h(n) =

{~

for -1 sn s 1
otherwise

Then

~ Ih(n )z -illizi
liz-II
=-+-+-=
3 3
3

.J

11 _ _

THE zTRANSFORM

175

TABLE 5.3

Relationship of the ROC to the discrete-time sequence type

Let us consider h(n), for - NF s n s N p , where N p may be infinite. Let p be the largest finite pole
radius of the system function H(z).
For finite-duration sequences (Np < 00), p = O.
For infiniteduration sequences (Np is infinite). p > O.
Region of convergence
Izl>p
(p<l)

For causal, stable systems (NF SO)

(all poles inside the unit circle)
For causal, unstable systems (NF S 0)
(at least one pole outside
the unit circle)
For noncausal, stable systems (NF > 0)
(NF poles at z = 00)

Izl>p
p < Izl < 00

(p>l)

(p < 1)

This sum is finite for Iz I < co and Izl > O. Hence, the ROC contains alI values of z
for which 0 < Izl < co, or the entire z plane except for z = 0 and z = co.
Example S.2O.

Region of convergence for the &rst-order recursive filter. Let

an
h(n) = { 0

for n 2: 0
otherwise

Then
~

2:

Ih(n)z-"I =

2: laz-II" = [l-Iaz-Ilr

,. ... _00

The sum converges for laz-II < 1 or Izl > 14 Hence, the ROC is the outside of a
circle of radius lal in the z plane as shown in Fig. 5.12.

DETERMINING THE STABILITY OF A DISCRETE-TIME SYSTEM IN THE

z-PLANE. The definition of a stable system was given in Chapter 2 in terms

of the unit-sample response sequence as

oc

L Ih(n)1 <00
n=-et;

(5.29)

FIGURE SOU
Region of convergence (ROC) for first-order recursive filter.

176

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

Systems having a finite-duration unit-sample response are always stable

because the sum of a finite number of finite-valued coefficients cannot become
infinite. Hence, the stability question is addressed to the infinite-duration
component, if any, of the unit-sample response. This component is generated
by the poles of the system function. The question of stability can then be
addressed in terms of the pole locations in the z plane.
Let us consider the simplest form of the general digital filter that contains
only finite-valued poles and all the zeros at z = 0, given by

H(z)

1/1- ~l

ak

z-k

(5.30)

To show explicitly the locations of the poles, we factor the denominator into
the product-of-terms form and perform a partial fraction expansion to obtain
M

H(z)

-ak

= ~11- dkz-1

(5.31)

where both crk and dk may be complex-valued. Since the z-transform is linear,
the unit-sample response is equal to the sum of that produced by each term in
the sum, or
M

h(n)

= L hk(n)

for n ~O

(5.32)

k=l

The unit-sample response of the kth first-order recursive filter is given by

(5.33)
For the system to be stable, each component sequence {hk(n)} must satisfy the
infinite sum condition, or
(5.34)
Applying this condition, we have
00

00

00

L Ihk(n)1 = L IcrkdZI = IcrklL Idkl n

n=O

n=O

(5.35)

n=O

For the last sum, composed of a geometric series, to be finite, the magnitude
of each term must be less than.unity, or Idkl < 1. Hence, we have shown that
each pole of a stable system must lie within the unit circle in the z plane. This is
illustrated in the following examples of two unstable systems.
Example 5.21. First-order uustable system. Consider the unit-sample response
given by
h(n) =

{~

forn~O

otherwise

TIlE z-TRANSFORM

177

Testing for stability, we compute the sum gjven by

~

L Ih(n)l= L 1

,.-0

n-O

which is not finite by inspection. Hence, the system is not stable. The system
function is gjven by
~

n-O

,.-0

L h(n)z-n = L z-n

H(z) =

Applying the infinite geometric sum formula, we get

1
H(z) = - 1
-\

-z

The pole of H(z) lies on the unit circle, at z = 1.

Example S.22. Unstable secondorder system. Consider the unit-sample response given by
h(n) =

{~S(.8n)

for n:2: 0
otherwise

Testing for stability, we compute the sum given by

~

L Ih(n)1 = ,.-0
L Icos(,Bn)1
.

n-O

which is not finite by inspection. Hence, the system is not stable. The system
function is given by
~

H(z) =

L h(n)z-n = L cos(,Bn)z-n

Applying the Euler identities, we have

H(z) = eJ/lnz-n + e-Jllnz- n

2n_0
2n_0

Applying the infinite geometric sum formula to each term, we get

Hz =
1/2 +
1/2
() 1 - eJ/lz -\ 1 - e J/lz \

The poles of H(z) lie on the unit circle, at z = ej/l.

DETERMINING THE ROC FROM THE UNITSAMPLE RESPONSE. The ROC

can also be determined directly from the properties of the unit-sample
response. The filter having a finite-duration unit-sample response has a system
function that can be expressed as a polynomial given by
Np

H(z)

= L

n=-Np

h(n) z-n

(5.36)

178

INI'RODUCITON TO DIGITAL SIGNAL PROCESSING

For such a polynomial system function, poles can exist only at z = 0 and z = 00,
with those at z = 00 being present only for sequences that are noncausal, i.e.
for NF > O. For infinite-duration sequences extending for n ;::: 0, typified by the
unit-sample response of a recursive filter, the system function will contain at
least one denominator term producing a finite pole in the z plane in the region
0< Iz I < 00. Then the ROC is outside the pole having the largest radius.
Infinite-duration sequences extending into negative time result when the ROC
is inside the smallest pole in the z plane. Since these have little practical utility,
they will not be considered in this book.

5.9 DETERMINING THE FRTER

COEFFICIENTS FROM THE
SINGULARITY LOCATIONS
For the filter design procedures described in the following chapters, it will be
useful to determine the filter coefficients from the singularities in the z plane.
As indicated by Eq. (5.19), any filter can be implemented as the cascade
combination of first- and second-order sections. The types of strictly recursive
and nonrecursive elementary filters that are needed for this application are
shown in Fig. 5.13, along with their respective z plane patterns. Let us define ro
to be a real number and (00 a frequency value, where 0 S (00 S 1r. Then the
elementary filters generate singularities of the following four types.

1. A single zero on the real axis at z = ro, where 0 < Irol < 00, with the
corresponding pole at z = O. The system function of the nonrecursive filter
that generates this pattern is given by
H1(z) = 1- ToZ- 1
(5.37)
Z. A single pole on the real axis at z = ro, where 0 < Irol < 1, with the
corresponding zero at z = O. The system function of the recursive filter that
generates this pattern is given by

= [1- roz-1]-1
A pair of complex-conjugate zeros at z = ro ejroo,
H2 (z)

3.

(5.38)

where 0 < ro < 00, with the

two corresponding poles at z = O. The system function of the filter that
generates this pattern is given by
H3(Z) = (1- ro ei roo z- 1)(1- To e-irooz- 1)
(5.39)

4. A pair of complex-conjugate poles at z. = ro ejroo, where 0 < ro < 1, with the

two corresponding zeros at z = O. The system function of the filter that
generates this pattern is given by
Hiz) = [(1- roeiroo z- 1)(1- roe-iCOoz-1)]-1

(5.40)

nm

z-TRANSFORM

179

1m

NON RECURSIVE FILTERS:

FIRST-ORDER
x(n) 0

- - - - l__-o-Re

o-Y(n)
Z-I
-TO

1m

SECOND-ORDER
x(n)

9 r oe ;'"

y(n)

IRe

OrOe- JCdIl

1m

RECURSIVE FILTERS:
FIRST-ORDER
x(n)

}-----.--oy(n)

1m

SECOND-ORDER
x(n)
~--_--y(n)

--_o---Re
2

nGURE 5-13
Relationsbips of simple filters to their pole/zero patterns_

Having implemented the singularities lying in 0 < Izl < 00, those remaining at
z = 0 and z = 00 are then considered. There will always remain an equal
number of opposite singularities at z = 0 and z = 00. If the system is noncausal,
recognized by the unit-sample response starting at n = - NF , then there will be
NF poles at z = 00_ The elementary system function component is then

Hs(z) = ZNF

(5.41)

If the system is causal, but has a delay in the unit-sample response, starting at
=M > 0, then there will be M zeros at z = 00. The elementary system
function component is then
n

(5.42)

The system function of the total filter is then proportional to the product

180

INfRODUcnON TO DIGITAL SIGNAL PROCESSING

of these elementary system functions, or

K

H(z)=An Hk(z)

(5.43)

k=l

where A is the filter gain constant, which cannot be determined from the
pole/zero pattern, and K is the number of elementary systems: By breaking up
a given z plane pole/zero pattern into these elementary types, the coefficients
of the filters generating these singularities can be determined by inspection.
If a pole or zero is needed at z = 0, but not available, a pole/zero
pair can be created there. Then the singularity that is needed is taken, leaving
the opposite singularity at z = 0. The singularities remaining at z = 0, after all
those located in 0 < Izl < co have been taken, have mates at z = co. If there are
N excess poles at z = 0, then these cause a net delay of N samples in the
system function, or z -N. If there are N excess zeros at z = 0; then these cause
a net advance, or ZN.
Example 5.23. Filter coefficients from the z plane singularities. Consider the
pole/zero pattern shown in Fig. 5.14. The system function will be represented by

We first collect the finite, real-valued singularities.

(1) The zero at z = d is paired with a pole at z
H1(z) = 1- dz-

= 0 to produce

Since there was no existing pole at z = 0, a pole/zero pair was created at z

the pole was taken and the zero remained, leaving two zeros at z = o.
(2) The pole at z = a is paired with a zero at z = 0 to produce

= 0,

H2(z) = [1- az-1r l

1m

Z= 00

-----4--~------~----------+_--~~Re

F1GURE 5-14
Pole/zero pattern to be
implemented by a digital
filter.

THE ,TRANSFORM

181

The finite, complex-conjugate singularities are then collected.

(3) The zero at z = e and z = e* and two poles at z = 0 produce
H3(Z) = (1- ez- 1)(1- e*z-l) = 1- (e + e*)z-l + lel 2 Z-2

Since there were no existing poles at z = 0, two pole/zero pairs were created at
z = 0, the poles were taken and the zeros remained, leaving three zeros at z =O.
(4) The poles at z = band z = b* and two zeros at z = 0 produce
H4(Z) = [(1- bz- 1 )(1- b*Z-lW 1 = [1- (b + b*)Z-l + IW z- 2

(5) The remaining pole at z = 00 indicates a noncausal system. The pole at

z = 00 and the zero at z =0 produces
Hs(z) = z

The system function gain factor A cannot be determined from the pole/zero
pattern, but must be given separately.

PROGRAMS TO DETERMINE THE FILTER COEFFICIENTS FROM THE

z-PLANE SINGULARITIES. To determine the filter coefficients from the
locations of the singularities, the following two subroutines are helpful. The
first determines the nonrecursive filter coefficients, and is given by

ZERO(R,DEG,BCOEF,NB)
where
R
DEG

(real input) is the radius of the zero (R > 0);

(real input) is the angle in degrees of the zero with respect to
the real axis (0 s DEG s 180);
BCOEF
(real output array) is the coefficient array for the nonrecursive
filter that implements the zero or zeros; and
NB
(integer output) is the size of the array BCOEF.
The angle is specified in degrees, rather than radians, since it is easier to
specify the singularity location in degrees.
The second program determines the recursive filter coefficients, and is
given by
.

POLE(R,DEG,ACOEF,NA)
where
R
DEG;

(real input) is the radius of the pole (0 < R < 1);

(real input) is the angle in degrees of the pole with respect to
the real axis (0 s DEG s 180);
ACOEF
(real output array) is' the coC(fficiellt array for the recursive filter
that implements the pole or' poles; and
NA
(integer output) is the size of the array ACOEF.
If DEG = 0 or 180, then the singularity lies on the real axis, at z = R or
z = -R, respectively. If DEG*O or 180, then the singularity is complexvalued. In this case, the subroutines should then also include the corresponding complex-conjugate singularity to produce real-valued coefficients. The

182

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

singularities at both z = R ejn(DEG)/180 are then implemented, as demonstrated

in the following examples.
The coefficients produced by ZERO can be used directly by the
NONREC filter subroutine, while those produced by POLE are used by REC.
The complete pole/zero pattern can be generated by cascading these first- and
second-order filter sections. In this case, the output of the previous section
becomes the input to the next.
Example 5.24. Coefficients for a real-valued zero. We want the filter
coefficients that implement a zero at z = roo Then R = Irol. If ro > 0, then
DEG = 0, or if ro < 0, then DEG = 180. The subroutine should determine that the
zero is produced by the filter having the system function

H(z) = 1- roz- I
The call to ZERO(R, DEG, BCOEF, NB) should generate BCOEF(l) = 1,
BCOEF(2) = -ro and NB = 2. The corresponding pole is generated at z = O.
Example 5.25. Coefficient for a real-valued pole. We want the filter coefficients
that implement a pole at z = roo Then R = Irol. If ro> 0, then DEG = 0, or if
ro < 0, then DEG = 180. The subroutine should determine that the pole is
produced by the filter having the system function

H(z)= [1- roz-1r l

The call to POLE(R, DEG, ACOEF, NA) should generate ACOEF(l) = ro and
NA = 1. The corresponding zero is generated at z = O.
Example 5.26. Coefficients for a complex-valued zero. We want the filter
coefficients that implement a zero at z = ro eJWO , where ro > O. Since we want a
filter having real-valued coefficients, the filter should also implement the
complex-conjugate zero at z = ro e -tWo. Then R = ro and DEG = 180wo/:re. The
desired system function is then given by

H(z) = (1- ro eiwoz - 1)(1 - ro e-JwOz- 1 )

= 1 - 2ro cos(WO)Z-l

+ r~z-2

The call to ZERO(R, DEG, BCOEF, NB) should generate BCOEF(l) = 1,

BCOEF(2) = - 2ro cos( 00), BCOEF(3) = r~ and NB = 3. The two corresponding
poles are generated at z = O.
Example 5.27. Coefficients for a complex-valued pole. We want the filter
coefficients that implement a pole at z = ro eiwo , where 0 < ro < 1. Since we want a
filter having real-valued coefficients, the filter should also implement the
complex-conjugate pole at z = ro e -JWO. Then R = ro and DEG = 180w o/:re. The
desired system function is then given by
H(z) = [(1 - ro ei wo z- 1 )(1 - ro e-JwOz-l)rl
= [1- 2ro cos(WO)Z-1

+ r~-2rl

The call to POLE(R,DEG,ACOEF,NA) should generate ACOEF(l) =

2ro cos(wo), ACOEF(2) = -r~ and NA = 2. The two corresponding zeros are
,generated at z = O.

TIlE zTRANSFORM

183

5.10 GEOMETRIC EVALUATION OF

mE z-TRANSFORM IN THE z PLANE
In this section, we indicate how to evaluate R(z) at any point in the z plane
using geometric principles. This procedure is then extended to the evaluation
of the Fourier transform in the next section.
First, let us define a finite singularity as one that is located at a finite value
of z, Le. not at z = 00. The general system function, given by
R(z)

=AzNFNrr;NF
l!l (l- Ck Z - I ) /

DI(l-dk z-

(5.44)

has NF + Np finite zeros located at Cb 1::; k ::; Np + NF, M finite poles located at
dk , 1::; k::; M, and M - N p zeros at z = O. The value of R(z) at some point

z = Zo in the z plane is a complex number that can be expressed as

R(zo) = IR(zo)1 ejArg[H(zol]

(5.45)

As shown in Fig. 5.15, the magnitude can be expressed in terms of the

distances from the finite singularities to the point z = Zo and the phase in terms
of the relative angles.
To provide a formula for the evaluation of the magnitude and phase,
consider the system function having Kp finite poles located at z =Pk for
1::; k ::; K p , and K z finite zeros located at z = qk for 1::; k ::; K z Then the
magnitude is given by
K,

/K

IR(Zo)1 = IAIII D[qb zo] Dl D[Pb zo]

(5.46)

where D[x, y] is the distance between the points x and y in the complex z
plane. The square of this distance is computed by

D2[X, y] = (x - y)(x - y)*

= [Re(x) - Re(y)]2 + [Im(x) - Im(y)]2

(5.47)

1m

FIGURE 515
Geometric evaluation of the system function
H(z) at point z = zoo

184

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

The phase of the complex number is equal to

K.

Arg[H(Zo) =

Kp

L L[qk, Zo) - L Lfpk> zo]

k=l

(5.48)

k=l

where L[x, y] is the angle between the positive real axis and the line
connecting points x and y. This angle can be computed by
L[x, y]

Im(y) - Im(x)

= arctan ( Re(y) _ Re(x)

(5.49)

The value of A cannot be determined from the pole/zero pattern, since the
pattern is not affected by a constant multiplier of the system function. We will
apply this evaluation procedure to determine the Fourier transform in the z
plane in the next section.

5.11 RELATIONSHIP BETWEEN mE

FOURIER TRANSFORM AND mE
z-TRANSFORM
In this section we show the relationship between the transfer function
H(&aJ) and the system function H(z) of a filter. This relationship indicates the
procedure to evaluate the .Fourier transform in the z plane.
The complex variable z can be expressed in the complex-vector form
given by z = r &aJ. Including this form in the definition of the z-transform, we
have

H(r eiaJ)

=L

h(n)r- n e-iaJn

(5.50)

n=-CIO

Note the similarity of this analytic form and the definition of the Fourier
transform of {hen)} given by

H(eiaJ ) =

..

L hen) e-icon
n=-oo

(5.51)

In fact, the two definitions are identical when r = 1. In the z plane, this
corresponds to the locus of points for which Izl = 1. This locus defines a circle
about the origin with radius equal to 1, called the unit circle. Hence, H(e iaJ ) is
equal to H(z) evaluated along the unit circle, or

H(e iCO ) = H(z)lz=ei"

(5.52)
iaJ
For H(e ) to exist, the ROC of H(z) must include the unit circle. As
shown previously, the ROC for a causal filter system function is the region
outside the pole having the largest radius. Since the Fourier transform is
evaluated on the unit circle, this implies that all the poles of a causal, stable
filter must lie inside the unit circle.
To determine the Fourier transform at a particular frequency 0 = 00, we

TIiB .TRANSFORM

185

must evaluate the system function at the point z = ei"'o. In the z plane, this
point is located at the intersection of the unit circle and the line passing
through the origin and making an angle Wo with respect to the positive real
axis. The evaluation is illustrated in the following examples.
Example 5.28. Evaluation of the Fourier transform in the
firstorder recursive filter. Let

plane for the

ROC: Izl > lal

There is a zero at z = 0 and a pole at z = a. For H(e l "') to exist, the ROC must
include the unit circle, which implies that lal < 1. For any frequency w, we
evaluate the z-transform at the point on the unit circle z = e l "'.
Let us consider the particular frequency wo, as shown in Fig. 5.16. The
distance from the zero is

and that from the pole is

D2[a, ei.,O] = (a - cOS(WO2 + sin2(wo)

Hence,
IH(ei"'O)1

= [(a - cOS(WO2 + sin2(woW l12

=[1- 2a cos(wo) + a 2rl12

This is the same result obtained in Example 3.7.

The angle from z =0 to ei"o is equal to wo, while the angle from the pole is

L[a, ei"o] = arctan[sin(wo)/(cos(w o) - a)]

Hence,
Arg[H(el"O)] = Wo - arctan[sin(wo)/(cos(w o) - a)].

Example 5.29. Evaluation of the Fourier transform in the zplane for the
3-sample averager. Let

1m

arctan (

sin(wo) )
IlGs(w o)- a

-I

nGURE 516
-j

Geometric evaluation of the Fourier

transform along the unit circle in the
z plane for the first -order recursive
filter.

186

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

1m

Xz=;x

-+---7'Yi-----'---+- Re
-}

FIGURE 5-17

Geometric evaluation of the Fourier transform

along the unit circle in the z plane for the 3-sample
averager.

-j

From Example 5.13, there are zeros at z = ei2"/3 and a finite pole at z = 0, as
shown in Fig. 5.17. The distance from the pole to eJw is
D2[0, eJW ] = cos 2 ( w) + sin2 ( w) = 1
The product of the distances from each of the zeros to eJw is
D2[ ei2>T/3, el 01] D2[ e - j2,,/3, eJW]
= [2 - ei(W-2lr/3) _ e- J(w-2"13)][2 _

eJ(w+2"/3)

= 3 + 2 cos(2w) + 4 cos(w) = 1 + 4([1

e- J(w+2"/3)]

+ cos(2w)]/2) + 4 cos(w)

= 1 + 4cos (w) + 4 cos(w) = (1 + 2cos(w)?

2

This result is 9 times the value for IH(e iw )12 in Example 5.13. Hence, A =~.
The angle from the pole at z = to eiw is equal to w. The angle from each
of the zeros is equal to

L [e

)2,,/3

-i2lr /3

'W]_

,el

(Sin(2;r/3)-Sin(W))

- arctan cos;r
(2 13) -cos ()
w

(-Sin(2;r/3) - Sin(W))
- arctan cos;r
(2 13) -cos ()
w

JW] _

,e

Then
Arg[H(eiW )] = L[eiZ"/3, eiW ] + L[e- i2 ,,/3, eJW ]

Evaluating the phase, we find

A,g[H(~)l ~ { -;

for -2;r/3 < w < 2;r/3

for 2;r 13 < w ~;r

for -;r < W < -2;r/3

From the previous examples, note the following general points.

Point 1. Since the distance from the origin to the unit circle is unity, the
singularities at the origin do not contribute to. the magnitude response. They
do, however, contribute to the phase response.
Point 2. The magnitude response is zero only when there is at least one
zero of the system function lying on the unit circle at that frequency location.

TIlE zTRANSFORM

187

Izl=l

<[e'(''',e''''''-]

FIGURE S18
Evaluation of the phase response around a zero
located on the unit circle.

Point 3. The phase jump of Jr radians results as the frequency passes

through each zero lying on the unit circle. Let the zero exist at z = eiwo , and let
wi) = Wo - E and wt = Wo + E, for some small E > O. As shown in Fig. 5.18, if a
single zero exists at z = eiwo , then the phase value at w = wi) is Jr radians less
than that at w = wt. If there are multiple zeros, say M, at z = eiwo , then there
is a jump of MJr in the phase as the w passes from wi) to wt. If.M is an
even-valued integer, then the jump is an integral number of 2Jr. In this case,
when the phase is plotted within the principal value range of -Jr to Jr, then no
net change of the phase value is observed.
Point 4. As w increases from 0 to Jr, the phase generated by each zero
that lies strictly inside the unit circle increases by Jr radians. The phase
generated by each pole inside the unit circle decreases by Jr. If the number of
such zeros is Nz and Np is the number of poles, the net increase in phase as w
goes from 0 to Jr is equal to (Nz - Np)Jr.

The same analytic results could have been obtained with much less effort
by merely substituting eiw for z in H(z) to obtain the Fourier transform. The
justification for this geometric evaluation is that the approximate shape of the
magnitude response of a complicated Fourier transform can be readily inferred
from the pole/zero pattern. For example, if a singularity is located close to the
unit circle, it will dominate the frequency response in the frequency range
adjacent to the singularity. This is true because the magnitude response is determined by the ratio of distances, given in Bq. (5.46). If the singularity is close
to the unit circle, then the distances from it can change radically for small
changes in frequency, which are increments along the unit circle, as shown in
Fig. 5.19. This dominant singularity idea is helpful not only for determining the
approximate shape of the frequency response, but also for the placement of
poles and zeros to achieve a desired transfer function, as demonstrated in the
following chapters.

188

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

DOMINANT POLE
1m

DOMINANT ZERO
1m

~'~

~-."

reJ(t)u

re Jwn

Re

Re

/H(e'''')/

/H(e'''')/

1\ ,

,,
\

,
W

1T

Wo

Wo

1T

FIGURE 519
Magnitude response'features for dominant zeros and poles. (r = 0.99).

S.12 mE zT~ANSFORM OF
SYMMETRIC SEQUENCES
In this section, we consider the pole/zero pattern for linear, time-invariant
systems hav.ing unit-sample response sequences that exhibit a point of
symmetry. In Chapter 3, symmetric and antisymmetric sequences were
discussed, with special cases being the even and odd sequences. To simplify the
analysis, we first consider these speci~l cases and then extend these results to
the general symmetric and antisymmetric cases.
EVEN SEQUENCES. Let us consider an even sequence, denoted by {hE(n)},
having duration N = 2M + 1, for some integer M. By the definition of an even
sequence, h(n) = h( -n). The z-transform of {hdn)} is given by
M

HE(z)

= L

n=-M
-1

=
Substituting k

n=-M

h(n)z-n

h(n)z~n

+ h(O) + L h(n)z-n
n=l

(5.53)

= -n in the first sum, we get

M

HE(z)

=L

k=1

h(k)Zk + h(O) +
M

=h(O) + L

n=1

h(n)[zn

L h(n)z-n

n=1

+ z-n]

(5.54)

1HE zTRANSFORM

189

HE(z) is a function of (zn + z-n). :H&nce, if z were replaced by lIz, there is no

change in the function. This condition can be expressed analytically as
HE(z) = HE(l/z)
In terms of the pole/zero pattern, if a zero is located at Zo = '0 ejwo then
another zero is also located at l/zo = e-jw!ro. Hence, real zeros occur in
reciprocal pairs and the complex zeros occur in reciprocal, complex-conjugate
quadruplets. When the zeros lie on the unit circle, they form their own
reciprocals and occur only as complex conjugates. Since the number of poles is
always equal to the number of zeros, half of the poles of HE(z) are located at
z = 0 and the other half lie at z = co. These latter poles correspond to the
noncausal elements in {hE(n)}.
Example 5.30. Pole/zero pattern for an even sequence. Let
h(-2) = 1

h(-l) = -2.5

h(O) = 5.25

h(l) = -2.5
h(2) = 1
and h(n) = 0, otherwise

The system function is equal to

H(z) = Z2 - 2.5z + 5.25 - 2.5z- 1 +

Z-2

= z2(1- 0.5 el2.n/3z-1)(1 - 0.5 e- i2"l3z -1)(1_ 2 el2.n/3z-1)(1_ 2 e- 12.n/3 z -1)

The pole/zero pattern in shown in Fig. 5.20. Note the complex-conjugate
reciprocal pattern of the zeros and the two poles at z = 00.
Example 5.31. Pole/zero pattern for an even sequence witb zeros on tbe unit
circle. Let
h(-l)=l
h(O)=-!
h(l)=l
andh(n)=O, otherwise
The system function is given by
H(z)=Z_~+Z-1

= z(l- e1"/4 z -1)(1- e- 1nl4z- 1)

The pole/zero pattern is shown in Fig. 5.21. Note the zeros on/ the unit circle
form their own reciprocals.

at z =

00

--~----~-L--~--Re

FIGURE 520
Pole/zero pattern for an even sennen"'"

190

INI'RODUcnON TO DIGITAL SIGNAL PROCESSING

1m

at z=oo

--~----~~---+---Re

FIGURE 521
Pole/zero pattern for an even sequence having
zeros on the unit circle.

CAUSAL SYMMETRIC SEQUENCES. For these sequences we must consider

two cases, depending on whether the point of symmetry falls on a sample point
or between sample points. If the point of symmetry falls on a sample point,
this causal sequence, denoted by {hs(n)}, can be obtained from an even
sequence by applying sufficient delay that the first nonzero element starts at
n = O. For the even sequence above, a delay of M samples is needed, or
for O$n $2M
Since delaying a sequence by k samples is equivalent to multiplying the
z-transform by Z-k, we get
(5.55)

This last equation indicates that delaying a sequence by M time units generates
M poles at z = 0 and M zeros at z = 00. These M zeros at z = 00 cancel the M
poles of the even sequence that were there. Hence, all the poles of the
z-transform for a causal finite-duration sequence are located at z = O. The
ROC is then defined by the entire z-plane, except z = 0, or Izl > O.
Example 5.32. Pole/zero configuration for a symmetric sequence having an odd
number of elements. Let
h(O) = 1

h(1) = -2.5

h(2) =5.25

h(3) = -2.5

h(4) = 1
and h(n) = 0, otherwise

Then the system function is equal to

H(z) = 1- 2.5z- 1 + 5.25z- 2 - 2.5z- 3+ Z-4
= (1- 0.5eJ2Jr/3z-I)(1- 0.5e- 12"/3z -I)(1- 2eI2 "l3z -1)(1- 2e-I2JJ/3z-1)

The pole/zero pattern is shown in Fig. 5.22.

If the point of symmetry falls between sample points, the analysis

becomes more complicated. However, we can arrive at the desired result by
the following argument. Let the symmetric sequence hs(n), for O$n $N -1,
where N is an even-valued integer, have the z-transform Hs(z). The point of
symmetry is then Ns = (N - 1)/2, which does not fall on a sample point, as

TIlE z-TRANSFORM

191

1m

-4----------~~~----~--Re

FIGURE 5.22
Pole/zero pattern for a causal symmetric sequence.

shown in Fig. 5.23. Let us now consider the square of the system function, or
H'(z)

= Hs(z) Hs(z)

(5.56)

H'(z) has the same singularity locations as Hs(z), although there are now
twice as many at each point. The equivalent time-domain operation is the
linear convolution of {h(n)} with itself, or
{h'(n)}

= {hs(n)} * {hs(n)}

(5.57)

By performing the convolution of an N-sample duration, symmetric function

with itself, the resulting sequence {h' (n)} has length 2N - 1 and is symmetric
about the sample point n = N - 1, as shown in Fig. 5.23. Hence, H'(z) has
zeros that occur in complex-conjugate reciprocal quadruplets. By the squaring
operation given in Eq. (5.56), the singularities of H'(z) occur an even number
of times at each singularity location. Half of these singularities belong to Hs(z)
when we factor H'(z). Hence, we have shown that Hs{z) also has zeros that
occur in complex-~onjugate reciprocal quadruplets.
Example 5.33. Pole/zero pattern for sequences baving symmetry between
sample points. Let
for O:s n :s 3
hs(n)={~ otherwise
{h(n)}

_ 0--,----:,--1-L-...I.I
1_0_ _ _
;0 II

-I
{h'(nJ}

3 4

3 4

= {h(n)}.:h(II)1

-I

5 6

1/

FIGURE 5-23
The self-convolution of a symmetric sequence that has its point
of symmetry between sampling
points produces a sequence tha,
has its point of symmetry on a
sampling POIDt.

192

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

Im

~--------~~3--------~~Re

Iz I= 1

FIGURE 5-24
Pole/zero pattern for N-point symmetric
sequence, where N is even integer.

shown in Fig. 5.23. The self-convolution sequence is then equal to

for OSn s3
for 4sn s6
otherwise

n+l
h'(n)=

~-n

Note that {h'(n)} is symmetric about n = 3. Hence, H'(z) has zeros that are
complex-conjugate quadruplets with an even number of zeros at each singularity
location. The z-transform of {h'(n)} is equal to

1- 2z- +
4

H'(z) =

Hence,

Z-8

I - 2z- 1 + Z

(1 _

Z-4)2

= -1- Z_-l

l-z- 4
Hs(z) = - 1
-I

-z

The numerator and denominator terms produce the pole/zero pattern shown in
Fig. 5.24. Note that a pole/zero cancellation occurs at z = 1.

The previous example illustrates a general result when N is even. Let us

consider a causal symmetric sequence hs(n), for 0:5 n :5 N - 1, when N is an
even integer. In this case, the highest power of Z-l in Hs(z) is Z-(N-l). In the
previous example N = 4. The terms in Hs(z) having the same coefficient value
can be grouped together as
(N~-l

Hs(z)

= 2.. h(n)[z-n + z-(N-l.-n)]

(5.58)

n=O

Note that when n is even, (N - 1 - n) is odd. Hence, when z = -1, the term in
the brackets is zero. This demonstrates that when N is even, Hs(z) will always
have a zero at z = -1.
ODD SEQUENCES. Applying the same analysis as for the even sequence to an
odd sequence, and recalling that ho(O) = 0, we obtain
M

Ho(z)

= ~ ho(n)[z-n - zn]
n=l

(5.59)

TIlE ,TRANSFORM

193

If z is replaced by liz, we get

M

Ho(1/z) =

2: ho(n)[zn -

z-n]

= -Ho(z)

(5.60)

n=1

Since multiplying the z-transform by the constant A = -1 does not change the
locations of the singularities in the z plane, the result here is the same as the
case for the even sequence: if there is a zero at z = Zo, then there is also one at
z = l/zo. Hence, the zeros for an odd sequence also occur in complexconjugate quadruplets.
Example 5.34. Pole/zero pattem for an odd sequence. Let
h(-3) = 1

h(-2)=2

h(-1)=3

h(O) = 0
h(3) =-1

h(1) =-3

h(2) =-2
and h(n) = 0, otherwise

The z-transform is then given by

H(z) =

Z3

+ 2Z2 + 3z - 3z- 1 - 2z- 2 - Z-3

The locations of the zeros are at

z = 1, -1, 1. 7ejO.644", and (1/1.7)ejO.644"
and three poles are at z = 0 and three (due to the noncausal component) are at
z = 00, as shown Fig. 5.25.

CAUSAL ANn-SYMMETRIC SEQUENCES. As above, an odd sequence can

be converted into a causal antisymmetric sequence, denoted by {hA(n)} , by

(h(n)}

----~~----~~----~~----Re

ftGURE S25
Pole/zero pattern for an odd-symmetric
seouence.

194

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

applying sufficient delay:

hA(n) = ho(n - M)

for OSn s2M

Then

HA(z) = z-MHo(z)
The M zeros at z = 00 produced by Z-M cancel all the poles of the odd sequence
that were there. As expected, all the poles of a causal finite-duration sequence
are located at z = O.
Again, we must consider the two cases when the point of antisymmetry
falls on an index value and when it falls between index values. The results are
exactly analogous to the causal symmetric results above. We present an
example of the latter, more complicated, case.
Example 5.35. Pole/zero pattem for an antisymmetric sequence. Let
h(O) = 1

h(1)=2

h(2)=-2

h(3)=-1

andh(n)=O,

otherwise

The self-convolution sequence is then equal to

h'(O) = 1

h'(1)=4

h'(2) = 0

h'(3) =-10
h'(6)"; 1

h'(4) = 0

h'(5) = 4

and h'(n) = 0, otherwise

and is shown in Fig. 5.26. The sequence {h'(n)} is symmetric about n = 3. Hence,
H'(z) has zeros that are complex-conjugate quadruplets with an even number of
zeros at each singularity location. The z-transfol'Ill of {h'(n)} is equal to

H'(z) = 1 + 4z- 1 -lOz- 3 + 4z-' + Z-6

The locations of the zeros were found by finding the roots of the polynomial. The
{h(n)}

{h'(n}}={h(n}}

* {h(n)}
4

-1

0 1 2

4 5 6 7 8

-10

FIGURE 5-26
The self-convolution of an antisymmetrc sequence that has its
point of symmetry between sampling points produces a sequence
that has its point of symmetry on
a sampling point.

THE ,TRANSFORM

l.95

1m

H'(z)

Izl=1
2

~~--------r-----~~--~---'2~Re

-2.618

1m

~----------~----~~------~--Re

-2.618

nGURE 5-rI
Pole/zero pattern for an
antisymmetric
sequence
whose point of symmetry
falls between sample points.

locations of the double zeros are at

z=l

-2.618

and

-1/2.168

and six poles are at z = 0, as shown in Fig. 5.27. Hence, the z-transform of the
original sequence, given by
HA(z) = 1 + 2Z-I- 2z- 1 - Z-3
has zeros that occur in reciprocal pairs (zeros on the unit circle form their own
reciprocals) .

One important consequence of the above result is that we can determine

by inspection of the pole/zero pattern whether a finite-duration sequence is
symmetric. For the z-transform of a finite-duration symmetric sequence, the
singularities in the region 0 < Iz I < 00 are all zeros and occur in complexconjugate reciprocal quadruplets, and all the poles lie at z = 0 or at z = 00.
Zeros at z = 1 and z = -1 form their own compfex-conjugates and reciprocals.
APPLICATION TO LINEAR-PHASE FILTERS. In Chapter 3 it was shown that
both symmetric and antisymmetric unit-sample responses produce linear-phase
responses. The above results indicate the pole/zero patterns corresponding to
symmetric or antisymmetric unit-sample responses can easily be recognized.
Hence, the system function of a linear-phase filter has a pole/zero pattern that
consists of zeros that occur in complex-conjugate quadruplets. The poles lie at

196

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

z = 0 and z = 00, the number at z = 00 being equal to N F , the number of

noncausal elements in the unit sample response.

S.13 THE INVERSE z TRANSFORM

In this section, we consider the converse problem: Given H(z) and the ROC,
how do we determine the discrete-time sequence {h(n)}? This question is
important for determining the unit-sample response of a system from its
pole/zero pattern in the z plane. From the definition of the z-transform, we
note that h(n) is the coefficient of z-n. Hence, if H(z) is expressed as a
polynomial in powers of z, determining the sequence is a trivial matter, since
we need only to equate h(n) with the coefficient of z-n, as shown in the
following example.
Example 5.36. Inverse operadon on a polynomial z-transform. Let

H(z) = az + b + cz- I
From the definition of the z-transform, we find by inspection that h( -1) = a,
2: 2 and n 5 -2.

h(O) = band h(l) = c, and h(n) = 0 for n

For more complicated functions of the z-transform, three methods are

usually employed to determine the values of the discrete-time sequence: long
division, Taylor expansion and application of the residue theorem from
complex-variable theory. The first method is the simplest for determining the
first few terms of the sequence. The second and third methods are more
complicated, but provide the functional form of the sequence. This functional
form then allows the value of any term of the sequence to be computed and
also indicates the trends that are present in the sequence. Each method will be
described separately below and applied to the same z-transforms for
comparison.
LONG-DIVISION METHOD. Since h(n) is defined as the coefficient of z-n, we

need only to expand H(z) into a polynomial in powers of Z-l. The coefficient
of z-n is then equal to h(n). We illustrate the application of this method to
various types of sequences in the following examples.
Example 5.37. Long-division method for causal sequences. Let
H(z)=l"

-az

-'. z
_I

z-a

ROC:lzl>a

Since the ROC is outside the pole, we should obtain a right-Sided sequence, i.e. a
polynomial in z-n for n 2: O. To accomplish this, we perform the following
division:

TIlE .1RANSFORM

197

Since h(n) is the coefficient of z-", we find

a"

h(n) = { 0

for n?:O
otherwise

EDIIlpie 5.38. Long-division method for noncausai sequence. Let

ZK
H(z)=1_bz-1

ZK+l
z-b

ROC: b<lzl<oo

Since the ROC is outside the pole, we should obtain a right-sided sequence, but
because there are K poles at z = 00, the sequence starts at n = - K. To accomplish
this, we perform the foilowing division:
ZK+l
__ = ZK + bZ K- 1+ b2z K- 2 + ... + bK + b K+1Z- 1+ bK+2Z- 2+ ...
z-b

Since h(n) is the coefficient of Z-II, we find

b ll + K
h(n) = { 0

for n?:-K
otherwise

Example 5.39. Long-division method for a second-order system. Let

1- b 2z- 2
H(z) = 1 + a2z-2

ROC:

Izi > a

Since the ROC defines the region outside the pole, we should expect an infinite
duration sequence extending for positive values of n. To obtain such a right-sided
sequence, we perform the division to obtain
H(z) = 1- (a 2 + b2)z-2 + a2(a 2+ b2)z-4 - ..
Equating the coefficient of Z-II )vith h(n), we have

h(O) = 1
h(1) = 0
h(2) = _(a 2 + b 2 )
h(3) = 0
h(4) = a 2 (a 2 + b 2 )

This result can be expressed analytically as

~ )~"'-~a' +.)

h(n) - { -1

forn=O
for n > 0 and even
otherWise

The long-division method is robust in that it can always be applied to any

rational function of z. By observing several terms of the sequence the analytic
form of the series can often be guessed, as in the examples above, or
approximated by some simple functional form. The two methods described

198

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

below provide analytic expressions for h(n), from which any element of the
sequence can then be determined.
TAYLOR SERIES EXPANSION, MEmOD. Recall that for a complex number

c, where

lei < 1,

then the Taylor series expansion is given by

1
'"
-=
en
1- e n=O

2:

(5.61)

The trick in successfully applying this method is to express the z-transfotm in

the appropriate form. To apply the Taylor series method, the denominator of
the z-transform must be factored to find the locations of the poles. Then a
partial fraction expansion can be performed to obtain the sum of first-order
terms with possibly complex coefficients. The series expansion can then be
calculated for each term. Since the z-transform is a linear transformation, the
coefficients of the z-n term from each expansion can be added to find h(n).
This is illustrated in the following examples.
Example 5.40. Taylor series method for determining the
transform. Let
1
ROC: Izl >a l / 2
H(z)=1
2

inverse

z-

-az

In the ROC, laz- 1< 1. Hence, the formula can be applied directly to express
H(z) in the following form:
~
~
2
H(z) = L (az- )" = L a"z-2n
"=0

PI-O

Since h(n) is the coefficient of z-", we can write

ann

for n 2= 0 and n even

otherwise

h(n)= { 0

Example 5.41. Taylor series method applied to a second-order system. Let

1- b 2z- 2
H(Z)=1
22
+az

ROC:lzl>a

The z-transform above can be factored into

1
H(z)=I+a 2z
~

b 2 z- 2
l+a2z 2
~

(_a 2z- 2)" - b 2z- 2L (-a 2z- 2)"

"=0

Combining terms having the same exponent value, we get

H(z) = 1- (a 2 + b 2)z-2 + a2(a 2+ b2)z-4 _ a4(a 2+ b2)z-6 + ...
This is the same result produced by the long-division method.

mE ,TRANSFORM

199

The following example shows that a partial fraction expansion can

simplify the application of the Taylor series method .
Example 5.42. The inverse z-transform of a second-order system. Let
, sin(wo)

Z- I

H(z) - ---""----'-'--:-----::---::
-1- 2, cos(wo) z 1+ ,2 Z 2
Factoring the denominator into the product of two terms and then performing the
partial fraction expansion , we get

Applying the Taylor series expansion to each term , we have

Combining the coefficients of z-' , we get

2: ,. sine won)z - '

.-0

Equating the coefficient of z - . with h (n), we finally get

hen) = {~. sin(won)

for n 2: 0
othenvise

RESIDUE THEOREM METHOD. This method is analytically the most

complicated of the three, using results from complex-variable theory. In a
first-level course, this section can be skipped without loss in continuity'. From
complex-variable theory, it can be shown that a direct relationship exists
between the nth element of a discrete-time sequence {hen)} and its ztransform H(z). This relationship is given by
hen)

= ~ 1 H(z)zn-l dz
2JrJ jc

(5.62)

where the integration is taken in a counterclockwise direction along the

contour C that lies in the region of convergence of H(z) and completely
encloses the origin, as shown in Fig. 5.28. The direct evaluation of the above
integral is usually difficult, and so a second result, called the Residue Theorem ,
is usually employed. This theorem is stated without proof, which can be found
in the references.

200

INTRODUcnON TO DIGrrAL SIGNAL PROCESSING

1m

Re
FIGURE 5-28
Contour of integration in the region of .convergence
for determining the inverse z-transfonn.

Residue theorem. If H(z)zn-l has N poles, denoted by PI' P2, ... PH. then
h(n) =

L Res[H(z)zn-l

at pole PI]

(5.63)

1-1

where Res[ . ] is the residue.

To define the residue, let H(z)zn-l be a rational function (the ratio of
two polynomials) of z, having an mth-order pole at z =Zo. We can then
express this mth-order pole explicitly as

H( z )z n-l = ( P(z))m
Z-Zo

(5.64)

=Zo.

The residue of H(z)zn-l at

where P(z) does not have any poles at z

z = Zo is then defined to be
Res[H(z)zn-l

at z = Zo] = (

1
dm-lp(z)
-1)1 d m-l
m
.
Z
z-zo

(5.65)

If m = 1, there is only one pole at z = Zo. The evaluation then simplifies to

Res[H(z)zn-l at z = Zo] = P(Zo)

(5.66)

The following examples illustrate the application of the Residue Theorem.

Eumple 5.43_ Residae theorem method for a noncausal sequence. Let

H(z) = z + a + Z-I
The contour C is located in the ROC, which is the entire z plane except for z = 0
and z = 00. Then

H(z) Z"-1 = z" + az"-I + Z"-2

For ,.=-2
P(z)

"'7

mE zTRANSFORM

201

There is only a fourth-order pole at z = O. Evaluating the residue, we get

h(-2)=~ d P\z) I
3

3. dz

=0

z-o

The same results would be obtained for n < - 2.

For II =-1

P(z)

7
There is only a third-order pole at z = O. Hence

h(-l) =1-d-P(z)
2=1
2 dz
z=o
For II =0

P(z)

7
There is only a second-order pole at z = O. Hence

h(O) = dP(z)
= a
dz z-o
For 11=1

H(z)=z+a+z- 1 =

z2+az + 1 P(z)
=-

There is only a first-order pole at z = 0 that is inside C. Hence h(l) = p(z)lz_o = 1

For 11=2

H(z) z =

Z2

+ az + 1

There are no poles inside C. Hence h(2) = O. This is also true for n

EUIIlple 5.44. Residue theorem method for

1
H(z)=l_az

2.

causal sequence. Let

ROC: Izl>a

To enclose all the poles, the contour C is situated in the region of convergence, as
shown in Fig. 5.29. Then

z"

H(z) Z,,-l = - z-a

For n > 0, C encloses only one pole, at z = a. Then

P(z) = H(Z)Z,,-l(Z - a) = z"

Evaluating the residue, we get

h(n) = P(a) = a"

for n >0

202

INTRODUCI10N TO DIGITAL SIGNAL PROCESSING

1m

-+-----.--*--f-- Re
FIGURE 5-29
Contour of integration for the. first-order recursive
filter.

For n < 0, C also encloses additional poles at z = O. For n = -1:

1
z(z-a)

H(Z)Z"-l=~_...,.

Then,

h(-l) = Res[H(z)z .. -l at z =0] + Res[H(z)z,,-l at z =a]

The residue at z = 0 is equal to
1

P(z)=z-a
The residue at z

P(O) = -l/a

=a is equal to
P(a) = l/a

P(z) = l/z

Adding the results of the two residues, we get

h(-l) =0
In a similar manner, it can be shown that h(n) =0 for n < O.
Example 5.45.

Residue theorem method for a second-order system. Let

1- b 2z- 2

H(z)=1+a 2z

ROC: Izl>a

If H(z) contains Np poles, then

h(n) = ~ Res[H(z) Z,,-l at pole Pk]

k-l

We first start with

H(Z)Z,,-l

(Z2 _ b 2)z,,-1

=-:-,--,.....,..:--...,.
(z + ja)(z - ja)

For n <0, h(n) = 0, because the sequence is right-sided.

For n

= 0,

H(z) Z,,-l has poles at z

= 0,

ja and -ja. Hence, the sum of the

TIlE zTRANSFORM

residues is
2
Z2 _ b 1
h(O) =-2--2
Z

+0

z-o

+b

(Z2 _

+ JO)
O~ + b 2

Z(Z

b 2)

z-ja

b 2)

(Z2 -

.)

Z(Z - JO

203

I
z=-ja

=-1
2 +-2 +--=
2
20

20

For n > 1, H(z) Z,,-l has poles only at Z = jo and -jo. Hence, for n > 1
h(n) =

(Z2 _

b 2)z"-1

(Z + JO.)

(Z2 -

+ (Z
Z-JI'

b 2 )z,,-1
-

.)
JO

2
= (0 -: b ) [( _jO)"-l _ (jO)"-l]
2Jo

z=-ja

for n > 1

Evaluating this equation produces the same values for {h(n)} as given in
Example 5.39.

5.14 SUMMARY
In this chapter, procedures for determining the z-transform of a discrete-time
sequence and for performing the inverse operation have been described. The
system function of a digital filter has been defined as the z-transform of its
unit-sample response. Cascade and parallel combinations of elemental digital
filters can be formed by factoring the system function.
The complex z plane was defined. The configuration of the poles and
zeros of the system function in the complex z plane were used to determine
whether the filter is recursive or nonrecursive. Stable causal filters were found
to have all their poles inside the unit circle, while stable noncausal filters also
had poles located at z = co. The region of convergence (ROC) was employed to
determine the type of the discrete-time sequence, either causal or noncausal,
finite-duration or infinite-duration. The Fourier transform was shown to be the
evaluation of the z-tr~sform along the unit circle in the z plane. Three
methods to determine the discrete-time sequence from the z -transform were
also presented.

FURTHER READING
Antoniou, Andreas: Digital Filters: Analysis and Design, McGraw-Hill, New York, 1979.
Hamming, R. W.: Digital Filters, 2d ed., Prentice-Hall, Englewood Oiffs, NJ, 1983.
Jong, M. T.: Methods of Discrete Signal and System Analysis, McGraw-Hill, New York, 1982.
Jury, E. I.: Theory and Application of the z-Transform Method, Wiley, New York, 1964.
Oppenheim, Alan V., and Ronald W. Schafer: Digital Signal Processing, Prentice-Hall,
Englewood CIi1Is, NJ, 1975.
Peled, A., and B. liu: Digital Signal Processing: Theory, Design and Implementation, Wiley, New
York, 1976.
Rabiner, Lawrence R., and Bernard Gold: Theory and Application of Digital Signal Processing,
Prentice-Hall, Englewood Cliffs, NJ, 1975.

204

INfRODUcnON TO DIGITAL SIGNAL PROCESSING

REFERENCE TO TOPIC
Complex variables
Churchill, R. V.: Complex Variables and Applications. McGraw-Hill. New York. 1960.

PROBLEMS
5.1. Given the following unit-sample response sequences, compute the z-transform,
sketch the pole/zero pattern in the z plane and the filter block diagram.
(a) h(-I) = 1, h(I) = 1, h(n) =0, otherwise
(b) h(O) = 1, h(I)=-I, h(n) =0, otherwise
(c) h(-I)=I, h(0)=-2, h(I)=I, h(n)=O, otherwise
(d) h n)={r"Sin(won), forn~O,
(
0 otherwise. Let Wo = n/8 and r = 0.9
5.2. For the following difference equations, compute the z-transform of {h(n)},
sketch the pole/zero pattern in the z plane and the filter block diagram.
(a) y(n)=x(n)-x(n-N). LetN=5
(b) y(n)=x(n)+x(n-N). LetN=5
(c) y(n) = 2rcos(8)y(n -I)+r2 y(n -2)+x(n). Let (J=n/4 and r=0.9
5.3. For the digital filter structures shown in Fig. PZ.ll, compute the z-transform,
sketch the pole/zero pattern in the z plane and the filter block diagram.
5.4. For the following magnitude and phase functions, compute the z-transform,
sketch the pole/zero pattern in the z plane and the filter block diagram.
(a) IH(eiw)1 =2sin(w/2)
for O:s w :s2n
iW
Arg[H(e )] = n/2 - w/2
for 0 < w < 21r
(b) IH(eiW)I=[5+4cos(2w)]112,
Arg[H(ei",)] = arctan[ -sin(2w) ],
2+cos(2w)

Arg[H(eiO)) = 0

(c) IH(ei"')1 = [17 - 8 COS(2W)]112,

j ..
_
[4Sin(2W)]
Arg[H(e )]-arctan I-4cos(2w) ,

Arg[H(ei?)] = 0

(d) IH(eJW)1 = [1.81-1.8 COS(W)]ll2,

..
[ 0.9 sin(w) ]
Arg[H(e1 )]=w+arctan I-0.9cos(w)'

Arg[H(ei~] =

5.5. For the z plane pole/zero patterns shown in Fig. P5.5, compute the z-transform,
sketch the filter block diagram, determine the transfer function and sketch the
unit-sample response sequence. Let the filter gain A = 1.
5.6. Determine the pole locations of the systems having the following unit-sample
response and determine whether they are stable.
(a) h(n) = {O(-l)"
(b) h(n) = {O(-l)"

for O:sn:sN -1, for any finite value of N

otherwise

forn~O
otherwise

TIm zTRANSFORM

(a)
1m

(b)

Izl=1

1m

...().---l~--4--+--<l

ROC:lzl>a

1m

ROC: Izl>O

1m

(Ii)

--~--~~---4-Re

--~--4~-~-Re

z=-1

ROC: IzI>O
1m

ROC: Izl> 0

1m

z=oo

-+---.......--~.Re

ROC: O<lzl< 00

Re

-2

(c)

205

Re

Izl=1
ROC:lzl>o

FIGURE PS.5
Pole/ufO patterns.

5.7. Determine the possible values of the phase response at I) = O. Indicate the
pole/zero pattern that can cause it. Repeat for I) = n.
5.8. Demonstrate that the z-transform of any antisymmetric sequence has a zero at
z=1.
5.9. Consider the z-transform given by
X(z) = sin(1/z)
Find a sequence {x(n)} that has this z-transform by performing a Taylor series
expansion.
5.10. Consider the sequence {x(n)} and its z-transform X(z).

206

INTRODUcnON TO DIOITAL SIONAL PROCESSING

(a) Find the z-transform of the sequence

{y(n)}

= {a"x(n)}

where

11'

is a positive real number

in terms of X(z).
(b) If X(z) has a pole inside the unit circle at radius r, what is the value of 11',
such that the pole of Y(z) falls on the unit circle?
5.11. Consider the filter with unit-sample response {hen)} and input {x(n)} given by
h(n)=

1
-01

forn=O
forn=8
otherwise

and
forn~O

x(n)=l

=0

otherwise

(a) Determine the output {yen)} by performing the discrete-time convolution.

(b) Determine the output {yen)} from the inverse z-transform Y(z) =
H(z) X(z). Note that the Fourier transform of {x(n)} does not exist.
S.U. Consider the system function given by
1

H(z) = 1_p- Z-l

1- pZ-l

for O<p< 1

(a) Draw the pole/zero pattern of this system function.

(b) Show that IH(eiOl ) I has a constant value for all w.
S.13. Consider a second-order filter having a finite-duration unit-sample response. Let
its magnitude response IH(eIOl)I, over -1r:S W < 1r, equal zero at two frequency
points w = Wo and w = - Wo.
(a) Show that the magnitude response for all frequencies can be determined to
within a gain constant from this information.
(b) Draw the pole/zero pattern for the filter that I}as both poles at z = o.

COMPUTER PROJECfS
5.1. DetermiDing the filter coefficients from the t: plane singularity location
Object. To determine the filter coefficients and the time-domain sequence from
the singularities in the z plane.

Write the two subroutines described in this chapter given by

ZERO(R,DEG,BCOEF,NB)
and

POLE(R,DEG,ACOEF,NA)
Verify the operation of your subroutines by generating the coefficients for the
following causal filters. For each filter provide the following:
(i) the sketch of the block diagram of the digital filter
(ii) the sketch of the z plane pole/zero pattern

TIlE zTRANSFORM

2fY1

(iii) the computer graph of the unit-sample response

(iv) the computer graphs of the first 65 points of the magnitude and phase
response computed from the 128-point Off.
The filters ~e:
(a) the causal3-sample averager
(b) the first-order recursive filter, with a = -0.9,
(c) the second-order filter described by

y(n) =2rcos(wo)y(n -1) -ry(n -2) +x(n) -TCOS(wo)x(n -1)

with Wo = n /10 and T = 0.95.
Generate 128 points of {h(n)}, but plot only the first 55.
Sol. Mapltude and phue response CUl'Ye8 for simple poles and zeros
Object. To determine the magnitude and phase response curves for simple poles
and zeros in the z plane.
AaaIytk results
(a) Consider a filter having a real zero located at z = To and a pole at z = O. Show
that the magnitude response of this filter is identical to within a gain constant
to that of the filter having a real zero at z = l/To and pole at z = O. Repeat the
analysis for a pair of complex-conjugate zeros located at z = Toejeuo and two
poles at z = O.
(b) Determine the magnitude response of a filter having a zero at z = To and a pole
at z = (l/To), where ITol > 1.
Computer results

Use ZERO and POLE to determine the filter coefficients for the following
singularity locations. Use NONREC or REC to determine the unit-sample
response (since all the filters are causal, use the unit-sample sequence without any
delay). Use FFf to find the 128-sample Fourier transform, followed by MAGPHS
to find the magnitude and phase response over the 65 points covering 0 s w s n.
Use PL01MR and PLOTPR to plot the curves.
For each part below include the following:
sketch of the filter block diagram
plot of the unit-sample response
analytic form of the transfer function and system function
sketch of the pole/zero pattern in the z plane
the magnitude and phase response plots
(1) Real zero at z = i and pole at z =O. Repeat for zero at z = t. Explain the
difference in the two phase responses.
(2) Complex-conjugate zeros at z = 0.9 ejJJl8 and two poles at z = O. Repeat for
zeros at z = 1/(0.9 ejJJ,,) and two poles at z = O. Explain the difference in the
two phase responses.
(3) Real pole at z = -0.9 and zero at z = O. Repeat by moving the zero from z = (')
to z = -1/0.9.
(4) ~mplex-conjugate poles at z =0.95 ej..18 and two zeros at z =O. Repeat by
moving one of the zeros from z = 0 to z = 0.95 cos(n/8).

CHAPTER

6
DIGITAL
FILTER
STRUCTURES

6.1 INfRODUcnON
In this chapter, we describe the most commonly employed methods to
impleme~t a digital filter structure from either its di~erence equation, its
unit-sample response, or its z-transform. We start by defining the terms
conventionally used for describing digital filters and relating them to the
system functiQn coefficients. The digital filter structure determined directly
from either the difference equation or the system function is called the Direct
Form I. An alternate view of the same equations results in the memoryefficient structure, called the Direct Form II. Digital filters structured as
cascade and parallel combinations of second-order components are shown to
have advantages in terms of hardware implementation and numerical sen.sitivity. The above methods are employed to implement the general filter
structure, one that contains both poles and zeros. An all-zero filter having a
linear-phase response is shown to have a special nonrecursive structure that
reduces the number of multiplications by approximately one-half. The
frequency-sampling structure is presented as an alternative to the tapped-delay
line structure for all-zero filters.

DIGITAL FILTER STRUcruRES

209

6.2 SYSTEM-DESCRIBING EQUATIONS

The equations that describe the input! output relations in the time and
z-transform domains have been defined in the previous chapters. They are
repeated here for reference.
In the time domain, the convolution equation is given by

2::""

yen) =

h(k)x(n - k)

(6.1)

k=-tIC

and the difference equation is given by

y(n)=faky(n-k)+

2 bkx(n-k)

(6.2)

k=-~

k=l

In the z-transform domain, the system function can be expressed in two useful
forms. The first is the sum-of-products form given by

H(z) =

~
k=-~

bk Z-k/1-

ak Z-k

(6.3)

k=l

The second is the product-of-terms form given by

H(z) =Az~

Nft (1 - CkZ-1)/ Ii: (1 - dkz-1)

k-l

(6.4)

k-l

We describe several different structures that can be used to implement the

digital filter from anyone of these defining equations. First, we discuss the
terms commonly used for classifying the filter types.

6.3 FILTER CATEGORIES

Digital filters are often categorized either by the duration of their unit-sample
response or by their structure. When a filter produces a unit-sample response
that has an infinite duration, it is called an infinite-impulse response (IIR) filter.
'Ibis term originated when it was popular to refer to the unit-sample response
as the impulse response, the conventional term for analog filters. If the digital
filter has a unit-sample response having a finite duration, then it is called a
finite-impulse response (FIR) filter.
Digital filter classification can also be based on the filter structure. In
general, the output of a filter can be a function of future, current and past
input values, as well as past output values. If the output is a function of the
past outputs, then a feedback, or recursive, mechanism from the output must
exist. Hence, such a filter is referred to as a recursive filter. A recursive filter
can be recognized from the defining equations, since at least one ak coefficient,
for 1:5: k:5: M, in Eqs. (6.2) and (6.3) is nonzero, and at least one dk
coefficient, for 1:5: k:5: M, in Eq. (6.4) is nonzero.
If the filter output value is a fu~ction of only the input sequence values, it

210

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

is called a nonrecursive filter. A nonrecursive filter can be easily recognized

from the defining equations, for which ak 0 in Eqs. (6.2) and (6.3), or dk = 0
in Eq. (6.4), for all k. As shown in the previous chapter, all the poles of a
nonrecursive filter are at either z = 0 or z = 00. In the geometric evaluation of
the transfer function, they did not contribute to the magnitude response.
Hence, they are said not to be relevant, and the nonrecursive filter is also
called an all-zero filter. The poles at z = 0, however, do contribute to the
phase response, by introducing linear phase.
To achieve an infinite-duration unit-sample response, some form of
recursive structure is necessary. Hence, the terms IIR and recursive are
commonly accepted as being interchangeable. Similarly, finite-duration unitsample responses are typically implemented with nonrecursive filters. Hence,
FIR and nonrecursive are also usually interchangeable. An exception to this
latter case is the frequency-sampling structure for FIR filters, described later in
this chapter. In this filter, a recursive structure is employed to generate a
finite-duration unit-sample response.
Several additional terms, used mostly in signal modeling, statistics and
system identification, are autoregressive (AR), moving-average (MA) and
autoregressive/moving-average (ARMA). The FIR filter can be viewed as an
MA system, while the general IIR filter, having both poles and zeros, is an
ARMA system. An AR system is an IIR filter whose system function has all its
zeros at z = 0, that is bk = 0 for k =I: 0 and Ck = 0 for all k. An AR filter is also
commonly called an all-pole filter, since all the zeros are located at z = 0, and
hence are not relevant, in that they do not contribute to the filter magnitude
response.

6.4 THE DIRECT FORM I AND

STRUCfURES

The digital filter structures implemented with the procedures described in the
previous chapters provide separate delay buffers for both the input and output
sequences, as shown in Fig. 6.1. This implementation is caIled the Direct Form

x(n)_-+---i

wen)

ftGURE 6-1
Direct Form I implementation of a digital filter.

~---_... y(n)

DIGITAL F1LTER STRUCfURES

211

I. Since it shows explicitly the delayed values of the input and output, the
Direct Form I structure is a convenient way to implement a digital filter with a
computer program. One subroutine (NONREC) is used to implement the
zeros of the system function, followed by a second (REC) to implement the
poles. However, this form does not make the most efficient use of memory. A
second structure that performs the same transformation but uses fewer
memory elements is considered next.
This alternate structure is most easily achieved by considering the system
function expression given in Eq. (6.3).
The system function is then divided into two parts connected in cascade,
the first part containing only the zeros, followed by the part containing only
the poles. As shown in Fig. 6.1, we can introduce an intermediate sequence
{w(n)} representing the output of the first part and the input to the second.
Hence, we write
(6.5)

and
M

y(n) =

L ak y(n -

k)

+ w(n)

(6.6)

k=l

In the z-transform domain, the equivalent relationships are given by

W(z) = X(z)

bkz- k

(6.7)

k=-iYF

and
(6.8)
Since we are dealing with linear systems, the order of these two parts can be
interchanged, as shown in Fig. 6.2(a), to generate the same overall system
function, but through a different intermediate sequence {u(n)}. The equations
for this second system are then given in the time-domain by
M

u(n) =

L ak u(n -

k) +x(n)

(6.9)

k=l

and

Np

y(n)

= L

bku(n - k)

(6.10)

(6.11)

k--iYF

In the z-transform domain, we have

U(z) = X(Z)/I-

k-l

akz-k

212

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

(a)

x(n)

yIn)

nGURE 6-2
Two step procedure to obtain the Direct Form II digital filter structure. a) Interchange recursive
and nonrecursive parts. b) Combine common delays.

DIGITAL FILTER STRUCfURES

213

and
Np

Y(Z)

= U(Z) 2:

bkz- k

(6.12)

k--Np

By examining the outputs of the delays in Fig. 6.2(0), we note that their
contents are identical. Hence, we can eliminate one set of delays, as indicated
in Fig. 6.2(b), to obtain a filter structure that uses fewer delay elements. This
filter structure, called the Direct Form II, uses a smaller number of delay
elements, equal to the maximum of M or Np + Np , than the Direct Form I,
which uses M + Np + Np elements. Since we need K delays to implement a
system function term containing Z-K, the Direct Form II uses the minimum
number of delays required to implement 0 given system function. It is therefore
said to be a canonical structure.
EuIapIe 6.1. Dired Fona n secoad-order filter Implementation. Let
y(n) = 2r cos(coo) y(n -1) - r2 y(n - 2) + x(n) - r cos(coo)x(n -1)

The corresponding system function is given by

Hz =
l-rcos(coo)z-1
() 1-2rcos(coo)z 1+,-2z-2

Y(z)
X(z)

The Direct Form I structure is obtained by factoring the system function to give

where H 1(z) is the numerator of H(z) and H 2(z) is the remainder. The
intermediate sequence {w(n)}, having z-transform W(z), is introduced between
these two sections as
H 1(z)

=1- rcos(coo) Z-1 = ~g?

and
H 2(z)

= (1- 2r cos(coo) Z-1 + ,-2z-2)-1 =;~:~

The order of the component filters is reversed to give

Y(z) = H2(z) Hl(Z)X(Z)

A second intermediate sequence {u(n)}, with z-transform U(z), is introduced

between these two sections to give
H2(z) = (1- 2r cos(co o) Z-1 + ,-lz-2)-1 = ~g~

and

214

JNTR.0DUCI10N TO DIGrrAL SIGNAL PROCESSING

zen)

y(n)

FIGURE 6-3
Implementation

of

second-order filter as a
Direct
Form
structure.

These latter two systems' are implemented as two Direct Form I filters and the
common delayed values of {u(n)} are combined to eliminate one set of delays, as
shown in Fig. 6.3. The result is the Direct Form II structure implementation of
the system function H(z).

The Direct Forms I and II are commonly employed to implement digital

filters, both in hardware and in programs. However, they have two shortcomings. First, they lack hardware flexibility, in that filters of different orders,
having a different number of multipliers and delay elements, require different
hardware structures. Second, the sensitivity of the coefficients to quantization
effects that occur when using finite-precision arithmetic, described in Chapter
10, increases with the order of the filter. This sensitivity is especially important
when generating poles, since small deviations in the coefficient values may
significantly change the form of the frequency response, or even cause the filter
to become unstable, as shown in Chapter 10. To overcome these effects, we
will next consider implementing a digital filter by combining filter building
blocks consisting of second-order sections, either in cascade or in parallel. The
cascade form has the advantage that the overall filter system function is easier
to determine. Being the product of the components, the overall system
function has the same zeros and poles as the individual components. The
parallel implementation does not have this property with its zeros, but does
provide an easier control of the dynamic range of the signal values. In the
cascade form, intermediate results can exceed the range of the number system
in some cases. The parallel implementation also allows the filter operation to
be performed in parallel. This feature is important for high-speed filtering
applications, in which a separate computer can be used to implement each
parallel structure and the processing can be performed simultaneously.

6.5 CASCADE COMBINADON OF

SECONDORDER SECI10NS
To obtain a cascade implementation, the system function H(z) is factored into
a product of second-order terms in both the numerator and denominator, after
the noncausal part is factored out. If M, the number of delays in the

DIGITAL P1LTER S'I1lUClUU!S

denominator, is greater than Np

write

+ Hp,

2lS

the number in the numerator, we c::an

[Nil)

H(z) =Az*

n Hk(z)

(6.13)

k-t

where
(6.14)
The constant A is a gain term, ZNp is the noncausal advance component and
[M 12] is the integer obtained by rounding up M 12 to the next integer value.
For example, if M=5, then [MI2] =3. H IVp+Hp>M, then the upper limit
would be [(Np + Np )/2].
For the filter coefficients to be real, two complex-conjugate zeros of the
system function H(z) are combined to form each numerator term and two
complex-conjugate poles are combined for each denominator term. Two
real-valued singularities can also be combined to form a second-order term as
wen. H the number of poles or zeros is odd, the remaining single singularity is
implemented by making b2lc =0 or a2lc =0 for one of the sections. Each Hk(z)
term describes a coscade second-order section (CSOS), that c::an be efficiently
implemented using the Direct Form n, as shown in Fig. 6.4. The defining
equations for the kth CSOS are given by

Yk(n) =alkYk(n -1) +a2lcYk(n - 2) +xk(n) + bllcxk(n -1) + b2lcxk(n -2)

(6.15)
in the time domain, and in the z domain by

Yk(z) = alkz - 1 Yk(z) + a2lcz-2 Yk(z) + Xk(z) + blkz- 1 Xk(z) + b2lc z- 2Xk(z)
(6.16)
Configuring a digital filter in this cascade form allows us to put an the

IIGURE 6-4
Cascade second-order section (CSOS) structure.

216

IN'IROOUCJ10N TO DIGITAL SIGNAL PROCESSING

design effort into an efficient implementation of a single second-order section

in hardware. Then, to implement any desired filter, the required number of
these identical sections are simply cascaded. The multiplier coefficient values
for a particular filter design can be stored and accessed from a read-only
memory (ROM). Modifications in the filter design can then easily be
accomplished by changing the contents of this ROM.
EumpIe 6.2. IJaplementiDc 8 filter witIa CSOS compoaelds. Let
H(z) =z/3 + 5/12 + 5z- 1/12 + z-2/12
1- z-I/2 + z-2/4
This system function can be factored into

The filter can be implemented in two ways, depending on which numerator term
is combined with the denominator to form a CSOS. The first is

1
H(z) = 3" z(1

1 + -1 +
z-I/4) 1- Z-!/2

+;

-2

2/4

which is shown in Fig. 6.5(a). The second is

_
1 +z-I/4
H(z)=lz(1+z l+z-~1_z 1/2+% 2/4
which is shown in Fig. 6.S(b).
In theory, there is no particular reason to prefer one structure over the
other. In practice, however, the ordering is important when the limited dynamic
range or the limited precision of the number system of the hardware is
considered. These topics are discussed in Chapter 10.

BARDWARE SAVINGS wrru IDGH-8PEED CIRCUITS. Implementing a filter

with CSOS components can also lead to savings in the amount of hardware
required. Four multiplications and additions are performed to execute the
general CSOS operation. Let the time for this procedure take Tp seconds. With
current special-purpose signal processing chips, in which multiplication times
are 200 nanoseconds, typical values for Tp are approximately 1 microsecond. If
a new sample arrives every 18 seconds, real-time operation, in which the
required processing is completed before the arrival of the next sample, occurs
when Tp < 18. To implement a high-order filter, we could use a large number
of identical CSOS' hardware sections.
If Tp is much smaller than 18, the hardware performing the processing is
very fast and is sitting idle most of the time. In this case, an alternate

DIGITAL FILTER STRUCIURES

217

(/I)

Y(II)

y(1I)

ftGURE 6-S
Two implementations of a digital filter using CSOS structures.

implementation is to have only one CSOS hardware structure that is used

repeatedly. H Tp 1S. the input data can be processed by the structure that
has coefficients (obtained from a ROM) corresponding to the first section of
the system function. The output of this section and the contents of the delays
are then stored in a read/write memory (RAM). The output of this section
then becomes the input for the next section of the system function. This second
section is implemented by the same hardware esos with the appropriate
coefficients read from ROM and whose delay contents are read from RAM.
This procedure is repeated until the required number of sections have been
employed to implement the filter. The process is again initiated upon the
arrival of the next sample.
EDmpIe 6.3. Processiag telepbone-quaHty speech in real time. Telephonequality speech is defined to have frequency components between 300 and
3500 Hz. The sampling rate commonly employed is approximately 8 kHz, or
Ts = 125lls. If the filter is implemented ll~ a cascade of [MI2] CSOS-sections,
then for real-time operation [MI2] must be less than Ts/Tp. When Tp < IllS, Ts/Tp
is greater than 100. Typical values for the filter order M rarely exceed 12. Hence,
real-time operation with a single CSOS structure is readily achieved.
When TsITp [MI2], as above, then not only can the same CSOS
hardware structure be used to implement the required filt~r, but it is fast
enough to process the data on additional voice channels. In the case above, as
many as 16 voice channels can be ac;eommodated with a single CSOS structure.

Z18

JNTRODUCl1ON TO DlGrrAL SIGNAL PROCl!SSlNd

6.6 PARALLEL COMBINADON OF

SECOND-ORDER SECI10NS
An alternate configuration that is useful for implementing digital filters is the
combination of parallel second-order sections (PSOS). This structure will be
useful for the implementation of the frequency-sampling FIR filter structure,
described later in this chapter. If M #ir-Np + NF , then H(z) can be factored into
[Mil)

H(z) =zNp ~ Gk(z)

(6.17)

k-l

where

_
gOA: + guz - 1
Gk (z ) -1
-1
-auz

-alkZ

-2

and [.] indicates the next largest integer value. This filter configuration is
shown in FIg. 6.6. For Np + Np > M, the upper limit of the sum becomes
[(Np + NF)/2]. The defining equations for the kth PSOS filter are given by
Yk(n)

=aUYk(n -1) +alkYk(n -

2) + gOA:Xk(n) + glkxk(n -1)

(6.18)

in the time domain, and in the z domain by

Yk(z)

=alkz-1 Yk(z) + alkz-2 Yk(z) + gOA: Xk(z) + guz- 1Xk(z)

(6.19)

Since the input sequence {x(n)} is applied simultaneously to all the PSOS
sections, a scale factor is necessary for each section. This scale factor is
included in the numerator eoemcients gOA: and glk' The poles in the PSOS
system function Eq. (6.17) are the same as those of the CSOS given in Eq.
(6.14). However, there is only one zero in each PSOS, as opposed to a pair of
complex-conjugate poles in the CSOS. This occurs because additional zeros
are generated in the total system function in making the parallel connections,
as shown in the following example. In general, determining the locations of
the zeros of a parallel combination of filters is not a simple problem.

It(lI)

I1GUIlE 6-6
Parallel second..order section (PSOS) structure.

DIGITAL FILlER STRUCTURES

219

Combining over a common denominator, we have

2-(a+b)z-1
H2(z) = (1- az-l)(I- bz 1)
H2(z) contains the same poles as the individual terms of H1(z), but there is a zero
at z =(a + b)/2 that is explicitly expressed in H 2(z). Of course, this zero is also
present in H 1(z), but is not obvious by inspection.
EuIIIpIe 6.5. Geaeratioa of COaaplell-coaJ...te zeros. Let us consider the
implementation of the system function given in Example 6.2 as a combination cif
PSOS structures. Let
H( ) = z/3 + 5/12 + 5z- 1/12 + z-2/12

1- z 1/2 + z- 2 /4

This can be factored into the following

H( ) = z(1/3 + 5z- 1/12+ 5z- 2/12 + z- 3 /12)

z
l-z 1/2+z-7./4
To obtain the form of H(z) given in Eq. (6.17), we divide the denominator into
the numerator, to get

7 Z-1
7/3 + 5z- 1/4 ]
H(z)=z [ 3+T+l-z- 1/2+z- 2 /4 .
The PSOS implementation is shown in Fig. 6.7.

The above cascade and parallel forms can be employed to implement

both nonrecursive and recursive filters. For filters having finite-duration
unit-sample responses, alternate structures, described in the next section, also
have some advantages .

.t(n)

nGURE 6-7
Digital filter implementation using PSOS structures.

y(n)

220

INI'RODucnON TO DIGITAL SIGNAL PROCESSING

6.7 LINEAR-PHASE FIR FILTER

STRUCfURES
As shown in Chapter 3, for a digital filter to have linear phase, its unit-sample
response sequence must be either symmetric or antisymmetric about some
point in time. Since the FIR filter coefficients are identical to the elements of
the unit-sample response, pairs of multipliers have coefficients with identical
magnitudes. To save a multiplication operation, the inputs to these multipliers
can be combined and then the product computed. Doing this reduces the
number of multiplications by approximately one-half. Since performing a
multiplication is the most time-consuming operation in a' digital filter, this
reduction in computation time is significant, possibly permitting real-time
operation to occur.
To illustrate this procedure, let us consider a noncausal FIR filter having
an even unit-sample response extending over - M s, n s, M. Then its system
function is given by
M

H(z)=

,,--M

h(n)z-"

(6.20)

The Direct Form I implementation of this filter shown in Fig. 6.8(a) requires
2M + 1 multiplications. For an even unit-sample response, the equal-valued
coefficients can be combined to give

H(z) = h(O) +

L h(n)[z" + z-"]

(6.21)

This last equation illustrates that the coefficient h(n) can be applied to tile sum
of the outputs of two delay elements, as shown in Fig. 6.8(b). This requires
only M + 1 multiplications to accomplish the same result as before.
Eumple 6.6. Lineu-phase FIR stradare. Let

H(z) = z/2 + 1 + z-I/2

This linear interpolator is implemented in the direct form structure in Fig. 6.9. To
exploit the even symmetry of the unit-sample response, we can write

H(z) = 1 + [z + z-I]/2
and implement,this filter with only one multiplication, as shown.

The same procedure can be applied to a causal FIR filter. The causal
filter is first implemented as a Direct Form I structure. The contents of the
delays leading to multipliers having the same coefficient magnitudes are then
routed to summing junctions before the multiplication, as in the next example.
Example 6.7. Causal 6near-phase FIR filter structure. Let

H(z) = 1/2 + Z-1 + z-2/2

Combining the outputs of the delays that lead to multipliers having the same

DIGITAL FILTER STRUcruRES

l!21

h(M)

y(n)

(b)

x(n)

y(n)

nGURE 6-8
(a) Nonrecursive Direct Form I structure of a noncausal FIR filter; (b) efficient noncausal
linear-phase FIR structure that reduces the multiplications by approximately one-half.

coefficient magnitUdes, we can express the system function as

H(z) =

Z-l

+ [1 + z-2]/2

These equivalent structures are shown in Fig. 6.10.

6.8 FREQUENCY-SAMPLING
STRUcrURE FOR THE FIR FILTER
In this section, we consider an alternate method of implementing an FIR filter.
This structure is interesting in that it employs the cascade combination of a

221:

INTRODUCfION TO DIGITAL SIGNAL PROCESSING

x(n) o-----.---~

x(n) o---.....-~--...c

y(n)

y(n)

FIGURE 6-9
Two equivalent structures for a noncausallinear-phase FIR filter.

nonrecursive filter and a bank of recursive filters to generate a finite-duration

unit-sample response. Since the unit-sample response of the final filter has finite
duration, it is by definition an FIR filter, even though it uses a recursive
structure. To illustrate the approach, we consider first the implementation of
simple first-and second-order filter sections and then combine these to form
more complicated filters. This approach is then generalized to be applicable to
any FIR filter.
Let us consider the following FIR filter having a causal finite-duration
unit-sample response containing N elements of constant value, or
for 0:5 n :5 N - 1 .
otherwise

(6.22)

The corresponding filter system function is equal to

G

Ho(z)

=;

N-l

~o z-n

x(n) 0---.......--+---,.,

(6.23)

y(n)

FIGURE 6-10
Two equivalent structures for a causal linear-phase FIR filter.
(-

DIGITAL FILTER STRUcnJRES

223

L----------T~---------JIlL-------------l~-----------J

COMB FILTER

RESONATOR

FIGURE 6-11
Comb filter followed by a resonator to implement an FIR filter.

Using the finite geometric sum formula, this sum can be evaluated to give
Ho(z)

Go 1- Z-N

= N 1- Z-I

(6.24)

This analytic form of the system function suggests a novel way to implement
the above filter: as the two-stage cascade structure shown in Fig. 6.11. The
system function of this cascade combination can be written as
Ho(z)

= Hdz) HR,o(z)

(6.25)

where
Hdz)=

1 -z -N
N

and

Go

H R,O( z-1
) - - -I
-

-z

Hdz) is the system function of the comb filter; the reason for this
name will become apparent shortly. The comb filter has the (N + I)-point unitsample response, only two elements of which are nonzero, h(O) = I/N and
h(N) = -1/N, as shown in Fig. 6.12 for N = 8. These nonzero elements of

FIGURE 6-12
FIR filter implementation using the comb filtl> with resonator structure.

224

INIRODUcnON TO DIGITAL SIGNAL PROCESSING

FIGURE 613
Magnitude response of comb filter (N = 8).

alternate signs produce the sinusoidal transfer function given by

.
1_e-iNw
Hde IW ) = - - N

2
N

=.1 e- 1Nw12 sin(NOJ/2)

(6.26)

which is shown in Fig. 6.13 for N = 8. HdeiW ) has N zeros equally spaced over
the frequency range O:s OJ < 2;r. Because the magnitude response resembles a
comb, this filter has become known as a comb filter. This form of the transfer
function can also be explained by considering the pole/zero pattern of the
system function Hdz). This is most easily seen by expressing Hdz) as
1 N-l
.
Hdz)=(l_ el 2n"kINz -l)
N k=O

IT

(6.27)

This last form explicitly indicates that there are N equally spaced zeros over
the unit circle with the first zero at z = 1 and all the poles are at z = O. This is
shown in Fig. 6.14 for N = 8.
HR,k(Z) is the system function of a resonator, or a filter that has poles on
the unit circle at frequency OJk = 2;rk/N. Here HR,o is the resonator at OJ = O.
1m

Izl=l
~------~------~Re

ei'

FIGURE 6-14
Pole/zero pattern for comb filter for N

= 8.

DIGITAL FILTER STR\JCI1JRES

-225

The resonator HR,o(z) has a pole on the unit circle, at z = 1. Since this pole is
not strictly within the unit circle, this filter is not stable, in that its unit-sample
response does not decay to zero with time. One common application of such a
filter is to implement an oscillator.
When this resonator is cascaded with the comb filter, however, the zero
of the comb filter located at z = 1 cancels the resonator pole, making the
combination stable, since the unit-sample response of the total filter has a finite
duration.
The transfer function of this two-stage filter is equal to
U( j<u)::c: Go 1- e-j~
uoe
N l-e-lro

= GOe_i(N_l)rol2sin(Nw/2)
N

(6.28)

sin( W /2)

As shown in Fig. 6.15 for N = 8, IHo(eiOO)1 has a maximum at W = 0, equal to

Go, and is equal to zero at Wk = 2;rk/ N, for 1::s; k :5 N - 1. As a consequence,
we can set the gain of the filter at W = 0 to any desired value Go without
effecting the gains at the other Wt points, since they are zero. We will exploit
this result later by connecting additional resonators in parallel with HR,o(z), to
set the gains at these other values of Wk to arbitrary values.
The behavior of this filter combination can also be explained in the time
domain. The comb filter energizes the resonator with the nonzero sample at
time n = O. The resonator is lossless, because the feedback coefficient has a
magnitude equal to unity. Hence, it maintains the output value until the comb
filter discharges the resonator with a nonzero sample of opposite sign at time
n = N. This two-stage filter combination produces a rectangular pulse having N
samples.
Let us extend the above two-stage filter by considering & complex-valued
resonator whose pole is situated at the next zero location of the comb filter,
located at frequency WI = 2:r./N. The system function of this resonator is given

FlGURE 6-15
Magnitude response of comb filter plub
resonator (N =8).

226

INfRODUcnON TO DIGITAL SIGNAL PROCESSING

by
(6.29)

where G1 is now a complex-valued gain. This constant is set to the complex

value of the desired transfer function H(e i "') at w = 2;r/N, or
G1 = H(e i21rIN)

(6.30)

To obtain real-valued coefficients in the filter, we need a second

complex-valued resonator, whose pole is the complex conjugate of that of
HR.1(Z). This pole also cancels the (N -1)th zero of the comb filter. The
system function of this second resonator is given by
GN-I

HR,N-l ( Z )

= 1 -e -J'7J<IN-I
z

(6.31)

where GN - 1 is complex-valued gain equal to the desired H(ei"') at w =

21c(N -1)/N = -21c/N. In Chapter 3, the transfer function of a filter having a
real-valued unit-sample response was shown to have the property that
(6.32)

Hence, ~-l = G;. Connecting these two resonators in parallel, the resulting
system function is given by
HR.1(Z)

G1
GN - 1
+ HR,N-l(Z) = 1 - ei21rINZ -I + 1 -e-j21rIN-I
Z
G1+ Gi - [G1e- j21rIN + Gi ei21r/,z-l
1- 2 cos(21c/N) Z-l + Z-2
_ 2 Re[G11- 2 Re[Gl e- j21r/,z-l
- 1- 2 cos(21c/N) Z 1 + Z-2

This combination forms a second-order resonator, whose coefficients are all

real-valued.
This procedure can be generalized to include a complex-valued resonator
at each of the N zero locations of the comb filter. These resonators are all
connected in parallel and this parallel combination is then connected in
cascade with the comb filter to produce the total filter H-r(z), given by
(6.34)

Example 6.8 illustrates the application of this technique.

Example 6.8. Comb and resonator strudare for an F1R filter. Let us consider
implementing the causal linear interpolator with the comb and resonator
structure. The unit-sample response is given by

h(O) =!

h(l) = 1

h(2) = ~

hen) = 0,

otherwise

DIGITAL FILTER STRUCTURES

227

The transfer function is given by

H(e1j = (1 + cos(O e-l "

Since the number of elements in the unit-sample response is equal to 3, we
choose N = 3. The comb filter system function is then

= (1- z-3)/3
Hc{z) has three zeros, located at z = ei:brk/3, for k =0, 1 and 2. The zero at z = 1
Hc(z)

will be canceled by a real resonator, and the zeros at z = e:tj:br/3 will be canceled
by a pair of complex resonators. The gains of the resonators are equal to

Go = H(eiD) = 2
G1 = H(~/3) = l e-I:br/3

G2 = H(eJ4'</3) =
Note that Go is real and ~
function is given by

=Gr.

! ~/3

The real-valued first-order resonator system

2
HRo(z)=-.
1- Z-1
The system function of the parallel combination of the complex-valued first-order
resonators is

l e-j:br/3
Hll,l(Z) + H...,2(z) = 1 -c.1:br/3Z

! ~/3

+ 1 -e-j:br/3z -1

2 ReB e-j:br/3] - 2 Re[i e-jU/3]z-l

1-2cos(2Jr/3)z-1+Z-l
cos(2Jr/3) - cos(4n/3) Z-1

= 1-2cos(2Jr/3)z-1+Z-Z
-i(1-z- 1)
= 1 +Z-I+ Z -2

The coefticients of the resUlting second-order resonator are real. The filter
structure is shown in Fig. 6.16. The unit-sample response of the two resonators
are shown in the figure. Their sum is equal to the desired unit-sample response.

This filter structure will be useful for the frequency-sampJing procedure

for FIR synthesis described in Chapter 9.

6.9 SUMMARY
In this chapter, we have described several common methods to implement a
digital filter structure from either its difference equation, its unit-sample
response or its z-transform. We started by defining the terms commonly used
for describing digital filters and then considering the various structures. The
structure determined directly from either the difference equation or the system
function is called the Direct Form I. An alternate view of the same equations
results in the more memory-efficient Direct Form II structure. Digital filters

INI'R.ODUcnON TO DIGITAL SIGNAL PROCESSING

.TIT.

x(n)

o--iH~~

FIGURE 6-16
Implementation of FIR digital filter using the comb filter with a bank of resonatoD.

structured as casCade and parallel combinations of second-order components

have advantages in terms of hardware implementation and numerical sensitivity. The above methods can be employed to implement the general filter
structure, which contains both poles and zeros. FIR filters having a linearphase response were shown to have a special structure that reduces the number
of multiplications by approximately one-half; The frequency-sampling structure is shown to be an alternative to the tapped-delay line structure for FIR
filters.

PROBLEMS
6.1. Implement the following system functions in the Direct Form I, Direct Form II,
cascade and parallel structures. All coefficients are real.
1
(a) H(z) =(1 + az-I)(I- bz I)

(use first-order sections)

1
(b) H(z) =(1- az-I)3

()
C

H()
z

I-bcos(8)z-1
(1- az- I )(I- 2b cos( 8)' z 1+ b2Z-~

1
1
(d) H(z) =(1- az- I)2 + (1- bZ-I)2
6.2. Implement the following system functions as Direct Form I, Direct Form II,
cascade, parallel and frequency-sampling structures. All the coefficients are real.

DIGITAL Fn.TER STRUCIURES

(b) H(z)=1_2rcos(O)z-2+rz-4

1-z-2
(c) H(z)=(l-2rcos(O)z l+rz-~2
6.3. Implement the following difference equations in structures indicated by Direct
Form I, Direct Form II, cascade of second-order sections and parallel of
second-order sections. Assume all the coefficients in the difference equations and
filter are real.
(a) (See Example 7.5)

A(l + z-I)(l + z-2)(1 + Z-1 + z-~

H(z) =(1- 0.7z 1)(1-1.2z-1 + 0.8z- 2)(1-1.2z- 1+ 0.6z-~(1- 0.9z-1+ O.4z-~
(b) (See Example 7.6)

A(l- z-2)(1 + 1.8z- 1+ z-~(l- 1.8z- 1+ Z-2)

(z) =(1 + 0.lz- 1+ 0.5z-~(1- 0.6z- 1+ 0.8z- 2)(1 + 0.7z-1+ 0.8z- 2)

(c) (See Example 8.7)

A

H(z) =(1-1.7z-1 + 0.7z-2)(1-1.7z-1 +0.9z-~

(d) (See Example 8.8)

1 +Z-1
1 +2z- 1+z-2
H(z) =A 1-0.8z-11-1.6z 1+0.8z- 2
(e) (See Example 8.11)
1-z~

1-z~

1-z~

H(z) =A--..........,..---=---...,.----=-:--":""":'--:---:--:---::;
1- 0.4z- 1 + 0:9z-2 1 +0.lz- 1+0.7z- 21 +0.6z 1+ 0.9z-2
(f) (See Example 8.12)

Hz =A 1-0.lz-1+z-2 1-0.lz- 1+z-2 1"-0.lz-1+z-2

()
1+0.lz-1+O.2z 21-0.4z 1+0.9z-21+0.6z 1+0.9z-2
6.4. Find the unit-sample response at the junction of the comb/first-order resonator
(pole at z = 1) combination, when they are connected in the reverse order.

CHAPTER

7
FROM
ANALYSIS
TO SYNTHESIS

7.1 INTRODUcnON
In this chapter we extend the analytic techniques discussed in the previous
chapters to be useful for the synthesis of digital filters. We then begin the filter
design section of the book by considering some general issues that should be
addressed at the initiation of a filter design. An alternative to designing the
filter analytically is the intuitive approach of pole/zero placement considered in
this chapter. Intuition dictates that the dips and peaks in the desired magnitude
response can be accomplished by suitable placement of zeros and poles in the z
plane. However, even though a filter can be designed in this manner, there is
no guarantee that another filter, having fewer delays or operations, may not
alSo satisfy the same requirement. This intuitive approach will form the
starting point for the optimal filter synthesis procedures described in the
following chapters. This intuitive and interactive procedure provides a fundamental understanding of the effects produced by poles and zeros on the
frequency transfer function.
An additional benefit of this interactive procedure is that it solves two
problems that exist With the conventional design procedures, described in the
next chapter. The first is that the filtet magnitude response resulting from
conventional design approaches can often be improved by proper placement of

FROM ANALYSIS TO SYNTHI!SIS

231

additional poles and zeros. The intuitive procedure helps to place such
additional singularities. The second problem occurs when the filter poles lie
very close to the unit circle. Filter design computations require a high precision
in the calculations. If the precision is not sufficient, the locations of the
singularities deviate slightly from those desired and can significantly alter the
magnitude response. The intuitive design procedure offers an alternative. It
indicates how to shift the resulting singularities in the z plane to satisfy the
specification without repeating the tedious computations.

7.2 INTERRELADONSIUPS OF
ANALYTIC METHODS
As an overview, let us consider the flow of information given in Fig. 7.1, in
which the five main categories of analytic techniques are shown interrelated by
the solid arrows. In principle, the results presented in the previous chapters
should allow us to start at any category, the starting point being the
specification to be met by the digital filter, and to determine any other. The
end point for synthesis is usually the digital filter structure and coefficient
values. As the figure implies, the easiest path is usually through the
z-transform. The numbers in the figure indicate examples in the book that
illustrate the corresponding path.
The important practical case in which only the desired magnitude
response is specified is shown connected by a broken arrow to the z plane and
to the complex frequency transfer function. In general, finding the z plane

(h(II)}

5.1- 5.4

FILTER
STRUCfURE

5.36 -45

5.5 - 5.8

7.1-2
H(e"")

H(z)

5.28 - 29

5.10,5.12
5.23- 27
7.3

5.13 - 5.18

POUYZERO
PATTERN
IN

1+ - -7.5.7.6
---:-

111.5
1

~
I
I
(JO>

He)

PLANE

FIGURE 7-1
Relationship of analytic methods. Numbers refer to Examples in the book that illustrate each path.

232

INTRODucnON TO DIGrrAL SIGNAL PROCESSING

pole/zero Configuration from only the magnitude response, with no phase

specification given, is not a well-defined analytical problem. The results of this
chapter will demonstrate that it does not have a unique solution; that is, many
filters can have the same magnitude response, but have different phase
responses. In this chapter, poles and zeros will be placed in the z plane
interactively to achieve a sufficiently accurate approximation to IH(ei"')I. Han
appropriate phase response constraint, called the minimum-phase condition, is
applied to the magnitude of the DFT sequence {IH(k)l}, then a unique phase
response sequence {Arg[H(k)]} can be determined. From these two, the
complex - DFf sequence {H(k)} can be found. The explanation of this
procedure is delayed until Chapter 11, since it involves the discrete Hilbert
transform.
Let us now consider the paths indicated by the solid arrows in Fig. 7.1.
The two paths from the Fourier transform to the z-transform and from the
pole/zero pattern to the z-transform have not been explicitly described
previously. Since these are important for filter synthesis, they are considered
now.
FROM FOURIER TRANSFORM TO z-TRANSF.ORM. The Fourier transform is

usually given in terms of the magnitude and phase response functions, which
cart be expressed as

ane

IH(ej "')12 = ~(ej"') + Hl(ei "')

=H(ei"')H*(eiu,)

(7.1)
(7.2)

Any common factor in H1(ei"') and HR(ei"'), such as the filter gain, has usually
been eliminated in the argument of the arctangent function through cancellation. The following procedure can then be employed to retrieve this common
factor.
OJ
Let N(ei ) and D(ei "') represent the numerator and denominator
polynoIoials in Arg[H(ei"')] that have any common factors removed, or
Arg[H(e iOJ)) = arctan[N(eiOJ)/D(ei"')
(7.3)
Then the imaginary and real parts of H(ei "') are given by
HI(ei "') = aero) N(ei"'),
HR(eiOJ) = a(ro)D(ei"')
where aero) is the real-valued common factor, that may be a function of ro.
We can find a( ro) from the specified magnitude response by substituting these
forms of the real and imaginary parts into Eq. (7.1), to find

IH(ei "') 12
+ D2(ei

a2(ro) = N2(eiw)

W)

(7.4)

the value of aero), the sq,""e root must be taken, allowing both

FROM ANALYSIS TO SYNlHESIS

Phase=O @ro=O

233

Phase = in @ro=O
1m

1m

H,=O

----~~----~-Re

---~-----Re

H.<O

Phase =

Phase =J @ro = O'

1m

-! @Ol = O
1m

----t------Re
H,<O

FIGURE 7-%
Possible values of phase at co = 0 or co = 0+ and the corresponding relationship of the real and
imaginary parts of the Fourier transform.

a(co) as equally valid solutions. The correct sign must be determined from
the value of the phase at co = 0 or co = 0+, as shown'in Fig. 7.2. If the phase is
0, then H I(eiO) = 0 and the sign of a( co) should be taken to make H R( ei~ > O.
If the phase is n, then HI(e iO) = 0 and the sign should be chosen to make
HR(eiO) < O. If the phase at co = 0+ i~ n/2, HR(ei~ =:! 0 and the sign should be
chosen to make HI(ei~ positive and if the phase is - n /2, H I(ejO) must be
negative.
The complex-valued Fourier transform is then equal to
H(ei"') = a(co)D(ei "') + ja(co)N(ei"')

(7.5)

where a( co) contains the correct sign. The Euler identiti~s can then be applied
to convert the sine and cosine functions in the analytic expression of H(ei"')
into complex exponentials. The z-transform is then obtained by replacing
a e-iOJll with aZ-n.
EumpIe 7.1. From Fourier traDsform to z-tnosform. Let
4

jOl 2
IH(e )1 -1z,a cos(co) + a 2

with

lal < 1

234

lNTRODUcnON TO DIGITAL SIGNAL PROCESSING

and

. = arctan [ -a sine w)
Arg[H(e''')]
( )]
l-acosw

. I De'''')
(. ]
= arctan[
N(e'''')

with Arg[H(eiO)] =:It.

Computing the squared magnitude with N(eiw) and D(eiW), we get
IH(eiw)12 = a2(w)[N2(eiw ) + D2(ei "')]
= a2(w)[(a sin(W2 + (1- a cos(w)Y]

= a2( w )[1 - 2a cos( w) + a2(cos2( w) + sin2( w))]

= a2( w)[1 - 2a cos( w) + a2j

This result must equal the specified magnitude response. Equating the two and
taking the square root, we get

a(w) =

1 - 2a cos( w) + a 2

The real part is then equal to a( w )D(ei "'), or

HR(e iw) = 2(1- a cos(w

1- 2a cos(w) + a2
Since the phase is equal to :rc at w = 0, HR(e~ must be negative (note that
=0). Hence, the negative square root is the correct choice, and

HI(~

H(ei

W
)

= a(w)D(eiw)

+ ja(w)N(ei "')

_ -2(1- a cos(w) - ja sin(w

1 - 2a cos( w) + a2

-2(I-a ei"')

-2
l-ae

i'"

Finally, the desired system function is given by

H(z) =1

-2
-az -1

Occasionally, the evaluation of Eq. (7.4) results in a complicated

trigonometric expression. An alternate procedure to determine a(c.o) follows
from expressing the Fourier transform in the following form:

H(ei"') = a(c.o)D(eim ) + ja(c.o)N(ei m)

=a(c.o)[D(ei"') + jN(ei m)]

(7.6)

The term in the square brackets can then be simplified by converting the
trigonometric terms into complex exponentials by applying the Euler identities. Then the squared magnitude of the riglit hand side of Eq. (7.6) can be
computed and compared with the specified IH(eimW to find a(c.o).

FROM ANA(YSIS TO SYNTHESIS

235

Example 7.2. Alternate approach to solving Example 7.1. From the numerator
and denominator terms of the phase response in Example 7.1, we have
H(ej ",) = (O)[l-a cos(o) - ja sin(O)]
= (O)[l-a

ei"']

Then the squared-magnitude response should have the following fonn:

IH(ej.,W = ~(O )[1 - 20 cos( (0) + a9

This result must equal the specified function given by

.
2
4
IH(e1jl - --~---:-~2
-1-20 cos(O) +a

Equating the two and taking the square root, we find

2
(0) = --~---:-~2
1- 20 cos(O) +a

which is equal to the result in the previous example.

FROM POLE/ZERO PAtTERN IN THE % PLANE TO %-TRANSFORM. In

Chapter 5, the system function of adigital filter was given by

H(z}=AzNp ~Np
11 (1-C/cZ-l)
/C-l

1MIT

(1-d/cz-l)

(7.7)

/c-l

where A is the gain constant. Since the pole/zero pattern is not effected by a
constant gain factor in the system function, the value of A must be specified
separately to determine a unique system from the pole/zero pattern.
To find the system function from the pole/zero pattern, the singularities
in the z plane should be considered one by one. When the singularities at z = 0
and z = 00 are included, the number of poles is equal to the number of zeros.
Let us first consider the singularities in the z plane that are not located at z == 0
and z = 00. A zero located at z = Cb for C/c =1= 0 or 00, produces a numerator
term (z - c/c) in the system function. Each such zero should be combined with
a pole at z = 0, to produce the term (z - c/c}/z, or (1- C/cZ-l), for one of the
terms in the numerator of Eq. (7.7). If there is no pole at z = 0, then a
pole/zero pair should be created there and the pole paired with the zero at
z = CIe, leaving the zero at z = O. A pole located at z = d/c, for d/c =1= 0 or 00,
produces a term 1/(z - die}' Each such pole should be combined with a zero at
z = 0, to produce the term z/(z - d/c} = 1/(1- d/cz-l} in Eq. (7.7). If there is
no zero at z = 0, then a pole/zero pair should be created there and the zero
paired with the pole at z = die, leaving the pole at z = O.
After all the singularities lying in 0 < Iz 1< 00 have beeQ eliminated, an
equal number of singularities of opposite types will remain at z = 0 and z = 00.
If there are Mp poles at z = 0, each of them produces a Z-1 term, or a z-Mr
term in the numerator of Eq. (7.7). Hence, Np = -Mp, indicating that the

236

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

1m
Z= OCJ

jO.9

~~------~------~~Re

nGURE 73
Pole/zero pattern to be implemented as a
digital filter.

pole/zero pattern is that of a causal filter whose first nonzero term in the
unit-sample response is h(Mp). If there are Mz zeros at z = 0, each of them
produces a z term, or a ZMz term in the numerator of Eq. (7.7). Hence,
NF = M z , indicating a noncausal filter whose unit-sample response has its first
nonzero term as h(-Mz ).
Example 7.3. From pole/zero pattem to z-traasform. Let us consider the z
plane pole/zero pattern shown in Fig. 7.3. Applying the procedure described
above first to the zeros in 0 < Izi < 00, then to the poles in 0 < Izl < 00, and finally
to the singularities at z = 0 and z = 00, we have
H(z) = z(1 + O.9Z"1)(1- O.9Z"1)
(1 + jO.9Z"1)(1- jO.9Z"1)
= -3 z(1- O.81z"2)
(1 + O.81z"2)

Note that the gain A = -3 must be specified separately.

7.3 PRACfICAL MAGNITUDE

RESPONSE SPECIFICATIONS
An important situation commonly arising in practice is one in which only the
desired magnitude response IH(ei"')1 is specified and the phase response is left
unspecified or arbitrary. Let us consider the ideal magnitude response
specification of the lowpass filter characteristic having the cutoff frequency Wc
shown in Fig. 7.4(a). In the ideal filter, frequencies smaller than Wc are
transmitted with no attenuation and frequencies greater than Wc are completely attenuated. We will see later that a characteristic having perfectly flat
transmission and attenuation regions and a sharp discontinuity between the
two cannot be realized in practice, but only approximated. This approximation
procedure is the art of filter design, and"'the goal is to become a good artist.
In practice, the desired IH(ei"')1 is usually specified to lie within allowable
regions. An example of an approxiUlation to an ideal lowpass filter is shown in
I

FROM ANALYSIS TO SYNTHESIS

237

(a)

\H(eJ")\

IT

It)

STOPBAND

PASSBAND

(b)

\H(eJOl)\

VV'flf----------- --..,

It)

__r.__~I~\--~.--~I\~--~.r--J

PASSBAND

TRANSmON
BAND

STOPBAND

FIGURE 7-4
(a) ldeallowpass filter magnitude
response; (b) practical specification to be satisfied by a digital

filter_

Fig. 7 .4(b). The cutoff frequency We is replaced by a passband frequency lOp

and a stopband frequency lOS' The passband is defined as the frequency range
in which the transfer function magnitude is approximately equal to unity, or
1- :s IH(eifll)l:S 1, for some small > O. The stopband is the range in which
the transfer function magnitude is small, or IH{ei"')I:s 6, for some small 6> O.
The location of We is conventionally taken as the frequency at which the
magnitude IH(eifll)1 is equal to 1/V2. The power transfer response IH(eifllW is
then equal to ! at we. Hence, it is also referred to as the half -power frequency.

Z38

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

7.4 LOGMAGNITUDE RESPONSE

CURVES
.
When the magnitude response is plotted on a linear scale, the features in the
stopband are difficult to observe when fJ is small. This attenuation region can
be expanded by plotting the magnitude response in logarithmic units, decibels
(dB), defined by
IH(eiOJ)ldB = 2010g1O IH(ejOJ)1 = 10 10g10 IH(ejOJ)1 2
(7.8)
An example is shown in Fig. 7.5, in which the magnitude responses of two
lowpass filters are compared. The first filter system function has a single pole at
z = 0.95 and the second has a double pole. The difference is more apparent in
the log-magnitude response curves.
It is conventional to normalize the log-magnitude response by vertically
shifting it such that the maximum value of 20 10g1O IH(eiro)1 over 0 s co S.1f is
1m

1m

,H(e''''> ,
20

,H(e'''> ,

400

L-=:::::======:::::I

IT

co

'H(e'' >'.B

oI\.
-20

"-'=

oLO...::::..----------IIT co

-20

-40

-40

-60

-60
I

-80 0

co

'H(e">'.B

!\
\

'",

--I

co

IT

FIGURE 7-S
Comparison of magnitude and log-magnitude response curves for first- and second-order recursive
filters.

FROM ANALYSIS TO SYNlHESIS

239

equal to 0 dB. The curve then shows the degree of attenuation and is simple to
interpret. When the gain A of the filter is included in the response, it produces
only a vertical shift without otherwise changing the shape of the curve.
A problem in the evaluation of Eq. (7.8) arises when IH(eiQ)1 = 0, which
results in an infinite logarithmic value and an execution error in the computer
program. This can occur when a zero is located on the unit circle in the z
plane. This problem can be avoided by checking for small values of IH(eiQ)1
and limiting them to be no less than some suitably small positive number. A
value equal to 10- 4 times the maximum value of IH(eiQ)1 over O:s (i):S 1r is
usually adequate. This procedure prevents the argument of the logarithm
function from being zero without significantly affecting the response of most
practical filters. This value of the constant allows the magnitude response to be
plotted over an 80 dB range, from 0 dB to -80 dB, which is adequate for the
filter designs in this book.
To show clearly both the passband and stopband characteristics, the
log-magnitude response must usually be plotted on two different scales: an
expanded dB scale to show the response in the passband, where the
attenuation is small-typically between 0 and -4 dB-and a compressed dB
scale to show the response in the stopband, where the attenuation is
high-typically IH(eiQ)ldB < -60 dB.
The practical design utility of the log-magnitude response comes in
cascading filter sections. The total filter log-magnitude response IH-r(eiOldB is
then equal to the sum of the log-magnitudes of the individual sections
IH/(eiQ)ldB. If there are N sections, then the overall filter response is given by
IHT(eiQ)ldB =

2: IHj(eiQ)ldB

(7.9)

Functions that have relatively simple forms, like those produced by single and
complex-conjugate singularities, are more readily visualized when added than
when they are multiplied. This observation will prove to be useful for the
intuitive filter design procedure presented later in this chapter.

7.5 PROGRAMS TO COMPUTE TIlE

LOGMAGNITUDE RESPONSE OF POLES
AND ZEROS
As shown in Chapter 5, the magnitude response of a first- or second-order
filter section is a function of the locations of its singularities in the z plane.
Viewing the magnitude response in this manner helps to perform efficiently the
interactive procedure described below. In this section, we describe programs to
compute and display the log-magnitude response. The first routine is
LOGMAG that computes the log-magnitude response from the magnitude
sequence generated by MAGPHS. To display the log-magnitude response, we
describe the routine PLOTLM, which is given in Appendix A.

240

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

In Chapter 5, the coefficients of the filter section producing real and

complex-conjugate singularities were determined with the subprograms ZERO
and POLE. These coefficients are then used as inputs to NONREC, if ZERO
is called, and to REe, if POLE is called. The unit-sample sequence is then
used as the input to the filter to determine its unit-sample response. The FFf
subprogram computes the discrete Fourier transform of the unit-sample
response and MAGPHS calculates the magnitude and phase responses. What
is still needed is a routine to compute the magnitude response in units of dB.
This subprogram is called

LOGMAG (HMAG, 1m t HDB )

where
HMAG
NH

lIDB

(input real array) is the magnitude response array

(integer input) is the number of spectral values over 0 S W S 1C
(typically, 65 when a 128-point FFr is computed)
(output real array) is the log-magnitude array in units of dB.

This routine performs the transformation given by

HDB(N) = 20 10810 [HMAG(N)]

for IsNsNH

(7.10)

This equation is invalid when HMAG(N) = 0 which can occur when a zero lies
on the unit circle at WN = 1C(N - 1)/(NH -1). Hence, values of HMAG(N)
less than some small number E should be set to E. The value E = 10-6 (equal to
-120 dB) is adequately small.
This output of LOGMAG should be plotted twice, once in expanded
vertical scale for the passband specification, and a second time in compressed
scale for the stopband specification. This plot is performed with the subprogram given in Appendix A named

PLOTLM(HTOT,DBVAL,SPEC,IFl,IF2,NPRNT)
where
IITOT

(input real array) is the log-magnitude response sequence

containing 65 elements to be plotted
DBYAL
(input real) is the lower limit in dB of the plot (typical values
are -4 for passband and -60 for stopband)
SPEC
(input real) is the specification value in dB that is marked as a
horizontal line on the plot (typical values are -1 for passband
and -50 for stopband),
IFl,IF2
(input integers) are the frequency points that are marked by
vertical lines on the plot (1 s IFl,IF2 s 65, and is equal to
frequency W = 2n;(k -1)/128, where k is either IFI or IF2. H
IFI = 0 or IF2 = 0, no mark is made), and
NPRNT
(input integer) is set to 1 if a printout of the terminal display is
desired.
The PL01LM subroutine automatically shifts the log-magnitude re-

FROM ANALYSIS TO SYNTHESIS

a-

UI'8 ...

1....

I:~;!!ft
I

pole at ae

L ... M

_
_ l _. _ _ _ _ _ _ _ _ _ _ _
I _ _ _-.

-I."

II
.....Ij

..

MrO,t .........

tr~eAC1

L ... 00

.. _ _ _ _ _ _ _--::=-.
.

-----I

II

-lO.'I

"'j
I
I
I

.....!--_._----_.._ - - -

'S.O+

I
I
I

I
I

-10."

-15.e+

-U.'-I
II
-1.....-_.-

tIMid

"'1
~I----~--~---~I

~.~
I

I
1

-u.~

I -21 ,,_

-1....I

I
I
I

-M.'

.......................
treq\l.ac:)'

pole .. t 1- 0 .... L ".OG

-!S." I

N.N

1:~;'~I..' - - - - - - . - - - - - -

I
I

'.

I
I

I:~:!!ft
I

'.Nt L

241

"
-..

j-'-'

I
I

__

.!

_ _ _ __.
_ _ _ __

FIGURE 7~
Log-magnitude response curves produced with PLOTLM subroutine for elemental sections. a)
zeros. b) poles. DBVAL= -20, SPEC = -3, IFl=IF2=O.

sponse curve, so that the maximum value of {IITOT(N)} is plotted to

correspond to OdB.
To illustrate the effects of singularity position on the magnitude response,
the log-magnitude response curves for single and complex-conjugate singularities have been compu,ted with LOGMAG and, plotted with PLOTLM.
The results are shown in Fig. 7.6. The peaks produced by the poles and the
troughs by the zeros have a sharpness that is related to the distance of the
singularity from the unit circle. These response curves can be considered to be
those of elemental filters.
With the above procedure, the log-magnitude response is computed for
each first and second-order elemental filter section that corresponds to the real
or comp.ex-conjugate singularities, respectively. The total filter structure is
then obtained by cascading these elemental sections. The log-magnitude
response sequence of this total filter, denoted by {IITOT(N)} is then obtained
by accumulating the log-magnitude responses of these elemental sections, as
given in Eq. (7.9).
To complete the design procedure, the ga4t constant A of the filter must
be determined. This procedure is described next.
Determining the v~ue of the filter gain constant A. In Chapter 5, it was shown
that the pole/zero pattern in the z plane is not affected by the gain constant of
the system fu~ction. Conversely, wh~.n the digital filter is designed by placing

Z42

INTRODUcnON TO DIGrrAL SIGNAL PROCESSING

poles and zeros in the z plane, the gain constant A of the filter must be
computed to meet the specification. In this section, we describe how this gain
constant is determined.
For illustration, let us consider how the gain constant of th-e digital filter
is computed in order that the maximum value of its magnitude response is equal
to one. We start with a digital filter pole/zero pattern whose magnitude
response satisfies the specification to within a gain constant. For the pole/zero
pattern having zeros at z = ek, for 1:s; k :s; N, and poles at z = db for
1 :s; k :s; M, the transfer function is given by

(7.11)

=AHpz(ei"')
where H pz(ei "') is the transfer function computed directly from the poles and
zeros by assuming that the gain constant is equal to one. The gain constant A
needs to be determined, such that the maximum value of IH(ei"')1 is equal to
one. ~o find the exact value of A, we must find the frequency lI)m for which
IHpz(e'QI)1 takes on its maximum value. Then the gain factor is equal to
(7.12)
The standard way to find lI)m is to compute the derivative d IHp.. (eiQl)I/dll), find
the values of lI) for which it is equal to zero, and evaluate IHp .. (ei "')1 at these
values to find the global maximum over O:s; lI) :s; n. In practice, however, the
procedure is very tedious because of the complicated analytic form of the
typical IHp ..(ei QI) 1 and of the presence of multiple maxima and minima.
Fortunately, there are two methods to find reasonably accurate approximations to the exact gain value that are usually satisfactory in practice.
FIrst, it may be sufficient to let the filter magnitude response be equal to one at
some convenient frequency in the passband and disregard the maximum value,
which may be slightly greater than one. For examJ>le, for lowpass filters, it is
reasonable to specify that the magnitude response at lI) = 0 be equal to one.
This can be easily accomplished by setting
A = 1/IHp.. (eiO)I.
(7.13)
The gain is readily evaluated since

(7.14)

For highpass filters, the magnitude response at lI) = n can be set to one in a

FROM ANALYSIS TO SYNTIfESIS

243

similar manner. For bandpass filters the evaluation is slightly more complicated
because the terms in Eq. (7.14) are complex-valued.
The second method of determining an approximate value for A involves
computing the DFT sequence {Hpz(k)} , finding t1)e maximum value of the
magnitude response sequence {IHpz(k)l}, denoted by IHpzlmax, and setting

= 1/IHpzlmax.

(7.15)

In the previous section, the log-magnitude response sequence of the filter

{HTOT(N)} was determined by accumulating the contributions of the elemental sections. If HMAX represents the maximum value of {HTOT(N)} in dB,
then the gain to make the maximum magnitude response unity is equal to

= lO-HMAXI2O

(7.16)

Determining the gain constant value from the DFT sequence is reasonably accurate if the DFT is computed at a sufficiently dense set of points, or
such that the form of IHpzC ei 0I is readily apparent from a graph of the DFT
sequence {lHpz(k)l. Then IHpzlmax will be approximately equal to the maximum value of IHpz(eiOI. Since computing the DFT is readily accomplished on
a computer, we will use this latter approximate method for determining the
value of A in the filter designs in this book.

7.6 PHASE RESPONSE

CONSIDERATIONS
The magnitude response of a filter is a description of the attenuation of the
various frequency components in the signal that is produced by the filter. The
phase response describes how the filter modifies the time relationship between
the frequency components passing through the filter. When a filter design is
performed to satisfy a magnitude response specification, the phase response is
usually left arbitrary. In this section, we provide the motivation for specifying a
linear phase response.
Let us start with the frequency domain description of the filtering
operation given by
(7.17)
Let us further consider the simple case of the bandpass magnitude response
given by
for WISW SW2
(7.18)
forOs W <WI and W2< W s;c
Applying the superposition principle, we can separate the input signal into two
parts
{x(n)} = {xI(n) +x2(n)}
(7.19)
The corresponding spectra are then given by

X(eIO

= X 1(ei "') + X 2(ei o

(7.20)

244

INI'RODUC11ON TO DIGITAL SIONAL PROCESSING

This separation is performed such that the desired part of {x(n)}, given by
{xl(n)}, lies in the passband of the filter, while the undesired part {x2(n)} lies

outside the passband. That is, XI(eiQ is nonzero only for WI:i w:i W2, and
X 2(ei Q is zero in this range. For example, {Xt(n)} may be the signal on a
desired radio channel and {x~n)} represents the signals on the other channels.
Since only X1(eiQ passes unattenuated through the filter, the output
spectrum is given by
Y(eiQ =X1(eiCO ) e i Ars(HCel-
(7.21)
H the filter has a linear phase response in the passband, or

Arg[H(ejQ] = -kw

for Wl:i W :i W2

(7.22)

then the output spectrum is given by

= X1(ejCO) e-ikco

Y(eill

(7.23)

In Chapter 3, we showed that a linear phase response having a slope equal to

-k corresponds to a mere time delay of k samples. Hence, the output
sequence is equal to
(7.24)
{yen)} = {xt(n- k)}
and is identical to the desired part of the input sequence, except for the time
delay of k sample periods experienced by passing through the filter. That is, a
linear phase response maintains the phase relationship of the frequency
components of the input signal in the passband, as illustrated below in
Example 7.4. In the stopband, the shape of the phase response is less
important since these signal components are highly attenuated or even
eliminated from the output. However, while a linear phase response is
desirable, it is often difficult to achieve in practice when speed and cost
considerations are applied. In these cases, a relatively smooth phase response
that approximates a linear function over the passband is often satisfactory.
EuJapIe 7.4.
given by

meet

of phase distortioa. Let the desired input sequence be

{Xl{n)} = {cos(coon) + cos(2coon)}

H the filter has a transfer function with a linear phase response
H{eIOl)

then

{y{n)}

= {cos coo{n -

= e-;.t..
k + cos(2coo{n - k}

As shown in Fig. 7.7, the shape of the signal is maintained but only a delay of k
samples is experienced.
Now consider a filter having the same magnitude response, equal to 1, but
having the phase response given by
Arg[H{ejOl)] =

f -rc/4

.1.-rc

for 0 $: co $: 3coo/2,
for 3coo/2 < co $: rc

FROM ANALYSIS TO SYNTHESIS

24S

(x,(n))

...

LINEAR-PHASE OUTPUT

NON-LINEAR PHASE OUTPUT

..

FIGURE 7-7
Illustration of phase distortion.

The output is then equal to

{y(n)} = {cos(WOn -;;c/4) + Cos(2wOn -;;c)}
and is also shown in Fig. 7.7. The shape of the sequence has been altered
appreciably, even though there was no amplitude distortion. This illustrates the
effect of phase distortion.

246

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

7.7 STEPS IN PERFORMING A FILTER

DESIGN
When designing a filter to meet a given specification, three steps must usually
be taken:
1. The choice of the filter type (IIR .or FIR).
2. The determination of the filter coefficients that meet the specification.
3. The verification that the resulting design indeed satisfies the specification.
We describe the implications of the first step in the next section. In this
chapter, the second step is accomplished by an intuitive pole/zero placement.
The conventional procedures for designing IIR filters are discussed in the next
chapter, and for designing FIR filters in the following chapter. The problems
encountered in the third step are due to finite accuracy arithmetic employed
either in the calculations or in the hardware implementation. These effects are
described in Chapter 10.

7.8 CHOICE OF FILTER TYPE

To illustrate the consequences of choosing between an IIR and FIR filter
design, let us define a relevant singularity as one that contributes to the
magnitude response of the filter. From the geometric evaluation of the
magnitude response in the z plane, discussed in Chapter 5, the rell!vant
singularities are those that are located in 0 < Izl < 00. The IIR filter system
function can be expressed as a rational function in z - 1, or
HIIR(Z)=AzNp

ND~ (l-CkZ-I)/fl (l-dkz-

(7.25)

This IIR system function has Np + NF relevant zeros at z = Ck> for l::s; k ::s;
Np + NF , that can lie anywhere in the z plane, and M relevant poles at z = dk ,
for l::s; k::s; M, that must be located within the unit circle for stability. The FIR
filter system function is more restricted, in that it is expressed as only a
polynominal in
or

Z-I,

Np+NF

HFIR(Z)=AzNp

IT

(l-CkZ- I )

(7.26)

k=1

This FIR system function has Np + NF relevant zeros that can lie anywhere in
the finite z plane, and no relevant pales.
We now give some general observations regarding digital filter design.
Let us' consider a specification that can be satisfied e~ther by an IIR filter
having M poles and N zeros, where N = Np + NF in (7.25), or by an FIR filter
having K zeros, where K = Np + NF in (7.26). Then the following general
observations apply.

FROM ANALYSIS TO SYNTIlESIS

247

Obseroation 1. The IIR filter is usually the more efficiellt design in terms
of computation time and memory requirements. The IIR filter is faster because
M + N will typically be less than K to satisfy a desired specification. The
number of multiplications for an IIR filter is equal to M + N compared with K
for the general FIR filter. As shown in Chapter 6, for a linear-phase FIR filter,
a special structure can be employed that requires only (K/2) + 1 multiplications. Since K delays are required to implement an FIR filter, while only the
maximum of NF + M or NF + N p in (7.25) for an IIR filter, the IIR filter also
requires less memory.
Obseroation 2. The practical specification can approach that of the ideal
filter by making E, 6, and (os - Op) in Fig. 7.4 all approach zero. However,
the values of both K and M + N increase as the specification approaches the
ideal. This, in turn, increases the computation time and memory requirements
of the design. Hence, there is tradeoff between the accuracy of the approximation and the computation time.
Obseroation 3. FIR filters are always stable, since the poles of the FIR
system function are located at z = 0 (and at z = 00 for noncausal filters). The
design of IIR filters with poles close to the unit circle must be designed with
care to ensure that instabilities do not occur. It is not uncommon to have poles
that were intended to be close to the unit circle actually fall outside the unit
circle in the implementation, owing to round-off errors in the calculations. In
this case, the implemented filter is unstable.

These general rules are illustrated in the design methods described in this
and the following chapters. The first method is the interactive and intuitive
procedure for designing a filter described in the next section. This procedure
does not generally result in an optimal filter, or one that has the minimum
number of delays or multiplications. However, it does lead to an intuitive
understanding of the effects of poles and zeros, and provides a better
appreciation of the conventional design procedures described in the following
chapters. An additional benefit of this interactive procedure is that it solves a
problem that exists with the conventional design procedures. The filter
magnitude response resulting from conventional design approaches can be
improved by proper placement of additional poles and zeros. The intuitive
procedure helps in deciding where to place such additional singularities.

7.9 INTERACTIVE FILTER DESIGN BY

INTUmVE POLE/ZERO PLACEMENT
In this section, we determine the filter structure and its coefficients to meet a
desired magnitude response by placing poles and zeros in the z plane. The
filter design is accomplished interactively by successively adding real and
complex-conjugate singularities in the z plane and observing whether the
magnitude of the resulting filter is approaching the desired magnitude

248

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

specification. The final filter structure is then the cascade connection of the
elemental filter sections that generate these real and complex-conjugate
singularities.
To assist in the design procedure, the following rules form the basis for
pole/zero placement:
Rule 1. The magnitude response specifications will consist of passband
and stopband regions. These bands are mapped onto the unit circle, as shown
in Fig. 7.8, to assist in the placement of the singularities. If a complex
singularity is specified, then it is understood that the complex-conjugate
singularity is alsO required, The log-magnitude response of the singularities are
computed and compared to the specification at 65 points equally spaced over
0:$ W :$ Jr. For easier interpretation, the angles of the singularities with respect
to the real axis will be given in degrees, rather than radians .
. Rule 2. All the poles in the design will have a radius rp lying in the range
0.6:$ rp < 0.96. The upper limit guarantees that the unit-sample response of
each elemental section will not have significant values beyond the 128th point.
This prevents truncation errors from affecting the log-magnitude response of a
particular elemental filter section. The lower limit is meant as a guide to avoid
poles having anemic magnitude responses.
Rule 3. The filter design is performed interactively by examining the
log-magnitude response after each additional singularity is added. In some
cases, the singularity position must be modified or removed. The effect of an
existing singularity can be canceled by specifying the opposite singularity at the
same position. Of course, such pole/zero cancellation pairs are not included in
the final filter design.

The design is initiated by placing a single pole in the middle of each

passband at radius rp, and a single zero in the middle of each stopband on the
unit circle. The object of this design procedure is to satisfy the specification by
pole/zero placement, not to find the filter having the minimum number of
elements. The total number of singularities in the designs below will typically
be le!.s than 2P, including the complex conjugates. It may be possible to find
another filter, having a smaller number of elements, that also satisfies the
specification. The procedure to find these simpler filters will be described in
the next two chapters.
Based on the log-magnitude response of poles and zeros, the following
hints may be useful:
Hint 1. The magnitude response over a narrow band of frequencies can
be increased by placing a pole at the angle corresponding to the center of the
band, but inside the unit circle. The radius of the pole determines both the size
of the bandwidth and the amount of gain. These two effects, however, are
inversely related, that is, for a single pole, high gains are usually achieved only
over narrow bandwidths. For wider bandwidths, additional poles must be used.

FROM ANALYSIS TO SYNTHESIS

HIGHPASS SPECIFICATION
IH(e'''')1

LOWPASS SPECIFICATION

IH(el"') I
1-

!~~?'llllJ

1-

1m

1m

Izl=1

~--------~--------~Re

BANDPASS SPECIFICATION

BANDREJECf SPECIFICATION
IH(e!"')1

IH(el"') I
1
1-

I-~~~~

W/UiH/iJ4

~~--~----~~""aA~~w
II

1m

~----~~~------~Re

FIGURE '-8
Mapping frequency regions in the specification opto the unit circle in the z plane.
t

249

150

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

Hint 2. The magnitude response over a narrow band of frequ~ncies can

be decreased by placing a zero at the angle corresponding to the center of the

band. For maximum a~enuation, the zero is placed on the unit circle.
Hint 3. When the log-magnitude responses of two singularities are added
together, the location of the peak (for poles) or of the valley (for zeros) may
shift owing to their interaction. This effect is shown in Fig. 7.9. In general,
when two poles are close to the unit circle or distant from each other, there is
<a)

(b)

pole at

0._ L

00

pol'. at p

I
I

..
I

.."_

_------------_

sa e._ L

o
__
-D.,";,--------~-__

pole III ... O.toO

dO.OO

-I
I

-1100;

_A~.----------~.-q-

------I

L IOO.GO

1:~:)la

':~I.

~1

."~
-

-~'i
U~

1,1

III

I
--t
-........---------:-..---------------!
L 10.00

aZII1 at 0.100

-::1

"

- . I

pol. at. .. 0.100

-".0.I

- .

-~'L.-

-10.0+

I
I
I
I

. ..

-"j

~'i

pole - '

0.100 L 10.00

,:,:,10
I

~~ldB

L 100.00

I
I

i - II

-15.0.-

II
II
-JO.~-------,_--------!

DOle -*- .. 0._ L . . . ad .& .....100 L IOO.GO

I:~'IGI

lac.J1a

00'1

~I----~--------I
~~

II

II

-10'"

-'''OJ

I
I

I
I

I
-~-------------o

I
I
I

....""

~I-----~--~-----~

-t'j
I
-11.0+

-u.o;

-"

-~~~.---------~-~-.---------

FIGURE 7-1)
Interactions of two poles. a) Interaction produces two peaks when poles are close to the unit
circle. b) Interaction produces one peak when poles are farther from the unit circle.

FROM ANALYSIS TO SYNTHESIS

251

little interaction. When the poles are far from the unit circle or close together,
their interaction is greater.
Two eamples of performing the filter designs are given next. An
algorithm for programming this procedure is suggested in Ute next section.
Example 7.5. Interactive design of a Iowpass filter. Let us consider the
following lowpass specification:
passband:
for 0 ~ I) ~ 1t/4
stopband:

IH(el"')ldB < -50

The pole/zero pattern indicating the placement of singularities is shown in Fig.
7 .10a. The sequence of magnitude response curves are shown in Fig. 7.10b. These
curves were produced by the log-magnitude plotting routine PLOTLM with
DBVAL = -4, SPEC = -1, IFI = 17 and IF2 =0 for the passband plots, and for
the stopband plots DBVAL = -60, SPEC = -50, IF1 =33 and IF2 =O.
The following sequence of poles and zeros resulted in a filter design that
met the specification:
SIDguJarities 1 and 2: A pole is placed in the middle of the passband at
radius :: 0.7 and a zero is placed in the middle of the stopband on the unit circle.

lin
(a)

./'" PASSBAND

--~------------~'-~~----*----r--Re

FIGURE 710
Interactive design of lowpass filter. a) Pole/zero pattern in z plane. b) Sequence of log-magnitude
response curves. c) Digital filter structure.

252

INTRODUcnON 1'0 DIGITAL SIGNAL PROCESSING

dQplar1tt Il ". PDh at. so- O. 'fOCI L 0.00

a1qphI'1tJ ft b 1*1"0 lit . . 1.000 L .to.OO

I:~!s.

1:~:'la

I
I

,-!

>"'1

>"j

__

- .ot'

I
I

~.,~.i~. ______~~~MMMM~~~M_MM~_~"~"~"""""~!
IIlap1.&ntJ "

. . . UI'O . , . . 1.000

--

,. u

pole

as ..

,I

. I,

I
I
I
I

....., ---"
, ,
I
I
I

..... 0.

a1IIplaritJ' "

18

UI"O

>1'' ,1
I

-'.0+

-j--

-I-

--

-40.0

-I

at .. 1.000 L 121.00

~1!s._

18(.".

,,
......,
I

--

I
I
I
I
-I
I
-I

-JCl.O+

I
I

,
I,
I

>Ui

-1.0

-J.o.

I
I
I
I
I
I
I
I

-" ,

--

I
I
...0.0

I
I
I
I

I-

~~.

0.0

>"',

I
I
I
I
I
I
I
I
I
I
I
I
-II
-II
I

0.100 L 10.00

lat-J1a

I-

--

.......

-'.0+

'1

- -II
I
-I

-..~.'--------I,--------~

I
I
I
I

I
I

-10.0.

II

I
I
I
I

-1.0+

"',,-

.......I

~"

,--

~"'j

I
I
I
I
-lI.o.
I
I
I
I

II

.....0

>U'j

I:~

I
II
I

-1.0

I,

--!-

L 00

lae.llem
0.0

,
I-

-I
I

~..I---1
.. :----~~*~*------**~----~

DGURE MOb

>U'j

,I

--

--

.. j

-,

.....'j

--I

~I-----------------~~---------------~
______________ ____MM__________

I~

fROM ANALYSIS TO SYNTHESIS

dIplWItr "

...

tole III a- O. HID L 00

1:~:)t91

t:~:)la
I

I
I
I

!..

-. a!

...

-u.o+

-i

-10.0+

-fJ.o+
- . .,

III

.o----i------------.;
--

-4....

~"""'.poHllta-O.@L

--

~A
..- - - - - - -____~........................__~

11I1."g

'.'\

"1J.Oj

-"'1

I
I
I

-....I
-..

..

..oa

I:~-t---------------110

Z53

-10.0+

. t_

I1

-4 ....- - - -

---~-....- -....- - - - . ;

FIGURE 71Ob (Continued)

-00'I

--

~------~~--------~

.I

....--....-----w

.qA~~------------~--

0.0027

z(n)

y(n)

FIGURE 7tOe

2S4

1NTR0DUcnON TO DIGITAL SIGNAL PROCESSING

Singularity 3: A zero is placed at the edge of the stopband on the unit circle.
Singularity 4: A pole is placed at z = 0.ge~j5O" to boost the magnitude
response in the passband. A mmPmum is produced slightly past the edge of the
passband. Having such a maximum at the edge simplifies the procedure to satisfy
the passband requirement.
Slnplarity 5: A zero is placed on the unit circle at the location of the
maximum of the response in the stopband. The stopband specification is then
nearly satisfied, but the response at the edge of the passband has been reduced.
SingularIty 6: A pole is situated at z = O. 75e~j4O" to boost the response at the
edge of the passband.
SIDgularity 7: A pole is added at z =O.6e#J8" to boost the response in the
middle of the passband. A check of the stopband indicates that it too is satisfied.
The final filter system function is given by'
0.0027(1 + z-I)(1 + z-~(1 + 1.03z- 1 + Z-2)
H(z) = (1-0. 7z- 1)(I-1.16z- 1 +0.81z-2)(1-1.15z- 1 +O.56z 2)(1_ 0.95z- 1 + 0.36%-2)

The filter structure is shown in Fig. 7.1Oc. The gain constant 0.0027 was
determined by finding the maximum value of the ~og-magnitude response
sequence and applying Eq. (7.16).

Eumple 7.6. JnteradiYe design of a bandpass mter. Let us consider the

following bandpass specification:
passband:

-1 < IH(ei"')ldB s O.

for O.42n' s

l.r)

S 0.61.1r.

stopband:

lH(eiOldB < -SO,

for Os l.r) sO.16.1r and for O.frl1f s

l.r)

S.1r.

The pole/zero pattern indicating the placement of singularities is shown in Fig.

7.lla. The sequence of magnitude response curves are shown in Fig. 7.11b.
The following sequence of poles and zeros resulted in a filter design that
met the specification:
SIDpIaritIes 1-3: A pole is placed in the center of the passband and zeros
are placed at the centers of the stopbands on the unit circle.
SfDpIarWes 4 and 5: Poles are placed at the edges of the passband to
increase the gain. The magnitude response indicates maxima at the edges of the
passband, in anticipation of the reduction that will occur when zeros are added in
the stopbands.
SJnpIarities 6 .... 7: Zeros are placed at the edges of the stopbands to satisfy
the stopband specification. The magbitude response at the edges of the passband
has also been reduced so that the passband requirement is also met.
The final filter system function is given by

The filter structure is shown in Fig. 7.lIe.

25S

FROM ANALYSIS TO SYN1HESIS

_STOPBAND

STOPBAND'"-.

--~------------~~~~-L--------6---Re

<a)

=::::::::=::::1
'!::'
a ..
abpbrttJ 0

"""I..
1.1
I

-s

b ....0

1.1iOO

L lID

~I
I

II

.I

I" '.1
I
!

I
I

I'

-1."

I
I

-I ...

....I
~

--

:1

.......----.

1.c.,I.
I
I
I
I
-u ...

--

I
I
I

II

0
1 0

J..-

-I
I
I
I

-11.01>

-<J.G<I

__.LI

10

II

II

I
I
I
I

I
I
I
I

'0

II

il
I I
--I I

"-

01
01

II

I1--0II

71..,
.

: ............... L
III pole at .. 1.110 L w

~ "

lat.J1ca

rl
I
I

-u.j

-SOo-;
I

I :

"I
I

0_0

--

I-I
I II

0-1

1-

I'
I

I
I
I
I
I
1
I
I
I
-to.O;-li----_------,I~

-<J.Oj:
0

(b)
FIGURE 711
Interactive design of bandpass filter. a) Pole/zero pattern in z plane. b) Sequence of log-magnitude
response curves. c) Digital filter structure.

256

INI'RODUcnON TO DIGITAL SIGNAL PROCESSING

II1IIp1....str "
~tr

,., b

&W'O al . .
a N at. . .

1.000 L tn.OII
1.000 L 11.00

.
-1"1
I
-a.OtI

I:~~I'--~--~--I
I
I

1II

_~

-10."1
I
I
I

......!

-'.04-I

--
...
0 --.:~-.--~I__
I---""';

I
I"
,-

I
.
I

I
I'
I.

I.

I
-, .
-I

I'I
I

1
I

I
I
I

.I
-I

-,-

1-

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

fiGURE ,.1lh (Continued)

y(n)

x(n)

FIGURE

'Dc

7.10 WRITING AN INTERACllVE

DESIGN PROGRAM
To achieve a flexible operation, the interactive filter design program should be
composed to include an additional single real or complex-conjugate pair of
singularities at each iteration. The log-magnitude response of the filter section
generating these singularities is computed and. added to the total filter
log-magnitude response. The running total log-magnitud~ response is then
compared to the specification as each singularity is added. When the
specification is met, the filter design is completed. .
The logic diagram of this procedure is shown in Fig. 7.12. The procedure
starts with the filter log-magnitude response specification indicating the
passband and stopband locations and their respective attenuation levels. These
values are inserted into the PLOTLM arguments to indicate these specified
frequency and attenuation limits on the log-magnitude response plot. The filter
response is monitored with the HTOT(N) array, for 1 s N s 65, that contains
the current valu~ of the total filter log-magnitude response for 0 S CI) S n. This
total filter log-magnitude response. is initialized to 0 dB. The responses of the

FROM ANALYSIS TO SYNTHESIS

Specify pas..band and stopband values for PLOTLM

Inillahze {HTOT} to zero

yes

determine {BCOEF} as If slngulanty is zero

(ZERO)

Set umt-sample respoose {H} equal to {BCOEF}

compute log-magnitude response {HDB}

\FFT).\MAGPHS). [LOGMAGI

no

} indicate array names

I indicate subroutines

FIGURE 7-U
Logic diagram for interactive filtet design

prog~jlID.

257

258

INfRODUcnON TO DIGITAL SIGNAL PROCESSING

sections are computed and accumulated in {HTOT}. The values in the HTOT
array are superimposed on the specification curve with PLOTLM.
The interactive filter design procedure should perform the following
steps.
Step 1. Input the desired singularity type (pole or zero) and position.
Step 2. Compute the corresponding filter section log-magnitude response

and add this section response to {HTOT}.

Step 3. Compare the current values in {HTOT} with the desired specification by plotting with PLOTLM. If the specification is not met, return to Step 1
to include additional singularities. Otherwise, the filter design is done.
The logic diagram illustrates an efficient procedure for determining the
log-magnitude response. Further, this procedure does not depend on the
duration of the unit-sample response of the poles. It uses the feature that
the poles contribute terms to the denominator of the system function. As
such, the inverse elemental magnitude response contribution is computed by
treating the pole as a zero, determining the log-magnitude response of this
zero, and then subtracting it from {HTOT}. This subtraction is accomplished
simply by assigning the correct value to the sign variable: if the specified
singularity is a zero, sign = 1, but if the singularity is a pole, sign = -1.
The filter structure should be implemented as a cascade of first and
second-order sections which has the minimum number of delay components.
To determine the value of the filter gain A, the maximum value HMAX of the
final log-magnitude response sequence {HTOT} should be found and substituted into Eq. (7.16).
This same program can be used for evaluating the IIR and FIR filters
described in the next two chapters. In those cases, however, the iterative
feature should not be necessary.

7.11 SUMMARY
In this chapter, the analytic techniques discussed in the previous chapters were
summarized in such a way as to be useful for the synthesis of digital filters. The
filter design section of the book began by considering some general issues that
should be addressed, before: a filter design is initiated. A procedure for
designing a digital filter by intuitive placement of poles and zeros in the z plane
was described. This procedure provided an understanding of the effects of
poles and zeros on the transfer function and an appreciation of the conventional filter design techniques to be studied in later chapters.
In the next two chapters, we will present the theory for placing the poles
and zeros in the z plane from the magnitude response specification. This
theory will help to formalize the intuitive procedure.

FROM ANALYSIS TO SYNTHESIS

PROBLEMS
7.1. Determine the system function H(z) for each of the following .magnitude and
phase responses (Let Arg[H(ei<)] =0):
a) IH(ei",W= 1/[5 -4cos(w)]
Arg[H(ei"')] = arctan[-2 sin(w)/(2cos(w) -1)]
b) IH(ei OJ)1 2 = 17 - 8cos(2w)
Arg[H(eJ"')] = arctan[4sin(2w)/(1-4cos(2w]
c) IH(ei"')1 2 = 4/[1- 2a 2 cos(2w) + a4 ]
Arg[H(eij] = arctan[-a 2 sin(2C1)/(1- a 2 cos(2w]
7.2. Determine the system function H(z) for each of the following magnitude and
phase responses:
a) IH(ei"')1 =2[1-cos(w)]
Arg(H(eiCD)] = -re, for all w
b) IH(ei")1 = 2Isin(w/2)1
Arg[H(ei"')]=(re - w)/2, for 0< w:s re
=(-re - w)/2, for -re < w <0
c) IH(eICD)I=4Isin(w)1
Arg[H(ei"')]=-re/2, for 0 < Ql < re
=re/2, for --re < w < 0
7.3. Determine a valid system function H(z) for each of the following magnitude
responses:
a) IH(eJaJ)1 = Icos(w)1
b) IH(ei"')1 = Isin(2w)1
7.4. Implement three digital filters that have the same structure but different coefficient
v8Iues and that have the same magnitude response, given by

IH(ei"')1 = 1 +cos(w)/2.
Hint: assume zero phase and determine the pole/zero pattern.
7.5. Implement the digital filters in Examples 7.5 and 7.6 in the following forms: a)
Direct Form II structure, and b) parallel connection of second-order sections.
7.6. Consider a second-order all-pole filter having poles at z = re~i8 and zeros at z = O.
Find a general analytic expression for the frequency at which the magnitude
response is a maximum. Evaluate this frequency and the maximum magnitude
when (J =rc/6 and re/2 and r =0.5 and 0.9.

COMPUTER PROJEcrs
7.1. Magnitude response curves fqr poles and zeros
.
Object. Determine the log-magnitude response curves for single and complexconjugate singularities as a function of radius and angle in the z plane.

260

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

Write the subroutine that computes the log-magnitude response of the

magnitude response array given by

LOGMAG(HMAG,NH,HDB).
ADaIytic results
(a) Determine the log-magnitude and phase response for first and second-order zeros
and poles given below.
(b) Determine the maximum value of the magnitude response.
(c) Verify the locations of the half-power frequencies for the singularities below.
Convert these frequency values to the nearest index values in the DFf sequence
and indicate these on the computer plots below.
Computer results
(a) Compute and plot the log-magnitude responses using PLOlLM with
DBVAL= -20., SPEC = -3. and 1Ft =1F2=0.
(b) For each singularity draw the digital filter block diagram. Determine and
include the filter gain A that makes the maximum of IH( ei")1 over O:S; Cl) :$ .1t'
equal to 1. Indicate the procedure you used to determine A.

Real-valued singularities (opposite singularity located at z = 0):

(i)
(ii)
(iii)
(iv)

zero at z = t
zero at z = -0.75
pole at z = 0.96
pole at z = -0.6

Complex-valued singularities (opposite singularities located at z = 0):

(i)
(ii)
(iii)
(iv)

zeros at z = epr/4 and at z = e -i",/4

zeros at z = eJIf/8 and at z = e- i,./8
poles at z = 0.8ei>r/4 and at z = 0.8e- Jlf/4
poles at z = O.8eilf/8 and at z = O.8e-JIf/8

Instructions for the FDter Design Projects

In the projects below include the following for the final filter:
the filter block diagram
the system function
the pole/zero pattern in the z plane
the log-magnitude response plots
the phase response plot (indicate the causes of the phase jumps).
For these interactive design projects, once a successful filter design has
been achieved, include the sequence of log-magnitude response plots obtained
as the filter converged to the final design. To economize on paper, show only
the log-magnitude plots with each additional two singularities and ignore any
singularities that were canceled and not used in the final filter. Below each
intermediate log-magnitude response plot, sketch the corresponding z plane
pattern of poles and zeros and indicate the two new singularities that were
.
added.

FROM ANALYSIS TO SThlHESIS

261

7.2. Intuitive filter design

Object. Design a filter that satisfies a given specification by pole and zero
placement in the z plane.
Computer results. Design a digital filter that meets one of the specifications given
in Appendix B. The specification is considered satisfied when, in the plot produced
by PLOTLM, all the points of the rilter log-magnitude response lie completely
inside the passband and stopband limits.
7.3. Intuitive linear-phase filter design
Object. Design a linear-phase filter that satisfies a given specification by zero
plaCement in the z plane.
A linear-phase filter is an all-zero filter whose zeros in the z plane form
complex-conjugate quadruplets. When the zero is located on the unit circle it
forms its own reciprocal. When the zero falls on the real axis, it forms its own
complex conjugate.
Write an interactive program to design a filter by using only the ZERO and
NONREC routines. When a zero is entered that is not on the unit circle, have the
program automatically generate the reciprocal zero.
Computer results. Design a digital filter that meets one of the specifications given
in Appendix B.

CHAPTER

8
INFINITE-IMPULSE
RESPONSE
FILTER
DESIGN
TECHNIQUES

8.1 INfRODUcnON
Infinite-Impulse Response (UR) filter design procedures are extensions of
those originally developed for analog filters and start with the design of the
appropriate analog filter in the analog frequency domain. In fact, IIR digital
filters are commonly used to replace existing analog filters. The complexvalued Laplace variable, s = (J + jQ, plays the same role for the analysis of
analog filters as the variable z for discrete-time filters. The three most
commonly used analog filter synthesis techniques for designing a lowpass filter
in the s plane are the Butterworth, Chebyshev and elliptic techniques. These
three are described and their advantages and disadvantages are compared. The
impulse-invariance and the bilinear transformation methods are then presented
for translating the s plane singularities of the analog filter into the z plane. A
digital filter can then be implemented from these z plane singularities using the
methods described in the previous chapters. Frequency transformations are
employed to convert lowpass digital filter designs into highpass, bandpass and
bandreject digital filters. All-pass filters are employed to alter only the phase
response of the IIR digital filter to approximate a linear phase response over
the passband.
262 .

INFINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

263

8.2 ANALOG FILTER SYSTEM

FUNCflON AND FREQUENCY
RESPONSE
In this section, we establish the notation for the analog <lesign techniques_ The
system function for an analog filter will be denoted by H(s), where s is the
Laplace complex variable s = (1 + jQ, with Q the analog frequency_ The
frequency transfer function of an analog filter, denoted by HOQ), is obtained
by evaluating the system function H(s) in the s plane along the frequency axis,
or
(8.1)
H(jQ) = H(s)\s=jQ
The squared magnitude response, IH(jQ)12, indicates the power transfer
characteristic of the filter and is the function commonly employed for analog
filter synthesis. The following equivalent forms are useful in the design of
analog filters:

IH(jQW = H(jQ) H*(jQ) = H(s) H( -s)\s=jQ

(8.2)

The analytic form of the last identity indicates that the singularities of
H(s) H( -s) in the s plane are symmetric about the jQ axis. To obtain a stable
system function from H(s)H(-s), the poles that lie in the left half of the s
plane are assigned to H(s). The assignment of the zeros is arbitrary.
Eumple 8.1. Analog lowpass filter. Let us consider the lowpass power transfer
function given by
IH(jQ)12 = (a 2 + Q2)-1

To obtain the system function, we replace Q with s/j = -js, to get

H(s) H(-s) = (a 2 -s2t 1 = [(a +s)(a -sW 1

The poles of H(s) H( -s) are located at s = a. Assigning the pole in the left half
of the s plane to H(s) for stability, we have

H(s) = [s + a]-l

8.3 ANALOG LOWPASS FILTER DESIGN

TECHNIQUES
In this section, we consider the problem of placing poles and zeros in the s
plane to satisfy a given magnitude response specification. The Butterworth,
Chebyshev and elliptic filter design techniques are described. Each is related to
the type of approximation to the desired magnitude response that is employed.
Each approach is illustrated by designing a lowpass filter to meet the
specification shown in Fig. 8.1. In the next section, two methods are described
to change these analog lowpass filters into digitallowpass filters. The following
section then describes frequency transformations to convert the digital lowpass
filter into one that is either highpass, bandpass or bandreject.

264

INTRODUcnON TO DIGITAl. SIGNAL PROCESSING

IHun) IdB

-1-~~

Op
(14041t)

Os

(82681t)

These analog lowpass filters are commonly employed as anti-aliasing

filters, that are applied to continuous-time signals before analog-to-digital
conversion. They are also suitable as interpolation filters to convert pulses,
having amplitudes proportional to the value of the elements in the discretetime sequence, into signals that are continuous in time. Both these applications
were described in Chapter 3.
BU'ITERWORm FILTER DESIGN PROCEDURE. Of the three filter design
procedures described in this section, the Butterworth is the easiest to apply. Its
simplicity lies in the straightforward pole/zero configuration of the filter
transfer function in the s plane: all the poles lie on a circle centered at the
origin of the s plane and all the zeros lie at infinity. The Butterworth lowpass
filter magnitude response is also the smoothest, decreasing monotonically with
frequency.
The Butterworth lowpass filter of order N is defined by

(8.3)
where Qe is the cutoff frequency. When s = jQe, we have IHB(jQdl z = !, or in
logarithmic units, IHB(jQdldB = -3. Since the highest power of s in the
denominator of Eq. (8.3) is 2N, HB(S) HB( -s) has 2N poles. These are located
in the s plane at the values of s for which
(s/jQd 2N = -1

The roots can be found by mUltiplying both sides of Eq. (8.4) by (jQd
give

(8.4)
2N

to

(8.5)

Since the value of a complex number does not change when it is multiplied by
eiZ"m, for integer values of m, we have
S2N

= ei"!Qe ei"I2)2N ei2m"

=

o.t! ei,,(N+2m+l)

(8.6)

INFINITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

265

jn

11

=3

,,

n=4
\
\
\
\

\
\

~~-----/~/\-\--+-~*-~a
I
I

--~----~~--~-+~a

,..,.. ....... '

... '

.......................
\

........

\
\

FlGURE 8-2
Pole pattern in the s plane for HB(S) HB ( -s) for N = 3 and N = 4.

The distinct locations of the poles are found by taking the 2Nth root, to give
Sm

. (N +2m
= gcexP [ )11"
2N +

1)J

for 0 $ m $ 2N - 1.

(8.7)

The mth pole, for 0 $ m $ 2N - 1, lies on a circle of radius gc and at an angle

with respect to the real axis equal to (N + 2m + 1):rr/2N. The pole configurations in the S plane for N = 3 and N = 4 are shown in Fig. 8.2. The poles are
equally spaced on a circle of radius Qc, but no poles lie on the jQ axis. The
poles of HB(S) HB ( -s) in the left half-plane are those belonging to the desired
system function HB(s), while those in the right half-plane are from HB(-S).
The 2N zeros of HB(S) H B( -s) lie at S = 00, and half of these are assigned to
HB(S). Later in this chapter, we present two methods for transferring these S
plane singularities into the z plane to obtain a digital Butterworth filter.
The power transfer function of the ButterWorth filter is obtained by
substituting jQ for S into Eq. (8.3), to give

IHBOQ)I' ~

(~

1+ Qc

(8.8)

For large g, the magnitude response decreases as Q-N, indicating the lowpass
nature of this filter. The log-magnitude response is computed as
IHB(jQ)ldB = 10 loglo IHB(jQ)12
= -10 loglo [1 + (Q/Qc)2N]

For large g, the log-magnitude response is approximately equal to

(8.9)
As Q increases by a factor of 10 or a decade, the log-magnitude response
drops 20 dB for e~ery pole, or -26N dB/decade. If Q increases by a factor of

266

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

_~r---=:::::,....

N= I (-6 dB/octave)

'~:---~~--

-20

-40

-wr-------r-------------------~~~----

_~~------L-------------------------__.

Oc

radls

FIGURE 8-3
Analog log-magnitude response curves for Butterworth lowpass filters of order N.

two, or an octave, the response drops by 201og10 2, or -6N dB/octave. The

log-magnitude responses for different values of N are shown in Fig. 8.3. All the
curves pass through -3 dB at Q = Qc, and the steepness of the log-magnitude
curve increases with the value of N. As N is increased, the magnitude response
approaches that of the ideal lowpass filter. The value of N required for a
particular application is determined by the passband and stopband specifications, as shown in the next example.
Example 8.2. Butterworth Iowpass analog filter design. Let us consider the
following lowpass specification:
passband:

-1 < IH{jO)ldB:s 0
stopband:
IH(jO)ldB < -60

for 02:8268n-rad/s (Os = 8268n-)

DeterminiDg the valae of N. If the cutoff frequency Oc were given, then we

could solve the following inequality:

IHB(jOs)1 2 = [1

+ (Os/Od2Nr 1 < 10-6

(-60 dB)

for N. The solution is then

N> log(l~ -1)

2log(Os/Od
Since Oc is not given, a guess must be made. The Butterworth log-magnitude
response approaches a slope of -6N dB/octave at high frequencies. The
specification calls for a drop of - 59 dB in the frequency range from the edge of
the passband (1404n-) to the edge of the stopband (8286n-). The frequency

lNFINlTE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

267

difference is eq~al to log2(8268/1404), or 2.56 octaves. Hence, an initial guess is

found from
I
I

-6N = -59/2.56

or

= 3.8

Since the order f the filter must be an integer, we choose the next higher integer
value, or N = 4.
We must qow compute the value for the cutoff frequency Qe, which is also
the radius of th~ Butterworth circle. One value for Q e is obtained by solving the
inequality at thel stopband edge with N = 4, or
IHBUQsW = [1 + (Q s/Qc)2Nrl < 10- 6
or
Qe < Qs X 1O-- 6I2N = Qs/5.62 = 1470.3Jr
Let us choose de = 1470Jr. This inequality assures that the stopband specification
will be met with! these values of Nand Q e . We need to test whether the passband
specification is ~lso satisfied, that is, we must test whether
'fBOQpW = [1 + (QpfQc)2Nrl :> 0.794

( = -1 dB)

Evaluating, we jind
[1 + (1404/1470)8r 1 = 1/1.692 = 0.59
The result is be~ow the passband specification. Hence, a fourth-order filter is not
sufficient. This is not unreasonable, since we employed a large Q argument to find
a first guess for N.
Let us inQrease N to 5. The stopband specification gives the Q c value:
Qe < Q s X 1O- 6I2N = Qs/3.981 =2076.8Jr
Using Q e = 2076Jr to test the passband specification, we get
,lHBUQpW = [1 + (1404/2076) lOr 1= 1/1.02 = 0.98
Hence, the stdpband specification is met and the passband specification is
exceeded with N= 5 and Q e = 2076Jr.
We could .also obtain a filter in which the passband specification is met and
the stopband specification is exceeded by using N = 5 and solving the passband
specification for: the value of Qe . Doing this we get
IHBUQpW = [1 + (Qp/Qc)1'1-1 > 0.794
or
10glO Qe > 10glO Qp -

1~ 10g(0.;94 -

1)

to give Qe> 1607Jr. We then use this value to find the stopband attenuation
IHt.jQS)12 = [1 + (Q s/Qc)2Nrl = 7.69 x 10-& (= -71 dB)
Both filters m et or exceed the specification. The particular one to choose
depends on ot er considerations in the design. To continue the example, we
choose N = 5 and Q e = 2076Jr.
Butterwot pole locations. The locations of the N poles on the Butterworth
I

268

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

circle of radius
Sl

Qc

and in the left half of the s plane are given by

= -2076Jr,

S2.3 = 2076Jr(cos(4Jr/5) j sin(4Jr(5 = -1680Jr j1220Jr = 2076Jr ejl44'

S4.5

= 2076Jr(cos(3Jr/5) j sin(3Jr/5 = -641.5Jr j1974Jr = 2076Jr ejlO8"

These are shown in Fig. 8.4(a). The system function is then given by
R (s) _
B

(2076Jr)S
[s + 2076Jr][S2 + 3359ru + (2076Jr)2][s2 + 1283ru + (2076Jr)2]

The log-magnitude and phase responses for this filter are shown in Fig. 8.4(b).

There are several interesting features that are peculiar to the Butterworth
magnitude response:
1. It decreases monotonically as the frequency 0 increases from 0 to 00. This
characteristic is evident from the functional form given in Eq. (8.8). It is
due to the configuration of the poles lying on a circle in the s plane. The
other two analog filter design procedures described below will not have this
smooth magnitude response charactersitic.
2. The magnitude response is maximally flat about 0 = 0, in that all its
derivatives up to order N are equal to zero at 0 = O. To illustrate this, let u~
make Oc = 1 and N = 1 for simplicity. We then have

IHB UO)1 2 =

(1 + 0 2)-1

(8.10)

jn

j20761t

(
(a)

/
FIGURE 8-4
Fifth-order Butterworth analog
lowpass filter results. a) Pole/zero
pattern in the s plane. b) Logmagnitude and phase responses.

INFINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

269

PASSBAND:
IHB(jn)ldB

o
-I~-~-+----------

-2
-3
4L-__

L-~~

________________

1404n

10,000 It

n (radls)
STOPBAND:
IHBun)ldB

0----

-10
-20

-30
-40

-50
-rol~---~-----~~~--

-70
_801L-_ _ _ _ _-------~-----J

8268n 10,000 It

n (radls)
Arg[HBun)]
It

_ _ _...;;.;10;,,::,000 It

-It

FIGURE 8-4b

The first deljivative is then equal to

d 1H8(jQW
dQ

which is zert at Q

= 0_

-2Q
(1 + Q2)2

(8.11)

The second derivative is equal to

d2 1H80 Q
dQ 2

W_ -2(1 + Q2)2 + 8Q 2(1 + Q2)

(1 + Q2)4

Note that tl1e second derivative is

e~ual

to -2 at

= O.

270

INTRODUCl10N TO DIGITAL SIGNAL PROCESSING

3. The phase response curve approaches - N:n: /2 for large 0, where N is the
number of poles on the Butterworth circle in the left half s plane. When the
value of N is large, the discontinuities in the principal value of the phase
response occasionally cause difficulties in the interpretation of the phase
curve.
The Butterworth filter is the easiest one to design. It is commonly used
because the smoothness of the magnitude response. This smoothness is bought
at a price, however. The Butterworth characteristic requires a relatively large
transition range between the passband and stopband. We next consider the
Chebyshev filter design that has a smaller transition band, but its magnitude
response will not be monotonically decreasing with frequency.
CHEBYSHEV FILTER DESIGN PROCEDURE. The Chebyshev filter is optimum in that for the given passband and stopband levels, it has the smallest
transition region, Os - Op, of any filter that consists only of finite poles. Such
all-pole filters have all their zeros at s = 00. The Chebyshev filter is defined by

(8.12)
where fJ is a measure of the allowable deviation in the passband, as shown
below, and
(8.13)
is the Nth-order Chebyshev polynomial. The corresponding power transfer
function is given by
(8.14)
The Chebyshev polynomials can be determined from the following recursion
formula:
for N;;:: 1
with Co(x) = 1 and C 1 (x)
Table 8.1.

= x.

The first five Chebyshev polynomials are given in

TABLE 8.1

Chebyshev polynomials CN(x)

N

CN(x)

0
1
2
3
4
5

x
2x 2 -1
4x 3 -3x
8x 4 _8x2 + 1
16x s - 20x 3 + 5x

(8.15)

INFINITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

271

The following two features of Chebyshev polynomials are important for

filter design.
1. ICN(x)1 S 1 for Ixl S 1. Inserting this result into Eq. (8.14) indicates that, in
the passband, the power transfer function lies in the range

(1 +Jl 2)-lS IHe(jOWs 1

forOsOsOp

(8.16)

Whereas the frequency value important for the design of the Butterworth
filter was the cutoff frequency Oe, the relevant frequency for the Chebyshev filter is the edge of the passband Op. This allows the Chebyshev filter
to be specified directly in terms of the two important design parameters
consisting of the passband frequency and the allowable deviation in the
passband.
2. For Ixl 1, ICN(x) I increases as the Nth power of x. This indicates that for
0 Op the magnitude response decreases as O-N, or the log-magnitude
response decreases by 6N dB/octave. This large-frequency result is identical
to that for the Butterworth filter.
Where the Butterworth filter poles lie on a circle in the s plane, it can be
shown that the poles of Hc(s) lie on an ellipse. The ellipse is defined by two
circles determining the major and minor axes, as shown in Fig. 8.5a. If we
define the parameter p as
P=Jl-1+Y1+Jl 2

then r, the radius of the circle defining the minor axis, is equal to
r = Op(pliN _ p-lIN)/2

(B. 17)

(8.18)

where N is the order of the filter. The radius R of the circle defining the major
axis, is equal to
(8.19)
The locations of the poles on the ellipse are found by first finding the
Butterworth pole locations. These are then projected onto the ellipse by
keeping the frequency constant, as shown in Fig. B.Sb.
The locations of the Chebyshev poles can be determined without actually
constructing the ellipse. As shown in Fig. 8.Sb, the ordinate (jO) locations of
the poles are defined by the intersections of the N lines with the larger circle.
The abscissa (0) locations are defined by the intersections of the lines with the
smaller circle. For example, let us consider the line that makes an angle 8 with
respect to the positive real axis. The corresponding pole of the Chebyshev
filter is located in the left half-plane at s =Sp, where
Sp = r cos( 8) + jR sine 8)

(8.20)

As with the Butterworth filter, the poles of Hc(s) Hc(-s) that lie in the
left half of the s plane are assigned to the system function Hc(s), and all the
zeros of the Chebyshev filter lie at s = 00.

272

"""""UCfiON

ro DK>IT"-

""~""""""

(a)

"!l

CHEBYSHEV
ELLIPSE

-+------~----~~----~----_++u

(b)

jn

BUTIERWORTH POLES

\
ftGURE 8-S
Chebyshev ellipse. a) Construction of ellipse. b) Pole pattern in the s plane for
He<.s) He<. -s) for N = 3.

Example 8.3. Chebyshev lowpass analog filter design. Consider the same
lowpass specification given in Example 8.2:
passband:
-1 < IH(jQ)ldB :s 0

for O:s Q:s 1404n rad/s (Q p = 1404n)

stopband:
IH(jQ)ldB < -60

for Q ~ 8268n rad/s (Qs = 8268n).

INFINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

273

Value of fl. The value of /1. is determined from the passband ripple, or
10 10g(1 + /1. 2)-1 > -1 dB
This gives the value as
/1. < [10. 1 _1]112 = 0.508
We will take /1. = 0.508.
Value of N. The order of the Chebyshev filter is determined from the
stopband inequality

IHdj Qs)1 2 = [1 + /1.2CU Q smpW I < 10-6 (-60 dB)

Since Qs/Q p = 5_9, we evaluate CN (5.9) for increasing N until the stopband
inequality is satisfied, or
C (5.9) > [(106 - 1)//1. 2]112 = 1969.
N

Evaluating, we find C3(5.9) = 804, C4 (5.9) = 9416. Hence, N = 4 is sufficien~

Since this last inequality is easily satisfied with N = 4 the value of /1. can be
reduced to as small as 0.11, to decrease passband ripple while satisfying the
stopband_ The value /1. = 0.4 provides a margin in both the passband and
stopband_ We proceed with the design with /1. = 0.508 to show the 1 dB ripple in
the passband.
Axes of ellipse. The radii of the circles defining the major and minor axes of
the ellipse are determined from the constant p, which is equal to
p = 0.508- 1 + (1 + 0.508- 2)112 = 4.17
The radii are equal to
R=

14~n (4.17 + 4.171/4

114

= 702n(L43 + 0.67) = 1473n

and
r = 1404n (4.17 114 - 4.17- 114) = 702n(1.43 - 0.67) = 533n.
2
jn

(a)

FIGURE 8-6
Fourth-order Chebyshev analog lowpass
filter results. a) Pole/zero pattern in the
s plane. b) Log-magnitude and phase
responses.

274

INl'RODUcnON TO DIGITAL SIGNAL PROCESSING

PASSBAND:

IHc(fil)ldB

-2
-3

-40

1404 It

10.000 It

n (radls)
b(i)
STOPBAND:

IHcun) IdB

n (radls)
b(ii)

Arg[Hc(fil)]
It

10.000 It

n (radls)
-I[

ftGURE 8-6b

b{iii)

Pole IocatiollS. Since there are four poles, the four lines dividing the s-plane
have angles equal to 7:Jr/8 and 5:Jr/8 as shown in Fig. 8.6(a). The four poles of
the Chebyshev filter are then at
SI.2

= 533:Jr cos(7:Jr/8) j1473:Jr sin(7:Jr/8) = -492:Jr j564:Jr = 748:Jr ejl3l

and
S3,4

= 533:Jr cos(5:Jr/8) j1473:Jr sin(5:Jr/8) = -204:Jr j1361:Jr = 1376:Jr ej98S

INFINITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

275

The system function is then given by

L06Jt'4 x 1012
[S2 - 490Jt's + (748Jt')2][S2 - 203.4n:s + (1376nY]

R~)-~---------=~------~~=
C

The log-magnitude and phase responses for this Chebyshev filter are shown in
Fig. 8.6(b).

Owing to the closer proximity of the Chebyshev filter poles to the jQ axis
illan those in the Butterworth filter, the magnitude response of the Chebyshev
filter exhibits a ripple in the passband. There is a peak in the passband for each
pole in the filter, located approximately at the ordinate value of the pole. Even
though the passband contains ripples, the magnitude response beyond the
passband is a smooth function that decreases monotonically with frequency.
But, of importance of implementing the filter having the smallest number of
components, the Chebyshev magnitude response exhibits a smaller transition
region to reach the desired attenuation in the stopband, when compared to the
Butterworth filter.
The Chebyshev phase response is similar to that of the Butterworth filter,
in that it decreases monotonically to -NJr/2. In the example, this asymptote is
equal to -2Jr.
Because of the proximity of the Chebyshev filter poles to the jQ axis,
small errors in their locations, caused by numerical round-off in the computations, can result in significant changes in the magnitude response. This is
especially true in the passband. Choosing a smaller value of IJ will provide
some margin for keeping the ripples within the passband specification.
However, too small a value for IJ may require an increase in the filter order.
In *e Butterworth and Chebyshev filters, we were constrained to use
only finite ~les to achieve the magnitude specification. It is reasonable to
expect that if relevant (finite) zeros were included in the system function, a
lower-order filter can be found to satisfy the specification. These relevant zeros
.could serve to achieve additional attenuation in the stopband. The elliptic
filter, described l1ext, does exactly this.
ELLIPTIC FlLTE, DESIGN PROCEDURE. The elliptic filter has thp smallest

transition band of the three approaches described here, but is the most
complicated analytically. The elliptic filter is defined by
(8.21)

where E is a measure of the allowable deviation in the passband, Q p is the edge

of the passband, and EN(X) is the Jacobian elliptic function of order N. The
analysis of this filter is rather complicated and beyond the scope of this book,
but can be found in the references. It is sufficient to state here that, unlike the
Chebyshev polynomial, EN(S) can be considered to be a rational function,
consisting of both a numerator polynomial and a denominator polynomial in s.
The denominator polynomial of EMs/jQ p ) would then appear in the numera-

276

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

tor of HE(S) HE( -s) and its roots would then become the zeros of the system
function. For example, although simpler than the first Jacobian elliptic
function, let us consider the following rational function,

(S2 -1)
(s) = (S2 + 1)

(8.22)

Then

(8.23)
The system function is then given by

H(s)H(-s) =

1
2(3'2-1)2
1 + (S2+ 1)2

_
(S2+ 1)2
- (S2 + 1)2 + (S2 - 1)2

(8.24)

Note that H(s) has zeros at s = j, and that these zeros came from the
denominator of (s).
The elliptic filter follows an obvious extension of the pole/zero placement
procedure of the Butterworth and Chebyshev filters. The Butterworth transfer
function decreases monotonically over the entire frequency range. By a1lo~g
ripples to occur in the passband with the Chebyshev filter, the transition region
is reduced and the attenuation decreases monotonically in the stopband.
Following this logic, the transition region can be reduced still further by also
allowing a ripple to occur in the stopband. This ripple is accomplished in the
elliptic filter by placing the zeros on the jQ axis in the stopband. For the
Butterworth and Chebyshev filters, all the zeros were located at s = 00.
The zeros of HE(S) HBC -s) appear in sets of pairs on the finite jQ axis.
One of each pair is assigned to HE(s), when HE(S) H E( -s) is factored and the
singularities are distributed between HE(S) and HE( -s). In addition, for the
coefficients to be real-valued, the zeros must also occur in complex-conjugate
pairs. Hence, if there are two zeros at S = jQo, then there are also two at
s = -jQo. If the order of the system function N is odd, BE(S) has one zero at
s = 00. For a lowpass filter, there are no zeros at S = O.
This additional complication with the placement of the zeros makes the
design of the elliptic filter very complicated analytically. To circumvent the
mathematics, but yet achieve an approximate elliptic filter magnitude response, we use the insights gained by the intuitive pole/zero placement
described in the previous chapter. The elliptic filter design will be approximated by initially designing a Chebyshev filter, to determine the approximate
location of the poles. Then the pole/zero pattern is modified by placing zeros
on the jQ axis to achieve additional stopband attenuation. These additional
zeros, however, also reduce the magnitude response in the passband. If the
passband specification is not meyafter these additional zeros are included, the

INFINlTEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

277

l()le locations must be slightly modified to satisfy the specification. This

rocedure is illustrated in the following example.
Eumple 8.4. ApproJdmate elliptic analo& filter design. Let us consider the
previous lowpass specification:
passband:
-1 < IH(jO)ldB ~ 0

for 0 sO s 1404n rad/s (Op = 1404n)

stopband:
for 02: 826&r rad/s (Os = 8268n).

IH(jO)ldB < -60

To approximate the elliptic filter, we design a Chebyshev filter having a lower

order than that in the previous example, and then add zeros to satisfy the
specification.
v... of f&o The value of #J was determined from the passband ripple in
.
Example 8.3, to give #J = 0.508.
V... of N. The order of the Chebyshev filter in the previous example was
4. Let us consider a third-order Chebyshev filter.
Axes of tile eDipse. The radii of the circles defining the major and minor axis
of the ellipse are determined from the constant p. In Example 8.3, p = 4.17. The
radii are equal to
R=

1~n (4.171/3 + 4.17- 113) = 702Jf(1.61 + 0.62) = 1566n

and
r

=1~n (4.17

113

4.17-113) =702Jf(1.61- 0.62) =694n

Pole 1ocadoIIs. Since there are three poles, the three lines dividing the 5
plane have angles equal to n and 231:/3, as shown mFig. 8.7(a). The poles are
then located at

= -694n
52.3 =694n 008(231:/3) j1566n sin(23I:/3) = -347n j1356n = 1400n ejl04.,
51

The log-magnitude response for this Chebyshev filter is shown in Fig. 8.7(b}. N.
expected, the stopband specification is not met with this third-order filter,
although the passband specification is satisfied.
Zero Iocatioas. To achieve the stopband specification, we move two of the
three zeros of the Chebyshev filter, which are all located at 0 = 00, to finite values
on the jO axis. An interactive pole/zero placement procedure, similar to that
described in the previous chapter, was employed to place the zeros at
5 Z1 Z2

= j9000n

These zeros also reduced the magnitude response in the passband, such that the
passband specification is not met. The locations of the complex-conjugate
Chebyshev poles were changed, uptil both the passband and stopband specifica-

278

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

jO

(a)
PASSBAND:

IHcUo)ldB

-2
-3

10,000 It

o (radls)
STOPBAND:

IHc(jO)ldB

o
-20
-40
-ro~--------------------+_--~

-SO'';:o-----------...,,.-J.----,J
8268lt 10,000 It

(radls)
(b)

OGURE 8-7
Approximate elliptic filter
design results. a) Thirdorder Chebyshev analog
lowpass filter pole/zero
pattern in the s plane. b)
Log-magnitude response of
third-order filter. c) Logmagnitude and phase responses after zeros are included at s = j90001r.

INFINflE-IMPULSE RESPONSE FlL11!R DESIGN TEOOOQUES

PASSBAND:

IHEUO)lcIB

-2
-3
-40

1404x

10,000 It

n (radls)
STOPBAND:

IHEUO)lcIB

n (radls)
x

10,000 x

n (radls)
FIGURE 8-'c

-x

tions we{e met. 1be final pole locations are given by

SPl

= -7501r

sn.P3 = 1450 e*jlOS

The system function is then given by

H. ( )
E S

[s

61[s2 + (9()()(br )2j

+ 7501r][S2 -751m + (14501r)2j

The log-magnitude and phase responses are shown in Fig. 8.7(c).

%79

280

INTRODUCTION 10 DIOITAL SIONAL PROCESSINO

In the previous example, the stopband magnitude response is no longer

monotonically decreasing, but rather shows a sharp dip at the frequency where
the zero is located. The presence of this zero causes a phase jump of 1r radians
in the phase response curve. However, this occurs in the stopband, where
there is little effect on the output signal. In the passband, the phase response is
similar to that of the previous two filters.
Having considered three analog filter design techniq~es for determining
the locations of the poles and zeros in the s plane, we now describe two
procedures for transferring these poles and zeros to the z plane. These z plane
singularities are then used to implement the corresponding digital filter.

8.4 METHODS TO CONVERT ANALOG

FILTERS INTO DIGITAL FILTERS
To convert an analog filter into a digital filter, the singularities in the s plane
must be mapped into equivalent singularities in the z plane. This section
describes two such methods: the impulse-invariance method and the bilinear
transform method.
TIlE IMPULSE-INVARIANCE MEmOD. When replacing an analog filter with
its digital counterpart, it would be intuitively pleasing if the unit-sample
response of the digital filter resembled the sampled version of the impulse
response of the analog filter. If such a resemblance were present in the time
domain, then, intuitively, the two filters should perform in a similar manner in
the frequency domain. The impulse-invariance method does exactly this. It
produces the digital filter whose unit-sample response is equal to samples of
the analog filter impulse response, or
(8.25)

where T. is the sampling period.

We now compare the transfer functions of the original analog filter and
the resulting digital filter obtained with this method. Let us recall the
relationship between the spectrum of a continuous-time signal and that of the
sampled signal, discussed in Chapter 3. Here .the analog signal is the impulse
response of the analog filter. Its spectrum is the filter transfer function,
HAGQ), which is found by taking the continuous-time Fourier transform of the
impulse response, given by
(8.26)

For sampled-data applications, we must scale the discrete-time values to the

analog frequencies by dividing CJ) by T.. Hence, the digital filter transfer
function is denoted by H(&W'T.). By sampling the impulse response to obtain
the unit-sample response, the transfer function of the digital filtet becomes the

INFlNIlEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

281

periodic extension of that of the analog filter:

H(ejO)/T,)

= L'"

HA(jw/T. + j2lrk/T.)

(8.27)

k--oo

If HAOO) is limited to the frequency range -OM:S O:S OM, then no

aliasing would occur if the sampling period T. is chosen to be less than lr/OM'
However, since practical analog filters are composed of a finite number of R, L
and C components, HA(jO) can be zero at no more than N frequency points,
where N is the order of the filter. These points correspond to the locations of
the zeros that may be located on the finite jO axis. Hence, HA(jO) cannot be
truly band-limited, and some degree of aliasing will always occur. For filters
having a IHA(jO)1 that approaches zero as 0 approaches infinity, such as those
of lowpass and bandpass filters, this aliasing error decreases as the sampling
period T. approaches O.
EumpIe 8.5. lmpIIIse-iDvariance method applied to a first-order filter. Let us
consider the RC lowpass filter shown in Fig. 8.8. If we let a = lIRC, the impulse
response is equal to
fort~O

otherwise
The analog system function is defined as the Laplace transform of hA(t), or

HA(S) = [ hA(t) e-n dt = (s + a)-I

The analog transfer function is the system function evaluated at s = jQ, or

HA(jQ) = (a + jQ)-1

X(.)~)

II

Q.. JlllllflllllfllJ

IIUltm tn,

FIGURE 8-8
Comparison of analog and digital filters using the impulseinvariance method. Filter
structures and time-domain respoFS we shown.

282

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

MAGNITUDE:

00

Q-

10.000 It
(radls)

wIT' -+-

PHASE:
It

."

QT'-

(radls)

IO.!XXI It

Arg(HA(jQ)
Arg(H(e..... I ..)

(a)
MAGNITUDE:

iH(eJ<o>'T')i

iH.UQ)i
4O.!XlOlt

PHASE:
It

Q-

,." _

(radls)

T,

[~
40'<)()(llt
Ol~~~

\
-It

Arg[H(e..... T')
Arg[HAUQ)

(b)

FIGURE ..,
Comparison of frequency responses of digital and analog filters.
a)
= 10-4 seconds. b)
=
0.25 x 10-4 seconds.

:r.

:r.

INFINITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

283

The unit-sample response derived by sampling hA(t) every T. seconds is given by

e--1i
h(nT.) = { 0

for n ~o
otherwise

The digital frequency response is given by

H(ei OJ/1i) =

L e-""1i e- """1i = (1- e--1i e-J"'IT,r

I

"-0

The digital filter system function is determined by substituting z for eiOJ/T" or

H(z) =

..

L h(nT.)z-" = (1- e-"T,z-lr

The digital filter that replaces the RC lowpass filter is shown in Fig. 8.8.
Comparing the values of the analog and digital frequency responses at zero
frequency, we find that HA(O) is lla and H(ej~ = 1/(1- e-T,). For small values
of T., H(el"):= I/(aT.).
A comparison of IH(eIOJ/T,)1 and IHAGQ)I is shown in Fig. 8.9 for different
values of T.. As T. approaches zero, {hen)} appears more similar to hA(t),
indicating that the aliasing is less severe, which, in tum, implies that H(eiOJ/T,)
approaches HAGQ) over the range -nlT.::s wiT. < niT..

Relationship of the s plane and the % plane. We now relate locations of the
singularities in the s plane to those in the z plane when the impulse-invariance
method is employed. In the z plane, the frequency "axis" corresponds to the
unit circle, with co =0 located at z = 1. In the s plane, the frequencies are
located on the vertical axis with Q = 0 at the origin. These zero-frequency
points in the two planes should coincide. By equating individual frequency
points in the range -nIT. ~ Q ~ nIT. with -nIT. ~ colT. ~ nIT., as shown in
Fig. 8.10, we find that the jQ-axis wraps around upon the unit circle.

FIGURE 8-10
Relationship of the s plane to the z plane for the impulse-invariance method.

284

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

Continuing this point-by-point mapping, we find that the interval (k1)JrIT.::s Q < (k + 1)JrIT., for any integer k, maps onto -JrIT.::s wiT. < JrIT..
With this mapping, the left half of the s plane becomes the inside of the unit
circle and the right half of the s plane becomes the outside of the unit circle.
This relationship between the s plane and the z plane is reasonable, since, for.
stability, the poles of the analog filter must lie in the left half of the s plane,
while those for stable digital filters must lie within the unit circle.
Pole-mapping procedure. We now derive the mapping rule that transfers the
poles from the s plane to the z plane. Let us factor HA(S) into the sum of
first-order system functions given by
HA(S)

=L

_k_

(8.28)

k=l S +ak

where ak may be a complex number. The total impulse response is then equal
to the sum of the responses of the individual first-order systems, or
N

hA(t) =

ak e-aJ<l

(8.29)

k=l

Sampling this impulse response, we obtain the discrete-time sequence given by

h(nT.) =

ak e-aknT,

(8.30)

k=l

When viewed as an exponentially decreasing unit-sample response, the kth

sequence could be generated by a first-order recursive digital filter having its
pole located at z = e-akT,. The linearity of the z-transform allows us to sum
these components to obtain the digital filter system function given by
N

H(z)

= ~l 1- e~:kT,z-l

(8.31)

By comparing the pole locations in this last equation with those in (8.28), we
arrive at the pole-mapping rule of the impulse-invariance method. This rule
can be stated as
a pole located at s = Sp in the
plane located at z = eS'T,

plane is transformed into a pole in the z

In the above formuiation, Sp = -ak' Examining the magnitude of the pole in z

plane, we find that poles lying in the left half of th!! s plane are mapped into
the inside of the unit circle. Hence, a stable analog filter converts into a stable
digital filter after the mapping.
In the derivation of the rule, HA(S) was factored into the sum of terms in
(8.28), corresponding to a parallel connection of filters. As discussed in
Chapter 5, the location of the finite zeros of a parallel connection of filters
does not relate in such a simple manner, but is rather complicated. It can be
r

INFINITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

shown, however, that the zeros at s = co map to z

example.

= O.

28S

We now consider an

Example 8.6. bnpuJse-invariance method applied to a second-order filter. Let

s+a
HA(S) = (s + a)2 + b 2
s+a
(s + a + jb )(s + a - jb)
The inverse Laplace transform yields
for t 2:0
otherwise
Sampling this function produces
h(nT.) =

{~-""r, cos(bnT.)

for n 2:0
otherwise

The z-transform of h(nT.) is equal to

1- e-ar, cos(bT.) Z-I
H(z) = ---=----'-~-o--__=
1 - 2e- ar, cos(bT.)Z-' + Z-2

(1- e-(a+ j b)T,z-')(l_ e (a-Jb)r,z-I)

Comparing the two system functions, we see that the pole located in the s
plane at s = p is transformed into a pole in the z plane at z = efT., as before.
However, the finite zero located in the s plane at s = -a was not converted into a
zero in the z plane at z = e-aT., although the zero at s = 00 was placed at z = O.
The digital filteutructure is shown in Fig. 8.11.

We illustrate the application of the impulse invariance method to the

design of a Chebyshev lowpass filter in the next example.
Example 8.7. Chebyshev lowpass digital filter using impuJse-invariance
method. From Example 8.3, the system function poles of the Chebyshev filter
x(n)

FIGURE 8-11
Second-order digital filter implementation using ~pulse-invariance method.

y(n)

l86

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

were found to be
S 1.2

= -492n

j564n

We must now choose a value for 7;. The frequency responses for the analog filters
were plotted over the frequency range from 0 to IO,OOOn rad/s. To set the
discrete-time frequency range [0, n/7;] equal to this analog range, we choose
7; = 10- 4 seconds. With this value for 7;, the poles are mapped to
ZI.2 = e$1.2r. = e-o.ISSjo.ln = 0.857ejIO.2Z3.4 = ell . r. = e- o 064jO.428 = 0.938ej24.S-

and the four zeros at s =

digital filter is given by

00

are all mapped to z = O. The system function of the

7.23 x 10- 3
H(z) = (1-1.687z-1 + O. 734z- 2 )(I-1. 707z 1 + O.88z 2)
The pole/zero pattern, the digital filter structure, the log-magnitude and phase
responses are shown in Fig. 8.12. The gain of the digital filter is set so that that
maximum value in the passband is equal to 1 (0 dB).
1m

(a)

-r------------~~~----------~Re

7.23

x(n)

lO.l

t----_- y(n)

(b)

HGURE 8-12
Chebyshev digitallowpass filter results when using impulse-invariance method. a) Pole/zero
pattern for Hdz). b) Digital filter structure. c) Log-magnitude and phase responses.

1NF1NITE-IMPULSE RESPONSE FILTER DESIGN TEOiNIQUES

1Jf1

PASSBAND:

IH(eJOllTs)ldB

-2
-3

10,000 It
"" T.

(radls)

STOPBAND:

IH (ejoolTs)ldB

-~~o--------------------~--~
8268 It 10,000 It
"'Ts (radls)
Arg(H(eJOllT,)

"Irs (radls)

-It

ftGURE 1-l2c

Although useful for implementing lowpass and bandpass filters, the

impulse-invariance method is unsuccessful for implementing digital filters for
which IHAOg)1 does not approach zero for large g, such as a highpass filter_
In these cases, IHA(jg)1 does not have small values for high frequencies and an
appropriate sampling period cannot be found to prevent catastrophic aliasing_
Hence, the aliasing error terms will dominate the transfer function, as shown
in Fig_ 8.13. Because of these aliasing error terms, a highpass digital filter does
not result when the impulse-invariance method is employed_ To overcome this
problem, we need a method in which the entire jaaxis, for -00 < g < 00, maps

288

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

HA(jQ.)

(a)

III

"n

{}
HA(e""'T,)

(b)

I
0

-2,/Ts

311

CJJ/n

"'ITs

:II WITs

2'1f.ITf>

HA (ell" TS+21t1Ts

-2J[/

I
0

Ts

2./
'Ts

HA(ellli> TS-2><iT.

I
-2'1(/Ts

2KITs

II
H(e""'T')
(c)

---,

n
I
-2J[/Ts

I
0

I
2x/ T<:.

'"Irs

FIGURE 8-13
Aliasing effects prevents the application of the impulse-invariance method to implement bighpass
filters. a) Original analog bighpass filter frequency response. b) Components in the periodic
extension after sampling the impulse response. c) Resulting digital filter magnitude response.

uniquely onto the unit circle, for -H / To < w / To < H / To. This is accomplished
by the bilinear transform method described in the next section.
BILINEAR TRANSFORM METHOD. The bilinear transform method provides
a nonlinear one-to-one mapping of the frequency points on the jQ axis in the s
plane to those on the unit circle in the z plane. This procedure also allows us to

NFINITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

289

implement digital highpass filters from their analog counterparts. To illustrate

the method, let us consider the following simple bilinear transformation:

2z-1
s=--T.z+1

(8.32)

To find H(z), each occurrence of sin HA(S) is replaced by (21T.)(z -1)/(z

The inverse transformation is given by

sT.12 + 1
z = ---:;-:---sT.12 - 1

+ 1).

(8.33)

We now explore the properties of this bilinear transformation that are

useful for digital filter synthesis.
Relationship of the s plane and z plane. To find the mapping of the
frequencies from Q to w, we set s = jQ and z = eia> in Eq. (8.33) to get
iw _ jQT.12 + 1
(Q 2(T.12)2 + 1)112 ei arctan[CT,I2)
e - jQT.12-1
(Q 2(T.12)2+ 1)1I2eiarctan[ OT,I2)
=

ei2 arctan[QT,I2)

(8.34)

Equating the exponents, we can relate the values of Q to w by

w = 2 arctan[QT.12]

(8.35)

This transformation is shown in Fig. 8.14. The entire jQ axis in the s plane,
-00 < jQ :5 00 maps exactly once onto the unit circle -:rr < w :5:rr, such that
there is a one-to-one correspondence between the continuous-time and
discrete-time frequency points. It is this one-to-one mapping that allows analog
highpass filters to be implemented in digital filter form. As in the impulseinvariance method, the left half of the s plane maps onto the inside of the unit
circle in the z plane, and the right half of the s plane maps onto the outside.

Cll
1[

1[

FIGURE 8-14
The arctangent relationship between
continuous-time and discrete-time frequencies that occurs in applying the bilinear
transform .

Z90

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

The inverse relationship, relating 0 to Q, is given by

Q

= T. tan[wI2]

(8.36)

For small values of frequency, this relationship reduces to

Q= 2sin(w/2)
T. cos(wI2)

2(0/2-0 3 /8+" .)
T.(l- 02/4 + ... ) = wiT. for small 0

This result indicates that the mapping is approximately linear for small Q and
w. For larger frequency values, the nonlinear compression that occurs in the
mapping of Q to 0 is more apparent. This compression causes the transfer
function at the high Q frequencies to be highly distorted when it is translated
to the w-domain.
Prewarping procedure. For some special types of transfer functions, the effect
of this nonlinear compression of Q to 0 values can be compensated. When the
desired magnitude response is piece-wise constant over frequency, this compression can be compensated by introducing a suitable prescaling, or
prewarping to the Q frequency scale. For the bilinear transform, the Q scale is
converted into the Q* scale, where
2
(QT.)
Q*=-tan
T.
2

(8.37)

This scale conversion is shown in Fig. 8.15.

An analog filter design is then performed to determine IHA(jQ*)1 that
satisfies the specifications on the Q* scale. When the bilinear transformation is
applied, the digital filter magnitude response will satisfy the original specifications. This can be seen by substituting Q* into Eq. (8.35): 0 =
2 arctan[Q* T./2] = QT..
For example, if the original specification calls for a passband frequency of
Qp, then the prewarped frequency value, to which the analog filter is designed,
occurs at Q; = (2/T.) tan(Qp/T.). When this prewarped value is employed in
the bilinear transform method, the passband of the resulting digital filter
design is located at wp/T. = Qp, or the original desired value.
Pole mapping procedure. We now derive the rule by which the poles are
mapped from the s plane to the z plane. Let us consider the analog system
function having a pole at s = sP' where sp < 0, given by
(8.38)

Applying the bilinear transform, we get

2 !_Z-l
)
H(z) = 1/ ( - - - - s
T.! + Z-l
p

INFINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

291

n (radls)

IHA(jn)1
lloo.:---"..-+--

l-Et-"''''''''--~

n (radls)
HGURE 1-15
Frequency prewarping procedure. Original specification in n scale is modified by prewarping to
produce the specification in the C scale.
Comparing the pole location in the s plane with that in the z plane, we can
state the pole mapping rule for the bilinear transform as
A pole at s = sp in the S plane gets mapped into ~ zero at z = -1 and a pole
at z = (2 + sp 7;)/(2 - sp 7;) in the z plane when the bilinear transform is
applied.
By examining the magnitude of the pole in the z plane, we find that poles lying
in the left half of the s plane are mapped into the inside of the unit circle.
Hence, a stable analog filter produces a stable digital filter after lhe mapping.
The following examples illustrate the application of the bilinear
trahsform.
Example 8.8. Chebyshev Iowpass filter design using the
trusformation. Consider the previous lowpass specification given by
passband:
-1 < IH(jQ)ldB s 0

for 0 s Q S 1404.1t' rad/s = 4411 rad/s

stopband:
IH(jQ)ldB < -60

for

Q ~ 8268.1t'

and let the sampling period T. = 10- seconds.

4

rad/s = 25 975 rad/s

bilinear

292

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

Prewarped rreqaeacy values. Since we intend to employ the bilinear transform

method, we must prewarp these frequencies. The prewarped values are given by

O~ =

T.2 tan[OpT./2] = 2 x let. tan[O.0702n] = 4484 rad/s

s

and
2
0: =-tan[OsT./2]
= 2 x let tan[0.4134n] = 71690 rad/s
T.
The modified specifications to which the analog filter is to be designed are given
by
passband:

-1 < IH(jO*)ldB s 0

for 0 s 0* s 4484 rad/s

stopband:
for 0* ~ 71 690 rad/s
IH(jO*)ldB < -60
Value of II. The value of I" is determined from the passband ripple:
I" =O.S08.
Value of N. The order of the Chebyshev filter N is determined from
IHdjQ:W = [1 + 1"2C~(O: /O;>r l < 10-6

(-60 dB)

Since OUO~ = 16, we evaluate CN (16) for increasing N until

CN (16) > [(lW - 1)/ ~]112 = 1969.
Doing this, we find C3 (16) = 16301. Hence, N = 3 is sufficient. When using the
impulse-invariance method a value of N = 4 was required. The prewarping of the
frequencies has resulted in a reduction in the filter order.
Axes of the ellipse. The radii of the circles defining the major and minor axes
of the ellipse are determined from the constant defined in Eq. (8.17) to give
p = 4.17. The radii are equal to
R = ~ (4.17 114 +4.17- 114) = SOOl
and
r = ~84 (4.17"4 _ 4.17- 114) = 2216
Since there are three poles, the three lines dividing the s plane have angles equal
to nand 2n/3. The three poles of the Chebyshev filter are then at
Sl = -2216
S2.3

= 2216 cos(2n/3) jSOOl sin(2n/3) = -1108 j4331 = 4470ejl04.4

The analog system function is then given by

R

_
4.43 X 1010
dS) - [s + 2216][s2 + 2223s + (.4470)2]

~oIe mapping. We now apply the pole mapping dictated by Eq. (8.39) to
each pole in the s plane. For s = SI in the S plane, it1 the z plane th~re is a zero at

INFINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

293

z = -1 and a pole at
z =2 + (- 2216 x 10-

4
)

=0 801

2-(-2216x1O 4)
For s =

S2.3,

there are two zeros at z = -1 and poles at

_ 2 + (-1108 j4331) x 10- 4 _

_
%j24.S"
- 2 _ (-1108 j4331) x 10-4 - 0.819 J0.373 - 0.9Oe

The digital filter system function is then given by

H(z) = 4.29 X 10- 3

1+

1 + 2z- ' + Z-2

-1

0~
1 - . 1z

1 - 1.638z

+ 0.81z

The pole/zero pattern, filter block diagram, the log-magnitude response

and the phase response are shown in Fig. 8.16. Only four multipliers and three
delays are needed to implement this filter. This is a reduction over the
requirements for the filter designed with the impulse-in variance method.
1m
(a)

-+__________~__rRe

~3*-____________

(b)

4.29
x(n)

10.3
y(n)

nGURE 8-16
Chebyshev digital lowpass filter results when using the bilinear transformation method. a)
Pole/zero pattern for Hdz). b) Digital filter structure. c) Log-magnitude and phase
responses.

294

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

PASSBAND:
IH(eJO>'T')ldB

-l~~r------------------------

-2-

-3 t-I

-4~O~14~~~n---------------------lO~,~~ n

""T. (radls)
STOPBAND:
IH(ejollT')ldB

""T.

(radls)

Arg[H(ejollTs)]

10,~n

-n

FIGURE 8-1tM:

The bilinear transform for the lowpass filter maps the zeros at g = 00 to
the zeros at OJ = 1C, as illustrated in the previous example. These zeros provide
additional attenuation in the stopband at the high frequencies. In contrast, by
the impulse invariance method these zeros are located at z = 0 and have no
effect on tbe magnitude response. Hence, the bilinear transform is a more
effective, albeit more complicated, procedure for obtaining digital filters from
analog filters.

INFINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

295

In designing highpass digital filters, it is more common to start with an

analog lowpass filter, determine the digital lowpass filter and then employ a
lowpass-to-highpass frequency transformation that then results in a highpass
filter. This and other common transformations are described in the next
section.

8.5 FREQUENCY TRANSFORMATIONS

FOR CONVERTING LOWPASS FILTERS
INTO OTHER TYPES
Once we have designed the digital lowpass filter, we can employ frequency
transformations to produce lowpass filters with different cutoff frequency
values, highpass filters, bandpass filters, or bandreject filters. The transformations that generate these filters are given in Table 8.2. These transformations
are applied by replacing each occurrence of z -1 in the prototype lowpass filter
with the entry in the table.
The lowpass and highpass transformations have analytic forms that are
similar to the bilinear transformation. In fact, the transformations act by
stretching and relocating the frequency points on the unit circle onto itself to
produce the desired filter. The characteristic frequency values in the table
serve as reference points, where the magnitude response has the same value in
both the prototype and transformed filter. Typically, the edge of the passband
is the appropriate characteristic frequency, to assure that the passband
characteristics match.
TABLE 8.1

Frequency transformations to convert.a lowpass filter of characteristic frequency p to a filter having characteristic frequency 8
Z-I_

Lowpass

Z-I+

Highpass

-1
-2

Bandpass

a=

1- az- 1

cosH(fJ + 8)]
a = - cosWfJ - 8)]

+ az- 1

2ak

k-l

-1

-k+i z +k+i

k - 1 -2 2ak -1 1
k+1 z -k+1 z +

(aL is lower characteristic frequency

au is upper characteristic frequency)
-2
Z

Bandreject

sin[HfJ - 8)]
sinB(fJ + 8)]

2a

-f+kz

-1

k = cotB(au - a L )] tan(fJ/2)

1- k

+f+k

1-k -2
2a -1 1
1+kz -1+k z +

a - cosH(au + ad]
- cosH(au - aL)]
k

= tan[Hau -

ad] tan(fJ/2)

296

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

POLE-MAPPING RlJLES. We now consider the pole-mapping rules for the

various transformations. In the discussions below, we consider two elemental

system functions. The first generates a zero at Z = c and a pole at Z = 0 and is
given by

(S.40)
The second elemental system function generates a pole at z
z = 0 and is given by

= d and a zero at

Hp(z)

(S.41)

= 1 _ dz- 1

Both c and d can be complex-valued numbers.

POLE-MAPPING FOR THE LOWPASS-TO-LOWPASS FILTER. Applying the
lowpass-to-Iowpass transformation to Hz(z), we get HLZ(z) given by

Z-1 - 11'

HLZ(z)

= 1- c 1 -az -1

= ( 1 + ca)

1- [(11' + c)/(l +ca)]z-1

1
1

- az

This relationship indicates that the lowpass zero at z

zero at z = Cl, where

(S.42)

= c is transformed into a

a+c
1 + ca

(S.43)

Cl=--

The pole at z = 0 is transformed into a pole at z = 11'. The location of the zero
is related to c, the location of the lowpass zero, while the location of the pole
is not. If we do the same to Hp(z), then we get the transformed system
function that is similar to the inverse of HLZ(z) above, given by
HLP(z)

= (1 + da)(l- [(11' + d)/(l + ad)] z

1)

(S.44)

This relationship indicates that the pole at z = d is transformed into a pole at

z = (11' + d)/(l + ad) and the zero at z = 0 is transformed into a zero at z = 11'.
The location of the pole is related to d, the location of the lowpass pole, while
the location of the zero is not.
If the original lowpass system function contains both relevant poles and
zeros, a cancellation of the singularities that are produced at z = 11' will occur
after the transformation is accomplished.
Since the poles for practical filters are typically close to the unit circle,
the coefficient values determined by these transformations are very sensitive to
numerical round-off. This coefficient error can result in unexpected distortions
in the magnitude response that cause the passband specification not to be
satisfied: To reduce this sensitivity
, in applying the frequency transformations,

INFINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

297

one pole should be transferred at a time. Even though a pole may be complex,
the analytic form for a first-order complex pole is still much simpler than that
for the second-order complex-conjugate pair. After the transformation is
accomplished, the complex-conjugate poles can then be recombined into a
second-order section with real-valued coefficients. This procedure is illustrated
in Example 8.9.
EDDlpie 8.9. Application of frequency
transformations-iowpass-t01lowpass. Consider the transformation of the Chebyshev lowpass digital filter,
given in Example 8.8, into a Chebyshev lowpass digital filter having the passband
specification, given by

for 0 :S OJ :S 0.2808Jr (8 = O. 2808n)

For the prototype lowpass filter, the passband edge is at
(fJ = O.I404n). First, we must evaluate the constant a-:

OJ

= O.I404Jr

a- = sinU(fJ - 8)] = sin( -0.0702Jr) = -0 356

sin[HfJ + 8)] sin(0.2106Jr)
.
The system function of the prototype Chebyshev filter can be written as
1 + Z-I 1 + Z-I 1 + Z-I
Hc(z) = A - - - ;1-----:- - - - ;
1- ZIZ 1- z2z-11- Z3 Z 1
which contains three zeros at Z = -1 and three relevant poles, at Z = ZI, Z = Z2
and Z = Z3' Each of the zeros at Z = -1 is transformed into a zero at

a--1
I-a-

z=--=-1

and each of the corresponding poles at Z = 0 is transformed into a pole at

Z = a-= -0.356
The pole of the prototype lowpass filter at Z = Z; is transformed into a pole
located at
a-+Z;
c=-, 1 + a-z;
and the corresponding zero at z = 0 is transformed into a zero at Z = a- =
-0.3561. These latter zeros cancel the poles at Z = a- that resulted from the zeros
of the prototype lowpass filter.
The pole locations are given by:
for ZI = 0.801, CI = 0.622
for ~ =0.819 + jO.373
C2 = (0.462 + jO.173)/(0.709 - jO.133)
= 0.824e J49 .S
for Z3 = 0.819 - jO.373
The system function of the new lowpass filter is given by
1 + Z-I
1 + 2z- 1 + Z-2
H(z) = 0.029 1 _ 0.622z- 11-1.07z- 1+ 0.674z

298

INTRODUcnON TlY DIGITAL SIGNAL PROCESSING

The pole/zero pattern in the z plane, the filter block diagram and the
log-magnitude and phase responses are shown in Fig. 8.17. For the prototype
lowpass filter, the log-magnitude response fell below -60 dB for w > O. 68.1r (see
Fig. 8.16). For the new filter, this occurs for w>O.83.1r. The stopband
performance of the new filter must be checked, since there is no control over it
when the passband frequency is used as the characteristic frequency in the
transformation. If the stopband is important, the stopband frequency should be
used in the transformation.
POLEMAPPING FOR THE LOWPAS5-TOHIGHPASS FILTER. Applying the
lowpass to bighpass transformation to Hz(z), we get HHZ(Z) given by

Z-l+ a
-1
1 +az
1
) 1 + [(c + a)/(1 + ca)] Z-l
= ( +ca
1 +az 1

HHZ(Z)=1+C

(8.45)

1m

(a)

~------------~------~~--~Re

(b)
x(n)

yen)

FIGURE 8-17
Results of applying lowpass-to-Iowpass frequency transformation to the digital filter shown in Fig.
8.16. a) Pole/zero pattern for H(z). b) Digital filter structure. c) Log-magnitude and phase
responses.

INFINfI'E.IMPULSE RESPONSE F1LTER DESIGN TECHNIQUES

299

PASSBAND:

IH(e.loldB

-2
-3
-4~0--~~~--------~-

0.2808 It

(J)

It

STOPBAND:

IH(el"')ld8
Or-----I....
-20

-40

-ro~----~--------------~---~~0-----~----------~~~1t~+(J)

Arg[H(e""))
It

~-----4-------------~--+(J)
It

FIGURE 8-17c

-It

This relationship indicates that the lowpass zero at z = c is transformed into a

zero at z = -(c + 0')/(1 + cO') and the pole at z = 0 is transformed into a pole
at z = -a. The location of the zero is related to c, the location of the lowpass
zero, while the location of the pole is not. If we do the same to Hp(z), then we
get the transformed system function that is similar to the inverse of HHZ(Z)
above, given by

1+O'z- 1
HP( ) - 1 + dO' (1 + [(d + 0')/(1 + O'd)]z 1)
Z _ _1_

(8.46)

300

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

This relationship indicates that the pole at z = d is transformed into a pole at

z = -(d + a)/(l + ad) and the zero at z = 0 is transformed into a zero at
z = -a. The location of the pole is related to d, the location of the lowpass
pole, while the location of the zero is not.
If the original lowpass system function contains both relevant poles and
zeros, a cancellation of the singularities that are produced at z = -a will occur
after the highpass transformation is accomplished.
Example 8.10. Application of frequency transfonnation&-lowpass-to-highpass.
Consider the transformation of the Chebyshev lowpass digital filter, given in
Example 8.8, into a Chebyshev highpass digital filter having the passband
specification given by

-1 < /H(eJ"')/dBSO

for ;r/2s w S;r (8 =0.5;r)

For the prototype lowpass filter, the passband edge is at w = 0.1404;r (13 =
0.1404;r). First, we must evaluate the constant (1':
(1'

= _ cosH(f3 + 8)] = _
cosH(f3 - 8)]

cos(0.3202;r) = -0.634
cos( -0. 1798;r)

The system function of the prototype Chebyshev filter can be written as

=A l+z-1 l+z- 1 l+z- 1
H.c (Z )
1 -ZIZ II -Z2Z II -Z3Z 1
Hc(z) contains three relevant zeros at z = -1 and 'three relevant poles, at z =
= Z2 and Z = Z3' Each of the zeros at Z = -1 is transformed into a zero at

ZI'

-1+a
z= - - - - = 1
1- (1'

and each of the corresponding poles at Z = 0 are transformed into a pole at

z = - (1' = 0.6337. The pole of the prototype lowpass filter at z = z, is transformed
into a pole located at
z,+(1'
c=--,
1 +az,
and the corresponding zero at z = 0 is transformed into a zero at z = - (1' =
0.6337. These latter zeros cancel the poles at z = -(1' that resulted from the zeros
of the prototype lowpass filter.
The pole locations are given by:
for
for

ZI

= 0.801
C I = -0.339
Z2 = 0.819 + jO.373
C2 = -(0.1848) + jO.373)/(0.4813 - jO.2364)
=0.777e- J90ZO

for

Z3

= 0.819 - j0.373

The system function of the high pass filter is given by

l-z- 1 1-2z- l +z- 2
H(z) = 0.132 1 + 0.339z I 1 + 0.603z 2

INFINITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

301

The pole/zero pattern in the z plane, the filter block diagram and the
log-magnitude and phase responses are shown in Fig. 8.18.

POLEMAPPING FOR TIlE LOWPASSoTO-BANDPASS FILTER. Applying the

lowpass-to-bandpass transformation to Hz(z), we get HBZ(z) given by
Z-2- [2akl(k + 1)JZ-l + [(k -1)/(k + 1)J
HBZ(z) = 1 + C[(k _ 1)/(k + 1)Jz 2 _ [2akl(k + 1)Jz 1 + 1
_k
-

+ 1 + c(k -

1) - 2ak(c + 1)Z-1 + [k - 1 + c(k + 1)Z-2

k+1-2akz 1+(k-1)z 2

_ [k + 1 + c(k -1)](1- CIZ-1)(1- C2Z-1)

(k + 1)(1 - d1z- 1)(1 - d2 z- 1 )

(8.47)

This relationship indicates that the lowpass zero at z

=c is transf6rmed into

1m

(a)

~-------'---+----------~~3Re

(b)
0.132
x(n)

y(n)

FIGURE 8-U
Results of applying lowpass-to-highpass frequency transformation to the digital filter shown in Fig.
8.16. a) Pole/zero pattern for H(z). b) Digital filter structure. c) Log-magnitude and phase
responses.

302

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

PASSBAND:
IH(ei"')ldB

o
-11--------#---"".-.::---

-2
-3

1t

STOPBAND:
IH(e"")ldB

o
-20
-40
-rol--~----~-------

1t
(J)

Arg(H(eJOl))
1t

O~-------~r-----------~

1t

(J)

ftGURE 8-18c

two zeros, at z =
c 1-

and

Cl

and z = C2, where

ak(c + 1) + {a2k2(C + 1)2 - [k + 1 + c(k -1)][k -1 + c(k

k + 1 + c(k - 1)

+ 1)]}112
(8.48)

ak(c + 1) - {a 2k 2(c + 1)2 - [k + 1 + c(k -1)][k -1 + c(k + 1)]}112

c ---~-~~--~-~~------~-~-----~-~2k+l+c(k-1)

303

INFINfI'E.IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

The pole at

=0 is transformed into two poles, at z = d 1 and z =d 2 ,

d _ irk
1-

and

+ [a2k

where

(k + 1)(k - 1)]112
k+1

(8.49)'

12
d _ irk - [a2k - (k + 1)(k -1)r
2k+ 1
2

The locations of the zeros are related to c, the location of the lowpass zero,
while the locations of the poles are not.
If we do the same to the elemental system function Hp(z) then we get the
transformed system function that is similar to the inverse of HBZ(z) above,
given by
(k + 1)(1 - d 1 z- 1 )(1 - d 2 z- 1)
BP(Z) = [k + 1 + c(k - 1)](1 - CI Z 1)(1- C2 Z 1).

H,

The above relationship indicates that the lowpass pole at

into two poles at Z = Cl and Z = C2' where
irk(d
Cl

+ 1) + {a2k 2 (d + 1)2 -

(8.50)

= d is transformed

+ 1 + d(k -1)][k -1 + d(k + 1)]}112

k + 1 + d(k - 1)
,

[k

(8.51)
and
C2

irk(d + 1) - {a2k 2(d + 1)2 - [k + 1 + d(k -1)][k - 1 + d(k

k + 1 + d(k - 1)

The zero at

= 0 is transformed into two zeros at Z = d 1 and Z = d 2 ,

d _ irk
1-

and

+ 1)]}112

+ [ir

2 2

(k + 1)(k - 1)]112
k+1

.
where
(8.52)

d _ irk - [a2k 2 - (k + 1)(k _1)]112

2k+1

The locations of the poles are related to d, the location of the lowpass
pole, while the locati,ons of the zeros are not. If the original lowpass system
function contains both relevant poles and zeros, a cancellation of the
singularities that are produced at z = d 1 and z =d 2 will occur after the
transformation is accomplished.
Eumple 8.11. App6catioD of frequency tnmsformations--lowplISS-to-bandpllSS.
Consider the transformation of the Chebyshev lowpass digital filter, given in
Example 8.8, into a Chebyshev bandpass digital filter having the passband
specification given by
bandpass:
-1 < IH(ei"')ldB sO

for 0.42l.n" s

ClJ

s O.608Jr

304

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

(OL = 0.42l.1r and

Ou = 0.608n). For the prototype lowpass filter, the passband

edge is at w = O.I404n (f3 = O.I404n).
First, we must evaluate the constants a and k:

a = cosU(ou + od] = cos(0.515n) = -0 0486

cosU(ou - od] cos(0.094n)
.
tan(0.070n)
k = cotU( Ou - OL)] tan(f3/2) = tan(0.094n)
~

0.7403

The product ak is equal to -0.0360. The system function of the prototype

Chebyshev filter can be written as
1 + Z-I 1 + Z-I 1 + Z-I
Hc(z) = A -1 ----11
-I 1
-I
- Zl Z
- Z2Z
- Z3Z
Hc(z) contains three relevant zeros at z = -1 and three relevant poles, at z = Zt>
Z = ~ and z = Z3' Each o( the zeros at z = -1 is transformed into two zeros, at
Z = CI and z = C2. where
C 1,2

{ -[k + 1- (k -1)][k -1- (k + 1)]}112

k + 1 _ (k _ 1)
= l

Each of the corresponding poles at z = 0 is transformed into two poles at z = d l

and z = d 2 where
d _ ak [a 2k 2 _ (k + l)(k _1)]112
1,2-

k+ 1

= -0.036 [0.0129 + 0.4519]112 = -0 021

1.7403

0 392

The pole of the prototype lowpass filter at z = Z; is transformed into two poles
located at
C1.1.2

-O.036(z; + 1) {O.013(z. + 1)2 - [1. 740 - 0.260z;][ -0.260 + 1. 74Oz;]} 112

1. 740 - 0.260z;

The zero at z = 0 is transformed into two zeros at z = d l and Z = d 2 , where

d
1.2

= ak [~k2_ (k + l)(k -1)]112

k +1

-002

0392

1 .

These latter zeros cancel the poles that resulted from the zeros of the prototype
lowpass filter.
The pole locations are given by:
for ZI = 0.801

CI.I.2 = -0.0422 jO.859 = 0.86Oej92.8"

for Z2 = 0.819 + j0.373 ,

C2.1 = 0.221- jO.906 = 0.932e- j76.3"

jlO9.I"
C2.2 = -0.305 + jO.883 = 0.934e

for Z3 = 0.819 - j0.373

C3.1 = 0.221 + jO.906 = 0.932ej76.3"

C3.2 = -0.305 - jO.883 = 0.934e-jl09.lo

INFINITEIMPULSE RESPONSE ALTER DESIGN TECHNIQUES

305

The system function of the bandpass filter is given by

3

-2

H(z) = 9.58 X 10- 1- O.441z- tZ+ O.869z-2

1-z~

1-z~

x-----:-----=
1 + O.084z- + O.74z 21 + O.611z
t

+ O.872z

The pole/zero pattern in the z plane, the filter block diagram and the
log-magnitude and phase responses are shown in Fig. 8.19.
1m
(a)

~---------+-------~~Re

(b)
9.58

10.3

x(n)

y(n)

-0.869

-\

0.74

.\

-0.872

.\

nGURE 8-19
Results of applying lowpass-to-bandpass frequency transformation to the digital filter shown in Fig.
8.16. a) Pole/zero pattern for H(z). b) Digital filter structure. c) Log-magnitude and phase

responses.

306

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

PASSBAND:

IH(e"")ldB

o
-1~-----P'---l,,+----

-2
-3

-40

0.4211r

0.608 1r
OJ

STOPBAND:

IH(e"")ldB

o
-20
-40
-W~~------r---1-----~-

-~~O~----~O.7.~~l-1r~O~.608~1r----~-n
OJ

.Arg(H(eJOJ)]

O~-------~~-+----~

-n

FIGURE 8-19c

POLEMAPPING FOR THE LOWPASS-TOBANDRElECf FILTER. Applying

the lowpass-to-bandreject transformation to Hz{z), we get HRZ{z) given by

Z-2 - [2a/(k + 1)]z-1 + [(1- k)/{k + 1)]

HRZ{z) = 1- c [(1- k)Z-2/{k + 1)] - [2a/{k + 1)]z-1 + 1
_ [k + 1 + c{k - 1)]{1 - c1z-1){1 - C2Z-1)
(k + 1){1 - d 1z 1){1 - d 2z 1)

(8.53)

INFINITE-IMPULSE RESPONSE FlLTER DESIGN TECHNIQUES

This relationship indicates that the lowpass zero at

two zeros, at z = Cl and z = C2, where

= c is

307

transformed into

-a(c -1) + {a2(c _1)2 - [k + 1 + c(k -1)][1- k - c(k + 1)]}112

Cl =
k + 1 + c(k - 1)
~

~~
-a(c -1) - {a2(c -1i - [k + 1 + c(k -1)][1- k - c(k + 1)]}112
c 2k + 1 + c(k -1)

The pole at z

= 0 is transformed into two poles, at z = d 1 and z = d 2 ,

d - a+ [a2- (k + 1)(1-k)JII2
1k+1

where
(8.55\
'/

and
d _a-[a2-(k+1)(1-k)j1''2
2k+1

The locations of the zeros are related to c, the location of the lowpass zero,
while the locations of the poles are not.
If we do the same to the elemental system function Hp(z) that generates
a pole at z = d and a zero at z = 0, then we get the transformed system
function that is similar to the inverse of HRZ(z) above, given by
(k + 1)(1- d 1z- 1)(1- d2 z- 1)
HRP(z) = [k + 1 + d(k -1)J(1- CI Z 1)(1- C2 Z 1)

(8.56)

This relationship indicates that the lowpass pole at z = d is transformed into

two poles at z = Cl and z = C2' where
Cl

-a(d -1) + {a2(d _1)2 - [k + 1 + d(k -1)][1- k - d(k + 1)]}112

k + 1 + d(k - 1)

and

(8.57)
-a(d -1) - {a2(d _1)2 - [k + 1 + d(k -1)][1- k - d(k + 1)]}112
C2 =
k + 1 + d(k - 1)

The zero at z = 0 is transformed into two zeros at z

= d 1 and z = d 2 ,

d _a+[a2-(k+1)(1-k)JII2
1k+ 1

where
(8.58)

and
d _ a - [a2- (k + 1)(1- k)JII2
2k+1

The locations of the poles are related to d, the location of the lowpass pole,
while the locations of the zeros are not.
If the original lowpass system function contains both relevant poles and
zeros, a cancellation of the singularities that are produced at z = d 1 and z = d 2
will occur after the transformation is accomplished.

308

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

Example 8.12. Application of frequency transformations-lowpass-tobandreject. Consider the transformation of the Chebyshev lowpass digital filter,
given in Example 8.8, into a Chebyshev bandreject digital filter having the
passband specification given by

for O:s w :s 0.42l.7r and 0.608n :s w :s n.

(aL = 0.421n and au = 0.608n). For the prototype lowpass filter, the passband
edge is at w = 0.1404n ({3 = 0.1404n).
First, we must evaluate the constants a and k:

a = cosH(au + ad] = cos(0.515n) = -0 0486

cosU{au - ad] cos(0.094n)
.
k = tanU(au - ad] tan({3/2) = tan(0.094n) tan(0.070n) = 0.068

- The system function of the prototype Chebyshev filter can be written as

1
1
1
Hc(z)=A 1+z- l+z- l+z1
1 - ZI Z 1_ Z2 Z 11 - Z3 Z 1
Hc(z) contains three relevant zeros at z = -1 and three relevant poles, at z =ZI.
z = ~ and Z = Z3' Each of the zeros at z = -1 is transformed into two zeros, at
z = Cl and Z = C2, where
2a {4~ - [k
1
C .2

+ 1- (k -1)][1- k + k + 1)]}112 a (~_1)112

k + 1- (k -1)

= -0.0486 jO.999 =ej928"

The corresponding pole at z = 0 is transformed into two poles at z = d 1 and
z =d2 , where
d _a[~-(k+l)(I-k)]112
1.2 k +1
-0.0486 [0.0024 + 0.995]112
1.068

-0.0455 0.934

The pole of the prototype lowpass filter at z = z, is transformed into two poles
located at
O.0486(z/ -1) {O.OO24(z; _1)2 - [1.068 - 0.932z;][0.932 -1.068z/W12
Cl.2=

1. 068 - 0.932z;

The corresponding zero at z = 0 is transformed into two zeros at z = d 1 and

z =d2 , where
2
d = a [a - (k + 1)(k _1)]112 -0.0455 0.934
1.2
k +1
These latter zeros cancel the poles that resulted from the zeros of the prototype
lowpass filter. The pole locations are given by:
for
for

ZI

= 0.801

Z2

= 0.819 + jO.373

C1.l.2

= 0.48gei93S
C2.1

= 0.932e- iH19 .2"

C2.2

= 9931ei76.2"

INFINITEIMPULSE RESPONSE FILlER DESIGN lECHNIQUES

for

Z3

= 0.819 - j0.373 = z;

309

0.932eJ109 2"
j76 r
C3,2 = 0.931e- .

C3,1

The system function of the bandpass filter is given by

1-0.lz-1 +z- 2
H(z) = 0.497 1 +O.06z 1+0.239z- 2
1-0.lz- 1+z- 2

1-0.lz- 1 +z- 2
x 1 _ O.444z 1+ 0.867z 21 + O.613z 1 + O.869z 2

The pole/zero pattern in the z plane, the filter block diagram and the
log-magnitude and phase responses are shown in Fig. 8.20.

(a)

~-------------+-------------;-Re

(b)

x(n)

y(n)

FIGURE 8-20

Results of applying lowpass-to-bandreject frequency transformation to the digital filter shown in

Fig. 8.16. a) Pole/zero pattern for H(z). b) Digital filter structure. c) Log-magnitude and phase
responses.

310

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

PASSBAND:
IH(eJOldB

o
-1t----~-t--___jp..L----

-2

-3

4~O------~O~.4~2~1~K~-0~.~OO~8-K------~K

STOPBAND:
IH(eJO IdB
0

-20
40

-60

-80

K
co

Arg!H(eJO)

co

FIGURE 8-ZOc

8.6 ALL-PASS FILTERS FOR PHASE

RESPONSE COMPENSATION
The phase response curves of the IIR filters above are not linear with
frequency. Since the unit-sample response has an infinite duration, a point of
symmetry cannot exist. Since this is a necessary condition for linear phase, IIR
filters cannot have exactly linear phase. In practice, however, it is usually
desirable to have approximately linear phase over the passband. In the

INFINITE-IMPULSE RESPONSE ALTER DESIGN TECHNIQUES

311

stopband, the phase response is less important since these components get
eliminated anyway. In this section, we investigate the use of all-pass filters to
modify the phase response in the passband. These modify the phase by being
connected in cascade with the existing filter. Since these filters have a constant
magnitude response, that of the existing filter is left unchanged.
An all-pass filter is one whose magnitude response is constant for all
frequencies, but whose phase response is not identically zero. We consider
both first and second-order filter sections that can be connected in cascade with
the IIR filters discussed previously in this chapter to modify the overall phase
characteristic.
First-order all-pass filter. The system function of a first-order all-pass filter is
given by
z
1 -r -1-1
HAP(Z) = .
-1
1-rz

for Irl < 1

(8.59)

This system function has a pole at z = r and a zero at z = r- 1. For stability, r

must be less than one. We can easily show that the magnitude response is
constant by computing the squared magnitude response, given by
1
1
H (i"') H (-i"') _ (1- r- e- i "')(1 - r- ei "')
AP e
AP e
- _ (1 _ re i"')(1 - rei"')
cos(w) + r- 2
1- 2r cos(w) + r2

=1- 2r-

r- 2

(8.60)

To compute the phase response, we write the Fourier transform as

(1- r- 1e- iro )(1- rei"')
HAP(e "') = (1- re- i "')(1- rei"')
.

_ 2 - (r + r-1)cos(w) - j(r - r-1)sin(w)

1-2rcos(w)+r2
-

(8.61)

The phase response is given by

Arg(HAP(z)] = arctan [ 2

-(r-r-1)sin(w) ]
(
1) (
-r+r cosw)

(8.62)

The phase responses for different values of r are shown in Fig. 8.21. When
0< r < 1, the zero lies on the positive real axis. The phase over 0 < OJ < n is
positive, at OJ = 0 it is equal to n and decreases until w = n, where it is zero.
When -1 < r < 0, the zero lies on the negative real axis. The phase over
0< OJ < n is negative, starting at 0 for OJ = 0 and decreasing to -n at OJ = n.
Since the zero is outside the unit circle, the phase will always decrease by n as
OJ goes from 0 to n. The shape of the phase response is related to the radius r.
Second-order all-pass filter. Second-order all-pass filters are obtained by
allowing r to take on complex values and including the complex-conjugate

312

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

1m

+-------r-----*+-9----Re
lIr

PHASE.

FlGURE 8.21
Pole/zero pattern and phase response for
first-order all-pass filter.

singularity locations. Doing this we have the second-order all-pass filter system
function given by
(1 - r- 1z- 1)(1- r-hz- 1)
HAP(Z) = (1 _ rz- 1 )(1- r*z-l)
1 - 2 Re(r-1)z-1 + Ir- 1 12 Z-2
1- 2 Re(r)z 1 + Irl2 Z-2

(8.63)

The phase responses for different values of r are shown in Fig. 8.22. Since
there are two zeros outside the unit circle, the phase at (J) = 0 is 2.1r, or
equivalently, equal to zero, when r > O. When r < 0, the phase at (J) = 0 is zero
again. Since the two zeros lie outside the unit-circle, the phase will decrease by
2.1r as CIJ goes from 0 to .1r.
The phase response of a given filter can be tailored in a narrow band of
frequencies by varying the value of r and including additional all-pass filters.
However, the type of tailoring that can be done is limited, since the phase
responses of all-pass filters all show a negative trend with frequency.

8.7 SUMMARY
The Butterworth, Chebyshev and elliptic filter synthesis techniques for
designing analog lowpass filters in the s plane have been presented. For a given
lowpass specification, it was found that the Butterworth filter was the simplest,
but required the highest order. The elliptic filter provided the lowest order, but
was the most difficult to design analytically. The Chebyshev filter prov(des a
reasonable tradeoff between filter order (hardware cost) versus design effort
(design cost).

INFINITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

313

1m

PHASE:

7t

-71

FIGURE 8.l2
Pole/zero pattern and phase response for
second-order all-pass filter.

The impulse-invariance and the bilinear transformation methods were

described for translating the s plane singularities into the z plane. The digital
filter could then be implemented from these z plane singularities as described
in the previous chapters. For the lowpass filter specification, the bilinear
transform provided better stopband results because the zeros were mapped to
z = -1, rather than z = 0 with the impulse-invariance method.
Frequency transformations were presented to convert the prototype
lowpass digital filter designs into highpass, bandpass and bandreject digital
filters. All-pass filters were used to compensate the phase response of IIR
filters. An exactly linear phase is impossible to obtain with a filter having an
infinite duration unit-sample response. In practice, an approximately linear
phase response can be achieved over the passband of the original filter by
including all-pass filters.

FURmER READING
Antoniou, Andreas: Digital Filters: Analysis and Design, McGraw-Hill, New York, 1979.
Gold, Bernard, and Charles M. Rader: Digital Processing of Signals, McGraw-Hill, New York,
1969.
Hamming, R. W.: Digital Filters, 2nd ed., Prentice-Hall, Englewood Cliffs, NJ, 1983.
Oppenheim, Alan V., and Ronald W. Schafer: Digital Signal Processing. Prentice-Hall,
Englewood Cliffs, NJ, 1975.
Rabiner, Lawrence R., and Bernard Gold: Theory and Application of Digital Signal Processing,
Prentice-Hall, Englewood Cliffs, NJ, 1975.

314

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

REFERENCE TO TOPIC
Elliptic filter reference

Rader, C. M., and B. Gold: "Digital filter design techniques

IEEE, 55: 149-171 (1967).

i~

the frequency domain," Proc.

PROBLEMS
8.1. In replacing an analog filter with a digital filter, aliasing effects must be eliminated
by including an anti-aliasing analog lowpass filter prior to the analog-to-digital
converter. The Butterworth and Chebyshev filters are suitable candidates in
practice. A possible specification for an anti-aliasing filter is to have a negligible
attenuation (between 0 and 0.1 dB) in the passband and to have a large
attenuation (greater than 60 dB) at the analog frequency corresponding to the near
edge of the passband of the next period of IH(eiw1T,)I, that is, the replica of the
magnitude response centered at w = 2n/T..
Consider the analog lowpass filter specification that is to be satisfied by the
digital filter to be given by
passband:
for O:s Q:s 10 rad/s
stopband:
IHUQ)ldB < -60

for Q 2= 50 rad/s

Let the sampling period T. = 10- 2 s.

(a) Sketch the digital filter magnitude response specification over the frequency
range -2n/T.:s w/T.:s 2Jr/T..
(b) Determine the order of the Butterworth analog filter that acts as a suitable
anti-aliasing filter.
(c) Determine the order of the Chebyshev analog filter that acts as a suitable
anti-aliasing filter.
8.2. Analog filter design. Design the Butterworth analog lowpass filter that satisfies the
analog specification given in Table Proj8.1. Use one set of the frequency values
given in the table. Show the steps in the design procedure. Sketch the pole/zero
pattern in the s plane and give the system function HA(S).
Digital filter design. Design a digital lowpass filter to replace the analog filter
above. Let the sampling period be T. = 10- 4 seconds.

Impulse invariance method. Determine the pole/zero pattern obtained by

performing the impulse-invariance mapping. Write the digital filter system
function H(z) and sketch the block diagram of the filter.
Bilinear transform method. Determine the pole/zero pattern obtained by
performing the bilinear mapping. Write the digital filter system function H(z) and
sketch the block diagram of the filter.
8.3. Repeat the steps in Problem 8.2 with a Chebyshev filter design.

INFINITEIMPULSE RESPONSE FILTER DESIGN TEOINIQUES

315

8.4. Consider the analog lowpass filter specification given by

passband:

-1 < IH(jO)ldB s 0

for Os Os 10rad/s

stopband:

IH(jO)ldB < -60

Let the sampling period

for 0

~ 50 rad/s

T. = 10- s.
2

(a) Determine the order of the Butterworth filter that satisfies this specification
and draw the s plane pole/zero pattern. Transfer the poles to the z plane using
the impulse invariance method and implement the digital filter.
(b) Determine the order of the Chebyshev filter that satisfies this specification and
draw the s plane pole/zero pattern. Transfer the poles to the z plane using the
impulse invariance method and implement the digital filter.
(c) Compare the order of the two filters.
8.5. Bilinear-transform method. Consider the analog lowpass filter specification given
by
passband:
for 0 sO s 10 rad/s
stopband:

IH(jO)ldB < -60

Let the sampling period

for 0

~ 50 rad/s

T. =10- s.
2

(a) Determine the specification in terms of the prewarped frequenc.-y values.

(b) Determine the order of the Butterworth filter that satisfies this specification
and draw the s plane pole/zero pattern.
(c) Determine the order of the Chebyshev filter that satisfies this specification and
draw the s plane pole/zero pattern.

COMPUTER PROJECTS
8.1. Comparison of impuJse..invariance and bilinear transform methods for a Butterworth lowpass IDter
Object. Design a Butterworth analog lowpass filter and compare the magnitude
and phase responses of the digital filters produced by using the impulse-invariance
and bilinear transform methods.

Analytic results
Analog filter design. Design the Butterworth analog lowpass filter that
satisfies the analog specification given in Table Proj8.1. Use one set of the
frequency values given in the table. Show the steps in the design procedure.
Sketch the pole/zero pattern in the s plane and give the system function HA(S).
Digital filter design. Design a digitallowpass filter to replace the analog filter
above. Let the sampling period be T. = 10- 4 s.
Impulse invariance method. Determine the pole/zero pattern obtained by
performing the impulse-invariance mapping. Write the digital filter system
function H(z) and sketch the block diagram of the filter.

316

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

TABLE Proj8.1

Frequency values for prototype analog lowpass specification (values are

given in radians/serond)
AIIaIog IowplSS spedficatioa

Possible values:

01
1094Jr
l094n
l094n
l094n
l094n

01

Az

5938n'

1250n
1250n
1250n
1250n
1250n

5938n

6094n
6250n
6406Jt

6563n

6094n
6250n
6406Jt
6563n

01
1406n
l406Jt
l406Jt
1406n
1406n

0%

01

liz

5938Jr

1563n
1563n
1563n
1563n
1563n

6094n
6250n

6094n
6250n
6406Jt
6563n

6406Jt
6563n
671fJn

01
1719x
1719x
1719x
1719x
1719x

liz
6094"
625On:
6406x

6563"
671fJx

Bilinear transform method. Determine the pole/zero pattern obtained by

performing the bilinear mapping. Write the digital filter system function H(z)
and sketch the block diagram of the filter.
Computer results
Implement both filters and plot the following
(a) the unit-sample response.
(b) the log-magnitude responses using PLOTI..M with DBVAL= -4 and -60.
The specification limits of -1 and -50 dB should be shown.
(c) the phase response.

Comparison
Compare the stopband performance of the two methods. Explain the phase
response in terms of the pole/zero pattern in the z plane.
8.2. Comparison of bnpnlse-invariance and bilinear transform methods for a Chebyshev lowpass ruter.
Object. Design a Chebyshev analog lowpass filter and compare the magnitu<a", and
phase responses of the digital filters produced by using the impulse:;invariance and
.'
bilinear transform methods.
Perform the steps listed in Project 8.1 for the Chebyshev filter design.
IHAUQ)ldB

_~~~Z'mlm

INFINITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

317

8.3. Butterworth DR filter design

Objed. Design an analog Butterworth filter that satisfies the lowpass specification;
use the bilinear transform to obtain the digital filter; and use frequency
transformations to obtain me digital filter to meet a bandpass, highpass or
bandreject specification.
Do the design by performing the following steps (let T. = 10- 4 s):

1. Design the analog Butterworth filter using the values for the analog lowpass
filter found in Table Proj8.1. These values have not been prewarped. Show the
an8.Iog Butterworth lowpass filter design procedure. Include sketches showing
the frequency prewarping transformation.
2. Design the digital lowpass filter by performing a bilinear transformation. For
the digital lowpass filter determine the system function obtained by the bilinear
transform and sketch the pole/zero pattern. Sketch the block diagram and plot
the unit-sample response. Plot the log-magnitude response with DVBAL = -4
and -60 with the specification limits shown. Plot the phase response.
3. Design the digital filter that satisfies one of the highpass, bandpass or band
reject specifications given in Appendix B. Use the frequency transformations
with the edge of the passband as the characteristic frequency. Sketch the
pole/zero pattern, and the block diagram. Plot the log-magnitude response of
the digital bandpass filter with DBVAL = -4 and -60, andwith the specification limits indicated. Also plot the phase response.
(a) If your stopband specification is not met after performing the frequency
transformation, satisfy the specification by adding zeros to the filter system
function to achieve additional attenuation.
(b) If your specification is met after performing the frequency transformation,
add zeros so that either of the following is achieved:
(1) a reduction in the transition band by a factor of 2, or
(2) a reduction in the stopband attenuation by' 6 dB.
Provide the following for the final filter: the system function, the filter block
diagram and the z plane pole/zero pattern.
8.4. Chebyshev DR filter design
Objed. Design an analog Chebyshev filter that satisfies the lowpass specification;
use the bilinear transform to obtain the digital filter; and use frequency
transformations to obtain the digital filter to meet a bandpass, highpass or
bandreject specification.
Perform the steps listed in Project 8.3 for a Chebyshev filter. Numerical
round-off errors in your computations may shift the implemented pole position so
that the specification is not satisfied. If this occurs, manually shift the pole
positions in your final filter until the specification is satisfied. Do this interactively
by examining the magnitude response after a modification is made. List the steps
that you took.

CHAPTER

9
FINITE-IMPULSE
RESPONSE
FILTER
DESIGN
TECHNIQUES

9.1 INfRODUCfION
In this chapter, we describe procedures for implementing finite-impulse
response (FIR) 1,ilters. In contrast to infinite-impulse response (IIR) filters, the
FIR filter is characterized by a unit-sample response that has a finite duration.
Understanding the performance of FIR filters is important because the FIR
filter is the discrete-time-domain equivalent to processing data in the frequency
domain with the fast Fourier transform (FFf) algorithm.
FIR filters have two advantages over IIR filters. First, the design
procedures for IIR filters are well established only for magnitude responses
that are piece-wise constant over frequency, such as lowpass, highpass,
bandpass and bandreject filters. In contrast, the magnitude response of an FIR
filter can be designed to have an arbitrary shape. Hence, if the magnitude
response specification is other than piece-wise constant, the FIR filter is the
filter of choice. The second advantage of FIR filters is that they can be
designed to have a phase response that is exactly linear with frequency. In
Chapter 6, it was shown that a special FIR linear-phase filter structure can be
used to reduce the number of multiplications by approximately one-half. The
disadvantage of FIR filters is that, if an IIR filter can be found that satisfies the
same specification as an FIR filter, the IIR filter will typically be faster and
require less memory.
318

FlNITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

319

Three FIR design procedures are described in this chapter. The one to be
used in a particular case depends on whether the desired unit-sample response
{hD(n)} is known and whether it has a finite duration. If {hD(n)} is known and
has a finite duration, the FIR design is straightforward and can be implemented directly by using the structures considered in Chapter 6. If the {hD(n)}
is known, but has an infinite duration, the response must be truncated to form
the finite-duration response {hN(n)} having N samples, where N is a finite
integer. To do this, the windowing method for designing FIR filters is
described. It employs an N-point tapered window sequence to perform the
truncation. When the form of the unit-sample response is not known, {hN(n)}
must be det6rmined from the desired magnitude specification IHo(ei"')1 and an
appropriate phase response. With the DFT method, {hN(n)} is determined
from the inverse discrete Fourier transform of N samples of the desired
magnitude response {IHo(k)l} and an appropriate linear phase response. The
windowing and DFT methods are combined to find the filter having the
shortest-duration unit-sample response that still satisfies the specification. The
third FIR filter design procedure is the frequency-sampling method. It
produces an interesting filter structure by using {IHo(k)l} and an appropriate
phase response. With this third method any continuous magnitude response
having an arbitrary shape can be achieved. The resulting filter structure
contains a parallel combination of second-order resonators, which allows the
processing to be performed in parallel. This structure is most useful for
applications in which processing speed must be reduced to a minimum. The
chapter concludes with a discussion on how to use the FFT algorithm to
perform filtering tasks on long data sequences. A comparison of multiplication
counts for performing the filtering operation with a digital filter and with the
FFT algorithm indicates when one is to be preferred over the other.

9.2 DESIGNING FIR FILTERS WITH mE

WINDOWING METHOD
As described in Chapter 5, the system function of an FIR filter can be written
as
Np

HFIR(Z) =

bkz- k

(9.1)

n=-NF

Using the nonrecursive structure, this filter can be implemented directly from
the coefficient values bk , for -NFsksNp Further, the unit-sample response
of the FIR filter is equal to this coefficient sequence, or
h(n) = b n

(9.2)

Hence, the FIR filter can be implemented directly from the unit-sample
response. We first consider implementing the FIR filter to approximate a
desired unit-sample response sequence. Then we consider the case of requiring
the FIR filter to satisfy a desired magnitude specification.

320

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

To take advantage of the efficient nonrecursive structure for linear-phase

FIR filters, described in Chapter 6, we will design FIR filters to a have a
linear-phase response. Let {ho(n)} denote the desired unit-sample response
sequence of a linear-phase filter. Further, let {ho(n)} be a finite-duration
symmetric sequence extending over - NF S n s Np , where NF = Np by symmetry. This desired sequence can be achieved exactly by an FIR filter by
setting the coefficients equal to ho(n), or
(9.3)

A problem arises for filters, such as the ideal lowpass filter, which have a
unit-sample response {ho(n)} that has an infinite duration. The sequence
{ho(n)} has an infinite duration whenever the deSlfed magnitude response
IH(eiw)l, or any of its derivatives exhibit a discontinuity. For such cases, we
must truncate the infinite-duration {ho(n)} to produce one that has a finite
duration of N samples, denoted by {hN(n)}.
As shown in Chapter 3, to have a linear-phase response, the unit-sample
response must be symmetric or antisymmetric about some point in time. To
simplify the analysis below, we consider {ho(n)} to be symmetric about n = 0,
or
(9.4)
for all n
To retain the linear-phase response in the FIR filter having the response
{hN(n)} , the truncation must be performed symmetrically around n = 0 to
maintain the symmetry in {hN(n)}, as shown in Fig. 9.1. The N-point
{ho(n)}

,.,

for N=7

I I

II
I 1111I1111
-2 -1 0 1 2

I I

o.

{h,{n))

.. 1111111
-3 -2 -\ 0 1 2 3 4

FlGURE 9-1
Comparison of desired unit-sample response {ho(n)} having an infinite duration and the 7-point
finite-duration approximation {h7(n)}.

FINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

321

unit-sample response, where N is an odd-valued integer, is then given by

for -(N - 1)/2:S. n :S. (N -1)/2
otherwise

(9.5)

If a causal linear-phase filter is desired, {hN(n)} can be delayed by (N -1)/2,

to start at n = O. In the truncation above, the value of N is taken to be an
odd-valued integer. In this case, the point of symmetry falls on a sample point.
When the point of symmetry in {ho(n)} falls between two sample points, N
must be an even-valued integer.
To analyze the truncation process, we model it as a multiplication of the
desired sequence by a finite-duration window sequence, denoted by {w(n)}.
The general truncation of an even sequence can be written as
for all n

(9.6)

where
w(n)

= w(-n):#O

for

=0

for

Inl :S. (N - 1)/2

Inl > (N -1)/2

(9.7)

The sequence {w(n)} has a duration of N samples and is symmetric about

n = 0 to maintain the symmetry that is present in {ho(n)}. The goal of the
windowing synthesis approach is, given the desired filter unit-sample response
{ho(n)} , to determine the values of {w(n)}, such that the finite-duration
{hN(n)} results in the FIR filter that satisfies the desired magnitude response
specification. Since the number of multiplications in the FIR filter increases
with N, the window having the smallest duration is desired.
It was shown in Chapter 3 that the multiplication of two discrete-time
sequences corresponds to a convolution of their Fourier transforms, or spectra.
If Ho(ei "') is the spectrum of {ho(n)}, and W(e i"') is the spectrum of {w(n)},
then the transfer function of the FIR filter HN(e 1"') is given by
HN(ei ",)

= Ho(ei "') @ W(e i ",)

(9.8)

Since the two functions in the integral are periodic, a circular convolution
results and the limits are taken over one period. Trying to evaluate this
convolution in general is difficult because the spectra are complex-valued and
tend to have complicated analytic forms involving sine and cosine functions.
However, if all the discrete-time sequences are symmetric about n = 0, then all
the spectra are real-valued functions of w. This property helps us to visualize
the effects of convolution in the following discussion.
To illustrate the effects of this convolution, let us consider the problem of
approximating an ideal lowpass filter by using the windowing method. For the

322

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

I
IT

IT

{horn)

~TT"T"",.L.JL...L.."""T""T...,J...LJ.-f.J....LJ4-..,...~L1....L.o'T'"lr"T""1_""'

n
FlGURE 9-2
Ideallowpass filter magnitude response IHD(eJ"')1 and the corresponding infinite-duration desired
unit-sample response {hD(n)}. Case for (lJc = lr/4 is shown.

cutoff frequency specified as

We'

the desired transfer function is equal to

for -weswswe
for -nS W < -we and coe < W <no

(9.9)

This is shown in Fig. 9.2 for We = n/4. The unit-sample response of the ideal
lowpass filter is given by the inverse Fourier transform, or

(9.10)
The desired unit-sample response is an even sequence that extends to both plus
and minus infinity. We will consider several different finite-duration approximations to this ideal response by employing several common window
sequences: the rectangular, triangular, and cosine windows.
TIlE RECfANGULAR WINDOW. The rectangular window sequence is defined
by
for -(N - 1)12 S n :S (N - 1)/2
(9.11)
~(n) = {~ otherwise

FlNITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

323

An example is shown in Fig. 9.3 for N = 31. The infinite-duration unit-sample

response of the ideal lowpass filter is approximated by a finite-duration
sequence containing N elements by multiplying it with the rectangular window,
or
hN(n)

= wR(n) hD(n)
=

sin ( wcn)

for all n
for -(N -1)/2::;; n::;; (N -1)/2

(9.12)

The real-valued transfer function HN(e]OJ) of the FIR approximation is shown

in Fig. 9.4. The shape of HN(e jOJ ) can be explained in terms of the convolution
(a)

(w.(n)} (N=31)

..- 11111"
11111111111!1111"
111111.
..n
15
0
15
(b)

1t
Q)

-lOO~----~----~'1-2-----L----~1t
Q)

FIGURE 9-]
Rectangular data window (N = 31).
(a) Window sequence; (b) spectrum;
(e) log-magnitude spectrum.

324

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

0 .....- -.....

-50

-100~----~-----U~----~----~
OJ

FIGURE 9-4
Results of applying a 31-point rectangular window to the ideallowpass filter unit-sample response.
(a) Finite-duration unit-sample response; (b) magnitude response of approximation to ideal
lowpass filter; (c) log-magnitude response.

of WR(e jCLl ) and HD(e jCLl ) by examining the features of WR(e jCLl ) that are
responsible for the deviations from the ideal response. The spectrum of the
rectangular window is equal to
WR(ejCLl )

(N-l)/2

n=-(N-l)12

N-l

e- jCLln = eJCLl (N-l)12 L e- jwn

n=O

(9.13)

RNITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

325

The finite geometric sum formula can be applied to evaluate the sum, to give
1- e- icoN
W, (e lW ) = elW (N-I)/2
.
(9.14)
R
1-e- lco
To simplify this analytic form, we express the numerator and denominator as
sine functions, or

sin(Nw)
sin(w/2)

(9.15)

The spectrum WR(eieo) for N = 31 is shown in Fig. 9.3. The spectrum WR(e lCO )
has two features that are important for this application, the width of the main
lobe and the sidelobe amplitude. The main-lobe width is defined as the
distance between the two points closest to w = 0 wh'ere WR ( eieo ) is zero. For
the rectangular window, the main-lobe width is equal to 4,,/ N. The maximum
side lobe magnitude for WR(eieo ) occurs for the first sidelobe and is equal to
approximately 22% of the main-lobe amplitude, or -13 dB relative to the
maximum value at w = O.
We now consider the effect of these two features in the convolution to
produce the lowpass FIR filter transfer function HN(elW ). The transfer function
HN(eiW ) and the log-magnitude response IHN(eieo)ldB are shown in Fig. 9.4 for
the 31-point rectangular window. The filter response differs from the desired
response in several ways. The sharp transition in the ideal response at w = We
has been converted into a gradual transition. In the passband, a series of
overshoots and undershoots occur, and in the stopband, where the desired
response is zero, the FIR filter has a nonzero response, called leakage. These
effects can be explained in terms of the features of the window spectrum.
The main lobe of WR(eieo ) causes the smearing of the desired transfer
function features, especially noticeable at frequencies where the desired
magnitude response is discontinuous. The discontinuity in HD(eieo ) is converted
into a gradual transition in HN(eieo ), whose width is related to the width of the
main lobe of WR(e ieo ). Since the main-lobe width of WR(e ieo ) is equal to 4Jr/N,
the size of this transition region can be reduced to any desired size by
increasing the size of the window sequence. The forms of HN(eieo ) for different
values of N are shown in Figs. 9.5 and 9.6. However, this increase in N also
increases the number of computations necessary to implement the FIR filter.
The sidelobes of WR(e leo ) cause a different distortion to the filter
magnitude response. Since the side lobes of WR{e ieo ) extend over a wide
frequency range, large amplitude components in HD(e lW ) become smeared
over a wide range of frequencies in HN{eico ). This effect is most evident in the
frequency ranges in which the HD(e ieo ) is either constant (in the passband) or
has a small magnitude (in the stopband). In the passband, these sidelobe
effects appear both as overshoots apd undershoots to the desired response. In

326
(a)

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

{hlS(n)}

(b)

It
OJ

-50

-100~0------~------~----~------~1t
OJ

FIGURE '-S
Results of applying a 1S-point rectangular window to the ideal lowpass filter
unit-sample response. (a) Finiteduration unit-sample response; (b)
magnitude response of approximation
to ideal lowpass filter; (c) logmagnitude response.

the stopband, these effects appear as a nonzero response. Unfortunately, and

unexpectedly, these sidelobe effects do not diminish significantly, but remain
almost constant, as the duration of the rectangular window is increased, as
shown in Figs. 9.4 through 9.6. These results indicate that no matter how many
elements of the {hD(n)} are included in {hN(n)} , the magnitudes of the
overshoot and leakage will not change significantly when the rectangular
window is used. This result is known as the Gibbs phenomenon, after the
American mathematician Josiah Willard Gibbs of Yale, who first noted this
effect.

FINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

327

(a)

-30

30
n

(b)

H", (e"")

L-

(c)

IH", (e"") IdB

OF---",

-50

-100~O------~---L--L-----~------~n

FIGURE 9-6

Results of applying a 61-point rectangular window to the ideal lowpass filter

unit-sample response. (a) Finiteduration unit-sample response; (b) magnitude response of approximation to
ideal lowpass filter; (c) log-magnitude
response.

To reduce these sidelobe effects, we must consider alternate window

sequences having spectra exhibiting smaller sidelobes. TQ indicate how the
form of the window should change, we observe that the sidelobes of the
window spectrum W(eioo ) represent the contribution of the high-frequency
components in the window sequence. For the rectangular window, these
high-frequency components are due to the sharp transitions from zero to one
at the edges of the window sequence. Hence, the amplitudes of these
high-frequency components, the sidelobe levels, can be reduced by replacing

328

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

these sharp transitions by more gradual ones. This is the motivation for
considering the triangular, or Bartlett, window described next.
THE TRIANGULAR, OR BARTLETT, WINDOW. Let us consider the effect of

tapering the rectangular window sequence linearly from the middle to the
ends. We then obtain the.N-point triangular window given by
WT(n)

= 1 -21nl
-N-1

for -(N - 1)/2 $ n $ (N - 1)/2

(9.16)

This window and its spectrum WT(e ieo ) are shown in Fig. 9.7. As expected, the
(a)

..," II1I111111111111!1111111ItI...
- 15

15

(b)

(c)

-lOO~~

__~__~__+-~__~__~-;

01
w

FIGURE 9-7
Triangular data window results
(N = 31). (a) Window sequence;
(b) spectrum; (c) log-magnitude
spectrum.

FINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

329

sidelobe level is smaller than that of the rectangular window, being reduced
from -13 to -25 dB relative to the' maximum_ However, the main-lobe width
is now 87'/N, or twice that of the rectangular window having the same
duration. This result illustrates that there is a tradeoff between main-lobe
width and sidelabe level.
The FIR lowpass magnitude response obtained by applying the triangular
window to the ideal lowpass filter unit-sample response is shown in Fig. 9.8.
Comparing it with that obtained by using the rectangular window, we find that
the triangular window produces a smoother magnitude response. The transition from passband to stopband is not as steep as that for the rectangular
(a)

-15

15

o~---+~===F====~--~

IT

IH3,(e'''lidB

(e)

01----.....

-50

-100L-----~------~------~----~

'/4

II

FIGURE 9-8
Results of applying a 31-point triangular window to the ideal lowpass
filter unit-sample response. (a)
Fmite-duration unit-sample response;
(b) magnitude response of approximation to ideal lowpass filter;
(c) log-magnitude response,

330

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

window. In the stopband, the response is smoother, but the attenuation is less
than that produced by the rectangular window. Because of these characteristics, the triangular window is not usually a good choice.
We next consider a class of raised-cosine windows that, compared with
the triangular window, are smoother at the ends, but c1os~r to one at the
middle. The smoother taper at the ends should reduce the sidelobe level, while
the broader middle section produces less distortion of {hD(n)} around n = 0.
RAISED-COSINE WINDOWS. To reduce the sidelobe level further, we can
consider an even more gradual taper at the ends of the window sequence by
using the raised-cosine sequence. We now discuss three such windows. The
first is the Hanning, or simply raised-cosine window, denoted by {wdn)} ,
given by
for -(N - 1)/2 s n S (N - 1)/2
+ 0.5 cos(21rn/(N -1
We n otherwise
o
(9.17)

( )_{o.s

This window and its spectrum WdeiO) are shown in Fig. 9.9 for N = 31. The
magnitude of the first sidelobe level is -31 dB relative to the maximum value,
an improvement of 6 dB over that for the triangular window. Since the
main-lobe width is the same, the Hanning window is preferable to the
triangular.
This reduction in the sidelobe level can be explained by considering
{wdn)} itself to be an infinite-duration raised-cosine sequence that is
multiplied by the rectangular window, or
wc(n) = [0.5 -+: 0.5 cos(N~nl) ]~(n)
=c(n)wR(n)

for all n
for all n,

(9.18)

where {c( n)} is the infinite-duration raised-cosine sequence as shown in Fig.

9.10(a). This multiplication in discrete-time is equivalent to the circular
convolution of the two Fourier transforms. Since {c(n)} is composed of a
constant term and a cosine term at W = 21r/(N -1), its Fourier transform can
be expressed in terms of three Dirac delta functions, given by
C(eiO) = 1rb(w)

+O.S1r b(W -~)

+O.S1rb(W +~)
N-l
N-l
for -1r S

W S

1r (9.19)

When convolving WR(eiO) with C(eiO), versions of WR(eiO) are shifted to the
locations of these impulses, the shifted spectra are scaled by the weight of the
impulse, and the results are added to produce the final curve, as shown in Fig.
9.10(b). The shifts caused by the impulses from the cosine function locate the
sidelobes of WR(eiO) so that they tend to cancel, thus reducing the sidelobe
magnitudes of the resulting spectrum Wc(eiO).

FINITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

331

(a)

-15

15

nGun
w

9-9
Raised-cosine (Hanning) window results
(N = 31), (a) Window sequence; (b)
spectrum; (c) log-magnitude spectrum.

The discrete-time and frequency apVfOximations to the ideal lowpass

filter produced by applying the Hanning window are shown in Fig. 9.11. Most
notable is the improved stopband attenuation characteristic. The largest peak
is approximately -44 dB relative to the passband level. At higher frequencies
the stopband attenuation is even greater.
Hamming noted that a reduction in the first sidelobe level can be
achieved by adding a small constant value to the cosine window. This produces

332

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

-15

15

{c(n)}={1 +COSCnI15)}

-15

15

{wc(n)}

+ n

.LLLI.&U.II....._ _ _

-15

15

FIGURE 9-10
Interpreting the raised-cosine
(Hanning) window results. (a)
Time domain; (b) frequency
domain.

the Hamming window, denoted by {WJI(n)}, and given by

WJI(n)

= {0054 + 0.46 cos(2Jr:nl(N -

1))

for -(N - 1)/2:$ n :$ (N - 1)/2

otherwise
(9.20)

This reduction in the first sidelobe level can be explained in terms of the
addition of the shifted and scaled window spectra WR(ei "'), shown in Fig.
9.10(b). For the raised-cosine window, the Dirac delta functions at co = 0 and
at co = n/15 formed a ratio of two to one. This ratio resulted in a negative
amplitude for the first side lobe in We( ei "'). By changing this ratio of delta
functions to become 2.35 to one (0.54 to 0.46/2), the magnitude of the first
sidelobe in We(ei "') can be reduced. The Hamming window and its spectrum

FINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

333

(b)

0("'''')
It

"1

-"1

"1

I I

-l~

N-I

N-I

10)
"1

FIGURE 9-IOb

WH(eJ "') are shown in Fig. 9.12 for N = 31. The magnitude of the first sidelobe
has been reduced to -41 dB, a reduction of 10 dB relative to the Hanning
window. But this reduction is achieved at the expense of the sidelobe
magnitudes at the higher frequencies, which are almost constant with
frequency. With the Hanning window, the sidelobe amplitudes decrease with
frequency.
The lowpass filter magnitude and log-magnitude responses when using
the Hamming window are shown in Fig. 9.13. It is noted that the first sidelobe
peak is -51 dB, -7 dB with respect to the Hanning window filter. However, as
the frequency increases, the stopband attenuation does not increase as much as
with the filter produced by the Hanning window.
The stopband attenuation in the lowpass filter magnitude response is

334

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

(a)

-15

(e)

!H31 (e"')!

0..--_

-4

ncURE 9-11
Results of applying a 31-point
raised-cosioe (Hanning) window
to the ideal lowpass filter unitsample response. (a) Finiteduration unit-sample response;
(b) magnitude response of approximation to ideal lowpass filter; (c) log-magnitude response.

limited by the sidelobe level of the window function. Even though the
Hamming window achieved an attenuation of 51 dB for our lowpass filter, it
may not be sufficiently small for some applications. A window having smaller
magnitude sidelobes is needed. Hamming reduced the sidelobe magnitude
while maintaining the main-lobe width, equal to 4:tr/N. The following window
of the main-lobe width.
reduces the sidelobe level at the expense
,

FINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

335

(a)

""111111111111111111111111111""
.n
0
-~

IT

(c)

-100L-__

~~

____- L____

____~

IT

co

ftGURE 9-12
Hanning window results (N = 31).
(a)
Window sequence; (b)
spectrum; (c) log-magnitude
spectrum.

Blackman noted that the sidelobes can be reduced further by including an

additional cosine term in the window sequence _The Blackman window is given
by

Wa(n) = 0.42 + 0_5 cos(27rn/(N - 1 + 0.08 cos(41rn/(N - 1

for -(N - 1)/2:5 n :5 (N - 1)/2
=0
otherwise.

(9.21)

336

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

(a)

15

- 15

o~----~~~--+-----~------~

1T

I H 31 (e IW ) IdB

(c)

0,---_.

-50

-l00L-----~------~----~~--~u

It

FIGURE 9-13
Results of applying a 31-point Hamming
window to the ideal lowpass filter unitsample response. (a) Finite-duration
unit-sample response; (b) magnitude
response of approximation to ideal lowpass filter; (c) log-magnitude response.

This window and its spectrum WB(ei "') are shown in Fig. 9.14. The magnitude
of the first sidelobe level has been reduced to - 57 dB. However, the main-lobe
width has increased to 1211" IN. This is another illustration of the tradeoff
between the sidelobe magnitude and the main-lobe width.
This reduction in the sidelobe level can be explained in a manner similar
to the raised-cosine window. The Blackman window can be written as the

FINITEIMPULSE RESPONSE FILTER DESIGN TECHNIOUES

337

(a)

-15

(b)

15

Wo(e'"')

It

-so
-57

-100 0
It
(J)

ftGURE 9-14
Blackman window results (N =31).
(a) Window sequence; (b) spectrum;
(c) log-magnitude spectrum.

product of an infinite-duration sequence and a rectangular window, or

WJt(n) = [0.42 + 0.5 cos(21rn/(N -1 + 0.08 cos(41rn/(N -1))]~(n)

for all n
for all n

(9.22)

where {~(n)} is the N-point rectangular window. As before, the multiplication of two sequences in the time-domain corresponds to a convolution in the

338

INfRODUcnON TO DIGITAL SIGNAL PROCESSING

frequency domain of their spectra. Since {c2(n)} is composed of a constant

term and two cosine terms at frequencies co = 2lr/(N - 1) and co = 4lr/(N - 1),
its Fourier transform can be expressed as
C2 (ei "') = O.84lro(co) + o.SlrO( co +O.04lrO(CO for

N2~ 1) + o.SlrO( co + N2~ 1)

N4~ 1) + O. 04lro ( co + N4~ 1)

-lr :5 co :5 lr.

(9.23)

This spectrum is shown in Fig. 9.1S(a). The individual shifted and scaled
versions of WR(ei "') produced by the convolution with C2(ei "') are shown in
Fig. 9.1S(b). Their sum produces the spectrum of the Blackman window
WB(ei "') shown in Fig. 9.1S(c). The main-lobe width is extended by the
components centered at co = 4lr/(N -1), but these components serve to
reduce the sidelobe levels as well.
The digital lowpass filter results obtained when applying the Blackman
window are shown in Fig. 9.16. The minimum attenuation in the stopband is
an iIJ'lpressive -74 dB, but it occurs for co> lr/2. Such a large transition region
is necessitated by the wide main-lobe width of the Blackman window
spectrum.
The windows discussed above are compared in terms of their main-lobe
width and the maximum sidelobe level in Table 9.l.
Although other more complicated windows have been proposed, their
performance is only slightly better than that produced by the windows above.
These other windows are not discussed here, but can be found in the
references. Another reason for discussing only these simple windows is that the
shape of the window spectrum itself is not the important parameter, but rather
its effect on the transfer function of the final FIR filter that is being
approximated, so that different windows may be better for different filter
types. Hence, it is impossible to make a general statement regarding the
appropriateness of a particular window for the general filter. It is more
important to be aware of the general effects produced by the window spectrum
main-lobe and sidelobe structures.
(a)

O.Srr

O.04rr

O.Srr

O.04rr

FIGURE '15
Frequency-domain interpretation of the Blackman window
results. (a) Spectrum of infinite-duration sequence; (b)
shifted and scaled versions of
the rectangular window spectrum generated by the convolution; (c) spectrum of the
Blackman window is the sum
of the components in (b).

FINITE-IMPlTLSE aESPONSE FILTER DESIGN TECHNIOUES

339

FIGURE 9-15b

FIGURE 9-&

The main problem with applying the windowing method is that the
desired unit-sample response {hD(n)} must be known. For a particular desired
magnitude response, this unit-sample response may not be known. In the next
section, we describe the DFf method that can be used to obtain a
finite-rluration approximation of the desired unit-sample response. In the
following section, we will apply the windowing method to this approximation.

340

lNTRODUcnON TO DIGITAL SIGNAL PROCESSING

o~----~~---+------~--~
It

(c)

I H31 (eI"') IdB

O_-~

-50
nGURE 9-16
Results of applying a 31-point Blackman
window to the ideal lowpass filter unitsample response. (a) Finite-duration
unit-sample response; (b) magnitude
response of approximation to ideal lowpass filter; (c) log-magnitude response.

-74

TABLE 9.1

Spectral properties of N-point windows

Muim... lidelobe Ienl
(relative to muimum), dB

Window

Rectangular
Bartlett
Hanning
Hamming
Blackman

41r/N
81r/N
81r/N
81r/N
121r/N

-13

-25
-31
-41

-57

FINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

341

9.3 THE DFf METHOD FOR

APPROXIMATING mE DESIRED
UNIT-SAMPLE RESPONSE
In this section, we describe a procedure to obtain a finite-duration approximation, denoted by {hin)}, to the desired unit-sample response {ho(n)} by
sampling the desired magnitude response 1Ho(ei "')1 and performing an inverse
Off. To compute the inverse Off, we must also specify a phase response
sequence, for which we will consider a Iinear-with-frequency response. Since
the inverse Off is employed, the approximation to {ho(n)} is finite-duration,
having Nd samples when an Nd-point Off is employed. In this case, Nd should
be much larger than N, the duration of the unit-sample response of the final
FIR filter. Having determined an approximation to {ho(n)}, we can then
apply the windowing method for designing the final FIR filter.
The Off method is important because it is the discrete-time equivalent
of performing signal processing in the frequency domain by operating directly
on the spectrum of the input signal, most often by using the FFf algorithm. By
examining this frequency-domain operation as the equivalent time-domain
filtering operation, we obtain a better understanding of the effects associated
with the frequency-domain approach. The Off method also serves as an
introduction to the frequency-sampling filter design considered later in this
chapter.
To find {hd(n)}, we need to compute the inverse of an Nd-point Off
sequence. The inverse Off produces a sequence defined over 0::;; n ::;; Nd - 1.
In Chapter 4, it was shown that the Nd-point sequence is one period of the
periodic sequence {iid(n)}, which is the periodic extension of the desired
sequence {ho(n)}, or
co

iiin)

= L
k=-oo

ho(n

+ kNd)

(9.24)

The relationships of {ho(n)}, {hin)} and {iid(n)} for the ideallowpass filter
are shown in Fig. 9.17. If {ho(n)} has an infinite duration, some degree of
time-aliasing will occur. For a stable filter, {ho(n)} is equal to or approaches
zero for both positive and negative values of n. Hence, the time-aliasing effect
can be made negligible by choosing Nd sufficiently large. In the design
procedures below, we will consider hd(n) = ho(n) for the Nd points over which
{hd(n)} is defined.
To illustrate the Off ..approach, let us consider the ideal lowpass filter.
To obtain this Nd-point Off sequence, the magnitude response is sampled at
frequency points Wk = 21fk/ Nd, for 0::;; k ::;; Nd - 1. Let us define the sequence
{Ho(k)} to be
(9.25)
for 0 ::;; k ::;; Nd - 1
Applied to the ideallowpass magnitude response, we have
for 0::;; k::; NL_and for Nd - NL ::;; k ::;;Nd -1, and
<
otherwise

342

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

(hDCn)}

...

(hCn)}

INVERSEDFT

FIGURE 9-17
Relationships of desired unit-sample response {ho(n)}, its periodic extension {hd(n)} (Nd 0= 64),
the sequence produced by the inverse DFf assuming a zero phase sequence, and the approximate
unit-sample response {hd(n)} for the non-causal FIR filter.

ANITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

(a)

343

{IHD(k)l}

III . .
o

. II .

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

{Arg[Ho(';')j}

o t'5 k

(b)

{h(n)}

(c)

~---------~~--------~)

FIGURE 9-18
(a) Magnitude and phase DFf sequences for approximating the ideallowpass filter (N = 16); (b)
resulting 16-point inverse DFf .<!quence; (c) approximate unit-sample response for the non-causal
FIR filter. Parts (b) and (c) indicate the shift of samples to be performed.

344

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

where NL defines the edge of the passband, as shown in Fig. 9.18 for Nd = 16
and NL = 2. To determine the unit-sample response from the inverse OFT, we
must also specify a phase response sequence. Let us first choose the simplest,
the zero-phase response sequence, given by
Arg[Ho(k)] = 0

for aU k

For this zero-phase response sequence, the real and imaginary sequences are
equal to
and
Hence, the OFT element values are given by
Ho(k) = HR(k) + jHI(k) = HR(k)

(9.26)

The Nd-point unit-sample response {hin)} is then related "to the inverse OFT
sequence {h(n)} computed as
1 N.-l
h(n) = - 2: Ho(k) ei 21<knlNd
(9.27)
Nd k=O
This sequence is shown in Fig.9.18(b). The analytic evaluation of the inverse
OFT can be simplified by noting that, for the zero-phase response sequence,
{HD(k)} is an even sequence of k. The limits of the sum can then be taken
from -NL to NL , over which Ho(k) = 1. Hence,
1

NL

h (n) = 2: ej21tknlNd
Nd k=-NL
e -j21tM.nINd 2M.

2:

Nd

ej21tknlNd

k=O

e- j21tM.nlNd 1- ej21t(2M.+l)nINd
= Nd
1 - ei21tnlNd
sin[2.n-(NL + !)n/Nd ]
(9.28)
Nd sin[.n-n / Nd]
Since a zero-phase response sequence has been specified, the unit-sample
response should be an even sequence of n. The Nd-point noncausal sequence,
corresponding to the zero-phase response sequence, can be obtained from the
inverse OFT results by shifting the second half of the sequence, the values for
(Nd/2) + 1 s n s Nd - 1, into negative values of n, as shown in Fig. 9.18(c).
This then defines {hin)}, the FIR approximation to the desired zero-phase
lowpass filter. However, even after this shift is accomplished, the Nd-point
sequence may not be exactly symmetric about n = 0 because Nd is an even
integer and there may be one additional nonzero element without a partner at
n = Nd/2. Since this symmetry is a necessary condition for a linear-phase
response, even though the sequence {Arg[Ho(k)]} was specified to be
zero-valued, the phase response function of the filter Arg[Ho(ei "')] cannot be
equal to zero, except at the samplin.,g points.
=--'''---'--.:::--=~-=

FINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

345

-It

FIGURE 919
Comparison of the Fourier transform of {hd(n)} shown in Fig_ 918 and the DFf sequence values_
The DFf samples are shown as heavy dots.

To resolve this dilemma, let us define Hd(ei "') to be the filter transfer
function determined by computing the Fourier transform of {hd(n)}, or
N.,;2

Hie iw ) =

hd(n)e- iwn

(9.29)

n=-Ndf2 + 1

This transfer function is plotted in Fig. 9.19. Note that the function passes
through the sample values of {Ho(k)}, but is not zero-phase.
We now want to relate Hd(eiw ) to the desired transfer function Ho(eIW).
The connection between the two is that {hin)} is one period of the periodic
extension of {ho(n)}. Expressing {hd(n)} as the rectangular window truncation of this periodic extension, we have
00

Hie iw ) =

wR(n) hd(n) e- iwn

(9.30)

n=-IX)

where {WR(n)} is the Nd-point rectangular window and {hi n)} is the periodic
extension of {ho(n)}. As shown in Chapter 3, the multiplication of two
discrete-time sequences corresponds to the circular convolution of the two
corresponding Fourier transforms. We then have
(9.31)

346

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

where WR(e iW ) is the spectrum of the window sequence, and

00

H(eiW )

2:

hd(n) e- iwn

(9.32)

n=-oc

Since {hd(n)} has an infinite number of nonzero elements, it is not absolutely

summable and a true functional expression does not exist for H(eiW).
However, it can be shown that H(eiw ) can be expressed as the set of Dirac
delta functions located at the frequency sample points Wk = 21rk/ Nd , for
O:os k :5 Nd - 1. The weight of the delta function at W = Wk is equal to the DFf
sequence value Ho(k), as shown in Fig. 9.20. When convolving any function,
in this case WR(e iw ), with a Dirac impulse, only a shift is produced in the
function to the location of the impulse. Because integration is a linear
operation, each delta function in H(eiw ) can be considered separately and the
contributions from all the delta functions are then summed. The resultant
continuous-frequency transfer function is shown in Fig. 9.21b. As expected,
the DFf sequence {Ho(k)} is equal to Hd(e iW ) at the sampling points, but
Hieiw ) exhibits overshoots and leakage terms very similar to those produced
by the rectang'llar window method described in the previous section. In fact,
the windowin!; operation was performed when {wR(n)} extracted the Nd-point
unit-sample response from the periodic extension in Eq. (9.30).
It is important to realize that the relevant function to be specified for a
digital filter is the transfer function Hd(e iw ), and not the DFf sequence
{Ho(k)}. Even though the Nd-point DFf sequence Ho(k), for 0:5 k:5 Nd - 1,
is a ~ampled version of the desired transfer function Ho(e iw ) over 0:5 W < 21r,
{iid(n)}

H(e'"')

'I II
FIGURE 9-20
Periodic extension of {ho(n)} forms the infinite-duration sequence {hd(n)}. The spectrum of the
infinite-duration sequence H(e JW ) is composed of Dirac delta functions.

FINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

347

thevalues of the FIR filter transfer function Hd(e JW ) between the sample points
may deviate significantly from HD(eiw ). The differences are most significant
around frequencies for which IHD(eiw)1 is discontinuous, as illustrated by the
edge of the passband in the ideal lowpass filter. When IHD(eJW)1 is a smooth
function of frequency, the sampled sequence {IHD(k)1} is a good representation of the form of IHD(eiW)I, provided the sampling is dense enough,
which' occurs for large values of Nd In this case, the filter transfer function
Hd(e IW ) will be essentially equal to HD(e iW ).
The above analysis also illustrates the problems that can occur when
filtering is performed by taking the DFf of the input data and merely
eliminating the undesired frequency components by setting them equal to zero.
In effect, a zero-phase filter is assumed having a symmetric finite-duration
unit-sample response {hin)}.
,In the next section, we co,nbine the windowing and DFf methods for the
design of practical FIR filters.
TIME-DOMAIN OPERATIONS
(hd(n)}

_.........~."
.llIlllll~IIlIIlII....._.......... ~
-N/ Z

+1

N/2

FIGURE '21
Interpreting the overshoots and leakage effects in Hie JOJ ). (a) Time domain; (b) frequency
domain.

348

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

H(e"")

1,

ftGURE !)2Ib

9.4 DESIGNING FIR FILTERS BY

COMBINING mE DFf AND WINDOW
MEmODS
In this section, we combine the windowing and OFT methods to design FIR
filters. This combination eliminates the disadvantages of each method. For the
windowing method, we needed to know the form of the desired unit-sample
response {hD(n)}. The OFT method produces {hd(n)} , the finite-duration
approximation to {hD(n)} containing Nd-points, where Nd is a large number.
When applying the OFT method directly, overshoots and leakage effects are
observed in the FIR filter transfer function Hd(eiOJ ). The windowing method
can be employed to reduce these effects. The windowing procedure allows the
duration of {hin)} to be reduced, to take advantage of the tolerance limits in
the magnitude specification. This cannot be done easily by using the OFT
method directly. We first consider the method of obtaining a desired

FINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

349

magnitude response sequence {IHo(k)l} from the magnitude response specification, followed by the procedure for specifying an appropriate phase
response. An inverse DFf then produces the Nd-point unit-sample response
{hd(n)}. The application of an N-point window, where N < Nd, to obtain
{hN(n)} from {hin)} is then discussed. Finally, a step-by-step procedure for
designing an FIR filter is presented.
In this section, we will describe the DFf/windowing procedure for
implementing lowpass, highpass, bandpass and bandreject filters. When
discussing the application of windows earlier in this chapter, we observed that
the resulting filter magnitude response was a distorted version of desired
magnitude response IHo(eIW)I. This distortion was caused by the convolution
of Ho( e iw ) with the spectrum of the window. This distortion prevents the
application of this procedure to designing filters having an arbitrary magnitude
response shape, such as linear-with-frequency, for example. A magnitude
response having a specific shape is more easily achieved with the frequencysampling method described in the next section .

SPECIFYING THE APPROPRIATE MAGNITUDE RESPONSE SEQUENCE. In

the design of lowpass filters by using windows, the resulting magnitude
response exhibited a reduction in the passband. This reduction is one
consequence of the ~onvolution with the main-lobe of the window spectrum
with the desired response Ho(eiw ). With the idea that a reduction will occur in
the final lowpass or bandpass filter magnitude response, the desired ideal
magnitude response IHo(eiW)1 should have its cutoff frequencies extended into
the transition regions of the specification, as shown in Fig. 9.22. The size of
this extension depends upon the main-lobe width of the final window
spectrum, which is related to the duration of the final unit-sample response and
window type employed. Since this duration is not known at the beginning of
the design process, a guess must be made. A good starting value for this
extension is equal to 30% of the transition region.
The magnitude response IHo(e i"')I gen'erated in this fashion is then
sampled at Nd points over 0:5 W < 2:r to generate the magnitude response
sequence {IHo(k)l}. The value Nd is much larger then the anticipated duration
of the final filter, with Nd = 128 usually being sufficient for most practical
filters.
The procedure to define the magnitude response sequence that is least
prone to error is to sample IHo(eiw)1 at (Nd/2) + 1 points over 0:5 w:5:r and
apply the even symmetry property of the magnitude sequence to obtain the
entire Nd-point sequence. Doing this, we have

IHo(k)1

= IHo(ei"')I..,=ei2d'Nd

IHo(k)1 = IHo(Nd - k)1

for 0 :5 k :5 Nd/2
fOr (Nd/2)

+ 1:5 k :5 Nd -

1.

(9.33)

SPECIFYING
THE
APPROPRIATE
LINEAR-PHASE
RESPONSE
SEQUENCE. From {IHo(k)l} and a ,~'pecified phase sequence {Arg[Ho(k)]},

350

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

1-- _______ J

IHo(eI"') I

r-------.,

ofI
O>p

0>,

Wc

0>

1t

IH(eI"') I
IfI-Ef-

L.. _ _ _ _ _ .J

/llio(eI"')1

,.-...

o -,

(Os!

ro CI c.o PI

I
rop~

CJ)c~

Ws~

1t

0>

FIGURE 922
The desired magnitude response for lowpass
and bandpass filters is obtained by extending the passband into the transition regions.
The specification is shown as dashed lines.

the unit-sample response {hd(n)} can be obtained by performing an inverse

discrete Fourier transform. The causal unit-sample respon~ e, having Nd points,
of a linear-phase filter should show a point of symmetr, (or 'antisymmetry)
about the half-way point, or at Ns = (Nd -1)/2. As shown in Chapter 3, four
possible cases of symmetry can occur: the sequence can be symmetric or
antisymmetric, and, for each of these, the point of symmetry can either fall on
a sample point or between sample points. When a conventional Nd-point FFT
algorithm is employed to compute the inverse DFT, Nd is an even integer.
Hence, the point of exact symmetry cannot fall on a sampling point in general.
In the ideal lowpass filter case considered in the previous section, when the
phase response sequence was zero-valued, the Nd-point unit-sample response
was almost symmetric about n = 0, except that a point without a partner
occurred at n = Nd/2. When the N-point window is applied, for N < Nd, this
extra point is deleted and the unit-sample response {hN(n)} can be made to be
symmetric about a sampling point.
The appropriate linear-phase sequence that is assigned to match the
magnitude response is determined by the symmetry properties of the unitsample response. The phase responses corresponding to the different symmetries in {hin)} are shown in Table 9.2. The entire phase value, rather than
its principal value, is given. In computing the real and imaginary parts from the
magnitude and phase, this entire phase value is sufficient.

F1NITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

351

TABLE 9.2

Linear phase responses for the various types of symmetry of the Nd-point
unit-sample response {hd(n)} (Nd is even integer)
Even symmetric with the point of symmetry on sample point Ns = Nd/2
Arg[Ho(eJW )] = -wNd /2

for 0 ~ w ~ H

Even symmetric with the point of symmetry between sample points, Ns = (Nd -1)/2
Arg[Ho(eiw )] = -W(Nd -1)/2

for O~ w'os; H

Antisymmetric with the point of symmetry on sample point Ns = Nd/2

Arg[Ho(eJW )] = -wNd /2 + rt/2

for

O~

w os;H

Antisymmetric with the point of symmetry between sample points, Ns = (Nd

Arg[Ho(eiw )]

= -w(Nd - 1)/2 + H/2

1)/2

for 0 ~ w ~ H

The phase at frequency points for which the magnitude response is zero-is
undefined and can be set to an arbitrary value. To simplify the specification of
the phase, we set it equal to the value of the linear trend. Then, at W = 0, the
phase value of antisymmetric sequences is defined to be equal to n /2. As
shown in Chapter 5, even symmetric sequences that have an even number of
elements have a z-transform that always produces a zero at z = -1, or a
magnitude response that is zero-valued at W = n. This condition occurs when
the point of symmetry falls between sample points at Ns = (Nd - 1)/2. Then, at
w = n, the phase value of even symmetric sequences is defined to be equal to
-(Nd - l)n/2.

The procedure to define the phase sequence that is least prone to error is
to sample Arg[Ho(eiCLI )] at (Nd/2) + 1 points over 0:$ w:$n and apply the odd
symmetry property of the phase sequence to obtain the entire Nd-point
sequence. Doing this, we have
Arg[Ho(k)] = Arg[Ho(eiCLI)] ICLI=eJ2KWNd
Arg[Ho(k)] = -Arg[Ho(Nd - k)]

for 0:$ k :$ Ndl2

for (Ndl2) + 1 :$ k

:$

Nd - 1.
(9.34)

As an example, let us consider the ideal lowpass filter magnitude

response with We = n/4, which is sampled over 0:$ W < 2n at Nd points for
Nd = 32. Since the transfer functions of antisymmetric unit-sample responses
have a zero at (J) = 0, they are not appropriate for the lowpass filter. The two
types of symmetric unit-sample responses from a 32-point inverse DFf are
shown in Fig. 9.23.
APPLYING A WINDOW. The above procedure provides an Nd-point ap-

proximation {hd(n)} to the desired unit-sample response {ho(n)}. We do not

want to use {hd(n)} directly, because it does not take advantage of the allowed

352

IN1"RODUcnON TO DIGITAL SIGNAL PROCESSING

{HDCk)}
(a)

11111 ...................... llll..

2 3 4 2 M 29 3() 31

(b) Point of symmetry at N,'Nd/2

{hd(n)

(e)

Point of symmetry at N,=(N u l)/2

{hd(n)

15 ;

FIGURE 923
Effect of phase response slope on unit-sample response. (a) 32-point magnitude response used in
the inverse DFf computations; (b) unit-sample response sequence result when phase slope is -16;
(c) unit-sample response sequence result when phase slope is -15.5.

tolerances in the specification. However, by reducing the duration of the

Nd-point sequence with a window sequence to N points, an FIR filter can be
implemented that meets the desired specification.
For the final FIR filter to have a linear-phase response, the windowed
unit-sample response {hN(n)} must be symmetric. To maintain the symmetry
that is present in {hd(n)}, an appropriate N-point window must be chosen
accordingly. If the point of symmetry of {hi n )} was chosen to fall on a sample
point, the window duration N must be an odd integer. If the point of symmetry
falls between sample points, N must be an even integer. Examples are shown
in Fig. 9.24.

FINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

353

POINT OF SYMMETRY Ns
ON SAMPLING POINT
{wH(n)}

(hd(n)}

- .. nlIl

l
IlIIIIIIII

III Iltl fI, ...."

Ns

POINT OF SYMMETRY Ns
BETWEEN SAMPLING POINTS

(hd(n)}

II

.. ,tI'IIIlIllll.lllllJlllIlt
.....
t
n

Ns

FIGURE 9-24
Positioning a Hamming window to maintain the symmetry that is present in the
original sequence_

Having described these separate parts, we can now integrate them into a
procedure to design an FIR filter.
STEPS IN DESIGNING A FILTER USING mE OFT/WINDOWING
METHOD. The steps to design a linear-phase FIR filter with the

DFf/windowing method are summarized in Table 9.3. To illustrate this

354

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

TABLE 9.3

Steps in performing DFf /windowing FlR filter design

Step 1. Determine the approximate Nd-point unit-sample response {hin)}
(a) Sketch the desired magnitude response /Ho(eJW )/.
(b) Choose Nd to be greater than the duration of the final filter unit-sample
response. Nd = 128 is a good choice.
(c) Assign the appropriate linear phase response Arg[Ho(ejW)].
(d) Sample /Ho(ei"')/ and Arg[Ho(eJW)] to obtain Nd-point sequences (/Ho(k)/}
and {Arg[Ho(k)]}.
(e) Compute the Nd-point real and imaginary sequences {HR(k)} and {H,(k)}.
(f) Evaluate the Nd-point inverse OFT. The sequence {hd(n)} is the real-valued
result.
Step 2. Determine the N-point unit-sample response {hN(n)}.
(a) Choose N (good initial guess: N = Ni2).
(b) Multiply {hd(n)} with N-point window to obtain {hN(n)}.
(c) Compute the Nd-point OFT of {hN(n)} to obtain {HNik)} (padding with
zeros is necessary).
(d) Compare {/HNd(k)1} with desired specification.
(e) If specification is satisfied, try to reduce N further, if specification is not
satisfied increase N. Repeat (b) through (e) until minimum value for N is found.
Step 3. Linear-phase FIR filter implementation
(a) Find the maximum value in {/HNd(k)/}. denoted by HMAX.
(b) Scale {hN(n)} by I/HMAX to find final filter coefficients.
(c) Shift coefficient sequence to start at n = 0 for causal filter. For non-causal
filter, shift {hN(n)} to be symmetric bout n = O.
(d) Implement the filter using special linear-phase FIR structure.

procedure, we consider the Nd-point sequence {hd(n)} to be symmetric about

the point n = Nd/2. In this case, the set of variable-duration windows are also
aU symmetric about this point. This feature simplifies the specification and
application of the window, as shown below. The case of {hin)} whose point
of symmetry lies between sample points can be handled by making small
modifications to the procedures given below.
Step

1. Determine

the

approximate

desired

unit-sample

response

{hin)}. The desired magnitude response, IHo(ei"')1 is determined by extend-

ing the cutoff frequency Wc into the transition region. The desired magnitude
response is sampled to produce the magnitude response sequence {IHo(k)I}.
The value of Nd should be chosen to be some large number, greater than the
duration of the final filter design, to make the time-aliasing effects in {hd(n)}
negligible. Most of the FIR filter specifications in this book will be satisfied
with durations of less that sixty samples. Hence, Nd = 128 is sufficient. The
linear-phase response sequence {Arg[Ho(k)]} is specified according to Table
9.2. For ease of specifying and applying windows, the point of symmetry of
{hd(n)} should be specified to be n = Nd/2.
In evaluating the inverse DFT, the real and imaginary parts of the DFT

FINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

355

sequence, {HR(k)} and {HI(k)}, must be computed_ These are given by

= IHo(k)1 cos(Arg[Ho(k)])
and
HI(k) = IHo(k)1 sin(Arg[Ho(k)])

HR(k)

for 0:5 k :5 Nd - 1 (9.35)

Since both the cosine and sine functions are periodic, it is not necessary to
express the phase in the principal value range of [-Jr, Jr] in the above
procedure for determining the phase response.

Step 2. Determine the values of the N-point unit-sample response

{h~n)}. By choosing {hd(n)} to be symmetric about n = Nd/2, the variableduration windows can be specified and applied in a direct fashion, as shown in
Fig. 9.25. The window is then defined to be symmetric about n = Nd/2 and has
an odd number of points to maintain the even symmetry in the N-point
windowed sequence {hN(n)}. For example, the N-point Hamming window that
is symmetric about Nd/2 is given by
wH(n)

0.54 + 0.46 cos[2Jr(n - Nd/2)/(N -1)]

for Nd/2 - (N - 1)/2:5 n:5 Nd/2 + (N -1)/2 (9.36)
{
otherwise.

The product of {wH(n)} and {hd(n)} is computed to determine the

shorter N-point windowed sequence {hN(n)} , where N < N d, which is a
candidate for the FIR filter unit-sample response. A reasonable first choice is
N = Nd/2. The goal is to find the minimum value of N such that {hN(n)} is the
unit-sample response of the FIR filter that satisfies the magnitude response
specification.
The magnitude response of the resulting FIR filter can be verified by
computing the Nd-point OFT of the N-point windowed umt-sample sequence
{hN(n)}, denoted {HNik)}. Let {hr(n)} and {hj(n)} be the real and imaginary
Nd-point sequences to be applied to the OFT. To compute the OFT, {hr(n)}
contains {hN(n)} and the zero values that were produced by the windowing
procedure, or hr(n) = hin) KH(n) for 0:5 n :5 Nd - 1. The imaginary sequence
{hi(n)} is set to zero.
The magnitude response sequence {IHNd(k)l} is then compared with the
specification to verify that both the passband and stopband specifications are
met. Because the Nd-point sequence {lHNik)l} is computed from a sequence
having only N points, where N < Nd, it appears to define a smooth curve. In
this case, we can be fairly confident that the digital filter magnitude response
IH(eiw)1 is accurately represented by the magnitude response sequence
{IHNik)I}
If {IHNik)l} satisfies the specification, an attempt to reduce the value of

N should be made. This will insure that a shorter duration unit-sample

response does not also satisfy the specification and that the minimum duration
sequence has been found. If {IHNik)l} does not satisfy the specification the

356

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

o
49-point window

33-point window

,. . . . . "~m~~~~llillfi~. . . . . ,,
o

Nfl

Na !

17-point window

'""...... "". dil l~l l 'l i. ""....."" . . . ., '

Nfl

Nd-!

FIGURE 9-25
By locating point of symmetry of {hd(n)} at n = Nd/2, variable length windows can be applied
directly. When Nd/2 is an integer, windows having an odd-integer duration must be used.

F1NITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

357

value of N should be increased. If N becomes greater than Nd/2, the initial

number of samples, this initial value was too smalL Then, Nd should be
increased by two and the entire procedure repeated.
Step 3. Linear-phase FIR filter implementation. Because of the windowing
procedure, the scale of the magnitude response may have been reduced. It
must then be adjusted to meet the specification. For example, let IHD(e j "')I
have a maximum value equal to 1. Because the duration and values of {hin)}
are reduced in the windowing operation, the maximum value of the filter
magnitude response IH(ej"')I, as approximated by the final {IHN.(k)l}, may no
longer be equal to 1. To make it equal to 1, the maximum value of {IHN.(k)I},
denoted by H MAX , should be determined and the values of the unit-sample
response should be multiplied by II H MAX
To implement a causal filter, the windowed unit-sample response that is
symmetric about n = Nd/2 must be shifted, so that the first nonzero value
occurs at n = O. This eliminates unnecessary delays in the filter structure. For
non-causal filters, {hN(n)} is shifted so that its point of symmetry occurs at
n = O. Because of the time symmetry of {hN( n)}, the special linear-phase FIR
filter structure can be used to reduce the number of multiplications reqcired to
accomplish the filtering operation.
To illustrate this procedure, let us consider the design of a filter by
combining the DFT and windowing methods.
E:umple 9.1. Design of linear-phase FIR filter with the DFl/windowing
procedure. Let us consider the bandpass specification shown in Fig. 9.26(a),
given by
passband: -I ~ IH(eiW)ldB~O
stopband: IH(ej")ldB < -60

for
for

O.42n ~ W

~ O.61n

O~w~O.16n

and

O.87n~w~n.

Step L Determine {hin)}.

Since the passband response will be reduced when a window is applied, the
cutoff frequencies of the desired magnitude response are extended into the
transition region by 30%, so that WeI = O.34n and WC2 = O.69n, as shown in Fig.
9.26(b).
The value Nd = 64 was chosen to make the figures easier to interpret. A
larger value of Nd could have been chosen, but the time sequences are more
easily drawn for Nd = 64.
To show the four possible different 64-point {hd(n)} , the corresponding
four possible phase characteristics given in Table 9.2 were all invoked. The point
of symmetry was chosen to be either Ns = 32 or Ns = 31.5. The magnitude and
phase functions were defined over 0 ~ Q) ~ n. The Nd-point real and imaginary
Off sequences are constructed from these in a way to insure their symmetry
properties, or
HR(k) = IH(k)1 cos(Arg[H(k)])
HR(k) = HR(Nd - k)

for O~ k ~Nd/2
for (Nd/2) + 1 ~ k

Nd - 1

358

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

(a)

IH(e'"') I

1
0.89
(-1 dB)

O.OOI(~ dB)LL._-:-~
o

:--_~IL,,- --:--:!r===t~_~-+

_ _--,,,..-L:1
0.16n:
0.42n:

(b)

__

0.6In:

0.87n:

n:

..

0)

IHo(e'"') I
If-

____

~L,,-

_______

0.34n:

~~

_____

~w

0.69n:

FIGURE 9-26
Initial frequency curves for Example 9.1. (a) Specification; (b) desired magnitude response
function.

and
HI(k) = IH(k)1 sin(Arg[H(k)])
HI(k) = -H1(Nd-k)

for O:s k :s Nd/2

for (Nd/2) + 1 :s k :s Nd - 1

The inverse DFf was evaluated with the four possible phase responses to
produce the four unit-sample responses shown in Fig. 9.27. To take advantage of
the zero-valued coefficient occurring at n = 32, thus eliminating a multiplication
in the digital filter, we can use the antisymmetric sequence that has its point of
symmetry on a sampling point as {hd(n)}. The magnitude response IH(ei "') I of
this sequence is shown in Fig. 9.28. This magnitude response was obtained by
computing the 512-point DFf of {hd(n)}. The overshoots and leakage effects due
to the 64-point rectangular window trunction are evident.
Step 2. Determine {hN(n)}.
A 33-point Hamming window was applied {hd(n)} to obtain {h33(n)}. The
corresponding magnitude response is shown in Fig. 9.29(a). We note that it is
much smoother than that of {hd(n)}, being convolved with the spectrum of the
Hamming window. The passband requirement is easily met, but the minimum

FINITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

359

(a)

(b)

315

(d)

31 5

'-17

FIGURE
Four unit-sample response sequences resulting from the four possible phase response types.
(a) Even symmetry with point of symmetry falling on sample point; (b) even symmetry with
point of symmetry falling between sample points. (C) Odd symmetry with point of symmetry
falling on sample point; (d) odd symmetry with point of symmetry falling between sample
points.

stopband attenuation is approximately -60 dB. Shorter Hamming windows

exceeded the stopband attenuation specification.
Since the stopband attenuation is governed by the sidelobe magnitude of
the window spectrum, the 33-point Blackman window was considered because of
its sidelobe characteristics. The magnitude response of the 33-point {hN(n)}
obtained with the Blackman window is shown in Fig. 9.29(b). Both the passband
and stopband specifications are met. In an attempt to find the minimum duration
window, shorter Blackman windows were tried. A 31-point window was found to
be the minimum duration that satisfies the specification, with results similar to

360

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

PASSBAND:
IH(e'''')ldB

o .
-lr-----------~Hr~~~+.Hr~------------

-2

-3

~~0------------L-~0~.4~2~n----~0.~6~ln~L---------~n ro

STOPBAND:
IH(e"")ldB

-20

~o

nGURE '28
Magnitude response corresponding to (hd(n)} shown in Fig. 9-27(c).

Fig. 9.29(b). Reducing the window further caused the low-frequency stopband to
be exceeded. Hence, N = 31 is the minimum duration, to give {h31(n)}.
Step 3. FIR filter implementation.
The maximum value of IH(e,W)1 for {h 31 (n)} was found to be equal to
HMAX = 0.9998. This value was close enough to one so that scaling the
coefficients by 1/HMAX was not necessary. The causal FIR filter structure is
shown in Fig. 9.30. Note that the output of the fifteenth delay is not connected to
a multipler because h 31 (15) = O. However, this savings of one multiplier is bought
at the cost of multiplying the outputs of 15 delays by -1 to achieve the
antisymmetric unit-sample response. If an even-symmetric {hd(n)} was used,
these negations would not be necessary. This filter would be efficient, however,
if the digital hardware allowed a subtraction to be performed as efficiently as an
addition.

FINITEIMPUlSE RESPONSE FILTER DESIGN TECHNIQUES

361

In the next section, we consider another FIR filter design procedure that
results in a structure which allows the processing to occur in parallel.

9.5 DESIGNING FIR FILTERS WIm THE

FREQUENCY-SAMPLING MEmOD
The frequency-sampling method is similar to the DFT method in that the
desired magnitude response is sampled and a linear-phase response is
(a)

33-point Hamming window results

PASSBAND:
IH(e"")ldB
0..-

f-

r2f-

c3ff-

-4

(i)

0.421t

0.611t

It

33-point Hamming window results

STOPBAND:
IH(e"")ldB

FIGURE 9-19
Magnitude response corresponding to windowed version of {hd(n)} shown in Fig. 9-27(c). (a)
33-point Hamming window results; (b) 33-poin(tBlackman window results.

362

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

(b)

33-point Blackman window results

PASSBAND:
IH(e'''')ldB

o
-I~--------------r+------~--------------

-2

-3

-4~O------------~~~----~~~----------~n

33-point Blackman window results

STOPBAND:
IH(e"")/dB

o
-20

-40

~~----~-------------------------H-----

~0~~~0~.1~6n------------------------~~~~

ftGURE '1911

specified. The difference is that, rather than computing the inverse Off to
determine the unit-sample response of the FIR filter, a special parallel filter
structure is used to implement each frequency sample directly. For a given
specification, the frequency-sampling procedure may produce a filter that is
faster and easier to implement in hardware.
Given a desired magnitude response IHD(&OI)I having an arbitrary shape,
the goal of this design procedure is to design an FIR filter having an N-point
unit-sample response and having a magnitude response IH(&OI)I, such that
IH(eiCOk)1

= IHD(eiCOk)1

FINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

x(n) - - - . . . - - - - - - - - ,

y(n)

FIGURE 930
linearpbase FIR filter structure produced by 31-point Blackman window.

363

364

INTROOUcnON TO DIGITAL SIGNAL PROCESSING

where C/)k = 2:rck1N for O:s: k:s: N -1. That is, we want to design an FIR filter
whose magnitude response exactly matches the desired magnitude response at
N equally spaced points over O:s: C/) < 2:rc.
We start by considering the FIR filter structure that is implemented by a
comb filter cascaded with a bank of resonators, described in Chapter 6. The
system function can then be written in the form:
=

H(z)

(1-NZ-N) ~o 1-Ho(e
ei2n;kIN -l
j2nkJN

N-l

(9.37)

This system function can be implemented as a camb filter, the term in

parentheses, in cascade with N complex first-order resonators that are
connected in parallel, as shown in Fig. 9.31(a). The gain of the resonator at
frequency C/)k = 2:rck IN, for O:s: k :s: N - 1, is equal to the value of the desired
complex-valued transfer function at that frequency Ho(ei2n;kIN). This complexvalued gain is computed from the desired magnitude response and the assigned
phase response.
For k=O (and for k=N/2, when N is an even-valued integer), the
first-order resonator must have a real-valued gain and feedback coefficient.
For other values of k, the coefficients are complex, but by combining
complex-conjugate terms, a set of real coefficients are obtained for each
second-order resonator section, denoted by HR,k(Z). Combining the k and
N - k first-order resonators, we have
Ho(ej 2n;kIN)
Ho(ej 2n;(N-k)lN)
HR,Jc(z) = 1- ei2n;kIN -l + 1- ei2n;(N-k)/Nz-I
Z

for l:s:k:s: (NI2) -1

(9.38)

This can be simplified, since

(9.39)

by the symmetry of the Fourier transform for real-valued sequences, and

ei2n;(N-k)/N = e-i 2n;kIN
by the periodicity of the complex exponential sequence. Combining Eq. (9.38)
over a common denominator and simplifying, we get
Ho(e j 2n;kIN)(l- e- j 2n;kINz -l) + H;'(ei'brlIN)(l _ ei2n;kINz-I)
HR,Jc(z) =
1- (e j 2n;kIN + e- j 2n;kIN)z-1 + z 2
2 Re[Ho(e j 2n;klA)] - 2 Re[Ho(e j 2n;kIN) e-i2n;kIN]z-l
=
1 - 2 cos(2:rckIN)z-1 + Z-2

(9.40)

where Re[ . ] is the real part of the complex quantity within the brackets. As
expected, the numerator is real-valued, but a function of both the real and
imaginary parts of Ho(e j 2n;kIN). The system function is then given by
H(z)

1- Z-N [Ho(ej~ Ho(d n ) (N~-I

]
N
1
-I + 1 + -I + ~ HR,k(Z) .
-z,/

k-l

(9.41)

FlNITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

36S

(a)

x(n)

yen)

(b)

x(n)

y(n)

ftGURE 9-31
Digital filter structures used when applying the frequency-sampling method. (a) General filter; (b)
simplified structure for linear-phase response.

366

INTRODUCllON TO DIGITAL SIGNAL PROCESSING

This result is valid for any desired magnitude and corresponding phase
response, but can be further simplified for the class of linear-phase filters
below.
CHOICE OF EVEN OR ODD VALUES FOR THE UNIT-SAMPLE RESPONSE
DURATION N. As shown in Chapter 6, the comb filter generates zeros that
are equally spaced on the unit circle. When N'is even, zeros are located at both
Z = 1 and z = -1, as shown in Fig. 9.32. For odd-valued N, there is a zero only
on the positive real axis at z = 1.
Not all of the zeros generated by the comb filter can be canceled by
resonators, since the pole/zero pattern is governed by the symmetry properties
of the unit-sample response. For example, it was shown previously that
antisymmetric sequences, since they have a zero-valued mean, always produce
a zero at w = 0 (z = 1). Hence, the resonator gain of the desired magnitude
response at w = 0 is always equal to zero. The system function for a
linear-phase filter having an antisymmetric unit-sample response HAS(z) is
given by
1- Z-N [HD(ei1l:) (NI2)-l
]
(9.42)
HAS(z) = N
1
-1 +
2: HR,k(Z)

+Z

k=l

Hence, if a zero is not desired at w = 0 in the final filter transfer function, as in

a lowpass filter for example, an antisymmetric unit-sample should not be
~o~n.
'
1m

1m
N=6

N=4

-.---~~--+-

Re

-'---~~--+-Re

1m

1m
N=3

N=5

--11---'-:----'-- Re

- t - - - i b - - - + - - Re

ftGURE 932
Pole/zero patterns in the z plane of comb filter for different values of N.

FlNITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

367

By a similar argument, symmetric sequences produce a zero at w = n

(z = -1) when N is even. Then the resonator gain of the desired magnitude
response at w = n is equal to zero. The system function for a linear-phase filter
having a symmetric unit-sample response Hs(z) is given by

Hs(z)=

1- Z-N
N

[H (eJo)

(N~-1

1~z 1+ 1:1 HRAz)

(9.43)

Hence, if a zero is not desired at w = n in the final filter transfer function, as in

a highpass filter for example, a symmetric unit-sample with an even duration
should not be chosen.
A SIMPLmCATION FOR LINEARPHASE FILTERS. A linear-phase filter can
have a unit-sample response that is either symmetric or antisymmetric about
some point in time. The associated phase responses were given in Table 9.2.
If the unit-sample response is symmetric about Ns = (N -1)/2, the
complex-valued samples of the resulting linear-phase transfer function can be
written as
Ho (k) = IH0 (eiro)1 e-i(N-l) ro12 1ro=21rkIN
(9.44)

where the linear-phase slope (N - 1)/2 corresponds to the point of symmetry

of the unit-sample response.
As shown in Chapter 5, an even symmetric sequence with an even
duration has a z-transform with a zero at z = - L The corresponding
magnitude response then has a zero at w = n. Hence, if N is an even integer,
Ho(N /2) = 0 and there is no resonator for k = N /2 or w = n. If N is an odd
integer, N /2 is not an integer and w = n is not a sample point, as shown in Fig.
9.32. So again, there is no resonator corresponding to w = n. To implement a
highpass filter, a zero at w = n is undesirable, so either N must be chosen to be
odd or an antisymmetric unit-sample response must be chosen. The simplification for the antisymmetric case is considered in the Problems.
Inserting Eq. (9.44) into the system function of the kth second-order
resonator Eq. (9.40), we get
2 IHo(ei21rkIN) I [cos(n(N -1)k/N) - cos(n(N - 1)k/N + 2nk/N) Z-I]
() =~~~--~~--~~--~~~--~~--~~------~~~

HR~Z

1- 2 cos(2nk/N) Z-1 + Z-2

(9.45)
The numerator can be simplified by adding 2kn to the argument of the first
cosine term, which does not change the value of the cosine function when k is
an integer. This gives

cos(n(N -1)k/N) = cos(n(N -1)k/N + 2kn) = cos(n(N + 1)k/N) (9.46)

The second cosine term simplifies to the same value since

cos(n(N -1)k/N + 2nk/N) = cos(n(N + 1)k/N)

368

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

The system function for the second-order resonators is then given by

1-z- 1
HR,k(Z) = Bk 1- 2 cos(2:Trk/N)z-1 + Z-2 for 15,. k 5,. (N /2) - 1 (9.47)

where
(9.48)

is the real-valued gain of the kth second-order resonator. Note that the cosine
factor makes some of these gains negative. Including this result in the total
system function Eq. (9.37), we get the filter shown in Fig. 9.31b and
_1- Z-N [Ho(ej~ (N~-l u ( ) ]
H( Z ) N
1
-1 + ~
flR,k Z
-

(9.49)

k=l

Before describing the frequency-sampling design procedure and presenting examples, we describe the procedure for choosing the gains of the
resonators in the transition region.
SPECIFYING THE MAGNITUDE RESPONSE IN THE TRANSmON
REGION. With the frequency-sampling procedure, the filter magnitude

response IH(ei"')1 is equal to the values of the desired response IHo(ei "')1 at the
sampling frequency points, The values of IH(ei"') 1 at other frequencies are
interpolated from these sample point values. Because the combfilter/resonator combination generates an N-point duration of a sinusoid, that
could be modeled as being obtained with an N-point rectangular window
applied ~o an infinite-duration sinusoid, the interpolation is achieved with
sin(Nw )/sin( w/2) functions, as shown previously in Fig. 9.21(b). As such,
sharp discontinuities in IHo(eico)1 produce overshoots and leakage effects in the
filter magnitude response IH(ei"')I. To reduce these effects, the samples in the
transition region are used to taper the desired response gradually between the
passband and stopband. Since each nonzero sample value, IHo(ei2nkIN)I, means
the addition of a resonator, the number of such samples in the transition
region should be minimized. Unfortunately, determining these transition
sample values is a complicated procedure involving linear programming
optimization procedures (see Rabiner and Gold). The values of the transition
points also depend on the number of samples N and the cutoff frequency, It is
claimed that, in lowpass filter designs, a maximum stopband attenuation
between 44 and 54 dB can be achieved when one transition point is allowed.
With two transition points, from 65 to 75 dB can be obtained, and with three
transition points between 85 to 95 dB is possible. Rather than do the
complicated optimization, or to rely on tabulated values, we present an
alternate procedure that is more intuitive. The cost involved with using this
procedure is that it may require more resonators that the optimum procedure.
To avoid the problems associated with discontinuities and to take
advantage of the tolerances in the specification, a smooth transition from the

FlNITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

369

passband to the stopband is introduced to define IHo(eiW)I. From our

experience with windows, the raised-cosine function is a good choice, as shown
in Fig. 9.33. To take advantage of the passband tolerance and to allow some
margin, the raised cosine curve passes through the tolerance interval at the
passband frequency. To obtain some margin from leakage effects, the desired
response is set to zero before the edge of the stopband. The resonator gains
are then determined from the (N /2) + 1 samples of IHo(ei"') I over 0::5 0 ::5 Jr.
In practice, determining N, the number of sample values of IHo(eiW)!, is
not very difficult. A rule of thumb using the above procedure indicates that
one transition point is needed in the transition region for each 10 dB of
attenuation. If NT transition points are needed, the value for N, the number of
samples of !Ho(eiw )! over 0::5 0 < 2Jr, can be determined. The frequency
sampling period, denoted by {)O, is then equal to
{)O

= !Os - Op!/NT

(9.50)

and
N

= 2Jr/{)O.

(9.51)

If there is more than one transition region, the smallest one determines the
value for N. Once N has been determined, the sample values o( {!Ho(ei2JrkIN)I}

falling in the stopband are set to zero. Initially, all the sample values of
IH(e"")1
\-

-----

Ii
( ) ' - - - - - " - - - - - - - ' - - - - - ' - - . (0

It

IH(e"")1

nGURE 9-33
The desired magnitude response for lowpass
and bandpass filters is obtained by connecting the passband and stopband with a
raised-cosine function.

370

INTRooucnON TO OIGITAL SIGNAL PROCESSING

{IHo(ei2J<kIN)I} falling in the passband are set to one. Then, the value at the
inside edge of the passband is reduced to fall on the raised-cosine taper, but
still within the limits. This allows us to take advantage of the passband
tolerance. For passband to stopband transitions, let kp + 1 be the number of
the first sample inside the edge of the passband and ks + 1 be that of the first
sample inside the stopband. Then the cosine taper should be equal to 1 for
k = kp and zero for k == k.. If we let Nps = k. - kp, then the cosine sequence
traverses 1' radians, including the end points, in Nps samples. The downward
transition sample values are then defined by
IHD( ei2J<kIN)1 =0.5(1 +cos(1'(kN.- kp))
ps

for kpsk s k.

(9.52)

For stopband to passband transitions, let kp - 1 be the number of the first

sample inside the edge of the passband and ks - 1 be that of the first sample
inside the stopband. Then the cosine taper should be equal to 1 for k = kp and
zero for k = k . If Nsp:::: k p - k., then the cosine sequence traverses 1' radians,
including the end points, in Nsp samples. The upward transition sample values
are then defined by
IHD (ei2J<kIN)1 =
1- cos(1'(~: ks))
for ks s k s k p
(9.53)

o.s(

E:umple 9.2. Determining transition region values for a bandpass filter. Let us
consider the bandpass specification shown in Fig. 9.34, given by
passband: -1 s H(eiW)ldB S 0
stopband: IH(eiW)ldB < -60

for O.421t S c.o S O.61Jr

for 0 S c.o S O.16Jr and O.81Jr S c.o S Jr.

oJ[
lo~l~==:::=?-I:--__-::-:lI:-----,..-,IL_ _---,r===-,C:=:=::lL-+'"
o

O.l6rr

O.42rr

O.61rr

O.87rr

(j)

rr

{Ho(k)}

1.... IIll!I!IIII. ......

o

1 2 3 4 5 6 7 8 9 JO II 12 13 14 15 16 17 IS 19 20 21

2~

23

nGURE 9-34
Frequence curves for Example 9.2,-(a) Specification; (b) desired magnitude response
sequence.

FINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

371

The two transition region sizes are

OA2.1r - 0_16n = O.26n

and

O.87n - O.6ln = O.26n.

Since 60 dB attenuation is required, the number of transition points is NT = 6.

Then the frequency sampling period is {)w = O.26n/6 = O.043n, and N =
2.1r/0.043n = 46.2. We will" take the value rounded to the nearest integer, or
N=46.
We can now determine the sample numbers in the stopbands and passband.
The sampling period with N = 46 is equal to {)w = 2n/46 = 0.043n.
The passband samples extend from
k = O.042.1r/O.043n = 9.

n. which is rounded to k = 10

to
k = 0.61n/0.043n = 14.19, which is rounded to k = 14.
The stopband samples extend from k = 0 to k = 0.16n /0.43n = 3.72, or
k = 4 for the low frequency stopband, and from k = O. 87n/0. 043n = 20.2, or
k =20 to k = N /2 = 23.
The values of {lHo(k)l} can now be evaluated.
Low-frequency stopband:
IHo(k)1 =0

for 0 s k s 4.

Stopband to passband transition:

For the upward transition, k. - 1 = 4, or k. = 5, and kp - 1 = 10, or kp = 11. Then
Nop=6 and
IHo(k)1 = O.S( 1-

cos(n(k6-S))

for S s k s 11.

As a check, IHo (10)1 should fall within the passband limits, or

IHo(10)1 =0.5(1- cos(Sn/6

=0.933

(=-0.60 dB).

Passband:
IHo(12)1 = 1
Passband to stopband transition:
For the downward transition, kp + 1 = 14, or kp = 13, and k. + 1 = 20, or k. = 19.
Then Npo = 6 and
IHo(k)1 = O.S(1 + cos (

n(k; 13))

for 13 s k s 19.

As a check, IHo(14)1 should fall within the passband limits, or

IHo(14)1 = O.S(1 + cos(n/6 =0.933 (=-0.60 dB).

High-frequency stopband:
IHo(k)I=!l

for20sks23.

372

INTRODucnON TO DIGITAL SIGNAL PROCESSING

TABLE 9.4

Steps for performing frequency-sampUng FIR filter design

Step 1. Choose the value of N, the number of samples of the desired magnitude response
IHD ( ejW)1 over 0 $ W < 2;r
Step Z. Determine the resonator gains
Step 3. Implement the second-order resonators and compute the unit-sample response
Step 4. Verify the filter design
Step S. Compare the number of computations with the linear-phase FIR structure

FREQUENCY-SAMPLING FILTER DESIGN PROCEDURE. We now give the

step-by-step procedure for designing a linear-phase FIR filter having a
unit-sample response duration of N samples. These steps are summarized in
Table 9.4.

Step 1. Choose the value of N. The desired magnitude response IHa(eiU1 is

plotted over 0::; (J) ::;:fr. For responses having discontinuities and transition
regions, raised-cosine curves are drawn in the transition regions to connect the
passband and stopband. The value of N is determined by the width of the
transition region and the desired attenuation in the stopband. A rule of thumb
is that one transition point is needed for each 10 dB of attenuation. For a
continuous magnitude response having an arbitrary shape, the number of
samples should be large enough that a good approximation to the desired
response is obtained by connecting the sample points with a smooth curve.
Since the number of resonators in the digital filter increases with N, the
smallest possible value should be chosen. If N is an even integer, zeros in the
transfer function are permitted at both (J) = 0 and (J) =:fr. If N is odd, a zero
cannot appear at (J) =:fr, since it is not a sample point.
Step 2. Determine the resonator gains. Since the simplifying equations have
been worked out for the linear-phase response corresponding to an N-point
unit-sample response that is even-symmetric about n = (N -1)/2, we will use
this phase response. For this case, the resonator gains are given by Eq. (9.48).
Step 3. implementation of the second-order resonators. Each nonzero value
of IHo(k)l, for 0::; k::; N /2 - 1, results in a resonator. If the desired
magnitude response is zero for a particular frequency sample, no resonator is
necessary. The filter is implemented by cascading the comb filter with the
parallel set of resonators. The filter unit-sample response is determined by
applying a unit-sample sequence to the input. The linear-phase response of the
filter can be verified in the time domain. by noting that the unit-sample
response {hN(n)} , for 0::; n::; N -1, exhibits a point of symmetry at n =
(N - 1)/2. If a problem occurs, each resonator and comb filter combination
should be examined separately to determine that its unit-sample response is
symmetric about n = (N -1)/2.

FINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

373

Step 4. Verify the filter design. The values of the filter magnitude response
at the frequencies other than the sampling points should be determined_ This is
done by computing the magnitude of the Nd-point OFT of the N-point
unit-sample response {hN(n)} , where Nd> N. The value of Nd should be
chosen large enough that the OFT samples, denoted by {HNd(k)}, are a good
representation of the continuous-frequency transfer function H(ejo. For
example, if N =32 then Nd = 128 is a good choice. The Nd-point OFT is
computed by padding the N-point unit-sample response with Nd - N zeros.
Upon verification, the magnitude response specification may not be met
because the allowed limits are exceeded at the frequencies between the
sampling points. If this happens, the specification must be sampled more
densely by increasing the value of N and the design procedure repeated. If the
specification iscexceeded only slightly, the value of N may need to be increased
by only a small number.
Step 5. Comparison of computation counts. Count the number of multiplications that are required to determine if the frequency-sampling filter structure is
more efficient than using the linear-phase nonrecursive implementation,
determined from {hN(n)}. However, the parallel structure of the frequencysampling implementation allows the computations to be performed in parallel,
which may mean that it is-a faster implementation.
Eumple 9.3. Desi&n of FIR bandpass filter widI frequency-samp1iDg
metlaod. Let us consider the bandpass specification given by
passband:

-1 s IH( ei"')ldB s 0

for 0.42JJ S

sO.61JJ

stopband:
for 0 s w S 0.16JJ and 0.87JJ s w S JJ.
From Example 9.2, we found N = 46. The resonator gains are then given by

B-.. = 2IHD(k)1 cos(45JJk/46)

from Eq. (9.48). Using the fifteen nonzero values from Example 9.2, we get

Bk'; (1- cos(JJ(k6- 5)) cos(45JJk/46)

Bk = 2 cos(11. 74JJ) = 1.37

for k

for 6sk S 11.

=12

Bk = (1 + cos ( JJ(k ~ 13)) cos(45JJk/46)

for 13 S k S 19.

The digital filter was implemented and the unit-sample response {h46(n)} was
computed and is shown in Fig. 9.35(a}. As expected, it has even symf!1etry about
n =22.5.
As a verification, the 512-point DFf of {h46(n)} was computed by padding
with zeros and the magnitude response was determined. The passband and
stopband results are shown to satisfy the specification in Fig. 9.35(b}.

374

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

{h..,,(n)

FIGURE 9-35
Unit-sample response sequence of filter satisfying
specification in Example
9.3.

225

To implement the filter 15 second-order sections are needed, one for each
nonzero IHo(k) I If multiplications by -1 are ignored, two multiplications are
required for each resonator. An additional mUltiplication is required by the
comb-filter for each output point. Hence, the frequency-sampling structure
requires 31 multiplications per output point. Since {h~n)} is even symmetric,
the linear-phase FIR structure can be impjemented to provide the same result
with 23 multiplications per output point. In this case, the frequency-sampling
structure requires more multiplications.

In comparing the results of the previous example with those from

Example 9.1, we find that a 31-point unit-sample response is sufficient to
satisfy the specification. Since the frequency-sampling method produces a
46 element unit-sample response, it is not as efficient. In part, this inefficiency
is due to the procedure for specifying the values in the transition region. For
lowpass, highpass, bandpass and bandreject filters, the DFT/windowing
method is the method of choice. The frequency-sampling method really shines,
however, when a magnitude response having a specific shape is to be
implemented, as described in the next example.
Eumple 9.4. Desiping. Unear-witb-trequency magnitude response with the
freqaency-suapling method. Let us consider the magnitude response given by

for O:s (J) :S Jr.

The N-point unit-sample response will be specified to have even symmetry, so
that the resonator gain equation (9.48) can be employed. To apply the
frequency-sampling procedure, we need only sample this magnitude response at
(N/2) + 1 points, when N is an even integer, or at (N + 1)/2 points when N is
odd. Doing this, we get

IHo(k)1 = 2k1r/N, for O:sk :sN/2 and N even

or

IHo(k)1 = 2k1r/(N + 1), for O:sk:s (N -1)/2 and N odd.

In this example, we explore the differences produced by even and odd N. and by
varying the value of N.
N = 64: Recall that if {hNJn)} has even symmetry and N is an even
(

FlNJTB.IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

375

PASSBAND:
IH(e"")ldB

II

1/\
I

-I

-2~
~

-3~
~

-4

III

O.61n

O.42n

(I

It

STOPBAND:
IH(t"')ldB

IL-~~~

__________________________

~~~~Ill

0.1611

FIGURE 9036
Magnitude response of filter satisfying specification in Example 9.3.

number, then its magnitude response has a zero at CJ) = n. Proceeding blindly
without this knowledge, we sample IHD ( ei ") I at 33 points over O:s CJ) :S n. The
sample values are shown in Fig. 9.37(a). The resonator gains are then given by
Bk = 4k1C cos(63nk/64)/N

for O:s k:s 32.

This filter has been implemented and the unit-sample response is shown in Fig.
9. 37(b). The even symmetry about n = 31.5 is apparent. However, when the
magnitude response of this filt~r is determined, it is linear over most of the
frequency range, but becomes zero at CJ) = n. This is due to the zero dictated by
the symmetry of the unit-sample respone. Upon closer examination, we also find
that
B32 = 2n cos(63n/2) = 2n cos(3n /2) = O.
Since an even symmetric sequence having an odd number of points does not have
a zero at CJ) = n, we should try N = 63.

,-

376

IN'tRODUcnON TO DIGITAL SIGNAL PROCESSING

(a)

111111

{iHo(k)i}

.' " I II
II I 2

II [Ill 11111111111111.

. '"

32

II

(b)

{h ..(n)}

(c)
iH(e"")i
1t

~------------------------------~~m
FIGURE 9-37
Results for Example 9.4 when N = 64. (a) Input samples of magnitude response sequence;
(b) unit-sample response of filter; (c) filter magnitude response.

N = 63: IHD ( ei..) I was sampled at 32 points over O:s Q) < n. These sample
values are shown in Fig. 9.38(a). The resonator gains are then given by
~

,.

= 4b cos(621rk/63)/(N + 1)

for O:s k :s 31.

This filter has been implemented and the unit-sample response is shown in Fig.
9 .38(b). The even symmetry about n = 31 is apparent. The magnitude response of
this filter, shown in Fig. 9.38(c), is more linear over the entire frequency range.
Having been convinced that an odd N is appropriate, we next observe the
results for N = 33 and 15, shown in Fig. 9.39. As N is reduced, the deviation of
IH(ei"')I from the desired linear function is more apparent, especially near Q) = 0
and Q) = n, where IHo(eiW)I has sharp comers. The minimum acceptable value for
N is determined by the specified tolerances, which may be in terms of maximum
deviation of IH(ei"')1 from a linear function, for example. This case is examined in
the projects.

FINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

377

(a) (IHo(k)\}

"'1111[11111111[11[11111111111[. k

II 1 2
II

:1CJJI'

(J)

62

II

31

(c) IH(tI"')1

~--------------------------~-+{J)

nGUU 9-38
Results for Example 9.4 when N =fi3. (a) Input samples of magnitude response sequence; (b)
unit-sample response of filter; (c) filter magnitude response.

OBSERVATIONS ABour TIlE FREQUENCY-SAMPLING FILTER DESIGN

METIIOD. The system function of each, second-order resonator in the
frequency-sampling structure contains a (1- Z-l) term in the numerator. This
can be factored out of the total system function to simplify the filter. This

cannot be done in general, because the first-order resonator, when it exists,

does not contain this term. 1f the filter is implemented with a hardware
second-order section, this term comes automatically, and since it does not
require a multiplication, is implemented with little loss in time. If the filter is
implemented as a program, factoring out this common term may save an
instruction per resonator, and speed up ,the execution a little.

378

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

(a) N=33 results

{h))(n)}

.12

(b)
N=15 results

FIGURE 9-39
Results for Example 9.4 for different values of N. (a) N = 33; (b) N

= 15.

FINITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES.

379

One advantage of the frequency-sampling filter structure is that the

computations can be performed in parallel. If it is important to minimize the
processing time, each of the resonators can be implemented by a separate
hardware circuit and all such circuits can process the incoming data
simultaneously.
One disadvantage of the frequency-sampling structure is that the cancellation performed by the comb filter must be exact. It is this exact
cancellation at n = N that causes {hN{n)} to have a finite duration. Otherwise,
the residual values circulate in the memory of the loss less resonator forever.
With time, these residuals can accumulate and introduce significant errors in
the filter output. To prevent this from happening in practice, the zeros of the
comb filter are placed slightly inside the unit circle, at radius r = 1 - a, for
some very small a > 0.. The system function of the comb filter is then equal to
Hdz)

= 1- rNz- N

(9.54)

The poles of the resonator are then also located inside the unit circle to cancel
these zeros. With their poles inside the unit circle, these resonators become
lossy. This loss allows any residuals that occur to decay to zero eventually.
No discussion of FIR filtering techniques is complete without a description of how to perform the filtering operation with the FFf algorithm. This
topic is described in the next section.

9.6 USE OF THE FFT ALGORITHM TO

PERFORM FIR FILTERING
In the preceding sections, we described the design of FIR filters and found
that, to satisfy a demanding specification, the number of points in the FIR
unit-sample response may be quite large. Since this large number is directly
related to the number of multiplications required per output point, FIR filters
typically require large amounts of computation time. In this section, rather
than implement a digital filter structure, we perform the filtering operations in
the frequency domain with the DFf. To be preferable to digital filter
structures, this DFf approach must be faster. The fast Fourier transform
(FFI) algorithm, described in Chapter 4 as an efficient computational
procedure for evaluating the OFf, is the main reason that much current signal
processing is performed in the frequency domain.
To make the comparison between an FIR filter and the FFf algorithm,
let us consider the FIR filter to have the rnite-duration unit-sample response,
denoted by h(n), for 0::; n ::; NH - 1. To implement the filter in a direct form
structure; the output value at time n is determined by the convolution
y(n)

= ~l h(k)x(n -

k)

(9.55)

k=O

For each output point, NH multiplications and additions are required. If the

380

INfRODUcnON TO DIGITAL SIGNAL PROCESSING

input sequence {x(n)} contains N x elements, then the output sequence {y(n)}
contains Ny = NH + N x - 1 points. Since a multiplication is typically the single
operation that requires the most time to perform in a general-purpose
computer, we will use the number of real-valued multiplications required to
perform a filtering operation as a measure of the efficiency of a filtering
m~thod. To perform the direct convolution above, NH multiplications are
required for each output value, or NyNH multiplications for the entire output
sequence.
The convolution above can be converted into a multiplicative operation
by taking the Ny-point discrete Fourier transform of the data and unit-sample
response and computing
Y(k)

= H(k)X(k)

(9.56)

The output sequence values y(n), for OSn sNy -1, are then determined
from the inverse OFT of {Y(k)}. The following steps must be performed to
accomplish the filtering operation in the frequency domain.
Step 1. Computation of the Ny-point OFTs of the unit-sample response and
the input sequence:
H(k) = ~I h(n) e- j 2n:nJcINy

for Osk sNy -1

(9.57)

n=O

and
N~I

X(k) = L x(n) e- j2 J1:nkINy

(9.58)

n=O

Step 2. Calculation of the product of the two Ny-point OFTs:

Y(k) = H(k) X(k)

(9.59)

Step 3. Computation of the Ny-point inverse OFT:

1 Ny-I
y(n) = Y(k)

Ny

L.

ej2JrnklNy

(9.60)

k=O

Under special circumstances, the above three-step procedure can be

modified to reduce the multiplication count. For example, if the same filter is
to be applied to many different input sequences, the transformation needed to
find {H(k)} is performed only once. However, when processing very long,
possibly infinite-duration, input sequences, the proper value for Ny is not
evident. For this case, the input sequence can be broken up into blocks of data
having a constant size. The input data can then be processed block by block.
We now examine this procedure and determine the size of the block that
minimizes the number of computations per output point.
FILTERING WITH THE FFT BY' USING mE OVERLAP-AND-ADD
MEmOD. When performing a filtering operation in the frequency domain,

FINITEIMPULSE RESPONSE FILTER DESIGN TECHNIQUES

381

the size of the DFf to be computed was determined by the duration of the
output sequence Ny = NH + Nx - 1. The DFfs of both the input and unitsample response had to be computed to this size. For long input data
sequences, such as the sampled audio signals from CD systems, Nx is typically
much greater than NH. To compute the Ny-point DFfs, {h(n)} would then
have to be padded with Nx - 1 zeros, a large number that requires a long
computation time. As an alternate procedure, we can divide the input sequence
into smaller blocks of data and determine the block size that minimizes the
number of multiplications required per output point. When processing data
with digital filter structures, this problem does not arise, since the number of
computations per output point is a constant, being determined by the structure
of the filter.
Let the sequence {x(n)} have an arbitrarily long duration, such that
N x NH, and break it up into blocks of No samples, as shown in Fig. 9.40.
The mth block is then denoted by {xm(n)} and defined by
xm(n) =x(n - mNo )

= 0,

for

O~n ~No

-1 and m = 0,1,2, ...

otherwise

(9.61)

When processed by the FIR filter having a unit-sample response containing NH

samples, the output for the mth block, denoted by {Ym(n)} is given by
Ym(n)

N~l

L h(n - k)xm(k)

for

k=O

O~

n ~NH + N o -1

(9.62)

Since the processing is linear, the superposition principle can be applied in an

interesting way. The entire input sequence can be considered to be the
superposition of the individual data blocks, or
x(n) =

2: xm(n m

mNo )

(9.63)

If {y(n)} is the linear convolution of {h(n)} and the entire sequence {x(n)},
then it is equal to the sum of the outputs for each block, or

(9.64)
m

Since NH + No - 1 is greater than No, a time overlap occurs in the successive

{Ym(n)} sequences, as shown in Fig. 9.40. The superposition principle, by Eq.
(9.64), states that the overlapping sections should be added to produce the
desired output. Hence, this method is called the overlap-and-add method.
To compute the number of multiplications required per output point
using this method, we let My = NH + No -1 and assume that the My-point
sequence {H(k)} has been already computed. This is reasonable for applications in which the same filter will be used repeatedly on input data. Then the
steps listed in Table 9.5 must be performed. The corresponding number of real
multiplications for each step is also listed.
Each application of these three steps generates My samples. But, because

382

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

{X(T .,I'Ie"

II'

11'1111*

i, .,I'leII'10lI'N~.~!

,,-.IIlI,

.III.

'II

'[U'

JIlt..
11'

........... ............. ..

tIll.

r" n

t .....tUr...11l'!"""" .. " .. .......... .......... ..

{Yo(n)}

sa.

{y(n~}
'0

'1},*

NH+NQ-l

'1.

1'1' 1'1

.i1..Ub..'.IIIIlIll' rllIlIh,

,IlIr

ow

'111*

ttIlt.

.nIh'I. ..

....

"GURE 9-40
Relationships of sequences used in the overlap-and-add method.

of the overlap, the valid output sequence, or those samples that require no
further manipulation, increases by No points. If we let R denote the number of
multiplications per valid output point, we have

= [4My iog2 My + 4My + 4My log2 My)l No

= [4(NH + No :- 1)(1 + 210g2 (NH + No - 1))]/No
(

(9.65)

FINITE IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

383

TABLE 9.S

Steps and multiplication count for performing a filtering operation with the
FFf
Step

Number of real multiplications

1. Compute the My-point DFf of {xm(n)}

Z. Compute the product Ym(k) = H(k)Xm(k)
3. Compute the My-point inverse DFf of {Ym(k)}

4Mylog2 My
4My
4Mylog2My

Examining this last equation, we find that for a given value of FIR filter
duration NH , there is a block size No that minimizes the value of R. The curves
of the R values for different values of No are shown in Fig. 9.41 for different
values of N H
To determine whether the FIR filter or the FFf approach is more
efficient, the value of R should be compared to N H , the number of
computations per output point required by the general Direct Form FIR filter
structure. For linear-phase filters, the value of R should be compared with
NH/2 to account for the linear-phase structure.
R

SOO
1024

4()O

300
256

200

12M

100

FIGURE 9-41
curves showing R, number of multiplications per output point, as a function of the block size No,
with size of filter duration NH as a parameter.

384

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

Figure 9.41 should be interpreted in the following manner. For a given

N H , the value of R is observed to have a minimum for some value No. For
applying this overlap-and-add procedure with the FFT, No should be chosen

such that
No + NH - 1 = My = 2K

for some integer K.

(9.66)

The input data sequences {xm(n)} must then be padded with My - NH + 1

zeros to compute the My-point DFf sequence {Xm(k)}. For small values of
NH 64), the computational overhead required to calculate the DFf
sequences makes this procedure less efficient than using the FIR filter structure
directly. For large values of NH , the efficiency of the FFT algorithm is put to
full advantage, but the choice for the value No becomes important. If No is
much smaller than NH , then an excessive number of zeros must be padded to
{xm(n)} and the computation of the My-point DFT yields only a few points at
the output. As No approaches infinity, Eq. (9.65) indicates that R also
becomes infinite, so that a minimum R exists for some finite value of No. The
curves indicate that No = 4NH is a reasonable choice that will provide savings
in the computations when NH > 256.

9.7 SUMMARY
In this chapter, three methods for synthesizing FIR filters were presented. The
first, called the windowing method, synthesizes the filter from the unit-sample
response sequence. This approach is most useful for designing an FIR filter
from a known unit-sample response sequence. One application of this case
occurs when replacing an existing analog filter with a digital FIR filter. The
impulse response of the analog filter can be sampled to produce the desired
unit-sample sequence. The second approach, call the DFT /windowing method,
the unit-sample response is determined from the inverse discrete Fourier
transform of the samples of the magnitude response and an assigned phase
response. A window sequence can then be applied to shorten the duration and
take advantage of the tolerances in the specification. When a linear-phase
response is specified, the special FIR structure can be implemented to reduce
the multiplication count by almost one-half. The third approach,. the
frequency-sampling method, synthesizes the filter directly from samples of the
magnitude response. This method results in a filter structure that consists of a
comb filter followed by a bank of parallel resonators. This parallel structure
can perform the computations on the input data simultaneously, thereby
reducing the computation time. This approach is most useful for designing a
filter to have a magnitude response with a specific shape.
We also considered performing the filtering operations in the frequency
domain with the FFT. The overlap-and-add method for performing a filter
operation on long data sequences divides the input data sequence into blocks.
The block size that minimizes the number of multiplications per output point
was determined. The number of ~uJtiplications can then be compared to that

FINITEIMPULSE; RESPONSE FILTER DESIGN TECHNIQUES

385

required for digital filter structures to perform the same function. It was
observed that the FFf method is not always the most efficient approach to
processing data.

FURTHER READING
Antoniou, Andreas: Digital Filters: Analysis and Design, McGraw-Hili, New York, 1979.
Bergland, G. D.: "A guided tour of the fast Fourier transform," IEEE Spectrum, 6: 41-52 (1969).
Burrus, C. S" and T. W. Parks: DFT / FFT and Convolution Algorithms, Wiley-Interscience, New
York, 1985. [This book provides a description of the various algorithms to implement a
discrete Fourier transform. Program listings are given in FORTRAN and assembly
language for the Texas Instruments TMS32010 signal processing chip.]
Cooley, J. W., and J. W. Tukey: "An algorithm for the machine computation of complex Fourier
series," Math. Comp., 19: 297-301 (1965).
Gold, Bernard, and Charles M. Rader: Digital Processing of Signals, McGraw-Hill, New York,
1969.
Hamming, R. W.: Digital Filters, 2d ed., Prentice-Hall, Englewood Cliffs, NJ, 1983.
McClellan, J. H., T. W., Parks, and L. R.,Rabiner: "A computer program for designing optimum
FIR digital filters," IEEE Trans. Audio Electroacoust" AU21: 506-526 (1973).
Oppenheim, Alan V., and Ronald W. Schafer: Digital Signal Processing, Prentice-Hall,
Englewood Cliffs, NJ, 1975.
Peled, A., and B. Liu: Digital Signal Processing: Theory, Design and Implementation. Wiley, New
York, 1976.
Rabiner, Lawrence R., and Bernard Gold: Theory and Application of Digital Signal Processing,
Prentice-Hall, Englewood Cliffs, NJ, 1975.
Taylor, F. J.: Digital Filter Design Handbook, Marcel Dekker, New York, 1983. [Advanced
topics, also covering ladder and wave filters. Includes hardware and &oftware
considerations]
Williams, Charles S.: Designing Digital Filters, Prentice-Hall, Englewood Cliffs, NJ, 1986.
[Beginning level text for juniors and seniors, relying very little on previous mathematical
knowledge.]

PROBLEMS
9.1. Consider the infinite-duration unit-sample responses given below. We have
determined that either 5-point or 6-point rectangular windows are sufficient to
implement linear-phase FIR filters that satisfy the specifications. For each
response, indicate the appropriate window duration and draw the causal
linear-phase FIR filter structure that performs multiplications most efficiently.
(a) h(n)=O.9"forn~O
h(n) =h(-n) for n <0
(b) h(n) = 0.9" for n > 0
h(O) = 0
h(n) = -h( -n) for n < 0
(c) h(n)=O.9"forn~O
h(n) = h( -n + 1) for n < 0
9.2. The triangular window can be analyzed as the convolution of two smallerduration rectangular windows. Compute the transfer function of the 5-point

386

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

triangular window by performing the multiplication of the transfer functions of

the two 3-point rectangular windows given by
for -1 ~n
otherwise

9.3.

9.4.

9.5.

9.6.

9.7.

9.S.
9.9.

Sketch and compare the transfer functions of the triangular and rectangular
windows.
Consider the ideallowpass filter having the cutoff frequency We = :r/2.
(a) Compute the desired unit-sample response {ho(n)}.
(b) Use a 3-point rectangular window and compute the transfer function of the
windowed sequence. Compare with the ideal response.
(c) The same sequence is obtained when a 5-point rectangular window is applied
(why?). Indicate by using sketches of the transfer functions why the same
windowed magnitude response is obtained.
Give simple examples that illustrate the four linear-phase responses given in
Table 9.2 by sketching representative unit-sample response sequences and their
corresponding phase responses.
Indicate the possible symmetries that the unit-sample response can have to
approximate a lowpass filter, a highpass, a bandpass filter and a bandreject filter.
Consider the locations of desirable and undesirable zeros of the transfer function.
Write the analytic expression for the set of variable-duration Hamming windows,
having duration N (even integer) and having the point of symmetry at
Ns = (Nd - 1)/2, where Ns is not an integer (N < Nd)'
Consider the desired magnitude response given by IHD(ei"')1 = Isin wi. Using the
Off method, determine the following 8-point unit-sample responses.
(a) Even-symmetric about n = 3.5
(b) Anti-symmetric about n = 3.5
(c) Even about n = 0 (except for unmatched point at n = 4)
(d) Odd about n = 0 (except for unmatched point at n = 4)
Design the four possible 8-point linear-phase FIR filters having the magnitude
response sequence given in Prob. 9.7 by using the frequency-sampling method.
Derive the transfer .t~i1ction of the second-order resonators used in the
frequency-sampling method for the case of an antisymmetric unit-sample
response.

COMPUTER PROJECfS
9.1. 'The Gaussian window
Object. Investigate the spectrum of the Gaussian window.
AnaIytk results. Consider the Gaussian window function defined by

w(t) = (2:rcrtll2 exp[-f/(2cr)]

for all t

Generate a 31-point window sequence {wG(n)} that is symmetric about n = O. Find

the value of 0 such that these 31 points are samples of the Gaussian curve for
-30~t~30.

FINITEIMPUlSE RESPONSE FILTER DESIGN TECHNIQUES

387

Computer results
(a) Compute the log-magnitude spectrum of this window and compare the
main-lobe width and sidelobe level with the windows given in Table 9.1.
(b) Try to invent your own window sequence that has better characteristics than
those above. Do this by modifying the values of ope of the window sequences
given in this chapter. An analytic form for the window sequence values is not
necessary.
9.2. Variable-duratioo windows
Object. Write a subroutine to generate a variable-duration window sequence.
This subroutine will be used in the following project. Write the subroutine

HAM(WH, NW)
where
WH (real output array) of size 128 that contains a Hamming window sequence
centered at N = 64 and has duration equal to NH. The rest of the WH
array is equal to zero.
NW (integer input) is the duration of the window. NW must be an odd
integer.
This WH array is appropriate for windowing unit-sample sequences that are
symmetric about a point. The windowing operation is accomplished by multiplying
HD(N), the 128-point approximation to the desired unit-sample response by the
128-point WH(N) array.
Repeat for the Blackman window by writing

BLACK(WB, NW)
Compoter results Verify the operation of HAM and BLACK by computing and
plotting the window sequences for NW = 31,63 and 127.
9.3. Linear-phase FIR filter design using the DFf/windowing method
Object. Design a causal linear-phase FIR filter that satisfies ore of the magnitude
response specifications given in Appendix B using the DFf/windowing procedure.
AnaIytk~

(a) Sketch the (linear, not log) magnitude response that is obtained by extending
the passbands.
(b) Choose the type of symmetry (even or odd) and the point of symmetry (on a
sample point or between sample points) of the causal unit-sample response.
Justify your choice.
(c) Sketch the appropriate linear-with-frequency phase response curve for the
filter. The principal value of the phase is not needed and should not be
sketched.
Computer results
(a) Compute and plot the 128-point approximate desired unit-sample response
{hd(n)} that is symmetric about n = 63. Ho(N) is symmetric about N = 64.
(b) Truncate {hin)} by using an Nw-point Hamming or Blackman window.
Justify your choice or compare the two.
Starting with Nw = 64, truncate Ho(N). Compute and plot the log-magnitude
response of the windowed unit-sample response. Reduce Nw until the specification
is no longer met. Provide log-magnitude response plots for all the values of Nw
used, i,n the order that they were tried, to indicate your strategy.

388

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

(c) For the filter design that meets the specification and has the smallest
unit-sample response duration, do the following:
(i) determine HMAX and scale the coefficients so that the maximum of the
magnitude response is equal to one.
(ii) plot the unit-sample response and indicate the symmetry on the plot.
(iii) plot the phase response and sketch the linear trend on the plot.
(iv) sketch the block diagram of the FIR filter structure.
9.4. Unear-phase FIR filter design using frequency-sampUng method

Object. Design a causal linear-phase FIR filter that satisfies one of the magnitude
response specifications given in Appendix B using the frequency-sampling procedure.
Analytic results
(a) Sketch the (linear, not log) magnitude response that is obtained by assigning
the response in the transition region to be a raised-cosine function that
connects the passband and the stopband.
(b) Determine the value for N, the duration of the filter unit-sample response.
Computer results
(a) Implement the comb filter and resonators. Compute and plot the unit-sample
response. Indicate the symmetry on the plot.
(b) Compute and plot the 128-point log-magnitude response of the filter. Circle
the points on the response curve that were specified in the frequency sampling
procedure.
(c) Sketch the filter block diagram. Compare the number of multiplications of this
frequency-sampling structure with that required for the nonrecursive FIR
linear-phase structure having the same unit-sample response.

9.S. Designing a filter with specific magnitude response shape with the fiequency.
sampHng method
Object. Design a filter whose magnitude response approximates a given analytic
form using the frequency-sampling method.
Implement the filter that approximates the desired magnitude response given
by one of the following:
(1) Linearly decreasing response
IHo(ej")1 = a(l- co/,,)

forOscos"

(2) Exponential response

IHo(ei")ldB = e- ....

forOscos"

(3) Gaussian response centered at co = a"

IHo(eJ")1 = exp[-(co -

a"YJ

for Os co s"

Let a be the number determined by placing a decimal point before the first
nonzero digit in your student ID number. For example if ID = 0123, then
a=0.123.
Error criterion. The filter magnitude response is allowed to deviate from this
specification by 2% of the maximum of the magnitude response.

FINITE-IMPULSE RESPONSE FILTER DESIGN TECHNIQUES

389

Analytic results
Sketch the desired magnitude response and indicate the limits that must contain
the final filter magnitude response.
Computer results
(a) Compute and plot
IHD(k)1 = IHD ( ej 2>rkll28)1

for 0 :S k

:S 64.

(b) Starting with N = 32, implement the comb filter and resonators. Compute and
plot the unit-sample response.
(c) Reduce the value of N. For each reduction, compute the 128-point magnitude
response and compare with the specification by calculating the error sequence
given by
E(k) = IHD(k)I-IH(k)1

for 0 :S k

:S 64.

The error element having the largest magnitude, denoted by IElm.,., should be
found. The filter design is successful when the minimum N is found, for which
IElm"" <O.02a.
(d) For the filter design that meets the specification and has the smallest number
of resonators, sketch the filter block diagram. Compute and plot the 12S-point
magnitude response of the filter. Circle the points on the response curve that
were specified in the frequency sampling procedure. Compute and plot the
unit-sample response. Compare the number of multiplications with that
required for the nonrecursive FIR linear-phase structure.

CHAPTER

10
FINITEPRECISION
EFFECTS

10.1 INTRODUCTION
Finite-precision effects are caused by the errors due to representing numbers to
a finite accuracy. These effects were previously noted in Chapter 8, when the
computed pole locations of IIR filters were different from those desired, the
difference being due to round-off effects in the calculations. These effects are
also important when implementing a filter structure on a computer or digital
hardware, which represents numbers as a finite set of binary values. In this
case, the input sequence and coefficients values must be represented in such a
finite-precision number system, as well as the results of the arithmetic
operations of addition and multiplication.
We start by defining the two most common forms of representing
numbers in the computer, fixed-point and floating-point notation. Then, three
effects that occur in finite-precision number systems are considered. The first is
the conversion of continuous-valued, or infinite-precision, signal values into
finite-precision values. This occurs in the analog-to-digital conversion operation. This effect is shown to be similar to adding noise to the input signal
values. The second effect occurs when the filter coefficient values, computed to
infinite precision in theory, are constrained to equal one of the allowable
numbers represented in the computer for the filter implementation. As a
result, the singularities in the z plane are limited to lie on a grid of allowable
390

FINITE PRECISION EFFECTS

391

locations. The deviations that are produced in the resulting transfer function
are considered. The third effect is the truncation that results when the sum or
product of two finite-precision numbers is itself represented as a finiteprecision number. If the filter being implemented is a nonrecursive structure,
this effect is similar to adding noise to the output of each multiplier. However,
if the filter has a recursive structure, the truncation of the product may cause
the output to exhibit undesired offsets or oscillations, called limit cycles.
Since these finite-accuracy effects are nonlinear and random, their
analysis is very complicated. To provide some insight, special cases are
analyze(l to indicate the type of behavior that can be expected. Procedures for
simulating and observing these effects on a computer are described. The intent
of this chapter is to present some points to be aware of when designing filters
in a finite-precision number system, but not to provide a comprehensive
analysis of these effects. Such an analysis is most difficult and beyond the scope
of this text. In the cases when the analysis can be performed, it requires results
from nonlinear system theory and probability theory. The analysis can be
found in the references at the end of the chapter.

10.2 FINITEPRECISION NUMBER

REPRESENTATION WITHIN THE
COMPUTER
In a computer, or any other digital hardware structure, numbers are
represented as combinations of a finite number of binary digits, or bits that
take on the values of 0 and 1. These bits are usually organized into bytes
containing 8 bits, or words containing 16 bits, but with 32-bit words becoming
increasingly common. We consider two common forms that are used to
represent numbers in a digital computer: fixed-point and floating-point
notation.
FIXED-POINT NOTATION. In fixed-point notation, the partition between
whole numbers and fractions, the binary-point, is constrained to lie at
particular, or fixed" position in the bit pattern, as shown in Fig. 10.1. The first
bit, called the sign bit, is reserved for indicating the sign of the number, usually
o for positive and 1 for negative. The magnitude of the number is then
expressed in powers of 2, with the binary point separating the positive

b l bits for integer

binary point
~ b l bits f~r fraction
---,'.1

,----_ _ _-A-.l _ _ _

SI I I I I I I I I I I I I I It
sign bit

least Significant
bit

FIGURE 10-1
Allocation of bits in a word for fixed-point representation (l6-bit word is shown).

392

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

exponents, including zero, and the negative exponents. For example, the
decimal value of the binary number 01.101 2 is determined as
01.101 2 = ( + )(1 x 2) + (1 x 2- 1)

+ (0 x 2- 2) + (1 x T

= 1.625\0'

The precision of the number system is defined as the increment between

two consecutive numbers and is determined by the value of the right-most bit,
called the least significant bit. In the number system given in the previous
example, a number can be expressed to the nearest 2-3 or 0.125. The range of
the number system is defined as the interval between the most negative and
most positive numbers that can be expressed. The binary point determines a
tradeoff between precision and range. When the binary point is defined to lie
at the right of the bit pattern, there is no fraction and only integer values are
allowed. This corresponds to the INTEGER variable in FORTRAN. For
digital filters, we will find that it is most convenient to have the binary point lie
one or two bit positions to the right of the sign bit, thus allowing fractions to
occur.
Conversion of decimal numbers into binary numbers. In converting a decimal
number into a binary number, the integer part and fraction must be converted
separately. Only positive numbers will be considered, with the procedure to
represent negative numbers having a particular magnitude being described
later. Converting the integer part involves repeated division by two, with the
remainder at each division step forming the binary number. The binary
number is formed by starting at the binary point and proceeding left-ward,
with the final nonzero remainder being the most significant bit. Every finite
decimal integer can be represented in a finite number of bits.
The conversion of the decimal fractional part involves repeated multiplications by two, with the integer part of the product at each step forming the
binary number and the fractional part. being used in the continued multiplications. The binary number is formed by starting at the binary point and
proceeding right-ward, with the final entry being the least significant bit in the
binary representation in the number system. Not all decimal fractions can be
represented in a finite number of bits. This process may need to be terminated
when the number of bits available to express the fraction is exceeded.
In converting a decimal fraction into a finite-precision binary number, the
least significant bit can be generated by truncation or by round-off. In
performing truncation, the magnitude of the infinite-precision number is
converted into finite-precision by using the available number of bits. The
process terminates after the least significant bit has been determined. In
round-off, the conversion proceeds for an additional bit beyond the least
significant bit. If this additional bit is a one, the least significant bit is
incremented. If the additional bit is zero, the process terminates and produces
the same result as truncation. The round-off representation is more accurate,
but requires more operations.

FINITEPRECISION EFFECTS

393

Example 10.1. Decimal to binary conversion. Convert 6.8610 into binary

notation having 8-bits of which four bits are used to represent integers, including
sign, and four bits represent fractions.
Integer part:

remainder

~=3
~=1
~=O

1
1

Hence, 6)0 = 01102

Fraction. To compare the results obtained with truncation and round-off,
we will convert the number to 5 bits:
integer part

0.86 x2= 1.72

0.72 x 2=\:44
0.44 x 2 = 0.88
0.88x2= 1.76
0.76 x 2= 1-52

1
1

o
1
1

(least significant bit)

Hence,
0.8610 = 0.1101 2 { = 0.8125 1O} by truncation
= 0.11102{= 0.87510) by rounding
Combining these two results, we have
6.86\0 = 0110.1101 2 by truncation
= 0110.11102 by rounding

2's complement notation. One common fixed-point number system currently

used is 2'$ complement notation, because. of its efficiency for performing
computations. The bit pattern of a negative number is obtained by starting
with the magnitude expressed as a positive number. Then all the bits are
complemented and a binary one is added to the least significant bit.
Example 10.2. Decimal to binary conversion of negative numbers using 2's
complement notation. Convert -6.8610 into binary notation having 8-bits of
which four bits are used to represent integers, including sign, and four bits
represent fractions.
First, the magnitude of the number is converted into binary notation. Using
the round-off results from Example 10.1, we have
6.86)0 = 0110.11102
Complementing this result bit-by-bit, we get l00LOOOb Adding 0.00012 to this
result, we have
-6.86\0 = l001.00102( = -6.875 10)
Table 10.1 shows the entire 2's complement repertoire of patterns and their
decimal equivalents for 2, 3 and 4-bit number systems.

394

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

TABLE 10.1

Numbers expressed in 2's-complement fixedpoint notation

Binary .amber

latecer values
(biliary polIIt
at rigIat)

Fractious
(biliary poiat
after sign bit)

2-bit.8IIIbersystem
01
00
11
10

1
0
-1
-2

0.5
0
-0.5
-1

3-bit .....ber system

011
010
001
000
111
110
101
100

3
2
1
0
-1
-2
-3

-4

0.75
0.5
0.25
0
-0.25
-0.5
-0.75
-1

4-bit .1IIIIber system

0111
0110
0101
0100
0011
0010
0001

0000
1111
1110
1101
1100
1011
1010
1001
1000

7
6
5

4
3
2
1
0
-1
-2
-3

-4
-5
-6
-7

-8

0.875
0.75
0.625
0.5
0.375
0.25
0.125
0
-0.125
-0.25
-0.375
-0.5
-0.625
-0.75
-0.875
-1

FLOATING-POINI' NOTATION. The second form of expressing numbers is

floating-point notation, also called real notation. In this case, a number x is
expressed in the form
(10.1)
where m is the mantissa and e is the exponent. The notation is conventionally
defined to have 1/2 s 1m I < 1. For example, the number 1.5 is expressed as
(0.75)21 , Both m and e are express~d as fixed-point numbers, using a constant

"

FINITEPRECISION EFFECfS

5Sign of mantissa

I I I II I

395

5Sign of exponent

I II

L -________~-----------JA~_y--~

mantissa

exponent

ftGURE 10-2
Allocation of bits in a word for floating-point representation (l6-bit word is shown).

pattern within the word, as shown in Fig. 10.2. If the word contains b bits, let
b m be assigned to the mantissa and be to the exponent. Since b = be + b m , a
hardware designer must decide on the bit allocation between the mantissa and
exponent. The number of bits assigned to be determines the range of the
number system, while the number of bits assigned to b m determines the
precision. A common compromise is to set bm = 3b/4. In many minicomputers,
two 16-bit words are combined to express a floating-point number as a 32-bit
word, of which b m = 24 and be = 8.
In the next section, we consider the effect of representing the infiniteprecision values of a discrete-time sequence in such a finite-precision system.

10.3 QUANTIZATION ERROR IN

ANALOG-TO-DIGITAL CONVERSION
As described in Chapter 3, a sampling process converts a continuous-time
signal {x(t)} into a discrete-time sequence {x(n)}. In practice, this sampling is
performed with an analog-to-digital converter (ADC) that represents the
sampled value with a finite number of bits, usually in fixed-point notation.
Hence, the ADC also simultaneously converts the continuous-valued sample
x(n) into one that has been quantized, denoted by xq(n). The value of xq(n) is
assigned to one of the numbers represented by the finite number of bits in the
ADC. The process of converting the x(n) into xq(n) can be modeled by the
quantizer characteristic shown in Fig. 10.3. The quantizer range is expressed in
terms of 2's-complement numbers that span the amplitude range from - M to
M - 1 in K steps, each of size fl. = 2M I K. If the quantizer contains b bits,
K=2 b and
fl. = 2M

2b

(10.2)

For a given quantizer range, increasing the number of bits reduces the step
size, making the staircase pattern approach a straight line.
There are two types of errors that can be introduced by such a quantizer.
The first is granular error, which occurs when Ix(n)1 < M, and is denoted by
Eg(n), where
(10.3)
is always less than the step size fl.. The
Note that the magnitude of this en:pr
<

396

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

0111

0110
0101
0100

5
4

0011
0010

------~------------~~~----------+_~~n)

-M

-7

1l0I
1100
lOll
1010
1001

-8

1000

-4
-5
-6

fiGURE 10-3
Quantizer characteristic for a 4-bit quantizer.

second error occurs when Ix(n)1 > M, when the amplitude exceeds the
quantizer limits, and is called clipping error, denoted by Ec(n), where
_{x(n)-M -1
Ec(n) - x(n) + M

for x(n) > M-1

forx(n)< -M

(10.4)

Very large, theoretically infinite, clipping amplitude errors are possible,

making the analysis of these errors rather complicated. In the analysis below,
we set the quantizer limits large enough, or equivale~tly scale the input
sequence to have small enough values, so that this clipping error does not
occur. In this case, only the granular error is important.
To make optimum use of the quantizer, it is usually desired to have the
signal values distributed throughout the entire range of numbers represented
by the bits in the ADC. This is typically done by adjusting the maximum value
of the input signal magnitude so that it is equal to the maximum magnitude of
the number system. Signals encountered in practice, such as speech and sensor
outputs, are typically random in that their exact analytic form is not known.
Some random signals tend to have a large dynamic range, such that the signal
power is large at some times and very small at others. For such signals, a
different quantization strategy is usually employed, in which the quantizer
range is matched to the average power level of the signal. The relevant
parameter in this case, is the root-mean-square (rms) value of the signal,
denoted b~ X rms The value of X~ms can be calculated from the power in the

FlNITEPRECISION EFFECTS

397

N-point sequence {x(n)} by

1 N-l

XZrms = -

2: x2~n)

(10.5)

n=O

The rms values for many common signals, such as telephone speech and
recorded music levels on tape are usually specified or determinable by simple
measurements.
For speech processing applications, it is not uncommon to have the
quantizer limits set equal to 4Xrms ' This setting represents a compromise in
which an accurate representation of the small-amplitude signals, which occur
more commonly, is achieved at the expense of large errors for more rarely
occurring large-magnitude signals. Nonlinear quantizer characteristics are also
common, in which the step size varies with the magnitude of the input signal.
A further discussion of nonuniform and also time-varying quantization
schemes can be found in the reference by Jayant and Noll.
To analyze the quantization operation, we employ the additive error
model shown in Fig. lOA. The quantized value xq(n) is considered to be equal
to the true value x(n) plus a random additive noise term Eg(n), or
xq(n) =x(n)

+ Eg(n)

(10.6)

When matched to a random signal having a known average power level, the
step size t:. is equal to

t:. = 8Xnns
2b

(10.7)

The quantization error Eg{n) is commonly assumed to be random and

uniformly distributed between t:./2. The noise power, or variance, can then
be shown to be equal to t:.2/12. The ratio of the signa.J power to that of the
noise, called the signal-to-noise ratio, denoted SNR, is then given by
SNR = X;"'s

t:. 2/12

= 12 22b
64

(10.8)

The SNR is observed to improve exponentially with the number of bits in the
quantizer. Expressing the SNR value in decibels, we have
12 22b) = 6b - 7.3
SNRdB = 10 10glO ( 64

x(n)--~

FIGURE 10-4
Additive noise model for quantizer.

. (10.9)

398

INTRODUcnON TO DlGITAIrSIGNAL PROCESSING

This last result states that the signal-to-noise ratio improves by 6 dB for every
bit that is added to the ADC. The effect of the quantization error on the SNR
for a digital audio system is illustrated in the example below.
Example 10.3. Quantization error in a digital audio system. Let the input be
{x(n)} with an rms value equal to Xrms and let the quantizer limits span 4X""s.
If a 16-bit quantizer is employed, a SNR of greater than 88 dB can be achieved.
The current claim for commercial digital audio systems is approximately 84 dB.
The difference is due to the thermal noise in the analog amplifiers and the
digital-to-analog conversion, which were not considered in the analysis.

EFFECf OF QUANTIZATION NOISE AT THE FILTER OUTPUT. The quantization error in the ADC process was modeled as the addition of a noise
sequence to the infinite-precision sample values. To examine the contribution
of this noise sequence to the output sequence produced by a linear digital
filter, we apply the quantized signal to the filter and can express the output
sequence as
{Yq(n)}

= {xq(n)} * {h(n)}
= {x(n) + Eg{n)} * {h{n)}

(10.10)

Because the filter is linear, we can apply the superposition principle to get
{Yq(n)}

= {x(n)} * {h(n)} + {Eg(n)} * {h(n)}

={y(n)} + {ve(n)}

(10.11)

The sequence {y(n)} is the output that would have been observed for
infinite-precision samples and {ve (n)} is the output produced by the sequence
of noise values. These output noise values can be written explicitly as the sum
of the random errors scaled by the terms of the unit-sample response as
v.,(n)

""
~ h(k)Eg(n - k)

(10.12)

k=-co

Adding a sequence of random numbers produces a random result, about which

little can be said in general. However, if the random error sequence consists of
independent values having the same variance 0;, then the variance of the
output ~ is equal to
""

a~ = 0; ~ h 2 (k)

(10.13)

k=-oo

This simple relationship allows us to compute the output noise power in terms
of the input noise power and the filter unit-sample response.
Example 10.4. Noise variance at the output of the 3-sample avenger. The
difference equation of 3-sample average is given by
y(n)= l[x(n

+ 1) +x(n)+x(n -1)]

FINITEPRECISION EFFECTS

399

If the input sequence has been quantized to the set of integers, then the step size
is 1. If the range of the quantizer is large (> 1(0) and the input amplitude has
been matched to the quantizer, the quantization errors can be considered to be
= -b.. The errors can be
uniformly distributed. The error variance is then
considered independent if consecutive values of {x(n)} tend to occur in different
levels of the quantizer in a random fashion. The noise variance at the output is
then

a;

Example 10.5. Noise variance at the output of the 6rst-order recursive

filter. The difference equation of the first-order recursive filter is given by
y(n) = ay(n -1) +x(n)

If the input sequence has been quantized to the set of integers, then the step size
is 1. Under the same conditions as in Example lOA, the noise variance at the
output is then
~

o~ = 0;

2: a

= 1/[(12)(1 -

)]

k=O

Having considered the quantization of the input sequence, we now tum

our attention to the quantization of the coefficient values in the digital filter.

10.4 EXPRESSING COEFFICIENTS IN

FINITE PRECISION
When a digital filter is implemented in hardware, the coefficient values must be
quantized to the accuracy of the digital number system employed in the
hardware. In this section, we show that the effect of this quantization is to limit
the locations of the poles and zeros in the z plane to a grid of allowable
positions. If the desired singularity location does not lie on a grid point, then
the shift that must be made in these singularity locations causes the transfer
function H(e jtJJ ) of the hardware filter structure to be different from that
anticipated when the design was performed in infinite precision. For FIR
filters, this deviation occurs in the locations of the zeros. For IIR filters there is
a similar deviation in the pole locations. However, for filters having poles that
lie close to the unit circle, these deviations may even cause the IIR filter to be
unstable, as shown in the following example.
Example 10.6. Quantization of 11K filter coefficients. Let

H(z) = (1- 0.901z

1 - 1. 844z -I

1)~1_ O.943z
~ 0.849643z

I)
2

(cascade form)
(direct form)

If the coefficients are quantized by rounding to the nearest 0.05, then the cascade

400

INfRODUCI10N TO DIGITAL SIGNAL PROCESSING

form becomes
1

H(z) = (1- O.90z 1)(1_ 0.95z 1)

which has poles at z = 0.9 and 0.95. The resulting filter is stable. However, the
Direct form becomes
1

H(z) = 1-1.85z

+ 0.85z-2

which factors into poles located at z =0.85 and 1.0. The pole lying on the unit
circle makes the filter unstable. This example illustrates that the sensitivity of a
filter to coefficient quantization increases with the order of the filter.

nR FILTER. Let us consider the second-order FIR filter system function

factored into the form to show the zero locations in the z plane explicitly, or

H(z) =A(l - r ei8z- 1)(1- r e- i9z- l )

=A(l- 2r cos( 0) Z-l + ,.zZ-2)

(10.14)

where A is a gain constant. When represented in a finite-precision number

system, the coefficients 2r cos(O) and ,.z can take on only certain values,
determined by the precision of the number system, similar to the quantization of sample values in the ADC described in the previous section. To
illustrate the effects of this quantization, let us consider the zero locations to be
in the right half of the z plane, and quantized to 4 bits. The positive values are
listed in Table 10.2.
The effect of quantizing the coefficients on the system function is most
evident in the z plane. Recall that the coefficient r2 indicates that the
singularity lies on a circle defined by /z / = r. Quantizing ,.z to have only a set of
values defines a set of circles in the z plane on which the singularity must lie,
as shown in Fig. 10.5(a). The coefficient 2r cos( 0) defines a vertical line in the
TABLE 10.2

Allowable coeBident values for second-order filter with 4-bit

quantization
Radios of circle
in z plane

Bit paHera

Coellicient values
(Zr cos(8) or,z)

0000

0001
0010
0011
0100
0101
0110

0.125
0.25
0.375
0.5
0.625
0.75
0.875
1.0

0.354
0.5
0.611
0.707
0.791

0111
(multiplication by 1)

0.866
0.935
1.0

FINITEPRECISION EFFEcrs

401

(0)

1m

JUHM
jU7!17

Izl=l
jus
jU3S

0 ~-----+--~~~~~-Re
0.5
(b)

1m

r =0.5,9 =60"

I
1

Ir

=0.35,9 =45 0

r =0.28,9 =30

I
1
O~~------+----------Re

2r cos (9) (= 0.5)

FIGURE 10-5
Coefficient quantization effects in the z plane.
0) Quantizing ,2 produces circles of radius r.
b) Quantizing 2r cos( 8) produces vertical lines.

z plane passing though z =2r cos( 8), as shown in Fig. 10.5(b). This vertical
line indicates the pairs of allowable rand 8 values that can occur. Quantizing
this coefficient produces a set of corresponding vertical lines. The locations of
the allowable singularities for the second-order section can then fall only at the
points of intersection between these circles and vertical lines. The points of
intersection for the 4-bit quantization are shown in Fig. 10.6. Note that
quantizing the coefficients in this manner produces a non-uniform distribution
of the locations that tend to cluster around z = j. Alternate filter structures
have been proposed that produce a uniform distribution of the allowable
locations in the z plane. These are described in the references.
Because of the coefficient quantization, the position of the zero must be
shifted from that desired by theory to one of the four adjacent allowable
positions. The filter designer must choose one of these positions or increase the
number of bits in the finite-precision representation. This latter choice comes
with an increase in hardware cost and possibly slower operation, The following
example illustrates the implication of these results for FIR filter design.

402

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

1m

jo 866
jo 707

0
0
0
0
0

jo 5

jo 35

~_,-~-.~~~e-~-&~~---Re

0.5

FIGURE 10-6
Allowable locations in the right-hand
z plane for singularities when 4-bit
quantization is employed. Complexconjugate singularity locations are
not drawn.

Example 10.7. Effects of FIR filter coefficient quantization. Let us assume that
we want to implement a second-order FIR filter having 4-bit precision. Let the
desired filter transfer function be given by
HF1R(Z) = (1- O.9ei70"z-I)(1_ O.ge- i 7O"z-1)
The desired zero location at z = O.gei70" is shown in Fig. 1O.7(a). Because of
coefficient quantization, the zero position must be moved to one of the four
adjacent grid locations, labeled 1 through 4 in the figure, that are located at
z = O.866eJ64, Z = O.866e i73 , z = O.935ei66 and z = O.935eJ74 The magnitude
responses of these four zeros are shown in Fig. 1O.7(b). All the zeros produce a
valley in the magnitude response, but the location of the valley is related to the
angle of the zero, while the depth is related to the radius of the zero and the
interaction with the complex-conjugate zero.
DR FILTERS. For IIR filters, a similar effect occurs to that described above for

the FIR filters. However, since IIR filters can become unstable, an additional
consideration is required. As discussed in Chapter 5, an IIR filter can be
configured either as a cascade of second-order sections or in a direct form
implementation. In any case, the poles must be inside the unit circle for a
stable filter. For the cascade configuration of second-order sections, the
locations of the poles of each section can be easily verified or be made stable
by placing the poles at the set of allowable locations determined by the
quantization that lie strictly inside the unit circle. The desired pole must then
be transferred to the closest allowable pole location in the hardware
implementation. Because of this shift in the singularity location, a change in
either the resonant frequency or bandwidth occurs. As described in Chapter 6,
as the poles approach the unit circle, the changes in the magnitUde response
become increasingly drastic.
Example 10.S. Effects of DR filter coeffident quantization. Let us assume that
we want to implement a second-order IIR filter having 4-bit precision. Let the

FINITEPRECISION EFFEcrs

403

(a)

1m

DESIRED ZERO at z = O.ge J11r

40j
10

02

~--------------~Re

IH(eJ"')1

(b)

2.63

O~

____________________. -______

~,ro

FIGURE 107
Effect of quantization on
zero placement. a) Desired
zero location must be shifted
to one of lour closest neighbors when filter is implemented. b) Four magnitude
response curves corresponding to the four allowable
pole locations.

desired transfer function be given by

1
Hlla(z) = (1- 0.gei70"z-I)(1- 0.ge- i70"z-l)

The desired pole location at z = 0.geJ70" is shown in Fig. 10.8(a). Because of

coefficient quantization, the pole position must be moved to one of the four
adjacent grid locations, labeled 1 through 4 in the figure, that are located at
z = 0.866ei64", Z = 0.866eJ73", Z = 0.935eJ66 and z = 0.935ei74". The magnitude
responses of these four poles are shown in Fig. 1O.8(b). All the poles produce a
peak in the magnitude response, but the location of the peak is related to the
angle of the pole, while the height is related to the radius of the pole and the
interaction with the complex-conjugate pole.

For both FIR and IIR filters, as the order of the filter increases, the
placement of the singularities is complicated by their interactions. In practice,
an optimization procedure is usually performed to find the combination of
singularity positions that produces the best approximation to the desired

404
(a)

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

1m

/DESIRED POLE

4'1.
~
'1.'1.'1. 3

'1.

O L - - - - - - - - -..... Re
1

FIGURE IN

O~----------------~ro
o
n

Effect of quantization on pole

placement. a) Desired pole
location must be shifted to
one of four closest neighbors
when filter is implemented. b)
Four magnitude response
curves corresponding to the
four allowable pole locations.

specification. Such an optimization procedure is beyond the scope of this book,

but can be found in the references.

10.5 PERFORMING ARITHMETIC IN

FINITEPRECISION NUMBER SYSTEMS
Having considered the effects of quantization of the discrete-time
sequence and coefficient values, we now consider the effects of performing the
arithmetic operations of addition and multiplication needed to implement a
digital filter. We describe both fixed-point and floating-point representations
and discuss the advantages and disadvantages of performing these operations
next.
F1XEDPOINT ARJTIIME11C. When expressing the input sequence in fixed
point notation, each value in the sequence must be scaled so that the result

FlNITE'PREClSION EFFEcrs

405

does not fall outside, or overflow, the number range when additions and
multiplications are performed. We consider each operation separately.
Addition in fixed-point arithmetic. The addition of two fixed-point numbers
expressed in 2's-complement notation is simple: The two numbers are simply
added bit by bit starting from the right, with the carry bit being added to the
next bit location. The final carry bit is discarded. A subtraction is performed
by first negating the number to be subtracted and then adding.
Example 10.9. Addition of two fixed-point numbers. Using the 4-bit number
representatioR xxx.x, perform the following additions:
1.5\0 + 1\0:

OOLl
001.0
010.1

(= 2.5\O)

2.5\0 + (-21O):

OlD.l
110.0
1 000.1

(= 0.5\O)

Note that the final carry bit is ignored.

-2.5\0 + 2\0:

tOLl
010.0
111.1

(=-0.5\0)

011.0
011.1
110.1

(= -1.51O)

3+3.5:

Overflow error occurs in the last result because 6.510 cannot be represented
in this 4-bit number system.

When two numbers of b bits each are added and the sum cannot be
represented by b bits, an overflow is said to occur. To describe a procedure
that prevents it from occurring, let us consider an FIR filter structure that
contains a total of M additions, as shown in Fig. 10.9. Let the number system
have its binary point immediately to the right of the sign bit, so that only
fractions are represented. The values of the input sequence {x(n)} can then be
normalized to have magnitudes less than or equal to one. The multiplier
coefficient values can also be scaled to be less than 1. Then the sum of these M
terms is always less than M. To prevent overflow when the sum is computed,
the input sequence must be scaled so that all the values have magnitudes that
are less than 1/M. This scaling then ensures that the maximum possible total is
less than or equal to 1.

406

x(n)

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

--..._------l

y(n)

FIGURE 10-9
Filter structure having M inputs to adder.

The above procedure to prevent overflow indicates the limitations of

implementing a high-order filter, or one having a large M. If the filter is
implemented in one of the direct forms, a large M necessitates having a large
number of bits to ensure that the quantization error in the data sequence is
sufficiently small after the sequence has been scaled to have a maximum of
1/M. The filter can also be implemented as a cascade of second-order sections
in which a maximum of three additions is performed at anyone time. This
reduces the severity of the scaling operation.
Multiplication in fixed-point arithmetic. As in any number system, multiplication is a three step procedure. First the magnitude and sign components are
separated, then the magnitudes are multiplied and, finally, the sign of the
product is determined and applied to the result. These steps are especially
apparent in 2's-complement notation, since the magnitudes of positive and
negative numbers have different bit patterns. In finite-precision number
systems, the number of bits and position of the binary point must remain
constant. When a b-bit number multiplies another b-bit number, the product
may contain 2b-bits. For example, 112 x 112 = 1001 2, If the b-bits are organized into b = b i + b" where b i represents the integer part and b, represents
the fraction, then the product may contain 2bi + 2b, bits. The result is
represented as a b-bit number by retaining the b, bits to the right of the binary
point, through truncation or rounding, and the b, bits to the left.
Example 10.10. Multiplication of two fixedpoint numbers. Using the 4-bit
number representation xxx.x, perform the multiplication given below. In this

FINITE PRECISION EFFECTS

407

case, b =4, b, = 1 and bi =3.

2.5 10 x 1.510 :

010.1

x 001.1

0101
x 0011
0101
0101
0000
0000
0001111 = 00011.11 when the binary point is restored
Restoring the 4-bit representation, the product becomes
001l.1 2 ( = 3.5 10) by truncation, and
0100.02 ( = 4.010) by rounding

The truncation that occurs when the 2b-bit product is represented as a

b-bit result is one source of error that will be analyzed below.
Another alternative is to implement the filter in floating-point representation. We consider the problems associated with floating-point representations next.
FLOATING-POINT ARITHMETIC. Floating-point representation of numbers is
a partial solution to the overflow problem, because a larger range of numbers
can be represented. This representation reduces the need for scaling the input
and the worry about overflow errors. Let us consider performing additions and
multiplications with the floating-point representation to determine the
disadvantages.
Addition of two floating-point numbers. The addition of two floating-point
numbers is more complicated than that of two fixed-point numbers. To
perform an addition, the exponents of the two numbers must be set equal. To
avoid exceeding the range of the mantissa, the exponent of the smaller number
must be increased to equal that of the larger number. In doing this, the
mantissa of the smaller number must fall outside the range ! s 1m I < 1. The
two mantissas are then added using the fixed-point procedure. The sum is
re-normalized to be in the magnitude range [!, 1) by adjusting the exponent.
Example 10.11. Addition of two floating-point numbers. Let us consider an
8-bit number representation with 5 bits to represent the mantissa and 3 bits to
represent the exponent. We want to perform the addition 5.510 + (-1.25)10'
Expressing 5.510 in floating-point notation, we get 0.6875 x 23 or 01011 011 2 ,
Expressing -1.2510 in floating-point notation, we get -0.625 x 2'. To set
the exponents equal, we must express -1.25 10 as -0.15625 x 2' or 111011 011 2 ,

408

INTRODUcrION TO DIGITAL SIGNAL PROCESSING

Perfurming the addition of the 6-bit expanded mantissas, we get

010110
111011
1 010001 (= 0.53125 10)
The result obtained by truncation is 01000 011 2 = 410 , and the result
obtained by rounding is 01001011 2 = 4.5 10 ,

The floating-point representation is costly in terms of the time required to

compute a sum, since first the exponents must be made equal before the sum is
computed. This increase in computation time slows the filter operation and
may prevent it from operating in real time.
Multiplication of two floating-point numbers. In performing a mUltiplication of

two floating-point numbers, the two mantissas are multiplied using the
fixed-point multiplication procedure and the two exponents are added. The
exponent value is then adjusted, so that the magnitude of the product of the
mantissas falls within [t 1).
Example 10.12. Multiplication of two ftoating-point numbers. Let us consider
an 8-bit number representation with 5 bits to represent the mantissa and 3 bits to
represent the exponent. We want to perform the multiplication

2.73510 x (-1.25)10
Expressing 2.73510 in floating-point notation, we get 0.68375 x 22 or
01011 0102,
Since the second multiplier is negative, the product will be negative.
Expressing the magnitude 1.25 10 in floating-point notation, we get 0.625 x 21 or
01010001 2,
Performing the mUltiplication of the mantissas, we get
01011
01010
00000
01011
00000
01011
00000
001101110
which is equal to O. 0110111~( = 0.4296875 10 ) when the binary point is restored.
By truncation, we get 0.01102 (=0.375 10), and by rounding we get 0.01112(=
0.4375 10)'
Performing the addition of the exponents, we get 0102 + 001 2 = 011z( = 310),
After converting the mantissa into a negative number, the final result is
11010 011 2( = -310) by truncation or
11001 011 2 ( = -3.5 10) by rounding

FINITE-PRECISION EFFECTS

409

As the examples above indicate, truncations must be performed after

both addition and multiplication operations with floating-point arithmetic.
Moreover, these truncations have a variable magnitude, depending on the
value of the exponent. This variability makes the analysis of these effect~ very
difficult. To give an indication of the effects that can occur with truncation, we
will consider only fixed-point filter implementations.

10.6 EFFECTS OF FINITE PRECISION

ARITHMETIC ON blGITAL FILTERS
In this section, we describe the effects of finite-precision arithmetic operations
on the performance of FIR and IIR filters. To simplify the analysis, only
fixed-point notation is considered. The input sequence and coefficient values
can be scaled so that overflow cannot occur in the addition operation. Hence,
the only truncation error associated with computing the product in a
finite-precision number system need be considered. This truncation can be
modeled as having an additive error source at the output of each multiplier, as
shown in Fig. 10.10. We first consider the simpler FIR filter.
FINITE-PRECISION ARITHMETIC EFFECTS IN FIR FILTERS. Let us consider
the FIR filter implemented in the direct form nonrecursive structure shown in
Fig. 10.11. The truncation errors appear as an additive noise source at the
output of each multiplier. If rounding is used, the noise value at each
multiplier can be assumed to be uniformly distributed between E/2, where
E = Tbf is the precision of the number system of the computer. The variance,
or power, of this noise component is then
(10.15)
t".,(n)
x(n)

FIGURE 10-10
Multiplication truncation effect is modeled as an additive noise source at the output of each
multiplier for the general Direct Form II filter structure.

410

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

x(n) -----..-----f

>-----:~

yen)

FIGURE 10-11
Multiplication truncation effect is modeled as an additive noise source at the output of each
multiplier for the Direct Form II FIR filter structure.

If the noise values for the different multipliers can be considered to be

independent of each other, then it can be shown that the noise powers add.
For N multipliers, the output noise power is equal to

a~ut = N 12 = N2-

2b

f /12

(10.16)

Since the noise power increases with N, and N is also equal to the duration of
the FIR filter unit-sample response {h(n)}, this is another reason why it is
desirable to make the duration of {h(n)} as short as possible.
FINITE-PRECISION ARITHMETIC EFFECfS IN UR FILTERS. The analysis of
truncation effects for IIR filters is more complicated than for FIR filters
because of the feedback structure. Because of this feedback, the errors
associated with finite-precision multiplications and additions circulate back into
the filter and affect future values of the output. As described below, these
errors may produce unwanted oscillations, called limit cycles.
In general, the analysis of these limit cycles is very difficult. To
demonstrate their existence, we consider the simple special case of the limit
cycles that occurs in response to a unit-sample sequence. Since the input to the
filter is zero for values of n other than n = 0, these limit cycles are commonly
called zero-input limit cycles. These cycles exist when the unit-sample response
is prevented from decaying to zero because of multiplication truncation.
Example 10.13. Limit cycle in first-order recursive filter. Let us assume the
digital filter is implemented with the number system having precision equal to O.l.
Consider the unit-sample response of the first-order recursive filter described by
y(n) = -0.9 y(n -1) +x(n)

The unit-sample response of this hardware implementation is shown in Fig. 10.12.

FINITEPRECISION EFFEl TS

411

0.8
0.6

0.5

0.5

0.5

-\ 0

-0.5

-0.5 -0.5

-0.7
-0.9

LIMIT CYCLE

FlGURE 10-12
Example of a zeroinput limit cycle in the
unit-sample response of
a first -order recursive
filter implemented with
finite-precision
arithmetic.

The unit-sample response begins to follow an exponential decay, as expected,

until the magnitude of 0.5 is reached. After that point in time, the product
0.9 X 0.5 = 0.45 is rounded back to 0.5 at the output of the multiplier. The minus
sign in the feedback loop causes the output to alternate between +0.5 and -0.5.
This is one example of a zero-input limit cycle.

These limit cycles also appear as unwanted noise at the output. But they
are different from those previously considered in that, rather than being
random (producing a "hiss"), they are correlated (producing a "hum" or
"buzz"). Even for such simple cases, the complete analysis of these limit
cycles, which determines under what conditions they exist and what their
amplitude and frequency characteristics are, is beyond the scope of this book,
but is included in the references.
ParaUel versus cascade implementation of filters. Parallel combinations are
easier to analyze because they can be considered separately. For example, for
parallel implementations, all sections can exhibit zero-input limit cycles
independently. Cascade implementations are more difficult to analyze because
the output of the first section is then the input to the second. Hence only the
first section in a cascade connection can exhibit a zero-input limit cycle. The
second section then has this limit cycle as an input and, by definition, cannot
have a zero-input limit cycle. Since the nature of the limit cycle depends on the
first or second-order filter characteristics, and only the first section can exhibit
a zero-input limit cycle, the sequence observed at the output of the cascade
combination depends on the ordering of the sections in the filter implementation. This effect is considered in the projects.
Since the analysis of these quantization effects is very difficult, we now
tum to programming techniques to simulate these effects arid provide insights
into their behavior.

10.7 PROGRAMS TO SIMULATE

QUANfIZATION EFFECTS
We now describe two subroutines that can be added to the filter subroutines to
study the behavior of digital filters under finite-precision conditions. The first

412

INTRODucnON TO DIGITAL SIGNAL PROCESSING

subroutine performs the quantization of an infinite-precision number to a

finite-precision value. The second subroutine scales a discrete-time sequence to
a particular range to match an ADC quantizer characteri!.tic and also to
prevent overflow errors in the computations.
To generate more interesting input sequences than the unit-sample
sequence, we also describe the operation of a random number generator that is
usually available on most computer systems. The random input sequence is
used along with the unit-sample sequence and the sinusoidal sequence to
observe the operation of the filters in the projects.
QUANTIZER SUBROUTINE. To simulate finite-precision systems, it is useful
to have a subroutine that transforms an infinite-precision value into the closest
finite-precision value represented in the number system. This subroutine can
be applied to values of the discrete-time sequence to simulate a quantizer. It
can also be applied to the results of multiplication and addition to simulate
rounding and overflow effects.
Let X be aninfinite-precisi<;>n number that is represented as a floatingpoint number on the computer. The value of X is to be converted into a
second floating-point number XQ, that is equal to the finite-precision value,
closest to X, that can be represented by a 2's-complement fixed-point number
having NBIT bits. Of these NBIT bits, NI are used to represent the integer
part, including the sign, and NF ( = NBIT - NI) bits are used to represent the
fractional part. To perform this operation we want to write the subroutine

QUANT(X,NI,NF,XQ)
where X (real input), NI and NF (integer inputs) and XQ (real output) are
defined above. The precision is equal to 2- NF, and the range of integer values
is then from _2N1 - 1 to 2NI - 1 - 2-NF
The simplest and fastest way of implementing this conversion is to use the
truncation function IFIX in FORTRAN. If Y is a floating-point number, then
N = IFIX(Y) is the integer part. For example, if Y = 2.8, then N =
IFIX(2.8) = 2. If Y> 0, by adding 0.5 to Y, we get a rounding to the nearest
integer, or N = IFIX(Y + 0.5). For example, N = IFIX(2.8 + 0.5) = 3. If Y < 0,
we must subtract 0.5. For example, if Y = -3.9, then N = IFIX( -3.9 - 0.5) =
-4. Using this IFIX function, the quantizer can be implemented by first
multiplying the value of X by a constant C, performing IFIX(CX + 0.5), and
then dividing the result by C. This sequence of steps produces the desired
value for XQ when C = 2NF. The value of NI is used to check whether the
number X is within the allowable range. If X exceeds the range, XQ should be
set to the maximum or minimum value in the number system, depending on
whether X exceeds the positive or negative limit. The following example
illustrates the conversion process.
Example 10.14. mustratioD of the QUANT cODvenioD process. Consider an
8-bit digital hardware system using 2's-complement notation with NI = 5 and
NF = 3. Quantize the values of X given below.

FINITEPRECISION EFFECI'S

413

Since NF = 3, the precision is 2- 3 = 0.125. Then C = 23 = 8. Hence, the

quantized value can be expressed as
XQ

= IFIX(8X + 0.5)

for X~O,

and
XQ = IFIX(8X - 0.5)

for X < 0

8
Since NI = 5, the range is from -16 (1000 000) to 15.875 (01111111).
For X = 3.1415926,
For X= -2.87,
For X = 37,

XQ = [IFIX(25.1327408 + 0.5)]/8 = 25/8 = 3.125.

XQ = [IFIX(-22.96-0.5)]/8 = -23/8= -2.875.
XQ = [1FIX(296 + 0.5)]/8 = 296/8 = 37
= 15.875 (maximum limit).

For X=-37,

XQ = (IFIX( -296 - 0.5)]/8 = -296/8 = -37

= -16 (minimum limit).

SCALE SUBROUl1NE. Before QUANT can be applied to the elements of an

arbitrary discrete-time sequence, the sequence values must be scaled, or

multiplied by a factor F, so that the range of the signal values matches the
range of the finite-precision number system. As discussed previously, one
scaling factor is appropriate for deterministic signals (sinusoids), while a
different factor is appropriate for random signals (speech). For deterministic
signals, the range of the signal is known. Let the values of the signal be in the
range [-R, R]. If the maximum magnitude of the quantizer is equal to M,
which occurs at the negative limit, the signal is scaled by F = M / R. For
random signals, the rms value X rm.. rather than the maximum value, is the
important parameter. One common scale factor is then F = M/(4Xrms).
The subroutine to perform this function is given by

SCALE(X,NX,F)
where
X
(real input/output array) is the discrete-time sequence to be scaled
NX
(integer input) is the size of the array
F
(real input) is the scale factor
The operation that is performed is given by
X(N) = F*X(N)

for 1 sNsNX

We next describe the sequences that can be used as the inputs to filters
implemented with finite-precision number systems.
INPUf SEQUENCES. Previously, t~e unit-sample sequence and sinusoidal
sequences have been employed as inputs to the various filters. These can also
be used for finite-precision filters. The unit-sample sequence shows the
existence of the zero-input liinit cycle for recursive filters. The sinusoidal
sequence having the frequency corresponding to the location of the filter

414

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

magnitude response maximum value can be used to ensure that the range of
the number system is not exceeded. We also consider a third type of input
sequence, the random sequence, that can be used to observe the filter output
behavior of signals that are more representative of actual signals. We first
consider the white random sequence, which on average has a spectrum that is
constant with frequency. We then filter this white sequence with a lowpass
filter to generate a lowpass correlated random sequence, and with a highpass
filter to generate a highpass correlated random sequence. A random number
generator is described next that can be used to generate the random
discrete-time sequence.
In FORTRAN, a random number can be generated by a call to the
pseudorandom number generator. On the VAX/VMS system, this is done with
the statement

A=RAN(NRANl,NRAN2).
where NRAN1 and NRAN2 are seeds used by the generator. The values for
these seeds must be specified to equal two arbitrary integers at the beginning
of the program, before the random number generator is called. Each time this
statement is executed, the variable A takes on a new value of a uniformlydistributed random number between 0 and 1. The average value of these
The nth element of
random numbers is then 0.5 and the variance is equal to
the N-point white random sequence {wen)} can be set to a zero-mean random
number with the same variance with the statement

n.

wen) = RAN(NRAN1,NRAN2) - 0.5,

for 0:5 n:5 M - 1

(10.17)

The meaning of white will be discussed in terms of the spectrum of {w(n)}

below. The sequence containing 64 random values is shown in Fig. 10. 13(a).
The random number generators in most languages and on most computers are
similar to that described above.
The power spectrum sequence, denoted by {P(k)}, of any N-point
sequence {x(n)} is obtained by computing the N-point DFf sequence {X(k)}
and then summing the squares of the real and imaginary parts, or
P(k)

= ~ [X~(k) + Xf(k)]

for 0:5 k :5 N - 1

(10.18)

If this is performed M times and the ith result is denoted by {P'(k)}; the
average value of the sequence, {Pavek)}, is given by

1 M
Pavek) = M ~ P;(k)

for 0 :5 k :5 N - 1

(10.19)

The random proce'ss {x(n)} is said to be white, when

lim Pavek)
M~oc

=a

for all k

(10.20)

where a is a constant. That is, the white sequence has equal components at
each frequency, on average. The power spectral sequence {P(k)} of a single

FINITE-PREOSION EFFEcrs

(0)

(w(n)

(b)

(u(n)

415

It

(c) {v(n)}

()

It

w
FIGURE 10-13
Examples of random sequences. The discrete-time sequences represent a typical realization of the
random process. The frequency curves represent the power spectrum of the realization and the
average power spectrum. a) White random process. b) Lowpass random process. c) Highpass
random process.

N-point realization of {w(n)}, however, usually appears quite random itself,

as shown in Fig. 10. 13(a).
Correlated random sequences are obtained by filtering the white process
{w(n)}. One example of a lowpass random process {u(n)} can be defined as
u(n)=![w(n)+w(n+1)]

for0:5n:5N-1

(10.21)

An example of this process is shown in Fig. 1O. 13(b). To determine an N-point

sequence for {u(n)}, N + 1 points of the white sequence {w(n)} must be

416

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

generated. Performing the same type of averaging procedure produces a power

spectral sequence that has the property that
lim Paik) = (1 + cos(21rk/N))/2

for all k

(10.22)

M-+oo

This power spectral sequence has most of its power concentrated at low
frequencies. One example of a highpass random process {v(n)} can be defined
as
v(n)

= Uw(n) -

w(n

+ 1)]

for OSnsN-1

(10.23)

An example of this process is shown in Fig. 1O.13(c). Performing the same

type of averaging procedure produces a power spectral sequence that has the

property that
lim Pav(k) = (1- cos(21rk/N))/2

for all k

(10.24)

M--.oo

This power spectral sequence ,has most of its power concentrated at high
frequencies.
These white and correlated sequences represent interesting classes of
signals that approximate signals that occur in practice. To observe the
finite-precision effects, these sequences are first applied to a digital filter that is
implemented in an infinite-precision number system, approximated by the
REAL variable in a 16-bit or 32-bit computer system. The output sequence
produced in this case can be compared with that produced by the same filter
structure but implemented with a finite-precision hardware. The effects of
finite-precision can be observed by taking the difference between these two
output sequences. The projects at the end of this chapter illustrate this
procedure.

10.8 SUMMARY
Quantization effects are caused by the errors that are due to representing
numbers with a finite precision. Two finite-precision number systems have
been described, the fixed-point and floating-point notation. Three effects due
to finite-precision were considered. The first occurs when the continuousvalued signal amplitude is represented with a finite number of bits in the
analog-to-digital conversion operation. This effect was shown to be similar to
adding noise to the signal. The second effect occurs when the filter coefficient
values are expressed in finite precision. This effect was shown to limit the
locations of the singularities in the z plane to lie on a grid. The corresponding
deviations of the magnitude response were illustrated for poles and zeros. The
third effect is the truncation that results when the sum or product of two
finite-precision numbers is itself represented as a finite-precision number. For
FIR filters implemented in the direct-form nonrecursive structure, this effect
was shown to introduce additive noise to the output sequence. For IIR filters,

FINITEPRECISION EFFECTS

417

however, these errors circulate in the memory of the filter and can produce
undesirable oscillations, called limit cycles. The chapter ends by describing
programming techniques to simulate these finite-precision effects, which are
very difficult to determine analytically.
These finite-precision effects are most important in three practical cases:
1. when specifying the number of bits in an ADC;
2. when implementing a filter in special purpose hardware to minimize the
number of bits in the finite-precision number system;
3. when implementing filters whose poles lie near the unit circle.

REFERENCES TO TOPICS
Zero-input Omit cycles

Blackman, R. B.: Linear Data-smoothing and Prediction in Theory and Practice, Addison-Wesley,
Reading, MA, 1965.
Jackson, L. B.: "Roundoff-noise analysis for fixed-point..digital filters realized in cascade or
parallel form," IEEE Trans. Audio Electroacoust., AU-IS: 107-122, (1970).
Jackson, L. B.: "Roundoff noise analysis in cascade realizations of finite impulse response digital
filters," Bell Sys. Tech. J., 52: 329-345 (1973).
FDter optimization proceclares

Avenhaus, E.: "On the design of digital filters with coefficients of limited word length," IEEE
Trans. Audio Electroacoust., AU-20: 206-212 (1972).
Oppenheim, Alan V., and Ronald W. Schafer: Digital Signal Processing, Prentice-Hall,
Englewood Cliffs, NJ, 1975.
Rabiner, Lawrence R., and Bernard Gold: Theory and Application of Digital Signal Processing,
Prentice-Hall, Englewood Cliffs, NJ, 1975.
Taylor, F. J.: Digital Filter Design Handbook, Marcel Dekker, New York. 1983. Advanced topics
also covering ladder and wave filters. Includes hardware and software considerations.
~

conversion

Jayant. N. S. and Peter Noll: Digital Coding of Waveforms, Prentice-Hall, Englewood etiffs. NJ.
1984.

Mano, M. Morris: Digital Logic and Computer Design, Prentice-Hall, Englewood Cliffs, NJ. 1979.

PROBLEMS
10.1. Range IUld preCISIon using 2's-c:omplement fixed-point notation. Consider a
computer with an 8-bit word. Determine the range of values and the increment
between numbers that can be expressed when the binary point is placed at the
following locations:
(a) after the right-most bit
(b) after the sign bit
(c) in the middle of the word
10.2. Repeat Problem 10.1 for a 16-bit word and a 32-bit word.

418

1N'l1l0DucnON TO DIGITAL SIGNAL PROCESSING

18.3. Rup aacI predsioa using ftoating-point notation. Determine the range of
values and the increment between numbers that can be expressed for a b-bit
word containing bID bits for the mantissa and be bits for the exponent for the
following values of b, bID and be:
(a) b = 8, b m = 6 and be = 2
(b) b=8, b",=2 andbe =6
(c) b= 16, b", = 12 and be=4
(d) b =32, bID =24 and be =8
1&.4. Convert the following decimal numbers to binary representation using 16-bit
notation in which the fraction is represented by 8 bits. Indicate the results
obtained both by rounding and truncation:
(a) 117
(b) 0.0726
(c) 43.34
(d) -18.68
10.5. Perform the following additions in binary representation using an 8-bit notation
in which the fraction is represented by 4 bits. Indicate the results obtained both
by rounding and truncation.
(a) 2.510 + 3.2510
.
(b) 2.510 + -3.2510
10.6. Perform the following multiplications in binary representation using an 8-bit
notation in which the fraction is represented by 4 bits. Indicate the results
obtained both by rounding and truncation.
(a) 2.5 10 x 3.2510
(b) 2.5 10 x -3.25 10
10.7. Perform the following calculations with binary numbers using both 2'scompleml!nt fixed-point with 8-bit words in which three bits are assigned to the
fraction. Perform the truncations by rounding.
(a) 1.5 - 3
(b) 2xn.
(c) 0.33 x 0.8
(d) -0.9 x 5.6
10.8. Perform the calculations given in Problem 10.7 by using floating-point notation
with 8-bit words in which three bits are assigned to the exponent (including the
sign bit of the exponent).
10.9. Express n as an 8-bit word using 2's complement notation, when the range of
the number system is [-8, 7].
10.10. Derive the SNR value similar to Eq. (10.8) for a sinusoidal signal whose
amplitude is equal to the maximum level of the quantizer.
10.11. Consider the all-pass system function given by

H(z) = 1- p-1 Z -l
1- pz

forO<p<1

If the filter coefficients are quantized to 8-bits in 2's-complement notation, does

the resulting filter still represent an all-pass system?
10.12. Consider the digital filter defined by the difference equation

y(n) = 0.9y(n -1) +x(n)

FINITEPRECISlON EFFECIS

419

(a) Determine the range to which the input sequence must be scaled such that
overflow is avoided for any sinusoidal input.
(b) Determine the range to which the input sequence must be scaled such that
overflow is avoided for any bounded input, that is, Ix (n)1 < M, for all n.

COMPUTER PROJECfS
10.1. Spectral whiteness of random nmnber generator
Object. Test that the random number generator on your computer produces a
white process.
Compater results. Generate a 128-point sequence of the white random sequence
{w(n)}.
To show the random shape of one realization of the power spectral
sequence, compute and plot 65 points of the power spectral sequence corresponding to O:s co :S re.
To show the limiting behavior of the averaging process, generate 10
independent {w(n)} sequences, compute the power spectral sequence of each,
compute and plot the average. To save on memory, compute a running average
by accumulating the individual spectral sequences and then 'dividing by the
number. Repeat with 100 independent sequences.
Comment on the "whiteness" of the spectral averages.
10.1. Spectral content of correlated random processes
Object. Test that the correlated random processes have lowpass and highpass
spectral shapes.
Computer results
(1) Generate a 128-point sequence of the lowpass rand(.m sequence {u(n)}
defined by Eq. (10.21). To show the random shape of one realization of the
po~ spectral sequence, compute and plot 65 points of the power spectral
sequence corresponding to O:s (J) :S re. To show the limiting behavior of the
averaging process, generate 10 independent {u(n)} sequences, compute the
power spectral sequence of each, compute and plot the average. To save on
memory, compute a running average by accumulating the individual spectral
sequences and then dividing by the number. Repeat with 100 independent
sequences. Comment on the agreement of the computer results with Eq.
(10.22) as the number of sequences increases.
(2) Repeat Part (1) for the highpass random process {v(n)}.
10.3. ADC conversion errors
Object. Compute the quantization errors produced by an analog-to-digital
converter and to compare the. noise power to the predicted value.
Consider an 8-bit quantizer characteristic that has the range [-8,7].
Implement the subroutines QUANT, to perform the quantization, and SCALE,
to match the input sequence to the quantizer. Two 128-point input sequences are
to be generated: {sin(ren/32)} and {w(n)} (the white random sequence).

420

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

Aualytic resds
(1) Determine the scale factors to be applied to the two input sequences to match
each to the quantizer.
(2) Compute the variance of the additive noise that is predicted if the
quantization errors are uniformly distributed and independent.

Coatpater resaIts. For each sequence:

(a) Compute and plot the original infinite-precision sequence. Use PLOT.
(b) Scale the sequence to match the quantizer. Denote this input sequence by
{X (n)}. Compute and plot the output sequence of the quantizer {xq(n)}
when {x(n)} is applied.
(c) Compute and plot the quantization error sequence e(n) by
e(n) =xq(n) - x(n)

for

O~n ~

127

(d) Estimate the power of the error sequence, denoted by p, given by

1 127
p=-~ e2(n)
128 n-O

and compare with the predicted value.

10.4. Deviation of mapitacle response due to coef6clent quantization
Object. Observe qualitatively the degradation of the magnitude response due to
finite-precision effects on the filter coefficients.
For this project use an IIR and/or FIR filter that you have designed in a
previous project and for which you have the coefficient values.
Coatpnter resaIts
(1) For comparison, plot the magnitude response of the filter with the coefficients
expressed in infinite-precision (approximated by the FLOATING POINT
variable in the computer).
(2) Apply the coefficients to the QUANT subroutine that simulales a 16-bit word
with the integer part being represented by 4 bits. Compute and plot the
magnitude response. Repeat while reducing the word size by one bit for each
iteration, until the magnitude response is grossly distorted or, for IIR filters,
an instability is encountered (overflow error occurs).
(3) Determine the minimum number of I bits needed to satisfy the magnitude
specification.
10.5. Zero-inpat 6mit cycle for an DR filter
Object. Observe the limit cycle behavior of an IIR filter.
For this project use one of the IIR filters that you have designed in a
previous project. Determine the coefficients for both the direct form and cascade
of second-order sections. Multiplication truncation is accomplished by using the
QUANT subroutine after any multiplication of two terms is pedormed. This can
be done most easily by implementing a subroutine SUMPQ, which is identical to
SUMP in Olapter 2 except fol including the QUANT subroutine.
AuIytlc raaICs. Determine the proper range of the quantizer to prevent
overflow errors from occurring when a unit-sample sequence is applied to the
filter. Generating the true unit-sample response may be helpful.

FlNlTEPRECISION EfFECTS

421

Computer results
(1) Limit-cycle of direct form implementation.

(a) Starting with 8-bits in the quantizer, matched to the range determined
above, plot the unit-samplr response that is produced when all the
products are truncated. Demonstrate the existence of a zero-input limit
cycle.
(b) Reduce the quantizer word size by one bit, while maintaining the same
range, and repeat Part (a).
(c) Repeat part (b) until the quantizer has four bits.
(2) Limit-cycle of cascade of second-order section implementation. Repeat Part (1)
under the foUowing conditions:

by using only the first section of the esos structure.

by using the entire esos structure.
by using a di1Ierent section of the esos structure.
by using the entire CSOS structure with this second section in the first
positi9n of the cascade structure.
Since only the first section can have a zero-input limit cycle, the two total filter
limit cycles observed in Parts (b) and (d) should be different.
(a)
(b)
(c)
(d)

CHAPTER

11
INVERSE
FILTERING

11.1 INTRODUcnON
In this chapter, we discuss the interesting and important topic in signal
processing called inverse filtering. This topic integrates many of the digital
signal processing concepts presented in the previous chapters to solve the
problem of removing distortions from observed signals that are introduced by
measuring instruments or by physical transmission media. This inverse filter is
also commonly called a deconvolution filter in the fields of geophysics and
bioengineering. Inverse filters work best when applied to a system that can be
modeled by a system function Hm(z) that has the minimum-phase property.
This property'is defined in terms of the pole/zero pattern of Hm(z), the phase
response Arg[Hm(eiOJ )] and the unit-sample response {hm(n)}. Two procedures
for designing the inverse filter are presented. The first performs the design
directly in the z plane by placing the poles of the inverse filter at the locations
of the model zeros and the filter zeros at the location of the model poles. This
procedure is limited to models for which either the system function or the
pole/zero pattern is known. In the second design procedure, an FIR filter
approximation to the ideal inverse filter is obtained. Although not exact, these
approximations have useful applications in practice. Modifications to this
design procedure are discussed to include additive noise in the observed signal.

422

INVERSE FlLTERlNG

423

A common alternate procedure for designing inverse filters in the time

domain requires concepts that are not difficult, but different from the topics
covered in this text. These time-domain design procedures are not discussed
here, but can be found in the references.

11.2 APPLICATIONS OF INVERSE

FILTERS
In many signal processing applications, it is common to encounter three
sources of distortion in the observed signals. The first involves the limitations
of the signal detector. If the detector magnitude response is not Oat and the
phase response is not linear with frequency, then the detector acts like a filter
that distorts the signal being sensed. Most transducers, such as microphones
, and antennas, exhibit this effect. The second source of distortion is due to the
physical medium through which the signal propagates. For example, acoustic
signals can pass through air (speech signals), water (SONAR signals),
biological tissues (diagnostic ultrasound signals) or through the earth (geoseismic signals). Each medium has its own effect on the acoustic signals. The effect
of the medium on the signal can be modeled as a filtering operation. Inverse
filters serve to compensate for the filtering actions produced by the detector
and/or the medium. The third source of distortion is additive noise. Since its
effect on inverse filters is complicated, its description will be delayed to the end
of the chapter. We now illustrate two representative cases of the first two
sources of distortion.
RADAR PROBLEM. Consider the radar problem shown in Fig. 11.1. The
range to an airplane is determined by transmitting an electromagnetic pulse
and detecting the arrival time of the echo. If to is the elapsed time from pulse
transmission, the range is equal to cto/2, where c is the pulse propagation
speed.
For individual airplanes, this procedure works quite well. The radar

ECHO

~-----Jo Time

:t-IE_TE_CrE_D_SI_G_NA_L_ _ _ _

10

FIGURE 111
Model of a radar system.

424

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

SINGLE ECHO

DESIRED SIGNAL

---1~_~_~L_VT_E:_i_'E~~ 6~--~l------~~
ECHOES FROM
MULTIPLE TARGETS

OUTPUT INDICATING MULTIPLE

TARGET RANGES

~ -1. ._~_~_~TE_ER_i_E--,~

....lt2----~

--'t!.. I!-l

01-1

FIGURE 11-2
Desired operation of an inverse filter for the radar problem.

system is limited, however, by the finite bandwidth of the antenna system,

which prevents the transmitted pulse from being impulsive. As a consequence,
when two airplanes are in close proximity, the two echoes may not appear as
separate waveforms, but merge into one as shown in Fig. 11.2. In this case, an
inverse filter is applied to the detected signals in an attempt to perform pulse
compression, or to compress the pulse waveform produced by the system into
one having a shorter duration, to improve the resolution. Since the more
complicated signal from multiple aircraft can be considered to be the
superposition of many pulse waveforms, the inverse filter is applied to the
observed waveform to transform it into multiple shorter-duration waveforms,
thus resolving the individual airplanes. This example can easily be extended to
typify many problems in bioengineering and geophysics in which an overlap in
identical pulse waveforms occurs.
REVERBERANT TRANSMISSION CHANNEL. To illustrate the distortion

produced by a physical medium, let us consider the data transmission channel

between two modems shown in Fig. 11.3. When the receiving modem is
terminated with an impedance value that does not match the characteristic
impedance of the channel, a part of the received signal is reflected back into
the channel, causing an echo to propagate back from the receiver to the
transmitter. If an improper termination exists at the transmitter end of the
transmission line, a reflection occurs there as well. If an impulsive signal is
transmitted, the signal detected at the receiving end includes the transmitted
signal and the set of echoes, as shown in Fig. 11.3. Since only a partial
reflection occurs at each end, the amplitudes of the echoes decrease with time.
When a nonimpulsive signal is transmitted, the received signal is simply the
convolution of the transmitted signal and the impulsive received signal. In
conventional communication systems, the entire received signal produced by
one transmission must decay before a second signal is transmitted. This limits
the data transmission rate across the channel.
To compensate for these echoes, we can model the channel as a filter

INVERSE FILTERING

TRANSMImNG
MODEM
,

TRANSMISSION
CHANNEL

425

RECEIVING
MODEM

TRANSMITTED SIGNAL

1-1

RECEIVED SIGNAL

III'III!

II"

~OES

1111"1~
I

I Ii

FIGURE 113
Model of a reverberant data transmission channel that exhibits echoes.

whose unit-sample response is identical to the detected signal when a single

impulse is transmitted. In this case, the channel can be modeled as a first-order
recursive filter, shown in Fig. 11.4. An inverse filter, called an echo suppressor,
is applied in an attempt to compensate for the behavior of the channel. The
input to the inverse filter is the entire set of detected echoes and the desired
output is the single pulse that was transmitted, as shown in Fig. 11.5.
We will find below that an exact inverse filter can be implemented only
for systems that have the minimum-phase property. In the next section, we
define this property.
TRANSMTITED SIGNAL

1o

CHANNEL MODEL

RECEIVED SIGNAL
1

Iii

I ,

FIGURE 11-4
Digital filter model for the reverberant channel.
DESIRED OUTPUT

RECEIVED SIGNAL

INVERSE FILTER
(ECHO SUPPRESSOR)

- ...- - - + . (

FIGURE 115
Desired operation of an inverse filter for the reverberant channel.

426

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

11.3 MINIMUMPHASE SYSTEMS

In this section, we model the causal and stable detector or physical medium in
terms of the Z plane pole/zero pattern of its system function Hm(z), its phase
response Arg[Hm(eiW )] and its unit-sample response {hm(n)}. The subscript m
denotes model. We can then define the three corresponding criteria by which
we can recognize a minimum-phase model.
THE z PLANE CRITERION. In Chapter 5, the system function for a stable and
causal discrete-time filter was shown to have its poles within the unit circle and
its zeros anywhere in the z plane. We now define three types of model system
functions Hm(z), depending on the location of the zeros. The three cases are
shown in Fig. 11.6. If all the zeros of Hm(z) lie inside the unit circle, the system
model is called minimum-phase, if all the zeros lie outside the unit circle, the
model is called maximum-phase, and if zeros lie both inside and outside the
unit circle, the model is called mixed-phase. As shown below, only for
minimum-phase models is the inverse filter both stable and causal.
THE PHASE RESPONSE CRITERION. In the discussion below, we consider
Arg[H(eiw )] to be the entire phase response of the model, rather than the
principal value. If we are given the principal value, we can construct the entire
phase function by adding or subtracting 2H radians to the principal value at the
points of discontinuity, as shown in Fig. 11. 7. This procedure is called phase
unwrapping.
In Chapter 5, the phase response of a filter was determined in the Z plane
by adding the contributions from its individual singularities. As (J) went from 0
to H, a pole decreases the phase response by H radians, while a zero inside the
unit circle increases the phase response by H radians. A zero outside the unit
circle made no net contribution to the phase response as (J) went from 0 to H. If
all the singularities are inside the unit circle, which is the case for a causal and
stable minimum-phase model, the phase response Arg[Hm(eieo )] starts at zero
at (J) = 0, is a continuous-function of (J), and returns to zero at (J) = H. Zeros on

/z/=!

1m

-+<>--..-----t-Re

MINIMUM - PHASE
MIXED- PHASE
MAXIMUM - PHASE
FIGURE 11-6
Pole/zero pattern for minimum-phase, mixed-phase and maximum-phase systems.

INVERSE FILTERING

PRINCIPAL VALUE
OF PHASE RESPONSE

ENTIRE
PHASE RESPONSE

"tL \V\ 7
_:!\
11

41:7

OJ

1C------+--~7t-+UJ

FIGURE 11-'
Comparison of principal value
of the phase response, which is
produced by the DFf, and the
entire phase response.

the unit circle cause discontinuous jumps of H radians in the phase response. If
Mz zeros lie outside the unit circle, the phase response is equal to - MzH at
co = H.
One other point must be considered before the phase response criterion
of a minimum-phase model is given. In the description of all-pass filters in
Chapter 8, it was shown that when a zero of the system function, located at
z = Zo, is moved to z = l/Zo, the shape of the magnitude response of the
system is unaltered, but the phase response changes. Hence, a given
magnitude response can have many valid phase response functions.
From the behavior of the magnitude and phase responses described
above, we can form the following definition of a minimum-phase system:
For a given magnitude response, the minimum-phase system is the causal
system that has the smallest magnitude phase at every frequency co.
That is, in the set of causal and stable filters having the same magnitude
response, the minimum-phase response exhibits the smallest deviation from
zero phase. Note that the zero-phase filters described previously in the text
cannot be minimum-phase because they are noncausal and their system
functions have poles outside the unit circle. By adding sufficient delay to the
unit-sample response, these filters become causal and have a linear phase
response. The minimum-phase filter having the same magnitude response will
have a phase response that has a smaller magnitude at each frequency.
mE UNIT-SAMPLE RESPONSE CRITERION. In the time-domain, we can
find the condition on a causal unit-sample response {hm(n)}, by which we can
recognize a minimum-phase system. Recall that for a causal system, hm(n) = 0,
for n < O. Let us define the energy evolution factor E(M) to be
M

E(M)

= 2:
n=O

h;,,(n)

for Os;M < 00

(11.1 )

428

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

2 o

MINIMUM - PHASE

MAXIMUM -PHASE

MIXED- PHASE
(LINEAR - PHASE)

FIGURE 11-8
Pole/zero pattern for minimum-phase, mixed-phase (linear-phase) and maximum-phase systems in
Example 11.1.

IH(eJ"')I

12.55

O~======~~-L

____ ____--Jw
~

It

Arg [H(eJ<O)
It

MINIMUM-PHASE

-It

-21t

-31t

-41t

FIGURE 11-9
Magnitude and phase responses for
minimum-phase, mixed-phase (Iinearphase) and maximum-phase systems in
Example 11.1.

INVERSE FILTERING

429

For a particular value of M, E(M) measures the accumulated energy in the

filter unit-sample response sequence from n = 0 to n = M. For minimum-phase
filters, E(M) takes on the maximum possible value for each M, in the set of
filters having the same magnitude response. That is, for the minimum-phase
filter, the energy in the unit-sample response is concentrated toward the
beginning of the sequence.
Example 11.1. MiDimum-, linear- and maximum-phase systems. Since the
minimum-phase property of causal and stable systems is determined by the
locations of the zeros, consider the fourth-order all-zero system containing a
double complex-conjugate set of zeros located at z = O.7ei"/4.
The minimum-, linear- and maximum-phase system pole zero patterns
having the identical magnitude response are shown in Fig. 11.8.
The magnitude response and the phase responses of the three systems are
shown in Fig. 11.9. The minimum-phase system is seen to have the phase with the
smallest deviation from zero at each frequency.
The unit-sample responses for the three systems is shown in Fig. 11.~ For
the fourth-order FIR system, the unit-sample response has at most fivenonzero
values that occur for 0 ~ n ~ 4. The energy evolution factors for the three
different systems have the same value for M 2: 4. For the minimum-phase system,
however, E(M) takes on the maximum values for O~M~3.

11.4 FORMULATING THE PROBLEM

FOR APPLYING AN INVERSE FILTER
To apply an inverse filter, we must first produce a discrete-time model of the
distorting physical system. The inverse filter then serves to compensate this
distortion. To illustrate the approach, let us consider the model of a
bandlimited detector, denoted by Hm(z) and shown in Fig. 11.11. To observe
the limitations of the detector, we apply a unit-sample sequence {d(n)} to the
input of the model and observe the output sequence. {y(n)}. In the z domain,
since D(z) = 1, the response function is given by
Y(z)

= Hm(z)

(11.2)

To compensate for the bandlimited detector, we apply {y(n)} to the input of

an inverse filter, having the system function H;b). The desired output of the
inverse filter is a replica of the input sequence applied to the detector, or
{Yd(n)} = {d(n)}. In the z-domain, we have }d(z) = 1. Since the desired
system response is equal to
Yd(z) =Hiz) Hm(z)

(11.3)

the system function of the inverse filter is equal to

(11.4)

Hence, the system function of the inverse filter is the reciprocal, or inverse, of
the model, from which the name inverse filter is derived.
Let the system function of the detector model be given by
H ( ) = N(z)
m z
D(z)

(11.5)

430

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

(a)

MINIMUM-PHASE

.5

MIXED-PHASE
(LINEAR-PHASE)

r
0

I
I I
3

)on

MAXIMUM-PHASE

(b)

(M)

MINIMUM-PHASE

9.76
~--7''----

MIXED-PHASE
(LINEAR-PHASE)

FIGURE 1110
a) Unit-sample responses for mmlmum-

MAXIMUM-PHASE
2

5 M

phase, mixed-phase (linear-phase) and

maximum-phase systems in Example 11.1.
b) Comparison of energy evolution factors.

DESIRED OUTPUT
.-_--.(yin)}=(d(n)}

(d(n)}

-+n

MODEL OF
BAND LIMITED
DETECTOR
Hm(z)

,fT
~

.9_

INVERSE
FILTER

!--.
n

H,,(z)

FIGURE 1111
Model of bandlimited detector and application of inverse filter.

INVERSE FILTERING

where N(z) and D(z) are polynomials in

function is then given by
D(z)
H;t<.z) = N(z)

Z-l.

431

The inverse filter system

(11.6)

This result indicates that the poles of the inverse filter system function
correspond to the zeros of the model and the zeros of the inverse filter are the
poles of the model. For a stable and causal inverse filter, the poles of the
inverse filter system function must lie within the unit circle. Hence, a causal
and stable inverse filter can be implemented only if the physical system is
minimum-phase. The following examples illustrate this point.
Example 11.2. Inverse filter for the radar problem. Let us assume that the
radar pulse can be accurately given by
r"
hm(n) =-.-(-) sin[wo(n + 1)]
sm Wo

for n ~O

Using the results of Example 5.42, the radar pulse sequence is modeled by the
unit-sample response of the second-order filter defined by
1
H (z) = - - - - - - : - - : - - -2 : :
m
1- 2r cos(wo)Z 1 + r z 2

The system function of the inverse filter that compresses {hm(n)} into
{d(n)} is

Example 11.3. Inverse filter for the transmission channel problem. Let the
transmission channel exhibit only one echo, having amplitude a, where lal < 1,
and occurring K sample periods after the transmitted signal is received. Then the
unit-sample response of the channel is given by
and

hm(n) = 0, otherwise

The channel model system function is then equal to

Hm(z) = 1 + az- K

The system function of the inverse filter is then

HJ..z) = l/Hm(z) = 1/(1 + az- K )
It is left as an exercise for the student to show that the inverse filter converts
{hm(n)} into the unit-sample sequence {d(n)}.

Example 11.4. Inverse filter for nonminimum-pbase system. Let us consider a

physical system whose model system function is given by
Hm(z) = 1 + az- 1

where lal > 1

This system model is not minimum-phase because it has a zero outside the unit
circle at z = -a. The inverse filter system function is given by
H,,(z) = 1/Hm(z) = 1/(1 + az- 1)

432

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

Since the pole of Hiz) is outside the unit circle, the inverse filter is unstable, but
it still converts {hm(n)} into {d(n)} (show this). However, if this conversion is
not done exactly, or if there is any noise present, the output sequence of this
inverse filter will approach infinity.

The formulation presented above provides a procedure for designing

inverse filters in the z plane. However, in practice, it may not be possible to
determine the system function representation Hm{z) of the physical system,
even if the unit-sample response can be observed. We next describe how to
obtain an FIR filter approximation to the inverse filter from the observed
unit-sample response of the physical system.

11.S AN FIR FILTER APPROXIMATION

TO THE INVERSE FILTER
We now describe a method for determining an FIR filter approximation to the
inverse filter from an N-point observation of the model unit-sample response
hm{n), for OsnsN-1, denoted by {hm,N{n)}. Using {hm,N{n)}, we can
compute an Nd-point OFT, where NdN. We show below why it is desirable
to make the duration of the OFT sequence much larger than the number of
points in {hm,N(n)}. From the OFT sequence, we can determine the Nd-point
magnitude response sequence {IHm.N{k)1} and the phase response sequence
{Arg[Hm,N{k)]}. Since the exact inverse filter is related to the model system
function by H;r(z) = l/Hm(z), we define the magnitude and phase response
sequences of the FIR approximate inverse filter by
(11. 7)

and
Arg[Hifa{k)]

= -Arg[Hm,N{k)]

(11.8)

The Nd-point unit-sample response of the FIR approximate inverse filter,

denoted by {hifa{n)} is computed by performing the inverse OFT on the real
and imaginary sequences obtained from {1H;fa(k)1} and {Arg[Hifa{k)]}.
This procedure generates an inverse filter that is only an approximation
for the following reasons. For systems having finite-duration unit-sample
responses, we can acquire the entire unit-sample response and have
{hm,N{n)} = {hm{n)}. In this case, Hm(z) has only zeros and the exact inverse
filter is all-pole. Hence, the unit-sample response of the exact inverse filter
{hHCn)} has an infinite duration. The unit-sample response {hifa{n)} obtained
with the above procedure is a finite-duration approximation to {h;r(n)}. As
described in Chapter 4, since the OFT sequence {H;fa{k)} is the sampled
version of the exact inverse filter transfer function H;f{ej{J), then {hifa{n)} is the
first period of the periodic extension of {h;r(n)}, or
00

hifa{n) =

k=-oo

hin + kNd )

(11.9)

INVERSE FILTERING

433

To minimize the associated time-aliasing effects, the value of Nd , the size of the
Nd-point Off, should be made as large as possible.
In the case that {hm(n)} has an infinite-duration, the observed sequence
{hm,N(n)} is equal to the first N samples, or

hm,N(n)=hm(n)

forO:sn:SN-1

(11.10)

This is equivalent to windowing the infinite-duration sequence with an N-point

rectangular window. The Off sequence {Hm,N(k)} is then equal to the
samples of the convolved transfer function, described in Chapter 9, given by

Hm,N(e iw ) = 21
jf

r Hm(ei8)WR(eJ(W-8)

L1<

d8

(11.11)

where WR(eiW ) is the spectrum of the N-point rectangular window. This

convolution causes the transfer function of the approximate inverse filter to
deviate from that desired. In general, the agreement between Hm(ei "') and
Hm,N(e i",) improves as N is increased. The required value of N depends on the
application.
In practice, several problems could exist with this design procedure.
First, the values of the Off sequence {Hm,N(k)} computed from the observed
N-point unit-sample response may not be accurate for the spectral values that
are small. In the inverse filter, these small magnitudes correspond to large
gains when, the reciprocal is computed. This effect is especially true for
bandlimited models. A second problem occurs when only the magnitude
response of the system can be observed, rather than {hm,N(n)}. For example,
the characteristics of many physical systems such as transmission channels, or
the acoustic properties of biological tissue for diagnostic ultrasound applications, are known only in terms of their magnitude response. Another example
occurs in speech processing, in which speech sounds are classified according to
their power spectral properties. A third problem occurs when the observed
signals are corrupted with noise. In this case, only the average power spectral
sequence, which is usually modeled as the magnitude-squared response of the
model IHm(ei",W, can be computed. An approximation to the magnitude
response sequence can be found by taking the square-root of the average
power spectral sequence.
In the next section, we describe the procedure to compute the phase
response sequence from the magnitude response sequence for minimum-phase
models. The effects of noise will be considered in the last section.

11.6 DISCRETE mLBERT TRANSFORM

RELATIONSIHP FOR MINIMUM-PHASE
SYSTEMS
For discrete-time minimum-phase systems, the log-magnitude response and
phase response can be shown to form a discrete-time Hilbert transform pair.

434

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

Here we will only give the definition of this transform as it relates to designing
inverse filters. The derivation and other properties of the discrete Hilbert
transform are given in the references. It is sufficient to state that the phase
response sequence of the minimum-phase FIR filter can be determined from
the observed magnitude response sequence by employing the discrete-time
Hilbert transform. The observed magnitude response and the computed phase
response sequences can then be employed in the procedure previously
described in Section 11.5 to implement the inverse filter approximation.
Using the results of the discrete Hilbert transform, the phase response is
given by
Arg[H(dQl)) = 2~P

L: lo~

IH(ei8 )1 cot[(O - w)/2] dO

(11.12)

where the symbol P indicates the Cauchy principal value of the integral. The
Cauchy principal value is needed here because the cotangent function becomes
infinite when its argument becomes zero, or when 0 = w. In this case, the
integral must be evaluated through the limiting procedure given by
Arg[H(eiQl )] = lim 21
1C

--+0

[L

ro
-

lo~ IH(ei8)1 cot[(O -

w)/2] dO

(11.13)

-1<

f+e lo~ IH(ei~1

cot[(O - w)/2]

dO]

The forms of the two functions that are multiplied and then integrated, for a
particular value of w, are shown in Fig. 11.12. Because of the complicated
analytic forms, the evaluation of the above integral must usually be performed
numerically. The log-magnitude response sequence is then used to determine
the values of the phase response sequence. A simple program to evaluate this
integral is given in the Projects.

-IT

cot( '-;')
0

IIIIII~IT

'-8

--~~~~----~O~--~W--------IT~--8
I

I
I
I

I
I
I

FIGURE UU
Relationship of the functions for
evaluating the discrete Hilbert
transform.

INVERSE Fn..TERING

43S

From the integral equation, one restriction is obvious: IH(~QI)I cannot be

equal to zero, since the logarithm would then be infinite and the integral
unbounded. The Cauchy principal value argument is not applicable here
because a sign change does not occur in the logarithm as its argument passes
through zero. Hence, H(z) cannot contain zeros on the unit circle. But for
minimum-phase systems, all the singularities are inside the unit circle, and this
procedure can be employed to find the minimum-phase response.
Example 11.5. Computing the minimum-phase response by using the discrete
Hilbert transform. Consider the magnitude response given by
IHm(eiw)1 = e-.Biwi

for

- l r S W S lr

This magnitude response is a common model for the acoustic attenuation of

biological soft tissue employed in diagnostic ultrasound tissue characterization
described in the references.
The log-magnitude response is the linear-with-frequency function shown in
Fig. 11.13 and given by
log.IHm(eiW)1 = -(Jlwl

lr

11:

(e)

.. ,111IIIII/fllllllllll........... "

for 0 S w S

211:

FlGURE 11-13
Results of Example 11.5. a)
Logmagnitude
response
employed in Hilbert transform. b) Phase response
produced by Hilbert transform. c) The unit-sample
response determined by inverse DFf.

436

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

This function was sampled at 1024 points over the range 0:5 OJ :5 2n to obtain the
log-magnitude response sequence. The integral to determine the phase response
was performed numerically using the DHT subroutine listed in the projects. The
resulting phase response {Arg[Hm(eJW}]} of the minimum-phase model is shown
in the figure. The phase response has the minimum-phase characteristics of being
continuous-valued and returning to zero at OJ = n. The unit-sample response
{hm(n)} shown in the figure was obtained by computing the inverse DFT of the
real and imaginary sequences computed from these magnitude and phase
response sequences.

11.7 DESIGNING INVERSE FILTERS

WITH THE DISCRETE HILBERT
TRANSFORM
With the aid of the discrete Hilbert transform, we can compute the minimum
phase response from the magnitude response. For designing inverse filters, the
magnitude response employed in Hilbert transform IHinv(eiOO )I is equal to the
inverse of the model magnitude response, IHinV(eiOO)1 = IHm(eioo)I- 1 If
loSe IHinv(eiOO)1 is employed in Eq. (11.12), then Arg[HinV(eiOO )], the phase
response of the minimum-phase inverse filter, is obtained. Performing the
integration numerically produces the N-point sequence {Arg[Hinv(k)]}. From
this sequence and {IHmv(k)l}, the real and imaginary DFf sequences can be
determined. The real part of the inverse DFf is the N-point unit-sample
response of the inverse filter. The imaginary part should be zero. This
procedure is illustrated in the following examples.
Example 11.6. Inverse filter design with the discrete Hilbert transform for
aU-pole system model. Let us consider the second-order all-pole system model
given by
1
H (z) = ----:---:---=---=

1 _ 2r cos( 8)z

+ rZz

with r = 0.9 and 8 = niS. The magnitude response is then given by

IHm(eJW)1 = [1

+ a2 + b 2 + a(2 + b) cos(OJ) + 2b cOS(2OJWll2

where a = -2r cos(8) = -1.66 and b = r2 = O.S1. The "inverse magnitude response
function IHin.(eiW)1 = IHm(eiw)I- 1 This inverse magnitude response is shown in
Fig. 11.l4(a) and lo~ IHin.(eJW)1 is shown in Fig. 11.l4(b). Applying the discrete
Hilbert transform to this latter function produces the phase response shown in
Fig. 11.14(c). Using this phase response and IHin.(eiw)I, we find the real and
imaginary DFT sequences. Performing an inverse DFT produces the unitsample response {hinv(n)} shown in Fig. 1l.14(d). The sequence values that
are significantly different than zero are hin.(O) = 1, h in.(1) = a = -1.66 and
h inv(2) = b = 0.S1, or those fpr the exact inverse filter. The other values in
{hin.(n)} should be zero, but are not exactly zero because of errors in the
finite-precision arithmetic.

INVERSE FILTERING

(a)

IH,nie)"')1

(b)

log. IHon,(e)"')1

(e)

Arg [H,n,(e)O'

437

-It

(d)

(hon,(n)}

I I . . . . .. . . . . .
o

II

FIGURE 1114
Results of Example 11.6. a) Magnitude response
of inverse filter. b) Log-magnitude response
employed in Hilbert transform. c) Phase response produced by Hilbert transform. d) The
unit-sample response determined by inverse
DFf. 16-point sequences were used.

The previous example illustrates than when the system model contains
only poles, the inverse filter contains only zeros, and the inverse filter
unit-sample response has a finite duration. The size of the sequences used in
the discrete-Hilbert transform and the inverse OFT must then be equal to or
larger than the duration of the unit-sample response. If, however, the system
model contains a zero, then the inverse filter will contain the corresponding
pole. The unit-sample response of the exact inverse filter will then have an
infinite duration. To obtain a valid approximation, the size of the sequences
used in the Hilbert transform and inverse OFT computations should be large

438

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

enough that time-aliasing effects are not significant. To determine the size of
these sequences that will make the time-aliasing effect negligible, we examine
the pole in the inverse filter system function having the largest radius, equal to
rm. The unit-sample response will then decay on the order of r::'. The value of
N should be chosen such that ~ < 0.001 or 0.1 % of the maximum value.
The approximate value of the duration N can also be determined from the
bandwidth of the peak in the inverse filter magnitude response, as was done in
Chapter 7. As an alternate procedure, several values of N can be tried and the
unit-sample response sequence can be examined for each. If the values of the
unit-sample response do not change significantly as N is increased, then
time-aliasing effects may be neglected.
Example 11.7. Inverse filter design with the discrete Hilbert transform for
all-zero system model. Let us consider the second-order all-zero system model
given by
Hm(z) = 1 - 2r cos(6)z-1 + r2z- 2
with r = 0.9 and 6 = 1&/8. The magnitude response is then given by
IHm(eiw)1 = [1 +a 2 + b 2 + a(2+b) cos(w)+2b cos(2w)]112
where a = -2r cos(6) = -1.66 and b = r2 = 0.81. The inverse magnitude response
function IHmv(eiW)1 = IHm(eiw)rl. This inverse magnitUde response is shown in
Fig. 11.15(a) and lo~ IHinV(eiw)1 is shown in Fig. 11.15(b). Applying the discrete
Hilbert transform to this latter function produces the phase response shown in
Fig. 11.15(c). Using this phase response and IHinv(eIW)I, we find the real and
imaginary DFT sequences.
To determine the size of these sequences, we determine from the bandwidth
of IHinv(eiOl)1 that rm = 0.9. The value of N is then equal to
N = 10~[0.OOl]/lo~[rm] = 66

Performing a 64-point inverse DFT produces {hmv(n)}. The first 32 samples of

{hmv(n)} are shown in Fig. 11.15(d).

INVERSE FILTERS FOR NONMINIMUMPHASE SYSTEM MODELS. If the

system model is not minimum-phase, because a zero lies outside the unit circle,
the resulting filter is not a pure inverse-filter in the sense that a unit-sample
sequence is produced at the output when the unit-sample response of the
system model is applied. The output, however, does have a magnitude
spectrum that is constant with frequency. In Chapter 5, it was shown that a
zero outside the unit circle at radius Izl = rz > 1, generates the same form of
magnitude response, to within a constant, as a zero at radius Izl =.l/rz. Since
the discrete Hilbert transform employs the magnitude response to determine
the minimum-phase response, the resulting inverse filter is minimum-phase and
its corresponding pole appears at Iz I = l/rz.
In Chapter 8, we observed that all-pass filters have the property of a flat
magnitude response and a unit-sample response that is not a unit-sample

INVERSE FILTERING

439

IHonv(eJ"')1

(a)

It

log., IHonv(e'OI)1

(b)

o
(c) Arg (Honv(elW)]
It

-It

(d) {h;nv(n)}

11[[11,.IlIlll'."111,.. ....31.. n

FtGURE 11-15
Results of Example 11.7. a) Magnitude response of inverse filter. b) Log-magnitude response employed in Hilbert transform. c)
Phase response produced by Hilbert transform.
d) The unit-sample response determined by
inverse DFf.

sequence. The output of the inverse filter determined with the above
procedure, when the unit-sample response of the nonminimum-phase system
model is applied at the input, will be equal to the unit-sample response of an
all-pass filter.
EUIIlple U.8. Inverse filter design with the discrete HObert transform for a
Domuinhnam-pbase systelD lDodel. Let us consider the second-order all-zero
system model given by

440

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

with r = 1.11 and 8 = n/8. The magnitude response is then given by

IHm(ei"')1 = [1 + a2 + b 2 + a(2 + b) cos(w) + 2b cos(2w)]112
where a = - 2r cos( 8) = -2. OS and b = r2 = 1. 23. The inverse magnitude response
function IHmv(ej"')1 = IHm(eiOJ)rl. This inverse magnitude response is to within a
gain constant of that shown previously in Fig. 11.1S(a). This gain constant adds a
constant value to 108e IHinv(ei")I. Since the cotangent function has a zero mean
value, this constant additive term does not contribute to the evaluation of the
integral. Applying the discrete Hilbert transform to this latter function produces
the minimum-phase response sequence. Using this phase response and lH;nv(ei"')I,
we find the real and imaginary DFf sequences. Performing a 64-point inverse
DFf produces the same {hinv(n)} shown in Fig. 11.1S(d). When the unit-sample
response of the non-minimum phase system {hnm(n)} is applied to this inverse
filter, it produces the (y(n)} shown in Fig. 11.16. Note that (y(n)} has the form
of the unit-sample response of a second-order all-pass filter. As such, its
magnitude spectrum is constant with frequency, but its phase spectrum is not
identically zero. Hence, the inverse filter has corrected the magnitude response
characteristics of the system model, but a phase distortion still exists.

In the next section, we examine the effects of additive noise in the design
and performance of inverse filters.
(a)

(hnm(n)}

(b)

(c)

31

{h,nv(n)}

(y(n)}

~rr"'''''''''''''nTl''n'''''---~n

FIGURE 11-16
Results of Example 11.8 .. a) Unit-sample response of nonminimum-phase system. b) The unitsample response of inverse filter. c) Output of
inverse filter is identical to the unit-sample response of all-pass filter.

INVERSE FILTERING

441

11.8 THE EFFECTS OF NOISE ON

INVERSE FILTERS
In addition to the requirement that the physical system be minimum-phase, the
limitations of inverse filters for system compensation and pulse compression
are usually encountered in the presence of noise. To illustrate these limitations, let us consider the noise to be white, implying that the power spectral
sequence, defined in Chapter 10, is constant with frequency, on average. For
example, consider the observed sequence {x(n)} to be the sum of the signal
sequence {s(n)} and a white noise sequence {w(n)}, given by
{x(n)}

= {s(n)} + {w(n)}

(11.14)

This case is shown in Fig. 11.17.

For random signals, we must compute average power spectra. Typically,
an N-point data window, like the Hamming window, is applied to the observed
-random signal to obtain an N-point observed sequence. If W;(e Jeo ) is the
Fourier transform of the ith N-point observed sequence {wi(n)}, and M such
independent sequences are observed, then the average power spectrum,
Pw(eieo ), is defined as

= lim

Pw(eieo )

M ....""

= lim
M ....""

[_1_
f W;(e
MN

ieo )

W;"(ei eo )]

i=1

[-

1M

.]
L 1W;(eJeoW

(11.15)

MNi=l

The division by N ensures that the Fourier transform remains finite for any
value of N. Since the noise sequence {w(n)} is random, the observed signal
{x(n)} is also random. A similar expression can be written for the power
spectrum of {x(n)}, denoted Px(eioo ), given by
Px(e ioo )

= lim
M ....""

[_1_ X,(e JW )
MNi=l

= lim [_1_
M ....""

x:(e

JW

)]

f IX,(eJOOW ]

(11.16)

MN,=!

{w(n))
I

{s(n)}

.o,I h. 111'

'"o
,It

l,t

t
t ...
""~n

".1.~. ~
n

FIGURE 1117

An example of adding random noise to a pulse waveform.

{x(n)}

"nIh "11"...

~n

442

INTRODUcnON TO DIGrrAL SIGNAL PROCESSING

For the nonrandom sequence {s(n)}, the power spectrum Ps(eiQ is equal to
Ps(eiQ = S(ei"') S*(ei"')
(11.17)

= IS(ei"'W
In many practical cases, it can be argued that the noise is statistically
independent of the signal. The power spectrum of the observed signal is then
equal to
(11.18)
The inverse filter design procedures considered above can be caUed
noiseless, since noise effects were not considered. The noiseless inverse filter is
designed to compensate for the spectrum of the signal and neglects the noise.
The magnitude response of the noiseless inverse filter is then given by
(11.19)
This result shows that the noiseless inverse filter gain is a maximum at the
frequency where the signal spectrum is a minimum. When we try to apply this
noiseless inverse filter to the signal with noise, the noise components in the
signal are amplified in the output, and the results are usually disappointing.
INCLUDING NOISE EFFECfS INfO THE DESIGN OF mE INVERSE

FILTER. If the characteristics of the noise are included in the design of the
inverse filter, the performance can be improved slightly. The gain of the
noiseless inverse filter is a maximum at the frequencies where the signal
spectrum is a minimum. If there are noise components at the frequency range
where the signal spectrum is low, these will dominate the output. There is a
trade off between the amount of pulse compression that can be obtained and
the amount of noise that is present in the output.
To illustrate this trade off, let us denote the inverse filter magnitude
response that includes the noise effects by IH.+w(ei "') I and let it be defined by
.

IHs+w(el"')1

= Ps(ei"') + APw(ei"')

(11.20)

where A is a parameter that controls the amount of compression. If A= 0, the

noiseless inverse filter results, providing the maximum pulse compression, but
also allowing the maximum effect of the noise to occur at the output. If
APw(ei"')>> Ps(eiQ, IHs+w(eiQ)W = 1/(APw(eiQ) and the filter acts to suppress
the noise without providing significant pulse compression. For the important
practical case of white noise, the noise power spectrum is equal to a constant
over frequency, or Pw(ei "') = wij.
The design of this inverse filter uses the same procedure as described in the
previous section, the only changes being that IHs+w(eIW)1 is used rather than
I/S(eiW) and that a value for A must be specified. The following example
illustrates this procedure.

INVERSE Fn.TERING

443

Desiping an inverse liter for signal with noise. Let us consider

the system model described by
(1- 0.95einf4z- 1)(1- 0.95e- inf4z- 1)

EuJDple 11.9.

Hm(z) = (1- 0.C)ei-z-l)(1- O.ge- ja16z 1)(1- 0.geln13z 1)(1_ 0.ge

j"I3Z -I)

Let us assume that the observed signal is equivalent to having a white noise
sequence {w(n)}, that is uniformly distributed between -0.5 and 0.5, added to
the output of this system. The noise spectrum is then constant with frequency and
Pw(ei .. )

= W~ =-b.

We will investigate the effect of choosing the value of A by examining the

performance of the filter under both noise-free and added noise conditions when
A takes on the values 0, 0.5, 2 and 8.
The first step is to determine the magnitude response of the inverse filter
that is given by
IHHW(ei")1 = [Ps(e i ")

~ AW~j112

where the pulse spectrum is given by

Ps(e l " ) = IHm(ei"W

The forms of fH.+w(eIW)1 for the different values of A are shown in Fig.
11.18(a). We note that the zero in Hm(z) at Z =0.95einf4 produces a peak in these
transfer functions at (JJ = ,,/4. The gain is also large for high frequencies to
compensate for the general lowpass characteristic of Hm(eIW ). The A = 0 curve
corresponds to the magnitude response of the noiseless inverse filter. As A
increases, the values of the maximum gains decrease to reflect the dominance of
noise in those frequency ranges. As A increases, the curves become flatter,
indicating the increasing dominance of the constant spectral component AW~
These magnitude curves were sampled at 128 points to generate the
sequence {lH.+ ..(k)l}. The logarithm of this sequence was then employed in Eq.
(11.12) to determine the phase response sequence {Arg[HHw(k))}. The phase
sequences for the different values of A are shown in Fig. 11.18(b). As A increases,
the phase response tends toward zero.
Having {IH.+ ..(k)l} and {Arg[HH ..(k))}, the 128-point inverse Off can be
computed to determine the unit-sample response of the inverse filter {hHW(n)}.
The first 32 points of the sequences obtained for the different values of A are
shown in Fig. 11.19. We note that the zero in Hm(z) produces an infinite-duration
component in the sequences. As A increases, these unit-sample responses
approach the unit-sample sequence {d(n)}. a result which is consistent with the
flattening of the magnitude and phase spectra.
The pulse compression behavior of these filters can be examined by their
response to the noiseless pulse sequence {hm(n - no)}, shown in Fig. 11.20(a).
The output sequences are shown in Fig. 11.20( b) for the different values of A. For
A = 0, the unit-sample seque~ce occurring at n = no is to be expected. Upon
closer examination, we note a slight residual component for n > no. This residual
is due to the time-aliasing effect produced with the 128-point sequences and can
be reduced by increasing the size of the DFf sequences. As A increases, the pulse

444

INTRODUcnON TO DIGITAL

'';',AL PROCESSING

A=

OL-______

________________________

"'.

ro

It

It

A=

A=

It

--It

ftGURE 1118
Response curves of inverse filter for signal in
additive noise, for different values of A. a) Magnitude responses employed
in Hilbert transform. b)
Phase response produced
by Hilbert transform.

compression capability of the inverse filter decreases and the form of the output
seqlJ.ences approaches that of {hm(n - no)}.
The behavior of the inverse filter in the presence of noise can now be
examined. The input to the filter is the delayed pulse sequence plus noise
{hm(n - no) + w(n)} shown in Fig. 1l.21(a). The output sequences that are
observed when this input is applied to the different filters are shown in Fig.
1l.21(b). As A decreases the effects of the noise become more apparent in the
output.
To quantify this observation, the variance of the noise in the output W2
was computed. This was done by subtracting the noiseless outputs, shown if'

INVERSE FILTERING

445

A. =0
~~~~~~~~~~~UU~rr-n

31

A. =05

A. =2

A. =8

..I...
01.0

.11.

on

.. .. .

.ILn

31

FIGURE 1119
Unit-sample responses determined
by inverse DFf for different values
orA.

Fig. 1l.20(b), from the respective sequences in Fig. 11.21(b). The average of the
squared residuals then provides an estimate of the noise variance W 2 The
computed values of W 2 are given in the figure for the different values of A.. As A.
increases, the value of W 2 approaches W~, the variance of the noise in the input
sequence.
Summarizing these results, we note that when noise is present there is a
trade off between the ability to compress the pulse and to suppress the noise.
When the maximum pulse compression, {hm(n - no)} -+ {d(n - no)}, is attempted, the noise component in the output has the highest power. The degree of
pulse compression that should be attempted depends on the particular application
and specification to be achieved.

446

INTRODUcrION TO DIGITAL SIGNAL PROCESSING

(a) Input: (hm(n-no)}

(b) Output: (y(n)}

A =0

ft

we

no

A =0.5

.u

A =2

. , . n

.... I. ....,i'"'. ... ..... ... ..

no

A =8
no

ill'

...

FIGURE 1120
Noiseless input results. 0) Input is a delayed unit-sample response of the minimum-phase system.
b) Output of inverse filters having unit-sample response shown in Fig. 11.19.

INVERSE FILTERING

(a) Input: (hm(n-nll)

447

+ w(n)

(wi =

0.083)

(b) Output: (yen)}

)..=0

),,=0.5

),,=2

)..=8

. IItt

(W2=0.09)

II

IT. I .

11'Ino I III 'iI Ii 111

FIGURE 11-21
Results with input contain:ng
additive noise. a) Input is a
delayed unit-sample response
of minimum-phase system plus
white noise. b) Output of inverse filters having unitsample response shown in Fig.
11.19. The variance of the
output noise is equal to W 2

448

INTROOUCllON TO DIGITAL SIGNAL PROCESSING

11.9 SUMMARY
This chapter began by considering two comm<fn distortions that occur to
observed signals and by defining discrete-time models that generate these
distortions. Inverse filters were found to work best on models whose system
functions have the minimum-phase property. The minimum-phase property
was defined in terms of the pole/zero pattern in the z plane of the system
function of the model Hm(z), the phase response Arg{Hm(ei "')} and the
unit-sample response {hm(n)}. Two procedures for designing the inverse filter
were described. The first performed the design in the z plane and the second in
the frequency domain. The z plane procedure produced the system function of
the inverse filter directly from the model pole/zero pattern: the poles of the
inverse filter were located at the positions of the model zeros and the filter
zeros at the positions of the model poles. This procedure is limited to models
for which either the system function or the pole/zero pattern is known. The
frequency-domain approach i~volved computing the DFf sequence of the
observed unit-sample response of the model to implement an approximate
inverse filter. The types of approximations that resulted by employing this
procedure were described. In cases for which only the magnitude response of
the model is known, an FIR approximation to the inverse filter can be
designed by using the discrete-time Hilbert transform. For minimum-phase
systems, the log-magnitude response sequence is the discrete-time Hilbert
transform of the phase response sequence. This useful result was employed for
designing inverse filters. The chapter ended by considering the problem of
designing inverse filters for signals in the presence of noise.
A common alternate procedure for designing inverse filters in the time
domain requires concepts that are not difficult, but sufficiently different from
the topics covered in this text that they were not covered here. These
time-domain design procedures can be found in the references.

REFERENCES TO TOPICS
General inverse filtering applications
Arya, V. K. and J. K. Aggarwal: Deconvolution of Seismic Data, Hutchinson Ross, Stroudsburg,
PA,1982.
Mendel, J. M.: Optimal Seismic Deconvolution, Academic Press, New York, 1983.
Anderson, B. D. O. and J. B. Moore: Optimal Filtering, Prentice-Hall, Englewood Cliffs, NJ,
1979.
Time-domain design 01 inverse filters
Makhoul, J.: "Linear prediction: A tutorial review," Proc. IEEE, 63, 561-580 (1975).
Robinson, E. A., and S. Treitel: Geophysical Signal Processing, Prentice-Hall, Englewood Cliffs,
NJ,1980.
Schwartz, Mischa, and Leonard Shaw: Signal Processing: Discrete Spectral Analysis, Detection,
and Estimation, McGraw-Hill, New York, 1975.
Treitel, S., and E. A. Robinson: "The design of high-resolution digital filters," IEEE Trans.
Geosci. Electron., GE-4: 25-38 (1966).

INVERSE FILTERING

449

Discrete-l6Ibert tnDIfona
Oppenheim, Alan V. and Ronald W. Schafer: Digital Signal Processing, Prentice-Hall, Englewood
Cliffs, NJ. 1975.
Adaptive filters

Cowan, C. F. N., and P. M. Grant: Adaptive Filters, Prentice-Hall, Englewood Cliffs, NJ, 1985.
Goodwin, G. C., and K. S. Sin: Adaptive Filtering, Prediction and Control, Prentice-Hall,
Englewood Cliffs, NJ, 1984.
..
Haykin, S.: Adaptive Filter Theory, Prentice-Hall, Englewood Cliffs, N.J, 1986.
Honig: M. L., and D. G. Messerschmitt: Adaptive Filters, Kluwer, Boston, 1984.
System aaodeliDg

Kuc, R.: "Modeling acoustic attenuation of soft tissue with a minimum-phase filter." Ultrasonic
Imaging Journal, 6: 24-36 (1984).

PROBLEMS
11.1. Show that a linear-phase filter cannot be minimum-phase by considering the
pole/zero pattern of this system.
11.2. Determine whether the following unit-sample responses are from a minimumphase filter. If not, indicate why not.
(a) hen) = 1
for O~n ~N-l
(b) h(n)=n+l
forO~n~N-l
(c) h(n)=r"cos(wo)n
forn~O, h(n)=O, otherwise
(d) h(n) = r" sine wo)n
for n ~ 0, h (n) = 0, otherwise
11.3. Consider a minimum-phase FIR filter having the unit-sample response given by
hmin(n), for 0 ~ n ~ N - 1. Let the maximum-phase FIR filter, with the identical
magnitude response, have the unit-sample response hmax(n), for 0 ~ n ~ N - 1.
Show that {hmax(n)} has the same values as {hmin(n)}, but in time-reversed
order, or
hmo.(n) = hmin(N -1- n)
for O~n ~N-l
11.4. Given any linear-phase FIR filter, how can this filter be transformed into a
minimum-phase filter without changing the magnitude response?
11.S. Given the unit-sample response of a minimum-phase filter {hen)}, show that a
second filter having the unit-sample response {h2(n)} = {hen - k)}, where k is
any nonzero integer, is not minimum-phase.
11.6. Consider the causal unit-sample response {hen)}, for which h(O) = o. Show that
this filter cannot be minimum-phase.
11.7. Consider a physical system whose model system functionis given by

Hm(z) = 1 + az- I

where lal > 1

and whose unit-sample response is given by {hm(n)}. This system model is not
minimum-phase because it has a zero outside the unit circle at Z = -a. The
inverse filter system function given by

H.b) = 1/ Hm(z) = 1/(1 + az-I)

450

lNTRODUcnON TO DIGITAL SIGNAL PROCESSING

is then unstable. Consider using instead the filter defined by

H(z) = (1/a)/[1

+ (1/a)z-1]

(a) What is the output sequence of H(z) when {h",(n)} is the input sequence?
(b) What is the magnitude spectrum of this output sequence?
11.8. Consider the fourth-order minimum-phase FIR system Hmia(z) having all four
zeros at z = 0.9.
(a) Compute the magnitude response of this system.
(b) For each of the possible phase responses having this magnitude response
form, calculate and sketch the phase response and the unit-sample response.
11.9. Inverse filter of an ARMA model. Let the system function of an ARMA model
be given by
1

H (z) = 1- O.5zm
1+0.5z 1

Determine and sketch the following:

(a) the z plane pole/zero pattern of the ARMA model
(b) the unit-sample response of this model
(c) the pole/zero pattern and block diagram of the inverse filter
11.10. Repeat Problem 11.9 for the system function of an ARMA model give:J by

1-0.81z- 2
Hm (z)=1_1.27z 1+0.9z

COMPUTER PROJEcrs
11.1. Inverse filter for the radar system
Objed. Simulate a radar system and to perform inverse filter design in the z
plane.
Generation of the ideal radar return. Let r(n), for O:$n :$63, be the
realization of an ideal radar return having the following values:
r(1) = 1

(isolated return)

r(31) = 1

r(32) = -1

r(n) = 0

otherwise

r(33) = 1 (multiple return)

The obseroed radar return, denoted by {s(n)}, is the first 64 samples of the
convolution of {r(n)} with a pulse waveform {p(n)}. Let {p(n)} be the
unit-sample response of the AR system described by the difference equation
given by
y(n) = 1.8 cos(.n"/8)y(n -1) - O.81y(n - 2) + x(n)

-1.9cos(n/4)x(n -1) +O.9025x(n -2)

The goal is to retrieve {r(n)} from {s(n)}.

INVERSE FILTERING

451

Analytic results
(a) Determine the Z plane pole/zero pattern of the AR system.
(b) Determine and sketch the block diagram for the inverse filter.

Computer results
(a) For the pulse waveform {p(n)}, use only the first 30 samples of the
unit-sample response of the AR system.
(b) Verify the operation of the inverse filter by computing and plotting the
output when {p(n)} is the input. Why is the output for n =30 not equal to
zero?
(c) Generate and plot the observed radar sequence {s(n)}.
(d) Compute and plot the output of the inverse filter when {s(n)} is the input.
11.2. IDverse fiber of the reverbennt eommunkatiou channel
Object. Simulate a reverberant communication channel and construct an inverse
filter.
Communication channel model. Let the system function of a reverberant
channel be given by

Hm(z) = 1- 0.81z- 2
AaaIytic results. Determine and sketch the following for this channel model:
(a) the z plane pole/zero pattern
(b) the unit-sample response
(c) the pole/zero pattern and block diagram of the inverse filter

Computer results. Implement the channel model and inverse filter. Perform the
following.
(a) Plot the unit-sample response of the channel model.
(b) Plot the output of the inverse filter when sequence in Part (a) is the inut.
(c) Plot the output of the channel filter when the FSK sequence, generated in
Chapter 2, is the input.
(d) Plot the output of the inverse filter when the sequence in Part (c) is the input.
11.3. Approldmate FIR inverse fiber for an FIR system model
Object. Determine the approximations of this inverse filter design procedure.
Consider the FIR system model to be given by

Hm(z) = 1- 0.9z- 1
We want to find the inverse filter using the frequency-domain procedure.
Analytic: results. Determine the true inverse filter and sketch its unit-sample
response {hin)}.

Computer results
(a) Plot the model unit-sample response {hm(n)}. Compute and plot the Nd-point
magnitude and phase sequences {IHm(k)l} and {Arg[Hm(k)]} for Nd = 16.
(b) Compute and plot {hif.(n)} , the unit-sample response of the FIR approximate inverse filter determined from {lHm(k)I-I} and {-Arg[Hm(k)]}.
Compare with {hif(n)}. Verify that {h'f.(n)} corresponds to the first period of
the periodic extension of {hin)}.
(c) Repeat for Nd = 64.

452

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

11.4. ApproJdmate FIR invene filter for an UR system model

Object. Determine the approximations of this inverse filter design procedure.
Consider the IIR system model to be given by

Hm(z) = 1/(1'- 0.9z- 1)

We want to find the inverse filter using the frequency-domain procedure.
Analytic results. Determine the true inverse filter and sketch its unit-sample
response {h;t{n)}.
Computer results
(a) Compute and plot the unit-sample response {hm,N(n)}, where
for O::S:n ::s:N-1
Let N = Nd'= 16. Compute and plot the N-point magnitude and phase
sequences {lHm,N(k)l} and {Arg[Hm,N(k)]}.
(b) Compute and plot {hifa(n)}, the unit-sample response of the FIR approximate inverse filter determined from {lHm,N(k)l-l} and {-Arg[Hm,N(k)]}. To
observe the pulse compression performance, compute and plot 55 points of
the output of this approximate inverse filter when {hm(n)} is the input.
(c) Repeat for N=Nd=32.
11.5. Inverse filter design using the Hilbert transform for an FIR system
Object. Determine performance of the inverse filter design procedure for FIR
systems.
Consider the FIR system model to be given by

Hm(z) = 1- 0.9z- 1
Let us assume that we only know the magnitude response of this system, from
which we want to design the inverse filter.
Analytic results
(a) Determine the true inverse filter and sketch its unit-sample response {hir{n)},
(b) Calculate and sketch the magnitude response IHm(eiw)1 of this system.
Computer results
(a) Compute and plot the N-point log-magnitude response sequence {lHm(k)ldB}'
Let N= 16.
(b) Use the discrete-Hilbert transform subroutine DHT, given below, to compute
the phase response. Plot the result and compare with that of the true filter.
(c) Compute and plot the unit-sample response of the FIR approximate inverse
filter, Compare with {hin)}.
(d) Repeat for N = 64.
SUBROUTINE DHT(XLMAG, XPHS, N)
C
C
C
C
C
C

XLMAG (REAL INPUT ARRAY) CONTAINS THE LOG-MAGNITUDE

SEQUENCE IN DB
XPHS (REAL OUTPUT ARRAY) IS THE COMPUTED PHASE SEQUENCE
N (INTEGER INPUT) ISS I ZE OF ARRAY (POWER OF 2
CORRESPONDING TO NUMBER OF POINT OVER 0 oS (J) < 2Jl'. )

INVERSE FILTERING

453

DIMENSION XLMAG(N) ,XPHS(N)

N2=N/2
N2Pl=N2+1
NP2=N+2
PI=4. *ATAN(l.)
DBTONP=l. /8.69
DOMGA=2.*PI/FLOAT(N)
XPHS(l)=O.
XPHS(N2Pl)=O.

CONVERSION FROM DB TO NAPERIAN LOG

FREQ INCREMENT
PHASE AT w=O AND w=;r IS ZERO

DO 1 I=2,N2
II=I
TYPE* ,II, I / ' ,N2
SUM=O.
OM=FLOAT(I-l) * DOMGA

INDICATES PROGRAM PROGRESS

C INTEGRATION LOOP
C

2
1

DO 2 J=l,N
IF ( J EQ. I) GOTO 2
THETA=FLOAT(J-l) *DOMGA
ALPHA=(THETA-OM)/2.
COTAN=COS(ALPHA)/SIN(ALPHA)
SUM=SUM+DBTONP*XLMAG(J)*COTAN
XPHS(I)=SUM/FLOAT(N)
XPHS(NP2-I)=-XPHS(I)
RETURN
END

JUMP OVER COTAN(O) POINT

PHASE IS ODD SEQUENCE

11.6. Inverse filter design using the Hilbert transfonn for an IIR system
Object. Determine performance of the inverse filter design procedure for FIR
systems.
Consider the IIR system model to be given by

Hm(z) = 1/(1- O.9z- 1)

Let us assume that we only know the magnitude response of this system, from
which we want to design the inverse filter.
Analytic results
(a) Determine the true inverse filter and sketch its unit-sample response {hin)}.
(b) Calculate and sketch the magnitude response IHm(eIW)I.
Computer results
(a) For N = 32, compute and plot the log-magnitude response sequence
{IHm(k)ldB}'
(b) Use the discrete-Hilbert transform subroutine DHT to compute the phase
response. Plot the result and compare with that of the true filter.
(c) Compute and plot the unit-sample response of the FIR approximate inverse
filter. Compare with {hAn)}.

"This page is Intentionally Left Blank"

APPENDIX A

PROGRAM LISTINGS
The computer programs listed below evaluate the DFf using the fast
Fourier transform and provide the graphics routines for displaying arrays on
the terminal and transferring them to a file for printer hard copy. The FFT
routines have been adapted from the Cooley-Tukey fast Fourier transform
algorithm (ref: Cooley, Lewis and Welch, IEEE Trans. Education, 12,27-34,
1969). Programs to test these routines are listed and sample print-outs are
presented. The routines include:
PLOT - to display an array of arbitrary size with a vertical abscissa.
FFf - to compute the DFT with the fast Fourier transform algorithm.

PLOTMR - to display the 65 unique values of a 12S-point magnitude

response.
PLOTPR - to display the 65 unique values of a 128-point phase
response.
PLOTLM - to display the 65 unique values of a 128-point logmagnitude response.
The FORTRAN and C-Ianguage programs have been written on a
45S

456

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

VAX* -780 computer system using the VMS operating system. The Pascal
programs have been written on a DEC Rainbow* PC with MS-DOS and using
Turbo Pascal, available from Borland International, 4585 Scotts Valley Drive,
Scotts Valley, CA 95066.
Inputs to programs:
X - real-valued array.
XR and XI - arrays containing the real and imaginary components of
the DFf.
NX - size of array.
NPRNT (FORTRAN and Pascal) - causes graph to be printed when
NPRNT= 1.
DSK (Pascal) - disk file name for data output.
LOG2N - integer equal to log2(N) for N-point DFf.
NTYPE - integer, when = 1, DFf is performed; when = - 1, IDFf is
performed.
DBSCAL - 0) defines lower limit of log-magnitude graph.
SPEC - 0) defines specification level to be marked on log-magnitude
graph.
IF1 and IF2 - two integers indicating DFf bin numbers to be marked
on graph.

A trademark of Digital Equipment Corporation.

APPENDIX A

CALL PLOT(X,45,NPRNT)
TYPE*, 'GO AGAIN? (1/0)'
ACCEPT*, IANS
IF(IANS.EQ.1) GOTO 1
STOP
END

FORTRAN programs.
SUBROUTINE PLOT(X,NX,NPRNT)
DIMENSION X( NX)
BYTE ARR(71) ,MARK
C FIND MINIMUM AND MAXIMUM VALUES IN ARRAY
XMAX=O.
XMIN=O.
DO 1 I=1,NX
IF(X(I).GT.XMAX) XMAX=X(I)
IF(X(I).LT.XMIN) XMIN=X(I)
IF(XMAX.LT.-XMIN) XMAX=-XMIN
IF(XMAX.EQ.O.) XMAX=1.
TYPE 200,XMAX
IF(NPRNT.EQ.1) PRINT 200,XMAX
200 FORMAT(' MAX VAL- ',F6.2)
C IPOS IS DOTTED LINE POSITION
IPOS=36
IF(XMIN.GE.O.) IPOS-1
IF(XMAX.LE.O.) IPOS-71
C SCALE TO PRINTER COORDINATES
SCALE=70./XMAX
IF(IPOS.EQ.36) SCALE-35./XMAX
MARK-'-'
DO 2 I=I,NX
RNDOFF=.5001
IF(X(I).LT.O.) RNDOFF--.5001
N=IFIX(X(I)*SCALE+RNDOFF)+IPOS
DO 5 J=I,71
5
ARR(J)=MARK
ARR(IPOS) =' I '
ARR(N)='*'
TYPE 100,I-I,ARR
IF(NPRNT.EQ.I) PRINT 100,I-I,ARR
100 FORMAT(' ',I3,'-',7IAI)
2
MARK-"
RETURN
END -

5
6

Program to test PLOT.FOR:

C TESTO FORTRAN
DIMENSION X(45)
TYPE*,' FOR SINE FCN - ENTER 0'
TYPE*,' FOR COS FCN - ENTER l '
ACCEPT*, ITYPE
TYPE*,' ENTER FREQ (0<F<0.5)'
ACCEPT*, F
TYPE*,'PRINT GRAPH (1/0)'
ACCEPT*, NPRNT
PI=4.*ATAN(I.)
DO 2 1=1,45
ALPHA-2.*PI*F*FLOAT(I-I)
IF(ITYPE.EQ.I) X(I)=COS(ALPHA)
IF(ITYPE.EQ.O) X(I)-SIN(ALPHA)
2
TYPE 102,I-I,X(I)
102 FORMAT(' X(',I2,')=',F6.2)

457

10

20

30

SUBROUTINE FFT (LOG2N,XR,XI,NTYPE)

DIMENSION XR(I),XI(1)
SIGNa-I.
IF(NTYPE.LT.O) SIGN=I.
N=2**LOG2N
NV2=N/2
NMI =N-I
J=l
DO 7 I=I,NMI
IF(I.GE.J) GOTO 5
TR=XR(J)
TI=XI(J)
XR(J)=XR(I)
XI(J)=XI( I)
XR(I)=TR
XI( I)=II
K=NV2
IF(K.GE.J) GOTO 7
J=J-K
K=K/2
GOTO 6
J=J+K
P1=4. *AIAN( 1.)
DO 20 L-I,LOG2N
LE-2**L
LEI-LE/2
URal.
UI=O.
WR=COS(PI/LEI)
WI-SIGN*SIN(PI/LE1)
DO 20 J-I,LEI
DO 10 I-J,N,LE
IP-I+LEI
TR=XR(IP)*UR - XI(IP)*UI
TI-XR(IP)*UI + XI(IP)*UR
XR(IP)-XR(I) - IR
XI(IP)-XI(I) - II
XR( I) =XR( I) + IR
XI(I)-XI(I) + II
IR-UR*WR - UI*WI
TI=UR*WI + UI*WR
UR-TR
UI-Tl
IF(NTYPE.GI.O) REIURN
AIN=I./N
DO 30 I-I,N
XR( I)=XR( I)*AIN
XI(I)-XI(I)*AIN
REIURN
END

458

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

SUBROUTINE PLOTMR(X,NPRNT)
DIMENSION X(65)
BYTE ARR(65) ,MARK
C FIND MAXIMUM OF X ARRAY - XMAX
TYPE 200
IF(NPRNT.EQ.l) PRINT 200
200 FORMAT(' MAGNITUDE RESPONSE')
XMAX=O.
DO 1 1=1,65
XMAX=AMAXl(XMAX,X(I
SCAL = XMAX/20.
SCALD2=SCAL/2.
DO 3 1=1,21
C LEVEL OF GRAPH LINE
GLVL=FLOAT(21-I)*SCAL
MARK=' ,
IF(I.EQ.l.0R.I.EQ.21) MARK='-'
DO 4 J s l,65
4
ARR(J)-MARK
ARR(l )=' I'
ARR(65)-' I'
DO 5 J-l,65
A-ABS(X(J)-GLVL)
5
IF(A.LT.SCALD2) ARR(J)-'*'
IF(NPRNT.EQ.l) PRI~100'GLVL.ARR
3
TYPE 100,GLVL,ARR
100 FORMAT(' 'F6.2,lX,65 1)
RETURN
'
END
SUBROUTINE PLOTPR(X,NPRNT)
DIMENSION X(65)
BYTE ARR(65) ,MARK
PI=4. *ATAN( 1.)
SCAL-PI/10.
TYPE 101
IF(NPRNT.EQ.1) PRINT 101
101 FORMAT(' PHASE RESPONSE')
SCALD2-SCAL/2.
DO 3 '1=1,21
PHLVL=FLOAT(ll-I)*SCAL
MARK=' ,
IF(I.EQ.1) MARK-'-'
IF(I.EQ.11) MARK-'-'
IF(I.EQ.21) MARK='-'
DO 4 J=l,65
ARR(J)=MARK
ARR(l )=' I'
ARR(65)-' I'
DO 5 J-1,65
A-ABS(X(J)-PHLVL)
5
IF(A.LT.SCALD2) ARR(J)-'*'
IF(NPRNT.EQ.1) PRINT 100,PHLVL,ARR
3
TYPE 100,PHLVL,ARR
100 FORMAT(' 'F6.2,lX,65Al)
RETURN
END

SUBROUTINE PLOTLM(X,DBSCAL,SPEC,
+
IFl, IF2 ,NPRNT)
DIMENSION X(65) ,Y(65)
BYTE ARR(65) ,MARK
TYPE 101
IF(NPRNT.EQ.l) PRINT 101
101 FORMAT(' LOG-MAGNITUDE RESPONSE')
XMAX=-1000.
DO 1 1=1,65
XMAX=AMAX1(XMAX,X(I
C SCALE ARRAY BETWEEN 0 AND DBSC~L
DO 2 1=1,65
Y(1)=X(1)-XMAX
2
Y(I)=AMAXl(Y(I),DBSCAL)
SCAL=-DBSCAL/20.
SCALD2=SCAL/2.
DO 3 1=1,21
MARK=' ,
IF(I.EQ.1.0R.I.EQ.21) MARK='-'
DO 4 J=l,65
4
ARR(J)=MARK
C INDICATE FREQUENCY MARKS AND LIMITS
ARR(l )=' I'
IF(IF1.GT.0.) ARR(IFl)='I'
IF(IF2.GT.0.) ARR(IF2)-'I'
ARR(65)=' I'
C DB LEVEL OF GRAPH LINE
DBINC=-FLOAT(I-1)*SCAL
DO 5 J-l,65
C SPECIFICATION LINE
A-ABS(SPEC-DBINC)
IF(A.LT.SCALD2) ARR(J)-'-'
C DATA POINT
A-ABS(Y(J)-DBINC)
5
IF(A.LT.SCALD2) ARR(J)='*'
IF(NPRNT.EQ.1)PRINT 100,DBINC,ARR
3
TYPE 100, DBINC,ARR
100 FORMAT(' 'F6.2,lX,65A1)
RETURN
END
Program to test FFT.FOR, PLOTMR.FOR,
PLOTPR.FOR and PLOTLM.FOR:
C

TEST1 FORTRAN
DIMENSION X(128),Y(128)
DO 1 1=1,128
X(1)=O.
Y(I)-O.
DO 2 1=1,8
X(1)-1.
CALL FFT(7,X,Y,l)
DO 3 1-1,65
A-X(1)
B-Y( 1)
X(I)-SQRT(A**2+B**2)
PHS-D.

APPENDIX A

1F(A.NE.O AND.B.NE.O.) PHS=ATAN2(B,A)

3

459

if x[i] < 0 then rndoff := -0.5001;

ival:~trunc(x[i]*scale+rndoff)+offset;

y(I)~PHS

NPRNT-I
CALL PLOTMR(X,NPRNT)
CALL PLOTPR(Y,NPRNT)
DBSCAL--40
SPEC--20
IFl-20
1F2-30
DO 4 1-1,65
1F(X(1).LE.O.OOOI) X(1)=.OOOI
X(1)=20.*ALOGIO(X(I
CALL PLOTLM(X,DBSCAL,
+
SPEC,IFI,IF2,NPRNT)
STOP
END

Pascal programs '

procedure plot(x:vector; nx,nprnt:integer;

var dsk:text);
var
i, j. zpt. ival, m, offset: integer;
xmx, xmn, scale, rndoff: real;

arr: array[I 71] of char;

dot: char;
begin
writeln; if nprnt = I then writeln(dsk);

for j := I to 71 do begin
if
I then dot := ' - '
else dot

;=-=

'

';

if
= z Pt the n dot : = ' I ' ;
if J = ivaI then dot := '*';
write(dot);
if nprnt = I then write(dsk,dot);
end;

writeln; if nprnt

1 then writeln(dsk)

end;

writeln; if nprnt =
then writeln(dsk);
if nprnt = 1 then begin
for j := I to 71 do write(dsk ,dot) ;
end; {empty print buffer}
end ;{plot}
Program to test plot.pas:
program testO(output);
type
vector = array [1 80] of real;
var
blnk: char;
dsk: text;
y: vector;

x,f,an: real;
n, nx J nprnt. se: integer;

xmx :- 0.; xmn :- 0.;

for i

:~

I to nx do begin

if xli] > XIDX then XIDX := xli];

if xli] < xmn then xmn

xli];

end;

if xmn - O. then begin

{zero axis at left}
zpt :- I; scale :-70./xmx; offset := 1;
end i

if XIDX - O. then xmx :- 1.;

if (XIDX > 0.0) and (xmn < 0.0) then begin
{zero axis at middle}
if -xmn > xmx then xmx := -xmn;
zpt :- 36; scale .~ 70./(2.*xIDX);
offset := 36;
end;
write('maximum value=',xlDX :6:2); writeln;
if nprnt-l then begin
write(dsk,'maximum value=',xmx :6:2);

writeln(dsk);
end;
for i :~ I to nx do begin
dot :='-';
write( i-I :3 ,dot) ;
if nprnt = I then
write(dsk, i-I :3,dot);
rndoff := 0.5001;

procedure fopen(var dsk:text);

var
filnam:string[14];
begin
write('enter filename '); read(filnam);
writeln;
assign(dsk,filnam); rewrite(dsk);
end; {fopen}
{$i plot
begin
write('f = (0<f<0.5) '); read();
write(' sine or cosine? (l/O) ');
read(se); writeln;
f := 2.*pi*f;
for n := I to 45 do begin
an := n-1.

if se = 0 then yIn] := eos(f*an);

if se = I then yIn] := sin{f*an);
end;

write('print? (I/O)'); read(nprnt);

writelnj

if nprnt = I then fopen(dsk);

plot(y, 45, nprnt, dsk); blnk :=
if nprnt=1 then

460

INTRODUcnON TO DIGITAL SIGNAL PROCESSING

for n:=1 to 160 do write(dsk.blnk);

end.

xiii] := xili]*inv;
end i

end;

end {fft};
procedure fft(var xr,xi:vector;

10g2n.ntype:integer);
const

pi = 3.1415926;
var

procedure plotmr(x:vector; nprnt:integer

var dsk:text);
const

n,n2,i,id,j,k,le,lel: integer;
ur,ui,wr,wi,tr,ti,sign,inv,omega: real;

begin
sign := -1.;
if ntype < 0 then sign :=1.;
n :- 1; for i := 1 to 10g2n do n := n*2;
n2 := n div 2;
j := 1; {time decimation}
for i := 1 to n-l do begin
if i < j then begin {switch}
tr
:= xrlj]; ti
:= xilj];
xrlj] :- xrli]; xilj] :- xiIi];
xrli] := tr;
xili]:= ti;
k

while k < j begin

j :- j-k; k := k div 2;
end i

:- j+k;
end;

for k :- 1 to 10g2n do begin

{butterfly computations}
Ie : - 1; for i : - 1 to k do Ie : - le*2;
leI :- Ie div 2;
ur :- 1.; ui :- 0 .

omega := pi/leI;
wr :- eos(omega);
wi :- sign * sin(omega);
for
:- 1 to leI do begin
i :- j; while i <- n begin
id .- i + leI;
tr .- xrlid] * ur - xilid] * ui;
ti . xrlid] * ui + xi!id] * ur;
xrlid] := xrli] - tr;
xilid] :- xiIi] - ti;
xr!i] := xrli] + tr;
xiIi] :- xi!i]-+ ti;
i :E i+le;
end;

tr .= ur*wr - ui*wi;
ti :- ur*wi + ui*wr;
:~

var
dot, mark: chari
i, j. zpt, ival, m, offset: integer;
xmx, zero, scale, scale2, inc: real;

begin
writeln('magnitude response'); writeln;
if nprnt = 1 then begin
writeln(dsk.'magnitude response');
writeln(dsk);
end;
xmx := 0.;
for i := 1

to num do begin
if xli] > xmx then xmx :- x!i];

end;
:a n2;

ur

b Ink - ' ';

65;

num ,.

trj ui := tis

end i

end;

if ntype < 0 then begin

inv :- 1./n;
for i :- 1 to n do begin
xrli] :- xrli]*inv;

end

scale := xmx/20.; scale2 :- seale/2.;

for i := 1 to 21 do begin
inc :- (21-i)*scale;
if (i - 1) or (i - 21)
then begin
mark :- '-';

if i 1 then begin
write(blnk.xmx :6:2);
if nprnt - 1 then
write(dsk.blnk,xmx :6:2);
end;

if i - 21 then begin
zero :- 0.;
write(blnk,zero :6:0);
if nprnt 1 then
write(dsk,blnk,zero :6:0);
end;

end
else begin
mark := blnk;
for j := 1 to 7 do write(blnk);
if nprnt = 1 then
for j := 1 to 7 do
write(dsk,blnk);
end i

for j :. 1 to num do begin

if (j - 1) or (j - num) then begin
dot := '1';
if (i 1) or (i - 21)
then dot :c '.';
end;

if abs(xlj]-inc) < scale2

APPENDIX A

then dot := '*';

write{dot);
if nprnt - 1 then write{dsk,dot);

if nprnt = I then write{dsk,dot);

dot := mark;
end;

writeln;
if nprnt = 1 then writeln{dsk);

dot :- mark;
end

write 1n;

if nprnt

=1

461

end;

then writeln{dsk);

end; {plotpr}

end;

end; {plotmr}
procedure plotpr{x:vector; nprnt:integer;
var dsk:text);
const
b Ink = '

';

num = 65;
var
dot, mark: char;
i, j, zpt, ivai, m, offset: integer;
xmx, pi, val, scale, scale2, inc: real;

begin
writeln{'phase response'); writeln;
if nprnt = 1 then begin
writeln{dsk,'phase response');
writeln{dsk);
end;
pi := 4.*arctan{I);
xmx :a pi;

scale .= xmx/lO.; scale2 := scale/2.;

for i := 1 to 21 do begin
inc := {21-i)*scale - pi;
if (i = I) or (i = 21) or (i=II)
then begin
mark :Z '-';
val :E xrnx;
if i = 11 then val := 0.;

if i = 21 then val := -XIUX;

write{blnk,val :6:2);
if nprnt = 1 .
then write{dsk,blnk,val :6:2);
end
else begin
mark := blnk;
for j := 1 to 7 do write{blnk);
if nprnt = 1 then
for j := 1 to 7 do
wnte{dsk,blnk) ;
end;

for J := 1 to'num do begin

if (j = 1) or (j = num) then begln
dot := '1';
if (i = 1) or (i
21)
then dot := '+';
end;

if abs(x[J]-inc) < scale2

then dot := '*';

write(dot) ;

procedure plotlm{x:vector;
dbscal, spec :real;
ifI,if2,nprnt:integer;
var dsk: text) ;
var
blnk,mark: char;
arr: array [1 65] of char;
i, j, k, zpt, ivai, m, offset: integer;
xmx, xmn, scale, scale2, dbinc: real;

begin {plot 1m}

writeln{'log-magnitude response');
writeln;
if nprnt = 1 then begin
writeln{dsk,'log-magnitude response');
writeln{dsk);
end;
b Ink
m :=

:= ' ,
65;
{find max of x array}
XIUX := -300;
for i : = 1 to m do
if xli] > xmx then XIUX := xli];
{normalize to 0 dB}
for i := 1 to m do begin
Xll] := xli] - XIUX;
if xli] < dbscal then xli] :- dbscal;
end;

scale := dbscal/20.;
scale2 := abs{scale)/2.;
for i := 1 to 21 do begin
mark :- ... ';
if (i = 1) or (i

= 21) then mark :=

for j := 1 to 65 do arr!j] :- mark;
arrlil := '1';
if if! > 0 then arr!ifl] :- '1';
if if2 > 0 then arr!if2] := '1';
arrl65] := '1';
dbinc := (i-I}*scale;
for j := 1 to 65 do begin
if abs{spec-dbinc) < scale2
then arr[j] := ' - ' ;
if abs{x[jl-dbinc) < scale2
then arr[j] :- '*';
end;

J := 5 * {(i-I) div 5);

if J = i-I
then begin
write(blnk,dbinc :6:I};

462

INTRODUCTION TD DIGITAL SIGNAL PROCESSING

if arr[1] = '1' then arr[1] := '+';

if nprnt = 1
then write(dsk,b1nk,dbinc :6:1);
end
else begin
for k:=
to 7 do write(blnk);
if nprnt
1
then for k := 1 to 1
do write(dsk,blnk);
end;

for j:=
to 65 do write(arrlj]);
if nprnt
1 then begin
for j := 1 to 65
do write(dsk,arr[j]);
writeln(dsk) ;
end;

writeln;
end;

end; {plot 1m}

end;

if x[i] < 0 then begin

if y[i] < O. then b := b-pi;
if y[i] >= O. then b := b+pi;
end'
x[i]
8; y[i] := b;

;=

end;

plotmr(x,nprnt,dsk);
plotpr(y,nprnt,dsk);
i1 := 20; if2 := 30;
dbscal := -40.; spec := -20.;
for i := 1 to 65 do begin
if x[i] <= 0.0001 then
xli] :- 0.0001;
x[i] := 8.6S*ln(x[i]);
end;

plotlm(x,dbscal,spec,if1,if2,nprnt,dsk);
blnk := ' ' ;
if nprnt = 1 then for i := 1 to 160 do
write(dsk,blnk); {clear print buffer}
end.

Program to test fft.pas, plotmr.pas,

plotpr.pas and plotlm.pas:
program testl (input, output);
type
vector = array [1 128] of real;
var
dsk: text;
blnk: char;
x,y: vector;
i,nprnt,n,ifl,if2: integer;
a,b,pi,dbscal,spec: real;

{$i fopen }
{$i Ht }
{$i plotmr
{$i plotpr
{$i plotlm
begin
write('print (I/O)? '); read(nprnt);
writeln;

if nprnt = 1 then fopen(dsk);

for i :~ 1 to 128 do begin
x[iI := 0.; y[i] := 0.;
end;

for i := 1 to 8 do x[i] := 1.;

ft(x,y,7 ,1);
a := 1.; pi := 4.*arctan(a);
for i := 1 to 65 do begin
a := sqrt(sqr(x[i]) + sqr(y[i];
b := 0.;
i f x[i] <> O.

then b :- arctan(y[i]/x[i]);
if xli] = O. then begin
if y[i] > O. then b :- pi/2.;
if y[i] < O. then b := -pi/2;

C-language programse

void plot(x,nx,nprnt,fp)
FILE
*fp;
float
xl];
nx,nprnt;
int
{
int

i,j,ival,zpt,offset;
xmx,xmo,scale,rndoff;

float
char

dot;

0.; xmn

xmx

0 .

for (i = 0; i < nx; i++) {

if (x[iI > xmx) xmx
xli];
if [xli] < xmn) xmn = xli];
}

if (xmx < -xmn) xmx = -xmn;

if (xmx
0.) xmx = 1.;
printf("max val- %6.2f \n",xmx);
if (nprnt == 1)
fprint( fp, "max val= %6.2f \n" ,xmx);
if (xmn == 0.) {
zpt - 0; scale = 70./xmx;
offset - 0;

-=

}
if (xmx

> O. && xmn < 0.) {

if (-xmn > xmx) xmx - -xmn;
zpt - 35; scale - 35./xmx;
offset - 35;
}

for (i - 0; i < nx; i++) {

printf("%3d-",i);

APPENDIX A

if (nprnt -- 1) fprintf(fp,"%3d-",i);
for (j - 0; j < 71; j++) {
if (i == 0) dot =
,
else dot = ' "
if (j -= zpt) dot - '1';
rndoff = .5001;
if (xli] < 0.) rndoff = -.5001;
ivaI = offset + (rndoff + xli]*scale);
if (ivaI == j) dot = '*',
printf("%lc",dot) ,
if (nprnt
1) fprintf(fp,"Zlc",dot);

-=

printf("\n") ;
if (nprnt
1) fprintf(fp,"\n");

=-

return;
}

Program to test plot.c,

463

void fft(log2n,xr,xi,ntype)
float

*xr, *xi;

int

log2n, ntype,
j ,k, 1 J Ie J lel , id , i~

int

n J n1 J n2 J i

double

ur,ui,wr,wi,tr,ti,
pi-4. *atan(l.) ;
fm,fl,sign,in;

float

sign = -1.; if (ntype < 0) sign =1;

fm = 10g2n; n = pow(2.,fm);
n2 = n/2; n1 - n-1, j
I,
for (i-I; i<=n1, iH) {
ip - i-I,
if (i<j) {
id - j-1;
tr = *(xr+id); ti = *(xi+id),
*(xr+id)
*(xr+ip),
*(xi+id) - *(xi+ip);
*(xr+ip)
tr; *(xi+ip) = ti,
}

/* testO.c */
#include <math.h>
'include <stdio.h>
#inc1ude <p10t.c>
main()

k=n2;
while (k<j) {
j-k; k - k/2;
j+k;

float

*fp;
x1128],f,
pi4.*atan(1.);

int

ans,i,j,sc,nx,nprnt;

void

plotO,

FILE

fp - fopen("testO.dat","w");

start' printf("sin or cos? (1/0) H);

scanf("%d",&sc);
printf("f- (0<f<0.5) H);
scanf( "%f" ,&f);
printf("print results? (1/0) H);
scan( "%d" ,&nprnt);
if (nprtn =- 1) {
if (s~ =- 1) fprintf(fp,"sine fcn ");
if (sc -- 0) fprintf(fp,"cos fcn H);
fprintf( fp, "with freq %6.3f \n", f),
}

f - 2.*pi*f; nx - 45;
for (i - 0; i < nx; i++)
if (sc
1) xli] - sin(f*i);
if (sc
0) xli] - cos(f*i);
}

plot(x,nx,nprnt,fp);
printf("go again? (1/0)");
scan("%d",&ans) ;
if (ans -- 1) goto start;
fc lose( fp);

for (k=l; k<=log2n, k++) {

fl - k; Ie - pow(2.,fl);
leI = le/2; ur 1.; ui = 0.;
wr - cos(pi/lel);
wi = sign*sin(pi/le1);
for (j-1; j<-le1; j++)
for (i-j; i<-n; i-i+le)
ip = i-I, id - ip+le1;
tr - *(xr+id)*ur - *(xi+id)*uij
ti = *(xr+id)*ui + *(xi+id)*ur;
*(xr+id) - *(xr+ip) - tr;
*(xi+id) = *(xi+ip) - til
*(xr+ip)
*(xr+ip) + tr;
*(xi+ip) - *(xi+ip) + til
}

tr
ur*wr
ui*wi;
ti = ur*wi + ui*wr;
ur

tr; ui = tit

if (ntype < 0) {
in - l./n;
for (i-O; i<n; i++)
*(xr+i)
*(xr+i) * in,
*(xi+n) = *(xi+n) * in;
}

return;
}

464

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

void plotmr(x.nprnt.fp)
*fp;
float
x[ J;

FILE

int

nprnt;

int
float
char

i,j,nx,ival,zpt,offsetj
xmx,xmn,scale,rndoff;
dot;

& 65; xmx mO.; xmn = 0.;

for (i = 0; i < nx; i ++) {
if(x[i]>xmx)xmx
x[i];
if [x[i] < xmn) xmn = x[i];

nx

if (xmx < -xmn) xmx = -xmn;

if (xmx == 0.) xmx = 1.;
printf("magnitude response\n");
if (nprnt == 1)
fprintf( fp. "magnitude response\n");
printf("max val= %6.2f \n".xmx);
if (nprnt == 1)
fprintf(fp."max val= %6.2f \n".xmx);
if (xmn == 0.) {
zpt = 20; scale = 20./xmx; offset = 0;

nx

= 65;

xmx

= pi;

print( "phase response \n") ;

if (uprnt == 1)
fprintf(fp."phase response\n");
zpt = 10; scale = IO./xmx; offset
10;
for (i = 0; i <- 20; i ++) {
for (J = 0; j < nx; j++) {
if (j == 0 " J == nx-I)
dot - ' I ' ; else dot = ' ,
if (i == 0 " i == 20 " i == 10)
dot

rndoff = .5001;
if (xlJ] < 0.) rndoff
-.5001;
ivaI = offset + (rndoff + x[j l*scale'
if (ivaI == 20-i) dot =
printf( "%lc" .dot);
if (nprnt == 1) fprintf(fp."Zlc".dot
}

printf( "\n") ;
if (nprnt == I) fprintfCfp,"\n");
}
re turn;
}

if

(XDlX > o. && xmn < 0.) {

if (-xmn > xmx) xmx = -xmn;
zpt = 10; scale = 10./xmx; offset = 10;

for (i = 0; i <= 20; i ++) {

for (j = 0; j < nx; j ++) {
if (j == 0 " j == nx-l)
dot ",-"'; else dot = ' ';
if (i =- 0 " i == 20) dot =
rndoff = .5001;
ivaI = offset + (rndoff + x[j]*scale);
if (ivaI == 20-i) dot = '*';
printf( "%lc" .dot};
if (nprnt == 1) fprintf(fp."l:lc",dot);
}

void plotlm
(x,dbscal.spec.ifl.if2.nprnt.fp)
FILE
*fp;
float
x[].dbscal.spec;
int
ifl.if2.nprnt;
{
int

float
char

i,j,nx,zpt,offset;
xt[65].xmx.scale,scale2.
rndoff .dbinc;
dot;

nx = 65; xmx = -1000.;

for (i = 0; i < nx; i++){
if (x[i] > xmx) xmx = x[il;
}

printf("\n") ;
if (nprnt == 1) fprintf(fp."\n");
}
return;
}

for (i = 0; i < nx; i++) {

xt[i] = xli] - xmx;
if (xt[i] < dbscal) xt[i] = dbscal;
}

int

nprot;

printf("log-magnitude response
");
print( "dbscal- %6.2f
",dbscal);
printf("spec= %6.2f\n".spec);
if (nprnt == l){
fprintH fp, "log-magnitude response ");
fprintH fp. "dbscal- %6.2 f
" .dbscal) ;
fprintf(fp,"spec= %6.2f\n",spec);

int

i,j,nx,ival,zpt,offset;
xmx,scale,rndoff;
pi = 4.*atan(I.);

scale = dbscal/20.;
scale2 = fabs(scale/2.);
for (i = 0; i <- 20; i++)

void plotpr(x.nprnt.fp)
*fp;
float
x[ J;

FILE

float
double
char

dot;

dbinc - i*scale i

for (j = 0; j < nx; j++) {

APPENDIX A

dot - " '.

if (j 0

II j nx-1
II j if 1 I I j if2)

dot ' I';

if (i 0 I I i 20) dot '-';
if (fabs(apec-dbine) < Beale2)
dot '-"i

if (fabs(xtlj l-dbinc:) < acale2)

dot '*'i

printf("X1c",dot);
if (nprnt 1)
fprintf(fp,"%lc",dot);
print( "\n");
if (nprnt 1) fprintf(fp,"\n");
)

return;
} ,

Program to test fft.c, plotmr.c,

p10tpr.c and plotlm.c:
/* test1.c */
'include <math.h>
linclude <stdio.h>
linc lude <fft.c >
linc lude <plotmr.c >
linc lude <p lotpr.c >
linclude <plotlm.c>
main()
(

*fp;
float x[128],y[128],dbscal,spee;
double a,b,f,
pi -4.*atan(I.);
int
i,j,if1,if2,nprnt,nx,ntype,log2n;
void
fft(),plotmr() ,plotpr() ,plotlm();
FILE

fp fopen("testl.dat","w");
nx 128; nprnt 1;
for (i 0; i < nx; i++) {
xli] -0.; y[i] -0.;
}

for (i 0; i < 8; i++)

xli] 1.;
}

10g2n 7; ntype 1;
fft(log2n,x,y,ntype);
nx 65;
for (i 0; i < nx; i++)
a x[i]; b y[i];
xli] sqrt(a*a + b*b);
if (a

o.

&& b 0.)

y[il

else y[i] atan2(b,a);

0.;

465

plotmr(x,nprnt,fp),
plotpr(y,nprnt,fp),
for (i 0, i < nx, i++) {
if (x[i] < 0.0001) xli] 0.0001;
xli] 20.*log10(x[i])1
}

dbscal -40., spec -20.,

if 1 19, if2 29;
plotlm(x,dbscal,apec,ifl,if2,nprnt,fp),
fc lose( fp) ,

466

JNI'llODucnON TO DIGrrAL SIGNAL PROCESSING

Re.ult. produced

by

te.tO:

aax1_ value 1.00

.--------

0--

123-

4-

56-

7B91011121314151617181920212223-
24- 25-26272B29303132333435363738394041-424344-

--

-- -

- r
I
I
I
I
I
I
I
I

L9t

+---------------------------------------------------------------+.t,1

,1I

..

..

. . . ,

..,

1
,

..,

.-------- -----.---------- ---.-------------.--------------.000

I .

,
1
1
1

1
1

1
1

+---------------------------------------------------------------+u
8sruodsr8.1 8sr'8qd

--------------.---------------.------------.------------+0
I

I

I
1
1
1

00 8
+------------------------------------------------..

V XICINlIdcIV

468

INTRODUCTION TO DIGITAL SIGNAL PROCESSING

log-magnitude response

0.0------------1---------1--------------------------------I
1
1
1
I
I

I
1
I

I
I
I
I
I

I
I
I
-10.0+

I
I
I
1
1
1
1
1
1
11
1I

I..

11...

-20.0-------------------------------------------------I

I
1I
-30.0+
1
1

1
II
1
I
1

I
I

1
1

I
II
I
I
1
1
I

I
II
I
1
I

I
1

-40.0+-----------------1---------1------------------------------.

APPENDIX B

FILTER SPECIFICATION FOR PROJECTS TABLE B-1

Frequency values for the filter specification shown in Fig. B.l. The frequency
values below are given in terms of k, the 128-point DFf bin number. The
value k = 1 corresponds to w = O. The frequency is then Wk = 2n(k - 1)/128.
The lowpass and highpass filters are specified with the pair of values k} and k2
that define Wk, and Wk2 The bandpass and bandreject filters are specified with
one set of values for kl1 k2' k3 and k 4
Lowpass and highpass freqaeacy maes

kl
4
5
6
7
8
9
10
11

12
13
14

k2
21

kl
33

k2

22
23

34

49

35

50

24

36

25

37
38
39

51
52
53

26
27
28
29
30

31

40

41
42
43

48

54

55
56
57
58

kl
6
7
8
9
10

k2
19
20
21

kl
31
32
33

Bandpass and baadreject frequency values

k2
46

47
48

22

34

49

23

35

50

kl
4
5
6
7
8
9
10

11

24

36

12
13
14
15
16

25

37

26
27
28
29

38

51
52
53

39

-54:

12

40

55
56

13

41

11

14

k2
21
22

k3
33
34

k.
48
49

23

35

50

24

36

25

37
38
39

51
52
53

26
27
28
29
30
31

11

6
7
8
9
10

k2
19
20
21

k3
31
32
33

22
23

34

49

35

50

51
52
53
54
55
56

11

24

36

54

12

25

40

55

13

41
42
43

56

14
15
16

26
27
28
29

37
38
39

57
58

40

41

469

k.
46

47
48

470

APPENDIX B

LOWPASS SPECIFICATION
IH(e'''')ldB

-~.~~

HIGHPASS SPECIFICATION
IH(e'''')ldB

o
-1

-50L_ _L-_...mi!~~ ro

o
BANDPASS SPECIFICATION
IH(e'''')ldB

o
-1

BAND REJECT SPECIFICATION

IH(e'''')ldB

O~!m

-If""

nGURE Bl
Lowpass, highpass, bandpass and bandreject specifications for the projects.

SUBJECT

INDEX

A
ADC (see analog-to-digital conversIOn)
Adder component, 3, 41,169
Advance component, 3, 41, 169
Aliasing
frequency domain, 103
time domain, 125, 127,341,433,438
Allpass filter, 310
Amplitude function, 75
Analog filter design methods, 263
Butterworth filter, 264
Chebyshev filter, 270
elliptic filter, 275
Analog-to-digital conversion (ADC), 12
Nyquist criterion, 102
quantization effects, 395
sampling of anlog signals, 103,280
Anti-aliasing filter, 105,264
Antisymmetric sequences, 86, 89
(also see Symmetric sequences)
B
Bartlett window, 328
Bilinear transform, 288
Binary point, 391
Blackman window, 335
Butterfty computation, 147
Butterworth filter, 264

C
Cascade second-order section, 215
Cascade structure of filters, 164, 215

Causal system, 40
Chebyshev filter, 270
Circular convolution, 95,129,130,321
GOefticient quantization, 399
FIR filter, 400
IIR filter, 402
Comb filter, 224, 364
Conversion of decimal to binary, 392
Convolution
circular, 95,129,130,321
linear, 31, 132
sum evaluation, 32
CSOS (see cascade second-order section)
Cutoff frequency, 236, 264
D
DAC (see digital-ta-analog conversion)
Decirnation-i'l-time FFT algorithm, 142
Delay component, 3, 41, 169
DFf (see Discrete Fourier transform)
DFf program, 135
DitJerence equation, 41,162
Digital filter (see FIR or I1R filter)
Digital-to-analog conversion, 12, 106
Direct form structures, 210
Discrete Fourier transform (DFf)
computation of, 123
definition, 119, 123
evaluating with fast Fourier transform, 142
Goertzel algorithm, 139
linear convolution with, 132
of infinite duration sequences, 127
of noncausal sequences, 125
program (DFf), 135

471

472

SUBJECf INDEX

Discrete Fourier transform (DFf) (contd.)

properties, 129
reducing computation time, 136
relationship to Fourier transform, 119
Discrete Hilbert transform
application to filter design, 436
definition, 433
program (DHT), 452
Discrete-time sequences
complex exponential, 23
unit-sample response, 27
unit-sample, 21
sinusoidal, 22
Discrete-time systems
descriptions
time-domain:
convolution, 31
difference equation, 41,162
unit-sample response, 21
frequency-domain:
transfer function, 67, 83
z-domain:
system function, 162
linear time-invariant (LTI) system
causality, 40
stability, 38
Dominant singularity, 187

E
Elliptic filter, 275
Energy evolution factor, 427
Euler's identities, 25
Even sequence, 86, 188, 189
F

Fast Fourier transform, 142

butterfly pattern, 147
decimation-in-time algorithm, 142
overlap-and-add method, 379
FFf (see fast Fourier transform)
Filter implementation
from difference equation, 41
from system function, 169
Filter structures
cascade, 164
cascade second-order section, 215
direct forms, 210
frequency-sampling, 221
parallel, 167
parallel second-order section, 218
Finite geometric sum formula, 36

Finite-impulse response (FIR) filter, 209

approximation to inverse filter, 432
design techniques, 318
DFf method, 341
DFf/windowing method, 348, 354
frequency-sampling method, 361, 372
windowing method, 319
finite-precision effects, 400, 409
frequency-sampling structure, 221
implementation with FFf, 379
linear-phase structures, 220
Finite-precision arithmetic effects, 391
FIR (see Finite impulse response filter)
Fixed-point
addition, 405
number representation, 391
multiplication, 406
Floating-point
addition, 407
number representation, 394
multiplication, 408
Fourier transform
definition, 67
linear-phase form, 75
of continuous-time signals, 99
of real-valued sequences, 72
of symmetric sequences, 86
of delayed sequences, 85
of product of two sequences, 94
program (Ff), 81
properties, 69
relationship to z-transform, 184, 232
Frequency-sampling
filter structure, 221
FIR filter design method, 361
Frequency transformation for lowpass filters,
295
Ff program, 81
G
Gain constant, 241
Geometric evaluation of z-transform, 183
Gibbs phenomenon, 326
Goertzel algorithm, 139
H

Hamming window, 332

Hanning window, 330
Hibert transform (see Discrete Hilbert
tr~sform)

SUBJECt INDEX

IDIT (see inverse discrete Fourier transform)

IIR (see Infinite-impulse response filter)
Impulse-invariance method, 280
Infinite-impulse response filter, 209
design techniques:
impulse-invariance method, 280
bilinear-transform method, 288
transformations of lowpass filter, 295
finite-precision effects, 402, 410
limit cycle, 410
Infinite geometric sum formula, 36
Interactive filter design, 247, 256
Interpolation, 108, 264
Inverse filter, 422
Inverse Fourier transform, 92
Inverse discrete Fourier transform, 123
Inverse z-transform, 196

Multiplication
fixed-point, 406
floating-point, 408
Multiplier component, 3, 41,169
N

Noncausal filter
definition, 40, 41
filter structure, 169
implementation of, 54
Nonrecursive filter, 5, 210
program (NONREC), 48
NONREC program, 48
NULL program, 45
Number representations
Fixed-point, 391
Floating-point, 394
Nyquist criterion, 102

Limit cycle, 410

Linear difference equations, 41
Linear conv91ution, 32
Linear-phase filter
energy evolution of, 427
FIR filter structure, 220unit-sample response Qf, 220
z plane pole/zero pattern of, 195
Linear-phase form of transfer function, 75
Linear-phase response, 243, 349, 426
Linear-phase system, 426
Linear system, 26
causality, 40
stability, 38
Log-magnitude response
definition, 238
program (LOGMAG), 240
LOGMAG program, 240
Lowpass frequency transformations, 295
to bandpass, 301
to bandreject, 306
to highpass, 298
to lowpass, 296

Order, 5
Overflow, 405
Overlap-and-add method, 379

M
Magnitude response, 71, 236
Magnitude-phase program (MAGPHS), 81
MAGPHS program, 81
Maximum-phase system, 426
Minimum-phase system, 426, 433

473

P
Padding with zeros, 123
Parallel second-order section, 218
Parallel structure, 167, 218
Periodic extension
of spectra, 102
of time sequences, 120, 125, 127, 341, 432
Phase distortion, 244, 440
Phase response, 71, 243
jumps in, 76
of minimum-phase system, 426
of maximum-phase system, 426
principal value, 73
PL01LM program, 240
POLE program, 181
Pole/zero pattern in z plane, 170
of linear-phase systems, 195
of maximum-phase sequences, 426
of minimum-phase sequences, 426
of symmetric sequences, 189
to z-transform, 235
Prewarping, 290
Program descriptions:
DIT,135
IT,81
LOGMAG,240

474

SUBJECT INDEX

Program descriptions: (contd.)

MAGPHS,81
NONREC,48
NULL, 45
PLOTLM,240
POLE, 181
QUANT, 412
REC,49
SCALE, 413
SHIFf, 54
Programming style, 45
PSOS (see parallel second-order section)

Q
Quantization
inADC,395
of filter coefficients, 399
simulation of, 412
Quantizer characteristic, 395
QUANT program, 412

T
Time-invariant system, 31
Transfer function, 67, 83
Transition region, 368
Triangular window, 328
Truncation of arithmetic results, 392, 406, 409
Truncation of sequences, 321
Two's-complement notation, 393

Unit circle, 189

Unit-sample response, 27
Unit-sample sequence, 21
Unit-step sequence, 22

v
Vector notation of complex numbers, 23

R
Raised-cosine windows, 330
Random number generator, 414
Rectangular window, 322
Reconstruction filter, 105
REC program, 49
Recursive filter, 5, 209
program (REC), 49
Relevant singularity, 246
Region of convergence, 174
Resonator, 224
Round-off, 392, 406

S
Sampling analog signals, 2, 96
Second-order sections, 215, 218, 400, 402
SCALE program, 413
SHFfRG program, 47
SHIFf progrl'm, 54
Shift register program (SHFfRG), 47
Spectrum, 67, 71
Stability criterion, 38
Sum-of-products program (SUMP)
SUMP program, 47
Superposition principle, 26, 381, 398
Symmetric sequences, 86
Fourier transform of, 86
pole/zero pattern of, 189
z-transform of, 188
System function, 162

Window sequences, 321

Bartlett, 328
Blackman, 335
comparison, 340
FIR filter design with, 319
rectangular, 322
Hamming, 332
Hanning, 330
main-lobe width, 325, 340
raised-cosine, 330
side-lobe magnitude, 325, 340
triangular, 328

ZERO program, 181

Zplane, 169
Z-transform
definition, 158
filter implementation from, 167
from difference equation, 162
inverse, 196
of symmetric sequences, 188
pole/zero pattern, 170,235
properties, 160
region of convergence, 174
relation to Fourier transform, 184