DIGITAL FILTERS: ANALYSIS AND DESIGN me tois AND DESIGN” Andreas Antoniou Professor and Chairman, Department of Electrical Engineering Concordia University McGraw-Hill Book Company New York St Louis Sax Francisco Auckland Bogots Dieter! Johannesburg London Madrid Mevico Montreal New Delht Panama Paris Sio Paulo Singapore Swiney Tokyo TorontoDIGITAL FILTERS: ANALYSIS AND DESIGN Copyright © 1979 by McGraw-Hill, ne. Al ihts reserved Pinte in the United Sites of America. No pat of this publication Imay be reproduce, stored in a retieval system, or tansmited in any form o by any means, electrons, mechanical, photocopying recording, oF ‘otherwise, without the prior writen permission ofthe publeher 78910 HDHD 8987654 ees AN GS-UL g4 Library of Congress Cataloging in Publication Data Antonio, Andreas, date Digital ters (Mecraw-til electrical engineering series: Communications and information theory section) Includes bibliographical references and inden |, Digital ters (Mathematics) L Tike I. Series. TKIBTFSAG “e138 78.13998 ISBN Oo7-0m2117-1 This Book was sen Monophoto, The editor was Frank J. Ceera {te Production supervisor was Donna Pir: the cover was designed by Rafael Herandes PIBLIOTEOA ate tetera To MY MOTHER AND FATHER| | i 12 13 14 1s 16 17 1B 19 24 22 23 24 25 26 Preface Introduction Elementary Analysis Introduction ‘Types of Discrete-Time Signals The Digital Filter as a System Characterization of Digital Filters Digital-Fiter Networks Introduction to Time-Domain Analysis Convolution Summation Stability State-Space Analysis References Problems The z Transform Introduction Definition ‘Theorems One-Sided 2 Transform Inverse 2 Transform Complex Convolution References, Problems CONTENTSa 32 33 aa 35 aa 42 43 4a 45 46 47 5 Su 52 33 3a 55 56 37 6a 62 63 6a 65 66 67 ray 2 B The Application of the z Transform Introduction ‘The Discrete-Time Transfer Function Stability ‘Time-Domain Analysis Frequency-Domain Analysis References Problems Realization Diect Realization Direct Canonic Realization Cascade Realization Parallel Realization Ladder Realization ‘eduction ‘Topological Properties References Additional References Problems Analog-Filter Approximations Introduction Basie Concepts Butterworth Approximation Tschebyschef Approximation Elliptic Approximation Bessel Approximation “Transformations References Problems Continuous-Time, Sampled, and Discrete-Time Signals Introdvetion The Fourier Transform Generalized Functions Sempled Siguals The Sampling Theorer Interrelations ‘The Processing of Continuous-Time Signals References Probiems Approximations for Recursive Filters Introduetion Realizability Constraints Invariant-Impulse-Response Method st st SL 34 58 9 65 65 0 70 1 16 %6 n * a 92 2 100 tor os 107 mn 29 0, 1a 13 a hr 17 1a 48 131 34 155 163 163 167 167 16k 168 7A 7 76 1” a 82 83 8a BS ou 92 93 94 95 36 10 10.1 102 103 104 105 106 107 108 109 il Md n2 HB ha ms convents xi Modified Invariant-Impulse-Response Method m1 Matched-z-Transformation Method 1s Bilineer-Transformation Method 178 Digita-Filter Transformations 186 References 90 ‘Additional References ist Problems 12 Recursive Filters Satisfying Prescribed Specifications 196 Introduction 196 Design Procedure 17 Design Formulas 198 Design Using the Formulas and Tables 207 Delay Equalization 2s Reference 21s, Problems 216 Design of Nonrecursive Filters 28 Introduction 218 Properties of Nonrecusive Fiers 218 Design Using the Fourier Series na Use of Window Functions 26 Design Based on Numerical-Analysis Formulas 2s Comparison Between Recursive and Nonrecursive Designs 298 References 209 ‘Additional References 29 Problems 230 Random Signals ass Introduction 288 Random Variables 286 Random Processes 258 First and Second-Order Stati 260 Moments and Autocoreelation 28 Stationary Processes 263 Frequeney-Domain Representation 264 Discrete-Time Random Processes 267 Filtering of Discrete-Time Random Signals 268 References 270 Problems 20 Effects of Finite Word Length in Digital Filters mm Introdetion 2 Number Representation 205 Coeticient Quantization 284 Produet Quantization 288 290 Signal Scaling| 116 2 Ra 22 123 124 125 126 127 128 R9 1210 Rut 13 Ba 132 83 ba 135 136 17 138 Be B10 1 2 14 144 142 M43 144 145 146 149 M8 ‘The Deadband Eifee: References ‘Additional References Problems Wave Digital Filters Introduction Sensitivity Considerations Wave Network Characterization Element Realizations Digital-Flter Realization Wave Digital Filters Satisfying Prescribed Specifications Reduction in the Number of Digital Elements Frequency-Domain Analysis Alternative Approach to the Synthesis of Wave Digital Filters ‘A Cascade Synthesis Based on the Wave Characterization Choice of Structure References Additional References Problems The Discrete Fourier Transform Introduction Definition Inverse DFT Properties Interrelation Between the DFT and the z Transform Interrelation Between the DPT and the CFT Interrelation Between the DFT and the Fouriee Series Nonzecursive Approximations Through the Use of the DFT Simplified Notation Periodic Convolutions Fast-Fourier-Transform Algorithms Digital-Filter Implemestation References Additional References Probiems Hardware Implementation Introduetion Boolean AJgebra Combinational Cireuits Flip-Flops, Registers, and Counters Sequential Circuits IC Families Digital-Filter Implemeniations Applications of Digital Filters, References Additional References Problems 296 308 308 305 309 309 SLL 3183 320 322 226 329 331 33 3a 33 348 34a 349 349 349 350 351 353 358 361 363 367 367 369 379 383 383 386 388 388 388 32 408 410 20 an 44 434 a5 436 AL Az Aa Aa AS AG AT AS BL B2 BB Ba BS Appendixes Elliptic Functions Introduction Elliptic Integral ofthe First Kind Elliptic Functions Imaginary Argument Formulas Periodicity Transformation Series Representation References ‘Computer Programs Introduction. System Configuration Program Descriptions Program Listings ‘Typical Runs Index 4a 4a 4a 4a 445 407 448 490 482 453 4s 454 484 454 460 490 suee PREFACE $$ COSEREFACE This book attempts to put together theories, techniques, and procedures which can be used to analyze, design, and implement digital filters. The digital filter is Viewed as a system which can be implemented by means of software or hardware In other words, the book is concerned beth with the construction of algorithms (or computer programs) that can be used to filter recorded signals and also with the design of dedicated digital hardware that can be used to perform real-time filtering tasks like the many required in communications systems ‘The prerequisite knowledge is a typcal undergraduate mathematics back- ‘ground of calculus, complex variables, and simple differential equations. Section 5.5 entails a basic understanding of elliptic functions. Since this topic is normally excluded from undergraduate curricula, Appendix A provides a brief but adequate {treatment of these functions, Chapter 1 requires a basic understanding of random variables and processes, which are reviewed in Chap. 10, Chapter 1 introduces the digital filter as a discrete-time system which can be linear or nonlinear, causal or noncausal, and so on. Time-domain analysis is then introduced at an elementary level. The analysis is accomplished by solving the difference equation of the filter by induction, Although this technique is unlikely to be employed in practice itis of pedagogical value as it provides the newcomer with a solid grasp of the physical nature of a digital filer. The chapter concludes with an alternative and more advanced time-domain analysis based on a state-space characterization, ‘The basic mathematical tool for the aralysis of digital filters, the z transform, is described in Chap. 2. Its application in the time-domain and frequency-domain analysis of linear, time-invariant filters is considered in Chap. 3 ‘The realization of digital filters, namely the process of translating the transfer fanction into a digital-flter network, is discussed in Chap. 4. Later in this chapter 8 very useful topological theorem, known as Tellegen’s theorem, is described, It is then used to develop the concepts of reciprocity, interreciprocity, and trans-| Position. The chapter concludes with a convenient sensitivity analysis based on Tellegen’s theorem. Transfer-funetion approsimations for digital filters of the recursive type are almost invariably obtained indirectly by using analog-ilter approximations (eg. Tschebyschef, elliptic, etc) Such approximations are considered in Chap. 5. The elliptic approximation is treated in detail because this is the most efficient of {he available approximatiors. Although the derivation of this approximation {urns out (0 be quite involved, a simple and easy-to-apply formulation is possible, as demonstrated at the end of Sec. 5.5. Up to Chapter 5, the digital filter is treated as a distinct entity with its own methods of analysis. In Chap 6, a theoretical link is established between diserete, {ime and continuous-time signals, which leads to a direct interrelation between the 2 (eansform and the Fourier transform. By this means, the enormous wealth of analog techniques can be applied in the analysis and design of digital filters; 'n addition, digital filters can be applied in the processing of analog signals Chapter 7 deals with the approximation problem for recursive filters. Methods are described by which a given continuous-time transfer function can be trans. formed into a corresponding discrete-time transfer function, eg, impulse invariant, response method and bilinear transformation method. A detailed procedure which can be used to design Butterworth, Tschebyscheff, and elliptic filers satisfying Prescribed specifications is found in Chap. 8. This can be readily programmed Nonrecutsive filters and taeir approximations are considered in Chap. 9. The Use of window functions is described in detail, Emphasis is placed on ‘Kaiser's Window function, which has proved very versatile. In Chap. 10 the concept of a random process is introduced as a means of representing random signals. Such signals arise in digital filters because of the inevitable signal quantization. The effects of finite word length in digital fiters along with appropriate, up-to-date methods of analysis are discussed in Chap. 11. The topics considered include Coefficient quantization, product quantization, sealing, and limit-eycle bounds, A relatively recent innova-ion in the domain of digital filtering has been the introduction of wave digital filters. Filters of this type can have certain attractive features, ¢4. low sensitivity, and are considered in Chap. 12. The chapter includes step-by-step procedures by which filters satisfying prescribed specifications can be designed. The chapter concludes with a list of guidelines that can be used in the choice of a digital-fiter network (or structure) Chapter 13 presents the discrete Fourier transform and the associated fast Fouriertransform method as mathematical tovls in the software implementation of digital filters. Much effort is spent in clearly conveying the exact interrelations between the discrete Fourier transform and (1) the z transform, (2) the continuous Fourier transform, and (3) the Fourier series because if these interrelations are not thoroughly understood, tke user of the fast Fourier-transform method is likely to finish with masses of meaningless numbers Chapter 14,which concludes the book, deals with the hardware implementation of digital filter. It starts with a brief review of boolean algebra and then proceeds {0 4 detailed description of the various types of combinational and sequential veuits that 1 components in the implementation, Later the main ‘stegrated-circuitiamilies are discussed and their features compared. Subsequently, in Sec. 14.7, three specific approaches to the implementation are described and their merits and demerits discussed. The chapter concludes with a brief outline Of past, present, and future trends in the domain of digital filtering, Most of the techniques are illustrated by examples and selected sets of Problems are included throughout the book. Also 17 useful computer programs are included in Appendix B, which can be used by the filter designer in real-life designs and by the student in developing projects and in solving some of the challenging problems included. ‘The book can serve as a text for a sequence of two one-semester courses on digital filters for senior undergraduate or first-year graduate students, Tt could also serve for a one-semester course for groups of students with an adequate background of discrete-time system theory, analog filters, the Fourier” \ransform, and the theory of random variables and processes. Such @ course could comprise the following: 1. Review of Chap. 6 2. Brief discussion of analog-flter approximations (Chap. 5) with the emphasis placed on application rather than derivation 3. Chapters 7 to 9 4. Chapter 11, possibly in conjunction with Sec, 4.7 5. A choice from Chaps. 12 to 14, depending on the course orientation desired The book should also be of interest to the filter designer, in particular Appendix B, and also to the scientist who has his own discrete-time signal to rocess, Pete book is Supported by @ comprehensive Solutions Manual and also by a Tape Cartridge, which contains the programs given in Appendix B. The latter will be made available to institutions adopting the book upon the payment of a nominal handling fe. I wish to thank V. Bhargava, A. G. Constantinides, R. Crochiere, L. B, Jackson, ©, Monkewich, A. Papoulis, W. Saraga, and L. Weinberg for reading parts of the manuscript and for suggesting numerous improvements; S. W. Director, R. F. J. Filipowsky, and S. K. Mitra for their detailed reviews and suggestions; E. I. Jury for suggesting a number of improvements for the second printing of the book; J.C. Callaghan and M. N. S. Swamy for encouraging and supporting this work; June Anderson for typing the manuscript; Concordia University for supporting this work: and the National Research Council of Canada for supporting the research that led to many of the new results presented Last but not least, T wish to thank my wife Rosemary and also Anthony, David, Constantine, and Helen for their sacrifices and understanding Andreas Antonionae INTRODUCTION Signals arise in almost every field of science and engineering, e.g. in acoustics, biomedical engineering, communications, control systems, radar, physics, seismol. ogy, and telemetry. Two general classes of signals can be identified, namely continuous-time and discrete-time signals. ‘A continuous-time signal is one that is defined at each and every instant of time. Typical examples are a voltage wave'orm and the velocity of a space vehicle as a function of time. A discrete-time signal, on the other hand, is one that is defined at discrete instants of time, perhaps every millisecond, second, or day. Examples of this type of signal are the closing price ofa particular commodity on the Stock Exchange and the daily precipitation as functions of time A discrete-time signal, like a continuous-time signal, can be represented by a unique function of frequeney referred to as the frequency spectrum of the signal. This is a description of the frequency content of the signal Filtering is a process by which the frequency spectrum of a signal can be modified, reshaped, or manipulated according to some desired specification. It may entail amplifying or attenuating a range of frequency components, rejecting vr isolating one specific frequency component, ete. The uses of filtering are mani- fold, cg, to eliminate signal contamination such as noise, to remove signal distor- tion brought about by an imperfect transmission channel or by inaccuracies in measurement, to separate two or more distinet signals which were purposely mixed in order to maximize channel utilization, to resolve signals into their frequency components, to demodulate signals, to convert discrete-time signals into continuous-time signals, and to bandlimit signals. The digital filter is a digital system that can be used to filter discrete-time signals. It can be implemented by means cf software (computer programs) or bySem, or aicated hardware, and in ether case it ean be used to filter real-time Signals or non-real-time (recorded) signals, Software digital fiters made their appearance along with the frst digital the midrates Bann ortes although the name digital fier did not emerge Cait the midsixties, Early in the history of the digital computer, many of the classical cary ad analysis formulas of Newton, Stirling, Everett, and others were teats Carty out interpolation, diferentiation, and integration of functions (signals) rep- none bY means of sequences of numbers (discrete-time signals) Since interpo- lation, differentiation, or integration ofa signal represents ¢ manipulation of the Frequency spectrum of the signal, the subroutines or programe constructed to ine ane Prelction, pattern cognition, electrocardiogram processing, and spec- becom is: In fact, as time goes on, interest in the software digital file i becoming progressively more intense while its applications are increasing at an exponential rate, A pandlmited continuous-time signal can be transformed into a dscretetime Beal gy means of sampling. Conversely, the iserte-ime signal so gencested ns be used to regenerate the orginal continuous-time signal by meses of interpola tian by virtue of Shannon's sampling theorem. As a consequence, hardware digi- datanets Gam be used to perform real-time filtering tasks which in the nok ey distant past were performed almost exclusively by analog filters The advantages senceatined are the waditional advantages associated with digital sysereein Reneral 1. Component tolerances are uncritical 2 Component deft and spurous environmental signals have no influence on the system performance, 3. Accuracy is high 4, Physical size is small 5. Reliability is high, ‘A very important addtioral advantage of digital ites isthe ease with which jut arameters can be changed in order to change the fier character aut allows one to design programmable iters which can be wsed topechor i tasks, Also one can design new types of filters such a adaniive files. The main disadvantage ofhardware digital ters at preent einen {atively high cost. However, with the tremendous advancements nthe demic Grastatn ne eBration. the cost of hardware digital fiers is likely to deep Faentally in the not too distant future. When this happens, hardware date, filters will replace analog filters in many more applications CHAPTER ONE _——CONE ELEMENTARY ANALYSIS 1.1 INTRODUCTION A digital filter, like an analog filter, can be represented by a network which comprises a collection of interconnected elements, Analysis ofa digital filter i the Process of determining the response of the filter network to a given excitation, Design of a digital filer, on the other hand, is the process of synthetizing and implementing a filter network so that a sel of prescribed excitations results in a set of desired responses. ‘This chapter is an introduction to the analysis of digital filters. First. the fundamental concepts of time invariance, causality, etc, a8 applied to digital filters are discussed, Then an elementary time-domain analysis is described. The chapter concludes with a more advanced time-domain analysis based on a state-space approach. 1.2 TYPES OF DISCRETE-TIME SIGNALS ‘A continuous-time signal can be represented by a function v(t) whose domain is a Fange of numbers (f,, 2), where — 00 <¢. and t, < oo, Similarly. a discrete-time Signal can be represented by a function x'n7), where T is a constant and nis an integer in the range (ny, 1) uch that ~ co k Qa) This criterion is illustrated in Fig. 4a ard b for a causal filter. i fe T 3 x(nT) and (6) to a delayed excitation x(a —kT) Ki ar Example 1.2 (a) A digital filter is represented by HOT) = Ax(nT) = Ix(nT — 27) + 3x(n7 + 27) Check the filter for causality. (b) Repeat part (a) if nT) = Rx(nT) = 3x(nT — T) — 3x(n 7 — 27) war-KT) SouuTioN (a) Assume that x,(n7) and x,(nT) satisfy Eq. (1.1), For n= k Bx, (07) = 3x (kT — 27) + 3,(kT + 27) Bx, (UT) = 3xy(kT — 27) + 3xy(kT + 27) HH but since 3x,(KT + 2T) # 3xg(kT +27) it follows that Ax,(nT) # Axz(nT) for ice, the filter is noncausal (b) For this case x (nT) = 3xy(nT — T)~ 3x,nT— 27) and AX,(0T) = 3x4(nT ~ 7) ~ 3x,(0T ~ 27) In sk, then n~1,n—2: Aax(nT) # a(n) | 5 and therefore the filter is nonlinear. i | (0) For this case ; 2 | Ala (nT) + Prxa(nT)] = (nT Plax, (nT + 27) + Bxy(nT + 27)) i é = a(nT}x,(nT + 27) + plnTPxy(nT + 27) ? = ax y(n) + BHX,(NT) i ie, the filter is linear. In this book we shall be concerned almost exclusively with linear, causal, and time-invariant filters.1.4 CHARACTERIZATION OF DIGITAL FILTERS Analog filters are characterized in terms of diferential equations. Digital filters, on the other hand, are characterized in terms of difference equations. Two types of sigital fiters can be identified, nonrecursive and recursive filters Nonrecursive Filters The response of a nonrecursve filter at instant nT is of the form WT)=SL..xOT— 1) x07), xo + 7), Hf we assume linearity and time invariance, y(n) can be expressed as wT) = ¥ ax(nt - iT) (12) where a, represent constants. Now by assuming causality and then using the causality criterion defined earlier, we can show that ay2a.e=0 and so WoT) = ¥o,x(ur = iT) 1h in addition, x(n7) =0 for n <0 and a, = 0 for i> N, woT)= Saxnt—it)+ ¥ ayxlnt -i7) Savor ins ¥ ax Sax ir) (13) ‘Therefore, a linear, time-invariant, causal, nonrecursive filter can be represented by an Nth-order linear difference equation, N is the order of the filter. Recursive Filters ‘The response of a recursive fier is 2 function of elements in the excitation as well as the response sequence. In the case of a linear, time-invariant, causal filter WOT) = Lax(n? —i7)— Yb,y(7 = 7) (14) i.e, iFinstant nT*is taken to bethe present, the present response isa function of the present and past N values of the excitation as well as the past N values of the response. Note that Eq, (14) simplifies to Eq, (13) if bj =0, and essentially the nonrecursive filter is a special case of the recursive one, 15 DIGITAL-FILTER NETWORKS ‘The basic digital-filter elements are the unit delay, the adder and the multiplier ‘Their characterizations and symbols are given in Table 1.1. Their implementation ‘can assume various forms, depending on the representation of the signals to be Processed, Ifthe signals are sequences of inary numbers, the unit delay will bein the form ofa shift register, whereas the adcer and multiplier will be combinational ‘or sequential networks comprising NAND of NOR gates. Digital-flter networks are collections of interconnected unit delays, adders, and multipliers. Their analysis is usually simple and can be carried out by using the element equations in Table 1. Example 1.4 (a) Analyze the network of Fig. 1Sa. (b) Repeat part (a) for the . network of Fig. 1.56. SoLutIon (a) The signals at nodes A and B are y(nT — T) and e'y(nT — T), respectively. Thus nT) = xinT) + y(n — 7) Table 1.1 Digital-filter elements ‘Symbol Equation - . xo) Mer)sie —— 0 vr) ae L 6] \ = je “OE Lg 1 | wer) « Figuce 18 Digital-fiter networks (Example 1.4): (a) first-order filter and (b) second-order fier (6) The analysis in this case is simplified by using the shift operator 6, which is defined by 8-'ulnT) = w(nT — rT) From Fig. 156 uy(1T) =m, (nT) + maug(nT) + mgus(nT) (XT) =F ug(nT) us) = 6 y(n) HOT) = mgus(nT) + mgus(n7) * and on eliminating uy(n7, us(n7), and us(07), we have HHT) = a,x(0T — T) + by yu — 7) + by y(n — 27) mim, y= my + mg by = mg mg — my mg | where a, Evidently, the analysis of digital-filter networks consists of solving a set of Simultaneous equations. Hence signal flow-graph techniques can be used [2.3] | 1.6 INTRODUCTION TO TIME-DOMAIN ANALYSIS The time-domain analysis of analog systems is facilitated by using several elemen- {ary funetions such as the unit impulse, the unit step, etc. Similar functions can be Used for digital filters. Such functions are defined in Table 1.2. The discrete-time Unit step, unit ramp, exponential, and sinusoid are generated by letting t= nT in the corresponding continuous-time functions. The discrete-time unit impulse S{nT), however is generated by letting t= 7 in the pulse function p,(t) given by 1 for|t|0. | ao Braise | ull Figure 16 Impulse response of nT fisvorder filter (Example 13a) For a = 0, using "Hospitals rule, we gst | nT) = tim Me" ide | | | | WaT) = Yd This is a geometric progression, and hence JOT) ~ ery(uT) = 1 = ea* on yor) te =nti wo d(l— eda and hence y(n) oo as n+ 00. For «> 0 lim wnt) = 5 +00 (see Fig. 1.7)| | | “alll ull l Ath cual ee ‘aT Mitsvorder ter (Example 135) Example 16 Find the response of the filter in Fig. 15a if _ hin on x07) = fh Assume that 2 <0, forn>0 otherwise ELEMENTARY ANALYSIS 17. SOLUTION The filter is linear, and so Wheel ae y(n) = & sin cont = ae x ) 1 1 gp-tar 1 i =hacmr 1 gett Ny ary ly ea) BPO — FEIT mF OT) — F vleT) ‘With the filter initially relaxed, the use of Eq. (15) gives yal) = eb n0)= 2 +ey,(-T)=1 UT) =e" bey) = et + eet QT) = "7 + ety, (T]= e% 4 ettH0" 4. eitor mL bet gg getter) meet Fete ten and, as in part (b) of Example 1.5, plan rae toe Af we let Ho) = where M(o)=|H(io)| = : SUNN Ta ee ot a1 _sin oF 8(w) = atg Hijo) = oT — tan“! POT yalnT) becomes 07) = AM(ettor=n ge Amy-ony By replacing w by ~w in y\(nT) we ge: YalrT) = M(—efelA-o-o — geetnssmemreany and since, according to Eqs. (1.7) and (1.8), M(—e)= Mo) {-e) = ~H(w) (16) (1.7) (18) (is) (1.10)lll ll IW lll, li POW see (1.6), (1.9), and (1.:0) yield Meo) WoT) = (elmore — gamer nr th oter-oT1_ g= Ame) -wy = M(o) sin [on + O(o)] ~ M(w}e**™ sin [0(o) ~ wT] ‘AS m+ co, the second term reduces to zero since # <0, and therefore lim y(n) = M(o) sin [on + 6(e0)} Im effect, the steady-state response of the fiter to a sinusoid with amplitude of unity isa displaced sinasoid with amplitude M(«), as illustrated in Fig. 18 1.7 CONVOLUTION SUMMATION The response ofa digital filter to an arbitrary excitation can be expressed in terms of the impulse response of the filter. An excitation x(nT) can be written as ELEMENTARY ANALYSIS 19 xT)= ¥ x(n) [sT) forum where ‘T)= fp otherwise Alternatively ayn) = x(k TIBOT — kT) and hence sor) S suryior —a7 y Now consider a linear time-invariant filter in which #6(nT)=hnT) and From Eq. (1.11) yonT)=@ F xkTp(nT — KT) ¥ xk kT) y(n) = &x(nT) ET) R(T KT) Ifthe filter is also causal, h(n) = 0 for n <0 and 207) = ¥ xT kT = Emer — kT) (ny Therefore, if (nT) = 0 for n <0, we have wot) = FT WT = KT) = F MT)<(uT — KT) (113) ‘This relation i of significant importance in tre characterization as well as analysis of digital filters. It is known as the convolution summation. Graphical Interpretation ‘The convolution summation is illustrated in Fig. 1.9. The impulse response (kT) is first folded about the y axis, as in Fig. 1.9e, and is then shifted to the right by a time interval nT, as in Fig. 1.9d, to yield h(n? — kT), Then x(KT) is multiplied by (nT — kT), as in Fig. 1.9e. The sum of al val filter at instant 97, lues in Fig. 1.9 is the response of the08 EES TILVERS! ANALYSIS AND DESIGN: ELEMENTARY ANALYSIS 21 “7 an) SOLUTION (a) From part (a) of Example LS \| Wit) = i | Hence from Eq, (1.13) i | yOT) = Punt) = ¥ eunT — kT) = See 7 x i (faeee no @ ” lo otherwise (b) For this part, we observe that | | a x(a) = u(nT) — uu = $7) brorr-ar) and so Asa) = RuluT\— ul — 57) [eaeeee fornca | onda Ths — . = aoc forn> 4 te @ © “ ' 18 STABILITY A cigital filter is said to be stable if and only iTany bounded excitation results ina kT MNT ~ AT) bounded response, icc., if [s(uT)| 0 MoT) = CA®-"B5(O) + CA"-BI(T) +--+ DA(nT) Therefore har) =| % for n=0 [CAB otherwise (1.23) Similarly, the unit-step response of the filter is ST) = CS A-BAT) + Duo) = Example 1.10 Assuming that the filter of Example 1.9 is initially relaxed, find nT). SouuTion From Eq, (123) A(ITT) = CAMB By forming A2, A4, A, and then A'®, we have 610987 [7 5] | 33 32708 |/0) 1076 als a 1597 |[1] = 363,148 B56 REFERENCES |: R 2 Schwarz and B, Frland, ° Linear Systems.” MeGraw-Hil, New York, 1968 | 2.1. R, Abrahams and G. P. Coverky “Signal Flow Analysis Petgamom, New York, 1965, 3 BC. Kuo, “Automatic Control Systems," Prentice Hall, Englewood Clty NJ. 1962, ELEMENTARY ANALYSIS 27 PROBLEMS 1.1 By using appropriate tests, check the filters characterized by the following equations fo invariance, causality, and linearity ) asin) = 20097) where 9g >0 _|&eT-S7) for) <6 “Wxior~ $7) for x7) > 6 () axer)= (or + aT) 37) (aint) = sorter) (0) eo) = xfer 437) (9) eto) = x67) sin on @) As0T)= Ky AxlnT) where AxlaT) = s(aT + 7)— x47) () asin) (0) x07) = K, Volar) where Vale) = eT) — eT ~ 7) () AMHT) = sto + THT ASOT) = 807 + Te sin oT 12 Analyze the filter networks shown in Fig PL2, wal T + ror Pre) wun) AC G0 sry28 DIGITAL FILTERS: ANALYSIS AND DESIGN ELEMENTARY ANALYSIS 29 15 Tig. PLS shows network with three inputs and are outputs. Derive a set of equations character: ining the network xP) nary) vinr) a \ 4 a vir) my ” Figure PLS: ° 16 By using appropriate tests, check the filters of Fg. PL 6for time invariance linearity, and causality (a) The filter of Fig, Pia uses a device N whose response is given by Figure P1266) Ax(eT) = |sto7}] 000 op aa) T 3 Two second-order fer sections ef the type shown in Fig. PL2c. are connected in cascade asin (b) The filter of Fig, PL6b uses a multiplier M whose parameter is given by Fig PL. The parameters of two stetions aT ayy. dy. —byu ~Byy aN yy 235. — Pass Dye respectively. Deuce the characterization of the combined fet m= ONT) (c) The flterof Fig, PL.e uses a multiplier M whose parameter i given by sinT)o—— j—| |—»—0 snr lye) en where 7) na independent signal 444 The two sections in Proh, 13 are connected in parallel as in Fig P14. Obtain the dilference “equation ofthe combined fer crio—-| Oo I Piguee Pia24 DIGATAL FILTERS: ANALYSIS AND DESIGN PHL | ow wrote -26 San Figure Pia) MT}0—» T oh EK Figure Pie) 17 Show that OT) HITT) ole) = wot) seer) = Tale) here Mle) = 20 p07) : niarya [TT tornz0 14 The fiter of Fig PLR i intalyeelaned Find the time-domain response for n= 0, finan? form =o to otherwise where w= a/67 and T= 1. OT) 0» TFL | ' 1 O—ormns | 4 $m --[F—! Figure Pa ELEMENTARY ANALYSIS 31 1.9 (0) Show that ° form <0 "OTN TE wT kT) ernie (b) By using this relation obtain the unt-ramp response ofthe iter shown in Fig, Sain closed form, The filter is initially relaxed, (6) Sketch th response for > 0, x = 0, and « <0, 1.10 The excitation in Fig 150 i [t foroznse MoT)= 52 for n> (0 tom nao Find the response in close form. LAL Repeat Prob. 110 for an exsitation (0 fornwo XOT)= 30 forn=0, I forn>a 2.3.4 12 Fig. PLZ shows a second-order recursive iter Compute the unitstep response for 0 0 Alternatively, for any value of fin) = unt ~ TYAS" THE = TRANSFORM 43 Because ofthe uniqueness of the Laurent series ina given annulus of conver- gence, the inverse z transform can also be obtained in other ways. For example, by equating coeficients, by performing long division, by expressing F(z) in terms of binomial or partial fraction expansions, or by using the convolution theorem. Use of Binomial Expansion ‘The binomial expansion can be used in simple one-pole z transforms. Example 2.3 Find the inverse z transform of SowuTion F(z) = £ (i= we (1+ wen! 4 wz Kot we? 4 wt) Emer = TK" Hence (oT) = un — T)Kw"* Use of Partial Fractions The z transform can be expressed in terms of partial fractions as F@)= DF) where F(z), F(z)... are simple one-pole z transforms, Thus. 2 Fe) = LX F (2) where each inverse z transform on the righthand side can be obtained from Table 2. Example 24 Find the inverse z translorms of (a) (0) FGSoLUTION (a) F(z) can be expressed as Fe) and from Example 23 f(nT) = QulnT ~ TIP! = un — TAY“ = au(nT ~ TYGY - QP] (6) By expanding Fle) whee =e 1 and so =P From Table 2.1 S(T) = u(nT)(—J05 + ip) = Gye (amt — em re} San = 24)%u(nT) sin Alternatively, one coud expand F(z) into partial fractions, as in part (a). Use of Convolution Theorem From Theorem 2.6(a) 2 TFE)GE)] =i SURT)g(nT — kT) Thus, if a 2 transform can be split into a produet of two z transforms whose inverses are available, pecforming the convolution summation will yield the desired inverse, Example 25 Find the inverse transforms of THE 2 TRANSFORM 45 Sotution (a) Let Fle) ad G)= From Table 2.1 FQwT)=ulnT) and g(nT)=u(nT — 7) and hence for n < 0, the convolution summation yields y(n) = 0, For n > 0 HOT) = Suk Tut ~ 7 kT) = uO)u(nT = 7) + (Tun ~ 27) + + (nT — Th(0) + wn T)u(—7) =Telte $1 +0en Therefore WoT) = nun) (6) For this example we can write in which case ST) = nus) a(oT) = unt — 7) Thus for n <0, y(n) =0. For n > 0 ku(kT)u(nT — T — kT) fun T — T)] + Uu(uT — 2T)] +--+ + (n — 1u(O) + mu(— 7) SOF1424-4n-140 =Tk Now by using the following simple manipulation Sale 2b geet + Thm bende gd 2D tng enwe obtain HOT) = 5 k= inn — 1) Therefore y(n) = Ju(a — juin) 2.6 COMPLEX CONVOLUTION ‘The inverse = transform of a produet of z transforms can be formed by using the real convolution. By anclogy, the z transform of a product af twa time-domain fanctions can be formed by using the complex convolution, ‘Theorem 2.7: Complex convolution If 2 (nT) = Fl TylnT) = G(e)= Y glnt)s-* then Ye)= 21 oT mor) = 3. f Fina?) a = aul F{F}otex é where Ts (or T,) is @ contour in the common region of convergence of F(e) and G(e/e) [or F(z/o) and G(o)] Proor From Eq, (24) hs (jon) ¥(e)= IFT, is circle inthe region of convergence and v= pe! and z= re, the above integral can be put in the form Vee) = (rele cre oem a0 THE 2 TRANSFORM. 47 which is recognized as a convolution integral The complex convolution will be used later in the design of nonrecursive filters. Example 2.6 Find the = transform of (nT) = u(nT)eo™* sin on T SouuTion Let SWT)=ulnT)e"™ — and —_ gn T) = u(nT) sin on From Table 2.1 Feo= sin oT and hence ¥(; F(z/v) and G(v) converge in the regions lel <|ze"] and fo] >t respectively, as illustrated in Fig. 23. By evaluating the residues at v= e*/°? and eT we have 7 sin oT VO)" arn Teos oT FEET Figure2.3 Region of convergence of F(t) and (te) (Example 26i 1 48 DIGITAL FILTERS: ANALYSIS AND DESIGN REFERENCES Jury, “Theory and Application of the =Transform Method,” Wily, New York, 1964, EL. E Kreysg, “Advanced Engineering Mathematics” Wiley, New York, 1968 1 2 PROBLEMS 21 (a) Find all Lauren series of with center = (6) Find all Laurent series of with center 2 =0 (c) Illustrate the region of convergence foreach ofthe above seis 22 Find the: transforms of he fllonng functions: {o) ulrT\Q+ Se") (0) er) ~ 007) fe) #07) (4) Yule7) {ewe + (Ine (7) aot) sin (on + Y) {a}. wot) cosh ax (i) wer () worn? () woryer? y) weryere-r demi therein of eonvergence in each ease 23 Find the = eansfoems of he olowing dseretsime sana i sore mraenst 10 otherste wy xpryelt® (ordsnss teense for <0 forO 0 0 for co ar forcn 'H(2) = a [ H(e)er* de For n=0 and for n > 0 HT) = Ses [H(ek"! 5 res (H(2))pt"" With re such that nSty Eres [H(e)}4~ team] lH + $F res ate) a-* G8) Since H(z) is analytic in the neighborhood of point z = 0 and each point z ‘h(0) as well as the residues of H(z) are hime, and so H()| Rew P wstailty cee ome re43 Peemisib tion of the poles of plane cepion for the lca‘56 DIGITAL FILTERS: ANALYSIS AND DESIGN 1m nonrecursive filters the poles are located at the origin of the = plane; ie, these filters are always stable. Example 3.2 Check the filter of Fig. 34 for stability, SoLuTION The transfer function of the filter is — aa res Bal corn cor here pursndeA Since Inked <1 the filter is stable. Stability Criterion Any digital filter can be checked for stability by finding the poles of the transfer function. An alternative and quicker approach, however, is to use a stability criterion due to Jury [I], This is analogous to the Routh-Hurwitz eriterion in analog systems [2]. Consider a filter cha-acterized by NG He 9) where De)= She"! and assume that by > 0. The coefficients of D{2) can be used to construct an array (of numbers, as in Table 3.1. The firs two rows of the array are formed by entering “4 o£ ot | Figure 34 Second-order ecunive iter (Example 3.2). THE APPLICATION OF THE: TRANSFORM 87 Table 3.1 Jury's array Row Coeficents L by by 2 De ee oe 3 @ 4 fet 4 ee! s 4 ay tes 6 Bye dey oo dp Wes non on the coefficients of D(z) directly. The elements of the third and fourth rows are computed as bo bys) on f LN=1 by | those ofthe fit and sinth rows as d, da ‘or i=0, 1, N-2 and so on until 2N ~3 rows are obtained. The last row will comprise three elements, say ro. ry, and rp Jurys stability eriterion states that the filter characterized by Eq. (3.9) is stable if and only ifthe following conditions are satisfied (i) Day>o (ii) (-1)D(-1) >0 Gil) by > [oul Heol > eval [dal > [dys] iro] > Ira Example 3.3 (a) A digital filter is characterized by‘58 DIGITAL FILTERS: ANALYSIS AND DESIGN Check the filter for stability. (b) Repeat part (a) it Pew NO Tey ee TS SouuTion (a) Jury's algorithm gives the following table: ve eee if ooo oo ae et 6 ons a 219 Since Di)= 11 (~1)*0(=1) = and bo > [bal |¢o| > fea]; Idol > [da|, the filter is stable. (b) In this case (=1'0(-1) = -1 ie, condition (ii) is siolated and the filter is unstable 34 TIME-DOMAIN ANALYSIS Tee limerdomain response ofa digital filter to any excitation x(n) canbe readily obtained from Eq. (3.1) as HoT) = 2-"[H(e)X (2)] ‘Any one of the inversion techniques described in See. 2 can be used Fxample 34 Find the unt-step response ofthe filter shown in Fig. 34 SoLuTION From Example 3.2 yee HO Ne =m) weet neta at ‘THE APPLICATION OF THE > TRANSFORM 59 20 Mon Figure 35 Unit-step response nT (Example 34) and from Table 2.1 Hence y(n) = -*[H2)X(2)] _ oe aT et-24 4. goto sree) = 2unT) + = ulnT) cos |e - 94] Vat ‘The response of the filter is plotted in Fig. 3. 3.5 FREQUENCY-DOMAIN ANALYSIS ‘An analog filter characterized by a continuous-time transfer function H(s) has a steady-state sinusoidal response of the form im Rule) sin cot = M(c) sin [wt + O(e)} rane where M(o) = | He) 5 iter at frequency «. The plots of M(ov)is the gain and 0(e)is the phase shilt ofthe filter at frequency Wis) and yw wi athe pate aa Pane ome im y(t) Ooo) = arg H(jeo)(60 DIGITAL FILTERS: ANALYSIS AND DESIGN respectively, constitute the basic frequency-domain characterization ofthe filter. If these plots are available, one can determine the steady-state response not only to any sinusoidal waveform but also to any other waveform that can be expressed as a linear combination of sinusoids. Let us consider an Nth-order digital filter characterized by H(2), and for the sake of simplicity let us exclude multiple poles. The sinusoidal response of such a filter is HAT) = 2-"[H()X (2) z sin oT where X(2) = ul) sin con? — (=e —e I errs t or WOT) = 55 HEIN) ds ve = Yores [H(2)X(z)2"""] Gul) For » > 0, Eqs. (3.10) and (3.11) yield i ylrT) = ¥ res [Hle\]X(p.)ph* zy [a(e'*")elT — He 7-7] Ina stable filter |p;| <1 for i= 1,2,..., Nand hence as n -» co, the summation part in the above equation tends to zero, Therefore, at steady state HOT) = 5 Lilet — Hehehe] Now He’) =F hart | z women tor|” a reer and if H(eh") = Mlwe™™ where M(o)= |H(e"")| Gl) = arg He") the response of the filer ean be expressed as aT) = M (wo) sin [on + O00] Clearly, the effect ofa digital filter on a sinusoidal excitation, like that of an analog filter, is to introduce a gain M(o) and a phase shift (a). Therefore, a digital filter, like an analog filter, can te represented in the frequency domain by an amplitude response and a phase response. The main difference between analog and digital filters is that in the first case the transfer funetion is evaluated on the imaginary axis of the s plane wheress in the second case this is evaluated on the unit circle |2| = 1 of the z plane THE APPLICATION OF THE 2 THANSFORM 61 For a transfer funetion expressed in terms ofits zeros and poles as in Eq. (3:3) Ho [](e”" — 2) (ec) = M(a)e™ and by letting (2) (13) we obtain Lhe De ‘Thus M(q) and 6(~) can be determined by drawing the phasors of Eqs. (3.12) and (3.13) in the 2 plane and then measuriag their magnitudes and angles. This Procedure is illustrated in Fig. 3.6 for a second-order filter. plane igre 36 Phasor diagram for a second.onder recurve filer(62 DIGETAL FILTERS: ANALYHS AND DESIGN The amplitude and phase responses can be obtained by repeating the above procedure for a number of points on the unit circle |2| = 1 Points 4 and C in Fig. 36 correspond to frequencies 0 and n/T, and one complete revolution of e about the origin corresponds to a frequency increment of w, = 2n/T rad/s, Since T is the period between samples, co, is said to be the sampling frequency. T phasor e in Fig, 3.6 is rotated k complete revolutions, the values of M(x) and 60) will obviously remain unchanged, and as a result He") 1m effect, H(e") is a periodic function of frequency with a period «,. The Physical reason for this is easy to identify. Discrete-time signals sin onT and sin {(o + k,)nT} are numerically identical, as demonstrated in Fig. 3.7, and there- fore the filter must respond to both of them in exactly the same way. A typical set of amplitude and phase responses is illustrated in Fig. 3.8. In future we shall be concemed almost exclusively with the fundamental period of He"), which extends from —c,/2 to «,/2. This frequency range is called the baseband. The frequency «,/2, witich corresponds to point C in Fig, 46, is often called the Nyquist frequency. In later chapters we shall be dealing with lowpass, highpass, bandpass, and bandstop filters. In digital filters these terms are defined in relation to the baseband, Hleko tow) Example 35 The constants in Fig. 29 are Ao=04 A, =0303 4, = 0.0935 Find the amplitude and phase responses of the filter if « SOLUTION The transfer function of the filter is PEA H(z) = 222 + Ase Ay +A or sin Heo + 4a, a°P) Figure 87 Plots of sin on} and sin [fr + o4)0T] versus nT: ‘THE APPLICATION OF THE 2 TRANSFORM 63 Ma) Figure 88 Typical frequeney response for a digital filter: (a) amplitude response and (4) phase response Hence rT 4 eRe) 4 Ale + HT) 4 Ay He") ee 24; 008 20T + 24, cos wT + Ay and so Mw) = |2A, cos 20T +24, cos wT + Ag| Olen) = 0 — oTt | een —EE Figure 19 Fourth-order, noncecursive filter (Example 3.5) [0 i 24, cos 207 + 2A, 008 wT + Ay 20 where y= |x otherwise ‘The amplitude and phase responses are plotted in Fig. 3.10. Figure 340 Frequency response (Example 3.5): (a) amplitude reponse and (b) phase response {THE APPLICATION OF THE 2 TRANSFORM 65 © Figure 310 (continued) REFERENCES |. E I. Jury, "Theory and Application of the Transform Method," Wiley. New York, 1968 “Linear Systems” MeGraw-Hil, New York, 1965, 2 RJ. Schwarz and B, Freda PROBLEMS [3M Derive the transfer functions of the fites in Prcbs. 1.2 to 14 132 A recursive filter is characterized by the equations OT) =I 6T) 4 Syl 7) $iy(0 27) NOT) = x07) + ATT) Obtain the tansfer function ofthe ite. 33 A iter is represented by QUoT + T) = AgloT) + BefeT) — y{nT)= Calor) + Dx(nT) o 1 o 0 nb t 1 Deduce the transfer faction of the fer | c-8 4 D-Uy= — ‘© DIGITAL FILTERS: ANALYSIS AND DESIGN ‘THE APPLICATION OF THE = TRANSFORM 67 34 Show that ‘310 Find the permissible ange fr m in Fig. P2.10if the iter isto be stable HG) Zi : se Shah! represents ac fier only if = N. 35 (o) Find the impuise response of the filter shown in Fig 32. (6) Repeat pact (a) forthe filter of Fig 34. ‘36 Obiain the impuise response of the fiter in Prob, 3.2, Sketch the response, [37 Starting fom frst principles, show ‘represents a stable filter. 8 (a) A cecursive tris represented by * Figure P310 oe ey G4 SAF FP FI ae (Cheek the iter for tabi (6) Repeat pr ot ‘MI Find the unitstep respons ofthe filter shownin Fig 32, ‘312 Find the unitramp response of the fiter shown in Fig. 34 i T= 1 tao ‘313 The input exciation in Fig. 3.2 is He) Be +5443 + Ed ts rosns2 39 Oban a he taser functn, (the pulse response and (the secsty conn for wens idon frzcn se stability forthe filter of Fig P39. The constants mand m, are given by 0 for na oe | one Dace he epee or 05 by wing the: asform, M4 Repeat Prob. 112 by using the = transform. For each of the three eases deduce the exact Trequency of ringing if T = 1 and also the steady-state vale ofthe response. BMS A filter has a transfer function 1 aes (nd esos : (nT) = a(n 7) sin wT (b) Dedce the steady-state sinusoidal response. U6 A filter i characterized by 1 stor} LF 5 J+! Hela where |r| <1. Show that the seady-statesinusoida response is given by yin lo) | ate") oe MoT) = Mo) sin (omT + Of} where Mle ™ [Mle 40) = arg He")317 (0) Show that (0) = Maye () By using this relation show that er ‘epresents am allpas ke, ie, ene with constant gain at all requencies, 38 Figure P3.18 shows a nontecursve filer, (a) Derive expressions forthe gain and phase shit (0) Determine the transmision zeror ofthe fe, ie, zeurguin frequencies. (6) Sketeh the amplitude ard phase response. wiry LG Tr + Mary 2605 uy oe ‘M9 The group delay in digital asin analog fers x dened by 42h) do Figure Pas ‘Show that the equation ser) nT) + 2007 7) + 3x(WP — 27) 4 4ufn— 37) + Sx(0T — 47) + Dein — 57) 4.x(ur — 67) fepresenis a consantelay fier. ‘320 Derive ruretions forthe gais and ple shi ofthe fier shown in Fig 34 Jt able PR21 gives the rant: function coeTiens of fur files labeled 4 to. By ploting “20g lo) vers fo values onthe rage Ota 50 re tely ove et bandpass, and a bandstop filter. Each filter has a transfer function of the. form — 8nd the sampling frequency 10 rads in each case. ‘322 Show thatthe gain and phase shift in a digital filter satisfy the relations M(o,—)= Mla) and, — 0) = ~oho) ‘THE APPLICATION OF THE 2 TRANSFORM 69 Table P3.21 Filer 1 Ay a Be By Lo 2222865 x 10"! 4.445001 x 1? 4520149 x 10? 1561833 x 10 AD AO8S38G 10-6177 I-14 509715 x 10-2 168171 x 10"! u=10 end eee eee 71226400 x 10" 4984635 « 10-* 1 5490566 9752985, B 2 STOR x 10") — 1.042887 1226400 « 10-4 94634 « 10-* H = 2816456 « 107? ——— SS—S—sesessssSSSSs 1787144 10-4 LSIT2IO KC 1-_5741567 x 10-* 1.224608 © 2 1399382 1214846 x 19-7 5741567 10-1 =1.224608 H = 8912809 % 10" 1 9308915 LS6IBOL x19 S087094% 197 —1291110 D2 2300089 Later 092186 « 10-* — — 1069291 1H = 6.665086 x 10-*| cuartee FOUR _ REALIZATION 4.1 INTRODUCTION ‘The design of digital filters comprises four general steps: 1. Approximation 2. Realization 3. Study of arithmetic errors 4, Implementation ‘The approximation stepis the process of generating a transfer function satisiy~ ing a set of desired specifications, which may concern the amplitude, phase, and Possibly the time-domain response ofthe filter. The realization step isthe process (of converting the transfer function into a filter network. The approximation and realization steps are carried out on the assumption that the arithmetic devices to be employed are of infinite precision. Since practical deviees are of finive prevision, it becomes necessary to study the effects of arithmetic errors on the performance of the filter The implementation of a digital filter can assume two forms, software and hardware. In the first case, implementation involves the simulation of the filter network on a digital computer. In the second ease, on the other hand, this involves the conversion of the filter network into a dedicated piece of hardware We consider here the second step in the design process, namely the realiza- tion, In addition, we discuss some important topological properties of digital-ilter networks in general ™ Realization can be accomplished by using the following methods. 1. Direct 2, Direct ex 3. Caseade 4, Parallel 5. Ladder 6. Wave ‘The fist five oF these methonts will he considered in the following wetions ‘The wave etext oy we wy the sues alse ity of Hh yin thn yl dn samvidevva faley, A Chay 1 4.2 DIRECT REALIZATION Let us consider an Nth-order transfer funetion such as HG) (a) where (42) and We can write Y(2) = Uy(e}+ Usle) where Vie) = NEWx(2) (43) and Ux(2)= ~DE)Y(2) and hence the realization of H(z) can be broken down into the realization of two simpler transfer functions, N(2) and —D’(), as illustrated in Fig. 41 Ye) Melo} ney; 90 0121 wy | -v |S Figure 4.4 Decomposition of H(=) into two simpler tanslerfunetions.| Consider the realization of N(z). From Eqs. (4.2) and (4.3) Uy(2)= [ao +N AEIXE) where MG)= Sat and thus N(z) can be realized by using a multiplier with parameter ay in parallel with a network characterized by 2~1N,(z). In turn, z~!N,(z) can be realized by using a unit delay in cascade with a network characterized by N (z)-Since the unit delay may precede or follaw the realization of N,(z), two possibilities exist for (2), as depicted in Fig, 42. ‘The above procedure can now be applied to N,(z). That is, N,(z) can be expressed as N,(z)= a, +274N3(2) where N,(z) land as before two networks can be obtained for Ny(z). Clearly, there are four networks for N(z). Two of them are shown in Fig. 43, ‘This eycle of events ear be repeated N times, whereupon Ny(2) will reduce to a single multiplier. In each eycle of the procedure there are two possibilities, and since there are N cycles, a total of 2" distinct networks can be deduced for N(z). Three of them are shown in Fig. 44. —D'(z) can be realized in exactly the same way. Networks for — D'(2) can be obtained by replacing ag, a,, a3, .. in Fig. 44 by 0, —by, ~ba, Finally the realization of H(z) can be accomplished by interconnecting real- izations of N(z) and — D(z) according to Fig 4.1. Figure 42 Two realizations of N=} xe 6 — al Figure 43 Two of four posible realizations of N= x0) Figure 4 Three of 2" possible realizations of N(2} REALIZATION 73 He© j igure 4. (continued) Example 4.1 Realize the transfer funetion SoLTION Two realizations of H(z) can be readily obtained from Fig, 44a and 6, as shown in Fig 45a and b REALIZATION 75 o Figure 48 Two realizations of fc) (Example 41)i 4.3 DIRECT CANONIC REALIZATION . A Gigital network is said to be canonic if the number of unit delays employed is equal (0 the order of the transfer function, Equation (4.1) can be expressed as Y@)=NE)"@) where Ye) = X(2) - D(e)r(@) in which case 1(2) can be wealized as in Fig. 46 by using the network of Fig. 44a for N(z) as well as - D'2) Evidently the signals at nodes 4’, B,... are equal to {he respective signals at nodes 4, B, .... Therefore, the unit delays in path AB’ can be eliminated to yield a canonie realization for H(z) 4.4 CASCADE REALIZATION ‘The transfer function can be factored into a product of second-order transfer functions as He) = [ine H{z)= Goi + ay,27* + 3,27? Zh where THB ue + bye? 409 Figure 46 Derivation ofthe canoni: realization of H(2). REALIZATION 77 Ho) me) fe fy i) ) Figure 4.7 (a) Cascade realization of H(z); (6) canonic second-order section. Hence (2) = [H,(z)X(z)JHa(2) ++ Hy(2) = [a(2)% (2) 3(2) -~ Hale) = Hul2)¥u—1@) where Ye) = HAe)X(e) Mie)= HG) fori=2,3,...,M=1 In this way, H(z) can be realized by using the cascade configuration of Fig. 4.7a. The individual sections can be realized by employing the canonic second-order network of Fig, 4.7b, 4.5 PARALLEL REALIZATION Alternatively, the transfer function can be expanded into partial fractions as ste) = Steeier +)- Myto Figure 48 Parallel celization of (2). where In this case and so the parallel configuration of Fig. 4.8 is obtained. ‘An alternative parallel realization can be readily obtained by expanding H(2V/2 into partial fractions. 4.6 LADDER REALIZATION “The last realization method :0 be considered here isthe ladder realization, due to Mitra and Sherwood [1]. This is based on the configuration of Fig. 49a With N= 4 in Fig. 49a straightforward analysis yields YE} = MEG) + HaeNe) (aa) REALIZATION 79 re Figure 49) Ladder realization of (a) H(z) = (—1)'/D(2)and (6) He) = M(cVD(2) hey| ‘WY DIOITAL FILTERS: ANALYSIS AND DESIGN MG 2)? where Hy(2)= 1 mrt mrt mre 1 WO=55 and since Y@=-Ye) Ea) gives t H(z) xe) "9" Te ‘Similarly, for any value of N it can be shown that eye) 1 Dalz) myz+ mz where k is the largest integer equal to or less than N/2; that is, wa [tl for =1,4,5,8,9, Cy oe Let us consider the transfer function | De) ~1+D@) wine where Di Dy(z) ~ rigmymym,z* + (mgm + mam, + mzm,je* +1 (45) We can write y He) Oop Gi4 DE) (6) where Od D(z) and Ev D(z) denote the odd part and even part of D(z), respee- tively. By comparing Eqs. (45) and (4.6) the following identifications can be made: 04 D(z) y= |EPDE) for even N | BPE for odd w for even N ’ for odd N Now by expressing H,(z) as a continued-fraction expansion about infinity we have Hy(2)= az+— (47) and therefore if maa the configuration of Fig. 49a becomes a realization of H(c). ‘The synthesis can be extended to any transfer functions of the form by modifying the basic configuration as illustrated in Fig. 496. For N= 4 and ‘m= ¢;, analysis yields 2 din(z) He)=where nG=1 ny()= G2 41 nis(z) = ~(eye,2? + 62 +1) ~Lerczesz? + e20y27 + (ey + 4)2 + I] By assigning Ne)= Sdn(e) =cees 0 0 OV fae) fas city ~eren 0 Off ds| _ fay late) -e ¢ Offa} = fa 48) “1 -t 1 aflal [as can be formed. The solution of this equation yields the necessary values of dy, 4, ‘The realization method relies on the existence of the continued-fraction ex- pansion in Eq, (4.7), and as a consequence it sometimes breaks down, eg, if H(z) has poles on the imaginary axis of the z plane [1]. Example 4.2 Realize the transfer function 3.517 + 0.6652 + 0.6652? — 3.5172" 3.2662 + 3.7392? — 15329 using the ladder method. SOLUTION H() can be expressed as 0,0882 + 0.0817: ~ 0.07932? = 32662 + 3.7392" = 1532 = 9.02299 +H) and hence f(z) can be realized by using a multiplier in parallel with a network haraterized by He. Sings the order of 1/2) ig odd, we can write H(z) = 002299 + 3.7392 +1 15327 + 3.2662 1 0.40922 + —————__ i ~ 1309241 1300+ 5562 REALIZATION 83, and thus my =e, = -04092 m= c= -1309 m= cy = —2856 For N= 3, analysis yields n()==1 ng(2) = ~(e,2 +1) ns(2)= eyegz? + ez +1 and so dy, d3, and d, are given by (Pt AE or oa 2B Lory 0 © dy ~ ay = 0.184 4.7 TOPOLOGICAL PROPERTIES Digital-filter networks can be represented in terms of flow graphs which have certain topological properties of theoretical as well as of practical interest. We discuss some of these properties here. Signal Flow-Graph Representation Consider the canonie network of Fig. 4.7b, and let y4(n). y(n)... be the signals flowing out of nodes 1, 2,.... We can write alr) = in) Faltt) = it) ~ by ys(o} ~ bayaln) sn) = *ya(n) valn) = &"ys(n) Isl) = ao yaln) + ay yon) + a2 yal) o(n) = ys(n) where i(n) and o(1) are the input and output, respectively. Thus a conventional flow graph can be constructed for the network comprising a number of nodes, a number of directed branches, a source i(7), and a sink o(n), as illustrated in Fig. 4.102. The source and sink can be treated as branches by using the sel-loop of|84 DIGITAL FILTERS: ANALYSIS AND DESIGN « xine atn) Fitwe 4.10 (a) Flow-graph representation of the iter depicted in Fig. 7b (6) Representation of @ source ora sink by means of eléloop Fig 4.106, where qin) isan independent quantity for a source and g,(n) = O for a GK By designating the signal entering node j ofthe network via path asx, and the signal flowing out of node j as y, the preceding equations can be put in the concise form ye Yxy forj=2.,5 where xy = 0 ifno direct path exists between nodes w and v, Similarly, for a digital network with N nodes Y= Yxy forj=tuwN Tellegen’s theorem Consider two distinct sets of signals § and S' defined by s Panes ams Disooee Ya} 8 (has ces Man Yo vos Doh ‘These may pertain to one and the same network, in which ease their dilfe ences may be due to dllerences in the input signals or possibly to variations in one OF more mulliplier constants. Alternatively, § and S? may pertain to REALIZATION 85 distinct but topologically compatible networks (networks which have the ‘same topology). Tellegen’s theorem as applied to digital networks [2, 3] states that the elements of $ and $' satisfy the general relation E Doni-ve)-0 Proor The validity of this theorem can be established by writing the left- hand side of the relation as and then replacing by yi and y,, respectively. Similarly, if Wy=Xy Ly=Y, the preceding steps yield LOX Vix) =0 (49) mm ie, Tellegen’s theorem holds in both the time domain and z domain, Reciprocity A flow graph with M accessible nodes is sid to be a multipole or M-pole). Such a flow graph can be represented as depicted in Fig. 4.11, where F is a source-ftee flow graph, Consider an M-pole, and let SH (Xe Xa Yo Mad S' Pee eee Yih be possible sets of signals. The M-pole is said to be reciprocal 2] it D&yi- xin) =0 (4.10) for all possible pairs of $ and S. ‘The M-pole can be represented by the set of equations HgX) fori=1,2,...,M (4.11 aX, m a.| {86 DIOKTAL FILTERS: ANALYSIS AND DESIGN O20 Figure 4.11 Multipoe low graph, where Hy is the transfer function from node j to node i. From Eqs. (4.10) and (4.11) Seu aai-xEnxjeo or ES ayxixy— SF Layx,xi 0 and on interchanging subscripts in the second summation we have XY Ly WyXiXj=0 Therefore, an M-pole is a reciprocal flow graph if and only if for all values of i and j. Interreciprocity Now consider two distine: M-poles G and G’ having the same number of acces- sible nodes, and let $ and S' as defined above be sets of signals pertaining to Gand G, respectively. The two M-poles are said to be interreciprocal [2] if LOs¥i= xiy=0 (4.12) for all possible pairs of $ and S" | REALIZATION 87 G and G' can be described by Y= Dax, fori= ly QoM (413) and Yee DX; forim2..0M (44) respectively. From Eqs. (4.12) to (4.14) E Sen, — myx and therefore G and G’ are interreciprocal flow graphs if and only if Hy= Hy (415) for all values of i and j ‘Transposition Given a flow graph G, a corresponding flow graph G’ can be derived by reversing the direction in each and every branch in G. The flow graph so derived is said to be the transpose (or adjoint) of G [2-4]. If the transmittance of branch ij in G is designated as By and that of branch ji in G as B, then by definition : (4.16) Let S= (Kar eey Xs Vaveees Wah S'= (Xba coe Xow Ve ees PN be possible sets of signals pertaining to G and G’, respectively, and assume that nodes 1 to M are accessible. From Eq, (4.9) LOX VEX) + z ix, = Ht) +E F 04x, — 1 0 ‘at (4.17) ‘The first summation represents contributions due to external self-loops, ie, sources and sinks, whereas the second and third summations represent contribu: tions due to internal sel-loops and internal branches, respectively. For both inter- nal sellsloops and branches we can write Xy= 4% and and hence EEX N= YE Oy-anjyixo iat fs YON VX = 1 (Bi BAY‘88 DIGITAL FILTERS: ANALYSS AND DESIGN according Eq, (4.16). For external sell-loops Xu=X, according to Fig. 4.1, and as a result Eq, (4.17) simplifies to Eq, (4.12), Therefore G and its transpose Gare interreciprocal flow graphs. Consequently, if H,, is the transfer function from node i to node jin G and His the transfer function from node j to node i in G’, then Hy= Because of this property, t-ansposition can serve as a means of deriving alternative digital filter networks. Example 4.3 Form the transpose of the canonie network of Fig, 47h SouUTION The transzose flow graph is readily obtained, as shown in Fig. 4.124, by using Fig. 4.10a The corresponding network is shown in Fig, 4.126 a © — *% : ; oo oO ARC ao! T 20 aQ a _f a ‘ @ Figure 4.12 (a) Transpose flow gaph of the network depicted in Fig. 7b (Example 43) (6) Cor ‘eipanding transpose network, REALIZATION 89 Sensitivity Analysis In the study of arithmetic errors (see Chap. 11) we shall be concerned with the sensitivities of digital-ilter networks to variations in the multiplier constants, In a network characterized by H(z) =S(z, my. my...) where m,, m, ... are multiplier constants, :he sensitivities are given by oH(e) Sale) = oy, fori= 22, Although differentiation ofthe transfer fanction will ultimately yield the sensi- tivities, much manipulation is often necessary, especially in complicated networks. Fortunately, however, differentiation can beavoided through the concept of trans: position [2, 3], as we shall now show. The network of Fig, 4.13a, the transpose of this network, and the network with m changed to m + Am can be represented by the 2-poles of Fig. 4.13b, c,d, respectively. By virtue of Tellegen's theorem yw=0 (4.18) where Wa Y (VjX— YEG) For MYX VN + ¥ (By Bal¥y en (4.19) mt since XueXj=0 Xpex=Q By For i=2 : Was YiXa2 — VaXia + Y (Bry — Ba) Vj¥3 = QY3 (4.20) since Xn=X:=Q Xn=Xj=0 By=By Fori Wy = VaNa ~ VyXt5 + 5 (8)— By = ~AmYs YE (4.21) since Xye=mVy X= (m+ mY By Finally, for i> 4 Wem (8, Bayyi 2)©) Figure4.13 Derivation of sensitivities: a) digital fiter network, (b) two-pole representation, (c) ‘o-polerepreentaton the transpose network, (a) wo-pole representation of the network with im changed to m+ Am since fori24 Consequently, from Eqs. (4.18) to (4.22) Wi+ + =0 Vi or yy-y,=" Now 2h = Hi, — Hy = and from Eqs. (4.23) and (4.24) AH ‘Am As Am 0, Hig -+ Hya, and therefore AR, Sale) = im, aa His Hi = Ayia Aya(2)H32(2) (4.23) (4.24) (425) ie, the sensitivity to variations in multiplier constant m ean be formed by mul- tiplying the transfer function from the input of the network to the input of the multiplier by the transfer function from the output of the multiplier to the output of the network. This technique can eliminate considerable manipulation in com- plicated networks. Example 44 Find the sensitivities for the network of Fig. 4.14, ts $ 3 Figure 414 Second-order fer (Example 44).js eet arnt OE DIGENG FILTERS; ANALYSIS AND DESIGN SoLUTION Straightforward analysis yields Aaa(2) = Hoa(z) = Hence from Eq. (425) 0 2 +1? ” (+1pP SMO rae MSE = ee ET REFERENCES 1S K. Mitra and R, J. Sherwood, Digital Ladder Networks, IEEE Trans. Audio Electroacoust, vol AU-2, pp. 30-36, Februny 1973 2 A. Feitweis, A General Thecrem for Signal-Flow Networks, with Applications, Arch. Elektron Uebertrag, vol. 28, pp. 557-56, 1971 3. RE. Seviora and M. Sablatash, A Tellegen’s Theorem for Digital Firs, IEEE Trans. Circuit Theory, vol. CT-8, pp. 201-03, January 1971 ‘4 LB. Jackson, On the Interaction of Roundoff Noise and Dynamic Range in Digital Filters, Bell Syst Tech J, vol 49, pp. 156-184, February 1970, ADDITIONAL REFERENCES Antoniou, A. Realization of Digtal Filters, IEEE Trans, Audio Electroacoust vol. AU-20, pp. 95-91, March 1972, (Ceochere, RE: Computational Methods for Sensitivity Analysis of Digital Fier, ML.T: Res Lab, Electron. Q. Prog. Rep. 19, pp. 113-123, Apel 18,1973, Director, $. W. and R.A. Rohrer: The Generalized Adjont Network and Network Sensitivities, EEE Trans, Cirewit Theory, vol. CT-16, pp. 318-323, August 1968, Mitra, S. K. P. S. Kata, and D, C. Huey: Cascade Lattice Realization of Digital Filters, Int. J Creat Theory Appl Yo. S,pp. 3-11, 1977 Seetupok, J, and 8. K. Mira: Digital Filter Realization Using Successive Muliplier-Extraction Approach, IEEE Trans. Acoust, Spech, Signal Process, voLASSP-23, pp. 235-239, April 1915, and Detection, Location, and Removal of Delay-Free Loops in Digital Fillet Configurations, IEEE Trans. Acoust, Speech, Signal Proces, VoL ASSP-23, pp. $88-S62, Decem. ber 1975, PROBLEMS (Goninon mt ei loi er con oy et OMe # Mage Gy wig te cede din he prelate the long ae ten ie + Ne ©) Mel imaee +3) REALIZATION 93 43 A iter is characterized by the equations AloT +7) = Aglor) + Bx(n7) 0 10 at -H2 Obtain a eanonic realization forthe fier. e Realize the transfer function Oolsde? + 004622 + 0.0462: + 0.0158 OTS 19900" + 15a — OARS WaT) = Calo) + Dx(e7) ° [ cl a 1 He) by using the ladder method. (Plo. Consrat a ow hat othe software implementation of an ston parle iter, (0) Repeat part (a) for an N-seation eascade Ble 46 Given a continuous-time transfer funetion Hs) a corresponding discretetime transfer fonction H(z) can be formed as Hole) = Hl) (Gor Sec. 76) (2) Obtain a digta-fiter network by using the transfer function 1" Fist (0) Evaluate the gin ofthe fer for = 0 and o = /T- 447 (a) The flow graph of Fig. PA. represents recersive iter, Deduce the transfer fnetion ofthe fier (0) Repeat part (a) forthe fow graph of Fig. Pe xe ye Figure P42 (0)