DISTORTION ANALYSIS OF ANALOG INTEGRATED CIRCUITS,THE KLUWER INTERNATIONAL SERIES IN ENGINEERING AND COMPUTER SCIENCE ANALOG CIRCUITS AND SIGNAL PROCESSING Consulting Editor: Mohammed Ismail. Ohio State University Related Titles: NEUROMORPHIC SYSTEMS ENGINEERING: Neural Networks in Silicon, edited by Tor Sver Lande; ISBN: 0-7923-8158-0 DESIGN OF MODULATORS FOR OVERSAMPLED CONVERTERS, Feng Wang, Rame{ Harjani, ISBN: 0-7923-8063-0 SYMBOLIC ANALYSIS IN ANALOG INTEGRATED CIRCUIT DESIGN, Henrik Floberg, ISBI 0-7923.9969-2 i SWITCHED-CURRENT DESIGN AND IMPLEMENTATION OF OVERSAMPLING Aj CONVERTERS, Nianxiong Tan, ISBN: 0-7923-9963-3 CMOS WIRELESS TRANSCEIVER DESIGN, Jan Crols, Michiel Steyaert, ISBN: 0-7923-9960 DESIGN OF LOW-VOLTAGE, LOW-POWER OPERATIONAL AMPLIFIER CELLS, Rj Hogervorst, Johan H. Huijsing, ISBN: 0-7923-9781-9 VLSI-COMPATIBLE IMPLEMENTATIONS FOR ARTIFICIAL NEURAL NETWORKS, Si Mehdi Fakhraie, Kenneth Carless Smith, ISBN: 0-7923-9825-4 CHARACTERIZATION METHODS FOR SUBMICRON MOSFETs, edited by Hisham He ISBN: 0-7923-9695-2 LOW-VOLTAGE LOW-POWER ANALOG INTEGRATED CIRCUITS, edited by Wouter Ser ISBN: 0-7923-9608-1 INTEGRATED VIDEO-FREQUENCY CONTINUOUS-TIME FILTERS: High-Perfo Realizations in BiCMOS, Scott D. Willingham, Ken Martin, ISBN: 0-7923-9595-6 FEED-FORWARD NEURAL NETWORKS: Vector Decomposition Analysis, Modelling and Anah Implementation, Anne-Johan Annema, ISBN: 0-7923-9567-0 FREQUENCY COMPENSATION TECHNIQUES LOW-POWER OPERATIONAL AMPLIFIERS, Ruud Easchauzier, Johan Huijsing, ISBN: 0-7923-9565-4 ANALOG SIGNAL GENERATION FOR BIST OF MIXED-SIGNAL INTEGRA‘ CIRCUITS, Gordon W. Roberts, Albert K. Lu, ISBN: 0-7923-9564-6 INTEGRATED FIBER-OPTIC RECEIVERS, Aaron Buchwald, Kenneth W. Martin, ISBN: 0-75 9549-2 MODELING WITH AN ANALOG HARDWARE DESCRIPTION LANGUAGE, H. Al Mantooth,Mike Fiegenbaum, ISBN: 0-7923-9516-6 LOW-VOLTAGE CMOS OPERATIONAL AMPLIFIERS: Theory, Design and Implementatiq Satoshi Sakurai, Mohammed Ismail, ISBN: 0-7923-9507-T ANALYSIS AND SYNTHESIS OF MOS TRANSLINEAR CIRCUITS, Remco J. Wiegerink, ISB} 0-7923-9390-2 COMPUTER-AIDED DESIGN OF ANALOG CIRCUITS AND SYSTEMS, L. Richard Carl{ Ronald S. Gyuresik, ISBN: 0-7923-9351-1 HIGH-PERFORMANCE CMOS CONTINUOUS-TIME FILTERS, José Silva-Martinez, “ Steyaert, Willy Sansen, ISBN: 0-7923-9339-2 SYMBOLIC ANALYSIS OF ANALOG CIRCUITS: Techniques and Applications, Lawrence Huelsman, Georges G. E. Gielen, ISBN: 0-7923-9324-4 DESIGN OF LOW-VOLTAGE BIPOLAR OPERATIONAL AMPLIFIERS, M. Jeroen Fonder, Johan H. Huijsing, ISBN: 0-7923-9317-1 STATISTICAL MODELING FOR COMPUTER-AIDED DESIGN OF MOS VLSI CIRCUI Christopher Michael, Mohammed Ismail, ISBN: 0-7923-9299-X. SELECTIVE LINEAR-PHASE SWITCHED-CAPACITOR AND DIGITAL FILTERS. HusseDISTORTION ANALYSIS OF ANALOG INTEGRATED CIRCUITS by Piet Wambacq IMEC, Leuven, Belgium and Willy Sansen Katholieke Universiteit Leuven, Belgium ae KLUWER ACADEMIC PUBLISHERS BOSTON / DORDRECHT / LONDONACLLP. Catalogue record for this book is available from the Library of Congress. ISBN 0-7923-8186-6 Published by Kluwer Academic Publishers, P.O. Box 17, 3300 AA Dordrecht, The Netherlands. Sold and distributed in North, Central and South America by Kluwer Academic Publishers, 101 Philip Drive, Norwell, MA 02061, U.S.A. In all other countries, sold and distributed by Kluwer Academic Publishers, P.O. Box 322, 3300 AH Dordrecht, The Netherlands. Printed on acid-free paper All Rights Reserved © 1998 Kluwer Academic Publishers, Boston No part of the material protected by this copyright notice may be reproduced or utilized in any form or by any means, electronic or mechanical, including photocopying, recording or by any information storage and retrieval system, without written permission from the copyright owner. Printed in the NetherlandsForeword ‘The analysis and prediction of nonlinear behavior in electronic circuits has long been a topic of concern for analog circuit designers. The recent explosion of interest in portable electronics such as cellular telephones, cordless telephones and other applications has served to reinforce the importance of these issues. The need now often arises to predict and optimize the distortion performance of diverse electronic circuit configurations operating in the gigahertz frequency range, where nonlinear reactive effects often dominate. However, there have historically been few sources available from which design engineers could obtain information on analysis tech- niques suitable for tackling these important problems. Tam sure that the analog circuit design community will thus welcome this work by Dr. Wambacq and Professor Sansen as a major contribution to the analog circuit design literature in the area of distortion analysis of electronic circuits, I am personally looking forward to hav- ing a copy readily available for reference when designing integrated circuits for communication systems. Robert G. Meyer Professor Electrical Engineering and Computer Sciences University of California Berkeley, California 1998Preface In the world of electronics nowadays very advanced systems can be integrated on one chip. This is mainly possible by the ability to build complex functions with digital VLST. However, not ev- ery functionality can be achieved using digital electronics. For example, some applications might require signal processing at a high frequency that is too high to process with digital circuitry. In that case, the signals must be processed with analog circuitry which can be integrated completely or partially. Other applications require an interface between the digital electronics and the outside world, which behaves in an analog way. As a result, such interface functions are implemented with analog electronics, which again can be integrated. ‘The above considerations indicate that analog integrated circuits are required, not only as integrated circuits on their own, but also as part of large mixed-signal integrated circuits. In - such mixed-signal systems more and more functions are implemented in a digital way. The _ specifications of the circuitry that remains to be designed in the analog domain are often very tough. As a result, circuit designers not only need to concentrate on the first-order characteristics of analog circuits, which can already be very complicated and which are most often attributed to the behavior of the linearized circuit. In addition, characteristics such as nonlinear behavior may become very critical. In addition to mixed-signal applications that demand tough specifications for the analog blocks, there are some applications where the suppression of nonlinear behavior is of utmost importance. Examples are audio applications and telecommunication applications. In the latter applications, nonlinear behavior of the circuits induces intermodulation products, which together with the noise, increase the amount of “unwanted signals”, thereby lowering the performance. In the last few years, an increased interest is seen in the integration of analog high-frequency front-ends for telecommunication applications, both in CMOS and in BiCMOS. These technolo- gies are quickly scaling to smaller dimensions, such that they can be used at ever increasing frequencies. In this way, these silicon technologies form a cheap alternative for GaAs. CMOS technologies are cheaper than BiCMOS technologies, but the bipolar (npn) transistors of the BiCMOS technologies are in general superior at high frequencies than the MOS transistors. , Integrated silicon RF front ends are found in literature for wireless communications [Long 95], used for example in the GSM standard [Seven 91, Stet 95], the DECT standard [Daw 97], and the | Japanese standard for the personal handy-phone system (PHS) {Sato 96], further in GPS (Global Positioning System) receivers [Herm 91], wireless local area networks (WLAN) [Madi 96, Har 96), in applications in the ISM bands (Industrial-Scientifie-Medical) around 900MHz [Hull 96], ,2.4GHz [Mey 97] or 5.7GHz [Voi 97], in the North American Digital Cellular (NADC) hand- set [Kara 96], and so on In analog RF front-ends for communication circuits, specifications that are related to nonlin- ear circuit behavior are very important, For example, if a receiver front-end is not very linear, then large incoming signals from the antenna will induce much extra distortion, Such large in- coming signal may be a wanted signal but it can also be a strong unwanted signal at an adjacent frequency. The increased efforts of the analog design community in the silicon integration of analog high-frequency RF front-ends has been one of the motivations to write this book. Not only the nonlinear behavior of the RF part of a communication circuit is important to control. In many communication circuits analog integrated filters are used, for example to per- form the anti-aliasing filtering function right before an analog-digital converter in a receiver. The nonlinear distortion of these filters can degrade the overall performance of a transceiver. Very of- ten gm-C filters are used in transceivers {Gopi 90, Chang 97], since they can achieve high speeds. This high speed is achieved at the expense of a reduced nonlinearity. In essence, a gzy-C cell is an open-loop transconductor, We shall see in this book that the conversion of a voltage into a current, which is realized by a transconductor, is difficult to realize with a high linearity without an overall feedback with a large loop gain. Hence, nonlinear distortion is an important aspect in the design of an active analog integrated filter. In many applications of analog circuits one is only interested in the circuits’ steady-state be- havior in response to a sinusoidal excitation or a combination of such excitations. Indeed, many circuit aspects are easier to characterize in the steady state. This is partially due to the fact that an extremely large class of analog circuits can be approximated very well by a linear system. Since sinusoidal functions are eigenfunctions of linear systems, the latter ones can be easily character- ized in terms of responses to sinusoidal excitations. Examples of quantities that characterize a circuit in steady state are transfer characteristics like gain or impedances. These characteristics are also best measured when a circuit is in steady state. Gain and impedances are mainly due to the behavior of the linearized circuit, However, most analog integrated circuits behave weakly nonlinearly, This means that, when a sinusoidal signal or a combination of sinusoids is applied ‘0 a circuit, the output spectrum does not only contain signals with the same frequency of the input signals, as one would expect of a linear system: in addition, the output spectrum contains small components — usually unwanted — at frequencies other than the input signal frequencies. If one sinusoidal signal is applied at the input, then these unwanted signals occur at multiples of one of the input frequencies. In this case they are termed as harmonics. In the case of an excita tion with more than one sinusoidal signal, unwanted components occur at frequencies which are linear combinations of the input frequencies. These components are denoted as intermodulation products. ‘The weakly nonlinear behavior of analog integrated circuits is caused by the slight curvature of the characteristics of the devices of the circuit around the operating point. This behavior contrasts with strongly nonlinear behavior, where devices such as transistors switch between an on-state and an off-state. Nonlinear behavior, both weakly and strongly nonlinear, are not always unwanted. For example, oscillators and mixers explicitly rely on nonlinearities for a Suitable operation, This book is concentrated on weakly nonlinear behavior only. In this way, the majority of continuous-time analog integrated building blocks is covered.The harmonics or intermodulation products characterize the amount of nonlinearity of a given circuit, Since sinusoidal signals and sums of sinusoids are frequently used as inputs, a sinusoidal steady-state analysis of weakly nonlinear circuits is certainly not restrictive. Such analysis is carried out in the frequency domain. ‘The familiarity of circuit engineers with linear systems has given rise to many useful insights and design rules for circuits that can be approximated as linear. A circuit engineer is able to derive closed-form expressions for characteristics of a linearized circuit, which he can interpret and use afterwards during the synthesis of the circuit. If the circuit or its simplified schematic is too large to analyze with hand calculations, he can resort to a symbolic analyzer for linear(ized) circuits, but he still remains able to some extent to reason about the linear circuit’s behavior even without explicitly having expressions for the characteristics. On the other hand, the analysis and synthesis of circuits in which nonlinearities play a role, is difficult. Indeed, for such circuits just a few design rules exist in the analog design world. This has several reasons First, circuit designers are trained to reason only about linear systems, but not about nonlinear ones. Secondly, it is not easy to analyze nonlinear effects by hand calculations, The most studious designers use Taylor series, but this approach is only feasible in small circuits at low frequencies, with a very small number of nonlinearities. Usually, nonlinear effects are analyzed with tedious time-domain simulations (so-called transient simulations), followed by a Fourier analysis. Other approaches such as the harmonic balance techniques, although very useful, are not (yet) universally used. With both approaches the simulation results are numbers. These numbers can be plotted onto a graph, which can give valuable information, but they do not indicate the fundamental circuit parameters that determine the observed performance. As a result, circuit designers often do not know in which way a circuit can be improved in order to meet the specifications related to nonlinear behavior. The above problem could be relieved if it were possible to indicate to circuit designers which circuit elements or which effects are mainly responsible for the observed nonlinear behavior. Such insight will be offered in this book for several building blocks of analog integrated circuits. In addition some general concepts of nonlinear behavior of analog integrated circuits will be studied as well. This book is intended as a guide to learn designers of analog integrated circuits to reason about nonlinear phenomena in weakly nonlinear, continuous-time analog ICs, The required background to read this book is an understanding of analog integrated circuit design. A prior knowledge about theory of nonlinear systems, such as the theory of Volterra series, is not re- quired. The background of Volterra series that is required to understand some essential concepts, is contained in this book. When browsing through this book, the reader will notice the large number of formulas in this book. When one wants to reason about nonlinear phenomena, then a minimum amount of mathematics is required. It is virtually impossible to write a comprehensible text on nonlinear effects without mathematics. However, no special advanced mathematical techniques are used in this book. Also, the mathematics are explained as clearly as possible, they are interpreted and illustrated with examples, and tedious derivatives are moved to appendices at the end of the book. This book is devoted to the analysis in the frequency domain of weakly nonlinear circuits. The circuits that are addressed are building blocks of analog integrated circuits, both in bipolarand CMOS technologies. The emphasis is on getting insight, both in the nonlinearities in the transistors themselves, as in the nonlinear behavior of transistor circuits. The outline of the text is as follows: - Chapter | presents an overview of the approach that is followed in this book. Further, some assumptions are made about the nonlinear circuits that will be analyzed in this book, and the scope of the book will be outlined. - In the analog design community, many definitions and keywords are used to characterize the nonlinear behavior of analog circuits in the frequency domain. This basic terminology is described in Chapter 2. ~ In order to analyze the nonlinear behavior of a circuit, we need to describe the different nonlinearities in the circuit under consideration with a sufficient accuracy. To this pur- pose, we will make use of power series expansions of the model equations that describe a nonlinear device. It will turn out that a device such as a transistor consists of several basic nonlinear elements, such as nonlinear conductances, transconductances and capac- itors. Each of these elements can be described with a power series. The coefficients of these power series are proportional to the derivatives of the model equations that describe these basic nonlinear elements. These power series coefficients, further in the text referred to as nonlinearity coefficients determine the nonlinear behavior behavior of a circuit. In Chapter 3, these power series are presented together with some simple examples. - A very useful technique to describe the nonlinear behavior of weakly nonlinear circuits is the Volterra series approach, which is covered in Chapter 4. With Volterra series, it is possible to take into account frequency effects into the calculations of harmonics and intermodulation products. This is absolutely necessary if one wants to study circuits with capacitors, both linear and nonlinear. Further, we will use Volterra series to study some general concepts in nonlinear circuits: the use of feedback, both linear and nonlinear, the exploitation of symmetry for the suppression of even-order or odd-order harmonics, the effect of cascading nonlinear circuits, and pre- and post-distortion will be studied. - Calculation methods for harmonies and intermodulation products are discussed in Chap- ter 5. The emphasis is on methods that allow to generate closed-form expressions for harmonics or intermodulation products. Numerical methods will be discussed briefly. For the generation of closed-form expressions a calculation method can be used that makes use of Volterra series. This method is explained with an example. Derivations can be found in literature [Buss 74, Chua 79b]. Further, an alternative method is developed in this, chapter that yields the same results without making use of Volterra series. With both methods the circuit is analyzed in the frequency domain. In fact, we perform an AC analysis on a circuit in which the nonlinear elements are not only represented by their linearized equivalent: in addition, the second- and third-order nonlinear behavior are taken into account as well.Despite the availability of different methods to obtain closed-form expressions for har- monics or intermodulation products, the use of these methods for hand calculations is too tedious, even for very small circuits. With a symbolic network analysis program, however, it is possible to automate these hand calculations and to obtain a closed-form expression. Symbolic network analysis programs can compute a closed-form expression for the AC be- havior of a linearized circuit as a function of the symbolic small-signal parameters of the circuit and of the complex frequency variable. Modern symbolic analysis programs, such as the program ISAAC [Giel 89, Giel 91}, can even generate approximate symbolic ex- pressions. The approximation is made because the exact expression is usually too lengthy, such that it cannot be easily interpreted. An approximation based upon numerical values for the circuit parameters can retain the few dominant terms of an expression, such that the resulting expression becomes interpretable. In Chapter 5 an extension of the program ISAAC is described for the generation of approximate, interpretable expressions for non- linear distortion - The nonlinearity coefficients of the different nonlinearities in a bipolar transistor and a MOS transistor are discussed in Chapters 6 and 7, respectively. Whereas for the bipolar transistor the Gummel-Poon model is still widely used or it forms the basis of more recent models [Mc And 95], the situation for a MOS transistor is more complicated. Due to the rapid scaling of a MOS transistor, effects that were recognized as second-order effects in older technologies, become dominant in modern devices. If these effects are not included in a transistor model or if they are badly modeled, then this may lead to very large errors on the harmonics or intermodulation products. The reason is that the nonlinearity coefficients are proportional to higher-order derivatives. The error on these derivatives tends to increase dramatically when the model equation is inaccurate. In Chapter 7 some shortcomings of widely used transistor models will be discussed. Further, a model for the drain current will be presented that will be used to derive closed-form expressions for the nonlinearity coefficients. - Apart from the examples that have been used throughout the individual chapters, some more applications are described in Chapter 8. Distortion will be analyzed for the follow- ing circuits: a single-transistor amplifier, both a bipolar and a MOS version, a bipolar and a MOS differential pair, a source follower, an emitter follower, a bipolar transistor with emitter degeneration, a common-base bipolar and a common-gate MOS transistor, a bipolar and a MOS current mirror, a Miller-compensated operational amplifier, a bipolar double-balanced mixer, and a CMOS upconverter. Hereby, use will be made of the extension of the program ISAAC to generate closed-form expressions for distortion. - Instead of computing the nonlinearity coefficients, it is also possible to measure these coefficients. A measurement procedure and measurement results on bipolar transistors are given in Chapter 9. Finally the authors would like to thank the following persons for their discussions and com- ments; Peter Kinget from Lucent Technologies, Murray Hill, Yannis Tsividis from the Univer- sity of Columbia, New York, Stéphane Donnay, Hugo De Man and Luc Dupas from IMEC,ix Leuven, Georges Gielen, Walter Daems and Joos Vandewalle from the ESAT Laboratories of the Catholic University of Leuven, Petr Dobrovolny from the University of Brno, Czech Re- public, Guang-Ming Yin and Frederic Stubbe from Rockwell Semiconductor Systems, Newport Beach, California, Frank Op ’t Eynde from Alcatel Microelectronics, Brussels, Paco Ferndndez from IMSE-CNM, Seville, Spain, Jan Vanthienen, and, last but not least, Kat Frangois for her support and patience, Piet Wambacq Willy SansenTO OUR FAMILIES Kaat, Lien, and Elli Hadewych, Katrien, Sara, and MarjanList of symbols and abbreviations ‘The SPICE model parameters are not included. They can be found in [Hspi 96, Anto 88]. arg(z) BIT multiplication symbol in symbolic expre: sions. Normally omitted, only used for clar- ity. multiplication symbol in numbers, for exam- ple 1.5x107", the complex conjugate of the complex num- ber a. transistor “beta”: maximum value of ratio be- tween the collector current and the base cur- rent of a bipolar transistor. body-effect coefficient of a MOS transistor. 3.4531x10-4-F/m, dielectric permittivity of SiO>. 1.0359x10-"F/m, dielectric permittivity of SiOo channel-length modulation factor (MOS tran- sistor). surface mobility (MOS transistor). surface inversion potential (MOS transistor). mobility-reduction coefficient (MOS transis- tor). phase of the complex number z. bipolar junction transistor. symbol for a capacitor. base-collector capacitance (bipolar transis- tor). base-emitter capacitance (bipolar transistor).Csp Che Coz CMRR det(s) F(z), BIS aE GBW Gm. Gmb go on Hy HD», HDs Hy(Gurry jas +++ jen) Ful» 725+ » Tm) collector-substrate capacitance (bipolar tran- sistor), bulk-drain capacitance (MOS transistor). gate-bulk capacitance (MOS transistor). gate-drain capacitance (MOS transistor). gate-source capacitance (MOS transistor). bulk-source capacitance (MOS transistor). MOS gate oxide capacitance per unit area. total MOS gate oxide capacitance. common-mode rejection ratio. determinant of the admittance matrix of a lint ear network as a function of the complex fy, quency variable s. i critical electric field (MOS transistor). average normal electric field that is experi enced by carriers in the inversion layer (MOS transistor). two notations for the derivative of f with re spect to x. symbols that are used for conductances. gain-bandwidth product. transistor transconductance (MOS and bip@ lar transistor), bulk transconductance (MOS transistor). transistor output conductance (MOS bipolar transistor). incremental base-emitter conductance (bit lar transistor). nth-order Volterra operator. second- and third-order harmonic distort n-dimensional Fourier transform of the a! order Volterra kernel, also denoted as order transfer function, nth-order Volterra kernel (time-domain sentation).iour Tour tout Tout ip, iy, Ips Tp ig, te, Ie. Te tp, tas Ip, Ta tpsar, tdsat Ipsars Tésat Te IM, IMy IMFDR IP on, IP 3p, IP oi, IP 34 Isp total current through a component (time do- main). DC component of the current through a com- ponent. AC component of the current through a com- ponent (time domain). Hence igyr = Iour+ tout phasor of the current through a component in the steady state (sinusoidal excitation). total value (time domain), AC value, DC value and phasor respectively of the base cur- rent of a bipolar transistor. total value (time domain), AC value, DC value and phasor respectively of the collector current of a bipolar transistor. total value (time domain), AC value, DC value and phasor respectively of the drain cur- rent of a MOS transistor in the triode region. total value (time domain), AC value, DC value and phasor respectively of the drain cur- rent of a MOS transistor in saturation. forward knee current (bipolar transistor). second- and third-order intermodulation dis- tortion, intermodulation-free dynamic range. second- and third-order intercept point for harmonics. second- and third-order intercept point for in- termodulation products. saturation current (bipolar transistor). base-emitter leakage saturation current (bipo- Jar transistor). voi 1,38062x10-" J/K,, Boltzmann’s constant.xiv Kony ™ jor te(m—Jaz Bemjostgateim—i-H)es nth-order nonlinearity coefficient in power series expansion of the function th describes the nonlinear relationship betw. the current and the controlling voltage for ei ther a nonlinear conductance, transcond tance or capacitor. The symbol « represent the linearized equivalent of the nonlinear el ment. normalized nonlinearity coefficient: Kn, vided by x. mth-order nonlinearity coefficient in the t dimensional power series expansion of function that describes the nonlinear relatic ship between the current and the cont Jing voltages u and v for a two-dimensior transconductance. The symbols 9; and go resent the coefficients in the power series the first-order terms in u and v, respectivel If the nonlinear relationship is expressed af i= f(u,0), then King. em sigs 18 Biven by 2 anf) 11 audv™=5 "F (m— jh If j or (m — j) are equal to one, then they are usually omitted, like in K2,,.,,+ mth-order nonlinearity coefficient in tht three-dimensional power series expansion 0 the function that describes the nonlinear re lationship between the current and the con trolling voltages u,v and w for a threg dimensional transconductance. The symbol 15 92 and gg represent the coefficients in tl power series of the first-order terms in a and w, respectively. If the nonlinear tionship is expressed as i = f(u,v, w), the Kmukoxlm=j-bins 8 BINED BY * an f(ujv,w) 1 1 1 Suidvkaw™ IF | 7 km —j — B)! Note that, if j, k or (m — j — k) are equ to one, then they are usually omitted, like | K. : Sor&onkon effective channel length of a MOS transisto:ne ne nm Para» P_sap g Qp Qe0 Q(z) T.R Ra,te Reco, "Bex Roi. TB To TF i, soutput xv acceptor concentration in the bulk region of a MOS transistor. base-emitter emission coefficient (bipolar transistor). forward emission coefficient (bipolar transis- tor). intrinsic carrier concentration (1.45x10!cm~ at room temperature). 1dB or 3dB compression point. 1.6022 C, elementary charge. majority charge in the neutral base region (bipolar transistor). majority charge in the neutral base region at vac = OV. inversion layer charge per unit area at the po- sition © in the channel of a MOS transistor (0 (44) (2.18) no THD = ‘The value of the total harmonic distortion is often specified in dB or in %. Equation (2.18) reveals that under low-distortion conditions the total harmonic distortion increases as the input! amplitude increases. 1 Whereas total harmonic distortion is very often not the most relevant parameter to charac- terize the nonlinear behavior of a circuit in the application for which the circuit is designed, the total harmonic distortion of circuits is very often specified. This is then considered as a measure’ of “how nonlinear a circuit is at a given input amplitude” and is often used to compare different’ circuits for a given application. Examples of THD The general definition of total harmonic distortion given in equation (2.17) is illustrated with the computation of THD for the square wave shown in Figure 2.3. This signal could correspond to the output of a circuit that is driven by a large sinusoidal input signal such that the output stage switches between the positive and the negative power supply. This corresponds to strongly nonlinear way. In order to compute THD, we start from the Fourier series of this square-wave signal: set (sine + Foin@t) + Esin(5t) +. ) (2.19)2.2 Single-frequency excitation 15 Figure 2.3: A square-wave signal. The amplitudes of the different harmonics in this equation can be substituted in the definition of THD fromequation (2.17). This yields (2.20) It is seen that in this example THD is independent of the input amplitude. It is a very high value compared to the total harmonic distortion of so-called low-distortion circuits. For example, in [Smit 94] an operational amplifier is reported that produces harmonic distortion that is 954B Jower than the wanted signal, when the amplifier drives a 1009 load and the input voltage has a peak-to-peak value of 1V) As another example, we compute THD for the triangular waveform of Figure 2.4. Figure 2.4: A triangular waveform The Fourier series of the triangular waveform is given by so=3 (saa ~ Fesint3s) + A sings — ) 21)16 Basic terminolog Substituting the amplitudes of the different harmonics into equation (2.17) yields = 7 Loar | Gyn 1 12.1% = ~18.34B od n n=l THD = which is seen to be about 124B (about a factor four) smaller than the total harmonic distort of a square-wave signal. 2.3. Even-order and odd-order distortion only It has already been mentioned above that no even harmonics appear at the output of a balanoy circuit when this is driven by two input signals of equal amplitude and opposite phase. TY absence of even-order harmonics can also be observed in the output waveform as indicated { the following theorem. THEOREM 2.1. If fora signal f(t) without DC component and with a period T = w/(2n),t following condition holds f= F004 oy then no even-order harmonics of w appear in f(t). ‘This theorem can be proven easily by computing the coefficients of the Fourier series of f and taking into account equation (2.23), We have already seen two waveforms for which equation (2.23) holds, namely the sqi waveform of Figure 2.3 and the triangle waveform of Figure 2.4. The Fourier series of these waveforms, equations (2.19) and (2.21), respectively, indeed do not contain any even harmoni¢ ‘A similar theorem holds for the absence of odd-order harmonics: i ‘THEOREM 2.2. If for a signal f(t) without DC component and with a period T = w/(2n), th following condition holds od ‘An example of such waveform is given by the absolute value of sin(w¢). The Fourier seq for this waveform is given by fw=su+4 then no odd-order harmonics of w appear in f(¢). cos(2t) — 13 3.5 7 #() = [sin(t)| = 4 (5 cos(4t) — senso") (24 and it is indeed seen that no odd-order harmonics are present. Notice that there is no compont at the fundamental frequency neither (this is also an odd-order component).2.4 Two-frequency excitation 17 2.4 Two-frequency excitation In much the same way as in the previous section, the test circuit under consideration is now excited with two sinusoids A; cos(wt) and Ay cos(w2t), both applied at the same input port. ‘When A; and Ag are sufficiently low, then the output spectrum contains two signals above the noise floor at the fundamental frequencies w and w, due to the circuit’s linear behavior, Since in a linear circuit the superposition principle is valid, the two excitations do not produce any interfering signal. However, when A; and Az become larger, then, apart from the harmonics of w; and w,, interfering signals grow above the noise floor at the frequencies w + wa, lay ~ wel, Quy + we, [2wW, — we], wy + Zw and | —w + 2w9|. The signals at |w; + w| are caused by second- order nonlinear behavior and are called second-order intermodulation products. They increase with the first power of both A, and Ap. The other signals originate from third-order nonlinear behavior and are denoted as third-order intermodulation products. The signals at |2u1 + we| increase with the square of A, and with the first power of Ap. The signals at Jw, + 2ws| increase proportionally to A, and with the square of A», Just as in the case of the single-tone excitation, at higher input amplitudes the dependence of the intermodulation products on the input amplitudes can deviate from the dependence described in the previous paragraph. For example, when A, is high, then the second-order intermodulation product at the frequency |, -+ ws| no longer increases proportionally with Ay. This again results in compression or expansion just as we had in the case of the single-tone excitation. ‘The different frequency components that can be observed at the output of the test circuit that is driven by two sinusoidal signals, are shown in Figure 2.5. It is assumed that the input signals are small enough such that the circuit behaves in a weakly nonlinear way. In order not to overload the figure, components caused by nonlinear behavior of order higher than three has not been shown. The frequencies 20 —wa| and Ju — 2w2], where third-order intermodulation products occur, can be quite close to the fundamental frequencies, which are usually the frequencies of interest. As a result, the intermodulation products can cause severe interference. This is a problem especially in communication circuits. wy tu Le 202 Ie frequency Figure 2.5: The different frequencies at the output of a weakly nonlinear analog circuit excited by two sinusoidal signals at frequencies w, and w,. The numbers between brackets indicate the order of nonlinearity by which the signal is determined.18 Basic terminology In order to determine which frequency components are generated by nth-order nonlinear behavior, the following rule of thumb can be followed. Assume that k different frequencies wy, 9, ..- ye are applied to the nonlinear circuit. Consider the array of frequencies Wry, “Wi, Wey “Way oes Why “Wie When from this array n frequencies are selected whereby a frequency may be chosen more than, once, then the sum of these selected frequencies yields a frequency which can have a compo- nent caused by nth-order behavior. For example, assume that two sinusoidal input signals are applied to a circuit with frequencies w, and wy. In order to know which frequency compo- nents are generated by third-order nonlinear behavior, we select three frequencies from the array wi, “Ww, We, —w2. Doing so, we find that third-order nonlinear behavior gives rise to signals at the following frequencies: tw, tw, $3uy, b3u, tw, $2, +2w we ‘The responses at negative frequencies are the complex conjugate of the responses at the corre: sponding positive frequencies, such that the sum of the two signals is real. For analog circuits such as amplifiers, the intermodulation products are usually unwanted Therefore, they are denoted as intermodulation distortion. In communication circuits, these un wanted products are often denoted as spurious responses (Wein 80]. When the input amplitudes y ‘Ay and Ay are taken equal, the second-order intermodulation distortion IM , and the third-order i intermodulation distortion IM are defined as the ratio of a second- or third-order intermodula-t tion product to the fundamental response. When no frequency effects are taken into account and when low-distortion conditions apply, then it will be shown below that the response at w; + Wa is equal to the response at fr — wa]. Also, the response at 2u + w is the same as the responses; at [201 — wy), 2a» + 21 and [20% — |. Then My and IMs are given by 4 Ve IM, = | 2.26) 2 Yet “ IM, =. |Yont2itt| lm togeris | Vo 2 enn Vout.s0 mit 40 ‘The above definitions can be clarified by considering again the test circuit with the input? output relationship given by equation (2.4). When the input signal 2(t) consists of two signald of equal amplitude and with a different frequency w; and w: a(t) = Acos(wit) + Acos(wet) (2.28) then the different responses are shown in Figure 2.6. It is assumed that the amplitude A is sufficiently low such that the circuit behaves in a weakly nonlinear way. The frequencies w, and wz have been given the value 2rx10MHz and 27x11 MHz. It is seen that harmonics of w; and «us are present at the output as well as intermodulation products. The amplitudes of the different harmonics and intermodulation products can be computed using the formulas of Appendix A. It2.4 Two-frequency excitation 19 MHz 3KA° KA? KA 0 | 0 1 Ka? DC shift 1) @-o Ka? IM, 9 | 20,-0, 1 ‘ | 3K,a® mM, fo | o K,A +2K,A° fundamental Wo | & { K,A +2K,A° fundamental 2 | 20-0, : 2Ka IM, or +608 | 2 | 20, : : 2K,a? HD, A | ato, . . : KA? IM, 2 | 2, : | EKA? HD, 8X or +9.54 dB 30 | 30, —~ ' | £Ka* HD, 31 | 20,40, . : : 3K,a? IM, 2 | a +20, : ‘ : 3K,A® IM, B | 30, 1 | : 1 EKA? HD, Figure 2.6: The different frequency components at the output of a weakly nonlinear circuit with the input-output relationship given by (2.4) to a combination of two sinusoidal signals with the same amplitude A. The frequency of the two input signals is 1MHz and 11 MHz. is seen that indeed the response at w1 +w is equal to the response at |; —w2|. Also, the response at 2, + wo is the same as the responses at |2w1 — w|, 2w2 + w and |2w — wi). From Figure 2.6 we can compute the second- and third-order intermodulation products for20 Basic terminology Comparing these values with the harmonic distortion figures of equation (2.13) and (2.14), we find IMy = 2HD2 (2.31) IM; = 3HDs (2.32) ‘This is an important relationship that is valid under low- (2.3.6) Similarly, we find with equations (2.16) and (2.35) IP sy V3 ‘Ona logarithmic scale the ratio of 1/3 between the two intercept points is found as a difference of -4.7dB. IPs: = (2.37) Intermodulation-free dynamic range Figure 2.7 shows yet another parameter: the inter modulation-free dynamic range IMF DRs. This is the ratio of the largest and the smallest sigmal] level the circuit under consideration can handle, where the intermodulation products are to Te-! main below the noise level. The smallest signal level the circuit can handle is usually also taken equal to the noise level. The parameter IMFDRs is read from the 2-axis (at the input side) in Figure 2.7, but it can also be read from the y-axis (at the output side) as indicated. The reason is that the slope of the curve that represents the response at the fundamental frequency is 45 degrees, on a double logarithmic scale. Desensitization We consider the following experiment: the input signal A, cos(wyt) is the input signal while Ay cos(w,t) is considered as an unwanted signal that is also applied to input. Ina communication circuit the latter signal can be a disturbing signal that is so close to the} wanted signal such that any bandpass filtering before our test circuit cannot reject the disturbing, signal. Assume now that the amplitude Ay is increased gradually. In a linear circuit the amplitude of the wanted response at w, would be unaffected by the presence of the second signal. Howewer,) this is no longer true in a nonlinear circuit. Indeed, when Ap exceeds a critical level, then the} response at .2; decreases. The reason is that the strong unwanted signal drives the input stage’ of the circuit into strongly nonlinear behavior, resulting in a lower gain for the wanted signal.; This effect is called desensitization. Desensitization is important in communication systems) where a small signal of interest (corresponding to A,) is influenced by a strong unwanted signal’ (corresponding to Az). 4 Example The need for low intermodulation distortion in communication circuits can be ilftus- trated with the following example. Assume that the receiver front-end of Figure 1.1 has to2,5 Mixer definitions 3 receive a signal that is modulated in some way with a carrier at frequency f, = 100.1MHz. Sup- pose further thatthe strength of this signal is low, whereas two strong adjacent signals are present at fo = 100.2MHz and at fz = 100.3MHz. The signals at fz and fz are so close to the wanted signal that they cannot be removed by a bandpass filter around 100.1 MHz. Instead, the unwanted signals will be removed at lower frequencies further in the receiver circuitry. Due to third-order intermodulation distortion in the front-end, the two strong signals give rise to a intermodulation product at 2f2 — fs = 100.1.MHz which is exactly the frequency of our wanted signal. If third- order intermodulation distortion in our front-end is very high, then the intermodulation product might have a signal strength that is comparable to our wanted signal. This of course complicates the detection (demodulation) of the wanted signal. . 2.5 Mixer definitions ‘A mixer performs a frequency translation either from a high-frequency input signal to a low- frequency signal or vice versa. A mixer has two inputs: a signal input and a local oscillator input. Very often, these inputs are applied at different input ports. A frequency translation to low frequencies is performed by applying to the mixer a high-frequency signal and a high- frequency local oscillator signal. The output signal of interest is at the frequency that is equal to the difference of the input signal frequency and the local oscillator frequency. This kind of frequency translation is used in receiver circuits. It is referred to as downconversion. In the case of a frequency translation to high frequencies, the input signal is at low frequen- cies, while the frequency of the local oscillator is again high. The output signal of interest is at the frequency which is the sum or difference of the two frequencies. This frequency translation is used in transmitter circuits. It is referred to as upconversion. A deep study of different mixer configurations is beyond the scope of this book. An excellent survey can be found for example in [Gill 96]. At this point it is only important to mention that mixer configurations can be classified in two categories. In the first category, the local oscillator signal is very large such that the mixer behaves in a strongly nonlinear way with respect to the local oscillator signal. The transistors in this circuit are switched on and off at a rate determined by the local oscillator signal. Roughly speaking, this gives rise to a square-wave signal. This square wave is modulated by the input signal, and the resulting waveform contains a frequency component at the sum or difference frequency. This category is denoted as the class of switching mixers. In a second class of mixers, the amplitude of the two input signals is kept fairly low, such that the mixer behaves in a weakly nonlinear way. The output signal at the sum or difference frequency is caused by second-order nonlinear behavior of the mixer. Indeed, in Figure 2.6 it is seen that second-order nonlinear behavior of a circuit that is excited by two signals of a different frequency, gives rise to a response at the sum or difference frequency. However, the situation is ‘more complicated here, since the input signal and the local oscillator signal are usually applied at different input ports. This category of mixers is referred to as weakly nonlinear mixers. In the literature both switching mixers [Mey 94, Sato 96, Madi 96] and weakly nonlinear mixers [Crols 95a, King 97, Borre 97] are found nowadays. In this book only weakly nonlinear24 Basic terminology mixers will be addressed. An analytic treatment of switching mixers is only possible by making simplifications [Mey 86] ‘The wanted signal at the output of a mixer is the second-order intermodulation product at the sum or difference frequency, while both the linear response and intermodulation products of order higher than two are unwanted. Assume that the input signal of a mixer is a sine wave, The conversion gain of a mixer is the ratio of the amplitude of the wanted second-order inter- modulation product at the output to the amplitude of the input signal. For a switching mixer the conversion gain is proportional to the amplitude of the local oscillator signal. Other mixer definitions such as compression point and third-order intercept point are identical to the definitions presented in Section 2.4. 2.6 Cross modulation Cross modulation is the nonlinear effect whereby modulation from one carrier is transferred to another. Assume that the input 2(t) to a nonlinear circuit is an amplitude-modulated signal at w together with an unmodulated carrier with frequency wp: a(t) = A(1+ my cos wnt) coswyt + Acos wat (2.38) In this equation m, is the modulation index and A is the amplitude of both signals. If the cit cuit has a cubic nonlinearity, then a distorted version of the modulation of the signal at w1 is transferred to the carrier at w. The effect of cross-modulation can be understood by computing the response of the test circuit with the input-output relationship of equation (2.4), to the amplitude-modulated signal of equation (2.38). The response can be computed using simple algebra. Among the different terms of the response, the following two are interest: response to AM-signal =... + AK, cos(w2t) + 6A°K3 cos*(w t)my Cos(wmt) cos(wet) +.- (2.39) With the trigonometric relationships of Appendix A we find cos(2uit) 40 ; (2.40) 1 c08"(wt) = 3t In this way, an amplitude-modulated signal is found in the response: response to AM-signal =... + AK; (1+ 12008 wat) cos(w2t) + @4l) in which the modulation index mz is given by 3Kym, A? i (2.42) m=7 Summary 25 The cross-modulation factor CM is the ratio of the two modulation indices mz and m: BK yA? om = ™ = 34 2.43) mK There is a simple relationship between the cross modulation factor and third-order intermodula- tion distortion. Indeed, combining equations (2.30) and (2.43) yields CM Ti; (2.44) ‘At low frequencies, the modulated signal at w» is amplitude-modulated. At high frequencies, however, both amplitude and phase modulation can occur at «, (Mey 72]. In order to provide quantitative measures for cross modulation at high frequencies, a knowledge of Volterra series is required. Therefore this topic will be resumed in Chapter 4. 2.7 Summary In this chapter, several performance parameters have been defined that are used to quantify the nonlinear behavior of a circuit. Most definitions are valid both for weakly and strongly nonlinear behavior. When the scope of the defined parameters is limited to weakly nonlinear circuits, then at low frequencies these parameters can be expressed in terms of first-order, second-order and third-order coefficients that describe the input-output relationship of the circuit, In Chapters 4 and 5 it will be explained how these coefficients can be computed when an explicit expression for the input-output relationship cannot be computed. Also, the above definitions will be extended such that the influence of frequency can be taken into account.Chapter 3 Description of nonlinearities in analog integrated circuits 3.1 Introduction Prior to the analysis of nonlinear behavior of analog circuits, it is necessary to describe the nonlinear devices that are present in analog integrated circuits. The devices most commonly used in silicon analog integrated circuits are transistors, resistors, capacitors and diodes. In the last few years self inductances and mutual inductances have been integrated as well [Ngu 90, Ash 96, Long 95, Long 95}. Their nonlinear behavior is not considered in this book. In circuit analysis the devices mentioned above are described using an equivalent circuit. This equivalent circuit can be as simple as one circuit element (e.g. one resistor), or it consist of several circuit elements (¢.g. a transistor). The elements of such equivalent circuit are nonlinear in general, The following circuit elements are used in numerical simulation of analog integrated circuits as a part of the equivalent circuit of a device: a nonlinear conductance: the current through this element is an algebraic function of the voltage over the element; * a nonlinear transconductance: the current through this element is an algebraic function of a voltage other than the voltage over the element. anonlinear resistance: the voltage over this element is an algebraic function of the current through this element; ‘© a nonlinear transresistance: the voltage over this clement is an algebraic function of a current different from the current through this element; ‘¢ a multidimensional nonlinear conductance or transconductance: the current through this element is a function of more than one voltage: a multidimensional nonlinear resistance or transresistance: the voltage across this element is a function of more than one current;3.1 Introduction 27 « anonlinear capacitance: the charge on this element is an algebraic function of the voltage across the element; # a nonlinear transcapacitance: the charge on this element is an algebraic function of a volt- age other than the voltage across the element; These circuit elements are referred to as basic nonlinearities, since they are the building elements for nonlinear equivalent circuits of devices such as transistors, diodes, integrated resistors, . Nonlinear voltage-controlled voltage sources and current-controlled current sources are seldom used in equivalent circuits of devices in analog integrated circuits. In this book a nonlinear circuit is analyzed by studying excursions around a quiescent point. ‘This means that every basic nonlinearity undergoes such excursion. A small excursion around the quiescent point of a basic nonlinearity can be described as a power series of the algebraic function that describes the basic nonlinearity. When those excursions are small enough, then this power series can be broken down after the first few terms. For very small excursions one can even do with the first term of every power series: since this term describes the linear behavior of acircuit element, this case corresponds to a linearization of the circuit. In Chapters 4 and 5 it will become clear that a power series description of basic nonlinearities is elegant when harmonics and intermodulation products are studied that are caused by weakly nonlinear behavior. ‘A power series description of a basic nonlinearity contains the derivatives of the output quan- tity (current or voltage) with respect to the controlling quantities. These derivatives are evaluated in the quiescent point. An accurate description of a nonlinear device in terms of power series requires that the different derivatives that are considered! be accurate, These derivatives are a function of the controlling quantities and of the model parameters that describe the nonlinear device. It is clear that a poor model for a nonlinear element — either due to inaccurate model equations or to an inaccurate parameter extraction — can give rise to dramatic errors on the derivatives. This is illustrated in Figure 3.1. This figure shows a possible situation where the measured quantity is approximated by a model using a least-squares criterion, Although the error on the quantity itself might be small, it is clear that the derivative of the fitted curve seriously differs from the real derivative. This problem has been noticed by many researchers in analog design, especially with respect to the modelling of MOS transistors [Tsiv 88, Tsiv 93b, Enz 95, Groen 94, Tsiv 81]. ‘The majority of performance parameters of an analog integrated circuit is related to the small- signal behavior of the circuit. Since the small-signal parameters are first-order derivatives, it is clear that an accurate analysis requires accurate first-order derivatives. Although many MOS models provide reasonable values for the drain current, they can give inaccurate values for some small-signal parameters. Especially the value of the output conductance is often erroneous. For distortion analysis this problem is even more severe: here, derivatives are required of order higher than one, and the deviation between a model and the real behavior tends to increase with the order of the derivative. ‘Usually, one does not consider nonlinear effects of order higher than three or four, hence derivatives of order higher than four do not need to be considered. In most cases one even limits to order three.in analog output quantity input quantity Figure 3.1: Fit of a model equation with a least-squares approximation (solid line) to measuig- ‘ment results (dashed line), Deviations between a model and the reality due to a poor parameter extraction can be mini- mized if in the parameter extraction procedure not only the deviation of the current is minimized but also the deviation on the derivatives, as suggested in (Tsiv 88]. This of course requires that the derivatives can be measured. In [Groen 94] the derivatives are computed by taking differences of the measured current. However, it is possible to directly measure the derivatives as described in [Iked 72} for the first-order derivatives and in Chapter 9 for higher-order derivatives. Of course, efforts in the parameter extraction phase are not useful if the model itself is not adequate. Inaccuracies in the models can have several reasons. First, a model is always a simplifi- cation of the real situation, Models are kept relatively simple for several reasons: a simple model can be quickly evaluated in numerical circuit simulation, such that the computation times remain acceptable. Secondly, a simple model yields insight in the behavior of the device. Transistors, for example, are usually modeled in such a way that the current flows into one direction and that the electric fields in the transistors are one-dimensional. Due to the scaling of the devices, this assumption is more and more violated, and two-dimensional or three-dimensional effects be~ come important, Also, equivalent circuits of devices are composed of lumped elements, whereas in reality a device is composed of distributed elements such as RC-transmission lines. An ac- curate simulation can only be performed by three-dimensional computer simulations. Since this is not efficient, much effort has been done in the world of device modeling to make acceptable simplifications by using empirical or semi-empirical approaches. In these approaches the two- or three-dimensional effects are broken down into simple, separate effects examined one at a time. Some simplifying assumptions are then made, which are sometimes difficult to justify, and relatively simple relationships are derived. The danger exists that oversimplification yields model equations that again describe the output quantity of the device (either a current or a volt- age) relatively accurate, while the inaccuracies on the derivatives are unacceptable. This danger is especially present in the modelling of MOS transistors, as will be discussed in Chapter 7. For-3,2 Power series description of basic nonlinearities 29 tunately, only the lowest-order derivatives are required in the analyses carried out in this book and with some corrections on widely used models, insight can still be obtained in combination with a good agreement between calculations and measurements. This will also be illustrated in Chapters 9 and 8 Devices with different operating regions are sometimes modeled with different model equa- tions. Examples are the different model equations for the weak and strong inversion region for 4 MOS transistor that are used in many transistor models. The models ate usually constructed in such a way that there is no “jump” in the output quantity (in this case the drain current) when switching from one model equation to another. Unfortunately, the derivatives often exhibit a di continuity at the transition point. This can cause problems in the computation of the DC solution of a circuit or in time-domain simulations, and it will also give rise to erroneous values of the small-signal parameters in the vicinity of the transition point. Higher-order derivatives, of course suffer from the same problem. This problem is not solved in this book. It is circumvented by assuming that the bias point of a device is far enough from the transition point and that the signal swings are small enough such that the device remains in the same operating region. Of course, this problem does not occur for devices that are described with one model equation, such as the bipolar transistor, For MOS transistors, recent research efforts have resulted in the modelling of the MOS current with one equation as well [BSIM 95, Foty 96]. With the power series approach the DC solution is split from the AC part, Since it has already been pointed out that many device models do a decent job for the computation of the DC solution whereas they are inaccurate for the derivatives, one can use such model for the DC solution and a separate model for the AC solution. This has also been suggested in (Tsiv 83, Tsiv 88] This chapter is organized as follows: in Section 3.2 the basic nonlinearities are described in terms of power series. These basic nonlinearities are the composing elements of equivalent circuits of devices such as diodes and transistors. Next, the nonlinearity of integrated resistors and capacitors is discussed in Section 3.2.2 and 3.2.3, respectively. The basic nonlinearities can be used to build transistor models, as will be mentioned in Section 3.5. Nevertheless, a detailed discussion of the different nonlinearities in bipolar and MOS transistors is quite involved. This is discussed in separate chapters, namely Chapters 6 and 7. The diode is not discussed as a separate device: its nonlinear behavior can be extracted from Chapter 6, 3.2 Power series description of basic nonlinearities In the introduction an enumeration has been given of the different nonlinear circuit elements that can make part of the equivalent circuit of a device. For most applications, one can do with & more restrictive sct of circuit elements: a nonlinear resistance, a nonlinear conductance, a Nonlinear transconductance, a nonlinear capacitance, and a nonlincar current source depending upon several controlling voltages, often indicated as a multidimensional transconductance”. They are depicted in Figure 3.2. These nonlinearities are denoted as the basic nonlinearities. Theit general description in terms of power series is given below. These basic nonlinearities are the Although the term transconductance is used, one of the controlling voltages can be the voltage over the element itself,30 Description of nonlinearities in analog integrated circuits v= f(a) ( fi (eu = f(v) (+ i= flv) 5 ij (a) (b) Co) (: ( = f(u,2) 2 i= sf) aa } (d) () Figure 3.2: The basic nonlinearities used for the analysis of the weakly nonlinear behavior of analog integrated circuits: (a) a nonlinear resistance, (b) a nonlinear conductance, (c) a nonlin= ear transconductance, (d) a multidimensional transconductance and (e) a nonlinear capacitance. building blocks of nonlinear models of devices such as transistors, as described in the subsequent sections. Multidimensional transconductances are seldom used in distortion calculations found in lit- erature on computations of distortion in analog integrated circuits. The reason is that most com- putations found in literature are simplified that much such that only one AC voltage that controls a multidimensional transconductance, differs from zero. For more complicated computations, however, one must allow that all controlling voltages of a multidimensional transconductance are different from zero. This can only be described accurately with multidimensional transcon- ductances. It will be seen in Sections 3.2.4 and 3.2.5 that in the power series description of multidimensional transconductance terms arise that are due to the simultaneous presence of two or three controlling voltages (in AC). These terms, in fact cross-terms, do not occur in purely linear circuits, and they cannot be obtained by describing a multidimensional transconductance as the sum of one-dimensional (trans)conductances, ‘A description of nonlinear transcapacitances is omitted in this book. Their description is similar to the description of nonlinear capacitances. Transcapacitances are used for example in the high-frequency modelling of MOS transistors [Tsiv 88] In the modelling of devices that occur in analog integrated circuits, the nonlinear elements without memory (no capacitor or inductor) are most often described as yoltage-controlled ele- ments, Current-controlled elements occur quite seldom. An example is a nonlinear resistance that suffers from current crowding, such as the base resistance of a bipolar transistor. In most cases, however, it is possible to construct a power series in terms of voltages for a circuit element that is originally described as a resistance. Conversion formulas are given in Section 3.2.2.3,2 Power series description of basic nonlineari Bie 3.2.1 Nonlinear conductance and transconductance For a nonlinear conductance or transconductance, the current through the element, ioyr(t), is ‘a nonlinear function f; of the controlling voltage vcowre(t). For a conductance this is the voltage over the element itself, whereas for a transconductance this is a voltage elsewhere in the circuit. This function can be expanded into a power series around the quiescent point Ioyr = fa(Veowrr)? tour(t) = f(veowre(t)) = F(Voonre + Voor (t)) =f( Ubonie(t) GB) In this equation igyr({) is the total value of the current, which is the sum of the DC and the AC current. The voltage Uggair(t) is the AC voltage that controls the conductance. The second term in equation (3.1) is a power series representing the AC part of the current. ‘When the analysis of a circuit that contains a nonlinear conductance is limited to first-, second- and third-order nonlinear behavior, then the power series in equation (3.1) can be broken down after the third term. Defining the following coefficients n (3.2) 1 & fw) Kay, = 5 Hoe sevoonne (3.3) - 1 BFW) Bo = 31 Gus je=Veowra G4) in which we omitted the time dependence for simplicity, and in general 1 a Fo) ve a ou" v=VoonTR es) leads to the expression of the AC current through the conductance ioue(t) = 91-Veontr(t) + Kag, Veontr(t) + Kg, Voonte(t) + (3.6) In this expression, g; is the small-signal (trans)conductance of the linearized element. The coef- ficients in the second and third term, Ky, and K’3,, are respectively the second- and third-order Nonlinearity coefficients that describe the nonlinear element. Similarly, the small-signal con- ductance is often referred to as the first-order coefficient. The subscript for K and K is the symbol that represents the linearized element, in this case g). This convention will be followed for the other basic nonlinearities as well.32 Description of nonlinearities in analog integrated circuits Sometimes it is interesting to normalize the second- and third-order nonlinearity coefficients with respect to the first-order coefficient. In that case, Kg, = Koy, / 9 G2) is the second-order normalized nonlinearity coefficient, and ig, = Kag,/ (3.8) is the third-order normalized nonlinearity coefficient. Normalized nonlinearity coefficients are interesting to consider since, according to equation (2.13) and (2.14) they are proportional to the harmonic distortion of the current through the nonlinear conductance as a result of a sinusoidal controlling voltage. The units of the coefficients gs, Kay, and Ky, are A/V, A/V? and A/V®, respectivety. The units for the normalized nonlinearity coefficients K,, and Ks, are V-! and V2, respectively. It is important to note that the polarity of the nonlinear conductance must be taken into: account. For a linear conductance it is the same which of both terminals is the positive node. If in Figure 3.2 for a linear element node i is chosen as the positive node, then the relation between the current and the controlling voltage is exactly the same as if node j were the positive node. On the other hand, if for a nonlinear conductance node i is chosen as the positive node, then the current flows from node i to j, and the current is described as a function 7; of vu, — u;. For node J chosen as the positive node, the current flows from node j to i and is represented by a function iz of vj — vj. Generally, the (nonlinear) relationship between i, and v; — v is not the same as the relation between i, and v; — v). For example, consider a transconductance that is described by the relationship ty = exp (vj — vj) (3.9) If the polarity is reversed, then the transconductance is described as ig = —exp(—(v; —u)) (3.10) # exp(vj— v4) GAD Developing the functions of both equations (3.10) and (3.11) into a power series, reveals that the terms of the odd powers are equal while the ones of the even powers are opposite. If, however, the function that describes the nonlinearity is an odd function, then no even powers are present in the power series expansion and the polarity can be reversed without altering the functional relationship. 3.2.1.1 Example: collector current of a BJT A simplified model of the collector current ic of a bipolar transistor is given by ie =I5 exp(“BE .12) ‘3,2 Power series description of basic nonlinearities 33 in which Is, vpe and V; are the transistor saturation current, the base-emitter voltage and the thermal voltage, respectively. This is a nonlinear transconductance: the current which flows between collector and emitter is controlled by the base-emitter voltage difference, ‘The first derivative of the collector current with respect to the base-emitter voltage is the tran- sistor transconductance g,,. Hence, the nonlinearity coefficients that describe the nonlinearity of the collector current are denoted by K2,,, and Ks,,,. Using the definitions of equations (3.2) through (3.4) in conjunction with equation (3.12) yields Gm (3.13) Gm Koy, m 3.14) Im Kg, ev2 (3.15) and in general In a 3.16) Koon = TVR VT (3.16) The normalized nonlinearity coefficients are then given by 1 1 Koon = et @.17) For a quiescent current of ImA and with V, = 0.0258V, one obtains gm, = 0.0387A/V, Ky, = 0.751A/V? and K3,,, = 9.70A/V%. The second- and third-order normalized non- linearity coefficients are found to be K}, = 19.4V—! and K5,,, = 250.6V 2 Z 3.2.1.2 Example: base current of a BJT The base current of a bipolar transistor is an example of a one-dimensional conductance. The first-order model of the base current is given by (3.18) in which Gp, the transistor beta [Lak 94] is considered as a constant. Comparing this simple Model to the simple model of the collector current (equation (3.12)) itis seen that the two currents differ by a constant. Hence, the shape of the two nonlinearities is identical. The first-order derivative of the base current with respect to the base-emitter voltage is the inverse of the small-signal resistance r,, denoted as g,. In this way, the nonlinearity coefficients44 Description of nonlinearities in analog integrated circuits are (3.19) (3.20) G21) (3.22), Koy. = Gua oP 623) wave ~ 24) and in general : Kuge = Gita = ET 25), = a= (3.26) 1 More advanced descriptions of the collector and base current of a bipolar transistor will be presented in Section 3.5. 3.2.2. Nonlinear resistance A nonlinear resistance is a current-controlled element: the voltage over this element is a nonlinear function f of the current through the element. This function can again be expanded into a power series around the quiescent point: Vour = f(Iconrr) vour(t) = fliconralt)) = fTconrn + veonr(t)) + Hbontr(t) (3.27) i=loonre =f(Icowrr) + ‘The second term in equation (3.27) is a power series representing the AC part of the voltage over the clement. Since this voltage is the output quantity, it has the subscript “out”, whereas the controlling quantity is the current, which has a subscript “cont When the analysis of a circuit that contains a nonlinear resistance is limited to first-, second- and third-order nonlinear behavior, then the power series in equation (3.27) can be broken down3.2 Power series description of basic nonlinearities 35 after the third term. Defining the following coefficients (3.28) (3.29) (3.30) and in general 1 aslv) nm = a Bae sounm (3.31) leads to the expression of the AC voltage Vout(t) = 11 iconte (t) ++ Kop “Contr (t) + Keay ieoneelt) + ++ 3.32) In this expression, ionie(t) denotes the AC value of the controlling current. For a resistance, this is the current over the element, while for a transresistance, this is an AC current elsewhere in the circuit. The coefficient of the first term in the power series above, r), is the small-signal (trans)resistance of the linearized element. The coefficients in the second and third term, Ko, and Ks,, are respectively the second- and third-order nonlinearity coefficients that describe the nonlinear resistance. Similarly, the small-signal (trans)resistance is often referred to as the first-order nonlinearity coefficient. The units of the coefficients r1, Kz and K's are 2, 2/A and Q/A?, respectively. Sometimes it is interesting to normalize the second- and third-order nonlinearity coefficients with respect to the first-order coefficient. In that case, Kj, = Ky,, /1 is the second-order nor- malized nonlinearity coefficient, and Ks, = Ks,, /r; the third-order normalized nonlinearity coefficient. The units for the second- and third-order normalized nonlinearity coefficients are Av! and A ®, respectively. 3.2.2.1 Conversion formulas between a voltage-controlled description and a current-controlled description Many practical circuit elements can be described both as voltage-controlled and as current- Controlled elements. Usually, the physical origin of a nonlinearity gives rise to one of the two options: for example, the value of the base resistance changes with the base current due to current crowding. This appeals for a description of the base resistance as a current-controlled element. Nevertheless, in many situations, a description in terms of a controlling voltage is mathemat- ically correct as well, and will lead to the same results. Indeed, if, for example, a nonlinear Conductance is described as i= f(v) (3.33)36 Description of nonlinearities in analog integrated circuits} 1 then the element can be described as well as ves") (3.34) if at least the inverse function f~! exists. In practice, one of the two descriptions (3.33) or (3.34)? is an explicit relationship derived from physical considerations. The other relationship, on the other hand, cannot always be written explicitly. Nevertheless, as will be shown below, it i possible under certain conditions to find the derivatives even without having to determine thi relationship explicitly. These derivatives can then be used in the power series description of th nonlinear element. The determination of the first-order derivative can be performed using th following theorem. In this theorem we denote the derivative of a function with a prime. THEOREM 3.1. Let the function f be a continuous function over an interval (1, y] such th the function value for every element inside this interval is unique. If f can be differentiated in f7'(a) for a € [x,y] and f’ (f-1(a)) # 0, then f~! can be differentiated ina and =~ 1 (PY = 79=R—H (3.35) This theorem establishes a relationship between the derivative of the inverse function and the derivative of the function itself. This theorem will be useful for the derivation of the power seri coefficients of the inverse function. If the conditions stated in Theorem 3.1 are not only fulfilled for f but also for ’ and f", which is the second derivative, then one can easily find “ayy PME Ma) (OO) = CTA O78 0 = ae (-7 "(Ma) + 8 ayy ) ox These expressions can now be used to compute the nonlinearity coefficients of a voltage- controlled description from the nonlinearity coefficients of a current-controlled description. As-: sume that a current-controlled description is given by Vs tae t+ Kon P+ Ky B+ (3.38) while the equivalent voltage-controlled description is given by Gac V+ Koy, 1? + Kay, P+. (3.39) then, using Theorem 3.1 and equations (3.36) and (3.37), one obtains jue = 3.40) - Ban : Kay. = = G.41) K3go¢ = (- (pe + 2 (3.42)3.2 Power series description of basic nonlinearities 37 and, conversely, Tac = 1 (3.43) w= 43) - Kane Ko, = a (3.44) 7 1 (Kaya)? Knee = (-*. +2 (3.45) ‘These conversion formulas will be used for example in Chapter 6 for computations which involve the nonlinearity of the base resistance. 3.2.2.2 Difference between DC and AC resistance For a nonlinear resistance it is interesting to note that the DC value is different from the AC value. This is illustrated for a nonlinear resistor in Figure 3.3, which shows the voltage over a nonlinear resistor as a function of the current through the resistor. The DC and the AC resistance Figure 3.3: Voltage as a function of current for a nonlinear resistor. The DC and AC resistances are given by the indicated slopes. ‘te indicated on the figure as well. The DC resistance is the ratio of the voltage over the resistor nd the current that flows through the element. When this voltage is evaluated as a function of. current, then the slope — or derivative — of this curve is the AC resistance. The second- and third-order derivatives of this curve are equal to 2!K»,,., and 31K's,,,+ Fespectively. It i ‘Important to note that for a linear resistor there is no difference between the AC and DC value.38 Description of nonlinearities in analog integrated circuits Indeed, for a linear resistor, the curve that represents the voltage as a function of current is a straight line. Hence, the ratio of the voltage and the current is the same as the slope of this curve, Ina similar way as for resistances, a difference can be made between the DC value and the AC value of a nonlinear conductance. 3.2.3. Nonlinear capacitance Capacitors in analog integrated circuits can be parasitic or they can be wanted components. In the first case, they occur for example in the equivalent circuit of a transistor. Integrated capacitors, on the other hand, are wanted components. Both kinds exhibit nonlinear behavior. For a nonlinear capacitor the charge is considered as the controlled quantity, and the voltagd over the capacitor is the controlling quantity. The nonlinear relationship between the chargé dour and the voltage vconr over the capacitor is given by gour(t) = f(vcowralt)) = f(Veonre + Veoner(t)) wy pe ' = f(Vconrn) + > ce af rhe homag tant (3.46) 1 ‘The power series in equation (3.46) represents the AC part of the charge upon the capacitor, Defining the first, second- and third-order nonlinearity coefficients as = 20) G47; Ov |. -veowse - 1 fv) Kae, = Re -Veovne (3.48) 1 A f(v) . ei 0" lo=voowre 64 the AC part of the capacitor charge is written as Gout = C1 Veontr(t) + Kag, Veonie(t) + Kae, Yronte(t) + (3.50) The first-order coefficient C; represents the small-signal value of the linearized element. The| units for Cy, Kae, and Ks, are respectively F, F/V and F/V?. Sometimes it is interesting to normalize the second- and third-order nonlinearity coefficients| with respect to the first-order coefficient. In that case, Kj, = Ka, /C1 is the second-order nor-: malized nonlinearity coefficient, and Kt, = Kye, /C, ‘the third-order normalized coefficient of the nonlinear capacitor. The units for the second: and third-order normalized nonlinearity coefficients are V~? and V~, respectively. The AC current ipu:(t) through a capacitor is found by taking the time-derivative of the ad charge upon the capacitor: 1 d ina t) = 5 (Cr-tamir(t) + Kae, Veour(t) + Bay Lner(t) +++) sty3.2 Power series description of basic nonlinearities 39 3.2.3.1 Example: diffusion capacitance ‘As an example, consider the simplified model of the diffusion capacitance between the base and ‘emitter of a bipolar transistor. The nonlinear relationship between the capacitor charge Q and the voltage vr over the clement is given by [Getr 76, Anto 88, Lak 94]: vee ve Qp = tric = Tr Is expl (3.52) in which 7p is the transit time, ic is the collector current, Is is the saturation current of the transistor and Vj is the thermal voltage. Noting the first derivative of the charge with respect to the controlling voltage as C’p, one obtains the following nonlinearity coefficients: = = Te Om (3.53) TE Im nyt (3.54) For a collector current of Im and a forward transit time of 50ps, the following values are obtained at T=300K: Cp = 1.94pF, Ko,,, = 37.6pF/V and Ka, = 485pF/V*. . 3.2.4 Two-dimensional transconductance A two-dimensional transconductance is an element the current of which is controlled by two different voltages. In other words, the current ioyr(t) is a function f of two voltages ucowrr and vcowres which can be expressed in terms of AC values using a two-dimensional power series expansion around the quiescent point Jour = f(Ucowrr, Veowrr) tour(t) =f(uconralt),vcowrelt)) = f(Ucowre + tecontr(t)s Voorn + Yeontr(t)) =f (Uconrr, Veonrr) + 2 | an flue) LX | arave mal n=O ) _ Woniet) | Ueontn(t) u=Uconra ‘~ nl" (=n)! v= Veonrr (3.55) The AC part of the current corresponds to the second term of equation (3.55), which is a two- dimensional power series. This series can be split into three series iy, ig and ig, each correspond- ing to a part of the total AC current. The first two series, i; and iy contain powers of one single Voltage and so they are similar to the series described in Section 3.2.1: iy = gu eontr + Kg, Cénie + Kg, “Moonee Fo (3.56) and ig = G2-Veontr + Kagy Veonte + aga” (3.57)40 Description of nonlinearities in analog integrated circuit Ae The time dependence of teonur and Veonur has been omitted for simplicity. The third series, ig, contains nothing but cross-terms, which are terms that contain a nonzero power of both tents and Ucontr? is = Ko, uy, ow, . 4 . ve, rpitegy * Weontr “ Veonir + Kany, gegy ° Weontr * Veontr + K3,, png, “ Uoontr * Ueontr (3.58) ‘The meaning of the subscripts in the nonlinearity coefficients defined above is as follows. Sup- pose that the first-order derivative of the total current with respect to u and v are respectively gy’ and go. Then a coefficient KE Kinjy, gimjyq, With ™ and j positive integers and rm > j means a fluyv) Ld Kmiq,nn-ine = Dusdu™I | Fl j™m-a (3.59) If j or (m ~ j) are equal to one, then they are usually omitted as a subscript, like in Ke, Note that the derivatives are evaluated for u = Ucowre and v = Veowrr. Just as with the above basic nonlinearities, one can again consider normalized nonlinearity coefficients by taking the appropriate ratios. 3.2.4.1 Example: two-dimensional collector current In order to clarify the above definitions, consider the simple equation for the collector current of a bipolar transistor including the Early effect: (3.60) in which Vr denotes the Early voltage for the forward active mode of operation and tox: is the collector-emitter voltage. Clearly, the collector current is a function of two controlling voltages. The first derivatives with respect to the controlling voltages vp and vce are the transconduc- tance gm and the output conductance g,, respectively. Using these symbols, the AC value of the collector current can be written as the sum of a series containing powers of uye only, a series depending only on tice and a series with nothing but cross-terms: be = Im Vve + Korg, the + Ksgg, Ube +++ + Go"Vee + Kag, Veg + Kg, Ute + ++ +K Bu ( “Upp go Vee Vee + Kaye, seg, Vbe"Vee + May, nrg, Ube’ Yee G61) in which the time dependence has been omitted for clarity. The meaning of the coefficients in this power series is as discussed above and in Section 3.2.1. The definition of the coefficients is repeated for clarity in Table 3.1. The nonlinearity coefficients Kz,,, and Ka, have already been discussed earlier in the ex- ample of the one-dimensional collector current (see Section 3.2.1). The coefficient g, is the output conductance and is given by Ic/Vap. The coefficients K2,, and K's), are zero: since the3.2 Power series description of basic nonlinearities 41 7 1H In Kom | Gough Pic Kay Ky —— | 2m Yond | Dupndven Koon | zomtesn Ga | Momse2ze | , | Table 3.1: Definition of the nonlinearity coefficients of the collector current of a bipolar transis- tor. simple model for the collector current assumes a linear dependence on the collector-emitter volt- age difference, the higher-order derivatives of ic with respect to vcr are zero, The coefficients of the cross-terms in equation (3.61) are given by Im Mgutege = Varn (3.62) c, Gra Koeantege = Var (3.63) Kg ang, =9 (3.64) For an Early voltage Vr of SOV and a collector current of 1m, one finds gy = 2x10-°A/V, Kay gg, = T:TSx10-" A/V? and Ky, 4, = 0.015 A/V* ‘A more advanced model of the collector current than the one from equation (3.60) will be discussed in Chapter 6. 7 3.2.5 Three-dimensional transconductance A three-dimensional transconductance is a current source that is controlled by three voltages. In Other words, the current is a function f of three voltages u(t), v(t) and 1w(t). Using a power Series expansion around the quiescent value, the total value of the current can be split into a| 5 42 Description of nonlinearities in analog integrated circuit quiescent part Iovr = f(Ucowrr, Veonrr, Weowrn) and an AC part. This AC part is given by tour(t) = f(uconra(t), vconra(t), Wconralt)) = fUcowrn + teonte(t); Veowrr + Veonte(t)s Weonrr + Weontr(t)) =f Ceowrn Veowrr, Wconrr) + & F(u, ‘contr (t) SS | se elered wel A keI 150 9=0 mort) whi (t) (8-H OG In this series the derivatives are evaluated for u = Ucowrr, 0 = Veonrr and w = Weowrr, The AC current can be split into distinct parts: first, there are three power series that only contai powers of one single voltage. These series correspond to a one-dimensional nonlinear condu tance as discussed in Section 3.2.1. Next, there are three power series containing only cross-te1 in exactly two voltages, similar to the two-dimensional series described in Section 3.2.4. Finall there is a power series that only contains cross-terms in three voltages at the same time. Wh only nonlinear effects up to the third order are considered, then from this last series only the fi term of this series is taken into account. This series implies the introduction of the followin, nonlinearity coefficients: a" f(u,», ymkkooklmn—j-Ko ~ GusavkOw™ o 1 Kn 'magoe 3.66 When the positive integers j,k or (m ~ J ~ K) are equal to one, then they are omitted in th notation, like in Ks, nkeoakeo 3.2.5.1 Example: drain current of a MOS transistor ‘The meaning of the newly defined coefficients is illustrated with a simple model for the drai current of an n-MOS transistor in saturation. Taking into account bulk effect and Early effect, the drain current is given by iv = 8 (was — Ve) +A wns) 3.67 with Vr = Vro +7 (Vise + 6~ V8) 3.68) and B= Kp v 66) The parameters A, 7 and @ are the channel-length modulation factor, the body-effect coefficient and the surface inversion potential, respectively. These parameters will be discussed further in: Chapter 7.nonlinearities 3.2 Power series description of bi a ‘The first derivatives of the current with respect to the controlling voltages tas, Usp and vps are the small-signal parameters gyn, Jin and go. The symbols of these parameters will serve as the subscript of the different nonlinearity coefficients of order two and three, Using the notations introduced in equations (3.5), (3.59) and (3.66), the AC value of the drain current is given by ia = Gas + Kagy Us + Kage Vas # ++ +9o"Vas + Koy, Vis + Ks, viet.. er = dy Pa ~ Keaggy Vs — Kisggy Uy + + Rog trang 28°08 + R826, ggg Uae’ Ub T M5, bys B98 Mab T+ (3.70) 4 Kay gg, tetas + Koy eg, Uae Uas + Kp ag, Uae ts + “2, wy ( svg 02, FR oy sogy Ua Os F Kng pty Vad ds +H, saz Pad Mada t -- + Kay, tggahagy B40 Bab Mas o> In this equation, the first three lines correspond to series that describe the dependence of the AC drain current on one single AC voltage. The following three lines represent the variation of the current when two voltages change at the same time. Hereby we make use of the nonlinearity coefficients that have been defined to describe two-dimensional conductances. The last line in equation (3.70) describes the variation of the current with the three controlling voltages at the same time. ILis seen that the terms that depend on vg, only, are preceded by a minus sign. The reason is that the bulk transconductance gg is usually represented as a controlled source flowing from the soutce to the drain, which is opposite to the direction of the source gym Uyy- This yields a positive value for gp. Hence, dri is the first-order derivative times -1. For consistency, the coefficients Koy,,, and K’39,,, are adjusted in the same way, Equation (3.70) reveals that a description of the second- and third-order nonlinearity of the three-dimensional drain current in general requires sixteen second- and third-order coefficients. However, for the simple drain current model of equation (3.67) the coefficients that are computed by deriving twice with respect to ups are zero because the current is linearly dependent on vps. Also, coefficients obtained by deriving three times with respect to ugs are zero because of the quadratic dependency of the drain current on v¢ Table 3.2 lists the expressions for the coefficients that describe the nonlinearity of the drain current according to equation (3.70). The expressions have been derived using the simple square- law model of equation (3.67). The table also includes numerical values for an n-MOS transistor. The parameters of this transistor are: W = l0um, L = 3um, Ves = 1.45V, Vos = 2V, Veg = 1.5V, Veo = 0.9V, 7 = 0.3V"2, ¢ = 0.7V, A “1, KP = 5x10-°A/V?. With these parameters, an effective threshold voltage of 1.09V and a drain current of 10.9644 ate found.44 Description of nonlinearities in analog integrated circuit, Table 3.2: The first-, second- and third-order coefficients for the description of the nonlinear drain (AC) current, using the model of equation (3.67) B(L+ Vos)(Ves — Vr) dm 6.16«10°8 A/V Koy, fa +A Vos) 8.65x10°5 A/V? Kun, 0 0.0 Als B.A Vos)(Ves = Vr) 6.23x10-6 A/V Sine 2(6-+ Veo)? ke —B 10+ 3 Vos) Vas — Veo +18) =1.59%10-* A/V? os 8(64 Van) ke Ba (+3 Vos) (Vas ~Vro + 18) 3.16x10"7 A/V3 Pane 16 (6 + Ven)?! be Bs (Vas — Vr)? 2.01x10-? A/V Ky, 6 0.0 Av? 3 Ky, 0 0.0 A/V —3(+A Vos) 7 =1.74x10-5 A/V? Kaa toms 26+ Vs5 tata At : 6 4/y2 Kigay, [8 Was ~¥r) LABx10-8 A/V" - — 7 Was = Vr) 114x107 A/V? Kegs atem V6 + Vee weal 0.0 Ajvs Keontom | K Sgn 8 20mb By +A Vos) 8(6 + Vsn)9/? 1.99x1078 A/VS 339840 BN 2 1.58x10-° A/V8 Ks,,, 2d0 0 0.0 A/v*3.2 Power series description of basic nonlinearities 45 says | Bq 800 291x108 A/V : | Kin tcamakete 0.0 Alves Syn 20 | —3.20x10-7 A/V* For transistor models that are more advanced, the analytic expressions for the different deriva- tives of the drain current are very complicated. The expressions listed above can seriously differ from the more exact values. This will be discussed in Chapter 7 . 3.2.6 Tracking nonlinearities In Section 3.2.1 we saw that the first-order expressions of the collector current and the base current of a bipolar transistor as given in equation (3.12) and (3.18), only differ by a constant factor, In general, if a nonlinear conductance or transconductance is described by the equation tour: = g(vconra) G71) and a second conductance is described by jourz = ag(vconrne) (3.72) in which a is a constant, then it is said that the two nonlinearities “track”. The reason for this nomenclature can be explained as follows. If for both nonlinearities the current through the el- cement is plotted versus the controlling voltage on a logarithmic scale, then the two curves are identical, apart from a translation, as shown in Figure 3.4. For the determination of the nonlin- earity coefficients, the relationship (3.71) is developed into a power series around its quiescent Point, after which the AC part can be identified with tout = iveonirs + Korg, Venter + Kg, Veonens ++ (3.73) in which, according to the above definitions, di n= (3.74) VCONT 1 dd or) Ku, = 1 fioun (3.75) (3.76) Doing the same for the relationship equation (3.72), we obtain iguta = 92° Veontr2 + Kagy “ Veontra + Kgn * Utontra + Bm| | | 46 Description of nonlinearities in analog integrated circuit, output current (log) #2 = ag(vconrei) #1 = g(vconrm) controlling vi age log) Figure ’.4: The input-output relationship of two nonlinearities that track. with di = tour (3.78 duconrre 1 Piour 2 Cours (3.79 2 dveonrre 1 @ioura (3.8¢ 5.2 6 dveonrne If the quiescent value of the controlling voltage is identical for both nonlinearities, then it is cles from Figure 3.4, that the derivatives that determine the above nonlinearity coefficients only diffe by a factor a, which yields f= 491 381 Koy, = aK, (3.82 Kay, = a Kay, (3.82 The normalized nonlinearity coefficients can now be derived. For the nonlinear conductance described by equation (3.71) we find Piours Ki, = (3.84 and Pioun Ki = —Wéoonrm 3.8 39, = 6913.3 Integrated resistors ” If for the second nonlinearity, described by equation (3.72), the quiescent value of the controlling yoltage is the same as for the first nonlinearity, then we find for the normalized nonlinearity coefficients (3.86) (3.87) Here we come to the important conclusion that tracking nonlinearities with identical quiescent conditions have the same normalized nonlinearity coefficients, or, in other words, they produce the same distortion The first-order expressions of the base and the collector current of a bipolar transistor were the first example we saw of tracking nonlinearities. Tracking nonlinearities will occur later in this book as well. The concept of tracking nonlinearities can, of course, be generalized to other basic nonlinearities as well. 3.3 Integrated resistors In bipolar processes and analog CMOS processes integrated resistors of acceptable quality can be realized in different ways. For each of the possible implementations, foundries specify a sheet resistance, the absolute accuracy, a temperature coefficient, a breakdown voltage and a voltage coefficient. The voltage coefficient indicates how much the value of the resistance varies when the voltage over the resistor changes. This variation is normalized to the nominal value of the resistor. In this way, the voltage coefficient is expressed in percent per Volt. With the specification of a voltage coefficient no distinction is made between the DC and AC value of a resistance. In Section 3.2.2.2 it has been pointed out that there is a difference between the AC and DC value when the resistor is nonlinear, The voltage coefficients that are computed in this book address the change of the AC resistance and not the DC resistance, whereas voltage Coefficients of integrated resistors as they are specified by a foundry, usually address the DC value. Moreover, the specification of the nonlinearity of a resistor by means of a voltage coeffi- cient is incomplete. Only one value of the voltage coefficient is specified at a fixed DC voltage over the resistor. Voltage coefficients for resistors range from 0.5%/V for diffused resistors fabricated with the emitter diffusion (bipolar process) or with the source and drain diffusion (CMOS process), to 0.02%/V for polysilicon resistors [Lak 94]. Resistors that are isolated from their substrate by means of a depletion layer have a significant Voltage coefficient. In the next section, this voltage coefficient will be estimated in terms of the 8eometry of the resistor and of the process parameters. Also, the voltage coefficient at bias Voltages different from zero will be evaluated and the third-order nonlinearity coefficient will be computed. A Voltage coefficient could as well be defined for AC values of an element. For a voltage- Controlled element, such as a nonlinear conductance, this voltage coefficient indicates the relative Variation of the AC conductance with the applied voltage. Since the AC conductance is nothing48 Description of nonlinearities in analog integrated ci else but the first derivative of the voltage-current relationship that describes the conduct the variation of the AC conductance corresponds to the second derivative of the voltage-c relationship. If this variation is normalized with respect to the AC conductance, then it is that the voltage coefficient of an AC conductance gor is equal to two times the second-o normalized nonlinearity coefficient K3,_/2. ‘The factor two arises from the fact that Kp, only half of the second-order derivative: 3.3.1 Nonlinearity coefficients of an implanted or diffused resistor , In this section we will compute the nonlinearity coefficients of an implanted or diffused resis First, we will derive a relationship between the applied voltage over the resistor and the cu through the resistor. We will assume that the voltage is the controlling quantity and the cu the controlled one. Hence, we will represent the element as a conductance instead of a resist Figure 3.5 depicts an implanted or diffused resistor consisting of n-type material with a form doping concentration Nyes. The applied voltage over the resistor is denoted as vcowrrs VcoNnTR 0 Figure 3.5: A diffused or implanted resistor. the current through the element is written as tour. The resistor is embedded in uniform p-type material with a concentration Ny. The verti distance between the surface and the junction is denoted as D. The horizontal width of resistance, from junction to junction is W’, When a voltage vcown is applied over the resist current will flow through the resistor. It is assumed that this flow is only in the x—direction. length of the resistor is L. “The effective depth and width of the resistor are reduced by the extension of the deplet layer inside the resistor. Since the width of a depletion layer depends on the voltage over pn-junction, the size of the resistor, and hence its value, is voltage-dependent. In the subseq) calculations, the nonlinearity caused by this effect will be computed.3,3 Integrated resistors » In the calculations below, the effective depth and width will be denoted by d and w, respec- tively. These are given by d=D-%q (3.88) w=W 2a (3.89) in which 24 denoted the extension of the depletion layer inside the n-type material. For an abrupt junction, this is given by [Sze 85] wy =AVVi te (3.90) Here V; is the junction potential and v is the voltage on an «r—position between 0 and L, referred tothe voltage on the place x = L. The factor A is given by 25/1 1 h As 1). 3.91 oon (a x) Ny + Nres G91) The junction potential V, depends on the doping of the resistor Nc, and of the surrounding petype material, 1V,: KT —In (3.92) 4q It is assumed that the current in the resistor is totally determined by the majority carriers, in this case the electrons that drift due to the applied voltage. No recombination of generation of electron-hole pairs is considered. The current density of electrons that flow into the direction is then given by [Sze 85] det) doe dr pa de in(2) = Onn (3.93) in which n indicates the electron concentration per volume unit and jz, is the mobility of elec- trons; pa is the bulk resistivity. The total current as a function of z is found to be vw pet dul iour(x) = -f [ 4 ate) (3.94) Jy=o Je=0 Pn ce According to the above assumptions, the integrand is constant, such that 1 du (a ade! ) (3.95) iour(a) de = whew du(x) (3.96)50 sidered as being constant between x = 0 and « = L. Also, the bulk resistivity is assumed to constant in the direction. Integrating both sides of equation (3.95) from x = 0 to x = L, using equations (3.88) and (3.89) then yields 4 jour = = rr (Ww -24V% ¥u) (D- AVY Fe) do aa ‘After some algebra one obtains 1 jour =7r [WD veowma — 34 QD + W) (Vj + ecowre)" - +24°V; -vconra + A’ UGonre) & Equation (3.98) is an admittance description of the nonlinear resistor, expressing the cu through the device as a function of the voltage over the device. This relationship is nonlinear. however, the extension of the depletion layer inside the resistor is neglected, which correspor to setting A = 0, then the resistor becomes linear, as can be seen from the admittance descripti '‘D tour = “UCONTR Gs pL For the determination of the nonlinearity coefficients equation (3.98) is developed int power series around the quiescent point (our, Veowrr)- This yields 1 pp, AD + W) (Vj + Vin)? 7 iour =lour + al(™ Dig 8G + Veownn) ) veo 9 a (e-A@DEW ) ey A/V; + Voowrr ‘The AC part of this series can be identified with the power series expansion of the AC through a nonlinear conductance (see Section 3.2.1): jiout = Gac* Veonts + Rog, * Vout + Bagge * Vlonte + which yields (WD A@D+W) VGH Vin +24? (Hs + Voowsa)) 4 pL Ky, —-1 fe A(2D+W) we ph 4\/V; + Veowrr Ke -k AD +W) ome pl \ 24 (V+ Veowra)’3.3 Integrated resistors a 33.1 Example ‘As an example, the nonlinearity coefficients as derived above are evaluated for an implanted resistor of 20k@ which is realized with a phosphor or arsenic implantation in a p-type material. The depth D of the (n-type) implanted region is 4.5m. The impurity concentration Nj of the implanted region is 1.1x10'°crn~*, whereas the concentration N, of the underlying p-type material is 8.1x10'cm*. The bulk resistivity p of the implanted region is 0.42 em. The width W of the region is 10m, The junction potential (see equation (3.92)) is found to be 0.637V. ‘The factor A of equation (3.91) is equal to 9.08x10-8m/V/?, If the extension of the depletion layer inside the n-type material is neglected, then the resis- tance is given by pL/(WD) (see the first term in the right-hand side of equation (3.98). If this value is set equal to the required 20kQ, then L is found to be 25pm. We will now see what the influence of the extension of the depletion layer is, not only on the value of the resistor, but also on the nonlinearity coefficients. Figure 3.6 shows the conductance of the resistor as a function of the voltage vcowrr over the element. If the depletion layer would not extend at all into the implanted region, then the conductance would be 5x1074/V. Due to this extension, the conductance is smaller as seen from Figure 3.6. Also, it is seen that the conductance decreases as ucow'e increases. This is duc to the increase of the extension of the depletion layer into the implanted region. 4 conductance vcontrV) Figure 3.6: Variation of the conductance of an implanted resistor as a function of the voltage Over the resistor Ih order to have an idea about the nonlinearity of the integrated resistor, the normalized Ronlinearity coefficients of order two and three have been computed as a function of the input Voltage as shown in Figure 3.7. The voltage coefficient that expresses the nonlinearity of the AC conductance is found by multiplying £3, with two. In this way, we find from Figure 3.7 that the vollage coefficient of Tesistor varies between 2.4% and 0.8%.(A/V?) 2g K. 52 Description of nonlinearities in analog integrated Sireut 0.0035 " 0.0030 0.0025 0.0020 | 0.0015 0.0010 0.0005 0.0000 -0.004 | -0.006 -- 0.008 0.010 -0.012 0.014 veonre(V) veonrn(V) : igure 3.7: Normalized nonlinearity coefficient K,, (left) and K,, (right) for the implanted resistor as a function of the voltage over the resistor Further it is seen that both the second- and third-order nonlinearity coefficients becomes, smaller in absolute value as vcowrr increases. This means that the resistor becomes more lineatt as the voltage over the element increases. i 3.4 Integrated capacitors ‘The capacitors encountered in integrated circuits can roughly be divided in two types: oxide capacitors and junction capacitors. Oxide capacitors are consist of two conducting layers wit an oxide in between. Examples are capacitors between a metal layer and a polysilicon layer, and capacitors between two different polysilicon layers. Junction capacitors, on the other hand, fall apart in two parts: the capacitor of the depletion layer around the junction, and the diffusion} capacitor. The latter is only important when the junction is forwardly biased. In this section we! concentrate on the nonlinearity of junction capacitors. ! ‘The junction capacitance per unit area is simply given by 5 Est = 3.105) S Bap + in (3.10: y i in which xrg, and x, are the extension of the depletion layer into the p-type material and into the: n-type material, respectively. For an abrupt junction, xi» is given by equation (3.90), which is reformulated as 4 Lan = (Wt iy =) q Nn Np in which vp is the reverse voltage over the junction, Vj the junction potential given by equa+ tion (3.92) and NV, and Np are the impurity concentrations in the n-type material and the p-type material, respectively. The expression for x4, is obtained by interchanging the role of p and n ist equation (3.106).3.4 Integrated capacitors 53 Very often, the values N, and N,, differ largely. If, for example, Ny > Np, then gn > tap, and the junction capacitance per unit area can be approximated by (3.107) (3.108) In many practical implementations a junction is not abrupt. In those situations, the junction capacitance per area unit is given by Sp (+#)" Vj referred to as the junction grading coefficient, For an abrupt junction, mj = 0.5 as explained above, and for a linear junction m; = 0.33 [Sze 85, Lak 94]. It is seen in equation (3.109) that C; goes to infinity for vg = —V;. This does not occur in practice. In [Poon 69] it is shown that expression (3.107) is no longer valid at forward bias voltages around the junction potential. Nevertheless, the value of a junction capacitance at con- siderable forward bias voltages (say larger than V;/2) is not important, since under forward bias the junction capacitance is much smaller than the diffusion capacitance of the junction. As a result, distortion caused by junction capacitors under forward bias conditions is neglected com- pared to distortion caused by diffusion capacitors. In Figure 3.8 the variation of the junction capacitor is shown as a function of the reverse Voltage vp over the junction for Cy being 30fF, m, = 0.33 and 0.7V. The scale of the vertical axis is logarithmic. An interesting observation from Figure 3.8 is that at large reverse bias voltages the junction capacitance not only decreases, but it also changes less with the applied Teverse voltage. Hence it becomes more linear at large reverse bias values. For distortion computations it is interesting to derive the nonlinearity coefficients that de- scribe a junction capacitance. To this purpose, equation (3.109) is expanded into a power series around the quiescent point. This yields (3.109) + Hey tf dug 2 = C,(Vr) ut. (3.110) duh in which Q is the charge associated with the depletion layer, Ve is the quiescent value of the Teverse voltage vp and v, is the incremental value of the reverse voltage. The derivatives in €quation (3.110) are evaluated at Vp.54 Description of nonlinearities in analog integrated circuits! 4e-14 " 3.2e-14 | -- 2.5e-14 C;(F) 2e-14 |---- 1.6e-14 136-14 tet4 0 Figure 3.8: Variation of the junction capacitance as a function of the reverse voltage. The AC current through the capacitor C; is dQ dQ dup dv, dey = AR MO Wn Os Looe ae fou Gt dug du, dt? dt Gall Substitution of equation (3.110) into equation (3.111) yields z gi day, 1 0,4 du, 3 fou = Cid G+ 5° Goeat OF) +5 ao ae) + Gara This is identified with , 42 dus tou = Ci(VR aya + Bag, @ (vt) + Kag, a (7) + B.113 which yields 1d, =p ot 3.114} Ke, = 3" don on 1 &C; a, =i. 3.1 Kag, = § ah GANS G.116, GALT3,5 Weakly nonlinear transistor models: introduction 55 or, for the normalized nonlinearity coefficients, (3.118) (3.119) It is seen that the normalized nonlinearity coefficients decrease when the reverse voltage in- creases. This is in correspondence with the observation made from Figure 3.8: a junction capac- itor becomes more linear as the reverse voltage increases. ‘As an example, consider a base-collector junction with a junction potential Vj of 0.7V and a grading coefficient m, of 0.33. Then one obtains K, = —0.23V-) and Ky, = 0.15V~? for a reverse voltage of OV, For a reverse voltage of 1V one obtains Kj, = —0.09V-! and Ky, = 0.025V 2 Compared to other integrated capacitors, a junction capacitor is quite nonlinear. This is due to the change of the width of the depletion layer with the applied voltage over the junction. For example, the voltage coefficient of an integrated poly-substrate capacitor is 0.05%/V, and for a poly-poly capacitor 0.005%/V [Lak 94]. 3.5 Weakly nonlinear transistor models: introduction ‘The different basic nonlinearities described in the previous sections can now be tailored together to construct nonlinear equivalent circuits for transistors. These are straightforward extensions of the linear equivalent circuits [Gray 93, Lak 94]. Before discussing silicon bipolar and MOS models in depth in Chapters 6 and 7, a few preliminary concepts are discussed here. Much like many transistor models that are used in SPICE-like circuit simulators, the transistor models that will be used in this book contain some simplifications. In this way, some insight can be obtained in the nonlinear phenomena that are the major sources of distortion caused by a transistor, More accurate models are sometimes too complicated to provide such insights and Tequire iteration or finite-element analysis for their evaluation. A considerable simplification both for bipolar and MOS models is achieved by considering the current flow in a transistor as one-dimensional. Explicit two-dimensional effects are then modeled as extensions to those one- dimensional models. Also, distributed elements are lumped into just a few circuit elements. An example is the base resistance of a bipolar transistor which is shown in Figure 3.9. It is seen that the base resistance is distributed over the width of the base-emitter junction. Such distributed representation is not practical to handle in circuit design. Instead, a lumped Tepresentation is used. Apart from the simplifications as the ones mentioned above, some further simplifications are Performed in practical device models, still in order to end up with tractable analytic expressions.56 ion of nonlinearities in analog integrated circuitay r=0 rab consist of distributed elements (left). In bipolar transistor models for circuit design these et ements are lumped into only two circuit elements (right). The extrinsic base resistance (ne shown) is put in series with Rp. Figure 3.9: The intrinsic base resistance and the base-emitter junction of a bipolar | For the analysis of a linearized circuit first-order derivatives of the model equations are requit and — as pointed out in this chapter — for distortion computations higher-order derivative are required. For example, the transconductance g,, of a MOS transistor is obtained by takin the derivative of the model equation for the drain current with respect to vgs. The second: and third-order nonlinearity coefficients K»,,, and Ks,,, are — according to the results fro Section 3.2.1 — obtained by taking the second- and third-order derivative of the drain curret However, some effects which do not seem to be important for the computation of the current of 4m could become important when the higher-order derivatives are considered. Hence, one shoul be careful not to oversimplify the models and to introduce fit factors too rigorously. However, accurate computation of derivatives of order five or higher is not achievable due to the inevitabl simplifications that have been made to obtain analytic model equations Existing circuit simulators provide values for the transistor current and its first derivatives which are the small-signal parameters, but the higher-order derivatives — which are proportion: to the nonlinearity coefficients — are not computed. In order to obtain a value for the nonlin: earity coefficients, two approaches can be followed. First, the derivatives can be computed b} numerical differentiation [Num 92]. This approach can lead to inaccuracies, since numerical dif- ferentiation is a process that can lead to numerical instabilities. Furthermore, this approach d not yield enough insight: apart from the numerical values, it is not known which physical effec! determine the nonlinearity coefficient under consideration. In order to avoid the disadvantages of numerical differentiation, an alternative approach been developed. In this approach, analytical expressions for the derivatives are determined by computing symbolically the derivatives of the model equations with respect to the different con- trolling voltages or currents. This process can be speeded up by making use of symbolic algel programs such as MAPLE [Map 91], MACSYMA [Macs 87] or MATHEMATICA [Wolf 91} that can compute derivatives symbolically. These programs provide facilities to dump a giv expression into source code for C or Fortran. Using this approach, C routines have been devel oped. In these routines the different derivatives are computed as a function of the bias voltage or currents, the transistor dimensions and the model parameters. The correctness of the routine3.6 Summary 57 has been checked by a comparison with the derivatives that have been computed by numerical differentiation. An additional test for the correctness is a comparison with simulation results from a nu- merical circuit simulator. As already mentioned above, these circuit simulators usually do not compute derivatives of order higher than one. Nevertheless, a comparison of the current and the first derivatives with simulation results already give a good indication of the correctness. This comparison, of course, is only possible for models that have also been implemented in the circuit simulator. The expressions of derivatives of model equations are usually very complicated. Insight can be obtained from those expressions by plotting them as a function of bias, transistor dimensions, model parameters, .... Nevertheless, the insight in distortion phenomena would be greatly en- hanced if a closed-form expression — even if it is approximate — could be obtained for the derivatives of interest. To this purpose, the C routines mentioned above have been extended by routines that trace the dominant contributions to a given nonlinearity coefficient. An interpreta- tion of these dominant contributions makes it possible to identify the dominant physical effects that determine a given nonlinearity coefficient. A tracing of the dominant terms of a derivative is possible since the transistor current is usually a sum or a product of several composed functions. For example, in Chapter 7 it will be shown that the drain current of a MOS transistor in strong inversion (triode region) can be written as a product of three functions: ip = mobred (ues, ns sp) hot (vas, ons, ven) *large (was,vps, vse) 3.120) in which the three functions mobred, hot and large are in turn functions that may contain sums, products, powers and exponential functions of the transistor terminal voltages vas, vps and vsp. The meaning of these three functions will be explained in Chapter 7. Since the derivative of a product with respect to a variable (in this case a voltage) is a sum of two functions, it is easy to see that the nonlinearity coefficients, which are proportional to higher-order derivatives of the current with respect to the terminal voltages, will consist of a sum with many terms. An identification of the dominant terms in this sum will yield an approximate, interpretable expression of the nonlinearity coefficient under consideration. This approach will be illustrated in Section 7.7.4. The C routines that are developed to evaluate the nonlinearity coefficients and to identify the dominant contributions could be combined with a numerical circuit simulator, Moreover, it is not mandatory that the transistor model used by the circuit simulator is exactly the same as the model that is used in the computations of the nonlinearity coefficients: the circuit simulator only needs to find the DC solution, providing the bias values needed to evaluate the nonlinearity Coefficients. 3.6 Summary In this chapter we defined the nonlinearity coefficients of the basic nonlinearities that will be Used in the next sections to describe the nonlinear behavior of bipolar and MOS transistors. In Addition, we discussed the nonlinearity of integrated capacitors and resistors.58 Description of nonlinearities in analog integrated circuits’ The nonlinearity coefficients of order two and three are proportional to second- and third- order derivatives of the model equation with respect to the controlling voltage(s) or current(s). Since model equations can already be quite complicated, the expressions of the derivatives are even more complicated such that they cannot be interpreted easily by a design engineer. In: order to overcome this problem, an approach has been discussed to derive the dominant terms of, the nonlinearity coefficients. This results in shorter expressions which can be better interpreted,! This approach will be used in Chapters 6 and 7. iChapter 4 Volterra series and their applications to analog integrated circuit design 4.1 Introduction In this chapter it is investigated how weakly nonlinear behavior of an analog integrated circuit can be computed. The emphasis here is on obtaining insight. Therefore, less attention is paid to numerical methods such as time-domain analysis followed by a Fourier transform, which is the classical SPICE approach [Royc 89, Hspi 96], or harmonic balance methods. Such methods compute the response of a nonlinear circuit by iteration, and the final result is a list of numbers, which do not indicate which nonlinearities in the circuit are mainly responsible for the observed nonlinear behavior. Hence such methods are suitable for verification of circuits that have already been designed. When simulations show that the specifications regarding the nonlinear behavior are not met, then these methods do not present information from which designers can derive which circuit parameters or circuit elements they have to modify in order to obtain the required specifications. Such valuable information can be obtained with the use of Volterra series. The price that is paid for this extra insight is that the analysis is limited to weakly nonlinear behavior only. Volterra series have already been used for distortion computations [Nar 67, Nar 70, Mey 72, Sans 72, Kuo 73, Nar 73, Buss 74, Khad 74, Kuo 77, Rud 78, Chua 79a, Chua 79b, Wein 80, Sale 82, Chua 82, Maas 88, Wamb 90], also with SPICE [Chis 73, Royc 89]. In several imple- mentations of SPICE these computations can be performed with the .DISTO command. Never- theless, Volterra series are seldom used by IC designers. This is explained by the fact that the use of Volterra series is limited to weakly nonlinear behavior, and the simulation results are presented in the same way as for most numerical simulators, namely as a list of numbers, from which it is not clear what is behind those numbers. In this book we try to perform distortion computations with Volterra series in a way that insight can be gained. Volterra series describe the output of a nonlinear system as the sum of the response of a first- order operator, a second-order one, a third-order one and so on [Sche 80]. These operators are shown in the block-diagram representation of Figure 4.1. Every operator is described either in60 Volterra series and their applications to analog integrated circuit d x(t) +| He y(t) Figure 4.1: Schematic representation of a system characterized by a Volterra series, the time domain or in the frequency domain with a kind of transfer function, called a Volterra kernel. Just as for linear circuits, the frequency-domain representation is preferred over the time- domain representation for many circuit analysis aspects. It has to be remarked here that the block that is represented in Figure 4.1 with Fi, represents the linearized system. In fact a Volterra series describes a nonlinear system in a way which is equivalent to the way Taylor series approximate an analytic function. A nonlinear system which is excited by a signal with small amplitude can be described by a Volterra series which can be broken down after the first few terms. The higher the input amplitude, the more terms of that series need to be taken into account in order to describe the system behavior properly. For very high amplitudes, the series diverges, just as Taylor series. This is often due to the occurrence of a strong nonlinearity at high amplitude levels, such as the cut-off of a transistor. Hence, Volterra series are only suitable for the analysis of weakly nonlinear circuits. The Volterra series approach has been proven to be very attractive for hand calculations of small transistor networks, Several studies of effects like intermodulation or harmonic distortion and cross modulation in such circuits have been reported [Nar 67, Mey 72, Poon 72, Sans 72, Nar 73, Khad 74, Abra 76]. Since Volterra kernels retain phase information, they are especially useful for high-frequency analysis and for effects like AM to PM conversion. Volterra series give ‘a general characterization of a nonlinear circuit in the sense that once the Volterra kernels of a circuit is known, ils output can be found for any input. For example, the response of a nonlinear system to noise can be studied with Volterra series [Bedr 71, Rud 78, Sche 80]. In Section 4.2 some definitions concerning Volterra series are given. ‘The mathematical foundations of those definitions are given in Appendix B. ‘The Volterra series representation is not only an explicit nonlinear representation of the sys- tem response in terms of the input, but also provides insight into the system operation. ‘This insight is readily obtained from the block-diagram representation of the Volterra kernels. Using this representation some interesting applications can be considered like cascading of nonlinear circuits, distortion cancellation, linear and nonlinear feedback in weakly nonlinear circuits,....4.2 Basics of Volterra series 61 ‘These topics are covered in the Sections 4.5, 4.6, 4.7 and 4.8 “The reported hand calculations of Volterra kernels were limited to small circuits only. In the seventies an algorithm has been developed for the numerical calculation of Volterra kernels in the frequency domain for larger nonlinear networks. The development of the original algorithm by Bussgang, Ehrman and Graham [Buss 74] was granted by the U.S. Air Forces for the anal- ysis of their communication receivers. Later, this algorithm has been generalized by Chua and Ng [Chua 79b]. This algorithm is explained in Chapter 5. It is interesting to note that a sim- ilar approach for time-discrete circuits such as switched-capacitor circuits, has been described in [VdWal 83]. In Chapter 5 some alternative methods will be discussed that circumvent the use of Volterra series, but the approach is basically the same. ‘The Volterra series approach is not the only technique to obtain approximate closed-form expressions for nonlinear behavior. Another technique is the method of the describing func- tion [Ath 75]. With this technique it is possible to obtain closed-form expressions for a feedback system that contains an isolated static nonlinearity in the feedback loop. Since it is not possi- ble in general to map an analog integrated circuit to such a feedback system, the method of the describing function is not used here. 4.2 Basics of Volterra series In this section the fundamentals of Volterra series are briefly discussed, both in the time domain and in the frequency domain. A thorough study of Volterra series can be found in [Sche 80] The theory of Volterra series can be viewed as an extension of the theory of linear, first-order systems to weakly nonlinear systems [Sche 80]. In the Volterra series description, such a system is considered as the combination of different operators of different order as shown in Figure 4.1 Every block Hy. Hy and H,, represents an operator of order 1, 2,..., respectively. When a nonlinear system is excited by a signal with very low input amplitude, then the output can be described accurately by taking into account only the first-order behavior of the system, which is the linear behavior, represented by block Hj. In the frequency domain this block is characterized by the transfer function of the linearized circuit. When the input ampli- tude increases, then a substantial part of the output signal is caused by nonlinear effects, For a sufficiently low input amplitude, these nonlinear effects can be described accurately by taking into account second- and third-order effects only, which are modeled by the operators Hy and H, of Figure 4.1. The total output is the sum of the outputs of every operator individually. When the input amplitude increases even more, then higher-order effects occur, which are modeled by higher-order operators. For a nonlinear system excited by a sine wave with a limited amplitude this model roughly corresponds to a decomposition of the output signal into the different harmonics. The nth har- monic is primarily determined by the nth-order operator of Figure 4.1. Just as for linear circuits, the analysis can be performed in the time domain as well as in the frequency domain. The use of Volterra series for nonlinear systems can be compared to the use of Taylor series for analytic functions. In the latter case, small excursions of the function arguments around a fixed point can be described accurately, if at least the excursions fall within the convergence62 Volterra series and their applications to analog integrated circuit desigt radius of the Taylor series. The larger the excursions are, the more terms of the Taylor serie: must be taken into account for a sufficiently accurate approximation. Also, some Taylor serie: around a certain point have a convergence radius of zero. Equivalently, Volterra series have i convergence radius which can be zero. The latter case corresponds to systems which cannot bi described at all with a converging Volterra series such as an ideal clipper [Sche 80] or an idea comparator. Figure 4.2 shows the input-output relationship of such clipper. When the inpu output 0 input Figure 4.2: Input-output relationship of an ideal clipper. signal is zero, then the output is zero as well. The least deviation of the input signal from zero i cither direction gives an output of +1 or 1. It is clear that this is a strongly nonlinear effect. 4.2.1 Volterra operators Consider a second-order nonlinear system or, briefly, a second-order nonlinearity. Its operatic can be seen as follows. The second-order nonlinearity combines two signals, which eventual) are identical, and produces with these signals a second-order signal. A third-order nonlinearit combines three signals, which can be identical, then multiplies them to produce a third-ord: signal. The combination of signals can be just a multiplication, but in Sections 4.3.1 and 4.3 we will see that the two signals can be combined in a more complex way. Assume now that a single sinusoidal signal with a frequency w is applied to a second-ordk nonlinearity. In this case the second-order nonlinearity combines two times this signal to produc a second-order signal. This signal has two frequency components, one at w -+ wy = 20 ar one at w; — w = 0 Hz. When two sinusoidal signals with frequency w; and w. are applied : this second-order nonlinearity, then the second-order signal combines two times the first sign to produce a DC term and a component at 21, which is denoted as a second harmonic. Als the second-order system combines two times the input signal at w2 to produce a DC term and component at 2. Finally, the nonlinearity also combines the sine wave at w with the sine war at wy, to produce a sine wave at the frequencies w; + w and |w; — w2|. The latter two respons: are intermodulation products.4.2 Basics of Volterra series 63 Consider now a third-order nonlinearity. This nonlinearity will combine three signals, some of which can be identical, and produce with these signals a third-order signal. When one sinu- soidal signal at frequency .; is applied to this nonlinearity, then the latter produces sine waves at the frequencies | + wy +t w ++ w1|. This corresponds to two frequencies, namely a; and ws With the above considerations we obtain components at the same frequencies we found in Chapter 2 for a simple nonlinear system. Although the considerations look quite intuitive, they can be formalized using second-order, third-order and in general nth-order operators. This is explained in Appendix B An operator performs a transformation of the input signal that results in an output signal. The operator that corresponds to the second-order nonlinearity that we discussed above, is denoted here as the second-order Volterra operator and its notation is Hy. The output of this second- order operator to which an input signal x(t) — which can be a sine wave, a combination of sine waves, ... — is applied, is then denoted as Hy[2(?)]. Similarly, an nth-order nonlinearity is described with an nth-order Volterra operator H.,, and the output of this nth-order nonlinearity is denoted as H,{z()]. Finally, one should not forget to mode! the linear behavior of the system as well. This is performed with a first-order Volterra operator. This operator is identical to the operator that describes a Jincar system, and its computation can be performed using the theory of linear systems. From this theory we know that the transformation that is performed by this first-order Volterra operator is a convolution of the input signal with the impulse response of the linear o linearized system. 4.2.2. Time-domain fundamentals With the Volterra series model of Figure 4.1 a nonlinear system consists of several Volterra operators in parallel. In the previous section we explained qualitatively how such operators act on an input signal. In this section we will explain this in a quantitative manner. In Appendix B it is shown that the transformation on an input signal performed by a second- order Volterra operator is given by opt H,l0(0)] | [ a(n, 72)e(t — 71)0(t — 72)dndry 4.) This two-dimensional integral is recognized as a two-dimensional convolution integral. Here we see a similarity with a linear system: whereas the output of a linear system is computed by taking the convolution of the input signal with the impulse response of the linear system, the output of a second-order nonlinearity is a two-dimensional convolution of the function p(T, 72) with the input signal. The function h2(r;, 7) is denoted as the second-order Volterra kernel. In [Sche 80] itis shown that the second-order Volterra kernel can be considered as a two-dimensional impulse Tesponse. Consider now a third-order Volterra operator. The mathematical representation of the trans- formation that is performed by a third-order Volterra operator, is given by Hyl2(2)] -f “fe [- ig(rista, reat —n)alt—n)e(t—re)dridmdrs (42)64 Volterra series and their applications to analog integrated circuit design and, for an nth-order Volterra operator Hyeecol= « [talon oe sty)e(t —n)a(t = 12) -+-20(t— m)dridry .. dry (4.3) The n-dimensional integral is seen to be an nth-order convolution integral. The function Iin(T1s72y+++ Tn) i8 an nth-order Volterra kernel. In (Sche 80] it is proven that this can considered as an nth-order impulse response. Equation (4.3) indicates that the output signal Sf an nth-order Volterra operator can be computed by taking the n-dimensional convolution of this n-dimensional impulse response with the input signal Recall now that with the formalism of the Volterra series the output of a nonlinear system is represented as the sum of the output of a first-order Volterra operator with the output of a second-order one, a third-order one, and so on, as shown in Figure 4.1. Using equations (4.1), (4.2) and (4.3) we find that the complete output of the complete nonlinear system is given by y(t) = Hy [x(t)] + He[e(t)] + Ha[z(t)] +... + Hal t] +. (4.4) This equation is a Volterra series representation of the nonlinear system. Let us consider the nth-order Volterra kernel fn (71 7a) in more detail. It is seen that a Volterra kernel of order n is a function of n time-domain variables. In Appendix B it is shown that a kernel can always be made symmetric, which means that the value of the kernel does not change when the arguments are interchanged. All Volterra kernels that will be used throughout this book are assumed to be symmetric unless it is explicitly mentioned they are not symmetric. In this book we are reasoning only with causal systems. This means that the response does not depend on the future of the input. For such systems it can be proven [Sche 80] that the value of the Volterra kernels for any negative argument is zero: for any 1; <0, 7 (4.5) Let us now consider the shape of a Volterra series. The Volterra series is a power series. This can be seen by changing the input by a factor a, resulting in a new input ax(t). Using equation (4.3) and (4.4) the new output is uo) = SOHslax(e) = Sah (ol0) 46) a Moreover, a Volterra series is a series with memory. Indeed, the integrals for H,, are convo- lutions. This is obvious for the first term of the series in equation (4.4), which is nothing else but the convolution integral of a linear system. ‘The terms after the first one are higher-order convolutions [Sche 80]. This means that Volterra series —in contrast with power series— canBasics of Volterra series 6 describe weakly nonlinear circuits with capacitors and inductors. If the system described by the Volterra series of equation (4.4), has no memory at all, then faaltis Tos =0 — forany 7) > 0, G=U Qn (47) and the Volterra series reduces to the power series u(t) = Yo hn(0,0,..- 0)2%(t) (48) ‘Asa result of its power series character, there are some limitations associated with the use of Volterra series to nonlinear problems. Just as with the Taylor series representation of a function, problems of convergence may occur. When the input amplitude increases, the higher-order terms in the Volterra series relatively increase more than the lower-order ones. At a certain amplitude, divergence may occur. Hence, a convergence radius can be defined [Boyd 84]. The computation of a convergence radius is quite complicated and it will not be considered in this book. Instead it will be assumed that the signals that are applied to the circuits under consideration are small enough such that the Volterra series converges. This is a realistic assumption in most situations. 4.2.3. Frequency-domain representation From the theory of linear systems we know that the output of a linear system can be computed in the time domain by performing a (one-dimensional) convolution of the time-domain representa- tion of the input signal with the system’s impulse response. In the frequency domain the response of linear system can be computed by multiplying the Fourier transform of the impulse response with the Fourier transform of the input signal. This approach can now be extended to second- order, third-order, ... operators. It has already been mentioned above that an nth-order Volterra kernel can be considered as an nth-order impulse response. Consequently, if we draw the parallel with a linear system, we expect that the Fourier transform of the output of an nith-order Volterra Operator involves a multiplication with the Fourier transform of the nth-order Volterra kernel. This is considered now in more detail. First we consider the frequency domain representation of an nth-order Volterra kernel. To this purpose a multidimensional Laplace and Fourier transform are defined. The Laplace transform of the nth-order Volterra kernel kernel is called the nth-order nonlinear transfer function or the nth-order kernel transform. The analogy with a linear transfer function is again due to the fact that an nth-order Volterra kernel can be considered as an n-dimensional impulse response. Very often, when no confusion is possible, the frequency domain representation of the kernel is also called a Volterra kernel. Since the nth-order kernel is a function of n time variables, a complete frequency domain representation requires n frequency variables. In Appendix B it is shown that the multidimensional Laplace transform H,( 15q) of the kernel h,(ri,.-. ,7n) is given by soon Halsiseee 86) = fo PO iltasene TOM dry AB)66 Volterra series and their applications to analog integrated circuit design in which 8, = 0, + jw; (¢ = 1,2,...,n) is a complex number. The multidimensional Fourier transform is obtained from the Laplace transform by making all 0; zero in equation (4.9)'. Itcan be proven easily that the nth-order transfer function is symmetric when its time-domain equivalent is symmetric. This is in fact very natural: a nonlinear operator cannot distinguish the different applied frequencies. An interesting property for symmetric nth-order transfer functions is that its complex conjugate can be obtained by changing the sign of the frequency arguments: He (—Jay~Jway- ++ 1 Jn) = AR Gewr, Jerr, --- » Fen) (4.10) t Equivalent to the linear transfer function, the nth-order nonlinear transfer function can be used to find the nth-order system’s output as a result of a sinusoidal excitation, Consider for example a second-order system that corresponds to the operator H in Figure 4.1. In Appendix B it is proven that the output of such a system excited by a sinusoidal excitation A, cos wt can be written in terms of the second-order nonlinear transfer function H2(1, 82) as follows: A 2 AR elt) = a Re (Ha(jwe, jwe)e") + F Re (Haliws, —jw:)) A a = > |Ho(jwe, jwe)| + cos (Quiet + arg (Ho(jwe, jw2))) + “F Haliwe, —Jus) (4.11) Clearly, the output of this second-order system is a constant and a sinusoid at frequency 2uz. These frequency components were found as well in Chapter 2 for a memoryless second-order nonlinearity. An analysis of the response to a sum of sinusoidal signals is postponed to Sec- tion 4.4. For a nonlinear system corresponding to a third-order Volterra operator we find in Ap- pendix B that the output to the sinusoidal excitation A, cos(wst) in terms of the third-order nonlinear transfer function H(s1, s2, 83) is equal to Al sot) , 342 | just wn(0) = SE Re (Hal jue, te, jude") + —F* Re (Hs jets, jee, —Jwz)e™') Al = Sf [Hal jute, jue, js)| + 608 (Buel + are Ha( jee, jute, Jve))) 34h + FE [Ha Gee, judas —jees)| C08 (wat + arg(Hs (jure, jeter —jwz))) (4.12) It is seen that the response consist of a component at the frequencies 3w, and w,. These fre- quencies have been found as well in Chapter 2 for a memoryless third-order nonlinearity. The difference between the results obtained with Volterra series and the results from Chapter 2 is that the result given here is more general: the Fourier transforms H(jw, jw2) and Ha(jw1, jw, jws) are complex numbers that model phase shifts on the harmonics due to capacitors or inductors in the nonlinear network. This was not taken into account in the simple calculations of Chapter 2, where we only calculated with real numbers instead of complex numbers. For nonlinear circuits where capacitive or inductive effects can be neglected, for example in amplifiers at low frequen- cies, it is clear that the responses computed in Chapter 2 must be the same as the responses . oF course, requires that the region of absolute convergence includes the w1, wz, ... and wy axes.4.2 Basics of Volterra series 67 expressed in terms of Volterra series. This will be discussed further in Sections 4.3.1 and 4.3.2. Farther, the computation of the Fourier transforms H( jus, jus) and Hs (juer, jw2, jus) in a non- Jinear circuit is explained in Chapter 5. For the sake of completeness we repeat that the response of a linear or first-order system toa sinusoidal excitation A, cos wet is given by yi(t) = ARe (Hi (jwz)e*") = A|Hi(jws)| cos (wet + arg(Hi(jws))) a) 4.2.4 Weakly nonlinear circuit behavior revisited At this point, weakly nonlinear circuit behavior can be defined in terms of Volterra series. The most general definition would be that a circuit excited by a given source behaves weakly nonlin- early if its response can be represented by a converging Volterra series. Although in some cases many terms of the Volterra series are required to accurately describe the circuit response even at moderate input levels [VdEi 89], the response of many circuits can most often be described by the first few terms of the Volterra series that describes the circuit under consideration. Therefore, the following more restrictive definition is more practical: A circuit behaves weakly nonlinearly if, for the applied input signal, it can be accurately described by the first three terms of its (converging) Volterra series. This definition requires that for the characterization of a circuit its linear behavior be computed together with its lowest even- and odd-order nonlinear behavior. This yields a quite complete description of the behavior of a large class of analog integrated circuits under normal signal con- ditions, One could argue that the above definition is mathematically not exact: it is not described how accurately the response of the circuit must be described. The required accuracy depends on the requirements of the circuit engineer. Most often a fairly low accuracy (for example errors up toa few dB) are acceptable if at least the results obtained using Volterra series yield insight in a circuit's nonlinear behavior. For both the general and the more practical definition, the problem remains that it is in general not known in advance if the Volterra series will converge, since the radius of convergence can only be computed once the Volterra kernels are known. If the series converges, then it is generally Not known how many terms are required for an accurate description [VdEi 89]. For the circuits 0 which the computations in the subsequent chapters are addressed, however, this is seldom 8 problem, as will be demonstrated with numerical simulations with other techniques that are Not limited to weakly nonlinear behavior. Hence, the practical but more restrictive definition of {enlinear behavior can be used for most analog integrated circuits instead of the more general lefinition. Finally, it must be noted that strictly speaking, weakly nonlinear behavior is not a property of the circuit alone, since it depends also on the amplitude of the excitation(s). However, many Circuits are called weakly nonlinear or quasi-linear [Chua 75]. This means that they behave Weakly nonlinearly when they are excited by practical signal levels for which they have been lesigned.68 Volterra series and their applications to analog integrated cireuit desiga, 4.3 Examples of Volterra kernels The mathematical treatment of Volterra series in the previous sections is now clarified with a few, examples. 4.3.1 Basic second-order system. In this section we concentrate on a general block diagram representation of a second-o Volterra operator, which corresponds to the block Hh in Figure 4.1. It is clear that a seconc order operator performs at least one multiplication, since we stated that a second-order nonli earity combines two signals to produce a second-order signal. The most general second~ system with only one multiplication is shown in Figure 4.3. ko) 2a n(t) “ * y2() KO Figure 4 system. + Block-diagram representation that illustrates the operation of a simple sec ondt-ord With this simple second-order system the operation of a second-order nonlinearity can explained generally as follows: the incoming signal 2(t) is first fed to two linear blocks wi impulse response k,(t) and y(t), respectively, yielding the outputs z,(/) and 2,(t). The I signals are now combined by the multiplier block that produces a second-order signal, which turn is fed to a Jinear system characterized by the impulse response k,(t). The output of linear system is the overall output y2(t). The reader should be aware that the representation of Figure 4.3 is a general block-diagrag representation in which the different sub-blocks do not necessarily correspond to “physical” sul blocks in a practical nonlinear circuit. The block diagram has merely been used to explain h in general a second-order nonlinearity operates. It is not difficult to prove [Sche 80] that the (second-order) Volterra kemel of this gener system is given by halts / kelo)ka (11 — 0) ke (1» ~ 7) do arq ~ | In this book we are more interested in a frequency-domain representation where we make “9 of Fourier transforms. The Fourier transform of ha(r1,72). denoted as H(jwi, jw) is found after some algebra — to be aja) = Ka jn) Ko jee) Keli + Si) ag {4.3 Examples of Volterra kernels o where Ka(jw), Ky(jw) and K.(jw) denote the Fourier transforms of the linear subsystems de- scribed by the impulse responses k,(t), kp(t) and k,(t), respectively. Let us now consider some interesting simplifications of the general block diagram of Fig- ure 4.3. Assume that the the systems described by ka(t), ky(2) are not present, while k.(t) cor responds to a multiplication with a constant scale factor Kp. This leads to the simplified block diagram of Figure 4.4. This corresponds to a second-order nonlinearity without memory, the same as we used in Chapter 2 in our simple calculations. x) 0) | KS Figure 4.4: Block-diagram representation of a simple memoryless second-order system. Comparing Figure 4.4 to Figure 4.3, it is seen that the systems k,(t) and p(t) have been replaced by simple “through-connections”. The response of such through-connection to a unit impulse is the impulse itself, since this connection does not alter a signal at all. From system theory we know that the Fourier transform of a unit impulse is equal to one. Hence, from equa- tion (4.15) we find that the Fourier transform of the system of Figure 4.4 is given by Hy (js, jr) = Ko (4.16) This means that the second-order kernel of a nonlinear circuit without memory (no capacitive or inductive effects) for which the input-output relationship is described by a power series of the form y(t) = Kyx(t) + Ko(x(t))? + Ka(a(t))® +... 4.17) the second-order transform is Kp. 43.2 Basic third-order system Similarly to the previous section we consider now a general block diagram representation of a third-order Volterra operator. We will consider a block diagram with linear systems and multi- Pliers that multiply two signals at a time, as we did in the second-order system of Figure 4.3. It is clear that a third-order system will require two such multipliers. In this way a third-order sys- tem combines three signals to produce a third-order signal. The most general third-order system With only two multipliers is shown in Figure 4.5. As shown, x(t) is the common input to the Second-order system F, with the output z,(¢) and to the linear system with the impulse response a(t) and output z(t). The second-order system F,, is identical to the one shown in Figure 4.3. is system already combines two signals such that the last multiplier (in front of the linear sys- '™M with impulse response k¢(t)) effectively combines three signals. Once again we remark that70 Volterra series and their applications to analog integrated circuit designy 240 | kyo \ ‘ kO - x(t) Figure 4.5: Block-diagram representation that illustrates the operation of a simple third-onie system. the parts in Figure 4.5 do not necessarily respond to “physical” parts of a practical ea | nonlinearity. In [Sche 80] it is proven that the third-order kernel transform of this system is given by (Jur, ja, Js) = Ka( jor) Ko(jre) Ke(jor + jlrr) Kal js) Ke(jer + je + jus) Al Let us now simplify the block diagram of Figure 4.5 to a memoryless system. This is by replacing the blocks corresponding to ka(t), k(t), he(t) and ku(®) by through-connectio while the block that corresponds to k(t) is replaced by a block that performs a scaling with. factor K. The resulting block diagram is shown in Figure 4.6. With the same reasoning as 8 3D Figure 4.6: Block-diagram representation of a simple memoryless third-order system. the previous section we find that the third-order kernel transform reduces to Hs(jw, jr, jws) = Ks at4.3 Examples of Volterra kernels nm ‘This result could have been predicted with the knowledge of the previous section. 43.3 Application: a nonlinear amplifier We now analyze the circuit of Figure 4.7: it consists of a nonlinear amplifier with an RC circuit at its input and its output. Capacitive effects inside the amplifier itself are neglected. “ Olan K, a vv + y= Pour YO Figure 4.7; A nonlinear amplifier with an RC network at its input and output. ‘The input impedance of the amplifier is assumed to be infinite and the output impedance is zero. Since capacitive (and inductive) effects in the amplifier are neglected, we can describe the amplifier with nonlinearity coefficients that are independent of frequency. The relationship between the input voltage and the output voltage of the amplifier is given by Vout(t) = A vin(t) + Ka, (vin(t))” + Bay (in(t))? (4.20) Itis clear that the amplifier nonlinearity is memoryless. We will handle this example further in the frequency domain. With simple network analysis, we find that the ratio between the input voltage Vj, and the voltage Vj at the input of the amplifier is given by 1 Vin 1+ jwRiCy (4.21) With this result it is easy to see that the transfer function H, (jw) of the overall linearized circuit is given by Vou _ A Ay(j = oe Ge) = = Ce jeRiC) + ole) (4.22) We now consider the second-order kernel transform of the circuit. This kernel is nonzero due to the second-order nonlinearity of the amplifier. This second-order nonlinearity is fed by the Voltage over the capacitor C or, in other words, by the output of the RC network R,—C. The Second-order memoryless nonlinearity of the amplifier can be represented like the block diagram of Figure 4.4. Combining this representation with blocks that represent the RC networks at the Put and the output yields the configuration of Figure 4.8 from which the second-order response2 Volterra series and their applications to analog integrated circuit design| Ky }—= output T+ jOR,C, ” JORG RC network input —>| 11 RC network 2nd-order nonlinearity of amplifier i Figure 4.8: Block diagram used for the computation of the second-order response of the circuit of Figure 4.7. can be computed. Clearly, this block diagram can be transformed to the block diagram of Fig- ture 4.9, The operation of this diagram can be explained easily: the second-order nonlinearity of the amplifier squares the voltage over the capacitor to produce a second-order output signal. This second-order signal then propagates through the IRC network at the output. In this block scheme} the scale factor K,, of the second-order nonlinearity has been combined into one block withy the representation of the RC network at the output. This block diagram can be identified wit 1 T+ joR,C aN Ky output input =) OO] T#J0R,c, > ore 1 / ~ T+jOR,C, Figure 4.9: Equivalent of the block diagram of Figure 4.8. the block diagram of Figure 4.3. For the latter we can immediately write down the second-order kernel transform (jw, jw2), using equation (4.15): Ko, H: ny = oro om en des) = FRAC UF doaRsCi) (1+ Her +e) FCs) (4.23), Let us now consider the third-order kernel transform which is caused by the third-order non~ linearity of the amplifier. This third-order nonlinearity is fed by the voltage over the capacitor Cy. Using the general block representation of a memoryless third-order nonlinearity that is give in Figure 4.6, the block diagram from which the third-order response can be computed, can set up easily, as shown in Figure 4.9. This diagram can be transformed into the diagram of Figure 4.11. Identifying this diagram with the general representation of Figure 4.5, we can uExamples of Volterra kernels LL " 3rd-order nonlinearity of the amplifier Figure 4.10: Block diagram used for the computation of the third-order response of the circuit of Figure 4.7. input > 7+j0R,C, Figure 4.11: Equivalent of the block diagram of Figure 4.10. equation (4.18) to write down the third-order kernel transform Hs(ji1, juv2, jus) of the complete circuit: Hy (juny, jury, js) = Koa Fea F8) ~ GSTS RO) (1 + jal Ci) (1 + JwshaCr) (1 + Hea + oe + wos) C2) (4.24) From the second- and third-order kernel transform we can now compute the response to a sinusoidal signal. To this purpose we take a sinusoidal input signal vin(t) Vin C08 (wet) (4.25)74 Volterra series and their applications to analog integrated circuit desis First-order response The first-order response y; (t) can be found by combining the expressio of the linear transfer function, equation (4.22) with the general expression of a linear system to sinusoidal response, equation (4.13). This yields Vin « A t) 008 (wrt — atan (WRC) — atan (we R2C2)) (4.26) anit) 1 wiRIC? 1 RCE ( (weRiC1) ( Cr) ¢ ‘The ~ symbol in equation (4.26) is due to the fact that the response at the frequency w, is m only caused by the linear behavior as indicated in this equation, but also by third-order an higher-order behavior as we saw before. Second-order response For the computation of the second-order response 1j2(t) to the sinu: soidal signal of equation (4.25) we combine the general equation for the response of a second! order system to a sinusoidal signal (equation (4.11)) with the second-order kernel transform the circuit (see equation (4.23). This yields after some algebra in’ Ko walt a 2(1 + w2 RIC?) /1 + 402 R3CF 4 aly 304 RC?) Clearly, this response is seen to consist of a DC shift and a second harmonic. The magnitude the two components depends on the frequency w,. Note that the DC shift is independent of RC; ‘The reason is that this DC shift is a signal that is produced by the second-order nonlinearity the amplifier. This signal propagates at 0Hz through the RC network at the output withot attenuation. ‘The second harmonic distortion can now be found by taking the ratio of the response at 2g and at ws. This yields + [L + cos (Quiet — 2atan(w.,.R,C\) — atan(2w,RCy! (42 Vin Ko 1 1+ jwsRoC> | i H. fim B24 | | IB 4.28} Pea laa 1+ j2w,RsCo G28 __. (4.2 14+ w2 RIC? i It is seen that the second harmonic distortion depends on the normalized second-order nonli arity coefficient A> ,/A of the amplifier. Further it is seen that the second harmonic distort decreases with increasing frequency. This is due to the fact that the second harmonic decre: faster with increasing frequency than the fundamental response, since the former contains square of (1 + jwC; in the denominator, while the latter contains only the first power (1+ jweRtsC)) in its denominator. In addition, the second harmonic is a component of a secon order signal that is produced by the second-order nonlinearity and propagates at a frequency 2 to the output, whereas the first-order response is due to the input signal that propagates at a quency w, in a linear way to the output, Ifw, > 1/(J22C3) then the ratio of the second harmonit and the first-order response is a factor of 6dB smaller than at low frequencies.