eee Panel PEIDNEAS VN3g-Z3HINYS, ria ° & RosBust SYSTEMS nme | ae sta) Dar ACA ane Waren ochre eon ee es Sea) een eer eee eat phe iene j Leetonia) ener ane A n See) Korte et ae eee ara Mario SZNAIER WY ciincreasingly important subject Co ee Le ae See Pee enn een ees CO eae ee es are complemented by elaborated examples and a group of worked: De eee reac et nee Ce ee ee eon tet Ce eet eed ee) ‘expressly a5 textbook for master’s and first-year PhD ete REINA d-ZaHONYS TNE) ces basic robustness concepts inthe context of SISO systems described by Laplace nets eet ke eae ee ee ee Se eee ea a ene ee] oad es the four basic problems in robust contol and the Loop shaping design method ROBUST SYSTEMS THEORY ano APPLICATIONS ee ee ee Ce eee ee) = ee ee ee ee Te Se ee eC ee snr aes oe er and Le eee) ek ke ee eee s the basic tools of model order reduction Se eee ‘numerous endot-chapter problems and workec-out examples of robust contol See cd Activities (CONAE) and Professor of Contol Systems atthe School of Engneesng at eee ee een eee ere Kel S STORES eee a ering at Pennsylvania State University, University Patk, USA Mario SZNAIER WON wi WILSAS 1sngoyROBUST SYSTEMS THEORY AND APPLICATIONSAdaptive and Learning Systems for Signal Processing, Communications, and Control Editor: Simon Haykin Werbos / THE ROOTS OF BACKPROPAGATION: From Ordered Derivatives to Neural Networks and Political Forecasting Kislié, Kaneliakopoulos, and Kokotovié / NONLINEAR AND ADAPTIVE CONTROL DESIGN Nikias and Shao / SIGNAL PROCESSING WITH ALPHA-STABLE DISTRIBUTIONS AND APPLICATIONS Diamantaras and Kung / PRINCIPAL COMPONENT NEURAL NETWORKS: THEORY AND APPLICATIONS Tao and Kokotovié / ADAPTIVE CONTROL OF SYSTEMS WITH ACTUATOR AND SENSOR NONUNEARITIES Tsoukalas / FUZZY AND NEURAL APPROACHES IN ENGINEERING Hryce / NEUROCONTROL: TOWARDS AN INDUSTRIAL CONTROL METHODOLOGY Beckerman / ADAPTIVE COOPERATIVE SYSTEMS Cherkassky and Mulier / LEARNING FROM DATA: CONCEPTS, THEORY, AND METHODS Passino and Burgess / STABILITY OF DISCRETE EVENT SYSTEMS SGnchez-Pefia and Sznaier / ROBUST SYSTEMS THEORY AND ‘APPLICATIONS Vapnik / STATISTICAL LEARNING THEORY ROBUST SYSTEMS THEORY AND APPLICATIONS Ricardo S. Sanchez-Pena Mario Sznaier ® ‘A Wiley-Interscience Publication JOHN WILEY & SONS, INC. New York / Chichester / Weinheim / Brisbane / Singapore / Toronto‘This book i printed on aides paper. @) Copyright © 1998 by John Wiley & Sons Ic. Allright reserved Published simakancousy in Canada No past of this publication may be reproduced stored in revival system or tranamited in any form or by any meas, cto, mechanical. photocopying. record. eamnng cpherwise, except as permitted under Section 107 oF TOR ofthe 1976 United Sites Copyright Drive, Danvers, MA 01923, (978) 750.840 fax (978 7504744 Requents to the Publades lor Permission should be adiressed 10 the Permissions Department ke Wie) & Seon fe (05 Third Avenue, New York, NY 101580012, (212) 8504011 fax 212) 850e008 E-Mail: PERMREQ@WILEY COM, Library of Congress Caaloging-in Publication Data: Sinches Pena. Ricardo boom Includes bibliographical references and inde. ISBN 0471176073) |. Automatic control 2. Control theory. 3. System ‘entiation. L'Sensier, Mario Tile TAISTISe 1998 e9s-de2I 9.6820 c Printed inthe United States of Ameria WORTH SE TIT To our parents Lolita and Miguel, Rebeca and MarcosCONTENTS | Introduction 1 1.1 General Control Problem / 1 LL Experimental Phase / 2 11.2. Simulation Phase / 3 113. Theoretical Phase / 5 1.2. Why Feedback? / 11 13 Feedback Loop Trade-olf / 14 14 Objectives of an Applied Theory / 17 15. Objectives of This Book / 19 15.1 General Remarks / 19 152. Scope / 19 153. How to Use This Book / 21 SISO Systems 23 21 Introduction / 23 22 Well Posedness / 24 23. Nominal Internal Stability / 26 24 Robust Stability / 29 24.1 Phase and Gain Margins / 29 242 Global Dynamic Uncertainty / 32 25. Nominal Performance / 42 25.1 Known Disturbance/Noise/Reference / 42 252 Bounded Disturbances at the Output / 44 253 Other Performance Criteria / 47 26 Robust Performance / 50 2.1 Extension to MIMO Systems / 55 28 Problems / $73 CCONTENIS Stabilization 6 3.1 Introduction / 61 3.2 Well Posedness and Internal Stability / 63 32.1 Well Posedness / 63 32.2 Internal Stability. / 64 33. Open-Loop Stable Plants / 65 34° ‘The General Case / 68 34.1 Special Problems / 69 342 The Output Feedback Case / 75 35 Controller Structure and Separation Principle / 78 3.6 Closed-Loop Mappings / 79 37 A Coprime Factorization Approach / 80 37.1 Coprime Factorizations / 80 38 LFTs and Stability / $8 3.9 Problems / 90) Loop Shaping 93 4.1 Introduction / 93 42 Nominal Pesfosmance 97 43° Robust Stability / 98 44 Nominal Performance and Robust Stability / 99 45, Robust Performance / 100 45.1 Sensor Uncertainty / 101 45.2 Actuator Uncertainty / 102 46 Design Procedure / 105 4.7 Examples / 109 47.1 Permanent Magnet Stepper Motor / 109 47.2 Loop Shaping Q(s) / 119 48 Related Design Procedures. 123 49° Problems / 125 %_ Optimal Control 127 5.1 Introduction / 127 5.2. The Classical Linear Quadratic Regulator Problem / 128 53° The Standard 4, Problem / 133 54 Relaxing Some of the Assumptions / 141 5.5 Closed-Loop Properties / 142 55.1 The LOR Case: Kalman’s Inequality / 143 CONTENTS x 55.2 Some Consequences of Kalman’s Inequality / 145 553. Stability Margins of Optimal 4» Controllers / 146 56 Related Problems / 149 57 Problems / 152 Hoo Control 157 6. Introduction / 157 62. The Standard 4, Problem / 158 62.1 Background: Hankel and Mixed Hankel-Toeplitz Operators / 161 62.2 Proof of Theorem 6.1 / 176 63. Relaxing Some of the Assumptions / 179 64 LMI Approach to H. Control / 180 64.1 Characterization of All Output Feedback Controllers / 183 642 Connections with the DGKF Results / 188 65 Limiting Behavior / 192 66. The Youla Parametrization Approach / 193 67 Problems 203 Structured Uncertainty 207 7.1 Introduction / 207 TAA Stability Margin / 210 72. Structured Dynamic Uncertainty / 215 72.1. Computation / 215 72.2. Analysis and Design / 219 73. Parametric Uncertainty / 223 731 Introduction / 223 732. Research Directions / 225 733. Kharitonov's Theorem / 230 734 Mapping Theorem / 233 74. Mixed Type Uncertainty / 237 74.1 Introduction / 237 742 Mixed w / 239 75 Problems / 242 Control 248, 8.1 Introduction / 24510 ‘CONTENTS 82. Robust Stability Revisited / 247 82.1 Robust Stability Under LTV Perturbations / 250 82.2. Stability Under Time Invariant Perturbations / 253 83 _A Solution to the SISO ¢' Control Problem / 254 83.1 Properties of the Solution / 262 83.2. ‘The MIMO Case / 266 84° Approximate Solutions / 267 84.1 An Upper Bound of the é Norm / 268 842 The + Norm / 269 843. Full State Feedback / 273 844 All Output Feedback Controllers for Optimal «Norm Problems / 276 85. The Continuous Time Case / 282 85.1 Solution Via Duality / 282 852 Rational Approximations to the Optimal £, Controller / 285 8.6 Problems / 288 Model Order Reduction 9.1 Introduction / 293 92 Geometry of State-Space Realizations / 295 92.1 Controllable/Unobservable Spaces / 295 922. Principal Components / 297 93 Hankel Singular Values / 304 93.1 Continuous Systems / 304 93.2 Discrete Systems / 306 94 Model Reduction / 310 94.1 Introduction / 310 9.4.2 Hankel Operator Reduction / 310 943 Balanced Realizations / 312 944 Balanced Truncation / 314 95 Algorithms / 318 95.1 Approximation Error / 319 9.6 Problems / 321 Robust Identification 10.1 Introduction / 323 102 General Setup / 324 293 323 103 104 105 contents xi 02.1 Input Data / 325 10.22 Consistency / 326 102.3 Identification Error / 328 10.24 Convergence / 334 102.5 Validation / 338 Frequency-Domain Identification / 339 103. Preliminaries / 340 1032 Sampling Procedure / 342 1033 Consistency / 344 10.34 Identification Procedures / 346 ‘Time-Domain Identification / 361 1041 Preliminaries / 362 10.42. Idemification Procedures / 366 Further Research Topics / 372 105.1 Unstable Systems / 372 1052 Nonuniformly Spaced Experimental Points / 372 1053. Model Reduction / 373 1054 Continuous Time Plants / 373 155 Sample Complexity / 373 105.6 Mixed Time/Frequency Experiments / 374 105.7 Mixed Parametric/Nonparametric Models / 374 Application Examples 377 ie n2 u3 na SAC-C Attitude Control Analysis / 377 141 Introduction / 377 11.4.2 Linear Model / 378 11.13 Design Constraints / 381 11.14 Robustness Analysis / 383 115. Simulations / 385 Controller Design for a DO Plant / 389 11.2.1 Model of the Plant / 389 11.22. Robustness Analysis / 391 11.23 Controller Design / 397 X-29 Parametric Analysis / 400 113.1 Linear Model / 400 1132 Results / 405 Control of a DC-to-DC Resonant Converter / 407 11.4.1 Introduction / 407xl CONTENTS CONTENTS xi 11.42. The Conventional Parallel Resonant Converter / 407 BO Inverse / 469 1143 Small Signal Model / 409 B.10 Linear Fractional Transformations / 470 1144 Control Objectives / 410 BAL Norm Computations / 475 1145 Analysis of the Plant / 411 B.IL1 7% Norm Computation / 475 1146 Control Design / 414 B112 Me Norm Computation / 476 1147 Controller Synthesis / 420 B13. € Norm Computation / 476 1148. Simulation Results / 421 B.I2 Problems / 479 Bibliography 427 ¢ Riccati Equations ar ‘A Mathematical Background 445 Index 487 Al Algebraic Structures / 445 ALI Field / 445 A.L2 Linear Vector Space / 446 AL3 Metric, Norm, and Inner Products / 447 ‘A2 Function Spaces / 449 21 Introduction / 449 22. Banach and Hilbert Spaces / 449 23. Operator and Signal A24 Isomorphism / 452 A25 Induced Norms / 453 A26 Some Important Induced System Norms / 455 AB. Duality and Dual Spaces / 457 A31_ The Dual Space / 457 32 Minimum Norm Problems / 458 A4 Singular Values / 460 AAI Definition / 460 AA2 Properties and Applications / 461 paves / 450 B System Computations 465 BI Series / 465 B2 Change of Variables / 466 B3_ State Feedback / 467 Ba ‘timation / 467 B.S Transpose System / 467 B.6 Conjugate System / 467 BT Addition / 468 BS Output Feedback / 468 tatePREFACE Robustness against disturbances and model uncertainty is at the heart of control practice, Indeed, in the (completely unrealistic) case where both all external disturbances and a model of the system to be controlled are exactly knowin, there is no need for feedback: Optimal performance can be achieved with an open loop controller. The main ingredients of present day robust control theory were already present inthe classical work of Bode and in many popular frequency domain: based design techniques. With the advent of state-space methods in the mid 1960s the issue of robustness took a backseat t0 other topics, but was never completely abandoned by the contzol community. especially control practi tioners, Interest in robust control rose again in the late 1970s where it was shown that many popular control methods (including optimal LOR control and controller design based on the cascade of an observer and state feedback) Jed to closed-loop systems very sensitive to model perturbations, Moreover, widely accepted “ad-hoc recipes for “improving robustness.” such as using "nifiiallylarge noise levels in the design, had precisely the opposite effect. Robust control has undergone extensive developments in the past «wo decades, leading to powerful formalisms, such as 7, w-synthesis/analysis and, more recently "optimal contro, that, coupled with newly developed control-oriented identification techniques, have been successfully applied to challenging practical problems. A salient feature of the framework is that itis oriented towards applications and thus is based on “practical,” realistic assumptions “There are many excellent books that cover specialized topics (Hz, Ma ©, parametric uncertainty linear matrix inequalities) with others scattered in the technical journals, Our intention in writing this book is to provide a self contained overview of robust control that illustrates all the issues involved, ranging from the transformation of experimental signals from the physica plant toa set of models (robust identification), to the synthesis of a controller {or that set of models (robust control), The purpose of the book is twofold to serve as a textbook for courses at the Masters/beginning Ph.D. level and as a refesence for control practitioners. It assumes that the reader has a background inclassial and state-space control methods. In order to keep thexvi PREFACE text size at a manageable level, in some cases only basic results ate covered, and the reader is referred to more specialized literature for further coverage In all cases we have strived to link the theory with practical applications. To this end, in addition to the examples covered throughout the book, the last chapter contains several worked out application problems that stress specific practical issues: nonlinearities, unknown time delays, infinite dimen- sional plants, actuator and sensor limitations. They are all extracted from our practical experience in different engineering fields. Furthermore, due to the fact that most of the problems to be solved by the theory presented here are computer-intensive, we stress the algorithmic and computational aspect of the solution along with the mathematical theory. We (and our graduate students) have tried to eliminate obvious mistakes, However, as anyone who has tried knows, itis virtually impossible to make 1 “perfect” book. We encourage readers to send corrections, comments and general feedback to either one of the authors. Finally, as altested by the list of more than 300 references, we have tried to give credit where it is due, However, owing to the sheer volume of literature published on robust control, we may have inadvertently failed to do so on occasion, We apologize in advance to readers or authors who may fee! that this is the case, and we encourage them to send us comments, ACKNOWLEDOMENTS We owe a special debt to Professor Thanasis Sideris from whom both of us learned a great deal (not only about robust control) and who was very influential in shaping our subsequent research. We would also like (0 a: knowledge the important contribution that Professor Manfred Morari had in the development of the concepts presented in this book, and in advising Ricardo S, Sanchez Pea to make the right decisions during hard times in his career. Mario Sznaier is specially indebted to Professor Mark Damborg. for his mentoring. Among Professor Damborg’s many contributions, the two that stand out are attracting him to the field of control and making sure that he got a Ph.D. degree when he showed signs of enjoying life too much as a sraduate student The reviews by Professors Peter Dorato and Roy Smith at initial stages « of this book have been instrumental in encouraging the authors to carry out the project and in shaping its final form. Preliminary versions of the book were tested in courses offered by the au- thors at Penn State University and the University of Buenos Aires, Feedback from students enrolled in these courses was fundamental to its fine tuning We are also indebted to our graduate and post-doctoral students, Takeshi ‘Amishima, Pablo Anigstein, Juanyu Bu, Tamer Inanc, Cecilia Mazzaro, Pablo Parnlo and Zi-Qin Wang, for reviewing the book in great detail and provid- ing some of the material, In particular, Cecilia developed very efficiently the PREFACE xv ‘examples in Chapter 10 and Takeshi, Juanyu, and Zi-Qin contributed some of the material in Chapters 8 and 11 ‘The book was shaped to a large extent by interactions with our colleagues. In particular, many of the ideas and developments in Chapter 8 arose from research carried out jointly with Professor Franco Blanchini. Dr. Hector Rot- stein influenced the presentation of the, material in Chapter 6, Profes- sors S. P, Bhattacharyya, D. Bernstein, N. K. Bose, O. Crisalle, M. Dahleh, M.A. Dabileh, J. C. Doyle, M. Fan, C. V. Hollot, M. Khammash, A. Megret ski, A. Packard, J. Boyd Pearson, A. Ray, A. Saberi, A. Stoorvogel, R. Suarez, R. Tempo and J. Vagners contributed (directly or indirectly) to this book through many discussions and comments during the years. ‘The first author would like to thank his colleagues at LAE, School of Engineering of the University of Buenos Aires and at the Argentine Space ‘Agency (CONAE), especially Professors C. Godifrid and A. Perez, and Ing. Beto Alonso for important support. The financial support of CONAE, the University of Buenos Aires and Fundacién Antorchas during his stays at Penn State in 1994 and 1996 are also gratefully acknowledged, “Mario Sznaier would like to thank the National Science Foundation, and in particular Dr. Kishan Baheti, Director of the Engineering Systems Program, for supporting his research program. This book would have not been possible ‘without this support We are also indebted to Ms. Lisa Van Horn and Mr. George Telecki at John Wiley and Sons, Inc. for their assistance throughout this project. Finally we would like to thank our children, Gadiel, Lucila and Pablo, and spouses, Ménica and Octavia, for their infinite patience, support, and understanding for the time this book has stolen from them. . Ricanpo $. SANCHEZ-PESA Manio S2NAieR Buenos Aires, Argentina University Park, PennsylvaniaROBUST SYSTEMS THEORY AND APPLICATIONSINTRODUCTION 1.1 GENERAL CONTROL PROBLEM The purpose of this book isto introduce the reader to the theory of control systems, with particular emphasis on the applicability ofthe results to pract- cal problems. Therefore it considers not only the control viewpoint but also the ‘denifcation of systems as part of the problem. Also, as for any theory of systems oriented toward practical applications, robustness is essential and will be the underlying concept throughout the book, For these reasons the “tle of the book includes the following words: Theory, Applications, Robust, and Systems. It is important to start with a wide angle view of the general control picture, so that we can locate ourselves in the particular approach focused by this book. For this reason, the general problem of implementing a con- troller, which aetuates over a physical process in a prescribed way, will be divided into three well defined phases. These phases are all equally impor tant but involve very different echnical cools, which go from pure mathe- ‘matics to microprocessor technology. In today’s practical control problems, all phases should be considered and itis very difficult to have a clear knowl edge of all the technical fields involved. Although in this book we con centrale on the theoretical and computational aspects of the whole problem, we first make a general overview of the disciplines involved at teach stage and the interrelation among them. This is the basic objective ‘of robust identification and control theory: to produce theoretical results that are closely related to the computational and experimental aspects Of the control problem. Tn the next subsection we will describe each phase in detail2 INTRODUCTION 1.1.1 Experimental Phase the control problem is generated and where the last test ‘on the controller is performed. This means that the actual control problem to be solved comes always from a physical setup: an aircraft, a distillation plant, a robotic manipulator. The performance specifications are also formulated in terms of variables involved in this physical experience: stable dynamics at high angle of attack, constant concentration of the distilled product, fast response to commanded trajectories. Also, in many cases, once the practical problem has been formulated, an experiment should be performed to:obtain ‘a mathematical model as “close” as possible to the real physical system to be controlled. Finally the last test to prove the effectiveness of the controller is performed on the physical system, ‘The experimental phase is basically technological. It involves engineer- ing disciplines (electrical, mechanical, chemical, acronautical, ete.) of even biology or economics, depending on what we define as the system to Be con ‘rolled. It also involves knowledge on the sensors, actuators, and technical aspects of microprocessors and 1/O devices, for computer control. We may as well include in this phase the low- and high-level programming, as in the case of communication programs in Assembler and control algorithm coding, respectively. ‘The interaction of the physical plant with its environment is performed by “real world” signals provided by the plant to the sensors and application ‘of physical stimulations to the plant through the actuators, which modify its behavior. These input and output signals, as well as the physical laws that describe the system behavior, provide the only information we have of the physical plant. In many cases there are no physical laws that explain the plant behavior or they are extremely complex. A trial and error contol, or direct experimentation using the input and output signals (without any analysis), is untealizable, at least for the class of problems in which we are interested. ‘A more elaborate and “scientific” procedure should (and can) be performed {0 control a given plant. Furthermore, in many cases trial and error would not even be practical because of costs (chemical plants) or simply to prevent accidents (aircraft, nuclear power plant). ‘The information on the system—input-output signals and physical laws— is never complete (to complicate things). The reason is that the physical laws, which may be applied to the (sensor, actuator, plant) behavior describe only’. 4 certain aspect of the problem, but not all. To make this point, consider the following example. To describe the output voltage on a given resistor through which a given current is circulating, we may use Ohm's law. Nevertheless, the temperature, gravitational field, and magnetic field could also influence the ‘measurement, as well as the physics of the voltage sensor and current supply. A description of a particular physical process using all the many aspects involved could, to say the least, be impractical.' Sometimes, also for practical is an epistemological question sf complete mathematical picture of physical rely GENERAL CONTROL PROBLEM 3 reasons, a simplification of the physical laws should be applied to describe the system. Actually, the laws are not modified, but instead new simplifying assumptions are added The above facts determine that all practical control problems have a cer lain amount of uncertainty in its more general meaning. It should be noted that this uncertainty is a consequence of the observation of the system and its environment, not due to the system itself. In particular, in the control jargon, this lack of knowledge is called noise when applied to the sensing, proce: dure, disturbance for the external stimulations to the plant, and uncertainty when it applies to the description of the dynamical behavior of the plant by ‘a mathematical model. These three elements—in particular, the latter—play 4 fundamental role in the orientation of control theory since the 1980s and in the motivation for Robust Identification and Control. For many different reasons it is highly convenient to have a computational ‘model of the plant, By this we mean a representation, in a computer, of the relevant inputs and outputs of the physical system and their relationship. This representation can be very elaborate as in a hardware in the loop scheme, where parts of the system are represented computationally and others are ‘connected physically to the computer I/O ports? Otherwise it can be a com- plete computational model of the plant obtained from a mathematical model that includes all relevant physical phenomena involved “To illustrate some of the many reasons justifying a simulation of the plant, we cite the following “real world” examples, which will be described in greater detail in Chapter 11 |.2 Simulation Phase 1, In many cases direct experimentation with the hardware is impossible, as in the case of satellites. In these situations, itis usual to start with a high-fidelity simulation, which includes models of the system (satellite), actuators (inertial wheels, gas jets, magnetic torque coils), sensors (s0- lar and Earth magnetic field sensors), and environment (Sun Ephemeris and Earth magnetic field model along the orbit). These models include the nonlinear dynamics, flexible phenomena, and all the necessary ele- ‘ments to represent the system. According to required tests to be per formed along the project, pieces of real hardware may be connected to this hi-fi simulator replacing their simulations, that is, sensors or actua- tors. 2. The cost of experimentation could be a strong reason to simulate a plant, ike the case of many chemical processes, where the startup of 1, could he argued that in thi case we fave » combination of experimental and simulation phases, What we mean hy simulation in ths book i im genera an eect that isnot exaetly the eal plat or envionment In any case. the important point st realize that there ae diferent pasts the onto problem,4 IntRODUCTION the process could involve many millions of dotlars. The high costs could be a consequence of very expensive materials or long experimentation, ‘Testing over a computer model provides a very fast and inexpensive way of verifying a controller (see DzO plant in Chapter 11 and [228)), 3. A very important reason for simulation is when human lives are in volved, as in the case of an experimental aircraft, Therefore a com- puter simulation gives a safe means of testing high-risk maneuvers (see NASA's X-29 experimental aircraft in Chapter 11), The disciplines involved are related to the areas of computer science, op- timization, and numerical methods. The mathematical models to be imple- mented by a computer code could be obtained from physical laws or from direct experimentation on the plant. In this last case, the classical Parameter Identification procedures ({184, 293)) or the very recent Robust Identification ((146)) techniques (see Chapter 10) could be used. The main goal of the simulation phase is to have a copy as “close” as possible to the experimental plant. The word experimental has been empha- sized to distinguish it from the true plant, which is not the same concept, as ‘mentioned in the previous subsection and detailed next. For simulation, we are interested only in certain aspects of the physical system but not in all of them. Take, for example, the dynamics of the longitudinal axis of a rigid aircraft, which is a partienlar aspect ta he considered not at all related to the ‘thermal gradients of the wings of the same aircraft. Therefore, for simulation purposes, we only consider certain input and output signals that are relevant ‘o our problem. Furthermore, even for a particular aspect of the problem, we may consider different levels of detail. Take the same example, where in cone case we may consider the aircraft as a rigid body and in a more detailed analysis we may add the flexible modes. Therefore, for simulation purposes, we only consider the experimental plant, which is a particular aspect of the ‘rue plant, that arises in the experimental setup. ‘The trade-off in this phase of the problem is defined in terms of how “close” we want the computational model to represent the plant versus how “big” a computer will be needed to perform this task. In the latter, the main limitations are computation time and available memory. As the computa- tional model describes the plant better and better, the computation time will grow, which could pose a serious limitation in many real-time applications. For example, even for reasonable size minicomputers, a complete modef of the orbit, Earth magnetic field, Sun Ephemeris, gravitational and acro- dynamic disturbance models, and satellite dynamics could take much more time for simulation than for real time. This is the case of low-orbit satellites with short orbital periods (SAC-C in Chapter 11 has a 90 minute period), Also, in many cases. a “good” computational model could involve a great amount of memory, like in many finite element representations of physical systems, Im any case, this trade-off determines that certain simplifying assumptions should he made when representing a physical system by a computer code, GENERAL CONTROL PROBLEM 5 Typical assumptions for the nonlinear model of an aircraft are rigid body dynamics and symmetry. The simplifying assumptions mentioned above, in the simulation phase, are only related to the computational constraints. In stead, for the theoretical analysis and synthesis phase, new hypothesis should be added to simplify the problem even more. This is related to the available mathematical tools to design and analyze a feedback loop, as will be seen next 1.1.3 Theoretical Phase Introduction In history and, in particular, in the history of science, humans have tried to solve first the simpler problems and continue with the more complicated ones. This, which seems almost a natural law of human behavior, also applies to the theory of control systems. For this reason in classical con- trol theory, systems were first described as single-input single-output (SISO), ‘while later on, the more general multiple-input multiple-output (MIMO) de- scription was introduced by modern control theory through the state-space ‘description. In the same line of thought, the optimization of a quadratic func- tional (LOG optimal control) appeared earlier than the Ha: optimal control problem. In this latter case, the linear 1. control problem was stated ((329]) flmost ten years before the first solutions (|22, 157, 309, 310)) to the Ax nonlinear control problem were produced. At this point, there are not yet fully developed analysis and design tools to deal with control systems described by coupled nonlinear partial differ tential equation descriptions with (possibly) uncertain parameters. Therefore the theoretical phase of a control problem, which will be the main concern .of this book, deals with a much more simplified mathematical model of the plant. Historically, the most usual system description for design and analysis of feedback systems has been finite-dimensional (lumped), linear, time invari- ant (FDLTI) mathematical models. For this class of models there is a well established theory, which also covers many of the most usual applications? ‘A visual idea of all the simplifying assumptions the designer needs to make to obtain a FDLTI mathematical model from the physical system may bbe seen in Figure 1.1. The usual steps that should be taken to compute a FDLTI model, from a very complex mathematical representation of a partic- ular plant, are presented there, Some applications, which will be detailed in Chapter 11, illustrate a few of the steps. System Model _At this point we will distinguish two different approaches that compute a FDLTI model from a physical system. The first one obtains ‘4 simplified version of a very complex mathematical model. This could be a set of partial differential equations or a time delayed or a high order set of ‘ordinary differential equations. This complex mathematical model is usually "ot hecause physical systems ae themselses FDL ho Dect for moet practical purposes6 ntRoDUCTION Physical System (Cavs of Physics)/(Tdentification) | ed] | Nonlinear O.D.E. } 4 Nonlinear P.D.E. Linear PDE. Ll | LC | (Gtotat (ocat linorisation) linearisation). {Ayprosimation) EO) Eo Physical System = FDLTI Model + Uncertainty Figure 1.1, “Transtaton” ofa physical system to a FDLT model ‘obtained from physical laws describing the system. This model simplification procedure is called approximation and is purely mathematical, once the model Of the system, obtained from physical laws, is known, GENERAL CONTROL PROBLEM 7 ‘There are a variety of methods that approximate systems described by partial differential equations (infinite-dimensional systems) by ordinary dif- ferential equations. Among, others, we can mention finite elements, finite dif- ferences, averaging techniques, splines, and FFT methods ((24, 25, 139, 169)). ‘More recently, the balanced realization and truncation of infnite-dimensional ‘models provides simplified finite-dimensional models ({79, 134)). In the case of systems with time delays, which are also infinite dimensional, although the above methods can be used, the most usual and practical approach is to use Padé approximants. TThe second approach is based on direct experimentation over the physical system and is called in general identificarion. This approach consists in com- puting a mathematical model, usually FDLTI, from the data measured at the system's output when a particular input signal is applied. The classical param: ‘eter identification procedures (|184, 293)) assume a particular mathematical structure for the model and, through an optimization method, obtain a set of parameters that minimize the “difference” between the model and the system's output, for the same applied input. More recent techniques called robust identification obtain ditectly a FDLT1 model, without assuming any parameter structure, from the output data of the system when a particular Sel of signals are applied to the input. These latter techniques are related di rectly to robust control and will be described in greater detail in Chapter 10, From Figure 1.1 we observe that sometimes we shoul linearize a nonlinear ‘model of the system. The classical linearization technique consists in comput- ing a linear version of the nonlinear model for a particular set of parameters (operating point) of the latter. This is called a local model ({189)), which is approximately valid only in a small neighborhood of the selected values of these parameters. In some applications, the parameters have fixed values and, “ for that reason, there is only one working point and one linear model ((268)) Otherwise, ifthe parameters can take different values, the designer should synthesize a linear controller for each linear model. Nex, a test should be made (analytical, by simulation or experimentation) to guarantee a smooth transition from one controller to another (see Example below). This is called {ain scheduling and is a well known procedure to control nonlinear plants Recent results in this area attack this problem through the use of linear parameter varying (LPV) controllers ([222)) ‘Ako, for certain classes of nonlinear models (linear analytic models), a state transformation and a nonlinear state feedback can be computed, which produce a (global) linear mode! ((89, 153, 56)). Although this is not @ robust result! it is very useful in many practical applications ([1, 2, 44, 106, 107)), for example, robotic manipulators, aitcraft/spaceeraft control, and motor control. “The state-feedhack lw, which linearis the mode, depends on exact knowledge of the non linear modck therefore Fedhack lineation wa robot against uncertainty inthe onlinest8 introDucTON Disturbances The designer not only assumes certain properties ofthe sy ‘em (e., linearity, time invariance) to obtain a simpler model but also needs to make certain hypotheses on the disturbances applied to the system by is environment. In classical and modem control, the disturbances were usually assumed to be certain classes of signals: steps, ramps, impulses, sinusoids, In 4 stochastic approach, the covariance matrix of the stochastic proce’s repre- senting the disturbance was assumed to he known exactly. Very seldom does the designer have a clear knowledge of the disturb- ing signals acting on the system. Therefore the less one assumes about their nature, the higher the possibility of producing an elfective control, despite these disturbances. As a matter of fact, a control design procedure that op- timally rejects a step disturbing signal will no longer be optimal if the actual disturbance is a square wave. For this reason, in robust control, the only assumption on the disturbances is the fact that they are bounded (usually in energy), which is a very realistic hypothesis. Ifthe designer has some extra in- formation on the frequency distribution ofthe disturbances, that information could be incorporated into the design procedure and produce a less conser- vative result. Note that even in tis eas, there is no assumption on the form of the disturbance Uncertainty Due to the assumptions on the system and the disturbances, a fair amount of uncertainty between the mathematical model and the exper ‘mental results will appear. Inthe simulation phase, the simplifying hypotheses fre made due to computational restritions (time, memory). Iasead, in the theoretical phase, the assumptions that generate the uncertainty ate a con- sequence of the mathematical tools thatthe analyst/designer has decided to ‘We can evaluate separately the effect on the feedback loop of the uncer- tainty due to the modeling ofthe system and due to the external disturbances. From_basic contro courses we know that the latter has no effect on the sta- lity of the loop, at least in linear systems,° although it can degrade the performance. On the other hand, the unce ¢ the design of a stabilizing controller has been based on an erroneous plant ‘model, the loop can be unstable. This is clear from the basics of control systems, as in the case of models with uncertain gains or phases. Take, for example, a controller designed for a model of the plant with a gain that eX- actly doubles the gain of the actual plant. If this model has a gain margin smaller than 2 (and the designer doesnt know it), the controller could make the loop unstable, Sn nantinear systems, bath the uncertainty ia the model and the umbnown dsturances can Ustahiive the enedonp system, GENERAL CONIROL PROBLEM 9 Surprisingly, mode! uncertainty has only started to be considered quantita Lively in the assumptions of control theary since the 1980s. This recent theory, which considers uncertainty as part of the problem, is called robust control and, together with robust identification, is the main subject of this book. ‘Throughout its chapters, the following fact will be stressed: the design of a controller should be based not only on a particular model, but also on the un certainty this model has with respect to the actual physical plant, Uncertainty is related directly to feedback, as will be seen in Section 1.2. Example The three phases described above are closely related and equally important. The final solution of the control problem, from the purely specu- lative theoretical step to the final physical implementation of the controller ‘of the plant, depends on all of them. Each phase has its own characteristics Although in the experimental phase there are no approximations, it is not possible to elaborate any control strategy directly from the input and output signals. The computational model used in simulations is always an intermedi- ate step, because it already contains simplifying assumptions, which limit its applicability. Finally, no matter how strong (necessary and sulficient) the con- S(s) as) ‘This variation may represent uncertainty in the modeling Ifthe variation is not infinitesimal, we can compute the relation between open- and closed-loop ‘models considering the influence ofa signal us(s) on the output of the plant ‘model y(s). These signals are related to the relative uncertainty of the loop as follows: ts(o) = BOs 65)16 IwiroouctiON shee the uncertain loop (1+ L(s}/L(s)] Lt). Inside the feedback toop both signals are related by AL(s) vals) = T45)-0g |, SEED ) (0) = Teo) [FE (6 This interpretation of model uncertainty will be explained in greater detait in Chaptets 2 and 4 From their definitions, it is easy to see that S(s) and T(s) satisfy the following equation [TOssO=T (17) where 1 is the identity matrix. This equation places a serious constraint when designing a controller that should guarantee stability and performance, as well as robustness to model uncertainty. According 10 equations (1.3) and (1.4), for disturbance rejection at the ‘output, to minimize the influence of reference and disturbance in the track- ing error, or the (infinitesimal) robustness problem of equation (1.5), the designer should minimize the “size” of S(s). Instead, if the objective is Te- jection of measurement noise at the output, at the tracking error signal e, or the robustness problem of equation (1.6). the “size” of T(s) should he ‘minimized, From the restriction in (1.7) this is not possible simultaneously. To solve this problem, first we should define an adequate “size” for a MIMO transfer matrix, particularly the sensitivity matrices, and by means of this measure we should try to solve the trade-off in (1.7). In later chapters we will define this measure and present different methodologies to solve the trade-off. Some of them are based on a more intuitive approach, as is the case of loop shaping (Chapter 4). The others—1a (Chapter 5) and Hx, optimal control (Chapter 6), y-synthesis (Chapter 7), and ¢! (Chapter 8) optimal contro!—solve the problem by means of an algorithmic procedure, The important point of this section is to realize that any well defined control problem should include implicitly in its requirements this trade-off, Otherwise the problem may be trivial a5 inthe following “textbook” prob Problem 11 Design a stabilizing controler for the system g(s) = (s+ As~2), which also rejets disturbing steps atthe output Clearly a controller k(s) = q/s with q > 2 solves the problem. This means that an infinite control action may be proposed being that q has no upper ‘bound. From the practical point of view, this is obviously unrealizable because of actuator bounds and high-frequency mode! uncertainty. The point is that ven in the formulation of a “theoretical” problem, the trade-off should be stated implicitly, (OBJECTIVES OF AN APPLEDTHEORY 17 1.4 OBJECTIVES OF AN APPLIED THEORY In the preceding sections we presented some ideas concerning model uncer: tainty in “real world” applications. the use of feedback, and the connections between the three hasie phases of a control problem. From them we can ex tract minimal requirements for a systems theory to be applicable in practical problems. In particular, we will concentrate on how Robust System Theory satisfies these requirements ++ Practical Assumptions. This requirement reflects the need to have sim: plifying assumptions that are related not only to the mathematical sim plicity but also to the applicability of the theory. In general, it means that the standing assumptions should be as unrestrictive as possible, in the sense of being more “realistic.” In the case of robust control it pertains to the connection between the three phases described in Section 1.1. As an example, assuming distur bbances as hounded energy signals is always more realistic than assuming they are steps. Also, from the practical point of view, instead of having 4 particular model as a description of the system, We can represent it by a family of models. This family is described by a nominal model and uses the information on the uncertainty bound, Strong Results. From the mathematical point of view, by strong we mean necessary and sufficient conditions. These conditions are therefore equiv- alent to a desired result, which can be verified (analysis) or implemented (synthesis), From a more practical point of view, we seek conditions that are neither unrealistic (too optimistic) nor conservative. In particular, the conditions we are interested in are on stability and performance of a certain control loop. As an example, the verification of ‘a necessary condition for stability is not enough to guarantee it; there fore itis optimistic. On the other hand, there could be no controller that satisfies a sufficient condition for achieving performance. That does not ‘mean there is no controller achieving that performance, therefore a con servative result. In both cases the results are inconclusive. Instead, we look for a strong mathematical condition (60 it can be tested), which is equivalent to a particular desired result (stability and/or performance) ‘Computability, The results obtained from the analysis and synthesis pro- cedures should have not only a mathematical representation but also ‘a computer representation. This means that the analysis conditions to be tested should be computable and the synthesis procedures imple mentable by means of an algorithmic code. ‘In particular, at this point we may discuss what “computable” means. exactly. We could argue that, although a code that implements an algo: rithm can be written, it may happen that either the code or the algorith are very inefficient, This could lead to a more general discussion between tractable and untractable computational problems ([226]) and efficiency18 INTRODUCTION Of algorithms and codes. Therefore the main point could be clarified as follows. The computational implementation of a control problem (analysis or synthesis) should not be the bottleneck of the solution to the problem. In the following example we will see how robust system theory meets the above requirements. : Example Consider the closed-loop system of Figure 1.8, The system S is represented by a family of models of the form S= (Go(s)[1+W(s)4] | Al] <1} (8) ‘This description considers specifically the uncertainty between the system and the set of models by means of A and the weight W(s), The output data from an experiment performed on the actual system have been processed by 2 robust identification algorithm (Chapter 10). From this procedure we com- pute the nominal model Gy(s) and W(s). Therefore we have quantified the relation between the actual system, the nominal model, and the uncertainty. The objective is to achieve closed-loop stability for all members of the family, defined as robust stability. If from the experimental data we obtain a family of models that “covers” the actual plant, robust stability will guarantec stability of the actual closed-loop system with the controller connected to the real plant. As will be explained in Chapters 2 and 4, at equivalent condition for robust stability is the following: [ROG Us KEG WE], <1 9) By means of the 71x. optimal procedure (Chapter 6), a controller that achieves the minimum of the above analysis condition can be calculated, Both the analysis condition and the synthesis algorithm can be computed efficiently by means of commercial software ({20, 126, 262) From the point of view of the three phases described in Section 1.1, both the analysis and synthesis theoretical results provide implementable condi- tions to guarantee closed-loop stability in the final experimental verification phase, r+ w(s)a\ COBECTIVES OF THSBOOK 19 1.5 OBJECTIVES OF THIS BOOK 1.5.1 General Remarks ‘The main objective of this book is to introduce the techniques used in the theoretical phase of the control problem from the point of view of Robust System Theory. Therefore, according to the discussion in Section 1.1, the theoretical results will take into account, as well, the problems that arise in the other two phases. ‘The title of the book includes Robust System Theory because it not only considers robust control but also has a chapter devoted to robust identi cation. The words Theory and Applications appear in the title because not only theory is presented but also practical “real world” problems. This is consistent with the objective of bridging the gap between both. The book is oriented to students of engineering departments with a back ‘ground in classical and modern control. ‘The main analysis and design methods are complemented by an elabo- rated example, either in the same chapter or in Chapter 11, The latter has a ‘group of worked out applications that stress particular practical issues: nonlin- earities, unknown time delays, infinite-dimensional plants, and actuator and sensor limitations. The three appendices—Mathematical Background, System ‘Computations, and Riccati Equations—make the book as self-contained as possible 1.5.2 Scope Chapter 2 gives the first robustness concepts applied to SISO control systems described by Laplace transforms. It relates these new ideas to the classical control stability, performance, and robustness concepts. It is a simple intro- duction to the main issues of robust control, for students or engineers with a classical control background. ‘Chapter 3 presents the internal stabilization problem from two different points of view. The first i a state-space approach based on solving four special problems first introduced in {104}. This also serves as an introduction to the tools used in Chapters 5 and 6. The second is a more classical approach, based on the use of coprime factorizations. As a by-product we obtain a parametrization of all stabilizing controllers and achievable closed-loop maps. ‘These results are instrumental in solving the €" control problem addressed in Chapter 8, ‘Chapter 4 introduces the four basic problems in robust control: nominal and robust stability and nominal and robust performance for MIMO systems, in a linear fractional transformation setup. The objective is to present a prac- tical synthesis methodology that is not mathematically involved and could be helpful to the student /engineer who developed an “intuition” designing with classical control tools (Bode plots). The fact that this methodology for mult20° INTRODUCTION variable systems is not a simple generalization of the procedure for SISO systems (Chapter 2) is also explained. This is the case of combined plant input uncertainty (uncertain actuator models) and high condition number plants, presented in its last section, Chapters 5 and 6 present the optimal 4) and 21x, control problems. These problems are solved first using a state-space approach using the tools intro dduced in Chapter 3. Alternatively, by using the Youla parametrization the problems are recast into an approximation form and solved using either pro- jections (in the 7p case) or Nehar's theorem (H.). Finally, in the Ma. ease ‘we present a third approach based on the bounded real lemma and the use of linear matrix inequalities. This approach allows us to easly address general (even singular) output feedback problems. Chapter 7 discusses more general types of uncertainties. In the case of structured dynamic, it defines the structured singular value (2) and presents its computation and use in both analysis and design. In the case of parametric uncertainty or mixed type uncertainties, it covers the analysis, the synthesis results being under development. This chapter also gives a good idea of the ‘open problems in this area, which are stil undergoing intense research Chapter 8 presents an overview of the relatively new subject of optimal € control theory. While the problem was formulated in 1986 [315], with the first solution methods for the simpler SISO case appearing around that time 180, 81. 83). this is presently a very active research area, where complete theory is just starting to emerge and where most of the results appear only in the specialized literature. The chapter starts by discussing the ¢! framework following the same induced-norm paradigm used in Chapter 2 to address stability and performance in the Has context. The € optimal problem is ‘motivated as the problem of rejecting bounded persistent disturbances (as ‘opposed to bounded energy signals in the A framework) or the problem of achieving stability against linear time varying model uncertainty. For sim- plicity most of the chapter is devoted to SISO systems, showing that in this simpler case the problem can be recast, through the use of duality as a finite- dimensional linear programming problem. Since this theory may result in very large order controllers, the chapter contains some very recent results on synthesizing suboptimal controllers using an LMI-based approach as an alter- native to the usually high-order optimal controllers. Finally, tis shown that in the continuous time case * optimal controllers are infnite-dimensional and the issue of fnite-dimensional approximations is addressed. Chapter 9 has a threefold objective. First, it gives a nice geometrical i terpretation ofthe structural properties of the state-space models in terms of principal components and directions, as related to controllability and observ ability. Second, this same interpretation is useful when evaluating numerical issues in state-space descriptions: For example, how “near” is the model from uncontrollability? This is useful for the student to distinguish between the “theoretical” model and the “numerical” model implemented in a com puter, This is also related to the “measure” given by the Hankel singular OBJECTIVES OF THSBOOK — 21 Table 1.1. Road map for self-study, by topic Robust Indentification ‘Stabilization Robust Analysis Robust Synthesis! 2 3 3 A 9 3 6 6 8 10 4 naz 7 n values. Finally, the third objective is to provide an introduction to balanced realizations as a part of model order reduction. A classical model reduction algorithm is presented. This material is closely related to the one presented in the next chapter. ‘Chapter 10 is a tutorial on a very recent (circa 1990) research subject that ‘connects the robust control methodology with the experimental modeling of systems. It presents the conceptual problem of worst-case deterministic identification but also details some of the main algorithmic procedures. Both time- and frequency-based experiments are considered, and a flexible beam example illustrates the use of some of these procedures along the chapter. It also provides a wide range of bibliographic references for the researcher. Chapter 11 gives a good idea of the issues that appear in practical prob- lems, It presents the implications of applying robust idemtiication and control procedures to several practical application problems, 1.5.3. How to Use This Book ‘This book is intended to be used either as a reference for control engineers or as a graduate level textbook. Table 1.1 gives an outline by topic of how to use the book for self-study. If used as a texbook, there are several al- tematives depending on the background of the students and the amount of twowerester courte Second Course Preteq: First Course 7 e 9 10 in * Optional ebapter22 mtRoouction time available. At the simplest level it could be used as a one-semester, be- ginning graduate level course serving as an introduction to robust control and requiring as a prerequisite a standard course on linear systems theory. This one-semester course would cover Chapters 2-5, portions of Chapters 6 and 7, and some of the examples in Chapter 11. Since Chapter 8 is largely independent, it could replace Chapter 5 if the students have already taken an optimal control course. As a follow-up to this course, a second advanced graduate level course would cover Chapters 7-11. These options are outlined. in Table 1.2, SISO SYSTEMS 2.1 INTRODUCTION In this chapter we address the issues of nominal and robust stability and per- formance probleme in single-input single-output (SISO) systems, represented either by a transfer function or by state-space equations. The analysis is per- formed proceeding from stability of the nominal model of the plant to the final objective of robust control: robust performance. ‘As mentioned in the introductory chapter, robustness measures are related to the type of uncertainty considered. The classical uncertainty description for SISO models, gain and phase margins, are presented first, along with an analysis of their limitations. To overcome these limitations, a more general robustness analysis based on more realistic descriptions is presented. Per- formance is based on rejection or tracking of sets of disturbances, noises, or reference signals, which lead to conditions on the Toop sensitivity func- tions. The above conditions for both performance and stability will affect the selection of the loop transfer function as mentioned in Chapter 1 For the sake of simplicity and clarity, in this chapter the definitions of uncertainty and performance are restricted to SISO systems. Generalization to MIMO systems is postponed until Chapter 4, where we will also intro- duce additional issues that do not arise in the simpler SISO case. Following the standard notation, we will use lowercase to denote SISO transfer func: tions and uppercase for MIMO systems. Throughout the chapter we will use packed notation to represent state-space realizations, that is, worncer-avrerve | Ate] 2324 sisosystEMs 2.2. WELL POSEDNESS Consider the feedback loop of Figure 2.1, where the input signals [uy(s),us(s)}. the outputs [y1(s),y2(5)}, and the errors fe1(3),ex(8)] satisfy the following equations: Jils) = als)er(s); yals) = K(ser(s)s es) =ui(s)—yas) A) ex(s) = uals) + y1(5) @2) Next, we define well posedness of a feedback loop, a standing assumption that should be made in any control problem. This avoids trivial situations in the loop, as in the case where g(s) = —-'(s) Definition 2.1 The feedback loop of Figure 2.1 is said to be well posed if and only if all possible transfer functions between the inputs ui(s) and wn(s) and a Outputs ex(s) e3)_yul6), and y2{5) exist and are proper. In other words, we can say that a feedback system is well posed when for well defined inputs, we obtain well defined outputs. In fact, it can be proved (Problem 1) that it is sufficient to verify well defined outputs [ey(s),€2(s)] or Lvi(s),y2(8))- It should be clear that well posedness of the feedback loop is a mathematical prerequisite for any control problem in which stability and performance are to be defined. It is also clear that in practical situations, feedback loops are always well posed. Before proving an equivalent condition for well posedness, we state in compact form the relationship between the inputs u; and errors e; for the feedback system of Figure 2.1 [28] = [eto “P] [88] ro [28 @3) wig als) 7 h Hs) ak Figure 2.1. Feedback loop model fo evatuate well posedness and (nominal internal ssobilty WELLPOSEONESS 25, ‘Therefore to have well defined outputs e\(s) and ¢2(s), we need the inverse of (5) to exist and to be proper. Otherwise, either the outputs will not be well defined oF could be nonproper. We will illustrate this by means of a simple example, Example 2.1 Consider the following model of a plant and a controller, con nected as in the feedback loop of Figure 2.1 as _ +7 2 = Gey KO Gay ea) The transfer function between tz and e3 is <1 _ (842s +5) Bacar! = ST es) whichis improper. Therefore the loop is not well posed. Note that, in practice, the model of the plant is stricily proper. In that case, iti easy 10 See Tha the above warsfér famction Would fave been proper Next, we state necessary and sufficient conditions for well posedness. As- sume that the plant and controller have the following state-space realizations: as) = [2 A 26) an Then we have the following result. Lemma 2.1 The following conditions are all equivalent: 1, The system in Figure 2.1 is well posed. 2. P(s) has a proper inverse. 3. 14+8(s)K(3) has a proper inverse. 4 Proof. All transfer functions between 1(s) and ¢)(s),f,)= 1,2 in (23) are well defined if and only ifthe inverse of P(s) is proper and exists for almost alls. The later and the fact that to guarantee well posediness it is enough that these four transfer functions are well defined (Problem 1) prove the equivalence between items 1 and 2. It is clear that if both the plant and the controller are proper, P(s) will have a proper inverse if and only if its determinant, (1+ g(3)k(8), has a proper inverse; which proves 2 and 3 are26 sIsosySTEMS equivalent. For properness, (1 + g(s)k(s)]! should exist for almost all s and, in particular, for s ~+ co. Equivalently, ‘im 1 + g(s)k(s) 40 28) ‘The state-space representation of this limit is 1+ DpDg #0, which proves 3 and 4 are equivalent, The above arguments stilT hold in the case where we consider yy and y2 as the outputs of the closed loop (Problem 1), a 2.3 NOMINAL INTERNAL STABILITY The classical concept of stability of a linear time invariant system, described either as a transfer function or by state-space equations, between a given input-output pair of signals, is well known. Therefore we will not repeat here these definitions already stated by classical and modern control theories, Nevertheless, we will point out the idea of internal stability of a feedback loop, which was not stated explicitly. To this end we will clearly distinguish the stability from an input-output point of view from the stability of a feedback loop. Consider again the feedback loop of Figure 2.1. Its not difficult to see that the stability of a certain input-output pair (all the poles of the corresponding, transfer function in the open left-half complex plane C_) does not guarantee that all input-output pairs will be stable (in the same sense), This is illustrated by the following simple example. Example 22 Consider the following system and controler: (+1) (s-1) ene *= wt) oD connected as in Figure 2.1. The transfer function from the input u; to the ouput 1 is given by Ty, =1/(s+4), which is stable in the usual sense. However, the transfer function between the input uy and the output yi Ty, = (5+ 1)/(6~1)(s +4), which is obviously unstable. As we will see next, this is caused by the cancellation of the unstable plant pole at s = 1 by a zero of the controller. ‘This example shows that there is a difference between the stability of a certain system, considered as a mapping between its input and output,! which we define as input-output stability, and stability of a feedback loop, Which will be defined next. In the latter, we must guarantee that all possible input-output pairs be stable, which leads to the concept of internal stability "Even in the MIMO case with several inputs and outputs. NOMINALINTERNALSTABIITY 27 Definition 2.2. The feedback loop of Figure 2.1 is internally stable if and only if all wansfer funciions obtained from all input-oulpua pairs have ther poles nC. (input-ouput stable). Itis easy to prove (Problem 2) that to verify internal stability i is sufficient to check the input-output stability of the four transfer functions between the inputs [us(s),u2(5)] and the outputs fe:(s),e2(s)]- This is summarized in the following lemma. Lemma 22 Let ow = [245+] 9) bbe a minimal state-space realization of the closed-loop system of Figure 2.1, between the inputs |us(s),ux(s)] and outputs {ey(s),ex(s)] (or [pi(s).yal)]) Then the following conditions are equivalent: The feedback loop in Figure 2.1 is internally stable The eigenvalues of Ac are in the open left-half complex plane C The four transfer functions obtained between inputs |us(s),43(8)] and outputs {ey(s),€2(s)] (or lyi(s),y2(8))) have their poles in C_; that is, Ci(s) is input” output stable, Proof, 1 <> 3isleft as an exercise (Problem 2).2 = 3 follows immediately from Definition 2.2. Finally, ifthe four transfer functions are stable, then all the eigenvalues of A. are in C_. since minimality of (2.9) precludes pole-zero cancellations. This shows that 3 = 2. a Is important to note that the concept of internal stability, although not explicitly stated, has existed among practicing engineers for many years. A. proof of this is the basic controller design rule of no unstable pole/non- Minimum phase zero cancellaion Between plant and coniroller should be allowed. A misleading explanation for this rule is that, numerically, its im- Ossble to cancel exaetly a zero and a pole (why is it misleading?). According to the definition of internal stability, we now know that the argument behind this “practical” rule is the following: with unstable pole-zero cancellations, there will always be an unstable input-output pair in the loop (and hence ‘an internal signal could grow unbounded, damaging the system). This will be proved next. Lemma 23 The feedback loop in Figure 2.1 is internally stable if and only if (1+ e(s)k(3)|" is stable and there are no right-half plane (RHP) pole-zero cancellations between the plant and the controller.28 sisosystems Proof, First factorize both plant and controller into coprime numerator and denominator polynomials (which do not share common roots), as follows: =) mute) als) a6)" (s) = als) (2.10) We assume that estoy = tesla) en) ing(odne(s) + de(s)dg(6) is stable. If (s) and de(s) share common roots in the RHP, these will also be roots of msimu(s) + dy(s)d,la)-The same occurs if ny(s) and ds) share contin roots in the RHP. Therefore in the first case, the transfer function y(s)ru(s) ie(s)ru(s) + da(s)dg(s) is unstable, because the RHP roots of the denominator will not be can- celed-by-the numerator [this is precluded by the coprimeness of the pairs (ro(s) (3) and ((5) dys). In the second case the transfer function [1+ g(s)k(s)}"1 (6) will be unstable for the same reason. Note that the com plementary sensitivity (1 + (s)k(s)}-"g(s)K(s) remains always stable, with or without RHP cancellations ifthe sensitivity is stable. On the other hand, if there are no RHP pole-zero cancellations, all these transfer functions are stable. As these are all possible transfer functions, the loop is internally stable, which concludes the proof. o H(s) [1+ g(R(S)™ (212) Note also that, according to the definition, any stable pole-zero cancel lation in the loop will not modify the internal stability. However, from the point of view of performance, the cancellation could be performed over a stable although undesirable mode of the system. This means that a certain input-output pair of Figure 2.1 will have “poor” performance. This can be understood better by means of an example. Example 2.3 Consider the following system and controller: (1 = s)(s? +0.2s +1) (s+ Ips +5) 89) = Cerner aes) MO = 3670241) connected as in Figure 2.1. The design objectives are to internally stabilize the Toop and to track with yy @ step input injected at uz with not more than 25% ‘overshoot. The controler selected meets the performance specification, since the closed loop is internally stable and y,(s}/us(s) = (I~ $)/(s +1), which thas an adequate time response. However the transfer function y(s) uals) ROBUSTSTABILITY 29 (5+ 3)(s-45)/(s + 1s? +025 +1) presents undesirable oscillations due to the lightly damped poles at s = —O.1¥), This situation can be avoided by simply extending the definition of inter- nal stability to encompass more general regions, rather than just the (open) leftchalf complex plane. Therefore, according to the desired location of the closed-loop poles, an adequate region in C_ can be defined, and no pole-zero cancellation between plant and controller should be performed outside this region. 2.4 ROBUST STABILITY As mentioned in Chapter 1, a given robustness margin is related to a specific type of model uncertainty. Therefore by first evaluating the stability robust- ness margins used in classical control theory we can analize their limitations and propose more realistic and general uncertainty descriptions. 2.4, Phase and Gain Margins The phase dy and gain margins gq are used as robustness measures to evalu- ‘Toosely speaking, {he abilty OF syslem To Tespond adequately, n terms of performance and stability, even when the open-loop model é(s) = (s)k(s) used in the design does not exactly match the physical system, due to the existence of uncer- tainty Both measures can be clearly interpreted in terms of the Nyquist plot and the Nyquist stability criteria, In this plot, dy and gm Fepresent the “distance” in angle and absolute value, respectively, to the critical point z= —1, as shown in Figure 22. ‘The usual practice of evaluating both margins separately assumes implic- ily that both types of uncertainty (phase and gain) act on the loop one at a time. As a consequence, these margins are effective as analysis tools only wen the model ofthe plat as ber phase or ginancereiny Care must be taken When using ¢, and gq for stability robustness analysis of SISO closed-loop systems, because they do not guarantee robust stability for the more realistic situation where both phase and gain are alfected by uncertainty. Here robust stability means the stabil sible models that are described by the combination of the nominal model 4go(s) and the uncertainty 5= ce, where cand @ are uncertain values Bontained inside the intervals Te = [émnsCmar] and Ly = [4mie: Prue» FESPEC tively (Figure 2.3). This set of plant models can be defined 38 the family of models F © {g(s):8(s) = go(s)[ee"*), C1. & € Ly}. We will clarify these concepts by means ofthe following examples.30 siSOSYSTEMS. Figure 2.2. Phase and goin margins. Example 24 Consider the Nyquist plot of Figure 2.4, in which both $m and Bm have adequate values. Nevertheless, with small simultaneous perturbations in the phase and gain of the oop, the plot will encircle the critical point z = 1. Therefore, in this case, the minimum distance from the Nyquist plot to the critical point (the minimum return difference) gives a much better assessment of the robust stability properties of the system than dy OF Bm Example 25 Consider the Nyquist plot of Figure 2.5. In this case, the gain ‘margin gm —» 02, although the phase margin is very small. If we know that the only source of uncertainty in the loop is in the gain, a small b, poses no threat to stability. Therefore, in this case, the controller designed guarantees robust stability Figure 2.3. Phase and goin margins retated to the uncertainty in the model ROBUST STABILITY 31 Im(:) Figure 2.4, Simutianeous uncertainty n both gain and phase, Im(s) Re(s) Figure 2.5, Uncorointy inthe opentoop goin,32. SISO SvSTEMS From both examples we conclude the following: ‘+ A given stability margin is a good indicator of the robust stability prop. Of gain and phase margins in the analysis of control systems is limited to the cases where there is @ priori knowledge that the ut s either on (that isnot ously. ‘© The uncertain system can be analyzed as a family of models that can be described mathematically by a nominal model and a bounded uncer- tainty set. In the Nyquist plot this corresponds to a nominal plot and an uncertainty region surrounding it «If there is simultaneous uncertainty in phase and gain, itis important to analyze the type of uncertainty considered, because it will determine the form of the set of Nyquist plots in the complex plane. «© The stability margin can be computed as the minimum distance between the family of models and the critical point in the Nyquist plot. Therefore it is directly related with the region describing the model set in C and as a consequence, with the type of uncertainty. This can be expressed as tollows: ‘Type of uncertainty | ~ Phase and gain margins are special measures that are useful only when either the uncertainty is “angular” or “modular,” respectively. They can be considered as special cases of a more general (and realistic) uncer- tainty description that will be presented next. 24.2. Global Dynamic Uncertainty Mote tals descriptions of mode! uncertainty have been developed, lead ing to controller designs tat perform better in practice, One seh Sesrip Mon the global dynamic unctiang. The ame arses snce thin uncertainty is related «9 the uncertainty athe system dynamics and covers globally the compete model ofthe plant. This ype of uncertainy description can be wed ‘when the order of the differential equations describing the plant is unknown ? This situation arises in practice when certain physical phenomena (last fenibliy) are not included inthe plant deseription or when the physical description of the ssem isto complicated to Tead to Wecable problems 2A ciferent branch of control theory sso dealing with uncertain systems adaptive contol IW has as one of ts base assumptions thatthe order ofthe equations describing the system is nova. Today, thee i an important effort geared toward relaxing this assumption ROBUSTSTABLTY 33 In many cases it describes situations where a constraint in the computation time (real-time control ofa fast process) imposes a limit on the order of the ‘mathematical models that can be handled. As a consequence, these models are approximated by a set of equations of lower order, and the neglected dynamics are lumped into the uncertainty Global dynamic uncertainty can also be used to describe lin errors, where a nonlinear system is linearized around a nominal point, Although the resulting model is valid only in a neighborhood of the ‘nominal operating point, the nonlinear effects outside this neighborhood can be bounded and interpreted as dynamic uncertainty. Typical examples of this situation appear in the modeling of aircrafts, which are described by a general nonlinear mathematical model. The latter is linearized at different altitudes, velocities, and angles of attack, yielding a set of linear models, each ‘one valid at a particular operating point. An uncertainty bound “covering” the linearization errors in a neighborhood of each linearization point can be interpreted as dynamic uncertainty Finally, many systems may be described by partial differential equations or time delayed equations, both called infinite dimensional systems. In these cases, an approximation ([24,25, 79, 134, 169)) of the general equation should bbe made so that a finite-dimensional model is obtained. The approximation error can be interpreted as dynamic uncertainty because the approximation process eliminates all the high-order dynamics of the general mathematical Iodel. In particular, uncertainty inthe time delay can be modeled as dynamic uncertainty ((268)) in a very simple way. Examples of the latter and of the ‘dynamic uncertainty of a flexible structure due to its approximation error are presented in Chapters 10 and 11 ‘The approach adopted by robust control theory is to describe a physical “system by means of an uncertain model. The latter is defined as a set of models and described in terms of a nominal plant together with bounded "uncertainty. In classical and modem control theories the controller is designed using the nominal model of the plant. Instead, robust control theory attempts to design a single controller guaranteeing that certain properties such as stability and a given level of performance are achieved for all members of the set of models. In this case, these properties are said to be robust. In the {following subsections, different sets or families of models will be described. Multiplicative Dynamic Uncertainty An uncertain model with multi- plicative dynamic uncertainty is shown in Figure 2.6. It describes a physical system as a set G of mathematical models, as follows: G = felsh:als) = gu(s){1+3W6)], BEC, I< 1} 13) The set G is the family of models and is characterized by a nominal plant (8), fixed weighting function W(s), and a class of bounded uncertainty 5 Te nominal model g,(s) corresponds to the case where there is no uncer tainty, that i, 8 — 0. Without loss of generality, the hound on the uncertainty34° ssosysteMs [Family of Moses ii + 5W(s)}-+ 96(s) | SEY Figure 2.6, Dynamic mutipticative uncertain 6.can be taken to be one, because any other bound can be absorbed into the weight W(s), “The weighting function W/(s) represents the “dynamics” of the uncertainty, or in other words its “frequency distribution.” A graphical interpretation of this uncertainty weight can be scen in Figure 2.7. In this plot we can observe a \ypical weighting function, with 20% uncertainty at low frequencies and 100% at higher ones. A complete lack of knowledge of the system (100% error) ‘means in practice that there is no phase information (ie., the plant transfer function g(s) can have either a positive or negative real part). Thus, in order to guarantee stability of the closed-loop system, the controller must render the nominal loop function go(ja)k(jw) small enough at frequencies above , to guarantee that the Nyquist plot does not encifcle the critical point, regardless of the phase of g(jm)k (jon) This point illustrates an important limitation for robust stabilization that should be taken into account in any controller design, The situation where the uncertainty is small at low frequencies and in- creases at higher frequencies (although not nearly’as dranvatically-as in the example discussed above) is the usual case arising in practice. For instance, for mechanical systems a better model can be obtained to describe slow Figure 2.7. Frequency distibution of uncerionty, ROBUSTSTABLITY 38, mechanical resonances than faster ones (involving mechanical Nexibility) This situation can also be observed in models of electrical cicuits. For appli- cations in lower frequency bands, the model is simpler and usually provides a ‘good description of the physical realty. To deseribe the response of the circuit at higher frequencies, other physical phenomena should be included in the ‘model (parasitic capacitances and inductances) increasing its complexity. Al- tematively, these additional dynamics can be treated as dynamic uncertainty. The weighting function W(s) can be obtained explicitly in many applications {see Chapter 11) or by using the techniques known as robust identification, which will be described in Chapter 10. An important each mem ber of unste SraTHeat based on the Noquist graphical criteria (whichis a necessary and sufficient condition for stability of a single model) as follows. Suppose that any two members of a family have a different number of RHP poles, then their Nyquist plots will encircle the critical point a different number of times. [As a consequence, the region between both plots will include the critical point z = —1. For the family of models considered, the set of Nyquist plots of this family “covers” this region and also includes z = 1. On the other hand, a model that goes from (closed-loop) stability to instability should have its Nyquist plot go across the critical point, because it changes the number of cncirclements. Thus it will be difficult to detect if an element of the family is (closed-loop) unstable since the whole set of models already includes the point z = —1. Hence we cannot derive equivalent conditions for (closed-loop) stability of all members, based only on their frequeney response (Nyquist plot). For this reason, to guarantee that the number of RHP poles of all members of @ remains constant, we will assume, without loss of generality, that W(s) is stable (why?) and that no unstable pole-zero cancellations take place when forming gq(s)|1 + 8W(s)). Perturbations & satisfying this later condition will be termed allowable. ‘A more general uncertainty description can be obtained by replacing 8 € C by a teal rational transfer function 4(s), analytic in the RHP and such that 1A(s)Ilw = sup,, |Aw)] <1. As before, we require that no unstable pole— zero cancellations occur when forming g.(s)[1 + A(s)W(s)]. As will be seen next, robust stability of this general model set just described and the one in (2.13) yield the same condition. Thus, for simplicity, in most of this chap- ter we adopt the description in (2.13), which ass “frequency” distribution of uncertainty to W(s), and the bound to 8 € C. Finally, note that, in order for the problem to make sense, we should always assume bounded uncertainty, otherwise there is no possible design that can stabilize the whole set simultaneously (why?), A condition guarantecing stability of all elements of the family G, that is, robust stability of gu(8), is derived next.36 sISOSvSTEMS ‘Theorem 2.1 Assume the nominal model go(s) is (internally) stabilized by a controller kis) (Figure 28). Then all members of the family G will Be (inter: nally) stabilized by the same controller if and only if the following condition is satisfied: | ITEWEIe *supl7Gw) Ge) <1 | easy Le with T(s) © go(s)k(s)[1 + go(s)k(s)]~" the complementary sensitivity function of the closed-loop system. Proof. From Lemma 23, it follows that: 1. {1 +g(s)k(s)}" stable for all members g(s) € ¢. 2. No RHP pole-zero cancellation between k(s) and any member g(s) € 6. are equivalent to robust (internal) stability. First we prove that (2.14) is equivalent to condition 1 above, when the nominal loop is internally table [1 +g(s)k(s)]* stable Vals) eG (2.15) > MH gG)KON+WE)] ZO MBl 1+ gols)k(s) # go(s)k(s)W(s)8 —¥/8J<1,seC, (2.17) gt g BoD K(S)WE) # T+ go(SK(G) o (enon <1 vec. 219) = IT(s)W(s)I] S10 (2.20) Pej bead | Ka) = Vidl 418 €,|8|<1 can vay be found sch that equality holds The “only i partis vial For equation (220) we have sed the maximum modulus eo. Fem which states that since PEpW (a) analyca Ce BORE oT eR of the nominal closed Toop-and W (3), IT @)W (3). achie the bounty of Cy thatthe yo an Neve we complete hep tiene (=) It is clear that robust (internal) stability implies nominal (internal) stability and equation (2.14) (equivalent to condition 1), (on) On te other hand t's ear that nominal (tera) ability and equation (220) imply condition 1. To complete the proof we need ts how tht eunition 2 abo bolas Flom Lemna 23 we hove that nena ser nal sbity implies no RHP pole-2ere cancelation between) and Ke) Hence for any RHP pol of the open logp pv € Gs Ti.) = Listy?) Use (218), whichis equivalent to Sondtioo TRS have Mos) < ana hence 13 Gp) £0.09] 1 Becane paring, this imbisno RAP poe zero cancelation between Ka) tnd any merther glahe 6 (eomion 9) : ols)k(s)W(s) T+ go(5)K(5) itsmaximunron f of the Using Figure 2.9, we can interpret condition (2.20) graphically, in terms of the family of Nyquist plots corresponding to the set of loops. First observe that (2.20) is equivalent to [1+ go(@)K(s)| 2 lso(S)K(G)WOS)], Y= jo 221) For a given frequency w, the locus of all points z(sw) = go(Jo)k(Jo) + BoJo)kyw)Ww)6, 6 €C, [8] <1 isa disk Dw), centered at g,(Jo)k(Jo) with radius r= |go(jo)k(Jo)W a)]. Since [1 + go(7w)k(w)| is the distance between the critical point and the point of the nominal Nyquist plot corre- sponding to the frequency jw, it follows that condition (2.20) is equivalent to requiring that, for each frequency «, the uncertainty disk D(a) excludes the critical point z = ~1. Therefore robust stability for SISO systems can be checked graphically by drawing the envelope of all Nyquist plots formed by the set of circles centered at the nominal plot, with radi [@,(/w)k(j«)W (a) and checking whether or not this envelope encloses the critical point 2 = —1 In the MIMO case, although an equivalent condition can also be obtained (Gee Chapter 4), therg is no such graphical interpretation The frequency at which |W(jw.)] = 1 corresponds to the limit for which there is 100% uncertainty in the model of the plant. Therefore another in terpretation of condition (2.14) is that frequency w, is the upper limit for38 siso systems Figure 2.9. Sot of Nyquist plots of the fomily of models. the bandwidth of the complementary sensitivity function T's). At frequen- cies above w., the gain |7(ju)| should be less than one, to achieve robust stability Gain and phase uncertainty in the loop can be considered as special cases of the dynamic multiplicative description, W(s) = 1 and 8 € R, [8| <1 4 t0_gain uncertainty. On the other hand W(s) = }€ C6 € 0,2) = |3| <1 corresponds to phase uncertainty. Th both cases there is an exira structure on the uncertainty 6, which should be a consequence of the extra « priori information on model uncertainty that is, ‘only phase or gain is uncertain. For this reason, the condition of Theorem 2.1 could be conservative if applied to this case and if the a prior information is correct. Otherwise itis more prudent to consider the unstructured type of ‘uncertainty described by the set &. Several examples will serve o illustrate the above concepts Example 26 Consider the following nominal model go(s) and a second pos sible plant g,(s): 300 Gree H+ TO} oe HE g165) = (2.22) Bul3) = (2.23) ROBUSTSTABLTY 39 These two plants can be described using the following family, characterized by ‘multiplicative uncertainty represented by a weight W(s) and a bound on 5: fs(sh-e6s)~gols)[1+ 3WG)}, [81<1} 24) wo) = a (225) Its easily verified that gy corresponds 10 8 = 1. Note that the set G also includes many other plant models, For instance, for 6 = we obain as +50 ale) = ett) 2.26) GF IG ++ 100) Therefore if we only need to include the models g,(s) and go(s) or even a finite set, the description could be unnecessarily conservative. As a consequence, any design that applies 10 all members of G could be conservative as well (what will be sacrificed?). In the next example we illustrate the use of multiplicative uncertainty to describe uncertainty in the location of the zeros of a transfer function. Example 2.7 Consider a set of models with uncertainty in the high-frequency dynamics (w > 120 rad/s) of the numerator polynomial. o= (ewe ca} any fol) ag for 5=0 (228) CHD ‘This uncertainty can be represented as multiplicative dynamic uncertainty using the weighting function W(s) =0.2{1 + (s/20)]. Its graphical representation is similar to the one of Figure 2.7. The system is known with a 20% relative error up (0 10 rad/s. Above 100 rad/s the model has no information on the system ‘that may be useful for control design (|W (ja)| > 1). According to condition (2.20), this frequency is the upper limit for the bandwidth of the complementary sensitivity function T(s), to achieve robust stability. Other Uncertainty Descriptions Additional types of dynamic uncer- tainty descriptions, similar to (2.13), can be formulated as follows: Gu = (Gols) + 8Wa(s), 15] <1} (2.29) Gy 3 {eols) {1 + BWa(s)11, [3] <1} (230) G, = {eols) [+ 8MiCs)Ro(8)1"", [91< 1} 31)40 siso systems ‘These descriptions are defined as additive, quotient, and inverse dynamic uncertainties, respectively, The difference between the description of uncer- tainty By ary of the above depends on the specific application under consid- eration, Take, for example, the case where a high-order model (even infinite dimensional) g(s) must be approximated by a lower-order one g,(s). The approximation error can be considered as additive dynamic uncertainty. The set of models “centered” at the nominal g,(s), which includes the high-order ‘one g(s), can be defined as follows: Go S {arls) + 8Wals), [5] <1} (232) ‘The weight W,(s) can be obtained from the frequency responses of the approximation error (or its upper bound) and g(s). A systematic way to ‘obtain this uncertainty description and its corresponding uncertainty weight will be presented in Chapter 10. In the same way that multiplicative uncertainty seems to be the “natural” structure to describe uncertainty in the zero locations, quotient dynamic un- certainty is the least conservative way to describe open-loop pole uncertainty. ‘This will be illustrated in the following example. Example 2.8 Consider the following set of mathematical models, which can be described as a family with quotient dynamic uncertainty 1 % {moar wali} 233) & feo(s)[L+ Wolo)", lol <1} (234) which leads to the following nominal model and uncertaimy weight 1 3 me MWe (235) Bo(5) If, instead, the same set would have been described by a family of multi plicative dynamic uncertain models, there would have been an unnecessarily large number of models, and as a consequence the description would have been conservative. If we take the same nominal model, the new uncertainty weight Wm(s) and bound ym should be defined such that {BmWnls), Bm EC, [Bul < Yh (236) “covers” the following set: 38 - {etry 86 ici} (237) ROBUSTSTABILITY 41 ‘obtained by considering (g(s) ~ go(5))/Bo(s). (5) € Gy- The weight Wo(s) = 3/{8 +3) and bound y= 1 achieve the above condition but include in the new set of models Gy second-order systems, which were not included in Gy We conclude that, depending on the specific application, the description of model uncertainty should be selected appropriately among the above de- scriptions. If ail the uncertainty is included in one particular description, itis called global dynamic uncertainty. Otherwise, in many cases, we may adopt ‘wo or more uncertainty descriptions to better “fit” the family of models, By this we mean that, taking advantage of the information on the “structure” of, the uncertainty in the model, we select the smallest possible set. This is im- portant because, in general, a robust controller should stabilize and provide performance for all possible plants in the set of models; therefore the larger this set, the more conservative the design. The type of uncertainty description that combines information of different portions of the plant and adopts different uncertainties for each part, that is, different 3's, is called a structured dynamic uncertainty description. For ‘example, consider the following set of models with both uncertain zero and uncertain pole locations: §+2(1+0.36,) o = {STS lal <, Fe (zn) (238) 2 fo cll +6) 7 ; 8 {eon lal<1, ra} (239) In Chapter 7, robustness analysis and design tools for this more general ype of uncertainty description willbe presented. Following the proof of Theorem?.1, necessary and sufficient conditions for the robust stabity of the model sets described in Tale 21 can be obtained (Problem 3), where S(s) © [1+ gu(9)(6)-" is the sensitivity function of the nominal closed-loop system. We can conclude from ths table that different stability conditions correspond to different uncertainty descnptions. This coherent with the fat stated before, that sabity margin ae related to 8 Table 2.1 Global dynamic uncerlainty descriptions and the conesponding robust sabilty condition (6 <1) @ Robust Stability £a(9)[1¢ Wal) WAT, <1 ss) + 3446) HWEEe)kLS) SIL < 1 sols) [1 + 8W,(5))! WW, (s)S(syIL. <1 [1+ BWSR) IMC SE), = 142 ssosvstems specific type of uncertainty description, This was the case with the classical phase and gain margins. The robust stability conditions in Table 2.1 can be interpreted also as stability margins. Take the case of multiplicative uncertainty and suppose we bound it as Bo] < yn, With y%, positive real number. Then its easy to prove that the equivalent robust stability condition for this new set of models is the following: IPOW Oh <5 240) The above means that [|7(5)W (s)|q, can be interpreted as a stability margin for the set of models with multipicative dynamic uncertainty in the same way gn i the tabiliy margin fora set of models with gun uncertainty. TBs isso Because the infinity norm ofthis particular function gives the measure of how much dynamic multiplicative uncertainty can be tolerated before there exists at least one model inthe set that is closed oop unstable Te follows thatthe controller that yields the largest stabity margin (with respect to multiplicative uncertainty) can be found by solving the following ‘Optimization problem: waa, yp ITO¥ Oe Gan ‘The above is an Mx. optimal control, which will be analyzed in detail in Chapter 6, 2.5 NOMINAL PERFORMANCE Performance will be defined in the frequency domain in terms of the sen: sitivity functions, according to the compromises stated in Chapter 1. First, the classical and modern control concept of performance as rejection (track ing) of known disturbances and/or noise (references) is stated. Based on the practical limitations of this approach, we state a more realistic definition of performance as rejection (tracking) of sets of disturbances and/or noise (references) 2.5.1, Known Disturbance/Noise/Reference In classical control theory, a measure of performance of a closed-loop system is based ofits ability to reject (in steady state) known disturbances, noise, ‘oF measureitient errors, which may appear at differeni parts of thé loop, that is, sensors, actuators, or outputs. In this context, we define a known signal as one having a particular form (sine, step, impulse, etc.), which is known beforehand by the designer, athough it snot known at what particular NOMINAL PERFORMANCE 43, time it will disturb the system. Similarly, a measure of performance may be posed as the ability of the loop to follow known reference signals with zet0 steady-state error. Itis a well knowin fact that rejection of exogenous disturbances and tracking references are equivalent problems. If we consider input signals that are polynomial in time, then we can clusify stable systems according to the highest degree of the input that can be tracked with finite steady-state error. An internally stable system is said to be of type n if it can track (or reject) a polynomial input of degree up t0 7: with zero steady state error I€18 easy to show that a type m FDLTI system must satisty the condition lim, .s"a(s)K(s) = K 4 0, For instance, in order to track a step input (or reject a step disturbance), the system must be at least of type I (ie, lim, 0 59(s)k(s) © K, #0). Similarly, to track a ramp reference (or reject a ramp disturbance), the system must be at least of type 2 (lim,_.9 s°g(s)k(s) K, #0), In modern control theory, the same concept of performance has been used: rejection of known disturbances in the state of output of the system or tracking a known reference signal. In addition, the concept of optimal control is introduced. In this framework the contrller seeks to minimize a certain functional, which quantifies the compromise between tracking (rejec- tion, stabilization) speed and the control signal energy. I is possible to state this optimization problem in both a deterministic and a stochastic context Inthe latter, the perturbing signat 1s considered as a stochastie process of Anown covariance. In either case, the optimal controller that minimizes the functional will remain optimal as long as the disturbance or the reference signal matches exactly the assumptions made in the design process. If this hypothesis is not met, there is no guarantee of optimal, or even “good.” _ disturbance rejection or tracking (100). We may conclude that, in both cases, the fact that the designer should have a clear knowledge of the disturbance, noise, or reference to be tracked poses a serious constraint. If he/she designs for the rejection of a particular sig- ral, but it tums out that the disturbing signal is different, performance could seriously be degraded. In the same case, if the design has been performed by an optimization method, the performance no longer will be optimal. Con- ceptually, this is the same limitation we found for closed-loop stability. If we stabilize a nominal model, there is no guarantee that the closed loop will re- main stable if the mode! is changed. For these reasons, classical and modern control techniques require a clear knowledge of the nominal plant model and the external signals for correct (or optimal) stabilization and performance, In practice, the designer does not have such detailed information. For this reason, the assumptions on the system and external signals should be relaxed. For the case of uncertainty in the system description, the approach pursued in the last section was to consider a family of models, rather than a single nom- inal model. Similarly, in the case of performance, a family of disturbances, noise, or references will be considered. An example of a practical situation44 sso systems Be He) Fale) Figure 2.10. Feedback loop with output disturbance, where this relaxed assumption is necessary is the case of robotic manipula- tors, when different trajectories need to be tracked in situations where the ‘workspace is changing with time. Also, we could have sinusoidal disturbances ‘with frequencies contained in a certain uncertainty band, or even less infor- ‘mation, as the case of energy-bounded signals. In the next section we present the performance analysis when defined with respect to a family of external signals, acting on the nominal model of the system. 2.5.2 Bounded Disturbances at the Output Consider the block diagram of Figure 2.10, with output y(s) = S(s)d(s) = [1 +.g(s)k(s)}"" d(s). If the performance objective is to minimize the effect Of the disturbance d(s) at the output y(s), a stabilizing controller that makes Sis) as “small” as possible needs to be designed. The trivial solution to this problem is k(s) + 00, s0 that S(s) — 0, but there is no guarantee of nominal stability. Furthermore, as mentioned in the introductory chapter, @ fondamental constraint of any feedback loop is the equation S(s) + T(s) = 1 ‘Therefore in this case we would have T(s) —+ 1 at all frequencies. According to condition (2.14) ofthe last section, this implies that robust stability.can be ‘guaranteed only if the uncertainty in the plant model is less than 100% error at all frequen : = a ‘AS a consequence, we should seek a controller that makes $(s) “small” in a certain frequency range of interest for the particular application on hapd (Gee Figure 2.11). This could be the case. when output disturbance rejection needs to be achieved at low frequencies, as in the position control of Targe mechanical systems with very low natural resonance frequencies. Further- 15()| Figure 2.11. Desired frequency distibution fr 5(s), NOMINALPERFORMANCE 45 more, there could be situations where there is a certain a priori knowledge of the frequency contents of the disturbances. The frequency range where ‘S(s) needs to be “small” or the frequency content of the disturbances can be represented by the weights W,(s) and Wg(s), respectively. These weights are dynamic systems, which represent the knowledge of the frequency bands of interest for performance and the disturbance frequency content. Both can be incorporated into the block diagram of the system as in Figure 2.12. ‘Our next step is to define precisely the meaning of “small” S(s). The size of a transfer function, in particular the sensitivity function, will be measured by its induced norm. For a general LTI operator A: x A.+x the induced norm is defined as follows: Alles = sup A sala (2.42) In particular, the operation « will be equivalent to convolution, if both input ‘and output are time signals, or product, in the case where they are repre- sented by their Laplace transforms. Therefore the induced norm will depend. entirely on the norms we adopt to measure both the input and the output. For convenience, in most of this text, we will adopt the energy of the signals as a ‘measure of their size. In this case, according to Parseval’s theorem, for both representations of a signal, time or frequency, their norm will be the same, that is, x(0)l2 = l/x(s)[i. With this in mind, we define nominal performance as follows: Definition 23 Nominal performance of the feedback loop of Figure 2.12 is achieved if and only if the weighted output remains bounded by unity, that is, HW, (s)y(s)lla <4, for all disturbances in the set {4 € £2, |\d2 <1}, and for all other external inputs to the system equal t0 zero. Without loss of generality, we have considered unitary bounds in both ceases, because any other bound can be absorbed into the weights W,(s) and Wa(s), by linearity, Wale) Wale) = Figure 2.12. Augmented feedbock loop with performance weights46 ssosvstems Checking nominal performance by using Definition 23 directly requires a search over all bounded £; disturbances, which is clearly not possible. Hence, as in the case of robust stability, it is necessary to find a computationally verifiably equivalent condition. This condition will be obtained by exploiting the relationship between the £2 to £) induced norm of a LTI system g and its frequency response, Theorem 2.2 The feedback system of Figure 2.12 achieves nominal perform ance, as defined in Definition 2.3, if and only if (1W4(s)S(s)Wa(s)|],, © sup |W,Ga)Syo)Wa(yo)] <1 (2.43) Proof. The operator that maps the input signal d(s) to the weighted output W, (s)y(3) is W,(3)S(s)W,(s). The induced norm from £; to £3 isthe infinity norm (see Appendix A). Therefore are WWyylle = Wyler cS (2.44) => [Ws (5)S(s)Walslao <1 (245) Hence nominal performance is equivalent to equation (2.43), 0 Note that this condition is similar to the one for robust stability (2.14), although applied to a different transfer function. Nevertheless, there is an important difference between both conditions. The one for robust stability can be interpreted qualitatively. In other words, either all members of Tare closed-Toop stable, or there is at least one model that is not. There is 1o~ “gradual” boundary separafion-between the set that satisfies (2.14) and its complement. Instead, nominal performance can gradually be relaxed, since it is a quantitative property of the closed loop. Specifically, this means that” the upper bound of the weighted output to be bounded (or minimized) can bbe changed to a number y > 1. This new bound will “gradually” degrade the performance, This separation between qualitative and quantitative properties of systems is well known and has nothing to do with robustness but with the definitions of stability and performance. It has been pointed out, because it will be important in the next section, when considering robust performance. ‘As before, the optimal controller, in the sense of providing optimal dis- tubance rejection, can be found by solving an H.. optimal control problem, in this case of the form stl pain [MOSH]. (2.46) The solution to this type of problem will be discussed in detail in Chapter 6. NOMINALPERFORMANCE 47 Note that the only assumption on the disturbances is that they have bounded energy, There is no hypothesis whatsoever on the shape of these signals, as in the usual setting of classical and modern control. The weight- ing functions W(s) and W,(s) only provide an a priori knowledge of the frequency content of these signals but do not determine their shape. ‘A graphical interpretation of the nominal performance condition can be ‘obtained by means of a Nyquist plot (see Figure 2.13). To this end, define W(s) 2 Wy(s)- Wy(s) and note that (2.43) is equivalent to [W(yo)| <|1+gol@)k(Yo)|, Yo (247) Consider, for each frequency J, @ disk Diya) centered at 2 = ~1, with radius 1 =|W(w)]. Then (2.47) can be interpreted graphically as nominal perform: ance being achieved if and only if, for every frequency jw, the disk D(jw) |W), vo @si) where W(s) © Wy(s) Wa(s). Hence, in this case, nominal performance is achieved if and only if, at each frequency, the disk centered at z= —1 with radius r = |W (Jw)| does not include the inverse Nyquist plot of go(je)k (Jw) Finally, we like to remark that, due to the fact that the bounds for in- put and output signals have been stated in terms of their energy, all the Conditions for nominal performance are expressed in terms of the infinity norm. An equivalent formulation can be obtained when the input signals are taken as bounded persistant signals. In this case, the natural norm to measure the “size” of the inputs and the outputs is the time-domain £,, norm (i.e. the sup,). Since the La. {0 Lay induced norm is the L; operator norm (see ‘Appendix A), in this case the conditions for nominal performance will have the form of a weighted £; norm. The problem of optimizing this norm (C, ‘optimal control theory) will be addressed in Chapter 8. 2.6 ROBUST PERFORMANCE ‘The final goal of robust control is to achieve the performance requirement on all members of the family of models (j.e., robust performance), with a single controller. Clearly, robust stability and nominal performance are pre: requisites for achieving robust performance (why?). Next, we will establish a necesary and sufficient condition for robust performance, by making use of the conditions for nominal performance and robust stability. To this end we consider the case of performance defined as disturbance rejection at the out- ut for a family of models described by dynamic multiplicative uncertainty as represented in Figure 2.16, Definition2.6 The feedback loop of Figure 2.16 achieves robust performance if and only if |W,(s)y(s)ll2 <1, for all possible disturbances in the set {d € £| ld < 1}, for all inputs to the system equal to zero, and for all models in the set G = {g: g(s) = [1+ Wa(s)8]go(s), 8€C, [6] <1), Recall that the necessary and sufficient conditions for nominal perform- ance (2.43) and (2.14) and robust stability are parominal , | <> Wals)S(ly <1 (252) REY | <> W6(S)TUIl <1 253) ROBUST PERFORMANCE 81 $ System Figure 2.16. Disturbance ejection atthe output for afar of model wih multiplicative uncertainty, Robust performance is equivalent to condition (2.43), but with the nomi- nal model go(s) replaced by the family of models g = go [1+ Ws(s)8], 5 € , [3] < 1, We use this fact to obtain a necessary and sufficient condition for robust performance, ‘Thearem 2.3 A necessary and sufficient condition for robust performance of the family of models in Figure 2.16 is |1W2ue)Sue + )7Ge)Ws(r0)| |] <1 254) Proof. Replacing the nominal model ga(s) in (2.43) by the family of models £2(8) yields the following condition for robust performance: retort +e(s) wa(s)8))"I <1, |aj |Walyw)S(y0) 1+ Tyo)We(y0)8"| <1, [8)< 1 > [Wa(y0)SU«)] < 1+ TUW)We0)8], 18] <1, Yo => |Wa(soo)S{ya)| S1—|T(yo)We(yo)], Yo (2.55) = W0)S(9)| +I TGOWsK0)1 <1, Yo 256) > | mar0)S0)) + 17Gw)WoUe) | <2 esr) where fw) = ge(w)k(w). To prove the “if” part of (2.55), we consider 5 = 6. such that T(jw)We(sw)8, is real and negative, The “only if” partis immediate, 052 sisosYsTEMS Figure 2.17. Nyquis pot for bust performance interpretation, As before, a graphical interpretation of the robust performance condition can be obtained by means of the Nyquist plot of Figure 2.17. Note that (2.54) is equivalent to 11+ €(40)| ~ (Ws(@)€(G0)] + Waal) 20 (258) From the figure we see the robust performance requirement combines both the graphical conditions for robust stability of Figure 29 and nominal per formance of Figure 2.13. Robust performance is equivalent to the disk cen- tered at 2) = ~1 with radius ; = [W,(jo)] and the disk centered at z> = (ju) with radius ro = |Ws(jw)E(jw)] being disjoint. Clearly, this is more restrictive than achieving the robust stability and nominal performance conditions sep- arately. "Nevertheless, in the SISO case, by scaling the nominal performance and/or robust stability conditions conveniently, we can guarantee robust perform: ance. From the practical point of view, there is an important difference when scaling one or the other. Inthe first case, as we mentioned before, the condi- tion can be relaxed by changing the upper bound of the output to be hounded. In the second case, condition (2.14) is exclusive, therefore either we guaran. tee internal closed-loop stability for all members of the set or not. Sealing can be accomplished by means of a new “control-oriented” identification of the plant, yielding a new nominal model and a “smaller” uncertainty region, This isnot a trivial task, as will be seen in Chapter 10. ROBUST PERFORMANCE 53. ‘Therefore the usual practice is to scale performance, once condition (2.14) is achieved. This is performed by means of a parameter 7, as follows [71¥0e)sy0))+1W0e)T eI] <1 ye Wt 259) [As we have seen before, nominal performance and robust stability can be verified by computing the infinity norm of an appropriately weighted closed- loop transfer function. Synthesis of an optimal controller, which maximizes cither nominal performance or robust stability, isa wel established procedure, known as Hq, optimal control (see Chapter 6). However, the condition for robust performance (2.54) cannot be expressed in terms of a closed-loop weighted infinity norm. Rather, a new measure j., the structured singular value, must be used. This measure will be defined in Chapter 7, where we will indicate how to compute it, a procedure known as jeanalysis. The synthesis of controliers that optimize this measure, synthesis, will also be explained in the same chapter. In the next paragraph we briefly describe the relation between robust performance and structured uncertainty. Consider the block diagram in Figure 2.18, where two different types of uncertainty have been considered, Ws, and W,d2. The set of models involv- ing both types of uncertainties is the following % {so(s)[1 + Wals)d,] "(1 + Wols)&], 18 <1, i= 1,2} 2.60) Next, we obtain a condition for the robust stability of this family of mod- cls. The proof follows the one in Theorem 2.1, replacing the multiplicative uncertainty set by (2.60). Lemma 2.4 Robust performance of the family of models in Figure 2.16 is Wb Ws] Figure 2.18. Robust performance interpreted as robust stabity of sets with structured, uncertainty,54 siso systems equivalent to the robust stability of the family of models described by equation (2.60), shown in Figure 2.18. Proof, Robust stability of all members of the set G, is equivalent to 1+ €Uw) (1 + Weo)6i) [1 + Wee)8]' #0, Wor, VB) <1 || <1 = 1+ e(w) + £40)We(j0)6) + Wa0)6: #0, Yo, VIB <1 [6] <1 > [1+ G0) ~ [eGo)We0)] ~ Walo)| 20, Vo => |TQw)We(w)| + |Sw)Wew)| <1, Yo = |lroweol+iseymaef, <1 where £(s) = g(s)k(s) is the feedback loop. The above equation is identical to (254) o It follows that robust performance can alwayss be cast into robust stability of a set of models with structured uncertainty ((102]). Another interpretation is the following. The problem of rejecting a family of disturbances weighted by W,(s) can be incorporated as 4 new type of model uncertainty, as in the block diagram of Figure 2.18, As mentioned before, the measure over the closed-loop system corre- sponding to this problem is not a usual system norm. The computation of this new measure should take into account the structural information on the uncertainty of the system (see Chapter 7). The transformation from robust performance to structured robust stability can be represented by means of linear fractional transformations (LFT) as in Figure 2.19 (see Appendix B). In this case & is called the performance block. [5 3] P(e) Figure 2.19. Structured robus! stably wih extra performance Block hi EXIENSIONTO MIMO SYSTEMS 55 2.7. EXTENSION TO MIMO SYSTEMS. Concerning the extension of the above concepts to MIMO systems, we can state the following: « There is an equivalent concept for pole-zero cancellations, which is not straightforward as in SISO systems. © The infinity norm is also used to evaluate nominal performance and robust stability, by means of the singular values. «# The robust stability equivalent condition changes when the uncertainty in the model appears at the input (e.g., actuator uncertainty) or at the ‘output (¢., sensor uncertainty). The condition number of the nominal ‘model plays an important role in the former case. « In general, scaling by y the nominal performance condition, as a method to achieve robust performance, is not as effective as in the SISO case. ‘© The robust performance condition cannot be computed directly as in equation (2.54). Instead, a minimization problem should be solved. We conclude that the tools for the analysis of stability and performance robustness in MIMO systems are not a simple extension of the ones used in SISO systems, although we can find many similarities, ‘A usual procedure to operate with multivariable systems is to apply the tools for SISO models to each input-output pair of the MIMO system. We will show, by means of a simple example, that, in general, this is not correct. Example 2.9 In this example we show how the analysis by “loop at a time” ‘may lead to wrong results in the case of a MIMO control problem. Consider the following nominal plant and controller (see Figure 2.20): 5 ~ 2641) Gs) = 61) G+NG+2) 1 # 3 Theerain System Figure 2.20. Feedback loop of uncertain system56 SISO SvsTEMS l : K(s) 542 Zs+Div+2) (2.62) 1 For the above design, the nominal loop L(s) = G(s)K(s) and the complemen: tary sensitivity functions are uy tf, ° tft a 1 HOS ml ‘| eo) The “loop at a time” analysis is based on using the tools for SISO systems ‘on the problem. To this end, each loop is analyzed by means of the gain and phase margins, while the other loop is closed. This can be seen in Figure 2.21, where the first loop has been opened, while the second one remains closed. The equations in this case are the following: mao ae a es) Similarly, losing the first loop and evaluating the second one, we obtain y2 = ~us/s. As a consequence of this analysis, thee is an infinite gain margin and 90° phase margin in both loops. This analysis is based on the fact that the uncertainties in the system do not appear simultaneously in both channels. "Next, we show that with small uncertain gains in both channels, the closed-loop system can be unstable To this end, take the multiplicative uncertainty of Figure 2.20. If we con- sider a perturbed model (I + 8)G(s) instead of the nominal, the characteristic polynomial will be modified. Take, for example, an uncertainty of the form 4 A= G a (2.66) Figure 2.21. “Loop ata ime” anaiysis, PROBLEMS 87 The new characteristic polynomial of the closed-loop system is the following: s+ 5(24 8, +) + (148 +8) =0 (261) nis easy to verify that the following uncertainty destabilizes the closed loop: [29] as The perturbed model produces a closed-loop pole ats = 0. The “size” ofthe perturbation can be measured by means of a matrix norm. In particular, we may use the maximum singular value, which results in &(S.) = v2/2. We have considered muliplicariveuncertanty, therefore we might aswell ry to use he tools described in his chapter for SISO systems. To this end we mea sure the stability margin using the equivalent condition (2.14) (the weight in this case is W(s) = 1) If we use asthe spatial norm || for the complementary Sensitivity the maximum singular value, we obtain [Tih = 1618. Therefore the “smatles” (measured with the same norm) perturbation that destabilzes the loop ist inverse hati, (Sx) ~ 0618. Ths minimum perturbation can be found i we alow & tobe a fll complex matrix. Nevertheless, by using the special structure ofthe uncertainty, i can be shown tha the smallest destabilz ing perturbation i indeed (268), which has norm a(S.) ~ 2/2. This resul i nat as opimisicas we could have expected from the “loop tate” analysis thor i tas pessimistic as the analysis neglecting the special structure of. This is reasonable because we are now allowing a much ricer (and more practical) Set of uncertainties, which affect the plant model. We will sce in Chapter 4 that the equivalent condition (2.14) for robust stability sil holds for MIMO systems, by replacing the magnitude by a matrix 2.8 PROBLEMS 1. Prove that well posedness can be verified by considering only outputs e and ¢) of Figure 2.1, 2. Prove that to verify internal stability, it is enough to check the input- ‘output stability of the four transfer functions Fis) :u;—+e) with i,j = ea 3. Prove all entries of Table 2.1 4, Represent as a linear fractional transformation diagram ((M(s), A] struc: ture, see Appendix B), specifying M(s) and A, the following set of models: (a) A model g(s) with a positive feedback of 6.58 sisosystems (b) A model g(s) multiplied at the input and output by blocks (1 + 4) and (1 + 53), respectively. (€) A model g(s) multiplied at the input and output by blocks (1 + 6;) and (1 + 8:)-*, respectively In all cases there is a controller k(s) that stabilizes the nominal model. Is it possible to find necessary and sufficient conditions for robust stability in all the cases? 5. Consider the following family of models: : neg quo l 1 batet\t < spe yle {ake wai 0. Then a simple root-locus argu: ‘ment shows that in this case the set of stabilizing controllers of this form is not even connected, 2. Even in cases where the set of stabilizing controllers can be charac- terized, the resulting optimization problem is hard to solve since the objective function (3.1) is a nonconvex function of the optimization variable [in this case K(s)]. | 4 : a a | OSes ee eae, eal Axis Figure 3.2. Root locus iustrating « nonconnacted set of stabilizing controliors. WELL POSEDNESS AND INTERNALSTABIITY 63 Alternatively, we could attempt to use tools from optimal control theory such as LOR. However, tis will require recasting the performance specifi cations in terms of Q and R, the state and control weighting matrices used in the LOR performance index. Since there are no rules that will allow for exactly expressing specifications of the form (3.1) in terms of Q and R, in practice this leads to a “box-with-a-crank” approach, where the weighting matrices are adjusted through multiple trial-and-error type iterations In order to avoid these difficulties, we will parametize all the controllers that stabilize a given plant in terms of a free parameter Q(s) (a stable proper transfer matt). This parametrization, orginally developed by Wouls ba oe workers in the continuous time case (323, 324 (se also [94] for an alternative derivation) and, independently, by Kucera [178] for discrete time systems, is usually known as the Youla or YBJK parametrization, although sometimes itis also denoted as the Q-parametrization. A similar parametrization for ‘multidimensional systems was developed by Guiver and Bose [140] ‘The main result ofthe chapter shows that all the closed-loop mappings are affine functions ofthe parameter Q(s). Thus any convex specification given in terms of a closed-loop transfer function translates into a convex specification in Q(s), and the resulting synthesis problem reduces to a convex (albeit infinite-dimensional) optimization problem [58]. We will use these results in later chapters to solve the 743, lax, and £" optimal control problems. The fist step in achieving the parametrization isto extend the concepts of well posedness and internal stability introduced in Chapter 2 to the general case of MIMO plants 3.2. WELL POSEDNESS AND INTERNAL STABILITY ‘While the concepts of well posedness and internal stability of MIMO systems are very similar to those introduced in Chapter 2 for the SISO case, there are nevertheless some differences, mostly related to the tools used to assess whether or not a given loop exhibits these properties. In this section we briefly reexamine these concepts and introduce the appropriate changes. 3.2.1, Well Posedness For multivariable systems, the definition of well posedness is the same as in the scalar case (see Definition 21). Thus Lemma 2.1 can still be used to assess whether or not the feedback loop is well posed, provided that the scalar 1 there is replaced by the identity matrix F and in item 4, different from zer0 (#0) is interpreted as nonsingular. Nevertheless in spite the similarity with the SISO case, the luck of well posedness in MIMO systems may not be detected easily, as illustrated by the following example64 STABILIZATION Example 32 Consider the following plant model and controller: 32) ow) pale 1 rei The elements of the corresponding ouput sensitivity function S,(3) are all improper: (s#1) {s #1) [+ G)K()7 2 63) Go Se) xo =[4, 9] 6s the sensitivity function is not well defined, because + G(s)K(s) is singular for all's 3.2.2 Internal Stability ‘The same concept of internal stability presented in Definition 2.2, as well as Lemma 22, can be applied to the multivariable case. The loop intercon- nection shown in Figure 33 is internally stable if and only if all possible input-output transfer functions are stable. On the other hand, the general- ization of Lemma 2.3 is not straightforward. The difficulty here is that the concept of RHP pole-zero cancellation (as interpreted in SISO systems) in hg et as) w » K(®) anh, ws Figure 3.3. Loop interconnection to assess well posedness ond intornal stability, COFENLOOPSIABIE PLANTS 65 transfer matrices does not always compromise internal stability. In fact, entirely possible Tor a transfer matrix location of € (but not in the same input-output “channel"), as we illustrate with the following example. Example 33. Consider the plant +1 = 0 Gs) = s42 65) sel From the definitions of multivariable poles and zeros, we have that both are “simultaneously” located at 2; = py = —1 and 23 = py = -2. However, it can easily be verified that a realization containing the modes hy = ~1 and Ay = ~2 is both controllable and observable and therefore minimal. Moreover, a loop having G(s) as the closed-loop transfer function from (u; ug] to [ey €) is certainly internally sable. A generalization of Lemma 2.3 will be deferred until Section 3.7, where we introduce the concept of coprime factorizations. In the sequel, to develop 4 {state-space based) parametrization of all stabilizing controllers we oily need the following MIMO version of Lemma 22> cate [4h] 06 be a minimal state-space realization of the closed-loop system of Figure 3.3, between the inputs [ui(s),u2(s)} and outputs [ey(s),€2(8)} (or [ya(s),y2(6)]) Then the following conditions are equivalent: 1. The feedback loop is internally stable. 2. The eigenvalues of A. are in the open lefhalf complex plane C-. 3. The four transfer functions obtained between the ints js) 43(3)] and outputs |e;(s),e2(s)] (or yi (s),¥2(s)|) have their poles in C_ 4. The waner matic [2, “Ki inertbl in RP, Proof. The proof is virtually identical to the proof of Lemma 2.2, 0 3.3 OPEN-LOOP STABLE PLANTS In the case of open-loop stable plants, the set of all linear time invariant, finite-dimensional controllers that internally stabilize the plant has a simple expression, easily interpreted66 STABIUZATION ‘Theorem 3.1 Assume that the interconnection shown in Figure 3.3 is well posed and G(s) € RM, Then the set of all FDLTI controllers that internally stabilize the loop is given by K(s) = Q(s)1- GOUT" 6 Qls) € RA, det[f — Gloo)Q(oo)] #0 Proof. (=>) Assume that the interconnection is well posed and that K(s) internally stabilizes G(s). We need to show that there exists Q(s) € RNs. such that K(s) = O(s) |! ~ G(s)Q(s)]-*. To this effect define as) 2 By assumption, Q(s) is well defined and in RM... Solving equation (3.8) for K yields (8) = + KG) 'K = K(+ GK) Ga) K(s) = Q(s)|f — G(s) Q(s)}"' G9) (<) To complete the proof we need to show that any controller of the form (3.7) internally stabilizes the loop. Recall that checking internal stability is equivalent to checking that the four transfer matrices obtained between the inputs {ui(s),ua(s)] and the outputs [y1(5), ya(s)] in Figure 3.3 are in RH. Straightforward calculations using (3.7) yield vn] 2 [F-GQG Ge]fu [e)- ("28° [a] om ‘The proof now follows immediately from the fact that by assumption both G(s) and Q(s) are in RMLs. 0 A simple physical interpretation of the parametrization (37) can be ob- tained by elementary block manipulation techniques as follows. First note ae) Gls) Figure 3.4. Reolling O(3)1- G00)" OPENLOOPSTABLE PLANTS 67 2. a || a te) FY ate) | @ © Figure 3.5. Physica interpretation of the paromettization for stable plants that Q(s) {7 ~ G(s)Q(s)|" can be realized with the block diagram shown in Figure 3.4. Replacing the block K(s) in Figure 3.1 with this realization yields the block diagram in Figure 3.5a, Finally elementary loop manipulations lead to the block diagram in Figure 3.5b. From this last block diagram we can see that, in the case of a stable plant, the parametrization (3.7) amounts to can: celling out the plant (by subtracting in the feedback signal) and replacing it with any arbitrary stable dynamics Q(s), ‘This interpretation also highlights the reason why this approach will fail tor the case of open-loop unstable plants. While ~G(3) cancels the unstable Als) “Fi Gals) Figure 3.6. Presiabiluing an openoop unstable plant, Gay denotes the prestabilzed plant (46a) "6.68 STABILZATION dynamics G(s) in some of the paths (such as Ty), there are some others (©, Ty,4) Where this cancellation does not take place, yielding an unstable closed-loop transfer function. Thus a different approach is needed to deal with the general case where the plant is not necessarily open-loop stable. Note in passing that one may attempt to paramettize all the controllers that stabilize an open-loop unstable plant by first prestabilizing the plant and then applying the ra Figuré_36. However, it can be shown that while this approach generates stabilizing controllers, it does not generate all the controllers that stabilize the plant, unless the prestabilizing controller K,(s) is itself open-loop stable. 3.4 THE GENERAL CASE In this section we develop a parametrization ofall stabilizing controllers for the general case where the plant is not necessarily open-loop stable. The main result shows the existence of a separation-like structure, where all the stabilizing controllers have a familiar observer-based structure. The state- space approach that we use follows after the recent paper {185}, although the observer-based structure of all stabilizing controllers was already pointed out a5 early a5 1984 [103} In order to use the results of this section in later chapters, we consider the slightly mote general structure shown in Figure 3.7. Here the inputs to the plant have been partitioned into two groups: w, containing exogenous dissrbances and reference igmals anu the contol ingus Tikes, the ‘utputs have Beer partitioned inib z, containing outpuis subject to perform ance specifications, and y, containing the outputs available to the controller. Throughout this section we assume that the corresponding state-space real- w es) 2 K() Figure 3.7. the generalized pont THEGENERALCASE 69 ization of G(s) is given by sow A| BB v) Dn Dia Gan) yl G1] Dn Da yl Gl dn 0: where the pairs (A,B) and (C:,A) are stabilizable and detectable, respec- tively. Clearly, this assumption is required for the stabilization problem to be solvable. 3.4.1, Special Problems In the sequel we consider four special cases of the general structure (3.11), ‘commonly referred to as full information (FI), disturbance feedforward (DE), full control (FC), and output estimation desired parametrization for the general ouput feedback (OF) case will be obtained by combining these cases. Full information This problem corresponds to the case where the plant thas the following form: A | BB CG | Du Dy Ger=| oh | To) Te @.12) ol | ly lo Note that the outputs available to the controller include both the states of te plant and tie exogenous perturbation, that —— he plant and tife exogenous perturbation, _ fxn w= [30 hence the name. Rather than attempting to parametrize all the controllers that stabilize a given FI plant, in this subsection we obtain a weaker result, namely, a parametrization of all the possible control actions. Specifically, we define two controllers K and K’ to be equivalent (denoted K ~ K’) ifthe corresponding closed-loop matrices are identical, that K~K! = F(G,K)=FG.K) G3) ‘This equivalence relation induces a partition of all the stabilizing controllers into equivalence classes (an equivalence class is formed by all the controllers that generate the same control action). As we will see in the sequel, in order 0 obtain the set of all output feedback stabilizing controllers it suffices 10 obtain the set of all equivalence classes.70 STABILIZATION “ Lemma 3.2_ Let F be a constant feedback matrix such that A + BpF is stable. Then the set of all stabilizing FI controllers! can be parametrized as Kei(s)=|F Q(s)], QE Re G4) Proof, (=) Simple algebra shows that closing the loop with the controller (4) yields vy sf). Te) = TE) +806) [C+D | Div | n= espe] ME No [ A+BF | By 2 Oe Dak T Diz which is stable since A + B2F and Q are stable. (<) Assume now that a FI stabilizing controller Kp) = [Ky Kz] is given, To complete the proof we need to show that there exists Q € RM, such that Krr~[F Q)-To this effect, make the change of variable v =u ~ Fx, and let Q denote the closed-loop transfer function from w to v obtained when closing the loop with Ky. Since Key stabilizes Gr, 2 € RHay. Moreover, we have that un Peso = ke[3] uo =1F ai[5] Hence Kr, ~|F Q) o Disturbance Feedforward In this case, the plant has the following state space realization: f A| BB Gor(s)=2| Gr [Dn Dia G17) ylal| ro In addition, we assume that A ~ ByC; is stableAs we show next, with this additional hypothesis the FT and DF problems are equivalent. Lemma 33. Consider Gop and assume that A B\Cz is stable. Then the following results hold: 1. The feedback matrix Kpr(s) internally stabilizes Gpr if and only if Kop [CxT-intemally: stabilizes Gey. In addition, Fy(Gor, Kor Fu(Gry Kor(C_ 1) "nthe sense of generating all posible contol action, as dscussed before THE GENERALCASE 71 2. Given any feedback system Kyy(s) that stabilizes Gy,, then controller Koe = Fi(Pop, Krs) internally stabilizes Gor, where G.18) Moreover, Fe(Ger, Kes) = Fe (Gor, Fe(Por. Kei) Proof. Result | follows immediately from the relationship between the FI and DF problems, illustrated in Figure 38. For result 2, note that in this case the state and output equations can be solved to obtain fe] in terms of EE |. yielding @.19) Denote by é the states of the “observer” Py and by w the estimated value of the disturbance. From Figure 3.9 we have that a state-space realization of oF) cle |Grr(s)|_ Kor [oe 7) Figure 3.8. Relationship between the Fl and DF probiems.72 stanuzation Figure 3.9 Inlorconnocting Gor(s)ond Pos). the interconnection Gar—Por is given by A(t) = Ax(0) + Byw(t) + Byult) ee) = (A~ BCE) 2(0) = Crx()+ Durw(e) + Diault) 620) H(t) = x()—e@) (0) = Cree) + w(t) where ¢ 2 x ~ 2, Finally, eliminating the (stable) noncontrollable mode ¢(t) yields A| BB Du Da r}o0 0 ol 7 6 Thus F,(Gris Ker) = Fr [Fe(Gor.Por):Kei]. The desired result follows now from the fact that Fr (Fe(Gor,Por), Kei) = Fe(Gor,Fe(Por.Ken)] (see ‘Appendix B), o FuGor,Por) = = Gr 21) ‘This lemma shows that as long as the additional assumption concerning the stability of A ~ BC; holds, the FI and DF problems are equivalent. This ‘equivalence is exploited next to parametrize all the stabilizing controllers for the DF case. Note in passing that the proof of the lemma clearly shows the necessity of the additional assumption, since the matrix A — ByC> is precisely the error dynamics in (3.20). THEGENERALCASE 73 Combining Lemmas 32 and 33, we have that a class of stabilizing con- trollers for the DF problem is given by 3.22) where F is any matrix such that A + ByF is stable. Moreover, simple algebra shows that (3.22) can be rewritten nee Kor(s) = For, Q) OE RH» \..4¥** (3.23) A+ BAF ~BiC: | By By , Jor F oer *% (3.24) -G 70 © Indeed a stronger result holds: (3.23) parametrizes all the stabilizing DF controllers. This result is summarized in the following lemma, where Lemma 34 Assume that A ~ B,C3 is stable. Then (3.23) parametrizes all the controllers that internally stabilize Gr Proof, (=) The fact that Kpr(s) internally stabilizes Gpp follows immedi- ately from Lemmas 32 and 33. (4) Consider now a DF stabilizing controller Kpr. We need to show that Kpp can be written in the form (3.23) for some Q € RM. Let - A |B B Jor =| =F 0T ¥ 25) and define Q, = FeVne, Kor). Since Jor and Gpr share the same triple (A,Bz,C2) and since Kpr stabilizes Gpp, it follows that Qy € RH. More- over, by using the formulas for composition of LFTs given in Appendix B we have that FiUpr, Qo) = Fi [Jor, For, Kor)| = FeUimp, Kor), where A+ BrP ~ BC x _ =BiCe Jing = oT e ro A-BG ~BF tsk | 0? Aver hey - 0 =F why 0G Thus F)Upe, Qn) = FrUinp, Kor) = Kar o74 STABIUZATION Full Control in this case the plant has the following realization: A {l 0 Gre=|C | Du [OTT 8.27) G | Dy (0 0 The name stems from the fact that the control input wo= [283] affects independently the states x and the performance outputs z. Rather than obtaining directly a parametrization of all stabilizing con- rollers for this case, which would essentially follow the same steps of the previous two cases, we will exploit these results by using the concept of duality. given a plant G(s), its algebraic dual is obtained by simply transpos- ing it. Consider now the interconnection of Figure 3.7 and the duals of both the controller and the plant. It can easily be checked that, (FAG, K)}! = Fx(GT,KT) ) (328) internally stabilizes a po at_K internally sta t G if. and only if its dual KT stabilizes G™. Thus a parametrization of all controllers that stabilize a given plant G can be obtained by simply transposing the corresponding parametrization for its dual G”. In particular, comparing the realizations (3.12) and (3.27) itis immediate that the FI and FC problems are algebraic duals. It follows that @ parametrization of all FC stabilizing controllers can ‘be obtained by simply transposing (3.14). This result is summarized in the following lemma, Lemma 35. Let L be a constant feedback matrix such that A + LCz is stable. FC controllers? can be parametrized as 329) Output Estimation Finally, we consider the case where the plant has the realization 830) Note that if B; = 0, then this reduces to the problem of estimating the output z from the noisy measurements y, hence the name. nthe same sense asin the F cae a at 4 epeyrate Bi A=B:C| 6 [1 -B,) y Por = Cr 0 a (3.31) a 4 , \\ THEGENERALCASE 75 It is easily seen that the DF and OE estimation problems are algebraic duals. Thus the relationship between OE and FC is similar to that of the DF and FI problems. In particular, the following properties hold, Lemma 3.6 Assume that A — B3C, is stable. Then: 1. The feedback matrix Koe(s) internally stabilizes Goe(s) if and only if [Jw meaty sbi Got Moen hse have that F: (re [5] Kor) 2. Given any feedback matrix Kc(s) that stabilizes Grc(s), then controller Fe(Poe, Kec) internally stabilizes Gor(s), where Fi (Gor, Kor) G 1 [0 0} Blas Hdy (Goes Fe(Por.Kro)] Glee] 3. The set ofall internally stabilizing OE controllers can be parametrized a8 Kor = For: Q), Q € RH, where aa A+LG~ BC | L_-B or oT 3.32) GQ 10 3.4.2 The Output Feedback Case ‘We return now to the problem of parametrizing all the controllers that in- ternally stabilize the plant (3.11). This parametrization will be obtained by combining some of the four special cases discussed before, Jor = Theorem 32 Consider the general output feedback problem for the plant (3.11). Let F and L be constant matrices such that A+ LC; and A+ ByF are stable, Then all the controllers that internally stabilize G are given by K()=FU,Q), QE RM, det[l + DO(o0)] #0 (3.33) A+B E+LG+LDuk | -L_By+LDy Js F oT (Cr+ Dak) 1 Dn where (334)76 STARIUZATION Proof, For simplicity, we will assume for the time being that Dz> = 0. This, assumption will be removed at the end of the proof. Consider the state-space realization of G(s), which under this assumption will be denoted G(s): HO = Ax()+ Byw() + Boule) 2(0) = Cix(0) + Durw(0) + Drault) 3.35) Y(8) = Cax(0) + Dw Using the change of variable v= u— Fx, we have that (3.35) can be patti tioned into two subsystems as illustrated in Figure 3.10, where 5 [_AtBF | Bi Be oar a Br] Ca) By By Gump =| =F T 837) G | Duo and Since Gy is stable by choice of F, it follows that a controller K(s) internally stabilizes G(s) if and only ift stabilizes Gay. Moreover, Ging ha an op estimation structure, since A + BaF is stable by construction. Hence, from Lemma 36, we have that al the controllers that stabilize Ginp (and therefore K(s) Figure 3.10. Output feedback os « cascade of an OF and a stable stern THEGENERALCASE 77 J =| (338) oc | ‘To complete the proof we need to remove the assumption D22 = 0. This can be achieved through a standard loop shifting technique, consisting in using the change of variable pay-Daur This change of variable subtracts Dzau from the output of the plant and wraps a feedback Day around the controler, resulting in the struture shown in Figure 3.11, where G has Dz, =O. Hence the controller X is given by K = FeG,0), The original controller K can be recovered by wrapping the feedback ~Dzy around X as shown in Figure 3.12. Thus, inthe general OF case, we have that K = F,(J,Q), where simple algebra shows that [oe -L_By+LDn Je F 7 T “(C+ Dak) 1 -Dn 3.39) Finally, det(! + DzxO(ce)) #0 is a necessary and sulficient condition for the interconnection of Figure 312 to be well posed 0 Gs) G(s) * . ron Da kK) LAAT Ri) Figure 3.11. Loop shiting transtormation.78 STABILIZATION 4) Da « iy Je) as) Figure 3.12. Undoing the loop shiting lranstormation. 3.5 CONTROLLER STRUCTURE AND SEPARATION PRINCIPLE Consider again the parametrization of all OF stabilizing controllers developed in last section, K(s) = FJ, Q). A block diagram of this parametrization is shown in Figure 3.13, Note that J has a very familiar structure. Indeed, if ‘we neglect for a moment the additional output and input associated with Q (e and v), then J is precisely a classical observer-based output feedback controller, having F and L as the controller and observer gains, respectively. «| ow T() K(s) ae) Figure 3.13, Porometerzation of all sabilaing contolers. CLOSED-LOOP MAPPINGS 79 Thus the parametrization (3.33) can be interpreted as an observer-based con- troller, where a nominal classical observer is modified so that it produces an additional error signal ¢£ y — yon. Al stabilizing controllers are obtained ty Slering this eror with arbitrary sable dynamics Q and injecting the resul- tng signalo = Qe back into the observer. whee it gets aed tothe contol Sipnal u-~ Pi Recall from standard linear systems theory that the estimation es —— Sece no feedback, Sine the nominal controller stabilizes the plant and since is sable in hindsight i is aot surprising that the family of augmented Sonvlirs also intrmall stabiizes the plan The surprising est is the fact that this modified observer paradigm generates all stabiling controls “the existence ofa seperation property now apparent All outpat feed er —“ Cee edback (or F) problem to obiain a matt F such that A + BP is Saupe eatmation (or PC] poem in sian Lach at CIpREMIGL A TLC are stable. All controllers ate obtained by com: ihe tofu Thee problems to form the modified observer J and conngcting arbitrary stable dynamics Q between ¢ and v. Alternatively, this err ‘was obtained essentially using state-feedback and Ginp corresponds 10 an Ne rpoee eyes 3.6 CLOSED-LOOP MAPPINGS Recall from the example in the introduction that one of the reasons for seek- ing a parametrization of all stabilizing controllers is to recast performance specifications given in terms of closed-loop transfer functions into @ form having a simpler dependence on the free parameter (in this case Q). As we show next, an additional advantage of the parametrization is that the resulting closed-loop mapping are affine in Q. To see this, consider again the block di- gram of Figure 3.13, We have that Tew = Fe [G, FiJ,Q)] = Fe(T,Q), where renos[h th] Since, as pointed out before, ¢ is uncontrollable from v, Tap =0. Thus Te reduces to Tew = Tu + TQ ~ T2Q)"Tn = Ta + T2Ta. (3.40) ‘an affine function of Q. Simple algebra using the explicit expressions given in Appendix B yields the following state-space realization for T:80 STABILZATION Bh ea . BiskDy 0 ~ Di Da 3.7 A COPRIME FACTORIZATION APPROACH In this section we provide an alternative derivation of the Youla paramettiza- tion based on the use of coprime factorizations [94, 324, 314). Besides its historical importance, this approach allows for a simple derivation of some well known results on the simultaneous stabilization of multiple plants and the related problem of strong stabilization, that is, stabilization with stable controllers. 3.7.1 Coprime Factorizations Definition 3.1 Two transfer matrices M, and Ny in RMxs are right coprime over Re if there exist two matrices X, € RH and ¥, € RHx such that the following Bezout identity holds: X,M, + Y,)Ny (3.42) Similary, Mj, N; € RHao are said to be left coprime over RU. if there exist matrices X; € RHae and Y; € RH, such that MiXi+ MY, =1 G43) Remark From these definitions it can easily be shown that if M, and N; have a common right factor U, € RM (ie, Mp = MU, and N,=NU,), then Uz! € Rte. A similar property holds for let-coprime matrices, Thus coprimeness among elements of the ting of stable, proper transfer matrices is.a generalization of the SISO concept of no commrion RHP zeros between numerator aid denominator (|72, 314). Every proper, real rational transfer matrix G(s) can be factored as G(s) = N(s)M(s)"! (NO) G4) where N,M\N, and M are in Ra, and where N and M ate tight coprime and NV and M are left coprime. These factorizations are known as a right- coprime_and_a lefi-coprime factorization of Gis), respectively. In the next Jemma we provide a state-space based algorithm to obtain such factorizations starting from a stabilizable and detectable realization of G(s), ACOPRIMEFACIORUANONAPPROACH 8 Lemma 37° Ler Ge = | (345) be a stabilizable and detectable realization of G(s) and F and L any matrices such that A+ BF and A+LC are stable, Define Wit stone [AeEeb ese 1) aay Then (3:46) and (3:47) are right- and left-coprime factorizaions of G(s). respectively. [Me] = (3.46) NOs) and Proof. Since A+ BF and A + LC are Hurwitz by construction, it follows that all the transfer matrices defined in (3.45) and (3.47) are in RHa.. Moreover, it is easily seen that M and Af are not singular. Using the formulas for transfer ‘matrix inversion and multiplication given in Appendix B, we have that eo a ao and ctl Y A+SBF BF| B i N(S\M“4s) = - A “C¥DF DFT D AsBE 0] 0 oO A B | =G(s) C¥DF-=cTD_| where the lat equality was obtained by using the similarity transformation Smee ef sre vy a : a9) Pete tee we{ ee ———— itary ican easly be shown that M-"(0)N() ~ G(s) To show thatthe pats M,N and AV are right and left coprime, respectively, define (X46) Y= eS G50)82 STABILZATION and yun) Anette oe 0 es) WO" | cA) | 1 Cleary, these four matrices are in RH. Using the formulas for transfer ‘matrix multiplication yields (after some tedious algebra and eliminating un- ‘observable and uncontrollable modes) yu Xs) ¥e(s)] [M(s)_ ~¥ils) ro 8 HSI 28 1-[63] om RGIS Ra Hime, 0 mA Remark _A factorization of the form (3.52) is called a double-coprime factor- ization of the transfer matrix G(s). The state-space formulas presented here for computing double-coprime factorizations are due to Nett et al. (215}. Using these concepts we are ready now to present a MIMO extension of Lemma 23. Indes adnate foloush opine hactaatene Ope NON M~"(s)N(s), K(s) = Nu(s)M,'(s) = Mz'(s)Ng(s). Assuming well posedness, 1, The feedback loop is internally stable. MN, 2 [Me Gis] eimentiin mt My Ny 4 2 [Ae] mente cquivalent to the tanser mates [/) —*] being invertible in RPL. Expl [677-2 21 ay Using the (right) coprimeness of M,N and Mj, Nj, it can be shown that the 53) matrices Moo a [MM Oa ee ees are right coprime. i ACOPRIME FACTORIZATION APPROACH 83. Since M, My € RMx. and the matrices in the right-hand side of (3.53) only share common elements invertible in RM... it follows that: E K MN nes (3.54) | J erm “The equivalence of items 1 and 3 can be established proceeding in the same way using the let-coprime factorizations of G and K by noting that [6 K] erm — [% By] cen. cu tad (355) o Next, we show that given a stabilizable plant G(s) there exists an in- ternally stabilizing controller K,(3) = U,Vz! = Vz"U, such that the factors Up, Vo, Uo, and V, ae precisely the components of a double-coprime factor ization of G. This result constitutes the first step in the alternative derivation of the Youla paramettization. Lemma38__[f G(s) is stabilizable, and detectable then there existright-coprime {factors Up, Ve and lefi-coprime factors Ua, Vo in Rx. such that Ve Oo] [M Uv.) [1 0 -N M|[N ¥ or Moreover, the controller Kals) = Us(s\W=H(s) = Vs"s)to(s) internally stabilizes the loop. (836) Proof. The first part of the proof follows immediately from Lemma 3.7 by selecting Vp = X,, Uo = Y;, Vo = X;, and Uy = ¥;. To show that the con- troller K internally stabilizes the loop, note that with this choice of factors we have that : M -U]'_[V_ Up [x ve] =[5s ii)er Internal stability ofthe loop follows now from Theorem 33. ° G57) Lemma 38 provides a simple way of finding a stabilizing controller if a double-coprime factorization of the plant is known. As we show next, this ‘controller is exactly the familiar observer-based controller obtained by cas- cading an observer with gain L with a state-feedback gain F. A state-space realization of such an observer-based controller is given by (see Figure 3.14) a = At+ Bu+L iy -(-y)) = Ck+ Du (358) ua Fe84 STABILZATION a) = y He Ys Bod 2 FOO A+ BP , \c-+ pr| re xe) o ow) Hrht s Bods = - i L A+ BF i en , A+ BF : c+ DF | ) 0) - . Figure 3.14, The controller of lemma 38 as an observer-based controller: (a) standard realization ond (b) factored as UaV! [ACOPRIME FACTORIZATION APPROACH 85 Define ¢ © j + and replace w by the last equation: ASBF | L w= Atle = Use 639) AsBF_[L (C: TIS = Vo(sle Ts = Us(s}e = Uals)Volsy!y (G.60) which coincides with the controller obtained in Lemma 38, Next, we use the results of Theorem 33 and Lemma 38 to obtain an alternative derivation of the Youla parametrization, Theorem 3.4 The set of all controllers that internally stabilize the plant G(s) can be paramernzed as fg _ayg)(1ANO K() 1) The fact that U-U =F implies that |Uazlho £1 (since Uss is a compression of U and matrix compressions do not increase the norm). Since ||Q].o <1 then a simple small gain argument shows that (I - U@)-! € RH. Since both U,Q € RH, well posedness and internal stability follow immediately. To show that l|Tvullm <1, note that, since UF is inner, then WEIE-16) Hel $< wi ~~ HON IMI (3.78) Since || Qlao < 1. it follows that [lz < [lw], or equivalently [Tell <1 (1 = 2) Assume now that the interconnection is internally stable and such that [[Tz. fs <1. To show that Qj <1 assume to the contrary that there exists a frequency «, and a constant vector f such that 77) bh be which implies OCian)s 2 Ill (3.79)